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>
129 #include "blk-mq-tag.h"
130 #include "blk-mq-sched.h"
131 #include "bfq-iosched.h"
134 #define BFQ_BFQQ_FNS(name) \
135 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
137 __set_bit(BFQQF_##name, &(bfqq)->flags); \
139 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
141 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
143 int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
145 return test_bit(BFQQF_##name, &(bfqq)->flags); \
148 BFQ_BFQQ_FNS(just_created);
150 BFQ_BFQQ_FNS(wait_request);
151 BFQ_BFQQ_FNS(non_blocking_wait_rq);
152 BFQ_BFQQ_FNS(fifo_expire);
153 BFQ_BFQQ_FNS(has_short_ttime);
155 BFQ_BFQQ_FNS(IO_bound);
156 BFQ_BFQQ_FNS(in_large_burst);
158 BFQ_BFQQ_FNS(split_coop);
159 BFQ_BFQQ_FNS(softrt_update);
160 BFQ_BFQQ_FNS(has_waker);
161 #undef BFQ_BFQQ_FNS \
163 /* Expiration time of sync (0) and async (1) requests, in ns. */
164 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
166 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
167 static const int bfq_back_max = 16 * 1024;
169 /* Penalty of a backwards seek, in number of sectors. */
170 static const int bfq_back_penalty = 2;
172 /* Idling period duration, in ns. */
173 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
175 /* Minimum number of assigned budgets for which stats are safe to compute. */
176 static const int bfq_stats_min_budgets = 194;
178 /* Default maximum budget values, in sectors and number of requests. */
179 static const int bfq_default_max_budget = 16 * 1024;
182 * When a sync request is dispatched, the queue that contains that
183 * request, and all the ancestor entities of that queue, are charged
184 * with the number of sectors of the request. In contrast, if the
185 * request is async, then the queue and its ancestor entities are
186 * charged with the number of sectors of the request, multiplied by
187 * the factor below. This throttles the bandwidth for async I/O,
188 * w.r.t. to sync I/O, and it is done to counter the tendency of async
189 * writes to steal I/O throughput to reads.
191 * The current value of this parameter is the result of a tuning with
192 * several hardware and software configurations. We tried to find the
193 * lowest value for which writes do not cause noticeable problems to
194 * reads. In fact, the lower this parameter, the stabler I/O control,
195 * in the following respect. The lower this parameter is, the less
196 * the bandwidth enjoyed by a group decreases
197 * - when the group does writes, w.r.t. to when it does reads;
198 * - when other groups do reads, w.r.t. to when they do writes.
200 static const int bfq_async_charge_factor = 3;
202 /* Default timeout values, in jiffies, approximating CFQ defaults. */
203 const int bfq_timeout = HZ / 8;
206 * Time limit for merging (see comments in bfq_setup_cooperator). Set
207 * to the slowest value that, in our tests, proved to be effective in
208 * removing false positives, while not causing true positives to miss
211 * As can be deduced from the low time limit below, queue merging, if
212 * successful, happens at the very beginning of the I/O of the involved
213 * cooperating processes, as a consequence of the arrival of the very
214 * first requests from each cooperator. After that, there is very
215 * little chance to find cooperators.
217 static const unsigned long bfq_merge_time_limit = HZ/10;
219 static struct kmem_cache *bfq_pool;
221 /* Below this threshold (in ns), we consider thinktime immediate. */
222 #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
224 /* hw_tag detection: parallel requests threshold and min samples needed. */
225 #define BFQ_HW_QUEUE_THRESHOLD 3
226 #define BFQ_HW_QUEUE_SAMPLES 32
228 #define BFQQ_SEEK_THR (sector_t)(8 * 100)
229 #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
230 #define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \
231 (get_sdist(last_pos, rq) > \
233 (!blk_queue_nonrot(bfqd->queue) || \
234 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT))
235 #define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
236 #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19)
238 * Sync random I/O is likely to be confused with soft real-time I/O,
239 * because it is characterized by limited throughput and apparently
240 * isochronous arrival pattern. To avoid false positives, queues
241 * containing only random (seeky) I/O are prevented from being tagged
244 #define BFQQ_TOTALLY_SEEKY(bfqq) (bfqq->seek_history == -1)
246 /* Min number of samples required to perform peak-rate update */
247 #define BFQ_RATE_MIN_SAMPLES 32
248 /* Min observation time interval required to perform a peak-rate update (ns) */
249 #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
250 /* Target observation time interval for a peak-rate update (ns) */
251 #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
254 * Shift used for peak-rate fixed precision calculations.
256 * - the current shift: 16 positions
257 * - the current type used to store rate: u32
258 * - the current unit of measure for rate: [sectors/usec], or, more precisely,
259 * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
260 * the range of rates that can be stored is
261 * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
262 * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
263 * [15, 65G] sectors/sec
264 * Which, assuming a sector size of 512B, corresponds to a range of
267 #define BFQ_RATE_SHIFT 16
270 * When configured for computing the duration of the weight-raising
271 * for interactive queues automatically (see the comments at the
272 * beginning of this file), BFQ does it using the following formula:
273 * duration = (ref_rate / r) * ref_wr_duration,
274 * where r is the peak rate of the device, and ref_rate and
275 * ref_wr_duration are two reference parameters. In particular,
276 * ref_rate is the peak rate of the reference storage device (see
277 * below), and ref_wr_duration is about the maximum time needed, with
278 * BFQ and while reading two files in parallel, to load typical large
279 * applications on the reference device (see the comments on
280 * max_service_from_wr below, for more details on how ref_wr_duration
281 * is obtained). In practice, the slower/faster the device at hand
282 * is, the more/less it takes to load applications with respect to the
283 * reference device. Accordingly, the longer/shorter BFQ grants
284 * weight raising to interactive applications.
286 * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
287 * depending on whether the device is rotational or non-rotational.
289 * In the following definitions, ref_rate[0] and ref_wr_duration[0]
290 * are the reference values for a rotational device, whereas
291 * ref_rate[1] and ref_wr_duration[1] are the reference values for a
292 * non-rotational device. The reference rates are not the actual peak
293 * rates of the devices used as a reference, but slightly lower
294 * values. The reason for using slightly lower values is that the
295 * peak-rate estimator tends to yield slightly lower values than the
296 * actual peak rate (it can yield the actual peak rate only if there
297 * is only one process doing I/O, and the process does sequential
300 * The reference peak rates are measured in sectors/usec, left-shifted
303 static int ref_rate[2] = {14000, 33000};
305 * To improve readability, a conversion function is used to initialize
306 * the following array, which entails that the array can be
307 * initialized only in a function.
309 static int ref_wr_duration[2];
312 * BFQ uses the above-detailed, time-based weight-raising mechanism to
313 * privilege interactive tasks. This mechanism is vulnerable to the
314 * following false positives: I/O-bound applications that will go on
315 * doing I/O for much longer than the duration of weight
316 * raising. These applications have basically no benefit from being
317 * weight-raised at the beginning of their I/O. On the opposite end,
318 * while being weight-raised, these applications
319 * a) unjustly steal throughput to applications that may actually need
321 * b) make BFQ uselessly perform device idling; device idling results
322 * in loss of device throughput with most flash-based storage, and may
323 * increase latencies when used purposelessly.
325 * BFQ tries to reduce these problems, by adopting the following
326 * countermeasure. To introduce this countermeasure, we need first to
327 * finish explaining how the duration of weight-raising for
328 * interactive tasks is computed.
330 * For a bfq_queue deemed as interactive, the duration of weight
331 * raising is dynamically adjusted, as a function of the estimated
332 * peak rate of the device, so as to be equal to the time needed to
333 * execute the 'largest' interactive task we benchmarked so far. By
334 * largest task, we mean the task for which each involved process has
335 * to do more I/O than for any of the other tasks we benchmarked. This
336 * reference interactive task is the start-up of LibreOffice Writer,
337 * and in this task each process/bfq_queue needs to have at most ~110K
338 * sectors transferred.
340 * This last piece of information enables BFQ to reduce the actual
341 * duration of weight-raising for at least one class of I/O-bound
342 * applications: those doing sequential or quasi-sequential I/O. An
343 * example is file copy. In fact, once started, the main I/O-bound
344 * processes of these applications usually consume the above 110K
345 * sectors in much less time than the processes of an application that
346 * is starting, because these I/O-bound processes will greedily devote
347 * almost all their CPU cycles only to their target,
348 * throughput-friendly I/O operations. This is even more true if BFQ
349 * happens to be underestimating the device peak rate, and thus
350 * overestimating the duration of weight raising. But, according to
351 * our measurements, once transferred 110K sectors, these processes
352 * have no right to be weight-raised any longer.
354 * Basing on the last consideration, BFQ ends weight-raising for a
355 * bfq_queue if the latter happens to have received an amount of
356 * service at least equal to the following constant. The constant is
357 * set to slightly more than 110K, to have a minimum safety margin.
359 * This early ending of weight-raising reduces the amount of time
360 * during which interactive false positives cause the two problems
361 * described at the beginning of these comments.
363 static const unsigned long max_service_from_wr = 120000;
365 #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
366 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
368 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
370 return bic->bfqq[is_sync];
373 void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
375 bic->bfqq[is_sync] = bfqq;
378 struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
380 return bic->icq.q->elevator->elevator_data;
384 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
385 * @icq: the iocontext queue.
387 static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
389 /* bic->icq is the first member, %NULL will convert to %NULL */
390 return container_of(icq, struct bfq_io_cq, icq);
394 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
395 * @bfqd: the lookup key.
396 * @ioc: the io_context of the process doing I/O.
397 * @q: the request queue.
399 static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd,
400 struct io_context *ioc,
401 struct request_queue *q)
405 struct bfq_io_cq *icq;
407 spin_lock_irqsave(&q->queue_lock, flags);
408 icq = icq_to_bic(ioc_lookup_icq(ioc, q));
409 spin_unlock_irqrestore(&q->queue_lock, flags);
418 * Scheduler run of queue, if there are requests pending and no one in the
419 * driver that will restart queueing.
421 void bfq_schedule_dispatch(struct bfq_data *bfqd)
423 lockdep_assert_held(&bfqd->lock);
425 if (bfqd->queued != 0) {
426 bfq_log(bfqd, "schedule dispatch");
427 blk_mq_run_hw_queues(bfqd->queue, true);
431 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
432 #define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT)
434 #define bfq_sample_valid(samples) ((samples) > 80)
437 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
438 * We choose the request that is closer to the head right now. Distance
439 * behind the head is penalized and only allowed to a certain extent.
441 static struct request *bfq_choose_req(struct bfq_data *bfqd,
446 sector_t s1, s2, d1 = 0, d2 = 0;
447 unsigned long back_max;
448 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
449 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
450 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
452 if (!rq1 || rq1 == rq2)
457 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
459 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
461 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
463 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
466 s1 = blk_rq_pos(rq1);
467 s2 = blk_rq_pos(rq2);
470 * By definition, 1KiB is 2 sectors.
472 back_max = bfqd->bfq_back_max * 2;
475 * Strict one way elevator _except_ in the case where we allow
476 * short backward seeks which are biased as twice the cost of a
477 * similar forward seek.
481 else if (s1 + back_max >= last)
482 d1 = (last - s1) * bfqd->bfq_back_penalty;
484 wrap |= BFQ_RQ1_WRAP;
488 else if (s2 + back_max >= last)
489 d2 = (last - s2) * bfqd->bfq_back_penalty;
491 wrap |= BFQ_RQ2_WRAP;
493 /* Found required data */
496 * By doing switch() on the bit mask "wrap" we avoid having to
497 * check two variables for all permutations: --> faster!
500 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
515 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
518 * Since both rqs are wrapped,
519 * start with the one that's further behind head
520 * (--> only *one* back seek required),
521 * since back seek takes more time than forward.
531 * Async I/O can easily starve sync I/O (both sync reads and sync
532 * writes), by consuming all tags. Similarly, storms of sync writes,
533 * such as those that sync(2) may trigger, can starve sync reads.
534 * Limit depths of async I/O and sync writes so as to counter both
537 static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
539 struct bfq_data *bfqd = data->q->elevator->elevator_data;
541 if (op_is_sync(op) && !op_is_write(op))
544 data->shallow_depth =
545 bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)];
547 bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
548 __func__, bfqd->wr_busy_queues, op_is_sync(op),
549 data->shallow_depth);
552 static struct bfq_queue *
553 bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
554 sector_t sector, struct rb_node **ret_parent,
555 struct rb_node ***rb_link)
557 struct rb_node **p, *parent;
558 struct bfq_queue *bfqq = NULL;
566 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
569 * Sort strictly based on sector. Smallest to the left,
570 * largest to the right.
572 if (sector > blk_rq_pos(bfqq->next_rq))
574 else if (sector < blk_rq_pos(bfqq->next_rq))
582 *ret_parent = parent;
586 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
587 (unsigned long long)sector,
588 bfqq ? bfqq->pid : 0);
593 static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
595 return bfqq->service_from_backlogged > 0 &&
596 time_is_before_jiffies(bfqq->first_IO_time +
597 bfq_merge_time_limit);
601 * The following function is not marked as __cold because it is
602 * actually cold, but for the same performance goal described in the
603 * comments on the likely() at the beginning of
604 * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
605 * execution time for the case where this function is not invoked, we
606 * had to add an unlikely() in each involved if().
609 bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
611 struct rb_node **p, *parent;
612 struct bfq_queue *__bfqq;
614 if (bfqq->pos_root) {
615 rb_erase(&bfqq->pos_node, bfqq->pos_root);
616 bfqq->pos_root = NULL;
619 /* oom_bfqq does not participate in queue merging */
620 if (bfqq == &bfqd->oom_bfqq)
624 * bfqq cannot be merged any longer (see comments in
625 * bfq_setup_cooperator): no point in adding bfqq into the
628 if (bfq_too_late_for_merging(bfqq))
631 if (bfq_class_idle(bfqq))
636 bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
637 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
638 blk_rq_pos(bfqq->next_rq), &parent, &p);
640 rb_link_node(&bfqq->pos_node, parent, p);
641 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
643 bfqq->pos_root = NULL;
647 * The following function returns false either if every active queue
648 * must receive the same share of the throughput (symmetric scenario),
649 * or, as a special case, if bfqq must receive a share of the
650 * throughput lower than or equal to the share that every other active
651 * queue must receive. If bfqq does sync I/O, then these are the only
652 * two cases where bfqq happens to be guaranteed its share of the
653 * throughput even if I/O dispatching is not plugged when bfqq remains
654 * temporarily empty (for more details, see the comments in the
655 * function bfq_better_to_idle()). For this reason, the return value
656 * of this function is used to check whether I/O-dispatch plugging can
659 * The above first case (symmetric scenario) occurs when:
660 * 1) all active queues have the same weight,
661 * 2) all active queues belong to the same I/O-priority class,
662 * 3) all active groups at the same level in the groups tree have the same
664 * 4) all active groups at the same level in the groups tree have the same
665 * number of children.
667 * Unfortunately, keeping the necessary state for evaluating exactly
668 * the last two symmetry sub-conditions above would be quite complex
669 * and time consuming. Therefore this function evaluates, instead,
670 * only the following stronger three sub-conditions, for which it is
671 * much easier to maintain the needed state:
672 * 1) all active queues have the same weight,
673 * 2) all active queues belong to the same I/O-priority class,
674 * 3) there are no active groups.
675 * In particular, the last condition is always true if hierarchical
676 * support or the cgroups interface are not enabled, thus no state
677 * needs to be maintained in this case.
679 static bool bfq_asymmetric_scenario(struct bfq_data *bfqd,
680 struct bfq_queue *bfqq)
682 bool smallest_weight = bfqq &&
683 bfqq->weight_counter &&
684 bfqq->weight_counter ==
686 rb_first_cached(&bfqd->queue_weights_tree),
687 struct bfq_weight_counter,
691 * For queue weights to differ, queue_weights_tree must contain
692 * at least two nodes.
694 bool varied_queue_weights = !smallest_weight &&
695 !RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) &&
696 (bfqd->queue_weights_tree.rb_root.rb_node->rb_left ||
697 bfqd->queue_weights_tree.rb_root.rb_node->rb_right);
699 bool multiple_classes_busy =
700 (bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
701 (bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
702 (bfqd->busy_queues[1] && bfqd->busy_queues[2]);
704 return varied_queue_weights || multiple_classes_busy
705 #ifdef CONFIG_BFQ_GROUP_IOSCHED
706 || bfqd->num_groups_with_pending_reqs > 0
712 * If the weight-counter tree passed as input contains no counter for
713 * the weight of the input queue, then add that counter; otherwise just
714 * increment the existing counter.
716 * Note that weight-counter trees contain few nodes in mostly symmetric
717 * scenarios. For example, if all queues have the same weight, then the
718 * weight-counter tree for the queues may contain at most one node.
719 * This holds even if low_latency is on, because weight-raised queues
720 * are not inserted in the tree.
721 * In most scenarios, the rate at which nodes are created/destroyed
724 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq,
725 struct rb_root_cached *root)
727 struct bfq_entity *entity = &bfqq->entity;
728 struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
729 bool leftmost = true;
732 * Do not insert if the queue is already associated with a
733 * counter, which happens if:
734 * 1) a request arrival has caused the queue to become both
735 * non-weight-raised, and hence change its weight, and
736 * backlogged; in this respect, each of the two events
737 * causes an invocation of this function,
738 * 2) this is the invocation of this function caused by the
739 * second event. This second invocation is actually useless,
740 * and we handle this fact by exiting immediately. More
741 * efficient or clearer solutions might possibly be adopted.
743 if (bfqq->weight_counter)
747 struct bfq_weight_counter *__counter = container_of(*new,
748 struct bfq_weight_counter,
752 if (entity->weight == __counter->weight) {
753 bfqq->weight_counter = __counter;
756 if (entity->weight < __counter->weight)
757 new = &((*new)->rb_left);
759 new = &((*new)->rb_right);
764 bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
768 * In the unlucky event of an allocation failure, we just
769 * exit. This will cause the weight of queue to not be
770 * considered in bfq_asymmetric_scenario, which, in its turn,
771 * causes the scenario to be deemed wrongly symmetric in case
772 * bfqq's weight would have been the only weight making the
773 * scenario asymmetric. On the bright side, no unbalance will
774 * however occur when bfqq becomes inactive again (the
775 * invocation of this function is triggered by an activation
776 * of queue). In fact, bfq_weights_tree_remove does nothing
777 * if !bfqq->weight_counter.
779 if (unlikely(!bfqq->weight_counter))
782 bfqq->weight_counter->weight = entity->weight;
783 rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
784 rb_insert_color_cached(&bfqq->weight_counter->weights_node, root,
788 bfqq->weight_counter->num_active++;
793 * Decrement the weight counter associated with the queue, and, if the
794 * counter reaches 0, remove the counter from the tree.
795 * See the comments to the function bfq_weights_tree_add() for considerations
798 void __bfq_weights_tree_remove(struct bfq_data *bfqd,
799 struct bfq_queue *bfqq,
800 struct rb_root_cached *root)
802 if (!bfqq->weight_counter)
805 bfqq->weight_counter->num_active--;
806 if (bfqq->weight_counter->num_active > 0)
807 goto reset_entity_pointer;
809 rb_erase_cached(&bfqq->weight_counter->weights_node, root);
810 kfree(bfqq->weight_counter);
812 reset_entity_pointer:
813 bfqq->weight_counter = NULL;
818 * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
819 * of active groups for each queue's inactive parent entity.
821 void bfq_weights_tree_remove(struct bfq_data *bfqd,
822 struct bfq_queue *bfqq)
824 struct bfq_entity *entity = bfqq->entity.parent;
826 for_each_entity(entity) {
827 struct bfq_sched_data *sd = entity->my_sched_data;
829 if (sd->next_in_service || sd->in_service_entity) {
831 * entity is still active, because either
832 * next_in_service or in_service_entity is not
833 * NULL (see the comments on the definition of
834 * next_in_service for details on why
835 * in_service_entity must be checked too).
837 * As a consequence, its parent entities are
838 * active as well, and thus this loop must
845 * The decrement of num_groups_with_pending_reqs is
846 * not performed immediately upon the deactivation of
847 * entity, but it is delayed to when it also happens
848 * that the first leaf descendant bfqq of entity gets
849 * all its pending requests completed. The following
850 * instructions perform this delayed decrement, if
851 * needed. See the comments on
852 * num_groups_with_pending_reqs for details.
854 if (entity->in_groups_with_pending_reqs) {
855 entity->in_groups_with_pending_reqs = false;
856 bfqd->num_groups_with_pending_reqs--;
861 * Next function is invoked last, because it causes bfqq to be
862 * freed if the following holds: bfqq is not in service and
863 * has no dispatched request. DO NOT use bfqq after the next
864 * function invocation.
866 __bfq_weights_tree_remove(bfqd, bfqq,
867 &bfqd->queue_weights_tree);
871 * Return expired entry, or NULL to just start from scratch in rbtree.
873 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
874 struct request *last)
878 if (bfq_bfqq_fifo_expire(bfqq))
881 bfq_mark_bfqq_fifo_expire(bfqq);
883 rq = rq_entry_fifo(bfqq->fifo.next);
885 if (rq == last || ktime_get_ns() < rq->fifo_time)
888 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
892 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
893 struct bfq_queue *bfqq,
894 struct request *last)
896 struct rb_node *rbnext = rb_next(&last->rb_node);
897 struct rb_node *rbprev = rb_prev(&last->rb_node);
898 struct request *next, *prev = NULL;
900 /* Follow expired path, else get first next available. */
901 next = bfq_check_fifo(bfqq, last);
906 prev = rb_entry_rq(rbprev);
909 next = rb_entry_rq(rbnext);
911 rbnext = rb_first(&bfqq->sort_list);
912 if (rbnext && rbnext != &last->rb_node)
913 next = rb_entry_rq(rbnext);
916 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
919 /* see the definition of bfq_async_charge_factor for details */
920 static unsigned long bfq_serv_to_charge(struct request *rq,
921 struct bfq_queue *bfqq)
923 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
924 bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
925 return blk_rq_sectors(rq);
927 return blk_rq_sectors(rq) * bfq_async_charge_factor;
931 * bfq_updated_next_req - update the queue after a new next_rq selection.
932 * @bfqd: the device data the queue belongs to.
933 * @bfqq: the queue to update.
935 * If the first request of a queue changes we make sure that the queue
936 * has enough budget to serve at least its first request (if the
937 * request has grown). We do this because if the queue has not enough
938 * budget for its first request, it has to go through two dispatch
939 * rounds to actually get it dispatched.
941 static void bfq_updated_next_req(struct bfq_data *bfqd,
942 struct bfq_queue *bfqq)
944 struct bfq_entity *entity = &bfqq->entity;
945 struct request *next_rq = bfqq->next_rq;
946 unsigned long new_budget;
951 if (bfqq == bfqd->in_service_queue)
953 * In order not to break guarantees, budgets cannot be
954 * changed after an entity has been selected.
958 new_budget = max_t(unsigned long,
959 max_t(unsigned long, bfqq->max_budget,
960 bfq_serv_to_charge(next_rq, bfqq)),
962 if (entity->budget != new_budget) {
963 entity->budget = new_budget;
964 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
966 bfq_requeue_bfqq(bfqd, bfqq, false);
970 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
974 if (bfqd->bfq_wr_max_time > 0)
975 return bfqd->bfq_wr_max_time;
977 dur = bfqd->rate_dur_prod;
978 do_div(dur, bfqd->peak_rate);
981 * Limit duration between 3 and 25 seconds. The upper limit
982 * has been conservatively set after the following worst case:
983 * on a QEMU/KVM virtual machine
984 * - running in a slow PC
985 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
986 * - serving a heavy I/O workload, such as the sequential reading
988 * mplayer took 23 seconds to start, if constantly weight-raised.
990 * As for higher values than that accommodating the above bad
991 * scenario, tests show that higher values would often yield
992 * the opposite of the desired result, i.e., would worsen
993 * responsiveness by allowing non-interactive applications to
994 * preserve weight raising for too long.
996 * On the other end, lower values than 3 seconds make it
997 * difficult for most interactive tasks to complete their jobs
998 * before weight-raising finishes.
1000 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
1003 /* switch back from soft real-time to interactive weight raising */
1004 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
1005 struct bfq_data *bfqd)
1007 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1008 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1009 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
1013 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
1014 struct bfq_io_cq *bic, bool bfq_already_existing)
1016 unsigned int old_wr_coeff = bfqq->wr_coeff;
1017 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
1019 if (bic->saved_has_short_ttime)
1020 bfq_mark_bfqq_has_short_ttime(bfqq);
1022 bfq_clear_bfqq_has_short_ttime(bfqq);
1024 if (bic->saved_IO_bound)
1025 bfq_mark_bfqq_IO_bound(bfqq);
1027 bfq_clear_bfqq_IO_bound(bfqq);
1029 bfqq->entity.new_weight = bic->saved_weight;
1030 bfqq->ttime = bic->saved_ttime;
1031 bfqq->wr_coeff = bic->saved_wr_coeff;
1032 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
1033 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
1034 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
1036 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1037 time_is_before_jiffies(bfqq->last_wr_start_finish +
1038 bfqq->wr_cur_max_time))) {
1039 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1040 !bfq_bfqq_in_large_burst(bfqq) &&
1041 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1042 bfq_wr_duration(bfqd))) {
1043 switch_back_to_interactive_wr(bfqq, bfqd);
1046 bfq_log_bfqq(bfqq->bfqd, bfqq,
1047 "resume state: switching off wr");
1051 /* make sure weight will be updated, however we got here */
1052 bfqq->entity.prio_changed = 1;
1057 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1058 bfqd->wr_busy_queues++;
1059 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1060 bfqd->wr_busy_queues--;
1063 static int bfqq_process_refs(struct bfq_queue *bfqq)
1065 return bfqq->ref - bfqq->allocated - bfqq->entity.on_st -
1066 (bfqq->weight_counter != NULL);
1069 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1070 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1072 struct bfq_queue *item;
1073 struct hlist_node *n;
1075 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1076 hlist_del_init(&item->burst_list_node);
1079 * Start the creation of a new burst list only if there is no
1080 * active queue. See comments on the conditional invocation of
1081 * bfq_handle_burst().
