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/elevator.h>
121#include <linux/ktime.h>
122#include <linux/rbtree.h>
123#include <linux/ioprio.h>
124#include <linux/sbitmap.h>
125#include <linux/delay.h>
126
127#include "blk.h"
128#include "blk-mq.h"
129#include "blk-mq-tag.h"
130#include "blk-mq-sched.h"
131#include "bfq-iosched.h"
132#include "blk-wbt.h"
133
134#define BFQ_BFQQ_FNS(name) \
135void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
136{ \
137 __set_bit(BFQQF_##name, &(bfqq)->flags); \
138} \
139void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
140{ \
141 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
142} \
143int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
144{ \
145 return test_bit(BFQQF_##name, &(bfqq)->flags); \
146}
147
148BFQ_BFQQ_FNS(just_created);
149BFQ_BFQQ_FNS(busy);
150BFQ_BFQQ_FNS(wait_request);
151BFQ_BFQQ_FNS(non_blocking_wait_rq);
152BFQ_BFQQ_FNS(fifo_expire);
153BFQ_BFQQ_FNS(has_short_ttime);
154BFQ_BFQQ_FNS(sync);
155BFQ_BFQQ_FNS(IO_bound);
156BFQ_BFQQ_FNS(in_large_burst);
157BFQ_BFQQ_FNS(coop);
158BFQ_BFQQ_FNS(split_coop);
159BFQ_BFQQ_FNS(softrt_update);
160BFQ_BFQQ_FNS(has_waker);
161#undef BFQ_BFQQ_FNS \
162
163/* Expiration time of sync (0) and async (1) requests, in ns. */
164static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
165
166/* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
167static const int bfq_back_max = 16 * 1024;
168
169/* Penalty of a backwards seek, in number of sectors. */
170static const int bfq_back_penalty = 2;
171
172/* Idling period duration, in ns. */
173static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
174
175/* Minimum number of assigned budgets for which stats are safe to compute. */
176static const int bfq_stats_min_budgets = 194;
177
178/* Default maximum budget values, in sectors and number of requests. */
179static const int bfq_default_max_budget = 16 * 1024;
180
181/*
182 * When a sync request is dispatched, the queue that contains that
183 * request, and all the ancestor entities of that queue, are charged
184 * with the number of sectors of the request. In contrast, if the
185 * request is async, then the queue and its ancestor entities are
186 * charged with the number of sectors of the request, multiplied by
187 * the factor below. This throttles the bandwidth for async I/O,
188 * w.r.t. to sync I/O, and it is done to counter the tendency of async
189 * writes to steal I/O throughput to reads.
190 *
191 * The current value of this parameter is the result of a tuning with
192 * several hardware and software configurations. We tried to find the
193 * lowest value for which writes do not cause noticeable problems to
194 * reads. In fact, the lower this parameter, the stabler I/O control,
195 * in the following respect. The lower this parameter is, the less
196 * the bandwidth enjoyed by a group decreases
197 * - when the group does writes, w.r.t. to when it does reads;
198 * - when other groups do reads, w.r.t. to when they do writes.
199 */
200static const int bfq_async_charge_factor = 3;
201
202/* Default timeout values, in jiffies, approximating CFQ defaults. */
203const int bfq_timeout = HZ / 8;
204
205/*
206 * Time limit for merging (see comments in bfq_setup_cooperator). Set
207 * to the slowest value that, in our tests, proved to be effective in
208 * removing false positives, while not causing true positives to miss
209 * queue merging.
210 *
211 * As can be deduced from the low time limit below, queue merging, if
212 * successful, happens at the very beginning of the I/O of the involved
213 * cooperating processes, as a consequence of the arrival of the very
214 * first requests from each cooperator. After that, there is very
215 * little chance to find cooperators.
216 */
217static const unsigned long bfq_merge_time_limit = HZ/10;
218
219static struct kmem_cache *bfq_pool;
220
221/* Below this threshold (in ns), we consider thinktime immediate. */
222#define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
223
224/* hw_tag detection: parallel requests threshold and min samples needed. */
225#define BFQ_HW_QUEUE_THRESHOLD 3
226#define BFQ_HW_QUEUE_SAMPLES 32
227
228#define BFQQ_SEEK_THR (sector_t)(8 * 100)
229#define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
230#define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \
231 (get_sdist(last_pos, rq) > \
232 BFQQ_SEEK_THR && \
233 (!blk_queue_nonrot(bfqd->queue) || \
234 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT))
235#define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
236#define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19)
237/*
238 * Sync random I/O is likely to be confused with soft real-time I/O,
239 * because it is characterized by limited throughput and apparently
240 * isochronous arrival pattern. To avoid false positives, queues
241 * containing only random (seeky) I/O are prevented from being tagged
242 * as soft real-time.
243 */
244#define BFQQ_TOTALLY_SEEKY(bfqq) (bfqq->seek_history == -1)
245
246/* Min number of samples required to perform peak-rate update */
247#define BFQ_RATE_MIN_SAMPLES 32
248/* Min observation time interval required to perform a peak-rate update (ns) */
249#define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
250/* Target observation time interval for a peak-rate update (ns) */
251#define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
252
253/*
254 * Shift used for peak-rate fixed precision calculations.
255 * With
256 * - the current shift: 16 positions
257 * - the current type used to store rate: u32
258 * - the current unit of measure for rate: [sectors/usec], or, more precisely,
259 * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
260 * the range of rates that can be stored is
261 * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
262 * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
263 * [15, 65G] sectors/sec
264 * Which, assuming a sector size of 512B, corresponds to a range of
265 * [7.5K, 33T] B/sec
266 */
267#define BFQ_RATE_SHIFT 16
268
269/*
270 * When configured for computing the duration of the weight-raising
271 * for interactive queues automatically (see the comments at the
272 * beginning of this file), BFQ does it using the following formula:
273 * duration = (ref_rate / r) * ref_wr_duration,
274 * where r is the peak rate of the device, and ref_rate and
275 * ref_wr_duration are two reference parameters. In particular,
276 * ref_rate is the peak rate of the reference storage device (see
277 * below), and ref_wr_duration is about the maximum time needed, with
278 * BFQ and while reading two files in parallel, to load typical large
279 * applications on the reference device (see the comments on
280 * max_service_from_wr below, for more details on how ref_wr_duration
281 * is obtained). In practice, the slower/faster the device at hand
282 * is, the more/less it takes to load applications with respect to the
283 * reference device. Accordingly, the longer/shorter BFQ grants
284 * weight raising to interactive applications.
285 *
286 * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
287 * depending on whether the device is rotational or non-rotational.
288 *
289 * In the following definitions, ref_rate[0] and ref_wr_duration[0]
290 * are the reference values for a rotational device, whereas
291 * ref_rate[1] and ref_wr_duration[1] are the reference values for a
292 * non-rotational device. The reference rates are not the actual peak
293 * rates of the devices used as a reference, but slightly lower
294 * values. The reason for using slightly lower values is that the
295 * peak-rate estimator tends to yield slightly lower values than the
296 * actual peak rate (it can yield the actual peak rate only if there
297 * is only one process doing I/O, and the process does sequential
298 * I/O).
299 *
300 * The reference peak rates are measured in sectors/usec, left-shifted
301 * by BFQ_RATE_SHIFT.
302 */
303static int ref_rate[2] = {14000, 33000};
304/*
305 * To improve readability, a conversion function is used to initialize
306 * the following array, which entails that the array can be
307 * initialized only in a function.
308 */
309static int ref_wr_duration[2];
310
311/*
312 * BFQ uses the above-detailed, time-based weight-raising mechanism to
313 * privilege interactive tasks. This mechanism is vulnerable to the
314 * following false positives: I/O-bound applications that will go on
315 * doing I/O for much longer than the duration of weight
316 * raising. These applications have basically no benefit from being
317 * weight-raised at the beginning of their I/O. On the opposite end,
318 * while being weight-raised, these applications
319 * a) unjustly steal throughput to applications that may actually need
320 * low latency;
321 * b) make BFQ uselessly perform device idling; device idling results
322 * in loss of device throughput with most flash-based storage, and may
323 * increase latencies when used purposelessly.
324 *
325 * BFQ tries to reduce these problems, by adopting the following
326 * countermeasure. To introduce this countermeasure, we need first to
327 * finish explaining how the duration of weight-raising for
328 * interactive tasks is computed.
329 *
330 * For a bfq_queue deemed as interactive, the duration of weight
331 * raising is dynamically adjusted, as a function of the estimated
332 * peak rate of the device, so as to be equal to the time needed to
333 * execute the 'largest' interactive task we benchmarked so far. By
334 * largest task, we mean the task for which each involved process has
335 * to do more I/O than for any of the other tasks we benchmarked. This
336 * reference interactive task is the start-up of LibreOffice Writer,
337 * and in this task each process/bfq_queue needs to have at most ~110K
338 * sectors transferred.
339 *
340 * This last piece of information enables BFQ to reduce the actual
341 * duration of weight-raising for at least one class of I/O-bound
342 * applications: those doing sequential or quasi-sequential I/O. An
343 * example is file copy. In fact, once started, the main I/O-bound
344 * processes of these applications usually consume the above 110K
345 * sectors in much less time than the processes of an application that
346 * is starting, because these I/O-bound processes will greedily devote
347 * almost all their CPU cycles only to their target,
348 * throughput-friendly I/O operations. This is even more true if BFQ
349 * happens to be underestimating the device peak rate, and thus
350 * overestimating the duration of weight raising. But, according to
351 * our measurements, once transferred 110K sectors, these processes
352 * have no right to be weight-raised any longer.
353 *
354 * Basing on the last consideration, BFQ ends weight-raising for a
355 * bfq_queue if the latter happens to have received an amount of
356 * service at least equal to the following constant. The constant is
357 * set to slightly more than 110K, to have a minimum safety margin.
358 *
359 * This early ending of weight-raising reduces the amount of time
360 * during which interactive false positives cause the two problems
361 * described at the beginning of these comments.
362 */
363static const unsigned long max_service_from_wr = 120000;
364
365#define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
366#define RQ_BFQQ(rq) ((rq)->elv.priv[1])
367
368struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
369{
370 return bic->bfqq[is_sync];
371}
372
373void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
374{
375 bic->bfqq[is_sync] = bfqq;
376}
377
378struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
379{
380 return bic->icq.q->elevator->elevator_data;
381}
382
383/**
384 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
385 * @icq: the iocontext queue.
386 */
387static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
388{
389 /* bic->icq is the first member, %NULL will convert to %NULL */
390 return container_of(icq, struct bfq_io_cq, icq);
391}
392
393/**
394 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
395 * @bfqd: the lookup key.
396 * @ioc: the io_context of the process doing I/O.
397 * @q: the request queue.
398 */
399static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd,
400 struct io_context *ioc,
401 struct request_queue *q)
402{
403 if (ioc) {
404 unsigned long flags;
405 struct bfq_io_cq *icq;
406
407 spin_lock_irqsave(&q->queue_lock, flags);
408 icq = icq_to_bic(ioc_lookup_icq(ioc, q));
409 spin_unlock_irqrestore(&q->queue_lock, flags);
410
411 return icq;
412 }
413
414 return NULL;
415}
416
417/*
418 * Scheduler run of queue, if there are requests pending and no one in the
419 * driver that will restart queueing.
420 */
421void bfq_schedule_dispatch(struct bfq_data *bfqd)
422{
423 if (bfqd->queued != 0) {
424 bfq_log(bfqd, "schedule dispatch");
425 blk_mq_run_hw_queues(bfqd->queue, true);
426 }
427}
428
429#define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
430#define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT)
431
432#define bfq_sample_valid(samples) ((samples) > 80)
433
434/*
435 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
436 * We choose the request that is closer to the head right now. Distance
437 * behind the head is penalized and only allowed to a certain extent.
438 */
439static struct request *bfq_choose_req(struct bfq_data *bfqd,
440 struct request *rq1,
441 struct request *rq2,
442 sector_t last)
443{
444 sector_t s1, s2, d1 = 0, d2 = 0;
445 unsigned long back_max;
446#define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
447#define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
448 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
449
450 if (!rq1 || rq1 == rq2)
451 return rq2;
452 if (!rq2)
453 return rq1;
454
455 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
456 return rq1;
457 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
458 return rq2;
459 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
460 return rq1;
461 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
462 return rq2;
463
464 s1 = blk_rq_pos(rq1);
465 s2 = blk_rq_pos(rq2);
466
467 /*
468 * By definition, 1KiB is 2 sectors.
469 */
470 back_max = bfqd->bfq_back_max * 2;
471
472 /*
473 * Strict one way elevator _except_ in the case where we allow
474 * short backward seeks which are biased as twice the cost of a
475 * similar forward seek.
476 */
477 if (s1 >= last)
478 d1 = s1 - last;
479 else if (s1 + back_max >= last)
480 d1 = (last - s1) * bfqd->bfq_back_penalty;
481 else
482 wrap |= BFQ_RQ1_WRAP;
483
484 if (s2 >= last)
485 d2 = s2 - last;
486 else if (s2 + back_max >= last)
487 d2 = (last - s2) * bfqd->bfq_back_penalty;
488 else
489 wrap |= BFQ_RQ2_WRAP;
490
491 /* Found required data */
492
493 /*
494 * By doing switch() on the bit mask "wrap" we avoid having to
495 * check two variables for all permutations: --> faster!
496 */
497 switch (wrap) {
498 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
499 if (d1 < d2)
500 return rq1;
501 else if (d2 < d1)
502 return rq2;
503
504 if (s1 >= s2)
505 return rq1;
506 else
507 return rq2;
508
509 case BFQ_RQ2_WRAP:
510 return rq1;
511 case BFQ_RQ1_WRAP:
512 return rq2;
513 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
514 default:
515 /*
516 * Since both rqs are wrapped,
517 * start with the one that's further behind head
518 * (--> only *one* back seek required),
519 * since back seek takes more time than forward.
520 */
521 if (s1 <= s2)
522 return rq1;
523 else
524 return rq2;
525 }
526}
527
528/*
529 * Async I/O can easily starve sync I/O (both sync reads and sync
530 * writes), by consuming all tags. Similarly, storms of sync writes,
531 * such as those that sync(2) may trigger, can starve sync reads.
532 * Limit depths of async I/O and sync writes so as to counter both
533 * problems.
534 */
535static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
536{
537 struct bfq_data *bfqd = data->q->elevator->elevator_data;
538
539 if (op_is_sync(op) && !op_is_write(op))
540 return;
541
542 data->shallow_depth =
543 bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)];
544
545 bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
546 __func__, bfqd->wr_busy_queues, op_is_sync(op),
547 data->shallow_depth);
548}
549
550static struct bfq_queue *
551bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
552 sector_t sector, struct rb_node **ret_parent,
553 struct rb_node ***rb_link)
554{
555 struct rb_node **p, *parent;
556 struct bfq_queue *bfqq = NULL;
557
558 parent = NULL;
559 p = &root->rb_node;
560 while (*p) {
561 struct rb_node **n;
562
563 parent = *p;
564 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
565
566 /*
567 * Sort strictly based on sector. Smallest to the left,
568 * largest to the right.
569 */
570 if (sector > blk_rq_pos(bfqq->next_rq))
571 n = &(*p)->rb_right;
572 else if (sector < blk_rq_pos(bfqq->next_rq))
573 n = &(*p)->rb_left;
574 else
575 break;
576 p = n;
577 bfqq = NULL;
578 }
579
580 *ret_parent = parent;
581 if (rb_link)
582 *rb_link = p;
583
584 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
585 (unsigned long long)sector,
586 bfqq ? bfqq->pid : 0);
587
588 return bfqq;
589}
590
591static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
592{
593 return bfqq->service_from_backlogged > 0 &&
594 time_is_before_jiffies(bfqq->first_IO_time +
595 bfq_merge_time_limit);
596}
597
598/*
599 * The following function is not marked as __cold because it is
600 * actually cold, but for the same performance goal described in the
601 * comments on the likely() at the beginning of
602 * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
603 * execution time for the case where this function is not invoked, we
604 * had to add an unlikely() in each involved if().
605 */
606void __cold
607bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
608{
609 struct rb_node **p, *parent;
610 struct bfq_queue *__bfqq;
611
612 if (bfqq->pos_root) {
613 rb_erase(&bfqq->pos_node, bfqq->pos_root);
614 bfqq->pos_root = NULL;
615 }
616
617 /*
618 * bfqq cannot be merged any longer (see comments in
619 * bfq_setup_cooperator): no point in adding bfqq into the
620 * position tree.
621 */
622 if (bfq_too_late_for_merging(bfqq))
623 return;
624
625 if (bfq_class_idle(bfqq))
626 return;
627 if (!bfqq->next_rq)
628 return;
629
630 bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
631 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
632 blk_rq_pos(bfqq->next_rq), &parent, &p);
633 if (!__bfqq) {
634 rb_link_node(&bfqq->pos_node, parent, p);
635 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
636 } else
637 bfqq->pos_root = NULL;
638}
639
640/*
641 * The following function returns false either if every active queue
642 * must receive the same share of the throughput (symmetric scenario),
643 * or, as a special case, if bfqq must receive a share of the
644 * throughput lower than or equal to the share that every other active
645 * queue must receive. If bfqq does sync I/O, then these are the only
646 * two cases where bfqq happens to be guaranteed its share of the
647 * throughput even if I/O dispatching is not plugged when bfqq remains
648 * temporarily empty (for more details, see the comments in the
649 * function bfq_better_to_idle()). For this reason, the return value
650 * of this function is used to check whether I/O-dispatch plugging can
651 * be avoided.
652 *
653 * The above first case (symmetric scenario) occurs when:
654 * 1) all active queues have the same weight,
655 * 2) all active queues belong to the same I/O-priority class,
656 * 3) all active groups at the same level in the groups tree have the same
657 * weight,
658 * 4) all active groups at the same level in the groups tree have the same
659 * number of children.
660 *
661 * Unfortunately, keeping the necessary state for evaluating exactly
662 * the last two symmetry sub-conditions above would be quite complex
663 * and time consuming. Therefore this function evaluates, instead,
664 * only the following stronger three sub-conditions, for which it is
665 * much easier to maintain the needed state:
666 * 1) all active queues have the same weight,
667 * 2) all active queues belong to the same I/O-priority class,
668 * 3) there are no active groups.
669 * In particular, the last condition is always true if hierarchical
670 * support or the cgroups interface are not enabled, thus no state
671 * needs to be maintained in this case.
672 */
673static bool bfq_asymmetric_scenario(struct bfq_data *bfqd,
674 struct bfq_queue *bfqq)
675{
676 bool smallest_weight = bfqq &&
677 bfqq->weight_counter &&
678 bfqq->weight_counter ==
679 container_of(
680 rb_first_cached(&bfqd->queue_weights_tree),
681 struct bfq_weight_counter,
682 weights_node);
683
684 /*
685 * For queue weights to differ, queue_weights_tree must contain
686 * at least two nodes.
687 */
688 bool varied_queue_weights = !smallest_weight &&
689 !RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) &&
690 (bfqd->queue_weights_tree.rb_root.rb_node->rb_left ||
691 bfqd->queue_weights_tree.rb_root.rb_node->rb_right);
692
693 bool multiple_classes_busy =
694 (bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
695 (bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
696 (bfqd->busy_queues[1] && bfqd->busy_queues[2]);
697
698 return varied_queue_weights || multiple_classes_busy
699#ifdef CONFIG_BFQ_GROUP_IOSCHED
700 || bfqd->num_groups_with_pending_reqs > 0
701#endif
702 ;
703}
704
705/*
706 * If the weight-counter tree passed as input contains no counter for
707 * the weight of the input queue, then add that counter; otherwise just
708 * increment the existing counter.
709 *
710 * Note that weight-counter trees contain few nodes in mostly symmetric
711 * scenarios. For example, if all queues have the same weight, then the
712 * weight-counter tree for the queues may contain at most one node.
713 * This holds even if low_latency is on, because weight-raised queues
714 * are not inserted in the tree.
715 * In most scenarios, the rate at which nodes are created/destroyed
716 * should be low too.
717 */
718void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq,
719 struct rb_root_cached *root)
720{
721 struct bfq_entity *entity = &bfqq->entity;
722 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
723 bool leftmost = true;
724
725 /*
726 * Do not insert if the queue is already associated with a
727 * counter, which happens if:
728 * 1) a request arrival has caused the queue to become both
729 * non-weight-raised, and hence change its weight, and
730 * backlogged; in this respect, each of the two events
731 * causes an invocation of this function,
732 * 2) this is the invocation of this function caused by the
733 * second event. This second invocation is actually useless,
734 * and we handle this fact by exiting immediately. More
735 * efficient or clearer solutions might possibly be adopted.
736 */
737 if (bfqq->weight_counter)
738 return;
739
740 while (*new) {
741 struct bfq_weight_counter *__counter = container_of(*new,
742 struct bfq_weight_counter,
743 weights_node);
744 parent = *new;
745
746 if (entity->weight == __counter->weight) {
747 bfqq->weight_counter = __counter;
748 goto inc_counter;
749 }
750 if (entity->weight < __counter->weight)
751 new = &((*new)->rb_left);
752 else {
753 new = &((*new)->rb_right);
754 leftmost = false;
755 }
756 }
757
758 bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
759 GFP_ATOMIC);
760
761 /*
762 * In the unlucky event of an allocation failure, we just
763 * exit. This will cause the weight of queue to not be
764 * considered in bfq_asymmetric_scenario, which, in its turn,
765 * causes the scenario to be deemed wrongly symmetric in case
766 * bfqq's weight would have been the only weight making the
767 * scenario asymmetric. On the bright side, no unbalance will
768 * however occur when bfqq becomes inactive again (the
769 * invocation of this function is triggered by an activation
770 * of queue). In fact, bfq_weights_tree_remove does nothing
771 * if !bfqq->weight_counter.
772 */
773 if (unlikely(!bfqq->weight_counter))
774 return;
775
776 bfqq->weight_counter->weight = entity->weight;
777 rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
778 rb_insert_color_cached(&bfqq->weight_counter->weights_node, root,
779 leftmost);
780
781inc_counter:
782 bfqq->weight_counter->num_active++;
783 bfqq->ref++;
784}
785
786/*
787 * Decrement the weight counter associated with the queue, and, if the
788 * counter reaches 0, remove the counter from the tree.
789 * See the comments to the function bfq_weights_tree_add() for considerations
790 * about overhead.
791 */
792void __bfq_weights_tree_remove(struct bfq_data *bfqd,
793 struct bfq_queue *bfqq,
794 struct rb_root_cached *root)
795{
796 if (!bfqq->weight_counter)
797 return;
798
799 bfqq->weight_counter->num_active--;
800 if (bfqq->weight_counter->num_active > 0)
801 goto reset_entity_pointer;
802
803 rb_erase_cached(&bfqq->weight_counter->weights_node, root);
804 kfree(bfqq->weight_counter);
805
806reset_entity_pointer:
807 bfqq->weight_counter = NULL;
808 bfq_put_queue(bfqq);
809}
810
811/*
812 * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
813 * of active groups for each queue's inactive parent entity.
814 */
815void bfq_weights_tree_remove(struct bfq_data *bfqd,
816 struct bfq_queue *bfqq)
817{
818 struct bfq_entity *entity = bfqq->entity.parent;
819
820 for_each_entity(entity) {
821 struct bfq_sched_data *sd = entity->my_sched_data;
822
823 if (sd->next_in_service || sd->in_service_entity) {
824 /*
825 * entity is still active, because either
826 * next_in_service or in_service_entity is not
827 * NULL (see the comments on the definition of
828 * next_in_service for details on why
829 * in_service_entity must be checked too).
830 *
831 * As a consequence, its parent entities are
832 * active as well, and thus this loop must
833 * stop here.
834 */
835 break;
836 }
837
838 /*
839 * The decrement of num_groups_with_pending_reqs is
840 * not performed immediately upon the deactivation of
841 * entity, but it is delayed to when it also happens
842 * that the first leaf descendant bfqq of entity gets
843 * all its pending requests completed. The following
844 * instructions perform this delayed decrement, if
845 * needed. See the comments on
846 * num_groups_with_pending_reqs for details.
847 */
848 if (entity->in_groups_with_pending_reqs) {
849 entity->in_groups_with_pending_reqs = false;
850 bfqd->num_groups_with_pending_reqs--;
851 }
852 }
853
854 /*
855 * Next function is invoked last, because it causes bfqq to be
856 * freed if the following holds: bfqq is not in service and
857 * has no dispatched request. DO NOT use bfqq after the next
858 * function invocation.
859 */
860 __bfq_weights_tree_remove(bfqd, bfqq,
861 &bfqd->queue_weights_tree);
862}
863
864/*
865 * Return expired entry, or NULL to just start from scratch in rbtree.
866 */
867static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
868 struct request *last)
869{
870 struct request *rq;
871
872 if (bfq_bfqq_fifo_expire(bfqq))
873 return NULL;
874
875 bfq_mark_bfqq_fifo_expire(bfqq);
876
877 rq = rq_entry_fifo(bfqq->fifo.next);
878
879 if (rq == last || ktime_get_ns() < rq->fifo_time)
880 return NULL;
881
882 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
883 return rq;
884}
885
886static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
887 struct bfq_queue *bfqq,
888 struct request *last)
889{
890 struct rb_node *rbnext = rb_next(&last->rb_node);
891 struct rb_node *rbprev = rb_prev(&last->rb_node);
892 struct request *next, *prev = NULL;
893
894 /* Follow expired path, else get first next available. */
895 next = bfq_check_fifo(bfqq, last);
896 if (next)
897 return next;
898
899 if (rbprev)
900 prev = rb_entry_rq(rbprev);
901
902 if (rbnext)
903 next = rb_entry_rq(rbnext);
904 else {
905 rbnext = rb_first(&bfqq->sort_list);
906 if (rbnext && rbnext != &last->rb_node)
907 next = rb_entry_rq(rbnext);
908 }
909
910 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
911}
912
913/* see the definition of bfq_async_charge_factor for details */
914static unsigned long bfq_serv_to_charge(struct request *rq,
915 struct bfq_queue *bfqq)
916{
917 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
918 bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
919 return blk_rq_sectors(rq);
920
921 return blk_rq_sectors(rq) * bfq_async_charge_factor;
922}
923
924/**
925 * bfq_updated_next_req - update the queue after a new next_rq selection.
926 * @bfqd: the device data the queue belongs to.
927 * @bfqq: the queue to update.
928 *
929 * If the first request of a queue changes we make sure that the queue
930 * has enough budget to serve at least its first request (if the
931 * request has grown). We do this because if the queue has not enough
932 * budget for its first request, it has to go through two dispatch
933 * rounds to actually get it dispatched.
934 */
935static void bfq_updated_next_req(struct bfq_data *bfqd,
936 struct bfq_queue *bfqq)
937{
938 struct bfq_entity *entity = &bfqq->entity;
939 struct request *next_rq = bfqq->next_rq;
940 unsigned long new_budget;
941
942 if (!next_rq)
943 return;
944
945 if (bfqq == bfqd->in_service_queue)
946 /*
947 * In order not to break guarantees, budgets cannot be
948 * changed after an entity has been selected.
949 */
950 return;
951
952 new_budget = max_t(unsigned long,
953 max_t(unsigned long, bfqq->max_budget,
954 bfq_serv_to_charge(next_rq, bfqq)),
955 entity->service);
956 if (entity->budget != new_budget) {
957 entity->budget = new_budget;
958 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
959 new_budget);
960 bfq_requeue_bfqq(bfqd, bfqq, false);
961 }
962}
963
964static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
965{
966 u64 dur;
967
968 if (bfqd->bfq_wr_max_time > 0)
969 return bfqd->bfq_wr_max_time;
970
971 dur = bfqd->rate_dur_prod;
972 do_div(dur, bfqd->peak_rate);
973
974 /*
975 * Limit duration between 3 and 25 seconds. The upper limit
976 * has been conservatively set after the following worst case:
977 * on a QEMU/KVM virtual machine
978 * - running in a slow PC
979 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
980 * - serving a heavy I/O workload, such as the sequential reading
981 * of several files
982 * mplayer took 23 seconds to start, if constantly weight-raised.
983 *
984 * As for higher values than that accommodating the above bad
985 * scenario, tests show that higher values would often yield
986 * the opposite of the desired result, i.e., would worsen
987 * responsiveness by allowing non-interactive applications to
988 * preserve weight raising for too long.
989 *
990 * On the other end, lower values than 3 seconds make it
991 * difficult for most interactive tasks to complete their jobs
992 * before weight-raising finishes.
993 */
994 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
995}
996
997/* switch back from soft real-time to interactive weight raising */
998static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
999 struct bfq_data *bfqd)
1000{
1001 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1002 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1003 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
1004}
1005
1006static void
1007bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
1008 struct bfq_io_cq *bic, bool bfq_already_existing)
1009{
1010 unsigned int old_wr_coeff = bfqq->wr_coeff;
1011 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
1012
1013 if (bic->saved_has_short_ttime)
1014 bfq_mark_bfqq_has_short_ttime(bfqq);
1015 else
1016 bfq_clear_bfqq_has_short_ttime(bfqq);
1017
1018 if (bic->saved_IO_bound)
1019 bfq_mark_bfqq_IO_bound(bfqq);
1020 else
1021 bfq_clear_bfqq_IO_bound(bfqq);
1022
1023 bfqq->entity.new_weight = bic->saved_weight;
1024 bfqq->ttime = bic->saved_ttime;
1025 bfqq->wr_coeff = bic->saved_wr_coeff;
1026 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
1027 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
1028 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
1029
1030 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1031 time_is_before_jiffies(bfqq->last_wr_start_finish +
1032 bfqq->wr_cur_max_time))) {
1033 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1034 !bfq_bfqq_in_large_burst(bfqq) &&
1035 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1036 bfq_wr_duration(bfqd))) {
1037 switch_back_to_interactive_wr(bfqq, bfqd);
1038 } else {
1039 bfqq->wr_coeff = 1;
1040 bfq_log_bfqq(bfqq->bfqd, bfqq,
1041 "resume state: switching off wr");
1042 }
1043 }
1044
1045 /* make sure weight will be updated, however we got here */
1046 bfqq->entity.prio_changed = 1;
1047
1048 if (likely(!busy))
1049 return;
1050
1051 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1052 bfqd->wr_busy_queues++;
1053 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1054 bfqd->wr_busy_queues--;
1055}
1056
1057static int bfqq_process_refs(struct bfq_queue *bfqq)
1058{
1059 return bfqq->ref - bfqq->allocated - bfqq->entity.on_st -
1060 (bfqq->weight_counter != NULL);
1061}
1062
1063/* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1064static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1065{
1066 struct bfq_queue *item;
1067 struct hlist_node *n;
1068
1069 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1070 hlist_del_init(&item->burst_list_node);
1071
1072 /*
1073 * Start the creation of a new burst list only if there is no
1074 * active queue. See comments on the conditional invocation of
1075 * bfq_handle_burst().
