1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Budget Fair Queueing (BFQ) I/O scheduler.
5 * Based on ideas and code from CFQ:
6 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9 * Paolo Valente <paolo.valente@unimore.it>
11 * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
12 * Arianna Avanzini <avanzini@google.com>
14 * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
16 * BFQ is a proportional-share I/O scheduler, with some extra
17 * low-latency capabilities. BFQ also supports full hierarchical
18 * scheduling through cgroups. Next paragraphs provide an introduction
19 * on BFQ inner workings. Details on BFQ benefits, usage and
20 * limitations can be found in Documentation/block/bfq-iosched.rst.
22 * BFQ is a proportional-share storage-I/O scheduling algorithm based
23 * on the slice-by-slice service scheme of CFQ. But BFQ assigns
24 * budgets, measured in number of sectors, to processes instead of
25 * time slices. The device is not granted to the in-service process
26 * for a given time slice, but until it has exhausted its assigned
27 * budget. This change from the time to the service domain enables BFQ
28 * to distribute the device throughput among processes as desired,
29 * without any distortion due to throughput fluctuations, or to device
30 * internal queueing. BFQ uses an ad hoc internal scheduler, called
31 * B-WF2Q+, to schedule processes according to their budgets. More
32 * precisely, BFQ schedules queues associated with processes. Each
33 * process/queue is assigned a user-configurable weight, and B-WF2Q+
34 * guarantees that each queue receives a fraction of the throughput
35 * proportional to its weight. Thanks to the accurate policy of
36 * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
37 * processes issuing sequential requests (to boost the throughput),
38 * and yet guarantee a low latency to interactive and soft real-time
41 * In particular, to provide these low-latency guarantees, BFQ
42 * explicitly privileges the I/O of two classes of time-sensitive
43 * applications: interactive and soft real-time. In more detail, BFQ
44 * behaves this way if the low_latency parameter is set (default
45 * configuration). This feature enables BFQ to provide applications in
46 * these classes with a very low latency.
48 * To implement this feature, BFQ constantly tries to detect whether
49 * the I/O requests in a bfq_queue come from an interactive or a soft
50 * real-time application. For brevity, in these cases, the queue is
51 * said to be interactive or soft real-time. In both cases, BFQ
52 * privileges the service of the queue, over that of non-interactive
53 * and non-soft-real-time queues. This privileging is performed,
54 * mainly, by raising the weight of the queue. So, for brevity, we
55 * call just weight-raising periods the time periods during which a
56 * queue is privileged, because deemed interactive or soft real-time.
58 * The detection of soft real-time queues/applications is described in
59 * detail in the comments on the function
60 * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
61 * interactive queue works as follows: a queue is deemed interactive
62 * if it is constantly non empty only for a limited time interval,
63 * after which it does become empty. The queue may be deemed
64 * interactive again (for a limited time), if it restarts being
65 * constantly non empty, provided that this happens only after the
66 * queue has remained empty for a given minimum idle time.
68 * By default, BFQ computes automatically the above maximum time
69 * interval, i.e., the time interval after which a constantly
70 * non-empty queue stops being deemed interactive. Since a queue is
71 * weight-raised while it is deemed interactive, this maximum time
72 * interval happens to coincide with the (maximum) duration of the
73 * weight-raising for interactive queues.
75 * Finally, BFQ also features additional heuristics for
76 * preserving both a low latency and a high throughput on NCQ-capable,
77 * rotational or flash-based devices, and to get the job done quickly
78 * for applications consisting in many I/O-bound processes.
80 * NOTE: if the main or only goal, with a given device, is to achieve
81 * the maximum-possible throughput at all times, then do switch off
82 * all low-latency heuristics for that device, by setting low_latency
85 * BFQ is described in [1], where also a reference to the initial,
86 * more theoretical paper on BFQ can be found. The interested reader
87 * can find in the latter paper full details on the main algorithm, as
88 * well as formulas of the guarantees and formal proofs of all the
89 * properties. With respect to the version of BFQ presented in these
90 * papers, this implementation adds a few more heuristics, such as the
91 * ones that guarantee a low latency to interactive and soft real-time
92 * applications, and a hierarchical extension based on H-WF2Q+.
94 * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
95 * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
96 * with O(log N) complexity derives from the one introduced with EEVDF
99 * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
100 * Scheduler", Proceedings of the First Workshop on Mobile System
101 * Technologies (MST-2015), May 2015.
102 * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
104 * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
105 * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
108 * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
110 * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
111 * First: A Flexible and Accurate Mechanism for Proportional Share
112 * Resource Allocation", technical report.
114 * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
116 #include <linux/module.h>
117 #include <linux/slab.h>
118 #include <linux/blkdev.h>
119 #include <linux/cgroup.h>
120 #include <linux/elevator.h>
121 #include <linux/ktime.h>
122 #include <linux/rbtree.h>
123 #include <linux/ioprio.h>
124 #include <linux/sbitmap.h>
125 #include <linux/delay.h>
126 #include <linux/backing-dev.h>
130 #include "blk-mq-tag.h"
131 #include "blk-mq-sched.h"
132 #include "bfq-iosched.h"
135 #define BFQ_BFQQ_FNS(name) \
136 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
138 __set_bit(BFQQF_##name, &(bfqq)->flags); \
140 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
142 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
144 int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
146 return test_bit(BFQQF_##name, &(bfqq)->flags); \
149 BFQ_BFQQ_FNS(just_created);
151 BFQ_BFQQ_FNS(wait_request);
152 BFQ_BFQQ_FNS(non_blocking_wait_rq);
153 BFQ_BFQQ_FNS(fifo_expire);
154 BFQ_BFQQ_FNS(has_short_ttime);
156 BFQ_BFQQ_FNS(IO_bound);
157 BFQ_BFQQ_FNS(in_large_burst);
159 BFQ_BFQQ_FNS(split_coop);
160 BFQ_BFQQ_FNS(softrt_update);
161 BFQ_BFQQ_FNS(has_waker);
162 #undef BFQ_BFQQ_FNS \
164 /* Expiration time of sync (0) and async (1) requests, in ns. */
165 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
167 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
168 static const int bfq_back_max = 16 * 1024;
170 /* Penalty of a backwards seek, in number of sectors. */
171 static const int bfq_back_penalty = 2;
173 /* Idling period duration, in ns. */
174 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
176 /* Minimum number of assigned budgets for which stats are safe to compute. */
177 static const int bfq_stats_min_budgets = 194;
179 /* Default maximum budget values, in sectors and number of requests. */
180 static const int bfq_default_max_budget = 16 * 1024;
183 * When a sync request is dispatched, the queue that contains that
184 * request, and all the ancestor entities of that queue, are charged
185 * with the number of sectors of the request. In contrast, if the
186 * request is async, then the queue and its ancestor entities are
187 * charged with the number of sectors of the request, multiplied by
188 * the factor below. This throttles the bandwidth for async I/O,
189 * w.r.t. to sync I/O, and it is done to counter the tendency of async
190 * writes to steal I/O throughput to reads.
192 * The current value of this parameter is the result of a tuning with
193 * several hardware and software configurations. We tried to find the
194 * lowest value for which writes do not cause noticeable problems to
195 * reads. In fact, the lower this parameter, the stabler I/O control,
196 * in the following respect. The lower this parameter is, the less
197 * the bandwidth enjoyed by a group decreases
198 * - when the group does writes, w.r.t. to when it does reads;
199 * - when other groups do reads, w.r.t. to when they do writes.
201 static const int bfq_async_charge_factor = 3;
203 /* Default timeout values, in jiffies, approximating CFQ defaults. */
204 const int bfq_timeout = HZ / 8;
207 * Time limit for merging (see comments in bfq_setup_cooperator). Set
208 * to the slowest value that, in our tests, proved to be effective in
209 * removing false positives, while not causing true positives to miss
212 * As can be deduced from the low time limit below, queue merging, if
213 * successful, happens at the very beginning of the I/O of the involved
214 * cooperating processes, as a consequence of the arrival of the very
215 * first requests from each cooperator. After that, there is very
216 * little chance to find cooperators.
218 static const unsigned long bfq_merge_time_limit = HZ/10;
220 static struct kmem_cache *bfq_pool;
222 /* Below this threshold (in ns), we consider thinktime immediate. */
223 #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
225 /* hw_tag detection: parallel requests threshold and min samples needed. */
226 #define BFQ_HW_QUEUE_THRESHOLD 3
227 #define BFQ_HW_QUEUE_SAMPLES 32
229 #define BFQQ_SEEK_THR (sector_t)(8 * 100)
230 #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
231 #define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \
232 (get_sdist(last_pos, rq) > \
234 (!blk_queue_nonrot(bfqd->queue) || \
235 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT))
236 #define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
237 #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19)
239 * Sync random I/O is likely to be confused with soft real-time I/O,
240 * because it is characterized by limited throughput and apparently
241 * isochronous arrival pattern. To avoid false positives, queues
242 * containing only random (seeky) I/O are prevented from being tagged
245 #define BFQQ_TOTALLY_SEEKY(bfqq) (bfqq->seek_history == -1)
247 /* Min number of samples required to perform peak-rate update */
248 #define BFQ_RATE_MIN_SAMPLES 32
249 /* Min observation time interval required to perform a peak-rate update (ns) */
250 #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
251 /* Target observation time interval for a peak-rate update (ns) */
252 #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
255 * Shift used for peak-rate fixed precision calculations.
257 * - the current shift: 16 positions
258 * - the current type used to store rate: u32
259 * - the current unit of measure for rate: [sectors/usec], or, more precisely,
260 * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
261 * the range of rates that can be stored is
262 * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
263 * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
264 * [15, 65G] sectors/sec
265 * Which, assuming a sector size of 512B, corresponds to a range of
268 #define BFQ_RATE_SHIFT 16
271 * When configured for computing the duration of the weight-raising
272 * for interactive queues automatically (see the comments at the
273 * beginning of this file), BFQ does it using the following formula:
274 * duration = (ref_rate / r) * ref_wr_duration,
275 * where r is the peak rate of the device, and ref_rate and
276 * ref_wr_duration are two reference parameters. In particular,
277 * ref_rate is the peak rate of the reference storage device (see
278 * below), and ref_wr_duration is about the maximum time needed, with
279 * BFQ and while reading two files in parallel, to load typical large
280 * applications on the reference device (see the comments on
281 * max_service_from_wr below, for more details on how ref_wr_duration
282 * is obtained). In practice, the slower/faster the device at hand
283 * is, the more/less it takes to load applications with respect to the
284 * reference device. Accordingly, the longer/shorter BFQ grants
285 * weight raising to interactive applications.
287 * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
288 * depending on whether the device is rotational or non-rotational.
290 * In the following definitions, ref_rate[0] and ref_wr_duration[0]
291 * are the reference values for a rotational device, whereas
292 * ref_rate[1] and ref_wr_duration[1] are the reference values for a
293 * non-rotational device. The reference rates are not the actual peak
294 * rates of the devices used as a reference, but slightly lower
295 * values. The reason for using slightly lower values is that the
296 * peak-rate estimator tends to yield slightly lower values than the
297 * actual peak rate (it can yield the actual peak rate only if there
298 * is only one process doing I/O, and the process does sequential
301 * The reference peak rates are measured in sectors/usec, left-shifted
304 static int ref_rate[2] = {14000, 33000};
306 * To improve readability, a conversion function is used to initialize
307 * the following array, which entails that the array can be
308 * initialized only in a function.
310 static int ref_wr_duration[2];
313 * BFQ uses the above-detailed, time-based weight-raising mechanism to
314 * privilege interactive tasks. This mechanism is vulnerable to the
315 * following false positives: I/O-bound applications that will go on
316 * doing I/O for much longer than the duration of weight
317 * raising. These applications have basically no benefit from being
318 * weight-raised at the beginning of their I/O. On the opposite end,
319 * while being weight-raised, these applications
320 * a) unjustly steal throughput to applications that may actually need
322 * b) make BFQ uselessly perform device idling; device idling results
323 * in loss of device throughput with most flash-based storage, and may
324 * increase latencies when used purposelessly.
326 * BFQ tries to reduce these problems, by adopting the following
327 * countermeasure. To introduce this countermeasure, we need first to
328 * finish explaining how the duration of weight-raising for
329 * interactive tasks is computed.
331 * For a bfq_queue deemed as interactive, the duration of weight
332 * raising is dynamically adjusted, as a function of the estimated
333 * peak rate of the device, so as to be equal to the time needed to
334 * execute the 'largest' interactive task we benchmarked so far. By
335 * largest task, we mean the task for which each involved process has
336 * to do more I/O than for any of the other tasks we benchmarked. This
337 * reference interactive task is the start-up of LibreOffice Writer,
338 * and in this task each process/bfq_queue needs to have at most ~110K
339 * sectors transferred.
341 * This last piece of information enables BFQ to reduce the actual
342 * duration of weight-raising for at least one class of I/O-bound
343 * applications: those doing sequential or quasi-sequential I/O. An
344 * example is file copy. In fact, once started, the main I/O-bound
345 * processes of these applications usually consume the above 110K
346 * sectors in much less time than the processes of an application that
347 * is starting, because these I/O-bound processes will greedily devote
348 * almost all their CPU cycles only to their target,
349 * throughput-friendly I/O operations. This is even more true if BFQ
350 * happens to be underestimating the device peak rate, and thus
351 * overestimating the duration of weight raising. But, according to
352 * our measurements, once transferred 110K sectors, these processes
353 * have no right to be weight-raised any longer.
355 * Basing on the last consideration, BFQ ends weight-raising for a
356 * bfq_queue if the latter happens to have received an amount of
357 * service at least equal to the following constant. The constant is
358 * set to slightly more than 110K, to have a minimum safety margin.
360 * This early ending of weight-raising reduces the amount of time
361 * during which interactive false positives cause the two problems
362 * described at the beginning of these comments.
364 static const unsigned long max_service_from_wr = 120000;
366 #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
367 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
369 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
371 return bic->bfqq[is_sync];
374 void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
376 bic->bfqq[is_sync] = bfqq;
379 struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
381 return bic->icq.q->elevator->elevator_data;
385 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
386 * @icq: the iocontext queue.
388 static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
390 /* bic->icq is the first member, %NULL will convert to %NULL */
391 return container_of(icq, struct bfq_io_cq, icq);
395 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
396 * @bfqd: the lookup key.
397 * @ioc: the io_context of the process doing I/O.
398 * @q: the request queue.
400 static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd,
401 struct io_context *ioc,
402 struct request_queue *q)
406 struct bfq_io_cq *icq;
408 spin_lock_irqsave(&q->queue_lock, flags);
409 icq = icq_to_bic(ioc_lookup_icq(ioc, q));
410 spin_unlock_irqrestore(&q->queue_lock, flags);
419 * Scheduler run of queue, if there are requests pending and no one in the
420 * driver that will restart queueing.
422 void bfq_schedule_dispatch(struct bfq_data *bfqd)
424 if (bfqd->queued != 0) {
425 bfq_log(bfqd, "schedule dispatch");
426 blk_mq_run_hw_queues(bfqd->queue, true);
430 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
432 #define bfq_sample_valid(samples) ((samples) > 80)
435 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
436 * We choose the request that is closer to the head right now. Distance
437 * behind the head is penalized and only allowed to a certain extent.
439 static struct request *bfq_choose_req(struct bfq_data *bfqd,
444 sector_t s1, s2, d1 = 0, d2 = 0;
445 unsigned long back_max;
446 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
447 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
448 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
450 if (!rq1 || rq1 == rq2)
455 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
457 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
459 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
461 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
464 s1 = blk_rq_pos(rq1);
465 s2 = blk_rq_pos(rq2);
468 * By definition, 1KiB is 2 sectors.
470 back_max = bfqd->bfq_back_max * 2;
473 * Strict one way elevator _except_ in the case where we allow
474 * short backward seeks which are biased as twice the cost of a
475 * similar forward seek.
479 else if (s1 + back_max >= last)
480 d1 = (last - s1) * bfqd->bfq_back_penalty;
482 wrap |= BFQ_RQ1_WRAP;
486 else if (s2 + back_max >= last)
487 d2 = (last - s2) * bfqd->bfq_back_penalty;
489 wrap |= BFQ_RQ2_WRAP;
491 /* Found required data */
494 * By doing switch() on the bit mask "wrap" we avoid having to
495 * check two variables for all permutations: --> faster!
498 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
513 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
516 * Since both rqs are wrapped,
517 * start with the one that's further behind head
518 * (--> only *one* back seek required),
519 * since back seek takes more time than forward.
529 * Async I/O can easily starve sync I/O (both sync reads and sync
530 * writes), by consuming all tags. Similarly, storms of sync writes,
531 * such as those that sync(2) may trigger, can starve sync reads.
532 * Limit depths of async I/O and sync writes so as to counter both
535 static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
537 struct bfq_data *bfqd = data->q->elevator->elevator_data;
539 if (op_is_sync(op) && !op_is_write(op))
542 data->shallow_depth =
543 bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)];
545 bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
546 __func__, bfqd->wr_busy_queues, op_is_sync(op),
547 data->shallow_depth);
550 static struct bfq_queue *
551 bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
552 sector_t sector, struct rb_node **ret_parent,
553 struct rb_node ***rb_link)
555 struct rb_node **p, *parent;
556 struct bfq_queue *bfqq = NULL;
564 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
567 * Sort strictly based on sector. Smallest to the left,
568 * largest to the right.
570 if (sector > blk_rq_pos(bfqq->next_rq))
572 else if (sector < blk_rq_pos(bfqq->next_rq))
580 *ret_parent = parent;
584 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
585 (unsigned long long)sector,
586 bfqq ? bfqq->pid : 0);
591 static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
593 return bfqq->service_from_backlogged > 0 &&
594 time_is_before_jiffies(bfqq->first_IO_time +
595 bfq_merge_time_limit);
599 * The following function is not marked as __cold because it is
600 * actually cold, but for the same performance goal described in the
601 * comments on the likely() at the beginning of
602 * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
603 * execution time for the case where this function is not invoked, we
604 * had to add an unlikely() in each involved if().
607 bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
609 struct rb_node **p, *parent;
610 struct bfq_queue *__bfqq;
612 if (bfqq->pos_root) {
613 rb_erase(&bfqq->pos_node, bfqq->pos_root);
614 bfqq->pos_root = NULL;
617 /* oom_bfqq does not participate in queue merging */
618 if (bfqq == &bfqd->oom_bfqq)
622 * bfqq cannot be merged any longer (see comments in
623 * bfq_setup_cooperator): no point in adding bfqq into the
626 if (bfq_too_late_for_merging(bfqq))
629 if (bfq_class_idle(bfqq))
634 bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
635 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
636 blk_rq_pos(bfqq->next_rq), &parent, &p);
638 rb_link_node(&bfqq->pos_node, parent, p);
639 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
641 bfqq->pos_root = NULL;
645 * The following function returns false either if every active queue
646 * must receive the same share of the throughput (symmetric scenario),
647 * or, as a special case, if bfqq must receive a share of the
648 * throughput lower than or equal to the share that every other active
649 * queue must receive. If bfqq does sync I/O, then these are the only
650 * two cases where bfqq happens to be guaranteed its share of the
651 * throughput even if I/O dispatching is not plugged when bfqq remains
652 * temporarily empty (for more details, see the comments in the
653 * function bfq_better_to_idle()). For this reason, the return value
654 * of this function is used to check whether I/O-dispatch plugging can
657 * The above first case (symmetric scenario) occurs when:
658 * 1) all active queues have the same weight,
659 * 2) all active queues belong to the same I/O-priority class,
660 * 3) all active groups at the same level in the groups tree have the same
662 * 4) all active groups at the same level in the groups tree have the same
663 * number of children.
665 * Unfortunately, keeping the necessary state for evaluating exactly
666 * the last two symmetry sub-conditions above would be quite complex
667 * and time consuming. Therefore this function evaluates, instead,
668 * only the following stronger three sub-conditions, for which it is
669 * much easier to maintain the needed state:
670 * 1) all active queues have the same weight,
671 * 2) all active queues belong to the same I/O-priority class,
672 * 3) there are no active groups.
673 * In particular, the last condition is always true if hierarchical
674 * support or the cgroups interface are not enabled, thus no state
675 * needs to be maintained in this case.
677 static bool bfq_asymmetric_scenario(struct bfq_data *bfqd,
678 struct bfq_queue *bfqq)
680 bool smallest_weight = bfqq &&
681 bfqq->weight_counter &&
682 bfqq->weight_counter ==
684 rb_first_cached(&bfqd->queue_weights_tree),
685 struct bfq_weight_counter,
689 * For queue weights to differ, queue_weights_tree must contain
690 * at least two nodes.
692 bool varied_queue_weights = !smallest_weight &&
693 !RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) &&
694 (bfqd->queue_weights_tree.rb_root.rb_node->rb_left ||
695 bfqd->queue_weights_tree.rb_root.rb_node->rb_right);
697 bool multiple_classes_busy =
698 (bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
699 (bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
700 (bfqd->busy_queues[1] && bfqd->busy_queues[2]);
702 return varied_queue_weights || multiple_classes_busy
703 #ifdef CONFIG_BFQ_GROUP_IOSCHED
704 || bfqd->num_groups_with_pending_reqs > 0
710 * If the weight-counter tree passed as input contains no counter for
711 * the weight of the input queue, then add that counter; otherwise just
712 * increment the existing counter.
714 * Note that weight-counter trees contain few nodes in mostly symmetric
715 * scenarios. For example, if all queues have the same weight, then the
716 * weight-counter tree for the queues may contain at most one node.
717 * This holds even if low_latency is on, because weight-raised queues
718 * are not inserted in the tree.
719 * In most scenarios, the rate at which nodes are created/destroyed
722 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq,
723 struct rb_root_cached *root)
725 struct bfq_entity *entity = &bfqq->entity;
726 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
727 bool leftmost = true;
730 * Do not insert if the queue is already associated with a
731 * counter, which happens if:
732 * 1) a request arrival has caused the queue to become both
733 * non-weight-raised, and hence change its weight, and
734 * backlogged; in this respect, each of the two events
735 * causes an invocation of this function,
736 * 2) this is the invocation of this function caused by the
737 * second event. This second invocation is actually useless,
738 * and we handle this fact by exiting immediately. More
739 * efficient or clearer solutions might possibly be adopted.
741 if (bfqq->weight_counter)
745 struct bfq_weight_counter *__counter = container_of(*new,
746 struct bfq_weight_counter,
750 if (entity->weight == __counter->weight) {
751 bfqq->weight_counter = __counter;
754 if (entity->weight < __counter->weight)
755 new = &((*new)->rb_left);
757 new = &((*new)->rb_right);
762 bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
766 * In the unlucky event of an allocation failure, we just
767 * exit. This will cause the weight of queue to not be
768 * considered in bfq_asymmetric_scenario, which, in its turn,
769 * causes the scenario to be deemed wrongly symmetric in case
770 * bfqq's weight would have been the only weight making the
771 * scenario asymmetric. On the bright side, no unbalance will
772 * however occur when bfqq becomes inactive again (the
773 * invocation of this function is triggered by an activation
774 * of queue). In fact, bfq_weights_tree_remove does nothing
775 * if !bfqq->weight_counter.
777 if (unlikely(!bfqq->weight_counter))
780 bfqq->weight_counter->weight = entity->weight;
781 rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
782 rb_insert_color_cached(&bfqq->weight_counter->weights_node, root,
786 bfqq->weight_counter->num_active++;
791 * Decrement the weight counter associated with the queue, and, if the
792 * counter reaches 0, remove the counter from the tree.
793 * See the comments to the function bfq_weights_tree_add() for considerations
796 void __bfq_weights_tree_remove(struct bfq_data *bfqd,
797 struct bfq_queue *bfqq,
798 struct rb_root_cached *root)
800 if (!bfqq->weight_counter)
803 bfqq->weight_counter->num_active--;
804 if (bfqq->weight_counter->num_active > 0)
805 goto reset_entity_pointer;
807 rb_erase_cached(&bfqq->weight_counter->weights_node, root);
808 kfree(bfqq->weight_counter);
810 reset_entity_pointer:
811 bfqq->weight_counter = NULL;
816 * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
817 * of active groups for each queue's inactive parent entity.
819 void bfq_weights_tree_remove(struct bfq_data *bfqd,
820 struct bfq_queue *bfqq)
822 struct bfq_entity *entity = bfqq->entity.parent;
824 for_each_entity(entity) {
825 struct bfq_sched_data *sd = entity->my_sched_data;
827 if (sd->next_in_service || sd->in_service_entity) {
829 * entity is still active, because either
830 * next_in_service or in_service_entity is not
831 * NULL (see the comments on the definition of
832 * next_in_service for details on why
833 * in_service_entity must be checked too).
835 * As a consequence, its parent entities are
836 * active as well, and thus this loop must
843 * The decrement of num_groups_with_pending_reqs is
844 * not performed immediately upon the deactivation of
845 * entity, but it is delayed to when it also happens
846 * that the first leaf descendant bfqq of entity gets
847 * all its pending requests completed. The following
848 * instructions perform this delayed decrement, if
849 * needed. See the comments on
850 * num_groups_with_pending_reqs for details.
852 if (entity->in_groups_with_pending_reqs) {
853 entity->in_groups_with_pending_reqs = false;
854 bfqd->num_groups_with_pending_reqs--;
859 * Next function is invoked last, because it causes bfqq to be
860 * freed if the following holds: bfqq is not in service and
861 * has no dispatched request. DO NOT use bfqq after the next
862 * function invocation.
864 __bfq_weights_tree_remove(bfqd, bfqq,
865 &bfqd->queue_weights_tree);
869 * Return expired entry, or NULL to just start from scratch in rbtree.
871 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
872 struct request *last)
876 if (bfq_bfqq_fifo_expire(bfqq))
879 bfq_mark_bfqq_fifo_expire(bfqq);
881 rq = rq_entry_fifo(bfqq->fifo.next);
883 if (rq == last || ktime_get_ns() < rq->fifo_time)
886 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
890 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
891 struct bfq_queue *bfqq,
892 struct request *last)
894 struct rb_node *rbnext = rb_next(&last->rb_node);
895 struct rb_node *rbprev = rb_prev(&last->rb_node);
896 struct request *next, *prev = NULL;
898 /* Follow expired path, else get first next available. */
899 next = bfq_check_fifo(bfqq, last);
904 prev = rb_entry_rq(rbprev);
907 next = rb_entry_rq(rbnext);
909 rbnext = rb_first(&bfqq->sort_list);
910 if (rbnext && rbnext != &last->rb_node)
911 next = rb_entry_rq(rbnext);
914 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
917 /* see the definition of bfq_async_charge_factor for details */
918 static unsigned long bfq_serv_to_charge(struct request *rq,
919 struct bfq_queue *bfqq)
921 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
922 bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
923 return blk_rq_sectors(rq);
925 return blk_rq_sectors(rq) * bfq_async_charge_factor;
929 * bfq_updated_next_req - update the queue after a new next_rq selection.
