2 * Budget Fair Queueing (BFQ) I/O scheduler.
4 * Based on ideas and code from CFQ:
5 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
8 * Paolo Valente <paolo.valente@unimore.it>
10 * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
11 * Arianna Avanzini <avanzini@google.com>
13 * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License as
17 * published by the Free Software Foundation; either version 2 of the
18 * License, or (at your option) any later version.
20 * This program is distributed in the hope that it will be useful,
21 * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
23 * General Public License for more details.
25 * BFQ is a proportional-share I/O scheduler, with some extra
26 * low-latency capabilities. BFQ also supports full hierarchical
27 * scheduling through cgroups. Next paragraphs provide an introduction
28 * on BFQ inner workings. Details on BFQ benefits, usage and
29 * limitations can be found in Documentation/block/bfq-iosched.txt.
31 * BFQ is a proportional-share storage-I/O scheduling algorithm based
32 * on the slice-by-slice service scheme of CFQ. But BFQ assigns
33 * budgets, measured in number of sectors, to processes instead of
34 * time slices. The device is not granted to the in-service process
35 * for a given time slice, but until it has exhausted its assigned
36 * budget. This change from the time to the service domain enables BFQ
37 * to distribute the device throughput among processes as desired,
38 * without any distortion due to throughput fluctuations, or to device
39 * internal queueing. BFQ uses an ad hoc internal scheduler, called
40 * B-WF2Q+, to schedule processes according to their budgets. More
41 * precisely, BFQ schedules queues associated with processes. Each
42 * process/queue is assigned a user-configurable weight, and B-WF2Q+
43 * guarantees that each queue receives a fraction of the throughput
44 * proportional to its weight. Thanks to the accurate policy of
45 * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
46 * processes issuing sequential requests (to boost the throughput),
47 * and yet guarantee a low latency to interactive and soft real-time
50 * In particular, to provide these low-latency guarantees, BFQ
51 * explicitly privileges the I/O of two classes of time-sensitive
52 * applications: interactive and soft real-time. In more detail, BFQ
53 * behaves this way if the low_latency parameter is set (default
54 * configuration). This feature enables BFQ to provide applications in
55 * these classes with a very low latency.
57 * To implement this feature, BFQ constantly tries to detect whether
58 * the I/O requests in a bfq_queue come from an interactive or a soft
59 * real-time application. For brevity, in these cases, the queue is
60 * said to be interactive or soft real-time. In both cases, BFQ
61 * privileges the service of the queue, over that of non-interactive
62 * and non-soft-real-time queues. This privileging is performed,
63 * mainly, by raising the weight of the queue. So, for brevity, we
64 * call just weight-raising periods the time periods during which a
65 * queue is privileged, because deemed interactive or soft real-time.
67 * The detection of soft real-time queues/applications is described in
68 * detail in the comments on the function
69 * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
70 * interactive queue works as follows: a queue is deemed interactive
71 * if it is constantly non empty only for a limited time interval,
72 * after which it does become empty. The queue may be deemed
73 * interactive again (for a limited time), if it restarts being
74 * constantly non empty, provided that this happens only after the
75 * queue has remained empty for a given minimum idle time.
77 * By default, BFQ computes automatically the above maximum time
78 * interval, i.e., the time interval after which a constantly
79 * non-empty queue stops being deemed interactive. Since a queue is
80 * weight-raised while it is deemed interactive, this maximum time
81 * interval happens to coincide with the (maximum) duration of the
82 * weight-raising for interactive queues.
84 * Finally, BFQ also features additional heuristics for
85 * preserving both a low latency and a high throughput on NCQ-capable,
86 * rotational or flash-based devices, and to get the job done quickly
87 * for applications consisting in many I/O-bound processes.
89 * NOTE: if the main or only goal, with a given device, is to achieve
90 * the maximum-possible throughput at all times, then do switch off
91 * all low-latency heuristics for that device, by setting low_latency
94 * BFQ is described in [1], where also a reference to the initial,
95 * more theoretical paper on BFQ can be found. The interested reader
96 * can find in the latter paper full details on the main algorithm, as
97 * well as formulas of the guarantees and formal proofs of all the
98 * properties. With respect to the version of BFQ presented in these
99 * papers, this implementation adds a few more heuristics, such as the
100 * ones that guarantee a low latency to interactive and soft real-time
101 * applications, and a hierarchical extension based on H-WF2Q+.
103 * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
104 * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
105 * with O(log N) complexity derives from the one introduced with EEVDF
108 * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
109 * Scheduler", Proceedings of the First Workshop on Mobile System
110 * Technologies (MST-2015), May 2015.
111 * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
113 * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
114 * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
117 * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
119 * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
120 * First: A Flexible and Accurate Mechanism for Proportional Share
121 * Resource Allocation", technical report.
123 * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
125 #include <linux/module.h>
126 #include <linux/slab.h>
127 #include <linux/blkdev.h>
128 #include <linux/cgroup.h>
129 #include <linux/elevator.h>
130 #include <linux/ktime.h>
131 #include <linux/rbtree.h>
132 #include <linux/ioprio.h>
133 #include <linux/sbitmap.h>
134 #include <linux/delay.h>
138 #include "blk-mq-tag.h"
139 #include "blk-mq-sched.h"
140 #include "bfq-iosched.h"
143 #define BFQ_BFQQ_FNS(name) \
144 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
146 __set_bit(BFQQF_##name, &(bfqq)->flags); \
148 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
150 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
152 int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
154 return test_bit(BFQQF_##name, &(bfqq)->flags); \
157 BFQ_BFQQ_FNS(just_created);
159 BFQ_BFQQ_FNS(wait_request);
160 BFQ_BFQQ_FNS(non_blocking_wait_rq);
161 BFQ_BFQQ_FNS(fifo_expire);
162 BFQ_BFQQ_FNS(has_short_ttime);
164 BFQ_BFQQ_FNS(IO_bound);
165 BFQ_BFQQ_FNS(in_large_burst);
167 BFQ_BFQQ_FNS(split_coop);
168 BFQ_BFQQ_FNS(softrt_update);
169 #undef BFQ_BFQQ_FNS \
171 /* Expiration time of sync (0) and async (1) requests, in ns. */
172 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
174 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
175 static const int bfq_back_max = 16 * 1024;
177 /* Penalty of a backwards seek, in number of sectors. */
178 static const int bfq_back_penalty = 2;
180 /* Idling period duration, in ns. */
181 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
183 /* Minimum number of assigned budgets for which stats are safe to compute. */
184 static const int bfq_stats_min_budgets = 194;
186 /* Default maximum budget values, in sectors and number of requests. */
187 static const int bfq_default_max_budget = 16 * 1024;
190 * When a sync request is dispatched, the queue that contains that
191 * request, and all the ancestor entities of that queue, are charged
192 * with the number of sectors of the request. In constrast, if the
193 * request is async, then the queue and its ancestor entities are
194 * charged with the number of sectors of the request, multiplied by
195 * the factor below. This throttles the bandwidth for async I/O,
196 * w.r.t. to sync I/O, and it is done to counter the tendency of async
197 * writes to steal I/O throughput to reads.
199 * The current value of this parameter is the result of a tuning with
200 * several hardware and software configurations. We tried to find the
201 * lowest value for which writes do not cause noticeable problems to
202 * reads. In fact, the lower this parameter, the stabler I/O control,
203 * in the following respect. The lower this parameter is, the less
204 * the bandwidth enjoyed by a group decreases
205 * - when the group does writes, w.r.t. to when it does reads;
206 * - when other groups do reads, w.r.t. to when they do writes.
208 static const int bfq_async_charge_factor = 3;
210 /* Default timeout values, in jiffies, approximating CFQ defaults. */
211 const int bfq_timeout = HZ / 8;
214 * Time limit for merging (see comments in bfq_setup_cooperator). Set
215 * to the slowest value that, in our tests, proved to be effective in
216 * removing false positives, while not causing true positives to miss
219 * As can be deduced from the low time limit below, queue merging, if
220 * successful, happens at the very beggining of the I/O of the involved
221 * cooperating processes, as a consequence of the arrival of the very
222 * first requests from each cooperator. After that, there is very
223 * little chance to find cooperators.
225 static const unsigned long bfq_merge_time_limit = HZ/10;
227 static struct kmem_cache *bfq_pool;
229 /* Below this threshold (in ns), we consider thinktime immediate. */
230 #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
232 /* hw_tag detection: parallel requests threshold and min samples needed. */
233 #define BFQ_HW_QUEUE_THRESHOLD 4
234 #define BFQ_HW_QUEUE_SAMPLES 32
236 #define BFQQ_SEEK_THR (sector_t)(8 * 100)
237 #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
238 #define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
239 #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19)
241 /* Min number of samples required to perform peak-rate update */
242 #define BFQ_RATE_MIN_SAMPLES 32
243 /* Min observation time interval required to perform a peak-rate update (ns) */
244 #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
245 /* Target observation time interval for a peak-rate update (ns) */
246 #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
249 * Shift used for peak-rate fixed precision calculations.
251 * - the current shift: 16 positions
252 * - the current type used to store rate: u32
253 * - the current unit of measure for rate: [sectors/usec], or, more precisely,
254 * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
255 * the range of rates that can be stored is
256 * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
257 * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
258 * [15, 65G] sectors/sec
259 * Which, assuming a sector size of 512B, corresponds to a range of
262 #define BFQ_RATE_SHIFT 16
265 * When configured for computing the duration of the weight-raising
266 * for interactive queues automatically (see the comments at the
267 * beginning of this file), BFQ does it using the following formula:
268 * duration = (ref_rate / r) * ref_wr_duration,
269 * where r is the peak rate of the device, and ref_rate and
270 * ref_wr_duration are two reference parameters. In particular,
271 * ref_rate is the peak rate of the reference storage device (see
272 * below), and ref_wr_duration is about the maximum time needed, with
273 * BFQ and while reading two files in parallel, to load typical large
274 * applications on the reference device (see the comments on
275 * max_service_from_wr below, for more details on how ref_wr_duration
276 * is obtained). In practice, the slower/faster the device at hand
277 * is, the more/less it takes to load applications with respect to the
278 * reference device. Accordingly, the longer/shorter BFQ grants
279 * weight raising to interactive applications.
281 * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
282 * depending on whether the device is rotational or non-rotational.
284 * In the following definitions, ref_rate[0] and ref_wr_duration[0]
285 * are the reference values for a rotational device, whereas
286 * ref_rate[1] and ref_wr_duration[1] are the reference values for a
287 * non-rotational device. The reference rates are not the actual peak
288 * rates of the devices used as a reference, but slightly lower
289 * values. The reason for using slightly lower values is that the
290 * peak-rate estimator tends to yield slightly lower values than the
291 * actual peak rate (it can yield the actual peak rate only if there
292 * is only one process doing I/O, and the process does sequential
295 * The reference peak rates are measured in sectors/usec, left-shifted
298 static int ref_rate[2] = {14000, 33000};
300 * To improve readability, a conversion function is used to initialize
301 * the following array, which entails that the array can be
302 * initialized only in a function.
304 static int ref_wr_duration[2];
307 * BFQ uses the above-detailed, time-based weight-raising mechanism to
308 * privilege interactive tasks. This mechanism is vulnerable to the
309 * following false positives: I/O-bound applications that will go on
310 * doing I/O for much longer than the duration of weight
311 * raising. These applications have basically no benefit from being
312 * weight-raised at the beginning of their I/O. On the opposite end,
313 * while being weight-raised, these applications
314 * a) unjustly steal throughput to applications that may actually need
316 * b) make BFQ uselessly perform device idling; device idling results
317 * in loss of device throughput with most flash-based storage, and may
318 * increase latencies when used purposelessly.
320 * BFQ tries to reduce these problems, by adopting the following
321 * countermeasure. To introduce this countermeasure, we need first to
322 * finish explaining how the duration of weight-raising for
323 * interactive tasks is computed.
325 * For a bfq_queue deemed as interactive, the duration of weight
326 * raising is dynamically adjusted, as a function of the estimated
327 * peak rate of the device, so as to be equal to the time needed to
328 * execute the 'largest' interactive task we benchmarked so far. By
329 * largest task, we mean the task for which each involved process has
330 * to do more I/O than for any of the other tasks we benchmarked. This
331 * reference interactive task is the start-up of LibreOffice Writer,
332 * and in this task each process/bfq_queue needs to have at most ~110K
333 * sectors transferred.
335 * This last piece of information enables BFQ to reduce the actual
336 * duration of weight-raising for at least one class of I/O-bound
337 * applications: those doing sequential or quasi-sequential I/O. An
338 * example is file copy. In fact, once started, the main I/O-bound
339 * processes of these applications usually consume the above 110K
340 * sectors in much less time than the processes of an application that
341 * is starting, because these I/O-bound processes will greedily devote
342 * almost all their CPU cycles only to their target,
343 * throughput-friendly I/O operations. This is even more true if BFQ
344 * happens to be underestimating the device peak rate, and thus
345 * overestimating the duration of weight raising. But, according to
346 * our measurements, once transferred 110K sectors, these processes
347 * have no right to be weight-raised any longer.
349 * Basing on the last consideration, BFQ ends weight-raising for a
350 * bfq_queue if the latter happens to have received an amount of
351 * service at least equal to the following constant. The constant is
352 * set to slightly more than 110K, to have a minimum safety margin.
354 * This early ending of weight-raising reduces the amount of time
355 * during which interactive false positives cause the two problems
356 * described at the beginning of these comments.
358 static const unsigned long max_service_from_wr = 120000;
360 #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
361 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
363 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
365 return bic->bfqq[is_sync];
368 void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
370 bic->bfqq[is_sync] = bfqq;
373 struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
375 return bic->icq.q->elevator->elevator_data;
379 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
380 * @icq: the iocontext queue.
382 static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
384 /* bic->icq is the first member, %NULL will convert to %NULL */
385 return container_of(icq, struct bfq_io_cq, icq);
389 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
390 * @bfqd: the lookup key.
391 * @ioc: the io_context of the process doing I/O.
392 * @q: the request queue.
394 static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd,
395 struct io_context *ioc,
396 struct request_queue *q)
400 struct bfq_io_cq *icq;
402 spin_lock_irqsave(&q->queue_lock, flags);
403 icq = icq_to_bic(ioc_lookup_icq(ioc, q));
404 spin_unlock_irqrestore(&q->queue_lock, flags);
413 * Scheduler run of queue, if there are requests pending and no one in the
414 * driver that will restart queueing.
416 void bfq_schedule_dispatch(struct bfq_data *bfqd)
418 if (bfqd->queued != 0) {
419 bfq_log(bfqd, "schedule dispatch");
420 blk_mq_run_hw_queues(bfqd->queue, true);
424 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
425 #define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT)
427 #define bfq_sample_valid(samples) ((samples) > 80)
430 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
431 * We choose the request that is closesr to the head right now. Distance
432 * behind the head is penalized and only allowed to a certain extent.
434 static struct request *bfq_choose_req(struct bfq_data *bfqd,
439 sector_t s1, s2, d1 = 0, d2 = 0;
440 unsigned long back_max;
441 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
442 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
443 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
445 if (!rq1 || rq1 == rq2)
450 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
452 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
454 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
456 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
459 s1 = blk_rq_pos(rq1);
460 s2 = blk_rq_pos(rq2);
463 * By definition, 1KiB is 2 sectors.
465 back_max = bfqd->bfq_back_max * 2;
468 * Strict one way elevator _except_ in the case where we allow
469 * short backward seeks which are biased as twice the cost of a
470 * similar forward seek.
474 else if (s1 + back_max >= last)
475 d1 = (last - s1) * bfqd->bfq_back_penalty;
477 wrap |= BFQ_RQ1_WRAP;
481 else if (s2 + back_max >= last)
482 d2 = (last - s2) * bfqd->bfq_back_penalty;
484 wrap |= BFQ_RQ2_WRAP;
486 /* Found required data */
489 * By doing switch() on the bit mask "wrap" we avoid having to
490 * check two variables for all permutations: --> faster!
493 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
508 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
511 * Since both rqs are wrapped,
512 * start with the one that's further behind head
513 * (--> only *one* back seek required),
514 * since back seek takes more time than forward.
524 * Async I/O can easily starve sync I/O (both sync reads and sync
525 * writes), by consuming all tags. Similarly, storms of sync writes,
526 * such as those that sync(2) may trigger, can starve sync reads.
527 * Limit depths of async I/O and sync writes so as to counter both
530 static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
532 struct bfq_data *bfqd = data->q->elevator->elevator_data;
534 if (op_is_sync(op) && !op_is_write(op))
537 data->shallow_depth =
538 bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)];
540 bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
541 __func__, bfqd->wr_busy_queues, op_is_sync(op),
542 data->shallow_depth);
545 static struct bfq_queue *
546 bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
547 sector_t sector, struct rb_node **ret_parent,
548 struct rb_node ***rb_link)
550 struct rb_node **p, *parent;
551 struct bfq_queue *bfqq = NULL;
559 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
562 * Sort strictly based on sector. Smallest to the left,
563 * largest to the right.
565 if (sector > blk_rq_pos(bfqq->next_rq))
567 else if (sector < blk_rq_pos(bfqq->next_rq))
575 *ret_parent = parent;
579 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
580 (unsigned long long)sector,
581 bfqq ? bfqq->pid : 0);
586 static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
588 return bfqq->service_from_backlogged > 0 &&
589 time_is_before_jiffies(bfqq->first_IO_time +
590 bfq_merge_time_limit);
593 void bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
595 struct rb_node **p, *parent;
596 struct bfq_queue *__bfqq;
598 if (bfqq->pos_root) {
599 rb_erase(&bfqq->pos_node, bfqq->pos_root);
600 bfqq->pos_root = NULL;
604 * bfqq cannot be merged any longer (see comments in
605 * bfq_setup_cooperator): no point in adding bfqq into the
608 if (bfq_too_late_for_merging(bfqq))
611 if (bfq_class_idle(bfqq))
616 bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
617 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
618 blk_rq_pos(bfqq->next_rq), &parent, &p);
620 rb_link_node(&bfqq->pos_node, parent, p);
621 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
623 bfqq->pos_root = NULL;
627 * The following function returns true if every queue must receive the
628 * same share of the throughput (this condition is used when deciding
629 * whether idling may be disabled, see the comments in the function
630 * bfq_better_to_idle()).
632 * Such a scenario occurs when:
633 * 1) all active queues have the same weight,
634 * 2) all active queues belong to the same I/O-priority class,
635 * 3) all active groups at the same level in the groups tree have the same
637 * 4) all active groups at the same level in the groups tree have the same
638 * number of children.
640 * Unfortunately, keeping the necessary state for evaluating exactly
641 * the last two symmetry sub-conditions above would be quite complex
642 * and time consuming. Therefore this function evaluates, instead,
643 * only the following stronger three sub-conditions, for which it is
644 * much easier to maintain the needed state:
645 * 1) all active queues have the same weight,
646 * 2) all active queues belong to the same I/O-priority class,
647 * 3) there are no active groups.
648 * In particular, the last condition is always true if hierarchical
649 * support or the cgroups interface are not enabled, thus no state
650 * needs to be maintained in this case.
652 static bool bfq_symmetric_scenario(struct bfq_data *bfqd)
655 * For queue weights to differ, queue_weights_tree must contain
656 * at least two nodes.
658 bool varied_queue_weights = !RB_EMPTY_ROOT(&bfqd->queue_weights_tree) &&
659 (bfqd->queue_weights_tree.rb_node->rb_left ||
660 bfqd->queue_weights_tree.rb_node->rb_right);
662 bool multiple_classes_busy =
663 (bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
664 (bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
665 (bfqd->busy_queues[1] && bfqd->busy_queues[2]);
668 * For queue weights to differ, queue_weights_tree must contain
669 * at least two nodes.
671 return !(varied_queue_weights || multiple_classes_busy
672 #ifdef BFQ_GROUP_IOSCHED_ENABLED
673 || bfqd->num_groups_with_pending_reqs > 0
679 * If the weight-counter tree passed as input contains no counter for
680 * the weight of the input queue, then add that counter; otherwise just
681 * increment the existing counter.
683 * Note that weight-counter trees contain few nodes in mostly symmetric
684 * scenarios. For example, if all queues have the same weight, then the
685 * weight-counter tree for the queues may contain at most one node.
686 * This holds even if low_latency is on, because weight-raised queues
687 * are not inserted in the tree.
688 * In most scenarios, the rate at which nodes are created/destroyed
691 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq,
692 struct rb_root *root)
694 struct bfq_entity *entity = &bfqq->entity;
695 struct rb_node **new = &(root->rb_node), *parent = NULL;
698 * Do not insert if the queue is already associated with a
699 * counter, which happens if:
700 * 1) a request arrival has caused the queue to become both
701 * non-weight-raised, and hence change its weight, and
702 * backlogged; in this respect, each of the two events
703 * causes an invocation of this function,
704 * 2) this is the invocation of this function caused by the
705 * second event. This second invocation is actually useless,
706 * and we handle this fact by exiting immediately. More
707 * efficient or clearer solutions might possibly be adopted.
709 if (bfqq->weight_counter)
713 struct bfq_weight_counter *__counter = container_of(*new,
714 struct bfq_weight_counter,
718 if (entity->weight == __counter->weight) {
719 bfqq->weight_counter = __counter;
722 if (entity->weight < __counter->weight)
723 new = &((*new)->rb_left);
725 new = &((*new)->rb_right);
728 bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
732 * In the unlucky event of an allocation failure, we just
733 * exit. This will cause the weight of queue to not be
734 * considered in bfq_symmetric_scenario, which, in its turn,
735 * causes the scenario to be deemed wrongly symmetric in case
736 * bfqq's weight would have been the only weight making the
737 * scenario asymmetric. On the bright side, no unbalance will
738 * however occur when bfqq becomes inactive again (the
739 * invocation of this function is triggered by an activation
740 * of queue). In fact, bfq_weights_tree_remove does nothing
741 * if !bfqq->weight_counter.