1076 */
1077 if (bfq_tot_busy_queues(bfqd) == 0) {
1078 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1079 bfqd->burst_size = 1;
1080 } else
1081 bfqd->burst_size = 0;
1082
1083 bfqd->burst_parent_entity = bfqq->entity.parent;
1084}
1085
1086/* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1087static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1088{
1089 /* Increment burst size to take into account also bfqq */
1090 bfqd->burst_size++;
1091
1092 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1093 struct bfq_queue *pos, *bfqq_item;
1094 struct hlist_node *n;
1095
1096 /*
1097 * Enough queues have been activated shortly after each
1098 * other to consider this burst as large.
1099 */
1100 bfqd->large_burst = true;
1101
1102 /*
1103 * We can now mark all queues in the burst list as
1104 * belonging to a large burst.
1105 */
1106 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1107 burst_list_node)
1108 bfq_mark_bfqq_in_large_burst(bfqq_item);
1109 bfq_mark_bfqq_in_large_burst(bfqq);
1110
1111 /*
1112 * From now on, and until the current burst finishes, any
1113 * new queue being activated shortly after the last queue
1114 * was inserted in the burst can be immediately marked as
1115 * belonging to a large burst. So the burst list is not
1116 * needed any more. Remove it.
1117 */
1118 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1119 burst_list_node)
1120 hlist_del_init(&pos->burst_list_node);
1121 } else /*
1122 * Burst not yet large: add bfqq to the burst list. Do
1123 * not increment the ref counter for bfqq, because bfqq
1124 * is removed from the burst list before freeing bfqq
1125 * in put_queue.
1126 */
1127 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1128}
1129
1130/*
1131 * If many queues belonging to the same group happen to be created
1132 * shortly after each other, then the processes associated with these
1133 * queues have typically a common goal. In particular, bursts of queue
1134 * creations are usually caused by services or applications that spawn
1135 * many parallel threads/processes. Examples are systemd during boot,
1136 * or git grep. To help these processes get their job done as soon as
1137 * possible, it is usually better to not grant either weight-raising
1138 * or device idling to their queues, unless these queues must be
1139 * protected from the I/O flowing through other active queues.
1140 *
1141 * In this comment we describe, firstly, the reasons why this fact
1142 * holds, and, secondly, the next function, which implements the main
1143 * steps needed to properly mark these queues so that they can then be
1144 * treated in a different way.
1145 *
1146 * The above services or applications benefit mostly from a high
1147 * throughput: the quicker the requests of the activated queues are
1148 * cumulatively served, the sooner the target job of these queues gets
1149 * completed. As a consequence, weight-raising any of these queues,
1150 * which also implies idling the device for it, is almost always
1151 * counterproductive, unless there are other active queues to isolate
1152 * these new queues from. If there no other active queues, then
1153 * weight-raising these new queues just lowers throughput in most
1154 * cases.
1155 *
1156 * On the other hand, a burst of queue creations may be caused also by
1157 * the start of an application that does not consist of a lot of
1158 * parallel I/O-bound threads. In fact, with a complex application,
1159 * several short processes may need to be executed to start-up the
1160 * application. In this respect, to start an application as quickly as
1161 * possible, the best thing to do is in any case to privilege the I/O
1162 * related to the application with respect to all other
1163 * I/O. Therefore, the best strategy to start as quickly as possible
1164 * an application that causes a burst of queue creations is to
1165 * weight-raise all the queues created during the burst. This is the
1166 * exact opposite of the best strategy for the other type of bursts.
1167 *
1168 * In the end, to take the best action for each of the two cases, the
1169 * two types of bursts need to be distinguished. Fortunately, this
1170 * seems relatively easy, by looking at the sizes of the bursts. In
1171 * particular, we found a threshold such that only bursts with a
1172 * larger size than that threshold are apparently caused by
1173 * services or commands such as systemd or git grep. For brevity,
1174 * hereafter we call just 'large' these bursts. BFQ *does not*
1175 * weight-raise queues whose creation occurs in a large burst. In
1176 * addition, for each of these queues BFQ performs or does not perform
1177 * idling depending on which choice boosts the throughput more. The
1178 * exact choice depends on the device and request pattern at
1179 * hand.
1180 *
1181 * Unfortunately, false positives may occur while an interactive task
1182 * is starting (e.g., an application is being started). The
1183 * consequence is that the queues associated with the task do not
1184 * enjoy weight raising as expected. Fortunately these false positives
1185 * are very rare. They typically occur if some service happens to
1186 * start doing I/O exactly when the interactive task starts.
1187 *
1188 * Turning back to the next function, it is invoked only if there are
1189 * no active queues (apart from active queues that would belong to the
1190 * same, possible burst bfqq would belong to), and it implements all
1191 * the steps needed to detect the occurrence of a large burst and to
1192 * properly mark all the queues belonging to it (so that they can then
1193 * be treated in a different way). This goal is achieved by
1194 * maintaining a "burst list" that holds, temporarily, the queues that
1195 * belong to the burst in progress. The list is then used to mark
1196 * these queues as belonging to a large burst if the burst does become
1197 * large. The main steps are the following.
1198 *
1199 * . when the very first queue is created, the queue is inserted into the
1200 * list (as it could be the first queue in a possible burst)
1201 *
1202 * . if the current burst has not yet become large, and a queue Q that does
1203 * not yet belong to the burst is activated shortly after the last time
1204 * at which a new queue entered the burst list, then the function appends
1205 * Q to the burst list
1206 *
1207 * . if, as a consequence of the previous step, the burst size reaches
1208 * the large-burst threshold, then
1209 *
1210 * . all the queues in the burst list are marked as belonging to a
1211 * large burst
1212 *
1213 * . the burst list is deleted; in fact, the burst list already served
1214 * its purpose (keeping temporarily track of the queues in a burst,
1215 * so as to be able to mark them as belonging to a large burst in the
1216 * previous sub-step), and now is not needed any more
1217 *
1218 * . the device enters a large-burst mode
1219 *
1220 * . if a queue Q that does not belong to the burst is created while
1221 * the device is in large-burst mode and shortly after the last time
1222 * at which a queue either entered the burst list or was marked as
1223 * belonging to the current large burst, then Q is immediately marked
1224 * as belonging to a large burst.
1225 *
1226 * . if a queue Q that does not belong to the burst is created a while
1227 * later, i.e., not shortly after, than the last time at which a queue
1228 * either entered the burst list or was marked as belonging to the
1229 * current large burst, then the current burst is deemed as finished and:
1230 *
1231 * . the large-burst mode is reset if set
1232 *
1233 * . the burst list is emptied
1234 *
1235 * . Q is inserted in the burst list, as Q may be the first queue
1236 * in a possible new burst (then the burst list contains just Q
1237 * after this step).
1238 */
1239static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1240{
1241 /*
1242 * If bfqq is already in the burst list or is part of a large
1243 * burst, or finally has just been split, then there is
1244 * nothing else to do.
1245 */
1246 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1247 bfq_bfqq_in_large_burst(bfqq) ||
1248 time_is_after_eq_jiffies(bfqq->split_time +
1249 msecs_to_jiffies(10)))
1250 return;
1251
1252 /*
1253 * If bfqq's creation happens late enough, or bfqq belongs to
1254 * a different group than the burst group, then the current
1255 * burst is finished, and related data structures must be
1256 * reset.
1257 *
1258 * In this respect, consider the special case where bfqq is
1259 * the very first queue created after BFQ is selected for this
1260 * device. In this case, last_ins_in_burst and
1261 * burst_parent_entity are not yet significant when we get
1262 * here. But it is easy to verify that, whether or not the
1263 * following condition is true, bfqq will end up being
1264 * inserted into the burst list. In particular the list will
1265 * happen to contain only bfqq. And this is exactly what has
1266 * to happen, as bfqq may be the first queue of the first
1267 * burst.
1268 */
1269 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1270 bfqd->bfq_burst_interval) ||
1271 bfqq->entity.parent != bfqd->burst_parent_entity) {
1272 bfqd->large_burst = false;
1273 bfq_reset_burst_list(bfqd, bfqq);
1274 goto end;
1275 }
1276
1277 /*
1278 * If we get here, then bfqq is being activated shortly after the
1279 * last queue. So, if the current burst is also large, we can mark
1280 * bfqq as belonging to this large burst immediately.
1281 */
1282 if (bfqd->large_burst) {
1283 bfq_mark_bfqq_in_large_burst(bfqq);
1284 goto end;
1285 }
1286
1287 /*
1288 * If we get here, then a large-burst state has not yet been
1289 * reached, but bfqq is being activated shortly after the last
1290 * queue. Then we add bfqq to the burst.
1291 */
1292 bfq_add_to_burst(bfqd, bfqq);
1293end:
1294 /*
1295 * At this point, bfqq either has been added to the current
1296 * burst or has caused the current burst to terminate and a
1297 * possible new burst to start. In particular, in the second
1298 * case, bfqq has become the first queue in the possible new
1299 * burst. In both cases last_ins_in_burst needs to be moved
1300 * forward.
1301 */
1302 bfqd->last_ins_in_burst = jiffies;
1303}
1304
1305static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1306{
1307 struct bfq_entity *entity = &bfqq->entity;
1308
1309 return entity->budget - entity->service;
1310}
1311
1312/*
1313 * If enough samples have been computed, return the current max budget
1314 * stored in bfqd, which is dynamically updated according to the
1315 * estimated disk peak rate; otherwise return the default max budget
1316 */
1317static int bfq_max_budget(struct bfq_data *bfqd)
1318{
1319 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1320 return bfq_default_max_budget;
1321 else
1322 return bfqd->bfq_max_budget;
1323}
1324
1325/*
1326 * Return min budget, which is a fraction of the current or default
1327 * max budget (trying with 1/32)
1328 */
1329static int bfq_min_budget(struct bfq_data *bfqd)
1330{
1331 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1332 return bfq_default_max_budget / 32;
1333 else
1334 return bfqd->bfq_max_budget / 32;
1335}
1336
1337/*
1338 * The next function, invoked after the input queue bfqq switches from
1339 * idle to busy, updates the budget of bfqq. The function also tells
1340 * whether the in-service queue should be expired, by returning
1341 * true. The purpose of expiring the in-service queue is to give bfqq
1342 * the chance to possibly preempt the in-service queue, and the reason
1343 * for preempting the in-service queue is to achieve one of the two
1344 * goals below.
1345 *
1346 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1347 * expired because it has remained idle. In particular, bfqq may have
1348 * expired for one of the following two reasons:
1349 *
1350 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1351 * and did not make it to issue a new request before its last
1352 * request was served;
1353 *
1354 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1355 * a new request before the expiration of the idling-time.
1356 *
1357 * Even if bfqq has expired for one of the above reasons, the process
1358 * associated with the queue may be however issuing requests greedily,
1359 * and thus be sensitive to the bandwidth it receives (bfqq may have
1360 * remained idle for other reasons: CPU high load, bfqq not enjoying
1361 * idling, I/O throttling somewhere in the path from the process to
1362 * the I/O scheduler, ...). But if, after every expiration for one of
1363 * the above two reasons, bfqq has to wait for the service of at least
1364 * one full budget of another queue before being served again, then
1365 * bfqq is likely to get a much lower bandwidth or resource time than
1366 * its reserved ones. To address this issue, two countermeasures need
1367 * to be taken.
1368 *
1369 * First, the budget and the timestamps of bfqq need to be updated in
1370 * a special way on bfqq reactivation: they need to be updated as if
1371 * bfqq did not remain idle and did not expire. In fact, if they are
1372 * computed as if bfqq expired and remained idle until reactivation,
1373 * then the process associated with bfqq is treated as if, instead of
1374 * being greedy, it stopped issuing requests when bfqq remained idle,
1375 * and restarts issuing requests only on this reactivation. In other
1376 * words, the scheduler does not help the process recover the "service
1377 * hole" between bfqq expiration and reactivation. As a consequence,
1378 * the process receives a lower bandwidth than its reserved one. In
1379 * contrast, to recover this hole, the budget must be updated as if
1380 * bfqq was not expired at all before this reactivation, i.e., it must
1381 * be set to the value of the remaining budget when bfqq was
1382 * expired. Along the same line, timestamps need to be assigned the
1383 * value they had the last time bfqq was selected for service, i.e.,
1384 * before last expiration. Thus timestamps need to be back-shifted
1385 * with respect to their normal computation (see [1] for more details
1386 * on this tricky aspect).
1387 *
1388 * Secondly, to allow the process to recover the hole, the in-service
1389 * queue must be expired too, to give bfqq the chance to preempt it
1390 * immediately. In fact, if bfqq has to wait for a full budget of the
1391 * in-service queue to be completed, then it may become impossible to
1392 * let the process recover the hole, even if the back-shifted
1393 * timestamps of bfqq are lower than those of the in-service queue. If
1394 * this happens for most or all of the holes, then the process may not
1395 * receive its reserved bandwidth. In this respect, it is worth noting
1396 * that, being the service of outstanding requests unpreemptible, a
1397 * little fraction of the holes may however be unrecoverable, thereby
1398 * causing a little loss of bandwidth.
1399 *
1400 * The last important point is detecting whether bfqq does need this
1401 * bandwidth recovery. In this respect, the next function deems the
1402 * process associated with bfqq greedy, and thus allows it to recover
1403 * the hole, if: 1) the process is waiting for the arrival of a new
1404 * request (which implies that bfqq expired for one of the above two
1405 * reasons), and 2) such a request has arrived soon. The first
1406 * condition is controlled through the flag non_blocking_wait_rq,
1407 * while the second through the flag arrived_in_time. If both
1408 * conditions hold, then the function computes the budget in the
1409 * above-described special way, and signals that the in-service queue
1410 * should be expired. Timestamp back-shifting is done later in
1411 * __bfq_activate_entity.
1412 *
1413 * 2. Reduce latency. Even if timestamps are not backshifted to let
1414 * the process associated with bfqq recover a service hole, bfqq may
1415 * however happen to have, after being (re)activated, a lower finish
1416 * timestamp than the in-service queue. That is, the next budget of
1417 * bfqq may have to be completed before the one of the in-service
1418 * queue. If this is the case, then preempting the in-service queue
1419 * allows this goal to be achieved, apart from the unpreemptible,
1420 * outstanding requests mentioned above.
1421 *
1422 * Unfortunately, regardless of which of the above two goals one wants
1423 * to achieve, service trees need first to be updated to know whether
1424 * the in-service queue must be preempted. To have service trees
1425 * correctly updated, the in-service queue must be expired and
1426 * rescheduled, and bfqq must be scheduled too. This is one of the
1427 * most costly operations (in future versions, the scheduling
1428 * mechanism may be re-designed in such a way to make it possible to
1429 * know whether preemption is needed without needing to update service
1430 * trees). In addition, queue preemptions almost always cause random
1431 * I/O, which may in turn cause loss of throughput. Finally, there may
1432 * even be no in-service queue when the next function is invoked (so,
1433 * no queue to compare timestamps with). Because of these facts, the
1434 * next function adopts the following simple scheme to avoid costly
1435 * operations, too frequent preemptions and too many dependencies on
1436 * the state of the scheduler: it requests the expiration of the
1437 * in-service queue (unconditionally) only for queues that need to
1438 * recover a hole. Then it delegates to other parts of the code the
1439 * responsibility of handling the above case 2.
1440 */
1441static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1442 struct bfq_queue *bfqq,
1443 bool arrived_in_time)
1444{
1445 struct bfq_entity *entity = &bfqq->entity;
1446
1447 /*
1448 * In the next compound condition, we check also whether there
1449 * is some budget left, because otherwise there is no point in
1450 * trying to go on serving bfqq with this same budget: bfqq
1451 * would be expired immediately after being selected for
1452 * service. This would only cause useless overhead.
1453 */
1454 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1455 bfq_bfqq_budget_left(bfqq) > 0) {
1456 /*
1457 * We do not clear the flag non_blocking_wait_rq here, as
1458 * the latter is used in bfq_activate_bfqq to signal
1459 * that timestamps need to be back-shifted (and is
1460 * cleared right after).
1461 */
1462
1463 /*
1464 * In next assignment we rely on that either
1465 * entity->service or entity->budget are not updated
1466 * on expiration if bfqq is empty (see
1467 * __bfq_bfqq_recalc_budget). Thus both quantities
1468 * remain unchanged after such an expiration, and the
1469 * following statement therefore assigns to
1470 * entity->budget the remaining budget on such an
1471 * expiration.
1472 */
1473 entity->budget = min_t(unsigned long,
1474 bfq_bfqq_budget_left(bfqq),
1475 bfqq->max_budget);
1476
1477 /*
1478 * At this point, we have used entity->service to get
1479 * the budget left (needed for updating
1480 * entity->budget). Thus we finally can, and have to,
1481 * reset entity->service. The latter must be reset
1482 * because bfqq would otherwise be charged again for
1483 * the service it has received during its previous
1484 * service slot(s).
1485 */
1486 entity->service = 0;
1487
1488 return true;
1489 }
1490
1491 /*
1492 * We can finally complete expiration, by setting service to 0.
1493 */
1494 entity->service = 0;
1495 entity->budget = max_t(unsigned long, bfqq->max_budget,
1496 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1497 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1498 return false;
1499}
1500
1501/*
1502 * Return the farthest past time instant according to jiffies
1503 * macros.
1504 */
1505static unsigned long bfq_smallest_from_now(void)
1506{
1507 return jiffies - MAX_JIFFY_OFFSET;
1508}
1509
1510static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1511 struct bfq_queue *bfqq,
1512 unsigned int old_wr_coeff,
1513 bool wr_or_deserves_wr,
1514 bool interactive,
1515 bool in_burst,
1516 bool soft_rt)
1517{
1518 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1519 /* start a weight-raising period */
1520 if (interactive) {
1521 bfqq->service_from_wr = 0;
1522 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1523 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1524 } else {
1525 /*
1526 * No interactive weight raising in progress
1527 * here: assign minus infinity to
1528 * wr_start_at_switch_to_srt, to make sure
1529 * that, at the end of the soft-real-time
1530 * weight raising periods that is starting
1531 * now, no interactive weight-raising period
1532 * may be wrongly considered as still in
1533 * progress (and thus actually started by
1534 * mistake).
1535 */
1536 bfqq->wr_start_at_switch_to_srt =
1537 bfq_smallest_from_now();
1538 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1539 BFQ_SOFTRT_WEIGHT_FACTOR;
1540 bfqq->wr_cur_max_time =
1541 bfqd->bfq_wr_rt_max_time;
1542 }
1543
1544 /*
1545 * If needed, further reduce budget to make sure it is
1546 * close to bfqq's backlog, so as to reduce the
1547 * scheduling-error component due to a too large
1548 * budget. Do not care about throughput consequences,
1549 * but only about latency. Finally, do not assign a
1550 * too small budget either, to avoid increasing
1551 * latency by causing too frequent expirations.
1552 */
1553 bfqq->entity.budget = min_t(unsigned long,
1554 bfqq->entity.budget,
1555 2 * bfq_min_budget(bfqd));
1556 } else if (old_wr_coeff > 1) {
1557 if (interactive) { /* update wr coeff and duration */
1558 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1559 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1560 } else if (in_burst)
1561 bfqq->wr_coeff = 1;
1562 else if (soft_rt) {
1563 /*
1564 * The application is now or still meeting the
1565 * requirements for being deemed soft rt. We
1566 * can then correctly and safely (re)charge
1567 * the weight-raising duration for the
1568 * application with the weight-raising
1569 * duration for soft rt applications.
1570 *
1571 * In particular, doing this recharge now, i.e.,
1572 * before the weight-raising period for the
1573 * application finishes, reduces the probability
1574 * of the following negative scenario:
1575 * 1) the weight of a soft rt application is
1576 * raised at startup (as for any newly
1577 * created application),
1578 * 2) since the application is not interactive,
1579 * at a certain time weight-raising is
1580 * stopped for the application,
1581 * 3) at that time the application happens to
1582 * still have pending requests, and hence
1583 * is destined to not have a chance to be
1584 * deemed soft rt before these requests are
1585 * completed (see the comments to the
1586 * function bfq_bfqq_softrt_next_start()
1587 * for details on soft rt detection),
1588 * 4) these pending requests experience a high
1589 * latency because the application is not
1590 * weight-raised while they are pending.
1591 */
1592 if (bfqq->wr_cur_max_time !=
1593 bfqd->bfq_wr_rt_max_time) {
1594 bfqq->wr_start_at_switch_to_srt =
1595 bfqq->last_wr_start_finish;
1596
1597 bfqq->wr_cur_max_time =
1598 bfqd->bfq_wr_rt_max_time;
1599 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1600 BFQ_SOFTRT_WEIGHT_FACTOR;
1601 }
1602 bfqq->last_wr_start_finish = jiffies;
1603 }
1604 }
1605}
1606
1607static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1608 struct bfq_queue *bfqq)
1609{
1610 return bfqq->dispatched == 0 &&
1611 time_is_before_jiffies(
1612 bfqq->budget_timeout +
1613 bfqd->bfq_wr_min_idle_time);
1614}
1615
1616
1617/*
1618 * Return true if bfqq is in a higher priority class, or has a higher
1619 * weight than the in-service queue.
1620 */
1621static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq,
1622 struct bfq_queue *in_serv_bfqq)
1623{
1624 int bfqq_weight, in_serv_weight;
1625
1626 if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class)
1627 return true;
1628
1629 if (in_serv_bfqq->entity.parent == bfqq->entity.parent) {
1630 bfqq_weight = bfqq->entity.weight;
1631 in_serv_weight = in_serv_bfqq->entity.weight;
1632 } else {
1633 if (bfqq->entity.parent)
1634 bfqq_weight = bfqq->entity.parent->weight;
1635 else
1636 bfqq_weight = bfqq->entity.weight;
1637 if (in_serv_bfqq->entity.parent)
1638 in_serv_weight = in_serv_bfqq->entity.parent->weight;
1639 else
1640 in_serv_weight = in_serv_bfqq->entity.weight;
1641 }
1642
1643 return bfqq_weight > in_serv_weight;
1644}
1645
1646static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1647 struct bfq_queue *bfqq,
1648 int old_wr_coeff,
1649 struct request *rq,
1650 bool *interactive)
1651{
1652 bool soft_rt, in_burst, wr_or_deserves_wr,
1653 bfqq_wants_to_preempt,
1654 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1655 /*
1656 * See the comments on
1657 * bfq_bfqq_update_budg_for_activation for
1658 * details on the usage of the next variable.
1659 */
1660 arrived_in_time = ktime_get_ns() <=
1661 bfqq->ttime.last_end_request +
1662 bfqd->bfq_slice_idle * 3;
1663
1664
1665 /*
1666 * bfqq deserves to be weight-raised if:
1667 * - it is sync,
1668 * - it does not belong to a large burst,
1669 * - it has been idle for enough time or is soft real-time,
1670 * - is linked to a bfq_io_cq (it is not shared in any sense).
1671 */
1672 in_burst = bfq_bfqq_in_large_burst(bfqq);
1673 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1674 !BFQQ_TOTALLY_SEEKY(bfqq) &&
1675 !in_burst &&
1676 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1677 bfqq->dispatched == 0;
1678 *interactive = !in_burst && idle_for_long_time;
1679 wr_or_deserves_wr = bfqd->low_latency &&
1680 (bfqq->wr_coeff > 1 ||
1681 (bfq_bfqq_sync(bfqq) &&
1682 bfqq->bic && (*interactive || soft_rt)));
1683
1684 /*
1685 * Using the last flag, update budget and check whether bfqq
1686 * may want to preempt the in-service queue.
1687 */
1688 bfqq_wants_to_preempt =
1689 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1690 arrived_in_time);
1691
1692 /*
1693 * If bfqq happened to be activated in a burst, but has been
1694 * idle for much more than an interactive queue, then we
1695 * assume that, in the overall I/O initiated in the burst, the
1696 * I/O associated with bfqq is finished. So bfqq does not need
1697 * to be treated as a queue belonging to a burst
1698 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1699 * if set, and remove bfqq from the burst list if it's
1700 * there. We do not decrement burst_size, because the fact
1701 * that bfqq does not need to belong to the burst list any
1702 * more does not invalidate the fact that bfqq was created in
1703 * a burst.
1704 */
1705 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1706 idle_for_long_time &&
1707 time_is_before_jiffies(
1708 bfqq->budget_timeout +
1709 msecs_to_jiffies(10000))) {
1710 hlist_del_init(&bfqq->burst_list_node);
1711 bfq_clear_bfqq_in_large_burst(bfqq);
1712 }
1713
1714 bfq_clear_bfqq_just_created(bfqq);
1715
1716
1717 if (!bfq_bfqq_IO_bound(bfqq)) {
1718 if (arrived_in_time) {
1719 bfqq->requests_within_timer++;
1720 if (bfqq->requests_within_timer >=
1721 bfqd->bfq_requests_within_timer)
1722 bfq_mark_bfqq_IO_bound(bfqq);
1723 } else
1724 bfqq->requests_within_timer = 0;
1725 }
1726
1727 if (bfqd->low_latency) {
1728 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1729 /* wraparound */
1730 bfqq->split_time =
1731 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1732
1733 if (time_is_before_jiffies(bfqq->split_time +
1734 bfqd->bfq_wr_min_idle_time)) {
1735 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1736 old_wr_coeff,
1737 wr_or_deserves_wr,
1738 *interactive,
1739 in_burst,
1740 soft_rt);
1741
1742 if (old_wr_coeff != bfqq->wr_coeff)
1743 bfqq->entity.prio_changed = 1;
1744 }
1745 }
1746
1747 bfqq->last_idle_bklogged = jiffies;
1748 bfqq->service_from_backlogged = 0;
1749 bfq_clear_bfqq_softrt_update(bfqq);
1750
1751 bfq_add_bfqq_busy(bfqd, bfqq);
1752
1753 /*
1754 * Expire in-service queue only if preemption may be needed
1755 * for guarantees. In particular, we care only about two
1756 * cases. The first is that bfqq has to recover a service
1757 * hole, as explained in the comments on
1758 * bfq_bfqq_update_budg_for_activation(), i.e., that
1759 * bfqq_wants_to_preempt is true. However, if bfqq does not
1760 * carry time-critical I/O, then bfqq's bandwidth is less
1761 * important than that of queues that carry time-critical I/O.
1762 * So, as a further constraint, we consider this case only if
1763 * bfqq is at least as weight-raised, i.e., at least as time
1764 * critical, as the in-service queue.
1765 *
1766 * The second case is that bfqq is in a higher priority class,
1767 * or has a higher weight than the in-service queue. If this
1768 * condition does not hold, we don't care because, even if
1769 * bfqq does not start to be served immediately, the resulting
1770 * delay for bfqq's I/O is however lower or much lower than
1771 * the ideal completion time to be guaranteed to bfqq's I/O.
1772 *
1773 * In both cases, preemption is needed only if, according to
1774 * the timestamps of both bfqq and of the in-service queue,
1775 * bfqq actually is the next queue to serve. So, to reduce
1776 * useless preemptions, the return value of
1777 * next_queue_may_preempt() is considered in the next compound
1778 * condition too. Yet next_queue_may_preempt() just checks a
1779 * simple, necessary condition for bfqq to be the next queue
1780 * to serve. In fact, to evaluate a sufficient condition, the
1781 * timestamps of the in-service queue would need to be
1782 * updated, and this operation is quite costly (see the
1783 * comments on bfq_bfqq_update_budg_for_activation()).
1784 */
1785 if (bfqd->in_service_queue &&
1786 ((bfqq_wants_to_preempt &&
1787 bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) ||
1788 bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue)) &&
1789 next_queue_may_preempt(bfqd))
1790 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1791 false, BFQQE_PREEMPTED);
1792}
1793
1794static void bfq_reset_inject_limit(struct bfq_data *bfqd,
1795 struct bfq_queue *bfqq)
1796{
1797 /* invalidate baseline total service time */
1798 bfqq->last_serv_time_ns = 0;
1799
1800 /*
1801 * Reset pointer in case we are waiting for
1802 * some request completion.
1803 */
1804 bfqd->waited_rq = NULL;
1805
1806 /*
1807 * If bfqq has a short think time, then start by setting the
1808 * inject limit to 0 prudentially, because the service time of
1809 * an injected I/O request may be higher than the think time
1810 * of bfqq, and therefore, if one request was injected when
1811 * bfqq remains empty, this injected request might delay the
1812 * service of the next I/O request for bfqq significantly. In
1813 * case bfqq can actually tolerate some injection, then the
1814 * adaptive update will however raise the limit soon. This
1815 * lucky circumstance holds exactly because bfqq has a short
1816 * think time, and thus, after remaining empty, is likely to
1817 * get new I/O enqueued---and then completed---before being
1818 * expired. This is the very pattern that gives the
1819 * limit-update algorithm the chance to measure the effect of
1820 * injection on request service times, and then to update the
1821 * limit accordingly.
