2 * Budget Fair Queueing (BFQ) I/O scheduler.
4 * Based on ideas and code from CFQ:
5 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
8 * Paolo Valente <paolo.valente@unimore.it>
10 * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
11 * Arianna Avanzini <avanzini@google.com>
13 * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License as
17 * published by the Free Software Foundation; either version 2 of the
18 * License, or (at your option) any later version.
20 * This program is distributed in the hope that it will be useful,
21 * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
23 * General Public License for more details.
25 * BFQ is a proportional-share I/O scheduler, with some extra
26 * low-latency capabilities. BFQ also supports full hierarchical
27 * scheduling through cgroups. Next paragraphs provide an introduction
28 * on BFQ inner workings. Details on BFQ benefits, usage and
29 * limitations can be found in Documentation/block/bfq-iosched.txt.
31 * BFQ is a proportional-share storage-I/O scheduling algorithm based
32 * on the slice-by-slice service scheme of CFQ. But BFQ assigns
33 * budgets, measured in number of sectors, to processes instead of
34 * time slices. The device is not granted to the in-service process
35 * for a given time slice, but until it has exhausted its assigned
36 * budget. This change from the time to the service domain enables BFQ
37 * to distribute the device throughput among processes as desired,
38 * without any distortion due to throughput fluctuations, or to device
39 * internal queueing. BFQ uses an ad hoc internal scheduler, called
40 * B-WF2Q+, to schedule processes according to their budgets. More
41 * precisely, BFQ schedules queues associated with processes. Each
42 * process/queue is assigned a user-configurable weight, and B-WF2Q+
43 * guarantees that each queue receives a fraction of the throughput
44 * proportional to its weight. Thanks to the accurate policy of
45 * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
46 * processes issuing sequential requests (to boost the throughput),
47 * and yet guarantee a low latency to interactive and soft real-time
50 * In particular, to provide these low-latency guarantees, BFQ
51 * explicitly privileges the I/O of two classes of time-sensitive
52 * applications: interactive and soft real-time. In more detail, BFQ
53 * behaves this way if the low_latency parameter is set (default
54 * configuration). This feature enables BFQ to provide applications in
55 * these classes with a very low latency.
57 * To implement this feature, BFQ constantly tries to detect whether
58 * the I/O requests in a bfq_queue come from an interactive or a soft
59 * real-time application. For brevity, in these cases, the queue is
60 * said to be interactive or soft real-time. In both cases, BFQ
61 * privileges the service of the queue, over that of non-interactive
62 * and non-soft-real-time queues. This privileging is performed,
63 * mainly, by raising the weight of the queue. So, for brevity, we
64 * call just weight-raising periods the time periods during which a
65 * queue is privileged, because deemed interactive or soft real-time.
67 * The detection of soft real-time queues/applications is described in
68 * detail in the comments on the function
69 * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
70 * interactive queue works as follows: a queue is deemed interactive
71 * if it is constantly non empty only for a limited time interval,
72 * after which it does become empty. The queue may be deemed
73 * interactive again (for a limited time), if it restarts being
74 * constantly non empty, provided that this happens only after the
75 * queue has remained empty for a given minimum idle time.
77 * By default, BFQ computes automatically the above maximum time
78 * interval, i.e., the time interval after which a constantly
79 * non-empty queue stops being deemed interactive. Since a queue is
80 * weight-raised while it is deemed interactive, this maximum time
81 * interval happens to coincide with the (maximum) duration of the
82 * weight-raising for interactive queues.
84 * Finally, BFQ also features additional heuristics for
85 * preserving both a low latency and a high throughput on NCQ-capable,
86 * rotational or flash-based devices, and to get the job done quickly
87 * for applications consisting in many I/O-bound processes.
89 * NOTE: if the main or only goal, with a given device, is to achieve
90 * the maximum-possible throughput at all times, then do switch off
91 * all low-latency heuristics for that device, by setting low_latency
94 * BFQ is described in [1], where also a reference to the initial,
95 * more theoretical paper on BFQ can be found. The interested reader
96 * can find in the latter paper full details on the main algorithm, as
97 * well as formulas of the guarantees and formal proofs of all the
98 * properties. With respect to the version of BFQ presented in these
99 * papers, this implementation adds a few more heuristics, such as the
100 * ones that guarantee a low latency to interactive and soft real-time
101 * applications, and a hierarchical extension based on H-WF2Q+.
103 * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
104 * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
105 * with O(log N) complexity derives from the one introduced with EEVDF
108 * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
109 * Scheduler", Proceedings of the First Workshop on Mobile System
110 * Technologies (MST-2015), May 2015.
111 * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
113 * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
114 * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
117 * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
119 * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
120 * First: A Flexible and Accurate Mechanism for Proportional Share
121 * Resource Allocation", technical report.
123 * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
125 #include <linux/module.h>
126 #include <linux/slab.h>
127 #include <linux/blkdev.h>
128 #include <linux/cgroup.h>
129 #include <linux/elevator.h>
130 #include <linux/ktime.h>
131 #include <linux/rbtree.h>
132 #include <linux/ioprio.h>
133 #include <linux/sbitmap.h>
134 #include <linux/delay.h>
138 #include "blk-mq-tag.h"
139 #include "blk-mq-sched.h"
140 #include "bfq-iosched.h"
143 #define BFQ_BFQQ_FNS(name) \
144 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
146 __set_bit(BFQQF_##name, &(bfqq)->flags); \
148 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
150 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
152 int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
154 return test_bit(BFQQF_##name, &(bfqq)->flags); \
157 BFQ_BFQQ_FNS(just_created);
159 BFQ_BFQQ_FNS(wait_request);
160 BFQ_BFQQ_FNS(non_blocking_wait_rq);
161 BFQ_BFQQ_FNS(fifo_expire);
162 BFQ_BFQQ_FNS(has_short_ttime);
164 BFQ_BFQQ_FNS(IO_bound);
165 BFQ_BFQQ_FNS(in_large_burst);
167 BFQ_BFQQ_FNS(split_coop);
168 BFQ_BFQQ_FNS(softrt_update);
169 #undef BFQ_BFQQ_FNS \
171 /* Expiration time of sync (0) and async (1) requests, in ns. */
172 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
174 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
175 static const int bfq_back_max = 16 * 1024;
177 /* Penalty of a backwards seek, in number of sectors. */
178 static const int bfq_back_penalty = 2;
180 /* Idling period duration, in ns. */
181 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
183 /* Minimum number of assigned budgets for which stats are safe to compute. */
184 static const int bfq_stats_min_budgets = 194;
186 /* Default maximum budget values, in sectors and number of requests. */
187 static const int bfq_default_max_budget = 16 * 1024;
190 * Async to sync throughput distribution is controlled as follows:
191 * when an async request is served, the entity is charged the number
192 * of sectors of the request, multiplied by the factor below
194 static const int bfq_async_charge_factor = 10;
196 /* Default timeout values, in jiffies, approximating CFQ defaults. */
197 const int bfq_timeout = HZ / 8;
200 * Time limit for merging (see comments in bfq_setup_cooperator). Set
201 * to the slowest value that, in our tests, proved to be effective in
202 * removing false positives, while not causing true positives to miss
205 * As can be deduced from the low time limit below, queue merging, if
206 * successful, happens at the very beggining of the I/O of the involved
207 * cooperating processes, as a consequence of the arrival of the very
208 * first requests from each cooperator. After that, there is very
209 * little chance to find cooperators.
211 static const unsigned long bfq_merge_time_limit = HZ/10;
213 static struct kmem_cache *bfq_pool;
215 /* Below this threshold (in ns), we consider thinktime immediate. */
216 #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
218 /* hw_tag detection: parallel requests threshold and min samples needed. */
219 #define BFQ_HW_QUEUE_THRESHOLD 4
220 #define BFQ_HW_QUEUE_SAMPLES 32
222 #define BFQQ_SEEK_THR (sector_t)(8 * 100)
223 #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
224 #define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
225 #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19)
227 /* Min number of samples required to perform peak-rate update */
228 #define BFQ_RATE_MIN_SAMPLES 32
229 /* Min observation time interval required to perform a peak-rate update (ns) */
230 #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
231 /* Target observation time interval for a peak-rate update (ns) */
232 #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
235 * Shift used for peak-rate fixed precision calculations.
237 * - the current shift: 16 positions
238 * - the current type used to store rate: u32
239 * - the current unit of measure for rate: [sectors/usec], or, more precisely,
240 * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
241 * the range of rates that can be stored is
242 * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
243 * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
244 * [15, 65G] sectors/sec
245 * Which, assuming a sector size of 512B, corresponds to a range of
248 #define BFQ_RATE_SHIFT 16
251 * When configured for computing the duration of the weight-raising
252 * for interactive queues automatically (see the comments at the
253 * beginning of this file), BFQ does it using the following formula:
254 * duration = (ref_rate / r) * ref_wr_duration,
255 * where r is the peak rate of the device, and ref_rate and
256 * ref_wr_duration are two reference parameters. In particular,
257 * ref_rate is the peak rate of the reference storage device (see
258 * below), and ref_wr_duration is about the maximum time needed, with
259 * BFQ and while reading two files in parallel, to load typical large
260 * applications on the reference device (see the comments on
261 * max_service_from_wr below, for more details on how ref_wr_duration
262 * is obtained). In practice, the slower/faster the device at hand
263 * is, the more/less it takes to load applications with respect to the
264 * reference device. Accordingly, the longer/shorter BFQ grants
265 * weight raising to interactive applications.
267 * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
268 * depending on whether the device is rotational or non-rotational.
270 * In the following definitions, ref_rate[0] and ref_wr_duration[0]
271 * are the reference values for a rotational device, whereas
272 * ref_rate[1] and ref_wr_duration[1] are the reference values for a
273 * non-rotational device. The reference rates are not the actual peak
274 * rates of the devices used as a reference, but slightly lower
275 * values. The reason for using slightly lower values is that the
276 * peak-rate estimator tends to yield slightly lower values than the
277 * actual peak rate (it can yield the actual peak rate only if there
278 * is only one process doing I/O, and the process does sequential
281 * The reference peak rates are measured in sectors/usec, left-shifted
284 static int ref_rate[2] = {14000, 33000};
286 * To improve readability, a conversion function is used to initialize
287 * the following array, which entails that the array can be
288 * initialized only in a function.
290 static int ref_wr_duration[2];
293 * BFQ uses the above-detailed, time-based weight-raising mechanism to
294 * privilege interactive tasks. This mechanism is vulnerable to the
295 * following false positives: I/O-bound applications that will go on
296 * doing I/O for much longer than the duration of weight
297 * raising. These applications have basically no benefit from being
298 * weight-raised at the beginning of their I/O. On the opposite end,
299 * while being weight-raised, these applications
300 * a) unjustly steal throughput to applications that may actually need
302 * b) make BFQ uselessly perform device idling; device idling results
303 * in loss of device throughput with most flash-based storage, and may
304 * increase latencies when used purposelessly.
306 * BFQ tries to reduce these problems, by adopting the following
307 * countermeasure. To introduce this countermeasure, we need first to
308 * finish explaining how the duration of weight-raising for
309 * interactive tasks is computed.
311 * For a bfq_queue deemed as interactive, the duration of weight
312 * raising is dynamically adjusted, as a function of the estimated
313 * peak rate of the device, so as to be equal to the time needed to
314 * execute the 'largest' interactive task we benchmarked so far. By
315 * largest task, we mean the task for which each involved process has
316 * to do more I/O than for any of the other tasks we benchmarked. This
317 * reference interactive task is the start-up of LibreOffice Writer,
318 * and in this task each process/bfq_queue needs to have at most ~110K
319 * sectors transferred.
321 * This last piece of information enables BFQ to reduce the actual
322 * duration of weight-raising for at least one class of I/O-bound
323 * applications: those doing sequential or quasi-sequential I/O. An
324 * example is file copy. In fact, once started, the main I/O-bound
325 * processes of these applications usually consume the above 110K
326 * sectors in much less time than the processes of an application that
327 * is starting, because these I/O-bound processes will greedily devote
328 * almost all their CPU cycles only to their target,
329 * throughput-friendly I/O operations. This is even more true if BFQ
330 * happens to be underestimating the device peak rate, and thus
331 * overestimating the duration of weight raising. But, according to
332 * our measurements, once transferred 110K sectors, these processes
333 * have no right to be weight-raised any longer.
335 * Basing on the last consideration, BFQ ends weight-raising for a
336 * bfq_queue if the latter happens to have received an amount of
337 * service at least equal to the following constant. The constant is
338 * set to slightly more than 110K, to have a minimum safety margin.
340 * This early ending of weight-raising reduces the amount of time
341 * during which interactive false positives cause the two problems
342 * described at the beginning of these comments.
344 static const unsigned long max_service_from_wr = 120000;
346 #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
347 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
349 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
351 return bic->bfqq[is_sync];
354 void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
356 bic->bfqq[is_sync] = bfqq;
359 struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
361 return bic->icq.q->elevator->elevator_data;
365 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
366 * @icq: the iocontext queue.
368 static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
370 /* bic->icq is the first member, %NULL will convert to %NULL */
371 return container_of(icq, struct bfq_io_cq, icq);
375 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
376 * @bfqd: the lookup key.
377 * @ioc: the io_context of the process doing I/O.
378 * @q: the request queue.
380 static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd,
381 struct io_context *ioc,
382 struct request_queue *q)
386 struct bfq_io_cq *icq;
388 spin_lock_irqsave(q->queue_lock, flags);
389 icq = icq_to_bic(ioc_lookup_icq(ioc, q));
390 spin_unlock_irqrestore(q->queue_lock, flags);
399 * Scheduler run of queue, if there are requests pending and no one in the
400 * driver that will restart queueing.
402 void bfq_schedule_dispatch(struct bfq_data *bfqd)
404 if (bfqd->queued != 0) {
405 bfq_log(bfqd, "schedule dispatch");
406 blk_mq_run_hw_queues(bfqd->queue, true);
410 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
411 #define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT)
413 #define bfq_sample_valid(samples) ((samples) > 80)
416 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
417 * We choose the request that is closesr to the head right now. Distance
418 * behind the head is penalized and only allowed to a certain extent.
420 static struct request *bfq_choose_req(struct bfq_data *bfqd,
425 sector_t s1, s2, d1 = 0, d2 = 0;
426 unsigned long back_max;
427 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
428 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
429 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
431 if (!rq1 || rq1 == rq2)
436 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
438 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
440 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
442 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
445 s1 = blk_rq_pos(rq1);
446 s2 = blk_rq_pos(rq2);
449 * By definition, 1KiB is 2 sectors.
451 back_max = bfqd->bfq_back_max * 2;
454 * Strict one way elevator _except_ in the case where we allow
455 * short backward seeks which are biased as twice the cost of a
456 * similar forward seek.
460 else if (s1 + back_max >= last)
461 d1 = (last - s1) * bfqd->bfq_back_penalty;
463 wrap |= BFQ_RQ1_WRAP;
467 else if (s2 + back_max >= last)
468 d2 = (last - s2) * bfqd->bfq_back_penalty;
470 wrap |= BFQ_RQ2_WRAP;
472 /* Found required data */
475 * By doing switch() on the bit mask "wrap" we avoid having to
476 * check two variables for all permutations: --> faster!
479 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
494 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
497 * Since both rqs are wrapped,
498 * start with the one that's further behind head
499 * (--> only *one* back seek required),
500 * since back seek takes more time than forward.
510 * Async I/O can easily starve sync I/O (both sync reads and sync
511 * writes), by consuming all tags. Similarly, storms of sync writes,
512 * such as those that sync(2) may trigger, can starve sync reads.
513 * Limit depths of async I/O and sync writes so as to counter both
516 static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
518 struct bfq_data *bfqd = data->q->elevator->elevator_data;
520 if (op_is_sync(op) && !op_is_write(op))
523 data->shallow_depth =
524 bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)];
526 bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
527 __func__, bfqd->wr_busy_queues, op_is_sync(op),
528 data->shallow_depth);
531 static struct bfq_queue *
532 bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
533 sector_t sector, struct rb_node **ret_parent,
534 struct rb_node ***rb_link)
536 struct rb_node **p, *parent;
537 struct bfq_queue *bfqq = NULL;
545 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
548 * Sort strictly based on sector. Smallest to the left,
549 * largest to the right.
551 if (sector > blk_rq_pos(bfqq->next_rq))
553 else if (sector < blk_rq_pos(bfqq->next_rq))
561 *ret_parent = parent;
565 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
566 (unsigned long long)sector,
567 bfqq ? bfqq->pid : 0);
572 static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
574 return bfqq->service_from_backlogged > 0 &&
575 time_is_before_jiffies(bfqq->first_IO_time +
576 bfq_merge_time_limit);
579 void bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
581 struct rb_node **p, *parent;
582 struct bfq_queue *__bfqq;
584 if (bfqq->pos_root) {
585 rb_erase(&bfqq->pos_node, bfqq->pos_root);
586 bfqq->pos_root = NULL;
590 * bfqq cannot be merged any longer (see comments in
591 * bfq_setup_cooperator): no point in adding bfqq into the
594 if (bfq_too_late_for_merging(bfqq))
597 if (bfq_class_idle(bfqq))
602 bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
603 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
604 blk_rq_pos(bfqq->next_rq), &parent, &p);
606 rb_link_node(&bfqq->pos_node, parent, p);
607 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
609 bfqq->pos_root = NULL;
613 * Tell whether there are active queues or groups with differentiated weights.
615 static bool bfq_differentiated_weights(struct bfq_data *bfqd)
618 * For weights to differ, at least one of the trees must contain
619 * at least two nodes.
621 return (!RB_EMPTY_ROOT(&bfqd->queue_weights_tree) &&
622 (bfqd->queue_weights_tree.rb_node->rb_left ||
623 bfqd->queue_weights_tree.rb_node->rb_right)
624 #ifdef CONFIG_BFQ_GROUP_IOSCHED
626 (!RB_EMPTY_ROOT(&bfqd->group_weights_tree) &&
627 (bfqd->group_weights_tree.rb_node->rb_left ||
628 bfqd->group_weights_tree.rb_node->rb_right)
634 * The following function returns true if every queue must receive the
635 * same share of the throughput (this condition is used when deciding
636 * whether idling may be disabled, see the comments in the function
637 * bfq_better_to_idle()).
639 * Such a scenario occurs when:
640 * 1) all active queues have the same weight,
641 * 2) all active groups at the same level in the groups tree have the same
643 * 3) all active groups at the same level in the groups tree have the same
644 * number of children.
646 * Unfortunately, keeping the necessary state for evaluating exactly the
647 * above symmetry conditions would be quite complex and time-consuming.
648 * Therefore this function evaluates, instead, the following stronger
649 * sub-conditions, for which it is much easier to maintain the needed
651 * 1) all active queues have the same weight,
652 * 2) all active groups have the same weight,
653 * 3) all active groups have at most one active child each.
654 * In particular, the last two conditions are always true if hierarchical
655 * support and the cgroups interface are not enabled, thus no state needs
656 * to be maintained in this case.
658 static bool bfq_symmetric_scenario(struct bfq_data *bfqd)
660 return !bfq_differentiated_weights(bfqd);
664 * If the weight-counter tree passed as input contains no counter for
665 * the weight of the input entity, then add that counter; otherwise just
666 * increment the existing counter.
668 * Note that weight-counter trees contain few nodes in mostly symmetric
669 * scenarios. For example, if all queues have the same weight, then the
670 * weight-counter tree for the queues may contain at most one node.
671 * This holds even if low_latency is on, because weight-raised queues
672 * are not inserted in the tree.
673 * In most scenarios, the rate at which nodes are created/destroyed
676 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_entity *entity,
677 struct rb_root *root)
679 struct rb_node **new = &(root->rb_node), *parent = NULL;
682 * Do not insert if the entity is already associated with a
683 * counter, which happens if:
684 * 1) the entity is associated with a queue,
685 * 2) a request arrival has caused the queue to become both
686 * non-weight-raised, and hence change its weight, and
687 * backlogged; in this respect, each of the two events
688 * causes an invocation of this function,
689 * 3) this is the invocation of this function caused by the
690 * second event. This second invocation is actually useless,
691 * and we handle this fact by exiting immediately. More
692 * efficient or clearer solutions might possibly be adopted.
