Loading...
Note: File does not exist in v4.6.
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_data *bfqd,
586 struct bfq_io_cq *bic, blk_opf_t opf,
587 unsigned int act_idx, int limit)
588{
589 struct bfq_entity *inline_entities[BFQ_LIMIT_INLINE_DEPTH];
590 struct bfq_entity **entities = inline_entities;
591 int alloc_depth = BFQ_LIMIT_INLINE_DEPTH;
592 struct bfq_sched_data *sched_data;
593 struct bfq_entity *entity;
594 struct bfq_queue *bfqq;
595 unsigned long wsum;
596 bool ret = false;
597 int depth;
598 int level;
599
600retry:
601 spin_lock_irq(&bfqd->lock);
602 bfqq = bic_to_bfqq(bic, op_is_sync(opf), act_idx);
603 if (!bfqq)
604 goto out;
605
606 entity = &bfqq->entity;
607 if (!entity->on_st_or_in_serv)
608 goto out;
609
610 /* +1 for bfqq entity, root cgroup not included */
611 depth = bfqg_to_blkg(bfqq_group(bfqq))->blkcg->css.cgroup->level + 1;
612 if (depth > alloc_depth) {
613 spin_unlock_irq(&bfqd->lock);
614 if (entities != inline_entities)
615 kfree(entities);
616 entities = kmalloc_array(depth, sizeof(*entities), GFP_NOIO);
617 if (!entities)
618 return false;
619 alloc_depth = depth;
620 goto retry;
621 }
622
623 sched_data = entity->sched_data;
624 /* Gather our ancestors as we need to traverse them in reverse order */
625 level = 0;
626 for_each_entity(entity) {
627 /*
628 * If at some level entity is not even active, allow request
629 * queueing so that BFQ knows there's work to do and activate
630 * entities.
631 */
632 if (!entity->on_st_or_in_serv)
633 goto out;
634 /* Uh, more parents than cgroup subsystem thinks? */
635 if (WARN_ON_ONCE(level >= depth))
636 break;
637 entities[level++] = entity;
638 }
639 WARN_ON_ONCE(level != depth);
640 for (level--; level >= 0; level--) {
641 entity = entities[level];
642 if (level > 0) {
643 wsum = bfq_entity_service_tree(entity)->wsum;
644 } else {
645 int i;
646 /*
647 * For bfqq itself we take into account service trees
648 * of all higher priority classes and multiply their
649 * weights so that low prio queue from higher class
650 * gets more requests than high prio queue from lower
651 * class.
652 */
653 wsum = 0;
654 for (i = 0; i <= bfqq->ioprio_class - 1; i++) {
655 wsum = wsum * IOPRIO_BE_NR +
656 sched_data->service_tree[i].wsum;
657 }
658 }
659 if (!wsum)
660 continue;
661 limit = DIV_ROUND_CLOSEST(limit * entity->weight, wsum);
662 if (entity->allocated >= limit) {
663 bfq_log_bfqq(bfqq->bfqd, bfqq,
664 "too many requests: allocated %d limit %d level %d",
665 entity->allocated, limit, level);
666 ret = true;
667 break;
668 }
669 }
670out:
671 spin_unlock_irq(&bfqd->lock);
672 if (entities != inline_entities)
673 kfree(entities);
674 return ret;
675}
676#else
677static bool bfqq_request_over_limit(struct bfq_data *bfqd,
678 struct bfq_io_cq *bic, blk_opf_t opf,
679 unsigned int act_idx, int limit)
680{
681 return false;
682}
683#endif
684
685/*
686 * Async I/O can easily starve sync I/O (both sync reads and sync
687 * writes), by consuming all tags. Similarly, storms of sync writes,
688 * such as those that sync(2) may trigger, can starve sync reads.
689 * Limit depths of async I/O and sync writes so as to counter both
690 * problems.
691 *
692 * Also if a bfq queue or its parent cgroup consume more tags than would be
693 * appropriate for their weight, we trim the available tag depth to 1. This
694 * avoids a situation where one cgroup can starve another cgroup from tags and
695 * thus block service differentiation among cgroups. Note that because the
696 * queue / cgroup already has many requests allocated and queued, this does not
697 * significantly affect service guarantees coming from the BFQ scheduling
698 * algorithm.
699 */
700static void bfq_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
701{
702 struct bfq_data *bfqd = data->q->elevator->elevator_data;
703 struct bfq_io_cq *bic = bfq_bic_lookup(data->q);
704 int depth;
705 unsigned limit = data->q->nr_requests;
706 unsigned int act_idx;
707
708 /* Sync reads have full depth available */
709 if (op_is_sync(opf) && !op_is_write(opf)) {
710 depth = 0;
711 } else {
712 depth = bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(opf)];
713 limit = (limit * depth) >> bfqd->full_depth_shift;
714 }
715
716 for (act_idx = 0; bic && act_idx < bfqd->num_actuators; act_idx++) {
717 /* Fast path to check if bfqq is already allocated. */
718 if (!bic_to_bfqq(bic, op_is_sync(opf), act_idx))
719 continue;
720
721 /*
722 * Does queue (or any parent entity) exceed number of
723 * requests that should be available to it? Heavily
724 * limit depth so that it cannot consume more
725 * available requests and thus starve other entities.
726 */
727 if (bfqq_request_over_limit(bfqd, bic, opf, act_idx, limit)) {
728 depth = 1;
729 break;
730 }
731 }
732 bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
733 __func__, bfqd->wr_busy_queues, op_is_sync(opf), depth);
734 if (depth)
735 data->shallow_depth = depth;
736}
737
738static struct bfq_queue *
739bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
740 sector_t sector, struct rb_node **ret_parent,
741 struct rb_node ***rb_link)
742{
743 struct rb_node **p, *parent;
744 struct bfq_queue *bfqq = NULL;
745
746 parent = NULL;
747 p = &root->rb_node;
748 while (*p) {
749 struct rb_node **n;
750
751 parent = *p;
752 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
753
754 /*
755 * Sort strictly based on sector. Smallest to the left,
756 * largest to the right.
757 */
758 if (sector > blk_rq_pos(bfqq->next_rq))
759 n = &(*p)->rb_right;
760 else if (sector < blk_rq_pos(bfqq->next_rq))
761 n = &(*p)->rb_left;
762 else
763 break;
764 p = n;
765 bfqq = NULL;
766 }
767
768 *ret_parent = parent;
769 if (rb_link)
770 *rb_link = p;
771
772 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
773 (unsigned long long)sector,
774 bfqq ? bfqq->pid : 0);
775
776 return bfqq;
777}
778
779static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
780{
781 return bfqq->service_from_backlogged > 0 &&
782 time_is_before_jiffies(bfqq->first_IO_time +
783 bfq_merge_time_limit);
784}
785
786/*
787 * The following function is not marked as __cold because it is
788 * actually cold, but for the same performance goal described in the
789 * comments on the likely() at the beginning of
790 * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
791 * execution time for the case where this function is not invoked, we
792 * had to add an unlikely() in each involved if().
793 */
794void __cold
795bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
796{
797 struct rb_node **p, *parent;
798 struct bfq_queue *__bfqq;
799
800 if (bfqq->pos_root) {
801 rb_erase(&bfqq->pos_node, bfqq->pos_root);
802 bfqq->pos_root = NULL;
803 }
804
805 /* oom_bfqq does not participate in queue merging */
806 if (bfqq == &bfqd->oom_bfqq)
807 return;
808
809 /*
810 * bfqq cannot be merged any longer (see comments in
811 * bfq_setup_cooperator): no point in adding bfqq into the
812 * position tree.
813 */
814 if (bfq_too_late_for_merging(bfqq))
815 return;
816
817 if (bfq_class_idle(bfqq))
818 return;
819 if (!bfqq->next_rq)
820 return;
821
822 bfqq->pos_root = &bfqq_group(bfqq)->rq_pos_tree;
823 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
824 blk_rq_pos(bfqq->next_rq), &parent, &p);
825 if (!__bfqq) {
826 rb_link_node(&bfqq->pos_node, parent, p);
827 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
828 } else
829 bfqq->pos_root = NULL;
830}
831
832/*
833 * The following function returns false either if every active queue
834 * must receive the same share of the throughput (symmetric scenario),
835 * or, as a special case, if bfqq must receive a share of the
836 * throughput lower than or equal to the share that every other active
837 * queue must receive. If bfqq does sync I/O, then these are the only
838 * two cases where bfqq happens to be guaranteed its share of the
839 * throughput even if I/O dispatching is not plugged when bfqq remains
840 * temporarily empty (for more details, see the comments in the
841 * function bfq_better_to_idle()). For this reason, the return value
842 * of this function is used to check whether I/O-dispatch plugging can
843 * be avoided.
844 *
845 * The above first case (symmetric scenario) occurs when:
846 * 1) all active queues have the same weight,
847 * 2) all active queues belong to the same I/O-priority class,
848 * 3) all active groups at the same level in the groups tree have the same
849 * weight,
850 * 4) all active groups at the same level in the groups tree have the same
851 * number of children.
852 *
853 * Unfortunately, keeping the necessary state for evaluating exactly
854 * the last two symmetry sub-conditions above would be quite complex
855 * and time consuming. Therefore this function evaluates, instead,
856 * only the following stronger three sub-conditions, for which it is
857 * much easier to maintain the needed state:
858 * 1) all active queues have the same weight,
859 * 2) all active queues belong to the same I/O-priority class,
860 * 3) there is at most one active group.
861 * In particular, the last condition is always true if hierarchical
862 * support or the cgroups interface are not enabled, thus no state
863 * needs to be maintained in this case.
864 */
865static bool bfq_asymmetric_scenario(struct bfq_data *bfqd,
866 struct bfq_queue *bfqq)
867{
868 bool smallest_weight = bfqq &&
869 bfqq->weight_counter &&
870 bfqq->weight_counter ==
871 container_of(
872 rb_first_cached(&bfqd->queue_weights_tree),
873 struct bfq_weight_counter,
874 weights_node);
875
876 /*
877 * For queue weights to differ, queue_weights_tree must contain
878 * at least two nodes.
879 */
880 bool varied_queue_weights = !smallest_weight &&
881 !RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) &&
882 (bfqd->queue_weights_tree.rb_root.rb_node->rb_left ||
883 bfqd->queue_weights_tree.rb_root.rb_node->rb_right);
884
885 bool multiple_classes_busy =
886 (bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
887 (bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
888 (bfqd->busy_queues[1] && bfqd->busy_queues[2]);
889
890 return varied_queue_weights || multiple_classes_busy
891#ifdef CONFIG_BFQ_GROUP_IOSCHED
892 || bfqd->num_groups_with_pending_reqs > 1
893#endif
894 ;
895}
896
897/*
898 * If the weight-counter tree passed as input contains no counter for
899 * the weight of the input queue, then add that counter; otherwise just
900 * increment the existing counter.
901 *
902 * Note that weight-counter trees contain few nodes in mostly symmetric
903 * scenarios. For example, if all queues have the same weight, then the
904 * weight-counter tree for the queues may contain at most one node.
905 * This holds even if low_latency is on, because weight-raised queues
906 * are not inserted in the tree.
907 * In most scenarios, the rate at which nodes are created/destroyed
908 * should be low too.
909 */
910void bfq_weights_tree_add(struct bfq_queue *bfqq)
911{
912 struct rb_root_cached *root = &bfqq->bfqd->queue_weights_tree;
913 struct bfq_entity *entity = &bfqq->entity;
914 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
915 bool leftmost = true;
916
917 /*
918 * Do not insert if the queue is already associated with a
919 * counter, which happens if:
920 * 1) a request arrival has caused the queue to become both
921 * non-weight-raised, and hence change its weight, and
922 * backlogged; in this respect, each of the two events
923 * causes an invocation of this function,
924 * 2) this is the invocation of this function caused by the
925 * second event. This second invocation is actually useless,
926 * and we handle this fact by exiting immediately. More
927 * efficient or clearer solutions might possibly be adopted.
928 */
929 if (bfqq->weight_counter)
930 return;
931
932 while (*new) {
933 struct bfq_weight_counter *__counter = container_of(*new,
934 struct bfq_weight_counter,
935 weights_node);
936 parent = *new;
937
938 if (entity->weight == __counter->weight) {
939 bfqq->weight_counter = __counter;
940 goto inc_counter;
941 }
942 if (entity->weight < __counter->weight)
943 new = &((*new)->rb_left);
944 else {
945 new = &((*new)->rb_right);
946 leftmost = false;
947 }
948 }
949
950 bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
951 GFP_ATOMIC);
952
953 /*
954 * In the unlucky event of an allocation failure, we just
955 * exit. This will cause the weight of queue to not be
956 * considered in bfq_asymmetric_scenario, which, in its turn,
957 * causes the scenario to be deemed wrongly symmetric in case
958 * bfqq's weight would have been the only weight making the
959 * scenario asymmetric. On the bright side, no unbalance will
960 * however occur when bfqq becomes inactive again (the
961 * invocation of this function is triggered by an activation
962 * of queue). In fact, bfq_weights_tree_remove does nothing
963 * if !bfqq->weight_counter.
964 */
965 if (unlikely(!bfqq->weight_counter))
966 return;
967
968 bfqq->weight_counter->weight = entity->weight;
969 rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
970 rb_insert_color_cached(&bfqq->weight_counter->weights_node, root,
971 leftmost);
972
973inc_counter:
974 bfqq->weight_counter->num_active++;
975 bfqq->ref++;
976}
977
978/*
979 * Decrement the weight counter associated with the queue, and, if the
980 * counter reaches 0, remove the counter from the tree.
981 * See the comments to the function bfq_weights_tree_add() for considerations
982 * about overhead.
983 */
984void bfq_weights_tree_remove(struct bfq_queue *bfqq)
985{
986 struct rb_root_cached *root;
987
988 if (!bfqq->weight_counter)
989 return;
990
991 root = &bfqq->bfqd->queue_weights_tree;
992 bfqq->weight_counter->num_active--;
993 if (bfqq->weight_counter->num_active > 0)
994 goto reset_entity_pointer;
995
996 rb_erase_cached(&bfqq->weight_counter->weights_node, root);
997 kfree(bfqq->weight_counter);
998
999reset_entity_pointer:
1000 bfqq->weight_counter = NULL;
1001 bfq_put_queue(bfqq);
1002}
1003
1004/*
1005 * Return expired entry, or NULL to just start from scratch in rbtree.
1006 */
1007static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
1008 struct request *last)
1009{
1010 struct request *rq;
1011
1012 if (bfq_bfqq_fifo_expire(bfqq))
1013 return NULL;
1014
1015 bfq_mark_bfqq_fifo_expire(bfqq);
1016
1017 rq = rq_entry_fifo(bfqq->fifo.next);
1018
1019 if (rq == last || blk_time_get_ns() < rq->fifo_time)
1020 return NULL;
1021
1022 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
1023 return rq;
1024}
1025
1026static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
1027 struct bfq_queue *bfqq,
1028 struct request *last)
1029{
1030 struct rb_node *rbnext = rb_next(&last->rb_node);
1031 struct rb_node *rbprev = rb_prev(&last->rb_node);
1032 struct request *next, *prev = NULL;
1033
1034 /* Follow expired path, else get first next available. */
1035 next = bfq_check_fifo(bfqq, last);
1036 if (next)
1037 return next;
1038
1039 if (rbprev)
1040 prev = rb_entry_rq(rbprev);
1041
1042 if (rbnext)
1043 next = rb_entry_rq(rbnext);
1044 else {
1045 rbnext = rb_first(&bfqq->sort_list);
1046 if (rbnext && rbnext != &last->rb_node)
1047 next = rb_entry_rq(rbnext);
1048 }
1049
1050 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
1051}
1052
1053/* see the definition of bfq_async_charge_factor for details */
1054static unsigned long bfq_serv_to_charge(struct request *rq,
1055 struct bfq_queue *bfqq)
1056{
1057 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
1058 bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
1059 return blk_rq_sectors(rq);
1060
1061 return blk_rq_sectors(rq) * bfq_async_charge_factor;
1062}
1063
1064/**
1065 * bfq_updated_next_req - update the queue after a new next_rq selection.
1066 * @bfqd: the device data the queue belongs to.
1067 * @bfqq: the queue to update.
1068 *
1069 * If the first request of a queue changes we make sure that the queue
1070 * has enough budget to serve at least its first request (if the
1071 * request has grown). We do this because if the queue has not enough
1072 * budget for its first request, it has to go through two dispatch
1073 * rounds to actually get it dispatched.
1074 */
1075static void bfq_updated_next_req(struct bfq_data *bfqd,
1076 struct bfq_queue *bfqq)
1077{
1078 struct bfq_entity *entity = &bfqq->entity;
1079 struct request *next_rq = bfqq->next_rq;
1080 unsigned long new_budget;
1081
1082 if (!next_rq)
1083 return;
1084
1085 if (bfqq == bfqd->in_service_queue)
1086 /*
1087 * In order not to break guarantees, budgets cannot be
1088 * changed after an entity has been selected.
1089 */
1090 return;
1091
1092 new_budget = max_t(unsigned long,
1093 max_t(unsigned long, bfqq->max_budget,
1094 bfq_serv_to_charge(next_rq, bfqq)),
1095 entity->service);
1096 if (entity->budget != new_budget) {
1097 entity->budget = new_budget;
1098 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
1099 new_budget);
1100 bfq_requeue_bfqq(bfqd, bfqq, false);
1101 }
1102}
1103
1104static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
1105{
1106 u64 dur;
1107
1108 dur = bfqd->rate_dur_prod;
1109 do_div(dur, bfqd->peak_rate);
1110
1111 /*
1112 * Limit duration between 3 and 25 seconds. The upper limit
1113 * has been conservatively set after the following worst case:
1114 * on a QEMU/KVM virtual machine
1115 * - running in a slow PC
1116 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
1117 * - serving a heavy I/O workload, such as the sequential reading
1118 * of several files
1119 * mplayer took 23 seconds to start, if constantly weight-raised.
1120 *
1121 * As for higher values than that accommodating the above bad
1122 * scenario, tests show that higher values would often yield
1123 * the opposite of the desired result, i.e., would worsen
1124 * responsiveness by allowing non-interactive applications to
1125 * preserve weight raising for too long.
1126 *
1127 * On the other end, lower values than 3 seconds make it
1128 * difficult for most interactive tasks to complete their jobs
1129 * before weight-raising finishes.
1130 */
1131 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
1132}
1133
1134/* switch back from soft real-time to interactive weight raising */
1135static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
1136 struct bfq_data *bfqd)
1137{
1138 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1139 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1140 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
1141}
1142
1143static void
1144bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
1145 struct bfq_io_cq *bic, bool bfq_already_existing)
1146{
1147 unsigned int old_wr_coeff = 1;
1148 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
1149 unsigned int a_idx = bfqq->actuator_idx;
1150 struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[a_idx];
1151
1152 if (bfqq_data->saved_has_short_ttime)
1153 bfq_mark_bfqq_has_short_ttime(bfqq);
1154 else
1155 bfq_clear_bfqq_has_short_ttime(bfqq);
1156
1157 if (bfqq_data->saved_IO_bound)
1158 bfq_mark_bfqq_IO_bound(bfqq);
1159 else
1160 bfq_clear_bfqq_IO_bound(bfqq);
1161
1162 bfqq->last_serv_time_ns = bfqq_data->saved_last_serv_time_ns;
1163 bfqq->inject_limit = bfqq_data->saved_inject_limit;
1164 bfqq->decrease_time_jif = bfqq_data->saved_decrease_time_jif;
1165
1166 bfqq->entity.new_weight = bfqq_data->saved_weight;
1167 bfqq->ttime = bfqq_data->saved_ttime;
1168 bfqq->io_start_time = bfqq_data->saved_io_start_time;
1169 bfqq->tot_idle_time = bfqq_data->saved_tot_idle_time;
1170 /*
1171 * Restore weight coefficient only if low_latency is on
1172 */
1173 if (bfqd->low_latency) {
1174 old_wr_coeff = bfqq->wr_coeff;
1175 bfqq->wr_coeff = bfqq_data->saved_wr_coeff;
1176 }
1177 bfqq->service_from_wr = bfqq_data->saved_service_from_wr;
1178 bfqq->wr_start_at_switch_to_srt =
1179 bfqq_data->saved_wr_start_at_switch_to_srt;
1180 bfqq->last_wr_start_finish = bfqq_data->saved_last_wr_start_finish;
1181 bfqq->wr_cur_max_time = bfqq_data->saved_wr_cur_max_time;
1182
1183 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1184 time_is_before_jiffies(bfqq->last_wr_start_finish +
1185 bfqq->wr_cur_max_time))) {
1186 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1187 !bfq_bfqq_in_large_burst(bfqq) &&
1188 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1189 bfq_wr_duration(bfqd))) {
1190 switch_back_to_interactive_wr(bfqq, bfqd);
1191 } else {
1192 bfqq->wr_coeff = 1;
1193 bfq_log_bfqq(bfqq->bfqd, bfqq,
1194 "resume state: switching off wr");
1195 }
1196 }
1197
1198 /* make sure weight will be updated, however we got here */
1199 bfqq->entity.prio_changed = 1;
1200
1201 if (likely(!busy))
1202 return;
1203
1204 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1205 bfqd->wr_busy_queues++;
1206 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1207 bfqd->wr_busy_queues--;
1208}
1209
1210static int bfqq_process_refs(struct bfq_queue *bfqq)
1211{
1212 return bfqq->ref - bfqq->entity.allocated -
1213 bfqq->entity.on_st_or_in_serv -
1214 (bfqq->weight_counter != NULL) - bfqq->stable_ref;
1215}
1216
1217/* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1218static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1219{
1220 struct bfq_queue *item;
1221 struct hlist_node *n;
1222
1223 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1224 hlist_del_init(&item->burst_list_node);
1225
1226 /*
1227 * Start the creation of a new burst list only if there is no
1228 * active queue. See comments on the conditional invocation of
1229 * bfq_handle_burst().
1230 */
1231 if (bfq_tot_busy_queues(bfqd) == 0) {
1232 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1233 bfqd->burst_size = 1;
1234 } else
1235 bfqd->burst_size = 0;
1236
1237 bfqd->burst_parent_entity = bfqq->entity.parent;
1238}
1239
1240/* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1241static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1242{
1243 /* Increment burst size to take into account also bfqq */
1244 bfqd->burst_size++;
1245
1246 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1247 struct bfq_queue *pos, *bfqq_item;
1248 struct hlist_node *n;
1249
1250 /*
1251 * Enough queues have been activated shortly after each
1252 * other to consider this burst as large.
1253 */
1254 bfqd->large_burst = true;
1255
1256 /*
1257 * We can now mark all queues in the burst list as
1258 * belonging to a large burst.
1259 */
1260 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1261 burst_list_node)
1262 bfq_mark_bfqq_in_large_burst(bfqq_item);
1263 bfq_mark_bfqq_in_large_burst(bfqq);
1264
1265 /*
1266 * From now on, and until the current burst finishes, any
1267 * new queue being activated shortly after the last queue
1268 * was inserted in the burst can be immediately marked as
1269 * belonging to a large burst. So the burst list is not
1270 * needed any more. Remove it.
1271 */
1272 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1273 burst_list_node)
1274 hlist_del_init(&pos->burst_list_node);
1275 } else /*
1276 * Burst not yet large: add bfqq to the burst list. Do
1277 * not increment the ref counter for bfqq, because bfqq
1278 * is removed from the burst list before freeing bfqq
1279 * in put_queue.
1280 */
1281 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1282}
1283
1284/*
1285 * If many queues belonging to the same group happen to be created
1286 * shortly after each other, then the processes associated with these
1287 * queues have typically a common goal. In particular, bursts of queue
1288 * creations are usually caused by services or applications that spawn
1289 * many parallel threads/processes. Examples are systemd during boot,
1290 * or git grep. To help these processes get their job done as soon as
1291 * possible, it is usually better to not grant either weight-raising
1292 * or device idling to their queues, unless these queues must be
1293 * protected from the I/O flowing through other active queues.
1294 *
1295 * In this comment we describe, firstly, the reasons why this fact
1296 * holds, and, secondly, the next function, which implements the main
1297 * steps needed to properly mark these queues so that they can then be
1298 * treated in a different way.
1299 *
1300 * The above services or applications benefit mostly from a high
1301 * throughput: the quicker the requests of the activated queues are
1302 * cumulatively served, the sooner the target job of these queues gets
1303 * completed. As a consequence, weight-raising any of these queues,
1304 * which also implies idling the device for it, is almost always
1305 * counterproductive, unless there are other active queues to isolate
1306 * these new queues from. If there no other active queues, then
1307 * weight-raising these new queues just lowers throughput in most
1308 * cases.
1309 *
1310 * On the other hand, a burst of queue creations may be caused also by
1311 * the start of an application that does not consist of a lot of
1312 * parallel I/O-bound threads. In fact, with a complex application,
1313 * several short processes may need to be executed to start-up the
1314 * application. In this respect, to start an application as quickly as
1315 * possible, the best thing to do is in any case to privilege the I/O
1316 * related to the application with respect to all other
1317 * I/O. Therefore, the best strategy to start as quickly as possible
1318 * an application that causes a burst of queue creations is to
1319 * weight-raise all the queues created during the burst. This is the
1320 * exact opposite of the best strategy for the other type of bursts.
1321 *
1322 * In the end, to take the best action for each of the two cases, the
1323 * two types of bursts need to be distinguished. Fortunately, this
1324 * seems relatively easy, by looking at the sizes of the bursts. In
1325 * particular, we found a threshold such that only bursts with a
1326 * larger size than that threshold are apparently caused by
1327 * services or commands such as systemd or git grep. For brevity,
1328 * hereafter we call just 'large' these bursts. BFQ *does not*
1329 * weight-raise queues whose creation occurs in a large burst. In
1330 * addition, for each of these queues BFQ performs or does not perform
1331 * idling depending on which choice boosts the throughput more. The
1332 * exact choice depends on the device and request pattern at
1333 * hand.
1334 *
1335 * Unfortunately, false positives may occur while an interactive task
1336 * is starting (e.g., an application is being started). The
1337 * consequence is that the queues associated with the task do not
1338 * enjoy weight raising as expected. Fortunately these false positives
1339 * are very rare. They typically occur if some service happens to
1340 * start doing I/O exactly when the interactive task starts.
1341 *
1342 * Turning back to the next function, it is invoked only if there are
1343 * no active queues (apart from active queues that would belong to the
1344 * same, possible burst bfqq would belong to), and it implements all
1345 * the steps needed to detect the occurrence of a large burst and to
1346 * properly mark all the queues belonging to it (so that they can then
1347 * be treated in a different way). This goal is achieved by
1348 * maintaining a "burst list" that holds, temporarily, the queues that
1349 * belong to the burst in progress. The list is then used to mark
1350 * these queues as belonging to a large burst if the burst does become
1351 * large. The main steps are the following.
1352 *
1353 * . when the very first queue is created, the queue is inserted into the
1354 * list (as it could be the first queue in a possible burst)
1355 *
1356 * . if the current burst has not yet become large, and a queue Q that does
1357 * not yet belong to the burst is activated shortly after the last time
1358 * at which a new queue entered the burst list, then the function appends
1359 * Q to the burst list
1360 *
1361 * . if, as a consequence of the previous step, the burst size reaches
1362 * the large-burst threshold, then
1363 *
1364 * . all the queues in the burst list are marked as belonging to a
1365 * large burst
1366 *
1367 * . the burst list is deleted; in fact, the burst list already served
1368 * its purpose (keeping temporarily track of the queues in a burst,
1369 * so as to be able to mark them as belonging to a large burst in the
1370 * previous sub-step), and now is not needed any more
1371 *
1372 * . the device enters a large-burst mode
1373 *
1374 * . if a queue Q that does not belong to the burst is created while
1375 * the device is in large-burst mode and shortly after the last time
1376 * at which a queue either entered the burst list or was marked as
1377 * belonging to the current large burst, then Q is immediately marked
1378 * as belonging to a large burst.
1379 *
1380 * . if a queue Q that does not belong to the burst is created a while
1381 * later, i.e., not shortly after, than the last time at which a queue
1382 * either entered the burst list or was marked as belonging to the
1383 * current large burst, then the current burst is deemed as finished and:
1384 *
1385 * . the large-burst mode is reset if set
1386 *
1387 * . the burst list is emptied
1388 *
1389 * . Q is inserted in the burst list, as Q may be the first queue
1390 * in a possible new burst (then the burst list contains just Q
1391 * after this step).
1392 */
1393static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1394{
1395 /*
1396 * If bfqq is already in the burst list or is part of a large
1397 * burst, or finally has just been split, then there is
1398 * nothing else to do.
1399 */
1400 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1401 bfq_bfqq_in_large_burst(bfqq) ||
1402 time_is_after_eq_jiffies(bfqq->split_time +
1403 msecs_to_jiffies(10)))
1404 return;
1405
1406 /*
1407 * If bfqq's creation happens late enough, or bfqq belongs to
1408 * a different group than the burst group, then the current
1409 * burst is finished, and related data structures must be
1410 * reset.
1411 *
1412 * In this respect, consider the special case where bfqq is
1413 * the very first queue created after BFQ is selected for this
1414 * device. In this case, last_ins_in_burst and
1415 * burst_parent_entity are not yet significant when we get
1416 * here. But it is easy to verify that, whether or not the
1417 * following condition is true, bfqq will end up being
1418 * inserted into the burst list. In particular the list will
1419 * happen to contain only bfqq. And this is exactly what has
1420 * to happen, as bfqq may be the first queue of the first
1421 * burst.
1422 */
1423 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1424 bfqd->bfq_burst_interval) ||
1425 bfqq->entity.parent != bfqd->burst_parent_entity) {
1426 bfqd->large_burst = false;
1427 bfq_reset_burst_list(bfqd, bfqq);
1428 goto end;
1429 }
1430
1431 /*
1432 * If we get here, then bfqq is being activated shortly after the
1433 * last queue. So, if the current burst is also large, we can mark
1434 * bfqq as belonging to this large burst immediately.
1435 */
1436 if (bfqd->large_burst) {
1437 bfq_mark_bfqq_in_large_burst(bfqq);
1438 goto end;
1439 }
1440
1441 /*
1442 * If we get here, then a large-burst state has not yet been
1443 * reached, but bfqq is being activated shortly after the last
1444 * queue. Then we add bfqq to the burst.
1445 */
1446 bfq_add_to_burst(bfqd, bfqq);
1447end:
1448 /*
1449 * At this point, bfqq either has been added to the current
1450 * burst or has caused the current burst to terminate and a
1451 * possible new burst to start. In particular, in the second
1452 * case, bfqq has become the first queue in the possible new
1453 * burst. In both cases last_ins_in_burst needs to be moved
1454 * forward.
1455 */
1456 bfqd->last_ins_in_burst = jiffies;
1457}
1458
1459static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1460{
1461 struct bfq_entity *entity = &bfqq->entity;
1462
1463 return entity->budget - entity->service;
1464}
1465
1466/*
1467 * If enough samples have been computed, return the current max budget
1468 * stored in bfqd, which is dynamically updated according to the
1469 * estimated disk peak rate; otherwise return the default max budget
1470 */
1471static int bfq_max_budget(struct bfq_data *bfqd)
1472{
1473 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1474 return bfq_default_max_budget;
1475 else
1476 return bfqd->bfq_max_budget;
1477}
1478
1479/*
1480 * Return min budget, which is a fraction of the current or default
1481 * max budget (trying with 1/32)
1482 */
1483static int bfq_min_budget(struct bfq_data *bfqd)
1484{
1485 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1486 return bfq_default_max_budget / 32;
1487 else
1488 return bfqd->bfq_max_budget / 32;
1489}
1490
1491/*
1492 * The next function, invoked after the input queue bfqq switches from
1493 * idle to busy, updates the budget of bfqq. The function also tells
1494 * whether the in-service queue should be expired, by returning
1495 * true. The purpose of expiring the in-service queue is to give bfqq
1496 * the chance to possibly preempt the in-service queue, and the reason
1497 * for preempting the in-service queue is to achieve one of the two
1498 * goals below.
1499 *
1500 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1501 * expired because it has remained idle. In particular, bfqq may have
1502 * expired for one of the following two reasons:
1503 *
1504 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1505 * and did not make it to issue a new request before its last
1506 * request was served;
1507 *
1508 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1509 * a new request before the expiration of the idling-time.
1510 *
1511 * Even if bfqq has expired for one of the above reasons, the process
1512 * associated with the queue may be however issuing requests greedily,
1513 * and thus be sensitive to the bandwidth it receives (bfqq may have
1514 * remained idle for other reasons: CPU high load, bfqq not enjoying
1515 * idling, I/O throttling somewhere in the path from the process to
1516 * the I/O scheduler, ...). But if, after every expiration for one of
1517 * the above two reasons, bfqq has to wait for the service of at least
1518 * one full budget of another queue before being served again, then
1519 * bfqq is likely to get a much lower bandwidth or resource time than
1520 * its reserved ones. To address this issue, two countermeasures need
1521 * to be taken.