1822 *
1823 * However, in the following special case, the inject limit is
1824 * left to 1 even if the think time is short: bfqq's I/O is
1825 * synchronized with that of some other queue, i.e., bfqq may
1826 * receive new I/O only after the I/O of the other queue is
1827 * completed. Keeping the inject limit to 1 allows the
1828 * blocking I/O to be served while bfqq is in service. And
1829 * this is very convenient both for bfqq and for overall
1830 * throughput, as explained in detail in the comments in
1831 * bfq_update_has_short_ttime().
1832 *
1833 * On the opposite end, if bfqq has a long think time, then
1834 * start directly by 1, because:
1835 * a) on the bright side, keeping at most one request in
1836 * service in the drive is unlikely to cause any harm to the
1837 * latency of bfqq's requests, as the service time of a single
1838 * request is likely to be lower than the think time of bfqq;
1839 * b) on the downside, after becoming empty, bfqq is likely to
1840 * expire before getting its next request. With this request
1841 * arrival pattern, it is very hard to sample total service
1842 * times and update the inject limit accordingly (see comments
1843 * on bfq_update_inject_limit()). So the limit is likely to be
1844 * never, or at least seldom, updated. As a consequence, by
1845 * setting the limit to 1, we avoid that no injection ever
1846 * occurs with bfqq. On the downside, this proactive step
1847 * further reduces chances to actually compute the baseline
1848 * total service time. Thus it reduces chances to execute the
1849 * limit-update algorithm and possibly raise the limit to more
1850 * than 1.
1851 */
1852 if (bfq_bfqq_has_short_ttime(bfqq))
1853 bfqq->inject_limit = 0;
1854 else
1855 bfqq->inject_limit = 1;
1856
1857 bfqq->decrease_time_jif = jiffies;
1858}
1859
1860static void bfq_add_request(struct request *rq)
1861{
1862 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1863 struct bfq_data *bfqd = bfqq->bfqd;
1864 struct request *next_rq, *prev;
1865 unsigned int old_wr_coeff = bfqq->wr_coeff;
1866 bool interactive = false;
1867
1868 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1869 bfqq->queued[rq_is_sync(rq)]++;
1870 bfqd->queued++;
1871
1872 if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_bfqq_sync(bfqq)) {
1873 /*
1874 * Detect whether bfqq's I/O seems synchronized with
1875 * that of some other queue, i.e., whether bfqq, after
1876 * remaining empty, happens to receive new I/O only
1877 * right after some I/O request of the other queue has
1878 * been completed. We call waker queue the other
1879 * queue, and we assume, for simplicity, that bfqq may
1880 * have at most one waker queue.
1881 *
1882 * A remarkable throughput boost can be reached by
1883 * unconditionally injecting the I/O of the waker
1884 * queue, every time a new bfq_dispatch_request
1885 * happens to be invoked while I/O is being plugged
1886 * for bfqq. In addition to boosting throughput, this
1887 * unblocks bfqq's I/O, thereby improving bandwidth
1888 * and latency for bfqq. Note that these same results
1889 * may be achieved with the general injection
1890 * mechanism, but less effectively. For details on
1891 * this aspect, see the comments on the choice of the
1892 * queue for injection in bfq_select_queue().
1893 *
1894 * Turning back to the detection of a waker queue, a
1895 * queue Q is deemed as a waker queue for bfqq if, for
1896 * two consecutive times, bfqq happens to become non
1897 * empty right after a request of Q has been
1898 * completed. In particular, on the first time, Q is
1899 * tentatively set as a candidate waker queue, while
1900 * on the second time, the flag
1901 * bfq_bfqq_has_waker(bfqq) is set to confirm that Q
1902 * is a waker queue for bfqq. These detection steps
1903 * are performed only if bfqq has a long think time,
1904 * so as to make it more likely that bfqq's I/O is
1905 * actually being blocked by a synchronization. This
1906 * last filter, plus the above two-times requirement,
1907 * make false positives less likely.
1908 *
1909 * NOTE
1910 *
1911 * The sooner a waker queue is detected, the sooner
1912 * throughput can be boosted by injecting I/O from the
1913 * waker queue. Fortunately, detection is likely to be
1914 * actually fast, for the following reasons. While
1915 * blocked by synchronization, bfqq has a long think
1916 * time. This implies that bfqq's inject limit is at
1917 * least equal to 1 (see the comments in
1918 * bfq_update_inject_limit()). So, thanks to
1919 * injection, the waker queue is likely to be served
1920 * during the very first I/O-plugging time interval
1921 * for bfqq. This triggers the first step of the
1922 * detection mechanism. Thanks again to injection, the
1923 * candidate waker queue is then likely to be
1924 * confirmed no later than during the next
1925 * I/O-plugging interval for bfqq.
1926 */
1927 if (bfqd->last_completed_rq_bfqq &&
1928 !bfq_bfqq_has_short_ttime(bfqq) &&
1929 ktime_get_ns() - bfqd->last_completion <
1930 200 * NSEC_PER_USEC) {
1931 if (bfqd->last_completed_rq_bfqq != bfqq &&
1932 bfqd->last_completed_rq_bfqq !=
1933 bfqq->waker_bfqq) {
1934 /*
1935 * First synchronization detected with
1936 * a candidate waker queue, or with a
1937 * different candidate waker queue
1938 * from the current one.
1939 */
1940 bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
1941
1942 /*
1943 * If the waker queue disappears, then
1944 * bfqq->waker_bfqq must be reset. To
1945 * this goal, we maintain in each
1946 * waker queue a list, woken_list, of
1947 * all the queues that reference the
1948 * waker queue through their
1949 * waker_bfqq pointer. When the waker
1950 * queue exits, the waker_bfqq pointer
1951 * of all the queues in the woken_list
1952 * is reset.
1953 *
1954 * In addition, if bfqq is already in
1955 * the woken_list of a waker queue,
1956 * then, before being inserted into
1957 * the woken_list of a new waker
1958 * queue, bfqq must be removed from
1959 * the woken_list of the old waker
1960 * queue.
1961 */
1962 if (!hlist_unhashed(&bfqq->woken_list_node))
1963 hlist_del_init(&bfqq->woken_list_node);
1964 hlist_add_head(&bfqq->woken_list_node,
1965 &bfqd->last_completed_rq_bfqq->woken_list);
1966
1967 bfq_clear_bfqq_has_waker(bfqq);
1968 } else if (bfqd->last_completed_rq_bfqq ==
1969 bfqq->waker_bfqq &&
1970 !bfq_bfqq_has_waker(bfqq)) {
1971 /*
1972 * synchronization with waker_bfqq
1973 * seen for the second time
1974 */
1975 bfq_mark_bfqq_has_waker(bfqq);
1976 }
1977 }
1978
1979 /*
1980 * Periodically reset inject limit, to make sure that
1981 * the latter eventually drops in case workload
1982 * changes, see step (3) in the comments on
1983 * bfq_update_inject_limit().
1984 */
1985 if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
1986 msecs_to_jiffies(1000)))
1987 bfq_reset_inject_limit(bfqd, bfqq);
1988
1989 /*
1990 * The following conditions must hold to setup a new
1991 * sampling of total service time, and then a new
1992 * update of the inject limit:
1993 * - bfqq is in service, because the total service
1994 * time is evaluated only for the I/O requests of
1995 * the queues in service;
1996 * - this is the right occasion to compute or to
1997 * lower the baseline total service time, because
1998 * there are actually no requests in the drive,
1999 * or
2000 * the baseline total service time is available, and
2001 * this is the right occasion to compute the other
2002 * quantity needed to update the inject limit, i.e.,
2003 * the total service time caused by the amount of
2004 * injection allowed by the current value of the
2005 * limit. It is the right occasion because injection
2006 * has actually been performed during the service
2007 * hole, and there are still in-flight requests,
2008 * which are very likely to be exactly the injected
2009 * requests, or part of them;
2010 * - the minimum interval for sampling the total
2011 * service time and updating the inject limit has
2012 * elapsed.
2013 */
2014 if (bfqq == bfqd->in_service_queue &&
2015 (bfqd->rq_in_driver == 0 ||
2016 (bfqq->last_serv_time_ns > 0 &&
2017 bfqd->rqs_injected && bfqd->rq_in_driver > 0)) &&
2018 time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2019 msecs_to_jiffies(10))) {
2020 bfqd->last_empty_occupied_ns = ktime_get_ns();
2021 /*
2022 * Start the state machine for measuring the
2023 * total service time of rq: setting
2024 * wait_dispatch will cause bfqd->waited_rq to
2025 * be set when rq will be dispatched.
2026 */
2027 bfqd->wait_dispatch = true;
2028 /*
2029 * If there is no I/O in service in the drive,
2030 * then possible injection occurred before the
2031 * arrival of rq will not affect the total
2032 * service time of rq. So the injection limit
2033 * must not be updated as a function of such
2034 * total service time, unless new injection
2035 * occurs before rq is completed. To have the
2036 * injection limit updated only in the latter
2037 * case, reset rqs_injected here (rqs_injected
2038 * will be set in case injection is performed
2039 * on bfqq before rq is completed).
2040 */
2041 if (bfqd->rq_in_driver == 0)
2042 bfqd->rqs_injected = false;
2043 }
2044 }
2045
2046 elv_rb_add(&bfqq->sort_list, rq);
2047
2048 /*
2049 * Check if this request is a better next-serve candidate.
2050 */
2051 prev = bfqq->next_rq;
2052 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
2053 bfqq->next_rq = next_rq;
2054
2055 /*
2056 * Adjust priority tree position, if next_rq changes.
2057 * See comments on bfq_pos_tree_add_move() for the unlikely().
2058 */
2059 if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
2060 bfq_pos_tree_add_move(bfqd, bfqq);
2061
2062 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
2063 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
2064 rq, &interactive);
2065 else {
2066 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
2067 time_is_before_jiffies(
2068 bfqq->last_wr_start_finish +
2069 bfqd->bfq_wr_min_inter_arr_async)) {
2070 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
2071 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
2072
2073 bfqd->wr_busy_queues++;
2074 bfqq->entity.prio_changed = 1;
2075 }
2076 if (prev != bfqq->next_rq)
2077 bfq_updated_next_req(bfqd, bfqq);
2078 }
2079
2080 /*
2081 * Assign jiffies to last_wr_start_finish in the following
2082 * cases:
2083 *
2084 * . if bfqq is not going to be weight-raised, because, for
2085 * non weight-raised queues, last_wr_start_finish stores the
2086 * arrival time of the last request; as of now, this piece
2087 * of information is used only for deciding whether to
2088 * weight-raise async queues
2089 *
2090 * . if bfqq is not weight-raised, because, if bfqq is now
2091 * switching to weight-raised, then last_wr_start_finish
2092 * stores the time when weight-raising starts
2093 *
2094 * . if bfqq is interactive, because, regardless of whether
2095 * bfqq is currently weight-raised, the weight-raising
2096 * period must start or restart (this case is considered
2097 * separately because it is not detected by the above
2098 * conditions, if bfqq is already weight-raised)
2099 *
2100 * last_wr_start_finish has to be updated also if bfqq is soft
2101 * real-time, because the weight-raising period is constantly
2102 * restarted on idle-to-busy transitions for these queues, but
2103 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2104 * needed.
2105 */
2106 if (bfqd->low_latency &&
2107 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
2108 bfqq->last_wr_start_finish = jiffies;
2109}
2110
2111static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
2112 struct bio *bio,
2113 struct request_queue *q)
2114{
2115 struct bfq_queue *bfqq = bfqd->bio_bfqq;
2116
2117
2118 if (bfqq)
2119 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
2120
2121 return NULL;
2122}
2123
2124static sector_t get_sdist(sector_t last_pos, struct request *rq)
2125{
2126 if (last_pos)
2127 return abs(blk_rq_pos(rq) - last_pos);
2128
2129 return 0;
2130}
2131
2132#if 0 /* Still not clear if we can do without next two functions */
2133static void bfq_activate_request(struct request_queue *q, struct request *rq)
2134{
2135 struct bfq_data *bfqd = q->elevator->elevator_data;
2136
2137 bfqd->rq_in_driver++;
2138}
2139
2140static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
2141{
2142 struct bfq_data *bfqd = q->elevator->elevator_data;
2143
2144 bfqd->rq_in_driver--;
2145}
2146#endif
2147
2148static void bfq_remove_request(struct request_queue *q,
2149 struct request *rq)
2150{
2151 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2152 struct bfq_data *bfqd = bfqq->bfqd;
2153 const int sync = rq_is_sync(rq);
2154
2155 if (bfqq->next_rq == rq) {
2156 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
2157 bfq_updated_next_req(bfqd, bfqq);
2158 }
2159
2160 if (rq->queuelist.prev != &rq->queuelist)
2161 list_del_init(&rq->queuelist);
2162 bfqq->queued[sync]--;
2163 bfqd->queued--;
2164 elv_rb_del(&bfqq->sort_list, rq);
2165
2166 elv_rqhash_del(q, rq);
2167 if (q->last_merge == rq)
2168 q->last_merge = NULL;
2169
2170 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2171 bfqq->next_rq = NULL;
2172
2173 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
2174 bfq_del_bfqq_busy(bfqd, bfqq, false);
2175 /*
2176 * bfqq emptied. In normal operation, when
2177 * bfqq is empty, bfqq->entity.service and
2178 * bfqq->entity.budget must contain,
2179 * respectively, the service received and the
2180 * budget used last time bfqq emptied. These
2181 * facts do not hold in this case, as at least
2182 * this last removal occurred while bfqq is
2183 * not in service. To avoid inconsistencies,
2184 * reset both bfqq->entity.service and
2185 * bfqq->entity.budget, if bfqq has still a
2186 * process that may issue I/O requests to it.
2187 */
2188 bfqq->entity.budget = bfqq->entity.service = 0;
2189 }
2190
2191 /*
2192 * Remove queue from request-position tree as it is empty.
2193 */
2194 if (bfqq->pos_root) {
2195 rb_erase(&bfqq->pos_node, bfqq->pos_root);
2196 bfqq->pos_root = NULL;
2197 }
2198 } else {
2199 /* see comments on bfq_pos_tree_add_move() for the unlikely() */
2200 if (unlikely(!bfqd->nonrot_with_queueing))
2201 bfq_pos_tree_add_move(bfqd, bfqq);
2202 }
2203
2204 if (rq->cmd_flags & REQ_META)
2205 bfqq->meta_pending--;
2206
2207}
2208
2209static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio,
2210 unsigned int nr_segs)
2211{
2212 struct request_queue *q = hctx->queue;
2213 struct bfq_data *bfqd = q->elevator->elevator_data;
2214 struct request *free = NULL;
2215 /*
2216 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2217 * store its return value for later use, to avoid nesting
2218 * queue_lock inside the bfqd->lock. We assume that the bic
2219 * returned by bfq_bic_lookup does not go away before
2220 * bfqd->lock is taken.
2221 */
2222 struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
2223 bool ret;
2224
2225 spin_lock_irq(&bfqd->lock);
2226
2227 if (bic)
2228 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
2229 else
2230 bfqd->bio_bfqq = NULL;
2231 bfqd->bio_bic = bic;
2232
2233 ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
2234
2235 if (free)
2236 blk_mq_free_request(free);
2237 spin_unlock_irq(&bfqd->lock);
2238
2239 return ret;
2240}
2241
2242static int bfq_request_merge(struct request_queue *q, struct request **req,
2243 struct bio *bio)
2244{
2245 struct bfq_data *bfqd = q->elevator->elevator_data;
2246 struct request *__rq;
2247
2248 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
2249 if (__rq && elv_bio_merge_ok(__rq, bio)) {
2250 *req = __rq;
2251 return ELEVATOR_FRONT_MERGE;
2252 }
2253
2254 return ELEVATOR_NO_MERGE;
2255}
2256
2257static struct bfq_queue *bfq_init_rq(struct request *rq);
2258
2259static void bfq_request_merged(struct request_queue *q, struct request *req,
2260 enum elv_merge type)
2261{
2262 if (type == ELEVATOR_FRONT_MERGE &&
2263 rb_prev(&req->rb_node) &&
2264 blk_rq_pos(req) <
2265 blk_rq_pos(container_of(rb_prev(&req->rb_node),
2266 struct request, rb_node))) {
2267 struct bfq_queue *bfqq = bfq_init_rq(req);
2268 struct bfq_data *bfqd;
2269 struct request *prev, *next_rq;
2270
2271 if (!bfqq)
2272 return;
2273
2274 bfqd = bfqq->bfqd;
2275
2276 /* Reposition request in its sort_list */
2277 elv_rb_del(&bfqq->sort_list, req);
2278 elv_rb_add(&bfqq->sort_list, req);
2279
2280 /* Choose next request to be served for bfqq */
2281 prev = bfqq->next_rq;
2282 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
2283 bfqd->last_position);
2284 bfqq->next_rq = next_rq;
2285 /*
2286 * If next_rq changes, update both the queue's budget to
2287 * fit the new request and the queue's position in its
2288 * rq_pos_tree.
2289 */
2290 if (prev != bfqq->next_rq) {
2291 bfq_updated_next_req(bfqd, bfqq);
2292 /*
2293 * See comments on bfq_pos_tree_add_move() for
2294 * the unlikely().
2295 */
2296 if (unlikely(!bfqd->nonrot_with_queueing))
2297 bfq_pos_tree_add_move(bfqd, bfqq);
2298 }
2299 }
2300}
2301
2302/*
2303 * This function is called to notify the scheduler that the requests
2304 * rq and 'next' have been merged, with 'next' going away. BFQ
2305 * exploits this hook to address the following issue: if 'next' has a
2306 * fifo_time lower that rq, then the fifo_time of rq must be set to
2307 * the value of 'next', to not forget the greater age of 'next'.
2308 *
2309 * NOTE: in this function we assume that rq is in a bfq_queue, basing
2310 * on that rq is picked from the hash table q->elevator->hash, which,
2311 * in its turn, is filled only with I/O requests present in
2312 * bfq_queues, while BFQ is in use for the request queue q. In fact,
2313 * the function that fills this hash table (elv_rqhash_add) is called
2314 * only by bfq_insert_request.
2315 */
2316static void bfq_requests_merged(struct request_queue *q, struct request *rq,
2317 struct request *next)
2318{
2319 struct bfq_queue *bfqq = bfq_init_rq(rq),
2320 *next_bfqq = bfq_init_rq(next);
2321
2322 if (!bfqq)
2323 return;
2324
2325 /*
2326 * If next and rq belong to the same bfq_queue and next is older
2327 * than rq, then reposition rq in the fifo (by substituting next
2328 * with rq). Otherwise, if next and rq belong to different
2329 * bfq_queues, never reposition rq: in fact, we would have to
2330 * reposition it with respect to next's position in its own fifo,
2331 * which would most certainly be too expensive with respect to
2332 * the benefits.
2333 */
2334 if (bfqq == next_bfqq &&
2335 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2336 next->fifo_time < rq->fifo_time) {
2337 list_del_init(&rq->queuelist);
2338 list_replace_init(&next->queuelist, &rq->queuelist);
2339 rq->fifo_time = next->fifo_time;
2340 }
2341
2342 if (bfqq->next_rq == next)
2343 bfqq->next_rq = rq;
2344
2345 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
2346}
2347
2348/* Must be called with bfqq != NULL */
2349static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
2350{
2351 if (bfq_bfqq_busy(bfqq))
2352 bfqq->bfqd->wr_busy_queues--;
2353 bfqq->wr_coeff = 1;
2354 bfqq->wr_cur_max_time = 0;
2355 bfqq->last_wr_start_finish = jiffies;
2356 /*
2357 * Trigger a weight change on the next invocation of
2358 * __bfq_entity_update_weight_prio.
2359 */
2360 bfqq->entity.prio_changed = 1;
2361}
2362
2363void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2364 struct bfq_group *bfqg)
2365{
2366 int i, j;
2367
2368 for (i = 0; i < 2; i++)
2369 for (j = 0; j < IOPRIO_BE_NR; j++)
2370 if (bfqg->async_bfqq[i][j])
2371 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2372 if (bfqg->async_idle_bfqq)
2373 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2374}
2375
2376static void bfq_end_wr(struct bfq_data *bfqd)
2377{
2378 struct bfq_queue *bfqq;
2379
2380 spin_lock_irq(&bfqd->lock);
2381
2382 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2383 bfq_bfqq_end_wr(bfqq);
2384 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2385 bfq_bfqq_end_wr(bfqq);
2386 bfq_end_wr_async(bfqd);
2387
2388 spin_unlock_irq(&bfqd->lock);
2389}
2390
2391static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2392{
2393 if (request)
2394 return blk_rq_pos(io_struct);
2395 else
2396 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2397}
2398
2399static int bfq_rq_close_to_sector(void *io_struct, bool request,
2400 sector_t sector)
2401{
2402 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2403 BFQQ_CLOSE_THR;
2404}
2405
2406static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2407 struct bfq_queue *bfqq,
2408 sector_t sector)
2409{
2410 struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
2411 struct rb_node *parent, *node;
2412 struct bfq_queue *__bfqq;
2413
2414 if (RB_EMPTY_ROOT(root))
2415 return NULL;
2416
2417 /*
2418 * First, if we find a request starting at the end of the last
2419 * request, choose it.
2420 */
2421 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2422 if (__bfqq)
2423 return __bfqq;
2424
2425 /*
2426 * If the exact sector wasn't found, the parent of the NULL leaf
2427 * will contain the closest sector (rq_pos_tree sorted by
2428 * next_request position).
2429 */
2430 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2431 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2432 return __bfqq;
2433
2434 if (blk_rq_pos(__bfqq->next_rq) < sector)
2435 node = rb_next(&__bfqq->pos_node);
2436 else
2437 node = rb_prev(&__bfqq->pos_node);
2438 if (!node)
2439 return NULL;
2440
2441 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2442 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2443 return __bfqq;
2444
2445 return NULL;
2446}
2447
2448static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2449 struct bfq_queue *cur_bfqq,
2450 sector_t sector)
2451{
2452 struct bfq_queue *bfqq;
2453
2454 /*
2455 * We shall notice if some of the queues are cooperating,
2456 * e.g., working closely on the same area of the device. In
2457 * that case, we can group them together and: 1) don't waste
2458 * time idling, and 2) serve the union of their requests in
2459 * the best possible order for throughput.
2460 */
2461 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2462 if (!bfqq || bfqq == cur_bfqq)
2463 return NULL;
2464
2465 return bfqq;
2466}
2467
2468static struct bfq_queue *
2469bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2470{
2471 int process_refs, new_process_refs;
2472 struct bfq_queue *__bfqq;
2473
2474 /*
2475 * If there are no process references on the new_bfqq, then it is
2476 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2477 * may have dropped their last reference (not just their last process
2478 * reference).
2479 */
2480 if (!bfqq_process_refs(new_bfqq))
2481 return NULL;
2482
2483 /* Avoid a circular list and skip interim queue merges. */
2484 while ((__bfqq = new_bfqq->new_bfqq)) {
2485 if (__bfqq == bfqq)
2486 return NULL;
2487 new_bfqq = __bfqq;
2488 }
2489
2490 process_refs = bfqq_process_refs(bfqq);
2491 new_process_refs = bfqq_process_refs(new_bfqq);
2492 /*
2493 * If the process for the bfqq has gone away, there is no
2494 * sense in merging the queues.
2495 */
2496 if (process_refs == 0 || new_process_refs == 0)
2497 return NULL;
2498
2499 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2500 new_bfqq->pid);
2501
2502 /*
2503 * Merging is just a redirection: the requests of the process
2504 * owning one of the two queues are redirected to the other queue.
2505 * The latter queue, in its turn, is set as shared if this is the
2506 * first time that the requests of some process are redirected to
2507 * it.
2508 *
2509 * We redirect bfqq to new_bfqq and not the opposite, because
2510 * we are in the context of the process owning bfqq, thus we
2511 * have the io_cq of this process. So we can immediately
2512 * configure this io_cq to redirect the requests of the
2513 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2514 * not available any more (new_bfqq->bic == NULL).
2515 *
2516 * Anyway, even in case new_bfqq coincides with the in-service
2517 * queue, redirecting requests the in-service queue is the
2518 * best option, as we feed the in-service queue with new
2519 * requests close to the last request served and, by doing so,
2520 * are likely to increase the throughput.
2521 */
2522 bfqq->new_bfqq = new_bfqq;
2523 new_bfqq->ref += process_refs;
2524 return new_bfqq;
2525}
2526
2527static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2528 struct bfq_queue *new_bfqq)
2529{
2530 if (bfq_too_late_for_merging(new_bfqq))
2531 return false;
2532
2533 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2534 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2535 return false;
2536
2537 /*
2538 * If either of the queues has already been detected as seeky,
2539 * then merging it with the other queue is unlikely to lead to
2540 * sequential I/O.
2541 */
2542 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2543 return false;
2544
2545 /*
2546 * Interleaved I/O is known to be done by (some) applications
2547 * only for reads, so it does not make sense to merge async
2548 * queues.
2549 */
2550 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2551 return false;
2552
2553 return true;
2554}
2555
2556/*
2557 * Attempt to schedule a merge of bfqq with the currently in-service
2558 * queue or with a close queue among the scheduled queues. Return
2559 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2560 * structure otherwise.
2561 *
2562 * The OOM queue is not allowed to participate to cooperation: in fact, since
2563 * the requests temporarily redirected to the OOM queue could be redirected
2564 * again to dedicated queues at any time, the state needed to correctly
2565 * handle merging with the OOM queue would be quite complex and expensive
2566 * to maintain. Besides, in such a critical condition as an out of memory,
2567 * the benefits of queue merging may be little relevant, or even negligible.
2568 *
2569 * WARNING: queue merging may impair fairness among non-weight raised
2570 * queues, for at least two reasons: 1) the original weight of a
2571 * merged queue may change during the merged state, 2) even being the
2572 * weight the same, a merged queue may be bloated with many more
2573 * requests than the ones produced by its originally-associated
2574 * process.
2575 */
2576static struct bfq_queue *
2577bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2578 void *io_struct, bool request)
2579{
2580 struct bfq_queue *in_service_bfqq, *new_bfqq;
2581
2582 /*
2583 * Do not perform queue merging if the device is non
2584 * rotational and performs internal queueing. In fact, such a
2585 * device reaches a high speed through internal parallelism
2586 * and pipelining. This means that, to reach a high
2587 * throughput, it must have many requests enqueued at the same
2588 * time. But, in this configuration, the internal scheduling
2589 * algorithm of the device does exactly the job of queue
2590 * merging: it reorders requests so as to obtain as much as
2591 * possible a sequential I/O pattern. As a consequence, with
2592 * the workload generated by processes doing interleaved I/O,
2593 * the throughput reached by the device is likely to be the
2594 * same, with and without queue merging.
2595 *
2596 * Disabling merging also provides a remarkable benefit in
2597 * terms of throughput. Merging tends to make many workloads
2598 * artificially more uneven, because of shared queues
2599 * remaining non empty for incomparably more time than
2600 * non-merged queues. This may accentuate workload
2601 * asymmetries. For example, if one of the queues in a set of
2602 * merged queues has a higher weight than a normal queue, then
2603 * the shared queue may inherit such a high weight and, by
2604 * staying almost always active, may force BFQ to perform I/O
2605 * plugging most of the time. This evidently makes it harder
2606 * for BFQ to let the device reach a high throughput.
2607 *
2608 * Finally, the likely() macro below is not used because one
2609 * of the two branches is more likely than the other, but to
2610 * have the code path after the following if() executed as
2611 * fast as possible for the case of a non rotational device
2612 * with queueing. We want it because this is the fastest kind
2613 * of device. On the opposite end, the likely() may lengthen
2614 * the execution time of BFQ for the case of slower devices
2615 * (rotational or at least without queueing). But in this case
2616 * the execution time of BFQ matters very little, if not at
2617 * all.
2618 */
2619 if (likely(bfqd->nonrot_with_queueing))
2620 return NULL;
2621
2622 /*
2623 * Prevent bfqq from being merged if it has been created too
2624 * long ago. The idea is that true cooperating processes, and
2625 * thus their associated bfq_queues, are supposed to be
2626 * created shortly after each other. This is the case, e.g.,
2627 * for KVM/QEMU and dump I/O threads. Basing on this
2628 * assumption, the following filtering greatly reduces the
2629 * probability that two non-cooperating processes, which just
2630 * happen to do close I/O for some short time interval, have
2631 * their queues merged by mistake.
2632 */
2633 if (bfq_too_late_for_merging(bfqq))
2634 return NULL;
2635
2636 if (bfqq->new_bfqq)
2637 return bfqq->new_bfqq;
2638
2639 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2640 return NULL;
2641
2642 /* If there is only one backlogged queue, don't search. */
2643 if (bfq_tot_busy_queues(bfqd) == 1)
2644 return NULL;
2645
2646 in_service_bfqq = bfqd->in_service_queue;
2647
2648 if (in_service_bfqq && in_service_bfqq != bfqq &&
2649 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2650 bfq_rq_close_to_sector(io_struct, request,
2651 bfqd->in_serv_last_pos) &&
2652 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2653 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2654 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2655 if (new_bfqq)
2656 return new_bfqq;
2657 }
2658 /*
2659 * Check whether there is a cooperator among currently scheduled
2660 * queues. The only thing we need is that the bio/request is not
2661 * NULL, as we need it to establish whether a cooperator exists.
2662 */
2663 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2664 bfq_io_struct_pos(io_struct, request));
2665
2666 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2667 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2668 return bfq_setup_merge(bfqq, new_bfqq);
2669
2670 return NULL;
2671}
2672
2673static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2674{
2675 struct bfq_io_cq *bic = bfqq->bic;
2676
2677 /*
2678 * If !bfqq->bic, the queue is already shared or its requests
2679 * have already been redirected to a shared queue; both idle window
2680 * and weight raising state have already been saved. Do nothing.
