2 * Budget Fair Queueing (BFQ) I/O scheduler.
4 * Based on ideas and code from CFQ:
5 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
8 * Paolo Valente <paolo.valente@unimore.it>
10 * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
11 * Arianna Avanzini <avanzini@google.com>
13 * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License as
17 * published by the Free Software Foundation; either version 2 of the
18 * License, or (at your option) any later version.
20 * This program is distributed in the hope that it will be useful,
21 * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
23 * General Public License for more details.
25 * BFQ is a proportional-share I/O scheduler, with some extra
26 * low-latency capabilities. BFQ also supports full hierarchical
27 * scheduling through cgroups. Next paragraphs provide an introduction
28 * on BFQ inner workings. Details on BFQ benefits, usage and
29 * limitations can be found in Documentation/block/bfq-iosched.txt.
31 * BFQ is a proportional-share storage-I/O scheduling algorithm based
32 * on the slice-by-slice service scheme of CFQ. But BFQ assigns
33 * budgets, measured in number of sectors, to processes instead of
34 * time slices. The device is not granted to the in-service process
35 * for a given time slice, but until it has exhausted its assigned
36 * budget. This change from the time to the service domain enables BFQ
37 * to distribute the device throughput among processes as desired,
38 * without any distortion due to throughput fluctuations, or to device
39 * internal queueing. BFQ uses an ad hoc internal scheduler, called
40 * B-WF2Q+, to schedule processes according to their budgets. More
41 * precisely, BFQ schedules queues associated with processes. Each
42 * process/queue is assigned a user-configurable weight, and B-WF2Q+
43 * guarantees that each queue receives a fraction of the throughput
44 * proportional to its weight. Thanks to the accurate policy of
45 * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
46 * processes issuing sequential requests (to boost the throughput),
47 * and yet guarantee a low latency to interactive and soft real-time
50 * In particular, to provide these low-latency guarantees, BFQ
51 * explicitly privileges the I/O of two classes of time-sensitive
52 * applications: interactive and soft real-time. In more detail, BFQ
53 * behaves this way if the low_latency parameter is set (default
54 * configuration). This feature enables BFQ to provide applications in
55 * these classes with a very low latency.
57 * To implement this feature, BFQ constantly tries to detect whether
58 * the I/O requests in a bfq_queue come from an interactive or a soft
59 * real-time application. For brevity, in these cases, the queue is
60 * said to be interactive or soft real-time. In both cases, BFQ
61 * privileges the service of the queue, over that of non-interactive
62 * and non-soft-real-time queues. This privileging is performed,
63 * mainly, by raising the weight of the queue. So, for brevity, we
64 * call just weight-raising periods the time periods during which a
65 * queue is privileged, because deemed interactive or soft real-time.
67 * The detection of soft real-time queues/applications is described in
68 * detail in the comments on the function
69 * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
70 * interactive queue works as follows: a queue is deemed interactive
71 * if it is constantly non empty only for a limited time interval,
72 * after which it does become empty. The queue may be deemed
73 * interactive again (for a limited time), if it restarts being
74 * constantly non empty, provided that this happens only after the
75 * queue has remained empty for a given minimum idle time.
77 * By default, BFQ computes automatically the above maximum time
78 * interval, i.e., the time interval after which a constantly
79 * non-empty queue stops being deemed interactive. Since a queue is
80 * weight-raised while it is deemed interactive, this maximum time
81 * interval happens to coincide with the (maximum) duration of the
82 * weight-raising for interactive queues.
84 * Finally, BFQ also features additional heuristics for
85 * preserving both a low latency and a high throughput on NCQ-capable,
86 * rotational or flash-based devices, and to get the job done quickly
87 * for applications consisting in many I/O-bound processes.
89 * NOTE: if the main or only goal, with a given device, is to achieve
90 * the maximum-possible throughput at all times, then do switch off
91 * all low-latency heuristics for that device, by setting low_latency
94 * BFQ is described in [1], where also a reference to the initial,
95 * more theoretical paper on BFQ can be found. The interested reader
96 * can find in the latter paper full details on the main algorithm, as
97 * well as formulas of the guarantees and formal proofs of all the
98 * properties. With respect to the version of BFQ presented in these
99 * papers, this implementation adds a few more heuristics, such as the
100 * ones that guarantee a low latency to interactive and soft real-time
101 * applications, and a hierarchical extension based on H-WF2Q+.
103 * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
104 * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
105 * with O(log N) complexity derives from the one introduced with EEVDF
108 * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
109 * Scheduler", Proceedings of the First Workshop on Mobile System
110 * Technologies (MST-2015), May 2015.
111 * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
113 * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
114 * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
117 * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
119 * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
120 * First: A Flexible and Accurate Mechanism for Proportional Share
121 * Resource Allocation", technical report.
123 * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
125 #include <linux/module.h>
126 #include <linux/slab.h>
127 #include <linux/blkdev.h>
128 #include <linux/cgroup.h>
129 #include <linux/elevator.h>
130 #include <linux/ktime.h>
131 #include <linux/rbtree.h>
132 #include <linux/ioprio.h>
133 #include <linux/sbitmap.h>
134 #include <linux/delay.h>
138 #include "blk-mq-tag.h"
139 #include "blk-mq-sched.h"
140 #include "bfq-iosched.h"
143 #define BFQ_BFQQ_FNS(name) \
144 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
146 __set_bit(BFQQF_##name, &(bfqq)->flags); \
148 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
150 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
152 int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
154 return test_bit(BFQQF_##name, &(bfqq)->flags); \
157 BFQ_BFQQ_FNS(just_created);
159 BFQ_BFQQ_FNS(wait_request);
160 BFQ_BFQQ_FNS(non_blocking_wait_rq);
161 BFQ_BFQQ_FNS(fifo_expire);
162 BFQ_BFQQ_FNS(has_short_ttime);
164 BFQ_BFQQ_FNS(IO_bound);
165 BFQ_BFQQ_FNS(in_large_burst);
167 BFQ_BFQQ_FNS(split_coop);
168 BFQ_BFQQ_FNS(softrt_update);
169 #undef BFQ_BFQQ_FNS \
171 /* Expiration time of sync (0) and async (1) requests, in ns. */
172 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
174 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
175 static const int bfq_back_max = 16 * 1024;
177 /* Penalty of a backwards seek, in number of sectors. */
178 static const int bfq_back_penalty = 2;
180 /* Idling period duration, in ns. */
181 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
183 /* Minimum number of assigned budgets for which stats are safe to compute. */
184 static const int bfq_stats_min_budgets = 194;
186 /* Default maximum budget values, in sectors and number of requests. */
187 static const int bfq_default_max_budget = 16 * 1024;
190 * When a sync request is dispatched, the queue that contains that
191 * request, and all the ancestor entities of that queue, are charged
192 * with the number of sectors of the request. In constrast, if the
193 * request is async, then the queue and its ancestor entities are
194 * charged with the number of sectors of the request, multiplied by
195 * the factor below. This throttles the bandwidth for async I/O,
196 * w.r.t. to sync I/O, and it is done to counter the tendency of async
197 * writes to steal I/O throughput to reads.
199 * The current value of this parameter is the result of a tuning with
200 * several hardware and software configurations. We tried to find the
201 * lowest value for which writes do not cause noticeable problems to
202 * reads. In fact, the lower this parameter, the stabler I/O control,
203 * in the following respect. The lower this parameter is, the less
204 * the bandwidth enjoyed by a group decreases
205 * - when the group does writes, w.r.t. to when it does reads;
206 * - when other groups do reads, w.r.t. to when they do writes.
208 static const int bfq_async_charge_factor = 3;
210 /* Default timeout values, in jiffies, approximating CFQ defaults. */
211 const int bfq_timeout = HZ / 8;
214 * Time limit for merging (see comments in bfq_setup_cooperator). Set
215 * to the slowest value that, in our tests, proved to be effective in
216 * removing false positives, while not causing true positives to miss
219 * As can be deduced from the low time limit below, queue merging, if
220 * successful, happens at the very beggining of the I/O of the involved
221 * cooperating processes, as a consequence of the arrival of the very
222 * first requests from each cooperator. After that, there is very
223 * little chance to find cooperators.
225 static const unsigned long bfq_merge_time_limit = HZ/10;
227 static struct kmem_cache *bfq_pool;
229 /* Below this threshold (in ns), we consider thinktime immediate. */
230 #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
232 /* hw_tag detection: parallel requests threshold and min samples needed. */
233 #define BFQ_HW_QUEUE_THRESHOLD 4
234 #define BFQ_HW_QUEUE_SAMPLES 32
236 #define BFQQ_SEEK_THR (sector_t)(8 * 100)
237 #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
238 #define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
239 #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19)
241 /* Min number of samples required to perform peak-rate update */
242 #define BFQ_RATE_MIN_SAMPLES 32
243 /* Min observation time interval required to perform a peak-rate update (ns) */
244 #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
245 /* Target observation time interval for a peak-rate update (ns) */
246 #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
249 * Shift used for peak-rate fixed precision calculations.
251 * - the current shift: 16 positions
252 * - the current type used to store rate: u32
253 * - the current unit of measure for rate: [sectors/usec], or, more precisely,
254 * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
255 * the range of rates that can be stored is
256 * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
257 * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
258 * [15, 65G] sectors/sec
259 * Which, assuming a sector size of 512B, corresponds to a range of
262 #define BFQ_RATE_SHIFT 16
265 * When configured for computing the duration of the weight-raising
266 * for interactive queues automatically (see the comments at the
267 * beginning of this file), BFQ does it using the following formula:
268 * duration = (ref_rate / r) * ref_wr_duration,
269 * where r is the peak rate of the device, and ref_rate and
270 * ref_wr_duration are two reference parameters. In particular,
271 * ref_rate is the peak rate of the reference storage device (see
272 * below), and ref_wr_duration is about the maximum time needed, with
273 * BFQ and while reading two files in parallel, to load typical large
274 * applications on the reference device (see the comments on
275 * max_service_from_wr below, for more details on how ref_wr_duration
276 * is obtained). In practice, the slower/faster the device at hand
277 * is, the more/less it takes to load applications with respect to the
278 * reference device. Accordingly, the longer/shorter BFQ grants
279 * weight raising to interactive applications.
281 * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
282 * depending on whether the device is rotational or non-rotational.
284 * In the following definitions, ref_rate[0] and ref_wr_duration[0]
285 * are the reference values for a rotational device, whereas
286 * ref_rate[1] and ref_wr_duration[1] are the reference values for a
287 * non-rotational device. The reference rates are not the actual peak
288 * rates of the devices used as a reference, but slightly lower
289 * values. The reason for using slightly lower values is that the
290 * peak-rate estimator tends to yield slightly lower values than the
291 * actual peak rate (it can yield the actual peak rate only if there
292 * is only one process doing I/O, and the process does sequential
295 * The reference peak rates are measured in sectors/usec, left-shifted
298 static int ref_rate[2] = {14000, 33000};
300 * To improve readability, a conversion function is used to initialize
301 * the following array, which entails that the array can be
302 * initialized only in a function.
304 static int ref_wr_duration[2];
307 * BFQ uses the above-detailed, time-based weight-raising mechanism to
308 * privilege interactive tasks. This mechanism is vulnerable to the
309 * following false positives: I/O-bound applications that will go on
310 * doing I/O for much longer than the duration of weight
311 * raising. These applications have basically no benefit from being
312 * weight-raised at the beginning of their I/O. On the opposite end,
313 * while being weight-raised, these applications
314 * a) unjustly steal throughput to applications that may actually need
316 * b) make BFQ uselessly perform device idling; device idling results
317 * in loss of device throughput with most flash-based storage, and may
318 * increase latencies when used purposelessly.
320 * BFQ tries to reduce these problems, by adopting the following
321 * countermeasure. To introduce this countermeasure, we need first to
322 * finish explaining how the duration of weight-raising for
323 * interactive tasks is computed.
325 * For a bfq_queue deemed as interactive, the duration of weight
326 * raising is dynamically adjusted, as a function of the estimated
327 * peak rate of the device, so as to be equal to the time needed to
328 * execute the 'largest' interactive task we benchmarked so far. By
329 * largest task, we mean the task for which each involved process has
330 * to do more I/O than for any of the other tasks we benchmarked. This
331 * reference interactive task is the start-up of LibreOffice Writer,
332 * and in this task each process/bfq_queue needs to have at most ~110K
333 * sectors transferred.
335 * This last piece of information enables BFQ to reduce the actual
336 * duration of weight-raising for at least one class of I/O-bound
337 * applications: those doing sequential or quasi-sequential I/O. An
338 * example is file copy. In fact, once started, the main I/O-bound
339 * processes of these applications usually consume the above 110K
340 * sectors in much less time than the processes of an application that
341 * is starting, because these I/O-bound processes will greedily devote
342 * almost all their CPU cycles only to their target,
343 * throughput-friendly I/O operations. This is even more true if BFQ
344 * happens to be underestimating the device peak rate, and thus
345 * overestimating the duration of weight raising. But, according to
346 * our measurements, once transferred 110K sectors, these processes
347 * have no right to be weight-raised any longer.
349 * Basing on the last consideration, BFQ ends weight-raising for a
350 * bfq_queue if the latter happens to have received an amount of
351 * service at least equal to the following constant. The constant is
352 * set to slightly more than 110K, to have a minimum safety margin.
354 * This early ending of weight-raising reduces the amount of time
355 * during which interactive false positives cause the two problems
356 * described at the beginning of these comments.
358 static const unsigned long max_service_from_wr = 120000;
360 #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
361 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
363 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
365 return bic->bfqq[is_sync];
368 void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
370 bic->bfqq[is_sync] = bfqq;
373 struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
375 return bic->icq.q->elevator->elevator_data;
379 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
380 * @icq: the iocontext queue.
382 static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
384 /* bic->icq is the first member, %NULL will convert to %NULL */
385 return container_of(icq, struct bfq_io_cq, icq);
389 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
390 * @bfqd: the lookup key.
391 * @ioc: the io_context of the process doing I/O.
392 * @q: the request queue.
394 static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd,
395 struct io_context *ioc,
396 struct request_queue *q)
400 struct bfq_io_cq *icq;
402 spin_lock_irqsave(q->queue_lock, flags);
403 icq = icq_to_bic(ioc_lookup_icq(ioc, q));
404 spin_unlock_irqrestore(q->queue_lock, flags);
413 * Scheduler run of queue, if there are requests pending and no one in the
414 * driver that will restart queueing.
416 void bfq_schedule_dispatch(struct bfq_data *bfqd)
418 if (bfqd->queued != 0) {
419 bfq_log(bfqd, "schedule dispatch");
420 blk_mq_run_hw_queues(bfqd->queue, true);
424 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
425 #define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT)
427 #define bfq_sample_valid(samples) ((samples) > 80)
430 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
431 * We choose the request that is closesr to the head right now. Distance
432 * behind the head is penalized and only allowed to a certain extent.
434 static struct request *bfq_choose_req(struct bfq_data *bfqd,
439 sector_t s1, s2, d1 = 0, d2 = 0;
440 unsigned long back_max;
441 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
442 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
443 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
445 if (!rq1 || rq1 == rq2)
450 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
452 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
454 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
456 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
459 s1 = blk_rq_pos(rq1);
460 s2 = blk_rq_pos(rq2);
463 * By definition, 1KiB is 2 sectors.
465 back_max = bfqd->bfq_back_max * 2;
468 * Strict one way elevator _except_ in the case where we allow
469 * short backward seeks which are biased as twice the cost of a
470 * similar forward seek.
474 else if (s1 + back_max >= last)
475 d1 = (last - s1) * bfqd->bfq_back_penalty;
477 wrap |= BFQ_RQ1_WRAP;
481 else if (s2 + back_max >= last)
482 d2 = (last - s2) * bfqd->bfq_back_penalty;
484 wrap |= BFQ_RQ2_WRAP;
486 /* Found required data */
489 * By doing switch() on the bit mask "wrap" we avoid having to
490 * check two variables for all permutations: --> faster!
493 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
508 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
511 * Since both rqs are wrapped,
512 * start with the one that's further behind head
513 * (--> only *one* back seek required),
514 * since back seek takes more time than forward.
524 * Async I/O can easily starve sync I/O (both sync reads and sync
525 * writes), by consuming all tags. Similarly, storms of sync writes,
526 * such as those that sync(2) may trigger, can starve sync reads.
527 * Limit depths of async I/O and sync writes so as to counter both
530 static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
532 struct bfq_data *bfqd = data->q->elevator->elevator_data;
534 if (op_is_sync(op) && !op_is_write(op))
537 data->shallow_depth =
538 bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)];
540 bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
541 __func__, bfqd->wr_busy_queues, op_is_sync(op),
542 data->shallow_depth);
545 static struct bfq_queue *
546 bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
547 sector_t sector, struct rb_node **ret_parent,
548 struct rb_node ***rb_link)
550 struct rb_node **p, *parent;
551 struct bfq_queue *bfqq = NULL;
559 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
562 * Sort strictly based on sector. Smallest to the left,
563 * largest to the right.
565 if (sector > blk_rq_pos(bfqq->next_rq))
567 else if (sector < blk_rq_pos(bfqq->next_rq))
575 *ret_parent = parent;
579 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
580 (unsigned long long)sector,
581 bfqq ? bfqq->pid : 0);
586 static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
588 return bfqq->service_from_backlogged > 0 &&
589 time_is_before_jiffies(bfqq->first_IO_time +
590 bfq_merge_time_limit);
593 void bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
595 struct rb_node **p, *parent;
596 struct bfq_queue *__bfqq;
598 if (bfqq->pos_root) {
599 rb_erase(&bfqq->pos_node, bfqq->pos_root);
600 bfqq->pos_root = NULL;
604 * bfqq cannot be merged any longer (see comments in
605 * bfq_setup_cooperator): no point in adding bfqq into the
608 if (bfq_too_late_for_merging(bfqq))
611 if (bfq_class_idle(bfqq))
616 bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
617 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
618 blk_rq_pos(bfqq->next_rq), &parent, &p);
620 rb_link_node(&bfqq->pos_node, parent, p);
621 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
623 bfqq->pos_root = NULL;
627 * Tell whether there are active queues or groups with differentiated weights.
629 static bool bfq_differentiated_weights(struct bfq_data *bfqd)
632 * For weights to differ, at least one of the trees must contain
633 * at least two nodes.
635 return (!RB_EMPTY_ROOT(&bfqd->queue_weights_tree) &&
636 (bfqd->queue_weights_tree.rb_node->rb_left ||
637 bfqd->queue_weights_tree.rb_node->rb_right)
638 #ifdef CONFIG_BFQ_GROUP_IOSCHED
640 (!RB_EMPTY_ROOT(&bfqd->group_weights_tree) &&
641 (bfqd->group_weights_tree.rb_node->rb_left ||
642 bfqd->group_weights_tree.rb_node->rb_right)
648 * The following function returns true if every queue must receive the
649 * same share of the throughput (this condition is used when deciding
650 * whether idling may be disabled, see the comments in the function
651 * bfq_better_to_idle()).
653 * Such a scenario occurs when:
654 * 1) all active queues have the same weight,
655 * 2) all active groups at the same level in the groups tree have the same
657 * 3) all active groups at the same level in the groups tree have the same
658 * number of children.
660 * Unfortunately, keeping the necessary state for evaluating exactly the
661 * above symmetry conditions would be quite complex and time-consuming.
662 * Therefore this function evaluates, instead, the following stronger
663 * sub-conditions, for which it is much easier to maintain the needed
665 * 1) all active queues have the same weight,
666 * 2) all active groups have the same weight,
667 * 3) all active groups have at most one active child each.
668 * In particular, the last two conditions are always true if hierarchical
669 * support and the cgroups interface are not enabled, thus no state needs
670 * to be maintained in this case.
672 static bool bfq_symmetric_scenario(struct bfq_data *bfqd)
674 return !bfq_differentiated_weights(bfqd);
678 * If the weight-counter tree passed as input contains no counter for
679 * the weight of the input entity, then add that counter; otherwise just
680 * increment the existing counter.
682 * Note that weight-counter trees contain few nodes in mostly symmetric
683 * scenarios. For example, if all queues have the same weight, then the
684 * weight-counter tree for the queues may contain at most one node.
685 * This holds even if low_latency is on, because weight-raised queues
686 * are not inserted in the tree.
687 * In most scenarios, the rate at which nodes are created/destroyed
690 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_entity *entity,
691 struct rb_root *root)
693 struct rb_node **new = &(root->rb_node), *parent = NULL;
696 * Do not insert if the entity is already associated with a
697 * counter, which happens if:
698 * 1) the entity is associated with a queue,
699 * 2) a request arrival has caused the queue to become both
700 * non-weight-raised, and hence change its weight, and
701 * backlogged; in this respect, each of the two events
702 * causes an invocation of this function,
703 * 3) this is the invocation of this function caused by the
704 * second event. This second invocation is actually useless,
705 * and we handle this fact by exiting immediately. More
706 * efficient or clearer solutions might possibly be adopted.
