1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Budget Fair Queueing (BFQ) I/O scheduler.
5 * Based on ideas and code from CFQ:
6 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9 * Paolo Valente <paolo.valente@unimore.it>
11 * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
12 * Arianna Avanzini <avanzini@google.com>
14 * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
16 * BFQ is a proportional-share I/O scheduler, with some extra
17 * low-latency capabilities. BFQ also supports full hierarchical
18 * scheduling through cgroups. Next paragraphs provide an introduction
19 * on BFQ inner workings. Details on BFQ benefits, usage and
20 * limitations can be found in Documentation/block/bfq-iosched.rst.
22 * BFQ is a proportional-share storage-I/O scheduling algorithm based
23 * on the slice-by-slice service scheme of CFQ. But BFQ assigns
24 * budgets, measured in number of sectors, to processes instead of
25 * time slices. The device is not granted to the in-service process
26 * for a given time slice, but until it has exhausted its assigned
27 * budget. This change from the time to the service domain enables BFQ
28 * to distribute the device throughput among processes as desired,
29 * without any distortion due to throughput fluctuations, or to device
30 * internal queueing. BFQ uses an ad hoc internal scheduler, called
31 * B-WF2Q+, to schedule processes according to their budgets. More
32 * precisely, BFQ schedules queues associated with processes. Each
33 * process/queue is assigned a user-configurable weight, and B-WF2Q+
34 * guarantees that each queue receives a fraction of the throughput
35 * proportional to its weight. Thanks to the accurate policy of
36 * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
37 * processes issuing sequential requests (to boost the throughput),
38 * and yet guarantee a low latency to interactive and soft real-time
41 * In particular, to provide these low-latency guarantees, BFQ
42 * explicitly privileges the I/O of two classes of time-sensitive
43 * applications: interactive and soft real-time. In more detail, BFQ
44 * behaves this way if the low_latency parameter is set (default
45 * configuration). This feature enables BFQ to provide applications in
46 * these classes with a very low latency.
48 * To implement this feature, BFQ constantly tries to detect whether
49 * the I/O requests in a bfq_queue come from an interactive or a soft
50 * real-time application. For brevity, in these cases, the queue is
51 * said to be interactive or soft real-time. In both cases, BFQ
52 * privileges the service of the queue, over that of non-interactive
53 * and non-soft-real-time queues. This privileging is performed,
54 * mainly, by raising the weight of the queue. So, for brevity, we
55 * call just weight-raising periods the time periods during which a
56 * queue is privileged, because deemed interactive or soft real-time.
58 * The detection of soft real-time queues/applications is described in
59 * detail in the comments on the function
60 * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
61 * interactive queue works as follows: a queue is deemed interactive
62 * if it is constantly non empty only for a limited time interval,
63 * after which it does become empty. The queue may be deemed
64 * interactive again (for a limited time), if it restarts being
65 * constantly non empty, provided that this happens only after the
66 * queue has remained empty for a given minimum idle time.
68 * By default, BFQ computes automatically the above maximum time
69 * interval, i.e., the time interval after which a constantly
70 * non-empty queue stops being deemed interactive. Since a queue is
71 * weight-raised while it is deemed interactive, this maximum time
72 * interval happens to coincide with the (maximum) duration of the
73 * weight-raising for interactive queues.
75 * Finally, BFQ also features additional heuristics for
76 * preserving both a low latency and a high throughput on NCQ-capable,
77 * rotational or flash-based devices, and to get the job done quickly
78 * for applications consisting in many I/O-bound processes.
80 * NOTE: if the main or only goal, with a given device, is to achieve
81 * the maximum-possible throughput at all times, then do switch off
82 * all low-latency heuristics for that device, by setting low_latency
85 * BFQ is described in [1], where also a reference to the initial,
86 * more theoretical paper on BFQ can be found. The interested reader
87 * can find in the latter paper full details on the main algorithm, as
88 * well as formulas of the guarantees and formal proofs of all the
89 * properties. With respect to the version of BFQ presented in these
90 * papers, this implementation adds a few more heuristics, such as the
91 * ones that guarantee a low latency to interactive and soft real-time
92 * applications, and a hierarchical extension based on H-WF2Q+.
94 * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
95 * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
96 * with O(log N) complexity derives from the one introduced with EEVDF
99 * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
100 * Scheduler", Proceedings of the First Workshop on Mobile System
101 * Technologies (MST-2015), May 2015.
102 * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
104 * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
105 * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
108 * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
110 * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
111 * First: A Flexible and Accurate Mechanism for Proportional Share
112 * Resource Allocation", technical report.
114 * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
116 #include <linux/module.h>
117 #include <linux/slab.h>
118 #include <linux/blkdev.h>
119 #include <linux/cgroup.h>
120 #include <linux/elevator.h>
121 #include <linux/ktime.h>
122 #include <linux/rbtree.h>
123 #include <linux/ioprio.h>
124 #include <linux/sbitmap.h>
125 #include <linux/delay.h>
126 #include <linux/backing-dev.h>
128 #include <trace/events/block.h>
132 #include "blk-mq-tag.h"
133 #include "blk-mq-sched.h"
134 #include "bfq-iosched.h"
137 #define BFQ_BFQQ_FNS(name) \
138 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
140 __set_bit(BFQQF_##name, &(bfqq)->flags); \
142 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
144 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
146 int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
148 return test_bit(BFQQF_##name, &(bfqq)->flags); \
151 BFQ_BFQQ_FNS(just_created
);
153 BFQ_BFQQ_FNS(wait_request
);
154 BFQ_BFQQ_FNS(non_blocking_wait_rq
);
155 BFQ_BFQQ_FNS(fifo_expire
);
156 BFQ_BFQQ_FNS(has_short_ttime
);
158 BFQ_BFQQ_FNS(IO_bound
);
159 BFQ_BFQQ_FNS(in_large_burst
);
161 BFQ_BFQQ_FNS(split_coop
);
162 BFQ_BFQQ_FNS(softrt_update
);
163 #undef BFQ_BFQQ_FNS \
165 /* Expiration time of async (0) and sync (1) requests, in ns. */
166 static const u64 bfq_fifo_expire
[2] = { NSEC_PER_SEC
/ 4, NSEC_PER_SEC
/ 8 };
168 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
169 static const int bfq_back_max
= 16 * 1024;
171 /* Penalty of a backwards seek, in number of sectors. */
172 static const int bfq_back_penalty
= 2;
174 /* Idling period duration, in ns. */
175 static u64 bfq_slice_idle
= NSEC_PER_SEC
/ 125;
177 /* Minimum number of assigned budgets for which stats are safe to compute. */
178 static const int bfq_stats_min_budgets
= 194;
180 /* Default maximum budget values, in sectors and number of requests. */
181 static const int bfq_default_max_budget
= 16 * 1024;
184 * When a sync request is dispatched, the queue that contains that
185 * request, and all the ancestor entities of that queue, are charged
186 * with the number of sectors of the request. In contrast, if the
187 * request is async, then the queue and its ancestor entities are
188 * charged with the number of sectors of the request, multiplied by
189 * the factor below. This throttles the bandwidth for async I/O,
190 * w.r.t. to sync I/O, and it is done to counter the tendency of async
191 * writes to steal I/O throughput to reads.
193 * The current value of this parameter is the result of a tuning with
194 * several hardware and software configurations. We tried to find the
195 * lowest value for which writes do not cause noticeable problems to
196 * reads. In fact, the lower this parameter, the stabler I/O control,
197 * in the following respect. The lower this parameter is, the less
198 * the bandwidth enjoyed by a group decreases
199 * - when the group does writes, w.r.t. to when it does reads;
200 * - when other groups do reads, w.r.t. to when they do writes.
202 static const int bfq_async_charge_factor
= 3;
204 /* Default timeout values, in jiffies, approximating CFQ defaults. */
205 const int bfq_timeout
= HZ
/ 8;
208 * Time limit for merging (see comments in bfq_setup_cooperator). Set
209 * to the slowest value that, in our tests, proved to be effective in
210 * removing false positives, while not causing true positives to miss
213 * As can be deduced from the low time limit below, queue merging, if
214 * successful, happens at the very beginning of the I/O of the involved
215 * cooperating processes, as a consequence of the arrival of the very
216 * first requests from each cooperator. After that, there is very
217 * little chance to find cooperators.
219 static const unsigned long bfq_merge_time_limit
= HZ
/10;
221 static struct kmem_cache
*bfq_pool
;
223 /* Below this threshold (in ns), we consider thinktime immediate. */
224 #define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
226 /* hw_tag detection: parallel requests threshold and min samples needed. */
227 #define BFQ_HW_QUEUE_THRESHOLD 3
228 #define BFQ_HW_QUEUE_SAMPLES 32
230 #define BFQQ_SEEK_THR (sector_t)(8 * 100)
231 #define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
232 #define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \
233 (get_sdist(last_pos, rq) > \
235 (!blk_queue_nonrot(bfqd->queue) || \
236 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT))
237 #define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
238 #define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19)
240 * Sync random I/O is likely to be confused with soft real-time I/O,
241 * because it is characterized by limited throughput and apparently
242 * isochronous arrival pattern. To avoid false positives, queues
243 * containing only random (seeky) I/O are prevented from being tagged
246 #define BFQQ_TOTALLY_SEEKY(bfqq) (bfqq->seek_history == -1)
248 /* Min number of samples required to perform peak-rate update */
249 #define BFQ_RATE_MIN_SAMPLES 32
250 /* Min observation time interval required to perform a peak-rate update (ns) */
251 #define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
252 /* Target observation time interval for a peak-rate update (ns) */
253 #define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
256 * Shift used for peak-rate fixed precision calculations.
258 * - the current shift: 16 positions
259 * - the current type used to store rate: u32
260 * - the current unit of measure for rate: [sectors/usec], or, more precisely,
261 * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
262 * the range of rates that can be stored is
263 * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
264 * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
265 * [15, 65G] sectors/sec
266 * Which, assuming a sector size of 512B, corresponds to a range of
269 #define BFQ_RATE_SHIFT 16
272 * When configured for computing the duration of the weight-raising
273 * for interactive queues automatically (see the comments at the
274 * beginning of this file), BFQ does it using the following formula:
275 * duration = (ref_rate / r) * ref_wr_duration,
276 * where r is the peak rate of the device, and ref_rate and
277 * ref_wr_duration are two reference parameters. In particular,
278 * ref_rate is the peak rate of the reference storage device (see
279 * below), and ref_wr_duration is about the maximum time needed, with
280 * BFQ and while reading two files in parallel, to load typical large
281 * applications on the reference device (see the comments on
282 * max_service_from_wr below, for more details on how ref_wr_duration
283 * is obtained). In practice, the slower/faster the device at hand
284 * is, the more/less it takes to load applications with respect to the
285 * reference device. Accordingly, the longer/shorter BFQ grants
286 * weight raising to interactive applications.
288 * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
289 * depending on whether the device is rotational or non-rotational.
291 * In the following definitions, ref_rate[0] and ref_wr_duration[0]
292 * are the reference values for a rotational device, whereas
293 * ref_rate[1] and ref_wr_duration[1] are the reference values for a
294 * non-rotational device. The reference rates are not the actual peak
295 * rates of the devices used as a reference, but slightly lower
296 * values. The reason for using slightly lower values is that the
297 * peak-rate estimator tends to yield slightly lower values than the
298 * actual peak rate (it can yield the actual peak rate only if there
299 * is only one process doing I/O, and the process does sequential
302 * The reference peak rates are measured in sectors/usec, left-shifted
305 static int ref_rate
[2] = {14000, 33000};
307 * To improve readability, a conversion function is used to initialize
308 * the following array, which entails that the array can be
309 * initialized only in a function.
311 static int ref_wr_duration
[2];
314 * BFQ uses the above-detailed, time-based weight-raising mechanism to
315 * privilege interactive tasks. This mechanism is vulnerable to the
316 * following false positives: I/O-bound applications that will go on
317 * doing I/O for much longer than the duration of weight
318 * raising. These applications have basically no benefit from being
319 * weight-raised at the beginning of their I/O. On the opposite end,
320 * while being weight-raised, these applications
321 * a) unjustly steal throughput to applications that may actually need
323 * b) make BFQ uselessly perform device idling; device idling results
324 * in loss of device throughput with most flash-based storage, and may
325 * increase latencies when used purposelessly.
327 * BFQ tries to reduce these problems, by adopting the following
328 * countermeasure. To introduce this countermeasure, we need first to
329 * finish explaining how the duration of weight-raising for
330 * interactive tasks is computed.
332 * For a bfq_queue deemed as interactive, the duration of weight
333 * raising is dynamically adjusted, as a function of the estimated
334 * peak rate of the device, so as to be equal to the time needed to
335 * execute the 'largest' interactive task we benchmarked so far. By
336 * largest task, we mean the task for which each involved process has
337 * to do more I/O than for any of the other tasks we benchmarked. This
338 * reference interactive task is the start-up of LibreOffice Writer,
339 * and in this task each process/bfq_queue needs to have at most ~110K
340 * sectors transferred.
342 * This last piece of information enables BFQ to reduce the actual
343 * duration of weight-raising for at least one class of I/O-bound
344 * applications: those doing sequential or quasi-sequential I/O. An
345 * example is file copy. In fact, once started, the main I/O-bound
346 * processes of these applications usually consume the above 110K
347 * sectors in much less time than the processes of an application that
348 * is starting, because these I/O-bound processes will greedily devote
349 * almost all their CPU cycles only to their target,
350 * throughput-friendly I/O operations. This is even more true if BFQ
351 * happens to be underestimating the device peak rate, and thus
352 * overestimating the duration of weight raising. But, according to
353 * our measurements, once transferred 110K sectors, these processes
354 * have no right to be weight-raised any longer.
356 * Basing on the last consideration, BFQ ends weight-raising for a
357 * bfq_queue if the latter happens to have received an amount of
358 * service at least equal to the following constant. The constant is
359 * set to slightly more than 110K, to have a minimum safety margin.
361 * This early ending of weight-raising reduces the amount of time
362 * during which interactive false positives cause the two problems
363 * described at the beginning of these comments.
365 static const unsigned long max_service_from_wr
= 120000;
367 #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
368 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
370 struct bfq_queue
*bic_to_bfqq(struct bfq_io_cq
*bic
, bool is_sync
)
372 return bic
->bfqq
[is_sync
];
375 void bic_set_bfqq(struct bfq_io_cq
*bic
, struct bfq_queue
*bfqq
, bool is_sync
)
377 bic
->bfqq
[is_sync
] = bfqq
;
380 struct bfq_data
*bic_to_bfqd(struct bfq_io_cq
*bic
)
382 return bic
->icq
.q
->elevator
->elevator_data
;
386 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
387 * @icq: the iocontext queue.
389 static struct bfq_io_cq
*icq_to_bic(struct io_cq
*icq
)
391 /* bic->icq is the first member, %NULL will convert to %NULL */
392 return container_of(icq
, struct bfq_io_cq
, icq
);
396 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
397 * @bfqd: the lookup key.
398 * @ioc: the io_context of the process doing I/O.
399 * @q: the request queue.
401 static struct bfq_io_cq
*bfq_bic_lookup(struct bfq_data
*bfqd
,
402 struct io_context
*ioc
,
403 struct request_queue
*q
)
407 struct bfq_io_cq
*icq
;
409 spin_lock_irqsave(&q
->queue_lock
, flags
);
410 icq
= icq_to_bic(ioc_lookup_icq(ioc
, q
));
411 spin_unlock_irqrestore(&q
->queue_lock
, flags
);
420 * Scheduler run of queue, if there are requests pending and no one in the
421 * driver that will restart queueing.
423 void bfq_schedule_dispatch(struct bfq_data
*bfqd
)
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)
433 #define bfq_sample_valid(samples) ((samples) > 80)
436 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
437 * We choose the request that is closer to the head right now. Distance
438 * behind the head is penalized and only allowed to a certain extent.
440 static struct request
*bfq_choose_req(struct bfq_data
*bfqd
,
445 sector_t s1
, s2
, d1
= 0, d2
= 0;
446 unsigned long back_max
;
447 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
448 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
449 unsigned int wrap
= 0; /* bit mask: requests behind the disk head? */
451 if (!rq1
|| rq1
== rq2
)
456 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
458 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
460 if ((rq1
->cmd_flags
& REQ_META
) && !(rq2
->cmd_flags
& REQ_META
))
462 else if ((rq2
->cmd_flags
& REQ_META
) && !(rq1
->cmd_flags
& REQ_META
))
465 s1
= blk_rq_pos(rq1
);
466 s2
= blk_rq_pos(rq2
);
469 * By definition, 1KiB is 2 sectors.
471 back_max
= bfqd
->bfq_back_max
* 2;
474 * Strict one way elevator _except_ in the case where we allow
475 * short backward seeks which are biased as twice the cost of a
476 * similar forward seek.
480 else if (s1
+ back_max
>= last
)
481 d1
= (last
- s1
) * bfqd
->bfq_back_penalty
;
483 wrap
|= BFQ_RQ1_WRAP
;
487 else if (s2
+ back_max
>= last
)
488 d2
= (last
- s2
) * bfqd
->bfq_back_penalty
;
490 wrap
|= BFQ_RQ2_WRAP
;
492 /* Found required data */
495 * By doing switch() on the bit mask "wrap" we avoid having to
496 * check two variables for all permutations: --> faster!
499 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
514 case BFQ_RQ1_WRAP
|BFQ_RQ2_WRAP
: /* both rqs wrapped */
517 * Since both rqs are wrapped,
518 * start with the one that's further behind head
519 * (--> only *one* back seek required),
520 * since back seek takes more time than forward.
530 * Async I/O can easily starve sync I/O (both sync reads and sync
531 * writes), by consuming all tags. Similarly, storms of sync writes,
532 * such as those that sync(2) may trigger, can starve sync reads.
533 * Limit depths of async I/O and sync writes so as to counter both
536 static void bfq_limit_depth(unsigned int op
, struct blk_mq_alloc_data
*data
)
538 struct bfq_data
*bfqd
= data
->q
->elevator
->elevator_data
;
540 if (op_is_sync(op
) && !op_is_write(op
))
543 data
->shallow_depth
=
544 bfqd
->word_depths
[!!bfqd
->wr_busy_queues
][op_is_sync(op
)];
546 bfq_log(bfqd
, "[%s] wr_busy %d sync %d depth %u",
547 __func__
, bfqd
->wr_busy_queues
, op_is_sync(op
),
548 data
->shallow_depth
);
551 static struct bfq_queue
*
552 bfq_rq_pos_tree_lookup(struct bfq_data
*bfqd
, struct rb_root
*root
,
553 sector_t sector
, struct rb_node
**ret_parent
,
554 struct rb_node
***rb_link
)
556 struct rb_node
**p
, *parent
;
557 struct bfq_queue
*bfqq
= NULL
;
565 bfqq
= rb_entry(parent
, struct bfq_queue
, pos_node
);
568 * Sort strictly based on sector. Smallest to the left,
569 * largest to the right.
571 if (sector
> blk_rq_pos(bfqq
->next_rq
))
573 else if (sector
< blk_rq_pos(bfqq
->next_rq
))
581 *ret_parent
= parent
;
585 bfq_log(bfqd
, "rq_pos_tree_lookup %llu: returning %d",
586 (unsigned long long)sector
,
587 bfqq
? bfqq
->pid
: 0);
592 static bool bfq_too_late_for_merging(struct bfq_queue
*bfqq
)
594 return bfqq
->service_from_backlogged
> 0 &&
595 time_is_before_jiffies(bfqq
->first_IO_time
+
596 bfq_merge_time_limit
);
600 * The following function is not marked as __cold because it is
601 * actually cold, but for the same performance goal described in the
602 * comments on the likely() at the beginning of
603 * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
604 * execution time for the case where this function is not invoked, we
605 * had to add an unlikely() in each involved if().
608 bfq_pos_tree_add_move(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
610 struct rb_node
**p
, *parent
;
611 struct bfq_queue
*__bfqq
;
613 if (bfqq
->pos_root
) {
614 rb_erase(&bfqq
->pos_node
, bfqq
->pos_root
);
615 bfqq
->pos_root
= NULL
;
618 /* oom_bfqq does not participate in queue merging */
619 if (bfqq
== &bfqd
->oom_bfqq
)
623 * bfqq cannot be merged any longer (see comments in
624 * bfq_setup_cooperator): no point in adding bfqq into the
627 if (bfq_too_late_for_merging(bfqq
))
630 if (bfq_class_idle(bfqq
))
635 bfqq
->pos_root
= &bfq_bfqq_to_bfqg(bfqq
)->rq_pos_tree
;
636 __bfqq
= bfq_rq_pos_tree_lookup(bfqd
, bfqq
->pos_root
,
637 blk_rq_pos(bfqq
->next_rq
), &parent
, &p
);
639 rb_link_node(&bfqq
->pos_node
, parent
, p
);
640 rb_insert_color(&bfqq
->pos_node
, bfqq
->pos_root
);
642 bfqq
->pos_root
= NULL
;
646 * The following function returns false either if every active queue
647 * must receive the same share of the throughput (symmetric scenario),
648 * or, as a special case, if bfqq must receive a share of the
649 * throughput lower than or equal to the share that every other active
650 * queue must receive. If bfqq does sync I/O, then these are the only
651 * two cases where bfqq happens to be guaranteed its share of the
652 * throughput even if I/O dispatching is not plugged when bfqq remains
653 * temporarily empty (for more details, see the comments in the
654 * function bfq_better_to_idle()). For this reason, the return value
655 * of this function is used to check whether I/O-dispatch plugging can
658 * The above first case (symmetric scenario) occurs when:
659 * 1) all active queues have the same weight,
660 * 2) all active queues belong to the same I/O-priority class,
661 * 3) all active groups at the same level in the groups tree have the same
663 * 4) all active groups at the same level in the groups tree have the same
664 * number of children.
666 * Unfortunately, keeping the necessary state for evaluating exactly
667 * the last two symmetry sub-conditions above would be quite complex
668 * and time consuming. Therefore this function evaluates, instead,
669 * only the following stronger three sub-conditions, for which it is
670 * much easier to maintain the needed state:
671 * 1) all active queues have the same weight,
672 * 2) all active queues belong to the same I/O-priority class,
673 * 3) there are no active groups.
674 * In particular, the last condition is always true if hierarchical
675 * support or the cgroups interface are not enabled, thus no state
676 * needs to be maintained in this case.
678 static bool bfq_asymmetric_scenario(struct bfq_data
*bfqd
,
679 struct bfq_queue
*bfqq
)
681 bool smallest_weight
= bfqq
&&
682 bfqq
->weight_counter
&&
683 bfqq
->weight_counter
==
685 rb_first_cached(&bfqd
->queue_weights_tree
),
686 struct bfq_weight_counter
,
690 * For queue weights to differ, queue_weights_tree must contain
691 * at least two nodes.
693 bool varied_queue_weights
= !smallest_weight
&&
694 !RB_EMPTY_ROOT(&bfqd
->queue_weights_tree
.rb_root
) &&
695 (bfqd
->queue_weights_tree
.rb_root
.rb_node
->rb_left
||
696 bfqd
->queue_weights_tree
.rb_root
.rb_node
->rb_right
);
698 bool multiple_classes_busy
=
699 (bfqd
->busy_queues
[0] && bfqd
->busy_queues
[1]) ||
700 (bfqd
->busy_queues
[0] && bfqd
->busy_queues
[2]) ||
701 (bfqd
->busy_queues
[1] && bfqd
->busy_queues
[2]);
703 return varied_queue_weights
|| multiple_classes_busy
704 #ifdef CONFIG_BFQ_GROUP_IOSCHED
705 || bfqd
->num_groups_with_pending_reqs
> 0
711 * If the weight-counter tree passed as input contains no counter for
712 * the weight of the input queue, then add that counter; otherwise just
713 * increment the existing counter.