1522 *
1523 * First, the budget and the timestamps of bfqq need to be updated in
1524 * a special way on bfqq reactivation: they need to be updated as if
1525 * bfqq did not remain idle and did not expire. In fact, if they are
1526 * computed as if bfqq expired and remained idle until reactivation,
1527 * then the process associated with bfqq is treated as if, instead of
1528 * being greedy, it stopped issuing requests when bfqq remained idle,
1529 * and restarts issuing requests only on this reactivation. In other
1530 * words, the scheduler does not help the process recover the "service
1531 * hole" between bfqq expiration and reactivation. As a consequence,
1532 * the process receives a lower bandwidth than its reserved one. In
1533 * contrast, to recover this hole, the budget must be updated as if
1534 * bfqq was not expired at all before this reactivation, i.e., it must
1535 * be set to the value of the remaining budget when bfqq was
1536 * expired. Along the same line, timestamps need to be assigned the
1537 * value they had the last time bfqq was selected for service, i.e.,
1538 * before last expiration. Thus timestamps need to be back-shifted
1539 * with respect to their normal computation (see [1] for more details
1540 * on this tricky aspect).
1541 *
1542 * Secondly, to allow the process to recover the hole, the in-service
1543 * queue must be expired too, to give bfqq the chance to preempt it
1544 * immediately. In fact, if bfqq has to wait for a full budget of the
1545 * in-service queue to be completed, then it may become impossible to
1546 * let the process recover the hole, even if the back-shifted
1547 * timestamps of bfqq are lower than those of the in-service queue. If
1548 * this happens for most or all of the holes, then the process may not
1549 * receive its reserved bandwidth. In this respect, it is worth noting
1550 * that, being the service of outstanding requests unpreemptible, a
1551 * little fraction of the holes may however be unrecoverable, thereby
1552 * causing a little loss of bandwidth.
1553 *
1554 * The last important point is detecting whether bfqq does need this
1555 * bandwidth recovery. In this respect, the next function deems the
1556 * process associated with bfqq greedy, and thus allows it to recover
1557 * the hole, if: 1) the process is waiting for the arrival of a new
1558 * request (which implies that bfqq expired for one of the above two
1559 * reasons), and 2) such a request has arrived soon. The first
1560 * condition is controlled through the flag non_blocking_wait_rq,
1561 * while the second through the flag arrived_in_time. If both
1562 * conditions hold, then the function computes the budget in the
1563 * above-described special way, and signals that the in-service queue
1564 * should be expired. Timestamp back-shifting is done later in
1565 * __bfq_activate_entity.
1566 *
1567 * 2. Reduce latency. Even if timestamps are not backshifted to let
1568 * the process associated with bfqq recover a service hole, bfqq may
1569 * however happen to have, after being (re)activated, a lower finish
1570 * timestamp than the in-service queue. That is, the next budget of
1571 * bfqq may have to be completed before the one of the in-service
1572 * queue. If this is the case, then preempting the in-service queue
1573 * allows this goal to be achieved, apart from the unpreemptible,
1574 * outstanding requests mentioned above.
1575 *
1576 * Unfortunately, regardless of which of the above two goals one wants
1577 * to achieve, service trees need first to be updated to know whether
1578 * the in-service queue must be preempted. To have service trees
1579 * correctly updated, the in-service queue must be expired and
1580 * rescheduled, and bfqq must be scheduled too. This is one of the
1581 * most costly operations (in future versions, the scheduling
1582 * mechanism may be re-designed in such a way to make it possible to
1583 * know whether preemption is needed without needing to update service
1584 * trees). In addition, queue preemptions almost always cause random
1585 * I/O, which may in turn cause loss of throughput. Finally, there may
1586 * even be no in-service queue when the next function is invoked (so,
1587 * no queue to compare timestamps with). Because of these facts, the
1588 * next function adopts the following simple scheme to avoid costly
1589 * operations, too frequent preemptions and too many dependencies on
1590 * the state of the scheduler: it requests the expiration of the
1591 * in-service queue (unconditionally) only for queues that need to
1592 * recover a hole. Then it delegates to other parts of the code the
1593 * responsibility of handling the above case 2.
1594 */
1595static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1596 struct bfq_queue *bfqq,
1597 bool arrived_in_time)
1598{
1599 struct bfq_entity *entity = &bfqq->entity;
1600
1601 /*
1602 * In the next compound condition, we check also whether there
1603 * is some budget left, because otherwise there is no point in
1604 * trying to go on serving bfqq with this same budget: bfqq
1605 * would be expired immediately after being selected for
1606 * service. This would only cause useless overhead.
1607 */
1608 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1609 bfq_bfqq_budget_left(bfqq) > 0) {
1610 /*
1611 * We do not clear the flag non_blocking_wait_rq here, as
1612 * the latter is used in bfq_activate_bfqq to signal
1613 * that timestamps need to be back-shifted (and is
1614 * cleared right after).
1615 */
1616
1617 /*
1618 * In next assignment we rely on that either
1619 * entity->service or entity->budget are not updated
1620 * on expiration if bfqq is empty (see
1621 * __bfq_bfqq_recalc_budget). Thus both quantities
1622 * remain unchanged after such an expiration, and the
1623 * following statement therefore assigns to
1624 * entity->budget the remaining budget on such an
1625 * expiration.
1626 */
1627 entity->budget = min_t(unsigned long,
1628 bfq_bfqq_budget_left(bfqq),
1629 bfqq->max_budget);
1630
1631 /*
1632 * At this point, we have used entity->service to get
1633 * the budget left (needed for updating
1634 * entity->budget). Thus we finally can, and have to,
1635 * reset entity->service. The latter must be reset
1636 * because bfqq would otherwise be charged again for
1637 * the service it has received during its previous
1638 * service slot(s).
1639 */
1640 entity->service = 0;
1641
1642 return true;
1643 }
1644
1645 /*
1646 * We can finally complete expiration, by setting service to 0.
1647 */
1648 entity->service = 0;
1649 entity->budget = max_t(unsigned long, bfqq->max_budget,
1650 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1651 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1652 return false;
1653}
1654
1655/*
1656 * Return the farthest past time instant according to jiffies
1657 * macros.
1658 */
1659static unsigned long bfq_smallest_from_now(void)
1660{
1661 return jiffies - MAX_JIFFY_OFFSET;
1662}
1663
1664static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1665 struct bfq_queue *bfqq,
1666 unsigned int old_wr_coeff,
1667 bool wr_or_deserves_wr,
1668 bool interactive,
1669 bool in_burst,
1670 bool soft_rt)
1671{
1672 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1673 /* start a weight-raising period */
1674 if (interactive) {
1675 bfqq->service_from_wr = 0;
1676 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1677 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1678 } else {
1679 /*
1680 * No interactive weight raising in progress
1681 * here: assign minus infinity to
1682 * wr_start_at_switch_to_srt, to make sure
1683 * that, at the end of the soft-real-time
1684 * weight raising periods that is starting
1685 * now, no interactive weight-raising period
1686 * may be wrongly considered as still in
1687 * progress (and thus actually started by
1688 * mistake).
1689 */
1690 bfqq->wr_start_at_switch_to_srt =
1691 bfq_smallest_from_now();
1692 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1693 BFQ_SOFTRT_WEIGHT_FACTOR;
1694 bfqq->wr_cur_max_time =
1695 bfqd->bfq_wr_rt_max_time;
1696 }
1697
1698 /*
1699 * If needed, further reduce budget to make sure it is
1700 * close to bfqq's backlog, so as to reduce the
1701 * scheduling-error component due to a too large
1702 * budget. Do not care about throughput consequences,
1703 * but only about latency. Finally, do not assign a
1704 * too small budget either, to avoid increasing
1705 * latency by causing too frequent expirations.
1706 */
1707 bfqq->entity.budget = min_t(unsigned long,
1708 bfqq->entity.budget,
1709 2 * bfq_min_budget(bfqd));
1710 } else if (old_wr_coeff > 1) {
1711 if (interactive) { /* update wr coeff and duration */
1712 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1713 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1714 } else if (in_burst)
1715 bfqq->wr_coeff = 1;
1716 else if (soft_rt) {
1717 /*
1718 * The application is now or still meeting the
1719 * requirements for being deemed soft rt. We
1720 * can then correctly and safely (re)charge
1721 * the weight-raising duration for the
1722 * application with the weight-raising
1723 * duration for soft rt applications.
1724 *
1725 * In particular, doing this recharge now, i.e.,
1726 * before the weight-raising period for the
1727 * application finishes, reduces the probability
1728 * of the following negative scenario:
1729 * 1) the weight of a soft rt application is
1730 * raised at startup (as for any newly
1731 * created application),
1732 * 2) since the application is not interactive,
1733 * at a certain time weight-raising is
1734 * stopped for the application,
1735 * 3) at that time the application happens to
1736 * still have pending requests, and hence
1737 * is destined to not have a chance to be
1738 * deemed soft rt before these requests are
1739 * completed (see the comments to the
1740 * function bfq_bfqq_softrt_next_start()
1741 * for details on soft rt detection),
1742 * 4) these pending requests experience a high
1743 * latency because the application is not
1744 * weight-raised while they are pending.
1745 */
1746 if (bfqq->wr_cur_max_time !=
1747 bfqd->bfq_wr_rt_max_time) {
1748 bfqq->wr_start_at_switch_to_srt =
1749 bfqq->last_wr_start_finish;
1750
1751 bfqq->wr_cur_max_time =
1752 bfqd->bfq_wr_rt_max_time;
1753 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1754 BFQ_SOFTRT_WEIGHT_FACTOR;
1755 }
1756 bfqq->last_wr_start_finish = jiffies;
1757 }
1758 }
1759}
1760
1761static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1762 struct bfq_queue *bfqq)
1763{
1764 return bfqq->dispatched == 0 &&
1765 time_is_before_jiffies(
1766 bfqq->budget_timeout +
1767 bfqd->bfq_wr_min_idle_time);
1768}
1769
1770
1771/*
1772 * Return true if bfqq is in a higher priority class, or has a higher
1773 * weight than the in-service queue.
1774 */
1775static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq,
1776 struct bfq_queue *in_serv_bfqq)
1777{
1778 int bfqq_weight, in_serv_weight;
1779
1780 if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class)
1781 return true;
1782
1783 if (in_serv_bfqq->entity.parent == bfqq->entity.parent) {
1784 bfqq_weight = bfqq->entity.weight;
1785 in_serv_weight = in_serv_bfqq->entity.weight;
1786 } else {
1787 if (bfqq->entity.parent)
1788 bfqq_weight = bfqq->entity.parent->weight;
1789 else
1790 bfqq_weight = bfqq->entity.weight;
1791 if (in_serv_bfqq->entity.parent)
1792 in_serv_weight = in_serv_bfqq->entity.parent->weight;
1793 else
1794 in_serv_weight = in_serv_bfqq->entity.weight;
1795 }
1796
1797 return bfqq_weight > in_serv_weight;
1798}
1799
1800/*
1801 * Get the index of the actuator that will serve bio.
1802 */
1803static unsigned int bfq_actuator_index(struct bfq_data *bfqd, struct bio *bio)
1804{
1805 unsigned int i;
1806 sector_t end;
1807
1808 /* no search needed if one or zero ranges present */
1809 if (bfqd->num_actuators == 1)
1810 return 0;
1811
1812 /* bio_end_sector(bio) gives the sector after the last one */
1813 end = bio_end_sector(bio) - 1;
1814
1815 for (i = 0; i < bfqd->num_actuators; i++) {
1816 if (end >= bfqd->sector[i] &&
1817 end < bfqd->sector[i] + bfqd->nr_sectors[i])
1818 return i;
1819 }
1820
1821 WARN_ONCE(true,
1822 "bfq_actuator_index: bio sector out of ranges: end=%llu\n",
1823 end);
1824 return 0;
1825}
1826
1827static bool bfq_better_to_idle(struct bfq_queue *bfqq);
1828
1829static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1830 struct bfq_queue *bfqq,
1831 int old_wr_coeff,
1832 struct request *rq,
1833 bool *interactive)
1834{
1835 bool soft_rt, in_burst, wr_or_deserves_wr,
1836 bfqq_wants_to_preempt,
1837 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1838 /*
1839 * See the comments on
1840 * bfq_bfqq_update_budg_for_activation for
1841 * details on the usage of the next variable.
1842 */
1843 arrived_in_time = blk_time_get_ns() <=
1844 bfqq->ttime.last_end_request +
1845 bfqd->bfq_slice_idle * 3;
1846 unsigned int act_idx = bfq_actuator_index(bfqd, rq->bio);
1847 bool bfqq_non_merged_or_stably_merged =
1848 bfqq->bic || RQ_BIC(rq)->bfqq_data[act_idx].stably_merged;
1849
1850 /*
1851 * bfqq deserves to be weight-raised if:
1852 * - it is sync,
1853 * - it does not belong to a large burst,
1854 * - it has been idle for enough time or is soft real-time,
1855 * - is linked to a bfq_io_cq (it is not shared in any sense),
1856 * - has a default weight (otherwise we assume the user wanted
1857 * to control its weight explicitly)
1858 */
1859 in_burst = bfq_bfqq_in_large_burst(bfqq);
1860 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1861 !BFQQ_TOTALLY_SEEKY(bfqq) &&
1862 !in_burst &&
1863 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1864 bfqq->dispatched == 0 &&
1865 bfqq->entity.new_weight == 40;
1866 *interactive = !in_burst && idle_for_long_time &&
1867 bfqq->entity.new_weight == 40;
1868 /*
1869 * Merged bfq_queues are kept out of weight-raising
1870 * (low-latency) mechanisms. The reason is that these queues
1871 * are usually created for non-interactive and
1872 * non-soft-real-time tasks. Yet this is not the case for
1873 * stably-merged queues. These queues are merged just because
1874 * they are created shortly after each other. So they may
1875 * easily serve the I/O of an interactive or soft-real time
1876 * application, if the application happens to spawn multiple
1877 * processes. So let also stably-merged queued enjoy weight
1878 * raising.
1879 */
1880 wr_or_deserves_wr = bfqd->low_latency &&
1881 (bfqq->wr_coeff > 1 ||
1882 (bfq_bfqq_sync(bfqq) && bfqq_non_merged_or_stably_merged &&
1883 (*interactive || soft_rt)));
1884
1885 /*
1886 * Using the last flag, update budget and check whether bfqq
1887 * may want to preempt the in-service queue.
1888 */
1889 bfqq_wants_to_preempt =
1890 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1891 arrived_in_time);
1892
1893 /*
1894 * If bfqq happened to be activated in a burst, but has been
1895 * idle for much more than an interactive queue, then we
1896 * assume that, in the overall I/O initiated in the burst, the
1897 * I/O associated with bfqq is finished. So bfqq does not need
1898 * to be treated as a queue belonging to a burst
1899 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1900 * if set, and remove bfqq from the burst list if it's
1901 * there. We do not decrement burst_size, because the fact
1902 * that bfqq does not need to belong to the burst list any
1903 * more does not invalidate the fact that bfqq was created in
1904 * a burst.
1905 */
1906 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1907 idle_for_long_time &&
1908 time_is_before_jiffies(
1909 bfqq->budget_timeout +
1910 msecs_to_jiffies(10000))) {
1911 hlist_del_init(&bfqq->burst_list_node);
1912 bfq_clear_bfqq_in_large_burst(bfqq);
1913 }
1914
1915 bfq_clear_bfqq_just_created(bfqq);
1916
1917 if (bfqd->low_latency) {
1918 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1919 /* wraparound */
1920 bfqq->split_time =
1921 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1922
1923 if (time_is_before_jiffies(bfqq->split_time +
1924 bfqd->bfq_wr_min_idle_time)) {
1925 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1926 old_wr_coeff,
1927 wr_or_deserves_wr,
1928 *interactive,
1929 in_burst,
1930 soft_rt);
1931
1932 if (old_wr_coeff != bfqq->wr_coeff)
1933 bfqq->entity.prio_changed = 1;
1934 }
1935 }
1936
1937 bfqq->last_idle_bklogged = jiffies;
1938 bfqq->service_from_backlogged = 0;
1939 bfq_clear_bfqq_softrt_update(bfqq);
1940
1941 bfq_add_bfqq_busy(bfqq);
1942
1943 /*
1944 * Expire in-service queue if preemption may be needed for
1945 * guarantees or throughput. As for guarantees, we care
1946 * explicitly about two cases. The first is that bfqq has to
1947 * recover a service hole, as explained in the comments on
1948 * bfq_bfqq_update_budg_for_activation(), i.e., that
1949 * bfqq_wants_to_preempt is true. However, if bfqq does not
1950 * carry time-critical I/O, then bfqq's bandwidth is less
1951 * important than that of queues that carry time-critical I/O.
1952 * So, as a further constraint, we consider this case only if
1953 * bfqq is at least as weight-raised, i.e., at least as time
1954 * critical, as the in-service queue.
1955 *
1956 * The second case is that bfqq is in a higher priority class,
1957 * or has a higher weight than the in-service queue. If this
1958 * condition does not hold, we don't care because, even if
1959 * bfqq does not start to be served immediately, the resulting
1960 * delay for bfqq's I/O is however lower or much lower than
1961 * the ideal completion time to be guaranteed to bfqq's I/O.
1962 *
1963 * In both cases, preemption is needed only if, according to
1964 * the timestamps of both bfqq and of the in-service queue,
1965 * bfqq actually is the next queue to serve. So, to reduce
1966 * useless preemptions, the return value of
1967 * next_queue_may_preempt() is considered in the next compound
1968 * condition too. Yet next_queue_may_preempt() just checks a
1969 * simple, necessary condition for bfqq to be the next queue
1970 * to serve. In fact, to evaluate a sufficient condition, the
1971 * timestamps of the in-service queue would need to be
1972 * updated, and this operation is quite costly (see the
1973 * comments on bfq_bfqq_update_budg_for_activation()).
1974 *
1975 * As for throughput, we ask bfq_better_to_idle() whether we
1976 * still need to plug I/O dispatching. If bfq_better_to_idle()
1977 * says no, then plugging is not needed any longer, either to
1978 * boost throughput or to perserve service guarantees. Then
1979 * the best option is to stop plugging I/O, as not doing so
1980 * would certainly lower throughput. We may end up in this
1981 * case if: (1) upon a dispatch attempt, we detected that it
1982 * was better to plug I/O dispatch, and to wait for a new
1983 * request to arrive for the currently in-service queue, but
1984 * (2) this switch of bfqq to busy changes the scenario.
1985 */
1986 if (bfqd->in_service_queue &&
1987 ((bfqq_wants_to_preempt &&
1988 bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) ||
1989 bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue) ||
1990 !bfq_better_to_idle(bfqd->in_service_queue)) &&
1991 next_queue_may_preempt(bfqd))
1992 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1993 false, BFQQE_PREEMPTED);
1994}
1995
1996static void bfq_reset_inject_limit(struct bfq_data *bfqd,
1997 struct bfq_queue *bfqq)
1998{
1999 /* invalidate baseline total service time */
2000 bfqq->last_serv_time_ns = 0;
2001
2002 /*
2003 * Reset pointer in case we are waiting for
2004 * some request completion.
2005 */
2006 bfqd->waited_rq = NULL;
2007
2008 /*
2009 * If bfqq has a short think time, then start by setting the
2010 * inject limit to 0 prudentially, because the service time of
2011 * an injected I/O request may be higher than the think time
2012 * of bfqq, and therefore, if one request was injected when
2013 * bfqq remains empty, this injected request might delay the
2014 * service of the next I/O request for bfqq significantly. In
2015 * case bfqq can actually tolerate some injection, then the
2016 * adaptive update will however raise the limit soon. This
2017 * lucky circumstance holds exactly because bfqq has a short
2018 * think time, and thus, after remaining empty, is likely to
2019 * get new I/O enqueued---and then completed---before being
2020 * expired. This is the very pattern that gives the
2021 * limit-update algorithm the chance to measure the effect of
2022 * injection on request service times, and then to update the
2023 * limit accordingly.
2024 *
2025 * However, in the following special case, the inject limit is
2026 * left to 1 even if the think time is short: bfqq's I/O is
2027 * synchronized with that of some other queue, i.e., bfqq may
2028 * receive new I/O only after the I/O of the other queue is
2029 * completed. Keeping the inject limit to 1 allows the
2030 * blocking I/O to be served while bfqq is in service. And
2031 * this is very convenient both for bfqq and for overall
2032 * throughput, as explained in detail in the comments in
2033 * bfq_update_has_short_ttime().
2034 *
2035 * On the opposite end, if bfqq has a long think time, then
2036 * start directly by 1, because:
2037 * a) on the bright side, keeping at most one request in
2038 * service in the drive is unlikely to cause any harm to the
2039 * latency of bfqq's requests, as the service time of a single
2040 * request is likely to be lower than the think time of bfqq;
2041 * b) on the downside, after becoming empty, bfqq is likely to
2042 * expire before getting its next request. With this request
2043 * arrival pattern, it is very hard to sample total service
2044 * times and update the inject limit accordingly (see comments
2045 * on bfq_update_inject_limit()). So the limit is likely to be
2046 * never, or at least seldom, updated. As a consequence, by
2047 * setting the limit to 1, we avoid that no injection ever
2048 * occurs with bfqq. On the downside, this proactive step
2049 * further reduces chances to actually compute the baseline
2050 * total service time. Thus it reduces chances to execute the
2051 * limit-update algorithm and possibly raise the limit to more
2052 * than 1.
2053 */
2054 if (bfq_bfqq_has_short_ttime(bfqq))
2055 bfqq->inject_limit = 0;
2056 else
2057 bfqq->inject_limit = 1;
2058
2059 bfqq->decrease_time_jif = jiffies;
2060}
2061
2062static void bfq_update_io_intensity(struct bfq_queue *bfqq, u64 now_ns)
2063{
2064 u64 tot_io_time = now_ns - bfqq->io_start_time;
2065
2066 if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfqq->dispatched == 0)
2067 bfqq->tot_idle_time +=
2068 now_ns - bfqq->ttime.last_end_request;
2069
2070 if (unlikely(bfq_bfqq_just_created(bfqq)))
2071 return;
2072
2073 /*
2074 * Must be busy for at least about 80% of the time to be
2075 * considered I/O bound.
2076 */
2077 if (bfqq->tot_idle_time * 5 > tot_io_time)
2078 bfq_clear_bfqq_IO_bound(bfqq);
2079 else
2080 bfq_mark_bfqq_IO_bound(bfqq);
2081
2082 /*
2083 * Keep an observation window of at most 200 ms in the past
2084 * from now.
2085 */
2086 if (tot_io_time > 200 * NSEC_PER_MSEC) {
2087 bfqq->io_start_time = now_ns - (tot_io_time>>1);
2088 bfqq->tot_idle_time >>= 1;
2089 }
2090}
2091
2092/*
2093 * Detect whether bfqq's I/O seems synchronized with that of some
2094 * other queue, i.e., whether bfqq, after remaining empty, happens to
2095 * receive new I/O only right after some I/O request of the other
2096 * queue has been completed. We call waker queue the other queue, and
2097 * we assume, for simplicity, that bfqq may have at most one waker
2098 * queue.
2099 *
2100 * A remarkable throughput boost can be reached by unconditionally
2101 * injecting the I/O of the waker queue, every time a new
2102 * bfq_dispatch_request happens to be invoked while I/O is being
2103 * plugged for bfqq. In addition to boosting throughput, this
2104 * unblocks bfqq's I/O, thereby improving bandwidth and latency for
2105 * bfqq. Note that these same results may be achieved with the general
2106 * injection mechanism, but less effectively. For details on this
2107 * aspect, see the comments on the choice of the queue for injection
2108 * in bfq_select_queue().
2109 *
2110 * Turning back to the detection of a waker queue, a queue Q is deemed as a
2111 * waker queue for bfqq if, for three consecutive times, bfqq happens to become
2112 * non empty right after a request of Q has been completed within given
2113 * timeout. In this respect, even if bfqq is empty, we do not check for a waker
2114 * if it still has some in-flight I/O. In fact, in this case bfqq is actually
2115 * still being served by the drive, and may receive new I/O on the completion
2116 * of some of the in-flight requests. In particular, on the first time, Q is
2117 * tentatively set as a candidate waker queue, while on the third consecutive
2118 * time that Q is detected, the field waker_bfqq is set to Q, to confirm that Q
2119 * is a waker queue for bfqq. These detection steps are performed only if bfqq
2120 * has a long think time, so as to make it more likely that bfqq's I/O is
2121 * actually being blocked by a synchronization. This last filter, plus the
2122 * above three-times requirement and time limit for detection, make false
2123 * positives less likely.
2124 *
2125 * NOTE
2126 *
2127 * The sooner a waker queue is detected, the sooner throughput can be
2128 * boosted by injecting I/O from the waker queue. Fortunately,
2129 * detection is likely to be actually fast, for the following
2130 * reasons. While blocked by synchronization, bfqq has a long think
2131 * time. This implies that bfqq's inject limit is at least equal to 1
2132 * (see the comments in bfq_update_inject_limit()). So, thanks to
2133 * injection, the waker queue is likely to be served during the very
2134 * first I/O-plugging time interval for bfqq. This triggers the first
2135 * step of the detection mechanism. Thanks again to injection, the
2136 * candidate waker queue is then likely to be confirmed no later than
2137 * during the next I/O-plugging interval for bfqq.
2138 *
2139 * ISSUE
2140 *
2141 * On queue merging all waker information is lost.
2142 */
2143static void bfq_check_waker(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2144 u64 now_ns)
2145{
2146 char waker_name[MAX_BFQQ_NAME_LENGTH];
2147
2148 if (!bfqd->last_completed_rq_bfqq ||
2149 bfqd->last_completed_rq_bfqq == bfqq ||
2150 bfq_bfqq_has_short_ttime(bfqq) ||
2151 now_ns - bfqd->last_completion >= 4 * NSEC_PER_MSEC ||
2152 bfqd->last_completed_rq_bfqq == &bfqd->oom_bfqq ||
2153 bfqq == &bfqd->oom_bfqq)
2154 return;
2155
2156 /*
2157 * We reset waker detection logic also if too much time has passed
2158 * since the first detection. If wakeups are rare, pointless idling
2159 * doesn't hurt throughput that much. The condition below makes sure
2160 * we do not uselessly idle blocking waker in more than 1/64 cases.
2161 */
2162 if (bfqd->last_completed_rq_bfqq !=
2163 bfqq->tentative_waker_bfqq ||
2164 now_ns > bfqq->waker_detection_started +
2165 128 * (u64)bfqd->bfq_slice_idle) {
2166 /*
2167 * First synchronization detected with a
2168 * candidate waker queue, or with a different
2169 * candidate waker queue from the current one.
2170 */
2171 bfqq->tentative_waker_bfqq =
2172 bfqd->last_completed_rq_bfqq;
2173 bfqq->num_waker_detections = 1;
2174 bfqq->waker_detection_started = now_ns;
2175 bfq_bfqq_name(bfqq->tentative_waker_bfqq, waker_name,
2176 MAX_BFQQ_NAME_LENGTH);
2177 bfq_log_bfqq(bfqd, bfqq, "set tentative waker %s", waker_name);
2178 } else /* Same tentative waker queue detected again */
2179 bfqq->num_waker_detections++;
2180
2181 if (bfqq->num_waker_detections == 3) {
2182 bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
2183 bfqq->tentative_waker_bfqq = NULL;
2184 bfq_bfqq_name(bfqq->waker_bfqq, waker_name,
2185 MAX_BFQQ_NAME_LENGTH);
2186 bfq_log_bfqq(bfqd, bfqq, "set waker %s", waker_name);
2187
2188 /*
2189 * If the waker queue disappears, then
2190 * bfqq->waker_bfqq must be reset. To
2191 * this goal, we maintain in each
2192 * waker queue a list, woken_list, of
2193 * all the queues that reference the
2194 * waker queue through their
2195 * waker_bfqq pointer. When the waker
2196 * queue exits, the waker_bfqq pointer
2197 * of all the queues in the woken_list
2198 * is reset.
2199 *
2200 * In addition, if bfqq is already in
2201 * the woken_list of a waker queue,
2202 * then, before being inserted into
2203 * the woken_list of a new waker
2204 * queue, bfqq must be removed from
2205 * the woken_list of the old waker
2206 * queue.
2207 */
2208 if (!hlist_unhashed(&bfqq->woken_list_node))
2209 hlist_del_init(&bfqq->woken_list_node);
2210 hlist_add_head(&bfqq->woken_list_node,
2211 &bfqd->last_completed_rq_bfqq->woken_list);
2212 }
2213}
2214
2215static void bfq_add_request(struct request *rq)
2216{
2217 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2218 struct bfq_data *bfqd = bfqq->bfqd;
2219 struct request *next_rq, *prev;
2220 unsigned int old_wr_coeff = bfqq->wr_coeff;
2221 bool interactive = false;
2222 u64 now_ns = blk_time_get_ns();
2223
2224 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
2225 bfqq->queued[rq_is_sync(rq)]++;
2226 /*
2227 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2228 * may be read without holding the lock in bfq_has_work().
2229 */
2230 WRITE_ONCE(bfqd->queued, bfqd->queued + 1);
2231
2232 if (bfq_bfqq_sync(bfqq) && RQ_BIC(rq)->requests <= 1) {
2233 bfq_check_waker(bfqd, bfqq, now_ns);
2234
2235 /*
2236 * Periodically reset inject limit, to make sure that
2237 * the latter eventually drops in case workload
2238 * changes, see step (3) in the comments on
2239 * bfq_update_inject_limit().
2240 */
2241 if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2242 msecs_to_jiffies(1000)))
2243 bfq_reset_inject_limit(bfqd, bfqq);
2244
2245 /*
2246 * The following conditions must hold to setup a new
2247 * sampling of total service time, and then a new
2248 * update of the inject limit:
2249 * - bfqq is in service, because the total service
2250 * time is evaluated only for the I/O requests of
2251 * the queues in service;
2252 * - this is the right occasion to compute or to
2253 * lower the baseline total service time, because
2254 * there are actually no requests in the drive,
2255 * or
2256 * the baseline total service time is available, and
2257 * this is the right occasion to compute the other
2258 * quantity needed to update the inject limit, i.e.,
2259 * the total service time caused by the amount of
2260 * injection allowed by the current value of the
2261 * limit. It is the right occasion because injection
2262 * has actually been performed during the service
2263 * hole, and there are still in-flight requests,
2264 * which are very likely to be exactly the injected
2265 * requests, or part of them;
2266 * - the minimum interval for sampling the total
2267 * service time and updating the inject limit has
2268 * elapsed.
2269 */
2270 if (bfqq == bfqd->in_service_queue &&
2271 (bfqd->tot_rq_in_driver == 0 ||
2272 (bfqq->last_serv_time_ns > 0 &&
2273 bfqd->rqs_injected && bfqd->tot_rq_in_driver > 0)) &&
2274 time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2275 msecs_to_jiffies(10))) {
2276 bfqd->last_empty_occupied_ns = blk_time_get_ns();
2277 /*
2278 * Start the state machine for measuring the
2279 * total service time of rq: setting
2280 * wait_dispatch will cause bfqd->waited_rq to
2281 * be set when rq will be dispatched.
2282 */
2283 bfqd->wait_dispatch = true;
2284 /*
2285 * If there is no I/O in service in the drive,
2286 * then possible injection occurred before the
2287 * arrival of rq will not affect the total
2288 * service time of rq. So the injection limit
2289 * must not be updated as a function of such
2290 * total service time, unless new injection
2291 * occurs before rq is completed. To have the
2292 * injection limit updated only in the latter
2293 * case, reset rqs_injected here (rqs_injected
2294 * will be set in case injection is performed
2295 * on bfqq before rq is completed).
2296 */
2297 if (bfqd->tot_rq_in_driver == 0)
2298 bfqd->rqs_injected = false;
2299 }
2300 }
2301
2302 if (bfq_bfqq_sync(bfqq))
2303 bfq_update_io_intensity(bfqq, now_ns);
2304
2305 elv_rb_add(&bfqq->sort_list, rq);
2306
2307 /*
2308 * Check if this request is a better next-serve candidate.
2309 */
2310 prev = bfqq->next_rq;
2311 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
2312 bfqq->next_rq = next_rq;
2313
2314 /*
2315 * Adjust priority tree position, if next_rq changes.
2316 * See comments on bfq_pos_tree_add_move() for the unlikely().
2317 */
2318 if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
2319 bfq_pos_tree_add_move(bfqd, bfqq);
2320
2321 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
2322 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
2323 rq, &interactive);
2324 else {
2325 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
2326 time_is_before_jiffies(
2327 bfqq->last_wr_start_finish +
2328 bfqd->bfq_wr_min_inter_arr_async)) {
2329 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
2330 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
2331
2332 bfqd->wr_busy_queues++;
2333 bfqq->entity.prio_changed = 1;
2334 }
2335 if (prev != bfqq->next_rq)
2336 bfq_updated_next_req(bfqd, bfqq);
2337 }
2338
2339 /*
2340 * Assign jiffies to last_wr_start_finish in the following
2341 * cases:
2342 *
2343 * . if bfqq is not going to be weight-raised, because, for
2344 * non weight-raised queues, last_wr_start_finish stores the
2345 * arrival time of the last request; as of now, this piece
2346 * of information is used only for deciding whether to
2347 * weight-raise async queues
2348 *
2349 * . if bfqq is not weight-raised, because, if bfqq is now
2350 * switching to weight-raised, then last_wr_start_finish
2351 * stores the time when weight-raising starts
2352 *
2353 * . if bfqq is interactive, because, regardless of whether
2354 * bfqq is currently weight-raised, the weight-raising
2355 * period must start or restart (this case is considered
2356 * separately because it is not detected by the above
2357 * conditions, if bfqq is already weight-raised)
2358 *
2359 * last_wr_start_finish has to be updated also if bfqq is soft
2360 * real-time, because the weight-raising period is constantly
2361 * restarted on idle-to-busy transitions for these queues, but
2362 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2363 * needed.
