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>
135 #include <linux/backing-dev.h>
139 #include "blk-mq-tag.h"
140 #include "blk-mq-sched.h"
141 #include "bfq-iosched.h"
144 #define BFQ_BFQQ_FNS(name) \
145 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
147 __set_bit(BFQQF_##name, &(bfqq)->flags); \
149 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
151 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
153 int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
155 return test_bit(BFQQF_##name, &(bfqq)->flags); \
158 BFQ_BFQQ_FNS(just_created);
160 BFQ_BFQQ_FNS(wait_request);
161 BFQ_BFQQ_FNS(non_blocking_wait_rq);
162 BFQ_BFQQ_FNS(fifo_expire);
163 BFQ_BFQQ_FNS(has_short_ttime);
165 BFQ_BFQQ_FNS(IO_bound);
166 BFQ_BFQQ_FNS(in_large_burst);
168 BFQ_BFQQ_FNS(split_coop);
169 BFQ_BFQQ_FNS(softrt_update);
170 #undef BFQ_BFQQ_FNS \
172 /* Expiration time of sync (0) and async (1) requests, in ns. */
173 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
175 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
176 static const int bfq_back_max = 16 * 1024;
178 /* Penalty of a backwards seek, in number of sectors. */
179 static const int bfq_back_penalty = 2;
181 /* Idling period duration, in ns. */
182 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
184 /* Minimum number of assigned budgets for which stats are safe to compute. */
185 static const int bfq_stats_min_budgets = 194;
187 /* Default maximum budget values, in sectors and number of requests. */
188 static const int bfq_default_max_budget = 16 * 1024;
191 * When a sync request is dispatched, the queue that contains that
192 * request, and all the ancestor entities of that queue, are charged
193 * with the number of sectors of the request. In constrast, if the
194 * request is async, then the queue and its ancestor entities are
195 * charged with the number of sectors of the request, multiplied by
196 * the factor below. This throttles the bandwidth for async I/O,
197 * w.r.t. to sync I/O, and it is done to counter the tendency of async
198 * writes to steal I/O throughput to reads.
200 * The current value of this parameter is the result of a tuning with
201 * several hardware and software configurations. We tried to find the
202 * lowest value for which writes do not cause noticeable problems to
203 * reads. In fact, the lower this parameter, the stabler I/O control,
204 * in the following respect. The lower this parameter is, the less
205 * the bandwidth enjoyed by a group decreases
206 * - when the group does writes, w.r.t. to when it does reads;
207 * - when other groups do reads, w.r.t. to when they do writes.
209 static const int bfq_async_charge_factor = 3;
211 /* Default timeout values, in jiffies, approximating CFQ defaults. */
212 const int bfq_timeout = HZ / 8;
215 * Time limit for merging (see comments in bfq_setup_cooperator). Set
216 * to the slowest value that, in our tests, proved to be effective in
217 * removing false positives, while not causing true positives to miss
220 * As can be deduced from the low time limit below, queue merging, if
221 * successful, happens at the very beggining of the I/O of the involved
222 * cooperating processes, as a consequence of the arrival of the very
223 * first requests from each cooperator. After that, there is very
224 * little chance to find cooperators.
226 static const unsigned long bfq_merge_time_limit = HZ/10;
228 static struct kmem_cache *bfq_pool;
230 /* Below this threshold (in ns), we consider thinktime immediate. */
231 #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
233 /* hw_tag detection: parallel requests threshold and min samples needed. */
234 #define BFQ_HW_QUEUE_THRESHOLD 4
235 #define BFQ_HW_QUEUE_SAMPLES 32
237 #define BFQQ_SEEK_THR (sector_t)(8 * 100)
238 #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
239 #define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
240 #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19)
242 /* Min number of samples required to perform peak-rate update */
243 #define BFQ_RATE_MIN_SAMPLES 32
244 /* Min observation time interval required to perform a peak-rate update (ns) */
245 #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
246 /* Target observation time interval for a peak-rate update (ns) */
247 #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
250 * Shift used for peak-rate fixed precision calculations.
252 * - the current shift: 16 positions
253 * - the current type used to store rate: u32
254 * - the current unit of measure for rate: [sectors/usec], or, more precisely,
255 * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
256 * the range of rates that can be stored is
257 * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
258 * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
259 * [15, 65G] sectors/sec
260 * Which, assuming a sector size of 512B, corresponds to a range of
263 #define BFQ_RATE_SHIFT 16
266 * When configured for computing the duration of the weight-raising
267 * for interactive queues automatically (see the comments at the
268 * beginning of this file), BFQ does it using the following formula:
269 * duration = (ref_rate / r) * ref_wr_duration,
270 * where r is the peak rate of the device, and ref_rate and
271 * ref_wr_duration are two reference parameters. In particular,
272 * ref_rate is the peak rate of the reference storage device (see
273 * below), and ref_wr_duration is about the maximum time needed, with
274 * BFQ and while reading two files in parallel, to load typical large
275 * applications on the reference device (see the comments on
276 * max_service_from_wr below, for more details on how ref_wr_duration
277 * is obtained). In practice, the slower/faster the device at hand
278 * is, the more/less it takes to load applications with respect to the
279 * reference device. Accordingly, the longer/shorter BFQ grants
280 * weight raising to interactive applications.
282 * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
283 * depending on whether the device is rotational or non-rotational.
285 * In the following definitions, ref_rate[0] and ref_wr_duration[0]
286 * are the reference values for a rotational device, whereas
287 * ref_rate[1] and ref_wr_duration[1] are the reference values for a
288 * non-rotational device. The reference rates are not the actual peak
289 * rates of the devices used as a reference, but slightly lower
290 * values. The reason for using slightly lower values is that the
291 * peak-rate estimator tends to yield slightly lower values than the
292 * actual peak rate (it can yield the actual peak rate only if there
293 * is only one process doing I/O, and the process does sequential
296 * The reference peak rates are measured in sectors/usec, left-shifted
299 static int ref_rate[2] = {14000, 33000};
301 * To improve readability, a conversion function is used to initialize
302 * the following array, which entails that the array can be
303 * initialized only in a function.
305 static int ref_wr_duration[2];
308 * BFQ uses the above-detailed, time-based weight-raising mechanism to
309 * privilege interactive tasks. This mechanism is vulnerable to the
310 * following false positives: I/O-bound applications that will go on
311 * doing I/O for much longer than the duration of weight
312 * raising. These applications have basically no benefit from being
313 * weight-raised at the beginning of their I/O. On the opposite end,
314 * while being weight-raised, these applications
315 * a) unjustly steal throughput to applications that may actually need
317 * b) make BFQ uselessly perform device idling; device idling results
318 * in loss of device throughput with most flash-based storage, and may
319 * increase latencies when used purposelessly.
321 * BFQ tries to reduce these problems, by adopting the following
322 * countermeasure. To introduce this countermeasure, we need first to
323 * finish explaining how the duration of weight-raising for
324 * interactive tasks is computed.
326 * For a bfq_queue deemed as interactive, the duration of weight
327 * raising is dynamically adjusted, as a function of the estimated
328 * peak rate of the device, so as to be equal to the time needed to
329 * execute the 'largest' interactive task we benchmarked so far. By
330 * largest task, we mean the task for which each involved process has
331 * to do more I/O than for any of the other tasks we benchmarked. This
332 * reference interactive task is the start-up of LibreOffice Writer,
333 * and in this task each process/bfq_queue needs to have at most ~110K
334 * sectors transferred.
336 * This last piece of information enables BFQ to reduce the actual
337 * duration of weight-raising for at least one class of I/O-bound
338 * applications: those doing sequential or quasi-sequential I/O. An
339 * example is file copy. In fact, once started, the main I/O-bound
340 * processes of these applications usually consume the above 110K
341 * sectors in much less time than the processes of an application that
342 * is starting, because these I/O-bound processes will greedily devote
343 * almost all their CPU cycles only to their target,
344 * throughput-friendly I/O operations. This is even more true if BFQ
345 * happens to be underestimating the device peak rate, and thus
346 * overestimating the duration of weight raising. But, according to
347 * our measurements, once transferred 110K sectors, these processes
348 * have no right to be weight-raised any longer.
350 * Basing on the last consideration, BFQ ends weight-raising for a
351 * bfq_queue if the latter happens to have received an amount of
352 * service at least equal to the following constant. The constant is
353 * set to slightly more than 110K, to have a minimum safety margin.
355 * This early ending of weight-raising reduces the amount of time
356 * during which interactive false positives cause the two problems
357 * described at the beginning of these comments.
359 static const unsigned long max_service_from_wr = 120000;
361 #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
362 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
364 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
366 return bic->bfqq[is_sync];
369 void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
371 bic->bfqq[is_sync] = bfqq;
374 struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
376 return bic->icq.q->elevator->elevator_data;
380 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
381 * @icq: the iocontext queue.
383 static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
385 /* bic->icq is the first member, %NULL will convert to %NULL */
386 return container_of(icq, struct bfq_io_cq, icq);
390 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
391 * @bfqd: the lookup key.
392 * @ioc: the io_context of the process doing I/O.
393 * @q: the request queue.
395 static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd,
396 struct io_context *ioc,
397 struct request_queue *q)
401 struct bfq_io_cq *icq;
403 spin_lock_irqsave(q->queue_lock, flags);
404 icq = icq_to_bic(ioc_lookup_icq(ioc, q));
405 spin_unlock_irqrestore(q->queue_lock, flags);
414 * Scheduler run of queue, if there are requests pending and no one in the
415 * driver that will restart queueing.
417 void bfq_schedule_dispatch(struct bfq_data *bfqd)
419 if (bfqd->queued != 0) {
420 bfq_log(bfqd, "schedule dispatch");
421 blk_mq_run_hw_queues(bfqd->queue, true);
425 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
426 #define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT)
428 #define bfq_sample_valid(samples) ((samples) > 80)
431 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
432 * We choose the request that is closesr to the head right now. Distance
433 * behind the head is penalized and only allowed to a certain extent.
435 static struct request *bfq_choose_req(struct bfq_data *bfqd,
440 sector_t s1, s2, d1 = 0, d2 = 0;
441 unsigned long back_max;
442 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
443 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
444 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
446 if (!rq1 || rq1 == rq2)
451 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
453 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
455 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
457 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
460 s1 = blk_rq_pos(rq1);
461 s2 = blk_rq_pos(rq2);
464 * By definition, 1KiB is 2 sectors.
466 back_max = bfqd->bfq_back_max * 2;
469 * Strict one way elevator _except_ in the case where we allow
470 * short backward seeks which are biased as twice the cost of a
471 * similar forward seek.
475 else if (s1 + back_max >= last)
476 d1 = (last - s1) * bfqd->bfq_back_penalty;
478 wrap |= BFQ_RQ1_WRAP;
482 else if (s2 + back_max >= last)
483 d2 = (last - s2) * bfqd->bfq_back_penalty;
485 wrap |= BFQ_RQ2_WRAP;
487 /* Found required data */
490 * By doing switch() on the bit mask "wrap" we avoid having to
491 * check two variables for all permutations: --> faster!
494 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
509 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
512 * Since both rqs are wrapped,
513 * start with the one that's further behind head
514 * (--> only *one* back seek required),
515 * since back seek takes more time than forward.
525 * Async I/O can easily starve sync I/O (both sync reads and sync
526 * writes), by consuming all tags. Similarly, storms of sync writes,
527 * such as those that sync(2) may trigger, can starve sync reads.
528 * Limit depths of async I/O and sync writes so as to counter both
531 static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
533 struct bfq_data *bfqd = data->q->elevator->elevator_data;
535 if (op_is_sync(op) && !op_is_write(op))
538 data->shallow_depth =
539 bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)];
541 bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
542 __func__, bfqd->wr_busy_queues, op_is_sync(op),
543 data->shallow_depth);
546 static struct bfq_queue *
547 bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
548 sector_t sector, struct rb_node **ret_parent,
549 struct rb_node ***rb_link)
551 struct rb_node **p, *parent;
552 struct bfq_queue *bfqq = NULL;
560 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
563 * Sort strictly based on sector. Smallest to the left,
564 * largest to the right.
566 if (sector > blk_rq_pos(bfqq->next_rq))
568 else if (sector < blk_rq_pos(bfqq->next_rq))
576 *ret_parent = parent;
580 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
581 (unsigned long long)sector,
582 bfqq ? bfqq->pid : 0);
587 static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
589 return bfqq->service_from_backlogged > 0 &&
590 time_is_before_jiffies(bfqq->first_IO_time +
591 bfq_merge_time_limit);
594 void bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
596 struct rb_node **p, *parent;
597 struct bfq_queue *__bfqq;
599 if (bfqq->pos_root) {
600 rb_erase(&bfqq->pos_node, bfqq->pos_root);
601 bfqq->pos_root = NULL;
605 * bfqq cannot be merged any longer (see comments in
606 * bfq_setup_cooperator): no point in adding bfqq into the
609 if (bfq_too_late_for_merging(bfqq))
612 if (bfq_class_idle(bfqq))
617 bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
618 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
619 blk_rq_pos(bfqq->next_rq), &parent, &p);
621 rb_link_node(&bfqq->pos_node, parent, p);
622 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
624 bfqq->pos_root = NULL;
628 * Tell whether there are active queues with different weights or
631 static bool bfq_varied_queue_weights_or_active_groups(struct bfq_data *bfqd)
634 * For queue weights to differ, queue_weights_tree must contain
635 * at least two nodes.
637 return (!RB_EMPTY_ROOT(&bfqd->queue_weights_tree) &&
638 (bfqd->queue_weights_tree.rb_node->rb_left ||
639 bfqd->queue_weights_tree.rb_node->rb_right)
640 #ifdef CONFIG_BFQ_GROUP_IOSCHED
642 (bfqd->num_groups_with_pending_reqs > 0
648 * The following function returns true if every queue must receive the
649 * same share of the throughput (this condition is used when deciding
650 * whether idling may be disabled, see the comments in the function
651 * bfq_better_to_idle()).
653 * Such a scenario occurs when:
654 * 1) all active queues have the same weight,
655 * 2) all active groups at the same level in the groups tree have the same
657 * 3) all active groups at the same level in the groups tree have the same
658 * number of children.
660 * Unfortunately, keeping the necessary state for evaluating exactly
661 * the last two symmetry sub-conditions above would be quite complex
662 * and time consuming. Therefore this function evaluates, instead,
663 * only the following stronger two sub-conditions, for which it is
664 * much easier to maintain the needed state:
665 * 1) all active queues have the same weight,
666 * 2) there are no active groups.
667 * In particular, the last condition is always true if hierarchical
668 * support or the cgroups interface are not enabled, thus no state
669 * needs to be maintained in this case.
671 static bool bfq_symmetric_scenario(struct bfq_data *bfqd)
673 return !bfq_varied_queue_weights_or_active_groups(bfqd);
677 * If the weight-counter tree passed as input contains no counter for
678 * the weight of the input queue, then add that counter; otherwise just
679 * increment the existing counter.
681 * Note that weight-counter trees contain few nodes in mostly symmetric
682 * scenarios. For example, if all queues have the same weight, then the
683 * weight-counter tree for the queues may contain at most one node.
684 * This holds even if low_latency is on, because weight-raised queues
685 * are not inserted in the tree.
686 * In most scenarios, the rate at which nodes are created/destroyed
689 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq,
690 struct rb_root *root)
692 struct bfq_entity *entity = &bfqq->entity;
693 struct rb_node **new = &(root->rb_node), *parent = NULL;
696 * Do not insert if the queue is already associated with a
697 * counter, which happens if:
698 * 1) a request arrival has caused the queue to become both
699 * non-weight-raised, and hence change its weight, and
700 * backlogged; in this respect, each of the two events
701 * causes an invocation of this function,
702 * 2) this is the invocation of this function caused by the
703 * second event. This second invocation is actually useless,
704 * and we handle this fact by exiting immediately. More
705 * efficient or clearer solutions might possibly be adopted.
707 if (bfqq->weight_counter)
711 struct bfq_weight_counter *__counter = container_of(*new,
712 struct bfq_weight_counter,
716 if (entity->weight == __counter->weight) {
717 bfqq->weight_counter = __counter;
720 if (entity->weight < __counter->weight)
721 new = &((*new)->rb_left);
723 new = &((*new)->rb_right);
726 bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
730 * In the unlucky event of an allocation failure, we just
731 * exit. This will cause the weight of queue to not be
732 * considered in bfq_varied_queue_weights_or_active_groups,
733 * which, in its turn, causes the scenario to be deemed
734 * wrongly symmetric in case bfqq's weight would have been
735 * the only weight making the scenario asymmetric. On the
736 * bright side, no unbalance will however occur when bfqq
737 * becomes inactive again (the invocation of this function
738 * is triggered by an activation of queue). In fact,
739 * bfq_weights_tree_remove does nothing if
740 * !bfqq->weight_counter.
742 if (unlikely(!bfqq->weight_counter))
745 bfqq->weight_counter->weight = entity->weight;
746 rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
747 rb_insert_color(&bfqq->weight_counter->weights_node, root);
750 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;
780 * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
781 * of active groups for each queue's inactive parent entity.
783 void bfq_weights_tree_remove(struct bfq_data *bfqd,
784 struct bfq_queue *bfqq)
786 struct bfq_entity *entity = bfqq->entity.parent;
788 for_each_entity(entity) {
789 struct bfq_sched_data *sd = entity->my_sched_data;
791 if (sd->next_in_service || sd->in_service_entity) {
793 * entity is still active, because either
794 * next_in_service or in_service_entity is not
795 * NULL (see the comments on the definition of
796 * next_in_service for details on why
797 * in_service_entity must be checked too).
799 * As a consequence, its parent entities are
800 * active as well, and thus this loop must
807 * The decrement of num_groups_with_pending_reqs is
808 * not performed immediately upon the deactivation of
809 * entity, but it is delayed to when it also happens
810 * that the first leaf descendant bfqq of entity gets
811 * all its pending requests completed. The following
812 * instructions perform this delayed decrement, if
813 * needed. See the comments on
814 * num_groups_with_pending_reqs for details.
816 if (entity->in_groups_with_pending_reqs) {
817 entity->in_groups_with_pending_reqs = false;
818 bfqd->num_groups_with_pending_reqs--;
823 * Next function is invoked last, because it causes bfqq to be
824 * freed if the following holds: bfqq is not in service and
825 * has no dispatched request. DO NOT use bfqq after the next
826 * function invocation.
828 __bfq_weights_tree_remove(bfqd, bfqq,
829 &bfqd->queue_weights_tree);
833 * Return expired entry, or NULL to just start from scratch in rbtree.
835 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
836 struct request *last)
840 if (bfq_bfqq_fifo_expire(bfqq))
843 bfq_mark_bfqq_fifo_expire(bfqq);
845 rq = rq_entry_fifo(bfqq->fifo.next);
847 if (rq == last || ktime_get_ns() < rq->fifo_time)
850 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
854 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
855 struct bfq_queue *bfqq,
856 struct request *last)
858 struct rb_node *rbnext = rb_next(&last->rb_node);
859 struct rb_node *rbprev = rb_prev(&last->rb_node);
860 struct request *next, *prev = NULL;
862 /* Follow expired path, else get first next available. */
863 next = bfq_check_fifo(bfqq, last);
868 prev = rb_entry_rq(rbprev);
871 next = rb_entry_rq(rbnext);
873 rbnext = rb_first(&bfqq->sort_list);
874 if (rbnext && rbnext != &last->rb_node)
875 next = rb_entry_rq(rbnext);
878 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
881 /* see the definition of bfq_async_charge_factor for details */
882 static unsigned long bfq_serv_to_charge(struct request *rq,
883 struct bfq_queue *bfqq)
885 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1)
886 return blk_rq_sectors(rq);
888 return blk_rq_sectors(rq) * bfq_async_charge_factor;
892 * bfq_updated_next_req - update the queue after a new next_rq selection.
893 * @bfqd: the device data the queue belongs to.
894 * @bfqq: the queue to update.
896 * If the first request of a queue changes we make sure that the queue
897 * has enough budget to serve at least its first request (if the
898 * request has grown). We do this because if the queue has not enough
899 * budget for its first request, it has to go through two dispatch
900 * rounds to actually get it dispatched.
902 static void bfq_updated_next_req(struct bfq_data *bfqd,
903 struct bfq_queue *bfqq)
905 struct bfq_entity *entity = &bfqq->entity;
906 struct request *next_rq = bfqq->next_rq;
907 unsigned long new_budget;
912 if (bfqq == bfqd->in_service_queue)
914 * In order not to break guarantees, budgets cannot be
915 * changed after an entity has been selected.