930 * @bfqd: the device data the queue belongs to.
931 * @bfqq: the queue to update.
933 * If the first request of a queue changes we make sure that the queue
934 * has enough budget to serve at least its first request (if the
935 * request has grown). We do this because if the queue has not enough
936 * budget for its first request, it has to go through two dispatch
937 * rounds to actually get it dispatched.
939 static void bfq_updated_next_req(struct bfq_data *bfqd,
940 struct bfq_queue *bfqq)
942 struct bfq_entity *entity = &bfqq->entity;
943 struct request *next_rq = bfqq->next_rq;
944 unsigned long new_budget;
949 if (bfqq == bfqd->in_service_queue)
951 * In order not to break guarantees, budgets cannot be
952 * changed after an entity has been selected.
956 new_budget = max_t(unsigned long,
957 max_t(unsigned long, bfqq->max_budget,
958 bfq_serv_to_charge(next_rq, bfqq)),
960 if (entity->budget != new_budget) {
961 entity->budget = new_budget;
962 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
964 bfq_requeue_bfqq(bfqd, bfqq, false);
968 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
972 if (bfqd->bfq_wr_max_time > 0)
973 return bfqd->bfq_wr_max_time;
975 dur = bfqd->rate_dur_prod;
976 do_div(dur, bfqd->peak_rate);
979 * Limit duration between 3 and 25 seconds. The upper limit
980 * has been conservatively set after the following worst case:
981 * on a QEMU/KVM virtual machine
982 * - running in a slow PC
983 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
984 * - serving a heavy I/O workload, such as the sequential reading
986 * mplayer took 23 seconds to start, if constantly weight-raised.
988 * As for higher values than that accommodating the above bad
989 * scenario, tests show that higher values would often yield
990 * the opposite of the desired result, i.e., would worsen
991 * responsiveness by allowing non-interactive applications to
992 * preserve weight raising for too long.
994 * On the other end, lower values than 3 seconds make it
995 * difficult for most interactive tasks to complete their jobs
996 * before weight-raising finishes.
998 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
1001 /* switch back from soft real-time to interactive weight raising */
1002 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
1003 struct bfq_data *bfqd)
1005 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1006 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1007 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
1011 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
1012 struct bfq_io_cq *bic, bool bfq_already_existing)
1014 unsigned int old_wr_coeff = bfqq->wr_coeff;
1015 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
1017 if (bic->saved_has_short_ttime)
1018 bfq_mark_bfqq_has_short_ttime(bfqq);
1020 bfq_clear_bfqq_has_short_ttime(bfqq);
1022 if (bic->saved_IO_bound)
1023 bfq_mark_bfqq_IO_bound(bfqq);
1025 bfq_clear_bfqq_IO_bound(bfqq);
1027 bfqq->entity.new_weight = bic->saved_weight;
1028 bfqq->ttime = bic->saved_ttime;
1029 bfqq->wr_coeff = bic->saved_wr_coeff;
1030 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
1031 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
1032 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
1034 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1035 time_is_before_jiffies(bfqq->last_wr_start_finish +
1036 bfqq->wr_cur_max_time))) {
1037 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1038 !bfq_bfqq_in_large_burst(bfqq) &&
1039 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1040 bfq_wr_duration(bfqd))) {
1041 switch_back_to_interactive_wr(bfqq, bfqd);
1044 bfq_log_bfqq(bfqq->bfqd, bfqq,
1045 "resume state: switching off wr");
1049 /* make sure weight will be updated, however we got here */
1050 bfqq->entity.prio_changed = 1;
1055 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1056 bfqd->wr_busy_queues++;
1057 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1058 bfqd->wr_busy_queues--;
1061 static int bfqq_process_refs(struct bfq_queue *bfqq)
1063 return bfqq->ref - bfqq->allocated - bfqq->entity.on_st_or_in_serv -
1064 (bfqq->weight_counter != NULL);
1067 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1068 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1070 struct bfq_queue *item;
1071 struct hlist_node *n;
1073 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1074 hlist_del_init(&item->burst_list_node);
1077 * Start the creation of a new burst list only if there is no
1078 * active queue. See comments on the conditional invocation of
1079 * bfq_handle_burst().
1081 if (bfq_tot_busy_queues(bfqd) == 0) {
1082 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1083 bfqd->burst_size = 1;
1085 bfqd->burst_size = 0;
1087 bfqd->burst_parent_entity = bfqq->entity.parent;
1090 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1091 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1093 /* Increment burst size to take into account also bfqq */
1096 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1097 struct bfq_queue *pos, *bfqq_item;
1098 struct hlist_node *n;
1101 * Enough queues have been activated shortly after each
1102 * other to consider this burst as large.
1104 bfqd->large_burst = true;
1107 * We can now mark all queues in the burst list as
1108 * belonging to a large burst.
1110 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1112 bfq_mark_bfqq_in_large_burst(bfqq_item);
1113 bfq_mark_bfqq_in_large_burst(bfqq);
1116 * From now on, and until the current burst finishes, any
1117 * new queue being activated shortly after the last queue
1118 * was inserted in the burst can be immediately marked as
1119 * belonging to a large burst. So the burst list is not
1120 * needed any more. Remove it.
1122 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1124 hlist_del_init(&pos->burst_list_node);
1126 * Burst not yet large: add bfqq to the burst list. Do
1127 * not increment the ref counter for bfqq, because bfqq
1128 * is removed from the burst list before freeing bfqq
1131 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1135 * If many queues belonging to the same group happen to be created
1136 * shortly after each other, then the processes associated with these
1137 * queues have typically a common goal. In particular, bursts of queue
1138 * creations are usually caused by services or applications that spawn
1139 * many parallel threads/processes. Examples are systemd during boot,
1140 * or git grep. To help these processes get their job done as soon as
1141 * possible, it is usually better to not grant either weight-raising
1142 * or device idling to their queues, unless these queues must be
1143 * protected from the I/O flowing through other active queues.
1145 * In this comment we describe, firstly, the reasons why this fact
1146 * holds, and, secondly, the next function, which implements the main
1147 * steps needed to properly mark these queues so that they can then be
1148 * treated in a different way.
1150 * The above services or applications benefit mostly from a high
1151 * throughput: the quicker the requests of the activated queues are
1152 * cumulatively served, the sooner the target job of these queues gets
1153 * completed. As a consequence, weight-raising any of these queues,
1154 * which also implies idling the device for it, is almost always
1155 * counterproductive, unless there are other active queues to isolate
1156 * these new queues from. If there no other active queues, then
1157 * weight-raising these new queues just lowers throughput in most
1160 * On the other hand, a burst of queue creations may be caused also by
1161 * the start of an application that does not consist of a lot of
1162 * parallel I/O-bound threads. In fact, with a complex application,
1163 * several short processes may need to be executed to start-up the
1164 * application. In this respect, to start an application as quickly as
1165 * possible, the best thing to do is in any case to privilege the I/O
1166 * related to the application with respect to all other
1167 * I/O. Therefore, the best strategy to start as quickly as possible
1168 * an application that causes a burst of queue creations is to
1169 * weight-raise all the queues created during the burst. This is the
1170 * exact opposite of the best strategy for the other type of bursts.
1172 * In the end, to take the best action for each of the two cases, the
1173 * two types of bursts need to be distinguished. Fortunately, this
1174 * seems relatively easy, by looking at the sizes of the bursts. In
1175 * particular, we found a threshold such that only bursts with a
1176 * larger size than that threshold are apparently caused by
1177 * services or commands such as systemd or git grep. For brevity,
1178 * hereafter we call just 'large' these bursts. BFQ *does not*
1179 * weight-raise queues whose creation occurs in a large burst. In
1180 * addition, for each of these queues BFQ performs or does not perform
1181 * idling depending on which choice boosts the throughput more. The
1182 * exact choice depends on the device and request pattern at
1185 * Unfortunately, false positives may occur while an interactive task
1186 * is starting (e.g., an application is being started). The
1187 * consequence is that the queues associated with the task do not
1188 * enjoy weight raising as expected. Fortunately these false positives
1189 * are very rare. They typically occur if some service happens to
1190 * start doing I/O exactly when the interactive task starts.
1192 * Turning back to the next function, it is invoked only if there are
1193 * no active queues (apart from active queues that would belong to the
1194 * same, possible burst bfqq would belong to), and it implements all
1195 * the steps needed to detect the occurrence of a large burst and to
1196 * properly mark all the queues belonging to it (so that they can then
1197 * be treated in a different way). This goal is achieved by
1198 * maintaining a "burst list" that holds, temporarily, the queues that
1199 * belong to the burst in progress. The list is then used to mark
1200 * these queues as belonging to a large burst if the burst does become
1201 * large. The main steps are the following.
1203 * . when the very first queue is created, the queue is inserted into the
1204 * list (as it could be the first queue in a possible burst)
1206 * . if the current burst has not yet become large, and a queue Q that does
1207 * not yet belong to the burst is activated shortly after the last time
1208 * at which a new queue entered the burst list, then the function appends
1209 * Q to the burst list
1211 * . if, as a consequence of the previous step, the burst size reaches
1212 * the large-burst threshold, then
1214 * . all the queues in the burst list are marked as belonging to a
1217 * . the burst list is deleted; in fact, the burst list already served
1218 * its purpose (keeping temporarily track of the queues in a burst,
1219 * so as to be able to mark them as belonging to a large burst in the
1220 * previous sub-step), and now is not needed any more
1222 * . the device enters a large-burst mode
1224 * . if a queue Q that does not belong to the burst is created while
1225 * the device is in large-burst mode and shortly after the last time
1226 * at which a queue either entered the burst list or was marked as
1227 * belonging to the current large burst, then Q is immediately marked
1228 * as belonging to a large burst.
1230 * . if a queue Q that does not belong to the burst is created a while
1231 * later, i.e., not shortly after, than the last time at which a queue
1232 * either entered the burst list or was marked as belonging to the
1233 * current large burst, then the current burst is deemed as finished and:
1235 * . the large-burst mode is reset if set
1237 * . the burst list is emptied
1239 * . Q is inserted in the burst list, as Q may be the first queue
1240 * in a possible new burst (then the burst list contains just Q
1243 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1246 * If bfqq is already in the burst list or is part of a large
1247 * burst, or finally has just been split, then there is
1248 * nothing else to do.
1250 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1251 bfq_bfqq_in_large_burst(bfqq) ||
1252 time_is_after_eq_jiffies(bfqq->split_time +
1253 msecs_to_jiffies(10)))
1257 * If bfqq's creation happens late enough, or bfqq belongs to
1258 * a different group than the burst group, then the current
1259 * burst is finished, and related data structures must be
1262 * In this respect, consider the special case where bfqq is
1263 * the very first queue created after BFQ is selected for this
1264 * device. In this case, last_ins_in_burst and
1265 * burst_parent_entity are not yet significant when we get
1266 * here. But it is easy to verify that, whether or not the
1267 * following condition is true, bfqq will end up being
1268 * inserted into the burst list. In particular the list will
1269 * happen to contain only bfqq. And this is exactly what has
1270 * to happen, as bfqq may be the first queue of the first
1273 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1274 bfqd->bfq_burst_interval) ||
1275 bfqq->entity.parent != bfqd->burst_parent_entity) {
1276 bfqd->large_burst = false;
1277 bfq_reset_burst_list(bfqd, bfqq);
1282 * If we get here, then bfqq is being activated shortly after the
1283 * last queue. So, if the current burst is also large, we can mark
1284 * bfqq as belonging to this large burst immediately.
1286 if (bfqd->large_burst) {
1287 bfq_mark_bfqq_in_large_burst(bfqq);
1292 * If we get here, then a large-burst state has not yet been
1293 * reached, but bfqq is being activated shortly after the last
1294 * queue. Then we add bfqq to the burst.
1296 bfq_add_to_burst(bfqd, bfqq);
1299 * At this point, bfqq either has been added to the current
1300 * burst or has caused the current burst to terminate and a
1301 * possible new burst to start. In particular, in the second
1302 * case, bfqq has become the first queue in the possible new
1303 * burst. In both cases last_ins_in_burst needs to be moved
1306 bfqd->last_ins_in_burst = jiffies;
1309 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1311 struct bfq_entity *entity = &bfqq->entity;
1313 return entity->budget - entity->service;
1317 * If enough samples have been computed, return the current max budget
1318 * stored in bfqd, which is dynamically updated according to the
1319 * estimated disk peak rate; otherwise return the default max budget
1321 static int bfq_max_budget(struct bfq_data *bfqd)
1323 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1324 return bfq_default_max_budget;
1326 return bfqd->bfq_max_budget;
1330 * Return min budget, which is a fraction of the current or default
1331 * max budget (trying with 1/32)
1333 static int bfq_min_budget(struct bfq_data *bfqd)
1335 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1336 return bfq_default_max_budget / 32;
1338 return bfqd->bfq_max_budget / 32;
1342 * The next function, invoked after the input queue bfqq switches from
1343 * idle to busy, updates the budget of bfqq. The function also tells
1344 * whether the in-service queue should be expired, by returning
1345 * true. The purpose of expiring the in-service queue is to give bfqq
1346 * the chance to possibly preempt the in-service queue, and the reason
1347 * for preempting the in-service queue is to achieve one of the two
1350 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1351 * expired because it has remained idle. In particular, bfqq may have
1352 * expired for one of the following two reasons:
1354 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1355 * and did not make it to issue a new request before its last
1356 * request was served;
1358 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1359 * a new request before the expiration of the idling-time.
1361 * Even if bfqq has expired for one of the above reasons, the process
1362 * associated with the queue may be however issuing requests greedily,
1363 * and thus be sensitive to the bandwidth it receives (bfqq may have
1364 * remained idle for other reasons: CPU high load, bfqq not enjoying
1365 * idling, I/O throttling somewhere in the path from the process to
1366 * the I/O scheduler, ...). But if, after every expiration for one of
1367 * the above two reasons, bfqq has to wait for the service of at least
1368 * one full budget of another queue before being served again, then
1369 * bfqq is likely to get a much lower bandwidth or resource time than
1370 * its reserved ones. To address this issue, two countermeasures need
1373 * First, the budget and the timestamps of bfqq need to be updated in
1374 * a special way on bfqq reactivation: they need to be updated as if
1375 * bfqq did not remain idle and did not expire. In fact, if they are
1376 * computed as if bfqq expired and remained idle until reactivation,
1377 * then the process associated with bfqq is treated as if, instead of
1378 * being greedy, it stopped issuing requests when bfqq remained idle,
1379 * and restarts issuing requests only on this reactivation. In other
1380 * words, the scheduler does not help the process recover the "service
1381 * hole" between bfqq expiration and reactivation. As a consequence,
1382 * the process receives a lower bandwidth than its reserved one. In
1383 * contrast, to recover this hole, the budget must be updated as if
1384 * bfqq was not expired at all before this reactivation, i.e., it must
1385 * be set to the value of the remaining budget when bfqq was
1386 * expired. Along the same line, timestamps need to be assigned the
1387 * value they had the last time bfqq was selected for service, i.e.,
1388 * before last expiration. Thus timestamps need to be back-shifted
1389 * with respect to their normal computation (see [1] for more details
1390 * on this tricky aspect).
1392 * Secondly, to allow the process to recover the hole, the in-service
1393 * queue must be expired too, to give bfqq the chance to preempt it
1394 * immediately. In fact, if bfqq has to wait for a full budget of the
1395 * in-service queue to be completed, then it may become impossible to
1396 * let the process recover the hole, even if the back-shifted
1397 * timestamps of bfqq are lower than those of the in-service queue. If
1398 * this happens for most or all of the holes, then the process may not
1399 * receive its reserved bandwidth. In this respect, it is worth noting
1400 * that, being the service of outstanding requests unpreemptible, a
1401 * little fraction of the holes may however be unrecoverable, thereby
1402 * causing a little loss of bandwidth.
1404 * The last important point is detecting whether bfqq does need this
1405 * bandwidth recovery. In this respect, the next function deems the
1406 * process associated with bfqq greedy, and thus allows it to recover
1407 * the hole, if: 1) the process is waiting for the arrival of a new
1408 * request (which implies that bfqq expired for one of the above two
1409 * reasons), and 2) such a request has arrived soon. The first
1410 * condition is controlled through the flag non_blocking_wait_rq,
1411 * while the second through the flag arrived_in_time. If both
1412 * conditions hold, then the function computes the budget in the
1413 * above-described special way, and signals that the in-service queue
1414 * should be expired. Timestamp back-shifting is done later in
1415 * __bfq_activate_entity.
1417 * 2. Reduce latency. Even if timestamps are not backshifted to let
1418 * the process associated with bfqq recover a service hole, bfqq may
1419 * however happen to have, after being (re)activated, a lower finish
1420 * timestamp than the in-service queue. That is, the next budget of
1421 * bfqq may have to be completed before the one of the in-service
1422 * queue. If this is the case, then preempting the in-service queue
1423 * allows this goal to be achieved, apart from the unpreemptible,
1424 * outstanding requests mentioned above.
1426 * Unfortunately, regardless of which of the above two goals one wants
1427 * to achieve, service trees need first to be updated to know whether
1428 * the in-service queue must be preempted. To have service trees
1429 * correctly updated, the in-service queue must be expired and
1430 * rescheduled, and bfqq must be scheduled too. This is one of the
1431 * most costly operations (in future versions, the scheduling
1432 * mechanism may be re-designed in such a way to make it possible to
1433 * know whether preemption is needed without needing to update service
1434 * trees). In addition, queue preemptions almost always cause random
1435 * I/O, which may in turn cause loss of throughput. Finally, there may
1436 * even be no in-service queue when the next function is invoked (so,
1437 * no queue to compare timestamps with). Because of these facts, the
1438 * next function adopts the following simple scheme to avoid costly
1439 * operations, too frequent preemptions and too many dependencies on
1440 * the state of the scheduler: it requests the expiration of the
1441 * in-service queue (unconditionally) only for queues that need to
1442 * recover a hole. Then it delegates to other parts of the code the
1443 * responsibility of handling the above case 2.
1445 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1446 struct bfq_queue *bfqq,
1447 bool arrived_in_time)
1449 struct bfq_entity *entity = &bfqq->entity;
1452 * In the next compound condition, we check also whether there
1453 * is some budget left, because otherwise there is no point in
1454 * trying to go on serving bfqq with this same budget: bfqq
1455 * would be expired immediately after being selected for
1456 * service. This would only cause useless overhead.
1458 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1459 bfq_bfqq_budget_left(bfqq) > 0) {
1461 * We do not clear the flag non_blocking_wait_rq here, as
1462 * the latter is used in bfq_activate_bfqq to signal
1463 * that timestamps need to be back-shifted (and is
1464 * cleared right after).
1468 * In next assignment we rely on that either
1469 * entity->service or entity->budget are not updated
1470 * on expiration if bfqq is empty (see
1471 * __bfq_bfqq_recalc_budget). Thus both quantities
1472 * remain unchanged after such an expiration, and the
1473 * following statement therefore assigns to
1474 * entity->budget the remaining budget on such an
1477 entity->budget = min_t(unsigned long,
1478 bfq_bfqq_budget_left(bfqq),
1482 * At this point, we have used entity->service to get
1483 * the budget left (needed for updating
1484 * entity->budget). Thus we finally can, and have to,
1485 * reset entity->service. The latter must be reset
1486 * because bfqq would otherwise be charged again for
1487 * the service it has received during its previous
1490 entity->service = 0;
1496 * We can finally complete expiration, by setting service to 0.
1498 entity->service = 0;
1499 entity->budget = max_t(unsigned long, bfqq->max_budget,
1500 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1501 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1506 * Return the farthest past time instant according to jiffies
1509 static unsigned long bfq_smallest_from_now(void)
1511 return jiffies - MAX_JIFFY_OFFSET;
1514 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1515 struct bfq_queue *bfqq,
1516 unsigned int old_wr_coeff,
1517 bool wr_or_deserves_wr,
1522 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1523 /* start a weight-raising period */
1525 bfqq->service_from_wr = 0;
1526 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1527 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1530 * No interactive weight raising in progress
1531 * here: assign minus infinity to
1532 * wr_start_at_switch_to_srt, to make sure
1533 * that, at the end of the soft-real-time
1534 * weight raising periods that is starting
1535 * now, no interactive weight-raising period
1536 * may be wrongly considered as still in
1537 * progress (and thus actually started by
1540 bfqq->wr_start_at_switch_to_srt =
1541 bfq_smallest_from_now();
1542 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1543 BFQ_SOFTRT_WEIGHT_FACTOR;
1544 bfqq->wr_cur_max_time =
1545 bfqd->bfq_wr_rt_max_time;
1549 * If needed, further reduce budget to make sure it is
1550 * close to bfqq's backlog, so as to reduce the
1551 * scheduling-error component due to a too large
1552 * budget. Do not care about throughput consequences,
1553 * but only about latency. Finally, do not assign a
1554 * too small budget either, to avoid increasing
1555 * latency by causing too frequent expirations.
1557 bfqq->entity.budget = min_t(unsigned long,
1558 bfqq->entity.budget,
1559 2 * bfq_min_budget(bfqd));
1560 } else if (old_wr_coeff > 1) {
1561 if (interactive) { /* update wr coeff and duration */
1562 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1563 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1564 } else if (in_burst)
1568 * The application is now or still meeting the
1569 * requirements for being deemed soft rt. We
1570 * can then correctly and safely (re)charge
1571 * the weight-raising duration for the
1572 * application with the weight-raising
1573 * duration for soft rt applications.
1575 * In particular, doing this recharge now, i.e.,
1576 * before the weight-raising period for the
1577 * application finishes, reduces the probability
1578 * of the following negative scenario:
1579 * 1) the weight of a soft rt application is
1580 * raised at startup (as for any newly
1581 * created application),
1582 * 2) since the application is not interactive,
1583 * at a certain time weight-raising is
1584 * stopped for the application,
1585 * 3) at that time the application happens to
1586 * still have pending requests, and hence
1587 * is destined to not have a chance to be
1588 * deemed soft rt before these requests are
1589 * completed (see the comments to the
1590 * function bfq_bfqq_softrt_next_start()
1591 * for details on soft rt detection),
1592 * 4) these pending requests experience a high
1593 * latency because the application is not
1594 * weight-raised while they are pending.
1596 if (bfqq->wr_cur_max_time !=
1597 bfqd->bfq_wr_rt_max_time) {
1598 bfqq->wr_start_at_switch_to_srt =
1599 bfqq->last_wr_start_finish;
1601 bfqq->wr_cur_max_time =
1602 bfqd->bfq_wr_rt_max_time;
1603 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1604 BFQ_SOFTRT_WEIGHT_FACTOR;
1606 bfqq->last_wr_start_finish = jiffies;
1611 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1612 struct bfq_queue *bfqq)
1614 return bfqq->dispatched == 0 &&
1615 time_is_before_jiffies(
1616 bfqq->budget_timeout +
1617 bfqd->bfq_wr_min_idle_time);
1622 * Return true if bfqq is in a higher priority class, or has a higher
1623 * weight than the in-service queue.
1625 static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq,
1626 struct bfq_queue *in_serv_bfqq)
1628 int bfqq_weight, in_serv_weight;
1630 if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class)
1633 if (in_serv_bfqq->entity.parent == bfqq->entity.parent) {
1634 bfqq_weight = bfqq->entity.weight;
1635 in_serv_weight = in_serv_bfqq->entity.weight;
1637 if (bfqq->entity.parent)
1638 bfqq_weight = bfqq->entity.parent->weight;
1640 bfqq_weight = bfqq->entity.weight;
1641 if (in_serv_bfqq->entity.parent)
1642 in_serv_weight = in_serv_bfqq->entity.parent->weight;
1644 in_serv_weight = in_serv_bfqq->entity.weight;
1647 return bfqq_weight > in_serv_weight;
1650 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1651 struct bfq_queue *bfqq,
1656 bool soft_rt, in_burst, wr_or_deserves_wr,
1657 bfqq_wants_to_preempt,
1658 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1660 * See the comments on
1661 * bfq_bfqq_update_budg_for_activation for
1662 * details on the usage of the next variable.
1664 arrived_in_time = ktime_get_ns() <=
1665 bfqq->ttime.last_end_request +
1666 bfqd->bfq_slice_idle * 3;
1670 * bfqq deserves to be weight-raised if:
1672 * - it does not belong to a large burst,
1673 * - it has been idle for enough time or is soft real-time,
1674 * - is linked to a bfq_io_cq (it is not shared in any sense),
1675 * - has a default weight (otherwise we assume the user wanted
1676 * to control its weight explicitly)
1678 in_burst = bfq_bfqq_in_large_burst(bfqq);
1679 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1680 !BFQQ_TOTALLY_SEEKY(bfqq) &&
1682 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1683 bfqq->dispatched == 0 &&
1684 bfqq->entity.new_weight == 40;
1685 *interactive = !in_burst && idle_for_long_time &&
1686 bfqq->entity.new_weight == 40;
1687 wr_or_deserves_wr = bfqd->low_latency &&
1688 (bfqq->wr_coeff > 1 ||
1689 (bfq_bfqq_sync(bfqq) &&
1690 bfqq->bic && (*interactive || soft_rt)));
1693 * Using the last flag, update budget and check whether bfqq
1694 * may want to preempt the in-service queue.
