Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Budget Fair Queueing (BFQ) I/O scheduler.
4 *
5 * Based on ideas and code from CFQ:
6 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7 *
8 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9 * Paolo Valente <paolo.valente@unimore.it>
10 *
11 * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
12 * Arianna Avanzini <avanzini@google.com>
13 *
14 * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
15 *
16 * BFQ is a proportional-share I/O scheduler, with some extra
17 * low-latency capabilities. BFQ also supports full hierarchical
18 * scheduling through cgroups. Next paragraphs provide an introduction
19 * on BFQ inner workings. Details on BFQ benefits, usage and
20 * limitations can be found in Documentation/block/bfq-iosched.rst.
21 *
22 * BFQ is a proportional-share storage-I/O scheduling algorithm based
23 * on the slice-by-slice service scheme of CFQ. But BFQ assigns
24 * budgets, measured in number of sectors, to processes instead of
25 * time slices. The device is not granted to the in-service process
26 * for a given time slice, but until it has exhausted its assigned
27 * budget. This change from the time to the service domain enables BFQ
28 * to distribute the device throughput among processes as desired,
29 * without any distortion due to throughput fluctuations, or to device
30 * internal queueing. BFQ uses an ad hoc internal scheduler, called
31 * B-WF2Q+, to schedule processes according to their budgets. More
32 * precisely, BFQ schedules queues associated with processes. Each
33 * process/queue is assigned a user-configurable weight, and B-WF2Q+
34 * guarantees that each queue receives a fraction of the throughput
35 * proportional to its weight. Thanks to the accurate policy of
36 * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
37 * processes issuing sequential requests (to boost the throughput),
38 * and yet guarantee a low latency to interactive and soft real-time
39 * applications.
40 *
41 * In particular, to provide these low-latency guarantees, BFQ
42 * explicitly privileges the I/O of two classes of time-sensitive
43 * applications: interactive and soft real-time. In more detail, BFQ
44 * behaves this way if the low_latency parameter is set (default
45 * configuration). This feature enables BFQ to provide applications in
46 * these classes with a very low latency.
47 *
48 * To implement this feature, BFQ constantly tries to detect whether
49 * the I/O requests in a bfq_queue come from an interactive or a soft
50 * real-time application. For brevity, in these cases, the queue is
51 * said to be interactive or soft real-time. In both cases, BFQ
52 * privileges the service of the queue, over that of non-interactive
53 * and non-soft-real-time queues. This privileging is performed,
54 * mainly, by raising the weight of the queue. So, for brevity, we
55 * call just weight-raising periods the time periods during which a
56 * queue is privileged, because deemed interactive or soft real-time.
57 *
58 * The detection of soft real-time queues/applications is described in
59 * detail in the comments on the function
60 * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
61 * interactive queue works as follows: a queue is deemed interactive
62 * if it is constantly non empty only for a limited time interval,
63 * after which it does become empty. The queue may be deemed
64 * interactive again (for a limited time), if it restarts being
65 * constantly non empty, provided that this happens only after the
66 * queue has remained empty for a given minimum idle time.
67 *
68 * By default, BFQ computes automatically the above maximum time
69 * interval, i.e., the time interval after which a constantly
70 * non-empty queue stops being deemed interactive. Since a queue is
71 * weight-raised while it is deemed interactive, this maximum time
72 * interval happens to coincide with the (maximum) duration of the
73 * weight-raising for interactive queues.
74 *
75 * Finally, BFQ also features additional heuristics for
76 * preserving both a low latency and a high throughput on NCQ-capable,
77 * rotational or flash-based devices, and to get the job done quickly
78 * for applications consisting in many I/O-bound processes.
79 *
80 * NOTE: if the main or only goal, with a given device, is to achieve
81 * the maximum-possible throughput at all times, then do switch off
82 * all low-latency heuristics for that device, by setting low_latency
83 * to 0.
84 *
85 * BFQ is described in [1], where also a reference to the initial,
86 * more theoretical paper on BFQ can be found. The interested reader
87 * can find in the latter paper full details on the main algorithm, as
88 * well as formulas of the guarantees and formal proofs of all the
89 * properties. With respect to the version of BFQ presented in these
90 * papers, this implementation adds a few more heuristics, such as the
91 * ones that guarantee a low latency to interactive and soft real-time
92 * applications, and a hierarchical extension based on H-WF2Q+.
93 *
94 * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
95 * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
96 * with O(log N) complexity derives from the one introduced with EEVDF
97 * in [3].
98 *
99 * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
100 * Scheduler", Proceedings of the First Workshop on Mobile System
101 * Technologies (MST-2015), May 2015.
102 * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
103 *
104 * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
105 * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
106 * Oct 1997.
107 *
108 * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
109 *
110 * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
111 * First: A Flexible and Accurate Mechanism for Proportional Share
112 * Resource Allocation", technical report.
113 *
114 * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
115 */
116#include <linux/module.h>
117#include <linux/slab.h>
118#include <linux/blkdev.h>
119#include <linux/cgroup.h>
120#include <linux/ktime.h>
121#include <linux/rbtree.h>
122#include <linux/ioprio.h>
123#include <linux/sbitmap.h>
124#include <linux/delay.h>
125#include <linux/backing-dev.h>
126
127#include <trace/events/block.h>
128
129#include "elevator.h"
130#include "blk.h"
131#include "blk-mq.h"
132#include "blk-mq-tag.h"
133#include "blk-mq-sched.h"
134#include "bfq-iosched.h"
135#include "blk-wbt.h"
136
137#define BFQ_BFQQ_FNS(name) \
138void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
139{ \
140 __set_bit(BFQQF_##name, &(bfqq)->flags); \
141} \
142void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
143{ \
144 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
145} \
146int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
147{ \
148 return test_bit(BFQQF_##name, &(bfqq)->flags); \
149}
150
151BFQ_BFQQ_FNS(just_created);
152BFQ_BFQQ_FNS(busy);
153BFQ_BFQQ_FNS(wait_request);
154BFQ_BFQQ_FNS(non_blocking_wait_rq);
155BFQ_BFQQ_FNS(fifo_expire);
156BFQ_BFQQ_FNS(has_short_ttime);
157BFQ_BFQQ_FNS(sync);
158BFQ_BFQQ_FNS(IO_bound);
159BFQ_BFQQ_FNS(in_large_burst);
160BFQ_BFQQ_FNS(coop);
161BFQ_BFQQ_FNS(split_coop);
162BFQ_BFQQ_FNS(softrt_update);
163#undef BFQ_BFQQ_FNS \
164
165/* Expiration time of async (0) and sync (1) requests, in ns. */
166static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
167
168/* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
169static const int bfq_back_max = 16 * 1024;
170
171/* Penalty of a backwards seek, in number of sectors. */
172static const int bfq_back_penalty = 2;
173
174/* Idling period duration, in ns. */
175static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
176
177/* Minimum number of assigned budgets for which stats are safe to compute. */
178static const int bfq_stats_min_budgets = 194;
179
180/* Default maximum budget values, in sectors and number of requests. */
181static const int bfq_default_max_budget = 16 * 1024;
182
183/*
184 * When a sync request is dispatched, the queue that contains that
185 * request, and all the ancestor entities of that queue, are charged
186 * with the number of sectors of the request. In contrast, if the
187 * request is async, then the queue and its ancestor entities are
188 * charged with the number of sectors of the request, multiplied by
189 * the factor below. This throttles the bandwidth for async I/O,
190 * w.r.t. to sync I/O, and it is done to counter the tendency of async
191 * writes to steal I/O throughput to reads.
192 *
193 * The current value of this parameter is the result of a tuning with
194 * several hardware and software configurations. We tried to find the
195 * lowest value for which writes do not cause noticeable problems to
196 * reads. In fact, the lower this parameter, the stabler I/O control,
197 * in the following respect. The lower this parameter is, the less
198 * the bandwidth enjoyed by a group decreases
199 * - when the group does writes, w.r.t. to when it does reads;
200 * - when other groups do reads, w.r.t. to when they do writes.
201 */
202static const int bfq_async_charge_factor = 3;
203
204/* Default timeout values, in jiffies, approximating CFQ defaults. */
205const int bfq_timeout = HZ / 8;
206
207/*
208 * Time limit for merging (see comments in bfq_setup_cooperator). Set
209 * to the slowest value that, in our tests, proved to be effective in
210 * removing false positives, while not causing true positives to miss
211 * queue merging.
212 *
213 * As can be deduced from the low time limit below, queue merging, if
214 * successful, happens at the very beginning of the I/O of the involved
215 * cooperating processes, as a consequence of the arrival of the very
216 * first requests from each cooperator. After that, there is very
217 * little chance to find cooperators.
218 */
219static const unsigned long bfq_merge_time_limit = HZ/10;
220
221static struct kmem_cache *bfq_pool;
222
223/* Below this threshold (in ns), we consider thinktime immediate. */
224#define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
225
226/* hw_tag detection: parallel requests threshold and min samples needed. */
227#define BFQ_HW_QUEUE_THRESHOLD 3
228#define BFQ_HW_QUEUE_SAMPLES 32
229
230#define BFQQ_SEEK_THR (sector_t)(8 * 100)
231#define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
232#define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \
233 (get_sdist(last_pos, rq) > \
234 BFQQ_SEEK_THR && \
235 (!blk_queue_nonrot(bfqd->queue) || \
236 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT))
237#define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
238#define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19)
239/*
240 * Sync random I/O is likely to be confused with soft real-time I/O,
241 * because it is characterized by limited throughput and apparently
242 * isochronous arrival pattern. To avoid false positives, queues
243 * containing only random (seeky) I/O are prevented from being tagged
244 * as soft real-time.
245 */
246#define BFQQ_TOTALLY_SEEKY(bfqq) (bfqq->seek_history == -1)
247
248/* Min number of samples required to perform peak-rate update */
249#define BFQ_RATE_MIN_SAMPLES 32
250/* Min observation time interval required to perform a peak-rate update (ns) */
251#define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
252/* Target observation time interval for a peak-rate update (ns) */
253#define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
254
255/*
256 * Shift used for peak-rate fixed precision calculations.
257 * With
258 * - the current shift: 16 positions
259 * - the current type used to store rate: u32
260 * - the current unit of measure for rate: [sectors/usec], or, more precisely,
261 * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
262 * the range of rates that can be stored is
263 * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
264 * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
265 * [15, 65G] sectors/sec
266 * Which, assuming a sector size of 512B, corresponds to a range of
267 * [7.5K, 33T] B/sec
268 */
269#define BFQ_RATE_SHIFT 16
270
271/*
272 * When configured for computing the duration of the weight-raising
273 * for interactive queues automatically (see the comments at the
274 * beginning of this file), BFQ does it using the following formula:
275 * duration = (ref_rate / r) * ref_wr_duration,
276 * where r is the peak rate of the device, and ref_rate and
277 * ref_wr_duration are two reference parameters. In particular,
278 * ref_rate is the peak rate of the reference storage device (see
279 * below), and ref_wr_duration is about the maximum time needed, with
280 * BFQ and while reading two files in parallel, to load typical large
281 * applications on the reference device (see the comments on
282 * max_service_from_wr below, for more details on how ref_wr_duration
283 * is obtained). In practice, the slower/faster the device at hand
284 * is, the more/less it takes to load applications with respect to the
285 * reference device. Accordingly, the longer/shorter BFQ grants
286 * weight raising to interactive applications.
287 *
288 * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
289 * depending on whether the device is rotational or non-rotational.
290 *
291 * In the following definitions, ref_rate[0] and ref_wr_duration[0]
292 * are the reference values for a rotational device, whereas
293 * ref_rate[1] and ref_wr_duration[1] are the reference values for a
294 * non-rotational device. The reference rates are not the actual peak
295 * rates of the devices used as a reference, but slightly lower
296 * values. The reason for using slightly lower values is that the
297 * peak-rate estimator tends to yield slightly lower values than the
298 * actual peak rate (it can yield the actual peak rate only if there
299 * is only one process doing I/O, and the process does sequential
300 * I/O).
301 *
302 * The reference peak rates are measured in sectors/usec, left-shifted
303 * by BFQ_RATE_SHIFT.
304 */
305static int ref_rate[2] = {14000, 33000};
306/*
307 * To improve readability, a conversion function is used to initialize
308 * the following array, which entails that the array can be
309 * initialized only in a function.
310 */
311static int ref_wr_duration[2];
312
313/*
314 * BFQ uses the above-detailed, time-based weight-raising mechanism to
315 * privilege interactive tasks. This mechanism is vulnerable to the
316 * following false positives: I/O-bound applications that will go on
317 * doing I/O for much longer than the duration of weight
318 * raising. These applications have basically no benefit from being
319 * weight-raised at the beginning of their I/O. On the opposite end,
320 * while being weight-raised, these applications
321 * a) unjustly steal throughput to applications that may actually need
322 * low latency;
323 * b) make BFQ uselessly perform device idling; device idling results
324 * in loss of device throughput with most flash-based storage, and may
325 * increase latencies when used purposelessly.
326 *
327 * BFQ tries to reduce these problems, by adopting the following
328 * countermeasure. To introduce this countermeasure, we need first to
329 * finish explaining how the duration of weight-raising for
330 * interactive tasks is computed.
331 *
332 * For a bfq_queue deemed as interactive, the duration of weight
333 * raising is dynamically adjusted, as a function of the estimated
334 * peak rate of the device, so as to be equal to the time needed to
335 * execute the 'largest' interactive task we benchmarked so far. By
336 * largest task, we mean the task for which each involved process has
337 * to do more I/O than for any of the other tasks we benchmarked. This
338 * reference interactive task is the start-up of LibreOffice Writer,
339 * and in this task each process/bfq_queue needs to have at most ~110K
340 * sectors transferred.
341 *
342 * This last piece of information enables BFQ to reduce the actual
343 * duration of weight-raising for at least one class of I/O-bound
344 * applications: those doing sequential or quasi-sequential I/O. An
345 * example is file copy. In fact, once started, the main I/O-bound
346 * processes of these applications usually consume the above 110K
347 * sectors in much less time than the processes of an application that
348 * is starting, because these I/O-bound processes will greedily devote
349 * almost all their CPU cycles only to their target,
350 * throughput-friendly I/O operations. This is even more true if BFQ
351 * happens to be underestimating the device peak rate, and thus
352 * overestimating the duration of weight raising. But, according to
353 * our measurements, once transferred 110K sectors, these processes
354 * have no right to be weight-raised any longer.
355 *
356 * Basing on the last consideration, BFQ ends weight-raising for a
357 * bfq_queue if the latter happens to have received an amount of
358 * service at least equal to the following constant. The constant is
359 * set to slightly more than 110K, to have a minimum safety margin.
360 *
361 * This early ending of weight-raising reduces the amount of time
362 * during which interactive false positives cause the two problems
363 * described at the beginning of these comments.
364 */
365static const unsigned long max_service_from_wr = 120000;
366
367/*
368 * Maximum time between the creation of two queues, for stable merge
369 * to be activated (in ms)
370 */
371static const unsigned long bfq_activation_stable_merging = 600;
372/*
373 * Minimum time to be waited before evaluating delayed stable merge (in ms)
374 */
375static const unsigned long bfq_late_stable_merging = 600;
376
377#define RQ_BIC(rq) ((struct bfq_io_cq *)((rq)->elv.priv[0]))
378#define RQ_BFQQ(rq) ((rq)->elv.priv[1])
379
380struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
381{
382 return bic->bfqq[is_sync];
383}
384
385static void bfq_put_stable_ref(struct bfq_queue *bfqq);
386
387void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
388{
389 struct bfq_queue *old_bfqq = bic->bfqq[is_sync];
390
391 /* Clear bic pointer if bfqq is detached from this bic */
392 if (old_bfqq && old_bfqq->bic == bic)
393 old_bfqq->bic = NULL;
394
395 /*
396 * If bfqq != NULL, then a non-stable queue merge between
397 * bic->bfqq and bfqq is happening here. This causes troubles
398 * in the following case: bic->bfqq has also been scheduled
399 * for a possible stable merge with bic->stable_merge_bfqq,
400 * and bic->stable_merge_bfqq == bfqq happens to
401 * hold. Troubles occur because bfqq may then undergo a split,
402 * thereby becoming eligible for a stable merge. Yet, if
403 * bic->stable_merge_bfqq points exactly to bfqq, then bfqq
404 * would be stably merged with itself. To avoid this anomaly,
405 * we cancel the stable merge if
406 * bic->stable_merge_bfqq == bfqq.
407 */
408 bic->bfqq[is_sync] = bfqq;
409
410 if (bfqq && bic->stable_merge_bfqq == bfqq) {
411 /*
412 * Actually, these same instructions are executed also
413 * in bfq_setup_cooperator, in case of abort or actual
414 * execution of a stable merge. We could avoid
415 * repeating these instructions there too, but if we
416 * did so, we would nest even more complexity in this
417 * function.
418 */
419 bfq_put_stable_ref(bic->stable_merge_bfqq);
420
421 bic->stable_merge_bfqq = NULL;
422 }
423}
424
425struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
426{
427 return bic->icq.q->elevator->elevator_data;
428}
429
430/**
431 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
432 * @icq: the iocontext queue.
433 */
434static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
435{
436 /* bic->icq is the first member, %NULL will convert to %NULL */
437 return container_of(icq, struct bfq_io_cq, icq);
438}
439
440/**
441 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
442 * @q: the request queue.
443 */
444static struct bfq_io_cq *bfq_bic_lookup(struct request_queue *q)
445{
446 struct bfq_io_cq *icq;
447 unsigned long flags;
448
449 if (!current->io_context)
450 return NULL;
451
452 spin_lock_irqsave(&q->queue_lock, flags);
453 icq = icq_to_bic(ioc_lookup_icq(q));
454 spin_unlock_irqrestore(&q->queue_lock, flags);
455
456 return icq;
457}
458
459/*
460 * Scheduler run of queue, if there are requests pending and no one in the
461 * driver that will restart queueing.
462 */
463void bfq_schedule_dispatch(struct bfq_data *bfqd)
464{
465 lockdep_assert_held(&bfqd->lock);
466
467 if (bfqd->queued != 0) {
468 bfq_log(bfqd, "schedule dispatch");
469 blk_mq_run_hw_queues(bfqd->queue, true);
470 }
471}
472
473#define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
474
475#define bfq_sample_valid(samples) ((samples) > 80)
476
477/*
478 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
479 * We choose the request that is closer to the head right now. Distance
480 * behind the head is penalized and only allowed to a certain extent.
481 */
482static struct request *bfq_choose_req(struct bfq_data *bfqd,
483 struct request *rq1,
484 struct request *rq2,
485 sector_t last)
486{
487 sector_t s1, s2, d1 = 0, d2 = 0;
488 unsigned long back_max;
489#define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
490#define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
491 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
492
493 if (!rq1 || rq1 == rq2)
494 return rq2;
495 if (!rq2)
496 return rq1;
497
498 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
499 return rq1;
500 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
501 return rq2;
502 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
503 return rq1;
504 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
505 return rq2;
506
507 s1 = blk_rq_pos(rq1);
508 s2 = blk_rq_pos(rq2);
509
510 /*
511 * By definition, 1KiB is 2 sectors.
512 */
513 back_max = bfqd->bfq_back_max * 2;
514
515 /*
516 * Strict one way elevator _except_ in the case where we allow
517 * short backward seeks which are biased as twice the cost of a
518 * similar forward seek.
519 */
520 if (s1 >= last)
521 d1 = s1 - last;
522 else if (s1 + back_max >= last)
523 d1 = (last - s1) * bfqd->bfq_back_penalty;
524 else
525 wrap |= BFQ_RQ1_WRAP;
526
527 if (s2 >= last)
528 d2 = s2 - last;
529 else if (s2 + back_max >= last)
530 d2 = (last - s2) * bfqd->bfq_back_penalty;
531 else
532 wrap |= BFQ_RQ2_WRAP;
533
534 /* Found required data */
535
536 /*
537 * By doing switch() on the bit mask "wrap" we avoid having to
538 * check two variables for all permutations: --> faster!
539 */
540 switch (wrap) {
541 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
542 if (d1 < d2)
543 return rq1;
544 else if (d2 < d1)
545 return rq2;
546
547 if (s1 >= s2)
548 return rq1;
549 else
550 return rq2;
551
552 case BFQ_RQ2_WRAP:
553 return rq1;
554 case BFQ_RQ1_WRAP:
555 return rq2;
556 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
557 default:
558 /*
559 * Since both rqs are wrapped,
560 * start with the one that's further behind head
561 * (--> only *one* back seek required),
562 * since back seek takes more time than forward.
563 */
564 if (s1 <= s2)
565 return rq1;
566 else
567 return rq2;
568 }
569}
570
571#define BFQ_LIMIT_INLINE_DEPTH 16
572
573#ifdef CONFIG_BFQ_GROUP_IOSCHED
574static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
575{
576 struct bfq_data *bfqd = bfqq->bfqd;
577 struct bfq_entity *entity = &bfqq->entity;
578 struct bfq_entity *inline_entities[BFQ_LIMIT_INLINE_DEPTH];
579 struct bfq_entity **entities = inline_entities;
580 int depth, level, alloc_depth = BFQ_LIMIT_INLINE_DEPTH;
581 int class_idx = bfqq->ioprio_class - 1;
582 struct bfq_sched_data *sched_data;
583 unsigned long wsum;
584 bool ret = false;
585
586 if (!entity->on_st_or_in_serv)
587 return false;
588
589retry:
590 spin_lock_irq(&bfqd->lock);
591 /* +1 for bfqq entity, root cgroup not included */
592 depth = bfqg_to_blkg(bfqq_group(bfqq))->blkcg->css.cgroup->level + 1;
593 if (depth > alloc_depth) {
594 spin_unlock_irq(&bfqd->lock);
595 if (entities != inline_entities)
596 kfree(entities);
597 entities = kmalloc_array(depth, sizeof(*entities), GFP_NOIO);
598 if (!entities)
599 return false;
600 alloc_depth = depth;
601 goto retry;
602 }
603
604 sched_data = entity->sched_data;
605 /* Gather our ancestors as we need to traverse them in reverse order */
606 level = 0;
607 for_each_entity(entity) {
608 /*
609 * If at some level entity is not even active, allow request
610 * queueing so that BFQ knows there's work to do and activate
611 * entities.
612 */
613 if (!entity->on_st_or_in_serv)
614 goto out;
615 /* Uh, more parents than cgroup subsystem thinks? */
616 if (WARN_ON_ONCE(level >= depth))
617 break;
618 entities[level++] = entity;
619 }
620 WARN_ON_ONCE(level != depth);
621 for (level--; level >= 0; level--) {
622 entity = entities[level];
623 if (level > 0) {
624 wsum = bfq_entity_service_tree(entity)->wsum;
625 } else {
626 int i;
627 /*
628 * For bfqq itself we take into account service trees
629 * of all higher priority classes and multiply their
630 * weights so that low prio queue from higher class
631 * gets more requests than high prio queue from lower
632 * class.
633 */
634 wsum = 0;
635 for (i = 0; i <= class_idx; i++) {
636 wsum = wsum * IOPRIO_BE_NR +
637 sched_data->service_tree[i].wsum;
638 }
639 }
640 limit = DIV_ROUND_CLOSEST(limit * entity->weight, wsum);
641 if (entity->allocated >= limit) {
642 bfq_log_bfqq(bfqq->bfqd, bfqq,
643 "too many requests: allocated %d limit %d level %d",
644 entity->allocated, limit, level);
645 ret = true;
646 break;
647 }
648 }
649out:
650 spin_unlock_irq(&bfqd->lock);
651 if (entities != inline_entities)
652 kfree(entities);
653 return ret;
654}
655#else
656static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
657{
658 return false;
659}
660#endif
661
662/*
663 * Async I/O can easily starve sync I/O (both sync reads and sync
664 * writes), by consuming all tags. Similarly, storms of sync writes,
665 * such as those that sync(2) may trigger, can starve sync reads.
666 * Limit depths of async I/O and sync writes so as to counter both
667 * problems.
668 *
669 * Also if a bfq queue or its parent cgroup consume more tags than would be
670 * appropriate for their weight, we trim the available tag depth to 1. This
671 * avoids a situation where one cgroup can starve another cgroup from tags and
672 * thus block service differentiation among cgroups. Note that because the
673 * queue / cgroup already has many requests allocated and queued, this does not
674 * significantly affect service guarantees coming from the BFQ scheduling
675 * algorithm.
676 */
677static void bfq_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
678{
679 struct bfq_data *bfqd = data->q->elevator->elevator_data;
680 struct bfq_io_cq *bic = bfq_bic_lookup(data->q);
681 struct bfq_queue *bfqq = bic ? bic_to_bfqq(bic, op_is_sync(opf)) : NULL;
682 int depth;
683 unsigned limit = data->q->nr_requests;
684
685 /* Sync reads have full depth available */
686 if (op_is_sync(opf) && !op_is_write(opf)) {
687 depth = 0;
688 } else {
689 depth = bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(opf)];
690 limit = (limit * depth) >> bfqd->full_depth_shift;
691 }
692
693 /*
694 * Does queue (or any parent entity) exceed number of requests that
695 * should be available to it? Heavily limit depth so that it cannot
696 * consume more available requests and thus starve other entities.
697 */
698 if (bfqq && bfqq_request_over_limit(bfqq, limit))
699 depth = 1;
700
701 bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
702 __func__, bfqd->wr_busy_queues, op_is_sync(opf), depth);
703 if (depth)
704 data->shallow_depth = depth;
705}
706
707static struct bfq_queue *
708bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
709 sector_t sector, struct rb_node **ret_parent,
710 struct rb_node ***rb_link)
711{
712 struct rb_node **p, *parent;
713 struct bfq_queue *bfqq = NULL;
714
715 parent = NULL;
716 p = &root->rb_node;
717 while (*p) {
718 struct rb_node **n;
719
720 parent = *p;
721 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
722
723 /*
724 * Sort strictly based on sector. Smallest to the left,
725 * largest to the right.
726 */
727 if (sector > blk_rq_pos(bfqq->next_rq))
728 n = &(*p)->rb_right;
729 else if (sector < blk_rq_pos(bfqq->next_rq))
730 n = &(*p)->rb_left;
731 else
732 break;
733 p = n;
734 bfqq = NULL;
735 }
736
737 *ret_parent = parent;
738 if (rb_link)
739 *rb_link = p;
740
741 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
742 (unsigned long long)sector,
743 bfqq ? bfqq->pid : 0);
744
745 return bfqq;
746}
747
748static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
749{
750 return bfqq->service_from_backlogged > 0 &&
751 time_is_before_jiffies(bfqq->first_IO_time +
752 bfq_merge_time_limit);
753}
754
755/*
756 * The following function is not marked as __cold because it is
757 * actually cold, but for the same performance goal described in the
758 * comments on the likely() at the beginning of
759 * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
760 * execution time for the case where this function is not invoked, we
761 * had to add an unlikely() in each involved if().
762 */
763void __cold
764bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
765{
766 struct rb_node **p, *parent;
767 struct bfq_queue *__bfqq;
768
769 if (bfqq->pos_root) {
770 rb_erase(&bfqq->pos_node, bfqq->pos_root);
771 bfqq->pos_root = NULL;
772 }
773
774 /* oom_bfqq does not participate in queue merging */
775 if (bfqq == &bfqd->oom_bfqq)
776 return;
777
778 /*
779 * bfqq cannot be merged any longer (see comments in
780 * bfq_setup_cooperator): no point in adding bfqq into the
781 * position tree.
782 */
783 if (bfq_too_late_for_merging(bfqq))
784 return;
785
786 if (bfq_class_idle(bfqq))
787 return;
788 if (!bfqq->next_rq)
789 return;
790
791 bfqq->pos_root = &bfqq_group(bfqq)->rq_pos_tree;
792 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
793 blk_rq_pos(bfqq->next_rq), &parent, &p);
794 if (!__bfqq) {
795 rb_link_node(&bfqq->pos_node, parent, p);
796 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
797 } else
798 bfqq->pos_root = NULL;
799}
800
801/*
802 * The following function returns false either if every active queue
803 * must receive the same share of the throughput (symmetric scenario),
804 * or, as a special case, if bfqq must receive a share of the
805 * throughput lower than or equal to the share that every other active
806 * queue must receive. If bfqq does sync I/O, then these are the only
807 * two cases where bfqq happens to be guaranteed its share of the
808 * throughput even if I/O dispatching is not plugged when bfqq remains
809 * temporarily empty (for more details, see the comments in the
810 * function bfq_better_to_idle()). For this reason, the return value
811 * of this function is used to check whether I/O-dispatch plugging can
812 * be avoided.
813 *
814 * The above first case (symmetric scenario) occurs when:
815 * 1) all active queues have the same weight,
816 * 2) all active queues belong to the same I/O-priority class,
817 * 3) all active groups at the same level in the groups tree have the same
818 * weight,
819 * 4) all active groups at the same level in the groups tree have the same
820 * number of children.
821 *
822 * Unfortunately, keeping the necessary state for evaluating exactly
823 * the last two symmetry sub-conditions above would be quite complex
824 * and time consuming. Therefore this function evaluates, instead,
825 * only the following stronger three sub-conditions, for which it is
826 * much easier to maintain the needed state:
827 * 1) all active queues have the same weight,
828 * 2) all active queues belong to the same I/O-priority class,
829 * 3) there is at most one active group.
830 * In particular, the last condition is always true if hierarchical
831 * support or the cgroups interface are not enabled, thus no state
832 * needs to be maintained in this case.
833 */
834static bool bfq_asymmetric_scenario(struct bfq_data *bfqd,
835 struct bfq_queue *bfqq)
836{
837 bool smallest_weight = bfqq &&
838 bfqq->weight_counter &&
839 bfqq->weight_counter ==
840 container_of(
841 rb_first_cached(&bfqd->queue_weights_tree),
842 struct bfq_weight_counter,
843 weights_node);
844
845 /*
846 * For queue weights to differ, queue_weights_tree must contain
847 * at least two nodes.
848 */
849 bool varied_queue_weights = !smallest_weight &&
850 !RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) &&
851 (bfqd->queue_weights_tree.rb_root.rb_node->rb_left ||
852 bfqd->queue_weights_tree.rb_root.rb_node->rb_right);
853
854 bool multiple_classes_busy =
855 (bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
856 (bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
857 (bfqd->busy_queues[1] && bfqd->busy_queues[2]);
858
859 return varied_queue_weights || multiple_classes_busy
860#ifdef CONFIG_BFQ_GROUP_IOSCHED
861 || bfqd->num_groups_with_pending_reqs > 1
862#endif
863 ;
864}
865
866/*
867 * If the weight-counter tree passed as input contains no counter for
868 * the weight of the input queue, then add that counter; otherwise just
869 * increment the existing counter.
870 *
871 * Note that weight-counter trees contain few nodes in mostly symmetric
872 * scenarios. For example, if all queues have the same weight, then the
873 * weight-counter tree for the queues may contain at most one node.
874 * This holds even if low_latency is on, because weight-raised queues
875 * are not inserted in the tree.
876 * In most scenarios, the rate at which nodes are created/destroyed
877 * should be low too.
878 */
879void bfq_weights_tree_add(struct bfq_queue *bfqq)
880{
881 struct rb_root_cached *root = &bfqq->bfqd->queue_weights_tree;
882 struct bfq_entity *entity = &bfqq->entity;
883 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
884 bool leftmost = true;
885
886 /*
887 * Do not insert if the queue is already associated with a
888 * counter, which happens if:
889 * 1) a request arrival has caused the queue to become both
890 * non-weight-raised, and hence change its weight, and
891 * backlogged; in this respect, each of the two events
892 * causes an invocation of this function,
893 * 2) this is the invocation of this function caused by the
894 * second event. This second invocation is actually useless,
895 * and we handle this fact by exiting immediately. More
896 * efficient or clearer solutions might possibly be adopted.
897 */
898 if (bfqq->weight_counter)
899 return;
900
901 while (*new) {
902 struct bfq_weight_counter *__counter = container_of(*new,
903 struct bfq_weight_counter,
904 weights_node);
905 parent = *new;
906
907 if (entity->weight == __counter->weight) {
908 bfqq->weight_counter = __counter;
909 goto inc_counter;
910 }
911 if (entity->weight < __counter->weight)
912 new = &((*new)->rb_left);
913 else {
914 new = &((*new)->rb_right);
915 leftmost = false;
916 }
917 }
918
919 bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
920 GFP_ATOMIC);
921
922 /*
923 * In the unlucky event of an allocation failure, we just
924 * exit. This will cause the weight of queue to not be
925 * considered in bfq_asymmetric_scenario, which, in its turn,
926 * causes the scenario to be deemed wrongly symmetric in case
927 * bfqq's weight would have been the only weight making the
928 * scenario asymmetric. On the bright side, no unbalance will
929 * however occur when bfqq becomes inactive again (the
930 * invocation of this function is triggered by an activation
931 * of queue). In fact, bfq_weights_tree_remove does nothing
932 * if !bfqq->weight_counter.
933 */
934 if (unlikely(!bfqq->weight_counter))
935 return;
936
937 bfqq->weight_counter->weight = entity->weight;
938 rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
939 rb_insert_color_cached(&bfqq->weight_counter->weights_node, root,
940 leftmost);
941
942inc_counter:
943 bfqq->weight_counter->num_active++;
944 bfqq->ref++;
945}
946
947/*
948 * Decrement the weight counter associated with the queue, and, if the
949 * counter reaches 0, remove the counter from the tree.
950 * See the comments to the function bfq_weights_tree_add() for considerations
951 * about overhead.
952 */
953void bfq_weights_tree_remove(struct bfq_queue *bfqq)
954{
955 struct rb_root_cached *root;
956
957 if (!bfqq->weight_counter)
958 return;
959
960 root = &bfqq->bfqd->queue_weights_tree;
961 bfqq->weight_counter->num_active--;
962 if (bfqq->weight_counter->num_active > 0)
963 goto reset_entity_pointer;
964
965 rb_erase_cached(&bfqq->weight_counter->weights_node, root);
966 kfree(bfqq->weight_counter);
967
968reset_entity_pointer:
969 bfqq->weight_counter = NULL;
970 bfq_put_queue(bfqq);
971}
972
973/*
974 * Return expired entry, or NULL to just start from scratch in rbtree.
975 */
976static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
977 struct request *last)
978{
979 struct request *rq;
980
981 if (bfq_bfqq_fifo_expire(bfqq))
982 return NULL;
983
984 bfq_mark_bfqq_fifo_expire(bfqq);
985
986 rq = rq_entry_fifo(bfqq->fifo.next);
987
988 if (rq == last || ktime_get_ns() < rq->fifo_time)
989 return NULL;
990
991 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
992 return rq;
993}
994
995static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
996 struct bfq_queue *bfqq,
997 struct request *last)
998{
999 struct rb_node *rbnext = rb_next(&last->rb_node);
1000 struct rb_node *rbprev = rb_prev(&last->rb_node);
1001 struct request *next, *prev = NULL;
1002
1003 /* Follow expired path, else get first next available. */
1004 next = bfq_check_fifo(bfqq, last);
1005 if (next)
1006 return next;
1007
1008 if (rbprev)
1009 prev = rb_entry_rq(rbprev);
1010
1011 if (rbnext)
1012 next = rb_entry_rq(rbnext);
1013 else {
1014 rbnext = rb_first(&bfqq->sort_list);
1015 if (rbnext && rbnext != &last->rb_node)
1016 next = rb_entry_rq(rbnext);
1017 }
1018
1019 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
1020}
1021
1022/* see the definition of bfq_async_charge_factor for details */
1023static unsigned long bfq_serv_to_charge(struct request *rq,
1024 struct bfq_queue *bfqq)
1025{
1026 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
1027 bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
1028 return blk_rq_sectors(rq);
1029
1030 return blk_rq_sectors(rq) * bfq_async_charge_factor;
1031}
1032
1033/**
1034 * bfq_updated_next_req - update the queue after a new next_rq selection.
1035 * @bfqd: the device data the queue belongs to.
1036 * @bfqq: the queue to update.
1037 *
1038 * If the first request of a queue changes we make sure that the queue
1039 * has enough budget to serve at least its first request (if the
1040 * request has grown). We do this because if the queue has not enough
1041 * budget for its first request, it has to go through two dispatch
1042 * rounds to actually get it dispatched.
1043 */
1044static void bfq_updated_next_req(struct bfq_data *bfqd,
1045 struct bfq_queue *bfqq)
1046{
1047 struct bfq_entity *entity = &bfqq->entity;
1048 struct request *next_rq = bfqq->next_rq;
1049 unsigned long new_budget;
1050
1051 if (!next_rq)
1052 return;
1053
1054 if (bfqq == bfqd->in_service_queue)
1055 /*
1056 * In order not to break guarantees, budgets cannot be
1057 * changed after an entity has been selected.
1058 */
1059 return;
1060
1061 new_budget = max_t(unsigned long,
1062 max_t(unsigned long, bfqq->max_budget,
1063 bfq_serv_to_charge(next_rq, bfqq)),
1064 entity->service);
1065 if (entity->budget != new_budget) {
1066 entity->budget = new_budget;
1067 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
1068 new_budget);
1069 bfq_requeue_bfqq(bfqd, bfqq, false);
1070 }
1071}
1072
1073static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
1074{
1075 u64 dur;
1076
1077 if (bfqd->bfq_wr_max_time > 0)
1078 return bfqd->bfq_wr_max_time;
1079
1080 dur = bfqd->rate_dur_prod;
1081 do_div(dur, bfqd->peak_rate);
1082
1083 /*
1084 * Limit duration between 3 and 25 seconds. The upper limit
1085 * has been conservatively set after the following worst case:
1086 * on a QEMU/KVM virtual machine
1087 * - running in a slow PC
1088 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
1089 * - serving a heavy I/O workload, such as the sequential reading
1090 * of several files
1091 * mplayer took 23 seconds to start, if constantly weight-raised.
1092 *
1093 * As for higher values than that accommodating the above bad
1094 * scenario, tests show that higher values would often yield
1095 * the opposite of the desired result, i.e., would worsen
1096 * responsiveness by allowing non-interactive applications to
1097 * preserve weight raising for too long.
1098 *
1099 * On the other end, lower values than 3 seconds make it
1100 * difficult for most interactive tasks to complete their jobs
1101 * before weight-raising finishes.
1102 */
1103 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
1104}
1105
1106/* switch back from soft real-time to interactive weight raising */
1107static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
1108 struct bfq_data *bfqd)
1109{
1110 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1111 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1112 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
1113}
1114
1115static void
1116bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
1117 struct bfq_io_cq *bic, bool bfq_already_existing)
1118{
1119 unsigned int old_wr_coeff = 1;
1120 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
1121
1122 if (bic->saved_has_short_ttime)
1123 bfq_mark_bfqq_has_short_ttime(bfqq);
1124 else
1125 bfq_clear_bfqq_has_short_ttime(bfqq);
1126
1127 if (bic->saved_IO_bound)
1128 bfq_mark_bfqq_IO_bound(bfqq);
1129 else
1130 bfq_clear_bfqq_IO_bound(bfqq);
1131
1132 bfqq->last_serv_time_ns = bic->saved_last_serv_time_ns;
1133 bfqq->inject_limit = bic->saved_inject_limit;
1134 bfqq->decrease_time_jif = bic->saved_decrease_time_jif;
1135
1136 bfqq->entity.new_weight = bic->saved_weight;
1137 bfqq->ttime = bic->saved_ttime;
1138 bfqq->io_start_time = bic->saved_io_start_time;
1139 bfqq->tot_idle_time = bic->saved_tot_idle_time;
1140 /*
1141 * Restore weight coefficient only if low_latency is on
1142 */
1143 if (bfqd->low_latency) {
1144 old_wr_coeff = bfqq->wr_coeff;
1145 bfqq->wr_coeff = bic->saved_wr_coeff;
1146 }
1147 bfqq->service_from_wr = bic->saved_service_from_wr;
1148 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
1149 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
1150 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
1151
1152 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1153 time_is_before_jiffies(bfqq->last_wr_start_finish +
1154 bfqq->wr_cur_max_time))) {
1155 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1156 !bfq_bfqq_in_large_burst(bfqq) &&
1157 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1158 bfq_wr_duration(bfqd))) {
1159 switch_back_to_interactive_wr(bfqq, bfqd);
1160 } else {
1161 bfqq->wr_coeff = 1;
1162 bfq_log_bfqq(bfqq->bfqd, bfqq,
1163 "resume state: switching off wr");
1164 }
1165 }
1166
1167 /* make sure weight will be updated, however we got here */
1168 bfqq->entity.prio_changed = 1;
1169
1170 if (likely(!busy))
1171 return;
1172
1173 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1174 bfqd->wr_busy_queues++;
1175 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1176 bfqd->wr_busy_queues--;
1177}
1178
1179static int bfqq_process_refs(struct bfq_queue *bfqq)
1180{
1181 return bfqq->ref - bfqq->entity.allocated -
1182 bfqq->entity.on_st_or_in_serv -
1183 (bfqq->weight_counter != NULL) - bfqq->stable_ref;
1184}
1185
1186/* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1187static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1188{
1189 struct bfq_queue *item;
1190 struct hlist_node *n;
1191
1192 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1193 hlist_del_init(&item->burst_list_node);
1194
1195 /*
1196 * Start the creation of a new burst list only if there is no
1197 * active queue. See comments on the conditional invocation of
1198 * bfq_handle_burst().
1199 */
1200 if (bfq_tot_busy_queues(bfqd) == 0) {
1201 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1202 bfqd->burst_size = 1;
1203 } else
1204 bfqd->burst_size = 0;
1205
1206 bfqd->burst_parent_entity = bfqq->entity.parent;
1207}
1208
1209/* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1210static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1211{
1212 /* Increment burst size to take into account also bfqq */
1213 bfqd->burst_size++;
1214
1215 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1216 struct bfq_queue *pos, *bfqq_item;
1217 struct hlist_node *n;
1218
1219 /*
1220 * Enough queues have been activated shortly after each
1221 * other to consider this burst as large.
1222 */
1223 bfqd->large_burst = true;
1224
1225 /*
1226 * We can now mark all queues in the burst list as
1227 * belonging to a large burst.
1228 */
1229 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1230 burst_list_node)
1231 bfq_mark_bfqq_in_large_burst(bfqq_item);
1232 bfq_mark_bfqq_in_large_burst(bfqq);
1233
1234 /*
1235 * From now on, and until the current burst finishes, any
1236 * new queue being activated shortly after the last queue
1237 * was inserted in the burst can be immediately marked as
1238 * belonging to a large burst. So the burst list is not
1239 * needed any more. Remove it.
1240 */
1241 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1242 burst_list_node)
1243 hlist_del_init(&pos->burst_list_node);
1244 } else /*
1245 * Burst not yet large: add bfqq to the burst list. Do
1246 * not increment the ref counter for bfqq, because bfqq
1247 * is removed from the burst list before freeing bfqq
1248 * in put_queue.
1249 */
1250 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1251}
1252
1253/*
1254 * If many queues belonging to the same group happen to be created
1255 * shortly after each other, then the processes associated with these
1256 * queues have typically a common goal. In particular, bursts of queue
1257 * creations are usually caused by services or applications that spawn
1258 * many parallel threads/processes. Examples are systemd during boot,
1259 * or git grep. To help these processes get their job done as soon as
1260 * possible, it is usually better to not grant either weight-raising
1261 * or device idling to their queues, unless these queues must be
1262 * protected from the I/O flowing through other active queues.
1263 *
1264 * In this comment we describe, firstly, the reasons why this fact
1265 * holds, and, secondly, the next function, which implements the main
1266 * steps needed to properly mark these queues so that they can then be
1267 * treated in a different way.
1268 *
1269 * The above services or applications benefit mostly from a high
1270 * throughput: the quicker the requests of the activated queues are
1271 * cumulatively served, the sooner the target job of these queues gets
1272 * completed. As a consequence, weight-raising any of these queues,
1273 * which also implies idling the device for it, is almost always
1274 * counterproductive, unless there are other active queues to isolate
1275 * these new queues from. If there no other active queues, then
1276 * weight-raising these new queues just lowers throughput in most
1277 * cases.
1278 *
1279 * On the other hand, a burst of queue creations may be caused also by
1280 * the start of an application that does not consist of a lot of
1281 * parallel I/O-bound threads. In fact, with a complex application,
1282 * several short processes may need to be executed to start-up the
1283 * application. In this respect, to start an application as quickly as
1284 * possible, the best thing to do is in any case to privilege the I/O
1285 * related to the application with respect to all other
1286 * I/O. Therefore, the best strategy to start as quickly as possible
1287 * an application that causes a burst of queue creations is to
1288 * weight-raise all the queues created during the burst. This is the
1289 * exact opposite of the best strategy for the other type of bursts.
1290 *
1291 * In the end, to take the best action for each of the two cases, the
1292 * two types of bursts need to be distinguished. Fortunately, this
1293 * seems relatively easy, by looking at the sizes of the bursts. In
1294 * particular, we found a threshold such that only bursts with a
1295 * larger size than that threshold are apparently caused by
1296 * services or commands such as systemd or git grep. For brevity,
1297 * hereafter we call just 'large' these bursts. BFQ *does not*
1298 * weight-raise queues whose creation occurs in a large burst. In
1299 * addition, for each of these queues BFQ performs or does not perform
1300 * idling depending on which choice boosts the throughput more. The
1301 * exact choice depends on the device and request pattern at
1302 * hand.
1303 *
1304 * Unfortunately, false positives may occur while an interactive task
1305 * is starting (e.g., an application is being started). The
1306 * consequence is that the queues associated with the task do not
1307 * enjoy weight raising as expected. Fortunately these false positives
1308 * are very rare. They typically occur if some service happens to
1309 * start doing I/O exactly when the interactive task starts.
1310 *
1311 * Turning back to the next function, it is invoked only if there are
1312 * no active queues (apart from active queues that would belong to the
1313 * same, possible burst bfqq would belong to), and it implements all
1314 * the steps needed to detect the occurrence of a large burst and to
1315 * properly mark all the queues belonging to it (so that they can then
1316 * be treated in a different way). This goal is achieved by
1317 * maintaining a "burst list" that holds, temporarily, the queues that
1318 * belong to the burst in progress. The list is then used to mark
1319 * these queues as belonging to a large burst if the burst does become
1320 * large. The main steps are the following.
1321 *
1322 * . when the very first queue is created, the queue is inserted into the
1323 * list (as it could be the first queue in a possible burst)
1324 *
1325 * . if the current burst has not yet become large, and a queue Q that does
1326 * not yet belong to the burst is activated shortly after the last time
1327 * at which a new queue entered the burst list, then the function appends
1328 * Q to the burst list
1329 *
1330 * . if, as a consequence of the previous step, the burst size reaches
1331 * the large-burst threshold, then
1332 *
1333 * . all the queues in the burst list are marked as belonging to a
1334 * large burst
1335 *
1336 * . the burst list is deleted; in fact, the burst list already served
1337 * its purpose (keeping temporarily track of the queues in a burst,
1338 * so as to be able to mark them as belonging to a large burst in the
1339 * previous sub-step), and now is not needed any more
1340 *
1341 * . the device enters a large-burst mode
1342 *
1343 * . if a queue Q that does not belong to the burst is created while
1344 * the device is in large-burst mode and shortly after the last time
1345 * at which a queue either entered the burst list or was marked as
1346 * belonging to the current large burst, then Q is immediately marked
1347 * as belonging to a large burst.
1348 *
1349 * . if a queue Q that does not belong to the burst is created a while
1350 * later, i.e., not shortly after, than the last time at which a queue
1351 * either entered the burst list or was marked as belonging to the
1352 * current large burst, then the current burst is deemed as finished and:
1353 *
1354 * . the large-burst mode is reset if set
1355 *
1356 * . the burst list is emptied
1357 *
1358 * . Q is inserted in the burst list, as Q may be the first queue
1359 * in a possible new burst (then the burst list contains just Q
1360 * after this step).
1361 */
1362static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1363{
1364 /*
1365 * If bfqq is already in the burst list or is part of a large
1366 * burst, or finally has just been split, then there is
1367 * nothing else to do.
1368 */
1369 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1370 bfq_bfqq_in_large_burst(bfqq) ||
1371 time_is_after_eq_jiffies(bfqq->split_time +
1372 msecs_to_jiffies(10)))
1373 return;
1374
1375 /*
1376 * If bfqq's creation happens late enough, or bfqq belongs to
1377 * a different group than the burst group, then the current
1378 * burst is finished, and related data structures must be
1379 * reset.
1380 *
1381 * In this respect, consider the special case where bfqq is
1382 * the very first queue created after BFQ is selected for this
1383 * device. In this case, last_ins_in_burst and
1384 * burst_parent_entity are not yet significant when we get
1385 * here. But it is easy to verify that, whether or not the
1386 * following condition is true, bfqq will end up being
1387 * inserted into the burst list. In particular the list will
1388 * happen to contain only bfqq. And this is exactly what has
1389 * to happen, as bfqq may be the first queue of the first
1390 * burst.
1391 */
1392 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1393 bfqd->bfq_burst_interval) ||
1394 bfqq->entity.parent != bfqd->burst_parent_entity) {
1395 bfqd->large_burst = false;
1396 bfq_reset_burst_list(bfqd, bfqq);
1397 goto end;
1398 }
1399
1400 /*
1401 * If we get here, then bfqq is being activated shortly after the
1402 * last queue. So, if the current burst is also large, we can mark
1403 * bfqq as belonging to this large burst immediately.
1404 */
1405 if (bfqd->large_burst) {
1406 bfq_mark_bfqq_in_large_burst(bfqq);
1407 goto end;
1408 }
1409
1410 /*
1411 * If we get here, then a large-burst state has not yet been
1412 * reached, but bfqq is being activated shortly after the last
1413 * queue. Then we add bfqq to the burst.
1414 */
1415 bfq_add_to_burst(bfqd, bfqq);
1416end:
1417 /*
1418 * At this point, bfqq either has been added to the current
1419 * burst or has caused the current burst to terminate and a
1420 * possible new burst to start. In particular, in the second
1421 * case, bfqq has become the first queue in the possible new
1422 * burst. In both cases last_ins_in_burst needs to be moved
1423 * forward.
1424 */
1425 bfqd->last_ins_in_burst = jiffies;
1426}
1427
1428static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1429{
1430 struct bfq_entity *entity = &bfqq->entity;
1431
1432 return entity->budget - entity->service;
1433}
1434
1435/*
1436 * If enough samples have been computed, return the current max budget
1437 * stored in bfqd, which is dynamically updated according to the
1438 * estimated disk peak rate; otherwise return the default max budget
1439 */
1440static int bfq_max_budget(struct bfq_data *bfqd)
1441{
1442 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1443 return bfq_default_max_budget;
1444 else
1445 return bfqd->bfq_max_budget;
1446}
1447
1448/*
1449 * Return min budget, which is a fraction of the current or default
1450 * max budget (trying with 1/32)
1451 */
1452static int bfq_min_budget(struct bfq_data *bfqd)
1453{
1454 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1455 return bfq_default_max_budget / 32;
1456 else
1457 return bfqd->bfq_max_budget / 32;
1458}
1459
1460/*
1461 * The next function, invoked after the input queue bfqq switches from
1462 * idle to busy, updates the budget of bfqq. The function also tells
1463 * whether the in-service queue should be expired, by returning
1464 * true. The purpose of expiring the in-service queue is to give bfqq
1465 * the chance to possibly preempt the in-service queue, and the reason
1466 * for preempting the in-service queue is to achieve one of the two
1467 * goals below.
1468 *
1469 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1470 * expired because it has remained idle. In particular, bfqq may have
1471 * expired for one of the following two reasons:
1472 *
1473 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1474 * and did not make it to issue a new request before its last
1475 * request was served;
1476 *
1477 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1478 * a new request before the expiration of the idling-time.
1479 *
1480 * Even if bfqq has expired for one of the above reasons, the process
1481 * associated with the queue may be however issuing requests greedily,
1482 * and thus be sensitive to the bandwidth it receives (bfqq may have
1483 * remained idle for other reasons: CPU high load, bfqq not enjoying
1484 * idling, I/O throttling somewhere in the path from the process to
1485 * the I/O scheduler, ...). But if, after every expiration for one of
1486 * the above two reasons, bfqq has to wait for the service of at least
1487 * one full budget of another queue before being served again, then
1488 * bfqq is likely to get a much lower bandwidth or resource time than
1489 * its reserved ones. To address this issue, two countermeasures need
1490 * to be taken.
1491 *
1492 * First, the budget and the timestamps of bfqq need to be updated in
1493 * a special way on bfqq reactivation: they need to be updated as if
1494 * bfqq did not remain idle and did not expire. In fact, if they are
1495 * computed as if bfqq expired and remained idle until reactivation,
1496 * then the process associated with bfqq is treated as if, instead of
1497 * being greedy, it stopped issuing requests when bfqq remained idle,
1498 * and restarts issuing requests only on this reactivation. In other
1499 * words, the scheduler does not help the process recover the "service
1500 * hole" between bfqq expiration and reactivation. As a consequence,
1501 * the process receives a lower bandwidth than its reserved one. In
1502 * contrast, to recover this hole, the budget must be updated as if
1503 * bfqq was not expired at all before this reactivation, i.e., it must
1504 * be set to the value of the remaining budget when bfqq was
1505 * expired. Along the same line, timestamps need to be assigned the
1506 * value they had the last time bfqq was selected for service, i.e.,
1507 * before last expiration. Thus timestamps need to be back-shifted
1508 * with respect to their normal computation (see [1] for more details
1509 * on this tricky aspect).
1510 *
1511 * Secondly, to allow the process to recover the hole, the in-service
1512 * queue must be expired too, to give bfqq the chance to preempt it
1513 * immediately. In fact, if bfqq has to wait for a full budget of the
1514 * in-service queue to be completed, then it may become impossible to
1515 * let the process recover the hole, even if the back-shifted
1516 * timestamps of bfqq are lower than those of the in-service queue. If
1517 * this happens for most or all of the holes, then the process may not
1518 * receive its reserved bandwidth. In this respect, it is worth noting
1519 * that, being the service of outstanding requests unpreemptible, a
1520 * little fraction of the holes may however be unrecoverable, thereby
1521 * causing a little loss of bandwidth.
1522 *
1523 * The last important point is detecting whether bfqq does need this
1524 * bandwidth recovery. In this respect, the next function deems the
1525 * process associated with bfqq greedy, and thus allows it to recover
1526 * the hole, if: 1) the process is waiting for the arrival of a new
1527 * request (which implies that bfqq expired for one of the above two
1528 * reasons), and 2) such a request has arrived soon. The first
1529 * condition is controlled through the flag non_blocking_wait_rq,
1530 * while the second through the flag arrived_in_time. If both
1531 * conditions hold, then the function computes the budget in the
1532 * above-described special way, and signals that the in-service queue
1533 * should be expired. Timestamp back-shifting is done later in
1534 * __bfq_activate_entity.
1535 *
1536 * 2. Reduce latency. Even if timestamps are not backshifted to let
1537 * the process associated with bfqq recover a service hole, bfqq may
1538 * however happen to have, after being (re)activated, a lower finish
1539 * timestamp than the in-service queue. That is, the next budget of
1540 * bfqq may have to be completed before the one of the in-service
1541 * queue. If this is the case, then preempting the in-service queue
1542 * allows this goal to be achieved, apart from the unpreemptible,
1543 * outstanding requests mentioned above.
1544 *
1545 * Unfortunately, regardless of which of the above two goals one wants
1546 * to achieve, service trees need first to be updated to know whether
1547 * the in-service queue must be preempted. To have service trees
1548 * correctly updated, the in-service queue must be expired and
1549 * rescheduled, and bfqq must be scheduled too. This is one of the
1550 * most costly operations (in future versions, the scheduling
1551 * mechanism may be re-designed in such a way to make it possible to
1552 * know whether preemption is needed without needing to update service
1553 * trees). In addition, queue preemptions almost always cause random
1554 * I/O, which may in turn cause loss of throughput. Finally, there may
1555 * even be no in-service queue when the next function is invoked (so,
1556 * no queue to compare timestamps with). Because of these facts, the
1557 * next function adopts the following simple scheme to avoid costly
1558 * operations, too frequent preemptions and too many dependencies on
1559 * the state of the scheduler: it requests the expiration of the
1560 * in-service queue (unconditionally) only for queues that need to
1561 * recover a hole. Then it delegates to other parts of the code the
1562 * responsibility of handling the above case 2.
1563 */
1564static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1565 struct bfq_queue *bfqq,
1566 bool arrived_in_time)
1567{
1568 struct bfq_entity *entity = &bfqq->entity;
1569
1570 /*
1571 * In the next compound condition, we check also whether there
1572 * is some budget left, because otherwise there is no point in
1573 * trying to go on serving bfqq with this same budget: bfqq
1574 * would be expired immediately after being selected for
1575 * service. This would only cause useless overhead.
1576 */
1577 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1578 bfq_bfqq_budget_left(bfqq) > 0) {
1579 /*
1580 * We do not clear the flag non_blocking_wait_rq here, as
1581 * the latter is used in bfq_activate_bfqq to signal
1582 * that timestamps need to be back-shifted (and is
1583 * cleared right after).
1584 */
1585
1586 /*
1587 * In next assignment we rely on that either
1588 * entity->service or entity->budget are not updated
1589 * on expiration if bfqq is empty (see
1590 * __bfq_bfqq_recalc_budget). Thus both quantities
1591 * remain unchanged after such an expiration, and the
1592 * following statement therefore assigns to
1593 * entity->budget the remaining budget on such an
1594 * expiration.
1595 */
1596 entity->budget = min_t(unsigned long,
1597 bfq_bfqq_budget_left(bfqq),
1598 bfqq->max_budget);
1599
1600 /*
1601 * At this point, we have used entity->service to get
1602 * the budget left (needed for updating
1603 * entity->budget). Thus we finally can, and have to,
1604 * reset entity->service. The latter must be reset
1605 * because bfqq would otherwise be charged again for
1606 * the service it has received during its previous
1607 * service slot(s).
1608 */
1609 entity->service = 0;
1610
1611 return true;
1612 }
1613
1614 /*
1615 * We can finally complete expiration, by setting service to 0.
1616 */
1617 entity->service = 0;
1618 entity->budget = max_t(unsigned long, bfqq->max_budget,
1619 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1620 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1621 return false;
1622}
1623
1624/*
1625 * Return the farthest past time instant according to jiffies
1626 * macros.
1627 */
1628static unsigned long bfq_smallest_from_now(void)
1629{
1630 return jiffies - MAX_JIFFY_OFFSET;
1631}
1632
1633static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1634 struct bfq_queue *bfqq,
1635 unsigned int old_wr_coeff,
1636 bool wr_or_deserves_wr,
1637 bool interactive,
1638 bool in_burst,
1639 bool soft_rt)
1640{
1641 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1642 /* start a weight-raising period */
1643 if (interactive) {
1644 bfqq->service_from_wr = 0;
1645 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1646 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1647 } else {
1648 /*
1649 * No interactive weight raising in progress
1650 * here: assign minus infinity to
1651 * wr_start_at_switch_to_srt, to make sure
1652 * that, at the end of the soft-real-time
1653 * weight raising periods that is starting
1654 * now, no interactive weight-raising period
1655 * may be wrongly considered as still in
1656 * progress (and thus actually started by
1657 * mistake).
1658 */
1659 bfqq->wr_start_at_switch_to_srt =
1660 bfq_smallest_from_now();
1661 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1662 BFQ_SOFTRT_WEIGHT_FACTOR;
1663 bfqq->wr_cur_max_time =
1664 bfqd->bfq_wr_rt_max_time;
1665 }
1666
1667 /*
1668 * If needed, further reduce budget to make sure it is
1669 * close to bfqq's backlog, so as to reduce the
1670 * scheduling-error component due to a too large
1671 * budget. Do not care about throughput consequences,
1672 * but only about latency. Finally, do not assign a
1673 * too small budget either, to avoid increasing
1674 * latency by causing too frequent expirations.
1675 */
1676 bfqq->entity.budget = min_t(unsigned long,
1677 bfqq->entity.budget,
1678 2 * bfq_min_budget(bfqd));
1679 } else if (old_wr_coeff > 1) {
1680 if (interactive) { /* update wr coeff and duration */
1681 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1682 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1683 } else if (in_burst)
1684 bfqq->wr_coeff = 1;
1685 else if (soft_rt) {
1686 /*
1687 * The application is now or still meeting the
1688 * requirements for being deemed soft rt. We
1689 * can then correctly and safely (re)charge
1690 * the weight-raising duration for the
1691 * application with the weight-raising
1692 * duration for soft rt applications.
1693 *
1694 * In particular, doing this recharge now, i.e.,
1695 * before the weight-raising period for the
1696 * application finishes, reduces the probability
1697 * of the following negative scenario:
1698 * 1) the weight of a soft rt application is
1699 * raised at startup (as for any newly
1700 * created application),
1701 * 2) since the application is not interactive,
1702 * at a certain time weight-raising is
1703 * stopped for the application,
1704 * 3) at that time the application happens to
1705 * still have pending requests, and hence
1706 * is destined to not have a chance to be
1707 * deemed soft rt before these requests are
1708 * completed (see the comments to the
1709 * function bfq_bfqq_softrt_next_start()
1710 * for details on soft rt detection),
1711 * 4) these pending requests experience a high
1712 * latency because the application is not
1713 * weight-raised while they are pending.
1714 */
1715 if (bfqq->wr_cur_max_time !=
1716 bfqd->bfq_wr_rt_max_time) {
1717 bfqq->wr_start_at_switch_to_srt =
1718 bfqq->last_wr_start_finish;
1719
1720 bfqq->wr_cur_max_time =
1721 bfqd->bfq_wr_rt_max_time;
1722 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1723 BFQ_SOFTRT_WEIGHT_FACTOR;
1724 }
1725 bfqq->last_wr_start_finish = jiffies;
1726 }
1727 }
1728}
1729
1730static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1731 struct bfq_queue *bfqq)
1732{
1733 return bfqq->dispatched == 0 &&
1734 time_is_before_jiffies(
1735 bfqq->budget_timeout +
1736 bfqd->bfq_wr_min_idle_time);
1737}
1738
1739
1740/*
1741 * Return true if bfqq is in a higher priority class, or has a higher
1742 * weight than the in-service queue.
1743 */
1744static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq,
1745 struct bfq_queue *in_serv_bfqq)
1746{
1747 int bfqq_weight, in_serv_weight;
1748
1749 if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class)
1750 return true;
1751
1752 if (in_serv_bfqq->entity.parent == bfqq->entity.parent) {
1753 bfqq_weight = bfqq->entity.weight;
1754 in_serv_weight = in_serv_bfqq->entity.weight;
1755 } else {
1756 if (bfqq->entity.parent)
1757 bfqq_weight = bfqq->entity.parent->weight;
1758 else
1759 bfqq_weight = bfqq->entity.weight;
1760 if (in_serv_bfqq->entity.parent)
1761 in_serv_weight = in_serv_bfqq->entity.parent->weight;
1762 else
1763 in_serv_weight = in_serv_bfqq->entity.weight;
1764 }
1765
1766 return bfqq_weight > in_serv_weight;
1767}
1768
1769static bool bfq_better_to_idle(struct bfq_queue *bfqq);
1770
1771static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1772 struct bfq_queue *bfqq,
1773 int old_wr_coeff,
1774 struct request *rq,
1775 bool *interactive)
1776{
1777 bool soft_rt, in_burst, wr_or_deserves_wr,
1778 bfqq_wants_to_preempt,
1779 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1780 /*
1781 * See the comments on
1782 * bfq_bfqq_update_budg_for_activation for
1783 * details on the usage of the next variable.
1784 */
1785 arrived_in_time = ktime_get_ns() <=
1786 bfqq->ttime.last_end_request +
1787 bfqd->bfq_slice_idle * 3;
1788
1789
1790 /*
1791 * bfqq deserves to be weight-raised if:
1792 * - it is sync,
1793 * - it does not belong to a large burst,
1794 * - it has been idle for enough time or is soft real-time,
1795 * - is linked to a bfq_io_cq (it is not shared in any sense),
1796 * - has a default weight (otherwise we assume the user wanted
1797 * to control its weight explicitly)
1798 */
1799 in_burst = bfq_bfqq_in_large_burst(bfqq);
1800 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1801 !BFQQ_TOTALLY_SEEKY(bfqq) &&
1802 !in_burst &&
1803 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1804 bfqq->dispatched == 0 &&
1805 bfqq->entity.new_weight == 40;
1806 *interactive = !in_burst && idle_for_long_time &&
1807 bfqq->entity.new_weight == 40;
1808 /*
1809 * Merged bfq_queues are kept out of weight-raising
1810 * (low-latency) mechanisms. The reason is that these queues
1811 * are usually created for non-interactive and
1812 * non-soft-real-time tasks. Yet this is not the case for
1813 * stably-merged queues. These queues are merged just because
1814 * they are created shortly after each other. So they may
1815 * easily serve the I/O of an interactive or soft-real time
1816 * application, if the application happens to spawn multiple
1817 * processes. So let also stably-merged queued enjoy weight
1818 * raising.
1819 */
1820 wr_or_deserves_wr = bfqd->low_latency &&
1821 (bfqq->wr_coeff > 1 ||
1822 (bfq_bfqq_sync(bfqq) &&
1823 (bfqq->bic || RQ_BIC(rq)->stably_merged) &&
1824 (*interactive || soft_rt)));
1825
1826 /*
1827 * Using the last flag, update budget and check whether bfqq
1828 * may want to preempt the in-service queue.
1829 */
1830 bfqq_wants_to_preempt =
1831 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1832 arrived_in_time);
1833
1834 /*
1835 * If bfqq happened to be activated in a burst, but has been
1836 * idle for much more than an interactive queue, then we
1837 * assume that, in the overall I/O initiated in the burst, the
1838 * I/O associated with bfqq is finished. So bfqq does not need
1839 * to be treated as a queue belonging to a burst
1840 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1841 * if set, and remove bfqq from the burst list if it's
1842 * there. We do not decrement burst_size, because the fact
1843 * that bfqq does not need to belong to the burst list any
1844 * more does not invalidate the fact that bfqq was created in
1845 * a burst.
1846 */
1847 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1848 idle_for_long_time &&
1849 time_is_before_jiffies(
1850 bfqq->budget_timeout +
1851 msecs_to_jiffies(10000))) {
1852 hlist_del_init(&bfqq->burst_list_node);
1853 bfq_clear_bfqq_in_large_burst(bfqq);
1854 }
1855
1856 bfq_clear_bfqq_just_created(bfqq);
1857
1858 if (bfqd->low_latency) {
1859 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1860 /* wraparound */
1861 bfqq->split_time =
1862 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1863
1864 if (time_is_before_jiffies(bfqq->split_time +
1865 bfqd->bfq_wr_min_idle_time)) {
1866 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1867 old_wr_coeff,
1868 wr_or_deserves_wr,
1869 *interactive,
1870 in_burst,
1871 soft_rt);
1872
1873 if (old_wr_coeff != bfqq->wr_coeff)
1874 bfqq->entity.prio_changed = 1;
1875 }
1876 }
1877
1878 bfqq->last_idle_bklogged = jiffies;
1879 bfqq->service_from_backlogged = 0;
1880 bfq_clear_bfqq_softrt_update(bfqq);
1881
1882 bfq_add_bfqq_busy(bfqq);
1883
1884 /*
1885 * Expire in-service queue if preemption may be needed for
1886 * guarantees or throughput. As for guarantees, we care
1887 * explicitly about two cases. The first is that bfqq has to
1888 * recover a service hole, as explained in the comments on
1889 * bfq_bfqq_update_budg_for_activation(), i.e., that
1890 * bfqq_wants_to_preempt is true. However, if bfqq does not
1891 * carry time-critical I/O, then bfqq's bandwidth is less
1892 * important than that of queues that carry time-critical I/O.
1893 * So, as a further constraint, we consider this case only if
1894 * bfqq is at least as weight-raised, i.e., at least as time
1895 * critical, as the in-service queue.
1896 *
1897 * The second case is that bfqq is in a higher priority class,
1898 * or has a higher weight than the in-service queue. If this
1899 * condition does not hold, we don't care because, even if
1900 * bfqq does not start to be served immediately, the resulting
1901 * delay for bfqq's I/O is however lower or much lower than
1902 * the ideal completion time to be guaranteed to bfqq's I/O.
1903 *
1904 * In both cases, preemption is needed only if, according to
1905 * the timestamps of both bfqq and of the in-service queue,
1906 * bfqq actually is the next queue to serve. So, to reduce
1907 * useless preemptions, the return value of
1908 * next_queue_may_preempt() is considered in the next compound
1909 * condition too. Yet next_queue_may_preempt() just checks a
1910 * simple, necessary condition for bfqq to be the next queue
1911 * to serve. In fact, to evaluate a sufficient condition, the
1912 * timestamps of the in-service queue would need to be
1913 * updated, and this operation is quite costly (see the
1914 * comments on bfq_bfqq_update_budg_for_activation()).