708 if (entity->weight_counter)
712 struct bfq_weight_counter *__counter = container_of(*new,
713 struct bfq_weight_counter,
717 if (entity->weight == __counter->weight) {
718 entity->weight_counter = __counter;
721 if (entity->weight < __counter->weight)
722 new = &((*new)->rb_left);
724 new = &((*new)->rb_right);
727 entity->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
731 * In the unlucky event of an allocation failure, we just
732 * exit. This will cause the weight of entity to not be
733 * considered in bfq_differentiated_weights, which, in its
734 * turn, causes the scenario to be deemed wrongly symmetric in
735 * case entity's weight would have been the only weight making
736 * the scenario asymmetric. On the bright side, no unbalance
737 * will however occur when entity becomes inactive again (the
738 * invocation of this function is triggered by an activation
739 * of entity). In fact, bfq_weights_tree_remove does nothing
740 * if !entity->weight_counter.
742 if (unlikely(!entity->weight_counter))
745 entity->weight_counter->weight = entity->weight;
746 rb_link_node(&entity->weight_counter->weights_node, parent, new);
747 rb_insert_color(&entity->weight_counter->weights_node, root);
750 entity->weight_counter->num_active++;
754 * Decrement the weight counter associated with the entity, and, if the
755 * counter reaches 0, remove the counter from the tree.
756 * See the comments to the function bfq_weights_tree_add() for considerations
759 void __bfq_weights_tree_remove(struct bfq_data *bfqd,
760 struct bfq_entity *entity,
761 struct rb_root *root)
763 if (!entity->weight_counter)
766 entity->weight_counter->num_active--;
767 if (entity->weight_counter->num_active > 0)
768 goto reset_entity_pointer;
770 rb_erase(&entity->weight_counter->weights_node, root);
771 kfree(entity->weight_counter);
773 reset_entity_pointer:
774 entity->weight_counter = NULL;
778 * Invoke __bfq_weights_tree_remove on bfqq and all its inactive
781 void bfq_weights_tree_remove(struct bfq_data *bfqd,
782 struct bfq_queue *bfqq)
784 struct bfq_entity *entity = bfqq->entity.parent;
786 __bfq_weights_tree_remove(bfqd, &bfqq->entity,
787 &bfqd->queue_weights_tree);
789 for_each_entity(entity) {
790 struct bfq_sched_data *sd = entity->my_sched_data;
792 if (sd->next_in_service || sd->in_service_entity) {
794 * entity is still active, because either
795 * next_in_service or in_service_entity is not
796 * NULL (see the comments on the definition of
797 * next_in_service for details on why
798 * in_service_entity must be checked too).
800 * As a consequence, the weight of entity is
801 * not to be removed. In addition, if entity
802 * is active, then its parent entities are
803 * active as well, and thus their weights are
804 * not to be removed either. In the end, this
805 * loop must stop here.
809 __bfq_weights_tree_remove(bfqd, entity,
810 &bfqd->group_weights_tree);
815 * Return expired entry, or NULL to just start from scratch in rbtree.
817 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
818 struct request *last)
822 if (bfq_bfqq_fifo_expire(bfqq))
825 bfq_mark_bfqq_fifo_expire(bfqq);
827 rq = rq_entry_fifo(bfqq->fifo.next);
829 if (rq == last || ktime_get_ns() < rq->fifo_time)
832 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
836 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
837 struct bfq_queue *bfqq,
838 struct request *last)
840 struct rb_node *rbnext = rb_next(&last->rb_node);
841 struct rb_node *rbprev = rb_prev(&last->rb_node);
842 struct request *next, *prev = NULL;
844 /* Follow expired path, else get first next available. */
845 next = bfq_check_fifo(bfqq, last);
850 prev = rb_entry_rq(rbprev);
853 next = rb_entry_rq(rbnext);
855 rbnext = rb_first(&bfqq->sort_list);
856 if (rbnext && rbnext != &last->rb_node)
857 next = rb_entry_rq(rbnext);
860 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
863 /* see the definition of bfq_async_charge_factor for details */
864 static unsigned long bfq_serv_to_charge(struct request *rq,
865 struct bfq_queue *bfqq)
867 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1)
868 return blk_rq_sectors(rq);
870 return blk_rq_sectors(rq) * bfq_async_charge_factor;
874 * bfq_updated_next_req - update the queue after a new next_rq selection.
875 * @bfqd: the device data the queue belongs to.
876 * @bfqq: the queue to update.
878 * If the first request of a queue changes we make sure that the queue
879 * has enough budget to serve at least its first request (if the
880 * request has grown). We do this because if the queue has not enough
881 * budget for its first request, it has to go through two dispatch
882 * rounds to actually get it dispatched.
884 static void bfq_updated_next_req(struct bfq_data *bfqd,
885 struct bfq_queue *bfqq)
887 struct bfq_entity *entity = &bfqq->entity;
888 struct request *next_rq = bfqq->next_rq;
889 unsigned long new_budget;
894 if (bfqq == bfqd->in_service_queue)
896 * In order not to break guarantees, budgets cannot be
897 * changed after an entity has been selected.
901 new_budget = max_t(unsigned long, bfqq->max_budget,
902 bfq_serv_to_charge(next_rq, bfqq));
903 if (entity->budget != new_budget) {
904 entity->budget = new_budget;
905 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
907 bfq_requeue_bfqq(bfqd, bfqq, false);
911 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
915 if (bfqd->bfq_wr_max_time > 0)
916 return bfqd->bfq_wr_max_time;
918 dur = bfqd->rate_dur_prod;
919 do_div(dur, bfqd->peak_rate);
922 * Limit duration between 3 and 25 seconds. The upper limit
923 * has been conservatively set after the following worst case:
924 * on a QEMU/KVM virtual machine
925 * - running in a slow PC
926 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
927 * - serving a heavy I/O workload, such as the sequential reading
929 * mplayer took 23 seconds to start, if constantly weight-raised.
931 * As for higher values than that accomodating the above bad
932 * scenario, tests show that higher values would often yield
933 * the opposite of the desired result, i.e., would worsen
934 * responsiveness by allowing non-interactive applications to
935 * preserve weight raising for too long.
937 * On the other end, lower values than 3 seconds make it
938 * difficult for most interactive tasks to complete their jobs
939 * before weight-raising finishes.
941 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
944 /* switch back from soft real-time to interactive weight raising */
945 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
946 struct bfq_data *bfqd)
948 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
949 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
950 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
954 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
955 struct bfq_io_cq *bic, bool bfq_already_existing)
957 unsigned int old_wr_coeff = bfqq->wr_coeff;
958 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
960 if (bic->saved_has_short_ttime)
961 bfq_mark_bfqq_has_short_ttime(bfqq);
963 bfq_clear_bfqq_has_short_ttime(bfqq);
965 if (bic->saved_IO_bound)
966 bfq_mark_bfqq_IO_bound(bfqq);
968 bfq_clear_bfqq_IO_bound(bfqq);
970 bfqq->ttime = bic->saved_ttime;
971 bfqq->wr_coeff = bic->saved_wr_coeff;
972 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
973 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
974 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
976 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
977 time_is_before_jiffies(bfqq->last_wr_start_finish +
978 bfqq->wr_cur_max_time))) {
979 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
980 !bfq_bfqq_in_large_burst(bfqq) &&
981 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
982 bfq_wr_duration(bfqd))) {
983 switch_back_to_interactive_wr(bfqq, bfqd);
986 bfq_log_bfqq(bfqq->bfqd, bfqq,
987 "resume state: switching off wr");
991 /* make sure weight will be updated, however we got here */
992 bfqq->entity.prio_changed = 1;
997 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
998 bfqd->wr_busy_queues++;
999 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1000 bfqd->wr_busy_queues--;
1003 static int bfqq_process_refs(struct bfq_queue *bfqq)
1005 return bfqq->ref - bfqq->allocated - bfqq->entity.on_st;
1008 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1009 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1011 struct bfq_queue *item;
1012 struct hlist_node *n;
1014 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1015 hlist_del_init(&item->burst_list_node);
1016 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1017 bfqd->burst_size = 1;
1018 bfqd->burst_parent_entity = bfqq->entity.parent;
1021 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1022 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1024 /* Increment burst size to take into account also bfqq */
1027 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1028 struct bfq_queue *pos, *bfqq_item;
1029 struct hlist_node *n;
1032 * Enough queues have been activated shortly after each
1033 * other to consider this burst as large.
1035 bfqd->large_burst = true;
1038 * We can now mark all queues in the burst list as
1039 * belonging to a large burst.
1041 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1043 bfq_mark_bfqq_in_large_burst(bfqq_item);
1044 bfq_mark_bfqq_in_large_burst(bfqq);
1047 * From now on, and until the current burst finishes, any
1048 * new queue being activated shortly after the last queue
1049 * was inserted in the burst can be immediately marked as
1050 * belonging to a large burst. So the burst list is not
1051 * needed any more. Remove it.
1053 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1055 hlist_del_init(&pos->burst_list_node);
1057 * Burst not yet large: add bfqq to the burst list. Do
1058 * not increment the ref counter for bfqq, because bfqq
1059 * is removed from the burst list before freeing bfqq
1062 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1066 * If many queues belonging to the same group happen to be created
1067 * shortly after each other, then the processes associated with these
1068 * queues have typically a common goal. In particular, bursts of queue
1069 * creations are usually caused by services or applications that spawn
1070 * many parallel threads/processes. Examples are systemd during boot,
1071 * or git grep. To help these processes get their job done as soon as
1072 * possible, it is usually better to not grant either weight-raising
1073 * or device idling to their queues.
1075 * In this comment we describe, firstly, the reasons why this fact
1076 * holds, and, secondly, the next function, which implements the main
1077 * steps needed to properly mark these queues so that they can then be
1078 * treated in a different way.
1080 * The above services or applications benefit mostly from a high
1081 * throughput: the quicker the requests of the activated queues are
1082 * cumulatively served, the sooner the target job of these queues gets
1083 * completed. As a consequence, weight-raising any of these queues,
1084 * which also implies idling the device for it, is almost always
1085 * counterproductive. In most cases it just lowers throughput.
1087 * On the other hand, a burst of queue creations may be caused also by
1088 * the start of an application that does not consist of a lot of
1089 * parallel I/O-bound threads. In fact, with a complex application,
1090 * several short processes may need to be executed to start-up the
1091 * application. In this respect, to start an application as quickly as
1092 * possible, the best thing to do is in any case to privilege the I/O
1093 * related to the application with respect to all other
1094 * I/O. Therefore, the best strategy to start as quickly as possible
1095 * an application that causes a burst of queue creations is to
1096 * weight-raise all the queues created during the burst. This is the
1097 * exact opposite of the best strategy for the other type of bursts.
1099 * In the end, to take the best action for each of the two cases, the
1100 * two types of bursts need to be distinguished. Fortunately, this
1101 * seems relatively easy, by looking at the sizes of the bursts. In
1102 * particular, we found a threshold such that only bursts with a
1103 * larger size than that threshold are apparently caused by
1104 * services or commands such as systemd or git grep. For brevity,
1105 * hereafter we call just 'large' these bursts. BFQ *does not*
1106 * weight-raise queues whose creation occurs in a large burst. In
1107 * addition, for each of these queues BFQ performs or does not perform
1108 * idling depending on which choice boosts the throughput more. The
1109 * exact choice depends on the device and request pattern at
1112 * Unfortunately, false positives may occur while an interactive task
1113 * is starting (e.g., an application is being started). The
1114 * consequence is that the queues associated with the task do not
1115 * enjoy weight raising as expected. Fortunately these false positives
1116 * are very rare. They typically occur if some service happens to
1117 * start doing I/O exactly when the interactive task starts.
1119 * Turning back to the next function, it implements all the steps
1120 * needed to detect the occurrence of a large burst and to properly
1121 * mark all the queues belonging to it (so that they can then be
1122 * treated in a different way). This goal is achieved by maintaining a
1123 * "burst list" that holds, temporarily, the queues that belong to the
1124 * burst in progress. The list is then used to mark these queues as
1125 * belonging to a large burst if the burst does become large. The main
1126 * steps are the following.
1128 * . when the very first queue is created, the queue is inserted into the
1129 * list (as it could be the first queue in a possible burst)
1131 * . if the current burst has not yet become large, and a queue Q that does
1132 * not yet belong to the burst is activated shortly after the last time
1133 * at which a new queue entered the burst list, then the function appends
1134 * Q to the burst list
1136 * . if, as a consequence of the previous step, the burst size reaches
1137 * the large-burst threshold, then
1139 * . all the queues in the burst list are marked as belonging to a
1142 * . the burst list is deleted; in fact, the burst list already served
1143 * its purpose (keeping temporarily track of the queues in a burst,
1144 * so as to be able to mark them as belonging to a large burst in the
1145 * previous sub-step), and now is not needed any more
1147 * . the device enters a large-burst mode
1149 * . if a queue Q that does not belong to the burst is created while
1150 * the device is in large-burst mode and shortly after the last time
1151 * at which a queue either entered the burst list or was marked as
1152 * belonging to the current large burst, then Q is immediately marked
1153 * as belonging to a large burst.
1155 * . if a queue Q that does not belong to the burst is created a while
1156 * later, i.e., not shortly after, than the last time at which a queue
1157 * either entered the burst list or was marked as belonging to the
1158 * current large burst, then the current burst is deemed as finished and:
1160 * . the large-burst mode is reset if set
1162 * . the burst list is emptied
1164 * . Q is inserted in the burst list, as Q may be the first queue
1165 * in a possible new burst (then the burst list contains just Q
1168 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1171 * If bfqq is already in the burst list or is part of a large
1172 * burst, or finally has just been split, then there is
1173 * nothing else to do.
1175 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1176 bfq_bfqq_in_large_burst(bfqq) ||
1177 time_is_after_eq_jiffies(bfqq->split_time +
1178 msecs_to_jiffies(10)))
1182 * If bfqq's creation happens late enough, or bfqq belongs to
1183 * a different group than the burst group, then the current
1184 * burst is finished, and related data structures must be
1187 * In this respect, consider the special case where bfqq is
1188 * the very first queue created after BFQ is selected for this
1189 * device. In this case, last_ins_in_burst and
1190 * burst_parent_entity are not yet significant when we get
1191 * here. But it is easy to verify that, whether or not the
1192 * following condition is true, bfqq will end up being
1193 * inserted into the burst list. In particular the list will
1194 * happen to contain only bfqq. And this is exactly what has
1195 * to happen, as bfqq may be the first queue of the first
1198 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1199 bfqd->bfq_burst_interval) ||
1200 bfqq->entity.parent != bfqd->burst_parent_entity) {
1201 bfqd->large_burst = false;
1202 bfq_reset_burst_list(bfqd, bfqq);
1207 * If we get here, then bfqq is being activated shortly after the
1208 * last queue. So, if the current burst is also large, we can mark
1209 * bfqq as belonging to this large burst immediately.
1211 if (bfqd->large_burst) {
1212 bfq_mark_bfqq_in_large_burst(bfqq);
1217 * If we get here, then a large-burst state has not yet been
1218 * reached, but bfqq is being activated shortly after the last
1219 * queue. Then we add bfqq to the burst.
1221 bfq_add_to_burst(bfqd, bfqq);
1224 * At this point, bfqq either has been added to the current
1225 * burst or has caused the current burst to terminate and a
1226 * possible new burst to start. In particular, in the second
1227 * case, bfqq has become the first queue in the possible new
1228 * burst. In both cases last_ins_in_burst needs to be moved
1231 bfqd->last_ins_in_burst = jiffies;
1234 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1236 struct bfq_entity *entity = &bfqq->entity;
1238 return entity->budget - entity->service;
1242 * If enough samples have been computed, return the current max budget
1243 * stored in bfqd, which is dynamically updated according to the
1244 * estimated disk peak rate; otherwise return the default max budget
1246 static int bfq_max_budget(struct bfq_data *bfqd)
1248 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1249 return bfq_default_max_budget;
1251 return bfqd->bfq_max_budget;
1255 * Return min budget, which is a fraction of the current or default
1256 * max budget (trying with 1/32)
1258 static int bfq_min_budget(struct bfq_data *bfqd)
1260 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1261 return bfq_default_max_budget / 32;
1263 return bfqd->bfq_max_budget / 32;
1267 * The next function, invoked after the input queue bfqq switches from
1268 * idle to busy, updates the budget of bfqq. The function also tells
1269 * whether the in-service queue should be expired, by returning
1270 * true. The purpose of expiring the in-service queue is to give bfqq
1271 * the chance to possibly preempt the in-service queue, and the reason
1272 * for preempting the in-service queue is to achieve one of the two
1275 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1276 * expired because it has remained idle. In particular, bfqq may have
1277 * expired for one of the following two reasons:
1279 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1280 * and did not make it to issue a new request before its last
1281 * request was served;
1283 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1284 * a new request before the expiration of the idling-time.
1286 * Even if bfqq has expired for one of the above reasons, the process
1287 * associated with the queue may be however issuing requests greedily,
1288 * and thus be sensitive to the bandwidth it receives (bfqq may have
1289 * remained idle for other reasons: CPU high load, bfqq not enjoying
1290 * idling, I/O throttling somewhere in the path from the process to
1291 * the I/O scheduler, ...). But if, after every expiration for one of
1292 * the above two reasons, bfqq has to wait for the service of at least
1293 * one full budget of another queue before being served again, then
1294 * bfqq is likely to get a much lower bandwidth or resource time than
1295 * its reserved ones. To address this issue, two countermeasures need
1298 * First, the budget and the timestamps of bfqq need to be updated in
1299 * a special way on bfqq reactivation: they need to be updated as if
1300 * bfqq did not remain idle and did not expire. In fact, if they are
1301 * computed as if bfqq expired and remained idle until reactivation,
1302 * then the process associated with bfqq is treated as if, instead of
1303 * being greedy, it stopped issuing requests when bfqq remained idle,
1304 * and restarts issuing requests only on this reactivation. In other
1305 * words, the scheduler does not help the process recover the "service
1306 * hole" between bfqq expiration and reactivation. As a consequence,
1307 * the process receives a lower bandwidth than its reserved one. In
1308 * contrast, to recover this hole, the budget must be updated as if
1309 * bfqq was not expired at all before this reactivation, i.e., it must
1310 * be set to the value of the remaining budget when bfqq was
1311 * expired. Along the same line, timestamps need to be assigned the
1312 * value they had the last time bfqq was selected for service, i.e.,
1313 * before last expiration. Thus timestamps need to be back-shifted
1314 * with respect to their normal computation (see [1] for more details
1315 * on this tricky aspect).
1317 * Secondly, to allow the process to recover the hole, the in-service
1318 * queue must be expired too, to give bfqq the chance to preempt it
1319 * immediately. In fact, if bfqq has to wait for a full budget of the
1320 * in-service queue to be completed, then it may become impossible to
1321 * let the process recover the hole, even if the back-shifted
1322 * timestamps of bfqq are lower than those of the in-service queue. If
1323 * this happens for most or all of the holes, then the process may not
1324 * receive its reserved bandwidth. In this respect, it is worth noting
1325 * that, being the service of outstanding requests unpreemptible, a
1326 * little fraction of the holes may however be unrecoverable, thereby
1327 * causing a little loss of bandwidth.
1329 * The last important point is detecting whether bfqq does need this
1330 * bandwidth recovery. In this respect, the next function deems the
1331 * process associated with bfqq greedy, and thus allows it to recover
1332 * the hole, if: 1) the process is waiting for the arrival of a new
1333 * request (which implies that bfqq expired for one of the above two
1334 * reasons), and 2) such a request has arrived soon. The first
1335 * condition is controlled through the flag non_blocking_wait_rq,
1336 * while the second through the flag arrived_in_time. If both
1337 * conditions hold, then the function computes the budget in the
1338 * above-described special way, and signals that the in-service queue
1339 * should be expired. Timestamp back-shifting is done later in
1340 * __bfq_activate_entity.
1342 * 2. Reduce latency. Even if timestamps are not backshifted to let
1343 * the process associated with bfqq recover a service hole, bfqq may
1344 * however happen to have, after being (re)activated, a lower finish
1345 * timestamp than the in-service queue. That is, the next budget of
1346 * bfqq may have to be completed before the one of the in-service
1347 * queue. If this is the case, then preempting the in-service queue
1348 * allows this goal to be achieved, apart from the unpreemptible,
1349 * outstanding requests mentioned above.