1696 bfqq_wants_to_preempt =
1697 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1701 * If bfqq happened to be activated in a burst, but has been
1702 * idle for much more than an interactive queue, then we
1703 * assume that, in the overall I/O initiated in the burst, the
1704 * I/O associated with bfqq is finished. So bfqq does not need
1705 * to be treated as a queue belonging to a burst
1706 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1707 * if set, and remove bfqq from the burst list if it's
1708 * there. We do not decrement burst_size, because the fact
1709 * that bfqq does not need to belong to the burst list any
1710 * more does not invalidate the fact that bfqq was created in
1713 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1714 idle_for_long_time &&
1715 time_is_before_jiffies(
1716 bfqq->budget_timeout +
1717 msecs_to_jiffies(10000))) {
1718 hlist_del_init(&bfqq->burst_list_node);
1719 bfq_clear_bfqq_in_large_burst(bfqq);
1722 bfq_clear_bfqq_just_created(bfqq);
1725 if (!bfq_bfqq_IO_bound(bfqq)) {
1726 if (arrived_in_time) {
1727 bfqq->requests_within_timer++;
1728 if (bfqq->requests_within_timer >=
1729 bfqd->bfq_requests_within_timer)
1730 bfq_mark_bfqq_IO_bound(bfqq);
1732 bfqq->requests_within_timer = 0;
1735 if (bfqd->low_latency) {
1736 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1739 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1741 if (time_is_before_jiffies(bfqq->split_time +
1742 bfqd->bfq_wr_min_idle_time)) {
1743 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1750 if (old_wr_coeff != bfqq->wr_coeff)
1751 bfqq->entity.prio_changed = 1;
1755 bfqq->last_idle_bklogged = jiffies;
1756 bfqq->service_from_backlogged = 0;
1757 bfq_clear_bfqq_softrt_update(bfqq);
1759 bfq_add_bfqq_busy(bfqd, bfqq);
1762 * Expire in-service queue only if preemption may be needed
1763 * for guarantees. In particular, we care only about two
1764 * cases. The first is that bfqq has to recover a service
1765 * hole, as explained in the comments on
1766 * bfq_bfqq_update_budg_for_activation(), i.e., that
1767 * bfqq_wants_to_preempt is true. However, if bfqq does not
1768 * carry time-critical I/O, then bfqq's bandwidth is less
1769 * important than that of queues that carry time-critical I/O.
1770 * So, as a further constraint, we consider this case only if
1771 * bfqq is at least as weight-raised, i.e., at least as time
1772 * critical, as the in-service queue.
1774 * The second case is that bfqq is in a higher priority class,
1775 * or has a higher weight than the in-service queue. If this
1776 * condition does not hold, we don't care because, even if
1777 * bfqq does not start to be served immediately, the resulting
1778 * delay for bfqq's I/O is however lower or much lower than
1779 * the ideal completion time to be guaranteed to bfqq's I/O.
1781 * In both cases, preemption is needed only if, according to
1782 * the timestamps of both bfqq and of the in-service queue,
1783 * bfqq actually is the next queue to serve. So, to reduce
1784 * useless preemptions, the return value of
1785 * next_queue_may_preempt() is considered in the next compound
1786 * condition too. Yet next_queue_may_preempt() just checks a
1787 * simple, necessary condition for bfqq to be the next queue
1788 * to serve. In fact, to evaluate a sufficient condition, the
1789 * timestamps of the in-service queue would need to be
1790 * updated, and this operation is quite costly (see the
1791 * comments on bfq_bfqq_update_budg_for_activation()).
1793 if (bfqd->in_service_queue &&
1794 ((bfqq_wants_to_preempt &&
1795 bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) ||
1796 bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue)) &&
1797 next_queue_may_preempt(bfqd))
1798 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1799 false, BFQQE_PREEMPTED);
1802 static void bfq_reset_inject_limit(struct bfq_data *bfqd,
1803 struct bfq_queue *bfqq)
1805 /* invalidate baseline total service time */
1806 bfqq->last_serv_time_ns = 0;
1809 * Reset pointer in case we are waiting for
1810 * some request completion.
1812 bfqd->waited_rq = NULL;
1815 * If bfqq has a short think time, then start by setting the
1816 * inject limit to 0 prudentially, because the service time of
1817 * an injected I/O request may be higher than the think time
1818 * of bfqq, and therefore, if one request was injected when
1819 * bfqq remains empty, this injected request might delay the
1820 * service of the next I/O request for bfqq significantly. In
1821 * case bfqq can actually tolerate some injection, then the
1822 * adaptive update will however raise the limit soon. This
1823 * lucky circumstance holds exactly because bfqq has a short
1824 * think time, and thus, after remaining empty, is likely to
1825 * get new I/O enqueued---and then completed---before being
1826 * expired. This is the very pattern that gives the
1827 * limit-update algorithm the chance to measure the effect of
1828 * injection on request service times, and then to update the
1829 * limit accordingly.
1831 * However, in the following special case, the inject limit is
1832 * left to 1 even if the think time is short: bfqq's I/O is
1833 * synchronized with that of some other queue, i.e., bfqq may
1834 * receive new I/O only after the I/O of the other queue is
1835 * completed. Keeping the inject limit to 1 allows the
1836 * blocking I/O to be served while bfqq is in service. And
1837 * this is very convenient both for bfqq and for overall
1838 * throughput, as explained in detail in the comments in
1839 * bfq_update_has_short_ttime().
1841 * On the opposite end, if bfqq has a long think time, then
1842 * start directly by 1, because:
1843 * a) on the bright side, keeping at most one request in
1844 * service in the drive is unlikely to cause any harm to the
1845 * latency of bfqq's requests, as the service time of a single
1846 * request is likely to be lower than the think time of bfqq;
1847 * b) on the downside, after becoming empty, bfqq is likely to
1848 * expire before getting its next request. With this request
1849 * arrival pattern, it is very hard to sample total service
1850 * times and update the inject limit accordingly (see comments
1851 * on bfq_update_inject_limit()). So the limit is likely to be
1852 * never, or at least seldom, updated. As a consequence, by
1853 * setting the limit to 1, we avoid that no injection ever
1854 * occurs with bfqq. On the downside, this proactive step
1855 * further reduces chances to actually compute the baseline
1856 * total service time. Thus it reduces chances to execute the
1857 * limit-update algorithm and possibly raise the limit to more
1860 if (bfq_bfqq_has_short_ttime(bfqq))
1861 bfqq->inject_limit = 0;
1863 bfqq->inject_limit = 1;
1865 bfqq->decrease_time_jif = jiffies;
1868 static void bfq_add_request(struct request *rq)
1870 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1871 struct bfq_data *bfqd = bfqq->bfqd;
1872 struct request *next_rq, *prev;
1873 unsigned int old_wr_coeff = bfqq->wr_coeff;
1874 bool interactive = false;
1876 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1877 bfqq->queued[rq_is_sync(rq)]++;
1880 if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_bfqq_sync(bfqq)) {
1882 * Detect whether bfqq's I/O seems synchronized with
1883 * that of some other queue, i.e., whether bfqq, after
1884 * remaining empty, happens to receive new I/O only
1885 * right after some I/O request of the other queue has
1886 * been completed. We call waker queue the other
1887 * queue, and we assume, for simplicity, that bfqq may
1888 * have at most one waker queue.
1890 * A remarkable throughput boost can be reached by
1891 * unconditionally injecting the I/O of the waker
1892 * queue, every time a new bfq_dispatch_request
1893 * happens to be invoked while I/O is being plugged
1894 * for bfqq. In addition to boosting throughput, this
1895 * unblocks bfqq's I/O, thereby improving bandwidth
1896 * and latency for bfqq. Note that these same results
1897 * may be achieved with the general injection
1898 * mechanism, but less effectively. For details on
1899 * this aspect, see the comments on the choice of the
1900 * queue for injection in bfq_select_queue().
1902 * Turning back to the detection of a waker queue, a
1903 * queue Q is deemed as a waker queue for bfqq if, for
1904 * two consecutive times, bfqq happens to become non
1905 * empty right after a request of Q has been
1906 * completed. In particular, on the first time, Q is
1907 * tentatively set as a candidate waker queue, while
1908 * on the second time, the flag
1909 * bfq_bfqq_has_waker(bfqq) is set to confirm that Q
1910 * is a waker queue for bfqq. These detection steps
1911 * are performed only if bfqq has a long think time,
1912 * so as to make it more likely that bfqq's I/O is
1913 * actually being blocked by a synchronization. This
1914 * last filter, plus the above two-times requirement,
1915 * make false positives less likely.
1919 * The sooner a waker queue is detected, the sooner
1920 * throughput can be boosted by injecting I/O from the
1921 * waker queue. Fortunately, detection is likely to be
1922 * actually fast, for the following reasons. While
1923 * blocked by synchronization, bfqq has a long think
1924 * time. This implies that bfqq's inject limit is at
1925 * least equal to 1 (see the comments in
1926 * bfq_update_inject_limit()). So, thanks to
1927 * injection, the waker queue is likely to be served
1928 * during the very first I/O-plugging time interval
1929 * for bfqq. This triggers the first step of the
1930 * detection mechanism. Thanks again to injection, the
1931 * candidate waker queue is then likely to be
1932 * confirmed no later than during the next
1933 * I/O-plugging interval for bfqq.
1935 if (bfqd->last_completed_rq_bfqq &&
1936 !bfq_bfqq_has_short_ttime(bfqq) &&
1937 ktime_get_ns() - bfqd->last_completion <
1938 4 * NSEC_PER_MSEC) {
1939 if (bfqd->last_completed_rq_bfqq != bfqq &&
1940 bfqd->last_completed_rq_bfqq !=
1943 * First synchronization detected with
1944 * a candidate waker queue, or with a
1945 * different candidate waker queue
1946 * from the current one.
1948 bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
1951 * If the waker queue disappears, then
1952 * bfqq->waker_bfqq must be reset. To
1953 * this goal, we maintain in each
1954 * waker queue a list, woken_list, of
1955 * all the queues that reference the
1956 * waker queue through their
1957 * waker_bfqq pointer. When the waker
1958 * queue exits, the waker_bfqq pointer
1959 * of all the queues in the woken_list
1962 * In addition, if bfqq is already in
1963 * the woken_list of a waker queue,
1964 * then, before being inserted into
1965 * the woken_list of a new waker
1966 * queue, bfqq must be removed from
1967 * the woken_list of the old waker
1970 if (!hlist_unhashed(&bfqq->woken_list_node))
1971 hlist_del_init(&bfqq->woken_list_node);
1972 hlist_add_head(&bfqq->woken_list_node,
1973 &bfqd->last_completed_rq_bfqq->woken_list);
1975 bfq_clear_bfqq_has_waker(bfqq);
1976 } else if (bfqd->last_completed_rq_bfqq ==
1978 !bfq_bfqq_has_waker(bfqq)) {
1980 * synchronization with waker_bfqq
1981 * seen for the second time
1983 bfq_mark_bfqq_has_waker(bfqq);
1988 * Periodically reset inject limit, to make sure that
1989 * the latter eventually drops in case workload
1990 * changes, see step (3) in the comments on
1991 * bfq_update_inject_limit().
1993 if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
1994 msecs_to_jiffies(1000)))
1995 bfq_reset_inject_limit(bfqd, bfqq);
1998 * The following conditions must hold to setup a new
1999 * sampling of total service time, and then a new
2000 * update of the inject limit:
2001 * - bfqq is in service, because the total service
2002 * time is evaluated only for the I/O requests of
2003 * the queues in service;
2004 * - this is the right occasion to compute or to
2005 * lower the baseline total service time, because
2006 * there are actually no requests in the drive,
2008 * the baseline total service time is available, and
2009 * this is the right occasion to compute the other
2010 * quantity needed to update the inject limit, i.e.,
2011 * the total service time caused by the amount of
2012 * injection allowed by the current value of the
2013 * limit. It is the right occasion because injection
2014 * has actually been performed during the service
2015 * hole, and there are still in-flight requests,
2016 * which are very likely to be exactly the injected
2017 * requests, or part of them;
2018 * - the minimum interval for sampling the total
2019 * service time and updating the inject limit has
2022 if (bfqq == bfqd->in_service_queue &&
2023 (bfqd->rq_in_driver == 0 ||
2024 (bfqq->last_serv_time_ns > 0 &&
2025 bfqd->rqs_injected && bfqd->rq_in_driver > 0)) &&
2026 time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2027 msecs_to_jiffies(10))) {
2028 bfqd->last_empty_occupied_ns = ktime_get_ns();
2030 * Start the state machine for measuring the
2031 * total service time of rq: setting
2032 * wait_dispatch will cause bfqd->waited_rq to
2033 * be set when rq will be dispatched.
2035 bfqd->wait_dispatch = true;
2037 * If there is no I/O in service in the drive,
2038 * then possible injection occurred before the
2039 * arrival of rq will not affect the total
2040 * service time of rq. So the injection limit
2041 * must not be updated as a function of such
2042 * total service time, unless new injection
2043 * occurs before rq is completed. To have the
2044 * injection limit updated only in the latter
2045 * case, reset rqs_injected here (rqs_injected
2046 * will be set in case injection is performed
2047 * on bfqq before rq is completed).
2049 if (bfqd->rq_in_driver == 0)
2050 bfqd->rqs_injected = false;
2054 elv_rb_add(&bfqq->sort_list, rq);
2057 * Check if this request is a better next-serve candidate.
2059 prev = bfqq->next_rq;
2060 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
2061 bfqq->next_rq = next_rq;
2064 * Adjust priority tree position, if next_rq changes.
2065 * See comments on bfq_pos_tree_add_move() for the unlikely().
2067 if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
2068 bfq_pos_tree_add_move(bfqd, bfqq);
2070 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
2071 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
2074 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
2075 time_is_before_jiffies(
2076 bfqq->last_wr_start_finish +
2077 bfqd->bfq_wr_min_inter_arr_async)) {
2078 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
2079 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
2081 bfqd->wr_busy_queues++;
2082 bfqq->entity.prio_changed = 1;
2084 if (prev != bfqq->next_rq)
2085 bfq_updated_next_req(bfqd, bfqq);
2089 * Assign jiffies to last_wr_start_finish in the following
2092 * . if bfqq is not going to be weight-raised, because, for
2093 * non weight-raised queues, last_wr_start_finish stores the
2094 * arrival time of the last request; as of now, this piece
2095 * of information is used only for deciding whether to
2096 * weight-raise async queues
2098 * . if bfqq is not weight-raised, because, if bfqq is now
2099 * switching to weight-raised, then last_wr_start_finish
2100 * stores the time when weight-raising starts
2102 * . if bfqq is interactive, because, regardless of whether
2103 * bfqq is currently weight-raised, the weight-raising
2104 * period must start or restart (this case is considered
2105 * separately because it is not detected by the above
2106 * conditions, if bfqq is already weight-raised)
2108 * last_wr_start_finish has to be updated also if bfqq is soft
2109 * real-time, because the weight-raising period is constantly
2110 * restarted on idle-to-busy transitions for these queues, but
2111 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2114 if (bfqd->low_latency &&
2115 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
2116 bfqq->last_wr_start_finish = jiffies;
2119 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
2121 struct request_queue *q)
2123 struct bfq_queue *bfqq = bfqd->bio_bfqq;
2127 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
2132 static sector_t get_sdist(sector_t last_pos, struct request *rq)
2135 return abs(blk_rq_pos(rq) - last_pos);
2140 #if 0 /* Still not clear if we can do without next two functions */
2141 static void bfq_activate_request(struct request_queue *q, struct request *rq)
2143 struct bfq_data *bfqd = q->elevator->elevator_data;
2145 bfqd->rq_in_driver++;
2148 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
2150 struct bfq_data *bfqd = q->elevator->elevator_data;
2152 bfqd->rq_in_driver--;
2156 static void bfq_remove_request(struct request_queue *q,
2159 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2160 struct bfq_data *bfqd = bfqq->bfqd;
2161 const int sync = rq_is_sync(rq);
2163 if (bfqq->next_rq == rq) {
2164 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
2165 bfq_updated_next_req(bfqd, bfqq);
2168 if (rq->queuelist.prev != &rq->queuelist)
2169 list_del_init(&rq->queuelist);
2170 bfqq->queued[sync]--;
2172 elv_rb_del(&bfqq->sort_list, rq);
2174 elv_rqhash_del(q, rq);
2175 if (q->last_merge == rq)
2176 q->last_merge = NULL;
2178 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2179 bfqq->next_rq = NULL;
2181 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
2182 bfq_del_bfqq_busy(bfqd, bfqq, false);
2184 * bfqq emptied. In normal operation, when
2185 * bfqq is empty, bfqq->entity.service and
2186 * bfqq->entity.budget must contain,
2187 * respectively, the service received and the
2188 * budget used last time bfqq emptied. These
2189 * facts do not hold in this case, as at least
2190 * this last removal occurred while bfqq is
2191 * not in service. To avoid inconsistencies,
2192 * reset both bfqq->entity.service and
2193 * bfqq->entity.budget, if bfqq has still a
2194 * process that may issue I/O requests to it.
2196 bfqq->entity.budget = bfqq->entity.service = 0;
2200 * Remove queue from request-position tree as it is empty.
2202 if (bfqq->pos_root) {
2203 rb_erase(&bfqq->pos_node, bfqq->pos_root);
2204 bfqq->pos_root = NULL;
2207 /* see comments on bfq_pos_tree_add_move() for the unlikely() */
2208 if (unlikely(!bfqd->nonrot_with_queueing))
2209 bfq_pos_tree_add_move(bfqd, bfqq);
2212 if (rq->cmd_flags & REQ_META)
2213 bfqq->meta_pending--;
2217 static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio,
2218 unsigned int nr_segs)
2220 struct request_queue *q = hctx->queue;
2221 struct bfq_data *bfqd = q->elevator->elevator_data;
2222 struct request *free = NULL;
2224 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2225 * store its return value for later use, to avoid nesting
2226 * queue_lock inside the bfqd->lock. We assume that the bic
2227 * returned by bfq_bic_lookup does not go away before
2228 * bfqd->lock is taken.
2230 struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
2233 spin_lock_irq(&bfqd->lock);
2236 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
2238 bfqd->bio_bfqq = NULL;
2239 bfqd->bio_bic = bic;
2241 ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
2244 blk_mq_free_request(free);
2245 spin_unlock_irq(&bfqd->lock);
2250 static int bfq_request_merge(struct request_queue *q, struct request **req,
2253 struct bfq_data *bfqd = q->elevator->elevator_data;
2254 struct request *__rq;
2256 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
2257 if (__rq && elv_bio_merge_ok(__rq, bio)) {
2259 return ELEVATOR_FRONT_MERGE;
2262 return ELEVATOR_NO_MERGE;
2265 static struct bfq_queue *bfq_init_rq(struct request *rq);
2267 static void bfq_request_merged(struct request_queue *q, struct request *req,
2268 enum elv_merge type)
2270 if (type == ELEVATOR_FRONT_MERGE &&
2271 rb_prev(&req->rb_node) &&
2273 blk_rq_pos(container_of(rb_prev(&req->rb_node),
2274 struct request, rb_node))) {
2275 struct bfq_queue *bfqq = bfq_init_rq(req);
2276 struct bfq_data *bfqd;
2277 struct request *prev, *next_rq;
2284 /* Reposition request in its sort_list */
2285 elv_rb_del(&bfqq->sort_list, req);
2286 elv_rb_add(&bfqq->sort_list, req);
2288 /* Choose next request to be served for bfqq */
2289 prev = bfqq->next_rq;
2290 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
2291 bfqd->last_position);
2292 bfqq->next_rq = next_rq;
2294 * If next_rq changes, update both the queue's budget to
2295 * fit the new request and the queue's position in its
2298 if (prev != bfqq->next_rq) {
2299 bfq_updated_next_req(bfqd, bfqq);
2301 * See comments on bfq_pos_tree_add_move() for
2304 if (unlikely(!bfqd->nonrot_with_queueing))
2305 bfq_pos_tree_add_move(bfqd, bfqq);
2311 * This function is called to notify the scheduler that the requests
2312 * rq and 'next' have been merged, with 'next' going away. BFQ
2313 * exploits this hook to address the following issue: if 'next' has a
2314 * fifo_time lower that rq, then the fifo_time of rq must be set to
2315 * the value of 'next', to not forget the greater age of 'next'.
2317 * NOTE: in this function we assume that rq is in a bfq_queue, basing
2318 * on that rq is picked from the hash table q->elevator->hash, which,
2319 * in its turn, is filled only with I/O requests present in
2320 * bfq_queues, while BFQ is in use for the request queue q. In fact,
2321 * the function that fills this hash table (elv_rqhash_add) is called
2322 * only by bfq_insert_request.
2324 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
2325 struct request *next)
2327 struct bfq_queue *bfqq = bfq_init_rq(rq),
2328 *next_bfqq = bfq_init_rq(next);
2334 * If next and rq belong to the same bfq_queue and next is older
2335 * than rq, then reposition rq in the fifo (by substituting next
2336 * with rq). Otherwise, if next and rq belong to different
2337 * bfq_queues, never reposition rq: in fact, we would have to
2338 * reposition it with respect to next's position in its own fifo,
2339 * which would most certainly be too expensive with respect to
2342 if (bfqq == next_bfqq &&
2343 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2344 next->fifo_time < rq->fifo_time) {
2345 list_del_init(&rq->queuelist);
2346 list_replace_init(&next->queuelist, &rq->queuelist);
2347 rq->fifo_time = next->fifo_time;
2350 if (bfqq->next_rq == next)
2353 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
2356 /* Must be called with bfqq != NULL */
2357 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
2360 * If bfqq has been enjoying interactive weight-raising, then
2361 * reset soft_rt_next_start. We do it for the following
2362 * reason. bfqq may have been conveying the I/O needed to load
2363 * a soft real-time application. Such an application actually
2364 * exhibits a soft real-time I/O pattern after it finishes
2365 * loading, and finally starts doing its job. But, if bfqq has
2366 * been receiving a lot of bandwidth so far (likely to happen
2367 * on a fast device), then soft_rt_next_start now contains a
2368 * high value that. So, without this reset, bfqq would be
2369 * prevented from being possibly considered as soft_rt for a
2373 if (bfqq->wr_cur_max_time !=
2374 bfqq->bfqd->bfq_wr_rt_max_time)
2375 bfqq->soft_rt_next_start = jiffies;
2377 if (bfq_bfqq_busy(bfqq))
2378 bfqq->bfqd->wr_busy_queues--;
2380 bfqq->wr_cur_max_time = 0;
2381 bfqq->last_wr_start_finish = jiffies;
2383 * Trigger a weight change on the next invocation of
2384 * __bfq_entity_update_weight_prio.
2386 bfqq->entity.prio_changed = 1;
2389 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2390 struct bfq_group *bfqg)
2394 for (i = 0; i < 2; i++)
2395 for (j = 0; j < IOPRIO_BE_NR; j++)
2396 if (bfqg->async_bfqq[i][j])
2397 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2398 if (bfqg->async_idle_bfqq)
2399 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2402 static void bfq_end_wr(struct bfq_data *bfqd)
2404 struct bfq_queue *bfqq;
2406 spin_lock_irq(&bfqd->lock);
2408 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2409 bfq_bfqq_end_wr(bfqq);
2410 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2411 bfq_bfqq_end_wr(bfqq);
2412 bfq_end_wr_async(bfqd);
2414 spin_unlock_irq(&bfqd->lock);
2417 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2420 return blk_rq_pos(io_struct);
2422 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2425 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2428 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2432 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2433 struct bfq_queue *bfqq,
2436 struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
2437 struct rb_node *parent, *node;
2438 struct bfq_queue *__bfqq;
2440 if (RB_EMPTY_ROOT(root))
2444 * First, if we find a request starting at the end of the last
2445 * request, choose it.
2447 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2452 * If the exact sector wasn't found, the parent of the NULL leaf
2453 * will contain the closest sector (rq_pos_tree sorted by
2454 * next_request position).
2456 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2457 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2460 if (blk_rq_pos(__bfqq->next_rq) < sector)
2461 node = rb_next(&__bfqq->pos_node);
2463 node = rb_prev(&__bfqq->pos_node);
2467 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2468 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2474 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2475 struct bfq_queue *cur_bfqq,
2478 struct bfq_queue *bfqq;
2481 * We shall notice if some of the queues are cooperating,
2482 * e.g., working closely on the same area of the device. In
2483 * that case, we can group them together and: 1) don't waste
2484 * time idling, and 2) serve the union of their requests in
2485 * the best possible order for throughput.
2487 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2488 if (!bfqq || bfqq == cur_bfqq)
2494 static struct bfq_queue *
2495 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2497 int process_refs, new_process_refs;
2498 struct bfq_queue *__bfqq;
2501 * If there are no process references on the new_bfqq, then it is
2502 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2503 * may have dropped their last reference (not just their last process
2506 if (!bfqq_process_refs(new_bfqq))
2509 /* Avoid a circular list and skip interim queue merges. */
2510 while ((__bfqq = new_bfqq->new_bfqq)) {
2516 process_refs = bfqq_process_refs(bfqq);
2517 new_process_refs = bfqq_process_refs(new_bfqq);
2519 * If the process for the bfqq has gone away, there is no
2520 * sense in merging the queues.
2522 if (process_refs == 0 || new_process_refs == 0)
2525 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2529 * Merging is just a redirection: the requests of the process
2530 * owning one of the two queues are redirected to the other queue.
2531 * The latter queue, in its turn, is set as shared if this is the
2532 * first time that the requests of some process are redirected to
2535 * We redirect bfqq to new_bfqq and not the opposite, because
2536 * we are in the context of the process owning bfqq, thus we
2537 * have the io_cq of this process. So we can immediately
2538 * configure this io_cq to redirect the requests of the
2539 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2540 * not available any more (new_bfqq->bic == NULL).
2542 * Anyway, even in case new_bfqq coincides with the in-service
2543 * queue, redirecting requests the in-service queue is the
2544 * best option, as we feed the in-service queue with new
2545 * requests close to the last request served and, by doing so,
2546 * are likely to increase the throughput.
2548 bfqq->new_bfqq = new_bfqq;
2549 new_bfqq->ref += process_refs;
2553 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2554 struct bfq_queue *new_bfqq)
2556 if (bfq_too_late_for_merging(new_bfqq))
2559 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2560 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2564 * If either of the queues has already been detected as seeky,
2565 * then merging it with the other queue is unlikely to lead to
2568 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2572 * Interleaved I/O is known to be done by (some) applications
2573 * only for reads, so it does not make sense to merge async
2576 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2583 * Attempt to schedule a merge of bfqq with the currently in-service
2584 * queue or with a close queue among the scheduled queues. Return
2585 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2586 * structure otherwise.
2588 * The OOM queue is not allowed to participate to cooperation: in fact, since
2589 * the requests temporarily redirected to the OOM queue could be redirected
2590 * again to dedicated queues at any time, the state needed to correctly
2591 * handle merging with the OOM queue would be quite complex and expensive
2592 * to maintain. Besides, in such a critical condition as an out of memory,
2593 * the benefits of queue merging may be little relevant, or even negligible.