694 if (entity->weight_counter)
698 struct bfq_weight_counter *__counter = container_of(*new,
699 struct bfq_weight_counter,
703 if (entity->weight == __counter->weight) {
704 entity->weight_counter = __counter;
707 if (entity->weight < __counter->weight)
708 new = &((*new)->rb_left);
710 new = &((*new)->rb_right);
713 entity->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
717 * In the unlucky event of an allocation failure, we just
718 * exit. This will cause the weight of entity to not be
719 * considered in bfq_differentiated_weights, which, in its
720 * turn, causes the scenario to be deemed wrongly symmetric in
721 * case entity's weight would have been the only weight making
722 * the scenario asymmetric. On the bright side, no unbalance
723 * will however occur when entity becomes inactive again (the
724 * invocation of this function is triggered by an activation
725 * of entity). In fact, bfq_weights_tree_remove does nothing
726 * if !entity->weight_counter.
728 if (unlikely(!entity->weight_counter))
731 entity->weight_counter->weight = entity->weight;
732 rb_link_node(&entity->weight_counter->weights_node, parent, new);
733 rb_insert_color(&entity->weight_counter->weights_node, root);
736 entity->weight_counter->num_active++;
740 * Decrement the weight counter associated with the entity, and, if the
741 * counter reaches 0, remove the counter from the tree.
742 * See the comments to the function bfq_weights_tree_add() for considerations
745 void __bfq_weights_tree_remove(struct bfq_data *bfqd,
746 struct bfq_entity *entity,
747 struct rb_root *root)
749 if (!entity->weight_counter)
752 entity->weight_counter->num_active--;
753 if (entity->weight_counter->num_active > 0)
754 goto reset_entity_pointer;
756 rb_erase(&entity->weight_counter->weights_node, root);
757 kfree(entity->weight_counter);
759 reset_entity_pointer:
760 entity->weight_counter = NULL;
764 * Invoke __bfq_weights_tree_remove on bfqq and all its inactive
767 void bfq_weights_tree_remove(struct bfq_data *bfqd,
768 struct bfq_queue *bfqq)
770 struct bfq_entity *entity = bfqq->entity.parent;
772 __bfq_weights_tree_remove(bfqd, &bfqq->entity,
773 &bfqd->queue_weights_tree);
775 for_each_entity(entity) {
776 struct bfq_sched_data *sd = entity->my_sched_data;
778 if (sd->next_in_service || sd->in_service_entity) {
780 * entity is still active, because either
781 * next_in_service or in_service_entity is not
782 * NULL (see the comments on the definition of
783 * next_in_service for details on why
784 * in_service_entity must be checked too).
786 * As a consequence, the weight of entity is
787 * not to be removed. In addition, if entity
788 * is active, then its parent entities are
789 * active as well, and thus their weights are
790 * not to be removed either. In the end, this
791 * loop must stop here.
795 __bfq_weights_tree_remove(bfqd, entity,
796 &bfqd->group_weights_tree);
801 * Return expired entry, or NULL to just start from scratch in rbtree.
803 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
804 struct request *last)
808 if (bfq_bfqq_fifo_expire(bfqq))
811 bfq_mark_bfqq_fifo_expire(bfqq);
813 rq = rq_entry_fifo(bfqq->fifo.next);
815 if (rq == last || ktime_get_ns() < rq->fifo_time)
818 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
822 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
823 struct bfq_queue *bfqq,
824 struct request *last)
826 struct rb_node *rbnext = rb_next(&last->rb_node);
827 struct rb_node *rbprev = rb_prev(&last->rb_node);
828 struct request *next, *prev = NULL;
830 /* Follow expired path, else get first next available. */
831 next = bfq_check_fifo(bfqq, last);
836 prev = rb_entry_rq(rbprev);
839 next = rb_entry_rq(rbnext);
841 rbnext = rb_first(&bfqq->sort_list);
842 if (rbnext && rbnext != &last->rb_node)
843 next = rb_entry_rq(rbnext);
846 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
849 /* see the definition of bfq_async_charge_factor for details */
850 static unsigned long bfq_serv_to_charge(struct request *rq,
851 struct bfq_queue *bfqq)
853 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1)
854 return blk_rq_sectors(rq);
857 * If there are no weight-raised queues, then amplify service
858 * by just the async charge factor; otherwise amplify service
859 * by twice the async charge factor, to further reduce latency
860 * for weight-raised queues.
862 if (bfqq->bfqd->wr_busy_queues == 0)
863 return blk_rq_sectors(rq) * bfq_async_charge_factor;
865 return blk_rq_sectors(rq) * 2 * bfq_async_charge_factor;
869 * bfq_updated_next_req - update the queue after a new next_rq selection.
870 * @bfqd: the device data the queue belongs to.
871 * @bfqq: the queue to update.
873 * If the first request of a queue changes we make sure that the queue
874 * has enough budget to serve at least its first request (if the
875 * request has grown). We do this because if the queue has not enough
876 * budget for its first request, it has to go through two dispatch
877 * rounds to actually get it dispatched.
879 static void bfq_updated_next_req(struct bfq_data *bfqd,
880 struct bfq_queue *bfqq)
882 struct bfq_entity *entity = &bfqq->entity;
883 struct request *next_rq = bfqq->next_rq;
884 unsigned long new_budget;
889 if (bfqq == bfqd->in_service_queue)
891 * In order not to break guarantees, budgets cannot be
892 * changed after an entity has been selected.
896 new_budget = max_t(unsigned long, bfqq->max_budget,
897 bfq_serv_to_charge(next_rq, bfqq));
898 if (entity->budget != new_budget) {
899 entity->budget = new_budget;
900 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
902 bfq_requeue_bfqq(bfqd, bfqq, false);
906 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
910 if (bfqd->bfq_wr_max_time > 0)
911 return bfqd->bfq_wr_max_time;
913 dur = bfqd->rate_dur_prod;
914 do_div(dur, bfqd->peak_rate);
917 * Limit duration between 3 and 25 seconds. The upper limit
918 * has been conservatively set after the following worst case:
919 * on a QEMU/KVM virtual machine
920 * - running in a slow PC
921 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
922 * - serving a heavy I/O workload, such as the sequential reading
924 * mplayer took 23 seconds to start, if constantly weight-raised.
926 * As for higher values than that accomodating the above bad
927 * scenario, tests show that higher values would often yield
928 * the opposite of the desired result, i.e., would worsen
929 * responsiveness by allowing non-interactive applications to
930 * preserve weight raising for too long.
932 * On the other end, lower values than 3 seconds make it
933 * difficult for most interactive tasks to complete their jobs
934 * before weight-raising finishes.
936 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
939 /* switch back from soft real-time to interactive weight raising */
940 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
941 struct bfq_data *bfqd)
943 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
944 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
945 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
949 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
950 struct bfq_io_cq *bic, bool bfq_already_existing)
952 unsigned int old_wr_coeff = bfqq->wr_coeff;
953 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
955 if (bic->saved_has_short_ttime)
956 bfq_mark_bfqq_has_short_ttime(bfqq);
958 bfq_clear_bfqq_has_short_ttime(bfqq);
960 if (bic->saved_IO_bound)
961 bfq_mark_bfqq_IO_bound(bfqq);
963 bfq_clear_bfqq_IO_bound(bfqq);
965 bfqq->ttime = bic->saved_ttime;
966 bfqq->wr_coeff = bic->saved_wr_coeff;
967 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
968 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
969 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
971 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
972 time_is_before_jiffies(bfqq->last_wr_start_finish +
973 bfqq->wr_cur_max_time))) {
974 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
975 !bfq_bfqq_in_large_burst(bfqq) &&
976 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
977 bfq_wr_duration(bfqd))) {
978 switch_back_to_interactive_wr(bfqq, bfqd);
981 bfq_log_bfqq(bfqq->bfqd, bfqq,
982 "resume state: switching off wr");
986 /* make sure weight will be updated, however we got here */
987 bfqq->entity.prio_changed = 1;
992 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
993 bfqd->wr_busy_queues++;
994 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
995 bfqd->wr_busy_queues--;
998 static int bfqq_process_refs(struct bfq_queue *bfqq)
1000 return bfqq->ref - bfqq->allocated - bfqq->entity.on_st;
1003 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1004 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1006 struct bfq_queue *item;
1007 struct hlist_node *n;
1009 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1010 hlist_del_init(&item->burst_list_node);
1011 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1012 bfqd->burst_size = 1;
1013 bfqd->burst_parent_entity = bfqq->entity.parent;
1016 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1017 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1019 /* Increment burst size to take into account also bfqq */
1022 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1023 struct bfq_queue *pos, *bfqq_item;
1024 struct hlist_node *n;
1027 * Enough queues have been activated shortly after each
1028 * other to consider this burst as large.
1030 bfqd->large_burst = true;
1033 * We can now mark all queues in the burst list as
1034 * belonging to a large burst.
1036 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1038 bfq_mark_bfqq_in_large_burst(bfqq_item);
1039 bfq_mark_bfqq_in_large_burst(bfqq);
1042 * From now on, and until the current burst finishes, any
1043 * new queue being activated shortly after the last queue
1044 * was inserted in the burst can be immediately marked as
1045 * belonging to a large burst. So the burst list is not
1046 * needed any more. Remove it.
1048 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1050 hlist_del_init(&pos->burst_list_node);
1052 * Burst not yet large: add bfqq to the burst list. Do
1053 * not increment the ref counter for bfqq, because bfqq
1054 * is removed from the burst list before freeing bfqq
1057 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1061 * If many queues belonging to the same group happen to be created
1062 * shortly after each other, then the processes associated with these
1063 * queues have typically a common goal. In particular, bursts of queue
1064 * creations are usually caused by services or applications that spawn
1065 * many parallel threads/processes. Examples are systemd during boot,
1066 * or git grep. To help these processes get their job done as soon as
1067 * possible, it is usually better to not grant either weight-raising
1068 * or device idling to their queues.
1070 * In this comment we describe, firstly, the reasons why this fact
1071 * holds, and, secondly, the next function, which implements the main
1072 * steps needed to properly mark these queues so that they can then be
1073 * treated in a different way.
1075 * The above services or applications benefit mostly from a high
1076 * throughput: the quicker the requests of the activated queues are
1077 * cumulatively served, the sooner the target job of these queues gets
1078 * completed. As a consequence, weight-raising any of these queues,
1079 * which also implies idling the device for it, is almost always
1080 * counterproductive. In most cases it just lowers throughput.
1082 * On the other hand, a burst of queue creations may be caused also by
1083 * the start of an application that does not consist of a lot of
1084 * parallel I/O-bound threads. In fact, with a complex application,
1085 * several short processes may need to be executed to start-up the
1086 * application. In this respect, to start an application as quickly as
1087 * possible, the best thing to do is in any case to privilege the I/O
1088 * related to the application with respect to all other
1089 * I/O. Therefore, the best strategy to start as quickly as possible
1090 * an application that causes a burst of queue creations is to
1091 * weight-raise all the queues created during the burst. This is the
1092 * exact opposite of the best strategy for the other type of bursts.
1094 * In the end, to take the best action for each of the two cases, the
1095 * two types of bursts need to be distinguished. Fortunately, this
1096 * seems relatively easy, by looking at the sizes of the bursts. In
1097 * particular, we found a threshold such that only bursts with a
1098 * larger size than that threshold are apparently caused by
1099 * services or commands such as systemd or git grep. For brevity,
1100 * hereafter we call just 'large' these bursts. BFQ *does not*
1101 * weight-raise queues whose creation occurs in a large burst. In
1102 * addition, for each of these queues BFQ performs or does not perform
1103 * idling depending on which choice boosts the throughput more. The
1104 * exact choice depends on the device and request pattern at
1107 * Unfortunately, false positives may occur while an interactive task
1108 * is starting (e.g., an application is being started). The
1109 * consequence is that the queues associated with the task do not
1110 * enjoy weight raising as expected. Fortunately these false positives
1111 * are very rare. They typically occur if some service happens to
1112 * start doing I/O exactly when the interactive task starts.
1114 * Turning back to the next function, it implements all the steps
1115 * needed to detect the occurrence of a large burst and to properly
1116 * mark all the queues belonging to it (so that they can then be
1117 * treated in a different way). This goal is achieved by maintaining a
1118 * "burst list" that holds, temporarily, the queues that belong to the
1119 * burst in progress. The list is then used to mark these queues as
1120 * belonging to a large burst if the burst does become large. The main
1121 * steps are the following.
1123 * . when the very first queue is created, the queue is inserted into the
1124 * list (as it could be the first queue in a possible burst)
1126 * . if the current burst has not yet become large, and a queue Q that does
1127 * not yet belong to the burst is activated shortly after the last time
1128 * at which a new queue entered the burst list, then the function appends
1129 * Q to the burst list
1131 * . if, as a consequence of the previous step, the burst size reaches
1132 * the large-burst threshold, then
1134 * . all the queues in the burst list are marked as belonging to a
1137 * . the burst list is deleted; in fact, the burst list already served
1138 * its purpose (keeping temporarily track of the queues in a burst,
1139 * so as to be able to mark them as belonging to a large burst in the
1140 * previous sub-step), and now is not needed any more
1142 * . the device enters a large-burst mode
1144 * . if a queue Q that does not belong to the burst is created while
1145 * the device is in large-burst mode and shortly after the last time
1146 * at which a queue either entered the burst list or was marked as
1147 * belonging to the current large burst, then Q is immediately marked
1148 * as belonging to a large burst.
1150 * . if a queue Q that does not belong to the burst is created a while
1151 * later, i.e., not shortly after, than the last time at which a queue
1152 * either entered the burst list or was marked as belonging to the
1153 * current large burst, then the current burst is deemed as finished and:
1155 * . the large-burst mode is reset if set
1157 * . the burst list is emptied
1159 * . Q is inserted in the burst list, as Q may be the first queue
1160 * in a possible new burst (then the burst list contains just Q
1163 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1166 * If bfqq is already in the burst list or is part of a large
1167 * burst, or finally has just been split, then there is
1168 * nothing else to do.
1170 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1171 bfq_bfqq_in_large_burst(bfqq) ||
1172 time_is_after_eq_jiffies(bfqq->split_time +
1173 msecs_to_jiffies(10)))
1177 * If bfqq's creation happens late enough, or bfqq belongs to
1178 * a different group than the burst group, then the current
1179 * burst is finished, and related data structures must be
1182 * In this respect, consider the special case where bfqq is
1183 * the very first queue created after BFQ is selected for this
1184 * device. In this case, last_ins_in_burst and
1185 * burst_parent_entity are not yet significant when we get
1186 * here. But it is easy to verify that, whether or not the
1187 * following condition is true, bfqq will end up being
1188 * inserted into the burst list. In particular the list will
1189 * happen to contain only bfqq. And this is exactly what has
1190 * to happen, as bfqq may be the first queue of the first
1193 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1194 bfqd->bfq_burst_interval) ||
1195 bfqq->entity.parent != bfqd->burst_parent_entity) {
1196 bfqd->large_burst = false;
1197 bfq_reset_burst_list(bfqd, bfqq);
1202 * If we get here, then bfqq is being activated shortly after the
1203 * last queue. So, if the current burst is also large, we can mark
1204 * bfqq as belonging to this large burst immediately.
1206 if (bfqd->large_burst) {
1207 bfq_mark_bfqq_in_large_burst(bfqq);
1212 * If we get here, then a large-burst state has not yet been
1213 * reached, but bfqq is being activated shortly after the last
1214 * queue. Then we add bfqq to the burst.
1216 bfq_add_to_burst(bfqd, bfqq);
1219 * At this point, bfqq either has been added to the current
1220 * burst or has caused the current burst to terminate and a
1221 * possible new burst to start. In particular, in the second
1222 * case, bfqq has become the first queue in the possible new
1223 * burst. In both cases last_ins_in_burst needs to be moved
1226 bfqd->last_ins_in_burst = jiffies;
1229 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1231 struct bfq_entity *entity = &bfqq->entity;
1233 return entity->budget - entity->service;
1237 * If enough samples have been computed, return the current max budget
1238 * stored in bfqd, which is dynamically updated according to the
1239 * estimated disk peak rate; otherwise return the default max budget
1241 static int bfq_max_budget(struct bfq_data *bfqd)
1243 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1244 return bfq_default_max_budget;
1246 return bfqd->bfq_max_budget;
1250 * Return min budget, which is a fraction of the current or default
1251 * max budget (trying with 1/32)
1253 static int bfq_min_budget(struct bfq_data *bfqd)
1255 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1256 return bfq_default_max_budget / 32;
1258 return bfqd->bfq_max_budget / 32;
1262 * The next function, invoked after the input queue bfqq switches from
1263 * idle to busy, updates the budget of bfqq. The function also tells
1264 * whether the in-service queue should be expired, by returning
1265 * true. The purpose of expiring the in-service queue is to give bfqq
1266 * the chance to possibly preempt the in-service queue, and the reason
1267 * for preempting the in-service queue is to achieve one of the two
1270 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1271 * expired because it has remained idle. In particular, bfqq may have
1272 * expired for one of the following two reasons:
1274 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1275 * and did not make it to issue a new request before its last
1276 * request was served;
1278 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1279 * a new request before the expiration of the idling-time.
1281 * Even if bfqq has expired for one of the above reasons, the process
1282 * associated with the queue may be however issuing requests greedily,
1283 * and thus be sensitive to the bandwidth it receives (bfqq may have
1284 * remained idle for other reasons: CPU high load, bfqq not enjoying
1285 * idling, I/O throttling somewhere in the path from the process to
1286 * the I/O scheduler, ...). But if, after every expiration for one of
1287 * the above two reasons, bfqq has to wait for the service of at least
1288 * one full budget of another queue before being served again, then
1289 * bfqq is likely to get a much lower bandwidth or resource time than
1290 * its reserved ones. To address this issue, two countermeasures need
1293 * First, the budget and the timestamps of bfqq need to be updated in
1294 * a special way on bfqq reactivation: they need to be updated as if
1295 * bfqq did not remain idle and did not expire. In fact, if they are
1296 * computed as if bfqq expired and remained idle until reactivation,
1297 * then the process associated with bfqq is treated as if, instead of
1298 * being greedy, it stopped issuing requests when bfqq remained idle,
1299 * and restarts issuing requests only on this reactivation. In other
1300 * words, the scheduler does not help the process recover the "service
1301 * hole" between bfqq expiration and reactivation. As a consequence,
1302 * the process receives a lower bandwidth than its reserved one. In
1303 * contrast, to recover this hole, the budget must be updated as if
1304 * bfqq was not expired at all before this reactivation, i.e., it must
1305 * be set to the value of the remaining budget when bfqq was
1306 * expired. Along the same line, timestamps need to be assigned the
1307 * value they had the last time bfqq was selected for service, i.e.,
1308 * before last expiration. Thus timestamps need to be back-shifted
1309 * with respect to their normal computation (see [1] for more details
1310 * on this tricky aspect).
1312 * Secondly, to allow the process to recover the hole, the in-service
1313 * queue must be expired too, to give bfqq the chance to preempt it
1314 * immediately. In fact, if bfqq has to wait for a full budget of the
1315 * in-service queue to be completed, then it may become impossible to
1316 * let the process recover the hole, even if the back-shifted
1317 * timestamps of bfqq are lower than those of the in-service queue. If
1318 * this happens for most or all of the holes, then the process may not
1319 * receive its reserved bandwidth. In this respect, it is worth noting
1320 * that, being the service of outstanding requests unpreemptible, a
1321 * little fraction of the holes may however be unrecoverable, thereby
1322 * causing a little loss of bandwidth.
1324 * The last important point is detecting whether bfqq does need this
1325 * bandwidth recovery. In this respect, the next function deems the
1326 * process associated with bfqq greedy, and thus allows it to recover
1327 * the hole, if: 1) the process is waiting for the arrival of a new
1328 * request (which implies that bfqq expired for one of the above two
1329 * reasons), and 2) such a request has arrived soon. The first
1330 * condition is controlled through the flag non_blocking_wait_rq,
1331 * while the second through the flag arrived_in_time. If both
1332 * conditions hold, then the function computes the budget in the
1333 * above-described special way, and signals that the in-service queue
1334 * should be expired. Timestamp back-shifting is done later in
1335 * __bfq_activate_entity.
1337 * 2. Reduce latency. Even if timestamps are not backshifted to let
1338 * the process associated with bfqq recover a service hole, bfqq may
1339 * however happen to have, after being (re)activated, a lower finish
1340 * timestamp than the in-service queue. That is, the next budget of
1341 * bfqq may have to be completed before the one of the in-service
1342 * queue. If this is the case, then preempting the in-service queue
1343 * allows this goal to be achieved, apart from the unpreemptible,
1344 * outstanding requests mentioned above.