1915 *
1916 * As for throughput, we ask bfq_better_to_idle() whether we
1917 * still need to plug I/O dispatching. If bfq_better_to_idle()
1918 * says no, then plugging is not needed any longer, either to
1919 * boost throughput or to perserve service guarantees. Then
1920 * the best option is to stop plugging I/O, as not doing so
1921 * would certainly lower throughput. We may end up in this
1922 * case if: (1) upon a dispatch attempt, we detected that it
1923 * was better to plug I/O dispatch, and to wait for a new
1924 * request to arrive for the currently in-service queue, but
1925 * (2) this switch of bfqq to busy changes the scenario.
1926 */
1927 if (bfqd->in_service_queue &&
1928 ((bfqq_wants_to_preempt &&
1929 bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) ||
1930 bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue) ||
1931 !bfq_better_to_idle(bfqd->in_service_queue)) &&
1932 next_queue_may_preempt(bfqd))
1933 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1934 false, BFQQE_PREEMPTED);
1935}
1936
1937static void bfq_reset_inject_limit(struct bfq_data *bfqd,
1938 struct bfq_queue *bfqq)
1939{
1940 /* invalidate baseline total service time */
1941 bfqq->last_serv_time_ns = 0;
1942
1943 /*
1944 * Reset pointer in case we are waiting for
1945 * some request completion.
1946 */
1947 bfqd->waited_rq = NULL;
1948
1949 /*
1950 * If bfqq has a short think time, then start by setting the
1951 * inject limit to 0 prudentially, because the service time of
1952 * an injected I/O request may be higher than the think time
1953 * of bfqq, and therefore, if one request was injected when
1954 * bfqq remains empty, this injected request might delay the
1955 * service of the next I/O request for bfqq significantly. In
1956 * case bfqq can actually tolerate some injection, then the
1957 * adaptive update will however raise the limit soon. This
1958 * lucky circumstance holds exactly because bfqq has a short
1959 * think time, and thus, after remaining empty, is likely to
1960 * get new I/O enqueued---and then completed---before being
1961 * expired. This is the very pattern that gives the
1962 * limit-update algorithm the chance to measure the effect of
1963 * injection on request service times, and then to update the
1964 * limit accordingly.
1965 *
1966 * However, in the following special case, the inject limit is
1967 * left to 1 even if the think time is short: bfqq's I/O is
1968 * synchronized with that of some other queue, i.e., bfqq may
1969 * receive new I/O only after the I/O of the other queue is
1970 * completed. Keeping the inject limit to 1 allows the
1971 * blocking I/O to be served while bfqq is in service. And
1972 * this is very convenient both for bfqq and for overall
1973 * throughput, as explained in detail in the comments in
1974 * bfq_update_has_short_ttime().
1975 *
1976 * On the opposite end, if bfqq has a long think time, then
1977 * start directly by 1, because:
1978 * a) on the bright side, keeping at most one request in
1979 * service in the drive is unlikely to cause any harm to the
1980 * latency of bfqq's requests, as the service time of a single
1981 * request is likely to be lower than the think time of bfqq;
1982 * b) on the downside, after becoming empty, bfqq is likely to
1983 * expire before getting its next request. With this request
1984 * arrival pattern, it is very hard to sample total service
1985 * times and update the inject limit accordingly (see comments
1986 * on bfq_update_inject_limit()). So the limit is likely to be
1987 * never, or at least seldom, updated. As a consequence, by
1988 * setting the limit to 1, we avoid that no injection ever
1989 * occurs with bfqq. On the downside, this proactive step
1990 * further reduces chances to actually compute the baseline
1991 * total service time. Thus it reduces chances to execute the
1992 * limit-update algorithm and possibly raise the limit to more
1993 * than 1.
1994 */
1995 if (bfq_bfqq_has_short_ttime(bfqq))
1996 bfqq->inject_limit = 0;
1997 else
1998 bfqq->inject_limit = 1;
1999
2000 bfqq->decrease_time_jif = jiffies;
2001}
2002
2003static void bfq_update_io_intensity(struct bfq_queue *bfqq, u64 now_ns)
2004{
2005 u64 tot_io_time = now_ns - bfqq->io_start_time;
2006
2007 if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfqq->dispatched == 0)
2008 bfqq->tot_idle_time +=
2009 now_ns - bfqq->ttime.last_end_request;
2010
2011 if (unlikely(bfq_bfqq_just_created(bfqq)))
2012 return;
2013
2014 /*
2015 * Must be busy for at least about 80% of the time to be
2016 * considered I/O bound.
2017 */
2018 if (bfqq->tot_idle_time * 5 > tot_io_time)
2019 bfq_clear_bfqq_IO_bound(bfqq);
2020 else
2021 bfq_mark_bfqq_IO_bound(bfqq);
2022
2023 /*
2024 * Keep an observation window of at most 200 ms in the past
2025 * from now.
2026 */
2027 if (tot_io_time > 200 * NSEC_PER_MSEC) {
2028 bfqq->io_start_time = now_ns - (tot_io_time>>1);
2029 bfqq->tot_idle_time >>= 1;
2030 }
2031}
2032
2033/*
2034 * Detect whether bfqq's I/O seems synchronized with that of some
2035 * other queue, i.e., whether bfqq, after remaining empty, happens to
2036 * receive new I/O only right after some I/O request of the other
2037 * queue has been completed. We call waker queue the other queue, and
2038 * we assume, for simplicity, that bfqq may have at most one waker
2039 * queue.
2040 *
2041 * A remarkable throughput boost can be reached by unconditionally
2042 * injecting the I/O of the waker queue, every time a new
2043 * bfq_dispatch_request happens to be invoked while I/O is being
2044 * plugged for bfqq. In addition to boosting throughput, this
2045 * unblocks bfqq's I/O, thereby improving bandwidth and latency for
2046 * bfqq. Note that these same results may be achieved with the general
2047 * injection mechanism, but less effectively. For details on this
2048 * aspect, see the comments on the choice of the queue for injection
2049 * in bfq_select_queue().
2050 *
2051 * Turning back to the detection of a waker queue, a queue Q is deemed as a
2052 * waker queue for bfqq if, for three consecutive times, bfqq happens to become
2053 * non empty right after a request of Q has been completed within given
2054 * timeout. In this respect, even if bfqq is empty, we do not check for a waker
2055 * if it still has some in-flight I/O. In fact, in this case bfqq is actually
2056 * still being served by the drive, and may receive new I/O on the completion
2057 * of some of the in-flight requests. In particular, on the first time, Q is
2058 * tentatively set as a candidate waker queue, while on the third consecutive
2059 * time that Q is detected, the field waker_bfqq is set to Q, to confirm that Q
2060 * is a waker queue for bfqq. These detection steps are performed only if bfqq
2061 * has a long think time, so as to make it more likely that bfqq's I/O is
2062 * actually being blocked by a synchronization. This last filter, plus the
2063 * above three-times requirement and time limit for detection, make false
2064 * positives less likely.
2065 *
2066 * NOTE
2067 *
2068 * The sooner a waker queue is detected, the sooner throughput can be
2069 * boosted by injecting I/O from the waker queue. Fortunately,
2070 * detection is likely to be actually fast, for the following
2071 * reasons. While blocked by synchronization, bfqq has a long think
2072 * time. This implies that bfqq's inject limit is at least equal to 1
2073 * (see the comments in bfq_update_inject_limit()). So, thanks to
2074 * injection, the waker queue is likely to be served during the very
2075 * first I/O-plugging time interval for bfqq. This triggers the first
2076 * step of the detection mechanism. Thanks again to injection, the
2077 * candidate waker queue is then likely to be confirmed no later than
2078 * during the next I/O-plugging interval for bfqq.
2079 *
2080 * ISSUE
2081 *
2082 * On queue merging all waker information is lost.
2083 */
2084static void bfq_check_waker(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2085 u64 now_ns)
2086{
2087 char waker_name[MAX_BFQQ_NAME_LENGTH];
2088
2089 if (!bfqd->last_completed_rq_bfqq ||
2090 bfqd->last_completed_rq_bfqq == bfqq ||
2091 bfq_bfqq_has_short_ttime(bfqq) ||
2092 now_ns - bfqd->last_completion >= 4 * NSEC_PER_MSEC ||
2093 bfqd->last_completed_rq_bfqq == &bfqd->oom_bfqq ||
2094 bfqq == &bfqd->oom_bfqq)
2095 return;
2096
2097 /*
2098 * We reset waker detection logic also if too much time has passed
2099 * since the first detection. If wakeups are rare, pointless idling
2100 * doesn't hurt throughput that much. The condition below makes sure
2101 * we do not uselessly idle blocking waker in more than 1/64 cases.
2102 */
2103 if (bfqd->last_completed_rq_bfqq !=
2104 bfqq->tentative_waker_bfqq ||
2105 now_ns > bfqq->waker_detection_started +
2106 128 * (u64)bfqd->bfq_slice_idle) {
2107 /*
2108 * First synchronization detected with a
2109 * candidate waker queue, or with a different
2110 * candidate waker queue from the current one.
2111 */
2112 bfqq->tentative_waker_bfqq =
2113 bfqd->last_completed_rq_bfqq;
2114 bfqq->num_waker_detections = 1;
2115 bfqq->waker_detection_started = now_ns;
2116 bfq_bfqq_name(bfqq->tentative_waker_bfqq, waker_name,
2117 MAX_BFQQ_NAME_LENGTH);
2118 bfq_log_bfqq(bfqd, bfqq, "set tentative waker %s", waker_name);
2119 } else /* Same tentative waker queue detected again */
2120 bfqq->num_waker_detections++;
2121
2122 if (bfqq->num_waker_detections == 3) {
2123 bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
2124 bfqq->tentative_waker_bfqq = NULL;
2125 bfq_bfqq_name(bfqq->waker_bfqq, waker_name,
2126 MAX_BFQQ_NAME_LENGTH);
2127 bfq_log_bfqq(bfqd, bfqq, "set waker %s", waker_name);
2128
2129 /*
2130 * If the waker queue disappears, then
2131 * bfqq->waker_bfqq must be reset. To
2132 * this goal, we maintain in each
2133 * waker queue a list, woken_list, of
2134 * all the queues that reference the
2135 * waker queue through their
2136 * waker_bfqq pointer. When the waker
2137 * queue exits, the waker_bfqq pointer
2138 * of all the queues in the woken_list
2139 * is reset.
2140 *
2141 * In addition, if bfqq is already in
2142 * the woken_list of a waker queue,
2143 * then, before being inserted into
2144 * the woken_list of a new waker
2145 * queue, bfqq must be removed from
2146 * the woken_list of the old waker
2147 * queue.
2148 */
2149 if (!hlist_unhashed(&bfqq->woken_list_node))
2150 hlist_del_init(&bfqq->woken_list_node);
2151 hlist_add_head(&bfqq->woken_list_node,
2152 &bfqd->last_completed_rq_bfqq->woken_list);
2153 }
2154}
2155
2156static void bfq_add_request(struct request *rq)
2157{
2158 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2159 struct bfq_data *bfqd = bfqq->bfqd;
2160 struct request *next_rq, *prev;
2161 unsigned int old_wr_coeff = bfqq->wr_coeff;
2162 bool interactive = false;
2163 u64 now_ns = ktime_get_ns();
2164
2165 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
2166 bfqq->queued[rq_is_sync(rq)]++;
2167 /*
2168 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2169 * may be read without holding the lock in bfq_has_work().
2170 */
2171 WRITE_ONCE(bfqd->queued, bfqd->queued + 1);
2172
2173 if (bfq_bfqq_sync(bfqq) && RQ_BIC(rq)->requests <= 1) {
2174 bfq_check_waker(bfqd, bfqq, now_ns);
2175
2176 /*
2177 * Periodically reset inject limit, to make sure that
2178 * the latter eventually drops in case workload
2179 * changes, see step (3) in the comments on
2180 * bfq_update_inject_limit().
2181 */
2182 if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2183 msecs_to_jiffies(1000)))
2184 bfq_reset_inject_limit(bfqd, bfqq);
2185
2186 /*
2187 * The following conditions must hold to setup a new
2188 * sampling of total service time, and then a new
2189 * update of the inject limit:
2190 * - bfqq is in service, because the total service
2191 * time is evaluated only for the I/O requests of
2192 * the queues in service;
2193 * - this is the right occasion to compute or to
2194 * lower the baseline total service time, because
2195 * there are actually no requests in the drive,
2196 * or
2197 * the baseline total service time is available, and
2198 * this is the right occasion to compute the other
2199 * quantity needed to update the inject limit, i.e.,
2200 * the total service time caused by the amount of
2201 * injection allowed by the current value of the
2202 * limit. It is the right occasion because injection
2203 * has actually been performed during the service
2204 * hole, and there are still in-flight requests,
2205 * which are very likely to be exactly the injected
2206 * requests, or part of them;
2207 * - the minimum interval for sampling the total
2208 * service time and updating the inject limit has
2209 * elapsed.
2210 */
2211 if (bfqq == bfqd->in_service_queue &&
2212 (bfqd->rq_in_driver == 0 ||
2213 (bfqq->last_serv_time_ns > 0 &&
2214 bfqd->rqs_injected && bfqd->rq_in_driver > 0)) &&
2215 time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2216 msecs_to_jiffies(10))) {
2217 bfqd->last_empty_occupied_ns = ktime_get_ns();
2218 /*
2219 * Start the state machine for measuring the
2220 * total service time of rq: setting
2221 * wait_dispatch will cause bfqd->waited_rq to
2222 * be set when rq will be dispatched.
2223 */
2224 bfqd->wait_dispatch = true;
2225 /*
2226 * If there is no I/O in service in the drive,
2227 * then possible injection occurred before the
2228 * arrival of rq will not affect the total
2229 * service time of rq. So the injection limit
2230 * must not be updated as a function of such
2231 * total service time, unless new injection
2232 * occurs before rq is completed. To have the
2233 * injection limit updated only in the latter
2234 * case, reset rqs_injected here (rqs_injected
2235 * will be set in case injection is performed
2236 * on bfqq before rq is completed).
2237 */
2238 if (bfqd->rq_in_driver == 0)
2239 bfqd->rqs_injected = false;
2240 }
2241 }
2242
2243 if (bfq_bfqq_sync(bfqq))
2244 bfq_update_io_intensity(bfqq, now_ns);
2245
2246 elv_rb_add(&bfqq->sort_list, rq);
2247
2248 /*
2249 * Check if this request is a better next-serve candidate.
2250 */
2251 prev = bfqq->next_rq;
2252 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
2253 bfqq->next_rq = next_rq;
2254
2255 /*
2256 * Adjust priority tree position, if next_rq changes.
2257 * See comments on bfq_pos_tree_add_move() for the unlikely().
2258 */
2259 if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
2260 bfq_pos_tree_add_move(bfqd, bfqq);
2261
2262 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
2263 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
2264 rq, &interactive);
2265 else {
2266 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
2267 time_is_before_jiffies(
2268 bfqq->last_wr_start_finish +
2269 bfqd->bfq_wr_min_inter_arr_async)) {
2270 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
2271 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
2272
2273 bfqd->wr_busy_queues++;
2274 bfqq->entity.prio_changed = 1;
2275 }
2276 if (prev != bfqq->next_rq)
2277 bfq_updated_next_req(bfqd, bfqq);
2278 }
2279
2280 /*
2281 * Assign jiffies to last_wr_start_finish in the following
2282 * cases:
2283 *
2284 * . if bfqq is not going to be weight-raised, because, for
2285 * non weight-raised queues, last_wr_start_finish stores the
2286 * arrival time of the last request; as of now, this piece
2287 * of information is used only for deciding whether to
2288 * weight-raise async queues
2289 *
2290 * . if bfqq is not weight-raised, because, if bfqq is now
2291 * switching to weight-raised, then last_wr_start_finish
2292 * stores the time when weight-raising starts
2293 *
2294 * . if bfqq is interactive, because, regardless of whether
2295 * bfqq is currently weight-raised, the weight-raising
2296 * period must start or restart (this case is considered
2297 * separately because it is not detected by the above
2298 * conditions, if bfqq is already weight-raised)
2299 *
2300 * last_wr_start_finish has to be updated also if bfqq is soft
2301 * real-time, because the weight-raising period is constantly
2302 * restarted on idle-to-busy transitions for these queues, but
2303 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2304 * needed.
2305 */
2306 if (bfqd->low_latency &&
2307 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
2308 bfqq->last_wr_start_finish = jiffies;
2309}
2310
2311static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
2312 struct bio *bio,
2313 struct request_queue *q)
2314{
2315 struct bfq_queue *bfqq = bfqd->bio_bfqq;
2316
2317
2318 if (bfqq)
2319 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
2320
2321 return NULL;
2322}
2323
2324static sector_t get_sdist(sector_t last_pos, struct request *rq)
2325{
2326 if (last_pos)
2327 return abs(blk_rq_pos(rq) - last_pos);
2328
2329 return 0;
2330}
2331
2332static void bfq_remove_request(struct request_queue *q,
2333 struct request *rq)
2334{
2335 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2336 struct bfq_data *bfqd = bfqq->bfqd;
2337 const int sync = rq_is_sync(rq);
2338
2339 if (bfqq->next_rq == rq) {
2340 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
2341 bfq_updated_next_req(bfqd, bfqq);
2342 }
2343
2344 if (rq->queuelist.prev != &rq->queuelist)
2345 list_del_init(&rq->queuelist);
2346 bfqq->queued[sync]--;
2347 /*
2348 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2349 * may be read without holding the lock in bfq_has_work().
2350 */
2351 WRITE_ONCE(bfqd->queued, bfqd->queued - 1);
2352 elv_rb_del(&bfqq->sort_list, rq);
2353
2354 elv_rqhash_del(q, rq);
2355 if (q->last_merge == rq)
2356 q->last_merge = NULL;
2357
2358 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2359 bfqq->next_rq = NULL;
2360
2361 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
2362 bfq_del_bfqq_busy(bfqq, false);
2363 /*
2364 * bfqq emptied. In normal operation, when
2365 * bfqq is empty, bfqq->entity.service and
2366 * bfqq->entity.budget must contain,
2367 * respectively, the service received and the
2368 * budget used last time bfqq emptied. These
2369 * facts do not hold in this case, as at least
2370 * this last removal occurred while bfqq is
2371 * not in service. To avoid inconsistencies,
2372 * reset both bfqq->entity.service and
2373 * bfqq->entity.budget, if bfqq has still a
2374 * process that may issue I/O requests to it.
2375 */
2376 bfqq->entity.budget = bfqq->entity.service = 0;
2377 }
2378
2379 /*
2380 * Remove queue from request-position tree as it is empty.
2381 */
2382 if (bfqq->pos_root) {
2383 rb_erase(&bfqq->pos_node, bfqq->pos_root);
2384 bfqq->pos_root = NULL;
2385 }
2386 } else {
2387 /* see comments on bfq_pos_tree_add_move() for the unlikely() */
2388 if (unlikely(!bfqd->nonrot_with_queueing))
2389 bfq_pos_tree_add_move(bfqd, bfqq);
2390 }
2391
2392 if (rq->cmd_flags & REQ_META)
2393 bfqq->meta_pending--;
2394
2395}
2396
2397static bool bfq_bio_merge(struct request_queue *q, struct bio *bio,
2398 unsigned int nr_segs)
2399{
2400 struct bfq_data *bfqd = q->elevator->elevator_data;
2401 struct request *free = NULL;
2402 /*
2403 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2404 * store its return value for later use, to avoid nesting
2405 * queue_lock inside the bfqd->lock. We assume that the bic
2406 * returned by bfq_bic_lookup does not go away before
2407 * bfqd->lock is taken.
2408 */
2409 struct bfq_io_cq *bic = bfq_bic_lookup(q);
2410 bool ret;
2411
2412 spin_lock_irq(&bfqd->lock);
2413
2414 if (bic) {
2415 /*
2416 * Make sure cgroup info is uptodate for current process before
2417 * considering the merge.
2418 */
2419 bfq_bic_update_cgroup(bic, bio);
2420
2421 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
2422 } else {
2423 bfqd->bio_bfqq = NULL;
2424 }
2425 bfqd->bio_bic = bic;
2426
2427 ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
2428
2429 spin_unlock_irq(&bfqd->lock);
2430 if (free)
2431 blk_mq_free_request(free);
2432
2433 return ret;
2434}
2435
2436static int bfq_request_merge(struct request_queue *q, struct request **req,
2437 struct bio *bio)
2438{
2439 struct bfq_data *bfqd = q->elevator->elevator_data;
2440 struct request *__rq;
2441
2442 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
2443 if (__rq && elv_bio_merge_ok(__rq, bio)) {
2444 *req = __rq;
2445
2446 if (blk_discard_mergable(__rq))
2447 return ELEVATOR_DISCARD_MERGE;
2448 return ELEVATOR_FRONT_MERGE;
2449 }
2450
2451 return ELEVATOR_NO_MERGE;
2452}
2453
2454static void bfq_request_merged(struct request_queue *q, struct request *req,
2455 enum elv_merge type)
2456{
2457 if (type == ELEVATOR_FRONT_MERGE &&
2458 rb_prev(&req->rb_node) &&
2459 blk_rq_pos(req) <
2460 blk_rq_pos(container_of(rb_prev(&req->rb_node),
2461 struct request, rb_node))) {
2462 struct bfq_queue *bfqq = RQ_BFQQ(req);
2463 struct bfq_data *bfqd;
2464 struct request *prev, *next_rq;
2465
2466 if (!bfqq)
2467 return;
2468
2469 bfqd = bfqq->bfqd;
2470
2471 /* Reposition request in its sort_list */
2472 elv_rb_del(&bfqq->sort_list, req);
2473 elv_rb_add(&bfqq->sort_list, req);
2474
2475 /* Choose next request to be served for bfqq */
2476 prev = bfqq->next_rq;
2477 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
2478 bfqd->last_position);
2479 bfqq->next_rq = next_rq;
2480 /*
2481 * If next_rq changes, update both the queue's budget to
2482 * fit the new request and the queue's position in its
2483 * rq_pos_tree.
2484 */
2485 if (prev != bfqq->next_rq) {
2486 bfq_updated_next_req(bfqd, bfqq);
2487 /*
2488 * See comments on bfq_pos_tree_add_move() for
2489 * the unlikely().
2490 */
2491 if (unlikely(!bfqd->nonrot_with_queueing))
2492 bfq_pos_tree_add_move(bfqd, bfqq);
2493 }
2494 }
2495}
2496
2497/*
2498 * This function is called to notify the scheduler that the requests
2499 * rq and 'next' have been merged, with 'next' going away. BFQ
2500 * exploits this hook to address the following issue: if 'next' has a
2501 * fifo_time lower that rq, then the fifo_time of rq must be set to
2502 * the value of 'next', to not forget the greater age of 'next'.
2503 *
2504 * NOTE: in this function we assume that rq is in a bfq_queue, basing
2505 * on that rq is picked from the hash table q->elevator->hash, which,
2506 * in its turn, is filled only with I/O requests present in
2507 * bfq_queues, while BFQ is in use for the request queue q. In fact,
2508 * the function that fills this hash table (elv_rqhash_add) is called
2509 * only by bfq_insert_request.
2510 */
2511static void bfq_requests_merged(struct request_queue *q, struct request *rq,
2512 struct request *next)
2513{
2514 struct bfq_queue *bfqq = RQ_BFQQ(rq),
2515 *next_bfqq = RQ_BFQQ(next);
2516
2517 if (!bfqq)
2518 goto remove;
2519
2520 /*
2521 * If next and rq belong to the same bfq_queue and next is older
2522 * than rq, then reposition rq in the fifo (by substituting next
2523 * with rq). Otherwise, if next and rq belong to different
2524 * bfq_queues, never reposition rq: in fact, we would have to
2525 * reposition it with respect to next's position in its own fifo,
2526 * which would most certainly be too expensive with respect to
2527 * the benefits.
2528 */
2529 if (bfqq == next_bfqq &&
2530 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2531 next->fifo_time < rq->fifo_time) {
2532 list_del_init(&rq->queuelist);
2533 list_replace_init(&next->queuelist, &rq->queuelist);
2534 rq->fifo_time = next->fifo_time;
2535 }
2536
2537 if (bfqq->next_rq == next)
2538 bfqq->next_rq = rq;
2539
2540 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
2541remove:
2542 /* Merged request may be in the IO scheduler. Remove it. */
2543 if (!RB_EMPTY_NODE(&next->rb_node)) {
2544 bfq_remove_request(next->q, next);
2545 if (next_bfqq)
2546 bfqg_stats_update_io_remove(bfqq_group(next_bfqq),
2547 next->cmd_flags);
2548 }
2549}
2550
2551/* Must be called with bfqq != NULL */
2552static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
2553{
2554 /*
2555 * If bfqq has been enjoying interactive weight-raising, then
2556 * reset soft_rt_next_start. We do it for the following
2557 * reason. bfqq may have been conveying the I/O needed to load
2558 * a soft real-time application. Such an application actually
2559 * exhibits a soft real-time I/O pattern after it finishes
2560 * loading, and finally starts doing its job. But, if bfqq has
2561 * been receiving a lot of bandwidth so far (likely to happen
2562 * on a fast device), then soft_rt_next_start now contains a
2563 * high value that. So, without this reset, bfqq would be
2564 * prevented from being possibly considered as soft_rt for a
2565 * very long time.
2566 */
2567
2568 if (bfqq->wr_cur_max_time !=
2569 bfqq->bfqd->bfq_wr_rt_max_time)
2570 bfqq->soft_rt_next_start = jiffies;
2571
2572 if (bfq_bfqq_busy(bfqq))
2573 bfqq->bfqd->wr_busy_queues--;
2574 bfqq->wr_coeff = 1;
2575 bfqq->wr_cur_max_time = 0;
2576 bfqq->last_wr_start_finish = jiffies;
2577 /*
2578 * Trigger a weight change on the next invocation of
2579 * __bfq_entity_update_weight_prio.
2580 */
2581 bfqq->entity.prio_changed = 1;
2582}
2583
2584void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2585 struct bfq_group *bfqg)
2586{
2587 int i, j;
2588
2589 for (i = 0; i < 2; i++)
2590 for (j = 0; j < IOPRIO_NR_LEVELS; j++)
2591 if (bfqg->async_bfqq[i][j])
2592 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2593 if (bfqg->async_idle_bfqq)
2594 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2595}
2596
2597static void bfq_end_wr(struct bfq_data *bfqd)
2598{
2599 struct bfq_queue *bfqq;
2600
2601 spin_lock_irq(&bfqd->lock);
2602
2603 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2604 bfq_bfqq_end_wr(bfqq);
2605 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2606 bfq_bfqq_end_wr(bfqq);
2607 bfq_end_wr_async(bfqd);
2608
2609 spin_unlock_irq(&bfqd->lock);
2610}
2611
2612static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2613{
2614 if (request)
2615 return blk_rq_pos(io_struct);
2616 else
2617 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2618}
2619
2620static int bfq_rq_close_to_sector(void *io_struct, bool request,
2621 sector_t sector)
2622{
2623 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2624 BFQQ_CLOSE_THR;
2625}
2626
2627static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2628 struct bfq_queue *bfqq,
2629 sector_t sector)
2630{
2631 struct rb_root *root = &bfqq_group(bfqq)->rq_pos_tree;
2632 struct rb_node *parent, *node;
2633 struct bfq_queue *__bfqq;
2634
2635 if (RB_EMPTY_ROOT(root))
2636 return NULL;
2637
2638 /*
2639 * First, if we find a request starting at the end of the last
2640 * request, choose it.
2641 */
2642 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2643 if (__bfqq)
2644 return __bfqq;
2645
2646 /*
2647 * If the exact sector wasn't found, the parent of the NULL leaf
2648 * will contain the closest sector (rq_pos_tree sorted by
2649 * next_request position).
2650 */
2651 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2652 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2653 return __bfqq;
2654
2655 if (blk_rq_pos(__bfqq->next_rq) < sector)
2656 node = rb_next(&__bfqq->pos_node);
2657 else
2658 node = rb_prev(&__bfqq->pos_node);
2659 if (!node)
2660 return NULL;
2661
2662 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2663 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2664 return __bfqq;
2665
2666 return NULL;
2667}
2668
2669static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2670 struct bfq_queue *cur_bfqq,
2671 sector_t sector)
2672{
2673 struct bfq_queue *bfqq;
2674
2675 /*
2676 * We shall notice if some of the queues are cooperating,
2677 * e.g., working closely on the same area of the device. In
2678 * that case, we can group them together and: 1) don't waste
2679 * time idling, and 2) serve the union of their requests in
2680 * the best possible order for throughput.
2681 */
2682 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2683 if (!bfqq || bfqq == cur_bfqq)
2684 return NULL;
2685
2686 return bfqq;
2687}
2688
2689static struct bfq_queue *
2690bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2691{
2692 int process_refs, new_process_refs;
2693 struct bfq_queue *__bfqq;
2694
2695 /*
2696 * If there are no process references on the new_bfqq, then it is
2697 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2698 * may have dropped their last reference (not just their last process
2699 * reference).
2700 */
2701 if (!bfqq_process_refs(new_bfqq))
2702 return NULL;
2703
2704 /* Avoid a circular list and skip interim queue merges. */
2705 while ((__bfqq = new_bfqq->new_bfqq)) {
2706 if (__bfqq == bfqq)
2707 return NULL;
2708 new_bfqq = __bfqq;
2709 }
2710
2711 process_refs = bfqq_process_refs(bfqq);
2712 new_process_refs = bfqq_process_refs(new_bfqq);
2713 /*
2714 * If the process for the bfqq has gone away, there is no
2715 * sense in merging the queues.
2716 */
2717 if (process_refs == 0 || new_process_refs == 0)
2718 return NULL;
2719
2720 /*
2721 * Make sure merged queues belong to the same parent. Parents could
2722 * have changed since the time we decided the two queues are suitable
2723 * for merging.
2724 */
2725 if (new_bfqq->entity.parent != bfqq->entity.parent)
2726 return NULL;
2727
2728 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2729 new_bfqq->pid);
2730
2731 /*
2732 * Merging is just a redirection: the requests of the process
2733 * owning one of the two queues are redirected to the other queue.
2734 * The latter queue, in its turn, is set as shared if this is the
2735 * first time that the requests of some process are redirected to
2736 * it.
2737 *
2738 * We redirect bfqq to new_bfqq and not the opposite, because
2739 * we are in the context of the process owning bfqq, thus we
2740 * have the io_cq of this process. So we can immediately
2741 * configure this io_cq to redirect the requests of the
2742 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2743 * not available any more (new_bfqq->bic == NULL).
2744 *
2745 * Anyway, even in case new_bfqq coincides with the in-service
2746 * queue, redirecting requests the in-service queue is the
2747 * best option, as we feed the in-service queue with new
2748 * requests close to the last request served and, by doing so,
2749 * are likely to increase the throughput.
2750 */
2751 bfqq->new_bfqq = new_bfqq;
2752 /*
2753 * The above assignment schedules the following redirections:
2754 * each time some I/O for bfqq arrives, the process that
2755 * generated that I/O is disassociated from bfqq and
2756 * associated with new_bfqq. Here we increases new_bfqq->ref
2757 * in advance, adding the number of processes that are
2758 * expected to be associated with new_bfqq as they happen to
2759 * issue I/O.
2760 */
2761 new_bfqq->ref += process_refs;
2762 return new_bfqq;
2763}
2764
2765static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2766 struct bfq_queue *new_bfqq)
2767{
2768 if (bfq_too_late_for_merging(new_bfqq))
2769 return false;
2770
2771 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2772 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2773 return false;
2774
2775 /*
2776 * If either of the queues has already been detected as seeky,
2777 * then merging it with the other queue is unlikely to lead to
2778 * sequential I/O.
2779 */
2780 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2781 return false;
2782
2783 /*
2784 * Interleaved I/O is known to be done by (some) applications
2785 * only for reads, so it does not make sense to merge async
2786 * queues.
2787 */
2788 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2789 return false;
2790
2791 return true;
2792}
2793
2794static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
2795 struct bfq_queue *bfqq);
2796
2797/*
2798 * Attempt to schedule a merge of bfqq with the currently in-service
2799 * queue or with a close queue among the scheduled queues. Return
2800 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2801 * structure otherwise.
2802 *
2803 * The OOM queue is not allowed to participate to cooperation: in fact, since
2804 * the requests temporarily redirected to the OOM queue could be redirected
2805 * again to dedicated queues at any time, the state needed to correctly
2806 * handle merging with the OOM queue would be quite complex and expensive
2807 * to maintain. Besides, in such a critical condition as an out of memory,
2808 * the benefits of queue merging may be little relevant, or even negligible.
2809 *
2810 * WARNING: queue merging may impair fairness among non-weight raised
2811 * queues, for at least two reasons: 1) the original weight of a
2812 * merged queue may change during the merged state, 2) even being the
2813 * weight the same, a merged queue may be bloated with many more
2814 * requests than the ones produced by its originally-associated
2815 * process.
2816 */
2817static struct bfq_queue *
2818bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2819 void *io_struct, bool request, struct bfq_io_cq *bic)
2820{
2821 struct bfq_queue *in_service_bfqq, *new_bfqq;
2822
2823 /* if a merge has already been setup, then proceed with that first */
2824 if (bfqq->new_bfqq)
2825 return bfqq->new_bfqq;
2826
2827 /*
2828 * Check delayed stable merge for rotational or non-queueing
2829 * devs. For this branch to be executed, bfqq must not be
2830 * currently merged with some other queue (i.e., bfqq->bic
2831 * must be non null). If we considered also merged queues,
2832 * then we should also check whether bfqq has already been
2833 * merged with bic->stable_merge_bfqq. But this would be
2834 * costly and complicated.
2835 */
2836 if (unlikely(!bfqd->nonrot_with_queueing)) {
2837 /*
2838 * Make sure also that bfqq is sync, because
2839 * bic->stable_merge_bfqq may point to some queue (for
2840 * stable merging) also if bic is associated with a
2841 * sync queue, but this bfqq is async
2842 */
2843 if (bfq_bfqq_sync(bfqq) && bic->stable_merge_bfqq &&
2844 !bfq_bfqq_just_created(bfqq) &&
2845 time_is_before_jiffies(bfqq->split_time +
2846 msecs_to_jiffies(bfq_late_stable_merging)) &&
2847 time_is_before_jiffies(bfqq->creation_time +
2848 msecs_to_jiffies(bfq_late_stable_merging))) {
2849 struct bfq_queue *stable_merge_bfqq =
2850 bic->stable_merge_bfqq;
2851 int proc_ref = min(bfqq_process_refs(bfqq),
2852 bfqq_process_refs(stable_merge_bfqq));
2853
2854 /* deschedule stable merge, because done or aborted here */
2855 bfq_put_stable_ref(stable_merge_bfqq);
2856
2857 bic->stable_merge_bfqq = NULL;
2858
2859 if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
2860 proc_ref > 0) {
2861 /* next function will take at least one ref */
2862 struct bfq_queue *new_bfqq =
2863 bfq_setup_merge(bfqq, stable_merge_bfqq);
2864
2865 if (new_bfqq) {
2866 bic->stably_merged = true;
2867 if (new_bfqq->bic)
2868 new_bfqq->bic->stably_merged =
2869 true;
2870 }
2871 return new_bfqq;
2872 } else
2873 return NULL;
2874 }
2875 }
2876
2877 /*
2878 * Do not perform queue merging if the device is non
2879 * rotational and performs internal queueing. In fact, such a
2880 * device reaches a high speed through internal parallelism
2881 * and pipelining. This means that, to reach a high
2882 * throughput, it must have many requests enqueued at the same
2883 * time. But, in this configuration, the internal scheduling
2884 * algorithm of the device does exactly the job of queue
2885 * merging: it reorders requests so as to obtain as much as
2886 * possible a sequential I/O pattern. As a consequence, with
2887 * the workload generated by processes doing interleaved I/O,
2888 * the throughput reached by the device is likely to be the
2889 * same, with and without queue merging.
2890 *
2891 * Disabling merging also provides a remarkable benefit in
2892 * terms of throughput. Merging tends to make many workloads
2893 * artificially more uneven, because of shared queues
2894 * remaining non empty for incomparably more time than
2895 * non-merged queues. This may accentuate workload
2896 * asymmetries. For example, if one of the queues in a set of
2897 * merged queues has a higher weight than a normal queue, then
2898 * the shared queue may inherit such a high weight and, by
2899 * staying almost always active, may force BFQ to perform I/O
2900 * plugging most of the time. This evidently makes it harder
2901 * for BFQ to let the device reach a high throughput.
2902 *
2903 * Finally, the likely() macro below is not used because one
2904 * of the two branches is more likely than the other, but to
2905 * have the code path after the following if() executed as
2906 * fast as possible for the case of a non rotational device
2907 * with queueing. We want it because this is the fastest kind
2908 * of device. On the opposite end, the likely() may lengthen
2909 * the execution time of BFQ for the case of slower devices
2910 * (rotational or at least without queueing). But in this case
2911 * the execution time of BFQ matters very little, if not at
2912 * all.
2913 */
2914 if (likely(bfqd->nonrot_with_queueing))
2915 return NULL;
2916
2917 /*
2918 * Prevent bfqq from being merged if it has been created too
2919 * long ago. The idea is that true cooperating processes, and
2920 * thus their associated bfq_queues, are supposed to be
2921 * created shortly after each other. This is the case, e.g.,
2922 * for KVM/QEMU and dump I/O threads. Basing on this
2923 * assumption, the following filtering greatly reduces the
2924 * probability that two non-cooperating processes, which just
2925 * happen to do close I/O for some short time interval, have
2926 * their queues merged by mistake.
2927 */
2928 if (bfq_too_late_for_merging(bfqq))
2929 return NULL;
2930
2931 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2932 return NULL;
2933
2934 /* If there is only one backlogged queue, don't search. */
2935 if (bfq_tot_busy_queues(bfqd) == 1)
2936 return NULL;
2937
2938 in_service_bfqq = bfqd->in_service_queue;
2939
2940 if (in_service_bfqq && in_service_bfqq != bfqq &&
2941 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2942 bfq_rq_close_to_sector(io_struct, request,
2943 bfqd->in_serv_last_pos) &&
2944 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2945 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2946 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2947 if (new_bfqq)
2948 return new_bfqq;
2949 }
2950 /*
2951 * Check whether there is a cooperator among currently scheduled
2952 * queues. The only thing we need is that the bio/request is not
2953 * NULL, as we need it to establish whether a cooperator exists.
2954 */
2955 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2956 bfq_io_struct_pos(io_struct, request));
2957
2958 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2959 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2960 return bfq_setup_merge(bfqq, new_bfqq);
2961
2962 return NULL;
2963}
2964
2965static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2966{
2967 struct bfq_io_cq *bic = bfqq->bic;
2968
2969 /*
2970 * If !bfqq->bic, the queue is already shared or its requests
2971 * have already been redirected to a shared queue; both idle window
2972 * and weight raising state have already been saved. Do nothing.
2973 */
2974 if (!bic)
2975 return;
2976
2977 bic->saved_last_serv_time_ns = bfqq->last_serv_time_ns;
2978 bic->saved_inject_limit = bfqq->inject_limit;
2979 bic->saved_decrease_time_jif = bfqq->decrease_time_jif;
2980
2981 bic->saved_weight = bfqq->entity.orig_weight;
2982 bic->saved_ttime = bfqq->ttime;
2983 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2984 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2985 bic->saved_io_start_time = bfqq->io_start_time;
2986 bic->saved_tot_idle_time = bfqq->tot_idle_time;
2987 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2988 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2989 if (unlikely(bfq_bfqq_just_created(bfqq) &&
2990 !bfq_bfqq_in_large_burst(bfqq) &&
2991 bfqq->bfqd->low_latency)) {
2992 /*
2993 * bfqq being merged right after being created: bfqq
2994 * would have deserved interactive weight raising, but
2995 * did not make it to be set in a weight-raised state,
2996 * because of this early merge. Store directly the
2997 * weight-raising state that would have been assigned
2998 * to bfqq, so that to avoid that bfqq unjustly fails
2999 * to enjoy weight raising if split soon.
3000 */
3001 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
3002 bic->saved_wr_start_at_switch_to_srt = bfq_smallest_from_now();
3003 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
3004 bic->saved_last_wr_start_finish = jiffies;
3005 } else {
3006 bic->saved_wr_coeff = bfqq->wr_coeff;
3007 bic->saved_wr_start_at_switch_to_srt =
3008 bfqq->wr_start_at_switch_to_srt;
3009 bic->saved_service_from_wr = bfqq->service_from_wr;
3010 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
3011 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
3012 }
3013}
3014
3015
3016static void
3017bfq_reassign_last_bfqq(struct bfq_queue *cur_bfqq, struct bfq_queue *new_bfqq)
3018{
3019 if (cur_bfqq->entity.parent &&
3020 cur_bfqq->entity.parent->last_bfqq_created == cur_bfqq)
3021 cur_bfqq->entity.parent->last_bfqq_created = new_bfqq;
3022 else if (cur_bfqq->bfqd && cur_bfqq->bfqd->last_bfqq_created == cur_bfqq)
3023 cur_bfqq->bfqd->last_bfqq_created = new_bfqq;
3024}
3025
3026void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3027{
3028 /*
3029 * To prevent bfqq's service guarantees from being violated,
3030 * bfqq may be left busy, i.e., queued for service, even if
3031 * empty (see comments in __bfq_bfqq_expire() for
3032 * details). But, if no process will send requests to bfqq any
3033 * longer, then there is no point in keeping bfqq queued for
3034 * service. In addition, keeping bfqq queued for service, but
3035 * with no process ref any longer, may have caused bfqq to be
3036 * freed when dequeued from service. But this is assumed to
3037 * never happen.
3038 */
3039 if (bfq_bfqq_busy(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) &&
3040 bfqq != bfqd->in_service_queue)
3041 bfq_del_bfqq_busy(bfqq, false);
3042
3043 bfq_reassign_last_bfqq(bfqq, NULL);
3044
3045 bfq_put_queue(bfqq);
3046}
3047
3048static void
3049bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
3050 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
3051{
3052 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
3053 (unsigned long)new_bfqq->pid);
3054 /* Save weight raising and idle window of the merged queues */
3055 bfq_bfqq_save_state(bfqq);
3056 bfq_bfqq_save_state(new_bfqq);
3057 if (bfq_bfqq_IO_bound(bfqq))
3058 bfq_mark_bfqq_IO_bound(new_bfqq);
3059 bfq_clear_bfqq_IO_bound(bfqq);
3060
3061 /*
3062 * The processes associated with bfqq are cooperators of the
3063 * processes associated with new_bfqq. So, if bfqq has a
3064 * waker, then assume that all these processes will be happy
3065 * to let bfqq's waker freely inject I/O when they have no
3066 * I/O.
3067 */
3068 if (bfqq->waker_bfqq && !new_bfqq->waker_bfqq &&
3069 bfqq->waker_bfqq != new_bfqq) {
3070 new_bfqq->waker_bfqq = bfqq->waker_bfqq;
3071 new_bfqq->tentative_waker_bfqq = NULL;
3072
3073 /*
3074 * If the waker queue disappears, then
3075 * new_bfqq->waker_bfqq must be reset. So insert
3076 * new_bfqq into the woken_list of the waker. See
3077 * bfq_check_waker for details.
3078 */
3079 hlist_add_head(&new_bfqq->woken_list_node,
3080 &new_bfqq->waker_bfqq->woken_list);
3081
3082 }
3083
3084 /*
3085 * If bfqq is weight-raised, then let new_bfqq inherit
3086 * weight-raising. To reduce false positives, neglect the case
3087 * where bfqq has just been created, but has not yet made it
3088 * to be weight-raised (which may happen because EQM may merge
3089 * bfqq even before bfq_add_request is executed for the first
3090 * time for bfqq). Handling this case would however be very
3091 * easy, thanks to the flag just_created.
3092 */
3093 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
3094 new_bfqq->wr_coeff = bfqq->wr_coeff;
3095 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
3096 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
3097 new_bfqq->wr_start_at_switch_to_srt =
3098 bfqq->wr_start_at_switch_to_srt;
3099 if (bfq_bfqq_busy(new_bfqq))
3100 bfqd->wr_busy_queues++;
3101 new_bfqq->entity.prio_changed = 1;
3102 }
3103
3104 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
3105 bfqq->wr_coeff = 1;
3106 bfqq->entity.prio_changed = 1;
3107 if (bfq_bfqq_busy(bfqq))
3108 bfqd->wr_busy_queues--;
3109 }
3110
3111 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
3112 bfqd->wr_busy_queues);
3113
3114 /*
3115 * Merge queues (that is, let bic redirect its requests to new_bfqq)
3116 */
3117 bic_set_bfqq(bic, new_bfqq, true);
3118 bfq_mark_bfqq_coop(new_bfqq);
3119 /*
3120 * new_bfqq now belongs to at least two bics (it is a shared queue):
3121 * set new_bfqq->bic to NULL. bfqq either:
3122 * - does not belong to any bic any more, and hence bfqq->bic must
3123 * be set to NULL, or
3124 * - is a queue whose owning bics have already been redirected to a
3125 * different queue, hence the queue is destined to not belong to
3126 * any bic soon and bfqq->bic is already NULL (therefore the next
3127 * assignment causes no harm).
3128 */
3129 new_bfqq->bic = NULL;
3130 /*
3131 * If the queue is shared, the pid is the pid of one of the associated
3132 * processes. Which pid depends on the exact sequence of merge events
3133 * the queue underwent. So printing such a pid is useless and confusing
3134 * because it reports a random pid between those of the associated
3135 * processes.
3136 * We mark such a queue with a pid -1, and then print SHARED instead of
3137 * a pid in logging messages.
3138 */
3139 new_bfqq->pid = -1;
3140 bfqq->bic = NULL;
3141
3142 bfq_reassign_last_bfqq(bfqq, new_bfqq);
3143
3144 bfq_release_process_ref(bfqd, bfqq);
3145}
3146
3147static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
3148 struct bio *bio)
3149{
3150 struct bfq_data *bfqd = q->elevator->elevator_data;
3151 bool is_sync = op_is_sync(bio->bi_opf);
3152 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
3153
3154 /*
3155 * Disallow merge of a sync bio into an async request.
3156 */
3157 if (is_sync && !rq_is_sync(rq))
3158 return false;
3159
3160 /*
3161 * Lookup the bfqq that this bio will be queued with. Allow
3162 * merge only if rq is queued there.
3163 */
3164 if (!bfqq)
3165 return false;
3166
3167 /*
3168 * We take advantage of this function to perform an early merge
3169 * of the queues of possible cooperating processes.
3170 */
3171 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false, bfqd->bio_bic);
3172 if (new_bfqq) {
3173 /*
3174 * bic still points to bfqq, then it has not yet been
3175 * redirected to some other bfq_queue, and a queue
3176 * merge between bfqq and new_bfqq can be safely
3177 * fulfilled, i.e., bic can be redirected to new_bfqq
3178 * and bfqq can be put.
3179 */
3180 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
3181 new_bfqq);
3182 /*
3183 * If we get here, bio will be queued into new_queue,
3184 * so use new_bfqq to decide whether bio and rq can be
3185 * merged.
3186 */
3187 bfqq = new_bfqq;
3188
3189 /*
3190 * Change also bqfd->bio_bfqq, as
3191 * bfqd->bio_bic now points to new_bfqq, and
3192 * this function may be invoked again (and then may
3193 * use again bqfd->bio_bfqq).
3194 */
3195 bfqd->bio_bfqq = bfqq;
3196 }
3197
3198 return bfqq == RQ_BFQQ(rq);
3199}
3200
3201/*
3202 * Set the maximum time for the in-service queue to consume its
3203 * budget. This prevents seeky processes from lowering the throughput.
3204 * In practice, a time-slice service scheme is used with seeky
3205 * processes.
3206 */
3207static void bfq_set_budget_timeout(struct bfq_data *bfqd,
3208 struct bfq_queue *bfqq)
3209{
3210 unsigned int timeout_coeff;
3211
3212 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
3213 timeout_coeff = 1;
3214 else
3215 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
3216
3217 bfqd->last_budget_start = ktime_get();
3218
3219 bfqq->budget_timeout = jiffies +
3220 bfqd->bfq_timeout * timeout_coeff;
3221}
3222
3223static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
3224 struct bfq_queue *bfqq)
3225{
3226 if (bfqq) {
3227 bfq_clear_bfqq_fifo_expire(bfqq);
3228
3229 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
3230
3231 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
3232 bfqq->wr_coeff > 1 &&
3233 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
3234 time_is_before_jiffies(bfqq->budget_timeout)) {
3235 /*
3236 * For soft real-time queues, move the start
3237 * of the weight-raising period forward by the
3238 * time the queue has not received any
3239 * service. Otherwise, a relatively long
3240 * service delay is likely to cause the
3241 * weight-raising period of the queue to end,
3242 * because of the short duration of the
3243 * weight-raising period of a soft real-time
3244 * queue. It is worth noting that this move
3245 * is not so dangerous for the other queues,
3246 * because soft real-time queues are not
3247 * greedy.
3248 *
3249 * To not add a further variable, we use the
3250 * overloaded field budget_timeout to
3251 * determine for how long the queue has not
3252 * received service, i.e., how much time has
3253 * elapsed since the queue expired. However,
3254 * this is a little imprecise, because
3255 * budget_timeout is set to jiffies if bfqq
3256 * not only expires, but also remains with no
3257 * request.
3258 */
3259 if (time_after(bfqq->budget_timeout,
3260 bfqq->last_wr_start_finish))
3261 bfqq->last_wr_start_finish +=
3262 jiffies - bfqq->budget_timeout;
3263 else
3264 bfqq->last_wr_start_finish = jiffies;
3265 }
3266
3267 bfq_set_budget_timeout(bfqd, bfqq);
3268 bfq_log_bfqq(bfqd, bfqq,
3269 "set_in_service_queue, cur-budget = %d",
3270 bfqq->entity.budget);
3271 }
3272
3273 bfqd->in_service_queue = bfqq;
3274 bfqd->in_serv_last_pos = 0;
3275}
3276
3277/*
3278 * Get and set a new queue for service.
3279 */
3280static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
3281{
3282 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
3283
3284 __bfq_set_in_service_queue(bfqd, bfqq);
3285 return bfqq;
3286}
3287
3288static void bfq_arm_slice_timer(struct bfq_data *bfqd)
3289{
3290 struct bfq_queue *bfqq = bfqd->in_service_queue;
3291 u32 sl;
3292
3293 bfq_mark_bfqq_wait_request(bfqq);
3294
3295 /*
3296 * We don't want to idle for seeks, but we do want to allow
3297 * fair distribution of slice time for a process doing back-to-back
3298 * seeks. So allow a little bit of time for him to submit a new rq.
3299 */
3300 sl = bfqd->bfq_slice_idle;
3301 /*
3302 * Unless the queue is being weight-raised or the scenario is
3303 * asymmetric, grant only minimum idle time if the queue
3304 * is seeky. A long idling is preserved for a weight-raised
3305 * queue, or, more in general, in an asymmetric scenario,
3306 * because a long idling is needed for guaranteeing to a queue
3307 * its reserved share of the throughput (in particular, it is
3308 * needed if the queue has a higher weight than some other
3309 * queue).
3310 */
3311 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3312 !bfq_asymmetric_scenario(bfqd, bfqq))
3313 sl = min_t(u64, sl, BFQ_MIN_TT);
3314 else if (bfqq->wr_coeff > 1)
3315 sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
3316
3317 bfqd->last_idling_start = ktime_get();
3318 bfqd->last_idling_start_jiffies = jiffies;
3319
3320 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
3321 HRTIMER_MODE_REL);
3322 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
3323}
3324
3325/*
3326 * In autotuning mode, max_budget is dynamically recomputed as the
3327 * amount of sectors transferred in timeout at the estimated peak
3328 * rate. This enables BFQ to utilize a full timeslice with a full
3329 * budget, even if the in-service queue is served at peak rate. And
3330 * this maximises throughput with sequential workloads.
3331 */
3332static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
3333{
3334 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
3335 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
3336}
3337
3338/*
3339 * Update parameters related to throughput and responsiveness, as a
3340 * function of the estimated peak rate. See comments on
3341 * bfq_calc_max_budget(), and on the ref_wr_duration array.
3342 */
3343static void update_thr_responsiveness_params(struct bfq_data *bfqd)
3344{
3345 if (bfqd->bfq_user_max_budget == 0) {
3346 bfqd->bfq_max_budget =
3347 bfq_calc_max_budget(bfqd);
3348 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
3349 }
3350}
3351
3352static void bfq_reset_rate_computation(struct bfq_data *bfqd,
3353 struct request *rq)
3354{
3355 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
3356 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
3357 bfqd->peak_rate_samples = 1;
3358 bfqd->sequential_samples = 0;
3359 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
3360 blk_rq_sectors(rq);
3361 } else /* no new rq dispatched, just reset the number of samples */
3362 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
3363
3364 bfq_log(bfqd,
3365 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
3366 bfqd->peak_rate_samples, bfqd->sequential_samples,
3367 bfqd->tot_sectors_dispatched);
3368}
3369
3370static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
3371{
3372 u32 rate, weight, divisor;
3373
3374 /*
3375 * For the convergence property to hold (see comments on
3376 * bfq_update_peak_rate()) and for the assessment to be
3377 * reliable, a minimum number of samples must be present, and
3378 * a minimum amount of time must have elapsed. If not so, do
3379 * not compute new rate. Just reset parameters, to get ready
3380 * for a new evaluation attempt.
3381 */
3382 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
3383 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
3384 goto reset_computation;
3385
3386 /*
3387 * If a new request completion has occurred after last
3388 * dispatch, then, to approximate the rate at which requests
3389 * have been served by the device, it is more precise to
3390 * extend the observation interval to the last completion.
3391 */
3392 bfqd->delta_from_first =
3393 max_t(u64, bfqd->delta_from_first,
3394 bfqd->last_completion - bfqd->first_dispatch);
3395
3396 /*
3397 * Rate computed in sects/usec, and not sects/nsec, for
3398 * precision issues.
3399 */
3400 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
3401 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
3402
3403 /*
3404 * Peak rate not updated if:
3405 * - the percentage of sequential dispatches is below 3/4 of the
3406 * total, and rate is below the current estimated peak rate
3407 * - rate is unreasonably high (> 20M sectors/sec)
3408 */
3409 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
3410 rate <= bfqd->peak_rate) ||
3411 rate > 20<<BFQ_RATE_SHIFT)
3412 goto reset_computation;
3413
3414 /*
3415 * We have to update the peak rate, at last! To this purpose,
3416 * we use a low-pass filter. We compute the smoothing constant
3417 * of the filter as a function of the 'weight' of the new
3418 * measured rate.
3419 *
3420 * As can be seen in next formulas, we define this weight as a
3421 * quantity proportional to how sequential the workload is,
3422 * and to how long the observation time interval is.
3423 *
3424 * The weight runs from 0 to 8. The maximum value of the
3425 * weight, 8, yields the minimum value for the smoothing
3426 * constant. At this minimum value for the smoothing constant,
3427 * the measured rate contributes for half of the next value of
3428 * the estimated peak rate.
3429 *
3430 * So, the first step is to compute the weight as a function
3431 * of how sequential the workload is. Note that the weight
3432 * cannot reach 9, because bfqd->sequential_samples cannot
3433 * become equal to bfqd->peak_rate_samples, which, in its
3434 * turn, holds true because bfqd->sequential_samples is not
3435 * incremented for the first sample.
3436 */
3437 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
3438
3439 /*
3440 * Second step: further refine the weight as a function of the
3441 * duration of the observation interval.
3442 */
3443 weight = min_t(u32, 8,
3444 div_u64(weight * bfqd->delta_from_first,
3445 BFQ_RATE_REF_INTERVAL));
3446
3447 /*
3448 * Divisor ranging from 10, for minimum weight, to 2, for
3449 * maximum weight.
3450 */
3451 divisor = 10 - weight;
3452
3453 /*
3454 * Finally, update peak rate:
3455 *
3456 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
3457 */
3458 bfqd->peak_rate *= divisor-1;
3459 bfqd->peak_rate /= divisor;
3460 rate /= divisor; /* smoothing constant alpha = 1/divisor */
3461
3462 bfqd->peak_rate += rate;
3463
3464 /*
3465 * For a very slow device, bfqd->peak_rate can reach 0 (see
3466 * the minimum representable values reported in the comments
3467 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3468 * divisions by zero where bfqd->peak_rate is used as a
3469 * divisor.
3470 */
3471 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
3472
3473 update_thr_responsiveness_params(bfqd);
3474
3475reset_computation:
3476 bfq_reset_rate_computation(bfqd, rq);
3477}
3478
3479/*
3480 * Update the read/write peak rate (the main quantity used for
3481 * auto-tuning, see update_thr_responsiveness_params()).
3482 *
3483 * It is not trivial to estimate the peak rate (correctly): because of
3484 * the presence of sw and hw queues between the scheduler and the
3485 * device components that finally serve I/O requests, it is hard to
3486 * say exactly when a given dispatched request is served inside the
3487 * device, and for how long. As a consequence, it is hard to know
3488 * precisely at what rate a given set of requests is actually served
3489 * by the device.
3490 *
3491 * On the opposite end, the dispatch time of any request is trivially
3492 * available, and, from this piece of information, the "dispatch rate"
3493 * of requests can be immediately computed. So, the idea in the next
3494 * function is to use what is known, namely request dispatch times
3495 * (plus, when useful, request completion times), to estimate what is
3496 * unknown, namely in-device request service rate.
3497 *
3498 * The main issue is that, because of the above facts, the rate at
3499 * which a certain set of requests is dispatched over a certain time
3500 * interval can vary greatly with respect to the rate at which the
3501 * same requests are then served. But, since the size of any
3502 * intermediate queue is limited, and the service scheme is lossless
3503 * (no request is silently dropped), the following obvious convergence
3504 * property holds: the number of requests dispatched MUST become
3505 * closer and closer to the number of requests completed as the
3506 * observation interval grows. This is the key property used in
3507 * the next function to estimate the peak service rate as a function
3508 * of the observed dispatch rate. The function assumes to be invoked
3509 * on every request dispatch.
3510 */
3511static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
3512{
3513 u64 now_ns = ktime_get_ns();
3514
3515 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
3516 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
3517 bfqd->peak_rate_samples);
3518 bfq_reset_rate_computation(bfqd, rq);
3519 goto update_last_values; /* will add one sample */
3520 }
3521
3522 /*
3523 * Device idle for very long: the observation interval lasting
3524 * up to this dispatch cannot be a valid observation interval
3525 * for computing a new peak rate (similarly to the late-
3526 * completion event in bfq_completed_request()). Go to
3527 * update_rate_and_reset to have the following three steps
3528 * taken:
3529 * - close the observation interval at the last (previous)
3530 * request dispatch or completion
3531 * - compute rate, if possible, for that observation interval
3532 * - start a new observation interval with this dispatch
3533 */
3534 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
3535 bfqd->rq_in_driver == 0)
3536 goto update_rate_and_reset;
3537
3538 /* Update sampling information */
3539 bfqd->peak_rate_samples++;
3540
3541 if ((bfqd->rq_in_driver > 0 ||
3542 now_ns - bfqd->last_completion < BFQ_MIN_TT)
3543 && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
3544 bfqd->sequential_samples++;
3545
3546 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
3547
3548 /* Reset max observed rq size every 32 dispatches */
3549 if (likely(bfqd->peak_rate_samples % 32))
3550 bfqd->last_rq_max_size =
3551 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
3552 else
3553 bfqd->last_rq_max_size = blk_rq_sectors(rq);
3554
3555 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
3556
3557 /* Target observation interval not yet reached, go on sampling */
3558 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
3559 goto update_last_values;
3560
3561update_rate_and_reset:
3562 bfq_update_rate_reset(bfqd, rq);
3563update_last_values:
3564 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
3565 if (RQ_BFQQ(rq) == bfqd->in_service_queue)
3566 bfqd->in_serv_last_pos = bfqd->last_position;
3567 bfqd->last_dispatch = now_ns;
3568}
3569
3570/*
3571 * Remove request from internal lists.
3572 */
3573static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
3574{
3575 struct bfq_queue *bfqq = RQ_BFQQ(rq);
3576
3577 /*
3578 * For consistency, the next instruction should have been
3579 * executed after removing the request from the queue and
3580 * dispatching it. We execute instead this instruction before
3581 * bfq_remove_request() (and hence introduce a temporary
3582 * inconsistency), for efficiency. In fact, should this
3583 * dispatch occur for a non in-service bfqq, this anticipated
3584 * increment prevents two counters related to bfqq->dispatched
3585 * from risking to be, first, uselessly decremented, and then
3586 * incremented again when the (new) value of bfqq->dispatched
3587 * happens to be taken into account.
3588 */
3589 bfqq->dispatched++;
3590 bfq_update_peak_rate(q->elevator->elevator_data, rq);
3591
3592 bfq_remove_request(q, rq);
3593}
3594
3595/*
3596 * There is a case where idling does not have to be performed for
3597 * throughput concerns, but to preserve the throughput share of
3598 * the process associated with bfqq.
3599 *
3600 * To introduce this case, we can note that allowing the drive
3601 * to enqueue more than one request at a time, and hence
3602 * delegating de facto final scheduling decisions to the
3603 * drive's internal scheduler, entails loss of control on the
3604 * actual request service order. In particular, the critical
3605 * situation is when requests from different processes happen
3606 * to be present, at the same time, in the internal queue(s)
3607 * of the drive. In such a situation, the drive, by deciding
3608 * the service order of the internally-queued requests, does
3609 * determine also the actual throughput distribution among
3610 * these processes. But the drive typically has no notion or
3611 * concern about per-process throughput distribution, and
3612 * makes its decisions only on a per-request basis. Therefore,
3613 * the service distribution enforced by the drive's internal
3614 * scheduler is likely to coincide with the desired throughput
3615 * distribution only in a completely symmetric, or favorably
3616 * skewed scenario where:
3617 * (i-a) each of these processes must get the same throughput as
3618 * the others,
3619 * (i-b) in case (i-a) does not hold, it holds that the process
3620 * associated with bfqq must receive a lower or equal
3621 * throughput than any of the other processes;
3622 * (ii) the I/O of each process has the same properties, in
3623 * terms of locality (sequential or random), direction
3624 * (reads or writes), request sizes, greediness
3625 * (from I/O-bound to sporadic), and so on;
3626
3627 * In fact, in such a scenario, the drive tends to treat the requests
3628 * of each process in about the same way as the requests of the
3629 * others, and thus to provide each of these processes with about the
3630 * same throughput. This is exactly the desired throughput
3631 * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3632 * even more convenient distribution for (the process associated with)
3633 * bfqq.
3634 *
3635 * In contrast, in any asymmetric or unfavorable scenario, device
3636 * idling (I/O-dispatch plugging) is certainly needed to guarantee
3637 * that bfqq receives its assigned fraction of the device throughput
3638 * (see [1] for details).
3639 *
3640 * The problem is that idling may significantly reduce throughput with
3641 * certain combinations of types of I/O and devices. An important
3642 * example is sync random I/O on flash storage with command
3643 * queueing. So, unless bfqq falls in cases where idling also boosts
3644 * throughput, it is important to check conditions (i-a), i(-b) and
3645 * (ii) accurately, so as to avoid idling when not strictly needed for
3646 * service guarantees.
3647 *
3648 * Unfortunately, it is extremely difficult to thoroughly check
3649 * condition (ii). And, in case there are active groups, it becomes
3650 * very difficult to check conditions (i-a) and (i-b) too. In fact,
3651 * if there are active groups, then, for conditions (i-a) or (i-b) to
3652 * become false 'indirectly', it is enough that an active group
3653 * contains more active processes or sub-groups than some other active
3654 * group. More precisely, for conditions (i-a) or (i-b) to become
3655 * false because of such a group, it is not even necessary that the
3656 * group is (still) active: it is sufficient that, even if the group
3657 * has become inactive, some of its descendant processes still have
3658 * some request already dispatched but still waiting for
3659 * completion. In fact, requests have still to be guaranteed their
3660 * share of the throughput even after being dispatched. In this
3661 * respect, it is easy to show that, if a group frequently becomes
3662 * inactive while still having in-flight requests, and if, when this
3663 * happens, the group is not considered in the calculation of whether
3664 * the scenario is asymmetric, then the group may fail to be
3665 * guaranteed its fair share of the throughput (basically because
3666 * idling may not be performed for the descendant processes of the
3667 * group, but it had to be). We address this issue with the following
3668 * bi-modal behavior, implemented in the function
3669 * bfq_asymmetric_scenario().
3670 *
3671 * If there are groups with requests waiting for completion
3672 * (as commented above, some of these groups may even be
3673 * already inactive), then the scenario is tagged as
3674 * asymmetric, conservatively, without checking any of the
3675 * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3676 * This behavior matches also the fact that groups are created
3677 * exactly if controlling I/O is a primary concern (to
3678 * preserve bandwidth and latency guarantees).
3679 *
3680 * On the opposite end, if there are no groups with requests waiting
3681 * for completion, then only conditions (i-a) and (i-b) are actually
3682 * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3683 * idling is not performed, regardless of whether condition (ii)
3684 * holds. In other words, only if conditions (i-a) and (i-b) do not
3685 * hold, then idling is allowed, and the device tends to be prevented
3686 * from queueing many requests, possibly of several processes. Since
3687 * there are no groups with requests waiting for completion, then, to
3688 * control conditions (i-a) and (i-b) it is enough to check just
3689 * whether all the queues with requests waiting for completion also
3690 * have the same weight.
3691 *
3692 * Not checking condition (ii) evidently exposes bfqq to the
3693 * risk of getting less throughput than its fair share.
3694 * However, for queues with the same weight, a further
3695 * mechanism, preemption, mitigates or even eliminates this
3696 * problem. And it does so without consequences on overall
3697 * throughput. This mechanism and its benefits are explained
3698 * in the next three paragraphs.
3699 *
3700 * Even if a queue, say Q, is expired when it remains idle, Q
3701 * can still preempt the new in-service queue if the next
3702 * request of Q arrives soon (see the comments on
3703 * bfq_bfqq_update_budg_for_activation). If all queues and
3704 * groups have the same weight, this form of preemption,
3705 * combined with the hole-recovery heuristic described in the
3706 * comments on function bfq_bfqq_update_budg_for_activation,
3707 * are enough to preserve a correct bandwidth distribution in
3708 * the mid term, even without idling. In fact, even if not
3709 * idling allows the internal queues of the device to contain
3710 * many requests, and thus to reorder requests, we can rather
3711 * safely assume that the internal scheduler still preserves a
3712 * minimum of mid-term fairness.
3713 *
3714 * More precisely, this preemption-based, idleless approach
3715 * provides fairness in terms of IOPS, and not sectors per
3716 * second. This can be seen with a simple example. Suppose
3717 * that there are two queues with the same weight, but that
3718 * the first queue receives requests of 8 sectors, while the
3719 * second queue receives requests of 1024 sectors. In
3720 * addition, suppose that each of the two queues contains at
3721 * most one request at a time, which implies that each queue
3722 * always remains idle after it is served. Finally, after
3723 * remaining idle, each queue receives very quickly a new
3724 * request. It follows that the two queues are served
3725 * alternatively, preempting each other if needed. This
3726 * implies that, although both queues have the same weight,
3727 * the queue with large requests receives a service that is
3728 * 1024/8 times as high as the service received by the other
3729 * queue.
3730 *
3731 * The motivation for using preemption instead of idling (for
3732 * queues with the same weight) is that, by not idling,
3733 * service guarantees are preserved (completely or at least in
3734 * part) without minimally sacrificing throughput. And, if
3735 * there is no active group, then the primary expectation for
3736 * this device is probably a high throughput.
3737 *
3738 * We are now left only with explaining the two sub-conditions in the
3739 * additional compound condition that is checked below for deciding
3740 * whether the scenario is asymmetric. To explain the first
3741 * sub-condition, we need to add that the function
3742 * bfq_asymmetric_scenario checks the weights of only
3743 * non-weight-raised queues, for efficiency reasons (see comments on
3744 * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3745 * is checked explicitly here. More precisely, the compound condition
3746 * below takes into account also the fact that, even if bfqq is being
3747 * weight-raised, the scenario is still symmetric if all queues with
3748 * requests waiting for completion happen to be
3749 * weight-raised. Actually, we should be even more precise here, and
3750 * differentiate between interactive weight raising and soft real-time
3751 * weight raising.
3752 *
3753 * The second sub-condition checked in the compound condition is
3754 * whether there is a fair amount of already in-flight I/O not
3755 * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3756 * following reason. The drive may decide to serve in-flight
3757 * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3758 * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3759 * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3760 * basically uncontrolled amount of I/O from other queues may be
3761 * dispatched too, possibly causing the service of bfqq's I/O to be
3762 * delayed even longer in the drive. This problem gets more and more
3763 * serious as the speed and the queue depth of the drive grow,
3764 * because, as these two quantities grow, the probability to find no
3765 * queue busy but many requests in flight grows too. By contrast,
3766 * plugging I/O dispatching minimizes the delay induced by already
3767 * in-flight I/O, and enables bfqq to recover the bandwidth it may
3768 * lose because of this delay.