715 * Note that weight-counter trees contain few nodes in mostly symmetric
716 * scenarios. For example, if all queues have the same weight, then the
717 * weight-counter tree for the queues may contain at most one node.
718 * This holds even if low_latency is on, because weight-raised queues
719 * are not inserted in the tree.
720 * In most scenarios, the rate at which nodes are created/destroyed
723 void bfq_weights_tree_add(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
724 struct rb_root_cached
*root
)
726 struct bfq_entity
*entity
= &bfqq
->entity
;
727 struct rb_node
**new = &(root
->rb_root
.rb_node
), *parent
= NULL
;
728 bool leftmost
= true;
731 * Do not insert if the queue is already associated with a
732 * counter, which happens if:
733 * 1) a request arrival has caused the queue to become both
734 * non-weight-raised, and hence change its weight, and
735 * backlogged; in this respect, each of the two events
736 * causes an invocation of this function,
737 * 2) this is the invocation of this function caused by the
738 * second event. This second invocation is actually useless,
739 * and we handle this fact by exiting immediately. More
740 * efficient or clearer solutions might possibly be adopted.
742 if (bfqq
->weight_counter
)
746 struct bfq_weight_counter
*__counter
= container_of(*new,
747 struct bfq_weight_counter
,
751 if (entity
->weight
== __counter
->weight
) {
752 bfqq
->weight_counter
= __counter
;
755 if (entity
->weight
< __counter
->weight
)
756 new = &((*new)->rb_left
);
758 new = &((*new)->rb_right
);
763 bfqq
->weight_counter
= kzalloc(sizeof(struct bfq_weight_counter
),
767 * In the unlucky event of an allocation failure, we just
768 * exit. This will cause the weight of queue to not be
769 * considered in bfq_asymmetric_scenario, which, in its turn,
770 * causes the scenario to be deemed wrongly symmetric in case
771 * bfqq's weight would have been the only weight making the
772 * scenario asymmetric. On the bright side, no unbalance will
773 * however occur when bfqq becomes inactive again (the
774 * invocation of this function is triggered by an activation
775 * of queue). In fact, bfq_weights_tree_remove does nothing
776 * if !bfqq->weight_counter.
778 if (unlikely(!bfqq
->weight_counter
))
781 bfqq
->weight_counter
->weight
= entity
->weight
;
782 rb_link_node(&bfqq
->weight_counter
->weights_node
, parent
, new);
783 rb_insert_color_cached(&bfqq
->weight_counter
->weights_node
, root
,
787 bfqq
->weight_counter
->num_active
++;
792 * Decrement the weight counter associated with the queue, and, if the
793 * counter reaches 0, remove the counter from the tree.
794 * See the comments to the function bfq_weights_tree_add() for considerations
797 void __bfq_weights_tree_remove(struct bfq_data
*bfqd
,
798 struct bfq_queue
*bfqq
,
799 struct rb_root_cached
*root
)
801 if (!bfqq
->weight_counter
)
804 bfqq
->weight_counter
->num_active
--;
805 if (bfqq
->weight_counter
->num_active
> 0)
806 goto reset_entity_pointer
;
808 rb_erase_cached(&bfqq
->weight_counter
->weights_node
, root
);
809 kfree(bfqq
->weight_counter
);
811 reset_entity_pointer
:
812 bfqq
->weight_counter
= NULL
;
817 * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
818 * of active groups for each queue's inactive parent entity.
820 void bfq_weights_tree_remove(struct bfq_data
*bfqd
,
821 struct bfq_queue
*bfqq
)
823 struct bfq_entity
*entity
= bfqq
->entity
.parent
;
825 for_each_entity(entity
) {
826 struct bfq_sched_data
*sd
= entity
->my_sched_data
;
828 if (sd
->next_in_service
|| sd
->in_service_entity
) {
830 * entity is still active, because either
831 * next_in_service or in_service_entity is not
832 * NULL (see the comments on the definition of
833 * next_in_service for details on why
834 * in_service_entity must be checked too).
836 * As a consequence, its parent entities are
837 * active as well, and thus this loop must
844 * The decrement of num_groups_with_pending_reqs is
845 * not performed immediately upon the deactivation of
846 * entity, but it is delayed to when it also happens
847 * that the first leaf descendant bfqq of entity gets
848 * all its pending requests completed. The following
849 * instructions perform this delayed decrement, if
850 * needed. See the comments on
851 * num_groups_with_pending_reqs for details.
853 if (entity
->in_groups_with_pending_reqs
) {
854 entity
->in_groups_with_pending_reqs
= false;
855 bfqd
->num_groups_with_pending_reqs
--;
860 * Next function is invoked last, because it causes bfqq to be
861 * freed if the following holds: bfqq is not in service and
862 * has no dispatched request. DO NOT use bfqq after the next
863 * function invocation.
865 __bfq_weights_tree_remove(bfqd
, bfqq
,
866 &bfqd
->queue_weights_tree
);
870 * Return expired entry, or NULL to just start from scratch in rbtree.
872 static struct request
*bfq_check_fifo(struct bfq_queue
*bfqq
,
873 struct request
*last
)
877 if (bfq_bfqq_fifo_expire(bfqq
))
880 bfq_mark_bfqq_fifo_expire(bfqq
);
882 rq
= rq_entry_fifo(bfqq
->fifo
.next
);
884 if (rq
== last
|| ktime_get_ns() < rq
->fifo_time
)
887 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "check_fifo: returned %p", rq
);
891 static struct request
*bfq_find_next_rq(struct bfq_data
*bfqd
,
892 struct bfq_queue
*bfqq
,
893 struct request
*last
)
895 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
896 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
897 struct request
*next
, *prev
= NULL
;
899 /* Follow expired path, else get first next available. */
900 next
= bfq_check_fifo(bfqq
, last
);
905 prev
= rb_entry_rq(rbprev
);
908 next
= rb_entry_rq(rbnext
);
910 rbnext
= rb_first(&bfqq
->sort_list
);
911 if (rbnext
&& rbnext
!= &last
->rb_node
)
912 next
= rb_entry_rq(rbnext
);
915 return bfq_choose_req(bfqd
, next
, prev
, blk_rq_pos(last
));
918 /* see the definition of bfq_async_charge_factor for details */
919 static unsigned long bfq_serv_to_charge(struct request
*rq
,
920 struct bfq_queue
*bfqq
)
922 if (bfq_bfqq_sync(bfqq
) || bfqq
->wr_coeff
> 1 ||
923 bfq_asymmetric_scenario(bfqq
->bfqd
, bfqq
))
924 return blk_rq_sectors(rq
);
926 return blk_rq_sectors(rq
) * bfq_async_charge_factor
;
930 * bfq_updated_next_req - update the queue after a new next_rq selection.
931 * @bfqd: the device data the queue belongs to.
932 * @bfqq: the queue to update.
934 * If the first request of a queue changes we make sure that the queue
935 * has enough budget to serve at least its first request (if the
936 * request has grown). We do this because if the queue has not enough
937 * budget for its first request, it has to go through two dispatch
938 * rounds to actually get it dispatched.
940 static void bfq_updated_next_req(struct bfq_data
*bfqd
,
941 struct bfq_queue
*bfqq
)
943 struct bfq_entity
*entity
= &bfqq
->entity
;
944 struct request
*next_rq
= bfqq
->next_rq
;
945 unsigned long new_budget
;
950 if (bfqq
== bfqd
->in_service_queue
)
952 * In order not to break guarantees, budgets cannot be
953 * changed after an entity has been selected.
957 new_budget
= max_t(unsigned long,
958 max_t(unsigned long, bfqq
->max_budget
,
959 bfq_serv_to_charge(next_rq
, bfqq
)),
961 if (entity
->budget
!= new_budget
) {
962 entity
->budget
= new_budget
;
963 bfq_log_bfqq(bfqd
, bfqq
, "updated next rq: new budget %lu",
965 bfq_requeue_bfqq(bfqd
, bfqq
, false);
969 static unsigned int bfq_wr_duration(struct bfq_data
*bfqd
)
973 if (bfqd
->bfq_wr_max_time
> 0)
974 return bfqd
->bfq_wr_max_time
;
976 dur
= bfqd
->rate_dur_prod
;
977 do_div(dur
, bfqd
->peak_rate
);
980 * Limit duration between 3 and 25 seconds. The upper limit
981 * has been conservatively set after the following worst case:
982 * on a QEMU/KVM virtual machine
983 * - running in a slow PC
984 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
985 * - serving a heavy I/O workload, such as the sequential reading
987 * mplayer took 23 seconds to start, if constantly weight-raised.
989 * As for higher values than that accommodating the above bad
990 * scenario, tests show that higher values would often yield
991 * the opposite of the desired result, i.e., would worsen
992 * responsiveness by allowing non-interactive applications to
993 * preserve weight raising for too long.
995 * On the other end, lower values than 3 seconds make it
996 * difficult for most interactive tasks to complete their jobs
997 * before weight-raising finishes.
999 return clamp_val(dur
, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
1002 /* switch back from soft real-time to interactive weight raising */
1003 static void switch_back_to_interactive_wr(struct bfq_queue
*bfqq
,
1004 struct bfq_data
*bfqd
)
1006 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1007 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1008 bfqq
->last_wr_start_finish
= bfqq
->wr_start_at_switch_to_srt
;
1012 bfq_bfqq_resume_state(struct bfq_queue
*bfqq
, struct bfq_data
*bfqd
,
1013 struct bfq_io_cq
*bic
, bool bfq_already_existing
)
1015 unsigned int old_wr_coeff
= 1;
1016 bool busy
= bfq_already_existing
&& bfq_bfqq_busy(bfqq
);
1018 if (bic
->saved_has_short_ttime
)
1019 bfq_mark_bfqq_has_short_ttime(bfqq
);
1021 bfq_clear_bfqq_has_short_ttime(bfqq
);
1023 if (bic
->saved_IO_bound
)
1024 bfq_mark_bfqq_IO_bound(bfqq
);
1026 bfq_clear_bfqq_IO_bound(bfqq
);
1028 bfqq
->last_serv_time_ns
= bic
->saved_last_serv_time_ns
;
1029 bfqq
->inject_limit
= bic
->saved_inject_limit
;
1030 bfqq
->decrease_time_jif
= bic
->saved_decrease_time_jif
;
1032 bfqq
->entity
.new_weight
= bic
->saved_weight
;
1033 bfqq
->ttime
= bic
->saved_ttime
;
1034 bfqq
->io_start_time
= bic
->saved_io_start_time
;
1035 bfqq
->tot_idle_time
= bic
->saved_tot_idle_time
;
1037 * Restore weight coefficient only if low_latency is on
1039 if (bfqd
->low_latency
) {
1040 old_wr_coeff
= bfqq
->wr_coeff
;
1041 bfqq
->wr_coeff
= bic
->saved_wr_coeff
;
1043 bfqq
->service_from_wr
= bic
->saved_service_from_wr
;
1044 bfqq
->wr_start_at_switch_to_srt
= bic
->saved_wr_start_at_switch_to_srt
;
1045 bfqq
->last_wr_start_finish
= bic
->saved_last_wr_start_finish
;
1046 bfqq
->wr_cur_max_time
= bic
->saved_wr_cur_max_time
;
1048 if (bfqq
->wr_coeff
> 1 && (bfq_bfqq_in_large_burst(bfqq
) ||
1049 time_is_before_jiffies(bfqq
->last_wr_start_finish
+
1050 bfqq
->wr_cur_max_time
))) {
1051 if (bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
&&
1052 !bfq_bfqq_in_large_burst(bfqq
) &&
1053 time_is_after_eq_jiffies(bfqq
->wr_start_at_switch_to_srt
+
1054 bfq_wr_duration(bfqd
))) {
1055 switch_back_to_interactive_wr(bfqq
, bfqd
);
1058 bfq_log_bfqq(bfqq
->bfqd
, bfqq
,
1059 "resume state: switching off wr");
1063 /* make sure weight will be updated, however we got here */
1064 bfqq
->entity
.prio_changed
= 1;
1069 if (old_wr_coeff
== 1 && bfqq
->wr_coeff
> 1)
1070 bfqd
->wr_busy_queues
++;
1071 else if (old_wr_coeff
> 1 && bfqq
->wr_coeff
== 1)
1072 bfqd
->wr_busy_queues
--;
1075 static int bfqq_process_refs(struct bfq_queue
*bfqq
)
1077 return bfqq
->ref
- bfqq
->allocated
- bfqq
->entity
.on_st_or_in_serv
-
1078 (bfqq
->weight_counter
!= NULL
) - bfqq
->stable_ref
;
1081 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1082 static void bfq_reset_burst_list(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
1084 struct bfq_queue
*item
;
1085 struct hlist_node
*n
;
1087 hlist_for_each_entry_safe(item
, n
, &bfqd
->burst_list
, burst_list_node
)
1088 hlist_del_init(&item
->burst_list_node
);
1091 * Start the creation of a new burst list only if there is no
1092 * active queue. See comments on the conditional invocation of
1093 * bfq_handle_burst().
1095 if (bfq_tot_busy_queues(bfqd
) == 0) {
1096 hlist_add_head(&bfqq
->burst_list_node
, &bfqd
->burst_list
);
1097 bfqd
->burst_size
= 1;
1099 bfqd
->burst_size
= 0;
1101 bfqd
->burst_parent_entity
= bfqq
->entity
.parent
;
1104 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1105 static void bfq_add_to_burst(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
1107 /* Increment burst size to take into account also bfqq */
1110 if (bfqd
->burst_size
== bfqd
->bfq_large_burst_thresh
) {
1111 struct bfq_queue
*pos
, *bfqq_item
;
1112 struct hlist_node
*n
;
1115 * Enough queues have been activated shortly after each
1116 * other to consider this burst as large.
1118 bfqd
->large_burst
= true;
1121 * We can now mark all queues in the burst list as
1122 * belonging to a large burst.
1124 hlist_for_each_entry(bfqq_item
, &bfqd
->burst_list
,
1126 bfq_mark_bfqq_in_large_burst(bfqq_item
);
1127 bfq_mark_bfqq_in_large_burst(bfqq
);
1130 * From now on, and until the current burst finishes, any
1131 * new queue being activated shortly after the last queue
1132 * was inserted in the burst can be immediately marked as
1133 * belonging to a large burst. So the burst list is not
1134 * needed any more. Remove it.
1136 hlist_for_each_entry_safe(pos
, n
, &bfqd
->burst_list
,
1138 hlist_del_init(&pos
->burst_list_node
);
1140 * Burst not yet large: add bfqq to the burst list. Do
1141 * not increment the ref counter for bfqq, because bfqq
1142 * is removed from the burst list before freeing bfqq
1145 hlist_add_head(&bfqq
->burst_list_node
, &bfqd
->burst_list
);
1149 * If many queues belonging to the same group happen to be created
1150 * shortly after each other, then the processes associated with these
1151 * queues have typically a common goal. In particular, bursts of queue
1152 * creations are usually caused by services or applications that spawn
1153 * many parallel threads/processes. Examples are systemd during boot,
1154 * or git grep. To help these processes get their job done as soon as
1155 * possible, it is usually better to not grant either weight-raising
1156 * or device idling to their queues, unless these queues must be
1157 * protected from the I/O flowing through other active queues.
1159 * In this comment we describe, firstly, the reasons why this fact
1160 * holds, and, secondly, the next function, which implements the main
1161 * steps needed to properly mark these queues so that they can then be
1162 * treated in a different way.
1164 * The above services or applications benefit mostly from a high
1165 * throughput: the quicker the requests of the activated queues are
1166 * cumulatively served, the sooner the target job of these queues gets
1167 * completed. As a consequence, weight-raising any of these queues,
1168 * which also implies idling the device for it, is almost always
1169 * counterproductive, unless there are other active queues to isolate
1170 * these new queues from. If there no other active queues, then
1171 * weight-raising these new queues just lowers throughput in most
1174 * On the other hand, a burst of queue creations may be caused also by
1175 * the start of an application that does not consist of a lot of
1176 * parallel I/O-bound threads. In fact, with a complex application,
1177 * several short processes may need to be executed to start-up the
1178 * application. In this respect, to start an application as quickly as
1179 * possible, the best thing to do is in any case to privilege the I/O
1180 * related to the application with respect to all other
1181 * I/O. Therefore, the best strategy to start as quickly as possible
1182 * an application that causes a burst of queue creations is to
1183 * weight-raise all the queues created during the burst. This is the
1184 * exact opposite of the best strategy for the other type of bursts.
1186 * In the end, to take the best action for each of the two cases, the
1187 * two types of bursts need to be distinguished. Fortunately, this
1188 * seems relatively easy, by looking at the sizes of the bursts. In
1189 * particular, we found a threshold such that only bursts with a
1190 * larger size than that threshold are apparently caused by
1191 * services or commands such as systemd or git grep. For brevity,
1192 * hereafter we call just 'large' these bursts. BFQ *does not*
1193 * weight-raise queues whose creation occurs in a large burst. In
1194 * addition, for each of these queues BFQ performs or does not perform
1195 * idling depending on which choice boosts the throughput more. The
1196 * exact choice depends on the device and request pattern at
1199 * Unfortunately, false positives may occur while an interactive task
1200 * is starting (e.g., an application is being started). The
1201 * consequence is that the queues associated with the task do not
1202 * enjoy weight raising as expected. Fortunately these false positives
1203 * are very rare. They typically occur if some service happens to
1204 * start doing I/O exactly when the interactive task starts.
1206 * Turning back to the next function, it is invoked only if there are
1207 * no active queues (apart from active queues that would belong to the
1208 * same, possible burst bfqq would belong to), and it implements all
1209 * the steps needed to detect the occurrence of a large burst and to
1210 * properly mark all the queues belonging to it (so that they can then
1211 * be treated in a different way). This goal is achieved by
1212 * maintaining a "burst list" that holds, temporarily, the queues that
1213 * belong to the burst in progress. The list is then used to mark
1214 * these queues as belonging to a large burst if the burst does become
1215 * large. The main steps are the following.
1217 * . when the very first queue is created, the queue is inserted into the
1218 * list (as it could be the first queue in a possible burst)
1220 * . if the current burst has not yet become large, and a queue Q that does
1221 * not yet belong to the burst is activated shortly after the last time
1222 * at which a new queue entered the burst list, then the function appends
1223 * Q to the burst list
1225 * . if, as a consequence of the previous step, the burst size reaches
1226 * the large-burst threshold, then
1228 * . all the queues in the burst list are marked as belonging to a
1231 * . the burst list is deleted; in fact, the burst list already served
1232 * its purpose (keeping temporarily track of the queues in a burst,
1233 * so as to be able to mark them as belonging to a large burst in the
1234 * previous sub-step), and now is not needed any more
1236 * . the device enters a large-burst mode
1238 * . if a queue Q that does not belong to the burst is created while
1239 * the device is in large-burst mode and shortly after the last time
1240 * at which a queue either entered the burst list or was marked as
1241 * belonging to the current large burst, then Q is immediately marked
1242 * as belonging to a large burst.
1244 * . if a queue Q that does not belong to the burst is created a while
1245 * later, i.e., not shortly after, than the last time at which a queue
1246 * either entered the burst list or was marked as belonging to the
1247 * current large burst, then the current burst is deemed as finished and:
1249 * . the large-burst mode is reset if set
1251 * . the burst list is emptied
1253 * . Q is inserted in the burst list, as Q may be the first queue
1254 * in a possible new burst (then the burst list contains just Q
1257 static void bfq_handle_burst(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
1260 * If bfqq is already in the burst list or is part of a large
1261 * burst, or finally has just been split, then there is
1262 * nothing else to do.
1264 if (!hlist_unhashed(&bfqq
->burst_list_node
) ||
1265 bfq_bfqq_in_large_burst(bfqq
) ||
1266 time_is_after_eq_jiffies(bfqq
->split_time
+
1267 msecs_to_jiffies(10)))
1271 * If bfqq's creation happens late enough, or bfqq belongs to
1272 * a different group than the burst group, then the current
1273 * burst is finished, and related data structures must be
1276 * In this respect, consider the special case where bfqq is
1277 * the very first queue created after BFQ is selected for this
1278 * device. In this case, last_ins_in_burst and
1279 * burst_parent_entity are not yet significant when we get
1280 * here. But it is easy to verify that, whether or not the
1281 * following condition is true, bfqq will end up being
1282 * inserted into the burst list. In particular the list will
1283 * happen to contain only bfqq. And this is exactly what has
1284 * to happen, as bfqq may be the first queue of the first
1287 if (time_is_before_jiffies(bfqd
->last_ins_in_burst
+
1288 bfqd
->bfq_burst_interval
) ||
1289 bfqq
->entity
.parent
!= bfqd
->burst_parent_entity
) {
1290 bfqd
->large_burst
= false;
1291 bfq_reset_burst_list(bfqd
, bfqq
);
1296 * If we get here, then bfqq is being activated shortly after the
1297 * last queue. So, if the current burst is also large, we can mark
1298 * bfqq as belonging to this large burst immediately.
1300 if (bfqd
->large_burst
) {
1301 bfq_mark_bfqq_in_large_burst(bfqq
);
1306 * If we get here, then a large-burst state has not yet been
1307 * reached, but bfqq is being activated shortly after the last
1308 * queue. Then we add bfqq to the burst.
1310 bfq_add_to_burst(bfqd
, bfqq
);
1313 * At this point, bfqq either has been added to the current
1314 * burst or has caused the current burst to terminate and a
1315 * possible new burst to start. In particular, in the second
1316 * case, bfqq has become the first queue in the possible new
1317 * burst. In both cases last_ins_in_burst needs to be moved
1320 bfqd
->last_ins_in_burst
= jiffies
;
1323 static int bfq_bfqq_budget_left(struct bfq_queue
*bfqq
)
1325 struct bfq_entity
*entity
= &bfqq
->entity
;
1327 return entity
->budget
- entity
->service
;
1331 * If enough samples have been computed, return the current max budget
1332 * stored in bfqd, which is dynamically updated according to the
1333 * estimated disk peak rate; otherwise return the default max budget
1335 static int bfq_max_budget(struct bfq_data
*bfqd
)
1337 if (bfqd
->budgets_assigned
< bfq_stats_min_budgets
)
1338 return bfq_default_max_budget
;
1340 return bfqd
->bfq_max_budget
;
1344 * Return min budget, which is a fraction of the current or default
1345 * max budget (trying with 1/32)
1347 static int bfq_min_budget(struct bfq_data
*bfqd
)
1349 if (bfqd
->budgets_assigned
< bfq_stats_min_budgets
)
1350 return bfq_default_max_budget
/ 32;
1352 return bfqd
->bfq_max_budget
/ 32;
1356 * The next function, invoked after the input queue bfqq switches from
1357 * idle to busy, updates the budget of bfqq. The function also tells
1358 * whether the in-service queue should be expired, by returning
1359 * true. The purpose of expiring the in-service queue is to give bfqq
1360 * the chance to possibly preempt the in-service queue, and the reason
1361 * for preempting the in-service queue is to achieve one of the two
1364 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1365 * expired because it has remained idle. In particular, bfqq may have
1366 * expired for one of the following two reasons:
1368 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1369 * and did not make it to issue a new request before its last
1370 * request was served;
1372 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1373 * a new request before the expiration of the idling-time.