743 if (unlikely(!bfqq->weight_counter))
746 bfqq->weight_counter->weight = entity->weight;
747 rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
748 rb_insert_color(&bfqq->weight_counter->weights_node, root);
751 bfqq->weight_counter->num_active++;
755 * Decrement the weight counter associated with the queue, and, if the
756 * counter reaches 0, remove the counter from the tree.
757 * See the comments to the function bfq_weights_tree_add() for considerations
760 void __bfq_weights_tree_remove(struct bfq_data *bfqd,
761 struct bfq_queue *bfqq,
762 struct rb_root *root)
764 if (!bfqq->weight_counter)
767 bfqq->weight_counter->num_active--;
768 if (bfqq->weight_counter->num_active > 0)
769 goto reset_entity_pointer;
771 rb_erase(&bfqq->weight_counter->weights_node, root);
772 kfree(bfqq->weight_counter);
774 reset_entity_pointer:
775 bfqq->weight_counter = NULL;
779 * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
780 * of active groups for each queue's inactive parent entity.
782 void bfq_weights_tree_remove(struct bfq_data *bfqd,
783 struct bfq_queue *bfqq)
785 struct bfq_entity *entity = bfqq->entity.parent;
787 __bfq_weights_tree_remove(bfqd, bfqq,
788 &bfqd->queue_weights_tree);
790 for_each_entity(entity) {
791 struct bfq_sched_data *sd = entity->my_sched_data;
793 if (sd->next_in_service || sd->in_service_entity) {
795 * entity is still active, because either
796 * next_in_service or in_service_entity is not
797 * NULL (see the comments on the definition of
798 * next_in_service for details on why
799 * in_service_entity must be checked too).
801 * As a consequence, its parent entities are
802 * active as well, and thus this loop must
809 * The decrement of num_groups_with_pending_reqs is
810 * not performed immediately upon the deactivation of
811 * entity, but it is delayed to when it also happens
812 * that the first leaf descendant bfqq of entity gets
813 * all its pending requests completed. The following
814 * instructions perform this delayed decrement, if
815 * needed. See the comments on
816 * num_groups_with_pending_reqs for details.
818 if (entity->in_groups_with_pending_reqs) {
819 entity->in_groups_with_pending_reqs = false;
820 bfqd->num_groups_with_pending_reqs--;
826 * Return expired entry, or NULL to just start from scratch in rbtree.
828 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
829 struct request *last)
833 if (bfq_bfqq_fifo_expire(bfqq))
836 bfq_mark_bfqq_fifo_expire(bfqq);
838 rq = rq_entry_fifo(bfqq->fifo.next);
840 if (rq == last || ktime_get_ns() < rq->fifo_time)
843 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
847 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
848 struct bfq_queue *bfqq,
849 struct request *last)
851 struct rb_node *rbnext = rb_next(&last->rb_node);
852 struct rb_node *rbprev = rb_prev(&last->rb_node);
853 struct request *next, *prev = NULL;
855 /* Follow expired path, else get first next available. */
856 next = bfq_check_fifo(bfqq, last);
861 prev = rb_entry_rq(rbprev);
864 next = rb_entry_rq(rbnext);
866 rbnext = rb_first(&bfqq->sort_list);
867 if (rbnext && rbnext != &last->rb_node)
868 next = rb_entry_rq(rbnext);
871 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
874 /* see the definition of bfq_async_charge_factor for details */
875 static unsigned long bfq_serv_to_charge(struct request *rq,
876 struct bfq_queue *bfqq)
878 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1)
879 return blk_rq_sectors(rq);
881 return blk_rq_sectors(rq) * bfq_async_charge_factor;
885 * bfq_updated_next_req - update the queue after a new next_rq selection.
886 * @bfqd: the device data the queue belongs to.
887 * @bfqq: the queue to update.
889 * If the first request of a queue changes we make sure that the queue
890 * has enough budget to serve at least its first request (if the
891 * request has grown). We do this because if the queue has not enough
892 * budget for its first request, it has to go through two dispatch
893 * rounds to actually get it dispatched.
895 static void bfq_updated_next_req(struct bfq_data *bfqd,
896 struct bfq_queue *bfqq)
898 struct bfq_entity *entity = &bfqq->entity;
899 struct request *next_rq = bfqq->next_rq;
900 unsigned long new_budget;
905 if (bfqq == bfqd->in_service_queue)
907 * In order not to break guarantees, budgets cannot be
908 * changed after an entity has been selected.
912 new_budget = max_t(unsigned long,
913 max_t(unsigned long, bfqq->max_budget,
914 bfq_serv_to_charge(next_rq, bfqq)),
916 if (entity->budget != new_budget) {
917 entity->budget = new_budget;
918 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
920 bfq_requeue_bfqq(bfqd, bfqq, false);
924 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
928 if (bfqd->bfq_wr_max_time > 0)
929 return bfqd->bfq_wr_max_time;
931 dur = bfqd->rate_dur_prod;
932 do_div(dur, bfqd->peak_rate);
935 * Limit duration between 3 and 25 seconds. The upper limit
936 * has been conservatively set after the following worst case:
937 * on a QEMU/KVM virtual machine
938 * - running in a slow PC
939 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
940 * - serving a heavy I/O workload, such as the sequential reading
942 * mplayer took 23 seconds to start, if constantly weight-raised.
944 * As for higher values than that accomodating the above bad
945 * scenario, tests show that higher values would often yield
946 * the opposite of the desired result, i.e., would worsen
947 * responsiveness by allowing non-interactive applications to
948 * preserve weight raising for too long.
950 * On the other end, lower values than 3 seconds make it
951 * difficult for most interactive tasks to complete their jobs
952 * before weight-raising finishes.
954 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
957 /* switch back from soft real-time to interactive weight raising */
958 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
959 struct bfq_data *bfqd)
961 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
962 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
963 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
967 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
968 struct bfq_io_cq *bic, bool bfq_already_existing)
970 unsigned int old_wr_coeff = bfqq->wr_coeff;
971 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
973 if (bic->saved_has_short_ttime)
974 bfq_mark_bfqq_has_short_ttime(bfqq);
976 bfq_clear_bfqq_has_short_ttime(bfqq);
978 if (bic->saved_IO_bound)
979 bfq_mark_bfqq_IO_bound(bfqq);
981 bfq_clear_bfqq_IO_bound(bfqq);
983 bfqq->ttime = bic->saved_ttime;
984 bfqq->wr_coeff = bic->saved_wr_coeff;
985 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
986 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
987 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
989 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
990 time_is_before_jiffies(bfqq->last_wr_start_finish +
991 bfqq->wr_cur_max_time))) {
992 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
993 !bfq_bfqq_in_large_burst(bfqq) &&
994 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
995 bfq_wr_duration(bfqd))) {
996 switch_back_to_interactive_wr(bfqq, bfqd);
999 bfq_log_bfqq(bfqq->bfqd, bfqq,
1000 "resume state: switching off wr");
1004 /* make sure weight will be updated, however we got here */
1005 bfqq->entity.prio_changed = 1;
1010 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1011 bfqd->wr_busy_queues++;
1012 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1013 bfqd->wr_busy_queues--;
1016 static int bfqq_process_refs(struct bfq_queue *bfqq)
1018 return bfqq->ref - bfqq->allocated - bfqq->entity.on_st;
1021 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1022 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1024 struct bfq_queue *item;
1025 struct hlist_node *n;
1027 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1028 hlist_del_init(&item->burst_list_node);
1029 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1030 bfqd->burst_size = 1;
1031 bfqd->burst_parent_entity = bfqq->entity.parent;
1034 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1035 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1037 /* Increment burst size to take into account also bfqq */
1040 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1041 struct bfq_queue *pos, *bfqq_item;
1042 struct hlist_node *n;
1045 * Enough queues have been activated shortly after each
1046 * other to consider this burst as large.
1048 bfqd->large_burst = true;
1051 * We can now mark all queues in the burst list as
1052 * belonging to a large burst.
1054 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1056 bfq_mark_bfqq_in_large_burst(bfqq_item);
1057 bfq_mark_bfqq_in_large_burst(bfqq);
1060 * From now on, and until the current burst finishes, any
1061 * new queue being activated shortly after the last queue
1062 * was inserted in the burst can be immediately marked as
1063 * belonging to a large burst. So the burst list is not
1064 * needed any more. Remove it.
1066 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1068 hlist_del_init(&pos->burst_list_node);
1070 * Burst not yet large: add bfqq to the burst list. Do
1071 * not increment the ref counter for bfqq, because bfqq
1072 * is removed from the burst list before freeing bfqq
1075 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1079 * If many queues belonging to the same group happen to be created
1080 * shortly after each other, then the processes associated with these
1081 * queues have typically a common goal. In particular, bursts of queue
1082 * creations are usually caused by services or applications that spawn
1083 * many parallel threads/processes. Examples are systemd during boot,
1084 * or git grep. To help these processes get their job done as soon as
1085 * possible, it is usually better to not grant either weight-raising
1086 * or device idling to their queues.
1088 * In this comment we describe, firstly, the reasons why this fact
1089 * holds, and, secondly, the next function, which implements the main
1090 * steps needed to properly mark these queues so that they can then be
1091 * treated in a different way.
1093 * The above services or applications benefit mostly from a high
1094 * throughput: the quicker the requests of the activated queues are
1095 * cumulatively served, the sooner the target job of these queues gets
1096 * completed. As a consequence, weight-raising any of these queues,
1097 * which also implies idling the device for it, is almost always
1098 * counterproductive. In most cases it just lowers throughput.
1100 * On the other hand, a burst of queue creations may be caused also by
1101 * the start of an application that does not consist of a lot of
1102 * parallel I/O-bound threads. In fact, with a complex application,
1103 * several short processes may need to be executed to start-up the
1104 * application. In this respect, to start an application as quickly as
1105 * possible, the best thing to do is in any case to privilege the I/O
1106 * related to the application with respect to all other
1107 * I/O. Therefore, the best strategy to start as quickly as possible
1108 * an application that causes a burst of queue creations is to
1109 * weight-raise all the queues created during the burst. This is the
1110 * exact opposite of the best strategy for the other type of bursts.
1112 * In the end, to take the best action for each of the two cases, the
1113 * two types of bursts need to be distinguished. Fortunately, this
1114 * seems relatively easy, by looking at the sizes of the bursts. In
1115 * particular, we found a threshold such that only bursts with a
1116 * larger size than that threshold are apparently caused by
1117 * services or commands such as systemd or git grep. For brevity,
1118 * hereafter we call just 'large' these bursts. BFQ *does not*
1119 * weight-raise queues whose creation occurs in a large burst. In
1120 * addition, for each of these queues BFQ performs or does not perform
1121 * idling depending on which choice boosts the throughput more. The
1122 * exact choice depends on the device and request pattern at
1125 * Unfortunately, false positives may occur while an interactive task
1126 * is starting (e.g., an application is being started). The
1127 * consequence is that the queues associated with the task do not
1128 * enjoy weight raising as expected. Fortunately these false positives
1129 * are very rare. They typically occur if some service happens to
1130 * start doing I/O exactly when the interactive task starts.
1132 * Turning back to the next function, it implements all the steps
1133 * needed to detect the occurrence of a large burst and to properly
1134 * mark all the queues belonging to it (so that they can then be
1135 * treated in a different way). This goal is achieved by maintaining a
1136 * "burst list" that holds, temporarily, the queues that belong to the
1137 * burst in progress. The list is then used to mark these queues as
1138 * belonging to a large burst if the burst does become large. The main
1139 * steps are the following.
1141 * . when the very first queue is created, the queue is inserted into the
1142 * list (as it could be the first queue in a possible burst)
1144 * . if the current burst has not yet become large, and a queue Q that does
1145 * not yet belong to the burst is activated shortly after the last time
1146 * at which a new queue entered the burst list, then the function appends
1147 * Q to the burst list
1149 * . if, as a consequence of the previous step, the burst size reaches
1150 * the large-burst threshold, then
1152 * . all the queues in the burst list are marked as belonging to a
1155 * . the burst list is deleted; in fact, the burst list already served
1156 * its purpose (keeping temporarily track of the queues in a burst,
1157 * so as to be able to mark them as belonging to a large burst in the
1158 * previous sub-step), and now is not needed any more
1160 * . the device enters a large-burst mode
1162 * . if a queue Q that does not belong to the burst is created while
1163 * the device is in large-burst mode and shortly after the last time
1164 * at which a queue either entered the burst list or was marked as
1165 * belonging to the current large burst, then Q is immediately marked
1166 * as belonging to a large burst.
1168 * . if a queue Q that does not belong to the burst is created a while
1169 * later, i.e., not shortly after, than the last time at which a queue
1170 * either entered the burst list or was marked as belonging to the
1171 * current large burst, then the current burst is deemed as finished and:
1173 * . the large-burst mode is reset if set
1175 * . the burst list is emptied
1177 * . Q is inserted in the burst list, as Q may be the first queue
1178 * in a possible new burst (then the burst list contains just Q
1181 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1184 * If bfqq is already in the burst list or is part of a large
1185 * burst, or finally has just been split, then there is
1186 * nothing else to do.
1188 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1189 bfq_bfqq_in_large_burst(bfqq) ||
1190 time_is_after_eq_jiffies(bfqq->split_time +
1191 msecs_to_jiffies(10)))
1195 * If bfqq's creation happens late enough, or bfqq belongs to
1196 * a different group than the burst group, then the current
1197 * burst is finished, and related data structures must be
1200 * In this respect, consider the special case where bfqq is
1201 * the very first queue created after BFQ is selected for this
1202 * device. In this case, last_ins_in_burst and
1203 * burst_parent_entity are not yet significant when we get
1204 * here. But it is easy to verify that, whether or not the
1205 * following condition is true, bfqq will end up being
1206 * inserted into the burst list. In particular the list will
1207 * happen to contain only bfqq. And this is exactly what has
1208 * to happen, as bfqq may be the first queue of the first
1211 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1212 bfqd->bfq_burst_interval) ||
1213 bfqq->entity.parent != bfqd->burst_parent_entity) {
1214 bfqd->large_burst = false;
1215 bfq_reset_burst_list(bfqd, bfqq);
1220 * If we get here, then bfqq is being activated shortly after the
1221 * last queue. So, if the current burst is also large, we can mark
1222 * bfqq as belonging to this large burst immediately.
1224 if (bfqd->large_burst) {
1225 bfq_mark_bfqq_in_large_burst(bfqq);
1230 * If we get here, then a large-burst state has not yet been
1231 * reached, but bfqq is being activated shortly after the last
1232 * queue. Then we add bfqq to the burst.
1234 bfq_add_to_burst(bfqd, bfqq);
1237 * At this point, bfqq either has been added to the current
1238 * burst or has caused the current burst to terminate and a
1239 * possible new burst to start. In particular, in the second
1240 * case, bfqq has become the first queue in the possible new
1241 * burst. In both cases last_ins_in_burst needs to be moved
1244 bfqd->last_ins_in_burst = jiffies;
1247 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1249 struct bfq_entity *entity = &bfqq->entity;
1251 return entity->budget - entity->service;
1255 * If enough samples have been computed, return the current max budget
1256 * stored in bfqd, which is dynamically updated according to the
1257 * estimated disk peak rate; otherwise return the default max budget
1259 static int bfq_max_budget(struct bfq_data *bfqd)
1261 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1262 return bfq_default_max_budget;
1264 return bfqd->bfq_max_budget;
1268 * Return min budget, which is a fraction of the current or default
1269 * max budget (trying with 1/32)
1271 static int bfq_min_budget(struct bfq_data *bfqd)
1273 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1274 return bfq_default_max_budget / 32;
1276 return bfqd->bfq_max_budget / 32;
1280 * The next function, invoked after the input queue bfqq switches from
1281 * idle to busy, updates the budget of bfqq. The function also tells
1282 * whether the in-service queue should be expired, by returning
1283 * true. The purpose of expiring the in-service queue is to give bfqq
1284 * the chance to possibly preempt the in-service queue, and the reason
1285 * for preempting the in-service queue is to achieve one of the two
1288 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1289 * expired because it has remained idle. In particular, bfqq may have
1290 * expired for one of the following two reasons:
1292 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1293 * and did not make it to issue a new request before its last
1294 * request was served;
1296 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1297 * a new request before the expiration of the idling-time.
1299 * Even if bfqq has expired for one of the above reasons, the process
1300 * associated with the queue may be however issuing requests greedily,
1301 * and thus be sensitive to the bandwidth it receives (bfqq may have
1302 * remained idle for other reasons: CPU high load, bfqq not enjoying
1303 * idling, I/O throttling somewhere in the path from the process to
1304 * the I/O scheduler, ...). But if, after every expiration for one of
1305 * the above two reasons, bfqq has to wait for the service of at least
1306 * one full budget of another queue before being served again, then
1307 * bfqq is likely to get a much lower bandwidth or resource time than
1308 * its reserved ones. To address this issue, two countermeasures need
1311 * First, the budget and the timestamps of bfqq need to be updated in
1312 * a special way on bfqq reactivation: they need to be updated as if
1313 * bfqq did not remain idle and did not expire. In fact, if they are
1314 * computed as if bfqq expired and remained idle until reactivation,
1315 * then the process associated with bfqq is treated as if, instead of
1316 * being greedy, it stopped issuing requests when bfqq remained idle,
1317 * and restarts issuing requests only on this reactivation. In other
1318 * words, the scheduler does not help the process recover the "service
1319 * hole" between bfqq expiration and reactivation. As a consequence,
1320 * the process receives a lower bandwidth than its reserved one. In
1321 * contrast, to recover this hole, the budget must be updated as if
1322 * bfqq was not expired at all before this reactivation, i.e., it must
1323 * be set to the value of the remaining budget when bfqq was
1324 * expired. Along the same line, timestamps need to be assigned the
1325 * value they had the last time bfqq was selected for service, i.e.,
1326 * before last expiration. Thus timestamps need to be back-shifted
1327 * with respect to their normal computation (see [1] for more details
1328 * on this tricky aspect).
1330 * Secondly, to allow the process to recover the hole, the in-service
1331 * queue must be expired too, to give bfqq the chance to preempt it
1332 * immediately. In fact, if bfqq has to wait for a full budget of the
1333 * in-service queue to be completed, then it may become impossible to
1334 * let the process recover the hole, even if the back-shifted
1335 * timestamps of bfqq are lower than those of the in-service queue. If
1336 * this happens for most or all of the holes, then the process may not
1337 * receive its reserved bandwidth. In this respect, it is worth noting
1338 * that, being the service of outstanding requests unpreemptible, a
1339 * little fraction of the holes may however be unrecoverable, thereby
1340 * causing a little loss of bandwidth.
1342 * The last important point is detecting whether bfqq does need this
1343 * bandwidth recovery. In this respect, the next function deems the
1344 * process associated with bfqq greedy, and thus allows it to recover
1345 * the hole, if: 1) the process is waiting for the arrival of a new
1346 * request (which implies that bfqq expired for one of the above two
1347 * reasons), and 2) such a request has arrived soon. The first
1348 * condition is controlled through the flag non_blocking_wait_rq,
1349 * while the second through the flag arrived_in_time. If both
1350 * conditions hold, then the function computes the budget in the
1351 * above-described special way, and signals that the in-service queue
1352 * should be expired. Timestamp back-shifting is done later in
1353 * __bfq_activate_entity.
1355 * 2. Reduce latency. Even if timestamps are not backshifted to let
1356 * the process associated with bfqq recover a service hole, bfqq may
1357 * however happen to have, after being (re)activated, a lower finish
1358 * timestamp than the in-service queue. That is, the next budget of
1359 * bfqq may have to be completed before the one of the in-service
1360 * queue. If this is the case, then preempting the in-service queue
1361 * allows this goal to be achieved, apart from the unpreemptible,
1362 * outstanding requests mentioned above.
1364 * Unfortunately, regardless of which of the above two goals one wants
1365 * to achieve, service trees need first to be updated to know whether
1366 * the in-service queue must be preempted. To have service trees
1367 * correctly updated, the in-service queue must be expired and
1368 * rescheduled, and bfqq must be scheduled too. This is one of the
1369 * most costly operations (in future versions, the scheduling
1370 * mechanism may be re-designed in such a way to make it possible to
1371 * know whether preemption is needed without needing to update service
1372 * trees). In addition, queue preemptions almost always cause random
1373 * I/O, and thus loss of throughput. Because of these facts, the next
1374 * function adopts the following simple scheme to avoid both costly
1375 * operations and too frequent preemptions: it requests the expiration
1376 * of the in-service queue (unconditionally) only for queues that need
1377 * to recover a hole, or that either are weight-raised or deserve to
1380 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1381 struct bfq_queue *bfqq,
1382 bool arrived_in_time,
1383 bool wr_or_deserves_wr)
1385 struct bfq_entity *entity = &bfqq->entity;
1388 * In the next compound condition, we check also whether there
1389 * is some budget left, because otherwise there is no point in
1390 * trying to go on serving bfqq with this same budget: bfqq
1391 * would be expired immediately after being selected for
1392 * service. This would only cause useless overhead.