2681 */
2682 if (!bic)
2683 return;
2684
2685 bic->saved_weight = bfqq->entity.orig_weight;
2686 bic->saved_ttime = bfqq->ttime;
2687 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2688 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2689 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2690 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2691 if (unlikely(bfq_bfqq_just_created(bfqq) &&
2692 !bfq_bfqq_in_large_burst(bfqq) &&
2693 bfqq->bfqd->low_latency)) {
2694 /*
2695 * bfqq being merged right after being created: bfqq
2696 * would have deserved interactive weight raising, but
2697 * did not make it to be set in a weight-raised state,
2698 * because of this early merge. Store directly the
2699 * weight-raising state that would have been assigned
2700 * to bfqq, so that to avoid that bfqq unjustly fails
2701 * to enjoy weight raising if split soon.
2702 */
2703 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
2704 bic->saved_wr_start_at_switch_to_srt = bfq_smallest_from_now();
2705 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2706 bic->saved_last_wr_start_finish = jiffies;
2707 } else {
2708 bic->saved_wr_coeff = bfqq->wr_coeff;
2709 bic->saved_wr_start_at_switch_to_srt =
2710 bfqq->wr_start_at_switch_to_srt;
2711 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2712 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2713 }
2714}
2715
2716
2717static
2718void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
2719{
2720 /*
2721 * To prevent bfqq's service guarantees from being violated,
2722 * bfqq may be left busy, i.e., queued for service, even if
2723 * empty (see comments in __bfq_bfqq_expire() for
2724 * details). But, if no process will send requests to bfqq any
2725 * longer, then there is no point in keeping bfqq queued for
2726 * service. In addition, keeping bfqq queued for service, but
2727 * with no process ref any longer, may have caused bfqq to be
2728 * freed when dequeued from service. But this is assumed to
2729 * never happen.
2730 */
2731 if (bfq_bfqq_busy(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) &&
2732 bfqq != bfqd->in_service_queue)
2733 bfq_del_bfqq_busy(bfqd, bfqq, false);
2734
2735 bfq_put_queue(bfqq);
2736}
2737
2738static void
2739bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2740 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2741{
2742 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2743 (unsigned long)new_bfqq->pid);
2744 /* Save weight raising and idle window of the merged queues */
2745 bfq_bfqq_save_state(bfqq);
2746 bfq_bfqq_save_state(new_bfqq);
2747 if (bfq_bfqq_IO_bound(bfqq))
2748 bfq_mark_bfqq_IO_bound(new_bfqq);
2749 bfq_clear_bfqq_IO_bound(bfqq);
2750
2751 /*
2752 * If bfqq is weight-raised, then let new_bfqq inherit
2753 * weight-raising. To reduce false positives, neglect the case
2754 * where bfqq has just been created, but has not yet made it
2755 * to be weight-raised (which may happen because EQM may merge
2756 * bfqq even before bfq_add_request is executed for the first
2757 * time for bfqq). Handling this case would however be very
2758 * easy, thanks to the flag just_created.
2759 */
2760 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2761 new_bfqq->wr_coeff = bfqq->wr_coeff;
2762 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2763 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2764 new_bfqq->wr_start_at_switch_to_srt =
2765 bfqq->wr_start_at_switch_to_srt;
2766 if (bfq_bfqq_busy(new_bfqq))
2767 bfqd->wr_busy_queues++;
2768 new_bfqq->entity.prio_changed = 1;
2769 }
2770
2771 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2772 bfqq->wr_coeff = 1;
2773 bfqq->entity.prio_changed = 1;
2774 if (bfq_bfqq_busy(bfqq))
2775 bfqd->wr_busy_queues--;
2776 }
2777
2778 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2779 bfqd->wr_busy_queues);
2780
2781 /*
2782 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2783 */
2784 bic_set_bfqq(bic, new_bfqq, 1);
2785 bfq_mark_bfqq_coop(new_bfqq);
2786 /*
2787 * new_bfqq now belongs to at least two bics (it is a shared queue):
2788 * set new_bfqq->bic to NULL. bfqq either:
2789 * - does not belong to any bic any more, and hence bfqq->bic must
2790 * be set to NULL, or
2791 * - is a queue whose owning bics have already been redirected to a
2792 * different queue, hence the queue is destined to not belong to
2793 * any bic soon and bfqq->bic is already NULL (therefore the next
2794 * assignment causes no harm).
2795 */
2796 new_bfqq->bic = NULL;
2797 /*
2798 * If the queue is shared, the pid is the pid of one of the associated
2799 * processes. Which pid depends on the exact sequence of merge events
2800 * the queue underwent. So printing such a pid is useless and confusing
2801 * because it reports a random pid between those of the associated
2802 * processes.
2803 * We mark such a queue with a pid -1, and then print SHARED instead of
2804 * a pid in logging messages.
2805 */
2806 new_bfqq->pid = -1;
2807 bfqq->bic = NULL;
2808 bfq_release_process_ref(bfqd, bfqq);
2809}
2810
2811static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2812 struct bio *bio)
2813{
2814 struct bfq_data *bfqd = q->elevator->elevator_data;
2815 bool is_sync = op_is_sync(bio->bi_opf);
2816 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2817
2818 /*
2819 * Disallow merge of a sync bio into an async request.
2820 */
2821 if (is_sync && !rq_is_sync(rq))
2822 return false;
2823
2824 /*
2825 * Lookup the bfqq that this bio will be queued with. Allow
2826 * merge only if rq is queued there.
2827 */
2828 if (!bfqq)
2829 return false;
2830
2831 /*
2832 * We take advantage of this function to perform an early merge
2833 * of the queues of possible cooperating processes.
2834 */
2835 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2836 if (new_bfqq) {
2837 /*
2838 * bic still points to bfqq, then it has not yet been
2839 * redirected to some other bfq_queue, and a queue
2840 * merge between bfqq and new_bfqq can be safely
2841 * fulfilled, i.e., bic can be redirected to new_bfqq
2842 * and bfqq can be put.
2843 */
2844 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2845 new_bfqq);
2846 /*
2847 * If we get here, bio will be queued into new_queue,
2848 * so use new_bfqq to decide whether bio and rq can be
2849 * merged.
2850 */
2851 bfqq = new_bfqq;
2852
2853 /*
2854 * Change also bqfd->bio_bfqq, as
2855 * bfqd->bio_bic now points to new_bfqq, and
2856 * this function may be invoked again (and then may
2857 * use again bqfd->bio_bfqq).
2858 */
2859 bfqd->bio_bfqq = bfqq;
2860 }
2861
2862 return bfqq == RQ_BFQQ(rq);
2863}
2864
2865/*
2866 * Set the maximum time for the in-service queue to consume its
2867 * budget. This prevents seeky processes from lowering the throughput.
2868 * In practice, a time-slice service scheme is used with seeky
2869 * processes.
2870 */
2871static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2872 struct bfq_queue *bfqq)
2873{
2874 unsigned int timeout_coeff;
2875
2876 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2877 timeout_coeff = 1;
2878 else
2879 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2880
2881 bfqd->last_budget_start = ktime_get();
2882
2883 bfqq->budget_timeout = jiffies +
2884 bfqd->bfq_timeout * timeout_coeff;
2885}
2886
2887static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2888 struct bfq_queue *bfqq)
2889{
2890 if (bfqq) {
2891 bfq_clear_bfqq_fifo_expire(bfqq);
2892
2893 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2894
2895 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2896 bfqq->wr_coeff > 1 &&
2897 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2898 time_is_before_jiffies(bfqq->budget_timeout)) {
2899 /*
2900 * For soft real-time queues, move the start
2901 * of the weight-raising period forward by the
2902 * time the queue has not received any
2903 * service. Otherwise, a relatively long
2904 * service delay is likely to cause the
2905 * weight-raising period of the queue to end,
2906 * because of the short duration of the
2907 * weight-raising period of a soft real-time
2908 * queue. It is worth noting that this move
2909 * is not so dangerous for the other queues,
2910 * because soft real-time queues are not
2911 * greedy.
2912 *
2913 * To not add a further variable, we use the
2914 * overloaded field budget_timeout to
2915 * determine for how long the queue has not
2916 * received service, i.e., how much time has
2917 * elapsed since the queue expired. However,
2918 * this is a little imprecise, because
2919 * budget_timeout is set to jiffies if bfqq
2920 * not only expires, but also remains with no
2921 * request.
2922 */
2923 if (time_after(bfqq->budget_timeout,
2924 bfqq->last_wr_start_finish))
2925 bfqq->last_wr_start_finish +=
2926 jiffies - bfqq->budget_timeout;
2927 else
2928 bfqq->last_wr_start_finish = jiffies;
2929 }
2930
2931 bfq_set_budget_timeout(bfqd, bfqq);
2932 bfq_log_bfqq(bfqd, bfqq,
2933 "set_in_service_queue, cur-budget = %d",
2934 bfqq->entity.budget);
2935 }
2936
2937 bfqd->in_service_queue = bfqq;
2938}
2939
2940/*
2941 * Get and set a new queue for service.
2942 */
2943static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2944{
2945 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2946
2947 __bfq_set_in_service_queue(bfqd, bfqq);
2948 return bfqq;
2949}
2950
2951static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2952{
2953 struct bfq_queue *bfqq = bfqd->in_service_queue;
2954 u32 sl;
2955
2956 bfq_mark_bfqq_wait_request(bfqq);
2957
2958 /*
2959 * We don't want to idle for seeks, but we do want to allow
2960 * fair distribution of slice time for a process doing back-to-back
2961 * seeks. So allow a little bit of time for him to submit a new rq.
2962 */
2963 sl = bfqd->bfq_slice_idle;
2964 /*
2965 * Unless the queue is being weight-raised or the scenario is
2966 * asymmetric, grant only minimum idle time if the queue
2967 * is seeky. A long idling is preserved for a weight-raised
2968 * queue, or, more in general, in an asymmetric scenario,
2969 * because a long idling is needed for guaranteeing to a queue
2970 * its reserved share of the throughput (in particular, it is
2971 * needed if the queue has a higher weight than some other
2972 * queue).
2973 */
2974 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
2975 !bfq_asymmetric_scenario(bfqd, bfqq))
2976 sl = min_t(u64, sl, BFQ_MIN_TT);
2977 else if (bfqq->wr_coeff > 1)
2978 sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
2979
2980 bfqd->last_idling_start = ktime_get();
2981 bfqd->last_idling_start_jiffies = jiffies;
2982
2983 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
2984 HRTIMER_MODE_REL);
2985 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
2986}
2987
2988/*
2989 * In autotuning mode, max_budget is dynamically recomputed as the
2990 * amount of sectors transferred in timeout at the estimated peak
2991 * rate. This enables BFQ to utilize a full timeslice with a full
2992 * budget, even if the in-service queue is served at peak rate. And
2993 * this maximises throughput with sequential workloads.
2994 */
2995static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
2996{
2997 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
2998 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
2999}
3000
3001/*
3002 * Update parameters related to throughput and responsiveness, as a
3003 * function of the estimated peak rate. See comments on
3004 * bfq_calc_max_budget(), and on the ref_wr_duration array.
3005 */
3006static void update_thr_responsiveness_params(struct bfq_data *bfqd)
3007{
3008 if (bfqd->bfq_user_max_budget == 0) {
3009 bfqd->bfq_max_budget =
3010 bfq_calc_max_budget(bfqd);
3011 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
3012 }
3013}
3014
3015static void bfq_reset_rate_computation(struct bfq_data *bfqd,
3016 struct request *rq)
3017{
3018 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
3019 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
3020 bfqd->peak_rate_samples = 1;
3021 bfqd->sequential_samples = 0;
3022 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
3023 blk_rq_sectors(rq);
3024 } else /* no new rq dispatched, just reset the number of samples */
3025 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
3026
3027 bfq_log(bfqd,
3028 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
3029 bfqd->peak_rate_samples, bfqd->sequential_samples,
3030 bfqd->tot_sectors_dispatched);
3031}
3032
3033static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
3034{
3035 u32 rate, weight, divisor;
3036
3037 /*
3038 * For the convergence property to hold (see comments on
3039 * bfq_update_peak_rate()) and for the assessment to be
3040 * reliable, a minimum number of samples must be present, and
3041 * a minimum amount of time must have elapsed. If not so, do
3042 * not compute new rate. Just reset parameters, to get ready
3043 * for a new evaluation attempt.
3044 */
3045 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
3046 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
3047 goto reset_computation;
3048
3049 /*
3050 * If a new request completion has occurred after last
3051 * dispatch, then, to approximate the rate at which requests
3052 * have been served by the device, it is more precise to
3053 * extend the observation interval to the last completion.
3054 */
3055 bfqd->delta_from_first =
3056 max_t(u64, bfqd->delta_from_first,
3057 bfqd->last_completion - bfqd->first_dispatch);
3058
3059 /*
3060 * Rate computed in sects/usec, and not sects/nsec, for
3061 * precision issues.
3062 */
3063 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
3064 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
3065
3066 /*
3067 * Peak rate not updated if:
3068 * - the percentage of sequential dispatches is below 3/4 of the
3069 * total, and rate is below the current estimated peak rate
3070 * - rate is unreasonably high (> 20M sectors/sec)
3071 */
3072 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
3073 rate <= bfqd->peak_rate) ||
3074 rate > 20<<BFQ_RATE_SHIFT)
3075 goto reset_computation;
3076
3077 /*
3078 * We have to update the peak rate, at last! To this purpose,
3079 * we use a low-pass filter. We compute the smoothing constant
3080 * of the filter as a function of the 'weight' of the new
3081 * measured rate.
3082 *
3083 * As can be seen in next formulas, we define this weight as a
3084 * quantity proportional to how sequential the workload is,
3085 * and to how long the observation time interval is.
3086 *
3087 * The weight runs from 0 to 8. The maximum value of the
3088 * weight, 8, yields the minimum value for the smoothing
3089 * constant. At this minimum value for the smoothing constant,
3090 * the measured rate contributes for half of the next value of
3091 * the estimated peak rate.
3092 *
3093 * So, the first step is to compute the weight as a function
3094 * of how sequential the workload is. Note that the weight
3095 * cannot reach 9, because bfqd->sequential_samples cannot
3096 * become equal to bfqd->peak_rate_samples, which, in its
3097 * turn, holds true because bfqd->sequential_samples is not
3098 * incremented for the first sample.
3099 */
3100 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
3101
3102 /*
3103 * Second step: further refine the weight as a function of the
3104 * duration of the observation interval.
3105 */
3106 weight = min_t(u32, 8,
3107 div_u64(weight * bfqd->delta_from_first,
3108 BFQ_RATE_REF_INTERVAL));
3109
3110 /*
3111 * Divisor ranging from 10, for minimum weight, to 2, for
3112 * maximum weight.
3113 */
3114 divisor = 10 - weight;
3115
3116 /*
3117 * Finally, update peak rate:
3118 *
3119 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
3120 */
3121 bfqd->peak_rate *= divisor-1;
3122 bfqd->peak_rate /= divisor;
3123 rate /= divisor; /* smoothing constant alpha = 1/divisor */
3124
3125 bfqd->peak_rate += rate;
3126
3127 /*
3128 * For a very slow device, bfqd->peak_rate can reach 0 (see
3129 * the minimum representable values reported in the comments
3130 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3131 * divisions by zero where bfqd->peak_rate is used as a
3132 * divisor.
3133 */
3134 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
3135
3136 update_thr_responsiveness_params(bfqd);
3137
3138reset_computation:
3139 bfq_reset_rate_computation(bfqd, rq);
3140}
3141
3142/*
3143 * Update the read/write peak rate (the main quantity used for
3144 * auto-tuning, see update_thr_responsiveness_params()).
3145 *
3146 * It is not trivial to estimate the peak rate (correctly): because of
3147 * the presence of sw and hw queues between the scheduler and the
3148 * device components that finally serve I/O requests, it is hard to
3149 * say exactly when a given dispatched request is served inside the
3150 * device, and for how long. As a consequence, it is hard to know
3151 * precisely at what rate a given set of requests is actually served
3152 * by the device.
3153 *
3154 * On the opposite end, the dispatch time of any request is trivially
3155 * available, and, from this piece of information, the "dispatch rate"
3156 * of requests can be immediately computed. So, the idea in the next
3157 * function is to use what is known, namely request dispatch times
3158 * (plus, when useful, request completion times), to estimate what is
3159 * unknown, namely in-device request service rate.
3160 *
3161 * The main issue is that, because of the above facts, the rate at
3162 * which a certain set of requests is dispatched over a certain time
3163 * interval can vary greatly with respect to the rate at which the
3164 * same requests are then served. But, since the size of any
3165 * intermediate queue is limited, and the service scheme is lossless
3166 * (no request is silently dropped), the following obvious convergence
3167 * property holds: the number of requests dispatched MUST become
3168 * closer and closer to the number of requests completed as the
3169 * observation interval grows. This is the key property used in
3170 * the next function to estimate the peak service rate as a function
3171 * of the observed dispatch rate. The function assumes to be invoked
3172 * on every request dispatch.
3173 */
3174static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
3175{
3176 u64 now_ns = ktime_get_ns();
3177
3178 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
3179 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
3180 bfqd->peak_rate_samples);
3181 bfq_reset_rate_computation(bfqd, rq);
3182 goto update_last_values; /* will add one sample */
3183 }
3184
3185 /*
3186 * Device idle for very long: the observation interval lasting
3187 * up to this dispatch cannot be a valid observation interval
3188 * for computing a new peak rate (similarly to the late-
3189 * completion event in bfq_completed_request()). Go to
3190 * update_rate_and_reset to have the following three steps
3191 * taken:
3192 * - close the observation interval at the last (previous)
3193 * request dispatch or completion
3194 * - compute rate, if possible, for that observation interval
3195 * - start a new observation interval with this dispatch
3196 */
3197 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
3198 bfqd->rq_in_driver == 0)
3199 goto update_rate_and_reset;
3200
3201 /* Update sampling information */
3202 bfqd->peak_rate_samples++;
3203
3204 if ((bfqd->rq_in_driver > 0 ||
3205 now_ns - bfqd->last_completion < BFQ_MIN_TT)
3206 && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
3207 bfqd->sequential_samples++;
3208
3209 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
3210
3211 /* Reset max observed rq size every 32 dispatches */
3212 if (likely(bfqd->peak_rate_samples % 32))
3213 bfqd->last_rq_max_size =
3214 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
3215 else
3216 bfqd->last_rq_max_size = blk_rq_sectors(rq);
3217
3218 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
3219
3220 /* Target observation interval not yet reached, go on sampling */
3221 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
3222 goto update_last_values;
3223
3224update_rate_and_reset:
3225 bfq_update_rate_reset(bfqd, rq);
3226update_last_values:
3227 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
3228 if (RQ_BFQQ(rq) == bfqd->in_service_queue)
3229 bfqd->in_serv_last_pos = bfqd->last_position;
3230 bfqd->last_dispatch = now_ns;
3231}
3232
3233/*
3234 * Remove request from internal lists.
3235 */
3236static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
3237{
3238 struct bfq_queue *bfqq = RQ_BFQQ(rq);
3239
3240 /*
3241 * For consistency, the next instruction should have been
3242 * executed after removing the request from the queue and
3243 * dispatching it. We execute instead this instruction before
3244 * bfq_remove_request() (and hence introduce a temporary
3245 * inconsistency), for efficiency. In fact, should this
3246 * dispatch occur for a non in-service bfqq, this anticipated
3247 * increment prevents two counters related to bfqq->dispatched
3248 * from risking to be, first, uselessly decremented, and then
3249 * incremented again when the (new) value of bfqq->dispatched
3250 * happens to be taken into account.
3251 */
3252 bfqq->dispatched++;
3253 bfq_update_peak_rate(q->elevator->elevator_data, rq);
3254
3255 bfq_remove_request(q, rq);
3256}
3257
3258/*
3259 * There is a case where idling does not have to be performed for
3260 * throughput concerns, but to preserve the throughput share of
3261 * the process associated with bfqq.
3262 *
3263 * To introduce this case, we can note that allowing the drive
3264 * to enqueue more than one request at a time, and hence
3265 * delegating de facto final scheduling decisions to the
3266 * drive's internal scheduler, entails loss of control on the
3267 * actual request service order. In particular, the critical
3268 * situation is when requests from different processes happen
3269 * to be present, at the same time, in the internal queue(s)
3270 * of the drive. In such a situation, the drive, by deciding
3271 * the service order of the internally-queued requests, does
3272 * determine also the actual throughput distribution among
3273 * these processes. But the drive typically has no notion or
3274 * concern about per-process throughput distribution, and
3275 * makes its decisions only on a per-request basis. Therefore,
3276 * the service distribution enforced by the drive's internal
3277 * scheduler is likely to coincide with the desired throughput
3278 * distribution only in a completely symmetric, or favorably
3279 * skewed scenario where:
3280 * (i-a) each of these processes must get the same throughput as
3281 * the others,
3282 * (i-b) in case (i-a) does not hold, it holds that the process
3283 * associated with bfqq must receive a lower or equal
3284 * throughput than any of the other processes;
3285 * (ii) the I/O of each process has the same properties, in
3286 * terms of locality (sequential or random), direction
3287 * (reads or writes), request sizes, greediness
3288 * (from I/O-bound to sporadic), and so on;
3289
3290 * In fact, in such a scenario, the drive tends to treat the requests
3291 * of each process in about the same way as the requests of the
3292 * others, and thus to provide each of these processes with about the
3293 * same throughput. This is exactly the desired throughput
3294 * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3295 * even more convenient distribution for (the process associated with)
3296 * bfqq.
3297 *
3298 * In contrast, in any asymmetric or unfavorable scenario, device
3299 * idling (I/O-dispatch plugging) is certainly needed to guarantee
3300 * that bfqq receives its assigned fraction of the device throughput
3301 * (see [1] for details).
3302 *
3303 * The problem is that idling may significantly reduce throughput with
3304 * certain combinations of types of I/O and devices. An important
3305 * example is sync random I/O on flash storage with command
3306 * queueing. So, unless bfqq falls in cases where idling also boosts
3307 * throughput, it is important to check conditions (i-a), i(-b) and
3308 * (ii) accurately, so as to avoid idling when not strictly needed for
3309 * service guarantees.
3310 *
3311 * Unfortunately, it is extremely difficult to thoroughly check
3312 * condition (ii). And, in case there are active groups, it becomes
3313 * very difficult to check conditions (i-a) and (i-b) too. In fact,
3314 * if there are active groups, then, for conditions (i-a) or (i-b) to
3315 * become false 'indirectly', it is enough that an active group
3316 * contains more active processes or sub-groups than some other active
3317 * group. More precisely, for conditions (i-a) or (i-b) to become
3318 * false because of such a group, it is not even necessary that the
3319 * group is (still) active: it is sufficient that, even if the group
3320 * has become inactive, some of its descendant processes still have
3321 * some request already dispatched but still waiting for
3322 * completion. In fact, requests have still to be guaranteed their
3323 * share of the throughput even after being dispatched. In this
3324 * respect, it is easy to show that, if a group frequently becomes
3325 * inactive while still having in-flight requests, and if, when this
3326 * happens, the group is not considered in the calculation of whether
3327 * the scenario is asymmetric, then the group may fail to be
3328 * guaranteed its fair share of the throughput (basically because
3329 * idling may not be performed for the descendant processes of the
3330 * group, but it had to be). We address this issue with the following
3331 * bi-modal behavior, implemented in the function
3332 * bfq_asymmetric_scenario().
3333 *
3334 * If there are groups with requests waiting for completion
3335 * (as commented above, some of these groups may even be
3336 * already inactive), then the scenario is tagged as
3337 * asymmetric, conservatively, without checking any of the
3338 * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3339 * This behavior matches also the fact that groups are created
3340 * exactly if controlling I/O is a primary concern (to
3341 * preserve bandwidth and latency guarantees).
3342 *
3343 * On the opposite end, if there are no groups with requests waiting
3344 * for completion, then only conditions (i-a) and (i-b) are actually
3345 * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3346 * idling is not performed, regardless of whether condition (ii)
3347 * holds. In other words, only if conditions (i-a) and (i-b) do not
3348 * hold, then idling is allowed, and the device tends to be prevented
3349 * from queueing many requests, possibly of several processes. Since
3350 * there are no groups with requests waiting for completion, then, to
3351 * control conditions (i-a) and (i-b) it is enough to check just
3352 * whether all the queues with requests waiting for completion also
3353 * have the same weight.
3354 *
3355 * Not checking condition (ii) evidently exposes bfqq to the
3356 * risk of getting less throughput than its fair share.
3357 * However, for queues with the same weight, a further
3358 * mechanism, preemption, mitigates or even eliminates this
3359 * problem. And it does so without consequences on overall
3360 * throughput. This mechanism and its benefits are explained
3361 * in the next three paragraphs.
3362 *
3363 * Even if a queue, say Q, is expired when it remains idle, Q
3364 * can still preempt the new in-service queue if the next
3365 * request of Q arrives soon (see the comments on
3366 * bfq_bfqq_update_budg_for_activation). If all queues and
3367 * groups have the same weight, this form of preemption,
3368 * combined with the hole-recovery heuristic described in the
3369 * comments on function bfq_bfqq_update_budg_for_activation,
3370 * are enough to preserve a correct bandwidth distribution in
3371 * the mid term, even without idling. In fact, even if not
3372 * idling allows the internal queues of the device to contain
3373 * many requests, and thus to reorder requests, we can rather
3374 * safely assume that the internal scheduler still preserves a
3375 * minimum of mid-term fairness.
3376 *
3377 * More precisely, this preemption-based, idleless approach
3378 * provides fairness in terms of IOPS, and not sectors per
3379 * second. This can be seen with a simple example. Suppose
3380 * that there are two queues with the same weight, but that
3381 * the first queue receives requests of 8 sectors, while the
3382 * second queue receives requests of 1024 sectors. In
3383 * addition, suppose that each of the two queues contains at
3384 * most one request at a time, which implies that each queue
3385 * always remains idle after it is served. Finally, after
3386 * remaining idle, each queue receives very quickly a new
3387 * request. It follows that the two queues are served
3388 * alternatively, preempting each other if needed. This
3389 * implies that, although both queues have the same weight,
3390 * the queue with large requests receives a service that is
3391 * 1024/8 times as high as the service received by the other
3392 * queue.
3393 *
3394 * The motivation for using preemption instead of idling (for
3395 * queues with the same weight) is that, by not idling,
3396 * service guarantees are preserved (completely or at least in
3397 * part) without minimally sacrificing throughput. And, if
3398 * there is no active group, then the primary expectation for
3399 * this device is probably a high throughput.
3400 *
3401 * We are now left only with explaining the two sub-conditions in the
3402 * additional compound condition that is checked below for deciding
3403 * whether the scenario is asymmetric. To explain the first
3404 * sub-condition, we need to add that the function
3405 * bfq_asymmetric_scenario checks the weights of only
3406 * non-weight-raised queues, for efficiency reasons (see comments on
3407 * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3408 * is checked explicitly here. More precisely, the compound condition
3409 * below takes into account also the fact that, even if bfqq is being
3410 * weight-raised, the scenario is still symmetric if all queues with
3411 * requests waiting for completion happen to be
3412 * weight-raised. Actually, we should be even more precise here, and
3413 * differentiate between interactive weight raising and soft real-time
3414 * weight raising.
3415 *
3416 * The second sub-condition checked in the compound condition is
3417 * whether there is a fair amount of already in-flight I/O not
3418 * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3419 * following reason. The drive may decide to serve in-flight
3420 * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3421 * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3422 * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3423 * basically uncontrolled amount of I/O from other queues may be
3424 * dispatched too, possibly causing the service of bfqq's I/O to be
3425 * delayed even longer in the drive. This problem gets more and more
3426 * serious as the speed and the queue depth of the drive grow,
3427 * because, as these two quantities grow, the probability to find no
3428 * queue busy but many requests in flight grows too. By contrast,
3429 * plugging I/O dispatching minimizes the delay induced by already
3430 * in-flight I/O, and enables bfqq to recover the bandwidth it may
3431 * lose because of this delay.
3432 *
3433 * As a side note, it is worth considering that the above
3434 * device-idling countermeasures may however fail in the following
3435 * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3436 * in a time period during which all symmetry sub-conditions hold, and
3437 * therefore the device is allowed to enqueue many requests, but at
3438 * some later point in time some sub-condition stops to hold, then it
3439 * may become impossible to make requests be served in the desired
3440 * order until all the requests already queued in the device have been
3441 * served. The last sub-condition commented above somewhat mitigates
3442 * this problem for weight-raised queues.
3443 */
3444static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3445 struct bfq_queue *bfqq)
3446{
3447 return (bfqq->wr_coeff > 1 &&
3448 (bfqd->wr_busy_queues <
3449 bfq_tot_busy_queues(bfqd) ||
3450 bfqd->rq_in_driver >=
3451 bfqq->dispatched + 4)) ||
3452 bfq_asymmetric_scenario(bfqd, bfqq);
3453}
3454
3455static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3456 enum bfqq_expiration reason)
3457{
3458 /*
3459 * If this bfqq is shared between multiple processes, check
3460 * to make sure that those processes are still issuing I/Os
3461 * within the mean seek distance. If not, it may be time to
3462 * break the queues apart again.
3463 */
3464 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
3465 bfq_mark_bfqq_split_coop(bfqq);
3466
3467 /*
3468 * Consider queues with a higher finish virtual time than
3469 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3470 * true, then bfqq's bandwidth would be violated if an
3471 * uncontrolled amount of I/O from these queues were
3472 * dispatched while bfqq is waiting for its new I/O to
3473 * arrive. This is exactly what may happen if this is a forced
3474 * expiration caused by a preemption attempt, and if bfqq is
3475 * not re-scheduled. To prevent this from happening, re-queue
3476 * bfqq if it needs I/O-dispatch plugging, even if it is
3477 * empty. By doing so, bfqq is granted to be served before the
3478 * above queues (provided that bfqq is of course eligible).