2595 * WARNING: queue merging may impair fairness among non-weight raised
2596 * queues, for at least two reasons: 1) the original weight of a
2597 * merged queue may change during the merged state, 2) even being the
2598 * weight the same, a merged queue may be bloated with many more
2599 * requests than the ones produced by its originally-associated
2602 static struct bfq_queue *
2603 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2604 void *io_struct, bool request)
2606 struct bfq_queue *in_service_bfqq, *new_bfqq;
2609 * Do not perform queue merging if the device is non
2610 * rotational and performs internal queueing. In fact, such a
2611 * device reaches a high speed through internal parallelism
2612 * and pipelining. This means that, to reach a high
2613 * throughput, it must have many requests enqueued at the same
2614 * time. But, in this configuration, the internal scheduling
2615 * algorithm of the device does exactly the job of queue
2616 * merging: it reorders requests so as to obtain as much as
2617 * possible a sequential I/O pattern. As a consequence, with
2618 * the workload generated by processes doing interleaved I/O,
2619 * the throughput reached by the device is likely to be the
2620 * same, with and without queue merging.
2622 * Disabling merging also provides a remarkable benefit in
2623 * terms of throughput. Merging tends to make many workloads
2624 * artificially more uneven, because of shared queues
2625 * remaining non empty for incomparably more time than
2626 * non-merged queues. This may accentuate workload
2627 * asymmetries. For example, if one of the queues in a set of
2628 * merged queues has a higher weight than a normal queue, then
2629 * the shared queue may inherit such a high weight and, by
2630 * staying almost always active, may force BFQ to perform I/O
2631 * plugging most of the time. This evidently makes it harder
2632 * for BFQ to let the device reach a high throughput.
2634 * Finally, the likely() macro below is not used because one
2635 * of the two branches is more likely than the other, but to
2636 * have the code path after the following if() executed as
2637 * fast as possible for the case of a non rotational device
2638 * with queueing. We want it because this is the fastest kind
2639 * of device. On the opposite end, the likely() may lengthen
2640 * the execution time of BFQ for the case of slower devices
2641 * (rotational or at least without queueing). But in this case
2642 * the execution time of BFQ matters very little, if not at
2645 if (likely(bfqd->nonrot_with_queueing))
2649 * Prevent bfqq from being merged if it has been created too
2650 * long ago. The idea is that true cooperating processes, and
2651 * thus their associated bfq_queues, are supposed to be
2652 * created shortly after each other. This is the case, e.g.,
2653 * for KVM/QEMU and dump I/O threads. Basing on this
2654 * assumption, the following filtering greatly reduces the
2655 * probability that two non-cooperating processes, which just
2656 * happen to do close I/O for some short time interval, have
2657 * their queues merged by mistake.
2659 if (bfq_too_late_for_merging(bfqq))
2663 return bfqq->new_bfqq;
2665 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2668 /* If there is only one backlogged queue, don't search. */
2669 if (bfq_tot_busy_queues(bfqd) == 1)
2672 in_service_bfqq = bfqd->in_service_queue;
2674 if (in_service_bfqq && in_service_bfqq != bfqq &&
2675 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2676 bfq_rq_close_to_sector(io_struct, request,
2677 bfqd->in_serv_last_pos) &&
2678 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2679 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2680 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2685 * Check whether there is a cooperator among currently scheduled
2686 * queues. The only thing we need is that the bio/request is not
2687 * NULL, as we need it to establish whether a cooperator exists.
2689 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2690 bfq_io_struct_pos(io_struct, request));
2692 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2693 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2694 return bfq_setup_merge(bfqq, new_bfqq);
2699 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2701 struct bfq_io_cq *bic = bfqq->bic;
2704 * If !bfqq->bic, the queue is already shared or its requests
2705 * have already been redirected to a shared queue; both idle window
2706 * and weight raising state have already been saved. Do nothing.
2711 bic->saved_weight = bfqq->entity.orig_weight;
2712 bic->saved_ttime = bfqq->ttime;
2713 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2714 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2715 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2716 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2717 if (unlikely(bfq_bfqq_just_created(bfqq) &&
2718 !bfq_bfqq_in_large_burst(bfqq) &&
2719 bfqq->bfqd->low_latency)) {
2721 * bfqq being merged right after being created: bfqq
2722 * would have deserved interactive weight raising, but
2723 * did not make it to be set in a weight-raised state,
2724 * because of this early merge. Store directly the
2725 * weight-raising state that would have been assigned
2726 * to bfqq, so that to avoid that bfqq unjustly fails
2727 * to enjoy weight raising if split soon.
2729 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
2730 bic->saved_wr_start_at_switch_to_srt = bfq_smallest_from_now();
2731 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2732 bic->saved_last_wr_start_finish = jiffies;
2734 bic->saved_wr_coeff = bfqq->wr_coeff;
2735 bic->saved_wr_start_at_switch_to_srt =
2736 bfqq->wr_start_at_switch_to_srt;
2737 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2738 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2742 void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
2745 * To prevent bfqq's service guarantees from being violated,
2746 * bfqq may be left busy, i.e., queued for service, even if
2747 * empty (see comments in __bfq_bfqq_expire() for
2748 * details). But, if no process will send requests to bfqq any
2749 * longer, then there is no point in keeping bfqq queued for
2750 * service. In addition, keeping bfqq queued for service, but
2751 * with no process ref any longer, may have caused bfqq to be
2752 * freed when dequeued from service. But this is assumed to
2755 if (bfq_bfqq_busy(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) &&
2756 bfqq != bfqd->in_service_queue)
2757 bfq_del_bfqq_busy(bfqd, bfqq, false);
2759 bfq_put_queue(bfqq);
2763 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2764 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2766 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2767 (unsigned long)new_bfqq->pid);
2768 /* Save weight raising and idle window of the merged queues */
2769 bfq_bfqq_save_state(bfqq);
2770 bfq_bfqq_save_state(new_bfqq);
2771 if (bfq_bfqq_IO_bound(bfqq))
2772 bfq_mark_bfqq_IO_bound(new_bfqq);
2773 bfq_clear_bfqq_IO_bound(bfqq);
2776 * If bfqq is weight-raised, then let new_bfqq inherit
2777 * weight-raising. To reduce false positives, neglect the case
2778 * where bfqq has just been created, but has not yet made it
2779 * to be weight-raised (which may happen because EQM may merge
2780 * bfqq even before bfq_add_request is executed for the first
2781 * time for bfqq). Handling this case would however be very
2782 * easy, thanks to the flag just_created.
2784 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2785 new_bfqq->wr_coeff = bfqq->wr_coeff;
2786 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2787 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2788 new_bfqq->wr_start_at_switch_to_srt =
2789 bfqq->wr_start_at_switch_to_srt;
2790 if (bfq_bfqq_busy(new_bfqq))
2791 bfqd->wr_busy_queues++;
2792 new_bfqq->entity.prio_changed = 1;
2795 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2797 bfqq->entity.prio_changed = 1;
2798 if (bfq_bfqq_busy(bfqq))
2799 bfqd->wr_busy_queues--;
2802 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2803 bfqd->wr_busy_queues);
2806 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2808 bic_set_bfqq(bic, new_bfqq, 1);
2809 bfq_mark_bfqq_coop(new_bfqq);
2811 * new_bfqq now belongs to at least two bics (it is a shared queue):
2812 * set new_bfqq->bic to NULL. bfqq either:
2813 * - does not belong to any bic any more, and hence bfqq->bic must
2814 * be set to NULL, or
2815 * - is a queue whose owning bics have already been redirected to a
2816 * different queue, hence the queue is destined to not belong to
2817 * any bic soon and bfqq->bic is already NULL (therefore the next
2818 * assignment causes no harm).
2820 new_bfqq->bic = NULL;
2822 * If the queue is shared, the pid is the pid of one of the associated
2823 * processes. Which pid depends on the exact sequence of merge events
2824 * the queue underwent. So printing such a pid is useless and confusing
2825 * because it reports a random pid between those of the associated
2827 * We mark such a queue with a pid -1, and then print SHARED instead of
2828 * a pid in logging messages.
2832 bfq_release_process_ref(bfqd, bfqq);
2835 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2838 struct bfq_data *bfqd = q->elevator->elevator_data;
2839 bool is_sync = op_is_sync(bio->bi_opf);
2840 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2843 * Disallow merge of a sync bio into an async request.
2845 if (is_sync && !rq_is_sync(rq))
2849 * Lookup the bfqq that this bio will be queued with. Allow
2850 * merge only if rq is queued there.
2856 * We take advantage of this function to perform an early merge
2857 * of the queues of possible cooperating processes.
2859 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2862 * bic still points to bfqq, then it has not yet been
2863 * redirected to some other bfq_queue, and a queue
2864 * merge between bfqq and new_bfqq can be safely
2865 * fulfilled, i.e., bic can be redirected to new_bfqq
2866 * and bfqq can be put.
2868 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2871 * If we get here, bio will be queued into new_queue,
2872 * so use new_bfqq to decide whether bio and rq can be
2878 * Change also bqfd->bio_bfqq, as
2879 * bfqd->bio_bic now points to new_bfqq, and
2880 * this function may be invoked again (and then may
2881 * use again bqfd->bio_bfqq).
2883 bfqd->bio_bfqq = bfqq;
2886 return bfqq == RQ_BFQQ(rq);
2890 * Set the maximum time for the in-service queue to consume its
2891 * budget. This prevents seeky processes from lowering the throughput.
2892 * In practice, a time-slice service scheme is used with seeky
2895 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2896 struct bfq_queue *bfqq)
2898 unsigned int timeout_coeff;
2900 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2903 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2905 bfqd->last_budget_start = ktime_get();
2907 bfqq->budget_timeout = jiffies +
2908 bfqd->bfq_timeout * timeout_coeff;
2911 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2912 struct bfq_queue *bfqq)
2915 bfq_clear_bfqq_fifo_expire(bfqq);
2917 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2919 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2920 bfqq->wr_coeff > 1 &&
2921 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2922 time_is_before_jiffies(bfqq->budget_timeout)) {
2924 * For soft real-time queues, move the start
2925 * of the weight-raising period forward by the
2926 * time the queue has not received any
2927 * service. Otherwise, a relatively long
2928 * service delay is likely to cause the
2929 * weight-raising period of the queue to end,
2930 * because of the short duration of the
2931 * weight-raising period of a soft real-time
2932 * queue. It is worth noting that this move
2933 * is not so dangerous for the other queues,
2934 * because soft real-time queues are not
2937 * To not add a further variable, we use the
2938 * overloaded field budget_timeout to
2939 * determine for how long the queue has not
2940 * received service, i.e., how much time has
2941 * elapsed since the queue expired. However,
2942 * this is a little imprecise, because
2943 * budget_timeout is set to jiffies if bfqq
2944 * not only expires, but also remains with no
2947 if (time_after(bfqq->budget_timeout,
2948 bfqq->last_wr_start_finish))
2949 bfqq->last_wr_start_finish +=
2950 jiffies - bfqq->budget_timeout;
2952 bfqq->last_wr_start_finish = jiffies;
2955 bfq_set_budget_timeout(bfqd, bfqq);
2956 bfq_log_bfqq(bfqd, bfqq,
2957 "set_in_service_queue, cur-budget = %d",
2958 bfqq->entity.budget);
2961 bfqd->in_service_queue = bfqq;
2965 * Get and set a new queue for service.
2967 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2969 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2971 __bfq_set_in_service_queue(bfqd, bfqq);
2975 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2977 struct bfq_queue *bfqq = bfqd->in_service_queue;
2980 bfq_mark_bfqq_wait_request(bfqq);
2983 * We don't want to idle for seeks, but we do want to allow
2984 * fair distribution of slice time for a process doing back-to-back
2985 * seeks. So allow a little bit of time for him to submit a new rq.
2987 sl = bfqd->bfq_slice_idle;
2989 * Unless the queue is being weight-raised or the scenario is
2990 * asymmetric, grant only minimum idle time if the queue
2991 * is seeky. A long idling is preserved for a weight-raised
2992 * queue, or, more in general, in an asymmetric scenario,
2993 * because a long idling is needed for guaranteeing to a queue
2994 * its reserved share of the throughput (in particular, it is
2995 * needed if the queue has a higher weight than some other
2998 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
2999 !bfq_asymmetric_scenario(bfqd, bfqq))
3000 sl = min_t(u64, sl, BFQ_MIN_TT);
3001 else if (bfqq->wr_coeff > 1)
3002 sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
3004 bfqd->last_idling_start = ktime_get();
3005 bfqd->last_idling_start_jiffies = jiffies;
3007 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
3009 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
3013 * In autotuning mode, max_budget is dynamically recomputed as the
3014 * amount of sectors transferred in timeout at the estimated peak
3015 * rate. This enables BFQ to utilize a full timeslice with a full
3016 * budget, even if the in-service queue is served at peak rate. And
3017 * this maximises throughput with sequential workloads.
3019 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
3021 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
3022 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
3026 * Update parameters related to throughput and responsiveness, as a
3027 * function of the estimated peak rate. See comments on
3028 * bfq_calc_max_budget(), and on the ref_wr_duration array.
3030 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
3032 if (bfqd->bfq_user_max_budget == 0) {
3033 bfqd->bfq_max_budget =
3034 bfq_calc_max_budget(bfqd);
3035 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
3039 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
3042 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
3043 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
3044 bfqd->peak_rate_samples = 1;
3045 bfqd->sequential_samples = 0;
3046 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
3048 } else /* no new rq dispatched, just reset the number of samples */
3049 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
3052 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
3053 bfqd->peak_rate_samples, bfqd->sequential_samples,
3054 bfqd->tot_sectors_dispatched);
3057 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
3059 u32 rate, weight, divisor;
3062 * For the convergence property to hold (see comments on
3063 * bfq_update_peak_rate()) and for the assessment to be
3064 * reliable, a minimum number of samples must be present, and
3065 * a minimum amount of time must have elapsed. If not so, do
3066 * not compute new rate. Just reset parameters, to get ready
3067 * for a new evaluation attempt.
3069 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
3070 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
3071 goto reset_computation;
3074 * If a new request completion has occurred after last
3075 * dispatch, then, to approximate the rate at which requests
3076 * have been served by the device, it is more precise to
3077 * extend the observation interval to the last completion.
3079 bfqd->delta_from_first =
3080 max_t(u64, bfqd->delta_from_first,
3081 bfqd->last_completion - bfqd->first_dispatch);
3084 * Rate computed in sects/usec, and not sects/nsec, for
3087 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
3088 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
3091 * Peak rate not updated if:
3092 * - the percentage of sequential dispatches is below 3/4 of the
3093 * total, and rate is below the current estimated peak rate
3094 * - rate is unreasonably high (> 20M sectors/sec)
3096 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
3097 rate <= bfqd->peak_rate) ||
3098 rate > 20<<BFQ_RATE_SHIFT)
3099 goto reset_computation;
3102 * We have to update the peak rate, at last! To this purpose,
3103 * we use a low-pass filter. We compute the smoothing constant
3104 * of the filter as a function of the 'weight' of the new
3107 * As can be seen in next formulas, we define this weight as a
3108 * quantity proportional to how sequential the workload is,
3109 * and to how long the observation time interval is.
3111 * The weight runs from 0 to 8. The maximum value of the
3112 * weight, 8, yields the minimum value for the smoothing
3113 * constant. At this minimum value for the smoothing constant,
3114 * the measured rate contributes for half of the next value of
3115 * the estimated peak rate.
3117 * So, the first step is to compute the weight as a function
3118 * of how sequential the workload is. Note that the weight
3119 * cannot reach 9, because bfqd->sequential_samples cannot
3120 * become equal to bfqd->peak_rate_samples, which, in its
3121 * turn, holds true because bfqd->sequential_samples is not
3122 * incremented for the first sample.
3124 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
3127 * Second step: further refine the weight as a function of the
3128 * duration of the observation interval.
3130 weight = min_t(u32, 8,
3131 div_u64(weight * bfqd->delta_from_first,
3132 BFQ_RATE_REF_INTERVAL));
3135 * Divisor ranging from 10, for minimum weight, to 2, for
3138 divisor = 10 - weight;
3141 * Finally, update peak rate:
3143 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
3145 bfqd->peak_rate *= divisor-1;
3146 bfqd->peak_rate /= divisor;
3147 rate /= divisor; /* smoothing constant alpha = 1/divisor */
3149 bfqd->peak_rate += rate;
3152 * For a very slow device, bfqd->peak_rate can reach 0 (see
3153 * the minimum representable values reported in the comments
3154 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3155 * divisions by zero where bfqd->peak_rate is used as a
3158 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
3160 update_thr_responsiveness_params(bfqd);
3163 bfq_reset_rate_computation(bfqd, rq);
3167 * Update the read/write peak rate (the main quantity used for
3168 * auto-tuning, see update_thr_responsiveness_params()).
3170 * It is not trivial to estimate the peak rate (correctly): because of
3171 * the presence of sw and hw queues between the scheduler and the
3172 * device components that finally serve I/O requests, it is hard to
3173 * say exactly when a given dispatched request is served inside the
3174 * device, and for how long. As a consequence, it is hard to know
3175 * precisely at what rate a given set of requests is actually served
3178 * On the opposite end, the dispatch time of any request is trivially
3179 * available, and, from this piece of information, the "dispatch rate"
3180 * of requests can be immediately computed. So, the idea in the next
3181 * function is to use what is known, namely request dispatch times
3182 * (plus, when useful, request completion times), to estimate what is
3183 * unknown, namely in-device request service rate.
3185 * The main issue is that, because of the above facts, the rate at
3186 * which a certain set of requests is dispatched over a certain time
3187 * interval can vary greatly with respect to the rate at which the
3188 * same requests are then served. But, since the size of any
3189 * intermediate queue is limited, and the service scheme is lossless
3190 * (no request is silently dropped), the following obvious convergence
3191 * property holds: the number of requests dispatched MUST become
3192 * closer and closer to the number of requests completed as the
3193 * observation interval grows. This is the key property used in
3194 * the next function to estimate the peak service rate as a function
3195 * of the observed dispatch rate. The function assumes to be invoked
3196 * on every request dispatch.
3198 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
3200 u64 now_ns = ktime_get_ns();
3202 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
3203 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
3204 bfqd->peak_rate_samples);
3205 bfq_reset_rate_computation(bfqd, rq);
3206 goto update_last_values; /* will add one sample */
3210 * Device idle for very long: the observation interval lasting
3211 * up to this dispatch cannot be a valid observation interval
3212 * for computing a new peak rate (similarly to the late-
3213 * completion event in bfq_completed_request()). Go to
3214 * update_rate_and_reset to have the following three steps
3216 * - close the observation interval at the last (previous)
3217 * request dispatch or completion
3218 * - compute rate, if possible, for that observation interval
3219 * - start a new observation interval with this dispatch
3221 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
3222 bfqd->rq_in_driver == 0)
3223 goto update_rate_and_reset;
3225 /* Update sampling information */
3226 bfqd->peak_rate_samples++;
3228 if ((bfqd->rq_in_driver > 0 ||
3229 now_ns - bfqd->last_completion < BFQ_MIN_TT)
3230 && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
3231 bfqd->sequential_samples++;
3233 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
3235 /* Reset max observed rq size every 32 dispatches */
3236 if (likely(bfqd->peak_rate_samples % 32))
3237 bfqd->last_rq_max_size =
3238 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
3240 bfqd->last_rq_max_size = blk_rq_sectors(rq);
3242 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
3244 /* Target observation interval not yet reached, go on sampling */
3245 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
3246 goto update_last_values;
3248 update_rate_and_reset:
3249 bfq_update_rate_reset(bfqd, rq);
3251 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
3252 if (RQ_BFQQ(rq) == bfqd->in_service_queue)
3253 bfqd->in_serv_last_pos = bfqd->last_position;
3254 bfqd->last_dispatch = now_ns;
3258 * Remove request from internal lists.
3260 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
3262 struct bfq_queue *bfqq = RQ_BFQQ(rq);
3265 * For consistency, the next instruction should have been
3266 * executed after removing the request from the queue and
3267 * dispatching it. We execute instead this instruction before
3268 * bfq_remove_request() (and hence introduce a temporary
3269 * inconsistency), for efficiency. In fact, should this
3270 * dispatch occur for a non in-service bfqq, this anticipated
3271 * increment prevents two counters related to bfqq->dispatched
3272 * from risking to be, first, uselessly decremented, and then
3273 * incremented again when the (new) value of bfqq->dispatched
3274 * happens to be taken into account.
3277 bfq_update_peak_rate(q->elevator->elevator_data, rq);
3279 bfq_remove_request(q, rq);
3283 * There is a case where idling does not have to be performed for
3284 * throughput concerns, but to preserve the throughput share of
3285 * the process associated with bfqq.
3287 * To introduce this case, we can note that allowing the drive
3288 * to enqueue more than one request at a time, and hence
3289 * delegating de facto final scheduling decisions to the
3290 * drive's internal scheduler, entails loss of control on the
3291 * actual request service order. In particular, the critical
3292 * situation is when requests from different processes happen
3293 * to be present, at the same time, in the internal queue(s)
3294 * of the drive. In such a situation, the drive, by deciding
3295 * the service order of the internally-queued requests, does
3296 * determine also the actual throughput distribution among
3297 * these processes. But the drive typically has no notion or
3298 * concern about per-process throughput distribution, and
3299 * makes its decisions only on a per-request basis. Therefore,
3300 * the service distribution enforced by the drive's internal
3301 * scheduler is likely to coincide with the desired throughput
3302 * distribution only in a completely symmetric, or favorably
3303 * skewed scenario where:
3304 * (i-a) each of these processes must get the same throughput as
3306 * (i-b) in case (i-a) does not hold, it holds that the process
3307 * associated with bfqq must receive a lower or equal
3308 * throughput than any of the other processes;
3309 * (ii) the I/O of each process has the same properties, in
3310 * terms of locality (sequential or random), direction
3311 * (reads or writes), request sizes, greediness
3312 * (from I/O-bound to sporadic), and so on;
3314 * In fact, in such a scenario, the drive tends to treat the requests
3315 * of each process in about the same way as the requests of the
3316 * others, and thus to provide each of these processes with about the
3317 * same throughput. This is exactly the desired throughput
3318 * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3319 * even more convenient distribution for (the process associated with)
3322 * In contrast, in any asymmetric or unfavorable scenario, device
3323 * idling (I/O-dispatch plugging) is certainly needed to guarantee
3324 * that bfqq receives its assigned fraction of the device throughput
3325 * (see [1] for details).
3327 * The problem is that idling may significantly reduce throughput with
3328 * certain combinations of types of I/O and devices. An important
3329 * example is sync random I/O on flash storage with command
3330 * queueing. So, unless bfqq falls in cases where idling also boosts
3331 * throughput, it is important to check conditions (i-a), i(-b) and
3332 * (ii) accurately, so as to avoid idling when not strictly needed for
3333 * service guarantees.
3335 * Unfortunately, it is extremely difficult to thoroughly check
3336 * condition (ii). And, in case there are active groups, it becomes
3337 * very difficult to check conditions (i-a) and (i-b) too. In fact,
3338 * if there are active groups, then, for conditions (i-a) or (i-b) to
3339 * become false 'indirectly', it is enough that an active group
3340 * contains more active processes or sub-groups than some other active
3341 * group. More precisely, for conditions (i-a) or (i-b) to become
3342 * false because of such a group, it is not even necessary that the
3343 * group is (still) active: it is sufficient that, even if the group
3344 * has become inactive, some of its descendant processes still have
3345 * some request already dispatched but still waiting for
3346 * completion. In fact, requests have still to be guaranteed their
3347 * share of the throughput even after being dispatched. In this
3348 * respect, it is easy to show that, if a group frequently becomes
3349 * inactive while still having in-flight requests, and if, when this
3350 * happens, the group is not considered in the calculation of whether
3351 * the scenario is asymmetric, then the group may fail to be
3352 * guaranteed its fair share of the throughput (basically because
3353 * idling may not be performed for the descendant processes of the
3354 * group, but it had to be). We address this issue with the following
3355 * bi-modal behavior, implemented in the function
3356 * bfq_asymmetric_scenario().
3358 * If there are groups with requests waiting for completion
3359 * (as commented above, some of these groups may even be
3360 * already inactive), then the scenario is tagged as
3361 * asymmetric, conservatively, without checking any of the
3362 * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3363 * This behavior matches also the fact that groups are created
3364 * exactly if controlling I/O is a primary concern (to
3365 * preserve bandwidth and latency guarantees).
3367 * On the opposite end, if there are no groups with requests waiting
3368 * for completion, then only conditions (i-a) and (i-b) are actually
3369 * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3370 * idling is not performed, regardless of whether condition (ii)
3371 * holds. In other words, only if conditions (i-a) and (i-b) do not
3372 * hold, then idling is allowed, and the device tends to be prevented
3373 * from queueing many requests, possibly of several processes. Since
3374 * there are no groups with requests waiting for completion, then, to
3375 * control conditions (i-a) and (i-b) it is enough to check just
3376 * whether all the queues with requests waiting for completion also
3377 * have the same weight.
3379 * Not checking condition (ii) evidently exposes bfqq to the
3380 * risk of getting less throughput than its fair share.
3381 * However, for queues with the same weight, a further
3382 * mechanism, preemption, mitigates or even eliminates this
3383 * problem. And it does so without consequences on overall
3384 * throughput. This mechanism and its benefits are explained
3385 * in the next three paragraphs.
3387 * Even if a queue, say Q, is expired when it remains idle, Q
3388 * can still preempt the new in-service queue if the next
3389 * request of Q arrives soon (see the comments on
3390 * bfq_bfqq_update_budg_for_activation). If all queues and
3391 * groups have the same weight, this form of preemption,
3392 * combined with the hole-recovery heuristic described in the
3393 * comments on function bfq_bfqq_update_budg_for_activation,
3394 * are enough to preserve a correct bandwidth distribution in
3395 * the mid term, even without idling. In fact, even if not
3396 * idling allows the internal queues of the device to contain
3397 * many requests, and thus to reorder requests, we can rather
3398 * safely assume that the internal scheduler still preserves a
3399 * minimum of mid-term fairness.
3401 * More precisely, this preemption-based, idleless approach
3402 * provides fairness in terms of IOPS, and not sectors per
3403 * second. This can be seen with a simple example. Suppose
3404 * that there are two queues with the same weight, but that
3405 * the first queue receives requests of 8 sectors, while the
3406 * second queue receives requests of 1024 sectors. In
3407 * addition, suppose that each of the two queues contains at
3408 * most one request at a time, which implies that each queue
3409 * always remains idle after it is served. Finally, after
3410 * remaining idle, each queue receives very quickly a new
3411 * request. It follows that the two queues are served
3412 * alternatively, preempting each other if needed. This
3413 * implies that, although both queues have the same weight,
3414 * the queue with large requests receives a service that is
3415 * 1024/8 times as high as the service received by the other
3418 * The motivation for using preemption instead of idling (for
3419 * queues with the same weight) is that, by not idling,
3420 * service guarantees are preserved (completely or at least in
3421 * part) without minimally sacrificing throughput. And, if
3422 * there is no active group, then the primary expectation for
3423 * this device is probably a high throughput.