1375 * Even if bfqq has expired for one of the above reasons, the process
1376 * associated with the queue may be however issuing requests greedily,
1377 * and thus be sensitive to the bandwidth it receives (bfqq may have
1378 * remained idle for other reasons: CPU high load, bfqq not enjoying
1379 * idling, I/O throttling somewhere in the path from the process to
1380 * the I/O scheduler, ...). But if, after every expiration for one of
1381 * the above two reasons, bfqq has to wait for the service of at least
1382 * one full budget of another queue before being served again, then
1383 * bfqq is likely to get a much lower bandwidth or resource time than
1384 * its reserved ones. To address this issue, two countermeasures need
1387 * First, the budget and the timestamps of bfqq need to be updated in
1388 * a special way on bfqq reactivation: they need to be updated as if
1389 * bfqq did not remain idle and did not expire. In fact, if they are
1390 * computed as if bfqq expired and remained idle until reactivation,
1391 * then the process associated with bfqq is treated as if, instead of
1392 * being greedy, it stopped issuing requests when bfqq remained idle,
1393 * and restarts issuing requests only on this reactivation. In other
1394 * words, the scheduler does not help the process recover the "service
1395 * hole" between bfqq expiration and reactivation. As a consequence,
1396 * the process receives a lower bandwidth than its reserved one. In
1397 * contrast, to recover this hole, the budget must be updated as if
1398 * bfqq was not expired at all before this reactivation, i.e., it must
1399 * be set to the value of the remaining budget when bfqq was
1400 * expired. Along the same line, timestamps need to be assigned the
1401 * value they had the last time bfqq was selected for service, i.e.,
1402 * before last expiration. Thus timestamps need to be back-shifted
1403 * with respect to their normal computation (see [1] for more details
1404 * on this tricky aspect).
1406 * Secondly, to allow the process to recover the hole, the in-service
1407 * queue must be expired too, to give bfqq the chance to preempt it
1408 * immediately. In fact, if bfqq has to wait for a full budget of the
1409 * in-service queue to be completed, then it may become impossible to
1410 * let the process recover the hole, even if the back-shifted
1411 * timestamps of bfqq are lower than those of the in-service queue. If
1412 * this happens for most or all of the holes, then the process may not
1413 * receive its reserved bandwidth. In this respect, it is worth noting
1414 * that, being the service of outstanding requests unpreemptible, a
1415 * little fraction of the holes may however be unrecoverable, thereby
1416 * causing a little loss of bandwidth.
1418 * The last important point is detecting whether bfqq does need this
1419 * bandwidth recovery. In this respect, the next function deems the
1420 * process associated with bfqq greedy, and thus allows it to recover
1421 * the hole, if: 1) the process is waiting for the arrival of a new
1422 * request (which implies that bfqq expired for one of the above two
1423 * reasons), and 2) such a request has arrived soon. The first
1424 * condition is controlled through the flag non_blocking_wait_rq,
1425 * while the second through the flag arrived_in_time. If both
1426 * conditions hold, then the function computes the budget in the
1427 * above-described special way, and signals that the in-service queue
1428 * should be expired. Timestamp back-shifting is done later in
1429 * __bfq_activate_entity.
1431 * 2. Reduce latency. Even if timestamps are not backshifted to let
1432 * the process associated with bfqq recover a service hole, bfqq may
1433 * however happen to have, after being (re)activated, a lower finish
1434 * timestamp than the in-service queue. That is, the next budget of
1435 * bfqq may have to be completed before the one of the in-service
1436 * queue. If this is the case, then preempting the in-service queue
1437 * allows this goal to be achieved, apart from the unpreemptible,
1438 * outstanding requests mentioned above.
1440 * Unfortunately, regardless of which of the above two goals one wants
1441 * to achieve, service trees need first to be updated to know whether
1442 * the in-service queue must be preempted. To have service trees
1443 * correctly updated, the in-service queue must be expired and
1444 * rescheduled, and bfqq must be scheduled too. This is one of the
1445 * most costly operations (in future versions, the scheduling
1446 * mechanism may be re-designed in such a way to make it possible to
1447 * know whether preemption is needed without needing to update service
1448 * trees). In addition, queue preemptions almost always cause random
1449 * I/O, which may in turn cause loss of throughput. Finally, there may
1450 * even be no in-service queue when the next function is invoked (so,
1451 * no queue to compare timestamps with). Because of these facts, the
1452 * next function adopts the following simple scheme to avoid costly
1453 * operations, too frequent preemptions and too many dependencies on
1454 * the state of the scheduler: it requests the expiration of the
1455 * in-service queue (unconditionally) only for queues that need to
1456 * recover a hole. Then it delegates to other parts of the code the
1457 * responsibility of handling the above case 2.
1459 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data
*bfqd
,
1460 struct bfq_queue
*bfqq
,
1461 bool arrived_in_time
)
1463 struct bfq_entity
*entity
= &bfqq
->entity
;
1466 * In the next compound condition, we check also whether there
1467 * is some budget left, because otherwise there is no point in
1468 * trying to go on serving bfqq with this same budget: bfqq
1469 * would be expired immediately after being selected for
1470 * service. This would only cause useless overhead.
1472 if (bfq_bfqq_non_blocking_wait_rq(bfqq
) && arrived_in_time
&&
1473 bfq_bfqq_budget_left(bfqq
) > 0) {
1475 * We do not clear the flag non_blocking_wait_rq here, as
1476 * the latter is used in bfq_activate_bfqq to signal
1477 * that timestamps need to be back-shifted (and is
1478 * cleared right after).
1482 * In next assignment we rely on that either
1483 * entity->service or entity->budget are not updated
1484 * on expiration if bfqq is empty (see
1485 * __bfq_bfqq_recalc_budget). Thus both quantities
1486 * remain unchanged after such an expiration, and the
1487 * following statement therefore assigns to
1488 * entity->budget the remaining budget on such an
1491 entity
->budget
= min_t(unsigned long,
1492 bfq_bfqq_budget_left(bfqq
),
1496 * At this point, we have used entity->service to get
1497 * the budget left (needed for updating
1498 * entity->budget). Thus we finally can, and have to,
1499 * reset entity->service. The latter must be reset
1500 * because bfqq would otherwise be charged again for
1501 * the service it has received during its previous
1504 entity
->service
= 0;
1510 * We can finally complete expiration, by setting service to 0.
1512 entity
->service
= 0;
1513 entity
->budget
= max_t(unsigned long, bfqq
->max_budget
,
1514 bfq_serv_to_charge(bfqq
->next_rq
, bfqq
));
1515 bfq_clear_bfqq_non_blocking_wait_rq(bfqq
);
1520 * Return the farthest past time instant according to jiffies
1523 static unsigned long bfq_smallest_from_now(void)
1525 return jiffies
- MAX_JIFFY_OFFSET
;
1528 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data
*bfqd
,
1529 struct bfq_queue
*bfqq
,
1530 unsigned int old_wr_coeff
,
1531 bool wr_or_deserves_wr
,
1536 if (old_wr_coeff
== 1 && wr_or_deserves_wr
) {
1537 /* start a weight-raising period */
1539 bfqq
->service_from_wr
= 0;
1540 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1541 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1544 * No interactive weight raising in progress
1545 * here: assign minus infinity to
1546 * wr_start_at_switch_to_srt, to make sure
1547 * that, at the end of the soft-real-time
1548 * weight raising periods that is starting
1549 * now, no interactive weight-raising period
1550 * may be wrongly considered as still in
1551 * progress (and thus actually started by
1554 bfqq
->wr_start_at_switch_to_srt
=
1555 bfq_smallest_from_now();
1556 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
*
1557 BFQ_SOFTRT_WEIGHT_FACTOR
;
1558 bfqq
->wr_cur_max_time
=
1559 bfqd
->bfq_wr_rt_max_time
;
1563 * If needed, further reduce budget to make sure it is
1564 * close to bfqq's backlog, so as to reduce the
1565 * scheduling-error component due to a too large
1566 * budget. Do not care about throughput consequences,
1567 * but only about latency. Finally, do not assign a
1568 * too small budget either, to avoid increasing
1569 * latency by causing too frequent expirations.
1571 bfqq
->entity
.budget
= min_t(unsigned long,
1572 bfqq
->entity
.budget
,
1573 2 * bfq_min_budget(bfqd
));
1574 } else if (old_wr_coeff
> 1) {
1575 if (interactive
) { /* update wr coeff and duration */
1576 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1577 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1578 } else if (in_burst
)
1582 * The application is now or still meeting the
1583 * requirements for being deemed soft rt. We
1584 * can then correctly and safely (re)charge
1585 * the weight-raising duration for the
1586 * application with the weight-raising
1587 * duration for soft rt applications.
1589 * In particular, doing this recharge now, i.e.,
1590 * before the weight-raising period for the
1591 * application finishes, reduces the probability
1592 * of the following negative scenario:
1593 * 1) the weight of a soft rt application is
1594 * raised at startup (as for any newly
1595 * created application),
1596 * 2) since the application is not interactive,
1597 * at a certain time weight-raising is
1598 * stopped for the application,
1599 * 3) at that time the application happens to
1600 * still have pending requests, and hence
1601 * is destined to not have a chance to be
1602 * deemed soft rt before these requests are
1603 * completed (see the comments to the
1604 * function bfq_bfqq_softrt_next_start()
1605 * for details on soft rt detection),
1606 * 4) these pending requests experience a high
1607 * latency because the application is not
1608 * weight-raised while they are pending.
1610 if (bfqq
->wr_cur_max_time
!=
1611 bfqd
->bfq_wr_rt_max_time
) {
1612 bfqq
->wr_start_at_switch_to_srt
=
1613 bfqq
->last_wr_start_finish
;
1615 bfqq
->wr_cur_max_time
=
1616 bfqd
->bfq_wr_rt_max_time
;
1617 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
*
1618 BFQ_SOFTRT_WEIGHT_FACTOR
;
1620 bfqq
->last_wr_start_finish
= jiffies
;
1625 static bool bfq_bfqq_idle_for_long_time(struct bfq_data
*bfqd
,
1626 struct bfq_queue
*bfqq
)
1628 return bfqq
->dispatched
== 0 &&
1629 time_is_before_jiffies(
1630 bfqq
->budget_timeout
+
1631 bfqd
->bfq_wr_min_idle_time
);
1636 * Return true if bfqq is in a higher priority class, or has a higher
1637 * weight than the in-service queue.
1639 static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue
*bfqq
,
1640 struct bfq_queue
*in_serv_bfqq
)
1642 int bfqq_weight
, in_serv_weight
;
1644 if (bfqq
->ioprio_class
< in_serv_bfqq
->ioprio_class
)
1647 if (in_serv_bfqq
->entity
.parent
== bfqq
->entity
.parent
) {
1648 bfqq_weight
= bfqq
->entity
.weight
;
1649 in_serv_weight
= in_serv_bfqq
->entity
.weight
;
1651 if (bfqq
->entity
.parent
)
1652 bfqq_weight
= bfqq
->entity
.parent
->weight
;
1654 bfqq_weight
= bfqq
->entity
.weight
;
1655 if (in_serv_bfqq
->entity
.parent
)
1656 in_serv_weight
= in_serv_bfqq
->entity
.parent
->weight
;
1658 in_serv_weight
= in_serv_bfqq
->entity
.weight
;
1661 return bfqq_weight
> in_serv_weight
;
1664 static bool bfq_better_to_idle(struct bfq_queue
*bfqq
);
1666 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data
*bfqd
,
1667 struct bfq_queue
*bfqq
,
1672 bool soft_rt
, in_burst
, wr_or_deserves_wr
,
1673 bfqq_wants_to_preempt
,
1674 idle_for_long_time
= bfq_bfqq_idle_for_long_time(bfqd
, bfqq
),
1676 * See the comments on
1677 * bfq_bfqq_update_budg_for_activation for
1678 * details on the usage of the next variable.
1680 arrived_in_time
= ktime_get_ns() <=
1681 bfqq
->ttime
.last_end_request
+
1682 bfqd
->bfq_slice_idle
* 3;
1686 * bfqq deserves to be weight-raised if:
1688 * - it does not belong to a large burst,
1689 * - it has been idle for enough time or is soft real-time,
1690 * - is linked to a bfq_io_cq (it is not shared in any sense),
1691 * - has a default weight (otherwise we assume the user wanted
1692 * to control its weight explicitly)
1694 in_burst
= bfq_bfqq_in_large_burst(bfqq
);
1695 soft_rt
= bfqd
->bfq_wr_max_softrt_rate
> 0 &&
1696 !BFQQ_TOTALLY_SEEKY(bfqq
) &&
1698 time_is_before_jiffies(bfqq
->soft_rt_next_start
) &&
1699 bfqq
->dispatched
== 0 &&
1700 bfqq
->entity
.new_weight
== 40;
1701 *interactive
= !in_burst
&& idle_for_long_time
&&
1702 bfqq
->entity
.new_weight
== 40;
1703 wr_or_deserves_wr
= bfqd
->low_latency
&&
1704 (bfqq
->wr_coeff
> 1 ||
1705 (bfq_bfqq_sync(bfqq
) &&
1706 bfqq
->bic
&& (*interactive
|| soft_rt
)));
1709 * Using the last flag, update budget and check whether bfqq
1710 * may want to preempt the in-service queue.
1712 bfqq_wants_to_preempt
=
1713 bfq_bfqq_update_budg_for_activation(bfqd
, bfqq
,
1717 * If bfqq happened to be activated in a burst, but has been
1718 * idle for much more than an interactive queue, then we
1719 * assume that, in the overall I/O initiated in the burst, the
1720 * I/O associated with bfqq is finished. So bfqq does not need
1721 * to be treated as a queue belonging to a burst
1722 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1723 * if set, and remove bfqq from the burst list if it's
1724 * there. We do not decrement burst_size, because the fact
1725 * that bfqq does not need to belong to the burst list any
1726 * more does not invalidate the fact that bfqq was created in
1729 if (likely(!bfq_bfqq_just_created(bfqq
)) &&
1730 idle_for_long_time
&&
1731 time_is_before_jiffies(
1732 bfqq
->budget_timeout
+
1733 msecs_to_jiffies(10000))) {
1734 hlist_del_init(&bfqq
->burst_list_node
);
1735 bfq_clear_bfqq_in_large_burst(bfqq
);
1738 bfq_clear_bfqq_just_created(bfqq
);
1740 if (bfqd
->low_latency
) {
1741 if (unlikely(time_is_after_jiffies(bfqq
->split_time
)))
1744 jiffies
- bfqd
->bfq_wr_min_idle_time
- 1;
1746 if (time_is_before_jiffies(bfqq
->split_time
+
1747 bfqd
->bfq_wr_min_idle_time
)) {
1748 bfq_update_bfqq_wr_on_rq_arrival(bfqd
, bfqq
,
1755 if (old_wr_coeff
!= bfqq
->wr_coeff
)
1756 bfqq
->entity
.prio_changed
= 1;
1760 bfqq
->last_idle_bklogged
= jiffies
;
1761 bfqq
->service_from_backlogged
= 0;
1762 bfq_clear_bfqq_softrt_update(bfqq
);
1764 bfq_add_bfqq_busy(bfqd
, bfqq
);
1767 * Expire in-service queue if preemption may be needed for
1768 * guarantees or throughput. As for guarantees, we care
1769 * explicitly about two cases. The first is that bfqq has to
1770 * recover a service hole, as explained in the comments on
1771 * bfq_bfqq_update_budg_for_activation(), i.e., that
1772 * bfqq_wants_to_preempt is true. However, if bfqq does not
1773 * carry time-critical I/O, then bfqq's bandwidth is less
1774 * important than that of queues that carry time-critical I/O.
1775 * So, as a further constraint, we consider this case only if
1776 * bfqq is at least as weight-raised, i.e., at least as time
1777 * critical, as the in-service queue.
1779 * The second case is that bfqq is in a higher priority class,
1780 * or has a higher weight than the in-service queue. If this
1781 * condition does not hold, we don't care because, even if
1782 * bfqq does not start to be served immediately, the resulting
1783 * delay for bfqq's I/O is however lower or much lower than
1784 * the ideal completion time to be guaranteed to bfqq's I/O.
1786 * In both cases, preemption is needed only if, according to
1787 * the timestamps of both bfqq and of the in-service queue,
1788 * bfqq actually is the next queue to serve. So, to reduce
1789 * useless preemptions, the return value of
1790 * next_queue_may_preempt() is considered in the next compound
1791 * condition too. Yet next_queue_may_preempt() just checks a
1792 * simple, necessary condition for bfqq to be the next queue
1793 * to serve. In fact, to evaluate a sufficient condition, the
1794 * timestamps of the in-service queue would need to be
1795 * updated, and this operation is quite costly (see the
1796 * comments on bfq_bfqq_update_budg_for_activation()).
1798 * As for throughput, we ask bfq_better_to_idle() whether we
1799 * still need to plug I/O dispatching. If bfq_better_to_idle()
1800 * says no, then plugging is not needed any longer, either to
1801 * boost throughput or to perserve service guarantees. Then
1802 * the best option is to stop plugging I/O, as not doing so
1803 * would certainly lower throughput. We may end up in this
1804 * case if: (1) upon a dispatch attempt, we detected that it
1805 * was better to plug I/O dispatch, and to wait for a new
1806 * request to arrive for the currently in-service queue, but
1807 * (2) this switch of bfqq to busy changes the scenario.
1809 if (bfqd
->in_service_queue
&&
1810 ((bfqq_wants_to_preempt
&&
1811 bfqq
->wr_coeff
>= bfqd
->in_service_queue
->wr_coeff
) ||
1812 bfq_bfqq_higher_class_or_weight(bfqq
, bfqd
->in_service_queue
) ||
1813 !bfq_better_to_idle(bfqd
->in_service_queue
)) &&
1814 next_queue_may_preempt(bfqd
))
1815 bfq_bfqq_expire(bfqd
, bfqd
->in_service_queue
,
1816 false, BFQQE_PREEMPTED
);
1819 static void bfq_reset_inject_limit(struct bfq_data
*bfqd
,
1820 struct bfq_queue
*bfqq
)
1822 /* invalidate baseline total service time */
1823 bfqq
->last_serv_time_ns
= 0;
1826 * Reset pointer in case we are waiting for
1827 * some request completion.
1829 bfqd
->waited_rq
= NULL
;
1832 * If bfqq has a short think time, then start by setting the
1833 * inject limit to 0 prudentially, because the service time of
1834 * an injected I/O request may be higher than the think time
1835 * of bfqq, and therefore, if one request was injected when
1836 * bfqq remains empty, this injected request might delay the
1837 * service of the next I/O request for bfqq significantly. In
1838 * case bfqq can actually tolerate some injection, then the
1839 * adaptive update will however raise the limit soon. This
1840 * lucky circumstance holds exactly because bfqq has a short
1841 * think time, and thus, after remaining empty, is likely to
1842 * get new I/O enqueued---and then completed---before being
1843 * expired. This is the very pattern that gives the
1844 * limit-update algorithm the chance to measure the effect of
1845 * injection on request service times, and then to update the
1846 * limit accordingly.
1848 * However, in the following special case, the inject limit is
1849 * left to 1 even if the think time is short: bfqq's I/O is
1850 * synchronized with that of some other queue, i.e., bfqq may
1851 * receive new I/O only after the I/O of the other queue is
1852 * completed. Keeping the inject limit to 1 allows the
1853 * blocking I/O to be served while bfqq is in service. And
1854 * this is very convenient both for bfqq and for overall
1855 * throughput, as explained in detail in the comments in
1856 * bfq_update_has_short_ttime().
1858 * On the opposite end, if bfqq has a long think time, then
1859 * start directly by 1, because:
1860 * a) on the bright side, keeping at most one request in
1861 * service in the drive is unlikely to cause any harm to the
1862 * latency of bfqq's requests, as the service time of a single
1863 * request is likely to be lower than the think time of bfqq;
1864 * b) on the downside, after becoming empty, bfqq is likely to
1865 * expire before getting its next request. With this request
1866 * arrival pattern, it is very hard to sample total service
1867 * times and update the inject limit accordingly (see comments
1868 * on bfq_update_inject_limit()). So the limit is likely to be
1869 * never, or at least seldom, updated. As a consequence, by
1870 * setting the limit to 1, we avoid that no injection ever
1871 * occurs with bfqq. On the downside, this proactive step
1872 * further reduces chances to actually compute the baseline
1873 * total service time. Thus it reduces chances to execute the
1874 * limit-update algorithm and possibly raise the limit to more
1877 if (bfq_bfqq_has_short_ttime(bfqq
))
1878 bfqq
->inject_limit
= 0;
1880 bfqq
->inject_limit
= 1;
1882 bfqq
->decrease_time_jif
= jiffies
;
1885 static void bfq_update_io_intensity(struct bfq_queue
*bfqq
, u64 now_ns
)
1887 u64 tot_io_time
= now_ns
- bfqq
->io_start_time
;
1889 if (RB_EMPTY_ROOT(&bfqq
->sort_list
) && bfqq
->dispatched
== 0)
1890 bfqq
->tot_idle_time
+=
1891 now_ns
- bfqq
->ttime
.last_end_request
;
1893 if (unlikely(bfq_bfqq_just_created(bfqq
)))
1897 * Must be busy for at least about 80% of the time to be
1898 * considered I/O bound.
1900 if (bfqq
->tot_idle_time
* 5 > tot_io_time
)
1901 bfq_clear_bfqq_IO_bound(bfqq
);
1903 bfq_mark_bfqq_IO_bound(bfqq
);
1906 * Keep an observation window of at most 200 ms in the past
1909 if (tot_io_time
> 200 * NSEC_PER_MSEC
) {
1910 bfqq
->io_start_time
= now_ns
- (tot_io_time
>>1);
1911 bfqq
->tot_idle_time
>>= 1;
1916 * Detect whether bfqq's I/O seems synchronized with that of some
1917 * other queue, i.e., whether bfqq, after remaining empty, happens to
1918 * receive new I/O only right after some I/O request of the other
1919 * queue has been completed. We call waker queue the other queue, and
1920 * we assume, for simplicity, that bfqq may have at most one waker
1923 * A remarkable throughput boost can be reached by unconditionally
1924 * injecting the I/O of the waker queue, every time a new
1925 * bfq_dispatch_request happens to be invoked while I/O is being
1926 * plugged for bfqq. In addition to boosting throughput, this
1927 * unblocks bfqq's I/O, thereby improving bandwidth and latency for
1928 * bfqq. Note that these same results may be achieved with the general
1929 * injection mechanism, but less effectively. For details on this
1930 * aspect, see the comments on the choice of the queue for injection
1931 * in bfq_select_queue().
1933 * Turning back to the detection of a waker queue, a queue Q is deemed
1934 * as a waker queue for bfqq if, for three consecutive times, bfqq
1935 * happens to become non empty right after a request of Q has been
1936 * completed. In particular, on the first time, Q is tentatively set
1937 * as a candidate waker queue, while on the third consecutive time
1938 * that Q is detected, the field waker_bfqq is set to Q, to confirm
1939 * that Q is a waker queue for bfqq. These detection steps are
1940 * performed only if bfqq has a long think time, so as to make it more
1941 * likely that bfqq's I/O is actually being blocked by a
1942 * synchronization. This last filter, plus the above three-times
1943 * requirement, make false positives less likely.