1351 * Unfortunately, regardless of which of the above two goals one wants
1352 * to achieve, service trees need first to be updated to know whether
1353 * the in-service queue must be preempted. To have service trees
1354 * correctly updated, the in-service queue must be expired and
1355 * rescheduled, and bfqq must be scheduled too. This is one of the
1356 * most costly operations (in future versions, the scheduling
1357 * mechanism may be re-designed in such a way to make it possible to
1358 * know whether preemption is needed without needing to update service
1359 * trees). In addition, queue preemptions almost always cause random
1360 * I/O, and thus loss of throughput. Because of these facts, the next
1361 * function adopts the following simple scheme to avoid both costly
1362 * operations and too frequent preemptions: it requests the expiration
1363 * of the in-service queue (unconditionally) only for queues that need
1364 * to recover a hole, or that either are weight-raised or deserve to
1367 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1368 struct bfq_queue *bfqq,
1369 bool arrived_in_time,
1370 bool wr_or_deserves_wr)
1372 struct bfq_entity *entity = &bfqq->entity;
1374 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time) {
1376 * We do not clear the flag non_blocking_wait_rq here, as
1377 * the latter is used in bfq_activate_bfqq to signal
1378 * that timestamps need to be back-shifted (and is
1379 * cleared right after).
1383 * In next assignment we rely on that either
1384 * entity->service or entity->budget are not updated
1385 * on expiration if bfqq is empty (see
1386 * __bfq_bfqq_recalc_budget). Thus both quantities
1387 * remain unchanged after such an expiration, and the
1388 * following statement therefore assigns to
1389 * entity->budget the remaining budget on such an
1392 entity->budget = min_t(unsigned long,
1393 bfq_bfqq_budget_left(bfqq),
1397 * At this point, we have used entity->service to get
1398 * the budget left (needed for updating
1399 * entity->budget). Thus we finally can, and have to,
1400 * reset entity->service. The latter must be reset
1401 * because bfqq would otherwise be charged again for
1402 * the service it has received during its previous
1405 entity->service = 0;
1411 * We can finally complete expiration, by setting service to 0.
1413 entity->service = 0;
1414 entity->budget = max_t(unsigned long, bfqq->max_budget,
1415 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1416 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1417 return wr_or_deserves_wr;
1421 * Return the farthest past time instant according to jiffies
1424 static unsigned long bfq_smallest_from_now(void)
1426 return jiffies - MAX_JIFFY_OFFSET;
1429 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1430 struct bfq_queue *bfqq,
1431 unsigned int old_wr_coeff,
1432 bool wr_or_deserves_wr,
1437 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1438 /* start a weight-raising period */
1440 bfqq->service_from_wr = 0;
1441 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1442 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1445 * No interactive weight raising in progress
1446 * here: assign minus infinity to
1447 * wr_start_at_switch_to_srt, to make sure
1448 * that, at the end of the soft-real-time
1449 * weight raising periods that is starting
1450 * now, no interactive weight-raising period
1451 * may be wrongly considered as still in
1452 * progress (and thus actually started by
1455 bfqq->wr_start_at_switch_to_srt =
1456 bfq_smallest_from_now();
1457 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1458 BFQ_SOFTRT_WEIGHT_FACTOR;
1459 bfqq->wr_cur_max_time =
1460 bfqd->bfq_wr_rt_max_time;
1464 * If needed, further reduce budget to make sure it is
1465 * close to bfqq's backlog, so as to reduce the
1466 * scheduling-error component due to a too large
1467 * budget. Do not care about throughput consequences,
1468 * but only about latency. Finally, do not assign a
1469 * too small budget either, to avoid increasing
1470 * latency by causing too frequent expirations.
1472 bfqq->entity.budget = min_t(unsigned long,
1473 bfqq->entity.budget,
1474 2 * bfq_min_budget(bfqd));
1475 } else if (old_wr_coeff > 1) {
1476 if (interactive) { /* update wr coeff and duration */
1477 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1478 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1479 } else if (in_burst)
1483 * The application is now or still meeting the
1484 * requirements for being deemed soft rt. We
1485 * can then correctly and safely (re)charge
1486 * the weight-raising duration for the
1487 * application with the weight-raising
1488 * duration for soft rt applications.
1490 * In particular, doing this recharge now, i.e.,
1491 * before the weight-raising period for the
1492 * application finishes, reduces the probability
1493 * of the following negative scenario:
1494 * 1) the weight of a soft rt application is
1495 * raised at startup (as for any newly
1496 * created application),
1497 * 2) since the application is not interactive,
1498 * at a certain time weight-raising is
1499 * stopped for the application,
1500 * 3) at that time the application happens to
1501 * still have pending requests, and hence
1502 * is destined to not have a chance to be
1503 * deemed soft rt before these requests are
1504 * completed (see the comments to the
1505 * function bfq_bfqq_softrt_next_start()
1506 * for details on soft rt detection),
1507 * 4) these pending requests experience a high
1508 * latency because the application is not
1509 * weight-raised while they are pending.
1511 if (bfqq->wr_cur_max_time !=
1512 bfqd->bfq_wr_rt_max_time) {
1513 bfqq->wr_start_at_switch_to_srt =
1514 bfqq->last_wr_start_finish;
1516 bfqq->wr_cur_max_time =
1517 bfqd->bfq_wr_rt_max_time;
1518 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1519 BFQ_SOFTRT_WEIGHT_FACTOR;
1521 bfqq->last_wr_start_finish = jiffies;
1526 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1527 struct bfq_queue *bfqq)
1529 return bfqq->dispatched == 0 &&
1530 time_is_before_jiffies(
1531 bfqq->budget_timeout +
1532 bfqd->bfq_wr_min_idle_time);
1535 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1536 struct bfq_queue *bfqq,
1541 bool soft_rt, in_burst, wr_or_deserves_wr,
1542 bfqq_wants_to_preempt,
1543 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1545 * See the comments on
1546 * bfq_bfqq_update_budg_for_activation for
1547 * details on the usage of the next variable.
1549 arrived_in_time = ktime_get_ns() <=
1550 bfqq->ttime.last_end_request +
1551 bfqd->bfq_slice_idle * 3;
1555 * bfqq deserves to be weight-raised if:
1557 * - it does not belong to a large burst,
1558 * - it has been idle for enough time or is soft real-time,
1559 * - is linked to a bfq_io_cq (it is not shared in any sense).
1561 in_burst = bfq_bfqq_in_large_burst(bfqq);
1562 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1564 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1565 bfqq->dispatched == 0;
1566 *interactive = !in_burst && idle_for_long_time;
1567 wr_or_deserves_wr = bfqd->low_latency &&
1568 (bfqq->wr_coeff > 1 ||
1569 (bfq_bfqq_sync(bfqq) &&
1570 bfqq->bic && (*interactive || soft_rt)));
1573 * Using the last flag, update budget and check whether bfqq
1574 * may want to preempt the in-service queue.
1576 bfqq_wants_to_preempt =
1577 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1582 * If bfqq happened to be activated in a burst, but has been
1583 * idle for much more than an interactive queue, then we
1584 * assume that, in the overall I/O initiated in the burst, the
1585 * I/O associated with bfqq is finished. So bfqq does not need
1586 * to be treated as a queue belonging to a burst
1587 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1588 * if set, and remove bfqq from the burst list if it's
1589 * there. We do not decrement burst_size, because the fact
1590 * that bfqq does not need to belong to the burst list any
1591 * more does not invalidate the fact that bfqq was created in
1594 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1595 idle_for_long_time &&
1596 time_is_before_jiffies(
1597 bfqq->budget_timeout +
1598 msecs_to_jiffies(10000))) {
1599 hlist_del_init(&bfqq->burst_list_node);
1600 bfq_clear_bfqq_in_large_burst(bfqq);
1603 bfq_clear_bfqq_just_created(bfqq);
1606 if (!bfq_bfqq_IO_bound(bfqq)) {
1607 if (arrived_in_time) {
1608 bfqq->requests_within_timer++;
1609 if (bfqq->requests_within_timer >=
1610 bfqd->bfq_requests_within_timer)
1611 bfq_mark_bfqq_IO_bound(bfqq);
1613 bfqq->requests_within_timer = 0;
1616 if (bfqd->low_latency) {
1617 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1620 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1622 if (time_is_before_jiffies(bfqq->split_time +
1623 bfqd->bfq_wr_min_idle_time)) {
1624 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1631 if (old_wr_coeff != bfqq->wr_coeff)
1632 bfqq->entity.prio_changed = 1;
1636 bfqq->last_idle_bklogged = jiffies;
1637 bfqq->service_from_backlogged = 0;
1638 bfq_clear_bfqq_softrt_update(bfqq);
1640 bfq_add_bfqq_busy(bfqd, bfqq);
1643 * Expire in-service queue only if preemption may be needed
1644 * for guarantees. In this respect, the function
1645 * next_queue_may_preempt just checks a simple, necessary
1646 * condition, and not a sufficient condition based on
1647 * timestamps. In fact, for the latter condition to be
1648 * evaluated, timestamps would need first to be updated, and
1649 * this operation is quite costly (see the comments on the
1650 * function bfq_bfqq_update_budg_for_activation).
1652 if (bfqd->in_service_queue && bfqq_wants_to_preempt &&
1653 bfqd->in_service_queue->wr_coeff < bfqq->wr_coeff &&
1654 next_queue_may_preempt(bfqd))
1655 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1656 false, BFQQE_PREEMPTED);
1659 static void bfq_add_request(struct request *rq)
1661 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1662 struct bfq_data *bfqd = bfqq->bfqd;
1663 struct request *next_rq, *prev;
1664 unsigned int old_wr_coeff = bfqq->wr_coeff;
1665 bool interactive = false;
1667 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1668 bfqq->queued[rq_is_sync(rq)]++;
1671 elv_rb_add(&bfqq->sort_list, rq);
1674 * Check if this request is a better next-serve candidate.
1676 prev = bfqq->next_rq;
1677 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
1678 bfqq->next_rq = next_rq;
1681 * Adjust priority tree position, if next_rq changes.
1683 if (prev != bfqq->next_rq)
1684 bfq_pos_tree_add_move(bfqd, bfqq);
1686 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
1687 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
1690 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
1691 time_is_before_jiffies(
1692 bfqq->last_wr_start_finish +
1693 bfqd->bfq_wr_min_inter_arr_async)) {
1694 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1695 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1697 bfqd->wr_busy_queues++;
1698 bfqq->entity.prio_changed = 1;
1700 if (prev != bfqq->next_rq)
1701 bfq_updated_next_req(bfqd, bfqq);
1705 * Assign jiffies to last_wr_start_finish in the following
1708 * . if bfqq is not going to be weight-raised, because, for
1709 * non weight-raised queues, last_wr_start_finish stores the
1710 * arrival time of the last request; as of now, this piece
1711 * of information is used only for deciding whether to
1712 * weight-raise async queues
1714 * . if bfqq is not weight-raised, because, if bfqq is now
1715 * switching to weight-raised, then last_wr_start_finish
1716 * stores the time when weight-raising starts
1718 * . if bfqq is interactive, because, regardless of whether
1719 * bfqq is currently weight-raised, the weight-raising
1720 * period must start or restart (this case is considered
1721 * separately because it is not detected by the above
1722 * conditions, if bfqq is already weight-raised)
1724 * last_wr_start_finish has to be updated also if bfqq is soft
1725 * real-time, because the weight-raising period is constantly
1726 * restarted on idle-to-busy transitions for these queues, but
1727 * this is already done in bfq_bfqq_handle_idle_busy_switch if
1730 if (bfqd->low_latency &&
1731 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
1732 bfqq->last_wr_start_finish = jiffies;
1735 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
1737 struct request_queue *q)
1739 struct bfq_queue *bfqq = bfqd->bio_bfqq;
1743 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
1748 static sector_t get_sdist(sector_t last_pos, struct request *rq)
1751 return abs(blk_rq_pos(rq) - last_pos);
1756 #if 0 /* Still not clear if we can do without next two functions */
1757 static void bfq_activate_request(struct request_queue *q, struct request *rq)
1759 struct bfq_data *bfqd = q->elevator->elevator_data;
1761 bfqd->rq_in_driver++;
1764 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
1766 struct bfq_data *bfqd = q->elevator->elevator_data;
1768 bfqd->rq_in_driver--;
1772 static void bfq_remove_request(struct request_queue *q,
1775 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1776 struct bfq_data *bfqd = bfqq->bfqd;
1777 const int sync = rq_is_sync(rq);
1779 if (bfqq->next_rq == rq) {
1780 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
1781 bfq_updated_next_req(bfqd, bfqq);
1784 if (rq->queuelist.prev != &rq->queuelist)
1785 list_del_init(&rq->queuelist);
1786 bfqq->queued[sync]--;
1788 elv_rb_del(&bfqq->sort_list, rq);
1790 elv_rqhash_del(q, rq);
1791 if (q->last_merge == rq)
1792 q->last_merge = NULL;
1794 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
1795 bfqq->next_rq = NULL;
1797 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
1798 bfq_del_bfqq_busy(bfqd, bfqq, false);
1800 * bfqq emptied. In normal operation, when
1801 * bfqq is empty, bfqq->entity.service and
1802 * bfqq->entity.budget must contain,
1803 * respectively, the service received and the
1804 * budget used last time bfqq emptied. These
1805 * facts do not hold in this case, as at least
1806 * this last removal occurred while bfqq is
1807 * not in service. To avoid inconsistencies,
1808 * reset both bfqq->entity.service and
1809 * bfqq->entity.budget, if bfqq has still a
1810 * process that may issue I/O requests to it.
1812 bfqq->entity.budget = bfqq->entity.service = 0;
1816 * Remove queue from request-position tree as it is empty.
1818 if (bfqq->pos_root) {
1819 rb_erase(&bfqq->pos_node, bfqq->pos_root);
1820 bfqq->pos_root = NULL;
1823 bfq_pos_tree_add_move(bfqd, bfqq);
1826 if (rq->cmd_flags & REQ_META)
1827 bfqq->meta_pending--;
1831 static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio)
1833 struct request_queue *q = hctx->queue;
1834 struct bfq_data *bfqd = q->elevator->elevator_data;
1835 struct request *free = NULL;
1837 * bfq_bic_lookup grabs the queue_lock: invoke it now and
1838 * store its return value for later use, to avoid nesting
1839 * queue_lock inside the bfqd->lock. We assume that the bic
1840 * returned by bfq_bic_lookup does not go away before
1841 * bfqd->lock is taken.
1843 struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
1846 spin_lock_irq(&bfqd->lock);
1849 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
1851 bfqd->bio_bfqq = NULL;
1852 bfqd->bio_bic = bic;
1854 ret = blk_mq_sched_try_merge(q, bio, &free);
1857 blk_mq_free_request(free);
1858 spin_unlock_irq(&bfqd->lock);
1863 static int bfq_request_merge(struct request_queue *q, struct request **req,
1866 struct bfq_data *bfqd = q->elevator->elevator_data;
1867 struct request *__rq;
1869 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
1870 if (__rq && elv_bio_merge_ok(__rq, bio)) {
1872 return ELEVATOR_FRONT_MERGE;
1875 return ELEVATOR_NO_MERGE;
1878 static struct bfq_queue *bfq_init_rq(struct request *rq);
1880 static void bfq_request_merged(struct request_queue *q, struct request *req,
1881 enum elv_merge type)
1883 if (type == ELEVATOR_FRONT_MERGE &&
1884 rb_prev(&req->rb_node) &&
1886 blk_rq_pos(container_of(rb_prev(&req->rb_node),
1887 struct request, rb_node))) {
1888 struct bfq_queue *bfqq = bfq_init_rq(req);
1889 struct bfq_data *bfqd = bfqq->bfqd;
1890 struct request *prev, *next_rq;
1892 /* Reposition request in its sort_list */
1893 elv_rb_del(&bfqq->sort_list, req);
1894 elv_rb_add(&bfqq->sort_list, req);
1896 /* Choose next request to be served for bfqq */
1897 prev = bfqq->next_rq;
1898 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
1899 bfqd->last_position);
1900 bfqq->next_rq = next_rq;
1902 * If next_rq changes, update both the queue's budget to
1903 * fit the new request and the queue's position in its
1906 if (prev != bfqq->next_rq) {
1907 bfq_updated_next_req(bfqd, bfqq);
1908 bfq_pos_tree_add_move(bfqd, bfqq);
1914 * This function is called to notify the scheduler that the requests
1915 * rq and 'next' have been merged, with 'next' going away. BFQ
1916 * exploits this hook to address the following issue: if 'next' has a
1917 * fifo_time lower that rq, then the fifo_time of rq must be set to
1918 * the value of 'next', to not forget the greater age of 'next'.
1920 * NOTE: in this function we assume that rq is in a bfq_queue, basing
1921 * on that rq is picked from the hash table q->elevator->hash, which,
1922 * in its turn, is filled only with I/O requests present in
1923 * bfq_queues, while BFQ is in use for the request queue q. In fact,
1924 * the function that fills this hash table (elv_rqhash_add) is called
1925 * only by bfq_insert_request.
1927 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
1928 struct request *next)
1930 struct bfq_queue *bfqq = bfq_init_rq(rq),
1931 *next_bfqq = bfq_init_rq(next);
1934 * If next and rq belong to the same bfq_queue and next is older
1935 * than rq, then reposition rq in the fifo (by substituting next
1936 * with rq). Otherwise, if next and rq belong to different
1937 * bfq_queues, never reposition rq: in fact, we would have to
1938 * reposition it with respect to next's position in its own fifo,
1939 * which would most certainly be too expensive with respect to
1942 if (bfqq == next_bfqq &&
1943 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1944 next->fifo_time < rq->fifo_time) {
1945 list_del_init(&rq->queuelist);
1946 list_replace_init(&next->queuelist, &rq->queuelist);
1947 rq->fifo_time = next->fifo_time;
1950 if (bfqq->next_rq == next)
1953 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
1956 /* Must be called with bfqq != NULL */
1957 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
1959 if (bfq_bfqq_busy(bfqq))
1960 bfqq->bfqd->wr_busy_queues--;
1962 bfqq->wr_cur_max_time = 0;
1963 bfqq->last_wr_start_finish = jiffies;
1965 * Trigger a weight change on the next invocation of
1966 * __bfq_entity_update_weight_prio.
1968 bfqq->entity.prio_changed = 1;
1971 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
1972 struct bfq_group *bfqg)
1976 for (i = 0; i < 2; i++)
1977 for (j = 0; j < IOPRIO_BE_NR; j++)
1978 if (bfqg->async_bfqq[i][j])
1979 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
1980 if (bfqg->async_idle_bfqq)
1981 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
1984 static void bfq_end_wr(struct bfq_data *bfqd)
1986 struct bfq_queue *bfqq;
1988 spin_lock_irq(&bfqd->lock);
1990 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
1991 bfq_bfqq_end_wr(bfqq);
1992 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
1993 bfq_bfqq_end_wr(bfqq);
1994 bfq_end_wr_async(bfqd);
1996 spin_unlock_irq(&bfqd->lock);
1999 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2002 return blk_rq_pos(io_struct);
2004 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2007 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2010 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2014 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2015 struct bfq_queue *bfqq,
2018 struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
2019 struct rb_node *parent, *node;
2020 struct bfq_queue *__bfqq;
2022 if (RB_EMPTY_ROOT(root))
2026 * First, if we find a request starting at the end of the last
2027 * request, choose it.
2029 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2034 * If the exact sector wasn't found, the parent of the NULL leaf
2035 * will contain the closest sector (rq_pos_tree sorted by
2036 * next_request position).
2038 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2039 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2042 if (blk_rq_pos(__bfqq->next_rq) < sector)
2043 node = rb_next(&__bfqq->pos_node);
2045 node = rb_prev(&__bfqq->pos_node);
2049 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2050 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2056 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2057 struct bfq_queue *cur_bfqq,
2060 struct bfq_queue *bfqq;
2063 * We shall notice if some of the queues are cooperating,
2064 * e.g., working closely on the same area of the device. In
2065 * that case, we can group them together and: 1) don't waste
2066 * time idling, and 2) serve the union of their requests in
2067 * the best possible order for throughput.
2069 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2070 if (!bfqq || bfqq == cur_bfqq)
2076 static struct bfq_queue *
2077 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2079 int process_refs, new_process_refs;
2080 struct bfq_queue *__bfqq;
2083 * If there are no process references on the new_bfqq, then it is
2084 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2085 * may have dropped their last reference (not just their last process
2088 if (!bfqq_process_refs(new_bfqq))
2091 /* Avoid a circular list and skip interim queue merges. */
2092 while ((__bfqq = new_bfqq->new_bfqq)) {
2098 process_refs = bfqq_process_refs(bfqq);
2099 new_process_refs = bfqq_process_refs(new_bfqq);
2101 * If the process for the bfqq has gone away, there is no
2102 * sense in merging the queues.
2104 if (process_refs == 0 || new_process_refs == 0)
2107 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2111 * Merging is just a redirection: the requests of the process
2112 * owning one of the two queues are redirected to the other queue.