2364 */
2365 if (bfqd->low_latency &&
2366 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
2367 bfqq->last_wr_start_finish = jiffies;
2368}
2369
2370static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
2371 struct bio *bio,
2372 struct request_queue *q)
2373{
2374 struct bfq_queue *bfqq = bfqd->bio_bfqq;
2375
2376
2377 if (bfqq)
2378 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
2379
2380 return NULL;
2381}
2382
2383static sector_t get_sdist(sector_t last_pos, struct request *rq)
2384{
2385 if (last_pos)
2386 return abs(blk_rq_pos(rq) - last_pos);
2387
2388 return 0;
2389}
2390
2391static void bfq_remove_request(struct request_queue *q,
2392 struct request *rq)
2393{
2394 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2395 struct bfq_data *bfqd = bfqq->bfqd;
2396 const int sync = rq_is_sync(rq);
2397
2398 if (bfqq->next_rq == rq) {
2399 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
2400 bfq_updated_next_req(bfqd, bfqq);
2401 }
2402
2403 if (rq->queuelist.prev != &rq->queuelist)
2404 list_del_init(&rq->queuelist);
2405 bfqq->queued[sync]--;
2406 /*
2407 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2408 * may be read without holding the lock in bfq_has_work().
2409 */
2410 WRITE_ONCE(bfqd->queued, bfqd->queued - 1);
2411 elv_rb_del(&bfqq->sort_list, rq);
2412
2413 elv_rqhash_del(q, rq);
2414 if (q->last_merge == rq)
2415 q->last_merge = NULL;
2416
2417 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2418 bfqq->next_rq = NULL;
2419
2420 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
2421 bfq_del_bfqq_busy(bfqq, false);
2422 /*
2423 * bfqq emptied. In normal operation, when
2424 * bfqq is empty, bfqq->entity.service and
2425 * bfqq->entity.budget must contain,
2426 * respectively, the service received and the
2427 * budget used last time bfqq emptied. These
2428 * facts do not hold in this case, as at least
2429 * this last removal occurred while bfqq is
2430 * not in service. To avoid inconsistencies,
2431 * reset both bfqq->entity.service and
2432 * bfqq->entity.budget, if bfqq has still a
2433 * process that may issue I/O requests to it.
2434 */
2435 bfqq->entity.budget = bfqq->entity.service = 0;
2436 }
2437
2438 /*
2439 * Remove queue from request-position tree as it is empty.
2440 */
2441 if (bfqq->pos_root) {
2442 rb_erase(&bfqq->pos_node, bfqq->pos_root);
2443 bfqq->pos_root = NULL;
2444 }
2445 } else {
2446 /* see comments on bfq_pos_tree_add_move() for the unlikely() */
2447 if (unlikely(!bfqd->nonrot_with_queueing))
2448 bfq_pos_tree_add_move(bfqd, bfqq);
2449 }
2450
2451 if (rq->cmd_flags & REQ_META)
2452 bfqq->meta_pending--;
2453
2454}
2455
2456static bool bfq_bio_merge(struct request_queue *q, struct bio *bio,
2457 unsigned int nr_segs)
2458{
2459 struct bfq_data *bfqd = q->elevator->elevator_data;
2460 struct request *free = NULL;
2461 /*
2462 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2463 * store its return value for later use, to avoid nesting
2464 * queue_lock inside the bfqd->lock. We assume that the bic
2465 * returned by bfq_bic_lookup does not go away before
2466 * bfqd->lock is taken.
2467 */
2468 struct bfq_io_cq *bic = bfq_bic_lookup(q);
2469 bool ret;
2470
2471 spin_lock_irq(&bfqd->lock);
2472
2473 if (bic) {
2474 /*
2475 * Make sure cgroup info is uptodate for current process before
2476 * considering the merge.
2477 */
2478 bfq_bic_update_cgroup(bic, bio);
2479
2480 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf),
2481 bfq_actuator_index(bfqd, bio));
2482 } else {
2483 bfqd->bio_bfqq = NULL;
2484 }
2485 bfqd->bio_bic = bic;
2486
2487 ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
2488
2489 spin_unlock_irq(&bfqd->lock);
2490 if (free)
2491 blk_mq_free_request(free);
2492
2493 return ret;
2494}
2495
2496static int bfq_request_merge(struct request_queue *q, struct request **req,
2497 struct bio *bio)
2498{
2499 struct bfq_data *bfqd = q->elevator->elevator_data;
2500 struct request *__rq;
2501
2502 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
2503 if (__rq && elv_bio_merge_ok(__rq, bio)) {
2504 *req = __rq;
2505
2506 if (blk_discard_mergable(__rq))
2507 return ELEVATOR_DISCARD_MERGE;
2508 return ELEVATOR_FRONT_MERGE;
2509 }
2510
2511 return ELEVATOR_NO_MERGE;
2512}
2513
2514static void bfq_request_merged(struct request_queue *q, struct request *req,
2515 enum elv_merge type)
2516{
2517 if (type == ELEVATOR_FRONT_MERGE &&
2518 rb_prev(&req->rb_node) &&
2519 blk_rq_pos(req) <
2520 blk_rq_pos(container_of(rb_prev(&req->rb_node),
2521 struct request, rb_node))) {
2522 struct bfq_queue *bfqq = RQ_BFQQ(req);
2523 struct bfq_data *bfqd;
2524 struct request *prev, *next_rq;
2525
2526 if (!bfqq)
2527 return;
2528
2529 bfqd = bfqq->bfqd;
2530
2531 /* Reposition request in its sort_list */
2532 elv_rb_del(&bfqq->sort_list, req);
2533 elv_rb_add(&bfqq->sort_list, req);
2534
2535 /* Choose next request to be served for bfqq */
2536 prev = bfqq->next_rq;
2537 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
2538 bfqd->last_position);
2539 bfqq->next_rq = next_rq;
2540 /*
2541 * If next_rq changes, update both the queue's budget to
2542 * fit the new request and the queue's position in its
2543 * rq_pos_tree.
2544 */
2545 if (prev != bfqq->next_rq) {
2546 bfq_updated_next_req(bfqd, bfqq);
2547 /*
2548 * See comments on bfq_pos_tree_add_move() for
2549 * the unlikely().
2550 */
2551 if (unlikely(!bfqd->nonrot_with_queueing))
2552 bfq_pos_tree_add_move(bfqd, bfqq);
2553 }
2554 }
2555}
2556
2557/*
2558 * This function is called to notify the scheduler that the requests
2559 * rq and 'next' have been merged, with 'next' going away. BFQ
2560 * exploits this hook to address the following issue: if 'next' has a
2561 * fifo_time lower that rq, then the fifo_time of rq must be set to
2562 * the value of 'next', to not forget the greater age of 'next'.
2563 *
2564 * NOTE: in this function we assume that rq is in a bfq_queue, basing
2565 * on that rq is picked from the hash table q->elevator->hash, which,
2566 * in its turn, is filled only with I/O requests present in
2567 * bfq_queues, while BFQ is in use for the request queue q. In fact,
2568 * the function that fills this hash table (elv_rqhash_add) is called
2569 * only by bfq_insert_request.
2570 */
2571static void bfq_requests_merged(struct request_queue *q, struct request *rq,
2572 struct request *next)
2573{
2574 struct bfq_queue *bfqq = RQ_BFQQ(rq),
2575 *next_bfqq = RQ_BFQQ(next);
2576
2577 if (!bfqq)
2578 goto remove;
2579
2580 /*
2581 * If next and rq belong to the same bfq_queue and next is older
2582 * than rq, then reposition rq in the fifo (by substituting next
2583 * with rq). Otherwise, if next and rq belong to different
2584 * bfq_queues, never reposition rq: in fact, we would have to
2585 * reposition it with respect to next's position in its own fifo,
2586 * which would most certainly be too expensive with respect to
2587 * the benefits.
2588 */
2589 if (bfqq == next_bfqq &&
2590 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2591 next->fifo_time < rq->fifo_time) {
2592 list_del_init(&rq->queuelist);
2593 list_replace_init(&next->queuelist, &rq->queuelist);
2594 rq->fifo_time = next->fifo_time;
2595 }
2596
2597 if (bfqq->next_rq == next)
2598 bfqq->next_rq = rq;
2599
2600 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
2601remove:
2602 /* Merged request may be in the IO scheduler. Remove it. */
2603 if (!RB_EMPTY_NODE(&next->rb_node)) {
2604 bfq_remove_request(next->q, next);
2605 if (next_bfqq)
2606 bfqg_stats_update_io_remove(bfqq_group(next_bfqq),
2607 next->cmd_flags);
2608 }
2609}
2610
2611/* Must be called with bfqq != NULL */
2612static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
2613{
2614 /*
2615 * If bfqq has been enjoying interactive weight-raising, then
2616 * reset soft_rt_next_start. We do it for the following
2617 * reason. bfqq may have been conveying the I/O needed to load
2618 * a soft real-time application. Such an application actually
2619 * exhibits a soft real-time I/O pattern after it finishes
2620 * loading, and finally starts doing its job. But, if bfqq has
2621 * been receiving a lot of bandwidth so far (likely to happen
2622 * on a fast device), then soft_rt_next_start now contains a
2623 * high value that. So, without this reset, bfqq would be
2624 * prevented from being possibly considered as soft_rt for a
2625 * very long time.
2626 */
2627
2628 if (bfqq->wr_cur_max_time !=
2629 bfqq->bfqd->bfq_wr_rt_max_time)
2630 bfqq->soft_rt_next_start = jiffies;
2631
2632 if (bfq_bfqq_busy(bfqq))
2633 bfqq->bfqd->wr_busy_queues--;
2634 bfqq->wr_coeff = 1;
2635 bfqq->wr_cur_max_time = 0;
2636 bfqq->last_wr_start_finish = jiffies;
2637 /*
2638 * Trigger a weight change on the next invocation of
2639 * __bfq_entity_update_weight_prio.
2640 */
2641 bfqq->entity.prio_changed = 1;
2642}
2643
2644void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2645 struct bfq_group *bfqg)
2646{
2647 int i, j, k;
2648
2649 for (k = 0; k < bfqd->num_actuators; k++) {
2650 for (i = 0; i < 2; i++)
2651 for (j = 0; j < IOPRIO_NR_LEVELS; j++)
2652 if (bfqg->async_bfqq[i][j][k])
2653 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j][k]);
2654 if (bfqg->async_idle_bfqq[k])
2655 bfq_bfqq_end_wr(bfqg->async_idle_bfqq[k]);
2656 }
2657}
2658
2659static void bfq_end_wr(struct bfq_data *bfqd)
2660{
2661 struct bfq_queue *bfqq;
2662 int i;
2663
2664 spin_lock_irq(&bfqd->lock);
2665
2666 for (i = 0; i < bfqd->num_actuators; i++) {
2667 list_for_each_entry(bfqq, &bfqd->active_list[i], bfqq_list)
2668 bfq_bfqq_end_wr(bfqq);
2669 }
2670 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2671 bfq_bfqq_end_wr(bfqq);
2672 bfq_end_wr_async(bfqd);
2673
2674 spin_unlock_irq(&bfqd->lock);
2675}
2676
2677static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2678{
2679 if (request)
2680 return blk_rq_pos(io_struct);
2681 else
2682 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2683}
2684
2685static int bfq_rq_close_to_sector(void *io_struct, bool request,
2686 sector_t sector)
2687{
2688 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2689 BFQQ_CLOSE_THR;
2690}
2691
2692static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2693 struct bfq_queue *bfqq,
2694 sector_t sector)
2695{
2696 struct rb_root *root = &bfqq_group(bfqq)->rq_pos_tree;
2697 struct rb_node *parent, *node;
2698 struct bfq_queue *__bfqq;
2699
2700 if (RB_EMPTY_ROOT(root))
2701 return NULL;
2702
2703 /*
2704 * First, if we find a request starting at the end of the last
2705 * request, choose it.
2706 */
2707 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2708 if (__bfqq)
2709 return __bfqq;
2710
2711 /*
2712 * If the exact sector wasn't found, the parent of the NULL leaf
2713 * will contain the closest sector (rq_pos_tree sorted by
2714 * next_request position).
2715 */
2716 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2717 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2718 return __bfqq;
2719
2720 if (blk_rq_pos(__bfqq->next_rq) < sector)
2721 node = rb_next(&__bfqq->pos_node);
2722 else
2723 node = rb_prev(&__bfqq->pos_node);
2724 if (!node)
2725 return NULL;
2726
2727 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2728 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2729 return __bfqq;
2730
2731 return NULL;
2732}
2733
2734static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2735 struct bfq_queue *cur_bfqq,
2736 sector_t sector)
2737{
2738 struct bfq_queue *bfqq;
2739
2740 /*
2741 * We shall notice if some of the queues are cooperating,
2742 * e.g., working closely on the same area of the device. In
2743 * that case, we can group them together and: 1) don't waste
2744 * time idling, and 2) serve the union of their requests in
2745 * the best possible order for throughput.
2746 */
2747 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2748 if (!bfqq || bfqq == cur_bfqq)
2749 return NULL;
2750
2751 return bfqq;
2752}
2753
2754static struct bfq_queue *
2755bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2756{
2757 int process_refs, new_process_refs;
2758 struct bfq_queue *__bfqq;
2759
2760 /*
2761 * If there are no process references on the new_bfqq, then it is
2762 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2763 * may have dropped their last reference (not just their last process
2764 * reference).
2765 */
2766 if (!bfqq_process_refs(new_bfqq))
2767 return NULL;
2768
2769 /* Avoid a circular list and skip interim queue merges. */
2770 while ((__bfqq = new_bfqq->new_bfqq)) {
2771 if (__bfqq == bfqq)
2772 return NULL;
2773 new_bfqq = __bfqq;
2774 }
2775
2776 process_refs = bfqq_process_refs(bfqq);
2777 new_process_refs = bfqq_process_refs(new_bfqq);
2778 /*
2779 * If the process for the bfqq has gone away, there is no
2780 * sense in merging the queues.
2781 */
2782 if (process_refs == 0 || new_process_refs == 0)
2783 return NULL;
2784
2785 /*
2786 * Make sure merged queues belong to the same parent. Parents could
2787 * have changed since the time we decided the two queues are suitable
2788 * for merging.
2789 */
2790 if (new_bfqq->entity.parent != bfqq->entity.parent)
2791 return NULL;
2792
2793 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2794 new_bfqq->pid);
2795
2796 /*
2797 * Merging is just a redirection: the requests of the process
2798 * owning one of the two queues are redirected to the other queue.
2799 * The latter queue, in its turn, is set as shared if this is the
2800 * first time that the requests of some process are redirected to
2801 * it.
2802 *
2803 * We redirect bfqq to new_bfqq and not the opposite, because
2804 * we are in the context of the process owning bfqq, thus we
2805 * have the io_cq of this process. So we can immediately
2806 * configure this io_cq to redirect the requests of the
2807 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2808 * not available any more (new_bfqq->bic == NULL).
2809 *
2810 * Anyway, even in case new_bfqq coincides with the in-service
2811 * queue, redirecting requests the in-service queue is the
2812 * best option, as we feed the in-service queue with new
2813 * requests close to the last request served and, by doing so,
2814 * are likely to increase the throughput.
2815 */
2816 bfqq->new_bfqq = new_bfqq;
2817 /*
2818 * The above assignment schedules the following redirections:
2819 * each time some I/O for bfqq arrives, the process that
2820 * generated that I/O is disassociated from bfqq and
2821 * associated with new_bfqq. Here we increases new_bfqq->ref
2822 * in advance, adding the number of processes that are
2823 * expected to be associated with new_bfqq as they happen to
2824 * issue I/O.
2825 */
2826 new_bfqq->ref += process_refs;
2827 return new_bfqq;
2828}
2829
2830static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2831 struct bfq_queue *new_bfqq)
2832{
2833 if (bfq_too_late_for_merging(new_bfqq))
2834 return false;
2835
2836 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2837 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2838 return false;
2839
2840 /*
2841 * If either of the queues has already been detected as seeky,
2842 * then merging it with the other queue is unlikely to lead to
2843 * sequential I/O.
2844 */
2845 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2846 return false;
2847
2848 /*
2849 * Interleaved I/O is known to be done by (some) applications
2850 * only for reads, so it does not make sense to merge async
2851 * queues.
2852 */
2853 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2854 return false;
2855
2856 return true;
2857}
2858
2859static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
2860 struct bfq_queue *bfqq);
2861
2862static struct bfq_queue *
2863bfq_setup_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2864 struct bfq_queue *stable_merge_bfqq,
2865 struct bfq_iocq_bfqq_data *bfqq_data)
2866{
2867 int proc_ref = min(bfqq_process_refs(bfqq),
2868 bfqq_process_refs(stable_merge_bfqq));
2869 struct bfq_queue *new_bfqq = NULL;
2870
2871 bfqq_data->stable_merge_bfqq = NULL;
2872 if (idling_boosts_thr_without_issues(bfqd, bfqq) || proc_ref == 0)
2873 goto out;
2874
2875 /* next function will take at least one ref */
2876 new_bfqq = bfq_setup_merge(bfqq, stable_merge_bfqq);
2877
2878 if (new_bfqq) {
2879 bfqq_data->stably_merged = true;
2880 if (new_bfqq->bic) {
2881 unsigned int new_a_idx = new_bfqq->actuator_idx;
2882 struct bfq_iocq_bfqq_data *new_bfqq_data =
2883 &new_bfqq->bic->bfqq_data[new_a_idx];
2884
2885 new_bfqq_data->stably_merged = true;
2886 }
2887 }
2888
2889out:
2890 /* deschedule stable merge, because done or aborted here */
2891 bfq_put_stable_ref(stable_merge_bfqq);
2892
2893 return new_bfqq;
2894}
2895
2896/*
2897 * Attempt to schedule a merge of bfqq with the currently in-service
2898 * queue or with a close queue among the scheduled queues. Return
2899 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2900 * structure otherwise.
2901 *
2902 * The OOM queue is not allowed to participate to cooperation: in fact, since
2903 * the requests temporarily redirected to the OOM queue could be redirected
2904 * again to dedicated queues at any time, the state needed to correctly
2905 * handle merging with the OOM queue would be quite complex and expensive
2906 * to maintain. Besides, in such a critical condition as an out of memory,
2907 * the benefits of queue merging may be little relevant, or even negligible.
2908 *
2909 * WARNING: queue merging may impair fairness among non-weight raised
2910 * queues, for at least two reasons: 1) the original weight of a
2911 * merged queue may change during the merged state, 2) even being the
2912 * weight the same, a merged queue may be bloated with many more
2913 * requests than the ones produced by its originally-associated
2914 * process.
2915 */
2916static struct bfq_queue *
2917bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2918 void *io_struct, bool request, struct bfq_io_cq *bic)
2919{
2920 struct bfq_queue *in_service_bfqq, *new_bfqq;
2921 unsigned int a_idx = bfqq->actuator_idx;
2922 struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[a_idx];
2923
2924 /* if a merge has already been setup, then proceed with that first */
2925 new_bfqq = bfqq->new_bfqq;
2926 if (new_bfqq) {
2927 while (new_bfqq->new_bfqq)
2928 new_bfqq = new_bfqq->new_bfqq;
2929 return new_bfqq;
2930 }
2931
2932 /*
2933 * Check delayed stable merge for rotational or non-queueing
2934 * devs. For this branch to be executed, bfqq must not be
2935 * currently merged with some other queue (i.e., bfqq->bic
2936 * must be non null). If we considered also merged queues,
2937 * then we should also check whether bfqq has already been
2938 * merged with bic->stable_merge_bfqq. But this would be
2939 * costly and complicated.
2940 */
2941 if (unlikely(!bfqd->nonrot_with_queueing)) {
2942 /*
2943 * Make sure also that bfqq is sync, because
2944 * bic->stable_merge_bfqq may point to some queue (for
2945 * stable merging) also if bic is associated with a
2946 * sync queue, but this bfqq is async
2947 */
2948 if (bfq_bfqq_sync(bfqq) && bfqq_data->stable_merge_bfqq &&
2949 !bfq_bfqq_just_created(bfqq) &&
2950 time_is_before_jiffies(bfqq->split_time +
2951 msecs_to_jiffies(bfq_late_stable_merging)) &&
2952 time_is_before_jiffies(bfqq->creation_time +
2953 msecs_to_jiffies(bfq_late_stable_merging))) {
2954 struct bfq_queue *stable_merge_bfqq =
2955 bfqq_data->stable_merge_bfqq;
2956
2957 return bfq_setup_stable_merge(bfqd, bfqq,
2958 stable_merge_bfqq,
2959 bfqq_data);
2960 }
2961 }
2962
2963 /*
2964 * Do not perform queue merging if the device is non
2965 * rotational and performs internal queueing. In fact, such a
2966 * device reaches a high speed through internal parallelism
2967 * and pipelining. This means that, to reach a high
2968 * throughput, it must have many requests enqueued at the same
2969 * time. But, in this configuration, the internal scheduling
2970 * algorithm of the device does exactly the job of queue
2971 * merging: it reorders requests so as to obtain as much as
2972 * possible a sequential I/O pattern. As a consequence, with
2973 * the workload generated by processes doing interleaved I/O,
2974 * the throughput reached by the device is likely to be the
2975 * same, with and without queue merging.
2976 *
2977 * Disabling merging also provides a remarkable benefit in
2978 * terms of throughput. Merging tends to make many workloads
2979 * artificially more uneven, because of shared queues
2980 * remaining non empty for incomparably more time than
2981 * non-merged queues. This may accentuate workload
2982 * asymmetries. For example, if one of the queues in a set of
2983 * merged queues has a higher weight than a normal queue, then
2984 * the shared queue may inherit such a high weight and, by
2985 * staying almost always active, may force BFQ to perform I/O
2986 * plugging most of the time. This evidently makes it harder
2987 * for BFQ to let the device reach a high throughput.
2988 *
2989 * Finally, the likely() macro below is not used because one
2990 * of the two branches is more likely than the other, but to
2991 * have the code path after the following if() executed as
2992 * fast as possible for the case of a non rotational device
2993 * with queueing. We want it because this is the fastest kind
2994 * of device. On the opposite end, the likely() may lengthen
2995 * the execution time of BFQ for the case of slower devices
2996 * (rotational or at least without queueing). But in this case
2997 * the execution time of BFQ matters very little, if not at
2998 * all.
2999 */
3000 if (likely(bfqd->nonrot_with_queueing))
3001 return NULL;
3002
3003 /*
3004 * Prevent bfqq from being merged if it has been created too
3005 * long ago. The idea is that true cooperating processes, and
3006 * thus their associated bfq_queues, are supposed to be
3007 * created shortly after each other. This is the case, e.g.,
3008 * for KVM/QEMU and dump I/O threads. Basing on this
3009 * assumption, the following filtering greatly reduces the
3010 * probability that two non-cooperating processes, which just
3011 * happen to do close I/O for some short time interval, have
3012 * their queues merged by mistake.
3013 */
3014 if (bfq_too_late_for_merging(bfqq))
3015 return NULL;
3016
3017 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
3018 return NULL;
3019
3020 /* If there is only one backlogged queue, don't search. */
3021 if (bfq_tot_busy_queues(bfqd) == 1)
3022 return NULL;
3023
3024 in_service_bfqq = bfqd->in_service_queue;
3025
3026 if (in_service_bfqq && in_service_bfqq != bfqq &&
3027 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
3028 bfq_rq_close_to_sector(io_struct, request,
3029 bfqd->in_serv_last_pos) &&
3030 bfqq->entity.parent == in_service_bfqq->entity.parent &&
3031 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
3032 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
3033 if (new_bfqq)
3034 return new_bfqq;
3035 }
3036 /*
3037 * Check whether there is a cooperator among currently scheduled
3038 * queues. The only thing we need is that the bio/request is not
3039 * NULL, as we need it to establish whether a cooperator exists.
3040 */
3041 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
3042 bfq_io_struct_pos(io_struct, request));
3043
3044 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
3045 bfq_may_be_close_cooperator(bfqq, new_bfqq))
3046 return bfq_setup_merge(bfqq, new_bfqq);
3047
3048 return NULL;
3049}
3050
3051static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
3052{
3053 struct bfq_io_cq *bic = bfqq->bic;
3054 unsigned int a_idx = bfqq->actuator_idx;
3055 struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[a_idx];
3056
3057 /*
3058 * If !bfqq->bic, the queue is already shared or its requests
3059 * have already been redirected to a shared queue; both idle window
3060 * and weight raising state have already been saved. Do nothing.
3061 */
3062 if (!bic)
3063 return;
3064
3065 bfqq_data->saved_last_serv_time_ns = bfqq->last_serv_time_ns;
3066 bfqq_data->saved_inject_limit = bfqq->inject_limit;
3067 bfqq_data->saved_decrease_time_jif = bfqq->decrease_time_jif;
3068
3069 bfqq_data->saved_weight = bfqq->entity.orig_weight;
3070 bfqq_data->saved_ttime = bfqq->ttime;
3071 bfqq_data->saved_has_short_ttime =
3072 bfq_bfqq_has_short_ttime(bfqq);
3073 bfqq_data->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
3074 bfqq_data->saved_io_start_time = bfqq->io_start_time;
3075 bfqq_data->saved_tot_idle_time = bfqq->tot_idle_time;
3076 bfqq_data->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
3077 bfqq_data->was_in_burst_list =
3078 !hlist_unhashed(&bfqq->burst_list_node);
3079
3080 if (unlikely(bfq_bfqq_just_created(bfqq) &&
3081 !bfq_bfqq_in_large_burst(bfqq) &&
3082 bfqq->bfqd->low_latency)) {
3083 /*
3084 * bfqq being merged right after being created: bfqq
3085 * would have deserved interactive weight raising, but
3086 * did not make it to be set in a weight-raised state,
3087 * because of this early merge. Store directly the
3088 * weight-raising state that would have been assigned
3089 * to bfqq, so that to avoid that bfqq unjustly fails
3090 * to enjoy weight raising if split soon.
3091 */
3092 bfqq_data->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
3093 bfqq_data->saved_wr_start_at_switch_to_srt =
3094 bfq_smallest_from_now();
3095 bfqq_data->saved_wr_cur_max_time =
3096 bfq_wr_duration(bfqq->bfqd);
3097 bfqq_data->saved_last_wr_start_finish = jiffies;
3098 } else {
3099 bfqq_data->saved_wr_coeff = bfqq->wr_coeff;
3100 bfqq_data->saved_wr_start_at_switch_to_srt =
3101 bfqq->wr_start_at_switch_to_srt;
3102 bfqq_data->saved_service_from_wr =
3103 bfqq->service_from_wr;
3104 bfqq_data->saved_last_wr_start_finish =
3105 bfqq->last_wr_start_finish;
3106 bfqq_data->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
3107 }
3108}
3109
3110
3111void bfq_reassign_last_bfqq(struct bfq_queue *cur_bfqq,
3112 struct bfq_queue *new_bfqq)
3113{
3114 if (cur_bfqq->entity.parent &&
3115 cur_bfqq->entity.parent->last_bfqq_created == cur_bfqq)
3116 cur_bfqq->entity.parent->last_bfqq_created = new_bfqq;
3117 else if (cur_bfqq->bfqd && cur_bfqq->bfqd->last_bfqq_created == cur_bfqq)
3118 cur_bfqq->bfqd->last_bfqq_created = new_bfqq;
3119}
3120
3121void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3122{
3123 /*
3124 * To prevent bfqq's service guarantees from being violated,
3125 * bfqq may be left busy, i.e., queued for service, even if
3126 * empty (see comments in __bfq_bfqq_expire() for
3127 * details). But, if no process will send requests to bfqq any
3128 * longer, then there is no point in keeping bfqq queued for
3129 * service. In addition, keeping bfqq queued for service, but
3130 * with no process ref any longer, may have caused bfqq to be
3131 * freed when dequeued from service. But this is assumed to
3132 * never happen.
3133 */
3134 if (bfq_bfqq_busy(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) &&
3135 bfqq != bfqd->in_service_queue)
3136 bfq_del_bfqq_busy(bfqq, false);
3137
3138 bfq_reassign_last_bfqq(bfqq, NULL);
3139
3140 bfq_put_queue(bfqq);
3141}
3142
3143static struct bfq_queue *bfq_merge_bfqqs(struct bfq_data *bfqd,
3144 struct bfq_io_cq *bic,
3145 struct bfq_queue *bfqq)
3146{
3147 struct bfq_queue *new_bfqq = bfqq->new_bfqq;
3148
3149 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
3150 (unsigned long)new_bfqq->pid);
3151 /* Save weight raising and idle window of the merged queues */
3152 bfq_bfqq_save_state(bfqq);
3153 bfq_bfqq_save_state(new_bfqq);
3154 if (bfq_bfqq_IO_bound(bfqq))
3155 bfq_mark_bfqq_IO_bound(new_bfqq);
3156 bfq_clear_bfqq_IO_bound(bfqq);
3157
3158 /*
3159 * The processes associated with bfqq are cooperators of the
3160 * processes associated with new_bfqq. So, if bfqq has a
3161 * waker, then assume that all these processes will be happy
3162 * to let bfqq's waker freely inject I/O when they have no
3163 * I/O.
3164 */
3165 if (bfqq->waker_bfqq && !new_bfqq->waker_bfqq &&
3166 bfqq->waker_bfqq != new_bfqq) {
3167 new_bfqq->waker_bfqq = bfqq->waker_bfqq;
3168 new_bfqq->tentative_waker_bfqq = NULL;
3169
3170 /*
3171 * If the waker queue disappears, then
3172 * new_bfqq->waker_bfqq must be reset. So insert
3173 * new_bfqq into the woken_list of the waker. See
3174 * bfq_check_waker for details.
3175 */
3176 hlist_add_head(&new_bfqq->woken_list_node,
3177 &new_bfqq->waker_bfqq->woken_list);
3178
3179 }
3180
3181 /*
3182 * If bfqq is weight-raised, then let new_bfqq inherit
3183 * weight-raising. To reduce false positives, neglect the case
3184 * where bfqq has just been created, but has not yet made it
3185 * to be weight-raised (which may happen because EQM may merge
3186 * bfqq even before bfq_add_request is executed for the first
3187 * time for bfqq). Handling this case would however be very
3188 * easy, thanks to the flag just_created.
3189 */
3190 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
3191 new_bfqq->wr_coeff = bfqq->wr_coeff;
3192 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
3193 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
3194 new_bfqq->wr_start_at_switch_to_srt =
3195 bfqq->wr_start_at_switch_to_srt;
3196 if (bfq_bfqq_busy(new_bfqq))
3197 bfqd->wr_busy_queues++;
3198 new_bfqq->entity.prio_changed = 1;
3199 }
3200
3201 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
3202 bfqq->wr_coeff = 1;
3203 bfqq->entity.prio_changed = 1;
3204 if (bfq_bfqq_busy(bfqq))
3205 bfqd->wr_busy_queues--;
3206 }
3207
3208 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
3209 bfqd->wr_busy_queues);
3210
3211 /*
3212 * Merge queues (that is, let bic redirect its requests to new_bfqq)
3213 */
3214 bic_set_bfqq(bic, new_bfqq, true, bfqq->actuator_idx);
3215 bfq_mark_bfqq_coop(new_bfqq);
3216 /*
3217 * new_bfqq now belongs to at least two bics (it is a shared queue):
3218 * set new_bfqq->bic to NULL. bfqq either:
3219 * - does not belong to any bic any more, and hence bfqq->bic must
3220 * be set to NULL, or
3221 * - is a queue whose owning bics have already been redirected to a
3222 * different queue, hence the queue is destined to not belong to
3223 * any bic soon and bfqq->bic is already NULL (therefore the next
3224 * assignment causes no harm).
3225 */
3226 new_bfqq->bic = NULL;
3227 /*
3228 * If the queue is shared, the pid is the pid of one of the associated
3229 * processes. Which pid depends on the exact sequence of merge events
3230 * the queue underwent. So printing such a pid is useless and confusing
3231 * because it reports a random pid between those of the associated
3232 * processes.
3233 * We mark such a queue with a pid -1, and then print SHARED instead of
3234 * a pid in logging messages.
3235 */
3236 new_bfqq->pid = -1;
3237 bfqq->bic = NULL;
3238
3239 bfq_reassign_last_bfqq(bfqq, new_bfqq);
3240
3241 bfq_release_process_ref(bfqd, bfqq);
3242
3243 return new_bfqq;
3244}
3245
3246static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
3247 struct bio *bio)
3248{
3249 struct bfq_data *bfqd = q->elevator->elevator_data;
3250 bool is_sync = op_is_sync(bio->bi_opf);
3251 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
3252
3253 /*
3254 * Disallow merge of a sync bio into an async request.