1394 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1395 bfq_bfqq_budget_left(bfqq) > 0) {
1397 * We do not clear the flag non_blocking_wait_rq here, as
1398 * the latter is used in bfq_activate_bfqq to signal
1399 * that timestamps need to be back-shifted (and is
1400 * cleared right after).
1404 * In next assignment we rely on that either
1405 * entity->service or entity->budget are not updated
1406 * on expiration if bfqq is empty (see
1407 * __bfq_bfqq_recalc_budget). Thus both quantities
1408 * remain unchanged after such an expiration, and the
1409 * following statement therefore assigns to
1410 * entity->budget the remaining budget on such an
1413 entity->budget = min_t(unsigned long,
1414 bfq_bfqq_budget_left(bfqq),
1418 * At this point, we have used entity->service to get
1419 * the budget left (needed for updating
1420 * entity->budget). Thus we finally can, and have to,
1421 * reset entity->service. The latter must be reset
1422 * because bfqq would otherwise be charged again for
1423 * the service it has received during its previous
1426 entity->service = 0;
1432 * We can finally complete expiration, by setting service to 0.
1434 entity->service = 0;
1435 entity->budget = max_t(unsigned long, bfqq->max_budget,
1436 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1437 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1438 return wr_or_deserves_wr;
1442 * Return the farthest past time instant according to jiffies
1445 static unsigned long bfq_smallest_from_now(void)
1447 return jiffies - MAX_JIFFY_OFFSET;
1450 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1451 struct bfq_queue *bfqq,
1452 unsigned int old_wr_coeff,
1453 bool wr_or_deserves_wr,
1458 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1459 /* start a weight-raising period */
1461 bfqq->service_from_wr = 0;
1462 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1463 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1466 * No interactive weight raising in progress
1467 * here: assign minus infinity to
1468 * wr_start_at_switch_to_srt, to make sure
1469 * that, at the end of the soft-real-time
1470 * weight raising periods that is starting
1471 * now, no interactive weight-raising period
1472 * may be wrongly considered as still in
1473 * progress (and thus actually started by
1476 bfqq->wr_start_at_switch_to_srt =
1477 bfq_smallest_from_now();
1478 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1479 BFQ_SOFTRT_WEIGHT_FACTOR;
1480 bfqq->wr_cur_max_time =
1481 bfqd->bfq_wr_rt_max_time;
1485 * If needed, further reduce budget to make sure it is
1486 * close to bfqq's backlog, so as to reduce the
1487 * scheduling-error component due to a too large
1488 * budget. Do not care about throughput consequences,
1489 * but only about latency. Finally, do not assign a
1490 * too small budget either, to avoid increasing
1491 * latency by causing too frequent expirations.
1493 bfqq->entity.budget = min_t(unsigned long,
1494 bfqq->entity.budget,
1495 2 * bfq_min_budget(bfqd));
1496 } else if (old_wr_coeff > 1) {
1497 if (interactive) { /* update wr coeff and duration */
1498 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1499 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1500 } else if (in_burst)
1504 * The application is now or still meeting the
1505 * requirements for being deemed soft rt. We
1506 * can then correctly and safely (re)charge
1507 * the weight-raising duration for the
1508 * application with the weight-raising
1509 * duration for soft rt applications.
1511 * In particular, doing this recharge now, i.e.,
1512 * before the weight-raising period for the
1513 * application finishes, reduces the probability
1514 * of the following negative scenario:
1515 * 1) the weight of a soft rt application is
1516 * raised at startup (as for any newly
1517 * created application),
1518 * 2) since the application is not interactive,
1519 * at a certain time weight-raising is
1520 * stopped for the application,
1521 * 3) at that time the application happens to
1522 * still have pending requests, and hence
1523 * is destined to not have a chance to be
1524 * deemed soft rt before these requests are
1525 * completed (see the comments to the
1526 * function bfq_bfqq_softrt_next_start()
1527 * for details on soft rt detection),
1528 * 4) these pending requests experience a high
1529 * latency because the application is not
1530 * weight-raised while they are pending.
1532 if (bfqq->wr_cur_max_time !=
1533 bfqd->bfq_wr_rt_max_time) {
1534 bfqq->wr_start_at_switch_to_srt =
1535 bfqq->last_wr_start_finish;
1537 bfqq->wr_cur_max_time =
1538 bfqd->bfq_wr_rt_max_time;
1539 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1540 BFQ_SOFTRT_WEIGHT_FACTOR;
1542 bfqq->last_wr_start_finish = jiffies;
1547 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1548 struct bfq_queue *bfqq)
1550 return bfqq->dispatched == 0 &&
1551 time_is_before_jiffies(
1552 bfqq->budget_timeout +
1553 bfqd->bfq_wr_min_idle_time);
1556 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1557 struct bfq_queue *bfqq,
1562 bool soft_rt, in_burst, wr_or_deserves_wr,
1563 bfqq_wants_to_preempt,
1564 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1566 * See the comments on
1567 * bfq_bfqq_update_budg_for_activation for
1568 * details on the usage of the next variable.
1570 arrived_in_time = ktime_get_ns() <=
1571 bfqq->ttime.last_end_request +
1572 bfqd->bfq_slice_idle * 3;
1576 * bfqq deserves to be weight-raised if:
1578 * - it does not belong to a large burst,
1579 * - it has been idle for enough time or is soft real-time,
1580 * - is linked to a bfq_io_cq (it is not shared in any sense).
1582 in_burst = bfq_bfqq_in_large_burst(bfqq);
1583 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1585 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1586 bfqq->dispatched == 0;
1587 *interactive = !in_burst && idle_for_long_time;
1588 wr_or_deserves_wr = bfqd->low_latency &&
1589 (bfqq->wr_coeff > 1 ||
1590 (bfq_bfqq_sync(bfqq) &&
1591 bfqq->bic && (*interactive || soft_rt)));
1594 * Using the last flag, update budget and check whether bfqq
1595 * may want to preempt the in-service queue.
1597 bfqq_wants_to_preempt =
1598 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1603 * If bfqq happened to be activated in a burst, but has been
1604 * idle for much more than an interactive queue, then we
1605 * assume that, in the overall I/O initiated in the burst, the
1606 * I/O associated with bfqq is finished. So bfqq does not need
1607 * to be treated as a queue belonging to a burst
1608 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1609 * if set, and remove bfqq from the burst list if it's
1610 * there. We do not decrement burst_size, because the fact
1611 * that bfqq does not need to belong to the burst list any
1612 * more does not invalidate the fact that bfqq was created in
1615 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1616 idle_for_long_time &&
1617 time_is_before_jiffies(
1618 bfqq->budget_timeout +
1619 msecs_to_jiffies(10000))) {
1620 hlist_del_init(&bfqq->burst_list_node);
1621 bfq_clear_bfqq_in_large_burst(bfqq);
1624 bfq_clear_bfqq_just_created(bfqq);
1627 if (!bfq_bfqq_IO_bound(bfqq)) {
1628 if (arrived_in_time) {
1629 bfqq->requests_within_timer++;
1630 if (bfqq->requests_within_timer >=
1631 bfqd->bfq_requests_within_timer)
1632 bfq_mark_bfqq_IO_bound(bfqq);
1634 bfqq->requests_within_timer = 0;
1637 if (bfqd->low_latency) {
1638 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1641 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1643 if (time_is_before_jiffies(bfqq->split_time +
1644 bfqd->bfq_wr_min_idle_time)) {
1645 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1652 if (old_wr_coeff != bfqq->wr_coeff)
1653 bfqq->entity.prio_changed = 1;
1657 bfqq->last_idle_bklogged = jiffies;
1658 bfqq->service_from_backlogged = 0;
1659 bfq_clear_bfqq_softrt_update(bfqq);
1661 bfq_add_bfqq_busy(bfqd, bfqq);
1664 * Expire in-service queue only if preemption may be needed
1665 * for guarantees. In this respect, the function
1666 * next_queue_may_preempt just checks a simple, necessary
1667 * condition, and not a sufficient condition based on
1668 * timestamps. In fact, for the latter condition to be
1669 * evaluated, timestamps would need first to be updated, and
1670 * this operation is quite costly (see the comments on the
1671 * function bfq_bfqq_update_budg_for_activation).
1673 if (bfqd->in_service_queue && bfqq_wants_to_preempt &&
1674 bfqd->in_service_queue->wr_coeff < bfqq->wr_coeff &&
1675 next_queue_may_preempt(bfqd))
1676 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1677 false, BFQQE_PREEMPTED);
1680 static void bfq_add_request(struct request *rq)
1682 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1683 struct bfq_data *bfqd = bfqq->bfqd;
1684 struct request *next_rq, *prev;
1685 unsigned int old_wr_coeff = bfqq->wr_coeff;
1686 bool interactive = false;
1688 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1689 bfqq->queued[rq_is_sync(rq)]++;
1692 elv_rb_add(&bfqq->sort_list, rq);
1695 * Check if this request is a better next-serve candidate.
1697 prev = bfqq->next_rq;
1698 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
1699 bfqq->next_rq = next_rq;
1702 * Adjust priority tree position, if next_rq changes.
1704 if (prev != bfqq->next_rq)
1705 bfq_pos_tree_add_move(bfqd, bfqq);
1707 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
1708 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
1711 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
1712 time_is_before_jiffies(
1713 bfqq->last_wr_start_finish +
1714 bfqd->bfq_wr_min_inter_arr_async)) {
1715 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1716 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1718 bfqd->wr_busy_queues++;
1719 bfqq->entity.prio_changed = 1;
1721 if (prev != bfqq->next_rq)
1722 bfq_updated_next_req(bfqd, bfqq);
1726 * Assign jiffies to last_wr_start_finish in the following
1729 * . if bfqq is not going to be weight-raised, because, for
1730 * non weight-raised queues, last_wr_start_finish stores the
1731 * arrival time of the last request; as of now, this piece
1732 * of information is used only for deciding whether to
1733 * weight-raise async queues
1735 * . if bfqq is not weight-raised, because, if bfqq is now
1736 * switching to weight-raised, then last_wr_start_finish
1737 * stores the time when weight-raising starts
1739 * . if bfqq is interactive, because, regardless of whether
1740 * bfqq is currently weight-raised, the weight-raising
1741 * period must start or restart (this case is considered
1742 * separately because it is not detected by the above
1743 * conditions, if bfqq is already weight-raised)
1745 * last_wr_start_finish has to be updated also if bfqq is soft
1746 * real-time, because the weight-raising period is constantly
1747 * restarted on idle-to-busy transitions for these queues, but
1748 * this is already done in bfq_bfqq_handle_idle_busy_switch if
1751 if (bfqd->low_latency &&
1752 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
1753 bfqq->last_wr_start_finish = jiffies;
1756 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
1758 struct request_queue *q)
1760 struct bfq_queue *bfqq = bfqd->bio_bfqq;
1764 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
1769 static sector_t get_sdist(sector_t last_pos, struct request *rq)
1772 return abs(blk_rq_pos(rq) - last_pos);
1777 #if 0 /* Still not clear if we can do without next two functions */
1778 static void bfq_activate_request(struct request_queue *q, struct request *rq)
1780 struct bfq_data *bfqd = q->elevator->elevator_data;
1782 bfqd->rq_in_driver++;
1785 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
1787 struct bfq_data *bfqd = q->elevator->elevator_data;
1789 bfqd->rq_in_driver--;
1793 static void bfq_remove_request(struct request_queue *q,
1796 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1797 struct bfq_data *bfqd = bfqq->bfqd;
1798 const int sync = rq_is_sync(rq);
1800 if (bfqq->next_rq == rq) {
1801 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
1802 bfq_updated_next_req(bfqd, bfqq);
1805 if (rq->queuelist.prev != &rq->queuelist)
1806 list_del_init(&rq->queuelist);
1807 bfqq->queued[sync]--;
1809 elv_rb_del(&bfqq->sort_list, rq);
1811 elv_rqhash_del(q, rq);
1812 if (q->last_merge == rq)
1813 q->last_merge = NULL;
1815 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
1816 bfqq->next_rq = NULL;
1818 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
1819 bfq_del_bfqq_busy(bfqd, bfqq, false);
1821 * bfqq emptied. In normal operation, when
1822 * bfqq is empty, bfqq->entity.service and
1823 * bfqq->entity.budget must contain,
1824 * respectively, the service received and the
1825 * budget used last time bfqq emptied. These
1826 * facts do not hold in this case, as at least
1827 * this last removal occurred while bfqq is
1828 * not in service. To avoid inconsistencies,
1829 * reset both bfqq->entity.service and
1830 * bfqq->entity.budget, if bfqq has still a
1831 * process that may issue I/O requests to it.
1833 bfqq->entity.budget = bfqq->entity.service = 0;
1837 * Remove queue from request-position tree as it is empty.
1839 if (bfqq->pos_root) {
1840 rb_erase(&bfqq->pos_node, bfqq->pos_root);
1841 bfqq->pos_root = NULL;
1844 bfq_pos_tree_add_move(bfqd, bfqq);
1847 if (rq->cmd_flags & REQ_META)
1848 bfqq->meta_pending--;
1852 static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio)
1854 struct request_queue *q = hctx->queue;
1855 struct bfq_data *bfqd = q->elevator->elevator_data;
1856 struct request *free = NULL;
1858 * bfq_bic_lookup grabs the queue_lock: invoke it now and
1859 * store its return value for later use, to avoid nesting
1860 * queue_lock inside the bfqd->lock. We assume that the bic
1861 * returned by bfq_bic_lookup does not go away before
1862 * bfqd->lock is taken.
1864 struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
1867 spin_lock_irq(&bfqd->lock);
1870 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
1872 bfqd->bio_bfqq = NULL;
1873 bfqd->bio_bic = bic;
1875 ret = blk_mq_sched_try_merge(q, bio, &free);
1878 blk_mq_free_request(free);
1879 spin_unlock_irq(&bfqd->lock);
1884 static int bfq_request_merge(struct request_queue *q, struct request **req,
1887 struct bfq_data *bfqd = q->elevator->elevator_data;
1888 struct request *__rq;
1890 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
1891 if (__rq && elv_bio_merge_ok(__rq, bio)) {
1893 return ELEVATOR_FRONT_MERGE;
1896 return ELEVATOR_NO_MERGE;
1899 static struct bfq_queue *bfq_init_rq(struct request *rq);
1901 static void bfq_request_merged(struct request_queue *q, struct request *req,
1902 enum elv_merge type)
1904 if (type == ELEVATOR_FRONT_MERGE &&
1905 rb_prev(&req->rb_node) &&
1907 blk_rq_pos(container_of(rb_prev(&req->rb_node),
1908 struct request, rb_node))) {
1909 struct bfq_queue *bfqq = bfq_init_rq(req);
1910 struct bfq_data *bfqd = bfqq->bfqd;
1911 struct request *prev, *next_rq;
1913 /* Reposition request in its sort_list */
1914 elv_rb_del(&bfqq->sort_list, req);
1915 elv_rb_add(&bfqq->sort_list, req);
1917 /* Choose next request to be served for bfqq */
1918 prev = bfqq->next_rq;
1919 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
1920 bfqd->last_position);
1921 bfqq->next_rq = next_rq;
1923 * If next_rq changes, update both the queue's budget to
1924 * fit the new request and the queue's position in its
1927 if (prev != bfqq->next_rq) {
1928 bfq_updated_next_req(bfqd, bfqq);
1929 bfq_pos_tree_add_move(bfqd, bfqq);
1935 * This function is called to notify the scheduler that the requests
1936 * rq and 'next' have been merged, with 'next' going away. BFQ
1937 * exploits this hook to address the following issue: if 'next' has a
1938 * fifo_time lower that rq, then the fifo_time of rq must be set to
1939 * the value of 'next', to not forget the greater age of 'next'.
1941 * NOTE: in this function we assume that rq is in a bfq_queue, basing
1942 * on that rq is picked from the hash table q->elevator->hash, which,
1943 * in its turn, is filled only with I/O requests present in
1944 * bfq_queues, while BFQ is in use for the request queue q. In fact,
1945 * the function that fills this hash table (elv_rqhash_add) is called
1946 * only by bfq_insert_request.
1948 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
1949 struct request *next)
1951 struct bfq_queue *bfqq = bfq_init_rq(rq),
1952 *next_bfqq = bfq_init_rq(next);
1955 * If next and rq belong to the same bfq_queue and next is older
1956 * than rq, then reposition rq in the fifo (by substituting next
1957 * with rq). Otherwise, if next and rq belong to different
1958 * bfq_queues, never reposition rq: in fact, we would have to
1959 * reposition it with respect to next's position in its own fifo,
1960 * which would most certainly be too expensive with respect to
1963 if (bfqq == next_bfqq &&
1964 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1965 next->fifo_time < rq->fifo_time) {
1966 list_del_init(&rq->queuelist);
1967 list_replace_init(&next->queuelist, &rq->queuelist);
1968 rq->fifo_time = next->fifo_time;
1971 if (bfqq->next_rq == next)
1974 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
1977 /* Must be called with bfqq != NULL */
1978 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
1980 if (bfq_bfqq_busy(bfqq))
1981 bfqq->bfqd->wr_busy_queues--;
1983 bfqq->wr_cur_max_time = 0;
1984 bfqq->last_wr_start_finish = jiffies;
1986 * Trigger a weight change on the next invocation of
1987 * __bfq_entity_update_weight_prio.
1989 bfqq->entity.prio_changed = 1;
1992 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
1993 struct bfq_group *bfqg)
1997 for (i = 0; i < 2; i++)
1998 for (j = 0; j < IOPRIO_BE_NR; j++)
1999 if (bfqg->async_bfqq[i][j])
2000 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2001 if (bfqg->async_idle_bfqq)
2002 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2005 static void bfq_end_wr(struct bfq_data *bfqd)
2007 struct bfq_queue *bfqq;
2009 spin_lock_irq(&bfqd->lock);
2011 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2012 bfq_bfqq_end_wr(bfqq);
2013 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2014 bfq_bfqq_end_wr(bfqq);
2015 bfq_end_wr_async(bfqd);
2017 spin_unlock_irq(&bfqd->lock);
2020 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2023 return blk_rq_pos(io_struct);
2025 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2028 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2031 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2035 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2036 struct bfq_queue *bfqq,
2039 struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
2040 struct rb_node *parent, *node;
2041 struct bfq_queue *__bfqq;
2043 if (RB_EMPTY_ROOT(root))
2047 * First, if we find a request starting at the end of the last
2048 * request, choose it.
2050 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2055 * If the exact sector wasn't found, the parent of the NULL leaf
2056 * will contain the closest sector (rq_pos_tree sorted by
2057 * next_request position).
2059 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2060 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2063 if (blk_rq_pos(__bfqq->next_rq) < sector)
2064 node = rb_next(&__bfqq->pos_node);
2066 node = rb_prev(&__bfqq->pos_node);
2070 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2071 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2077 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2078 struct bfq_queue *cur_bfqq,
2081 struct bfq_queue *bfqq;
2084 * We shall notice if some of the queues are cooperating,
2085 * e.g., working closely on the same area of the device. In
2086 * that case, we can group them together and: 1) don't waste
2087 * time idling, and 2) serve the union of their requests in
2088 * the best possible order for throughput.
2090 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2091 if (!bfqq || bfqq == cur_bfqq)
2097 static struct bfq_queue *
2098 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2100 int process_refs, new_process_refs;
2101 struct bfq_queue *__bfqq;
2104 * If there are no process references on the new_bfqq, then it is
2105 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2106 * may have dropped their last reference (not just their last process
2109 if (!bfqq_process_refs(new_bfqq))
2112 /* Avoid a circular list and skip interim queue merges. */
2113 while ((__bfqq = new_bfqq->new_bfqq)) {
2119 process_refs = bfqq_process_refs(bfqq);
2120 new_process_refs = bfqq_process_refs(new_bfqq);
2122 * If the process for the bfqq has gone away, there is no
2123 * sense in merging the queues.
2125 if (process_refs == 0 || new_process_refs == 0)
2128 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2132 * Merging is just a redirection: the requests of the process
2133 * owning one of the two queues are redirected to the other queue.
2134 * The latter queue, in its turn, is set as shared if this is the
2135 * first time that the requests of some process are redirected to
2138 * We redirect bfqq to new_bfqq and not the opposite, because
2139 * we are in the context of the process owning bfqq, thus we
2140 * have the io_cq of this process. So we can immediately
2141 * configure this io_cq to redirect the requests of the
2142 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2143 * not available any more (new_bfqq->bic == NULL).
2145 * Anyway, even in case new_bfqq coincides with the in-service
2146 * queue, redirecting requests the in-service queue is the
2147 * best option, as we feed the in-service queue with new
2148 * requests close to the last request served and, by doing so,
2149 * are likely to increase the throughput.
2151 bfqq->new_bfqq = new_bfqq;
2152 new_bfqq->ref += process_refs;
2156 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2157 struct bfq_queue *new_bfqq)
2159 if (bfq_too_late_for_merging(new_bfqq))
2162 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2163 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2167 * If either of the queues has already been detected as seeky,
2168 * then merging it with the other queue is unlikely to lead to
2171 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2175 * Interleaved I/O is known to be done by (some) applications
2176 * only for reads, so it does not make sense to merge async
2179 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2186 * Attempt to schedule a merge of bfqq with the currently in-service
2187 * queue or with a close queue among the scheduled queues. Return
2188 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2189 * structure otherwise.
2191 * The OOM queue is not allowed to participate to cooperation: in fact, since
2192 * the requests temporarily redirected to the OOM queue could be redirected
2193 * again to dedicated queues at any time, the state needed to correctly
2194 * handle merging with the OOM queue would be quite complex and expensive
2195 * to maintain. Besides, in such a critical condition as an out of memory,
2196 * the benefits of queue merging may be little relevant, or even negligible.