3425 * We are now left only with explaining the two sub-conditions in the
3426 * additional compound condition that is checked below for deciding
3427 * whether the scenario is asymmetric. To explain the first
3428 * sub-condition, we need to add that the function
3429 * bfq_asymmetric_scenario checks the weights of only
3430 * non-weight-raised queues, for efficiency reasons (see comments on
3431 * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3432 * is checked explicitly here. More precisely, the compound condition
3433 * below takes into account also the fact that, even if bfqq is being
3434 * weight-raised, the scenario is still symmetric if all queues with
3435 * requests waiting for completion happen to be
3436 * weight-raised. Actually, we should be even more precise here, and
3437 * differentiate between interactive weight raising and soft real-time
3440 * The second sub-condition checked in the compound condition is
3441 * whether there is a fair amount of already in-flight I/O not
3442 * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3443 * following reason. The drive may decide to serve in-flight
3444 * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3445 * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3446 * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3447 * basically uncontrolled amount of I/O from other queues may be
3448 * dispatched too, possibly causing the service of bfqq's I/O to be
3449 * delayed even longer in the drive. This problem gets more and more
3450 * serious as the speed and the queue depth of the drive grow,
3451 * because, as these two quantities grow, the probability to find no
3452 * queue busy but many requests in flight grows too. By contrast,
3453 * plugging I/O dispatching minimizes the delay induced by already
3454 * in-flight I/O, and enables bfqq to recover the bandwidth it may
3455 * lose because of this delay.
3457 * As a side note, it is worth considering that the above
3458 * device-idling countermeasures may however fail in the following
3459 * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3460 * in a time period during which all symmetry sub-conditions hold, and
3461 * therefore the device is allowed to enqueue many requests, but at
3462 * some later point in time some sub-condition stops to hold, then it
3463 * may become impossible to make requests be served in the desired
3464 * order until all the requests already queued in the device have been
3465 * served. The last sub-condition commented above somewhat mitigates
3466 * this problem for weight-raised queues.
3468 * However, as an additional mitigation for this problem, we preserve
3469 * plugging for a special symmetric case that may suddenly turn into
3470 * asymmetric: the case where only bfqq is busy. In this case, not
3471 * expiring bfqq does not cause any harm to any other queues in terms
3472 * of service guarantees. In contrast, it avoids the following unlucky
3473 * sequence of events: (1) bfqq is expired, (2) a new queue with a
3474 * lower weight than bfqq becomes busy (or more queues), (3) the new
3475 * queue is served until a new request arrives for bfqq, (4) when bfqq
3476 * is finally served, there are so many requests of the new queue in
3477 * the drive that the pending requests for bfqq take a lot of time to
3478 * be served. In particular, event (2) may case even already
3479 * dispatched requests of bfqq to be delayed, inside the drive. So, to
3480 * avoid this series of events, the scenario is preventively declared
3481 * as asymmetric also if bfqq is the only busy queues
3483 static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3484 struct bfq_queue *bfqq)
3486 int tot_busy_queues = bfq_tot_busy_queues(bfqd);
3488 /* No point in idling for bfqq if it won't get requests any longer */
3489 if (unlikely(!bfqq_process_refs(bfqq)))
3492 return (bfqq->wr_coeff > 1 &&
3493 (bfqd->wr_busy_queues <
3495 bfqd->rq_in_driver >=
3496 bfqq->dispatched + 4)) ||
3497 bfq_asymmetric_scenario(bfqd, bfqq) ||
3498 tot_busy_queues == 1;
3501 static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3502 enum bfqq_expiration reason)
3505 * If this bfqq is shared between multiple processes, check
3506 * to make sure that those processes are still issuing I/Os
3507 * within the mean seek distance. If not, it may be time to
3508 * break the queues apart again.
3510 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
3511 bfq_mark_bfqq_split_coop(bfqq);
3514 * Consider queues with a higher finish virtual time than
3515 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3516 * true, then bfqq's bandwidth would be violated if an
3517 * uncontrolled amount of I/O from these queues were
3518 * dispatched while bfqq is waiting for its new I/O to
3519 * arrive. This is exactly what may happen if this is a forced
3520 * expiration caused by a preemption attempt, and if bfqq is
3521 * not re-scheduled. To prevent this from happening, re-queue
3522 * bfqq if it needs I/O-dispatch plugging, even if it is
3523 * empty. By doing so, bfqq is granted to be served before the
3524 * above queues (provided that bfqq is of course eligible).
3526 if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
3527 !(reason == BFQQE_PREEMPTED &&
3528 idling_needed_for_service_guarantees(bfqd, bfqq))) {
3529 if (bfqq->dispatched == 0)
3531 * Overloading budget_timeout field to store
3532 * the time at which the queue remains with no
3533 * backlog and no outstanding request; used by
3534 * the weight-raising mechanism.
3536 bfqq->budget_timeout = jiffies;
3538 bfq_del_bfqq_busy(bfqd, bfqq, true);
3540 bfq_requeue_bfqq(bfqd, bfqq, true);
3542 * Resort priority tree of potential close cooperators.
3543 * See comments on bfq_pos_tree_add_move() for the unlikely().
3545 if (unlikely(!bfqd->nonrot_with_queueing &&
3546 !RB_EMPTY_ROOT(&bfqq->sort_list)))
3547 bfq_pos_tree_add_move(bfqd, bfqq);
3551 * All in-service entities must have been properly deactivated
3552 * or requeued before executing the next function, which
3553 * resets all in-service entities as no more in service. This
3554 * may cause bfqq to be freed. If this happens, the next
3555 * function returns true.
3557 return __bfq_bfqd_reset_in_service(bfqd);
3561 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3562 * @bfqd: device data.
3563 * @bfqq: queue to update.
3564 * @reason: reason for expiration.
3566 * Handle the feedback on @bfqq budget at queue expiration.
3567 * See the body for detailed comments.
3569 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
3570 struct bfq_queue *bfqq,
3571 enum bfqq_expiration reason)
3573 struct request *next_rq;
3574 int budget, min_budget;
3576 min_budget = bfq_min_budget(bfqd);
3578 if (bfqq->wr_coeff == 1)
3579 budget = bfqq->max_budget;
3581 * Use a constant, low budget for weight-raised queues,
3582 * to help achieve a low latency. Keep it slightly higher
3583 * than the minimum possible budget, to cause a little
3584 * bit fewer expirations.
3586 budget = 2 * min_budget;
3588 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
3589 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
3590 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
3591 budget, bfq_min_budget(bfqd));
3592 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
3593 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
3595 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
3598 * Caveat: in all the following cases we trade latency
3601 case BFQQE_TOO_IDLE:
3603 * This is the only case where we may reduce
3604 * the budget: if there is no request of the
3605 * process still waiting for completion, then
3606 * we assume (tentatively) that the timer has
3607 * expired because the batch of requests of
3608 * the process could have been served with a
3609 * smaller budget. Hence, betting that
3610 * process will behave in the same way when it
3611 * becomes backlogged again, we reduce its
3612 * next budget. As long as we guess right,
3613 * this budget cut reduces the latency
3614 * experienced by the process.
3616 * However, if there are still outstanding
3617 * requests, then the process may have not yet
3618 * issued its next request just because it is
3619 * still waiting for the completion of some of
3620 * the still outstanding ones. So in this
3621 * subcase we do not reduce its budget, on the
3622 * contrary we increase it to possibly boost
3623 * the throughput, as discussed in the
3624 * comments to the BUDGET_TIMEOUT case.
3626 if (bfqq->dispatched > 0) /* still outstanding reqs */
3627 budget = min(budget * 2, bfqd->bfq_max_budget);
3629 if (budget > 5 * min_budget)
3630 budget -= 4 * min_budget;
3632 budget = min_budget;
3635 case BFQQE_BUDGET_TIMEOUT:
3637 * We double the budget here because it gives
3638 * the chance to boost the throughput if this
3639 * is not a seeky process (and has bumped into
3640 * this timeout because of, e.g., ZBR).
3642 budget = min(budget * 2, bfqd->bfq_max_budget);
3644 case BFQQE_BUDGET_EXHAUSTED:
3646 * The process still has backlog, and did not
3647 * let either the budget timeout or the disk
3648 * idling timeout expire. Hence it is not
3649 * seeky, has a short thinktime and may be
3650 * happy with a higher budget too. So
3651 * definitely increase the budget of this good
3652 * candidate to boost the disk throughput.
3654 budget = min(budget * 4, bfqd->bfq_max_budget);
3656 case BFQQE_NO_MORE_REQUESTS:
3658 * For queues that expire for this reason, it
3659 * is particularly important to keep the
3660 * budget close to the actual service they
3661 * need. Doing so reduces the timestamp
3662 * misalignment problem described in the
3663 * comments in the body of
3664 * __bfq_activate_entity. In fact, suppose
3665 * that a queue systematically expires for
3666 * BFQQE_NO_MORE_REQUESTS and presents a
3667 * new request in time to enjoy timestamp
3668 * back-shifting. The larger the budget of the
3669 * queue is with respect to the service the
3670 * queue actually requests in each service
3671 * slot, the more times the queue can be
3672 * reactivated with the same virtual finish
3673 * time. It follows that, even if this finish
3674 * time is pushed to the system virtual time
3675 * to reduce the consequent timestamp
3676 * misalignment, the queue unjustly enjoys for
3677 * many re-activations a lower finish time
3678 * than all newly activated queues.
3680 * The service needed by bfqq is measured
3681 * quite precisely by bfqq->entity.service.
3682 * Since bfqq does not enjoy device idling,
3683 * bfqq->entity.service is equal to the number
3684 * of sectors that the process associated with
3685 * bfqq requested to read/write before waiting
3686 * for request completions, or blocking for
3689 budget = max_t(int, bfqq->entity.service, min_budget);
3694 } else if (!bfq_bfqq_sync(bfqq)) {
3696 * Async queues get always the maximum possible
3697 * budget, as for them we do not care about latency
3698 * (in addition, their ability to dispatch is limited
3699 * by the charging factor).
3701 budget = bfqd->bfq_max_budget;
3704 bfqq->max_budget = budget;
3706 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
3707 !bfqd->bfq_user_max_budget)
3708 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
3711 * If there is still backlog, then assign a new budget, making
3712 * sure that it is large enough for the next request. Since
3713 * the finish time of bfqq must be kept in sync with the
3714 * budget, be sure to call __bfq_bfqq_expire() *after* this
3717 * If there is no backlog, then no need to update the budget;
3718 * it will be updated on the arrival of a new request.
3720 next_rq = bfqq->next_rq;
3722 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
3723 bfq_serv_to_charge(next_rq, bfqq));
3725 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
3726 next_rq ? blk_rq_sectors(next_rq) : 0,
3727 bfqq->entity.budget);
3731 * Return true if the process associated with bfqq is "slow". The slow
3732 * flag is used, in addition to the budget timeout, to reduce the
3733 * amount of service provided to seeky processes, and thus reduce
3734 * their chances to lower the throughput. More details in the comments
3735 * on the function bfq_bfqq_expire().
3737 * An important observation is in order: as discussed in the comments
3738 * on the function bfq_update_peak_rate(), with devices with internal
3739 * queues, it is hard if ever possible to know when and for how long
3740 * an I/O request is processed by the device (apart from the trivial
3741 * I/O pattern where a new request is dispatched only after the
3742 * previous one has been completed). This makes it hard to evaluate
3743 * the real rate at which the I/O requests of each bfq_queue are
3744 * served. In fact, for an I/O scheduler like BFQ, serving a
3745 * bfq_queue means just dispatching its requests during its service
3746 * slot (i.e., until the budget of the queue is exhausted, or the
3747 * queue remains idle, or, finally, a timeout fires). But, during the
3748 * service slot of a bfq_queue, around 100 ms at most, the device may
3749 * be even still processing requests of bfq_queues served in previous
3750 * service slots. On the opposite end, the requests of the in-service
3751 * bfq_queue may be completed after the service slot of the queue
3754 * Anyway, unless more sophisticated solutions are used
3755 * (where possible), the sum of the sizes of the requests dispatched
3756 * during the service slot of a bfq_queue is probably the only
3757 * approximation available for the service received by the bfq_queue
3758 * during its service slot. And this sum is the quantity used in this
3759 * function to evaluate the I/O speed of a process.
3761 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3762 bool compensate, enum bfqq_expiration reason,
3763 unsigned long *delta_ms)
3765 ktime_t delta_ktime;
3767 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
3769 if (!bfq_bfqq_sync(bfqq))
3773 delta_ktime = bfqd->last_idling_start;
3775 delta_ktime = ktime_get();
3776 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
3777 delta_usecs = ktime_to_us(delta_ktime);
3779 /* don't use too short time intervals */
3780 if (delta_usecs < 1000) {
3781 if (blk_queue_nonrot(bfqd->queue))
3783 * give same worst-case guarantees as idling
3786 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
3787 else /* charge at least one seek */
3788 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
3793 *delta_ms = delta_usecs / USEC_PER_MSEC;
3796 * Use only long (> 20ms) intervals to filter out excessive
3797 * spikes in service rate estimation.
3799 if (delta_usecs > 20000) {
3801 * Caveat for rotational devices: processes doing I/O
3802 * in the slower disk zones tend to be slow(er) even
3803 * if not seeky. In this respect, the estimated peak
3804 * rate is likely to be an average over the disk
3805 * surface. Accordingly, to not be too harsh with
3806 * unlucky processes, a process is deemed slow only if
3807 * its rate has been lower than half of the estimated
3810 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
3813 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
3819 * To be deemed as soft real-time, an application must meet two
3820 * requirements. First, the application must not require an average
3821 * bandwidth higher than the approximate bandwidth required to playback or
3822 * record a compressed high-definition video.
3823 * The next function is invoked on the completion of the last request of a
3824 * batch, to compute the next-start time instant, soft_rt_next_start, such
3825 * that, if the next request of the application does not arrive before
3826 * soft_rt_next_start, then the above requirement on the bandwidth is met.
3828 * The second requirement is that the request pattern of the application is
3829 * isochronous, i.e., that, after issuing a request or a batch of requests,
3830 * the application stops issuing new requests until all its pending requests
3831 * have been completed. After that, the application may issue a new batch,
3833 * For this reason the next function is invoked to compute
3834 * soft_rt_next_start only for applications that meet this requirement,
3835 * whereas soft_rt_next_start is set to infinity for applications that do
3838 * Unfortunately, even a greedy (i.e., I/O-bound) application may
3839 * happen to meet, occasionally or systematically, both the above
3840 * bandwidth and isochrony requirements. This may happen at least in
3841 * the following circumstances. First, if the CPU load is high. The
3842 * application may stop issuing requests while the CPUs are busy
3843 * serving other processes, then restart, then stop again for a while,
3844 * and so on. The other circumstances are related to the storage
3845 * device: the storage device is highly loaded or reaches a low-enough
3846 * throughput with the I/O of the application (e.g., because the I/O
3847 * is random and/or the device is slow). In all these cases, the
3848 * I/O of the application may be simply slowed down enough to meet
3849 * the bandwidth and isochrony requirements. To reduce the probability
3850 * that greedy applications are deemed as soft real-time in these
3851 * corner cases, a further rule is used in the computation of
3852 * soft_rt_next_start: the return value of this function is forced to
3853 * be higher than the maximum between the following two quantities.
3855 * (a) Current time plus: (1) the maximum time for which the arrival
3856 * of a request is waited for when a sync queue becomes idle,
3857 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
3858 * postpone for a moment the reason for adding a few extra
3859 * jiffies; we get back to it after next item (b). Lower-bounding
3860 * the return value of this function with the current time plus
3861 * bfqd->bfq_slice_idle tends to filter out greedy applications,
3862 * because the latter issue their next request as soon as possible
3863 * after the last one has been completed. In contrast, a soft
3864 * real-time application spends some time processing data, after a
3865 * batch of its requests has been completed.
3867 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
3868 * above, greedy applications may happen to meet both the
3869 * bandwidth and isochrony requirements under heavy CPU or
3870 * storage-device load. In more detail, in these scenarios, these
3871 * applications happen, only for limited time periods, to do I/O
3872 * slowly enough to meet all the requirements described so far,
3873 * including the filtering in above item (a). These slow-speed
3874 * time intervals are usually interspersed between other time
3875 * intervals during which these applications do I/O at a very high
3876 * speed. Fortunately, exactly because of the high speed of the
3877 * I/O in the high-speed intervals, the values returned by this
3878 * function happen to be so high, near the end of any such
3879 * high-speed interval, to be likely to fall *after* the end of
3880 * the low-speed time interval that follows. These high values are
3881 * stored in bfqq->soft_rt_next_start after each invocation of
3882 * this function. As a consequence, if the last value of
3883 * bfqq->soft_rt_next_start is constantly used to lower-bound the
3884 * next value that this function may return, then, from the very
3885 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
3886 * likely to be constantly kept so high that any I/O request
3887 * issued during the low-speed interval is considered as arriving
3888 * to soon for the application to be deemed as soft
3889 * real-time. Then, in the high-speed interval that follows, the
3890 * application will not be deemed as soft real-time, just because
3891 * it will do I/O at a high speed. And so on.
3893 * Getting back to the filtering in item (a), in the following two
3894 * cases this filtering might be easily passed by a greedy
3895 * application, if the reference quantity was just
3896 * bfqd->bfq_slice_idle:
3897 * 1) HZ is so low that the duration of a jiffy is comparable to or
3898 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
3899 * devices with HZ=100. The time granularity may be so coarse
3900 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
3901 * is rather lower than the exact value.
3902 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
3903 * for a while, then suddenly 'jump' by several units to recover the lost
3904 * increments. This seems to happen, e.g., inside virtual machines.
3905 * To address this issue, in the filtering in (a) we do not use as a
3906 * reference time interval just bfqd->bfq_slice_idle, but
3907 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
3908 * minimum number of jiffies for which the filter seems to be quite
3909 * precise also in embedded systems and KVM/QEMU virtual machines.
3911 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
3912 struct bfq_queue *bfqq)
3914 return max3(bfqq->soft_rt_next_start,
3915 bfqq->last_idle_bklogged +
3916 HZ * bfqq->service_from_backlogged /
3917 bfqd->bfq_wr_max_softrt_rate,
3918 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
3922 * bfq_bfqq_expire - expire a queue.
3923 * @bfqd: device owning the queue.
3924 * @bfqq: the queue to expire.
3925 * @compensate: if true, compensate for the time spent idling.
3926 * @reason: the reason causing the expiration.
3928 * If the process associated with bfqq does slow I/O (e.g., because it
3929 * issues random requests), we charge bfqq with the time it has been
3930 * in service instead of the service it has received (see
3931 * bfq_bfqq_charge_time for details on how this goal is achieved). As
3932 * a consequence, bfqq will typically get higher timestamps upon
3933 * reactivation, and hence it will be rescheduled as if it had
3934 * received more service than what it has actually received. In the
3935 * end, bfqq receives less service in proportion to how slowly its
3936 * associated process consumes its budgets (and hence how seriously it
3937 * tends to lower the throughput). In addition, this time-charging
3938 * strategy guarantees time fairness among slow processes. In
3939 * contrast, if the process associated with bfqq is not slow, we
3940 * charge bfqq exactly with the service it has received.
3942 * Charging time to the first type of queues and the exact service to
3943 * the other has the effect of using the WF2Q+ policy to schedule the
3944 * former on a timeslice basis, without violating service domain
3945 * guarantees among the latter.
3947 void bfq_bfqq_expire(struct bfq_data *bfqd,
3948 struct bfq_queue *bfqq,
3950 enum bfqq_expiration reason)
3953 unsigned long delta = 0;
3954 struct bfq_entity *entity = &bfqq->entity;
3957 * Check whether the process is slow (see bfq_bfqq_is_slow).
3959 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
3962 * As above explained, charge slow (typically seeky) and
3963 * timed-out queues with the time and not the service
3964 * received, to favor sequential workloads.
3966 * Processes doing I/O in the slower disk zones will tend to
3967 * be slow(er) even if not seeky. Therefore, since the
3968 * estimated peak rate is actually an average over the disk
3969 * surface, these processes may timeout just for bad luck. To
3970 * avoid punishing them, do not charge time to processes that
3971 * succeeded in consuming at least 2/3 of their budget. This
3972 * allows BFQ to preserve enough elasticity to still perform
3973 * bandwidth, and not time, distribution with little unlucky
3974 * or quasi-sequential processes.
3976 if (bfqq->wr_coeff == 1 &&
3978 (reason == BFQQE_BUDGET_TIMEOUT &&
3979 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
3980 bfq_bfqq_charge_time(bfqd, bfqq, delta);
3982 if (reason == BFQQE_TOO_IDLE &&
3983 entity->service <= 2 * entity->budget / 10)
3984 bfq_clear_bfqq_IO_bound(bfqq);
3986 if (bfqd->low_latency && bfqq->wr_coeff == 1)
3987 bfqq->last_wr_start_finish = jiffies;
3989 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
3990 RB_EMPTY_ROOT(&bfqq->sort_list)) {
3992 * If we get here, and there are no outstanding
3993 * requests, then the request pattern is isochronous
3994 * (see the comments on the function
3995 * bfq_bfqq_softrt_next_start()). Therefore we can
3996 * compute soft_rt_next_start.
3998 * If, instead, the queue still has outstanding
3999 * requests, then we have to wait for the completion
4000 * of all the outstanding requests to discover whether
4001 * the request pattern is actually isochronous.
4003 if (bfqq->dispatched == 0)
4004 bfqq->soft_rt_next_start =
4005 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4006 else if (bfqq->dispatched > 0) {
4008 * Schedule an update of soft_rt_next_start to when
4009 * the task may be discovered to be isochronous.
4011 bfq_mark_bfqq_softrt_update(bfqq);
4015 bfq_log_bfqq(bfqd, bfqq,
4016 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
4017 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
4020 * bfqq expired, so no total service time needs to be computed
4021 * any longer: reset state machine for measuring total service
4024 bfqd->rqs_injected = bfqd->wait_dispatch = false;
4025 bfqd->waited_rq = NULL;
4028 * Increase, decrease or leave budget unchanged according to
4031 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
4032 if (__bfq_bfqq_expire(bfqd, bfqq, reason))
4033 /* bfqq is gone, no more actions on it */
4036 /* mark bfqq as waiting a request only if a bic still points to it */
4037 if (!bfq_bfqq_busy(bfqq) &&
4038 reason != BFQQE_BUDGET_TIMEOUT &&
4039 reason != BFQQE_BUDGET_EXHAUSTED) {
4040 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
4042 * Not setting service to 0, because, if the next rq
4043 * arrives in time, the queue will go on receiving
4044 * service with this same budget (as if it never expired)
4047 entity->service = 0;
4050 * Reset the received-service counter for every parent entity.
4051 * Differently from what happens with bfqq->entity.service,
4052 * the resetting of this counter never needs to be postponed
4053 * for parent entities. In fact, in case bfqq may have a
4054 * chance to go on being served using the last, partially
4055 * consumed budget, bfqq->entity.service needs to be kept,
4056 * because if bfqq then actually goes on being served using
4057 * the same budget, the last value of bfqq->entity.service is
4058 * needed to properly decrement bfqq->entity.budget by the
4059 * portion already consumed. In contrast, it is not necessary
4060 * to keep entity->service for parent entities too, because
4061 * the bubble up of the new value of bfqq->entity.budget will
4062 * make sure that the budgets of parent entities are correct,
4063 * even in case bfqq and thus parent entities go on receiving
4064 * service with the same budget.
4066 entity = entity->parent;
4067 for_each_entity(entity)
4068 entity->service = 0;
4072 * Budget timeout is not implemented through a dedicated timer, but
4073 * just checked on request arrivals and completions, as well as on
4074 * idle timer expirations.
4076 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
4078 return time_is_before_eq_jiffies(bfqq->budget_timeout);
4082 * If we expire a queue that is actively waiting (i.e., with the
4083 * device idled) for the arrival of a new request, then we may incur
4084 * the timestamp misalignment problem described in the body of the
4085 * function __bfq_activate_entity. Hence we return true only if this
4086 * condition does not hold, or if the queue is slow enough to deserve
4087 * only to be kicked off for preserving a high throughput.
4089 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
4091 bfq_log_bfqq(bfqq->bfqd, bfqq,
4092 "may_budget_timeout: wait_request %d left %d timeout %d",
4093 bfq_bfqq_wait_request(bfqq),
4094 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
4095 bfq_bfqq_budget_timeout(bfqq));
4097 return (!bfq_bfqq_wait_request(bfqq) ||
4098 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
4100 bfq_bfqq_budget_timeout(bfqq);
4103 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
4104 struct bfq_queue *bfqq)
4106 bool rot_without_queueing =
4107 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
4108 bfqq_sequential_and_IO_bound,
4111 /* No point in idling for bfqq if it won't get requests any longer */
4112 if (unlikely(!bfqq_process_refs(bfqq)))
4115 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
4116 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
4119 * The next variable takes into account the cases where idling
4120 * boosts the throughput.
4122 * The value of the variable is computed considering, first, that
4123 * idling is virtually always beneficial for the throughput if:
4124 * (a) the device is not NCQ-capable and rotational, or
4125 * (b) regardless of the presence of NCQ, the device is rotational and
4126 * the request pattern for bfqq is I/O-bound and sequential, or
4127 * (c) regardless of whether it is rotational, the device is
4128 * not NCQ-capable and the request pattern for bfqq is
4129 * I/O-bound and sequential.
4131 * Secondly, and in contrast to the above item (b), idling an
4132 * NCQ-capable flash-based device would not boost the
4133 * throughput even with sequential I/O; rather it would lower
4134 * the throughput in proportion to how fast the device
4135 * is. Accordingly, the next variable is true if any of the
4136 * above conditions (a), (b) or (c) is true, and, in
4137 * particular, happens to be false if bfqd is an NCQ-capable
4138 * flash-based device.
4140 idling_boosts_thr = rot_without_queueing ||
4141 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
4142 bfqq_sequential_and_IO_bound);
4145 * The return value of this function is equal to that of
4146 * idling_boosts_thr, unless a special case holds. In this
4147 * special case, described below, idling may cause problems to
4148 * weight-raised queues.