1346 * Unfortunately, regardless of which of the above two goals one wants
1347 * to achieve, service trees need first to be updated to know whether
1348 * the in-service queue must be preempted. To have service trees
1349 * correctly updated, the in-service queue must be expired and
1350 * rescheduled, and bfqq must be scheduled too. This is one of the
1351 * most costly operations (in future versions, the scheduling
1352 * mechanism may be re-designed in such a way to make it possible to
1353 * know whether preemption is needed without needing to update service
1354 * trees). In addition, queue preemptions almost always cause random
1355 * I/O, and thus loss of throughput. Because of these facts, the next
1356 * function adopts the following simple scheme to avoid both costly
1357 * operations and too frequent preemptions: it requests the expiration
1358 * of the in-service queue (unconditionally) only for queues that need
1359 * to recover a hole, or that either are weight-raised or deserve to
1362 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1363 struct bfq_queue *bfqq,
1364 bool arrived_in_time,
1365 bool wr_or_deserves_wr)
1367 struct bfq_entity *entity = &bfqq->entity;
1369 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time) {
1371 * We do not clear the flag non_blocking_wait_rq here, as
1372 * the latter is used in bfq_activate_bfqq to signal
1373 * that timestamps need to be back-shifted (and is
1374 * cleared right after).
1378 * In next assignment we rely on that either
1379 * entity->service or entity->budget are not updated
1380 * on expiration if bfqq is empty (see
1381 * __bfq_bfqq_recalc_budget). Thus both quantities
1382 * remain unchanged after such an expiration, and the
1383 * following statement therefore assigns to
1384 * entity->budget the remaining budget on such an
1387 entity->budget = min_t(unsigned long,
1388 bfq_bfqq_budget_left(bfqq),
1392 * At this point, we have used entity->service to get
1393 * the budget left (needed for updating
1394 * entity->budget). Thus we finally can, and have to,
1395 * reset entity->service. The latter must be reset
1396 * because bfqq would otherwise be charged again for
1397 * the service it has received during its previous
1400 entity->service = 0;
1406 * We can finally complete expiration, by setting service to 0.
1408 entity->service = 0;
1409 entity->budget = max_t(unsigned long, bfqq->max_budget,
1410 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1411 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1412 return wr_or_deserves_wr;
1416 * Return the farthest past time instant according to jiffies
1419 static unsigned long bfq_smallest_from_now(void)
1421 return jiffies - MAX_JIFFY_OFFSET;
1424 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1425 struct bfq_queue *bfqq,
1426 unsigned int old_wr_coeff,
1427 bool wr_or_deserves_wr,
1432 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1433 /* start a weight-raising period */
1435 bfqq->service_from_wr = 0;
1436 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1437 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1440 * No interactive weight raising in progress
1441 * here: assign minus infinity to
1442 * wr_start_at_switch_to_srt, to make sure
1443 * that, at the end of the soft-real-time
1444 * weight raising periods that is starting
1445 * now, no interactive weight-raising period
1446 * may be wrongly considered as still in
1447 * progress (and thus actually started by
1450 bfqq->wr_start_at_switch_to_srt =
1451 bfq_smallest_from_now();
1452 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1453 BFQ_SOFTRT_WEIGHT_FACTOR;
1454 bfqq->wr_cur_max_time =
1455 bfqd->bfq_wr_rt_max_time;
1459 * If needed, further reduce budget to make sure it is
1460 * close to bfqq's backlog, so as to reduce the
1461 * scheduling-error component due to a too large
1462 * budget. Do not care about throughput consequences,
1463 * but only about latency. Finally, do not assign a
1464 * too small budget either, to avoid increasing
1465 * latency by causing too frequent expirations.
1467 bfqq->entity.budget = min_t(unsigned long,
1468 bfqq->entity.budget,
1469 2 * bfq_min_budget(bfqd));
1470 } else if (old_wr_coeff > 1) {
1471 if (interactive) { /* update wr coeff and duration */
1472 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1473 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1474 } else if (in_burst)
1478 * The application is now or still meeting the
1479 * requirements for being deemed soft rt. We
1480 * can then correctly and safely (re)charge
1481 * the weight-raising duration for the
1482 * application with the weight-raising
1483 * duration for soft rt applications.
1485 * In particular, doing this recharge now, i.e.,
1486 * before the weight-raising period for the
1487 * application finishes, reduces the probability
1488 * of the following negative scenario:
1489 * 1) the weight of a soft rt application is
1490 * raised at startup (as for any newly
1491 * created application),
1492 * 2) since the application is not interactive,
1493 * at a certain time weight-raising is
1494 * stopped for the application,
1495 * 3) at that time the application happens to
1496 * still have pending requests, and hence
1497 * is destined to not have a chance to be
1498 * deemed soft rt before these requests are
1499 * completed (see the comments to the
1500 * function bfq_bfqq_softrt_next_start()
1501 * for details on soft rt detection),
1502 * 4) these pending requests experience a high
1503 * latency because the application is not
1504 * weight-raised while they are pending.
1506 if (bfqq->wr_cur_max_time !=
1507 bfqd->bfq_wr_rt_max_time) {
1508 bfqq->wr_start_at_switch_to_srt =
1509 bfqq->last_wr_start_finish;
1511 bfqq->wr_cur_max_time =
1512 bfqd->bfq_wr_rt_max_time;
1513 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1514 BFQ_SOFTRT_WEIGHT_FACTOR;
1516 bfqq->last_wr_start_finish = jiffies;
1521 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1522 struct bfq_queue *bfqq)
1524 return bfqq->dispatched == 0 &&
1525 time_is_before_jiffies(
1526 bfqq->budget_timeout +
1527 bfqd->bfq_wr_min_idle_time);
1530 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1531 struct bfq_queue *bfqq,
1536 bool soft_rt, in_burst, wr_or_deserves_wr,
1537 bfqq_wants_to_preempt,
1538 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1540 * See the comments on
1541 * bfq_bfqq_update_budg_for_activation for
1542 * details on the usage of the next variable.
1544 arrived_in_time = ktime_get_ns() <=
1545 bfqq->ttime.last_end_request +
1546 bfqd->bfq_slice_idle * 3;
1550 * bfqq deserves to be weight-raised if:
1552 * - it does not belong to a large burst,
1553 * - it has been idle for enough time or is soft real-time,
1554 * - is linked to a bfq_io_cq (it is not shared in any sense).
1556 in_burst = bfq_bfqq_in_large_burst(bfqq);
1557 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1559 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1560 bfqq->dispatched == 0;
1561 *interactive = !in_burst && idle_for_long_time;
1562 wr_or_deserves_wr = bfqd->low_latency &&
1563 (bfqq->wr_coeff > 1 ||
1564 (bfq_bfqq_sync(bfqq) &&
1565 bfqq->bic && (*interactive || soft_rt)));
1568 * Using the last flag, update budget and check whether bfqq
1569 * may want to preempt the in-service queue.
1571 bfqq_wants_to_preempt =
1572 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1577 * If bfqq happened to be activated in a burst, but has been
1578 * idle for much more than an interactive queue, then we
1579 * assume that, in the overall I/O initiated in the burst, the
1580 * I/O associated with bfqq is finished. So bfqq does not need
1581 * to be treated as a queue belonging to a burst
1582 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1583 * if set, and remove bfqq from the burst list if it's
1584 * there. We do not decrement burst_size, because the fact
1585 * that bfqq does not need to belong to the burst list any
1586 * more does not invalidate the fact that bfqq was created in
1589 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1590 idle_for_long_time &&
1591 time_is_before_jiffies(
1592 bfqq->budget_timeout +
1593 msecs_to_jiffies(10000))) {
1594 hlist_del_init(&bfqq->burst_list_node);
1595 bfq_clear_bfqq_in_large_burst(bfqq);
1598 bfq_clear_bfqq_just_created(bfqq);
1601 if (!bfq_bfqq_IO_bound(bfqq)) {
1602 if (arrived_in_time) {
1603 bfqq->requests_within_timer++;
1604 if (bfqq->requests_within_timer >=
1605 bfqd->bfq_requests_within_timer)
1606 bfq_mark_bfqq_IO_bound(bfqq);
1608 bfqq->requests_within_timer = 0;
1611 if (bfqd->low_latency) {
1612 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1615 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1617 if (time_is_before_jiffies(bfqq->split_time +
1618 bfqd->bfq_wr_min_idle_time)) {
1619 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1626 if (old_wr_coeff != bfqq->wr_coeff)
1627 bfqq->entity.prio_changed = 1;
1631 bfqq->last_idle_bklogged = jiffies;
1632 bfqq->service_from_backlogged = 0;
1633 bfq_clear_bfqq_softrt_update(bfqq);
1635 bfq_add_bfqq_busy(bfqd, bfqq);
1638 * Expire in-service queue only if preemption may be needed
1639 * for guarantees. In this respect, the function
1640 * next_queue_may_preempt just checks a simple, necessary
1641 * condition, and not a sufficient condition based on
1642 * timestamps. In fact, for the latter condition to be
1643 * evaluated, timestamps would need first to be updated, and
1644 * this operation is quite costly (see the comments on the
1645 * function bfq_bfqq_update_budg_for_activation).
1647 if (bfqd->in_service_queue && bfqq_wants_to_preempt &&
1648 bfqd->in_service_queue->wr_coeff < bfqq->wr_coeff &&
1649 next_queue_may_preempt(bfqd))
1650 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1651 false, BFQQE_PREEMPTED);
1654 static void bfq_add_request(struct request *rq)
1656 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1657 struct bfq_data *bfqd = bfqq->bfqd;
1658 struct request *next_rq, *prev;
1659 unsigned int old_wr_coeff = bfqq->wr_coeff;
1660 bool interactive = false;
1662 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1663 bfqq->queued[rq_is_sync(rq)]++;
1666 elv_rb_add(&bfqq->sort_list, rq);
1669 * Check if this request is a better next-serve candidate.
1671 prev = bfqq->next_rq;
1672 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
1673 bfqq->next_rq = next_rq;
1676 * Adjust priority tree position, if next_rq changes.
1678 if (prev != bfqq->next_rq)
1679 bfq_pos_tree_add_move(bfqd, bfqq);
1681 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
1682 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
1685 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
1686 time_is_before_jiffies(
1687 bfqq->last_wr_start_finish +
1688 bfqd->bfq_wr_min_inter_arr_async)) {
1689 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1690 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1692 bfqd->wr_busy_queues++;
1693 bfqq->entity.prio_changed = 1;
1695 if (prev != bfqq->next_rq)
1696 bfq_updated_next_req(bfqd, bfqq);
1700 * Assign jiffies to last_wr_start_finish in the following
1703 * . if bfqq is not going to be weight-raised, because, for
1704 * non weight-raised queues, last_wr_start_finish stores the
1705 * arrival time of the last request; as of now, this piece
1706 * of information is used only for deciding whether to
1707 * weight-raise async queues
1709 * . if bfqq is not weight-raised, because, if bfqq is now
1710 * switching to weight-raised, then last_wr_start_finish
1711 * stores the time when weight-raising starts
1713 * . if bfqq is interactive, because, regardless of whether
1714 * bfqq is currently weight-raised, the weight-raising
1715 * period must start or restart (this case is considered
1716 * separately because it is not detected by the above
1717 * conditions, if bfqq is already weight-raised)
1719 * last_wr_start_finish has to be updated also if bfqq is soft
1720 * real-time, because the weight-raising period is constantly
1721 * restarted on idle-to-busy transitions for these queues, but
1722 * this is already done in bfq_bfqq_handle_idle_busy_switch if
1725 if (bfqd->low_latency &&
1726 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
1727 bfqq->last_wr_start_finish = jiffies;
1730 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
1732 struct request_queue *q)
1734 struct bfq_queue *bfqq = bfqd->bio_bfqq;
1738 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
1743 static sector_t get_sdist(sector_t last_pos, struct request *rq)
1746 return abs(blk_rq_pos(rq) - last_pos);
1751 #if 0 /* Still not clear if we can do without next two functions */
1752 static void bfq_activate_request(struct request_queue *q, struct request *rq)
1754 struct bfq_data *bfqd = q->elevator->elevator_data;
1756 bfqd->rq_in_driver++;
1759 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
1761 struct bfq_data *bfqd = q->elevator->elevator_data;
1763 bfqd->rq_in_driver--;
1767 static void bfq_remove_request(struct request_queue *q,
1770 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1771 struct bfq_data *bfqd = bfqq->bfqd;
1772 const int sync = rq_is_sync(rq);
1774 if (bfqq->next_rq == rq) {
1775 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
1776 bfq_updated_next_req(bfqd, bfqq);
1779 if (rq->queuelist.prev != &rq->queuelist)
1780 list_del_init(&rq->queuelist);
1781 bfqq->queued[sync]--;
1783 elv_rb_del(&bfqq->sort_list, rq);
1785 elv_rqhash_del(q, rq);
1786 if (q->last_merge == rq)
1787 q->last_merge = NULL;
1789 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
1790 bfqq->next_rq = NULL;
1792 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
1793 bfq_del_bfqq_busy(bfqd, bfqq, false);
1795 * bfqq emptied. In normal operation, when
1796 * bfqq is empty, bfqq->entity.service and
1797 * bfqq->entity.budget must contain,
1798 * respectively, the service received and the
1799 * budget used last time bfqq emptied. These
1800 * facts do not hold in this case, as at least
1801 * this last removal occurred while bfqq is
1802 * not in service. To avoid inconsistencies,
1803 * reset both bfqq->entity.service and
1804 * bfqq->entity.budget, if bfqq has still a
1805 * process that may issue I/O requests to it.
1807 bfqq->entity.budget = bfqq->entity.service = 0;
1811 * Remove queue from request-position tree as it is empty.
1813 if (bfqq->pos_root) {
1814 rb_erase(&bfqq->pos_node, bfqq->pos_root);
1815 bfqq->pos_root = NULL;
1818 bfq_pos_tree_add_move(bfqd, bfqq);
1821 if (rq->cmd_flags & REQ_META)
1822 bfqq->meta_pending--;
1826 static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio)
1828 struct request_queue *q = hctx->queue;
1829 struct bfq_data *bfqd = q->elevator->elevator_data;
1830 struct request *free = NULL;
1832 * bfq_bic_lookup grabs the queue_lock: invoke it now and
1833 * store its return value for later use, to avoid nesting
1834 * queue_lock inside the bfqd->lock. We assume that the bic
1835 * returned by bfq_bic_lookup does not go away before
1836 * bfqd->lock is taken.
1838 struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
1841 spin_lock_irq(&bfqd->lock);
1844 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
1846 bfqd->bio_bfqq = NULL;
1847 bfqd->bio_bic = bic;
1849 ret = blk_mq_sched_try_merge(q, bio, &free);
1852 blk_mq_free_request(free);
1853 spin_unlock_irq(&bfqd->lock);
1858 static int bfq_request_merge(struct request_queue *q, struct request **req,
1861 struct bfq_data *bfqd = q->elevator->elevator_data;
1862 struct request *__rq;
1864 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
1865 if (__rq && elv_bio_merge_ok(__rq, bio)) {
1867 return ELEVATOR_FRONT_MERGE;
1870 return ELEVATOR_NO_MERGE;
1873 static struct bfq_queue *bfq_init_rq(struct request *rq);
1875 static void bfq_request_merged(struct request_queue *q, struct request *req,
1876 enum elv_merge type)
1878 if (type == ELEVATOR_FRONT_MERGE &&
1879 rb_prev(&req->rb_node) &&
1881 blk_rq_pos(container_of(rb_prev(&req->rb_node),
1882 struct request, rb_node))) {
1883 struct bfq_queue *bfqq = bfq_init_rq(req);
1884 struct bfq_data *bfqd = bfqq->bfqd;
1885 struct request *prev, *next_rq;
1887 /* Reposition request in its sort_list */
1888 elv_rb_del(&bfqq->sort_list, req);
1889 elv_rb_add(&bfqq->sort_list, req);
1891 /* Choose next request to be served for bfqq */
1892 prev = bfqq->next_rq;
1893 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
1894 bfqd->last_position);
1895 bfqq->next_rq = next_rq;
1897 * If next_rq changes, update both the queue's budget to
1898 * fit the new request and the queue's position in its
1901 if (prev != bfqq->next_rq) {
1902 bfq_updated_next_req(bfqd, bfqq);
1903 bfq_pos_tree_add_move(bfqd, bfqq);
1909 * This function is called to notify the scheduler that the requests
1910 * rq and 'next' have been merged, with 'next' going away. BFQ
1911 * exploits this hook to address the following issue: if 'next' has a
1912 * fifo_time lower that rq, then the fifo_time of rq must be set to
1913 * the value of 'next', to not forget the greater age of 'next'.
1915 * NOTE: in this function we assume that rq is in a bfq_queue, basing
1916 * on that rq is picked from the hash table q->elevator->hash, which,
1917 * in its turn, is filled only with I/O requests present in
1918 * bfq_queues, while BFQ is in use for the request queue q. In fact,
1919 * the function that fills this hash table (elv_rqhash_add) is called
1920 * only by bfq_insert_request.
1922 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
1923 struct request *next)
1925 struct bfq_queue *bfqq = bfq_init_rq(rq),
1926 *next_bfqq = bfq_init_rq(next);
1929 * If next and rq belong to the same bfq_queue and next is older
1930 * than rq, then reposition rq in the fifo (by substituting next
1931 * with rq). Otherwise, if next and rq belong to different
1932 * bfq_queues, never reposition rq: in fact, we would have to
1933 * reposition it with respect to next's position in its own fifo,
1934 * which would most certainly be too expensive with respect to
1937 if (bfqq == next_bfqq &&
1938 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1939 next->fifo_time < rq->fifo_time) {
1940 list_del_init(&rq->queuelist);
1941 list_replace_init(&next->queuelist, &rq->queuelist);
1942 rq->fifo_time = next->fifo_time;
1945 if (bfqq->next_rq == next)
1948 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
1951 /* Must be called with bfqq != NULL */
1952 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
1954 if (bfq_bfqq_busy(bfqq))
1955 bfqq->bfqd->wr_busy_queues--;
1957 bfqq->wr_cur_max_time = 0;
1958 bfqq->last_wr_start_finish = jiffies;
1960 * Trigger a weight change on the next invocation of
1961 * __bfq_entity_update_weight_prio.
1963 bfqq->entity.prio_changed = 1;
1966 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
1967 struct bfq_group *bfqg)
1971 for (i = 0; i < 2; i++)
1972 for (j = 0; j < IOPRIO_BE_NR; j++)
1973 if (bfqg->async_bfqq[i][j])
1974 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
1975 if (bfqg->async_idle_bfqq)
1976 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
1979 static void bfq_end_wr(struct bfq_data *bfqd)
1981 struct bfq_queue *bfqq;
1983 spin_lock_irq(&bfqd->lock);
1985 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
1986 bfq_bfqq_end_wr(bfqq);
1987 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
1988 bfq_bfqq_end_wr(bfqq);
1989 bfq_end_wr_async(bfqd);
1991 spin_unlock_irq(&bfqd->lock);
1994 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
1997 return blk_rq_pos(io_struct);
1999 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2002 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2005 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2009 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2010 struct bfq_queue *bfqq,
2013 struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
2014 struct rb_node *parent, *node;
2015 struct bfq_queue *__bfqq;
2017 if (RB_EMPTY_ROOT(root))
2021 * First, if we find a request starting at the end of the last
2022 * request, choose it.
2024 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2029 * If the exact sector wasn't found, the parent of the NULL leaf
2030 * will contain the closest sector (rq_pos_tree sorted by
2031 * next_request position).
2033 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2034 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2037 if (blk_rq_pos(__bfqq->next_rq) < sector)
2038 node = rb_next(&__bfqq->pos_node);
2040 node = rb_prev(&__bfqq->pos_node);
2044 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2045 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2051 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2052 struct bfq_queue *cur_bfqq,
2055 struct bfq_queue *bfqq;
2058 * We shall notice if some of the queues are cooperating,
2059 * e.g., working closely on the same area of the device. In
2060 * that case, we can group them together and: 1) don't waste
2061 * time idling, and 2) serve the union of their requests in
2062 * the best possible order for throughput.
2064 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2065 if (!bfqq || bfqq == cur_bfqq)
2071 static struct bfq_queue *
2072 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2074 int process_refs, new_process_refs;
2075 struct bfq_queue *__bfqq;
2078 * If there are no process references on the new_bfqq, then it is
2079 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2080 * may have dropped their last reference (not just their last process
2083 if (!bfqq_process_refs(new_bfqq))
2086 /* Avoid a circular list and skip interim queue merges. */
2087 while ((__bfqq = new_bfqq->new_bfqq)) {
2093 process_refs = bfqq_process_refs(bfqq);
2094 new_process_refs = bfqq_process_refs(new_bfqq);
2096 * If the process for the bfqq has gone away, there is no
2097 * sense in merging the queues.
2099 if (process_refs == 0 || new_process_refs == 0)
2102 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2106 * Merging is just a redirection: the requests of the process
2107 * owning one of the two queues are redirected to the other queue.