2198 * WARNING: queue merging may impair fairness among non-weight raised
2199 * queues, for at least two reasons: 1) the original weight of a
2200 * merged queue may change during the merged state, 2) even being the
2201 * weight the same, a merged queue may be bloated with many more
2202 * requests than the ones produced by its originally-associated
2205 static struct bfq_queue *
2206 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2207 void *io_struct, bool request)
2209 struct bfq_queue *in_service_bfqq, *new_bfqq;
2212 * Prevent bfqq from being merged if it has been created too
2213 * long ago. The idea is that true cooperating processes, and
2214 * thus their associated bfq_queues, are supposed to be
2215 * created shortly after each other. This is the case, e.g.,
2216 * for KVM/QEMU and dump I/O threads. Basing on this
2217 * assumption, the following filtering greatly reduces the
2218 * probability that two non-cooperating processes, which just
2219 * happen to do close I/O for some short time interval, have
2220 * their queues merged by mistake.
2222 if (bfq_too_late_for_merging(bfqq))
2226 return bfqq->new_bfqq;
2228 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2231 /* If there is only one backlogged queue, don't search. */
2232 if (bfq_tot_busy_queues(bfqd) == 1)
2235 in_service_bfqq = bfqd->in_service_queue;
2237 if (in_service_bfqq && in_service_bfqq != bfqq &&
2238 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2239 bfq_rq_close_to_sector(io_struct, request, bfqd->last_position) &&
2240 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2241 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2242 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2247 * Check whether there is a cooperator among currently scheduled
2248 * queues. The only thing we need is that the bio/request is not
2249 * NULL, as we need it to establish whether a cooperator exists.
2251 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2252 bfq_io_struct_pos(io_struct, request));
2254 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2255 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2256 return bfq_setup_merge(bfqq, new_bfqq);
2261 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2263 struct bfq_io_cq *bic = bfqq->bic;
2266 * If !bfqq->bic, the queue is already shared or its requests
2267 * have already been redirected to a shared queue; both idle window
2268 * and weight raising state have already been saved. Do nothing.
2273 bic->saved_ttime = bfqq->ttime;
2274 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2275 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2276 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2277 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2278 if (unlikely(bfq_bfqq_just_created(bfqq) &&
2279 !bfq_bfqq_in_large_burst(bfqq) &&
2280 bfqq->bfqd->low_latency)) {
2282 * bfqq being merged right after being created: bfqq
2283 * would have deserved interactive weight raising, but
2284 * did not make it to be set in a weight-raised state,
2285 * because of this early merge. Store directly the
2286 * weight-raising state that would have been assigned
2287 * to bfqq, so that to avoid that bfqq unjustly fails
2288 * to enjoy weight raising if split soon.
2290 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
2291 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2292 bic->saved_last_wr_start_finish = jiffies;
2294 bic->saved_wr_coeff = bfqq->wr_coeff;
2295 bic->saved_wr_start_at_switch_to_srt =
2296 bfqq->wr_start_at_switch_to_srt;
2297 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2298 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2303 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2304 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2306 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2307 (unsigned long)new_bfqq->pid);
2308 /* Save weight raising and idle window of the merged queues */
2309 bfq_bfqq_save_state(bfqq);
2310 bfq_bfqq_save_state(new_bfqq);
2311 if (bfq_bfqq_IO_bound(bfqq))
2312 bfq_mark_bfqq_IO_bound(new_bfqq);
2313 bfq_clear_bfqq_IO_bound(bfqq);
2316 * If bfqq is weight-raised, then let new_bfqq inherit
2317 * weight-raising. To reduce false positives, neglect the case
2318 * where bfqq has just been created, but has not yet made it
2319 * to be weight-raised (which may happen because EQM may merge
2320 * bfqq even before bfq_add_request is executed for the first
2321 * time for bfqq). Handling this case would however be very
2322 * easy, thanks to the flag just_created.
2324 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2325 new_bfqq->wr_coeff = bfqq->wr_coeff;
2326 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2327 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2328 new_bfqq->wr_start_at_switch_to_srt =
2329 bfqq->wr_start_at_switch_to_srt;
2330 if (bfq_bfqq_busy(new_bfqq))
2331 bfqd->wr_busy_queues++;
2332 new_bfqq->entity.prio_changed = 1;
2335 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2337 bfqq->entity.prio_changed = 1;
2338 if (bfq_bfqq_busy(bfqq))
2339 bfqd->wr_busy_queues--;
2342 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2343 bfqd->wr_busy_queues);
2346 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2348 bic_set_bfqq(bic, new_bfqq, 1);
2349 bfq_mark_bfqq_coop(new_bfqq);
2351 * new_bfqq now belongs to at least two bics (it is a shared queue):
2352 * set new_bfqq->bic to NULL. bfqq either:
2353 * - does not belong to any bic any more, and hence bfqq->bic must
2354 * be set to NULL, or
2355 * - is a queue whose owning bics have already been redirected to a
2356 * different queue, hence the queue is destined to not belong to
2357 * any bic soon and bfqq->bic is already NULL (therefore the next
2358 * assignment causes no harm).
2360 new_bfqq->bic = NULL;
2362 /* release process reference to bfqq */
2363 bfq_put_queue(bfqq);
2366 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2369 struct bfq_data *bfqd = q->elevator->elevator_data;
2370 bool is_sync = op_is_sync(bio->bi_opf);
2371 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2374 * Disallow merge of a sync bio into an async request.
2376 if (is_sync && !rq_is_sync(rq))
2380 * Lookup the bfqq that this bio will be queued with. Allow
2381 * merge only if rq is queued there.
2387 * We take advantage of this function to perform an early merge
2388 * of the queues of possible cooperating processes.
2390 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2393 * bic still points to bfqq, then it has not yet been
2394 * redirected to some other bfq_queue, and a queue
2395 * merge beween bfqq and new_bfqq can be safely
2396 * fulfillled, i.e., bic can be redirected to new_bfqq
2397 * and bfqq can be put.
2399 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2402 * If we get here, bio will be queued into new_queue,
2403 * so use new_bfqq to decide whether bio and rq can be
2409 * Change also bqfd->bio_bfqq, as
2410 * bfqd->bio_bic now points to new_bfqq, and
2411 * this function may be invoked again (and then may
2412 * use again bqfd->bio_bfqq).
2414 bfqd->bio_bfqq = bfqq;
2417 return bfqq == RQ_BFQQ(rq);
2421 * Set the maximum time for the in-service queue to consume its
2422 * budget. This prevents seeky processes from lowering the throughput.
2423 * In practice, a time-slice service scheme is used with seeky
2426 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2427 struct bfq_queue *bfqq)
2429 unsigned int timeout_coeff;
2431 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2434 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2436 bfqd->last_budget_start = ktime_get();
2438 bfqq->budget_timeout = jiffies +
2439 bfqd->bfq_timeout * timeout_coeff;
2442 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2443 struct bfq_queue *bfqq)
2446 bfq_clear_bfqq_fifo_expire(bfqq);
2448 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2450 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2451 bfqq->wr_coeff > 1 &&
2452 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2453 time_is_before_jiffies(bfqq->budget_timeout)) {
2455 * For soft real-time queues, move the start
2456 * of the weight-raising period forward by the
2457 * time the queue has not received any
2458 * service. Otherwise, a relatively long
2459 * service delay is likely to cause the
2460 * weight-raising period of the queue to end,
2461 * because of the short duration of the
2462 * weight-raising period of a soft real-time
2463 * queue. It is worth noting that this move
2464 * is not so dangerous for the other queues,
2465 * because soft real-time queues are not
2468 * To not add a further variable, we use the
2469 * overloaded field budget_timeout to
2470 * determine for how long the queue has not
2471 * received service, i.e., how much time has
2472 * elapsed since the queue expired. However,
2473 * this is a little imprecise, because
2474 * budget_timeout is set to jiffies if bfqq
2475 * not only expires, but also remains with no
2478 if (time_after(bfqq->budget_timeout,
2479 bfqq->last_wr_start_finish))
2480 bfqq->last_wr_start_finish +=
2481 jiffies - bfqq->budget_timeout;
2483 bfqq->last_wr_start_finish = jiffies;
2486 bfq_set_budget_timeout(bfqd, bfqq);
2487 bfq_log_bfqq(bfqd, bfqq,
2488 "set_in_service_queue, cur-budget = %d",
2489 bfqq->entity.budget);
2492 bfqd->in_service_queue = bfqq;
2496 * Get and set a new queue for service.
2498 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2500 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2502 __bfq_set_in_service_queue(bfqd, bfqq);
2506 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2508 struct bfq_queue *bfqq = bfqd->in_service_queue;
2511 bfq_mark_bfqq_wait_request(bfqq);
2514 * We don't want to idle for seeks, but we do want to allow
2515 * fair distribution of slice time for a process doing back-to-back
2516 * seeks. So allow a little bit of time for him to submit a new rq.
2518 sl = bfqd->bfq_slice_idle;
2520 * Unless the queue is being weight-raised or the scenario is
2521 * asymmetric, grant only minimum idle time if the queue
2522 * is seeky. A long idling is preserved for a weight-raised
2523 * queue, or, more in general, in an asymmetric scenario,
2524 * because a long idling is needed for guaranteeing to a queue
2525 * its reserved share of the throughput (in particular, it is
2526 * needed if the queue has a higher weight than some other
2529 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
2530 bfq_symmetric_scenario(bfqd))
2531 sl = min_t(u64, sl, BFQ_MIN_TT);
2533 bfqd->last_idling_start = ktime_get();
2534 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
2536 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
2540 * In autotuning mode, max_budget is dynamically recomputed as the
2541 * amount of sectors transferred in timeout at the estimated peak
2542 * rate. This enables BFQ to utilize a full timeslice with a full
2543 * budget, even if the in-service queue is served at peak rate. And
2544 * this maximises throughput with sequential workloads.
2546 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
2548 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
2549 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
2553 * Update parameters related to throughput and responsiveness, as a
2554 * function of the estimated peak rate. See comments on
2555 * bfq_calc_max_budget(), and on the ref_wr_duration array.
2557 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
2559 if (bfqd->bfq_user_max_budget == 0) {
2560 bfqd->bfq_max_budget =
2561 bfq_calc_max_budget(bfqd);
2562 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
2566 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
2569 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
2570 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
2571 bfqd->peak_rate_samples = 1;
2572 bfqd->sequential_samples = 0;
2573 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
2575 } else /* no new rq dispatched, just reset the number of samples */
2576 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
2579 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
2580 bfqd->peak_rate_samples, bfqd->sequential_samples,
2581 bfqd->tot_sectors_dispatched);
2584 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
2586 u32 rate, weight, divisor;
2589 * For the convergence property to hold (see comments on
2590 * bfq_update_peak_rate()) and for the assessment to be
2591 * reliable, a minimum number of samples must be present, and
2592 * a minimum amount of time must have elapsed. If not so, do
2593 * not compute new rate. Just reset parameters, to get ready
2594 * for a new evaluation attempt.
2596 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
2597 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
2598 goto reset_computation;
2601 * If a new request completion has occurred after last
2602 * dispatch, then, to approximate the rate at which requests
2603 * have been served by the device, it is more precise to
2604 * extend the observation interval to the last completion.
2606 bfqd->delta_from_first =
2607 max_t(u64, bfqd->delta_from_first,
2608 bfqd->last_completion - bfqd->first_dispatch);
2611 * Rate computed in sects/usec, and not sects/nsec, for
2614 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
2615 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
2618 * Peak rate not updated if:
2619 * - the percentage of sequential dispatches is below 3/4 of the
2620 * total, and rate is below the current estimated peak rate
2621 * - rate is unreasonably high (> 20M sectors/sec)
2623 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
2624 rate <= bfqd->peak_rate) ||
2625 rate > 20<<BFQ_RATE_SHIFT)
2626 goto reset_computation;
2629 * We have to update the peak rate, at last! To this purpose,
2630 * we use a low-pass filter. We compute the smoothing constant
2631 * of the filter as a function of the 'weight' of the new
2634 * As can be seen in next formulas, we define this weight as a
2635 * quantity proportional to how sequential the workload is,
2636 * and to how long the observation time interval is.
2638 * The weight runs from 0 to 8. The maximum value of the
2639 * weight, 8, yields the minimum value for the smoothing
2640 * constant. At this minimum value for the smoothing constant,
2641 * the measured rate contributes for half of the next value of
2642 * the estimated peak rate.
2644 * So, the first step is to compute the weight as a function
2645 * of how sequential the workload is. Note that the weight
2646 * cannot reach 9, because bfqd->sequential_samples cannot
2647 * become equal to bfqd->peak_rate_samples, which, in its
2648 * turn, holds true because bfqd->sequential_samples is not
2649 * incremented for the first sample.
2651 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
2654 * Second step: further refine the weight as a function of the
2655 * duration of the observation interval.
2657 weight = min_t(u32, 8,
2658 div_u64(weight * bfqd->delta_from_first,
2659 BFQ_RATE_REF_INTERVAL));
2662 * Divisor ranging from 10, for minimum weight, to 2, for
2665 divisor = 10 - weight;
2668 * Finally, update peak rate:
2670 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
2672 bfqd->peak_rate *= divisor-1;
2673 bfqd->peak_rate /= divisor;
2674 rate /= divisor; /* smoothing constant alpha = 1/divisor */
2676 bfqd->peak_rate += rate;
2679 * For a very slow device, bfqd->peak_rate can reach 0 (see
2680 * the minimum representable values reported in the comments
2681 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
2682 * divisions by zero where bfqd->peak_rate is used as a
2685 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
2687 update_thr_responsiveness_params(bfqd);
2690 bfq_reset_rate_computation(bfqd, rq);
2694 * Update the read/write peak rate (the main quantity used for
2695 * auto-tuning, see update_thr_responsiveness_params()).
2697 * It is not trivial to estimate the peak rate (correctly): because of
2698 * the presence of sw and hw queues between the scheduler and the
2699 * device components that finally serve I/O requests, it is hard to
2700 * say exactly when a given dispatched request is served inside the
2701 * device, and for how long. As a consequence, it is hard to know
2702 * precisely at what rate a given set of requests is actually served
2705 * On the opposite end, the dispatch time of any request is trivially
2706 * available, and, from this piece of information, the "dispatch rate"
2707 * of requests can be immediately computed. So, the idea in the next
2708 * function is to use what is known, namely request dispatch times
2709 * (plus, when useful, request completion times), to estimate what is
2710 * unknown, namely in-device request service rate.
2712 * The main issue is that, because of the above facts, the rate at
2713 * which a certain set of requests is dispatched over a certain time
2714 * interval can vary greatly with respect to the rate at which the
2715 * same requests are then served. But, since the size of any
2716 * intermediate queue is limited, and the service scheme is lossless
2717 * (no request is silently dropped), the following obvious convergence
2718 * property holds: the number of requests dispatched MUST become
2719 * closer and closer to the number of requests completed as the
2720 * observation interval grows. This is the key property used in
2721 * the next function to estimate the peak service rate as a function
2722 * of the observed dispatch rate. The function assumes to be invoked
2723 * on every request dispatch.
2725 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
2727 u64 now_ns = ktime_get_ns();
2729 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
2730 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
2731 bfqd->peak_rate_samples);
2732 bfq_reset_rate_computation(bfqd, rq);
2733 goto update_last_values; /* will add one sample */
2737 * Device idle for very long: the observation interval lasting
2738 * up to this dispatch cannot be a valid observation interval
2739 * for computing a new peak rate (similarly to the late-
2740 * completion event in bfq_completed_request()). Go to
2741 * update_rate_and_reset to have the following three steps
2743 * - close the observation interval at the last (previous)
2744 * request dispatch or completion
2745 * - compute rate, if possible, for that observation interval
2746 * - start a new observation interval with this dispatch
2748 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
2749 bfqd->rq_in_driver == 0)
2750 goto update_rate_and_reset;
2752 /* Update sampling information */
2753 bfqd->peak_rate_samples++;
2755 if ((bfqd->rq_in_driver > 0 ||
2756 now_ns - bfqd->last_completion < BFQ_MIN_TT)
2757 && get_sdist(bfqd->last_position, rq) < BFQQ_SEEK_THR)
2758 bfqd->sequential_samples++;
2760 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
2762 /* Reset max observed rq size every 32 dispatches */
2763 if (likely(bfqd->peak_rate_samples % 32))
2764 bfqd->last_rq_max_size =
2765 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
2767 bfqd->last_rq_max_size = blk_rq_sectors(rq);
2769 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
2771 /* Target observation interval not yet reached, go on sampling */
2772 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
2773 goto update_last_values;
2775 update_rate_and_reset:
2776 bfq_update_rate_reset(bfqd, rq);
2778 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2779 bfqd->last_dispatch = now_ns;
2783 * Remove request from internal lists.
2785 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
2787 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2790 * For consistency, the next instruction should have been
2791 * executed after removing the request from the queue and
2792 * dispatching it. We execute instead this instruction before
2793 * bfq_remove_request() (and hence introduce a temporary
2794 * inconsistency), for efficiency. In fact, should this
2795 * dispatch occur for a non in-service bfqq, this anticipated
2796 * increment prevents two counters related to bfqq->dispatched
2797 * from risking to be, first, uselessly decremented, and then
2798 * incremented again when the (new) value of bfqq->dispatched
2799 * happens to be taken into account.
2802 bfq_update_peak_rate(q->elevator->elevator_data, rq);
2804 bfq_remove_request(q, rq);
2807 static void __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq)
2810 * If this bfqq is shared between multiple processes, check
2811 * to make sure that those processes are still issuing I/Os
2812 * within the mean seek distance. If not, it may be time to
2813 * break the queues apart again.
2815 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
2816 bfq_mark_bfqq_split_coop(bfqq);
2818 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2819 if (bfqq->dispatched == 0)
2821 * Overloading budget_timeout field to store
2822 * the time at which the queue remains with no
2823 * backlog and no outstanding request; used by
2824 * the weight-raising mechanism.
2826 bfqq->budget_timeout = jiffies;
2828 bfq_del_bfqq_busy(bfqd, bfqq, true);
2830 bfq_requeue_bfqq(bfqd, bfqq, true);
2832 * Resort priority tree of potential close cooperators.
2834 bfq_pos_tree_add_move(bfqd, bfqq);
2838 * All in-service entities must have been properly deactivated
2839 * or requeued before executing the next function, which
2840 * resets all in-service entites as no more in service.
2842 __bfq_bfqd_reset_in_service(bfqd);
2846 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
2847 * @bfqd: device data.
2848 * @bfqq: queue to update.
2849 * @reason: reason for expiration.
2851 * Handle the feedback on @bfqq budget at queue expiration.
2852 * See the body for detailed comments.
2854 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
2855 struct bfq_queue *bfqq,
2856 enum bfqq_expiration reason)
2858 struct request *next_rq;
2859 int budget, min_budget;
2861 min_budget = bfq_min_budget(bfqd);
2863 if (bfqq->wr_coeff == 1)
2864 budget = bfqq->max_budget;
2866 * Use a constant, low budget for weight-raised queues,
2867 * to help achieve a low latency. Keep it slightly higher
2868 * than the minimum possible budget, to cause a little
2869 * bit fewer expirations.
2871 budget = 2 * min_budget;
2873 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
2874 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
2875 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
2876 budget, bfq_min_budget(bfqd));
2877 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
2878 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
2880 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
2883 * Caveat: in all the following cases we trade latency
2886 case BFQQE_TOO_IDLE:
2888 * This is the only case where we may reduce
2889 * the budget: if there is no request of the
2890 * process still waiting for completion, then
2891 * we assume (tentatively) that the timer has
2892 * expired because the batch of requests of
2893 * the process could have been served with a
2894 * smaller budget. Hence, betting that
2895 * process will behave in the same way when it
2896 * becomes backlogged again, we reduce its
2897 * next budget. As long as we guess right,
2898 * this budget cut reduces the latency
2899 * experienced by the process.
2901 * However, if there are still outstanding
2902 * requests, then the process may have not yet
2903 * issued its next request just because it is
2904 * still waiting for the completion of some of
2905 * the still outstanding ones. So in this
2906 * subcase we do not reduce its budget, on the
2907 * contrary we increase it to possibly boost
2908 * the throughput, as discussed in the
2909 * comments to the BUDGET_TIMEOUT case.
2911 if (bfqq->dispatched > 0) /* still outstanding reqs */
2912 budget = min(budget * 2, bfqd->bfq_max_budget);
2914 if (budget > 5 * min_budget)
2915 budget -= 4 * min_budget;
2917 budget = min_budget;
2920 case BFQQE_BUDGET_TIMEOUT:
2922 * We double the budget here because it gives
2923 * the chance to boost the throughput if this
2924 * is not a seeky process (and has bumped into
2925 * this timeout because of, e.g., ZBR).
2927 budget = min(budget * 2, bfqd->bfq_max_budget);
2929 case BFQQE_BUDGET_EXHAUSTED:
2931 * The process still has backlog, and did not
2932 * let either the budget timeout or the disk
2933 * idling timeout expire. Hence it is not
2934 * seeky, has a short thinktime and may be
2935 * happy with a higher budget too. So
2936 * definitely increase the budget of this good
2937 * candidate to boost the disk throughput.