1083 if (bfq_tot_busy_queues(bfqd) == 0) {
1084 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1085 bfqd->burst_size = 1;
1087 bfqd->burst_size = 0;
1089 bfqd->burst_parent_entity = bfqq->entity.parent;
1092 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1093 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1095 /* Increment burst size to take into account also bfqq */
1098 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1099 struct bfq_queue *pos, *bfqq_item;
1100 struct hlist_node *n;
1103 * Enough queues have been activated shortly after each
1104 * other to consider this burst as large.
1106 bfqd->large_burst = true;
1109 * We can now mark all queues in the burst list as
1110 * belonging to a large burst.
1112 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1114 bfq_mark_bfqq_in_large_burst(bfqq_item);
1115 bfq_mark_bfqq_in_large_burst(bfqq);
1118 * From now on, and until the current burst finishes, any
1119 * new queue being activated shortly after the last queue
1120 * was inserted in the burst can be immediately marked as
1121 * belonging to a large burst. So the burst list is not
1122 * needed any more. Remove it.
1124 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1126 hlist_del_init(&pos->burst_list_node);
1128 * Burst not yet large: add bfqq to the burst list. Do
1129 * not increment the ref counter for bfqq, because bfqq
1130 * is removed from the burst list before freeing bfqq
1133 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1137 * If many queues belonging to the same group happen to be created
1138 * shortly after each other, then the processes associated with these
1139 * queues have typically a common goal. In particular, bursts of queue
1140 * creations are usually caused by services or applications that spawn
1141 * many parallel threads/processes. Examples are systemd during boot,
1142 * or git grep. To help these processes get their job done as soon as
1143 * possible, it is usually better to not grant either weight-raising
1144 * or device idling to their queues, unless these queues must be
1145 * protected from the I/O flowing through other active queues.
1147 * In this comment we describe, firstly, the reasons why this fact
1148 * holds, and, secondly, the next function, which implements the main
1149 * steps needed to properly mark these queues so that they can then be
1150 * treated in a different way.
1152 * The above services or applications benefit mostly from a high
1153 * throughput: the quicker the requests of the activated queues are
1154 * cumulatively served, the sooner the target job of these queues gets
1155 * completed. As a consequence, weight-raising any of these queues,
1156 * which also implies idling the device for it, is almost always
1157 * counterproductive, unless there are other active queues to isolate
1158 * these new queues from. If there no other active queues, then
1159 * weight-raising these new queues just lowers throughput in most
1162 * On the other hand, a burst of queue creations may be caused also by
1163 * the start of an application that does not consist of a lot of
1164 * parallel I/O-bound threads. In fact, with a complex application,
1165 * several short processes may need to be executed to start-up the
1166 * application. In this respect, to start an application as quickly as
1167 * possible, the best thing to do is in any case to privilege the I/O
1168 * related to the application with respect to all other
1169 * I/O. Therefore, the best strategy to start as quickly as possible
1170 * an application that causes a burst of queue creations is to
1171 * weight-raise all the queues created during the burst. This is the
1172 * exact opposite of the best strategy for the other type of bursts.
1174 * In the end, to take the best action for each of the two cases, the
1175 * two types of bursts need to be distinguished. Fortunately, this
1176 * seems relatively easy, by looking at the sizes of the bursts. In
1177 * particular, we found a threshold such that only bursts with a
1178 * larger size than that threshold are apparently caused by
1179 * services or commands such as systemd or git grep. For brevity,
1180 * hereafter we call just 'large' these bursts. BFQ *does not*
1181 * weight-raise queues whose creation occurs in a large burst. In
1182 * addition, for each of these queues BFQ performs or does not perform
1183 * idling depending on which choice boosts the throughput more. The
1184 * exact choice depends on the device and request pattern at
1187 * Unfortunately, false positives may occur while an interactive task
1188 * is starting (e.g., an application is being started). The
1189 * consequence is that the queues associated with the task do not
1190 * enjoy weight raising as expected. Fortunately these false positives
1191 * are very rare. They typically occur if some service happens to
1192 * start doing I/O exactly when the interactive task starts.
1194 * Turning back to the next function, it is invoked only if there are
1195 * no active queues (apart from active queues that would belong to the
1196 * same, possible burst bfqq would belong to), and it implements all
1197 * the steps needed to detect the occurrence of a large burst and to
1198 * properly mark all the queues belonging to it (so that they can then
1199 * be treated in a different way). This goal is achieved by
1200 * maintaining a "burst list" that holds, temporarily, the queues that
1201 * belong to the burst in progress. The list is then used to mark
1202 * these queues as belonging to a large burst if the burst does become
1203 * large. The main steps are the following.
1205 * . when the very first queue is created, the queue is inserted into the
1206 * list (as it could be the first queue in a possible burst)
1208 * . if the current burst has not yet become large, and a queue Q that does
1209 * not yet belong to the burst is activated shortly after the last time
1210 * at which a new queue entered the burst list, then the function appends
1211 * Q to the burst list
1213 * . if, as a consequence of the previous step, the burst size reaches
1214 * the large-burst threshold, then
1216 * . all the queues in the burst list are marked as belonging to a
1219 * . the burst list is deleted; in fact, the burst list already served
1220 * its purpose (keeping temporarily track of the queues in a burst,
1221 * so as to be able to mark them as belonging to a large burst in the
1222 * previous sub-step), and now is not needed any more
1224 * . the device enters a large-burst mode
1226 * . if a queue Q that does not belong to the burst is created while
1227 * the device is in large-burst mode and shortly after the last time
1228 * at which a queue either entered the burst list or was marked as
1229 * belonging to the current large burst, then Q is immediately marked
1230 * as belonging to a large burst.
1232 * . if a queue Q that does not belong to the burst is created a while
1233 * later, i.e., not shortly after, than the last time at which a queue
1234 * either entered the burst list or was marked as belonging to the
1235 * current large burst, then the current burst is deemed as finished and:
1237 * . the large-burst mode is reset if set
1239 * . the burst list is emptied
1241 * . Q is inserted in the burst list, as Q may be the first queue
1242 * in a possible new burst (then the burst list contains just Q
1245 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1248 * If bfqq is already in the burst list or is part of a large
1249 * burst, or finally has just been split, then there is
1250 * nothing else to do.
1252 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1253 bfq_bfqq_in_large_burst(bfqq) ||
1254 time_is_after_eq_jiffies(bfqq->split_time +
1255 msecs_to_jiffies(10)))
1259 * If bfqq's creation happens late enough, or bfqq belongs to
1260 * a different group than the burst group, then the current
1261 * burst is finished, and related data structures must be
1264 * In this respect, consider the special case where bfqq is
1265 * the very first queue created after BFQ is selected for this
1266 * device. In this case, last_ins_in_burst and
1267 * burst_parent_entity are not yet significant when we get
1268 * here. But it is easy to verify that, whether or not the
1269 * following condition is true, bfqq will end up being
1270 * inserted into the burst list. In particular the list will
1271 * happen to contain only bfqq. And this is exactly what has
1272 * to happen, as bfqq may be the first queue of the first
1275 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1276 bfqd->bfq_burst_interval) ||
1277 bfqq->entity.parent != bfqd->burst_parent_entity) {
1278 bfqd->large_burst = false;
1279 bfq_reset_burst_list(bfqd, bfqq);
1284 * If we get here, then bfqq is being activated shortly after the
1285 * last queue. So, if the current burst is also large, we can mark
1286 * bfqq as belonging to this large burst immediately.
1288 if (bfqd->large_burst) {
1289 bfq_mark_bfqq_in_large_burst(bfqq);
1294 * If we get here, then a large-burst state has not yet been
1295 * reached, but bfqq is being activated shortly after the last
1296 * queue. Then we add bfqq to the burst.
1298 bfq_add_to_burst(bfqd, bfqq);
1301 * At this point, bfqq either has been added to the current
1302 * burst or has caused the current burst to terminate and a
1303 * possible new burst to start. In particular, in the second
1304 * case, bfqq has become the first queue in the possible new
1305 * burst. In both cases last_ins_in_burst needs to be moved
1308 bfqd->last_ins_in_burst = jiffies;
1311 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1313 struct bfq_entity *entity = &bfqq->entity;
1315 return entity->budget - entity->service;
1319 * If enough samples have been computed, return the current max budget
1320 * stored in bfqd, which is dynamically updated according to the
1321 * estimated disk peak rate; otherwise return the default max budget
1323 static int bfq_max_budget(struct bfq_data *bfqd)
1325 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1326 return bfq_default_max_budget;
1328 return bfqd->bfq_max_budget;
1332 * Return min budget, which is a fraction of the current or default
1333 * max budget (trying with 1/32)
1335 static int bfq_min_budget(struct bfq_data *bfqd)
1337 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1338 return bfq_default_max_budget / 32;
1340 return bfqd->bfq_max_budget / 32;
1344 * The next function, invoked after the input queue bfqq switches from
1345 * idle to busy, updates the budget of bfqq. The function also tells
1346 * whether the in-service queue should be expired, by returning
1347 * true. The purpose of expiring the in-service queue is to give bfqq
1348 * the chance to possibly preempt the in-service queue, and the reason
1349 * for preempting the in-service queue is to achieve one of the two
1352 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1353 * expired because it has remained idle. In particular, bfqq may have
1354 * expired for one of the following two reasons:
1356 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1357 * and did not make it to issue a new request before its last
1358 * request was served;
1360 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1361 * a new request before the expiration of the idling-time.
1363 * Even if bfqq has expired for one of the above reasons, the process
1364 * associated with the queue may be however issuing requests greedily,
1365 * and thus be sensitive to the bandwidth it receives (bfqq may have
1366 * remained idle for other reasons: CPU high load, bfqq not enjoying
1367 * idling, I/O throttling somewhere in the path from the process to
1368 * the I/O scheduler, ...). But if, after every expiration for one of
1369 * the above two reasons, bfqq has to wait for the service of at least
1370 * one full budget of another queue before being served again, then
1371 * bfqq is likely to get a much lower bandwidth or resource time than
1372 * its reserved ones. To address this issue, two countermeasures need
1375 * First, the budget and the timestamps of bfqq need to be updated in
1376 * a special way on bfqq reactivation: they need to be updated as if
1377 * bfqq did not remain idle and did not expire. In fact, if they are
1378 * computed as if bfqq expired and remained idle until reactivation,
1379 * then the process associated with bfqq is treated as if, instead of
1380 * being greedy, it stopped issuing requests when bfqq remained idle,
1381 * and restarts issuing requests only on this reactivation. In other
1382 * words, the scheduler does not help the process recover the "service
1383 * hole" between bfqq expiration and reactivation. As a consequence,
1384 * the process receives a lower bandwidth than its reserved one. In
1385 * contrast, to recover this hole, the budget must be updated as if
1386 * bfqq was not expired at all before this reactivation, i.e., it must
1387 * be set to the value of the remaining budget when bfqq was
1388 * expired. Along the same line, timestamps need to be assigned the
1389 * value they had the last time bfqq was selected for service, i.e.,
1390 * before last expiration. Thus timestamps need to be back-shifted
1391 * with respect to their normal computation (see [1] for more details
1392 * on this tricky aspect).
1394 * Secondly, to allow the process to recover the hole, the in-service
1395 * queue must be expired too, to give bfqq the chance to preempt it
1396 * immediately. In fact, if bfqq has to wait for a full budget of the
1397 * in-service queue to be completed, then it may become impossible to
1398 * let the process recover the hole, even if the back-shifted
1399 * timestamps of bfqq are lower than those of the in-service queue. If
1400 * this happens for most or all of the holes, then the process may not
1401 * receive its reserved bandwidth. In this respect, it is worth noting
1402 * that, being the service of outstanding requests unpreemptible, a
1403 * little fraction of the holes may however be unrecoverable, thereby
1404 * causing a little loss of bandwidth.
1406 * The last important point is detecting whether bfqq does need this
1407 * bandwidth recovery. In this respect, the next function deems the
1408 * process associated with bfqq greedy, and thus allows it to recover
1409 * the hole, if: 1) the process is waiting for the arrival of a new
1410 * request (which implies that bfqq expired for one of the above two
1411 * reasons), and 2) such a request has arrived soon. The first
1412 * condition is controlled through the flag non_blocking_wait_rq,
1413 * while the second through the flag arrived_in_time. If both
1414 * conditions hold, then the function computes the budget in the
1415 * above-described special way, and signals that the in-service queue
1416 * should be expired. Timestamp back-shifting is done later in
1417 * __bfq_activate_entity.
1419 * 2. Reduce latency. Even if timestamps are not backshifted to let
1420 * the process associated with bfqq recover a service hole, bfqq may
1421 * however happen to have, after being (re)activated, a lower finish
1422 * timestamp than the in-service queue. That is, the next budget of
1423 * bfqq may have to be completed before the one of the in-service
1424 * queue. If this is the case, then preempting the in-service queue
1425 * allows this goal to be achieved, apart from the unpreemptible,
1426 * outstanding requests mentioned above.
1428 * Unfortunately, regardless of which of the above two goals one wants
1429 * to achieve, service trees need first to be updated to know whether
1430 * the in-service queue must be preempted. To have service trees
1431 * correctly updated, the in-service queue must be expired and
1432 * rescheduled, and bfqq must be scheduled too. This is one of the
1433 * most costly operations (in future versions, the scheduling
1434 * mechanism may be re-designed in such a way to make it possible to
1435 * know whether preemption is needed without needing to update service
1436 * trees). In addition, queue preemptions almost always cause random
1437 * I/O, which may in turn cause loss of throughput. Finally, there may
1438 * even be no in-service queue when the next function is invoked (so,
1439 * no queue to compare timestamps with). Because of these facts, the
1440 * next function adopts the following simple scheme to avoid costly
1441 * operations, too frequent preemptions and too many dependencies on
1442 * the state of the scheduler: it requests the expiration of the
1443 * in-service queue (unconditionally) only for queues that need to
1444 * recover a hole. Then it delegates to other parts of the code the
1445 * responsibility of handling the above case 2.
1447 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1448 struct bfq_queue *bfqq,
1449 bool arrived_in_time)
1451 struct bfq_entity *entity = &bfqq->entity;
1454 * In the next compound condition, we check also whether there
1455 * is some budget left, because otherwise there is no point in
1456 * trying to go on serving bfqq with this same budget: bfqq
1457 * would be expired immediately after being selected for
1458 * service. This would only cause useless overhead.
1460 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1461 bfq_bfqq_budget_left(bfqq) > 0) {
1463 * We do not clear the flag non_blocking_wait_rq here, as
1464 * the latter is used in bfq_activate_bfqq to signal
1465 * that timestamps need to be back-shifted (and is
1466 * cleared right after).
1470 * In next assignment we rely on that either
1471 * entity->service or entity->budget are not updated
1472 * on expiration if bfqq is empty (see
1473 * __bfq_bfqq_recalc_budget). Thus both quantities
1474 * remain unchanged after such an expiration, and the
1475 * following statement therefore assigns to
1476 * entity->budget the remaining budget on such an
1479 entity->budget = min_t(unsigned long,
1480 bfq_bfqq_budget_left(bfqq),
1484 * At this point, we have used entity->service to get
1485 * the budget left (needed for updating
1486 * entity->budget). Thus we finally can, and have to,
1487 * reset entity->service. The latter must be reset
1488 * because bfqq would otherwise be charged again for
1489 * the service it has received during its previous
1492 entity->service = 0;
1498 * We can finally complete expiration, by setting service to 0.
1500 entity->service = 0;
1501 entity->budget = max_t(unsigned long, bfqq->max_budget,
1502 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1503 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1508 * Return the farthest past time instant according to jiffies
1511 static unsigned long bfq_smallest_from_now(void)
1513 return jiffies - MAX_JIFFY_OFFSET;
1516 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1517 struct bfq_queue *bfqq,
1518 unsigned int old_wr_coeff,
1519 bool wr_or_deserves_wr,
1524 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1525 /* start a weight-raising period */
1527 bfqq->service_from_wr = 0;
1528 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1529 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1532 * No interactive weight raising in progress
1533 * here: assign minus infinity to
1534 * wr_start_at_switch_to_srt, to make sure
1535 * that, at the end of the soft-real-time
1536 * weight raising periods that is starting
1537 * now, no interactive weight-raising period
1538 * may be wrongly considered as still in
1539 * progress (and thus actually started by
1542 bfqq->wr_start_at_switch_to_srt =
1543 bfq_smallest_from_now();
1544 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1545 BFQ_SOFTRT_WEIGHT_FACTOR;
1546 bfqq->wr_cur_max_time =
1547 bfqd->bfq_wr_rt_max_time;
1551 * If needed, further reduce budget to make sure it is
1552 * close to bfqq's backlog, so as to reduce the
1553 * scheduling-error component due to a too large
1554 * budget. Do not care about throughput consequences,
1555 * but only about latency. Finally, do not assign a
1556 * too small budget either, to avoid increasing
1557 * latency by causing too frequent expirations.
1559 bfqq->entity.budget = min_t(unsigned long,
1560 bfqq->entity.budget,
1561 2 * bfq_min_budget(bfqd));
1562 } else if (old_wr_coeff > 1) {
1563 if (interactive) { /* update wr coeff and duration */
1564 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1565 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1566 } else if (in_burst)
1570 * The application is now or still meeting the
1571 * requirements for being deemed soft rt. We
1572 * can then correctly and safely (re)charge
1573 * the weight-raising duration for the
1574 * application with the weight-raising
1575 * duration for soft rt applications.
1577 * In particular, doing this recharge now, i.e.,
1578 * before the weight-raising period for the
1579 * application finishes, reduces the probability
1580 * of the following negative scenario:
1581 * 1) the weight of a soft rt application is
1582 * raised at startup (as for any newly
1583 * created application),
1584 * 2) since the application is not interactive,
1585 * at a certain time weight-raising is
1586 * stopped for the application,
1587 * 3) at that time the application happens to
1588 * still have pending requests, and hence
1589 * is destined to not have a chance to be
1590 * deemed soft rt before these requests are
1591 * completed (see the comments to the
1592 * function bfq_bfqq_softrt_next_start()
1593 * for details on soft rt detection),
1594 * 4) these pending requests experience a high
1595 * latency because the application is not
1596 * weight-raised while they are pending.
1598 if (bfqq->wr_cur_max_time !=
1599 bfqd->bfq_wr_rt_max_time) {
1600 bfqq->wr_start_at_switch_to_srt =
1601 bfqq->last_wr_start_finish;
1603 bfqq->wr_cur_max_time =
1604 bfqd->bfq_wr_rt_max_time;
1605 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1606 BFQ_SOFTRT_WEIGHT_FACTOR;
1608 bfqq->last_wr_start_finish = jiffies;
1613 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1614 struct bfq_queue *bfqq)
1616 return bfqq->dispatched == 0 &&
1617 time_is_before_jiffies(
1618 bfqq->budget_timeout +
1619 bfqd->bfq_wr_min_idle_time);
1624 * Return true if bfqq is in a higher priority class, or has a higher
1625 * weight than the in-service queue.
1627 static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq,
1628 struct bfq_queue *in_serv_bfqq)
1630 int bfqq_weight, in_serv_weight;
1632 if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class)
1635 if (in_serv_bfqq->entity.parent == bfqq->entity.parent) {
1636 bfqq_weight = bfqq->entity.weight;
1637 in_serv_weight = in_serv_bfqq->entity.weight;
1639 if (bfqq->entity.parent)
1640 bfqq_weight = bfqq->entity.parent->weight;
1642 bfqq_weight = bfqq->entity.weight;
1643 if (in_serv_bfqq->entity.parent)
1644 in_serv_weight = in_serv_bfqq->entity.parent->weight;
1646 in_serv_weight = in_serv_bfqq->entity.weight;
1649 return bfqq_weight > in_serv_weight;
1652 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1653 struct bfq_queue *bfqq,
1658 bool soft_rt, in_burst, wr_or_deserves_wr,
1659 bfqq_wants_to_preempt,
1660 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1662 * See the comments on
1663 * bfq_bfqq_update_budg_for_activation for
1664 * details on the usage of the next variable.
1666 arrived_in_time = ktime_get_ns() <=
1667 bfqq->ttime.last_end_request +
1668 bfqd->bfq_slice_idle * 3;
1672 * bfqq deserves to be weight-raised if:
1674 * - it does not belong to a large burst,
1675 * - it has been idle for enough time or is soft real-time,
1676 * - is linked to a bfq_io_cq (it is not shared in any sense).
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 *interactive = !in_burst && idle_for_long_time;
1685 wr_or_deserves_wr = bfqd->low_latency &&
1686 (bfqq->wr_coeff > 1 ||
1687 (bfq_bfqq_sync(bfqq) &&
1688 bfqq->bic && (*interactive || soft_rt)));
1691 * Using the last flag, update budget and check whether bfqq
1692 * may want to preempt the in-service queue.
1694 bfqq_wants_to_preempt =
1695 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1699 * If bfqq happened to be activated in a burst, but has been
1700 * idle for much more than an interactive queue, then we
1701 * assume that, in the overall I/O initiated in the burst, the
1702 * I/O associated with bfqq is finished. So bfqq does not need
1703 * to be treated as a queue belonging to a burst
1704 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1705 * if set, and remove bfqq from the burst list if it's
1706 * there. We do not decrement burst_size, because the fact
1707 * that bfqq does not need to belong to the burst list any
1708 * more does not invalidate the fact that bfqq was created in
1711 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1712 idle_for_long_time &&
1713 time_is_before_jiffies(
1714 bfqq->budget_timeout +
1715 msecs_to_jiffies(10000))) {
1716 hlist_del_init(&bfqq->burst_list_node);
1717 bfq_clear_bfqq_in_large_burst(bfqq);
1720 bfq_clear_bfqq_just_created(bfqq);
1723 if (!bfq_bfqq_IO_bound(bfqq)) {
1724 if (arrived_in_time) {
1725 bfqq->requests_within_timer++;
1726 if (bfqq->requests_within_timer >=
1727 bfqd->bfq_requests_within_timer)
1728 bfq_mark_bfqq_IO_bound(bfqq);
1730 bfqq->requests_within_timer = 0;
1733 if (bfqd->low_latency) {
1734 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1737 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1739 if (time_is_before_jiffies(bfqq->split_time +
1740 bfqd->bfq_wr_min_idle_time)) {
1741 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1748 if (old_wr_coeff != bfqq->wr_coeff)
1749 bfqq->entity.prio_changed = 1;
1753 bfqq->last_idle_bklogged = jiffies;
1754 bfqq->service_from_backlogged = 0;
1755 bfq_clear_bfqq_softrt_update(bfqq);
1757 bfq_add_bfqq_busy(bfqd, bfqq);
1760 * Expire in-service queue only if preemption may be needed
1761 * for guarantees. In particular, we care only about two
1762 * cases. The first is that bfqq has to recover a service
1763 * hole, as explained in the comments on
1764 * bfq_bfqq_update_budg_for_activation(), i.e., that
1765 * bfqq_wants_to_preempt is true. However, if bfqq does not
1766 * carry time-critical I/O, then bfqq's bandwidth is less
1767 * important than that of queues that carry time-critical I/O.
1768 * So, as a further constraint, we consider this case only if
1769 * bfqq is at least as weight-raised, i.e., at least as time
1770 * critical, as the in-service queue.
1772 * The second case is that bfqq is in a higher priority class,
1773 * or has a higher weight than the in-service queue. If this
1774 * condition does not hold, we don't care because, even if
1775 * bfqq does not start to be served immediately, the resulting
1776 * delay for bfqq's I/O is however lower or much lower than
1777 * the ideal completion time to be guaranteed to bfqq's I/O.
1779 * In both cases, preemption is needed only if, according to
1780 * the timestamps of both bfqq and of the in-service queue,
1781 * bfqq actually is the next queue to serve. So, to reduce
1782 * useless preemptions, the return value of
1783 * next_queue_may_preempt() is considered in the next compound
1784 * condition too. Yet next_queue_may_preempt() just checks a
1785 * simple, necessary condition for bfqq to be the next queue
1786 * to serve. In fact, to evaluate a sufficient condition, the
1787 * timestamps of the in-service queue would need to be
1788 * updated, and this operation is quite costly (see the
1789 * comments on bfq_bfqq_update_budg_for_activation()).
1791 if (bfqd->in_service_queue &&
1792 ((bfqq_wants_to_preempt &&
1793 bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) ||
1794 bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue)) &&
1795 next_queue_may_preempt(bfqd))
1796 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1797 false, BFQQE_PREEMPTED);
1800 static void bfq_reset_inject_limit(struct bfq_data *bfqd,
1801 struct bfq_queue *bfqq)
1803 /* invalidate baseline total service time */
1804 bfqq->last_serv_time_ns = 0;
1807 * Reset pointer in case we are waiting for
1808 * some request completion.
1810 bfqd->waited_rq = NULL;
1813 * If bfqq has a short think time, then start by setting the
1814 * inject limit to 0 prudentially, because the service time of
1815 * an injected I/O request may be higher than the think time
1816 * of bfqq, and therefore, if one request was injected when
1817 * bfqq remains empty, this injected request might delay the
1818 * service of the next I/O request for bfqq significantly. In
1819 * case bfqq can actually tolerate some injection, then the
1820 * adaptive update will however raise the limit soon. This
1821 * lucky circumstance holds exactly because bfqq has a short
1822 * think time, and thus, after remaining empty, is likely to
1823 * get new I/O enqueued---and then completed---before being
1824 * expired. This is the very pattern that gives the
1825 * limit-update algorithm the chance to measure the effect of
1826 * injection on request service times, and then to update the
1827 * limit accordingly.