919 new_budget = max_t(unsigned long, bfqq->max_budget,
920 bfq_serv_to_charge(next_rq, bfqq));
921 if (entity->budget != new_budget) {
922 entity->budget = new_budget;
923 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
925 bfq_requeue_bfqq(bfqd, bfqq, false);
929 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
933 if (bfqd->bfq_wr_max_time > 0)
934 return bfqd->bfq_wr_max_time;
936 dur = bfqd->rate_dur_prod;
937 do_div(dur, bfqd->peak_rate);
940 * Limit duration between 3 and 25 seconds. The upper limit
941 * has been conservatively set after the following worst case:
942 * on a QEMU/KVM virtual machine
943 * - running in a slow PC
944 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
945 * - serving a heavy I/O workload, such as the sequential reading
947 * mplayer took 23 seconds to start, if constantly weight-raised.
949 * As for higher values than that accomodating the above bad
950 * scenario, tests show that higher values would often yield
951 * the opposite of the desired result, i.e., would worsen
952 * responsiveness by allowing non-interactive applications to
953 * preserve weight raising for too long.
955 * On the other end, lower values than 3 seconds make it
956 * difficult for most interactive tasks to complete their jobs
957 * before weight-raising finishes.
959 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
962 /* switch back from soft real-time to interactive weight raising */
963 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
964 struct bfq_data *bfqd)
966 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
967 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
968 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
972 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
973 struct bfq_io_cq *bic, bool bfq_already_existing)
975 unsigned int old_wr_coeff = bfqq->wr_coeff;
976 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
978 if (bic->saved_has_short_ttime)
979 bfq_mark_bfqq_has_short_ttime(bfqq);
981 bfq_clear_bfqq_has_short_ttime(bfqq);
983 if (bic->saved_IO_bound)
984 bfq_mark_bfqq_IO_bound(bfqq);
986 bfq_clear_bfqq_IO_bound(bfqq);
988 bfqq->ttime = bic->saved_ttime;
989 bfqq->wr_coeff = bic->saved_wr_coeff;
990 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
991 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
992 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
994 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
995 time_is_before_jiffies(bfqq->last_wr_start_finish +
996 bfqq->wr_cur_max_time))) {
997 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
998 !bfq_bfqq_in_large_burst(bfqq) &&
999 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1000 bfq_wr_duration(bfqd))) {
1001 switch_back_to_interactive_wr(bfqq, bfqd);
1004 bfq_log_bfqq(bfqq->bfqd, bfqq,
1005 "resume state: switching off wr");
1009 /* make sure weight will be updated, however we got here */
1010 bfqq->entity.prio_changed = 1;
1015 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1016 bfqd->wr_busy_queues++;
1017 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1018 bfqd->wr_busy_queues--;
1021 static int bfqq_process_refs(struct bfq_queue *bfqq)
1023 return bfqq->ref - bfqq->allocated - bfqq->entity.on_st -
1024 (bfqq->weight_counter != NULL);
1027 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1028 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1030 struct bfq_queue *item;
1031 struct hlist_node *n;
1033 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1034 hlist_del_init(&item->burst_list_node);
1035 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1036 bfqd->burst_size = 1;
1037 bfqd->burst_parent_entity = bfqq->entity.parent;
1040 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1041 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1043 /* Increment burst size to take into account also bfqq */
1046 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1047 struct bfq_queue *pos, *bfqq_item;
1048 struct hlist_node *n;
1051 * Enough queues have been activated shortly after each
1052 * other to consider this burst as large.
1054 bfqd->large_burst = true;
1057 * We can now mark all queues in the burst list as
1058 * belonging to a large burst.
1060 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1062 bfq_mark_bfqq_in_large_burst(bfqq_item);
1063 bfq_mark_bfqq_in_large_burst(bfqq);
1066 * From now on, and until the current burst finishes, any
1067 * new queue being activated shortly after the last queue
1068 * was inserted in the burst can be immediately marked as
1069 * belonging to a large burst. So the burst list is not
1070 * needed any more. Remove it.
1072 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1074 hlist_del_init(&pos->burst_list_node);
1076 * Burst not yet large: add bfqq to the burst list. Do
1077 * not increment the ref counter for bfqq, because bfqq
1078 * is removed from the burst list before freeing bfqq
1081 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1085 * If many queues belonging to the same group happen to be created
1086 * shortly after each other, then the processes associated with these
1087 * queues have typically a common goal. In particular, bursts of queue
1088 * creations are usually caused by services or applications that spawn
1089 * many parallel threads/processes. Examples are systemd during boot,
1090 * or git grep. To help these processes get their job done as soon as
1091 * possible, it is usually better to not grant either weight-raising
1092 * or device idling to their queues.
1094 * In this comment we describe, firstly, the reasons why this fact
1095 * holds, and, secondly, the next function, which implements the main
1096 * steps needed to properly mark these queues so that they can then be
1097 * treated in a different way.
1099 * The above services or applications benefit mostly from a high
1100 * throughput: the quicker the requests of the activated queues are
1101 * cumulatively served, the sooner the target job of these queues gets
1102 * completed. As a consequence, weight-raising any of these queues,
1103 * which also implies idling the device for it, is almost always
1104 * counterproductive. In most cases it just lowers throughput.
1106 * On the other hand, a burst of queue creations may be caused also by
1107 * the start of an application that does not consist of a lot of
1108 * parallel I/O-bound threads. In fact, with a complex application,
1109 * several short processes may need to be executed to start-up the
1110 * application. In this respect, to start an application as quickly as
1111 * possible, the best thing to do is in any case to privilege the I/O
1112 * related to the application with respect to all other
1113 * I/O. Therefore, the best strategy to start as quickly as possible
1114 * an application that causes a burst of queue creations is to
1115 * weight-raise all the queues created during the burst. This is the
1116 * exact opposite of the best strategy for the other type of bursts.
1118 * In the end, to take the best action for each of the two cases, the
1119 * two types of bursts need to be distinguished. Fortunately, this
1120 * seems relatively easy, by looking at the sizes of the bursts. In
1121 * particular, we found a threshold such that only bursts with a
1122 * larger size than that threshold are apparently caused by
1123 * services or commands such as systemd or git grep. For brevity,
1124 * hereafter we call just 'large' these bursts. BFQ *does not*
1125 * weight-raise queues whose creation occurs in a large burst. In
1126 * addition, for each of these queues BFQ performs or does not perform
1127 * idling depending on which choice boosts the throughput more. The
1128 * exact choice depends on the device and request pattern at
1131 * Unfortunately, false positives may occur while an interactive task
1132 * is starting (e.g., an application is being started). The
1133 * consequence is that the queues associated with the task do not
1134 * enjoy weight raising as expected. Fortunately these false positives
1135 * are very rare. They typically occur if some service happens to
1136 * start doing I/O exactly when the interactive task starts.
1138 * Turning back to the next function, it implements all the steps
1139 * needed to detect the occurrence of a large burst and to properly
1140 * mark all the queues belonging to it (so that they can then be
1141 * treated in a different way). This goal is achieved by maintaining a
1142 * "burst list" that holds, temporarily, the queues that belong to the
1143 * burst in progress. The list is then used to mark these queues as
1144 * belonging to a large burst if the burst does become large. The main
1145 * steps are the following.
1147 * . when the very first queue is created, the queue is inserted into the
1148 * list (as it could be the first queue in a possible burst)
1150 * . if the current burst has not yet become large, and a queue Q that does
1151 * not yet belong to the burst is activated shortly after the last time
1152 * at which a new queue entered the burst list, then the function appends
1153 * Q to the burst list
1155 * . if, as a consequence of the previous step, the burst size reaches
1156 * the large-burst threshold, then
1158 * . all the queues in the burst list are marked as belonging to a
1161 * . the burst list is deleted; in fact, the burst list already served
1162 * its purpose (keeping temporarily track of the queues in a burst,
1163 * so as to be able to mark them as belonging to a large burst in the
1164 * previous sub-step), and now is not needed any more
1166 * . the device enters a large-burst mode
1168 * . if a queue Q that does not belong to the burst is created while
1169 * the device is in large-burst mode and shortly after the last time
1170 * at which a queue either entered the burst list or was marked as
1171 * belonging to the current large burst, then Q is immediately marked
1172 * as belonging to a large burst.
1174 * . if a queue Q that does not belong to the burst is created a while
1175 * later, i.e., not shortly after, than the last time at which a queue
1176 * either entered the burst list or was marked as belonging to the
1177 * current large burst, then the current burst is deemed as finished and:
1179 * . the large-burst mode is reset if set
1181 * . the burst list is emptied
1183 * . Q is inserted in the burst list, as Q may be the first queue
1184 * in a possible new burst (then the burst list contains just Q
1187 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1190 * If bfqq is already in the burst list or is part of a large
1191 * burst, or finally has just been split, then there is
1192 * nothing else to do.
1194 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1195 bfq_bfqq_in_large_burst(bfqq) ||
1196 time_is_after_eq_jiffies(bfqq->split_time +
1197 msecs_to_jiffies(10)))
1201 * If bfqq's creation happens late enough, or bfqq belongs to
1202 * a different group than the burst group, then the current
1203 * burst is finished, and related data structures must be
1206 * In this respect, consider the special case where bfqq is
1207 * the very first queue created after BFQ is selected for this
1208 * device. In this case, last_ins_in_burst and
1209 * burst_parent_entity are not yet significant when we get
1210 * here. But it is easy to verify that, whether or not the
1211 * following condition is true, bfqq will end up being
1212 * inserted into the burst list. In particular the list will
1213 * happen to contain only bfqq. And this is exactly what has
1214 * to happen, as bfqq may be the first queue of the first
1217 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1218 bfqd->bfq_burst_interval) ||
1219 bfqq->entity.parent != bfqd->burst_parent_entity) {
1220 bfqd->large_burst = false;
1221 bfq_reset_burst_list(bfqd, bfqq);
1226 * If we get here, then bfqq is being activated shortly after the
1227 * last queue. So, if the current burst is also large, we can mark
1228 * bfqq as belonging to this large burst immediately.
1230 if (bfqd->large_burst) {
1231 bfq_mark_bfqq_in_large_burst(bfqq);
1236 * If we get here, then a large-burst state has not yet been
1237 * reached, but bfqq is being activated shortly after the last
1238 * queue. Then we add bfqq to the burst.
1240 bfq_add_to_burst(bfqd, bfqq);
1243 * At this point, bfqq either has been added to the current
1244 * burst or has caused the current burst to terminate and a
1245 * possible new burst to start. In particular, in the second
1246 * case, bfqq has become the first queue in the possible new
1247 * burst. In both cases last_ins_in_burst needs to be moved
1250 bfqd->last_ins_in_burst = jiffies;
1253 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1255 struct bfq_entity *entity = &bfqq->entity;
1257 return entity->budget - entity->service;
1261 * If enough samples have been computed, return the current max budget
1262 * stored in bfqd, which is dynamically updated according to the
1263 * estimated disk peak rate; otherwise return the default max budget
1265 static int bfq_max_budget(struct bfq_data *bfqd)
1267 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1268 return bfq_default_max_budget;
1270 return bfqd->bfq_max_budget;
1274 * Return min budget, which is a fraction of the current or default
1275 * max budget (trying with 1/32)
1277 static int bfq_min_budget(struct bfq_data *bfqd)
1279 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1280 return bfq_default_max_budget / 32;
1282 return bfqd->bfq_max_budget / 32;
1286 * The next function, invoked after the input queue bfqq switches from
1287 * idle to busy, updates the budget of bfqq. The function also tells
1288 * whether the in-service queue should be expired, by returning
1289 * true. The purpose of expiring the in-service queue is to give bfqq
1290 * the chance to possibly preempt the in-service queue, and the reason
1291 * for preempting the in-service queue is to achieve one of the two
1294 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1295 * expired because it has remained idle. In particular, bfqq may have
1296 * expired for one of the following two reasons:
1298 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1299 * and did not make it to issue a new request before its last
1300 * request was served;
1302 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1303 * a new request before the expiration of the idling-time.
1305 * Even if bfqq has expired for one of the above reasons, the process
1306 * associated with the queue may be however issuing requests greedily,
1307 * and thus be sensitive to the bandwidth it receives (bfqq may have
1308 * remained idle for other reasons: CPU high load, bfqq not enjoying
1309 * idling, I/O throttling somewhere in the path from the process to
1310 * the I/O scheduler, ...). But if, after every expiration for one of
1311 * the above two reasons, bfqq has to wait for the service of at least
1312 * one full budget of another queue before being served again, then
1313 * bfqq is likely to get a much lower bandwidth or resource time than
1314 * its reserved ones. To address this issue, two countermeasures need
1317 * First, the budget and the timestamps of bfqq need to be updated in
1318 * a special way on bfqq reactivation: they need to be updated as if
1319 * bfqq did not remain idle and did not expire. In fact, if they are
1320 * computed as if bfqq expired and remained idle until reactivation,
1321 * then the process associated with bfqq is treated as if, instead of
1322 * being greedy, it stopped issuing requests when bfqq remained idle,
1323 * and restarts issuing requests only on this reactivation. In other
1324 * words, the scheduler does not help the process recover the "service
1325 * hole" between bfqq expiration and reactivation. As a consequence,
1326 * the process receives a lower bandwidth than its reserved one. In
1327 * contrast, to recover this hole, the budget must be updated as if
1328 * bfqq was not expired at all before this reactivation, i.e., it must
1329 * be set to the value of the remaining budget when bfqq was
1330 * expired. Along the same line, timestamps need to be assigned the
1331 * value they had the last time bfqq was selected for service, i.e.,
1332 * before last expiration. Thus timestamps need to be back-shifted
1333 * with respect to their normal computation (see [1] for more details
1334 * on this tricky aspect).
1336 * Secondly, to allow the process to recover the hole, the in-service
1337 * queue must be expired too, to give bfqq the chance to preempt it
1338 * immediately. In fact, if bfqq has to wait for a full budget of the
1339 * in-service queue to be completed, then it may become impossible to
1340 * let the process recover the hole, even if the back-shifted
1341 * timestamps of bfqq are lower than those of the in-service queue. If
1342 * this happens for most or all of the holes, then the process may not
1343 * receive its reserved bandwidth. In this respect, it is worth noting
1344 * that, being the service of outstanding requests unpreemptible, a
1345 * little fraction of the holes may however be unrecoverable, thereby
1346 * causing a little loss of bandwidth.
1348 * The last important point is detecting whether bfqq does need this
1349 * bandwidth recovery. In this respect, the next function deems the
1350 * process associated with bfqq greedy, and thus allows it to recover
1351 * the hole, if: 1) the process is waiting for the arrival of a new
1352 * request (which implies that bfqq expired for one of the above two
1353 * reasons), and 2) such a request has arrived soon. The first
1354 * condition is controlled through the flag non_blocking_wait_rq,
1355 * while the second through the flag arrived_in_time. If both
1356 * conditions hold, then the function computes the budget in the
1357 * above-described special way, and signals that the in-service queue
1358 * should be expired. Timestamp back-shifting is done later in
1359 * __bfq_activate_entity.
1361 * 2. Reduce latency. Even if timestamps are not backshifted to let
1362 * the process associated with bfqq recover a service hole, bfqq may
1363 * however happen to have, after being (re)activated, a lower finish
1364 * timestamp than the in-service queue. That is, the next budget of
1365 * bfqq may have to be completed before the one of the in-service
1366 * queue. If this is the case, then preempting the in-service queue
1367 * allows this goal to be achieved, apart from the unpreemptible,
1368 * outstanding requests mentioned above.
1370 * Unfortunately, regardless of which of the above two goals one wants
1371 * to achieve, service trees need first to be updated to know whether
1372 * the in-service queue must be preempted. To have service trees
1373 * correctly updated, the in-service queue must be expired and
1374 * rescheduled, and bfqq must be scheduled too. This is one of the
1375 * most costly operations (in future versions, the scheduling
1376 * mechanism may be re-designed in such a way to make it possible to
1377 * know whether preemption is needed without needing to update service
1378 * trees). In addition, queue preemptions almost always cause random
1379 * I/O, and thus loss of throughput. Because of these facts, the next
1380 * function adopts the following simple scheme to avoid both costly
1381 * operations and too frequent preemptions: it requests the expiration
1382 * of the in-service queue (unconditionally) only for queues that need
1383 * to recover a hole, or that either are weight-raised or deserve to
1386 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1387 struct bfq_queue *bfqq,
1388 bool arrived_in_time,
1389 bool wr_or_deserves_wr)
1391 struct bfq_entity *entity = &bfqq->entity;
1393 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time) {
1395 * We do not clear the flag non_blocking_wait_rq here, as
1396 * the latter is used in bfq_activate_bfqq to signal
1397 * that timestamps need to be back-shifted (and is
1398 * cleared right after).
1402 * In next assignment we rely on that either
1403 * entity->service or entity->budget are not updated
1404 * on expiration if bfqq is empty (see
1405 * __bfq_bfqq_recalc_budget). Thus both quantities
1406 * remain unchanged after such an expiration, and the
1407 * following statement therefore assigns to
1408 * entity->budget the remaining budget on such an
1411 entity->budget = min_t(unsigned long,
1412 bfq_bfqq_budget_left(bfqq),
1416 * At this point, we have used entity->service to get
1417 * the budget left (needed for updating
1418 * entity->budget). Thus we finally can, and have to,
1419 * reset entity->service. The latter must be reset
1420 * because bfqq would otherwise be charged again for
1421 * the service it has received during its previous
1424 entity->service = 0;
1430 * We can finally complete expiration, by setting service to 0.
1432 entity->service = 0;
1433 entity->budget = max_t(unsigned long, bfqq->max_budget,
1434 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1435 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1436 return wr_or_deserves_wr;
1440 * Return the farthest past time instant according to jiffies
1443 static unsigned long bfq_smallest_from_now(void)
1445 return jiffies - MAX_JIFFY_OFFSET;
1448 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1449 struct bfq_queue *bfqq,
1450 unsigned int old_wr_coeff,
1451 bool wr_or_deserves_wr,
1456 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1457 /* start a weight-raising period */
1459 bfqq->service_from_wr = 0;
1460 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1461 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1464 * No interactive weight raising in progress
1465 * here: assign minus infinity to
1466 * wr_start_at_switch_to_srt, to make sure
1467 * that, at the end of the soft-real-time
1468 * weight raising periods that is starting
1469 * now, no interactive weight-raising period
1470 * may be wrongly considered as still in
1471 * progress (and thus actually started by
1474 bfqq->wr_start_at_switch_to_srt =
1475 bfq_smallest_from_now();
1476 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1477 BFQ_SOFTRT_WEIGHT_FACTOR;
1478 bfqq->wr_cur_max_time =
1479 bfqd->bfq_wr_rt_max_time;
1483 * If needed, further reduce budget to make sure it is
1484 * close to bfqq's backlog, so as to reduce the
1485 * scheduling-error component due to a too large
1486 * budget. Do not care about throughput consequences,
1487 * but only about latency. Finally, do not assign a
1488 * too small budget either, to avoid increasing
1489 * latency by causing too frequent expirations.
1491 bfqq->entity.budget = min_t(unsigned long,
1492 bfqq->entity.budget,
1493 2 * bfq_min_budget(bfqd));
1494 } else if (old_wr_coeff > 1) {
1495 if (interactive) { /* update wr coeff and duration */
1496 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1497 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1498 } else if (in_burst)
1502 * The application is now or still meeting the
1503 * requirements for being deemed soft rt. We
1504 * can then correctly and safely (re)charge
1505 * the weight-raising duration for the
1506 * application with the weight-raising
1507 * duration for soft rt applications.
1509 * In particular, doing this recharge now, i.e.,
1510 * before the weight-raising period for the
1511 * application finishes, reduces the probability
1512 * of the following negative scenario:
1513 * 1) the weight of a soft rt application is
1514 * raised at startup (as for any newly
1515 * created application),
1516 * 2) since the application is not interactive,
1517 * at a certain time weight-raising is
1518 * stopped for the application,
1519 * 3) at that time the application happens to
1520 * still have pending requests, and hence
1521 * is destined to not have a chance to be
1522 * deemed soft rt before these requests are
1523 * completed (see the comments to the
1524 * function bfq_bfqq_softrt_next_start()
1525 * for details on soft rt detection),
1526 * 4) these pending requests experience a high
1527 * latency because the application is not
1528 * weight-raised while they are pending.
1530 if (bfqq->wr_cur_max_time !=
1531 bfqd->bfq_wr_rt_max_time) {
1532 bfqq->wr_start_at_switch_to_srt =
1533 bfqq->last_wr_start_finish;
1535 bfqq->wr_cur_max_time =
1536 bfqd->bfq_wr_rt_max_time;
1537 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1538 BFQ_SOFTRT_WEIGHT_FACTOR;
1540 bfqq->last_wr_start_finish = jiffies;
1545 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1546 struct bfq_queue *bfqq)
1548 return bfqq->dispatched == 0 &&
1549 time_is_before_jiffies(
1550 bfqq->budget_timeout +
1551 bfqd->bfq_wr_min_idle_time);
1554 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1555 struct bfq_queue *bfqq,
1560 bool soft_rt, in_burst, wr_or_deserves_wr,
1561 bfqq_wants_to_preempt,
1562 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1564 * See the comments on
1565 * bfq_bfqq_update_budg_for_activation for
1566 * details on the usage of the next variable.