3769 *
3770 * As a side note, it is worth considering that the above
3771 * device-idling countermeasures may however fail in the following
3772 * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3773 * in a time period during which all symmetry sub-conditions hold, and
3774 * therefore the device is allowed to enqueue many requests, but at
3775 * some later point in time some sub-condition stops to hold, then it
3776 * may become impossible to make requests be served in the desired
3777 * order until all the requests already queued in the device have been
3778 * served. The last sub-condition commented above somewhat mitigates
3779 * this problem for weight-raised queues.
3780 *
3781 * However, as an additional mitigation for this problem, we preserve
3782 * plugging for a special symmetric case that may suddenly turn into
3783 * asymmetric: the case where only bfqq is busy. In this case, not
3784 * expiring bfqq does not cause any harm to any other queues in terms
3785 * of service guarantees. In contrast, it avoids the following unlucky
3786 * sequence of events: (1) bfqq is expired, (2) a new queue with a
3787 * lower weight than bfqq becomes busy (or more queues), (3) the new
3788 * queue is served until a new request arrives for bfqq, (4) when bfqq
3789 * is finally served, there are so many requests of the new queue in
3790 * the drive that the pending requests for bfqq take a lot of time to
3791 * be served. In particular, event (2) may case even already
3792 * dispatched requests of bfqq to be delayed, inside the drive. So, to
3793 * avoid this series of events, the scenario is preventively declared
3794 * as asymmetric also if bfqq is the only busy queues
3795 */
3796static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3797 struct bfq_queue *bfqq)
3798{
3799 int tot_busy_queues = bfq_tot_busy_queues(bfqd);
3800
3801 /* No point in idling for bfqq if it won't get requests any longer */
3802 if (unlikely(!bfqq_process_refs(bfqq)))
3803 return false;
3804
3805 return (bfqq->wr_coeff > 1 &&
3806 (bfqd->wr_busy_queues <
3807 tot_busy_queues ||
3808 bfqd->rq_in_driver >=
3809 bfqq->dispatched + 4)) ||
3810 bfq_asymmetric_scenario(bfqd, bfqq) ||
3811 tot_busy_queues == 1;
3812}
3813
3814static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3815 enum bfqq_expiration reason)
3816{
3817 /*
3818 * If this bfqq is shared between multiple processes, check
3819 * to make sure that those processes are still issuing I/Os
3820 * within the mean seek distance. If not, it may be time to
3821 * break the queues apart again.
3822 */
3823 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
3824 bfq_mark_bfqq_split_coop(bfqq);
3825
3826 /*
3827 * Consider queues with a higher finish virtual time than
3828 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3829 * true, then bfqq's bandwidth would be violated if an
3830 * uncontrolled amount of I/O from these queues were
3831 * dispatched while bfqq is waiting for its new I/O to
3832 * arrive. This is exactly what may happen if this is a forced
3833 * expiration caused by a preemption attempt, and if bfqq is
3834 * not re-scheduled. To prevent this from happening, re-queue
3835 * bfqq if it needs I/O-dispatch plugging, even if it is
3836 * empty. By doing so, bfqq is granted to be served before the
3837 * above queues (provided that bfqq is of course eligible).
3838 */
3839 if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
3840 !(reason == BFQQE_PREEMPTED &&
3841 idling_needed_for_service_guarantees(bfqd, bfqq))) {
3842 if (bfqq->dispatched == 0)
3843 /*
3844 * Overloading budget_timeout field to store
3845 * the time at which the queue remains with no
3846 * backlog and no outstanding request; used by
3847 * the weight-raising mechanism.
3848 */
3849 bfqq->budget_timeout = jiffies;
3850
3851 bfq_del_bfqq_busy(bfqq, true);
3852 } else {
3853 bfq_requeue_bfqq(bfqd, bfqq, true);
3854 /*
3855 * Resort priority tree of potential close cooperators.
3856 * See comments on bfq_pos_tree_add_move() for the unlikely().
3857 */
3858 if (unlikely(!bfqd->nonrot_with_queueing &&
3859 !RB_EMPTY_ROOT(&bfqq->sort_list)))
3860 bfq_pos_tree_add_move(bfqd, bfqq);
3861 }
3862
3863 /*
3864 * All in-service entities must have been properly deactivated
3865 * or requeued before executing the next function, which
3866 * resets all in-service entities as no more in service. This
3867 * may cause bfqq to be freed. If this happens, the next
3868 * function returns true.
3869 */
3870 return __bfq_bfqd_reset_in_service(bfqd);
3871}
3872
3873/**
3874 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3875 * @bfqd: device data.
3876 * @bfqq: queue to update.
3877 * @reason: reason for expiration.
3878 *
3879 * Handle the feedback on @bfqq budget at queue expiration.
3880 * See the body for detailed comments.
3881 */
3882static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
3883 struct bfq_queue *bfqq,
3884 enum bfqq_expiration reason)
3885{
3886 struct request *next_rq;
3887 int budget, min_budget;
3888
3889 min_budget = bfq_min_budget(bfqd);
3890
3891 if (bfqq->wr_coeff == 1)
3892 budget = bfqq->max_budget;
3893 else /*
3894 * Use a constant, low budget for weight-raised queues,
3895 * to help achieve a low latency. Keep it slightly higher
3896 * than the minimum possible budget, to cause a little
3897 * bit fewer expirations.
3898 */
3899 budget = 2 * min_budget;
3900
3901 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
3902 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
3903 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
3904 budget, bfq_min_budget(bfqd));
3905 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
3906 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
3907
3908 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
3909 switch (reason) {
3910 /*
3911 * Caveat: in all the following cases we trade latency
3912 * for throughput.
3913 */
3914 case BFQQE_TOO_IDLE:
3915 /*
3916 * This is the only case where we may reduce
3917 * the budget: if there is no request of the
3918 * process still waiting for completion, then
3919 * we assume (tentatively) that the timer has
3920 * expired because the batch of requests of
3921 * the process could have been served with a
3922 * smaller budget. Hence, betting that
3923 * process will behave in the same way when it
3924 * becomes backlogged again, we reduce its
3925 * next budget. As long as we guess right,
3926 * this budget cut reduces the latency
3927 * experienced by the process.
3928 *
3929 * However, if there are still outstanding
3930 * requests, then the process may have not yet
3931 * issued its next request just because it is
3932 * still waiting for the completion of some of
3933 * the still outstanding ones. So in this
3934 * subcase we do not reduce its budget, on the
3935 * contrary we increase it to possibly boost
3936 * the throughput, as discussed in the
3937 * comments to the BUDGET_TIMEOUT case.
3938 */
3939 if (bfqq->dispatched > 0) /* still outstanding reqs */
3940 budget = min(budget * 2, bfqd->bfq_max_budget);
3941 else {
3942 if (budget > 5 * min_budget)
3943 budget -= 4 * min_budget;
3944 else
3945 budget = min_budget;
3946 }
3947 break;
3948 case BFQQE_BUDGET_TIMEOUT:
3949 /*
3950 * We double the budget here because it gives
3951 * the chance to boost the throughput if this
3952 * is not a seeky process (and has bumped into
3953 * this timeout because of, e.g., ZBR).
3954 */
3955 budget = min(budget * 2, bfqd->bfq_max_budget);
3956 break;
3957 case BFQQE_BUDGET_EXHAUSTED:
3958 /*
3959 * The process still has backlog, and did not
3960 * let either the budget timeout or the disk
3961 * idling timeout expire. Hence it is not
3962 * seeky, has a short thinktime and may be
3963 * happy with a higher budget too. So
3964 * definitely increase the budget of this good
3965 * candidate to boost the disk throughput.
3966 */
3967 budget = min(budget * 4, bfqd->bfq_max_budget);
3968 break;
3969 case BFQQE_NO_MORE_REQUESTS:
3970 /*
3971 * For queues that expire for this reason, it
3972 * is particularly important to keep the
3973 * budget close to the actual service they
3974 * need. Doing so reduces the timestamp
3975 * misalignment problem described in the
3976 * comments in the body of
3977 * __bfq_activate_entity. In fact, suppose
3978 * that a queue systematically expires for
3979 * BFQQE_NO_MORE_REQUESTS and presents a
3980 * new request in time to enjoy timestamp
3981 * back-shifting. The larger the budget of the
3982 * queue is with respect to the service the
3983 * queue actually requests in each service
3984 * slot, the more times the queue can be
3985 * reactivated with the same virtual finish
3986 * time. It follows that, even if this finish
3987 * time is pushed to the system virtual time
3988 * to reduce the consequent timestamp
3989 * misalignment, the queue unjustly enjoys for
3990 * many re-activations a lower finish time
3991 * than all newly activated queues.
3992 *
3993 * The service needed by bfqq is measured
3994 * quite precisely by bfqq->entity.service.
3995 * Since bfqq does not enjoy device idling,
3996 * bfqq->entity.service is equal to the number
3997 * of sectors that the process associated with
3998 * bfqq requested to read/write before waiting
3999 * for request completions, or blocking for
4000 * other reasons.
4001 */
4002 budget = max_t(int, bfqq->entity.service, min_budget);
4003 break;
4004 default:
4005 return;
4006 }
4007 } else if (!bfq_bfqq_sync(bfqq)) {
4008 /*
4009 * Async queues get always the maximum possible
4010 * budget, as for them we do not care about latency
4011 * (in addition, their ability to dispatch is limited
4012 * by the charging factor).
4013 */
4014 budget = bfqd->bfq_max_budget;
4015 }
4016
4017 bfqq->max_budget = budget;
4018
4019 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
4020 !bfqd->bfq_user_max_budget)
4021 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
4022
4023 /*
4024 * If there is still backlog, then assign a new budget, making
4025 * sure that it is large enough for the next request. Since
4026 * the finish time of bfqq must be kept in sync with the
4027 * budget, be sure to call __bfq_bfqq_expire() *after* this
4028 * update.
4029 *
4030 * If there is no backlog, then no need to update the budget;
4031 * it will be updated on the arrival of a new request.
4032 */
4033 next_rq = bfqq->next_rq;
4034 if (next_rq)
4035 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
4036 bfq_serv_to_charge(next_rq, bfqq));
4037
4038 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
4039 next_rq ? blk_rq_sectors(next_rq) : 0,
4040 bfqq->entity.budget);
4041}
4042
4043/*
4044 * Return true if the process associated with bfqq is "slow". The slow
4045 * flag is used, in addition to the budget timeout, to reduce the
4046 * amount of service provided to seeky processes, and thus reduce
4047 * their chances to lower the throughput. More details in the comments
4048 * on the function bfq_bfqq_expire().
4049 *
4050 * An important observation is in order: as discussed in the comments
4051 * on the function bfq_update_peak_rate(), with devices with internal
4052 * queues, it is hard if ever possible to know when and for how long
4053 * an I/O request is processed by the device (apart from the trivial
4054 * I/O pattern where a new request is dispatched only after the
4055 * previous one has been completed). This makes it hard to evaluate
4056 * the real rate at which the I/O requests of each bfq_queue are
4057 * served. In fact, for an I/O scheduler like BFQ, serving a
4058 * bfq_queue means just dispatching its requests during its service
4059 * slot (i.e., until the budget of the queue is exhausted, or the
4060 * queue remains idle, or, finally, a timeout fires). But, during the
4061 * service slot of a bfq_queue, around 100 ms at most, the device may
4062 * be even still processing requests of bfq_queues served in previous
4063 * service slots. On the opposite end, the requests of the in-service
4064 * bfq_queue may be completed after the service slot of the queue
4065 * finishes.
4066 *
4067 * Anyway, unless more sophisticated solutions are used
4068 * (where possible), the sum of the sizes of the requests dispatched
4069 * during the service slot of a bfq_queue is probably the only
4070 * approximation available for the service received by the bfq_queue
4071 * during its service slot. And this sum is the quantity used in this
4072 * function to evaluate the I/O speed of a process.
4073 */
4074static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4075 bool compensate, enum bfqq_expiration reason,
4076 unsigned long *delta_ms)
4077{
4078 ktime_t delta_ktime;
4079 u32 delta_usecs;
4080 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
4081
4082 if (!bfq_bfqq_sync(bfqq))
4083 return false;
4084
4085 if (compensate)
4086 delta_ktime = bfqd->last_idling_start;
4087 else
4088 delta_ktime = ktime_get();
4089 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
4090 delta_usecs = ktime_to_us(delta_ktime);
4091
4092 /* don't use too short time intervals */
4093 if (delta_usecs < 1000) {
4094 if (blk_queue_nonrot(bfqd->queue))
4095 /*
4096 * give same worst-case guarantees as idling
4097 * for seeky
4098 */
4099 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
4100 else /* charge at least one seek */
4101 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
4102
4103 return slow;
4104 }
4105
4106 *delta_ms = delta_usecs / USEC_PER_MSEC;
4107
4108 /*
4109 * Use only long (> 20ms) intervals to filter out excessive
4110 * spikes in service rate estimation.
4111 */
4112 if (delta_usecs > 20000) {
4113 /*
4114 * Caveat for rotational devices: processes doing I/O
4115 * in the slower disk zones tend to be slow(er) even
4116 * if not seeky. In this respect, the estimated peak
4117 * rate is likely to be an average over the disk
4118 * surface. Accordingly, to not be too harsh with
4119 * unlucky processes, a process is deemed slow only if
4120 * its rate has been lower than half of the estimated
4121 * peak rate.
4122 */
4123 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
4124 }
4125
4126 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
4127
4128 return slow;
4129}
4130
4131/*
4132 * To be deemed as soft real-time, an application must meet two
4133 * requirements. First, the application must not require an average
4134 * bandwidth higher than the approximate bandwidth required to playback or
4135 * record a compressed high-definition video.
4136 * The next function is invoked on the completion of the last request of a
4137 * batch, to compute the next-start time instant, soft_rt_next_start, such
4138 * that, if the next request of the application does not arrive before
4139 * soft_rt_next_start, then the above requirement on the bandwidth is met.
4140 *
4141 * The second requirement is that the request pattern of the application is
4142 * isochronous, i.e., that, after issuing a request or a batch of requests,
4143 * the application stops issuing new requests until all its pending requests
4144 * have been completed. After that, the application may issue a new batch,
4145 * and so on.
4146 * For this reason the next function is invoked to compute
4147 * soft_rt_next_start only for applications that meet this requirement,
4148 * whereas soft_rt_next_start is set to infinity for applications that do
4149 * not.
4150 *
4151 * Unfortunately, even a greedy (i.e., I/O-bound) application may
4152 * happen to meet, occasionally or systematically, both the above
4153 * bandwidth and isochrony requirements. This may happen at least in
4154 * the following circumstances. First, if the CPU load is high. The
4155 * application may stop issuing requests while the CPUs are busy
4156 * serving other processes, then restart, then stop again for a while,
4157 * and so on. The other circumstances are related to the storage
4158 * device: the storage device is highly loaded or reaches a low-enough
4159 * throughput with the I/O of the application (e.g., because the I/O
4160 * is random and/or the device is slow). In all these cases, the
4161 * I/O of the application may be simply slowed down enough to meet
4162 * the bandwidth and isochrony requirements. To reduce the probability
4163 * that greedy applications are deemed as soft real-time in these
4164 * corner cases, a further rule is used in the computation of
4165 * soft_rt_next_start: the return value of this function is forced to
4166 * be higher than the maximum between the following two quantities.
4167 *
4168 * (a) Current time plus: (1) the maximum time for which the arrival
4169 * of a request is waited for when a sync queue becomes idle,
4170 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
4171 * postpone for a moment the reason for adding a few extra
4172 * jiffies; we get back to it after next item (b). Lower-bounding
4173 * the return value of this function with the current time plus
4174 * bfqd->bfq_slice_idle tends to filter out greedy applications,
4175 * because the latter issue their next request as soon as possible
4176 * after the last one has been completed. In contrast, a soft
4177 * real-time application spends some time processing data, after a
4178 * batch of its requests has been completed.
4179 *
4180 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
4181 * above, greedy applications may happen to meet both the
4182 * bandwidth and isochrony requirements under heavy CPU or
4183 * storage-device load. In more detail, in these scenarios, these
4184 * applications happen, only for limited time periods, to do I/O
4185 * slowly enough to meet all the requirements described so far,
4186 * including the filtering in above item (a). These slow-speed
4187 * time intervals are usually interspersed between other time
4188 * intervals during which these applications do I/O at a very high
4189 * speed. Fortunately, exactly because of the high speed of the
4190 * I/O in the high-speed intervals, the values returned by this
4191 * function happen to be so high, near the end of any such
4192 * high-speed interval, to be likely to fall *after* the end of
4193 * the low-speed time interval that follows. These high values are
4194 * stored in bfqq->soft_rt_next_start after each invocation of
4195 * this function. As a consequence, if the last value of
4196 * bfqq->soft_rt_next_start is constantly used to lower-bound the
4197 * next value that this function may return, then, from the very
4198 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
4199 * likely to be constantly kept so high that any I/O request
4200 * issued during the low-speed interval is considered as arriving
4201 * to soon for the application to be deemed as soft
4202 * real-time. Then, in the high-speed interval that follows, the
4203 * application will not be deemed as soft real-time, just because
4204 * it will do I/O at a high speed. And so on.
4205 *
4206 * Getting back to the filtering in item (a), in the following two
4207 * cases this filtering might be easily passed by a greedy
4208 * application, if the reference quantity was just
4209 * bfqd->bfq_slice_idle:
4210 * 1) HZ is so low that the duration of a jiffy is comparable to or
4211 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
4212 * devices with HZ=100. The time granularity may be so coarse
4213 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
4214 * is rather lower than the exact value.
4215 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
4216 * for a while, then suddenly 'jump' by several units to recover the lost
4217 * increments. This seems to happen, e.g., inside virtual machines.
4218 * To address this issue, in the filtering in (a) we do not use as a
4219 * reference time interval just bfqd->bfq_slice_idle, but
4220 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
4221 * minimum number of jiffies for which the filter seems to be quite
4222 * precise also in embedded systems and KVM/QEMU virtual machines.
4223 */
4224static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
4225 struct bfq_queue *bfqq)
4226{
4227 return max3(bfqq->soft_rt_next_start,
4228 bfqq->last_idle_bklogged +
4229 HZ * bfqq->service_from_backlogged /
4230 bfqd->bfq_wr_max_softrt_rate,
4231 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
4232}
4233
4234/**
4235 * bfq_bfqq_expire - expire a queue.
4236 * @bfqd: device owning the queue.
4237 * @bfqq: the queue to expire.
4238 * @compensate: if true, compensate for the time spent idling.
4239 * @reason: the reason causing the expiration.
4240 *
4241 * If the process associated with bfqq does slow I/O (e.g., because it
4242 * issues random requests), we charge bfqq with the time it has been
4243 * in service instead of the service it has received (see
4244 * bfq_bfqq_charge_time for details on how this goal is achieved). As
4245 * a consequence, bfqq will typically get higher timestamps upon
4246 * reactivation, and hence it will be rescheduled as if it had
4247 * received more service than what it has actually received. In the
4248 * end, bfqq receives less service in proportion to how slowly its
4249 * associated process consumes its budgets (and hence how seriously it
4250 * tends to lower the throughput). In addition, this time-charging
4251 * strategy guarantees time fairness among slow processes. In
4252 * contrast, if the process associated with bfqq is not slow, we
4253 * charge bfqq exactly with the service it has received.
4254 *
4255 * Charging time to the first type of queues and the exact service to
4256 * the other has the effect of using the WF2Q+ policy to schedule the
4257 * former on a timeslice basis, without violating service domain
4258 * guarantees among the latter.
4259 */
4260void bfq_bfqq_expire(struct bfq_data *bfqd,
4261 struct bfq_queue *bfqq,
4262 bool compensate,
4263 enum bfqq_expiration reason)
4264{
4265 bool slow;
4266 unsigned long delta = 0;
4267 struct bfq_entity *entity = &bfqq->entity;
4268
4269 /*
4270 * Check whether the process is slow (see bfq_bfqq_is_slow).
4271 */
4272 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
4273
4274 /*
4275 * As above explained, charge slow (typically seeky) and
4276 * timed-out queues with the time and not the service
4277 * received, to favor sequential workloads.
4278 *
4279 * Processes doing I/O in the slower disk zones will tend to
4280 * be slow(er) even if not seeky. Therefore, since the
4281 * estimated peak rate is actually an average over the disk
4282 * surface, these processes may timeout just for bad luck. To
4283 * avoid punishing them, do not charge time to processes that
4284 * succeeded in consuming at least 2/3 of their budget. This
4285 * allows BFQ to preserve enough elasticity to still perform
4286 * bandwidth, and not time, distribution with little unlucky
4287 * or quasi-sequential processes.
4288 */
4289 if (bfqq->wr_coeff == 1 &&
4290 (slow ||
4291 (reason == BFQQE_BUDGET_TIMEOUT &&
4292 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
4293 bfq_bfqq_charge_time(bfqd, bfqq, delta);
4294
4295 if (bfqd->low_latency && bfqq->wr_coeff == 1)
4296 bfqq->last_wr_start_finish = jiffies;
4297
4298 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
4299 RB_EMPTY_ROOT(&bfqq->sort_list)) {
4300 /*
4301 * If we get here, and there are no outstanding
4302 * requests, then the request pattern is isochronous
4303 * (see the comments on the function
4304 * bfq_bfqq_softrt_next_start()). Therefore we can
4305 * compute soft_rt_next_start.
4306 *
4307 * If, instead, the queue still has outstanding
4308 * requests, then we have to wait for the completion
4309 * of all the outstanding requests to discover whether
4310 * the request pattern is actually isochronous.
4311 */
4312 if (bfqq->dispatched == 0)
4313 bfqq->soft_rt_next_start =
4314 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4315 else if (bfqq->dispatched > 0) {
4316 /*
4317 * Schedule an update of soft_rt_next_start to when
4318 * the task may be discovered to be isochronous.
4319 */
4320 bfq_mark_bfqq_softrt_update(bfqq);
4321 }
4322 }
4323
4324 bfq_log_bfqq(bfqd, bfqq,
4325 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
4326 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
4327
4328 /*
4329 * bfqq expired, so no total service time needs to be computed
4330 * any longer: reset state machine for measuring total service
4331 * times.
4332 */
4333 bfqd->rqs_injected = bfqd->wait_dispatch = false;
4334 bfqd->waited_rq = NULL;
4335
4336 /*
4337 * Increase, decrease or leave budget unchanged according to
4338 * reason.
4339 */
4340 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
4341 if (__bfq_bfqq_expire(bfqd, bfqq, reason))
4342 /* bfqq is gone, no more actions on it */
4343 return;
4344
4345 /* mark bfqq as waiting a request only if a bic still points to it */
4346 if (!bfq_bfqq_busy(bfqq) &&
4347 reason != BFQQE_BUDGET_TIMEOUT &&
4348 reason != BFQQE_BUDGET_EXHAUSTED) {
4349 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
4350 /*
4351 * Not setting service to 0, because, if the next rq
4352 * arrives in time, the queue will go on receiving
4353 * service with this same budget (as if it never expired)
4354 */
4355 } else
4356 entity->service = 0;
4357
4358 /*
4359 * Reset the received-service counter for every parent entity.
4360 * Differently from what happens with bfqq->entity.service,
4361 * the resetting of this counter never needs to be postponed
4362 * for parent entities. In fact, in case bfqq may have a
4363 * chance to go on being served using the last, partially
4364 * consumed budget, bfqq->entity.service needs to be kept,
4365 * because if bfqq then actually goes on being served using
4366 * the same budget, the last value of bfqq->entity.service is
4367 * needed to properly decrement bfqq->entity.budget by the
4368 * portion already consumed. In contrast, it is not necessary
4369 * to keep entity->service for parent entities too, because
4370 * the bubble up of the new value of bfqq->entity.budget will
4371 * make sure that the budgets of parent entities are correct,
4372 * even in case bfqq and thus parent entities go on receiving
4373 * service with the same budget.
4374 */
4375 entity = entity->parent;
4376 for_each_entity(entity)
4377 entity->service = 0;
4378}
4379
4380/*
4381 * Budget timeout is not implemented through a dedicated timer, but
4382 * just checked on request arrivals and completions, as well as on
4383 * idle timer expirations.
4384 */
4385static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
4386{
4387 return time_is_before_eq_jiffies(bfqq->budget_timeout);
4388}
4389
4390/*
4391 * If we expire a queue that is actively waiting (i.e., with the
4392 * device idled) for the arrival of a new request, then we may incur
4393 * the timestamp misalignment problem described in the body of the
4394 * function __bfq_activate_entity. Hence we return true only if this
4395 * condition does not hold, or if the queue is slow enough to deserve
4396 * only to be kicked off for preserving a high throughput.
4397 */
4398static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
4399{
4400 bfq_log_bfqq(bfqq->bfqd, bfqq,
4401 "may_budget_timeout: wait_request %d left %d timeout %d",
4402 bfq_bfqq_wait_request(bfqq),
4403 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
4404 bfq_bfqq_budget_timeout(bfqq));
4405
4406 return (!bfq_bfqq_wait_request(bfqq) ||
4407 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
4408 &&
4409 bfq_bfqq_budget_timeout(bfqq);
4410}
4411
4412static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
4413 struct bfq_queue *bfqq)
4414{
4415 bool rot_without_queueing =
4416 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
4417 bfqq_sequential_and_IO_bound,
4418 idling_boosts_thr;
4419
4420 /* No point in idling for bfqq if it won't get requests any longer */
4421 if (unlikely(!bfqq_process_refs(bfqq)))
4422 return false;
4423
4424 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
4425 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
4426
4427 /*
4428 * The next variable takes into account the cases where idling
4429 * boosts the throughput.
4430 *
4431 * The value of the variable is computed considering, first, that
4432 * idling is virtually always beneficial for the throughput if:
4433 * (a) the device is not NCQ-capable and rotational, or
4434 * (b) regardless of the presence of NCQ, the device is rotational and
4435 * the request pattern for bfqq is I/O-bound and sequential, or
4436 * (c) regardless of whether it is rotational, the device is
4437 * not NCQ-capable and the request pattern for bfqq is
4438 * I/O-bound and sequential.
4439 *
4440 * Secondly, and in contrast to the above item (b), idling an
4441 * NCQ-capable flash-based device would not boost the
4442 * throughput even with sequential I/O; rather it would lower
4443 * the throughput in proportion to how fast the device
4444 * is. Accordingly, the next variable is true if any of the
4445 * above conditions (a), (b) or (c) is true, and, in
4446 * particular, happens to be false if bfqd is an NCQ-capable
4447 * flash-based device.
4448 */
4449 idling_boosts_thr = rot_without_queueing ||
4450 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
4451 bfqq_sequential_and_IO_bound);
4452
4453 /*
4454 * The return value of this function is equal to that of
4455 * idling_boosts_thr, unless a special case holds. In this
4456 * special case, described below, idling may cause problems to
4457 * weight-raised queues.
4458 *
4459 * When the request pool is saturated (e.g., in the presence
4460 * of write hogs), if the processes associated with
4461 * non-weight-raised queues ask for requests at a lower rate,
4462 * then processes associated with weight-raised queues have a
4463 * higher probability to get a request from the pool
4464 * immediately (or at least soon) when they need one. Thus
4465 * they have a higher probability to actually get a fraction
4466 * of the device throughput proportional to their high
4467 * weight. This is especially true with NCQ-capable drives,
4468 * which enqueue several requests in advance, and further
4469 * reorder internally-queued requests.
4470 *
4471 * For this reason, we force to false the return value if
4472 * there are weight-raised busy queues. In this case, and if
4473 * bfqq is not weight-raised, this guarantees that the device
4474 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4475 * then idling will be guaranteed by another variable, see
4476 * below). Combined with the timestamping rules of BFQ (see
4477 * [1] for details), this behavior causes bfqq, and hence any
4478 * sync non-weight-raised queue, to get a lower number of
4479 * requests served, and thus to ask for a lower number of
4480 * requests from the request pool, before the busy
4481 * weight-raised queues get served again. This often mitigates
4482 * starvation problems in the presence of heavy write
4483 * workloads and NCQ, thereby guaranteeing a higher
4484 * application and system responsiveness in these hostile
4485 * scenarios.
4486 */
4487 return idling_boosts_thr &&
4488 bfqd->wr_busy_queues == 0;
4489}
4490
4491/*
4492 * For a queue that becomes empty, device idling is allowed only if
4493 * this function returns true for that queue. As a consequence, since
4494 * device idling plays a critical role for both throughput boosting
4495 * and service guarantees, the return value of this function plays a
4496 * critical role as well.
4497 *
4498 * In a nutshell, this function returns true only if idling is
4499 * beneficial for throughput or, even if detrimental for throughput,
4500 * idling is however necessary to preserve service guarantees (low
4501 * latency, desired throughput distribution, ...). In particular, on
4502 * NCQ-capable devices, this function tries to return false, so as to
4503 * help keep the drives' internal queues full, whenever this helps the
4504 * device boost the throughput without causing any service-guarantee
4505 * issue.
4506 *
4507 * Most of the issues taken into account to get the return value of
4508 * this function are not trivial. We discuss these issues in the two
4509 * functions providing the main pieces of information needed by this
4510 * function.
4511 */
4512static bool bfq_better_to_idle(struct bfq_queue *bfqq)
4513{
4514 struct bfq_data *bfqd = bfqq->bfqd;
4515 bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
4516
4517 /* No point in idling for bfqq if it won't get requests any longer */
4518 if (unlikely(!bfqq_process_refs(bfqq)))
4519 return false;
4520
4521 if (unlikely(bfqd->strict_guarantees))
4522 return true;
4523
4524 /*
4525 * Idling is performed only if slice_idle > 0. In addition, we
4526 * do not idle if
4527 * (a) bfqq is async
4528 * (b) bfqq is in the idle io prio class: in this case we do
4529 * not idle because we want to minimize the bandwidth that
4530 * queues in this class can steal to higher-priority queues
4531 */
4532 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
4533 bfq_class_idle(bfqq))
4534 return false;
4535
4536 idling_boosts_thr_with_no_issue =
4537 idling_boosts_thr_without_issues(bfqd, bfqq);
4538
4539 idling_needed_for_service_guar =
4540 idling_needed_for_service_guarantees(bfqd, bfqq);
4541
4542 /*
4543 * We have now the two components we need to compute the
4544 * return value of the function, which is true only if idling
4545 * either boosts the throughput (without issues), or is
4546 * necessary to preserve service guarantees.
4547 */
4548 return idling_boosts_thr_with_no_issue ||
4549 idling_needed_for_service_guar;
4550}
4551
4552/*
4553 * If the in-service queue is empty but the function bfq_better_to_idle
4554 * returns true, then:
4555 * 1) the queue must remain in service and cannot be expired, and
4556 * 2) the device must be idled to wait for the possible arrival of a new
4557 * request for the queue.
4558 * See the comments on the function bfq_better_to_idle for the reasons
4559 * why performing device idling is the best choice to boost the throughput
4560 * and preserve service guarantees when bfq_better_to_idle itself
4561 * returns true.
4562 */
4563static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
4564{
4565 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
4566}
4567
4568/*
4569 * This function chooses the queue from which to pick the next extra
4570 * I/O request to inject, if it finds a compatible queue. See the
4571 * comments on bfq_update_inject_limit() for details on the injection
4572 * mechanism, and for the definitions of the quantities mentioned
4573 * below.
4574 */
4575static struct bfq_queue *
4576bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
4577{
4578 struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
4579 unsigned int limit = in_serv_bfqq->inject_limit;
4580 /*
4581 * If
4582 * - bfqq is not weight-raised and therefore does not carry
4583 * time-critical I/O,
4584 * or
4585 * - regardless of whether bfqq is weight-raised, bfqq has
4586 * however a long think time, during which it can absorb the
4587 * effect of an appropriate number of extra I/O requests
4588 * from other queues (see bfq_update_inject_limit for
4589 * details on the computation of this number);
4590 * then injection can be performed without restrictions.
4591 */
4592 bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
4593 !bfq_bfqq_has_short_ttime(in_serv_bfqq);
4594
4595 /*
4596 * If
4597 * - the baseline total service time could not be sampled yet,
4598 * so the inject limit happens to be still 0, and
4599 * - a lot of time has elapsed since the plugging of I/O
4600 * dispatching started, so drive speed is being wasted
4601 * significantly;
4602 * then temporarily raise inject limit to one request.
4603 */
4604 if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
4605 bfq_bfqq_wait_request(in_serv_bfqq) &&
4606 time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
4607 bfqd->bfq_slice_idle)
4608 )
4609 limit = 1;
4610
4611 if (bfqd->rq_in_driver >= limit)
4612 return NULL;
4613
4614 /*
4615 * Linear search of the source queue for injection; but, with
4616 * a high probability, very few steps are needed to find a
4617 * candidate queue, i.e., a queue with enough budget left for
4618 * its next request. In fact:
4619 * - BFQ dynamically updates the budget of every queue so as
4620 * to accommodate the expected backlog of the queue;
4621 * - if a queue gets all its requests dispatched as injected
4622 * service, then the queue is removed from the active list
4623 * (and re-added only if it gets new requests, but then it
4624 * is assigned again enough budget for its new backlog).
4625 */
4626 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
4627 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4628 (in_serv_always_inject || bfqq->wr_coeff > 1) &&
4629 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4630 bfq_bfqq_budget_left(bfqq)) {
4631 /*
4632 * Allow for only one large in-flight request
4633 * on non-rotational devices, for the
4634 * following reason. On non-rotationl drives,
4635 * large requests take much longer than
4636 * smaller requests to be served. In addition,
4637 * the drive prefers to serve large requests
4638 * w.r.t. to small ones, if it can choose. So,
4639 * having more than one large requests queued
4640 * in the drive may easily make the next first
4641 * request of the in-service queue wait for so
4642 * long to break bfqq's service guarantees. On
4643 * the bright side, large requests let the
4644 * drive reach a very high throughput, even if
4645 * there is only one in-flight large request
4646 * at a time.
4647 */
4648 if (blk_queue_nonrot(bfqd->queue) &&
4649 blk_rq_sectors(bfqq->next_rq) >=
4650 BFQQ_SECT_THR_NONROT)
4651 limit = min_t(unsigned int, 1, limit);
4652 else
4653 limit = in_serv_bfqq->inject_limit;
4654
4655 if (bfqd->rq_in_driver < limit) {
4656 bfqd->rqs_injected = true;
4657 return bfqq;
4658 }
4659 }
4660
4661 return NULL;
4662}
4663
4664/*
4665 * Select a queue for service. If we have a current queue in service,
4666 * check whether to continue servicing it, or retrieve and set a new one.
4667 */
4668static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
4669{
4670 struct bfq_queue *bfqq;
4671 struct request *next_rq;
4672 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
4673
4674 bfqq = bfqd->in_service_queue;
4675 if (!bfqq)
4676 goto new_queue;
4677
4678 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
4679
4680 /*
4681 * Do not expire bfqq for budget timeout if bfqq may be about
4682 * to enjoy device idling. The reason why, in this case, we
4683 * prevent bfqq from expiring is the same as in the comments
4684 * on the case where bfq_bfqq_must_idle() returns true, in
4685 * bfq_completed_request().
4686 */
4687 if (bfq_may_expire_for_budg_timeout(bfqq) &&
4688 !bfq_bfqq_must_idle(bfqq))
4689 goto expire;
4690
4691check_queue:
4692 /*
4693 * This loop is rarely executed more than once. Even when it
4694 * happens, it is much more convenient to re-execute this loop
4695 * than to return NULL and trigger a new dispatch to get a
4696 * request served.
4697 */
4698 next_rq = bfqq->next_rq;
4699 /*
4700 * If bfqq has requests queued and it has enough budget left to
4701 * serve them, keep the queue, otherwise expire it.
4702 */
4703 if (next_rq) {
4704 if (bfq_serv_to_charge(next_rq, bfqq) >
4705 bfq_bfqq_budget_left(bfqq)) {
4706 /*
4707 * Expire the queue for budget exhaustion,
4708 * which makes sure that the next budget is
4709 * enough to serve the next request, even if
4710 * it comes from the fifo expired path.
4711 */
4712 reason = BFQQE_BUDGET_EXHAUSTED;
4713 goto expire;
4714 } else {
4715 /*
4716 * The idle timer may be pending because we may
4717 * not disable disk idling even when a new request
4718 * arrives.
4719 */
4720 if (bfq_bfqq_wait_request(bfqq)) {
4721 /*
4722 * If we get here: 1) at least a new request
4723 * has arrived but we have not disabled the
4724 * timer because the request was too small,
4725 * 2) then the block layer has unplugged
4726 * the device, causing the dispatch to be
4727 * invoked.
4728 *
4729 * Since the device is unplugged, now the
4730 * requests are probably large enough to
4731 * provide a reasonable throughput.
4732 * So we disable idling.
4733 */
4734 bfq_clear_bfqq_wait_request(bfqq);
4735 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4736 }
4737 goto keep_queue;
4738 }
4739 }
4740
4741 /*
4742 * No requests pending. However, if the in-service queue is idling
4743 * for a new request, or has requests waiting for a completion and
4744 * may idle after their completion, then keep it anyway.
4745 *
4746 * Yet, inject service from other queues if it boosts
4747 * throughput and is possible.
4748 */
4749 if (bfq_bfqq_wait_request(bfqq) ||
4750 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
4751 struct bfq_queue *async_bfqq =
4752 bfqq->bic && bfqq->bic->bfqq[0] &&
4753 bfq_bfqq_busy(bfqq->bic->bfqq[0]) &&
4754 bfqq->bic->bfqq[0]->next_rq ?
4755 bfqq->bic->bfqq[0] : NULL;
4756 struct bfq_queue *blocked_bfqq =
4757 !hlist_empty(&bfqq->woken_list) ?
4758 container_of(bfqq->woken_list.first,
4759 struct bfq_queue,
4760 woken_list_node)
4761 : NULL;
4762
4763 /*
4764 * The next four mutually-exclusive ifs decide
4765 * whether to try injection, and choose the queue to
4766 * pick an I/O request from.
4767 *
4768 * The first if checks whether the process associated
4769 * with bfqq has also async I/O pending. If so, it
4770 * injects such I/O unconditionally. Injecting async
4771 * I/O from the same process can cause no harm to the
4772 * process. On the contrary, it can only increase
4773 * bandwidth and reduce latency for the process.
4774 *
4775 * The second if checks whether there happens to be a
4776 * non-empty waker queue for bfqq, i.e., a queue whose
4777 * I/O needs to be completed for bfqq to receive new
4778 * I/O. This happens, e.g., if bfqq is associated with
4779 * a process that does some sync. A sync generates
4780 * extra blocking I/O, which must be completed before
4781 * the process associated with bfqq can go on with its
4782 * I/O. If the I/O of the waker queue is not served,
4783 * then bfqq remains empty, and no I/O is dispatched,
4784 * until the idle timeout fires for bfqq. This is
4785 * likely to result in lower bandwidth and higher
4786 * latencies for bfqq, and in a severe loss of total
4787 * throughput. The best action to take is therefore to
4788 * serve the waker queue as soon as possible. So do it
4789 * (without relying on the third alternative below for
4790 * eventually serving waker_bfqq's I/O; see the last
4791 * paragraph for further details). This systematic
4792 * injection of I/O from the waker queue does not
4793 * cause any delay to bfqq's I/O. On the contrary,
4794 * next bfqq's I/O is brought forward dramatically,
4795 * for it is not blocked for milliseconds.
4796 *
4797 * The third if checks whether there is a queue woken
4798 * by bfqq, and currently with pending I/O. Such a
4799 * woken queue does not steal bandwidth from bfqq,
4800 * because it remains soon without I/O if bfqq is not
4801 * served. So there is virtually no risk of loss of
4802 * bandwidth for bfqq if this woken queue has I/O
4803 * dispatched while bfqq is waiting for new I/O.
4804 *
4805 * The fourth if checks whether bfqq is a queue for
4806 * which it is better to avoid injection. It is so if
4807 * bfqq delivers more throughput when served without
4808 * any further I/O from other queues in the middle, or
4809 * if the service times of bfqq's I/O requests both
4810 * count more than overall throughput, and may be
4811 * easily increased by injection (this happens if bfqq
4812 * has a short think time). If none of these
4813 * conditions holds, then a candidate queue for
4814 * injection is looked for through
4815 * bfq_choose_bfqq_for_injection(). Note that the
4816 * latter may return NULL (for example if the inject
4817 * limit for bfqq is currently 0).
4818 *
4819 * NOTE: motivation for the second alternative
4820 *
4821 * Thanks to the way the inject limit is updated in
4822 * bfq_update_has_short_ttime(), it is rather likely
4823 * that, if I/O is being plugged for bfqq and the
4824 * waker queue has pending I/O requests that are
4825 * blocking bfqq's I/O, then the fourth alternative
4826 * above lets the waker queue get served before the
4827 * I/O-plugging timeout fires. So one may deem the
4828 * second alternative superfluous. It is not, because
4829 * the fourth alternative may be way less effective in
4830 * case of a synchronization. For two main
4831 * reasons. First, throughput may be low because the
4832 * inject limit may be too low to guarantee the same
4833 * amount of injected I/O, from the waker queue or
4834 * other queues, that the second alternative
4835 * guarantees (the second alternative unconditionally
4836 * injects a pending I/O request of the waker queue
4837 * for each bfq_dispatch_request()). Second, with the
4838 * fourth alternative, the duration of the plugging,
4839 * i.e., the time before bfqq finally receives new I/O,
4840 * may not be minimized, because the waker queue may
4841 * happen to be served only after other queues.
4842 */
4843 if (async_bfqq &&
4844 icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
4845 bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
4846 bfq_bfqq_budget_left(async_bfqq))
4847 bfqq = bfqq->bic->bfqq[0];
4848 else if (bfqq->waker_bfqq &&
4849 bfq_bfqq_busy(bfqq->waker_bfqq) &&
4850 bfqq->waker_bfqq->next_rq &&
4851 bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
4852 bfqq->waker_bfqq) <=
4853 bfq_bfqq_budget_left(bfqq->waker_bfqq)
4854 )
4855 bfqq = bfqq->waker_bfqq;
4856 else if (blocked_bfqq &&
4857 bfq_bfqq_busy(blocked_bfqq) &&
4858 blocked_bfqq->next_rq &&
4859 bfq_serv_to_charge(blocked_bfqq->next_rq,
4860 blocked_bfqq) <=
4861 bfq_bfqq_budget_left(blocked_bfqq)
4862 )
4863 bfqq = blocked_bfqq;
4864 else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
4865 (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
4866 !bfq_bfqq_has_short_ttime(bfqq)))
4867 bfqq = bfq_choose_bfqq_for_injection(bfqd);
4868 else
4869 bfqq = NULL;
4870
4871 goto keep_queue;
4872 }
4873
4874 reason = BFQQE_NO_MORE_REQUESTS;
4875expire:
4876 bfq_bfqq_expire(bfqd, bfqq, false, reason);
4877new_queue:
4878 bfqq = bfq_set_in_service_queue(bfqd);
4879 if (bfqq) {
4880 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
4881 goto check_queue;
4882 }
4883keep_queue:
4884 if (bfqq)
4885 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
4886 else
4887 bfq_log(bfqd, "select_queue: no queue returned");
4888
4889 return bfqq;
4890}
4891
4892static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4893{
4894 struct bfq_entity *entity = &bfqq->entity;
4895
4896 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
4897 bfq_log_bfqq(bfqd, bfqq,
4898 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
4899 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
4900 jiffies_to_msecs(bfqq->wr_cur_max_time),
4901 bfqq->wr_coeff,
4902 bfqq->entity.weight, bfqq->entity.orig_weight);
4903
4904 if (entity->prio_changed)
4905 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
4906
4907 /*
4908 * If the queue was activated in a burst, or too much
4909 * time has elapsed from the beginning of this
4910 * weight-raising period, then end weight raising.
4911 */
4912 if (bfq_bfqq_in_large_burst(bfqq))
4913 bfq_bfqq_end_wr(bfqq);
4914 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
4915 bfqq->wr_cur_max_time)) {
4916 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
4917 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
4918 bfq_wr_duration(bfqd))) {
4919 /*
4920 * Either in interactive weight
4921 * raising, or in soft_rt weight
4922 * raising with the
4923 * interactive-weight-raising period
4924 * elapsed (so no switch back to
4925 * interactive weight raising).
4926 */
4927 bfq_bfqq_end_wr(bfqq);
4928 } else { /*
4929 * soft_rt finishing while still in
4930 * interactive period, switch back to
4931 * interactive weight raising
4932 */
4933 switch_back_to_interactive_wr(bfqq, bfqd);
4934 bfqq->entity.prio_changed = 1;
4935 }
4936 }
4937 if (bfqq->wr_coeff > 1 &&
4938 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
4939 bfqq->service_from_wr > max_service_from_wr) {
4940 /* see comments on max_service_from_wr */
4941 bfq_bfqq_end_wr(bfqq);
4942 }
4943 }
4944 /*
4945 * To improve latency (for this or other queues), immediately
4946 * update weight both if it must be raised and if it must be
4947 * lowered. Since, entity may be on some active tree here, and
4948 * might have a pending change of its ioprio class, invoke
4949 * next function with the last parameter unset (see the
4950 * comments on the function).
4951 */
4952 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
4953 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
4954 entity, false);
4955}
4956
4957/*
4958 * Dispatch next request from bfqq.
4959 */
4960static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
4961 struct bfq_queue *bfqq)
4962{
4963 struct request *rq = bfqq->next_rq;
4964 unsigned long service_to_charge;
4965
4966 service_to_charge = bfq_serv_to_charge(rq, bfqq);
4967
4968 bfq_bfqq_served(bfqq, service_to_charge);
4969
4970 if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
4971 bfqd->wait_dispatch = false;
4972 bfqd->waited_rq = rq;
4973 }
4974
4975 bfq_dispatch_remove(bfqd->queue, rq);
4976
4977 if (bfqq != bfqd->in_service_queue)
4978 goto return_rq;
4979
4980 /*
4981 * If weight raising has to terminate for bfqq, then next
4982 * function causes an immediate update of bfqq's weight,
4983 * without waiting for next activation. As a consequence, on
4984 * expiration, bfqq will be timestamped as if has never been
4985 * weight-raised during this service slot, even if it has
4986 * received part or even most of the service as a
4987 * weight-raised queue. This inflates bfqq's timestamps, which
4988 * is beneficial, as bfqq is then more willing to leave the
4989 * device immediately to possible other weight-raised queues.
4990 */
4991 bfq_update_wr_data(bfqd, bfqq);
4992
4993 /*
4994 * Expire bfqq, pretending that its budget expired, if bfqq
4995 * belongs to CLASS_IDLE and other queues are waiting for
4996 * service.
4997 */
4998 if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
4999 goto return_rq;
5000
5001 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
5002
5003return_rq:
5004 return rq;
5005}
5006
5007static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
5008{
5009 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5010
5011 /*
5012 * Avoiding lock: a race on bfqd->queued should cause at
5013 * most a call to dispatch for nothing
5014 */
5015 return !list_empty_careful(&bfqd->dispatch) ||
5016 READ_ONCE(bfqd->queued);
5017}
5018
5019static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5020{
5021 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5022 struct request *rq = NULL;
5023 struct bfq_queue *bfqq = NULL;
5024
5025 if (!list_empty(&bfqd->dispatch)) {
5026 rq = list_first_entry(&bfqd->dispatch, struct request,
5027 queuelist);
5028 list_del_init(&rq->queuelist);
5029
5030 bfqq = RQ_BFQQ(rq);
5031
5032 if (bfqq) {
5033 /*
5034 * Increment counters here, because this
5035 * dispatch does not follow the standard
5036 * dispatch flow (where counters are
5037 * incremented)
5038 */
5039 bfqq->dispatched++;
5040
5041 goto inc_in_driver_start_rq;
5042 }
5043
5044 /*
5045 * We exploit the bfq_finish_requeue_request hook to
5046 * decrement rq_in_driver, but
5047 * bfq_finish_requeue_request will not be invoked on
5048 * this request. So, to avoid unbalance, just start
5049 * this request, without incrementing rq_in_driver. As
5050 * a negative consequence, rq_in_driver is deceptively
5051 * lower than it should be while this request is in
5052 * service. This may cause bfq_schedule_dispatch to be
5053 * invoked uselessly.
5054 *
5055 * As for implementing an exact solution, the
5056 * bfq_finish_requeue_request hook, if defined, is
5057 * probably invoked also on this request. So, by
5058 * exploiting this hook, we could 1) increment
5059 * rq_in_driver here, and 2) decrement it in
5060 * bfq_finish_requeue_request. Such a solution would
5061 * let the value of the counter be always accurate,
5062 * but it would entail using an extra interface
5063 * function. This cost seems higher than the benefit,
5064 * being the frequency of non-elevator-private
5065 * requests very low.
5066 */
5067 goto start_rq;
5068 }
5069
5070 bfq_log(bfqd, "dispatch requests: %d busy queues",
5071 bfq_tot_busy_queues(bfqd));
5072
5073 if (bfq_tot_busy_queues(bfqd) == 0)
5074 goto exit;
5075
5076 /*
5077 * Force device to serve one request at a time if
5078 * strict_guarantees is true. Forcing this service scheme is
5079 * currently the ONLY way to guarantee that the request
5080 * service order enforced by the scheduler is respected by a
5081 * queueing device. Otherwise the device is free even to make
5082 * some unlucky request wait for as long as the device
5083 * wishes.
5084 *
5085 * Of course, serving one request at a time may cause loss of
5086 * throughput.
5087 */
5088 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
5089 goto exit;
5090
5091 bfqq = bfq_select_queue(bfqd);
5092 if (!bfqq)
5093 goto exit;
5094
5095 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
5096
5097 if (rq) {
5098inc_in_driver_start_rq:
5099 bfqd->rq_in_driver++;
5100start_rq:
5101 rq->rq_flags |= RQF_STARTED;
5102 }
5103exit:
5104 return rq;
5105}
5106
5107#ifdef CONFIG_BFQ_CGROUP_DEBUG
5108static void bfq_update_dispatch_stats(struct request_queue *q,
5109 struct request *rq,
5110 struct bfq_queue *in_serv_queue,
5111 bool idle_timer_disabled)
5112{
5113 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
5114
5115 if (!idle_timer_disabled && !bfqq)
5116 return;
5117
5118 /*
5119 * rq and bfqq are guaranteed to exist until this function
5120 * ends, for the following reasons. First, rq can be
5121 * dispatched to the device, and then can be completed and
5122 * freed, only after this function ends. Second, rq cannot be
5123 * merged (and thus freed because of a merge) any longer,
5124 * because it has already started. Thus rq cannot be freed
5125 * before this function ends, and, since rq has a reference to
5126 * bfqq, the same guarantee holds for bfqq too.
5127 *
5128 * In addition, the following queue lock guarantees that
5129 * bfqq_group(bfqq) exists as well.
5130 */
5131 spin_lock_irq(&q->queue_lock);
5132 if (idle_timer_disabled)
5133 /*
5134 * Since the idle timer has been disabled,
5135 * in_serv_queue contained some request when
5136 * __bfq_dispatch_request was invoked above, which
5137 * implies that rq was picked exactly from
5138 * in_serv_queue. Thus in_serv_queue == bfqq, and is
5139 * therefore guaranteed to exist because of the above
5140 * arguments.
5141 */
5142 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
5143 if (bfqq) {
5144 struct bfq_group *bfqg = bfqq_group(bfqq);
5145
5146 bfqg_stats_update_avg_queue_size(bfqg);
5147 bfqg_stats_set_start_empty_time(bfqg);
5148 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
5149 }
5150 spin_unlock_irq(&q->queue_lock);
5151}
5152#else
5153static inline void bfq_update_dispatch_stats(struct request_queue *q,
5154 struct request *rq,
5155 struct bfq_queue *in_serv_queue,
5156 bool idle_timer_disabled) {}
5157#endif /* CONFIG_BFQ_CGROUP_DEBUG */
5158
5159static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5160{
5161 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5162 struct request *rq;
5163 struct bfq_queue *in_serv_queue;
5164 bool waiting_rq, idle_timer_disabled = false;
5165
5166 spin_lock_irq(&bfqd->lock);
5167
5168 in_serv_queue = bfqd->in_service_queue;
5169 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
5170
5171 rq = __bfq_dispatch_request(hctx);
5172 if (in_serv_queue == bfqd->in_service_queue) {
5173 idle_timer_disabled =
5174 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
5175 }
5176
5177 spin_unlock_irq(&bfqd->lock);
5178 bfq_update_dispatch_stats(hctx->queue, rq,
5179 idle_timer_disabled ? in_serv_queue : NULL,
5180 idle_timer_disabled);
5181
5182 return rq;
5183}
5184
5185/*
5186 * Task holds one reference to the queue, dropped when task exits. Each rq
5187 * in-flight on this queue also holds a reference, dropped when rq is freed.
5188 *
5189 * Scheduler lock must be held here. Recall not to use bfqq after calling
5190 * this function on it.
5191 */
5192void bfq_put_queue(struct bfq_queue *bfqq)
5193{
5194 struct bfq_queue *item;
5195 struct hlist_node *n;
5196 struct bfq_group *bfqg = bfqq_group(bfqq);
5197
5198 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d", bfqq, bfqq->ref);
5199
5200 bfqq->ref--;
5201 if (bfqq->ref)
5202 return;
5203
5204 if (!hlist_unhashed(&bfqq->burst_list_node)) {
5205 hlist_del_init(&bfqq->burst_list_node);
5206 /*
5207 * Decrement also burst size after the removal, if the
5208 * process associated with bfqq is exiting, and thus
5209 * does not contribute to the burst any longer. This
5210 * decrement helps filter out false positives of large
5211 * bursts, when some short-lived process (often due to
5212 * the execution of commands by some service) happens
5213 * to start and exit while a complex application is
5214 * starting, and thus spawning several processes that
5215 * do I/O (and that *must not* be treated as a large
5216 * burst, see comments on bfq_handle_burst).
5217 *
5218 * In particular, the decrement is performed only if:
5219 * 1) bfqq is not a merged queue, because, if it is,
5220 * then this free of bfqq is not triggered by the exit
5221 * of the process bfqq is associated with, but exactly
5222 * by the fact that bfqq has just been merged.
5223 * 2) burst_size is greater than 0, to handle
5224 * unbalanced decrements. Unbalanced decrements may
5225 * happen in te following case: bfqq is inserted into
5226 * the current burst list--without incrementing
5227 * bust_size--because of a split, but the current
5228 * burst list is not the burst list bfqq belonged to
5229 * (see comments on the case of a split in
5230 * bfq_set_request).
5231 */
5232 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
5233 bfqq->bfqd->burst_size--;
5234 }
5235
5236 /*
5237 * bfqq does not exist any longer, so it cannot be woken by
5238 * any other queue, and cannot wake any other queue. Then bfqq
5239 * must be removed from the woken list of its possible waker
5240 * queue, and all queues in the woken list of bfqq must stop
5241 * having a waker queue. Strictly speaking, these updates
5242 * should be performed when bfqq remains with no I/O source
5243 * attached to it, which happens before bfqq gets freed. In
5244 * particular, this happens when the last process associated
5245 * with bfqq exits or gets associated with a different
5246 * queue. However, both events lead to bfqq being freed soon,
5247 * and dangling references would come out only after bfqq gets
5248 * freed. So these updates are done here, as a simple and safe
5249 * way to handle all cases.
5250 */
5251 /* remove bfqq from woken list */
5252 if (!hlist_unhashed(&bfqq->woken_list_node))
5253 hlist_del_init(&bfqq->woken_list_node);
5254
5255 /* reset waker for all queues in woken list */
5256 hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
5257 woken_list_node) {
5258 item->waker_bfqq = NULL;
5259 hlist_del_init(&item->woken_list_node);
5260 }
5261
5262 if (bfqq->bfqd->last_completed_rq_bfqq == bfqq)
5263 bfqq->bfqd->last_completed_rq_bfqq = NULL;
5264
5265 kmem_cache_free(bfq_pool, bfqq);
5266 bfqg_and_blkg_put(bfqg);
5267}
5268
5269static void bfq_put_stable_ref(struct bfq_queue *bfqq)
5270{
5271 bfqq->stable_ref--;
5272 bfq_put_queue(bfqq);
5273}
5274
5275void bfq_put_cooperator(struct bfq_queue *bfqq)
5276{
5277 struct bfq_queue *__bfqq, *next;
5278
5279 /*
5280 * If this queue was scheduled to merge with another queue, be
5281 * sure to drop the reference taken on that queue (and others in
5282 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
5283 */
5284 __bfqq = bfqq->new_bfqq;
5285 while (__bfqq) {
5286 if (__bfqq == bfqq)
5287 break;
5288 next = __bfqq->new_bfqq;
5289 bfq_put_queue(__bfqq);
5290 __bfqq = next;
5291 }
5292}
5293
5294static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
5295{
5296 if (bfqq == bfqd->in_service_queue) {
5297 __bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT);
5298 bfq_schedule_dispatch(bfqd);
5299 }
5300
5301 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
5302
5303 bfq_put_cooperator(bfqq);
5304
5305 bfq_release_process_ref(bfqd, bfqq);
5306}
5307
5308static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
5309{
5310 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
5311 struct bfq_data *bfqd;
5312
5313 if (bfqq)
5314 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
5315
5316 if (bfqq && bfqd) {
5317 unsigned long flags;
5318
5319 spin_lock_irqsave(&bfqd->lock, flags);
5320 bic_set_bfqq(bic, NULL, is_sync);
5321 bfq_exit_bfqq(bfqd, bfqq);
5322 spin_unlock_irqrestore(&bfqd->lock, flags);
5323 }
5324}
5325
5326static void bfq_exit_icq(struct io_cq *icq)
5327{
5328 struct bfq_io_cq *bic = icq_to_bic(icq);
5329
5330 if (bic->stable_merge_bfqq) {
5331 struct bfq_data *bfqd = bic->stable_merge_bfqq->bfqd;
5332
5333 /*
5334 * bfqd is NULL if scheduler already exited, and in
5335 * that case this is the last time bfqq is accessed.
5336 */
5337 if (bfqd) {
5338 unsigned long flags;
5339
5340 spin_lock_irqsave(&bfqd->lock, flags);
5341 bfq_put_stable_ref(bic->stable_merge_bfqq);
5342 spin_unlock_irqrestore(&bfqd->lock, flags);
5343 } else {
5344 bfq_put_stable_ref(bic->stable_merge_bfqq);
5345 }
5346 }
5347
5348 bfq_exit_icq_bfqq(bic, true);
5349 bfq_exit_icq_bfqq(bic, false);
5350}
5351
5352/*
5353 * Update the entity prio values; note that the new values will not
5354 * be used until the next (re)activation.
5355 */
5356static void
5357bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
5358{
5359 struct task_struct *tsk = current;
5360 int ioprio_class;
5361 struct bfq_data *bfqd = bfqq->bfqd;
5362
5363 if (!bfqd)
5364 return;
5365
5366 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5367 switch (ioprio_class) {
5368 default:
5369 pr_err("bdi %s: bfq: bad prio class %d\n",
5370 bdi_dev_name(bfqq->bfqd->queue->disk->bdi),
5371 ioprio_class);
5372 fallthrough;
5373 case IOPRIO_CLASS_NONE:
5374 /*
5375 * No prio set, inherit CPU scheduling settings.
5376 */
5377 bfqq->new_ioprio = task_nice_ioprio(tsk);
5378 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
5379 break;
5380 case IOPRIO_CLASS_RT:
5381 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5382 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
5383 break;
5384 case IOPRIO_CLASS_BE:
5385 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5386 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
5387 break;
5388 case IOPRIO_CLASS_IDLE:
5389 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
5390 bfqq->new_ioprio = 7;
5391 break;
5392 }
5393
5394 if (bfqq->new_ioprio >= IOPRIO_NR_LEVELS) {
5395 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
5396 bfqq->new_ioprio);
5397 bfqq->new_ioprio = IOPRIO_NR_LEVELS - 1;
5398 }
5399
5400 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
5401 bfq_log_bfqq(bfqd, bfqq, "new_ioprio %d new_weight %d",
5402 bfqq->new_ioprio, bfqq->entity.new_weight);
5403 bfqq->entity.prio_changed = 1;
5404}
5405
5406static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5407 struct bio *bio, bool is_sync,
5408 struct bfq_io_cq *bic,
5409 bool respawn);
5410
5411static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
5412{
5413 struct bfq_data *bfqd = bic_to_bfqd(bic);
5414 struct bfq_queue *bfqq;
5415 int ioprio = bic->icq.ioc->ioprio;
5416
5417 /*
5418 * This condition may trigger on a newly created bic, be sure to
5419 * drop the lock before returning.
5420 */
5421 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
5422 return;
5423
5424 bic->ioprio = ioprio;
5425
5426 bfqq = bic_to_bfqq(bic, false);
5427 if (bfqq) {
5428 struct bfq_queue *old_bfqq = bfqq;
5429
5430 bfqq = bfq_get_queue(bfqd, bio, false, bic, true);
5431 bic_set_bfqq(bic, bfqq, false);
5432 bfq_release_process_ref(bfqd, old_bfqq);
5433 }
5434
5435 bfqq = bic_to_bfqq(bic, true);
5436 if (bfqq)
5437 bfq_set_next_ioprio_data(bfqq, bic);
5438}
5439
5440static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5441 struct bfq_io_cq *bic, pid_t pid, int is_sync)
5442{
5443 u64 now_ns = ktime_get_ns();
5444
5445 RB_CLEAR_NODE(&bfqq->entity.rb_node);
5446 INIT_LIST_HEAD(&bfqq->fifo);
5447 INIT_HLIST_NODE(&bfqq->burst_list_node);
5448 INIT_HLIST_NODE(&bfqq->woken_list_node);
5449 INIT_HLIST_HEAD(&bfqq->woken_list);
5450
5451 bfqq->ref = 0;
5452 bfqq->bfqd = bfqd;
5453
5454 if (bic)
5455 bfq_set_next_ioprio_data(bfqq, bic);
5456
5457 if (is_sync) {
5458 /*
5459 * No need to mark as has_short_ttime if in
5460 * idle_class, because no device idling is performed
5461 * for queues in idle class
5462 */
5463 if (!bfq_class_idle(bfqq))
5464 /* tentatively mark as has_short_ttime */
5465 bfq_mark_bfqq_has_short_ttime(bfqq);
5466 bfq_mark_bfqq_sync(bfqq);
5467 bfq_mark_bfqq_just_created(bfqq);
5468 } else
5469 bfq_clear_bfqq_sync(bfqq);
5470
5471 /* set end request to minus infinity from now */
5472 bfqq->ttime.last_end_request = now_ns + 1;
5473
5474 bfqq->creation_time = jiffies;
5475
5476 bfqq->io_start_time = now_ns;
5477
5478 bfq_mark_bfqq_IO_bound(bfqq);
5479
5480 bfqq->pid = pid;
5481
5482 /* Tentative initial value to trade off between thr and lat */
5483 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
5484 bfqq->budget_timeout = bfq_smallest_from_now();
5485
5486 bfqq->wr_coeff = 1;
5487 bfqq->last_wr_start_finish = jiffies;
5488 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
5489 bfqq->split_time = bfq_smallest_from_now();
5490
5491 /*
5492 * To not forget the possibly high bandwidth consumed by a
5493 * process/queue in the recent past,
5494 * bfq_bfqq_softrt_next_start() returns a value at least equal
5495 * to the current value of bfqq->soft_rt_next_start (see
5496 * comments on bfq_bfqq_softrt_next_start). Set
5497 * soft_rt_next_start to now, to mean that bfqq has consumed
5498 * no bandwidth so far.
5499 */
5500 bfqq->soft_rt_next_start = jiffies;
5501
5502 /* first request is almost certainly seeky */
5503 bfqq->seek_history = 1;
5504}
5505
5506static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
5507 struct bfq_group *bfqg,
5508 int ioprio_class, int ioprio)
5509{
5510 switch (ioprio_class) {
5511 case IOPRIO_CLASS_RT:
5512 return &bfqg->async_bfqq[0][ioprio];
5513 case IOPRIO_CLASS_NONE:
5514 ioprio = IOPRIO_BE_NORM;
5515 fallthrough;
5516 case IOPRIO_CLASS_BE:
5517 return &bfqg->async_bfqq[1][ioprio];
5518 case IOPRIO_CLASS_IDLE:
5519 return &bfqg->async_idle_bfqq;
5520 default:
5521 return NULL;
5522 }
5523}
5524
5525static struct bfq_queue *
5526bfq_do_early_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5527 struct bfq_io_cq *bic,
5528 struct bfq_queue *last_bfqq_created)
5529{
5530 struct bfq_queue *new_bfqq =
5531 bfq_setup_merge(bfqq, last_bfqq_created);
5532
5533 if (!new_bfqq)
5534 return bfqq;
5535
5536 if (new_bfqq->bic)
5537 new_bfqq->bic->stably_merged = true;
5538 bic->stably_merged = true;
5539
5540 /*
5541 * Reusing merge functions. This implies that
5542 * bfqq->bic must be set too, for
5543 * bfq_merge_bfqqs to correctly save bfqq's
5544 * state before killing it.
5545 */
5546 bfqq->bic = bic;
5547 bfq_merge_bfqqs(bfqd, bic, bfqq, new_bfqq);
5548
5549 return new_bfqq;
5550}
5551
5552/*
5553 * Many throughput-sensitive workloads are made of several parallel
5554 * I/O flows, with all flows generated by the same application, or
5555 * more generically by the same task (e.g., system boot). The most
5556 * counterproductive action with these workloads is plugging I/O
5557 * dispatch when one of the bfq_queues associated with these flows
5558 * remains temporarily empty.
5559 *
5560 * To avoid this plugging, BFQ has been using a burst-handling
5561 * mechanism for years now. This mechanism has proven effective for
5562 * throughput, and not detrimental for service guarantees. The
5563 * following function pushes this mechanism a little bit further,
5564 * basing on the following two facts.
5565 *
5566 * First, all the I/O flows of a the same application or task
5567 * contribute to the execution/completion of that common application
5568 * or task. So the performance figures that matter are total
5569 * throughput of the flows and task-wide I/O latency. In particular,
5570 * these flows do not need to be protected from each other, in terms
5571 * of individual bandwidth or latency.
5572 *
5573 * Second, the above fact holds regardless of the number of flows.
5574 *
5575 * Putting these two facts together, this commits merges stably the
5576 * bfq_queues associated with these I/O flows, i.e., with the
5577 * processes that generate these IO/ flows, regardless of how many the
5578 * involved processes are.
5579 *
5580 * To decide whether a set of bfq_queues is actually associated with
5581 * the I/O flows of a common application or task, and to merge these
5582 * queues stably, this function operates as follows: given a bfq_queue,
5583 * say Q2, currently being created, and the last bfq_queue, say Q1,
5584 * created before Q2, Q2 is merged stably with Q1 if
5585 * - very little time has elapsed since when Q1 was created
5586 * - Q2 has the same ioprio as Q1
5587 * - Q2 belongs to the same group as Q1
5588 *
5589 * Merging bfq_queues also reduces scheduling overhead. A fio test
5590 * with ten random readers on /dev/nullb shows a throughput boost of
5591 * 40%, with a quadcore. Since BFQ's execution time amounts to ~50% of
5592 * the total per-request processing time, the above throughput boost
5593 * implies that BFQ's overhead is reduced by more than 50%.
5594 *
5595 * This new mechanism most certainly obsoletes the current
5596 * burst-handling heuristics. We keep those heuristics for the moment.
5597 */
5598static struct bfq_queue *bfq_do_or_sched_stable_merge(struct bfq_data *bfqd,
5599 struct bfq_queue *bfqq,
5600 struct bfq_io_cq *bic)
5601{
5602 struct bfq_queue **source_bfqq = bfqq->entity.parent ?
5603 &bfqq->entity.parent->last_bfqq_created :
5604 &bfqd->last_bfqq_created;
5605
5606 struct bfq_queue *last_bfqq_created = *source_bfqq;
5607
5608 /*
5609 * If last_bfqq_created has not been set yet, then init it. If
5610 * it has been set already, but too long ago, then move it
5611 * forward to bfqq. Finally, move also if bfqq belongs to a
5612 * different group than last_bfqq_created, or if bfqq has a
5613 * different ioprio or ioprio_class. If none of these
5614 * conditions holds true, then try an early stable merge or
5615 * schedule a delayed stable merge.
5616 *
5617 * A delayed merge is scheduled (instead of performing an
5618 * early merge), in case bfqq might soon prove to be more
5619 * throughput-beneficial if not merged. Currently this is
5620 * possible only if bfqd is rotational with no queueing. For
5621 * such a drive, not merging bfqq is better for throughput if
5622 * bfqq happens to contain sequential I/O. So, we wait a
5623 * little bit for enough I/O to flow through bfqq. After that,
5624 * if such an I/O is sequential, then the merge is
5625 * canceled. Otherwise the merge is finally performed.
5626 */
5627 if (!last_bfqq_created ||
5628 time_before(last_bfqq_created->creation_time +
5629 msecs_to_jiffies(bfq_activation_stable_merging),
5630 bfqq->creation_time) ||
5631 bfqq->entity.parent != last_bfqq_created->entity.parent ||
5632 bfqq->ioprio != last_bfqq_created->ioprio ||
5633 bfqq->ioprio_class != last_bfqq_created->ioprio_class)
5634 *source_bfqq = bfqq;
5635 else if (time_after_eq(last_bfqq_created->creation_time +
5636 bfqd->bfq_burst_interval,
5637 bfqq->creation_time)) {
5638 if (likely(bfqd->nonrot_with_queueing))
5639 /*
5640 * With this type of drive, leaving
5641 * bfqq alone may provide no
5642 * throughput benefits compared with
5643 * merging bfqq. So merge bfqq now.