1829 * However, in the following special case, the inject limit is
1830 * left to 1 even if the think time is short: bfqq's I/O is
1831 * synchronized with that of some other queue, i.e., bfqq may
1832 * receive new I/O only after the I/O of the other queue is
1833 * completed. Keeping the inject limit to 1 allows the
1834 * blocking I/O to be served while bfqq is in service. And
1835 * this is very convenient both for bfqq and for overall
1836 * throughput, as explained in detail in the comments in
1837 * bfq_update_has_short_ttime().
1839 * On the opposite end, if bfqq has a long think time, then
1840 * start directly by 1, because:
1841 * a) on the bright side, keeping at most one request in
1842 * service in the drive is unlikely to cause any harm to the
1843 * latency of bfqq's requests, as the service time of a single
1844 * request is likely to be lower than the think time of bfqq;
1845 * b) on the downside, after becoming empty, bfqq is likely to
1846 * expire before getting its next request. With this request
1847 * arrival pattern, it is very hard to sample total service
1848 * times and update the inject limit accordingly (see comments
1849 * on bfq_update_inject_limit()). So the limit is likely to be
1850 * never, or at least seldom, updated. As a consequence, by
1851 * setting the limit to 1, we avoid that no injection ever
1852 * occurs with bfqq. On the downside, this proactive step
1853 * further reduces chances to actually compute the baseline
1854 * total service time. Thus it reduces chances to execute the
1855 * limit-update algorithm and possibly raise the limit to more
1858 if (bfq_bfqq_has_short_ttime(bfqq))
1859 bfqq->inject_limit = 0;
1861 bfqq->inject_limit = 1;
1863 bfqq->decrease_time_jif = jiffies;
1866 static void bfq_add_request(struct request *rq)
1868 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1869 struct bfq_data *bfqd = bfqq->bfqd;
1870 struct request *next_rq, *prev;
1871 unsigned int old_wr_coeff = bfqq->wr_coeff;
1872 bool interactive = false;
1874 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1875 bfqq->queued[rq_is_sync(rq)]++;
1878 if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_bfqq_sync(bfqq)) {
1880 * Detect whether bfqq's I/O seems synchronized with
1881 * that of some other queue, i.e., whether bfqq, after
1882 * remaining empty, happens to receive new I/O only
1883 * right after some I/O request of the other queue has
1884 * been completed. We call waker queue the other
1885 * queue, and we assume, for simplicity, that bfqq may
1886 * have at most one waker queue.
1888 * A remarkable throughput boost can be reached by
1889 * unconditionally injecting the I/O of the waker
1890 * queue, every time a new bfq_dispatch_request
1891 * happens to be invoked while I/O is being plugged
1892 * for bfqq. In addition to boosting throughput, this
1893 * unblocks bfqq's I/O, thereby improving bandwidth
1894 * and latency for bfqq. Note that these same results
1895 * may be achieved with the general injection
1896 * mechanism, but less effectively. For details on
1897 * this aspect, see the comments on the choice of the
1898 * queue for injection in bfq_select_queue().
1900 * Turning back to the detection of a waker queue, a
1901 * queue Q is deemed as a waker queue for bfqq if, for
1902 * two consecutive times, bfqq happens to become non
1903 * empty right after a request of Q has been
1904 * completed. In particular, on the first time, Q is
1905 * tentatively set as a candidate waker queue, while
1906 * on the second time, the flag
1907 * bfq_bfqq_has_waker(bfqq) is set to confirm that Q
1908 * is a waker queue for bfqq. These detection steps
1909 * are performed only if bfqq has a long think time,
1910 * so as to make it more likely that bfqq's I/O is
1911 * actually being blocked by a synchronization. This
1912 * last filter, plus the above two-times requirement,
1913 * make false positives less likely.
1917 * The sooner a waker queue is detected, the sooner
1918 * throughput can be boosted by injecting I/O from the
1919 * waker queue. Fortunately, detection is likely to be
1920 * actually fast, for the following reasons. While
1921 * blocked by synchronization, bfqq has a long think
1922 * time. This implies that bfqq's inject limit is at
1923 * least equal to 1 (see the comments in
1924 * bfq_update_inject_limit()). So, thanks to
1925 * injection, the waker queue is likely to be served
1926 * during the very first I/O-plugging time interval
1927 * for bfqq. This triggers the first step of the
1928 * detection mechanism. Thanks again to injection, the
1929 * candidate waker queue is then likely to be
1930 * confirmed no later than during the next
1931 * I/O-plugging interval for bfqq.
1933 if (bfqd->last_completed_rq_bfqq &&
1934 !bfq_bfqq_has_short_ttime(bfqq) &&
1935 ktime_get_ns() - bfqd->last_completion <
1936 200 * NSEC_PER_USEC) {
1937 if (bfqd->last_completed_rq_bfqq != bfqq &&
1938 bfqd->last_completed_rq_bfqq !=
1941 * First synchronization detected with
1942 * a candidate waker queue, or with a
1943 * different candidate waker queue
1944 * from the current one.
1946 bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
1949 * If the waker queue disappears, then
1950 * bfqq->waker_bfqq must be reset. To
1951 * this goal, we maintain in each
1952 * waker queue a list, woken_list, of
1953 * all the queues that reference the
1954 * waker queue through their
1955 * waker_bfqq pointer. When the waker
1956 * queue exits, the waker_bfqq pointer
1957 * of all the queues in the woken_list
1960 * In addition, if bfqq is already in
1961 * the woken_list of a waker queue,
1962 * then, before being inserted into
1963 * the woken_list of a new waker
1964 * queue, bfqq must be removed from
1965 * the woken_list of the old waker
1968 if (!hlist_unhashed(&bfqq->woken_list_node))
1969 hlist_del_init(&bfqq->woken_list_node);
1970 hlist_add_head(&bfqq->woken_list_node,
1971 &bfqd->last_completed_rq_bfqq->woken_list);
1973 bfq_clear_bfqq_has_waker(bfqq);
1974 } else if (bfqd->last_completed_rq_bfqq ==
1976 !bfq_bfqq_has_waker(bfqq)) {
1978 * synchronization with waker_bfqq
1979 * seen for the second time
1981 bfq_mark_bfqq_has_waker(bfqq);
1986 * Periodically reset inject limit, to make sure that
1987 * the latter eventually drops in case workload
1988 * changes, see step (3) in the comments on
1989 * bfq_update_inject_limit().
1991 if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
1992 msecs_to_jiffies(1000)))
1993 bfq_reset_inject_limit(bfqd, bfqq);
1996 * The following conditions must hold to setup a new
1997 * sampling of total service time, and then a new
1998 * update of the inject limit:
1999 * - bfqq is in service, because the total service
2000 * time is evaluated only for the I/O requests of
2001 * the queues in service;
2002 * - this is the right occasion to compute or to
2003 * lower the baseline total service time, because
2004 * there are actually no requests in the drive,
2006 * the baseline total service time is available, and
2007 * this is the right occasion to compute the other
2008 * quantity needed to update the inject limit, i.e.,
2009 * the total service time caused by the amount of
2010 * injection allowed by the current value of the
2011 * limit. It is the right occasion because injection
2012 * has actually been performed during the service
2013 * hole, and there are still in-flight requests,
2014 * which are very likely to be exactly the injected
2015 * requests, or part of them;
2016 * - the minimum interval for sampling the total
2017 * service time and updating the inject limit has
2020 if (bfqq == bfqd->in_service_queue &&
2021 (bfqd->rq_in_driver == 0 ||
2022 (bfqq->last_serv_time_ns > 0 &&
2023 bfqd->rqs_injected && bfqd->rq_in_driver > 0)) &&
2024 time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2025 msecs_to_jiffies(10))) {
2026 bfqd->last_empty_occupied_ns = ktime_get_ns();
2028 * Start the state machine for measuring the
2029 * total service time of rq: setting
2030 * wait_dispatch will cause bfqd->waited_rq to
2031 * be set when rq will be dispatched.
2033 bfqd->wait_dispatch = true;
2035 * If there is no I/O in service in the drive,
2036 * then possible injection occurred before the
2037 * arrival of rq will not affect the total
2038 * service time of rq. So the injection limit
2039 * must not be updated as a function of such
2040 * total service time, unless new injection
2041 * occurs before rq is completed. To have the
2042 * injection limit updated only in the latter
2043 * case, reset rqs_injected here (rqs_injected
2044 * will be set in case injection is performed
2045 * on bfqq before rq is completed).
2047 if (bfqd->rq_in_driver == 0)
2048 bfqd->rqs_injected = false;
2052 elv_rb_add(&bfqq->sort_list, rq);
2055 * Check if this request is a better next-serve candidate.
2057 prev = bfqq->next_rq;
2058 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
2059 bfqq->next_rq = next_rq;
2062 * Adjust priority tree position, if next_rq changes.
2063 * See comments on bfq_pos_tree_add_move() for the unlikely().
2065 if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
2066 bfq_pos_tree_add_move(bfqd, bfqq);
2068 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
2069 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
2072 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
2073 time_is_before_jiffies(
2074 bfqq->last_wr_start_finish +
2075 bfqd->bfq_wr_min_inter_arr_async)) {
2076 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
2077 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
2079 bfqd->wr_busy_queues++;
2080 bfqq->entity.prio_changed = 1;
2082 if (prev != bfqq->next_rq)
2083 bfq_updated_next_req(bfqd, bfqq);
2087 * Assign jiffies to last_wr_start_finish in the following
2090 * . if bfqq is not going to be weight-raised, because, for
2091 * non weight-raised queues, last_wr_start_finish stores the
2092 * arrival time of the last request; as of now, this piece
2093 * of information is used only for deciding whether to
2094 * weight-raise async queues
2096 * . if bfqq is not weight-raised, because, if bfqq is now
2097 * switching to weight-raised, then last_wr_start_finish
2098 * stores the time when weight-raising starts
2100 * . if bfqq is interactive, because, regardless of whether
2101 * bfqq is currently weight-raised, the weight-raising
2102 * period must start or restart (this case is considered
2103 * separately because it is not detected by the above
2104 * conditions, if bfqq is already weight-raised)
2106 * last_wr_start_finish has to be updated also if bfqq is soft
2107 * real-time, because the weight-raising period is constantly
2108 * restarted on idle-to-busy transitions for these queues, but
2109 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2112 if (bfqd->low_latency &&
2113 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
2114 bfqq->last_wr_start_finish = jiffies;
2117 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
2119 struct request_queue *q)
2121 struct bfq_queue *bfqq = bfqd->bio_bfqq;
2125 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
2130 static sector_t get_sdist(sector_t last_pos, struct request *rq)
2133 return abs(blk_rq_pos(rq) - last_pos);
2138 #if 0 /* Still not clear if we can do without next two functions */
2139 static void bfq_activate_request(struct request_queue *q, struct request *rq)
2141 struct bfq_data *bfqd = q->elevator->elevator_data;
2143 bfqd->rq_in_driver++;
2146 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
2148 struct bfq_data *bfqd = q->elevator->elevator_data;
2150 bfqd->rq_in_driver--;
2154 static void bfq_remove_request(struct request_queue *q,
2157 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2158 struct bfq_data *bfqd = bfqq->bfqd;
2159 const int sync = rq_is_sync(rq);
2161 if (bfqq->next_rq == rq) {
2162 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
2163 bfq_updated_next_req(bfqd, bfqq);
2166 if (rq->queuelist.prev != &rq->queuelist)
2167 list_del_init(&rq->queuelist);
2168 bfqq->queued[sync]--;
2170 elv_rb_del(&bfqq->sort_list, rq);
2172 elv_rqhash_del(q, rq);
2173 if (q->last_merge == rq)
2174 q->last_merge = NULL;
2176 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2177 bfqq->next_rq = NULL;
2179 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
2180 bfq_del_bfqq_busy(bfqd, bfqq, false);
2182 * bfqq emptied. In normal operation, when
2183 * bfqq is empty, bfqq->entity.service and
2184 * bfqq->entity.budget must contain,
2185 * respectively, the service received and the
2186 * budget used last time bfqq emptied. These
2187 * facts do not hold in this case, as at least
2188 * this last removal occurred while bfqq is
2189 * not in service. To avoid inconsistencies,
2190 * reset both bfqq->entity.service and
2191 * bfqq->entity.budget, if bfqq has still a
2192 * process that may issue I/O requests to it.
2194 bfqq->entity.budget = bfqq->entity.service = 0;
2198 * Remove queue from request-position tree as it is empty.
2200 if (bfqq->pos_root) {
2201 rb_erase(&bfqq->pos_node, bfqq->pos_root);
2202 bfqq->pos_root = NULL;
2205 /* see comments on bfq_pos_tree_add_move() for the unlikely() */
2206 if (unlikely(!bfqd->nonrot_with_queueing))
2207 bfq_pos_tree_add_move(bfqd, bfqq);
2210 if (rq->cmd_flags & REQ_META)
2211 bfqq->meta_pending--;
2215 static bool bfq_bio_merge(struct request_queue *q, struct bio *bio,
2216 unsigned int nr_segs)
2218 struct bfq_data *bfqd = q->elevator->elevator_data;
2219 struct request *free = NULL;
2221 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2222 * store its return value for later use, to avoid nesting
2223 * queue_lock inside the bfqd->lock. We assume that the bic
2224 * returned by bfq_bic_lookup does not go away before
2225 * bfqd->lock is taken.
2227 struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
2230 spin_lock_irq(&bfqd->lock);
2234 * Make sure cgroup info is uptodate for current process before
2235 * considering the merge.
2237 bfq_bic_update_cgroup(bic, bio);
2239 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
2241 bfqd->bio_bfqq = NULL;
2243 bfqd->bio_bic = bic;
2245 ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
2248 blk_mq_free_request(free);
2249 spin_unlock_irq(&bfqd->lock);
2254 static int bfq_request_merge(struct request_queue *q, struct request **req,
2257 struct bfq_data *bfqd = q->elevator->elevator_data;
2258 struct request *__rq;
2260 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
2261 if (__rq && elv_bio_merge_ok(__rq, bio)) {
2264 if (blk_discard_mergable(__rq))
2265 return ELEVATOR_DISCARD_MERGE;
2266 return ELEVATOR_FRONT_MERGE;
2269 return ELEVATOR_NO_MERGE;
2272 static void bfq_request_merged(struct request_queue *q, struct request *req,
2273 enum elv_merge type)
2275 if (type == ELEVATOR_FRONT_MERGE &&
2276 rb_prev(&req->rb_node) &&
2278 blk_rq_pos(container_of(rb_prev(&req->rb_node),
2279 struct request, rb_node))) {
2280 struct bfq_queue *bfqq = RQ_BFQQ(req);
2281 struct bfq_data *bfqd;
2282 struct request *prev, *next_rq;
2289 /* Reposition request in its sort_list */
2290 elv_rb_del(&bfqq->sort_list, req);
2291 elv_rb_add(&bfqq->sort_list, req);
2293 /* Choose next request to be served for bfqq */
2294 prev = bfqq->next_rq;
2295 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
2296 bfqd->last_position);
2297 bfqq->next_rq = next_rq;
2299 * If next_rq changes, update both the queue's budget to
2300 * fit the new request and the queue's position in its
2303 if (prev != bfqq->next_rq) {
2304 bfq_updated_next_req(bfqd, bfqq);
2306 * See comments on bfq_pos_tree_add_move() for
2309 if (unlikely(!bfqd->nonrot_with_queueing))
2310 bfq_pos_tree_add_move(bfqd, bfqq);
2316 * This function is called to notify the scheduler that the requests
2317 * rq and 'next' have been merged, with 'next' going away. BFQ
2318 * exploits this hook to address the following issue: if 'next' has a
2319 * fifo_time lower that rq, then the fifo_time of rq must be set to
2320 * the value of 'next', to not forget the greater age of 'next'.
2322 * NOTE: in this function we assume that rq is in a bfq_queue, basing
2323 * on that rq is picked from the hash table q->elevator->hash, which,
2324 * in its turn, is filled only with I/O requests present in
2325 * bfq_queues, while BFQ is in use for the request queue q. In fact,
2326 * the function that fills this hash table (elv_rqhash_add) is called
2327 * only by bfq_insert_request.
2329 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
2330 struct request *next)
2332 struct bfq_queue *bfqq = RQ_BFQQ(rq),
2333 *next_bfqq = RQ_BFQQ(next);
2339 * If next and rq belong to the same bfq_queue and next is older
2340 * than rq, then reposition rq in the fifo (by substituting next
2341 * with rq). Otherwise, if next and rq belong to different
2342 * bfq_queues, never reposition rq: in fact, we would have to
2343 * reposition it with respect to next's position in its own fifo,
2344 * which would most certainly be too expensive with respect to
2347 if (bfqq == next_bfqq &&
2348 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2349 next->fifo_time < rq->fifo_time) {
2350 list_del_init(&rq->queuelist);
2351 list_replace_init(&next->queuelist, &rq->queuelist);
2352 rq->fifo_time = next->fifo_time;
2355 if (bfqq->next_rq == next)
2358 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
2361 /* Must be called with bfqq != NULL */
2362 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
2364 if (bfq_bfqq_busy(bfqq))
2365 bfqq->bfqd->wr_busy_queues--;
2367 bfqq->wr_cur_max_time = 0;
2368 bfqq->last_wr_start_finish = jiffies;
2370 * Trigger a weight change on the next invocation of
2371 * __bfq_entity_update_weight_prio.
2373 bfqq->entity.prio_changed = 1;
2376 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2377 struct bfq_group *bfqg)
2381 for (i = 0; i < 2; i++)
2382 for (j = 0; j < IOPRIO_BE_NR; j++)
2383 if (bfqg->async_bfqq[i][j])
2384 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2385 if (bfqg->async_idle_bfqq)
2386 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2389 static void bfq_end_wr(struct bfq_data *bfqd)
2391 struct bfq_queue *bfqq;
2393 spin_lock_irq(&bfqd->lock);
2395 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2396 bfq_bfqq_end_wr(bfqq);
2397 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2398 bfq_bfqq_end_wr(bfqq);
2399 bfq_end_wr_async(bfqd);
2401 spin_unlock_irq(&bfqd->lock);
2404 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2407 return blk_rq_pos(io_struct);
2409 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2412 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2415 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2419 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2420 struct bfq_queue *bfqq,
2423 struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
2424 struct rb_node *parent, *node;
2425 struct bfq_queue *__bfqq;
2427 if (RB_EMPTY_ROOT(root))
2431 * First, if we find a request starting at the end of the last
2432 * request, choose it.
2434 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2439 * If the exact sector wasn't found, the parent of the NULL leaf
2440 * will contain the closest sector (rq_pos_tree sorted by
2441 * next_request position).
2443 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2444 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2447 if (blk_rq_pos(__bfqq->next_rq) < sector)
2448 node = rb_next(&__bfqq->pos_node);
2450 node = rb_prev(&__bfqq->pos_node);
2454 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2455 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2461 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2462 struct bfq_queue *cur_bfqq,
2465 struct bfq_queue *bfqq;
2468 * We shall notice if some of the queues are cooperating,
2469 * e.g., working closely on the same area of the device. In
2470 * that case, we can group them together and: 1) don't waste
2471 * time idling, and 2) serve the union of their requests in
2472 * the best possible order for throughput.
2474 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2475 if (!bfqq || bfqq == cur_bfqq)
2481 static struct bfq_queue *
2482 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2484 int process_refs, new_process_refs;
2485 struct bfq_queue *__bfqq;
2488 * If there are no process references on the new_bfqq, then it is
2489 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2490 * may have dropped their last reference (not just their last process
2493 if (!bfqq_process_refs(new_bfqq))
2496 /* Avoid a circular list and skip interim queue merges. */
2497 while ((__bfqq = new_bfqq->new_bfqq)) {
2503 process_refs = bfqq_process_refs(bfqq);
2504 new_process_refs = bfqq_process_refs(new_bfqq);
2506 * If the process for the bfqq has gone away, there is no
2507 * sense in merging the queues.
2509 if (process_refs == 0 || new_process_refs == 0)
2513 * Make sure merged queues belong to the same parent. Parents could
2514 * have changed since the time we decided the two queues are suitable
2517 if (new_bfqq->entity.parent != bfqq->entity.parent)
2520 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2524 * Merging is just a redirection: the requests of the process
2525 * owning one of the two queues are redirected to the other queue.
2526 * The latter queue, in its turn, is set as shared if this is the
2527 * first time that the requests of some process are redirected to
2530 * We redirect bfqq to new_bfqq and not the opposite, because
2531 * we are in the context of the process owning bfqq, thus we
2532 * have the io_cq of this process. So we can immediately
2533 * configure this io_cq to redirect the requests of the
2534 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2535 * not available any more (new_bfqq->bic == NULL).
2537 * Anyway, even in case new_bfqq coincides with the in-service
2538 * queue, redirecting requests the in-service queue is the
2539 * best option, as we feed the in-service queue with new
2540 * requests close to the last request served and, by doing so,
2541 * are likely to increase the throughput.
2543 bfqq->new_bfqq = new_bfqq;
2545 * The above assignment schedules the following redirections:
2546 * each time some I/O for bfqq arrives, the process that
2547 * generated that I/O is disassociated from bfqq and
2548 * associated with new_bfqq. Here we increases new_bfqq->ref
2549 * in advance, adding the number of processes that are
2550 * expected to be associated with new_bfqq as they happen to
2553 new_bfqq->ref += process_refs;
2557 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2558 struct bfq_queue *new_bfqq)
2560 if (bfq_too_late_for_merging(new_bfqq))
2563 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2564 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2568 * If either of the queues has already been detected as seeky,
2569 * then merging it with the other queue is unlikely to lead to
2572 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2576 * Interleaved I/O is known to be done by (some) applications
2577 * only for reads, so it does not make sense to merge async
2580 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2587 * Attempt to schedule a merge of bfqq with the currently in-service
2588 * queue or with a close queue among the scheduled queues. Return
2589 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2590 * structure otherwise.
2592 * The OOM queue is not allowed to participate to cooperation: in fact, since
2593 * the requests temporarily redirected to the OOM queue could be redirected
2594 * again to dedicated queues at any time, the state needed to correctly
2595 * handle merging with the OOM queue would be quite complex and expensive
2596 * to maintain. Besides, in such a critical condition as an out of memory,
2597 * the benefits of queue merging may be little relevant, or even negligible.
2599 * WARNING: queue merging may impair fairness among non-weight raised
2600 * queues, for at least two reasons: 1) the original weight of a
2601 * merged queue may change during the merged state, 2) even being the
2602 * weight the same, a merged queue may be bloated with many more
2603 * requests than the ones produced by its originally-associated
2606 static struct bfq_queue *
2607 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2608 void *io_struct, bool request)
2610 struct bfq_queue *in_service_bfqq, *new_bfqq;
2612 /* if a merge has already been setup, then proceed with that first */
2614 return bfqq->new_bfqq;
2617 * Do not perform queue merging if the device is non
2618 * rotational and performs internal queueing. In fact, such a
2619 * device reaches a high speed through internal parallelism
2620 * and pipelining. This means that, to reach a high
2621 * throughput, it must have many requests enqueued at the same
2622 * time. But, in this configuration, the internal scheduling
2623 * algorithm of the device does exactly the job of queue
2624 * merging: it reorders requests so as to obtain as much as
2625 * possible a sequential I/O pattern. As a consequence, with
2626 * the workload generated by processes doing interleaved I/O,
2627 * the throughput reached by the device is likely to be the
2628 * same, with and without queue merging.
2630 * Disabling merging also provides a remarkable benefit in
2631 * terms of throughput. Merging tends to make many workloads
2632 * artificially more uneven, because of shared queues
2633 * remaining non empty for incomparably more time than
2634 * non-merged queues. This may accentuate workload
2635 * asymmetries. For example, if one of the queues in a set of
2636 * merged queues has a higher weight than a normal queue, then
2637 * the shared queue may inherit such a high weight and, by
2638 * staying almost always active, may force BFQ to perform I/O
2639 * plugging most of the time. This evidently makes it harder
2640 * for BFQ to let the device reach a high throughput.
2642 * Finally, the likely() macro below is not used because one
2643 * of the two branches is more likely than the other, but to
2644 * have the code path after the following if() executed as
2645 * fast as possible for the case of a non rotational device
2646 * with queueing. We want it because this is the fastest kind
2647 * of device. On the opposite end, the likely() may lengthen
2648 * the execution time of BFQ for the case of slower devices
2649 * (rotational or at least without queueing). But in this case
2650 * the execution time of BFQ matters very little, if not at
2653 if (likely(bfqd->nonrot_with_queueing))
2657 * Prevent bfqq from being merged if it has been created too
2658 * long ago. The idea is that true cooperating processes, and
2659 * thus their associated bfq_queues, are supposed to be
2660 * created shortly after each other. This is the case, e.g.,
2661 * for KVM/QEMU and dump I/O threads. Basing on this
2662 * assumption, the following filtering greatly reduces the
2663 * probability that two non-cooperating processes, which just
2664 * happen to do close I/O for some short time interval, have
2665 * their queues merged by mistake.
2667 if (bfq_too_late_for_merging(bfqq))
2670 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2673 /* If there is only one backlogged queue, don't search. */
2674 if (bfq_tot_busy_queues(bfqd) == 1)
2677 in_service_bfqq = bfqd->in_service_queue;
2679 if (in_service_bfqq && in_service_bfqq != bfqq &&
2680 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2681 bfq_rq_close_to_sector(io_struct, request,
2682 bfqd->in_serv_last_pos) &&
2683 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2684 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2685 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2690 * Check whether there is a cooperator among currently scheduled
2691 * queues. The only thing we need is that the bio/request is not
2692 * NULL, as we need it to establish whether a cooperator exists.