1568 arrived_in_time = ktime_get_ns() <=
1569 bfqq->ttime.last_end_request +
1570 bfqd->bfq_slice_idle * 3;
1574 * bfqq deserves to be weight-raised if:
1576 * - it does not belong to a large burst,
1577 * - it has been idle for enough time or is soft real-time,
1578 * - is linked to a bfq_io_cq (it is not shared in any sense).
1580 in_burst = bfq_bfqq_in_large_burst(bfqq);
1581 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1583 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1584 bfqq->dispatched == 0;
1585 *interactive = !in_burst && idle_for_long_time;
1586 wr_or_deserves_wr = bfqd->low_latency &&
1587 (bfqq->wr_coeff > 1 ||
1588 (bfq_bfqq_sync(bfqq) &&
1589 bfqq->bic && (*interactive || soft_rt)));
1592 * Using the last flag, update budget and check whether bfqq
1593 * may want to preempt the in-service queue.
1595 bfqq_wants_to_preempt =
1596 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1601 * If bfqq happened to be activated in a burst, but has been
1602 * idle for much more than an interactive queue, then we
1603 * assume that, in the overall I/O initiated in the burst, the
1604 * I/O associated with bfqq is finished. So bfqq does not need
1605 * to be treated as a queue belonging to a burst
1606 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1607 * if set, and remove bfqq from the burst list if it's
1608 * there. We do not decrement burst_size, because the fact
1609 * that bfqq does not need to belong to the burst list any
1610 * more does not invalidate the fact that bfqq was created in
1613 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1614 idle_for_long_time &&
1615 time_is_before_jiffies(
1616 bfqq->budget_timeout +
1617 msecs_to_jiffies(10000))) {
1618 hlist_del_init(&bfqq->burst_list_node);
1619 bfq_clear_bfqq_in_large_burst(bfqq);
1622 bfq_clear_bfqq_just_created(bfqq);
1625 if (!bfq_bfqq_IO_bound(bfqq)) {
1626 if (arrived_in_time) {
1627 bfqq->requests_within_timer++;
1628 if (bfqq->requests_within_timer >=
1629 bfqd->bfq_requests_within_timer)
1630 bfq_mark_bfqq_IO_bound(bfqq);
1632 bfqq->requests_within_timer = 0;
1635 if (bfqd->low_latency) {
1636 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1639 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1641 if (time_is_before_jiffies(bfqq->split_time +
1642 bfqd->bfq_wr_min_idle_time)) {
1643 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1650 if (old_wr_coeff != bfqq->wr_coeff)
1651 bfqq->entity.prio_changed = 1;
1655 bfqq->last_idle_bklogged = jiffies;
1656 bfqq->service_from_backlogged = 0;
1657 bfq_clear_bfqq_softrt_update(bfqq);
1659 bfq_add_bfqq_busy(bfqd, bfqq);
1662 * Expire in-service queue only if preemption may be needed
1663 * for guarantees. In this respect, the function
1664 * next_queue_may_preempt just checks a simple, necessary
1665 * condition, and not a sufficient condition based on
1666 * timestamps. In fact, for the latter condition to be
1667 * evaluated, timestamps would need first to be updated, and
1668 * this operation is quite costly (see the comments on the
1669 * function bfq_bfqq_update_budg_for_activation).
1671 if (bfqd->in_service_queue && bfqq_wants_to_preempt &&
1672 bfqd->in_service_queue->wr_coeff < bfqq->wr_coeff &&
1673 next_queue_may_preempt(bfqd))
1674 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1675 false, BFQQE_PREEMPTED);
1678 static void bfq_add_request(struct request *rq)
1680 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1681 struct bfq_data *bfqd = bfqq->bfqd;
1682 struct request *next_rq, *prev;
1683 unsigned int old_wr_coeff = bfqq->wr_coeff;
1684 bool interactive = false;
1686 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1687 bfqq->queued[rq_is_sync(rq)]++;
1690 elv_rb_add(&bfqq->sort_list, rq);
1693 * Check if this request is a better next-serve candidate.
1695 prev = bfqq->next_rq;
1696 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
1697 bfqq->next_rq = next_rq;
1700 * Adjust priority tree position, if next_rq changes.
1702 if (prev != bfqq->next_rq)
1703 bfq_pos_tree_add_move(bfqd, bfqq);
1705 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
1706 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
1709 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
1710 time_is_before_jiffies(
1711 bfqq->last_wr_start_finish +
1712 bfqd->bfq_wr_min_inter_arr_async)) {
1713 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1714 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1716 bfqd->wr_busy_queues++;
1717 bfqq->entity.prio_changed = 1;
1719 if (prev != bfqq->next_rq)
1720 bfq_updated_next_req(bfqd, bfqq);
1724 * Assign jiffies to last_wr_start_finish in the following
1727 * . if bfqq is not going to be weight-raised, because, for
1728 * non weight-raised queues, last_wr_start_finish stores the
1729 * arrival time of the last request; as of now, this piece
1730 * of information is used only for deciding whether to
1731 * weight-raise async queues
1733 * . if bfqq is not weight-raised, because, if bfqq is now
1734 * switching to weight-raised, then last_wr_start_finish
1735 * stores the time when weight-raising starts
1737 * . if bfqq is interactive, because, regardless of whether
1738 * bfqq is currently weight-raised, the weight-raising
1739 * period must start or restart (this case is considered
1740 * separately because it is not detected by the above
1741 * conditions, if bfqq is already weight-raised)
1743 * last_wr_start_finish has to be updated also if bfqq is soft
1744 * real-time, because the weight-raising period is constantly
1745 * restarted on idle-to-busy transitions for these queues, but
1746 * this is already done in bfq_bfqq_handle_idle_busy_switch if
1749 if (bfqd->low_latency &&
1750 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
1751 bfqq->last_wr_start_finish = jiffies;
1754 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
1756 struct request_queue *q)
1758 struct bfq_queue *bfqq = bfqd->bio_bfqq;
1762 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
1767 static sector_t get_sdist(sector_t last_pos, struct request *rq)
1770 return abs(blk_rq_pos(rq) - last_pos);
1775 #if 0 /* Still not clear if we can do without next two functions */
1776 static void bfq_activate_request(struct request_queue *q, struct request *rq)
1778 struct bfq_data *bfqd = q->elevator->elevator_data;
1780 bfqd->rq_in_driver++;
1783 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
1785 struct bfq_data *bfqd = q->elevator->elevator_data;
1787 bfqd->rq_in_driver--;
1791 static void bfq_remove_request(struct request_queue *q,
1794 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1795 struct bfq_data *bfqd = bfqq->bfqd;
1796 const int sync = rq_is_sync(rq);
1798 if (bfqq->next_rq == rq) {
1799 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
1800 bfq_updated_next_req(bfqd, bfqq);
1803 if (rq->queuelist.prev != &rq->queuelist)
1804 list_del_init(&rq->queuelist);
1805 bfqq->queued[sync]--;
1807 elv_rb_del(&bfqq->sort_list, rq);
1809 elv_rqhash_del(q, rq);
1810 if (q->last_merge == rq)
1811 q->last_merge = NULL;
1813 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
1814 bfqq->next_rq = NULL;
1816 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
1817 bfq_del_bfqq_busy(bfqd, bfqq, false);
1819 * bfqq emptied. In normal operation, when
1820 * bfqq is empty, bfqq->entity.service and
1821 * bfqq->entity.budget must contain,
1822 * respectively, the service received and the
1823 * budget used last time bfqq emptied. These
1824 * facts do not hold in this case, as at least
1825 * this last removal occurred while bfqq is
1826 * not in service. To avoid inconsistencies,
1827 * reset both bfqq->entity.service and
1828 * bfqq->entity.budget, if bfqq has still a
1829 * process that may issue I/O requests to it.
1831 bfqq->entity.budget = bfqq->entity.service = 0;
1835 * Remove queue from request-position tree as it is empty.
1837 if (bfqq->pos_root) {
1838 rb_erase(&bfqq->pos_node, bfqq->pos_root);
1839 bfqq->pos_root = NULL;
1842 bfq_pos_tree_add_move(bfqd, bfqq);
1845 if (rq->cmd_flags & REQ_META)
1846 bfqq->meta_pending--;
1850 static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio)
1852 struct request_queue *q = hctx->queue;
1853 struct bfq_data *bfqd = q->elevator->elevator_data;
1854 struct request *free = NULL;
1856 * bfq_bic_lookup grabs the queue_lock: invoke it now and
1857 * store its return value for later use, to avoid nesting
1858 * queue_lock inside the bfqd->lock. We assume that the bic
1859 * returned by bfq_bic_lookup does not go away before
1860 * bfqd->lock is taken.
1862 struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
1865 spin_lock_irq(&bfqd->lock);
1868 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
1870 bfqd->bio_bfqq = NULL;
1871 bfqd->bio_bic = bic;
1873 ret = blk_mq_sched_try_merge(q, bio, &free);
1876 blk_mq_free_request(free);
1877 spin_unlock_irq(&bfqd->lock);
1882 static int bfq_request_merge(struct request_queue *q, struct request **req,
1885 struct bfq_data *bfqd = q->elevator->elevator_data;
1886 struct request *__rq;
1888 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
1889 if (__rq && elv_bio_merge_ok(__rq, bio)) {
1891 return ELEVATOR_FRONT_MERGE;
1894 return ELEVATOR_NO_MERGE;
1897 static struct bfq_queue *bfq_init_rq(struct request *rq);
1899 static void bfq_request_merged(struct request_queue *q, struct request *req,
1900 enum elv_merge type)
1902 if (type == ELEVATOR_FRONT_MERGE &&
1903 rb_prev(&req->rb_node) &&
1905 blk_rq_pos(container_of(rb_prev(&req->rb_node),
1906 struct request, rb_node))) {
1907 struct bfq_queue *bfqq = bfq_init_rq(req);
1908 struct bfq_data *bfqd;
1909 struct request *prev, *next_rq;
1916 /* Reposition request in its sort_list */
1917 elv_rb_del(&bfqq->sort_list, req);
1918 elv_rb_add(&bfqq->sort_list, req);
1920 /* Choose next request to be served for bfqq */
1921 prev = bfqq->next_rq;
1922 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
1923 bfqd->last_position);
1924 bfqq->next_rq = next_rq;
1926 * If next_rq changes, update both the queue's budget to
1927 * fit the new request and the queue's position in its
1930 if (prev != bfqq->next_rq) {
1931 bfq_updated_next_req(bfqd, bfqq);
1932 bfq_pos_tree_add_move(bfqd, bfqq);
1938 * This function is called to notify the scheduler that the requests
1939 * rq and 'next' have been merged, with 'next' going away. BFQ
1940 * exploits this hook to address the following issue: if 'next' has a
1941 * fifo_time lower that rq, then the fifo_time of rq must be set to
1942 * the value of 'next', to not forget the greater age of 'next'.
1944 * NOTE: in this function we assume that rq is in a bfq_queue, basing
1945 * on that rq is picked from the hash table q->elevator->hash, which,
1946 * in its turn, is filled only with I/O requests present in
1947 * bfq_queues, while BFQ is in use for the request queue q. In fact,
1948 * the function that fills this hash table (elv_rqhash_add) is called
1949 * only by bfq_insert_request.
1951 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
1952 struct request *next)
1954 struct bfq_queue *bfqq = bfq_init_rq(rq),
1955 *next_bfqq = bfq_init_rq(next);
1961 * If next and rq belong to the same bfq_queue and next is older
1962 * than rq, then reposition rq in the fifo (by substituting next
1963 * with rq). Otherwise, if next and rq belong to different
1964 * bfq_queues, never reposition rq: in fact, we would have to
1965 * reposition it with respect to next's position in its own fifo,
1966 * which would most certainly be too expensive with respect to
1969 if (bfqq == next_bfqq &&
1970 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1971 next->fifo_time < rq->fifo_time) {
1972 list_del_init(&rq->queuelist);
1973 list_replace_init(&next->queuelist, &rq->queuelist);
1974 rq->fifo_time = next->fifo_time;
1977 if (bfqq->next_rq == next)
1980 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
1983 /* Must be called with bfqq != NULL */
1984 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
1986 if (bfq_bfqq_busy(bfqq))
1987 bfqq->bfqd->wr_busy_queues--;
1989 bfqq->wr_cur_max_time = 0;
1990 bfqq->last_wr_start_finish = jiffies;
1992 * Trigger a weight change on the next invocation of
1993 * __bfq_entity_update_weight_prio.
1995 bfqq->entity.prio_changed = 1;
1998 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
1999 struct bfq_group *bfqg)
2003 for (i = 0; i < 2; i++)
2004 for (j = 0; j < IOPRIO_BE_NR; j++)
2005 if (bfqg->async_bfqq[i][j])
2006 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2007 if (bfqg->async_idle_bfqq)
2008 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2011 static void bfq_end_wr(struct bfq_data *bfqd)
2013 struct bfq_queue *bfqq;
2015 spin_lock_irq(&bfqd->lock);
2017 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2018 bfq_bfqq_end_wr(bfqq);
2019 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2020 bfq_bfqq_end_wr(bfqq);
2021 bfq_end_wr_async(bfqd);
2023 spin_unlock_irq(&bfqd->lock);
2026 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2029 return blk_rq_pos(io_struct);
2031 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2034 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2037 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2041 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2042 struct bfq_queue *bfqq,
2045 struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
2046 struct rb_node *parent, *node;
2047 struct bfq_queue *__bfqq;
2049 if (RB_EMPTY_ROOT(root))
2053 * First, if we find a request starting at the end of the last
2054 * request, choose it.
2056 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2061 * If the exact sector wasn't found, the parent of the NULL leaf
2062 * will contain the closest sector (rq_pos_tree sorted by
2063 * next_request position).
2065 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2066 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2069 if (blk_rq_pos(__bfqq->next_rq) < sector)
2070 node = rb_next(&__bfqq->pos_node);
2072 node = rb_prev(&__bfqq->pos_node);
2076 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2077 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2083 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2084 struct bfq_queue *cur_bfqq,
2087 struct bfq_queue *bfqq;
2090 * We shall notice if some of the queues are cooperating,
2091 * e.g., working closely on the same area of the device. In
2092 * that case, we can group them together and: 1) don't waste
2093 * time idling, and 2) serve the union of their requests in
2094 * the best possible order for throughput.
2096 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2097 if (!bfqq || bfqq == cur_bfqq)
2103 static struct bfq_queue *
2104 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2106 int process_refs, new_process_refs;
2107 struct bfq_queue *__bfqq;
2110 * If there are no process references on the new_bfqq, then it is
2111 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2112 * may have dropped their last reference (not just their last process
2115 if (!bfqq_process_refs(new_bfqq))
2118 /* Avoid a circular list and skip interim queue merges. */
2119 while ((__bfqq = new_bfqq->new_bfqq)) {
2125 process_refs = bfqq_process_refs(bfqq);
2126 new_process_refs = bfqq_process_refs(new_bfqq);
2128 * If the process for the bfqq has gone away, there is no
2129 * sense in merging the queues.
2131 if (process_refs == 0 || new_process_refs == 0)
2134 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2138 * Merging is just a redirection: the requests of the process
2139 * owning one of the two queues are redirected to the other queue.
2140 * The latter queue, in its turn, is set as shared if this is the
2141 * first time that the requests of some process are redirected to
2144 * We redirect bfqq to new_bfqq and not the opposite, because
2145 * we are in the context of the process owning bfqq, thus we
2146 * have the io_cq of this process. So we can immediately
2147 * configure this io_cq to redirect the requests of the
2148 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2149 * not available any more (new_bfqq->bic == NULL).
2151 * Anyway, even in case new_bfqq coincides with the in-service
2152 * queue, redirecting requests the in-service queue is the
2153 * best option, as we feed the in-service queue with new
2154 * requests close to the last request served and, by doing so,
2155 * are likely to increase the throughput.
2157 bfqq->new_bfqq = new_bfqq;
2159 * The above assignment schedules the following redirections:
2160 * each time some I/O for bfqq arrives, the process that
2161 * generated that I/O is disassociated from bfqq and
2162 * associated with new_bfqq. Here we increases new_bfqq->ref
2163 * in advance, adding the number of processes that are
2164 * expected to be associated with new_bfqq as they happen to
2167 new_bfqq->ref += process_refs;
2171 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2172 struct bfq_queue *new_bfqq)
2174 if (bfq_too_late_for_merging(new_bfqq))
2177 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2178 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2182 * If either of the queues has already been detected as seeky,
2183 * then merging it with the other queue is unlikely to lead to
2186 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2190 * Interleaved I/O is known to be done by (some) applications
2191 * only for reads, so it does not make sense to merge async
2194 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2201 * Attempt to schedule a merge of bfqq with the currently in-service
2202 * queue or with a close queue among the scheduled queues. Return
2203 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2204 * structure otherwise.
2206 * The OOM queue is not allowed to participate to cooperation: in fact, since
2207 * the requests temporarily redirected to the OOM queue could be redirected
2208 * again to dedicated queues at any time, the state needed to correctly
2209 * handle merging with the OOM queue would be quite complex and expensive
2210 * to maintain. Besides, in such a critical condition as an out of memory,
2211 * the benefits of queue merging may be little relevant, or even negligible.
2213 * WARNING: queue merging may impair fairness among non-weight raised
2214 * queues, for at least two reasons: 1) the original weight of a
2215 * merged queue may change during the merged state, 2) even being the
2216 * weight the same, a merged queue may be bloated with many more
2217 * requests than the ones produced by its originally-associated
2220 static struct bfq_queue *
2221 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2222 void *io_struct, bool request)
2224 struct bfq_queue *in_service_bfqq, *new_bfqq;
2226 /* if a merge has already been setup, then proceed with that first */
2228 return bfqq->new_bfqq;
2231 * Prevent bfqq from being merged if it has been created too
2232 * long ago. The idea is that true cooperating processes, and
2233 * thus their associated bfq_queues, are supposed to be
2234 * created shortly after each other. This is the case, e.g.,
2235 * for KVM/QEMU and dump I/O threads. Basing on this
2236 * assumption, the following filtering greatly reduces the
2237 * probability that two non-cooperating processes, which just
2238 * happen to do close I/O for some short time interval, have
2239 * their queues merged by mistake.
2241 if (bfq_too_late_for_merging(bfqq))
2244 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2247 /* If there is only one backlogged queue, don't search. */
2248 if (bfqd->busy_queues == 1)
2251 in_service_bfqq = bfqd->in_service_queue;
2253 if (in_service_bfqq && in_service_bfqq != bfqq &&
2254 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2255 bfq_rq_close_to_sector(io_struct, request,
2256 bfqd->in_serv_last_pos) &&
2257 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2258 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2259 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2264 * Check whether there is a cooperator among currently scheduled
2265 * queues. The only thing we need is that the bio/request is not
2266 * NULL, as we need it to establish whether a cooperator exists.
2268 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2269 bfq_io_struct_pos(io_struct, request));
2271 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2272 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2273 return bfq_setup_merge(bfqq, new_bfqq);
2278 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2280 struct bfq_io_cq *bic = bfqq->bic;
2283 * If !bfqq->bic, the queue is already shared or its requests
2284 * have already been redirected to a shared queue; both idle window
2285 * and weight raising state have already been saved. Do nothing.
2290 bic->saved_ttime = bfqq->ttime;
2291 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2292 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2293 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2294 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2295 if (unlikely(bfq_bfqq_just_created(bfqq) &&
2296 !bfq_bfqq_in_large_burst(bfqq) &&
2297 bfqq->bfqd->low_latency)) {
2299 * bfqq being merged right after being created: bfqq
2300 * would have deserved interactive weight raising, but
2301 * did not make it to be set in a weight-raised state,
2302 * because of this early merge. Store directly the
2303 * weight-raising state that would have been assigned
2304 * to bfqq, so that to avoid that bfqq unjustly fails
2305 * to enjoy weight raising if split soon.
2307 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
2308 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2309 bic->saved_last_wr_start_finish = jiffies;
2311 bic->saved_wr_coeff = bfqq->wr_coeff;
2312 bic->saved_wr_start_at_switch_to_srt =
2313 bfqq->wr_start_at_switch_to_srt;
2314 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2315 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2320 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2321 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2323 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2324 (unsigned long)new_bfqq->pid);
2325 /* Save weight raising and idle window of the merged queues */
2326 bfq_bfqq_save_state(bfqq);
2327 bfq_bfqq_save_state(new_bfqq);
2328 if (bfq_bfqq_IO_bound(bfqq))
2329 bfq_mark_bfqq_IO_bound(new_bfqq);
2330 bfq_clear_bfqq_IO_bound(bfqq);
2333 * If bfqq is weight-raised, then let new_bfqq inherit
2334 * weight-raising. To reduce false positives, neglect the case
2335 * where bfqq has just been created, but has not yet made it
2336 * to be weight-raised (which may happen because EQM may merge
2337 * bfqq even before bfq_add_request is executed for the first
2338 * time for bfqq). Handling this case would however be very
2339 * easy, thanks to the flag just_created.
