1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Budget Fair Queueing (BFQ) I/O scheduler.
5 * Based on ideas and code from CFQ:
6 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9 * Paolo Valente <paolo.valente@unimore.it>
11 * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
12 * Arianna Avanzini <avanzini@google.com>
14 * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
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.txt.
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
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.
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.
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.
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.
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.
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
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+.
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
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
104 * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
105 * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
108 * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
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.
114 * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
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>
129 #include "blk-mq-tag.h"
130 #include "blk-mq-sched.h"
131 #include "bfq-iosched.h"
134 #define BFQ_BFQQ_FNS(name) \
135 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
137 __set_bit(BFQQF_##name, &(bfqq)->flags); \
139 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
141 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
143 int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
145 return test_bit(BFQQF_##name, &(bfqq)->flags); \
148 BFQ_BFQQ_FNS(just_created);
150 BFQ_BFQQ_FNS(wait_request);
151 BFQ_BFQQ_FNS(non_blocking_wait_rq);
152 BFQ_BFQQ_FNS(fifo_expire);
153 BFQ_BFQQ_FNS(has_short_ttime);
155 BFQ_BFQQ_FNS(IO_bound);
156 BFQ_BFQQ_FNS(in_large_burst);
158 BFQ_BFQQ_FNS(split_coop);
159 BFQ_BFQQ_FNS(softrt_update);
160 BFQ_BFQQ_FNS(has_waker);
161 #undef BFQ_BFQQ_FNS \
163 /* Expiration time of sync (0) and async (1) requests, in ns. */
164 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
166 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
167 static const int bfq_back_max = 16 * 1024;
169 /* Penalty of a backwards seek, in number of sectors. */
170 static const int bfq_back_penalty = 2;
172 /* Idling period duration, in ns. */
173 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
175 /* Minimum number of assigned budgets for which stats are safe to compute. */
176 static const int bfq_stats_min_budgets = 194;
178 /* Default maximum budget values, in sectors and number of requests. */
179 static const int bfq_default_max_budget = 16 * 1024;
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.
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.
200 static const int bfq_async_charge_factor = 3;
202 /* Default timeout values, in jiffies, approximating CFQ defaults. */
203 const int bfq_timeout = HZ / 8;
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
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.
217 static const unsigned long bfq_merge_time_limit = HZ/10;
219 static struct kmem_cache *bfq_pool;
221 /* Below this threshold (in ns), we consider thinktime immediate. */
222 #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
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
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) > \
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)
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
244 #define BFQQ_TOTALLY_SEEKY(bfqq) (bfqq->seek_history & -1)
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
254 * Shift used for peak-rate fixed precision calculations.
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
267 #define BFQ_RATE_SHIFT 16
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.
286 * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
287 * depending on whether the device is rotational or non-rotational.
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
300 * The reference peak rates are measured in sectors/usec, left-shifted
303 static int ref_rate[2] = {14000, 33000};
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.
309 static int ref_wr_duration[2];
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
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.
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.
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.
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.
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.
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.
363 static const unsigned long max_service_from_wr = 120000;
365 #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
366 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
368 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
370 return bic->bfqq[is_sync];
373 void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
375 bic->bfqq[is_sync] = bfqq;
378 struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
380 return bic->icq.q->elevator->elevator_data;
384 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
385 * @icq: the iocontext queue.
387 static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
389 /* bic->icq is the first member, %NULL will convert to %NULL */
390 return container_of(icq, struct bfq_io_cq, icq);
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.
399 static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd,
400 struct io_context *ioc,
401 struct request_queue *q)
405 struct bfq_io_cq *icq;
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);
418 * Scheduler run of queue, if there are requests pending and no one in the
419 * driver that will restart queueing.
421 void bfq_schedule_dispatch(struct bfq_data *bfqd)
423 if (bfqd->queued != 0) {
424 bfq_log(bfqd, "schedule dispatch");
425 blk_mq_run_hw_queues(bfqd->queue, true);
429 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
430 #define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT)
432 #define bfq_sample_valid(samples) ((samples) > 80)
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.
439 static struct request *bfq_choose_req(struct bfq_data *bfqd,
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? */
450 if (!rq1 || rq1 == rq2)
455 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
457 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
459 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
461 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
464 s1 = blk_rq_pos(rq1);
465 s2 = blk_rq_pos(rq2);
468 * By definition, 1KiB is 2 sectors.
470 back_max = bfqd->bfq_back_max * 2;
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.
479 else if (s1 + back_max >= last)
480 d1 = (last - s1) * bfqd->bfq_back_penalty;
482 wrap |= BFQ_RQ1_WRAP;
486 else if (s2 + back_max >= last)
487 d2 = (last - s2) * bfqd->bfq_back_penalty;
489 wrap |= BFQ_RQ2_WRAP;
491 /* Found required data */
494 * By doing switch() on the bit mask "wrap" we avoid having to
495 * check two variables for all permutations: --> faster!
498 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
513 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
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.
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
535 static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
537 struct bfq_data *bfqd = data->q->elevator->elevator_data;
539 if (op_is_sync(op) && !op_is_write(op))
542 data->shallow_depth =
543 bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)];
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);
550 static struct bfq_queue *
551 bfq_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)
555 struct rb_node **p, *parent;
556 struct bfq_queue *bfqq = NULL;
564 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
567 * Sort strictly based on sector. Smallest to the left,
568 * largest to the right.
570 if (sector > blk_rq_pos(bfqq->next_rq))
572 else if (sector < blk_rq_pos(bfqq->next_rq))
580 *ret_parent = parent;
584 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
585 (unsigned long long)sector,
586 bfqq ? bfqq->pid : 0);
591 static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
593 return bfqq->service_from_backlogged > 0 &&
594 time_is_before_jiffies(bfqq->first_IO_time +
595 bfq_merge_time_limit);
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().
607 bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
609 struct rb_node **p, *parent;
610 struct bfq_queue *__bfqq;
612 if (bfqq->pos_root) {
613 rb_erase(&bfqq->pos_node, bfqq->pos_root);
614 bfqq->pos_root = NULL;
618 * bfqq cannot be merged any longer (see comments in
619 * bfq_setup_cooperator): no point in adding bfqq into the
622 if (bfq_too_late_for_merging(bfqq))
625 if (bfq_class_idle(bfqq))
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);
634 rb_link_node(&bfqq->pos_node, parent, p);
635 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
637 bfqq->pos_root = NULL;
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
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
658 * 4) all active groups at the same level in the groups tree have the same
659 * number of children.
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.
673 static bool bfq_asymmetric_scenario(struct bfq_data *bfqd,
674 struct bfq_queue *bfqq)
676 bool smallest_weight = bfqq &&
677 bfqq->weight_counter &&
678 bfqq->weight_counter ==
680 rb_first_cached(&bfqd->queue_weights_tree),
681 struct bfq_weight_counter,
685 * For queue weights to differ, queue_weights_tree must contain
686 * at least two nodes.
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);
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]);
698 return varied_queue_weights || multiple_classes_busy
699 #ifdef CONFIG_BFQ_GROUP_IOSCHED
700 || bfqd->num_groups_with_pending_reqs > 0
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.
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
718 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq,
719 struct rb_root_cached *root)
721 struct bfq_entity *entity = &bfqq->entity;
722 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
723 bool leftmost = true;
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.
737 if (bfqq->weight_counter)
741 struct bfq_weight_counter *__counter = container_of(*new,
742 struct bfq_weight_counter,
746 if (entity->weight == __counter->weight) {
747 bfqq->weight_counter = __counter;
750 if (entity->weight < __counter->weight)
751 new = &((*new)->rb_left);
753 new = &((*new)->rb_right);
758 bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
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.
773 if (unlikely(!bfqq->weight_counter))
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,
782 bfqq->weight_counter->num_active++;
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
792 void __bfq_weights_tree_remove(struct bfq_data *bfqd,
793 struct bfq_queue *bfqq,
794 struct rb_root_cached *root)
796 if (!bfqq->weight_counter)
799 bfqq->weight_counter->num_active--;
800 if (bfqq->weight_counter->num_active > 0)
801 goto reset_entity_pointer;
803 rb_erase_cached(&bfqq->weight_counter->weights_node, root);
804 kfree(bfqq->weight_counter);
806 reset_entity_pointer:
807 bfqq->weight_counter = NULL;
812 * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
813 * of active groups for each queue's inactive parent entity.
815 void bfq_weights_tree_remove(struct bfq_data *bfqd,
816 struct bfq_queue *bfqq)
818 struct bfq_entity *entity = bfqq->entity.parent;
820 for_each_entity(entity) {
821 struct bfq_sched_data *sd = entity->my_sched_data;
823 if (sd->next_in_service || sd->in_service_entity) {
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).
831 * As a consequence, its parent entities are
832 * active as well, and thus this loop must
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.
848 if (entity->in_groups_with_pending_reqs) {
849 entity->in_groups_with_pending_reqs = false;
850 bfqd->num_groups_with_pending_reqs--;
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.
860 __bfq_weights_tree_remove(bfqd, bfqq,
861 &bfqd->queue_weights_tree);
865 * Return expired entry, or NULL to just start from scratch in rbtree.
867 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
868 struct request *last)
872 if (bfq_bfqq_fifo_expire(bfqq))
875 bfq_mark_bfqq_fifo_expire(bfqq);
877 rq = rq_entry_fifo(bfqq->fifo.next);
879 if (rq == last || ktime_get_ns() < rq->fifo_time)
882 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
886 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
887 struct bfq_queue *bfqq,
888 struct request *last)
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;
894 /* Follow expired path, else get first next available. */
895 next = bfq_check_fifo(bfqq, last);
900 prev = rb_entry_rq(rbprev);
903 next = rb_entry_rq(rbnext);
905 rbnext = rb_first(&bfqq->sort_list);
906 if (rbnext && rbnext != &last->rb_node)
907 next = rb_entry_rq(rbnext);
910 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
913 /* see the definition of bfq_async_charge_factor for details */
914 static unsigned long bfq_serv_to_charge(struct request *rq,
915 struct bfq_queue *bfqq)
917 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
918 bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
919 return blk_rq_sectors(rq);
921 return blk_rq_sectors(rq) * bfq_async_charge_factor;
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.
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.
935 static void bfq_updated_next_req(struct bfq_data *bfqd,
936 struct bfq_queue *bfqq)
938 struct bfq_entity *entity = &bfqq->entity;
939 struct request *next_rq = bfqq->next_rq;
940 unsigned long new_budget;
945 if (bfqq == bfqd->in_service_queue)
947 * In order not to break guarantees, budgets cannot be
948 * changed after an entity has been selected.
952 new_budget = max_t(unsigned long,
953 max_t(unsigned long, bfqq->max_budget,
954 bfq_serv_to_charge(next_rq, bfqq)),
956 if (entity->budget != new_budget) {
957 entity->budget = new_budget;
958 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
960 bfq_requeue_bfqq(bfqd, bfqq, false);
964 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
968 if (bfqd->bfq_wr_max_time > 0)
969 return bfqd->bfq_wr_max_time;
971 dur = bfqd->rate_dur_prod;
972 do_div(dur, bfqd->peak_rate);
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
982 * mplayer took 23 seconds to start, if constantly weight-raised.
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.
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.
994 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
997 /* switch back from soft real-time to interactive weight raising */
998 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
999 struct bfq_data *bfqd)
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;
1007 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
1008 struct bfq_io_cq *bic, bool bfq_already_existing)
1010 unsigned int old_wr_coeff = bfqq->wr_coeff;
1011 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
1013 if (bic->saved_has_short_ttime)
1014 bfq_mark_bfqq_has_short_ttime(bfqq);
1016 bfq_clear_bfqq_has_short_ttime(bfqq);
1018 if (bic->saved_IO_bound)
1019 bfq_mark_bfqq_IO_bound(bfqq);
1021 bfq_clear_bfqq_IO_bound(bfqq);
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;
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);
1040 bfq_log_bfqq(bfqq->bfqd, bfqq,
1041 "resume state: switching off wr");
1045 /* make sure weight will be updated, however we got here */
1046 bfqq->entity.prio_changed = 1;
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--;
1057 static int bfqq_process_refs(struct bfq_queue *bfqq)
1059 return bfqq->ref - bfqq->allocated - bfqq->entity.on_st -
1060 (bfqq->weight_counter != NULL);
1063 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1064 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1066 struct bfq_queue *item;
1067 struct hlist_node *n;
1069 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1070 hlist_del_init(&item->burst_list_node);
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().
1077 if (bfq_tot_busy_queues(bfqd) == 0) {
1078 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1079 bfqd->burst_size = 1;
1081 bfqd->burst_size = 0;
1083 bfqd->burst_parent_entity = bfqq->entity.parent;
1086 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1087 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1089 /* Increment burst size to take into account also bfqq */
1092 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1093 struct bfq_queue *pos, *bfqq_item;
1094 struct hlist_node *n;
1097 * Enough queues have been activated shortly after each
1098 * other to consider this burst as large.
1100 bfqd->large_burst = true;
1103 * We can now mark all queues in the burst list as
1104 * belonging to a large burst.
1106 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1108 bfq_mark_bfqq_in_large_burst(bfqq_item);
1109 bfq_mark_bfqq_in_large_burst(bfqq);
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.
1118 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1120 hlist_del_init(&pos->burst_list_node);
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
1127 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
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.
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.
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
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.
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
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.
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.
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)
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
1207 * . if, as a consequence of the previous step, the burst size reaches
1208 * the large-burst threshold, then
1210 * . all the queues in the burst list are marked as belonging to a
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
1218 * . the device enters a large-burst mode
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.
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:
1231 * . the large-burst mode is reset if set
1233 * . the burst list is emptied
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
1239 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
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.
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)))
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
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
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);
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.
1282 if (bfqd->large_burst) {
1283 bfq_mark_bfqq_in_large_burst(bfqq);
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.
1292 bfq_add_to_burst(bfqd, bfqq);
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
1302 bfqd->last_ins_in_burst = jiffies;
1305 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1307 struct bfq_entity *entity = &bfqq->entity;
1309 return entity->budget - entity->service;
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
1317 static int bfq_max_budget(struct bfq_data *bfqd)
1319 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1320 return bfq_default_max_budget;
1322 return bfqd->bfq_max_budget;
1326 * Return min budget, which is a fraction of the current or default
1327 * max budget (trying with 1/32)
1329 static int bfq_min_budget(struct bfq_data *bfqd)
1331 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1332 return bfq_default_max_budget / 32;
1334 return bfqd->bfq_max_budget / 32;
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
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:
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;
1354 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1355 * a new request before the expiration of the idling-time.
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
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).
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.
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.
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.
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, and thus loss of throughput. Because of these facts, the next
1432 * function adopts the following simple scheme to avoid both costly
1433 * operations and too frequent preemptions: it requests the expiration
1434 * of the in-service queue (unconditionally) only for queues that need
1435 * to recover a hole, or that either are weight-raised or deserve to
1438 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1439 struct bfq_queue *bfqq,
1440 bool arrived_in_time,
1441 bool wr_or_deserves_wr)
1443 struct bfq_entity *entity = &bfqq->entity;
1446 * In the next compound condition, we check also whether there
1447 * is some budget left, because otherwise there is no point in
1448 * trying to go on serving bfqq with this same budget: bfqq
1449 * would be expired immediately after being selected for
1450 * service. This would only cause useless overhead.
1452 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1453 bfq_bfqq_budget_left(bfqq) > 0) {
1455 * We do not clear the flag non_blocking_wait_rq here, as
1456 * the latter is used in bfq_activate_bfqq to signal
1457 * that timestamps need to be back-shifted (and is
1458 * cleared right after).
1462 * In next assignment we rely on that either
1463 * entity->service or entity->budget are not updated
1464 * on expiration if bfqq is empty (see
1465 * __bfq_bfqq_recalc_budget). Thus both quantities
1466 * remain unchanged after such an expiration, and the
1467 * following statement therefore assigns to
1468 * entity->budget the remaining budget on such an
1471 entity->budget = min_t(unsigned long,
1472 bfq_bfqq_budget_left(bfqq),
1476 * At this point, we have used entity->service to get
1477 * the budget left (needed for updating
1478 * entity->budget). Thus we finally can, and have to,
1479 * reset entity->service. The latter must be reset
1480 * because bfqq would otherwise be charged again for
1481 * the service it has received during its previous
1484 entity->service = 0;
1490 * We can finally complete expiration, by setting service to 0.
1492 entity->service = 0;
1493 entity->budget = max_t(unsigned long, bfqq->max_budget,
1494 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1495 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1496 return wr_or_deserves_wr;
1500 * Return the farthest past time instant according to jiffies
1503 static unsigned long bfq_smallest_from_now(void)
1505 return jiffies - MAX_JIFFY_OFFSET;
1508 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1509 struct bfq_queue *bfqq,
1510 unsigned int old_wr_coeff,
1511 bool wr_or_deserves_wr,
1516 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1517 /* start a weight-raising period */
1519 bfqq->service_from_wr = 0;
1520 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1521 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1524 * No interactive weight raising in progress
1525 * here: assign minus infinity to
1526 * wr_start_at_switch_to_srt, to make sure
1527 * that, at the end of the soft-real-time
1528 * weight raising periods that is starting
1529 * now, no interactive weight-raising period
1530 * may be wrongly considered as still in
1531 * progress (and thus actually started by
1534 bfqq->wr_start_at_switch_to_srt =
1535 bfq_smallest_from_now();
1536 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1537 BFQ_SOFTRT_WEIGHT_FACTOR;
1538 bfqq->wr_cur_max_time =
1539 bfqd->bfq_wr_rt_max_time;
1543 * If needed, further reduce budget to make sure it is
1544 * close to bfqq's backlog, so as to reduce the
1545 * scheduling-error component due to a too large
1546 * budget. Do not care about throughput consequences,
1547 * but only about latency. Finally, do not assign a
1548 * too small budget either, to avoid increasing
1549 * latency by causing too frequent expirations.
1551 bfqq->entity.budget = min_t(unsigned long,
1552 bfqq->entity.budget,
1553 2 * bfq_min_budget(bfqd));
1554 } else if (old_wr_coeff > 1) {
1555 if (interactive) { /* update wr coeff and duration */
1556 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1557 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1558 } else if (in_burst)
1562 * The application is now or still meeting the
1563 * requirements for being deemed soft rt. We
1564 * can then correctly and safely (re)charge
1565 * the weight-raising duration for the
1566 * application with the weight-raising
1567 * duration for soft rt applications.
1569 * In particular, doing this recharge now, i.e.,
1570 * before the weight-raising period for the
1571 * application finishes, reduces the probability
1572 * of the following negative scenario:
1573 * 1) the weight of a soft rt application is
1574 * raised at startup (as for any newly
1575 * created application),
1576 * 2) since the application is not interactive,
1577 * at a certain time weight-raising is
1578 * stopped for the application,
1579 * 3) at that time the application happens to
1580 * still have pending requests, and hence
1581 * is destined to not have a chance to be
1582 * deemed soft rt before these requests are
1583 * completed (see the comments to the
1584 * function bfq_bfqq_softrt_next_start()
1585 * for details on soft rt detection),
1586 * 4) these pending requests experience a high
1587 * latency because the application is not
1588 * weight-raised while they are pending.
1590 if (bfqq->wr_cur_max_time !=
1591 bfqd->bfq_wr_rt_max_time) {
1592 bfqq->wr_start_at_switch_to_srt =
1593 bfqq->last_wr_start_finish;
1595 bfqq->wr_cur_max_time =
1596 bfqd->bfq_wr_rt_max_time;
1597 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1598 BFQ_SOFTRT_WEIGHT_FACTOR;
1600 bfqq->last_wr_start_finish = jiffies;
1605 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1606 struct bfq_queue *bfqq)
1608 return bfqq->dispatched == 0 &&
1609 time_is_before_jiffies(
1610 bfqq->budget_timeout +
1611 bfqd->bfq_wr_min_idle_time);
1614 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1615 struct bfq_queue *bfqq,
1620 bool soft_rt, in_burst, wr_or_deserves_wr,
1621 bfqq_wants_to_preempt,
1622 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1624 * See the comments on
1625 * bfq_bfqq_update_budg_for_activation for
1626 * details on the usage of the next variable.
1628 arrived_in_time = ktime_get_ns() <=
1629 bfqq->ttime.last_end_request +
1630 bfqd->bfq_slice_idle * 3;
1634 * bfqq deserves to be weight-raised if:
1636 * - it does not belong to a large burst,
1637 * - it has been idle for enough time or is soft real-time,
1638 * - is linked to a bfq_io_cq (it is not shared in any sense).
1640 in_burst = bfq_bfqq_in_large_burst(bfqq);
1641 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1642 !BFQQ_TOTALLY_SEEKY(bfqq) &&
1644 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1645 bfqq->dispatched == 0;
1646 *interactive = !in_burst && idle_for_long_time;
1647 wr_or_deserves_wr = bfqd->low_latency &&
1648 (bfqq->wr_coeff > 1 ||
1649 (bfq_bfqq_sync(bfqq) &&
1650 bfqq->bic && (*interactive || soft_rt)));
1653 * Using the last flag, update budget and check whether bfqq
1654 * may want to preempt the in-service queue.