3479 */
3480 if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
3481 !(reason == BFQQE_PREEMPTED &&
3482 idling_needed_for_service_guarantees(bfqd, bfqq))) {
3483 if (bfqq->dispatched == 0)
3484 /*
3485 * Overloading budget_timeout field to store
3486 * the time at which the queue remains with no
3487 * backlog and no outstanding request; used by
3488 * the weight-raising mechanism.
3489 */
3490 bfqq->budget_timeout = jiffies;
3491
3492 bfq_del_bfqq_busy(bfqd, bfqq, true);
3493 } else {
3494 bfq_requeue_bfqq(bfqd, bfqq, true);
3495 /*
3496 * Resort priority tree of potential close cooperators.
3497 * See comments on bfq_pos_tree_add_move() for the unlikely().
3498 */
3499 if (unlikely(!bfqd->nonrot_with_queueing &&
3500 !RB_EMPTY_ROOT(&bfqq->sort_list)))
3501 bfq_pos_tree_add_move(bfqd, bfqq);
3502 }
3503
3504 /*
3505 * All in-service entities must have been properly deactivated
3506 * or requeued before executing the next function, which
3507 * resets all in-service entities as no more in service. This
3508 * may cause bfqq to be freed. If this happens, the next
3509 * function returns true.
3510 */
3511 return __bfq_bfqd_reset_in_service(bfqd);
3512}
3513
3514/**
3515 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3516 * @bfqd: device data.
3517 * @bfqq: queue to update.
3518 * @reason: reason for expiration.
3519 *
3520 * Handle the feedback on @bfqq budget at queue expiration.
3521 * See the body for detailed comments.
3522 */
3523static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
3524 struct bfq_queue *bfqq,
3525 enum bfqq_expiration reason)
3526{
3527 struct request *next_rq;
3528 int budget, min_budget;
3529
3530 min_budget = bfq_min_budget(bfqd);
3531
3532 if (bfqq->wr_coeff == 1)
3533 budget = bfqq->max_budget;
3534 else /*
3535 * Use a constant, low budget for weight-raised queues,
3536 * to help achieve a low latency. Keep it slightly higher
3537 * than the minimum possible budget, to cause a little
3538 * bit fewer expirations.
3539 */
3540 budget = 2 * min_budget;
3541
3542 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
3543 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
3544 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
3545 budget, bfq_min_budget(bfqd));
3546 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
3547 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
3548
3549 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
3550 switch (reason) {
3551 /*
3552 * Caveat: in all the following cases we trade latency
3553 * for throughput.
3554 */
3555 case BFQQE_TOO_IDLE:
3556 /*
3557 * This is the only case where we may reduce
3558 * the budget: if there is no request of the
3559 * process still waiting for completion, then
3560 * we assume (tentatively) that the timer has
3561 * expired because the batch of requests of
3562 * the process could have been served with a
3563 * smaller budget. Hence, betting that
3564 * process will behave in the same way when it
3565 * becomes backlogged again, we reduce its
3566 * next budget. As long as we guess right,
3567 * this budget cut reduces the latency
3568 * experienced by the process.
3569 *
3570 * However, if there are still outstanding
3571 * requests, then the process may have not yet
3572 * issued its next request just because it is
3573 * still waiting for the completion of some of
3574 * the still outstanding ones. So in this
3575 * subcase we do not reduce its budget, on the
3576 * contrary we increase it to possibly boost
3577 * the throughput, as discussed in the
3578 * comments to the BUDGET_TIMEOUT case.
3579 */
3580 if (bfqq->dispatched > 0) /* still outstanding reqs */
3581 budget = min(budget * 2, bfqd->bfq_max_budget);
3582 else {
3583 if (budget > 5 * min_budget)
3584 budget -= 4 * min_budget;
3585 else
3586 budget = min_budget;
3587 }
3588 break;
3589 case BFQQE_BUDGET_TIMEOUT:
3590 /*
3591 * We double the budget here because it gives
3592 * the chance to boost the throughput if this
3593 * is not a seeky process (and has bumped into
3594 * this timeout because of, e.g., ZBR).
3595 */
3596 budget = min(budget * 2, bfqd->bfq_max_budget);
3597 break;
3598 case BFQQE_BUDGET_EXHAUSTED:
3599 /*
3600 * The process still has backlog, and did not
3601 * let either the budget timeout or the disk
3602 * idling timeout expire. Hence it is not
3603 * seeky, has a short thinktime and may be
3604 * happy with a higher budget too. So
3605 * definitely increase the budget of this good
3606 * candidate to boost the disk throughput.
3607 */
3608 budget = min(budget * 4, bfqd->bfq_max_budget);
3609 break;
3610 case BFQQE_NO_MORE_REQUESTS:
3611 /*
3612 * For queues that expire for this reason, it
3613 * is particularly important to keep the
3614 * budget close to the actual service they
3615 * need. Doing so reduces the timestamp
3616 * misalignment problem described in the
3617 * comments in the body of
3618 * __bfq_activate_entity. In fact, suppose
3619 * that a queue systematically expires for
3620 * BFQQE_NO_MORE_REQUESTS and presents a
3621 * new request in time to enjoy timestamp
3622 * back-shifting. The larger the budget of the
3623 * queue is with respect to the service the
3624 * queue actually requests in each service
3625 * slot, the more times the queue can be
3626 * reactivated with the same virtual finish
3627 * time. It follows that, even if this finish
3628 * time is pushed to the system virtual time
3629 * to reduce the consequent timestamp
3630 * misalignment, the queue unjustly enjoys for
3631 * many re-activations a lower finish time
3632 * than all newly activated queues.
3633 *
3634 * The service needed by bfqq is measured
3635 * quite precisely by bfqq->entity.service.
3636 * Since bfqq does not enjoy device idling,
3637 * bfqq->entity.service is equal to the number
3638 * of sectors that the process associated with
3639 * bfqq requested to read/write before waiting
3640 * for request completions, or blocking for
3641 * other reasons.
3642 */
3643 budget = max_t(int, bfqq->entity.service, min_budget);
3644 break;
3645 default:
3646 return;
3647 }
3648 } else if (!bfq_bfqq_sync(bfqq)) {
3649 /*
3650 * Async queues get always the maximum possible
3651 * budget, as for them we do not care about latency
3652 * (in addition, their ability to dispatch is limited
3653 * by the charging factor).
3654 */
3655 budget = bfqd->bfq_max_budget;
3656 }
3657
3658 bfqq->max_budget = budget;
3659
3660 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
3661 !bfqd->bfq_user_max_budget)
3662 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
3663
3664 /*
3665 * If there is still backlog, then assign a new budget, making
3666 * sure that it is large enough for the next request. Since
3667 * the finish time of bfqq must be kept in sync with the
3668 * budget, be sure to call __bfq_bfqq_expire() *after* this
3669 * update.
3670 *
3671 * If there is no backlog, then no need to update the budget;
3672 * it will be updated on the arrival of a new request.
3673 */
3674 next_rq = bfqq->next_rq;
3675 if (next_rq)
3676 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
3677 bfq_serv_to_charge(next_rq, bfqq));
3678
3679 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
3680 next_rq ? blk_rq_sectors(next_rq) : 0,
3681 bfqq->entity.budget);
3682}
3683
3684/*
3685 * Return true if the process associated with bfqq is "slow". The slow
3686 * flag is used, in addition to the budget timeout, to reduce the
3687 * amount of service provided to seeky processes, and thus reduce
3688 * their chances to lower the throughput. More details in the comments
3689 * on the function bfq_bfqq_expire().
3690 *
3691 * An important observation is in order: as discussed in the comments
3692 * on the function bfq_update_peak_rate(), with devices with internal
3693 * queues, it is hard if ever possible to know when and for how long
3694 * an I/O request is processed by the device (apart from the trivial
3695 * I/O pattern where a new request is dispatched only after the
3696 * previous one has been completed). This makes it hard to evaluate
3697 * the real rate at which the I/O requests of each bfq_queue are
3698 * served. In fact, for an I/O scheduler like BFQ, serving a
3699 * bfq_queue means just dispatching its requests during its service
3700 * slot (i.e., until the budget of the queue is exhausted, or the
3701 * queue remains idle, or, finally, a timeout fires). But, during the
3702 * service slot of a bfq_queue, around 100 ms at most, the device may
3703 * be even still processing requests of bfq_queues served in previous
3704 * service slots. On the opposite end, the requests of the in-service
3705 * bfq_queue may be completed after the service slot of the queue
3706 * finishes.
3707 *
3708 * Anyway, unless more sophisticated solutions are used
3709 * (where possible), the sum of the sizes of the requests dispatched
3710 * during the service slot of a bfq_queue is probably the only
3711 * approximation available for the service received by the bfq_queue
3712 * during its service slot. And this sum is the quantity used in this
3713 * function to evaluate the I/O speed of a process.
3714 */
3715static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3716 bool compensate, enum bfqq_expiration reason,
3717 unsigned long *delta_ms)
3718{
3719 ktime_t delta_ktime;
3720 u32 delta_usecs;
3721 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
3722
3723 if (!bfq_bfqq_sync(bfqq))
3724 return false;
3725
3726 if (compensate)
3727 delta_ktime = bfqd->last_idling_start;
3728 else
3729 delta_ktime = ktime_get();
3730 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
3731 delta_usecs = ktime_to_us(delta_ktime);
3732
3733 /* don't use too short time intervals */
3734 if (delta_usecs < 1000) {
3735 if (blk_queue_nonrot(bfqd->queue))
3736 /*
3737 * give same worst-case guarantees as idling
3738 * for seeky
3739 */
3740 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
3741 else /* charge at least one seek */
3742 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
3743
3744 return slow;
3745 }
3746
3747 *delta_ms = delta_usecs / USEC_PER_MSEC;
3748
3749 /*
3750 * Use only long (> 20ms) intervals to filter out excessive
3751 * spikes in service rate estimation.
3752 */
3753 if (delta_usecs > 20000) {
3754 /*
3755 * Caveat for rotational devices: processes doing I/O
3756 * in the slower disk zones tend to be slow(er) even
3757 * if not seeky. In this respect, the estimated peak
3758 * rate is likely to be an average over the disk
3759 * surface. Accordingly, to not be too harsh with
3760 * unlucky processes, a process is deemed slow only if
3761 * its rate has been lower than half of the estimated
3762 * peak rate.
3763 */
3764 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
3765 }
3766
3767 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
3768
3769 return slow;
3770}
3771
3772/*
3773 * To be deemed as soft real-time, an application must meet two
3774 * requirements. First, the application must not require an average
3775 * bandwidth higher than the approximate bandwidth required to playback or
3776 * record a compressed high-definition video.
3777 * The next function is invoked on the completion of the last request of a
3778 * batch, to compute the next-start time instant, soft_rt_next_start, such
3779 * that, if the next request of the application does not arrive before
3780 * soft_rt_next_start, then the above requirement on the bandwidth is met.
3781 *
3782 * The second requirement is that the request pattern of the application is
3783 * isochronous, i.e., that, after issuing a request or a batch of requests,
3784 * the application stops issuing new requests until all its pending requests
3785 * have been completed. After that, the application may issue a new batch,
3786 * and so on.
3787 * For this reason the next function is invoked to compute
3788 * soft_rt_next_start only for applications that meet this requirement,
3789 * whereas soft_rt_next_start is set to infinity for applications that do
3790 * not.
3791 *
3792 * Unfortunately, even a greedy (i.e., I/O-bound) application may
3793 * happen to meet, occasionally or systematically, both the above
3794 * bandwidth and isochrony requirements. This may happen at least in
3795 * the following circumstances. First, if the CPU load is high. The
3796 * application may stop issuing requests while the CPUs are busy
3797 * serving other processes, then restart, then stop again for a while,
3798 * and so on. The other circumstances are related to the storage
3799 * device: the storage device is highly loaded or reaches a low-enough
3800 * throughput with the I/O of the application (e.g., because the I/O
3801 * is random and/or the device is slow). In all these cases, the
3802 * I/O of the application may be simply slowed down enough to meet
3803 * the bandwidth and isochrony requirements. To reduce the probability
3804 * that greedy applications are deemed as soft real-time in these
3805 * corner cases, a further rule is used in the computation of
3806 * soft_rt_next_start: the return value of this function is forced to
3807 * be higher than the maximum between the following two quantities.
3808 *
3809 * (a) Current time plus: (1) the maximum time for which the arrival
3810 * of a request is waited for when a sync queue becomes idle,
3811 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
3812 * postpone for a moment the reason for adding a few extra
3813 * jiffies; we get back to it after next item (b). Lower-bounding
3814 * the return value of this function with the current time plus
3815 * bfqd->bfq_slice_idle tends to filter out greedy applications,
3816 * because the latter issue their next request as soon as possible
3817 * after the last one has been completed. In contrast, a soft
3818 * real-time application spends some time processing data, after a
3819 * batch of its requests has been completed.
3820 *
3821 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
3822 * above, greedy applications may happen to meet both the
3823 * bandwidth and isochrony requirements under heavy CPU or
3824 * storage-device load. In more detail, in these scenarios, these
3825 * applications happen, only for limited time periods, to do I/O
3826 * slowly enough to meet all the requirements described so far,
3827 * including the filtering in above item (a). These slow-speed
3828 * time intervals are usually interspersed between other time
3829 * intervals during which these applications do I/O at a very high
3830 * speed. Fortunately, exactly because of the high speed of the
3831 * I/O in the high-speed intervals, the values returned by this
3832 * function happen to be so high, near the end of any such
3833 * high-speed interval, to be likely to fall *after* the end of
3834 * the low-speed time interval that follows. These high values are
3835 * stored in bfqq->soft_rt_next_start after each invocation of
3836 * this function. As a consequence, if the last value of
3837 * bfqq->soft_rt_next_start is constantly used to lower-bound the
3838 * next value that this function may return, then, from the very
3839 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
3840 * likely to be constantly kept so high that any I/O request
3841 * issued during the low-speed interval is considered as arriving
3842 * to soon for the application to be deemed as soft
3843 * real-time. Then, in the high-speed interval that follows, the
3844 * application will not be deemed as soft real-time, just because
3845 * it will do I/O at a high speed. And so on.
3846 *
3847 * Getting back to the filtering in item (a), in the following two
3848 * cases this filtering might be easily passed by a greedy
3849 * application, if the reference quantity was just
3850 * bfqd->bfq_slice_idle:
3851 * 1) HZ is so low that the duration of a jiffy is comparable to or
3852 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
3853 * devices with HZ=100. The time granularity may be so coarse
3854 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
3855 * is rather lower than the exact value.
3856 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
3857 * for a while, then suddenly 'jump' by several units to recover the lost
3858 * increments. This seems to happen, e.g., inside virtual machines.
3859 * To address this issue, in the filtering in (a) we do not use as a
3860 * reference time interval just bfqd->bfq_slice_idle, but
3861 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
3862 * minimum number of jiffies for which the filter seems to be quite
3863 * precise also in embedded systems and KVM/QEMU virtual machines.
3864 */
3865static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
3866 struct bfq_queue *bfqq)
3867{
3868 return max3(bfqq->soft_rt_next_start,
3869 bfqq->last_idle_bklogged +
3870 HZ * bfqq->service_from_backlogged /
3871 bfqd->bfq_wr_max_softrt_rate,
3872 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
3873}
3874
3875/**
3876 * bfq_bfqq_expire - expire a queue.
3877 * @bfqd: device owning the queue.
3878 * @bfqq: the queue to expire.
3879 * @compensate: if true, compensate for the time spent idling.
3880 * @reason: the reason causing the expiration.
3881 *
3882 * If the process associated with bfqq does slow I/O (e.g., because it
3883 * issues random requests), we charge bfqq with the time it has been
3884 * in service instead of the service it has received (see
3885 * bfq_bfqq_charge_time for details on how this goal is achieved). As
3886 * a consequence, bfqq will typically get higher timestamps upon
3887 * reactivation, and hence it will be rescheduled as if it had
3888 * received more service than what it has actually received. In the
3889 * end, bfqq receives less service in proportion to how slowly its
3890 * associated process consumes its budgets (and hence how seriously it
3891 * tends to lower the throughput). In addition, this time-charging
3892 * strategy guarantees time fairness among slow processes. In
3893 * contrast, if the process associated with bfqq is not slow, we
3894 * charge bfqq exactly with the service it has received.
3895 *
3896 * Charging time to the first type of queues and the exact service to
3897 * the other has the effect of using the WF2Q+ policy to schedule the
3898 * former on a timeslice basis, without violating service domain
3899 * guarantees among the latter.
3900 */
3901void bfq_bfqq_expire(struct bfq_data *bfqd,
3902 struct bfq_queue *bfqq,
3903 bool compensate,
3904 enum bfqq_expiration reason)
3905{
3906 bool slow;
3907 unsigned long delta = 0;
3908 struct bfq_entity *entity = &bfqq->entity;
3909
3910 /*
3911 * Check whether the process is slow (see bfq_bfqq_is_slow).
3912 */
3913 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
3914
3915 /*
3916 * As above explained, charge slow (typically seeky) and
3917 * timed-out queues with the time and not the service
3918 * received, to favor sequential workloads.
3919 *
3920 * Processes doing I/O in the slower disk zones will tend to
3921 * be slow(er) even if not seeky. Therefore, since the
3922 * estimated peak rate is actually an average over the disk
3923 * surface, these processes may timeout just for bad luck. To
3924 * avoid punishing them, do not charge time to processes that
3925 * succeeded in consuming at least 2/3 of their budget. This
3926 * allows BFQ to preserve enough elasticity to still perform
3927 * bandwidth, and not time, distribution with little unlucky
3928 * or quasi-sequential processes.
3929 */
3930 if (bfqq->wr_coeff == 1 &&
3931 (slow ||
3932 (reason == BFQQE_BUDGET_TIMEOUT &&
3933 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
3934 bfq_bfqq_charge_time(bfqd, bfqq, delta);
3935
3936 if (reason == BFQQE_TOO_IDLE &&
3937 entity->service <= 2 * entity->budget / 10)
3938 bfq_clear_bfqq_IO_bound(bfqq);
3939
3940 if (bfqd->low_latency && bfqq->wr_coeff == 1)
3941 bfqq->last_wr_start_finish = jiffies;
3942
3943 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
3944 RB_EMPTY_ROOT(&bfqq->sort_list)) {
3945 /*
3946 * If we get here, and there are no outstanding
3947 * requests, then the request pattern is isochronous
3948 * (see the comments on the function
3949 * bfq_bfqq_softrt_next_start()). Thus we can compute
3950 * soft_rt_next_start. And we do it, unless bfqq is in
3951 * interactive weight raising. We do not do it in the
3952 * latter subcase, for the following reason. bfqq may
3953 * be conveying the I/O needed to load a soft
3954 * real-time application. Such an application will
3955 * actually exhibit a soft real-time I/O pattern after
3956 * it finally starts doing its job. But, if
3957 * soft_rt_next_start is computed here for an
3958 * interactive bfqq, and bfqq had received a lot of
3959 * service before remaining with no outstanding
3960 * request (likely to happen on a fast device), then
3961 * soft_rt_next_start would be assigned such a high
3962 * value that, for a very long time, bfqq would be
3963 * prevented from being possibly considered as soft
3964 * real time.
3965 *
3966 * If, instead, the queue still has outstanding
3967 * requests, then we have to wait for the completion
3968 * of all the outstanding requests to discover whether
3969 * the request pattern is actually isochronous.
3970 */
3971 if (bfqq->dispatched == 0 &&
3972 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
3973 bfqq->soft_rt_next_start =
3974 bfq_bfqq_softrt_next_start(bfqd, bfqq);
3975 else if (bfqq->dispatched > 0) {
3976 /*
3977 * Schedule an update of soft_rt_next_start to when
3978 * the task may be discovered to be isochronous.
3979 */
3980 bfq_mark_bfqq_softrt_update(bfqq);
3981 }
3982 }
3983
3984 bfq_log_bfqq(bfqd, bfqq,
3985 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
3986 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
3987
3988 /*
3989 * bfqq expired, so no total service time needs to be computed
3990 * any longer: reset state machine for measuring total service
3991 * times.
3992 */
3993 bfqd->rqs_injected = bfqd->wait_dispatch = false;
3994 bfqd->waited_rq = NULL;
3995
3996 /*
3997 * Increase, decrease or leave budget unchanged according to
3998 * reason.
3999 */
4000 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
4001 if (__bfq_bfqq_expire(bfqd, bfqq, reason))
4002 /* bfqq is gone, no more actions on it */
4003 return;
4004
4005 /* mark bfqq as waiting a request only if a bic still points to it */
4006 if (!bfq_bfqq_busy(bfqq) &&
4007 reason != BFQQE_BUDGET_TIMEOUT &&
4008 reason != BFQQE_BUDGET_EXHAUSTED) {
4009 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
4010 /*
4011 * Not setting service to 0, because, if the next rq
4012 * arrives in time, the queue will go on receiving
4013 * service with this same budget (as if it never expired)
4014 */
4015 } else
4016 entity->service = 0;
4017
4018 /*
4019 * Reset the received-service counter for every parent entity.
4020 * Differently from what happens with bfqq->entity.service,
4021 * the resetting of this counter never needs to be postponed
4022 * for parent entities. In fact, in case bfqq may have a
4023 * chance to go on being served using the last, partially
4024 * consumed budget, bfqq->entity.service needs to be kept,
4025 * because if bfqq then actually goes on being served using
4026 * the same budget, the last value of bfqq->entity.service is
4027 * needed to properly decrement bfqq->entity.budget by the
4028 * portion already consumed. In contrast, it is not necessary
4029 * to keep entity->service for parent entities too, because
4030 * the bubble up of the new value of bfqq->entity.budget will
4031 * make sure that the budgets of parent entities are correct,
4032 * even in case bfqq and thus parent entities go on receiving
4033 * service with the same budget.
4034 */
4035 entity = entity->parent;
4036 for_each_entity(entity)
4037 entity->service = 0;
4038}
4039
4040/*
4041 * Budget timeout is not implemented through a dedicated timer, but
4042 * just checked on request arrivals and completions, as well as on
4043 * idle timer expirations.
4044 */
4045static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
4046{
4047 return time_is_before_eq_jiffies(bfqq->budget_timeout);
4048}
4049
4050/*
4051 * If we expire a queue that is actively waiting (i.e., with the
4052 * device idled) for the arrival of a new request, then we may incur
4053 * the timestamp misalignment problem described in the body of the
4054 * function __bfq_activate_entity. Hence we return true only if this
4055 * condition does not hold, or if the queue is slow enough to deserve
4056 * only to be kicked off for preserving a high throughput.
4057 */
4058static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
4059{
4060 bfq_log_bfqq(bfqq->bfqd, bfqq,
4061 "may_budget_timeout: wait_request %d left %d timeout %d",
4062 bfq_bfqq_wait_request(bfqq),
4063 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
4064 bfq_bfqq_budget_timeout(bfqq));
4065
4066 return (!bfq_bfqq_wait_request(bfqq) ||
4067 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
4068 &&
4069 bfq_bfqq_budget_timeout(bfqq);
4070}
4071
4072static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
4073 struct bfq_queue *bfqq)
4074{
4075 bool rot_without_queueing =
4076 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
4077 bfqq_sequential_and_IO_bound,
4078 idling_boosts_thr;
4079
4080 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
4081 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
4082
4083 /*
4084 * The next variable takes into account the cases where idling
4085 * boosts the throughput.
4086 *
4087 * The value of the variable is computed considering, first, that
4088 * idling is virtually always beneficial for the throughput if:
4089 * (a) the device is not NCQ-capable and rotational, or
4090 * (b) regardless of the presence of NCQ, the device is rotational and
4091 * the request pattern for bfqq is I/O-bound and sequential, or
4092 * (c) regardless of whether it is rotational, the device is
4093 * not NCQ-capable and the request pattern for bfqq is
4094 * I/O-bound and sequential.
4095 *
4096 * Secondly, and in contrast to the above item (b), idling an
4097 * NCQ-capable flash-based device would not boost the
4098 * throughput even with sequential I/O; rather it would lower
4099 * the throughput in proportion to how fast the device
4100 * is. Accordingly, the next variable is true if any of the
4101 * above conditions (a), (b) or (c) is true, and, in
4102 * particular, happens to be false if bfqd is an NCQ-capable
4103 * flash-based device.
4104 */
4105 idling_boosts_thr = rot_without_queueing ||
4106 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
4107 bfqq_sequential_and_IO_bound);
4108
4109 /*
4110 * The return value of this function is equal to that of
4111 * idling_boosts_thr, unless a special case holds. In this
4112 * special case, described below, idling may cause problems to
4113 * weight-raised queues.
4114 *
4115 * When the request pool is saturated (e.g., in the presence
4116 * of write hogs), if the processes associated with
4117 * non-weight-raised queues ask for requests at a lower rate,
4118 * then processes associated with weight-raised queues have a
4119 * higher probability to get a request from the pool
4120 * immediately (or at least soon) when they need one. Thus
4121 * they have a higher probability to actually get a fraction
4122 * of the device throughput proportional to their high
4123 * weight. This is especially true with NCQ-capable drives,
4124 * which enqueue several requests in advance, and further
4125 * reorder internally-queued requests.
4126 *
4127 * For this reason, we force to false the return value if
4128 * there are weight-raised busy queues. In this case, and if
4129 * bfqq is not weight-raised, this guarantees that the device
4130 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4131 * then idling will be guaranteed by another variable, see
4132 * below). Combined with the timestamping rules of BFQ (see
4133 * [1] for details), this behavior causes bfqq, and hence any
4134 * sync non-weight-raised queue, to get a lower number of
4135 * requests served, and thus to ask for a lower number of
4136 * requests from the request pool, before the busy
4137 * weight-raised queues get served again. This often mitigates
4138 * starvation problems in the presence of heavy write
4139 * workloads and NCQ, thereby guaranteeing a higher
4140 * application and system responsiveness in these hostile
4141 * scenarios.
4142 */
4143 return idling_boosts_thr &&
4144 bfqd->wr_busy_queues == 0;
4145}
4146
4147/*
4148 * For a queue that becomes empty, device idling is allowed only if
4149 * this function returns true for that queue. As a consequence, since
4150 * device idling plays a critical role for both throughput boosting
4151 * and service guarantees, the return value of this function plays a
4152 * critical role as well.
4153 *
4154 * In a nutshell, this function returns true only if idling is
4155 * beneficial for throughput or, even if detrimental for throughput,
4156 * idling is however necessary to preserve service guarantees (low
4157 * latency, desired throughput distribution, ...). In particular, on
4158 * NCQ-capable devices, this function tries to return false, so as to
4159 * help keep the drives' internal queues full, whenever this helps the
4160 * device boost the throughput without causing any service-guarantee
4161 * issue.
4162 *
4163 * Most of the issues taken into account to get the return value of
4164 * this function are not trivial. We discuss these issues in the two
4165 * functions providing the main pieces of information needed by this
4166 * function.
4167 */
4168static bool bfq_better_to_idle(struct bfq_queue *bfqq)
4169{
4170 struct bfq_data *bfqd = bfqq->bfqd;
4171 bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
4172
4173 if (unlikely(bfqd->strict_guarantees))
4174 return true;
4175
4176 /*
4177 * Idling is performed only if slice_idle > 0. In addition, we
4178 * do not idle if
4179 * (a) bfqq is async
4180 * (b) bfqq is in the idle io prio class: in this case we do
4181 * not idle because we want to minimize the bandwidth that
4182 * queues in this class can steal to higher-priority queues
4183 */
4184 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
4185 bfq_class_idle(bfqq))
4186 return false;
4187
4188 idling_boosts_thr_with_no_issue =
4189 idling_boosts_thr_without_issues(bfqd, bfqq);
4190
4191 idling_needed_for_service_guar =
4192 idling_needed_for_service_guarantees(bfqd, bfqq);
4193
4194 /*
4195 * We have now the two components we need to compute the
4196 * return value of the function, which is true only if idling
4197 * either boosts the throughput (without issues), or is
4198 * necessary to preserve service guarantees.
4199 */
4200 return idling_boosts_thr_with_no_issue ||
4201 idling_needed_for_service_guar;
4202}
4203
4204/*
4205 * If the in-service queue is empty but the function bfq_better_to_idle
4206 * returns true, then:
4207 * 1) the queue must remain in service and cannot be expired, and
4208 * 2) the device must be idled to wait for the possible arrival of a new
4209 * request for the queue.
4210 * See the comments on the function bfq_better_to_idle for the reasons
4211 * why performing device idling is the best choice to boost the throughput
4212 * and preserve service guarantees when bfq_better_to_idle itself
4213 * returns true.
4214 */
4215static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
4216{
4217 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
4218}
4219
4220/*
4221 * This function chooses the queue from which to pick the next extra
4222 * I/O request to inject, if it finds a compatible queue. See the
4223 * comments on bfq_update_inject_limit() for details on the injection
4224 * mechanism, and for the definitions of the quantities mentioned
4225 * below.
4226 */
4227static struct bfq_queue *
4228bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
4229{
4230 struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
4231 unsigned int limit = in_serv_bfqq->inject_limit;
4232 /*
4233 * If
4234 * - bfqq is not weight-raised and therefore does not carry
4235 * time-critical I/O,
4236 * or
4237 * - regardless of whether bfqq is weight-raised, bfqq has
4238 * however a long think time, during which it can absorb the
4239 * effect of an appropriate number of extra I/O requests
4240 * from other queues (see bfq_update_inject_limit for
4241 * details on the computation of this number);
4242 * then injection can be performed without restrictions.
4243 */
4244 bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
4245 !bfq_bfqq_has_short_ttime(in_serv_bfqq);
4246
4247 /*
4248 * If
4249 * - the baseline total service time could not be sampled yet,
4250 * so the inject limit happens to be still 0, and
4251 * - a lot of time has elapsed since the plugging of I/O
4252 * dispatching started, so drive speed is being wasted
4253 * significantly;
4254 * then temporarily raise inject limit to one request.