3255 */
3256 if (is_sync && !rq_is_sync(rq))
3257 return false;
3258
3259 /*
3260 * Lookup the bfqq that this bio will be queued with. Allow
3261 * merge only if rq is queued there.
3262 */
3263 if (!bfqq)
3264 return false;
3265
3266 /*
3267 * We take advantage of this function to perform an early merge
3268 * of the queues of possible cooperating processes.
3269 */
3270 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false, bfqd->bio_bic);
3271 if (new_bfqq) {
3272 /*
3273 * bic still points to bfqq, then it has not yet been
3274 * redirected to some other bfq_queue, and a queue
3275 * merge between bfqq and new_bfqq can be safely
3276 * fulfilled, i.e., bic can be redirected to new_bfqq
3277 * and bfqq can be put.
3278 */
3279 while (bfqq != new_bfqq)
3280 bfqq = bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq);
3281
3282 /*
3283 * Change also bqfd->bio_bfqq, as
3284 * bfqd->bio_bic now points to new_bfqq, and
3285 * this function may be invoked again (and then may
3286 * use again bqfd->bio_bfqq).
3287 */
3288 bfqd->bio_bfqq = bfqq;
3289 }
3290
3291 return bfqq == RQ_BFQQ(rq);
3292}
3293
3294/*
3295 * Set the maximum time for the in-service queue to consume its
3296 * budget. This prevents seeky processes from lowering the throughput.
3297 * In practice, a time-slice service scheme is used with seeky
3298 * processes.
3299 */
3300static void bfq_set_budget_timeout(struct bfq_data *bfqd,
3301 struct bfq_queue *bfqq)
3302{
3303 unsigned int timeout_coeff;
3304
3305 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
3306 timeout_coeff = 1;
3307 else
3308 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
3309
3310 bfqd->last_budget_start = blk_time_get();
3311
3312 bfqq->budget_timeout = jiffies +
3313 bfqd->bfq_timeout * timeout_coeff;
3314}
3315
3316static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
3317 struct bfq_queue *bfqq)
3318{
3319 if (bfqq) {
3320 bfq_clear_bfqq_fifo_expire(bfqq);
3321
3322 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
3323
3324 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
3325 bfqq->wr_coeff > 1 &&
3326 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
3327 time_is_before_jiffies(bfqq->budget_timeout)) {
3328 /*
3329 * For soft real-time queues, move the start
3330 * of the weight-raising period forward by the
3331 * time the queue has not received any
3332 * service. Otherwise, a relatively long
3333 * service delay is likely to cause the
3334 * weight-raising period of the queue to end,
3335 * because of the short duration of the
3336 * weight-raising period of a soft real-time
3337 * queue. It is worth noting that this move
3338 * is not so dangerous for the other queues,
3339 * because soft real-time queues are not
3340 * greedy.
3341 *
3342 * To not add a further variable, we use the
3343 * overloaded field budget_timeout to
3344 * determine for how long the queue has not
3345 * received service, i.e., how much time has
3346 * elapsed since the queue expired. However,
3347 * this is a little imprecise, because
3348 * budget_timeout is set to jiffies if bfqq
3349 * not only expires, but also remains with no
3350 * request.
3351 */
3352 if (time_after(bfqq->budget_timeout,
3353 bfqq->last_wr_start_finish))
3354 bfqq->last_wr_start_finish +=
3355 jiffies - bfqq->budget_timeout;
3356 else
3357 bfqq->last_wr_start_finish = jiffies;
3358 }
3359
3360 bfq_set_budget_timeout(bfqd, bfqq);
3361 bfq_log_bfqq(bfqd, bfqq,
3362 "set_in_service_queue, cur-budget = %d",
3363 bfqq->entity.budget);
3364 }
3365
3366 bfqd->in_service_queue = bfqq;
3367 bfqd->in_serv_last_pos = 0;
3368}
3369
3370/*
3371 * Get and set a new queue for service.
3372 */
3373static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
3374{
3375 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
3376
3377 __bfq_set_in_service_queue(bfqd, bfqq);
3378 return bfqq;
3379}
3380
3381static void bfq_arm_slice_timer(struct bfq_data *bfqd)
3382{
3383 struct bfq_queue *bfqq = bfqd->in_service_queue;
3384 u32 sl;
3385
3386 bfq_mark_bfqq_wait_request(bfqq);
3387
3388 /*
3389 * We don't want to idle for seeks, but we do want to allow
3390 * fair distribution of slice time for a process doing back-to-back
3391 * seeks. So allow a little bit of time for him to submit a new rq.
3392 */
3393 sl = bfqd->bfq_slice_idle;
3394 /*
3395 * Unless the queue is being weight-raised or the scenario is
3396 * asymmetric, grant only minimum idle time if the queue
3397 * is seeky. A long idling is preserved for a weight-raised
3398 * queue, or, more in general, in an asymmetric scenario,
3399 * because a long idling is needed for guaranteeing to a queue
3400 * its reserved share of the throughput (in particular, it is
3401 * needed if the queue has a higher weight than some other
3402 * queue).
3403 */
3404 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3405 !bfq_asymmetric_scenario(bfqd, bfqq))
3406 sl = min_t(u64, sl, BFQ_MIN_TT);
3407 else if (bfqq->wr_coeff > 1)
3408 sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
3409
3410 bfqd->last_idling_start = blk_time_get();
3411 bfqd->last_idling_start_jiffies = jiffies;
3412
3413 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
3414 HRTIMER_MODE_REL);
3415 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
3416}
3417
3418/*
3419 * In autotuning mode, max_budget is dynamically recomputed as the
3420 * amount of sectors transferred in timeout at the estimated peak
3421 * rate. This enables BFQ to utilize a full timeslice with a full
3422 * budget, even if the in-service queue is served at peak rate. And
3423 * this maximises throughput with sequential workloads.
3424 */
3425static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
3426{
3427 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
3428 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
3429}
3430
3431/*
3432 * Update parameters related to throughput and responsiveness, as a
3433 * function of the estimated peak rate. See comments on
3434 * bfq_calc_max_budget(), and on the ref_wr_duration array.
3435 */
3436static void update_thr_responsiveness_params(struct bfq_data *bfqd)
3437{
3438 if (bfqd->bfq_user_max_budget == 0) {
3439 bfqd->bfq_max_budget =
3440 bfq_calc_max_budget(bfqd);
3441 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
3442 }
3443}
3444
3445static void bfq_reset_rate_computation(struct bfq_data *bfqd,
3446 struct request *rq)
3447{
3448 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
3449 bfqd->last_dispatch = bfqd->first_dispatch = blk_time_get_ns();
3450 bfqd->peak_rate_samples = 1;
3451 bfqd->sequential_samples = 0;
3452 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
3453 blk_rq_sectors(rq);
3454 } else /* no new rq dispatched, just reset the number of samples */
3455 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
3456
3457 bfq_log(bfqd,
3458 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
3459 bfqd->peak_rate_samples, bfqd->sequential_samples,
3460 bfqd->tot_sectors_dispatched);
3461}
3462
3463static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
3464{
3465 u32 rate, weight, divisor;
3466
3467 /*
3468 * For the convergence property to hold (see comments on
3469 * bfq_update_peak_rate()) and for the assessment to be
3470 * reliable, a minimum number of samples must be present, and
3471 * a minimum amount of time must have elapsed. If not so, do
3472 * not compute new rate. Just reset parameters, to get ready
3473 * for a new evaluation attempt.
3474 */
3475 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
3476 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
3477 goto reset_computation;
3478
3479 /*
3480 * If a new request completion has occurred after last
3481 * dispatch, then, to approximate the rate at which requests
3482 * have been served by the device, it is more precise to
3483 * extend the observation interval to the last completion.
3484 */
3485 bfqd->delta_from_first =
3486 max_t(u64, bfqd->delta_from_first,
3487 bfqd->last_completion - bfqd->first_dispatch);
3488
3489 /*
3490 * Rate computed in sects/usec, and not sects/nsec, for
3491 * precision issues.
3492 */
3493 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
3494 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
3495
3496 /*
3497 * Peak rate not updated if:
3498 * - the percentage of sequential dispatches is below 3/4 of the
3499 * total, and rate is below the current estimated peak rate
3500 * - rate is unreasonably high (> 20M sectors/sec)
3501 */
3502 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
3503 rate <= bfqd->peak_rate) ||
3504 rate > 20<<BFQ_RATE_SHIFT)
3505 goto reset_computation;
3506
3507 /*
3508 * We have to update the peak rate, at last! To this purpose,
3509 * we use a low-pass filter. We compute the smoothing constant
3510 * of the filter as a function of the 'weight' of the new
3511 * measured rate.
3512 *
3513 * As can be seen in next formulas, we define this weight as a
3514 * quantity proportional to how sequential the workload is,
3515 * and to how long the observation time interval is.
3516 *
3517 * The weight runs from 0 to 8. The maximum value of the
3518 * weight, 8, yields the minimum value for the smoothing
3519 * constant. At this minimum value for the smoothing constant,
3520 * the measured rate contributes for half of the next value of
3521 * the estimated peak rate.
3522 *
3523 * So, the first step is to compute the weight as a function
3524 * of how sequential the workload is. Note that the weight
3525 * cannot reach 9, because bfqd->sequential_samples cannot
3526 * become equal to bfqd->peak_rate_samples, which, in its
3527 * turn, holds true because bfqd->sequential_samples is not
3528 * incremented for the first sample.
3529 */
3530 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
3531
3532 /*
3533 * Second step: further refine the weight as a function of the
3534 * duration of the observation interval.
3535 */
3536 weight = min_t(u32, 8,
3537 div_u64(weight * bfqd->delta_from_first,
3538 BFQ_RATE_REF_INTERVAL));
3539
3540 /*
3541 * Divisor ranging from 10, for minimum weight, to 2, for
3542 * maximum weight.
3543 */
3544 divisor = 10 - weight;
3545
3546 /*
3547 * Finally, update peak rate:
3548 *
3549 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
3550 */
3551 bfqd->peak_rate *= divisor-1;
3552 bfqd->peak_rate /= divisor;
3553 rate /= divisor; /* smoothing constant alpha = 1/divisor */
3554
3555 bfqd->peak_rate += rate;
3556
3557 /*
3558 * For a very slow device, bfqd->peak_rate can reach 0 (see
3559 * the minimum representable values reported in the comments
3560 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3561 * divisions by zero where bfqd->peak_rate is used as a
3562 * divisor.
3563 */
3564 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
3565
3566 update_thr_responsiveness_params(bfqd);
3567
3568reset_computation:
3569 bfq_reset_rate_computation(bfqd, rq);
3570}
3571
3572/*
3573 * Update the read/write peak rate (the main quantity used for
3574 * auto-tuning, see update_thr_responsiveness_params()).
3575 *
3576 * It is not trivial to estimate the peak rate (correctly): because of
3577 * the presence of sw and hw queues between the scheduler and the
3578 * device components that finally serve I/O requests, it is hard to
3579 * say exactly when a given dispatched request is served inside the
3580 * device, and for how long. As a consequence, it is hard to know
3581 * precisely at what rate a given set of requests is actually served
3582 * by the device.
3583 *
3584 * On the opposite end, the dispatch time of any request is trivially
3585 * available, and, from this piece of information, the "dispatch rate"
3586 * of requests can be immediately computed. So, the idea in the next
3587 * function is to use what is known, namely request dispatch times
3588 * (plus, when useful, request completion times), to estimate what is
3589 * unknown, namely in-device request service rate.
3590 *
3591 * The main issue is that, because of the above facts, the rate at
3592 * which a certain set of requests is dispatched over a certain time
3593 * interval can vary greatly with respect to the rate at which the
3594 * same requests are then served. But, since the size of any
3595 * intermediate queue is limited, and the service scheme is lossless
3596 * (no request is silently dropped), the following obvious convergence
3597 * property holds: the number of requests dispatched MUST become
3598 * closer and closer to the number of requests completed as the
3599 * observation interval grows. This is the key property used in
3600 * the next function to estimate the peak service rate as a function
3601 * of the observed dispatch rate. The function assumes to be invoked
3602 * on every request dispatch.
3603 */
3604static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
3605{
3606 u64 now_ns = blk_time_get_ns();
3607
3608 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
3609 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
3610 bfqd->peak_rate_samples);
3611 bfq_reset_rate_computation(bfqd, rq);
3612 goto update_last_values; /* will add one sample */
3613 }
3614
3615 /*
3616 * Device idle for very long: the observation interval lasting
3617 * up to this dispatch cannot be a valid observation interval
3618 * for computing a new peak rate (similarly to the late-
3619 * completion event in bfq_completed_request()). Go to
3620 * update_rate_and_reset to have the following three steps
3621 * taken:
3622 * - close the observation interval at the last (previous)
3623 * request dispatch or completion
3624 * - compute rate, if possible, for that observation interval
3625 * - start a new observation interval with this dispatch
3626 */
3627 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
3628 bfqd->tot_rq_in_driver == 0)
3629 goto update_rate_and_reset;
3630
3631 /* Update sampling information */
3632 bfqd->peak_rate_samples++;
3633
3634 if ((bfqd->tot_rq_in_driver > 0 ||
3635 now_ns - bfqd->last_completion < BFQ_MIN_TT)
3636 && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
3637 bfqd->sequential_samples++;
3638
3639 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
3640
3641 /* Reset max observed rq size every 32 dispatches */
3642 if (likely(bfqd->peak_rate_samples % 32))
3643 bfqd->last_rq_max_size =
3644 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
3645 else
3646 bfqd->last_rq_max_size = blk_rq_sectors(rq);
3647
3648 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
3649
3650 /* Target observation interval not yet reached, go on sampling */
3651 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
3652 goto update_last_values;
3653
3654update_rate_and_reset:
3655 bfq_update_rate_reset(bfqd, rq);
3656update_last_values:
3657 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
3658 if (RQ_BFQQ(rq) == bfqd->in_service_queue)
3659 bfqd->in_serv_last_pos = bfqd->last_position;
3660 bfqd->last_dispatch = now_ns;
3661}
3662
3663/*
3664 * Remove request from internal lists.
3665 */
3666static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
3667{
3668 struct bfq_queue *bfqq = RQ_BFQQ(rq);
3669
3670 /*
3671 * For consistency, the next instruction should have been
3672 * executed after removing the request from the queue and
3673 * dispatching it. We execute instead this instruction before
3674 * bfq_remove_request() (and hence introduce a temporary
3675 * inconsistency), for efficiency. In fact, should this
3676 * dispatch occur for a non in-service bfqq, this anticipated
3677 * increment prevents two counters related to bfqq->dispatched
3678 * from risking to be, first, uselessly decremented, and then
3679 * incremented again when the (new) value of bfqq->dispatched
3680 * happens to be taken into account.
3681 */
3682 bfqq->dispatched++;
3683 bfq_update_peak_rate(q->elevator->elevator_data, rq);
3684
3685 bfq_remove_request(q, rq);
3686}
3687
3688/*
3689 * There is a case where idling does not have to be performed for
3690 * throughput concerns, but to preserve the throughput share of
3691 * the process associated with bfqq.
3692 *
3693 * To introduce this case, we can note that allowing the drive
3694 * to enqueue more than one request at a time, and hence
3695 * delegating de facto final scheduling decisions to the
3696 * drive's internal scheduler, entails loss of control on the
3697 * actual request service order. In particular, the critical
3698 * situation is when requests from different processes happen
3699 * to be present, at the same time, in the internal queue(s)
3700 * of the drive. In such a situation, the drive, by deciding
3701 * the service order of the internally-queued requests, does
3702 * determine also the actual throughput distribution among
3703 * these processes. But the drive typically has no notion or
3704 * concern about per-process throughput distribution, and
3705 * makes its decisions only on a per-request basis. Therefore,
3706 * the service distribution enforced by the drive's internal
3707 * scheduler is likely to coincide with the desired throughput
3708 * distribution only in a completely symmetric, or favorably
3709 * skewed scenario where:
3710 * (i-a) each of these processes must get the same throughput as
3711 * the others,
3712 * (i-b) in case (i-a) does not hold, it holds that the process
3713 * associated with bfqq must receive a lower or equal
3714 * throughput than any of the other processes;
3715 * (ii) the I/O of each process has the same properties, in
3716 * terms of locality (sequential or random), direction
3717 * (reads or writes), request sizes, greediness
3718 * (from I/O-bound to sporadic), and so on;
3719
3720 * In fact, in such a scenario, the drive tends to treat the requests
3721 * of each process in about the same way as the requests of the
3722 * others, and thus to provide each of these processes with about the
3723 * same throughput. This is exactly the desired throughput
3724 * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3725 * even more convenient distribution for (the process associated with)
3726 * bfqq.
3727 *
3728 * In contrast, in any asymmetric or unfavorable scenario, device
3729 * idling (I/O-dispatch plugging) is certainly needed to guarantee
3730 * that bfqq receives its assigned fraction of the device throughput
3731 * (see [1] for details).
3732 *
3733 * The problem is that idling may significantly reduce throughput with
3734 * certain combinations of types of I/O and devices. An important
3735 * example is sync random I/O on flash storage with command
3736 * queueing. So, unless bfqq falls in cases where idling also boosts
3737 * throughput, it is important to check conditions (i-a), i(-b) and
3738 * (ii) accurately, so as to avoid idling when not strictly needed for
3739 * service guarantees.
3740 *
3741 * Unfortunately, it is extremely difficult to thoroughly check
3742 * condition (ii). And, in case there are active groups, it becomes
3743 * very difficult to check conditions (i-a) and (i-b) too. In fact,
3744 * if there are active groups, then, for conditions (i-a) or (i-b) to
3745 * become false 'indirectly', it is enough that an active group
3746 * contains more active processes or sub-groups than some other active
3747 * group. More precisely, for conditions (i-a) or (i-b) to become
3748 * false because of such a group, it is not even necessary that the
3749 * group is (still) active: it is sufficient that, even if the group
3750 * has become inactive, some of its descendant processes still have
3751 * some request already dispatched but still waiting for
3752 * completion. In fact, requests have still to be guaranteed their
3753 * share of the throughput even after being dispatched. In this
3754 * respect, it is easy to show that, if a group frequently becomes
3755 * inactive while still having in-flight requests, and if, when this
3756 * happens, the group is not considered in the calculation of whether
3757 * the scenario is asymmetric, then the group may fail to be
3758 * guaranteed its fair share of the throughput (basically because
3759 * idling may not be performed for the descendant processes of the
3760 * group, but it had to be). We address this issue with the following
3761 * bi-modal behavior, implemented in the function
3762 * bfq_asymmetric_scenario().
3763 *
3764 * If there are groups with requests waiting for completion
3765 * (as commented above, some of these groups may even be
3766 * already inactive), then the scenario is tagged as
3767 * asymmetric, conservatively, without checking any of the
3768 * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3769 * This behavior matches also the fact that groups are created
3770 * exactly if controlling I/O is a primary concern (to
3771 * preserve bandwidth and latency guarantees).
3772 *
3773 * On the opposite end, if there are no groups with requests waiting
3774 * for completion, then only conditions (i-a) and (i-b) are actually
3775 * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3776 * idling is not performed, regardless of whether condition (ii)
3777 * holds. In other words, only if conditions (i-a) and (i-b) do not
3778 * hold, then idling is allowed, and the device tends to be prevented
3779 * from queueing many requests, possibly of several processes. Since
3780 * there are no groups with requests waiting for completion, then, to
3781 * control conditions (i-a) and (i-b) it is enough to check just
3782 * whether all the queues with requests waiting for completion also
3783 * have the same weight.
3784 *
3785 * Not checking condition (ii) evidently exposes bfqq to the
3786 * risk of getting less throughput than its fair share.
3787 * However, for queues with the same weight, a further
3788 * mechanism, preemption, mitigates or even eliminates this
3789 * problem. And it does so without consequences on overall
3790 * throughput. This mechanism and its benefits are explained
3791 * in the next three paragraphs.
3792 *
3793 * Even if a queue, say Q, is expired when it remains idle, Q
3794 * can still preempt the new in-service queue if the next
3795 * request of Q arrives soon (see the comments on
3796 * bfq_bfqq_update_budg_for_activation). If all queues and
3797 * groups have the same weight, this form of preemption,
3798 * combined with the hole-recovery heuristic described in the
3799 * comments on function bfq_bfqq_update_budg_for_activation,
3800 * are enough to preserve a correct bandwidth distribution in
3801 * the mid term, even without idling. In fact, even if not
3802 * idling allows the internal queues of the device to contain
3803 * many requests, and thus to reorder requests, we can rather
3804 * safely assume that the internal scheduler still preserves a
3805 * minimum of mid-term fairness.
3806 *
3807 * More precisely, this preemption-based, idleless approach
3808 * provides fairness in terms of IOPS, and not sectors per
3809 * second. This can be seen with a simple example. Suppose
3810 * that there are two queues with the same weight, but that
3811 * the first queue receives requests of 8 sectors, while the
3812 * second queue receives requests of 1024 sectors. In
3813 * addition, suppose that each of the two queues contains at
3814 * most one request at a time, which implies that each queue
3815 * always remains idle after it is served. Finally, after
3816 * remaining idle, each queue receives very quickly a new
3817 * request. It follows that the two queues are served
3818 * alternatively, preempting each other if needed. This
3819 * implies that, although both queues have the same weight,
3820 * the queue with large requests receives a service that is
3821 * 1024/8 times as high as the service received by the other
3822 * queue.
3823 *
3824 * The motivation for using preemption instead of idling (for
3825 * queues with the same weight) is that, by not idling,
3826 * service guarantees are preserved (completely or at least in
3827 * part) without minimally sacrificing throughput. And, if
3828 * there is no active group, then the primary expectation for
3829 * this device is probably a high throughput.
3830 *
3831 * We are now left only with explaining the two sub-conditions in the
3832 * additional compound condition that is checked below for deciding
3833 * whether the scenario is asymmetric. To explain the first
3834 * sub-condition, we need to add that the function
3835 * bfq_asymmetric_scenario checks the weights of only
3836 * non-weight-raised queues, for efficiency reasons (see comments on
3837 * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3838 * is checked explicitly here. More precisely, the compound condition
3839 * below takes into account also the fact that, even if bfqq is being
3840 * weight-raised, the scenario is still symmetric if all queues with
3841 * requests waiting for completion happen to be
3842 * weight-raised. Actually, we should be even more precise here, and
3843 * differentiate between interactive weight raising and soft real-time
3844 * weight raising.
3845 *
3846 * The second sub-condition checked in the compound condition is
3847 * whether there is a fair amount of already in-flight I/O not
3848 * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3849 * following reason. The drive may decide to serve in-flight
3850 * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3851 * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3852 * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3853 * basically uncontrolled amount of I/O from other queues may be
3854 * dispatched too, possibly causing the service of bfqq's I/O to be
3855 * delayed even longer in the drive. This problem gets more and more
3856 * serious as the speed and the queue depth of the drive grow,
3857 * because, as these two quantities grow, the probability to find no
3858 * queue busy but many requests in flight grows too. By contrast,
3859 * plugging I/O dispatching minimizes the delay induced by already
3860 * in-flight I/O, and enables bfqq to recover the bandwidth it may
3861 * lose because of this delay.
3862 *
3863 * As a side note, it is worth considering that the above
3864 * device-idling countermeasures may however fail in the following
3865 * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3866 * in a time period during which all symmetry sub-conditions hold, and
3867 * therefore the device is allowed to enqueue many requests, but at
3868 * some later point in time some sub-condition stops to hold, then it
3869 * may become impossible to make requests be served in the desired
3870 * order until all the requests already queued in the device have been
3871 * served. The last sub-condition commented above somewhat mitigates
3872 * this problem for weight-raised queues.
3873 *
3874 * However, as an additional mitigation for this problem, we preserve
3875 * plugging for a special symmetric case that may suddenly turn into
3876 * asymmetric: the case where only bfqq is busy. In this case, not
3877 * expiring bfqq does not cause any harm to any other queues in terms
3878 * of service guarantees. In contrast, it avoids the following unlucky
3879 * sequence of events: (1) bfqq is expired, (2) a new queue with a
3880 * lower weight than bfqq becomes busy (or more queues), (3) the new
3881 * queue is served until a new request arrives for bfqq, (4) when bfqq
3882 * is finally served, there are so many requests of the new queue in
3883 * the drive that the pending requests for bfqq take a lot of time to
3884 * be served. In particular, event (2) may case even already
3885 * dispatched requests of bfqq to be delayed, inside the drive. So, to
3886 * avoid this series of events, the scenario is preventively declared
3887 * as asymmetric also if bfqq is the only busy queues
3888 */
3889static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3890 struct bfq_queue *bfqq)
3891{
3892 int tot_busy_queues = bfq_tot_busy_queues(bfqd);
3893
3894 /* No point in idling for bfqq if it won't get requests any longer */
3895 if (unlikely(!bfqq_process_refs(bfqq)))
3896 return false;
3897
3898 return (bfqq->wr_coeff > 1 &&
3899 (bfqd->wr_busy_queues < tot_busy_queues ||
3900 bfqd->tot_rq_in_driver >= bfqq->dispatched + 4)) ||
3901 bfq_asymmetric_scenario(bfqd, bfqq) ||
3902 tot_busy_queues == 1;
3903}
3904
3905static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3906 enum bfqq_expiration reason)
3907{
3908 /*
3909 * If this bfqq is shared between multiple processes, check
3910 * to make sure that those processes are still issuing I/Os
3911 * within the mean seek distance. If not, it may be time to
3912 * break the queues apart again.
3913 */
3914 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
3915 bfq_mark_bfqq_split_coop(bfqq);
3916
3917 /*
3918 * Consider queues with a higher finish virtual time than
3919 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3920 * true, then bfqq's bandwidth would be violated if an
3921 * uncontrolled amount of I/O from these queues were
3922 * dispatched while bfqq is waiting for its new I/O to
3923 * arrive. This is exactly what may happen if this is a forced
3924 * expiration caused by a preemption attempt, and if bfqq is
3925 * not re-scheduled. To prevent this from happening, re-queue
3926 * bfqq if it needs I/O-dispatch plugging, even if it is
3927 * empty. By doing so, bfqq is granted to be served before the
3928 * above queues (provided that bfqq is of course eligible).
3929 */
3930 if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
3931 !(reason == BFQQE_PREEMPTED &&
3932 idling_needed_for_service_guarantees(bfqd, bfqq))) {
3933 if (bfqq->dispatched == 0)
3934 /*
3935 * Overloading budget_timeout field to store
3936 * the time at which the queue remains with no
3937 * backlog and no outstanding request; used by
3938 * the weight-raising mechanism.
3939 */
3940 bfqq->budget_timeout = jiffies;
3941
3942 bfq_del_bfqq_busy(bfqq, true);
3943 } else {
3944 bfq_requeue_bfqq(bfqd, bfqq, true);
3945 /*
3946 * Resort priority tree of potential close cooperators.
3947 * See comments on bfq_pos_tree_add_move() for the unlikely().
3948 */
3949 if (unlikely(!bfqd->nonrot_with_queueing &&
3950 !RB_EMPTY_ROOT(&bfqq->sort_list)))
3951 bfq_pos_tree_add_move(bfqd, bfqq);
3952 }
3953
3954 /*
3955 * All in-service entities must have been properly deactivated
3956 * or requeued before executing the next function, which
3957 * resets all in-service entities as no more in service. This
3958 * may cause bfqq to be freed. If this happens, the next
3959 * function returns true.
3960 */
3961 return __bfq_bfqd_reset_in_service(bfqd);
3962}
3963
3964/**
3965 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3966 * @bfqd: device data.
3967 * @bfqq: queue to update.
3968 * @reason: reason for expiration.
3969 *
3970 * Handle the feedback on @bfqq budget at queue expiration.
3971 * See the body for detailed comments.
3972 */
3973static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
3974 struct bfq_queue *bfqq,
3975 enum bfqq_expiration reason)
3976{
3977 struct request *next_rq;
3978 int budget, min_budget;
3979
3980 min_budget = bfq_min_budget(bfqd);
3981
3982 if (bfqq->wr_coeff == 1)
3983 budget = bfqq->max_budget;
3984 else /*
3985 * Use a constant, low budget for weight-raised queues,
3986 * to help achieve a low latency. Keep it slightly higher
3987 * than the minimum possible budget, to cause a little
3988 * bit fewer expirations.
3989 */
3990 budget = 2 * min_budget;
3991
3992 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
3993 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
3994 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
3995 budget, bfq_min_budget(bfqd));
3996 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
3997 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
3998
3999 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
4000 switch (reason) {
4001 /*
4002 * Caveat: in all the following cases we trade latency
4003 * for throughput.
4004 */
4005 case BFQQE_TOO_IDLE:
4006 /*
4007 * This is the only case where we may reduce
4008 * the budget: if there is no request of the
4009 * process still waiting for completion, then
4010 * we assume (tentatively) that the timer has
4011 * expired because the batch of requests of
4012 * the process could have been served with a
4013 * smaller budget. Hence, betting that
4014 * process will behave in the same way when it
4015 * becomes backlogged again, we reduce its
4016 * next budget. As long as we guess right,
4017 * this budget cut reduces the latency
4018 * experienced by the process.
4019 *
4020 * However, if there are still outstanding
4021 * requests, then the process may have not yet
4022 * issued its next request just because it is
4023 * still waiting for the completion of some of
4024 * the still outstanding ones. So in this
4025 * subcase we do not reduce its budget, on the
4026 * contrary we increase it to possibly boost
4027 * the throughput, as discussed in the
4028 * comments to the BUDGET_TIMEOUT case.
4029 */
4030 if (bfqq->dispatched > 0) /* still outstanding reqs */
4031 budget = min(budget * 2, bfqd->bfq_max_budget);
4032 else {
4033 if (budget > 5 * min_budget)
4034 budget -= 4 * min_budget;
4035 else
4036 budget = min_budget;
4037 }
4038 break;
4039 case BFQQE_BUDGET_TIMEOUT:
4040 /*
4041 * We double the budget here because it gives
4042 * the chance to boost the throughput if this
4043 * is not a seeky process (and has bumped into
4044 * this timeout because of, e.g., ZBR).
4045 */
4046 budget = min(budget * 2, bfqd->bfq_max_budget);
4047 break;
4048 case BFQQE_BUDGET_EXHAUSTED:
4049 /*
4050 * The process still has backlog, and did not
4051 * let either the budget timeout or the disk
4052 * idling timeout expire. Hence it is not
4053 * seeky, has a short thinktime and may be
4054 * happy with a higher budget too. So
4055 * definitely increase the budget of this good
4056 * candidate to boost the disk throughput.
4057 */
4058 budget = min(budget * 4, bfqd->bfq_max_budget);
4059 break;
4060 case BFQQE_NO_MORE_REQUESTS:
4061 /*
4062 * For queues that expire for this reason, it
4063 * is particularly important to keep the
4064 * budget close to the actual service they
4065 * need. Doing so reduces the timestamp
4066 * misalignment problem described in the
4067 * comments in the body of
4068 * __bfq_activate_entity. In fact, suppose
4069 * that a queue systematically expires for
4070 * BFQQE_NO_MORE_REQUESTS and presents a
4071 * new request in time to enjoy timestamp
4072 * back-shifting. The larger the budget of the
4073 * queue is with respect to the service the
4074 * queue actually requests in each service
4075 * slot, the more times the queue can be
4076 * reactivated with the same virtual finish
4077 * time. It follows that, even if this finish
4078 * time is pushed to the system virtual time
4079 * to reduce the consequent timestamp
4080 * misalignment, the queue unjustly enjoys for
4081 * many re-activations a lower finish time
4082 * than all newly activated queues.
4083 *
4084 * The service needed by bfqq is measured
4085 * quite precisely by bfqq->entity.service.
4086 * Since bfqq does not enjoy device idling,
4087 * bfqq->entity.service is equal to the number
4088 * of sectors that the process associated with
4089 * bfqq requested to read/write before waiting
4090 * for request completions, or blocking for
4091 * other reasons.
4092 */
4093 budget = max_t(int, bfqq->entity.service, min_budget);
4094 break;
4095 default:
4096 return;
4097 }
4098 } else if (!bfq_bfqq_sync(bfqq)) {
4099 /*
4100 * Async queues get always the maximum possible
4101 * budget, as for them we do not care about latency
4102 * (in addition, their ability to dispatch is limited
4103 * by the charging factor).
4104 */
4105 budget = bfqd->bfq_max_budget;
4106 }
4107
4108 bfqq->max_budget = budget;
4109
4110 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
4111 !bfqd->bfq_user_max_budget)
4112 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
4113
4114 /*
4115 * If there is still backlog, then assign a new budget, making
4116 * sure that it is large enough for the next request. Since
4117 * the finish time of bfqq must be kept in sync with the
4118 * budget, be sure to call __bfq_bfqq_expire() *after* this
4119 * update.
4120 *
4121 * If there is no backlog, then no need to update the budget;
4122 * it will be updated on the arrival of a new request.