5644 */
5645 bfqq = bfq_do_early_stable_merge(bfqd, bfqq,
5646 bic,
5647 last_bfqq_created);
5648 else { /* schedule tentative stable merge */
5649 /*
5650 * get reference on last_bfqq_created,
5651 * to prevent it from being freed,
5652 * until we decide whether to merge
5653 */
5654 last_bfqq_created->ref++;
5655 /*
5656 * need to keep track of stable refs, to
5657 * compute process refs correctly
5658 */
5659 last_bfqq_created->stable_ref++;
5660 /*
5661 * Record the bfqq to merge to.
5662 */
5663 bic->stable_merge_bfqq = last_bfqq_created;
5664 }
5665 }
5666
5667 return bfqq;
5668}
5669
5670
5671static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5672 struct bio *bio, bool is_sync,
5673 struct bfq_io_cq *bic,
5674 bool respawn)
5675{
5676 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5677 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5678 struct bfq_queue **async_bfqq = NULL;
5679 struct bfq_queue *bfqq;
5680 struct bfq_group *bfqg;
5681
5682 bfqg = bfq_bio_bfqg(bfqd, bio);
5683 if (!is_sync) {
5684 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
5685 ioprio);
5686 bfqq = *async_bfqq;
5687 if (bfqq)
5688 goto out;
5689 }
5690
5691 bfqq = kmem_cache_alloc_node(bfq_pool,
5692 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
5693 bfqd->queue->node);
5694
5695 if (bfqq) {
5696 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
5697 is_sync);
5698 bfq_init_entity(&bfqq->entity, bfqg);
5699 bfq_log_bfqq(bfqd, bfqq, "allocated");
5700 } else {
5701 bfqq = &bfqd->oom_bfqq;
5702 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
5703 goto out;
5704 }
5705
5706 /*
5707 * Pin the queue now that it's allocated, scheduler exit will
5708 * prune it.
5709 */
5710 if (async_bfqq) {
5711 bfqq->ref++; /*
5712 * Extra group reference, w.r.t. sync
5713 * queue. This extra reference is removed
5714 * only if bfqq->bfqg disappears, to
5715 * guarantee that this queue is not freed
5716 * until its group goes away.
5717 */
5718 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
5719 bfqq, bfqq->ref);
5720 *async_bfqq = bfqq;
5721 }
5722
5723out:
5724 bfqq->ref++; /* get a process reference to this queue */
5725
5726 if (bfqq != &bfqd->oom_bfqq && is_sync && !respawn)
5727 bfqq = bfq_do_or_sched_stable_merge(bfqd, bfqq, bic);
5728 return bfqq;
5729}
5730
5731static void bfq_update_io_thinktime(struct bfq_data *bfqd,
5732 struct bfq_queue *bfqq)
5733{
5734 struct bfq_ttime *ttime = &bfqq->ttime;
5735 u64 elapsed;
5736
5737 /*
5738 * We are really interested in how long it takes for the queue to
5739 * become busy when there is no outstanding IO for this queue. So
5740 * ignore cases when the bfq queue has already IO queued.
5741 */
5742 if (bfqq->dispatched || bfq_bfqq_busy(bfqq))
5743 return;
5744 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
5745 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
5746
5747 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
5748 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
5749 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
5750 ttime->ttime_samples);
5751}
5752
5753static void
5754bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5755 struct request *rq)
5756{
5757 bfqq->seek_history <<= 1;
5758 bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
5759
5760 if (bfqq->wr_coeff > 1 &&
5761 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
5762 BFQQ_TOTALLY_SEEKY(bfqq)) {
5763 if (time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
5764 bfq_wr_duration(bfqd))) {
5765 /*
5766 * In soft_rt weight raising with the
5767 * interactive-weight-raising period
5768 * elapsed (so no switch back to
5769 * interactive weight raising).
5770 */
5771 bfq_bfqq_end_wr(bfqq);
5772 } else { /*
5773 * stopping soft_rt weight raising
5774 * while still in interactive period,
5775 * switch back to interactive weight
5776 * raising
5777 */
5778 switch_back_to_interactive_wr(bfqq, bfqd);
5779 bfqq->entity.prio_changed = 1;
5780 }
5781 }
5782}
5783
5784static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
5785 struct bfq_queue *bfqq,
5786 struct bfq_io_cq *bic)
5787{
5788 bool has_short_ttime = true, state_changed;
5789
5790 /*
5791 * No need to update has_short_ttime if bfqq is async or in
5792 * idle io prio class, or if bfq_slice_idle is zero, because
5793 * no device idling is performed for bfqq in this case.
5794 */
5795 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
5796 bfqd->bfq_slice_idle == 0)
5797 return;
5798
5799 /* Idle window just restored, statistics are meaningless. */
5800 if (time_is_after_eq_jiffies(bfqq->split_time +
5801 bfqd->bfq_wr_min_idle_time))
5802 return;
5803
5804 /* Think time is infinite if no process is linked to
5805 * bfqq. Otherwise check average think time to decide whether
5806 * to mark as has_short_ttime. To this goal, compare average
5807 * think time with half the I/O-plugging timeout.
5808 */
5809 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
5810 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
5811 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle>>1))
5812 has_short_ttime = false;
5813
5814 state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
5815
5816 if (has_short_ttime)
5817 bfq_mark_bfqq_has_short_ttime(bfqq);
5818 else
5819 bfq_clear_bfqq_has_short_ttime(bfqq);
5820
5821 /*
5822 * Until the base value for the total service time gets
5823 * finally computed for bfqq, the inject limit does depend on
5824 * the think-time state (short|long). In particular, the limit
5825 * is 0 or 1 if the think time is deemed, respectively, as
5826 * short or long (details in the comments in
5827 * bfq_update_inject_limit()). Accordingly, the next
5828 * instructions reset the inject limit if the think-time state
5829 * has changed and the above base value is still to be
5830 * computed.
5831 *
5832 * However, the reset is performed only if more than 100 ms
5833 * have elapsed since the last update of the inject limit, or
5834 * (inclusive) if the change is from short to long think
5835 * time. The reason for this waiting is as follows.
5836 *
5837 * bfqq may have a long think time because of a
5838 * synchronization with some other queue, i.e., because the
5839 * I/O of some other queue may need to be completed for bfqq
5840 * to receive new I/O. Details in the comments on the choice
5841 * of the queue for injection in bfq_select_queue().
5842 *
5843 * As stressed in those comments, if such a synchronization is
5844 * actually in place, then, without injection on bfqq, the
5845 * blocking I/O cannot happen to served while bfqq is in
5846 * service. As a consequence, if bfqq is granted
5847 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
5848 * is dispatched, until the idle timeout fires. This is likely
5849 * to result in lower bandwidth and higher latencies for bfqq,
5850 * and in a severe loss of total throughput.
5851 *
5852 * On the opposite end, a non-zero inject limit may allow the
5853 * I/O that blocks bfqq to be executed soon, and therefore
5854 * bfqq to receive new I/O soon.
5855 *
5856 * But, if the blocking gets actually eliminated, then the
5857 * next think-time sample for bfqq may be very low. This in
5858 * turn may cause bfqq's think time to be deemed
5859 * short. Without the 100 ms barrier, this new state change
5860 * would cause the body of the next if to be executed
5861 * immediately. But this would set to 0 the inject
5862 * limit. Without injection, the blocking I/O would cause the
5863 * think time of bfqq to become long again, and therefore the
5864 * inject limit to be raised again, and so on. The only effect
5865 * of such a steady oscillation between the two think-time
5866 * states would be to prevent effective injection on bfqq.
5867 *
5868 * In contrast, if the inject limit is not reset during such a
5869 * long time interval as 100 ms, then the number of short
5870 * think time samples can grow significantly before the reset
5871 * is performed. As a consequence, the think time state can
5872 * become stable before the reset. Therefore there will be no
5873 * state change when the 100 ms elapse, and no reset of the
5874 * inject limit. The inject limit remains steadily equal to 1
5875 * both during and after the 100 ms. So injection can be
5876 * performed at all times, and throughput gets boosted.
5877 *
5878 * An inject limit equal to 1 is however in conflict, in
5879 * general, with the fact that the think time of bfqq is
5880 * short, because injection may be likely to delay bfqq's I/O
5881 * (as explained in the comments in
5882 * bfq_update_inject_limit()). But this does not happen in
5883 * this special case, because bfqq's low think time is due to
5884 * an effective handling of a synchronization, through
5885 * injection. In this special case, bfqq's I/O does not get
5886 * delayed by injection; on the contrary, bfqq's I/O is
5887 * brought forward, because it is not blocked for
5888 * milliseconds.
5889 *
5890 * In addition, serving the blocking I/O much sooner, and much
5891 * more frequently than once per I/O-plugging timeout, makes
5892 * it much quicker to detect a waker queue (the concept of
5893 * waker queue is defined in the comments in
5894 * bfq_add_request()). This makes it possible to start sooner
5895 * to boost throughput more effectively, by injecting the I/O
5896 * of the waker queue unconditionally on every
5897 * bfq_dispatch_request().
5898 *
5899 * One last, important benefit of not resetting the inject
5900 * limit before 100 ms is that, during this time interval, the
5901 * base value for the total service time is likely to get
5902 * finally computed for bfqq, freeing the inject limit from
5903 * its relation with the think time.
5904 */
5905 if (state_changed && bfqq->last_serv_time_ns == 0 &&
5906 (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
5907 msecs_to_jiffies(100)) ||
5908 !has_short_ttime))
5909 bfq_reset_inject_limit(bfqd, bfqq);
5910}
5911
5912/*
5913 * Called when a new fs request (rq) is added to bfqq. Check if there's
5914 * something we should do about it.
5915 */
5916static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5917 struct request *rq)
5918{
5919 if (rq->cmd_flags & REQ_META)
5920 bfqq->meta_pending++;
5921
5922 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
5923
5924 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
5925 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
5926 blk_rq_sectors(rq) < 32;
5927 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
5928
5929 /*
5930 * There is just this request queued: if
5931 * - the request is small, and
5932 * - we are idling to boost throughput, and
5933 * - the queue is not to be expired,
5934 * then just exit.
5935 *
5936 * In this way, if the device is being idled to wait
5937 * for a new request from the in-service queue, we
5938 * avoid unplugging the device and committing the
5939 * device to serve just a small request. In contrast
5940 * we wait for the block layer to decide when to
5941 * unplug the device: hopefully, new requests will be
5942 * merged to this one quickly, then the device will be
5943 * unplugged and larger requests will be dispatched.
5944 */
5945 if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
5946 !budget_timeout)
5947 return;
5948
5949 /*
5950 * A large enough request arrived, or idling is being
5951 * performed to preserve service guarantees, or
5952 * finally the queue is to be expired: in all these
5953 * cases disk idling is to be stopped, so clear
5954 * wait_request flag and reset timer.
5955 */
5956 bfq_clear_bfqq_wait_request(bfqq);
5957 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
5958
5959 /*
5960 * The queue is not empty, because a new request just
5961 * arrived. Hence we can safely expire the queue, in
5962 * case of budget timeout, without risking that the
5963 * timestamps of the queue are not updated correctly.
5964 * See [1] for more details.
5965 */
5966 if (budget_timeout)
5967 bfq_bfqq_expire(bfqd, bfqq, false,
5968 BFQQE_BUDGET_TIMEOUT);
5969 }
5970}
5971
5972static void bfqq_request_allocated(struct bfq_queue *bfqq)
5973{
5974 struct bfq_entity *entity = &bfqq->entity;
5975
5976 for_each_entity(entity)
5977 entity->allocated++;
5978}
5979
5980static void bfqq_request_freed(struct bfq_queue *bfqq)
5981{
5982 struct bfq_entity *entity = &bfqq->entity;
5983
5984 for_each_entity(entity)
5985 entity->allocated--;
5986}
5987
5988/* returns true if it causes the idle timer to be disabled */
5989static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
5990{
5991 struct bfq_queue *bfqq = RQ_BFQQ(rq),
5992 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true,
5993 RQ_BIC(rq));
5994 bool waiting, idle_timer_disabled = false;
5995
5996 if (new_bfqq) {
5997 /*
5998 * Release the request's reference to the old bfqq
5999 * and make sure one is taken to the shared queue.
6000 */
6001 bfqq_request_allocated(new_bfqq);
6002 bfqq_request_freed(bfqq);
6003 new_bfqq->ref++;
6004 /*
6005 * If the bic associated with the process
6006 * issuing this request still points to bfqq
6007 * (and thus has not been already redirected
6008 * to new_bfqq or even some other bfq_queue),
6009 * then complete the merge and redirect it to
6010 * new_bfqq.
6011 */
6012 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
6013 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
6014 bfqq, new_bfqq);
6015
6016 bfq_clear_bfqq_just_created(bfqq);
6017 /*
6018 * rq is about to be enqueued into new_bfqq,
6019 * release rq reference on bfqq
6020 */
6021 bfq_put_queue(bfqq);
6022 rq->elv.priv[1] = new_bfqq;
6023 bfqq = new_bfqq;
6024 }
6025
6026 bfq_update_io_thinktime(bfqd, bfqq);
6027 bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
6028 bfq_update_io_seektime(bfqd, bfqq, rq);
6029
6030 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
6031 bfq_add_request(rq);
6032 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
6033
6034 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
6035 list_add_tail(&rq->queuelist, &bfqq->fifo);
6036
6037 bfq_rq_enqueued(bfqd, bfqq, rq);
6038
6039 return idle_timer_disabled;
6040}
6041
6042#ifdef CONFIG_BFQ_CGROUP_DEBUG
6043static void bfq_update_insert_stats(struct request_queue *q,
6044 struct bfq_queue *bfqq,
6045 bool idle_timer_disabled,
6046 blk_opf_t cmd_flags)
6047{
6048 if (!bfqq)
6049 return;
6050
6051 /*
6052 * bfqq still exists, because it can disappear only after
6053 * either it is merged with another queue, or the process it
6054 * is associated with exits. But both actions must be taken by
6055 * the same process currently executing this flow of
6056 * instructions.
6057 *
6058 * In addition, the following queue lock guarantees that
6059 * bfqq_group(bfqq) exists as well.
6060 */
6061 spin_lock_irq(&q->queue_lock);
6062 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
6063 if (idle_timer_disabled)
6064 bfqg_stats_update_idle_time(bfqq_group(bfqq));
6065 spin_unlock_irq(&q->queue_lock);
6066}
6067#else
6068static inline void bfq_update_insert_stats(struct request_queue *q,
6069 struct bfq_queue *bfqq,
6070 bool idle_timer_disabled,
6071 blk_opf_t cmd_flags) {}
6072#endif /* CONFIG_BFQ_CGROUP_DEBUG */
6073
6074static struct bfq_queue *bfq_init_rq(struct request *rq);
6075
6076static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
6077 bool at_head)
6078{
6079 struct request_queue *q = hctx->queue;
6080 struct bfq_data *bfqd = q->elevator->elevator_data;
6081 struct bfq_queue *bfqq;
6082 bool idle_timer_disabled = false;
6083 blk_opf_t cmd_flags;
6084 LIST_HEAD(free);
6085
6086#ifdef CONFIG_BFQ_GROUP_IOSCHED
6087 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) && rq->bio)
6088 bfqg_stats_update_legacy_io(q, rq);
6089#endif
6090 spin_lock_irq(&bfqd->lock);
6091 bfqq = bfq_init_rq(rq);
6092 if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
6093 spin_unlock_irq(&bfqd->lock);
6094 blk_mq_free_requests(&free);
6095 return;
6096 }
6097
6098 trace_block_rq_insert(rq);
6099
6100 if (!bfqq || at_head) {
6101 if (at_head)
6102 list_add(&rq->queuelist, &bfqd->dispatch);
6103 else
6104 list_add_tail(&rq->queuelist, &bfqd->dispatch);
6105 } else {
6106 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
6107 /*
6108 * Update bfqq, because, if a queue merge has occurred
6109 * in __bfq_insert_request, then rq has been
6110 * redirected into a new queue.
6111 */
6112 bfqq = RQ_BFQQ(rq);
6113
6114 if (rq_mergeable(rq)) {
6115 elv_rqhash_add(q, rq);
6116 if (!q->last_merge)
6117 q->last_merge = rq;
6118 }
6119 }
6120
6121 /*
6122 * Cache cmd_flags before releasing scheduler lock, because rq
6123 * may disappear afterwards (for example, because of a request
6124 * merge).
6125 */
6126 cmd_flags = rq->cmd_flags;
6127 spin_unlock_irq(&bfqd->lock);
6128
6129 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
6130 cmd_flags);
6131}
6132
6133static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
6134 struct list_head *list, bool at_head)
6135{
6136 while (!list_empty(list)) {
6137 struct request *rq;
6138
6139 rq = list_first_entry(list, struct request, queuelist);
6140 list_del_init(&rq->queuelist);
6141 bfq_insert_request(hctx, rq, at_head);
6142 }
6143}
6144
6145static void bfq_update_hw_tag(struct bfq_data *bfqd)
6146{
6147 struct bfq_queue *bfqq = bfqd->in_service_queue;
6148
6149 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
6150 bfqd->rq_in_driver);
6151
6152 if (bfqd->hw_tag == 1)
6153 return;
6154
6155 /*
6156 * This sample is valid if the number of outstanding requests
6157 * is large enough to allow a queueing behavior. Note that the
6158 * sum is not exact, as it's not taking into account deactivated
6159 * requests.
6160 */
6161 if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
6162 return;
6163
6164 /*
6165 * If active queue hasn't enough requests and can idle, bfq might not
6166 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
6167 * case
6168 */
6169 if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
6170 bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
6171 BFQ_HW_QUEUE_THRESHOLD &&
6172 bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
6173 return;
6174
6175 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
6176 return;
6177
6178 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
6179 bfqd->max_rq_in_driver = 0;
6180 bfqd->hw_tag_samples = 0;
6181
6182 bfqd->nonrot_with_queueing =
6183 blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
6184}
6185
6186static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
6187{
6188 u64 now_ns;
6189 u32 delta_us;
6190
6191 bfq_update_hw_tag(bfqd);
6192
6193 bfqd->rq_in_driver--;
6194 bfqq->dispatched--;
6195
6196 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
6197 /*
6198 * Set budget_timeout (which we overload to store the
6199 * time at which the queue remains with no backlog and
6200 * no outstanding request; used by the weight-raising
6201 * mechanism).
6202 */
6203 bfqq->budget_timeout = jiffies;
6204
6205 bfq_del_bfqq_in_groups_with_pending_reqs(bfqq);
6206 bfq_weights_tree_remove(bfqq);
6207 }
6208
6209 now_ns = ktime_get_ns();
6210
6211 bfqq->ttime.last_end_request = now_ns;
6212
6213 /*
6214 * Using us instead of ns, to get a reasonable precision in
6215 * computing rate in next check.
6216 */
6217 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
6218
6219 /*
6220 * If the request took rather long to complete, and, according
6221 * to the maximum request size recorded, this completion latency
6222 * implies that the request was certainly served at a very low
6223 * rate (less than 1M sectors/sec), then the whole observation
6224 * interval that lasts up to this time instant cannot be a
6225 * valid time interval for computing a new peak rate. Invoke
6226 * bfq_update_rate_reset to have the following three steps
6227 * taken:
6228 * - close the observation interval at the last (previous)
6229 * request dispatch or completion
6230 * - compute rate, if possible, for that observation interval
6231 * - reset to zero samples, which will trigger a proper
6232 * re-initialization of the observation interval on next
6233 * dispatch
6234 */
6235 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
6236 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
6237 1UL<<(BFQ_RATE_SHIFT - 10))
6238 bfq_update_rate_reset(bfqd, NULL);
6239 bfqd->last_completion = now_ns;
6240 /*
6241 * Shared queues are likely to receive I/O at a high
6242 * rate. This may deceptively let them be considered as wakers
6243 * of other queues. But a false waker will unjustly steal
6244 * bandwidth to its supposedly woken queue. So considering
6245 * also shared queues in the waking mechanism may cause more
6246 * control troubles than throughput benefits. Then reset
6247 * last_completed_rq_bfqq if bfqq is a shared queue.
6248 */
6249 if (!bfq_bfqq_coop(bfqq))
6250 bfqd->last_completed_rq_bfqq = bfqq;
6251 else
6252 bfqd->last_completed_rq_bfqq = NULL;
6253
6254 /*
6255 * If we are waiting to discover whether the request pattern
6256 * of the task associated with the queue is actually
6257 * isochronous, and both requisites for this condition to hold
6258 * are now satisfied, then compute soft_rt_next_start (see the
6259 * comments on the function bfq_bfqq_softrt_next_start()). We
6260 * do not compute soft_rt_next_start if bfqq is in interactive
6261 * weight raising (see the comments in bfq_bfqq_expire() for
6262 * an explanation). We schedule this delayed update when bfqq
6263 * expires, if it still has in-flight requests.
6264 */
6265 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
6266 RB_EMPTY_ROOT(&bfqq->sort_list) &&
6267 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
6268 bfqq->soft_rt_next_start =
6269 bfq_bfqq_softrt_next_start(bfqd, bfqq);
6270
6271 /*
6272 * If this is the in-service queue, check if it needs to be expired,
6273 * or if we want to idle in case it has no pending requests.
6274 */
6275 if (bfqd->in_service_queue == bfqq) {
6276 if (bfq_bfqq_must_idle(bfqq)) {
6277 if (bfqq->dispatched == 0)
6278 bfq_arm_slice_timer(bfqd);
6279 /*
6280 * If we get here, we do not expire bfqq, even
6281 * if bfqq was in budget timeout or had no
6282 * more requests (as controlled in the next
6283 * conditional instructions). The reason for
6284 * not expiring bfqq is as follows.
6285 *
6286 * Here bfqq->dispatched > 0 holds, but
6287 * bfq_bfqq_must_idle() returned true. This
6288 * implies that, even if no request arrives
6289 * for bfqq before bfqq->dispatched reaches 0,
6290 * bfqq will, however, not be expired on the
6291 * completion event that causes bfqq->dispatch
6292 * to reach zero. In contrast, on this event,
6293 * bfqq will start enjoying device idling
6294 * (I/O-dispatch plugging).
6295 *
6296 * But, if we expired bfqq here, bfqq would
6297 * not have the chance to enjoy device idling
6298 * when bfqq->dispatched finally reaches
6299 * zero. This would expose bfqq to violation
6300 * of its reserved service guarantees.
6301 */
6302 return;
6303 } else if (bfq_may_expire_for_budg_timeout(bfqq))
6304 bfq_bfqq_expire(bfqd, bfqq, false,
6305 BFQQE_BUDGET_TIMEOUT);
6306 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
6307 (bfqq->dispatched == 0 ||
6308 !bfq_better_to_idle(bfqq)))
6309 bfq_bfqq_expire(bfqd, bfqq, false,
6310 BFQQE_NO_MORE_REQUESTS);
6311 }
6312
6313 if (!bfqd->rq_in_driver)
6314 bfq_schedule_dispatch(bfqd);
6315}
6316
6317/*
6318 * The processes associated with bfqq may happen to generate their
6319 * cumulative I/O at a lower rate than the rate at which the device
6320 * could serve the same I/O. This is rather probable, e.g., if only
6321 * one process is associated with bfqq and the device is an SSD. It
6322 * results in bfqq becoming often empty while in service. In this
6323 * respect, if BFQ is allowed to switch to another queue when bfqq
6324 * remains empty, then the device goes on being fed with I/O requests,
6325 * and the throughput is not affected. In contrast, if BFQ is not
6326 * allowed to switch to another queue---because bfqq is sync and
6327 * I/O-dispatch needs to be plugged while bfqq is temporarily
6328 * empty---then, during the service of bfqq, there will be frequent
6329 * "service holes", i.e., time intervals during which bfqq gets empty
6330 * and the device can only consume the I/O already queued in its
6331 * hardware queues. During service holes, the device may even get to
6332 * remaining idle. In the end, during the service of bfqq, the device
6333 * is driven at a lower speed than the one it can reach with the kind
6334 * of I/O flowing through bfqq.
6335 *
6336 * To counter this loss of throughput, BFQ implements a "request
6337 * injection mechanism", which tries to fill the above service holes
6338 * with I/O requests taken from other queues. The hard part in this
6339 * mechanism is finding the right amount of I/O to inject, so as to
6340 * both boost throughput and not break bfqq's bandwidth and latency
6341 * guarantees. In this respect, the mechanism maintains a per-queue
6342 * inject limit, computed as below. While bfqq is empty, the injection
6343 * mechanism dispatches extra I/O requests only until the total number
6344 * of I/O requests in flight---i.e., already dispatched but not yet
6345 * completed---remains lower than this limit.
6346 *
6347 * A first definition comes in handy to introduce the algorithm by
6348 * which the inject limit is computed. We define as first request for
6349 * bfqq, an I/O request for bfqq that arrives while bfqq is in
6350 * service, and causes bfqq to switch from empty to non-empty. The
6351 * algorithm updates the limit as a function of the effect of
6352 * injection on the service times of only the first requests of
6353 * bfqq. The reason for this restriction is that these are the
6354 * requests whose service time is affected most, because they are the
6355 * first to arrive after injection possibly occurred.
6356 *
6357 * To evaluate the effect of injection, the algorithm measures the
6358 * "total service time" of first requests. We define as total service
6359 * time of an I/O request, the time that elapses since when the
6360 * request is enqueued into bfqq, to when it is completed. This
6361 * quantity allows the whole effect of injection to be measured. It is
6362 * easy to see why. Suppose that some requests of other queues are
6363 * actually injected while bfqq is empty, and that a new request R
6364 * then arrives for bfqq. If the device does start to serve all or
6365 * part of the injected requests during the service hole, then,
6366 * because of this extra service, it may delay the next invocation of
6367 * the dispatch hook of BFQ. Then, even after R gets eventually
6368 * dispatched, the device may delay the actual service of R if it is
6369 * still busy serving the extra requests, or if it decides to serve,
6370 * before R, some extra request still present in its queues. As a
6371 * conclusion, the cumulative extra delay caused by injection can be
6372 * easily evaluated by just comparing the total service time of first
6373 * requests with and without injection.
6374 *
6375 * The limit-update algorithm works as follows. On the arrival of a
6376 * first request of bfqq, the algorithm measures the total time of the
6377 * request only if one of the three cases below holds, and, for each
6378 * case, it updates the limit as described below:
6379 *
6380 * (1) If there is no in-flight request. This gives a baseline for the
6381 * total service time of the requests of bfqq. If the baseline has
6382 * not been computed yet, then, after computing it, the limit is
6383 * set to 1, to start boosting throughput, and to prepare the
6384 * ground for the next case. If the baseline has already been
6385 * computed, then it is updated, in case it results to be lower
6386 * than the previous value.
6387 *
6388 * (2) If the limit is higher than 0 and there are in-flight
6389 * requests. By comparing the total service time in this case with
6390 * the above baseline, it is possible to know at which extent the
6391 * current value of the limit is inflating the total service
6392 * time. If the inflation is below a certain threshold, then bfqq
6393 * is assumed to be suffering from no perceivable loss of its
6394 * service guarantees, and the limit is even tentatively
6395 * increased. If the inflation is above the threshold, then the
6396 * limit is decreased. Due to the lack of any hysteresis, this
6397 * logic makes the limit oscillate even in steady workload
6398 * conditions. Yet we opted for it, because it is fast in reaching
6399 * the best value for the limit, as a function of the current I/O
6400 * workload. To reduce oscillations, this step is disabled for a
6401 * short time interval after the limit happens to be decreased.
6402 *
6403 * (3) Periodically, after resetting the limit, to make sure that the
6404 * limit eventually drops in case the workload changes. This is
6405 * needed because, after the limit has gone safely up for a
6406 * certain workload, it is impossible to guess whether the
6407 * baseline total service time may have changed, without measuring
6408 * it again without injection. A more effective version of this
6409 * step might be to just sample the baseline, by interrupting
6410 * injection only once, and then to reset/lower the limit only if
6411 * the total service time with the current limit does happen to be
6412 * too large.
6413 *
6414 * More details on each step are provided in the comments on the
6415 * pieces of code that implement these steps: the branch handling the
6416 * transition from empty to non empty in bfq_add_request(), the branch
6417 * handling injection in bfq_select_queue(), and the function
6418 * bfq_choose_bfqq_for_injection(). These comments also explain some
6419 * exceptions, made by the injection mechanism in some special cases.
6420 */
6421static void bfq_update_inject_limit(struct bfq_data *bfqd,
6422 struct bfq_queue *bfqq)
6423{
6424 u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns;
6425 unsigned int old_limit = bfqq->inject_limit;
6426
6427 if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) {
6428 u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
6429
6430 if (tot_time_ns >= threshold && old_limit > 0) {
6431 bfqq->inject_limit--;
6432 bfqq->decrease_time_jif = jiffies;
6433 } else if (tot_time_ns < threshold &&
6434 old_limit <= bfqd->max_rq_in_driver)
6435 bfqq->inject_limit++;
6436 }
6437
6438 /*
6439 * Either we still have to compute the base value for the
6440 * total service time, and there seem to be the right
6441 * conditions to do it, or we can lower the last base value
6442 * computed.
6443 *
6444 * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
6445 * request in flight, because this function is in the code
6446 * path that handles the completion of a request of bfqq, and,
6447 * in particular, this function is executed before
6448 * bfqd->rq_in_driver is decremented in such a code path.
6449 */
6450 if ((bfqq->last_serv_time_ns == 0 && bfqd->rq_in_driver == 1) ||
6451 tot_time_ns < bfqq->last_serv_time_ns) {
6452 if (bfqq->last_serv_time_ns == 0) {
6453 /*
6454 * Now we certainly have a base value: make sure we
6455 * start trying injection.
6456 */
6457 bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
6458 }
6459 bfqq->last_serv_time_ns = tot_time_ns;
6460 } else if (!bfqd->rqs_injected && bfqd->rq_in_driver == 1)
6461 /*
6462 * No I/O injected and no request still in service in
6463 * the drive: these are the exact conditions for
6464 * computing the base value of the total service time
6465 * for bfqq. So let's update this value, because it is
6466 * rather variable. For example, it varies if the size
6467 * or the spatial locality of the I/O requests in bfqq
6468 * change.
6469 */
6470 bfqq->last_serv_time_ns = tot_time_ns;
6471
6472
6473 /* update complete, not waiting for any request completion any longer */
6474 bfqd->waited_rq = NULL;
6475 bfqd->rqs_injected = false;
6476}
6477
6478/*
6479 * Handle either a requeue or a finish for rq. The things to do are
6480 * the same in both cases: all references to rq are to be dropped. In
6481 * particular, rq is considered completed from the point of view of
6482 * the scheduler.
6483 */
6484static void bfq_finish_requeue_request(struct request *rq)
6485{
6486 struct bfq_queue *bfqq = RQ_BFQQ(rq);
6487 struct bfq_data *bfqd;
6488 unsigned long flags;
6489
6490 /*
6491 * rq either is not associated with any icq, or is an already
6492 * requeued request that has not (yet) been re-inserted into
6493 * a bfq_queue.
6494 */
6495 if (!rq->elv.icq || !bfqq)
6496 return;
6497
6498 bfqd = bfqq->bfqd;
6499
6500 if (rq->rq_flags & RQF_STARTED)
6501 bfqg_stats_update_completion(bfqq_group(bfqq),
6502 rq->start_time_ns,
6503 rq->io_start_time_ns,
6504 rq->cmd_flags);
6505
6506 spin_lock_irqsave(&bfqd->lock, flags);
6507 if (likely(rq->rq_flags & RQF_STARTED)) {
6508 if (rq == bfqd->waited_rq)
6509 bfq_update_inject_limit(bfqd, bfqq);
6510
6511 bfq_completed_request(bfqq, bfqd);
6512 }
6513 bfqq_request_freed(bfqq);
6514 bfq_put_queue(bfqq);
6515 RQ_BIC(rq)->requests--;
6516 spin_unlock_irqrestore(&bfqd->lock, flags);
6517
6518 /*
6519 * Reset private fields. In case of a requeue, this allows
6520 * this function to correctly do nothing if it is spuriously
6521 * invoked again on this same request (see the check at the
6522 * beginning of the function). Probably, a better general
6523 * design would be to prevent blk-mq from invoking the requeue
6524 * or finish hooks of an elevator, for a request that is not
6525 * referred by that elevator.
6526 *
6527 * Resetting the following fields would break the
6528 * request-insertion logic if rq is re-inserted into a bfq
6529 * internal queue, without a re-preparation. Here we assume
6530 * that re-insertions of requeued requests, without
6531 * re-preparation, can happen only for pass_through or at_head
6532 * requests (which are not re-inserted into bfq internal
6533 * queues).
6534 */
6535 rq->elv.priv[0] = NULL;
6536 rq->elv.priv[1] = NULL;
6537}
6538
6539static void bfq_finish_request(struct request *rq)
6540{
6541 bfq_finish_requeue_request(rq);
6542
6543 if (rq->elv.icq) {
6544 put_io_context(rq->elv.icq->ioc);
6545 rq->elv.icq = NULL;
6546 }
6547}
6548
6549/*
6550 * Removes the association between the current task and bfqq, assuming
6551 * that bic points to the bfq iocontext of the task.
6552 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
6553 * was the last process referring to that bfqq.
6554 */
6555static struct bfq_queue *
6556bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
6557{
6558 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
6559
6560 if (bfqq_process_refs(bfqq) == 1) {
6561 bfqq->pid = current->pid;
6562 bfq_clear_bfqq_coop(bfqq);
6563 bfq_clear_bfqq_split_coop(bfqq);
6564 return bfqq;
6565 }
6566
6567 bic_set_bfqq(bic, NULL, true);
6568
6569 bfq_put_cooperator(bfqq);
6570
6571 bfq_release_process_ref(bfqq->bfqd, bfqq);
6572 return NULL;
6573}
6574
6575static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
6576 struct bfq_io_cq *bic,
6577 struct bio *bio,
6578 bool split, bool is_sync,
6579 bool *new_queue)
6580{
6581 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
6582
6583 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
6584 return bfqq;
6585
6586 if (new_queue)
6587 *new_queue = true;
6588
6589 if (bfqq)
6590 bfq_put_queue(bfqq);
6591 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic, split);
6592
6593 bic_set_bfqq(bic, bfqq, is_sync);
6594 if (split && is_sync) {
6595 if ((bic->was_in_burst_list && bfqd->large_burst) ||
6596 bic->saved_in_large_burst)
6597 bfq_mark_bfqq_in_large_burst(bfqq);
6598 else {
6599 bfq_clear_bfqq_in_large_burst(bfqq);
6600 if (bic->was_in_burst_list)
6601 /*
6602 * If bfqq was in the current
6603 * burst list before being
6604 * merged, then we have to add
6605 * it back. And we do not need
6606 * to increase burst_size, as
6607 * we did not decrement
6608 * burst_size when we removed
6609 * bfqq from the burst list as
6610 * a consequence of a merge
6611 * (see comments in
6612 * bfq_put_queue). In this
6613 * respect, it would be rather
6614 * costly to know whether the
6615 * current burst list is still
6616 * the same burst list from
6617 * which bfqq was removed on
6618 * the merge. To avoid this
6619 * cost, if bfqq was in a
6620 * burst list, then we add
6621 * bfqq to the current burst
6622 * list without any further
6623 * check. This can cause
6624 * inappropriate insertions,
6625 * but rarely enough to not
6626 * harm the detection of large
6627 * bursts significantly.
6628 */
6629 hlist_add_head(&bfqq->burst_list_node,
6630 &bfqd->burst_list);
6631 }
6632 bfqq->split_time = jiffies;
6633 }
6634
6635 return bfqq;
6636}
6637
6638/*
6639 * Only reset private fields. The actual request preparation will be
6640 * performed by bfq_init_rq, when rq is either inserted or merged. See
6641 * comments on bfq_init_rq for the reason behind this delayed
6642 * preparation.
6643 */
6644static void bfq_prepare_request(struct request *rq)
6645{
6646 rq->elv.icq = ioc_find_get_icq(rq->q);
6647
6648 /*
6649 * Regardless of whether we have an icq attached, we have to
6650 * clear the scheduler pointers, as they might point to
6651 * previously allocated bic/bfqq structs.
6652 */
6653 rq->elv.priv[0] = rq->elv.priv[1] = NULL;
6654}
6655
6656/*
6657 * If needed, init rq, allocate bfq data structures associated with
6658 * rq, and increment reference counters in the destination bfq_queue
6659 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6660 * not associated with any bfq_queue.
6661 *
6662 * This function is invoked by the functions that perform rq insertion
6663 * or merging. One may have expected the above preparation operations
6664 * to be performed in bfq_prepare_request, and not delayed to when rq
6665 * is inserted or merged. The rationale behind this delayed
6666 * preparation is that, after the prepare_request hook is invoked for
6667 * rq, rq may still be transformed into a request with no icq, i.e., a
6668 * request not associated with any queue. No bfq hook is invoked to
6669 * signal this transformation. As a consequence, should these
6670 * preparation operations be performed when the prepare_request hook
6671 * is invoked, and should rq be transformed one moment later, bfq
6672 * would end up in an inconsistent state, because it would have
6673 * incremented some queue counters for an rq destined to
6674 * transformation, without any chance to correctly lower these
6675 * counters back. In contrast, no transformation can still happen for
6676 * rq after rq has been inserted or merged. So, it is safe to execute
6677 * these preparation operations when rq is finally inserted or merged.
6678 */
6679static struct bfq_queue *bfq_init_rq(struct request *rq)
6680{
6681 struct request_queue *q = rq->q;
6682 struct bio *bio = rq->bio;
6683 struct bfq_data *bfqd = q->elevator->elevator_data;
6684 struct bfq_io_cq *bic;
6685 const int is_sync = rq_is_sync(rq);
6686 struct bfq_queue *bfqq;
6687 bool new_queue = false;
6688 bool bfqq_already_existing = false, split = false;
6689
6690 if (unlikely(!rq->elv.icq))
6691 return NULL;
6692
6693 /*
6694 * Assuming that elv.priv[1] is set only if everything is set
6695 * for this rq. This holds true, because this function is
6696 * invoked only for insertion or merging, and, after such
6697 * events, a request cannot be manipulated any longer before
6698 * being removed from bfq.
6699 */
6700 if (rq->elv.priv[1])
6701 return rq->elv.priv[1];
6702
6703 bic = icq_to_bic(rq->elv.icq);
6704
6705 bfq_check_ioprio_change(bic, bio);
6706
6707 bfq_bic_update_cgroup(bic, bio);
6708
6709 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
6710 &new_queue);
6711
6712 if (likely(!new_queue)) {
6713 /* If the queue was seeky for too long, break it apart. */
6714 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq) &&
6715 !bic->stably_merged) {
6716 struct bfq_queue *old_bfqq = bfqq;
6717
6718 /* Update bic before losing reference to bfqq */
6719 if (bfq_bfqq_in_large_burst(bfqq))
6720 bic->saved_in_large_burst = true;
6721
6722 bfqq = bfq_split_bfqq(bic, bfqq);
6723 split = true;
6724
6725 if (!bfqq) {
6726 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
6727 true, is_sync,
6728 NULL);
6729 if (unlikely(bfqq == &bfqd->oom_bfqq))
6730 bfqq_already_existing = true;
6731 } else
6732 bfqq_already_existing = true;
6733
6734 if (!bfqq_already_existing) {
6735 bfqq->waker_bfqq = old_bfqq->waker_bfqq;
6736 bfqq->tentative_waker_bfqq = NULL;
6737
6738 /*
6739 * If the waker queue disappears, then
6740 * new_bfqq->waker_bfqq must be
6741 * reset. So insert new_bfqq into the
6742 * woken_list of the waker. See
6743 * bfq_check_waker for details.
6744 */
6745 if (bfqq->waker_bfqq)
6746 hlist_add_head(&bfqq->woken_list_node,
6747 &bfqq->waker_bfqq->woken_list);
6748 }
6749 }
6750 }
6751
6752 bfqq_request_allocated(bfqq);
6753 bfqq->ref++;
6754 bic->requests++;
6755 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
6756 rq, bfqq, bfqq->ref);
6757
6758 rq->elv.priv[0] = bic;
6759 rq->elv.priv[1] = bfqq;
6760
6761 /*
6762 * If a bfq_queue has only one process reference, it is owned
6763 * by only this bic: we can then set bfqq->bic = bic. in
6764 * addition, if the queue has also just been split, we have to
6765 * resume its state.
6766 */
6767 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
6768 bfqq->bic = bic;
6769 if (split) {
6770 /*
6771 * The queue has just been split from a shared
6772 * queue: restore the idle window and the
6773 * possible weight raising period.
6774 */
6775 bfq_bfqq_resume_state(bfqq, bfqd, bic,
6776 bfqq_already_existing);
6777 }
6778 }
6779
6780 /*
6781 * Consider bfqq as possibly belonging to a burst of newly
6782 * created queues only if:
6783 * 1) A burst is actually happening (bfqd->burst_size > 0)
6784 * or
6785 * 2) There is no other active queue. In fact, if, in
6786 * contrast, there are active queues not belonging to the
6787 * possible burst bfqq may belong to, then there is no gain
6788 * in considering bfqq as belonging to a burst, and
6789 * therefore in not weight-raising bfqq. See comments on
6790 * bfq_handle_burst().
6791 *
6792 * This filtering also helps eliminating false positives,
6793 * occurring when bfqq does not belong to an actual large
6794 * burst, but some background task (e.g., a service) happens
6795 * to trigger the creation of new queues very close to when
6796 * bfqq and its possible companion queues are created. See
6797 * comments on bfq_handle_burst() for further details also on
6798 * this issue.
6799 */
6800 if (unlikely(bfq_bfqq_just_created(bfqq) &&
6801 (bfqd->burst_size > 0 ||
6802 bfq_tot_busy_queues(bfqd) == 0)))
6803 bfq_handle_burst(bfqd, bfqq);
6804
6805 return bfqq;
6806}
6807
6808static void
6809bfq_idle_slice_timer_body(struct bfq_data *bfqd, struct bfq_queue *bfqq)
6810{
6811 enum bfqq_expiration reason;
6812 unsigned long flags;
6813
6814 spin_lock_irqsave(&bfqd->lock, flags);
6815
6816 /*
6817 * Considering that bfqq may be in race, we should firstly check
6818 * whether bfqq is in service before doing something on it. If
6819 * the bfqq in race is not in service, it has already been expired
6820 * through __bfq_bfqq_expire func and its wait_request flags has
6821 * been cleared in __bfq_bfqd_reset_in_service func.
6822 */
6823 if (bfqq != bfqd->in_service_queue) {
6824 spin_unlock_irqrestore(&bfqd->lock, flags);
6825 return;
6826 }
6827
6828 bfq_clear_bfqq_wait_request(bfqq);
6829
6830 if (bfq_bfqq_budget_timeout(bfqq))
6831 /*
6832 * Also here the queue can be safely expired
6833 * for budget timeout without wasting
6834 * guarantees
6835 */
6836 reason = BFQQE_BUDGET_TIMEOUT;
6837 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
6838 /*
6839 * The queue may not be empty upon timer expiration,
6840 * because we may not disable the timer when the
6841 * first request of the in-service queue arrives
6842 * during disk idling.
6843 */
6844 reason = BFQQE_TOO_IDLE;
6845 else
6846 goto schedule_dispatch;
6847
6848 bfq_bfqq_expire(bfqd, bfqq, true, reason);
6849
6850schedule_dispatch:
6851 bfq_schedule_dispatch(bfqd);
6852 spin_unlock_irqrestore(&bfqd->lock, flags);
6853}
6854
6855/*
6856 * Handler of the expiration of the timer running if the in-service queue
6857 * is idling inside its time slice.
6858 */
6859static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
6860{
6861 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
6862 idle_slice_timer);
6863 struct bfq_queue *bfqq = bfqd->in_service_queue;
6864
6865 /*
6866 * Theoretical race here: the in-service queue can be NULL or
6867 * different from the queue that was idling if a new request
6868 * arrives for the current queue and there is a full dispatch
6869 * cycle that changes the in-service queue. This can hardly
6870 * happen, but in the worst case we just expire a queue too
6871 * early.
6872 */
6873 if (bfqq)
6874 bfq_idle_slice_timer_body(bfqd, bfqq);
6875
6876 return HRTIMER_NORESTART;
6877}
6878
6879static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
6880 struct bfq_queue **bfqq_ptr)
6881{
6882 struct bfq_queue *bfqq = *bfqq_ptr;
6883
6884 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
6885 if (bfqq) {
6886 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
6887
6888 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
6889 bfqq, bfqq->ref);
6890 bfq_put_queue(bfqq);
6891 *bfqq_ptr = NULL;
6892 }
6893}
6894
6895/*
6896 * Release all the bfqg references to its async queues. If we are
6897 * deallocating the group these queues may still contain requests, so
6898 * we reparent them to the root cgroup (i.e., the only one that will
6899 * exist for sure until all the requests on a device are gone).
6900 */
6901void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
6902{
6903 int i, j;
6904
6905 for (i = 0; i < 2; i++)
6906 for (j = 0; j < IOPRIO_NR_LEVELS; j++)
6907 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
6908
6909 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
6910}
6911
6912/*
6913 * See the comments on bfq_limit_depth for the purpose of
6914 * the depths set in the function. Return minimum shallow depth we'll use.
6915 */
6916static void bfq_update_depths(struct bfq_data *bfqd, struct sbitmap_queue *bt)
6917{
6918 unsigned int depth = 1U << bt->sb.shift;
6919
6920 bfqd->full_depth_shift = bt->sb.shift;
6921 /*
6922 * In-word depths if no bfq_queue is being weight-raised:
6923 * leaving 25% of tags only for sync reads.
6924 *
6925 * In next formulas, right-shift the value
6926 * (1U<<bt->sb.shift), instead of computing directly
6927 * (1U<<(bt->sb.shift - something)), to be robust against
6928 * any possible value of bt->sb.shift, without having to
6929 * limit 'something'.
6930 */
6931 /* no more than 50% of tags for async I/O */
6932 bfqd->word_depths[0][0] = max(depth >> 1, 1U);
6933 /*
6934 * no more than 75% of tags for sync writes (25% extra tags
6935 * w.r.t. async I/O, to prevent async I/O from starving sync
6936 * writes)
6937 */
6938 bfqd->word_depths[0][1] = max((depth * 3) >> 2, 1U);
6939
6940 /*
6941 * In-word depths in case some bfq_queue is being weight-
6942 * raised: leaving ~63% of tags for sync reads. This is the
6943 * highest percentage for which, in our tests, application
6944 * start-up times didn't suffer from any regression due to tag
6945 * shortage.
6946 */
6947 /* no more than ~18% of tags for async I/O */
6948 bfqd->word_depths[1][0] = max((depth * 3) >> 4, 1U);
6949 /* no more than ~37% of tags for sync writes (~20% extra tags) */
6950 bfqd->word_depths[1][1] = max((depth * 6) >> 4, 1U);
6951}
6952
6953static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
6954{
6955 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
6956 struct blk_mq_tags *tags = hctx->sched_tags;
6957
6958 bfq_update_depths(bfqd, &tags->bitmap_tags);
6959 sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, 1);
6960}
6961
6962static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
6963{
6964 bfq_depth_updated(hctx);
6965 return 0;
6966}
6967
6968static void bfq_exit_queue(struct elevator_queue *e)
6969{
6970 struct bfq_data *bfqd = e->elevator_data;
6971 struct bfq_queue *bfqq, *n;
6972
6973 hrtimer_cancel(&bfqd->idle_slice_timer);
6974
6975 spin_lock_irq(&bfqd->lock);
6976 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
6977 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
6978 spin_unlock_irq(&bfqd->lock);
6979
6980 hrtimer_cancel(&bfqd->idle_slice_timer);
6981
6982 /* release oom-queue reference to root group */
6983 bfqg_and_blkg_put(bfqd->root_group);
6984
6985#ifdef CONFIG_BFQ_GROUP_IOSCHED
6986 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
6987#else
6988 spin_lock_irq(&bfqd->lock);
6989 bfq_put_async_queues(bfqd, bfqd->root_group);
6990 kfree(bfqd->root_group);
6991 spin_unlock_irq(&bfqd->lock);
6992#endif
6993
6994 blk_stat_disable_accounting(bfqd->queue);
6995 clear_bit(ELEVATOR_FLAG_DISABLE_WBT, &e->flags);
6996 wbt_enable_default(bfqd->queue);
6997
6998 kfree(bfqd);
6999}
7000
7001static void bfq_init_root_group(struct bfq_group *root_group,
7002 struct bfq_data *bfqd)
7003{
7004 int i;
7005
7006#ifdef CONFIG_BFQ_GROUP_IOSCHED
7007 root_group->entity.parent = NULL;
7008 root_group->my_entity = NULL;
7009 root_group->bfqd = bfqd;
7010#endif
7011 root_group->rq_pos_tree = RB_ROOT;
7012 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
7013 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
7014 root_group->sched_data.bfq_class_idle_last_service = jiffies;
7015}
7016
7017static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
7018{
7019 struct bfq_data *bfqd;
7020 struct elevator_queue *eq;
7021
7022 eq = elevator_alloc(q, e);
7023 if (!eq)
7024 return -ENOMEM;
7025
7026 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
7027 if (!bfqd) {
7028 kobject_put(&eq->kobj);
7029 return -ENOMEM;
7030 }
7031 eq->elevator_data = bfqd;
7032
7033 spin_lock_irq(&q->queue_lock);
7034 q->elevator = eq;
7035 spin_unlock_irq(&q->queue_lock);
7036
7037 /*
7038 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
7039 * Grab a permanent reference to it, so that the normal code flow
7040 * will not attempt to free it.
7041 */
7042 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
7043 bfqd->oom_bfqq.ref++;
7044 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
7045 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
7046 bfqd->oom_bfqq.entity.new_weight =
7047 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
7048
7049 /* oom_bfqq does not participate to bursts */
7050 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
7051
7052 /*
7053 * Trigger weight initialization, according to ioprio, at the
7054 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
7055 * class won't be changed any more.
7056 */
7057 bfqd->oom_bfqq.entity.prio_changed = 1;
7058
7059 bfqd->queue = q;
7060
7061 INIT_LIST_HEAD(&bfqd->dispatch);
7062
7063 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
7064 HRTIMER_MODE_REL);
7065 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
7066
7067 bfqd->queue_weights_tree = RB_ROOT_CACHED;
7068#ifdef CONFIG_BFQ_GROUP_IOSCHED
7069 bfqd->num_groups_with_pending_reqs = 0;
7070#endif
7071
7072 INIT_LIST_HEAD(&bfqd->active_list);
7073 INIT_LIST_HEAD(&bfqd->idle_list);
7074 INIT_HLIST_HEAD(&bfqd->burst_list);
7075
7076 bfqd->hw_tag = -1;
7077 bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
7078
7079 bfqd->bfq_max_budget = bfq_default_max_budget;
7080
7081 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
7082 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
7083 bfqd->bfq_back_max = bfq_back_max;
7084 bfqd->bfq_back_penalty = bfq_back_penalty;
7085 bfqd->bfq_slice_idle = bfq_slice_idle;
7086 bfqd->bfq_timeout = bfq_timeout;
7087
7088 bfqd->bfq_large_burst_thresh = 8;
7089 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
7090
7091 bfqd->low_latency = true;
7092
7093 /*
7094 * Trade-off between responsiveness and fairness.
7095 */
7096 bfqd->bfq_wr_coeff = 30;
7097 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
7098 bfqd->bfq_wr_max_time = 0;
7099 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
7100 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
7101 bfqd->bfq_wr_max_softrt_rate = 7000; /*
7102 * Approximate rate required
7103 * to playback or record a
7104 * high-definition compressed
7105 * video.
7106 */
7107 bfqd->wr_busy_queues = 0;
7108
7109 /*
7110 * Begin by assuming, optimistically, that the device peak
7111 * rate is equal to 2/3 of the highest reference rate.
7112 */
7113 bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
7114 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
7115 bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
7116
7117 spin_lock_init(&bfqd->lock);
7118
7119 /*
7120 * The invocation of the next bfq_create_group_hierarchy
7121 * function is the head of a chain of function calls
7122 * (bfq_create_group_hierarchy->blkcg_activate_policy->
7123 * blk_mq_freeze_queue) that may lead to the invocation of the
7124 * has_work hook function. For this reason,
7125 * bfq_create_group_hierarchy is invoked only after all
7126 * scheduler data has been initialized, apart from the fields
7127 * that can be initialized only after invoking
7128 * bfq_create_group_hierarchy. This, in particular, enables
7129 * has_work to correctly return false. Of course, to avoid
7130 * other inconsistencies, the blk-mq stack must then refrain
7131 * from invoking further scheduler hooks before this init
7132 * function is finished.
7133 */
7134 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
7135 if (!bfqd->root_group)
7136 goto out_free;
7137 bfq_init_root_group(bfqd->root_group, bfqd);
7138 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
7139
7140 /* We dispatch from request queue wide instead of hw queue */
7141 blk_queue_flag_set(QUEUE_FLAG_SQ_SCHED, q);
7142
7143 set_bit(ELEVATOR_FLAG_DISABLE_WBT, &eq->flags);
7144 wbt_disable_default(q);
7145 blk_stat_enable_accounting(q);
7146
7147 return 0;
7148
7149out_free:
7150 kfree(bfqd);
7151 kobject_put(&eq->kobj);
7152 return -ENOMEM;
7153}
7154
7155static void bfq_slab_kill(void)
7156{
7157 kmem_cache_destroy(bfq_pool);
7158}
7159
7160static int __init bfq_slab_setup(void)
7161{
7162 bfq_pool = KMEM_CACHE(bfq_queue, 0);
7163 if (!bfq_pool)
7164 return -ENOMEM;
7165 return 0;
7166}
7167
7168static ssize_t bfq_var_show(unsigned int var, char *page)
7169{
7170 return sprintf(page, "%u\n", var);
7171}
7172
7173static int bfq_var_store(unsigned long *var, const char *page)
7174{
7175 unsigned long new_val;
7176 int ret = kstrtoul(page, 10, &new_val);
7177
7178 if (ret)
7179 return ret;
7180 *var = new_val;
7181 return 0;
7182}
7183
7184#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
7185static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7186{ \
7187 struct bfq_data *bfqd = e->elevator_data; \
7188 u64 __data = __VAR; \
7189 if (__CONV == 1) \
7190 __data = jiffies_to_msecs(__data); \
7191 else if (__CONV == 2) \
7192 __data = div_u64(__data, NSEC_PER_MSEC); \
7193 return bfq_var_show(__data, (page)); \
7194}
7195SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
7196SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
7197SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
7198SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
7199SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
7200SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
7201SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
7202SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
7203SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
7204#undef SHOW_FUNCTION
7205
7206#define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
7207static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7208{ \
7209 struct bfq_data *bfqd = e->elevator_data; \
7210 u64 __data = __VAR; \
7211 __data = div_u64(__data, NSEC_PER_USEC); \
7212 return bfq_var_show(__data, (page)); \
7213}
7214USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
7215#undef USEC_SHOW_FUNCTION
7216
7217#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
7218static ssize_t \
7219__FUNC(struct elevator_queue *e, const char *page, size_t count) \
7220{ \
7221 struct bfq_data *bfqd = e->elevator_data; \
7222 unsigned long __data, __min = (MIN), __max = (MAX); \
7223 int ret; \
7224 \
7225 ret = bfq_var_store(&__data, (page)); \
7226 if (ret) \
7227 return ret; \
7228 if (__data < __min) \
7229 __data = __min; \
7230 else if (__data > __max) \
7231 __data = __max; \
7232 if (__CONV == 1) \
7233 *(__PTR) = msecs_to_jiffies(__data); \
7234 else if (__CONV == 2) \
7235 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
7236 else \
7237 *(__PTR) = __data; \
7238 return count; \
7239}
7240STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
7241 INT_MAX, 2);
7242STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
7243 INT_MAX, 2);
7244STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
7245STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
7246 INT_MAX, 0);
7247STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
7248#undef STORE_FUNCTION
7249
7250#define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
7251static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
7252{ \
7253 struct bfq_data *bfqd = e->elevator_data; \
7254 unsigned long __data, __min = (MIN), __max = (MAX); \
7255 int ret; \
7256 \
7257 ret = bfq_var_store(&__data, (page)); \
7258 if (ret) \
7259 return ret; \
7260 if (__data < __min) \
7261 __data = __min; \
7262 else if (__data > __max) \
7263 __data = __max; \
7264 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
7265 return count; \
7266}
7267USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
7268 UINT_MAX);
7269#undef USEC_STORE_FUNCTION
7270
7271static ssize_t bfq_max_budget_store(struct elevator_queue *e,
7272 const char *page, size_t count)
7273{
7274 struct bfq_data *bfqd = e->elevator_data;
7275 unsigned long __data;
7276 int ret;
7277
7278 ret = bfq_var_store(&__data, (page));
7279 if (ret)
7280 return ret;
7281
7282 if (__data == 0)
7283 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7284 else {
7285 if (__data > INT_MAX)
7286 __data = INT_MAX;
7287 bfqd->bfq_max_budget = __data;
7288 }
7289
7290 bfqd->bfq_user_max_budget = __data;
7291
7292 return count;
7293}
7294
7295/*
7296 * Leaving this name to preserve name compatibility with cfq
7297 * parameters, but this timeout is used for both sync and async.
7298 */
7299static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
7300 const char *page, size_t count)
7301{
7302 struct bfq_data *bfqd = e->elevator_data;
7303 unsigned long __data;
7304 int ret;
7305
7306 ret = bfq_var_store(&__data, (page));
7307 if (ret)
7308 return ret;
7309
7310 if (__data < 1)
7311 __data = 1;
7312 else if (__data > INT_MAX)
7313 __data = INT_MAX;
7314
7315 bfqd->bfq_timeout = msecs_to_jiffies(__data);
7316 if (bfqd->bfq_user_max_budget == 0)
7317 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7318
7319 return count;
7320}
7321
7322static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
7323 const char *page, size_t count)
7324{
7325 struct bfq_data *bfqd = e->elevator_data;
7326 unsigned long __data;
7327 int ret;
7328
7329 ret = bfq_var_store(&__data, (page));
7330 if (ret)
7331 return ret;
7332
7333 if (__data > 1)
7334 __data = 1;
7335 if (!bfqd->strict_guarantees && __data == 1
7336 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
7337 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
7338
7339 bfqd->strict_guarantees = __data;
7340
7341 return count;
7342}
7343
7344static ssize_t bfq_low_latency_store(struct elevator_queue *e,
7345 const char *page, size_t count)
7346{
7347 struct bfq_data *bfqd = e->elevator_data;
7348 unsigned long __data;
7349 int ret;
7350
7351 ret = bfq_var_store(&__data, (page));
7352 if (ret)
7353 return ret;
7354
7355 if (__data > 1)
7356 __data = 1;
7357 if (__data == 0 && bfqd->low_latency != 0)
7358 bfq_end_wr(bfqd);
7359 bfqd->low_latency = __data;
7360
7361 return count;
7362}
7363
7364#define BFQ_ATTR(name) \
7365 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
7366
7367static struct elv_fs_entry bfq_attrs[] = {
7368 BFQ_ATTR(fifo_expire_sync),
7369 BFQ_ATTR(fifo_expire_async),
7370 BFQ_ATTR(back_seek_max),
7371 BFQ_ATTR(back_seek_penalty),
7372 BFQ_ATTR(slice_idle),
7373 BFQ_ATTR(slice_idle_us),
7374 BFQ_ATTR(max_budget),
7375 BFQ_ATTR(timeout_sync),
7376 BFQ_ATTR(strict_guarantees),
7377 BFQ_ATTR(low_latency),
7378 __ATTR_NULL
7379};
7380
7381static struct elevator_type iosched_bfq_mq = {
7382 .ops = {
7383 .limit_depth = bfq_limit_depth,
7384 .prepare_request = bfq_prepare_request,
7385 .requeue_request = bfq_finish_requeue_request,
7386 .finish_request = bfq_finish_request,
7387 .exit_icq = bfq_exit_icq,
7388 .insert_requests = bfq_insert_requests,
7389 .dispatch_request = bfq_dispatch_request,
7390 .next_request = elv_rb_latter_request,
7391 .former_request = elv_rb_former_request,
7392 .allow_merge = bfq_allow_bio_merge,
7393 .bio_merge = bfq_bio_merge,
7394 .request_merge = bfq_request_merge,
7395 .requests_merged = bfq_requests_merged,
7396 .request_merged = bfq_request_merged,
7397 .has_work = bfq_has_work,
7398 .depth_updated = bfq_depth_updated,
7399 .init_hctx = bfq_init_hctx,
7400 .init_sched = bfq_init_queue,
7401 .exit_sched = bfq_exit_queue,
7402 },
7403
7404 .icq_size = sizeof(struct bfq_io_cq),
7405 .icq_align = __alignof__(struct bfq_io_cq),
7406 .elevator_attrs = bfq_attrs,
7407 .elevator_name = "bfq",
7408 .elevator_owner = THIS_MODULE,
7409};
7410MODULE_ALIAS("bfq-iosched");
7411
7412static int __init bfq_init(void)
7413{
7414 int ret;
7415
7416#ifdef CONFIG_BFQ_GROUP_IOSCHED
7417 ret = blkcg_policy_register(&blkcg_policy_bfq);
7418 if (ret)
7419 return ret;
7420#endif
7421
7422 ret = -ENOMEM;
7423 if (bfq_slab_setup())
7424 goto err_pol_unreg;
7425
7426 /*
7427 * Times to load large popular applications for the typical
7428 * systems installed on the reference devices (see the
7429 * comments before the definition of the next
7430 * array). Actually, we use slightly lower values, as the
7431 * estimated peak rate tends to be smaller than the actual
7432 * peak rate. The reason for this last fact is that estimates
7433 * are computed over much shorter time intervals than the long
7434 * intervals typically used for benchmarking. Why? First, to
7435 * adapt more quickly to variations. Second, because an I/O
7436 * scheduler cannot rely on a peak-rate-evaluation workload to
7437 * be run for a long time.
7438 */
7439 ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
7440 ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
7441
7442 ret = elv_register(&iosched_bfq_mq);
7443 if (ret)
7444 goto slab_kill;
7445
7446 return 0;
7447
7448slab_kill:
7449 bfq_slab_kill();
7450err_pol_unreg:
7451#ifdef CONFIG_BFQ_GROUP_IOSCHED
7452 blkcg_policy_unregister(&blkcg_policy_bfq);
7453#endif
7454 return ret;
7455}
7456
7457static void __exit bfq_exit(void)
7458{
7459 elv_unregister(&iosched_bfq_mq);
7460#ifdef CONFIG_BFQ_GROUP_IOSCHED
7461 blkcg_policy_unregister(&blkcg_policy_bfq);
7462#endif
7463 bfq_slab_kill();
7464}
7465
7466module_init(bfq_init);
7467module_exit(bfq_exit);
7468
7469MODULE_AUTHOR("Paolo Valente");
7470MODULE_LICENSE("GPL");
7471MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Budget Fair Queueing (BFQ) I/O scheduler.
4 *
5 * Based on ideas and code from CFQ:
6 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7 *
8 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9 * Paolo Valente <paolo.valente@unimore.it>
10 *
11 * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
12 * Arianna Avanzini <avanzini@google.com>
13 *
14 * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
15 *
16 * BFQ is a proportional-share I/O scheduler, with some extra
17 * low-latency capabilities. BFQ also supports full hierarchical
18 * scheduling through cgroups. Next paragraphs provide an introduction
19 * on BFQ inner workings. Details on BFQ benefits, usage and
20 * limitations can be found in Documentation/block/bfq-iosched.rst.
21 *
22 * BFQ is a proportional-share storage-I/O scheduling algorithm based
23 * on the slice-by-slice service scheme of CFQ. But BFQ assigns
24 * budgets, measured in number of sectors, to processes instead of
25 * time slices. The device is not granted to the in-service process
26 * for a given time slice, but until it has exhausted its assigned
27 * budget. This change from the time to the service domain enables BFQ
28 * to distribute the device throughput among processes as desired,
29 * without any distortion due to throughput fluctuations, or to device
30 * internal queueing. BFQ uses an ad hoc internal scheduler, called
31 * B-WF2Q+, to schedule processes according to their budgets. More
32 * precisely, BFQ schedules queues associated with processes. Each
33 * process/queue is assigned a user-configurable weight, and B-WF2Q+
34 * guarantees that each queue receives a fraction of the throughput
35 * proportional to its weight. Thanks to the accurate policy of
36 * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
37 * processes issuing sequential requests (to boost the throughput),
38 * and yet guarantee a low latency to interactive and soft real-time
39 * applications.
40 *
41 * In particular, to provide these low-latency guarantees, BFQ
42 * explicitly privileges the I/O of two classes of time-sensitive
43 * applications: interactive and soft real-time. In more detail, BFQ
44 * behaves this way if the low_latency parameter is set (default
45 * configuration). This feature enables BFQ to provide applications in
46 * these classes with a very low latency.
47 *
48 * To implement this feature, BFQ constantly tries to detect whether
49 * the I/O requests in a bfq_queue come from an interactive or a soft
50 * real-time application. For brevity, in these cases, the queue is
51 * said to be interactive or soft real-time. In both cases, BFQ
52 * privileges the service of the queue, over that of non-interactive
53 * and non-soft-real-time queues. This privileging is performed,
54 * mainly, by raising the weight of the queue. So, for brevity, we
55 * call just weight-raising periods the time periods during which a
56 * queue is privileged, because deemed interactive or soft real-time.
57 *
58 * The detection of soft real-time queues/applications is described in
59 * detail in the comments on the function
60 * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
61 * interactive queue works as follows: a queue is deemed interactive
62 * if it is constantly non empty only for a limited time interval,
63 * after which it does become empty. The queue may be deemed
64 * interactive again (for a limited time), if it restarts being
65 * constantly non empty, provided that this happens only after the
66 * queue has remained empty for a given minimum idle time.
67 *
68 * By default, BFQ computes automatically the above maximum time
69 * interval, i.e., the time interval after which a constantly
70 * non-empty queue stops being deemed interactive. Since a queue is
71 * weight-raised while it is deemed interactive, this maximum time
72 * interval happens to coincide with the (maximum) duration of the
73 * weight-raising for interactive queues.
74 *
75 * Finally, BFQ also features additional heuristics for
76 * preserving both a low latency and a high throughput on NCQ-capable,
77 * rotational or flash-based devices, and to get the job done quickly
78 * for applications consisting in many I/O-bound processes.
79 *
80 * NOTE: if the main or only goal, with a given device, is to achieve
81 * the maximum-possible throughput at all times, then do switch off
82 * all low-latency heuristics for that device, by setting low_latency
83 * to 0.
84 *
85 * BFQ is described in [1], where also a reference to the initial,
86 * more theoretical paper on BFQ can be found. The interested reader
87 * can find in the latter paper full details on the main algorithm, as
88 * well as formulas of the guarantees and formal proofs of all the
89 * properties. With respect to the version of BFQ presented in these
90 * papers, this implementation adds a few more heuristics, such as the
91 * ones that guarantee a low latency to interactive and soft real-time
92 * applications, and a hierarchical extension based on H-WF2Q+.
93 *
94 * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
95 * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
96 * with O(log N) complexity derives from the one introduced with EEVDF
97 * in [3].
98 *
99 * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
100 * Scheduler", Proceedings of the First Workshop on Mobile System
101 * Technologies (MST-2015), May 2015.
102 * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
103 *
104 * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
105 * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
106 * Oct 1997.
107 *
108 * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
109 *
110 * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
111 * First: A Flexible and Accurate Mechanism for Proportional Share
112 * Resource Allocation", technical report.