2341 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2342 new_bfqq->wr_coeff = bfqq->wr_coeff;
2343 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2344 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2345 new_bfqq->wr_start_at_switch_to_srt =
2346 bfqq->wr_start_at_switch_to_srt;
2347 if (bfq_bfqq_busy(new_bfqq))
2348 bfqd->wr_busy_queues++;
2349 new_bfqq->entity.prio_changed = 1;
2352 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2354 bfqq->entity.prio_changed = 1;
2355 if (bfq_bfqq_busy(bfqq))
2356 bfqd->wr_busy_queues--;
2359 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2360 bfqd->wr_busy_queues);
2363 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2365 bic_set_bfqq(bic, new_bfqq, 1);
2366 bfq_mark_bfqq_coop(new_bfqq);
2368 * new_bfqq now belongs to at least two bics (it is a shared queue):
2369 * set new_bfqq->bic to NULL. bfqq either:
2370 * - does not belong to any bic any more, and hence bfqq->bic must
2371 * be set to NULL, or
2372 * - is a queue whose owning bics have already been redirected to a
2373 * different queue, hence the queue is destined to not belong to
2374 * any bic soon and bfqq->bic is already NULL (therefore the next
2375 * assignment causes no harm).
2377 new_bfqq->bic = NULL;
2379 /* release process reference to bfqq */
2380 bfq_put_queue(bfqq);
2383 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2386 struct bfq_data *bfqd = q->elevator->elevator_data;
2387 bool is_sync = op_is_sync(bio->bi_opf);
2388 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2391 * Disallow merge of a sync bio into an async request.
2393 if (is_sync && !rq_is_sync(rq))
2397 * Lookup the bfqq that this bio will be queued with. Allow
2398 * merge only if rq is queued there.
2404 * We take advantage of this function to perform an early merge
2405 * of the queues of possible cooperating processes.
2407 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2410 * bic still points to bfqq, then it has not yet been
2411 * redirected to some other bfq_queue, and a queue
2412 * merge beween bfqq and new_bfqq can be safely
2413 * fulfillled, i.e., bic can be redirected to new_bfqq
2414 * and bfqq can be put.
2416 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2419 * If we get here, bio will be queued into new_queue,
2420 * so use new_bfqq to decide whether bio and rq can be
2426 * Change also bqfd->bio_bfqq, as
2427 * bfqd->bio_bic now points to new_bfqq, and
2428 * this function may be invoked again (and then may
2429 * use again bqfd->bio_bfqq).
2431 bfqd->bio_bfqq = bfqq;
2434 return bfqq == RQ_BFQQ(rq);
2438 * Set the maximum time for the in-service queue to consume its
2439 * budget. This prevents seeky processes from lowering the throughput.
2440 * In practice, a time-slice service scheme is used with seeky
2443 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2444 struct bfq_queue *bfqq)
2446 unsigned int timeout_coeff;
2448 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2451 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2453 bfqd->last_budget_start = ktime_get();
2455 bfqq->budget_timeout = jiffies +
2456 bfqd->bfq_timeout * timeout_coeff;
2459 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2460 struct bfq_queue *bfqq)
2463 bfq_clear_bfqq_fifo_expire(bfqq);
2465 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2467 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2468 bfqq->wr_coeff > 1 &&
2469 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2470 time_is_before_jiffies(bfqq->budget_timeout)) {
2472 * For soft real-time queues, move the start
2473 * of the weight-raising period forward by the
2474 * time the queue has not received any
2475 * service. Otherwise, a relatively long
2476 * service delay is likely to cause the
2477 * weight-raising period of the queue to end,
2478 * because of the short duration of the
2479 * weight-raising period of a soft real-time
2480 * queue. It is worth noting that this move
2481 * is not so dangerous for the other queues,
2482 * because soft real-time queues are not
2485 * To not add a further variable, we use the
2486 * overloaded field budget_timeout to
2487 * determine for how long the queue has not
2488 * received service, i.e., how much time has
2489 * elapsed since the queue expired. However,
2490 * this is a little imprecise, because
2491 * budget_timeout is set to jiffies if bfqq
2492 * not only expires, but also remains with no
2495 if (time_after(bfqq->budget_timeout,
2496 bfqq->last_wr_start_finish))
2497 bfqq->last_wr_start_finish +=
2498 jiffies - bfqq->budget_timeout;
2500 bfqq->last_wr_start_finish = jiffies;
2503 bfq_set_budget_timeout(bfqd, bfqq);
2504 bfq_log_bfqq(bfqd, bfqq,
2505 "set_in_service_queue, cur-budget = %d",
2506 bfqq->entity.budget);
2509 bfqd->in_service_queue = bfqq;
2510 bfqd->in_serv_last_pos = 0;
2514 * Get and set a new queue for service.
2516 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2518 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2520 __bfq_set_in_service_queue(bfqd, bfqq);
2524 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2526 struct bfq_queue *bfqq = bfqd->in_service_queue;
2529 bfq_mark_bfqq_wait_request(bfqq);
2532 * We don't want to idle for seeks, but we do want to allow
2533 * fair distribution of slice time for a process doing back-to-back
2534 * seeks. So allow a little bit of time for him to submit a new rq.
2536 sl = bfqd->bfq_slice_idle;
2538 * Unless the queue is being weight-raised or the scenario is
2539 * asymmetric, grant only minimum idle time if the queue
2540 * is seeky. A long idling is preserved for a weight-raised
2541 * queue, or, more in general, in an asymmetric scenario,
2542 * because a long idling is needed for guaranteeing to a queue
2543 * its reserved share of the throughput (in particular, it is
2544 * needed if the queue has a higher weight than some other
2547 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
2548 bfq_symmetric_scenario(bfqd))
2549 sl = min_t(u64, sl, BFQ_MIN_TT);
2550 else if (bfqq->wr_coeff > 1)
2551 sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
2553 bfqd->last_idling_start = ktime_get();
2554 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
2556 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
2560 * In autotuning mode, max_budget is dynamically recomputed as the
2561 * amount of sectors transferred in timeout at the estimated peak
2562 * rate. This enables BFQ to utilize a full timeslice with a full
2563 * budget, even if the in-service queue is served at peak rate. And
2564 * this maximises throughput with sequential workloads.
2566 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
2568 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
2569 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
2573 * Update parameters related to throughput and responsiveness, as a
2574 * function of the estimated peak rate. See comments on
2575 * bfq_calc_max_budget(), and on the ref_wr_duration array.
2577 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
2579 if (bfqd->bfq_user_max_budget == 0) {
2580 bfqd->bfq_max_budget =
2581 bfq_calc_max_budget(bfqd);
2582 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
2586 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
2589 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
2590 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
2591 bfqd->peak_rate_samples = 1;
2592 bfqd->sequential_samples = 0;
2593 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
2595 } else /* no new rq dispatched, just reset the number of samples */
2596 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
2599 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
2600 bfqd->peak_rate_samples, bfqd->sequential_samples,
2601 bfqd->tot_sectors_dispatched);
2604 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
2606 u32 rate, weight, divisor;
2609 * For the convergence property to hold (see comments on
2610 * bfq_update_peak_rate()) and for the assessment to be
2611 * reliable, a minimum number of samples must be present, and
2612 * a minimum amount of time must have elapsed. If not so, do
2613 * not compute new rate. Just reset parameters, to get ready
2614 * for a new evaluation attempt.
2616 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
2617 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
2618 goto reset_computation;
2621 * If a new request completion has occurred after last
2622 * dispatch, then, to approximate the rate at which requests
2623 * have been served by the device, it is more precise to
2624 * extend the observation interval to the last completion.
2626 bfqd->delta_from_first =
2627 max_t(u64, bfqd->delta_from_first,
2628 bfqd->last_completion - bfqd->first_dispatch);
2631 * Rate computed in sects/usec, and not sects/nsec, for
2634 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
2635 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
2638 * Peak rate not updated if:
2639 * - the percentage of sequential dispatches is below 3/4 of the
2640 * total, and rate is below the current estimated peak rate
2641 * - rate is unreasonably high (> 20M sectors/sec)
2643 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
2644 rate <= bfqd->peak_rate) ||
2645 rate > 20<<BFQ_RATE_SHIFT)
2646 goto reset_computation;
2649 * We have to update the peak rate, at last! To this purpose,
2650 * we use a low-pass filter. We compute the smoothing constant
2651 * of the filter as a function of the 'weight' of the new
2654 * As can be seen in next formulas, we define this weight as a
2655 * quantity proportional to how sequential the workload is,
2656 * and to how long the observation time interval is.
2658 * The weight runs from 0 to 8. The maximum value of the
2659 * weight, 8, yields the minimum value for the smoothing
2660 * constant. At this minimum value for the smoothing constant,
2661 * the measured rate contributes for half of the next value of
2662 * the estimated peak rate.
2664 * So, the first step is to compute the weight as a function
2665 * of how sequential the workload is. Note that the weight
2666 * cannot reach 9, because bfqd->sequential_samples cannot
2667 * become equal to bfqd->peak_rate_samples, which, in its
2668 * turn, holds true because bfqd->sequential_samples is not
2669 * incremented for the first sample.
2671 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
2674 * Second step: further refine the weight as a function of the
2675 * duration of the observation interval.
2677 weight = min_t(u32, 8,
2678 div_u64(weight * bfqd->delta_from_first,
2679 BFQ_RATE_REF_INTERVAL));
2682 * Divisor ranging from 10, for minimum weight, to 2, for
2685 divisor = 10 - weight;
2688 * Finally, update peak rate:
2690 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
2692 bfqd->peak_rate *= divisor-1;
2693 bfqd->peak_rate /= divisor;
2694 rate /= divisor; /* smoothing constant alpha = 1/divisor */
2696 bfqd->peak_rate += rate;
2699 * For a very slow device, bfqd->peak_rate can reach 0 (see
2700 * the minimum representable values reported in the comments
2701 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
2702 * divisions by zero where bfqd->peak_rate is used as a
2705 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
2707 update_thr_responsiveness_params(bfqd);
2710 bfq_reset_rate_computation(bfqd, rq);
2714 * Update the read/write peak rate (the main quantity used for
2715 * auto-tuning, see update_thr_responsiveness_params()).
2717 * It is not trivial to estimate the peak rate (correctly): because of
2718 * the presence of sw and hw queues between the scheduler and the
2719 * device components that finally serve I/O requests, it is hard to
2720 * say exactly when a given dispatched request is served inside the
2721 * device, and for how long. As a consequence, it is hard to know
2722 * precisely at what rate a given set of requests is actually served
2725 * On the opposite end, the dispatch time of any request is trivially
2726 * available, and, from this piece of information, the "dispatch rate"
2727 * of requests can be immediately computed. So, the idea in the next
2728 * function is to use what is known, namely request dispatch times
2729 * (plus, when useful, request completion times), to estimate what is
2730 * unknown, namely in-device request service rate.
2732 * The main issue is that, because of the above facts, the rate at
2733 * which a certain set of requests is dispatched over a certain time
2734 * interval can vary greatly with respect to the rate at which the
2735 * same requests are then served. But, since the size of any
2736 * intermediate queue is limited, and the service scheme is lossless
2737 * (no request is silently dropped), the following obvious convergence
2738 * property holds: the number of requests dispatched MUST become
2739 * closer and closer to the number of requests completed as the
2740 * observation interval grows. This is the key property used in
2741 * the next function to estimate the peak service rate as a function
2742 * of the observed dispatch rate. The function assumes to be invoked
2743 * on every request dispatch.
2745 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
2747 u64 now_ns = ktime_get_ns();
2749 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
2750 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
2751 bfqd->peak_rate_samples);
2752 bfq_reset_rate_computation(bfqd, rq);
2753 goto update_last_values; /* will add one sample */
2757 * Device idle for very long: the observation interval lasting
2758 * up to this dispatch cannot be a valid observation interval
2759 * for computing a new peak rate (similarly to the late-
2760 * completion event in bfq_completed_request()). Go to
2761 * update_rate_and_reset to have the following three steps
2763 * - close the observation interval at the last (previous)
2764 * request dispatch or completion
2765 * - compute rate, if possible, for that observation interval
2766 * - start a new observation interval with this dispatch
2768 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
2769 bfqd->rq_in_driver == 0)
2770 goto update_rate_and_reset;
2772 /* Update sampling information */
2773 bfqd->peak_rate_samples++;
2775 if ((bfqd->rq_in_driver > 0 ||
2776 now_ns - bfqd->last_completion < BFQ_MIN_TT)
2777 && get_sdist(bfqd->last_position, rq) < BFQQ_SEEK_THR)
2778 bfqd->sequential_samples++;
2780 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
2782 /* Reset max observed rq size every 32 dispatches */
2783 if (likely(bfqd->peak_rate_samples % 32))
2784 bfqd->last_rq_max_size =
2785 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
2787 bfqd->last_rq_max_size = blk_rq_sectors(rq);
2789 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
2791 /* Target observation interval not yet reached, go on sampling */
2792 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
2793 goto update_last_values;
2795 update_rate_and_reset:
2796 bfq_update_rate_reset(bfqd, rq);
2798 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2799 if (RQ_BFQQ(rq) == bfqd->in_service_queue)
2800 bfqd->in_serv_last_pos = bfqd->last_position;
2801 bfqd->last_dispatch = now_ns;
2805 * Remove request from internal lists.
2807 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
2809 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2812 * For consistency, the next instruction should have been
2813 * executed after removing the request from the queue and
2814 * dispatching it. We execute instead this instruction before
2815 * bfq_remove_request() (and hence introduce a temporary
2816 * inconsistency), for efficiency. In fact, should this
2817 * dispatch occur for a non in-service bfqq, this anticipated
2818 * increment prevents two counters related to bfqq->dispatched
2819 * from risking to be, first, uselessly decremented, and then
2820 * incremented again when the (new) value of bfqq->dispatched
2821 * happens to be taken into account.
2824 bfq_update_peak_rate(q->elevator->elevator_data, rq);
2826 bfq_remove_request(q, rq);
2829 static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq)
2832 * If this bfqq is shared between multiple processes, check
2833 * to make sure that those processes are still issuing I/Os
2834 * within the mean seek distance. If not, it may be time to
2835 * break the queues apart again.
2837 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
2838 bfq_mark_bfqq_split_coop(bfqq);
2840 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2841 if (bfqq->dispatched == 0)
2843 * Overloading budget_timeout field to store
2844 * the time at which the queue remains with no
2845 * backlog and no outstanding request; used by
2846 * the weight-raising mechanism.
2848 bfqq->budget_timeout = jiffies;
2850 bfq_del_bfqq_busy(bfqd, bfqq, true);
2852 bfq_requeue_bfqq(bfqd, bfqq, true);
2854 * Resort priority tree of potential close cooperators.
2856 bfq_pos_tree_add_move(bfqd, bfqq);
2860 * All in-service entities must have been properly deactivated
2861 * or requeued before executing the next function, which
2862 * resets all in-service entities as no more in service. This
2863 * may cause bfqq to be freed. If this happens, the next
2864 * function returns true.
2866 return __bfq_bfqd_reset_in_service(bfqd);
2870 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
2871 * @bfqd: device data.
2872 * @bfqq: queue to update.
2873 * @reason: reason for expiration.
2875 * Handle the feedback on @bfqq budget at queue expiration.
2876 * See the body for detailed comments.
2878 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
2879 struct bfq_queue *bfqq,
2880 enum bfqq_expiration reason)
2882 struct request *next_rq;
2883 int budget, min_budget;
2885 min_budget = bfq_min_budget(bfqd);
2887 if (bfqq->wr_coeff == 1)
2888 budget = bfqq->max_budget;
2890 * Use a constant, low budget for weight-raised queues,
2891 * to help achieve a low latency. Keep it slightly higher
2892 * than the minimum possible budget, to cause a little
2893 * bit fewer expirations.
2895 budget = 2 * min_budget;
2897 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
2898 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
2899 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
2900 budget, bfq_min_budget(bfqd));
2901 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
2902 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
2904 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
2907 * Caveat: in all the following cases we trade latency
2910 case BFQQE_TOO_IDLE:
2912 * This is the only case where we may reduce
2913 * the budget: if there is no request of the
2914 * process still waiting for completion, then
2915 * we assume (tentatively) that the timer has
2916 * expired because the batch of requests of
2917 * the process could have been served with a
2918 * smaller budget. Hence, betting that
2919 * process will behave in the same way when it
2920 * becomes backlogged again, we reduce its
2921 * next budget. As long as we guess right,
2922 * this budget cut reduces the latency
2923 * experienced by the process.
2925 * However, if there are still outstanding
2926 * requests, then the process may have not yet
2927 * issued its next request just because it is
2928 * still waiting for the completion of some of
2929 * the still outstanding ones. So in this
2930 * subcase we do not reduce its budget, on the
2931 * contrary we increase it to possibly boost
2932 * the throughput, as discussed in the
2933 * comments to the BUDGET_TIMEOUT case.
2935 if (bfqq->dispatched > 0) /* still outstanding reqs */
2936 budget = min(budget * 2, bfqd->bfq_max_budget);
2938 if (budget > 5 * min_budget)
2939 budget -= 4 * min_budget;
2941 budget = min_budget;
2944 case BFQQE_BUDGET_TIMEOUT:
2946 * We double the budget here because it gives
2947 * the chance to boost the throughput if this
2948 * is not a seeky process (and has bumped into
2949 * this timeout because of, e.g., ZBR).
2951 budget = min(budget * 2, bfqd->bfq_max_budget);
2953 case BFQQE_BUDGET_EXHAUSTED:
2955 * The process still has backlog, and did not
2956 * let either the budget timeout or the disk
2957 * idling timeout expire. Hence it is not
2958 * seeky, has a short thinktime and may be
2959 * happy with a higher budget too. So
2960 * definitely increase the budget of this good
2961 * candidate to boost the disk throughput.
2963 budget = min(budget * 4, bfqd->bfq_max_budget);
2965 case BFQQE_NO_MORE_REQUESTS:
2967 * For queues that expire for this reason, it
2968 * is particularly important to keep the
2969 * budget close to the actual service they
2970 * need. Doing so reduces the timestamp
2971 * misalignment problem described in the
2972 * comments in the body of
2973 * __bfq_activate_entity. In fact, suppose
2974 * that a queue systematically expires for
2975 * BFQQE_NO_MORE_REQUESTS and presents a
2976 * new request in time to enjoy timestamp
2977 * back-shifting. The larger the budget of the
2978 * queue is with respect to the service the
2979 * queue actually requests in each service
2980 * slot, the more times the queue can be
2981 * reactivated with the same virtual finish
2982 * time. It follows that, even if this finish
2983 * time is pushed to the system virtual time
2984 * to reduce the consequent timestamp
2985 * misalignment, the queue unjustly enjoys for
2986 * many re-activations a lower finish time
2987 * than all newly activated queues.
2989 * The service needed by bfqq is measured
2990 * quite precisely by bfqq->entity.service.
2991 * Since bfqq does not enjoy device idling,
2992 * bfqq->entity.service is equal to the number
2993 * of sectors that the process associated with
2994 * bfqq requested to read/write before waiting
2995 * for request completions, or blocking for
2998 budget = max_t(int, bfqq->entity.service, min_budget);
3003 } else if (!bfq_bfqq_sync(bfqq)) {
3005 * Async queues get always the maximum possible
3006 * budget, as for them we do not care about latency
3007 * (in addition, their ability to dispatch is limited
3008 * by the charging factor).
3010 budget = bfqd->bfq_max_budget;
3013 bfqq->max_budget = budget;
3015 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
3016 !bfqd->bfq_user_max_budget)
3017 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
3020 * If there is still backlog, then assign a new budget, making
3021 * sure that it is large enough for the next request. Since
3022 * the finish time of bfqq must be kept in sync with the
3023 * budget, be sure to call __bfq_bfqq_expire() *after* this
3026 * If there is no backlog, then no need to update the budget;
3027 * it will be updated on the arrival of a new request.
3029 next_rq = bfqq->next_rq;
3031 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
3032 bfq_serv_to_charge(next_rq, bfqq));
3034 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
3035 next_rq ? blk_rq_sectors(next_rq) : 0,
3036 bfqq->entity.budget);
3040 * Return true if the process associated with bfqq is "slow". The slow
3041 * flag is used, in addition to the budget timeout, to reduce the
3042 * amount of service provided to seeky processes, and thus reduce
3043 * their chances to lower the throughput. More details in the comments
3044 * on the function bfq_bfqq_expire().