1656 bfqq_wants_to_preempt =
1657 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1662 * If bfqq happened to be activated in a burst, but has been
1663 * idle for much more than an interactive queue, then we
1664 * assume that, in the overall I/O initiated in the burst, the
1665 * I/O associated with bfqq is finished. So bfqq does not need
1666 * to be treated as a queue belonging to a burst
1667 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1668 * if set, and remove bfqq from the burst list if it's
1669 * there. We do not decrement burst_size, because the fact
1670 * that bfqq does not need to belong to the burst list any
1671 * more does not invalidate the fact that bfqq was created in
1674 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1675 idle_for_long_time &&
1676 time_is_before_jiffies(
1677 bfqq->budget_timeout +
1678 msecs_to_jiffies(10000))) {
1679 hlist_del_init(&bfqq->burst_list_node);
1680 bfq_clear_bfqq_in_large_burst(bfqq);
1683 bfq_clear_bfqq_just_created(bfqq);
1686 if (!bfq_bfqq_IO_bound(bfqq)) {
1687 if (arrived_in_time) {
1688 bfqq->requests_within_timer++;
1689 if (bfqq->requests_within_timer >=
1690 bfqd->bfq_requests_within_timer)
1691 bfq_mark_bfqq_IO_bound(bfqq);
1693 bfqq->requests_within_timer = 0;
1696 if (bfqd->low_latency) {
1697 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1700 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1702 if (time_is_before_jiffies(bfqq->split_time +
1703 bfqd->bfq_wr_min_idle_time)) {
1704 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1711 if (old_wr_coeff != bfqq->wr_coeff)
1712 bfqq->entity.prio_changed = 1;
1716 bfqq->last_idle_bklogged = jiffies;
1717 bfqq->service_from_backlogged = 0;
1718 bfq_clear_bfqq_softrt_update(bfqq);
1720 bfq_add_bfqq_busy(bfqd, bfqq);
1723 * Expire in-service queue only if preemption may be needed
1724 * for guarantees. In this respect, the function
1725 * next_queue_may_preempt just checks a simple, necessary
1726 * condition, and not a sufficient condition based on
1727 * timestamps. In fact, for the latter condition to be
1728 * evaluated, timestamps would need first to be updated, and
1729 * this operation is quite costly (see the comments on the
1730 * function bfq_bfqq_update_budg_for_activation).
1732 if (bfqd->in_service_queue && bfqq_wants_to_preempt &&
1733 bfqd->in_service_queue->wr_coeff < bfqq->wr_coeff &&
1734 next_queue_may_preempt(bfqd))
1735 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1736 false, BFQQE_PREEMPTED);
1739 static void bfq_reset_inject_limit(struct bfq_data *bfqd,
1740 struct bfq_queue *bfqq)
1742 /* invalidate baseline total service time */
1743 bfqq->last_serv_time_ns = 0;
1746 * Reset pointer in case we are waiting for
1747 * some request completion.
1749 bfqd->waited_rq = NULL;
1752 * If bfqq has a short think time, then start by setting the
1753 * inject limit to 0 prudentially, because the service time of
1754 * an injected I/O request may be higher than the think time
1755 * of bfqq, and therefore, if one request was injected when
1756 * bfqq remains empty, this injected request might delay the
1757 * service of the next I/O request for bfqq significantly. In
1758 * case bfqq can actually tolerate some injection, then the
1759 * adaptive update will however raise the limit soon. This
1760 * lucky circumstance holds exactly because bfqq has a short
1761 * think time, and thus, after remaining empty, is likely to
1762 * get new I/O enqueued---and then completed---before being
1763 * expired. This is the very pattern that gives the
1764 * limit-update algorithm the chance to measure the effect of
1765 * injection on request service times, and then to update the
1766 * limit accordingly.
1768 * However, in the following special case, the inject limit is
1769 * left to 1 even if the think time is short: bfqq's I/O is
1770 * synchronized with that of some other queue, i.e., bfqq may
1771 * receive new I/O only after the I/O of the other queue is
1772 * completed. Keeping the inject limit to 1 allows the
1773 * blocking I/O to be served while bfqq is in service. And
1774 * this is very convenient both for bfqq and for overall
1775 * throughput, as explained in detail in the comments in
1776 * bfq_update_has_short_ttime().
1778 * On the opposite end, if bfqq has a long think time, then
1779 * start directly by 1, because:
1780 * a) on the bright side, keeping at most one request in
1781 * service in the drive is unlikely to cause any harm to the
1782 * latency of bfqq's requests, as the service time of a single
1783 * request is likely to be lower than the think time of bfqq;
1784 * b) on the downside, after becoming empty, bfqq is likely to
1785 * expire before getting its next request. With this request
1786 * arrival pattern, it is very hard to sample total service
1787 * times and update the inject limit accordingly (see comments
1788 * on bfq_update_inject_limit()). So the limit is likely to be
1789 * never, or at least seldom, updated. As a consequence, by
1790 * setting the limit to 1, we avoid that no injection ever
1791 * occurs with bfqq. On the downside, this proactive step
1792 * further reduces chances to actually compute the baseline
1793 * total service time. Thus it reduces chances to execute the
1794 * limit-update algorithm and possibly raise the limit to more
1797 if (bfq_bfqq_has_short_ttime(bfqq))
1798 bfqq->inject_limit = 0;
1800 bfqq->inject_limit = 1;
1802 bfqq->decrease_time_jif = jiffies;
1805 static void bfq_add_request(struct request *rq)
1807 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1808 struct bfq_data *bfqd = bfqq->bfqd;
1809 struct request *next_rq, *prev;
1810 unsigned int old_wr_coeff = bfqq->wr_coeff;
1811 bool interactive = false;
1813 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1814 bfqq->queued[rq_is_sync(rq)]++;
1817 if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_bfqq_sync(bfqq)) {
1819 * Detect whether bfqq's I/O seems synchronized with
1820 * that of some other queue, i.e., whether bfqq, after
1821 * remaining empty, happens to receive new I/O only
1822 * right after some I/O request of the other queue has
1823 * been completed. We call waker queue the other
1824 * queue, and we assume, for simplicity, that bfqq may
1825 * have at most one waker queue.
1827 * A remarkable throughput boost can be reached by
1828 * unconditionally injecting the I/O of the waker
1829 * queue, every time a new bfq_dispatch_request
1830 * happens to be invoked while I/O is being plugged
1831 * for bfqq. In addition to boosting throughput, this
1832 * unblocks bfqq's I/O, thereby improving bandwidth
1833 * and latency for bfqq. Note that these same results
1834 * may be achieved with the general injection
1835 * mechanism, but less effectively. For details on
1836 * this aspect, see the comments on the choice of the
1837 * queue for injection in bfq_select_queue().
1839 * Turning back to the detection of a waker queue, a
1840 * queue Q is deemed as a waker queue for bfqq if, for
1841 * two consecutive times, bfqq happens to become non
1842 * empty right after a request of Q has been
1843 * completed. In particular, on the first time, Q is
1844 * tentatively set as a candidate waker queue, while
1845 * on the second time, the flag
1846 * bfq_bfqq_has_waker(bfqq) is set to confirm that Q
1847 * is a waker queue for bfqq. These detection steps
1848 * are performed only if bfqq has a long think time,
1849 * so as to make it more likely that bfqq's I/O is
1850 * actually being blocked by a synchronization. This
1851 * last filter, plus the above two-times requirement,
1852 * make false positives less likely.
1856 * The sooner a waker queue is detected, the sooner
1857 * throughput can be boosted by injecting I/O from the
1858 * waker queue. Fortunately, detection is likely to be
1859 * actually fast, for the following reasons. While
1860 * blocked by synchronization, bfqq has a long think
1861 * time. This implies that bfqq's inject limit is at
1862 * least equal to 1 (see the comments in
1863 * bfq_update_inject_limit()). So, thanks to
1864 * injection, the waker queue is likely to be served
1865 * during the very first I/O-plugging time interval
1866 * for bfqq. This triggers the first step of the
1867 * detection mechanism. Thanks again to injection, the
1868 * candidate waker queue is then likely to be
1869 * confirmed no later than during the next
1870 * I/O-plugging interval for bfqq.
1872 if (!bfq_bfqq_has_short_ttime(bfqq) &&
1873 ktime_get_ns() - bfqd->last_completion <
1874 200 * NSEC_PER_USEC) {
1875 if (bfqd->last_completed_rq_bfqq != bfqq &&
1876 bfqd->last_completed_rq_bfqq !=
1879 * First synchronization detected with
1880 * a candidate waker queue, or with a
1881 * different candidate waker queue
1882 * from the current one.
1884 bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
1887 * If the waker queue disappears, then
1888 * bfqq->waker_bfqq must be reset. To
1889 * this goal, we maintain in each
1890 * waker queue a list, woken_list, of
1891 * all the queues that reference the
1892 * waker queue through their
1893 * waker_bfqq pointer. When the waker
1894 * queue exits, the waker_bfqq pointer
1895 * of all the queues in the woken_list
1898 * In addition, if bfqq is already in
1899 * the woken_list of a waker queue,
1900 * then, before being inserted into
1901 * the woken_list of a new waker
1902 * queue, bfqq must be removed from
1903 * the woken_list of the old waker
1906 if (!hlist_unhashed(&bfqq->woken_list_node))
1907 hlist_del_init(&bfqq->woken_list_node);
1908 hlist_add_head(&bfqq->woken_list_node,
1909 &bfqd->last_completed_rq_bfqq->woken_list);
1911 bfq_clear_bfqq_has_waker(bfqq);
1912 } else if (bfqd->last_completed_rq_bfqq ==
1914 !bfq_bfqq_has_waker(bfqq)) {
1916 * synchronization with waker_bfqq
1917 * seen for the second time
1919 bfq_mark_bfqq_has_waker(bfqq);
1924 * Periodically reset inject limit, to make sure that
1925 * the latter eventually drops in case workload
1926 * changes, see step (3) in the comments on
1927 * bfq_update_inject_limit().
1929 if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
1930 msecs_to_jiffies(1000)))
1931 bfq_reset_inject_limit(bfqd, bfqq);
1934 * The following conditions must hold to setup a new
1935 * sampling of total service time, and then a new
1936 * update of the inject limit:
1937 * - bfqq is in service, because the total service
1938 * time is evaluated only for the I/O requests of
1939 * the queues in service;
1940 * - this is the right occasion to compute or to
1941 * lower the baseline total service time, because
1942 * there are actually no requests in the drive,
1944 * the baseline total service time is available, and
1945 * this is the right occasion to compute the other
1946 * quantity needed to update the inject limit, i.e.,
1947 * the total service time caused by the amount of
1948 * injection allowed by the current value of the
1949 * limit. It is the right occasion because injection
1950 * has actually been performed during the service
1951 * hole, and there are still in-flight requests,
1952 * which are very likely to be exactly the injected
1953 * requests, or part of them;
1954 * - the minimum interval for sampling the total
1955 * service time and updating the inject limit has
1958 if (bfqq == bfqd->in_service_queue &&
1959 (bfqd->rq_in_driver == 0 ||
1960 (bfqq->last_serv_time_ns > 0 &&
1961 bfqd->rqs_injected && bfqd->rq_in_driver > 0)) &&
1962 time_is_before_eq_jiffies(bfqq->decrease_time_jif +
1963 msecs_to_jiffies(100))) {
1964 bfqd->last_empty_occupied_ns = ktime_get_ns();
1966 * Start the state machine for measuring the
1967 * total service time of rq: setting
1968 * wait_dispatch will cause bfqd->waited_rq to
1969 * be set when rq will be dispatched.
1971 bfqd->wait_dispatch = true;
1972 bfqd->rqs_injected = false;
1976 elv_rb_add(&bfqq->sort_list, rq);
1979 * Check if this request is a better next-serve candidate.
1981 prev = bfqq->next_rq;
1982 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
1983 bfqq->next_rq = next_rq;
1986 * Adjust priority tree position, if next_rq changes.
1987 * See comments on bfq_pos_tree_add_move() for the unlikely().
1989 if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
1990 bfq_pos_tree_add_move(bfqd, bfqq);
1992 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
1993 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
1996 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
1997 time_is_before_jiffies(
1998 bfqq->last_wr_start_finish +
1999 bfqd->bfq_wr_min_inter_arr_async)) {
2000 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
2001 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
2003 bfqd->wr_busy_queues++;
2004 bfqq->entity.prio_changed = 1;
2006 if (prev != bfqq->next_rq)
2007 bfq_updated_next_req(bfqd, bfqq);
2011 * Assign jiffies to last_wr_start_finish in the following
2014 * . if bfqq is not going to be weight-raised, because, for
2015 * non weight-raised queues, last_wr_start_finish stores the
2016 * arrival time of the last request; as of now, this piece
2017 * of information is used only for deciding whether to
2018 * weight-raise async queues
2020 * . if bfqq is not weight-raised, because, if bfqq is now
2021 * switching to weight-raised, then last_wr_start_finish
2022 * stores the time when weight-raising starts
2024 * . if bfqq is interactive, because, regardless of whether
2025 * bfqq is currently weight-raised, the weight-raising
2026 * period must start or restart (this case is considered
2027 * separately because it is not detected by the above
2028 * conditions, if bfqq is already weight-raised)
2030 * last_wr_start_finish has to be updated also if bfqq is soft
2031 * real-time, because the weight-raising period is constantly
2032 * restarted on idle-to-busy transitions for these queues, but
2033 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2036 if (bfqd->low_latency &&
2037 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
2038 bfqq->last_wr_start_finish = jiffies;
2041 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
2043 struct request_queue *q)
2045 struct bfq_queue *bfqq = bfqd->bio_bfqq;
2049 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
2054 static sector_t get_sdist(sector_t last_pos, struct request *rq)
2057 return abs(blk_rq_pos(rq) - last_pos);
2062 #if 0 /* Still not clear if we can do without next two functions */
2063 static void bfq_activate_request(struct request_queue *q, struct request *rq)
2065 struct bfq_data *bfqd = q->elevator->elevator_data;
2067 bfqd->rq_in_driver++;
2070 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
2072 struct bfq_data *bfqd = q->elevator->elevator_data;
2074 bfqd->rq_in_driver--;
2078 static void bfq_remove_request(struct request_queue *q,
2081 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2082 struct bfq_data *bfqd = bfqq->bfqd;
2083 const int sync = rq_is_sync(rq);
2085 if (bfqq->next_rq == rq) {
2086 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
2087 bfq_updated_next_req(bfqd, bfqq);
2090 if (rq->queuelist.prev != &rq->queuelist)
2091 list_del_init(&rq->queuelist);
2092 bfqq->queued[sync]--;
2094 elv_rb_del(&bfqq->sort_list, rq);
2096 elv_rqhash_del(q, rq);
2097 if (q->last_merge == rq)
2098 q->last_merge = NULL;
2100 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2101 bfqq->next_rq = NULL;
2103 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
2104 bfq_del_bfqq_busy(bfqd, bfqq, false);
2106 * bfqq emptied. In normal operation, when
2107 * bfqq is empty, bfqq->entity.service and
2108 * bfqq->entity.budget must contain,
2109 * respectively, the service received and the
2110 * budget used last time bfqq emptied. These
2111 * facts do not hold in this case, as at least
2112 * this last removal occurred while bfqq is
2113 * not in service. To avoid inconsistencies,
2114 * reset both bfqq->entity.service and
2115 * bfqq->entity.budget, if bfqq has still a
2116 * process that may issue I/O requests to it.
2118 bfqq->entity.budget = bfqq->entity.service = 0;
2122 * Remove queue from request-position tree as it is empty.
2124 if (bfqq->pos_root) {
2125 rb_erase(&bfqq->pos_node, bfqq->pos_root);
2126 bfqq->pos_root = NULL;
2129 /* see comments on bfq_pos_tree_add_move() for the unlikely() */
2130 if (unlikely(!bfqd->nonrot_with_queueing))
2131 bfq_pos_tree_add_move(bfqd, bfqq);
2134 if (rq->cmd_flags & REQ_META)
2135 bfqq->meta_pending--;
2139 static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio,
2140 unsigned int nr_segs)
2142 struct request_queue *q = hctx->queue;
2143 struct bfq_data *bfqd = q->elevator->elevator_data;
2144 struct request *free = NULL;
2146 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2147 * store its return value for later use, to avoid nesting
2148 * queue_lock inside the bfqd->lock. We assume that the bic
2149 * returned by bfq_bic_lookup does not go away before
2150 * bfqd->lock is taken.
2152 struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
2155 spin_lock_irq(&bfqd->lock);
2158 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
2160 bfqd->bio_bfqq = NULL;
2161 bfqd->bio_bic = bic;
2163 ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
2166 blk_mq_free_request(free);
2167 spin_unlock_irq(&bfqd->lock);
2172 static int bfq_request_merge(struct request_queue *q, struct request **req,
2175 struct bfq_data *bfqd = q->elevator->elevator_data;
2176 struct request *__rq;
2178 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
2179 if (__rq && elv_bio_merge_ok(__rq, bio)) {
2181 return ELEVATOR_FRONT_MERGE;
2184 return ELEVATOR_NO_MERGE;
2187 static struct bfq_queue *bfq_init_rq(struct request *rq);
2189 static void bfq_request_merged(struct request_queue *q, struct request *req,
2190 enum elv_merge type)
2192 if (type == ELEVATOR_FRONT_MERGE &&
2193 rb_prev(&req->rb_node) &&
2195 blk_rq_pos(container_of(rb_prev(&req->rb_node),
2196 struct request, rb_node))) {
2197 struct bfq_queue *bfqq = bfq_init_rq(req);
2198 struct bfq_data *bfqd = bfqq->bfqd;
2199 struct request *prev, *next_rq;
2201 /* Reposition request in its sort_list */
2202 elv_rb_del(&bfqq->sort_list, req);
2203 elv_rb_add(&bfqq->sort_list, req);
2205 /* Choose next request to be served for bfqq */
2206 prev = bfqq->next_rq;
2207 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
2208 bfqd->last_position);
2209 bfqq->next_rq = next_rq;
2211 * If next_rq changes, update both the queue's budget to
2212 * fit the new request and the queue's position in its
2215 if (prev != bfqq->next_rq) {
2216 bfq_updated_next_req(bfqd, bfqq);
2218 * See comments on bfq_pos_tree_add_move() for
2221 if (unlikely(!bfqd->nonrot_with_queueing))
2222 bfq_pos_tree_add_move(bfqd, bfqq);
2228 * This function is called to notify the scheduler that the requests
2229 * rq and 'next' have been merged, with 'next' going away. BFQ
2230 * exploits this hook to address the following issue: if 'next' has a
2231 * fifo_time lower that rq, then the fifo_time of rq must be set to
2232 * the value of 'next', to not forget the greater age of 'next'.
2234 * NOTE: in this function we assume that rq is in a bfq_queue, basing
2235 * on that rq is picked from the hash table q->elevator->hash, which,
2236 * in its turn, is filled only with I/O requests present in
2237 * bfq_queues, while BFQ is in use for the request queue q. In fact,
2238 * the function that fills this hash table (elv_rqhash_add) is called
2239 * only by bfq_insert_request.
2241 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
2242 struct request *next)
2244 struct bfq_queue *bfqq = bfq_init_rq(rq),
2245 *next_bfqq = bfq_init_rq(next);
2248 * If next and rq belong to the same bfq_queue and next is older
2249 * than rq, then reposition rq in the fifo (by substituting next
2250 * with rq). Otherwise, if next and rq belong to different
2251 * bfq_queues, never reposition rq: in fact, we would have to
2252 * reposition it with respect to next's position in its own fifo,
2253 * which would most certainly be too expensive with respect to
2256 if (bfqq == next_bfqq &&
2257 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2258 next->fifo_time < rq->fifo_time) {
2259 list_del_init(&rq->queuelist);
2260 list_replace_init(&next->queuelist, &rq->queuelist);
2261 rq->fifo_time = next->fifo_time;
2264 if (bfqq->next_rq == next)
2267 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
2270 /* Must be called with bfqq != NULL */
2271 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
2273 if (bfq_bfqq_busy(bfqq))
2274 bfqq->bfqd->wr_busy_queues--;
2276 bfqq->wr_cur_max_time = 0;
2277 bfqq->last_wr_start_finish = jiffies;
2279 * Trigger a weight change on the next invocation of
2280 * __bfq_entity_update_weight_prio.
2282 bfqq->entity.prio_changed = 1;
2285 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2286 struct bfq_group *bfqg)
2290 for (i = 0; i < 2; i++)
2291 for (j = 0; j < IOPRIO_BE_NR; j++)
2292 if (bfqg->async_bfqq[i][j])
2293 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2294 if (bfqg->async_idle_bfqq)
2295 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2298 static void bfq_end_wr(struct bfq_data *bfqd)
2300 struct bfq_queue *bfqq;
2302 spin_lock_irq(&bfqd->lock);
2304 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2305 bfq_bfqq_end_wr(bfqq);
2306 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2307 bfq_bfqq_end_wr(bfqq);
2308 bfq_end_wr_async(bfqd);
2310 spin_unlock_irq(&bfqd->lock);
2313 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2316 return blk_rq_pos(io_struct);
2318 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2321 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2324 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2328 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2329 struct bfq_queue *bfqq,
2332 struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
2333 struct rb_node *parent, *node;
2334 struct bfq_queue *__bfqq;
2336 if (RB_EMPTY_ROOT(root))
2340 * First, if we find a request starting at the end of the last
2341 * request, choose it.
2343 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2348 * If the exact sector wasn't found, the parent of the NULL leaf
2349 * will contain the closest sector (rq_pos_tree sorted by
2350 * next_request position).
2352 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2353 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2356 if (blk_rq_pos(__bfqq->next_rq) < sector)
2357 node = rb_next(&__bfqq->pos_node);
2359 node = rb_prev(&__bfqq->pos_node);
2363 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2364 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2370 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2371 struct bfq_queue *cur_bfqq,
2374 struct bfq_queue *bfqq;
2377 * We shall notice if some of the queues are cooperating,
2378 * e.g., working closely on the same area of the device. In
2379 * that case, we can group them together and: 1) don't waste
2380 * time idling, and 2) serve the union of their requests in
2381 * the best possible order for throughput.