2108 * The latter queue, in its turn, is set as shared if this is the
2109 * first time that the requests of some process are redirected to
2112 * We redirect bfqq to new_bfqq and not the opposite, because
2113 * we are in the context of the process owning bfqq, thus we
2114 * have the io_cq of this process. So we can immediately
2115 * configure this io_cq to redirect the requests of the
2116 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2117 * not available any more (new_bfqq->bic == NULL).
2119 * Anyway, even in case new_bfqq coincides with the in-service
2120 * queue, redirecting requests the in-service queue is the
2121 * best option, as we feed the in-service queue with new
2122 * requests close to the last request served and, by doing so,
2123 * are likely to increase the throughput.
2125 bfqq->new_bfqq = new_bfqq;
2126 new_bfqq->ref += process_refs;
2130 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2131 struct bfq_queue *new_bfqq)
2133 if (bfq_too_late_for_merging(new_bfqq))
2136 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2137 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2141 * If either of the queues has already been detected as seeky,
2142 * then merging it with the other queue is unlikely to lead to
2145 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2149 * Interleaved I/O is known to be done by (some) applications
2150 * only for reads, so it does not make sense to merge async
2153 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2160 * Attempt to schedule a merge of bfqq with the currently in-service
2161 * queue or with a close queue among the scheduled queues. Return
2162 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2163 * structure otherwise.
2165 * The OOM queue is not allowed to participate to cooperation: in fact, since
2166 * the requests temporarily redirected to the OOM queue could be redirected
2167 * again to dedicated queues at any time, the state needed to correctly
2168 * handle merging with the OOM queue would be quite complex and expensive
2169 * to maintain. Besides, in such a critical condition as an out of memory,
2170 * the benefits of queue merging may be little relevant, or even negligible.
2172 * WARNING: queue merging may impair fairness among non-weight raised
2173 * queues, for at least two reasons: 1) the original weight of a
2174 * merged queue may change during the merged state, 2) even being the
2175 * weight the same, a merged queue may be bloated with many more
2176 * requests than the ones produced by its originally-associated
2179 static struct bfq_queue *
2180 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2181 void *io_struct, bool request)
2183 struct bfq_queue *in_service_bfqq, *new_bfqq;
2186 * Prevent bfqq from being merged if it has been created too
2187 * long ago. The idea is that true cooperating processes, and
2188 * thus their associated bfq_queues, are supposed to be
2189 * created shortly after each other. This is the case, e.g.,
2190 * for KVM/QEMU and dump I/O threads. Basing on this
2191 * assumption, the following filtering greatly reduces the
2192 * probability that two non-cooperating processes, which just
2193 * happen to do close I/O for some short time interval, have
2194 * their queues merged by mistake.
2196 if (bfq_too_late_for_merging(bfqq))
2200 return bfqq->new_bfqq;
2202 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2205 /* If there is only one backlogged queue, don't search. */
2206 if (bfqd->busy_queues == 1)
2209 in_service_bfqq = bfqd->in_service_queue;
2211 if (in_service_bfqq && in_service_bfqq != bfqq &&
2212 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2213 bfq_rq_close_to_sector(io_struct, request, bfqd->last_position) &&
2214 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2215 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2216 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2221 * Check whether there is a cooperator among currently scheduled
2222 * queues. The only thing we need is that the bio/request is not
2223 * NULL, as we need it to establish whether a cooperator exists.
2225 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2226 bfq_io_struct_pos(io_struct, request));
2228 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2229 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2230 return bfq_setup_merge(bfqq, new_bfqq);
2235 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2237 struct bfq_io_cq *bic = bfqq->bic;
2240 * If !bfqq->bic, the queue is already shared or its requests
2241 * have already been redirected to a shared queue; both idle window
2242 * and weight raising state have already been saved. Do nothing.
2247 bic->saved_ttime = bfqq->ttime;
2248 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2249 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2250 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2251 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2252 if (unlikely(bfq_bfqq_just_created(bfqq) &&
2253 !bfq_bfqq_in_large_burst(bfqq) &&
2254 bfqq->bfqd->low_latency)) {
2256 * bfqq being merged right after being created: bfqq
2257 * would have deserved interactive weight raising, but
2258 * did not make it to be set in a weight-raised state,
2259 * because of this early merge. Store directly the
2260 * weight-raising state that would have been assigned
2261 * to bfqq, so that to avoid that bfqq unjustly fails
2262 * to enjoy weight raising if split soon.
2264 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
2265 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2266 bic->saved_last_wr_start_finish = jiffies;
2268 bic->saved_wr_coeff = bfqq->wr_coeff;
2269 bic->saved_wr_start_at_switch_to_srt =
2270 bfqq->wr_start_at_switch_to_srt;
2271 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2272 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2277 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2278 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2280 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2281 (unsigned long)new_bfqq->pid);
2282 /* Save weight raising and idle window of the merged queues */
2283 bfq_bfqq_save_state(bfqq);
2284 bfq_bfqq_save_state(new_bfqq);
2285 if (bfq_bfqq_IO_bound(bfqq))
2286 bfq_mark_bfqq_IO_bound(new_bfqq);
2287 bfq_clear_bfqq_IO_bound(bfqq);
2290 * If bfqq is weight-raised, then let new_bfqq inherit
2291 * weight-raising. To reduce false positives, neglect the case
2292 * where bfqq has just been created, but has not yet made it
2293 * to be weight-raised (which may happen because EQM may merge
2294 * bfqq even before bfq_add_request is executed for the first
2295 * time for bfqq). Handling this case would however be very
2296 * easy, thanks to the flag just_created.
2298 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2299 new_bfqq->wr_coeff = bfqq->wr_coeff;
2300 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2301 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2302 new_bfqq->wr_start_at_switch_to_srt =
2303 bfqq->wr_start_at_switch_to_srt;
2304 if (bfq_bfqq_busy(new_bfqq))
2305 bfqd->wr_busy_queues++;
2306 new_bfqq->entity.prio_changed = 1;
2309 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2311 bfqq->entity.prio_changed = 1;
2312 if (bfq_bfqq_busy(bfqq))
2313 bfqd->wr_busy_queues--;
2316 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2317 bfqd->wr_busy_queues);
2320 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2322 bic_set_bfqq(bic, new_bfqq, 1);
2323 bfq_mark_bfqq_coop(new_bfqq);
2325 * new_bfqq now belongs to at least two bics (it is a shared queue):
2326 * set new_bfqq->bic to NULL. bfqq either:
2327 * - does not belong to any bic any more, and hence bfqq->bic must
2328 * be set to NULL, or
2329 * - is a queue whose owning bics have already been redirected to a
2330 * different queue, hence the queue is destined to not belong to
2331 * any bic soon and bfqq->bic is already NULL (therefore the next
2332 * assignment causes no harm).
2334 new_bfqq->bic = NULL;
2336 /* release process reference to bfqq */
2337 bfq_put_queue(bfqq);
2340 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2343 struct bfq_data *bfqd = q->elevator->elevator_data;
2344 bool is_sync = op_is_sync(bio->bi_opf);
2345 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2348 * Disallow merge of a sync bio into an async request.
2350 if (is_sync && !rq_is_sync(rq))
2354 * Lookup the bfqq that this bio will be queued with. Allow
2355 * merge only if rq is queued there.
2361 * We take advantage of this function to perform an early merge
2362 * of the queues of possible cooperating processes.
2364 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2367 * bic still points to bfqq, then it has not yet been
2368 * redirected to some other bfq_queue, and a queue
2369 * merge beween bfqq and new_bfqq can be safely
2370 * fulfillled, i.e., bic can be redirected to new_bfqq
2371 * and bfqq can be put.
2373 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2376 * If we get here, bio will be queued into new_queue,
2377 * so use new_bfqq to decide whether bio and rq can be
2383 * Change also bqfd->bio_bfqq, as
2384 * bfqd->bio_bic now points to new_bfqq, and
2385 * this function may be invoked again (and then may
2386 * use again bqfd->bio_bfqq).
2388 bfqd->bio_bfqq = bfqq;
2391 return bfqq == RQ_BFQQ(rq);
2395 * Set the maximum time for the in-service queue to consume its
2396 * budget. This prevents seeky processes from lowering the throughput.
2397 * In practice, a time-slice service scheme is used with seeky
2400 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2401 struct bfq_queue *bfqq)
2403 unsigned int timeout_coeff;
2405 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2408 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2410 bfqd->last_budget_start = ktime_get();
2412 bfqq->budget_timeout = jiffies +
2413 bfqd->bfq_timeout * timeout_coeff;
2416 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2417 struct bfq_queue *bfqq)
2420 bfq_clear_bfqq_fifo_expire(bfqq);
2422 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2424 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2425 bfqq->wr_coeff > 1 &&
2426 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2427 time_is_before_jiffies(bfqq->budget_timeout)) {
2429 * For soft real-time queues, move the start
2430 * of the weight-raising period forward by the
2431 * time the queue has not received any
2432 * service. Otherwise, a relatively long
2433 * service delay is likely to cause the
2434 * weight-raising period of the queue to end,
2435 * because of the short duration of the
2436 * weight-raising period of a soft real-time
2437 * queue. It is worth noting that this move
2438 * is not so dangerous for the other queues,
2439 * because soft real-time queues are not
2442 * To not add a further variable, we use the
2443 * overloaded field budget_timeout to
2444 * determine for how long the queue has not
2445 * received service, i.e., how much time has
2446 * elapsed since the queue expired. However,
2447 * this is a little imprecise, because
2448 * budget_timeout is set to jiffies if bfqq
2449 * not only expires, but also remains with no
2452 if (time_after(bfqq->budget_timeout,
2453 bfqq->last_wr_start_finish))
2454 bfqq->last_wr_start_finish +=
2455 jiffies - bfqq->budget_timeout;
2457 bfqq->last_wr_start_finish = jiffies;
2460 bfq_set_budget_timeout(bfqd, bfqq);
2461 bfq_log_bfqq(bfqd, bfqq,
2462 "set_in_service_queue, cur-budget = %d",
2463 bfqq->entity.budget);
2466 bfqd->in_service_queue = bfqq;
2470 * Get and set a new queue for service.
2472 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2474 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2476 __bfq_set_in_service_queue(bfqd, bfqq);
2480 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2482 struct bfq_queue *bfqq = bfqd->in_service_queue;
2485 bfq_mark_bfqq_wait_request(bfqq);
2488 * We don't want to idle for seeks, but we do want to allow
2489 * fair distribution of slice time for a process doing back-to-back
2490 * seeks. So allow a little bit of time for him to submit a new rq.
2492 sl = bfqd->bfq_slice_idle;
2494 * Unless the queue is being weight-raised or the scenario is
2495 * asymmetric, grant only minimum idle time if the queue
2496 * is seeky. A long idling is preserved for a weight-raised
2497 * queue, or, more in general, in an asymmetric scenario,
2498 * because a long idling is needed for guaranteeing to a queue
2499 * its reserved share of the throughput (in particular, it is
2500 * needed if the queue has a higher weight than some other
2503 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
2504 bfq_symmetric_scenario(bfqd))
2505 sl = min_t(u64, sl, BFQ_MIN_TT);
2507 bfqd->last_idling_start = ktime_get();
2508 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
2510 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
2514 * In autotuning mode, max_budget is dynamically recomputed as the
2515 * amount of sectors transferred in timeout at the estimated peak
2516 * rate. This enables BFQ to utilize a full timeslice with a full
2517 * budget, even if the in-service queue is served at peak rate. And
2518 * this maximises throughput with sequential workloads.
2520 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
2522 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
2523 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
2527 * Update parameters related to throughput and responsiveness, as a
2528 * function of the estimated peak rate. See comments on
2529 * bfq_calc_max_budget(), and on the ref_wr_duration array.
2531 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
2533 if (bfqd->bfq_user_max_budget == 0) {
2534 bfqd->bfq_max_budget =
2535 bfq_calc_max_budget(bfqd);
2536 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
2540 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
2543 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
2544 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
2545 bfqd->peak_rate_samples = 1;
2546 bfqd->sequential_samples = 0;
2547 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
2549 } else /* no new rq dispatched, just reset the number of samples */
2550 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
2553 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
2554 bfqd->peak_rate_samples, bfqd->sequential_samples,
2555 bfqd->tot_sectors_dispatched);
2558 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
2560 u32 rate, weight, divisor;
2563 * For the convergence property to hold (see comments on
2564 * bfq_update_peak_rate()) and for the assessment to be
2565 * reliable, a minimum number of samples must be present, and
2566 * a minimum amount of time must have elapsed. If not so, do
2567 * not compute new rate. Just reset parameters, to get ready
2568 * for a new evaluation attempt.
2570 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
2571 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
2572 goto reset_computation;
2575 * If a new request completion has occurred after last
2576 * dispatch, then, to approximate the rate at which requests
2577 * have been served by the device, it is more precise to
2578 * extend the observation interval to the last completion.
2580 bfqd->delta_from_first =
2581 max_t(u64, bfqd->delta_from_first,
2582 bfqd->last_completion - bfqd->first_dispatch);
2585 * Rate computed in sects/usec, and not sects/nsec, for
2588 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
2589 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
2592 * Peak rate not updated if:
2593 * - the percentage of sequential dispatches is below 3/4 of the
2594 * total, and rate is below the current estimated peak rate
2595 * - rate is unreasonably high (> 20M sectors/sec)
2597 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
2598 rate <= bfqd->peak_rate) ||
2599 rate > 20<<BFQ_RATE_SHIFT)
2600 goto reset_computation;
2603 * We have to update the peak rate, at last! To this purpose,
2604 * we use a low-pass filter. We compute the smoothing constant
2605 * of the filter as a function of the 'weight' of the new
2608 * As can be seen in next formulas, we define this weight as a
2609 * quantity proportional to how sequential the workload is,
2610 * and to how long the observation time interval is.
2612 * The weight runs from 0 to 8. The maximum value of the
2613 * weight, 8, yields the minimum value for the smoothing
2614 * constant. At this minimum value for the smoothing constant,
2615 * the measured rate contributes for half of the next value of
2616 * the estimated peak rate.
2618 * So, the first step is to compute the weight as a function
2619 * of how sequential the workload is. Note that the weight
2620 * cannot reach 9, because bfqd->sequential_samples cannot
2621 * become equal to bfqd->peak_rate_samples, which, in its
2622 * turn, holds true because bfqd->sequential_samples is not
2623 * incremented for the first sample.
2625 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
2628 * Second step: further refine the weight as a function of the
2629 * duration of the observation interval.
2631 weight = min_t(u32, 8,
2632 div_u64(weight * bfqd->delta_from_first,
2633 BFQ_RATE_REF_INTERVAL));
2636 * Divisor ranging from 10, for minimum weight, to 2, for
2639 divisor = 10 - weight;
2642 * Finally, update peak rate:
2644 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
2646 bfqd->peak_rate *= divisor-1;
2647 bfqd->peak_rate /= divisor;
2648 rate /= divisor; /* smoothing constant alpha = 1/divisor */
2650 bfqd->peak_rate += rate;
2653 * For a very slow device, bfqd->peak_rate can reach 0 (see
2654 * the minimum representable values reported in the comments
2655 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
2656 * divisions by zero where bfqd->peak_rate is used as a
2659 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
2661 update_thr_responsiveness_params(bfqd);
2664 bfq_reset_rate_computation(bfqd, rq);
2668 * Update the read/write peak rate (the main quantity used for
2669 * auto-tuning, see update_thr_responsiveness_params()).
2671 * It is not trivial to estimate the peak rate (correctly): because of
2672 * the presence of sw and hw queues between the scheduler and the
2673 * device components that finally serve I/O requests, it is hard to
2674 * say exactly when a given dispatched request is served inside the
2675 * device, and for how long. As a consequence, it is hard to know
2676 * precisely at what rate a given set of requests is actually served
2679 * On the opposite end, the dispatch time of any request is trivially
2680 * available, and, from this piece of information, the "dispatch rate"
2681 * of requests can be immediately computed. So, the idea in the next
2682 * function is to use what is known, namely request dispatch times
2683 * (plus, when useful, request completion times), to estimate what is
2684 * unknown, namely in-device request service rate.
2686 * The main issue is that, because of the above facts, the rate at
2687 * which a certain set of requests is dispatched over a certain time
2688 * interval can vary greatly with respect to the rate at which the
2689 * same requests are then served. But, since the size of any
2690 * intermediate queue is limited, and the service scheme is lossless
2691 * (no request is silently dropped), the following obvious convergence
2692 * property holds: the number of requests dispatched MUST become
2693 * closer and closer to the number of requests completed as the
2694 * observation interval grows. This is the key property used in
2695 * the next function to estimate the peak service rate as a function
2696 * of the observed dispatch rate. The function assumes to be invoked
2697 * on every request dispatch.
2699 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
2701 u64 now_ns = ktime_get_ns();
2703 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
2704 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
2705 bfqd->peak_rate_samples);
2706 bfq_reset_rate_computation(bfqd, rq);
2707 goto update_last_values; /* will add one sample */
2711 * Device idle for very long: the observation interval lasting
2712 * up to this dispatch cannot be a valid observation interval
2713 * for computing a new peak rate (similarly to the late-
2714 * completion event in bfq_completed_request()). Go to
2715 * update_rate_and_reset to have the following three steps
2717 * - close the observation interval at the last (previous)
2718 * request dispatch or completion
2719 * - compute rate, if possible, for that observation interval
2720 * - start a new observation interval with this dispatch
2722 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
2723 bfqd->rq_in_driver == 0)
2724 goto update_rate_and_reset;
2726 /* Update sampling information */
2727 bfqd->peak_rate_samples++;
2729 if ((bfqd->rq_in_driver > 0 ||
2730 now_ns - bfqd->last_completion < BFQ_MIN_TT)
2731 && get_sdist(bfqd->last_position, rq) < BFQQ_SEEK_THR)
2732 bfqd->sequential_samples++;
2734 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
2736 /* Reset max observed rq size every 32 dispatches */
2737 if (likely(bfqd->peak_rate_samples % 32))
2738 bfqd->last_rq_max_size =
2739 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
2741 bfqd->last_rq_max_size = blk_rq_sectors(rq);
2743 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
2745 /* Target observation interval not yet reached, go on sampling */
2746 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
2747 goto update_last_values;
2749 update_rate_and_reset:
2750 bfq_update_rate_reset(bfqd, rq);
2752 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2753 bfqd->last_dispatch = now_ns;
2757 * Remove request from internal lists.
2759 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
2761 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2764 * For consistency, the next instruction should have been
2765 * executed after removing the request from the queue and
2766 * dispatching it. We execute instead this instruction before
2767 * bfq_remove_request() (and hence introduce a temporary
2768 * inconsistency), for efficiency. In fact, should this
2769 * dispatch occur for a non in-service bfqq, this anticipated
2770 * increment prevents two counters related to bfqq->dispatched
2771 * from risking to be, first, uselessly decremented, and then
2772 * incremented again when the (new) value of bfqq->dispatched
2773 * happens to be taken into account.
2776 bfq_update_peak_rate(q->elevator->elevator_data, rq);
2778 bfq_remove_request(q, rq);
2781 static void __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq)
2784 * If this bfqq is shared between multiple processes, check
2785 * to make sure that those processes are still issuing I/Os
2786 * within the mean seek distance. If not, it may be time to
2787 * break the queues apart again.
2789 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
2790 bfq_mark_bfqq_split_coop(bfqq);
2792 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2793 if (bfqq->dispatched == 0)
2795 * Overloading budget_timeout field to store
2796 * the time at which the queue remains with no
2797 * backlog and no outstanding request; used by
2798 * the weight-raising mechanism.
2800 bfqq->budget_timeout = jiffies;
2802 bfq_del_bfqq_busy(bfqd, bfqq, true);
2804 bfq_requeue_bfqq(bfqd, bfqq, true);
2806 * Resort priority tree of potential close cooperators.
2808 bfq_pos_tree_add_move(bfqd, bfqq);
2812 * All in-service entities must have been properly deactivated
2813 * or requeued before executing the next function, which
2814 * resets all in-service entites as no more in service.
2816 __bfq_bfqd_reset_in_service(bfqd);
2820 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
2821 * @bfqd: device data.
2822 * @bfqq: queue to update.
2823 * @reason: reason for expiration.
2825 * Handle the feedback on @bfqq budget at queue expiration.
2826 * See the body for detailed comments.
2828 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
2829 struct bfq_queue *bfqq,
2830 enum bfqq_expiration reason)
2832 struct request *next_rq;
2833 int budget, min_budget;
2835 min_budget = bfq_min_budget(bfqd);
2837 if (bfqq->wr_coeff == 1)
2838 budget = bfqq->max_budget;
2840 * Use a constant, low budget for weight-raised queues,
2841 * to help achieve a low latency. Keep it slightly higher
2842 * than the minimum possible budget, to cause a little
2843 * bit fewer expirations.