4123 */
4124 next_rq = bfqq->next_rq;
4125 if (next_rq)
4126 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
4127 bfq_serv_to_charge(next_rq, bfqq));
4128
4129 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
4130 next_rq ? blk_rq_sectors(next_rq) : 0,
4131 bfqq->entity.budget);
4132}
4133
4134/*
4135 * Return true if the process associated with bfqq is "slow". The slow
4136 * flag is used, in addition to the budget timeout, to reduce the
4137 * amount of service provided to seeky processes, and thus reduce
4138 * their chances to lower the throughput. More details in the comments
4139 * on the function bfq_bfqq_expire().
4140 *
4141 * An important observation is in order: as discussed in the comments
4142 * on the function bfq_update_peak_rate(), with devices with internal
4143 * queues, it is hard if ever possible to know when and for how long
4144 * an I/O request is processed by the device (apart from the trivial
4145 * I/O pattern where a new request is dispatched only after the
4146 * previous one has been completed). This makes it hard to evaluate
4147 * the real rate at which the I/O requests of each bfq_queue are
4148 * served. In fact, for an I/O scheduler like BFQ, serving a
4149 * bfq_queue means just dispatching its requests during its service
4150 * slot (i.e., until the budget of the queue is exhausted, or the
4151 * queue remains idle, or, finally, a timeout fires). But, during the
4152 * service slot of a bfq_queue, around 100 ms at most, the device may
4153 * be even still processing requests of bfq_queues served in previous
4154 * service slots. On the opposite end, the requests of the in-service
4155 * bfq_queue may be completed after the service slot of the queue
4156 * finishes.
4157 *
4158 * Anyway, unless more sophisticated solutions are used
4159 * (where possible), the sum of the sizes of the requests dispatched
4160 * during the service slot of a bfq_queue is probably the only
4161 * approximation available for the service received by the bfq_queue
4162 * during its service slot. And this sum is the quantity used in this
4163 * function to evaluate the I/O speed of a process.
4164 */
4165static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4166 bool compensate, unsigned long *delta_ms)
4167{
4168 ktime_t delta_ktime;
4169 u32 delta_usecs;
4170 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
4171
4172 if (!bfq_bfqq_sync(bfqq))
4173 return false;
4174
4175 if (compensate)
4176 delta_ktime = bfqd->last_idling_start;
4177 else
4178 delta_ktime = blk_time_get();
4179 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
4180 delta_usecs = ktime_to_us(delta_ktime);
4181
4182 /* don't use too short time intervals */
4183 if (delta_usecs < 1000) {
4184 if (blk_queue_nonrot(bfqd->queue))
4185 /*
4186 * give same worst-case guarantees as idling
4187 * for seeky
4188 */
4189 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
4190 else /* charge at least one seek */
4191 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
4192
4193 return slow;
4194 }
4195
4196 *delta_ms = delta_usecs / USEC_PER_MSEC;
4197
4198 /*
4199 * Use only long (> 20ms) intervals to filter out excessive
4200 * spikes in service rate estimation.
4201 */
4202 if (delta_usecs > 20000) {
4203 /*
4204 * Caveat for rotational devices: processes doing I/O
4205 * in the slower disk zones tend to be slow(er) even
4206 * if not seeky. In this respect, the estimated peak
4207 * rate is likely to be an average over the disk
4208 * surface. Accordingly, to not be too harsh with
4209 * unlucky processes, a process is deemed slow only if
4210 * its rate has been lower than half of the estimated
4211 * peak rate.
4212 */
4213 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
4214 }
4215
4216 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
4217
4218 return slow;
4219}
4220
4221/*
4222 * To be deemed as soft real-time, an application must meet two
4223 * requirements. First, the application must not require an average
4224 * bandwidth higher than the approximate bandwidth required to playback or
4225 * record a compressed high-definition video.
4226 * The next function is invoked on the completion of the last request of a
4227 * batch, to compute the next-start time instant, soft_rt_next_start, such
4228 * that, if the next request of the application does not arrive before
4229 * soft_rt_next_start, then the above requirement on the bandwidth is met.
4230 *
4231 * The second requirement is that the request pattern of the application is
4232 * isochronous, i.e., that, after issuing a request or a batch of requests,
4233 * the application stops issuing new requests until all its pending requests
4234 * have been completed. After that, the application may issue a new batch,
4235 * and so on.
4236 * For this reason the next function is invoked to compute
4237 * soft_rt_next_start only for applications that meet this requirement,
4238 * whereas soft_rt_next_start is set to infinity for applications that do
4239 * not.
4240 *
4241 * Unfortunately, even a greedy (i.e., I/O-bound) application may
4242 * happen to meet, occasionally or systematically, both the above
4243 * bandwidth and isochrony requirements. This may happen at least in
4244 * the following circumstances. First, if the CPU load is high. The
4245 * application may stop issuing requests while the CPUs are busy
4246 * serving other processes, then restart, then stop again for a while,
4247 * and so on. The other circumstances are related to the storage
4248 * device: the storage device is highly loaded or reaches a low-enough
4249 * throughput with the I/O of the application (e.g., because the I/O
4250 * is random and/or the device is slow). In all these cases, the
4251 * I/O of the application may be simply slowed down enough to meet
4252 * the bandwidth and isochrony requirements. To reduce the probability
4253 * that greedy applications are deemed as soft real-time in these
4254 * corner cases, a further rule is used in the computation of
4255 * soft_rt_next_start: the return value of this function is forced to
4256 * be higher than the maximum between the following two quantities.
4257 *
4258 * (a) Current time plus: (1) the maximum time for which the arrival
4259 * of a request is waited for when a sync queue becomes idle,
4260 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
4261 * postpone for a moment the reason for adding a few extra
4262 * jiffies; we get back to it after next item (b). Lower-bounding
4263 * the return value of this function with the current time plus
4264 * bfqd->bfq_slice_idle tends to filter out greedy applications,
4265 * because the latter issue their next request as soon as possible
4266 * after the last one has been completed. In contrast, a soft
4267 * real-time application spends some time processing data, after a
4268 * batch of its requests has been completed.
4269 *
4270 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
4271 * above, greedy applications may happen to meet both the
4272 * bandwidth and isochrony requirements under heavy CPU or
4273 * storage-device load. In more detail, in these scenarios, these
4274 * applications happen, only for limited time periods, to do I/O
4275 * slowly enough to meet all the requirements described so far,
4276 * including the filtering in above item (a). These slow-speed
4277 * time intervals are usually interspersed between other time
4278 * intervals during which these applications do I/O at a very high
4279 * speed. Fortunately, exactly because of the high speed of the
4280 * I/O in the high-speed intervals, the values returned by this
4281 * function happen to be so high, near the end of any such
4282 * high-speed interval, to be likely to fall *after* the end of
4283 * the low-speed time interval that follows. These high values are
4284 * stored in bfqq->soft_rt_next_start after each invocation of
4285 * this function. As a consequence, if the last value of
4286 * bfqq->soft_rt_next_start is constantly used to lower-bound the
4287 * next value that this function may return, then, from the very
4288 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
4289 * likely to be constantly kept so high that any I/O request
4290 * issued during the low-speed interval is considered as arriving
4291 * to soon for the application to be deemed as soft
4292 * real-time. Then, in the high-speed interval that follows, the
4293 * application will not be deemed as soft real-time, just because
4294 * it will do I/O at a high speed. And so on.
4295 *
4296 * Getting back to the filtering in item (a), in the following two
4297 * cases this filtering might be easily passed by a greedy
4298 * application, if the reference quantity was just
4299 * bfqd->bfq_slice_idle:
4300 * 1) HZ is so low that the duration of a jiffy is comparable to or
4301 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
4302 * devices with HZ=100. The time granularity may be so coarse
4303 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
4304 * is rather lower than the exact value.
4305 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
4306 * for a while, then suddenly 'jump' by several units to recover the lost
4307 * increments. This seems to happen, e.g., inside virtual machines.
4308 * To address this issue, in the filtering in (a) we do not use as a
4309 * reference time interval just bfqd->bfq_slice_idle, but
4310 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
4311 * minimum number of jiffies for which the filter seems to be quite
4312 * precise also in embedded systems and KVM/QEMU virtual machines.
4313 */
4314static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
4315 struct bfq_queue *bfqq)
4316{
4317 return max3(bfqq->soft_rt_next_start,
4318 bfqq->last_idle_bklogged +
4319 HZ * bfqq->service_from_backlogged /
4320 bfqd->bfq_wr_max_softrt_rate,
4321 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
4322}
4323
4324/**
4325 * bfq_bfqq_expire - expire a queue.
4326 * @bfqd: device owning the queue.
4327 * @bfqq: the queue to expire.
4328 * @compensate: if true, compensate for the time spent idling.
4329 * @reason: the reason causing the expiration.
4330 *
4331 * If the process associated with bfqq does slow I/O (e.g., because it
4332 * issues random requests), we charge bfqq with the time it has been
4333 * in service instead of the service it has received (see
4334 * bfq_bfqq_charge_time for details on how this goal is achieved). As
4335 * a consequence, bfqq will typically get higher timestamps upon
4336 * reactivation, and hence it will be rescheduled as if it had
4337 * received more service than what it has actually received. In the
4338 * end, bfqq receives less service in proportion to how slowly its
4339 * associated process consumes its budgets (and hence how seriously it
4340 * tends to lower the throughput). In addition, this time-charging
4341 * strategy guarantees time fairness among slow processes. In
4342 * contrast, if the process associated with bfqq is not slow, we
4343 * charge bfqq exactly with the service it has received.
4344 *
4345 * Charging time to the first type of queues and the exact service to
4346 * the other has the effect of using the WF2Q+ policy to schedule the
4347 * former on a timeslice basis, without violating service domain
4348 * guarantees among the latter.
4349 */
4350void bfq_bfqq_expire(struct bfq_data *bfqd,
4351 struct bfq_queue *bfqq,
4352 bool compensate,
4353 enum bfqq_expiration reason)
4354{
4355 bool slow;
4356 unsigned long delta = 0;
4357 struct bfq_entity *entity = &bfqq->entity;
4358
4359 /*
4360 * Check whether the process is slow (see bfq_bfqq_is_slow).
4361 */
4362 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, &delta);
4363
4364 /*
4365 * As above explained, charge slow (typically seeky) and
4366 * timed-out queues with the time and not the service
4367 * received, to favor sequential workloads.
4368 *
4369 * Processes doing I/O in the slower disk zones will tend to
4370 * be slow(er) even if not seeky. Therefore, since the
4371 * estimated peak rate is actually an average over the disk
4372 * surface, these processes may timeout just for bad luck. To
4373 * avoid punishing them, do not charge time to processes that
4374 * succeeded in consuming at least 2/3 of their budget. This
4375 * allows BFQ to preserve enough elasticity to still perform
4376 * bandwidth, and not time, distribution with little unlucky
4377 * or quasi-sequential processes.
4378 */
4379 if (bfqq->wr_coeff == 1 &&
4380 (slow ||
4381 (reason == BFQQE_BUDGET_TIMEOUT &&
4382 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
4383 bfq_bfqq_charge_time(bfqd, bfqq, delta);
4384
4385 if (bfqd->low_latency && bfqq->wr_coeff == 1)
4386 bfqq->last_wr_start_finish = jiffies;
4387
4388 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
4389 RB_EMPTY_ROOT(&bfqq->sort_list)) {
4390 /*
4391 * If we get here, and there are no outstanding
4392 * requests, then the request pattern is isochronous
4393 * (see the comments on the function
4394 * bfq_bfqq_softrt_next_start()). Therefore we can
4395 * compute soft_rt_next_start.
4396 *
4397 * If, instead, the queue still has outstanding
4398 * requests, then we have to wait for the completion
4399 * of all the outstanding requests to discover whether
4400 * the request pattern is actually isochronous.
4401 */
4402 if (bfqq->dispatched == 0)
4403 bfqq->soft_rt_next_start =
4404 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4405 else if (bfqq->dispatched > 0) {
4406 /*
4407 * Schedule an update of soft_rt_next_start to when
4408 * the task may be discovered to be isochronous.
4409 */
4410 bfq_mark_bfqq_softrt_update(bfqq);
4411 }
4412 }
4413
4414 bfq_log_bfqq(bfqd, bfqq,
4415 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
4416 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
4417
4418 /*
4419 * bfqq expired, so no total service time needs to be computed
4420 * any longer: reset state machine for measuring total service
4421 * times.
4422 */
4423 bfqd->rqs_injected = bfqd->wait_dispatch = false;
4424 bfqd->waited_rq = NULL;
4425
4426 /*
4427 * Increase, decrease or leave budget unchanged according to
4428 * reason.
4429 */
4430 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
4431 if (__bfq_bfqq_expire(bfqd, bfqq, reason))
4432 /* bfqq is gone, no more actions on it */
4433 return;
4434
4435 /* mark bfqq as waiting a request only if a bic still points to it */
4436 if (!bfq_bfqq_busy(bfqq) &&
4437 reason != BFQQE_BUDGET_TIMEOUT &&
4438 reason != BFQQE_BUDGET_EXHAUSTED) {
4439 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
4440 /*
4441 * Not setting service to 0, because, if the next rq
4442 * arrives in time, the queue will go on receiving
4443 * service with this same budget (as if it never expired)
4444 */
4445 } else
4446 entity->service = 0;
4447
4448 /*
4449 * Reset the received-service counter for every parent entity.
4450 * Differently from what happens with bfqq->entity.service,
4451 * the resetting of this counter never needs to be postponed
4452 * for parent entities. In fact, in case bfqq may have a
4453 * chance to go on being served using the last, partially
4454 * consumed budget, bfqq->entity.service needs to be kept,
4455 * because if bfqq then actually goes on being served using
4456 * the same budget, the last value of bfqq->entity.service is
4457 * needed to properly decrement bfqq->entity.budget by the
4458 * portion already consumed. In contrast, it is not necessary
4459 * to keep entity->service for parent entities too, because
4460 * the bubble up of the new value of bfqq->entity.budget will
4461 * make sure that the budgets of parent entities are correct,
4462 * even in case bfqq and thus parent entities go on receiving
4463 * service with the same budget.
4464 */
4465 entity = entity->parent;
4466 for_each_entity(entity)
4467 entity->service = 0;
4468}
4469
4470/*
4471 * Budget timeout is not implemented through a dedicated timer, but
4472 * just checked on request arrivals and completions, as well as on
4473 * idle timer expirations.
4474 */
4475static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
4476{
4477 return time_is_before_eq_jiffies(bfqq->budget_timeout);
4478}
4479
4480/*
4481 * If we expire a queue that is actively waiting (i.e., with the
4482 * device idled) for the arrival of a new request, then we may incur
4483 * the timestamp misalignment problem described in the body of the
4484 * function __bfq_activate_entity. Hence we return true only if this
4485 * condition does not hold, or if the queue is slow enough to deserve
4486 * only to be kicked off for preserving a high throughput.
4487 */
4488static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
4489{
4490 bfq_log_bfqq(bfqq->bfqd, bfqq,
4491 "may_budget_timeout: wait_request %d left %d timeout %d",
4492 bfq_bfqq_wait_request(bfqq),
4493 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
4494 bfq_bfqq_budget_timeout(bfqq));
4495
4496 return (!bfq_bfqq_wait_request(bfqq) ||
4497 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
4498 &&
4499 bfq_bfqq_budget_timeout(bfqq);
4500}
4501
4502static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
4503 struct bfq_queue *bfqq)
4504{
4505 bool rot_without_queueing =
4506 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
4507 bfqq_sequential_and_IO_bound,
4508 idling_boosts_thr;
4509
4510 /* No point in idling for bfqq if it won't get requests any longer */
4511 if (unlikely(!bfqq_process_refs(bfqq)))
4512 return false;
4513
4514 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
4515 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
4516
4517 /*
4518 * The next variable takes into account the cases where idling
4519 * boosts the throughput.
4520 *
4521 * The value of the variable is computed considering, first, that
4522 * idling is virtually always beneficial for the throughput if:
4523 * (a) the device is not NCQ-capable and rotational, or
4524 * (b) regardless of the presence of NCQ, the device is rotational and
4525 * the request pattern for bfqq is I/O-bound and sequential, or
4526 * (c) regardless of whether it is rotational, the device is
4527 * not NCQ-capable and the request pattern for bfqq is
4528 * I/O-bound and sequential.
4529 *
4530 * Secondly, and in contrast to the above item (b), idling an
4531 * NCQ-capable flash-based device would not boost the
4532 * throughput even with sequential I/O; rather it would lower
4533 * the throughput in proportion to how fast the device
4534 * is. Accordingly, the next variable is true if any of the
4535 * above conditions (a), (b) or (c) is true, and, in
4536 * particular, happens to be false if bfqd is an NCQ-capable
4537 * flash-based device.
4538 */
4539 idling_boosts_thr = rot_without_queueing ||
4540 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
4541 bfqq_sequential_and_IO_bound);
4542
4543 /*
4544 * The return value of this function is equal to that of
4545 * idling_boosts_thr, unless a special case holds. In this
4546 * special case, described below, idling may cause problems to
4547 * weight-raised queues.
4548 *
4549 * When the request pool is saturated (e.g., in the presence
4550 * of write hogs), if the processes associated with
4551 * non-weight-raised queues ask for requests at a lower rate,
4552 * then processes associated with weight-raised queues have a
4553 * higher probability to get a request from the pool
4554 * immediately (or at least soon) when they need one. Thus
4555 * they have a higher probability to actually get a fraction
4556 * of the device throughput proportional to their high
4557 * weight. This is especially true with NCQ-capable drives,
4558 * which enqueue several requests in advance, and further
4559 * reorder internally-queued requests.
4560 *
4561 * For this reason, we force to false the return value if
4562 * there are weight-raised busy queues. In this case, and if
4563 * bfqq is not weight-raised, this guarantees that the device
4564 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4565 * then idling will be guaranteed by another variable, see
4566 * below). Combined with the timestamping rules of BFQ (see
4567 * [1] for details), this behavior causes bfqq, and hence any
4568 * sync non-weight-raised queue, to get a lower number of
4569 * requests served, and thus to ask for a lower number of
4570 * requests from the request pool, before the busy
4571 * weight-raised queues get served again. This often mitigates
4572 * starvation problems in the presence of heavy write
4573 * workloads and NCQ, thereby guaranteeing a higher
4574 * application and system responsiveness in these hostile
4575 * scenarios.
4576 */
4577 return idling_boosts_thr &&
4578 bfqd->wr_busy_queues == 0;
4579}
4580
4581/*
4582 * For a queue that becomes empty, device idling is allowed only if
4583 * this function returns true for that queue. As a consequence, since
4584 * device idling plays a critical role for both throughput boosting
4585 * and service guarantees, the return value of this function plays a
4586 * critical role as well.
4587 *
4588 * In a nutshell, this function returns true only if idling is
4589 * beneficial for throughput or, even if detrimental for throughput,
4590 * idling is however necessary to preserve service guarantees (low
4591 * latency, desired throughput distribution, ...). In particular, on
4592 * NCQ-capable devices, this function tries to return false, so as to
4593 * help keep the drives' internal queues full, whenever this helps the
4594 * device boost the throughput without causing any service-guarantee
4595 * issue.
4596 *
4597 * Most of the issues taken into account to get the return value of
4598 * this function are not trivial. We discuss these issues in the two
4599 * functions providing the main pieces of information needed by this
4600 * function.
4601 */
4602static bool bfq_better_to_idle(struct bfq_queue *bfqq)
4603{
4604 struct bfq_data *bfqd = bfqq->bfqd;
4605 bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
4606
4607 /* No point in idling for bfqq if it won't get requests any longer */
4608 if (unlikely(!bfqq_process_refs(bfqq)))
4609 return false;
4610
4611 if (unlikely(bfqd->strict_guarantees))
4612 return true;
4613
4614 /*
4615 * Idling is performed only if slice_idle > 0. In addition, we
4616 * do not idle if
4617 * (a) bfqq is async
4618 * (b) bfqq is in the idle io prio class: in this case we do
4619 * not idle because we want to minimize the bandwidth that
4620 * queues in this class can steal to higher-priority queues
4621 */
4622 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
4623 bfq_class_idle(bfqq))
4624 return false;
4625
4626 idling_boosts_thr_with_no_issue =
4627 idling_boosts_thr_without_issues(bfqd, bfqq);
4628
4629 idling_needed_for_service_guar =
4630 idling_needed_for_service_guarantees(bfqd, bfqq);
4631
4632 /*
4633 * We have now the two components we need to compute the
4634 * return value of the function, which is true only if idling
4635 * either boosts the throughput (without issues), or is
4636 * necessary to preserve service guarantees.
4637 */
4638 return idling_boosts_thr_with_no_issue ||
4639 idling_needed_for_service_guar;
4640}
4641
4642/*
4643 * If the in-service queue is empty but the function bfq_better_to_idle
4644 * returns true, then:
4645 * 1) the queue must remain in service and cannot be expired, and
4646 * 2) the device must be idled to wait for the possible arrival of a new
4647 * request for the queue.
4648 * See the comments on the function bfq_better_to_idle for the reasons
4649 * why performing device idling is the best choice to boost the throughput
4650 * and preserve service guarantees when bfq_better_to_idle itself
4651 * returns true.
4652 */
4653static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
4654{
4655 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
4656}
4657
4658/*
4659 * This function chooses the queue from which to pick the next extra
4660 * I/O request to inject, if it finds a compatible queue. See the
4661 * comments on bfq_update_inject_limit() for details on the injection
4662 * mechanism, and for the definitions of the quantities mentioned
4663 * below.
4664 */
4665static struct bfq_queue *
4666bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
4667{
4668 struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
4669 unsigned int limit = in_serv_bfqq->inject_limit;
4670 int i;
4671
4672 /*
4673 * If
4674 * - bfqq is not weight-raised and therefore does not carry
4675 * time-critical I/O,
4676 * or
4677 * - regardless of whether bfqq is weight-raised, bfqq has
4678 * however a long think time, during which it can absorb the
4679 * effect of an appropriate number of extra I/O requests
4680 * from other queues (see bfq_update_inject_limit for
4681 * details on the computation of this number);
4682 * then injection can be performed without restrictions.
4683 */
4684 bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
4685 !bfq_bfqq_has_short_ttime(in_serv_bfqq);
4686
4687 /*
4688 * If
4689 * - the baseline total service time could not be sampled yet,
4690 * so the inject limit happens to be still 0, and
4691 * - a lot of time has elapsed since the plugging of I/O
4692 * dispatching started, so drive speed is being wasted
4693 * significantly;
4694 * then temporarily raise inject limit to one request.
4695 */
4696 if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
4697 bfq_bfqq_wait_request(in_serv_bfqq) &&
4698 time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
4699 bfqd->bfq_slice_idle)
4700 )
4701 limit = 1;
4702
4703 if (bfqd->tot_rq_in_driver >= limit)
4704 return NULL;
4705
4706 /*
4707 * Linear search of the source queue for injection; but, with
4708 * a high probability, very few steps are needed to find a
4709 * candidate queue, i.e., a queue with enough budget left for
4710 * its next request. In fact:
4711 * - BFQ dynamically updates the budget of every queue so as
4712 * to accommodate the expected backlog of the queue;
4713 * - if a queue gets all its requests dispatched as injected
4714 * service, then the queue is removed from the active list
4715 * (and re-added only if it gets new requests, but then it
4716 * is assigned again enough budget for its new backlog).
4717 */
4718 for (i = 0; i < bfqd->num_actuators; i++) {
4719 list_for_each_entry(bfqq, &bfqd->active_list[i], bfqq_list)
4720 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4721 (in_serv_always_inject || bfqq->wr_coeff > 1) &&
4722 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4723 bfq_bfqq_budget_left(bfqq)) {
4724 /*
4725 * Allow for only one large in-flight request
4726 * on non-rotational devices, for the
4727 * following reason. On non-rotationl drives,
4728 * large requests take much longer than
4729 * smaller requests to be served. In addition,
4730 * the drive prefers to serve large requests
4731 * w.r.t. to small ones, if it can choose. So,
4732 * having more than one large requests queued
4733 * in the drive may easily make the next first
4734 * request of the in-service queue wait for so
4735 * long to break bfqq's service guarantees. On
4736 * the bright side, large requests let the
4737 * drive reach a very high throughput, even if
4738 * there is only one in-flight large request
4739 * at a time.
4740 */
4741 if (blk_queue_nonrot(bfqd->queue) &&
4742 blk_rq_sectors(bfqq->next_rq) >=
4743 BFQQ_SECT_THR_NONROT &&
4744 bfqd->tot_rq_in_driver >= 1)
4745 continue;
4746 else {
4747 bfqd->rqs_injected = true;
4748 return bfqq;
4749 }
4750 }
4751 }
4752
4753 return NULL;
4754}
4755
4756static struct bfq_queue *
4757bfq_find_active_bfqq_for_actuator(struct bfq_data *bfqd, int idx)
4758{
4759 struct bfq_queue *bfqq;
4760
4761 if (bfqd->in_service_queue &&
4762 bfqd->in_service_queue->actuator_idx == idx)
4763 return bfqd->in_service_queue;
4764
4765 list_for_each_entry(bfqq, &bfqd->active_list[idx], bfqq_list) {
4766 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4767 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4768 bfq_bfqq_budget_left(bfqq)) {
4769 return bfqq;
4770 }
4771 }
4772
4773 return NULL;
4774}
4775
4776/*
4777 * Perform a linear scan of each actuator, until an actuator is found
4778 * for which the following three conditions hold: the load of the
4779 * actuator is below the threshold (see comments on
4780 * actuator_load_threshold for details) and lower than that of the
4781 * next actuator (comments on this extra condition below), and there
4782 * is a queue that contains I/O for that actuator. On success, return
4783 * that queue.
4784 *
4785 * Performing a plain linear scan entails a prioritization among
4786 * actuators. The extra condition above breaks this prioritization and
4787 * tends to distribute injection uniformly across actuators.
4788 */
4789static struct bfq_queue *
4790bfq_find_bfqq_for_underused_actuator(struct bfq_data *bfqd)
4791{
4792 int i;
4793
4794 for (i = 0 ; i < bfqd->num_actuators; i++) {
4795 if (bfqd->rq_in_driver[i] < bfqd->actuator_load_threshold &&
4796 (i == bfqd->num_actuators - 1 ||
4797 bfqd->rq_in_driver[i] < bfqd->rq_in_driver[i+1])) {
4798 struct bfq_queue *bfqq =
4799 bfq_find_active_bfqq_for_actuator(bfqd, i);
4800
4801 if (bfqq)
4802 return bfqq;
4803 }
4804 }
4805
4806 return NULL;
4807}
4808
4809
4810/*
4811 * Select a queue for service. If we have a current queue in service,
4812 * check whether to continue servicing it, or retrieve and set a new one.
4813 */
4814static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
4815{
4816 struct bfq_queue *bfqq, *inject_bfqq;
4817 struct request *next_rq;
4818 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
4819
4820 bfqq = bfqd->in_service_queue;
4821 if (!bfqq)
4822 goto new_queue;
4823
4824 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
4825
4826 /*
4827 * Do not expire bfqq for budget timeout if bfqq may be about
4828 * to enjoy device idling. The reason why, in this case, we
4829 * prevent bfqq from expiring is the same as in the comments
4830 * on the case where bfq_bfqq_must_idle() returns true, in
4831 * bfq_completed_request().
4832 */
4833 if (bfq_may_expire_for_budg_timeout(bfqq) &&
4834 !bfq_bfqq_must_idle(bfqq))
4835 goto expire;
4836
4837check_queue:
4838 /*
4839 * If some actuator is underutilized, but the in-service
4840 * queue does not contain I/O for that actuator, then try to
4841 * inject I/O for that actuator.
4842 */
4843 inject_bfqq = bfq_find_bfqq_for_underused_actuator(bfqd);
4844 if (inject_bfqq && inject_bfqq != bfqq)
4845 return inject_bfqq;
4846
4847 /*
4848 * This loop is rarely executed more than once. Even when it
4849 * happens, it is much more convenient to re-execute this loop
4850 * than to return NULL and trigger a new dispatch to get a
4851 * request served.
4852 */
4853 next_rq = bfqq->next_rq;
4854 /*
4855 * If bfqq has requests queued and it has enough budget left to
4856 * serve them, keep the queue, otherwise expire it.
4857 */
4858 if (next_rq) {
4859 if (bfq_serv_to_charge(next_rq, bfqq) >
4860 bfq_bfqq_budget_left(bfqq)) {
4861 /*
4862 * Expire the queue for budget exhaustion,
4863 * which makes sure that the next budget is
4864 * enough to serve the next request, even if
4865 * it comes from the fifo expired path.
4866 */
4867 reason = BFQQE_BUDGET_EXHAUSTED;
4868 goto expire;
4869 } else {
4870 /*
4871 * The idle timer may be pending because we may
4872 * not disable disk idling even when a new request
4873 * arrives.
4874 */
4875 if (bfq_bfqq_wait_request(bfqq)) {
4876 /*
4877 * If we get here: 1) at least a new request
4878 * has arrived but we have not disabled the
4879 * timer because the request was too small,
4880 * 2) then the block layer has unplugged
4881 * the device, causing the dispatch to be
4882 * invoked.
4883 *
4884 * Since the device is unplugged, now the
4885 * requests are probably large enough to
4886 * provide a reasonable throughput.
4887 * So we disable idling.
4888 */
4889 bfq_clear_bfqq_wait_request(bfqq);
4890 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4891 }
4892 goto keep_queue;
4893 }
4894 }
4895
4896 /*
4897 * No requests pending. However, if the in-service queue is idling
4898 * for a new request, or has requests waiting for a completion and
4899 * may idle after their completion, then keep it anyway.
4900 *
4901 * Yet, inject service from other queues if it boosts
4902 * throughput and is possible.
4903 */
4904 if (bfq_bfqq_wait_request(bfqq) ||
4905 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
4906 unsigned int act_idx = bfqq->actuator_idx;
4907 struct bfq_queue *async_bfqq = NULL;
4908 struct bfq_queue *blocked_bfqq =
4909 !hlist_empty(&bfqq->woken_list) ?
4910 container_of(bfqq->woken_list.first,
4911 struct bfq_queue,
4912 woken_list_node)
4913 : NULL;
4914
4915 if (bfqq->bic && bfqq->bic->bfqq[0][act_idx] &&
4916 bfq_bfqq_busy(bfqq->bic->bfqq[0][act_idx]) &&
4917 bfqq->bic->bfqq[0][act_idx]->next_rq)
4918 async_bfqq = bfqq->bic->bfqq[0][act_idx];
4919 /*
4920 * The next four mutually-exclusive ifs decide
4921 * whether to try injection, and choose the queue to
4922 * pick an I/O request from.
4923 *
4924 * The first if checks whether the process associated
4925 * with bfqq has also async I/O pending. If so, it
4926 * injects such I/O unconditionally. Injecting async
4927 * I/O from the same process can cause no harm to the
4928 * process. On the contrary, it can only increase
4929 * bandwidth and reduce latency for the process.
4930 *
4931 * The second if checks whether there happens to be a
4932 * non-empty waker queue for bfqq, i.e., a queue whose
4933 * I/O needs to be completed for bfqq to receive new
4934 * I/O. This happens, e.g., if bfqq is associated with
4935 * a process that does some sync. A sync generates
4936 * extra blocking I/O, which must be completed before
4937 * the process associated with bfqq can go on with its
4938 * I/O. If the I/O of the waker queue is not served,
4939 * then bfqq remains empty, and no I/O is dispatched,
4940 * until the idle timeout fires for bfqq. This is
4941 * likely to result in lower bandwidth and higher
4942 * latencies for bfqq, and in a severe loss of total
4943 * throughput. The best action to take is therefore to
4944 * serve the waker queue as soon as possible. So do it
4945 * (without relying on the third alternative below for
4946 * eventually serving waker_bfqq's I/O; see the last
4947 * paragraph for further details). This systematic
4948 * injection of I/O from the waker queue does not
4949 * cause any delay to bfqq's I/O. On the contrary,
4950 * next bfqq's I/O is brought forward dramatically,
4951 * for it is not blocked for milliseconds.
4952 *
4953 * The third if checks whether there is a queue woken
4954 * by bfqq, and currently with pending I/O. Such a
4955 * woken queue does not steal bandwidth from bfqq,
4956 * because it remains soon without I/O if bfqq is not
4957 * served. So there is virtually no risk of loss of
4958 * bandwidth for bfqq if this woken queue has I/O
4959 * dispatched while bfqq is waiting for new I/O.
4960 *
4961 * The fourth if checks whether bfqq is a queue for
4962 * which it is better to avoid injection. It is so if
4963 * bfqq delivers more throughput when served without
4964 * any further I/O from other queues in the middle, or
4965 * if the service times of bfqq's I/O requests both
4966 * count more than overall throughput, and may be
4967 * easily increased by injection (this happens if bfqq
4968 * has a short think time). If none of these
4969 * conditions holds, then a candidate queue for
4970 * injection is looked for through
4971 * bfq_choose_bfqq_for_injection(). Note that the
4972 * latter may return NULL (for example if the inject
4973 * limit for bfqq is currently 0).