3046 * An important observation is in order: as discussed in the comments
3047 * on the function bfq_update_peak_rate(), with devices with internal
3048 * queues, it is hard if ever possible to know when and for how long
3049 * an I/O request is processed by the device (apart from the trivial
3050 * I/O pattern where a new request is dispatched only after the
3051 * previous one has been completed). This makes it hard to evaluate
3052 * the real rate at which the I/O requests of each bfq_queue are
3053 * served. In fact, for an I/O scheduler like BFQ, serving a
3054 * bfq_queue means just dispatching its requests during its service
3055 * slot (i.e., until the budget of the queue is exhausted, or the
3056 * queue remains idle, or, finally, a timeout fires). But, during the
3057 * service slot of a bfq_queue, around 100 ms at most, the device may
3058 * be even still processing requests of bfq_queues served in previous
3059 * service slots. On the opposite end, the requests of the in-service
3060 * bfq_queue may be completed after the service slot of the queue
3063 * Anyway, unless more sophisticated solutions are used
3064 * (where possible), the sum of the sizes of the requests dispatched
3065 * during the service slot of a bfq_queue is probably the only
3066 * approximation available for the service received by the bfq_queue
3067 * during its service slot. And this sum is the quantity used in this
3068 * function to evaluate the I/O speed of a process.
3070 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3071 bool compensate, enum bfqq_expiration reason,
3072 unsigned long *delta_ms)
3074 ktime_t delta_ktime;
3076 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
3078 if (!bfq_bfqq_sync(bfqq))
3082 delta_ktime = bfqd->last_idling_start;
3084 delta_ktime = ktime_get();
3085 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
3086 delta_usecs = ktime_to_us(delta_ktime);
3088 /* don't use too short time intervals */
3089 if (delta_usecs < 1000) {
3090 if (blk_queue_nonrot(bfqd->queue))
3092 * give same worst-case guarantees as idling
3095 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
3096 else /* charge at least one seek */
3097 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
3102 *delta_ms = delta_usecs / USEC_PER_MSEC;
3105 * Use only long (> 20ms) intervals to filter out excessive
3106 * spikes in service rate estimation.
3108 if (delta_usecs > 20000) {
3110 * Caveat for rotational devices: processes doing I/O
3111 * in the slower disk zones tend to be slow(er) even
3112 * if not seeky. In this respect, the estimated peak
3113 * rate is likely to be an average over the disk
3114 * surface. Accordingly, to not be too harsh with
3115 * unlucky processes, a process is deemed slow only if
3116 * its rate has been lower than half of the estimated
3119 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
3122 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
3128 * To be deemed as soft real-time, an application must meet two
3129 * requirements. First, the application must not require an average
3130 * bandwidth higher than the approximate bandwidth required to playback or
3131 * record a compressed high-definition video.
3132 * The next function is invoked on the completion of the last request of a
3133 * batch, to compute the next-start time instant, soft_rt_next_start, such
3134 * that, if the next request of the application does not arrive before
3135 * soft_rt_next_start, then the above requirement on the bandwidth is met.
3137 * The second requirement is that the request pattern of the application is
3138 * isochronous, i.e., that, after issuing a request or a batch of requests,
3139 * the application stops issuing new requests until all its pending requests
3140 * have been completed. After that, the application may issue a new batch,
3142 * For this reason the next function is invoked to compute
3143 * soft_rt_next_start only for applications that meet this requirement,
3144 * whereas soft_rt_next_start is set to infinity for applications that do
3147 * Unfortunately, even a greedy (i.e., I/O-bound) application may
3148 * happen to meet, occasionally or systematically, both the above
3149 * bandwidth and isochrony requirements. This may happen at least in
3150 * the following circumstances. First, if the CPU load is high. The
3151 * application may stop issuing requests while the CPUs are busy
3152 * serving other processes, then restart, then stop again for a while,
3153 * and so on. The other circumstances are related to the storage
3154 * device: the storage device is highly loaded or reaches a low-enough
3155 * throughput with the I/O of the application (e.g., because the I/O
3156 * is random and/or the device is slow). In all these cases, the
3157 * I/O of the application may be simply slowed down enough to meet
3158 * the bandwidth and isochrony requirements. To reduce the probability
3159 * that greedy applications are deemed as soft real-time in these
3160 * corner cases, a further rule is used in the computation of
3161 * soft_rt_next_start: the return value of this function is forced to
3162 * be higher than the maximum between the following two quantities.
3164 * (a) Current time plus: (1) the maximum time for which the arrival
3165 * of a request is waited for when a sync queue becomes idle,
3166 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
3167 * postpone for a moment the reason for adding a few extra
3168 * jiffies; we get back to it after next item (b). Lower-bounding
3169 * the return value of this function with the current time plus
3170 * bfqd->bfq_slice_idle tends to filter out greedy applications,
3171 * because the latter issue their next request as soon as possible
3172 * after the last one has been completed. In contrast, a soft
3173 * real-time application spends some time processing data, after a
3174 * batch of its requests has been completed.
3176 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
3177 * above, greedy applications may happen to meet both the
3178 * bandwidth and isochrony requirements under heavy CPU or
3179 * storage-device load. In more detail, in these scenarios, these
3180 * applications happen, only for limited time periods, to do I/O
3181 * slowly enough to meet all the requirements described so far,
3182 * including the filtering in above item (a). These slow-speed
3183 * time intervals are usually interspersed between other time
3184 * intervals during which these applications do I/O at a very high
3185 * speed. Fortunately, exactly because of the high speed of the
3186 * I/O in the high-speed intervals, the values returned by this
3187 * function happen to be so high, near the end of any such
3188 * high-speed interval, to be likely to fall *after* the end of
3189 * the low-speed time interval that follows. These high values are
3190 * stored in bfqq->soft_rt_next_start after each invocation of
3191 * this function. As a consequence, if the last value of
3192 * bfqq->soft_rt_next_start is constantly used to lower-bound the
3193 * next value that this function may return, then, from the very
3194 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
3195 * likely to be constantly kept so high that any I/O request
3196 * issued during the low-speed interval is considered as arriving
3197 * to soon for the application to be deemed as soft
3198 * real-time. Then, in the high-speed interval that follows, the
3199 * application will not be deemed as soft real-time, just because
3200 * it will do I/O at a high speed. And so on.
3202 * Getting back to the filtering in item (a), in the following two
3203 * cases this filtering might be easily passed by a greedy
3204 * application, if the reference quantity was just
3205 * bfqd->bfq_slice_idle:
3206 * 1) HZ is so low that the duration of a jiffy is comparable to or
3207 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
3208 * devices with HZ=100. The time granularity may be so coarse
3209 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
3210 * is rather lower than the exact value.
3211 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
3212 * for a while, then suddenly 'jump' by several units to recover the lost
3213 * increments. This seems to happen, e.g., inside virtual machines.
3214 * To address this issue, in the filtering in (a) we do not use as a
3215 * reference time interval just bfqd->bfq_slice_idle, but
3216 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
3217 * minimum number of jiffies for which the filter seems to be quite
3218 * precise also in embedded systems and KVM/QEMU virtual machines.
3220 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
3221 struct bfq_queue *bfqq)
3223 return max3(bfqq->soft_rt_next_start,
3224 bfqq->last_idle_bklogged +
3225 HZ * bfqq->service_from_backlogged /
3226 bfqd->bfq_wr_max_softrt_rate,
3227 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
3230 static bool bfq_bfqq_injectable(struct bfq_queue *bfqq)
3232 return BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3233 blk_queue_nonrot(bfqq->bfqd->queue) &&
3238 * bfq_bfqq_expire - expire a queue.
3239 * @bfqd: device owning the queue.
3240 * @bfqq: the queue to expire.
3241 * @compensate: if true, compensate for the time spent idling.
3242 * @reason: the reason causing the expiration.
3244 * If the process associated with bfqq does slow I/O (e.g., because it
3245 * issues random requests), we charge bfqq with the time it has been
3246 * in service instead of the service it has received (see
3247 * bfq_bfqq_charge_time for details on how this goal is achieved). As
3248 * a consequence, bfqq will typically get higher timestamps upon
3249 * reactivation, and hence it will be rescheduled as if it had
3250 * received more service than what it has actually received. In the
3251 * end, bfqq receives less service in proportion to how slowly its
3252 * associated process consumes its budgets (and hence how seriously it
3253 * tends to lower the throughput). In addition, this time-charging
3254 * strategy guarantees time fairness among slow processes. In
3255 * contrast, if the process associated with bfqq is not slow, we
3256 * charge bfqq exactly with the service it has received.
3258 * Charging time to the first type of queues and the exact service to
3259 * the other has the effect of using the WF2Q+ policy to schedule the
3260 * former on a timeslice basis, without violating service domain
3261 * guarantees among the latter.
3263 void bfq_bfqq_expire(struct bfq_data *bfqd,
3264 struct bfq_queue *bfqq,
3266 enum bfqq_expiration reason)
3269 unsigned long delta = 0;
3270 struct bfq_entity *entity = &bfqq->entity;
3273 * Check whether the process is slow (see bfq_bfqq_is_slow).
3275 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
3278 * As above explained, charge slow (typically seeky) and
3279 * timed-out queues with the time and not the service
3280 * received, to favor sequential workloads.
3282 * Processes doing I/O in the slower disk zones will tend to
3283 * be slow(er) even if not seeky. Therefore, since the
3284 * estimated peak rate is actually an average over the disk
3285 * surface, these processes may timeout just for bad luck. To
3286 * avoid punishing them, do not charge time to processes that
3287 * succeeded in consuming at least 2/3 of their budget. This
3288 * allows BFQ to preserve enough elasticity to still perform
3289 * bandwidth, and not time, distribution with little unlucky
3290 * or quasi-sequential processes.
3292 if (bfqq->wr_coeff == 1 &&
3294 (reason == BFQQE_BUDGET_TIMEOUT &&
3295 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
3296 bfq_bfqq_charge_time(bfqd, bfqq, delta);
3298 if (reason == BFQQE_TOO_IDLE &&
3299 entity->service <= 2 * entity->budget / 10)
3300 bfq_clear_bfqq_IO_bound(bfqq);
3302 if (bfqd->low_latency && bfqq->wr_coeff == 1)
3303 bfqq->last_wr_start_finish = jiffies;
3305 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
3306 RB_EMPTY_ROOT(&bfqq->sort_list)) {
3308 * If we get here, and there are no outstanding
3309 * requests, then the request pattern is isochronous
3310 * (see the comments on the function
3311 * bfq_bfqq_softrt_next_start()). Thus we can compute
3312 * soft_rt_next_start. If, instead, the queue still
3313 * has outstanding requests, then we have to wait for
3314 * the completion of all the outstanding requests to
3315 * discover whether the request pattern is actually
3318 if (bfqq->dispatched == 0)
3319 bfqq->soft_rt_next_start =
3320 bfq_bfqq_softrt_next_start(bfqd, bfqq);
3323 * Schedule an update of soft_rt_next_start to when
3324 * the task may be discovered to be isochronous.
3326 bfq_mark_bfqq_softrt_update(bfqq);
3330 bfq_log_bfqq(bfqd, bfqq,
3331 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
3332 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
3335 * Increase, decrease or leave budget unchanged according to
3338 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
3339 if (__bfq_bfqq_expire(bfqd, bfqq))
3340 /* bfqq is gone, no more actions on it */
3343 bfqq->injected_service = 0;
3345 /* mark bfqq as waiting a request only if a bic still points to it */
3346 if (!bfq_bfqq_busy(bfqq) &&
3347 reason != BFQQE_BUDGET_TIMEOUT &&
3348 reason != BFQQE_BUDGET_EXHAUSTED) {
3349 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
3351 * Not setting service to 0, because, if the next rq
3352 * arrives in time, the queue will go on receiving
3353 * service with this same budget (as if it never expired)
3356 entity->service = 0;
3359 * Reset the received-service counter for every parent entity.
3360 * Differently from what happens with bfqq->entity.service,
3361 * the resetting of this counter never needs to be postponed
3362 * for parent entities. In fact, in case bfqq may have a
3363 * chance to go on being served using the last, partially
3364 * consumed budget, bfqq->entity.service needs to be kept,
3365 * because if bfqq then actually goes on being served using
3366 * the same budget, the last value of bfqq->entity.service is
3367 * needed to properly decrement bfqq->entity.budget by the
3368 * portion already consumed. In contrast, it is not necessary
3369 * to keep entity->service for parent entities too, because
3370 * the bubble up of the new value of bfqq->entity.budget will
3371 * make sure that the budgets of parent entities are correct,
3372 * even in case bfqq and thus parent entities go on receiving
3373 * service with the same budget.
3375 entity = entity->parent;
3376 for_each_entity(entity)
3377 entity->service = 0;
3381 * Budget timeout is not implemented through a dedicated timer, but
3382 * just checked on request arrivals and completions, as well as on
3383 * idle timer expirations.
3385 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
3387 return time_is_before_eq_jiffies(bfqq->budget_timeout);
3391 * If we expire a queue that is actively waiting (i.e., with the
3392 * device idled) for the arrival of a new request, then we may incur
3393 * the timestamp misalignment problem described in the body of the
3394 * function __bfq_activate_entity. Hence we return true only if this
3395 * condition does not hold, or if the queue is slow enough to deserve
3396 * only to be kicked off for preserving a high throughput.
3398 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
3400 bfq_log_bfqq(bfqq->bfqd, bfqq,
3401 "may_budget_timeout: wait_request %d left %d timeout %d",
3402 bfq_bfqq_wait_request(bfqq),
3403 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
3404 bfq_bfqq_budget_timeout(bfqq));
3406 return (!bfq_bfqq_wait_request(bfqq) ||
3407 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
3409 bfq_bfqq_budget_timeout(bfqq);
3413 * For a queue that becomes empty, device idling is allowed only if
3414 * this function returns true for the queue. As a consequence, since
3415 * device idling plays a critical role in both throughput boosting and
3416 * service guarantees, the return value of this function plays a
3417 * critical role in both these aspects as well.
3419 * In a nutshell, this function returns true only if idling is
3420 * beneficial for throughput or, even if detrimental for throughput,
3421 * idling is however necessary to preserve service guarantees (low
3422 * latency, desired throughput distribution, ...). In particular, on
3423 * NCQ-capable devices, this function tries to return false, so as to
3424 * help keep the drives' internal queues full, whenever this helps the
3425 * device boost the throughput without causing any service-guarantee
3428 * In more detail, the return value of this function is obtained by,
3429 * first, computing a number of boolean variables that take into
3430 * account throughput and service-guarantee issues, and, then,
3431 * combining these variables in a logical expression. Most of the
3432 * issues taken into account are not trivial. We discuss these issues
3433 * individually while introducing the variables.
3435 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
3437 struct bfq_data *bfqd = bfqq->bfqd;
3438 bool rot_without_queueing =
3439 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
3440 bfqq_sequential_and_IO_bound,
3441 idling_boosts_thr, idling_boosts_thr_without_issues,
3442 idling_needed_for_service_guarantees,
3443 asymmetric_scenario;
3445 if (bfqd->strict_guarantees)
3449 * Idling is performed only if slice_idle > 0. In addition, we
3452 * (b) bfqq is in the idle io prio class: in this case we do
3453 * not idle because we want to minimize the bandwidth that
3454 * queues in this class can steal to higher-priority queues
3456 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
3457 bfq_class_idle(bfqq))
3460 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
3461 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
3464 * The next variable takes into account the cases where idling
3465 * boosts the throughput.
3467 * The value of the variable is computed considering, first, that
3468 * idling is virtually always beneficial for the throughput if:
3469 * (a) the device is not NCQ-capable and rotational, or
3470 * (b) regardless of the presence of NCQ, the device is rotational and
3471 * the request pattern for bfqq is I/O-bound and sequential, or
3472 * (c) regardless of whether it is rotational, the device is
3473 * not NCQ-capable and the request pattern for bfqq is
3474 * I/O-bound and sequential.
3476 * Secondly, and in contrast to the above item (b), idling an
3477 * NCQ-capable flash-based device would not boost the
3478 * throughput even with sequential I/O; rather it would lower
3479 * the throughput in proportion to how fast the device
3480 * is. Accordingly, the next variable is true if any of the
3481 * above conditions (a), (b) or (c) is true, and, in
3482 * particular, happens to be false if bfqd is an NCQ-capable
3483 * flash-based device.
3485 idling_boosts_thr = rot_without_queueing ||
3486 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
3487 bfqq_sequential_and_IO_bound);
3490 * The value of the next variable,
3491 * idling_boosts_thr_without_issues, is equal to that of
3492 * idling_boosts_thr, unless a special case holds. In this
3493 * special case, described below, idling may cause problems to
3494 * weight-raised queues.
3496 * When the request pool is saturated (e.g., in the presence
3497 * of write hogs), if the processes associated with
3498 * non-weight-raised queues ask for requests at a lower rate,
3499 * then processes associated with weight-raised queues have a
3500 * higher probability to get a request from the pool
3501 * immediately (or at least soon) when they need one. Thus
3502 * they have a higher probability to actually get a fraction
3503 * of the device throughput proportional to their high
3504 * weight. This is especially true with NCQ-capable drives,
3505 * which enqueue several requests in advance, and further
3506 * reorder internally-queued requests.
3508 * For this reason, we force to false the value of
3509 * idling_boosts_thr_without_issues if there are weight-raised
3510 * busy queues. In this case, and if bfqq is not weight-raised,
3511 * this guarantees that the device is not idled for bfqq (if,
3512 * instead, bfqq is weight-raised, then idling will be
3513 * guaranteed by another variable, see below). Combined with
3514 * the timestamping rules of BFQ (see [1] for details), this
3515 * behavior causes bfqq, and hence any sync non-weight-raised
3516 * queue, to get a lower number of requests served, and thus
3517 * to ask for a lower number of requests from the request
3518 * pool, before the busy weight-raised queues get served
3519 * again. This often mitigates starvation problems in the
3520 * presence of heavy write workloads and NCQ, thereby
3521 * guaranteeing a higher application and system responsiveness
3522 * in these hostile scenarios.
3524 idling_boosts_thr_without_issues = idling_boosts_thr &&
3525 bfqd->wr_busy_queues == 0;
3528 * There is then a case where idling must be performed not
3529 * for throughput concerns, but to preserve service
3532 * To introduce this case, we can note that allowing the drive
3533 * to enqueue more than one request at a time, and hence
3534 * delegating de facto final scheduling decisions to the
3535 * drive's internal scheduler, entails loss of control on the
3536 * actual request service order. In particular, the critical
3537 * situation is when requests from different processes happen
3538 * to be present, at the same time, in the internal queue(s)
3539 * of the drive. In such a situation, the drive, by deciding
3540 * the service order of the internally-queued requests, does
3541 * determine also the actual throughput distribution among
3542 * these processes. But the drive typically has no notion or
3543 * concern about per-process throughput distribution, and
3544 * makes its decisions only on a per-request basis. Therefore,
3545 * the service distribution enforced by the drive's internal
3546 * scheduler is likely to coincide with the desired
3547 * device-throughput distribution only in a completely
3548 * symmetric scenario where:
3549 * (i) each of these processes must get the same throughput as
3551 * (ii) the I/O of each process has the same properties, in
3552 * terms of locality (sequential or random), direction
3553 * (reads or writes), request sizes, greediness
3554 * (from I/O-bound to sporadic), and so on.
3555 * In fact, in such a scenario, the drive tends to treat
3556 * the requests of each of these processes in about the same
3557 * way as the requests of the others, and thus to provide
3558 * each of these processes with about the same throughput
3559 * (which is exactly the desired throughput distribution). In
3560 * contrast, in any asymmetric scenario, device idling is
3561 * certainly needed to guarantee that bfqq receives its
3562 * assigned fraction of the device throughput (see [1] for
3564 * The problem is that idling may significantly reduce
3565 * throughput with certain combinations of types of I/O and
3566 * devices. An important example is sync random I/O, on flash
3567 * storage with command queueing. So, unless bfqq falls in the
3568 * above cases where idling also boosts throughput, it would
3569 * be important to check conditions (i) and (ii) accurately,
3570 * so as to avoid idling when not strictly needed for service
3573 * Unfortunately, it is extremely difficult to thoroughly
3574 * check condition (ii). And, in case there are active groups,
3575 * it becomes very difficult to check condition (i) too. In
3576 * fact, if there are active groups, then, for condition (i)
3577 * to become false, it is enough that an active group contains
3578 * more active processes or sub-groups than some other active
3579 * group. More precisely, for condition (i) to hold because of
3580 * such a group, it is not even necessary that the group is
3581 * (still) active: it is sufficient that, even if the group
3582 * has become inactive, some of its descendant processes still
3583 * have some request already dispatched but still waiting for
3584 * completion. In fact, requests have still to be guaranteed
3585 * their share of the throughput even after being
3586 * dispatched. In this respect, it is easy to show that, if a
3587 * group frequently becomes inactive while still having
3588 * in-flight requests, and if, when this happens, the group is
3589 * not considered in the calculation of whether the scenario
3590 * is asymmetric, then the group may fail to be guaranteed its
3591 * fair share of the throughput (basically because idling may
3592 * not be performed for the descendant processes of the group,
3593 * but it had to be). We address this issue with the
3594 * following bi-modal behavior, implemented in the function
3595 * bfq_symmetric_scenario().