2383 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2384 if (!bfqq || bfqq == cur_bfqq)
2390 static struct bfq_queue *
2391 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2393 int process_refs, new_process_refs;
2394 struct bfq_queue *__bfqq;
2397 * If there are no process references on the new_bfqq, then it is
2398 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2399 * may have dropped their last reference (not just their last process
2402 if (!bfqq_process_refs(new_bfqq))
2405 /* Avoid a circular list and skip interim queue merges. */
2406 while ((__bfqq = new_bfqq->new_bfqq)) {
2412 process_refs = bfqq_process_refs(bfqq);
2413 new_process_refs = bfqq_process_refs(new_bfqq);
2415 * If the process for the bfqq has gone away, there is no
2416 * sense in merging the queues.
2418 if (process_refs == 0 || new_process_refs == 0)
2421 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2425 * Merging is just a redirection: the requests of the process
2426 * owning one of the two queues are redirected to the other queue.
2427 * The latter queue, in its turn, is set as shared if this is the
2428 * first time that the requests of some process are redirected to
2431 * We redirect bfqq to new_bfqq and not the opposite, because
2432 * we are in the context of the process owning bfqq, thus we
2433 * have the io_cq of this process. So we can immediately
2434 * configure this io_cq to redirect the requests of the
2435 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2436 * not available any more (new_bfqq->bic == NULL).
2438 * Anyway, even in case new_bfqq coincides with the in-service
2439 * queue, redirecting requests the in-service queue is the
2440 * best option, as we feed the in-service queue with new
2441 * requests close to the last request served and, by doing so,
2442 * are likely to increase the throughput.
2444 bfqq->new_bfqq = new_bfqq;
2445 new_bfqq->ref += process_refs;
2449 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2450 struct bfq_queue *new_bfqq)
2452 if (bfq_too_late_for_merging(new_bfqq))
2455 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2456 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2460 * If either of the queues has already been detected as seeky,
2461 * then merging it with the other queue is unlikely to lead to
2464 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2468 * Interleaved I/O is known to be done by (some) applications
2469 * only for reads, so it does not make sense to merge async
2472 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2479 * Attempt to schedule a merge of bfqq with the currently in-service
2480 * queue or with a close queue among the scheduled queues. Return
2481 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2482 * structure otherwise.
2484 * The OOM queue is not allowed to participate to cooperation: in fact, since
2485 * the requests temporarily redirected to the OOM queue could be redirected
2486 * again to dedicated queues at any time, the state needed to correctly
2487 * handle merging with the OOM queue would be quite complex and expensive
2488 * to maintain. Besides, in such a critical condition as an out of memory,
2489 * the benefits of queue merging may be little relevant, or even negligible.
2491 * WARNING: queue merging may impair fairness among non-weight raised
2492 * queues, for at least two reasons: 1) the original weight of a
2493 * merged queue may change during the merged state, 2) even being the
2494 * weight the same, a merged queue may be bloated with many more
2495 * requests than the ones produced by its originally-associated
2498 static struct bfq_queue *
2499 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2500 void *io_struct, bool request)
2502 struct bfq_queue *in_service_bfqq, *new_bfqq;
2505 * Do not perform queue merging if the device is non
2506 * rotational and performs internal queueing. In fact, such a
2507 * device reaches a high speed through internal parallelism
2508 * and pipelining. This means that, to reach a high
2509 * throughput, it must have many requests enqueued at the same
2510 * time. But, in this configuration, the internal scheduling
2511 * algorithm of the device does exactly the job of queue
2512 * merging: it reorders requests so as to obtain as much as
2513 * possible a sequential I/O pattern. As a consequence, with
2514 * the workload generated by processes doing interleaved I/O,
2515 * the throughput reached by the device is likely to be the
2516 * same, with and without queue merging.
2518 * Disabling merging also provides a remarkable benefit in
2519 * terms of throughput. Merging tends to make many workloads
2520 * artificially more uneven, because of shared queues
2521 * remaining non empty for incomparably more time than
2522 * non-merged queues. This may accentuate workload
2523 * asymmetries. For example, if one of the queues in a set of
2524 * merged queues has a higher weight than a normal queue, then
2525 * the shared queue may inherit such a high weight and, by
2526 * staying almost always active, may force BFQ to perform I/O
2527 * plugging most of the time. This evidently makes it harder
2528 * for BFQ to let the device reach a high throughput.
2530 * Finally, the likely() macro below is not used because one
2531 * of the two branches is more likely than the other, but to
2532 * have the code path after the following if() executed as
2533 * fast as possible for the case of a non rotational device
2534 * with queueing. We want it because this is the fastest kind
2535 * of device. On the opposite end, the likely() may lengthen
2536 * the execution time of BFQ for the case of slower devices
2537 * (rotational or at least without queueing). But in this case
2538 * the execution time of BFQ matters very little, if not at
2541 if (likely(bfqd->nonrot_with_queueing))
2545 * Prevent bfqq from being merged if it has been created too
2546 * long ago. The idea is that true cooperating processes, and
2547 * thus their associated bfq_queues, are supposed to be
2548 * created shortly after each other. This is the case, e.g.,
2549 * for KVM/QEMU and dump I/O threads. Basing on this
2550 * assumption, the following filtering greatly reduces the
2551 * probability that two non-cooperating processes, which just
2552 * happen to do close I/O for some short time interval, have
2553 * their queues merged by mistake.
2555 if (bfq_too_late_for_merging(bfqq))
2559 return bfqq->new_bfqq;
2561 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2564 /* If there is only one backlogged queue, don't search. */
2565 if (bfq_tot_busy_queues(bfqd) == 1)
2568 in_service_bfqq = bfqd->in_service_queue;
2570 if (in_service_bfqq && in_service_bfqq != bfqq &&
2571 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2572 bfq_rq_close_to_sector(io_struct, request,
2573 bfqd->in_serv_last_pos) &&
2574 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2575 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2576 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2581 * Check whether there is a cooperator among currently scheduled
2582 * queues. The only thing we need is that the bio/request is not
2583 * NULL, as we need it to establish whether a cooperator exists.
2585 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2586 bfq_io_struct_pos(io_struct, request));
2588 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2589 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2590 return bfq_setup_merge(bfqq, new_bfqq);
2595 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2597 struct bfq_io_cq *bic = bfqq->bic;
2600 * If !bfqq->bic, the queue is already shared or its requests
2601 * have already been redirected to a shared queue; both idle window
2602 * and weight raising state have already been saved. Do nothing.
2607 bic->saved_weight = bfqq->entity.orig_weight;
2608 bic->saved_ttime = bfqq->ttime;
2609 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2610 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2611 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2612 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2613 if (unlikely(bfq_bfqq_just_created(bfqq) &&
2614 !bfq_bfqq_in_large_burst(bfqq) &&
2615 bfqq->bfqd->low_latency)) {
2617 * bfqq being merged right after being created: bfqq
2618 * would have deserved interactive weight raising, but
2619 * did not make it to be set in a weight-raised state,
2620 * because of this early merge. Store directly the
2621 * weight-raising state that would have been assigned
2622 * to bfqq, so that to avoid that bfqq unjustly fails
2623 * to enjoy weight raising if split soon.
2625 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
2626 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2627 bic->saved_last_wr_start_finish = jiffies;
2629 bic->saved_wr_coeff = bfqq->wr_coeff;
2630 bic->saved_wr_start_at_switch_to_srt =
2631 bfqq->wr_start_at_switch_to_srt;
2632 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2633 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2638 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2639 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2641 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2642 (unsigned long)new_bfqq->pid);
2643 /* Save weight raising and idle window of the merged queues */
2644 bfq_bfqq_save_state(bfqq);
2645 bfq_bfqq_save_state(new_bfqq);
2646 if (bfq_bfqq_IO_bound(bfqq))
2647 bfq_mark_bfqq_IO_bound(new_bfqq);
2648 bfq_clear_bfqq_IO_bound(bfqq);
2651 * If bfqq is weight-raised, then let new_bfqq inherit
2652 * weight-raising. To reduce false positives, neglect the case
2653 * where bfqq has just been created, but has not yet made it
2654 * to be weight-raised (which may happen because EQM may merge
2655 * bfqq even before bfq_add_request is executed for the first
2656 * time for bfqq). Handling this case would however be very
2657 * easy, thanks to the flag just_created.
2659 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2660 new_bfqq->wr_coeff = bfqq->wr_coeff;
2661 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2662 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2663 new_bfqq->wr_start_at_switch_to_srt =
2664 bfqq->wr_start_at_switch_to_srt;
2665 if (bfq_bfqq_busy(new_bfqq))
2666 bfqd->wr_busy_queues++;
2667 new_bfqq->entity.prio_changed = 1;
2670 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2672 bfqq->entity.prio_changed = 1;
2673 if (bfq_bfqq_busy(bfqq))
2674 bfqd->wr_busy_queues--;
2677 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2678 bfqd->wr_busy_queues);
2681 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2683 bic_set_bfqq(bic, new_bfqq, 1);
2684 bfq_mark_bfqq_coop(new_bfqq);
2686 * new_bfqq now belongs to at least two bics (it is a shared queue):
2687 * set new_bfqq->bic to NULL. bfqq either:
2688 * - does not belong to any bic any more, and hence bfqq->bic must
2689 * be set to NULL, or
2690 * - is a queue whose owning bics have already been redirected to a
2691 * different queue, hence the queue is destined to not belong to
2692 * any bic soon and bfqq->bic is already NULL (therefore the next
2693 * assignment causes no harm).
2695 new_bfqq->bic = NULL;
2697 * If the queue is shared, the pid is the pid of one of the associated
2698 * processes. Which pid depends on the exact sequence of merge events
2699 * the queue underwent. So printing such a pid is useless and confusing
2700 * because it reports a random pid between those of the associated
2702 * We mark such a queue with a pid -1, and then print SHARED instead of
2703 * a pid in logging messages.
2707 /* release process reference to bfqq */
2708 bfq_put_queue(bfqq);
2711 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2714 struct bfq_data *bfqd = q->elevator->elevator_data;
2715 bool is_sync = op_is_sync(bio->bi_opf);
2716 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2719 * Disallow merge of a sync bio into an async request.
2721 if (is_sync && !rq_is_sync(rq))
2725 * Lookup the bfqq that this bio will be queued with. Allow
2726 * merge only if rq is queued there.
2732 * We take advantage of this function to perform an early merge
2733 * of the queues of possible cooperating processes.
2735 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2738 * bic still points to bfqq, then it has not yet been
2739 * redirected to some other bfq_queue, and a queue
2740 * merge between bfqq and new_bfqq can be safely
2741 * fulfilled, i.e., bic can be redirected to new_bfqq
2742 * and bfqq can be put.
2744 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2747 * If we get here, bio will be queued into new_queue,
2748 * so use new_bfqq to decide whether bio and rq can be
2754 * Change also bqfd->bio_bfqq, as
2755 * bfqd->bio_bic now points to new_bfqq, and
2756 * this function may be invoked again (and then may
2757 * use again bqfd->bio_bfqq).
2759 bfqd->bio_bfqq = bfqq;
2762 return bfqq == RQ_BFQQ(rq);
2766 * Set the maximum time for the in-service queue to consume its
2767 * budget. This prevents seeky processes from lowering the throughput.
2768 * In practice, a time-slice service scheme is used with seeky
2771 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2772 struct bfq_queue *bfqq)
2774 unsigned int timeout_coeff;
2776 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2779 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2781 bfqd->last_budget_start = ktime_get();
2783 bfqq->budget_timeout = jiffies +
2784 bfqd->bfq_timeout * timeout_coeff;
2787 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2788 struct bfq_queue *bfqq)
2791 bfq_clear_bfqq_fifo_expire(bfqq);
2793 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2795 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2796 bfqq->wr_coeff > 1 &&
2797 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2798 time_is_before_jiffies(bfqq->budget_timeout)) {
2800 * For soft real-time queues, move the start
2801 * of the weight-raising period forward by the
2802 * time the queue has not received any
2803 * service. Otherwise, a relatively long
2804 * service delay is likely to cause the
2805 * weight-raising period of the queue to end,
2806 * because of the short duration of the
2807 * weight-raising period of a soft real-time
2808 * queue. It is worth noting that this move
2809 * is not so dangerous for the other queues,
2810 * because soft real-time queues are not
2813 * To not add a further variable, we use the
2814 * overloaded field budget_timeout to
2815 * determine for how long the queue has not
2816 * received service, i.e., how much time has
2817 * elapsed since the queue expired. However,
2818 * this is a little imprecise, because
2819 * budget_timeout is set to jiffies if bfqq
2820 * not only expires, but also remains with no
2823 if (time_after(bfqq->budget_timeout,
2824 bfqq->last_wr_start_finish))
2825 bfqq->last_wr_start_finish +=
2826 jiffies - bfqq->budget_timeout;
2828 bfqq->last_wr_start_finish = jiffies;
2831 bfq_set_budget_timeout(bfqd, bfqq);
2832 bfq_log_bfqq(bfqd, bfqq,
2833 "set_in_service_queue, cur-budget = %d",
2834 bfqq->entity.budget);
2837 bfqd->in_service_queue = bfqq;
2841 * Get and set a new queue for service.
2843 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2845 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2847 __bfq_set_in_service_queue(bfqd, bfqq);
2851 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2853 struct bfq_queue *bfqq = bfqd->in_service_queue;
2856 bfq_mark_bfqq_wait_request(bfqq);
2859 * We don't want to idle for seeks, but we do want to allow
2860 * fair distribution of slice time for a process doing back-to-back
2861 * seeks. So allow a little bit of time for him to submit a new rq.
2863 sl = bfqd->bfq_slice_idle;
2865 * Unless the queue is being weight-raised or the scenario is
2866 * asymmetric, grant only minimum idle time if the queue
2867 * is seeky. A long idling is preserved for a weight-raised
2868 * queue, or, more in general, in an asymmetric scenario,
2869 * because a long idling is needed for guaranteeing to a queue
2870 * its reserved share of the throughput (in particular, it is
2871 * needed if the queue has a higher weight than some other
2874 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
2875 !bfq_asymmetric_scenario(bfqd, bfqq))
2876 sl = min_t(u64, sl, BFQ_MIN_TT);
2877 else if (bfqq->wr_coeff > 1)
2878 sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
2880 bfqd->last_idling_start = ktime_get();
2881 bfqd->last_idling_start_jiffies = jiffies;
2883 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
2885 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
2889 * In autotuning mode, max_budget is dynamically recomputed as the
2890 * amount of sectors transferred in timeout at the estimated peak
2891 * rate. This enables BFQ to utilize a full timeslice with a full
2892 * budget, even if the in-service queue is served at peak rate. And
2893 * this maximises throughput with sequential workloads.
2895 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
2897 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
2898 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
2902 * Update parameters related to throughput and responsiveness, as a
2903 * function of the estimated peak rate. See comments on
2904 * bfq_calc_max_budget(), and on the ref_wr_duration array.
2906 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
2908 if (bfqd->bfq_user_max_budget == 0) {
2909 bfqd->bfq_max_budget =
2910 bfq_calc_max_budget(bfqd);
2911 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
2915 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
2918 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
2919 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
2920 bfqd->peak_rate_samples = 1;
2921 bfqd->sequential_samples = 0;
2922 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
2924 } else /* no new rq dispatched, just reset the number of samples */
2925 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
2928 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
2929 bfqd->peak_rate_samples, bfqd->sequential_samples,
2930 bfqd->tot_sectors_dispatched);
2933 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
2935 u32 rate, weight, divisor;
2938 * For the convergence property to hold (see comments on
2939 * bfq_update_peak_rate()) and for the assessment to be
2940 * reliable, a minimum number of samples must be present, and
2941 * a minimum amount of time must have elapsed. If not so, do
2942 * not compute new rate. Just reset parameters, to get ready
2943 * for a new evaluation attempt.
2945 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
2946 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
2947 goto reset_computation;
2950 * If a new request completion has occurred after last
2951 * dispatch, then, to approximate the rate at which requests
2952 * have been served by the device, it is more precise to
2953 * extend the observation interval to the last completion.
2955 bfqd->delta_from_first =
2956 max_t(u64, bfqd->delta_from_first,
2957 bfqd->last_completion - bfqd->first_dispatch);
2960 * Rate computed in sects/usec, and not sects/nsec, for
2963 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
2964 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
2967 * Peak rate not updated if:
2968 * - the percentage of sequential dispatches is below 3/4 of the
2969 * total, and rate is below the current estimated peak rate
2970 * - rate is unreasonably high (> 20M sectors/sec)
2972 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
2973 rate <= bfqd->peak_rate) ||
2974 rate > 20<<BFQ_RATE_SHIFT)
2975 goto reset_computation;
2978 * We have to update the peak rate, at last! To this purpose,
2979 * we use a low-pass filter. We compute the smoothing constant
2980 * of the filter as a function of the 'weight' of the new
2983 * As can be seen in next formulas, we define this weight as a
2984 * quantity proportional to how sequential the workload is,
2985 * and to how long the observation time interval is.
2987 * The weight runs from 0 to 8. The maximum value of the
2988 * weight, 8, yields the minimum value for the smoothing
2989 * constant. At this minimum value for the smoothing constant,
2990 * the measured rate contributes for half of the next value of
2991 * the estimated peak rate.
2993 * So, the first step is to compute the weight as a function
2994 * of how sequential the workload is. Note that the weight
2995 * cannot reach 9, because bfqd->sequential_samples cannot
2996 * become equal to bfqd->peak_rate_samples, which, in its
2997 * turn, holds true because bfqd->sequential_samples is not
2998 * incremented for the first sample.
3000 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
3003 * Second step: further refine the weight as a function of the
3004 * duration of the observation interval.
3006 weight = min_t(u32, 8,
3007 div_u64(weight * bfqd->delta_from_first,
3008 BFQ_RATE_REF_INTERVAL));
3011 * Divisor ranging from 10, for minimum weight, to 2, for
3014 divisor = 10 - weight;
3017 * Finally, update peak rate:
3019 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
3021 bfqd->peak_rate *= divisor-1;
3022 bfqd->peak_rate /= divisor;
3023 rate /= divisor; /* smoothing constant alpha = 1/divisor */
3025 bfqd->peak_rate += rate;
3028 * For a very slow device, bfqd->peak_rate can reach 0 (see
3029 * the minimum representable values reported in the comments
3030 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3031 * divisions by zero where bfqd->peak_rate is used as a
3034 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
3036 update_thr_responsiveness_params(bfqd);
3039 bfq_reset_rate_computation(bfqd, rq);
3043 * Update the read/write peak rate (the main quantity used for
3044 * auto-tuning, see update_thr_responsiveness_params()).
3046 * It is not trivial to estimate the peak rate (correctly): because of
3047 * the presence of sw and hw queues between the scheduler and the
3048 * device components that finally serve I/O requests, it is hard to
3049 * say exactly when a given dispatched request is served inside the
3050 * device, and for how long. As a consequence, it is hard to know
3051 * precisely at what rate a given set of requests is actually served
3054 * On the opposite end, the dispatch time of any request is trivially
3055 * available, and, from this piece of information, the "dispatch rate"
3056 * of requests can be immediately computed. So, the idea in the next
3057 * function is to use what is known, namely request dispatch times
3058 * (plus, when useful, request completion times), to estimate what is
3059 * unknown, namely in-device request service rate.
3061 * The main issue is that, because of the above facts, the rate at
3062 * which a certain set of requests is dispatched over a certain time
3063 * interval can vary greatly with respect to the rate at which the
3064 * same requests are then served. But, since the size of any
3065 * intermediate queue is limited, and the service scheme is lossless
3066 * (no request is silently dropped), the following obvious convergence
3067 * property holds: the number of requests dispatched MUST become
3068 * closer and closer to the number of requests completed as the
3069 * observation interval grows. This is the key property used in
3070 * the next function to estimate the peak service rate as a function
3071 * of the observed dispatch rate. The function assumes to be invoked
3072 * on every request dispatch.
3074 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
3076 u64 now_ns = ktime_get_ns();
3078 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
3079 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
3080 bfqd->peak_rate_samples);
3081 bfq_reset_rate_computation(bfqd, rq);
3082 goto update_last_values; /* will add one sample */
3086 * Device idle for very long: the observation interval lasting
3087 * up to this dispatch cannot be a valid observation interval
3088 * for computing a new peak rate (similarly to the late-
3089 * completion event in bfq_completed_request()). Go to
3090 * update_rate_and_reset to have the following three steps
3092 * - close the observation interval at the last (previous)
3093 * request dispatch or completion
3094 * - compute rate, if possible, for that observation interval
3095 * - start a new observation interval with this dispatch
3097 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
3098 bfqd->rq_in_driver == 0)
3099 goto update_rate_and_reset;
3101 /* Update sampling information */
3102 bfqd->peak_rate_samples++;
3104 if ((bfqd->rq_in_driver > 0 ||
3105 now_ns - bfqd->last_completion < BFQ_MIN_TT)
3106 && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
3107 bfqd->sequential_samples++;
3109 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
3111 /* Reset max observed rq size every 32 dispatches */
3112 if (likely(bfqd->peak_rate_samples % 32))
3113 bfqd->last_rq_max_size =
3114 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
3116 bfqd->last_rq_max_size = blk_rq_sectors(rq);
3118 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
3120 /* Target observation interval not yet reached, go on sampling */
3121 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
3122 goto update_last_values;
3124 update_rate_and_reset:
3125 bfq_update_rate_reset(bfqd, rq);
3127 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
3128 if (RQ_BFQQ(rq) == bfqd->in_service_queue)
3129 bfqd->in_serv_last_pos = bfqd->last_position;
3130 bfqd->last_dispatch = now_ns;
3134 * Remove request from internal lists.
3136 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
3138 struct bfq_queue *bfqq = RQ_BFQQ(rq);
3141 * For consistency, the next instruction should have been
3142 * executed after removing the request from the queue and
3143 * dispatching it. We execute instead this instruction before
3144 * bfq_remove_request() (and hence introduce a temporary
3145 * inconsistency), for efficiency. In fact, should this
3146 * dispatch occur for a non in-service bfqq, this anticipated
3147 * increment prevents two counters related to bfqq->dispatched
3148 * from risking to be, first, uselessly decremented, and then
3149 * incremented again when the (new) value of bfqq->dispatched
3150 * happens to be taken into account.