4150 * When the request pool is saturated (e.g., in the presence
4151 * of write hogs), if the processes associated with
4152 * non-weight-raised queues ask for requests at a lower rate,
4153 * then processes associated with weight-raised queues have a
4154 * higher probability to get a request from the pool
4155 * immediately (or at least soon) when they need one. Thus
4156 * they have a higher probability to actually get a fraction
4157 * of the device throughput proportional to their high
4158 * weight. This is especially true with NCQ-capable drives,
4159 * which enqueue several requests in advance, and further
4160 * reorder internally-queued requests.
4162 * For this reason, we force to false the return value if
4163 * there are weight-raised busy queues. In this case, and if
4164 * bfqq is not weight-raised, this guarantees that the device
4165 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4166 * then idling will be guaranteed by another variable, see
4167 * below). Combined with the timestamping rules of BFQ (see
4168 * [1] for details), this behavior causes bfqq, and hence any
4169 * sync non-weight-raised queue, to get a lower number of
4170 * requests served, and thus to ask for a lower number of
4171 * requests from the request pool, before the busy
4172 * weight-raised queues get served again. This often mitigates
4173 * starvation problems in the presence of heavy write
4174 * workloads and NCQ, thereby guaranteeing a higher
4175 * application and system responsiveness in these hostile
4178 return idling_boosts_thr &&
4179 bfqd->wr_busy_queues == 0;
4183 * For a queue that becomes empty, device idling is allowed only if
4184 * this function returns true for that queue. As a consequence, since
4185 * device idling plays a critical role for both throughput boosting
4186 * and service guarantees, the return value of this function plays a
4187 * critical role as well.
4189 * In a nutshell, this function returns true only if idling is
4190 * beneficial for throughput or, even if detrimental for throughput,
4191 * idling is however necessary to preserve service guarantees (low
4192 * latency, desired throughput distribution, ...). In particular, on
4193 * NCQ-capable devices, this function tries to return false, so as to
4194 * help keep the drives' internal queues full, whenever this helps the
4195 * device boost the throughput without causing any service-guarantee
4198 * Most of the issues taken into account to get the return value of
4199 * this function are not trivial. We discuss these issues in the two
4200 * functions providing the main pieces of information needed by this
4203 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
4205 struct bfq_data *bfqd = bfqq->bfqd;
4206 bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
4208 /* No point in idling for bfqq if it won't get requests any longer */
4209 if (unlikely(!bfqq_process_refs(bfqq)))
4212 if (unlikely(bfqd->strict_guarantees))
4216 * Idling is performed only if slice_idle > 0. In addition, we
4219 * (b) bfqq is in the idle io prio class: in this case we do
4220 * not idle because we want to minimize the bandwidth that
4221 * queues in this class can steal to higher-priority queues
4223 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
4224 bfq_class_idle(bfqq))
4227 idling_boosts_thr_with_no_issue =
4228 idling_boosts_thr_without_issues(bfqd, bfqq);
4230 idling_needed_for_service_guar =
4231 idling_needed_for_service_guarantees(bfqd, bfqq);
4234 * We have now the two components we need to compute the
4235 * return value of the function, which is true only if idling
4236 * either boosts the throughput (without issues), or is
4237 * necessary to preserve service guarantees.
4239 return idling_boosts_thr_with_no_issue ||
4240 idling_needed_for_service_guar;
4244 * If the in-service queue is empty but the function bfq_better_to_idle
4245 * returns true, then:
4246 * 1) the queue must remain in service and cannot be expired, and
4247 * 2) the device must be idled to wait for the possible arrival of a new
4248 * request for the queue.
4249 * See the comments on the function bfq_better_to_idle for the reasons
4250 * why performing device idling is the best choice to boost the throughput
4251 * and preserve service guarantees when bfq_better_to_idle itself
4254 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
4256 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
4260 * This function chooses the queue from which to pick the next extra
4261 * I/O request to inject, if it finds a compatible queue. See the
4262 * comments on bfq_update_inject_limit() for details on the injection
4263 * mechanism, and for the definitions of the quantities mentioned
4266 static struct bfq_queue *
4267 bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
4269 struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
4270 unsigned int limit = in_serv_bfqq->inject_limit;
4273 * - bfqq is not weight-raised and therefore does not carry
4274 * time-critical I/O,
4276 * - regardless of whether bfqq is weight-raised, bfqq has
4277 * however a long think time, during which it can absorb the
4278 * effect of an appropriate number of extra I/O requests
4279 * from other queues (see bfq_update_inject_limit for
4280 * details on the computation of this number);
4281 * then injection can be performed without restrictions.
4283 bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
4284 !bfq_bfqq_has_short_ttime(in_serv_bfqq);
4288 * - the baseline total service time could not be sampled yet,
4289 * so the inject limit happens to be still 0, and
4290 * - a lot of time has elapsed since the plugging of I/O
4291 * dispatching started, so drive speed is being wasted
4293 * then temporarily raise inject limit to one request.
4295 if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
4296 bfq_bfqq_wait_request(in_serv_bfqq) &&
4297 time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
4298 bfqd->bfq_slice_idle)
4302 if (bfqd->rq_in_driver >= limit)
4306 * Linear search of the source queue for injection; but, with
4307 * a high probability, very few steps are needed to find a
4308 * candidate queue, i.e., a queue with enough budget left for
4309 * its next request. In fact:
4310 * - BFQ dynamically updates the budget of every queue so as
4311 * to accommodate the expected backlog of the queue;
4312 * - if a queue gets all its requests dispatched as injected
4313 * service, then the queue is removed from the active list
4314 * (and re-added only if it gets new requests, but then it
4315 * is assigned again enough budget for its new backlog).
4317 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
4318 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4319 (in_serv_always_inject || bfqq->wr_coeff > 1) &&
4320 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4321 bfq_bfqq_budget_left(bfqq)) {
4323 * Allow for only one large in-flight request
4324 * on non-rotational devices, for the
4325 * following reason. On non-rotationl drives,
4326 * large requests take much longer than
4327 * smaller requests to be served. In addition,
4328 * the drive prefers to serve large requests
4329 * w.r.t. to small ones, if it can choose. So,
4330 * having more than one large requests queued
4331 * in the drive may easily make the next first
4332 * request of the in-service queue wait for so
4333 * long to break bfqq's service guarantees. On
4334 * the bright side, large requests let the
4335 * drive reach a very high throughput, even if
4336 * there is only one in-flight large request
4339 if (blk_queue_nonrot(bfqd->queue) &&
4340 blk_rq_sectors(bfqq->next_rq) >=
4341 BFQQ_SECT_THR_NONROT)
4342 limit = min_t(unsigned int, 1, limit);
4344 limit = in_serv_bfqq->inject_limit;
4346 if (bfqd->rq_in_driver < limit) {
4347 bfqd->rqs_injected = true;
4356 * Select a queue for service. If we have a current queue in service,
4357 * check whether to continue servicing it, or retrieve and set a new one.
4359 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
4361 struct bfq_queue *bfqq;
4362 struct request *next_rq;
4363 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
4365 bfqq = bfqd->in_service_queue;
4369 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
4372 * Do not expire bfqq for budget timeout if bfqq may be about
4373 * to enjoy device idling. The reason why, in this case, we
4374 * prevent bfqq from expiring is the same as in the comments
4375 * on the case where bfq_bfqq_must_idle() returns true, in
4376 * bfq_completed_request().
4378 if (bfq_may_expire_for_budg_timeout(bfqq) &&
4379 !bfq_bfqq_must_idle(bfqq))
4384 * This loop is rarely executed more than once. Even when it
4385 * happens, it is much more convenient to re-execute this loop
4386 * than to return NULL and trigger a new dispatch to get a
4389 next_rq = bfqq->next_rq;
4391 * If bfqq has requests queued and it has enough budget left to
4392 * serve them, keep the queue, otherwise expire it.
4395 if (bfq_serv_to_charge(next_rq, bfqq) >
4396 bfq_bfqq_budget_left(bfqq)) {
4398 * Expire the queue for budget exhaustion,
4399 * which makes sure that the next budget is
4400 * enough to serve the next request, even if
4401 * it comes from the fifo expired path.
4403 reason = BFQQE_BUDGET_EXHAUSTED;
4407 * The idle timer may be pending because we may
4408 * not disable disk idling even when a new request
4411 if (bfq_bfqq_wait_request(bfqq)) {
4413 * If we get here: 1) at least a new request
4414 * has arrived but we have not disabled the
4415 * timer because the request was too small,
4416 * 2) then the block layer has unplugged
4417 * the device, causing the dispatch to be
4420 * Since the device is unplugged, now the
4421 * requests are probably large enough to
4422 * provide a reasonable throughput.
4423 * So we disable idling.
4425 bfq_clear_bfqq_wait_request(bfqq);
4426 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4433 * No requests pending. However, if the in-service queue is idling
4434 * for a new request, or has requests waiting for a completion and
4435 * may idle after their completion, then keep it anyway.
4437 * Yet, inject service from other queues if it boosts
4438 * throughput and is possible.
4440 if (bfq_bfqq_wait_request(bfqq) ||
4441 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
4442 struct bfq_queue *async_bfqq =
4443 bfqq->bic && bfqq->bic->bfqq[0] &&
4444 bfq_bfqq_busy(bfqq->bic->bfqq[0]) &&
4445 bfqq->bic->bfqq[0]->next_rq ?
4446 bfqq->bic->bfqq[0] : NULL;
4449 * The next three mutually-exclusive ifs decide
4450 * whether to try injection, and choose the queue to
4451 * pick an I/O request from.
4453 * The first if checks whether the process associated
4454 * with bfqq has also async I/O pending. If so, it
4455 * injects such I/O unconditionally. Injecting async
4456 * I/O from the same process can cause no harm to the
4457 * process. On the contrary, it can only increase
4458 * bandwidth and reduce latency for the process.
4460 * The second if checks whether there happens to be a
4461 * non-empty waker queue for bfqq, i.e., a queue whose
4462 * I/O needs to be completed for bfqq to receive new
4463 * I/O. This happens, e.g., if bfqq is associated with
4464 * a process that does some sync. A sync generates
4465 * extra blocking I/O, which must be completed before
4466 * the process associated with bfqq can go on with its
4467 * I/O. If the I/O of the waker queue is not served,
4468 * then bfqq remains empty, and no I/O is dispatched,
4469 * until the idle timeout fires for bfqq. This is
4470 * likely to result in lower bandwidth and higher
4471 * latencies for bfqq, and in a severe loss of total
4472 * throughput. The best action to take is therefore to
4473 * serve the waker queue as soon as possible. So do it
4474 * (without relying on the third alternative below for
4475 * eventually serving waker_bfqq's I/O; see the last
4476 * paragraph for further details). This systematic
4477 * injection of I/O from the waker queue does not
4478 * cause any delay to bfqq's I/O. On the contrary,
4479 * next bfqq's I/O is brought forward dramatically,
4480 * for it is not blocked for milliseconds.
4482 * The third if checks whether bfqq is a queue for
4483 * which it is better to avoid injection. It is so if
4484 * bfqq delivers more throughput when served without
4485 * any further I/O from other queues in the middle, or
4486 * if the service times of bfqq's I/O requests both
4487 * count more than overall throughput, and may be
4488 * easily increased by injection (this happens if bfqq
4489 * has a short think time). If none of these
4490 * conditions holds, then a candidate queue for
4491 * injection is looked for through
4492 * bfq_choose_bfqq_for_injection(). Note that the
4493 * latter may return NULL (for example if the inject
4494 * limit for bfqq is currently 0).
4496 * NOTE: motivation for the second alternative
4498 * Thanks to the way the inject limit is updated in
4499 * bfq_update_has_short_ttime(), it is rather likely
4500 * that, if I/O is being plugged for bfqq and the
4501 * waker queue has pending I/O requests that are
4502 * blocking bfqq's I/O, then the third alternative
4503 * above lets the waker queue get served before the
4504 * I/O-plugging timeout fires. So one may deem the
4505 * second alternative superfluous. It is not, because
4506 * the third alternative may be way less effective in
4507 * case of a synchronization. For two main
4508 * reasons. First, throughput may be low because the
4509 * inject limit may be too low to guarantee the same
4510 * amount of injected I/O, from the waker queue or
4511 * other queues, that the second alternative
4512 * guarantees (the second alternative unconditionally
4513 * injects a pending I/O request of the waker queue
4514 * for each bfq_dispatch_request()). Second, with the
4515 * third alternative, the duration of the plugging,
4516 * i.e., the time before bfqq finally receives new I/O,
4517 * may not be minimized, because the waker queue may
4518 * happen to be served only after other queues.
4521 icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
4522 bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
4523 bfq_bfqq_budget_left(async_bfqq))
4524 bfqq = bfqq->bic->bfqq[0];
4525 else if (bfq_bfqq_has_waker(bfqq) &&
4526 bfq_bfqq_busy(bfqq->waker_bfqq) &&
4527 bfqq->waker_bfqq->next_rq &&
4528 bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
4529 bfqq->waker_bfqq) <=
4530 bfq_bfqq_budget_left(bfqq->waker_bfqq)
4532 bfqq = bfqq->waker_bfqq;
4533 else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
4534 (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
4535 !bfq_bfqq_has_short_ttime(bfqq)))
4536 bfqq = bfq_choose_bfqq_for_injection(bfqd);
4543 reason = BFQQE_NO_MORE_REQUESTS;
4545 bfq_bfqq_expire(bfqd, bfqq, false, reason);
4547 bfqq = bfq_set_in_service_queue(bfqd);
4549 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
4554 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
4556 bfq_log(bfqd, "select_queue: no queue returned");
4561 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4563 struct bfq_entity *entity = &bfqq->entity;
4565 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
4566 bfq_log_bfqq(bfqd, bfqq,
4567 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
4568 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
4569 jiffies_to_msecs(bfqq->wr_cur_max_time),
4571 bfqq->entity.weight, bfqq->entity.orig_weight);
4573 if (entity->prio_changed)
4574 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
4577 * If the queue was activated in a burst, or too much
4578 * time has elapsed from the beginning of this
4579 * weight-raising period, then end weight raising.
4581 if (bfq_bfqq_in_large_burst(bfqq))
4582 bfq_bfqq_end_wr(bfqq);
4583 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
4584 bfqq->wr_cur_max_time)) {
4585 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
4586 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
4587 bfq_wr_duration(bfqd))) {
4589 * Either in interactive weight
4590 * raising, or in soft_rt weight
4592 * interactive-weight-raising period
4593 * elapsed (so no switch back to
4594 * interactive weight raising).
4596 bfq_bfqq_end_wr(bfqq);
4598 * soft_rt finishing while still in
4599 * interactive period, switch back to
4600 * interactive weight raising
4602 switch_back_to_interactive_wr(bfqq, bfqd);
4603 bfqq->entity.prio_changed = 1;
4606 if (bfqq->wr_coeff > 1 &&
4607 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
4608 bfqq->service_from_wr > max_service_from_wr) {
4609 /* see comments on max_service_from_wr */
4610 bfq_bfqq_end_wr(bfqq);
4614 * To improve latency (for this or other queues), immediately
4615 * update weight both if it must be raised and if it must be
4616 * lowered. Since, entity may be on some active tree here, and
4617 * might have a pending change of its ioprio class, invoke
4618 * next function with the last parameter unset (see the
4619 * comments on the function).
4621 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
4622 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
4627 * Dispatch next request from bfqq.
4629 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
4630 struct bfq_queue *bfqq)
4632 struct request *rq = bfqq->next_rq;
4633 unsigned long service_to_charge;
4635 service_to_charge = bfq_serv_to_charge(rq, bfqq);
4637 bfq_bfqq_served(bfqq, service_to_charge);
4639 if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
4640 bfqd->wait_dispatch = false;
4641 bfqd->waited_rq = rq;
4644 bfq_dispatch_remove(bfqd->queue, rq);
4646 if (bfqq != bfqd->in_service_queue)
4650 * If weight raising has to terminate for bfqq, then next
4651 * function causes an immediate update of bfqq's weight,
4652 * without waiting for next activation. As a consequence, on
4653 * expiration, bfqq will be timestamped as if has never been
4654 * weight-raised during this service slot, even if it has
4655 * received part or even most of the service as a
4656 * weight-raised queue. This inflates bfqq's timestamps, which
4657 * is beneficial, as bfqq is then more willing to leave the
4658 * device immediately to possible other weight-raised queues.
4660 bfq_update_wr_data(bfqd, bfqq);
4663 * Expire bfqq, pretending that its budget expired, if bfqq
4664 * belongs to CLASS_IDLE and other queues are waiting for
4667 if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
4670 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
4676 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
4678 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4681 * Avoiding lock: a race on bfqd->busy_queues should cause at
4682 * most a call to dispatch for nothing
4684 return !list_empty_careful(&bfqd->dispatch) ||
4685 bfq_tot_busy_queues(bfqd) > 0;
4688 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
4690 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4691 struct request *rq = NULL;
4692 struct bfq_queue *bfqq = NULL;
4694 if (!list_empty(&bfqd->dispatch)) {
4695 rq = list_first_entry(&bfqd->dispatch, struct request,
4697 list_del_init(&rq->queuelist);
4703 * Increment counters here, because this
4704 * dispatch does not follow the standard
4705 * dispatch flow (where counters are
4710 goto inc_in_driver_start_rq;
4714 * We exploit the bfq_finish_requeue_request hook to
4715 * decrement rq_in_driver, but
4716 * bfq_finish_requeue_request will not be invoked on
4717 * this request. So, to avoid unbalance, just start
4718 * this request, without incrementing rq_in_driver. As
4719 * a negative consequence, rq_in_driver is deceptively
4720 * lower than it should be while this request is in
4721 * service. This may cause bfq_schedule_dispatch to be
4722 * invoked uselessly.
4724 * As for implementing an exact solution, the
4725 * bfq_finish_requeue_request hook, if defined, is
4726 * probably invoked also on this request. So, by
4727 * exploiting this hook, we could 1) increment
4728 * rq_in_driver here, and 2) decrement it in
4729 * bfq_finish_requeue_request. Such a solution would
4730 * let the value of the counter be always accurate,
4731 * but it would entail using an extra interface
4732 * function. This cost seems higher than the benefit,
4733 * being the frequency of non-elevator-private
4734 * requests very low.
4739 bfq_log(bfqd, "dispatch requests: %d busy queues",
4740 bfq_tot_busy_queues(bfqd));
4742 if (bfq_tot_busy_queues(bfqd) == 0)
4746 * Force device to serve one request at a time if
4747 * strict_guarantees is true. Forcing this service scheme is
4748 * currently the ONLY way to guarantee that the request
4749 * service order enforced by the scheduler is respected by a
4750 * queueing device. Otherwise the device is free even to make
4751 * some unlucky request wait for as long as the device
4754 * Of course, serving one request at a time may cause loss of
4757 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
4760 bfqq = bfq_select_queue(bfqd);
4764 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
4767 inc_in_driver_start_rq:
4768 bfqd->rq_in_driver++;
4770 rq->rq_flags |= RQF_STARTED;
4776 #ifdef CONFIG_BFQ_CGROUP_DEBUG
4777 static void bfq_update_dispatch_stats(struct request_queue *q,
4779 struct bfq_queue *in_serv_queue,
4780 bool idle_timer_disabled)
4782 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
4784 if (!idle_timer_disabled && !bfqq)
4788 * rq and bfqq are guaranteed to exist until this function
4789 * ends, for the following reasons. First, rq can be
4790 * dispatched to the device, and then can be completed and
4791 * freed, only after this function ends. Second, rq cannot be
4792 * merged (and thus freed because of a merge) any longer,
4793 * because it has already started. Thus rq cannot be freed
4794 * before this function ends, and, since rq has a reference to
4795 * bfqq, the same guarantee holds for bfqq too.
4797 * In addition, the following queue lock guarantees that
4798 * bfqq_group(bfqq) exists as well.
4800 spin_lock_irq(&q->queue_lock);
4801 if (idle_timer_disabled)
4803 * Since the idle timer has been disabled,
4804 * in_serv_queue contained some request when
4805 * __bfq_dispatch_request was invoked above, which
4806 * implies that rq was picked exactly from
4807 * in_serv_queue. Thus in_serv_queue == bfqq, and is
4808 * therefore guaranteed to exist because of the above
4811 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
4813 struct bfq_group *bfqg = bfqq_group(bfqq);
4815 bfqg_stats_update_avg_queue_size(bfqg);
4816 bfqg_stats_set_start_empty_time(bfqg);
4817 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
4819 spin_unlock_irq(&q->queue_lock);
4822 static inline void bfq_update_dispatch_stats(struct request_queue *q,
4824 struct bfq_queue *in_serv_queue,
4825 bool idle_timer_disabled) {}
4826 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
4828 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
4830 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4832 struct bfq_queue *in_serv_queue;
4833 bool waiting_rq, idle_timer_disabled;
4835 spin_lock_irq(&bfqd->lock);
4837 in_serv_queue = bfqd->in_service_queue;
4838 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
4840 rq = __bfq_dispatch_request(hctx);
4842 idle_timer_disabled =
4843 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
4845 spin_unlock_irq(&bfqd->lock);
4847 bfq_update_dispatch_stats(hctx->queue, rq, in_serv_queue,
4848 idle_timer_disabled);
4854 * Task holds one reference to the queue, dropped when task exits. Each rq
4855 * in-flight on this queue also holds a reference, dropped when rq is freed.
4857 * Scheduler lock must be held here. Recall not to use bfqq after calling
4858 * this function on it.
4860 void bfq_put_queue(struct bfq_queue *bfqq)
4862 struct bfq_queue *item;
4863 struct hlist_node *n;
4864 struct bfq_group *bfqg = bfqq_group(bfqq);
4867 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
4874 if (!hlist_unhashed(&bfqq->burst_list_node)) {
4875 hlist_del_init(&bfqq->burst_list_node);
4877 * Decrement also burst size after the removal, if the
4878 * process associated with bfqq is exiting, and thus
4879 * does not contribute to the burst any longer. This
4880 * decrement helps filter out false positives of large
4881 * bursts, when some short-lived process (often due to
4882 * the execution of commands by some service) happens
4883 * to start and exit while a complex application is
4884 * starting, and thus spawning several processes that
4885 * do I/O (and that *must not* be treated as a large
4886 * burst, see comments on bfq_handle_burst).
4888 * In particular, the decrement is performed only if:
4889 * 1) bfqq is not a merged queue, because, if it is,
4890 * then this free of bfqq is not triggered by the exit
4891 * of the process bfqq is associated with, but exactly
4892 * by the fact that bfqq has just been merged.
4893 * 2) burst_size is greater than 0, to handle
4894 * unbalanced decrements. Unbalanced decrements may
4895 * happen in te following case: bfqq is inserted into
4896 * the current burst list--without incrementing
4897 * bust_size--because of a split, but the current
4898 * burst list is not the burst list bfqq belonged to
4899 * (see comments on the case of a split in
4902 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
4903 bfqq->bfqd->burst_size--;
4907 * bfqq does not exist any longer, so it cannot be woken by
4908 * any other queue, and cannot wake any other queue. Then bfqq
4909 * must be removed from the woken list of its possible waker
4910 * queue, and all queues in the woken list of bfqq must stop
4911 * having a waker queue. Strictly speaking, these updates
4912 * should be performed when bfqq remains with no I/O source
4913 * attached to it, which happens before bfqq gets freed. In
4914 * particular, this happens when the last process associated
4915 * with bfqq exits or gets associated with a different
4916 * queue. However, both events lead to bfqq being freed soon,
4917 * and dangling references would come out only after bfqq gets
4918 * freed. So these updates are done here, as a simple and safe
4919 * way to handle all cases.
4921 /* remove bfqq from woken list */
4922 if (!hlist_unhashed(&bfqq->woken_list_node))
4923 hlist_del_init(&bfqq->woken_list_node);
4925 /* reset waker for all queues in woken list */
4926 hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
4928 item->waker_bfqq = NULL;
4929 bfq_clear_bfqq_has_waker(item);
4930 hlist_del_init(&item->woken_list_node);
4933 if (bfqq->bfqd && bfqq->bfqd->last_completed_rq_bfqq == bfqq)
4934 bfqq->bfqd->last_completed_rq_bfqq = NULL;
4936 kmem_cache_free(bfq_pool, bfqq);
4937 bfqg_and_blkg_put(bfqg);
4940 static void bfq_put_cooperator(struct bfq_queue *bfqq)
4942 struct bfq_queue *__bfqq, *next;
4945 * If this queue was scheduled to merge with another queue, be
4946 * sure to drop the reference taken on that queue (and others in
4947 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
4949 __bfqq = bfqq->new_bfqq;
4953 next = __bfqq->new_bfqq;
4954 bfq_put_queue(__bfqq);
4959 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4961 if (bfqq == bfqd->in_service_queue) {
4962 __bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT);
4963 bfq_schedule_dispatch(bfqd);
4966 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
4968 bfq_put_cooperator(bfqq);
4970 bfq_release_process_ref(bfqd, bfqq);
4973 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
4975 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4976 struct bfq_data *bfqd;
4979 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
4982 unsigned long flags;
4984 spin_lock_irqsave(&bfqd->lock, flags);
4986 bfq_exit_bfqq(bfqd, bfqq);
4987 bic_set_bfqq(bic, NULL, is_sync);
4988 spin_unlock_irqrestore(&bfqd->lock, flags);
4992 static void bfq_exit_icq(struct io_cq *icq)
4994 struct bfq_io_cq *bic = icq_to_bic(icq);
4996 bfq_exit_icq_bfqq(bic, true);
4997 bfq_exit_icq_bfqq(bic, false);
5001 * Update the entity prio values; note that the new values will not
5002 * be used until the next (re)activation.
5005 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
5007 struct task_struct *tsk = current;
5009 struct bfq_data *bfqd = bfqq->bfqd;
5014 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5015 switch (ioprio_class) {
5017 pr_err("bdi %s: bfq: bad prio class %d\n",
5018 bdi_dev_name(bfqq->bfqd->queue->backing_dev_info),
5021 case IOPRIO_CLASS_NONE:
5023 * No prio set, inherit CPU scheduling settings.