2845 budget = 2 * min_budget;
2847 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
2848 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
2849 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
2850 budget, bfq_min_budget(bfqd));
2851 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
2852 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
2854 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
2857 * Caveat: in all the following cases we trade latency
2860 case BFQQE_TOO_IDLE:
2862 * This is the only case where we may reduce
2863 * the budget: if there is no request of the
2864 * process still waiting for completion, then
2865 * we assume (tentatively) that the timer has
2866 * expired because the batch of requests of
2867 * the process could have been served with a
2868 * smaller budget. Hence, betting that
2869 * process will behave in the same way when it
2870 * becomes backlogged again, we reduce its
2871 * next budget. As long as we guess right,
2872 * this budget cut reduces the latency
2873 * experienced by the process.
2875 * However, if there are still outstanding
2876 * requests, then the process may have not yet
2877 * issued its next request just because it is
2878 * still waiting for the completion of some of
2879 * the still outstanding ones. So in this
2880 * subcase we do not reduce its budget, on the
2881 * contrary we increase it to possibly boost
2882 * the throughput, as discussed in the
2883 * comments to the BUDGET_TIMEOUT case.
2885 if (bfqq->dispatched > 0) /* still outstanding reqs */
2886 budget = min(budget * 2, bfqd->bfq_max_budget);
2888 if (budget > 5 * min_budget)
2889 budget -= 4 * min_budget;
2891 budget = min_budget;
2894 case BFQQE_BUDGET_TIMEOUT:
2896 * We double the budget here because it gives
2897 * the chance to boost the throughput if this
2898 * is not a seeky process (and has bumped into
2899 * this timeout because of, e.g., ZBR).
2901 budget = min(budget * 2, bfqd->bfq_max_budget);
2903 case BFQQE_BUDGET_EXHAUSTED:
2905 * The process still has backlog, and did not
2906 * let either the budget timeout or the disk
2907 * idling timeout expire. Hence it is not
2908 * seeky, has a short thinktime and may be
2909 * happy with a higher budget too. So
2910 * definitely increase the budget of this good
2911 * candidate to boost the disk throughput.
2913 budget = min(budget * 4, bfqd->bfq_max_budget);
2915 case BFQQE_NO_MORE_REQUESTS:
2917 * For queues that expire for this reason, it
2918 * is particularly important to keep the
2919 * budget close to the actual service they
2920 * need. Doing so reduces the timestamp
2921 * misalignment problem described in the
2922 * comments in the body of
2923 * __bfq_activate_entity. In fact, suppose
2924 * that a queue systematically expires for
2925 * BFQQE_NO_MORE_REQUESTS and presents a
2926 * new request in time to enjoy timestamp
2927 * back-shifting. The larger the budget of the
2928 * queue is with respect to the service the
2929 * queue actually requests in each service
2930 * slot, the more times the queue can be
2931 * reactivated with the same virtual finish
2932 * time. It follows that, even if this finish
2933 * time is pushed to the system virtual time
2934 * to reduce the consequent timestamp
2935 * misalignment, the queue unjustly enjoys for
2936 * many re-activations a lower finish time
2937 * than all newly activated queues.
2939 * The service needed by bfqq is measured
2940 * quite precisely by bfqq->entity.service.
2941 * Since bfqq does not enjoy device idling,
2942 * bfqq->entity.service is equal to the number
2943 * of sectors that the process associated with
2944 * bfqq requested to read/write before waiting
2945 * for request completions, or blocking for
2948 budget = max_t(int, bfqq->entity.service, min_budget);
2953 } else if (!bfq_bfqq_sync(bfqq)) {
2955 * Async queues get always the maximum possible
2956 * budget, as for them we do not care about latency
2957 * (in addition, their ability to dispatch is limited
2958 * by the charging factor).
2960 budget = bfqd->bfq_max_budget;
2963 bfqq->max_budget = budget;
2965 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
2966 !bfqd->bfq_user_max_budget)
2967 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
2970 * If there is still backlog, then assign a new budget, making
2971 * sure that it is large enough for the next request. Since
2972 * the finish time of bfqq must be kept in sync with the
2973 * budget, be sure to call __bfq_bfqq_expire() *after* this
2976 * If there is no backlog, then no need to update the budget;
2977 * it will be updated on the arrival of a new request.
2979 next_rq = bfqq->next_rq;
2981 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
2982 bfq_serv_to_charge(next_rq, bfqq));
2984 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
2985 next_rq ? blk_rq_sectors(next_rq) : 0,
2986 bfqq->entity.budget);
2990 * Return true if the process associated with bfqq is "slow". The slow
2991 * flag is used, in addition to the budget timeout, to reduce the
2992 * amount of service provided to seeky processes, and thus reduce
2993 * their chances to lower the throughput. More details in the comments
2994 * on the function bfq_bfqq_expire().
2996 * An important observation is in order: as discussed in the comments
2997 * on the function bfq_update_peak_rate(), with devices with internal
2998 * queues, it is hard if ever possible to know when and for how long
2999 * an I/O request is processed by the device (apart from the trivial
3000 * I/O pattern where a new request is dispatched only after the
3001 * previous one has been completed). This makes it hard to evaluate
3002 * the real rate at which the I/O requests of each bfq_queue are
3003 * served. In fact, for an I/O scheduler like BFQ, serving a
3004 * bfq_queue means just dispatching its requests during its service
3005 * slot (i.e., until the budget of the queue is exhausted, or the
3006 * queue remains idle, or, finally, a timeout fires). But, during the
3007 * service slot of a bfq_queue, around 100 ms at most, the device may
3008 * be even still processing requests of bfq_queues served in previous
3009 * service slots. On the opposite end, the requests of the in-service
3010 * bfq_queue may be completed after the service slot of the queue
3013 * Anyway, unless more sophisticated solutions are used
3014 * (where possible), the sum of the sizes of the requests dispatched
3015 * during the service slot of a bfq_queue is probably the only
3016 * approximation available for the service received by the bfq_queue
3017 * during its service slot. And this sum is the quantity used in this
3018 * function to evaluate the I/O speed of a process.
3020 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3021 bool compensate, enum bfqq_expiration reason,
3022 unsigned long *delta_ms)
3024 ktime_t delta_ktime;
3026 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
3028 if (!bfq_bfqq_sync(bfqq))
3032 delta_ktime = bfqd->last_idling_start;
3034 delta_ktime = ktime_get();
3035 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
3036 delta_usecs = ktime_to_us(delta_ktime);
3038 /* don't use too short time intervals */
3039 if (delta_usecs < 1000) {
3040 if (blk_queue_nonrot(bfqd->queue))
3042 * give same worst-case guarantees as idling
3045 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
3046 else /* charge at least one seek */
3047 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
3052 *delta_ms = delta_usecs / USEC_PER_MSEC;
3055 * Use only long (> 20ms) intervals to filter out excessive
3056 * spikes in service rate estimation.
3058 if (delta_usecs > 20000) {
3060 * Caveat for rotational devices: processes doing I/O
3061 * in the slower disk zones tend to be slow(er) even
3062 * if not seeky. In this respect, the estimated peak
3063 * rate is likely to be an average over the disk
3064 * surface. Accordingly, to not be too harsh with
3065 * unlucky processes, a process is deemed slow only if
3066 * its rate has been lower than half of the estimated
3069 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
3072 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
3078 * To be deemed as soft real-time, an application must meet two
3079 * requirements. First, the application must not require an average
3080 * bandwidth higher than the approximate bandwidth required to playback or
3081 * record a compressed high-definition video.
3082 * The next function is invoked on the completion of the last request of a
3083 * batch, to compute the next-start time instant, soft_rt_next_start, such
3084 * that, if the next request of the application does not arrive before
3085 * soft_rt_next_start, then the above requirement on the bandwidth is met.
3087 * The second requirement is that the request pattern of the application is
3088 * isochronous, i.e., that, after issuing a request or a batch of requests,
3089 * the application stops issuing new requests until all its pending requests
3090 * have been completed. After that, the application may issue a new batch,
3092 * For this reason the next function is invoked to compute
3093 * soft_rt_next_start only for applications that meet this requirement,
3094 * whereas soft_rt_next_start is set to infinity for applications that do
3097 * Unfortunately, even a greedy (i.e., I/O-bound) application may
3098 * happen to meet, occasionally or systematically, both the above
3099 * bandwidth and isochrony requirements. This may happen at least in
3100 * the following circumstances. First, if the CPU load is high. The
3101 * application may stop issuing requests while the CPUs are busy
3102 * serving other processes, then restart, then stop again for a while,
3103 * and so on. The other circumstances are related to the storage
3104 * device: the storage device is highly loaded or reaches a low-enough
3105 * throughput with the I/O of the application (e.g., because the I/O
3106 * is random and/or the device is slow). In all these cases, the
3107 * I/O of the application may be simply slowed down enough to meet
3108 * the bandwidth and isochrony requirements. To reduce the probability
3109 * that greedy applications are deemed as soft real-time in these
3110 * corner cases, a further rule is used in the computation of
3111 * soft_rt_next_start: the return value of this function is forced to
3112 * be higher than the maximum between the following two quantities.
3114 * (a) Current time plus: (1) the maximum time for which the arrival
3115 * of a request is waited for when a sync queue becomes idle,
3116 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
3117 * postpone for a moment the reason for adding a few extra
3118 * jiffies; we get back to it after next item (b). Lower-bounding
3119 * the return value of this function with the current time plus
3120 * bfqd->bfq_slice_idle tends to filter out greedy applications,
3121 * because the latter issue their next request as soon as possible
3122 * after the last one has been completed. In contrast, a soft
3123 * real-time application spends some time processing data, after a
3124 * batch of its requests has been completed.
3126 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
3127 * above, greedy applications may happen to meet both the
3128 * bandwidth and isochrony requirements under heavy CPU or
3129 * storage-device load. In more detail, in these scenarios, these
3130 * applications happen, only for limited time periods, to do I/O
3131 * slowly enough to meet all the requirements described so far,
3132 * including the filtering in above item (a). These slow-speed
3133 * time intervals are usually interspersed between other time
3134 * intervals during which these applications do I/O at a very high
3135 * speed. Fortunately, exactly because of the high speed of the
3136 * I/O in the high-speed intervals, the values returned by this
3137 * function happen to be so high, near the end of any such
3138 * high-speed interval, to be likely to fall *after* the end of
3139 * the low-speed time interval that follows. These high values are
3140 * stored in bfqq->soft_rt_next_start after each invocation of
3141 * this function. As a consequence, if the last value of
3142 * bfqq->soft_rt_next_start is constantly used to lower-bound the
3143 * next value that this function may return, then, from the very
3144 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
3145 * likely to be constantly kept so high that any I/O request
3146 * issued during the low-speed interval is considered as arriving
3147 * to soon for the application to be deemed as soft
3148 * real-time. Then, in the high-speed interval that follows, the
3149 * application will not be deemed as soft real-time, just because
3150 * it will do I/O at a high speed. And so on.
3152 * Getting back to the filtering in item (a), in the following two
3153 * cases this filtering might be easily passed by a greedy
3154 * application, if the reference quantity was just
3155 * bfqd->bfq_slice_idle:
3156 * 1) HZ is so low that the duration of a jiffy is comparable to or
3157 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
3158 * devices with HZ=100. The time granularity may be so coarse
3159 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
3160 * is rather lower than the exact value.
3161 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
3162 * for a while, then suddenly 'jump' by several units to recover the lost
3163 * increments. This seems to happen, e.g., inside virtual machines.
3164 * To address this issue, in the filtering in (a) we do not use as a
3165 * reference time interval just bfqd->bfq_slice_idle, but
3166 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
3167 * minimum number of jiffies for which the filter seems to be quite
3168 * precise also in embedded systems and KVM/QEMU virtual machines.
3170 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
3171 struct bfq_queue *bfqq)
3173 return max3(bfqq->soft_rt_next_start,
3174 bfqq->last_idle_bklogged +
3175 HZ * bfqq->service_from_backlogged /
3176 bfqd->bfq_wr_max_softrt_rate,
3177 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
3181 * bfq_bfqq_expire - expire a queue.
3182 * @bfqd: device owning the queue.
3183 * @bfqq: the queue to expire.
3184 * @compensate: if true, compensate for the time spent idling.
3185 * @reason: the reason causing the expiration.
3187 * If the process associated with bfqq does slow I/O (e.g., because it
3188 * issues random requests), we charge bfqq with the time it has been
3189 * in service instead of the service it has received (see
3190 * bfq_bfqq_charge_time for details on how this goal is achieved). As
3191 * a consequence, bfqq will typically get higher timestamps upon
3192 * reactivation, and hence it will be rescheduled as if it had
3193 * received more service than what it has actually received. In the
3194 * end, bfqq receives less service in proportion to how slowly its
3195 * associated process consumes its budgets (and hence how seriously it
3196 * tends to lower the throughput). In addition, this time-charging
3197 * strategy guarantees time fairness among slow processes. In
3198 * contrast, if the process associated with bfqq is not slow, we
3199 * charge bfqq exactly with the service it has received.
3201 * Charging time to the first type of queues and the exact service to
3202 * the other has the effect of using the WF2Q+ policy to schedule the
3203 * former on a timeslice basis, without violating service domain
3204 * guarantees among the latter.
3206 void bfq_bfqq_expire(struct bfq_data *bfqd,
3207 struct bfq_queue *bfqq,
3209 enum bfqq_expiration reason)
3212 unsigned long delta = 0;
3213 struct bfq_entity *entity = &bfqq->entity;
3217 * Check whether the process is slow (see bfq_bfqq_is_slow).
3219 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
3222 * As above explained, charge slow (typically seeky) and
3223 * timed-out queues with the time and not the service
3224 * received, to favor sequential workloads.
3226 * Processes doing I/O in the slower disk zones will tend to
3227 * be slow(er) even if not seeky. Therefore, since the
3228 * estimated peak rate is actually an average over the disk
3229 * surface, these processes may timeout just for bad luck. To
3230 * avoid punishing them, do not charge time to processes that
3231 * succeeded in consuming at least 2/3 of their budget. This
3232 * allows BFQ to preserve enough elasticity to still perform
3233 * bandwidth, and not time, distribution with little unlucky
3234 * or quasi-sequential processes.
3236 if (bfqq->wr_coeff == 1 &&
3238 (reason == BFQQE_BUDGET_TIMEOUT &&
3239 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
3240 bfq_bfqq_charge_time(bfqd, bfqq, delta);
3242 if (reason == BFQQE_TOO_IDLE &&
3243 entity->service <= 2 * entity->budget / 10)
3244 bfq_clear_bfqq_IO_bound(bfqq);
3246 if (bfqd->low_latency && bfqq->wr_coeff == 1)
3247 bfqq->last_wr_start_finish = jiffies;
3249 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
3250 RB_EMPTY_ROOT(&bfqq->sort_list)) {
3252 * If we get here, and there are no outstanding
3253 * requests, then the request pattern is isochronous
3254 * (see the comments on the function
3255 * bfq_bfqq_softrt_next_start()). Thus we can compute
3256 * soft_rt_next_start. If, instead, the queue still
3257 * has outstanding requests, then we have to wait for
3258 * the completion of all the outstanding requests to
3259 * discover whether the request pattern is actually
3262 if (bfqq->dispatched == 0)
3263 bfqq->soft_rt_next_start =
3264 bfq_bfqq_softrt_next_start(bfqd, bfqq);
3267 * Schedule an update of soft_rt_next_start to when
3268 * the task may be discovered to be isochronous.
3270 bfq_mark_bfqq_softrt_update(bfqq);
3274 bfq_log_bfqq(bfqd, bfqq,
3275 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
3276 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
3279 * Increase, decrease or leave budget unchanged according to
3282 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
3284 __bfq_bfqq_expire(bfqd, bfqq);
3286 if (ref == 1) /* bfqq is gone, no more actions on it */
3289 /* mark bfqq as waiting a request only if a bic still points to it */
3290 if (!bfq_bfqq_busy(bfqq) &&
3291 reason != BFQQE_BUDGET_TIMEOUT &&
3292 reason != BFQQE_BUDGET_EXHAUSTED) {
3293 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
3295 * Not setting service to 0, because, if the next rq
3296 * arrives in time, the queue will go on receiving
3297 * service with this same budget (as if it never expired)
3300 entity->service = 0;
3304 * Budget timeout is not implemented through a dedicated timer, but
3305 * just checked on request arrivals and completions, as well as on
3306 * idle timer expirations.
3308 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
3310 return time_is_before_eq_jiffies(bfqq->budget_timeout);
3314 * If we expire a queue that is actively waiting (i.e., with the
3315 * device idled) for the arrival of a new request, then we may incur
3316 * the timestamp misalignment problem described in the body of the
3317 * function __bfq_activate_entity. Hence we return true only if this
3318 * condition does not hold, or if the queue is slow enough to deserve
3319 * only to be kicked off for preserving a high throughput.
3321 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
3323 bfq_log_bfqq(bfqq->bfqd, bfqq,
3324 "may_budget_timeout: wait_request %d left %d timeout %d",
3325 bfq_bfqq_wait_request(bfqq),
3326 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
3327 bfq_bfqq_budget_timeout(bfqq));
3329 return (!bfq_bfqq_wait_request(bfqq) ||
3330 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
3332 bfq_bfqq_budget_timeout(bfqq);
3336 * For a queue that becomes empty, device idling is allowed only if
3337 * this function returns true for the queue. As a consequence, since
3338 * device idling plays a critical role in both throughput boosting and
3339 * service guarantees, the return value of this function plays a
3340 * critical role in both these aspects as well.
3342 * In a nutshell, this function returns true only if idling is
3343 * beneficial for throughput or, even if detrimental for throughput,
3344 * idling is however necessary to preserve service guarantees (low
3345 * latency, desired throughput distribution, ...). In particular, on
3346 * NCQ-capable devices, this function tries to return false, so as to
3347 * help keep the drives' internal queues full, whenever this helps the
3348 * device boost the throughput without causing any service-guarantee
3351 * In more detail, the return value of this function is obtained by,
3352 * first, computing a number of boolean variables that take into
3353 * account throughput and service-guarantee issues, and, then,
3354 * combining these variables in a logical expression. Most of the
3355 * issues taken into account are not trivial. We discuss these issues
3356 * individually while introducing the variables.
3358 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
3360 struct bfq_data *bfqd = bfqq->bfqd;
3361 bool rot_without_queueing =
3362 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
3363 bfqq_sequential_and_IO_bound,
3364 idling_boosts_thr, idling_boosts_thr_without_issues,
3365 idling_needed_for_service_guarantees,
3366 asymmetric_scenario;
3368 if (bfqd->strict_guarantees)
3372 * Idling is performed only if slice_idle > 0. In addition, we
3375 * (b) bfqq is in the idle io prio class: in this case we do
3376 * not idle because we want to minimize the bandwidth that
3377 * queues in this class can steal to higher-priority queues
3379 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
3380 bfq_class_idle(bfqq))
3383 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
3384 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
3387 * The next variable takes into account the cases where idling
3388 * boosts the throughput.
3390 * The value of the variable is computed considering, first, that
3391 * idling is virtually always beneficial for the throughput if:
3392 * (a) the device is not NCQ-capable and rotational, or
3393 * (b) regardless of the presence of NCQ, the device is rotational and
3394 * the request pattern for bfqq is I/O-bound and sequential, or
3395 * (c) regardless of whether it is rotational, the device is
3396 * not NCQ-capable and the request pattern for bfqq is
3397 * I/O-bound and sequential.
3399 * Secondly, and in contrast to the above item (b), idling an
3400 * NCQ-capable flash-based device would not boost the
3401 * throughput even with sequential I/O; rather it would lower
3402 * the throughput in proportion to how fast the device
3403 * is. Accordingly, the next variable is true if any of the
3404 * above conditions (a), (b) or (c) is true, and, in
3405 * particular, happens to be false if bfqd is an NCQ-capable
3406 * flash-based device.
3408 idling_boosts_thr = rot_without_queueing ||
3409 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
3410 bfqq_sequential_and_IO_bound);
3413 * The value of the next variable,
3414 * idling_boosts_thr_without_issues, is equal to that of
3415 * idling_boosts_thr, unless a special case holds. In this
3416 * special case, described below, idling may cause problems to
3417 * weight-raised queues.