1947 * The sooner a waker queue is detected, the sooner throughput can be
1948 * boosted by injecting I/O from the waker queue. Fortunately,
1949 * detection is likely to be actually fast, for the following
1950 * reasons. While blocked by synchronization, bfqq has a long think
1951 * time. This implies that bfqq's inject limit is at least equal to 1
1952 * (see the comments in bfq_update_inject_limit()). So, thanks to
1953 * injection, the waker queue is likely to be served during the very
1954 * first I/O-plugging time interval for bfqq. This triggers the first
1955 * step of the detection mechanism. Thanks again to injection, the
1956 * candidate waker queue is then likely to be confirmed no later than
1957 * during the next I/O-plugging interval for bfqq.
1961 * On queue merging all waker information is lost.
1963 static void bfq_check_waker(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
1966 if (!bfqd
->last_completed_rq_bfqq
||
1967 bfqd
->last_completed_rq_bfqq
== bfqq
||
1968 bfq_bfqq_has_short_ttime(bfqq
) ||
1969 now_ns
- bfqd
->last_completion
>= 4 * NSEC_PER_MSEC
||
1970 bfqd
->last_completed_rq_bfqq
== bfqq
->waker_bfqq
)
1973 if (bfqd
->last_completed_rq_bfqq
!=
1974 bfqq
->tentative_waker_bfqq
) {
1976 * First synchronization detected with a
1977 * candidate waker queue, or with a different
1978 * candidate waker queue from the current one.
1980 bfqq
->tentative_waker_bfqq
=
1981 bfqd
->last_completed_rq_bfqq
;
1982 bfqq
->num_waker_detections
= 1;
1983 } else /* Same tentative waker queue detected again */
1984 bfqq
->num_waker_detections
++;
1986 if (bfqq
->num_waker_detections
== 3) {
1987 bfqq
->waker_bfqq
= bfqd
->last_completed_rq_bfqq
;
1988 bfqq
->tentative_waker_bfqq
= NULL
;
1991 * If the waker queue disappears, then
1992 * bfqq->waker_bfqq must be reset. To
1993 * this goal, we maintain in each
1994 * waker queue a list, woken_list, of
1995 * all the queues that reference the
1996 * waker queue through their
1997 * waker_bfqq pointer. When the waker
1998 * queue exits, the waker_bfqq pointer
1999 * of all the queues in the woken_list
2002 * In addition, if bfqq is already in
2003 * the woken_list of a waker queue,
2004 * then, before being inserted into
2005 * the woken_list of a new waker
2006 * queue, bfqq must be removed from
2007 * the woken_list of the old waker
2010 if (!hlist_unhashed(&bfqq
->woken_list_node
))
2011 hlist_del_init(&bfqq
->woken_list_node
);
2012 hlist_add_head(&bfqq
->woken_list_node
,
2013 &bfqd
->last_completed_rq_bfqq
->woken_list
);
2017 static void bfq_add_request(struct request
*rq
)
2019 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
2020 struct bfq_data
*bfqd
= bfqq
->bfqd
;
2021 struct request
*next_rq
, *prev
;
2022 unsigned int old_wr_coeff
= bfqq
->wr_coeff
;
2023 bool interactive
= false;
2024 u64 now_ns
= ktime_get_ns();
2026 bfq_log_bfqq(bfqd
, bfqq
, "add_request %d", rq_is_sync(rq
));
2027 bfqq
->queued
[rq_is_sync(rq
)]++;
2030 if (RB_EMPTY_ROOT(&bfqq
->sort_list
) && bfq_bfqq_sync(bfqq
)) {
2031 bfq_check_waker(bfqd
, bfqq
, now_ns
);
2034 * Periodically reset inject limit, to make sure that
2035 * the latter eventually drops in case workload
2036 * changes, see step (3) in the comments on
2037 * bfq_update_inject_limit().
2039 if (time_is_before_eq_jiffies(bfqq
->decrease_time_jif
+
2040 msecs_to_jiffies(1000)))
2041 bfq_reset_inject_limit(bfqd
, bfqq
);
2044 * The following conditions must hold to setup a new
2045 * sampling of total service time, and then a new
2046 * update of the inject limit:
2047 * - bfqq is in service, because the total service
2048 * time is evaluated only for the I/O requests of
2049 * the queues in service;
2050 * - this is the right occasion to compute or to
2051 * lower the baseline total service time, because
2052 * there are actually no requests in the drive,
2054 * the baseline total service time is available, and
2055 * this is the right occasion to compute the other
2056 * quantity needed to update the inject limit, i.e.,
2057 * the total service time caused by the amount of
2058 * injection allowed by the current value of the
2059 * limit. It is the right occasion because injection
2060 * has actually been performed during the service
2061 * hole, and there are still in-flight requests,
2062 * which are very likely to be exactly the injected
2063 * requests, or part of them;
2064 * - the minimum interval for sampling the total
2065 * service time and updating the inject limit has
2068 if (bfqq
== bfqd
->in_service_queue
&&
2069 (bfqd
->rq_in_driver
== 0 ||
2070 (bfqq
->last_serv_time_ns
> 0 &&
2071 bfqd
->rqs_injected
&& bfqd
->rq_in_driver
> 0)) &&
2072 time_is_before_eq_jiffies(bfqq
->decrease_time_jif
+
2073 msecs_to_jiffies(10))) {
2074 bfqd
->last_empty_occupied_ns
= ktime_get_ns();
2076 * Start the state machine for measuring the
2077 * total service time of rq: setting
2078 * wait_dispatch will cause bfqd->waited_rq to
2079 * be set when rq will be dispatched.
2081 bfqd
->wait_dispatch
= true;
2083 * If there is no I/O in service in the drive,
2084 * then possible injection occurred before the
2085 * arrival of rq will not affect the total
2086 * service time of rq. So the injection limit
2087 * must not be updated as a function of such
2088 * total service time, unless new injection
2089 * occurs before rq is completed. To have the
2090 * injection limit updated only in the latter
2091 * case, reset rqs_injected here (rqs_injected
2092 * will be set in case injection is performed
2093 * on bfqq before rq is completed).
2095 if (bfqd
->rq_in_driver
== 0)
2096 bfqd
->rqs_injected
= false;
2100 if (bfq_bfqq_sync(bfqq
))
2101 bfq_update_io_intensity(bfqq
, now_ns
);
2103 elv_rb_add(&bfqq
->sort_list
, rq
);
2106 * Check if this request is a better next-serve candidate.
2108 prev
= bfqq
->next_rq
;
2109 next_rq
= bfq_choose_req(bfqd
, bfqq
->next_rq
, rq
, bfqd
->last_position
);
2110 bfqq
->next_rq
= next_rq
;
2113 * Adjust priority tree position, if next_rq changes.
2114 * See comments on bfq_pos_tree_add_move() for the unlikely().
2116 if (unlikely(!bfqd
->nonrot_with_queueing
&& prev
!= bfqq
->next_rq
))
2117 bfq_pos_tree_add_move(bfqd
, bfqq
);
2119 if (!bfq_bfqq_busy(bfqq
)) /* switching to busy ... */
2120 bfq_bfqq_handle_idle_busy_switch(bfqd
, bfqq
, old_wr_coeff
,
2123 if (bfqd
->low_latency
&& old_wr_coeff
== 1 && !rq_is_sync(rq
) &&
2124 time_is_before_jiffies(
2125 bfqq
->last_wr_start_finish
+
2126 bfqd
->bfq_wr_min_inter_arr_async
)) {
2127 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
2128 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
2130 bfqd
->wr_busy_queues
++;
2131 bfqq
->entity
.prio_changed
= 1;
2133 if (prev
!= bfqq
->next_rq
)
2134 bfq_updated_next_req(bfqd
, bfqq
);
2138 * Assign jiffies to last_wr_start_finish in the following
2141 * . if bfqq is not going to be weight-raised, because, for
2142 * non weight-raised queues, last_wr_start_finish stores the
2143 * arrival time of the last request; as of now, this piece
2144 * of information is used only for deciding whether to
2145 * weight-raise async queues
2147 * . if bfqq is not weight-raised, because, if bfqq is now
2148 * switching to weight-raised, then last_wr_start_finish
2149 * stores the time when weight-raising starts
2151 * . if bfqq is interactive, because, regardless of whether
2152 * bfqq is currently weight-raised, the weight-raising
2153 * period must start or restart (this case is considered
2154 * separately because it is not detected by the above
2155 * conditions, if bfqq is already weight-raised)
2157 * last_wr_start_finish has to be updated also if bfqq is soft
2158 * real-time, because the weight-raising period is constantly
2159 * restarted on idle-to-busy transitions for these queues, but
2160 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2163 if (bfqd
->low_latency
&&
2164 (old_wr_coeff
== 1 || bfqq
->wr_coeff
== 1 || interactive
))
2165 bfqq
->last_wr_start_finish
= jiffies
;
2168 static struct request
*bfq_find_rq_fmerge(struct bfq_data
*bfqd
,
2170 struct request_queue
*q
)
2172 struct bfq_queue
*bfqq
= bfqd
->bio_bfqq
;
2176 return elv_rb_find(&bfqq
->sort_list
, bio_end_sector(bio
));
2181 static sector_t
get_sdist(sector_t last_pos
, struct request
*rq
)
2184 return abs(blk_rq_pos(rq
) - last_pos
);
2189 #if 0 /* Still not clear if we can do without next two functions */
2190 static void bfq_activate_request(struct request_queue
*q
, struct request
*rq
)
2192 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
2194 bfqd
->rq_in_driver
++;
2197 static void bfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
2199 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
2201 bfqd
->rq_in_driver
--;
2205 static void bfq_remove_request(struct request_queue
*q
,
2208 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
2209 struct bfq_data
*bfqd
= bfqq
->bfqd
;
2210 const int sync
= rq_is_sync(rq
);
2212 if (bfqq
->next_rq
== rq
) {
2213 bfqq
->next_rq
= bfq_find_next_rq(bfqd
, bfqq
, rq
);
2214 bfq_updated_next_req(bfqd
, bfqq
);
2217 if (rq
->queuelist
.prev
!= &rq
->queuelist
)
2218 list_del_init(&rq
->queuelist
);
2219 bfqq
->queued
[sync
]--;
2221 elv_rb_del(&bfqq
->sort_list
, rq
);
2223 elv_rqhash_del(q
, rq
);
2224 if (q
->last_merge
== rq
)
2225 q
->last_merge
= NULL
;
2227 if (RB_EMPTY_ROOT(&bfqq
->sort_list
)) {
2228 bfqq
->next_rq
= NULL
;
2230 if (bfq_bfqq_busy(bfqq
) && bfqq
!= bfqd
->in_service_queue
) {
2231 bfq_del_bfqq_busy(bfqd
, bfqq
, false);
2233 * bfqq emptied. In normal operation, when
2234 * bfqq is empty, bfqq->entity.service and
2235 * bfqq->entity.budget must contain,
2236 * respectively, the service received and the
2237 * budget used last time bfqq emptied. These
2238 * facts do not hold in this case, as at least
2239 * this last removal occurred while bfqq is
2240 * not in service. To avoid inconsistencies,
2241 * reset both bfqq->entity.service and
2242 * bfqq->entity.budget, if bfqq has still a
2243 * process that may issue I/O requests to it.
2245 bfqq
->entity
.budget
= bfqq
->entity
.service
= 0;
2249 * Remove queue from request-position tree as it is empty.
2251 if (bfqq
->pos_root
) {
2252 rb_erase(&bfqq
->pos_node
, bfqq
->pos_root
);
2253 bfqq
->pos_root
= NULL
;
2256 /* see comments on bfq_pos_tree_add_move() for the unlikely() */
2257 if (unlikely(!bfqd
->nonrot_with_queueing
))
2258 bfq_pos_tree_add_move(bfqd
, bfqq
);
2261 if (rq
->cmd_flags
& REQ_META
)
2262 bfqq
->meta_pending
--;
2266 static bool bfq_bio_merge(struct blk_mq_hw_ctx
*hctx
, struct bio
*bio
,
2267 unsigned int nr_segs
)
2269 struct request_queue
*q
= hctx
->queue
;
2270 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
2271 struct request
*free
= NULL
;
2273 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2274 * store its return value for later use, to avoid nesting
2275 * queue_lock inside the bfqd->lock. We assume that the bic
2276 * returned by bfq_bic_lookup does not go away before
2277 * bfqd->lock is taken.
2279 struct bfq_io_cq
*bic
= bfq_bic_lookup(bfqd
, current
->io_context
, q
);
2282 spin_lock_irq(&bfqd
->lock
);
2285 bfqd
->bio_bfqq
= bic_to_bfqq(bic
, op_is_sync(bio
->bi_opf
));
2287 bfqd
->bio_bfqq
= NULL
;
2288 bfqd
->bio_bic
= bic
;
2290 ret
= blk_mq_sched_try_merge(q
, bio
, nr_segs
, &free
);
2293 blk_mq_free_request(free
);
2294 spin_unlock_irq(&bfqd
->lock
);
2299 static int bfq_request_merge(struct request_queue
*q
, struct request
**req
,
2302 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
2303 struct request
*__rq
;
2305 __rq
= bfq_find_rq_fmerge(bfqd
, bio
, q
);
2306 if (__rq
&& elv_bio_merge_ok(__rq
, bio
)) {
2308 return ELEVATOR_FRONT_MERGE
;
2311 return ELEVATOR_NO_MERGE
;
2314 static struct bfq_queue
*bfq_init_rq(struct request
*rq
);
2316 static void bfq_request_merged(struct request_queue
*q
, struct request
*req
,
2317 enum elv_merge type
)
2319 if (type
== ELEVATOR_FRONT_MERGE
&&
2320 rb_prev(&req
->rb_node
) &&
2322 blk_rq_pos(container_of(rb_prev(&req
->rb_node
),
2323 struct request
, rb_node
))) {
2324 struct bfq_queue
*bfqq
= bfq_init_rq(req
);
2325 struct bfq_data
*bfqd
;
2326 struct request
*prev
, *next_rq
;
2333 /* Reposition request in its sort_list */
2334 elv_rb_del(&bfqq
->sort_list
, req
);
2335 elv_rb_add(&bfqq
->sort_list
, req
);
2337 /* Choose next request to be served for bfqq */
2338 prev
= bfqq
->next_rq
;
2339 next_rq
= bfq_choose_req(bfqd
, bfqq
->next_rq
, req
,
2340 bfqd
->last_position
);
2341 bfqq
->next_rq
= next_rq
;
2343 * If next_rq changes, update both the queue's budget to
2344 * fit the new request and the queue's position in its
2347 if (prev
!= bfqq
->next_rq
) {
2348 bfq_updated_next_req(bfqd
, bfqq
);
2350 * See comments on bfq_pos_tree_add_move() for
2353 if (unlikely(!bfqd
->nonrot_with_queueing
))
2354 bfq_pos_tree_add_move(bfqd
, bfqq
);
2360 * This function is called to notify the scheduler that the requests
2361 * rq and 'next' have been merged, with 'next' going away. BFQ
2362 * exploits this hook to address the following issue: if 'next' has a
2363 * fifo_time lower that rq, then the fifo_time of rq must be set to
2364 * the value of 'next', to not forget the greater age of 'next'.
2366 * NOTE: in this function we assume that rq is in a bfq_queue, basing
2367 * on that rq is picked from the hash table q->elevator->hash, which,
2368 * in its turn, is filled only with I/O requests present in
2369 * bfq_queues, while BFQ is in use for the request queue q. In fact,
2370 * the function that fills this hash table (elv_rqhash_add) is called
2371 * only by bfq_insert_request.
2373 static void bfq_requests_merged(struct request_queue
*q
, struct request
*rq
,
2374 struct request
*next
)
2376 struct bfq_queue
*bfqq
= bfq_init_rq(rq
),
2377 *next_bfqq
= bfq_init_rq(next
);
2383 * If next and rq belong to the same bfq_queue and next is older
2384 * than rq, then reposition rq in the fifo (by substituting next
2385 * with rq). Otherwise, if next and rq belong to different
2386 * bfq_queues, never reposition rq: in fact, we would have to
2387 * reposition it with respect to next's position in its own fifo,
2388 * which would most certainly be too expensive with respect to
2391 if (bfqq
== next_bfqq
&&
2392 !list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
2393 next
->fifo_time
< rq
->fifo_time
) {
2394 list_del_init(&rq
->queuelist
);
2395 list_replace_init(&next
->queuelist
, &rq
->queuelist
);
2396 rq
->fifo_time
= next
->fifo_time
;
2399 if (bfqq
->next_rq
== next
)
2402 bfqg_stats_update_io_merged(bfqq_group(bfqq
), next
->cmd_flags
);
2405 /* Must be called with bfqq != NULL */
2406 static void bfq_bfqq_end_wr(struct bfq_queue
*bfqq
)
2409 * If bfqq has been enjoying interactive weight-raising, then
2410 * reset soft_rt_next_start. We do it for the following
2411 * reason. bfqq may have been conveying the I/O needed to load
2412 * a soft real-time application. Such an application actually
2413 * exhibits a soft real-time I/O pattern after it finishes
2414 * loading, and finally starts doing its job. But, if bfqq has
2415 * been receiving a lot of bandwidth so far (likely to happen
2416 * on a fast device), then soft_rt_next_start now contains a
2417 * high value that. So, without this reset, bfqq would be
2418 * prevented from being possibly considered as soft_rt for a
2422 if (bfqq
->wr_cur_max_time
!=
2423 bfqq
->bfqd
->bfq_wr_rt_max_time
)
2424 bfqq
->soft_rt_next_start
= jiffies
;
2426 if (bfq_bfqq_busy(bfqq
))
2427 bfqq
->bfqd
->wr_busy_queues
--;
2429 bfqq
->wr_cur_max_time
= 0;
2430 bfqq
->last_wr_start_finish
= jiffies
;
2432 * Trigger a weight change on the next invocation of
2433 * __bfq_entity_update_weight_prio.
2435 bfqq
->entity
.prio_changed
= 1;
2438 void bfq_end_wr_async_queues(struct bfq_data
*bfqd
,
2439 struct bfq_group
*bfqg
)
2443 for (i
= 0; i
< 2; i
++)
2444 for (j
= 0; j
< IOPRIO_BE_NR
; j
++)
2445 if (bfqg
->async_bfqq
[i
][j
])
2446 bfq_bfqq_end_wr(bfqg
->async_bfqq
[i
][j
]);
2447 if (bfqg
->async_idle_bfqq
)
2448 bfq_bfqq_end_wr(bfqg
->async_idle_bfqq
);
2451 static void bfq_end_wr(struct bfq_data
*bfqd
)
2453 struct bfq_queue
*bfqq
;
2455 spin_lock_irq(&bfqd
->lock
);
2457 list_for_each_entry(bfqq
, &bfqd
->active_list
, bfqq_list
)
2458 bfq_bfqq_end_wr(bfqq
);
2459 list_for_each_entry(bfqq
, &bfqd
->idle_list
, bfqq_list
)
2460 bfq_bfqq_end_wr(bfqq
);
2461 bfq_end_wr_async(bfqd
);
2463 spin_unlock_irq(&bfqd
->lock
);
2466 static sector_t
bfq_io_struct_pos(void *io_struct
, bool request
)
2469 return blk_rq_pos(io_struct
);
2471 return ((struct bio
*)io_struct
)->bi_iter
.bi_sector
;
2474 static int bfq_rq_close_to_sector(void *io_struct
, bool request
,
2477 return abs(bfq_io_struct_pos(io_struct
, request
) - sector
) <=
2481 static struct bfq_queue
*bfqq_find_close(struct bfq_data
*bfqd
,
2482 struct bfq_queue
*bfqq
,
2485 struct rb_root
*root
= &bfq_bfqq_to_bfqg(bfqq
)->rq_pos_tree
;
2486 struct rb_node
*parent
, *node
;
2487 struct bfq_queue
*__bfqq
;
2489 if (RB_EMPTY_ROOT(root
))
2493 * First, if we find a request starting at the end of the last
2494 * request, choose it.
2496 __bfqq
= bfq_rq_pos_tree_lookup(bfqd
, root
, sector
, &parent
, NULL
);
2501 * If the exact sector wasn't found, the parent of the NULL leaf
2502 * will contain the closest sector (rq_pos_tree sorted by
2503 * next_request position).
2505 __bfqq
= rb_entry(parent
, struct bfq_queue
, pos_node
);
2506 if (bfq_rq_close_to_sector(__bfqq
->next_rq
, true, sector
))
2509 if (blk_rq_pos(__bfqq
->next_rq
) < sector
)
2510 node
= rb_next(&__bfqq
->pos_node
);
2512 node
= rb_prev(&__bfqq
->pos_node
);
2516 __bfqq
= rb_entry(node
, struct bfq_queue
, pos_node
);
2517 if (bfq_rq_close_to_sector(__bfqq
->next_rq
, true, sector
))
2523 static struct bfq_queue
*bfq_find_close_cooperator(struct bfq_data
*bfqd
,
2524 struct bfq_queue
*cur_bfqq
,
2527 struct bfq_queue
*bfqq
;
2530 * We shall notice if some of the queues are cooperating,
2531 * e.g., working closely on the same area of the device. In
2532 * that case, we can group them together and: 1) don't waste
2533 * time idling, and 2) serve the union of their requests in
2534 * the best possible order for throughput.
2536 bfqq
= bfqq_find_close(bfqd
, cur_bfqq
, sector
);
2537 if (!bfqq
|| bfqq
== cur_bfqq
)
2543 static struct bfq_queue
*
2544 bfq_setup_merge(struct bfq_queue
*bfqq
, struct bfq_queue
*new_bfqq
)
2546 int process_refs
, new_process_refs
;
2547 struct bfq_queue
*__bfqq
;
2550 * If there are no process references on the new_bfqq, then it is
2551 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2552 * may have dropped their last reference (not just their last process
2555 if (!bfqq_process_refs(new_bfqq
))
2558 /* Avoid a circular list and skip interim queue merges. */
2559 while ((__bfqq
= new_bfqq
->new_bfqq
)) {
2565 process_refs
= bfqq_process_refs(bfqq
);
2566 new_process_refs
= bfqq_process_refs(new_bfqq
);
2568 * If the process for the bfqq has gone away, there is no
2569 * sense in merging the queues.
2571 if (process_refs
== 0 || new_process_refs
== 0)
2574 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "scheduling merge with queue %d",
2578 * Merging is just a redirection: the requests of the process
2579 * owning one of the two queues are redirected to the other queue.
2580 * The latter queue, in its turn, is set as shared if this is the
2581 * first time that the requests of some process are redirected to
2584 * We redirect bfqq to new_bfqq and not the opposite, because
2585 * we are in the context of the process owning bfqq, thus we
2586 * have the io_cq of this process. So we can immediately
2587 * configure this io_cq to redirect the requests of the
2588 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2589 * not available any more (new_bfqq->bic == NULL).
2591 * Anyway, even in case new_bfqq coincides with the in-service
2592 * queue, redirecting requests the in-service queue is the
2593 * best option, as we feed the in-service queue with new
2594 * requests close to the last request served and, by doing so,
2595 * are likely to increase the throughput.