2694 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2695 bfq_io_struct_pos(io_struct, request));
2697 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2698 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2699 return bfq_setup_merge(bfqq, new_bfqq);
2704 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2706 struct bfq_io_cq *bic = bfqq->bic;
2709 * If !bfqq->bic, the queue is already shared or its requests
2710 * have already been redirected to a shared queue; both idle window
2711 * and weight raising state have already been saved. Do nothing.
2716 bic->saved_weight = bfqq->entity.orig_weight;
2717 bic->saved_ttime = bfqq->ttime;
2718 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2719 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2720 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2721 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2722 if (unlikely(bfq_bfqq_just_created(bfqq) &&
2723 !bfq_bfqq_in_large_burst(bfqq) &&
2724 bfqq->bfqd->low_latency)) {
2726 * bfqq being merged right after being created: bfqq
2727 * would have deserved interactive weight raising, but
2728 * did not make it to be set in a weight-raised state,
2729 * because of this early merge. Store directly the
2730 * weight-raising state that would have been assigned
2731 * to bfqq, so that to avoid that bfqq unjustly fails
2732 * to enjoy weight raising if split soon.
2734 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
2735 bic->saved_wr_start_at_switch_to_srt = bfq_smallest_from_now();
2736 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2737 bic->saved_last_wr_start_finish = jiffies;
2739 bic->saved_wr_coeff = bfqq->wr_coeff;
2740 bic->saved_wr_start_at_switch_to_srt =
2741 bfqq->wr_start_at_switch_to_srt;
2742 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2743 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2747 void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
2750 * To prevent bfqq's service guarantees from being violated,
2751 * bfqq may be left busy, i.e., queued for service, even if
2752 * empty (see comments in __bfq_bfqq_expire() for
2753 * details). But, if no process will send requests to bfqq any
2754 * longer, then there is no point in keeping bfqq queued for
2755 * service. In addition, keeping bfqq queued for service, but
2756 * with no process ref any longer, may have caused bfqq to be
2757 * freed when dequeued from service. But this is assumed to
2760 if (bfq_bfqq_busy(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) &&
2761 bfqq != bfqd->in_service_queue)
2762 bfq_del_bfqq_busy(bfqd, bfqq, false);
2764 bfq_put_queue(bfqq);
2768 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2769 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2771 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2772 (unsigned long)new_bfqq->pid);
2773 /* Save weight raising and idle window of the merged queues */
2774 bfq_bfqq_save_state(bfqq);
2775 bfq_bfqq_save_state(new_bfqq);
2776 if (bfq_bfqq_IO_bound(bfqq))
2777 bfq_mark_bfqq_IO_bound(new_bfqq);
2778 bfq_clear_bfqq_IO_bound(bfqq);
2781 * If bfqq is weight-raised, then let new_bfqq inherit
2782 * weight-raising. To reduce false positives, neglect the case
2783 * where bfqq has just been created, but has not yet made it
2784 * to be weight-raised (which may happen because EQM may merge
2785 * bfqq even before bfq_add_request is executed for the first
2786 * time for bfqq). Handling this case would however be very
2787 * easy, thanks to the flag just_created.
2789 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2790 new_bfqq->wr_coeff = bfqq->wr_coeff;
2791 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2792 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2793 new_bfqq->wr_start_at_switch_to_srt =
2794 bfqq->wr_start_at_switch_to_srt;
2795 if (bfq_bfqq_busy(new_bfqq))
2796 bfqd->wr_busy_queues++;
2797 new_bfqq->entity.prio_changed = 1;
2800 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2802 bfqq->entity.prio_changed = 1;
2803 if (bfq_bfqq_busy(bfqq))
2804 bfqd->wr_busy_queues--;
2807 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2808 bfqd->wr_busy_queues);
2811 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2813 bic_set_bfqq(bic, new_bfqq, 1);
2814 bfq_mark_bfqq_coop(new_bfqq);
2816 * new_bfqq now belongs to at least two bics (it is a shared queue):
2817 * set new_bfqq->bic to NULL. bfqq either:
2818 * - does not belong to any bic any more, and hence bfqq->bic must
2819 * be set to NULL, or
2820 * - is a queue whose owning bics have already been redirected to a
2821 * different queue, hence the queue is destined to not belong to
2822 * any bic soon and bfqq->bic is already NULL (therefore the next
2823 * assignment causes no harm).
2825 new_bfqq->bic = NULL;
2827 * If the queue is shared, the pid is the pid of one of the associated
2828 * processes. Which pid depends on the exact sequence of merge events
2829 * the queue underwent. So printing such a pid is useless and confusing
2830 * because it reports a random pid between those of the associated
2832 * We mark such a queue with a pid -1, and then print SHARED instead of
2833 * a pid in logging messages.
2837 bfq_release_process_ref(bfqd, bfqq);
2840 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2843 struct bfq_data *bfqd = q->elevator->elevator_data;
2844 bool is_sync = op_is_sync(bio->bi_opf);
2845 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2848 * Disallow merge of a sync bio into an async request.
2850 if (is_sync && !rq_is_sync(rq))
2854 * Lookup the bfqq that this bio will be queued with. Allow
2855 * merge only if rq is queued there.
2861 * We take advantage of this function to perform an early merge
2862 * of the queues of possible cooperating processes.
2864 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2867 * bic still points to bfqq, then it has not yet been
2868 * redirected to some other bfq_queue, and a queue
2869 * merge between bfqq and new_bfqq can be safely
2870 * fulfilled, i.e., bic can be redirected to new_bfqq
2871 * and bfqq can be put.
2873 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2876 * If we get here, bio will be queued into new_queue,
2877 * so use new_bfqq to decide whether bio and rq can be
2883 * Change also bqfd->bio_bfqq, as
2884 * bfqd->bio_bic now points to new_bfqq, and
2885 * this function may be invoked again (and then may
2886 * use again bqfd->bio_bfqq).
2888 bfqd->bio_bfqq = bfqq;
2891 return bfqq == RQ_BFQQ(rq);
2895 * Set the maximum time for the in-service queue to consume its
2896 * budget. This prevents seeky processes from lowering the throughput.
2897 * In practice, a time-slice service scheme is used with seeky
2900 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2901 struct bfq_queue *bfqq)
2903 unsigned int timeout_coeff;
2905 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2908 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2910 bfqd->last_budget_start = ktime_get();
2912 bfqq->budget_timeout = jiffies +
2913 bfqd->bfq_timeout * timeout_coeff;
2916 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2917 struct bfq_queue *bfqq)
2920 bfq_clear_bfqq_fifo_expire(bfqq);
2922 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2924 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2925 bfqq->wr_coeff > 1 &&
2926 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2927 time_is_before_jiffies(bfqq->budget_timeout)) {
2929 * For soft real-time queues, move the start
2930 * of the weight-raising period forward by the
2931 * time the queue has not received any
2932 * service. Otherwise, a relatively long
2933 * service delay is likely to cause the
2934 * weight-raising period of the queue to end,
2935 * because of the short duration of the
2936 * weight-raising period of a soft real-time
2937 * queue. It is worth noting that this move
2938 * is not so dangerous for the other queues,
2939 * because soft real-time queues are not
2942 * To not add a further variable, we use the
2943 * overloaded field budget_timeout to
2944 * determine for how long the queue has not
2945 * received service, i.e., how much time has
2946 * elapsed since the queue expired. However,
2947 * this is a little imprecise, because
2948 * budget_timeout is set to jiffies if bfqq
2949 * not only expires, but also remains with no
2952 if (time_after(bfqq->budget_timeout,
2953 bfqq->last_wr_start_finish))
2954 bfqq->last_wr_start_finish +=
2955 jiffies - bfqq->budget_timeout;
2957 bfqq->last_wr_start_finish = jiffies;
2960 bfq_set_budget_timeout(bfqd, bfqq);
2961 bfq_log_bfqq(bfqd, bfqq,
2962 "set_in_service_queue, cur-budget = %d",
2963 bfqq->entity.budget);
2966 bfqd->in_service_queue = bfqq;
2967 bfqd->in_serv_last_pos = 0;
2971 * Get and set a new queue for service.
2973 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2975 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2977 __bfq_set_in_service_queue(bfqd, bfqq);
2981 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2983 struct bfq_queue *bfqq = bfqd->in_service_queue;
2986 bfq_mark_bfqq_wait_request(bfqq);
2989 * We don't want to idle for seeks, but we do want to allow
2990 * fair distribution of slice time for a process doing back-to-back
2991 * seeks. So allow a little bit of time for him to submit a new rq.
2993 sl = bfqd->bfq_slice_idle;
2995 * Unless the queue is being weight-raised or the scenario is
2996 * asymmetric, grant only minimum idle time if the queue
2997 * is seeky. A long idling is preserved for a weight-raised
2998 * queue, or, more in general, in an asymmetric scenario,
2999 * because a long idling is needed for guaranteeing to a queue
3000 * its reserved share of the throughput (in particular, it is
3001 * needed if the queue has a higher weight than some other
3004 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3005 !bfq_asymmetric_scenario(bfqd, bfqq))
3006 sl = min_t(u64, sl, BFQ_MIN_TT);
3007 else if (bfqq->wr_coeff > 1)
3008 sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
3010 bfqd->last_idling_start = ktime_get();
3011 bfqd->last_idling_start_jiffies = jiffies;
3013 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
3015 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
3019 * In autotuning mode, max_budget is dynamically recomputed as the
3020 * amount of sectors transferred in timeout at the estimated peak
3021 * rate. This enables BFQ to utilize a full timeslice with a full
3022 * budget, even if the in-service queue is served at peak rate. And
3023 * this maximises throughput with sequential workloads.
3025 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
3027 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
3028 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
3032 * Update parameters related to throughput and responsiveness, as a
3033 * function of the estimated peak rate. See comments on
3034 * bfq_calc_max_budget(), and on the ref_wr_duration array.
3036 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
3038 if (bfqd->bfq_user_max_budget == 0) {
3039 bfqd->bfq_max_budget =
3040 bfq_calc_max_budget(bfqd);
3041 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
3045 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
3048 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
3049 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
3050 bfqd->peak_rate_samples = 1;
3051 bfqd->sequential_samples = 0;
3052 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
3054 } else /* no new rq dispatched, just reset the number of samples */
3055 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
3058 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
3059 bfqd->peak_rate_samples, bfqd->sequential_samples,
3060 bfqd->tot_sectors_dispatched);
3063 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
3065 u32 rate, weight, divisor;
3068 * For the convergence property to hold (see comments on
3069 * bfq_update_peak_rate()) and for the assessment to be
3070 * reliable, a minimum number of samples must be present, and
3071 * a minimum amount of time must have elapsed. If not so, do
3072 * not compute new rate. Just reset parameters, to get ready
3073 * for a new evaluation attempt.
3075 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
3076 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
3077 goto reset_computation;
3080 * If a new request completion has occurred after last
3081 * dispatch, then, to approximate the rate at which requests
3082 * have been served by the device, it is more precise to
3083 * extend the observation interval to the last completion.
3085 bfqd->delta_from_first =
3086 max_t(u64, bfqd->delta_from_first,
3087 bfqd->last_completion - bfqd->first_dispatch);
3090 * Rate computed in sects/usec, and not sects/nsec, for
3093 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
3094 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
3097 * Peak rate not updated if:
3098 * - the percentage of sequential dispatches is below 3/4 of the
3099 * total, and rate is below the current estimated peak rate
3100 * - rate is unreasonably high (> 20M sectors/sec)
3102 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
3103 rate <= bfqd->peak_rate) ||
3104 rate > 20<<BFQ_RATE_SHIFT)
3105 goto reset_computation;
3108 * We have to update the peak rate, at last! To this purpose,
3109 * we use a low-pass filter. We compute the smoothing constant
3110 * of the filter as a function of the 'weight' of the new
3113 * As can be seen in next formulas, we define this weight as a
3114 * quantity proportional to how sequential the workload is,
3115 * and to how long the observation time interval is.
3117 * The weight runs from 0 to 8. The maximum value of the
3118 * weight, 8, yields the minimum value for the smoothing
3119 * constant. At this minimum value for the smoothing constant,
3120 * the measured rate contributes for half of the next value of
3121 * the estimated peak rate.
3123 * So, the first step is to compute the weight as a function
3124 * of how sequential the workload is. Note that the weight
3125 * cannot reach 9, because bfqd->sequential_samples cannot
3126 * become equal to bfqd->peak_rate_samples, which, in its
3127 * turn, holds true because bfqd->sequential_samples is not
3128 * incremented for the first sample.
3130 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
3133 * Second step: further refine the weight as a function of the
3134 * duration of the observation interval.
3136 weight = min_t(u32, 8,
3137 div_u64(weight * bfqd->delta_from_first,
3138 BFQ_RATE_REF_INTERVAL));
3141 * Divisor ranging from 10, for minimum weight, to 2, for
3144 divisor = 10 - weight;
3147 * Finally, update peak rate:
3149 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
3151 bfqd->peak_rate *= divisor-1;
3152 bfqd->peak_rate /= divisor;
3153 rate /= divisor; /* smoothing constant alpha = 1/divisor */
3155 bfqd->peak_rate += rate;
3158 * For a very slow device, bfqd->peak_rate can reach 0 (see
3159 * the minimum representable values reported in the comments
3160 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3161 * divisions by zero where bfqd->peak_rate is used as a
3164 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
3166 update_thr_responsiveness_params(bfqd);
3169 bfq_reset_rate_computation(bfqd, rq);
3173 * Update the read/write peak rate (the main quantity used for
3174 * auto-tuning, see update_thr_responsiveness_params()).
3176 * It is not trivial to estimate the peak rate (correctly): because of
3177 * the presence of sw and hw queues between the scheduler and the
3178 * device components that finally serve I/O requests, it is hard to
3179 * say exactly when a given dispatched request is served inside the
3180 * device, and for how long. As a consequence, it is hard to know
3181 * precisely at what rate a given set of requests is actually served
3184 * On the opposite end, the dispatch time of any request is trivially
3185 * available, and, from this piece of information, the "dispatch rate"
3186 * of requests can be immediately computed. So, the idea in the next
3187 * function is to use what is known, namely request dispatch times
3188 * (plus, when useful, request completion times), to estimate what is
3189 * unknown, namely in-device request service rate.
3191 * The main issue is that, because of the above facts, the rate at
3192 * which a certain set of requests is dispatched over a certain time
3193 * interval can vary greatly with respect to the rate at which the
3194 * same requests are then served. But, since the size of any
3195 * intermediate queue is limited, and the service scheme is lossless
3196 * (no request is silently dropped), the following obvious convergence
3197 * property holds: the number of requests dispatched MUST become
3198 * closer and closer to the number of requests completed as the
3199 * observation interval grows. This is the key property used in
3200 * the next function to estimate the peak service rate as a function
3201 * of the observed dispatch rate. The function assumes to be invoked
3202 * on every request dispatch.
3204 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
3206 u64 now_ns = ktime_get_ns();
3208 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
3209 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
3210 bfqd->peak_rate_samples);
3211 bfq_reset_rate_computation(bfqd, rq);
3212 goto update_last_values; /* will add one sample */
3216 * Device idle for very long: the observation interval lasting
3217 * up to this dispatch cannot be a valid observation interval
3218 * for computing a new peak rate (similarly to the late-
3219 * completion event in bfq_completed_request()). Go to
3220 * update_rate_and_reset to have the following three steps
3222 * - close the observation interval at the last (previous)
3223 * request dispatch or completion
3224 * - compute rate, if possible, for that observation interval
3225 * - start a new observation interval with this dispatch
3227 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
3228 bfqd->rq_in_driver == 0)
3229 goto update_rate_and_reset;
3231 /* Update sampling information */
3232 bfqd->peak_rate_samples++;
3234 if ((bfqd->rq_in_driver > 0 ||
3235 now_ns - bfqd->last_completion < BFQ_MIN_TT)
3236 && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
3237 bfqd->sequential_samples++;
3239 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
3241 /* Reset max observed rq size every 32 dispatches */
3242 if (likely(bfqd->peak_rate_samples % 32))
3243 bfqd->last_rq_max_size =
3244 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
3246 bfqd->last_rq_max_size = blk_rq_sectors(rq);
3248 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
3250 /* Target observation interval not yet reached, go on sampling */
3251 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
3252 goto update_last_values;
3254 update_rate_and_reset:
3255 bfq_update_rate_reset(bfqd, rq);
3257 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
3258 if (RQ_BFQQ(rq) == bfqd->in_service_queue)
3259 bfqd->in_serv_last_pos = bfqd->last_position;
3260 bfqd->last_dispatch = now_ns;
3264 * Remove request from internal lists.
3266 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
3268 struct bfq_queue *bfqq = RQ_BFQQ(rq);
3271 * For consistency, the next instruction should have been
3272 * executed after removing the request from the queue and
3273 * dispatching it. We execute instead this instruction before
3274 * bfq_remove_request() (and hence introduce a temporary
3275 * inconsistency), for efficiency. In fact, should this
3276 * dispatch occur for a non in-service bfqq, this anticipated
3277 * increment prevents two counters related to bfqq->dispatched
3278 * from risking to be, first, uselessly decremented, and then
3279 * incremented again when the (new) value of bfqq->dispatched
3280 * happens to be taken into account.
3283 bfq_update_peak_rate(q->elevator->elevator_data, rq);
3285 bfq_remove_request(q, rq);
3289 * There is a case where idling does not have to be performed for
3290 * throughput concerns, but to preserve the throughput share of
3291 * the process associated with bfqq.
3293 * To introduce this case, we can note that allowing the drive
3294 * to enqueue more than one request at a time, and hence
3295 * delegating de facto final scheduling decisions to the
3296 * drive's internal scheduler, entails loss of control on the
3297 * actual request service order. In particular, the critical
3298 * situation is when requests from different processes happen
3299 * to be present, at the same time, in the internal queue(s)
3300 * of the drive. In such a situation, the drive, by deciding
3301 * the service order of the internally-queued requests, does
3302 * determine also the actual throughput distribution among
3303 * these processes. But the drive typically has no notion or
3304 * concern about per-process throughput distribution, and
3305 * makes its decisions only on a per-request basis. Therefore,
3306 * the service distribution enforced by the drive's internal
3307 * scheduler is likely to coincide with the desired throughput
3308 * distribution only in a completely symmetric, or favorably
3309 * skewed scenario where:
3310 * (i-a) each of these processes must get the same throughput as
3312 * (i-b) in case (i-a) does not hold, it holds that the process
3313 * associated with bfqq must receive a lower or equal
3314 * throughput than any of the other processes;
3315 * (ii) the I/O of each process has the same properties, in
3316 * terms of locality (sequential or random), direction
3317 * (reads or writes), request sizes, greediness
3318 * (from I/O-bound to sporadic), and so on;
3320 * In fact, in such a scenario, the drive tends to treat the requests
3321 * of each process in about the same way as the requests of the
3322 * others, and thus to provide each of these processes with about the
3323 * same throughput. This is exactly the desired throughput
3324 * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3325 * even more convenient distribution for (the process associated with)
3328 * In contrast, in any asymmetric or unfavorable scenario, device
3329 * idling (I/O-dispatch plugging) is certainly needed to guarantee
3330 * that bfqq receives its assigned fraction of the device throughput
3331 * (see [1] for details).
3333 * The problem is that idling may significantly reduce throughput with
3334 * certain combinations of types of I/O and devices. An important
3335 * example is sync random I/O on flash storage with command
3336 * queueing. So, unless bfqq falls in cases where idling also boosts
3337 * throughput, it is important to check conditions (i-a), i(-b) and
3338 * (ii) accurately, so as to avoid idling when not strictly needed for
3339 * service guarantees.
3341 * Unfortunately, it is extremely difficult to thoroughly check
3342 * condition (ii). And, in case there are active groups, it becomes
3343 * very difficult to check conditions (i-a) and (i-b) too. In fact,
3344 * if there are active groups, then, for conditions (i-a) or (i-b) to
3345 * become false 'indirectly', it is enough that an active group
3346 * contains more active processes or sub-groups than some other active
3347 * group. More precisely, for conditions (i-a) or (i-b) to become
3348 * false because of such a group, it is not even necessary that the
3349 * group is (still) active: it is sufficient that, even if the group
3350 * has become inactive, some of its descendant processes still have
3351 * some request already dispatched but still waiting for
3352 * completion. In fact, requests have still to be guaranteed their
3353 * share of the throughput even after being dispatched. In this
3354 * respect, it is easy to show that, if a group frequently becomes
3355 * inactive while still having in-flight requests, and if, when this
3356 * happens, the group is not considered in the calculation of whether
3357 * the scenario is asymmetric, then the group may fail to be
3358 * guaranteed its fair share of the throughput (basically because
3359 * idling may not be performed for the descendant processes of the
3360 * group, but it had to be). We address this issue with the following
3361 * bi-modal behavior, implemented in the function
3362 * bfq_asymmetric_scenario().
3364 * If there are groups with requests waiting for completion
3365 * (as commented above, some of these groups may even be
3366 * already inactive), then the scenario is tagged as
3367 * asymmetric, conservatively, without checking any of the
3368 * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3369 * This behavior matches also the fact that groups are created
3370 * exactly if controlling I/O is a primary concern (to
3371 * preserve bandwidth and latency guarantees).
3373 * On the opposite end, if there are no groups with requests waiting
3374 * for completion, then only conditions (i-a) and (i-b) are actually
3375 * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3376 * idling is not performed, regardless of whether condition (ii)
3377 * holds. In other words, only if conditions (i-a) and (i-b) do not
3378 * hold, then idling is allowed, and the device tends to be prevented
3379 * from queueing many requests, possibly of several processes. Since
3380 * there are no groups with requests waiting for completion, then, to
3381 * control conditions (i-a) and (i-b) it is enough to check just
3382 * whether all the queues with requests waiting for completion also
3383 * have the same weight.
3385 * Not checking condition (ii) evidently exposes bfqq to the
3386 * risk of getting less throughput than its fair share.
3387 * However, for queues with the same weight, a further
3388 * mechanism, preemption, mitigates or even eliminates this
3389 * problem. And it does so without consequences on overall
3390 * throughput. This mechanism and its benefits are explained
3391 * in the next three paragraphs.
3393 * Even if a queue, say Q, is expired when it remains idle, Q
3394 * can still preempt the new in-service queue if the next
3395 * request of Q arrives soon (see the comments on
3396 * bfq_bfqq_update_budg_for_activation). If all queues and
3397 * groups have the same weight, this form of preemption,
3398 * combined with the hole-recovery heuristic described in the
3399 * comments on function bfq_bfqq_update_budg_for_activation,
3400 * are enough to preserve a correct bandwidth distribution in
3401 * the mid term, even without idling. In fact, even if not
3402 * idling allows the internal queues of the device to contain
3403 * many requests, and thus to reorder requests, we can rather
3404 * safely assume that the internal scheduler still preserves a
3405 * minimum of mid-term fairness.
3407 * More precisely, this preemption-based, idleless approach
3408 * provides fairness in terms of IOPS, and not sectors per
3409 * second. This can be seen with a simple example. Suppose
3410 * that there are two queues with the same weight, but that
3411 * the first queue receives requests of 8 sectors, while the
3412 * second queue receives requests of 1024 sectors. In
3413 * addition, suppose that each of the two queues contains at
3414 * most one request at a time, which implies that each queue
3415 * always remains idle after it is served. Finally, after
3416 * remaining idle, each queue receives very quickly a new
3417 * request. It follows that the two queues are served
3418 * alternatively, preempting each other if needed. This
3419 * implies that, although both queues have the same weight,
3420 * the queue with large requests receives a service that is
3421 * 1024/8 times as high as the service received by the other
3424 * The motivation for using preemption instead of idling (for
3425 * queues with the same weight) is that, by not idling,
3426 * service guarantees are preserved (completely or at least in
3427 * part) without minimally sacrificing throughput. And, if
3428 * there is no active group, then the primary expectation for
3429 * this device is probably a high throughput.
3431 * We are now left only with explaining the two sub-conditions in the
3432 * additional compound condition that is checked below for deciding
3433 * whether the scenario is asymmetric. To explain the first
3434 * sub-condition, we need to add that the function
3435 * bfq_asymmetric_scenario checks the weights of only
3436 * non-weight-raised queues, for efficiency reasons (see comments on
3437 * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3438 * is checked explicitly here. More precisely, the compound condition
3439 * below takes into account also the fact that, even if bfqq is being
3440 * weight-raised, the scenario is still symmetric if all queues with
3441 * requests waiting for completion happen to be
3442 * weight-raised. Actually, we should be even more precise here, and
3443 * differentiate between interactive weight raising and soft real-time
3446 * The second sub-condition checked in the compound condition is
3447 * whether there is a fair amount of already in-flight I/O not
3448 * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3449 * following reason. The drive may decide to serve in-flight
3450 * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3451 * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3452 * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3453 * basically uncontrolled amount of I/O from other queues may be
3454 * dispatched too, possibly causing the service of bfqq's I/O to be
3455 * delayed even longer in the drive. This problem gets more and more
3456 * serious as the speed and the queue depth of the drive grow,
3457 * because, as these two quantities grow, the probability to find no
3458 * queue busy but many requests in flight grows too. By contrast,
3459 * plugging I/O dispatching minimizes the delay induced by already
3460 * in-flight I/O, and enables bfqq to recover the bandwidth it may
3461 * lose because of this delay.
3463 * As a side note, it is worth considering that the above
3464 * device-idling countermeasures may however fail in the following
3465 * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3466 * in a time period during which all symmetry sub-conditions hold, and
3467 * therefore the device is allowed to enqueue many requests, but at
3468 * some later point in time some sub-condition stops to hold, then it
3469 * may become impossible to make requests be served in the desired
3470 * order until all the requests already queued in the device have been
3471 * served. The last sub-condition commented above somewhat mitigates
3472 * this problem for weight-raised queues.