4255 */
4256 if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
4257 bfq_bfqq_wait_request(in_serv_bfqq) &&
4258 time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
4259 bfqd->bfq_slice_idle)
4260 )
4261 limit = 1;
4262
4263 if (bfqd->rq_in_driver >= limit)
4264 return NULL;
4265
4266 /*
4267 * Linear search of the source queue for injection; but, with
4268 * a high probability, very few steps are needed to find a
4269 * candidate queue, i.e., a queue with enough budget left for
4270 * its next request. In fact:
4271 * - BFQ dynamically updates the budget of every queue so as
4272 * to accommodate the expected backlog of the queue;
4273 * - if a queue gets all its requests dispatched as injected
4274 * service, then the queue is removed from the active list
4275 * (and re-added only if it gets new requests, but then it
4276 * is assigned again enough budget for its new backlog).
4277 */
4278 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
4279 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4280 (in_serv_always_inject || bfqq->wr_coeff > 1) &&
4281 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4282 bfq_bfqq_budget_left(bfqq)) {
4283 /*
4284 * Allow for only one large in-flight request
4285 * on non-rotational devices, for the
4286 * following reason. On non-rotationl drives,
4287 * large requests take much longer than
4288 * smaller requests to be served. In addition,
4289 * the drive prefers to serve large requests
4290 * w.r.t. to small ones, if it can choose. So,
4291 * having more than one large requests queued
4292 * in the drive may easily make the next first
4293 * request of the in-service queue wait for so
4294 * long to break bfqq's service guarantees. On
4295 * the bright side, large requests let the
4296 * drive reach a very high throughput, even if
4297 * there is only one in-flight large request
4298 * at a time.
4299 */
4300 if (blk_queue_nonrot(bfqd->queue) &&
4301 blk_rq_sectors(bfqq->next_rq) >=
4302 BFQQ_SECT_THR_NONROT)
4303 limit = min_t(unsigned int, 1, limit);
4304 else
4305 limit = in_serv_bfqq->inject_limit;
4306
4307 if (bfqd->rq_in_driver < limit) {
4308 bfqd->rqs_injected = true;
4309 return bfqq;
4310 }
4311 }
4312
4313 return NULL;
4314}
4315
4316/*
4317 * Select a queue for service. If we have a current queue in service,
4318 * check whether to continue servicing it, or retrieve and set a new one.
4319 */
4320static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
4321{
4322 struct bfq_queue *bfqq;
4323 struct request *next_rq;
4324 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
4325
4326 bfqq = bfqd->in_service_queue;
4327 if (!bfqq)
4328 goto new_queue;
4329
4330 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
4331
4332 /*
4333 * Do not expire bfqq for budget timeout if bfqq may be about
4334 * to enjoy device idling. The reason why, in this case, we
4335 * prevent bfqq from expiring is the same as in the comments
4336 * on the case where bfq_bfqq_must_idle() returns true, in
4337 * bfq_completed_request().
4338 */
4339 if (bfq_may_expire_for_budg_timeout(bfqq) &&
4340 !bfq_bfqq_must_idle(bfqq))
4341 goto expire;
4342
4343check_queue:
4344 /*
4345 * This loop is rarely executed more than once. Even when it
4346 * happens, it is much more convenient to re-execute this loop
4347 * than to return NULL and trigger a new dispatch to get a
4348 * request served.
4349 */
4350 next_rq = bfqq->next_rq;
4351 /*
4352 * If bfqq has requests queued and it has enough budget left to
4353 * serve them, keep the queue, otherwise expire it.
4354 */
4355 if (next_rq) {
4356 if (bfq_serv_to_charge(next_rq, bfqq) >
4357 bfq_bfqq_budget_left(bfqq)) {
4358 /*
4359 * Expire the queue for budget exhaustion,
4360 * which makes sure that the next budget is
4361 * enough to serve the next request, even if
4362 * it comes from the fifo expired path.
4363 */
4364 reason = BFQQE_BUDGET_EXHAUSTED;
4365 goto expire;
4366 } else {
4367 /*
4368 * The idle timer may be pending because we may
4369 * not disable disk idling even when a new request
4370 * arrives.
4371 */
4372 if (bfq_bfqq_wait_request(bfqq)) {
4373 /*
4374 * If we get here: 1) at least a new request
4375 * has arrived but we have not disabled the
4376 * timer because the request was too small,
4377 * 2) then the block layer has unplugged
4378 * the device, causing the dispatch to be
4379 * invoked.
4380 *
4381 * Since the device is unplugged, now the
4382 * requests are probably large enough to
4383 * provide a reasonable throughput.
4384 * So we disable idling.
4385 */
4386 bfq_clear_bfqq_wait_request(bfqq);
4387 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4388 }
4389 goto keep_queue;
4390 }
4391 }
4392
4393 /*
4394 * No requests pending. However, if the in-service queue is idling
4395 * for a new request, or has requests waiting for a completion and
4396 * may idle after their completion, then keep it anyway.
4397 *
4398 * Yet, inject service from other queues if it boosts
4399 * throughput and is possible.
4400 */
4401 if (bfq_bfqq_wait_request(bfqq) ||
4402 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
4403 struct bfq_queue *async_bfqq =
4404 bfqq->bic && bfqq->bic->bfqq[0] &&
4405 bfq_bfqq_busy(bfqq->bic->bfqq[0]) &&
4406 bfqq->bic->bfqq[0]->next_rq ?
4407 bfqq->bic->bfqq[0] : NULL;
4408
4409 /*
4410 * The next three mutually-exclusive ifs decide
4411 * whether to try injection, and choose the queue to
4412 * pick an I/O request from.
4413 *
4414 * The first if checks whether the process associated
4415 * with bfqq has also async I/O pending. If so, it
4416 * injects such I/O unconditionally. Injecting async
4417 * I/O from the same process can cause no harm to the
4418 * process. On the contrary, it can only increase
4419 * bandwidth and reduce latency for the process.
4420 *
4421 * The second if checks whether there happens to be a
4422 * non-empty waker queue for bfqq, i.e., a queue whose
4423 * I/O needs to be completed for bfqq to receive new
4424 * I/O. This happens, e.g., if bfqq is associated with
4425 * a process that does some sync. A sync generates
4426 * extra blocking I/O, which must be completed before
4427 * the process associated with bfqq can go on with its
4428 * I/O. If the I/O of the waker queue is not served,
4429 * then bfqq remains empty, and no I/O is dispatched,
4430 * until the idle timeout fires for bfqq. This is
4431 * likely to result in lower bandwidth and higher
4432 * latencies for bfqq, and in a severe loss of total
4433 * throughput. The best action to take is therefore to
4434 * serve the waker queue as soon as possible. So do it
4435 * (without relying on the third alternative below for
4436 * eventually serving waker_bfqq's I/O; see the last
4437 * paragraph for further details). This systematic
4438 * injection of I/O from the waker queue does not
4439 * cause any delay to bfqq's I/O. On the contrary,
4440 * next bfqq's I/O is brought forward dramatically,
4441 * for it is not blocked for milliseconds.
4442 *
4443 * The third if checks whether bfqq is a queue for
4444 * which it is better to avoid injection. It is so if
4445 * bfqq delivers more throughput when served without
4446 * any further I/O from other queues in the middle, or
4447 * if the service times of bfqq's I/O requests both
4448 * count more than overall throughput, and may be
4449 * easily increased by injection (this happens if bfqq
4450 * has a short think time). If none of these
4451 * conditions holds, then a candidate queue for
4452 * injection is looked for through
4453 * bfq_choose_bfqq_for_injection(). Note that the
4454 * latter may return NULL (for example if the inject
4455 * limit for bfqq is currently 0).
4456 *
4457 * NOTE: motivation for the second alternative
4458 *
4459 * Thanks to the way the inject limit is updated in
4460 * bfq_update_has_short_ttime(), it is rather likely
4461 * that, if I/O is being plugged for bfqq and the
4462 * waker queue has pending I/O requests that are
4463 * blocking bfqq's I/O, then the third alternative
4464 * above lets the waker queue get served before the
4465 * I/O-plugging timeout fires. So one may deem the
4466 * second alternative superfluous. It is not, because
4467 * the third alternative may be way less effective in
4468 * case of a synchronization. For two main
4469 * reasons. First, throughput may be low because the
4470 * inject limit may be too low to guarantee the same
4471 * amount of injected I/O, from the waker queue or
4472 * other queues, that the second alternative
4473 * guarantees (the second alternative unconditionally
4474 * injects a pending I/O request of the waker queue
4475 * for each bfq_dispatch_request()). Second, with the
4476 * third alternative, the duration of the plugging,
4477 * i.e., the time before bfqq finally receives new I/O,
4478 * may not be minimized, because the waker queue may
4479 * happen to be served only after other queues.
4480 */
4481 if (async_bfqq &&
4482 icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
4483 bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
4484 bfq_bfqq_budget_left(async_bfqq))
4485 bfqq = bfqq->bic->bfqq[0];
4486 else if (bfq_bfqq_has_waker(bfqq) &&
4487 bfq_bfqq_busy(bfqq->waker_bfqq) &&
4488 bfqq->next_rq &&
4489 bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
4490 bfqq->waker_bfqq) <=
4491 bfq_bfqq_budget_left(bfqq->waker_bfqq)
4492 )
4493 bfqq = bfqq->waker_bfqq;
4494 else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
4495 (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
4496 !bfq_bfqq_has_short_ttime(bfqq)))
4497 bfqq = bfq_choose_bfqq_for_injection(bfqd);
4498 else
4499 bfqq = NULL;
4500
4501 goto keep_queue;
4502 }
4503
4504 reason = BFQQE_NO_MORE_REQUESTS;
4505expire:
4506 bfq_bfqq_expire(bfqd, bfqq, false, reason);
4507new_queue:
4508 bfqq = bfq_set_in_service_queue(bfqd);
4509 if (bfqq) {
4510 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
4511 goto check_queue;
4512 }
4513keep_queue:
4514 if (bfqq)
4515 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
4516 else
4517 bfq_log(bfqd, "select_queue: no queue returned");
4518
4519 return bfqq;
4520}
4521
4522static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4523{
4524 struct bfq_entity *entity = &bfqq->entity;
4525
4526 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
4527 bfq_log_bfqq(bfqd, bfqq,
4528 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
4529 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
4530 jiffies_to_msecs(bfqq->wr_cur_max_time),
4531 bfqq->wr_coeff,
4532 bfqq->entity.weight, bfqq->entity.orig_weight);
4533
4534 if (entity->prio_changed)
4535 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
4536
4537 /*
4538 * If the queue was activated in a burst, or too much
4539 * time has elapsed from the beginning of this
4540 * weight-raising period, then end weight raising.
4541 */
4542 if (bfq_bfqq_in_large_burst(bfqq))
4543 bfq_bfqq_end_wr(bfqq);
4544 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
4545 bfqq->wr_cur_max_time)) {
4546 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
4547 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
4548 bfq_wr_duration(bfqd)))
4549 bfq_bfqq_end_wr(bfqq);
4550 else {
4551 switch_back_to_interactive_wr(bfqq, bfqd);
4552 bfqq->entity.prio_changed = 1;
4553 }
4554 }
4555 if (bfqq->wr_coeff > 1 &&
4556 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
4557 bfqq->service_from_wr > max_service_from_wr) {
4558 /* see comments on max_service_from_wr */
4559 bfq_bfqq_end_wr(bfqq);
4560 }
4561 }
4562 /*
4563 * To improve latency (for this or other queues), immediately
4564 * update weight both if it must be raised and if it must be
4565 * lowered. Since, entity may be on some active tree here, and
4566 * might have a pending change of its ioprio class, invoke
4567 * next function with the last parameter unset (see the
4568 * comments on the function).
4569 */
4570 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
4571 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
4572 entity, false);
4573}
4574
4575/*
4576 * Dispatch next request from bfqq.
4577 */
4578static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
4579 struct bfq_queue *bfqq)
4580{
4581 struct request *rq = bfqq->next_rq;
4582 unsigned long service_to_charge;
4583
4584 service_to_charge = bfq_serv_to_charge(rq, bfqq);
4585
4586 bfq_bfqq_served(bfqq, service_to_charge);
4587
4588 if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
4589 bfqd->wait_dispatch = false;
4590 bfqd->waited_rq = rq;
4591 }
4592
4593 bfq_dispatch_remove(bfqd->queue, rq);
4594
4595 if (bfqq != bfqd->in_service_queue)
4596 goto return_rq;
4597
4598 /*
4599 * If weight raising has to terminate for bfqq, then next
4600 * function causes an immediate update of bfqq's weight,
4601 * without waiting for next activation. As a consequence, on
4602 * expiration, bfqq will be timestamped as if has never been
4603 * weight-raised during this service slot, even if it has
4604 * received part or even most of the service as a
4605 * weight-raised queue. This inflates bfqq's timestamps, which
4606 * is beneficial, as bfqq is then more willing to leave the
4607 * device immediately to possible other weight-raised queues.
4608 */
4609 bfq_update_wr_data(bfqd, bfqq);
4610
4611 /*
4612 * Expire bfqq, pretending that its budget expired, if bfqq
4613 * belongs to CLASS_IDLE and other queues are waiting for
4614 * service.
4615 */
4616 if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
4617 goto return_rq;
4618
4619 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
4620
4621return_rq:
4622 return rq;
4623}
4624
4625static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
4626{
4627 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4628
4629 /*
4630 * Avoiding lock: a race on bfqd->busy_queues should cause at
4631 * most a call to dispatch for nothing
4632 */
4633 return !list_empty_careful(&bfqd->dispatch) ||
4634 bfq_tot_busy_queues(bfqd) > 0;
4635}
4636
4637static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
4638{
4639 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4640 struct request *rq = NULL;
4641 struct bfq_queue *bfqq = NULL;
4642
4643 if (!list_empty(&bfqd->dispatch)) {
4644 rq = list_first_entry(&bfqd->dispatch, struct request,
4645 queuelist);
4646 list_del_init(&rq->queuelist);
4647
4648 bfqq = RQ_BFQQ(rq);
4649
4650 if (bfqq) {
4651 /*
4652 * Increment counters here, because this
4653 * dispatch does not follow the standard
4654 * dispatch flow (where counters are
4655 * incremented)
4656 */
4657 bfqq->dispatched++;
4658
4659 goto inc_in_driver_start_rq;
4660 }
4661
4662 /*
4663 * We exploit the bfq_finish_requeue_request hook to
4664 * decrement rq_in_driver, but
4665 * bfq_finish_requeue_request will not be invoked on
4666 * this request. So, to avoid unbalance, just start
4667 * this request, without incrementing rq_in_driver. As
4668 * a negative consequence, rq_in_driver is deceptively
4669 * lower than it should be while this request is in
4670 * service. This may cause bfq_schedule_dispatch to be
4671 * invoked uselessly.
4672 *
4673 * As for implementing an exact solution, the
4674 * bfq_finish_requeue_request hook, if defined, is
4675 * probably invoked also on this request. So, by
4676 * exploiting this hook, we could 1) increment
4677 * rq_in_driver here, and 2) decrement it in
4678 * bfq_finish_requeue_request. Such a solution would
4679 * let the value of the counter be always accurate,
4680 * but it would entail using an extra interface
4681 * function. This cost seems higher than the benefit,
4682 * being the frequency of non-elevator-private
4683 * requests very low.
4684 */
4685 goto start_rq;
4686 }
4687
4688 bfq_log(bfqd, "dispatch requests: %d busy queues",
4689 bfq_tot_busy_queues(bfqd));
4690
4691 if (bfq_tot_busy_queues(bfqd) == 0)
4692 goto exit;
4693
4694 /*
4695 * Force device to serve one request at a time if
4696 * strict_guarantees is true. Forcing this service scheme is
4697 * currently the ONLY way to guarantee that the request
4698 * service order enforced by the scheduler is respected by a
4699 * queueing device. Otherwise the device is free even to make
4700 * some unlucky request wait for as long as the device
4701 * wishes.
4702 *
4703 * Of course, serving one request at at time may cause loss of
4704 * throughput.
4705 */
4706 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
4707 goto exit;
4708
4709 bfqq = bfq_select_queue(bfqd);
4710 if (!bfqq)
4711 goto exit;
4712
4713 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
4714
4715 if (rq) {
4716inc_in_driver_start_rq:
4717 bfqd->rq_in_driver++;
4718start_rq:
4719 rq->rq_flags |= RQF_STARTED;
4720 }
4721exit:
4722 return rq;
4723}
4724
4725#ifdef CONFIG_BFQ_CGROUP_DEBUG
4726static void bfq_update_dispatch_stats(struct request_queue *q,
4727 struct request *rq,
4728 struct bfq_queue *in_serv_queue,
4729 bool idle_timer_disabled)
4730{
4731 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
4732
4733 if (!idle_timer_disabled && !bfqq)
4734 return;
4735
4736 /*
4737 * rq and bfqq are guaranteed to exist until this function
4738 * ends, for the following reasons. First, rq can be
4739 * dispatched to the device, and then can be completed and
4740 * freed, only after this function ends. Second, rq cannot be
4741 * merged (and thus freed because of a merge) any longer,
4742 * because it has already started. Thus rq cannot be freed
4743 * before this function ends, and, since rq has a reference to
4744 * bfqq, the same guarantee holds for bfqq too.
4745 *
4746 * In addition, the following queue lock guarantees that
4747 * bfqq_group(bfqq) exists as well.
4748 */
4749 spin_lock_irq(&q->queue_lock);
4750 if (idle_timer_disabled)
4751 /*
4752 * Since the idle timer has been disabled,
4753 * in_serv_queue contained some request when
4754 * __bfq_dispatch_request was invoked above, which
4755 * implies that rq was picked exactly from
4756 * in_serv_queue. Thus in_serv_queue == bfqq, and is
4757 * therefore guaranteed to exist because of the above
4758 * arguments.
4759 */
4760 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
4761 if (bfqq) {
4762 struct bfq_group *bfqg = bfqq_group(bfqq);
4763
4764 bfqg_stats_update_avg_queue_size(bfqg);
4765 bfqg_stats_set_start_empty_time(bfqg);
4766 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
4767 }
4768 spin_unlock_irq(&q->queue_lock);
4769}
4770#else
4771static inline void bfq_update_dispatch_stats(struct request_queue *q,
4772 struct request *rq,
4773 struct bfq_queue *in_serv_queue,
4774 bool idle_timer_disabled) {}
4775#endif /* CONFIG_BFQ_CGROUP_DEBUG */
4776
4777static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
4778{
4779 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4780 struct request *rq;
4781 struct bfq_queue *in_serv_queue;
4782 bool waiting_rq, idle_timer_disabled;
4783
4784 spin_lock_irq(&bfqd->lock);
4785
4786 in_serv_queue = bfqd->in_service_queue;
4787 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
4788
4789 rq = __bfq_dispatch_request(hctx);
4790
4791 idle_timer_disabled =
4792 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
4793
4794 spin_unlock_irq(&bfqd->lock);
4795
4796 bfq_update_dispatch_stats(hctx->queue, rq, in_serv_queue,
4797 idle_timer_disabled);
4798
4799 return rq;
4800}
4801
4802/*
4803 * Task holds one reference to the queue, dropped when task exits. Each rq
4804 * in-flight on this queue also holds a reference, dropped when rq is freed.
4805 *
4806 * Scheduler lock must be held here. Recall not to use bfqq after calling
4807 * this function on it.
4808 */
4809void bfq_put_queue(struct bfq_queue *bfqq)
4810{
4811 struct bfq_queue *item;
4812 struct hlist_node *n;
4813#ifdef CONFIG_BFQ_GROUP_IOSCHED
4814 struct bfq_group *bfqg = bfqq_group(bfqq);
4815#endif
4816
4817 if (bfqq->bfqd)
4818 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
4819 bfqq, bfqq->ref);
4820
4821 bfqq->ref--;
4822 if (bfqq->ref)
4823 return;
4824
4825 if (!hlist_unhashed(&bfqq->burst_list_node)) {
4826 hlist_del_init(&bfqq->burst_list_node);
4827 /*
4828 * Decrement also burst size after the removal, if the
4829 * process associated with bfqq is exiting, and thus
4830 * does not contribute to the burst any longer. This
4831 * decrement helps filter out false positives of large
4832 * bursts, when some short-lived process (often due to
4833 * the execution of commands by some service) happens
4834 * to start and exit while a complex application is
4835 * starting, and thus spawning several processes that
4836 * do I/O (and that *must not* be treated as a large
4837 * burst, see comments on bfq_handle_burst).
4838 *
4839 * In particular, the decrement is performed only if:
4840 * 1) bfqq is not a merged queue, because, if it is,
4841 * then this free of bfqq is not triggered by the exit
4842 * of the process bfqq is associated with, but exactly
4843 * by the fact that bfqq has just been merged.
4844 * 2) burst_size is greater than 0, to handle
4845 * unbalanced decrements. Unbalanced decrements may
4846 * happen in te following case: bfqq is inserted into
4847 * the current burst list--without incrementing
4848 * bust_size--because of a split, but the current
4849 * burst list is not the burst list bfqq belonged to
4850 * (see comments on the case of a split in
4851 * bfq_set_request).
4852 */
4853 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
4854 bfqq->bfqd->burst_size--;
4855 }
4856
4857 /*
4858 * bfqq does not exist any longer, so it cannot be woken by
4859 * any other queue, and cannot wake any other queue. Then bfqq
4860 * must be removed from the woken list of its possible waker
4861 * queue, and all queues in the woken list of bfqq must stop
4862 * having a waker queue. Strictly speaking, these updates
4863 * should be performed when bfqq remains with no I/O source
4864 * attached to it, which happens before bfqq gets freed. In
4865 * particular, this happens when the last process associated
4866 * with bfqq exits or gets associated with a different
4867 * queue. However, both events lead to bfqq being freed soon,
4868 * and dangling references would come out only after bfqq gets
4869 * freed. So these updates are done here, as a simple and safe
4870 * way to handle all cases.
4871 */
4872 /* remove bfqq from woken list */
4873 if (!hlist_unhashed(&bfqq->woken_list_node))
4874 hlist_del_init(&bfqq->woken_list_node);
4875
4876 /* reset waker for all queues in woken list */
4877 hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
4878 woken_list_node) {
4879 item->waker_bfqq = NULL;
4880 bfq_clear_bfqq_has_waker(item);
4881 hlist_del_init(&item->woken_list_node);
4882 }
4883
4884 if (bfqq->bfqd && bfqq->bfqd->last_completed_rq_bfqq == bfqq)
4885 bfqq->bfqd->last_completed_rq_bfqq = NULL;
4886
4887 kmem_cache_free(bfq_pool, bfqq);
4888#ifdef CONFIG_BFQ_GROUP_IOSCHED
4889 bfqg_and_blkg_put(bfqg);
4890#endif
4891}
4892
4893static void bfq_put_cooperator(struct bfq_queue *bfqq)
4894{
4895 struct bfq_queue *__bfqq, *next;
4896
4897 /*
4898 * If this queue was scheduled to merge with another queue, be
4899 * sure to drop the reference taken on that queue (and others in
4900 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
4901 */
4902 __bfqq = bfqq->new_bfqq;
4903 while (__bfqq) {
4904 if (__bfqq == bfqq)
4905 break;
4906 next = __bfqq->new_bfqq;
4907 bfq_put_queue(__bfqq);
4908 __bfqq = next;
4909 }
4910}
4911
4912static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4913{
4914 if (bfqq == bfqd->in_service_queue) {
4915 __bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT);
4916 bfq_schedule_dispatch(bfqd);
4917 }
4918
4919 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
4920
4921 bfq_put_cooperator(bfqq);
4922
4923 bfq_release_process_ref(bfqd, bfqq);
4924}
4925
4926static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
4927{
4928 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4929 struct bfq_data *bfqd;
4930
4931 if (bfqq)
4932 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
4933
4934 if (bfqq && bfqd) {
4935 unsigned long flags;
4936
4937 spin_lock_irqsave(&bfqd->lock, flags);
4938 bfqq->bic = NULL;
4939 bfq_exit_bfqq(bfqd, bfqq);
4940 bic_set_bfqq(bic, NULL, is_sync);
4941 spin_unlock_irqrestore(&bfqd->lock, flags);
4942 }
4943}
4944
4945static void bfq_exit_icq(struct io_cq *icq)
4946{
4947 struct bfq_io_cq *bic = icq_to_bic(icq);
4948
4949 bfq_exit_icq_bfqq(bic, true);
4950 bfq_exit_icq_bfqq(bic, false);
4951}
4952
4953/*
4954 * Update the entity prio values; note that the new values will not
4955 * be used until the next (re)activation.
4956 */
4957static void
4958bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
4959{
4960 struct task_struct *tsk = current;
4961 int ioprio_class;
4962 struct bfq_data *bfqd = bfqq->bfqd;
4963
4964 if (!bfqd)
4965 return;
4966
4967 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4968 switch (ioprio_class) {
4969 default:
4970 dev_err(bfqq->bfqd->queue->backing_dev_info->dev,
4971 "bfq: bad prio class %d\n", ioprio_class);
4972 /* fall through */
4973 case IOPRIO_CLASS_NONE:
4974 /*
4975 * No prio set, inherit CPU scheduling settings.
4976 */
4977 bfqq->new_ioprio = task_nice_ioprio(tsk);
4978 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
4979 break;
4980 case IOPRIO_CLASS_RT:
4981 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4982 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
4983 break;
4984 case IOPRIO_CLASS_BE:
4985 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4986 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
4987 break;
4988 case IOPRIO_CLASS_IDLE:
4989 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
4990 bfqq->new_ioprio = 7;
4991 break;
4992 }
4993
4994 if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
4995 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
4996 bfqq->new_ioprio);
4997 bfqq->new_ioprio = IOPRIO_BE_NR;
4998 }
4999
5000 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
5001 bfqq->entity.prio_changed = 1;
5002}
5003
5004static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5005 struct bio *bio, bool is_sync,
5006 struct bfq_io_cq *bic);
5007
5008static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
5009{
5010 struct bfq_data *bfqd = bic_to_bfqd(bic);
5011 struct bfq_queue *bfqq;
5012 int ioprio = bic->icq.ioc->ioprio;
5013
5014 /*
5015 * This condition may trigger on a newly created bic, be sure to
5016 * drop the lock before returning.
5017 */
5018 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
5019 return;
5020
5021 bic->ioprio = ioprio;
5022
5023 bfqq = bic_to_bfqq(bic, false);
5024 if (bfqq) {
5025 bfq_release_process_ref(bfqd, bfqq);
5026 bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
5027 bic_set_bfqq(bic, bfqq, false);
5028 }
5029
5030 bfqq = bic_to_bfqq(bic, true);
5031 if (bfqq)
5032 bfq_set_next_ioprio_data(bfqq, bic);
5033}
5034
5035static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5036 struct bfq_io_cq *bic, pid_t pid, int is_sync)
5037{
5038 RB_CLEAR_NODE(&bfqq->entity.rb_node);
5039 INIT_LIST_HEAD(&bfqq->fifo);
5040 INIT_HLIST_NODE(&bfqq->burst_list_node);
5041 INIT_HLIST_NODE(&bfqq->woken_list_node);
5042 INIT_HLIST_HEAD(&bfqq->woken_list);
5043
5044 bfqq->ref = 0;
5045 bfqq->bfqd = bfqd;
5046
5047 if (bic)
5048 bfq_set_next_ioprio_data(bfqq, bic);
5049
5050 if (is_sync) {
5051 /*
5052 * No need to mark as has_short_ttime if in
5053 * idle_class, because no device idling is performed
5054 * for queues in idle class
5055 */
5056 if (!bfq_class_idle(bfqq))
5057 /* tentatively mark as has_short_ttime */
5058 bfq_mark_bfqq_has_short_ttime(bfqq);
5059 bfq_mark_bfqq_sync(bfqq);
5060 bfq_mark_bfqq_just_created(bfqq);
5061 } else
5062 bfq_clear_bfqq_sync(bfqq);
5063
5064 /* set end request to minus infinity from now */
5065 bfqq->ttime.last_end_request = ktime_get_ns() + 1;
5066
5067 bfq_mark_bfqq_IO_bound(bfqq);
5068
5069 bfqq->pid = pid;
5070
5071 /* Tentative initial value to trade off between thr and lat */
5072 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
5073 bfqq->budget_timeout = bfq_smallest_from_now();
5074
5075 bfqq->wr_coeff = 1;
5076 bfqq->last_wr_start_finish = jiffies;
5077 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
5078 bfqq->split_time = bfq_smallest_from_now();
5079
5080 /*
5081 * To not forget the possibly high bandwidth consumed by a
5082 * process/queue in the recent past,
5083 * bfq_bfqq_softrt_next_start() returns a value at least equal
5084 * to the current value of bfqq->soft_rt_next_start (see
5085 * comments on bfq_bfqq_softrt_next_start). Set
5086 * soft_rt_next_start to now, to mean that bfqq has consumed
5087 * no bandwidth so far.