2597 bfqq
->new_bfqq
= new_bfqq
;
2598 new_bfqq
->ref
+= process_refs
;
2602 static bool bfq_may_be_close_cooperator(struct bfq_queue
*bfqq
,
2603 struct bfq_queue
*new_bfqq
)
2605 if (bfq_too_late_for_merging(new_bfqq
))
2608 if (bfq_class_idle(bfqq
) || bfq_class_idle(new_bfqq
) ||
2609 (bfqq
->ioprio_class
!= new_bfqq
->ioprio_class
))
2613 * If either of the queues has already been detected as seeky,
2614 * then merging it with the other queue is unlikely to lead to
2617 if (BFQQ_SEEKY(bfqq
) || BFQQ_SEEKY(new_bfqq
))
2621 * Interleaved I/O is known to be done by (some) applications
2622 * only for reads, so it does not make sense to merge async
2625 if (!bfq_bfqq_sync(bfqq
) || !bfq_bfqq_sync(new_bfqq
))
2631 static bool idling_boosts_thr_without_issues(struct bfq_data
*bfqd
,
2632 struct bfq_queue
*bfqq
);
2634 static void bfq_put_stable_ref(struct bfq_queue
*bfqq
);
2637 * Attempt to schedule a merge of bfqq with the currently in-service
2638 * queue or with a close queue among the scheduled queues. Return
2639 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2640 * structure otherwise.
2642 * The OOM queue is not allowed to participate to cooperation: in fact, since
2643 * the requests temporarily redirected to the OOM queue could be redirected
2644 * again to dedicated queues at any time, the state needed to correctly
2645 * handle merging with the OOM queue would be quite complex and expensive
2646 * to maintain. Besides, in such a critical condition as an out of memory,
2647 * the benefits of queue merging may be little relevant, or even negligible.
2649 * WARNING: queue merging may impair fairness among non-weight raised
2650 * queues, for at least two reasons: 1) the original weight of a
2651 * merged queue may change during the merged state, 2) even being the
2652 * weight the same, a merged queue may be bloated with many more
2653 * requests than the ones produced by its originally-associated
2656 static struct bfq_queue
*
2657 bfq_setup_cooperator(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
2658 void *io_struct
, bool request
, struct bfq_io_cq
*bic
)
2660 struct bfq_queue
*in_service_bfqq
, *new_bfqq
;
2663 * Check delayed stable merge for rotational or non-queueing
2664 * devs. For this branch to be executed, bfqq must not be
2665 * currently merged with some other queue (i.e., bfqq->bic
2666 * must be non null). If we considered also merged queues,
2667 * then we should also check whether bfqq has already been
2668 * merged with bic->stable_merge_bfqq. But this would be
2669 * costly and complicated.
2671 if (unlikely(!bfqd
->nonrot_with_queueing
)) {
2672 if (bic
->stable_merge_bfqq
&&
2673 !bfq_bfqq_just_created(bfqq
) &&
2674 time_is_after_jiffies(bfqq
->split_time
+
2675 msecs_to_jiffies(200))) {
2676 struct bfq_queue
*stable_merge_bfqq
=
2677 bic
->stable_merge_bfqq
;
2678 int proc_ref
= min(bfqq_process_refs(bfqq
),
2679 bfqq_process_refs(stable_merge_bfqq
));
2681 /* deschedule stable merge, because done or aborted here */
2682 bfq_put_stable_ref(stable_merge_bfqq
);
2684 bic
->stable_merge_bfqq
= NULL
;
2686 if (!idling_boosts_thr_without_issues(bfqd
, bfqq
) &&
2688 /* next function will take at least one ref */
2689 struct bfq_queue
*new_bfqq
=
2690 bfq_setup_merge(bfqq
, stable_merge_bfqq
);
2692 bic
->stably_merged
= true;
2693 if (new_bfqq
&& new_bfqq
->bic
)
2694 new_bfqq
->bic
->stably_merged
= true;
2702 * Do not perform queue merging if the device is non
2703 * rotational and performs internal queueing. In fact, such a
2704 * device reaches a high speed through internal parallelism
2705 * and pipelining. This means that, to reach a high
2706 * throughput, it must have many requests enqueued at the same
2707 * time. But, in this configuration, the internal scheduling
2708 * algorithm of the device does exactly the job of queue
2709 * merging: it reorders requests so as to obtain as much as
2710 * possible a sequential I/O pattern. As a consequence, with
2711 * the workload generated by processes doing interleaved I/O,
2712 * the throughput reached by the device is likely to be the
2713 * same, with and without queue merging.
2715 * Disabling merging also provides a remarkable benefit in
2716 * terms of throughput. Merging tends to make many workloads
2717 * artificially more uneven, because of shared queues
2718 * remaining non empty for incomparably more time than
2719 * non-merged queues. This may accentuate workload
2720 * asymmetries. For example, if one of the queues in a set of
2721 * merged queues has a higher weight than a normal queue, then
2722 * the shared queue may inherit such a high weight and, by
2723 * staying almost always active, may force BFQ to perform I/O
2724 * plugging most of the time. This evidently makes it harder
2725 * for BFQ to let the device reach a high throughput.
2727 * Finally, the likely() macro below is not used because one
2728 * of the two branches is more likely than the other, but to
2729 * have the code path after the following if() executed as
2730 * fast as possible for the case of a non rotational device
2731 * with queueing. We want it because this is the fastest kind
2732 * of device. On the opposite end, the likely() may lengthen
2733 * the execution time of BFQ for the case of slower devices
2734 * (rotational or at least without queueing). But in this case
2735 * the execution time of BFQ matters very little, if not at
2738 if (likely(bfqd
->nonrot_with_queueing
))
2742 * Prevent bfqq from being merged if it has been created too
2743 * long ago. The idea is that true cooperating processes, and
2744 * thus their associated bfq_queues, are supposed to be
2745 * created shortly after each other. This is the case, e.g.,
2746 * for KVM/QEMU and dump I/O threads. Basing on this
2747 * assumption, the following filtering greatly reduces the
2748 * probability that two non-cooperating processes, which just
2749 * happen to do close I/O for some short time interval, have
2750 * their queues merged by mistake.
2752 if (bfq_too_late_for_merging(bfqq
))
2756 return bfqq
->new_bfqq
;
2758 if (!io_struct
|| unlikely(bfqq
== &bfqd
->oom_bfqq
))
2761 /* If there is only one backlogged queue, don't search. */
2762 if (bfq_tot_busy_queues(bfqd
) == 1)
2765 in_service_bfqq
= bfqd
->in_service_queue
;
2767 if (in_service_bfqq
&& in_service_bfqq
!= bfqq
&&
2768 likely(in_service_bfqq
!= &bfqd
->oom_bfqq
) &&
2769 bfq_rq_close_to_sector(io_struct
, request
,
2770 bfqd
->in_serv_last_pos
) &&
2771 bfqq
->entity
.parent
== in_service_bfqq
->entity
.parent
&&
2772 bfq_may_be_close_cooperator(bfqq
, in_service_bfqq
)) {
2773 new_bfqq
= bfq_setup_merge(bfqq
, in_service_bfqq
);
2778 * Check whether there is a cooperator among currently scheduled
2779 * queues. The only thing we need is that the bio/request is not
2780 * NULL, as we need it to establish whether a cooperator exists.
2782 new_bfqq
= bfq_find_close_cooperator(bfqd
, bfqq
,
2783 bfq_io_struct_pos(io_struct
, request
));
2785 if (new_bfqq
&& likely(new_bfqq
!= &bfqd
->oom_bfqq
) &&
2786 bfq_may_be_close_cooperator(bfqq
, new_bfqq
))
2787 return bfq_setup_merge(bfqq
, new_bfqq
);
2792 static void bfq_bfqq_save_state(struct bfq_queue
*bfqq
)
2794 struct bfq_io_cq
*bic
= bfqq
->bic
;
2797 * If !bfqq->bic, the queue is already shared or its requests
2798 * have already been redirected to a shared queue; both idle window
2799 * and weight raising state have already been saved. Do nothing.
2804 bic
->saved_last_serv_time_ns
= bfqq
->last_serv_time_ns
;
2805 bic
->saved_inject_limit
= bfqq
->inject_limit
;
2806 bic
->saved_decrease_time_jif
= bfqq
->decrease_time_jif
;
2808 bic
->saved_weight
= bfqq
->entity
.orig_weight
;
2809 bic
->saved_ttime
= bfqq
->ttime
;
2810 bic
->saved_has_short_ttime
= bfq_bfqq_has_short_ttime(bfqq
);
2811 bic
->saved_IO_bound
= bfq_bfqq_IO_bound(bfqq
);
2812 bic
->saved_io_start_time
= bfqq
->io_start_time
;
2813 bic
->saved_tot_idle_time
= bfqq
->tot_idle_time
;
2814 bic
->saved_in_large_burst
= bfq_bfqq_in_large_burst(bfqq
);
2815 bic
->was_in_burst_list
= !hlist_unhashed(&bfqq
->burst_list_node
);
2816 if (unlikely(bfq_bfqq_just_created(bfqq
) &&
2817 !bfq_bfqq_in_large_burst(bfqq
) &&
2818 bfqq
->bfqd
->low_latency
)) {
2820 * bfqq being merged right after being created: bfqq
2821 * would have deserved interactive weight raising, but
2822 * did not make it to be set in a weight-raised state,
2823 * because of this early merge. Store directly the
2824 * weight-raising state that would have been assigned
2825 * to bfqq, so that to avoid that bfqq unjustly fails
2826 * to enjoy weight raising if split soon.
2828 bic
->saved_wr_coeff
= bfqq
->bfqd
->bfq_wr_coeff
;
2829 bic
->saved_wr_start_at_switch_to_srt
= bfq_smallest_from_now();
2830 bic
->saved_wr_cur_max_time
= bfq_wr_duration(bfqq
->bfqd
);
2831 bic
->saved_last_wr_start_finish
= jiffies
;
2833 bic
->saved_wr_coeff
= bfqq
->wr_coeff
;
2834 bic
->saved_wr_start_at_switch_to_srt
=
2835 bfqq
->wr_start_at_switch_to_srt
;
2836 bic
->saved_service_from_wr
= bfqq
->service_from_wr
;
2837 bic
->saved_last_wr_start_finish
= bfqq
->last_wr_start_finish
;
2838 bic
->saved_wr_cur_max_time
= bfqq
->wr_cur_max_time
;
2844 bfq_reassign_last_bfqq(struct bfq_queue
*cur_bfqq
, struct bfq_queue
*new_bfqq
)
2846 if (cur_bfqq
->entity
.parent
&&
2847 cur_bfqq
->entity
.parent
->last_bfqq_created
== cur_bfqq
)
2848 cur_bfqq
->entity
.parent
->last_bfqq_created
= new_bfqq
;
2849 else if (cur_bfqq
->bfqd
&& cur_bfqq
->bfqd
->last_bfqq_created
== cur_bfqq
)
2850 cur_bfqq
->bfqd
->last_bfqq_created
= new_bfqq
;
2853 void bfq_release_process_ref(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
2856 * To prevent bfqq's service guarantees from being violated,
2857 * bfqq may be left busy, i.e., queued for service, even if
2858 * empty (see comments in __bfq_bfqq_expire() for
2859 * details). But, if no process will send requests to bfqq any
2860 * longer, then there is no point in keeping bfqq queued for
2861 * service. In addition, keeping bfqq queued for service, but
2862 * with no process ref any longer, may have caused bfqq to be
2863 * freed when dequeued from service. But this is assumed to
2866 if (bfq_bfqq_busy(bfqq
) && RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
2867 bfqq
!= bfqd
->in_service_queue
)
2868 bfq_del_bfqq_busy(bfqd
, bfqq
, false);
2870 bfq_reassign_last_bfqq(bfqq
, NULL
);
2872 bfq_put_queue(bfqq
);
2876 bfq_merge_bfqqs(struct bfq_data
*bfqd
, struct bfq_io_cq
*bic
,
2877 struct bfq_queue
*bfqq
, struct bfq_queue
*new_bfqq
)
2879 bfq_log_bfqq(bfqd
, bfqq
, "merging with queue %lu",
2880 (unsigned long)new_bfqq
->pid
);
2881 /* Save weight raising and idle window of the merged queues */
2882 bfq_bfqq_save_state(bfqq
);
2883 bfq_bfqq_save_state(new_bfqq
);
2884 if (bfq_bfqq_IO_bound(bfqq
))
2885 bfq_mark_bfqq_IO_bound(new_bfqq
);
2886 bfq_clear_bfqq_IO_bound(bfqq
);
2889 * The processes associated with bfqq are cooperators of the
2890 * processes associated with new_bfqq. So, if bfqq has a
2891 * waker, then assume that all these processes will be happy
2892 * to let bfqq's waker freely inject I/O when they have no
2895 if (bfqq
->waker_bfqq
&& !new_bfqq
->waker_bfqq
&&
2896 bfqq
->waker_bfqq
!= new_bfqq
) {
2897 new_bfqq
->waker_bfqq
= bfqq
->waker_bfqq
;
2898 new_bfqq
->tentative_waker_bfqq
= NULL
;
2901 * If the waker queue disappears, then
2902 * new_bfqq->waker_bfqq must be reset. So insert
2903 * new_bfqq into the woken_list of the waker. See
2904 * bfq_check_waker for details.
2906 hlist_add_head(&new_bfqq
->woken_list_node
,
2907 &new_bfqq
->waker_bfqq
->woken_list
);
2912 * If bfqq is weight-raised, then let new_bfqq inherit
2913 * weight-raising. To reduce false positives, neglect the case
2914 * where bfqq has just been created, but has not yet made it
2915 * to be weight-raised (which may happen because EQM may merge
2916 * bfqq even before bfq_add_request is executed for the first
2917 * time for bfqq). Handling this case would however be very
2918 * easy, thanks to the flag just_created.
2920 if (new_bfqq
->wr_coeff
== 1 && bfqq
->wr_coeff
> 1) {
2921 new_bfqq
->wr_coeff
= bfqq
->wr_coeff
;
2922 new_bfqq
->wr_cur_max_time
= bfqq
->wr_cur_max_time
;
2923 new_bfqq
->last_wr_start_finish
= bfqq
->last_wr_start_finish
;
2924 new_bfqq
->wr_start_at_switch_to_srt
=
2925 bfqq
->wr_start_at_switch_to_srt
;
2926 if (bfq_bfqq_busy(new_bfqq
))
2927 bfqd
->wr_busy_queues
++;
2928 new_bfqq
->entity
.prio_changed
= 1;
2931 if (bfqq
->wr_coeff
> 1) { /* bfqq has given its wr to new_bfqq */
2933 bfqq
->entity
.prio_changed
= 1;
2934 if (bfq_bfqq_busy(bfqq
))
2935 bfqd
->wr_busy_queues
--;
2938 bfq_log_bfqq(bfqd
, new_bfqq
, "merge_bfqqs: wr_busy %d",
2939 bfqd
->wr_busy_queues
);
2942 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2944 bic_set_bfqq(bic
, new_bfqq
, 1);
2945 bfq_mark_bfqq_coop(new_bfqq
);
2947 * new_bfqq now belongs to at least two bics (it is a shared queue):
2948 * set new_bfqq->bic to NULL. bfqq either:
2949 * - does not belong to any bic any more, and hence bfqq->bic must
2950 * be set to NULL, or
2951 * - is a queue whose owning bics have already been redirected to a
2952 * different queue, hence the queue is destined to not belong to
2953 * any bic soon and bfqq->bic is already NULL (therefore the next
2954 * assignment causes no harm).
2956 new_bfqq
->bic
= NULL
;
2958 * If the queue is shared, the pid is the pid of one of the associated
2959 * processes. Which pid depends on the exact sequence of merge events
2960 * the queue underwent. So printing such a pid is useless and confusing
2961 * because it reports a random pid between those of the associated
2963 * We mark such a queue with a pid -1, and then print SHARED instead of
2964 * a pid in logging messages.
2969 bfq_reassign_last_bfqq(bfqq
, new_bfqq
);
2971 bfq_release_process_ref(bfqd
, bfqq
);
2974 static bool bfq_allow_bio_merge(struct request_queue
*q
, struct request
*rq
,
2977 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
2978 bool is_sync
= op_is_sync(bio
->bi_opf
);
2979 struct bfq_queue
*bfqq
= bfqd
->bio_bfqq
, *new_bfqq
;
2982 * Disallow merge of a sync bio into an async request.
2984 if (is_sync
&& !rq_is_sync(rq
))
2988 * Lookup the bfqq that this bio will be queued with. Allow
2989 * merge only if rq is queued there.
2995 * We take advantage of this function to perform an early merge
2996 * of the queues of possible cooperating processes.
2998 new_bfqq
= bfq_setup_cooperator(bfqd
, bfqq
, bio
, false, bfqd
->bio_bic
);
3001 * bic still points to bfqq, then it has not yet been
3002 * redirected to some other bfq_queue, and a queue
3003 * merge between bfqq and new_bfqq can be safely
3004 * fulfilled, i.e., bic can be redirected to new_bfqq
3005 * and bfqq can be put.
3007 bfq_merge_bfqqs(bfqd
, bfqd
->bio_bic
, bfqq
,
3010 * If we get here, bio will be queued into new_queue,
3011 * so use new_bfqq to decide whether bio and rq can be
3017 * Change also bqfd->bio_bfqq, as
3018 * bfqd->bio_bic now points to new_bfqq, and
3019 * this function may be invoked again (and then may
3020 * use again bqfd->bio_bfqq).
3022 bfqd
->bio_bfqq
= bfqq
;
3025 return bfqq
== RQ_BFQQ(rq
);
3029 * Set the maximum time for the in-service queue to consume its
3030 * budget. This prevents seeky processes from lowering the throughput.
3031 * In practice, a time-slice service scheme is used with seeky
3034 static void bfq_set_budget_timeout(struct bfq_data
*bfqd
,
3035 struct bfq_queue
*bfqq
)
3037 unsigned int timeout_coeff
;
3039 if (bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
)
3042 timeout_coeff
= bfqq
->entity
.weight
/ bfqq
->entity
.orig_weight
;
3044 bfqd
->last_budget_start
= ktime_get();
3046 bfqq
->budget_timeout
= jiffies
+
3047 bfqd
->bfq_timeout
* timeout_coeff
;
3050 static void __bfq_set_in_service_queue(struct bfq_data
*bfqd
,
3051 struct bfq_queue
*bfqq
)
3054 bfq_clear_bfqq_fifo_expire(bfqq
);
3056 bfqd
->budgets_assigned
= (bfqd
->budgets_assigned
* 7 + 256) / 8;
3058 if (time_is_before_jiffies(bfqq
->last_wr_start_finish
) &&
3059 bfqq
->wr_coeff
> 1 &&
3060 bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
&&
3061 time_is_before_jiffies(bfqq
->budget_timeout
)) {
3063 * For soft real-time queues, move the start
3064 * of the weight-raising period forward by the
3065 * time the queue has not received any
3066 * service. Otherwise, a relatively long
3067 * service delay is likely to cause the
3068 * weight-raising period of the queue to end,
3069 * because of the short duration of the
3070 * weight-raising period of a soft real-time
3071 * queue. It is worth noting that this move
3072 * is not so dangerous for the other queues,
3073 * because soft real-time queues are not
3076 * To not add a further variable, we use the
3077 * overloaded field budget_timeout to
3078 * determine for how long the queue has not
3079 * received service, i.e., how much time has
3080 * elapsed since the queue expired. However,
3081 * this is a little imprecise, because
3082 * budget_timeout is set to jiffies if bfqq
3083 * not only expires, but also remains with no
3086 if (time_after(bfqq
->budget_timeout
,
3087 bfqq
->last_wr_start_finish
))
3088 bfqq
->last_wr_start_finish
+=
3089 jiffies
- bfqq
->budget_timeout
;
3091 bfqq
->last_wr_start_finish
= jiffies
;
3094 bfq_set_budget_timeout(bfqd
, bfqq
);
3095 bfq_log_bfqq(bfqd
, bfqq
,
3096 "set_in_service_queue, cur-budget = %d",
3097 bfqq
->entity
.budget
);
3100 bfqd
->in_service_queue
= bfqq
;
3101 bfqd
->in_serv_last_pos
= 0;
3105 * Get and set a new queue for service.
3107 static struct bfq_queue
*bfq_set_in_service_queue(struct bfq_data
*bfqd
)
3109 struct bfq_queue
*bfqq
= bfq_get_next_queue(bfqd
);
3111 __bfq_set_in_service_queue(bfqd
, bfqq
);
3115 static void bfq_arm_slice_timer(struct bfq_data
*bfqd
)
3117 struct bfq_queue
*bfqq
= bfqd
->in_service_queue
;
3120 bfq_mark_bfqq_wait_request(bfqq
);
3123 * We don't want to idle for seeks, but we do want to allow
3124 * fair distribution of slice time for a process doing back-to-back
3125 * seeks. So allow a little bit of time for him to submit a new rq.
3127 sl
= bfqd
->bfq_slice_idle
;
3129 * Unless the queue is being weight-raised or the scenario is
3130 * asymmetric, grant only minimum idle time if the queue
3131 * is seeky. A long idling is preserved for a weight-raised
3132 * queue, or, more in general, in an asymmetric scenario,
3133 * because a long idling is needed for guaranteeing to a queue
3134 * its reserved share of the throughput (in particular, it is
3135 * needed if the queue has a higher weight than some other
3138 if (BFQQ_SEEKY(bfqq
) && bfqq
->wr_coeff
== 1 &&
3139 !bfq_asymmetric_scenario(bfqd
, bfqq
))
3140 sl
= min_t(u64
, sl
, BFQ_MIN_TT
);
3141 else if (bfqq
->wr_coeff
> 1)
3142 sl
= max_t(u32
, sl
, 20ULL * NSEC_PER_MSEC
);
3144 bfqd
->last_idling_start
= ktime_get();
3145 bfqd
->last_idling_start_jiffies
= jiffies
;
3147 hrtimer_start(&bfqd
->idle_slice_timer
, ns_to_ktime(sl
),
3149 bfqg_stats_set_start_idle_time(bfqq_group(bfqq
));
3153 * In autotuning mode, max_budget is dynamically recomputed as the
3154 * amount of sectors transferred in timeout at the estimated peak
3155 * rate. This enables BFQ to utilize a full timeslice with a full
3156 * budget, even if the in-service queue is served at peak rate. And
3157 * this maximises throughput with sequential workloads.
3159 static unsigned long bfq_calc_max_budget(struct bfq_data
*bfqd
)
3161 return (u64
)bfqd
->peak_rate
* USEC_PER_MSEC
*
3162 jiffies_to_msecs(bfqd
->bfq_timeout
)>>BFQ_RATE_SHIFT
;
3166 * Update parameters related to throughput and responsiveness, as a
3167 * function of the estimated peak rate. See comments on
3168 * bfq_calc_max_budget(), and on the ref_wr_duration array.