3419 * When the request pool is saturated (e.g., in the presence
3420 * of write hogs), if the processes associated with
3421 * non-weight-raised queues ask for requests at a lower rate,
3422 * then processes associated with weight-raised queues have a
3423 * higher probability to get a request from the pool
3424 * immediately (or at least soon) when they need one. Thus
3425 * they have a higher probability to actually get a fraction
3426 * of the device throughput proportional to their high
3427 * weight. This is especially true with NCQ-capable drives,
3428 * which enqueue several requests in advance, and further
3429 * reorder internally-queued requests.
3431 * For this reason, we force to false the value of
3432 * idling_boosts_thr_without_issues if there are weight-raised
3433 * busy queues. In this case, and if bfqq is not weight-raised,
3434 * this guarantees that the device is not idled for bfqq (if,
3435 * instead, bfqq is weight-raised, then idling will be
3436 * guaranteed by another variable, see below). Combined with
3437 * the timestamping rules of BFQ (see [1] for details), this
3438 * behavior causes bfqq, and hence any sync non-weight-raised
3439 * queue, to get a lower number of requests served, and thus
3440 * to ask for a lower number of requests from the request
3441 * pool, before the busy weight-raised queues get served
3442 * again. This often mitigates starvation problems in the
3443 * presence of heavy write workloads and NCQ, thereby
3444 * guaranteeing a higher application and system responsiveness
3445 * in these hostile scenarios.
3447 idling_boosts_thr_without_issues = idling_boosts_thr &&
3448 bfqd->wr_busy_queues == 0;
3451 * There is then a case where idling must be performed not
3452 * for throughput concerns, but to preserve service
3455 * To introduce this case, we can note that allowing the drive
3456 * to enqueue more than one request at a time, and hence
3457 * delegating de facto final scheduling decisions to the
3458 * drive's internal scheduler, entails loss of control on the
3459 * actual request service order. In particular, the critical
3460 * situation is when requests from different processes happen
3461 * to be present, at the same time, in the internal queue(s)
3462 * of the drive. In such a situation, the drive, by deciding
3463 * the service order of the internally-queued requests, does
3464 * determine also the actual throughput distribution among
3465 * these processes. But the drive typically has no notion or
3466 * concern about per-process throughput distribution, and
3467 * makes its decisions only on a per-request basis. Therefore,
3468 * the service distribution enforced by the drive's internal
3469 * scheduler is likely to coincide with the desired
3470 * device-throughput distribution only in a completely
3471 * symmetric scenario where:
3472 * (i) each of these processes must get the same throughput as
3474 * (ii) all these processes have the same I/O pattern
3475 (either sequential or random).
3476 * In fact, in such a scenario, the drive will tend to treat
3477 * the requests of each of these processes in about the same
3478 * way as the requests of the others, and thus to provide
3479 * each of these processes with about the same throughput
3480 * (which is exactly the desired throughput distribution). In
3481 * contrast, in any asymmetric scenario, device idling is
3482 * certainly needed to guarantee that bfqq receives its
3483 * assigned fraction of the device throughput (see [1] for
3486 * We address this issue by controlling, actually, only the
3487 * symmetry sub-condition (i), i.e., provided that
3488 * sub-condition (i) holds, idling is not performed,
3489 * regardless of whether sub-condition (ii) holds. In other
3490 * words, only if sub-condition (i) holds, then idling is
3491 * allowed, and the device tends to be prevented from queueing
3492 * many requests, possibly of several processes. The reason
3493 * for not controlling also sub-condition (ii) is that we
3494 * exploit preemption to preserve guarantees in case of
3495 * symmetric scenarios, even if (ii) does not hold, as
3496 * explained in the next two paragraphs.
3498 * Even if a queue, say Q, is expired when it remains idle, Q
3499 * can still preempt the new in-service queue if the next
3500 * request of Q arrives soon (see the comments on
3501 * bfq_bfqq_update_budg_for_activation). If all queues and
3502 * groups have the same weight, this form of preemption,
3503 * combined with the hole-recovery heuristic described in the
3504 * comments on function bfq_bfqq_update_budg_for_activation,
3505 * are enough to preserve a correct bandwidth distribution in
3506 * the mid term, even without idling. In fact, even if not
3507 * idling allows the internal queues of the device to contain
3508 * many requests, and thus to reorder requests, we can rather
3509 * safely assume that the internal scheduler still preserves a
3510 * minimum of mid-term fairness. The motivation for using
3511 * preemption instead of idling is that, by not idling,
3512 * service guarantees are preserved without minimally
3513 * sacrificing throughput. In other words, both a high
3514 * throughput and its desired distribution are obtained.
3516 * More precisely, this preemption-based, idleless approach
3517 * provides fairness in terms of IOPS, and not sectors per
3518 * second. This can be seen with a simple example. Suppose
3519 * that there are two queues with the same weight, but that
3520 * the first queue receives requests of 8 sectors, while the
3521 * second queue receives requests of 1024 sectors. In
3522 * addition, suppose that each of the two queues contains at
3523 * most one request at a time, which implies that each queue
3524 * always remains idle after it is served. Finally, after
3525 * remaining idle, each queue receives very quickly a new
3526 * request. It follows that the two queues are served
3527 * alternatively, preempting each other if needed. This
3528 * implies that, although both queues have the same weight,
3529 * the queue with large requests receives a service that is
3530 * 1024/8 times as high as the service received by the other
3533 * On the other hand, device idling is performed, and thus
3534 * pure sector-domain guarantees are provided, for the
3535 * following queues, which are likely to need stronger
3536 * throughput guarantees: weight-raised queues, and queues
3537 * with a higher weight than other queues. When such queues
3538 * are active, sub-condition (i) is false, which triggers
3541 * According to the above considerations, the next variable is
3542 * true (only) if sub-condition (i) holds. To compute the
3543 * value of this variable, we not only use the return value of
3544 * the function bfq_symmetric_scenario(), but also check
3545 * whether bfqq is being weight-raised, because
3546 * bfq_symmetric_scenario() does not take into account also
3547 * weight-raised queues (see comments on
3548 * bfq_weights_tree_add()).
3550 * As a side note, it is worth considering that the above
3551 * device-idling countermeasures may however fail in the
3552 * following unlucky scenario: if idling is (correctly)
3553 * disabled in a time period during which all symmetry
3554 * sub-conditions hold, and hence the device is allowed to
3555 * enqueue many requests, but at some later point in time some
3556 * sub-condition stops to hold, then it may become impossible
3557 * to let requests be served in the desired order until all
3558 * the requests already queued in the device have been served.
3560 asymmetric_scenario = bfqq->wr_coeff > 1 ||
3561 !bfq_symmetric_scenario(bfqd);
3564 * Finally, there is a case where maximizing throughput is the
3565 * best choice even if it may cause unfairness toward
3566 * bfqq. Such a case is when bfqq became active in a burst of
3567 * queue activations. Queues that became active during a large
3568 * burst benefit only from throughput, as discussed in the
3569 * comments on bfq_handle_burst. Thus, if bfqq became active
3570 * in a burst and not idling the device maximizes throughput,
3571 * then the device must no be idled, because not idling the
3572 * device provides bfqq and all other queues in the burst with
3573 * maximum benefit. Combining this and the above case, we can
3574 * now establish when idling is actually needed to preserve
3575 * service guarantees.
3577 idling_needed_for_service_guarantees =
3578 asymmetric_scenario && !bfq_bfqq_in_large_burst(bfqq);
3581 * We have now all the components we need to compute the
3582 * return value of the function, which is true only if idling
3583 * either boosts the throughput (without issues), or is
3584 * necessary to preserve service guarantees.
3586 return idling_boosts_thr_without_issues ||
3587 idling_needed_for_service_guarantees;
3591 * If the in-service queue is empty but the function bfq_better_to_idle
3592 * returns true, then:
3593 * 1) the queue must remain in service and cannot be expired, and
3594 * 2) the device must be idled to wait for the possible arrival of a new
3595 * request for the queue.
3596 * See the comments on the function bfq_better_to_idle for the reasons
3597 * why performing device idling is the best choice to boost the throughput
3598 * and preserve service guarantees when bfq_better_to_idle itself
3601 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
3603 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
3607 * Select a queue for service. If we have a current queue in service,
3608 * check whether to continue servicing it, or retrieve and set a new one.
3610 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
3612 struct bfq_queue *bfqq;
3613 struct request *next_rq;
3614 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
3616 bfqq = bfqd->in_service_queue;
3620 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
3623 * Do not expire bfqq for budget timeout if bfqq may be about
3624 * to enjoy device idling. The reason why, in this case, we
3625 * prevent bfqq from expiring is the same as in the comments
3626 * on the case where bfq_bfqq_must_idle() returns true, in
3627 * bfq_completed_request().
3629 if (bfq_may_expire_for_budg_timeout(bfqq) &&
3630 !bfq_bfqq_must_idle(bfqq))
3635 * This loop is rarely executed more than once. Even when it
3636 * happens, it is much more convenient to re-execute this loop
3637 * than to return NULL and trigger a new dispatch to get a
3640 next_rq = bfqq->next_rq;
3642 * If bfqq has requests queued and it has enough budget left to
3643 * serve them, keep the queue, otherwise expire it.
3646 if (bfq_serv_to_charge(next_rq, bfqq) >
3647 bfq_bfqq_budget_left(bfqq)) {
3649 * Expire the queue for budget exhaustion,
3650 * which makes sure that the next budget is
3651 * enough to serve the next request, even if
3652 * it comes from the fifo expired path.
3654 reason = BFQQE_BUDGET_EXHAUSTED;
3658 * The idle timer may be pending because we may
3659 * not disable disk idling even when a new request
3662 if (bfq_bfqq_wait_request(bfqq)) {
3664 * If we get here: 1) at least a new request
3665 * has arrived but we have not disabled the
3666 * timer because the request was too small,
3667 * 2) then the block layer has unplugged
3668 * the device, causing the dispatch to be
3671 * Since the device is unplugged, now the
3672 * requests are probably large enough to
3673 * provide a reasonable throughput.
3674 * So we disable idling.
3676 bfq_clear_bfqq_wait_request(bfqq);
3677 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
3684 * No requests pending. However, if the in-service queue is idling
3685 * for a new request, or has requests waiting for a completion and
3686 * may idle after their completion, then keep it anyway.
3688 if (bfq_bfqq_wait_request(bfqq) ||
3689 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
3694 reason = BFQQE_NO_MORE_REQUESTS;
3696 bfq_bfqq_expire(bfqd, bfqq, false, reason);
3698 bfqq = bfq_set_in_service_queue(bfqd);
3700 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
3705 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
3707 bfq_log(bfqd, "select_queue: no queue returned");
3712 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3714 struct bfq_entity *entity = &bfqq->entity;
3716 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
3717 bfq_log_bfqq(bfqd, bfqq,
3718 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
3719 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
3720 jiffies_to_msecs(bfqq->wr_cur_max_time),
3722 bfqq->entity.weight, bfqq->entity.orig_weight);
3724 if (entity->prio_changed)
3725 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
3728 * If the queue was activated in a burst, or too much
3729 * time has elapsed from the beginning of this
3730 * weight-raising period, then end weight raising.
3732 if (bfq_bfqq_in_large_burst(bfqq))
3733 bfq_bfqq_end_wr(bfqq);
3734 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
3735 bfqq->wr_cur_max_time)) {
3736 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
3737 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
3738 bfq_wr_duration(bfqd)))
3739 bfq_bfqq_end_wr(bfqq);
3741 switch_back_to_interactive_wr(bfqq, bfqd);
3742 bfqq->entity.prio_changed = 1;
3745 if (bfqq->wr_coeff > 1 &&
3746 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
3747 bfqq->service_from_wr > max_service_from_wr) {
3748 /* see comments on max_service_from_wr */
3749 bfq_bfqq_end_wr(bfqq);
3753 * To improve latency (for this or other queues), immediately
3754 * update weight both if it must be raised and if it must be
3755 * lowered. Since, entity may be on some active tree here, and
3756 * might have a pending change of its ioprio class, invoke
3757 * next function with the last parameter unset (see the
3758 * comments on the function).
3760 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
3761 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
3766 * Dispatch next request from bfqq.
3768 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
3769 struct bfq_queue *bfqq)
3771 struct request *rq = bfqq->next_rq;
3772 unsigned long service_to_charge;
3774 service_to_charge = bfq_serv_to_charge(rq, bfqq);
3776 bfq_bfqq_served(bfqq, service_to_charge);
3778 bfq_dispatch_remove(bfqd->queue, rq);
3781 * If weight raising has to terminate for bfqq, then next
3782 * function causes an immediate update of bfqq's weight,
3783 * without waiting for next activation. As a consequence, on
3784 * expiration, bfqq will be timestamped as if has never been
3785 * weight-raised during this service slot, even if it has
3786 * received part or even most of the service as a
3787 * weight-raised queue. This inflates bfqq's timestamps, which
3788 * is beneficial, as bfqq is then more willing to leave the
3789 * device immediately to possible other weight-raised queues.
3791 bfq_update_wr_data(bfqd, bfqq);
3794 * Expire bfqq, pretending that its budget expired, if bfqq
3795 * belongs to CLASS_IDLE and other queues are waiting for
3798 if (bfqd->busy_queues > 1 && bfq_class_idle(bfqq))
3804 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
3808 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
3810 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3813 * Avoiding lock: a race on bfqd->busy_queues should cause at
3814 * most a call to dispatch for nothing
3816 return !list_empty_careful(&bfqd->dispatch) ||
3817 bfqd->busy_queues > 0;
3820 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3822 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3823 struct request *rq = NULL;
3824 struct bfq_queue *bfqq = NULL;
3826 if (!list_empty(&bfqd->dispatch)) {
3827 rq = list_first_entry(&bfqd->dispatch, struct request,
3829 list_del_init(&rq->queuelist);
3835 * Increment counters here, because this
3836 * dispatch does not follow the standard
3837 * dispatch flow (where counters are
3842 goto inc_in_driver_start_rq;
3846 * We exploit the bfq_finish_requeue_request hook to
3847 * decrement rq_in_driver, but
3848 * bfq_finish_requeue_request will not be invoked on
3849 * this request. So, to avoid unbalance, just start
3850 * this request, without incrementing rq_in_driver. As
3851 * a negative consequence, rq_in_driver is deceptively
3852 * lower than it should be while this request is in
3853 * service. This may cause bfq_schedule_dispatch to be
3854 * invoked uselessly.
3856 * As for implementing an exact solution, the
3857 * bfq_finish_requeue_request hook, if defined, is
3858 * probably invoked also on this request. So, by
3859 * exploiting this hook, we could 1) increment
3860 * rq_in_driver here, and 2) decrement it in
3861 * bfq_finish_requeue_request. Such a solution would
3862 * let the value of the counter be always accurate,
3863 * but it would entail using an extra interface
3864 * function. This cost seems higher than the benefit,
3865 * being the frequency of non-elevator-private
3866 * requests very low.
3871 bfq_log(bfqd, "dispatch requests: %d busy queues", bfqd->busy_queues);
3873 if (bfqd->busy_queues == 0)
3877 * Force device to serve one request at a time if
3878 * strict_guarantees is true. Forcing this service scheme is
3879 * currently the ONLY way to guarantee that the request
3880 * service order enforced by the scheduler is respected by a
3881 * queueing device. Otherwise the device is free even to make
3882 * some unlucky request wait for as long as the device
3885 * Of course, serving one request at at time may cause loss of
3888 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
3891 bfqq = bfq_select_queue(bfqd);
3895 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
3898 inc_in_driver_start_rq:
3899 bfqd->rq_in_driver++;
3901 rq->rq_flags |= RQF_STARTED;
3907 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
3908 static void bfq_update_dispatch_stats(struct request_queue *q,
3910 struct bfq_queue *in_serv_queue,
3911 bool idle_timer_disabled)
3913 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
3915 if (!idle_timer_disabled && !bfqq)
3919 * rq and bfqq are guaranteed to exist until this function
3920 * ends, for the following reasons. First, rq can be
3921 * dispatched to the device, and then can be completed and
3922 * freed, only after this function ends. Second, rq cannot be
3923 * merged (and thus freed because of a merge) any longer,
3924 * because it has already started. Thus rq cannot be freed
3925 * before this function ends, and, since rq has a reference to
3926 * bfqq, the same guarantee holds for bfqq too.
3928 * In addition, the following queue lock guarantees that
3929 * bfqq_group(bfqq) exists as well.
3931 spin_lock_irq(q->queue_lock);
3932 if (idle_timer_disabled)
3934 * Since the idle timer has been disabled,
3935 * in_serv_queue contained some request when
3936 * __bfq_dispatch_request was invoked above, which
3937 * implies that rq was picked exactly from
3938 * in_serv_queue. Thus in_serv_queue == bfqq, and is
3939 * therefore guaranteed to exist because of the above
3942 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
3944 struct bfq_group *bfqg = bfqq_group(bfqq);
3946 bfqg_stats_update_avg_queue_size(bfqg);
3947 bfqg_stats_set_start_empty_time(bfqg);
3948 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
3950 spin_unlock_irq(q->queue_lock);
3953 static inline void bfq_update_dispatch_stats(struct request_queue *q,
3955 struct bfq_queue *in_serv_queue,
3956 bool idle_timer_disabled) {}
3959 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3961 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3963 struct bfq_queue *in_serv_queue;
3964 bool waiting_rq, idle_timer_disabled;
3966 spin_lock_irq(&bfqd->lock);
3968 in_serv_queue = bfqd->in_service_queue;
3969 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
3971 rq = __bfq_dispatch_request(hctx);
3973 idle_timer_disabled =
3974 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
3976 spin_unlock_irq(&bfqd->lock);
3978 bfq_update_dispatch_stats(hctx->queue, rq, in_serv_queue,
3979 idle_timer_disabled);
3985 * Task holds one reference to the queue, dropped when task exits. Each rq
3986 * in-flight on this queue also holds a reference, dropped when rq is freed.
3988 * Scheduler lock must be held here. Recall not to use bfqq after calling
3989 * this function on it.
3991 void bfq_put_queue(struct bfq_queue *bfqq)
3993 #ifdef CONFIG_BFQ_GROUP_IOSCHED
3994 struct bfq_group *bfqg = bfqq_group(bfqq);
3998 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
4005 if (!hlist_unhashed(&bfqq->burst_list_node)) {
4006 hlist_del_init(&bfqq->burst_list_node);
4008 * Decrement also burst size after the removal, if the
4009 * process associated with bfqq is exiting, and thus
4010 * does not contribute to the burst any longer. This
4011 * decrement helps filter out false positives of large
4012 * bursts, when some short-lived process (often due to
4013 * the execution of commands by some service) happens
4014 * to start and exit while a complex application is
4015 * starting, and thus spawning several processes that
4016 * do I/O (and that *must not* be treated as a large
4017 * burst, see comments on bfq_handle_burst).
4019 * In particular, the decrement is performed only if:
4020 * 1) bfqq is not a merged queue, because, if it is,
4021 * then this free of bfqq is not triggered by the exit
4022 * of the process bfqq is associated with, but exactly
4023 * by the fact that bfqq has just been merged.
4024 * 2) burst_size is greater than 0, to handle
4025 * unbalanced decrements. Unbalanced decrements may
4026 * happen in te following case: bfqq is inserted into
4027 * the current burst list--without incrementing
4028 * bust_size--because of a split, but the current
4029 * burst list is not the burst list bfqq belonged to
4030 * (see comments on the case of a split in
4033 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
4034 bfqq->bfqd->burst_size--;
4037 kmem_cache_free(bfq_pool, bfqq);
4038 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4039 bfqg_and_blkg_put(bfqg);
4043 static void bfq_put_cooperator(struct bfq_queue *bfqq)
4045 struct bfq_queue *__bfqq, *next;
4048 * If this queue was scheduled to merge with another queue, be
4049 * sure to drop the reference taken on that queue (and others in
4050 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
4052 __bfqq = bfqq->new_bfqq;
4056 next = __bfqq->new_bfqq;
4057 bfq_put_queue(__bfqq);
4062 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4064 if (bfqq == bfqd->in_service_queue) {
4065 __bfq_bfqq_expire(bfqd, bfqq);
4066 bfq_schedule_dispatch(bfqd);
4069 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
4071 bfq_put_cooperator(bfqq);
4073 bfq_put_queue(bfqq); /* release process reference */
4076 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
4078 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4079 struct bfq_data *bfqd;
4082 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
4085 unsigned long flags;
4087 spin_lock_irqsave(&bfqd->lock, flags);
4088 bfq_exit_bfqq(bfqd, bfqq);
4089 bic_set_bfqq(bic, NULL, is_sync);
4090 spin_unlock_irqrestore(&bfqd->lock, flags);
4094 static void bfq_exit_icq(struct io_cq *icq)
4096 struct bfq_io_cq *bic = icq_to_bic(icq);
4098 bfq_exit_icq_bfqq(bic, true);
4099 bfq_exit_icq_bfqq(bic, false);
4103 * Update the entity prio values; note that the new values will not
4104 * be used until the next (re)activation.