3474 static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3475 struct bfq_queue *bfqq)
3477 /* No point in idling for bfqq if it won't get requests any longer */
3478 if (unlikely(!bfqq_process_refs(bfqq)))
3481 return (bfqq->wr_coeff > 1 &&
3482 (bfqd->wr_busy_queues <
3483 bfq_tot_busy_queues(bfqd) ||
3484 bfqd->rq_in_driver >=
3485 bfqq->dispatched + 4)) ||
3486 bfq_asymmetric_scenario(bfqd, bfqq);
3489 static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3490 enum bfqq_expiration reason)
3493 * If this bfqq is shared between multiple processes, check
3494 * to make sure that those processes are still issuing I/Os
3495 * within the mean seek distance. If not, it may be time to
3496 * break the queues apart again.
3498 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
3499 bfq_mark_bfqq_split_coop(bfqq);
3502 * Consider queues with a higher finish virtual time than
3503 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3504 * true, then bfqq's bandwidth would be violated if an
3505 * uncontrolled amount of I/O from these queues were
3506 * dispatched while bfqq is waiting for its new I/O to
3507 * arrive. This is exactly what may happen if this is a forced
3508 * expiration caused by a preemption attempt, and if bfqq is
3509 * not re-scheduled. To prevent this from happening, re-queue
3510 * bfqq if it needs I/O-dispatch plugging, even if it is
3511 * empty. By doing so, bfqq is granted to be served before the
3512 * above queues (provided that bfqq is of course eligible).
3514 if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
3515 !(reason == BFQQE_PREEMPTED &&
3516 idling_needed_for_service_guarantees(bfqd, bfqq))) {
3517 if (bfqq->dispatched == 0)
3519 * Overloading budget_timeout field to store
3520 * the time at which the queue remains with no
3521 * backlog and no outstanding request; used by
3522 * the weight-raising mechanism.
3524 bfqq->budget_timeout = jiffies;
3526 bfq_del_bfqq_busy(bfqd, bfqq, true);
3528 bfq_requeue_bfqq(bfqd, bfqq, true);
3530 * Resort priority tree of potential close cooperators.
3531 * See comments on bfq_pos_tree_add_move() for the unlikely().
3533 if (unlikely(!bfqd->nonrot_with_queueing &&
3534 !RB_EMPTY_ROOT(&bfqq->sort_list)))
3535 bfq_pos_tree_add_move(bfqd, bfqq);
3539 * All in-service entities must have been properly deactivated
3540 * or requeued before executing the next function, which
3541 * resets all in-service entities as no more in service. This
3542 * may cause bfqq to be freed. If this happens, the next
3543 * function returns true.
3545 return __bfq_bfqd_reset_in_service(bfqd);
3549 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3550 * @bfqd: device data.
3551 * @bfqq: queue to update.
3552 * @reason: reason for expiration.
3554 * Handle the feedback on @bfqq budget at queue expiration.
3555 * See the body for detailed comments.
3557 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
3558 struct bfq_queue *bfqq,
3559 enum bfqq_expiration reason)
3561 struct request *next_rq;
3562 int budget, min_budget;
3564 min_budget = bfq_min_budget(bfqd);
3566 if (bfqq->wr_coeff == 1)
3567 budget = bfqq->max_budget;
3569 * Use a constant, low budget for weight-raised queues,
3570 * to help achieve a low latency. Keep it slightly higher
3571 * than the minimum possible budget, to cause a little
3572 * bit fewer expirations.
3574 budget = 2 * min_budget;
3576 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
3577 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
3578 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
3579 budget, bfq_min_budget(bfqd));
3580 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
3581 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
3583 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
3586 * Caveat: in all the following cases we trade latency
3589 case BFQQE_TOO_IDLE:
3591 * This is the only case where we may reduce
3592 * the budget: if there is no request of the
3593 * process still waiting for completion, then
3594 * we assume (tentatively) that the timer has
3595 * expired because the batch of requests of
3596 * the process could have been served with a
3597 * smaller budget. Hence, betting that
3598 * process will behave in the same way when it
3599 * becomes backlogged again, we reduce its
3600 * next budget. As long as we guess right,
3601 * this budget cut reduces the latency
3602 * experienced by the process.
3604 * However, if there are still outstanding
3605 * requests, then the process may have not yet
3606 * issued its next request just because it is
3607 * still waiting for the completion of some of
3608 * the still outstanding ones. So in this
3609 * subcase we do not reduce its budget, on the
3610 * contrary we increase it to possibly boost
3611 * the throughput, as discussed in the
3612 * comments to the BUDGET_TIMEOUT case.
3614 if (bfqq->dispatched > 0) /* still outstanding reqs */
3615 budget = min(budget * 2, bfqd->bfq_max_budget);
3617 if (budget > 5 * min_budget)
3618 budget -= 4 * min_budget;
3620 budget = min_budget;
3623 case BFQQE_BUDGET_TIMEOUT:
3625 * We double the budget here because it gives
3626 * the chance to boost the throughput if this
3627 * is not a seeky process (and has bumped into
3628 * this timeout because of, e.g., ZBR).
3630 budget = min(budget * 2, bfqd->bfq_max_budget);
3632 case BFQQE_BUDGET_EXHAUSTED:
3634 * The process still has backlog, and did not
3635 * let either the budget timeout or the disk
3636 * idling timeout expire. Hence it is not
3637 * seeky, has a short thinktime and may be
3638 * happy with a higher budget too. So
3639 * definitely increase the budget of this good
3640 * candidate to boost the disk throughput.
3642 budget = min(budget * 4, bfqd->bfq_max_budget);
3644 case BFQQE_NO_MORE_REQUESTS:
3646 * For queues that expire for this reason, it
3647 * is particularly important to keep the
3648 * budget close to the actual service they
3649 * need. Doing so reduces the timestamp
3650 * misalignment problem described in the
3651 * comments in the body of
3652 * __bfq_activate_entity. In fact, suppose
3653 * that a queue systematically expires for
3654 * BFQQE_NO_MORE_REQUESTS and presents a
3655 * new request in time to enjoy timestamp
3656 * back-shifting. The larger the budget of the
3657 * queue is with respect to the service the
3658 * queue actually requests in each service
3659 * slot, the more times the queue can be
3660 * reactivated with the same virtual finish
3661 * time. It follows that, even if this finish
3662 * time is pushed to the system virtual time
3663 * to reduce the consequent timestamp
3664 * misalignment, the queue unjustly enjoys for
3665 * many re-activations a lower finish time
3666 * than all newly activated queues.
3668 * The service needed by bfqq is measured
3669 * quite precisely by bfqq->entity.service.
3670 * Since bfqq does not enjoy device idling,
3671 * bfqq->entity.service is equal to the number
3672 * of sectors that the process associated with
3673 * bfqq requested to read/write before waiting
3674 * for request completions, or blocking for
3677 budget = max_t(int, bfqq->entity.service, min_budget);
3682 } else if (!bfq_bfqq_sync(bfqq)) {
3684 * Async queues get always the maximum possible
3685 * budget, as for them we do not care about latency
3686 * (in addition, their ability to dispatch is limited
3687 * by the charging factor).
3689 budget = bfqd->bfq_max_budget;
3692 bfqq->max_budget = budget;
3694 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
3695 !bfqd->bfq_user_max_budget)
3696 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
3699 * If there is still backlog, then assign a new budget, making
3700 * sure that it is large enough for the next request. Since
3701 * the finish time of bfqq must be kept in sync with the
3702 * budget, be sure to call __bfq_bfqq_expire() *after* this
3705 * If there is no backlog, then no need to update the budget;
3706 * it will be updated on the arrival of a new request.
3708 next_rq = bfqq->next_rq;
3710 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
3711 bfq_serv_to_charge(next_rq, bfqq));
3713 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
3714 next_rq ? blk_rq_sectors(next_rq) : 0,
3715 bfqq->entity.budget);
3719 * Return true if the process associated with bfqq is "slow". The slow
3720 * flag is used, in addition to the budget timeout, to reduce the
3721 * amount of service provided to seeky processes, and thus reduce
3722 * their chances to lower the throughput. More details in the comments
3723 * on the function bfq_bfqq_expire().
3725 * An important observation is in order: as discussed in the comments
3726 * on the function bfq_update_peak_rate(), with devices with internal
3727 * queues, it is hard if ever possible to know when and for how long
3728 * an I/O request is processed by the device (apart from the trivial
3729 * I/O pattern where a new request is dispatched only after the
3730 * previous one has been completed). This makes it hard to evaluate
3731 * the real rate at which the I/O requests of each bfq_queue are
3732 * served. In fact, for an I/O scheduler like BFQ, serving a
3733 * bfq_queue means just dispatching its requests during its service
3734 * slot (i.e., until the budget of the queue is exhausted, or the
3735 * queue remains idle, or, finally, a timeout fires). But, during the
3736 * service slot of a bfq_queue, around 100 ms at most, the device may
3737 * be even still processing requests of bfq_queues served in previous
3738 * service slots. On the opposite end, the requests of the in-service
3739 * bfq_queue may be completed after the service slot of the queue
3742 * Anyway, unless more sophisticated solutions are used
3743 * (where possible), the sum of the sizes of the requests dispatched
3744 * during the service slot of a bfq_queue is probably the only
3745 * approximation available for the service received by the bfq_queue
3746 * during its service slot. And this sum is the quantity used in this
3747 * function to evaluate the I/O speed of a process.
3749 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3750 bool compensate, enum bfqq_expiration reason,
3751 unsigned long *delta_ms)
3753 ktime_t delta_ktime;
3755 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
3757 if (!bfq_bfqq_sync(bfqq))
3761 delta_ktime = bfqd->last_idling_start;
3763 delta_ktime = ktime_get();
3764 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
3765 delta_usecs = ktime_to_us(delta_ktime);
3767 /* don't use too short time intervals */
3768 if (delta_usecs < 1000) {
3769 if (blk_queue_nonrot(bfqd->queue))
3771 * give same worst-case guarantees as idling
3774 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
3775 else /* charge at least one seek */
3776 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
3781 *delta_ms = delta_usecs / USEC_PER_MSEC;
3784 * Use only long (> 20ms) intervals to filter out excessive
3785 * spikes in service rate estimation.
3787 if (delta_usecs > 20000) {
3789 * Caveat for rotational devices: processes doing I/O
3790 * in the slower disk zones tend to be slow(er) even
3791 * if not seeky. In this respect, the estimated peak
3792 * rate is likely to be an average over the disk
3793 * surface. Accordingly, to not be too harsh with
3794 * unlucky processes, a process is deemed slow only if
3795 * its rate has been lower than half of the estimated
3798 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
3801 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
3807 * To be deemed as soft real-time, an application must meet two
3808 * requirements. First, the application must not require an average
3809 * bandwidth higher than the approximate bandwidth required to playback or
3810 * record a compressed high-definition video.
3811 * The next function is invoked on the completion of the last request of a
3812 * batch, to compute the next-start time instant, soft_rt_next_start, such
3813 * that, if the next request of the application does not arrive before
3814 * soft_rt_next_start, then the above requirement on the bandwidth is met.
3816 * The second requirement is that the request pattern of the application is
3817 * isochronous, i.e., that, after issuing a request or a batch of requests,
3818 * the application stops issuing new requests until all its pending requests
3819 * have been completed. After that, the application may issue a new batch,
3821 * For this reason the next function is invoked to compute
3822 * soft_rt_next_start only for applications that meet this requirement,
3823 * whereas soft_rt_next_start is set to infinity for applications that do
3826 * Unfortunately, even a greedy (i.e., I/O-bound) application may
3827 * happen to meet, occasionally or systematically, both the above
3828 * bandwidth and isochrony requirements. This may happen at least in
3829 * the following circumstances. First, if the CPU load is high. The
3830 * application may stop issuing requests while the CPUs are busy
3831 * serving other processes, then restart, then stop again for a while,
3832 * and so on. The other circumstances are related to the storage
3833 * device: the storage device is highly loaded or reaches a low-enough
3834 * throughput with the I/O of the application (e.g., because the I/O
3835 * is random and/or the device is slow). In all these cases, the
3836 * I/O of the application may be simply slowed down enough to meet
3837 * the bandwidth and isochrony requirements. To reduce the probability
3838 * that greedy applications are deemed as soft real-time in these
3839 * corner cases, a further rule is used in the computation of
3840 * soft_rt_next_start: the return value of this function is forced to
3841 * be higher than the maximum between the following two quantities.
3843 * (a) Current time plus: (1) the maximum time for which the arrival
3844 * of a request is waited for when a sync queue becomes idle,
3845 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
3846 * postpone for a moment the reason for adding a few extra
3847 * jiffies; we get back to it after next item (b). Lower-bounding
3848 * the return value of this function with the current time plus
3849 * bfqd->bfq_slice_idle tends to filter out greedy applications,
3850 * because the latter issue their next request as soon as possible
3851 * after the last one has been completed. In contrast, a soft
3852 * real-time application spends some time processing data, after a
3853 * batch of its requests has been completed.
3855 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
3856 * above, greedy applications may happen to meet both the
3857 * bandwidth and isochrony requirements under heavy CPU or
3858 * storage-device load. In more detail, in these scenarios, these
3859 * applications happen, only for limited time periods, to do I/O
3860 * slowly enough to meet all the requirements described so far,
3861 * including the filtering in above item (a). These slow-speed
3862 * time intervals are usually interspersed between other time
3863 * intervals during which these applications do I/O at a very high
3864 * speed. Fortunately, exactly because of the high speed of the
3865 * I/O in the high-speed intervals, the values returned by this
3866 * function happen to be so high, near the end of any such
3867 * high-speed interval, to be likely to fall *after* the end of
3868 * the low-speed time interval that follows. These high values are
3869 * stored in bfqq->soft_rt_next_start after each invocation of
3870 * this function. As a consequence, if the last value of
3871 * bfqq->soft_rt_next_start is constantly used to lower-bound the
3872 * next value that this function may return, then, from the very
3873 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
3874 * likely to be constantly kept so high that any I/O request
3875 * issued during the low-speed interval is considered as arriving
3876 * to soon for the application to be deemed as soft
3877 * real-time. Then, in the high-speed interval that follows, the
3878 * application will not be deemed as soft real-time, just because
3879 * it will do I/O at a high speed. And so on.
3881 * Getting back to the filtering in item (a), in the following two
3882 * cases this filtering might be easily passed by a greedy
3883 * application, if the reference quantity was just
3884 * bfqd->bfq_slice_idle:
3885 * 1) HZ is so low that the duration of a jiffy is comparable to or
3886 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
3887 * devices with HZ=100. The time granularity may be so coarse
3888 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
3889 * is rather lower than the exact value.
3890 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
3891 * for a while, then suddenly 'jump' by several units to recover the lost
3892 * increments. This seems to happen, e.g., inside virtual machines.
3893 * To address this issue, in the filtering in (a) we do not use as a
3894 * reference time interval just bfqd->bfq_slice_idle, but
3895 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
3896 * minimum number of jiffies for which the filter seems to be quite
3897 * precise also in embedded systems and KVM/QEMU virtual machines.
3899 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
3900 struct bfq_queue *bfqq)
3902 return max3(bfqq->soft_rt_next_start,
3903 bfqq->last_idle_bklogged +
3904 HZ * bfqq->service_from_backlogged /
3905 bfqd->bfq_wr_max_softrt_rate,
3906 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
3910 * bfq_bfqq_expire - expire a queue.
3911 * @bfqd: device owning the queue.
3912 * @bfqq: the queue to expire.
3913 * @compensate: if true, compensate for the time spent idling.
3914 * @reason: the reason causing the expiration.
3916 * If the process associated with bfqq does slow I/O (e.g., because it
3917 * issues random requests), we charge bfqq with the time it has been
3918 * in service instead of the service it has received (see
3919 * bfq_bfqq_charge_time for details on how this goal is achieved). As
3920 * a consequence, bfqq will typically get higher timestamps upon
3921 * reactivation, and hence it will be rescheduled as if it had
3922 * received more service than what it has actually received. In the
3923 * end, bfqq receives less service in proportion to how slowly its
3924 * associated process consumes its budgets (and hence how seriously it
3925 * tends to lower the throughput). In addition, this time-charging
3926 * strategy guarantees time fairness among slow processes. In
3927 * contrast, if the process associated with bfqq is not slow, we
3928 * charge bfqq exactly with the service it has received.
3930 * Charging time to the first type of queues and the exact service to
3931 * the other has the effect of using the WF2Q+ policy to schedule the
3932 * former on a timeslice basis, without violating service domain
3933 * guarantees among the latter.
3935 void bfq_bfqq_expire(struct bfq_data *bfqd,
3936 struct bfq_queue *bfqq,
3938 enum bfqq_expiration reason)
3941 unsigned long delta = 0;
3942 struct bfq_entity *entity = &bfqq->entity;
3945 * Check whether the process is slow (see bfq_bfqq_is_slow).
3947 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
3950 * As above explained, charge slow (typically seeky) and
3951 * timed-out queues with the time and not the service
3952 * received, to favor sequential workloads.
3954 * Processes doing I/O in the slower disk zones will tend to
3955 * be slow(er) even if not seeky. Therefore, since the
3956 * estimated peak rate is actually an average over the disk
3957 * surface, these processes may timeout just for bad luck. To
3958 * avoid punishing them, do not charge time to processes that
3959 * succeeded in consuming at least 2/3 of their budget. This
3960 * allows BFQ to preserve enough elasticity to still perform
3961 * bandwidth, and not time, distribution with little unlucky
3962 * or quasi-sequential processes.
3964 if (bfqq->wr_coeff == 1 &&
3966 (reason == BFQQE_BUDGET_TIMEOUT &&
3967 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
3968 bfq_bfqq_charge_time(bfqd, bfqq, delta);
3970 if (reason == BFQQE_TOO_IDLE &&
3971 entity->service <= 2 * entity->budget / 10)
3972 bfq_clear_bfqq_IO_bound(bfqq);
3974 if (bfqd->low_latency && bfqq->wr_coeff == 1)
3975 bfqq->last_wr_start_finish = jiffies;
3977 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
3978 RB_EMPTY_ROOT(&bfqq->sort_list)) {
3980 * If we get here, and there are no outstanding
3981 * requests, then the request pattern is isochronous
3982 * (see the comments on the function
3983 * bfq_bfqq_softrt_next_start()). Thus we can compute
3984 * soft_rt_next_start. And we do it, unless bfqq is in
3985 * interactive weight raising. We do not do it in the
3986 * latter subcase, for the following reason. bfqq may
3987 * be conveying the I/O needed to load a soft
3988 * real-time application. Such an application will
3989 * actually exhibit a soft real-time I/O pattern after
3990 * it finally starts doing its job. But, if
3991 * soft_rt_next_start is computed here for an
3992 * interactive bfqq, and bfqq had received a lot of
3993 * service before remaining with no outstanding
3994 * request (likely to happen on a fast device), then
3995 * soft_rt_next_start would be assigned such a high
3996 * value that, for a very long time, bfqq would be
3997 * prevented from being possibly considered as soft
4000 * If, instead, the queue still has outstanding
4001 * requests, then we have to wait for the completion
4002 * of all the outstanding requests to discover whether
4003 * the request pattern is actually isochronous.
4005 if (bfqq->dispatched == 0 &&
4006 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
4007 bfqq->soft_rt_next_start =
4008 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4009 else if (bfqq->dispatched > 0) {
4011 * Schedule an update of soft_rt_next_start to when
4012 * the task may be discovered to be isochronous.
4014 bfq_mark_bfqq_softrt_update(bfqq);
4018 bfq_log_bfqq(bfqd, bfqq,
4019 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
4020 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
4023 * bfqq expired, so no total service time needs to be computed
4024 * any longer: reset state machine for measuring total service
4027 bfqd->rqs_injected = bfqd->wait_dispatch = false;
4028 bfqd->waited_rq = NULL;
4031 * Increase, decrease or leave budget unchanged according to
4034 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
4035 if (__bfq_bfqq_expire(bfqd, bfqq, reason))
4036 /* bfqq is gone, no more actions on it */
4039 /* mark bfqq as waiting a request only if a bic still points to it */
4040 if (!bfq_bfqq_busy(bfqq) &&
4041 reason != BFQQE_BUDGET_TIMEOUT &&
4042 reason != BFQQE_BUDGET_EXHAUSTED) {
4043 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
4045 * Not setting service to 0, because, if the next rq
4046 * arrives in time, the queue will go on receiving
4047 * service with this same budget (as if it never expired)
4050 entity->service = 0;
4053 * Reset the received-service counter for every parent entity.
4054 * Differently from what happens with bfqq->entity.service,
4055 * the resetting of this counter never needs to be postponed
4056 * for parent entities. In fact, in case bfqq may have a
4057 * chance to go on being served using the last, partially
4058 * consumed budget, bfqq->entity.service needs to be kept,
4059 * because if bfqq then actually goes on being served using
4060 * the same budget, the last value of bfqq->entity.service is
4061 * needed to properly decrement bfqq->entity.budget by the
4062 * portion already consumed. In contrast, it is not necessary
4063 * to keep entity->service for parent entities too, because
4064 * the bubble up of the new value of bfqq->entity.budget will
4065 * make sure that the budgets of parent entities are correct,
4066 * even in case bfqq and thus parent entities go on receiving
4067 * service with the same budget.
4069 entity = entity->parent;
4070 for_each_entity(entity)
4071 entity->service = 0;
4075 * Budget timeout is not implemented through a dedicated timer, but
4076 * just checked on request arrivals and completions, as well as on
4077 * idle timer expirations.
4079 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
4081 return time_is_before_eq_jiffies(bfqq->budget_timeout);
4085 * If we expire a queue that is actively waiting (i.e., with the
4086 * device idled) for the arrival of a new request, then we may incur
4087 * the timestamp misalignment problem described in the body of the
4088 * function __bfq_activate_entity. Hence we return true only if this
4089 * condition does not hold, or if the queue is slow enough to deserve
4090 * only to be kicked off for preserving a high throughput.
4092 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
4094 bfq_log_bfqq(bfqq->bfqd, bfqq,
4095 "may_budget_timeout: wait_request %d left %d timeout %d",
4096 bfq_bfqq_wait_request(bfqq),
4097 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
4098 bfq_bfqq_budget_timeout(bfqq));
4100 return (!bfq_bfqq_wait_request(bfqq) ||
4101 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
4103 bfq_bfqq_budget_timeout(bfqq);
4106 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
4107 struct bfq_queue *bfqq)
4109 bool rot_without_queueing =
4110 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
4111 bfqq_sequential_and_IO_bound,
4114 /* No point in idling for bfqq if it won't get requests any longer */
4115 if (unlikely(!bfqq_process_refs(bfqq)))
4118 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
4119 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
4122 * The next variable takes into account the cases where idling
4123 * boosts the throughput.
4125 * The value of the variable is computed considering, first, that
4126 * idling is virtually always beneficial for the throughput if:
4127 * (a) the device is not NCQ-capable and rotational, or
4128 * (b) regardless of the presence of NCQ, the device is rotational and
4129 * the request pattern for bfqq is I/O-bound and sequential, or
4130 * (c) regardless of whether it is rotational, the device is
4131 * not NCQ-capable and the request pattern for bfqq is
4132 * I/O-bound and sequential.
4134 * Secondly, and in contrast to the above item (b), idling an
4135 * NCQ-capable flash-based device would not boost the
4136 * throughput even with sequential I/O; rather it would lower
4137 * the throughput in proportion to how fast the device
4138 * is. Accordingly, the next variable is true if any of the
4139 * above conditions (a), (b) or (c) is true, and, in
4140 * particular, happens to be false if bfqd is an NCQ-capable
4141 * flash-based device.
4143 idling_boosts_thr = rot_without_queueing ||
4144 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
4145 bfqq_sequential_and_IO_bound);
4148 * The return value of this function is equal to that of
4149 * idling_boosts_thr, unless a special case holds. In this
4150 * special case, described below, idling may cause problems to
4151 * weight-raised queues.
4153 * When the request pool is saturated (e.g., in the presence
4154 * of write hogs), if the processes associated with
4155 * non-weight-raised queues ask for requests at a lower rate,
4156 * then processes associated with weight-raised queues have a
4157 * higher probability to get a request from the pool
4158 * immediately (or at least soon) when they need one. Thus
4159 * they have a higher probability to actually get a fraction
4160 * of the device throughput proportional to their high
4161 * weight. This is especially true with NCQ-capable drives,
4162 * which enqueue several requests in advance, and further
4163 * reorder internally-queued requests.
4165 * For this reason, we force to false the return value if
4166 * there are weight-raised busy queues. In this case, and if
4167 * bfqq is not weight-raised, this guarantees that the device
4168 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4169 * then idling will be guaranteed by another variable, see
4170 * below). Combined with the timestamping rules of BFQ (see
4171 * [1] for details), this behavior causes bfqq, and hence any
4172 * sync non-weight-raised queue, to get a lower number of
4173 * requests served, and thus to ask for a lower number of
4174 * requests from the request pool, before the busy
4175 * weight-raised queues get served again. This often mitigates
4176 * starvation problems in the presence of heavy write
4177 * workloads and NCQ, thereby guaranteeing a higher
4178 * application and system responsiveness in these hostile
4181 return idling_boosts_thr &&
4182 bfqd->wr_busy_queues == 0;
4186 * For a queue that becomes empty, device idling is allowed only if
4187 * this function returns true for that queue. As a consequence, since
4188 * device idling plays a critical role for both throughput boosting
4189 * and service guarantees, the return value of this function plays a
4190 * critical role as well.