4974 *
4975 * NOTE: motivation for the second alternative
4976 *
4977 * Thanks to the way the inject limit is updated in
4978 * bfq_update_has_short_ttime(), it is rather likely
4979 * that, if I/O is being plugged for bfqq and the
4980 * waker queue has pending I/O requests that are
4981 * blocking bfqq's I/O, then the fourth alternative
4982 * above lets the waker queue get served before the
4983 * I/O-plugging timeout fires. So one may deem the
4984 * second alternative superfluous. It is not, because
4985 * the fourth alternative may be way less effective in
4986 * case of a synchronization. For two main
4987 * reasons. First, throughput may be low because the
4988 * inject limit may be too low to guarantee the same
4989 * amount of injected I/O, from the waker queue or
4990 * other queues, that the second alternative
4991 * guarantees (the second alternative unconditionally
4992 * injects a pending I/O request of the waker queue
4993 * for each bfq_dispatch_request()). Second, with the
4994 * fourth alternative, the duration of the plugging,
4995 * i.e., the time before bfqq finally receives new I/O,
4996 * may not be minimized, because the waker queue may
4997 * happen to be served only after other queues.
4998 */
4999 if (async_bfqq &&
5000 icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
5001 bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
5002 bfq_bfqq_budget_left(async_bfqq))
5003 bfqq = async_bfqq;
5004 else if (bfqq->waker_bfqq &&
5005 bfq_bfqq_busy(bfqq->waker_bfqq) &&
5006 bfqq->waker_bfqq->next_rq &&
5007 bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
5008 bfqq->waker_bfqq) <=
5009 bfq_bfqq_budget_left(bfqq->waker_bfqq)
5010 )
5011 bfqq = bfqq->waker_bfqq;
5012 else if (blocked_bfqq &&
5013 bfq_bfqq_busy(blocked_bfqq) &&
5014 blocked_bfqq->next_rq &&
5015 bfq_serv_to_charge(blocked_bfqq->next_rq,
5016 blocked_bfqq) <=
5017 bfq_bfqq_budget_left(blocked_bfqq)
5018 )
5019 bfqq = blocked_bfqq;
5020 else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
5021 (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
5022 !bfq_bfqq_has_short_ttime(bfqq)))
5023 bfqq = bfq_choose_bfqq_for_injection(bfqd);
5024 else
5025 bfqq = NULL;
5026
5027 goto keep_queue;
5028 }
5029
5030 reason = BFQQE_NO_MORE_REQUESTS;
5031expire:
5032 bfq_bfqq_expire(bfqd, bfqq, false, reason);
5033new_queue:
5034 bfqq = bfq_set_in_service_queue(bfqd);
5035 if (bfqq) {
5036 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
5037 goto check_queue;
5038 }
5039keep_queue:
5040 if (bfqq)
5041 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
5042 else
5043 bfq_log(bfqd, "select_queue: no queue returned");
5044
5045 return bfqq;
5046}
5047
5048static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
5049{
5050 struct bfq_entity *entity = &bfqq->entity;
5051
5052 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
5053 bfq_log_bfqq(bfqd, bfqq,
5054 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
5055 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
5056 jiffies_to_msecs(bfqq->wr_cur_max_time),
5057 bfqq->wr_coeff,
5058 bfqq->entity.weight, bfqq->entity.orig_weight);
5059
5060 if (entity->prio_changed)
5061 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
5062
5063 /*
5064 * If the queue was activated in a burst, or too much
5065 * time has elapsed from the beginning of this
5066 * weight-raising period, then end weight raising.
5067 */
5068 if (bfq_bfqq_in_large_burst(bfqq))
5069 bfq_bfqq_end_wr(bfqq);
5070 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
5071 bfqq->wr_cur_max_time)) {
5072 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
5073 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
5074 bfq_wr_duration(bfqd))) {
5075 /*
5076 * Either in interactive weight
5077 * raising, or in soft_rt weight
5078 * raising with the
5079 * interactive-weight-raising period
5080 * elapsed (so no switch back to
5081 * interactive weight raising).
5082 */
5083 bfq_bfqq_end_wr(bfqq);
5084 } else { /*
5085 * soft_rt finishing while still in
5086 * interactive period, switch back to
5087 * interactive weight raising
5088 */
5089 switch_back_to_interactive_wr(bfqq, bfqd);
5090 bfqq->entity.prio_changed = 1;
5091 }
5092 }
5093 if (bfqq->wr_coeff > 1 &&
5094 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
5095 bfqq->service_from_wr > max_service_from_wr) {
5096 /* see comments on max_service_from_wr */
5097 bfq_bfqq_end_wr(bfqq);
5098 }
5099 }
5100 /*
5101 * To improve latency (for this or other queues), immediately
5102 * update weight both if it must be raised and if it must be
5103 * lowered. Since, entity may be on some active tree here, and
5104 * might have a pending change of its ioprio class, invoke
5105 * next function with the last parameter unset (see the
5106 * comments on the function).
5107 */
5108 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
5109 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
5110 entity, false);
5111}
5112
5113/*
5114 * Dispatch next request from bfqq.
5115 */
5116static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
5117 struct bfq_queue *bfqq)
5118{
5119 struct request *rq = bfqq->next_rq;
5120 unsigned long service_to_charge;
5121
5122 service_to_charge = bfq_serv_to_charge(rq, bfqq);
5123
5124 bfq_bfqq_served(bfqq, service_to_charge);
5125
5126 if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
5127 bfqd->wait_dispatch = false;
5128 bfqd->waited_rq = rq;
5129 }
5130
5131 bfq_dispatch_remove(bfqd->queue, rq);
5132
5133 if (bfqq != bfqd->in_service_queue)
5134 return rq;
5135
5136 /*
5137 * If weight raising has to terminate for bfqq, then next
5138 * function causes an immediate update of bfqq's weight,
5139 * without waiting for next activation. As a consequence, on
5140 * expiration, bfqq will be timestamped as if has never been
5141 * weight-raised during this service slot, even if it has
5142 * received part or even most of the service as a
5143 * weight-raised queue. This inflates bfqq's timestamps, which
5144 * is beneficial, as bfqq is then more willing to leave the
5145 * device immediately to possible other weight-raised queues.
5146 */
5147 bfq_update_wr_data(bfqd, bfqq);
5148
5149 /*
5150 * Expire bfqq, pretending that its budget expired, if bfqq
5151 * belongs to CLASS_IDLE and other queues are waiting for
5152 * service.
5153 */
5154 if (bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq))
5155 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
5156
5157 return rq;
5158}
5159
5160static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
5161{
5162 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5163
5164 /*
5165 * Avoiding lock: a race on bfqd->queued should cause at
5166 * most a call to dispatch for nothing
5167 */
5168 return !list_empty_careful(&bfqd->dispatch) ||
5169 READ_ONCE(bfqd->queued);
5170}
5171
5172static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5173{
5174 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5175 struct request *rq = NULL;
5176 struct bfq_queue *bfqq = NULL;
5177
5178 if (!list_empty(&bfqd->dispatch)) {
5179 rq = list_first_entry(&bfqd->dispatch, struct request,
5180 queuelist);
5181 list_del_init(&rq->queuelist);
5182
5183 bfqq = RQ_BFQQ(rq);
5184
5185 if (bfqq) {
5186 /*
5187 * Increment counters here, because this
5188 * dispatch does not follow the standard
5189 * dispatch flow (where counters are
5190 * incremented)
5191 */
5192 bfqq->dispatched++;
5193
5194 goto inc_in_driver_start_rq;
5195 }
5196
5197 /*
5198 * We exploit the bfq_finish_requeue_request hook to
5199 * decrement tot_rq_in_driver, but
5200 * bfq_finish_requeue_request will not be invoked on
5201 * this request. So, to avoid unbalance, just start
5202 * this request, without incrementing tot_rq_in_driver. As
5203 * a negative consequence, tot_rq_in_driver is deceptively
5204 * lower than it should be while this request is in
5205 * service. This may cause bfq_schedule_dispatch to be
5206 * invoked uselessly.
5207 *
5208 * As for implementing an exact solution, the
5209 * bfq_finish_requeue_request hook, if defined, is
5210 * probably invoked also on this request. So, by
5211 * exploiting this hook, we could 1) increment
5212 * tot_rq_in_driver here, and 2) decrement it in
5213 * bfq_finish_requeue_request. Such a solution would
5214 * let the value of the counter be always accurate,
5215 * but it would entail using an extra interface
5216 * function. This cost seems higher than the benefit,
5217 * being the frequency of non-elevator-private
5218 * requests very low.
5219 */
5220 goto start_rq;
5221 }
5222
5223 bfq_log(bfqd, "dispatch requests: %d busy queues",
5224 bfq_tot_busy_queues(bfqd));
5225
5226 if (bfq_tot_busy_queues(bfqd) == 0)
5227 goto exit;
5228
5229 /*
5230 * Force device to serve one request at a time if
5231 * strict_guarantees is true. Forcing this service scheme is
5232 * currently the ONLY way to guarantee that the request
5233 * service order enforced by the scheduler is respected by a
5234 * queueing device. Otherwise the device is free even to make
5235 * some unlucky request wait for as long as the device
5236 * wishes.
5237 *
5238 * Of course, serving one request at a time may cause loss of
5239 * throughput.
5240 */
5241 if (bfqd->strict_guarantees && bfqd->tot_rq_in_driver > 0)
5242 goto exit;
5243
5244 bfqq = bfq_select_queue(bfqd);
5245 if (!bfqq)
5246 goto exit;
5247
5248 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
5249
5250 if (rq) {
5251inc_in_driver_start_rq:
5252 bfqd->rq_in_driver[bfqq->actuator_idx]++;
5253 bfqd->tot_rq_in_driver++;
5254start_rq:
5255 rq->rq_flags |= RQF_STARTED;
5256 }
5257exit:
5258 return rq;
5259}
5260
5261#ifdef CONFIG_BFQ_CGROUP_DEBUG
5262static void bfq_update_dispatch_stats(struct request_queue *q,
5263 struct request *rq,
5264 struct bfq_queue *in_serv_queue,
5265 bool idle_timer_disabled)
5266{
5267 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
5268
5269 if (!idle_timer_disabled && !bfqq)
5270 return;
5271
5272 /*
5273 * rq and bfqq are guaranteed to exist until this function
5274 * ends, for the following reasons. First, rq can be
5275 * dispatched to the device, and then can be completed and
5276 * freed, only after this function ends. Second, rq cannot be
5277 * merged (and thus freed because of a merge) any longer,
5278 * because it has already started. Thus rq cannot be freed
5279 * before this function ends, and, since rq has a reference to
5280 * bfqq, the same guarantee holds for bfqq too.
5281 *
5282 * In addition, the following queue lock guarantees that
5283 * bfqq_group(bfqq) exists as well.
5284 */
5285 spin_lock_irq(&q->queue_lock);
5286 if (idle_timer_disabled)
5287 /*
5288 * Since the idle timer has been disabled,
5289 * in_serv_queue contained some request when
5290 * __bfq_dispatch_request was invoked above, which
5291 * implies that rq was picked exactly from
5292 * in_serv_queue. Thus in_serv_queue == bfqq, and is
5293 * therefore guaranteed to exist because of the above
5294 * arguments.
5295 */
5296 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
5297 if (bfqq) {
5298 struct bfq_group *bfqg = bfqq_group(bfqq);
5299
5300 bfqg_stats_update_avg_queue_size(bfqg);
5301 bfqg_stats_set_start_empty_time(bfqg);
5302 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
5303 }
5304 spin_unlock_irq(&q->queue_lock);
5305}
5306#else
5307static inline void bfq_update_dispatch_stats(struct request_queue *q,
5308 struct request *rq,
5309 struct bfq_queue *in_serv_queue,
5310 bool idle_timer_disabled) {}
5311#endif /* CONFIG_BFQ_CGROUP_DEBUG */
5312
5313static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5314{
5315 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5316 struct request *rq;
5317 struct bfq_queue *in_serv_queue;
5318 bool waiting_rq, idle_timer_disabled = false;
5319
5320 spin_lock_irq(&bfqd->lock);
5321
5322 in_serv_queue = bfqd->in_service_queue;
5323 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
5324
5325 rq = __bfq_dispatch_request(hctx);
5326 if (in_serv_queue == bfqd->in_service_queue) {
5327 idle_timer_disabled =
5328 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
5329 }
5330
5331 spin_unlock_irq(&bfqd->lock);
5332 bfq_update_dispatch_stats(hctx->queue, rq,
5333 idle_timer_disabled ? in_serv_queue : NULL,
5334 idle_timer_disabled);
5335
5336 return rq;
5337}
5338
5339/*
5340 * Task holds one reference to the queue, dropped when task exits. Each rq
5341 * in-flight on this queue also holds a reference, dropped when rq is freed.
5342 *
5343 * Scheduler lock must be held here. Recall not to use bfqq after calling
5344 * this function on it.
5345 */
5346void bfq_put_queue(struct bfq_queue *bfqq)
5347{
5348 struct bfq_queue *item;
5349 struct hlist_node *n;
5350 struct bfq_group *bfqg = bfqq_group(bfqq);
5351
5352 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d", bfqq, bfqq->ref);
5353
5354 bfqq->ref--;
5355 if (bfqq->ref)
5356 return;
5357
5358 if (!hlist_unhashed(&bfqq->burst_list_node)) {
5359 hlist_del_init(&bfqq->burst_list_node);
5360 /*
5361 * Decrement also burst size after the removal, if the
5362 * process associated with bfqq is exiting, and thus
5363 * does not contribute to the burst any longer. This
5364 * decrement helps filter out false positives of large
5365 * bursts, when some short-lived process (often due to
5366 * the execution of commands by some service) happens
5367 * to start and exit while a complex application is
5368 * starting, and thus spawning several processes that
5369 * do I/O (and that *must not* be treated as a large
5370 * burst, see comments on bfq_handle_burst).
5371 *
5372 * In particular, the decrement is performed only if:
5373 * 1) bfqq is not a merged queue, because, if it is,
5374 * then this free of bfqq is not triggered by the exit
5375 * of the process bfqq is associated with, but exactly
5376 * by the fact that bfqq has just been merged.
5377 * 2) burst_size is greater than 0, to handle
5378 * unbalanced decrements. Unbalanced decrements may
5379 * happen in te following case: bfqq is inserted into
5380 * the current burst list--without incrementing
5381 * bust_size--because of a split, but the current
5382 * burst list is not the burst list bfqq belonged to
5383 * (see comments on the case of a split in
5384 * bfq_set_request).
5385 */
5386 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
5387 bfqq->bfqd->burst_size--;
5388 }
5389
5390 /*
5391 * bfqq does not exist any longer, so it cannot be woken by
5392 * any other queue, and cannot wake any other queue. Then bfqq
5393 * must be removed from the woken list of its possible waker
5394 * queue, and all queues in the woken list of bfqq must stop
5395 * having a waker queue. Strictly speaking, these updates
5396 * should be performed when bfqq remains with no I/O source
5397 * attached to it, which happens before bfqq gets freed. In
5398 * particular, this happens when the last process associated
5399 * with bfqq exits or gets associated with a different
5400 * queue. However, both events lead to bfqq being freed soon,
5401 * and dangling references would come out only after bfqq gets
5402 * freed. So these updates are done here, as a simple and safe
5403 * way to handle all cases.
5404 */
5405 /* remove bfqq from woken list */
5406 if (!hlist_unhashed(&bfqq->woken_list_node))
5407 hlist_del_init(&bfqq->woken_list_node);
5408
5409 /* reset waker for all queues in woken list */
5410 hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
5411 woken_list_node) {
5412 item->waker_bfqq = NULL;
5413 hlist_del_init(&item->woken_list_node);
5414 }
5415
5416 if (bfqq->bfqd->last_completed_rq_bfqq == bfqq)
5417 bfqq->bfqd->last_completed_rq_bfqq = NULL;
5418
5419 WARN_ON_ONCE(!list_empty(&bfqq->fifo));
5420 WARN_ON_ONCE(!RB_EMPTY_ROOT(&bfqq->sort_list));
5421 WARN_ON_ONCE(bfqq->dispatched);
5422
5423 kmem_cache_free(bfq_pool, bfqq);
5424 bfqg_and_blkg_put(bfqg);
5425}
5426
5427static void bfq_put_stable_ref(struct bfq_queue *bfqq)
5428{
5429 bfqq->stable_ref--;
5430 bfq_put_queue(bfqq);
5431}
5432
5433void bfq_put_cooperator(struct bfq_queue *bfqq)
5434{
5435 struct bfq_queue *__bfqq, *next;
5436
5437 /*
5438 * If this queue was scheduled to merge with another queue, be
5439 * sure to drop the reference taken on that queue (and others in
5440 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
5441 */
5442 __bfqq = bfqq->new_bfqq;
5443 while (__bfqq) {
5444 next = __bfqq->new_bfqq;
5445 bfq_put_queue(__bfqq);
5446 __bfqq = next;
5447 }
5448}
5449
5450static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
5451{
5452 if (bfqq == bfqd->in_service_queue) {
5453 __bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT);
5454 bfq_schedule_dispatch(bfqd);
5455 }
5456
5457 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
5458
5459 bfq_put_cooperator(bfqq);
5460
5461 bfq_release_process_ref(bfqd, bfqq);
5462}
5463
5464static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync,
5465 unsigned int actuator_idx)
5466{
5467 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync, actuator_idx);
5468 struct bfq_data *bfqd;
5469
5470 if (bfqq)
5471 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
5472
5473 if (bfqq && bfqd) {
5474 bic_set_bfqq(bic, NULL, is_sync, actuator_idx);
5475 bfq_exit_bfqq(bfqd, bfqq);
5476 }
5477}
5478
5479static void _bfq_exit_icq(struct bfq_io_cq *bic, unsigned int num_actuators)
5480{
5481 struct bfq_iocq_bfqq_data *bfqq_data = bic->bfqq_data;
5482 unsigned int act_idx;
5483
5484 for (act_idx = 0; act_idx < num_actuators; act_idx++) {
5485 if (bfqq_data[act_idx].stable_merge_bfqq)
5486 bfq_put_stable_ref(bfqq_data[act_idx].stable_merge_bfqq);
5487
5488 bfq_exit_icq_bfqq(bic, true, act_idx);
5489 bfq_exit_icq_bfqq(bic, false, act_idx);
5490 }
5491}
5492
5493static void bfq_exit_icq(struct io_cq *icq)
5494{
5495 struct bfq_io_cq *bic = icq_to_bic(icq);
5496 struct bfq_data *bfqd = bic_to_bfqd(bic);
5497 unsigned long flags;
5498
5499 /*
5500 * If bfqd and thus bfqd->num_actuators is not available any
5501 * longer, then cycle over all possible per-actuator bfqqs in
5502 * next loop. We rely on bic being zeroed on creation, and
5503 * therefore on its unused per-actuator fields being NULL.
5504 *
5505 * bfqd is NULL if scheduler already exited, and in that case
5506 * this is the last time these queues are accessed.
5507 */
5508 if (bfqd) {
5509 spin_lock_irqsave(&bfqd->lock, flags);
5510 _bfq_exit_icq(bic, bfqd->num_actuators);
5511 spin_unlock_irqrestore(&bfqd->lock, flags);
5512 } else {
5513 _bfq_exit_icq(bic, BFQ_MAX_ACTUATORS);
5514 }
5515}
5516
5517/*
5518 * Update the entity prio values; note that the new values will not
5519 * be used until the next (re)activation.
5520 */
5521static void
5522bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
5523{
5524 struct task_struct *tsk = current;
5525 int ioprio_class;
5526 struct bfq_data *bfqd = bfqq->bfqd;
5527
5528 if (!bfqd)
5529 return;
5530
5531 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5532 switch (ioprio_class) {
5533 default:
5534 pr_err("bdi %s: bfq: bad prio class %d\n",
5535 bdi_dev_name(bfqq->bfqd->queue->disk->bdi),
5536 ioprio_class);
5537 fallthrough;
5538 case IOPRIO_CLASS_NONE:
5539 /*
5540 * No prio set, inherit CPU scheduling settings.
5541 */
5542 bfqq->new_ioprio = task_nice_ioprio(tsk);
5543 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
5544 break;
5545 case IOPRIO_CLASS_RT:
5546 bfqq->new_ioprio = IOPRIO_PRIO_LEVEL(bic->ioprio);
5547 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
5548 break;
5549 case IOPRIO_CLASS_BE:
5550 bfqq->new_ioprio = IOPRIO_PRIO_LEVEL(bic->ioprio);
5551 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
5552 break;
5553 case IOPRIO_CLASS_IDLE:
5554 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
5555 bfqq->new_ioprio = IOPRIO_NR_LEVELS - 1;
5556 break;
5557 }
5558
5559 if (bfqq->new_ioprio >= IOPRIO_NR_LEVELS) {
5560 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
5561 bfqq->new_ioprio);
5562 bfqq->new_ioprio = IOPRIO_NR_LEVELS - 1;
5563 }
5564
5565 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
5566 bfq_log_bfqq(bfqd, bfqq, "new_ioprio %d new_weight %d",
5567 bfqq->new_ioprio, bfqq->entity.new_weight);
5568 bfqq->entity.prio_changed = 1;
5569}
5570
5571static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5572 struct bio *bio, bool is_sync,
5573 struct bfq_io_cq *bic,
5574 bool respawn);
5575
5576static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
5577{
5578 struct bfq_data *bfqd = bic_to_bfqd(bic);
5579 struct bfq_queue *bfqq;
5580 int ioprio = bic->icq.ioc->ioprio;
5581
5582 /*
5583 * This condition may trigger on a newly created bic, be sure to
5584 * drop the lock before returning.
5585 */
5586 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
5587 return;
5588
5589 bic->ioprio = ioprio;
5590
5591 bfqq = bic_to_bfqq(bic, false, bfq_actuator_index(bfqd, bio));
5592 if (bfqq) {
5593 struct bfq_queue *old_bfqq = bfqq;
5594
5595 bfqq = bfq_get_queue(bfqd, bio, false, bic, true);
5596 bic_set_bfqq(bic, bfqq, false, bfq_actuator_index(bfqd, bio));
5597 bfq_release_process_ref(bfqd, old_bfqq);
5598 }
5599
5600 bfqq = bic_to_bfqq(bic, true, bfq_actuator_index(bfqd, bio));
5601 if (bfqq)
5602 bfq_set_next_ioprio_data(bfqq, bic);
5603}
5604
5605static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5606 struct bfq_io_cq *bic, pid_t pid, int is_sync,
5607 unsigned int act_idx)
5608{
5609 u64 now_ns = blk_time_get_ns();
5610
5611 bfqq->actuator_idx = act_idx;
5612 RB_CLEAR_NODE(&bfqq->entity.rb_node);
5613 INIT_LIST_HEAD(&bfqq->fifo);
5614 INIT_HLIST_NODE(&bfqq->burst_list_node);
5615 INIT_HLIST_NODE(&bfqq->woken_list_node);
5616 INIT_HLIST_HEAD(&bfqq->woken_list);
5617
5618 bfqq->ref = 0;
5619 bfqq->bfqd = bfqd;
5620
5621 if (bic)
5622 bfq_set_next_ioprio_data(bfqq, bic);
5623
5624 if (is_sync) {
5625 /*
5626 * No need to mark as has_short_ttime if in
5627 * idle_class, because no device idling is performed
5628 * for queues in idle class
5629 */
5630 if (!bfq_class_idle(bfqq))
5631 /* tentatively mark as has_short_ttime */
5632 bfq_mark_bfqq_has_short_ttime(bfqq);
5633 bfq_mark_bfqq_sync(bfqq);
5634 bfq_mark_bfqq_just_created(bfqq);
5635 } else
5636 bfq_clear_bfqq_sync(bfqq);
5637
5638 /* set end request to minus infinity from now */
5639 bfqq->ttime.last_end_request = now_ns + 1;
5640
5641 bfqq->creation_time = jiffies;
5642
5643 bfqq->io_start_time = now_ns;
5644
5645 bfq_mark_bfqq_IO_bound(bfqq);
5646
5647 bfqq->pid = pid;
5648
5649 /* Tentative initial value to trade off between thr and lat */
5650 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
5651 bfqq->budget_timeout = bfq_smallest_from_now();
5652
5653 bfqq->wr_coeff = 1;
5654 bfqq->last_wr_start_finish = jiffies;
5655 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
5656 bfqq->split_time = bfq_smallest_from_now();
5657
5658 /*
5659 * To not forget the possibly high bandwidth consumed by a
5660 * process/queue in the recent past,
5661 * bfq_bfqq_softrt_next_start() returns a value at least equal
5662 * to the current value of bfqq->soft_rt_next_start (see
5663 * comments on bfq_bfqq_softrt_next_start). Set
5664 * soft_rt_next_start to now, to mean that bfqq has consumed
5665 * no bandwidth so far.
5666 */
5667 bfqq->soft_rt_next_start = jiffies;
5668
5669 /* first request is almost certainly seeky */
5670 bfqq->seek_history = 1;
5671
5672 bfqq->decrease_time_jif = jiffies;
5673}
5674
5675static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
5676 struct bfq_group *bfqg,
5677 int ioprio_class, int ioprio, int act_idx)
5678{
5679 switch (ioprio_class) {
5680 case IOPRIO_CLASS_RT:
5681 return &bfqg->async_bfqq[0][ioprio][act_idx];
5682 case IOPRIO_CLASS_NONE:
5683 ioprio = IOPRIO_BE_NORM;
5684 fallthrough;
5685 case IOPRIO_CLASS_BE:
5686 return &bfqg->async_bfqq[1][ioprio][act_idx];
5687 case IOPRIO_CLASS_IDLE:
5688 return &bfqg->async_idle_bfqq[act_idx];
5689 default:
5690 return NULL;
5691 }
5692}
5693
5694static struct bfq_queue *
5695bfq_do_early_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5696 struct bfq_io_cq *bic,
5697 struct bfq_queue *last_bfqq_created)
5698{
5699 unsigned int a_idx = last_bfqq_created->actuator_idx;
5700 struct bfq_queue *new_bfqq =
5701 bfq_setup_merge(bfqq, last_bfqq_created);
5702
5703 if (!new_bfqq)
5704 return bfqq;
5705
5706 if (new_bfqq->bic)
5707 new_bfqq->bic->bfqq_data[a_idx].stably_merged = true;
5708 bic->bfqq_data[a_idx].stably_merged = true;
5709
5710 /*
5711 * Reusing merge functions. This implies that
5712 * bfqq->bic must be set too, for
5713 * bfq_merge_bfqqs to correctly save bfqq's
5714 * state before killing it.
5715 */
5716 bfqq->bic = bic;
5717 return bfq_merge_bfqqs(bfqd, bic, bfqq);
5718}
5719
5720/*
5721 * Many throughput-sensitive workloads are made of several parallel
5722 * I/O flows, with all flows generated by the same application, or
5723 * more generically by the same task (e.g., system boot). The most
5724 * counterproductive action with these workloads is plugging I/O
5725 * dispatch when one of the bfq_queues associated with these flows
5726 * remains temporarily empty.
5727 *
5728 * To avoid this plugging, BFQ has been using a burst-handling
5729 * mechanism for years now. This mechanism has proven effective for
5730 * throughput, and not detrimental for service guarantees. The
5731 * following function pushes this mechanism a little bit further,
5732 * basing on the following two facts.
5733 *
5734 * First, all the I/O flows of a the same application or task
5735 * contribute to the execution/completion of that common application
5736 * or task. So the performance figures that matter are total
5737 * throughput of the flows and task-wide I/O latency. In particular,
5738 * these flows do not need to be protected from each other, in terms
5739 * of individual bandwidth or latency.
5740 *
5741 * Second, the above fact holds regardless of the number of flows.
5742 *
5743 * Putting these two facts together, this commits merges stably the
5744 * bfq_queues associated with these I/O flows, i.e., with the
5745 * processes that generate these IO/ flows, regardless of how many the
5746 * involved processes are.
5747 *
5748 * To decide whether a set of bfq_queues is actually associated with
5749 * the I/O flows of a common application or task, and to merge these
5750 * queues stably, this function operates as follows: given a bfq_queue,
5751 * say Q2, currently being created, and the last bfq_queue, say Q1,
5752 * created before Q2, Q2 is merged stably with Q1 if
5753 * - very little time has elapsed since when Q1 was created
5754 * - Q2 has the same ioprio as Q1
5755 * - Q2 belongs to the same group as Q1
5756 *
5757 * Merging bfq_queues also reduces scheduling overhead. A fio test
5758 * with ten random readers on /dev/nullb shows a throughput boost of
5759 * 40%, with a quadcore. Since BFQ's execution time amounts to ~50% of
5760 * the total per-request processing time, the above throughput boost
5761 * implies that BFQ's overhead is reduced by more than 50%.
5762 *
5763 * This new mechanism most certainly obsoletes the current
5764 * burst-handling heuristics. We keep those heuristics for the moment.
5765 */
5766static struct bfq_queue *bfq_do_or_sched_stable_merge(struct bfq_data *bfqd,
5767 struct bfq_queue *bfqq,
5768 struct bfq_io_cq *bic)
5769{
5770 struct bfq_queue **source_bfqq = bfqq->entity.parent ?
5771 &bfqq->entity.parent->last_bfqq_created :
5772 &bfqd->last_bfqq_created;
5773
5774 struct bfq_queue *last_bfqq_created = *source_bfqq;
5775
5776 /*
5777 * If last_bfqq_created has not been set yet, then init it. If
5778 * it has been set already, but too long ago, then move it
5779 * forward to bfqq. Finally, move also if bfqq belongs to a
5780 * different group than last_bfqq_created, or if bfqq has a
5781 * different ioprio, ioprio_class or actuator_idx. If none of
5782 * these conditions holds true, then try an early stable merge
5783 * or schedule a delayed stable merge. As for the condition on
5784 * actuator_idx, the reason is that, if queues associated with
5785 * different actuators are merged, then control is lost on
5786 * each actuator. Therefore some actuator may be
5787 * underutilized, and throughput may decrease.
5788 *
5789 * A delayed merge is scheduled (instead of performing an
5790 * early merge), in case bfqq might soon prove to be more
5791 * throughput-beneficial if not merged. Currently this is
5792 * possible only if bfqd is rotational with no queueing. For
5793 * such a drive, not merging bfqq is better for throughput if
5794 * bfqq happens to contain sequential I/O. So, we wait a
5795 * little bit for enough I/O to flow through bfqq. After that,
5796 * if such an I/O is sequential, then the merge is
5797 * canceled. Otherwise the merge is finally performed.
5798 */
5799 if (!last_bfqq_created ||
5800 time_before(last_bfqq_created->creation_time +
5801 msecs_to_jiffies(bfq_activation_stable_merging),
5802 bfqq->creation_time) ||
5803 bfqq->entity.parent != last_bfqq_created->entity.parent ||
5804 bfqq->ioprio != last_bfqq_created->ioprio ||
5805 bfqq->ioprio_class != last_bfqq_created->ioprio_class ||
5806 bfqq->actuator_idx != last_bfqq_created->actuator_idx)
5807 *source_bfqq = bfqq;
5808 else if (time_after_eq(last_bfqq_created->creation_time +
5809 bfqd->bfq_burst_interval,
5810 bfqq->creation_time)) {
5811 if (likely(bfqd->nonrot_with_queueing))
5812 /*
5813 * With this type of drive, leaving
5814 * bfqq alone may provide no
5815 * throughput benefits compared with
5816 * merging bfqq. So merge bfqq now.
5817 */
5818 bfqq = bfq_do_early_stable_merge(bfqd, bfqq,
5819 bic,
5820 last_bfqq_created);
5821 else { /* schedule tentative stable merge */
5822 /*
5823 * get reference on last_bfqq_created,
5824 * to prevent it from being freed,
5825 * until we decide whether to merge
5826 */
5827 last_bfqq_created->ref++;
5828 /*
5829 * need to keep track of stable refs, to
5830 * compute process refs correctly
5831 */
5832 last_bfqq_created->stable_ref++;
5833 /*
5834 * Record the bfqq to merge to.
5835 */
5836 bic->bfqq_data[last_bfqq_created->actuator_idx].stable_merge_bfqq =
5837 last_bfqq_created;
5838 }
5839 }
5840
5841 return bfqq;
5842}
5843
5844
5845static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5846 struct bio *bio, bool is_sync,
5847 struct bfq_io_cq *bic,
5848 bool respawn)
5849{
5850 const int ioprio = IOPRIO_PRIO_LEVEL(bic->ioprio);
5851 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5852 struct bfq_queue **async_bfqq = NULL;
5853 struct bfq_queue *bfqq;
5854 struct bfq_group *bfqg;
5855
5856 bfqg = bfq_bio_bfqg(bfqd, bio);
5857 if (!is_sync) {
5858 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
5859 ioprio,
5860 bfq_actuator_index(bfqd, bio));
5861 bfqq = *async_bfqq;
5862 if (bfqq)
5863 goto out;
5864 }
5865
5866 bfqq = kmem_cache_alloc_node(bfq_pool,
5867 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
5868 bfqd->queue->node);
5869
5870 if (bfqq) {
5871 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
5872 is_sync, bfq_actuator_index(bfqd, bio));
5873 bfq_init_entity(&bfqq->entity, bfqg);
5874 bfq_log_bfqq(bfqd, bfqq, "allocated");
5875 } else {
5876 bfqq = &bfqd->oom_bfqq;
5877 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
5878 goto out;
5879 }
5880
5881 /*
5882 * Pin the queue now that it's allocated, scheduler exit will
5883 * prune it.