2113 * The latter queue, in its turn, is set as shared if this is the
2114 * first time that the requests of some process are redirected to
2117 * We redirect bfqq to new_bfqq and not the opposite, because
2118 * we are in the context of the process owning bfqq, thus we
2119 * have the io_cq of this process. So we can immediately
2120 * configure this io_cq to redirect the requests of the
2121 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2122 * not available any more (new_bfqq->bic == NULL).
2124 * Anyway, even in case new_bfqq coincides with the in-service
2125 * queue, redirecting requests the in-service queue is the
2126 * best option, as we feed the in-service queue with new
2127 * requests close to the last request served and, by doing so,
2128 * are likely to increase the throughput.
2130 bfqq->new_bfqq = new_bfqq;
2131 new_bfqq->ref += process_refs;
2135 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2136 struct bfq_queue *new_bfqq)
2138 if (bfq_too_late_for_merging(new_bfqq))
2141 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2142 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2146 * If either of the queues has already been detected as seeky,
2147 * then merging it with the other queue is unlikely to lead to
2150 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2154 * Interleaved I/O is known to be done by (some) applications
2155 * only for reads, so it does not make sense to merge async
2158 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2165 * Attempt to schedule a merge of bfqq with the currently in-service
2166 * queue or with a close queue among the scheduled queues. Return
2167 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2168 * structure otherwise.
2170 * The OOM queue is not allowed to participate to cooperation: in fact, since
2171 * the requests temporarily redirected to the OOM queue could be redirected
2172 * again to dedicated queues at any time, the state needed to correctly
2173 * handle merging with the OOM queue would be quite complex and expensive
2174 * to maintain. Besides, in such a critical condition as an out of memory,
2175 * the benefits of queue merging may be little relevant, or even negligible.
2177 * WARNING: queue merging may impair fairness among non-weight raised
2178 * queues, for at least two reasons: 1) the original weight of a
2179 * merged queue may change during the merged state, 2) even being the
2180 * weight the same, a merged queue may be bloated with many more
2181 * requests than the ones produced by its originally-associated
2184 static struct bfq_queue *
2185 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2186 void *io_struct, bool request)
2188 struct bfq_queue *in_service_bfqq, *new_bfqq;
2191 * Prevent bfqq from being merged if it has been created too
2192 * long ago. The idea is that true cooperating processes, and
2193 * thus their associated bfq_queues, are supposed to be
2194 * created shortly after each other. This is the case, e.g.,
2195 * for KVM/QEMU and dump I/O threads. Basing on this
2196 * assumption, the following filtering greatly reduces the
2197 * probability that two non-cooperating processes, which just
2198 * happen to do close I/O for some short time interval, have
2199 * their queues merged by mistake.
2201 if (bfq_too_late_for_merging(bfqq))
2205 return bfqq->new_bfqq;
2207 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2210 /* If there is only one backlogged queue, don't search. */
2211 if (bfqd->busy_queues == 1)
2214 in_service_bfqq = bfqd->in_service_queue;
2216 if (in_service_bfqq && in_service_bfqq != bfqq &&
2217 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2218 bfq_rq_close_to_sector(io_struct, request,
2219 bfqd->in_serv_last_pos) &&
2220 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2221 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2222 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2227 * Check whether there is a cooperator among currently scheduled
2228 * queues. The only thing we need is that the bio/request is not
2229 * NULL, as we need it to establish whether a cooperator exists.
2231 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2232 bfq_io_struct_pos(io_struct, request));
2234 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2235 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2236 return bfq_setup_merge(bfqq, new_bfqq);
2241 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2243 struct bfq_io_cq *bic = bfqq->bic;
2246 * If !bfqq->bic, the queue is already shared or its requests
2247 * have already been redirected to a shared queue; both idle window
2248 * and weight raising state have already been saved. Do nothing.
2253 bic->saved_ttime = bfqq->ttime;
2254 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2255 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2256 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2257 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2258 if (unlikely(bfq_bfqq_just_created(bfqq) &&
2259 !bfq_bfqq_in_large_burst(bfqq) &&
2260 bfqq->bfqd->low_latency)) {
2262 * bfqq being merged right after being created: bfqq
2263 * would have deserved interactive weight raising, but
2264 * did not make it to be set in a weight-raised state,
2265 * because of this early merge. Store directly the
2266 * weight-raising state that would have been assigned
2267 * to bfqq, so that to avoid that bfqq unjustly fails
2268 * to enjoy weight raising if split soon.
2270 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
2271 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2272 bic->saved_last_wr_start_finish = jiffies;
2274 bic->saved_wr_coeff = bfqq->wr_coeff;
2275 bic->saved_wr_start_at_switch_to_srt =
2276 bfqq->wr_start_at_switch_to_srt;
2277 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2278 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2283 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2284 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2286 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2287 (unsigned long)new_bfqq->pid);
2288 /* Save weight raising and idle window of the merged queues */
2289 bfq_bfqq_save_state(bfqq);
2290 bfq_bfqq_save_state(new_bfqq);
2291 if (bfq_bfqq_IO_bound(bfqq))
2292 bfq_mark_bfqq_IO_bound(new_bfqq);
2293 bfq_clear_bfqq_IO_bound(bfqq);
2296 * If bfqq is weight-raised, then let new_bfqq inherit
2297 * weight-raising. To reduce false positives, neglect the case
2298 * where bfqq has just been created, but has not yet made it
2299 * to be weight-raised (which may happen because EQM may merge
2300 * bfqq even before bfq_add_request is executed for the first
2301 * time for bfqq). Handling this case would however be very
2302 * easy, thanks to the flag just_created.
2304 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2305 new_bfqq->wr_coeff = bfqq->wr_coeff;
2306 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2307 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2308 new_bfqq->wr_start_at_switch_to_srt =
2309 bfqq->wr_start_at_switch_to_srt;
2310 if (bfq_bfqq_busy(new_bfqq))
2311 bfqd->wr_busy_queues++;
2312 new_bfqq->entity.prio_changed = 1;
2315 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2317 bfqq->entity.prio_changed = 1;
2318 if (bfq_bfqq_busy(bfqq))
2319 bfqd->wr_busy_queues--;
2322 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2323 bfqd->wr_busy_queues);
2326 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2328 bic_set_bfqq(bic, new_bfqq, 1);
2329 bfq_mark_bfqq_coop(new_bfqq);
2331 * new_bfqq now belongs to at least two bics (it is a shared queue):
2332 * set new_bfqq->bic to NULL. bfqq either:
2333 * - does not belong to any bic any more, and hence bfqq->bic must
2334 * be set to NULL, or
2335 * - is a queue whose owning bics have already been redirected to a
2336 * different queue, hence the queue is destined to not belong to
2337 * any bic soon and bfqq->bic is already NULL (therefore the next
2338 * assignment causes no harm).
2340 new_bfqq->bic = NULL;
2342 /* release process reference to bfqq */
2343 bfq_put_queue(bfqq);
2346 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2349 struct bfq_data *bfqd = q->elevator->elevator_data;
2350 bool is_sync = op_is_sync(bio->bi_opf);
2351 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2354 * Disallow merge of a sync bio into an async request.
2356 if (is_sync && !rq_is_sync(rq))
2360 * Lookup the bfqq that this bio will be queued with. Allow
2361 * merge only if rq is queued there.
2367 * We take advantage of this function to perform an early merge
2368 * of the queues of possible cooperating processes.
2370 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2373 * bic still points to bfqq, then it has not yet been
2374 * redirected to some other bfq_queue, and a queue
2375 * merge beween bfqq and new_bfqq can be safely
2376 * fulfillled, i.e., bic can be redirected to new_bfqq
2377 * and bfqq can be put.
2379 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2382 * If we get here, bio will be queued into new_queue,
2383 * so use new_bfqq to decide whether bio and rq can be
2389 * Change also bqfd->bio_bfqq, as
2390 * bfqd->bio_bic now points to new_bfqq, and
2391 * this function may be invoked again (and then may
2392 * use again bqfd->bio_bfqq).
2394 bfqd->bio_bfqq = bfqq;
2397 return bfqq == RQ_BFQQ(rq);
2401 * Set the maximum time for the in-service queue to consume its
2402 * budget. This prevents seeky processes from lowering the throughput.
2403 * In practice, a time-slice service scheme is used with seeky
2406 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2407 struct bfq_queue *bfqq)
2409 unsigned int timeout_coeff;
2411 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2414 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2416 bfqd->last_budget_start = ktime_get();
2418 bfqq->budget_timeout = jiffies +
2419 bfqd->bfq_timeout * timeout_coeff;
2422 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2423 struct bfq_queue *bfqq)
2426 bfq_clear_bfqq_fifo_expire(bfqq);
2428 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2430 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2431 bfqq->wr_coeff > 1 &&
2432 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2433 time_is_before_jiffies(bfqq->budget_timeout)) {
2435 * For soft real-time queues, move the start
2436 * of the weight-raising period forward by the
2437 * time the queue has not received any
2438 * service. Otherwise, a relatively long
2439 * service delay is likely to cause the
2440 * weight-raising period of the queue to end,
2441 * because of the short duration of the
2442 * weight-raising period of a soft real-time
2443 * queue. It is worth noting that this move
2444 * is not so dangerous for the other queues,
2445 * because soft real-time queues are not
2448 * To not add a further variable, we use the
2449 * overloaded field budget_timeout to
2450 * determine for how long the queue has not
2451 * received service, i.e., how much time has
2452 * elapsed since the queue expired. However,
2453 * this is a little imprecise, because
2454 * budget_timeout is set to jiffies if bfqq
2455 * not only expires, but also remains with no
2458 if (time_after(bfqq->budget_timeout,
2459 bfqq->last_wr_start_finish))
2460 bfqq->last_wr_start_finish +=
2461 jiffies - bfqq->budget_timeout;
2463 bfqq->last_wr_start_finish = jiffies;
2466 bfq_set_budget_timeout(bfqd, bfqq);
2467 bfq_log_bfqq(bfqd, bfqq,
2468 "set_in_service_queue, cur-budget = %d",
2469 bfqq->entity.budget);
2472 bfqd->in_service_queue = bfqq;
2476 * Get and set a new queue for service.
2478 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2480 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2482 __bfq_set_in_service_queue(bfqd, bfqq);
2486 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2488 struct bfq_queue *bfqq = bfqd->in_service_queue;
2491 bfq_mark_bfqq_wait_request(bfqq);
2494 * We don't want to idle for seeks, but we do want to allow
2495 * fair distribution of slice time for a process doing back-to-back
2496 * seeks. So allow a little bit of time for him to submit a new rq.
2498 sl = bfqd->bfq_slice_idle;
2500 * Unless the queue is being weight-raised or the scenario is
2501 * asymmetric, grant only minimum idle time if the queue
2502 * is seeky. A long idling is preserved for a weight-raised
2503 * queue, or, more in general, in an asymmetric scenario,
2504 * because a long idling is needed for guaranteeing to a queue
2505 * its reserved share of the throughput (in particular, it is
2506 * needed if the queue has a higher weight than some other
2509 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
2510 bfq_symmetric_scenario(bfqd))
2511 sl = min_t(u64, sl, BFQ_MIN_TT);
2513 bfqd->last_idling_start = ktime_get();
2514 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
2516 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
2520 * In autotuning mode, max_budget is dynamically recomputed as the
2521 * amount of sectors transferred in timeout at the estimated peak
2522 * rate. This enables BFQ to utilize a full timeslice with a full
2523 * budget, even if the in-service queue is served at peak rate. And
2524 * this maximises throughput with sequential workloads.
2526 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
2528 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
2529 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
2533 * Update parameters related to throughput and responsiveness, as a
2534 * function of the estimated peak rate. See comments on
2535 * bfq_calc_max_budget(), and on the ref_wr_duration array.
2537 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
2539 if (bfqd->bfq_user_max_budget == 0) {
2540 bfqd->bfq_max_budget =
2541 bfq_calc_max_budget(bfqd);
2542 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
2546 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
2549 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
2550 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
2551 bfqd->peak_rate_samples = 1;
2552 bfqd->sequential_samples = 0;
2553 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
2555 } else /* no new rq dispatched, just reset the number of samples */
2556 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
2559 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
2560 bfqd->peak_rate_samples, bfqd->sequential_samples,
2561 bfqd->tot_sectors_dispatched);
2564 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
2566 u32 rate, weight, divisor;
2569 * For the convergence property to hold (see comments on
2570 * bfq_update_peak_rate()) and for the assessment to be
2571 * reliable, a minimum number of samples must be present, and
2572 * a minimum amount of time must have elapsed. If not so, do
2573 * not compute new rate. Just reset parameters, to get ready
2574 * for a new evaluation attempt.
2576 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
2577 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
2578 goto reset_computation;
2581 * If a new request completion has occurred after last
2582 * dispatch, then, to approximate the rate at which requests
2583 * have been served by the device, it is more precise to
2584 * extend the observation interval to the last completion.
2586 bfqd->delta_from_first =
2587 max_t(u64, bfqd->delta_from_first,
2588 bfqd->last_completion - bfqd->first_dispatch);
2591 * Rate computed in sects/usec, and not sects/nsec, for
2594 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
2595 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
2598 * Peak rate not updated if:
2599 * - the percentage of sequential dispatches is below 3/4 of the
2600 * total, and rate is below the current estimated peak rate
2601 * - rate is unreasonably high (> 20M sectors/sec)
2603 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
2604 rate <= bfqd->peak_rate) ||
2605 rate > 20<<BFQ_RATE_SHIFT)
2606 goto reset_computation;
2609 * We have to update the peak rate, at last! To this purpose,
2610 * we use a low-pass filter. We compute the smoothing constant
2611 * of the filter as a function of the 'weight' of the new
2614 * As can be seen in next formulas, we define this weight as a
2615 * quantity proportional to how sequential the workload is,
2616 * and to how long the observation time interval is.
2618 * The weight runs from 0 to 8. The maximum value of the
2619 * weight, 8, yields the minimum value for the smoothing
2620 * constant. At this minimum value for the smoothing constant,
2621 * the measured rate contributes for half of the next value of
2622 * the estimated peak rate.
2624 * So, the first step is to compute the weight as a function
2625 * of how sequential the workload is. Note that the weight
2626 * cannot reach 9, because bfqd->sequential_samples cannot
2627 * become equal to bfqd->peak_rate_samples, which, in its
2628 * turn, holds true because bfqd->sequential_samples is not
2629 * incremented for the first sample.
2631 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
2634 * Second step: further refine the weight as a function of the
2635 * duration of the observation interval.
2637 weight = min_t(u32, 8,
2638 div_u64(weight * bfqd->delta_from_first,
2639 BFQ_RATE_REF_INTERVAL));
2642 * Divisor ranging from 10, for minimum weight, to 2, for
2645 divisor = 10 - weight;
2648 * Finally, update peak rate:
2650 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
2652 bfqd->peak_rate *= divisor-1;
2653 bfqd->peak_rate /= divisor;
2654 rate /= divisor; /* smoothing constant alpha = 1/divisor */
2656 bfqd->peak_rate += rate;
2659 * For a very slow device, bfqd->peak_rate can reach 0 (see
2660 * the minimum representable values reported in the comments
2661 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
2662 * divisions by zero where bfqd->peak_rate is used as a
2665 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
2667 update_thr_responsiveness_params(bfqd);
2670 bfq_reset_rate_computation(bfqd, rq);
2674 * Update the read/write peak rate (the main quantity used for
2675 * auto-tuning, see update_thr_responsiveness_params()).
2677 * It is not trivial to estimate the peak rate (correctly): because of
2678 * the presence of sw and hw queues between the scheduler and the
2679 * device components that finally serve I/O requests, it is hard to
2680 * say exactly when a given dispatched request is served inside the
2681 * device, and for how long. As a consequence, it is hard to know
2682 * precisely at what rate a given set of requests is actually served
2685 * On the opposite end, the dispatch time of any request is trivially
2686 * available, and, from this piece of information, the "dispatch rate"
2687 * of requests can be immediately computed. So, the idea in the next
2688 * function is to use what is known, namely request dispatch times
2689 * (plus, when useful, request completion times), to estimate what is
2690 * unknown, namely in-device request service rate.
2692 * The main issue is that, because of the above facts, the rate at
2693 * which a certain set of requests is dispatched over a certain time
2694 * interval can vary greatly with respect to the rate at which the
2695 * same requests are then served. But, since the size of any
2696 * intermediate queue is limited, and the service scheme is lossless
2697 * (no request is silently dropped), the following obvious convergence
2698 * property holds: the number of requests dispatched MUST become
2699 * closer and closer to the number of requests completed as the
2700 * observation interval grows. This is the key property used in
2701 * the next function to estimate the peak service rate as a function
2702 * of the observed dispatch rate. The function assumes to be invoked
2703 * on every request dispatch.
2705 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
2707 u64 now_ns = ktime_get_ns();
2709 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
2710 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
2711 bfqd->peak_rate_samples);
2712 bfq_reset_rate_computation(bfqd, rq);
2713 goto update_last_values; /* will add one sample */
2717 * Device idle for very long: the observation interval lasting
2718 * up to this dispatch cannot be a valid observation interval
2719 * for computing a new peak rate (similarly to the late-
2720 * completion event in bfq_completed_request()). Go to
2721 * update_rate_and_reset to have the following three steps
2723 * - close the observation interval at the last (previous)
2724 * request dispatch or completion
2725 * - compute rate, if possible, for that observation interval
2726 * - start a new observation interval with this dispatch
2728 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
2729 bfqd->rq_in_driver == 0)
2730 goto update_rate_and_reset;
2732 /* Update sampling information */
2733 bfqd->peak_rate_samples++;
2735 if ((bfqd->rq_in_driver > 0 ||
2736 now_ns - bfqd->last_completion < BFQ_MIN_TT)
2737 && get_sdist(bfqd->last_position, rq) < BFQQ_SEEK_THR)
2738 bfqd->sequential_samples++;
2740 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
2742 /* Reset max observed rq size every 32 dispatches */
2743 if (likely(bfqd->peak_rate_samples % 32))
2744 bfqd->last_rq_max_size =
2745 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
2747 bfqd->last_rq_max_size = blk_rq_sectors(rq);
2749 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
2751 /* Target observation interval not yet reached, go on sampling */
2752 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
2753 goto update_last_values;
2755 update_rate_and_reset:
2756 bfq_update_rate_reset(bfqd, rq);
2758 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2759 if (RQ_BFQQ(rq) == bfqd->in_service_queue)
2760 bfqd->in_serv_last_pos = bfqd->last_position;
2761 bfqd->last_dispatch = now_ns;
2765 * Remove request from internal lists.
2767 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
2769 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2772 * For consistency, the next instruction should have been
2773 * executed after removing the request from the queue and
2774 * dispatching it. We execute instead this instruction before
2775 * bfq_remove_request() (and hence introduce a temporary
2776 * inconsistency), for efficiency. In fact, should this
2777 * dispatch occur for a non in-service bfqq, this anticipated
2778 * increment prevents two counters related to bfqq->dispatched
2779 * from risking to be, first, uselessly decremented, and then
2780 * incremented again when the (new) value of bfqq->dispatched
2781 * happens to be taken into account.
2784 bfq_update_peak_rate(q->elevator->elevator_data, rq);
2786 bfq_remove_request(q, rq);
2789 static void __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq)
2792 * If this bfqq is shared between multiple processes, check
2793 * to make sure that those processes are still issuing I/Os
2794 * within the mean seek distance. If not, it may be time to
2795 * break the queues apart again.
2797 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
2798 bfq_mark_bfqq_split_coop(bfqq);
2800 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2801 if (bfqq->dispatched == 0)
2803 * Overloading budget_timeout field to store
2804 * the time at which the queue remains with no
2805 * backlog and no outstanding request; used by
2806 * the weight-raising mechanism.
2808 bfqq->budget_timeout = jiffies;
2810 bfq_del_bfqq_busy(bfqd, bfqq, true);
2812 bfq_requeue_bfqq(bfqd, bfqq, true);
2814 * Resort priority tree of potential close cooperators.
2816 bfq_pos_tree_add_move(bfqd, bfqq);
2820 * All in-service entities must have been properly deactivated
2821 * or requeued before executing the next function, which
2822 * resets all in-service entites as no more in service.
2824 __bfq_bfqd_reset_in_service(bfqd);
2828 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
2829 * @bfqd: device data.
2830 * @bfqq: queue to update.
2831 * @reason: reason for expiration.
2833 * Handle the feedback on @bfqq budget at queue expiration.
2834 * See the body for detailed comments.
2836 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
2837 struct bfq_queue *bfqq,
2838 enum bfqq_expiration reason)
2840 struct request *next_rq;
2841 int budget, min_budget;
2843 min_budget = bfq_min_budget(bfqd);
2845 if (bfqq->wr_coeff == 1)
2846 budget = bfqq->max_budget;
2848 * Use a constant, low budget for weight-raised queues,
2849 * to help achieve a low latency. Keep it slightly higher
2850 * than the minimum possible budget, to cause a little
2851 * bit fewer expirations.