4192 * In a nutshell, this function returns true only if idling is
4193 * beneficial for throughput or, even if detrimental for throughput,
4194 * idling is however necessary to preserve service guarantees (low
4195 * latency, desired throughput distribution, ...). In particular, on
4196 * NCQ-capable devices, this function tries to return false, so as to
4197 * help keep the drives' internal queues full, whenever this helps the
4198 * device boost the throughput without causing any service-guarantee
4201 * Most of the issues taken into account to get the return value of
4202 * this function are not trivial. We discuss these issues in the two
4203 * functions providing the main pieces of information needed by this
4206 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
4208 struct bfq_data *bfqd = bfqq->bfqd;
4209 bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
4211 /* No point in idling for bfqq if it won't get requests any longer */
4212 if (unlikely(!bfqq_process_refs(bfqq)))
4215 if (unlikely(bfqd->strict_guarantees))
4219 * Idling is performed only if slice_idle > 0. In addition, we
4222 * (b) bfqq is in the idle io prio class: in this case we do
4223 * not idle because we want to minimize the bandwidth that
4224 * queues in this class can steal to higher-priority queues
4226 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
4227 bfq_class_idle(bfqq))
4230 idling_boosts_thr_with_no_issue =
4231 idling_boosts_thr_without_issues(bfqd, bfqq);
4233 idling_needed_for_service_guar =
4234 idling_needed_for_service_guarantees(bfqd, bfqq);
4237 * We have now the two components we need to compute the
4238 * return value of the function, which is true only if idling
4239 * either boosts the throughput (without issues), or is
4240 * necessary to preserve service guarantees.
4242 return idling_boosts_thr_with_no_issue ||
4243 idling_needed_for_service_guar;
4247 * If the in-service queue is empty but the function bfq_better_to_idle
4248 * returns true, then:
4249 * 1) the queue must remain in service and cannot be expired, and
4250 * 2) the device must be idled to wait for the possible arrival of a new
4251 * request for the queue.
4252 * See the comments on the function bfq_better_to_idle for the reasons
4253 * why performing device idling is the best choice to boost the throughput
4254 * and preserve service guarantees when bfq_better_to_idle itself
4257 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
4259 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
4263 * This function chooses the queue from which to pick the next extra
4264 * I/O request to inject, if it finds a compatible queue. See the
4265 * comments on bfq_update_inject_limit() for details on the injection
4266 * mechanism, and for the definitions of the quantities mentioned
4269 static struct bfq_queue *
4270 bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
4272 struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
4273 unsigned int limit = in_serv_bfqq->inject_limit;
4276 * - bfqq is not weight-raised and therefore does not carry
4277 * time-critical I/O,
4279 * - regardless of whether bfqq is weight-raised, bfqq has
4280 * however a long think time, during which it can absorb the
4281 * effect of an appropriate number of extra I/O requests
4282 * from other queues (see bfq_update_inject_limit for
4283 * details on the computation of this number);
4284 * then injection can be performed without restrictions.
4286 bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
4287 !bfq_bfqq_has_short_ttime(in_serv_bfqq);
4291 * - the baseline total service time could not be sampled yet,
4292 * so the inject limit happens to be still 0, and
4293 * - a lot of time has elapsed since the plugging of I/O
4294 * dispatching started, so drive speed is being wasted
4296 * then temporarily raise inject limit to one request.
4298 if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
4299 bfq_bfqq_wait_request(in_serv_bfqq) &&
4300 time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
4301 bfqd->bfq_slice_idle)
4305 if (bfqd->rq_in_driver >= limit)
4309 * Linear search of the source queue for injection; but, with
4310 * a high probability, very few steps are needed to find a
4311 * candidate queue, i.e., a queue with enough budget left for
4312 * its next request. In fact:
4313 * - BFQ dynamically updates the budget of every queue so as
4314 * to accommodate the expected backlog of the queue;
4315 * - if a queue gets all its requests dispatched as injected
4316 * service, then the queue is removed from the active list
4317 * (and re-added only if it gets new requests, but then it
4318 * is assigned again enough budget for its new backlog).
4320 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
4321 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4322 (in_serv_always_inject || bfqq->wr_coeff > 1) &&
4323 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4324 bfq_bfqq_budget_left(bfqq)) {
4326 * Allow for only one large in-flight request
4327 * on non-rotational devices, for the
4328 * following reason. On non-rotationl drives,
4329 * large requests take much longer than
4330 * smaller requests to be served. In addition,
4331 * the drive prefers to serve large requests
4332 * w.r.t. to small ones, if it can choose. So,
4333 * having more than one large requests queued
4334 * in the drive may easily make the next first
4335 * request of the in-service queue wait for so
4336 * long to break bfqq's service guarantees. On
4337 * the bright side, large requests let the
4338 * drive reach a very high throughput, even if
4339 * there is only one in-flight large request
4342 if (blk_queue_nonrot(bfqd->queue) &&
4343 blk_rq_sectors(bfqq->next_rq) >=
4344 BFQQ_SECT_THR_NONROT)
4345 limit = min_t(unsigned int, 1, limit);
4347 limit = in_serv_bfqq->inject_limit;
4349 if (bfqd->rq_in_driver < limit) {
4350 bfqd->rqs_injected = true;
4359 * Select a queue for service. If we have a current queue in service,
4360 * check whether to continue servicing it, or retrieve and set a new one.
4362 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
4364 struct bfq_queue *bfqq;
4365 struct request *next_rq;
4366 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
4368 bfqq = bfqd->in_service_queue;
4372 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
4375 * Do not expire bfqq for budget timeout if bfqq may be about
4376 * to enjoy device idling. The reason why, in this case, we
4377 * prevent bfqq from expiring is the same as in the comments
4378 * on the case where bfq_bfqq_must_idle() returns true, in
4379 * bfq_completed_request().
4381 if (bfq_may_expire_for_budg_timeout(bfqq) &&
4382 !bfq_bfqq_must_idle(bfqq))
4387 * This loop is rarely executed more than once. Even when it
4388 * happens, it is much more convenient to re-execute this loop
4389 * than to return NULL and trigger a new dispatch to get a
4392 next_rq = bfqq->next_rq;
4394 * If bfqq has requests queued and it has enough budget left to
4395 * serve them, keep the queue, otherwise expire it.
4398 if (bfq_serv_to_charge(next_rq, bfqq) >
4399 bfq_bfqq_budget_left(bfqq)) {
4401 * Expire the queue for budget exhaustion,
4402 * which makes sure that the next budget is
4403 * enough to serve the next request, even if
4404 * it comes from the fifo expired path.
4406 reason = BFQQE_BUDGET_EXHAUSTED;
4410 * The idle timer may be pending because we may
4411 * not disable disk idling even when a new request
4414 if (bfq_bfqq_wait_request(bfqq)) {
4416 * If we get here: 1) at least a new request
4417 * has arrived but we have not disabled the
4418 * timer because the request was too small,
4419 * 2) then the block layer has unplugged
4420 * the device, causing the dispatch to be
4423 * Since the device is unplugged, now the
4424 * requests are probably large enough to
4425 * provide a reasonable throughput.
4426 * So we disable idling.
4428 bfq_clear_bfqq_wait_request(bfqq);
4429 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4436 * No requests pending. However, if the in-service queue is idling
4437 * for a new request, or has requests waiting for a completion and
4438 * may idle after their completion, then keep it anyway.
4440 * Yet, inject service from other queues if it boosts
4441 * throughput and is possible.
4443 if (bfq_bfqq_wait_request(bfqq) ||
4444 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
4445 struct bfq_queue *async_bfqq =
4446 bfqq->bic && bfqq->bic->bfqq[0] &&
4447 bfq_bfqq_busy(bfqq->bic->bfqq[0]) &&
4448 bfqq->bic->bfqq[0]->next_rq ?
4449 bfqq->bic->bfqq[0] : NULL;
4452 * The next three mutually-exclusive ifs decide
4453 * whether to try injection, and choose the queue to
4454 * pick an I/O request from.
4456 * The first if checks whether the process associated
4457 * with bfqq has also async I/O pending. If so, it
4458 * injects such I/O unconditionally. Injecting async
4459 * I/O from the same process can cause no harm to the
4460 * process. On the contrary, it can only increase
4461 * bandwidth and reduce latency for the process.
4463 * The second if checks whether there happens to be a
4464 * non-empty waker queue for bfqq, i.e., a queue whose
4465 * I/O needs to be completed for bfqq to receive new
4466 * I/O. This happens, e.g., if bfqq is associated with
4467 * a process that does some sync. A sync generates
4468 * extra blocking I/O, which must be completed before
4469 * the process associated with bfqq can go on with its
4470 * I/O. If the I/O of the waker queue is not served,
4471 * then bfqq remains empty, and no I/O is dispatched,
4472 * until the idle timeout fires for bfqq. This is
4473 * likely to result in lower bandwidth and higher
4474 * latencies for bfqq, and in a severe loss of total
4475 * throughput. The best action to take is therefore to
4476 * serve the waker queue as soon as possible. So do it
4477 * (without relying on the third alternative below for
4478 * eventually serving waker_bfqq's I/O; see the last
4479 * paragraph for further details). This systematic
4480 * injection of I/O from the waker queue does not
4481 * cause any delay to bfqq's I/O. On the contrary,
4482 * next bfqq's I/O is brought forward dramatically,
4483 * for it is not blocked for milliseconds.
4485 * The third if checks whether bfqq is a queue for
4486 * which it is better to avoid injection. It is so if
4487 * bfqq delivers more throughput when served without
4488 * any further I/O from other queues in the middle, or
4489 * if the service times of bfqq's I/O requests both
4490 * count more than overall throughput, and may be
4491 * easily increased by injection (this happens if bfqq
4492 * has a short think time). If none of these
4493 * conditions holds, then a candidate queue for
4494 * injection is looked for through
4495 * bfq_choose_bfqq_for_injection(). Note that the
4496 * latter may return NULL (for example if the inject
4497 * limit for bfqq is currently 0).
4499 * NOTE: motivation for the second alternative
4501 * Thanks to the way the inject limit is updated in
4502 * bfq_update_has_short_ttime(), it is rather likely
4503 * that, if I/O is being plugged for bfqq and the
4504 * waker queue has pending I/O requests that are
4505 * blocking bfqq's I/O, then the third alternative
4506 * above lets the waker queue get served before the
4507 * I/O-plugging timeout fires. So one may deem the
4508 * second alternative superfluous. It is not, because
4509 * the third alternative may be way less effective in
4510 * case of a synchronization. For two main
4511 * reasons. First, throughput may be low because the
4512 * inject limit may be too low to guarantee the same
4513 * amount of injected I/O, from the waker queue or
4514 * other queues, that the second alternative
4515 * guarantees (the second alternative unconditionally
4516 * injects a pending I/O request of the waker queue
4517 * for each bfq_dispatch_request()). Second, with the
4518 * third alternative, the duration of the plugging,
4519 * i.e., the time before bfqq finally receives new I/O,
4520 * may not be minimized, because the waker queue may
4521 * happen to be served only after other queues.
4524 icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
4525 bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
4526 bfq_bfqq_budget_left(async_bfqq))
4527 bfqq = bfqq->bic->bfqq[0];
4528 else if (bfq_bfqq_has_waker(bfqq) &&
4529 bfq_bfqq_busy(bfqq->waker_bfqq) &&
4531 bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
4532 bfqq->waker_bfqq) <=
4533 bfq_bfqq_budget_left(bfqq->waker_bfqq)
4535 bfqq = bfqq->waker_bfqq;
4536 else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
4537 (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
4538 !bfq_bfqq_has_short_ttime(bfqq)))
4539 bfqq = bfq_choose_bfqq_for_injection(bfqd);
4546 reason = BFQQE_NO_MORE_REQUESTS;
4548 bfq_bfqq_expire(bfqd, bfqq, false, reason);
4550 bfqq = bfq_set_in_service_queue(bfqd);
4552 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
4557 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
4559 bfq_log(bfqd, "select_queue: no queue returned");
4564 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4566 struct bfq_entity *entity = &bfqq->entity;
4568 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
4569 bfq_log_bfqq(bfqd, bfqq,
4570 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
4571 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
4572 jiffies_to_msecs(bfqq->wr_cur_max_time),
4574 bfqq->entity.weight, bfqq->entity.orig_weight);
4576 if (entity->prio_changed)
4577 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
4580 * If the queue was activated in a burst, or too much
4581 * time has elapsed from the beginning of this
4582 * weight-raising period, then end weight raising.
4584 if (bfq_bfqq_in_large_burst(bfqq))
4585 bfq_bfqq_end_wr(bfqq);
4586 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
4587 bfqq->wr_cur_max_time)) {
4588 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
4589 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
4590 bfq_wr_duration(bfqd)))
4591 bfq_bfqq_end_wr(bfqq);
4593 switch_back_to_interactive_wr(bfqq, bfqd);
4594 bfqq->entity.prio_changed = 1;
4597 if (bfqq->wr_coeff > 1 &&
4598 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
4599 bfqq->service_from_wr > max_service_from_wr) {
4600 /* see comments on max_service_from_wr */
4601 bfq_bfqq_end_wr(bfqq);
4605 * To improve latency (for this or other queues), immediately
4606 * update weight both if it must be raised and if it must be
4607 * lowered. Since, entity may be on some active tree here, and
4608 * might have a pending change of its ioprio class, invoke
4609 * next function with the last parameter unset (see the
4610 * comments on the function).
4612 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
4613 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
4618 * Dispatch next request from bfqq.
4620 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
4621 struct bfq_queue *bfqq)
4623 struct request *rq = bfqq->next_rq;
4624 unsigned long service_to_charge;
4626 service_to_charge = bfq_serv_to_charge(rq, bfqq);
4628 bfq_bfqq_served(bfqq, service_to_charge);
4630 if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
4631 bfqd->wait_dispatch = false;
4632 bfqd->waited_rq = rq;
4635 bfq_dispatch_remove(bfqd->queue, rq);
4637 if (bfqq != bfqd->in_service_queue)
4641 * If weight raising has to terminate for bfqq, then next
4642 * function causes an immediate update of bfqq's weight,
4643 * without waiting for next activation. As a consequence, on
4644 * expiration, bfqq will be timestamped as if has never been
4645 * weight-raised during this service slot, even if it has
4646 * received part or even most of the service as a
4647 * weight-raised queue. This inflates bfqq's timestamps, which
4648 * is beneficial, as bfqq is then more willing to leave the
4649 * device immediately to possible other weight-raised queues.
4651 bfq_update_wr_data(bfqd, bfqq);
4654 * Expire bfqq, pretending that its budget expired, if bfqq
4655 * belongs to CLASS_IDLE and other queues are waiting for
4658 if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
4661 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
4667 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
4669 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4672 * Avoiding lock: a race on bfqd->busy_queues should cause at
4673 * most a call to dispatch for nothing
4675 return !list_empty_careful(&bfqd->dispatch) ||
4676 bfq_tot_busy_queues(bfqd) > 0;
4679 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
4681 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4682 struct request *rq = NULL;
4683 struct bfq_queue *bfqq = NULL;
4685 if (!list_empty(&bfqd->dispatch)) {
4686 rq = list_first_entry(&bfqd->dispatch, struct request,
4688 list_del_init(&rq->queuelist);
4694 * Increment counters here, because this
4695 * dispatch does not follow the standard
4696 * dispatch flow (where counters are
4701 goto inc_in_driver_start_rq;
4705 * We exploit the bfq_finish_requeue_request hook to
4706 * decrement rq_in_driver, but
4707 * bfq_finish_requeue_request will not be invoked on
4708 * this request. So, to avoid unbalance, just start
4709 * this request, without incrementing rq_in_driver. As
4710 * a negative consequence, rq_in_driver is deceptively
4711 * lower than it should be while this request is in
4712 * service. This may cause bfq_schedule_dispatch to be
4713 * invoked uselessly.
4715 * As for implementing an exact solution, the
4716 * bfq_finish_requeue_request hook, if defined, is
4717 * probably invoked also on this request. So, by
4718 * exploiting this hook, we could 1) increment
4719 * rq_in_driver here, and 2) decrement it in
4720 * bfq_finish_requeue_request. Such a solution would
4721 * let the value of the counter be always accurate,
4722 * but it would entail using an extra interface
4723 * function. This cost seems higher than the benefit,
4724 * being the frequency of non-elevator-private
4725 * requests very low.
4730 bfq_log(bfqd, "dispatch requests: %d busy queues",
4731 bfq_tot_busy_queues(bfqd));
4733 if (bfq_tot_busy_queues(bfqd) == 0)
4737 * Force device to serve one request at a time if
4738 * strict_guarantees is true. Forcing this service scheme is
4739 * currently the ONLY way to guarantee that the request
4740 * service order enforced by the scheduler is respected by a
4741 * queueing device. Otherwise the device is free even to make
4742 * some unlucky request wait for as long as the device
4745 * Of course, serving one request at at time may cause loss of
4748 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
4751 bfqq = bfq_select_queue(bfqd);
4755 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
4758 inc_in_driver_start_rq:
4759 bfqd->rq_in_driver++;
4761 rq->rq_flags |= RQF_STARTED;
4767 #ifdef CONFIG_BFQ_CGROUP_DEBUG
4768 static void bfq_update_dispatch_stats(struct request_queue *q,
4770 struct bfq_queue *in_serv_queue,
4771 bool idle_timer_disabled)
4773 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
4775 if (!idle_timer_disabled && !bfqq)
4779 * rq and bfqq are guaranteed to exist until this function
4780 * ends, for the following reasons. First, rq can be
4781 * dispatched to the device, and then can be completed and
4782 * freed, only after this function ends. Second, rq cannot be
4783 * merged (and thus freed because of a merge) any longer,
4784 * because it has already started. Thus rq cannot be freed
4785 * before this function ends, and, since rq has a reference to
4786 * bfqq, the same guarantee holds for bfqq too.
4788 * In addition, the following queue lock guarantees that
4789 * bfqq_group(bfqq) exists as well.
4791 spin_lock_irq(&q->queue_lock);
4792 if (idle_timer_disabled)
4794 * Since the idle timer has been disabled,
4795 * in_serv_queue contained some request when
4796 * __bfq_dispatch_request was invoked above, which
4797 * implies that rq was picked exactly from
4798 * in_serv_queue. Thus in_serv_queue == bfqq, and is
4799 * therefore guaranteed to exist because of the above
4802 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
4804 struct bfq_group *bfqg = bfqq_group(bfqq);
4806 bfqg_stats_update_avg_queue_size(bfqg);
4807 bfqg_stats_set_start_empty_time(bfqg);
4808 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
4810 spin_unlock_irq(&q->queue_lock);
4813 static inline void bfq_update_dispatch_stats(struct request_queue *q,
4815 struct bfq_queue *in_serv_queue,
4816 bool idle_timer_disabled) {}
4817 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
4819 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
4821 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4823 struct bfq_queue *in_serv_queue;
4824 bool waiting_rq, idle_timer_disabled = false;
4826 spin_lock_irq(&bfqd->lock);
4828 in_serv_queue = bfqd->in_service_queue;
4829 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
4831 rq = __bfq_dispatch_request(hctx);
4832 if (in_serv_queue == bfqd->in_service_queue) {
4833 idle_timer_disabled =
4834 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
4837 spin_unlock_irq(&bfqd->lock);
4838 bfq_update_dispatch_stats(hctx->queue, rq,
4839 idle_timer_disabled ? in_serv_queue : NULL,
4840 idle_timer_disabled);
4846 * Task holds one reference to the queue, dropped when task exits. Each rq
4847 * in-flight on this queue also holds a reference, dropped when rq is freed.
4849 * Scheduler lock must be held here. Recall not to use bfqq after calling
4850 * this function on it.
4852 void bfq_put_queue(struct bfq_queue *bfqq)
4854 struct bfq_queue *item;
4855 struct hlist_node *n;
4856 struct bfq_group *bfqg = bfqq_group(bfqq);
4859 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
4866 if (!hlist_unhashed(&bfqq->burst_list_node)) {
4867 hlist_del_init(&bfqq->burst_list_node);
4869 * Decrement also burst size after the removal, if the
4870 * process associated with bfqq is exiting, and thus
4871 * does not contribute to the burst any longer. This
4872 * decrement helps filter out false positives of large
4873 * bursts, when some short-lived process (often due to
4874 * the execution of commands by some service) happens
4875 * to start and exit while a complex application is
4876 * starting, and thus spawning several processes that
4877 * do I/O (and that *must not* be treated as a large
4878 * burst, see comments on bfq_handle_burst).
4880 * In particular, the decrement is performed only if:
4881 * 1) bfqq is not a merged queue, because, if it is,
4882 * then this free of bfqq is not triggered by the exit
4883 * of the process bfqq is associated with, but exactly
4884 * by the fact that bfqq has just been merged.
4885 * 2) burst_size is greater than 0, to handle
4886 * unbalanced decrements. Unbalanced decrements may
4887 * happen in te following case: bfqq is inserted into
4888 * the current burst list--without incrementing
4889 * bust_size--because of a split, but the current
4890 * burst list is not the burst list bfqq belonged to
4891 * (see comments on the case of a split in
4894 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
4895 bfqq->bfqd->burst_size--;
4899 * bfqq does not exist any longer, so it cannot be woken by
4900 * any other queue, and cannot wake any other queue. Then bfqq
4901 * must be removed from the woken list of its possible waker
4902 * queue, and all queues in the woken list of bfqq must stop
4903 * having a waker queue. Strictly speaking, these updates
4904 * should be performed when bfqq remains with no I/O source
4905 * attached to it, which happens before bfqq gets freed. In
4906 * particular, this happens when the last process associated
4907 * with bfqq exits or gets associated with a different
4908 * queue. However, both events lead to bfqq being freed soon,
4909 * and dangling references would come out only after bfqq gets
4910 * freed. So these updates are done here, as a simple and safe
4911 * way to handle all cases.
4913 /* remove bfqq from woken list */
4914 if (!hlist_unhashed(&bfqq->woken_list_node))
4915 hlist_del_init(&bfqq->woken_list_node);
4917 /* reset waker for all queues in woken list */
4918 hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
4920 item->waker_bfqq = NULL;
4921 bfq_clear_bfqq_has_waker(item);
4922 hlist_del_init(&item->woken_list_node);
4925 if (bfqq->bfqd && bfqq->bfqd->last_completed_rq_bfqq == bfqq)
4926 bfqq->bfqd->last_completed_rq_bfqq = NULL;
4928 kmem_cache_free(bfq_pool, bfqq);
4929 bfqg_and_blkg_put(bfqg);
4932 void bfq_put_cooperator(struct bfq_queue *bfqq)
4934 struct bfq_queue *__bfqq, *next;
4937 * If this queue was scheduled to merge with another queue, be
4938 * sure to drop the reference taken on that queue (and others in
4939 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
4941 __bfqq = bfqq->new_bfqq;
4945 next = __bfqq->new_bfqq;
4946 bfq_put_queue(__bfqq);
4951 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4953 if (bfqq == bfqd->in_service_queue) {
4954 __bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT);
4955 bfq_schedule_dispatch(bfqd);
4958 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
4960 bfq_put_cooperator(bfqq);
4962 bfq_release_process_ref(bfqd, bfqq);
4965 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
4967 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4968 struct bfq_data *bfqd;
4971 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
4974 unsigned long flags;
4976 spin_lock_irqsave(&bfqd->lock, flags);
4978 bfq_exit_bfqq(bfqd, bfqq);
4979 bic_set_bfqq(bic, NULL, is_sync);
4980 spin_unlock_irqrestore(&bfqd->lock, flags);
4984 static void bfq_exit_icq(struct io_cq *icq)
4986 struct bfq_io_cq *bic = icq_to_bic(icq);
4988 bfq_exit_icq_bfqq(bic, true);
4989 bfq_exit_icq_bfqq(bic, false);
4993 * Update the entity prio values; note that the new values will not
4994 * be used until the next (re)activation.
4997 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
4999 struct task_struct *tsk = current;
5001 struct bfq_data *bfqd = bfqq->bfqd;
5006 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5007 switch (ioprio_class) {
5009 dev_err(bfqq->bfqd->queue->backing_dev_info->dev,
5010 "bfq: bad prio class %d\n", ioprio_class);
5012 case IOPRIO_CLASS_NONE:
5014 * No prio set, inherit CPU scheduling settings.
5016 bfqq->new_ioprio = task_nice_ioprio(tsk);
5017 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
5019 case IOPRIO_CLASS_RT:
5020 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5021 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
5023 case IOPRIO_CLASS_BE:
5024 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5025 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
5027 case IOPRIO_CLASS_IDLE:
5028 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
5029 bfqq->new_ioprio = 7;
5033 if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
5034 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
5036 bfqq->new_ioprio = IOPRIO_BE_NR - 1;
5039 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
5040 bfqq->entity.prio_changed = 1;
5043 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5044 struct bio *bio, bool is_sync,
5045 struct bfq_io_cq *bic);
5047 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
5049 struct bfq_data *bfqd = bic_to_bfqd(bic);
5050 struct bfq_queue *bfqq;
5051 int ioprio = bic->icq.ioc->ioprio;
5054 * This condition may trigger on a newly created bic, be sure to
5055 * drop the lock before returning.