5025 bfqq->new_ioprio = task_nice_ioprio(tsk);
5026 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
5028 case IOPRIO_CLASS_RT:
5029 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5030 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
5032 case IOPRIO_CLASS_BE:
5033 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5034 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
5036 case IOPRIO_CLASS_IDLE:
5037 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
5038 bfqq->new_ioprio = 7;
5042 if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
5043 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
5045 bfqq->new_ioprio = IOPRIO_BE_NR;
5048 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
5049 bfq_log_bfqq(bfqd, bfqq, "new_ioprio %d new_weight %d",
5050 bfqq->new_ioprio, bfqq->entity.new_weight);
5051 bfqq->entity.prio_changed = 1;
5054 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5055 struct bio *bio, bool is_sync,
5056 struct bfq_io_cq *bic);
5058 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
5060 struct bfq_data *bfqd = bic_to_bfqd(bic);
5061 struct bfq_queue *bfqq;
5062 int ioprio = bic->icq.ioc->ioprio;
5065 * This condition may trigger on a newly created bic, be sure to
5066 * drop the lock before returning.
5068 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
5071 bic->ioprio = ioprio;
5073 bfqq = bic_to_bfqq(bic, false);
5075 bfq_release_process_ref(bfqd, bfqq);
5076 bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
5077 bic_set_bfqq(bic, bfqq, false);
5080 bfqq = bic_to_bfqq(bic, true);
5082 bfq_set_next_ioprio_data(bfqq, bic);
5085 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5086 struct bfq_io_cq *bic, pid_t pid, int is_sync)
5088 RB_CLEAR_NODE(&bfqq->entity.rb_node);
5089 INIT_LIST_HEAD(&bfqq->fifo);
5090 INIT_HLIST_NODE(&bfqq->burst_list_node);
5091 INIT_HLIST_NODE(&bfqq->woken_list_node);
5092 INIT_HLIST_HEAD(&bfqq->woken_list);
5098 bfq_set_next_ioprio_data(bfqq, bic);
5102 * No need to mark as has_short_ttime if in
5103 * idle_class, because no device idling is performed
5104 * for queues in idle class
5106 if (!bfq_class_idle(bfqq))
5107 /* tentatively mark as has_short_ttime */
5108 bfq_mark_bfqq_has_short_ttime(bfqq);
5109 bfq_mark_bfqq_sync(bfqq);
5110 bfq_mark_bfqq_just_created(bfqq);
5112 bfq_clear_bfqq_sync(bfqq);
5114 /* set end request to minus infinity from now */
5115 bfqq->ttime.last_end_request = ktime_get_ns() + 1;
5117 bfq_mark_bfqq_IO_bound(bfqq);
5121 /* Tentative initial value to trade off between thr and lat */
5122 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
5123 bfqq->budget_timeout = bfq_smallest_from_now();
5126 bfqq->last_wr_start_finish = jiffies;
5127 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
5128 bfqq->split_time = bfq_smallest_from_now();
5131 * To not forget the possibly high bandwidth consumed by a
5132 * process/queue in the recent past,
5133 * bfq_bfqq_softrt_next_start() returns a value at least equal
5134 * to the current value of bfqq->soft_rt_next_start (see
5135 * comments on bfq_bfqq_softrt_next_start). Set
5136 * soft_rt_next_start to now, to mean that bfqq has consumed
5137 * no bandwidth so far.
5139 bfqq->soft_rt_next_start = jiffies;
5141 /* first request is almost certainly seeky */
5142 bfqq->seek_history = 1;
5145 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
5146 struct bfq_group *bfqg,
5147 int ioprio_class, int ioprio)
5149 switch (ioprio_class) {
5150 case IOPRIO_CLASS_RT:
5151 return &bfqg->async_bfqq[0][ioprio];
5152 case IOPRIO_CLASS_NONE:
5153 ioprio = IOPRIO_NORM;
5155 case IOPRIO_CLASS_BE:
5156 return &bfqg->async_bfqq[1][ioprio];
5157 case IOPRIO_CLASS_IDLE:
5158 return &bfqg->async_idle_bfqq;
5164 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5165 struct bio *bio, bool is_sync,
5166 struct bfq_io_cq *bic)
5168 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5169 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5170 struct bfq_queue **async_bfqq = NULL;
5171 struct bfq_queue *bfqq;
5172 struct bfq_group *bfqg;
5176 bfqg = bfq_find_set_group(bfqd, __bio_blkcg(bio));
5178 bfqq = &bfqd->oom_bfqq;
5183 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
5190 bfqq = kmem_cache_alloc_node(bfq_pool,
5191 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
5195 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
5197 bfq_init_entity(&bfqq->entity, bfqg);
5198 bfq_log_bfqq(bfqd, bfqq, "allocated");
5200 bfqq = &bfqd->oom_bfqq;
5201 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
5206 * Pin the queue now that it's allocated, scheduler exit will
5211 * Extra group reference, w.r.t. sync
5212 * queue. This extra reference is removed
5213 * only if bfqq->bfqg disappears, to
5214 * guarantee that this queue is not freed
5215 * until its group goes away.
5217 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
5223 bfqq->ref++; /* get a process reference to this queue */
5224 bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
5229 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
5230 struct bfq_queue *bfqq)
5232 struct bfq_ttime *ttime = &bfqq->ttime;
5233 u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
5235 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
5237 ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
5238 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
5239 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
5240 ttime->ttime_samples);
5244 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5247 bfqq->seek_history <<= 1;
5248 bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
5250 if (bfqq->wr_coeff > 1 &&
5251 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
5252 BFQQ_TOTALLY_SEEKY(bfqq))
5253 bfq_bfqq_end_wr(bfqq);
5256 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
5257 struct bfq_queue *bfqq,
5258 struct bfq_io_cq *bic)
5260 bool has_short_ttime = true, state_changed;
5263 * No need to update has_short_ttime if bfqq is async or in
5264 * idle io prio class, or if bfq_slice_idle is zero, because
5265 * no device idling is performed for bfqq in this case.
5267 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
5268 bfqd->bfq_slice_idle == 0)
5271 /* Idle window just restored, statistics are meaningless. */
5272 if (time_is_after_eq_jiffies(bfqq->split_time +
5273 bfqd->bfq_wr_min_idle_time))
5276 /* Think time is infinite if no process is linked to
5277 * bfqq. Otherwise check average think time to decide whether
5278 * to mark as has_short_ttime. To this goal, compare average
5279 * think time with half the I/O-plugging timeout.
5281 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
5282 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
5283 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle>>1))
5284 has_short_ttime = false;
5286 state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
5288 if (has_short_ttime)
5289 bfq_mark_bfqq_has_short_ttime(bfqq);
5291 bfq_clear_bfqq_has_short_ttime(bfqq);
5294 * Until the base value for the total service time gets
5295 * finally computed for bfqq, the inject limit does depend on
5296 * the think-time state (short|long). In particular, the limit
5297 * is 0 or 1 if the think time is deemed, respectively, as
5298 * short or long (details in the comments in
5299 * bfq_update_inject_limit()). Accordingly, the next
5300 * instructions reset the inject limit if the think-time state
5301 * has changed and the above base value is still to be
5304 * However, the reset is performed only if more than 100 ms
5305 * have elapsed since the last update of the inject limit, or
5306 * (inclusive) if the change is from short to long think
5307 * time. The reason for this waiting is as follows.
5309 * bfqq may have a long think time because of a
5310 * synchronization with some other queue, i.e., because the
5311 * I/O of some other queue may need to be completed for bfqq
5312 * to receive new I/O. Details in the comments on the choice
5313 * of the queue for injection in bfq_select_queue().
5315 * As stressed in those comments, if such a synchronization is
5316 * actually in place, then, without injection on bfqq, the
5317 * blocking I/O cannot happen to served while bfqq is in
5318 * service. As a consequence, if bfqq is granted
5319 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
5320 * is dispatched, until the idle timeout fires. This is likely
5321 * to result in lower bandwidth and higher latencies for bfqq,
5322 * and in a severe loss of total throughput.
5324 * On the opposite end, a non-zero inject limit may allow the
5325 * I/O that blocks bfqq to be executed soon, and therefore
5326 * bfqq to receive new I/O soon.
5328 * But, if the blocking gets actually eliminated, then the
5329 * next think-time sample for bfqq may be very low. This in
5330 * turn may cause bfqq's think time to be deemed
5331 * short. Without the 100 ms barrier, this new state change
5332 * would cause the body of the next if to be executed
5333 * immediately. But this would set to 0 the inject
5334 * limit. Without injection, the blocking I/O would cause the
5335 * think time of bfqq to become long again, and therefore the
5336 * inject limit to be raised again, and so on. The only effect
5337 * of such a steady oscillation between the two think-time
5338 * states would be to prevent effective injection on bfqq.
5340 * In contrast, if the inject limit is not reset during such a
5341 * long time interval as 100 ms, then the number of short
5342 * think time samples can grow significantly before the reset
5343 * is performed. As a consequence, the think time state can
5344 * become stable before the reset. Therefore there will be no
5345 * state change when the 100 ms elapse, and no reset of the
5346 * inject limit. The inject limit remains steadily equal to 1
5347 * both during and after the 100 ms. So injection can be
5348 * performed at all times, and throughput gets boosted.
5350 * An inject limit equal to 1 is however in conflict, in
5351 * general, with the fact that the think time of bfqq is
5352 * short, because injection may be likely to delay bfqq's I/O
5353 * (as explained in the comments in
5354 * bfq_update_inject_limit()). But this does not happen in
5355 * this special case, because bfqq's low think time is due to
5356 * an effective handling of a synchronization, through
5357 * injection. In this special case, bfqq's I/O does not get
5358 * delayed by injection; on the contrary, bfqq's I/O is
5359 * brought forward, because it is not blocked for
5362 * In addition, serving the blocking I/O much sooner, and much
5363 * more frequently than once per I/O-plugging timeout, makes
5364 * it much quicker to detect a waker queue (the concept of
5365 * waker queue is defined in the comments in
5366 * bfq_add_request()). This makes it possible to start sooner
5367 * to boost throughput more effectively, by injecting the I/O
5368 * of the waker queue unconditionally on every
5369 * bfq_dispatch_request().
5371 * One last, important benefit of not resetting the inject
5372 * limit before 100 ms is that, during this time interval, the
5373 * base value for the total service time is likely to get
5374 * finally computed for bfqq, freeing the inject limit from
5375 * its relation with the think time.
5377 if (state_changed && bfqq->last_serv_time_ns == 0 &&
5378 (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
5379 msecs_to_jiffies(100)) ||
5381 bfq_reset_inject_limit(bfqd, bfqq);
5385 * Called when a new fs request (rq) is added to bfqq. Check if there's
5386 * something we should do about it.
5388 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5391 if (rq->cmd_flags & REQ_META)
5392 bfqq->meta_pending++;
5394 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
5396 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
5397 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
5398 blk_rq_sectors(rq) < 32;
5399 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
5402 * There is just this request queued: if
5403 * - the request is small, and
5404 * - we are idling to boost throughput, and
5405 * - the queue is not to be expired,
5408 * In this way, if the device is being idled to wait
5409 * for a new request from the in-service queue, we
5410 * avoid unplugging the device and committing the
5411 * device to serve just a small request. In contrast
5412 * we wait for the block layer to decide when to
5413 * unplug the device: hopefully, new requests will be
5414 * merged to this one quickly, then the device will be
5415 * unplugged and larger requests will be dispatched.
5417 if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
5422 * A large enough request arrived, or idling is being
5423 * performed to preserve service guarantees, or
5424 * finally the queue is to be expired: in all these
5425 * cases disk idling is to be stopped, so clear
5426 * wait_request flag and reset timer.
5428 bfq_clear_bfqq_wait_request(bfqq);
5429 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
5432 * The queue is not empty, because a new request just
5433 * arrived. Hence we can safely expire the queue, in
5434 * case of budget timeout, without risking that the
5435 * timestamps of the queue are not updated correctly.
5436 * See [1] for more details.
5439 bfq_bfqq_expire(bfqd, bfqq, false,
5440 BFQQE_BUDGET_TIMEOUT);
5444 /* returns true if it causes the idle timer to be disabled */
5445 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
5447 struct bfq_queue *bfqq = RQ_BFQQ(rq),
5448 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
5449 bool waiting, idle_timer_disabled = false;
5453 * Release the request's reference to the old bfqq
5454 * and make sure one is taken to the shared queue.
5456 new_bfqq->allocated++;
5460 * If the bic associated with the process
5461 * issuing this request still points to bfqq
5462 * (and thus has not been already redirected
5463 * to new_bfqq or even some other bfq_queue),
5464 * then complete the merge and redirect it to
5467 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
5468 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
5471 bfq_clear_bfqq_just_created(bfqq);
5473 * rq is about to be enqueued into new_bfqq,
5474 * release rq reference on bfqq
5476 bfq_put_queue(bfqq);
5477 rq->elv.priv[1] = new_bfqq;
5481 bfq_update_io_thinktime(bfqd, bfqq);
5482 bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
5483 bfq_update_io_seektime(bfqd, bfqq, rq);
5485 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
5486 bfq_add_request(rq);
5487 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
5489 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
5490 list_add_tail(&rq->queuelist, &bfqq->fifo);
5492 bfq_rq_enqueued(bfqd, bfqq, rq);
5494 return idle_timer_disabled;
5497 #ifdef CONFIG_BFQ_CGROUP_DEBUG
5498 static void bfq_update_insert_stats(struct request_queue *q,
5499 struct bfq_queue *bfqq,
5500 bool idle_timer_disabled,
5501 unsigned int cmd_flags)
5507 * bfqq still exists, because it can disappear only after
5508 * either it is merged with another queue, or the process it
5509 * is associated with exits. But both actions must be taken by
5510 * the same process currently executing this flow of
5513 * In addition, the following queue lock guarantees that
5514 * bfqq_group(bfqq) exists as well.
5516 spin_lock_irq(&q->queue_lock);
5517 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
5518 if (idle_timer_disabled)
5519 bfqg_stats_update_idle_time(bfqq_group(bfqq));
5520 spin_unlock_irq(&q->queue_lock);
5523 static inline void bfq_update_insert_stats(struct request_queue *q,
5524 struct bfq_queue *bfqq,
5525 bool idle_timer_disabled,
5526 unsigned int cmd_flags) {}
5527 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
5529 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
5532 struct request_queue *q = hctx->queue;
5533 struct bfq_data *bfqd = q->elevator->elevator_data;
5534 struct bfq_queue *bfqq;
5535 bool idle_timer_disabled = false;
5536 unsigned int cmd_flags;
5538 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5539 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) && rq->bio)
5540 bfqg_stats_update_legacy_io(q, rq);
5542 spin_lock_irq(&bfqd->lock);
5543 if (blk_mq_sched_try_insert_merge(q, rq)) {
5544 spin_unlock_irq(&bfqd->lock);
5548 spin_unlock_irq(&bfqd->lock);
5550 blk_mq_sched_request_inserted(rq);
5552 spin_lock_irq(&bfqd->lock);
5553 bfqq = bfq_init_rq(rq);
5554 if (!bfqq || at_head || blk_rq_is_passthrough(rq)) {
5556 list_add(&rq->queuelist, &bfqd->dispatch);
5558 list_add_tail(&rq->queuelist, &bfqd->dispatch);
5560 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
5562 * Update bfqq, because, if a queue merge has occurred
5563 * in __bfq_insert_request, then rq has been
5564 * redirected into a new queue.
5568 if (rq_mergeable(rq)) {
5569 elv_rqhash_add(q, rq);
5576 * Cache cmd_flags before releasing scheduler lock, because rq
5577 * may disappear afterwards (for example, because of a request
5580 cmd_flags = rq->cmd_flags;
5582 spin_unlock_irq(&bfqd->lock);
5584 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
5588 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
5589 struct list_head *list, bool at_head)
5591 while (!list_empty(list)) {
5594 rq = list_first_entry(list, struct request, queuelist);
5595 list_del_init(&rq->queuelist);
5596 bfq_insert_request(hctx, rq, at_head);
5600 static void bfq_update_hw_tag(struct bfq_data *bfqd)
5602 struct bfq_queue *bfqq = bfqd->in_service_queue;
5604 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
5605 bfqd->rq_in_driver);
5607 if (bfqd->hw_tag == 1)
5611 * This sample is valid if the number of outstanding requests
5612 * is large enough to allow a queueing behavior. Note that the
5613 * sum is not exact, as it's not taking into account deactivated
5616 if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
5620 * If active queue hasn't enough requests and can idle, bfq might not
5621 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
5624 if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
5625 bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
5626 BFQ_HW_QUEUE_THRESHOLD &&
5627 bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
5630 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
5633 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
5634 bfqd->max_rq_in_driver = 0;
5635 bfqd->hw_tag_samples = 0;
5637 bfqd->nonrot_with_queueing =
5638 blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
5641 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
5646 bfq_update_hw_tag(bfqd);
5648 bfqd->rq_in_driver--;
5651 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
5653 * Set budget_timeout (which we overload to store the
5654 * time at which the queue remains with no backlog and
5655 * no outstanding request; used by the weight-raising
5658 bfqq->budget_timeout = jiffies;
5660 bfq_weights_tree_remove(bfqd, bfqq);
5663 now_ns = ktime_get_ns();
5665 bfqq->ttime.last_end_request = now_ns;
5668 * Using us instead of ns, to get a reasonable precision in
5669 * computing rate in next check.
5671 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
5674 * If the request took rather long to complete, and, according
5675 * to the maximum request size recorded, this completion latency
5676 * implies that the request was certainly served at a very low
5677 * rate (less than 1M sectors/sec), then the whole observation
5678 * interval that lasts up to this time instant cannot be a
5679 * valid time interval for computing a new peak rate. Invoke
5680 * bfq_update_rate_reset to have the following three steps
5682 * - close the observation interval at the last (previous)
5683 * request dispatch or completion
5684 * - compute rate, if possible, for that observation interval
5685 * - reset to zero samples, which will trigger a proper
5686 * re-initialization of the observation interval on next
5689 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
5690 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
5691 1UL<<(BFQ_RATE_SHIFT - 10))
5692 bfq_update_rate_reset(bfqd, NULL);
5693 bfqd->last_completion = now_ns;
5694 bfqd->last_completed_rq_bfqq = bfqq;
5697 * If we are waiting to discover whether the request pattern
5698 * of the task associated with the queue is actually
5699 * isochronous, and both requisites for this condition to hold
5700 * are now satisfied, then compute soft_rt_next_start (see the
5701 * comments on the function bfq_bfqq_softrt_next_start()). We
5702 * do not compute soft_rt_next_start if bfqq is in interactive
5703 * weight raising (see the comments in bfq_bfqq_expire() for
5704 * an explanation). We schedule this delayed update when bfqq
5705 * expires, if it still has in-flight requests.
5707 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
5708 RB_EMPTY_ROOT(&bfqq->sort_list) &&
5709 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
5710 bfqq->soft_rt_next_start =
5711 bfq_bfqq_softrt_next_start(bfqd, bfqq);
5714 * If this is the in-service queue, check if it needs to be expired,
5715 * or if we want to idle in case it has no pending requests.
5717 if (bfqd->in_service_queue == bfqq) {
5718 if (bfq_bfqq_must_idle(bfqq)) {
5719 if (bfqq->dispatched == 0)
5720 bfq_arm_slice_timer(bfqd);
5722 * If we get here, we do not expire bfqq, even
5723 * if bfqq was in budget timeout or had no
5724 * more requests (as controlled in the next
5725 * conditional instructions). The reason for
5726 * not expiring bfqq is as follows.
5728 * Here bfqq->dispatched > 0 holds, but
5729 * bfq_bfqq_must_idle() returned true. This
5730 * implies that, even if no request arrives
5731 * for bfqq before bfqq->dispatched reaches 0,
5732 * bfqq will, however, not be expired on the
5733 * completion event that causes bfqq->dispatch
5734 * to reach zero. In contrast, on this event,
5735 * bfqq will start enjoying device idling
5736 * (I/O-dispatch plugging).
5738 * But, if we expired bfqq here, bfqq would
5739 * not have the chance to enjoy device idling
5740 * when bfqq->dispatched finally reaches
5741 * zero. This would expose bfqq to violation
5742 * of its reserved service guarantees.
5745 } else if (bfq_may_expire_for_budg_timeout(bfqq))
5746 bfq_bfqq_expire(bfqd, bfqq, false,
5747 BFQQE_BUDGET_TIMEOUT);
5748 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
5749 (bfqq->dispatched == 0 ||
5750 !bfq_better_to_idle(bfqq)))
5751 bfq_bfqq_expire(bfqd, bfqq, false,
5752 BFQQE_NO_MORE_REQUESTS);
5755 if (!bfqd->rq_in_driver)
5756 bfq_schedule_dispatch(bfqd);
5759 static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq)
5763 bfq_put_queue(bfqq);
5767 * The processes associated with bfqq may happen to generate their
5768 * cumulative I/O at a lower rate than the rate at which the device
5769 * could serve the same I/O. This is rather probable, e.g., if only
5770 * one process is associated with bfqq and the device is an SSD. It
5771 * results in bfqq becoming often empty while in service. In this
5772 * respect, if BFQ is allowed to switch to another queue when bfqq
5773 * remains empty, then the device goes on being fed with I/O requests,
5774 * and the throughput is not affected. In contrast, if BFQ is not
5775 * allowed to switch to another queue---because bfqq is sync and
5776 * I/O-dispatch needs to be plugged while bfqq is temporarily
5777 * empty---then, during the service of bfqq, there will be frequent
5778 * "service holes", i.e., time intervals during which bfqq gets empty
5779 * and the device can only consume the I/O already queued in its
5780 * hardware queues. During service holes, the device may even get to
5781 * remaining idle. In the end, during the service of bfqq, the device
5782 * is driven at a lower speed than the one it can reach with the kind
5783 * of I/O flowing through bfqq.
5785 * To counter this loss of throughput, BFQ implements a "request
5786 * injection mechanism", which tries to fill the above service holes
5787 * with I/O requests taken from other queues. The hard part in this
5788 * mechanism is finding the right amount of I/O to inject, so as to
5789 * both boost throughput and not break bfqq's bandwidth and latency
5790 * guarantees. In this respect, the mechanism maintains a per-queue
5791 * inject limit, computed as below. While bfqq is empty, the injection
5792 * mechanism dispatches extra I/O requests only until the total number
5793 * of I/O requests in flight---i.e., already dispatched but not yet
5794 * completed---remains lower than this limit.
5796 * A first definition comes in handy to introduce the algorithm by
5797 * which the inject limit is computed. We define as first request for
5798 * bfqq, an I/O request for bfqq that arrives while bfqq is in
5799 * service, and causes bfqq to switch from empty to non-empty. The
5800 * algorithm updates the limit as a function of the effect of
5801 * injection on the service times of only the first requests of
5802 * bfqq. The reason for this restriction is that these are the
5803 * requests whose service time is affected most, because they are the
5804 * first to arrive after injection possibly occurred.
5806 * To evaluate the effect of injection, the algorithm measures the
5807 * "total service time" of first requests. We define as total service
5808 * time of an I/O request, the time that elapses since when the
5809 * request is enqueued into bfqq, to when it is completed. This
5810 * quantity allows the whole effect of injection to be measured. It is
5811 * easy to see why. Suppose that some requests of other queues are
5812 * actually injected while bfqq is empty, and that a new request R
5813 * then arrives for bfqq. If the device does start to serve all or
5814 * part of the injected requests during the service hole, then,
5815 * because of this extra service, it may delay the next invocation of
5816 * the dispatch hook of BFQ. Then, even after R gets eventually
5817 * dispatched, the device may delay the actual service of R if it is
5818 * still busy serving the extra requests, or if it decides to serve,
5819 * before R, some extra request still present in its queues. As a
5820 * conclusion, the cumulative extra delay caused by injection can be
5821 * easily evaluated by just comparing the total service time of first
5822 * requests with and without injection.
5824 * The limit-update algorithm works as follows. On the arrival of a
5825 * first request of bfqq, the algorithm measures the total time of the
5826 * request only if one of the three cases below holds, and, for each
5827 * case, it updates the limit as described below:
5829 * (1) If there is no in-flight request. This gives a baseline for the
5830 * total service time of the requests of bfqq. If the baseline has
5831 * not been computed yet, then, after computing it, the limit is
5832 * set to 1, to start boosting throughput, and to prepare the
5833 * ground for the next case. If the baseline has already been
5834 * computed, then it is updated, in case it results to be lower
5835 * than the previous value.
5837 * (2) If the limit is higher than 0 and there are in-flight
5838 * requests. By comparing the total service time in this case with
5839 * the above baseline, it is possible to know at which extent the
5840 * current value of the limit is inflating the total service
5841 * time. If the inflation is below a certain threshold, then bfqq
5842 * is assumed to be suffering from no perceivable loss of its
5843 * service guarantees, and the limit is even tentatively
5844 * increased. If the inflation is above the threshold, then the
5845 * limit is decreased. Due to the lack of any hysteresis, this
5846 * logic makes the limit oscillate even in steady workload
5847 * conditions. Yet we opted for it, because it is fast in reaching
5848 * the best value for the limit, as a function of the current I/O
5849 * workload. To reduce oscillations, this step is disabled for a
5850 * short time interval after the limit happens to be decreased.
5852 * (3) Periodically, after resetting the limit, to make sure that the
5853 * limit eventually drops in case the workload changes. This is
5854 * needed because, after the limit has gone safely up for a
5855 * certain workload, it is impossible to guess whether the
5856 * baseline total service time may have changed, without measuring
5857 * it again without injection. A more effective version of this
5858 * step might be to just sample the baseline, by interrupting
5859 * injection only once, and then to reset/lower the limit only if
5860 * the total service time with the current limit does happen to be
5863 * More details on each step are provided in the comments on the
5864 * pieces of code that implement these steps: the branch handling the
5865 * transition from empty to non empty in bfq_add_request(), the branch
5866 * handling injection in bfq_select_queue(), and the function
5867 * bfq_choose_bfqq_for_injection(). These comments also explain some
5868 * exceptions, made by the injection mechanism in some special cases.
5870 static void bfq_update_inject_limit(struct bfq_data *bfqd,
5871 struct bfq_queue *bfqq)
5873 u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns;
5874 unsigned int old_limit = bfqq->inject_limit;
5876 if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) {
5877 u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
5879 if (tot_time_ns >= threshold && old_limit > 0) {
5880 bfqq->inject_limit--;
5881 bfqq->decrease_time_jif = jiffies;
5882 } else if (tot_time_ns < threshold &&
5883 old_limit <= bfqd->max_rq_in_driver)
5884 bfqq->inject_limit++;
5888 * Either we still have to compute the base value for the
5889 * total service time, and there seem to be the right
5890 * conditions to do it, or we can lower the last base value
5893 * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
5894 * request in flight, because this function is in the code
5895 * path that handles the completion of a request of bfqq, and,
5896 * in particular, this function is executed before
5897 * bfqd->rq_in_driver is decremented in such a code path.