2853 budget = 2 * min_budget;
2855 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
2856 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
2857 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
2858 budget, bfq_min_budget(bfqd));
2859 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
2860 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
2862 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
2865 * Caveat: in all the following cases we trade latency
2868 case BFQQE_TOO_IDLE:
2870 * This is the only case where we may reduce
2871 * the budget: if there is no request of the
2872 * process still waiting for completion, then
2873 * we assume (tentatively) that the timer has
2874 * expired because the batch of requests of
2875 * the process could have been served with a
2876 * smaller budget. Hence, betting that
2877 * process will behave in the same way when it
2878 * becomes backlogged again, we reduce its
2879 * next budget. As long as we guess right,
2880 * this budget cut reduces the latency
2881 * experienced by the process.
2883 * However, if there are still outstanding
2884 * requests, then the process may have not yet
2885 * issued its next request just because it is
2886 * still waiting for the completion of some of
2887 * the still outstanding ones. So in this
2888 * subcase we do not reduce its budget, on the
2889 * contrary we increase it to possibly boost
2890 * the throughput, as discussed in the
2891 * comments to the BUDGET_TIMEOUT case.
2893 if (bfqq->dispatched > 0) /* still outstanding reqs */
2894 budget = min(budget * 2, bfqd->bfq_max_budget);
2896 if (budget > 5 * min_budget)
2897 budget -= 4 * min_budget;
2899 budget = min_budget;
2902 case BFQQE_BUDGET_TIMEOUT:
2904 * We double the budget here because it gives
2905 * the chance to boost the throughput if this
2906 * is not a seeky process (and has bumped into
2907 * this timeout because of, e.g., ZBR).
2909 budget = min(budget * 2, bfqd->bfq_max_budget);
2911 case BFQQE_BUDGET_EXHAUSTED:
2913 * The process still has backlog, and did not
2914 * let either the budget timeout or the disk
2915 * idling timeout expire. Hence it is not
2916 * seeky, has a short thinktime and may be
2917 * happy with a higher budget too. So
2918 * definitely increase the budget of this good
2919 * candidate to boost the disk throughput.
2921 budget = min(budget * 4, bfqd->bfq_max_budget);
2923 case BFQQE_NO_MORE_REQUESTS:
2925 * For queues that expire for this reason, it
2926 * is particularly important to keep the
2927 * budget close to the actual service they
2928 * need. Doing so reduces the timestamp
2929 * misalignment problem described in the
2930 * comments in the body of
2931 * __bfq_activate_entity. In fact, suppose
2932 * that a queue systematically expires for
2933 * BFQQE_NO_MORE_REQUESTS and presents a
2934 * new request in time to enjoy timestamp
2935 * back-shifting. The larger the budget of the
2936 * queue is with respect to the service the
2937 * queue actually requests in each service
2938 * slot, the more times the queue can be
2939 * reactivated with the same virtual finish
2940 * time. It follows that, even if this finish
2941 * time is pushed to the system virtual time
2942 * to reduce the consequent timestamp
2943 * misalignment, the queue unjustly enjoys for
2944 * many re-activations a lower finish time
2945 * than all newly activated queues.
2947 * The service needed by bfqq is measured
2948 * quite precisely by bfqq->entity.service.
2949 * Since bfqq does not enjoy device idling,
2950 * bfqq->entity.service is equal to the number
2951 * of sectors that the process associated with
2952 * bfqq requested to read/write before waiting
2953 * for request completions, or blocking for
2956 budget = max_t(int, bfqq->entity.service, min_budget);
2961 } else if (!bfq_bfqq_sync(bfqq)) {
2963 * Async queues get always the maximum possible
2964 * budget, as for them we do not care about latency
2965 * (in addition, their ability to dispatch is limited
2966 * by the charging factor).
2968 budget = bfqd->bfq_max_budget;
2971 bfqq->max_budget = budget;
2973 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
2974 !bfqd->bfq_user_max_budget)
2975 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
2978 * If there is still backlog, then assign a new budget, making
2979 * sure that it is large enough for the next request. Since
2980 * the finish time of bfqq must be kept in sync with the
2981 * budget, be sure to call __bfq_bfqq_expire() *after* this
2984 * If there is no backlog, then no need to update the budget;
2985 * it will be updated on the arrival of a new request.
2987 next_rq = bfqq->next_rq;
2989 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
2990 bfq_serv_to_charge(next_rq, bfqq));
2992 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
2993 next_rq ? blk_rq_sectors(next_rq) : 0,
2994 bfqq->entity.budget);
2998 * Return true if the process associated with bfqq is "slow". The slow
2999 * flag is used, in addition to the budget timeout, to reduce the
3000 * amount of service provided to seeky processes, and thus reduce
3001 * their chances to lower the throughput. More details in the comments
3002 * on the function bfq_bfqq_expire().
3004 * An important observation is in order: as discussed in the comments
3005 * on the function bfq_update_peak_rate(), with devices with internal
3006 * queues, it is hard if ever possible to know when and for how long
3007 * an I/O request is processed by the device (apart from the trivial
3008 * I/O pattern where a new request is dispatched only after the
3009 * previous one has been completed). This makes it hard to evaluate
3010 * the real rate at which the I/O requests of each bfq_queue are
3011 * served. In fact, for an I/O scheduler like BFQ, serving a
3012 * bfq_queue means just dispatching its requests during its service
3013 * slot (i.e., until the budget of the queue is exhausted, or the
3014 * queue remains idle, or, finally, a timeout fires). But, during the
3015 * service slot of a bfq_queue, around 100 ms at most, the device may
3016 * be even still processing requests of bfq_queues served in previous
3017 * service slots. On the opposite end, the requests of the in-service
3018 * bfq_queue may be completed after the service slot of the queue
3021 * Anyway, unless more sophisticated solutions are used
3022 * (where possible), the sum of the sizes of the requests dispatched
3023 * during the service slot of a bfq_queue is probably the only
3024 * approximation available for the service received by the bfq_queue
3025 * during its service slot. And this sum is the quantity used in this
3026 * function to evaluate the I/O speed of a process.
3028 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3029 bool compensate, enum bfqq_expiration reason,
3030 unsigned long *delta_ms)
3032 ktime_t delta_ktime;
3034 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
3036 if (!bfq_bfqq_sync(bfqq))
3040 delta_ktime = bfqd->last_idling_start;
3042 delta_ktime = ktime_get();
3043 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
3044 delta_usecs = ktime_to_us(delta_ktime);
3046 /* don't use too short time intervals */
3047 if (delta_usecs < 1000) {
3048 if (blk_queue_nonrot(bfqd->queue))
3050 * give same worst-case guarantees as idling
3053 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
3054 else /* charge at least one seek */
3055 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
3060 *delta_ms = delta_usecs / USEC_PER_MSEC;
3063 * Use only long (> 20ms) intervals to filter out excessive
3064 * spikes in service rate estimation.
3066 if (delta_usecs > 20000) {
3068 * Caveat for rotational devices: processes doing I/O
3069 * in the slower disk zones tend to be slow(er) even
3070 * if not seeky. In this respect, the estimated peak
3071 * rate is likely to be an average over the disk
3072 * surface. Accordingly, to not be too harsh with
3073 * unlucky processes, a process is deemed slow only if
3074 * its rate has been lower than half of the estimated
3077 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
3080 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
3086 * To be deemed as soft real-time, an application must meet two
3087 * requirements. First, the application must not require an average
3088 * bandwidth higher than the approximate bandwidth required to playback or
3089 * record a compressed high-definition video.
3090 * The next function is invoked on the completion of the last request of a
3091 * batch, to compute the next-start time instant, soft_rt_next_start, such
3092 * that, if the next request of the application does not arrive before
3093 * soft_rt_next_start, then the above requirement on the bandwidth is met.
3095 * The second requirement is that the request pattern of the application is
3096 * isochronous, i.e., that, after issuing a request or a batch of requests,
3097 * the application stops issuing new requests until all its pending requests
3098 * have been completed. After that, the application may issue a new batch,
3100 * For this reason the next function is invoked to compute
3101 * soft_rt_next_start only for applications that meet this requirement,
3102 * whereas soft_rt_next_start is set to infinity for applications that do
3105 * Unfortunately, even a greedy (i.e., I/O-bound) application may
3106 * happen to meet, occasionally or systematically, both the above
3107 * bandwidth and isochrony requirements. This may happen at least in
3108 * the following circumstances. First, if the CPU load is high. The
3109 * application may stop issuing requests while the CPUs are busy
3110 * serving other processes, then restart, then stop again for a while,
3111 * and so on. The other circumstances are related to the storage
3112 * device: the storage device is highly loaded or reaches a low-enough
3113 * throughput with the I/O of the application (e.g., because the I/O
3114 * is random and/or the device is slow). In all these cases, the
3115 * I/O of the application may be simply slowed down enough to meet
3116 * the bandwidth and isochrony requirements. To reduce the probability
3117 * that greedy applications are deemed as soft real-time in these
3118 * corner cases, a further rule is used in the computation of
3119 * soft_rt_next_start: the return value of this function is forced to
3120 * be higher than the maximum between the following two quantities.
3122 * (a) Current time plus: (1) the maximum time for which the arrival
3123 * of a request is waited for when a sync queue becomes idle,
3124 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
3125 * postpone for a moment the reason for adding a few extra
3126 * jiffies; we get back to it after next item (b). Lower-bounding
3127 * the return value of this function with the current time plus
3128 * bfqd->bfq_slice_idle tends to filter out greedy applications,
3129 * because the latter issue their next request as soon as possible
3130 * after the last one has been completed. In contrast, a soft
3131 * real-time application spends some time processing data, after a
3132 * batch of its requests has been completed.
3134 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
3135 * above, greedy applications may happen to meet both the
3136 * bandwidth and isochrony requirements under heavy CPU or
3137 * storage-device load. In more detail, in these scenarios, these
3138 * applications happen, only for limited time periods, to do I/O
3139 * slowly enough to meet all the requirements described so far,
3140 * including the filtering in above item (a). These slow-speed
3141 * time intervals are usually interspersed between other time
3142 * intervals during which these applications do I/O at a very high
3143 * speed. Fortunately, exactly because of the high speed of the
3144 * I/O in the high-speed intervals, the values returned by this
3145 * function happen to be so high, near the end of any such
3146 * high-speed interval, to be likely to fall *after* the end of
3147 * the low-speed time interval that follows. These high values are
3148 * stored in bfqq->soft_rt_next_start after each invocation of
3149 * this function. As a consequence, if the last value of
3150 * bfqq->soft_rt_next_start is constantly used to lower-bound the
3151 * next value that this function may return, then, from the very
3152 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
3153 * likely to be constantly kept so high that any I/O request
3154 * issued during the low-speed interval is considered as arriving
3155 * to soon for the application to be deemed as soft
3156 * real-time. Then, in the high-speed interval that follows, the
3157 * application will not be deemed as soft real-time, just because
3158 * it will do I/O at a high speed. And so on.
3160 * Getting back to the filtering in item (a), in the following two
3161 * cases this filtering might be easily passed by a greedy
3162 * application, if the reference quantity was just
3163 * bfqd->bfq_slice_idle:
3164 * 1) HZ is so low that the duration of a jiffy is comparable to or
3165 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
3166 * devices with HZ=100. The time granularity may be so coarse
3167 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
3168 * is rather lower than the exact value.
3169 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
3170 * for a while, then suddenly 'jump' by several units to recover the lost
3171 * increments. This seems to happen, e.g., inside virtual machines.
3172 * To address this issue, in the filtering in (a) we do not use as a
3173 * reference time interval just bfqd->bfq_slice_idle, but
3174 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
3175 * minimum number of jiffies for which the filter seems to be quite
3176 * precise also in embedded systems and KVM/QEMU virtual machines.
3178 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
3179 struct bfq_queue *bfqq)
3181 return max3(bfqq->soft_rt_next_start,
3182 bfqq->last_idle_bklogged +
3183 HZ * bfqq->service_from_backlogged /
3184 bfqd->bfq_wr_max_softrt_rate,
3185 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
3189 * bfq_bfqq_expire - expire a queue.
3190 * @bfqd: device owning the queue.
3191 * @bfqq: the queue to expire.
3192 * @compensate: if true, compensate for the time spent idling.
3193 * @reason: the reason causing the expiration.
3195 * If the process associated with bfqq does slow I/O (e.g., because it
3196 * issues random requests), we charge bfqq with the time it has been
3197 * in service instead of the service it has received (see
3198 * bfq_bfqq_charge_time for details on how this goal is achieved). As
3199 * a consequence, bfqq will typically get higher timestamps upon
3200 * reactivation, and hence it will be rescheduled as if it had
3201 * received more service than what it has actually received. In the
3202 * end, bfqq receives less service in proportion to how slowly its
3203 * associated process consumes its budgets (and hence how seriously it
3204 * tends to lower the throughput). In addition, this time-charging
3205 * strategy guarantees time fairness among slow processes. In
3206 * contrast, if the process associated with bfqq is not slow, we
3207 * charge bfqq exactly with the service it has received.
3209 * Charging time to the first type of queues and the exact service to
3210 * the other has the effect of using the WF2Q+ policy to schedule the
3211 * former on a timeslice basis, without violating service domain
3212 * guarantees among the latter.
3214 void bfq_bfqq_expire(struct bfq_data *bfqd,
3215 struct bfq_queue *bfqq,
3217 enum bfqq_expiration reason)
3220 unsigned long delta = 0;
3221 struct bfq_entity *entity = &bfqq->entity;
3225 * Check whether the process is slow (see bfq_bfqq_is_slow).
3227 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
3230 * As above explained, charge slow (typically seeky) and
3231 * timed-out queues with the time and not the service
3232 * received, to favor sequential workloads.
3234 * Processes doing I/O in the slower disk zones will tend to
3235 * be slow(er) even if not seeky. Therefore, since the
3236 * estimated peak rate is actually an average over the disk
3237 * surface, these processes may timeout just for bad luck. To
3238 * avoid punishing them, do not charge time to processes that
3239 * succeeded in consuming at least 2/3 of their budget. This
3240 * allows BFQ to preserve enough elasticity to still perform
3241 * bandwidth, and not time, distribution with little unlucky
3242 * or quasi-sequential processes.
3244 if (bfqq->wr_coeff == 1 &&
3246 (reason == BFQQE_BUDGET_TIMEOUT &&
3247 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
3248 bfq_bfqq_charge_time(bfqd, bfqq, delta);
3250 if (reason == BFQQE_TOO_IDLE &&
3251 entity->service <= 2 * entity->budget / 10)
3252 bfq_clear_bfqq_IO_bound(bfqq);
3254 if (bfqd->low_latency && bfqq->wr_coeff == 1)
3255 bfqq->last_wr_start_finish = jiffies;
3257 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
3258 RB_EMPTY_ROOT(&bfqq->sort_list)) {
3260 * If we get here, and there are no outstanding
3261 * requests, then the request pattern is isochronous
3262 * (see the comments on the function
3263 * bfq_bfqq_softrt_next_start()). Thus we can compute
3264 * soft_rt_next_start. If, instead, the queue still
3265 * has outstanding requests, then we have to wait for
3266 * the completion of all the outstanding requests to
3267 * discover whether the request pattern is actually
3270 if (bfqq->dispatched == 0)
3271 bfqq->soft_rt_next_start =
3272 bfq_bfqq_softrt_next_start(bfqd, bfqq);
3275 * Schedule an update of soft_rt_next_start to when
3276 * the task may be discovered to be isochronous.
3278 bfq_mark_bfqq_softrt_update(bfqq);
3282 bfq_log_bfqq(bfqd, bfqq,
3283 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
3284 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
3287 * Increase, decrease or leave budget unchanged according to
3290 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
3292 __bfq_bfqq_expire(bfqd, bfqq);
3294 if (ref == 1) /* bfqq is gone, no more actions on it */
3297 /* mark bfqq as waiting a request only if a bic still points to it */
3298 if (!bfq_bfqq_busy(bfqq) &&
3299 reason != BFQQE_BUDGET_TIMEOUT &&
3300 reason != BFQQE_BUDGET_EXHAUSTED) {
3301 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
3303 * Not setting service to 0, because, if the next rq
3304 * arrives in time, the queue will go on receiving
3305 * service with this same budget (as if it never expired)
3308 entity->service = 0;
3311 * Reset the received-service counter for every parent entity.
3312 * Differently from what happens with bfqq->entity.service,
3313 * the resetting of this counter never needs to be postponed
3314 * for parent entities. In fact, in case bfqq may have a
3315 * chance to go on being served using the last, partially
3316 * consumed budget, bfqq->entity.service needs to be kept,
3317 * because if bfqq then actually goes on being served using
3318 * the same budget, the last value of bfqq->entity.service is
3319 * needed to properly decrement bfqq->entity.budget by the
3320 * portion already consumed. In contrast, it is not necessary
3321 * to keep entity->service for parent entities too, because
3322 * the bubble up of the new value of bfqq->entity.budget will
3323 * make sure that the budgets of parent entities are correct,
3324 * even in case bfqq and thus parent entities go on receiving
3325 * service with the same budget.
3327 entity = entity->parent;
3328 for_each_entity(entity)
3329 entity->service = 0;
3333 * Budget timeout is not implemented through a dedicated timer, but
3334 * just checked on request arrivals and completions, as well as on
3335 * idle timer expirations.
3337 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
3339 return time_is_before_eq_jiffies(bfqq->budget_timeout);
3343 * If we expire a queue that is actively waiting (i.e., with the
3344 * device idled) for the arrival of a new request, then we may incur
3345 * the timestamp misalignment problem described in the body of the
3346 * function __bfq_activate_entity. Hence we return true only if this
3347 * condition does not hold, or if the queue is slow enough to deserve
3348 * only to be kicked off for preserving a high throughput.
3350 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
3352 bfq_log_bfqq(bfqq->bfqd, bfqq,
3353 "may_budget_timeout: wait_request %d left %d timeout %d",
3354 bfq_bfqq_wait_request(bfqq),
3355 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
3356 bfq_bfqq_budget_timeout(bfqq));
3358 return (!bfq_bfqq_wait_request(bfqq) ||
3359 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
3361 bfq_bfqq_budget_timeout(bfqq);
3365 * For a queue that becomes empty, device idling is allowed only if
3366 * this function returns true for the queue. As a consequence, since
3367 * device idling plays a critical role in both throughput boosting and
3368 * service guarantees, the return value of this function plays a
3369 * critical role in both these aspects as well.
3371 * In a nutshell, this function returns true only if idling is
3372 * beneficial for throughput or, even if detrimental for throughput,
3373 * idling is however necessary to preserve service guarantees (low
3374 * latency, desired throughput distribution, ...). In particular, on
3375 * NCQ-capable devices, this function tries to return false, so as to
3376 * help keep the drives' internal queues full, whenever this helps the
3377 * device boost the throughput without causing any service-guarantee
3380 * In more detail, the return value of this function is obtained by,
3381 * first, computing a number of boolean variables that take into
3382 * account throughput and service-guarantee issues, and, then,
3383 * combining these variables in a logical expression. Most of the
3384 * issues taken into account are not trivial. We discuss these issues
3385 * individually while introducing the variables.
3387 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
3389 struct bfq_data *bfqd = bfqq->bfqd;
3390 bool rot_without_queueing =
3391 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
3392 bfqq_sequential_and_IO_bound,
3393 idling_boosts_thr, idling_boosts_thr_without_issues,
3394 idling_needed_for_service_guarantees,
3395 asymmetric_scenario;
3397 if (bfqd->strict_guarantees)
3401 * Idling is performed only if slice_idle > 0. In addition, we
3404 * (b) bfqq is in the idle io prio class: in this case we do
3405 * not idle because we want to minimize the bandwidth that
3406 * queues in this class can steal to higher-priority queues
3408 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
3409 bfq_class_idle(bfqq))
3412 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
3413 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
3416 * The next variable takes into account the cases where idling
3417 * boosts the throughput.
3419 * The value of the variable is computed considering, first, that
3420 * idling is virtually always beneficial for the throughput if:
3421 * (a) the device is not NCQ-capable and rotational, or
3422 * (b) regardless of the presence of NCQ, the device is rotational and
3423 * the request pattern for bfqq is I/O-bound and sequential, or
3424 * (c) regardless of whether it is rotational, the device is
3425 * not NCQ-capable and the request pattern for bfqq is
3426 * I/O-bound and sequential.
3428 * Secondly, and in contrast to the above item (b), idling an
3429 * NCQ-capable flash-based device would not boost the
3430 * throughput even with sequential I/O; rather it would lower
3431 * the throughput in proportion to how fast the device
3432 * is. Accordingly, the next variable is true if any of the
3433 * above conditions (a), (b) or (c) is true, and, in
3434 * particular, happens to be false if bfqd is an NCQ-capable
3435 * flash-based device.