3597 * If there are groups with requests waiting for completion
3598 * (as commented above, some of these groups may even be
3599 * already inactive), then the scenario is tagged as
3600 * asymmetric, conservatively, without checking any of the
3601 * conditions (i) and (ii). So the device is idled for bfqq.
3602 * This behavior matches also the fact that groups are created
3603 * exactly if controlling I/O is a primary concern (to
3604 * preserve bandwidth and latency guarantees).
3606 * On the opposite end, if there are no groups with requests
3607 * waiting for completion, then only condition (i) is actually
3608 * controlled, i.e., provided that condition (i) holds, idling
3609 * is not performed, regardless of whether condition (ii)
3610 * holds. In other words, only if condition (i) does not hold,
3611 * then idling is allowed, and the device tends to be
3612 * prevented from queueing many requests, possibly of several
3613 * processes. Since there are no groups with requests waiting
3614 * for completion, then, to control condition (i) it is enough
3615 * to check just whether all the queues with requests waiting
3616 * for completion also have the same weight.
3618 * Not checking condition (ii) evidently exposes bfqq to the
3619 * risk of getting less throughput than its fair share.
3620 * However, for queues with the same weight, a further
3621 * mechanism, preemption, mitigates or even eliminates this
3622 * problem. And it does so without consequences on overall
3623 * throughput. This mechanism and its benefits are explained
3624 * in the next three paragraphs.
3626 * Even if a queue, say Q, is expired when it remains idle, Q
3627 * can still preempt the new in-service queue if the next
3628 * request of Q arrives soon (see the comments on
3629 * bfq_bfqq_update_budg_for_activation). If all queues and
3630 * groups have the same weight, this form of preemption,
3631 * combined with the hole-recovery heuristic described in the
3632 * comments on function bfq_bfqq_update_budg_for_activation,
3633 * are enough to preserve a correct bandwidth distribution in
3634 * the mid term, even without idling. In fact, even if not
3635 * idling allows the internal queues of the device to contain
3636 * many requests, and thus to reorder requests, we can rather
3637 * safely assume that the internal scheduler still preserves a
3638 * minimum of mid-term fairness.
3640 * More precisely, this preemption-based, idleless approach
3641 * provides fairness in terms of IOPS, and not sectors per
3642 * second. This can be seen with a simple example. Suppose
3643 * that there are two queues with the same weight, but that
3644 * the first queue receives requests of 8 sectors, while the
3645 * second queue receives requests of 1024 sectors. In
3646 * addition, suppose that each of the two queues contains at
3647 * most one request at a time, which implies that each queue
3648 * always remains idle after it is served. Finally, after
3649 * remaining idle, each queue receives very quickly a new
3650 * request. It follows that the two queues are served
3651 * alternatively, preempting each other if needed. This
3652 * implies that, although both queues have the same weight,
3653 * the queue with large requests receives a service that is
3654 * 1024/8 times as high as the service received by the other
3657 * The motivation for using preemption instead of idling (for
3658 * queues with the same weight) is that, by not idling,
3659 * service guarantees are preserved (completely or at least in
3660 * part) without minimally sacrificing throughput. And, if
3661 * there is no active group, then the primary expectation for
3662 * this device is probably a high throughput.
3664 * We are now left only with explaining the additional
3665 * compound condition that is checked below for deciding
3666 * whether the scenario is asymmetric. To explain this
3667 * compound condition, we need to add that the function
3668 * bfq_symmetric_scenario checks the weights of only
3669 * non-weight-raised queues, for efficiency reasons (see
3670 * comments on bfq_weights_tree_add()). Then the fact that
3671 * bfqq is weight-raised is checked explicitly here. More
3672 * precisely, the compound condition below takes into account
3673 * also the fact that, even if bfqq is being weight-raised,
3674 * the scenario is still symmetric if all queues with requests
3675 * waiting for completion happen to be
3676 * weight-raised. Actually, we should be even more precise
3677 * here, and differentiate between interactive weight raising
3678 * and soft real-time weight raising.
3680 * As a side note, it is worth considering that the above
3681 * device-idling countermeasures may however fail in the
3682 * following unlucky scenario: if idling is (correctly)
3683 * disabled in a time period during which all symmetry
3684 * sub-conditions hold, and hence the device is allowed to
3685 * enqueue many requests, but at some later point in time some
3686 * sub-condition stops to hold, then it may become impossible
3687 * to let requests be served in the desired order until all
3688 * the requests already queued in the device have been served.
3690 asymmetric_scenario = (bfqq->wr_coeff > 1 &&
3691 bfqd->wr_busy_queues < bfqd->busy_queues) ||
3692 !bfq_symmetric_scenario(bfqd);
3695 * Finally, there is a case where maximizing throughput is the
3696 * best choice even if it may cause unfairness toward
3697 * bfqq. Such a case is when bfqq became active in a burst of
3698 * queue activations. Queues that became active during a large
3699 * burst benefit only from throughput, as discussed in the
3700 * comments on bfq_handle_burst. Thus, if bfqq became active
3701 * in a burst and not idling the device maximizes throughput,
3702 * then the device must no be idled, because not idling the
3703 * device provides bfqq and all other queues in the burst with
3704 * maximum benefit. Combining this and the above case, we can
3705 * now establish when idling is actually needed to preserve
3706 * service guarantees.
3708 idling_needed_for_service_guarantees =
3709 asymmetric_scenario && !bfq_bfqq_in_large_burst(bfqq);
3712 * We have now all the components we need to compute the
3713 * return value of the function, which is true only if idling
3714 * either boosts the throughput (without issues), or is
3715 * necessary to preserve service guarantees.
3717 return idling_boosts_thr_without_issues ||
3718 idling_needed_for_service_guarantees;
3722 * If the in-service queue is empty but the function bfq_better_to_idle
3723 * returns true, then:
3724 * 1) the queue must remain in service and cannot be expired, and
3725 * 2) the device must be idled to wait for the possible arrival of a new
3726 * request for the queue.
3727 * See the comments on the function bfq_better_to_idle for the reasons
3728 * why performing device idling is the best choice to boost the throughput
3729 * and preserve service guarantees when bfq_better_to_idle itself
3732 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
3734 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
3737 static struct bfq_queue *bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
3739 struct bfq_queue *bfqq;
3742 * A linear search; but, with a high probability, very few
3743 * steps are needed to find a candidate queue, i.e., a queue
3744 * with enough budget left for its next request. In fact:
3745 * - BFQ dynamically updates the budget of every queue so as
3746 * to accommodate the expected backlog of the queue;
3747 * - if a queue gets all its requests dispatched as injected
3748 * service, then the queue is removed from the active list
3749 * (and re-added only if it gets new requests, but with
3750 * enough budget for its new backlog).
3752 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
3753 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
3754 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
3755 bfq_bfqq_budget_left(bfqq))
3762 * Select a queue for service. If we have a current queue in service,
3763 * check whether to continue servicing it, or retrieve and set a new one.
3765 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
3767 struct bfq_queue *bfqq;
3768 struct request *next_rq;
3769 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
3771 bfqq = bfqd->in_service_queue;
3775 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
3778 * Do not expire bfqq for budget timeout if bfqq may be about
3779 * to enjoy device idling. The reason why, in this case, we
3780 * prevent bfqq from expiring is the same as in the comments
3781 * on the case where bfq_bfqq_must_idle() returns true, in
3782 * bfq_completed_request().
3784 if (bfq_may_expire_for_budg_timeout(bfqq) &&
3785 !bfq_bfqq_must_idle(bfqq))
3790 * This loop is rarely executed more than once. Even when it
3791 * happens, it is much more convenient to re-execute this loop
3792 * than to return NULL and trigger a new dispatch to get a
3795 next_rq = bfqq->next_rq;
3797 * If bfqq has requests queued and it has enough budget left to
3798 * serve them, keep the queue, otherwise expire it.
3801 if (bfq_serv_to_charge(next_rq, bfqq) >
3802 bfq_bfqq_budget_left(bfqq)) {
3804 * Expire the queue for budget exhaustion,
3805 * which makes sure that the next budget is
3806 * enough to serve the next request, even if
3807 * it comes from the fifo expired path.
3809 reason = BFQQE_BUDGET_EXHAUSTED;
3813 * The idle timer may be pending because we may
3814 * not disable disk idling even when a new request
3817 if (bfq_bfqq_wait_request(bfqq)) {
3819 * If we get here: 1) at least a new request
3820 * has arrived but we have not disabled the
3821 * timer because the request was too small,
3822 * 2) then the block layer has unplugged
3823 * the device, causing the dispatch to be
3826 * Since the device is unplugged, now the
3827 * requests are probably large enough to
3828 * provide a reasonable throughput.
3829 * So we disable idling.
3831 bfq_clear_bfqq_wait_request(bfqq);
3832 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
3839 * No requests pending. However, if the in-service queue is idling
3840 * for a new request, or has requests waiting for a completion and
3841 * may idle after their completion, then keep it anyway.
3843 * Yet, to boost throughput, inject service from other queues if
3846 if (bfq_bfqq_wait_request(bfqq) ||
3847 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
3848 if (bfq_bfqq_injectable(bfqq) &&
3849 bfqq->injected_service * bfqq->inject_coeff <
3850 bfqq->entity.service * 10)
3851 bfqq = bfq_choose_bfqq_for_injection(bfqd);
3858 reason = BFQQE_NO_MORE_REQUESTS;
3860 bfq_bfqq_expire(bfqd, bfqq, false, reason);
3862 bfqq = bfq_set_in_service_queue(bfqd);
3864 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
3869 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
3871 bfq_log(bfqd, "select_queue: no queue returned");
3876 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3878 struct bfq_entity *entity = &bfqq->entity;
3880 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
3881 bfq_log_bfqq(bfqd, bfqq,
3882 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
3883 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
3884 jiffies_to_msecs(bfqq->wr_cur_max_time),
3886 bfqq->entity.weight, bfqq->entity.orig_weight);
3888 if (entity->prio_changed)
3889 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
3892 * If the queue was activated in a burst, or too much
3893 * time has elapsed from the beginning of this
3894 * weight-raising period, then end weight raising.
3896 if (bfq_bfqq_in_large_burst(bfqq))
3897 bfq_bfqq_end_wr(bfqq);
3898 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
3899 bfqq->wr_cur_max_time)) {
3900 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
3901 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
3902 bfq_wr_duration(bfqd)))
3903 bfq_bfqq_end_wr(bfqq);
3905 switch_back_to_interactive_wr(bfqq, bfqd);
3906 bfqq->entity.prio_changed = 1;
3909 if (bfqq->wr_coeff > 1 &&
3910 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
3911 bfqq->service_from_wr > max_service_from_wr) {
3912 /* see comments on max_service_from_wr */
3913 bfq_bfqq_end_wr(bfqq);
3917 * To improve latency (for this or other queues), immediately
3918 * update weight both if it must be raised and if it must be
3919 * lowered. Since, entity may be on some active tree here, and
3920 * might have a pending change of its ioprio class, invoke
3921 * next function with the last parameter unset (see the
3922 * comments on the function).
3924 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
3925 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
3930 * Dispatch next request from bfqq.
3932 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
3933 struct bfq_queue *bfqq)
3935 struct request *rq = bfqq->next_rq;
3936 unsigned long service_to_charge;
3938 service_to_charge = bfq_serv_to_charge(rq, bfqq);
3940 bfq_bfqq_served(bfqq, service_to_charge);
3942 bfq_dispatch_remove(bfqd->queue, rq);
3944 if (bfqq != bfqd->in_service_queue) {
3945 if (likely(bfqd->in_service_queue))
3946 bfqd->in_service_queue->injected_service +=
3947 bfq_serv_to_charge(rq, bfqq);
3953 * If weight raising has to terminate for bfqq, then next
3954 * function causes an immediate update of bfqq's weight,
3955 * without waiting for next activation. As a consequence, on
3956 * expiration, bfqq will be timestamped as if has never been
3957 * weight-raised during this service slot, even if it has
3958 * received part or even most of the service as a
3959 * weight-raised queue. This inflates bfqq's timestamps, which
3960 * is beneficial, as bfqq is then more willing to leave the
3961 * device immediately to possible other weight-raised queues.
3963 bfq_update_wr_data(bfqd, bfqq);
3966 * Expire bfqq, pretending that its budget expired, if bfqq
3967 * belongs to CLASS_IDLE and other queues are waiting for
3970 if (!(bfqd->busy_queues > 1 && bfq_class_idle(bfqq)))
3973 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
3979 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
3981 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3984 * Avoiding lock: a race on bfqd->busy_queues should cause at
3985 * most a call to dispatch for nothing
3987 return !list_empty_careful(&bfqd->dispatch) ||
3988 bfqd->busy_queues > 0;
3991 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3993 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3994 struct request *rq = NULL;
3995 struct bfq_queue *bfqq = NULL;
3997 if (!list_empty(&bfqd->dispatch)) {
3998 rq = list_first_entry(&bfqd->dispatch, struct request,
4000 list_del_init(&rq->queuelist);
4006 * Increment counters here, because this
4007 * dispatch does not follow the standard
4008 * dispatch flow (where counters are
4013 goto inc_in_driver_start_rq;
4017 * We exploit the bfq_finish_requeue_request hook to
4018 * decrement rq_in_driver, but
4019 * bfq_finish_requeue_request will not be invoked on
4020 * this request. So, to avoid unbalance, just start
4021 * this request, without incrementing rq_in_driver. As
4022 * a negative consequence, rq_in_driver is deceptively
4023 * lower than it should be while this request is in
4024 * service. This may cause bfq_schedule_dispatch to be
4025 * invoked uselessly.
4027 * As for implementing an exact solution, the
4028 * bfq_finish_requeue_request hook, if defined, is
4029 * probably invoked also on this request. So, by
4030 * exploiting this hook, we could 1) increment
4031 * rq_in_driver here, and 2) decrement it in
4032 * bfq_finish_requeue_request. Such a solution would
4033 * let the value of the counter be always accurate,
4034 * but it would entail using an extra interface
4035 * function. This cost seems higher than the benefit,
4036 * being the frequency of non-elevator-private
4037 * requests very low.
4042 bfq_log(bfqd, "dispatch requests: %d busy queues", bfqd->busy_queues);
4044 if (bfqd->busy_queues == 0)
4048 * Force device to serve one request at a time if
4049 * strict_guarantees is true. Forcing this service scheme is
4050 * currently the ONLY way to guarantee that the request
4051 * service order enforced by the scheduler is respected by a
4052 * queueing device. Otherwise the device is free even to make
4053 * some unlucky request wait for as long as the device
4056 * Of course, serving one request at at time may cause loss of
4059 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
4062 bfqq = bfq_select_queue(bfqd);
4066 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
4069 inc_in_driver_start_rq:
4070 bfqd->rq_in_driver++;
4072 rq->rq_flags |= RQF_STARTED;
4078 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
4079 static void bfq_update_dispatch_stats(struct request_queue *q,
4081 struct bfq_queue *in_serv_queue,
4082 bool idle_timer_disabled)
4084 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
4086 if (!idle_timer_disabled && !bfqq)
4090 * rq and bfqq are guaranteed to exist until this function
4091 * ends, for the following reasons. First, rq can be
4092 * dispatched to the device, and then can be completed and
4093 * freed, only after this function ends. Second, rq cannot be
4094 * merged (and thus freed because of a merge) any longer,
4095 * because it has already started. Thus rq cannot be freed
4096 * before this function ends, and, since rq has a reference to
4097 * bfqq, the same guarantee holds for bfqq too.
4099 * In addition, the following queue lock guarantees that
4100 * bfqq_group(bfqq) exists as well.
4102 spin_lock_irq(q->queue_lock);
4103 if (idle_timer_disabled)
4105 * Since the idle timer has been disabled,
4106 * in_serv_queue contained some request when
4107 * __bfq_dispatch_request was invoked above, which
4108 * implies that rq was picked exactly from
4109 * in_serv_queue. Thus in_serv_queue == bfqq, and is
4110 * therefore guaranteed to exist because of the above
4113 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
4115 struct bfq_group *bfqg = bfqq_group(bfqq);
4117 bfqg_stats_update_avg_queue_size(bfqg);
4118 bfqg_stats_set_start_empty_time(bfqg);
4119 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
4121 spin_unlock_irq(q->queue_lock);
4124 static inline void bfq_update_dispatch_stats(struct request_queue *q,
4126 struct bfq_queue *in_serv_queue,
4127 bool idle_timer_disabled) {}
4130 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
4132 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4134 struct bfq_queue *in_serv_queue;
4135 bool waiting_rq, idle_timer_disabled = false;
4137 spin_lock_irq(&bfqd->lock);
4139 in_serv_queue = bfqd->in_service_queue;
4140 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
4142 rq = __bfq_dispatch_request(hctx);
4143 if (in_serv_queue == bfqd->in_service_queue) {
4144 idle_timer_disabled =
4145 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
4148 spin_unlock_irq(&bfqd->lock);
4149 bfq_update_dispatch_stats(hctx->queue, rq,
4150 idle_timer_disabled ? in_serv_queue : NULL,
4151 idle_timer_disabled);
4157 * Task holds one reference to the queue, dropped when task exits. Each rq
4158 * in-flight on this queue also holds a reference, dropped when rq is freed.
4160 * Scheduler lock must be held here. Recall not to use bfqq after calling
4161 * this function on it.
4163 void bfq_put_queue(struct bfq_queue *bfqq)
4165 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4166 struct bfq_group *bfqg = bfqq_group(bfqq);
4170 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
4177 if (!hlist_unhashed(&bfqq->burst_list_node)) {
4178 hlist_del_init(&bfqq->burst_list_node);
4180 * Decrement also burst size after the removal, if the
4181 * process associated with bfqq is exiting, and thus
4182 * does not contribute to the burst any longer. This
4183 * decrement helps filter out false positives of large
4184 * bursts, when some short-lived process (often due to
4185 * the execution of commands by some service) happens
4186 * to start and exit while a complex application is
4187 * starting, and thus spawning several processes that
4188 * do I/O (and that *must not* be treated as a large
4189 * burst, see comments on bfq_handle_burst).
4191 * In particular, the decrement is performed only if:
4192 * 1) bfqq is not a merged queue, because, if it is,
4193 * then this free of bfqq is not triggered by the exit
4194 * of the process bfqq is associated with, but exactly
4195 * by the fact that bfqq has just been merged.
4196 * 2) burst_size is greater than 0, to handle
4197 * unbalanced decrements. Unbalanced decrements may
4198 * happen in te following case: bfqq is inserted into
4199 * the current burst list--without incrementing
4200 * bust_size--because of a split, but the current
4201 * burst list is not the burst list bfqq belonged to
4202 * (see comments on the case of a split in
4205 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
4206 bfqq->bfqd->burst_size--;
4209 kmem_cache_free(bfq_pool, bfqq);
4210 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4211 bfqg_and_blkg_put(bfqg);
4215 static void bfq_put_cooperator(struct bfq_queue *bfqq)
4217 struct bfq_queue *__bfqq, *next;
4220 * If this queue was scheduled to merge with another queue, be
4221 * sure to drop the reference taken on that queue (and others in
4222 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
4224 __bfqq = bfqq->new_bfqq;
4228 next = __bfqq->new_bfqq;
4229 bfq_put_queue(__bfqq);
4234 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4236 if (bfqq == bfqd->in_service_queue) {
4237 __bfq_bfqq_expire(bfqd, bfqq);
4238 bfq_schedule_dispatch(bfqd);
4241 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
4243 bfq_put_cooperator(bfqq);
4245 bfq_put_queue(bfqq); /* release process reference */
4248 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
4250 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4251 struct bfq_data *bfqd;
4254 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
4257 unsigned long flags;
4259 spin_lock_irqsave(&bfqd->lock, flags);
4261 bfq_exit_bfqq(bfqd, bfqq);
4262 bic_set_bfqq(bic, NULL, is_sync);
4263 spin_unlock_irqrestore(&bfqd->lock, flags);
4267 static void bfq_exit_icq(struct io_cq *icq)
4269 struct bfq_io_cq *bic = icq_to_bic(icq);
4271 bfq_exit_icq_bfqq(bic, true);
4272 bfq_exit_icq_bfqq(bic, false);
4276 * Update the entity prio values; note that the new values will not
4277 * be used until the next (re)activation.