5088 */
5089 bfqq->soft_rt_next_start = jiffies;
5090
5091 /* first request is almost certainly seeky */
5092 bfqq->seek_history = 1;
5093}
5094
5095static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
5096 struct bfq_group *bfqg,
5097 int ioprio_class, int ioprio)
5098{
5099 switch (ioprio_class) {
5100 case IOPRIO_CLASS_RT:
5101 return &bfqg->async_bfqq[0][ioprio];
5102 case IOPRIO_CLASS_NONE:
5103 ioprio = IOPRIO_NORM;
5104 /* fall through */
5105 case IOPRIO_CLASS_BE:
5106 return &bfqg->async_bfqq[1][ioprio];
5107 case IOPRIO_CLASS_IDLE:
5108 return &bfqg->async_idle_bfqq;
5109 default:
5110 return NULL;
5111 }
5112}
5113
5114static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5115 struct bio *bio, bool is_sync,
5116 struct bfq_io_cq *bic)
5117{
5118 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5119 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5120 struct bfq_queue **async_bfqq = NULL;
5121 struct bfq_queue *bfqq;
5122 struct bfq_group *bfqg;
5123
5124 rcu_read_lock();
5125
5126 bfqg = bfq_find_set_group(bfqd, __bio_blkcg(bio));
5127 if (!bfqg) {
5128 bfqq = &bfqd->oom_bfqq;
5129 goto out;
5130 }
5131
5132 if (!is_sync) {
5133 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
5134 ioprio);
5135 bfqq = *async_bfqq;
5136 if (bfqq)
5137 goto out;
5138 }
5139
5140 bfqq = kmem_cache_alloc_node(bfq_pool,
5141 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
5142 bfqd->queue->node);
5143
5144 if (bfqq) {
5145 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
5146 is_sync);
5147 bfq_init_entity(&bfqq->entity, bfqg);
5148 bfq_log_bfqq(bfqd, bfqq, "allocated");
5149 } else {
5150 bfqq = &bfqd->oom_bfqq;
5151 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
5152 goto out;
5153 }
5154
5155 /*
5156 * Pin the queue now that it's allocated, scheduler exit will
5157 * prune it.
5158 */
5159 if (async_bfqq) {
5160 bfqq->ref++; /*
5161 * Extra group reference, w.r.t. sync
5162 * queue. This extra reference is removed
5163 * only if bfqq->bfqg disappears, to
5164 * guarantee that this queue is not freed
5165 * until its group goes away.
5166 */
5167 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
5168 bfqq, bfqq->ref);
5169 *async_bfqq = bfqq;
5170 }
5171
5172out:
5173 bfqq->ref++; /* get a process reference to this queue */
5174 bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
5175 rcu_read_unlock();
5176 return bfqq;
5177}
5178
5179static void bfq_update_io_thinktime(struct bfq_data *bfqd,
5180 struct bfq_queue *bfqq)
5181{
5182 struct bfq_ttime *ttime = &bfqq->ttime;
5183 u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
5184
5185 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
5186
5187 ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
5188 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
5189 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
5190 ttime->ttime_samples);
5191}
5192
5193static void
5194bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5195 struct request *rq)
5196{
5197 bfqq->seek_history <<= 1;
5198 bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
5199
5200 if (bfqq->wr_coeff > 1 &&
5201 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
5202 BFQQ_TOTALLY_SEEKY(bfqq))
5203 bfq_bfqq_end_wr(bfqq);
5204}
5205
5206static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
5207 struct bfq_queue *bfqq,
5208 struct bfq_io_cq *bic)
5209{
5210 bool has_short_ttime = true, state_changed;
5211
5212 /*
5213 * No need to update has_short_ttime if bfqq is async or in
5214 * idle io prio class, or if bfq_slice_idle is zero, because
5215 * no device idling is performed for bfqq in this case.
5216 */
5217 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
5218 bfqd->bfq_slice_idle == 0)
5219 return;
5220
5221 /* Idle window just restored, statistics are meaningless. */
5222 if (time_is_after_eq_jiffies(bfqq->split_time +
5223 bfqd->bfq_wr_min_idle_time))
5224 return;
5225
5226 /* Think time is infinite if no process is linked to
5227 * bfqq. Otherwise check average think time to
5228 * decide whether to mark as has_short_ttime
5229 */
5230 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
5231 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
5232 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle))
5233 has_short_ttime = false;
5234
5235 state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
5236
5237 if (has_short_ttime)
5238 bfq_mark_bfqq_has_short_ttime(bfqq);
5239 else
5240 bfq_clear_bfqq_has_short_ttime(bfqq);
5241
5242 /*
5243 * Until the base value for the total service time gets
5244 * finally computed for bfqq, the inject limit does depend on
5245 * the think-time state (short|long). In particular, the limit
5246 * is 0 or 1 if the think time is deemed, respectively, as
5247 * short or long (details in the comments in
5248 * bfq_update_inject_limit()). Accordingly, the next
5249 * instructions reset the inject limit if the think-time state
5250 * has changed and the above base value is still to be
5251 * computed.
5252 *
5253 * However, the reset is performed only if more than 100 ms
5254 * have elapsed since the last update of the inject limit, or
5255 * (inclusive) if the change is from short to long think
5256 * time. The reason for this waiting is as follows.
5257 *
5258 * bfqq may have a long think time because of a
5259 * synchronization with some other queue, i.e., because the
5260 * I/O of some other queue may need to be completed for bfqq
5261 * to receive new I/O. Details in the comments on the choice
5262 * of the queue for injection in bfq_select_queue().
5263 *
5264 * As stressed in those comments, if such a synchronization is
5265 * actually in place, then, without injection on bfqq, the
5266 * blocking I/O cannot happen to served while bfqq is in
5267 * service. As a consequence, if bfqq is granted
5268 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
5269 * is dispatched, until the idle timeout fires. This is likely
5270 * to result in lower bandwidth and higher latencies for bfqq,
5271 * and in a severe loss of total throughput.
5272 *
5273 * On the opposite end, a non-zero inject limit may allow the
5274 * I/O that blocks bfqq to be executed soon, and therefore
5275 * bfqq to receive new I/O soon.
5276 *
5277 * But, if the blocking gets actually eliminated, then the
5278 * next think-time sample for bfqq may be very low. This in
5279 * turn may cause bfqq's think time to be deemed
5280 * short. Without the 100 ms barrier, this new state change
5281 * would cause the body of the next if to be executed
5282 * immediately. But this would set to 0 the inject
5283 * limit. Without injection, the blocking I/O would cause the
5284 * think time of bfqq to become long again, and therefore the
5285 * inject limit to be raised again, and so on. The only effect
5286 * of such a steady oscillation between the two think-time
5287 * states would be to prevent effective injection on bfqq.
5288 *
5289 * In contrast, if the inject limit is not reset during such a
5290 * long time interval as 100 ms, then the number of short
5291 * think time samples can grow significantly before the reset
5292 * is performed. As a consequence, the think time state can
5293 * become stable before the reset. Therefore there will be no
5294 * state change when the 100 ms elapse, and no reset of the
5295 * inject limit. The inject limit remains steadily equal to 1
5296 * both during and after the 100 ms. So injection can be
5297 * performed at all times, and throughput gets boosted.
5298 *
5299 * An inject limit equal to 1 is however in conflict, in
5300 * general, with the fact that the think time of bfqq is
5301 * short, because injection may be likely to delay bfqq's I/O
5302 * (as explained in the comments in
5303 * bfq_update_inject_limit()). But this does not happen in
5304 * this special case, because bfqq's low think time is due to
5305 * an effective handling of a synchronization, through
5306 * injection. In this special case, bfqq's I/O does not get
5307 * delayed by injection; on the contrary, bfqq's I/O is
5308 * brought forward, because it is not blocked for
5309 * milliseconds.
5310 *
5311 * In addition, serving the blocking I/O much sooner, and much
5312 * more frequently than once per I/O-plugging timeout, makes
5313 * it much quicker to detect a waker queue (the concept of
5314 * waker queue is defined in the comments in
5315 * bfq_add_request()). This makes it possible to start sooner
5316 * to boost throughput more effectively, by injecting the I/O
5317 * of the waker queue unconditionally on every
5318 * bfq_dispatch_request().
5319 *
5320 * One last, important benefit of not resetting the inject
5321 * limit before 100 ms is that, during this time interval, the
5322 * base value for the total service time is likely to get
5323 * finally computed for bfqq, freeing the inject limit from
5324 * its relation with the think time.
5325 */
5326 if (state_changed && bfqq->last_serv_time_ns == 0 &&
5327 (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
5328 msecs_to_jiffies(100)) ||
5329 !has_short_ttime))
5330 bfq_reset_inject_limit(bfqd, bfqq);
5331}
5332
5333/*
5334 * Called when a new fs request (rq) is added to bfqq. Check if there's
5335 * something we should do about it.
5336 */
5337static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5338 struct request *rq)
5339{
5340 if (rq->cmd_flags & REQ_META)
5341 bfqq->meta_pending++;
5342
5343 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
5344
5345 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
5346 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
5347 blk_rq_sectors(rq) < 32;
5348 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
5349
5350 /*
5351 * There is just this request queued: if
5352 * - the request is small, and
5353 * - we are idling to boost throughput, and
5354 * - the queue is not to be expired,
5355 * then just exit.
5356 *
5357 * In this way, if the device is being idled to wait
5358 * for a new request from the in-service queue, we
5359 * avoid unplugging the device and committing the
5360 * device to serve just a small request. In contrast
5361 * we wait for the block layer to decide when to
5362 * unplug the device: hopefully, new requests will be
5363 * merged to this one quickly, then the device will be
5364 * unplugged and larger requests will be dispatched.
5365 */
5366 if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
5367 !budget_timeout)
5368 return;
5369
5370 /*
5371 * A large enough request arrived, or idling is being
5372 * performed to preserve service guarantees, or
5373 * finally the queue is to be expired: in all these
5374 * cases disk idling is to be stopped, so clear
5375 * wait_request flag and reset timer.
5376 */
5377 bfq_clear_bfqq_wait_request(bfqq);
5378 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
5379
5380 /*
5381 * The queue is not empty, because a new request just
5382 * arrived. Hence we can safely expire the queue, in
5383 * case of budget timeout, without risking that the
5384 * timestamps of the queue are not updated correctly.
5385 * See [1] for more details.
5386 */
5387 if (budget_timeout)
5388 bfq_bfqq_expire(bfqd, bfqq, false,
5389 BFQQE_BUDGET_TIMEOUT);
5390 }
5391}
5392
5393/* returns true if it causes the idle timer to be disabled */
5394static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
5395{
5396 struct bfq_queue *bfqq = RQ_BFQQ(rq),
5397 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
5398 bool waiting, idle_timer_disabled = false;
5399
5400 if (new_bfqq) {
5401 /*
5402 * Release the request's reference to the old bfqq
5403 * and make sure one is taken to the shared queue.
5404 */
5405 new_bfqq->allocated++;
5406 bfqq->allocated--;
5407 new_bfqq->ref++;
5408 /*
5409 * If the bic associated with the process
5410 * issuing this request still points to bfqq
5411 * (and thus has not been already redirected
5412 * to new_bfqq or even some other bfq_queue),
5413 * then complete the merge and redirect it to
5414 * new_bfqq.
5415 */
5416 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
5417 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
5418 bfqq, new_bfqq);
5419
5420 bfq_clear_bfqq_just_created(bfqq);
5421 /*
5422 * rq is about to be enqueued into new_bfqq,
5423 * release rq reference on bfqq
5424 */
5425 bfq_put_queue(bfqq);
5426 rq->elv.priv[1] = new_bfqq;
5427 bfqq = new_bfqq;
5428 }
5429
5430 bfq_update_io_thinktime(bfqd, bfqq);
5431 bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
5432 bfq_update_io_seektime(bfqd, bfqq, rq);
5433
5434 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
5435 bfq_add_request(rq);
5436 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
5437
5438 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
5439 list_add_tail(&rq->queuelist, &bfqq->fifo);
5440
5441 bfq_rq_enqueued(bfqd, bfqq, rq);
5442
5443 return idle_timer_disabled;
5444}
5445
5446#ifdef CONFIG_BFQ_CGROUP_DEBUG
5447static void bfq_update_insert_stats(struct request_queue *q,
5448 struct bfq_queue *bfqq,
5449 bool idle_timer_disabled,
5450 unsigned int cmd_flags)
5451{
5452 if (!bfqq)
5453 return;
5454
5455 /*
5456 * bfqq still exists, because it can disappear only after
5457 * either it is merged with another queue, or the process it
5458 * is associated with exits. But both actions must be taken by
5459 * the same process currently executing this flow of
5460 * instructions.
5461 *
5462 * In addition, the following queue lock guarantees that
5463 * bfqq_group(bfqq) exists as well.
5464 */
5465 spin_lock_irq(&q->queue_lock);
5466 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
5467 if (idle_timer_disabled)
5468 bfqg_stats_update_idle_time(bfqq_group(bfqq));
5469 spin_unlock_irq(&q->queue_lock);
5470}
5471#else
5472static inline void bfq_update_insert_stats(struct request_queue *q,
5473 struct bfq_queue *bfqq,
5474 bool idle_timer_disabled,
5475 unsigned int cmd_flags) {}
5476#endif /* CONFIG_BFQ_CGROUP_DEBUG */
5477
5478static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
5479 bool at_head)
5480{
5481 struct request_queue *q = hctx->queue;
5482 struct bfq_data *bfqd = q->elevator->elevator_data;
5483 struct bfq_queue *bfqq;
5484 bool idle_timer_disabled = false;
5485 unsigned int cmd_flags;
5486
5487 spin_lock_irq(&bfqd->lock);
5488 if (blk_mq_sched_try_insert_merge(q, rq)) {
5489 spin_unlock_irq(&bfqd->lock);
5490 return;
5491 }
5492
5493 spin_unlock_irq(&bfqd->lock);
5494
5495 blk_mq_sched_request_inserted(rq);
5496
5497 spin_lock_irq(&bfqd->lock);
5498 bfqq = bfq_init_rq(rq);
5499 if (!bfqq || at_head || blk_rq_is_passthrough(rq)) {
5500 if (at_head)
5501 list_add(&rq->queuelist, &bfqd->dispatch);
5502 else
5503 list_add_tail(&rq->queuelist, &bfqd->dispatch);
5504 } else {
5505 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
5506 /*
5507 * Update bfqq, because, if a queue merge has occurred
5508 * in __bfq_insert_request, then rq has been
5509 * redirected into a new queue.
5510 */
5511 bfqq = RQ_BFQQ(rq);
5512
5513 if (rq_mergeable(rq)) {
5514 elv_rqhash_add(q, rq);
5515 if (!q->last_merge)
5516 q->last_merge = rq;
5517 }
5518 }
5519
5520 /*
5521 * Cache cmd_flags before releasing scheduler lock, because rq
5522 * may disappear afterwards (for example, because of a request
5523 * merge).
5524 */
5525 cmd_flags = rq->cmd_flags;
5526
5527 spin_unlock_irq(&bfqd->lock);
5528
5529 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
5530 cmd_flags);
5531}
5532
5533static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
5534 struct list_head *list, bool at_head)
5535{
5536 while (!list_empty(list)) {
5537 struct request *rq;
5538
5539 rq = list_first_entry(list, struct request, queuelist);
5540 list_del_init(&rq->queuelist);
5541 bfq_insert_request(hctx, rq, at_head);
5542 }
5543}
5544
5545static void bfq_update_hw_tag(struct bfq_data *bfqd)
5546{
5547 struct bfq_queue *bfqq = bfqd->in_service_queue;
5548
5549 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
5550 bfqd->rq_in_driver);
5551
5552 if (bfqd->hw_tag == 1)
5553 return;
5554
5555 /*
5556 * This sample is valid if the number of outstanding requests
5557 * is large enough to allow a queueing behavior. Note that the
5558 * sum is not exact, as it's not taking into account deactivated
5559 * requests.
5560 */
5561 if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
5562 return;
5563
5564 /*
5565 * If active queue hasn't enough requests and can idle, bfq might not
5566 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
5567 * case
5568 */
5569 if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
5570 bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
5571 BFQ_HW_QUEUE_THRESHOLD &&
5572 bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
5573 return;
5574
5575 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
5576 return;
5577
5578 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
5579 bfqd->max_rq_in_driver = 0;
5580 bfqd->hw_tag_samples = 0;
5581
5582 bfqd->nonrot_with_queueing =
5583 blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
5584}
5585
5586static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
5587{
5588 u64 now_ns;
5589 u32 delta_us;
5590
5591 bfq_update_hw_tag(bfqd);
5592
5593 bfqd->rq_in_driver--;
5594 bfqq->dispatched--;
5595
5596 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
5597 /*
5598 * Set budget_timeout (which we overload to store the
5599 * time at which the queue remains with no backlog and
5600 * no outstanding request; used by the weight-raising
5601 * mechanism).
5602 */
5603 bfqq->budget_timeout = jiffies;
5604
5605 bfq_weights_tree_remove(bfqd, bfqq);
5606 }
5607
5608 now_ns = ktime_get_ns();
5609
5610 bfqq->ttime.last_end_request = now_ns;
5611
5612 /*
5613 * Using us instead of ns, to get a reasonable precision in
5614 * computing rate in next check.
5615 */
5616 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
5617
5618 /*
5619 * If the request took rather long to complete, and, according
5620 * to the maximum request size recorded, this completion latency
5621 * implies that the request was certainly served at a very low
5622 * rate (less than 1M sectors/sec), then the whole observation
5623 * interval that lasts up to this time instant cannot be a
5624 * valid time interval for computing a new peak rate. Invoke
5625 * bfq_update_rate_reset to have the following three steps
5626 * taken:
5627 * - close the observation interval at the last (previous)
5628 * request dispatch or completion
5629 * - compute rate, if possible, for that observation interval
5630 * - reset to zero samples, which will trigger a proper
5631 * re-initialization of the observation interval on next
5632 * dispatch
5633 */
5634 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
5635 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
5636 1UL<<(BFQ_RATE_SHIFT - 10))
5637 bfq_update_rate_reset(bfqd, NULL);
5638 bfqd->last_completion = now_ns;
5639 bfqd->last_completed_rq_bfqq = bfqq;
5640
5641 /*
5642 * If we are waiting to discover whether the request pattern
5643 * of the task associated with the queue is actually
5644 * isochronous, and both requisites for this condition to hold
5645 * are now satisfied, then compute soft_rt_next_start (see the
5646 * comments on the function bfq_bfqq_softrt_next_start()). We
5647 * do not compute soft_rt_next_start if bfqq is in interactive
5648 * weight raising (see the comments in bfq_bfqq_expire() for
5649 * an explanation). We schedule this delayed update when bfqq
5650 * expires, if it still has in-flight requests.
5651 */
5652 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
5653 RB_EMPTY_ROOT(&bfqq->sort_list) &&
5654 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
5655 bfqq->soft_rt_next_start =
5656 bfq_bfqq_softrt_next_start(bfqd, bfqq);
5657
5658 /*
5659 * If this is the in-service queue, check if it needs to be expired,
5660 * or if we want to idle in case it has no pending requests.
5661 */
5662 if (bfqd->in_service_queue == bfqq) {
5663 if (bfq_bfqq_must_idle(bfqq)) {
5664 if (bfqq->dispatched == 0)
5665 bfq_arm_slice_timer(bfqd);
5666 /*
5667 * If we get here, we do not expire bfqq, even
5668 * if bfqq was in budget timeout or had no
5669 * more requests (as controlled in the next
5670 * conditional instructions). The reason for
5671 * not expiring bfqq is as follows.
5672 *
5673 * Here bfqq->dispatched > 0 holds, but
5674 * bfq_bfqq_must_idle() returned true. This
5675 * implies that, even if no request arrives
5676 * for bfqq before bfqq->dispatched reaches 0,
5677 * bfqq will, however, not be expired on the
5678 * completion event that causes bfqq->dispatch
5679 * to reach zero. In contrast, on this event,
5680 * bfqq will start enjoying device idling
5681 * (I/O-dispatch plugging).
5682 *
5683 * But, if we expired bfqq here, bfqq would
5684 * not have the chance to enjoy device idling
5685 * when bfqq->dispatched finally reaches
5686 * zero. This would expose bfqq to violation
5687 * of its reserved service guarantees.
5688 */
5689 return;
5690 } else if (bfq_may_expire_for_budg_timeout(bfqq))
5691 bfq_bfqq_expire(bfqd, bfqq, false,
5692 BFQQE_BUDGET_TIMEOUT);
5693 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
5694 (bfqq->dispatched == 0 ||
5695 !bfq_better_to_idle(bfqq)))
5696 bfq_bfqq_expire(bfqd, bfqq, false,
5697 BFQQE_NO_MORE_REQUESTS);
5698 }
5699
5700 if (!bfqd->rq_in_driver)
5701 bfq_schedule_dispatch(bfqd);
5702}
5703
5704static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq)
5705{
5706 bfqq->allocated--;
5707
5708 bfq_put_queue(bfqq);
5709}
5710
5711/*
5712 * The processes associated with bfqq may happen to generate their
5713 * cumulative I/O at a lower rate than the rate at which the device
5714 * could serve the same I/O. This is rather probable, e.g., if only
5715 * one process is associated with bfqq and the device is an SSD. It
5716 * results in bfqq becoming often empty while in service. In this
5717 * respect, if BFQ is allowed to switch to another queue when bfqq
5718 * remains empty, then the device goes on being fed with I/O requests,
5719 * and the throughput is not affected. In contrast, if BFQ is not
5720 * allowed to switch to another queue---because bfqq is sync and
5721 * I/O-dispatch needs to be plugged while bfqq is temporarily
5722 * empty---then, during the service of bfqq, there will be frequent
5723 * "service holes", i.e., time intervals during which bfqq gets empty
5724 * and the device can only consume the I/O already queued in its
5725 * hardware queues. During service holes, the device may even get to
5726 * remaining idle. In the end, during the service of bfqq, the device
5727 * is driven at a lower speed than the one it can reach with the kind
5728 * of I/O flowing through bfqq.
5729 *
5730 * To counter this loss of throughput, BFQ implements a "request
5731 * injection mechanism", which tries to fill the above service holes
5732 * with I/O requests taken from other queues. The hard part in this
5733 * mechanism is finding the right amount of I/O to inject, so as to
5734 * both boost throughput and not break bfqq's bandwidth and latency
5735 * guarantees. In this respect, the mechanism maintains a per-queue
5736 * inject limit, computed as below. While bfqq is empty, the injection
5737 * mechanism dispatches extra I/O requests only until the total number
5738 * of I/O requests in flight---i.e., already dispatched but not yet
5739 * completed---remains lower than this limit.
5740 *
5741 * A first definition comes in handy to introduce the algorithm by
5742 * which the inject limit is computed. We define as first request for
5743 * bfqq, an I/O request for bfqq that arrives while bfqq is in
5744 * service, and causes bfqq to switch from empty to non-empty. The
5745 * algorithm updates the limit as a function of the effect of
5746 * injection on the service times of only the first requests of
5747 * bfqq. The reason for this restriction is that these are the
5748 * requests whose service time is affected most, because they are the
5749 * first to arrive after injection possibly occurred.
5750 *
5751 * To evaluate the effect of injection, the algorithm measures the
5752 * "total service time" of first requests. We define as total service
5753 * time of an I/O request, the time that elapses since when the
5754 * request is enqueued into bfqq, to when it is completed. This
5755 * quantity allows the whole effect of injection to be measured. It is
5756 * easy to see why. Suppose that some requests of other queues are
5757 * actually injected while bfqq is empty, and that a new request R
5758 * then arrives for bfqq. If the device does start to serve all or
5759 * part of the injected requests during the service hole, then,
5760 * because of this extra service, it may delay the next invocation of
5761 * the dispatch hook of BFQ. Then, even after R gets eventually
5762 * dispatched, the device may delay the actual service of R if it is
5763 * still busy serving the extra requests, or if it decides to serve,
5764 * before R, some extra request still present in its queues. As a
5765 * conclusion, the cumulative extra delay caused by injection can be
5766 * easily evaluated by just comparing the total service time of first
5767 * requests with and without injection.
5768 *
5769 * The limit-update algorithm works as follows. On the arrival of a
5770 * first request of bfqq, the algorithm measures the total time of the
5771 * request only if one of the three cases below holds, and, for each
5772 * case, it updates the limit as described below:
5773 *
5774 * (1) If there is no in-flight request. This gives a baseline for the
5775 * total service time of the requests of bfqq. If the baseline has
5776 * not been computed yet, then, after computing it, the limit is
5777 * set to 1, to start boosting throughput, and to prepare the
5778 * ground for the next case. If the baseline has already been
5779 * computed, then it is updated, in case it results to be lower
5780 * than the previous value.
5781 *
5782 * (2) If the limit is higher than 0 and there are in-flight
5783 * requests. By comparing the total service time in this case with
5784 * the above baseline, it is possible to know at which extent the
5785 * current value of the limit is inflating the total service
5786 * time. If the inflation is below a certain threshold, then bfqq
5787 * is assumed to be suffering from no perceivable loss of its
5788 * service guarantees, and the limit is even tentatively
5789 * increased. If the inflation is above the threshold, then the
5790 * limit is decreased. Due to the lack of any hysteresis, this
5791 * logic makes the limit oscillate even in steady workload
5792 * conditions. Yet we opted for it, because it is fast in reaching
5793 * the best value for the limit, as a function of the current I/O
5794 * workload. To reduce oscillations, this step is disabled for a
5795 * short time interval after the limit happens to be decreased.
5796 *
5797 * (3) Periodically, after resetting the limit, to make sure that the
5798 * limit eventually drops in case the workload changes. This is
5799 * needed because, after the limit has gone safely up for a
5800 * certain workload, it is impossible to guess whether the
5801 * baseline total service time may have changed, without measuring
5802 * it again without injection. A more effective version of this
5803 * step might be to just sample the baseline, by interrupting
5804 * injection only once, and then to reset/lower the limit only if
5805 * the total service time with the current limit does happen to be
5806 * too large.
5807 *
5808 * More details on each step are provided in the comments on the
5809 * pieces of code that implement these steps: the branch handling the
5810 * transition from empty to non empty in bfq_add_request(), the branch
5811 * handling injection in bfq_select_queue(), and the function
5812 * bfq_choose_bfqq_for_injection(). These comments also explain some
5813 * exceptions, made by the injection mechanism in some special cases.
5814 */
5815static void bfq_update_inject_limit(struct bfq_data *bfqd,
5816 struct bfq_queue *bfqq)
5817{
5818 u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns;
5819 unsigned int old_limit = bfqq->inject_limit;
5820
5821 if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) {
5822 u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
5823
5824 if (tot_time_ns >= threshold && old_limit > 0) {
5825 bfqq->inject_limit--;
5826 bfqq->decrease_time_jif = jiffies;
5827 } else if (tot_time_ns < threshold &&
5828 old_limit <= bfqd->max_rq_in_driver)
5829 bfqq->inject_limit++;
5830 }
5831
5832 /*
5833 * Either we still have to compute the base value for the
5834 * total service time, and there seem to be the right
5835 * conditions to do it, or we can lower the last base value
5836 * computed.
5837 *
5838 * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
5839 * request in flight, because this function is in the code
5840 * path that handles the completion of a request of bfqq, and,
5841 * in particular, this function is executed before
5842 * bfqd->rq_in_driver is decremented in such a code path.
5843 */
5844 if ((bfqq->last_serv_time_ns == 0 && bfqd->rq_in_driver == 1) ||
5845 tot_time_ns < bfqq->last_serv_time_ns) {
5846 if (bfqq->last_serv_time_ns == 0) {
5847 /*
5848 * Now we certainly have a base value: make sure we
5849 * start trying injection.
5850 */
5851 bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
5852 }
5853 bfqq->last_serv_time_ns = tot_time_ns;
5854 } else if (!bfqd->rqs_injected && bfqd->rq_in_driver == 1)
5855 /*
5856 * No I/O injected and no request still in service in
5857 * the drive: these are the exact conditions for
5858 * computing the base value of the total service time
5859 * for bfqq. So let's update this value, because it is
5860 * rather variable. For example, it varies if the size
5861 * or the spatial locality of the I/O requests in bfqq
5862 * change.
5863 */
5864 bfqq->last_serv_time_ns = tot_time_ns;
5865
5866
5867 /* update complete, not waiting for any request completion any longer */
5868 bfqd->waited_rq = NULL;
5869 bfqd->rqs_injected = false;
5870}
5871
5872/*
5873 * Handle either a requeue or a finish for rq. The things to do are
5874 * the same in both cases: all references to rq are to be dropped. In
5875 * particular, rq is considered completed from the point of view of
5876 * the scheduler.
5877 */
5878static void bfq_finish_requeue_request(struct request *rq)
5879{
5880 struct bfq_queue *bfqq = RQ_BFQQ(rq);
5881 struct bfq_data *bfqd;
5882
5883 /*
5884 * Requeue and finish hooks are invoked in blk-mq without
5885 * checking whether the involved request is actually still
5886 * referenced in the scheduler. To handle this fact, the
5887 * following two checks make this function exit in case of
5888 * spurious invocations, for which there is nothing to do.
5889 *
5890 * First, check whether rq has nothing to do with an elevator.
5891 */
5892 if (unlikely(!(rq->rq_flags & RQF_ELVPRIV)))
5893 return;
5894
5895 /*
5896 * rq either is not associated with any icq, or is an already
5897 * requeued request that has not (yet) been re-inserted into
5898 * a bfq_queue.
5899 */
5900 if (!rq->elv.icq || !bfqq)
5901 return;
5902
5903 bfqd = bfqq->bfqd;
5904
5905 if (rq->rq_flags & RQF_STARTED)
5906 bfqg_stats_update_completion(bfqq_group(bfqq),
5907 rq->start_time_ns,
5908 rq->io_start_time_ns,
5909 rq->cmd_flags);
5910
5911 if (likely(rq->rq_flags & RQF_STARTED)) {
5912 unsigned long flags;
5913
5914 spin_lock_irqsave(&bfqd->lock, flags);
5915
5916 if (rq == bfqd->waited_rq)
5917 bfq_update_inject_limit(bfqd, bfqq);
5918
5919 bfq_completed_request(bfqq, bfqd);
5920 bfq_finish_requeue_request_body(bfqq);
5921
5922 spin_unlock_irqrestore(&bfqd->lock, flags);
5923 } else {
5924 /*
5925 * Request rq may be still/already in the scheduler,
5926 * in which case we need to remove it (this should
5927 * never happen in case of requeue). And we cannot
5928 * defer such a check and removal, to avoid
5929 * inconsistencies in the time interval from the end
5930 * of this function to the start of the deferred work.
5931 * This situation seems to occur only in process
5932 * context, as a consequence of a merge. In the
5933 * current version of the code, this implies that the
5934 * lock is held.