3437 idling_boosts_thr = rot_without_queueing ||
3438 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
3439 bfqq_sequential_and_IO_bound);
3442 * The value of the next variable,
3443 * idling_boosts_thr_without_issues, is equal to that of
3444 * idling_boosts_thr, unless a special case holds. In this
3445 * special case, described below, idling may cause problems to
3446 * weight-raised queues.
3448 * When the request pool is saturated (e.g., in the presence
3449 * of write hogs), if the processes associated with
3450 * non-weight-raised queues ask for requests at a lower rate,
3451 * then processes associated with weight-raised queues have a
3452 * higher probability to get a request from the pool
3453 * immediately (or at least soon) when they need one. Thus
3454 * they have a higher probability to actually get a fraction
3455 * of the device throughput proportional to their high
3456 * weight. This is especially true with NCQ-capable drives,
3457 * which enqueue several requests in advance, and further
3458 * reorder internally-queued requests.
3460 * For this reason, we force to false the value of
3461 * idling_boosts_thr_without_issues if there are weight-raised
3462 * busy queues. In this case, and if bfqq is not weight-raised,
3463 * this guarantees that the device is not idled for bfqq (if,
3464 * instead, bfqq is weight-raised, then idling will be
3465 * guaranteed by another variable, see below). Combined with
3466 * the timestamping rules of BFQ (see [1] for details), this
3467 * behavior causes bfqq, and hence any sync non-weight-raised
3468 * queue, to get a lower number of requests served, and thus
3469 * to ask for a lower number of requests from the request
3470 * pool, before the busy weight-raised queues get served
3471 * again. This often mitigates starvation problems in the
3472 * presence of heavy write workloads and NCQ, thereby
3473 * guaranteeing a higher application and system responsiveness
3474 * in these hostile scenarios.
3476 idling_boosts_thr_without_issues = idling_boosts_thr &&
3477 bfqd->wr_busy_queues == 0;
3480 * There is then a case where idling must be performed not
3481 * for throughput concerns, but to preserve service
3484 * To introduce this case, we can note that allowing the drive
3485 * to enqueue more than one request at a time, and hence
3486 * delegating de facto final scheduling decisions to the
3487 * drive's internal scheduler, entails loss of control on the
3488 * actual request service order. In particular, the critical
3489 * situation is when requests from different processes happen
3490 * to be present, at the same time, in the internal queue(s)
3491 * of the drive. In such a situation, the drive, by deciding
3492 * the service order of the internally-queued requests, does
3493 * determine also the actual throughput distribution among
3494 * these processes. But the drive typically has no notion or
3495 * concern about per-process throughput distribution, and
3496 * makes its decisions only on a per-request basis. Therefore,
3497 * the service distribution enforced by the drive's internal
3498 * scheduler is likely to coincide with the desired
3499 * device-throughput distribution only in a completely
3500 * symmetric scenario where:
3501 * (i) each of these processes must get the same throughput as
3503 * (ii) all these processes have the same I/O pattern
3504 (either sequential or random).
3505 * In fact, in such a scenario, the drive will tend to treat
3506 * the requests of each of these processes in about the same
3507 * way as the requests of the others, and thus to provide
3508 * each of these processes with about the same throughput
3509 * (which is exactly the desired throughput distribution). In
3510 * contrast, in any asymmetric scenario, device idling is
3511 * certainly needed to guarantee that bfqq receives its
3512 * assigned fraction of the device throughput (see [1] for
3515 * We address this issue by controlling, actually, only the
3516 * symmetry sub-condition (i), i.e., provided that
3517 * sub-condition (i) holds, idling is not performed,
3518 * regardless of whether sub-condition (ii) holds. In other
3519 * words, only if sub-condition (i) holds, then idling is
3520 * allowed, and the device tends to be prevented from queueing
3521 * many requests, possibly of several processes. The reason
3522 * for not controlling also sub-condition (ii) is that we
3523 * exploit preemption to preserve guarantees in case of
3524 * symmetric scenarios, even if (ii) does not hold, as
3525 * explained in the next two paragraphs.
3527 * Even if a queue, say Q, is expired when it remains idle, Q
3528 * can still preempt the new in-service queue if the next
3529 * request of Q arrives soon (see the comments on
3530 * bfq_bfqq_update_budg_for_activation). If all queues and
3531 * groups have the same weight, this form of preemption,
3532 * combined with the hole-recovery heuristic described in the
3533 * comments on function bfq_bfqq_update_budg_for_activation,
3534 * are enough to preserve a correct bandwidth distribution in
3535 * the mid term, even without idling. In fact, even if not
3536 * idling allows the internal queues of the device to contain
3537 * many requests, and thus to reorder requests, we can rather
3538 * safely assume that the internal scheduler still preserves a
3539 * minimum of mid-term fairness. The motivation for using
3540 * preemption instead of idling is that, by not idling,
3541 * service guarantees are preserved without minimally
3542 * sacrificing throughput. In other words, both a high
3543 * throughput and its desired distribution are obtained.
3545 * More precisely, this preemption-based, idleless approach
3546 * provides fairness in terms of IOPS, and not sectors per
3547 * second. This can be seen with a simple example. Suppose
3548 * that there are two queues with the same weight, but that
3549 * the first queue receives requests of 8 sectors, while the
3550 * second queue receives requests of 1024 sectors. In
3551 * addition, suppose that each of the two queues contains at
3552 * most one request at a time, which implies that each queue
3553 * always remains idle after it is served. Finally, after
3554 * remaining idle, each queue receives very quickly a new
3555 * request. It follows that the two queues are served
3556 * alternatively, preempting each other if needed. This
3557 * implies that, although both queues have the same weight,
3558 * the queue with large requests receives a service that is
3559 * 1024/8 times as high as the service received by the other
3562 * On the other hand, device idling is performed, and thus
3563 * pure sector-domain guarantees are provided, for the
3564 * following queues, which are likely to need stronger
3565 * throughput guarantees: weight-raised queues, and queues
3566 * with a higher weight than other queues. When such queues
3567 * are active, sub-condition (i) is false, which triggers
3570 * According to the above considerations, the next variable is
3571 * true (only) if sub-condition (i) holds. To compute the
3572 * value of this variable, we not only use the return value of
3573 * the function bfq_symmetric_scenario(), but also check
3574 * whether bfqq is being weight-raised, because
3575 * bfq_symmetric_scenario() does not take into account also
3576 * weight-raised queues (see comments on
3577 * bfq_weights_tree_add()).
3579 * As a side note, it is worth considering that the above
3580 * device-idling countermeasures may however fail in the
3581 * following unlucky scenario: if idling is (correctly)
3582 * disabled in a time period during which all symmetry
3583 * sub-conditions hold, and hence the device is allowed to
3584 * enqueue many requests, but at some later point in time some
3585 * sub-condition stops to hold, then it may become impossible
3586 * to let requests be served in the desired order until all
3587 * the requests already queued in the device have been served.
3589 asymmetric_scenario = bfqq->wr_coeff > 1 ||
3590 !bfq_symmetric_scenario(bfqd);
3593 * Finally, there is a case where maximizing throughput is the
3594 * best choice even if it may cause unfairness toward
3595 * bfqq. Such a case is when bfqq became active in a burst of
3596 * queue activations. Queues that became active during a large
3597 * burst benefit only from throughput, as discussed in the
3598 * comments on bfq_handle_burst. Thus, if bfqq became active
3599 * in a burst and not idling the device maximizes throughput,
3600 * then the device must no be idled, because not idling the
3601 * device provides bfqq and all other queues in the burst with
3602 * maximum benefit. Combining this and the above case, we can
3603 * now establish when idling is actually needed to preserve
3604 * service guarantees.
3606 idling_needed_for_service_guarantees =
3607 asymmetric_scenario && !bfq_bfqq_in_large_burst(bfqq);
3610 * We have now all the components we need to compute the
3611 * return value of the function, which is true only if idling
3612 * either boosts the throughput (without issues), or is
3613 * necessary to preserve service guarantees.
3615 return idling_boosts_thr_without_issues ||
3616 idling_needed_for_service_guarantees;
3620 * If the in-service queue is empty but the function bfq_better_to_idle
3621 * returns true, then:
3622 * 1) the queue must remain in service and cannot be expired, and
3623 * 2) the device must be idled to wait for the possible arrival of a new
3624 * request for the queue.
3625 * See the comments on the function bfq_better_to_idle for the reasons
3626 * why performing device idling is the best choice to boost the throughput
3627 * and preserve service guarantees when bfq_better_to_idle itself
3630 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
3632 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
3636 * Select a queue for service. If we have a current queue in service,
3637 * check whether to continue servicing it, or retrieve and set a new one.
3639 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
3641 struct bfq_queue *bfqq;
3642 struct request *next_rq;
3643 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
3645 bfqq = bfqd->in_service_queue;
3649 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
3652 * Do not expire bfqq for budget timeout if bfqq may be about
3653 * to enjoy device idling. The reason why, in this case, we
3654 * prevent bfqq from expiring is the same as in the comments
3655 * on the case where bfq_bfqq_must_idle() returns true, in
3656 * bfq_completed_request().
3658 if (bfq_may_expire_for_budg_timeout(bfqq) &&
3659 !bfq_bfqq_must_idle(bfqq))
3664 * This loop is rarely executed more than once. Even when it
3665 * happens, it is much more convenient to re-execute this loop
3666 * than to return NULL and trigger a new dispatch to get a
3669 next_rq = bfqq->next_rq;
3671 * If bfqq has requests queued and it has enough budget left to
3672 * serve them, keep the queue, otherwise expire it.
3675 if (bfq_serv_to_charge(next_rq, bfqq) >
3676 bfq_bfqq_budget_left(bfqq)) {
3678 * Expire the queue for budget exhaustion,
3679 * which makes sure that the next budget is
3680 * enough to serve the next request, even if
3681 * it comes from the fifo expired path.
3683 reason = BFQQE_BUDGET_EXHAUSTED;
3687 * The idle timer may be pending because we may
3688 * not disable disk idling even when a new request
3691 if (bfq_bfqq_wait_request(bfqq)) {
3693 * If we get here: 1) at least a new request
3694 * has arrived but we have not disabled the
3695 * timer because the request was too small,
3696 * 2) then the block layer has unplugged
3697 * the device, causing the dispatch to be
3700 * Since the device is unplugged, now the
3701 * requests are probably large enough to
3702 * provide a reasonable throughput.
3703 * So we disable idling.
3705 bfq_clear_bfqq_wait_request(bfqq);
3706 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
3713 * No requests pending. However, if the in-service queue is idling
3714 * for a new request, or has requests waiting for a completion and
3715 * may idle after their completion, then keep it anyway.
3717 if (bfq_bfqq_wait_request(bfqq) ||
3718 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
3723 reason = BFQQE_NO_MORE_REQUESTS;
3725 bfq_bfqq_expire(bfqd, bfqq, false, reason);
3727 bfqq = bfq_set_in_service_queue(bfqd);
3729 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
3734 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
3736 bfq_log(bfqd, "select_queue: no queue returned");
3741 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3743 struct bfq_entity *entity = &bfqq->entity;
3745 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
3746 bfq_log_bfqq(bfqd, bfqq,
3747 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
3748 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
3749 jiffies_to_msecs(bfqq->wr_cur_max_time),
3751 bfqq->entity.weight, bfqq->entity.orig_weight);
3753 if (entity->prio_changed)
3754 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
3757 * If the queue was activated in a burst, or too much
3758 * time has elapsed from the beginning of this
3759 * weight-raising period, then end weight raising.
3761 if (bfq_bfqq_in_large_burst(bfqq))
3762 bfq_bfqq_end_wr(bfqq);
3763 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
3764 bfqq->wr_cur_max_time)) {
3765 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
3766 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
3767 bfq_wr_duration(bfqd)))
3768 bfq_bfqq_end_wr(bfqq);
3770 switch_back_to_interactive_wr(bfqq, bfqd);
3771 bfqq->entity.prio_changed = 1;
3774 if (bfqq->wr_coeff > 1 &&
3775 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
3776 bfqq->service_from_wr > max_service_from_wr) {
3777 /* see comments on max_service_from_wr */
3778 bfq_bfqq_end_wr(bfqq);
3782 * To improve latency (for this or other queues), immediately
3783 * update weight both if it must be raised and if it must be
3784 * lowered. Since, entity may be on some active tree here, and
3785 * might have a pending change of its ioprio class, invoke
3786 * next function with the last parameter unset (see the
3787 * comments on the function).
3789 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
3790 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
3795 * Dispatch next request from bfqq.
3797 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
3798 struct bfq_queue *bfqq)
3800 struct request *rq = bfqq->next_rq;
3801 unsigned long service_to_charge;
3803 service_to_charge = bfq_serv_to_charge(rq, bfqq);
3805 bfq_bfqq_served(bfqq, service_to_charge);
3807 bfq_dispatch_remove(bfqd->queue, rq);
3810 * If weight raising has to terminate for bfqq, then next
3811 * function causes an immediate update of bfqq's weight,
3812 * without waiting for next activation. As a consequence, on
3813 * expiration, bfqq will be timestamped as if has never been
3814 * weight-raised during this service slot, even if it has
3815 * received part or even most of the service as a
3816 * weight-raised queue. This inflates bfqq's timestamps, which
3817 * is beneficial, as bfqq is then more willing to leave the
3818 * device immediately to possible other weight-raised queues.
3820 bfq_update_wr_data(bfqd, bfqq);
3823 * Expire bfqq, pretending that its budget expired, if bfqq
3824 * belongs to CLASS_IDLE and other queues are waiting for
3827 if (bfqd->busy_queues > 1 && bfq_class_idle(bfqq))
3833 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
3837 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
3839 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3842 * Avoiding lock: a race on bfqd->busy_queues should cause at
3843 * most a call to dispatch for nothing
3845 return !list_empty_careful(&bfqd->dispatch) ||
3846 bfqd->busy_queues > 0;
3849 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3851 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3852 struct request *rq = NULL;
3853 struct bfq_queue *bfqq = NULL;
3855 if (!list_empty(&bfqd->dispatch)) {
3856 rq = list_first_entry(&bfqd->dispatch, struct request,
3858 list_del_init(&rq->queuelist);
3864 * Increment counters here, because this
3865 * dispatch does not follow the standard
3866 * dispatch flow (where counters are
3871 goto inc_in_driver_start_rq;
3875 * We exploit the bfq_finish_requeue_request hook to
3876 * decrement rq_in_driver, but
3877 * bfq_finish_requeue_request will not be invoked on
3878 * this request. So, to avoid unbalance, just start
3879 * this request, without incrementing rq_in_driver. As
3880 * a negative consequence, rq_in_driver is deceptively
3881 * lower than it should be while this request is in
3882 * service. This may cause bfq_schedule_dispatch to be
3883 * invoked uselessly.
3885 * As for implementing an exact solution, the
3886 * bfq_finish_requeue_request hook, if defined, is
3887 * probably invoked also on this request. So, by
3888 * exploiting this hook, we could 1) increment
3889 * rq_in_driver here, and 2) decrement it in
3890 * bfq_finish_requeue_request. Such a solution would
3891 * let the value of the counter be always accurate,
3892 * but it would entail using an extra interface
3893 * function. This cost seems higher than the benefit,
3894 * being the frequency of non-elevator-private
3895 * requests very low.
3900 bfq_log(bfqd, "dispatch requests: %d busy queues", bfqd->busy_queues);
3902 if (bfqd->busy_queues == 0)
3906 * Force device to serve one request at a time if
3907 * strict_guarantees is true. Forcing this service scheme is
3908 * currently the ONLY way to guarantee that the request
3909 * service order enforced by the scheduler is respected by a
3910 * queueing device. Otherwise the device is free even to make
3911 * some unlucky request wait for as long as the device
3914 * Of course, serving one request at at time may cause loss of
3917 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
3920 bfqq = bfq_select_queue(bfqd);
3924 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
3927 inc_in_driver_start_rq:
3928 bfqd->rq_in_driver++;
3930 rq->rq_flags |= RQF_STARTED;
3936 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
3937 static void bfq_update_dispatch_stats(struct request_queue *q,
3939 struct bfq_queue *in_serv_queue,
3940 bool idle_timer_disabled)
3942 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
3944 if (!idle_timer_disabled && !bfqq)
3948 * rq and bfqq are guaranteed to exist until this function
3949 * ends, for the following reasons. First, rq can be
3950 * dispatched to the device, and then can be completed and
3951 * freed, only after this function ends. Second, rq cannot be
3952 * merged (and thus freed because of a merge) any longer,
3953 * because it has already started. Thus rq cannot be freed
3954 * before this function ends, and, since rq has a reference to
3955 * bfqq, the same guarantee holds for bfqq too.
3957 * In addition, the following queue lock guarantees that
3958 * bfqq_group(bfqq) exists as well.
3960 spin_lock_irq(q->queue_lock);
3961 if (idle_timer_disabled)
3963 * Since the idle timer has been disabled,
3964 * in_serv_queue contained some request when
3965 * __bfq_dispatch_request was invoked above, which
3966 * implies that rq was picked exactly from
3967 * in_serv_queue. Thus in_serv_queue == bfqq, and is
3968 * therefore guaranteed to exist because of the above
3971 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
3973 struct bfq_group *bfqg = bfqq_group(bfqq);
3975 bfqg_stats_update_avg_queue_size(bfqg);
3976 bfqg_stats_set_start_empty_time(bfqg);
3977 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
3979 spin_unlock_irq(q->queue_lock);
3982 static inline void bfq_update_dispatch_stats(struct request_queue *q,
3984 struct bfq_queue *in_serv_queue,
3985 bool idle_timer_disabled) {}
3988 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3990 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3992 struct bfq_queue *in_serv_queue;
3993 bool waiting_rq, idle_timer_disabled;
3995 spin_lock_irq(&bfqd->lock);
3997 in_serv_queue = bfqd->in_service_queue;
3998 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
4000 rq = __bfq_dispatch_request(hctx);
4002 idle_timer_disabled =
4003 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
4005 spin_unlock_irq(&bfqd->lock);
4007 bfq_update_dispatch_stats(hctx->queue, rq, in_serv_queue,
4008 idle_timer_disabled);
4014 * Task holds one reference to the queue, dropped when task exits. Each rq
4015 * in-flight on this queue also holds a reference, dropped when rq is freed.
4017 * Scheduler lock must be held here. Recall not to use bfqq after calling
4018 * this function on it.
4020 void bfq_put_queue(struct bfq_queue *bfqq)
4022 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4023 struct bfq_group *bfqg = bfqq_group(bfqq);
4027 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
4034 if (!hlist_unhashed(&bfqq->burst_list_node)) {
4035 hlist_del_init(&bfqq->burst_list_node);
4037 * Decrement also burst size after the removal, if the
4038 * process associated with bfqq is exiting, and thus
4039 * does not contribute to the burst any longer. This
4040 * decrement helps filter out false positives of large
4041 * bursts, when some short-lived process (often due to
4042 * the execution of commands by some service) happens
4043 * to start and exit while a complex application is
4044 * starting, and thus spawning several processes that
4045 * do I/O (and that *must not* be treated as a large
4046 * burst, see comments on bfq_handle_burst).
4048 * In particular, the decrement is performed only if:
4049 * 1) bfqq is not a merged queue, because, if it is,
4050 * then this free of bfqq is not triggered by the exit
4051 * of the process bfqq is associated with, but exactly
4052 * by the fact that bfqq has just been merged.
4053 * 2) burst_size is greater than 0, to handle
4054 * unbalanced decrements. Unbalanced decrements may
4055 * happen in te following case: bfqq is inserted into
4056 * the current burst list--without incrementing
4057 * bust_size--because of a split, but the current
4058 * burst list is not the burst list bfqq belonged to
4059 * (see comments on the case of a split in
4062 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
4063 bfqq->bfqd->burst_size--;
4066 kmem_cache_free(bfq_pool, bfqq);
4067 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4068 bfqg_and_blkg_put(bfqg);
4072 static void bfq_put_cooperator(struct bfq_queue *bfqq)
4074 struct bfq_queue *__bfqq, *next;
4077 * If this queue was scheduled to merge with another queue, be
4078 * sure to drop the reference taken on that queue (and others in
4079 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
4081 __bfqq = bfqq->new_bfqq;
4085 next = __bfqq->new_bfqq;
4086 bfq_put_queue(__bfqq);
4091 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4093 if (bfqq == bfqd->in_service_queue) {
4094 __bfq_bfqq_expire(bfqd, bfqq);
4095 bfq_schedule_dispatch(bfqd);
4098 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
4100 bfq_put_cooperator(bfqq);
4102 bfq_put_queue(bfqq); /* release process reference */
4105 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
4107 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4108 struct bfq_data *bfqd;
4111 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
4114 unsigned long flags;
4116 spin_lock_irqsave(&bfqd->lock, flags);
4117 bfq_exit_bfqq(bfqd, bfqq);
4118 bic_set_bfqq(bic, NULL, is_sync);
4119 spin_unlock_irqrestore(&bfqd->lock, flags);
4123 static void bfq_exit_icq(struct io_cq *icq)
4125 struct bfq_io_cq *bic = icq_to_bic(icq);
4127 bfq_exit_icq_bfqq(bic, true);
4128 bfq_exit_icq_bfqq(bic, false);
4132 * Update the entity prio values; note that the new values will not
4133 * be used until the next (re)activation.