4280 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
4282 struct task_struct *tsk = current;
4284 struct bfq_data *bfqd = bfqq->bfqd;
4289 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4290 switch (ioprio_class) {
4292 pr_err("bdi %s: bfq: bad prio class %d\n",
4293 bdi_dev_name(bfqq->bfqd->queue->backing_dev_info),
4296 case IOPRIO_CLASS_NONE:
4298 * No prio set, inherit CPU scheduling settings.
4300 bfqq->new_ioprio = task_nice_ioprio(tsk);
4301 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
4303 case IOPRIO_CLASS_RT:
4304 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4305 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
4307 case IOPRIO_CLASS_BE:
4308 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4309 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
4311 case IOPRIO_CLASS_IDLE:
4312 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
4313 bfqq->new_ioprio = 7;
4317 if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
4318 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
4320 bfqq->new_ioprio = IOPRIO_BE_NR - 1;
4323 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
4324 bfqq->entity.prio_changed = 1;
4327 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4328 struct bio *bio, bool is_sync,
4329 struct bfq_io_cq *bic);
4331 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
4333 struct bfq_data *bfqd = bic_to_bfqd(bic);
4334 struct bfq_queue *bfqq;
4335 int ioprio = bic->icq.ioc->ioprio;
4338 * This condition may trigger on a newly created bic, be sure to
4339 * drop the lock before returning.
4341 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
4344 bic->ioprio = ioprio;
4346 bfqq = bic_to_bfqq(bic, false);
4348 /* release process reference on this queue */
4349 bfq_put_queue(bfqq);
4350 bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
4351 bic_set_bfqq(bic, bfqq, false);
4354 bfqq = bic_to_bfqq(bic, true);
4356 bfq_set_next_ioprio_data(bfqq, bic);
4359 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4360 struct bfq_io_cq *bic, pid_t pid, int is_sync)
4362 RB_CLEAR_NODE(&bfqq->entity.rb_node);
4363 INIT_LIST_HEAD(&bfqq->fifo);
4364 INIT_HLIST_NODE(&bfqq->burst_list_node);
4370 bfq_set_next_ioprio_data(bfqq, bic);
4374 * No need to mark as has_short_ttime if in
4375 * idle_class, because no device idling is performed
4376 * for queues in idle class
4378 if (!bfq_class_idle(bfqq))
4379 /* tentatively mark as has_short_ttime */
4380 bfq_mark_bfqq_has_short_ttime(bfqq);
4381 bfq_mark_bfqq_sync(bfqq);
4382 bfq_mark_bfqq_just_created(bfqq);
4384 * Aggressively inject a lot of service: up to 90%.
4385 * This coefficient remains constant during bfqq life,
4386 * but this behavior might be changed, after enough
4387 * testing and tuning.
4389 bfqq->inject_coeff = 1;
4391 bfq_clear_bfqq_sync(bfqq);
4393 /* set end request to minus infinity from now */
4394 bfqq->ttime.last_end_request = ktime_get_ns() + 1;
4396 bfq_mark_bfqq_IO_bound(bfqq);
4400 /* Tentative initial value to trade off between thr and lat */
4401 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
4402 bfqq->budget_timeout = bfq_smallest_from_now();
4405 bfqq->last_wr_start_finish = jiffies;
4406 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
4407 bfqq->split_time = bfq_smallest_from_now();
4410 * To not forget the possibly high bandwidth consumed by a
4411 * process/queue in the recent past,
4412 * bfq_bfqq_softrt_next_start() returns a value at least equal
4413 * to the current value of bfqq->soft_rt_next_start (see
4414 * comments on bfq_bfqq_softrt_next_start). Set
4415 * soft_rt_next_start to now, to mean that bfqq has consumed
4416 * no bandwidth so far.
4418 bfqq->soft_rt_next_start = jiffies;
4420 /* first request is almost certainly seeky */
4421 bfqq->seek_history = 1;
4424 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
4425 struct bfq_group *bfqg,
4426 int ioprio_class, int ioprio)
4428 switch (ioprio_class) {
4429 case IOPRIO_CLASS_RT:
4430 return &bfqg->async_bfqq[0][ioprio];
4431 case IOPRIO_CLASS_NONE:
4432 ioprio = IOPRIO_NORM;
4434 case IOPRIO_CLASS_BE:
4435 return &bfqg->async_bfqq[1][ioprio];
4436 case IOPRIO_CLASS_IDLE:
4437 return &bfqg->async_idle_bfqq;
4443 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4444 struct bio *bio, bool is_sync,
4445 struct bfq_io_cq *bic)
4447 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4448 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4449 struct bfq_queue **async_bfqq = NULL;
4450 struct bfq_queue *bfqq;
4451 struct bfq_group *bfqg;
4455 bfqg = bfq_find_set_group(bfqd, bio_blkcg(bio));
4457 bfqq = &bfqd->oom_bfqq;
4462 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
4469 bfqq = kmem_cache_alloc_node(bfq_pool,
4470 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
4474 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
4476 bfq_init_entity(&bfqq->entity, bfqg);
4477 bfq_log_bfqq(bfqd, bfqq, "allocated");
4479 bfqq = &bfqd->oom_bfqq;
4480 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
4485 * Pin the queue now that it's allocated, scheduler exit will
4490 * Extra group reference, w.r.t. sync
4491 * queue. This extra reference is removed
4492 * only if bfqq->bfqg disappears, to
4493 * guarantee that this queue is not freed
4494 * until its group goes away.
4496 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
4502 bfqq->ref++; /* get a process reference to this queue */
4503 bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
4508 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
4509 struct bfq_queue *bfqq)
4511 struct bfq_ttime *ttime = &bfqq->ttime;
4512 u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
4514 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
4516 ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
4517 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
4518 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
4519 ttime->ttime_samples);
4523 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4526 bfqq->seek_history <<= 1;
4527 bfqq->seek_history |=
4528 get_sdist(bfqq->last_request_pos, rq) > BFQQ_SEEK_THR &&
4529 (!blk_queue_nonrot(bfqd->queue) ||
4530 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT);
4533 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
4534 struct bfq_queue *bfqq,
4535 struct bfq_io_cq *bic)
4537 bool has_short_ttime = true;
4540 * No need to update has_short_ttime if bfqq is async or in
4541 * idle io prio class, or if bfq_slice_idle is zero, because
4542 * no device idling is performed for bfqq in this case.
4544 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
4545 bfqd->bfq_slice_idle == 0)
4548 /* Idle window just restored, statistics are meaningless. */
4549 if (time_is_after_eq_jiffies(bfqq->split_time +
4550 bfqd->bfq_wr_min_idle_time))
4553 /* Think time is infinite if no process is linked to
4554 * bfqq. Otherwise check average think time to
4555 * decide whether to mark as has_short_ttime
4557 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
4558 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
4559 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle))
4560 has_short_ttime = false;
4562 bfq_log_bfqq(bfqd, bfqq, "update_has_short_ttime: has_short_ttime %d",
4565 if (has_short_ttime)
4566 bfq_mark_bfqq_has_short_ttime(bfqq);
4568 bfq_clear_bfqq_has_short_ttime(bfqq);
4572 * Called when a new fs request (rq) is added to bfqq. Check if there's
4573 * something we should do about it.
4575 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4578 struct bfq_io_cq *bic = RQ_BIC(rq);
4580 if (rq->cmd_flags & REQ_META)
4581 bfqq->meta_pending++;
4583 bfq_update_io_thinktime(bfqd, bfqq);
4584 bfq_update_has_short_ttime(bfqd, bfqq, bic);
4585 bfq_update_io_seektime(bfqd, bfqq, rq);
4587 bfq_log_bfqq(bfqd, bfqq,
4588 "rq_enqueued: has_short_ttime=%d (seeky %d)",
4589 bfq_bfqq_has_short_ttime(bfqq), BFQQ_SEEKY(bfqq));
4591 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4593 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
4594 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
4595 blk_rq_sectors(rq) < 32;
4596 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
4599 * There is just this request queued: if the request
4600 * is small and the queue is not to be expired, then
4603 * In this way, if the device is being idled to wait
4604 * for a new request from the in-service queue, we
4605 * avoid unplugging the device and committing the
4606 * device to serve just a small request. On the
4607 * contrary, we wait for the block layer to decide
4608 * when to unplug the device: hopefully, new requests
4609 * will be merged to this one quickly, then the device
4610 * will be unplugged and larger requests will be
4613 if (small_req && !budget_timeout)
4617 * A large enough request arrived, or the queue is to
4618 * be expired: in both cases disk idling is to be
4619 * stopped, so clear wait_request flag and reset
4622 bfq_clear_bfqq_wait_request(bfqq);
4623 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4626 * The queue is not empty, because a new request just
4627 * arrived. Hence we can safely expire the queue, in
4628 * case of budget timeout, without risking that the
4629 * timestamps of the queue are not updated correctly.
4630 * See [1] for more details.
4633 bfq_bfqq_expire(bfqd, bfqq, false,
4634 BFQQE_BUDGET_TIMEOUT);
4638 /* returns true if it causes the idle timer to be disabled */
4639 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
4641 struct bfq_queue *bfqq = RQ_BFQQ(rq),
4642 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
4643 bool waiting, idle_timer_disabled = false;
4646 if (bic_to_bfqq(RQ_BIC(rq), 1) != bfqq)
4647 new_bfqq = bic_to_bfqq(RQ_BIC(rq), 1);
4649 * Release the request's reference to the old bfqq
4650 * and make sure one is taken to the shared queue.
4652 new_bfqq->allocated++;
4656 * If the bic associated with the process
4657 * issuing this request still points to bfqq
4658 * (and thus has not been already redirected
4659 * to new_bfqq or even some other bfq_queue),
4660 * then complete the merge and redirect it to
4663 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
4664 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
4667 bfq_clear_bfqq_just_created(bfqq);
4669 * rq is about to be enqueued into new_bfqq,
4670 * release rq reference on bfqq
4672 bfq_put_queue(bfqq);
4673 rq->elv.priv[1] = new_bfqq;
4677 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
4678 bfq_add_request(rq);
4679 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
4681 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
4682 list_add_tail(&rq->queuelist, &bfqq->fifo);
4684 bfq_rq_enqueued(bfqd, bfqq, rq);
4686 return idle_timer_disabled;
4689 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
4690 static void bfq_update_insert_stats(struct request_queue *q,
4691 struct bfq_queue *bfqq,
4692 bool idle_timer_disabled,
4693 unsigned int cmd_flags)
4699 * bfqq still exists, because it can disappear only after
4700 * either it is merged with another queue, or the process it
4701 * is associated with exits. But both actions must be taken by
4702 * the same process currently executing this flow of
4705 * In addition, the following queue lock guarantees that
4706 * bfqq_group(bfqq) exists as well.
4708 spin_lock_irq(q->queue_lock);
4709 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
4710 if (idle_timer_disabled)
4711 bfqg_stats_update_idle_time(bfqq_group(bfqq));
4712 spin_unlock_irq(q->queue_lock);
4715 static inline void bfq_update_insert_stats(struct request_queue *q,
4716 struct bfq_queue *bfqq,
4717 bool idle_timer_disabled,
4718 unsigned int cmd_flags) {}
4721 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
4724 struct request_queue *q = hctx->queue;
4725 struct bfq_data *bfqd = q->elevator->elevator_data;
4726 struct bfq_queue *bfqq;
4727 bool idle_timer_disabled = false;
4728 unsigned int cmd_flags;
4730 spin_lock_irq(&bfqd->lock);
4731 if (blk_mq_sched_try_insert_merge(q, rq)) {
4732 spin_unlock_irq(&bfqd->lock);
4736 spin_unlock_irq(&bfqd->lock);
4738 blk_mq_sched_request_inserted(rq);
4740 spin_lock_irq(&bfqd->lock);
4741 bfqq = bfq_init_rq(rq);
4742 if (!bfqq || at_head || blk_rq_is_passthrough(rq)) {
4744 list_add(&rq->queuelist, &bfqd->dispatch);
4746 list_add_tail(&rq->queuelist, &bfqd->dispatch);
4748 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
4750 * Update bfqq, because, if a queue merge has occurred
4751 * in __bfq_insert_request, then rq has been
4752 * redirected into a new queue.
4756 if (rq_mergeable(rq)) {
4757 elv_rqhash_add(q, rq);
4764 * Cache cmd_flags before releasing scheduler lock, because rq
4765 * may disappear afterwards (for example, because of a request
4768 cmd_flags = rq->cmd_flags;
4770 spin_unlock_irq(&bfqd->lock);
4772 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
4776 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
4777 struct list_head *list, bool at_head)
4779 while (!list_empty(list)) {
4782 rq = list_first_entry(list, struct request, queuelist);
4783 list_del_init(&rq->queuelist);
4784 bfq_insert_request(hctx, rq, at_head);
4788 static void bfq_update_hw_tag(struct bfq_data *bfqd)
4790 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
4791 bfqd->rq_in_driver);
4793 if (bfqd->hw_tag == 1)
4797 * This sample is valid if the number of outstanding requests
4798 * is large enough to allow a queueing behavior. Note that the
4799 * sum is not exact, as it's not taking into account deactivated
4802 if (bfqd->rq_in_driver + bfqd->queued < BFQ_HW_QUEUE_THRESHOLD)
4805 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
4808 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
4809 bfqd->max_rq_in_driver = 0;
4810 bfqd->hw_tag_samples = 0;
4813 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
4818 bfq_update_hw_tag(bfqd);
4820 bfqd->rq_in_driver--;
4823 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
4825 * Set budget_timeout (which we overload to store the
4826 * time at which the queue remains with no backlog and
4827 * no outstanding request; used by the weight-raising
4830 bfqq->budget_timeout = jiffies;
4832 bfq_weights_tree_remove(bfqd, bfqq);
4835 now_ns = ktime_get_ns();
4837 bfqq->ttime.last_end_request = now_ns;
4840 * Using us instead of ns, to get a reasonable precision in
4841 * computing rate in next check.
4843 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
4846 * If the request took rather long to complete, and, according
4847 * to the maximum request size recorded, this completion latency
4848 * implies that the request was certainly served at a very low
4849 * rate (less than 1M sectors/sec), then the whole observation
4850 * interval that lasts up to this time instant cannot be a
4851 * valid time interval for computing a new peak rate. Invoke
4852 * bfq_update_rate_reset to have the following three steps
4854 * - close the observation interval at the last (previous)
4855 * request dispatch or completion
4856 * - compute rate, if possible, for that observation interval
4857 * - reset to zero samples, which will trigger a proper
4858 * re-initialization of the observation interval on next
4861 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
4862 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
4863 1UL<<(BFQ_RATE_SHIFT - 10))
4864 bfq_update_rate_reset(bfqd, NULL);
4865 bfqd->last_completion = now_ns;
4868 * If we are waiting to discover whether the request pattern
4869 * of the task associated with the queue is actually
4870 * isochronous, and both requisites for this condition to hold
4871 * are now satisfied, then compute soft_rt_next_start (see the
4872 * comments on the function bfq_bfqq_softrt_next_start()). We
4873 * schedule this delayed check when bfqq expires, if it still
4874 * has in-flight requests.
4876 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
4877 RB_EMPTY_ROOT(&bfqq->sort_list))
4878 bfqq->soft_rt_next_start =
4879 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4882 * If this is the in-service queue, check if it needs to be expired,
4883 * or if we want to idle in case it has no pending requests.
4885 if (bfqd->in_service_queue == bfqq) {
4886 if (bfq_bfqq_must_idle(bfqq)) {
4887 if (bfqq->dispatched == 0)
4888 bfq_arm_slice_timer(bfqd);
4890 * If we get here, we do not expire bfqq, even
4891 * if bfqq was in budget timeout or had no
4892 * more requests (as controlled in the next
4893 * conditional instructions). The reason for
4894 * not expiring bfqq is as follows.
4896 * Here bfqq->dispatched > 0 holds, but
4897 * bfq_bfqq_must_idle() returned true. This
4898 * implies that, even if no request arrives
4899 * for bfqq before bfqq->dispatched reaches 0,
4900 * bfqq will, however, not be expired on the
4901 * completion event that causes bfqq->dispatch
4902 * to reach zero. In contrast, on this event,
4903 * bfqq will start enjoying device idling
4904 * (I/O-dispatch plugging).
4906 * But, if we expired bfqq here, bfqq would
4907 * not have the chance to enjoy device idling
4908 * when bfqq->dispatched finally reaches
4909 * zero. This would expose bfqq to violation
4910 * of its reserved service guarantees.
4913 } else if (bfq_may_expire_for_budg_timeout(bfqq))
4914 bfq_bfqq_expire(bfqd, bfqq, false,
4915 BFQQE_BUDGET_TIMEOUT);
4916 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
4917 (bfqq->dispatched == 0 ||
4918 !bfq_better_to_idle(bfqq)))
4919 bfq_bfqq_expire(bfqd, bfqq, false,
4920 BFQQE_NO_MORE_REQUESTS);
4923 if (!bfqd->rq_in_driver)
4924 bfq_schedule_dispatch(bfqd);
4927 static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq)
4931 bfq_put_queue(bfqq);
4935 * Handle either a requeue or a finish for rq. The things to do are
4936 * the same in both cases: all references to rq are to be dropped. In
4937 * particular, rq is considered completed from the point of view of
4940 static void bfq_finish_requeue_request(struct request *rq)
4942 struct bfq_queue *bfqq = RQ_BFQQ(rq);
4943 struct bfq_data *bfqd;
4946 * Requeue and finish hooks are invoked in blk-mq without
4947 * checking whether the involved request is actually still
4948 * referenced in the scheduler. To handle this fact, the
4949 * following two checks make this function exit in case of
4950 * spurious invocations, for which there is nothing to do.
4952 * First, check whether rq has nothing to do with an elevator.
4954 if (unlikely(!(rq->rq_flags & RQF_ELVPRIV)))
4958 * rq either is not associated with any icq, or is an already
4959 * requeued request that has not (yet) been re-inserted into
4962 if (!rq->elv.icq || !bfqq)
4967 if (rq->rq_flags & RQF_STARTED)
4968 bfqg_stats_update_completion(bfqq_group(bfqq),
4970 rq->io_start_time_ns,
4973 if (likely(rq->rq_flags & RQF_STARTED)) {
4974 unsigned long flags;
4976 spin_lock_irqsave(&bfqd->lock, flags);
4978 bfq_completed_request(bfqq, bfqd);
4979 bfq_finish_requeue_request_body(bfqq);
4981 spin_unlock_irqrestore(&bfqd->lock, flags);
4984 * Request rq may be still/already in the scheduler,
4985 * in which case we need to remove it (this should
4986 * never happen in case of requeue). And we cannot
4987 * defer such a check and removal, to avoid
4988 * inconsistencies in the time interval from the end
4989 * of this function to the start of the deferred work.
4990 * This situation seems to occur only in process
4991 * context, as a consequence of a merge. In the
4992 * current version of the code, this implies that the
4996 if (!RB_EMPTY_NODE(&rq->rb_node)) {
4997 bfq_remove_request(rq->q, rq);
4998 bfqg_stats_update_io_remove(bfqq_group(bfqq),
5001 bfq_finish_requeue_request_body(bfqq);
5005 * Reset private fields. In case of a requeue, this allows
5006 * this function to correctly do nothing if it is spuriously
5007 * invoked again on this same request (see the check at the
5008 * beginning of the function). Probably, a better general
5009 * design would be to prevent blk-mq from invoking the requeue
5010 * or finish hooks of an elevator, for a request that is not
5011 * referred by that elevator.
5013 * Resetting the following fields would break the
5014 * request-insertion logic if rq is re-inserted into a bfq
5015 * internal queue, without a re-preparation. Here we assume
5016 * that re-insertions of requeued requests, without
5017 * re-preparation, can happen only for pass_through or at_head
5018 * requests (which are not re-inserted into bfq internal
5021 rq->elv.priv[0] = NULL;
5022 rq->elv.priv[1] = NULL;
5026 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
5027 * was the last process referring to that bfqq.
5029 static struct bfq_queue *
5030 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
5032 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
5034 if (bfqq_process_refs(bfqq) == 1) {
5035 bfqq->pid = current->pid;
5036 bfq_clear_bfqq_coop(bfqq);
5037 bfq_clear_bfqq_split_coop(bfqq);
5041 bic_set_bfqq(bic, NULL, 1);
5043 bfq_put_cooperator(bfqq);
5045 bfq_put_queue(bfqq);
5049 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
5050 struct bfq_io_cq *bic,
5052 bool split, bool is_sync,
5055 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
5057 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
5064 bfq_put_queue(bfqq);
5065 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
5067 bic_set_bfqq(bic, bfqq, is_sync);
5068 if (split && is_sync) {
5069 if ((bic->was_in_burst_list && bfqd->large_burst) ||
5070 bic->saved_in_large_burst)
5071 bfq_mark_bfqq_in_large_burst(bfqq);
5073 bfq_clear_bfqq_in_large_burst(bfqq);
5074 if (bic->was_in_burst_list)
5076 * If bfqq was in the current
5077 * burst list before being
5078 * merged, then we have to add
5079 * it back. And we do not need
5080 * to increase burst_size, as
5081 * we did not decrement
5082 * burst_size when we removed
5083 * bfqq from the burst list as
5084 * a consequence of a merge
5086 * bfq_put_queue). In this
5087 * respect, it would be rather
5088 * costly to know whether the
5089 * current burst list is still
5090 * the same burst list from
5091 * which bfqq was removed on
5092 * the merge. To avoid this
5093 * cost, if bfqq was in a
5094 * burst list, then we add
5095 * bfqq to the current burst
5096 * list without any further
5097 * check. This can cause
5098 * inappropriate insertions,
5099 * but rarely enough to not
5100 * harm the detection of large
5101 * bursts significantly.