3153 bfq_update_peak_rate(q->elevator->elevator_data, rq);
3155 bfq_remove_request(q, rq);
3158 static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3161 * If this bfqq is shared between multiple processes, check
3162 * to make sure that those processes are still issuing I/Os
3163 * within the mean seek distance. If not, it may be time to
3164 * break the queues apart again.
3166 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
3167 bfq_mark_bfqq_split_coop(bfqq);
3169 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
3170 if (bfqq->dispatched == 0)
3172 * Overloading budget_timeout field to store
3173 * the time at which the queue remains with no
3174 * backlog and no outstanding request; used by
3175 * the weight-raising mechanism.
3177 bfqq->budget_timeout = jiffies;
3179 bfq_del_bfqq_busy(bfqd, bfqq, true);
3181 bfq_requeue_bfqq(bfqd, bfqq, true);
3183 * Resort priority tree of potential close cooperators.
3184 * See comments on bfq_pos_tree_add_move() for the unlikely().
3186 if (unlikely(!bfqd->nonrot_with_queueing))
3187 bfq_pos_tree_add_move(bfqd, bfqq);
3191 * All in-service entities must have been properly deactivated
3192 * or requeued before executing the next function, which
3193 * resets all in-service entities as no more in service. This
3194 * may cause bfqq to be freed. If this happens, the next
3195 * function returns true.
3197 return __bfq_bfqd_reset_in_service(bfqd);
3201 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3202 * @bfqd: device data.
3203 * @bfqq: queue to update.
3204 * @reason: reason for expiration.
3206 * Handle the feedback on @bfqq budget at queue expiration.
3207 * See the body for detailed comments.
3209 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
3210 struct bfq_queue *bfqq,
3211 enum bfqq_expiration reason)
3213 struct request *next_rq;
3214 int budget, min_budget;
3216 min_budget = bfq_min_budget(bfqd);
3218 if (bfqq->wr_coeff == 1)
3219 budget = bfqq->max_budget;
3221 * Use a constant, low budget for weight-raised queues,
3222 * to help achieve a low latency. Keep it slightly higher
3223 * than the minimum possible budget, to cause a little
3224 * bit fewer expirations.
3226 budget = 2 * min_budget;
3228 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
3229 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
3230 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
3231 budget, bfq_min_budget(bfqd));
3232 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
3233 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
3235 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
3238 * Caveat: in all the following cases we trade latency
3241 case BFQQE_TOO_IDLE:
3243 * This is the only case where we may reduce
3244 * the budget: if there is no request of the
3245 * process still waiting for completion, then
3246 * we assume (tentatively) that the timer has
3247 * expired because the batch of requests of
3248 * the process could have been served with a
3249 * smaller budget. Hence, betting that
3250 * process will behave in the same way when it
3251 * becomes backlogged again, we reduce its
3252 * next budget. As long as we guess right,
3253 * this budget cut reduces the latency
3254 * experienced by the process.
3256 * However, if there are still outstanding
3257 * requests, then the process may have not yet
3258 * issued its next request just because it is
3259 * still waiting for the completion of some of
3260 * the still outstanding ones. So in this
3261 * subcase we do not reduce its budget, on the
3262 * contrary we increase it to possibly boost
3263 * the throughput, as discussed in the
3264 * comments to the BUDGET_TIMEOUT case.
3266 if (bfqq->dispatched > 0) /* still outstanding reqs */
3267 budget = min(budget * 2, bfqd->bfq_max_budget);
3269 if (budget > 5 * min_budget)
3270 budget -= 4 * min_budget;
3272 budget = min_budget;
3275 case BFQQE_BUDGET_TIMEOUT:
3277 * We double the budget here because it gives
3278 * the chance to boost the throughput if this
3279 * is not a seeky process (and has bumped into
3280 * this timeout because of, e.g., ZBR).
3282 budget = min(budget * 2, bfqd->bfq_max_budget);
3284 case BFQQE_BUDGET_EXHAUSTED:
3286 * The process still has backlog, and did not
3287 * let either the budget timeout or the disk
3288 * idling timeout expire. Hence it is not
3289 * seeky, has a short thinktime and may be
3290 * happy with a higher budget too. So
3291 * definitely increase the budget of this good
3292 * candidate to boost the disk throughput.
3294 budget = min(budget * 4, bfqd->bfq_max_budget);
3296 case BFQQE_NO_MORE_REQUESTS:
3298 * For queues that expire for this reason, it
3299 * is particularly important to keep the
3300 * budget close to the actual service they
3301 * need. Doing so reduces the timestamp
3302 * misalignment problem described in the
3303 * comments in the body of
3304 * __bfq_activate_entity. In fact, suppose
3305 * that a queue systematically expires for
3306 * BFQQE_NO_MORE_REQUESTS and presents a
3307 * new request in time to enjoy timestamp
3308 * back-shifting. The larger the budget of the
3309 * queue is with respect to the service the
3310 * queue actually requests in each service
3311 * slot, the more times the queue can be
3312 * reactivated with the same virtual finish
3313 * time. It follows that, even if this finish
3314 * time is pushed to the system virtual time
3315 * to reduce the consequent timestamp
3316 * misalignment, the queue unjustly enjoys for
3317 * many re-activations a lower finish time
3318 * than all newly activated queues.
3320 * The service needed by bfqq is measured
3321 * quite precisely by bfqq->entity.service.
3322 * Since bfqq does not enjoy device idling,
3323 * bfqq->entity.service is equal to the number
3324 * of sectors that the process associated with
3325 * bfqq requested to read/write before waiting
3326 * for request completions, or blocking for
3329 budget = max_t(int, bfqq->entity.service, min_budget);
3334 } else if (!bfq_bfqq_sync(bfqq)) {
3336 * Async queues get always the maximum possible
3337 * budget, as for them we do not care about latency
3338 * (in addition, their ability to dispatch is limited
3339 * by the charging factor).
3341 budget = bfqd->bfq_max_budget;
3344 bfqq->max_budget = budget;
3346 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
3347 !bfqd->bfq_user_max_budget)
3348 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
3351 * If there is still backlog, then assign a new budget, making
3352 * sure that it is large enough for the next request. Since
3353 * the finish time of bfqq must be kept in sync with the
3354 * budget, be sure to call __bfq_bfqq_expire() *after* this
3357 * If there is no backlog, then no need to update the budget;
3358 * it will be updated on the arrival of a new request.
3360 next_rq = bfqq->next_rq;
3362 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
3363 bfq_serv_to_charge(next_rq, bfqq));
3365 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
3366 next_rq ? blk_rq_sectors(next_rq) : 0,
3367 bfqq->entity.budget);
3371 * Return true if the process associated with bfqq is "slow". The slow
3372 * flag is used, in addition to the budget timeout, to reduce the
3373 * amount of service provided to seeky processes, and thus reduce
3374 * their chances to lower the throughput. More details in the comments
3375 * on the function bfq_bfqq_expire().
3377 * An important observation is in order: as discussed in the comments
3378 * on the function bfq_update_peak_rate(), with devices with internal
3379 * queues, it is hard if ever possible to know when and for how long
3380 * an I/O request is processed by the device (apart from the trivial
3381 * I/O pattern where a new request is dispatched only after the
3382 * previous one has been completed). This makes it hard to evaluate
3383 * the real rate at which the I/O requests of each bfq_queue are
3384 * served. In fact, for an I/O scheduler like BFQ, serving a
3385 * bfq_queue means just dispatching its requests during its service
3386 * slot (i.e., until the budget of the queue is exhausted, or the
3387 * queue remains idle, or, finally, a timeout fires). But, during the
3388 * service slot of a bfq_queue, around 100 ms at most, the device may
3389 * be even still processing requests of bfq_queues served in previous
3390 * service slots. On the opposite end, the requests of the in-service
3391 * bfq_queue may be completed after the service slot of the queue
3394 * Anyway, unless more sophisticated solutions are used
3395 * (where possible), the sum of the sizes of the requests dispatched
3396 * during the service slot of a bfq_queue is probably the only
3397 * approximation available for the service received by the bfq_queue
3398 * during its service slot. And this sum is the quantity used in this
3399 * function to evaluate the I/O speed of a process.
3401 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3402 bool compensate, enum bfqq_expiration reason,
3403 unsigned long *delta_ms)
3405 ktime_t delta_ktime;
3407 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
3409 if (!bfq_bfqq_sync(bfqq))
3413 delta_ktime = bfqd->last_idling_start;
3415 delta_ktime = ktime_get();
3416 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
3417 delta_usecs = ktime_to_us(delta_ktime);
3419 /* don't use too short time intervals */
3420 if (delta_usecs < 1000) {
3421 if (blk_queue_nonrot(bfqd->queue))
3423 * give same worst-case guarantees as idling
3426 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
3427 else /* charge at least one seek */
3428 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
3433 *delta_ms = delta_usecs / USEC_PER_MSEC;
3436 * Use only long (> 20ms) intervals to filter out excessive
3437 * spikes in service rate estimation.
3439 if (delta_usecs > 20000) {
3441 * Caveat for rotational devices: processes doing I/O
3442 * in the slower disk zones tend to be slow(er) even
3443 * if not seeky. In this respect, the estimated peak
3444 * rate is likely to be an average over the disk
3445 * surface. Accordingly, to not be too harsh with
3446 * unlucky processes, a process is deemed slow only if
3447 * its rate has been lower than half of the estimated
3450 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
3453 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
3459 * To be deemed as soft real-time, an application must meet two
3460 * requirements. First, the application must not require an average
3461 * bandwidth higher than the approximate bandwidth required to playback or
3462 * record a compressed high-definition video.
3463 * The next function is invoked on the completion of the last request of a
3464 * batch, to compute the next-start time instant, soft_rt_next_start, such
3465 * that, if the next request of the application does not arrive before
3466 * soft_rt_next_start, then the above requirement on the bandwidth is met.
3468 * The second requirement is that the request pattern of the application is
3469 * isochronous, i.e., that, after issuing a request or a batch of requests,
3470 * the application stops issuing new requests until all its pending requests
3471 * have been completed. After that, the application may issue a new batch,
3473 * For this reason the next function is invoked to compute
3474 * soft_rt_next_start only for applications that meet this requirement,
3475 * whereas soft_rt_next_start is set to infinity for applications that do
3478 * Unfortunately, even a greedy (i.e., I/O-bound) application may
3479 * happen to meet, occasionally or systematically, both the above
3480 * bandwidth and isochrony requirements. This may happen at least in
3481 * the following circumstances. First, if the CPU load is high. The
3482 * application may stop issuing requests while the CPUs are busy
3483 * serving other processes, then restart, then stop again for a while,
3484 * and so on. The other circumstances are related to the storage
3485 * device: the storage device is highly loaded or reaches a low-enough
3486 * throughput with the I/O of the application (e.g., because the I/O
3487 * is random and/or the device is slow). In all these cases, the
3488 * I/O of the application may be simply slowed down enough to meet
3489 * the bandwidth and isochrony requirements. To reduce the probability
3490 * that greedy applications are deemed as soft real-time in these
3491 * corner cases, a further rule is used in the computation of
3492 * soft_rt_next_start: the return value of this function is forced to
3493 * be higher than the maximum between the following two quantities.
3495 * (a) Current time plus: (1) the maximum time for which the arrival
3496 * of a request is waited for when a sync queue becomes idle,
3497 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
3498 * postpone for a moment the reason for adding a few extra
3499 * jiffies; we get back to it after next item (b). Lower-bounding
3500 * the return value of this function with the current time plus
3501 * bfqd->bfq_slice_idle tends to filter out greedy applications,
3502 * because the latter issue their next request as soon as possible
3503 * after the last one has been completed. In contrast, a soft
3504 * real-time application spends some time processing data, after a
3505 * batch of its requests has been completed.
3507 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
3508 * above, greedy applications may happen to meet both the
3509 * bandwidth and isochrony requirements under heavy CPU or
3510 * storage-device load. In more detail, in these scenarios, these
3511 * applications happen, only for limited time periods, to do I/O
3512 * slowly enough to meet all the requirements described so far,
3513 * including the filtering in above item (a). These slow-speed
3514 * time intervals are usually interspersed between other time
3515 * intervals during which these applications do I/O at a very high
3516 * speed. Fortunately, exactly because of the high speed of the
3517 * I/O in the high-speed intervals, the values returned by this
3518 * function happen to be so high, near the end of any such
3519 * high-speed interval, to be likely to fall *after* the end of
3520 * the low-speed time interval that follows. These high values are
3521 * stored in bfqq->soft_rt_next_start after each invocation of
3522 * this function. As a consequence, if the last value of
3523 * bfqq->soft_rt_next_start is constantly used to lower-bound the
3524 * next value that this function may return, then, from the very
3525 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
3526 * likely to be constantly kept so high that any I/O request
3527 * issued during the low-speed interval is considered as arriving
3528 * to soon for the application to be deemed as soft
3529 * real-time. Then, in the high-speed interval that follows, the
3530 * application will not be deemed as soft real-time, just because
3531 * it will do I/O at a high speed. And so on.
3533 * Getting back to the filtering in item (a), in the following two
3534 * cases this filtering might be easily passed by a greedy
3535 * application, if the reference quantity was just
3536 * bfqd->bfq_slice_idle:
3537 * 1) HZ is so low that the duration of a jiffy is comparable to or
3538 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
3539 * devices with HZ=100. The time granularity may be so coarse
3540 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
3541 * is rather lower than the exact value.
3542 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
3543 * for a while, then suddenly 'jump' by several units to recover the lost
3544 * increments. This seems to happen, e.g., inside virtual machines.
3545 * To address this issue, in the filtering in (a) we do not use as a
3546 * reference time interval just bfqd->bfq_slice_idle, but
3547 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
3548 * minimum number of jiffies for which the filter seems to be quite
3549 * precise also in embedded systems and KVM/QEMU virtual machines.
3551 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
3552 struct bfq_queue *bfqq)
3554 return max3(bfqq->soft_rt_next_start,
3555 bfqq->last_idle_bklogged +
3556 HZ * bfqq->service_from_backlogged /
3557 bfqd->bfq_wr_max_softrt_rate,
3558 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
3562 * bfq_bfqq_expire - expire a queue.
3563 * @bfqd: device owning the queue.
3564 * @bfqq: the queue to expire.
3565 * @compensate: if true, compensate for the time spent idling.
3566 * @reason: the reason causing the expiration.
3568 * If the process associated with bfqq does slow I/O (e.g., because it
3569 * issues random requests), we charge bfqq with the time it has been
3570 * in service instead of the service it has received (see
3571 * bfq_bfqq_charge_time for details on how this goal is achieved). As
3572 * a consequence, bfqq will typically get higher timestamps upon
3573 * reactivation, and hence it will be rescheduled as if it had
3574 * received more service than what it has actually received. In the
3575 * end, bfqq receives less service in proportion to how slowly its
3576 * associated process consumes its budgets (and hence how seriously it
3577 * tends to lower the throughput). In addition, this time-charging
3578 * strategy guarantees time fairness among slow processes. In
3579 * contrast, if the process associated with bfqq is not slow, we
3580 * charge bfqq exactly with the service it has received.
3582 * Charging time to the first type of queues and the exact service to
3583 * the other has the effect of using the WF2Q+ policy to schedule the
3584 * former on a timeslice basis, without violating service domain
3585 * guarantees among the latter.
3587 void bfq_bfqq_expire(struct bfq_data *bfqd,
3588 struct bfq_queue *bfqq,
3590 enum bfqq_expiration reason)
3593 unsigned long delta = 0;
3594 struct bfq_entity *entity = &bfqq->entity;
3597 * Check whether the process is slow (see bfq_bfqq_is_slow).
3599 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
3602 * As above explained, charge slow (typically seeky) and
3603 * timed-out queues with the time and not the service
3604 * received, to favor sequential workloads.
3606 * Processes doing I/O in the slower disk zones will tend to
3607 * be slow(er) even if not seeky. Therefore, since the
3608 * estimated peak rate is actually an average over the disk
3609 * surface, these processes may timeout just for bad luck. To
3610 * avoid punishing them, do not charge time to processes that
3611 * succeeded in consuming at least 2/3 of their budget. This
3612 * allows BFQ to preserve enough elasticity to still perform
3613 * bandwidth, and not time, distribution with little unlucky
3614 * or quasi-sequential processes.
3616 if (bfqq->wr_coeff == 1 &&
3618 (reason == BFQQE_BUDGET_TIMEOUT &&
3619 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
3620 bfq_bfqq_charge_time(bfqd, bfqq, delta);
3622 if (reason == BFQQE_TOO_IDLE &&
3623 entity->service <= 2 * entity->budget / 10)
3624 bfq_clear_bfqq_IO_bound(bfqq);
3626 if (bfqd->low_latency && bfqq->wr_coeff == 1)
3627 bfqq->last_wr_start_finish = jiffies;
3629 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
3630 RB_EMPTY_ROOT(&bfqq->sort_list)) {
3632 * If we get here, and there are no outstanding
3633 * requests, then the request pattern is isochronous
3634 * (see the comments on the function
3635 * bfq_bfqq_softrt_next_start()). Thus we can compute
3636 * soft_rt_next_start. And we do it, unless bfqq is in
3637 * interactive weight raising. We do not do it in the
3638 * latter subcase, for the following reason. bfqq may
3639 * be conveying the I/O needed to load a soft
3640 * real-time application. Such an application will
3641 * actually exhibit a soft real-time I/O pattern after
3642 * it finally starts doing its job. But, if
3643 * soft_rt_next_start is computed here for an
3644 * interactive bfqq, and bfqq had received a lot of
3645 * service before remaining with no outstanding
3646 * request (likely to happen on a fast device), then
3647 * soft_rt_next_start would be assigned such a high
3648 * value that, for a very long time, bfqq would be
3649 * prevented from being possibly considered as soft
3652 * If, instead, the queue still has outstanding
3653 * requests, then we have to wait for the completion
3654 * of all the outstanding requests to discover whether
3655 * the request pattern is actually isochronous.
3657 if (bfqq->dispatched == 0 &&
3658 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
3659 bfqq->soft_rt_next_start =
3660 bfq_bfqq_softrt_next_start(bfqd, bfqq);
3661 else if (bfqq->dispatched > 0) {
3663 * Schedule an update of soft_rt_next_start to when
3664 * the task may be discovered to be isochronous.
3666 bfq_mark_bfqq_softrt_update(bfqq);
3670 bfq_log_bfqq(bfqd, bfqq,
3671 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
3672 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
3675 * bfqq expired, so no total service time needs to be computed
3676 * any longer: reset state machine for measuring total service
3679 bfqd->rqs_injected = bfqd->wait_dispatch = false;
3680 bfqd->waited_rq = NULL;
3683 * Increase, decrease or leave budget unchanged according to
3686 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
3687 if (__bfq_bfqq_expire(bfqd, bfqq))
3688 /* bfqq is gone, no more actions on it */
3691 /* mark bfqq as waiting a request only if a bic still points to it */
3692 if (!bfq_bfqq_busy(bfqq) &&
3693 reason != BFQQE_BUDGET_TIMEOUT &&
3694 reason != BFQQE_BUDGET_EXHAUSTED) {
3695 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
3697 * Not setting service to 0, because, if the next rq
3698 * arrives in time, the queue will go on receiving
3699 * service with this same budget (as if it never expired)
3702 entity->service = 0;
3705 * Reset the received-service counter for every parent entity.
3706 * Differently from what happens with bfqq->entity.service,
3707 * the resetting of this counter never needs to be postponed
3708 * for parent entities. In fact, in case bfqq may have a
3709 * chance to go on being served using the last, partially
3710 * consumed budget, bfqq->entity.service needs to be kept,
3711 * because if bfqq then actually goes on being served using
3712 * the same budget, the last value of bfqq->entity.service is
3713 * needed to properly decrement bfqq->entity.budget by the
3714 * portion already consumed. In contrast, it is not necessary
3715 * to keep entity->service for parent entities too, because
3716 * the bubble up of the new value of bfqq->entity.budget will
3717 * make sure that the budgets of parent entities are correct,
3718 * even in case bfqq and thus parent entities go on receiving
3719 * service with the same budget.
3721 entity = entity->parent;
3722 for_each_entity(entity)
3723 entity->service = 0;
3727 * Budget timeout is not implemented through a dedicated timer, but
3728 * just checked on request arrivals and completions, as well as on
3729 * idle timer expirations.
3731 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
3733 return time_is_before_eq_jiffies(bfqq->budget_timeout);
3737 * If we expire a queue that is actively waiting (i.e., with the
3738 * device idled) for the arrival of a new request, then we may incur
3739 * the timestamp misalignment problem described in the body of the
3740 * function __bfq_activate_entity. Hence we return true only if this
3741 * condition does not hold, or if the queue is slow enough to deserve
3742 * only to be kicked off for preserving a high throughput.
3744 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
3746 bfq_log_bfqq(bfqq->bfqd, bfqq,
3747 "may_budget_timeout: wait_request %d left %d timeout %d",
3748 bfq_bfqq_wait_request(bfqq),
3749 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
3750 bfq_bfqq_budget_timeout(bfqq));
3752 return (!bfq_bfqq_wait_request(bfqq) ||
3753 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
3755 bfq_bfqq_budget_timeout(bfqq);
3758 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
3759 struct bfq_queue *bfqq)
3761 bool rot_without_queueing =
3762 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
3763 bfqq_sequential_and_IO_bound,
3766 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
3767 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
3770 * The next variable takes into account the cases where idling
3771 * boosts the throughput.