5899 if ((bfqq->last_serv_time_ns == 0 && bfqd->rq_in_driver == 1) ||
5900 tot_time_ns < bfqq->last_serv_time_ns) {
5901 if (bfqq->last_serv_time_ns == 0) {
5903 * Now we certainly have a base value: make sure we
5904 * start trying injection.
5906 bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
5908 bfqq->last_serv_time_ns = tot_time_ns;
5909 } else if (!bfqd->rqs_injected && bfqd->rq_in_driver == 1)
5911 * No I/O injected and no request still in service in
5912 * the drive: these are the exact conditions for
5913 * computing the base value of the total service time
5914 * for bfqq. So let's update this value, because it is
5915 * rather variable. For example, it varies if the size
5916 * or the spatial locality of the I/O requests in bfqq
5919 bfqq->last_serv_time_ns = tot_time_ns;
5922 /* update complete, not waiting for any request completion any longer */
5923 bfqd->waited_rq = NULL;
5924 bfqd->rqs_injected = false;
5928 * Handle either a requeue or a finish for rq. The things to do are
5929 * the same in both cases: all references to rq are to be dropped. In
5930 * particular, rq is considered completed from the point of view of
5933 static void bfq_finish_requeue_request(struct request *rq)
5935 struct bfq_queue *bfqq = RQ_BFQQ(rq);
5936 struct bfq_data *bfqd;
5939 * rq either is not associated with any icq, or is an already
5940 * requeued request that has not (yet) been re-inserted into
5943 if (!rq->elv.icq || !bfqq)
5948 if (rq->rq_flags & RQF_STARTED)
5949 bfqg_stats_update_completion(bfqq_group(bfqq),
5951 rq->io_start_time_ns,
5954 if (likely(rq->rq_flags & RQF_STARTED)) {
5955 unsigned long flags;
5957 spin_lock_irqsave(&bfqd->lock, flags);
5959 if (rq == bfqd->waited_rq)
5960 bfq_update_inject_limit(bfqd, bfqq);
5962 bfq_completed_request(bfqq, bfqd);
5963 bfq_finish_requeue_request_body(bfqq);
5965 spin_unlock_irqrestore(&bfqd->lock, flags);
5968 * Request rq may be still/already in the scheduler,
5969 * in which case we need to remove it (this should
5970 * never happen in case of requeue). And we cannot
5971 * defer such a check and removal, to avoid
5972 * inconsistencies in the time interval from the end
5973 * of this function to the start of the deferred work.
5974 * This situation seems to occur only in process
5975 * context, as a consequence of a merge. In the
5976 * current version of the code, this implies that the
5980 if (!RB_EMPTY_NODE(&rq->rb_node)) {
5981 bfq_remove_request(rq->q, rq);
5982 bfqg_stats_update_io_remove(bfqq_group(bfqq),
5985 bfq_finish_requeue_request_body(bfqq);
5989 * Reset private fields. In case of a requeue, this allows
5990 * this function to correctly do nothing if it is spuriously
5991 * invoked again on this same request (see the check at the
5992 * beginning of the function). Probably, a better general
5993 * design would be to prevent blk-mq from invoking the requeue
5994 * or finish hooks of an elevator, for a request that is not
5995 * referred by that elevator.
5997 * Resetting the following fields would break the
5998 * request-insertion logic if rq is re-inserted into a bfq
5999 * internal queue, without a re-preparation. Here we assume
6000 * that re-insertions of requeued requests, without
6001 * re-preparation, can happen only for pass_through or at_head
6002 * requests (which are not re-inserted into bfq internal
6005 rq->elv.priv[0] = NULL;
6006 rq->elv.priv[1] = NULL;
6010 * Removes the association between the current task and bfqq, assuming
6011 * that bic points to the bfq iocontext of the task.
6012 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
6013 * was the last process referring to that bfqq.
6015 static struct bfq_queue *
6016 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
6018 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
6020 if (bfqq_process_refs(bfqq) == 1) {
6021 bfqq->pid = current->pid;
6022 bfq_clear_bfqq_coop(bfqq);
6023 bfq_clear_bfqq_split_coop(bfqq);
6027 bic_set_bfqq(bic, NULL, 1);
6029 bfq_put_cooperator(bfqq);
6031 bfq_release_process_ref(bfqq->bfqd, bfqq);
6035 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
6036 struct bfq_io_cq *bic,
6038 bool split, bool is_sync,
6041 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
6043 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
6050 bfq_put_queue(bfqq);
6051 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
6053 bic_set_bfqq(bic, bfqq, is_sync);
6054 if (split && is_sync) {
6055 if ((bic->was_in_burst_list && bfqd->large_burst) ||
6056 bic->saved_in_large_burst)
6057 bfq_mark_bfqq_in_large_burst(bfqq);
6059 bfq_clear_bfqq_in_large_burst(bfqq);
6060 if (bic->was_in_burst_list)
6062 * If bfqq was in the current
6063 * burst list before being
6064 * merged, then we have to add
6065 * it back. And we do not need
6066 * to increase burst_size, as
6067 * we did not decrement
6068 * burst_size when we removed
6069 * bfqq from the burst list as
6070 * a consequence of a merge
6072 * bfq_put_queue). In this
6073 * respect, it would be rather
6074 * costly to know whether the
6075 * current burst list is still
6076 * the same burst list from
6077 * which bfqq was removed on
6078 * the merge. To avoid this
6079 * cost, if bfqq was in a
6080 * burst list, then we add
6081 * bfqq to the current burst
6082 * list without any further
6083 * check. This can cause
6084 * inappropriate insertions,
6085 * but rarely enough to not
6086 * harm the detection of large
6087 * bursts significantly.
6089 hlist_add_head(&bfqq->burst_list_node,
6092 bfqq->split_time = jiffies;
6099 * Only reset private fields. The actual request preparation will be
6100 * performed by bfq_init_rq, when rq is either inserted or merged. See
6101 * comments on bfq_init_rq for the reason behind this delayed
6104 static void bfq_prepare_request(struct request *rq)
6107 * Regardless of whether we have an icq attached, we have to
6108 * clear the scheduler pointers, as they might point to
6109 * previously allocated bic/bfqq structs.
6111 rq->elv.priv[0] = rq->elv.priv[1] = NULL;
6115 * If needed, init rq, allocate bfq data structures associated with
6116 * rq, and increment reference counters in the destination bfq_queue
6117 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6118 * not associated with any bfq_queue.
6120 * This function is invoked by the functions that perform rq insertion
6121 * or merging. One may have expected the above preparation operations
6122 * to be performed in bfq_prepare_request, and not delayed to when rq
6123 * is inserted or merged. The rationale behind this delayed
6124 * preparation is that, after the prepare_request hook is invoked for
6125 * rq, rq may still be transformed into a request with no icq, i.e., a
6126 * request not associated with any queue. No bfq hook is invoked to
6127 * signal this transformation. As a consequence, should these
6128 * preparation operations be performed when the prepare_request hook
6129 * is invoked, and should rq be transformed one moment later, bfq
6130 * would end up in an inconsistent state, because it would have
6131 * incremented some queue counters for an rq destined to
6132 * transformation, without any chance to correctly lower these
6133 * counters back. In contrast, no transformation can still happen for
6134 * rq after rq has been inserted or merged. So, it is safe to execute
6135 * these preparation operations when rq is finally inserted or merged.
6137 static struct bfq_queue *bfq_init_rq(struct request *rq)
6139 struct request_queue *q = rq->q;
6140 struct bio *bio = rq->bio;
6141 struct bfq_data *bfqd = q->elevator->elevator_data;
6142 struct bfq_io_cq *bic;
6143 const int is_sync = rq_is_sync(rq);
6144 struct bfq_queue *bfqq;
6145 bool new_queue = false;
6146 bool bfqq_already_existing = false, split = false;
6148 if (unlikely(!rq->elv.icq))
6152 * Assuming that elv.priv[1] is set only if everything is set
6153 * for this rq. This holds true, because this function is
6154 * invoked only for insertion or merging, and, after such
6155 * events, a request cannot be manipulated any longer before
6156 * being removed from bfq.
6158 if (rq->elv.priv[1])
6159 return rq->elv.priv[1];
6161 bic = icq_to_bic(rq->elv.icq);
6163 bfq_check_ioprio_change(bic, bio);
6165 bfq_bic_update_cgroup(bic, bio);
6167 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
6170 if (likely(!new_queue)) {
6171 /* If the queue was seeky for too long, break it apart. */
6172 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
6173 bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
6175 /* Update bic before losing reference to bfqq */
6176 if (bfq_bfqq_in_large_burst(bfqq))
6177 bic->saved_in_large_burst = true;
6179 bfqq = bfq_split_bfqq(bic, bfqq);
6183 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
6187 bfqq_already_existing = true;
6193 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
6194 rq, bfqq, bfqq->ref);
6196 rq->elv.priv[0] = bic;
6197 rq->elv.priv[1] = bfqq;
6200 * If a bfq_queue has only one process reference, it is owned
6201 * by only this bic: we can then set bfqq->bic = bic. in
6202 * addition, if the queue has also just been split, we have to
6205 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
6209 * The queue has just been split from a shared
6210 * queue: restore the idle window and the
6211 * possible weight raising period.
6213 bfq_bfqq_resume_state(bfqq, bfqd, bic,
6214 bfqq_already_existing);
6219 * Consider bfqq as possibly belonging to a burst of newly
6220 * created queues only if:
6221 * 1) A burst is actually happening (bfqd->burst_size > 0)
6223 * 2) There is no other active queue. In fact, if, in
6224 * contrast, there are active queues not belonging to the
6225 * possible burst bfqq may belong to, then there is no gain
6226 * in considering bfqq as belonging to a burst, and
6227 * therefore in not weight-raising bfqq. See comments on
6228 * bfq_handle_burst().
6230 * This filtering also helps eliminating false positives,
6231 * occurring when bfqq does not belong to an actual large
6232 * burst, but some background task (e.g., a service) happens
6233 * to trigger the creation of new queues very close to when
6234 * bfqq and its possible companion queues are created. See
6235 * comments on bfq_handle_burst() for further details also on
6238 if (unlikely(bfq_bfqq_just_created(bfqq) &&
6239 (bfqd->burst_size > 0 ||
6240 bfq_tot_busy_queues(bfqd) == 0)))
6241 bfq_handle_burst(bfqd, bfqq);
6247 bfq_idle_slice_timer_body(struct bfq_data *bfqd, struct bfq_queue *bfqq)
6249 enum bfqq_expiration reason;
6250 unsigned long flags;
6252 spin_lock_irqsave(&bfqd->lock, flags);
6255 * Considering that bfqq may be in race, we should firstly check
6256 * whether bfqq is in service before doing something on it. If
6257 * the bfqq in race is not in service, it has already been expired
6258 * through __bfq_bfqq_expire func and its wait_request flags has
6259 * been cleared in __bfq_bfqd_reset_in_service func.
6261 if (bfqq != bfqd->in_service_queue) {
6262 spin_unlock_irqrestore(&bfqd->lock, flags);
6266 bfq_clear_bfqq_wait_request(bfqq);
6268 if (bfq_bfqq_budget_timeout(bfqq))
6270 * Also here the queue can be safely expired
6271 * for budget timeout without wasting
6274 reason = BFQQE_BUDGET_TIMEOUT;
6275 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
6277 * The queue may not be empty upon timer expiration,
6278 * because we may not disable the timer when the
6279 * first request of the in-service queue arrives
6280 * during disk idling.
6282 reason = BFQQE_TOO_IDLE;
6284 goto schedule_dispatch;
6286 bfq_bfqq_expire(bfqd, bfqq, true, reason);
6289 spin_unlock_irqrestore(&bfqd->lock, flags);
6290 bfq_schedule_dispatch(bfqd);
6294 * Handler of the expiration of the timer running if the in-service queue
6295 * is idling inside its time slice.
6297 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
6299 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
6301 struct bfq_queue *bfqq = bfqd->in_service_queue;
6304 * Theoretical race here: the in-service queue can be NULL or
6305 * different from the queue that was idling if a new request
6306 * arrives for the current queue and there is a full dispatch
6307 * cycle that changes the in-service queue. This can hardly
6308 * happen, but in the worst case we just expire a queue too
6312 bfq_idle_slice_timer_body(bfqd, bfqq);
6314 return HRTIMER_NORESTART;
6317 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
6318 struct bfq_queue **bfqq_ptr)
6320 struct bfq_queue *bfqq = *bfqq_ptr;
6322 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
6324 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
6326 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
6328 bfq_put_queue(bfqq);
6334 * Release all the bfqg references to its async queues. If we are
6335 * deallocating the group these queues may still contain requests, so
6336 * we reparent them to the root cgroup (i.e., the only one that will
6337 * exist for sure until all the requests on a device are gone).
6339 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
6343 for (i = 0; i < 2; i++)
6344 for (j = 0; j < IOPRIO_BE_NR; j++)
6345 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
6347 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
6351 * See the comments on bfq_limit_depth for the purpose of
6352 * the depths set in the function. Return minimum shallow depth we'll use.
6354 static unsigned int bfq_update_depths(struct bfq_data *bfqd,
6355 struct sbitmap_queue *bt)
6357 unsigned int i, j, min_shallow = UINT_MAX;
6360 * In-word depths if no bfq_queue is being weight-raised:
6361 * leaving 25% of tags only for sync reads.
6363 * In next formulas, right-shift the value
6364 * (1U<<bt->sb.shift), instead of computing directly
6365 * (1U<<(bt->sb.shift - something)), to be robust against
6366 * any possible value of bt->sb.shift, without having to
6367 * limit 'something'.
6369 /* no more than 50% of tags for async I/O */
6370 bfqd->word_depths[0][0] = max(bt->sb.depth >> 1, 1U);
6372 * no more than 75% of tags for sync writes (25% extra tags
6373 * w.r.t. async I/O, to prevent async I/O from starving sync
6376 bfqd->word_depths[0][1] = max((bt->sb.depth * 3) >> 2, 1U);
6379 * In-word depths in case some bfq_queue is being weight-
6380 * raised: leaving ~63% of tags for sync reads. This is the
6381 * highest percentage for which, in our tests, application
6382 * start-up times didn't suffer from any regression due to tag
6385 /* no more than ~18% of tags for async I/O */
6386 bfqd->word_depths[1][0] = max((bt->sb.depth * 3) >> 4, 1U);
6387 /* no more than ~37% of tags for sync writes (~20% extra tags) */
6388 bfqd->word_depths[1][1] = max((bt->sb.depth * 6) >> 4, 1U);
6390 for (i = 0; i < 2; i++)
6391 for (j = 0; j < 2; j++)
6392 min_shallow = min(min_shallow, bfqd->word_depths[i][j]);
6397 static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
6399 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
6400 struct blk_mq_tags *tags = hctx->sched_tags;
6401 unsigned int min_shallow;
6403 min_shallow = bfq_update_depths(bfqd, tags->bitmap_tags);
6404 sbitmap_queue_min_shallow_depth(tags->bitmap_tags, min_shallow);
6407 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
6409 bfq_depth_updated(hctx);
6413 static void bfq_exit_queue(struct elevator_queue *e)
6415 struct bfq_data *bfqd = e->elevator_data;
6416 struct bfq_queue *bfqq, *n;
6418 hrtimer_cancel(&bfqd->idle_slice_timer);
6420 spin_lock_irq(&bfqd->lock);
6421 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
6422 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
6423 spin_unlock_irq(&bfqd->lock);
6425 hrtimer_cancel(&bfqd->idle_slice_timer);
6427 /* release oom-queue reference to root group */
6428 bfqg_and_blkg_put(bfqd->root_group);
6430 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6431 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
6433 spin_lock_irq(&bfqd->lock);
6434 bfq_put_async_queues(bfqd, bfqd->root_group);
6435 kfree(bfqd->root_group);
6436 spin_unlock_irq(&bfqd->lock);
6442 static void bfq_init_root_group(struct bfq_group *root_group,
6443 struct bfq_data *bfqd)
6447 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6448 root_group->entity.parent = NULL;
6449 root_group->my_entity = NULL;
6450 root_group->bfqd = bfqd;
6452 root_group->rq_pos_tree = RB_ROOT;
6453 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
6454 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
6455 root_group->sched_data.bfq_class_idle_last_service = jiffies;
6458 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
6460 struct bfq_data *bfqd;
6461 struct elevator_queue *eq;
6463 eq = elevator_alloc(q, e);
6467 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
6469 kobject_put(&eq->kobj);
6472 eq->elevator_data = bfqd;
6474 spin_lock_irq(&q->queue_lock);
6476 spin_unlock_irq(&q->queue_lock);
6479 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
6480 * Grab a permanent reference to it, so that the normal code flow
6481 * will not attempt to free it.
6483 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
6484 bfqd->oom_bfqq.ref++;
6485 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
6486 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
6487 bfqd->oom_bfqq.entity.new_weight =
6488 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
6490 /* oom_bfqq does not participate to bursts */
6491 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
6494 * Trigger weight initialization, according to ioprio, at the
6495 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
6496 * class won't be changed any more.
6498 bfqd->oom_bfqq.entity.prio_changed = 1;
6502 INIT_LIST_HEAD(&bfqd->dispatch);
6504 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
6506 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
6508 bfqd->queue_weights_tree = RB_ROOT_CACHED;
6509 bfqd->num_groups_with_pending_reqs = 0;
6511 INIT_LIST_HEAD(&bfqd->active_list);
6512 INIT_LIST_HEAD(&bfqd->idle_list);
6513 INIT_HLIST_HEAD(&bfqd->burst_list);
6516 bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
6518 bfqd->bfq_max_budget = bfq_default_max_budget;
6520 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
6521 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
6522 bfqd->bfq_back_max = bfq_back_max;
6523 bfqd->bfq_back_penalty = bfq_back_penalty;
6524 bfqd->bfq_slice_idle = bfq_slice_idle;
6525 bfqd->bfq_timeout = bfq_timeout;
6527 bfqd->bfq_requests_within_timer = 120;
6529 bfqd->bfq_large_burst_thresh = 8;
6530 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
6532 bfqd->low_latency = true;
6535 * Trade-off between responsiveness and fairness.
6537 bfqd->bfq_wr_coeff = 30;
6538 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
6539 bfqd->bfq_wr_max_time = 0;
6540 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
6541 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
6542 bfqd->bfq_wr_max_softrt_rate = 7000; /*
6543 * Approximate rate required
6544 * to playback or record a
6545 * high-definition compressed
6548 bfqd->wr_busy_queues = 0;
6551 * Begin by assuming, optimistically, that the device peak
6552 * rate is equal to 2/3 of the highest reference rate.
6554 bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
6555 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
6556 bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
6558 spin_lock_init(&bfqd->lock);
6561 * The invocation of the next bfq_create_group_hierarchy
6562 * function is the head of a chain of function calls
6563 * (bfq_create_group_hierarchy->blkcg_activate_policy->
6564 * blk_mq_freeze_queue) that may lead to the invocation of the
6565 * has_work hook function. For this reason,
6566 * bfq_create_group_hierarchy is invoked only after all
6567 * scheduler data has been initialized, apart from the fields
6568 * that can be initialized only after invoking
6569 * bfq_create_group_hierarchy. This, in particular, enables
6570 * has_work to correctly return false. Of course, to avoid
6571 * other inconsistencies, the blk-mq stack must then refrain
6572 * from invoking further scheduler hooks before this init
6573 * function is finished.
6575 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
6576 if (!bfqd->root_group)
6578 bfq_init_root_group(bfqd->root_group, bfqd);
6579 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
6581 wbt_disable_default(q);
6586 kobject_put(&eq->kobj);
6590 static void bfq_slab_kill(void)
6592 kmem_cache_destroy(bfq_pool);
6595 static int __init bfq_slab_setup(void)
6597 bfq_pool = KMEM_CACHE(bfq_queue, 0);
6603 static ssize_t bfq_var_show(unsigned int var, char *page)
6605 return sprintf(page, "%u\n", var);
6608 static int bfq_var_store(unsigned long *var, const char *page)
6610 unsigned long new_val;
6611 int ret = kstrtoul(page, 10, &new_val);
6619 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
6620 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
6622 struct bfq_data *bfqd = e->elevator_data; \
6623 u64 __data = __VAR; \
6625 __data = jiffies_to_msecs(__data); \
6626 else if (__CONV == 2) \
6627 __data = div_u64(__data, NSEC_PER_MSEC); \
6628 return bfq_var_show(__data, (page)); \
6630 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
6631 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
6632 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
6633 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
6634 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
6635 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
6636 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
6637 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
6638 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
6639 #undef SHOW_FUNCTION
6641 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
6642 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
6644 struct bfq_data *bfqd = e->elevator_data; \
6645 u64 __data = __VAR; \
6646 __data = div_u64(__data, NSEC_PER_USEC); \
6647 return bfq_var_show(__data, (page)); \
6649 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
6650 #undef USEC_SHOW_FUNCTION
6652 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
6654 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
6656 struct bfq_data *bfqd = e->elevator_data; \
6657 unsigned long __data, __min = (MIN), __max = (MAX); \
6660 ret = bfq_var_store(&__data, (page)); \
6663 if (__data < __min) \
6665 else if (__data > __max) \
6668 *(__PTR) = msecs_to_jiffies(__data); \
6669 else if (__CONV == 2) \
6670 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
6672 *(__PTR) = __data; \
6675 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
6677 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
6679 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
6680 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
6682 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
6683 #undef STORE_FUNCTION
6685 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
6686 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
6688 struct bfq_data *bfqd = e->elevator_data; \
6689 unsigned long __data, __min = (MIN), __max = (MAX); \
6692 ret = bfq_var_store(&__data, (page)); \
6695 if (__data < __min) \
6697 else if (__data > __max) \
6699 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
6702 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
6704 #undef USEC_STORE_FUNCTION
6706 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
6707 const char *page, size_t count)
6709 struct bfq_data *bfqd = e->elevator_data;
6710 unsigned long __data;
6713 ret = bfq_var_store(&__data, (page));
6718 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
6720 if (__data > INT_MAX)
6722 bfqd->bfq_max_budget = __data;
6725 bfqd->bfq_user_max_budget = __data;
6731 * Leaving this name to preserve name compatibility with cfq
6732 * parameters, but this timeout is used for both sync and async.
6734 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
6735 const char *page, size_t count)
6737 struct bfq_data *bfqd = e->elevator_data;
6738 unsigned long __data;
6741 ret = bfq_var_store(&__data, (page));
6747 else if (__data > INT_MAX)
6750 bfqd->bfq_timeout = msecs_to_jiffies(__data);
6751 if (bfqd->bfq_user_max_budget == 0)
6752 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
6757 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
6758 const char *page, size_t count)
6760 struct bfq_data *bfqd = e->elevator_data;
6761 unsigned long __data;
6764 ret = bfq_var_store(&__data, (page));
6770 if (!bfqd->strict_guarantees && __data == 1
6771 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
6772 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
6774 bfqd->strict_guarantees = __data;
6779 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
6780 const char *page, size_t count)
6782 struct bfq_data *bfqd = e->elevator_data;
6783 unsigned long __data;
6786 ret = bfq_var_store(&__data, (page));
6792 if (__data == 0 && bfqd->low_latency != 0)
6794 bfqd->low_latency = __data;
6799 #define BFQ_ATTR(name) \
6800 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
6802 static struct elv_fs_entry bfq_attrs[] = {
6803 BFQ_ATTR(fifo_expire_sync),
6804 BFQ_ATTR(fifo_expire_async),
6805 BFQ_ATTR(back_seek_max),
6806 BFQ_ATTR(back_seek_penalty),
6807 BFQ_ATTR(slice_idle),
6808 BFQ_ATTR(slice_idle_us),
6809 BFQ_ATTR(max_budget),
6810 BFQ_ATTR(timeout_sync),
6811 BFQ_ATTR(strict_guarantees),
6812 BFQ_ATTR(low_latency),
6816 static struct elevator_type iosched_bfq_mq = {
6818 .limit_depth = bfq_limit_depth,
6819 .prepare_request = bfq_prepare_request,
6820 .requeue_request = bfq_finish_requeue_request,
6821 .finish_request = bfq_finish_requeue_request,
6822 .exit_icq = bfq_exit_icq,
6823 .insert_requests = bfq_insert_requests,
6824 .dispatch_request = bfq_dispatch_request,
6825 .next_request = elv_rb_latter_request,
6826 .former_request = elv_rb_former_request,
6827 .allow_merge = bfq_allow_bio_merge,
6828 .bio_merge = bfq_bio_merge,
6829 .request_merge = bfq_request_merge,
6830 .requests_merged = bfq_requests_merged,
6831 .request_merged = bfq_request_merged,
6832 .has_work = bfq_has_work,
6833 .depth_updated = bfq_depth_updated,
6834 .init_hctx = bfq_init_hctx,
6835 .init_sched = bfq_init_queue,
6836 .exit_sched = bfq_exit_queue,
6839 .icq_size = sizeof(struct bfq_io_cq),
6840 .icq_align = __alignof__(struct bfq_io_cq),
6841 .elevator_attrs = bfq_attrs,
6842 .elevator_name = "bfq",
6843 .elevator_owner = THIS_MODULE,
6845 MODULE_ALIAS("bfq-iosched");
6847 static int __init bfq_init(void)
6851 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6852 ret = blkcg_policy_register(&blkcg_policy_bfq);
6858 if (bfq_slab_setup())
6862 * Times to load large popular applications for the typical
6863 * systems installed on the reference devices (see the
6864 * comments before the definition of the next
6865 * array). Actually, we use slightly lower values, as the
6866 * estimated peak rate tends to be smaller than the actual
6867 * peak rate. The reason for this last fact is that estimates
6868 * are computed over much shorter time intervals than the long
6869 * intervals typically used for benchmarking. Why? First, to
6870 * adapt more quickly to variations. Second, because an I/O
6871 * scheduler cannot rely on a peak-rate-evaluation workload to
6872 * be run for a long time.
6874 ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
6875 ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
6877 ret = elv_register(&iosched_bfq_mq);
6886 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6887 blkcg_policy_unregister(&blkcg_policy_bfq);
6892 static void __exit bfq_exit(void)
6894 elv_unregister(&iosched_bfq_mq);
6895 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6896 blkcg_policy_unregister(&blkcg_policy_bfq);
6901 module_init(bfq_init);
6902 module_exit(bfq_exit);
6904 MODULE_AUTHOR("Paolo Valente");
6905 MODULE_LICENSE("GPL");
6906 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");