4136 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
4138 struct task_struct *tsk = current;
4140 struct bfq_data *bfqd = bfqq->bfqd;
4145 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4146 switch (ioprio_class) {
4148 dev_err(bfqq->bfqd->queue->backing_dev_info->dev,
4149 "bfq: bad prio class %d\n", ioprio_class);
4151 case IOPRIO_CLASS_NONE:
4153 * No prio set, inherit CPU scheduling settings.
4155 bfqq->new_ioprio = task_nice_ioprio(tsk);
4156 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
4158 case IOPRIO_CLASS_RT:
4159 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4160 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
4162 case IOPRIO_CLASS_BE:
4163 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4164 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
4166 case IOPRIO_CLASS_IDLE:
4167 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
4168 bfqq->new_ioprio = 7;
4172 if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
4173 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
4175 bfqq->new_ioprio = IOPRIO_BE_NR;
4178 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
4179 bfqq->entity.prio_changed = 1;
4182 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4183 struct bio *bio, bool is_sync,
4184 struct bfq_io_cq *bic);
4186 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
4188 struct bfq_data *bfqd = bic_to_bfqd(bic);
4189 struct bfq_queue *bfqq;
4190 int ioprio = bic->icq.ioc->ioprio;
4193 * This condition may trigger on a newly created bic, be sure to
4194 * drop the lock before returning.
4196 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
4199 bic->ioprio = ioprio;
4201 bfqq = bic_to_bfqq(bic, false);
4203 /* release process reference on this queue */
4204 bfq_put_queue(bfqq);
4205 bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
4206 bic_set_bfqq(bic, bfqq, false);
4209 bfqq = bic_to_bfqq(bic, true);
4211 bfq_set_next_ioprio_data(bfqq, bic);
4214 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4215 struct bfq_io_cq *bic, pid_t pid, int is_sync)
4217 RB_CLEAR_NODE(&bfqq->entity.rb_node);
4218 INIT_LIST_HEAD(&bfqq->fifo);
4219 INIT_HLIST_NODE(&bfqq->burst_list_node);
4225 bfq_set_next_ioprio_data(bfqq, bic);
4229 * No need to mark as has_short_ttime if in
4230 * idle_class, because no device idling is performed
4231 * for queues in idle class
4233 if (!bfq_class_idle(bfqq))
4234 /* tentatively mark as has_short_ttime */
4235 bfq_mark_bfqq_has_short_ttime(bfqq);
4236 bfq_mark_bfqq_sync(bfqq);
4237 bfq_mark_bfqq_just_created(bfqq);
4239 bfq_clear_bfqq_sync(bfqq);
4241 /* set end request to minus infinity from now */
4242 bfqq->ttime.last_end_request = ktime_get_ns() + 1;
4244 bfq_mark_bfqq_IO_bound(bfqq);
4248 /* Tentative initial value to trade off between thr and lat */
4249 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
4250 bfqq->budget_timeout = bfq_smallest_from_now();
4253 bfqq->last_wr_start_finish = jiffies;
4254 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
4255 bfqq->split_time = bfq_smallest_from_now();
4258 * To not forget the possibly high bandwidth consumed by a
4259 * process/queue in the recent past,
4260 * bfq_bfqq_softrt_next_start() returns a value at least equal
4261 * to the current value of bfqq->soft_rt_next_start (see
4262 * comments on bfq_bfqq_softrt_next_start). Set
4263 * soft_rt_next_start to now, to mean that bfqq has consumed
4264 * no bandwidth so far.
4266 bfqq->soft_rt_next_start = jiffies;
4268 /* first request is almost certainly seeky */
4269 bfqq->seek_history = 1;
4272 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
4273 struct bfq_group *bfqg,
4274 int ioprio_class, int ioprio)
4276 switch (ioprio_class) {
4277 case IOPRIO_CLASS_RT:
4278 return &bfqg->async_bfqq[0][ioprio];
4279 case IOPRIO_CLASS_NONE:
4280 ioprio = IOPRIO_NORM;
4282 case IOPRIO_CLASS_BE:
4283 return &bfqg->async_bfqq[1][ioprio];
4284 case IOPRIO_CLASS_IDLE:
4285 return &bfqg->async_idle_bfqq;
4291 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4292 struct bio *bio, bool is_sync,
4293 struct bfq_io_cq *bic)
4295 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4296 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4297 struct bfq_queue **async_bfqq = NULL;
4298 struct bfq_queue *bfqq;
4299 struct bfq_group *bfqg;
4303 bfqg = bfq_find_set_group(bfqd, bio_blkcg(bio));
4305 bfqq = &bfqd->oom_bfqq;
4310 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
4317 bfqq = kmem_cache_alloc_node(bfq_pool,
4318 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
4322 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
4324 bfq_init_entity(&bfqq->entity, bfqg);
4325 bfq_log_bfqq(bfqd, bfqq, "allocated");
4327 bfqq = &bfqd->oom_bfqq;
4328 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
4333 * Pin the queue now that it's allocated, scheduler exit will
4338 * Extra group reference, w.r.t. sync
4339 * queue. This extra reference is removed
4340 * only if bfqq->bfqg disappears, to
4341 * guarantee that this queue is not freed
4342 * until its group goes away.
4344 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
4350 bfqq->ref++; /* get a process reference to this queue */
4351 bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
4356 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
4357 struct bfq_queue *bfqq)
4359 struct bfq_ttime *ttime = &bfqq->ttime;
4360 u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
4362 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
4364 ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
4365 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
4366 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
4367 ttime->ttime_samples);
4371 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4374 bfqq->seek_history <<= 1;
4375 bfqq->seek_history |=
4376 get_sdist(bfqq->last_request_pos, rq) > BFQQ_SEEK_THR &&
4377 (!blk_queue_nonrot(bfqd->queue) ||
4378 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT);
4381 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
4382 struct bfq_queue *bfqq,
4383 struct bfq_io_cq *bic)
4385 bool has_short_ttime = true;
4388 * No need to update has_short_ttime if bfqq is async or in
4389 * idle io prio class, or if bfq_slice_idle is zero, because
4390 * no device idling is performed for bfqq in this case.
4392 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
4393 bfqd->bfq_slice_idle == 0)
4396 /* Idle window just restored, statistics are meaningless. */
4397 if (time_is_after_eq_jiffies(bfqq->split_time +
4398 bfqd->bfq_wr_min_idle_time))
4401 /* Think time is infinite if no process is linked to
4402 * bfqq. Otherwise check average think time to
4403 * decide whether to mark as has_short_ttime
4405 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
4406 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
4407 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle))
4408 has_short_ttime = false;
4410 bfq_log_bfqq(bfqd, bfqq, "update_has_short_ttime: has_short_ttime %d",
4413 if (has_short_ttime)
4414 bfq_mark_bfqq_has_short_ttime(bfqq);
4416 bfq_clear_bfqq_has_short_ttime(bfqq);
4420 * Called when a new fs request (rq) is added to bfqq. Check if there's
4421 * something we should do about it.
4423 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4426 struct bfq_io_cq *bic = RQ_BIC(rq);
4428 if (rq->cmd_flags & REQ_META)
4429 bfqq->meta_pending++;
4431 bfq_update_io_thinktime(bfqd, bfqq);
4432 bfq_update_has_short_ttime(bfqd, bfqq, bic);
4433 bfq_update_io_seektime(bfqd, bfqq, rq);
4435 bfq_log_bfqq(bfqd, bfqq,
4436 "rq_enqueued: has_short_ttime=%d (seeky %d)",
4437 bfq_bfqq_has_short_ttime(bfqq), BFQQ_SEEKY(bfqq));
4439 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4441 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
4442 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
4443 blk_rq_sectors(rq) < 32;
4444 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
4447 * There is just this request queued: if the request
4448 * is small and the queue is not to be expired, then
4451 * In this way, if the device is being idled to wait
4452 * for a new request from the in-service queue, we
4453 * avoid unplugging the device and committing the
4454 * device to serve just a small request. On the
4455 * contrary, we wait for the block layer to decide
4456 * when to unplug the device: hopefully, new requests
4457 * will be merged to this one quickly, then the device
4458 * will be unplugged and larger requests will be
4461 if (small_req && !budget_timeout)
4465 * A large enough request arrived, or the queue is to
4466 * be expired: in both cases disk idling is to be
4467 * stopped, so clear wait_request flag and reset
4470 bfq_clear_bfqq_wait_request(bfqq);
4471 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4474 * The queue is not empty, because a new request just
4475 * arrived. Hence we can safely expire the queue, in
4476 * case of budget timeout, without risking that the
4477 * timestamps of the queue are not updated correctly.
4478 * See [1] for more details.
4481 bfq_bfqq_expire(bfqd, bfqq, false,
4482 BFQQE_BUDGET_TIMEOUT);
4486 /* returns true if it causes the idle timer to be disabled */
4487 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
4489 struct bfq_queue *bfqq = RQ_BFQQ(rq),
4490 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
4491 bool waiting, idle_timer_disabled = false;
4494 if (bic_to_bfqq(RQ_BIC(rq), 1) != bfqq)
4495 new_bfqq = bic_to_bfqq(RQ_BIC(rq), 1);
4497 * Release the request's reference to the old bfqq
4498 * and make sure one is taken to the shared queue.
4500 new_bfqq->allocated++;
4504 * If the bic associated with the process
4505 * issuing this request still points to bfqq
4506 * (and thus has not been already redirected
4507 * to new_bfqq or even some other bfq_queue),
4508 * then complete the merge and redirect it to
4511 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
4512 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
4515 bfq_clear_bfqq_just_created(bfqq);
4517 * rq is about to be enqueued into new_bfqq,
4518 * release rq reference on bfqq
4520 bfq_put_queue(bfqq);
4521 rq->elv.priv[1] = new_bfqq;
4525 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
4526 bfq_add_request(rq);
4527 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
4529 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
4530 list_add_tail(&rq->queuelist, &bfqq->fifo);
4532 bfq_rq_enqueued(bfqd, bfqq, rq);
4534 return idle_timer_disabled;
4537 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
4538 static void bfq_update_insert_stats(struct request_queue *q,
4539 struct bfq_queue *bfqq,
4540 bool idle_timer_disabled,
4541 unsigned int cmd_flags)
4547 * bfqq still exists, because it can disappear only after
4548 * either it is merged with another queue, or the process it
4549 * is associated with exits. But both actions must be taken by
4550 * the same process currently executing this flow of
4553 * In addition, the following queue lock guarantees that
4554 * bfqq_group(bfqq) exists as well.
4556 spin_lock_irq(q->queue_lock);
4557 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
4558 if (idle_timer_disabled)
4559 bfqg_stats_update_idle_time(bfqq_group(bfqq));
4560 spin_unlock_irq(q->queue_lock);
4563 static inline void bfq_update_insert_stats(struct request_queue *q,
4564 struct bfq_queue *bfqq,
4565 bool idle_timer_disabled,
4566 unsigned int cmd_flags) {}
4569 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
4572 struct request_queue *q = hctx->queue;
4573 struct bfq_data *bfqd = q->elevator->elevator_data;
4574 struct bfq_queue *bfqq;
4575 bool idle_timer_disabled = false;
4576 unsigned int cmd_flags;
4578 spin_lock_irq(&bfqd->lock);
4579 if (blk_mq_sched_try_insert_merge(q, rq)) {
4580 spin_unlock_irq(&bfqd->lock);
4584 spin_unlock_irq(&bfqd->lock);
4586 blk_mq_sched_request_inserted(rq);
4588 spin_lock_irq(&bfqd->lock);
4589 bfqq = bfq_init_rq(rq);
4590 if (at_head || blk_rq_is_passthrough(rq)) {
4592 list_add(&rq->queuelist, &bfqd->dispatch);
4594 list_add_tail(&rq->queuelist, &bfqd->dispatch);
4595 } else { /* bfqq is assumed to be non null here */
4596 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
4598 * Update bfqq, because, if a queue merge has occurred
4599 * in __bfq_insert_request, then rq has been
4600 * redirected into a new queue.
4604 if (rq_mergeable(rq)) {
4605 elv_rqhash_add(q, rq);
4612 * Cache cmd_flags before releasing scheduler lock, because rq
4613 * may disappear afterwards (for example, because of a request
4616 cmd_flags = rq->cmd_flags;
4618 spin_unlock_irq(&bfqd->lock);
4620 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
4624 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
4625 struct list_head *list, bool at_head)
4627 while (!list_empty(list)) {
4630 rq = list_first_entry(list, struct request, queuelist);
4631 list_del_init(&rq->queuelist);
4632 bfq_insert_request(hctx, rq, at_head);
4636 static void bfq_update_hw_tag(struct bfq_data *bfqd)
4638 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
4639 bfqd->rq_in_driver);
4641 if (bfqd->hw_tag == 1)
4645 * This sample is valid if the number of outstanding requests
4646 * is large enough to allow a queueing behavior. Note that the
4647 * sum is not exact, as it's not taking into account deactivated
4650 if (bfqd->rq_in_driver + bfqd->queued < BFQ_HW_QUEUE_THRESHOLD)
4653 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
4656 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
4657 bfqd->max_rq_in_driver = 0;
4658 bfqd->hw_tag_samples = 0;
4661 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
4666 bfq_update_hw_tag(bfqd);
4668 bfqd->rq_in_driver--;
4671 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
4673 * Set budget_timeout (which we overload to store the
4674 * time at which the queue remains with no backlog and
4675 * no outstanding request; used by the weight-raising
4678 bfqq->budget_timeout = jiffies;
4680 bfq_weights_tree_remove(bfqd, bfqq);
4683 now_ns = ktime_get_ns();
4685 bfqq->ttime.last_end_request = now_ns;
4688 * Using us instead of ns, to get a reasonable precision in
4689 * computing rate in next check.
4691 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
4694 * If the request took rather long to complete, and, according
4695 * to the maximum request size recorded, this completion latency
4696 * implies that the request was certainly served at a very low
4697 * rate (less than 1M sectors/sec), then the whole observation
4698 * interval that lasts up to this time instant cannot be a
4699 * valid time interval for computing a new peak rate. Invoke
4700 * bfq_update_rate_reset to have the following three steps
4702 * - close the observation interval at the last (previous)
4703 * request dispatch or completion
4704 * - compute rate, if possible, for that observation interval
4705 * - reset to zero samples, which will trigger a proper
4706 * re-initialization of the observation interval on next
4709 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
4710 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
4711 1UL<<(BFQ_RATE_SHIFT - 10))
4712 bfq_update_rate_reset(bfqd, NULL);
4713 bfqd->last_completion = now_ns;
4716 * If we are waiting to discover whether the request pattern
4717 * of the task associated with the queue is actually
4718 * isochronous, and both requisites for this condition to hold
4719 * are now satisfied, then compute soft_rt_next_start (see the
4720 * comments on the function bfq_bfqq_softrt_next_start()). We
4721 * schedule this delayed check when bfqq expires, if it still
4722 * has in-flight requests.
4724 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
4725 RB_EMPTY_ROOT(&bfqq->sort_list))
4726 bfqq->soft_rt_next_start =
4727 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4730 * If this is the in-service queue, check if it needs to be expired,
4731 * or if we want to idle in case it has no pending requests.
4733 if (bfqd->in_service_queue == bfqq) {
4734 if (bfq_bfqq_must_idle(bfqq)) {
4735 if (bfqq->dispatched == 0)
4736 bfq_arm_slice_timer(bfqd);
4738 * If we get here, we do not expire bfqq, even
4739 * if bfqq was in budget timeout or had no
4740 * more requests (as controlled in the next
4741 * conditional instructions). The reason for
4742 * not expiring bfqq is as follows.
4744 * Here bfqq->dispatched > 0 holds, but
4745 * bfq_bfqq_must_idle() returned true. This
4746 * implies that, even if no request arrives
4747 * for bfqq before bfqq->dispatched reaches 0,
4748 * bfqq will, however, not be expired on the
4749 * completion event that causes bfqq->dispatch
4750 * to reach zero. In contrast, on this event,
4751 * bfqq will start enjoying device idling
4752 * (I/O-dispatch plugging).
4754 * But, if we expired bfqq here, bfqq would
4755 * not have the chance to enjoy device idling
4756 * when bfqq->dispatched finally reaches
4757 * zero. This would expose bfqq to violation
4758 * of its reserved service guarantees.
4761 } else if (bfq_may_expire_for_budg_timeout(bfqq))
4762 bfq_bfqq_expire(bfqd, bfqq, false,
4763 BFQQE_BUDGET_TIMEOUT);
4764 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
4765 (bfqq->dispatched == 0 ||
4766 !bfq_better_to_idle(bfqq)))
4767 bfq_bfqq_expire(bfqd, bfqq, false,
4768 BFQQE_NO_MORE_REQUESTS);
4771 if (!bfqd->rq_in_driver)
4772 bfq_schedule_dispatch(bfqd);
4775 static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq)
4779 bfq_put_queue(bfqq);
4783 * Handle either a requeue or a finish for rq. The things to do are
4784 * the same in both cases: all references to rq are to be dropped. In
4785 * particular, rq is considered completed from the point of view of
4788 static void bfq_finish_requeue_request(struct request *rq)
4790 struct bfq_queue *bfqq = RQ_BFQQ(rq);
4791 struct bfq_data *bfqd;
4794 * Requeue and finish hooks are invoked in blk-mq without
4795 * checking whether the involved request is actually still
4796 * referenced in the scheduler. To handle this fact, the
4797 * following two checks make this function exit in case of
4798 * spurious invocations, for which there is nothing to do.
4800 * First, check whether rq has nothing to do with an elevator.
4802 if (unlikely(!(rq->rq_flags & RQF_ELVPRIV)))
4806 * rq either is not associated with any icq, or is an already
4807 * requeued request that has not (yet) been re-inserted into
4810 if (!rq->elv.icq || !bfqq)
4815 if (rq->rq_flags & RQF_STARTED)
4816 bfqg_stats_update_completion(bfqq_group(bfqq),
4818 rq->io_start_time_ns,
4821 if (likely(rq->rq_flags & RQF_STARTED)) {
4822 unsigned long flags;
4824 spin_lock_irqsave(&bfqd->lock, flags);
4826 bfq_completed_request(bfqq, bfqd);
4827 bfq_finish_requeue_request_body(bfqq);
4829 spin_unlock_irqrestore(&bfqd->lock, flags);
4832 * Request rq may be still/already in the scheduler,
4833 * in which case we need to remove it (this should
4834 * never happen in case of requeue). And we cannot
4835 * defer such a check and removal, to avoid
4836 * inconsistencies in the time interval from the end
4837 * of this function to the start of the deferred work.
4838 * This situation seems to occur only in process
4839 * context, as a consequence of a merge. In the
4840 * current version of the code, this implies that the
4844 if (!RB_EMPTY_NODE(&rq->rb_node)) {
4845 bfq_remove_request(rq->q, rq);
4846 bfqg_stats_update_io_remove(bfqq_group(bfqq),
4849 bfq_finish_requeue_request_body(bfqq);
4853 * Reset private fields. In case of a requeue, this allows
4854 * this function to correctly do nothing if it is spuriously
4855 * invoked again on this same request (see the check at the
4856 * beginning of the function). Probably, a better general
4857 * design would be to prevent blk-mq from invoking the requeue
4858 * or finish hooks of an elevator, for a request that is not
4859 * referred by that elevator.
4861 * Resetting the following fields would break the
4862 * request-insertion logic if rq is re-inserted into a bfq
4863 * internal queue, without a re-preparation. Here we assume
4864 * that re-insertions of requeued requests, without
4865 * re-preparation, can happen only for pass_through or at_head
4866 * requests (which are not re-inserted into bfq internal
4869 rq->elv.priv[0] = NULL;
4870 rq->elv.priv[1] = NULL;
4874 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
4875 * was the last process referring to that bfqq.