3773 * The value of the variable is computed considering, first, that
3774 * idling is virtually always beneficial for the throughput if:
3775 * (a) the device is not NCQ-capable and rotational, or
3776 * (b) regardless of the presence of NCQ, the device is rotational and
3777 * the request pattern for bfqq is I/O-bound and sequential, or
3778 * (c) regardless of whether it is rotational, the device is
3779 * not NCQ-capable and the request pattern for bfqq is
3780 * I/O-bound and sequential.
3782 * Secondly, and in contrast to the above item (b), idling an
3783 * NCQ-capable flash-based device would not boost the
3784 * throughput even with sequential I/O; rather it would lower
3785 * the throughput in proportion to how fast the device
3786 * is. Accordingly, the next variable is true if any of the
3787 * above conditions (a), (b) or (c) is true, and, in
3788 * particular, happens to be false if bfqd is an NCQ-capable
3789 * flash-based device.
3791 idling_boosts_thr = rot_without_queueing ||
3792 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
3793 bfqq_sequential_and_IO_bound);
3796 * The return value of this function is equal to that of
3797 * idling_boosts_thr, unless a special case holds. In this
3798 * special case, described below, idling may cause problems to
3799 * weight-raised queues.
3801 * When the request pool is saturated (e.g., in the presence
3802 * of write hogs), if the processes associated with
3803 * non-weight-raised queues ask for requests at a lower rate,
3804 * then processes associated with weight-raised queues have a
3805 * higher probability to get a request from the pool
3806 * immediately (or at least soon) when they need one. Thus
3807 * they have a higher probability to actually get a fraction
3808 * of the device throughput proportional to their high
3809 * weight. This is especially true with NCQ-capable drives,
3810 * which enqueue several requests in advance, and further
3811 * reorder internally-queued requests.
3813 * For this reason, we force to false the return value if
3814 * there are weight-raised busy queues. In this case, and if
3815 * bfqq is not weight-raised, this guarantees that the device
3816 * is not idled for bfqq (if, instead, bfqq is weight-raised,
3817 * then idling will be guaranteed by another variable, see
3818 * below). Combined with the timestamping rules of BFQ (see
3819 * [1] for details), this behavior causes bfqq, and hence any
3820 * sync non-weight-raised queue, to get a lower number of
3821 * requests served, and thus to ask for a lower number of
3822 * requests from the request pool, before the busy
3823 * weight-raised queues get served again. This often mitigates
3824 * starvation problems in the presence of heavy write
3825 * workloads and NCQ, thereby guaranteeing a higher
3826 * application and system responsiveness in these hostile
3829 return idling_boosts_thr &&
3830 bfqd->wr_busy_queues == 0;
3834 * There is a case where idling does not have to be performed for
3835 * throughput concerns, but to preserve the throughput share of
3836 * the process associated with bfqq.
3838 * To introduce this case, we can note that allowing the drive
3839 * to enqueue more than one request at a time, and hence
3840 * delegating de facto final scheduling decisions to the
3841 * drive's internal scheduler, entails loss of control on the
3842 * actual request service order. In particular, the critical
3843 * situation is when requests from different processes happen
3844 * to be present, at the same time, in the internal queue(s)
3845 * of the drive. In such a situation, the drive, by deciding
3846 * the service order of the internally-queued requests, does
3847 * determine also the actual throughput distribution among
3848 * these processes. But the drive typically has no notion or
3849 * concern about per-process throughput distribution, and
3850 * makes its decisions only on a per-request basis. Therefore,
3851 * the service distribution enforced by the drive's internal
3852 * scheduler is likely to coincide with the desired throughput
3853 * distribution only in a completely symmetric, or favorably
3854 * skewed scenario where:
3855 * (i-a) each of these processes must get the same throughput as
3857 * (i-b) in case (i-a) does not hold, it holds that the process
3858 * associated with bfqq must receive a lower or equal
3859 * throughput than any of the other processes;
3860 * (ii) the I/O of each process has the same properties, in
3861 * terms of locality (sequential or random), direction
3862 * (reads or writes), request sizes, greediness
3863 * (from I/O-bound to sporadic), and so on;
3865 * In fact, in such a scenario, the drive tends to treat the requests
3866 * of each process in about the same way as the requests of the
3867 * others, and thus to provide each of these processes with about the
3868 * same throughput. This is exactly the desired throughput
3869 * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3870 * even more convenient distribution for (the process associated with)
3873 * In contrast, in any asymmetric or unfavorable scenario, device
3874 * idling (I/O-dispatch plugging) is certainly needed to guarantee
3875 * that bfqq receives its assigned fraction of the device throughput
3876 * (see [1] for details).
3878 * The problem is that idling may significantly reduce throughput with
3879 * certain combinations of types of I/O and devices. An important
3880 * example is sync random I/O on flash storage with command
3881 * queueing. So, unless bfqq falls in cases where idling also boosts
3882 * throughput, it is important to check conditions (i-a), i(-b) and
3883 * (ii) accurately, so as to avoid idling when not strictly needed for
3884 * service guarantees.
3886 * Unfortunately, it is extremely difficult to thoroughly check
3887 * condition (ii). And, in case there are active groups, it becomes
3888 * very difficult to check conditions (i-a) and (i-b) too. In fact,
3889 * if there are active groups, then, for conditions (i-a) or (i-b) to
3890 * become false 'indirectly', it is enough that an active group
3891 * contains more active processes or sub-groups than some other active
3892 * group. More precisely, for conditions (i-a) or (i-b) to become
3893 * false because of such a group, it is not even necessary that the
3894 * group is (still) active: it is sufficient that, even if the group
3895 * has become inactive, some of its descendant processes still have
3896 * some request already dispatched but still waiting for
3897 * completion. In fact, requests have still to be guaranteed their
3898 * share of the throughput even after being dispatched. In this
3899 * respect, it is easy to show that, if a group frequently becomes
3900 * inactive while still having in-flight requests, and if, when this
3901 * happens, the group is not considered in the calculation of whether
3902 * the scenario is asymmetric, then the group may fail to be
3903 * guaranteed its fair share of the throughput (basically because
3904 * idling may not be performed for the descendant processes of the
3905 * group, but it had to be). We address this issue with the following
3906 * bi-modal behavior, implemented in the function
3907 * bfq_asymmetric_scenario().
3909 * If there are groups with requests waiting for completion
3910 * (as commented above, some of these groups may even be
3911 * already inactive), then the scenario is tagged as
3912 * asymmetric, conservatively, without checking any of the
3913 * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3914 * This behavior matches also the fact that groups are created
3915 * exactly if controlling I/O is a primary concern (to
3916 * preserve bandwidth and latency guarantees).
3918 * On the opposite end, if there are no groups with requests waiting
3919 * for completion, then only conditions (i-a) and (i-b) are actually
3920 * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3921 * idling is not performed, regardless of whether condition (ii)
3922 * holds. In other words, only if conditions (i-a) and (i-b) do not
3923 * hold, then idling is allowed, and the device tends to be prevented
3924 * from queueing many requests, possibly of several processes. Since
3925 * there are no groups with requests waiting for completion, then, to
3926 * control conditions (i-a) and (i-b) it is enough to check just
3927 * whether all the queues with requests waiting for completion also
3928 * have the same weight.
3930 * Not checking condition (ii) evidently exposes bfqq to the
3931 * risk of getting less throughput than its fair share.
3932 * However, for queues with the same weight, a further
3933 * mechanism, preemption, mitigates or even eliminates this
3934 * problem. And it does so without consequences on overall
3935 * throughput. This mechanism and its benefits are explained
3936 * in the next three paragraphs.
3938 * Even if a queue, say Q, is expired when it remains idle, Q
3939 * can still preempt the new in-service queue if the next
3940 * request of Q arrives soon (see the comments on
3941 * bfq_bfqq_update_budg_for_activation). If all queues and
3942 * groups have the same weight, this form of preemption,
3943 * combined with the hole-recovery heuristic described in the
3944 * comments on function bfq_bfqq_update_budg_for_activation,
3945 * are enough to preserve a correct bandwidth distribution in
3946 * the mid term, even without idling. In fact, even if not
3947 * idling allows the internal queues of the device to contain
3948 * many requests, and thus to reorder requests, we can rather
3949 * safely assume that the internal scheduler still preserves a
3950 * minimum of mid-term fairness.
3952 * More precisely, this preemption-based, idleless approach
3953 * provides fairness in terms of IOPS, and not sectors per
3954 * second. This can be seen with a simple example. Suppose
3955 * that there are two queues with the same weight, but that
3956 * the first queue receives requests of 8 sectors, while the
3957 * second queue receives requests of 1024 sectors. In
3958 * addition, suppose that each of the two queues contains at
3959 * most one request at a time, which implies that each queue
3960 * always remains idle after it is served. Finally, after
3961 * remaining idle, each queue receives very quickly a new
3962 * request. It follows that the two queues are served
3963 * alternatively, preempting each other if needed. This
3964 * implies that, although both queues have the same weight,
3965 * the queue with large requests receives a service that is
3966 * 1024/8 times as high as the service received by the other
3969 * The motivation for using preemption instead of idling (for
3970 * queues with the same weight) is that, by not idling,
3971 * service guarantees are preserved (completely or at least in
3972 * part) without minimally sacrificing throughput. And, if
3973 * there is no active group, then the primary expectation for
3974 * this device is probably a high throughput.
3976 * We are now left only with explaining the additional
3977 * compound condition that is checked below for deciding
3978 * whether the scenario is asymmetric. To explain this
3979 * compound condition, we need to add that the function
3980 * bfq_asymmetric_scenario checks the weights of only
3981 * non-weight-raised queues, for efficiency reasons (see
3982 * comments on bfq_weights_tree_add()). Then the fact that
3983 * bfqq is weight-raised is checked explicitly here. More
3984 * precisely, the compound condition below takes into account
3985 * also the fact that, even if bfqq is being weight-raised,
3986 * the scenario is still symmetric if all queues with requests
3987 * waiting for completion happen to be
3988 * weight-raised. Actually, we should be even more precise
3989 * here, and differentiate between interactive weight raising
3990 * and soft real-time weight raising.
3992 * As a side note, it is worth considering that the above
3993 * device-idling countermeasures may however fail in the
3994 * following unlucky scenario: if idling is (correctly)
3995 * disabled in a time period during which all symmetry
3996 * sub-conditions hold, and hence the device is allowed to
3997 * enqueue many requests, but at some later point in time some
3998 * sub-condition stops to hold, then it may become impossible
3999 * to let requests be served in the desired order until all
4000 * the requests already queued in the device have been served.
4002 static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
4003 struct bfq_queue *bfqq)
4005 return (bfqq->wr_coeff > 1 &&
4006 bfqd->wr_busy_queues <
4007 bfq_tot_busy_queues(bfqd)) ||
4008 bfq_asymmetric_scenario(bfqd, bfqq);
4012 * For a queue that becomes empty, device idling is allowed only if
4013 * this function returns true for that queue. As a consequence, since
4014 * device idling plays a critical role for both throughput boosting
4015 * and service guarantees, the return value of this function plays a
4016 * critical role as well.
4018 * In a nutshell, this function returns true only if idling is
4019 * beneficial for throughput or, even if detrimental for throughput,
4020 * idling is however necessary to preserve service guarantees (low
4021 * latency, desired throughput distribution, ...). In particular, on
4022 * NCQ-capable devices, this function tries to return false, so as to
4023 * help keep the drives' internal queues full, whenever this helps the
4024 * device boost the throughput without causing any service-guarantee
4027 * Most of the issues taken into account to get the return value of
4028 * this function are not trivial. We discuss these issues in the two
4029 * functions providing the main pieces of information needed by this
4032 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
4034 struct bfq_data *bfqd = bfqq->bfqd;
4035 bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
4037 if (unlikely(bfqd->strict_guarantees))
4041 * Idling is performed only if slice_idle > 0. In addition, we
4044 * (b) bfqq is in the idle io prio class: in this case we do
4045 * not idle because we want to minimize the bandwidth that
4046 * queues in this class can steal to higher-priority queues
4048 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
4049 bfq_class_idle(bfqq))
4052 idling_boosts_thr_with_no_issue =
4053 idling_boosts_thr_without_issues(bfqd, bfqq);
4055 idling_needed_for_service_guar =
4056 idling_needed_for_service_guarantees(bfqd, bfqq);
4059 * We have now the two components we need to compute the
4060 * return value of the function, which is true only if idling
4061 * either boosts the throughput (without issues), or is
4062 * necessary to preserve service guarantees.
4064 return idling_boosts_thr_with_no_issue ||
4065 idling_needed_for_service_guar;
4069 * If the in-service queue is empty but the function bfq_better_to_idle
4070 * returns true, then:
4071 * 1) the queue must remain in service and cannot be expired, and
4072 * 2) the device must be idled to wait for the possible arrival of a new
4073 * request for the queue.
4074 * See the comments on the function bfq_better_to_idle for the reasons
4075 * why performing device idling is the best choice to boost the throughput
4076 * and preserve service guarantees when bfq_better_to_idle itself
4079 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
4081 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
4085 * This function chooses the queue from which to pick the next extra
4086 * I/O request to inject, if it finds a compatible queue. See the
4087 * comments on bfq_update_inject_limit() for details on the injection
4088 * mechanism, and for the definitions of the quantities mentioned
4091 static struct bfq_queue *
4092 bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
4094 struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
4095 unsigned int limit = in_serv_bfqq->inject_limit;
4098 * - bfqq is not weight-raised and therefore does not carry
4099 * time-critical I/O,
4101 * - regardless of whether bfqq is weight-raised, bfqq has
4102 * however a long think time, during which it can absorb the
4103 * effect of an appropriate number of extra I/O requests
4104 * from other queues (see bfq_update_inject_limit for
4105 * details on the computation of this number);
4106 * then injection can be performed without restrictions.
4108 bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
4109 !bfq_bfqq_has_short_ttime(in_serv_bfqq);
4113 * - the baseline total service time could not be sampled yet,
4114 * so the inject limit happens to be still 0, and
4115 * - a lot of time has elapsed since the plugging of I/O
4116 * dispatching started, so drive speed is being wasted
4118 * then temporarily raise inject limit to one request.
4120 if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
4121 bfq_bfqq_wait_request(in_serv_bfqq) &&
4122 time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
4123 bfqd->bfq_slice_idle)
4127 if (bfqd->rq_in_driver >= limit)
4131 * Linear search of the source queue for injection; but, with
4132 * a high probability, very few steps are needed to find a
4133 * candidate queue, i.e., a queue with enough budget left for
4134 * its next request. In fact:
4135 * - BFQ dynamically updates the budget of every queue so as
4136 * to accommodate the expected backlog of the queue;
4137 * - if a queue gets all its requests dispatched as injected
4138 * service, then the queue is removed from the active list
4139 * (and re-added only if it gets new requests, but then it
4140 * is assigned again enough budget for its new backlog).
4142 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
4143 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4144 (in_serv_always_inject || bfqq->wr_coeff > 1) &&
4145 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4146 bfq_bfqq_budget_left(bfqq)) {
4148 * Allow for only one large in-flight request
4149 * on non-rotational devices, for the
4150 * following reason. On non-rotationl drives,
4151 * large requests take much longer than
4152 * smaller requests to be served. In addition,
4153 * the drive prefers to serve large requests
4154 * w.r.t. to small ones, if it can choose. So,
4155 * having more than one large requests queued
4156 * in the drive may easily make the next first
4157 * request of the in-service queue wait for so
4158 * long to break bfqq's service guarantees. On
4159 * the bright side, large requests let the
4160 * drive reach a very high throughput, even if
4161 * there is only one in-flight large request
4164 if (blk_queue_nonrot(bfqd->queue) &&
4165 blk_rq_sectors(bfqq->next_rq) >=
4166 BFQQ_SECT_THR_NONROT)
4167 limit = min_t(unsigned int, 1, limit);
4169 limit = in_serv_bfqq->inject_limit;
4171 if (bfqd->rq_in_driver < limit) {
4172 bfqd->rqs_injected = true;
4181 * Select a queue for service. If we have a current queue in service,
4182 * check whether to continue servicing it, or retrieve and set a new one.
4184 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
4186 struct bfq_queue *bfqq;
4187 struct request *next_rq;
4188 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
4190 bfqq = bfqd->in_service_queue;
4194 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
4197 * Do not expire bfqq for budget timeout if bfqq may be about
4198 * to enjoy device idling. The reason why, in this case, we
4199 * prevent bfqq from expiring is the same as in the comments
4200 * on the case where bfq_bfqq_must_idle() returns true, in
4201 * bfq_completed_request().
4203 if (bfq_may_expire_for_budg_timeout(bfqq) &&
4204 !bfq_bfqq_must_idle(bfqq))
4209 * This loop is rarely executed more than once. Even when it
4210 * happens, it is much more convenient to re-execute this loop
4211 * than to return NULL and trigger a new dispatch to get a
4214 next_rq = bfqq->next_rq;
4216 * If bfqq has requests queued and it has enough budget left to
4217 * serve them, keep the queue, otherwise expire it.
4220 if (bfq_serv_to_charge(next_rq, bfqq) >
4221 bfq_bfqq_budget_left(bfqq)) {
4223 * Expire the queue for budget exhaustion,
4224 * which makes sure that the next budget is
4225 * enough to serve the next request, even if
4226 * it comes from the fifo expired path.
4228 reason = BFQQE_BUDGET_EXHAUSTED;
4232 * The idle timer may be pending because we may
4233 * not disable disk idling even when a new request
4236 if (bfq_bfqq_wait_request(bfqq)) {
4238 * If we get here: 1) at least a new request
4239 * has arrived but we have not disabled the
4240 * timer because the request was too small,
4241 * 2) then the block layer has unplugged
4242 * the device, causing the dispatch to be
4245 * Since the device is unplugged, now the
4246 * requests are probably large enough to
4247 * provide a reasonable throughput.
4248 * So we disable idling.
4250 bfq_clear_bfqq_wait_request(bfqq);
4251 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4258 * No requests pending. However, if the in-service queue is idling
4259 * for a new request, or has requests waiting for a completion and
4260 * may idle after their completion, then keep it anyway.
4262 * Yet, inject service from other queues if it boosts
4263 * throughput and is possible.
4265 if (bfq_bfqq_wait_request(bfqq) ||
4266 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
4267 struct bfq_queue *async_bfqq =
4268 bfqq->bic && bfqq->bic->bfqq[0] &&
4269 bfq_bfqq_busy(bfqq->bic->bfqq[0]) ?
4270 bfqq->bic->bfqq[0] : NULL;
4273 * The next three mutually-exclusive ifs decide
4274 * whether to try injection, and choose the queue to
4275 * pick an I/O request from.
4277 * The first if checks whether the process associated
4278 * with bfqq has also async I/O pending. If so, it
4279 * injects such I/O unconditionally. Injecting async
4280 * I/O from the same process can cause no harm to the
4281 * process. On the contrary, it can only increase
4282 * bandwidth and reduce latency for the process.
4284 * The second if checks whether there happens to be a
4285 * non-empty waker queue for bfqq, i.e., a queue whose
4286 * I/O needs to be completed for bfqq to receive new
4287 * I/O. This happens, e.g., if bfqq is associated with
4288 * a process that does some sync. A sync generates
4289 * extra blocking I/O, which must be completed before
4290 * the process associated with bfqq can go on with its
4291 * I/O. If the I/O of the waker queue is not served,
4292 * then bfqq remains empty, and no I/O is dispatched,
4293 * until the idle timeout fires for bfqq. This is
4294 * likely to result in lower bandwidth and higher
4295 * latencies for bfqq, and in a severe loss of total
4296 * throughput. The best action to take is therefore to
4297 * serve the waker queue as soon as possible. So do it
4298 * (without relying on the third alternative below for
4299 * eventually serving waker_bfqq's I/O; see the last
4300 * paragraph for further details). This systematic
4301 * injection of I/O from the waker queue does not
4302 * cause any delay to bfqq's I/O. On the contrary,
4303 * next bfqq's I/O is brought forward dramatically,
4304 * for it is not blocked for milliseconds.
4306 * The third if checks whether bfqq is a queue for
4307 * which it is better to avoid injection. It is so if
4308 * bfqq delivers more throughput when served without
4309 * any further I/O from other queues in the middle, or
4310 * if the service times of bfqq's I/O requests both
4311 * count more than overall throughput, and may be
4312 * easily increased by injection (this happens if bfqq
4313 * has a short think time). If none of these
4314 * conditions holds, then a candidate queue for
4315 * injection is looked for through
4316 * bfq_choose_bfqq_for_injection(). Note that the
4317 * latter may return NULL (for example if the inject
4318 * limit for bfqq is currently 0).
4320 * NOTE: motivation for the second alternative
4322 * Thanks to the way the inject limit is updated in
4323 * bfq_update_has_short_ttime(), it is rather likely
4324 * that, if I/O is being plugged for bfqq and the
4325 * waker queue has pending I/O requests that are
4326 * blocking bfqq's I/O, then the third alternative
4327 * above lets the waker queue get served before the
4328 * I/O-plugging timeout fires. So one may deem the
4329 * second alternative superfluous. It is not, because
4330 * the third alternative may be way less effective in
4331 * case of a synchronization. For two main
4332 * reasons. First, throughput may be low because the
4333 * inject limit may be too low to guarantee the same
4334 * amount of injected I/O, from the waker queue or
4335 * other queues, that the second alternative
4336 * guarantees (the second alternative unconditionally
4337 * injects a pending I/O request of the waker queue
4338 * for each bfq_dispatch_request()). Second, with the
4339 * third alternative, the duration of the plugging,
4340 * i.e., the time before bfqq finally receives new I/O,
4341 * may not be minimized, because the waker queue may
4342 * happen to be served only after other queues.