3170 static void update_thr_responsiveness_params(struct bfq_data
*bfqd
)
3172 if (bfqd
->bfq_user_max_budget
== 0) {
3173 bfqd
->bfq_max_budget
=
3174 bfq_calc_max_budget(bfqd
);
3175 bfq_log(bfqd
, "new max_budget = %d", bfqd
->bfq_max_budget
);
3179 static void bfq_reset_rate_computation(struct bfq_data
*bfqd
,
3182 if (rq
!= NULL
) { /* new rq dispatch now, reset accordingly */
3183 bfqd
->last_dispatch
= bfqd
->first_dispatch
= ktime_get_ns();
3184 bfqd
->peak_rate_samples
= 1;
3185 bfqd
->sequential_samples
= 0;
3186 bfqd
->tot_sectors_dispatched
= bfqd
->last_rq_max_size
=
3188 } else /* no new rq dispatched, just reset the number of samples */
3189 bfqd
->peak_rate_samples
= 0; /* full re-init on next disp. */
3192 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
3193 bfqd
->peak_rate_samples
, bfqd
->sequential_samples
,
3194 bfqd
->tot_sectors_dispatched
);
3197 static void bfq_update_rate_reset(struct bfq_data
*bfqd
, struct request
*rq
)
3199 u32 rate
, weight
, divisor
;
3202 * For the convergence property to hold (see comments on
3203 * bfq_update_peak_rate()) and for the assessment to be
3204 * reliable, a minimum number of samples must be present, and
3205 * a minimum amount of time must have elapsed. If not so, do
3206 * not compute new rate. Just reset parameters, to get ready
3207 * for a new evaluation attempt.
3209 if (bfqd
->peak_rate_samples
< BFQ_RATE_MIN_SAMPLES
||
3210 bfqd
->delta_from_first
< BFQ_RATE_MIN_INTERVAL
)
3211 goto reset_computation
;
3214 * If a new request completion has occurred after last
3215 * dispatch, then, to approximate the rate at which requests
3216 * have been served by the device, it is more precise to
3217 * extend the observation interval to the last completion.
3219 bfqd
->delta_from_first
=
3220 max_t(u64
, bfqd
->delta_from_first
,
3221 bfqd
->last_completion
- bfqd
->first_dispatch
);
3224 * Rate computed in sects/usec, and not sects/nsec, for
3227 rate
= div64_ul(bfqd
->tot_sectors_dispatched
<<BFQ_RATE_SHIFT
,
3228 div_u64(bfqd
->delta_from_first
, NSEC_PER_USEC
));
3231 * Peak rate not updated if:
3232 * - the percentage of sequential dispatches is below 3/4 of the
3233 * total, and rate is below the current estimated peak rate
3234 * - rate is unreasonably high (> 20M sectors/sec)
3236 if ((bfqd
->sequential_samples
< (3 * bfqd
->peak_rate_samples
)>>2 &&
3237 rate
<= bfqd
->peak_rate
) ||
3238 rate
> 20<<BFQ_RATE_SHIFT
)
3239 goto reset_computation
;
3242 * We have to update the peak rate, at last! To this purpose,
3243 * we use a low-pass filter. We compute the smoothing constant
3244 * of the filter as a function of the 'weight' of the new
3247 * As can be seen in next formulas, we define this weight as a
3248 * quantity proportional to how sequential the workload is,
3249 * and to how long the observation time interval is.
3251 * The weight runs from 0 to 8. The maximum value of the
3252 * weight, 8, yields the minimum value for the smoothing
3253 * constant. At this minimum value for the smoothing constant,
3254 * the measured rate contributes for half of the next value of
3255 * the estimated peak rate.
3257 * So, the first step is to compute the weight as a function
3258 * of how sequential the workload is. Note that the weight
3259 * cannot reach 9, because bfqd->sequential_samples cannot
3260 * become equal to bfqd->peak_rate_samples, which, in its
3261 * turn, holds true because bfqd->sequential_samples is not
3262 * incremented for the first sample.
3264 weight
= (9 * bfqd
->sequential_samples
) / bfqd
->peak_rate_samples
;
3267 * Second step: further refine the weight as a function of the
3268 * duration of the observation interval.
3270 weight
= min_t(u32
, 8,
3271 div_u64(weight
* bfqd
->delta_from_first
,
3272 BFQ_RATE_REF_INTERVAL
));
3275 * Divisor ranging from 10, for minimum weight, to 2, for
3278 divisor
= 10 - weight
;
3281 * Finally, update peak rate:
3283 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
3285 bfqd
->peak_rate
*= divisor
-1;
3286 bfqd
->peak_rate
/= divisor
;
3287 rate
/= divisor
; /* smoothing constant alpha = 1/divisor */
3289 bfqd
->peak_rate
+= rate
;
3292 * For a very slow device, bfqd->peak_rate can reach 0 (see
3293 * the minimum representable values reported in the comments
3294 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3295 * divisions by zero where bfqd->peak_rate is used as a
3298 bfqd
->peak_rate
= max_t(u32
, 1, bfqd
->peak_rate
);
3300 update_thr_responsiveness_params(bfqd
);
3303 bfq_reset_rate_computation(bfqd
, rq
);
3307 * Update the read/write peak rate (the main quantity used for
3308 * auto-tuning, see update_thr_responsiveness_params()).
3310 * It is not trivial to estimate the peak rate (correctly): because of
3311 * the presence of sw and hw queues between the scheduler and the
3312 * device components that finally serve I/O requests, it is hard to
3313 * say exactly when a given dispatched request is served inside the
3314 * device, and for how long. As a consequence, it is hard to know
3315 * precisely at what rate a given set of requests is actually served
3318 * On the opposite end, the dispatch time of any request is trivially
3319 * available, and, from this piece of information, the "dispatch rate"
3320 * of requests can be immediately computed. So, the idea in the next
3321 * function is to use what is known, namely request dispatch times
3322 * (plus, when useful, request completion times), to estimate what is
3323 * unknown, namely in-device request service rate.
3325 * The main issue is that, because of the above facts, the rate at
3326 * which a certain set of requests is dispatched over a certain time
3327 * interval can vary greatly with respect to the rate at which the
3328 * same requests are then served. But, since the size of any
3329 * intermediate queue is limited, and the service scheme is lossless
3330 * (no request is silently dropped), the following obvious convergence
3331 * property holds: the number of requests dispatched MUST become
3332 * closer and closer to the number of requests completed as the
3333 * observation interval grows. This is the key property used in
3334 * the next function to estimate the peak service rate as a function
3335 * of the observed dispatch rate. The function assumes to be invoked
3336 * on every request dispatch.
3338 static void bfq_update_peak_rate(struct bfq_data
*bfqd
, struct request
*rq
)
3340 u64 now_ns
= ktime_get_ns();
3342 if (bfqd
->peak_rate_samples
== 0) { /* first dispatch */
3343 bfq_log(bfqd
, "update_peak_rate: goto reset, samples %d",
3344 bfqd
->peak_rate_samples
);
3345 bfq_reset_rate_computation(bfqd
, rq
);
3346 goto update_last_values
; /* will add one sample */
3350 * Device idle for very long: the observation interval lasting
3351 * up to this dispatch cannot be a valid observation interval
3352 * for computing a new peak rate (similarly to the late-
3353 * completion event in bfq_completed_request()). Go to
3354 * update_rate_and_reset to have the following three steps
3356 * - close the observation interval at the last (previous)
3357 * request dispatch or completion
3358 * - compute rate, if possible, for that observation interval
3359 * - start a new observation interval with this dispatch
3361 if (now_ns
- bfqd
->last_dispatch
> 100*NSEC_PER_MSEC
&&
3362 bfqd
->rq_in_driver
== 0)
3363 goto update_rate_and_reset
;
3365 /* Update sampling information */
3366 bfqd
->peak_rate_samples
++;
3368 if ((bfqd
->rq_in_driver
> 0 ||
3369 now_ns
- bfqd
->last_completion
< BFQ_MIN_TT
)
3370 && !BFQ_RQ_SEEKY(bfqd
, bfqd
->last_position
, rq
))
3371 bfqd
->sequential_samples
++;
3373 bfqd
->tot_sectors_dispatched
+= blk_rq_sectors(rq
);
3375 /* Reset max observed rq size every 32 dispatches */
3376 if (likely(bfqd
->peak_rate_samples
% 32))
3377 bfqd
->last_rq_max_size
=
3378 max_t(u32
, blk_rq_sectors(rq
), bfqd
->last_rq_max_size
);
3380 bfqd
->last_rq_max_size
= blk_rq_sectors(rq
);
3382 bfqd
->delta_from_first
= now_ns
- bfqd
->first_dispatch
;
3384 /* Target observation interval not yet reached, go on sampling */
3385 if (bfqd
->delta_from_first
< BFQ_RATE_REF_INTERVAL
)
3386 goto update_last_values
;
3388 update_rate_and_reset
:
3389 bfq_update_rate_reset(bfqd
, rq
);
3391 bfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
3392 if (RQ_BFQQ(rq
) == bfqd
->in_service_queue
)
3393 bfqd
->in_serv_last_pos
= bfqd
->last_position
;
3394 bfqd
->last_dispatch
= now_ns
;
3398 * Remove request from internal lists.
3400 static void bfq_dispatch_remove(struct request_queue
*q
, struct request
*rq
)
3402 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
3405 * For consistency, the next instruction should have been
3406 * executed after removing the request from the queue and
3407 * dispatching it. We execute instead this instruction before
3408 * bfq_remove_request() (and hence introduce a temporary
3409 * inconsistency), for efficiency. In fact, should this
3410 * dispatch occur for a non in-service bfqq, this anticipated
3411 * increment prevents two counters related to bfqq->dispatched
3412 * from risking to be, first, uselessly decremented, and then
3413 * incremented again when the (new) value of bfqq->dispatched
3414 * happens to be taken into account.
3417 bfq_update_peak_rate(q
->elevator
->elevator_data
, rq
);
3419 bfq_remove_request(q
, rq
);
3423 * There is a case where idling does not have to be performed for
3424 * throughput concerns, but to preserve the throughput share of
3425 * the process associated with bfqq.
3427 * To introduce this case, we can note that allowing the drive
3428 * to enqueue more than one request at a time, and hence
3429 * delegating de facto final scheduling decisions to the
3430 * drive's internal scheduler, entails loss of control on the
3431 * actual request service order. In particular, the critical
3432 * situation is when requests from different processes happen
3433 * to be present, at the same time, in the internal queue(s)
3434 * of the drive. In such a situation, the drive, by deciding
3435 * the service order of the internally-queued requests, does
3436 * determine also the actual throughput distribution among
3437 * these processes. But the drive typically has no notion or
3438 * concern about per-process throughput distribution, and
3439 * makes its decisions only on a per-request basis. Therefore,
3440 * the service distribution enforced by the drive's internal
3441 * scheduler is likely to coincide with the desired throughput
3442 * distribution only in a completely symmetric, or favorably
3443 * skewed scenario where:
3444 * (i-a) each of these processes must get the same throughput as
3446 * (i-b) in case (i-a) does not hold, it holds that the process
3447 * associated with bfqq must receive a lower or equal
3448 * throughput than any of the other processes;
3449 * (ii) the I/O of each process has the same properties, in
3450 * terms of locality (sequential or random), direction
3451 * (reads or writes), request sizes, greediness
3452 * (from I/O-bound to sporadic), and so on;
3454 * In fact, in such a scenario, the drive tends to treat the requests
3455 * of each process in about the same way as the requests of the
3456 * others, and thus to provide each of these processes with about the
3457 * same throughput. This is exactly the desired throughput
3458 * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3459 * even more convenient distribution for (the process associated with)
3462 * In contrast, in any asymmetric or unfavorable scenario, device
3463 * idling (I/O-dispatch plugging) is certainly needed to guarantee
3464 * that bfqq receives its assigned fraction of the device throughput
3465 * (see [1] for details).
3467 * The problem is that idling may significantly reduce throughput with
3468 * certain combinations of types of I/O and devices. An important
3469 * example is sync random I/O on flash storage with command
3470 * queueing. So, unless bfqq falls in cases where idling also boosts
3471 * throughput, it is important to check conditions (i-a), i(-b) and
3472 * (ii) accurately, so as to avoid idling when not strictly needed for
3473 * service guarantees.
3475 * Unfortunately, it is extremely difficult to thoroughly check
3476 * condition (ii). And, in case there are active groups, it becomes
3477 * very difficult to check conditions (i-a) and (i-b) too. In fact,
3478 * if there are active groups, then, for conditions (i-a) or (i-b) to
3479 * become false 'indirectly', it is enough that an active group
3480 * contains more active processes or sub-groups than some other active
3481 * group. More precisely, for conditions (i-a) or (i-b) to become
3482 * false because of such a group, it is not even necessary that the
3483 * group is (still) active: it is sufficient that, even if the group
3484 * has become inactive, some of its descendant processes still have
3485 * some request already dispatched but still waiting for
3486 * completion. In fact, requests have still to be guaranteed their
3487 * share of the throughput even after being dispatched. In this
3488 * respect, it is easy to show that, if a group frequently becomes
3489 * inactive while still having in-flight requests, and if, when this
3490 * happens, the group is not considered in the calculation of whether
3491 * the scenario is asymmetric, then the group may fail to be
3492 * guaranteed its fair share of the throughput (basically because
3493 * idling may not be performed for the descendant processes of the
3494 * group, but it had to be). We address this issue with the following
3495 * bi-modal behavior, implemented in the function
3496 * bfq_asymmetric_scenario().
3498 * If there are groups with requests waiting for completion
3499 * (as commented above, some of these groups may even be
3500 * already inactive), then the scenario is tagged as
3501 * asymmetric, conservatively, without checking any of the
3502 * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3503 * This behavior matches also the fact that groups are created
3504 * exactly if controlling I/O is a primary concern (to
3505 * preserve bandwidth and latency guarantees).
3507 * On the opposite end, if there are no groups with requests waiting
3508 * for completion, then only conditions (i-a) and (i-b) are actually
3509 * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3510 * idling is not performed, regardless of whether condition (ii)
3511 * holds. In other words, only if conditions (i-a) and (i-b) do not
3512 * hold, then idling is allowed, and the device tends to be prevented
3513 * from queueing many requests, possibly of several processes. Since
3514 * there are no groups with requests waiting for completion, then, to
3515 * control conditions (i-a) and (i-b) it is enough to check just
3516 * whether all the queues with requests waiting for completion also
3517 * have the same weight.
3519 * Not checking condition (ii) evidently exposes bfqq to the
3520 * risk of getting less throughput than its fair share.
3521 * However, for queues with the same weight, a further
3522 * mechanism, preemption, mitigates or even eliminates this
3523 * problem. And it does so without consequences on overall
3524 * throughput. This mechanism and its benefits are explained
3525 * in the next three paragraphs.
3527 * Even if a queue, say Q, is expired when it remains idle, Q
3528 * can still preempt the new in-service queue if the next
3529 * request of Q arrives soon (see the comments on
3530 * bfq_bfqq_update_budg_for_activation). If all queues and
3531 * groups have the same weight, this form of preemption,
3532 * combined with the hole-recovery heuristic described in the
3533 * comments on function bfq_bfqq_update_budg_for_activation,
3534 * are enough to preserve a correct bandwidth distribution in
3535 * the mid term, even without idling. In fact, even if not
3536 * idling allows the internal queues of the device to contain
3537 * many requests, and thus to reorder requests, we can rather
3538 * safely assume that the internal scheduler still preserves a
3539 * minimum of mid-term fairness.
3541 * More precisely, this preemption-based, idleless approach
3542 * provides fairness in terms of IOPS, and not sectors per
3543 * second. This can be seen with a simple example. Suppose
3544 * that there are two queues with the same weight, but that
3545 * the first queue receives requests of 8 sectors, while the
3546 * second queue receives requests of 1024 sectors. In
3547 * addition, suppose that each of the two queues contains at
3548 * most one request at a time, which implies that each queue
3549 * always remains idle after it is served. Finally, after
3550 * remaining idle, each queue receives very quickly a new
3551 * request. It follows that the two queues are served
3552 * alternatively, preempting each other if needed. This
3553 * implies that, although both queues have the same weight,
3554 * the queue with large requests receives a service that is
3555 * 1024/8 times as high as the service received by the other
3558 * The motivation for using preemption instead of idling (for
3559 * queues with the same weight) is that, by not idling,
3560 * service guarantees are preserved (completely or at least in
3561 * part) without minimally sacrificing throughput. And, if
3562 * there is no active group, then the primary expectation for
3563 * this device is probably a high throughput.
3565 * We are now left only with explaining the two sub-conditions in the
3566 * additional compound condition that is checked below for deciding
3567 * whether the scenario is asymmetric. To explain the first
3568 * sub-condition, we need to add that the function
3569 * bfq_asymmetric_scenario checks the weights of only
3570 * non-weight-raised queues, for efficiency reasons (see comments on
3571 * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3572 * is checked explicitly here. More precisely, the compound condition
3573 * below takes into account also the fact that, even if bfqq is being
3574 * weight-raised, the scenario is still symmetric if all queues with
3575 * requests waiting for completion happen to be
3576 * weight-raised. Actually, we should be even more precise here, and
3577 * differentiate between interactive weight raising and soft real-time
3580 * The second sub-condition checked in the compound condition is
3581 * whether there is a fair amount of already in-flight I/O not
3582 * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3583 * following reason. The drive may decide to serve in-flight
3584 * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3585 * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3586 * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3587 * basically uncontrolled amount of I/O from other queues may be
3588 * dispatched too, possibly causing the service of bfqq's I/O to be
3589 * delayed even longer in the drive. This problem gets more and more
3590 * serious as the speed and the queue depth of the drive grow,
3591 * because, as these two quantities grow, the probability to find no
3592 * queue busy but many requests in flight grows too. By contrast,
3593 * plugging I/O dispatching minimizes the delay induced by already
3594 * in-flight I/O, and enables bfqq to recover the bandwidth it may
3595 * lose because of this delay.
3597 * As a side note, it is worth considering that the above
3598 * device-idling countermeasures may however fail in the following
3599 * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3600 * in a time period during which all symmetry sub-conditions hold, and
3601 * therefore the device is allowed to enqueue many requests, but at
3602 * some later point in time some sub-condition stops to hold, then it
3603 * may become impossible to make requests be served in the desired
3604 * order until all the requests already queued in the device have been
3605 * served. The last sub-condition commented above somewhat mitigates
3606 * this problem for weight-raised queues.
3608 * However, as an additional mitigation for this problem, we preserve
3609 * plugging for a special symmetric case that may suddenly turn into
3610 * asymmetric: the case where only bfqq is busy. In this case, not
3611 * expiring bfqq does not cause any harm to any other queues in terms
3612 * of service guarantees. In contrast, it avoids the following unlucky
3613 * sequence of events: (1) bfqq is expired, (2) a new queue with a
3614 * lower weight than bfqq becomes busy (or more queues), (3) the new
3615 * queue is served until a new request arrives for bfqq, (4) when bfqq
3616 * is finally served, there are so many requests of the new queue in
3617 * the drive that the pending requests for bfqq take a lot of time to
3618 * be served. In particular, event (2) may case even already
3619 * dispatched requests of bfqq to be delayed, inside the drive. So, to
3620 * avoid this series of events, the scenario is preventively declared
3621 * as asymmetric also if bfqq is the only busy queues
3623 static bool idling_needed_for_service_guarantees(struct bfq_data
*bfqd
,
3624 struct bfq_queue
*bfqq
)
3626 int tot_busy_queues
= bfq_tot_busy_queues(bfqd
);
3628 /* No point in idling for bfqq if it won't get requests any longer */
3629 if (unlikely(!bfqq_process_refs(bfqq
)))
3632 return (bfqq
->wr_coeff
> 1 &&
3633 (bfqd
->wr_busy_queues
<
3635 bfqd
->rq_in_driver
>=
3636 bfqq
->dispatched
+ 4)) ||
3637 bfq_asymmetric_scenario(bfqd
, bfqq
) ||
3638 tot_busy_queues
== 1;
3641 static bool __bfq_bfqq_expire(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
3642 enum bfqq_expiration reason
)
3645 * If this bfqq is shared between multiple processes, check
3646 * to make sure that those processes are still issuing I/Os
3647 * within the mean seek distance. If not, it may be time to
3648 * break the queues apart again.
3650 if (bfq_bfqq_coop(bfqq
) && BFQQ_SEEKY(bfqq
))
3651 bfq_mark_bfqq_split_coop(bfqq
);
3654 * Consider queues with a higher finish virtual time than
3655 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3656 * true, then bfqq's bandwidth would be violated if an
3657 * uncontrolled amount of I/O from these queues were
3658 * dispatched while bfqq is waiting for its new I/O to
3659 * arrive. This is exactly what may happen if this is a forced
3660 * expiration caused by a preemption attempt, and if bfqq is
3661 * not re-scheduled. To prevent this from happening, re-queue
3662 * bfqq if it needs I/O-dispatch plugging, even if it is
3663 * empty. By doing so, bfqq is granted to be served before the
3664 * above queues (provided that bfqq is of course eligible).
3666 if (RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
3667 !(reason
== BFQQE_PREEMPTED
&&
3668 idling_needed_for_service_guarantees(bfqd
, bfqq
))) {
3669 if (bfqq
->dispatched
== 0)
3671 * Overloading budget_timeout field to store
3672 * the time at which the queue remains with no
3673 * backlog and no outstanding request; used by
3674 * the weight-raising mechanism.
3676 bfqq
->budget_timeout
= jiffies
;
3678 bfq_del_bfqq_busy(bfqd
, bfqq
, true);
3680 bfq_requeue_bfqq(bfqd
, bfqq
, true);
3682 * Resort priority tree of potential close cooperators.
3683 * See comments on bfq_pos_tree_add_move() for the unlikely().
3685 if (unlikely(!bfqd
->nonrot_with_queueing
&&
3686 !RB_EMPTY_ROOT(&bfqq
->sort_list
)))
3687 bfq_pos_tree_add_move(bfqd
, bfqq
);
3691 * All in-service entities must have been properly deactivated
3692 * or requeued before executing the next function, which
3693 * resets all in-service entities as no more in service. This
3694 * may cause bfqq to be freed. If this happens, the next
3695 * function returns true.
3697 return __bfq_bfqd_reset_in_service(bfqd
);
3701 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3702 * @bfqd: device data.
3703 * @bfqq: queue to update.
3704 * @reason: reason for expiration.
3706 * Handle the feedback on @bfqq budget at queue expiration.
3707 * See the body for detailed comments.
3709 static void __bfq_bfqq_recalc_budget(struct bfq_data
*bfqd
,
3710 struct bfq_queue
*bfqq
,
3711 enum bfqq_expiration reason
)
3713 struct request
*next_rq
;
3714 int budget
, min_budget
;
3716 min_budget
= bfq_min_budget(bfqd
);
3718 if (bfqq
->wr_coeff
== 1)
3719 budget
= bfqq
->max_budget
;
3721 * Use a constant, low budget for weight-raised queues,
3722 * to help achieve a low latency. Keep it slightly higher
3723 * than the minimum possible budget, to cause a little
3724 * bit fewer expirations.
3726 budget
= 2 * min_budget
;
3728 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: last budg %d, budg left %d",
3729 bfqq
->entity
.budget
, bfq_bfqq_budget_left(bfqq
));
3730 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: last max_budg %d, min budg %d",
3731 budget
, bfq_min_budget(bfqd
));
3732 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: sync %d, seeky %d",
3733 bfq_bfqq_sync(bfqq
), BFQQ_SEEKY(bfqd
->in_service_queue
));
3735 if (bfq_bfqq_sync(bfqq
) && bfqq
->wr_coeff
== 1) {
3738 * Caveat: in all the following cases we trade latency
3741 case BFQQE_TOO_IDLE
:
3743 * This is the only case where we may reduce
3744 * the budget: if there is no request of the
3745 * process still waiting for completion, then
3746 * we assume (tentatively) that the timer has
3747 * expired because the batch of requests of
3748 * the process could have been served with a
3749 * smaller budget. Hence, betting that
3750 * process will behave in the same way when it
3751 * becomes backlogged again, we reduce its
3752 * next budget. As long as we guess right,
3753 * this budget cut reduces the latency
3754 * experienced by the process.