4107 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
4109 struct task_struct *tsk = current;
4111 struct bfq_data *bfqd = bfqq->bfqd;
4116 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4117 switch (ioprio_class) {
4119 dev_err(bfqq->bfqd->queue->backing_dev_info->dev,
4120 "bfq: bad prio class %d\n", ioprio_class);
4122 case IOPRIO_CLASS_NONE:
4124 * No prio set, inherit CPU scheduling settings.
4126 bfqq->new_ioprio = task_nice_ioprio(tsk);
4127 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
4129 case IOPRIO_CLASS_RT:
4130 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4131 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
4133 case IOPRIO_CLASS_BE:
4134 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4135 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
4137 case IOPRIO_CLASS_IDLE:
4138 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
4139 bfqq->new_ioprio = 7;
4143 if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
4144 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
4146 bfqq->new_ioprio = IOPRIO_BE_NR;
4149 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
4150 bfqq->entity.prio_changed = 1;
4153 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4154 struct bio *bio, bool is_sync,
4155 struct bfq_io_cq *bic);
4157 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
4159 struct bfq_data *bfqd = bic_to_bfqd(bic);
4160 struct bfq_queue *bfqq;
4161 int ioprio = bic->icq.ioc->ioprio;
4164 * This condition may trigger on a newly created bic, be sure to
4165 * drop the lock before returning.
4167 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
4170 bic->ioprio = ioprio;
4172 bfqq = bic_to_bfqq(bic, false);
4174 /* release process reference on this queue */
4175 bfq_put_queue(bfqq);
4176 bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
4177 bic_set_bfqq(bic, bfqq, false);
4180 bfqq = bic_to_bfqq(bic, true);
4182 bfq_set_next_ioprio_data(bfqq, bic);
4185 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4186 struct bfq_io_cq *bic, pid_t pid, int is_sync)
4188 RB_CLEAR_NODE(&bfqq->entity.rb_node);
4189 INIT_LIST_HEAD(&bfqq->fifo);
4190 INIT_HLIST_NODE(&bfqq->burst_list_node);
4196 bfq_set_next_ioprio_data(bfqq, bic);
4200 * No need to mark as has_short_ttime if in
4201 * idle_class, because no device idling is performed
4202 * for queues in idle class
4204 if (!bfq_class_idle(bfqq))
4205 /* tentatively mark as has_short_ttime */
4206 bfq_mark_bfqq_has_short_ttime(bfqq);
4207 bfq_mark_bfqq_sync(bfqq);
4208 bfq_mark_bfqq_just_created(bfqq);
4210 bfq_clear_bfqq_sync(bfqq);
4212 /* set end request to minus infinity from now */
4213 bfqq->ttime.last_end_request = ktime_get_ns() + 1;
4215 bfq_mark_bfqq_IO_bound(bfqq);
4219 /* Tentative initial value to trade off between thr and lat */
4220 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
4221 bfqq->budget_timeout = bfq_smallest_from_now();
4224 bfqq->last_wr_start_finish = jiffies;
4225 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
4226 bfqq->split_time = bfq_smallest_from_now();
4229 * To not forget the possibly high bandwidth consumed by a
4230 * process/queue in the recent past,
4231 * bfq_bfqq_softrt_next_start() returns a value at least equal
4232 * to the current value of bfqq->soft_rt_next_start (see
4233 * comments on bfq_bfqq_softrt_next_start). Set
4234 * soft_rt_next_start to now, to mean that bfqq has consumed
4235 * no bandwidth so far.
4237 bfqq->soft_rt_next_start = jiffies;
4239 /* first request is almost certainly seeky */
4240 bfqq->seek_history = 1;
4243 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
4244 struct bfq_group *bfqg,
4245 int ioprio_class, int ioprio)
4247 switch (ioprio_class) {
4248 case IOPRIO_CLASS_RT:
4249 return &bfqg->async_bfqq[0][ioprio];
4250 case IOPRIO_CLASS_NONE:
4251 ioprio = IOPRIO_NORM;
4253 case IOPRIO_CLASS_BE:
4254 return &bfqg->async_bfqq[1][ioprio];
4255 case IOPRIO_CLASS_IDLE:
4256 return &bfqg->async_idle_bfqq;
4262 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4263 struct bio *bio, bool is_sync,
4264 struct bfq_io_cq *bic)
4266 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4267 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4268 struct bfq_queue **async_bfqq = NULL;
4269 struct bfq_queue *bfqq;
4270 struct bfq_group *bfqg;
4274 bfqg = bfq_find_set_group(bfqd, bio_blkcg(bio));
4276 bfqq = &bfqd->oom_bfqq;
4281 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
4288 bfqq = kmem_cache_alloc_node(bfq_pool,
4289 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
4293 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
4295 bfq_init_entity(&bfqq->entity, bfqg);
4296 bfq_log_bfqq(bfqd, bfqq, "allocated");
4298 bfqq = &bfqd->oom_bfqq;
4299 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
4304 * Pin the queue now that it's allocated, scheduler exit will
4309 * Extra group reference, w.r.t. sync
4310 * queue. This extra reference is removed
4311 * only if bfqq->bfqg disappears, to
4312 * guarantee that this queue is not freed
4313 * until its group goes away.
4315 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
4321 bfqq->ref++; /* get a process reference to this queue */
4322 bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
4327 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
4328 struct bfq_queue *bfqq)
4330 struct bfq_ttime *ttime = &bfqq->ttime;
4331 u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
4333 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
4335 ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
4336 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
4337 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
4338 ttime->ttime_samples);
4342 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4345 bfqq->seek_history <<= 1;
4346 bfqq->seek_history |=
4347 get_sdist(bfqq->last_request_pos, rq) > BFQQ_SEEK_THR &&
4348 (!blk_queue_nonrot(bfqd->queue) ||
4349 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT);
4352 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
4353 struct bfq_queue *bfqq,
4354 struct bfq_io_cq *bic)
4356 bool has_short_ttime = true;
4359 * No need to update has_short_ttime if bfqq is async or in
4360 * idle io prio class, or if bfq_slice_idle is zero, because
4361 * no device idling is performed for bfqq in this case.
4363 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
4364 bfqd->bfq_slice_idle == 0)
4367 /* Idle window just restored, statistics are meaningless. */
4368 if (time_is_after_eq_jiffies(bfqq->split_time +
4369 bfqd->bfq_wr_min_idle_time))
4372 /* Think time is infinite if no process is linked to
4373 * bfqq. Otherwise check average think time to
4374 * decide whether to mark as has_short_ttime
4376 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
4377 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
4378 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle))
4379 has_short_ttime = false;
4381 bfq_log_bfqq(bfqd, bfqq, "update_has_short_ttime: has_short_ttime %d",
4384 if (has_short_ttime)
4385 bfq_mark_bfqq_has_short_ttime(bfqq);
4387 bfq_clear_bfqq_has_short_ttime(bfqq);
4391 * Called when a new fs request (rq) is added to bfqq. Check if there's
4392 * something we should do about it.
4394 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4397 struct bfq_io_cq *bic = RQ_BIC(rq);
4399 if (rq->cmd_flags & REQ_META)
4400 bfqq->meta_pending++;
4402 bfq_update_io_thinktime(bfqd, bfqq);
4403 bfq_update_has_short_ttime(bfqd, bfqq, bic);
4404 bfq_update_io_seektime(bfqd, bfqq, rq);
4406 bfq_log_bfqq(bfqd, bfqq,
4407 "rq_enqueued: has_short_ttime=%d (seeky %d)",
4408 bfq_bfqq_has_short_ttime(bfqq), BFQQ_SEEKY(bfqq));
4410 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4412 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
4413 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
4414 blk_rq_sectors(rq) < 32;
4415 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
4418 * There is just this request queued: if the request
4419 * is small and the queue is not to be expired, then
4422 * In this way, if the device is being idled to wait
4423 * for a new request from the in-service queue, we
4424 * avoid unplugging the device and committing the
4425 * device to serve just a small request. On the
4426 * contrary, we wait for the block layer to decide
4427 * when to unplug the device: hopefully, new requests
4428 * will be merged to this one quickly, then the device
4429 * will be unplugged and larger requests will be
4432 if (small_req && !budget_timeout)
4436 * A large enough request arrived, or the queue is to
4437 * be expired: in both cases disk idling is to be
4438 * stopped, so clear wait_request flag and reset
4441 bfq_clear_bfqq_wait_request(bfqq);
4442 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4445 * The queue is not empty, because a new request just
4446 * arrived. Hence we can safely expire the queue, in
4447 * case of budget timeout, without risking that the
4448 * timestamps of the queue are not updated correctly.
4449 * See [1] for more details.
4452 bfq_bfqq_expire(bfqd, bfqq, false,
4453 BFQQE_BUDGET_TIMEOUT);
4457 /* returns true if it causes the idle timer to be disabled */
4458 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
4460 struct bfq_queue *bfqq = RQ_BFQQ(rq),
4461 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
4462 bool waiting, idle_timer_disabled = false;
4465 if (bic_to_bfqq(RQ_BIC(rq), 1) != bfqq)
4466 new_bfqq = bic_to_bfqq(RQ_BIC(rq), 1);
4468 * Release the request's reference to the old bfqq
4469 * and make sure one is taken to the shared queue.
4471 new_bfqq->allocated++;
4475 * If the bic associated with the process
4476 * issuing this request still points to bfqq
4477 * (and thus has not been already redirected
4478 * to new_bfqq or even some other bfq_queue),
4479 * then complete the merge and redirect it to
4482 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
4483 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
4486 bfq_clear_bfqq_just_created(bfqq);
4488 * rq is about to be enqueued into new_bfqq,
4489 * release rq reference on bfqq
4491 bfq_put_queue(bfqq);
4492 rq->elv.priv[1] = new_bfqq;
4496 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
4497 bfq_add_request(rq);
4498 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
4500 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
4501 list_add_tail(&rq->queuelist, &bfqq->fifo);
4503 bfq_rq_enqueued(bfqd, bfqq, rq);
4505 return idle_timer_disabled;
4508 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
4509 static void bfq_update_insert_stats(struct request_queue *q,
4510 struct bfq_queue *bfqq,
4511 bool idle_timer_disabled,
4512 unsigned int cmd_flags)
4518 * bfqq still exists, because it can disappear only after
4519 * either it is merged with another queue, or the process it
4520 * is associated with exits. But both actions must be taken by
4521 * the same process currently executing this flow of
4524 * In addition, the following queue lock guarantees that
4525 * bfqq_group(bfqq) exists as well.
4527 spin_lock_irq(q->queue_lock);
4528 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
4529 if (idle_timer_disabled)
4530 bfqg_stats_update_idle_time(bfqq_group(bfqq));
4531 spin_unlock_irq(q->queue_lock);
4534 static inline void bfq_update_insert_stats(struct request_queue *q,
4535 struct bfq_queue *bfqq,
4536 bool idle_timer_disabled,
4537 unsigned int cmd_flags) {}
4540 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
4543 struct request_queue *q = hctx->queue;
4544 struct bfq_data *bfqd = q->elevator->elevator_data;
4545 struct bfq_queue *bfqq;
4546 bool idle_timer_disabled = false;
4547 unsigned int cmd_flags;
4549 spin_lock_irq(&bfqd->lock);
4550 if (blk_mq_sched_try_insert_merge(q, rq)) {
4551 spin_unlock_irq(&bfqd->lock);
4555 spin_unlock_irq(&bfqd->lock);
4557 blk_mq_sched_request_inserted(rq);
4559 spin_lock_irq(&bfqd->lock);
4560 bfqq = bfq_init_rq(rq);
4561 if (at_head || blk_rq_is_passthrough(rq)) {
4563 list_add(&rq->queuelist, &bfqd->dispatch);
4565 list_add_tail(&rq->queuelist, &bfqd->dispatch);
4566 } else { /* bfqq is assumed to be non null here */
4567 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
4569 * Update bfqq, because, if a queue merge has occurred
4570 * in __bfq_insert_request, then rq has been
4571 * redirected into a new queue.
4575 if (rq_mergeable(rq)) {
4576 elv_rqhash_add(q, rq);
4583 * Cache cmd_flags before releasing scheduler lock, because rq
4584 * may disappear afterwards (for example, because of a request
4587 cmd_flags = rq->cmd_flags;
4589 spin_unlock_irq(&bfqd->lock);
4591 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
4595 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
4596 struct list_head *list, bool at_head)
4598 while (!list_empty(list)) {
4601 rq = list_first_entry(list, struct request, queuelist);
4602 list_del_init(&rq->queuelist);
4603 bfq_insert_request(hctx, rq, at_head);
4607 static void bfq_update_hw_tag(struct bfq_data *bfqd)
4609 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
4610 bfqd->rq_in_driver);
4612 if (bfqd->hw_tag == 1)
4616 * This sample is valid if the number of outstanding requests
4617 * is large enough to allow a queueing behavior. Note that the
4618 * sum is not exact, as it's not taking into account deactivated
4621 if (bfqd->rq_in_driver + bfqd->queued < BFQ_HW_QUEUE_THRESHOLD)
4624 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
4627 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
4628 bfqd->max_rq_in_driver = 0;
4629 bfqd->hw_tag_samples = 0;
4632 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
4637 bfq_update_hw_tag(bfqd);
4639 bfqd->rq_in_driver--;
4642 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
4644 * Set budget_timeout (which we overload to store the
4645 * time at which the queue remains with no backlog and
4646 * no outstanding request; used by the weight-raising
4649 bfqq->budget_timeout = jiffies;
4651 bfq_weights_tree_remove(bfqd, bfqq);
4654 now_ns = ktime_get_ns();
4656 bfqq->ttime.last_end_request = now_ns;
4659 * Using us instead of ns, to get a reasonable precision in
4660 * computing rate in next check.
4662 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
4665 * If the request took rather long to complete, and, according
4666 * to the maximum request size recorded, this completion latency
4667 * implies that the request was certainly served at a very low
4668 * rate (less than 1M sectors/sec), then the whole observation
4669 * interval that lasts up to this time instant cannot be a
4670 * valid time interval for computing a new peak rate. Invoke
4671 * bfq_update_rate_reset to have the following three steps
4673 * - close the observation interval at the last (previous)
4674 * request dispatch or completion
4675 * - compute rate, if possible, for that observation interval
4676 * - reset to zero samples, which will trigger a proper
4677 * re-initialization of the observation interval on next
4680 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
4681 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
4682 1UL<<(BFQ_RATE_SHIFT - 10))
4683 bfq_update_rate_reset(bfqd, NULL);
4684 bfqd->last_completion = now_ns;
4687 * If we are waiting to discover whether the request pattern
4688 * of the task associated with the queue is actually
4689 * isochronous, and both requisites for this condition to hold
4690 * are now satisfied, then compute soft_rt_next_start (see the
4691 * comments on the function bfq_bfqq_softrt_next_start()). We
4692 * schedule this delayed check when bfqq expires, if it still
4693 * has in-flight requests.
4695 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
4696 RB_EMPTY_ROOT(&bfqq->sort_list))
4697 bfqq->soft_rt_next_start =
4698 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4701 * If this is the in-service queue, check if it needs to be expired,
4702 * or if we want to idle in case it has no pending requests.
4704 if (bfqd->in_service_queue == bfqq) {
4705 if (bfq_bfqq_must_idle(bfqq)) {
4706 if (bfqq->dispatched == 0)
4707 bfq_arm_slice_timer(bfqd);
4709 * If we get here, we do not expire bfqq, even
4710 * if bfqq was in budget timeout or had no
4711 * more requests (as controlled in the next
4712 * conditional instructions). The reason for
4713 * not expiring bfqq is as follows.
4715 * Here bfqq->dispatched > 0 holds, but
4716 * bfq_bfqq_must_idle() returned true. This
4717 * implies that, even if no request arrives
4718 * for bfqq before bfqq->dispatched reaches 0,
4719 * bfqq will, however, not be expired on the
4720 * completion event that causes bfqq->dispatch
4721 * to reach zero. In contrast, on this event,
4722 * bfqq will start enjoying device idling
4723 * (I/O-dispatch plugging).
4725 * But, if we expired bfqq here, bfqq would
4726 * not have the chance to enjoy device idling
4727 * when bfqq->dispatched finally reaches
4728 * zero. This would expose bfqq to violation
4729 * of its reserved service guarantees.
4732 } else if (bfq_may_expire_for_budg_timeout(bfqq))
4733 bfq_bfqq_expire(bfqd, bfqq, false,
4734 BFQQE_BUDGET_TIMEOUT);
4735 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
4736 (bfqq->dispatched == 0 ||
4737 !bfq_better_to_idle(bfqq)))
4738 bfq_bfqq_expire(bfqd, bfqq, false,
4739 BFQQE_NO_MORE_REQUESTS);
4742 if (!bfqd->rq_in_driver)
4743 bfq_schedule_dispatch(bfqd);
4746 static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq)
4750 bfq_put_queue(bfqq);
4754 * Handle either a requeue or a finish for rq. The things to do are
4755 * the same in both cases: all references to rq are to be dropped. In
4756 * particular, rq is considered completed from the point of view of
4759 static void bfq_finish_requeue_request(struct request *rq)
4761 struct bfq_queue *bfqq = RQ_BFQQ(rq);
4762 struct bfq_data *bfqd;
4765 * Requeue and finish hooks are invoked in blk-mq without
4766 * checking whether the involved request is actually still
4767 * referenced in the scheduler. To handle this fact, the
4768 * following two checks make this function exit in case of
4769 * spurious invocations, for which there is nothing to do.
4771 * First, check whether rq has nothing to do with an elevator.
4773 if (unlikely(!(rq->rq_flags & RQF_ELVPRIV)))
4777 * rq either is not associated with any icq, or is an already
4778 * requeued request that has not (yet) been re-inserted into
4781 if (!rq->elv.icq || !bfqq)
4786 if (rq->rq_flags & RQF_STARTED)
4787 bfqg_stats_update_completion(bfqq_group(bfqq),
4789 rq->io_start_time_ns,
4792 if (likely(rq->rq_flags & RQF_STARTED)) {
4793 unsigned long flags;
4795 spin_lock_irqsave(&bfqd->lock, flags);
4797 bfq_completed_request(bfqq, bfqd);
4798 bfq_finish_requeue_request_body(bfqq);
4800 spin_unlock_irqrestore(&bfqd->lock, flags);
4803 * Request rq may be still/already in the scheduler,
4804 * in which case we need to remove it (this should
4805 * never happen in case of requeue). And we cannot
4806 * defer such a check and removal, to avoid
4807 * inconsistencies in the time interval from the end
4808 * of this function to the start of the deferred work.
4809 * This situation seems to occur only in process
4810 * context, as a consequence of a merge. In the
4811 * current version of the code, this implies that the
4815 if (!RB_EMPTY_NODE(&rq->rb_node)) {
4816 bfq_remove_request(rq->q, rq);
4817 bfqg_stats_update_io_remove(bfqq_group(bfqq),
4820 bfq_finish_requeue_request_body(bfqq);
4824 * Reset private fields. In case of a requeue, this allows
4825 * this function to correctly do nothing if it is spuriously
4826 * invoked again on this same request (see the check at the
4827 * beginning of the function). Probably, a better general
4828 * design would be to prevent blk-mq from invoking the requeue
4829 * or finish hooks of an elevator, for a request that is not
4830 * referred by that elevator.
4832 * Resetting the following fields would break the
4833 * request-insertion logic if rq is re-inserted into a bfq
4834 * internal queue, without a re-preparation. Here we assume
4835 * that re-insertions of requeued requests, without
4836 * re-preparation, can happen only for pass_through or at_head
4837 * requests (which are not re-inserted into bfq internal
4840 rq->elv.priv[0] = NULL;
4841 rq->elv.priv[1] = NULL;
4845 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
4846 * was the last process referring to that bfqq.