4345 icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
4346 bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
4347 bfq_bfqq_budget_left(async_bfqq))
4348 bfqq = bfqq->bic->bfqq[0];
4349 else if (bfq_bfqq_has_waker(bfqq) &&
4350 bfq_bfqq_busy(bfqq->waker_bfqq) &&
4351 bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
4352 bfqq->waker_bfqq) <=
4353 bfq_bfqq_budget_left(bfqq->waker_bfqq)
4355 bfqq = bfqq->waker_bfqq;
4356 else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
4357 (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
4358 !bfq_bfqq_has_short_ttime(bfqq)))
4359 bfqq = bfq_choose_bfqq_for_injection(bfqd);
4366 reason = BFQQE_NO_MORE_REQUESTS;
4368 bfq_bfqq_expire(bfqd, bfqq, false, reason);
4370 bfqq = bfq_set_in_service_queue(bfqd);
4372 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
4377 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
4379 bfq_log(bfqd, "select_queue: no queue returned");
4384 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4386 struct bfq_entity *entity = &bfqq->entity;
4388 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
4389 bfq_log_bfqq(bfqd, bfqq,
4390 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
4391 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
4392 jiffies_to_msecs(bfqq->wr_cur_max_time),
4394 bfqq->entity.weight, bfqq->entity.orig_weight);
4396 if (entity->prio_changed)
4397 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
4400 * If the queue was activated in a burst, or too much
4401 * time has elapsed from the beginning of this
4402 * weight-raising period, then end weight raising.
4404 if (bfq_bfqq_in_large_burst(bfqq))
4405 bfq_bfqq_end_wr(bfqq);
4406 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
4407 bfqq->wr_cur_max_time)) {
4408 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
4409 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
4410 bfq_wr_duration(bfqd)))
4411 bfq_bfqq_end_wr(bfqq);
4413 switch_back_to_interactive_wr(bfqq, bfqd);
4414 bfqq->entity.prio_changed = 1;
4417 if (bfqq->wr_coeff > 1 &&
4418 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
4419 bfqq->service_from_wr > max_service_from_wr) {
4420 /* see comments on max_service_from_wr */
4421 bfq_bfqq_end_wr(bfqq);
4425 * To improve latency (for this or other queues), immediately
4426 * update weight both if it must be raised and if it must be
4427 * lowered. Since, entity may be on some active tree here, and
4428 * might have a pending change of its ioprio class, invoke
4429 * next function with the last parameter unset (see the
4430 * comments on the function).
4432 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
4433 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
4438 * Dispatch next request from bfqq.
4440 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
4441 struct bfq_queue *bfqq)
4443 struct request *rq = bfqq->next_rq;
4444 unsigned long service_to_charge;
4446 service_to_charge = bfq_serv_to_charge(rq, bfqq);
4448 bfq_bfqq_served(bfqq, service_to_charge);
4450 if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
4451 bfqd->wait_dispatch = false;
4452 bfqd->waited_rq = rq;
4455 bfq_dispatch_remove(bfqd->queue, rq);
4457 if (bfqq != bfqd->in_service_queue)
4461 * If weight raising has to terminate for bfqq, then next
4462 * function causes an immediate update of bfqq's weight,
4463 * without waiting for next activation. As a consequence, on
4464 * expiration, bfqq will be timestamped as if has never been
4465 * weight-raised during this service slot, even if it has
4466 * received part or even most of the service as a
4467 * weight-raised queue. This inflates bfqq's timestamps, which
4468 * is beneficial, as bfqq is then more willing to leave the
4469 * device immediately to possible other weight-raised queues.
4471 bfq_update_wr_data(bfqd, bfqq);
4474 * Expire bfqq, pretending that its budget expired, if bfqq
4475 * belongs to CLASS_IDLE and other queues are waiting for
4478 if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
4481 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
4487 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
4489 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4492 * Avoiding lock: a race on bfqd->busy_queues should cause at
4493 * most a call to dispatch for nothing
4495 return !list_empty_careful(&bfqd->dispatch) ||
4496 bfq_tot_busy_queues(bfqd) > 0;
4499 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
4501 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4502 struct request *rq = NULL;
4503 struct bfq_queue *bfqq = NULL;
4505 if (!list_empty(&bfqd->dispatch)) {
4506 rq = list_first_entry(&bfqd->dispatch, struct request,
4508 list_del_init(&rq->queuelist);
4514 * Increment counters here, because this
4515 * dispatch does not follow the standard
4516 * dispatch flow (where counters are
4521 goto inc_in_driver_start_rq;
4525 * We exploit the bfq_finish_requeue_request hook to
4526 * decrement rq_in_driver, but
4527 * bfq_finish_requeue_request will not be invoked on
4528 * this request. So, to avoid unbalance, just start
4529 * this request, without incrementing rq_in_driver. As
4530 * a negative consequence, rq_in_driver is deceptively
4531 * lower than it should be while this request is in
4532 * service. This may cause bfq_schedule_dispatch to be
4533 * invoked uselessly.
4535 * As for implementing an exact solution, the
4536 * bfq_finish_requeue_request hook, if defined, is
4537 * probably invoked also on this request. So, by
4538 * exploiting this hook, we could 1) increment
4539 * rq_in_driver here, and 2) decrement it in
4540 * bfq_finish_requeue_request. Such a solution would
4541 * let the value of the counter be always accurate,
4542 * but it would entail using an extra interface
4543 * function. This cost seems higher than the benefit,
4544 * being the frequency of non-elevator-private
4545 * requests very low.
4550 bfq_log(bfqd, "dispatch requests: %d busy queues",
4551 bfq_tot_busy_queues(bfqd));
4553 if (bfq_tot_busy_queues(bfqd) == 0)
4557 * Force device to serve one request at a time if
4558 * strict_guarantees is true. Forcing this service scheme is
4559 * currently the ONLY way to guarantee that the request
4560 * service order enforced by the scheduler is respected by a
4561 * queueing device. Otherwise the device is free even to make
4562 * some unlucky request wait for as long as the device
4565 * Of course, serving one request at at time may cause loss of
4568 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
4571 bfqq = bfq_select_queue(bfqd);
4575 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
4578 inc_in_driver_start_rq:
4579 bfqd->rq_in_driver++;
4581 rq->rq_flags |= RQF_STARTED;
4587 #ifdef CONFIG_BFQ_CGROUP_DEBUG
4588 static void bfq_update_dispatch_stats(struct request_queue *q,
4590 struct bfq_queue *in_serv_queue,
4591 bool idle_timer_disabled)
4593 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
4595 if (!idle_timer_disabled && !bfqq)
4599 * rq and bfqq are guaranteed to exist until this function
4600 * ends, for the following reasons. First, rq can be
4601 * dispatched to the device, and then can be completed and
4602 * freed, only after this function ends. Second, rq cannot be
4603 * merged (and thus freed because of a merge) any longer,
4604 * because it has already started. Thus rq cannot be freed
4605 * before this function ends, and, since rq has a reference to
4606 * bfqq, the same guarantee holds for bfqq too.
4608 * In addition, the following queue lock guarantees that
4609 * bfqq_group(bfqq) exists as well.
4611 spin_lock_irq(&q->queue_lock);
4612 if (idle_timer_disabled)
4614 * Since the idle timer has been disabled,
4615 * in_serv_queue contained some request when
4616 * __bfq_dispatch_request was invoked above, which
4617 * implies that rq was picked exactly from
4618 * in_serv_queue. Thus in_serv_queue == bfqq, and is
4619 * therefore guaranteed to exist because of the above
4622 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
4624 struct bfq_group *bfqg = bfqq_group(bfqq);
4626 bfqg_stats_update_avg_queue_size(bfqg);
4627 bfqg_stats_set_start_empty_time(bfqg);
4628 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
4630 spin_unlock_irq(&q->queue_lock);
4633 static inline void bfq_update_dispatch_stats(struct request_queue *q,
4635 struct bfq_queue *in_serv_queue,
4636 bool idle_timer_disabled) {}
4637 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
4639 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
4641 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4643 struct bfq_queue *in_serv_queue;
4644 bool waiting_rq, idle_timer_disabled;
4646 spin_lock_irq(&bfqd->lock);
4648 in_serv_queue = bfqd->in_service_queue;
4649 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
4651 rq = __bfq_dispatch_request(hctx);
4653 idle_timer_disabled =
4654 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
4656 spin_unlock_irq(&bfqd->lock);
4658 bfq_update_dispatch_stats(hctx->queue, rq, in_serv_queue,
4659 idle_timer_disabled);
4665 * Task holds one reference to the queue, dropped when task exits. Each rq
4666 * in-flight on this queue also holds a reference, dropped when rq is freed.
4668 * Scheduler lock must be held here. Recall not to use bfqq after calling
4669 * this function on it.
4671 void bfq_put_queue(struct bfq_queue *bfqq)
4673 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4674 struct bfq_group *bfqg = bfqq_group(bfqq);
4678 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
4685 if (!hlist_unhashed(&bfqq->burst_list_node)) {
4686 hlist_del_init(&bfqq->burst_list_node);
4688 * Decrement also burst size after the removal, if the
4689 * process associated with bfqq is exiting, and thus
4690 * does not contribute to the burst any longer. This
4691 * decrement helps filter out false positives of large
4692 * bursts, when some short-lived process (often due to
4693 * the execution of commands by some service) happens
4694 * to start and exit while a complex application is
4695 * starting, and thus spawning several processes that
4696 * do I/O (and that *must not* be treated as a large
4697 * burst, see comments on bfq_handle_burst).
4699 * In particular, the decrement is performed only if:
4700 * 1) bfqq is not a merged queue, because, if it is,
4701 * then this free of bfqq is not triggered by the exit
4702 * of the process bfqq is associated with, but exactly
4703 * by the fact that bfqq has just been merged.
4704 * 2) burst_size is greater than 0, to handle
4705 * unbalanced decrements. Unbalanced decrements may
4706 * happen in te following case: bfqq is inserted into
4707 * the current burst list--without incrementing
4708 * bust_size--because of a split, but the current
4709 * burst list is not the burst list bfqq belonged to
4710 * (see comments on the case of a split in
4713 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
4714 bfqq->bfqd->burst_size--;
4717 kmem_cache_free(bfq_pool, bfqq);
4718 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4719 bfqg_and_blkg_put(bfqg);
4723 static void bfq_put_cooperator(struct bfq_queue *bfqq)
4725 struct bfq_queue *__bfqq, *next;
4728 * If this queue was scheduled to merge with another queue, be
4729 * sure to drop the reference taken on that queue (and others in
4730 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
4732 __bfqq = bfqq->new_bfqq;
4736 next = __bfqq->new_bfqq;
4737 bfq_put_queue(__bfqq);
4742 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4744 struct bfq_queue *item;
4745 struct hlist_node *n;
4747 if (bfqq == bfqd->in_service_queue) {
4748 __bfq_bfqq_expire(bfqd, bfqq);
4749 bfq_schedule_dispatch(bfqd);
4752 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
4754 bfq_put_cooperator(bfqq);
4756 /* remove bfqq from woken list */
4757 if (!hlist_unhashed(&bfqq->woken_list_node))
4758 hlist_del_init(&bfqq->woken_list_node);
4760 /* reset waker for all queues in woken list */
4761 hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
4763 item->waker_bfqq = NULL;
4764 bfq_clear_bfqq_has_waker(item);
4765 hlist_del_init(&item->woken_list_node);
4768 bfq_put_queue(bfqq); /* release process reference */
4771 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
4773 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4774 struct bfq_data *bfqd;
4777 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
4780 unsigned long flags;
4782 spin_lock_irqsave(&bfqd->lock, flags);
4783 bfq_exit_bfqq(bfqd, bfqq);
4784 bic_set_bfqq(bic, NULL, is_sync);
4785 spin_unlock_irqrestore(&bfqd->lock, flags);
4789 static void bfq_exit_icq(struct io_cq *icq)
4791 struct bfq_io_cq *bic = icq_to_bic(icq);
4793 bfq_exit_icq_bfqq(bic, true);
4794 bfq_exit_icq_bfqq(bic, false);
4798 * Update the entity prio values; note that the new values will not
4799 * be used until the next (re)activation.
4802 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
4804 struct task_struct *tsk = current;
4806 struct bfq_data *bfqd = bfqq->bfqd;
4811 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4812 switch (ioprio_class) {
4814 dev_err(bfqq->bfqd->queue->backing_dev_info->dev,
4815 "bfq: bad prio class %d\n", ioprio_class);
4817 case IOPRIO_CLASS_NONE:
4819 * No prio set, inherit CPU scheduling settings.
4821 bfqq->new_ioprio = task_nice_ioprio(tsk);
4822 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
4824 case IOPRIO_CLASS_RT:
4825 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4826 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
4828 case IOPRIO_CLASS_BE:
4829 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4830 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
4832 case IOPRIO_CLASS_IDLE:
4833 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
4834 bfqq->new_ioprio = 7;
4838 if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
4839 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
4841 bfqq->new_ioprio = IOPRIO_BE_NR;
4844 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
4845 bfqq->entity.prio_changed = 1;
4848 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4849 struct bio *bio, bool is_sync,
4850 struct bfq_io_cq *bic);
4852 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
4854 struct bfq_data *bfqd = bic_to_bfqd(bic);
4855 struct bfq_queue *bfqq;
4856 int ioprio = bic->icq.ioc->ioprio;
4859 * This condition may trigger on a newly created bic, be sure to
4860 * drop the lock before returning.
4862 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
4865 bic->ioprio = ioprio;
4867 bfqq = bic_to_bfqq(bic, false);
4869 /* release process reference on this queue */
4870 bfq_put_queue(bfqq);
4871 bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
4872 bic_set_bfqq(bic, bfqq, false);
4875 bfqq = bic_to_bfqq(bic, true);
4877 bfq_set_next_ioprio_data(bfqq, bic);
4880 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4881 struct bfq_io_cq *bic, pid_t pid, int is_sync)
4883 RB_CLEAR_NODE(&bfqq->entity.rb_node);
4884 INIT_LIST_HEAD(&bfqq->fifo);
4885 INIT_HLIST_NODE(&bfqq->burst_list_node);
4886 INIT_HLIST_NODE(&bfqq->woken_list_node);
4887 INIT_HLIST_HEAD(&bfqq->woken_list);
4893 bfq_set_next_ioprio_data(bfqq, bic);
4897 * No need to mark as has_short_ttime if in
4898 * idle_class, because no device idling is performed
4899 * for queues in idle class
4901 if (!bfq_class_idle(bfqq))
4902 /* tentatively mark as has_short_ttime */
4903 bfq_mark_bfqq_has_short_ttime(bfqq);
4904 bfq_mark_bfqq_sync(bfqq);
4905 bfq_mark_bfqq_just_created(bfqq);
4907 bfq_clear_bfqq_sync(bfqq);
4909 /* set end request to minus infinity from now */
4910 bfqq->ttime.last_end_request = ktime_get_ns() + 1;
4912 bfq_mark_bfqq_IO_bound(bfqq);
4916 /* Tentative initial value to trade off between thr and lat */
4917 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
4918 bfqq->budget_timeout = bfq_smallest_from_now();
4921 bfqq->last_wr_start_finish = jiffies;
4922 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
4923 bfqq->split_time = bfq_smallest_from_now();
4926 * To not forget the possibly high bandwidth consumed by a
4927 * process/queue in the recent past,
4928 * bfq_bfqq_softrt_next_start() returns a value at least equal
4929 * to the current value of bfqq->soft_rt_next_start (see
4930 * comments on bfq_bfqq_softrt_next_start). Set
4931 * soft_rt_next_start to now, to mean that bfqq has consumed
4932 * no bandwidth so far.
4934 bfqq->soft_rt_next_start = jiffies;
4936 /* first request is almost certainly seeky */
4937 bfqq->seek_history = 1;
4940 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
4941 struct bfq_group *bfqg,
4942 int ioprio_class, int ioprio)
4944 switch (ioprio_class) {
4945 case IOPRIO_CLASS_RT:
4946 return &bfqg->async_bfqq[0][ioprio];
4947 case IOPRIO_CLASS_NONE:
4948 ioprio = IOPRIO_NORM;
4950 case IOPRIO_CLASS_BE:
4951 return &bfqg->async_bfqq[1][ioprio];
4952 case IOPRIO_CLASS_IDLE:
4953 return &bfqg->async_idle_bfqq;
4959 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4960 struct bio *bio, bool is_sync,
4961 struct bfq_io_cq *bic)
4963 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4964 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4965 struct bfq_queue **async_bfqq = NULL;
4966 struct bfq_queue *bfqq;
4967 struct bfq_group *bfqg;
4971 bfqg = bfq_find_set_group(bfqd, __bio_blkcg(bio));
4973 bfqq = &bfqd->oom_bfqq;
4978 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
4985 bfqq = kmem_cache_alloc_node(bfq_pool,
4986 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
4990 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
4992 bfq_init_entity(&bfqq->entity, bfqg);
4993 bfq_log_bfqq(bfqd, bfqq, "allocated");
4995 bfqq = &bfqd->oom_bfqq;
4996 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
5001 * Pin the queue now that it's allocated, scheduler exit will
5006 * Extra group reference, w.r.t. sync
5007 * queue. This extra reference is removed
5008 * only if bfqq->bfqg disappears, to
5009 * guarantee that this queue is not freed
5010 * until its group goes away.
5012 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
5018 bfqq->ref++; /* get a process reference to this queue */
5019 bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
5024 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
5025 struct bfq_queue *bfqq)
5027 struct bfq_ttime *ttime = &bfqq->ttime;
5028 u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
5030 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
5032 ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
5033 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
5034 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
5035 ttime->ttime_samples);
5039 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5042 bfqq->seek_history <<= 1;
5043 bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
5045 if (bfqq->wr_coeff > 1 &&
5046 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
5047 BFQQ_TOTALLY_SEEKY(bfqq))
5048 bfq_bfqq_end_wr(bfqq);
5051 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
5052 struct bfq_queue *bfqq,
5053 struct bfq_io_cq *bic)
5055 bool has_short_ttime = true, state_changed;
5058 * No need to update has_short_ttime if bfqq is async or in
5059 * idle io prio class, or if bfq_slice_idle is zero, because
5060 * no device idling is performed for bfqq in this case.
5062 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
5063 bfqd->bfq_slice_idle == 0)
5066 /* Idle window just restored, statistics are meaningless. */
5067 if (time_is_after_eq_jiffies(bfqq->split_time +
5068 bfqd->bfq_wr_min_idle_time))
5071 /* Think time is infinite if no process is linked to
5072 * bfqq. Otherwise check average think time to
5073 * decide whether to mark as has_short_ttime
5075 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
5076 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
5077 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle))
5078 has_short_ttime = false;
5080 state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
5082 if (has_short_ttime)
5083 bfq_mark_bfqq_has_short_ttime(bfqq);
5085 bfq_clear_bfqq_has_short_ttime(bfqq);
5088 * Until the base value for the total service time gets
5089 * finally computed for bfqq, the inject limit does depend on
5090 * the think-time state (short|long). In particular, the limit
5091 * is 0 or 1 if the think time is deemed, respectively, as
5092 * short or long (details in the comments in
5093 * bfq_update_inject_limit()). Accordingly, the next
5094 * instructions reset the inject limit if the think-time state
5095 * has changed and the above base value is still to be
5098 * However, the reset is performed only if more than 100 ms
5099 * have elapsed since the last update of the inject limit, or
5100 * (inclusive) if the change is from short to long think
5101 * time. The reason for this waiting is as follows.
5103 * bfqq may have a long think time because of a
5104 * synchronization with some other queue, i.e., because the
5105 * I/O of some other queue may need to be completed for bfqq
5106 * to receive new I/O. Details in the comments on the choice
5107 * of the queue for injection in bfq_select_queue().
5109 * As stressed in those comments, if such a synchronization is
5110 * actually in place, then, without injection on bfqq, the
5111 * blocking I/O cannot happen to served while bfqq is in
5112 * service. As a consequence, if bfqq is granted
5113 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
5114 * is dispatched, until the idle timeout fires. This is likely
5115 * to result in lower bandwidth and higher latencies for bfqq,
5116 * and in a severe loss of total throughput.
5118 * On the opposite end, a non-zero inject limit may allow the
5119 * I/O that blocks bfqq to be executed soon, and therefore
5120 * bfqq to receive new I/O soon.
5122 * But, if the blocking gets actually eliminated, then the
5123 * next think-time sample for bfqq may be very low. This in
5124 * turn may cause bfqq's think time to be deemed
5125 * short. Without the 100 ms barrier, this new state change
5126 * would cause the body of the next if to be executed
5127 * immediately. But this would set to 0 the inject
5128 * limit. Without injection, the blocking I/O would cause the
5129 * think time of bfqq to become long again, and therefore the
5130 * inject limit to be raised again, and so on. The only effect
5131 * of such a steady oscillation between the two think-time
5132 * states would be to prevent effective injection on bfqq.
5134 * In contrast, if the inject limit is not reset during such a
5135 * long time interval as 100 ms, then the number of short
5136 * think time samples can grow significantly before the reset
5137 * is performed. As a consequence, the think time state can
5138 * become stable before the reset. Therefore there will be no
5139 * state change when the 100 ms elapse, and no reset of the
5140 * inject limit. The inject limit remains steadily equal to 1
5141 * both during and after the 100 ms. So injection can be
5142 * performed at all times, and throughput gets boosted.
5144 * An inject limit equal to 1 is however in conflict, in
5145 * general, with the fact that the think time of bfqq is
5146 * short, because injection may be likely to delay bfqq's I/O
5147 * (as explained in the comments in
5148 * bfq_update_inject_limit()). But this does not happen in
5149 * this special case, because bfqq's low think time is due to
5150 * an effective handling of a synchronization, through
5151 * injection. In this special case, bfqq's I/O does not get
5152 * delayed by injection; on the contrary, bfqq's I/O is
5153 * brought forward, because it is not blocked for
5156 * In addition, serving the blocking I/O much sooner, and much
5157 * more frequently than once per I/O-plugging timeout, makes
5158 * it much quicker to detect a waker queue (the concept of
5159 * waker queue is defined in the comments in
5160 * bfq_add_request()). This makes it possible to start sooner
5161 * to boost throughput more effectively, by injecting the I/O
5162 * of the waker queue unconditionally on every
5163 * bfq_dispatch_request().