2939 budget = min(budget * 4, bfqd->bfq_max_budget);
2941 case BFQQE_NO_MORE_REQUESTS:
2943 * For queues that expire for this reason, it
2944 * is particularly important to keep the
2945 * budget close to the actual service they
2946 * need. Doing so reduces the timestamp
2947 * misalignment problem described in the
2948 * comments in the body of
2949 * __bfq_activate_entity. In fact, suppose
2950 * that a queue systematically expires for
2951 * BFQQE_NO_MORE_REQUESTS and presents a
2952 * new request in time to enjoy timestamp
2953 * back-shifting. The larger the budget of the
2954 * queue is with respect to the service the
2955 * queue actually requests in each service
2956 * slot, the more times the queue can be
2957 * reactivated with the same virtual finish
2958 * time. It follows that, even if this finish
2959 * time is pushed to the system virtual time
2960 * to reduce the consequent timestamp
2961 * misalignment, the queue unjustly enjoys for
2962 * many re-activations a lower finish time
2963 * than all newly activated queues.
2965 * The service needed by bfqq is measured
2966 * quite precisely by bfqq->entity.service.
2967 * Since bfqq does not enjoy device idling,
2968 * bfqq->entity.service is equal to the number
2969 * of sectors that the process associated with
2970 * bfqq requested to read/write before waiting
2971 * for request completions, or blocking for
2974 budget = max_t(int, bfqq->entity.service, min_budget);
2979 } else if (!bfq_bfqq_sync(bfqq)) {
2981 * Async queues get always the maximum possible
2982 * budget, as for them we do not care about latency
2983 * (in addition, their ability to dispatch is limited
2984 * by the charging factor).
2986 budget = bfqd->bfq_max_budget;
2989 bfqq->max_budget = budget;
2991 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
2992 !bfqd->bfq_user_max_budget)
2993 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
2996 * If there is still backlog, then assign a new budget, making
2997 * sure that it is large enough for the next request. Since
2998 * the finish time of bfqq must be kept in sync with the
2999 * budget, be sure to call __bfq_bfqq_expire() *after* this
3002 * If there is no backlog, then no need to update the budget;
3003 * it will be updated on the arrival of a new request.
3005 next_rq = bfqq->next_rq;
3007 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
3008 bfq_serv_to_charge(next_rq, bfqq));
3010 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
3011 next_rq ? blk_rq_sectors(next_rq) : 0,
3012 bfqq->entity.budget);
3016 * Return true if the process associated with bfqq is "slow". The slow
3017 * flag is used, in addition to the budget timeout, to reduce the
3018 * amount of service provided to seeky processes, and thus reduce
3019 * their chances to lower the throughput. More details in the comments
3020 * on the function bfq_bfqq_expire().
3022 * An important observation is in order: as discussed in the comments
3023 * on the function bfq_update_peak_rate(), with devices with internal
3024 * queues, it is hard if ever possible to know when and for how long
3025 * an I/O request is processed by the device (apart from the trivial
3026 * I/O pattern where a new request is dispatched only after the
3027 * previous one has been completed). This makes it hard to evaluate
3028 * the real rate at which the I/O requests of each bfq_queue are
3029 * served. In fact, for an I/O scheduler like BFQ, serving a
3030 * bfq_queue means just dispatching its requests during its service
3031 * slot (i.e., until the budget of the queue is exhausted, or the
3032 * queue remains idle, or, finally, a timeout fires). But, during the
3033 * service slot of a bfq_queue, around 100 ms at most, the device may
3034 * be even still processing requests of bfq_queues served in previous
3035 * service slots. On the opposite end, the requests of the in-service
3036 * bfq_queue may be completed after the service slot of the queue
3039 * Anyway, unless more sophisticated solutions are used
3040 * (where possible), the sum of the sizes of the requests dispatched
3041 * during the service slot of a bfq_queue is probably the only
3042 * approximation available for the service received by the bfq_queue
3043 * during its service slot. And this sum is the quantity used in this
3044 * function to evaluate the I/O speed of a process.
3046 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3047 bool compensate, enum bfqq_expiration reason,
3048 unsigned long *delta_ms)
3050 ktime_t delta_ktime;
3052 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
3054 if (!bfq_bfqq_sync(bfqq))
3058 delta_ktime = bfqd->last_idling_start;
3060 delta_ktime = ktime_get();
3061 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
3062 delta_usecs = ktime_to_us(delta_ktime);
3064 /* don't use too short time intervals */
3065 if (delta_usecs < 1000) {
3066 if (blk_queue_nonrot(bfqd->queue))
3068 * give same worst-case guarantees as idling
3071 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
3072 else /* charge at least one seek */
3073 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
3078 *delta_ms = delta_usecs / USEC_PER_MSEC;
3081 * Use only long (> 20ms) intervals to filter out excessive
3082 * spikes in service rate estimation.
3084 if (delta_usecs > 20000) {
3086 * Caveat for rotational devices: processes doing I/O
3087 * in the slower disk zones tend to be slow(er) even
3088 * if not seeky. In this respect, the estimated peak
3089 * rate is likely to be an average over the disk
3090 * surface. Accordingly, to not be too harsh with
3091 * unlucky processes, a process is deemed slow only if
3092 * its rate has been lower than half of the estimated
3095 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
3098 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
3104 * To be deemed as soft real-time, an application must meet two
3105 * requirements. First, the application must not require an average
3106 * bandwidth higher than the approximate bandwidth required to playback or
3107 * record a compressed high-definition video.
3108 * The next function is invoked on the completion of the last request of a
3109 * batch, to compute the next-start time instant, soft_rt_next_start, such
3110 * that, if the next request of the application does not arrive before
3111 * soft_rt_next_start, then the above requirement on the bandwidth is met.
3113 * The second requirement is that the request pattern of the application is
3114 * isochronous, i.e., that, after issuing a request or a batch of requests,
3115 * the application stops issuing new requests until all its pending requests
3116 * have been completed. After that, the application may issue a new batch,
3118 * For this reason the next function is invoked to compute
3119 * soft_rt_next_start only for applications that meet this requirement,
3120 * whereas soft_rt_next_start is set to infinity for applications that do
3123 * Unfortunately, even a greedy (i.e., I/O-bound) application may
3124 * happen to meet, occasionally or systematically, both the above
3125 * bandwidth and isochrony requirements. This may happen at least in
3126 * the following circumstances. First, if the CPU load is high. The
3127 * application may stop issuing requests while the CPUs are busy
3128 * serving other processes, then restart, then stop again for a while,
3129 * and so on. The other circumstances are related to the storage
3130 * device: the storage device is highly loaded or reaches a low-enough
3131 * throughput with the I/O of the application (e.g., because the I/O
3132 * is random and/or the device is slow). In all these cases, the
3133 * I/O of the application may be simply slowed down enough to meet
3134 * the bandwidth and isochrony requirements. To reduce the probability
3135 * that greedy applications are deemed as soft real-time in these
3136 * corner cases, a further rule is used in the computation of
3137 * soft_rt_next_start: the return value of this function is forced to
3138 * be higher than the maximum between the following two quantities.
3140 * (a) Current time plus: (1) the maximum time for which the arrival
3141 * of a request is waited for when a sync queue becomes idle,
3142 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
3143 * postpone for a moment the reason for adding a few extra
3144 * jiffies; we get back to it after next item (b). Lower-bounding
3145 * the return value of this function with the current time plus
3146 * bfqd->bfq_slice_idle tends to filter out greedy applications,
3147 * because the latter issue their next request as soon as possible
3148 * after the last one has been completed. In contrast, a soft
3149 * real-time application spends some time processing data, after a
3150 * batch of its requests has been completed.
3152 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
3153 * above, greedy applications may happen to meet both the
3154 * bandwidth and isochrony requirements under heavy CPU or
3155 * storage-device load. In more detail, in these scenarios, these
3156 * applications happen, only for limited time periods, to do I/O
3157 * slowly enough to meet all the requirements described so far,
3158 * including the filtering in above item (a). These slow-speed
3159 * time intervals are usually interspersed between other time
3160 * intervals during which these applications do I/O at a very high
3161 * speed. Fortunately, exactly because of the high speed of the
3162 * I/O in the high-speed intervals, the values returned by this
3163 * function happen to be so high, near the end of any such
3164 * high-speed interval, to be likely to fall *after* the end of
3165 * the low-speed time interval that follows. These high values are
3166 * stored in bfqq->soft_rt_next_start after each invocation of
3167 * this function. As a consequence, if the last value of
3168 * bfqq->soft_rt_next_start is constantly used to lower-bound the
3169 * next value that this function may return, then, from the very
3170 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
3171 * likely to be constantly kept so high that any I/O request
3172 * issued during the low-speed interval is considered as arriving
3173 * to soon for the application to be deemed as soft
3174 * real-time. Then, in the high-speed interval that follows, the
3175 * application will not be deemed as soft real-time, just because
3176 * it will do I/O at a high speed. And so on.
3178 * Getting back to the filtering in item (a), in the following two
3179 * cases this filtering might be easily passed by a greedy
3180 * application, if the reference quantity was just
3181 * bfqd->bfq_slice_idle:
3182 * 1) HZ is so low that the duration of a jiffy is comparable to or
3183 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
3184 * devices with HZ=100. The time granularity may be so coarse
3185 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
3186 * is rather lower than the exact value.
3187 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
3188 * for a while, then suddenly 'jump' by several units to recover the lost
3189 * increments. This seems to happen, e.g., inside virtual machines.
3190 * To address this issue, in the filtering in (a) we do not use as a
3191 * reference time interval just bfqd->bfq_slice_idle, but
3192 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
3193 * minimum number of jiffies for which the filter seems to be quite
3194 * precise also in embedded systems and KVM/QEMU virtual machines.
3196 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
3197 struct bfq_queue *bfqq)
3199 return max3(bfqq->soft_rt_next_start,
3200 bfqq->last_idle_bklogged +
3201 HZ * bfqq->service_from_backlogged /
3202 bfqd->bfq_wr_max_softrt_rate,
3203 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
3206 static bool bfq_bfqq_injectable(struct bfq_queue *bfqq)
3208 return BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3209 blk_queue_nonrot(bfqq->bfqd->queue) &&
3214 * bfq_bfqq_expire - expire a queue.
3215 * @bfqd: device owning the queue.
3216 * @bfqq: the queue to expire.
3217 * @compensate: if true, compensate for the time spent idling.
3218 * @reason: the reason causing the expiration.
3220 * If the process associated with bfqq does slow I/O (e.g., because it
3221 * issues random requests), we charge bfqq with the time it has been
3222 * in service instead of the service it has received (see
3223 * bfq_bfqq_charge_time for details on how this goal is achieved). As
3224 * a consequence, bfqq will typically get higher timestamps upon
3225 * reactivation, and hence it will be rescheduled as if it had
3226 * received more service than what it has actually received. In the
3227 * end, bfqq receives less service in proportion to how slowly its
3228 * associated process consumes its budgets (and hence how seriously it
3229 * tends to lower the throughput). In addition, this time-charging
3230 * strategy guarantees time fairness among slow processes. In
3231 * contrast, if the process associated with bfqq is not slow, we
3232 * charge bfqq exactly with the service it has received.
3234 * Charging time to the first type of queues and the exact service to
3235 * the other has the effect of using the WF2Q+ policy to schedule the
3236 * former on a timeslice basis, without violating service domain
3237 * guarantees among the latter.
3239 void bfq_bfqq_expire(struct bfq_data *bfqd,
3240 struct bfq_queue *bfqq,
3242 enum bfqq_expiration reason)
3245 unsigned long delta = 0;
3246 struct bfq_entity *entity = &bfqq->entity;
3250 * Check whether the process is slow (see bfq_bfqq_is_slow).
3252 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
3255 * As above explained, charge slow (typically seeky) and
3256 * timed-out queues with the time and not the service
3257 * received, to favor sequential workloads.
3259 * Processes doing I/O in the slower disk zones will tend to
3260 * be slow(er) even if not seeky. Therefore, since the
3261 * estimated peak rate is actually an average over the disk
3262 * surface, these processes may timeout just for bad luck. To
3263 * avoid punishing them, do not charge time to processes that
3264 * succeeded in consuming at least 2/3 of their budget. This
3265 * allows BFQ to preserve enough elasticity to still perform
3266 * bandwidth, and not time, distribution with little unlucky
3267 * or quasi-sequential processes.
3269 if (bfqq->wr_coeff == 1 &&
3271 (reason == BFQQE_BUDGET_TIMEOUT &&
3272 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
3273 bfq_bfqq_charge_time(bfqd, bfqq, delta);
3275 if (reason == BFQQE_TOO_IDLE &&
3276 entity->service <= 2 * entity->budget / 10)
3277 bfq_clear_bfqq_IO_bound(bfqq);
3279 if (bfqd->low_latency && bfqq->wr_coeff == 1)
3280 bfqq->last_wr_start_finish = jiffies;
3282 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
3283 RB_EMPTY_ROOT(&bfqq->sort_list)) {
3285 * If we get here, and there are no outstanding
3286 * requests, then the request pattern is isochronous
3287 * (see the comments on the function
3288 * bfq_bfqq_softrt_next_start()). Thus we can compute
3289 * soft_rt_next_start. And we do it, unless bfqq is in
3290 * interactive weight raising. We do not do it in the
3291 * latter subcase, for the following reason. bfqq may
3292 * be conveying the I/O needed to load a soft
3293 * real-time application. Such an application will
3294 * actually exhibit a soft real-time I/O pattern after
3295 * it finally starts doing its job. But, if
3296 * soft_rt_next_start is computed here for an
3297 * interactive bfqq, and bfqq had received a lot of
3298 * service before remaining with no outstanding
3299 * request (likely to happen on a fast device), then
3300 * soft_rt_next_start would be assigned such a high
3301 * value that, for a very long time, bfqq would be
3302 * prevented from being possibly considered as soft
3305 * If, instead, the queue still has outstanding
3306 * requests, then we have to wait for the completion
3307 * of all the outstanding requests to discover whether
3308 * the request pattern is actually isochronous.
3310 if (bfqq->dispatched == 0 &&
3311 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
3312 bfqq->soft_rt_next_start =
3313 bfq_bfqq_softrt_next_start(bfqd, bfqq);
3314 else if (bfqq->dispatched > 0) {
3316 * Schedule an update of soft_rt_next_start to when
3317 * the task may be discovered to be isochronous.
3319 bfq_mark_bfqq_softrt_update(bfqq);
3323 bfq_log_bfqq(bfqd, bfqq,
3324 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
3325 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
3328 * Increase, decrease or leave budget unchanged according to
3331 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
3333 __bfq_bfqq_expire(bfqd, bfqq);
3335 if (ref == 1) /* bfqq is gone, no more actions on it */
3338 bfqq->injected_service = 0;
3340 /* mark bfqq as waiting a request only if a bic still points to it */
3341 if (!bfq_bfqq_busy(bfqq) &&
3342 reason != BFQQE_BUDGET_TIMEOUT &&
3343 reason != BFQQE_BUDGET_EXHAUSTED) {
3344 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
3346 * Not setting service to 0, because, if the next rq
3347 * arrives in time, the queue will go on receiving
3348 * service with this same budget (as if it never expired)
3351 entity->service = 0;
3354 * Reset the received-service counter for every parent entity.
3355 * Differently from what happens with bfqq->entity.service,
3356 * the resetting of this counter never needs to be postponed
3357 * for parent entities. In fact, in case bfqq may have a
3358 * chance to go on being served using the last, partially
3359 * consumed budget, bfqq->entity.service needs to be kept,
3360 * because if bfqq then actually goes on being served using
3361 * the same budget, the last value of bfqq->entity.service is
3362 * needed to properly decrement bfqq->entity.budget by the
3363 * portion already consumed. In contrast, it is not necessary
3364 * to keep entity->service for parent entities too, because
3365 * the bubble up of the new value of bfqq->entity.budget will
3366 * make sure that the budgets of parent entities are correct,
3367 * even in case bfqq and thus parent entities go on receiving
3368 * service with the same budget.
3370 entity = entity->parent;
3371 for_each_entity(entity)
3372 entity->service = 0;
3376 * Budget timeout is not implemented through a dedicated timer, but
3377 * just checked on request arrivals and completions, as well as on
3378 * idle timer expirations.
3380 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
3382 return time_is_before_eq_jiffies(bfqq->budget_timeout);
3386 * If we expire a queue that is actively waiting (i.e., with the
3387 * device idled) for the arrival of a new request, then we may incur
3388 * the timestamp misalignment problem described in the body of the
3389 * function __bfq_activate_entity. Hence we return true only if this
3390 * condition does not hold, or if the queue is slow enough to deserve
3391 * only to be kicked off for preserving a high throughput.
3393 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
3395 bfq_log_bfqq(bfqq->bfqd, bfqq,
3396 "may_budget_timeout: wait_request %d left %d timeout %d",
3397 bfq_bfqq_wait_request(bfqq),
3398 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
3399 bfq_bfqq_budget_timeout(bfqq));
3401 return (!bfq_bfqq_wait_request(bfqq) ||
3402 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
3404 bfq_bfqq_budget_timeout(bfqq);
3408 * For a queue that becomes empty, device idling is allowed only if
3409 * this function returns true for the queue. As a consequence, since
3410 * device idling plays a critical role in both throughput boosting and
3411 * service guarantees, the return value of this function plays a
3412 * critical role in both these aspects as well.
3414 * In a nutshell, this function returns true only if idling is
3415 * beneficial for throughput or, even if detrimental for throughput,
3416 * idling is however necessary to preserve service guarantees (low
3417 * latency, desired throughput distribution, ...). In particular, on
3418 * NCQ-capable devices, this function tries to return false, so as to
3419 * help keep the drives' internal queues full, whenever this helps the
3420 * device boost the throughput without causing any service-guarantee
3423 * In more detail, the return value of this function is obtained by,
3424 * first, computing a number of boolean variables that take into
3425 * account throughput and service-guarantee issues, and, then,
3426 * combining these variables in a logical expression. Most of the
3427 * issues taken into account are not trivial. We discuss these issues
3428 * individually while introducing the variables.
3430 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
3432 struct bfq_data *bfqd = bfqq->bfqd;
3433 bool rot_without_queueing =
3434 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
3435 bfqq_sequential_and_IO_bound,
3436 idling_boosts_thr, idling_boosts_thr_without_issues,
3437 idling_needed_for_service_guarantees,
3438 asymmetric_scenario;
3440 if (bfqd->strict_guarantees)
3444 * Idling is performed only if slice_idle > 0. In addition, we
3447 * (b) bfqq is in the idle io prio class: in this case we do
3448 * not idle because we want to minimize the bandwidth that
3449 * queues in this class can steal to higher-priority queues
3451 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
3452 bfq_class_idle(bfqq))
3455 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
3456 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
3459 * The next variable takes into account the cases where idling
3460 * boosts the throughput.
3462 * The value of the variable is computed considering, first, that
3463 * idling is virtually always beneficial for the throughput if:
3464 * (a) the device is not NCQ-capable and rotational, or
3465 * (b) regardless of the presence of NCQ, the device is rotational and
3466 * the request pattern for bfqq is I/O-bound and sequential, or
3467 * (c) regardless of whether it is rotational, the device is
3468 * not NCQ-capable and the request pattern for bfqq is
3469 * I/O-bound and sequential.
3471 * Secondly, and in contrast to the above item (b), idling an
3472 * NCQ-capable flash-based device would not boost the
3473 * throughput even with sequential I/O; rather it would lower
3474 * the throughput in proportion to how fast the device
3475 * is. Accordingly, the next variable is true if any of the
3476 * above conditions (a), (b) or (c) is true, and, in
3477 * particular, happens to be false if bfqd is an NCQ-capable
3478 * flash-based device.
3480 idling_boosts_thr = rot_without_queueing ||
3481 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
3482 bfqq_sequential_and_IO_bound);
3485 * The value of the next variable,
3486 * idling_boosts_thr_without_issues, is equal to that of
3487 * idling_boosts_thr, unless a special case holds. In this
3488 * special case, described below, idling may cause problems to
3489 * weight-raised queues.
3491 * When the request pool is saturated (e.g., in the presence
3492 * of write hogs), if the processes associated with
3493 * non-weight-raised queues ask for requests at a lower rate,
3494 * then processes associated with weight-raised queues have a
3495 * higher probability to get a request from the pool
3496 * immediately (or at least soon) when they need one. Thus
3497 * they have a higher probability to actually get a fraction
3498 * of the device throughput proportional to their high
3499 * weight. This is especially true with NCQ-capable drives,
3500 * which enqueue several requests in advance, and further
3501 * reorder internally-queued requests.
3503 * For this reason, we force to false the value of
3504 * idling_boosts_thr_without_issues if there are weight-raised
3505 * busy queues. In this case, and if bfqq is not weight-raised,
3506 * this guarantees that the device is not idled for bfqq (if,
3507 * instead, bfqq is weight-raised, then idling will be
3508 * guaranteed by another variable, see below). Combined with
3509 * the timestamping rules of BFQ (see [1] for details), this
3510 * behavior causes bfqq, and hence any sync non-weight-raised
3511 * queue, to get a lower number of requests served, and thus
3512 * to ask for a lower number of requests from the request
3513 * pool, before the busy weight-raised queues get served
3514 * again. This often mitigates starvation problems in the
3515 * presence of heavy write workloads and NCQ, thereby
3516 * guaranteeing a higher application and system responsiveness
3517 * in these hostile scenarios.