113 *
114 * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
115 */
116#include <linux/module.h>
117#include <linux/slab.h>
118#include <linux/blkdev.h>
119#include <linux/cgroup.h>
120#include <linux/ktime.h>
121#include <linux/rbtree.h>
122#include <linux/ioprio.h>
123#include <linux/sbitmap.h>
124#include <linux/delay.h>
125#include <linux/backing-dev.h>
126
127#include <trace/events/block.h>
128
129#include "elevator.h"
130#include "blk.h"
131#include "blk-mq.h"
132#include "blk-mq-sched.h"
133#include "bfq-iosched.h"
134#include "blk-wbt.h"
135
136#define BFQ_BFQQ_FNS(name) \
137void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
138{ \
139 __set_bit(BFQQF_##name, &(bfqq)->flags); \
140} \
141void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
142{ \
143 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
144} \
145int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
146{ \
147 return test_bit(BFQQF_##name, &(bfqq)->flags); \
148}
149
150BFQ_BFQQ_FNS(just_created);
151BFQ_BFQQ_FNS(busy);
152BFQ_BFQQ_FNS(wait_request);
153BFQ_BFQQ_FNS(non_blocking_wait_rq);
154BFQ_BFQQ_FNS(fifo_expire);
155BFQ_BFQQ_FNS(has_short_ttime);
156BFQ_BFQQ_FNS(sync);
157BFQ_BFQQ_FNS(IO_bound);
158BFQ_BFQQ_FNS(in_large_burst);
159BFQ_BFQQ_FNS(coop);
160BFQ_BFQQ_FNS(split_coop);
161BFQ_BFQQ_FNS(softrt_update);
162#undef BFQ_BFQQ_FNS \
163
164/* Expiration time of async (0) and sync (1) requests, in ns. */
165static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
166
167/* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
168static const int bfq_back_max = 16 * 1024;
169
170/* Penalty of a backwards seek, in number of sectors. */
171static const int bfq_back_penalty = 2;
172
173/* Idling period duration, in ns. */
174static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
175
176/* Minimum number of assigned budgets for which stats are safe to compute. */
177static const int bfq_stats_min_budgets = 194;
178
179/* Default maximum budget values, in sectors and number of requests. */
180static const int bfq_default_max_budget = 16 * 1024;
181
182/*
183 * When a sync request is dispatched, the queue that contains that
184 * request, and all the ancestor entities of that queue, are charged
185 * with the number of sectors of the request. In contrast, if the
186 * request is async, then the queue and its ancestor entities are
187 * charged with the number of sectors of the request, multiplied by
188 * the factor below. This throttles the bandwidth for async I/O,
189 * w.r.t. to sync I/O, and it is done to counter the tendency of async
190 * writes to steal I/O throughput to reads.
191 *
192 * The current value of this parameter is the result of a tuning with
193 * several hardware and software configurations. We tried to find the
194 * lowest value for which writes do not cause noticeable problems to
195 * reads. In fact, the lower this parameter, the stabler I/O control,
196 * in the following respect. The lower this parameter is, the less
197 * the bandwidth enjoyed by a group decreases
198 * - when the group does writes, w.r.t. to when it does reads;
199 * - when other groups do reads, w.r.t. to when they do writes.
200 */
201static const int bfq_async_charge_factor = 3;
202
203/* Default timeout values, in jiffies, approximating CFQ defaults. */
204const int bfq_timeout = HZ / 8;
205
206/*
207 * Time limit for merging (see comments in bfq_setup_cooperator). Set
208 * to the slowest value that, in our tests, proved to be effective in
209 * removing false positives, while not causing true positives to miss
210 * queue merging.
211 *
212 * As can be deduced from the low time limit below, queue merging, if
213 * successful, happens at the very beginning of the I/O of the involved
214 * cooperating processes, as a consequence of the arrival of the very
215 * first requests from each cooperator. After that, there is very
216 * little chance to find cooperators.
217 */
218static const unsigned long bfq_merge_time_limit = HZ/10;
219
220static struct kmem_cache *bfq_pool;
221
222/* Below this threshold (in ns), we consider thinktime immediate. */
223#define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
224
225/* hw_tag detection: parallel requests threshold and min samples needed. */
226#define BFQ_HW_QUEUE_THRESHOLD 3
227#define BFQ_HW_QUEUE_SAMPLES 32
228
229#define BFQQ_SEEK_THR (sector_t)(8 * 100)
230#define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
231#define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \
232 (get_sdist(last_pos, rq) > \
233 BFQQ_SEEK_THR && \
234 (!blk_queue_nonrot(bfqd->queue) || \
235 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT))
236#define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
237#define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19)
238/*
239 * Sync random I/O is likely to be confused with soft real-time I/O,
240 * because it is characterized by limited throughput and apparently
241 * isochronous arrival pattern. To avoid false positives, queues
242 * containing only random (seeky) I/O are prevented from being tagged
243 * as soft real-time.
244 */
245#define BFQQ_TOTALLY_SEEKY(bfqq) (bfqq->seek_history == -1)
246
247/* Min number of samples required to perform peak-rate update */
248#define BFQ_RATE_MIN_SAMPLES 32
249/* Min observation time interval required to perform a peak-rate update (ns) */
250#define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
251/* Target observation time interval for a peak-rate update (ns) */
252#define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
253
254/*
255 * Shift used for peak-rate fixed precision calculations.
256 * With
257 * - the current shift: 16 positions
258 * - the current type used to store rate: u32
259 * - the current unit of measure for rate: [sectors/usec], or, more precisely,
260 * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
261 * the range of rates that can be stored is
262 * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
263 * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
264 * [15, 65G] sectors/sec
265 * Which, assuming a sector size of 512B, corresponds to a range of
266 * [7.5K, 33T] B/sec
267 */
268#define BFQ_RATE_SHIFT 16
269
270/*
271 * When configured for computing the duration of the weight-raising
272 * for interactive queues automatically (see the comments at the
273 * beginning of this file), BFQ does it using the following formula:
274 * duration = (ref_rate / r) * ref_wr_duration,
275 * where r is the peak rate of the device, and ref_rate and
276 * ref_wr_duration are two reference parameters. In particular,
277 * ref_rate is the peak rate of the reference storage device (see
278 * below), and ref_wr_duration is about the maximum time needed, with
279 * BFQ and while reading two files in parallel, to load typical large
280 * applications on the reference device (see the comments on
281 * max_service_from_wr below, for more details on how ref_wr_duration
282 * is obtained). In practice, the slower/faster the device at hand
283 * is, the more/less it takes to load applications with respect to the
284 * reference device. Accordingly, the longer/shorter BFQ grants
285 * weight raising to interactive applications.
286 *
287 * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
288 * depending on whether the device is rotational or non-rotational.
289 *
290 * In the following definitions, ref_rate[0] and ref_wr_duration[0]
291 * are the reference values for a rotational device, whereas
292 * ref_rate[1] and ref_wr_duration[1] are the reference values for a
293 * non-rotational device. The reference rates are not the actual peak
294 * rates of the devices used as a reference, but slightly lower
295 * values. The reason for using slightly lower values is that the
296 * peak-rate estimator tends to yield slightly lower values than the
297 * actual peak rate (it can yield the actual peak rate only if there
298 * is only one process doing I/O, and the process does sequential
299 * I/O).
300 *
301 * The reference peak rates are measured in sectors/usec, left-shifted
302 * by BFQ_RATE_SHIFT.
303 */
304static int ref_rate[2] = {14000, 33000};
305/*
306 * To improve readability, a conversion function is used to initialize
307 * the following array, which entails that the array can be
308 * initialized only in a function.
309 */
310static int ref_wr_duration[2];
311
312/*
313 * BFQ uses the above-detailed, time-based weight-raising mechanism to
314 * privilege interactive tasks. This mechanism is vulnerable to the
315 * following false positives: I/O-bound applications that will go on
316 * doing I/O for much longer than the duration of weight
317 * raising. These applications have basically no benefit from being
318 * weight-raised at the beginning of their I/O. On the opposite end,
319 * while being weight-raised, these applications
320 * a) unjustly steal throughput to applications that may actually need
321 * low latency;
322 * b) make BFQ uselessly perform device idling; device idling results
323 * in loss of device throughput with most flash-based storage, and may
324 * increase latencies when used purposelessly.
325 *
326 * BFQ tries to reduce these problems, by adopting the following
327 * countermeasure. To introduce this countermeasure, we need first to
328 * finish explaining how the duration of weight-raising for
329 * interactive tasks is computed.
330 *
331 * For a bfq_queue deemed as interactive, the duration of weight
332 * raising is dynamically adjusted, as a function of the estimated
333 * peak rate of the device, so as to be equal to the time needed to
334 * execute the 'largest' interactive task we benchmarked so far. By
335 * largest task, we mean the task for which each involved process has
336 * to do more I/O than for any of the other tasks we benchmarked. This
337 * reference interactive task is the start-up of LibreOffice Writer,
338 * and in this task each process/bfq_queue needs to have at most ~110K
339 * sectors transferred.
340 *
341 * This last piece of information enables BFQ to reduce the actual
342 * duration of weight-raising for at least one class of I/O-bound
343 * applications: those doing sequential or quasi-sequential I/O. An
344 * example is file copy. In fact, once started, the main I/O-bound
345 * processes of these applications usually consume the above 110K
346 * sectors in much less time than the processes of an application that
347 * is starting, because these I/O-bound processes will greedily devote
348 * almost all their CPU cycles only to their target,
349 * throughput-friendly I/O operations. This is even more true if BFQ
350 * happens to be underestimating the device peak rate, and thus
351 * overestimating the duration of weight raising. But, according to
352 * our measurements, once transferred 110K sectors, these processes
353 * have no right to be weight-raised any longer.
354 *
355 * Basing on the last consideration, BFQ ends weight-raising for a
356 * bfq_queue if the latter happens to have received an amount of
357 * service at least equal to the following constant. The constant is
358 * set to slightly more than 110K, to have a minimum safety margin.
359 *
360 * This early ending of weight-raising reduces the amount of time
361 * during which interactive false positives cause the two problems
362 * described at the beginning of these comments.
363 */
364static const unsigned long max_service_from_wr = 120000;
365
366/*
367 * Maximum time between the creation of two queues, for stable merge
368 * to be activated (in ms)
369 */
370static const unsigned long bfq_activation_stable_merging = 600;
371/*
372 * Minimum time to be waited before evaluating delayed stable merge (in ms)
373 */
374static const unsigned long bfq_late_stable_merging = 600;
375
376#define RQ_BIC(rq) ((struct bfq_io_cq *)((rq)->elv.priv[0]))
377#define RQ_BFQQ(rq) ((rq)->elv.priv[1])
378
379struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync,
380 unsigned int actuator_idx)
381{
382 if (is_sync)
383 return bic->bfqq[1][actuator_idx];
384
385 return bic->bfqq[0][actuator_idx];
386}
387
388static void bfq_put_stable_ref(struct bfq_queue *bfqq);
389
390void bic_set_bfqq(struct bfq_io_cq *bic,
391 struct bfq_queue *bfqq,
392 bool is_sync,
393 unsigned int actuator_idx)
394{
395 struct bfq_queue *old_bfqq = bic->bfqq[is_sync][actuator_idx];
396
397 /*
398 * If bfqq != NULL, then a non-stable queue merge between
399 * bic->bfqq and bfqq is happening here. This causes troubles
400 * in the following case: bic->bfqq has also been scheduled
401 * for a possible stable merge with bic->stable_merge_bfqq,
402 * and bic->stable_merge_bfqq == bfqq happens to
403 * hold. Troubles occur because bfqq may then undergo a split,
404 * thereby becoming eligible for a stable merge. Yet, if
405 * bic->stable_merge_bfqq points exactly to bfqq, then bfqq
406 * would be stably merged with itself. To avoid this anomaly,
407 * we cancel the stable merge if
408 * bic->stable_merge_bfqq == bfqq.
409 */
410 struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[actuator_idx];
411
412 /* Clear bic pointer if bfqq is detached from this bic */
413 if (old_bfqq && old_bfqq->bic == bic)
414 old_bfqq->bic = NULL;
415
416 if (is_sync)
417 bic->bfqq[1][actuator_idx] = bfqq;
418 else
419 bic->bfqq[0][actuator_idx] = bfqq;
420
421 if (bfqq && bfqq_data->stable_merge_bfqq == bfqq) {
422 /*
423 * Actually, these same instructions are executed also
424 * in bfq_setup_cooperator, in case of abort or actual
425 * execution of a stable merge. We could avoid
426 * repeating these instructions there too, but if we
427 * did so, we would nest even more complexity in this
428 * function.
429 */
430 bfq_put_stable_ref(bfqq_data->stable_merge_bfqq);
431
432 bfqq_data->stable_merge_bfqq = NULL;
433 }
434}
435
436struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
437{
438 return bic->icq.q->elevator->elevator_data;
439}
440
441/**
442 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
443 * @icq: the iocontext queue.
444 */
445static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
446{
447 /* bic->icq is the first member, %NULL will convert to %NULL */
448 return container_of(icq, struct bfq_io_cq, icq);
449}
450
451/**
452 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
453 * @q: the request queue.
454 */
455static struct bfq_io_cq *bfq_bic_lookup(struct request_queue *q)
456{
457 struct bfq_io_cq *icq;
458 unsigned long flags;
459
460 if (!current->io_context)
461 return NULL;
462
463 spin_lock_irqsave(&q->queue_lock, flags);
464 icq = icq_to_bic(ioc_lookup_icq(q));
465 spin_unlock_irqrestore(&q->queue_lock, flags);
466
467 return icq;
468}
469
470/*
471 * Scheduler run of queue, if there are requests pending and no one in the
472 * driver that will restart queueing.
473 */
474void bfq_schedule_dispatch(struct bfq_data *bfqd)
475{
476 lockdep_assert_held(&bfqd->lock);
477
478 if (bfqd->queued != 0) {
479 bfq_log(bfqd, "schedule dispatch");
480 blk_mq_run_hw_queues(bfqd->queue, true);
481 }
482}
483
484#define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
485
486#define bfq_sample_valid(samples) ((samples) > 80)
487
488/*
489 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
490 * We choose the request that is closer to the head right now. Distance
491 * behind the head is penalized and only allowed to a certain extent.
492 */
493static struct request *bfq_choose_req(struct bfq_data *bfqd,
494 struct request *rq1,
495 struct request *rq2,
496 sector_t last)
497{
498 sector_t s1, s2, d1 = 0, d2 = 0;
499 unsigned long back_max;
500#define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
501#define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
502 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
503
504 if (!rq1 || rq1 == rq2)
505 return rq2;
506 if (!rq2)
507 return rq1;
508
509 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
510 return rq1;
511 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
512 return rq2;
513 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
514 return rq1;
515 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
516 return rq2;
517
518 s1 = blk_rq_pos(rq1);
519 s2 = blk_rq_pos(rq2);
520
521 /*
522 * By definition, 1KiB is 2 sectors.
523 */
524 back_max = bfqd->bfq_back_max * 2;
525
526 /*
527 * Strict one way elevator _except_ in the case where we allow
528 * short backward seeks which are biased as twice the cost of a
529 * similar forward seek.
530 */
531 if (s1 >= last)
532 d1 = s1 - last;
533 else if (s1 + back_max >= last)
534 d1 = (last - s1) * bfqd->bfq_back_penalty;
535 else
536 wrap |= BFQ_RQ1_WRAP;
537
538 if (s2 >= last)
539 d2 = s2 - last;
540 else if (s2 + back_max >= last)
541 d2 = (last - s2) * bfqd->bfq_back_penalty;
542 else
543 wrap |= BFQ_RQ2_WRAP;
544
545 /* Found required data */
546
547 /*
548 * By doing switch() on the bit mask "wrap" we avoid having to
549 * check two variables for all permutations: --> faster!
550 */
551 switch (wrap) {
552 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
553 if (d1 < d2)
554 return rq1;
555 else if (d2 < d1)
556 return rq2;
557
558 if (s1 >= s2)
559 return rq1;
560 else
561 return rq2;
562
563 case BFQ_RQ2_WRAP:
564 return rq1;
565 case BFQ_RQ1_WRAP:
566 return rq2;
567 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
568 default:
569 /*
570 * Since both rqs are wrapped,
571 * start with the one that's further behind head
572 * (--> only *one* back seek required),
573 * since back seek takes more time than forward.
574 */
575 if (s1 <= s2)
576 return rq1;
577 else
578 return rq2;
579 }
580}
581
582#define BFQ_LIMIT_INLINE_DEPTH 16
583
584#ifdef CONFIG_BFQ_GROUP_IOSCHED
585static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
586{
587 struct bfq_data *bfqd = bfqq->bfqd;
588 struct bfq_entity *entity = &bfqq->entity;
589 struct bfq_entity *inline_entities[BFQ_LIMIT_INLINE_DEPTH];
590 struct bfq_entity **entities = inline_entities;
591 int depth, level, alloc_depth = BFQ_LIMIT_INLINE_DEPTH;
592 int class_idx = bfqq->ioprio_class - 1;
593 struct bfq_sched_data *sched_data;
594 unsigned long wsum;
595 bool ret = false;
596
597 if (!entity->on_st_or_in_serv)
598 return false;
599
600retry:
601 spin_lock_irq(&bfqd->lock);
602 /* +1 for bfqq entity, root cgroup not included */
603 depth = bfqg_to_blkg(bfqq_group(bfqq))->blkcg->css.cgroup->level + 1;
604 if (depth > alloc_depth) {
605 spin_unlock_irq(&bfqd->lock);
606 if (entities != inline_entities)
607 kfree(entities);
608 entities = kmalloc_array(depth, sizeof(*entities), GFP_NOIO);
609 if (!entities)
610 return false;
611 alloc_depth = depth;
612 goto retry;
613 }
614
615 sched_data = entity->sched_data;
616 /* Gather our ancestors as we need to traverse them in reverse order */
617 level = 0;
618 for_each_entity(entity) {
619 /*
620 * If at some level entity is not even active, allow request
621 * queueing so that BFQ knows there's work to do and activate
622 * entities.
623 */
624 if (!entity->on_st_or_in_serv)
625 goto out;
626 /* Uh, more parents than cgroup subsystem thinks? */
627 if (WARN_ON_ONCE(level >= depth))
628 break;
629 entities[level++] = entity;
630 }
631 WARN_ON_ONCE(level != depth);
632 for (level--; level >= 0; level--) {
633 entity = entities[level];
634 if (level > 0) {
635 wsum = bfq_entity_service_tree(entity)->wsum;
636 } else {
637 int i;
638 /*
639 * For bfqq itself we take into account service trees
640 * of all higher priority classes and multiply their
641 * weights so that low prio queue from higher class
642 * gets more requests than high prio queue from lower
643 * class.
644 */
645 wsum = 0;
646 for (i = 0; i <= class_idx; i++) {
647 wsum = wsum * IOPRIO_BE_NR +
648 sched_data->service_tree[i].wsum;
649 }
650 }
651 if (!wsum)
652 continue;
653 limit = DIV_ROUND_CLOSEST(limit * entity->weight, wsum);
654 if (entity->allocated >= limit) {
655 bfq_log_bfqq(bfqq->bfqd, bfqq,
656 "too many requests: allocated %d limit %d level %d",
657 entity->allocated, limit, level);
658 ret = true;
659 break;
660 }
661 }
662out:
663 spin_unlock_irq(&bfqd->lock);
664 if (entities != inline_entities)
665 kfree(entities);
666 return ret;
667}
668#else
669static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
670{
671 return false;
672}
673#endif
674
675/*
676 * Async I/O can easily starve sync I/O (both sync reads and sync
677 * writes), by consuming all tags. Similarly, storms of sync writes,
678 * such as those that sync(2) may trigger, can starve sync reads.
679 * Limit depths of async I/O and sync writes so as to counter both
680 * problems.
681 *
682 * Also if a bfq queue or its parent cgroup consume more tags than would be
683 * appropriate for their weight, we trim the available tag depth to 1. This
684 * avoids a situation where one cgroup can starve another cgroup from tags and
685 * thus block service differentiation among cgroups. Note that because the
686 * queue / cgroup already has many requests allocated and queued, this does not
687 * significantly affect service guarantees coming from the BFQ scheduling
688 * algorithm.
689 */
690static void bfq_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
691{
692 struct bfq_data *bfqd = data->q->elevator->elevator_data;
693 struct bfq_io_cq *bic = bfq_bic_lookup(data->q);
694 int depth;
695 unsigned limit = data->q->nr_requests;
696 unsigned int act_idx;
697
698 /* Sync reads have full depth available */
699 if (op_is_sync(opf) && !op_is_write(opf)) {
700 depth = 0;
701 } else {
702 depth = bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(opf)];
703 limit = (limit * depth) >> bfqd->full_depth_shift;
704 }
705
706 for (act_idx = 0; bic && act_idx < bfqd->num_actuators; act_idx++) {
707 struct bfq_queue *bfqq =
708 bic_to_bfqq(bic, op_is_sync(opf), act_idx);
709
710 /*
711 * Does queue (or any parent entity) exceed number of
712 * requests that should be available to it? Heavily
713 * limit depth so that it cannot consume more
714 * available requests and thus starve other entities.
715 */
716 if (bfqq && bfqq_request_over_limit(bfqq, limit)) {
717 depth = 1;
718 break;
719 }
720 }
721 bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
722 __func__, bfqd->wr_busy_queues, op_is_sync(opf), depth);
723 if (depth)
724 data->shallow_depth = depth;
725}
726
727static struct bfq_queue *
728bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
729 sector_t sector, struct rb_node **ret_parent,
730 struct rb_node ***rb_link)
731{
732 struct rb_node **p, *parent;
733 struct bfq_queue *bfqq = NULL;
734
735 parent = NULL;
736 p = &root->rb_node;
737 while (*p) {
738 struct rb_node **n;
739
740 parent = *p;
741 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
742
743 /*
744 * Sort strictly based on sector. Smallest to the left,
745 * largest to the right.
746 */
747 if (sector > blk_rq_pos(bfqq->next_rq))
748 n = &(*p)->rb_right;
749 else if (sector < blk_rq_pos(bfqq->next_rq))
750 n = &(*p)->rb_left;
751 else
752 break;
753 p = n;
754 bfqq = NULL;
755 }
756
757 *ret_parent = parent;
758 if (rb_link)
759 *rb_link = p;
760
761 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
762 (unsigned long long)sector,
763 bfqq ? bfqq->pid : 0);
764
765 return bfqq;
766}
767
768static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
769{
770 return bfqq->service_from_backlogged > 0 &&
771 time_is_before_jiffies(bfqq->first_IO_time +
772 bfq_merge_time_limit);
773}
774
775/*
776 * The following function is not marked as __cold because it is
777 * actually cold, but for the same performance goal described in the
778 * comments on the likely() at the beginning of
779 * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
780 * execution time for the case where this function is not invoked, we
781 * had to add an unlikely() in each involved if().
782 */
783void __cold
784bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
785{
786 struct rb_node **p, *parent;
787 struct bfq_queue *__bfqq;
788
789 if (bfqq->pos_root) {
790 rb_erase(&bfqq->pos_node, bfqq->pos_root);
791 bfqq->pos_root = NULL;
792 }
793
794 /* oom_bfqq does not participate in queue merging */
795 if (bfqq == &bfqd->oom_bfqq)
796 return;
797
798 /*
799 * bfqq cannot be merged any longer (see comments in
800 * bfq_setup_cooperator): no point in adding bfqq into the
801 * position tree.
802 */
803 if (bfq_too_late_for_merging(bfqq))
804 return;
805
806 if (bfq_class_idle(bfqq))
807 return;
808 if (!bfqq->next_rq)
809 return;
810
811 bfqq->pos_root = &bfqq_group(bfqq)->rq_pos_tree;
812 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
813 blk_rq_pos(bfqq->next_rq), &parent, &p);
814 if (!__bfqq) {
815 rb_link_node(&bfqq->pos_node, parent, p);
816 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
817 } else
818 bfqq->pos_root = NULL;
819}
820
821/*
822 * The following function returns false either if every active queue
823 * must receive the same share of the throughput (symmetric scenario),
824 * or, as a special case, if bfqq must receive a share of the
825 * throughput lower than or equal to the share that every other active
826 * queue must receive. If bfqq does sync I/O, then these are the only
827 * two cases where bfqq happens to be guaranteed its share of the
828 * throughput even if I/O dispatching is not plugged when bfqq remains
829 * temporarily empty (for more details, see the comments in the
830 * function bfq_better_to_idle()). For this reason, the return value
831 * of this function is used to check whether I/O-dispatch plugging can
832 * be avoided.
833 *
834 * The above first case (symmetric scenario) occurs when:
835 * 1) all active queues have the same weight,
836 * 2) all active queues belong to the same I/O-priority class,
837 * 3) all active groups at the same level in the groups tree have the same
838 * weight,
839 * 4) all active groups at the same level in the groups tree have the same
840 * number of children.
841 *
842 * Unfortunately, keeping the necessary state for evaluating exactly
843 * the last two symmetry sub-conditions above would be quite complex
844 * and time consuming. Therefore this function evaluates, instead,
845 * only the following stronger three sub-conditions, for which it is
846 * much easier to maintain the needed state:
847 * 1) all active queues have the same weight,
848 * 2) all active queues belong to the same I/O-priority class,
849 * 3) there is at most one active group.
850 * In particular, the last condition is always true if hierarchical
851 * support or the cgroups interface are not enabled, thus no state
852 * needs to be maintained in this case.
853 */
854static bool bfq_asymmetric_scenario(struct bfq_data *bfqd,
855 struct bfq_queue *bfqq)
856{
857 bool smallest_weight = bfqq &&
858 bfqq->weight_counter &&
859 bfqq->weight_counter ==
860 container_of(
861 rb_first_cached(&bfqd->queue_weights_tree),
862 struct bfq_weight_counter,
863 weights_node);
864
865 /*
866 * For queue weights to differ, queue_weights_tree must contain
867 * at least two nodes.
868 */
869 bool varied_queue_weights = !smallest_weight &&
870 !RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) &&
871 (bfqd->queue_weights_tree.rb_root.rb_node->rb_left ||
872 bfqd->queue_weights_tree.rb_root.rb_node->rb_right);
873
874 bool multiple_classes_busy =
875 (bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
876 (bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
877 (bfqd->busy_queues[1] && bfqd->busy_queues[2]);
878
879 return varied_queue_weights || multiple_classes_busy
880#ifdef CONFIG_BFQ_GROUP_IOSCHED
881 || bfqd->num_groups_with_pending_reqs > 1
882#endif
883 ;
884}
885
886/*
887 * If the weight-counter tree passed as input contains no counter for
888 * the weight of the input queue, then add that counter; otherwise just
889 * increment the existing counter.
890 *
891 * Note that weight-counter trees contain few nodes in mostly symmetric
892 * scenarios. For example, if all queues have the same weight, then the
893 * weight-counter tree for the queues may contain at most one node.
894 * This holds even if low_latency is on, because weight-raised queues
895 * are not inserted in the tree.
896 * In most scenarios, the rate at which nodes are created/destroyed
897 * should be low too.
898 */
899void bfq_weights_tree_add(struct bfq_queue *bfqq)
900{
901 struct rb_root_cached *root = &bfqq->bfqd->queue_weights_tree;
902 struct bfq_entity *entity = &bfqq->entity;
903 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
904 bool leftmost = true;
905
906 /*
907 * Do not insert if the queue is already associated with a
908 * counter, which happens if:
909 * 1) a request arrival has caused the queue to become both
910 * non-weight-raised, and hence change its weight, and
911 * backlogged; in this respect, each of the two events
912 * causes an invocation of this function,
913 * 2) this is the invocation of this function caused by the
914 * second event. This second invocation is actually useless,
915 * and we handle this fact by exiting immediately. More
916 * efficient or clearer solutions might possibly be adopted.
917 */
918 if (bfqq->weight_counter)
919 return;
920
921 while (*new) {
922 struct bfq_weight_counter *__counter = container_of(*new,
923 struct bfq_weight_counter,
924 weights_node);
925 parent = *new;
926
927 if (entity->weight == __counter->weight) {
928 bfqq->weight_counter = __counter;
929 goto inc_counter;
930 }
931 if (entity->weight < __counter->weight)
932 new = &((*new)->rb_left);
933 else {
934 new = &((*new)->rb_right);
935 leftmost = false;
936 }
937 }
938
939 bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
940 GFP_ATOMIC);
941
942 /*
943 * In the unlucky event of an allocation failure, we just
944 * exit. This will cause the weight of queue to not be
945 * considered in bfq_asymmetric_scenario, which, in its turn,
946 * causes the scenario to be deemed wrongly symmetric in case
947 * bfqq's weight would have been the only weight making the
948 * scenario asymmetric. On the bright side, no unbalance will
949 * however occur when bfqq becomes inactive again (the
950 * invocation of this function is triggered by an activation
951 * of queue). In fact, bfq_weights_tree_remove does nothing
952 * if !bfqq->weight_counter.
953 */
954 if (unlikely(!bfqq->weight_counter))
955 return;
956
957 bfqq->weight_counter->weight = entity->weight;
958 rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
959 rb_insert_color_cached(&bfqq->weight_counter->weights_node, root,
960 leftmost);
961
962inc_counter:
963 bfqq->weight_counter->num_active++;
964 bfqq->ref++;
965}
966
967/*
968 * Decrement the weight counter associated with the queue, and, if the
969 * counter reaches 0, remove the counter from the tree.
970 * See the comments to the function bfq_weights_tree_add() for considerations
971 * about overhead.
972 */
973void bfq_weights_tree_remove(struct bfq_queue *bfqq)
974{
975 struct rb_root_cached *root;
976
977 if (!bfqq->weight_counter)
978 return;
979
980 root = &bfqq->bfqd->queue_weights_tree;
981 bfqq->weight_counter->num_active--;
982 if (bfqq->weight_counter->num_active > 0)
983 goto reset_entity_pointer;
984
985 rb_erase_cached(&bfqq->weight_counter->weights_node, root);
986 kfree(bfqq->weight_counter);
987
988reset_entity_pointer:
989 bfqq->weight_counter = NULL;
990 bfq_put_queue(bfqq);
991}
992
993/*
994 * Return expired entry, or NULL to just start from scratch in rbtree.
995 */
996static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
997 struct request *last)
998{
999 struct request *rq;
1000
1001 if (bfq_bfqq_fifo_expire(bfqq))
1002 return NULL;
1003
1004 bfq_mark_bfqq_fifo_expire(bfqq);
1005
1006 rq = rq_entry_fifo(bfqq->fifo.next);
1007
1008 if (rq == last || ktime_get_ns() < rq->fifo_time)
1009 return NULL;
1010
1011 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
1012 return rq;
1013}
1014
1015static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
1016 struct bfq_queue *bfqq,
1017 struct request *last)
1018{
1019 struct rb_node *rbnext = rb_next(&last->rb_node);
1020 struct rb_node *rbprev = rb_prev(&last->rb_node);
1021 struct request *next, *prev = NULL;
1022
1023 /* Follow expired path, else get first next available. */
1024 next = bfq_check_fifo(bfqq, last);
1025 if (next)
1026 return next;
1027
1028 if (rbprev)
1029 prev = rb_entry_rq(rbprev);
1030
1031 if (rbnext)
1032 next = rb_entry_rq(rbnext);
1033 else {
1034 rbnext = rb_first(&bfqq->sort_list);
1035 if (rbnext && rbnext != &last->rb_node)
1036 next = rb_entry_rq(rbnext);
1037 }
1038
1039 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
1040}
1041
1042/* see the definition of bfq_async_charge_factor for details */
1043static unsigned long bfq_serv_to_charge(struct request *rq,
1044 struct bfq_queue *bfqq)
1045{
1046 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
1047 bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
1048 return blk_rq_sectors(rq);
1049
1050 return blk_rq_sectors(rq) * bfq_async_charge_factor;
1051}
1052
1053/**
1054 * bfq_updated_next_req - update the queue after a new next_rq selection.
1055 * @bfqd: the device data the queue belongs to.
1056 * @bfqq: the queue to update.
1057 *
1058 * If the first request of a queue changes we make sure that the queue
1059 * has enough budget to serve at least its first request (if the
1060 * request has grown). We do this because if the queue has not enough
1061 * budget for its first request, it has to go through two dispatch
1062 * rounds to actually get it dispatched.
1063 */
1064static void bfq_updated_next_req(struct bfq_data *bfqd,
1065 struct bfq_queue *bfqq)
1066{
1067 struct bfq_entity *entity = &bfqq->entity;
1068 struct request *next_rq = bfqq->next_rq;
1069 unsigned long new_budget;
1070
1071 if (!next_rq)
1072 return;
1073
1074 if (bfqq == bfqd->in_service_queue)
1075 /*
1076 * In order not to break guarantees, budgets cannot be
1077 * changed after an entity has been selected.
1078 */
1079 return;
1080
1081 new_budget = max_t(unsigned long,
1082 max_t(unsigned long, bfqq->max_budget,
1083 bfq_serv_to_charge(next_rq, bfqq)),
1084 entity->service);
1085 if (entity->budget != new_budget) {
1086 entity->budget = new_budget;
1087 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
1088 new_budget);
1089 bfq_requeue_bfqq(bfqd, bfqq, false);
1090 }
1091}
1092
1093static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
1094{
1095 u64 dur;
1096
1097 dur = bfqd->rate_dur_prod;
1098 do_div(dur, bfqd->peak_rate);
1099
1100 /*
1101 * Limit duration between 3 and 25 seconds. The upper limit
1102 * has been conservatively set after the following worst case:
1103 * on a QEMU/KVM virtual machine
1104 * - running in a slow PC
1105 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
1106 * - serving a heavy I/O workload, such as the sequential reading
1107 * of several files
1108 * mplayer took 23 seconds to start, if constantly weight-raised.
1109 *
1110 * As for higher values than that accommodating the above bad
1111 * scenario, tests show that higher values would often yield
1112 * the opposite of the desired result, i.e., would worsen
1113 * responsiveness by allowing non-interactive applications to
1114 * preserve weight raising for too long.
1115 *
1116 * On the other end, lower values than 3 seconds make it
1117 * difficult for most interactive tasks to complete their jobs
1118 * before weight-raising finishes.
1119 */
1120 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
1121}
1122
1123/* switch back from soft real-time to interactive weight raising */
1124static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
1125 struct bfq_data *bfqd)
1126{
1127 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1128 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1129 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
1130}
1131
1132static void
1133bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
1134 struct bfq_io_cq *bic, bool bfq_already_existing)
1135{
1136 unsigned int old_wr_coeff = 1;
1137 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
1138 unsigned int a_idx = bfqq->actuator_idx;
1139 struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[a_idx];
1140
1141 if (bfqq_data->saved_has_short_ttime)
1142 bfq_mark_bfqq_has_short_ttime(bfqq);
1143 else
1144 bfq_clear_bfqq_has_short_ttime(bfqq);
1145
1146 if (bfqq_data->saved_IO_bound)
1147 bfq_mark_bfqq_IO_bound(bfqq);
1148 else
1149 bfq_clear_bfqq_IO_bound(bfqq);
1150
1151 bfqq->last_serv_time_ns = bfqq_data->saved_last_serv_time_ns;
1152 bfqq->inject_limit = bfqq_data->saved_inject_limit;
1153 bfqq->decrease_time_jif = bfqq_data->saved_decrease_time_jif;
1154
1155 bfqq->entity.new_weight = bfqq_data->saved_weight;
1156 bfqq->ttime = bfqq_data->saved_ttime;
1157 bfqq->io_start_time = bfqq_data->saved_io_start_time;
1158 bfqq->tot_idle_time = bfqq_data->saved_tot_idle_time;
1159 /*
1160 * Restore weight coefficient only if low_latency is on
1161 */
1162 if (bfqd->low_latency) {
1163 old_wr_coeff = bfqq->wr_coeff;
1164 bfqq->wr_coeff = bfqq_data->saved_wr_coeff;
1165 }
1166 bfqq->service_from_wr = bfqq_data->saved_service_from_wr;
1167 bfqq->wr_start_at_switch_to_srt =
1168 bfqq_data->saved_wr_start_at_switch_to_srt;
1169 bfqq->last_wr_start_finish = bfqq_data->saved_last_wr_start_finish;
1170 bfqq->wr_cur_max_time = bfqq_data->saved_wr_cur_max_time;
1171
1172 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1173 time_is_before_jiffies(bfqq->last_wr_start_finish +
1174 bfqq->wr_cur_max_time))) {
1175 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1176 !bfq_bfqq_in_large_burst(bfqq) &&
1177 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1178 bfq_wr_duration(bfqd))) {
1179 switch_back_to_interactive_wr(bfqq, bfqd);
1180 } else {
1181 bfqq->wr_coeff = 1;
1182 bfq_log_bfqq(bfqq->bfqd, bfqq,
1183 "resume state: switching off wr");
1184 }
1185 }
1186
1187 /* make sure weight will be updated, however we got here */
1188 bfqq->entity.prio_changed = 1;
1189
1190 if (likely(!busy))
1191 return;
1192
1193 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1194 bfqd->wr_busy_queues++;
1195 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1196 bfqd->wr_busy_queues--;
1197}
1198
1199static int bfqq_process_refs(struct bfq_queue *bfqq)
1200{
1201 return bfqq->ref - bfqq->entity.allocated -
1202 bfqq->entity.on_st_or_in_serv -
1203 (bfqq->weight_counter != NULL) - bfqq->stable_ref;
1204}
1205
1206/* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1207static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1208{
1209 struct bfq_queue *item;
1210 struct hlist_node *n;
1211
1212 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1213 hlist_del_init(&item->burst_list_node);
1214
1215 /*
1216 * Start the creation of a new burst list only if there is no
1217 * active queue. See comments on the conditional invocation of
1218 * bfq_handle_burst().
1219 */
1220 if (bfq_tot_busy_queues(bfqd) == 0) {
1221 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1222 bfqd->burst_size = 1;
1223 } else
1224 bfqd->burst_size = 0;
1225
1226 bfqd->burst_parent_entity = bfqq->entity.parent;
1227}
1228
1229/* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1230static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1231{
1232 /* Increment burst size to take into account also bfqq */
1233 bfqd->burst_size++;
1234
1235 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1236 struct bfq_queue *pos, *bfqq_item;
1237 struct hlist_node *n;
1238
1239 /*
1240 * Enough queues have been activated shortly after each
1241 * other to consider this burst as large.
1242 */
1243 bfqd->large_burst = true;
1244
1245 /*
1246 * We can now mark all queues in the burst list as
1247 * belonging to a large burst.
1248 */
1249 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1250 burst_list_node)
1251 bfq_mark_bfqq_in_large_burst(bfqq_item);
1252 bfq_mark_bfqq_in_large_burst(bfqq);
1253
1254 /*
1255 * From now on, and until the current burst finishes, any
1256 * new queue being activated shortly after the last queue
1257 * was inserted in the burst can be immediately marked as
1258 * belonging to a large burst. So the burst list is not
1259 * needed any more. Remove it.
1260 */
1261 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1262 burst_list_node)
1263 hlist_del_init(&pos->burst_list_node);
1264 } else /*
1265 * Burst not yet large: add bfqq to the burst list. Do
1266 * not increment the ref counter for bfqq, because bfqq
1267 * is removed from the burst list before freeing bfqq
1268 * in put_queue.
1269 */
1270 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1271}
1272
1273/*
1274 * If many queues belonging to the same group happen to be created
1275 * shortly after each other, then the processes associated with these
1276 * queues have typically a common goal. In particular, bursts of queue
1277 * creations are usually caused by services or applications that spawn
1278 * many parallel threads/processes. Examples are systemd during boot,
1279 * or git grep. To help these processes get their job done as soon as
1280 * possible, it is usually better to not grant either weight-raising
1281 * or device idling to their queues, unless these queues must be
1282 * protected from the I/O flowing through other active queues.
1283 *
1284 * In this comment we describe, firstly, the reasons why this fact
1285 * holds, and, secondly, the next function, which implements the main
1286 * steps needed to properly mark these queues so that they can then be
1287 * treated in a different way.
1288 *
1289 * The above services or applications benefit mostly from a high
1290 * throughput: the quicker the requests of the activated queues are
1291 * cumulatively served, the sooner the target job of these queues gets
1292 * completed. As a consequence, weight-raising any of these queues,
1293 * which also implies idling the device for it, is almost always
1294 * counterproductive, unless there are other active queues to isolate
1295 * these new queues from. If there no other active queues, then
1296 * weight-raising these new queues just lowers throughput in most
1297 * cases.
1298 *
1299 * On the other hand, a burst of queue creations may be caused also by
1300 * the start of an application that does not consist of a lot of
1301 * parallel I/O-bound threads. In fact, with a complex application,
1302 * several short processes may need to be executed to start-up the
1303 * application. In this respect, to start an application as quickly as
1304 * possible, the best thing to do is in any case to privilege the I/O
1305 * related to the application with respect to all other
1306 * I/O. Therefore, the best strategy to start as quickly as possible
1307 * an application that causes a burst of queue creations is to
1308 * weight-raise all the queues created during the burst. This is the
1309 * exact opposite of the best strategy for the other type of bursts.
1310 *
1311 * In the end, to take the best action for each of the two cases, the
1312 * two types of bursts need to be distinguished. Fortunately, this
1313 * seems relatively easy, by looking at the sizes of the bursts. In
1314 * particular, we found a threshold such that only bursts with a
1315 * larger size than that threshold are apparently caused by
1316 * services or commands such as systemd or git grep. For brevity,
1317 * hereafter we call just 'large' these bursts. BFQ *does not*
1318 * weight-raise queues whose creation occurs in a large burst. In
1319 * addition, for each of these queues BFQ performs or does not perform
1320 * idling depending on which choice boosts the throughput more. The
1321 * exact choice depends on the device and request pattern at
1322 * hand.
1323 *
1324 * Unfortunately, false positives may occur while an interactive task
1325 * is starting (e.g., an application is being started). The
1326 * consequence is that the queues associated with the task do not
1327 * enjoy weight raising as expected. Fortunately these false positives
1328 * are very rare. They typically occur if some service happens to
1329 * start doing I/O exactly when the interactive task starts.
1330 *
1331 * Turning back to the next function, it is invoked only if there are
1332 * no active queues (apart from active queues that would belong to the
1333 * same, possible burst bfqq would belong to), and it implements all
1334 * the steps needed to detect the occurrence of a large burst and to
1335 * properly mark all the queues belonging to it (so that they can then
1336 * be treated in a different way). This goal is achieved by
1337 * maintaining a "burst list" that holds, temporarily, the queues that
1338 * belong to the burst in progress. The list is then used to mark
1339 * these queues as belonging to a large burst if the burst does become
1340 * large. The main steps are the following.
1341 *
1342 * . when the very first queue is created, the queue is inserted into the
1343 * list (as it could be the first queue in a possible burst)
1344 *
1345 * . if the current burst has not yet become large, and a queue Q that does
1346 * not yet belong to the burst is activated shortly after the last time
1347 * at which a new queue entered the burst list, then the function appends
1348 * Q to the burst list
1349 *
1350 * . if, as a consequence of the previous step, the burst size reaches
1351 * the large-burst threshold, then
1352 *
1353 * . all the queues in the burst list are marked as belonging to a
1354 * large burst
1355 *
1356 * . the burst list is deleted; in fact, the burst list already served
1357 * its purpose (keeping temporarily track of the queues in a burst,
1358 * so as to be able to mark them as belonging to a large burst in the
1359 * previous sub-step), and now is not needed any more
1360 *
1361 * . the device enters a large-burst mode
1362 *
1363 * . if a queue Q that does not belong to the burst is created while
1364 * the device is in large-burst mode and shortly after the last time
1365 * at which a queue either entered the burst list or was marked as
1366 * belonging to the current large burst, then Q is immediately marked
1367 * as belonging to a large burst.
1368 *
1369 * . if a queue Q that does not belong to the burst is created a while
1370 * later, i.e., not shortly after, than the last time at which a queue
1371 * either entered the burst list or was marked as belonging to the
1372 * current large burst, then the current burst is deemed as finished and:
1373 *
1374 * . the large-burst mode is reset if set
1375 *
1376 * . the burst list is emptied
1377 *
1378 * . Q is inserted in the burst list, as Q may be the first queue
1379 * in a possible new burst (then the burst list contains just Q
1380 * after this step).
1381 */
1382static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1383{
1384 /*
1385 * If bfqq is already in the burst list or is part of a large
1386 * burst, or finally has just been split, then there is
1387 * nothing else to do.
1388 */
1389 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1390 bfq_bfqq_in_large_burst(bfqq) ||
1391 time_is_after_eq_jiffies(bfqq->split_time +
1392 msecs_to_jiffies(10)))
1393 return;
1394
1395 /*
1396 * If bfqq's creation happens late enough, or bfqq belongs to
1397 * a different group than the burst group, then the current
1398 * burst is finished, and related data structures must be
1399 * reset.
1400 *
1401 * In this respect, consider the special case where bfqq is
1402 * the very first queue created after BFQ is selected for this
1403 * device. In this case, last_ins_in_burst and
1404 * burst_parent_entity are not yet significant when we get
1405 * here. But it is easy to verify that, whether or not the
1406 * following condition is true, bfqq will end up being
1407 * inserted into the burst list. In particular the list will
1408 * happen to contain only bfqq. And this is exactly what has
1409 * to happen, as bfqq may be the first queue of the first
1410 * burst.
1411 */
1412 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1413 bfqd->bfq_burst_interval) ||
1414 bfqq->entity.parent != bfqd->burst_parent_entity) {
1415 bfqd->large_burst = false;
1416 bfq_reset_burst_list(bfqd, bfqq);
1417 goto end;
1418 }
1419
1420 /*
1421 * If we get here, then bfqq is being activated shortly after the
1422 * last queue. So, if the current burst is also large, we can mark
1423 * bfqq as belonging to this large burst immediately.
1424 */
1425 if (bfqd->large_burst) {
1426 bfq_mark_bfqq_in_large_burst(bfqq);
1427 goto end;
1428 }
1429
1430 /*
1431 * If we get here, then a large-burst state has not yet been
1432 * reached, but bfqq is being activated shortly after the last
1433 * queue. Then we add bfqq to the burst.
1434 */
1435 bfq_add_to_burst(bfqd, bfqq);
1436end:
1437 /*
1438 * At this point, bfqq either has been added to the current
1439 * burst or has caused the current burst to terminate and a
1440 * possible new burst to start. In particular, in the second
1441 * case, bfqq has become the first queue in the possible new
1442 * burst. In both cases last_ins_in_burst needs to be moved
1443 * forward.
1444 */
1445 bfqd->last_ins_in_burst = jiffies;
1446}
1447
1448static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1449{
1450 struct bfq_entity *entity = &bfqq->entity;
1451
1452 return entity->budget - entity->service;
1453}
1454
1455/*
1456 * If enough samples have been computed, return the current max budget
1457 * stored in bfqd, which is dynamically updated according to the
1458 * estimated disk peak rate; otherwise return the default max budget
1459 */
1460static int bfq_max_budget(struct bfq_data *bfqd)
1461{
1462 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1463 return bfq_default_max_budget;
1464 else
1465 return bfqd->bfq_max_budget;
1466}
1467
1468/*
1469 * Return min budget, which is a fraction of the current or default
1470 * max budget (trying with 1/32)
1471 */
1472static int bfq_min_budget(struct bfq_data *bfqd)
1473{
1474 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1475 return bfq_default_max_budget / 32;
1476 else
1477 return bfqd->bfq_max_budget / 32;
1478}
1479
1480/*
1481 * The next function, invoked after the input queue bfqq switches from
1482 * idle to busy, updates the budget of bfqq. The function also tells
1483 * whether the in-service queue should be expired, by returning
1484 * true. The purpose of expiring the in-service queue is to give bfqq
1485 * the chance to possibly preempt the in-service queue, and the reason
1486 * for preempting the in-service queue is to achieve one of the two
1487 * goals below.
1488 *
1489 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1490 * expired because it has remained idle. In particular, bfqq may have
1491 * expired for one of the following two reasons:
1492 *
1493 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1494 * and did not make it to issue a new request before its last
1495 * request was served;
1496 *
1497 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1498 * a new request before the expiration of the idling-time.
1499 *
1500 * Even if bfqq has expired for one of the above reasons, the process
1501 * associated with the queue may be however issuing requests greedily,
1502 * and thus be sensitive to the bandwidth it receives (bfqq may have
1503 * remained idle for other reasons: CPU high load, bfqq not enjoying
1504 * idling, I/O throttling somewhere in the path from the process to
1505 * the I/O scheduler, ...). But if, after every expiration for one of
1506 * the above two reasons, bfqq has to wait for the service of at least
1507 * one full budget of another queue before being served again, then
1508 * bfqq is likely to get a much lower bandwidth or resource time than
1509 * its reserved ones. To address this issue, two countermeasures need
1510 * to be taken.
1511 *
1512 * First, the budget and the timestamps of bfqq need to be updated in
1513 * a special way on bfqq reactivation: they need to be updated as if
1514 * bfqq did not remain idle and did not expire. In fact, if they are
1515 * computed as if bfqq expired and remained idle until reactivation,
1516 * then the process associated with bfqq is treated as if, instead of
1517 * being greedy, it stopped issuing requests when bfqq remained idle,
1518 * and restarts issuing requests only on this reactivation. In other
1519 * words, the scheduler does not help the process recover the "service
1520 * hole" between bfqq expiration and reactivation. As a consequence,
1521 * the process receives a lower bandwidth than its reserved one. In
1522 * contrast, to recover this hole, the budget must be updated as if
1523 * bfqq was not expired at all before this reactivation, i.e., it must
1524 * be set to the value of the remaining budget when bfqq was
1525 * expired. Along the same line, timestamps need to be assigned the
1526 * value they had the last time bfqq was selected for service, i.e.,
1527 * before last expiration. Thus timestamps need to be back-shifted
1528 * with respect to their normal computation (see [1] for more details
1529 * on this tricky aspect).
1530 *
1531 * Secondly, to allow the process to recover the hole, the in-service
1532 * queue must be expired too, to give bfqq the chance to preempt it
1533 * immediately. In fact, if bfqq has to wait for a full budget of the
1534 * in-service queue to be completed, then it may become impossible to
1535 * let the process recover the hole, even if the back-shifted
1536 * timestamps of bfqq are lower than those of the in-service queue. If
1537 * this happens for most or all of the holes, then the process may not
1538 * receive its reserved bandwidth. In this respect, it is worth noting
1539 * that, being the service of outstanding requests unpreemptible, a
1540 * little fraction of the holes may however be unrecoverable, thereby
1541 * causing a little loss of bandwidth.
1542 *
1543 * The last important point is detecting whether bfqq does need this
1544 * bandwidth recovery. In this respect, the next function deems the
1545 * process associated with bfqq greedy, and thus allows it to recover
1546 * the hole, if: 1) the process is waiting for the arrival of a new
1547 * request (which implies that bfqq expired for one of the above two
1548 * reasons), and 2) such a request has arrived soon. The first
1549 * condition is controlled through the flag non_blocking_wait_rq,
1550 * while the second through the flag arrived_in_time. If both
1551 * conditions hold, then the function computes the budget in the
1552 * above-described special way, and signals that the in-service queue
1553 * should be expired. Timestamp back-shifting is done later in
1554 * __bfq_activate_entity.
1555 *
1556 * 2. Reduce latency. Even if timestamps are not backshifted to let
1557 * the process associated with bfqq recover a service hole, bfqq may
1558 * however happen to have, after being (re)activated, a lower finish
1559 * timestamp than the in-service queue. That is, the next budget of
1560 * bfqq may have to be completed before the one of the in-service
1561 * queue. If this is the case, then preempting the in-service queue
1562 * allows this goal to be achieved, apart from the unpreemptible,
1563 * outstanding requests mentioned above.
1564 *
1565 * Unfortunately, regardless of which of the above two goals one wants
1566 * to achieve, service trees need first to be updated to know whether
1567 * the in-service queue must be preempted. To have service trees
1568 * correctly updated, the in-service queue must be expired and
1569 * rescheduled, and bfqq must be scheduled too. This is one of the
1570 * most costly operations (in future versions, the scheduling
1571 * mechanism may be re-designed in such a way to make it possible to
1572 * know whether preemption is needed without needing to update service
1573 * trees). In addition, queue preemptions almost always cause random
1574 * I/O, which may in turn cause loss of throughput. Finally, there may
1575 * even be no in-service queue when the next function is invoked (so,
1576 * no queue to compare timestamps with). Because of these facts, the
1577 * next function adopts the following simple scheme to avoid costly
1578 * operations, too frequent preemptions and too many dependencies on
1579 * the state of the scheduler: it requests the expiration of the
1580 * in-service queue (unconditionally) only for queues that need to
1581 * recover a hole. Then it delegates to other parts of the code the
1582 * responsibility of handling the above case 2.
1583 */
1584static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1585 struct bfq_queue *bfqq,
1586 bool arrived_in_time)
1587{
1588 struct bfq_entity *entity = &bfqq->entity;
1589
1590 /*
1591 * In the next compound condition, we check also whether there
1592 * is some budget left, because otherwise there is no point in
1593 * trying to go on serving bfqq with this same budget: bfqq
1594 * would be expired immediately after being selected for
1595 * service. This would only cause useless overhead.
1596 */
1597 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1598 bfq_bfqq_budget_left(bfqq) > 0) {
1599 /*
1600 * We do not clear the flag non_blocking_wait_rq here, as
1601 * the latter is used in bfq_activate_bfqq to signal
1602 * that timestamps need to be back-shifted (and is
1603 * cleared right after).
1604 */
1605
1606 /*
1607 * In next assignment we rely on that either
1608 * entity->service or entity->budget are not updated
1609 * on expiration if bfqq is empty (see
1610 * __bfq_bfqq_recalc_budget). Thus both quantities
1611 * remain unchanged after such an expiration, and the
1612 * following statement therefore assigns to
1613 * entity->budget the remaining budget on such an
1614 * expiration.
1615 */
1616 entity->budget = min_t(unsigned long,
1617 bfq_bfqq_budget_left(bfqq),
1618 bfqq->max_budget);
1619
1620 /*
1621 * At this point, we have used entity->service to get
1622 * the budget left (needed for updating
1623 * entity->budget). Thus we finally can, and have to,
1624 * reset entity->service. The latter must be reset
1625 * because bfqq would otherwise be charged again for
1626 * the service it has received during its previous
1627 * service slot(s).
1628 */
1629 entity->service = 0;
1630
1631 return true;
1632 }
1633
1634 /*
1635 * We can finally complete expiration, by setting service to 0.
1636 */
1637 entity->service = 0;
1638 entity->budget = max_t(unsigned long, bfqq->max_budget,
1639 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1640 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1641 return false;
1642}
1643
1644/*
1645 * Return the farthest past time instant according to jiffies
1646 * macros.
1647 */
1648static unsigned long bfq_smallest_from_now(void)
1649{
1650 return jiffies - MAX_JIFFY_OFFSET;
1651}
1652
1653static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1654 struct bfq_queue *bfqq,
1655 unsigned int old_wr_coeff,
1656 bool wr_or_deserves_wr,
1657 bool interactive,
1658 bool in_burst,
1659 bool soft_rt)
1660{
1661 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1662 /* start a weight-raising period */
1663 if (interactive) {
1664 bfqq->service_from_wr = 0;
1665 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1666 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1667 } else {
1668 /*
1669 * No interactive weight raising in progress
1670 * here: assign minus infinity to
1671 * wr_start_at_switch_to_srt, to make sure
1672 * that, at the end of the soft-real-time
1673 * weight raising periods that is starting
1674 * now, no interactive weight-raising period
1675 * may be wrongly considered as still in
1676 * progress (and thus actually started by
1677 * mistake).
1678 */
1679 bfqq->wr_start_at_switch_to_srt =
1680 bfq_smallest_from_now();
1681 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1682 BFQ_SOFTRT_WEIGHT_FACTOR;
1683 bfqq->wr_cur_max_time =
1684 bfqd->bfq_wr_rt_max_time;
1685 }
1686
1687 /*
1688 * If needed, further reduce budget to make sure it is
1689 * close to bfqq's backlog, so as to reduce the
1690 * scheduling-error component due to a too large
1691 * budget. Do not care about throughput consequences,
1692 * but only about latency. Finally, do not assign a
1693 * too small budget either, to avoid increasing
1694 * latency by causing too frequent expirations.
1695 */
1696 bfqq->entity.budget = min_t(unsigned long,
1697 bfqq->entity.budget,
1698 2 * bfq_min_budget(bfqd));
1699 } else if (old_wr_coeff > 1) {
1700 if (interactive) { /* update wr coeff and duration */
1701 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1702 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1703 } else if (in_burst)
1704 bfqq->wr_coeff = 1;
1705 else if (soft_rt) {
1706 /*
1707 * The application is now or still meeting the
1708 * requirements for being deemed soft rt. We
1709 * can then correctly and safely (re)charge
1710 * the weight-raising duration for the
1711 * application with the weight-raising
1712 * duration for soft rt applications.
1713 *
1714 * In particular, doing this recharge now, i.e.,
1715 * before the weight-raising period for the
1716 * application finishes, reduces the probability
1717 * of the following negative scenario:
1718 * 1) the weight of a soft rt application is
1719 * raised at startup (as for any newly
1720 * created application),
1721 * 2) since the application is not interactive,
1722 * at a certain time weight-raising is
1723 * stopped for the application,
1724 * 3) at that time the application happens to
1725 * still have pending requests, and hence
1726 * is destined to not have a chance to be
1727 * deemed soft rt before these requests are
1728 * completed (see the comments to the
1729 * function bfq_bfqq_softrt_next_start()
1730 * for details on soft rt detection),
1731 * 4) these pending requests experience a high
1732 * latency because the application is not
1733 * weight-raised while they are pending.
1734 */
1735 if (bfqq->wr_cur_max_time !=
1736 bfqd->bfq_wr_rt_max_time) {
1737 bfqq->wr_start_at_switch_to_srt =
1738 bfqq->last_wr_start_finish;
1739
1740 bfqq->wr_cur_max_time =
1741 bfqd->bfq_wr_rt_max_time;
1742 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1743 BFQ_SOFTRT_WEIGHT_FACTOR;
1744 }
1745 bfqq->last_wr_start_finish = jiffies;
1746 }
1747 }
1748}
1749
1750static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1751 struct bfq_queue *bfqq)
1752{
1753 return bfqq->dispatched == 0 &&
1754 time_is_before_jiffies(
1755 bfqq->budget_timeout +
1756 bfqd->bfq_wr_min_idle_time);
1757}
1758
1759
1760/*
1761 * Return true if bfqq is in a higher priority class, or has a higher
1762 * weight than the in-service queue.
1763 */
1764static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq,
1765 struct bfq_queue *in_serv_bfqq)
1766{
1767 int bfqq_weight, in_serv_weight;
1768
1769 if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class)
1770 return true;
1771
1772 if (in_serv_bfqq->entity.parent == bfqq->entity.parent) {
1773 bfqq_weight = bfqq->entity.weight;
1774 in_serv_weight = in_serv_bfqq->entity.weight;
1775 } else {
1776 if (bfqq->entity.parent)
1777 bfqq_weight = bfqq->entity.parent->weight;
1778 else
1779 bfqq_weight = bfqq->entity.weight;
1780 if (in_serv_bfqq->entity.parent)
1781 in_serv_weight = in_serv_bfqq->entity.parent->weight;
1782 else
1783 in_serv_weight = in_serv_bfqq->entity.weight;
1784 }
1785
1786 return bfqq_weight > in_serv_weight;
1787}
1788
1789/*
1790 * Get the index of the actuator that will serve bio.
1791 */
1792static unsigned int bfq_actuator_index(struct bfq_data *bfqd, struct bio *bio)
1793{
1794 unsigned int i;
1795 sector_t end;
1796
1797 /* no search needed if one or zero ranges present */
1798 if (bfqd->num_actuators == 1)
1799 return 0;
1800
1801 /* bio_end_sector(bio) gives the sector after the last one */
1802 end = bio_end_sector(bio) - 1;
1803
1804 for (i = 0; i < bfqd->num_actuators; i++) {
1805 if (end >= bfqd->sector[i] &&
1806 end < bfqd->sector[i] + bfqd->nr_sectors[i])
1807 return i;
1808 }
1809
1810 WARN_ONCE(true,
1811 "bfq_actuator_index: bio sector out of ranges: end=%llu\n",
1812 end);
1813 return 0;
1814}
1815
1816static bool bfq_better_to_idle(struct bfq_queue *bfqq);
1817
1818static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1819 struct bfq_queue *bfqq,
1820 int old_wr_coeff,
1821 struct request *rq,
1822 bool *interactive)
1823{
1824 bool soft_rt, in_burst, wr_or_deserves_wr,
1825 bfqq_wants_to_preempt,
1826 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1827 /*
1828 * See the comments on
1829 * bfq_bfqq_update_budg_for_activation for
1830 * details on the usage of the next variable.
1831 */
1832 arrived_in_time = ktime_get_ns() <=
1833 bfqq->ttime.last_end_request +
1834 bfqd->bfq_slice_idle * 3;
1835 unsigned int act_idx = bfq_actuator_index(bfqd, rq->bio);
1836 bool bfqq_non_merged_or_stably_merged =
1837 bfqq->bic || RQ_BIC(rq)->bfqq_data[act_idx].stably_merged;
1838
1839 /*
1840 * bfqq deserves to be weight-raised if:
1841 * - it is sync,
1842 * - it does not belong to a large burst,
1843 * - it has been idle for enough time or is soft real-time,
1844 * - is linked to a bfq_io_cq (it is not shared in any sense),
1845 * - has a default weight (otherwise we assume the user wanted
1846 * to control its weight explicitly)
1847 */
1848 in_burst = bfq_bfqq_in_large_burst(bfqq);
1849 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1850 !BFQQ_TOTALLY_SEEKY(bfqq) &&
1851 !in_burst &&
1852 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1853 bfqq->dispatched == 0 &&
1854 bfqq->entity.new_weight == 40;
1855 *interactive = !in_burst && idle_for_long_time &&
1856 bfqq->entity.new_weight == 40;
1857 /*
1858 * Merged bfq_queues are kept out of weight-raising
1859 * (low-latency) mechanisms. The reason is that these queues
1860 * are usually created for non-interactive and
1861 * non-soft-real-time tasks. Yet this is not the case for
1862 * stably-merged queues. These queues are merged just because
1863 * they are created shortly after each other. So they may
1864 * easily serve the I/O of an interactive or soft-real time
1865 * application, if the application happens to spawn multiple
1866 * processes. So let also stably-merged queued enjoy weight
1867 * raising.
1868 */
1869 wr_or_deserves_wr = bfqd->low_latency &&
1870 (bfqq->wr_coeff > 1 ||
1871 (bfq_bfqq_sync(bfqq) && bfqq_non_merged_or_stably_merged &&
1872 (*interactive || soft_rt)));
1873
1874 /*
1875 * Using the last flag, update budget and check whether bfqq
1876 * may want to preempt the in-service queue.
1877 */
1878 bfqq_wants_to_preempt =
1879 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1880 arrived_in_time);
1881
1882 /*
1883 * If bfqq happened to be activated in a burst, but has been
1884 * idle for much more than an interactive queue, then we
1885 * assume that, in the overall I/O initiated in the burst, the
1886 * I/O associated with bfqq is finished. So bfqq does not need
1887 * to be treated as a queue belonging to a burst
1888 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1889 * if set, and remove bfqq from the burst list if it's
1890 * there. We do not decrement burst_size, because the fact
1891 * that bfqq does not need to belong to the burst list any
1892 * more does not invalidate the fact that bfqq was created in
1893 * a burst.
1894 */
1895 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1896 idle_for_long_time &&
1897 time_is_before_jiffies(
1898 bfqq->budget_timeout +
1899 msecs_to_jiffies(10000))) {
1900 hlist_del_init(&bfqq->burst_list_node);
1901 bfq_clear_bfqq_in_large_burst(bfqq);
1902 }
1903
1904 bfq_clear_bfqq_just_created(bfqq);
1905
1906 if (bfqd->low_latency) {
1907 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1908 /* wraparound */
1909 bfqq->split_time =
1910 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1911
1912 if (time_is_before_jiffies(bfqq->split_time +
1913 bfqd->bfq_wr_min_idle_time)) {
1914 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1915 old_wr_coeff,
1916 wr_or_deserves_wr,
1917 *interactive,
1918 in_burst,
1919 soft_rt);
1920
1921 if (old_wr_coeff != bfqq->wr_coeff)
1922 bfqq->entity.prio_changed = 1;
1923 }
1924 }
1925
1926 bfqq->last_idle_bklogged = jiffies;
1927 bfqq->service_from_backlogged = 0;
1928 bfq_clear_bfqq_softrt_update(bfqq);
1929
1930 bfq_add_bfqq_busy(bfqq);
1931
1932 /*
1933 * Expire in-service queue if preemption may be needed for
1934 * guarantees or throughput. As for guarantees, we care
1935 * explicitly about two cases. The first is that bfqq has to
1936 * recover a service hole, as explained in the comments on
1937 * bfq_bfqq_update_budg_for_activation(), i.e., that
1938 * bfqq_wants_to_preempt is true. However, if bfqq does not
1939 * carry time-critical I/O, then bfqq's bandwidth is less
1940 * important than that of queues that carry time-critical I/O.
1941 * So, as a further constraint, we consider this case only if
1942 * bfqq is at least as weight-raised, i.e., at least as time
1943 * critical, as the in-service queue.
1944 *
1945 * The second case is that bfqq is in a higher priority class,
1946 * or has a higher weight than the in-service queue. If this
1947 * condition does not hold, we don't care because, even if
1948 * bfqq does not start to be served immediately, the resulting
1949 * delay for bfqq's I/O is however lower or much lower than
1950 * the ideal completion time to be guaranteed to bfqq's I/O.
1951 *
1952 * In both cases, preemption is needed only if, according to
1953 * the timestamps of both bfqq and of the in-service queue,
1954 * bfqq actually is the next queue to serve. So, to reduce
1955 * useless preemptions, the return value of
1956 * next_queue_may_preempt() is considered in the next compound
1957 * condition too. Yet next_queue_may_preempt() just checks a
1958 * simple, necessary condition for bfqq to be the next queue
1959 * to serve. In fact, to evaluate a sufficient condition, the
1960 * timestamps of the in-service queue would need to be
1961 * updated, and this operation is quite costly (see the
1962 * comments on bfq_bfqq_update_budg_for_activation()).
1963 *
1964 * As for throughput, we ask bfq_better_to_idle() whether we
1965 * still need to plug I/O dispatching. If bfq_better_to_idle()
1966 * says no, then plugging is not needed any longer, either to
1967 * boost throughput or to perserve service guarantees. Then
1968 * the best option is to stop plugging I/O, as not doing so
1969 * would certainly lower throughput. We may end up in this
1970 * case if: (1) upon a dispatch attempt, we detected that it
1971 * was better to plug I/O dispatch, and to wait for a new
1972 * request to arrive for the currently in-service queue, but
1973 * (2) this switch of bfqq to busy changes the scenario.
1974 */
1975 if (bfqd->in_service_queue &&
1976 ((bfqq_wants_to_preempt &&
1977 bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) ||
1978 bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue) ||
1979 !bfq_better_to_idle(bfqd->in_service_queue)) &&
1980 next_queue_may_preempt(bfqd))
1981 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1982 false, BFQQE_PREEMPTED);
1983}
1984
1985static void bfq_reset_inject_limit(struct bfq_data *bfqd,
1986 struct bfq_queue *bfqq)
1987{
1988 /* invalidate baseline total service time */
1989 bfqq->last_serv_time_ns = 0;
1990
1991 /*
1992 * Reset pointer in case we are waiting for
1993 * some request completion.
1994 */
1995 bfqd->waited_rq = NULL;
1996
1997 /*
1998 * If bfqq has a short think time, then start by setting the
1999 * inject limit to 0 prudentially, because the service time of
2000 * an injected I/O request may be higher than the think time
2001 * of bfqq, and therefore, if one request was injected when
2002 * bfqq remains empty, this injected request might delay the
2003 * service of the next I/O request for bfqq significantly. In
2004 * case bfqq can actually tolerate some injection, then the
2005 * adaptive update will however raise the limit soon. This
2006 * lucky circumstance holds exactly because bfqq has a short
2007 * think time, and thus, after remaining empty, is likely to
2008 * get new I/O enqueued---and then completed---before being
2009 * expired. This is the very pattern that gives the
2010 * limit-update algorithm the chance to measure the effect of
2011 * injection on request service times, and then to update the
2012 * limit accordingly.
2013 *
2014 * However, in the following special case, the inject limit is
2015 * left to 1 even if the think time is short: bfqq's I/O is
2016 * synchronized with that of some other queue, i.e., bfqq may
2017 * receive new I/O only after the I/O of the other queue is
2018 * completed. Keeping the inject limit to 1 allows the
2019 * blocking I/O to be served while bfqq is in service. And
2020 * this is very convenient both for bfqq and for overall
2021 * throughput, as explained in detail in the comments in
2022 * bfq_update_has_short_ttime().
2023 *
2024 * On the opposite end, if bfqq has a long think time, then
2025 * start directly by 1, because:
2026 * a) on the bright side, keeping at most one request in
2027 * service in the drive is unlikely to cause any harm to the
2028 * latency of bfqq's requests, as the service time of a single
2029 * request is likely to be lower than the think time of bfqq;
2030 * b) on the downside, after becoming empty, bfqq is likely to
2031 * expire before getting its next request. With this request
2032 * arrival pattern, it is very hard to sample total service
2033 * times and update the inject limit accordingly (see comments
2034 * on bfq_update_inject_limit()). So the limit is likely to be
2035 * never, or at least seldom, updated. As a consequence, by
2036 * setting the limit to 1, we avoid that no injection ever
2037 * occurs with bfqq. On the downside, this proactive step
2038 * further reduces chances to actually compute the baseline
2039 * total service time. Thus it reduces chances to execute the
2040 * limit-update algorithm and possibly raise the limit to more
2041 * than 1.
2042 */
2043 if (bfq_bfqq_has_short_ttime(bfqq))
2044 bfqq->inject_limit = 0;
2045 else
2046 bfqq->inject_limit = 1;
2047
2048 bfqq->decrease_time_jif = jiffies;
2049}
2050
2051static void bfq_update_io_intensity(struct bfq_queue *bfqq, u64 now_ns)
2052{
2053 u64 tot_io_time = now_ns - bfqq->io_start_time;
2054
2055 if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfqq->dispatched == 0)
2056 bfqq->tot_idle_time +=
2057 now_ns - bfqq->ttime.last_end_request;
2058
2059 if (unlikely(bfq_bfqq_just_created(bfqq)))
2060 return;
2061
2062 /*
2063 * Must be busy for at least about 80% of the time to be
2064 * considered I/O bound.