3756 * However, if there are still outstanding
3757 * requests, then the process may have not yet
3758 * issued its next request just because it is
3759 * still waiting for the completion of some of
3760 * the still outstanding ones. So in this
3761 * subcase we do not reduce its budget, on the
3762 * contrary we increase it to possibly boost
3763 * the throughput, as discussed in the
3764 * comments to the BUDGET_TIMEOUT case.
3766 if (bfqq
->dispatched
> 0) /* still outstanding reqs */
3767 budget
= min(budget
* 2, bfqd
->bfq_max_budget
);
3769 if (budget
> 5 * min_budget
)
3770 budget
-= 4 * min_budget
;
3772 budget
= min_budget
;
3775 case BFQQE_BUDGET_TIMEOUT
:
3777 * We double the budget here because it gives
3778 * the chance to boost the throughput if this
3779 * is not a seeky process (and has bumped into
3780 * this timeout because of, e.g., ZBR).
3782 budget
= min(budget
* 2, bfqd
->bfq_max_budget
);
3784 case BFQQE_BUDGET_EXHAUSTED
:
3786 * The process still has backlog, and did not
3787 * let either the budget timeout or the disk
3788 * idling timeout expire. Hence it is not
3789 * seeky, has a short thinktime and may be
3790 * happy with a higher budget too. So
3791 * definitely increase the budget of this good
3792 * candidate to boost the disk throughput.
3794 budget
= min(budget
* 4, bfqd
->bfq_max_budget
);
3796 case BFQQE_NO_MORE_REQUESTS
:
3798 * For queues that expire for this reason, it
3799 * is particularly important to keep the
3800 * budget close to the actual service they
3801 * need. Doing so reduces the timestamp
3802 * misalignment problem described in the
3803 * comments in the body of
3804 * __bfq_activate_entity. In fact, suppose
3805 * that a queue systematically expires for
3806 * BFQQE_NO_MORE_REQUESTS and presents a
3807 * new request in time to enjoy timestamp
3808 * back-shifting. The larger the budget of the
3809 * queue is with respect to the service the
3810 * queue actually requests in each service
3811 * slot, the more times the queue can be
3812 * reactivated with the same virtual finish
3813 * time. It follows that, even if this finish
3814 * time is pushed to the system virtual time
3815 * to reduce the consequent timestamp
3816 * misalignment, the queue unjustly enjoys for
3817 * many re-activations a lower finish time
3818 * than all newly activated queues.
3820 * The service needed by bfqq is measured
3821 * quite precisely by bfqq->entity.service.
3822 * Since bfqq does not enjoy device idling,
3823 * bfqq->entity.service is equal to the number
3824 * of sectors that the process associated with
3825 * bfqq requested to read/write before waiting
3826 * for request completions, or blocking for
3829 budget
= max_t(int, bfqq
->entity
.service
, min_budget
);
3834 } else if (!bfq_bfqq_sync(bfqq
)) {
3836 * Async queues get always the maximum possible
3837 * budget, as for them we do not care about latency
3838 * (in addition, their ability to dispatch is limited
3839 * by the charging factor).
3841 budget
= bfqd
->bfq_max_budget
;
3844 bfqq
->max_budget
= budget
;
3846 if (bfqd
->budgets_assigned
>= bfq_stats_min_budgets
&&
3847 !bfqd
->bfq_user_max_budget
)
3848 bfqq
->max_budget
= min(bfqq
->max_budget
, bfqd
->bfq_max_budget
);
3851 * If there is still backlog, then assign a new budget, making
3852 * sure that it is large enough for the next request. Since
3853 * the finish time of bfqq must be kept in sync with the
3854 * budget, be sure to call __bfq_bfqq_expire() *after* this
3857 * If there is no backlog, then no need to update the budget;
3858 * it will be updated on the arrival of a new request.
3860 next_rq
= bfqq
->next_rq
;
3862 bfqq
->entity
.budget
= max_t(unsigned long, bfqq
->max_budget
,
3863 bfq_serv_to_charge(next_rq
, bfqq
));
3865 bfq_log_bfqq(bfqd
, bfqq
, "head sect: %u, new budget %d",
3866 next_rq
? blk_rq_sectors(next_rq
) : 0,
3867 bfqq
->entity
.budget
);
3871 * Return true if the process associated with bfqq is "slow". The slow
3872 * flag is used, in addition to the budget timeout, to reduce the
3873 * amount of service provided to seeky processes, and thus reduce
3874 * their chances to lower the throughput. More details in the comments
3875 * on the function bfq_bfqq_expire().
3877 * An important observation is in order: as discussed in the comments
3878 * on the function bfq_update_peak_rate(), with devices with internal
3879 * queues, it is hard if ever possible to know when and for how long
3880 * an I/O request is processed by the device (apart from the trivial
3881 * I/O pattern where a new request is dispatched only after the
3882 * previous one has been completed). This makes it hard to evaluate
3883 * the real rate at which the I/O requests of each bfq_queue are
3884 * served. In fact, for an I/O scheduler like BFQ, serving a
3885 * bfq_queue means just dispatching its requests during its service
3886 * slot (i.e., until the budget of the queue is exhausted, or the
3887 * queue remains idle, or, finally, a timeout fires). But, during the
3888 * service slot of a bfq_queue, around 100 ms at most, the device may
3889 * be even still processing requests of bfq_queues served in previous
3890 * service slots. On the opposite end, the requests of the in-service
3891 * bfq_queue may be completed after the service slot of the queue
3894 * Anyway, unless more sophisticated solutions are used
3895 * (where possible), the sum of the sizes of the requests dispatched
3896 * during the service slot of a bfq_queue is probably the only
3897 * approximation available for the service received by the bfq_queue
3898 * during its service slot. And this sum is the quantity used in this
3899 * function to evaluate the I/O speed of a process.
3901 static bool bfq_bfqq_is_slow(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
3902 bool compensate
, enum bfqq_expiration reason
,
3903 unsigned long *delta_ms
)
3905 ktime_t delta_ktime
;
3907 bool slow
= BFQQ_SEEKY(bfqq
); /* if delta too short, use seekyness */
3909 if (!bfq_bfqq_sync(bfqq
))
3913 delta_ktime
= bfqd
->last_idling_start
;
3915 delta_ktime
= ktime_get();
3916 delta_ktime
= ktime_sub(delta_ktime
, bfqd
->last_budget_start
);
3917 delta_usecs
= ktime_to_us(delta_ktime
);
3919 /* don't use too short time intervals */
3920 if (delta_usecs
< 1000) {
3921 if (blk_queue_nonrot(bfqd
->queue
))
3923 * give same worst-case guarantees as idling
3926 *delta_ms
= BFQ_MIN_TT
/ NSEC_PER_MSEC
;
3927 else /* charge at least one seek */
3928 *delta_ms
= bfq_slice_idle
/ NSEC_PER_MSEC
;
3933 *delta_ms
= delta_usecs
/ USEC_PER_MSEC
;
3936 * Use only long (> 20ms) intervals to filter out excessive
3937 * spikes in service rate estimation.
3939 if (delta_usecs
> 20000) {
3941 * Caveat for rotational devices: processes doing I/O
3942 * in the slower disk zones tend to be slow(er) even
3943 * if not seeky. In this respect, the estimated peak
3944 * rate is likely to be an average over the disk
3945 * surface. Accordingly, to not be too harsh with
3946 * unlucky processes, a process is deemed slow only if
3947 * its rate has been lower than half of the estimated
3950 slow
= bfqq
->entity
.service
< bfqd
->bfq_max_budget
/ 2;
3953 bfq_log_bfqq(bfqd
, bfqq
, "bfq_bfqq_is_slow: slow %d", slow
);
3959 * To be deemed as soft real-time, an application must meet two
3960 * requirements. First, the application must not require an average
3961 * bandwidth higher than the approximate bandwidth required to playback or
3962 * record a compressed high-definition video.
3963 * The next function is invoked on the completion of the last request of a
3964 * batch, to compute the next-start time instant, soft_rt_next_start, such
3965 * that, if the next request of the application does not arrive before
3966 * soft_rt_next_start, then the above requirement on the bandwidth is met.
3968 * The second requirement is that the request pattern of the application is
3969 * isochronous, i.e., that, after issuing a request or a batch of requests,
3970 * the application stops issuing new requests until all its pending requests
3971 * have been completed. After that, the application may issue a new batch,
3973 * For this reason the next function is invoked to compute
3974 * soft_rt_next_start only for applications that meet this requirement,
3975 * whereas soft_rt_next_start is set to infinity for applications that do
3978 * Unfortunately, even a greedy (i.e., I/O-bound) application may
3979 * happen to meet, occasionally or systematically, both the above
3980 * bandwidth and isochrony requirements. This may happen at least in
3981 * the following circumstances. First, if the CPU load is high. The
3982 * application may stop issuing requests while the CPUs are busy
3983 * serving other processes, then restart, then stop again for a while,
3984 * and so on. The other circumstances are related to the storage
3985 * device: the storage device is highly loaded or reaches a low-enough
3986 * throughput with the I/O of the application (e.g., because the I/O
3987 * is random and/or the device is slow). In all these cases, the
3988 * I/O of the application may be simply slowed down enough to meet
3989 * the bandwidth and isochrony requirements. To reduce the probability
3990 * that greedy applications are deemed as soft real-time in these
3991 * corner cases, a further rule is used in the computation of
3992 * soft_rt_next_start: the return value of this function is forced to
3993 * be higher than the maximum between the following two quantities.
3995 * (a) Current time plus: (1) the maximum time for which the arrival
3996 * of a request is waited for when a sync queue becomes idle,
3997 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
3998 * postpone for a moment the reason for adding a few extra
3999 * jiffies; we get back to it after next item (b). Lower-bounding
4000 * the return value of this function with the current time plus
4001 * bfqd->bfq_slice_idle tends to filter out greedy applications,
4002 * because the latter issue their next request as soon as possible
4003 * after the last one has been completed. In contrast, a soft
4004 * real-time application spends some time processing data, after a
4005 * batch of its requests has been completed.
4007 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
4008 * above, greedy applications may happen to meet both the
4009 * bandwidth and isochrony requirements under heavy CPU or
4010 * storage-device load. In more detail, in these scenarios, these
4011 * applications happen, only for limited time periods, to do I/O
4012 * slowly enough to meet all the requirements described so far,
4013 * including the filtering in above item (a). These slow-speed
4014 * time intervals are usually interspersed between other time
4015 * intervals during which these applications do I/O at a very high
4016 * speed. Fortunately, exactly because of the high speed of the
4017 * I/O in the high-speed intervals, the values returned by this
4018 * function happen to be so high, near the end of any such
4019 * high-speed interval, to be likely to fall *after* the end of
4020 * the low-speed time interval that follows. These high values are
4021 * stored in bfqq->soft_rt_next_start after each invocation of
4022 * this function. As a consequence, if the last value of
4023 * bfqq->soft_rt_next_start is constantly used to lower-bound the
4024 * next value that this function may return, then, from the very
4025 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
4026 * likely to be constantly kept so high that any I/O request
4027 * issued during the low-speed interval is considered as arriving
4028 * to soon for the application to be deemed as soft
4029 * real-time. Then, in the high-speed interval that follows, the
4030 * application will not be deemed as soft real-time, just because
4031 * it will do I/O at a high speed. And so on.
4033 * Getting back to the filtering in item (a), in the following two
4034 * cases this filtering might be easily passed by a greedy
4035 * application, if the reference quantity was just
4036 * bfqd->bfq_slice_idle:
4037 * 1) HZ is so low that the duration of a jiffy is comparable to or
4038 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
4039 * devices with HZ=100. The time granularity may be so coarse
4040 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
4041 * is rather lower than the exact value.
4042 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
4043 * for a while, then suddenly 'jump' by several units to recover the lost
4044 * increments. This seems to happen, e.g., inside virtual machines.
4045 * To address this issue, in the filtering in (a) we do not use as a
4046 * reference time interval just bfqd->bfq_slice_idle, but
4047 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
4048 * minimum number of jiffies for which the filter seems to be quite
4049 * precise also in embedded systems and KVM/QEMU virtual machines.
4051 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data
*bfqd
,
4052 struct bfq_queue
*bfqq
)
4054 return max3(bfqq
->soft_rt_next_start
,
4055 bfqq
->last_idle_bklogged
+
4056 HZ
* bfqq
->service_from_backlogged
/
4057 bfqd
->bfq_wr_max_softrt_rate
,
4058 jiffies
+ nsecs_to_jiffies(bfqq
->bfqd
->bfq_slice_idle
) + 4);
4062 * bfq_bfqq_expire - expire a queue.
4063 * @bfqd: device owning the queue.
4064 * @bfqq: the queue to expire.
4065 * @compensate: if true, compensate for the time spent idling.
4066 * @reason: the reason causing the expiration.
4068 * If the process associated with bfqq does slow I/O (e.g., because it
4069 * issues random requests), we charge bfqq with the time it has been
4070 * in service instead of the service it has received (see
4071 * bfq_bfqq_charge_time for details on how this goal is achieved). As
4072 * a consequence, bfqq will typically get higher timestamps upon
4073 * reactivation, and hence it will be rescheduled as if it had
4074 * received more service than what it has actually received. In the
4075 * end, bfqq receives less service in proportion to how slowly its
4076 * associated process consumes its budgets (and hence how seriously it
4077 * tends to lower the throughput). In addition, this time-charging
4078 * strategy guarantees time fairness among slow processes. In
4079 * contrast, if the process associated with bfqq is not slow, we
4080 * charge bfqq exactly with the service it has received.
4082 * Charging time to the first type of queues and the exact service to
4083 * the other has the effect of using the WF2Q+ policy to schedule the
4084 * former on a timeslice basis, without violating service domain
4085 * guarantees among the latter.
4087 void bfq_bfqq_expire(struct bfq_data
*bfqd
,
4088 struct bfq_queue
*bfqq
,
4090 enum bfqq_expiration reason
)
4093 unsigned long delta
= 0;
4094 struct bfq_entity
*entity
= &bfqq
->entity
;
4097 * Check whether the process is slow (see bfq_bfqq_is_slow).
4099 slow
= bfq_bfqq_is_slow(bfqd
, bfqq
, compensate
, reason
, &delta
);
4102 * As above explained, charge slow (typically seeky) and
4103 * timed-out queues with the time and not the service
4104 * received, to favor sequential workloads.
4106 * Processes doing I/O in the slower disk zones will tend to
4107 * be slow(er) even if not seeky. Therefore, since the
4108 * estimated peak rate is actually an average over the disk
4109 * surface, these processes may timeout just for bad luck. To
4110 * avoid punishing them, do not charge time to processes that
4111 * succeeded in consuming at least 2/3 of their budget. This
4112 * allows BFQ to preserve enough elasticity to still perform
4113 * bandwidth, and not time, distribution with little unlucky
4114 * or quasi-sequential processes.
4116 if (bfqq
->wr_coeff
== 1 &&
4118 (reason
== BFQQE_BUDGET_TIMEOUT
&&
4119 bfq_bfqq_budget_left(bfqq
) >= entity
->budget
/ 3)))
4120 bfq_bfqq_charge_time(bfqd
, bfqq
, delta
);
4122 if (bfqd
->low_latency
&& bfqq
->wr_coeff
== 1)
4123 bfqq
->last_wr_start_finish
= jiffies
;
4125 if (bfqd
->low_latency
&& bfqd
->bfq_wr_max_softrt_rate
> 0 &&
4126 RB_EMPTY_ROOT(&bfqq
->sort_list
)) {
4128 * If we get here, and there are no outstanding
4129 * requests, then the request pattern is isochronous
4130 * (see the comments on the function
4131 * bfq_bfqq_softrt_next_start()). Therefore we can
4132 * compute soft_rt_next_start.
4134 * If, instead, the queue still has outstanding
4135 * requests, then we have to wait for the completion
4136 * of all the outstanding requests to discover whether
4137 * the request pattern is actually isochronous.
4139 if (bfqq
->dispatched
== 0)
4140 bfqq
->soft_rt_next_start
=
4141 bfq_bfqq_softrt_next_start(bfqd
, bfqq
);
4142 else if (bfqq
->dispatched
> 0) {
4144 * Schedule an update of soft_rt_next_start to when
4145 * the task may be discovered to be isochronous.
4147 bfq_mark_bfqq_softrt_update(bfqq
);
4151 bfq_log_bfqq(bfqd
, bfqq
,
4152 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason
,
4153 slow
, bfqq
->dispatched
, bfq_bfqq_has_short_ttime(bfqq
));
4156 * bfqq expired, so no total service time needs to be computed
4157 * any longer: reset state machine for measuring total service
4160 bfqd
->rqs_injected
= bfqd
->wait_dispatch
= false;
4161 bfqd
->waited_rq
= NULL
;
4164 * Increase, decrease or leave budget unchanged according to
4167 __bfq_bfqq_recalc_budget(bfqd
, bfqq
, reason
);
4168 if (__bfq_bfqq_expire(bfqd
, bfqq
, reason
))
4169 /* bfqq is gone, no more actions on it */
4172 /* mark bfqq as waiting a request only if a bic still points to it */
4173 if (!bfq_bfqq_busy(bfqq
) &&
4174 reason
!= BFQQE_BUDGET_TIMEOUT
&&
4175 reason
!= BFQQE_BUDGET_EXHAUSTED
) {
4176 bfq_mark_bfqq_non_blocking_wait_rq(bfqq
);
4178 * Not setting service to 0, because, if the next rq
4179 * arrives in time, the queue will go on receiving
4180 * service with this same budget (as if it never expired)
4183 entity
->service
= 0;
4186 * Reset the received-service counter for every parent entity.
4187 * Differently from what happens with bfqq->entity.service,
4188 * the resetting of this counter never needs to be postponed
4189 * for parent entities. In fact, in case bfqq may have a
4190 * chance to go on being served using the last, partially
4191 * consumed budget, bfqq->entity.service needs to be kept,
4192 * because if bfqq then actually goes on being served using
4193 * the same budget, the last value of bfqq->entity.service is
4194 * needed to properly decrement bfqq->entity.budget by the
4195 * portion already consumed. In contrast, it is not necessary
4196 * to keep entity->service for parent entities too, because
4197 * the bubble up of the new value of bfqq->entity.budget will
4198 * make sure that the budgets of parent entities are correct,
4199 * even in case bfqq and thus parent entities go on receiving
4200 * service with the same budget.
4202 entity
= entity
->parent
;
4203 for_each_entity(entity
)
4204 entity
->service
= 0;
4208 * Budget timeout is not implemented through a dedicated timer, but
4209 * just checked on request arrivals and completions, as well as on
4210 * idle timer expirations.
4212 static bool bfq_bfqq_budget_timeout(struct bfq_queue
*bfqq
)
4214 return time_is_before_eq_jiffies(bfqq
->budget_timeout
);
4218 * If we expire a queue that is actively waiting (i.e., with the
4219 * device idled) for the arrival of a new request, then we may incur
4220 * the timestamp misalignment problem described in the body of the
4221 * function __bfq_activate_entity. Hence we return true only if this
4222 * condition does not hold, or if the queue is slow enough to deserve
4223 * only to be kicked off for preserving a high throughput.
4225 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue
*bfqq
)
4227 bfq_log_bfqq(bfqq
->bfqd
, bfqq
,
4228 "may_budget_timeout: wait_request %d left %d timeout %d",
4229 bfq_bfqq_wait_request(bfqq
),
4230 bfq_bfqq_budget_left(bfqq
) >= bfqq
->entity
.budget
/ 3,
4231 bfq_bfqq_budget_timeout(bfqq
));
4233 return (!bfq_bfqq_wait_request(bfqq
) ||
4234 bfq_bfqq_budget_left(bfqq
) >= bfqq
->entity
.budget
/ 3)
4236 bfq_bfqq_budget_timeout(bfqq
);
4239 static bool idling_boosts_thr_without_issues(struct bfq_data
*bfqd
,
4240 struct bfq_queue
*bfqq
)
4242 bool rot_without_queueing
=
4243 !blk_queue_nonrot(bfqd
->queue
) && !bfqd
->hw_tag
,
4244 bfqq_sequential_and_IO_bound
,
4247 /* No point in idling for bfqq if it won't get requests any longer */
4248 if (unlikely(!bfqq_process_refs(bfqq
)))
4251 bfqq_sequential_and_IO_bound
= !BFQQ_SEEKY(bfqq
) &&
4252 bfq_bfqq_IO_bound(bfqq
) && bfq_bfqq_has_short_ttime(bfqq
);
4255 * The next variable takes into account the cases where idling
4256 * boosts the throughput.
4258 * The value of the variable is computed considering, first, that
4259 * idling is virtually always beneficial for the throughput if:
4260 * (a) the device is not NCQ-capable and rotational, or
4261 * (b) regardless of the presence of NCQ, the device is rotational and
4262 * the request pattern for bfqq is I/O-bound and sequential, or
4263 * (c) regardless of whether it is rotational, the device is
4264 * not NCQ-capable and the request pattern for bfqq is
4265 * I/O-bound and sequential.
4267 * Secondly, and in contrast to the above item (b), idling an
4268 * NCQ-capable flash-based device would not boost the
4269 * throughput even with sequential I/O; rather it would lower
4270 * the throughput in proportion to how fast the device
4271 * is. Accordingly, the next variable is true if any of the
4272 * above conditions (a), (b) or (c) is true, and, in
4273 * particular, happens to be false if bfqd is an NCQ-capable
4274 * flash-based device.
4276 idling_boosts_thr
= rot_without_queueing
||
4277 ((!blk_queue_nonrot(bfqd
->queue
) || !bfqd
->hw_tag
) &&
4278 bfqq_sequential_and_IO_bound
);
4281 * The return value of this function is equal to that of
4282 * idling_boosts_thr, unless a special case holds. In this
4283 * special case, described below, idling may cause problems to
4284 * weight-raised queues.
4286 * When the request pool is saturated (e.g., in the presence
4287 * of write hogs), if the processes associated with
4288 * non-weight-raised queues ask for requests at a lower rate,
4289 * then processes associated with weight-raised queues have a
4290 * higher probability to get a request from the pool
4291 * immediately (or at least soon) when they need one. Thus
4292 * they have a higher probability to actually get a fraction
4293 * of the device throughput proportional to their high
4294 * weight. This is especially true with NCQ-capable drives,
4295 * which enqueue several requests in advance, and further
4296 * reorder internally-queued requests.
4298 * For this reason, we force to false the return value if
4299 * there are weight-raised busy queues. In this case, and if
4300 * bfqq is not weight-raised, this guarantees that the device
4301 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4302 * then idling will be guaranteed by another variable, see
4303 * below). Combined with the timestamping rules of BFQ (see
4304 * [1] for details), this behavior causes bfqq, and hence any
4305 * sync non-weight-raised queue, to get a lower number of
4306 * requests served, and thus to ask for a lower number of
4307 * requests from the request pool, before the busy
4308 * weight-raised queues get served again. This often mitigates
4309 * starvation problems in the presence of heavy write
4310 * workloads and NCQ, thereby guaranteeing a higher
4311 * application and system responsiveness in these hostile
4314 return idling_boosts_thr
&&
4315 bfqd
->wr_busy_queues
== 0;
4319 * For a queue that becomes empty, device idling is allowed only if
4320 * this function returns true for that queue. As a consequence, since
4321 * device idling plays a critical role for both throughput boosting
4322 * and service guarantees, the return value of this function plays a
4323 * critical role as well.