3519 idling_boosts_thr_without_issues = idling_boosts_thr &&
3520 bfqd->wr_busy_queues == 0;
3523 * There is then a case where idling must be performed not
3524 * for throughput concerns, but to preserve service
3527 * To introduce this case, we can note that allowing the drive
3528 * to enqueue more than one request at a time, and hence
3529 * delegating de facto final scheduling decisions to the
3530 * drive's internal scheduler, entails loss of control on the
3531 * actual request service order. In particular, the critical
3532 * situation is when requests from different processes happen
3533 * to be present, at the same time, in the internal queue(s)
3534 * of the drive. In such a situation, the drive, by deciding
3535 * the service order of the internally-queued requests, does
3536 * determine also the actual throughput distribution among
3537 * these processes. But the drive typically has no notion or
3538 * concern about per-process throughput distribution, and
3539 * makes its decisions only on a per-request basis. Therefore,
3540 * the service distribution enforced by the drive's internal
3541 * scheduler is likely to coincide with the desired
3542 * device-throughput distribution only in a completely
3543 * symmetric scenario where:
3544 * (i) each of these processes must get the same throughput as
3546 * (ii) the I/O of each process has the same properties, in
3547 * terms of locality (sequential or random), direction
3548 * (reads or writes), request sizes, greediness
3549 * (from I/O-bound to sporadic), and so on.
3550 * In fact, in such a scenario, the drive tends to treat
3551 * the requests of each of these processes in about the same
3552 * way as the requests of the others, and thus to provide
3553 * each of these processes with about the same throughput
3554 * (which is exactly the desired throughput distribution). In
3555 * contrast, in any asymmetric scenario, device idling is
3556 * certainly needed to guarantee that bfqq receives its
3557 * assigned fraction of the device throughput (see [1] for
3559 * The problem is that idling may significantly reduce
3560 * throughput with certain combinations of types of I/O and
3561 * devices. An important example is sync random I/O, on flash
3562 * storage with command queueing. So, unless bfqq falls in the
3563 * above cases where idling also boosts throughput, it would
3564 * be important to check conditions (i) and (ii) accurately,
3565 * so as to avoid idling when not strictly needed for service
3568 * Unfortunately, it is extremely difficult to thoroughly
3569 * check condition (ii). And, in case there are active groups,
3570 * it becomes very difficult to check condition (i) too. In
3571 * fact, if there are active groups, then, for condition (i)
3572 * to become false, it is enough that an active group contains
3573 * more active processes or sub-groups than some other active
3574 * group. More precisely, for condition (i) to hold because of
3575 * such a group, it is not even necessary that the group is
3576 * (still) active: it is sufficient that, even if the group
3577 * has become inactive, some of its descendant processes still
3578 * have some request already dispatched but still waiting for
3579 * completion. In fact, requests have still to be guaranteed
3580 * their share of the throughput even after being
3581 * dispatched. In this respect, it is easy to show that, if a
3582 * group frequently becomes inactive while still having
3583 * in-flight requests, and if, when this happens, the group is
3584 * not considered in the calculation of whether the scenario
3585 * is asymmetric, then the group may fail to be guaranteed its
3586 * fair share of the throughput (basically because idling may
3587 * not be performed for the descendant processes of the group,
3588 * but it had to be). We address this issue with the
3589 * following bi-modal behavior, implemented in the function
3590 * bfq_symmetric_scenario().
3592 * If there are groups with requests waiting for completion
3593 * (as commented above, some of these groups may even be
3594 * already inactive), then the scenario is tagged as
3595 * asymmetric, conservatively, without checking any of the
3596 * conditions (i) and (ii). So the device is idled for bfqq.
3597 * This behavior matches also the fact that groups are created
3598 * exactly if controlling I/O is a primary concern (to
3599 * preserve bandwidth and latency guarantees).
3601 * On the opposite end, if there are no groups with requests
3602 * waiting for completion, then only condition (i) is actually
3603 * controlled, i.e., provided that condition (i) holds, idling
3604 * is not performed, regardless of whether condition (ii)
3605 * holds. In other words, only if condition (i) does not hold,
3606 * then idling is allowed, and the device tends to be
3607 * prevented from queueing many requests, possibly of several
3608 * processes. Since there are no groups with requests waiting
3609 * for completion, then, to control condition (i) it is enough
3610 * to check just whether all the queues with requests waiting
3611 * for completion also have the same weight.
3613 * Not checking condition (ii) evidently exposes bfqq to the
3614 * risk of getting less throughput than its fair share.
3615 * However, for queues with the same weight, a further
3616 * mechanism, preemption, mitigates or even eliminates this
3617 * problem. And it does so without consequences on overall
3618 * throughput. This mechanism and its benefits are explained
3619 * in the next three paragraphs.
3621 * Even if a queue, say Q, is expired when it remains idle, Q
3622 * can still preempt the new in-service queue if the next
3623 * request of Q arrives soon (see the comments on
3624 * bfq_bfqq_update_budg_for_activation). If all queues and
3625 * groups have the same weight, this form of preemption,
3626 * combined with the hole-recovery heuristic described in the
3627 * comments on function bfq_bfqq_update_budg_for_activation,
3628 * are enough to preserve a correct bandwidth distribution in
3629 * the mid term, even without idling. In fact, even if not
3630 * idling allows the internal queues of the device to contain
3631 * many requests, and thus to reorder requests, we can rather
3632 * safely assume that the internal scheduler still preserves a
3633 * minimum of mid-term fairness.
3635 * More precisely, this preemption-based, idleless approach
3636 * provides fairness in terms of IOPS, and not sectors per
3637 * second. This can be seen with a simple example. Suppose
3638 * that there are two queues with the same weight, but that
3639 * the first queue receives requests of 8 sectors, while the
3640 * second queue receives requests of 1024 sectors. In
3641 * addition, suppose that each of the two queues contains at
3642 * most one request at a time, which implies that each queue
3643 * always remains idle after it is served. Finally, after
3644 * remaining idle, each queue receives very quickly a new
3645 * request. It follows that the two queues are served
3646 * alternatively, preempting each other if needed. This
3647 * implies that, although both queues have the same weight,
3648 * the queue with large requests receives a service that is
3649 * 1024/8 times as high as the service received by the other
3652 * The motivation for using preemption instead of idling (for
3653 * queues with the same weight) is that, by not idling,
3654 * service guarantees are preserved (completely or at least in
3655 * part) without minimally sacrificing throughput. And, if
3656 * there is no active group, then the primary expectation for
3657 * this device is probably a high throughput.
3659 * We are now left only with explaining the additional
3660 * compound condition that is checked below for deciding
3661 * whether the scenario is asymmetric. To explain this
3662 * compound condition, we need to add that the function
3663 * bfq_symmetric_scenario checks the weights of only
3664 * non-weight-raised queues, for efficiency reasons (see
3665 * comments on bfq_weights_tree_add()). Then the fact that
3666 * bfqq is weight-raised is checked explicitly here. More
3667 * precisely, the compound condition below takes into account
3668 * also the fact that, even if bfqq is being weight-raised,
3669 * the scenario is still symmetric if all queues with requests
3670 * waiting for completion happen to be
3671 * weight-raised. Actually, we should be even more precise
3672 * here, and differentiate between interactive weight raising
3673 * and soft real-time weight raising.
3675 * As a side note, it is worth considering that the above
3676 * device-idling countermeasures may however fail in the
3677 * following unlucky scenario: if idling is (correctly)
3678 * disabled in a time period during which all symmetry
3679 * sub-conditions hold, and hence the device is allowed to
3680 * enqueue many requests, but at some later point in time some
3681 * sub-condition stops to hold, then it may become impossible
3682 * to let requests be served in the desired order until all
3683 * the requests already queued in the device have been served.
3685 asymmetric_scenario = (bfqq->wr_coeff > 1 &&
3686 bfqd->wr_busy_queues <
3687 bfq_tot_busy_queues(bfqd)) ||
3688 !bfq_symmetric_scenario(bfqd);
3691 * Finally, there is a case where maximizing throughput is the
3692 * best choice even if it may cause unfairness toward
3693 * bfqq. Such a case is when bfqq became active in a burst of
3694 * queue activations. Queues that became active during a large
3695 * burst benefit only from throughput, as discussed in the
3696 * comments on bfq_handle_burst. Thus, if bfqq became active
3697 * in a burst and not idling the device maximizes throughput,
3698 * then the device must no be idled, because not idling the
3699 * device provides bfqq and all other queues in the burst with
3700 * maximum benefit. Combining this and the above case, we can
3701 * now establish when idling is actually needed to preserve
3702 * service guarantees.
3704 idling_needed_for_service_guarantees =
3705 asymmetric_scenario && !bfq_bfqq_in_large_burst(bfqq);
3708 * We have now all the components we need to compute the
3709 * return value of the function, which is true only if idling
3710 * either boosts the throughput (without issues), or is
3711 * necessary to preserve service guarantees.
3713 return idling_boosts_thr_without_issues ||
3714 idling_needed_for_service_guarantees;
3718 * If the in-service queue is empty but the function bfq_better_to_idle
3719 * returns true, then:
3720 * 1) the queue must remain in service and cannot be expired, and
3721 * 2) the device must be idled to wait for the possible arrival of a new
3722 * request for the queue.
3723 * See the comments on the function bfq_better_to_idle for the reasons
3724 * why performing device idling is the best choice to boost the throughput
3725 * and preserve service guarantees when bfq_better_to_idle itself
3728 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
3730 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
3733 static struct bfq_queue *bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
3735 struct bfq_queue *bfqq;
3738 * A linear search; but, with a high probability, very few
3739 * steps are needed to find a candidate queue, i.e., a queue
3740 * with enough budget left for its next request. In fact:
3741 * - BFQ dynamically updates the budget of every queue so as
3742 * to accommodate the expected backlog of the queue;
3743 * - if a queue gets all its requests dispatched as injected
3744 * service, then the queue is removed from the active list
3745 * (and re-added only if it gets new requests, but with
3746 * enough budget for its new backlog).
3748 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
3749 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
3750 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
3751 bfq_bfqq_budget_left(bfqq))
3758 * Select a queue for service. If we have a current queue in service,
3759 * check whether to continue servicing it, or retrieve and set a new one.
3761 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
3763 struct bfq_queue *bfqq;
3764 struct request *next_rq;
3765 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
3767 bfqq = bfqd->in_service_queue;
3771 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
3774 * Do not expire bfqq for budget timeout if bfqq may be about
3775 * to enjoy device idling. The reason why, in this case, we
3776 * prevent bfqq from expiring is the same as in the comments
3777 * on the case where bfq_bfqq_must_idle() returns true, in
3778 * bfq_completed_request().
3780 if (bfq_may_expire_for_budg_timeout(bfqq) &&
3781 !bfq_bfqq_must_idle(bfqq))
3786 * This loop is rarely executed more than once. Even when it
3787 * happens, it is much more convenient to re-execute this loop
3788 * than to return NULL and trigger a new dispatch to get a
3791 next_rq = bfqq->next_rq;
3793 * If bfqq has requests queued and it has enough budget left to
3794 * serve them, keep the queue, otherwise expire it.
3797 if (bfq_serv_to_charge(next_rq, bfqq) >
3798 bfq_bfqq_budget_left(bfqq)) {
3800 * Expire the queue for budget exhaustion,
3801 * which makes sure that the next budget is
3802 * enough to serve the next request, even if
3803 * it comes from the fifo expired path.
3805 reason = BFQQE_BUDGET_EXHAUSTED;
3809 * The idle timer may be pending because we may
3810 * not disable disk idling even when a new request
3813 if (bfq_bfqq_wait_request(bfqq)) {
3815 * If we get here: 1) at least a new request
3816 * has arrived but we have not disabled the
3817 * timer because the request was too small,
3818 * 2) then the block layer has unplugged
3819 * the device, causing the dispatch to be
3822 * Since the device is unplugged, now the
3823 * requests are probably large enough to
3824 * provide a reasonable throughput.
3825 * So we disable idling.
3827 bfq_clear_bfqq_wait_request(bfqq);
3828 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
3835 * No requests pending. However, if the in-service queue is idling
3836 * for a new request, or has requests waiting for a completion and
3837 * may idle after their completion, then keep it anyway.
3839 * Yet, to boost throughput, inject service from other queues if
3842 if (bfq_bfqq_wait_request(bfqq) ||
3843 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
3844 if (bfq_bfqq_injectable(bfqq) &&
3845 bfqq->injected_service * bfqq->inject_coeff <
3846 bfqq->entity.service * 10)
3847 bfqq = bfq_choose_bfqq_for_injection(bfqd);
3854 reason = BFQQE_NO_MORE_REQUESTS;
3856 bfq_bfqq_expire(bfqd, bfqq, false, reason);
3858 bfqq = bfq_set_in_service_queue(bfqd);
3860 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
3865 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
3867 bfq_log(bfqd, "select_queue: no queue returned");
3872 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3874 struct bfq_entity *entity = &bfqq->entity;
3876 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
3877 bfq_log_bfqq(bfqd, bfqq,
3878 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
3879 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
3880 jiffies_to_msecs(bfqq->wr_cur_max_time),
3882 bfqq->entity.weight, bfqq->entity.orig_weight);
3884 if (entity->prio_changed)
3885 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
3888 * If the queue was activated in a burst, or too much
3889 * time has elapsed from the beginning of this
3890 * weight-raising period, then end weight raising.
3892 if (bfq_bfqq_in_large_burst(bfqq))
3893 bfq_bfqq_end_wr(bfqq);
3894 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
3895 bfqq->wr_cur_max_time)) {
3896 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
3897 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
3898 bfq_wr_duration(bfqd)))
3899 bfq_bfqq_end_wr(bfqq);
3901 switch_back_to_interactive_wr(bfqq, bfqd);
3902 bfqq->entity.prio_changed = 1;
3905 if (bfqq->wr_coeff > 1 &&
3906 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
3907 bfqq->service_from_wr > max_service_from_wr) {
3908 /* see comments on max_service_from_wr */
3909 bfq_bfqq_end_wr(bfqq);
3913 * To improve latency (for this or other queues), immediately
3914 * update weight both if it must be raised and if it must be
3915 * lowered. Since, entity may be on some active tree here, and
3916 * might have a pending change of its ioprio class, invoke
3917 * next function with the last parameter unset (see the
3918 * comments on the function).
3920 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
3921 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
3926 * Dispatch next request from bfqq.
3928 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
3929 struct bfq_queue *bfqq)
3931 struct request *rq = bfqq->next_rq;
3932 unsigned long service_to_charge;
3934 service_to_charge = bfq_serv_to_charge(rq, bfqq);
3936 bfq_bfqq_served(bfqq, service_to_charge);
3938 bfq_dispatch_remove(bfqd->queue, rq);
3940 if (bfqq != bfqd->in_service_queue) {
3941 if (likely(bfqd->in_service_queue))
3942 bfqd->in_service_queue->injected_service +=
3943 bfq_serv_to_charge(rq, bfqq);
3949 * If weight raising has to terminate for bfqq, then next
3950 * function causes an immediate update of bfqq's weight,
3951 * without waiting for next activation. As a consequence, on
3952 * expiration, bfqq will be timestamped as if has never been
3953 * weight-raised during this service slot, even if it has
3954 * received part or even most of the service as a
3955 * weight-raised queue. This inflates bfqq's timestamps, which
3956 * is beneficial, as bfqq is then more willing to leave the
3957 * device immediately to possible other weight-raised queues.
3959 bfq_update_wr_data(bfqd, bfqq);
3962 * Expire bfqq, pretending that its budget expired, if bfqq
3963 * belongs to CLASS_IDLE and other queues are waiting for
3966 if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
3969 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
3975 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
3977 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3980 * Avoiding lock: a race on bfqd->busy_queues should cause at
3981 * most a call to dispatch for nothing
3983 return !list_empty_careful(&bfqd->dispatch) ||
3984 bfq_tot_busy_queues(bfqd) > 0;
3987 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3989 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3990 struct request *rq = NULL;
3991 struct bfq_queue *bfqq = NULL;
3993 if (!list_empty(&bfqd->dispatch)) {
3994 rq = list_first_entry(&bfqd->dispatch, struct request,
3996 list_del_init(&rq->queuelist);
4002 * Increment counters here, because this
4003 * dispatch does not follow the standard
4004 * dispatch flow (where counters are
4009 goto inc_in_driver_start_rq;
4013 * We exploit the bfq_finish_requeue_request hook to
4014 * decrement rq_in_driver, but
4015 * bfq_finish_requeue_request will not be invoked on
4016 * this request. So, to avoid unbalance, just start
4017 * this request, without incrementing rq_in_driver. As
4018 * a negative consequence, rq_in_driver is deceptively
4019 * lower than it should be while this request is in
4020 * service. This may cause bfq_schedule_dispatch to be
4021 * invoked uselessly.
4023 * As for implementing an exact solution, the
4024 * bfq_finish_requeue_request hook, if defined, is
4025 * probably invoked also on this request. So, by
4026 * exploiting this hook, we could 1) increment
4027 * rq_in_driver here, and 2) decrement it in
4028 * bfq_finish_requeue_request. Such a solution would
4029 * let the value of the counter be always accurate,
4030 * but it would entail using an extra interface
4031 * function. This cost seems higher than the benefit,
4032 * being the frequency of non-elevator-private
4033 * requests very low.
4038 bfq_log(bfqd, "dispatch requests: %d busy queues",
4039 bfq_tot_busy_queues(bfqd));
4041 if (bfq_tot_busy_queues(bfqd) == 0)
4045 * Force device to serve one request at a time if
4046 * strict_guarantees is true. Forcing this service scheme is
4047 * currently the ONLY way to guarantee that the request
4048 * service order enforced by the scheduler is respected by a
4049 * queueing device. Otherwise the device is free even to make
4050 * some unlucky request wait for as long as the device
4053 * Of course, serving one request at at time may cause loss of
4056 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
4059 bfqq = bfq_select_queue(bfqd);
4063 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
4066 inc_in_driver_start_rq:
4067 bfqd->rq_in_driver++;
4069 rq->rq_flags |= RQF_STARTED;
4075 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
4076 static void bfq_update_dispatch_stats(struct request_queue *q,
4078 struct bfq_queue *in_serv_queue,
4079 bool idle_timer_disabled)
4081 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
4083 if (!idle_timer_disabled && !bfqq)
4087 * rq and bfqq are guaranteed to exist until this function
4088 * ends, for the following reasons. First, rq can be
4089 * dispatched to the device, and then can be completed and
4090 * freed, only after this function ends. Second, rq cannot be
4091 * merged (and thus freed because of a merge) any longer,
4092 * because it has already started. Thus rq cannot be freed
4093 * before this function ends, and, since rq has a reference to
4094 * bfqq, the same guarantee holds for bfqq too.
4096 * In addition, the following queue lock guarantees that
4097 * bfqq_group(bfqq) exists as well.
4099 spin_lock_irq(&q->queue_lock);
4100 if (idle_timer_disabled)
4102 * Since the idle timer has been disabled,
4103 * in_serv_queue contained some request when
4104 * __bfq_dispatch_request was invoked above, which
4105 * implies that rq was picked exactly from
4106 * in_serv_queue. Thus in_serv_queue == bfqq, and is
4107 * therefore guaranteed to exist because of the above
4110 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
4112 struct bfq_group *bfqg = bfqq_group(bfqq);
4114 bfqg_stats_update_avg_queue_size(bfqg);
4115 bfqg_stats_set_start_empty_time(bfqg);
4116 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
4118 spin_unlock_irq(&q->queue_lock);
4121 static inline void bfq_update_dispatch_stats(struct request_queue *q,
4123 struct bfq_queue *in_serv_queue,
4124 bool idle_timer_disabled) {}
4127 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
4129 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4131 struct bfq_queue *in_serv_queue;
4132 bool waiting_rq, idle_timer_disabled;
4134 spin_lock_irq(&bfqd->lock);
4136 in_serv_queue = bfqd->in_service_queue;
4137 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
4139 rq = __bfq_dispatch_request(hctx);
4141 idle_timer_disabled =
4142 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
4144 spin_unlock_irq(&bfqd->lock);
4146 bfq_update_dispatch_stats(hctx->queue, rq, in_serv_queue,
4147 idle_timer_disabled);
4153 * Task holds one reference to the queue, dropped when task exits. Each rq
4154 * in-flight on this queue also holds a reference, dropped when rq is freed.
4156 * Scheduler lock must be held here. Recall not to use bfqq after calling
4157 * this function on it.
4159 void bfq_put_queue(struct bfq_queue *bfqq)
4161 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4162 struct bfq_group *bfqg = bfqq_group(bfqq);
4166 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
4173 if (!hlist_unhashed(&bfqq->burst_list_node)) {
4174 hlist_del_init(&bfqq->burst_list_node);
4176 * Decrement also burst size after the removal, if the
4177 * process associated with bfqq is exiting, and thus
4178 * does not contribute to the burst any longer. This
4179 * decrement helps filter out false positives of large
4180 * bursts, when some short-lived process (often due to
4181 * the execution of commands by some service) happens
4182 * to start and exit while a complex application is
4183 * starting, and thus spawning several processes that
4184 * do I/O (and that *must not* be treated as a large
4185 * burst, see comments on bfq_handle_burst).
4187 * In particular, the decrement is performed only if:
4188 * 1) bfqq is not a merged queue, because, if it is,
4189 * then this free of bfqq is not triggered by the exit
4190 * of the process bfqq is associated with, but exactly
4191 * by the fact that bfqq has just been merged.