5103 hlist_add_head(&bfqq->burst_list_node,
5106 bfqq->split_time = jiffies;
5113 * Only reset private fields. The actual request preparation will be
5114 * performed by bfq_init_rq, when rq is either inserted or merged. See
5115 * comments on bfq_init_rq for the reason behind this delayed
5118 static void bfq_prepare_request(struct request *rq, struct bio *bio)
5121 * Regardless of whether we have an icq attached, we have to
5122 * clear the scheduler pointers, as they might point to
5123 * previously allocated bic/bfqq structs.
5125 rq->elv.priv[0] = rq->elv.priv[1] = NULL;
5129 * If needed, init rq, allocate bfq data structures associated with
5130 * rq, and increment reference counters in the destination bfq_queue
5131 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
5132 * not associated with any bfq_queue.
5134 * This function is invoked by the functions that perform rq insertion
5135 * or merging. One may have expected the above preparation operations
5136 * to be performed in bfq_prepare_request, and not delayed to when rq
5137 * is inserted or merged. The rationale behind this delayed
5138 * preparation is that, after the prepare_request hook is invoked for
5139 * rq, rq may still be transformed into a request with no icq, i.e., a
5140 * request not associated with any queue. No bfq hook is invoked to
5141 * signal this tranformation. As a consequence, should these
5142 * preparation operations be performed when the prepare_request hook
5143 * is invoked, and should rq be transformed one moment later, bfq
5144 * would end up in an inconsistent state, because it would have
5145 * incremented some queue counters for an rq destined to
5146 * transformation, without any chance to correctly lower these
5147 * counters back. In contrast, no transformation can still happen for
5148 * rq after rq has been inserted or merged. So, it is safe to execute
5149 * these preparation operations when rq is finally inserted or merged.
5151 static struct bfq_queue *bfq_init_rq(struct request *rq)
5153 struct request_queue *q = rq->q;
5154 struct bio *bio = rq->bio;
5155 struct bfq_data *bfqd = q->elevator->elevator_data;
5156 struct bfq_io_cq *bic;
5157 const int is_sync = rq_is_sync(rq);
5158 struct bfq_queue *bfqq;
5159 bool new_queue = false;
5160 bool bfqq_already_existing = false, split = false;
5162 if (unlikely(!rq->elv.icq))
5166 * Assuming that elv.priv[1] is set only if everything is set
5167 * for this rq. This holds true, because this function is
5168 * invoked only for insertion or merging, and, after such
5169 * events, a request cannot be manipulated any longer before
5170 * being removed from bfq.
5172 if (rq->elv.priv[1])
5173 return rq->elv.priv[1];
5175 bic = icq_to_bic(rq->elv.icq);
5177 bfq_check_ioprio_change(bic, bio);
5179 bfq_bic_update_cgroup(bic, bio);
5181 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
5184 if (likely(!new_queue)) {
5185 /* If the queue was seeky for too long, break it apart. */
5186 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
5187 bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
5189 /* Update bic before losing reference to bfqq */
5190 if (bfq_bfqq_in_large_burst(bfqq))
5191 bic->saved_in_large_burst = true;
5193 bfqq = bfq_split_bfqq(bic, bfqq);
5197 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
5201 bfqq_already_existing = true;
5207 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
5208 rq, bfqq, bfqq->ref);
5210 rq->elv.priv[0] = bic;
5211 rq->elv.priv[1] = bfqq;
5214 * If a bfq_queue has only one process reference, it is owned
5215 * by only this bic: we can then set bfqq->bic = bic. in
5216 * addition, if the queue has also just been split, we have to
5219 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
5223 * The queue has just been split from a shared
5224 * queue: restore the idle window and the
5225 * possible weight raising period.
5227 bfq_bfqq_resume_state(bfqq, bfqd, bic,
5228 bfqq_already_existing);
5232 if (unlikely(bfq_bfqq_just_created(bfqq)))
5233 bfq_handle_burst(bfqd, bfqq);
5239 bfq_idle_slice_timer_body(struct bfq_data *bfqd, struct bfq_queue *bfqq)
5241 enum bfqq_expiration reason;
5242 unsigned long flags;
5244 spin_lock_irqsave(&bfqd->lock, flags);
5247 * Considering that bfqq may be in race, we should firstly check
5248 * whether bfqq is in service before doing something on it. If
5249 * the bfqq in race is not in service, it has already been expired
5250 * through __bfq_bfqq_expire func and its wait_request flags has
5251 * been cleared in __bfq_bfqd_reset_in_service func.
5253 if (bfqq != bfqd->in_service_queue) {
5254 spin_unlock_irqrestore(&bfqd->lock, flags);
5258 bfq_clear_bfqq_wait_request(bfqq);
5260 if (bfq_bfqq_budget_timeout(bfqq))
5262 * Also here the queue can be safely expired
5263 * for budget timeout without wasting
5266 reason = BFQQE_BUDGET_TIMEOUT;
5267 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
5269 * The queue may not be empty upon timer expiration,
5270 * because we may not disable the timer when the
5271 * first request of the in-service queue arrives
5272 * during disk idling.
5274 reason = BFQQE_TOO_IDLE;
5276 goto schedule_dispatch;
5278 bfq_bfqq_expire(bfqd, bfqq, true, reason);
5281 spin_unlock_irqrestore(&bfqd->lock, flags);
5282 bfq_schedule_dispatch(bfqd);
5286 * Handler of the expiration of the timer running if the in-service queue
5287 * is idling inside its time slice.
5289 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
5291 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
5293 struct bfq_queue *bfqq = bfqd->in_service_queue;
5296 * Theoretical race here: the in-service queue can be NULL or
5297 * different from the queue that was idling if a new request
5298 * arrives for the current queue and there is a full dispatch
5299 * cycle that changes the in-service queue. This can hardly
5300 * happen, but in the worst case we just expire a queue too
5304 bfq_idle_slice_timer_body(bfqd, bfqq);
5306 return HRTIMER_NORESTART;
5309 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
5310 struct bfq_queue **bfqq_ptr)
5312 struct bfq_queue *bfqq = *bfqq_ptr;
5314 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
5316 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
5318 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
5320 bfq_put_queue(bfqq);
5326 * Release all the bfqg references to its async queues. If we are
5327 * deallocating the group these queues may still contain requests, so
5328 * we reparent them to the root cgroup (i.e., the only one that will
5329 * exist for sure until all the requests on a device are gone).
5331 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
5335 for (i = 0; i < 2; i++)
5336 for (j = 0; j < IOPRIO_BE_NR; j++)
5337 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
5339 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
5343 * See the comments on bfq_limit_depth for the purpose of
5344 * the depths set in the function. Return minimum shallow depth we'll use.
5346 static unsigned int bfq_update_depths(struct bfq_data *bfqd,
5347 struct sbitmap_queue *bt)
5349 unsigned int i, j, min_shallow = UINT_MAX;
5352 * In-word depths if no bfq_queue is being weight-raised:
5353 * leaving 25% of tags only for sync reads.
5355 * In next formulas, right-shift the value
5356 * (1U<<bt->sb.shift), instead of computing directly
5357 * (1U<<(bt->sb.shift - something)), to be robust against
5358 * any possible value of bt->sb.shift, without having to
5359 * limit 'something'.
5361 /* no more than 50% of tags for async I/O */
5362 bfqd->word_depths[0][0] = max((1U << bt->sb.shift) >> 1, 1U);
5364 * no more than 75% of tags for sync writes (25% extra tags
5365 * w.r.t. async I/O, to prevent async I/O from starving sync
5368 bfqd->word_depths[0][1] = max(((1U << bt->sb.shift) * 3) >> 2, 1U);
5371 * In-word depths in case some bfq_queue is being weight-
5372 * raised: leaving ~63% of tags for sync reads. This is the
5373 * highest percentage for which, in our tests, application
5374 * start-up times didn't suffer from any regression due to tag
5377 /* no more than ~18% of tags for async I/O */
5378 bfqd->word_depths[1][0] = max(((1U << bt->sb.shift) * 3) >> 4, 1U);
5379 /* no more than ~37% of tags for sync writes (~20% extra tags) */
5380 bfqd->word_depths[1][1] = max(((1U << bt->sb.shift) * 6) >> 4, 1U);
5382 for (i = 0; i < 2; i++)
5383 for (j = 0; j < 2; j++)
5384 min_shallow = min(min_shallow, bfqd->word_depths[i][j]);
5389 static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
5391 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5392 struct blk_mq_tags *tags = hctx->sched_tags;
5393 unsigned int min_shallow;
5395 min_shallow = bfq_update_depths(bfqd, &tags->bitmap_tags);
5396 sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, min_shallow);
5399 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
5401 bfq_depth_updated(hctx);
5405 static void bfq_exit_queue(struct elevator_queue *e)
5407 struct bfq_data *bfqd = e->elevator_data;
5408 struct bfq_queue *bfqq, *n;
5410 hrtimer_cancel(&bfqd->idle_slice_timer);
5412 spin_lock_irq(&bfqd->lock);
5413 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
5414 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
5415 spin_unlock_irq(&bfqd->lock);
5417 hrtimer_cancel(&bfqd->idle_slice_timer);
5419 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5420 /* release oom-queue reference to root group */
5421 bfqg_and_blkg_put(bfqd->root_group);
5423 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
5425 spin_lock_irq(&bfqd->lock);
5426 bfq_put_async_queues(bfqd, bfqd->root_group);
5427 kfree(bfqd->root_group);
5428 spin_unlock_irq(&bfqd->lock);
5431 wbt_enable_default(bfqd->queue);
5436 static void bfq_init_root_group(struct bfq_group *root_group,
5437 struct bfq_data *bfqd)
5441 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5442 root_group->entity.parent = NULL;
5443 root_group->my_entity = NULL;
5444 root_group->bfqd = bfqd;
5446 root_group->rq_pos_tree = RB_ROOT;
5447 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
5448 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
5449 root_group->sched_data.bfq_class_idle_last_service = jiffies;
5452 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
5454 struct bfq_data *bfqd;
5455 struct elevator_queue *eq;
5457 eq = elevator_alloc(q, e);
5461 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
5463 kobject_put(&eq->kobj);
5466 eq->elevator_data = bfqd;
5468 spin_lock_irq(q->queue_lock);
5470 spin_unlock_irq(q->queue_lock);
5473 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
5474 * Grab a permanent reference to it, so that the normal code flow
5475 * will not attempt to free it.
5477 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
5478 bfqd->oom_bfqq.ref++;
5479 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
5480 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
5481 bfqd->oom_bfqq.entity.new_weight =
5482 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
5484 /* oom_bfqq does not participate to bursts */
5485 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
5488 * Trigger weight initialization, according to ioprio, at the
5489 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
5490 * class won't be changed any more.
5492 bfqd->oom_bfqq.entity.prio_changed = 1;
5496 INIT_LIST_HEAD(&bfqd->dispatch);
5498 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
5500 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
5502 bfqd->queue_weights_tree = RB_ROOT;
5503 bfqd->num_groups_with_pending_reqs = 0;
5505 INIT_LIST_HEAD(&bfqd->active_list);
5506 INIT_LIST_HEAD(&bfqd->idle_list);
5507 INIT_HLIST_HEAD(&bfqd->burst_list);
5511 bfqd->bfq_max_budget = bfq_default_max_budget;
5513 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
5514 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
5515 bfqd->bfq_back_max = bfq_back_max;
5516 bfqd->bfq_back_penalty = bfq_back_penalty;
5517 bfqd->bfq_slice_idle = bfq_slice_idle;
5518 bfqd->bfq_timeout = bfq_timeout;
5520 bfqd->bfq_requests_within_timer = 120;
5522 bfqd->bfq_large_burst_thresh = 8;
5523 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
5525 bfqd->low_latency = true;
5528 * Trade-off between responsiveness and fairness.
5530 bfqd->bfq_wr_coeff = 30;
5531 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
5532 bfqd->bfq_wr_max_time = 0;
5533 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
5534 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
5535 bfqd->bfq_wr_max_softrt_rate = 7000; /*
5536 * Approximate rate required
5537 * to playback or record a
5538 * high-definition compressed
5541 bfqd->wr_busy_queues = 0;
5544 * Begin by assuming, optimistically, that the device peak
5545 * rate is equal to 2/3 of the highest reference rate.
5547 bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
5548 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
5549 bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
5551 spin_lock_init(&bfqd->lock);
5554 * The invocation of the next bfq_create_group_hierarchy
5555 * function is the head of a chain of function calls
5556 * (bfq_create_group_hierarchy->blkcg_activate_policy->
5557 * blk_mq_freeze_queue) that may lead to the invocation of the
5558 * has_work hook function. For this reason,
5559 * bfq_create_group_hierarchy is invoked only after all
5560 * scheduler data has been initialized, apart from the fields
5561 * that can be initialized only after invoking
5562 * bfq_create_group_hierarchy. This, in particular, enables
5563 * has_work to correctly return false. Of course, to avoid
5564 * other inconsistencies, the blk-mq stack must then refrain
5565 * from invoking further scheduler hooks before this init
5566 * function is finished.
5568 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
5569 if (!bfqd->root_group)
5571 bfq_init_root_group(bfqd->root_group, bfqd);
5572 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
5574 wbt_disable_default(q);
5579 kobject_put(&eq->kobj);
5583 static void bfq_slab_kill(void)
5585 kmem_cache_destroy(bfq_pool);
5588 static int __init bfq_slab_setup(void)
5590 bfq_pool = KMEM_CACHE(bfq_queue, 0);
5596 static ssize_t bfq_var_show(unsigned int var, char *page)
5598 return sprintf(page, "%u\n", var);
5601 static int bfq_var_store(unsigned long *var, const char *page)
5603 unsigned long new_val;
5604 int ret = kstrtoul(page, 10, &new_val);
5612 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
5613 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
5615 struct bfq_data *bfqd = e->elevator_data; \
5616 u64 __data = __VAR; \
5618 __data = jiffies_to_msecs(__data); \
5619 else if (__CONV == 2) \
5620 __data = div_u64(__data, NSEC_PER_MSEC); \
5621 return bfq_var_show(__data, (page)); \
5623 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
5624 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
5625 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
5626 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
5627 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
5628 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
5629 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
5630 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
5631 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
5632 #undef SHOW_FUNCTION
5634 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
5635 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
5637 struct bfq_data *bfqd = e->elevator_data; \
5638 u64 __data = __VAR; \
5639 __data = div_u64(__data, NSEC_PER_USEC); \
5640 return bfq_var_show(__data, (page)); \
5642 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
5643 #undef USEC_SHOW_FUNCTION
5645 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
5647 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
5649 struct bfq_data *bfqd = e->elevator_data; \
5650 unsigned long __data, __min = (MIN), __max = (MAX); \
5653 ret = bfq_var_store(&__data, (page)); \
5656 if (__data < __min) \
5658 else if (__data > __max) \
5661 *(__PTR) = msecs_to_jiffies(__data); \
5662 else if (__CONV == 2) \
5663 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
5665 *(__PTR) = __data; \
5668 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
5670 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
5672 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
5673 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
5675 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
5676 #undef STORE_FUNCTION
5678 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
5679 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
5681 struct bfq_data *bfqd = e->elevator_data; \
5682 unsigned long __data, __min = (MIN), __max = (MAX); \
5685 ret = bfq_var_store(&__data, (page)); \
5688 if (__data < __min) \
5690 else if (__data > __max) \
5692 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
5695 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
5697 #undef USEC_STORE_FUNCTION
5699 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
5700 const char *page, size_t count)
5702 struct bfq_data *bfqd = e->elevator_data;
5703 unsigned long __data;
5706 ret = bfq_var_store(&__data, (page));
5711 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5713 if (__data > INT_MAX)
5715 bfqd->bfq_max_budget = __data;
5718 bfqd->bfq_user_max_budget = __data;
5724 * Leaving this name to preserve name compatibility with cfq
5725 * parameters, but this timeout is used for both sync and async.
5727 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
5728 const char *page, size_t count)
5730 struct bfq_data *bfqd = e->elevator_data;
5731 unsigned long __data;
5734 ret = bfq_var_store(&__data, (page));
5740 else if (__data > INT_MAX)
5743 bfqd->bfq_timeout = msecs_to_jiffies(__data);
5744 if (bfqd->bfq_user_max_budget == 0)
5745 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5750 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
5751 const char *page, size_t count)
5753 struct bfq_data *bfqd = e->elevator_data;
5754 unsigned long __data;
5757 ret = bfq_var_store(&__data, (page));
5763 if (!bfqd->strict_guarantees && __data == 1
5764 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
5765 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
5767 bfqd->strict_guarantees = __data;
5772 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
5773 const char *page, size_t count)
5775 struct bfq_data *bfqd = e->elevator_data;
5776 unsigned long __data;
5779 ret = bfq_var_store(&__data, (page));
5785 if (__data == 0 && bfqd->low_latency != 0)
5787 bfqd->low_latency = __data;
5792 #define BFQ_ATTR(name) \
5793 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
5795 static struct elv_fs_entry bfq_attrs[] = {
5796 BFQ_ATTR(fifo_expire_sync),
5797 BFQ_ATTR(fifo_expire_async),
5798 BFQ_ATTR(back_seek_max),
5799 BFQ_ATTR(back_seek_penalty),
5800 BFQ_ATTR(slice_idle),
5801 BFQ_ATTR(slice_idle_us),
5802 BFQ_ATTR(max_budget),
5803 BFQ_ATTR(timeout_sync),
5804 BFQ_ATTR(strict_guarantees),
5805 BFQ_ATTR(low_latency),
5809 static struct elevator_type iosched_bfq_mq = {
5811 .limit_depth = bfq_limit_depth,
5812 .prepare_request = bfq_prepare_request,
5813 .requeue_request = bfq_finish_requeue_request,
5814 .finish_request = bfq_finish_requeue_request,
5815 .exit_icq = bfq_exit_icq,
5816 .insert_requests = bfq_insert_requests,
5817 .dispatch_request = bfq_dispatch_request,
5818 .next_request = elv_rb_latter_request,
5819 .former_request = elv_rb_former_request,
5820 .allow_merge = bfq_allow_bio_merge,
5821 .bio_merge = bfq_bio_merge,
5822 .request_merge = bfq_request_merge,
5823 .requests_merged = bfq_requests_merged,
5824 .request_merged = bfq_request_merged,
5825 .has_work = bfq_has_work,
5826 .depth_updated = bfq_depth_updated,
5827 .init_hctx = bfq_init_hctx,
5828 .init_sched = bfq_init_queue,
5829 .exit_sched = bfq_exit_queue,
5833 .icq_size = sizeof(struct bfq_io_cq),
5834 .icq_align = __alignof__(struct bfq_io_cq),
5835 .elevator_attrs = bfq_attrs,
5836 .elevator_name = "bfq",
5837 .elevator_owner = THIS_MODULE,
5839 MODULE_ALIAS("bfq-iosched");
5841 static int __init bfq_init(void)
5845 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5846 ret = blkcg_policy_register(&blkcg_policy_bfq);
5852 if (bfq_slab_setup())
5856 * Times to load large popular applications for the typical
5857 * systems installed on the reference devices (see the
5858 * comments before the definition of the next
5859 * array). Actually, we use slightly lower values, as the
5860 * estimated peak rate tends to be smaller than the actual
5861 * peak rate. The reason for this last fact is that estimates
5862 * are computed over much shorter time intervals than the long
5863 * intervals typically used for benchmarking. Why? First, to
5864 * adapt more quickly to variations. Second, because an I/O
5865 * scheduler cannot rely on a peak-rate-evaluation workload to
5866 * be run for a long time.
5868 ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
5869 ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
5871 ret = elv_register(&iosched_bfq_mq);
5880 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5881 blkcg_policy_unregister(&blkcg_policy_bfq);
5886 static void __exit bfq_exit(void)
5888 elv_unregister(&iosched_bfq_mq);
5889 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5890 blkcg_policy_unregister(&blkcg_policy_bfq);
5895 module_init(bfq_init);
5896 module_exit(bfq_exit);
5898 MODULE_AUTHOR("Paolo Valente");
5899 MODULE_LICENSE("GPL");
5900 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");