5057 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
5060 bic->ioprio = ioprio;
5062 bfqq = bic_to_bfqq(bic, false);
5064 bfq_release_process_ref(bfqd, bfqq);
5065 bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
5066 bic_set_bfqq(bic, bfqq, false);
5069 bfqq = bic_to_bfqq(bic, true);
5071 bfq_set_next_ioprio_data(bfqq, bic);
5074 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5075 struct bfq_io_cq *bic, pid_t pid, int is_sync)
5077 RB_CLEAR_NODE(&bfqq->entity.rb_node);
5078 INIT_LIST_HEAD(&bfqq->fifo);
5079 INIT_HLIST_NODE(&bfqq->burst_list_node);
5080 INIT_HLIST_NODE(&bfqq->woken_list_node);
5081 INIT_HLIST_HEAD(&bfqq->woken_list);
5087 bfq_set_next_ioprio_data(bfqq, bic);
5091 * No need to mark as has_short_ttime if in
5092 * idle_class, because no device idling is performed
5093 * for queues in idle class
5095 if (!bfq_class_idle(bfqq))
5096 /* tentatively mark as has_short_ttime */
5097 bfq_mark_bfqq_has_short_ttime(bfqq);
5098 bfq_mark_bfqq_sync(bfqq);
5099 bfq_mark_bfqq_just_created(bfqq);
5101 bfq_clear_bfqq_sync(bfqq);
5103 /* set end request to minus infinity from now */
5104 bfqq->ttime.last_end_request = ktime_get_ns() + 1;
5106 bfq_mark_bfqq_IO_bound(bfqq);
5110 /* Tentative initial value to trade off between thr and lat */
5111 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
5112 bfqq->budget_timeout = bfq_smallest_from_now();
5115 bfqq->last_wr_start_finish = jiffies;
5116 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
5117 bfqq->split_time = bfq_smallest_from_now();
5120 * To not forget the possibly high bandwidth consumed by a
5121 * process/queue in the recent past,
5122 * bfq_bfqq_softrt_next_start() returns a value at least equal
5123 * to the current value of bfqq->soft_rt_next_start (see
5124 * comments on bfq_bfqq_softrt_next_start). Set
5125 * soft_rt_next_start to now, to mean that bfqq has consumed
5126 * no bandwidth so far.
5128 bfqq->soft_rt_next_start = jiffies;
5130 /* first request is almost certainly seeky */
5131 bfqq->seek_history = 1;
5134 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
5135 struct bfq_group *bfqg,
5136 int ioprio_class, int ioprio)
5138 switch (ioprio_class) {
5139 case IOPRIO_CLASS_RT:
5140 return &bfqg->async_bfqq[0][ioprio];
5141 case IOPRIO_CLASS_NONE:
5142 ioprio = IOPRIO_NORM;
5144 case IOPRIO_CLASS_BE:
5145 return &bfqg->async_bfqq[1][ioprio];
5146 case IOPRIO_CLASS_IDLE:
5147 return &bfqg->async_idle_bfqq;
5153 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5154 struct bio *bio, bool is_sync,
5155 struct bfq_io_cq *bic)
5157 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5158 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5159 struct bfq_queue **async_bfqq = NULL;
5160 struct bfq_queue *bfqq;
5161 struct bfq_group *bfqg;
5163 bfqg = bfq_bio_bfqg(bfqd, bio);
5165 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
5172 bfqq = kmem_cache_alloc_node(bfq_pool,
5173 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
5177 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
5179 bfq_init_entity(&bfqq->entity, bfqg);
5180 bfq_log_bfqq(bfqd, bfqq, "allocated");
5182 bfqq = &bfqd->oom_bfqq;
5183 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
5188 * Pin the queue now that it's allocated, scheduler exit will
5193 * Extra group reference, w.r.t. sync
5194 * queue. This extra reference is removed
5195 * only if bfqq->bfqg disappears, to
5196 * guarantee that this queue is not freed
5197 * until its group goes away.
5199 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
5205 bfqq->ref++; /* get a process reference to this queue */
5206 bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
5210 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
5211 struct bfq_queue *bfqq)
5213 struct bfq_ttime *ttime = &bfqq->ttime;
5214 u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
5216 elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
5218 ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
5219 ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8);
5220 ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
5221 ttime->ttime_samples);
5225 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5228 bfqq->seek_history <<= 1;
5229 bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
5231 if (bfqq->wr_coeff > 1 &&
5232 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
5233 BFQQ_TOTALLY_SEEKY(bfqq))
5234 bfq_bfqq_end_wr(bfqq);
5237 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
5238 struct bfq_queue *bfqq,
5239 struct bfq_io_cq *bic)
5241 bool has_short_ttime = true, state_changed;
5244 * No need to update has_short_ttime if bfqq is async or in
5245 * idle io prio class, or if bfq_slice_idle is zero, because
5246 * no device idling is performed for bfqq in this case.
5248 if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
5249 bfqd->bfq_slice_idle == 0)
5252 /* Idle window just restored, statistics are meaningless. */
5253 if (time_is_after_eq_jiffies(bfqq->split_time +
5254 bfqd->bfq_wr_min_idle_time))
5257 /* Think time is infinite if no process is linked to
5258 * bfqq. Otherwise check average think time to
5259 * decide whether to mark as has_short_ttime
5261 if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
5262 (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
5263 bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle))
5264 has_short_ttime = false;
5266 state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
5268 if (has_short_ttime)
5269 bfq_mark_bfqq_has_short_ttime(bfqq);
5271 bfq_clear_bfqq_has_short_ttime(bfqq);
5274 * Until the base value for the total service time gets
5275 * finally computed for bfqq, the inject limit does depend on
5276 * the think-time state (short|long). In particular, the limit
5277 * is 0 or 1 if the think time is deemed, respectively, as
5278 * short or long (details in the comments in
5279 * bfq_update_inject_limit()). Accordingly, the next
5280 * instructions reset the inject limit if the think-time state
5281 * has changed and the above base value is still to be
5284 * However, the reset is performed only if more than 100 ms
5285 * have elapsed since the last update of the inject limit, or
5286 * (inclusive) if the change is from short to long think
5287 * time. The reason for this waiting is as follows.
5289 * bfqq may have a long think time because of a
5290 * synchronization with some other queue, i.e., because the
5291 * I/O of some other queue may need to be completed for bfqq
5292 * to receive new I/O. Details in the comments on the choice
5293 * of the queue for injection in bfq_select_queue().
5295 * As stressed in those comments, if such a synchronization is
5296 * actually in place, then, without injection on bfqq, the
5297 * blocking I/O cannot happen to served while bfqq is in
5298 * service. As a consequence, if bfqq is granted
5299 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
5300 * is dispatched, until the idle timeout fires. This is likely
5301 * to result in lower bandwidth and higher latencies for bfqq,
5302 * and in a severe loss of total throughput.
5304 * On the opposite end, a non-zero inject limit may allow the
5305 * I/O that blocks bfqq to be executed soon, and therefore
5306 * bfqq to receive new I/O soon.
5308 * But, if the blocking gets actually eliminated, then the
5309 * next think-time sample for bfqq may be very low. This in
5310 * turn may cause bfqq's think time to be deemed
5311 * short. Without the 100 ms barrier, this new state change
5312 * would cause the body of the next if to be executed
5313 * immediately. But this would set to 0 the inject
5314 * limit. Without injection, the blocking I/O would cause the
5315 * think time of bfqq to become long again, and therefore the
5316 * inject limit to be raised again, and so on. The only effect
5317 * of such a steady oscillation between the two think-time
5318 * states would be to prevent effective injection on bfqq.
5320 * In contrast, if the inject limit is not reset during such a
5321 * long time interval as 100 ms, then the number of short
5322 * think time samples can grow significantly before the reset
5323 * is performed. As a consequence, the think time state can
5324 * become stable before the reset. Therefore there will be no
5325 * state change when the 100 ms elapse, and no reset of the
5326 * inject limit. The inject limit remains steadily equal to 1
5327 * both during and after the 100 ms. So injection can be
5328 * performed at all times, and throughput gets boosted.
5330 * An inject limit equal to 1 is however in conflict, in
5331 * general, with the fact that the think time of bfqq is
5332 * short, because injection may be likely to delay bfqq's I/O
5333 * (as explained in the comments in
5334 * bfq_update_inject_limit()). But this does not happen in
5335 * this special case, because bfqq's low think time is due to
5336 * an effective handling of a synchronization, through
5337 * injection. In this special case, bfqq's I/O does not get
5338 * delayed by injection; on the contrary, bfqq's I/O is
5339 * brought forward, because it is not blocked for
5342 * In addition, serving the blocking I/O much sooner, and much
5343 * more frequently than once per I/O-plugging timeout, makes
5344 * it much quicker to detect a waker queue (the concept of
5345 * waker queue is defined in the comments in
5346 * bfq_add_request()). This makes it possible to start sooner
5347 * to boost throughput more effectively, by injecting the I/O
5348 * of the waker queue unconditionally on every
5349 * bfq_dispatch_request().
5351 * One last, important benefit of not resetting the inject
5352 * limit before 100 ms is that, during this time interval, the
5353 * base value for the total service time is likely to get
5354 * finally computed for bfqq, freeing the inject limit from
5355 * its relation with the think time.
5357 if (state_changed && bfqq->last_serv_time_ns == 0 &&
5358 (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
5359 msecs_to_jiffies(100)) ||
5361 bfq_reset_inject_limit(bfqd, bfqq);
5365 * Called when a new fs request (rq) is added to bfqq. Check if there's
5366 * something we should do about it.
5368 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5371 if (rq->cmd_flags & REQ_META)
5372 bfqq->meta_pending++;
5374 bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
5376 if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
5377 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
5378 blk_rq_sectors(rq) < 32;
5379 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
5382 * There is just this request queued: if
5383 * - the request is small, and
5384 * - we are idling to boost throughput, and
5385 * - the queue is not to be expired,
5388 * In this way, if the device is being idled to wait
5389 * for a new request from the in-service queue, we
5390 * avoid unplugging the device and committing the
5391 * device to serve just a small request. In contrast
5392 * we wait for the block layer to decide when to
5393 * unplug the device: hopefully, new requests will be
5394 * merged to this one quickly, then the device will be
5395 * unplugged and larger requests will be dispatched.
5397 if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
5402 * A large enough request arrived, or idling is being
5403 * performed to preserve service guarantees, or
5404 * finally the queue is to be expired: in all these
5405 * cases disk idling is to be stopped, so clear
5406 * wait_request flag and reset timer.
5408 bfq_clear_bfqq_wait_request(bfqq);
5409 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
5412 * The queue is not empty, because a new request just
5413 * arrived. Hence we can safely expire the queue, in
5414 * case of budget timeout, without risking that the
5415 * timestamps of the queue are not updated correctly.
5416 * See [1] for more details.
5419 bfq_bfqq_expire(bfqd, bfqq, false,
5420 BFQQE_BUDGET_TIMEOUT);
5424 /* returns true if it causes the idle timer to be disabled */
5425 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
5427 struct bfq_queue *bfqq = RQ_BFQQ(rq),
5428 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
5429 bool waiting, idle_timer_disabled = false;
5433 * Release the request's reference to the old bfqq
5434 * and make sure one is taken to the shared queue.
5436 new_bfqq->allocated++;
5440 * If the bic associated with the process
5441 * issuing this request still points to bfqq
5442 * (and thus has not been already redirected
5443 * to new_bfqq or even some other bfq_queue),
5444 * then complete the merge and redirect it to
5447 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
5448 bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
5451 bfq_clear_bfqq_just_created(bfqq);
5453 * rq is about to be enqueued into new_bfqq,
5454 * release rq reference on bfqq
5456 bfq_put_queue(bfqq);
5457 rq->elv.priv[1] = new_bfqq;
5461 bfq_update_io_thinktime(bfqd, bfqq);
5462 bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
5463 bfq_update_io_seektime(bfqd, bfqq, rq);
5465 waiting = bfqq && bfq_bfqq_wait_request(bfqq);
5466 bfq_add_request(rq);
5467 idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
5469 rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
5470 list_add_tail(&rq->queuelist, &bfqq->fifo);
5472 bfq_rq_enqueued(bfqd, bfqq, rq);
5474 return idle_timer_disabled;
5477 #ifdef CONFIG_BFQ_CGROUP_DEBUG
5478 static void bfq_update_insert_stats(struct request_queue *q,
5479 struct bfq_queue *bfqq,
5480 bool idle_timer_disabled,
5481 unsigned int cmd_flags)
5487 * bfqq still exists, because it can disappear only after
5488 * either it is merged with another queue, or the process it
5489 * is associated with exits. But both actions must be taken by
5490 * the same process currently executing this flow of
5493 * In addition, the following queue lock guarantees that
5494 * bfqq_group(bfqq) exists as well.
5496 spin_lock_irq(&q->queue_lock);
5497 bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
5498 if (idle_timer_disabled)
5499 bfqg_stats_update_idle_time(bfqq_group(bfqq));
5500 spin_unlock_irq(&q->queue_lock);
5503 static inline void bfq_update_insert_stats(struct request_queue *q,
5504 struct bfq_queue *bfqq,
5505 bool idle_timer_disabled,
5506 unsigned int cmd_flags) {}
5507 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
5509 static struct bfq_queue *bfq_init_rq(struct request *rq);
5511 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
5514 struct request_queue *q = hctx->queue;
5515 struct bfq_data *bfqd = q->elevator->elevator_data;
5516 struct bfq_queue *bfqq;
5517 bool idle_timer_disabled = false;
5518 unsigned int cmd_flags;
5520 spin_lock_irq(&bfqd->lock);
5521 bfqq = bfq_init_rq(rq);
5522 if (blk_mq_sched_try_insert_merge(q, rq)) {
5523 spin_unlock_irq(&bfqd->lock);
5527 blk_mq_sched_request_inserted(rq);
5529 if (!bfqq || at_head || blk_rq_is_passthrough(rq)) {
5531 list_add(&rq->queuelist, &bfqd->dispatch);
5533 list_add_tail(&rq->queuelist, &bfqd->dispatch);
5535 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
5537 * Update bfqq, because, if a queue merge has occurred
5538 * in __bfq_insert_request, then rq has been
5539 * redirected into a new queue.
5543 if (rq_mergeable(rq)) {
5544 elv_rqhash_add(q, rq);
5551 * Cache cmd_flags before releasing scheduler lock, because rq
5552 * may disappear afterwards (for example, because of a request
5555 cmd_flags = rq->cmd_flags;
5557 spin_unlock_irq(&bfqd->lock);
5559 bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
5563 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
5564 struct list_head *list, bool at_head)
5566 while (!list_empty(list)) {
5569 rq = list_first_entry(list, struct request, queuelist);
5570 list_del_init(&rq->queuelist);
5571 bfq_insert_request(hctx, rq, at_head);
5575 static void bfq_update_hw_tag(struct bfq_data *bfqd)
5577 struct bfq_queue *bfqq = bfqd->in_service_queue;
5579 bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
5580 bfqd->rq_in_driver);
5582 if (bfqd->hw_tag == 1)
5586 * This sample is valid if the number of outstanding requests
5587 * is large enough to allow a queueing behavior. Note that the
5588 * sum is not exact, as it's not taking into account deactivated
5591 if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
5595 * If active queue hasn't enough requests and can idle, bfq might not
5596 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
5599 if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
5600 bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
5601 BFQ_HW_QUEUE_THRESHOLD &&
5602 bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
5605 if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
5608 bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
5609 bfqd->max_rq_in_driver = 0;
5610 bfqd->hw_tag_samples = 0;
5612 bfqd->nonrot_with_queueing =
5613 blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
5616 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
5621 bfq_update_hw_tag(bfqd);
5623 bfqd->rq_in_driver--;
5626 if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
5628 * Set budget_timeout (which we overload to store the
5629 * time at which the queue remains with no backlog and
5630 * no outstanding request; used by the weight-raising
5633 bfqq->budget_timeout = jiffies;
5635 bfq_weights_tree_remove(bfqd, bfqq);
5638 now_ns = ktime_get_ns();
5640 bfqq->ttime.last_end_request = now_ns;
5643 * Using us instead of ns, to get a reasonable precision in
5644 * computing rate in next check.
5646 delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
5649 * If the request took rather long to complete, and, according
5650 * to the maximum request size recorded, this completion latency
5651 * implies that the request was certainly served at a very low
5652 * rate (less than 1M sectors/sec), then the whole observation
5653 * interval that lasts up to this time instant cannot be a
5654 * valid time interval for computing a new peak rate. Invoke
5655 * bfq_update_rate_reset to have the following three steps
5657 * - close the observation interval at the last (previous)
5658 * request dispatch or completion
5659 * - compute rate, if possible, for that observation interval
5660 * - reset to zero samples, which will trigger a proper
5661 * re-initialization of the observation interval on next
5664 if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
5665 (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
5666 1UL<<(BFQ_RATE_SHIFT - 10))
5667 bfq_update_rate_reset(bfqd, NULL);
5668 bfqd->last_completion = now_ns;
5669 bfqd->last_completed_rq_bfqq = bfqq;
5672 * If we are waiting to discover whether the request pattern
5673 * of the task associated with the queue is actually
5674 * isochronous, and both requisites for this condition to hold
5675 * are now satisfied, then compute soft_rt_next_start (see the
5676 * comments on the function bfq_bfqq_softrt_next_start()). We
5677 * do not compute soft_rt_next_start if bfqq is in interactive
5678 * weight raising (see the comments in bfq_bfqq_expire() for
5679 * an explanation). We schedule this delayed update when bfqq
5680 * expires, if it still has in-flight requests.
5682 if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
5683 RB_EMPTY_ROOT(&bfqq->sort_list) &&
5684 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
5685 bfqq->soft_rt_next_start =
5686 bfq_bfqq_softrt_next_start(bfqd, bfqq);
5689 * If this is the in-service queue, check if it needs to be expired,
5690 * or if we want to idle in case it has no pending requests.
5692 if (bfqd->in_service_queue == bfqq) {
5693 if (bfq_bfqq_must_idle(bfqq)) {
5694 if (bfqq->dispatched == 0)
5695 bfq_arm_slice_timer(bfqd);
5697 * If we get here, we do not expire bfqq, even
5698 * if bfqq was in budget timeout or had no
5699 * more requests (as controlled in the next
5700 * conditional instructions). The reason for
5701 * not expiring bfqq is as follows.
5703 * Here bfqq->dispatched > 0 holds, but
5704 * bfq_bfqq_must_idle() returned true. This
5705 * implies that, even if no request arrives
5706 * for bfqq before bfqq->dispatched reaches 0,
5707 * bfqq will, however, not be expired on the
5708 * completion event that causes bfqq->dispatch
5709 * to reach zero. In contrast, on this event,
5710 * bfqq will start enjoying device idling
5711 * (I/O-dispatch plugging).
5713 * But, if we expired bfqq here, bfqq would
5714 * not have the chance to enjoy device idling
5715 * when bfqq->dispatched finally reaches
5716 * zero. This would expose bfqq to violation
5717 * of its reserved service guarantees.
5720 } else if (bfq_may_expire_for_budg_timeout(bfqq))
5721 bfq_bfqq_expire(bfqd, bfqq, false,
5722 BFQQE_BUDGET_TIMEOUT);
5723 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
5724 (bfqq->dispatched == 0 ||
5725 !bfq_better_to_idle(bfqq)))
5726 bfq_bfqq_expire(bfqd, bfqq, false,
5727 BFQQE_NO_MORE_REQUESTS);
5730 if (!bfqd->rq_in_driver)
5731 bfq_schedule_dispatch(bfqd);
5734 static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq)
5738 bfq_put_queue(bfqq);
5742 * The processes associated with bfqq may happen to generate their
5743 * cumulative I/O at a lower rate than the rate at which the device
5744 * could serve the same I/O. This is rather probable, e.g., if only
5745 * one process is associated with bfqq and the device is an SSD. It
5746 * results in bfqq becoming often empty while in service. In this
5747 * respect, if BFQ is allowed to switch to another queue when bfqq
5748 * remains empty, then the device goes on being fed with I/O requests,
5749 * and the throughput is not affected. In contrast, if BFQ is not
5750 * allowed to switch to another queue---because bfqq is sync and
5751 * I/O-dispatch needs to be plugged while bfqq is temporarily
5752 * empty---then, during the service of bfqq, there will be frequent
5753 * "service holes", i.e., time intervals during which bfqq gets empty
5754 * and the device can only consume the I/O already queued in its
5755 * hardware queues. During service holes, the device may even get to
5756 * remaining idle. In the end, during the service of bfqq, the device
5757 * is driven at a lower speed than the one it can reach with the kind
5758 * of I/O flowing through bfqq.
5760 * To counter this loss of throughput, BFQ implements a "request
5761 * injection mechanism", which tries to fill the above service holes
5762 * with I/O requests taken from other queues. The hard part in this
5763 * mechanism is finding the right amount of I/O to inject, so as to
5764 * both boost throughput and not break bfqq's bandwidth and latency
5765 * guarantees. In this respect, the mechanism maintains a per-queue
5766 * inject limit, computed as below. While bfqq is empty, the injection
5767 * mechanism dispatches extra I/O requests only until the total number
5768 * of I/O requests in flight---i.e., already dispatched but not yet
5769 * completed---remains lower than this limit.
5771 * A first definition comes in handy to introduce the algorithm by
5772 * which the inject limit is computed. We define as first request for
5773 * bfqq, an I/O request for bfqq that arrives while bfqq is in
5774 * service, and causes bfqq to switch from empty to non-empty. The
5775 * algorithm updates the limit as a function of the effect of
5776 * injection on the service times of only the first requests of
5777 * bfqq. The reason for this restriction is that these are the
5778 * requests whose service time is affected most, because they are the
5779 * first to arrive after injection possibly occurred.
5781 * To evaluate the effect of injection, the algorithm measures the
5782 * "total service time" of first requests. We define as total service
5783 * time of an I/O request, the time that elapses since when the
5784 * request is enqueued into bfqq, to when it is completed. This
5785 * quantity allows the whole effect of injection to be measured. It is
5786 * easy to see why. Suppose that some requests of other queues are
5787 * actually injected while bfqq is empty, and that a new request R
5788 * then arrives for bfqq. If the device does start to serve all or
5789 * part of the injected requests during the service hole, then,
5790 * because of this extra service, it may delay the next invocation of
5791 * the dispatch hook of BFQ. Then, even after R gets eventually
5792 * dispatched, the device may delay the actual service of R if it is
5793 * still busy serving the extra requests, or if it decides to serve,
5794 * before R, some extra request still present in its queues. As a
5795 * conclusion, the cumulative extra delay caused by injection can be
5796 * easily evaluated by just comparing the total service time of first
5797 * requests with and without injection.
5799 * The limit-update algorithm works as follows. On the arrival of a
5800 * first request of bfqq, the algorithm measures the total time of the
5801 * request only if one of the three cases below holds, and, for each
5802 * case, it updates the limit as described below:
5804 * (1) If there is no in-flight request. This gives a baseline for the
5805 * total service time of the requests of bfqq. If the baseline has
5806 * not been computed yet, then, after computing it, the limit is
5807 * set to 1, to start boosting throughput, and to prepare the
5808 * ground for the next case. If the baseline has already been
5809 * computed, then it is updated, in case it results to be lower
5810 * than the previous value.
5812 * (2) If the limit is higher than 0 and there are in-flight
5813 * requests. By comparing the total service time in this case with
5814 * the above baseline, it is possible to know at which extent the
5815 * current value of the limit is inflating the total service
5816 * time. If the inflation is below a certain threshold, then bfqq
5817 * is assumed to be suffering from no perceivable loss of its
5818 * service guarantees, and the limit is even tentatively
5819 * increased. If the inflation is above the threshold, then the
5820 * limit is decreased. Due to the lack of any hysteresis, this
5821 * logic makes the limit oscillate even in steady workload
5822 * conditions. Yet we opted for it, because it is fast in reaching
5823 * the best value for the limit, as a function of the current I/O
5824 * workload. To reduce oscillations, this step is disabled for a
5825 * short time interval after the limit happens to be decreased.
5827 * (3) Periodically, after resetting the limit, to make sure that the
5828 * limit eventually drops in case the workload changes. This is
5829 * needed because, after the limit has gone safely up for a
5830 * certain workload, it is impossible to guess whether the
5831 * baseline total service time may have changed, without measuring
5832 * it again without injection. A more effective version of this
5833 * step might be to just sample the baseline, by interrupting
5834 * injection only once, and then to reset/lower the limit only if
5835 * the total service time with the current limit does happen to be
5838 * More details on each step are provided in the comments on the
5839 * pieces of code that implement these steps: the branch handling the
5840 * transition from empty to non empty in bfq_add_request(), the branch
5841 * handling injection in bfq_select_queue(), and the function
5842 * bfq_choose_bfqq_for_injection(). These comments also explain some
5843 * exceptions, made by the injection mechanism in some special cases.
5845 static void bfq_update_inject_limit(struct bfq_data *bfqd,
5846 struct bfq_queue *bfqq)
5848 u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns;
5849 unsigned int old_limit = bfqq->inject_limit;
5851 if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) {
5852 u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
5854 if (tot_time_ns >= threshold && old_limit > 0) {
5855 bfqq->inject_limit--;
5856 bfqq->decrease_time_jif = jiffies;
5857 } else if (tot_time_ns < threshold &&
5858 old_limit <= bfqd->max_rq_in_driver)
5859 bfqq->inject_limit++;
5863 * Either we still have to compute the base value for the
5864 * total service time, and there seem to be the right
5865 * conditions to do it, or we can lower the last base value
5868 * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
5869 * request in flight, because this function is in the code
5870 * path that handles the completion of a request of bfqq, and,
5871 * in particular, this function is executed before
5872 * bfqd->rq_in_driver is decremented in such a code path.
5874 if ((bfqq->last_serv_time_ns == 0 && bfqd->rq_in_driver == 1) ||
5875 tot_time_ns < bfqq->last_serv_time_ns) {
5876 if (bfqq->last_serv_time_ns == 0) {
5878 * Now we certainly have a base value: make sure we
5879 * start trying injection.