5165 * One last, important benefit of not resetting the inject
5166 * limit before 100 ms is that, during this time interval, the
5167 * base value for the total service time is likely to get
5168 * finally computed for bfqq, freeing the inject limit from
5169 * its relation with the think time.
5171 if (state_changed && bfqq->last_serv_time_ns == 0 &&
5172 (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
5173 msecs_to_jiffies(100)) ||
5175 bfq_reset_inject_limit(bfqd, bfqq);
5179 * Called when a new fs request (rq) is added to bfqq. Check if there's
5180 * something we should do about it.
5182 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5185 if (rq->cmd_flags & REQ_META)
5186 bfqq->meta_pending++;
5188 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
5190 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
5191 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
5192 blk_rq_sectors(rq) < 32;
5193 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
5196 * There is just this request queued: if
5197 * - the request is small, and
5198 * - we are idling to boost throughput, and
5199 * - the queue is not to be expired,
5202 * In this way, if the device is being idled to wait
5203 * for a new request from the in-service queue, we
5204 * avoid unplugging the device and committing the
5205 * device to serve just a small request. In contrast
5206 * we wait for the block layer to decide when to
5207 * unplug the device: hopefully, new requests will be
5208 * merged to this one quickly, then the device will be
5209 * unplugged and larger requests will be dispatched.
5211 if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
5216 * A large enough request arrived, or idling is being
5217 * performed to preserve service guarantees, or
5218 * finally the queue is to be expired: in all these
5219 * cases disk idling is to be stopped, so clear
5220 * wait_request flag and reset timer.
5222 bfq_clear_bfqq_wait_request(bfqq);
5223 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
5226 * The queue is not empty, because a new request just
5227 * arrived. Hence we can safely expire the queue, in
5228 * case of budget timeout, without risking that the
5229 * timestamps of the queue are not updated correctly.
5230 * See [1] for more details.
5233 bfq_bfqq_expire(bfqd, bfqq, false,
5234 BFQQE_BUDGET_TIMEOUT);
5238 /* returns true if it causes the idle timer to be disabled */
5239 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
5241 struct bfq_queue *bfqq = RQ_BFQQ(rq),
5242 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
5243 bool waiting, idle_timer_disabled = false;
5247 * Release the request's reference to the old bfqq
5248 * and make sure one is taken to the shared queue.
5250 new_bfqq->allocated++;
5254 * If the bic associated with the process
5255 * issuing this request still points to bfqq
5256 * (and thus has not been already redirected
5257 * to new_bfqq or even some other bfq_queue),
5258 * then complete the merge and redirect it to
5261 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
5262 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
5265 bfq_clear_bfqq_just_created(bfqq);
5267 * rq is about to be enqueued into new_bfqq,
5268 * release rq reference on bfqq
5270 bfq_put_queue(bfqq);
5271 rq->elv.priv[1] = new_bfqq;
5275 bfq_update_io_thinktime(bfqd, bfqq);
5276 bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
5277 bfq_update_io_seektime(bfqd, bfqq, rq);
5279 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
5280 bfq_add_request(rq);
5281 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
5283 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
5284 list_add_tail(&rq->queuelist, &bfqq->fifo);
5286 bfq_rq_enqueued(bfqd, bfqq, rq);
5288 return idle_timer_disabled;
5291 #ifdef CONFIG_BFQ_CGROUP_DEBUG
5292 static void bfq_update_insert_stats(struct request_queue *q,
5293 struct bfq_queue *bfqq,
5294 bool idle_timer_disabled,
5295 unsigned int cmd_flags)
5301 * bfqq still exists, because it can disappear only after
5302 * either it is merged with another queue, or the process it
5303 * is associated with exits. But both actions must be taken by
5304 * the same process currently executing this flow of
5307 * In addition, the following queue lock guarantees that
5308 * bfqq_group(bfqq) exists as well.
5310 spin_lock_irq(&q->queue_lock);
5311 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
5312 if (idle_timer_disabled)
5313 bfqg_stats_update_idle_time(bfqq_group(bfqq));
5314 spin_unlock_irq(&q->queue_lock);
5317 static inline void bfq_update_insert_stats(struct request_queue *q,
5318 struct bfq_queue *bfqq,
5319 bool idle_timer_disabled,
5320 unsigned int cmd_flags) {}
5321 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
5323 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
5326 struct request_queue *q = hctx->queue;
5327 struct bfq_data *bfqd = q->elevator->elevator_data;
5328 struct bfq_queue *bfqq;
5329 bool idle_timer_disabled = false;
5330 unsigned int cmd_flags;
5332 spin_lock_irq(&bfqd->lock);
5333 if (blk_mq_sched_try_insert_merge(q, rq)) {
5334 spin_unlock_irq(&bfqd->lock);
5338 spin_unlock_irq(&bfqd->lock);
5340 blk_mq_sched_request_inserted(rq);
5342 spin_lock_irq(&bfqd->lock);
5343 bfqq = bfq_init_rq(rq);
5344 if (at_head || blk_rq_is_passthrough(rq)) {
5346 list_add(&rq->queuelist, &bfqd->dispatch);
5348 list_add_tail(&rq->queuelist, &bfqd->dispatch);
5349 } else { /* bfqq is assumed to be non null here */
5350 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
5352 * Update bfqq, because, if a queue merge has occurred
5353 * in __bfq_insert_request, then rq has been
5354 * redirected into a new queue.
5358 if (rq_mergeable(rq)) {
5359 elv_rqhash_add(q, rq);
5366 * Cache cmd_flags before releasing scheduler lock, because rq
5367 * may disappear afterwards (for example, because of a request
5370 cmd_flags = rq->cmd_flags;
5372 spin_unlock_irq(&bfqd->lock);
5374 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
5378 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
5379 struct list_head *list, bool at_head)
5381 while (!list_empty(list)) {
5384 rq = list_first_entry(list, struct request, queuelist);
5385 list_del_init(&rq->queuelist);
5386 bfq_insert_request(hctx, rq, at_head);
5390 static void bfq_update_hw_tag(struct bfq_data *bfqd)
5392 struct bfq_queue *bfqq = bfqd->in_service_queue;
5394 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
5395 bfqd->rq_in_driver);
5397 if (bfqd->hw_tag == 1)
5401 * This sample is valid if the number of outstanding requests
5402 * is large enough to allow a queueing behavior. Note that the
5403 * sum is not exact, as it's not taking into account deactivated
5406 if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
5410 * If active queue hasn't enough requests and can idle, bfq might not
5411 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
5414 if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
5415 bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
5416 BFQ_HW_QUEUE_THRESHOLD &&
5417 bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
5420 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
5423 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
5424 bfqd->max_rq_in_driver = 0;
5425 bfqd->hw_tag_samples = 0;
5427 bfqd->nonrot_with_queueing =
5428 blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
5431 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
5436 bfq_update_hw_tag(bfqd);
5438 bfqd->rq_in_driver--;
5441 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
5443 * Set budget_timeout (which we overload to store the
5444 * time at which the queue remains with no backlog and
5445 * no outstanding request; used by the weight-raising
5448 bfqq->budget_timeout = jiffies;
5450 bfq_weights_tree_remove(bfqd, bfqq);
5453 now_ns = ktime_get_ns();
5455 bfqq->ttime.last_end_request = now_ns;
5458 * Using us instead of ns, to get a reasonable precision in
5459 * computing rate in next check.
5461 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
5464 * If the request took rather long to complete, and, according
5465 * to the maximum request size recorded, this completion latency
5466 * implies that the request was certainly served at a very low
5467 * rate (less than 1M sectors/sec), then the whole observation
5468 * interval that lasts up to this time instant cannot be a
5469 * valid time interval for computing a new peak rate. Invoke
5470 * bfq_update_rate_reset to have the following three steps
5472 * - close the observation interval at the last (previous)
5473 * request dispatch or completion
5474 * - compute rate, if possible, for that observation interval
5475 * - reset to zero samples, which will trigger a proper
5476 * re-initialization of the observation interval on next
5479 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
5480 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
5481 1UL<<(BFQ_RATE_SHIFT - 10))
5482 bfq_update_rate_reset(bfqd, NULL);
5483 bfqd->last_completion = now_ns;
5484 bfqd->last_completed_rq_bfqq = bfqq;
5487 * If we are waiting to discover whether the request pattern
5488 * of the task associated with the queue is actually
5489 * isochronous, and both requisites for this condition to hold
5490 * are now satisfied, then compute soft_rt_next_start (see the
5491 * comments on the function bfq_bfqq_softrt_next_start()). We
5492 * do not compute soft_rt_next_start if bfqq is in interactive
5493 * weight raising (see the comments in bfq_bfqq_expire() for
5494 * an explanation). We schedule this delayed update when bfqq
5495 * expires, if it still has in-flight requests.
5497 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
5498 RB_EMPTY_ROOT(&bfqq->sort_list) &&
5499 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
5500 bfqq->soft_rt_next_start =
5501 bfq_bfqq_softrt_next_start(bfqd, bfqq);
5504 * If this is the in-service queue, check if it needs to be expired,
5505 * or if we want to idle in case it has no pending requests.
5507 if (bfqd->in_service_queue == bfqq) {
5508 if (bfq_bfqq_must_idle(bfqq)) {
5509 if (bfqq->dispatched == 0)
5510 bfq_arm_slice_timer(bfqd);
5512 * If we get here, we do not expire bfqq, even
5513 * if bfqq was in budget timeout or had no
5514 * more requests (as controlled in the next
5515 * conditional instructions). The reason for
5516 * not expiring bfqq is as follows.
5518 * Here bfqq->dispatched > 0 holds, but
5519 * bfq_bfqq_must_idle() returned true. This
5520 * implies that, even if no request arrives
5521 * for bfqq before bfqq->dispatched reaches 0,
5522 * bfqq will, however, not be expired on the
5523 * completion event that causes bfqq->dispatch
5524 * to reach zero. In contrast, on this event,
5525 * bfqq will start enjoying device idling
5526 * (I/O-dispatch plugging).
5528 * But, if we expired bfqq here, bfqq would
5529 * not have the chance to enjoy device idling
5530 * when bfqq->dispatched finally reaches
5531 * zero. This would expose bfqq to violation
5532 * of its reserved service guarantees.
5535 } else if (bfq_may_expire_for_budg_timeout(bfqq))
5536 bfq_bfqq_expire(bfqd, bfqq, false,
5537 BFQQE_BUDGET_TIMEOUT);
5538 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
5539 (bfqq->dispatched == 0 ||
5540 !bfq_better_to_idle(bfqq)))
5541 bfq_bfqq_expire(bfqd, bfqq, false,
5542 BFQQE_NO_MORE_REQUESTS);
5545 if (!bfqd->rq_in_driver)
5546 bfq_schedule_dispatch(bfqd);
5549 static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq)
5553 bfq_put_queue(bfqq);
5557 * The processes associated with bfqq may happen to generate their
5558 * cumulative I/O at a lower rate than the rate at which the device
5559 * could serve the same I/O. This is rather probable, e.g., if only
5560 * one process is associated with bfqq and the device is an SSD. It
5561 * results in bfqq becoming often empty while in service. In this
5562 * respect, if BFQ is allowed to switch to another queue when bfqq
5563 * remains empty, then the device goes on being fed with I/O requests,
5564 * and the throughput is not affected. In contrast, if BFQ is not
5565 * allowed to switch to another queue---because bfqq is sync and
5566 * I/O-dispatch needs to be plugged while bfqq is temporarily
5567 * empty---then, during the service of bfqq, there will be frequent
5568 * "service holes", i.e., time intervals during which bfqq gets empty
5569 * and the device can only consume the I/O already queued in its
5570 * hardware queues. During service holes, the device may even get to
5571 * remaining idle. In the end, during the service of bfqq, the device
5572 * is driven at a lower speed than the one it can reach with the kind
5573 * of I/O flowing through bfqq.
5575 * To counter this loss of throughput, BFQ implements a "request
5576 * injection mechanism", which tries to fill the above service holes
5577 * with I/O requests taken from other queues. The hard part in this
5578 * mechanism is finding the right amount of I/O to inject, so as to
5579 * both boost throughput and not break bfqq's bandwidth and latency
5580 * guarantees. In this respect, the mechanism maintains a per-queue
5581 * inject limit, computed as below. While bfqq is empty, the injection
5582 * mechanism dispatches extra I/O requests only until the total number
5583 * of I/O requests in flight---i.e., already dispatched but not yet
5584 * completed---remains lower than this limit.
5586 * A first definition comes in handy to introduce the algorithm by
5587 * which the inject limit is computed. We define as first request for
5588 * bfqq, an I/O request for bfqq that arrives while bfqq is in
5589 * service, and causes bfqq to switch from empty to non-empty. The
5590 * algorithm updates the limit as a function of the effect of
5591 * injection on the service times of only the first requests of
5592 * bfqq. The reason for this restriction is that these are the
5593 * requests whose service time is affected most, because they are the
5594 * first to arrive after injection possibly occurred.
5596 * To evaluate the effect of injection, the algorithm measures the
5597 * "total service time" of first requests. We define as total service
5598 * time of an I/O request, the time that elapses since when the
5599 * request is enqueued into bfqq, to when it is completed. This
5600 * quantity allows the whole effect of injection to be measured. It is
5601 * easy to see why. Suppose that some requests of other queues are
5602 * actually injected while bfqq is empty, and that a new request R
5603 * then arrives for bfqq. If the device does start to serve all or
5604 * part of the injected requests during the service hole, then,
5605 * because of this extra service, it may delay the next invocation of
5606 * the dispatch hook of BFQ. Then, even after R gets eventually
5607 * dispatched, the device may delay the actual service of R if it is
5608 * still busy serving the extra requests, or if it decides to serve,
5609 * before R, some extra request still present in its queues. As a
5610 * conclusion, the cumulative extra delay caused by injection can be
5611 * easily evaluated by just comparing the total service time of first
5612 * requests with and without injection.
5614 * The limit-update algorithm works as follows. On the arrival of a
5615 * first request of bfqq, the algorithm measures the total time of the
5616 * request only if one of the three cases below holds, and, for each
5617 * case, it updates the limit as described below:
5619 * (1) If there is no in-flight request. This gives a baseline for the
5620 * total service time of the requests of bfqq. If the baseline has
5621 * not been computed yet, then, after computing it, the limit is
5622 * set to 1, to start boosting throughput, and to prepare the
5623 * ground for the next case. If the baseline has already been
5624 * computed, then it is updated, in case it results to be lower
5625 * than the previous value.
5627 * (2) If the limit is higher than 0 and there are in-flight
5628 * requests. By comparing the total service time in this case with
5629 * the above baseline, it is possible to know at which extent the
5630 * current value of the limit is inflating the total service
5631 * time. If the inflation is below a certain threshold, then bfqq
5632 * is assumed to be suffering from no perceivable loss of its
5633 * service guarantees, and the limit is even tentatively
5634 * increased. If the inflation is above the threshold, then the
5635 * limit is decreased. Due to the lack of any hysteresis, this
5636 * logic makes the limit oscillate even in steady workload
5637 * conditions. Yet we opted for it, because it is fast in reaching
5638 * the best value for the limit, as a function of the current I/O
5639 * workload. To reduce oscillations, this step is disabled for a
5640 * short time interval after the limit happens to be decreased.
5642 * (3) Periodically, after resetting the limit, to make sure that the
5643 * limit eventually drops in case the workload changes. This is
5644 * needed because, after the limit has gone safely up for a
5645 * certain workload, it is impossible to guess whether the
5646 * baseline total service time may have changed, without measuring
5647 * it again without injection. A more effective version of this
5648 * step might be to just sample the baseline, by interrupting
5649 * injection only once, and then to reset/lower the limit only if
5650 * the total service time with the current limit does happen to be
5653 * More details on each step are provided in the comments on the
5654 * pieces of code that implement these steps: the branch handling the
5655 * transition from empty to non empty in bfq_add_request(), the branch
5656 * handling injection in bfq_select_queue(), and the function
5657 * bfq_choose_bfqq_for_injection(). These comments also explain some
5658 * exceptions, made by the injection mechanism in some special cases.
5660 static void bfq_update_inject_limit(struct bfq_data *bfqd,
5661 struct bfq_queue *bfqq)
5663 u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns;
5664 unsigned int old_limit = bfqq->inject_limit;
5666 if (bfqq->last_serv_time_ns > 0) {
5667 u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
5669 if (tot_time_ns >= threshold && old_limit > 0) {
5670 bfqq->inject_limit--;
5671 bfqq->decrease_time_jif = jiffies;
5672 } else if (tot_time_ns < threshold &&
5673 old_limit < bfqd->max_rq_in_driver<<1)
5674 bfqq->inject_limit++;
5678 * Either we still have to compute the base value for the
5679 * total service time, and there seem to be the right
5680 * conditions to do it, or we can lower the last base value
5683 * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
5684 * request in flight, because this function is in the code
5685 * path that handles the completion of a request of bfqq, and,
5686 * in particular, this function is executed before
5687 * bfqd->rq_in_driver is decremented in such a code path.
5689 if ((bfqq->last_serv_time_ns == 0 && bfqd->rq_in_driver == 1) ||
5690 tot_time_ns < bfqq->last_serv_time_ns) {
5691 bfqq->last_serv_time_ns = tot_time_ns;
5693 * Now we certainly have a base value: make sure we
5694 * start trying injection.
5696 bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
5697 } else if (!bfqd->rqs_injected && bfqd->rq_in_driver == 1)
5699 * No I/O injected and no request still in service in
5700 * the drive: these are the exact conditions for
5701 * computing the base value of the total service time
5702 * for bfqq. So let's update this value, because it is
5703 * rather variable. For example, it varies if the size
5704 * or the spatial locality of the I/O requests in bfqq
5707 bfqq->last_serv_time_ns = tot_time_ns;
5710 /* update complete, not waiting for any request completion any longer */
5711 bfqd->waited_rq = NULL;
5715 * Handle either a requeue or a finish for rq. The things to do are
5716 * the same in both cases: all references to rq are to be dropped. In
5717 * particular, rq is considered completed from the point of view of
5720 static void bfq_finish_requeue_request(struct request *rq)
5722 struct bfq_queue *bfqq = RQ_BFQQ(rq);
5723 struct bfq_data *bfqd;
5726 * Requeue and finish hooks are invoked in blk-mq without
5727 * checking whether the involved request is actually still
5728 * referenced in the scheduler. To handle this fact, the
5729 * following two checks make this function exit in case of
5730 * spurious invocations, for which there is nothing to do.
5732 * First, check whether rq has nothing to do with an elevator.
5734 if (unlikely(!(rq->rq_flags & RQF_ELVPRIV)))
5738 * rq either is not associated with any icq, or is an already
5739 * requeued request that has not (yet) been re-inserted into
5742 if (!rq->elv.icq || !bfqq)
5747 if (rq->rq_flags & RQF_STARTED)
5748 bfqg_stats_update_completion(bfqq_group(bfqq),
5750 rq->io_start_time_ns,
5753 if (likely(rq->rq_flags & RQF_STARTED)) {
5754 unsigned long flags;
5756 spin_lock_irqsave(&bfqd->lock, flags);
5758 if (rq == bfqd->waited_rq)
5759 bfq_update_inject_limit(bfqd, bfqq);
5761 bfq_completed_request(bfqq, bfqd);
5762 bfq_finish_requeue_request_body(bfqq);
5764 spin_unlock_irqrestore(&bfqd->lock, flags);
5767 * Request rq may be still/already in the scheduler,
5768 * in which case we need to remove it (this should
5769 * never happen in case of requeue). And we cannot
5770 * defer such a check and removal, to avoid
5771 * inconsistencies in the time interval from the end
5772 * of this function to the start of the deferred work.
5773 * This situation seems to occur only in process
5774 * context, as a consequence of a merge. In the
5775 * current version of the code, this implies that the
5779 if (!RB_EMPTY_NODE(&rq->rb_node)) {
5780 bfq_remove_request(rq->q, rq);
5781 bfqg_stats_update_io_remove(bfqq_group(bfqq),
5784 bfq_finish_requeue_request_body(bfqq);
5788 * Reset private fields. In case of a requeue, this allows
5789 * this function to correctly do nothing if it is spuriously
5790 * invoked again on this same request (see the check at the
5791 * beginning of the function). Probably, a better general
5792 * design would be to prevent blk-mq from invoking the requeue
5793 * or finish hooks of an elevator, for a request that is not
5794 * referred by that elevator.
5796 * Resetting the following fields would break the
5797 * request-insertion logic if rq is re-inserted into a bfq
5798 * internal queue, without a re-preparation. Here we assume
5799 * that re-insertions of requeued requests, without
5800 * re-preparation, can happen only for pass_through or at_head
5801 * requests (which are not re-inserted into bfq internal
5804 rq->elv.priv[0] = NULL;
5805 rq->elv.priv[1] = NULL;
5809 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
5810 * was the last process referring to that bfqq.