2065 */
2066 if (bfqq->tot_idle_time * 5 > tot_io_time)
2067 bfq_clear_bfqq_IO_bound(bfqq);
2068 else
2069 bfq_mark_bfqq_IO_bound(bfqq);
2070
2071 /*
2072 * Keep an observation window of at most 200 ms in the past
2073 * from now.
2074 */
2075 if (tot_io_time > 200 * NSEC_PER_MSEC) {
2076 bfqq->io_start_time = now_ns - (tot_io_time>>1);
2077 bfqq->tot_idle_time >>= 1;
2078 }
2079}
2080
2081/*
2082 * Detect whether bfqq's I/O seems synchronized with that of some
2083 * other queue, i.e., whether bfqq, after remaining empty, happens to
2084 * receive new I/O only right after some I/O request of the other
2085 * queue has been completed. We call waker queue the other queue, and
2086 * we assume, for simplicity, that bfqq may have at most one waker
2087 * queue.
2088 *
2089 * A remarkable throughput boost can be reached by unconditionally
2090 * injecting the I/O of the waker queue, every time a new
2091 * bfq_dispatch_request happens to be invoked while I/O is being
2092 * plugged for bfqq. In addition to boosting throughput, this
2093 * unblocks bfqq's I/O, thereby improving bandwidth and latency for
2094 * bfqq. Note that these same results may be achieved with the general
2095 * injection mechanism, but less effectively. For details on this
2096 * aspect, see the comments on the choice of the queue for injection
2097 * in bfq_select_queue().
2098 *
2099 * Turning back to the detection of a waker queue, a queue Q is deemed as a
2100 * waker queue for bfqq if, for three consecutive times, bfqq happens to become
2101 * non empty right after a request of Q has been completed within given
2102 * timeout. In this respect, even if bfqq is empty, we do not check for a waker
2103 * if it still has some in-flight I/O. In fact, in this case bfqq is actually
2104 * still being served by the drive, and may receive new I/O on the completion
2105 * of some of the in-flight requests. In particular, on the first time, Q is
2106 * tentatively set as a candidate waker queue, while on the third consecutive
2107 * time that Q is detected, the field waker_bfqq is set to Q, to confirm that Q
2108 * is a waker queue for bfqq. These detection steps are performed only if bfqq
2109 * has a long think time, so as to make it more likely that bfqq's I/O is
2110 * actually being blocked by a synchronization. This last filter, plus the
2111 * above three-times requirement and time limit for detection, make false
2112 * positives less likely.
2113 *
2114 * NOTE
2115 *
2116 * The sooner a waker queue is detected, the sooner throughput can be
2117 * boosted by injecting I/O from the waker queue. Fortunately,
2118 * detection is likely to be actually fast, for the following
2119 * reasons. While blocked by synchronization, bfqq has a long think
2120 * time. This implies that bfqq's inject limit is at least equal to 1
2121 * (see the comments in bfq_update_inject_limit()). So, thanks to
2122 * injection, the waker queue is likely to be served during the very
2123 * first I/O-plugging time interval for bfqq. This triggers the first
2124 * step of the detection mechanism. Thanks again to injection, the
2125 * candidate waker queue is then likely to be confirmed no later than
2126 * during the next I/O-plugging interval for bfqq.
2127 *
2128 * ISSUE
2129 *
2130 * On queue merging all waker information is lost.
2131 */
2132static void bfq_check_waker(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2133 u64 now_ns)
2134{
2135 char waker_name[MAX_BFQQ_NAME_LENGTH];
2136
2137 if (!bfqd->last_completed_rq_bfqq ||
2138 bfqd->last_completed_rq_bfqq == bfqq ||
2139 bfq_bfqq_has_short_ttime(bfqq) ||
2140 now_ns - bfqd->last_completion >= 4 * NSEC_PER_MSEC ||
2141 bfqd->last_completed_rq_bfqq == &bfqd->oom_bfqq ||
2142 bfqq == &bfqd->oom_bfqq)
2143 return;
2144
2145 /*
2146 * We reset waker detection logic also if too much time has passed
2147 * since the first detection. If wakeups are rare, pointless idling
2148 * doesn't hurt throughput that much. The condition below makes sure
2149 * we do not uselessly idle blocking waker in more than 1/64 cases.
2150 */
2151 if (bfqd->last_completed_rq_bfqq !=
2152 bfqq->tentative_waker_bfqq ||
2153 now_ns > bfqq->waker_detection_started +
2154 128 * (u64)bfqd->bfq_slice_idle) {
2155 /*
2156 * First synchronization detected with a
2157 * candidate waker queue, or with a different
2158 * candidate waker queue from the current one.
2159 */
2160 bfqq->tentative_waker_bfqq =
2161 bfqd->last_completed_rq_bfqq;
2162 bfqq->num_waker_detections = 1;
2163 bfqq->waker_detection_started = now_ns;
2164 bfq_bfqq_name(bfqq->tentative_waker_bfqq, waker_name,
2165 MAX_BFQQ_NAME_LENGTH);
2166 bfq_log_bfqq(bfqd, bfqq, "set tentative waker %s", waker_name);
2167 } else /* Same tentative waker queue detected again */
2168 bfqq->num_waker_detections++;
2169
2170 if (bfqq->num_waker_detections == 3) {
2171 bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
2172 bfqq->tentative_waker_bfqq = NULL;
2173 bfq_bfqq_name(bfqq->waker_bfqq, waker_name,
2174 MAX_BFQQ_NAME_LENGTH);
2175 bfq_log_bfqq(bfqd, bfqq, "set waker %s", waker_name);
2176
2177 /*
2178 * If the waker queue disappears, then
2179 * bfqq->waker_bfqq must be reset. To
2180 * this goal, we maintain in each
2181 * waker queue a list, woken_list, of
2182 * all the queues that reference the
2183 * waker queue through their
2184 * waker_bfqq pointer. When the waker
2185 * queue exits, the waker_bfqq pointer
2186 * of all the queues in the woken_list
2187 * is reset.
2188 *
2189 * In addition, if bfqq is already in
2190 * the woken_list of a waker queue,
2191 * then, before being inserted into
2192 * the woken_list of a new waker
2193 * queue, bfqq must be removed from
2194 * the woken_list of the old waker
2195 * queue.
2196 */
2197 if (!hlist_unhashed(&bfqq->woken_list_node))
2198 hlist_del_init(&bfqq->woken_list_node);
2199 hlist_add_head(&bfqq->woken_list_node,
2200 &bfqd->last_completed_rq_bfqq->woken_list);
2201 }
2202}
2203
2204static void bfq_add_request(struct request *rq)
2205{
2206 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2207 struct bfq_data *bfqd = bfqq->bfqd;
2208 struct request *next_rq, *prev;
2209 unsigned int old_wr_coeff = bfqq->wr_coeff;
2210 bool interactive = false;
2211 u64 now_ns = ktime_get_ns();
2212
2213 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
2214 bfqq->queued[rq_is_sync(rq)]++;
2215 /*
2216 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2217 * may be read without holding the lock in bfq_has_work().
2218 */
2219 WRITE_ONCE(bfqd->queued, bfqd->queued + 1);
2220
2221 if (bfq_bfqq_sync(bfqq) && RQ_BIC(rq)->requests <= 1) {
2222 bfq_check_waker(bfqd, bfqq, now_ns);
2223
2224 /*
2225 * Periodically reset inject limit, to make sure that
2226 * the latter eventually drops in case workload
2227 * changes, see step (3) in the comments on
2228 * bfq_update_inject_limit().
2229 */
2230 if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2231 msecs_to_jiffies(1000)))
2232 bfq_reset_inject_limit(bfqd, bfqq);
2233
2234 /*
2235 * The following conditions must hold to setup a new
2236 * sampling of total service time, and then a new
2237 * update of the inject limit:
2238 * - bfqq is in service, because the total service
2239 * time is evaluated only for the I/O requests of
2240 * the queues in service;
2241 * - this is the right occasion to compute or to
2242 * lower the baseline total service time, because
2243 * there are actually no requests in the drive,
2244 * or
2245 * the baseline total service time is available, and
2246 * this is the right occasion to compute the other
2247 * quantity needed to update the inject limit, i.e.,
2248 * the total service time caused by the amount of
2249 * injection allowed by the current value of the
2250 * limit. It is the right occasion because injection
2251 * has actually been performed during the service
2252 * hole, and there are still in-flight requests,
2253 * which are very likely to be exactly the injected
2254 * requests, or part of them;
2255 * - the minimum interval for sampling the total
2256 * service time and updating the inject limit has
2257 * elapsed.
2258 */
2259 if (bfqq == bfqd->in_service_queue &&
2260 (bfqd->tot_rq_in_driver == 0 ||
2261 (bfqq->last_serv_time_ns > 0 &&
2262 bfqd->rqs_injected && bfqd->tot_rq_in_driver > 0)) &&
2263 time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2264 msecs_to_jiffies(10))) {
2265 bfqd->last_empty_occupied_ns = ktime_get_ns();
2266 /*
2267 * Start the state machine for measuring the
2268 * total service time of rq: setting
2269 * wait_dispatch will cause bfqd->waited_rq to
2270 * be set when rq will be dispatched.
2271 */
2272 bfqd->wait_dispatch = true;
2273 /*
2274 * If there is no I/O in service in the drive,
2275 * then possible injection occurred before the
2276 * arrival of rq will not affect the total
2277 * service time of rq. So the injection limit
2278 * must not be updated as a function of such
2279 * total service time, unless new injection
2280 * occurs before rq is completed. To have the
2281 * injection limit updated only in the latter
2282 * case, reset rqs_injected here (rqs_injected
2283 * will be set in case injection is performed
2284 * on bfqq before rq is completed).
2285 */
2286 if (bfqd->tot_rq_in_driver == 0)
2287 bfqd->rqs_injected = false;
2288 }
2289 }
2290
2291 if (bfq_bfqq_sync(bfqq))
2292 bfq_update_io_intensity(bfqq, now_ns);
2293
2294 elv_rb_add(&bfqq->sort_list, rq);
2295
2296 /*
2297 * Check if this request is a better next-serve candidate.
2298 */
2299 prev = bfqq->next_rq;
2300 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
2301 bfqq->next_rq = next_rq;
2302
2303 /*
2304 * Adjust priority tree position, if next_rq changes.
2305 * See comments on bfq_pos_tree_add_move() for the unlikely().
2306 */
2307 if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
2308 bfq_pos_tree_add_move(bfqd, bfqq);
2309
2310 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
2311 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
2312 rq, &interactive);
2313 else {
2314 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
2315 time_is_before_jiffies(
2316 bfqq->last_wr_start_finish +
2317 bfqd->bfq_wr_min_inter_arr_async)) {
2318 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
2319 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
2320
2321 bfqd->wr_busy_queues++;
2322 bfqq->entity.prio_changed = 1;
2323 }
2324 if (prev != bfqq->next_rq)
2325 bfq_updated_next_req(bfqd, bfqq);
2326 }
2327
2328 /*
2329 * Assign jiffies to last_wr_start_finish in the following
2330 * cases:
2331 *
2332 * . if bfqq is not going to be weight-raised, because, for
2333 * non weight-raised queues, last_wr_start_finish stores the
2334 * arrival time of the last request; as of now, this piece
2335 * of information is used only for deciding whether to
2336 * weight-raise async queues
2337 *
2338 * . if bfqq is not weight-raised, because, if bfqq is now
2339 * switching to weight-raised, then last_wr_start_finish
2340 * stores the time when weight-raising starts
2341 *
2342 * . if bfqq is interactive, because, regardless of whether
2343 * bfqq is currently weight-raised, the weight-raising
2344 * period must start or restart (this case is considered
2345 * separately because it is not detected by the above
2346 * conditions, if bfqq is already weight-raised)
2347 *
2348 * last_wr_start_finish has to be updated also if bfqq is soft
2349 * real-time, because the weight-raising period is constantly
2350 * restarted on idle-to-busy transitions for these queues, but
2351 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2352 * needed.
2353 */
2354 if (bfqd->low_latency &&
2355 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
2356 bfqq->last_wr_start_finish = jiffies;
2357}
2358
2359static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
2360 struct bio *bio,
2361 struct request_queue *q)
2362{
2363 struct bfq_queue *bfqq = bfqd->bio_bfqq;
2364
2365
2366 if (bfqq)
2367 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
2368
2369 return NULL;
2370}
2371
2372static sector_t get_sdist(sector_t last_pos, struct request *rq)
2373{
2374 if (last_pos)
2375 return abs(blk_rq_pos(rq) - last_pos);
2376
2377 return 0;
2378}
2379
2380static void bfq_remove_request(struct request_queue *q,
2381 struct request *rq)
2382{
2383 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2384 struct bfq_data *bfqd = bfqq->bfqd;
2385 const int sync = rq_is_sync(rq);
2386
2387 if (bfqq->next_rq == rq) {
2388 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
2389 bfq_updated_next_req(bfqd, bfqq);
2390 }
2391
2392 if (rq->queuelist.prev != &rq->queuelist)
2393 list_del_init(&rq->queuelist);
2394 bfqq->queued[sync]--;
2395 /*
2396 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2397 * may be read without holding the lock in bfq_has_work().
2398 */
2399 WRITE_ONCE(bfqd->queued, bfqd->queued - 1);
2400 elv_rb_del(&bfqq->sort_list, rq);
2401
2402 elv_rqhash_del(q, rq);
2403 if (q->last_merge == rq)
2404 q->last_merge = NULL;
2405
2406 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2407 bfqq->next_rq = NULL;
2408
2409 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
2410 bfq_del_bfqq_busy(bfqq, false);
2411 /*
2412 * bfqq emptied. In normal operation, when
2413 * bfqq is empty, bfqq->entity.service and
2414 * bfqq->entity.budget must contain,
2415 * respectively, the service received and the
2416 * budget used last time bfqq emptied. These
2417 * facts do not hold in this case, as at least
2418 * this last removal occurred while bfqq is
2419 * not in service. To avoid inconsistencies,
2420 * reset both bfqq->entity.service and
2421 * bfqq->entity.budget, if bfqq has still a
2422 * process that may issue I/O requests to it.
2423 */
2424 bfqq->entity.budget = bfqq->entity.service = 0;
2425 }
2426
2427 /*
2428 * Remove queue from request-position tree as it is empty.
2429 */
2430 if (bfqq->pos_root) {
2431 rb_erase(&bfqq->pos_node, bfqq->pos_root);
2432 bfqq->pos_root = NULL;
2433 }
2434 } else {
2435 /* see comments on bfq_pos_tree_add_move() for the unlikely() */
2436 if (unlikely(!bfqd->nonrot_with_queueing))
2437 bfq_pos_tree_add_move(bfqd, bfqq);
2438 }
2439
2440 if (rq->cmd_flags & REQ_META)
2441 bfqq->meta_pending--;
2442
2443}
2444
2445static bool bfq_bio_merge(struct request_queue *q, struct bio *bio,
2446 unsigned int nr_segs)
2447{
2448 struct bfq_data *bfqd = q->elevator->elevator_data;
2449 struct request *free = NULL;
2450 /*
2451 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2452 * store its return value for later use, to avoid nesting
2453 * queue_lock inside the bfqd->lock. We assume that the bic
2454 * returned by bfq_bic_lookup does not go away before
2455 * bfqd->lock is taken.
2456 */
2457 struct bfq_io_cq *bic = bfq_bic_lookup(q);
2458 bool ret;
2459
2460 spin_lock_irq(&bfqd->lock);
2461
2462 if (bic) {
2463 /*
2464 * Make sure cgroup info is uptodate for current process before
2465 * considering the merge.
2466 */
2467 bfq_bic_update_cgroup(bic, bio);
2468
2469 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf),
2470 bfq_actuator_index(bfqd, bio));
2471 } else {
2472 bfqd->bio_bfqq = NULL;
2473 }
2474 bfqd->bio_bic = bic;
2475
2476 ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
2477
2478 spin_unlock_irq(&bfqd->lock);
2479 if (free)
2480 blk_mq_free_request(free);
2481
2482 return ret;
2483}
2484
2485static int bfq_request_merge(struct request_queue *q, struct request **req,
2486 struct bio *bio)
2487{
2488 struct bfq_data *bfqd = q->elevator->elevator_data;
2489 struct request *__rq;
2490
2491 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
2492 if (__rq && elv_bio_merge_ok(__rq, bio)) {
2493 *req = __rq;
2494
2495 if (blk_discard_mergable(__rq))
2496 return ELEVATOR_DISCARD_MERGE;
2497 return ELEVATOR_FRONT_MERGE;
2498 }
2499
2500 return ELEVATOR_NO_MERGE;
2501}
2502
2503static void bfq_request_merged(struct request_queue *q, struct request *req,
2504 enum elv_merge type)
2505{
2506 if (type == ELEVATOR_FRONT_MERGE &&
2507 rb_prev(&req->rb_node) &&
2508 blk_rq_pos(req) <
2509 blk_rq_pos(container_of(rb_prev(&req->rb_node),
2510 struct request, rb_node))) {
2511 struct bfq_queue *bfqq = RQ_BFQQ(req);
2512 struct bfq_data *bfqd;
2513 struct request *prev, *next_rq;
2514
2515 if (!bfqq)
2516 return;
2517
2518 bfqd = bfqq->bfqd;
2519
2520 /* Reposition request in its sort_list */
2521 elv_rb_del(&bfqq->sort_list, req);
2522 elv_rb_add(&bfqq->sort_list, req);
2523
2524 /* Choose next request to be served for bfqq */
2525 prev = bfqq->next_rq;
2526 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
2527 bfqd->last_position);
2528 bfqq->next_rq = next_rq;
2529 /*
2530 * If next_rq changes, update both the queue's budget to
2531 * fit the new request and the queue's position in its
2532 * rq_pos_tree.
2533 */
2534 if (prev != bfqq->next_rq) {
2535 bfq_updated_next_req(bfqd, bfqq);
2536 /*
2537 * See comments on bfq_pos_tree_add_move() for
2538 * the unlikely().
2539 */
2540 if (unlikely(!bfqd->nonrot_with_queueing))
2541 bfq_pos_tree_add_move(bfqd, bfqq);
2542 }
2543 }
2544}
2545
2546/*
2547 * This function is called to notify the scheduler that the requests
2548 * rq and 'next' have been merged, with 'next' going away. BFQ
2549 * exploits this hook to address the following issue: if 'next' has a
2550 * fifo_time lower that rq, then the fifo_time of rq must be set to
2551 * the value of 'next', to not forget the greater age of 'next'.
2552 *
2553 * NOTE: in this function we assume that rq is in a bfq_queue, basing
2554 * on that rq is picked from the hash table q->elevator->hash, which,
2555 * in its turn, is filled only with I/O requests present in
2556 * bfq_queues, while BFQ is in use for the request queue q. In fact,
2557 * the function that fills this hash table (elv_rqhash_add) is called
2558 * only by bfq_insert_request.
2559 */
2560static void bfq_requests_merged(struct request_queue *q, struct request *rq,
2561 struct request *next)
2562{
2563 struct bfq_queue *bfqq = RQ_BFQQ(rq),
2564 *next_bfqq = RQ_BFQQ(next);
2565
2566 if (!bfqq)
2567 goto remove;
2568
2569 /*
2570 * If next and rq belong to the same bfq_queue and next is older
2571 * than rq, then reposition rq in the fifo (by substituting next
2572 * with rq). Otherwise, if next and rq belong to different
2573 * bfq_queues, never reposition rq: in fact, we would have to
2574 * reposition it with respect to next's position in its own fifo,
2575 * which would most certainly be too expensive with respect to
2576 * the benefits.
2577 */
2578 if (bfqq == next_bfqq &&
2579 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2580 next->fifo_time < rq->fifo_time) {
2581 list_del_init(&rq->queuelist);
2582 list_replace_init(&next->queuelist, &rq->queuelist);
2583 rq->fifo_time = next->fifo_time;
2584 }
2585
2586 if (bfqq->next_rq == next)
2587 bfqq->next_rq = rq;
2588
2589 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
2590remove:
2591 /* Merged request may be in the IO scheduler. Remove it. */
2592 if (!RB_EMPTY_NODE(&next->rb_node)) {
2593 bfq_remove_request(next->q, next);
2594 if (next_bfqq)
2595 bfqg_stats_update_io_remove(bfqq_group(next_bfqq),
2596 next->cmd_flags);
2597 }
2598}
2599
2600/* Must be called with bfqq != NULL */
2601static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
2602{
2603 /*
2604 * If bfqq has been enjoying interactive weight-raising, then
2605 * reset soft_rt_next_start. We do it for the following
2606 * reason. bfqq may have been conveying the I/O needed to load
2607 * a soft real-time application. Such an application actually
2608 * exhibits a soft real-time I/O pattern after it finishes
2609 * loading, and finally starts doing its job. But, if bfqq has
2610 * been receiving a lot of bandwidth so far (likely to happen
2611 * on a fast device), then soft_rt_next_start now contains a
2612 * high value that. So, without this reset, bfqq would be
2613 * prevented from being possibly considered as soft_rt for a
2614 * very long time.
2615 */
2616
2617 if (bfqq->wr_cur_max_time !=
2618 bfqq->bfqd->bfq_wr_rt_max_time)
2619 bfqq->soft_rt_next_start = jiffies;
2620
2621 if (bfq_bfqq_busy(bfqq))
2622 bfqq->bfqd->wr_busy_queues--;
2623 bfqq->wr_coeff = 1;
2624 bfqq->wr_cur_max_time = 0;
2625 bfqq->last_wr_start_finish = jiffies;
2626 /*
2627 * Trigger a weight change on the next invocation of
2628 * __bfq_entity_update_weight_prio.
2629 */
2630 bfqq->entity.prio_changed = 1;
2631}
2632
2633void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2634 struct bfq_group *bfqg)
2635{
2636 int i, j, k;
2637
2638 for (k = 0; k < bfqd->num_actuators; k++) {
2639 for (i = 0; i < 2; i++)
2640 for (j = 0; j < IOPRIO_NR_LEVELS; j++)
2641 if (bfqg->async_bfqq[i][j][k])
2642 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j][k]);
2643 if (bfqg->async_idle_bfqq[k])
2644 bfq_bfqq_end_wr(bfqg->async_idle_bfqq[k]);
2645 }
2646}
2647
2648static void bfq_end_wr(struct bfq_data *bfqd)
2649{
2650 struct bfq_queue *bfqq;
2651 int i;
2652
2653 spin_lock_irq(&bfqd->lock);
2654
2655 for (i = 0; i < bfqd->num_actuators; i++) {
2656 list_for_each_entry(bfqq, &bfqd->active_list[i], bfqq_list)
2657 bfq_bfqq_end_wr(bfqq);
2658 }
2659 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2660 bfq_bfqq_end_wr(bfqq);
2661 bfq_end_wr_async(bfqd);
2662
2663 spin_unlock_irq(&bfqd->lock);
2664}
2665
2666static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2667{
2668 if (request)
2669 return blk_rq_pos(io_struct);
2670 else
2671 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2672}
2673
2674static int bfq_rq_close_to_sector(void *io_struct, bool request,
2675 sector_t sector)
2676{
2677 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2678 BFQQ_CLOSE_THR;
2679}
2680
2681static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2682 struct bfq_queue *bfqq,
2683 sector_t sector)
2684{
2685 struct rb_root *root = &bfqq_group(bfqq)->rq_pos_tree;
2686 struct rb_node *parent, *node;
2687 struct bfq_queue *__bfqq;
2688
2689 if (RB_EMPTY_ROOT(root))
2690 return NULL;
2691
2692 /*
2693 * First, if we find a request starting at the end of the last
2694 * request, choose it.
2695 */
2696 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2697 if (__bfqq)
2698 return __bfqq;
2699
2700 /*
2701 * If the exact sector wasn't found, the parent of the NULL leaf
2702 * will contain the closest sector (rq_pos_tree sorted by
2703 * next_request position).
2704 */
2705 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2706 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2707 return __bfqq;
2708
2709 if (blk_rq_pos(__bfqq->next_rq) < sector)
2710 node = rb_next(&__bfqq->pos_node);
2711 else
2712 node = rb_prev(&__bfqq->pos_node);
2713 if (!node)
2714 return NULL;
2715
2716 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2717 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2718 return __bfqq;
2719
2720 return NULL;
2721}
2722
2723static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2724 struct bfq_queue *cur_bfqq,
2725 sector_t sector)
2726{
2727 struct bfq_queue *bfqq;
2728
2729 /*
2730 * We shall notice if some of the queues are cooperating,
2731 * e.g., working closely on the same area of the device. In
2732 * that case, we can group them together and: 1) don't waste
2733 * time idling, and 2) serve the union of their requests in
2734 * the best possible order for throughput.
2735 */
2736 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2737 if (!bfqq || bfqq == cur_bfqq)
2738 return NULL;
2739
2740 return bfqq;
2741}
2742
2743static struct bfq_queue *
2744bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2745{
2746 int process_refs, new_process_refs;
2747 struct bfq_queue *__bfqq;
2748
2749 /*
2750 * If there are no process references on the new_bfqq, then it is
2751 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2752 * may have dropped their last reference (not just their last process
2753 * reference).
2754 */
2755 if (!bfqq_process_refs(new_bfqq))
2756 return NULL;
2757
2758 /* Avoid a circular list and skip interim queue merges. */
2759 while ((__bfqq = new_bfqq->new_bfqq)) {
2760 if (__bfqq == bfqq)
2761 return NULL;
2762 new_bfqq = __bfqq;
2763 }
2764
2765 process_refs = bfqq_process_refs(bfqq);
2766 new_process_refs = bfqq_process_refs(new_bfqq);
2767 /*
2768 * If the process for the bfqq has gone away, there is no
2769 * sense in merging the queues.
2770 */
2771 if (process_refs == 0 || new_process_refs == 0)
2772 return NULL;
2773
2774 /*
2775 * Make sure merged queues belong to the same parent. Parents could
2776 * have changed since the time we decided the two queues are suitable
2777 * for merging.
2778 */
2779 if (new_bfqq->entity.parent != bfqq->entity.parent)
2780 return NULL;
2781
2782 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2783 new_bfqq->pid);
2784
2785 /*
2786 * Merging is just a redirection: the requests of the process
2787 * owning one of the two queues are redirected to the other queue.
2788 * The latter queue, in its turn, is set as shared if this is the
2789 * first time that the requests of some process are redirected to
2790 * it.
2791 *
2792 * We redirect bfqq to new_bfqq and not the opposite, because
2793 * we are in the context of the process owning bfqq, thus we
2794 * have the io_cq of this process. So we can immediately
2795 * configure this io_cq to redirect the requests of the
2796 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2797 * not available any more (new_bfqq->bic == NULL).
2798 *
2799 * Anyway, even in case new_bfqq coincides with the in-service
2800 * queue, redirecting requests the in-service queue is the
2801 * best option, as we feed the in-service queue with new
2802 * requests close to the last request served and, by doing so,
2803 * are likely to increase the throughput.
2804 */
2805 bfqq->new_bfqq = new_bfqq;
2806 /*
2807 * The above assignment schedules the following redirections:
2808 * each time some I/O for bfqq arrives, the process that
2809 * generated that I/O is disassociated from bfqq and
2810 * associated with new_bfqq. Here we increases new_bfqq->ref
2811 * in advance, adding the number of processes that are
2812 * expected to be associated with new_bfqq as they happen to
2813 * issue I/O.
2814 */
2815 new_bfqq->ref += process_refs;
2816 return new_bfqq;
2817}
2818
2819static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2820 struct bfq_queue *new_bfqq)
2821{
2822 if (bfq_too_late_for_merging(new_bfqq))
2823 return false;
2824
2825 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2826 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2827 return false;
2828
2829 /*
2830 * If either of the queues has already been detected as seeky,
2831 * then merging it with the other queue is unlikely to lead to
2832 * sequential I/O.
2833 */
2834 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2835 return false;
2836
2837 /*
2838 * Interleaved I/O is known to be done by (some) applications
2839 * only for reads, so it does not make sense to merge async
2840 * queues.
2841 */
2842 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2843 return false;
2844
2845 return true;
2846}
2847
2848static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
2849 struct bfq_queue *bfqq);
2850
2851static struct bfq_queue *
2852bfq_setup_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2853 struct bfq_queue *stable_merge_bfqq,
2854 struct bfq_iocq_bfqq_data *bfqq_data)
2855{
2856 int proc_ref = min(bfqq_process_refs(bfqq),
2857 bfqq_process_refs(stable_merge_bfqq));
2858 struct bfq_queue *new_bfqq = NULL;
2859
2860 bfqq_data->stable_merge_bfqq = NULL;
2861 if (idling_boosts_thr_without_issues(bfqd, bfqq) || proc_ref == 0)
2862 goto out;
2863
2864 /* next function will take at least one ref */
2865 new_bfqq = bfq_setup_merge(bfqq, stable_merge_bfqq);
2866
2867 if (new_bfqq) {
2868 bfqq_data->stably_merged = true;
2869 if (new_bfqq->bic) {
2870 unsigned int new_a_idx = new_bfqq->actuator_idx;
2871 struct bfq_iocq_bfqq_data *new_bfqq_data =
2872 &new_bfqq->bic->bfqq_data[new_a_idx];
2873
2874 new_bfqq_data->stably_merged = true;
2875 }
2876 }
2877
2878out:
2879 /* deschedule stable merge, because done or aborted here */
2880 bfq_put_stable_ref(stable_merge_bfqq);
2881
2882 return new_bfqq;
2883}
2884
2885/*
2886 * Attempt to schedule a merge of bfqq with the currently in-service
2887 * queue or with a close queue among the scheduled queues. Return
2888 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2889 * structure otherwise.
2890 *
2891 * The OOM queue is not allowed to participate to cooperation: in fact, since
2892 * the requests temporarily redirected to the OOM queue could be redirected
2893 * again to dedicated queues at any time, the state needed to correctly
2894 * handle merging with the OOM queue would be quite complex and expensive
2895 * to maintain. Besides, in such a critical condition as an out of memory,
2896 * the benefits of queue merging may be little relevant, or even negligible.
2897 *
2898 * WARNING: queue merging may impair fairness among non-weight raised
2899 * queues, for at least two reasons: 1) the original weight of a
2900 * merged queue may change during the merged state, 2) even being the
2901 * weight the same, a merged queue may be bloated with many more
2902 * requests than the ones produced by its originally-associated
2903 * process.
2904 */
2905static struct bfq_queue *
2906bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2907 void *io_struct, bool request, struct bfq_io_cq *bic)
2908{
2909 struct bfq_queue *in_service_bfqq, *new_bfqq;
2910 unsigned int a_idx = bfqq->actuator_idx;
2911 struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[a_idx];
2912
2913 /* if a merge has already been setup, then proceed with that first */
2914 if (bfqq->new_bfqq)
2915 return bfqq->new_bfqq;
2916
2917 /*
2918 * Check delayed stable merge for rotational or non-queueing
2919 * devs. For this branch to be executed, bfqq must not be
2920 * currently merged with some other queue (i.e., bfqq->bic
2921 * must be non null). If we considered also merged queues,
2922 * then we should also check whether bfqq has already been
2923 * merged with bic->stable_merge_bfqq. But this would be
2924 * costly and complicated.
2925 */
2926 if (unlikely(!bfqd->nonrot_with_queueing)) {
2927 /*
2928 * Make sure also that bfqq is sync, because
2929 * bic->stable_merge_bfqq may point to some queue (for
2930 * stable merging) also if bic is associated with a
2931 * sync queue, but this bfqq is async
2932 */
2933 if (bfq_bfqq_sync(bfqq) && bfqq_data->stable_merge_bfqq &&
2934 !bfq_bfqq_just_created(bfqq) &&
2935 time_is_before_jiffies(bfqq->split_time +
2936 msecs_to_jiffies(bfq_late_stable_merging)) &&
2937 time_is_before_jiffies(bfqq->creation_time +
2938 msecs_to_jiffies(bfq_late_stable_merging))) {
2939 struct bfq_queue *stable_merge_bfqq =
2940 bfqq_data->stable_merge_bfqq;
2941
2942 return bfq_setup_stable_merge(bfqd, bfqq,
2943 stable_merge_bfqq,
2944 bfqq_data);
2945 }
2946 }
2947
2948 /*
2949 * Do not perform queue merging if the device is non
2950 * rotational and performs internal queueing. In fact, such a
2951 * device reaches a high speed through internal parallelism
2952 * and pipelining. This means that, to reach a high
2953 * throughput, it must have many requests enqueued at the same
2954 * time. But, in this configuration, the internal scheduling
2955 * algorithm of the device does exactly the job of queue
2956 * merging: it reorders requests so as to obtain as much as
2957 * possible a sequential I/O pattern. As a consequence, with
2958 * the workload generated by processes doing interleaved I/O,
2959 * the throughput reached by the device is likely to be the
2960 * same, with and without queue merging.
2961 *
2962 * Disabling merging also provides a remarkable benefit in
2963 * terms of throughput. Merging tends to make many workloads
2964 * artificially more uneven, because of shared queues
2965 * remaining non empty for incomparably more time than
2966 * non-merged queues. This may accentuate workload
2967 * asymmetries. For example, if one of the queues in a set of
2968 * merged queues has a higher weight than a normal queue, then
2969 * the shared queue may inherit such a high weight and, by
2970 * staying almost always active, may force BFQ to perform I/O
2971 * plugging most of the time. This evidently makes it harder
2972 * for BFQ to let the device reach a high throughput.
2973 *
2974 * Finally, the likely() macro below is not used because one
2975 * of the two branches is more likely than the other, but to
2976 * have the code path after the following if() executed as
2977 * fast as possible for the case of a non rotational device
2978 * with queueing. We want it because this is the fastest kind
2979 * of device. On the opposite end, the likely() may lengthen
2980 * the execution time of BFQ for the case of slower devices
2981 * (rotational or at least without queueing). But in this case
2982 * the execution time of BFQ matters very little, if not at
2983 * all.
2984 */
2985 if (likely(bfqd->nonrot_with_queueing))
2986 return NULL;
2987
2988 /*
2989 * Prevent bfqq from being merged if it has been created too
2990 * long ago. The idea is that true cooperating processes, and
2991 * thus their associated bfq_queues, are supposed to be
2992 * created shortly after each other. This is the case, e.g.,
2993 * for KVM/QEMU and dump I/O threads. Basing on this
2994 * assumption, the following filtering greatly reduces the
2995 * probability that two non-cooperating processes, which just
2996 * happen to do close I/O for some short time interval, have
2997 * their queues merged by mistake.
2998 */
2999 if (bfq_too_late_for_merging(bfqq))
3000 return NULL;
3001
3002 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
3003 return NULL;
3004
3005 /* If there is only one backlogged queue, don't search. */
3006 if (bfq_tot_busy_queues(bfqd) == 1)
3007 return NULL;
3008
3009 in_service_bfqq = bfqd->in_service_queue;
3010
3011 if (in_service_bfqq && in_service_bfqq != bfqq &&
3012 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
3013 bfq_rq_close_to_sector(io_struct, request,
3014 bfqd->in_serv_last_pos) &&
3015 bfqq->entity.parent == in_service_bfqq->entity.parent &&
3016 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
3017 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
3018 if (new_bfqq)
3019 return new_bfqq;
3020 }
3021 /*
3022 * Check whether there is a cooperator among currently scheduled
3023 * queues. The only thing we need is that the bio/request is not
3024 * NULL, as we need it to establish whether a cooperator exists.
3025 */
3026 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
3027 bfq_io_struct_pos(io_struct, request));
3028
3029 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
3030 bfq_may_be_close_cooperator(bfqq, new_bfqq))
3031 return bfq_setup_merge(bfqq, new_bfqq);
3032
3033 return NULL;
3034}
3035
3036static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
3037{
3038 struct bfq_io_cq *bic = bfqq->bic;
3039 unsigned int a_idx = bfqq->actuator_idx;
3040 struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[a_idx];
3041
3042 /*
3043 * If !bfqq->bic, the queue is already shared or its requests
3044 * have already been redirected to a shared queue; both idle window
3045 * and weight raising state have already been saved. Do nothing.
3046 */
3047 if (!bic)
3048 return;
3049
3050 bfqq_data->saved_last_serv_time_ns = bfqq->last_serv_time_ns;
3051 bfqq_data->saved_inject_limit = bfqq->inject_limit;
3052 bfqq_data->saved_decrease_time_jif = bfqq->decrease_time_jif;
3053
3054 bfqq_data->saved_weight = bfqq->entity.orig_weight;
3055 bfqq_data->saved_ttime = bfqq->ttime;
3056 bfqq_data->saved_has_short_ttime =
3057 bfq_bfqq_has_short_ttime(bfqq);
3058 bfqq_data->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
3059 bfqq_data->saved_io_start_time = bfqq->io_start_time;
3060 bfqq_data->saved_tot_idle_time = bfqq->tot_idle_time;
3061 bfqq_data->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
3062 bfqq_data->was_in_burst_list =
3063 !hlist_unhashed(&bfqq->burst_list_node);
3064
3065 if (unlikely(bfq_bfqq_just_created(bfqq) &&
3066 !bfq_bfqq_in_large_burst(bfqq) &&
3067 bfqq->bfqd->low_latency)) {
3068 /*
3069 * bfqq being merged right after being created: bfqq
3070 * would have deserved interactive weight raising, but
3071 * did not make it to be set in a weight-raised state,
3072 * because of this early merge. Store directly the
3073 * weight-raising state that would have been assigned
3074 * to bfqq, so that to avoid that bfqq unjustly fails
3075 * to enjoy weight raising if split soon.
3076 */
3077 bfqq_data->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
3078 bfqq_data->saved_wr_start_at_switch_to_srt =
3079 bfq_smallest_from_now();
3080 bfqq_data->saved_wr_cur_max_time =
3081 bfq_wr_duration(bfqq->bfqd);
3082 bfqq_data->saved_last_wr_start_finish = jiffies;
3083 } else {
3084 bfqq_data->saved_wr_coeff = bfqq->wr_coeff;
3085 bfqq_data->saved_wr_start_at_switch_to_srt =
3086 bfqq->wr_start_at_switch_to_srt;
3087 bfqq_data->saved_service_from_wr =
3088 bfqq->service_from_wr;
3089 bfqq_data->saved_last_wr_start_finish =
3090 bfqq->last_wr_start_finish;
3091 bfqq_data->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
3092 }
3093}
3094
3095
3096static void
3097bfq_reassign_last_bfqq(struct bfq_queue *cur_bfqq, struct bfq_queue *new_bfqq)
3098{
3099 if (cur_bfqq->entity.parent &&
3100 cur_bfqq->entity.parent->last_bfqq_created == cur_bfqq)
3101 cur_bfqq->entity.parent->last_bfqq_created = new_bfqq;
3102 else if (cur_bfqq->bfqd && cur_bfqq->bfqd->last_bfqq_created == cur_bfqq)
3103 cur_bfqq->bfqd->last_bfqq_created = new_bfqq;
3104}
3105
3106void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3107{
3108 /*
3109 * To prevent bfqq's service guarantees from being violated,
3110 * bfqq may be left busy, i.e., queued for service, even if
3111 * empty (see comments in __bfq_bfqq_expire() for
3112 * details). But, if no process will send requests to bfqq any
3113 * longer, then there is no point in keeping bfqq queued for
3114 * service. In addition, keeping bfqq queued for service, but
3115 * with no process ref any longer, may have caused bfqq to be
3116 * freed when dequeued from service. But this is assumed to
3117 * never happen.
3118 */
3119 if (bfq_bfqq_busy(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) &&
3120 bfqq != bfqd->in_service_queue)
3121 bfq_del_bfqq_busy(bfqq, false);
3122
3123 bfq_reassign_last_bfqq(bfqq, NULL);
3124
3125 bfq_put_queue(bfqq);
3126}
3127
3128static void
3129bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
3130 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
3131{
3132 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
3133 (unsigned long)new_bfqq->pid);
3134 /* Save weight raising and idle window of the merged queues */
3135 bfq_bfqq_save_state(bfqq);
3136 bfq_bfqq_save_state(new_bfqq);
3137 if (bfq_bfqq_IO_bound(bfqq))
3138 bfq_mark_bfqq_IO_bound(new_bfqq);
3139 bfq_clear_bfqq_IO_bound(bfqq);
3140
3141 /*
3142 * The processes associated with bfqq are cooperators of the
3143 * processes associated with new_bfqq. So, if bfqq has a
3144 * waker, then assume that all these processes will be happy
3145 * to let bfqq's waker freely inject I/O when they have no
3146 * I/O.
3147 */
3148 if (bfqq->waker_bfqq && !new_bfqq->waker_bfqq &&
3149 bfqq->waker_bfqq != new_bfqq) {
3150 new_bfqq->waker_bfqq = bfqq->waker_bfqq;
3151 new_bfqq->tentative_waker_bfqq = NULL;
3152
3153 /*
3154 * If the waker queue disappears, then
3155 * new_bfqq->waker_bfqq must be reset. So insert
3156 * new_bfqq into the woken_list of the waker. See
3157 * bfq_check_waker for details.
3158 */
3159 hlist_add_head(&new_bfqq->woken_list_node,
3160 &new_bfqq->waker_bfqq->woken_list);
3161
3162 }
3163
3164 /*
3165 * If bfqq is weight-raised, then let new_bfqq inherit
3166 * weight-raising. To reduce false positives, neglect the case
3167 * where bfqq has just been created, but has not yet made it
3168 * to be weight-raised (which may happen because EQM may merge
3169 * bfqq even before bfq_add_request is executed for the first
3170 * time for bfqq). Handling this case would however be very
3171 * easy, thanks to the flag just_created.
3172 */
3173 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
3174 new_bfqq->wr_coeff = bfqq->wr_coeff;
3175 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
3176 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
3177 new_bfqq->wr_start_at_switch_to_srt =
3178 bfqq->wr_start_at_switch_to_srt;
3179 if (bfq_bfqq_busy(new_bfqq))
3180 bfqd->wr_busy_queues++;
3181 new_bfqq->entity.prio_changed = 1;
3182 }
3183
3184 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
3185 bfqq->wr_coeff = 1;
3186 bfqq->entity.prio_changed = 1;
3187 if (bfq_bfqq_busy(bfqq))
3188 bfqd->wr_busy_queues--;
3189 }
3190
3191 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
3192 bfqd->wr_busy_queues);
3193
3194 /*
3195 * Merge queues (that is, let bic redirect its requests to new_bfqq)
3196 */
3197 bic_set_bfqq(bic, new_bfqq, true, bfqq->actuator_idx);
3198 bfq_mark_bfqq_coop(new_bfqq);
3199 /*
3200 * new_bfqq now belongs to at least two bics (it is a shared queue):
3201 * set new_bfqq->bic to NULL. bfqq either:
3202 * - does not belong to any bic any more, and hence bfqq->bic must
3203 * be set to NULL, or
3204 * - is a queue whose owning bics have already been redirected to a
3205 * different queue, hence the queue is destined to not belong to
3206 * any bic soon and bfqq->bic is already NULL (therefore the next
3207 * assignment causes no harm).
3208 */
3209 new_bfqq->bic = NULL;
3210 /*
3211 * If the queue is shared, the pid is the pid of one of the associated
3212 * processes. Which pid depends on the exact sequence of merge events
3213 * the queue underwent. So printing such a pid is useless and confusing
3214 * because it reports a random pid between those of the associated
3215 * processes.
3216 * We mark such a queue with a pid -1, and then print SHARED instead of
3217 * a pid in logging messages.
3218 */
3219 new_bfqq->pid = -1;
3220 bfqq->bic = NULL;
3221
3222 bfq_reassign_last_bfqq(bfqq, new_bfqq);
3223
3224 bfq_release_process_ref(bfqd, bfqq);
3225}
3226
3227static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
3228 struct bio *bio)
3229{
3230 struct bfq_data *bfqd = q->elevator->elevator_data;
3231 bool is_sync = op_is_sync(bio->bi_opf);
3232 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
3233
3234 /*
3235 * Disallow merge of a sync bio into an async request.
3236 */
3237 if (is_sync && !rq_is_sync(rq))
3238 return false;
3239
3240 /*
3241 * Lookup the bfqq that this bio will be queued with. Allow
3242 * merge only if rq is queued there.
3243 */
3244 if (!bfqq)
3245 return false;
3246
3247 /*
3248 * We take advantage of this function to perform an early merge
3249 * of the queues of possible cooperating processes.
3250 */
3251 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false, bfqd->bio_bic);
3252 if (new_bfqq) {
3253 /*
3254 * bic still points to bfqq, then it has not yet been
3255 * redirected to some other bfq_queue, and a queue
3256 * merge between bfqq and new_bfqq can be safely
3257 * fulfilled, i.e., bic can be redirected to new_bfqq
3258 * and bfqq can be put.
3259 */
3260 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
3261 new_bfqq);
3262 /*
3263 * If we get here, bio will be queued into new_queue,
3264 * so use new_bfqq to decide whether bio and rq can be
3265 * merged.
3266 */
3267 bfqq = new_bfqq;
3268
3269 /*
3270 * Change also bqfd->bio_bfqq, as
3271 * bfqd->bio_bic now points to new_bfqq, and
3272 * this function may be invoked again (and then may
3273 * use again bqfd->bio_bfqq).
3274 */
3275 bfqd->bio_bfqq = bfqq;
3276 }
3277
3278 return bfqq == RQ_BFQQ(rq);
3279}
3280
3281/*
3282 * Set the maximum time for the in-service queue to consume its
3283 * budget. This prevents seeky processes from lowering the throughput.
3284 * In practice, a time-slice service scheme is used with seeky
3285 * processes.
3286 */
3287static void bfq_set_budget_timeout(struct bfq_data *bfqd,
3288 struct bfq_queue *bfqq)
3289{
3290 unsigned int timeout_coeff;
3291
3292 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
3293 timeout_coeff = 1;
3294 else
3295 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
3296
3297 bfqd->last_budget_start = ktime_get();
3298
3299 bfqq->budget_timeout = jiffies +
3300 bfqd->bfq_timeout * timeout_coeff;
3301}
3302
3303static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
3304 struct bfq_queue *bfqq)
3305{
3306 if (bfqq) {
3307 bfq_clear_bfqq_fifo_expire(bfqq);
3308
3309 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
3310
3311 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
3312 bfqq->wr_coeff > 1 &&
3313 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
3314 time_is_before_jiffies(bfqq->budget_timeout)) {
3315 /*
3316 * For soft real-time queues, move the start
3317 * of the weight-raising period forward by the
3318 * time the queue has not received any
3319 * service. Otherwise, a relatively long
3320 * service delay is likely to cause the
3321 * weight-raising period of the queue to end,
3322 * because of the short duration of the
3323 * weight-raising period of a soft real-time
3324 * queue. It is worth noting that this move
3325 * is not so dangerous for the other queues,
3326 * because soft real-time queues are not
3327 * greedy.
3328 *
3329 * To not add a further variable, we use the
3330 * overloaded field budget_timeout to
3331 * determine for how long the queue has not
3332 * received service, i.e., how much time has
3333 * elapsed since the queue expired. However,
3334 * this is a little imprecise, because
3335 * budget_timeout is set to jiffies if bfqq
3336 * not only expires, but also remains with no
3337 * request.
3338 */
3339 if (time_after(bfqq->budget_timeout,
3340 bfqq->last_wr_start_finish))
3341 bfqq->last_wr_start_finish +=
3342 jiffies - bfqq->budget_timeout;
3343 else
3344 bfqq->last_wr_start_finish = jiffies;
3345 }
3346
3347 bfq_set_budget_timeout(bfqd, bfqq);
3348 bfq_log_bfqq(bfqd, bfqq,
3349 "set_in_service_queue, cur-budget = %d",
3350 bfqq->entity.budget);
3351 }
3352
3353 bfqd->in_service_queue = bfqq;
3354 bfqd->in_serv_last_pos = 0;
3355}
3356
3357/*
3358 * Get and set a new queue for service.
3359 */
3360static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
3361{
3362 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
3363
3364 __bfq_set_in_service_queue(bfqd, bfqq);
3365 return bfqq;
3366}
3367
3368static void bfq_arm_slice_timer(struct bfq_data *bfqd)
3369{
3370 struct bfq_queue *bfqq = bfqd->in_service_queue;
3371 u32 sl;
3372
3373 bfq_mark_bfqq_wait_request(bfqq);
3374
3375 /*
3376 * We don't want to idle for seeks, but we do want to allow
3377 * fair distribution of slice time for a process doing back-to-back
3378 * seeks. So allow a little bit of time for him to submit a new rq.
3379 */
3380 sl = bfqd->bfq_slice_idle;
3381 /*
3382 * Unless the queue is being weight-raised or the scenario is
3383 * asymmetric, grant only minimum idle time if the queue
3384 * is seeky. A long idling is preserved for a weight-raised
3385 * queue, or, more in general, in an asymmetric scenario,
3386 * because a long idling is needed for guaranteeing to a queue
3387 * its reserved share of the throughput (in particular, it is
3388 * needed if the queue has a higher weight than some other
3389 * queue).
3390 */
3391 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3392 !bfq_asymmetric_scenario(bfqd, bfqq))
3393 sl = min_t(u64, sl, BFQ_MIN_TT);
3394 else if (bfqq->wr_coeff > 1)
3395 sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
3396
3397 bfqd->last_idling_start = ktime_get();
3398 bfqd->last_idling_start_jiffies = jiffies;
3399
3400 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
3401 HRTIMER_MODE_REL);
3402 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
3403}
3404
3405/*
3406 * In autotuning mode, max_budget is dynamically recomputed as the
3407 * amount of sectors transferred in timeout at the estimated peak
3408 * rate. This enables BFQ to utilize a full timeslice with a full
3409 * budget, even if the in-service queue is served at peak rate. And
3410 * this maximises throughput with sequential workloads.
3411 */
3412static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
3413{
3414 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
3415 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
3416}
3417
3418/*
3419 * Update parameters related to throughput and responsiveness, as a
3420 * function of the estimated peak rate. See comments on
3421 * bfq_calc_max_budget(), and on the ref_wr_duration array.
3422 */
3423static void update_thr_responsiveness_params(struct bfq_data *bfqd)
3424{
3425 if (bfqd->bfq_user_max_budget == 0) {
3426 bfqd->bfq_max_budget =
3427 bfq_calc_max_budget(bfqd);
3428 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
3429 }
3430}
3431
3432static void bfq_reset_rate_computation(struct bfq_data *bfqd,
3433 struct request *rq)
3434{
3435 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
3436 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
3437 bfqd->peak_rate_samples = 1;
3438 bfqd->sequential_samples = 0;
3439 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
3440 blk_rq_sectors(rq);
3441 } else /* no new rq dispatched, just reset the number of samples */
3442 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
3443
3444 bfq_log(bfqd,
3445 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
3446 bfqd->peak_rate_samples, bfqd->sequential_samples,
3447 bfqd->tot_sectors_dispatched);
3448}
3449
3450static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
3451{
3452 u32 rate, weight, divisor;
3453
3454 /*
3455 * For the convergence property to hold (see comments on
3456 * bfq_update_peak_rate()) and for the assessment to be
3457 * reliable, a minimum number of samples must be present, and
3458 * a minimum amount of time must have elapsed. If not so, do
3459 * not compute new rate. Just reset parameters, to get ready
3460 * for a new evaluation attempt.
3461 */
3462 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
3463 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
3464 goto reset_computation;
3465
3466 /*
3467 * If a new request completion has occurred after last
3468 * dispatch, then, to approximate the rate at which requests
3469 * have been served by the device, it is more precise to
3470 * extend the observation interval to the last completion.
3471 */
3472 bfqd->delta_from_first =
3473 max_t(u64, bfqd->delta_from_first,
3474 bfqd->last_completion - bfqd->first_dispatch);
3475
3476 /*
3477 * Rate computed in sects/usec, and not sects/nsec, for
3478 * precision issues.
3479 */
3480 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
3481 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
3482
3483 /*
3484 * Peak rate not updated if:
3485 * - the percentage of sequential dispatches is below 3/4 of the
3486 * total, and rate is below the current estimated peak rate
3487 * - rate is unreasonably high (> 20M sectors/sec)
3488 */
3489 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
3490 rate <= bfqd->peak_rate) ||
3491 rate > 20<<BFQ_RATE_SHIFT)
3492 goto reset_computation;
3493
3494 /*
3495 * We have to update the peak rate, at last! To this purpose,
3496 * we use a low-pass filter. We compute the smoothing constant
3497 * of the filter as a function of the 'weight' of the new
3498 * measured rate.
3499 *
3500 * As can be seen in next formulas, we define this weight as a
3501 * quantity proportional to how sequential the workload is,
3502 * and to how long the observation time interval is.
3503 *
3504 * The weight runs from 0 to 8. The maximum value of the
3505 * weight, 8, yields the minimum value for the smoothing
3506 * constant. At this minimum value for the smoothing constant,
3507 * the measured rate contributes for half of the next value of
3508 * the estimated peak rate.
3509 *
3510 * So, the first step is to compute the weight as a function
3511 * of how sequential the workload is. Note that the weight
3512 * cannot reach 9, because bfqd->sequential_samples cannot
3513 * become equal to bfqd->peak_rate_samples, which, in its
3514 * turn, holds true because bfqd->sequential_samples is not
3515 * incremented for the first sample.
3516 */
3517 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
3518
3519 /*
3520 * Second step: further refine the weight as a function of the
3521 * duration of the observation interval.
3522 */
3523 weight = min_t(u32, 8,
3524 div_u64(weight * bfqd->delta_from_first,
3525 BFQ_RATE_REF_INTERVAL));
3526
3527 /*
3528 * Divisor ranging from 10, for minimum weight, to 2, for
3529 * maximum weight.
3530 */
3531 divisor = 10 - weight;
3532
3533 /*
3534 * Finally, update peak rate:
3535 *
3536 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
3537 */
3538 bfqd->peak_rate *= divisor-1;
3539 bfqd->peak_rate /= divisor;
3540 rate /= divisor; /* smoothing constant alpha = 1/divisor */
3541
3542 bfqd->peak_rate += rate;
3543
3544 /*
3545 * For a very slow device, bfqd->peak_rate can reach 0 (see
3546 * the minimum representable values reported in the comments
3547 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3548 * divisions by zero where bfqd->peak_rate is used as a
3549 * divisor.
3550 */
3551 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
3552
3553 update_thr_responsiveness_params(bfqd);
3554
3555reset_computation:
3556 bfq_reset_rate_computation(bfqd, rq);
3557}
3558
3559/*
3560 * Update the read/write peak rate (the main quantity used for
3561 * auto-tuning, see update_thr_responsiveness_params()).
3562 *
3563 * It is not trivial to estimate the peak rate (correctly): because of
3564 * the presence of sw and hw queues between the scheduler and the
3565 * device components that finally serve I/O requests, it is hard to
3566 * say exactly when a given dispatched request is served inside the
3567 * device, and for how long. As a consequence, it is hard to know
3568 * precisely at what rate a given set of requests is actually served
3569 * by the device.
3570 *
3571 * On the opposite end, the dispatch time of any request is trivially
3572 * available, and, from this piece of information, the "dispatch rate"
3573 * of requests can be immediately computed. So, the idea in the next
3574 * function is to use what is known, namely request dispatch times
3575 * (plus, when useful, request completion times), to estimate what is
3576 * unknown, namely in-device request service rate.
3577 *
3578 * The main issue is that, because of the above facts, the rate at
3579 * which a certain set of requests is dispatched over a certain time
3580 * interval can vary greatly with respect to the rate at which the
3581 * same requests are then served. But, since the size of any
3582 * intermediate queue is limited, and the service scheme is lossless
3583 * (no request is silently dropped), the following obvious convergence
3584 * property holds: the number of requests dispatched MUST become
3585 * closer and closer to the number of requests completed as the
3586 * observation interval grows. This is the key property used in
3587 * the next function to estimate the peak service rate as a function
3588 * of the observed dispatch rate. The function assumes to be invoked
3589 * on every request dispatch.
3590 */
3591static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
3592{
3593 u64 now_ns = ktime_get_ns();
3594
3595 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
3596 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
3597 bfqd->peak_rate_samples);
3598 bfq_reset_rate_computation(bfqd, rq);
3599 goto update_last_values; /* will add one sample */
3600 }
3601
3602 /*
3603 * Device idle for very long: the observation interval lasting
3604 * up to this dispatch cannot be a valid observation interval
3605 * for computing a new peak rate (similarly to the late-
3606 * completion event in bfq_completed_request()). Go to
3607 * update_rate_and_reset to have the following three steps
3608 * taken:
3609 * - close the observation interval at the last (previous)
3610 * request dispatch or completion
3611 * - compute rate, if possible, for that observation interval
3612 * - start a new observation interval with this dispatch
3613 */
3614 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
3615 bfqd->tot_rq_in_driver == 0)
3616 goto update_rate_and_reset;
3617
3618 /* Update sampling information */
3619 bfqd->peak_rate_samples++;
3620
3621 if ((bfqd->tot_rq_in_driver > 0 ||
3622 now_ns - bfqd->last_completion < BFQ_MIN_TT)
3623 && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
3624 bfqd->sequential_samples++;
3625
3626 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
3627
3628 /* Reset max observed rq size every 32 dispatches */
3629 if (likely(bfqd->peak_rate_samples % 32))
3630 bfqd->last_rq_max_size =
3631 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
3632 else
3633 bfqd->last_rq_max_size = blk_rq_sectors(rq);
3634
3635 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
3636
3637 /* Target observation interval not yet reached, go on sampling */
3638 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
3639 goto update_last_values;
3640
3641update_rate_and_reset:
3642 bfq_update_rate_reset(bfqd, rq);
3643update_last_values:
3644 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
3645 if (RQ_BFQQ(rq) == bfqd->in_service_queue)
3646 bfqd->in_serv_last_pos = bfqd->last_position;
3647 bfqd->last_dispatch = now_ns;
3648}
3649
3650/*
3651 * Remove request from internal lists.
3652 */
3653static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
3654{
3655 struct bfq_queue *bfqq = RQ_BFQQ(rq);
3656
3657 /*
3658 * For consistency, the next instruction should have been
3659 * executed after removing the request from the queue and
3660 * dispatching it. We execute instead this instruction before
3661 * bfq_remove_request() (and hence introduce a temporary
3662 * inconsistency), for efficiency. In fact, should this
3663 * dispatch occur for a non in-service bfqq, this anticipated
3664 * increment prevents two counters related to bfqq->dispatched
3665 * from risking to be, first, uselessly decremented, and then
3666 * incremented again when the (new) value of bfqq->dispatched
3667 * happens to be taken into account.
3668 */
3669 bfqq->dispatched++;
3670 bfq_update_peak_rate(q->elevator->elevator_data, rq);
3671
3672 bfq_remove_request(q, rq);
3673}
3674
3675/*
3676 * There is a case where idling does not have to be performed for
3677 * throughput concerns, but to preserve the throughput share of
3678 * the process associated with bfqq.
3679 *
3680 * To introduce this case, we can note that allowing the drive
3681 * to enqueue more than one request at a time, and hence
3682 * delegating de facto final scheduling decisions to the
3683 * drive's internal scheduler, entails loss of control on the
3684 * actual request service order. In particular, the critical
3685 * situation is when requests from different processes happen
3686 * to be present, at the same time, in the internal queue(s)
3687 * of the drive. In such a situation, the drive, by deciding
3688 * the service order of the internally-queued requests, does
3689 * determine also the actual throughput distribution among
3690 * these processes. But the drive typically has no notion or
3691 * concern about per-process throughput distribution, and
3692 * makes its decisions only on a per-request basis. Therefore,
3693 * the service distribution enforced by the drive's internal
3694 * scheduler is likely to coincide with the desired throughput
3695 * distribution only in a completely symmetric, or favorably
3696 * skewed scenario where:
3697 * (i-a) each of these processes must get the same throughput as
3698 * the others,
3699 * (i-b) in case (i-a) does not hold, it holds that the process
3700 * associated with bfqq must receive a lower or equal
3701 * throughput than any of the other processes;
3702 * (ii) the I/O of each process has the same properties, in
3703 * terms of locality (sequential or random), direction
3704 * (reads or writes), request sizes, greediness
3705 * (from I/O-bound to sporadic), and so on;
3706
3707 * In fact, in such a scenario, the drive tends to treat the requests
3708 * of each process in about the same way as the requests of the
3709 * others, and thus to provide each of these processes with about the
3710 * same throughput. This is exactly the desired throughput
3711 * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3712 * even more convenient distribution for (the process associated with)
3713 * bfqq.
3714 *
3715 * In contrast, in any asymmetric or unfavorable scenario, device
3716 * idling (I/O-dispatch plugging) is certainly needed to guarantee
3717 * that bfqq receives its assigned fraction of the device throughput
3718 * (see [1] for details).
3719 *
3720 * The problem is that idling may significantly reduce throughput with
3721 * certain combinations of types of I/O and devices. An important
3722 * example is sync random I/O on flash storage with command
3723 * queueing. So, unless bfqq falls in cases where idling also boosts
3724 * throughput, it is important to check conditions (i-a), i(-b) and
3725 * (ii) accurately, so as to avoid idling when not strictly needed for
3726 * service guarantees.
3727 *
3728 * Unfortunately, it is extremely difficult to thoroughly check
3729 * condition (ii). And, in case there are active groups, it becomes
3730 * very difficult to check conditions (i-a) and (i-b) too. In fact,
3731 * if there are active groups, then, for conditions (i-a) or (i-b) to
3732 * become false 'indirectly', it is enough that an active group
3733 * contains more active processes or sub-groups than some other active
3734 * group. More precisely, for conditions (i-a) or (i-b) to become
3735 * false because of such a group, it is not even necessary that the
3736 * group is (still) active: it is sufficient that, even if the group
3737 * has become inactive, some of its descendant processes still have
3738 * some request already dispatched but still waiting for
3739 * completion. In fact, requests have still to be guaranteed their
3740 * share of the throughput even after being dispatched. In this
3741 * respect, it is easy to show that, if a group frequently becomes
3742 * inactive while still having in-flight requests, and if, when this
3743 * happens, the group is not considered in the calculation of whether
3744 * the scenario is asymmetric, then the group may fail to be
3745 * guaranteed its fair share of the throughput (basically because
3746 * idling may not be performed for the descendant processes of the
3747 * group, but it had to be). We address this issue with the following
3748 * bi-modal behavior, implemented in the function
3749 * bfq_asymmetric_scenario().
3750 *
3751 * If there are groups with requests waiting for completion
3752 * (as commented above, some of these groups may even be
3753 * already inactive), then the scenario is tagged as
3754 * asymmetric, conservatively, without checking any of the
3755 * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3756 * This behavior matches also the fact that groups are created
3757 * exactly if controlling I/O is a primary concern (to
3758 * preserve bandwidth and latency guarantees).
3759 *
3760 * On the opposite end, if there are no groups with requests waiting
3761 * for completion, then only conditions (i-a) and (i-b) are actually
3762 * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3763 * idling is not performed, regardless of whether condition (ii)
3764 * holds. In other words, only if conditions (i-a) and (i-b) do not
3765 * hold, then idling is allowed, and the device tends to be prevented
3766 * from queueing many requests, possibly of several processes. Since
3767 * there are no groups with requests waiting for completion, then, to
3768 * control conditions (i-a) and (i-b) it is enough to check just
3769 * whether all the queues with requests waiting for completion also
3770 * have the same weight.
3771 *
3772 * Not checking condition (ii) evidently exposes bfqq to the
3773 * risk of getting less throughput than its fair share.
3774 * However, for queues with the same weight, a further
3775 * mechanism, preemption, mitigates or even eliminates this
3776 * problem. And it does so without consequences on overall
3777 * throughput. This mechanism and its benefits are explained
3778 * in the next three paragraphs.
3779 *
3780 * Even if a queue, say Q, is expired when it remains idle, Q
3781 * can still preempt the new in-service queue if the next
3782 * request of Q arrives soon (see the comments on
3783 * bfq_bfqq_update_budg_for_activation). If all queues and
3784 * groups have the same weight, this form of preemption,
3785 * combined with the hole-recovery heuristic described in the
3786 * comments on function bfq_bfqq_update_budg_for_activation,
3787 * are enough to preserve a correct bandwidth distribution in
3788 * the mid term, even without idling. In fact, even if not
3789 * idling allows the internal queues of the device to contain
3790 * many requests, and thus to reorder requests, we can rather
3791 * safely assume that the internal scheduler still preserves a
3792 * minimum of mid-term fairness.
3793 *
3794 * More precisely, this preemption-based, idleless approach
3795 * provides fairness in terms of IOPS, and not sectors per
3796 * second. This can be seen with a simple example. Suppose
3797 * that there are two queues with the same weight, but that
3798 * the first queue receives requests of 8 sectors, while the
3799 * second queue receives requests of 1024 sectors. In
3800 * addition, suppose that each of the two queues contains at
3801 * most one request at a time, which implies that each queue
3802 * always remains idle after it is served. Finally, after
3803 * remaining idle, each queue receives very quickly a new
3804 * request. It follows that the two queues are served
3805 * alternatively, preempting each other if needed. This
3806 * implies that, although both queues have the same weight,
3807 * the queue with large requests receives a service that is
3808 * 1024/8 times as high as the service received by the other
3809 * queue.
3810 *
3811 * The motivation for using preemption instead of idling (for
3812 * queues with the same weight) is that, by not idling,
3813 * service guarantees are preserved (completely or at least in
3814 * part) without minimally sacrificing throughput. And, if
3815 * there is no active group, then the primary expectation for
3816 * this device is probably a high throughput.
3817 *
3818 * We are now left only with explaining the two sub-conditions in the
3819 * additional compound condition that is checked below for deciding
3820 * whether the scenario is asymmetric. To explain the first
3821 * sub-condition, we need to add that the function
3822 * bfq_asymmetric_scenario checks the weights of only
3823 * non-weight-raised queues, for efficiency reasons (see comments on
3824 * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3825 * is checked explicitly here. More precisely, the compound condition
3826 * below takes into account also the fact that, even if bfqq is being
3827 * weight-raised, the scenario is still symmetric if all queues with
3828 * requests waiting for completion happen to be
3829 * weight-raised. Actually, we should be even more precise here, and
3830 * differentiate between interactive weight raising and soft real-time
3831 * weight raising.
3832 *
3833 * The second sub-condition checked in the compound condition is
3834 * whether there is a fair amount of already in-flight I/O not
3835 * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3836 * following reason. The drive may decide to serve in-flight
3837 * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3838 * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3839 * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3840 * basically uncontrolled amount of I/O from other queues may be
3841 * dispatched too, possibly causing the service of bfqq's I/O to be
3842 * delayed even longer in the drive. This problem gets more and more
3843 * serious as the speed and the queue depth of the drive grow,
3844 * because, as these two quantities grow, the probability to find no
3845 * queue busy but many requests in flight grows too. By contrast,
3846 * plugging I/O dispatching minimizes the delay induced by already
3847 * in-flight I/O, and enables bfqq to recover the bandwidth it may
3848 * lose because of this delay.
3849 *
3850 * As a side note, it is worth considering that the above
3851 * device-idling countermeasures may however fail in the following
3852 * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3853 * in a time period during which all symmetry sub-conditions hold, and
3854 * therefore the device is allowed to enqueue many requests, but at
3855 * some later point in time some sub-condition stops to hold, then it
3856 * may become impossible to make requests be served in the desired
3857 * order until all the requests already queued in the device have been
3858 * served. The last sub-condition commented above somewhat mitigates
3859 * this problem for weight-raised queues.
3860 *
3861 * However, as an additional mitigation for this problem, we preserve
3862 * plugging for a special symmetric case that may suddenly turn into
3863 * asymmetric: the case where only bfqq is busy. In this case, not
3864 * expiring bfqq does not cause any harm to any other queues in terms
3865 * of service guarantees. In contrast, it avoids the following unlucky
3866 * sequence of events: (1) bfqq is expired, (2) a new queue with a
3867 * lower weight than bfqq becomes busy (or more queues), (3) the new
3868 * queue is served until a new request arrives for bfqq, (4) when bfqq
3869 * is finally served, there are so many requests of the new queue in
3870 * the drive that the pending requests for bfqq take a lot of time to
3871 * be served. In particular, event (2) may case even already
3872 * dispatched requests of bfqq to be delayed, inside the drive. So, to
3873 * avoid this series of events, the scenario is preventively declared
3874 * as asymmetric also if bfqq is the only busy queues
3875 */
3876static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3877 struct bfq_queue *bfqq)
3878{
3879 int tot_busy_queues = bfq_tot_busy_queues(bfqd);
3880
3881 /* No point in idling for bfqq if it won't get requests any longer */
3882 if (unlikely(!bfqq_process_refs(bfqq)))
3883 return false;
3884
3885 return (bfqq->wr_coeff > 1 &&
3886 (bfqd->wr_busy_queues < tot_busy_queues ||
3887 bfqd->tot_rq_in_driver >= bfqq->dispatched + 4)) ||
3888 bfq_asymmetric_scenario(bfqd, bfqq) ||
3889 tot_busy_queues == 1;
3890}
3891
3892static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3893 enum bfqq_expiration reason)
3894{
3895 /*
3896 * If this bfqq is shared between multiple processes, check
3897 * to make sure that those processes are still issuing I/Os
3898 * within the mean seek distance. If not, it may be time to
3899 * break the queues apart again.