5884 */
5885 if (async_bfqq) {
5886 bfqq->ref++; /*
5887 * Extra group reference, w.r.t. sync
5888 * queue. This extra reference is removed
5889 * only if bfqq->bfqg disappears, to
5890 * guarantee that this queue is not freed
5891 * until its group goes away.
5892 */
5893 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
5894 bfqq, bfqq->ref);
5895 *async_bfqq = bfqq;
5896 }
5897
5898out:
5899 bfqq->ref++; /* get a process reference to this queue */
5900
5901 if (bfqq != &bfqd->oom_bfqq && is_sync && !respawn)
5902 bfqq = bfq_do_or_sched_stable_merge(bfqd, bfqq, bic);
5903 return bfqq;
5904}
5905
5906static void bfq_update_io_thinktime(struct bfq_data *bfqd,
5907 struct bfq_queue *bfqq)
5908{
5909 struct bfq_ttime *ttime = &bfqq->ttime;
5910 u64 elapsed;
5911
5912 /*
5913 * We are really interested in how long it takes for the queue to
5914 * become busy when there is no outstanding IO for this queue. So
5915 * ignore cases when the bfq queue has already IO queued.
5916 */
5917 if (bfqq->dispatched || bfq_bfqq_busy(bfqq))
5918 return;
5919 elapsed = blk_time_get_ns() - bfqq->ttime.last_end_request;
5920 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
5921
5922 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
5923 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
5924 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
5925 ttime->ttime_samples);
5926}
5927
5928static void
5929bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5930 struct request *rq)
5931{
5932 bfqq->seek_history <<= 1;
5933 bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
5934
5935 if (bfqq->wr_coeff > 1 &&
5936 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
5937 BFQQ_TOTALLY_SEEKY(bfqq)) {
5938 if (time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
5939 bfq_wr_duration(bfqd))) {
5940 /*
5941 * In soft_rt weight raising with the
5942 * interactive-weight-raising period
5943 * elapsed (so no switch back to
5944 * interactive weight raising).
5945 */
5946 bfq_bfqq_end_wr(bfqq);
5947 } else { /*
5948 * stopping soft_rt weight raising
5949 * while still in interactive period,
5950 * switch back to interactive weight
5951 * raising
5952 */
5953 switch_back_to_interactive_wr(bfqq, bfqd);
5954 bfqq->entity.prio_changed = 1;
5955 }
5956 }
5957}
5958
5959static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
5960 struct bfq_queue *bfqq,
5961 struct bfq_io_cq *bic)
5962{
5963 bool has_short_ttime = true, state_changed;
5964
5965 /*
5966 * No need to update has_short_ttime if bfqq is async or in
5967 * idle io prio class, or if bfq_slice_idle is zero, because
5968 * no device idling is performed for bfqq in this case.
5969 */
5970 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
5971 bfqd->bfq_slice_idle == 0)
5972 return;
5973
5974 /* Idle window just restored, statistics are meaningless. */
5975 if (time_is_after_eq_jiffies(bfqq->split_time +
5976 bfqd->bfq_wr_min_idle_time))
5977 return;
5978
5979 /* Think time is infinite if no process is linked to
5980 * bfqq. Otherwise check average think time to decide whether
5981 * to mark as has_short_ttime. To this goal, compare average
5982 * think time with half the I/O-plugging timeout.
5983 */
5984 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
5985 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
5986 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle>>1))
5987 has_short_ttime = false;
5988
5989 state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
5990
5991 if (has_short_ttime)
5992 bfq_mark_bfqq_has_short_ttime(bfqq);
5993 else
5994 bfq_clear_bfqq_has_short_ttime(bfqq);
5995
5996 /*
5997 * Until the base value for the total service time gets
5998 * finally computed for bfqq, the inject limit does depend on
5999 * the think-time state (short|long). In particular, the limit
6000 * is 0 or 1 if the think time is deemed, respectively, as
6001 * short or long (details in the comments in
6002 * bfq_update_inject_limit()). Accordingly, the next
6003 * instructions reset the inject limit if the think-time state
6004 * has changed and the above base value is still to be
6005 * computed.
6006 *
6007 * However, the reset is performed only if more than 100 ms
6008 * have elapsed since the last update of the inject limit, or
6009 * (inclusive) if the change is from short to long think
6010 * time. The reason for this waiting is as follows.
6011 *
6012 * bfqq may have a long think time because of a
6013 * synchronization with some other queue, i.e., because the
6014 * I/O of some other queue may need to be completed for bfqq
6015 * to receive new I/O. Details in the comments on the choice
6016 * of the queue for injection in bfq_select_queue().
6017 *
6018 * As stressed in those comments, if such a synchronization is
6019 * actually in place, then, without injection on bfqq, the
6020 * blocking I/O cannot happen to served while bfqq is in
6021 * service. As a consequence, if bfqq is granted
6022 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
6023 * is dispatched, until the idle timeout fires. This is likely
6024 * to result in lower bandwidth and higher latencies for bfqq,
6025 * and in a severe loss of total throughput.
6026 *
6027 * On the opposite end, a non-zero inject limit may allow the
6028 * I/O that blocks bfqq to be executed soon, and therefore
6029 * bfqq to receive new I/O soon.
6030 *
6031 * But, if the blocking gets actually eliminated, then the
6032 * next think-time sample for bfqq may be very low. This in
6033 * turn may cause bfqq's think time to be deemed
6034 * short. Without the 100 ms barrier, this new state change
6035 * would cause the body of the next if to be executed
6036 * immediately. But this would set to 0 the inject
6037 * limit. Without injection, the blocking I/O would cause the
6038 * think time of bfqq to become long again, and therefore the
6039 * inject limit to be raised again, and so on. The only effect
6040 * of such a steady oscillation between the two think-time
6041 * states would be to prevent effective injection on bfqq.
6042 *
6043 * In contrast, if the inject limit is not reset during such a
6044 * long time interval as 100 ms, then the number of short
6045 * think time samples can grow significantly before the reset
6046 * is performed. As a consequence, the think time state can
6047 * become stable before the reset. Therefore there will be no
6048 * state change when the 100 ms elapse, and no reset of the
6049 * inject limit. The inject limit remains steadily equal to 1
6050 * both during and after the 100 ms. So injection can be
6051 * performed at all times, and throughput gets boosted.
6052 *
6053 * An inject limit equal to 1 is however in conflict, in
6054 * general, with the fact that the think time of bfqq is
6055 * short, because injection may be likely to delay bfqq's I/O
6056 * (as explained in the comments in
6057 * bfq_update_inject_limit()). But this does not happen in
6058 * this special case, because bfqq's low think time is due to
6059 * an effective handling of a synchronization, through
6060 * injection. In this special case, bfqq's I/O does not get
6061 * delayed by injection; on the contrary, bfqq's I/O is
6062 * brought forward, because it is not blocked for
6063 * milliseconds.
6064 *
6065 * In addition, serving the blocking I/O much sooner, and much
6066 * more frequently than once per I/O-plugging timeout, makes
6067 * it much quicker to detect a waker queue (the concept of
6068 * waker queue is defined in the comments in
6069 * bfq_add_request()). This makes it possible to start sooner
6070 * to boost throughput more effectively, by injecting the I/O
6071 * of the waker queue unconditionally on every
6072 * bfq_dispatch_request().
6073 *
6074 * One last, important benefit of not resetting the inject
6075 * limit before 100 ms is that, during this time interval, the
6076 * base value for the total service time is likely to get
6077 * finally computed for bfqq, freeing the inject limit from
6078 * its relation with the think time.
6079 */
6080 if (state_changed && bfqq->last_serv_time_ns == 0 &&
6081 (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
6082 msecs_to_jiffies(100)) ||
6083 !has_short_ttime))
6084 bfq_reset_inject_limit(bfqd, bfqq);
6085}
6086
6087/*
6088 * Called when a new fs request (rq) is added to bfqq. Check if there's
6089 * something we should do about it.
6090 */
6091static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
6092 struct request *rq)
6093{
6094 if (rq->cmd_flags & REQ_META)
6095 bfqq->meta_pending++;
6096
6097 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
6098
6099 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
6100 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
6101 blk_rq_sectors(rq) < 32;
6102 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
6103
6104 /*
6105 * There is just this request queued: if
6106 * - the request is small, and
6107 * - we are idling to boost throughput, and
6108 * - the queue is not to be expired,
6109 * then just exit.
6110 *
6111 * In this way, if the device is being idled to wait
6112 * for a new request from the in-service queue, we
6113 * avoid unplugging the device and committing the
6114 * device to serve just a small request. In contrast
6115 * we wait for the block layer to decide when to
6116 * unplug the device: hopefully, new requests will be
6117 * merged to this one quickly, then the device will be
6118 * unplugged and larger requests will be dispatched.
6119 */
6120 if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
6121 !budget_timeout)
6122 return;
6123
6124 /*
6125 * A large enough request arrived, or idling is being
6126 * performed to preserve service guarantees, or
6127 * finally the queue is to be expired: in all these
6128 * cases disk idling is to be stopped, so clear
6129 * wait_request flag and reset timer.
6130 */
6131 bfq_clear_bfqq_wait_request(bfqq);
6132 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
6133
6134 /*
6135 * The queue is not empty, because a new request just
6136 * arrived. Hence we can safely expire the queue, in
6137 * case of budget timeout, without risking that the
6138 * timestamps of the queue are not updated correctly.
6139 * See [1] for more details.
6140 */
6141 if (budget_timeout)
6142 bfq_bfqq_expire(bfqd, bfqq, false,
6143 BFQQE_BUDGET_TIMEOUT);
6144 }
6145}
6146
6147static void bfqq_request_allocated(struct bfq_queue *bfqq)
6148{
6149 struct bfq_entity *entity = &bfqq->entity;
6150
6151 for_each_entity(entity)
6152 entity->allocated++;
6153}
6154
6155static void bfqq_request_freed(struct bfq_queue *bfqq)
6156{
6157 struct bfq_entity *entity = &bfqq->entity;
6158
6159 for_each_entity(entity)
6160 entity->allocated--;
6161}
6162
6163/* returns true if it causes the idle timer to be disabled */
6164static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
6165{
6166 struct bfq_queue *bfqq = RQ_BFQQ(rq),
6167 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true,
6168 RQ_BIC(rq));
6169 bool waiting, idle_timer_disabled = false;
6170
6171 if (new_bfqq) {
6172 struct bfq_queue *old_bfqq = bfqq;
6173 /*
6174 * Release the request's reference to the old bfqq
6175 * and make sure one is taken to the shared queue.
6176 */
6177 bfqq_request_allocated(new_bfqq);
6178 bfqq_request_freed(bfqq);
6179 new_bfqq->ref++;
6180 /*
6181 * If the bic associated with the process
6182 * issuing this request still points to bfqq
6183 * (and thus has not been already redirected
6184 * to new_bfqq or even some other bfq_queue),
6185 * then complete the merge and redirect it to
6186 * new_bfqq.
6187 */
6188 if (bic_to_bfqq(RQ_BIC(rq), true,
6189 bfq_actuator_index(bfqd, rq->bio)) == bfqq) {
6190 while (bfqq != new_bfqq)
6191 bfqq = bfq_merge_bfqqs(bfqd, RQ_BIC(rq), bfqq);
6192 }
6193
6194 bfq_clear_bfqq_just_created(old_bfqq);
6195 /*
6196 * rq is about to be enqueued into new_bfqq,
6197 * release rq reference on bfqq
6198 */
6199 bfq_put_queue(old_bfqq);
6200 rq->elv.priv[1] = new_bfqq;
6201 }
6202
6203 bfq_update_io_thinktime(bfqd, bfqq);
6204 bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
6205 bfq_update_io_seektime(bfqd, bfqq, rq);
6206
6207 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
6208 bfq_add_request(rq);
6209 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
6210
6211 rq->fifo_time = blk_time_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
6212 list_add_tail(&rq->queuelist, &bfqq->fifo);
6213
6214 bfq_rq_enqueued(bfqd, bfqq, rq);
6215
6216 return idle_timer_disabled;
6217}
6218
6219#ifdef CONFIG_BFQ_CGROUP_DEBUG
6220static void bfq_update_insert_stats(struct request_queue *q,
6221 struct bfq_queue *bfqq,
6222 bool idle_timer_disabled,
6223 blk_opf_t cmd_flags)
6224{
6225 if (!bfqq)
6226 return;
6227
6228 /*
6229 * bfqq still exists, because it can disappear only after
6230 * either it is merged with another queue, or the process it
6231 * is associated with exits. But both actions must be taken by
6232 * the same process currently executing this flow of
6233 * instructions.
6234 *
6235 * In addition, the following queue lock guarantees that
6236 * bfqq_group(bfqq) exists as well.
6237 */
6238 spin_lock_irq(&q->queue_lock);
6239 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
6240 if (idle_timer_disabled)
6241 bfqg_stats_update_idle_time(bfqq_group(bfqq));
6242 spin_unlock_irq(&q->queue_lock);
6243}
6244#else
6245static inline void bfq_update_insert_stats(struct request_queue *q,
6246 struct bfq_queue *bfqq,
6247 bool idle_timer_disabled,
6248 blk_opf_t cmd_flags) {}
6249#endif /* CONFIG_BFQ_CGROUP_DEBUG */
6250
6251static struct bfq_queue *bfq_init_rq(struct request *rq);
6252
6253static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
6254 blk_insert_t flags)
6255{
6256 struct request_queue *q = hctx->queue;
6257 struct bfq_data *bfqd = q->elevator->elevator_data;
6258 struct bfq_queue *bfqq;
6259 bool idle_timer_disabled = false;
6260 blk_opf_t cmd_flags;
6261 LIST_HEAD(free);
6262
6263#ifdef CONFIG_BFQ_GROUP_IOSCHED
6264 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) && rq->bio)
6265 bfqg_stats_update_legacy_io(q, rq);
6266#endif
6267 spin_lock_irq(&bfqd->lock);
6268 bfqq = bfq_init_rq(rq);
6269 if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
6270 spin_unlock_irq(&bfqd->lock);
6271 blk_mq_free_requests(&free);
6272 return;
6273 }
6274
6275 trace_block_rq_insert(rq);
6276
6277 if (flags & BLK_MQ_INSERT_AT_HEAD) {
6278 list_add(&rq->queuelist, &bfqd->dispatch);
6279 } else if (!bfqq) {
6280 list_add_tail(&rq->queuelist, &bfqd->dispatch);
6281 } else {
6282 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
6283 /*
6284 * Update bfqq, because, if a queue merge has occurred
6285 * in __bfq_insert_request, then rq has been
6286 * redirected into a new queue.
6287 */
6288 bfqq = RQ_BFQQ(rq);
6289
6290 if (rq_mergeable(rq)) {
6291 elv_rqhash_add(q, rq);
6292 if (!q->last_merge)
6293 q->last_merge = rq;
6294 }
6295 }
6296
6297 /*
6298 * Cache cmd_flags before releasing scheduler lock, because rq
6299 * may disappear afterwards (for example, because of a request
6300 * merge).
6301 */
6302 cmd_flags = rq->cmd_flags;
6303 spin_unlock_irq(&bfqd->lock);
6304
6305 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
6306 cmd_flags);
6307}
6308
6309static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
6310 struct list_head *list,
6311 blk_insert_t flags)
6312{
6313 while (!list_empty(list)) {
6314 struct request *rq;
6315
6316 rq = list_first_entry(list, struct request, queuelist);
6317 list_del_init(&rq->queuelist);
6318 bfq_insert_request(hctx, rq, flags);
6319 }
6320}
6321
6322static void bfq_update_hw_tag(struct bfq_data *bfqd)
6323{
6324 struct bfq_queue *bfqq = bfqd->in_service_queue;
6325
6326 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
6327 bfqd->tot_rq_in_driver);
6328
6329 if (bfqd->hw_tag == 1)
6330 return;
6331
6332 /*
6333 * This sample is valid if the number of outstanding requests
6334 * is large enough to allow a queueing behavior. Note that the
6335 * sum is not exact, as it's not taking into account deactivated
6336 * requests.
6337 */
6338 if (bfqd->tot_rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
6339 return;
6340
6341 /*
6342 * If active queue hasn't enough requests and can idle, bfq might not
6343 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
6344 * case
6345 */
6346 if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
6347 bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
6348 BFQ_HW_QUEUE_THRESHOLD &&
6349 bfqd->tot_rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
6350 return;
6351
6352 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
6353 return;
6354
6355 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
6356 bfqd->max_rq_in_driver = 0;
6357 bfqd->hw_tag_samples = 0;
6358
6359 bfqd->nonrot_with_queueing =
6360 blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
6361}
6362
6363static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
6364{
6365 u64 now_ns;
6366 u32 delta_us;
6367
6368 bfq_update_hw_tag(bfqd);
6369
6370 bfqd->rq_in_driver[bfqq->actuator_idx]--;
6371 bfqd->tot_rq_in_driver--;
6372 bfqq->dispatched--;
6373
6374 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
6375 /*
6376 * Set budget_timeout (which we overload to store the
6377 * time at which the queue remains with no backlog and
6378 * no outstanding request; used by the weight-raising
6379 * mechanism).
6380 */
6381 bfqq->budget_timeout = jiffies;
6382
6383 bfq_del_bfqq_in_groups_with_pending_reqs(bfqq);
6384 bfq_weights_tree_remove(bfqq);
6385 }
6386
6387 now_ns = blk_time_get_ns();
6388
6389 bfqq->ttime.last_end_request = now_ns;
6390
6391 /*
6392 * Using us instead of ns, to get a reasonable precision in
6393 * computing rate in next check.
6394 */
6395 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
6396
6397 /*
6398 * If the request took rather long to complete, and, according
6399 * to the maximum request size recorded, this completion latency
6400 * implies that the request was certainly served at a very low
6401 * rate (less than 1M sectors/sec), then the whole observation
6402 * interval that lasts up to this time instant cannot be a
6403 * valid time interval for computing a new peak rate. Invoke
6404 * bfq_update_rate_reset to have the following three steps
6405 * taken:
6406 * - close the observation interval at the last (previous)
6407 * request dispatch or completion
6408 * - compute rate, if possible, for that observation interval
6409 * - reset to zero samples, which will trigger a proper
6410 * re-initialization of the observation interval on next
6411 * dispatch
6412 */
6413 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
6414 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
6415 1UL<<(BFQ_RATE_SHIFT - 10))
6416 bfq_update_rate_reset(bfqd, NULL);
6417 bfqd->last_completion = now_ns;
6418 /*
6419 * Shared queues are likely to receive I/O at a high
6420 * rate. This may deceptively let them be considered as wakers
6421 * of other queues. But a false waker will unjustly steal
6422 * bandwidth to its supposedly woken queue. So considering
6423 * also shared queues in the waking mechanism may cause more
6424 * control troubles than throughput benefits. Then reset
6425 * last_completed_rq_bfqq if bfqq is a shared queue.
6426 */
6427 if (!bfq_bfqq_coop(bfqq))
6428 bfqd->last_completed_rq_bfqq = bfqq;
6429 else
6430 bfqd->last_completed_rq_bfqq = NULL;
6431
6432 /*
6433 * If we are waiting to discover whether the request pattern
6434 * of the task associated with the queue is actually
6435 * isochronous, and both requisites for this condition to hold
6436 * are now satisfied, then compute soft_rt_next_start (see the
6437 * comments on the function bfq_bfqq_softrt_next_start()). We
6438 * do not compute soft_rt_next_start if bfqq is in interactive
6439 * weight raising (see the comments in bfq_bfqq_expire() for
6440 * an explanation). We schedule this delayed update when bfqq
6441 * expires, if it still has in-flight requests.
6442 */
6443 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
6444 RB_EMPTY_ROOT(&bfqq->sort_list) &&
6445 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
6446 bfqq->soft_rt_next_start =
6447 bfq_bfqq_softrt_next_start(bfqd, bfqq);
6448
6449 /*
6450 * If this is the in-service queue, check if it needs to be expired,
6451 * or if we want to idle in case it has no pending requests.
6452 */
6453 if (bfqd->in_service_queue == bfqq) {
6454 if (bfq_bfqq_must_idle(bfqq)) {
6455 if (bfqq->dispatched == 0)
6456 bfq_arm_slice_timer(bfqd);
6457 /*
6458 * If we get here, we do not expire bfqq, even
6459 * if bfqq was in budget timeout or had no
6460 * more requests (as controlled in the next
6461 * conditional instructions). The reason for
6462 * not expiring bfqq is as follows.
6463 *
6464 * Here bfqq->dispatched > 0 holds, but
6465 * bfq_bfqq_must_idle() returned true. This
6466 * implies that, even if no request arrives
6467 * for bfqq before bfqq->dispatched reaches 0,
6468 * bfqq will, however, not be expired on the
6469 * completion event that causes bfqq->dispatch
6470 * to reach zero. In contrast, on this event,
6471 * bfqq will start enjoying device idling
6472 * (I/O-dispatch plugging).
6473 *
6474 * But, if we expired bfqq here, bfqq would
6475 * not have the chance to enjoy device idling
6476 * when bfqq->dispatched finally reaches
6477 * zero. This would expose bfqq to violation
6478 * of its reserved service guarantees.
6479 */
6480 return;
6481 } else if (bfq_may_expire_for_budg_timeout(bfqq))
6482 bfq_bfqq_expire(bfqd, bfqq, false,
6483 BFQQE_BUDGET_TIMEOUT);
6484 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
6485 (bfqq->dispatched == 0 ||
6486 !bfq_better_to_idle(bfqq)))
6487 bfq_bfqq_expire(bfqd, bfqq, false,
6488 BFQQE_NO_MORE_REQUESTS);
6489 }
6490
6491 if (!bfqd->tot_rq_in_driver)
6492 bfq_schedule_dispatch(bfqd);
6493}
6494
6495/*
6496 * The processes associated with bfqq may happen to generate their
6497 * cumulative I/O at a lower rate than the rate at which the device
6498 * could serve the same I/O. This is rather probable, e.g., if only
6499 * one process is associated with bfqq and the device is an SSD. It
6500 * results in bfqq becoming often empty while in service. In this
6501 * respect, if BFQ is allowed to switch to another queue when bfqq
6502 * remains empty, then the device goes on being fed with I/O requests,
6503 * and the throughput is not affected. In contrast, if BFQ is not
6504 * allowed to switch to another queue---because bfqq is sync and
6505 * I/O-dispatch needs to be plugged while bfqq is temporarily
6506 * empty---then, during the service of bfqq, there will be frequent
6507 * "service holes", i.e., time intervals during which bfqq gets empty
6508 * and the device can only consume the I/O already queued in its
6509 * hardware queues. During service holes, the device may even get to
6510 * remaining idle. In the end, during the service of bfqq, the device
6511 * is driven at a lower speed than the one it can reach with the kind
6512 * of I/O flowing through bfqq.
6513 *
6514 * To counter this loss of throughput, BFQ implements a "request
6515 * injection mechanism", which tries to fill the above service holes
6516 * with I/O requests taken from other queues. The hard part in this
6517 * mechanism is finding the right amount of I/O to inject, so as to
6518 * both boost throughput and not break bfqq's bandwidth and latency
6519 * guarantees. In this respect, the mechanism maintains a per-queue
6520 * inject limit, computed as below. While bfqq is empty, the injection
6521 * mechanism dispatches extra I/O requests only until the total number
6522 * of I/O requests in flight---i.e., already dispatched but not yet
6523 * completed---remains lower than this limit.
6524 *
6525 * A first definition comes in handy to introduce the algorithm by
6526 * which the inject limit is computed. We define as first request for
6527 * bfqq, an I/O request for bfqq that arrives while bfqq is in
6528 * service, and causes bfqq to switch from empty to non-empty. The
6529 * algorithm updates the limit as a function of the effect of
6530 * injection on the service times of only the first requests of
6531 * bfqq. The reason for this restriction is that these are the
6532 * requests whose service time is affected most, because they are the
6533 * first to arrive after injection possibly occurred.
6534 *
6535 * To evaluate the effect of injection, the algorithm measures the
6536 * "total service time" of first requests. We define as total service
6537 * time of an I/O request, the time that elapses since when the
6538 * request is enqueued into bfqq, to when it is completed. This
6539 * quantity allows the whole effect of injection to be measured. It is
6540 * easy to see why. Suppose that some requests of other queues are
6541 * actually injected while bfqq is empty, and that a new request R
6542 * then arrives for bfqq. If the device does start to serve all or
6543 * part of the injected requests during the service hole, then,
6544 * because of this extra service, it may delay the next invocation of
6545 * the dispatch hook of BFQ. Then, even after R gets eventually
6546 * dispatched, the device may delay the actual service of R if it is
6547 * still busy serving the extra requests, or if it decides to serve,
6548 * before R, some extra request still present in its queues. As a
6549 * conclusion, the cumulative extra delay caused by injection can be
6550 * easily evaluated by just comparing the total service time of first
6551 * requests with and without injection.
6552 *
6553 * The limit-update algorithm works as follows. On the arrival of a
6554 * first request of bfqq, the algorithm measures the total time of the
6555 * request only if one of the three cases below holds, and, for each
6556 * case, it updates the limit as described below:
6557 *
6558 * (1) If there is no in-flight request. This gives a baseline for the
6559 * total service time of the requests of bfqq. If the baseline has
6560 * not been computed yet, then, after computing it, the limit is
6561 * set to 1, to start boosting throughput, and to prepare the
6562 * ground for the next case. If the baseline has already been
6563 * computed, then it is updated, in case it results to be lower
6564 * than the previous value.
6565 *
6566 * (2) If the limit is higher than 0 and there are in-flight
6567 * requests. By comparing the total service time in this case with
6568 * the above baseline, it is possible to know at which extent the
6569 * current value of the limit is inflating the total service
6570 * time. If the inflation is below a certain threshold, then bfqq
6571 * is assumed to be suffering from no perceivable loss of its
6572 * service guarantees, and the limit is even tentatively
6573 * increased. If the inflation is above the threshold, then the
6574 * limit is decreased. Due to the lack of any hysteresis, this
6575 * logic makes the limit oscillate even in steady workload
6576 * conditions. Yet we opted for it, because it is fast in reaching
6577 * the best value for the limit, as a function of the current I/O
6578 * workload. To reduce oscillations, this step is disabled for a
6579 * short time interval after the limit happens to be decreased.
6580 *
6581 * (3) Periodically, after resetting the limit, to make sure that the
6582 * limit eventually drops in case the workload changes. This is
6583 * needed because, after the limit has gone safely up for a
6584 * certain workload, it is impossible to guess whether the
6585 * baseline total service time may have changed, without measuring
6586 * it again without injection. A more effective version of this
6587 * step might be to just sample the baseline, by interrupting
6588 * injection only once, and then to reset/lower the limit only if
6589 * the total service time with the current limit does happen to be
6590 * too large.
6591 *
6592 * More details on each step are provided in the comments on the
6593 * pieces of code that implement these steps: the branch handling the
6594 * transition from empty to non empty in bfq_add_request(), the branch
6595 * handling injection in bfq_select_queue(), and the function
6596 * bfq_choose_bfqq_for_injection(). These comments also explain some
6597 * exceptions, made by the injection mechanism in some special cases.
6598 */
6599static void bfq_update_inject_limit(struct bfq_data *bfqd,
6600 struct bfq_queue *bfqq)
6601{
6602 u64 tot_time_ns = blk_time_get_ns() - bfqd->last_empty_occupied_ns;
6603 unsigned int old_limit = bfqq->inject_limit;
6604
6605 if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) {
6606 u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
6607
6608 if (tot_time_ns >= threshold && old_limit > 0) {
6609 bfqq->inject_limit--;
6610 bfqq->decrease_time_jif = jiffies;
6611 } else if (tot_time_ns < threshold &&
6612 old_limit <= bfqd->max_rq_in_driver)
6613 bfqq->inject_limit++;
6614 }
6615
6616 /*
6617 * Either we still have to compute the base value for the
6618 * total service time, and there seem to be the right
6619 * conditions to do it, or we can lower the last base value
6620 * computed.
6621 *
6622 * NOTE: (bfqd->tot_rq_in_driver == 1) means that there is no I/O
6623 * request in flight, because this function is in the code
6624 * path that handles the completion of a request of bfqq, and,
6625 * in particular, this function is executed before
6626 * bfqd->tot_rq_in_driver is decremented in such a code path.
6627 */
6628 if ((bfqq->last_serv_time_ns == 0 && bfqd->tot_rq_in_driver == 1) ||
6629 tot_time_ns < bfqq->last_serv_time_ns) {
6630 if (bfqq->last_serv_time_ns == 0) {
6631 /*
6632 * Now we certainly have a base value: make sure we
6633 * start trying injection.
6634 */
6635 bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
6636 }
6637 bfqq->last_serv_time_ns = tot_time_ns;
6638 } else if (!bfqd->rqs_injected && bfqd->tot_rq_in_driver == 1)
6639 /*
6640 * No I/O injected and no request still in service in
6641 * the drive: these are the exact conditions for
6642 * computing the base value of the total service time
6643 * for bfqq. So let's update this value, because it is
6644 * rather variable. For example, it varies if the size
6645 * or the spatial locality of the I/O requests in bfqq
6646 * change.
6647 */
6648 bfqq->last_serv_time_ns = tot_time_ns;
6649
6650
6651 /* update complete, not waiting for any request completion any longer */
6652 bfqd->waited_rq = NULL;
6653 bfqd->rqs_injected = false;
6654}
6655
6656/*
6657 * Handle either a requeue or a finish for rq. The things to do are
6658 * the same in both cases: all references to rq are to be dropped. In
6659 * particular, rq is considered completed from the point of view of
6660 * the scheduler.
6661 */
6662static void bfq_finish_requeue_request(struct request *rq)
6663{
6664 struct bfq_queue *bfqq = RQ_BFQQ(rq);
6665 struct bfq_data *bfqd;
6666 unsigned long flags;
6667
6668 /*
6669 * rq either is not associated with any icq, or is an already
6670 * requeued request that has not (yet) been re-inserted into
6671 * a bfq_queue.
6672 */
6673 if (!rq->elv.icq || !bfqq)
6674 return;
6675
6676 bfqd = bfqq->bfqd;
6677
6678 if (rq->rq_flags & RQF_STARTED)
6679 bfqg_stats_update_completion(bfqq_group(bfqq),
6680 rq->start_time_ns,
6681 rq->io_start_time_ns,
6682 rq->cmd_flags);
6683
6684 spin_lock_irqsave(&bfqd->lock, flags);
6685 if (likely(rq->rq_flags & RQF_STARTED)) {
6686 if (rq == bfqd->waited_rq)
6687 bfq_update_inject_limit(bfqd, bfqq);
6688
6689 bfq_completed_request(bfqq, bfqd);
6690 }
6691 bfqq_request_freed(bfqq);
6692 bfq_put_queue(bfqq);
6693 RQ_BIC(rq)->requests--;
6694 spin_unlock_irqrestore(&bfqd->lock, flags);
6695
6696 /*
6697 * Reset private fields. In case of a requeue, this allows
6698 * this function to correctly do nothing if it is spuriously
6699 * invoked again on this same request (see the check at the
6700 * beginning of the function). Probably, a better general
6701 * design would be to prevent blk-mq from invoking the requeue
6702 * or finish hooks of an elevator, for a request that is not
6703 * referred by that elevator.
6704 *
6705 * Resetting the following fields would break the
6706 * request-insertion logic if rq is re-inserted into a bfq
6707 * internal queue, without a re-preparation. Here we assume
6708 * that re-insertions of requeued requests, without
6709 * re-preparation, can happen only for pass_through or at_head
6710 * requests (which are not re-inserted into bfq internal
6711 * queues).
6712 */
6713 rq->elv.priv[0] = NULL;
6714 rq->elv.priv[1] = NULL;
6715}
6716
6717static void bfq_finish_request(struct request *rq)
6718{
6719 bfq_finish_requeue_request(rq);
6720
6721 if (rq->elv.icq) {
6722 put_io_context(rq->elv.icq->ioc);
6723 rq->elv.icq = NULL;
6724 }
6725}
6726
6727/*
6728 * Removes the association between the current task and bfqq, assuming
6729 * that bic points to the bfq iocontext of the task.
6730 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
6731 * was the last process referring to that bfqq.