4192 * 2) burst_size is greater than 0, to handle
4193 * unbalanced decrements. Unbalanced decrements may
4194 * happen in te following case: bfqq is inserted into
4195 * the current burst list--without incrementing
4196 * bust_size--because of a split, but the current
4197 * burst list is not the burst list bfqq belonged to
4198 * (see comments on the case of a split in
4201 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
4202 bfqq->bfqd->burst_size--;
4205 kmem_cache_free(bfq_pool, bfqq);
4206 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4207 bfqg_and_blkg_put(bfqg);
4211 static void bfq_put_cooperator(struct bfq_queue *bfqq)
4213 struct bfq_queue *__bfqq, *next;
4216 * If this queue was scheduled to merge with another queue, be
4217 * sure to drop the reference taken on that queue (and others in
4218 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
4220 __bfqq = bfqq->new_bfqq;
4224 next = __bfqq->new_bfqq;
4225 bfq_put_queue(__bfqq);
4230 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4232 if (bfqq == bfqd->in_service_queue) {
4233 __bfq_bfqq_expire(bfqd, bfqq);
4234 bfq_schedule_dispatch(bfqd);
4237 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
4239 bfq_put_cooperator(bfqq);
4241 bfq_put_queue(bfqq); /* release process reference */
4244 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
4246 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4247 struct bfq_data *bfqd;
4250 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
4253 unsigned long flags;
4255 spin_lock_irqsave(&bfqd->lock, flags);
4256 bfq_exit_bfqq(bfqd, bfqq);
4257 bic_set_bfqq(bic, NULL, is_sync);
4258 spin_unlock_irqrestore(&bfqd->lock, flags);
4262 static void bfq_exit_icq(struct io_cq *icq)
4264 struct bfq_io_cq *bic = icq_to_bic(icq);
4266 bfq_exit_icq_bfqq(bic, true);
4267 bfq_exit_icq_bfqq(bic, false);
4271 * Update the entity prio values; note that the new values will not
4272 * be used until the next (re)activation.
4275 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
4277 struct task_struct *tsk = current;
4279 struct bfq_data *bfqd = bfqq->bfqd;
4284 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4285 switch (ioprio_class) {
4287 dev_err(bfqq->bfqd->queue->backing_dev_info->dev,
4288 "bfq: bad prio class %d\n", ioprio_class);
4290 case IOPRIO_CLASS_NONE:
4292 * No prio set, inherit CPU scheduling settings.
4294 bfqq->new_ioprio = task_nice_ioprio(tsk);
4295 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
4297 case IOPRIO_CLASS_RT:
4298 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4299 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
4301 case IOPRIO_CLASS_BE:
4302 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4303 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
4305 case IOPRIO_CLASS_IDLE:
4306 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
4307 bfqq->new_ioprio = 7;
4311 if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
4312 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
4314 bfqq->new_ioprio = IOPRIO_BE_NR;
4317 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
4318 bfqq->entity.prio_changed = 1;
4321 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4322 struct bio *bio, bool is_sync,
4323 struct bfq_io_cq *bic);
4325 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
4327 struct bfq_data *bfqd = bic_to_bfqd(bic);
4328 struct bfq_queue *bfqq;
4329 int ioprio = bic->icq.ioc->ioprio;
4332 * This condition may trigger on a newly created bic, be sure to
4333 * drop the lock before returning.
4335 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
4338 bic->ioprio = ioprio;
4340 bfqq = bic_to_bfqq(bic, false);
4342 /* release process reference on this queue */
4343 bfq_put_queue(bfqq);
4344 bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
4345 bic_set_bfqq(bic, bfqq, false);
4348 bfqq = bic_to_bfqq(bic, true);
4350 bfq_set_next_ioprio_data(bfqq, bic);
4353 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4354 struct bfq_io_cq *bic, pid_t pid, int is_sync)
4356 RB_CLEAR_NODE(&bfqq->entity.rb_node);
4357 INIT_LIST_HEAD(&bfqq->fifo);
4358 INIT_HLIST_NODE(&bfqq->burst_list_node);
4364 bfq_set_next_ioprio_data(bfqq, bic);
4368 * No need to mark as has_short_ttime if in
4369 * idle_class, because no device idling is performed
4370 * for queues in idle class
4372 if (!bfq_class_idle(bfqq))
4373 /* tentatively mark as has_short_ttime */
4374 bfq_mark_bfqq_has_short_ttime(bfqq);
4375 bfq_mark_bfqq_sync(bfqq);
4376 bfq_mark_bfqq_just_created(bfqq);
4378 * Aggressively inject a lot of service: up to 90%.
4379 * This coefficient remains constant during bfqq life,
4380 * but this behavior might be changed, after enough
4381 * testing and tuning.
4383 bfqq->inject_coeff = 1;
4385 bfq_clear_bfqq_sync(bfqq);
4387 /* set end request to minus infinity from now */
4388 bfqq->ttime.last_end_request = ktime_get_ns() + 1;
4390 bfq_mark_bfqq_IO_bound(bfqq);
4394 /* Tentative initial value to trade off between thr and lat */
4395 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
4396 bfqq->budget_timeout = bfq_smallest_from_now();
4399 bfqq->last_wr_start_finish = jiffies;
4400 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
4401 bfqq->split_time = bfq_smallest_from_now();
4404 * To not forget the possibly high bandwidth consumed by a
4405 * process/queue in the recent past,
4406 * bfq_bfqq_softrt_next_start() returns a value at least equal
4407 * to the current value of bfqq->soft_rt_next_start (see
4408 * comments on bfq_bfqq_softrt_next_start). Set
4409 * soft_rt_next_start to now, to mean that bfqq has consumed
4410 * no bandwidth so far.
4412 bfqq->soft_rt_next_start = jiffies;
4414 /* first request is almost certainly seeky */
4415 bfqq->seek_history = 1;
4418 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
4419 struct bfq_group *bfqg,
4420 int ioprio_class, int ioprio)
4422 switch (ioprio_class) {
4423 case IOPRIO_CLASS_RT:
4424 return &bfqg->async_bfqq[0][ioprio];
4425 case IOPRIO_CLASS_NONE:
4426 ioprio = IOPRIO_NORM;
4428 case IOPRIO_CLASS_BE:
4429 return &bfqg->async_bfqq[1][ioprio];
4430 case IOPRIO_CLASS_IDLE:
4431 return &bfqg->async_idle_bfqq;
4437 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4438 struct bio *bio, bool is_sync,
4439 struct bfq_io_cq *bic)
4441 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4442 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4443 struct bfq_queue **async_bfqq = NULL;
4444 struct bfq_queue *bfqq;
4445 struct bfq_group *bfqg;
4449 bfqg = bfq_find_set_group(bfqd, __bio_blkcg(bio));
4451 bfqq = &bfqd->oom_bfqq;
4456 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
4463 bfqq = kmem_cache_alloc_node(bfq_pool,
4464 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
4468 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
4470 bfq_init_entity(&bfqq->entity, bfqg);
4471 bfq_log_bfqq(bfqd, bfqq, "allocated");
4473 bfqq = &bfqd->oom_bfqq;
4474 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
4479 * Pin the queue now that it's allocated, scheduler exit will
4484 * Extra group reference, w.r.t. sync
4485 * queue. This extra reference is removed
4486 * only if bfqq->bfqg disappears, to
4487 * guarantee that this queue is not freed
4488 * until its group goes away.
4490 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
4496 bfqq->ref++; /* get a process reference to this queue */
4497 bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
4502 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
4503 struct bfq_queue *bfqq)
4505 struct bfq_ttime *ttime = &bfqq->ttime;
4506 u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
4508 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
4510 ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
4511 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
4512 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
4513 ttime->ttime_samples);
4517 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4520 bfqq->seek_history <<= 1;
4521 bfqq->seek_history |=
4522 get_sdist(bfqq->last_request_pos, rq) > BFQQ_SEEK_THR &&
4523 (!blk_queue_nonrot(bfqd->queue) ||
4524 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT);
4527 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
4528 struct bfq_queue *bfqq,
4529 struct bfq_io_cq *bic)
4531 bool has_short_ttime = true;
4534 * No need to update has_short_ttime if bfqq is async or in
4535 * idle io prio class, or if bfq_slice_idle is zero, because
4536 * no device idling is performed for bfqq in this case.
4538 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
4539 bfqd->bfq_slice_idle == 0)
4542 /* Idle window just restored, statistics are meaningless. */
4543 if (time_is_after_eq_jiffies(bfqq->split_time +
4544 bfqd->bfq_wr_min_idle_time))
4547 /* Think time is infinite if no process is linked to
4548 * bfqq. Otherwise check average think time to
4549 * decide whether to mark as has_short_ttime
4551 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
4552 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
4553 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle))
4554 has_short_ttime = false;
4556 bfq_log_bfqq(bfqd, bfqq, "update_has_short_ttime: has_short_ttime %d",
4559 if (has_short_ttime)
4560 bfq_mark_bfqq_has_short_ttime(bfqq);
4562 bfq_clear_bfqq_has_short_ttime(bfqq);
4566 * Called when a new fs request (rq) is added to bfqq. Check if there's
4567 * something we should do about it.
4569 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4572 struct bfq_io_cq *bic = RQ_BIC(rq);
4574 if (rq->cmd_flags & REQ_META)
4575 bfqq->meta_pending++;
4577 bfq_update_io_thinktime(bfqd, bfqq);
4578 bfq_update_has_short_ttime(bfqd, bfqq, bic);
4579 bfq_update_io_seektime(bfqd, bfqq, rq);
4581 bfq_log_bfqq(bfqd, bfqq,
4582 "rq_enqueued: has_short_ttime=%d (seeky %d)",
4583 bfq_bfqq_has_short_ttime(bfqq), BFQQ_SEEKY(bfqq));
4585 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4587 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
4588 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
4589 blk_rq_sectors(rq) < 32;
4590 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
4593 * There is just this request queued: if the request
4594 * is small and the queue is not to be expired, then
4597 * In this way, if the device is being idled to wait
4598 * for a new request from the in-service queue, we
4599 * avoid unplugging the device and committing the
4600 * device to serve just a small request. On the
4601 * contrary, we wait for the block layer to decide
4602 * when to unplug the device: hopefully, new requests
4603 * will be merged to this one quickly, then the device
4604 * will be unplugged and larger requests will be
4607 if (small_req && !budget_timeout)
4611 * A large enough request arrived, or the queue is to
4612 * be expired: in both cases disk idling is to be
4613 * stopped, so clear wait_request flag and reset
4616 bfq_clear_bfqq_wait_request(bfqq);
4617 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4620 * The queue is not empty, because a new request just
4621 * arrived. Hence we can safely expire the queue, in
4622 * case of budget timeout, without risking that the
4623 * timestamps of the queue are not updated correctly.
4624 * See [1] for more details.
4627 bfq_bfqq_expire(bfqd, bfqq, false,
4628 BFQQE_BUDGET_TIMEOUT);
4632 /* returns true if it causes the idle timer to be disabled */
4633 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
4635 struct bfq_queue *bfqq = RQ_BFQQ(rq),
4636 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
4637 bool waiting, idle_timer_disabled = false;
4641 * Release the request's reference to the old bfqq
4642 * and make sure one is taken to the shared queue.
4644 new_bfqq->allocated++;
4648 * If the bic associated with the process
4649 * issuing this request still points to bfqq
4650 * (and thus has not been already redirected
4651 * to new_bfqq or even some other bfq_queue),
4652 * then complete the merge and redirect it to
4655 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
4656 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
4659 bfq_clear_bfqq_just_created(bfqq);
4661 * rq is about to be enqueued into new_bfqq,
4662 * release rq reference on bfqq
4664 bfq_put_queue(bfqq);
4665 rq->elv.priv[1] = new_bfqq;
4669 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
4670 bfq_add_request(rq);
4671 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
4673 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
4674 list_add_tail(&rq->queuelist, &bfqq->fifo);
4676 bfq_rq_enqueued(bfqd, bfqq, rq);
4678 return idle_timer_disabled;
4681 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
4682 static void bfq_update_insert_stats(struct request_queue *q,
4683 struct bfq_queue *bfqq,
4684 bool idle_timer_disabled,
4685 unsigned int cmd_flags)
4691 * bfqq still exists, because it can disappear only after
4692 * either it is merged with another queue, or the process it
4693 * is associated with exits. But both actions must be taken by
4694 * the same process currently executing this flow of
4697 * In addition, the following queue lock guarantees that
4698 * bfqq_group(bfqq) exists as well.
4700 spin_lock_irq(&q->queue_lock);
4701 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
4702 if (idle_timer_disabled)
4703 bfqg_stats_update_idle_time(bfqq_group(bfqq));
4704 spin_unlock_irq(&q->queue_lock);
4707 static inline void bfq_update_insert_stats(struct request_queue *q,
4708 struct bfq_queue *bfqq,
4709 bool idle_timer_disabled,
4710 unsigned int cmd_flags) {}
4713 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
4716 struct request_queue *q = hctx->queue;
4717 struct bfq_data *bfqd = q->elevator->elevator_data;
4718 struct bfq_queue *bfqq;
4719 bool idle_timer_disabled = false;
4720 unsigned int cmd_flags;
4722 spin_lock_irq(&bfqd->lock);
4723 if (blk_mq_sched_try_insert_merge(q, rq)) {
4724 spin_unlock_irq(&bfqd->lock);
4728 spin_unlock_irq(&bfqd->lock);
4730 blk_mq_sched_request_inserted(rq);
4732 spin_lock_irq(&bfqd->lock);
4733 bfqq = bfq_init_rq(rq);
4734 if (at_head || blk_rq_is_passthrough(rq)) {
4736 list_add(&rq->queuelist, &bfqd->dispatch);
4738 list_add_tail(&rq->queuelist, &bfqd->dispatch);
4739 } else { /* bfqq is assumed to be non null here */
4740 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
4742 * Update bfqq, because, if a queue merge has occurred
4743 * in __bfq_insert_request, then rq has been
4744 * redirected into a new queue.
4748 if (rq_mergeable(rq)) {
4749 elv_rqhash_add(q, rq);
4756 * Cache cmd_flags before releasing scheduler lock, because rq
4757 * may disappear afterwards (for example, because of a request
4760 cmd_flags = rq->cmd_flags;
4762 spin_unlock_irq(&bfqd->lock);
4764 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
4768 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
4769 struct list_head *list, bool at_head)
4771 while (!list_empty(list)) {
4774 rq = list_first_entry(list, struct request, queuelist);
4775 list_del_init(&rq->queuelist);
4776 bfq_insert_request(hctx, rq, at_head);
4780 static void bfq_update_hw_tag(struct bfq_data *bfqd)
4782 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
4783 bfqd->rq_in_driver);
4785 if (bfqd->hw_tag == 1)
4789 * This sample is valid if the number of outstanding requests
4790 * is large enough to allow a queueing behavior. Note that the
4791 * sum is not exact, as it's not taking into account deactivated
4794 if (bfqd->rq_in_driver + bfqd->queued < BFQ_HW_QUEUE_THRESHOLD)
4797 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
4800 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
4801 bfqd->max_rq_in_driver = 0;
4802 bfqd->hw_tag_samples = 0;
4805 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
4810 bfq_update_hw_tag(bfqd);
4812 bfqd->rq_in_driver--;
4815 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
4817 * Set budget_timeout (which we overload to store the
4818 * time at which the queue remains with no backlog and
4819 * no outstanding request; used by the weight-raising
4822 bfqq->budget_timeout = jiffies;
4824 bfq_weights_tree_remove(bfqd, bfqq);
4827 now_ns = ktime_get_ns();
4829 bfqq->ttime.last_end_request = now_ns;
4832 * Using us instead of ns, to get a reasonable precision in
4833 * computing rate in next check.
4835 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
4838 * If the request took rather long to complete, and, according
4839 * to the maximum request size recorded, this completion latency
4840 * implies that the request was certainly served at a very low
4841 * rate (less than 1M sectors/sec), then the whole observation
4842 * interval that lasts up to this time instant cannot be a
4843 * valid time interval for computing a new peak rate. Invoke
4844 * bfq_update_rate_reset to have the following three steps
4846 * - close the observation interval at the last (previous)
4847 * request dispatch or completion
4848 * - compute rate, if possible, for that observation interval
4849 * - reset to zero samples, which will trigger a proper
4850 * re-initialization of the observation interval on next
4853 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
4854 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
4855 1UL<<(BFQ_RATE_SHIFT - 10))
4856 bfq_update_rate_reset(bfqd, NULL);
4857 bfqd->last_completion = now_ns;
4860 * If we are waiting to discover whether the request pattern
4861 * of the task associated with the queue is actually
4862 * isochronous, and both requisites for this condition to hold
4863 * are now satisfied, then compute soft_rt_next_start (see the
4864 * comments on the function bfq_bfqq_softrt_next_start()). We
4865 * do not compute soft_rt_next_start if bfqq is in interactive
4866 * weight raising (see the comments in bfq_bfqq_expire() for
4867 * an explanation). We schedule this delayed update when bfqq
4868 * expires, if it still has in-flight requests.
4870 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
4871 RB_EMPTY_ROOT(&bfqq->sort_list) &&
4872 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
4873 bfqq->soft_rt_next_start =
4874 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4877 * If this is the in-service queue, check if it needs to be expired,
4878 * or if we want to idle in case it has no pending requests.
4880 if (bfqd->in_service_queue == bfqq) {
4881 if (bfq_bfqq_must_idle(bfqq)) {
4882 if (bfqq->dispatched == 0)
4883 bfq_arm_slice_timer(bfqd);
4885 * If we get here, we do not expire bfqq, even
4886 * if bfqq was in budget timeout or had no
4887 * more requests (as controlled in the next
4888 * conditional instructions). The reason for
4889 * not expiring bfqq is as follows.
4891 * Here bfqq->dispatched > 0 holds, but
4892 * bfq_bfqq_must_idle() returned true. This
4893 * implies that, even if no request arrives
4894 * for bfqq before bfqq->dispatched reaches 0,
4895 * bfqq will, however, not be expired on the
4896 * completion event that causes bfqq->dispatch
4897 * to reach zero. In contrast, on this event,
4898 * bfqq will start enjoying device idling
4899 * (I/O-dispatch plugging).
4901 * But, if we expired bfqq here, bfqq would
4902 * not have the chance to enjoy device idling
4903 * when bfqq->dispatched finally reaches
4904 * zero. This would expose bfqq to violation
4905 * of its reserved service guarantees.
4908 } else if (bfq_may_expire_for_budg_timeout(bfqq))
4909 bfq_bfqq_expire(bfqd, bfqq, false,
4910 BFQQE_BUDGET_TIMEOUT);
4911 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
4912 (bfqq->dispatched == 0 ||
4913 !bfq_better_to_idle(bfqq)))
4914 bfq_bfqq_expire(bfqd, bfqq, false,
4915 BFQQE_NO_MORE_REQUESTS);
4918 if (!bfqd->rq_in_driver)
4919 bfq_schedule_dispatch(bfqd);
4922 static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq)
4926 bfq_put_queue(bfqq);
4930 * Handle either a requeue or a finish for rq. The things to do are
4931 * the same in both cases: all references to rq are to be dropped. In
4932 * particular, rq is considered completed from the point of view of
4935 static void bfq_finish_requeue_request(struct request *rq)
4937 struct bfq_queue *bfqq = RQ_BFQQ(rq);
4938 struct bfq_data *bfqd;
4941 * Requeue and finish hooks are invoked in blk-mq without
4942 * checking whether the involved request is actually still
4943 * referenced in the scheduler. To handle this fact, the
4944 * following two checks make this function exit in case of
4945 * spurious invocations, for which there is nothing to do.
4947 * First, check whether rq has nothing to do with an elevator.
4949 if (unlikely(!(rq->rq_flags & RQF_ELVPRIV)))
4953 * rq either is not associated with any icq, or is an already
4954 * requeued request that has not (yet) been re-inserted into
4957 if (!rq->elv.icq || !bfqq)
4962 if (rq->rq_flags & RQF_STARTED)
4963 bfqg_stats_update_completion(bfqq_group(bfqq),
4965 rq->io_start_time_ns,
4968 if (likely(rq->rq_flags & RQF_STARTED)) {
4969 unsigned long flags;
4971 spin_lock_irqsave(&bfqd->lock, flags);
4973 bfq_completed_request(bfqq, bfqd);
4974 bfq_finish_requeue_request_body(bfqq);
4976 spin_unlock_irqrestore(&bfqd->lock, flags);
4979 * Request rq may be still/already in the scheduler,
4980 * in which case we need to remove it (this should
4981 * never happen in case of requeue). And we cannot
4982 * defer such a check and removal, to avoid
4983 * inconsistencies in the time interval from the end
4984 * of this function to the start of the deferred work.
4985 * This situation seems to occur only in process
4986 * context, as a consequence of a merge. In the
4987 * current version of the code, this implies that the
4991 if (!RB_EMPTY_NODE(&rq->rb_node)) {
4992 bfq_remove_request(rq->q, rq);
4993 bfqg_stats_update_io_remove(bfqq_group(bfqq),
4996 bfq_finish_requeue_request_body(bfqq);
5000 * Reset private fields. In case of a requeue, this allows
5001 * this function to correctly do nothing if it is spuriously
5002 * invoked again on this same request (see the check at the
5003 * beginning of the function). Probably, a better general
5004 * design would be to prevent blk-mq from invoking the requeue
5005 * or finish hooks of an elevator, for a request that is not
5006 * referred by that elevator.
5008 * Resetting the following fields would break the
5009 * request-insertion logic if rq is re-inserted into a bfq
5010 * internal queue, without a re-preparation. Here we assume
5011 * that re-insertions of requeued requests, without
5012 * re-preparation, can happen only for pass_through or at_head
5013 * requests (which are not re-inserted into bfq internal
5016 rq->elv.priv[0] = NULL;
5017 rq->elv.priv[1] = NULL;
5021 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
5022 * was the last process referring to that bfqq.