3900 */
3901 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
3902 bfq_mark_bfqq_split_coop(bfqq);
3903
3904 /*
3905 * Consider queues with a higher finish virtual time than
3906 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3907 * true, then bfqq's bandwidth would be violated if an
3908 * uncontrolled amount of I/O from these queues were
3909 * dispatched while bfqq is waiting for its new I/O to
3910 * arrive. This is exactly what may happen if this is a forced
3911 * expiration caused by a preemption attempt, and if bfqq is
3912 * not re-scheduled. To prevent this from happening, re-queue
3913 * bfqq if it needs I/O-dispatch plugging, even if it is
3914 * empty. By doing so, bfqq is granted to be served before the
3915 * above queues (provided that bfqq is of course eligible).
3916 */
3917 if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
3918 !(reason == BFQQE_PREEMPTED &&
3919 idling_needed_for_service_guarantees(bfqd, bfqq))) {
3920 if (bfqq->dispatched == 0)
3921 /*
3922 * Overloading budget_timeout field to store
3923 * the time at which the queue remains with no
3924 * backlog and no outstanding request; used by
3925 * the weight-raising mechanism.
3926 */
3927 bfqq->budget_timeout = jiffies;
3928
3929 bfq_del_bfqq_busy(bfqq, true);
3930 } else {
3931 bfq_requeue_bfqq(bfqd, bfqq, true);
3932 /*
3933 * Resort priority tree of potential close cooperators.
3934 * See comments on bfq_pos_tree_add_move() for the unlikely().
3935 */
3936 if (unlikely(!bfqd->nonrot_with_queueing &&
3937 !RB_EMPTY_ROOT(&bfqq->sort_list)))
3938 bfq_pos_tree_add_move(bfqd, bfqq);
3939 }
3940
3941 /*
3942 * All in-service entities must have been properly deactivated
3943 * or requeued before executing the next function, which
3944 * resets all in-service entities as no more in service. This
3945 * may cause bfqq to be freed. If this happens, the next
3946 * function returns true.
3947 */
3948 return __bfq_bfqd_reset_in_service(bfqd);
3949}
3950
3951/**
3952 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3953 * @bfqd: device data.
3954 * @bfqq: queue to update.
3955 * @reason: reason for expiration.
3956 *
3957 * Handle the feedback on @bfqq budget at queue expiration.
3958 * See the body for detailed comments.
3959 */
3960static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
3961 struct bfq_queue *bfqq,
3962 enum bfqq_expiration reason)
3963{
3964 struct request *next_rq;
3965 int budget, min_budget;
3966
3967 min_budget = bfq_min_budget(bfqd);
3968
3969 if (bfqq->wr_coeff == 1)
3970 budget = bfqq->max_budget;
3971 else /*
3972 * Use a constant, low budget for weight-raised queues,
3973 * to help achieve a low latency. Keep it slightly higher
3974 * than the minimum possible budget, to cause a little
3975 * bit fewer expirations.
3976 */
3977 budget = 2 * min_budget;
3978
3979 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
3980 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
3981 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
3982 budget, bfq_min_budget(bfqd));
3983 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
3984 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
3985
3986 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
3987 switch (reason) {
3988 /*
3989 * Caveat: in all the following cases we trade latency
3990 * for throughput.
3991 */
3992 case BFQQE_TOO_IDLE:
3993 /*
3994 * This is the only case where we may reduce
3995 * the budget: if there is no request of the
3996 * process still waiting for completion, then
3997 * we assume (tentatively) that the timer has
3998 * expired because the batch of requests of
3999 * the process could have been served with a
4000 * smaller budget. Hence, betting that
4001 * process will behave in the same way when it
4002 * becomes backlogged again, we reduce its
4003 * next budget. As long as we guess right,
4004 * this budget cut reduces the latency
4005 * experienced by the process.
4006 *
4007 * However, if there are still outstanding
4008 * requests, then the process may have not yet
4009 * issued its next request just because it is
4010 * still waiting for the completion of some of
4011 * the still outstanding ones. So in this
4012 * subcase we do not reduce its budget, on the
4013 * contrary we increase it to possibly boost
4014 * the throughput, as discussed in the
4015 * comments to the BUDGET_TIMEOUT case.
4016 */
4017 if (bfqq->dispatched > 0) /* still outstanding reqs */
4018 budget = min(budget * 2, bfqd->bfq_max_budget);
4019 else {
4020 if (budget > 5 * min_budget)
4021 budget -= 4 * min_budget;
4022 else
4023 budget = min_budget;
4024 }
4025 break;
4026 case BFQQE_BUDGET_TIMEOUT:
4027 /*
4028 * We double the budget here because it gives
4029 * the chance to boost the throughput if this
4030 * is not a seeky process (and has bumped into
4031 * this timeout because of, e.g., ZBR).
4032 */
4033 budget = min(budget * 2, bfqd->bfq_max_budget);
4034 break;
4035 case BFQQE_BUDGET_EXHAUSTED:
4036 /*
4037 * The process still has backlog, and did not
4038 * let either the budget timeout or the disk
4039 * idling timeout expire. Hence it is not
4040 * seeky, has a short thinktime and may be
4041 * happy with a higher budget too. So
4042 * definitely increase the budget of this good
4043 * candidate to boost the disk throughput.
4044 */
4045 budget = min(budget * 4, bfqd->bfq_max_budget);
4046 break;
4047 case BFQQE_NO_MORE_REQUESTS:
4048 /*
4049 * For queues that expire for this reason, it
4050 * is particularly important to keep the
4051 * budget close to the actual service they
4052 * need. Doing so reduces the timestamp
4053 * misalignment problem described in the
4054 * comments in the body of
4055 * __bfq_activate_entity. In fact, suppose
4056 * that a queue systematically expires for
4057 * BFQQE_NO_MORE_REQUESTS and presents a
4058 * new request in time to enjoy timestamp
4059 * back-shifting. The larger the budget of the
4060 * queue is with respect to the service the
4061 * queue actually requests in each service
4062 * slot, the more times the queue can be
4063 * reactivated with the same virtual finish
4064 * time. It follows that, even if this finish
4065 * time is pushed to the system virtual time
4066 * to reduce the consequent timestamp
4067 * misalignment, the queue unjustly enjoys for
4068 * many re-activations a lower finish time
4069 * than all newly activated queues.
4070 *
4071 * The service needed by bfqq is measured
4072 * quite precisely by bfqq->entity.service.
4073 * Since bfqq does not enjoy device idling,
4074 * bfqq->entity.service is equal to the number
4075 * of sectors that the process associated with
4076 * bfqq requested to read/write before waiting
4077 * for request completions, or blocking for
4078 * other reasons.
4079 */
4080 budget = max_t(int, bfqq->entity.service, min_budget);
4081 break;
4082 default:
4083 return;
4084 }
4085 } else if (!bfq_bfqq_sync(bfqq)) {
4086 /*
4087 * Async queues get always the maximum possible
4088 * budget, as for them we do not care about latency
4089 * (in addition, their ability to dispatch is limited
4090 * by the charging factor).
4091 */
4092 budget = bfqd->bfq_max_budget;
4093 }
4094
4095 bfqq->max_budget = budget;
4096
4097 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
4098 !bfqd->bfq_user_max_budget)
4099 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
4100
4101 /*
4102 * If there is still backlog, then assign a new budget, making
4103 * sure that it is large enough for the next request. Since
4104 * the finish time of bfqq must be kept in sync with the
4105 * budget, be sure to call __bfq_bfqq_expire() *after* this
4106 * update.
4107 *
4108 * If there is no backlog, then no need to update the budget;
4109 * it will be updated on the arrival of a new request.
4110 */
4111 next_rq = bfqq->next_rq;
4112 if (next_rq)
4113 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
4114 bfq_serv_to_charge(next_rq, bfqq));
4115
4116 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
4117 next_rq ? blk_rq_sectors(next_rq) : 0,
4118 bfqq->entity.budget);
4119}
4120
4121/*
4122 * Return true if the process associated with bfqq is "slow". The slow
4123 * flag is used, in addition to the budget timeout, to reduce the
4124 * amount of service provided to seeky processes, and thus reduce
4125 * their chances to lower the throughput. More details in the comments
4126 * on the function bfq_bfqq_expire().
4127 *
4128 * An important observation is in order: as discussed in the comments
4129 * on the function bfq_update_peak_rate(), with devices with internal
4130 * queues, it is hard if ever possible to know when and for how long
4131 * an I/O request is processed by the device (apart from the trivial
4132 * I/O pattern where a new request is dispatched only after the
4133 * previous one has been completed). This makes it hard to evaluate
4134 * the real rate at which the I/O requests of each bfq_queue are
4135 * served. In fact, for an I/O scheduler like BFQ, serving a
4136 * bfq_queue means just dispatching its requests during its service
4137 * slot (i.e., until the budget of the queue is exhausted, or the
4138 * queue remains idle, or, finally, a timeout fires). But, during the
4139 * service slot of a bfq_queue, around 100 ms at most, the device may
4140 * be even still processing requests of bfq_queues served in previous
4141 * service slots. On the opposite end, the requests of the in-service
4142 * bfq_queue may be completed after the service slot of the queue
4143 * finishes.
4144 *
4145 * Anyway, unless more sophisticated solutions are used
4146 * (where possible), the sum of the sizes of the requests dispatched
4147 * during the service slot of a bfq_queue is probably the only
4148 * approximation available for the service received by the bfq_queue
4149 * during its service slot. And this sum is the quantity used in this
4150 * function to evaluate the I/O speed of a process.
4151 */
4152static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4153 bool compensate, unsigned long *delta_ms)
4154{
4155 ktime_t delta_ktime;
4156 u32 delta_usecs;
4157 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
4158
4159 if (!bfq_bfqq_sync(bfqq))
4160 return false;
4161
4162 if (compensate)
4163 delta_ktime = bfqd->last_idling_start;
4164 else
4165 delta_ktime = ktime_get();
4166 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
4167 delta_usecs = ktime_to_us(delta_ktime);
4168
4169 /* don't use too short time intervals */
4170 if (delta_usecs < 1000) {
4171 if (blk_queue_nonrot(bfqd->queue))
4172 /*
4173 * give same worst-case guarantees as idling
4174 * for seeky
4175 */
4176 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
4177 else /* charge at least one seek */
4178 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
4179
4180 return slow;
4181 }
4182
4183 *delta_ms = delta_usecs / USEC_PER_MSEC;
4184
4185 /*
4186 * Use only long (> 20ms) intervals to filter out excessive
4187 * spikes in service rate estimation.
4188 */
4189 if (delta_usecs > 20000) {
4190 /*
4191 * Caveat for rotational devices: processes doing I/O
4192 * in the slower disk zones tend to be slow(er) even
4193 * if not seeky. In this respect, the estimated peak
4194 * rate is likely to be an average over the disk
4195 * surface. Accordingly, to not be too harsh with
4196 * unlucky processes, a process is deemed slow only if
4197 * its rate has been lower than half of the estimated
4198 * peak rate.
4199 */
4200 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
4201 }
4202
4203 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
4204
4205 return slow;
4206}
4207
4208/*
4209 * To be deemed as soft real-time, an application must meet two
4210 * requirements. First, the application must not require an average
4211 * bandwidth higher than the approximate bandwidth required to playback or
4212 * record a compressed high-definition video.
4213 * The next function is invoked on the completion of the last request of a
4214 * batch, to compute the next-start time instant, soft_rt_next_start, such
4215 * that, if the next request of the application does not arrive before
4216 * soft_rt_next_start, then the above requirement on the bandwidth is met.
4217 *
4218 * The second requirement is that the request pattern of the application is
4219 * isochronous, i.e., that, after issuing a request or a batch of requests,
4220 * the application stops issuing new requests until all its pending requests
4221 * have been completed. After that, the application may issue a new batch,
4222 * and so on.
4223 * For this reason the next function is invoked to compute
4224 * soft_rt_next_start only for applications that meet this requirement,
4225 * whereas soft_rt_next_start is set to infinity for applications that do
4226 * not.
4227 *
4228 * Unfortunately, even a greedy (i.e., I/O-bound) application may
4229 * happen to meet, occasionally or systematically, both the above
4230 * bandwidth and isochrony requirements. This may happen at least in
4231 * the following circumstances. First, if the CPU load is high. The
4232 * application may stop issuing requests while the CPUs are busy
4233 * serving other processes, then restart, then stop again for a while,
4234 * and so on. The other circumstances are related to the storage
4235 * device: the storage device is highly loaded or reaches a low-enough
4236 * throughput with the I/O of the application (e.g., because the I/O
4237 * is random and/or the device is slow). In all these cases, the
4238 * I/O of the application may be simply slowed down enough to meet
4239 * the bandwidth and isochrony requirements. To reduce the probability
4240 * that greedy applications are deemed as soft real-time in these
4241 * corner cases, a further rule is used in the computation of
4242 * soft_rt_next_start: the return value of this function is forced to
4243 * be higher than the maximum between the following two quantities.
4244 *
4245 * (a) Current time plus: (1) the maximum time for which the arrival
4246 * of a request is waited for when a sync queue becomes idle,
4247 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
4248 * postpone for a moment the reason for adding a few extra
4249 * jiffies; we get back to it after next item (b). Lower-bounding
4250 * the return value of this function with the current time plus
4251 * bfqd->bfq_slice_idle tends to filter out greedy applications,
4252 * because the latter issue their next request as soon as possible
4253 * after the last one has been completed. In contrast, a soft
4254 * real-time application spends some time processing data, after a
4255 * batch of its requests has been completed.
4256 *
4257 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
4258 * above, greedy applications may happen to meet both the
4259 * bandwidth and isochrony requirements under heavy CPU or
4260 * storage-device load. In more detail, in these scenarios, these
4261 * applications happen, only for limited time periods, to do I/O
4262 * slowly enough to meet all the requirements described so far,
4263 * including the filtering in above item (a). These slow-speed
4264 * time intervals are usually interspersed between other time
4265 * intervals during which these applications do I/O at a very high
4266 * speed. Fortunately, exactly because of the high speed of the
4267 * I/O in the high-speed intervals, the values returned by this
4268 * function happen to be so high, near the end of any such
4269 * high-speed interval, to be likely to fall *after* the end of
4270 * the low-speed time interval that follows. These high values are
4271 * stored in bfqq->soft_rt_next_start after each invocation of
4272 * this function. As a consequence, if the last value of
4273 * bfqq->soft_rt_next_start is constantly used to lower-bound the
4274 * next value that this function may return, then, from the very
4275 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
4276 * likely to be constantly kept so high that any I/O request
4277 * issued during the low-speed interval is considered as arriving
4278 * to soon for the application to be deemed as soft
4279 * real-time. Then, in the high-speed interval that follows, the
4280 * application will not be deemed as soft real-time, just because
4281 * it will do I/O at a high speed. And so on.
4282 *
4283 * Getting back to the filtering in item (a), in the following two
4284 * cases this filtering might be easily passed by a greedy
4285 * application, if the reference quantity was just
4286 * bfqd->bfq_slice_idle:
4287 * 1) HZ is so low that the duration of a jiffy is comparable to or
4288 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
4289 * devices with HZ=100. The time granularity may be so coarse
4290 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
4291 * is rather lower than the exact value.
4292 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
4293 * for a while, then suddenly 'jump' by several units to recover the lost
4294 * increments. This seems to happen, e.g., inside virtual machines.
4295 * To address this issue, in the filtering in (a) we do not use as a
4296 * reference time interval just bfqd->bfq_slice_idle, but
4297 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
4298 * minimum number of jiffies for which the filter seems to be quite
4299 * precise also in embedded systems and KVM/QEMU virtual machines.
4300 */
4301static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
4302 struct bfq_queue *bfqq)
4303{
4304 return max3(bfqq->soft_rt_next_start,
4305 bfqq->last_idle_bklogged +
4306 HZ * bfqq->service_from_backlogged /
4307 bfqd->bfq_wr_max_softrt_rate,
4308 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
4309}
4310
4311/**
4312 * bfq_bfqq_expire - expire a queue.
4313 * @bfqd: device owning the queue.
4314 * @bfqq: the queue to expire.
4315 * @compensate: if true, compensate for the time spent idling.
4316 * @reason: the reason causing the expiration.
4317 *
4318 * If the process associated with bfqq does slow I/O (e.g., because it
4319 * issues random requests), we charge bfqq with the time it has been
4320 * in service instead of the service it has received (see
4321 * bfq_bfqq_charge_time for details on how this goal is achieved). As
4322 * a consequence, bfqq will typically get higher timestamps upon
4323 * reactivation, and hence it will be rescheduled as if it had
4324 * received more service than what it has actually received. In the
4325 * end, bfqq receives less service in proportion to how slowly its
4326 * associated process consumes its budgets (and hence how seriously it
4327 * tends to lower the throughput). In addition, this time-charging
4328 * strategy guarantees time fairness among slow processes. In
4329 * contrast, if the process associated with bfqq is not slow, we
4330 * charge bfqq exactly with the service it has received.
4331 *
4332 * Charging time to the first type of queues and the exact service to
4333 * the other has the effect of using the WF2Q+ policy to schedule the
4334 * former on a timeslice basis, without violating service domain
4335 * guarantees among the latter.
4336 */
4337void bfq_bfqq_expire(struct bfq_data *bfqd,
4338 struct bfq_queue *bfqq,
4339 bool compensate,
4340 enum bfqq_expiration reason)
4341{
4342 bool slow;
4343 unsigned long delta = 0;
4344 struct bfq_entity *entity = &bfqq->entity;
4345
4346 /*
4347 * Check whether the process is slow (see bfq_bfqq_is_slow).
4348 */
4349 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, &delta);
4350
4351 /*
4352 * As above explained, charge slow (typically seeky) and
4353 * timed-out queues with the time and not the service
4354 * received, to favor sequential workloads.
4355 *
4356 * Processes doing I/O in the slower disk zones will tend to
4357 * be slow(er) even if not seeky. Therefore, since the
4358 * estimated peak rate is actually an average over the disk
4359 * surface, these processes may timeout just for bad luck. To
4360 * avoid punishing them, do not charge time to processes that
4361 * succeeded in consuming at least 2/3 of their budget. This
4362 * allows BFQ to preserve enough elasticity to still perform
4363 * bandwidth, and not time, distribution with little unlucky
4364 * or quasi-sequential processes.
4365 */
4366 if (bfqq->wr_coeff == 1 &&
4367 (slow ||
4368 (reason == BFQQE_BUDGET_TIMEOUT &&
4369 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
4370 bfq_bfqq_charge_time(bfqd, bfqq, delta);
4371
4372 if (bfqd->low_latency && bfqq->wr_coeff == 1)
4373 bfqq->last_wr_start_finish = jiffies;
4374
4375 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
4376 RB_EMPTY_ROOT(&bfqq->sort_list)) {
4377 /*
4378 * If we get here, and there are no outstanding
4379 * requests, then the request pattern is isochronous
4380 * (see the comments on the function
4381 * bfq_bfqq_softrt_next_start()). Therefore we can
4382 * compute soft_rt_next_start.
4383 *
4384 * If, instead, the queue still has outstanding
4385 * requests, then we have to wait for the completion
4386 * of all the outstanding requests to discover whether
4387 * the request pattern is actually isochronous.
4388 */
4389 if (bfqq->dispatched == 0)
4390 bfqq->soft_rt_next_start =
4391 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4392 else if (bfqq->dispatched > 0) {
4393 /*
4394 * Schedule an update of soft_rt_next_start to when
4395 * the task may be discovered to be isochronous.
4396 */
4397 bfq_mark_bfqq_softrt_update(bfqq);
4398 }
4399 }
4400
4401 bfq_log_bfqq(bfqd, bfqq,
4402 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
4403 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
4404
4405 /*
4406 * bfqq expired, so no total service time needs to be computed
4407 * any longer: reset state machine for measuring total service
4408 * times.
4409 */
4410 bfqd->rqs_injected = bfqd->wait_dispatch = false;
4411 bfqd->waited_rq = NULL;
4412
4413 /*
4414 * Increase, decrease or leave budget unchanged according to
4415 * reason.
4416 */
4417 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
4418 if (__bfq_bfqq_expire(bfqd, bfqq, reason))
4419 /* bfqq is gone, no more actions on it */
4420 return;
4421
4422 /* mark bfqq as waiting a request only if a bic still points to it */
4423 if (!bfq_bfqq_busy(bfqq) &&
4424 reason != BFQQE_BUDGET_TIMEOUT &&
4425 reason != BFQQE_BUDGET_EXHAUSTED) {
4426 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
4427 /*
4428 * Not setting service to 0, because, if the next rq
4429 * arrives in time, the queue will go on receiving
4430 * service with this same budget (as if it never expired)
4431 */
4432 } else
4433 entity->service = 0;
4434
4435 /*
4436 * Reset the received-service counter for every parent entity.
4437 * Differently from what happens with bfqq->entity.service,
4438 * the resetting of this counter never needs to be postponed
4439 * for parent entities. In fact, in case bfqq may have a
4440 * chance to go on being served using the last, partially
4441 * consumed budget, bfqq->entity.service needs to be kept,
4442 * because if bfqq then actually goes on being served using
4443 * the same budget, the last value of bfqq->entity.service is
4444 * needed to properly decrement bfqq->entity.budget by the
4445 * portion already consumed. In contrast, it is not necessary
4446 * to keep entity->service for parent entities too, because
4447 * the bubble up of the new value of bfqq->entity.budget will
4448 * make sure that the budgets of parent entities are correct,
4449 * even in case bfqq and thus parent entities go on receiving
4450 * service with the same budget.
4451 */
4452 entity = entity->parent;
4453 for_each_entity(entity)
4454 entity->service = 0;
4455}
4456
4457/*
4458 * Budget timeout is not implemented through a dedicated timer, but
4459 * just checked on request arrivals and completions, as well as on
4460 * idle timer expirations.
4461 */
4462static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
4463{
4464 return time_is_before_eq_jiffies(bfqq->budget_timeout);
4465}
4466
4467/*
4468 * If we expire a queue that is actively waiting (i.e., with the
4469 * device idled) for the arrival of a new request, then we may incur
4470 * the timestamp misalignment problem described in the body of the
4471 * function __bfq_activate_entity. Hence we return true only if this
4472 * condition does not hold, or if the queue is slow enough to deserve
4473 * only to be kicked off for preserving a high throughput.
4474 */
4475static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
4476{
4477 bfq_log_bfqq(bfqq->bfqd, bfqq,
4478 "may_budget_timeout: wait_request %d left %d timeout %d",
4479 bfq_bfqq_wait_request(bfqq),
4480 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
4481 bfq_bfqq_budget_timeout(bfqq));
4482
4483 return (!bfq_bfqq_wait_request(bfqq) ||
4484 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
4485 &&
4486 bfq_bfqq_budget_timeout(bfqq);
4487}
4488
4489static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
4490 struct bfq_queue *bfqq)
4491{
4492 bool rot_without_queueing =
4493 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
4494 bfqq_sequential_and_IO_bound,
4495 idling_boosts_thr;
4496
4497 /* No point in idling for bfqq if it won't get requests any longer */
4498 if (unlikely(!bfqq_process_refs(bfqq)))
4499 return false;
4500
4501 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
4502 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
4503
4504 /*
4505 * The next variable takes into account the cases where idling
4506 * boosts the throughput.
4507 *
4508 * The value of the variable is computed considering, first, that
4509 * idling is virtually always beneficial for the throughput if:
4510 * (a) the device is not NCQ-capable and rotational, or
4511 * (b) regardless of the presence of NCQ, the device is rotational and
4512 * the request pattern for bfqq is I/O-bound and sequential, or
4513 * (c) regardless of whether it is rotational, the device is
4514 * not NCQ-capable and the request pattern for bfqq is
4515 * I/O-bound and sequential.
4516 *
4517 * Secondly, and in contrast to the above item (b), idling an
4518 * NCQ-capable flash-based device would not boost the
4519 * throughput even with sequential I/O; rather it would lower
4520 * the throughput in proportion to how fast the device
4521 * is. Accordingly, the next variable is true if any of the
4522 * above conditions (a), (b) or (c) is true, and, in
4523 * particular, happens to be false if bfqd is an NCQ-capable
4524 * flash-based device.
4525 */
4526 idling_boosts_thr = rot_without_queueing ||
4527 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
4528 bfqq_sequential_and_IO_bound);
4529
4530 /*
4531 * The return value of this function is equal to that of
4532 * idling_boosts_thr, unless a special case holds. In this
4533 * special case, described below, idling may cause problems to
4534 * weight-raised queues.
4535 *
4536 * When the request pool is saturated (e.g., in the presence
4537 * of write hogs), if the processes associated with
4538 * non-weight-raised queues ask for requests at a lower rate,
4539 * then processes associated with weight-raised queues have a
4540 * higher probability to get a request from the pool
4541 * immediately (or at least soon) when they need one. Thus
4542 * they have a higher probability to actually get a fraction
4543 * of the device throughput proportional to their high
4544 * weight. This is especially true with NCQ-capable drives,
4545 * which enqueue several requests in advance, and further
4546 * reorder internally-queued requests.
4547 *
4548 * For this reason, we force to false the return value if
4549 * there are weight-raised busy queues. In this case, and if
4550 * bfqq is not weight-raised, this guarantees that the device
4551 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4552 * then idling will be guaranteed by another variable, see
4553 * below). Combined with the timestamping rules of BFQ (see
4554 * [1] for details), this behavior causes bfqq, and hence any
4555 * sync non-weight-raised queue, to get a lower number of
4556 * requests served, and thus to ask for a lower number of
4557 * requests from the request pool, before the busy
4558 * weight-raised queues get served again. This often mitigates
4559 * starvation problems in the presence of heavy write
4560 * workloads and NCQ, thereby guaranteeing a higher
4561 * application and system responsiveness in these hostile
4562 * scenarios.
4563 */
4564 return idling_boosts_thr &&
4565 bfqd->wr_busy_queues == 0;
4566}
4567
4568/*
4569 * For a queue that becomes empty, device idling is allowed only if
4570 * this function returns true for that queue. As a consequence, since
4571 * device idling plays a critical role for both throughput boosting
4572 * and service guarantees, the return value of this function plays a
4573 * critical role as well.
4574 *
4575 * In a nutshell, this function returns true only if idling is
4576 * beneficial for throughput or, even if detrimental for throughput,
4577 * idling is however necessary to preserve service guarantees (low
4578 * latency, desired throughput distribution, ...). In particular, on
4579 * NCQ-capable devices, this function tries to return false, so as to
4580 * help keep the drives' internal queues full, whenever this helps the
4581 * device boost the throughput without causing any service-guarantee
4582 * issue.
4583 *
4584 * Most of the issues taken into account to get the return value of
4585 * this function are not trivial. We discuss these issues in the two
4586 * functions providing the main pieces of information needed by this
4587 * function.
4588 */
4589static bool bfq_better_to_idle(struct bfq_queue *bfqq)
4590{
4591 struct bfq_data *bfqd = bfqq->bfqd;
4592 bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
4593
4594 /* No point in idling for bfqq if it won't get requests any longer */
4595 if (unlikely(!bfqq_process_refs(bfqq)))
4596 return false;
4597
4598 if (unlikely(bfqd->strict_guarantees))
4599 return true;
4600
4601 /*
4602 * Idling is performed only if slice_idle > 0. In addition, we
4603 * do not idle if
4604 * (a) bfqq is async
4605 * (b) bfqq is in the idle io prio class: in this case we do
4606 * not idle because we want to minimize the bandwidth that
4607 * queues in this class can steal to higher-priority queues
4608 */
4609 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
4610 bfq_class_idle(bfqq))
4611 return false;
4612
4613 idling_boosts_thr_with_no_issue =
4614 idling_boosts_thr_without_issues(bfqd, bfqq);
4615
4616 idling_needed_for_service_guar =
4617 idling_needed_for_service_guarantees(bfqd, bfqq);
4618
4619 /*
4620 * We have now the two components we need to compute the
4621 * return value of the function, which is true only if idling
4622 * either boosts the throughput (without issues), or is
4623 * necessary to preserve service guarantees.
4624 */
4625 return idling_boosts_thr_with_no_issue ||
4626 idling_needed_for_service_guar;
4627}
4628
4629/*
4630 * If the in-service queue is empty but the function bfq_better_to_idle
4631 * returns true, then:
4632 * 1) the queue must remain in service and cannot be expired, and
4633 * 2) the device must be idled to wait for the possible arrival of a new
4634 * request for the queue.
4635 * See the comments on the function bfq_better_to_idle for the reasons
4636 * why performing device idling is the best choice to boost the throughput
4637 * and preserve service guarantees when bfq_better_to_idle itself
4638 * returns true.
4639 */
4640static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
4641{
4642 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
4643}
4644
4645/*
4646 * This function chooses the queue from which to pick the next extra
4647 * I/O request to inject, if it finds a compatible queue. See the
4648 * comments on bfq_update_inject_limit() for details on the injection
4649 * mechanism, and for the definitions of the quantities mentioned
4650 * below.
4651 */
4652static struct bfq_queue *
4653bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
4654{
4655 struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
4656 unsigned int limit = in_serv_bfqq->inject_limit;
4657 int i;
4658
4659 /*
4660 * If
4661 * - bfqq is not weight-raised and therefore does not carry
4662 * time-critical I/O,
4663 * or
4664 * - regardless of whether bfqq is weight-raised, bfqq has
4665 * however a long think time, during which it can absorb the
4666 * effect of an appropriate number of extra I/O requests
4667 * from other queues (see bfq_update_inject_limit for
4668 * details on the computation of this number);
4669 * then injection can be performed without restrictions.
4670 */
4671 bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
4672 !bfq_bfqq_has_short_ttime(in_serv_bfqq);
4673
4674 /*
4675 * If
4676 * - the baseline total service time could not be sampled yet,
4677 * so the inject limit happens to be still 0, and
4678 * - a lot of time has elapsed since the plugging of I/O
4679 * dispatching started, so drive speed is being wasted
4680 * significantly;
4681 * then temporarily raise inject limit to one request.
4682 */
4683 if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
4684 bfq_bfqq_wait_request(in_serv_bfqq) &&
4685 time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
4686 bfqd->bfq_slice_idle)
4687 )
4688 limit = 1;
4689
4690 if (bfqd->tot_rq_in_driver >= limit)
4691 return NULL;
4692
4693 /*
4694 * Linear search of the source queue for injection; but, with
4695 * a high probability, very few steps are needed to find a
4696 * candidate queue, i.e., a queue with enough budget left for
4697 * its next request. In fact:
4698 * - BFQ dynamically updates the budget of every queue so as
4699 * to accommodate the expected backlog of the queue;
4700 * - if a queue gets all its requests dispatched as injected
4701 * service, then the queue is removed from the active list
4702 * (and re-added only if it gets new requests, but then it
4703 * is assigned again enough budget for its new backlog).
4704 */
4705 for (i = 0; i < bfqd->num_actuators; i++) {
4706 list_for_each_entry(bfqq, &bfqd->active_list[i], bfqq_list)
4707 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4708 (in_serv_always_inject || bfqq->wr_coeff > 1) &&
4709 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4710 bfq_bfqq_budget_left(bfqq)) {
4711 /*
4712 * Allow for only one large in-flight request
4713 * on non-rotational devices, for the
4714 * following reason. On non-rotationl drives,
4715 * large requests take much longer than
4716 * smaller requests to be served. In addition,
4717 * the drive prefers to serve large requests
4718 * w.r.t. to small ones, if it can choose. So,
4719 * having more than one large requests queued
4720 * in the drive may easily make the next first
4721 * request of the in-service queue wait for so
4722 * long to break bfqq's service guarantees. On
4723 * the bright side, large requests let the
4724 * drive reach a very high throughput, even if
4725 * there is only one in-flight large request
4726 * at a time.
4727 */
4728 if (blk_queue_nonrot(bfqd->queue) &&
4729 blk_rq_sectors(bfqq->next_rq) >=
4730 BFQQ_SECT_THR_NONROT &&
4731 bfqd->tot_rq_in_driver >= 1)
4732 continue;
4733 else {
4734 bfqd->rqs_injected = true;
4735 return bfqq;
4736 }
4737 }
4738 }
4739
4740 return NULL;
4741}
4742
4743static struct bfq_queue *
4744bfq_find_active_bfqq_for_actuator(struct bfq_data *bfqd, int idx)
4745{
4746 struct bfq_queue *bfqq;
4747
4748 if (bfqd->in_service_queue &&
4749 bfqd->in_service_queue->actuator_idx == idx)
4750 return bfqd->in_service_queue;
4751
4752 list_for_each_entry(bfqq, &bfqd->active_list[idx], bfqq_list) {
4753 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4754 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4755 bfq_bfqq_budget_left(bfqq)) {
4756 return bfqq;
4757 }
4758 }
4759
4760 return NULL;
4761}
4762
4763/*
4764 * Perform a linear scan of each actuator, until an actuator is found
4765 * for which the following three conditions hold: the load of the
4766 * actuator is below the threshold (see comments on
4767 * actuator_load_threshold for details) and lower than that of the
4768 * next actuator (comments on this extra condition below), and there
4769 * is a queue that contains I/O for that actuator. On success, return
4770 * that queue.
4771 *
4772 * Performing a plain linear scan entails a prioritization among
4773 * actuators. The extra condition above breaks this prioritization and
4774 * tends to distribute injection uniformly across actuators.
4775 */
4776static struct bfq_queue *
4777bfq_find_bfqq_for_underused_actuator(struct bfq_data *bfqd)
4778{
4779 int i;
4780
4781 for (i = 0 ; i < bfqd->num_actuators; i++) {
4782 if (bfqd->rq_in_driver[i] < bfqd->actuator_load_threshold &&
4783 (i == bfqd->num_actuators - 1 ||
4784 bfqd->rq_in_driver[i] < bfqd->rq_in_driver[i+1])) {
4785 struct bfq_queue *bfqq =
4786 bfq_find_active_bfqq_for_actuator(bfqd, i);
4787
4788 if (bfqq)
4789 return bfqq;
4790 }
4791 }
4792
4793 return NULL;
4794}
4795
4796
4797/*
4798 * Select a queue for service. If we have a current queue in service,
4799 * check whether to continue servicing it, or retrieve and set a new one.
4800 */
4801static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
4802{
4803 struct bfq_queue *bfqq, *inject_bfqq;
4804 struct request *next_rq;
4805 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
4806
4807 bfqq = bfqd->in_service_queue;
4808 if (!bfqq)
4809 goto new_queue;
4810
4811 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
4812
4813 /*
4814 * Do not expire bfqq for budget timeout if bfqq may be about
4815 * to enjoy device idling. The reason why, in this case, we
4816 * prevent bfqq from expiring is the same as in the comments
4817 * on the case where bfq_bfqq_must_idle() returns true, in
4818 * bfq_completed_request().
4819 */
4820 if (bfq_may_expire_for_budg_timeout(bfqq) &&
4821 !bfq_bfqq_must_idle(bfqq))
4822 goto expire;
4823
4824check_queue:
4825 /*
4826 * If some actuator is underutilized, but the in-service
4827 * queue does not contain I/O for that actuator, then try to
4828 * inject I/O for that actuator.
4829 */
4830 inject_bfqq = bfq_find_bfqq_for_underused_actuator(bfqd);
4831 if (inject_bfqq && inject_bfqq != bfqq)
4832 return inject_bfqq;
4833
4834 /*
4835 * This loop is rarely executed more than once. Even when it
4836 * happens, it is much more convenient to re-execute this loop
4837 * than to return NULL and trigger a new dispatch to get a
4838 * request served.
4839 */
4840 next_rq = bfqq->next_rq;
4841 /*
4842 * If bfqq has requests queued and it has enough budget left to
4843 * serve them, keep the queue, otherwise expire it.
4844 */
4845 if (next_rq) {
4846 if (bfq_serv_to_charge(next_rq, bfqq) >
4847 bfq_bfqq_budget_left(bfqq)) {
4848 /*
4849 * Expire the queue for budget exhaustion,
4850 * which makes sure that the next budget is
4851 * enough to serve the next request, even if
4852 * it comes from the fifo expired path.
4853 */
4854 reason = BFQQE_BUDGET_EXHAUSTED;
4855 goto expire;
4856 } else {
4857 /*
4858 * The idle timer may be pending because we may
4859 * not disable disk idling even when a new request
4860 * arrives.
4861 */
4862 if (bfq_bfqq_wait_request(bfqq)) {
4863 /*
4864 * If we get here: 1) at least a new request
4865 * has arrived but we have not disabled the
4866 * timer because the request was too small,
4867 * 2) then the block layer has unplugged
4868 * the device, causing the dispatch to be
4869 * invoked.
4870 *
4871 * Since the device is unplugged, now the
4872 * requests are probably large enough to
4873 * provide a reasonable throughput.
4874 * So we disable idling.
4875 */
4876 bfq_clear_bfqq_wait_request(bfqq);
4877 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4878 }
4879 goto keep_queue;
4880 }
4881 }
4882
4883 /*
4884 * No requests pending. However, if the in-service queue is idling
4885 * for a new request, or has requests waiting for a completion and
4886 * may idle after their completion, then keep it anyway.
4887 *
4888 * Yet, inject service from other queues if it boosts
4889 * throughput and is possible.
4890 */
4891 if (bfq_bfqq_wait_request(bfqq) ||
4892 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
4893 unsigned int act_idx = bfqq->actuator_idx;
4894 struct bfq_queue *async_bfqq = NULL;
4895 struct bfq_queue *blocked_bfqq =
4896 !hlist_empty(&bfqq->woken_list) ?
4897 container_of(bfqq->woken_list.first,
4898 struct bfq_queue,
4899 woken_list_node)
4900 : NULL;
4901
4902 if (bfqq->bic && bfqq->bic->bfqq[0][act_idx] &&
4903 bfq_bfqq_busy(bfqq->bic->bfqq[0][act_idx]) &&
4904 bfqq->bic->bfqq[0][act_idx]->next_rq)
4905 async_bfqq = bfqq->bic->bfqq[0][act_idx];
4906 /*
4907 * The next four mutually-exclusive ifs decide
4908 * whether to try injection, and choose the queue to
4909 * pick an I/O request from.
4910 *
4911 * The first if checks whether the process associated
4912 * with bfqq has also async I/O pending. If so, it
4913 * injects such I/O unconditionally. Injecting async
4914 * I/O from the same process can cause no harm to the
4915 * process. On the contrary, it can only increase
4916 * bandwidth and reduce latency for the process.
4917 *
4918 * The second if checks whether there happens to be a
4919 * non-empty waker queue for bfqq, i.e., a queue whose
4920 * I/O needs to be completed for bfqq to receive new
4921 * I/O. This happens, e.g., if bfqq is associated with
4922 * a process that does some sync. A sync generates
4923 * extra blocking I/O, which must be completed before
4924 * the process associated with bfqq can go on with its
4925 * I/O. If the I/O of the waker queue is not served,
4926 * then bfqq remains empty, and no I/O is dispatched,
4927 * until the idle timeout fires for bfqq. This is
4928 * likely to result in lower bandwidth and higher
4929 * latencies for bfqq, and in a severe loss of total
4930 * throughput. The best action to take is therefore to
4931 * serve the waker queue as soon as possible. So do it
4932 * (without relying on the third alternative below for
4933 * eventually serving waker_bfqq's I/O; see the last
4934 * paragraph for further details). This systematic
4935 * injection of I/O from the waker queue does not
4936 * cause any delay to bfqq's I/O. On the contrary,
4937 * next bfqq's I/O is brought forward dramatically,
4938 * for it is not blocked for milliseconds.
4939 *
4940 * The third if checks whether there is a queue woken
4941 * by bfqq, and currently with pending I/O. Such a
4942 * woken queue does not steal bandwidth from bfqq,
4943 * because it remains soon without I/O if bfqq is not
4944 * served. So there is virtually no risk of loss of
4945 * bandwidth for bfqq if this woken queue has I/O
4946 * dispatched while bfqq is waiting for new I/O.
4947 *
4948 * The fourth if checks whether bfqq is a queue for
4949 * which it is better to avoid injection. It is so if
4950 * bfqq delivers more throughput when served without
4951 * any further I/O from other queues in the middle, or
4952 * if the service times of bfqq's I/O requests both
4953 * count more than overall throughput, and may be
4954 * easily increased by injection (this happens if bfqq
4955 * has a short think time). If none of these
4956 * conditions holds, then a candidate queue for
4957 * injection is looked for through
4958 * bfq_choose_bfqq_for_injection(). Note that the
4959 * latter may return NULL (for example if the inject
4960 * limit for bfqq is currently 0).
4961 *
4962 * NOTE: motivation for the second alternative
4963 *
4964 * Thanks to the way the inject limit is updated in
4965 * bfq_update_has_short_ttime(), it is rather likely
4966 * that, if I/O is being plugged for bfqq and the
4967 * waker queue has pending I/O requests that are
4968 * blocking bfqq's I/O, then the fourth alternative
4969 * above lets the waker queue get served before the
4970 * I/O-plugging timeout fires. So one may deem the
4971 * second alternative superfluous. It is not, because
4972 * the fourth alternative may be way less effective in
4973 * case of a synchronization. For two main
4974 * reasons. First, throughput may be low because the
4975 * inject limit may be too low to guarantee the same
4976 * amount of injected I/O, from the waker queue or
4977 * other queues, that the second alternative
4978 * guarantees (the second alternative unconditionally
4979 * injects a pending I/O request of the waker queue
4980 * for each bfq_dispatch_request()). Second, with the
4981 * fourth alternative, the duration of the plugging,
4982 * i.e., the time before bfqq finally receives new I/O,
4983 * may not be minimized, because the waker queue may
4984 * happen to be served only after other queues.
4985 */
4986 if (async_bfqq &&
4987 icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
4988 bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
4989 bfq_bfqq_budget_left(async_bfqq))
4990 bfqq = async_bfqq;
4991 else if (bfqq->waker_bfqq &&
4992 bfq_bfqq_busy(bfqq->waker_bfqq) &&
4993 bfqq->waker_bfqq->next_rq &&
4994 bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
4995 bfqq->waker_bfqq) <=
4996 bfq_bfqq_budget_left(bfqq->waker_bfqq)
4997 )
4998 bfqq = bfqq->waker_bfqq;
4999 else if (blocked_bfqq &&
5000 bfq_bfqq_busy(blocked_bfqq) &&
5001 blocked_bfqq->next_rq &&
5002 bfq_serv_to_charge(blocked_bfqq->next_rq,
5003 blocked_bfqq) <=
5004 bfq_bfqq_budget_left(blocked_bfqq)
5005 )
5006 bfqq = blocked_bfqq;
5007 else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
5008 (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
5009 !bfq_bfqq_has_short_ttime(bfqq)))
5010 bfqq = bfq_choose_bfqq_for_injection(bfqd);
5011 else
5012 bfqq = NULL;
5013
5014 goto keep_queue;
5015 }
5016
5017 reason = BFQQE_NO_MORE_REQUESTS;
5018expire:
5019 bfq_bfqq_expire(bfqd, bfqq, false, reason);
5020new_queue:
5021 bfqq = bfq_set_in_service_queue(bfqd);
5022 if (bfqq) {
5023 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
5024 goto check_queue;
5025 }
5026keep_queue:
5027 if (bfqq)
5028 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
5029 else
5030 bfq_log(bfqd, "select_queue: no queue returned");
5031
5032 return bfqq;
5033}
5034
5035static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
5036{
5037 struct bfq_entity *entity = &bfqq->entity;
5038
5039 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
5040 bfq_log_bfqq(bfqd, bfqq,
5041 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
5042 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
5043 jiffies_to_msecs(bfqq->wr_cur_max_time),
5044 bfqq->wr_coeff,
5045 bfqq->entity.weight, bfqq->entity.orig_weight);
5046
5047 if (entity->prio_changed)
5048 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
5049
5050 /*
5051 * If the queue was activated in a burst, or too much
5052 * time has elapsed from the beginning of this
5053 * weight-raising period, then end weight raising.
5054 */
5055 if (bfq_bfqq_in_large_burst(bfqq))
5056 bfq_bfqq_end_wr(bfqq);
5057 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
5058 bfqq->wr_cur_max_time)) {
5059 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
5060 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
5061 bfq_wr_duration(bfqd))) {
5062 /*
5063 * Either in interactive weight
5064 * raising, or in soft_rt weight
5065 * raising with the
5066 * interactive-weight-raising period
5067 * elapsed (so no switch back to
5068 * interactive weight raising).
5069 */
5070 bfq_bfqq_end_wr(bfqq);
5071 } else { /*
5072 * soft_rt finishing while still in
5073 * interactive period, switch back to
5074 * interactive weight raising
5075 */
5076 switch_back_to_interactive_wr(bfqq, bfqd);
5077 bfqq->entity.prio_changed = 1;
5078 }
5079 }
5080 if (bfqq->wr_coeff > 1 &&
5081 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
5082 bfqq->service_from_wr > max_service_from_wr) {
5083 /* see comments on max_service_from_wr */
5084 bfq_bfqq_end_wr(bfqq);
5085 }
5086 }
5087 /*
5088 * To improve latency (for this or other queues), immediately
5089 * update weight both if it must be raised and if it must be
5090 * lowered. Since, entity may be on some active tree here, and
5091 * might have a pending change of its ioprio class, invoke
5092 * next function with the last parameter unset (see the
5093 * comments on the function).
5094 */
5095 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
5096 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
5097 entity, false);
5098}
5099
5100/*
5101 * Dispatch next request from bfqq.
5102 */
5103static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
5104 struct bfq_queue *bfqq)
5105{
5106 struct request *rq = bfqq->next_rq;
5107 unsigned long service_to_charge;
5108
5109 service_to_charge = bfq_serv_to_charge(rq, bfqq);
5110
5111 bfq_bfqq_served(bfqq, service_to_charge);
5112
5113 if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
5114 bfqd->wait_dispatch = false;
5115 bfqd->waited_rq = rq;
5116 }
5117
5118 bfq_dispatch_remove(bfqd->queue, rq);
5119
5120 if (bfqq != bfqd->in_service_queue)
5121 return rq;
5122
5123 /*
5124 * If weight raising has to terminate for bfqq, then next
5125 * function causes an immediate update of bfqq's weight,
5126 * without waiting for next activation. As a consequence, on
5127 * expiration, bfqq will be timestamped as if has never been
5128 * weight-raised during this service slot, even if it has
5129 * received part or even most of the service as a
5130 * weight-raised queue. This inflates bfqq's timestamps, which
5131 * is beneficial, as bfqq is then more willing to leave the
5132 * device immediately to possible other weight-raised queues.
5133 */
5134 bfq_update_wr_data(bfqd, bfqq);
5135
5136 /*
5137 * Expire bfqq, pretending that its budget expired, if bfqq
5138 * belongs to CLASS_IDLE and other queues are waiting for
5139 * service.
5140 */
5141 if (bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq))
5142 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
5143
5144 return rq;
5145}
5146
5147static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
5148{
5149 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5150
5151 /*
5152 * Avoiding lock: a race on bfqd->queued should cause at
5153 * most a call to dispatch for nothing
5154 */
5155 return !list_empty_careful(&bfqd->dispatch) ||
5156 READ_ONCE(bfqd->queued);
5157}
5158
5159static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5160{
5161 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5162 struct request *rq = NULL;
5163 struct bfq_queue *bfqq = NULL;
5164
5165 if (!list_empty(&bfqd->dispatch)) {
5166 rq = list_first_entry(&bfqd->dispatch, struct request,
5167 queuelist);
5168 list_del_init(&rq->queuelist);
5169
5170 bfqq = RQ_BFQQ(rq);
5171
5172 if (bfqq) {
5173 /*
5174 * Increment counters here, because this
5175 * dispatch does not follow the standard
5176 * dispatch flow (where counters are
5177 * incremented)
5178 */
5179 bfqq->dispatched++;
5180
5181 goto inc_in_driver_start_rq;
5182 }
5183
5184 /*
5185 * We exploit the bfq_finish_requeue_request hook to
5186 * decrement tot_rq_in_driver, but
5187 * bfq_finish_requeue_request will not be invoked on
5188 * this request. So, to avoid unbalance, just start
5189 * this request, without incrementing tot_rq_in_driver. As
5190 * a negative consequence, tot_rq_in_driver is deceptively
5191 * lower than it should be while this request is in
5192 * service. This may cause bfq_schedule_dispatch to be
5193 * invoked uselessly.
5194 *
5195 * As for implementing an exact solution, the
5196 * bfq_finish_requeue_request hook, if defined, is
5197 * probably invoked also on this request. So, by
5198 * exploiting this hook, we could 1) increment
5199 * tot_rq_in_driver here, and 2) decrement it in
5200 * bfq_finish_requeue_request. Such a solution would
5201 * let the value of the counter be always accurate,
5202 * but it would entail using an extra interface
5203 * function. This cost seems higher than the benefit,
5204 * being the frequency of non-elevator-private
5205 * requests very low.
5206 */
5207 goto start_rq;
5208 }
5209
5210 bfq_log(bfqd, "dispatch requests: %d busy queues",
5211 bfq_tot_busy_queues(bfqd));
5212
5213 if (bfq_tot_busy_queues(bfqd) == 0)
5214 goto exit;
5215
5216 /*
5217 * Force device to serve one request at a time if
5218 * strict_guarantees is true. Forcing this service scheme is
5219 * currently the ONLY way to guarantee that the request
5220 * service order enforced by the scheduler is respected by a
5221 * queueing device. Otherwise the device is free even to make
5222 * some unlucky request wait for as long as the device
5223 * wishes.
5224 *
5225 * Of course, serving one request at a time may cause loss of
5226 * throughput.
5227 */
5228 if (bfqd->strict_guarantees && bfqd->tot_rq_in_driver > 0)
5229 goto exit;
5230
5231 bfqq = bfq_select_queue(bfqd);
5232 if (!bfqq)
5233 goto exit;
5234
5235 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
5236
5237 if (rq) {
5238inc_in_driver_start_rq:
5239 bfqd->rq_in_driver[bfqq->actuator_idx]++;
5240 bfqd->tot_rq_in_driver++;
5241start_rq:
5242 rq->rq_flags |= RQF_STARTED;
5243 }
5244exit:
5245 return rq;
5246}
5247
5248#ifdef CONFIG_BFQ_CGROUP_DEBUG
5249static void bfq_update_dispatch_stats(struct request_queue *q,
5250 struct request *rq,
5251 struct bfq_queue *in_serv_queue,
5252 bool idle_timer_disabled)
5253{
5254 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
5255
5256 if (!idle_timer_disabled && !bfqq)
5257 return;
5258
5259 /*
5260 * rq and bfqq are guaranteed to exist until this function
5261 * ends, for the following reasons. First, rq can be
5262 * dispatched to the device, and then can be completed and
5263 * freed, only after this function ends. Second, rq cannot be
5264 * merged (and thus freed because of a merge) any longer,
5265 * because it has already started. Thus rq cannot be freed
5266 * before this function ends, and, since rq has a reference to
5267 * bfqq, the same guarantee holds for bfqq too.
5268 *
5269 * In addition, the following queue lock guarantees that
5270 * bfqq_group(bfqq) exists as well.
5271 */
5272 spin_lock_irq(&q->queue_lock);
5273 if (idle_timer_disabled)
5274 /*
5275 * Since the idle timer has been disabled,
5276 * in_serv_queue contained some request when
5277 * __bfq_dispatch_request was invoked above, which
5278 * implies that rq was picked exactly from
5279 * in_serv_queue. Thus in_serv_queue == bfqq, and is
5280 * therefore guaranteed to exist because of the above
5281 * arguments.
5282 */
5283 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
5284 if (bfqq) {
5285 struct bfq_group *bfqg = bfqq_group(bfqq);
5286
5287 bfqg_stats_update_avg_queue_size(bfqg);
5288 bfqg_stats_set_start_empty_time(bfqg);
5289 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
5290 }
5291 spin_unlock_irq(&q->queue_lock);
5292}
5293#else
5294static inline void bfq_update_dispatch_stats(struct request_queue *q,
5295 struct request *rq,
5296 struct bfq_queue *in_serv_queue,
5297 bool idle_timer_disabled) {}
5298#endif /* CONFIG_BFQ_CGROUP_DEBUG */
5299
5300static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5301{
5302 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5303 struct request *rq;
5304 struct bfq_queue *in_serv_queue;
5305 bool waiting_rq, idle_timer_disabled = false;
5306
5307 spin_lock_irq(&bfqd->lock);
5308
5309 in_serv_queue = bfqd->in_service_queue;
5310 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
5311
5312 rq = __bfq_dispatch_request(hctx);
5313 if (in_serv_queue == bfqd->in_service_queue) {
5314 idle_timer_disabled =
5315 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
5316 }
5317
5318 spin_unlock_irq(&bfqd->lock);
5319 bfq_update_dispatch_stats(hctx->queue, rq,
5320 idle_timer_disabled ? in_serv_queue : NULL,
5321 idle_timer_disabled);
5322
5323 return rq;
5324}
5325
5326/*
5327 * Task holds one reference to the queue, dropped when task exits. Each rq
5328 * in-flight on this queue also holds a reference, dropped when rq is freed.
5329 *
5330 * Scheduler lock must be held here. Recall not to use bfqq after calling
5331 * this function on it.
5332 */
5333void bfq_put_queue(struct bfq_queue *bfqq)
5334{
5335 struct bfq_queue *item;
5336 struct hlist_node *n;
5337 struct bfq_group *bfqg = bfqq_group(bfqq);
5338
5339 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d", bfqq, bfqq->ref);
5340
5341 bfqq->ref--;
5342 if (bfqq->ref)
5343 return;
5344
5345 if (!hlist_unhashed(&bfqq->burst_list_node)) {
5346 hlist_del_init(&bfqq->burst_list_node);
5347 /*
5348 * Decrement also burst size after the removal, if the
5349 * process associated with bfqq is exiting, and thus
5350 * does not contribute to the burst any longer. This
5351 * decrement helps filter out false positives of large
5352 * bursts, when some short-lived process (often due to
5353 * the execution of commands by some service) happens
5354 * to start and exit while a complex application is
5355 * starting, and thus spawning several processes that
5356 * do I/O (and that *must not* be treated as a large
5357 * burst, see comments on bfq_handle_burst).
5358 *
5359 * In particular, the decrement is performed only if:
5360 * 1) bfqq is not a merged queue, because, if it is,
5361 * then this free of bfqq is not triggered by the exit
5362 * of the process bfqq is associated with, but exactly
5363 * by the fact that bfqq has just been merged.
5364 * 2) burst_size is greater than 0, to handle
5365 * unbalanced decrements. Unbalanced decrements may
5366 * happen in te following case: bfqq is inserted into
5367 * the current burst list--without incrementing
5368 * bust_size--because of a split, but the current
5369 * burst list is not the burst list bfqq belonged to
5370 * (see comments on the case of a split in
5371 * bfq_set_request).
5372 */
5373 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
5374 bfqq->bfqd->burst_size--;
5375 }
5376
5377 /*
5378 * bfqq does not exist any longer, so it cannot be woken by
5379 * any other queue, and cannot wake any other queue. Then bfqq
5380 * must be removed from the woken list of its possible waker
5381 * queue, and all queues in the woken list of bfqq must stop
5382 * having a waker queue. Strictly speaking, these updates
5383 * should be performed when bfqq remains with no I/O source
5384 * attached to it, which happens before bfqq gets freed. In
5385 * particular, this happens when the last process associated
5386 * with bfqq exits or gets associated with a different
5387 * queue. However, both events lead to bfqq being freed soon,
5388 * and dangling references would come out only after bfqq gets
5389 * freed. So these updates are done here, as a simple and safe
5390 * way to handle all cases.
5391 */
5392 /* remove bfqq from woken list */
5393 if (!hlist_unhashed(&bfqq->woken_list_node))
5394 hlist_del_init(&bfqq->woken_list_node);
5395
5396 /* reset waker for all queues in woken list */
5397 hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
5398 woken_list_node) {
5399 item->waker_bfqq = NULL;
5400 hlist_del_init(&item->woken_list_node);
5401 }
5402
5403 if (bfqq->bfqd->last_completed_rq_bfqq == bfqq)
5404 bfqq->bfqd->last_completed_rq_bfqq = NULL;
5405
5406 WARN_ON_ONCE(!list_empty(&bfqq->fifo));
5407 WARN_ON_ONCE(!RB_EMPTY_ROOT(&bfqq->sort_list));
5408 WARN_ON_ONCE(bfqq->dispatched);
5409
5410 kmem_cache_free(bfq_pool, bfqq);
5411 bfqg_and_blkg_put(bfqg);
5412}
5413
5414static void bfq_put_stable_ref(struct bfq_queue *bfqq)
5415{
5416 bfqq->stable_ref--;
5417 bfq_put_queue(bfqq);
5418}
5419
5420void bfq_put_cooperator(struct bfq_queue *bfqq)
5421{
5422 struct bfq_queue *__bfqq, *next;
5423
5424 /*
5425 * If this queue was scheduled to merge with another queue, be
5426 * sure to drop the reference taken on that queue (and others in
5427 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
5428 */
5429 __bfqq = bfqq->new_bfqq;
5430 while (__bfqq) {
5431 next = __bfqq->new_bfqq;
5432 bfq_put_queue(__bfqq);
5433 __bfqq = next;
5434 }
5435}
5436
5437static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
5438{
5439 if (bfqq == bfqd->in_service_queue) {
5440 __bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT);
5441 bfq_schedule_dispatch(bfqd);
5442 }
5443
5444 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
5445
5446 bfq_put_cooperator(bfqq);
5447
5448 bfq_release_process_ref(bfqd, bfqq);
5449}
5450
5451static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync,
5452 unsigned int actuator_idx)
5453{
5454 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync, actuator_idx);
5455 struct bfq_data *bfqd;
5456
5457 if (bfqq)
5458 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
5459
5460 if (bfqq && bfqd) {
5461 bic_set_bfqq(bic, NULL, is_sync, actuator_idx);
5462 bfq_exit_bfqq(bfqd, bfqq);
5463 }
5464}
5465
5466static void bfq_exit_icq(struct io_cq *icq)
5467{
5468 struct bfq_io_cq *bic = icq_to_bic(icq);
5469 struct bfq_data *bfqd = bic_to_bfqd(bic);
5470 unsigned long flags;
5471 unsigned int act_idx;
5472 /*
5473 * If bfqd and thus bfqd->num_actuators is not available any
5474 * longer, then cycle over all possible per-actuator bfqqs in
5475 * next loop. We rely on bic being zeroed on creation, and
5476 * therefore on its unused per-actuator fields being NULL.
5477 */
5478 unsigned int num_actuators = BFQ_MAX_ACTUATORS;
5479 struct bfq_iocq_bfqq_data *bfqq_data = bic->bfqq_data;
5480
5481 /*
5482 * bfqd is NULL if scheduler already exited, and in that case
5483 * this is the last time these queues are accessed.
5484 */
5485 if (bfqd) {
5486 spin_lock_irqsave(&bfqd->lock, flags);
5487 num_actuators = bfqd->num_actuators;
5488 }
5489
5490 for (act_idx = 0; act_idx < num_actuators; act_idx++) {
5491 if (bfqq_data[act_idx].stable_merge_bfqq)
5492 bfq_put_stable_ref(bfqq_data[act_idx].stable_merge_bfqq);
5493
5494 bfq_exit_icq_bfqq(bic, true, act_idx);
5495 bfq_exit_icq_bfqq(bic, false, act_idx);
5496 }
5497
5498 if (bfqd)
5499 spin_unlock_irqrestore(&bfqd->lock, flags);
5500}
5501
5502/*
5503 * Update the entity prio values; note that the new values will not
5504 * be used until the next (re)activation.
5505 */
5506static void
5507bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
5508{
5509 struct task_struct *tsk = current;
5510 int ioprio_class;
5511 struct bfq_data *bfqd = bfqq->bfqd;
5512
5513 if (!bfqd)
5514 return;
5515
5516 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5517 switch (ioprio_class) {
5518 default:
5519 pr_err("bdi %s: bfq: bad prio class %d\n",
5520 bdi_dev_name(bfqq->bfqd->queue->disk->bdi),
5521 ioprio_class);
5522 fallthrough;
5523 case IOPRIO_CLASS_NONE:
5524 /*
5525 * No prio set, inherit CPU scheduling settings.
5526 */
5527 bfqq->new_ioprio = task_nice_ioprio(tsk);
5528 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
5529 break;
5530 case IOPRIO_CLASS_RT:
5531 bfqq->new_ioprio = IOPRIO_PRIO_LEVEL(bic->ioprio);
5532 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
5533 break;
5534 case IOPRIO_CLASS_BE:
5535 bfqq->new_ioprio = IOPRIO_PRIO_LEVEL(bic->ioprio);
5536 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
5537 break;
5538 case IOPRIO_CLASS_IDLE:
5539 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
5540 bfqq->new_ioprio = IOPRIO_NR_LEVELS - 1;
5541 break;
5542 }
5543
5544 if (bfqq->new_ioprio >= IOPRIO_NR_LEVELS) {
5545 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
5546 bfqq->new_ioprio);
5547 bfqq->new_ioprio = IOPRIO_NR_LEVELS - 1;
5548 }
5549
5550 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
5551 bfq_log_bfqq(bfqd, bfqq, "new_ioprio %d new_weight %d",
5552 bfqq->new_ioprio, bfqq->entity.new_weight);
5553 bfqq->entity.prio_changed = 1;
5554}
5555
5556static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5557 struct bio *bio, bool is_sync,
5558 struct bfq_io_cq *bic,
5559 bool respawn);
5560
5561static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
5562{
5563 struct bfq_data *bfqd = bic_to_bfqd(bic);
5564 struct bfq_queue *bfqq;
5565 int ioprio = bic->icq.ioc->ioprio;
5566
5567 /*
5568 * This condition may trigger on a newly created bic, be sure to
5569 * drop the lock before returning.
5570 */
5571 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
5572 return;
5573
5574 bic->ioprio = ioprio;
5575
5576 bfqq = bic_to_bfqq(bic, false, bfq_actuator_index(bfqd, bio));
5577 if (bfqq) {
5578 struct bfq_queue *old_bfqq = bfqq;
5579
5580 bfqq = bfq_get_queue(bfqd, bio, false, bic, true);
5581 bic_set_bfqq(bic, bfqq, false, bfq_actuator_index(bfqd, bio));
5582 bfq_release_process_ref(bfqd, old_bfqq);
5583 }
5584
5585 bfqq = bic_to_bfqq(bic, true, bfq_actuator_index(bfqd, bio));
5586 if (bfqq)
5587 bfq_set_next_ioprio_data(bfqq, bic);
5588}
5589
5590static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5591 struct bfq_io_cq *bic, pid_t pid, int is_sync,
5592 unsigned int act_idx)
5593{
5594 u64 now_ns = ktime_get_ns();
5595
5596 bfqq->actuator_idx = act_idx;
5597 RB_CLEAR_NODE(&bfqq->entity.rb_node);
5598 INIT_LIST_HEAD(&bfqq->fifo);
5599 INIT_HLIST_NODE(&bfqq->burst_list_node);
5600 INIT_HLIST_NODE(&bfqq->woken_list_node);
5601 INIT_HLIST_HEAD(&bfqq->woken_list);
5602
5603 bfqq->ref = 0;
5604 bfqq->bfqd = bfqd;
5605
5606 if (bic)
5607 bfq_set_next_ioprio_data(bfqq, bic);
5608
5609 if (is_sync) {
5610 /*
5611 * No need to mark as has_short_ttime if in
5612 * idle_class, because no device idling is performed
5613 * for queues in idle class
5614 */
5615 if (!bfq_class_idle(bfqq))
5616 /* tentatively mark as has_short_ttime */
5617 bfq_mark_bfqq_has_short_ttime(bfqq);
5618 bfq_mark_bfqq_sync(bfqq);
5619 bfq_mark_bfqq_just_created(bfqq);
5620 } else
5621 bfq_clear_bfqq_sync(bfqq);
5622
5623 /* set end request to minus infinity from now */
5624 bfqq->ttime.last_end_request = now_ns + 1;
5625
5626 bfqq->creation_time = jiffies;
5627
5628 bfqq->io_start_time = now_ns;
5629
5630 bfq_mark_bfqq_IO_bound(bfqq);
5631
5632 bfqq->pid = pid;
5633
5634 /* Tentative initial value to trade off between thr and lat */
5635 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
5636 bfqq->budget_timeout = bfq_smallest_from_now();
5637
5638 bfqq->wr_coeff = 1;
5639 bfqq->last_wr_start_finish = jiffies;
5640 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
5641 bfqq->split_time = bfq_smallest_from_now();
5642
5643 /*
5644 * To not forget the possibly high bandwidth consumed by a
5645 * process/queue in the recent past,
5646 * bfq_bfqq_softrt_next_start() returns a value at least equal
5647 * to the current value of bfqq->soft_rt_next_start (see
5648 * comments on bfq_bfqq_softrt_next_start). Set
5649 * soft_rt_next_start to now, to mean that bfqq has consumed
5650 * no bandwidth so far.
5651 */
5652 bfqq->soft_rt_next_start = jiffies;
5653
5654 /* first request is almost certainly seeky */
5655 bfqq->seek_history = 1;
5656
5657 bfqq->decrease_time_jif = jiffies;
5658}
5659
5660static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
5661 struct bfq_group *bfqg,
5662 int ioprio_class, int ioprio, int act_idx)
5663{
5664 switch (ioprio_class) {
5665 case IOPRIO_CLASS_RT:
5666 return &bfqg->async_bfqq[0][ioprio][act_idx];
5667 case IOPRIO_CLASS_NONE:
5668 ioprio = IOPRIO_BE_NORM;
5669 fallthrough;
5670 case IOPRIO_CLASS_BE:
5671 return &bfqg->async_bfqq[1][ioprio][act_idx];
5672 case IOPRIO_CLASS_IDLE:
5673 return &bfqg->async_idle_bfqq[act_idx];
5674 default:
5675 return NULL;
5676 }
5677}
5678
5679static struct bfq_queue *
5680bfq_do_early_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5681 struct bfq_io_cq *bic,
5682 struct bfq_queue *last_bfqq_created)
5683{
5684 unsigned int a_idx = last_bfqq_created->actuator_idx;
5685 struct bfq_queue *new_bfqq =
5686 bfq_setup_merge(bfqq, last_bfqq_created);
5687
5688 if (!new_bfqq)
5689 return bfqq;
5690
5691 if (new_bfqq->bic)
5692 new_bfqq->bic->bfqq_data[a_idx].stably_merged = true;
5693 bic->bfqq_data[a_idx].stably_merged = true;
5694
5695 /*
5696 * Reusing merge functions. This implies that
5697 * bfqq->bic must be set too, for
5698 * bfq_merge_bfqqs to correctly save bfqq's
5699 * state before killing it.
5700 */
5701 bfqq->bic = bic;
5702 bfq_merge_bfqqs(bfqd, bic, bfqq, new_bfqq);
5703
5704 return new_bfqq;
5705}
5706
5707/*
5708 * Many throughput-sensitive workloads are made of several parallel
5709 * I/O flows, with all flows generated by the same application, or
5710 * more generically by the same task (e.g., system boot). The most
5711 * counterproductive action with these workloads is plugging I/O
5712 * dispatch when one of the bfq_queues associated with these flows
5713 * remains temporarily empty.
5714 *
5715 * To avoid this plugging, BFQ has been using a burst-handling
5716 * mechanism for years now. This mechanism has proven effective for
5717 * throughput, and not detrimental for service guarantees. The
5718 * following function pushes this mechanism a little bit further,
5719 * basing on the following two facts.
5720 *
5721 * First, all the I/O flows of a the same application or task
5722 * contribute to the execution/completion of that common application
5723 * or task. So the performance figures that matter are total
5724 * throughput of the flows and task-wide I/O latency. In particular,
5725 * these flows do not need to be protected from each other, in terms
5726 * of individual bandwidth or latency.
5727 *
5728 * Second, the above fact holds regardless of the number of flows.
5729 *
5730 * Putting these two facts together, this commits merges stably the
5731 * bfq_queues associated with these I/O flows, i.e., with the
5732 * processes that generate these IO/ flows, regardless of how many the
5733 * involved processes are.
5734 *
5735 * To decide whether a set of bfq_queues is actually associated with
5736 * the I/O flows of a common application or task, and to merge these
5737 * queues stably, this function operates as follows: given a bfq_queue,
5738 * say Q2, currently being created, and the last bfq_queue, say Q1,
5739 * created before Q2, Q2 is merged stably with Q1 if
5740 * - very little time has elapsed since when Q1 was created
5741 * - Q2 has the same ioprio as Q1
5742 * - Q2 belongs to the same group as Q1
5743 *
5744 * Merging bfq_queues also reduces scheduling overhead. A fio test
5745 * with ten random readers on /dev/nullb shows a throughput boost of
5746 * 40%, with a quadcore. Since BFQ's execution time amounts to ~50% of
5747 * the total per-request processing time, the above throughput boost
5748 * implies that BFQ's overhead is reduced by more than 50%.