5881 bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
5883 bfqq->last_serv_time_ns = tot_time_ns;
5884 } else if (!bfqd->rqs_injected && bfqd->rq_in_driver == 1)
5886 * No I/O injected and no request still in service in
5887 * the drive: these are the exact conditions for
5888 * computing the base value of the total service time
5889 * for bfqq. So let's update this value, because it is
5890 * rather variable. For example, it varies if the size
5891 * or the spatial locality of the I/O requests in bfqq
5894 bfqq->last_serv_time_ns = tot_time_ns;
5897 /* update complete, not waiting for any request completion any longer */
5898 bfqd->waited_rq = NULL;
5899 bfqd->rqs_injected = false;
5903 * Handle either a requeue or a finish for rq. The things to do are
5904 * the same in both cases: all references to rq are to be dropped. In
5905 * particular, rq is considered completed from the point of view of
5908 static void bfq_finish_requeue_request(struct request *rq)
5910 struct bfq_queue *bfqq = RQ_BFQQ(rq);
5911 struct bfq_data *bfqd;
5914 * rq either is not associated with any icq, or is an already
5915 * requeued request that has not (yet) been re-inserted into
5918 if (!rq->elv.icq || !bfqq)
5923 if (rq->rq_flags & RQF_STARTED)
5924 bfqg_stats_update_completion(bfqq_group(bfqq),
5926 rq->io_start_time_ns,
5929 if (likely(rq->rq_flags & RQF_STARTED)) {
5930 unsigned long flags;
5932 spin_lock_irqsave(&bfqd->lock, flags);
5934 if (rq == bfqd->waited_rq)
5935 bfq_update_inject_limit(bfqd, bfqq);
5937 bfq_completed_request(bfqq, bfqd);
5938 bfq_finish_requeue_request_body(bfqq);
5940 spin_unlock_irqrestore(&bfqd->lock, flags);
5943 * Request rq may be still/already in the scheduler,
5944 * in which case we need to remove it (this should
5945 * never happen in case of requeue). And we cannot
5946 * defer such a check and removal, to avoid
5947 * inconsistencies in the time interval from the end
5948 * of this function to the start of the deferred work.
5949 * This situation seems to occur only in process
5950 * context, as a consequence of a merge. In the
5951 * current version of the code, this implies that the
5955 if (!RB_EMPTY_NODE(&rq->rb_node)) {
5956 bfq_remove_request(rq->q, rq);
5957 bfqg_stats_update_io_remove(bfqq_group(bfqq),
5960 bfq_finish_requeue_request_body(bfqq);
5964 * Reset private fields. In case of a requeue, this allows
5965 * this function to correctly do nothing if it is spuriously
5966 * invoked again on this same request (see the check at the
5967 * beginning of the function). Probably, a better general
5968 * design would be to prevent blk-mq from invoking the requeue
5969 * or finish hooks of an elevator, for a request that is not
5970 * referred by that elevator.
5972 * Resetting the following fields would break the
5973 * request-insertion logic if rq is re-inserted into a bfq
5974 * internal queue, without a re-preparation. Here we assume
5975 * that re-insertions of requeued requests, without
5976 * re-preparation, can happen only for pass_through or at_head
5977 * requests (which are not re-inserted into bfq internal
5980 rq->elv.priv[0] = NULL;
5981 rq->elv.priv[1] = NULL;
5985 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
5986 * was the last process referring to that bfqq.
5988 static struct bfq_queue *
5989 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
5991 bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
5993 if (bfqq_process_refs(bfqq) == 1) {
5994 bfqq->pid = current->pid;
5995 bfq_clear_bfqq_coop(bfqq);
5996 bfq_clear_bfqq_split_coop(bfqq);
6000 bic_set_bfqq(bic, NULL, 1);
6002 bfq_put_cooperator(bfqq);
6004 bfq_release_process_ref(bfqq->bfqd, bfqq);
6008 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
6009 struct bfq_io_cq *bic,
6011 bool split, bool is_sync,
6014 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
6016 if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
6023 bfq_put_queue(bfqq);
6024 bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
6026 bic_set_bfqq(bic, bfqq, is_sync);
6027 if (split && is_sync) {
6028 if ((bic->was_in_burst_list && bfqd->large_burst) ||
6029 bic->saved_in_large_burst)
6030 bfq_mark_bfqq_in_large_burst(bfqq);
6032 bfq_clear_bfqq_in_large_burst(bfqq);
6033 if (bic->was_in_burst_list)
6035 * If bfqq was in the current
6036 * burst list before being
6037 * merged, then we have to add
6038 * it back. And we do not need
6039 * to increase burst_size, as
6040 * we did not decrement
6041 * burst_size when we removed
6042 * bfqq from the burst list as
6043 * a consequence of a merge
6045 * bfq_put_queue). In this
6046 * respect, it would be rather
6047 * costly to know whether the
6048 * current burst list is still
6049 * the same burst list from
6050 * which bfqq was removed on
6051 * the merge. To avoid this
6052 * cost, if bfqq was in a
6053 * burst list, then we add
6054 * bfqq to the current burst
6055 * list without any further
6056 * check. This can cause
6057 * inappropriate insertions,
6058 * but rarely enough to not
6059 * harm the detection of large
6060 * bursts significantly.
6062 hlist_add_head(&bfqq->burst_list_node,
6065 bfqq->split_time = jiffies;
6072 * Only reset private fields. The actual request preparation will be
6073 * performed by bfq_init_rq, when rq is either inserted or merged. See
6074 * comments on bfq_init_rq for the reason behind this delayed
6077 static void bfq_prepare_request(struct request *rq, struct bio *bio)
6080 * Regardless of whether we have an icq attached, we have to
6081 * clear the scheduler pointers, as they might point to
6082 * previously allocated bic/bfqq structs.
6084 rq->elv.priv[0] = rq->elv.priv[1] = NULL;
6088 * If needed, init rq, allocate bfq data structures associated with
6089 * rq, and increment reference counters in the destination bfq_queue
6090 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6091 * not associated with any bfq_queue.
6093 * This function is invoked by the functions that perform rq insertion
6094 * or merging. One may have expected the above preparation operations
6095 * to be performed in bfq_prepare_request, and not delayed to when rq
6096 * is inserted or merged. The rationale behind this delayed
6097 * preparation is that, after the prepare_request hook is invoked for
6098 * rq, rq may still be transformed into a request with no icq, i.e., a
6099 * request not associated with any queue. No bfq hook is invoked to
6100 * signal this transformation. As a consequence, should these
6101 * preparation operations be performed when the prepare_request hook
6102 * is invoked, and should rq be transformed one moment later, bfq
6103 * would end up in an inconsistent state, because it would have
6104 * incremented some queue counters for an rq destined to
6105 * transformation, without any chance to correctly lower these
6106 * counters back. In contrast, no transformation can still happen for
6107 * rq after rq has been inserted or merged. So, it is safe to execute
6108 * these preparation operations when rq is finally inserted or merged.
6110 static struct bfq_queue *bfq_init_rq(struct request *rq)
6112 struct request_queue *q = rq->q;
6113 struct bio *bio = rq->bio;
6114 struct bfq_data *bfqd = q->elevator->elevator_data;
6115 struct bfq_io_cq *bic;
6116 const int is_sync = rq_is_sync(rq);
6117 struct bfq_queue *bfqq;
6118 bool new_queue = false;
6119 bool bfqq_already_existing = false, split = false;
6121 if (unlikely(!rq->elv.icq))
6125 * Assuming that elv.priv[1] is set only if everything is set
6126 * for this rq. This holds true, because this function is
6127 * invoked only for insertion or merging, and, after such
6128 * events, a request cannot be manipulated any longer before
6129 * being removed from bfq.
6131 if (rq->elv.priv[1])
6132 return rq->elv.priv[1];
6134 bic = icq_to_bic(rq->elv.icq);
6136 bfq_check_ioprio_change(bic, bio);
6138 bfq_bic_update_cgroup(bic, bio);
6140 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
6143 if (likely(!new_queue)) {
6144 /* If the queue was seeky for too long, break it apart. */
6145 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
6146 bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
6148 /* Update bic before losing reference to bfqq */
6149 if (bfq_bfqq_in_large_burst(bfqq))
6150 bic->saved_in_large_burst = true;
6152 bfqq = bfq_split_bfqq(bic, bfqq);
6156 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
6160 bfqq_already_existing = true;
6166 bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
6167 rq, bfqq, bfqq->ref);
6169 rq->elv.priv[0] = bic;
6170 rq->elv.priv[1] = bfqq;
6173 * If a bfq_queue has only one process reference, it is owned
6174 * by only this bic: we can then set bfqq->bic = bic. in
6175 * addition, if the queue has also just been split, we have to
6178 if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
6182 * The queue has just been split from a shared
6183 * queue: restore the idle window and the
6184 * possible weight raising period.
6186 bfq_bfqq_resume_state(bfqq, bfqd, bic,
6187 bfqq_already_existing);
6192 * Consider bfqq as possibly belonging to a burst of newly
6193 * created queues only if:
6194 * 1) A burst is actually happening (bfqd->burst_size > 0)
6196 * 2) There is no other active queue. In fact, if, in
6197 * contrast, there are active queues not belonging to the
6198 * possible burst bfqq may belong to, then there is no gain
6199 * in considering bfqq as belonging to a burst, and
6200 * therefore in not weight-raising bfqq. See comments on
6201 * bfq_handle_burst().
6203 * This filtering also helps eliminating false positives,
6204 * occurring when bfqq does not belong to an actual large
6205 * burst, but some background task (e.g., a service) happens
6206 * to trigger the creation of new queues very close to when
6207 * bfqq and its possible companion queues are created. See
6208 * comments on bfq_handle_burst() for further details also on
6211 if (unlikely(bfq_bfqq_just_created(bfqq) &&
6212 (bfqd->burst_size > 0 ||
6213 bfq_tot_busy_queues(bfqd) == 0)))
6214 bfq_handle_burst(bfqd, bfqq);
6220 bfq_idle_slice_timer_body(struct bfq_data *bfqd, struct bfq_queue *bfqq)
6222 enum bfqq_expiration reason;
6223 unsigned long flags;
6225 spin_lock_irqsave(&bfqd->lock, flags);
6228 * Considering that bfqq may be in race, we should firstly check
6229 * whether bfqq is in service before doing something on it. If
6230 * the bfqq in race is not in service, it has already been expired
6231 * through __bfq_bfqq_expire func and its wait_request flags has
6232 * been cleared in __bfq_bfqd_reset_in_service func.
6234 if (bfqq != bfqd->in_service_queue) {
6235 spin_unlock_irqrestore(&bfqd->lock, flags);
6239 bfq_clear_bfqq_wait_request(bfqq);
6241 if (bfq_bfqq_budget_timeout(bfqq))
6243 * Also here the queue can be safely expired
6244 * for budget timeout without wasting
6247 reason = BFQQE_BUDGET_TIMEOUT;
6248 else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
6250 * The queue may not be empty upon timer expiration,
6251 * because we may not disable the timer when the
6252 * first request of the in-service queue arrives
6253 * during disk idling.
6255 reason = BFQQE_TOO_IDLE;
6257 goto schedule_dispatch;
6259 bfq_bfqq_expire(bfqd, bfqq, true, reason);
6262 bfq_schedule_dispatch(bfqd);
6263 spin_unlock_irqrestore(&bfqd->lock, flags);
6267 * Handler of the expiration of the timer running if the in-service queue
6268 * is idling inside its time slice.
6270 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
6272 struct bfq_data *bfqd = container_of(timer, struct bfq_data,
6274 struct bfq_queue *bfqq = bfqd->in_service_queue;
6277 * Theoretical race here: the in-service queue can be NULL or
6278 * different from the queue that was idling if a new request
6279 * arrives for the current queue and there is a full dispatch
6280 * cycle that changes the in-service queue. This can hardly
6281 * happen, but in the worst case we just expire a queue too
6285 bfq_idle_slice_timer_body(bfqd, bfqq);
6287 return HRTIMER_NORESTART;
6290 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
6291 struct bfq_queue **bfqq_ptr)
6293 struct bfq_queue *bfqq = *bfqq_ptr;
6295 bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
6297 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
6299 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
6301 bfq_put_queue(bfqq);
6307 * Release all the bfqg references to its async queues. If we are
6308 * deallocating the group these queues may still contain requests, so
6309 * we reparent them to the root cgroup (i.e., the only one that will
6310 * exist for sure until all the requests on a device are gone).
6312 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
6316 for (i = 0; i < 2; i++)
6317 for (j = 0; j < IOPRIO_BE_NR; j++)
6318 __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
6320 __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
6324 * See the comments on bfq_limit_depth for the purpose of
6325 * the depths set in the function. Return minimum shallow depth we'll use.
6327 static unsigned int bfq_update_depths(struct bfq_data *bfqd,
6328 struct sbitmap_queue *bt)
6330 unsigned int i, j, min_shallow = UINT_MAX;
6333 * In-word depths if no bfq_queue is being weight-raised:
6334 * leaving 25% of tags only for sync reads.
6336 * In next formulas, right-shift the value
6337 * (1U<<bt->sb.shift), instead of computing directly
6338 * (1U<<(bt->sb.shift - something)), to be robust against
6339 * any possible value of bt->sb.shift, without having to
6340 * limit 'something'.
6342 /* no more than 50% of tags for async I/O */
6343 bfqd->word_depths[0][0] = max((1U << bt->sb.shift) >> 1, 1U);
6345 * no more than 75% of tags for sync writes (25% extra tags
6346 * w.r.t. async I/O, to prevent async I/O from starving sync
6349 bfqd->word_depths[0][1] = max(((1U << bt->sb.shift) * 3) >> 2, 1U);
6352 * In-word depths in case some bfq_queue is being weight-
6353 * raised: leaving ~63% of tags for sync reads. This is the
6354 * highest percentage for which, in our tests, application
6355 * start-up times didn't suffer from any regression due to tag
6358 /* no more than ~18% of tags for async I/O */
6359 bfqd->word_depths[1][0] = max(((1U << bt->sb.shift) * 3) >> 4, 1U);
6360 /* no more than ~37% of tags for sync writes (~20% extra tags) */
6361 bfqd->word_depths[1][1] = max(((1U << bt->sb.shift) * 6) >> 4, 1U);
6363 for (i = 0; i < 2; i++)
6364 for (j = 0; j < 2; j++)
6365 min_shallow = min(min_shallow, bfqd->word_depths[i][j]);
6370 static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
6372 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
6373 struct blk_mq_tags *tags = hctx->sched_tags;
6374 unsigned int min_shallow;
6376 min_shallow = bfq_update_depths(bfqd, &tags->bitmap_tags);
6377 sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, min_shallow);
6380 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
6382 bfq_depth_updated(hctx);
6386 static void bfq_exit_queue(struct elevator_queue *e)
6388 struct bfq_data *bfqd = e->elevator_data;
6389 struct bfq_queue *bfqq, *n;
6391 hrtimer_cancel(&bfqd->idle_slice_timer);
6393 spin_lock_irq(&bfqd->lock);
6394 list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
6395 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
6396 spin_unlock_irq(&bfqd->lock);
6398 hrtimer_cancel(&bfqd->idle_slice_timer);
6400 /* release oom-queue reference to root group */
6401 bfqg_and_blkg_put(bfqd->root_group);
6403 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6404 blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
6406 spin_lock_irq(&bfqd->lock);
6407 bfq_put_async_queues(bfqd, bfqd->root_group);
6408 kfree(bfqd->root_group);
6409 spin_unlock_irq(&bfqd->lock);
6412 wbt_enable_default(bfqd->queue);
6417 static void bfq_init_root_group(struct bfq_group *root_group,
6418 struct bfq_data *bfqd)
6422 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6423 root_group->entity.parent = NULL;
6424 root_group->my_entity = NULL;
6425 root_group->bfqd = bfqd;
6427 root_group->rq_pos_tree = RB_ROOT;
6428 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
6429 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
6430 root_group->sched_data.bfq_class_idle_last_service = jiffies;
6433 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
6435 struct bfq_data *bfqd;
6436 struct elevator_queue *eq;
6438 eq = elevator_alloc(q, e);
6442 bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
6444 kobject_put(&eq->kobj);
6447 eq->elevator_data = bfqd;
6449 spin_lock_irq(&q->queue_lock);
6451 spin_unlock_irq(&q->queue_lock);
6454 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
6455 * Grab a permanent reference to it, so that the normal code flow
6456 * will not attempt to free it.
6458 bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
6459 bfqd->oom_bfqq.ref++;
6460 bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
6461 bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
6462 bfqd->oom_bfqq.entity.new_weight =
6463 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
6465 /* oom_bfqq does not participate to bursts */
6466 bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
6469 * Trigger weight initialization, according to ioprio, at the
6470 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
6471 * class won't be changed any more.
6473 bfqd->oom_bfqq.entity.prio_changed = 1;
6477 INIT_LIST_HEAD(&bfqd->dispatch);
6479 hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
6481 bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
6483 bfqd->queue_weights_tree = RB_ROOT_CACHED;
6484 bfqd->num_groups_with_pending_reqs = 0;
6486 INIT_LIST_HEAD(&bfqd->active_list);
6487 INIT_LIST_HEAD(&bfqd->idle_list);
6488 INIT_HLIST_HEAD(&bfqd->burst_list);
6491 bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
6493 bfqd->bfq_max_budget = bfq_default_max_budget;
6495 bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
6496 bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
6497 bfqd->bfq_back_max = bfq_back_max;
6498 bfqd->bfq_back_penalty = bfq_back_penalty;
6499 bfqd->bfq_slice_idle = bfq_slice_idle;
6500 bfqd->bfq_timeout = bfq_timeout;
6502 bfqd->bfq_requests_within_timer = 120;
6504 bfqd->bfq_large_burst_thresh = 8;
6505 bfqd->bfq_burst_interval = msecs_to_jiffies(180);
6507 bfqd->low_latency = true;
6510 * Trade-off between responsiveness and fairness.
6512 bfqd->bfq_wr_coeff = 30;
6513 bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
6514 bfqd->bfq_wr_max_time = 0;
6515 bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
6516 bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
6517 bfqd->bfq_wr_max_softrt_rate = 7000; /*
6518 * Approximate rate required
6519 * to playback or record a
6520 * high-definition compressed
6523 bfqd->wr_busy_queues = 0;
6526 * Begin by assuming, optimistically, that the device peak
6527 * rate is equal to 2/3 of the highest reference rate.
6529 bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
6530 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
6531 bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
6533 spin_lock_init(&bfqd->lock);
6536 * The invocation of the next bfq_create_group_hierarchy
6537 * function is the head of a chain of function calls
6538 * (bfq_create_group_hierarchy->blkcg_activate_policy->
6539 * blk_mq_freeze_queue) that may lead to the invocation of the
6540 * has_work hook function. For this reason,
6541 * bfq_create_group_hierarchy is invoked only after all
6542 * scheduler data has been initialized, apart from the fields
6543 * that can be initialized only after invoking
6544 * bfq_create_group_hierarchy. This, in particular, enables
6545 * has_work to correctly return false. Of course, to avoid
6546 * other inconsistencies, the blk-mq stack must then refrain
6547 * from invoking further scheduler hooks before this init
6548 * function is finished.
6550 bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
6551 if (!bfqd->root_group)
6553 bfq_init_root_group(bfqd->root_group, bfqd);
6554 bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
6556 wbt_disable_default(q);
6561 kobject_put(&eq->kobj);
6565 static void bfq_slab_kill(void)
6567 kmem_cache_destroy(bfq_pool);
6570 static int __init bfq_slab_setup(void)
6572 bfq_pool = KMEM_CACHE(bfq_queue, 0);
6578 static ssize_t bfq_var_show(unsigned int var, char *page)
6580 return sprintf(page, "%u\n", var);
6583 static int bfq_var_store(unsigned long *var, const char *page)
6585 unsigned long new_val;
6586 int ret = kstrtoul(page, 10, &new_val);
6594 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
6595 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
6597 struct bfq_data *bfqd = e->elevator_data; \
6598 u64 __data = __VAR; \
6600 __data = jiffies_to_msecs(__data); \
6601 else if (__CONV == 2) \
6602 __data = div_u64(__data, NSEC_PER_MSEC); \
6603 return bfq_var_show(__data, (page)); \
6605 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
6606 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
6607 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
6608 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
6609 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
6610 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
6611 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
6612 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
6613 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
6614 #undef SHOW_FUNCTION
6616 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
6617 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
6619 struct bfq_data *bfqd = e->elevator_data; \
6620 u64 __data = __VAR; \
6621 __data = div_u64(__data, NSEC_PER_USEC); \
6622 return bfq_var_show(__data, (page)); \
6624 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
6625 #undef USEC_SHOW_FUNCTION
6627 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
6629 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
6631 struct bfq_data *bfqd = e->elevator_data; \
6632 unsigned long __data, __min = (MIN), __max = (MAX); \
6635 ret = bfq_var_store(&__data, (page)); \
6638 if (__data < __min) \
6640 else if (__data > __max) \
6643 *(__PTR) = msecs_to_jiffies(__data); \
6644 else if (__CONV == 2) \
6645 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
6647 *(__PTR) = __data; \
6650 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
6652 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
6654 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
6655 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
6657 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
6658 #undef STORE_FUNCTION
6660 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
6661 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
6663 struct bfq_data *bfqd = e->elevator_data; \
6664 unsigned long __data, __min = (MIN), __max = (MAX); \
6667 ret = bfq_var_store(&__data, (page)); \
6670 if (__data < __min) \
6672 else if (__data > __max) \
6674 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
6677 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
6679 #undef USEC_STORE_FUNCTION
6681 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
6682 const char *page, size_t count)
6684 struct bfq_data *bfqd = e->elevator_data;
6685 unsigned long __data;
6688 ret = bfq_var_store(&__data, (page));
6693 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
6695 if (__data > INT_MAX)
6697 bfqd->bfq_max_budget = __data;
6700 bfqd->bfq_user_max_budget = __data;
6706 * Leaving this name to preserve name compatibility with cfq
6707 * parameters, but this timeout is used for both sync and async.
6709 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
6710 const char *page, size_t count)
6712 struct bfq_data *bfqd = e->elevator_data;
6713 unsigned long __data;
6716 ret = bfq_var_store(&__data, (page));
6722 else if (__data > INT_MAX)
6725 bfqd->bfq_timeout = msecs_to_jiffies(__data);
6726 if (bfqd->bfq_user_max_budget == 0)
6727 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
6732 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
6733 const char *page, size_t count)
6735 struct bfq_data *bfqd = e->elevator_data;
6736 unsigned long __data;
6739 ret = bfq_var_store(&__data, (page));
6745 if (!bfqd->strict_guarantees && __data == 1
6746 && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
6747 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
6749 bfqd->strict_guarantees = __data;
6754 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
6755 const char *page, size_t count)
6757 struct bfq_data *bfqd = e->elevator_data;
6758 unsigned long __data;
6761 ret = bfq_var_store(&__data, (page));
6767 if (__data == 0 && bfqd->low_latency != 0)
6769 bfqd->low_latency = __data;
6774 #define BFQ_ATTR(name) \
6775 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
6777 static struct elv_fs_entry bfq_attrs[] = {
6778 BFQ_ATTR(fifo_expire_sync),
6779 BFQ_ATTR(fifo_expire_async),
6780 BFQ_ATTR(back_seek_max),
6781 BFQ_ATTR(back_seek_penalty),
6782 BFQ_ATTR(slice_idle),
6783 BFQ_ATTR(slice_idle_us),
6784 BFQ_ATTR(max_budget),
6785 BFQ_ATTR(timeout_sync),
6786 BFQ_ATTR(strict_guarantees),
6787 BFQ_ATTR(low_latency),
6791 static struct elevator_type iosched_bfq_mq = {
6793 .limit_depth = bfq_limit_depth,
6794 .prepare_request = bfq_prepare_request,
6795 .requeue_request = bfq_finish_requeue_request,
6796 .finish_request = bfq_finish_requeue_request,
6797 .exit_icq = bfq_exit_icq,
6798 .insert_requests = bfq_insert_requests,
6799 .dispatch_request = bfq_dispatch_request,
6800 .next_request = elv_rb_latter_request,
6801 .former_request = elv_rb_former_request,
6802 .allow_merge = bfq_allow_bio_merge,
6803 .bio_merge = bfq_bio_merge,
6804 .request_merge = bfq_request_merge,
6805 .requests_merged = bfq_requests_merged,
6806 .request_merged = bfq_request_merged,
6807 .has_work = bfq_has_work,
6808 .depth_updated = bfq_depth_updated,
6809 .init_hctx = bfq_init_hctx,
6810 .init_sched = bfq_init_queue,
6811 .exit_sched = bfq_exit_queue,
6814 .icq_size = sizeof(struct bfq_io_cq),
6815 .icq_align = __alignof__(struct bfq_io_cq),
6816 .elevator_attrs = bfq_attrs,
6817 .elevator_name = "bfq",
6818 .elevator_owner = THIS_MODULE,
6820 MODULE_ALIAS("bfq-iosched");
6822 static int __init bfq_init(void)
6826 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6827 ret = blkcg_policy_register(&blkcg_policy_bfq);
6833 if (bfq_slab_setup())
6837 * Times to load large popular applications for the typical
6838 * systems installed on the reference devices (see the
6839 * comments before the definition of the next
6840 * array). Actually, we use slightly lower values, as the
6841 * estimated peak rate tends to be smaller than the actual
6842 * peak rate. The reason for this last fact is that estimates
6843 * are computed over much shorter time intervals than the long
6844 * intervals typically used for benchmarking. Why? First, to
6845 * adapt more quickly to variations. Second, because an I/O
6846 * scheduler cannot rely on a peak-rate-evaluation workload to
6847 * be run for a long time.
6849 ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
6850 ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
6852 ret = elv_register(&iosched_bfq_mq);
6861 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6862 blkcg_policy_unregister(&blkcg_policy_bfq);
6867 static void __exit bfq_exit(void)
6869 elv_unregister(&iosched_bfq_mq);
6870 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6871 blkcg_policy_unregister(&blkcg_policy_bfq);
6876 module_init(bfq_init);
6877 module_exit(bfq_exit);
6879 MODULE_AUTHOR("Paolo Valente");
6880 MODULE_LICENSE("GPL");
6881 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");