4877 static struct bfq_queue *
4878 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
4880 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
4882 if (bfqq_process_refs(bfqq) == 1) {
4883 bfqq->pid = current->pid;
4884 bfq_clear_bfqq_coop(bfqq);
4885 bfq_clear_bfqq_split_coop(bfqq);
4889 bic_set_bfqq(bic, NULL, 1);
4891 bfq_put_cooperator(bfqq);
4893 bfq_put_queue(bfqq);
4897 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
4898 struct bfq_io_cq *bic,
4900 bool split, bool is_sync,
4903 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4905 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
4912 bfq_put_queue(bfqq);
4913 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
4915 bic_set_bfqq(bic, bfqq, is_sync);
4916 if (split && is_sync) {
4917 if ((bic->was_in_burst_list && bfqd->large_burst) ||
4918 bic->saved_in_large_burst)
4919 bfq_mark_bfqq_in_large_burst(bfqq);
4921 bfq_clear_bfqq_in_large_burst(bfqq);
4922 if (bic->was_in_burst_list)
4924 * If bfqq was in the current
4925 * burst list before being
4926 * merged, then we have to add
4927 * it back. And we do not need
4928 * to increase burst_size, as
4929 * we did not decrement
4930 * burst_size when we removed
4931 * bfqq from the burst list as
4932 * a consequence of a merge
4934 * bfq_put_queue). In this
4935 * respect, it would be rather
4936 * costly to know whether the
4937 * current burst list is still
4938 * the same burst list from
4939 * which bfqq was removed on
4940 * the merge. To avoid this
4941 * cost, if bfqq was in a
4942 * burst list, then we add
4943 * bfqq to the current burst
4944 * list without any further
4945 * check. This can cause
4946 * inappropriate insertions,
4947 * but rarely enough to not
4948 * harm the detection of large
4949 * bursts significantly.
4951 hlist_add_head(&bfqq->burst_list_node,
4954 bfqq->split_time = jiffies;
4961 * Only reset private fields. The actual request preparation will be
4962 * performed by bfq_init_rq, when rq is either inserted or merged. See
4963 * comments on bfq_init_rq for the reason behind this delayed
4966 static void bfq_prepare_request(struct request *rq, struct bio *bio)
4969 * Regardless of whether we have an icq attached, we have to
4970 * clear the scheduler pointers, as they might point to
4971 * previously allocated bic/bfqq structs.
4973 rq->elv.priv[0] = rq->elv.priv[1] = NULL;
4977 * If needed, init rq, allocate bfq data structures associated with
4978 * rq, and increment reference counters in the destination bfq_queue
4979 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
4980 * not associated with any bfq_queue.
4982 * This function is invoked by the functions that perform rq insertion
4983 * or merging. One may have expected the above preparation operations
4984 * to be performed in bfq_prepare_request, and not delayed to when rq
4985 * is inserted or merged. The rationale behind this delayed
4986 * preparation is that, after the prepare_request hook is invoked for
4987 * rq, rq may still be transformed into a request with no icq, i.e., a
4988 * request not associated with any queue. No bfq hook is invoked to
4989 * signal this tranformation. As a consequence, should these
4990 * preparation operations be performed when the prepare_request hook
4991 * is invoked, and should rq be transformed one moment later, bfq
4992 * would end up in an inconsistent state, because it would have
4993 * incremented some queue counters for an rq destined to
4994 * transformation, without any chance to correctly lower these
4995 * counters back. In contrast, no transformation can still happen for
4996 * rq after rq has been inserted or merged. So, it is safe to execute
4997 * these preparation operations when rq is finally inserted or merged.
4999 static struct bfq_queue *bfq_init_rq(struct request *rq)
5001 struct request_queue *q = rq->q;
5002 struct bio *bio = rq->bio;
5003 struct bfq_data *bfqd = q->elevator->elevator_data;
5004 struct bfq_io_cq *bic;
5005 const int is_sync = rq_is_sync(rq);
5006 struct bfq_queue *bfqq;
5007 bool new_queue = false;
5008 bool bfqq_already_existing = false, split = false;
5010 if (unlikely(!rq->elv.icq))
5014 * Assuming that elv.priv[1] is set only if everything is set
5015 * for this rq. This holds true, because this function is
5016 * invoked only for insertion or merging, and, after such
5017 * events, a request cannot be manipulated any longer before
5018 * being removed from bfq.
5020 if (rq->elv.priv[1])
5021 return rq->elv.priv[1];
5023 bic = icq_to_bic(rq->elv.icq);
5025 bfq_check_ioprio_change(bic, bio);
5027 bfq_bic_update_cgroup(bic, bio);
5029 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
5032 if (likely(!new_queue)) {
5033 /* If the queue was seeky for too long, break it apart. */
5034 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
5035 bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
5037 /* Update bic before losing reference to bfqq */
5038 if (bfq_bfqq_in_large_burst(bfqq))
5039 bic->saved_in_large_burst = true;
5041 bfqq = bfq_split_bfqq(bic, bfqq);
5045 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
5049 bfqq_already_existing = true;
5055 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
5056 rq, bfqq, bfqq->ref);
5058 rq->elv.priv[0] = bic;
5059 rq->elv.priv[1] = bfqq;
5062 * If a bfq_queue has only one process reference, it is owned
5063 * by only this bic: we can then set bfqq->bic = bic. in
5064 * addition, if the queue has also just been split, we have to
5067 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
5071 * The queue has just been split from a shared
5072 * queue: restore the idle window and the
5073 * possible weight raising period.
5075 bfq_bfqq_resume_state(bfqq, bfqd, bic,
5076 bfqq_already_existing);
5080 if (unlikely(bfq_bfqq_just_created(bfqq)))
5081 bfq_handle_burst(bfqd, bfqq);
5086 static void bfq_idle_slice_timer_body(struct bfq_queue *bfqq)
5088 struct bfq_data *bfqd = bfqq->bfqd;
5089 enum bfqq_expiration reason;
5090 unsigned long flags;
5092 spin_lock_irqsave(&bfqd->lock, flags);
5093 bfq_clear_bfqq_wait_request(bfqq);
5095 if (bfqq != bfqd->in_service_queue) {
5096 spin_unlock_irqrestore(&bfqd->lock, flags);
5100 if (bfq_bfqq_budget_timeout(bfqq))
5102 * Also here the queue can be safely expired
5103 * for budget timeout without wasting
5106 reason = BFQQE_BUDGET_TIMEOUT;
5107 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
5109 * The queue may not be empty upon timer expiration,
5110 * because we may not disable the timer when the
5111 * first request of the in-service queue arrives
5112 * during disk idling.
5114 reason = BFQQE_TOO_IDLE;
5116 goto schedule_dispatch;
5118 bfq_bfqq_expire(bfqd, bfqq, true, reason);
5121 spin_unlock_irqrestore(&bfqd->lock, flags);
5122 bfq_schedule_dispatch(bfqd);
5126 * Handler of the expiration of the timer running if the in-service queue
5127 * is idling inside its time slice.
5129 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
5131 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
5133 struct bfq_queue *bfqq = bfqd->in_service_queue;
5136 * Theoretical race here: the in-service queue can be NULL or
5137 * different from the queue that was idling if a new request
5138 * arrives for the current queue and there is a full dispatch
5139 * cycle that changes the in-service queue. This can hardly
5140 * happen, but in the worst case we just expire a queue too
5144 bfq_idle_slice_timer_body(bfqq);
5146 return HRTIMER_NORESTART;
5149 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
5150 struct bfq_queue **bfqq_ptr)
5152 struct bfq_queue *bfqq = *bfqq_ptr;
5154 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
5156 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
5158 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
5160 bfq_put_queue(bfqq);
5166 * Release all the bfqg references to its async queues. If we are
5167 * deallocating the group these queues may still contain requests, so
5168 * we reparent them to the root cgroup (i.e., the only one that will
5169 * exist for sure until all the requests on a device are gone).
5171 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
5175 for (i = 0; i < 2; i++)
5176 for (j = 0; j < IOPRIO_BE_NR; j++)
5177 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
5179 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
5183 * See the comments on bfq_limit_depth for the purpose of
5184 * the depths set in the function. Return minimum shallow depth we'll use.
5186 static unsigned int bfq_update_depths(struct bfq_data *bfqd,
5187 struct sbitmap_queue *bt)
5189 unsigned int i, j, min_shallow = UINT_MAX;
5192 * In-word depths if no bfq_queue is being weight-raised:
5193 * leaving 25% of tags only for sync reads.
5195 * In next formulas, right-shift the value
5196 * (1U<<bt->sb.shift), instead of computing directly
5197 * (1U<<(bt->sb.shift - something)), to be robust against
5198 * any possible value of bt->sb.shift, without having to
5199 * limit 'something'.
5201 /* no more than 50% of tags for async I/O */
5202 bfqd->word_depths[0][0] = max((1U << bt->sb.shift) >> 1, 1U);
5204 * no more than 75% of tags for sync writes (25% extra tags
5205 * w.r.t. async I/O, to prevent async I/O from starving sync
5208 bfqd->word_depths[0][1] = max(((1U << bt->sb.shift) * 3) >> 2, 1U);
5211 * In-word depths in case some bfq_queue is being weight-
5212 * raised: leaving ~63% of tags for sync reads. This is the
5213 * highest percentage for which, in our tests, application
5214 * start-up times didn't suffer from any regression due to tag
5217 /* no more than ~18% of tags for async I/O */
5218 bfqd->word_depths[1][0] = max(((1U << bt->sb.shift) * 3) >> 4, 1U);
5219 /* no more than ~37% of tags for sync writes (~20% extra tags) */
5220 bfqd->word_depths[1][1] = max(((1U << bt->sb.shift) * 6) >> 4, 1U);
5222 for (i = 0; i < 2; i++)
5223 for (j = 0; j < 2; j++)
5224 min_shallow = min(min_shallow, bfqd->word_depths[i][j]);
5229 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
5231 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5232 struct blk_mq_tags *tags = hctx->sched_tags;
5233 unsigned int min_shallow;
5235 min_shallow = bfq_update_depths(bfqd, &tags->bitmap_tags);
5236 sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, min_shallow);
5240 static void bfq_exit_queue(struct elevator_queue *e)
5242 struct bfq_data *bfqd = e->elevator_data;
5243 struct bfq_queue *bfqq, *n;
5245 hrtimer_cancel(&bfqd->idle_slice_timer);
5247 spin_lock_irq(&bfqd->lock);
5248 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
5249 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
5250 spin_unlock_irq(&bfqd->lock);
5252 hrtimer_cancel(&bfqd->idle_slice_timer);
5254 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5255 /* release oom-queue reference to root group */
5256 bfqg_and_blkg_put(bfqd->root_group);
5258 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
5260 spin_lock_irq(&bfqd->lock);
5261 bfq_put_async_queues(bfqd, bfqd->root_group);
5262 kfree(bfqd->root_group);
5263 spin_unlock_irq(&bfqd->lock);
5269 static void bfq_init_root_group(struct bfq_group *root_group,
5270 struct bfq_data *bfqd)
5274 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5275 root_group->entity.parent = NULL;
5276 root_group->my_entity = NULL;
5277 root_group->bfqd = bfqd;
5279 root_group->rq_pos_tree = RB_ROOT;
5280 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
5281 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
5282 root_group->sched_data.bfq_class_idle_last_service = jiffies;
5285 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
5287 struct bfq_data *bfqd;
5288 struct elevator_queue *eq;
5290 eq = elevator_alloc(q, e);
5294 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
5296 kobject_put(&eq->kobj);
5299 eq->elevator_data = bfqd;
5301 spin_lock_irq(q->queue_lock);
5303 spin_unlock_irq(q->queue_lock);
5306 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
5307 * Grab a permanent reference to it, so that the normal code flow
5308 * will not attempt to free it.
5310 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
5311 bfqd->oom_bfqq.ref++;
5312 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
5313 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
5314 bfqd->oom_bfqq.entity.new_weight =
5315 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
5317 /* oom_bfqq does not participate to bursts */
5318 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
5321 * Trigger weight initialization, according to ioprio, at the
5322 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
5323 * class won't be changed any more.
5325 bfqd->oom_bfqq.entity.prio_changed = 1;
5329 INIT_LIST_HEAD(&bfqd->dispatch);
5331 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
5333 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
5335 bfqd->queue_weights_tree = RB_ROOT;
5336 bfqd->group_weights_tree = RB_ROOT;
5338 INIT_LIST_HEAD(&bfqd->active_list);
5339 INIT_LIST_HEAD(&bfqd->idle_list);
5340 INIT_HLIST_HEAD(&bfqd->burst_list);
5344 bfqd->bfq_max_budget = bfq_default_max_budget;
5346 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
5347 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
5348 bfqd->bfq_back_max = bfq_back_max;
5349 bfqd->bfq_back_penalty = bfq_back_penalty;
5350 bfqd->bfq_slice_idle = bfq_slice_idle;
5351 bfqd->bfq_timeout = bfq_timeout;
5353 bfqd->bfq_requests_within_timer = 120;
5355 bfqd->bfq_large_burst_thresh = 8;
5356 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
5358 bfqd->low_latency = true;
5361 * Trade-off between responsiveness and fairness.
5363 bfqd->bfq_wr_coeff = 30;
5364 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
5365 bfqd->bfq_wr_max_time = 0;
5366 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
5367 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
5368 bfqd->bfq_wr_max_softrt_rate = 7000; /*
5369 * Approximate rate required
5370 * to playback or record a
5371 * high-definition compressed
5374 bfqd->wr_busy_queues = 0;
5377 * Begin by assuming, optimistically, that the device peak
5378 * rate is equal to 2/3 of the highest reference rate.
5380 bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
5381 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
5382 bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
5384 spin_lock_init(&bfqd->lock);
5387 * The invocation of the next bfq_create_group_hierarchy
5388 * function is the head of a chain of function calls
5389 * (bfq_create_group_hierarchy->blkcg_activate_policy->
5390 * blk_mq_freeze_queue) that may lead to the invocation of the
5391 * has_work hook function. For this reason,
5392 * bfq_create_group_hierarchy is invoked only after all
5393 * scheduler data has been initialized, apart from the fields
5394 * that can be initialized only after invoking
5395 * bfq_create_group_hierarchy. This, in particular, enables
5396 * has_work to correctly return false. Of course, to avoid
5397 * other inconsistencies, the blk-mq stack must then refrain
5398 * from invoking further scheduler hooks before this init
5399 * function is finished.
5401 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
5402 if (!bfqd->root_group)
5404 bfq_init_root_group(bfqd->root_group, bfqd);
5405 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
5407 wbt_disable_default(q);
5412 kobject_put(&eq->kobj);
5416 static void bfq_slab_kill(void)
5418 kmem_cache_destroy(bfq_pool);
5421 static int __init bfq_slab_setup(void)
5423 bfq_pool = KMEM_CACHE(bfq_queue, 0);
5429 static ssize_t bfq_var_show(unsigned int var, char *page)
5431 return sprintf(page, "%u\n", var);
5434 static int bfq_var_store(unsigned long *var, const char *page)
5436 unsigned long new_val;
5437 int ret = kstrtoul(page, 10, &new_val);
5445 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
5446 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
5448 struct bfq_data *bfqd = e->elevator_data; \
5449 u64 __data = __VAR; \
5451 __data = jiffies_to_msecs(__data); \
5452 else if (__CONV == 2) \
5453 __data = div_u64(__data, NSEC_PER_MSEC); \
5454 return bfq_var_show(__data, (page)); \
5456 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
5457 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
5458 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
5459 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
5460 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
5461 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
5462 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
5463 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
5464 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
5465 #undef SHOW_FUNCTION
5467 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
5468 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
5470 struct bfq_data *bfqd = e->elevator_data; \
5471 u64 __data = __VAR; \
5472 __data = div_u64(__data, NSEC_PER_USEC); \
5473 return bfq_var_show(__data, (page)); \
5475 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
5476 #undef USEC_SHOW_FUNCTION
5478 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
5480 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
5482 struct bfq_data *bfqd = e->elevator_data; \
5483 unsigned long __data, __min = (MIN), __max = (MAX); \
5486 ret = bfq_var_store(&__data, (page)); \
5489 if (__data < __min) \
5491 else if (__data > __max) \
5494 *(__PTR) = msecs_to_jiffies(__data); \
5495 else if (__CONV == 2) \
5496 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
5498 *(__PTR) = __data; \
5501 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
5503 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
5505 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
5506 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
5508 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
5509 #undef STORE_FUNCTION
5511 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
5512 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
5514 struct bfq_data *bfqd = e->elevator_data; \
5515 unsigned long __data, __min = (MIN), __max = (MAX); \
5518 ret = bfq_var_store(&__data, (page)); \
5521 if (__data < __min) \
5523 else if (__data > __max) \
5525 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
5528 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
5530 #undef USEC_STORE_FUNCTION
5532 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
5533 const char *page, size_t count)
5535 struct bfq_data *bfqd = e->elevator_data;
5536 unsigned long __data;
5539 ret = bfq_var_store(&__data, (page));
5544 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5546 if (__data > INT_MAX)
5548 bfqd->bfq_max_budget = __data;
5551 bfqd->bfq_user_max_budget = __data;
5557 * Leaving this name to preserve name compatibility with cfq
5558 * parameters, but this timeout is used for both sync and async.
5560 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
5561 const char *page, size_t count)
5563 struct bfq_data *bfqd = e->elevator_data;
5564 unsigned long __data;
5567 ret = bfq_var_store(&__data, (page));
5573 else if (__data > INT_MAX)
5576 bfqd->bfq_timeout = msecs_to_jiffies(__data);
5577 if (bfqd->bfq_user_max_budget == 0)
5578 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5583 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
5584 const char *page, size_t count)
5586 struct bfq_data *bfqd = e->elevator_data;
5587 unsigned long __data;
5590 ret = bfq_var_store(&__data, (page));
5596 if (!bfqd->strict_guarantees && __data == 1
5597 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
5598 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
5600 bfqd->strict_guarantees = __data;
5605 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
5606 const char *page, size_t count)
5608 struct bfq_data *bfqd = e->elevator_data;
5609 unsigned long __data;
5612 ret = bfq_var_store(&__data, (page));
5618 if (__data == 0 && bfqd->low_latency != 0)
5620 bfqd->low_latency = __data;
5625 #define BFQ_ATTR(name) \
5626 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
5628 static struct elv_fs_entry bfq_attrs[] = {
5629 BFQ_ATTR(fifo_expire_sync),
5630 BFQ_ATTR(fifo_expire_async),
5631 BFQ_ATTR(back_seek_max),
5632 BFQ_ATTR(back_seek_penalty),
5633 BFQ_ATTR(slice_idle),
5634 BFQ_ATTR(slice_idle_us),
5635 BFQ_ATTR(max_budget),
5636 BFQ_ATTR(timeout_sync),
5637 BFQ_ATTR(strict_guarantees),
5638 BFQ_ATTR(low_latency),
5642 static struct elevator_type iosched_bfq_mq = {
5644 .limit_depth = bfq_limit_depth,
5645 .prepare_request = bfq_prepare_request,
5646 .requeue_request = bfq_finish_requeue_request,
5647 .finish_request = bfq_finish_requeue_request,
5648 .exit_icq = bfq_exit_icq,
5649 .insert_requests = bfq_insert_requests,
5650 .dispatch_request = bfq_dispatch_request,
5651 .next_request = elv_rb_latter_request,
5652 .former_request = elv_rb_former_request,
5653 .allow_merge = bfq_allow_bio_merge,
5654 .bio_merge = bfq_bio_merge,
5655 .request_merge = bfq_request_merge,
5656 .requests_merged = bfq_requests_merged,
5657 .request_merged = bfq_request_merged,
5658 .has_work = bfq_has_work,
5659 .init_hctx = bfq_init_hctx,
5660 .init_sched = bfq_init_queue,
5661 .exit_sched = bfq_exit_queue,
5665 .icq_size = sizeof(struct bfq_io_cq),
5666 .icq_align = __alignof__(struct bfq_io_cq),
5667 .elevator_attrs = bfq_attrs,
5668 .elevator_name = "bfq",
5669 .elevator_owner = THIS_MODULE,
5671 MODULE_ALIAS("bfq-iosched");
5673 static int __init bfq_init(void)
5677 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5678 ret = blkcg_policy_register(&blkcg_policy_bfq);
5684 if (bfq_slab_setup())
5688 * Times to load large popular applications for the typical
5689 * systems installed on the reference devices (see the
5690 * comments before the definition of the next
5691 * array). Actually, we use slightly lower values, as the
5692 * estimated peak rate tends to be smaller than the actual
5693 * peak rate. The reason for this last fact is that estimates
5694 * are computed over much shorter time intervals than the long
5695 * intervals typically used for benchmarking. Why? First, to
5696 * adapt more quickly to variations. Second, because an I/O
5697 * scheduler cannot rely on a peak-rate-evaluation workload to
5698 * be run for a long time.
5700 ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
5701 ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
5703 ret = elv_register(&iosched_bfq_mq);
5712 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5713 blkcg_policy_unregister(&blkcg_policy_bfq);
5718 static void __exit bfq_exit(void)
5720 elv_unregister(&iosched_bfq_mq);
5721 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5722 blkcg_policy_unregister(&blkcg_policy_bfq);
5727 module_init(bfq_init);
5728 module_exit(bfq_exit);
5730 MODULE_AUTHOR("Paolo Valente");
5731 MODULE_LICENSE("GPL");
5732 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");