4848 static struct bfq_queue *
4849 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
4851 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
4853 if (bfqq_process_refs(bfqq) == 1) {
4854 bfqq->pid = current->pid;
4855 bfq_clear_bfqq_coop(bfqq);
4856 bfq_clear_bfqq_split_coop(bfqq);
4860 bic_set_bfqq(bic, NULL, 1);
4862 bfq_put_cooperator(bfqq);
4864 bfq_put_queue(bfqq);
4868 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
4869 struct bfq_io_cq *bic,
4871 bool split, bool is_sync,
4874 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4876 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
4883 bfq_put_queue(bfqq);
4884 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
4886 bic_set_bfqq(bic, bfqq, is_sync);
4887 if (split && is_sync) {
4888 if ((bic->was_in_burst_list && bfqd->large_burst) ||
4889 bic->saved_in_large_burst)
4890 bfq_mark_bfqq_in_large_burst(bfqq);
4892 bfq_clear_bfqq_in_large_burst(bfqq);
4893 if (bic->was_in_burst_list)
4895 * If bfqq was in the current
4896 * burst list before being
4897 * merged, then we have to add
4898 * it back. And we do not need
4899 * to increase burst_size, as
4900 * we did not decrement
4901 * burst_size when we removed
4902 * bfqq from the burst list as
4903 * a consequence of a merge
4905 * bfq_put_queue). In this
4906 * respect, it would be rather
4907 * costly to know whether the
4908 * current burst list is still
4909 * the same burst list from
4910 * which bfqq was removed on
4911 * the merge. To avoid this
4912 * cost, if bfqq was in a
4913 * burst list, then we add
4914 * bfqq to the current burst
4915 * list without any further
4916 * check. This can cause
4917 * inappropriate insertions,
4918 * but rarely enough to not
4919 * harm the detection of large
4920 * bursts significantly.
4922 hlist_add_head(&bfqq->burst_list_node,
4925 bfqq->split_time = jiffies;
4932 * Only reset private fields. The actual request preparation will be
4933 * performed by bfq_init_rq, when rq is either inserted or merged. See
4934 * comments on bfq_init_rq for the reason behind this delayed
4937 static void bfq_prepare_request(struct request *rq, struct bio *bio)
4940 * Regardless of whether we have an icq attached, we have to
4941 * clear the scheduler pointers, as they might point to
4942 * previously allocated bic/bfqq structs.
4944 rq->elv.priv[0] = rq->elv.priv[1] = NULL;
4948 * If needed, init rq, allocate bfq data structures associated with
4949 * rq, and increment reference counters in the destination bfq_queue
4950 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
4951 * not associated with any bfq_queue.
4953 * This function is invoked by the functions that perform rq insertion
4954 * or merging. One may have expected the above preparation operations
4955 * to be performed in bfq_prepare_request, and not delayed to when rq
4956 * is inserted or merged. The rationale behind this delayed
4957 * preparation is that, after the prepare_request hook is invoked for
4958 * rq, rq may still be transformed into a request with no icq, i.e., a
4959 * request not associated with any queue. No bfq hook is invoked to
4960 * signal this tranformation. As a consequence, should these
4961 * preparation operations be performed when the prepare_request hook
4962 * is invoked, and should rq be transformed one moment later, bfq
4963 * would end up in an inconsistent state, because it would have
4964 * incremented some queue counters for an rq destined to
4965 * transformation, without any chance to correctly lower these
4966 * counters back. In contrast, no transformation can still happen for
4967 * rq after rq has been inserted or merged. So, it is safe to execute
4968 * these preparation operations when rq is finally inserted or merged.
4970 static struct bfq_queue *bfq_init_rq(struct request *rq)
4972 struct request_queue *q = rq->q;
4973 struct bio *bio = rq->bio;
4974 struct bfq_data *bfqd = q->elevator->elevator_data;
4975 struct bfq_io_cq *bic;
4976 const int is_sync = rq_is_sync(rq);
4977 struct bfq_queue *bfqq;
4978 bool new_queue = false;
4979 bool bfqq_already_existing = false, split = false;
4981 if (unlikely(!rq->elv.icq))
4985 * Assuming that elv.priv[1] is set only if everything is set
4986 * for this rq. This holds true, because this function is
4987 * invoked only for insertion or merging, and, after such
4988 * events, a request cannot be manipulated any longer before
4989 * being removed from bfq.
4991 if (rq->elv.priv[1])
4992 return rq->elv.priv[1];
4994 bic = icq_to_bic(rq->elv.icq);
4996 bfq_check_ioprio_change(bic, bio);
4998 bfq_bic_update_cgroup(bic, bio);
5000 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
5003 if (likely(!new_queue)) {
5004 /* If the queue was seeky for too long, break it apart. */
5005 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
5006 bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
5008 /* Update bic before losing reference to bfqq */
5009 if (bfq_bfqq_in_large_burst(bfqq))
5010 bic->saved_in_large_burst = true;
5012 bfqq = bfq_split_bfqq(bic, bfqq);
5016 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
5020 bfqq_already_existing = true;
5026 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
5027 rq, bfqq, bfqq->ref);
5029 rq->elv.priv[0] = bic;
5030 rq->elv.priv[1] = bfqq;
5033 * If a bfq_queue has only one process reference, it is owned
5034 * by only this bic: we can then set bfqq->bic = bic. in
5035 * addition, if the queue has also just been split, we have to
5038 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
5042 * The queue has just been split from a shared
5043 * queue: restore the idle window and the
5044 * possible weight raising period.
5046 bfq_bfqq_resume_state(bfqq, bfqd, bic,
5047 bfqq_already_existing);
5051 if (unlikely(bfq_bfqq_just_created(bfqq)))
5052 bfq_handle_burst(bfqd, bfqq);
5057 static void bfq_idle_slice_timer_body(struct bfq_queue *bfqq)
5059 struct bfq_data *bfqd = bfqq->bfqd;
5060 enum bfqq_expiration reason;
5061 unsigned long flags;
5063 spin_lock_irqsave(&bfqd->lock, flags);
5064 bfq_clear_bfqq_wait_request(bfqq);
5066 if (bfqq != bfqd->in_service_queue) {
5067 spin_unlock_irqrestore(&bfqd->lock, flags);
5071 if (bfq_bfqq_budget_timeout(bfqq))
5073 * Also here the queue can be safely expired
5074 * for budget timeout without wasting
5077 reason = BFQQE_BUDGET_TIMEOUT;
5078 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
5080 * The queue may not be empty upon timer expiration,
5081 * because we may not disable the timer when the
5082 * first request of the in-service queue arrives
5083 * during disk idling.
5085 reason = BFQQE_TOO_IDLE;
5087 goto schedule_dispatch;
5089 bfq_bfqq_expire(bfqd, bfqq, true, reason);
5092 spin_unlock_irqrestore(&bfqd->lock, flags);
5093 bfq_schedule_dispatch(bfqd);
5097 * Handler of the expiration of the timer running if the in-service queue
5098 * is idling inside its time slice.
5100 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
5102 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
5104 struct bfq_queue *bfqq = bfqd->in_service_queue;
5107 * Theoretical race here: the in-service queue can be NULL or
5108 * different from the queue that was idling if a new request
5109 * arrives for the current queue and there is a full dispatch
5110 * cycle that changes the in-service queue. This can hardly
5111 * happen, but in the worst case we just expire a queue too
5115 bfq_idle_slice_timer_body(bfqq);
5117 return HRTIMER_NORESTART;
5120 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
5121 struct bfq_queue **bfqq_ptr)
5123 struct bfq_queue *bfqq = *bfqq_ptr;
5125 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
5127 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
5129 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
5131 bfq_put_queue(bfqq);
5137 * Release all the bfqg references to its async queues. If we are
5138 * deallocating the group these queues may still contain requests, so
5139 * we reparent them to the root cgroup (i.e., the only one that will
5140 * exist for sure until all the requests on a device are gone).
5142 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
5146 for (i = 0; i < 2; i++)
5147 for (j = 0; j < IOPRIO_BE_NR; j++)
5148 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
5150 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
5154 * See the comments on bfq_limit_depth for the purpose of
5155 * the depths set in the function. Return minimum shallow depth we'll use.
5157 static unsigned int bfq_update_depths(struct bfq_data *bfqd,
5158 struct sbitmap_queue *bt)
5160 unsigned int i, j, min_shallow = UINT_MAX;
5163 * In-word depths if no bfq_queue is being weight-raised:
5164 * leaving 25% of tags only for sync reads.
5166 * In next formulas, right-shift the value
5167 * (1U<<bt->sb.shift), instead of computing directly
5168 * (1U<<(bt->sb.shift - something)), to be robust against
5169 * any possible value of bt->sb.shift, without having to
5170 * limit 'something'.
5172 /* no more than 50% of tags for async I/O */
5173 bfqd->word_depths[0][0] = max((1U << bt->sb.shift) >> 1, 1U);
5175 * no more than 75% of tags for sync writes (25% extra tags
5176 * w.r.t. async I/O, to prevent async I/O from starving sync
5179 bfqd->word_depths[0][1] = max(((1U << bt->sb.shift) * 3) >> 2, 1U);
5182 * In-word depths in case some bfq_queue is being weight-
5183 * raised: leaving ~63% of tags for sync reads. This is the
5184 * highest percentage for which, in our tests, application
5185 * start-up times didn't suffer from any regression due to tag
5188 /* no more than ~18% of tags for async I/O */
5189 bfqd->word_depths[1][0] = max(((1U << bt->sb.shift) * 3) >> 4, 1U);
5190 /* no more than ~37% of tags for sync writes (~20% extra tags) */
5191 bfqd->word_depths[1][1] = max(((1U << bt->sb.shift) * 6) >> 4, 1U);
5193 for (i = 0; i < 2; i++)
5194 for (j = 0; j < 2; j++)
5195 min_shallow = min(min_shallow, bfqd->word_depths[i][j]);
5200 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
5202 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5203 struct blk_mq_tags *tags = hctx->sched_tags;
5204 unsigned int min_shallow;
5206 min_shallow = bfq_update_depths(bfqd, &tags->bitmap_tags);
5207 sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, min_shallow);
5211 static void bfq_exit_queue(struct elevator_queue *e)
5213 struct bfq_data *bfqd = e->elevator_data;
5214 struct bfq_queue *bfqq, *n;
5216 hrtimer_cancel(&bfqd->idle_slice_timer);
5218 spin_lock_irq(&bfqd->lock);
5219 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
5220 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
5221 spin_unlock_irq(&bfqd->lock);
5223 hrtimer_cancel(&bfqd->idle_slice_timer);
5225 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5226 /* release oom-queue reference to root group */
5227 bfqg_and_blkg_put(bfqd->root_group);
5229 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
5231 spin_lock_irq(&bfqd->lock);
5232 bfq_put_async_queues(bfqd, bfqd->root_group);
5233 kfree(bfqd->root_group);
5234 spin_unlock_irq(&bfqd->lock);
5240 static void bfq_init_root_group(struct bfq_group *root_group,
5241 struct bfq_data *bfqd)
5245 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5246 root_group->entity.parent = NULL;
5247 root_group->my_entity = NULL;
5248 root_group->bfqd = bfqd;
5250 root_group->rq_pos_tree = RB_ROOT;
5251 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
5252 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
5253 root_group->sched_data.bfq_class_idle_last_service = jiffies;
5256 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
5258 struct bfq_data *bfqd;
5259 struct elevator_queue *eq;
5261 eq = elevator_alloc(q, e);
5265 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
5267 kobject_put(&eq->kobj);
5270 eq->elevator_data = bfqd;
5272 spin_lock_irq(q->queue_lock);
5274 spin_unlock_irq(q->queue_lock);
5277 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
5278 * Grab a permanent reference to it, so that the normal code flow
5279 * will not attempt to free it.
5281 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
5282 bfqd->oom_bfqq.ref++;
5283 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
5284 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
5285 bfqd->oom_bfqq.entity.new_weight =
5286 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
5288 /* oom_bfqq does not participate to bursts */
5289 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
5292 * Trigger weight initialization, according to ioprio, at the
5293 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
5294 * class won't be changed any more.
5296 bfqd->oom_bfqq.entity.prio_changed = 1;
5300 INIT_LIST_HEAD(&bfqd->dispatch);
5302 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
5304 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
5306 bfqd->queue_weights_tree = RB_ROOT;
5307 bfqd->group_weights_tree = RB_ROOT;
5309 INIT_LIST_HEAD(&bfqd->active_list);
5310 INIT_LIST_HEAD(&bfqd->idle_list);
5311 INIT_HLIST_HEAD(&bfqd->burst_list);
5315 bfqd->bfq_max_budget = bfq_default_max_budget;
5317 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
5318 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
5319 bfqd->bfq_back_max = bfq_back_max;
5320 bfqd->bfq_back_penalty = bfq_back_penalty;
5321 bfqd->bfq_slice_idle = bfq_slice_idle;
5322 bfqd->bfq_timeout = bfq_timeout;
5324 bfqd->bfq_requests_within_timer = 120;
5326 bfqd->bfq_large_burst_thresh = 8;
5327 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
5329 bfqd->low_latency = true;
5332 * Trade-off between responsiveness and fairness.
5334 bfqd->bfq_wr_coeff = 30;
5335 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
5336 bfqd->bfq_wr_max_time = 0;
5337 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
5338 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
5339 bfqd->bfq_wr_max_softrt_rate = 7000; /*
5340 * Approximate rate required
5341 * to playback or record a
5342 * high-definition compressed
5345 bfqd->wr_busy_queues = 0;
5348 * Begin by assuming, optimistically, that the device peak
5349 * rate is equal to 2/3 of the highest reference rate.
5351 bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
5352 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
5353 bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
5355 spin_lock_init(&bfqd->lock);
5358 * The invocation of the next bfq_create_group_hierarchy
5359 * function is the head of a chain of function calls
5360 * (bfq_create_group_hierarchy->blkcg_activate_policy->
5361 * blk_mq_freeze_queue) that may lead to the invocation of the
5362 * has_work hook function. For this reason,
5363 * bfq_create_group_hierarchy is invoked only after all
5364 * scheduler data has been initialized, apart from the fields
5365 * that can be initialized only after invoking
5366 * bfq_create_group_hierarchy. This, in particular, enables
5367 * has_work to correctly return false. Of course, to avoid
5368 * other inconsistencies, the blk-mq stack must then refrain
5369 * from invoking further scheduler hooks before this init
5370 * function is finished.
5372 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
5373 if (!bfqd->root_group)
5375 bfq_init_root_group(bfqd->root_group, bfqd);
5376 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
5378 wbt_disable_default(q);
5383 kobject_put(&eq->kobj);
5387 static void bfq_slab_kill(void)
5389 kmem_cache_destroy(bfq_pool);
5392 static int __init bfq_slab_setup(void)
5394 bfq_pool = KMEM_CACHE(bfq_queue, 0);
5400 static ssize_t bfq_var_show(unsigned int var, char *page)
5402 return sprintf(page, "%u\n", var);
5405 static int bfq_var_store(unsigned long *var, const char *page)
5407 unsigned long new_val;
5408 int ret = kstrtoul(page, 10, &new_val);
5416 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
5417 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
5419 struct bfq_data *bfqd = e->elevator_data; \
5420 u64 __data = __VAR; \
5422 __data = jiffies_to_msecs(__data); \
5423 else if (__CONV == 2) \
5424 __data = div_u64(__data, NSEC_PER_MSEC); \
5425 return bfq_var_show(__data, (page)); \
5427 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
5428 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
5429 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
5430 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
5431 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
5432 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
5433 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
5434 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
5435 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
5436 #undef SHOW_FUNCTION
5438 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
5439 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
5441 struct bfq_data *bfqd = e->elevator_data; \
5442 u64 __data = __VAR; \
5443 __data = div_u64(__data, NSEC_PER_USEC); \
5444 return bfq_var_show(__data, (page)); \
5446 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
5447 #undef USEC_SHOW_FUNCTION
5449 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
5451 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
5453 struct bfq_data *bfqd = e->elevator_data; \
5454 unsigned long __data, __min = (MIN), __max = (MAX); \
5457 ret = bfq_var_store(&__data, (page)); \
5460 if (__data < __min) \
5462 else if (__data > __max) \
5465 *(__PTR) = msecs_to_jiffies(__data); \
5466 else if (__CONV == 2) \
5467 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
5469 *(__PTR) = __data; \
5472 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
5474 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
5476 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
5477 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
5479 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
5480 #undef STORE_FUNCTION
5482 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
5483 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
5485 struct bfq_data *bfqd = e->elevator_data; \
5486 unsigned long __data, __min = (MIN), __max = (MAX); \
5489 ret = bfq_var_store(&__data, (page)); \
5492 if (__data < __min) \
5494 else if (__data > __max) \
5496 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
5499 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
5501 #undef USEC_STORE_FUNCTION
5503 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
5504 const char *page, size_t count)
5506 struct bfq_data *bfqd = e->elevator_data;
5507 unsigned long __data;
5510 ret = bfq_var_store(&__data, (page));
5515 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5517 if (__data > INT_MAX)
5519 bfqd->bfq_max_budget = __data;
5522 bfqd->bfq_user_max_budget = __data;
5528 * Leaving this name to preserve name compatibility with cfq
5529 * parameters, but this timeout is used for both sync and async.
5531 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
5532 const char *page, size_t count)
5534 struct bfq_data *bfqd = e->elevator_data;
5535 unsigned long __data;
5538 ret = bfq_var_store(&__data, (page));
5544 else if (__data > INT_MAX)
5547 bfqd->bfq_timeout = msecs_to_jiffies(__data);
5548 if (bfqd->bfq_user_max_budget == 0)
5549 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5554 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
5555 const char *page, size_t count)
5557 struct bfq_data *bfqd = e->elevator_data;
5558 unsigned long __data;
5561 ret = bfq_var_store(&__data, (page));
5567 if (!bfqd->strict_guarantees && __data == 1
5568 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
5569 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
5571 bfqd->strict_guarantees = __data;
5576 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
5577 const char *page, size_t count)
5579 struct bfq_data *bfqd = e->elevator_data;
5580 unsigned long __data;
5583 ret = bfq_var_store(&__data, (page));
5589 if (__data == 0 && bfqd->low_latency != 0)
5591 bfqd->low_latency = __data;
5596 #define BFQ_ATTR(name) \
5597 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
5599 static struct elv_fs_entry bfq_attrs[] = {
5600 BFQ_ATTR(fifo_expire_sync),
5601 BFQ_ATTR(fifo_expire_async),
5602 BFQ_ATTR(back_seek_max),
5603 BFQ_ATTR(back_seek_penalty),
5604 BFQ_ATTR(slice_idle),
5605 BFQ_ATTR(slice_idle_us),
5606 BFQ_ATTR(max_budget),
5607 BFQ_ATTR(timeout_sync),
5608 BFQ_ATTR(strict_guarantees),
5609 BFQ_ATTR(low_latency),
5613 static struct elevator_type iosched_bfq_mq = {
5615 .limit_depth = bfq_limit_depth,
5616 .prepare_request = bfq_prepare_request,
5617 .requeue_request = bfq_finish_requeue_request,
5618 .finish_request = bfq_finish_requeue_request,
5619 .exit_icq = bfq_exit_icq,
5620 .insert_requests = bfq_insert_requests,
5621 .dispatch_request = bfq_dispatch_request,
5622 .next_request = elv_rb_latter_request,
5623 .former_request = elv_rb_former_request,
5624 .allow_merge = bfq_allow_bio_merge,
5625 .bio_merge = bfq_bio_merge,
5626 .request_merge = bfq_request_merge,
5627 .requests_merged = bfq_requests_merged,
5628 .request_merged = bfq_request_merged,
5629 .has_work = bfq_has_work,
5630 .init_hctx = bfq_init_hctx,
5631 .init_sched = bfq_init_queue,
5632 .exit_sched = bfq_exit_queue,
5636 .icq_size = sizeof(struct bfq_io_cq),
5637 .icq_align = __alignof__(struct bfq_io_cq),
5638 .elevator_attrs = bfq_attrs,
5639 .elevator_name = "bfq",
5640 .elevator_owner = THIS_MODULE,
5642 MODULE_ALIAS("bfq-iosched");
5644 static int __init bfq_init(void)
5648 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5649 ret = blkcg_policy_register(&blkcg_policy_bfq);
5655 if (bfq_slab_setup())
5659 * Times to load large popular applications for the typical
5660 * systems installed on the reference devices (see the
5661 * comments before the definition of the next
5662 * array). Actually, we use slightly lower values, as the
5663 * estimated peak rate tends to be smaller than the actual
5664 * peak rate. The reason for this last fact is that estimates
5665 * are computed over much shorter time intervals than the long
5666 * intervals typically used for benchmarking. Why? First, to
5667 * adapt more quickly to variations. Second, because an I/O
5668 * scheduler cannot rely on a peak-rate-evaluation workload to
5669 * be run for a long time.
5671 ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
5672 ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
5674 ret = elv_register(&iosched_bfq_mq);
5683 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5684 blkcg_policy_unregister(&blkcg_policy_bfq);
5689 static void __exit bfq_exit(void)
5691 elv_unregister(&iosched_bfq_mq);
5692 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5693 blkcg_policy_unregister(&blkcg_policy_bfq);
5698 module_init(bfq_init);
5699 module_exit(bfq_exit);
5701 MODULE_AUTHOR("Paolo Valente");
5702 MODULE_LICENSE("GPL");
5703 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");