5935 */
5936
5937 if (!RB_EMPTY_NODE(&rq->rb_node)) {
5938 bfq_remove_request(rq->q, rq);
5939 bfqg_stats_update_io_remove(bfqq_group(bfqq),
5940 rq->cmd_flags);
5941 }
5942 bfq_finish_requeue_request_body(bfqq);
5943 }
5944
5945 /*
5946 * Reset private fields. In case of a requeue, this allows
5947 * this function to correctly do nothing if it is spuriously
5948 * invoked again on this same request (see the check at the
5949 * beginning of the function). Probably, a better general
5950 * design would be to prevent blk-mq from invoking the requeue
5951 * or finish hooks of an elevator, for a request that is not
5952 * referred by that elevator.
5953 *
5954 * Resetting the following fields would break the
5955 * request-insertion logic if rq is re-inserted into a bfq
5956 * internal queue, without a re-preparation. Here we assume
5957 * that re-insertions of requeued requests, without
5958 * re-preparation, can happen only for pass_through or at_head
5959 * requests (which are not re-inserted into bfq internal
5960 * queues).
5961 */
5962 rq->elv.priv[0] = NULL;
5963 rq->elv.priv[1] = NULL;
5964}
5965
5966/*
5967 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
5968 * was the last process referring to that bfqq.
5969 */
5970static struct bfq_queue *
5971bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
5972{
5973 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
5974
5975 if (bfqq_process_refs(bfqq) == 1) {
5976 bfqq->pid = current->pid;
5977 bfq_clear_bfqq_coop(bfqq);
5978 bfq_clear_bfqq_split_coop(bfqq);
5979 return bfqq;
5980 }
5981
5982 bic_set_bfqq(bic, NULL, 1);
5983
5984 bfq_put_cooperator(bfqq);
5985
5986 bfq_release_process_ref(bfqq->bfqd, bfqq);
5987 return NULL;
5988}
5989
5990static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
5991 struct bfq_io_cq *bic,
5992 struct bio *bio,
5993 bool split, bool is_sync,
5994 bool *new_queue)
5995{
5996 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
5997
5998 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
5999 return bfqq;
6000
6001 if (new_queue)
6002 *new_queue = true;
6003
6004 if (bfqq)
6005 bfq_put_queue(bfqq);
6006 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
6007
6008 bic_set_bfqq(bic, bfqq, is_sync);
6009 if (split && is_sync) {
6010 if ((bic->was_in_burst_list && bfqd->large_burst) ||
6011 bic->saved_in_large_burst)
6012 bfq_mark_bfqq_in_large_burst(bfqq);
6013 else {
6014 bfq_clear_bfqq_in_large_burst(bfqq);
6015 if (bic->was_in_burst_list)
6016 /*
6017 * If bfqq was in the current
6018 * burst list before being
6019 * merged, then we have to add
6020 * it back. And we do not need
6021 * to increase burst_size, as
6022 * we did not decrement
6023 * burst_size when we removed
6024 * bfqq from the burst list as
6025 * a consequence of a merge
6026 * (see comments in
6027 * bfq_put_queue). In this
6028 * respect, it would be rather
6029 * costly to know whether the
6030 * current burst list is still
6031 * the same burst list from
6032 * which bfqq was removed on
6033 * the merge. To avoid this
6034 * cost, if bfqq was in a
6035 * burst list, then we add
6036 * bfqq to the current burst
6037 * list without any further
6038 * check. This can cause
6039 * inappropriate insertions,
6040 * but rarely enough to not
6041 * harm the detection of large
6042 * bursts significantly.
6043 */
6044 hlist_add_head(&bfqq->burst_list_node,
6045 &bfqd->burst_list);
6046 }
6047 bfqq->split_time = jiffies;
6048 }
6049
6050 return bfqq;
6051}
6052
6053/*
6054 * Only reset private fields. The actual request preparation will be
6055 * performed by bfq_init_rq, when rq is either inserted or merged. See
6056 * comments on bfq_init_rq for the reason behind this delayed
6057 * preparation.
6058 */
6059static void bfq_prepare_request(struct request *rq, struct bio *bio)
6060{
6061 /*
6062 * Regardless of whether we have an icq attached, we have to
6063 * clear the scheduler pointers, as they might point to
6064 * previously allocated bic/bfqq structs.
6065 */
6066 rq->elv.priv[0] = rq->elv.priv[1] = NULL;
6067}
6068
6069/*
6070 * If needed, init rq, allocate bfq data structures associated with
6071 * rq, and increment reference counters in the destination bfq_queue
6072 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6073 * not associated with any bfq_queue.
6074 *
6075 * This function is invoked by the functions that perform rq insertion
6076 * or merging. One may have expected the above preparation operations
6077 * to be performed in bfq_prepare_request, and not delayed to when rq
6078 * is inserted or merged. The rationale behind this delayed
6079 * preparation is that, after the prepare_request hook is invoked for
6080 * rq, rq may still be transformed into a request with no icq, i.e., a
6081 * request not associated with any queue. No bfq hook is invoked to
6082 * signal this transformation. As a consequence, should these
6083 * preparation operations be performed when the prepare_request hook
6084 * is invoked, and should rq be transformed one moment later, bfq
6085 * would end up in an inconsistent state, because it would have
6086 * incremented some queue counters for an rq destined to
6087 * transformation, without any chance to correctly lower these
6088 * counters back. In contrast, no transformation can still happen for
6089 * rq after rq has been inserted or merged. So, it is safe to execute
6090 * these preparation operations when rq is finally inserted or merged.
6091 */
6092static struct bfq_queue *bfq_init_rq(struct request *rq)
6093{
6094 struct request_queue *q = rq->q;
6095 struct bio *bio = rq->bio;
6096 struct bfq_data *bfqd = q->elevator->elevator_data;
6097 struct bfq_io_cq *bic;
6098 const int is_sync = rq_is_sync(rq);
6099 struct bfq_queue *bfqq;
6100 bool new_queue = false;
6101 bool bfqq_already_existing = false, split = false;
6102
6103 if (unlikely(!rq->elv.icq))
6104 return NULL;
6105
6106 /*
6107 * Assuming that elv.priv[1] is set only if everything is set
6108 * for this rq. This holds true, because this function is
6109 * invoked only for insertion or merging, and, after such
6110 * events, a request cannot be manipulated any longer before
6111 * being removed from bfq.
6112 */
6113 if (rq->elv.priv[1])
6114 return rq->elv.priv[1];
6115
6116 bic = icq_to_bic(rq->elv.icq);
6117
6118 bfq_check_ioprio_change(bic, bio);
6119
6120 bfq_bic_update_cgroup(bic, bio);
6121
6122 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
6123 &new_queue);
6124
6125 if (likely(!new_queue)) {
6126 /* If the queue was seeky for too long, break it apart. */
6127 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
6128 bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
6129
6130 /* Update bic before losing reference to bfqq */
6131 if (bfq_bfqq_in_large_burst(bfqq))
6132 bic->saved_in_large_burst = true;
6133
6134 bfqq = bfq_split_bfqq(bic, bfqq);
6135 split = true;
6136
6137 if (!bfqq)
6138 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
6139 true, is_sync,
6140 NULL);
6141 else
6142 bfqq_already_existing = true;
6143 }
6144 }
6145
6146 bfqq->allocated++;
6147 bfqq->ref++;
6148 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
6149 rq, bfqq, bfqq->ref);
6150
6151 rq->elv.priv[0] = bic;
6152 rq->elv.priv[1] = bfqq;
6153
6154 /*
6155 * If a bfq_queue has only one process reference, it is owned
6156 * by only this bic: we can then set bfqq->bic = bic. in
6157 * addition, if the queue has also just been split, we have to
6158 * resume its state.
6159 */
6160 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
6161 bfqq->bic = bic;
6162 if (split) {
6163 /*
6164 * The queue has just been split from a shared
6165 * queue: restore the idle window and the
6166 * possible weight raising period.
6167 */
6168 bfq_bfqq_resume_state(bfqq, bfqd, bic,
6169 bfqq_already_existing);
6170 }
6171 }
6172
6173 /*
6174 * Consider bfqq as possibly belonging to a burst of newly
6175 * created queues only if:
6176 * 1) A burst is actually happening (bfqd->burst_size > 0)
6177 * or
6178 * 2) There is no other active queue. In fact, if, in
6179 * contrast, there are active queues not belonging to the
6180 * possible burst bfqq may belong to, then there is no gain
6181 * in considering bfqq as belonging to a burst, and
6182 * therefore in not weight-raising bfqq. See comments on
6183 * bfq_handle_burst().
6184 *
6185 * This filtering also helps eliminating false positives,
6186 * occurring when bfqq does not belong to an actual large
6187 * burst, but some background task (e.g., a service) happens
6188 * to trigger the creation of new queues very close to when
6189 * bfqq and its possible companion queues are created. See
6190 * comments on bfq_handle_burst() for further details also on
6191 * this issue.
6192 */
6193 if (unlikely(bfq_bfqq_just_created(bfqq) &&
6194 (bfqd->burst_size > 0 ||
6195 bfq_tot_busy_queues(bfqd) == 0)))
6196 bfq_handle_burst(bfqd, bfqq);
6197
6198 return bfqq;
6199}
6200
6201static void bfq_idle_slice_timer_body(struct bfq_queue *bfqq)
6202{
6203 struct bfq_data *bfqd = bfqq->bfqd;
6204 enum bfqq_expiration reason;
6205 unsigned long flags;
6206
6207 spin_lock_irqsave(&bfqd->lock, flags);
6208 bfq_clear_bfqq_wait_request(bfqq);
6209
6210 if (bfqq != bfqd->in_service_queue) {
6211 spin_unlock_irqrestore(&bfqd->lock, flags);
6212 return;
6213 }
6214
6215 if (bfq_bfqq_budget_timeout(bfqq))
6216 /*
6217 * Also here the queue can be safely expired
6218 * for budget timeout without wasting
6219 * guarantees
6220 */
6221 reason = BFQQE_BUDGET_TIMEOUT;
6222 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
6223 /*
6224 * The queue may not be empty upon timer expiration,
6225 * because we may not disable the timer when the
6226 * first request of the in-service queue arrives
6227 * during disk idling.
6228 */
6229 reason = BFQQE_TOO_IDLE;
6230 else
6231 goto schedule_dispatch;
6232
6233 bfq_bfqq_expire(bfqd, bfqq, true, reason);
6234
6235schedule_dispatch:
6236 spin_unlock_irqrestore(&bfqd->lock, flags);
6237 bfq_schedule_dispatch(bfqd);
6238}
6239
6240/*
6241 * Handler of the expiration of the timer running if the in-service queue
6242 * is idling inside its time slice.
6243 */
6244static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
6245{
6246 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
6247 idle_slice_timer);
6248 struct bfq_queue *bfqq = bfqd->in_service_queue;
6249
6250 /*
6251 * Theoretical race here: the in-service queue can be NULL or
6252 * different from the queue that was idling if a new request
6253 * arrives for the current queue and there is a full dispatch
6254 * cycle that changes the in-service queue. This can hardly
6255 * happen, but in the worst case we just expire a queue too
6256 * early.
6257 */
6258 if (bfqq)
6259 bfq_idle_slice_timer_body(bfqq);
6260
6261 return HRTIMER_NORESTART;
6262}
6263
6264static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
6265 struct bfq_queue **bfqq_ptr)
6266{
6267 struct bfq_queue *bfqq = *bfqq_ptr;
6268
6269 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
6270 if (bfqq) {
6271 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
6272
6273 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
6274 bfqq, bfqq->ref);
6275 bfq_put_queue(bfqq);
6276 *bfqq_ptr = NULL;
6277 }
6278}
6279
6280/*
6281 * Release all the bfqg references to its async queues. If we are
6282 * deallocating the group these queues may still contain requests, so
6283 * we reparent them to the root cgroup (i.e., the only one that will
6284 * exist for sure until all the requests on a device are gone).
6285 */
6286void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
6287{
6288 int i, j;
6289
6290 for (i = 0; i < 2; i++)
6291 for (j = 0; j < IOPRIO_BE_NR; j++)
6292 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
6293
6294 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
6295}
6296
6297/*
6298 * See the comments on bfq_limit_depth for the purpose of
6299 * the depths set in the function. Return minimum shallow depth we'll use.
6300 */
6301static unsigned int bfq_update_depths(struct bfq_data *bfqd,
6302 struct sbitmap_queue *bt)
6303{
6304 unsigned int i, j, min_shallow = UINT_MAX;
6305
6306 /*
6307 * In-word depths if no bfq_queue is being weight-raised:
6308 * leaving 25% of tags only for sync reads.
6309 *
6310 * In next formulas, right-shift the value
6311 * (1U<<bt->sb.shift), instead of computing directly
6312 * (1U<<(bt->sb.shift - something)), to be robust against
6313 * any possible value of bt->sb.shift, without having to
6314 * limit 'something'.
6315 */
6316 /* no more than 50% of tags for async I/O */
6317 bfqd->word_depths[0][0] = max((1U << bt->sb.shift) >> 1, 1U);
6318 /*
6319 * no more than 75% of tags for sync writes (25% extra tags
6320 * w.r.t. async I/O, to prevent async I/O from starving sync
6321 * writes)
6322 */
6323 bfqd->word_depths[0][1] = max(((1U << bt->sb.shift) * 3) >> 2, 1U);
6324
6325 /*
6326 * In-word depths in case some bfq_queue is being weight-
6327 * raised: leaving ~63% of tags for sync reads. This is the
6328 * highest percentage for which, in our tests, application
6329 * start-up times didn't suffer from any regression due to tag
6330 * shortage.
6331 */
6332 /* no more than ~18% of tags for async I/O */
6333 bfqd->word_depths[1][0] = max(((1U << bt->sb.shift) * 3) >> 4, 1U);
6334 /* no more than ~37% of tags for sync writes (~20% extra tags) */
6335 bfqd->word_depths[1][1] = max(((1U << bt->sb.shift) * 6) >> 4, 1U);
6336
6337 for (i = 0; i < 2; i++)
6338 for (j = 0; j < 2; j++)
6339 min_shallow = min(min_shallow, bfqd->word_depths[i][j]);
6340
6341 return min_shallow;
6342}
6343
6344static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
6345{
6346 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
6347 struct blk_mq_tags *tags = hctx->sched_tags;
6348 unsigned int min_shallow;
6349
6350 min_shallow = bfq_update_depths(bfqd, &tags->bitmap_tags);
6351 sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, min_shallow);
6352}
6353
6354static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
6355{
6356 bfq_depth_updated(hctx);
6357 return 0;
6358}
6359
6360static void bfq_exit_queue(struct elevator_queue *e)
6361{
6362 struct bfq_data *bfqd = e->elevator_data;
6363 struct bfq_queue *bfqq, *n;
6364
6365 hrtimer_cancel(&bfqd->idle_slice_timer);
6366
6367 spin_lock_irq(&bfqd->lock);
6368 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
6369 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
6370 spin_unlock_irq(&bfqd->lock);
6371
6372 hrtimer_cancel(&bfqd->idle_slice_timer);
6373
6374#ifdef CONFIG_BFQ_GROUP_IOSCHED
6375 /* release oom-queue reference to root group */
6376 bfqg_and_blkg_put(bfqd->root_group);
6377
6378 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
6379#else
6380 spin_lock_irq(&bfqd->lock);
6381 bfq_put_async_queues(bfqd, bfqd->root_group);
6382 kfree(bfqd->root_group);
6383 spin_unlock_irq(&bfqd->lock);
6384#endif
6385
6386 kfree(bfqd);
6387}
6388
6389static void bfq_init_root_group(struct bfq_group *root_group,
6390 struct bfq_data *bfqd)
6391{
6392 int i;
6393
6394#ifdef CONFIG_BFQ_GROUP_IOSCHED
6395 root_group->entity.parent = NULL;
6396 root_group->my_entity = NULL;
6397 root_group->bfqd = bfqd;
6398#endif
6399 root_group->rq_pos_tree = RB_ROOT;
6400 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
6401 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
6402 root_group->sched_data.bfq_class_idle_last_service = jiffies;
6403}
6404
6405static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
6406{
6407 struct bfq_data *bfqd;
6408 struct elevator_queue *eq;
6409
6410 eq = elevator_alloc(q, e);
6411 if (!eq)
6412 return -ENOMEM;
6413
6414 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
6415 if (!bfqd) {
6416 kobject_put(&eq->kobj);
6417 return -ENOMEM;
6418 }
6419 eq->elevator_data = bfqd;
6420
6421 spin_lock_irq(&q->queue_lock);
6422 q->elevator = eq;
6423 spin_unlock_irq(&q->queue_lock);
6424
6425 /*
6426 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
6427 * Grab a permanent reference to it, so that the normal code flow
6428 * will not attempt to free it.
6429 */
6430 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
6431 bfqd->oom_bfqq.ref++;
6432 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
6433 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
6434 bfqd->oom_bfqq.entity.new_weight =
6435 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
6436
6437 /* oom_bfqq does not participate to bursts */
6438 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
6439
6440 /*
6441 * Trigger weight initialization, according to ioprio, at the
6442 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
6443 * class won't be changed any more.
6444 */
6445 bfqd->oom_bfqq.entity.prio_changed = 1;
6446
6447 bfqd->queue = q;
6448
6449 INIT_LIST_HEAD(&bfqd->dispatch);
6450
6451 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
6452 HRTIMER_MODE_REL);
6453 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
6454
6455 bfqd->queue_weights_tree = RB_ROOT_CACHED;
6456 bfqd->num_groups_with_pending_reqs = 0;
6457
6458 INIT_LIST_HEAD(&bfqd->active_list);
6459 INIT_LIST_HEAD(&bfqd->idle_list);
6460 INIT_HLIST_HEAD(&bfqd->burst_list);
6461
6462 bfqd->hw_tag = -1;
6463 bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
6464
6465 bfqd->bfq_max_budget = bfq_default_max_budget;
6466
6467 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
6468 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
6469 bfqd->bfq_back_max = bfq_back_max;
6470 bfqd->bfq_back_penalty = bfq_back_penalty;
6471 bfqd->bfq_slice_idle = bfq_slice_idle;
6472 bfqd->bfq_timeout = bfq_timeout;
6473
6474 bfqd->bfq_requests_within_timer = 120;
6475
6476 bfqd->bfq_large_burst_thresh = 8;
6477 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
6478
6479 bfqd->low_latency = true;
6480
6481 /*
6482 * Trade-off between responsiveness and fairness.
6483 */
6484 bfqd->bfq_wr_coeff = 30;
6485 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
6486 bfqd->bfq_wr_max_time = 0;
6487 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
6488 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
6489 bfqd->bfq_wr_max_softrt_rate = 7000; /*
6490 * Approximate rate required
6491 * to playback or record a
6492 * high-definition compressed
6493 * video.
6494 */
6495 bfqd->wr_busy_queues = 0;
6496
6497 /*
6498 * Begin by assuming, optimistically, that the device peak
6499 * rate is equal to 2/3 of the highest reference rate.
6500 */
6501 bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
6502 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
6503 bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
6504
6505 spin_lock_init(&bfqd->lock);
6506
6507 /*
6508 * The invocation of the next bfq_create_group_hierarchy
6509 * function is the head of a chain of function calls
6510 * (bfq_create_group_hierarchy->blkcg_activate_policy->
6511 * blk_mq_freeze_queue) that may lead to the invocation of the
6512 * has_work hook function. For this reason,
6513 * bfq_create_group_hierarchy is invoked only after all
6514 * scheduler data has been initialized, apart from the fields
6515 * that can be initialized only after invoking
6516 * bfq_create_group_hierarchy. This, in particular, enables
6517 * has_work to correctly return false. Of course, to avoid
6518 * other inconsistencies, the blk-mq stack must then refrain
6519 * from invoking further scheduler hooks before this init
6520 * function is finished.
6521 */
6522 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
6523 if (!bfqd->root_group)
6524 goto out_free;
6525 bfq_init_root_group(bfqd->root_group, bfqd);
6526 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
6527
6528 wbt_disable_default(q);
6529 return 0;
6530
6531out_free:
6532 kfree(bfqd);
6533 kobject_put(&eq->kobj);
6534 return -ENOMEM;
6535}
6536
6537static void bfq_slab_kill(void)
6538{
6539 kmem_cache_destroy(bfq_pool);
6540}
6541
6542static int __init bfq_slab_setup(void)
6543{
6544 bfq_pool = KMEM_CACHE(bfq_queue, 0);
6545 if (!bfq_pool)
6546 return -ENOMEM;
6547 return 0;
6548}
6549
6550static ssize_t bfq_var_show(unsigned int var, char *page)
6551{
6552 return sprintf(page, "%u\n", var);
6553}
6554
6555static int bfq_var_store(unsigned long *var, const char *page)
6556{
6557 unsigned long new_val;
6558 int ret = kstrtoul(page, 10, &new_val);
6559
6560 if (ret)
6561 return ret;
6562 *var = new_val;
6563 return 0;
6564}
6565
6566#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
6567static ssize_t __FUNC(struct elevator_queue *e, char *page) \
6568{ \
6569 struct bfq_data *bfqd = e->elevator_data; \
6570 u64 __data = __VAR; \
6571 if (__CONV == 1) \
6572 __data = jiffies_to_msecs(__data); \
6573 else if (__CONV == 2) \
6574 __data = div_u64(__data, NSEC_PER_MSEC); \
6575 return bfq_var_show(__data, (page)); \
6576}
6577SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
6578SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
6579SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
6580SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
6581SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
6582SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
6583SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
6584SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
6585SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
6586#undef SHOW_FUNCTION
6587
6588#define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
6589static ssize_t __FUNC(struct elevator_queue *e, char *page) \
6590{ \
6591 struct bfq_data *bfqd = e->elevator_data; \
6592 u64 __data = __VAR; \
6593 __data = div_u64(__data, NSEC_PER_USEC); \
6594 return bfq_var_show(__data, (page)); \
6595}
6596USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
6597#undef USEC_SHOW_FUNCTION
6598
6599#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
6600static ssize_t \
6601__FUNC(struct elevator_queue *e, const char *page, size_t count) \
6602{ \
6603 struct bfq_data *bfqd = e->elevator_data; \
6604 unsigned long __data, __min = (MIN), __max = (MAX); \
6605 int ret; \
6606 \
6607 ret = bfq_var_store(&__data, (page)); \
6608 if (ret) \
6609 return ret; \
6610 if (__data < __min) \
6611 __data = __min; \
6612 else if (__data > __max) \
6613 __data = __max; \
6614 if (__CONV == 1) \
6615 *(__PTR) = msecs_to_jiffies(__data); \
6616 else if (__CONV == 2) \
6617 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
6618 else \
6619 *(__PTR) = __data; \
6620 return count; \
6621}
6622STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
6623 INT_MAX, 2);
6624STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
6625 INT_MAX, 2);
6626STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
6627STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
6628 INT_MAX, 0);
6629STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
6630#undef STORE_FUNCTION
6631
6632#define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
6633static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
6634{ \
6635 struct bfq_data *bfqd = e->elevator_data; \
6636 unsigned long __data, __min = (MIN), __max = (MAX); \
6637 int ret; \
6638 \
6639 ret = bfq_var_store(&__data, (page)); \
6640 if (ret) \
6641 return ret; \
6642 if (__data < __min) \
6643 __data = __min; \
6644 else if (__data > __max) \
6645 __data = __max; \
6646 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
6647 return count; \
6648}
6649USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
6650 UINT_MAX);
6651#undef USEC_STORE_FUNCTION
6652
6653static ssize_t bfq_max_budget_store(struct elevator_queue *e,
6654 const char *page, size_t count)
6655{
6656 struct bfq_data *bfqd = e->elevator_data;
6657 unsigned long __data;
6658 int ret;
6659
6660 ret = bfq_var_store(&__data, (page));
6661 if (ret)
6662 return ret;
6663
6664 if (__data == 0)
6665 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
6666 else {
6667 if (__data > INT_MAX)
6668 __data = INT_MAX;
6669 bfqd->bfq_max_budget = __data;
6670 }
6671
6672 bfqd->bfq_user_max_budget = __data;
6673
6674 return count;
6675}
6676
6677/*
6678 * Leaving this name to preserve name compatibility with cfq
6679 * parameters, but this timeout is used for both sync and async.
6680 */
6681static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
6682 const char *page, size_t count)
6683{
6684 struct bfq_data *bfqd = e->elevator_data;
6685 unsigned long __data;
6686 int ret;
6687
6688 ret = bfq_var_store(&__data, (page));
6689 if (ret)
6690 return ret;
6691
6692 if (__data < 1)
6693 __data = 1;
6694 else if (__data > INT_MAX)
6695 __data = INT_MAX;
6696
6697 bfqd->bfq_timeout = msecs_to_jiffies(__data);
6698 if (bfqd->bfq_user_max_budget == 0)
6699 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
6700
6701 return count;
6702}
6703
6704static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
6705 const char *page, size_t count)
6706{
6707 struct bfq_data *bfqd = e->elevator_data;
6708 unsigned long __data;
6709 int ret;
6710
6711 ret = bfq_var_store(&__data, (page));
6712 if (ret)
6713 return ret;
6714
6715 if (__data > 1)
6716 __data = 1;
6717 if (!bfqd->strict_guarantees && __data == 1
6718 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
6719 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
6720
6721 bfqd->strict_guarantees = __data;
6722
6723 return count;
6724}
6725
6726static ssize_t bfq_low_latency_store(struct elevator_queue *e,
6727 const char *page, size_t count)
6728{
6729 struct bfq_data *bfqd = e->elevator_data;
6730 unsigned long __data;
6731 int ret;
6732
6733 ret = bfq_var_store(&__data, (page));
6734 if (ret)
6735 return ret;
6736
6737 if (__data > 1)
6738 __data = 1;
6739 if (__data == 0 && bfqd->low_latency != 0)
6740 bfq_end_wr(bfqd);
6741 bfqd->low_latency = __data;
6742
6743 return count;
6744}
6745
6746#define BFQ_ATTR(name) \
6747 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
6748
6749static struct elv_fs_entry bfq_attrs[] = {
6750 BFQ_ATTR(fifo_expire_sync),
6751 BFQ_ATTR(fifo_expire_async),
6752 BFQ_ATTR(back_seek_max),
6753 BFQ_ATTR(back_seek_penalty),
6754 BFQ_ATTR(slice_idle),
6755 BFQ_ATTR(slice_idle_us),
6756 BFQ_ATTR(max_budget),
6757 BFQ_ATTR(timeout_sync),
6758 BFQ_ATTR(strict_guarantees),
6759 BFQ_ATTR(low_latency),
6760 __ATTR_NULL
6761};
6762
6763static struct elevator_type iosched_bfq_mq = {
6764 .ops = {
6765 .limit_depth = bfq_limit_depth,
6766 .prepare_request = bfq_prepare_request,
6767 .requeue_request = bfq_finish_requeue_request,
6768 .finish_request = bfq_finish_requeue_request,
6769 .exit_icq = bfq_exit_icq,
6770 .insert_requests = bfq_insert_requests,
6771 .dispatch_request = bfq_dispatch_request,
6772 .next_request = elv_rb_latter_request,
6773 .former_request = elv_rb_former_request,
6774 .allow_merge = bfq_allow_bio_merge,
6775 .bio_merge = bfq_bio_merge,
6776 .request_merge = bfq_request_merge,
6777 .requests_merged = bfq_requests_merged,
6778 .request_merged = bfq_request_merged,
6779 .has_work = bfq_has_work,
6780 .depth_updated = bfq_depth_updated,
6781 .init_hctx = bfq_init_hctx,
6782 .init_sched = bfq_init_queue,
6783 .exit_sched = bfq_exit_queue,
6784 },
6785
6786 .icq_size = sizeof(struct bfq_io_cq),
6787 .icq_align = __alignof__(struct bfq_io_cq),
6788 .elevator_attrs = bfq_attrs,
6789 .elevator_name = "bfq",
6790 .elevator_owner = THIS_MODULE,
6791};
6792MODULE_ALIAS("bfq-iosched");
6793
6794static int __init bfq_init(void)
6795{
6796 int ret;
6797
6798#ifdef CONFIG_BFQ_GROUP_IOSCHED
6799 ret = blkcg_policy_register(&blkcg_policy_bfq);
6800 if (ret)
6801 return ret;
6802#endif
6803
6804 ret = -ENOMEM;
6805 if (bfq_slab_setup())
6806 goto err_pol_unreg;
6807
6808 /*
6809 * Times to load large popular applications for the typical
6810 * systems installed on the reference devices (see the
6811 * comments before the definition of the next
6812 * array). Actually, we use slightly lower values, as the
6813 * estimated peak rate tends to be smaller than the actual
6814 * peak rate. The reason for this last fact is that estimates
6815 * are computed over much shorter time intervals than the long
6816 * intervals typically used for benchmarking. Why? First, to
6817 * adapt more quickly to variations. Second, because an I/O
6818 * scheduler cannot rely on a peak-rate-evaluation workload to
6819 * be run for a long time.
6820 */
6821 ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
6822 ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
6823
6824 ret = elv_register(&iosched_bfq_mq);
6825 if (ret)
6826 goto slab_kill;
6827
6828 return 0;
6829
6830slab_kill:
6831 bfq_slab_kill();
6832err_pol_unreg:
6833#ifdef CONFIG_BFQ_GROUP_IOSCHED
6834 blkcg_policy_unregister(&blkcg_policy_bfq);
6835#endif
6836 return ret;
6837}
6838
6839static void __exit bfq_exit(void)
6840{
6841 elv_unregister(&iosched_bfq_mq);
6842#ifdef CONFIG_BFQ_GROUP_IOSCHED
6843 blkcg_policy_unregister(&blkcg_policy_bfq);
6844#endif
6845 bfq_slab_kill();
6846}
6847
6848module_init(bfq_init);
6849module_exit(bfq_exit);
6850
6851MODULE_AUTHOR("Paolo Valente");
6852MODULE_LICENSE("GPL");
6853MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");