5812 static struct bfq_queue *
5813 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
5815 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
5817 if (bfqq_process_refs(bfqq) == 1) {
5818 bfqq->pid = current->pid;
5819 bfq_clear_bfqq_coop(bfqq);
5820 bfq_clear_bfqq_split_coop(bfqq);
5824 bic_set_bfqq(bic, NULL, 1);
5826 bfq_put_cooperator(bfqq);
5828 bfq_put_queue(bfqq);
5832 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
5833 struct bfq_io_cq *bic,
5835 bool split, bool is_sync,
5838 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
5840 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
5847 bfq_put_queue(bfqq);
5848 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
5850 bic_set_bfqq(bic, bfqq, is_sync);
5851 if (split && is_sync) {
5852 if ((bic->was_in_burst_list && bfqd->large_burst) ||
5853 bic->saved_in_large_burst)
5854 bfq_mark_bfqq_in_large_burst(bfqq);
5856 bfq_clear_bfqq_in_large_burst(bfqq);
5857 if (bic->was_in_burst_list)
5859 * If bfqq was in the current
5860 * burst list before being
5861 * merged, then we have to add
5862 * it back. And we do not need
5863 * to increase burst_size, as
5864 * we did not decrement
5865 * burst_size when we removed
5866 * bfqq from the burst list as
5867 * a consequence of a merge
5869 * bfq_put_queue). In this
5870 * respect, it would be rather
5871 * costly to know whether the
5872 * current burst list is still
5873 * the same burst list from
5874 * which bfqq was removed on
5875 * the merge. To avoid this
5876 * cost, if bfqq was in a
5877 * burst list, then we add
5878 * bfqq to the current burst
5879 * list without any further
5880 * check. This can cause
5881 * inappropriate insertions,
5882 * but rarely enough to not
5883 * harm the detection of large
5884 * bursts significantly.
5886 hlist_add_head(&bfqq->burst_list_node,
5889 bfqq->split_time = jiffies;
5896 * Only reset private fields. The actual request preparation will be
5897 * performed by bfq_init_rq, when rq is either inserted or merged. See
5898 * comments on bfq_init_rq for the reason behind this delayed
5901 static void bfq_prepare_request(struct request *rq, struct bio *bio)
5904 * Regardless of whether we have an icq attached, we have to
5905 * clear the scheduler pointers, as they might point to
5906 * previously allocated bic/bfqq structs.
5908 rq->elv.priv[0] = rq->elv.priv[1] = NULL;
5912 * If needed, init rq, allocate bfq data structures associated with
5913 * rq, and increment reference counters in the destination bfq_queue
5914 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
5915 * not associated with any bfq_queue.
5917 * This function is invoked by the functions that perform rq insertion
5918 * or merging. One may have expected the above preparation operations
5919 * to be performed in bfq_prepare_request, and not delayed to when rq
5920 * is inserted or merged. The rationale behind this delayed
5921 * preparation is that, after the prepare_request hook is invoked for
5922 * rq, rq may still be transformed into a request with no icq, i.e., a
5923 * request not associated with any queue. No bfq hook is invoked to
5924 * signal this transformation. As a consequence, should these
5925 * preparation operations be performed when the prepare_request hook
5926 * is invoked, and should rq be transformed one moment later, bfq
5927 * would end up in an inconsistent state, because it would have
5928 * incremented some queue counters for an rq destined to
5929 * transformation, without any chance to correctly lower these
5930 * counters back. In contrast, no transformation can still happen for
5931 * rq after rq has been inserted or merged. So, it is safe to execute
5932 * these preparation operations when rq is finally inserted or merged.
5934 static struct bfq_queue *bfq_init_rq(struct request *rq)
5936 struct request_queue *q = rq->q;
5937 struct bio *bio = rq->bio;
5938 struct bfq_data *bfqd = q->elevator->elevator_data;
5939 struct bfq_io_cq *bic;
5940 const int is_sync = rq_is_sync(rq);
5941 struct bfq_queue *bfqq;
5942 bool new_queue = false;
5943 bool bfqq_already_existing = false, split = false;
5945 if (unlikely(!rq->elv.icq))
5949 * Assuming that elv.priv[1] is set only if everything is set
5950 * for this rq. This holds true, because this function is
5951 * invoked only for insertion or merging, and, after such
5952 * events, a request cannot be manipulated any longer before
5953 * being removed from bfq.
5955 if (rq->elv.priv[1])
5956 return rq->elv.priv[1];
5958 bic = icq_to_bic(rq->elv.icq);
5960 bfq_check_ioprio_change(bic, bio);
5962 bfq_bic_update_cgroup(bic, bio);
5964 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
5967 if (likely(!new_queue)) {
5968 /* If the queue was seeky for too long, break it apart. */
5969 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
5970 bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
5972 /* Update bic before losing reference to bfqq */
5973 if (bfq_bfqq_in_large_burst(bfqq))
5974 bic->saved_in_large_burst = true;
5976 bfqq = bfq_split_bfqq(bic, bfqq);
5980 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
5984 bfqq_already_existing = true;
5990 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
5991 rq, bfqq, bfqq->ref);
5993 rq->elv.priv[0] = bic;
5994 rq->elv.priv[1] = bfqq;
5997 * If a bfq_queue has only one process reference, it is owned
5998 * by only this bic: we can then set bfqq->bic = bic. in
5999 * addition, if the queue has also just been split, we have to
6002 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
6006 * The queue has just been split from a shared
6007 * queue: restore the idle window and the
6008 * possible weight raising period.
6010 bfq_bfqq_resume_state(bfqq, bfqd, bic,
6011 bfqq_already_existing);
6016 * Consider bfqq as possibly belonging to a burst of newly
6017 * created queues only if:
6018 * 1) A burst is actually happening (bfqd->burst_size > 0)
6020 * 2) There is no other active queue. In fact, if, in
6021 * contrast, there are active queues not belonging to the
6022 * possible burst bfqq may belong to, then there is no gain
6023 * in considering bfqq as belonging to a burst, and
6024 * therefore in not weight-raising bfqq. See comments on
6025 * bfq_handle_burst().
6027 * This filtering also helps eliminating false positives,
6028 * occurring when bfqq does not belong to an actual large
6029 * burst, but some background task (e.g., a service) happens
6030 * to trigger the creation of new queues very close to when
6031 * bfqq and its possible companion queues are created. See
6032 * comments on bfq_handle_burst() for further details also on
6035 if (unlikely(bfq_bfqq_just_created(bfqq) &&
6036 (bfqd->burst_size > 0 ||
6037 bfq_tot_busy_queues(bfqd) == 0)))
6038 bfq_handle_burst(bfqd, bfqq);
6043 static void bfq_idle_slice_timer_body(struct bfq_queue *bfqq)
6045 struct bfq_data *bfqd = bfqq->bfqd;
6046 enum bfqq_expiration reason;
6047 unsigned long flags;
6049 spin_lock_irqsave(&bfqd->lock, flags);
6050 bfq_clear_bfqq_wait_request(bfqq);
6052 if (bfqq != bfqd->in_service_queue) {
6053 spin_unlock_irqrestore(&bfqd->lock, flags);
6057 if (bfq_bfqq_budget_timeout(bfqq))
6059 * Also here the queue can be safely expired
6060 * for budget timeout without wasting
6063 reason = BFQQE_BUDGET_TIMEOUT;
6064 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
6066 * The queue may not be empty upon timer expiration,
6067 * because we may not disable the timer when the
6068 * first request of the in-service queue arrives
6069 * during disk idling.
6071 reason = BFQQE_TOO_IDLE;
6073 goto schedule_dispatch;
6075 bfq_bfqq_expire(bfqd, bfqq, true, reason);
6078 spin_unlock_irqrestore(&bfqd->lock, flags);
6079 bfq_schedule_dispatch(bfqd);
6083 * Handler of the expiration of the timer running if the in-service queue
6084 * is idling inside its time slice.
6086 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
6088 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
6090 struct bfq_queue *bfqq = bfqd->in_service_queue;
6093 * Theoretical race here: the in-service queue can be NULL or
6094 * different from the queue that was idling if a new request
6095 * arrives for the current queue and there is a full dispatch
6096 * cycle that changes the in-service queue. This can hardly
6097 * happen, but in the worst case we just expire a queue too
6101 bfq_idle_slice_timer_body(bfqq);
6103 return HRTIMER_NORESTART;
6106 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
6107 struct bfq_queue **bfqq_ptr)
6109 struct bfq_queue *bfqq = *bfqq_ptr;
6111 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
6113 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
6115 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
6117 bfq_put_queue(bfqq);
6123 * Release all the bfqg references to its async queues. If we are
6124 * deallocating the group these queues may still contain requests, so
6125 * we reparent them to the root cgroup (i.e., the only one that will
6126 * exist for sure until all the requests on a device are gone).
6128 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
6132 for (i = 0; i < 2; i++)
6133 for (j = 0; j < IOPRIO_BE_NR; j++)
6134 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
6136 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
6140 * See the comments on bfq_limit_depth for the purpose of
6141 * the depths set in the function. Return minimum shallow depth we'll use.
6143 static unsigned int bfq_update_depths(struct bfq_data *bfqd,
6144 struct sbitmap_queue *bt)
6146 unsigned int i, j, min_shallow = UINT_MAX;
6149 * In-word depths if no bfq_queue is being weight-raised:
6150 * leaving 25% of tags only for sync reads.
6152 * In next formulas, right-shift the value
6153 * (1U<<bt->sb.shift), instead of computing directly
6154 * (1U<<(bt->sb.shift - something)), to be robust against
6155 * any possible value of bt->sb.shift, without having to
6156 * limit 'something'.
6158 /* no more than 50% of tags for async I/O */
6159 bfqd->word_depths[0][0] = max((1U << bt->sb.shift) >> 1, 1U);
6161 * no more than 75% of tags for sync writes (25% extra tags
6162 * w.r.t. async I/O, to prevent async I/O from starving sync
6165 bfqd->word_depths[0][1] = max(((1U << bt->sb.shift) * 3) >> 2, 1U);
6168 * In-word depths in case some bfq_queue is being weight-
6169 * raised: leaving ~63% of tags for sync reads. This is the
6170 * highest percentage for which, in our tests, application
6171 * start-up times didn't suffer from any regression due to tag
6174 /* no more than ~18% of tags for async I/O */
6175 bfqd->word_depths[1][0] = max(((1U << bt->sb.shift) * 3) >> 4, 1U);
6176 /* no more than ~37% of tags for sync writes (~20% extra tags) */
6177 bfqd->word_depths[1][1] = max(((1U << bt->sb.shift) * 6) >> 4, 1U);
6179 for (i = 0; i < 2; i++)
6180 for (j = 0; j < 2; j++)
6181 min_shallow = min(min_shallow, bfqd->word_depths[i][j]);
6186 static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
6188 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
6189 struct blk_mq_tags *tags = hctx->sched_tags;
6190 unsigned int min_shallow;
6192 min_shallow = bfq_update_depths(bfqd, &tags->bitmap_tags);
6193 sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, min_shallow);
6196 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
6198 bfq_depth_updated(hctx);
6202 static void bfq_exit_queue(struct elevator_queue *e)
6204 struct bfq_data *bfqd = e->elevator_data;
6205 struct bfq_queue *bfqq, *n;
6207 hrtimer_cancel(&bfqd->idle_slice_timer);
6209 spin_lock_irq(&bfqd->lock);
6210 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
6211 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
6212 spin_unlock_irq(&bfqd->lock);
6214 hrtimer_cancel(&bfqd->idle_slice_timer);
6216 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6217 /* release oom-queue reference to root group */
6218 bfqg_and_blkg_put(bfqd->root_group);
6220 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
6222 spin_lock_irq(&bfqd->lock);
6223 bfq_put_async_queues(bfqd, bfqd->root_group);
6224 kfree(bfqd->root_group);
6225 spin_unlock_irq(&bfqd->lock);
6231 static void bfq_init_root_group(struct bfq_group *root_group,
6232 struct bfq_data *bfqd)
6236 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6237 root_group->entity.parent = NULL;
6238 root_group->my_entity = NULL;
6239 root_group->bfqd = bfqd;
6241 root_group->rq_pos_tree = RB_ROOT;
6242 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
6243 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
6244 root_group->sched_data.bfq_class_idle_last_service = jiffies;
6247 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
6249 struct bfq_data *bfqd;
6250 struct elevator_queue *eq;
6252 eq = elevator_alloc(q, e);
6256 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
6258 kobject_put(&eq->kobj);
6261 eq->elevator_data = bfqd;
6263 spin_lock_irq(&q->queue_lock);
6265 spin_unlock_irq(&q->queue_lock);
6268 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
6269 * Grab a permanent reference to it, so that the normal code flow
6270 * will not attempt to free it.
6272 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
6273 bfqd->oom_bfqq.ref++;
6274 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
6275 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
6276 bfqd->oom_bfqq.entity.new_weight =
6277 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
6279 /* oom_bfqq does not participate to bursts */
6280 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
6283 * Trigger weight initialization, according to ioprio, at the
6284 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
6285 * class won't be changed any more.
6287 bfqd->oom_bfqq.entity.prio_changed = 1;
6291 INIT_LIST_HEAD(&bfqd->dispatch);
6293 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
6295 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
6297 bfqd->queue_weights_tree = RB_ROOT_CACHED;
6298 bfqd->num_groups_with_pending_reqs = 0;
6300 INIT_LIST_HEAD(&bfqd->active_list);
6301 INIT_LIST_HEAD(&bfqd->idle_list);
6302 INIT_HLIST_HEAD(&bfqd->burst_list);
6305 bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
6307 bfqd->bfq_max_budget = bfq_default_max_budget;
6309 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
6310 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
6311 bfqd->bfq_back_max = bfq_back_max;
6312 bfqd->bfq_back_penalty = bfq_back_penalty;
6313 bfqd->bfq_slice_idle = bfq_slice_idle;
6314 bfqd->bfq_timeout = bfq_timeout;
6316 bfqd->bfq_requests_within_timer = 120;
6318 bfqd->bfq_large_burst_thresh = 8;
6319 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
6321 bfqd->low_latency = true;
6324 * Trade-off between responsiveness and fairness.
6326 bfqd->bfq_wr_coeff = 30;
6327 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
6328 bfqd->bfq_wr_max_time = 0;
6329 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
6330 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
6331 bfqd->bfq_wr_max_softrt_rate = 7000; /*
6332 * Approximate rate required
6333 * to playback or record a
6334 * high-definition compressed
6337 bfqd->wr_busy_queues = 0;
6340 * Begin by assuming, optimistically, that the device peak
6341 * rate is equal to 2/3 of the highest reference rate.
6343 bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
6344 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
6345 bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
6347 spin_lock_init(&bfqd->lock);
6350 * The invocation of the next bfq_create_group_hierarchy
6351 * function is the head of a chain of function calls
6352 * (bfq_create_group_hierarchy->blkcg_activate_policy->
6353 * blk_mq_freeze_queue) that may lead to the invocation of the
6354 * has_work hook function. For this reason,
6355 * bfq_create_group_hierarchy is invoked only after all
6356 * scheduler data has been initialized, apart from the fields
6357 * that can be initialized only after invoking
6358 * bfq_create_group_hierarchy. This, in particular, enables
6359 * has_work to correctly return false. Of course, to avoid
6360 * other inconsistencies, the blk-mq stack must then refrain
6361 * from invoking further scheduler hooks before this init
6362 * function is finished.
6364 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
6365 if (!bfqd->root_group)
6367 bfq_init_root_group(bfqd->root_group, bfqd);
6368 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
6370 wbt_disable_default(q);
6375 kobject_put(&eq->kobj);
6379 static void bfq_slab_kill(void)
6381 kmem_cache_destroy(bfq_pool);
6384 static int __init bfq_slab_setup(void)
6386 bfq_pool = KMEM_CACHE(bfq_queue, 0);
6392 static ssize_t bfq_var_show(unsigned int var, char *page)
6394 return sprintf(page, "%u\n", var);
6397 static int bfq_var_store(unsigned long *var, const char *page)
6399 unsigned long new_val;
6400 int ret = kstrtoul(page, 10, &new_val);
6408 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
6409 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
6411 struct bfq_data *bfqd = e->elevator_data; \
6412 u64 __data = __VAR; \
6414 __data = jiffies_to_msecs(__data); \
6415 else if (__CONV == 2) \
6416 __data = div_u64(__data, NSEC_PER_MSEC); \
6417 return bfq_var_show(__data, (page)); \
6419 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
6420 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
6421 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
6422 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
6423 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
6424 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
6425 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
6426 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
6427 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
6428 #undef SHOW_FUNCTION
6430 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
6431 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
6433 struct bfq_data *bfqd = e->elevator_data; \
6434 u64 __data = __VAR; \
6435 __data = div_u64(__data, NSEC_PER_USEC); \
6436 return bfq_var_show(__data, (page)); \
6438 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
6439 #undef USEC_SHOW_FUNCTION
6441 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
6443 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
6445 struct bfq_data *bfqd = e->elevator_data; \
6446 unsigned long __data, __min = (MIN), __max = (MAX); \
6449 ret = bfq_var_store(&__data, (page)); \
6452 if (__data < __min) \
6454 else if (__data > __max) \
6457 *(__PTR) = msecs_to_jiffies(__data); \
6458 else if (__CONV == 2) \
6459 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
6461 *(__PTR) = __data; \
6464 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
6466 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
6468 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
6469 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
6471 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
6472 #undef STORE_FUNCTION
6474 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
6475 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
6477 struct bfq_data *bfqd = e->elevator_data; \
6478 unsigned long __data, __min = (MIN), __max = (MAX); \
6481 ret = bfq_var_store(&__data, (page)); \
6484 if (__data < __min) \
6486 else if (__data > __max) \
6488 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
6491 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
6493 #undef USEC_STORE_FUNCTION
6495 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
6496 const char *page, size_t count)
6498 struct bfq_data *bfqd = e->elevator_data;
6499 unsigned long __data;
6502 ret = bfq_var_store(&__data, (page));
6507 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
6509 if (__data > INT_MAX)
6511 bfqd->bfq_max_budget = __data;
6514 bfqd->bfq_user_max_budget = __data;
6520 * Leaving this name to preserve name compatibility with cfq
6521 * parameters, but this timeout is used for both sync and async.
6523 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
6524 const char *page, size_t count)
6526 struct bfq_data *bfqd = e->elevator_data;
6527 unsigned long __data;
6530 ret = bfq_var_store(&__data, (page));
6536 else if (__data > INT_MAX)
6539 bfqd->bfq_timeout = msecs_to_jiffies(__data);
6540 if (bfqd->bfq_user_max_budget == 0)
6541 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
6546 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
6547 const char *page, size_t count)
6549 struct bfq_data *bfqd = e->elevator_data;
6550 unsigned long __data;
6553 ret = bfq_var_store(&__data, (page));
6559 if (!bfqd->strict_guarantees && __data == 1
6560 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
6561 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
6563 bfqd->strict_guarantees = __data;
6568 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
6569 const char *page, size_t count)
6571 struct bfq_data *bfqd = e->elevator_data;
6572 unsigned long __data;
6575 ret = bfq_var_store(&__data, (page));
6581 if (__data == 0 && bfqd->low_latency != 0)
6583 bfqd->low_latency = __data;
6588 #define BFQ_ATTR(name) \
6589 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
6591 static struct elv_fs_entry bfq_attrs[] = {
6592 BFQ_ATTR(fifo_expire_sync),
6593 BFQ_ATTR(fifo_expire_async),
6594 BFQ_ATTR(back_seek_max),
6595 BFQ_ATTR(back_seek_penalty),
6596 BFQ_ATTR(slice_idle),
6597 BFQ_ATTR(slice_idle_us),
6598 BFQ_ATTR(max_budget),
6599 BFQ_ATTR(timeout_sync),
6600 BFQ_ATTR(strict_guarantees),
6601 BFQ_ATTR(low_latency),
6605 static struct elevator_type iosched_bfq_mq = {
6607 .limit_depth = bfq_limit_depth,
6608 .prepare_request = bfq_prepare_request,
6609 .requeue_request = bfq_finish_requeue_request,
6610 .finish_request = bfq_finish_requeue_request,
6611 .exit_icq = bfq_exit_icq,
6612 .insert_requests = bfq_insert_requests,
6613 .dispatch_request = bfq_dispatch_request,
6614 .next_request = elv_rb_latter_request,
6615 .former_request = elv_rb_former_request,
6616 .allow_merge = bfq_allow_bio_merge,
6617 .bio_merge = bfq_bio_merge,
6618 .request_merge = bfq_request_merge,
6619 .requests_merged = bfq_requests_merged,
6620 .request_merged = bfq_request_merged,
6621 .has_work = bfq_has_work,
6622 .depth_updated = bfq_depth_updated,
6623 .init_hctx = bfq_init_hctx,
6624 .init_sched = bfq_init_queue,
6625 .exit_sched = bfq_exit_queue,
6628 .icq_size = sizeof(struct bfq_io_cq),
6629 .icq_align = __alignof__(struct bfq_io_cq),
6630 .elevator_attrs = bfq_attrs,
6631 .elevator_name = "bfq",
6632 .elevator_owner = THIS_MODULE,
6634 MODULE_ALIAS("bfq-iosched");
6636 static int __init bfq_init(void)
6640 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6641 ret = blkcg_policy_register(&blkcg_policy_bfq);
6647 if (bfq_slab_setup())
6651 * Times to load large popular applications for the typical
6652 * systems installed on the reference devices (see the
6653 * comments before the definition of the next
6654 * array). Actually, we use slightly lower values, as the
6655 * estimated peak rate tends to be smaller than the actual
6656 * peak rate. The reason for this last fact is that estimates
6657 * are computed over much shorter time intervals than the long
6658 * intervals typically used for benchmarking. Why? First, to
6659 * adapt more quickly to variations. Second, because an I/O
6660 * scheduler cannot rely on a peak-rate-evaluation workload to
6661 * be run for a long time.
6663 ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
6664 ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
6666 ret = elv_register(&iosched_bfq_mq);
6675 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6676 blkcg_policy_unregister(&blkcg_policy_bfq);
6681 static void __exit bfq_exit(void)
6683 elv_unregister(&iosched_bfq_mq);
6684 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6685 blkcg_policy_unregister(&blkcg_policy_bfq);
6690 module_init(bfq_init);
6691 module_exit(bfq_exit);
6693 MODULE_AUTHOR("Paolo Valente");
6694 MODULE_LICENSE("GPL");
6695 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");