5024 static struct bfq_queue *
5025 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
5027 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
5029 if (bfqq_process_refs(bfqq) == 1) {
5030 bfqq->pid = current->pid;
5031 bfq_clear_bfqq_coop(bfqq);
5032 bfq_clear_bfqq_split_coop(bfqq);
5036 bic_set_bfqq(bic, NULL, 1);
5038 bfq_put_cooperator(bfqq);
5040 bfq_put_queue(bfqq);
5044 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
5045 struct bfq_io_cq *bic,
5047 bool split, bool is_sync,
5050 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
5052 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
5059 bfq_put_queue(bfqq);
5060 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
5062 bic_set_bfqq(bic, bfqq, is_sync);
5063 if (split && is_sync) {
5064 if ((bic->was_in_burst_list && bfqd->large_burst) ||
5065 bic->saved_in_large_burst)
5066 bfq_mark_bfqq_in_large_burst(bfqq);
5068 bfq_clear_bfqq_in_large_burst(bfqq);
5069 if (bic->was_in_burst_list)
5071 * If bfqq was in the current
5072 * burst list before being
5073 * merged, then we have to add
5074 * it back. And we do not need
5075 * to increase burst_size, as
5076 * we did not decrement
5077 * burst_size when we removed
5078 * bfqq from the burst list as
5079 * a consequence of a merge
5081 * bfq_put_queue). In this
5082 * respect, it would be rather
5083 * costly to know whether the
5084 * current burst list is still
5085 * the same burst list from
5086 * which bfqq was removed on
5087 * the merge. To avoid this
5088 * cost, if bfqq was in a
5089 * burst list, then we add
5090 * bfqq to the current burst
5091 * list without any further
5092 * check. This can cause
5093 * inappropriate insertions,
5094 * but rarely enough to not
5095 * harm the detection of large
5096 * bursts significantly.
5098 hlist_add_head(&bfqq->burst_list_node,
5101 bfqq->split_time = jiffies;
5108 * Only reset private fields. The actual request preparation will be
5109 * performed by bfq_init_rq, when rq is either inserted or merged. See
5110 * comments on bfq_init_rq for the reason behind this delayed
5113 static void bfq_prepare_request(struct request *rq, struct bio *bio)
5116 * Regardless of whether we have an icq attached, we have to
5117 * clear the scheduler pointers, as they might point to
5118 * previously allocated bic/bfqq structs.
5120 rq->elv.priv[0] = rq->elv.priv[1] = NULL;
5124 * If needed, init rq, allocate bfq data structures associated with
5125 * rq, and increment reference counters in the destination bfq_queue
5126 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
5127 * not associated with any bfq_queue.
5129 * This function is invoked by the functions that perform rq insertion
5130 * or merging. One may have expected the above preparation operations
5131 * to be performed in bfq_prepare_request, and not delayed to when rq
5132 * is inserted or merged. The rationale behind this delayed
5133 * preparation is that, after the prepare_request hook is invoked for
5134 * rq, rq may still be transformed into a request with no icq, i.e., a
5135 * request not associated with any queue. No bfq hook is invoked to
5136 * signal this tranformation. As a consequence, should these
5137 * preparation operations be performed when the prepare_request hook
5138 * is invoked, and should rq be transformed one moment later, bfq
5139 * would end up in an inconsistent state, because it would have
5140 * incremented some queue counters for an rq destined to
5141 * transformation, without any chance to correctly lower these
5142 * counters back. In contrast, no transformation can still happen for
5143 * rq after rq has been inserted or merged. So, it is safe to execute
5144 * these preparation operations when rq is finally inserted or merged.
5146 static struct bfq_queue *bfq_init_rq(struct request *rq)
5148 struct request_queue *q = rq->q;
5149 struct bio *bio = rq->bio;
5150 struct bfq_data *bfqd = q->elevator->elevator_data;
5151 struct bfq_io_cq *bic;
5152 const int is_sync = rq_is_sync(rq);
5153 struct bfq_queue *bfqq;
5154 bool new_queue = false;
5155 bool bfqq_already_existing = false, split = false;
5157 if (unlikely(!rq->elv.icq))
5161 * Assuming that elv.priv[1] is set only if everything is set
5162 * for this rq. This holds true, because this function is
5163 * invoked only for insertion or merging, and, after such
5164 * events, a request cannot be manipulated any longer before
5165 * being removed from bfq.
5167 if (rq->elv.priv[1])
5168 return rq->elv.priv[1];
5170 bic = icq_to_bic(rq->elv.icq);
5172 bfq_check_ioprio_change(bic, bio);
5174 bfq_bic_update_cgroup(bic, bio);
5176 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
5179 if (likely(!new_queue)) {
5180 /* If the queue was seeky for too long, break it apart. */
5181 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
5182 bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
5184 /* Update bic before losing reference to bfqq */
5185 if (bfq_bfqq_in_large_burst(bfqq))
5186 bic->saved_in_large_burst = true;
5188 bfqq = bfq_split_bfqq(bic, bfqq);
5192 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
5196 bfqq_already_existing = true;
5202 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
5203 rq, bfqq, bfqq->ref);
5205 rq->elv.priv[0] = bic;
5206 rq->elv.priv[1] = bfqq;
5209 * If a bfq_queue has only one process reference, it is owned
5210 * by only this bic: we can then set bfqq->bic = bic. in
5211 * addition, if the queue has also just been split, we have to
5214 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
5218 * The queue has just been split from a shared
5219 * queue: restore the idle window and the
5220 * possible weight raising period.
5222 bfq_bfqq_resume_state(bfqq, bfqd, bic,
5223 bfqq_already_existing);
5227 if (unlikely(bfq_bfqq_just_created(bfqq)))
5228 bfq_handle_burst(bfqd, bfqq);
5233 static void bfq_idle_slice_timer_body(struct bfq_queue *bfqq)
5235 struct bfq_data *bfqd = bfqq->bfqd;
5236 enum bfqq_expiration reason;
5237 unsigned long flags;
5239 spin_lock_irqsave(&bfqd->lock, flags);
5240 bfq_clear_bfqq_wait_request(bfqq);
5242 if (bfqq != bfqd->in_service_queue) {
5243 spin_unlock_irqrestore(&bfqd->lock, flags);
5247 if (bfq_bfqq_budget_timeout(bfqq))
5249 * Also here the queue can be safely expired
5250 * for budget timeout without wasting
5253 reason = BFQQE_BUDGET_TIMEOUT;
5254 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
5256 * The queue may not be empty upon timer expiration,
5257 * because we may not disable the timer when the
5258 * first request of the in-service queue arrives
5259 * during disk idling.
5261 reason = BFQQE_TOO_IDLE;
5263 goto schedule_dispatch;
5265 bfq_bfqq_expire(bfqd, bfqq, true, reason);
5268 spin_unlock_irqrestore(&bfqd->lock, flags);
5269 bfq_schedule_dispatch(bfqd);
5273 * Handler of the expiration of the timer running if the in-service queue
5274 * is idling inside its time slice.
5276 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
5278 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
5280 struct bfq_queue *bfqq = bfqd->in_service_queue;
5283 * Theoretical race here: the in-service queue can be NULL or
5284 * different from the queue that was idling if a new request
5285 * arrives for the current queue and there is a full dispatch
5286 * cycle that changes the in-service queue. This can hardly
5287 * happen, but in the worst case we just expire a queue too
5291 bfq_idle_slice_timer_body(bfqq);
5293 return HRTIMER_NORESTART;
5296 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
5297 struct bfq_queue **bfqq_ptr)
5299 struct bfq_queue *bfqq = *bfqq_ptr;
5301 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
5303 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
5305 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
5307 bfq_put_queue(bfqq);
5313 * Release all the bfqg references to its async queues. If we are
5314 * deallocating the group these queues may still contain requests, so
5315 * we reparent them to the root cgroup (i.e., the only one that will
5316 * exist for sure until all the requests on a device are gone).
5318 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
5322 for (i = 0; i < 2; i++)
5323 for (j = 0; j < IOPRIO_BE_NR; j++)
5324 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
5326 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
5330 * See the comments on bfq_limit_depth for the purpose of
5331 * the depths set in the function. Return minimum shallow depth we'll use.
5333 static unsigned int bfq_update_depths(struct bfq_data *bfqd,
5334 struct sbitmap_queue *bt)
5336 unsigned int i, j, min_shallow = UINT_MAX;
5339 * In-word depths if no bfq_queue is being weight-raised:
5340 * leaving 25% of tags only for sync reads.
5342 * In next formulas, right-shift the value
5343 * (1U<<bt->sb.shift), instead of computing directly
5344 * (1U<<(bt->sb.shift - something)), to be robust against
5345 * any possible value of bt->sb.shift, without having to
5346 * limit 'something'.
5348 /* no more than 50% of tags for async I/O */
5349 bfqd->word_depths[0][0] = max((1U << bt->sb.shift) >> 1, 1U);
5351 * no more than 75% of tags for sync writes (25% extra tags
5352 * w.r.t. async I/O, to prevent async I/O from starving sync
5355 bfqd->word_depths[0][1] = max(((1U << bt->sb.shift) * 3) >> 2, 1U);
5358 * In-word depths in case some bfq_queue is being weight-
5359 * raised: leaving ~63% of tags for sync reads. This is the
5360 * highest percentage for which, in our tests, application
5361 * start-up times didn't suffer from any regression due to tag
5364 /* no more than ~18% of tags for async I/O */
5365 bfqd->word_depths[1][0] = max(((1U << bt->sb.shift) * 3) >> 4, 1U);
5366 /* no more than ~37% of tags for sync writes (~20% extra tags) */
5367 bfqd->word_depths[1][1] = max(((1U << bt->sb.shift) * 6) >> 4, 1U);
5369 for (i = 0; i < 2; i++)
5370 for (j = 0; j < 2; j++)
5371 min_shallow = min(min_shallow, bfqd->word_depths[i][j]);
5376 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
5378 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5379 struct blk_mq_tags *tags = hctx->sched_tags;
5380 unsigned int min_shallow;
5382 min_shallow = bfq_update_depths(bfqd, &tags->bitmap_tags);
5383 sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, min_shallow);
5387 static void bfq_exit_queue(struct elevator_queue *e)
5389 struct bfq_data *bfqd = e->elevator_data;
5390 struct bfq_queue *bfqq, *n;
5392 hrtimer_cancel(&bfqd->idle_slice_timer);
5394 spin_lock_irq(&bfqd->lock);
5395 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
5396 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
5397 spin_unlock_irq(&bfqd->lock);
5399 hrtimer_cancel(&bfqd->idle_slice_timer);
5401 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5402 /* release oom-queue reference to root group */
5403 bfqg_and_blkg_put(bfqd->root_group);
5405 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
5407 spin_lock_irq(&bfqd->lock);
5408 bfq_put_async_queues(bfqd, bfqd->root_group);
5409 kfree(bfqd->root_group);
5410 spin_unlock_irq(&bfqd->lock);
5416 static void bfq_init_root_group(struct bfq_group *root_group,
5417 struct bfq_data *bfqd)
5421 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5422 root_group->entity.parent = NULL;
5423 root_group->my_entity = NULL;
5424 root_group->bfqd = bfqd;
5426 root_group->rq_pos_tree = RB_ROOT;
5427 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
5428 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
5429 root_group->sched_data.bfq_class_idle_last_service = jiffies;
5432 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
5434 struct bfq_data *bfqd;
5435 struct elevator_queue *eq;
5437 eq = elevator_alloc(q, e);
5441 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
5443 kobject_put(&eq->kobj);
5446 eq->elevator_data = bfqd;
5448 spin_lock_irq(&q->queue_lock);
5450 spin_unlock_irq(&q->queue_lock);
5453 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
5454 * Grab a permanent reference to it, so that the normal code flow
5455 * will not attempt to free it.
5457 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
5458 bfqd->oom_bfqq.ref++;
5459 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
5460 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
5461 bfqd->oom_bfqq.entity.new_weight =
5462 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
5464 /* oom_bfqq does not participate to bursts */
5465 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
5468 * Trigger weight initialization, according to ioprio, at the
5469 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
5470 * class won't be changed any more.
5472 bfqd->oom_bfqq.entity.prio_changed = 1;
5476 INIT_LIST_HEAD(&bfqd->dispatch);
5478 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
5480 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
5482 bfqd->queue_weights_tree = RB_ROOT;
5483 bfqd->num_groups_with_pending_reqs = 0;
5485 INIT_LIST_HEAD(&bfqd->active_list);
5486 INIT_LIST_HEAD(&bfqd->idle_list);
5487 INIT_HLIST_HEAD(&bfqd->burst_list);
5491 bfqd->bfq_max_budget = bfq_default_max_budget;
5493 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
5494 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
5495 bfqd->bfq_back_max = bfq_back_max;
5496 bfqd->bfq_back_penalty = bfq_back_penalty;
5497 bfqd->bfq_slice_idle = bfq_slice_idle;
5498 bfqd->bfq_timeout = bfq_timeout;
5500 bfqd->bfq_requests_within_timer = 120;
5502 bfqd->bfq_large_burst_thresh = 8;
5503 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
5505 bfqd->low_latency = true;
5508 * Trade-off between responsiveness and fairness.
5510 bfqd->bfq_wr_coeff = 30;
5511 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
5512 bfqd->bfq_wr_max_time = 0;
5513 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
5514 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
5515 bfqd->bfq_wr_max_softrt_rate = 7000; /*
5516 * Approximate rate required
5517 * to playback or record a
5518 * high-definition compressed
5521 bfqd->wr_busy_queues = 0;
5524 * Begin by assuming, optimistically, that the device peak
5525 * rate is equal to 2/3 of the highest reference rate.
5527 bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
5528 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
5529 bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
5531 spin_lock_init(&bfqd->lock);
5534 * The invocation of the next bfq_create_group_hierarchy
5535 * function is the head of a chain of function calls
5536 * (bfq_create_group_hierarchy->blkcg_activate_policy->
5537 * blk_mq_freeze_queue) that may lead to the invocation of the
5538 * has_work hook function. For this reason,
5539 * bfq_create_group_hierarchy is invoked only after all
5540 * scheduler data has been initialized, apart from the fields
5541 * that can be initialized only after invoking
5542 * bfq_create_group_hierarchy. This, in particular, enables
5543 * has_work to correctly return false. Of course, to avoid
5544 * other inconsistencies, the blk-mq stack must then refrain
5545 * from invoking further scheduler hooks before this init
5546 * function is finished.
5548 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
5549 if (!bfqd->root_group)
5551 bfq_init_root_group(bfqd->root_group, bfqd);
5552 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
5554 wbt_disable_default(q);
5559 kobject_put(&eq->kobj);
5563 static void bfq_slab_kill(void)
5565 kmem_cache_destroy(bfq_pool);
5568 static int __init bfq_slab_setup(void)
5570 bfq_pool = KMEM_CACHE(bfq_queue, 0);
5576 static ssize_t bfq_var_show(unsigned int var, char *page)
5578 return sprintf(page, "%u\n", var);
5581 static int bfq_var_store(unsigned long *var, const char *page)
5583 unsigned long new_val;
5584 int ret = kstrtoul(page, 10, &new_val);
5592 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
5593 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
5595 struct bfq_data *bfqd = e->elevator_data; \
5596 u64 __data = __VAR; \
5598 __data = jiffies_to_msecs(__data); \
5599 else if (__CONV == 2) \
5600 __data = div_u64(__data, NSEC_PER_MSEC); \
5601 return bfq_var_show(__data, (page)); \
5603 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
5604 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
5605 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
5606 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
5607 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
5608 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
5609 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
5610 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
5611 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
5612 #undef SHOW_FUNCTION
5614 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
5615 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
5617 struct bfq_data *bfqd = e->elevator_data; \
5618 u64 __data = __VAR; \
5619 __data = div_u64(__data, NSEC_PER_USEC); \
5620 return bfq_var_show(__data, (page)); \
5622 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
5623 #undef USEC_SHOW_FUNCTION
5625 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
5627 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
5629 struct bfq_data *bfqd = e->elevator_data; \
5630 unsigned long __data, __min = (MIN), __max = (MAX); \
5633 ret = bfq_var_store(&__data, (page)); \
5636 if (__data < __min) \
5638 else if (__data > __max) \
5641 *(__PTR) = msecs_to_jiffies(__data); \
5642 else if (__CONV == 2) \
5643 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
5645 *(__PTR) = __data; \
5648 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
5650 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
5652 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
5653 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
5655 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
5656 #undef STORE_FUNCTION
5658 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
5659 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
5661 struct bfq_data *bfqd = e->elevator_data; \
5662 unsigned long __data, __min = (MIN), __max = (MAX); \
5665 ret = bfq_var_store(&__data, (page)); \
5668 if (__data < __min) \
5670 else if (__data > __max) \
5672 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
5675 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
5677 #undef USEC_STORE_FUNCTION
5679 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
5680 const char *page, size_t count)
5682 struct bfq_data *bfqd = e->elevator_data;
5683 unsigned long __data;
5686 ret = bfq_var_store(&__data, (page));
5691 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5693 if (__data > INT_MAX)
5695 bfqd->bfq_max_budget = __data;
5698 bfqd->bfq_user_max_budget = __data;
5704 * Leaving this name to preserve name compatibility with cfq
5705 * parameters, but this timeout is used for both sync and async.
5707 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
5708 const char *page, size_t count)
5710 struct bfq_data *bfqd = e->elevator_data;
5711 unsigned long __data;
5714 ret = bfq_var_store(&__data, (page));
5720 else if (__data > INT_MAX)
5723 bfqd->bfq_timeout = msecs_to_jiffies(__data);
5724 if (bfqd->bfq_user_max_budget == 0)
5725 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5730 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
5731 const char *page, size_t count)
5733 struct bfq_data *bfqd = e->elevator_data;
5734 unsigned long __data;
5737 ret = bfq_var_store(&__data, (page));
5743 if (!bfqd->strict_guarantees && __data == 1
5744 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
5745 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
5747 bfqd->strict_guarantees = __data;
5752 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
5753 const char *page, size_t count)
5755 struct bfq_data *bfqd = e->elevator_data;
5756 unsigned long __data;
5759 ret = bfq_var_store(&__data, (page));
5765 if (__data == 0 && bfqd->low_latency != 0)
5767 bfqd->low_latency = __data;
5772 #define BFQ_ATTR(name) \
5773 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
5775 static struct elv_fs_entry bfq_attrs[] = {
5776 BFQ_ATTR(fifo_expire_sync),
5777 BFQ_ATTR(fifo_expire_async),
5778 BFQ_ATTR(back_seek_max),
5779 BFQ_ATTR(back_seek_penalty),
5780 BFQ_ATTR(slice_idle),
5781 BFQ_ATTR(slice_idle_us),
5782 BFQ_ATTR(max_budget),
5783 BFQ_ATTR(timeout_sync),
5784 BFQ_ATTR(strict_guarantees),
5785 BFQ_ATTR(low_latency),
5789 static struct elevator_type iosched_bfq_mq = {
5791 .limit_depth = bfq_limit_depth,
5792 .prepare_request = bfq_prepare_request,
5793 .requeue_request = bfq_finish_requeue_request,
5794 .finish_request = bfq_finish_requeue_request,
5795 .exit_icq = bfq_exit_icq,
5796 .insert_requests = bfq_insert_requests,
5797 .dispatch_request = bfq_dispatch_request,
5798 .next_request = elv_rb_latter_request,
5799 .former_request = elv_rb_former_request,
5800 .allow_merge = bfq_allow_bio_merge,
5801 .bio_merge = bfq_bio_merge,
5802 .request_merge = bfq_request_merge,
5803 .requests_merged = bfq_requests_merged,
5804 .request_merged = bfq_request_merged,
5805 .has_work = bfq_has_work,
5806 .init_hctx = bfq_init_hctx,
5807 .init_sched = bfq_init_queue,
5808 .exit_sched = bfq_exit_queue,
5811 .icq_size = sizeof(struct bfq_io_cq),
5812 .icq_align = __alignof__(struct bfq_io_cq),
5813 .elevator_attrs = bfq_attrs,
5814 .elevator_name = "bfq",
5815 .elevator_owner = THIS_MODULE,
5817 MODULE_ALIAS("bfq-iosched");
5819 static int __init bfq_init(void)
5823 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5824 ret = blkcg_policy_register(&blkcg_policy_bfq);
5830 if (bfq_slab_setup())
5834 * Times to load large popular applications for the typical
5835 * systems installed on the reference devices (see the
5836 * comments before the definition of the next
5837 * array). Actually, we use slightly lower values, as the
5838 * estimated peak rate tends to be smaller than the actual
5839 * peak rate. The reason for this last fact is that estimates
5840 * are computed over much shorter time intervals than the long
5841 * intervals typically used for benchmarking. Why? First, to
5842 * adapt more quickly to variations. Second, because an I/O
5843 * scheduler cannot rely on a peak-rate-evaluation workload to
5844 * be run for a long time.
5846 ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
5847 ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
5849 ret = elv_register(&iosched_bfq_mq);
5858 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5859 blkcg_policy_unregister(&blkcg_policy_bfq);
5864 static void __exit bfq_exit(void)
5866 elv_unregister(&iosched_bfq_mq);
5867 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5868 blkcg_policy_unregister(&blkcg_policy_bfq);
5873 module_init(bfq_init);
5874 module_exit(bfq_exit);
5876 MODULE_AUTHOR("Paolo Valente");
5877 MODULE_LICENSE("GPL");
5878 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");