5749 *
5750 * This new mechanism most certainly obsoletes the current
5751 * burst-handling heuristics. We keep those heuristics for the moment.
5752 */
5753static struct bfq_queue *bfq_do_or_sched_stable_merge(struct bfq_data *bfqd,
5754 struct bfq_queue *bfqq,
5755 struct bfq_io_cq *bic)
5756{
5757 struct bfq_queue **source_bfqq = bfqq->entity.parent ?
5758 &bfqq->entity.parent->last_bfqq_created :
5759 &bfqd->last_bfqq_created;
5760
5761 struct bfq_queue *last_bfqq_created = *source_bfqq;
5762
5763 /*
5764 * If last_bfqq_created has not been set yet, then init it. If
5765 * it has been set already, but too long ago, then move it
5766 * forward to bfqq. Finally, move also if bfqq belongs to a
5767 * different group than last_bfqq_created, or if bfqq has a
5768 * different ioprio, ioprio_class or actuator_idx. If none of
5769 * these conditions holds true, then try an early stable merge
5770 * or schedule a delayed stable merge. As for the condition on
5771 * actuator_idx, the reason is that, if queues associated with
5772 * different actuators are merged, then control is lost on
5773 * each actuator. Therefore some actuator may be
5774 * underutilized, and throughput may decrease.
5775 *
5776 * A delayed merge is scheduled (instead of performing an
5777 * early merge), in case bfqq might soon prove to be more
5778 * throughput-beneficial if not merged. Currently this is
5779 * possible only if bfqd is rotational with no queueing. For
5780 * such a drive, not merging bfqq is better for throughput if
5781 * bfqq happens to contain sequential I/O. So, we wait a
5782 * little bit for enough I/O to flow through bfqq. After that,
5783 * if such an I/O is sequential, then the merge is
5784 * canceled. Otherwise the merge is finally performed.
5785 */
5786 if (!last_bfqq_created ||
5787 time_before(last_bfqq_created->creation_time +
5788 msecs_to_jiffies(bfq_activation_stable_merging),
5789 bfqq->creation_time) ||
5790 bfqq->entity.parent != last_bfqq_created->entity.parent ||
5791 bfqq->ioprio != last_bfqq_created->ioprio ||
5792 bfqq->ioprio_class != last_bfqq_created->ioprio_class ||
5793 bfqq->actuator_idx != last_bfqq_created->actuator_idx)
5794 *source_bfqq = bfqq;
5795 else if (time_after_eq(last_bfqq_created->creation_time +
5796 bfqd->bfq_burst_interval,
5797 bfqq->creation_time)) {
5798 if (likely(bfqd->nonrot_with_queueing))
5799 /*
5800 * With this type of drive, leaving
5801 * bfqq alone may provide no
5802 * throughput benefits compared with
5803 * merging bfqq. So merge bfqq now.
5804 */
5805 bfqq = bfq_do_early_stable_merge(bfqd, bfqq,
5806 bic,
5807 last_bfqq_created);
5808 else { /* schedule tentative stable merge */
5809 /*
5810 * get reference on last_bfqq_created,
5811 * to prevent it from being freed,
5812 * until we decide whether to merge
5813 */
5814 last_bfqq_created->ref++;
5815 /*
5816 * need to keep track of stable refs, to
5817 * compute process refs correctly
5818 */
5819 last_bfqq_created->stable_ref++;
5820 /*
5821 * Record the bfqq to merge to.
5822 */
5823 bic->bfqq_data[last_bfqq_created->actuator_idx].stable_merge_bfqq =
5824 last_bfqq_created;
5825 }
5826 }
5827
5828 return bfqq;
5829}
5830
5831
5832static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5833 struct bio *bio, bool is_sync,
5834 struct bfq_io_cq *bic,
5835 bool respawn)
5836{
5837 const int ioprio = IOPRIO_PRIO_LEVEL(bic->ioprio);
5838 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5839 struct bfq_queue **async_bfqq = NULL;
5840 struct bfq_queue *bfqq;
5841 struct bfq_group *bfqg;
5842
5843 bfqg = bfq_bio_bfqg(bfqd, bio);
5844 if (!is_sync) {
5845 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
5846 ioprio,
5847 bfq_actuator_index(bfqd, bio));
5848 bfqq = *async_bfqq;
5849 if (bfqq)
5850 goto out;
5851 }
5852
5853 bfqq = kmem_cache_alloc_node(bfq_pool,
5854 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
5855 bfqd->queue->node);
5856
5857 if (bfqq) {
5858 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
5859 is_sync, bfq_actuator_index(bfqd, bio));
5860 bfq_init_entity(&bfqq->entity, bfqg);
5861 bfq_log_bfqq(bfqd, bfqq, "allocated");
5862 } else {
5863 bfqq = &bfqd->oom_bfqq;
5864 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
5865 goto out;
5866 }
5867
5868 /*
5869 * Pin the queue now that it's allocated, scheduler exit will
5870 * prune it.
5871 */
5872 if (async_bfqq) {
5873 bfqq->ref++; /*
5874 * Extra group reference, w.r.t. sync
5875 * queue. This extra reference is removed
5876 * only if bfqq->bfqg disappears, to
5877 * guarantee that this queue is not freed
5878 * until its group goes away.
5879 */
5880 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
5881 bfqq, bfqq->ref);
5882 *async_bfqq = bfqq;
5883 }
5884
5885out:
5886 bfqq->ref++; /* get a process reference to this queue */
5887
5888 if (bfqq != &bfqd->oom_bfqq && is_sync && !respawn)
5889 bfqq = bfq_do_or_sched_stable_merge(bfqd, bfqq, bic);
5890 return bfqq;
5891}
5892
5893static void bfq_update_io_thinktime(struct bfq_data *bfqd,
5894 struct bfq_queue *bfqq)
5895{
5896 struct bfq_ttime *ttime = &bfqq->ttime;
5897 u64 elapsed;
5898
5899 /*
5900 * We are really interested in how long it takes for the queue to
5901 * become busy when there is no outstanding IO for this queue. So
5902 * ignore cases when the bfq queue has already IO queued.
5903 */
5904 if (bfqq->dispatched || bfq_bfqq_busy(bfqq))
5905 return;
5906 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
5907 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
5908
5909 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
5910 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
5911 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
5912 ttime->ttime_samples);
5913}
5914
5915static void
5916bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5917 struct request *rq)
5918{
5919 bfqq->seek_history <<= 1;
5920 bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
5921
5922 if (bfqq->wr_coeff > 1 &&
5923 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
5924 BFQQ_TOTALLY_SEEKY(bfqq)) {
5925 if (time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
5926 bfq_wr_duration(bfqd))) {
5927 /*
5928 * In soft_rt weight raising with the
5929 * interactive-weight-raising period
5930 * elapsed (so no switch back to
5931 * interactive weight raising).
5932 */
5933 bfq_bfqq_end_wr(bfqq);
5934 } else { /*
5935 * stopping soft_rt weight raising
5936 * while still in interactive period,
5937 * switch back to interactive weight
5938 * raising
5939 */
5940 switch_back_to_interactive_wr(bfqq, bfqd);
5941 bfqq->entity.prio_changed = 1;
5942 }
5943 }
5944}
5945
5946static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
5947 struct bfq_queue *bfqq,
5948 struct bfq_io_cq *bic)
5949{
5950 bool has_short_ttime = true, state_changed;
5951
5952 /*
5953 * No need to update has_short_ttime if bfqq is async or in
5954 * idle io prio class, or if bfq_slice_idle is zero, because
5955 * no device idling is performed for bfqq in this case.
5956 */
5957 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
5958 bfqd->bfq_slice_idle == 0)
5959 return;
5960
5961 /* Idle window just restored, statistics are meaningless. */
5962 if (time_is_after_eq_jiffies(bfqq->split_time +
5963 bfqd->bfq_wr_min_idle_time))
5964 return;
5965
5966 /* Think time is infinite if no process is linked to
5967 * bfqq. Otherwise check average think time to decide whether
5968 * to mark as has_short_ttime. To this goal, compare average
5969 * think time with half the I/O-plugging timeout.
5970 */
5971 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
5972 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
5973 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle>>1))
5974 has_short_ttime = false;
5975
5976 state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
5977
5978 if (has_short_ttime)
5979 bfq_mark_bfqq_has_short_ttime(bfqq);
5980 else
5981 bfq_clear_bfqq_has_short_ttime(bfqq);
5982
5983 /*
5984 * Until the base value for the total service time gets
5985 * finally computed for bfqq, the inject limit does depend on
5986 * the think-time state (short|long). In particular, the limit
5987 * is 0 or 1 if the think time is deemed, respectively, as
5988 * short or long (details in the comments in
5989 * bfq_update_inject_limit()). Accordingly, the next
5990 * instructions reset the inject limit if the think-time state
5991 * has changed and the above base value is still to be
5992 * computed.
5993 *
5994 * However, the reset is performed only if more than 100 ms
5995 * have elapsed since the last update of the inject limit, or
5996 * (inclusive) if the change is from short to long think
5997 * time. The reason for this waiting is as follows.
5998 *
5999 * bfqq may have a long think time because of a
6000 * synchronization with some other queue, i.e., because the
6001 * I/O of some other queue may need to be completed for bfqq
6002 * to receive new I/O. Details in the comments on the choice
6003 * of the queue for injection in bfq_select_queue().
6004 *
6005 * As stressed in those comments, if such a synchronization is
6006 * actually in place, then, without injection on bfqq, the
6007 * blocking I/O cannot happen to served while bfqq is in
6008 * service. As a consequence, if bfqq is granted
6009 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
6010 * is dispatched, until the idle timeout fires. This is likely
6011 * to result in lower bandwidth and higher latencies for bfqq,
6012 * and in a severe loss of total throughput.
6013 *
6014 * On the opposite end, a non-zero inject limit may allow the
6015 * I/O that blocks bfqq to be executed soon, and therefore
6016 * bfqq to receive new I/O soon.
6017 *
6018 * But, if the blocking gets actually eliminated, then the
6019 * next think-time sample for bfqq may be very low. This in
6020 * turn may cause bfqq's think time to be deemed
6021 * short. Without the 100 ms barrier, this new state change
6022 * would cause the body of the next if to be executed
6023 * immediately. But this would set to 0 the inject
6024 * limit. Without injection, the blocking I/O would cause the
6025 * think time of bfqq to become long again, and therefore the
6026 * inject limit to be raised again, and so on. The only effect
6027 * of such a steady oscillation between the two think-time
6028 * states would be to prevent effective injection on bfqq.
6029 *
6030 * In contrast, if the inject limit is not reset during such a
6031 * long time interval as 100 ms, then the number of short
6032 * think time samples can grow significantly before the reset
6033 * is performed. As a consequence, the think time state can
6034 * become stable before the reset. Therefore there will be no
6035 * state change when the 100 ms elapse, and no reset of the
6036 * inject limit. The inject limit remains steadily equal to 1
6037 * both during and after the 100 ms. So injection can be
6038 * performed at all times, and throughput gets boosted.
6039 *
6040 * An inject limit equal to 1 is however in conflict, in
6041 * general, with the fact that the think time of bfqq is
6042 * short, because injection may be likely to delay bfqq's I/O
6043 * (as explained in the comments in
6044 * bfq_update_inject_limit()). But this does not happen in
6045 * this special case, because bfqq's low think time is due to
6046 * an effective handling of a synchronization, through
6047 * injection. In this special case, bfqq's I/O does not get
6048 * delayed by injection; on the contrary, bfqq's I/O is
6049 * brought forward, because it is not blocked for
6050 * milliseconds.
6051 *
6052 * In addition, serving the blocking I/O much sooner, and much
6053 * more frequently than once per I/O-plugging timeout, makes
6054 * it much quicker to detect a waker queue (the concept of
6055 * waker queue is defined in the comments in
6056 * bfq_add_request()). This makes it possible to start sooner
6057 * to boost throughput more effectively, by injecting the I/O
6058 * of the waker queue unconditionally on every
6059 * bfq_dispatch_request().
6060 *
6061 * One last, important benefit of not resetting the inject
6062 * limit before 100 ms is that, during this time interval, the
6063 * base value for the total service time is likely to get
6064 * finally computed for bfqq, freeing the inject limit from
6065 * its relation with the think time.
6066 */
6067 if (state_changed && bfqq->last_serv_time_ns == 0 &&
6068 (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
6069 msecs_to_jiffies(100)) ||
6070 !has_short_ttime))
6071 bfq_reset_inject_limit(bfqd, bfqq);
6072}
6073
6074/*
6075 * Called when a new fs request (rq) is added to bfqq. Check if there's
6076 * something we should do about it.
6077 */
6078static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
6079 struct request *rq)
6080{
6081 if (rq->cmd_flags & REQ_META)
6082 bfqq->meta_pending++;
6083
6084 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
6085
6086 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
6087 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
6088 blk_rq_sectors(rq) < 32;
6089 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
6090
6091 /*
6092 * There is just this request queued: if
6093 * - the request is small, and
6094 * - we are idling to boost throughput, and
6095 * - the queue is not to be expired,
6096 * then just exit.
6097 *
6098 * In this way, if the device is being idled to wait
6099 * for a new request from the in-service queue, we
6100 * avoid unplugging the device and committing the
6101 * device to serve just a small request. In contrast
6102 * we wait for the block layer to decide when to
6103 * unplug the device: hopefully, new requests will be
6104 * merged to this one quickly, then the device will be
6105 * unplugged and larger requests will be dispatched.
6106 */
6107 if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
6108 !budget_timeout)
6109 return;
6110
6111 /*
6112 * A large enough request arrived, or idling is being
6113 * performed to preserve service guarantees, or
6114 * finally the queue is to be expired: in all these
6115 * cases disk idling is to be stopped, so clear
6116 * wait_request flag and reset timer.
6117 */
6118 bfq_clear_bfqq_wait_request(bfqq);
6119 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
6120
6121 /*
6122 * The queue is not empty, because a new request just
6123 * arrived. Hence we can safely expire the queue, in
6124 * case of budget timeout, without risking that the
6125 * timestamps of the queue are not updated correctly.
6126 * See [1] for more details.
6127 */
6128 if (budget_timeout)
6129 bfq_bfqq_expire(bfqd, bfqq, false,
6130 BFQQE_BUDGET_TIMEOUT);
6131 }
6132}
6133
6134static void bfqq_request_allocated(struct bfq_queue *bfqq)
6135{
6136 struct bfq_entity *entity = &bfqq->entity;
6137
6138 for_each_entity(entity)
6139 entity->allocated++;
6140}
6141
6142static void bfqq_request_freed(struct bfq_queue *bfqq)
6143{
6144 struct bfq_entity *entity = &bfqq->entity;
6145
6146 for_each_entity(entity)
6147 entity->allocated--;
6148}
6149
6150/* returns true if it causes the idle timer to be disabled */
6151static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
6152{
6153 struct bfq_queue *bfqq = RQ_BFQQ(rq),
6154 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true,
6155 RQ_BIC(rq));
6156 bool waiting, idle_timer_disabled = false;
6157
6158 if (new_bfqq) {
6159 /*
6160 * Release the request's reference to the old bfqq
6161 * and make sure one is taken to the shared queue.
6162 */
6163 bfqq_request_allocated(new_bfqq);
6164 bfqq_request_freed(bfqq);
6165 new_bfqq->ref++;
6166 /*
6167 * If the bic associated with the process
6168 * issuing this request still points to bfqq
6169 * (and thus has not been already redirected
6170 * to new_bfqq or even some other bfq_queue),
6171 * then complete the merge and redirect it to
6172 * new_bfqq.
6173 */
6174 if (bic_to_bfqq(RQ_BIC(rq), true,
6175 bfq_actuator_index(bfqd, rq->bio)) == bfqq)
6176 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
6177 bfqq, new_bfqq);
6178
6179 bfq_clear_bfqq_just_created(bfqq);
6180 /*
6181 * rq is about to be enqueued into new_bfqq,
6182 * release rq reference on bfqq
6183 */
6184 bfq_put_queue(bfqq);
6185 rq->elv.priv[1] = new_bfqq;
6186 bfqq = new_bfqq;
6187 }
6188
6189 bfq_update_io_thinktime(bfqd, bfqq);
6190 bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
6191 bfq_update_io_seektime(bfqd, bfqq, rq);
6192
6193 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
6194 bfq_add_request(rq);
6195 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
6196
6197 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
6198 list_add_tail(&rq->queuelist, &bfqq->fifo);
6199
6200 bfq_rq_enqueued(bfqd, bfqq, rq);
6201
6202 return idle_timer_disabled;
6203}
6204
6205#ifdef CONFIG_BFQ_CGROUP_DEBUG
6206static void bfq_update_insert_stats(struct request_queue *q,
6207 struct bfq_queue *bfqq,
6208 bool idle_timer_disabled,
6209 blk_opf_t cmd_flags)
6210{
6211 if (!bfqq)
6212 return;
6213
6214 /*
6215 * bfqq still exists, because it can disappear only after
6216 * either it is merged with another queue, or the process it
6217 * is associated with exits. But both actions must be taken by
6218 * the same process currently executing this flow of
6219 * instructions.
6220 *
6221 * In addition, the following queue lock guarantees that
6222 * bfqq_group(bfqq) exists as well.
6223 */
6224 spin_lock_irq(&q->queue_lock);
6225 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
6226 if (idle_timer_disabled)
6227 bfqg_stats_update_idle_time(bfqq_group(bfqq));
6228 spin_unlock_irq(&q->queue_lock);
6229}
6230#else
6231static inline void bfq_update_insert_stats(struct request_queue *q,
6232 struct bfq_queue *bfqq,
6233 bool idle_timer_disabled,
6234 blk_opf_t cmd_flags) {}
6235#endif /* CONFIG_BFQ_CGROUP_DEBUG */
6236
6237static struct bfq_queue *bfq_init_rq(struct request *rq);
6238
6239static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
6240 blk_insert_t flags)
6241{
6242 struct request_queue *q = hctx->queue;
6243 struct bfq_data *bfqd = q->elevator->elevator_data;
6244 struct bfq_queue *bfqq;
6245 bool idle_timer_disabled = false;
6246 blk_opf_t cmd_flags;
6247 LIST_HEAD(free);
6248
6249#ifdef CONFIG_BFQ_GROUP_IOSCHED
6250 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) && rq->bio)
6251 bfqg_stats_update_legacy_io(q, rq);
6252#endif
6253 spin_lock_irq(&bfqd->lock);
6254 bfqq = bfq_init_rq(rq);
6255 if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
6256 spin_unlock_irq(&bfqd->lock);
6257 blk_mq_free_requests(&free);
6258 return;
6259 }
6260
6261 trace_block_rq_insert(rq);
6262
6263 if (flags & BLK_MQ_INSERT_AT_HEAD) {
6264 list_add(&rq->queuelist, &bfqd->dispatch);
6265 } else if (!bfqq) {
6266 list_add_tail(&rq->queuelist, &bfqd->dispatch);
6267 } else {
6268 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
6269 /*
6270 * Update bfqq, because, if a queue merge has occurred
6271 * in __bfq_insert_request, then rq has been
6272 * redirected into a new queue.
6273 */
6274 bfqq = RQ_BFQQ(rq);
6275
6276 if (rq_mergeable(rq)) {
6277 elv_rqhash_add(q, rq);
6278 if (!q->last_merge)
6279 q->last_merge = rq;
6280 }
6281 }
6282
6283 /*
6284 * Cache cmd_flags before releasing scheduler lock, because rq
6285 * may disappear afterwards (for example, because of a request
6286 * merge).
6287 */
6288 cmd_flags = rq->cmd_flags;
6289 spin_unlock_irq(&bfqd->lock);
6290
6291 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
6292 cmd_flags);
6293}
6294
6295static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
6296 struct list_head *list,
6297 blk_insert_t flags)
6298{
6299 while (!list_empty(list)) {
6300 struct request *rq;
6301
6302 rq = list_first_entry(list, struct request, queuelist);
6303 list_del_init(&rq->queuelist);
6304 bfq_insert_request(hctx, rq, flags);
6305 }
6306}
6307
6308static void bfq_update_hw_tag(struct bfq_data *bfqd)
6309{
6310 struct bfq_queue *bfqq = bfqd->in_service_queue;
6311
6312 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
6313 bfqd->tot_rq_in_driver);
6314
6315 if (bfqd->hw_tag == 1)
6316 return;
6317
6318 /*
6319 * This sample is valid if the number of outstanding requests
6320 * is large enough to allow a queueing behavior. Note that the
6321 * sum is not exact, as it's not taking into account deactivated
6322 * requests.
6323 */
6324 if (bfqd->tot_rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
6325 return;
6326
6327 /*
6328 * If active queue hasn't enough requests and can idle, bfq might not
6329 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
6330 * case
6331 */
6332 if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
6333 bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
6334 BFQ_HW_QUEUE_THRESHOLD &&
6335 bfqd->tot_rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
6336 return;
6337
6338 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
6339 return;
6340
6341 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
6342 bfqd->max_rq_in_driver = 0;
6343 bfqd->hw_tag_samples = 0;
6344
6345 bfqd->nonrot_with_queueing =
6346 blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
6347}
6348
6349static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
6350{
6351 u64 now_ns;
6352 u32 delta_us;
6353
6354 bfq_update_hw_tag(bfqd);
6355
6356 bfqd->rq_in_driver[bfqq->actuator_idx]--;
6357 bfqd->tot_rq_in_driver--;
6358 bfqq->dispatched--;
6359
6360 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
6361 /*
6362 * Set budget_timeout (which we overload to store the
6363 * time at which the queue remains with no backlog and
6364 * no outstanding request; used by the weight-raising
6365 * mechanism).
6366 */
6367 bfqq->budget_timeout = jiffies;
6368
6369 bfq_del_bfqq_in_groups_with_pending_reqs(bfqq);
6370 bfq_weights_tree_remove(bfqq);
6371 }
6372
6373 now_ns = ktime_get_ns();
6374
6375 bfqq->ttime.last_end_request = now_ns;
6376
6377 /*
6378 * Using us instead of ns, to get a reasonable precision in
6379 * computing rate in next check.
6380 */
6381 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
6382
6383 /*
6384 * If the request took rather long to complete, and, according
6385 * to the maximum request size recorded, this completion latency
6386 * implies that the request was certainly served at a very low
6387 * rate (less than 1M sectors/sec), then the whole observation
6388 * interval that lasts up to this time instant cannot be a
6389 * valid time interval for computing a new peak rate. Invoke
6390 * bfq_update_rate_reset to have the following three steps
6391 * taken:
6392 * - close the observation interval at the last (previous)
6393 * request dispatch or completion
6394 * - compute rate, if possible, for that observation interval
6395 * - reset to zero samples, which will trigger a proper
6396 * re-initialization of the observation interval on next
6397 * dispatch
6398 */
6399 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
6400 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
6401 1UL<<(BFQ_RATE_SHIFT - 10))
6402 bfq_update_rate_reset(bfqd, NULL);
6403 bfqd->last_completion = now_ns;
6404 /*
6405 * Shared queues are likely to receive I/O at a high
6406 * rate. This may deceptively let them be considered as wakers
6407 * of other queues. But a false waker will unjustly steal
6408 * bandwidth to its supposedly woken queue. So considering
6409 * also shared queues in the waking mechanism may cause more
6410 * control troubles than throughput benefits. Then reset
6411 * last_completed_rq_bfqq if bfqq is a shared queue.
6412 */
6413 if (!bfq_bfqq_coop(bfqq))
6414 bfqd->last_completed_rq_bfqq = bfqq;
6415 else
6416 bfqd->last_completed_rq_bfqq = NULL;
6417
6418 /*
6419 * If we are waiting to discover whether the request pattern
6420 * of the task associated with the queue is actually
6421 * isochronous, and both requisites for this condition to hold
6422 * are now satisfied, then compute soft_rt_next_start (see the
6423 * comments on the function bfq_bfqq_softrt_next_start()). We
6424 * do not compute soft_rt_next_start if bfqq is in interactive
6425 * weight raising (see the comments in bfq_bfqq_expire() for
6426 * an explanation). We schedule this delayed update when bfqq
6427 * expires, if it still has in-flight requests.
6428 */
6429 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
6430 RB_EMPTY_ROOT(&bfqq->sort_list) &&
6431 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
6432 bfqq->soft_rt_next_start =
6433 bfq_bfqq_softrt_next_start(bfqd, bfqq);
6434
6435 /*
6436 * If this is the in-service queue, check if it needs to be expired,
6437 * or if we want to idle in case it has no pending requests.
6438 */
6439 if (bfqd->in_service_queue == bfqq) {
6440 if (bfq_bfqq_must_idle(bfqq)) {
6441 if (bfqq->dispatched == 0)
6442 bfq_arm_slice_timer(bfqd);
6443 /*
6444 * If we get here, we do not expire bfqq, even
6445 * if bfqq was in budget timeout or had no
6446 * more requests (as controlled in the next
6447 * conditional instructions). The reason for
6448 * not expiring bfqq is as follows.
6449 *
6450 * Here bfqq->dispatched > 0 holds, but
6451 * bfq_bfqq_must_idle() returned true. This
6452 * implies that, even if no request arrives
6453 * for bfqq before bfqq->dispatched reaches 0,
6454 * bfqq will, however, not be expired on the
6455 * completion event that causes bfqq->dispatch
6456 * to reach zero. In contrast, on this event,
6457 * bfqq will start enjoying device idling
6458 * (I/O-dispatch plugging).
6459 *
6460 * But, if we expired bfqq here, bfqq would
6461 * not have the chance to enjoy device idling
6462 * when bfqq->dispatched finally reaches
6463 * zero. This would expose bfqq to violation
6464 * of its reserved service guarantees.
6465 */
6466 return;
6467 } else if (bfq_may_expire_for_budg_timeout(bfqq))
6468 bfq_bfqq_expire(bfqd, bfqq, false,
6469 BFQQE_BUDGET_TIMEOUT);
6470 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
6471 (bfqq->dispatched == 0 ||
6472 !bfq_better_to_idle(bfqq)))
6473 bfq_bfqq_expire(bfqd, bfqq, false,
6474 BFQQE_NO_MORE_REQUESTS);
6475 }
6476
6477 if (!bfqd->tot_rq_in_driver)
6478 bfq_schedule_dispatch(bfqd);
6479}
6480
6481/*
6482 * The processes associated with bfqq may happen to generate their
6483 * cumulative I/O at a lower rate than the rate at which the device
6484 * could serve the same I/O. This is rather probable, e.g., if only
6485 * one process is associated with bfqq and the device is an SSD. It
6486 * results in bfqq becoming often empty while in service. In this
6487 * respect, if BFQ is allowed to switch to another queue when bfqq
6488 * remains empty, then the device goes on being fed with I/O requests,
6489 * and the throughput is not affected. In contrast, if BFQ is not
6490 * allowed to switch to another queue---because bfqq is sync and
6491 * I/O-dispatch needs to be plugged while bfqq is temporarily
6492 * empty---then, during the service of bfqq, there will be frequent
6493 * "service holes", i.e., time intervals during which bfqq gets empty
6494 * and the device can only consume the I/O already queued in its
6495 * hardware queues. During service holes, the device may even get to
6496 * remaining idle. In the end, during the service of bfqq, the device
6497 * is driven at a lower speed than the one it can reach with the kind
6498 * of I/O flowing through bfqq.
6499 *
6500 * To counter this loss of throughput, BFQ implements a "request
6501 * injection mechanism", which tries to fill the above service holes
6502 * with I/O requests taken from other queues. The hard part in this
6503 * mechanism is finding the right amount of I/O to inject, so as to
6504 * both boost throughput and not break bfqq's bandwidth and latency
6505 * guarantees. In this respect, the mechanism maintains a per-queue
6506 * inject limit, computed as below. While bfqq is empty, the injection
6507 * mechanism dispatches extra I/O requests only until the total number
6508 * of I/O requests in flight---i.e., already dispatched but not yet
6509 * completed---remains lower than this limit.
6510 *
6511 * A first definition comes in handy to introduce the algorithm by
6512 * which the inject limit is computed. We define as first request for
6513 * bfqq, an I/O request for bfqq that arrives while bfqq is in
6514 * service, and causes bfqq to switch from empty to non-empty. The
6515 * algorithm updates the limit as a function of the effect of
6516 * injection on the service times of only the first requests of
6517 * bfqq. The reason for this restriction is that these are the
6518 * requests whose service time is affected most, because they are the
6519 * first to arrive after injection possibly occurred.
6520 *
6521 * To evaluate the effect of injection, the algorithm measures the
6522 * "total service time" of first requests. We define as total service
6523 * time of an I/O request, the time that elapses since when the
6524 * request is enqueued into bfqq, to when it is completed. This
6525 * quantity allows the whole effect of injection to be measured. It is
6526 * easy to see why. Suppose that some requests of other queues are
6527 * actually injected while bfqq is empty, and that a new request R
6528 * then arrives for bfqq. If the device does start to serve all or
6529 * part of the injected requests during the service hole, then,
6530 * because of this extra service, it may delay the next invocation of
6531 * the dispatch hook of BFQ. Then, even after R gets eventually
6532 * dispatched, the device may delay the actual service of R if it is
6533 * still busy serving the extra requests, or if it decides to serve,
6534 * before R, some extra request still present in its queues. As a
6535 * conclusion, the cumulative extra delay caused by injection can be
6536 * easily evaluated by just comparing the total service time of first
6537 * requests with and without injection.
6538 *
6539 * The limit-update algorithm works as follows. On the arrival of a
6540 * first request of bfqq, the algorithm measures the total time of the
6541 * request only if one of the three cases below holds, and, for each
6542 * case, it updates the limit as described below:
6543 *
6544 * (1) If there is no in-flight request. This gives a baseline for the
6545 * total service time of the requests of bfqq. If the baseline has
6546 * not been computed yet, then, after computing it, the limit is
6547 * set to 1, to start boosting throughput, and to prepare the
6548 * ground for the next case. If the baseline has already been
6549 * computed, then it is updated, in case it results to be lower
6550 * than the previous value.
6551 *
6552 * (2) If the limit is higher than 0 and there are in-flight
6553 * requests. By comparing the total service time in this case with
6554 * the above baseline, it is possible to know at which extent the
6555 * current value of the limit is inflating the total service
6556 * time. If the inflation is below a certain threshold, then bfqq
6557 * is assumed to be suffering from no perceivable loss of its
6558 * service guarantees, and the limit is even tentatively
6559 * increased. If the inflation is above the threshold, then the
6560 * limit is decreased. Due to the lack of any hysteresis, this
6561 * logic makes the limit oscillate even in steady workload
6562 * conditions. Yet we opted for it, because it is fast in reaching
6563 * the best value for the limit, as a function of the current I/O
6564 * workload. To reduce oscillations, this step is disabled for a
6565 * short time interval after the limit happens to be decreased.
6566 *
6567 * (3) Periodically, after resetting the limit, to make sure that the
6568 * limit eventually drops in case the workload changes. This is
6569 * needed because, after the limit has gone safely up for a
6570 * certain workload, it is impossible to guess whether the
6571 * baseline total service time may have changed, without measuring
6572 * it again without injection. A more effective version of this
6573 * step might be to just sample the baseline, by interrupting
6574 * injection only once, and then to reset/lower the limit only if
6575 * the total service time with the current limit does happen to be
6576 * too large.
6577 *
6578 * More details on each step are provided in the comments on the
6579 * pieces of code that implement these steps: the branch handling the
6580 * transition from empty to non empty in bfq_add_request(), the branch
6581 * handling injection in bfq_select_queue(), and the function
6582 * bfq_choose_bfqq_for_injection(). These comments also explain some
6583 * exceptions, made by the injection mechanism in some special cases.
6584 */
6585static void bfq_update_inject_limit(struct bfq_data *bfqd,
6586 struct bfq_queue *bfqq)
6587{
6588 u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns;
6589 unsigned int old_limit = bfqq->inject_limit;
6590
6591 if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) {
6592 u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
6593
6594 if (tot_time_ns >= threshold && old_limit > 0) {
6595 bfqq->inject_limit--;
6596 bfqq->decrease_time_jif = jiffies;
6597 } else if (tot_time_ns < threshold &&
6598 old_limit <= bfqd->max_rq_in_driver)
6599 bfqq->inject_limit++;
6600 }
6601
6602 /*
6603 * Either we still have to compute the base value for the
6604 * total service time, and there seem to be the right
6605 * conditions to do it, or we can lower the last base value
6606 * computed.
6607 *
6608 * NOTE: (bfqd->tot_rq_in_driver == 1) means that there is no I/O
6609 * request in flight, because this function is in the code
6610 * path that handles the completion of a request of bfqq, and,
6611 * in particular, this function is executed before
6612 * bfqd->tot_rq_in_driver is decremented in such a code path.
6613 */
6614 if ((bfqq->last_serv_time_ns == 0 && bfqd->tot_rq_in_driver == 1) ||
6615 tot_time_ns < bfqq->last_serv_time_ns) {
6616 if (bfqq->last_serv_time_ns == 0) {
6617 /*
6618 * Now we certainly have a base value: make sure we
6619 * start trying injection.
6620 */
6621 bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
6622 }
6623 bfqq->last_serv_time_ns = tot_time_ns;
6624 } else if (!bfqd->rqs_injected && bfqd->tot_rq_in_driver == 1)
6625 /*
6626 * No I/O injected and no request still in service in
6627 * the drive: these are the exact conditions for
6628 * computing the base value of the total service time
6629 * for bfqq. So let's update this value, because it is
6630 * rather variable. For example, it varies if the size
6631 * or the spatial locality of the I/O requests in bfqq
6632 * change.
6633 */
6634 bfqq->last_serv_time_ns = tot_time_ns;
6635
6636
6637 /* update complete, not waiting for any request completion any longer */
6638 bfqd->waited_rq = NULL;
6639 bfqd->rqs_injected = false;
6640}
6641
6642/*
6643 * Handle either a requeue or a finish for rq. The things to do are
6644 * the same in both cases: all references to rq are to be dropped. In
6645 * particular, rq is considered completed from the point of view of
6646 * the scheduler.
6647 */
6648static void bfq_finish_requeue_request(struct request *rq)
6649{
6650 struct bfq_queue *bfqq = RQ_BFQQ(rq);
6651 struct bfq_data *bfqd;
6652 unsigned long flags;
6653
6654 /*
6655 * rq either is not associated with any icq, or is an already
6656 * requeued request that has not (yet) been re-inserted into
6657 * a bfq_queue.
6658 */
6659 if (!rq->elv.icq || !bfqq)
6660 return;
6661
6662 bfqd = bfqq->bfqd;
6663
6664 if (rq->rq_flags & RQF_STARTED)
6665 bfqg_stats_update_completion(bfqq_group(bfqq),
6666 rq->start_time_ns,
6667 rq->io_start_time_ns,
6668 rq->cmd_flags);
6669
6670 spin_lock_irqsave(&bfqd->lock, flags);
6671 if (likely(rq->rq_flags & RQF_STARTED)) {
6672 if (rq == bfqd->waited_rq)
6673 bfq_update_inject_limit(bfqd, bfqq);
6674
6675 bfq_completed_request(bfqq, bfqd);
6676 }
6677 bfqq_request_freed(bfqq);
6678 bfq_put_queue(bfqq);
6679 RQ_BIC(rq)->requests--;
6680 spin_unlock_irqrestore(&bfqd->lock, flags);
6681
6682 /*
6683 * Reset private fields. In case of a requeue, this allows
6684 * this function to correctly do nothing if it is spuriously
6685 * invoked again on this same request (see the check at the
6686 * beginning of the function). Probably, a better general
6687 * design would be to prevent blk-mq from invoking the requeue
6688 * or finish hooks of an elevator, for a request that is not
6689 * referred by that elevator.
6690 *
6691 * Resetting the following fields would break the
6692 * request-insertion logic if rq is re-inserted into a bfq
6693 * internal queue, without a re-preparation. Here we assume
6694 * that re-insertions of requeued requests, without
6695 * re-preparation, can happen only for pass_through or at_head
6696 * requests (which are not re-inserted into bfq internal
6697 * queues).
6698 */
6699 rq->elv.priv[0] = NULL;
6700 rq->elv.priv[1] = NULL;
6701}
6702
6703static void bfq_finish_request(struct request *rq)
6704{
6705 bfq_finish_requeue_request(rq);
6706
6707 if (rq->elv.icq) {
6708 put_io_context(rq->elv.icq->ioc);
6709 rq->elv.icq = NULL;
6710 }
6711}
6712
6713/*
6714 * Removes the association between the current task and bfqq, assuming
6715 * that bic points to the bfq iocontext of the task.
6716 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
6717 * was the last process referring to that bfqq.
6718 */
6719static struct bfq_queue *
6720bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
6721{
6722 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
6723
6724 if (bfqq_process_refs(bfqq) == 1) {
6725 bfqq->pid = current->pid;
6726 bfq_clear_bfqq_coop(bfqq);
6727 bfq_clear_bfqq_split_coop(bfqq);
6728 return bfqq;
6729 }
6730
6731 bic_set_bfqq(bic, NULL, true, bfqq->actuator_idx);
6732
6733 bfq_put_cooperator(bfqq);
6734
6735 bfq_release_process_ref(bfqq->bfqd, bfqq);
6736 return NULL;
6737}
6738
6739static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
6740 struct bfq_io_cq *bic,
6741 struct bio *bio,
6742 bool split, bool is_sync,
6743 bool *new_queue)
6744{
6745 unsigned int act_idx = bfq_actuator_index(bfqd, bio);
6746 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync, act_idx);
6747 struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[act_idx];
6748
6749 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
6750 return bfqq;
6751
6752 if (new_queue)
6753 *new_queue = true;
6754
6755 if (bfqq)
6756 bfq_put_queue(bfqq);
6757 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic, split);
6758
6759 bic_set_bfqq(bic, bfqq, is_sync, act_idx);
6760 if (split && is_sync) {
6761 if ((bfqq_data->was_in_burst_list && bfqd->large_burst) ||
6762 bfqq_data->saved_in_large_burst)
6763 bfq_mark_bfqq_in_large_burst(bfqq);
6764 else {
6765 bfq_clear_bfqq_in_large_burst(bfqq);
6766 if (bfqq_data->was_in_burst_list)
6767 /*
6768 * If bfqq was in the current
6769 * burst list before being
6770 * merged, then we have to add
6771 * it back. And we do not need
6772 * to increase burst_size, as
6773 * we did not decrement
6774 * burst_size when we removed
6775 * bfqq from the burst list as
6776 * a consequence of a merge
6777 * (see comments in
6778 * bfq_put_queue). In this
6779 * respect, it would be rather
6780 * costly to know whether the
6781 * current burst list is still
6782 * the same burst list from
6783 * which bfqq was removed on
6784 * the merge. To avoid this
6785 * cost, if bfqq was in a
6786 * burst list, then we add
6787 * bfqq to the current burst
6788 * list without any further
6789 * check. This can cause
6790 * inappropriate insertions,
6791 * but rarely enough to not
6792 * harm the detection of large
6793 * bursts significantly.
6794 */
6795 hlist_add_head(&bfqq->burst_list_node,
6796 &bfqd->burst_list);
6797 }
6798 bfqq->split_time = jiffies;
6799 }
6800
6801 return bfqq;
6802}
6803
6804/*
6805 * Only reset private fields. The actual request preparation will be
6806 * performed by bfq_init_rq, when rq is either inserted or merged. See
6807 * comments on bfq_init_rq for the reason behind this delayed
6808 * preparation.
6809 */
6810static void bfq_prepare_request(struct request *rq)
6811{
6812 rq->elv.icq = ioc_find_get_icq(rq->q);
6813
6814 /*
6815 * Regardless of whether we have an icq attached, we have to
6816 * clear the scheduler pointers, as they might point to
6817 * previously allocated bic/bfqq structs.
6818 */
6819 rq->elv.priv[0] = rq->elv.priv[1] = NULL;
6820}
6821
6822/*
6823 * If needed, init rq, allocate bfq data structures associated with
6824 * rq, and increment reference counters in the destination bfq_queue
6825 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6826 * not associated with any bfq_queue.
6827 *
6828 * This function is invoked by the functions that perform rq insertion
6829 * or merging. One may have expected the above preparation operations
6830 * to be performed in bfq_prepare_request, and not delayed to when rq
6831 * is inserted or merged. The rationale behind this delayed
6832 * preparation is that, after the prepare_request hook is invoked for
6833 * rq, rq may still be transformed into a request with no icq, i.e., a
6834 * request not associated with any queue. No bfq hook is invoked to
6835 * signal this transformation. As a consequence, should these
6836 * preparation operations be performed when the prepare_request hook
6837 * is invoked, and should rq be transformed one moment later, bfq
6838 * would end up in an inconsistent state, because it would have
6839 * incremented some queue counters for an rq destined to
6840 * transformation, without any chance to correctly lower these
6841 * counters back. In contrast, no transformation can still happen for
6842 * rq after rq has been inserted or merged. So, it is safe to execute
6843 * these preparation operations when rq is finally inserted or merged.
6844 */
6845static struct bfq_queue *bfq_init_rq(struct request *rq)
6846{
6847 struct request_queue *q = rq->q;
6848 struct bio *bio = rq->bio;
6849 struct bfq_data *bfqd = q->elevator->elevator_data;
6850 struct bfq_io_cq *bic;
6851 const int is_sync = rq_is_sync(rq);
6852 struct bfq_queue *bfqq;
6853 bool new_queue = false;
6854 bool bfqq_already_existing = false, split = false;
6855 unsigned int a_idx = bfq_actuator_index(bfqd, bio);
6856
6857 if (unlikely(!rq->elv.icq))
6858 return NULL;
6859
6860 /*
6861 * Assuming that RQ_BFQQ(rq) is set only if everything is set
6862 * for this rq. This holds true, because this function is
6863 * invoked only for insertion or merging, and, after such
6864 * events, a request cannot be manipulated any longer before
6865 * being removed from bfq.
6866 */
6867 if (RQ_BFQQ(rq))
6868 return RQ_BFQQ(rq);
6869
6870 bic = icq_to_bic(rq->elv.icq);
6871
6872 bfq_check_ioprio_change(bic, bio);
6873
6874 bfq_bic_update_cgroup(bic, bio);
6875
6876 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
6877 &new_queue);
6878
6879 if (likely(!new_queue)) {
6880 /* If the queue was seeky for too long, break it apart. */
6881 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq) &&
6882 !bic->bfqq_data[a_idx].stably_merged) {
6883 struct bfq_queue *old_bfqq = bfqq;
6884
6885 /* Update bic before losing reference to bfqq */
6886 if (bfq_bfqq_in_large_burst(bfqq))
6887 bic->bfqq_data[a_idx].saved_in_large_burst =
6888 true;
6889
6890 bfqq = bfq_split_bfqq(bic, bfqq);
6891 split = true;
6892
6893 if (!bfqq) {
6894 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
6895 true, is_sync,
6896 NULL);
6897 if (unlikely(bfqq == &bfqd->oom_bfqq))
6898 bfqq_already_existing = true;
6899 } else
6900 bfqq_already_existing = true;
6901
6902 if (!bfqq_already_existing) {
6903 bfqq->waker_bfqq = old_bfqq->waker_bfqq;
6904 bfqq->tentative_waker_bfqq = NULL;
6905
6906 /*
6907 * If the waker queue disappears, then
6908 * new_bfqq->waker_bfqq must be
6909 * reset. So insert new_bfqq into the
6910 * woken_list of the waker. See
6911 * bfq_check_waker for details.
6912 */
6913 if (bfqq->waker_bfqq)
6914 hlist_add_head(&bfqq->woken_list_node,
6915 &bfqq->waker_bfqq->woken_list);
6916 }
6917 }
6918 }
6919
6920 bfqq_request_allocated(bfqq);
6921 bfqq->ref++;
6922 bic->requests++;
6923 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
6924 rq, bfqq, bfqq->ref);
6925
6926 rq->elv.priv[0] = bic;
6927 rq->elv.priv[1] = bfqq;
6928
6929 /*
6930 * If a bfq_queue has only one process reference, it is owned
6931 * by only this bic: we can then set bfqq->bic = bic. in
6932 * addition, if the queue has also just been split, we have to
6933 * resume its state.
6934 */
6935 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
6936 bfqq->bic = bic;
6937 if (split) {
6938 /*
6939 * The queue has just been split from a shared
6940 * queue: restore the idle window and the
6941 * possible weight raising period.
6942 */
6943 bfq_bfqq_resume_state(bfqq, bfqd, bic,
6944 bfqq_already_existing);
6945 }
6946 }
6947
6948 /*
6949 * Consider bfqq as possibly belonging to a burst of newly
6950 * created queues only if:
6951 * 1) A burst is actually happening (bfqd->burst_size > 0)
6952 * or
6953 * 2) There is no other active queue. In fact, if, in
6954 * contrast, there are active queues not belonging to the
6955 * possible burst bfqq may belong to, then there is no gain
6956 * in considering bfqq as belonging to a burst, and
6957 * therefore in not weight-raising bfqq. See comments on
6958 * bfq_handle_burst().
6959 *
6960 * This filtering also helps eliminating false positives,
6961 * occurring when bfqq does not belong to an actual large
6962 * burst, but some background task (e.g., a service) happens
6963 * to trigger the creation of new queues very close to when
6964 * bfqq and its possible companion queues are created. See
6965 * comments on bfq_handle_burst() for further details also on
6966 * this issue.
6967 */
6968 if (unlikely(bfq_bfqq_just_created(bfqq) &&
6969 (bfqd->burst_size > 0 ||
6970 bfq_tot_busy_queues(bfqd) == 0)))
6971 bfq_handle_burst(bfqd, bfqq);
6972
6973 return bfqq;
6974}
6975
6976static void
6977bfq_idle_slice_timer_body(struct bfq_data *bfqd, struct bfq_queue *bfqq)
6978{
6979 enum bfqq_expiration reason;
6980 unsigned long flags;
6981
6982 spin_lock_irqsave(&bfqd->lock, flags);
6983
6984 /*
6985 * Considering that bfqq may be in race, we should firstly check
6986 * whether bfqq is in service before doing something on it. If
6987 * the bfqq in race is not in service, it has already been expired
6988 * through __bfq_bfqq_expire func and its wait_request flags has
6989 * been cleared in __bfq_bfqd_reset_in_service func.
6990 */
6991 if (bfqq != bfqd->in_service_queue) {
6992 spin_unlock_irqrestore(&bfqd->lock, flags);
6993 return;
6994 }
6995
6996 bfq_clear_bfqq_wait_request(bfqq);
6997
6998 if (bfq_bfqq_budget_timeout(bfqq))
6999 /*
7000 * Also here the queue can be safely expired
7001 * for budget timeout without wasting
7002 * guarantees
7003 */
7004 reason = BFQQE_BUDGET_TIMEOUT;
7005 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
7006 /*
7007 * The queue may not be empty upon timer expiration,
7008 * because we may not disable the timer when the
7009 * first request of the in-service queue arrives
7010 * during disk idling.
7011 */
7012 reason = BFQQE_TOO_IDLE;
7013 else
7014 goto schedule_dispatch;
7015
7016 bfq_bfqq_expire(bfqd, bfqq, true, reason);
7017
7018schedule_dispatch:
7019 bfq_schedule_dispatch(bfqd);
7020 spin_unlock_irqrestore(&bfqd->lock, flags);
7021}
7022
7023/*
7024 * Handler of the expiration of the timer running if the in-service queue
7025 * is idling inside its time slice.
7026 */
7027static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
7028{
7029 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
7030 idle_slice_timer);
7031 struct bfq_queue *bfqq = bfqd->in_service_queue;
7032
7033 /*
7034 * Theoretical race here: the in-service queue can be NULL or
7035 * different from the queue that was idling if a new request
7036 * arrives for the current queue and there is a full dispatch
7037 * cycle that changes the in-service queue. This can hardly
7038 * happen, but in the worst case we just expire a queue too
7039 * early.
7040 */
7041 if (bfqq)
7042 bfq_idle_slice_timer_body(bfqd, bfqq);
7043
7044 return HRTIMER_NORESTART;
7045}
7046
7047static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
7048 struct bfq_queue **bfqq_ptr)
7049{
7050 struct bfq_queue *bfqq = *bfqq_ptr;
7051
7052 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
7053 if (bfqq) {
7054 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
7055
7056 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
7057 bfqq, bfqq->ref);
7058 bfq_put_queue(bfqq);
7059 *bfqq_ptr = NULL;
7060 }
7061}
7062
7063/*
7064 * Release all the bfqg references to its async queues. If we are
7065 * deallocating the group these queues may still contain requests, so
7066 * we reparent them to the root cgroup (i.e., the only one that will
7067 * exist for sure until all the requests on a device are gone).
7068 */
7069void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
7070{
7071 int i, j, k;
7072
7073 for (k = 0; k < bfqd->num_actuators; k++) {
7074 for (i = 0; i < 2; i++)
7075 for (j = 0; j < IOPRIO_NR_LEVELS; j++)
7076 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j][k]);
7077
7078 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq[k]);
7079 }
7080}
7081
7082/*
7083 * See the comments on bfq_limit_depth for the purpose of
7084 * the depths set in the function. Return minimum shallow depth we'll use.
7085 */
7086static void bfq_update_depths(struct bfq_data *bfqd, struct sbitmap_queue *bt)
7087{
7088 unsigned int depth = 1U << bt->sb.shift;
7089
7090 bfqd->full_depth_shift = bt->sb.shift;
7091 /*
7092 * In-word depths if no bfq_queue is being weight-raised:
7093 * leaving 25% of tags only for sync reads.
7094 *
7095 * In next formulas, right-shift the value
7096 * (1U<<bt->sb.shift), instead of computing directly
7097 * (1U<<(bt->sb.shift - something)), to be robust against
7098 * any possible value of bt->sb.shift, without having to
7099 * limit 'something'.
7100 */
7101 /* no more than 50% of tags for async I/O */
7102 bfqd->word_depths[0][0] = max(depth >> 1, 1U);
7103 /*
7104 * no more than 75% of tags for sync writes (25% extra tags
7105 * w.r.t. async I/O, to prevent async I/O from starving sync
7106 * writes)
7107 */
7108 bfqd->word_depths[0][1] = max((depth * 3) >> 2, 1U);
7109
7110 /*
7111 * In-word depths in case some bfq_queue is being weight-
7112 * raised: leaving ~63% of tags for sync reads. This is the
7113 * highest percentage for which, in our tests, application
7114 * start-up times didn't suffer from any regression due to tag
7115 * shortage.
7116 */
7117 /* no more than ~18% of tags for async I/O */
7118 bfqd->word_depths[1][0] = max((depth * 3) >> 4, 1U);
7119 /* no more than ~37% of tags for sync writes (~20% extra tags) */
7120 bfqd->word_depths[1][1] = max((depth * 6) >> 4, 1U);
7121}
7122
7123static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
7124{
7125 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
7126 struct blk_mq_tags *tags = hctx->sched_tags;
7127
7128 bfq_update_depths(bfqd, &tags->bitmap_tags);
7129 sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, 1);
7130}
7131
7132static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
7133{
7134 bfq_depth_updated(hctx);
7135 return 0;
7136}
7137
7138static void bfq_exit_queue(struct elevator_queue *e)
7139{
7140 struct bfq_data *bfqd = e->elevator_data;
7141 struct bfq_queue *bfqq, *n;
7142 unsigned int actuator;
7143
7144 hrtimer_cancel(&bfqd->idle_slice_timer);
7145
7146 spin_lock_irq(&bfqd->lock);
7147 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
7148 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
7149 spin_unlock_irq(&bfqd->lock);
7150
7151 for (actuator = 0; actuator < bfqd->num_actuators; actuator++)
7152 WARN_ON_ONCE(bfqd->rq_in_driver[actuator]);
7153 WARN_ON_ONCE(bfqd->tot_rq_in_driver);
7154
7155 hrtimer_cancel(&bfqd->idle_slice_timer);
7156
7157 /* release oom-queue reference to root group */
7158 bfqg_and_blkg_put(bfqd->root_group);
7159
7160#ifdef CONFIG_BFQ_GROUP_IOSCHED
7161 blkcg_deactivate_policy(bfqd->queue->disk, &blkcg_policy_bfq);
7162#else
7163 spin_lock_irq(&bfqd->lock);
7164 bfq_put_async_queues(bfqd, bfqd->root_group);
7165 kfree(bfqd->root_group);
7166 spin_unlock_irq(&bfqd->lock);
7167#endif
7168
7169 blk_stat_disable_accounting(bfqd->queue);
7170 clear_bit(ELEVATOR_FLAG_DISABLE_WBT, &e->flags);
7171 wbt_enable_default(bfqd->queue->disk);
7172
7173 kfree(bfqd);
7174}
7175
7176static void bfq_init_root_group(struct bfq_group *root_group,
7177 struct bfq_data *bfqd)
7178{
7179 int i;
7180
7181#ifdef CONFIG_BFQ_GROUP_IOSCHED
7182 root_group->entity.parent = NULL;
7183 root_group->my_entity = NULL;
7184 root_group->bfqd = bfqd;
7185#endif
7186 root_group->rq_pos_tree = RB_ROOT;
7187 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
7188 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
7189 root_group->sched_data.bfq_class_idle_last_service = jiffies;
7190}
7191
7192static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
7193{
7194 struct bfq_data *bfqd;
7195 struct elevator_queue *eq;
7196 unsigned int i;
7197 struct blk_independent_access_ranges *ia_ranges = q->disk->ia_ranges;
7198
7199 eq = elevator_alloc(q, e);
7200 if (!eq)
7201 return -ENOMEM;
7202
7203 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
7204 if (!bfqd) {
7205 kobject_put(&eq->kobj);
7206 return -ENOMEM;
7207 }
7208 eq->elevator_data = bfqd;
7209
7210 spin_lock_irq(&q->queue_lock);
7211 q->elevator = eq;
7212 spin_unlock_irq(&q->queue_lock);
7213
7214 /*
7215 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
7216 * Grab a permanent reference to it, so that the normal code flow
7217 * will not attempt to free it.
7218 * Set zero as actuator index: we will pretend that
7219 * all I/O requests are for the same actuator.
7220 */
7221 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0, 0);
7222 bfqd->oom_bfqq.ref++;
7223 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
7224 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
7225 bfqd->oom_bfqq.entity.new_weight =
7226 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
7227
7228 /* oom_bfqq does not participate to bursts */
7229 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
7230
7231 /*
7232 * Trigger weight initialization, according to ioprio, at the
7233 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
7234 * class won't be changed any more.
7235 */
7236 bfqd->oom_bfqq.entity.prio_changed = 1;
7237
7238 bfqd->queue = q;
7239
7240 bfqd->num_actuators = 1;
7241 /*
7242 * If the disk supports multiple actuators, copy independent
7243 * access ranges from the request queue structure.
7244 */
7245 spin_lock_irq(&q->queue_lock);
7246 if (ia_ranges) {
7247 /*
7248 * Check if the disk ia_ranges size exceeds the current bfq
7249 * actuator limit.
7250 */
7251 if (ia_ranges->nr_ia_ranges > BFQ_MAX_ACTUATORS) {
7252 pr_crit("nr_ia_ranges higher than act limit: iars=%d, max=%d.\n",
7253 ia_ranges->nr_ia_ranges, BFQ_MAX_ACTUATORS);
7254 pr_crit("Falling back to single actuator mode.\n");
7255 } else {
7256 bfqd->num_actuators = ia_ranges->nr_ia_ranges;
7257
7258 for (i = 0; i < bfqd->num_actuators; i++) {
7259 bfqd->sector[i] = ia_ranges->ia_range[i].sector;
7260 bfqd->nr_sectors[i] =
7261 ia_ranges->ia_range[i].nr_sectors;
7262 }
7263 }
7264 }
7265
7266 /* Otherwise use single-actuator dev info */
7267 if (bfqd->num_actuators == 1) {
7268 bfqd->sector[0] = 0;
7269 bfqd->nr_sectors[0] = get_capacity(q->disk);
7270 }
7271 spin_unlock_irq(&q->queue_lock);
7272
7273 INIT_LIST_HEAD(&bfqd->dispatch);
7274
7275 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
7276 HRTIMER_MODE_REL);
7277 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
7278
7279 bfqd->queue_weights_tree = RB_ROOT_CACHED;
7280#ifdef CONFIG_BFQ_GROUP_IOSCHED
7281 bfqd->num_groups_with_pending_reqs = 0;
7282#endif
7283
7284 INIT_LIST_HEAD(&bfqd->active_list[0]);
7285 INIT_LIST_HEAD(&bfqd->active_list[1]);
7286 INIT_LIST_HEAD(&bfqd->idle_list);
7287 INIT_HLIST_HEAD(&bfqd->burst_list);
7288
7289 bfqd->hw_tag = -1;
7290 bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
7291
7292 bfqd->bfq_max_budget = bfq_default_max_budget;
7293
7294 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
7295 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
7296 bfqd->bfq_back_max = bfq_back_max;
7297 bfqd->bfq_back_penalty = bfq_back_penalty;
7298 bfqd->bfq_slice_idle = bfq_slice_idle;
7299 bfqd->bfq_timeout = bfq_timeout;
7300
7301 bfqd->bfq_large_burst_thresh = 8;
7302 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
7303
7304 bfqd->low_latency = true;
7305
7306 /*
7307 * Trade-off between responsiveness and fairness.
7308 */
7309 bfqd->bfq_wr_coeff = 30;
7310 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
7311 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
7312 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
7313 bfqd->bfq_wr_max_softrt_rate = 7000; /*
7314 * Approximate rate required
7315 * to playback or record a
7316 * high-definition compressed
7317 * video.
7318 */
7319 bfqd->wr_busy_queues = 0;
7320
7321 /*
7322 * Begin by assuming, optimistically, that the device peak
7323 * rate is equal to 2/3 of the highest reference rate.
7324 */
7325 bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
7326 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
7327 bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
7328
7329 /* see comments on the definition of next field inside bfq_data */
7330 bfqd->actuator_load_threshold = 4;
7331
7332 spin_lock_init(&bfqd->lock);
7333
7334 /*
7335 * The invocation of the next bfq_create_group_hierarchy
7336 * function is the head of a chain of function calls
7337 * (bfq_create_group_hierarchy->blkcg_activate_policy->
7338 * blk_mq_freeze_queue) that may lead to the invocation of the
7339 * has_work hook function. For this reason,
7340 * bfq_create_group_hierarchy is invoked only after all
7341 * scheduler data has been initialized, apart from the fields
7342 * that can be initialized only after invoking
7343 * bfq_create_group_hierarchy. This, in particular, enables
7344 * has_work to correctly return false. Of course, to avoid
7345 * other inconsistencies, the blk-mq stack must then refrain
7346 * from invoking further scheduler hooks before this init
7347 * function is finished.
7348 */
7349 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
7350 if (!bfqd->root_group)
7351 goto out_free;
7352 bfq_init_root_group(bfqd->root_group, bfqd);
7353 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
7354
7355 /* We dispatch from request queue wide instead of hw queue */
7356 blk_queue_flag_set(QUEUE_FLAG_SQ_SCHED, q);
7357
7358 set_bit(ELEVATOR_FLAG_DISABLE_WBT, &eq->flags);
7359 wbt_disable_default(q->disk);
7360 blk_stat_enable_accounting(q);
7361
7362 return 0;
7363
7364out_free:
7365 kfree(bfqd);
7366 kobject_put(&eq->kobj);
7367 return -ENOMEM;
7368}
7369
7370static void bfq_slab_kill(void)
7371{
7372 kmem_cache_destroy(bfq_pool);
7373}
7374
7375static int __init bfq_slab_setup(void)
7376{
7377 bfq_pool = KMEM_CACHE(bfq_queue, 0);
7378 if (!bfq_pool)
7379 return -ENOMEM;
7380 return 0;
7381}
7382
7383static ssize_t bfq_var_show(unsigned int var, char *page)
7384{
7385 return sprintf(page, "%u\n", var);
7386}
7387
7388static int bfq_var_store(unsigned long *var, const char *page)
7389{
7390 unsigned long new_val;
7391 int ret = kstrtoul(page, 10, &new_val);
7392
7393 if (ret)
7394 return ret;
7395 *var = new_val;
7396 return 0;
7397}
7398
7399#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
7400static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7401{ \
7402 struct bfq_data *bfqd = e->elevator_data; \
7403 u64 __data = __VAR; \
7404 if (__CONV == 1) \
7405 __data = jiffies_to_msecs(__data); \
7406 else if (__CONV == 2) \
7407 __data = div_u64(__data, NSEC_PER_MSEC); \
7408 return bfq_var_show(__data, (page)); \
7409}
7410SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
7411SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
7412SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
7413SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
7414SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
7415SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
7416SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
7417SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
7418SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
7419#undef SHOW_FUNCTION
7420
7421#define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
7422static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7423{ \
7424 struct bfq_data *bfqd = e->elevator_data; \
7425 u64 __data = __VAR; \
7426 __data = div_u64(__data, NSEC_PER_USEC); \
7427 return bfq_var_show(__data, (page)); \
7428}
7429USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
7430#undef USEC_SHOW_FUNCTION
7431
7432#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
7433static ssize_t \
7434__FUNC(struct elevator_queue *e, const char *page, size_t count) \
7435{ \
7436 struct bfq_data *bfqd = e->elevator_data; \
7437 unsigned long __data, __min = (MIN), __max = (MAX); \
7438 int ret; \
7439 \
7440 ret = bfq_var_store(&__data, (page)); \
7441 if (ret) \
7442 return ret; \
7443 if (__data < __min) \
7444 __data = __min; \
7445 else if (__data > __max) \
7446 __data = __max; \
7447 if (__CONV == 1) \
7448 *(__PTR) = msecs_to_jiffies(__data); \
7449 else if (__CONV == 2) \
7450 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
7451 else \
7452 *(__PTR) = __data; \
7453 return count; \
7454}
7455STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
7456 INT_MAX, 2);
7457STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
7458 INT_MAX, 2);
7459STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
7460STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
7461 INT_MAX, 0);
7462STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
7463#undef STORE_FUNCTION
7464
7465#define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
7466static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
7467{ \
7468 struct bfq_data *bfqd = e->elevator_data; \
7469 unsigned long __data, __min = (MIN), __max = (MAX); \
7470 int ret; \
7471 \
7472 ret = bfq_var_store(&__data, (page)); \
7473 if (ret) \
7474 return ret; \
7475 if (__data < __min) \
7476 __data = __min; \
7477 else if (__data > __max) \
7478 __data = __max; \
7479 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
7480 return count; \
7481}
7482USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
7483 UINT_MAX);
7484#undef USEC_STORE_FUNCTION
7485
7486static ssize_t bfq_max_budget_store(struct elevator_queue *e,
7487 const char *page, size_t count)
7488{
7489 struct bfq_data *bfqd = e->elevator_data;
7490 unsigned long __data;
7491 int ret;
7492
7493 ret = bfq_var_store(&__data, (page));
7494 if (ret)
7495 return ret;
7496
7497 if (__data == 0)
7498 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7499 else {
7500 if (__data > INT_MAX)
7501 __data = INT_MAX;
7502 bfqd->bfq_max_budget = __data;
7503 }
7504
7505 bfqd->bfq_user_max_budget = __data;
7506
7507 return count;
7508}
7509
7510/*
7511 * Leaving this name to preserve name compatibility with cfq
7512 * parameters, but this timeout is used for both sync and async.
7513 */
7514static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
7515 const char *page, size_t count)
7516{
7517 struct bfq_data *bfqd = e->elevator_data;
7518 unsigned long __data;
7519 int ret;
7520
7521 ret = bfq_var_store(&__data, (page));
7522 if (ret)
7523 return ret;
7524
7525 if (__data < 1)
7526 __data = 1;
7527 else if (__data > INT_MAX)
7528 __data = INT_MAX;
7529
7530 bfqd->bfq_timeout = msecs_to_jiffies(__data);
7531 if (bfqd->bfq_user_max_budget == 0)
7532 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7533
7534 return count;
7535}
7536
7537static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
7538 const char *page, size_t count)
7539{
7540 struct bfq_data *bfqd = e->elevator_data;
7541 unsigned long __data;
7542 int ret;
7543
7544 ret = bfq_var_store(&__data, (page));
7545 if (ret)
7546 return ret;
7547
7548 if (__data > 1)
7549 __data = 1;
7550 if (!bfqd->strict_guarantees && __data == 1
7551 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
7552 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
7553
7554 bfqd->strict_guarantees = __data;
7555
7556 return count;
7557}
7558
7559static ssize_t bfq_low_latency_store(struct elevator_queue *e,
7560 const char *page, size_t count)
7561{
7562 struct bfq_data *bfqd = e->elevator_data;
7563 unsigned long __data;
7564 int ret;
7565
7566 ret = bfq_var_store(&__data, (page));
7567 if (ret)
7568 return ret;
7569
7570 if (__data > 1)
7571 __data = 1;
7572 if (__data == 0 && bfqd->low_latency != 0)
7573 bfq_end_wr(bfqd);
7574 bfqd->low_latency = __data;
7575
7576 return count;
7577}
7578
7579#define BFQ_ATTR(name) \
7580 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
7581
7582static struct elv_fs_entry bfq_attrs[] = {
7583 BFQ_ATTR(fifo_expire_sync),
7584 BFQ_ATTR(fifo_expire_async),
7585 BFQ_ATTR(back_seek_max),
7586 BFQ_ATTR(back_seek_penalty),
7587 BFQ_ATTR(slice_idle),
7588 BFQ_ATTR(slice_idle_us),
7589 BFQ_ATTR(max_budget),
7590 BFQ_ATTR(timeout_sync),
7591 BFQ_ATTR(strict_guarantees),
7592 BFQ_ATTR(low_latency),
7593 __ATTR_NULL
7594};
7595
7596static struct elevator_type iosched_bfq_mq = {
7597 .ops = {
7598 .limit_depth = bfq_limit_depth,
7599 .prepare_request = bfq_prepare_request,
7600 .requeue_request = bfq_finish_requeue_request,
7601 .finish_request = bfq_finish_request,
7602 .exit_icq = bfq_exit_icq,
7603 .insert_requests = bfq_insert_requests,
7604 .dispatch_request = bfq_dispatch_request,
7605 .next_request = elv_rb_latter_request,
7606 .former_request = elv_rb_former_request,
7607 .allow_merge = bfq_allow_bio_merge,
7608 .bio_merge = bfq_bio_merge,
7609 .request_merge = bfq_request_merge,
7610 .requests_merged = bfq_requests_merged,
7611 .request_merged = bfq_request_merged,
7612 .has_work = bfq_has_work,
7613 .depth_updated = bfq_depth_updated,
7614 .init_hctx = bfq_init_hctx,
7615 .init_sched = bfq_init_queue,
7616 .exit_sched = bfq_exit_queue,
7617 },
7618
7619 .icq_size = sizeof(struct bfq_io_cq),
7620 .icq_align = __alignof__(struct bfq_io_cq),
7621 .elevator_attrs = bfq_attrs,
7622 .elevator_name = "bfq",
7623 .elevator_owner = THIS_MODULE,
7624};
7625MODULE_ALIAS("bfq-iosched");
7626
7627static int __init bfq_init(void)
7628{
7629 int ret;
7630
7631#ifdef CONFIG_BFQ_GROUP_IOSCHED
7632 ret = blkcg_policy_register(&blkcg_policy_bfq);
7633 if (ret)
7634 return ret;
7635#endif
7636
7637 ret = -ENOMEM;
7638 if (bfq_slab_setup())
7639 goto err_pol_unreg;
7640
7641 /*
7642 * Times to load large popular applications for the typical
7643 * systems installed on the reference devices (see the
7644 * comments before the definition of the next
7645 * array). Actually, we use slightly lower values, as the
7646 * estimated peak rate tends to be smaller than the actual
7647 * peak rate. The reason for this last fact is that estimates
7648 * are computed over much shorter time intervals than the long
7649 * intervals typically used for benchmarking. Why? First, to
7650 * adapt more quickly to variations. Second, because an I/O
7651 * scheduler cannot rely on a peak-rate-evaluation workload to
7652 * be run for a long time.
7653 */
7654 ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
7655 ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
7656
7657 ret = elv_register(&iosched_bfq_mq);
7658 if (ret)
7659 goto slab_kill;
7660
7661 return 0;
7662
7663slab_kill:
7664 bfq_slab_kill();
7665err_pol_unreg:
7666#ifdef CONFIG_BFQ_GROUP_IOSCHED
7667 blkcg_policy_unregister(&blkcg_policy_bfq);
7668#endif
7669 return ret;
7670}
7671
7672static void __exit bfq_exit(void)
7673{
7674 elv_unregister(&iosched_bfq_mq);
7675#ifdef CONFIG_BFQ_GROUP_IOSCHED
7676 blkcg_policy_unregister(&blkcg_policy_bfq);
7677#endif
7678 bfq_slab_kill();
7679}
7680
7681module_init(bfq_init);
7682module_exit(bfq_exit);
7683
7684MODULE_AUTHOR("Paolo Valente");
7685MODULE_LICENSE("GPL");
7686MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");