6732 */
6733static struct bfq_queue *
6734bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
6735{
6736 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
6737
6738 if (bfqq_process_refs(bfqq) == 1 && !bfqq->new_bfqq) {
6739 bfqq->pid = current->pid;
6740 bfq_clear_bfqq_coop(bfqq);
6741 bfq_clear_bfqq_split_coop(bfqq);
6742 return bfqq;
6743 }
6744
6745 bic_set_bfqq(bic, NULL, true, bfqq->actuator_idx);
6746
6747 bfq_put_cooperator(bfqq);
6748
6749 bfq_release_process_ref(bfqq->bfqd, bfqq);
6750 return NULL;
6751}
6752
6753static struct bfq_queue *
6754__bfq_get_bfqq_handle_split(struct bfq_data *bfqd, struct bfq_io_cq *bic,
6755 struct bio *bio, bool split, bool is_sync,
6756 bool *new_queue)
6757{
6758 unsigned int act_idx = bfq_actuator_index(bfqd, bio);
6759 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync, act_idx);
6760 struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[act_idx];
6761
6762 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
6763 return bfqq;
6764
6765 if (new_queue)
6766 *new_queue = true;
6767
6768 if (bfqq)
6769 bfq_put_queue(bfqq);
6770 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic, split);
6771
6772 bic_set_bfqq(bic, bfqq, is_sync, act_idx);
6773 if (split && is_sync) {
6774 if ((bfqq_data->was_in_burst_list && bfqd->large_burst) ||
6775 bfqq_data->saved_in_large_burst)
6776 bfq_mark_bfqq_in_large_burst(bfqq);
6777 else {
6778 bfq_clear_bfqq_in_large_burst(bfqq);
6779 if (bfqq_data->was_in_burst_list)
6780 /*
6781 * If bfqq was in the current
6782 * burst list before being
6783 * merged, then we have to add
6784 * it back. And we do not need
6785 * to increase burst_size, as
6786 * we did not decrement
6787 * burst_size when we removed
6788 * bfqq from the burst list as
6789 * a consequence of a merge
6790 * (see comments in
6791 * bfq_put_queue). In this
6792 * respect, it would be rather
6793 * costly to know whether the
6794 * current burst list is still
6795 * the same burst list from
6796 * which bfqq was removed on
6797 * the merge. To avoid this
6798 * cost, if bfqq was in a
6799 * burst list, then we add
6800 * bfqq to the current burst
6801 * list without any further
6802 * check. This can cause
6803 * inappropriate insertions,
6804 * but rarely enough to not
6805 * harm the detection of large
6806 * bursts significantly.
6807 */
6808 hlist_add_head(&bfqq->burst_list_node,
6809 &bfqd->burst_list);
6810 }
6811 bfqq->split_time = jiffies;
6812 }
6813
6814 return bfqq;
6815}
6816
6817/*
6818 * Only reset private fields. The actual request preparation will be
6819 * performed by bfq_init_rq, when rq is either inserted or merged. See
6820 * comments on bfq_init_rq for the reason behind this delayed
6821 * preparation.
6822 */
6823static void bfq_prepare_request(struct request *rq)
6824{
6825 rq->elv.icq = ioc_find_get_icq(rq->q);
6826
6827 /*
6828 * Regardless of whether we have an icq attached, we have to
6829 * clear the scheduler pointers, as they might point to
6830 * previously allocated bic/bfqq structs.
6831 */
6832 rq->elv.priv[0] = rq->elv.priv[1] = NULL;
6833}
6834
6835static struct bfq_queue *bfq_waker_bfqq(struct bfq_queue *bfqq)
6836{
6837 struct bfq_queue *new_bfqq = bfqq->new_bfqq;
6838 struct bfq_queue *waker_bfqq = bfqq->waker_bfqq;
6839
6840 if (!waker_bfqq)
6841 return NULL;
6842
6843 while (new_bfqq) {
6844 if (new_bfqq == waker_bfqq) {
6845 /*
6846 * If waker_bfqq is in the merge chain, and current
6847 * is the only process, waker_bfqq can be freed.
6848 */
6849 if (bfqq_process_refs(waker_bfqq) == 1)
6850 return NULL;
6851
6852 return waker_bfqq;
6853 }
6854
6855 new_bfqq = new_bfqq->new_bfqq;
6856 }
6857
6858 /*
6859 * If waker_bfqq is not in the merge chain, and it's procress reference
6860 * is 0, waker_bfqq can be freed.
6861 */
6862 if (bfqq_process_refs(waker_bfqq) == 0)
6863 return NULL;
6864
6865 return waker_bfqq;
6866}
6867
6868static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
6869 struct bfq_io_cq *bic,
6870 struct bio *bio,
6871 unsigned int idx,
6872 bool is_sync)
6873{
6874 struct bfq_queue *waker_bfqq;
6875 struct bfq_queue *bfqq;
6876 bool new_queue = false;
6877
6878 bfqq = __bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
6879 &new_queue);
6880 if (unlikely(new_queue))
6881 return bfqq;
6882
6883 /* If the queue was seeky for too long, break it apart. */
6884 if (!bfq_bfqq_coop(bfqq) || !bfq_bfqq_split_coop(bfqq) ||
6885 bic->bfqq_data[idx].stably_merged)
6886 return bfqq;
6887
6888 waker_bfqq = bfq_waker_bfqq(bfqq);
6889
6890 /* Update bic before losing reference to bfqq */
6891 if (bfq_bfqq_in_large_burst(bfqq))
6892 bic->bfqq_data[idx].saved_in_large_burst = true;
6893
6894 bfqq = bfq_split_bfqq(bic, bfqq);
6895 if (bfqq) {
6896 bfq_bfqq_resume_state(bfqq, bfqd, bic, true);
6897 return bfqq;
6898 }
6899
6900 bfqq = __bfq_get_bfqq_handle_split(bfqd, bic, bio, true, is_sync, NULL);
6901 if (unlikely(bfqq == &bfqd->oom_bfqq))
6902 return bfqq;
6903
6904 bfq_bfqq_resume_state(bfqq, bfqd, bic, false);
6905 bfqq->waker_bfqq = waker_bfqq;
6906 bfqq->tentative_waker_bfqq = NULL;
6907
6908 /*
6909 * If the waker queue disappears, then new_bfqq->waker_bfqq must be
6910 * reset. So insert new_bfqq into the
6911 * woken_list of the waker. See
6912 * bfq_check_waker for details.
6913 */
6914 if (waker_bfqq)
6915 hlist_add_head(&bfqq->woken_list_node,
6916 &bfqq->waker_bfqq->woken_list);
6917
6918 return bfqq;
6919}
6920
6921/*
6922 * If needed, init rq, allocate bfq data structures associated with
6923 * rq, and increment reference counters in the destination bfq_queue
6924 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6925 * not associated with any bfq_queue.
6926 *
6927 * This function is invoked by the functions that perform rq insertion
6928 * or merging. One may have expected the above preparation operations
6929 * to be performed in bfq_prepare_request, and not delayed to when rq
6930 * is inserted or merged. The rationale behind this delayed
6931 * preparation is that, after the prepare_request hook is invoked for
6932 * rq, rq may still be transformed into a request with no icq, i.e., a
6933 * request not associated with any queue. No bfq hook is invoked to
6934 * signal this transformation. As a consequence, should these
6935 * preparation operations be performed when the prepare_request hook
6936 * is invoked, and should rq be transformed one moment later, bfq
6937 * would end up in an inconsistent state, because it would have
6938 * incremented some queue counters for an rq destined to
6939 * transformation, without any chance to correctly lower these
6940 * counters back. In contrast, no transformation can still happen for
6941 * rq after rq has been inserted or merged. So, it is safe to execute
6942 * these preparation operations when rq is finally inserted or merged.
6943 */
6944static struct bfq_queue *bfq_init_rq(struct request *rq)
6945{
6946 struct request_queue *q = rq->q;
6947 struct bio *bio = rq->bio;
6948 struct bfq_data *bfqd = q->elevator->elevator_data;
6949 struct bfq_io_cq *bic;
6950 const int is_sync = rq_is_sync(rq);
6951 struct bfq_queue *bfqq;
6952 unsigned int a_idx = bfq_actuator_index(bfqd, bio);
6953
6954 if (unlikely(!rq->elv.icq))
6955 return NULL;
6956
6957 /*
6958 * Assuming that RQ_BFQQ(rq) is set only if everything is set
6959 * for this rq. This holds true, because this function is
6960 * invoked only for insertion or merging, and, after such
6961 * events, a request cannot be manipulated any longer before
6962 * being removed from bfq.
6963 */
6964 if (RQ_BFQQ(rq))
6965 return RQ_BFQQ(rq);
6966
6967 bic = icq_to_bic(rq->elv.icq);
6968 bfq_check_ioprio_change(bic, bio);
6969 bfq_bic_update_cgroup(bic, bio);
6970 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, a_idx, is_sync);
6971
6972 bfqq_request_allocated(bfqq);
6973 bfqq->ref++;
6974 bic->requests++;
6975 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
6976 rq, bfqq, bfqq->ref);
6977
6978 rq->elv.priv[0] = bic;
6979 rq->elv.priv[1] = bfqq;
6980
6981 /*
6982 * If a bfq_queue has only one process reference, it is owned
6983 * by only this bic: we can then set bfqq->bic = bic. in
6984 * addition, if the queue has also just been split, we have to
6985 * resume its state.
6986 */
6987 if (likely(bfqq != &bfqd->oom_bfqq) && !bfqq->new_bfqq &&
6988 bfqq_process_refs(bfqq) == 1)
6989 bfqq->bic = bic;
6990
6991 /*
6992 * Consider bfqq as possibly belonging to a burst of newly
6993 * created queues only if:
6994 * 1) A burst is actually happening (bfqd->burst_size > 0)
6995 * or
6996 * 2) There is no other active queue. In fact, if, in
6997 * contrast, there are active queues not belonging to the
6998 * possible burst bfqq may belong to, then there is no gain
6999 * in considering bfqq as belonging to a burst, and
7000 * therefore in not weight-raising bfqq. See comments on
7001 * bfq_handle_burst().
7002 *
7003 * This filtering also helps eliminating false positives,
7004 * occurring when bfqq does not belong to an actual large
7005 * burst, but some background task (e.g., a service) happens
7006 * to trigger the creation of new queues very close to when
7007 * bfqq and its possible companion queues are created. See
7008 * comments on bfq_handle_burst() for further details also on
7009 * this issue.
7010 */
7011 if (unlikely(bfq_bfqq_just_created(bfqq) &&
7012 (bfqd->burst_size > 0 ||
7013 bfq_tot_busy_queues(bfqd) == 0)))
7014 bfq_handle_burst(bfqd, bfqq);
7015
7016 return bfqq;
7017}
7018
7019static void
7020bfq_idle_slice_timer_body(struct bfq_data *bfqd, struct bfq_queue *bfqq)
7021{
7022 enum bfqq_expiration reason;
7023 unsigned long flags;
7024
7025 spin_lock_irqsave(&bfqd->lock, flags);
7026
7027 /*
7028 * Considering that bfqq may be in race, we should firstly check
7029 * whether bfqq is in service before doing something on it. If
7030 * the bfqq in race is not in service, it has already been expired
7031 * through __bfq_bfqq_expire func and its wait_request flags has
7032 * been cleared in __bfq_bfqd_reset_in_service func.
7033 */
7034 if (bfqq != bfqd->in_service_queue) {
7035 spin_unlock_irqrestore(&bfqd->lock, flags);
7036 return;
7037 }
7038
7039 bfq_clear_bfqq_wait_request(bfqq);
7040
7041 if (bfq_bfqq_budget_timeout(bfqq))
7042 /*
7043 * Also here the queue can be safely expired
7044 * for budget timeout without wasting
7045 * guarantees
7046 */
7047 reason = BFQQE_BUDGET_TIMEOUT;
7048 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
7049 /*
7050 * The queue may not be empty upon timer expiration,
7051 * because we may not disable the timer when the
7052 * first request of the in-service queue arrives
7053 * during disk idling.
7054 */
7055 reason = BFQQE_TOO_IDLE;
7056 else
7057 goto schedule_dispatch;
7058
7059 bfq_bfqq_expire(bfqd, bfqq, true, reason);
7060
7061schedule_dispatch:
7062 bfq_schedule_dispatch(bfqd);
7063 spin_unlock_irqrestore(&bfqd->lock, flags);
7064}
7065
7066/*
7067 * Handler of the expiration of the timer running if the in-service queue
7068 * is idling inside its time slice.
7069 */
7070static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
7071{
7072 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
7073 idle_slice_timer);
7074 struct bfq_queue *bfqq = bfqd->in_service_queue;
7075
7076 /*
7077 * Theoretical race here: the in-service queue can be NULL or
7078 * different from the queue that was idling if a new request
7079 * arrives for the current queue and there is a full dispatch
7080 * cycle that changes the in-service queue. This can hardly
7081 * happen, but in the worst case we just expire a queue too
7082 * early.
7083 */
7084 if (bfqq)
7085 bfq_idle_slice_timer_body(bfqd, bfqq);
7086
7087 return HRTIMER_NORESTART;
7088}
7089
7090static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
7091 struct bfq_queue **bfqq_ptr)
7092{
7093 struct bfq_queue *bfqq = *bfqq_ptr;
7094
7095 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
7096 if (bfqq) {
7097 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
7098
7099 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
7100 bfqq, bfqq->ref);
7101 bfq_put_queue(bfqq);
7102 *bfqq_ptr = NULL;
7103 }
7104}
7105
7106/*
7107 * Release all the bfqg references to its async queues. If we are
7108 * deallocating the group these queues may still contain requests, so
7109 * we reparent them to the root cgroup (i.e., the only one that will
7110 * exist for sure until all the requests on a device are gone).
7111 */
7112void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
7113{
7114 int i, j, k;
7115
7116 for (k = 0; k < bfqd->num_actuators; k++) {
7117 for (i = 0; i < 2; i++)
7118 for (j = 0; j < IOPRIO_NR_LEVELS; j++)
7119 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j][k]);
7120
7121 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq[k]);
7122 }
7123}
7124
7125/*
7126 * See the comments on bfq_limit_depth for the purpose of
7127 * the depths set in the function. Return minimum shallow depth we'll use.
7128 */
7129static void bfq_update_depths(struct bfq_data *bfqd, struct sbitmap_queue *bt)
7130{
7131 unsigned int depth = 1U << bt->sb.shift;
7132
7133 bfqd->full_depth_shift = bt->sb.shift;
7134 /*
7135 * In-word depths if no bfq_queue is being weight-raised:
7136 * leaving 25% of tags only for sync reads.
7137 *
7138 * In next formulas, right-shift the value
7139 * (1U<<bt->sb.shift), instead of computing directly
7140 * (1U<<(bt->sb.shift - something)), to be robust against
7141 * any possible value of bt->sb.shift, without having to
7142 * limit 'something'.
7143 */
7144 /* no more than 50% of tags for async I/O */
7145 bfqd->word_depths[0][0] = max(depth >> 1, 1U);
7146 /*
7147 * no more than 75% of tags for sync writes (25% extra tags
7148 * w.r.t. async I/O, to prevent async I/O from starving sync
7149 * writes)
7150 */
7151 bfqd->word_depths[0][1] = max((depth * 3) >> 2, 1U);
7152
7153 /*
7154 * In-word depths in case some bfq_queue is being weight-
7155 * raised: leaving ~63% of tags for sync reads. This is the
7156 * highest percentage for which, in our tests, application
7157 * start-up times didn't suffer from any regression due to tag
7158 * shortage.
7159 */
7160 /* no more than ~18% of tags for async I/O */
7161 bfqd->word_depths[1][0] = max((depth * 3) >> 4, 1U);
7162 /* no more than ~37% of tags for sync writes (~20% extra tags) */
7163 bfqd->word_depths[1][1] = max((depth * 6) >> 4, 1U);
7164}
7165
7166static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
7167{
7168 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
7169 struct blk_mq_tags *tags = hctx->sched_tags;
7170
7171 bfq_update_depths(bfqd, &tags->bitmap_tags);
7172 sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, 1);
7173}
7174
7175static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
7176{
7177 bfq_depth_updated(hctx);
7178 return 0;
7179}
7180
7181static void bfq_exit_queue(struct elevator_queue *e)
7182{
7183 struct bfq_data *bfqd = e->elevator_data;
7184 struct bfq_queue *bfqq, *n;
7185 unsigned int actuator;
7186
7187 hrtimer_cancel(&bfqd->idle_slice_timer);
7188
7189 spin_lock_irq(&bfqd->lock);
7190 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
7191 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
7192 spin_unlock_irq(&bfqd->lock);
7193
7194 for (actuator = 0; actuator < bfqd->num_actuators; actuator++)
7195 WARN_ON_ONCE(bfqd->rq_in_driver[actuator]);
7196 WARN_ON_ONCE(bfqd->tot_rq_in_driver);
7197
7198 hrtimer_cancel(&bfqd->idle_slice_timer);
7199
7200 /* release oom-queue reference to root group */
7201 bfqg_and_blkg_put(bfqd->root_group);
7202
7203#ifdef CONFIG_BFQ_GROUP_IOSCHED
7204 blkcg_deactivate_policy(bfqd->queue->disk, &blkcg_policy_bfq);
7205#else
7206 spin_lock_irq(&bfqd->lock);
7207 bfq_put_async_queues(bfqd, bfqd->root_group);
7208 kfree(bfqd->root_group);
7209 spin_unlock_irq(&bfqd->lock);
7210#endif
7211
7212 blk_stat_disable_accounting(bfqd->queue);
7213 clear_bit(ELEVATOR_FLAG_DISABLE_WBT, &e->flags);
7214 wbt_enable_default(bfqd->queue->disk);
7215
7216 kfree(bfqd);
7217}
7218
7219static void bfq_init_root_group(struct bfq_group *root_group,
7220 struct bfq_data *bfqd)
7221{
7222 int i;
7223
7224#ifdef CONFIG_BFQ_GROUP_IOSCHED
7225 root_group->entity.parent = NULL;
7226 root_group->my_entity = NULL;
7227 root_group->bfqd = bfqd;
7228#endif
7229 root_group->rq_pos_tree = RB_ROOT;
7230 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
7231 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
7232 root_group->sched_data.bfq_class_idle_last_service = jiffies;
7233}
7234
7235static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
7236{
7237 struct bfq_data *bfqd;
7238 struct elevator_queue *eq;
7239 unsigned int i;
7240 struct blk_independent_access_ranges *ia_ranges = q->disk->ia_ranges;
7241
7242 eq = elevator_alloc(q, e);
7243 if (!eq)
7244 return -ENOMEM;
7245
7246 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
7247 if (!bfqd) {
7248 kobject_put(&eq->kobj);
7249 return -ENOMEM;
7250 }
7251 eq->elevator_data = bfqd;
7252
7253 spin_lock_irq(&q->queue_lock);
7254 q->elevator = eq;
7255 spin_unlock_irq(&q->queue_lock);
7256
7257 /*
7258 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
7259 * Grab a permanent reference to it, so that the normal code flow
7260 * will not attempt to free it.
7261 * Set zero as actuator index: we will pretend that
7262 * all I/O requests are for the same actuator.
7263 */
7264 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0, 0);
7265 bfqd->oom_bfqq.ref++;
7266 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
7267 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
7268 bfqd->oom_bfqq.entity.new_weight =
7269 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
7270
7271 /* oom_bfqq does not participate to bursts */
7272 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
7273
7274 /*
7275 * Trigger weight initialization, according to ioprio, at the
7276 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
7277 * class won't be changed any more.
7278 */
7279 bfqd->oom_bfqq.entity.prio_changed = 1;
7280
7281 bfqd->queue = q;
7282
7283 bfqd->num_actuators = 1;
7284 /*
7285 * If the disk supports multiple actuators, copy independent
7286 * access ranges from the request queue structure.
7287 */
7288 spin_lock_irq(&q->queue_lock);
7289 if (ia_ranges) {
7290 /*
7291 * Check if the disk ia_ranges size exceeds the current bfq
7292 * actuator limit.
7293 */
7294 if (ia_ranges->nr_ia_ranges > BFQ_MAX_ACTUATORS) {
7295 pr_crit("nr_ia_ranges higher than act limit: iars=%d, max=%d.\n",
7296 ia_ranges->nr_ia_ranges, BFQ_MAX_ACTUATORS);
7297 pr_crit("Falling back to single actuator mode.\n");
7298 } else {
7299 bfqd->num_actuators = ia_ranges->nr_ia_ranges;
7300
7301 for (i = 0; i < bfqd->num_actuators; i++) {
7302 bfqd->sector[i] = ia_ranges->ia_range[i].sector;
7303 bfqd->nr_sectors[i] =
7304 ia_ranges->ia_range[i].nr_sectors;
7305 }
7306 }
7307 }
7308
7309 /* Otherwise use single-actuator dev info */
7310 if (bfqd->num_actuators == 1) {
7311 bfqd->sector[0] = 0;
7312 bfqd->nr_sectors[0] = get_capacity(q->disk);
7313 }
7314 spin_unlock_irq(&q->queue_lock);
7315
7316 INIT_LIST_HEAD(&bfqd->dispatch);
7317
7318 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
7319 HRTIMER_MODE_REL);
7320 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
7321
7322 bfqd->queue_weights_tree = RB_ROOT_CACHED;
7323#ifdef CONFIG_BFQ_GROUP_IOSCHED
7324 bfqd->num_groups_with_pending_reqs = 0;
7325#endif
7326
7327 INIT_LIST_HEAD(&bfqd->active_list[0]);
7328 INIT_LIST_HEAD(&bfqd->active_list[1]);
7329 INIT_LIST_HEAD(&bfqd->idle_list);
7330 INIT_HLIST_HEAD(&bfqd->burst_list);
7331
7332 bfqd->hw_tag = -1;
7333 bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
7334
7335 bfqd->bfq_max_budget = bfq_default_max_budget;
7336
7337 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
7338 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
7339 bfqd->bfq_back_max = bfq_back_max;
7340 bfqd->bfq_back_penalty = bfq_back_penalty;
7341 bfqd->bfq_slice_idle = bfq_slice_idle;
7342 bfqd->bfq_timeout = bfq_timeout;
7343
7344 bfqd->bfq_large_burst_thresh = 8;
7345 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
7346
7347 bfqd->low_latency = true;
7348
7349 /*
7350 * Trade-off between responsiveness and fairness.
7351 */
7352 bfqd->bfq_wr_coeff = 30;
7353 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
7354 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
7355 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
7356 bfqd->bfq_wr_max_softrt_rate = 7000; /*
7357 * Approximate rate required
7358 * to playback or record a
7359 * high-definition compressed
7360 * video.
7361 */
7362 bfqd->wr_busy_queues = 0;
7363
7364 /*
7365 * Begin by assuming, optimistically, that the device peak
7366 * rate is equal to 2/3 of the highest reference rate.
7367 */
7368 bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
7369 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
7370 bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
7371
7372 /* see comments on the definition of next field inside bfq_data */
7373 bfqd->actuator_load_threshold = 4;
7374
7375 spin_lock_init(&bfqd->lock);
7376
7377 /*
7378 * The invocation of the next bfq_create_group_hierarchy
7379 * function is the head of a chain of function calls
7380 * (bfq_create_group_hierarchy->blkcg_activate_policy->
7381 * blk_mq_freeze_queue) that may lead to the invocation of the
7382 * has_work hook function. For this reason,
7383 * bfq_create_group_hierarchy is invoked only after all
7384 * scheduler data has been initialized, apart from the fields
7385 * that can be initialized only after invoking
7386 * bfq_create_group_hierarchy. This, in particular, enables
7387 * has_work to correctly return false. Of course, to avoid
7388 * other inconsistencies, the blk-mq stack must then refrain
7389 * from invoking further scheduler hooks before this init
7390 * function is finished.
7391 */
7392 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
7393 if (!bfqd->root_group)
7394 goto out_free;
7395 bfq_init_root_group(bfqd->root_group, bfqd);
7396 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
7397
7398 /* We dispatch from request queue wide instead of hw queue */
7399 blk_queue_flag_set(QUEUE_FLAG_SQ_SCHED, q);
7400
7401 set_bit(ELEVATOR_FLAG_DISABLE_WBT, &eq->flags);
7402 wbt_disable_default(q->disk);
7403 blk_stat_enable_accounting(q);
7404
7405 return 0;
7406
7407out_free:
7408 kfree(bfqd);
7409 kobject_put(&eq->kobj);
7410 return -ENOMEM;
7411}
7412
7413static void bfq_slab_kill(void)
7414{
7415 kmem_cache_destroy(bfq_pool);
7416}
7417
7418static int __init bfq_slab_setup(void)
7419{
7420 bfq_pool = KMEM_CACHE(bfq_queue, 0);
7421 if (!bfq_pool)
7422 return -ENOMEM;
7423 return 0;
7424}
7425
7426static ssize_t bfq_var_show(unsigned int var, char *page)
7427{
7428 return sprintf(page, "%u\n", var);
7429}
7430
7431static int bfq_var_store(unsigned long *var, const char *page)
7432{
7433 unsigned long new_val;
7434 int ret = kstrtoul(page, 10, &new_val);
7435
7436 if (ret)
7437 return ret;
7438 *var = new_val;
7439 return 0;
7440}
7441
7442#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
7443static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7444{ \
7445 struct bfq_data *bfqd = e->elevator_data; \
7446 u64 __data = __VAR; \
7447 if (__CONV == 1) \
7448 __data = jiffies_to_msecs(__data); \
7449 else if (__CONV == 2) \
7450 __data = div_u64(__data, NSEC_PER_MSEC); \
7451 return bfq_var_show(__data, (page)); \
7452}
7453SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
7454SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
7455SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
7456SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
7457SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
7458SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
7459SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
7460SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
7461SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
7462#undef SHOW_FUNCTION
7463
7464#define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
7465static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7466{ \
7467 struct bfq_data *bfqd = e->elevator_data; \
7468 u64 __data = __VAR; \
7469 __data = div_u64(__data, NSEC_PER_USEC); \
7470 return bfq_var_show(__data, (page)); \
7471}
7472USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
7473#undef USEC_SHOW_FUNCTION
7474
7475#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
7476static ssize_t \
7477__FUNC(struct elevator_queue *e, const char *page, size_t count) \
7478{ \
7479 struct bfq_data *bfqd = e->elevator_data; \
7480 unsigned long __data, __min = (MIN), __max = (MAX); \
7481 int ret; \
7482 \
7483 ret = bfq_var_store(&__data, (page)); \
7484 if (ret) \
7485 return ret; \
7486 if (__data < __min) \
7487 __data = __min; \
7488 else if (__data > __max) \
7489 __data = __max; \
7490 if (__CONV == 1) \
7491 *(__PTR) = msecs_to_jiffies(__data); \
7492 else if (__CONV == 2) \
7493 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
7494 else \
7495 *(__PTR) = __data; \
7496 return count; \
7497}
7498STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
7499 INT_MAX, 2);
7500STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
7501 INT_MAX, 2);
7502STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
7503STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
7504 INT_MAX, 0);
7505STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
7506#undef STORE_FUNCTION
7507
7508#define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
7509static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
7510{ \
7511 struct bfq_data *bfqd = e->elevator_data; \
7512 unsigned long __data, __min = (MIN), __max = (MAX); \
7513 int ret; \
7514 \
7515 ret = bfq_var_store(&__data, (page)); \
7516 if (ret) \
7517 return ret; \
7518 if (__data < __min) \
7519 __data = __min; \
7520 else if (__data > __max) \
7521 __data = __max; \
7522 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
7523 return count; \
7524}
7525USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
7526 UINT_MAX);
7527#undef USEC_STORE_FUNCTION
7528
7529static ssize_t bfq_max_budget_store(struct elevator_queue *e,
7530 const char *page, size_t count)
7531{
7532 struct bfq_data *bfqd = e->elevator_data;
7533 unsigned long __data;
7534 int ret;
7535
7536 ret = bfq_var_store(&__data, (page));
7537 if (ret)
7538 return ret;
7539
7540 if (__data == 0)
7541 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7542 else {
7543 if (__data > INT_MAX)
7544 __data = INT_MAX;
7545 bfqd->bfq_max_budget = __data;
7546 }
7547
7548 bfqd->bfq_user_max_budget = __data;
7549
7550 return count;
7551}
7552
7553/*
7554 * Leaving this name to preserve name compatibility with cfq
7555 * parameters, but this timeout is used for both sync and async.
7556 */
7557static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
7558 const char *page, size_t count)
7559{
7560 struct bfq_data *bfqd = e->elevator_data;
7561 unsigned long __data;
7562 int ret;
7563
7564 ret = bfq_var_store(&__data, (page));
7565 if (ret)
7566 return ret;
7567
7568 if (__data < 1)
7569 __data = 1;
7570 else if (__data > INT_MAX)
7571 __data = INT_MAX;
7572
7573 bfqd->bfq_timeout = msecs_to_jiffies(__data);
7574 if (bfqd->bfq_user_max_budget == 0)
7575 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7576
7577 return count;
7578}
7579
7580static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
7581 const char *page, size_t count)
7582{
7583 struct bfq_data *bfqd = e->elevator_data;
7584 unsigned long __data;
7585 int ret;
7586
7587 ret = bfq_var_store(&__data, (page));
7588 if (ret)
7589 return ret;
7590
7591 if (__data > 1)
7592 __data = 1;
7593 if (!bfqd->strict_guarantees && __data == 1
7594 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
7595 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
7596
7597 bfqd->strict_guarantees = __data;
7598
7599 return count;
7600}
7601
7602static ssize_t bfq_low_latency_store(struct elevator_queue *e,
7603 const char *page, size_t count)
7604{
7605 struct bfq_data *bfqd = e->elevator_data;
7606 unsigned long __data;
7607 int ret;
7608
7609 ret = bfq_var_store(&__data, (page));
7610 if (ret)
7611 return ret;
7612
7613 if (__data > 1)
7614 __data = 1;
7615 if (__data == 0 && bfqd->low_latency != 0)
7616 bfq_end_wr(bfqd);
7617 bfqd->low_latency = __data;
7618
7619 return count;
7620}
7621
7622#define BFQ_ATTR(name) \
7623 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
7624
7625static struct elv_fs_entry bfq_attrs[] = {
7626 BFQ_ATTR(fifo_expire_sync),
7627 BFQ_ATTR(fifo_expire_async),
7628 BFQ_ATTR(back_seek_max),
7629 BFQ_ATTR(back_seek_penalty),
7630 BFQ_ATTR(slice_idle),
7631 BFQ_ATTR(slice_idle_us),
7632 BFQ_ATTR(max_budget),
7633 BFQ_ATTR(timeout_sync),
7634 BFQ_ATTR(strict_guarantees),
7635 BFQ_ATTR(low_latency),
7636 __ATTR_NULL
7637};
7638
7639static struct elevator_type iosched_bfq_mq = {
7640 .ops = {
7641 .limit_depth = bfq_limit_depth,
7642 .prepare_request = bfq_prepare_request,
7643 .requeue_request = bfq_finish_requeue_request,
7644 .finish_request = bfq_finish_request,
7645 .exit_icq = bfq_exit_icq,
7646 .insert_requests = bfq_insert_requests,
7647 .dispatch_request = bfq_dispatch_request,
7648 .next_request = elv_rb_latter_request,
7649 .former_request = elv_rb_former_request,
7650 .allow_merge = bfq_allow_bio_merge,
7651 .bio_merge = bfq_bio_merge,
7652 .request_merge = bfq_request_merge,
7653 .requests_merged = bfq_requests_merged,
7654 .request_merged = bfq_request_merged,
7655 .has_work = bfq_has_work,
7656 .depth_updated = bfq_depth_updated,
7657 .init_hctx = bfq_init_hctx,
7658 .init_sched = bfq_init_queue,
7659 .exit_sched = bfq_exit_queue,
7660 },
7661
7662 .icq_size = sizeof(struct bfq_io_cq),
7663 .icq_align = __alignof__(struct bfq_io_cq),
7664 .elevator_attrs = bfq_attrs,
7665 .elevator_name = "bfq",
7666 .elevator_owner = THIS_MODULE,
7667};
7668MODULE_ALIAS("bfq-iosched");
7669
7670static int __init bfq_init(void)
7671{
7672 int ret;
7673
7674#ifdef CONFIG_BFQ_GROUP_IOSCHED
7675 ret = blkcg_policy_register(&blkcg_policy_bfq);
7676 if (ret)
7677 return ret;
7678#endif
7679
7680 ret = -ENOMEM;
7681 if (bfq_slab_setup())
7682 goto err_pol_unreg;
7683
7684 /*
7685 * Times to load large popular applications for the typical
7686 * systems installed on the reference devices (see the
7687 * comments before the definition of the next
7688 * array). Actually, we use slightly lower values, as the
7689 * estimated peak rate tends to be smaller than the actual
7690 * peak rate. The reason for this last fact is that estimates
7691 * are computed over much shorter time intervals than the long
7692 * intervals typically used for benchmarking. Why? First, to
7693 * adapt more quickly to variations. Second, because an I/O
7694 * scheduler cannot rely on a peak-rate-evaluation workload to
7695 * be run for a long time.
7696 */
7697 ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
7698 ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
7699
7700 ret = elv_register(&iosched_bfq_mq);
7701 if (ret)
7702 goto slab_kill;
7703
7704 return 0;
7705
7706slab_kill:
7707 bfq_slab_kill();
7708err_pol_unreg:
7709#ifdef CONFIG_BFQ_GROUP_IOSCHED
7710 blkcg_policy_unregister(&blkcg_policy_bfq);
7711#endif
7712 return ret;
7713}
7714
7715static void __exit bfq_exit(void)
7716{
7717 elv_unregister(&iosched_bfq_mq);
7718#ifdef CONFIG_BFQ_GROUP_IOSCHED
7719 blkcg_policy_unregister(&blkcg_policy_bfq);
7720#endif
7721 bfq_slab_kill();
7722}
7723
7724module_init(bfq_init);
7725module_exit(bfq_exit);
7726
7727MODULE_AUTHOR("Paolo Valente");
7728MODULE_LICENSE("GPL");
7729MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");