4325 * In a nutshell, this function returns true only if idling is
4326 * beneficial for throughput or, even if detrimental for throughput,
4327 * idling is however necessary to preserve service guarantees (low
4328 * latency, desired throughput distribution, ...). In particular, on
4329 * NCQ-capable devices, this function tries to return false, so as to
4330 * help keep the drives' internal queues full, whenever this helps the
4331 * device boost the throughput without causing any service-guarantee
4334 * Most of the issues taken into account to get the return value of
4335 * this function are not trivial. We discuss these issues in the two
4336 * functions providing the main pieces of information needed by this
4339 static bool bfq_better_to_idle(struct bfq_queue
*bfqq
)
4341 struct bfq_data
*bfqd
= bfqq
->bfqd
;
4342 bool idling_boosts_thr_with_no_issue
, idling_needed_for_service_guar
;
4344 /* No point in idling for bfqq if it won't get requests any longer */
4345 if (unlikely(!bfqq_process_refs(bfqq
)))
4348 if (unlikely(bfqd
->strict_guarantees
))
4352 * Idling is performed only if slice_idle > 0. In addition, we
4355 * (b) bfqq is in the idle io prio class: in this case we do
4356 * not idle because we want to minimize the bandwidth that
4357 * queues in this class can steal to higher-priority queues
4359 if (bfqd
->bfq_slice_idle
== 0 || !bfq_bfqq_sync(bfqq
) ||
4360 bfq_class_idle(bfqq
))
4363 idling_boosts_thr_with_no_issue
=
4364 idling_boosts_thr_without_issues(bfqd
, bfqq
);
4366 idling_needed_for_service_guar
=
4367 idling_needed_for_service_guarantees(bfqd
, bfqq
);
4370 * We have now the two components we need to compute the
4371 * return value of the function, which is true only if idling
4372 * either boosts the throughput (without issues), or is
4373 * necessary to preserve service guarantees.
4375 return idling_boosts_thr_with_no_issue
||
4376 idling_needed_for_service_guar
;
4380 * If the in-service queue is empty but the function bfq_better_to_idle
4381 * returns true, then:
4382 * 1) the queue must remain in service and cannot be expired, and
4383 * 2) the device must be idled to wait for the possible arrival of a new
4384 * request for the queue.
4385 * See the comments on the function bfq_better_to_idle for the reasons
4386 * why performing device idling is the best choice to boost the throughput
4387 * and preserve service guarantees when bfq_better_to_idle itself
4390 static bool bfq_bfqq_must_idle(struct bfq_queue
*bfqq
)
4392 return RB_EMPTY_ROOT(&bfqq
->sort_list
) && bfq_better_to_idle(bfqq
);
4396 * This function chooses the queue from which to pick the next extra
4397 * I/O request to inject, if it finds a compatible queue. See the
4398 * comments on bfq_update_inject_limit() for details on the injection
4399 * mechanism, and for the definitions of the quantities mentioned
4402 static struct bfq_queue
*
4403 bfq_choose_bfqq_for_injection(struct bfq_data
*bfqd
)
4405 struct bfq_queue
*bfqq
, *in_serv_bfqq
= bfqd
->in_service_queue
;
4406 unsigned int limit
= in_serv_bfqq
->inject_limit
;
4409 * - bfqq is not weight-raised and therefore does not carry
4410 * time-critical I/O,
4412 * - regardless of whether bfqq is weight-raised, bfqq has
4413 * however a long think time, during which it can absorb the
4414 * effect of an appropriate number of extra I/O requests
4415 * from other queues (see bfq_update_inject_limit for
4416 * details on the computation of this number);
4417 * then injection can be performed without restrictions.
4419 bool in_serv_always_inject
= in_serv_bfqq
->wr_coeff
== 1 ||
4420 !bfq_bfqq_has_short_ttime(in_serv_bfqq
);
4424 * - the baseline total service time could not be sampled yet,
4425 * so the inject limit happens to be still 0, and
4426 * - a lot of time has elapsed since the plugging of I/O
4427 * dispatching started, so drive speed is being wasted
4429 * then temporarily raise inject limit to one request.
4431 if (limit
== 0 && in_serv_bfqq
->last_serv_time_ns
== 0 &&
4432 bfq_bfqq_wait_request(in_serv_bfqq
) &&
4433 time_is_before_eq_jiffies(bfqd
->last_idling_start_jiffies
+
4434 bfqd
->bfq_slice_idle
)
4438 if (bfqd
->rq_in_driver
>= limit
)
4442 * Linear search of the source queue for injection; but, with
4443 * a high probability, very few steps are needed to find a
4444 * candidate queue, i.e., a queue with enough budget left for
4445 * its next request. In fact:
4446 * - BFQ dynamically updates the budget of every queue so as
4447 * to accommodate the expected backlog of the queue;
4448 * - if a queue gets all its requests dispatched as injected
4449 * service, then the queue is removed from the active list
4450 * (and re-added only if it gets new requests, but then it
4451 * is assigned again enough budget for its new backlog).
4453 list_for_each_entry(bfqq
, &bfqd
->active_list
, bfqq_list
)
4454 if (!RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
4455 (in_serv_always_inject
|| bfqq
->wr_coeff
> 1) &&
4456 bfq_serv_to_charge(bfqq
->next_rq
, bfqq
) <=
4457 bfq_bfqq_budget_left(bfqq
)) {
4459 * Allow for only one large in-flight request
4460 * on non-rotational devices, for the
4461 * following reason. On non-rotationl drives,
4462 * large requests take much longer than
4463 * smaller requests to be served. In addition,
4464 * the drive prefers to serve large requests
4465 * w.r.t. to small ones, if it can choose. So,
4466 * having more than one large requests queued
4467 * in the drive may easily make the next first
4468 * request of the in-service queue wait for so
4469 * long to break bfqq's service guarantees. On
4470 * the bright side, large requests let the
4471 * drive reach a very high throughput, even if
4472 * there is only one in-flight large request
4475 if (blk_queue_nonrot(bfqd
->queue
) &&
4476 blk_rq_sectors(bfqq
->next_rq
) >=
4477 BFQQ_SECT_THR_NONROT
)
4478 limit
= min_t(unsigned int, 1, limit
);
4480 limit
= in_serv_bfqq
->inject_limit
;
4482 if (bfqd
->rq_in_driver
< limit
) {
4483 bfqd
->rqs_injected
= true;
4492 * Select a queue for service. If we have a current queue in service,
4493 * check whether to continue servicing it, or retrieve and set a new one.
4495 static struct bfq_queue
*bfq_select_queue(struct bfq_data
*bfqd
)
4497 struct bfq_queue
*bfqq
;
4498 struct request
*next_rq
;
4499 enum bfqq_expiration reason
= BFQQE_BUDGET_TIMEOUT
;
4501 bfqq
= bfqd
->in_service_queue
;
4505 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: already in-service queue");
4508 * Do not expire bfqq for budget timeout if bfqq may be about
4509 * to enjoy device idling. The reason why, in this case, we
4510 * prevent bfqq from expiring is the same as in the comments
4511 * on the case where bfq_bfqq_must_idle() returns true, in
4512 * bfq_completed_request().
4514 if (bfq_may_expire_for_budg_timeout(bfqq
) &&
4515 !bfq_bfqq_must_idle(bfqq
))
4520 * This loop is rarely executed more than once. Even when it
4521 * happens, it is much more convenient to re-execute this loop
4522 * than to return NULL and trigger a new dispatch to get a
4525 next_rq
= bfqq
->next_rq
;
4527 * If bfqq has requests queued and it has enough budget left to
4528 * serve them, keep the queue, otherwise expire it.
4531 if (bfq_serv_to_charge(next_rq
, bfqq
) >
4532 bfq_bfqq_budget_left(bfqq
)) {
4534 * Expire the queue for budget exhaustion,
4535 * which makes sure that the next budget is
4536 * enough to serve the next request, even if
4537 * it comes from the fifo expired path.
4539 reason
= BFQQE_BUDGET_EXHAUSTED
;
4543 * The idle timer may be pending because we may
4544 * not disable disk idling even when a new request
4547 if (bfq_bfqq_wait_request(bfqq
)) {
4549 * If we get here: 1) at least a new request
4550 * has arrived but we have not disabled the
4551 * timer because the request was too small,
4552 * 2) then the block layer has unplugged
4553 * the device, causing the dispatch to be
4556 * Since the device is unplugged, now the
4557 * requests are probably large enough to
4558 * provide a reasonable throughput.
4559 * So we disable idling.
4561 bfq_clear_bfqq_wait_request(bfqq
);
4562 hrtimer_try_to_cancel(&bfqd
->idle_slice_timer
);
4569 * No requests pending. However, if the in-service queue is idling
4570 * for a new request, or has requests waiting for a completion and
4571 * may idle after their completion, then keep it anyway.
4573 * Yet, inject service from other queues if it boosts
4574 * throughput and is possible.
4576 if (bfq_bfqq_wait_request(bfqq
) ||
4577 (bfqq
->dispatched
!= 0 && bfq_better_to_idle(bfqq
))) {
4578 struct bfq_queue
*async_bfqq
=
4579 bfqq
->bic
&& bfqq
->bic
->bfqq
[0] &&
4580 bfq_bfqq_busy(bfqq
->bic
->bfqq
[0]) &&
4581 bfqq
->bic
->bfqq
[0]->next_rq
?
4582 bfqq
->bic
->bfqq
[0] : NULL
;
4583 struct bfq_queue
*blocked_bfqq
=
4584 !hlist_empty(&bfqq
->woken_list
) ?
4585 container_of(bfqq
->woken_list
.first
,
4591 * The next four mutually-exclusive ifs decide
4592 * whether to try injection, and choose the queue to
4593 * pick an I/O request from.
4595 * The first if checks whether the process associated
4596 * with bfqq has also async I/O pending. If so, it
4597 * injects such I/O unconditionally. Injecting async
4598 * I/O from the same process can cause no harm to the
4599 * process. On the contrary, it can only increase
4600 * bandwidth and reduce latency for the process.
4602 * The second if checks whether there happens to be a
4603 * non-empty waker queue for bfqq, i.e., a queue whose
4604 * I/O needs to be completed for bfqq to receive new
4605 * I/O. This happens, e.g., if bfqq is associated with
4606 * a process that does some sync. A sync generates
4607 * extra blocking I/O, which must be completed before
4608 * the process associated with bfqq can go on with its
4609 * I/O. If the I/O of the waker queue is not served,
4610 * then bfqq remains empty, and no I/O is dispatched,
4611 * until the idle timeout fires for bfqq. This is
4612 * likely to result in lower bandwidth and higher
4613 * latencies for bfqq, and in a severe loss of total
4614 * throughput. The best action to take is therefore to
4615 * serve the waker queue as soon as possible. So do it
4616 * (without relying on the third alternative below for
4617 * eventually serving waker_bfqq's I/O; see the last
4618 * paragraph for further details). This systematic
4619 * injection of I/O from the waker queue does not
4620 * cause any delay to bfqq's I/O. On the contrary,
4621 * next bfqq's I/O is brought forward dramatically,
4622 * for it is not blocked for milliseconds.
4624 * The third if checks whether there is a queue woken
4625 * by bfqq, and currently with pending I/O. Such a
4626 * woken queue does not steal bandwidth from bfqq,
4627 * because it remains soon without I/O if bfqq is not
4628 * served. So there is virtually no risk of loss of
4629 * bandwidth for bfqq if this woken queue has I/O
4630 * dispatched while bfqq is waiting for new I/O.
4632 * The fourth if checks whether bfqq is a queue for
4633 * which it is better to avoid injection. It is so if
4634 * bfqq delivers more throughput when served without
4635 * any further I/O from other queues in the middle, or
4636 * if the service times of bfqq's I/O requests both
4637 * count more than overall throughput, and may be
4638 * easily increased by injection (this happens if bfqq
4639 * has a short think time). If none of these
4640 * conditions holds, then a candidate queue for
4641 * injection is looked for through
4642 * bfq_choose_bfqq_for_injection(). Note that the
4643 * latter may return NULL (for example if the inject
4644 * limit for bfqq is currently 0).
4646 * NOTE: motivation for the second alternative
4648 * Thanks to the way the inject limit is updated in
4649 * bfq_update_has_short_ttime(), it is rather likely
4650 * that, if I/O is being plugged for bfqq and the
4651 * waker queue has pending I/O requests that are
4652 * blocking bfqq's I/O, then the fourth alternative
4653 * above lets the waker queue get served before the
4654 * I/O-plugging timeout fires. So one may deem the
4655 * second alternative superfluous. It is not, because
4656 * the fourth alternative may be way less effective in
4657 * case of a synchronization. For two main
4658 * reasons. First, throughput may be low because the
4659 * inject limit may be too low to guarantee the same
4660 * amount of injected I/O, from the waker queue or
4661 * other queues, that the second alternative
4662 * guarantees (the second alternative unconditionally
4663 * injects a pending I/O request of the waker queue
4664 * for each bfq_dispatch_request()). Second, with the
4665 * fourth alternative, the duration of the plugging,
4666 * i.e., the time before bfqq finally receives new I/O,
4667 * may not be minimized, because the waker queue may
4668 * happen to be served only after other queues.
4671 icq_to_bic(async_bfqq
->next_rq
->elv
.icq
) == bfqq
->bic
&&
4672 bfq_serv_to_charge(async_bfqq
->next_rq
, async_bfqq
) <=
4673 bfq_bfqq_budget_left(async_bfqq
))
4674 bfqq
= bfqq
->bic
->bfqq
[0];
4675 else if (bfqq
->waker_bfqq
&&
4676 bfq_bfqq_busy(bfqq
->waker_bfqq
) &&
4677 bfqq
->waker_bfqq
->next_rq
&&
4678 bfq_serv_to_charge(bfqq
->waker_bfqq
->next_rq
,
4679 bfqq
->waker_bfqq
) <=
4680 bfq_bfqq_budget_left(bfqq
->waker_bfqq
)
4682 bfqq
= bfqq
->waker_bfqq
;
4683 else if (blocked_bfqq
&&
4684 bfq_bfqq_busy(blocked_bfqq
) &&
4685 blocked_bfqq
->next_rq
&&
4686 bfq_serv_to_charge(blocked_bfqq
->next_rq
,
4688 bfq_bfqq_budget_left(blocked_bfqq
)
4690 bfqq
= blocked_bfqq
;
4691 else if (!idling_boosts_thr_without_issues(bfqd
, bfqq
) &&
4692 (bfqq
->wr_coeff
== 1 || bfqd
->wr_busy_queues
> 1 ||
4693 !bfq_bfqq_has_short_ttime(bfqq
)))
4694 bfqq
= bfq_choose_bfqq_for_injection(bfqd
);
4701 reason
= BFQQE_NO_MORE_REQUESTS
;
4703 bfq_bfqq_expire(bfqd
, bfqq
, false, reason
);
4705 bfqq
= bfq_set_in_service_queue(bfqd
);
4707 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: checking new queue");
4712 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: returned this queue");
4714 bfq_log(bfqd
, "select_queue: no queue returned");
4719 static void bfq_update_wr_data(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
4721 struct bfq_entity
*entity
= &bfqq
->entity
;
4723 if (bfqq
->wr_coeff
> 1) { /* queue is being weight-raised */
4724 bfq_log_bfqq(bfqd
, bfqq
,
4725 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
4726 jiffies_to_msecs(jiffies
- bfqq
->last_wr_start_finish
),
4727 jiffies_to_msecs(bfqq
->wr_cur_max_time
),
4729 bfqq
->entity
.weight
, bfqq
->entity
.orig_weight
);
4731 if (entity
->prio_changed
)
4732 bfq_log_bfqq(bfqd
, bfqq
, "WARN: pending prio change");
4735 * If the queue was activated in a burst, or too much
4736 * time has elapsed from the beginning of this
4737 * weight-raising period, then end weight raising.
4739 if (bfq_bfqq_in_large_burst(bfqq
))
4740 bfq_bfqq_end_wr(bfqq
);
4741 else if (time_is_before_jiffies(bfqq
->last_wr_start_finish
+
4742 bfqq
->wr_cur_max_time
)) {
4743 if (bfqq
->wr_cur_max_time
!= bfqd
->bfq_wr_rt_max_time
||
4744 time_is_before_jiffies(bfqq
->wr_start_at_switch_to_srt
+
4745 bfq_wr_duration(bfqd
))) {
4747 * Either in interactive weight
4748 * raising, or in soft_rt weight
4750 * interactive-weight-raising period
4751 * elapsed (so no switch back to
4752 * interactive weight raising).
4754 bfq_bfqq_end_wr(bfqq
);
4756 * soft_rt finishing while still in
4757 * interactive period, switch back to
4758 * interactive weight raising
4760 switch_back_to_interactive_wr(bfqq
, bfqd
);
4761 bfqq
->entity
.prio_changed
= 1;
4764 if (bfqq
->wr_coeff
> 1 &&
4765 bfqq
->wr_cur_max_time
!= bfqd
->bfq_wr_rt_max_time
&&
4766 bfqq
->service_from_wr
> max_service_from_wr
) {
4767 /* see comments on max_service_from_wr */
4768 bfq_bfqq_end_wr(bfqq
);
4772 * To improve latency (for this or other queues), immediately
4773 * update weight both if it must be raised and if it must be
4774 * lowered. Since, entity may be on some active tree here, and
4775 * might have a pending change of its ioprio class, invoke
4776 * next function with the last parameter unset (see the
4777 * comments on the function).
4779 if ((entity
->weight
> entity
->orig_weight
) != (bfqq
->wr_coeff
> 1))
4780 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity
),
4785 * Dispatch next request from bfqq.
4787 static struct request
*bfq_dispatch_rq_from_bfqq(struct bfq_data
*bfqd
,
4788 struct bfq_queue
*bfqq
)
4790 struct request
*rq
= bfqq
->next_rq
;
4791 unsigned long service_to_charge
;
4793 service_to_charge
= bfq_serv_to_charge(rq
, bfqq
);
4795 bfq_bfqq_served(bfqq
, service_to_charge
);
4797 if (bfqq
== bfqd
->in_service_queue
&& bfqd
->wait_dispatch
) {
4798 bfqd
->wait_dispatch
= false;
4799 bfqd
->waited_rq
= rq
;
4802 bfq_dispatch_remove(bfqd
->queue
, rq
);
4804 if (bfqq
!= bfqd
->in_service_queue
)
4808 * If weight raising has to terminate for bfqq, then next
4809 * function causes an immediate update of bfqq's weight,
4810 * without waiting for next activation. As a consequence, on
4811 * expiration, bfqq will be timestamped as if has never been
4812 * weight-raised during this service slot, even if it has
4813 * received part or even most of the service as a
4814 * weight-raised queue. This inflates bfqq's timestamps, which
4815 * is beneficial, as bfqq is then more willing to leave the
4816 * device immediately to possible other weight-raised queues.
4818 bfq_update_wr_data(bfqd
, bfqq
);
4821 * Expire bfqq, pretending that its budget expired, if bfqq
4822 * belongs to CLASS_IDLE and other queues are waiting for
4825 if (!(bfq_tot_busy_queues(bfqd
) > 1 && bfq_class_idle(bfqq
)))
4828 bfq_bfqq_expire(bfqd
, bfqq
, false, BFQQE_BUDGET_EXHAUSTED
);
4834 static bool bfq_has_work(struct blk_mq_hw_ctx
*hctx
)
4836 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
4839 * Avoiding lock: a race on bfqd->busy_queues should cause at
4840 * most a call to dispatch for nothing
4842 return !list_empty_careful(&bfqd
->dispatch
) ||
4843 bfq_tot_busy_queues(bfqd
) > 0;
4846 static struct request
*__bfq_dispatch_request(struct blk_mq_hw_ctx
*hctx
)
4848 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
4849 struct request
*rq
= NULL
;
4850 struct bfq_queue
*bfqq
= NULL
;
4852 if (!list_empty(&bfqd
->dispatch
)) {
4853 rq
= list_first_entry(&bfqd
->dispatch
, struct request
,
4855 list_del_init(&rq
->queuelist
);
4861 * Increment counters here, because this
4862 * dispatch does not follow the standard
4863 * dispatch flow (where counters are
4868 goto inc_in_driver_start_rq
;
4872 * We exploit the bfq_finish_requeue_request hook to
4873 * decrement rq_in_driver, but
4874 * bfq_finish_requeue_request will not be invoked on
4875 * this request. So, to avoid unbalance, just start
4876 * this request, without incrementing rq_in_driver. As
4877 * a negative consequence, rq_in_driver is deceptively
4878 * lower than it should be while this request is in
4879 * service. This may cause bfq_schedule_dispatch to be
4880 * invoked uselessly.
4882 * As for implementing an exact solution, the
4883 * bfq_finish_requeue_request hook, if defined, is
4884 * probably invoked also on this request. So, by
4885 * exploiting this hook, we could 1) increment
4886 * rq_in_driver here, and 2) decrement it in
4887 * bfq_finish_requeue_request. Such a solution would
4888 * let the value of the counter be always accurate,
4889 * but it would entail using an extra interface
4890 * function. This cost seems higher than the benefit,
4891 * being the frequency of non-elevator-private
4892 * requests very low.
4897 bfq_log(bfqd
, "dispatch requests: %d busy queues",
4898 bfq_tot_busy_queues(bfqd
));
4900 if (bfq_tot_busy_queues(bfqd
) == 0)
4904 * Force device to serve one request at a time if
4905 * strict_guarantees is true. Forcing this service scheme is
4906 * currently the ONLY way to guarantee that the request
4907 * service order enforced by the scheduler is respected by a
4908 * queueing device. Otherwise the device is free even to make
4909 * some unlucky request wait for as long as the device
4912 * Of course, serving one request at a time may cause loss of
4915 if (bfqd
->strict_guarantees
&& bfqd
->rq_in_driver
> 0)
4918 bfqq
= bfq_select_queue(bfqd
);
4922 rq
= bfq_dispatch_rq_from_bfqq(bfqd
, bfqq
);
4925 inc_in_driver_start_rq
:
4926 bfqd
->rq_in_driver
++;
4928 rq
->rq_flags
|= RQF_STARTED
;
4934 #ifdef CONFIG_BFQ_CGROUP_DEBUG
4935 static void bfq_update_dispatch_stats(struct request_queue
*q
,
4937 struct bfq_queue
*in_serv_queue
,
4938 bool idle_timer_disabled
)
4940 struct bfq_queue
*bfqq
= rq
? RQ_BFQQ(rq
) : NULL
;
4942 if (!idle_timer_disabled
&& !bfqq
)
4946 * rq and bfqq are guaranteed to exist until this function
4947 * ends, for the following reasons. First, rq can be
4948 * dispatched to the device, and then can be completed and
4949 * freed, only after this function ends. Second, rq cannot be
4950 * merged (and thus freed because of a merge) any longer,
4951 * because it has already started. Thus rq cannot be freed
4952 * before this function ends, and, since rq has a reference to
4953 * bfqq, the same guarantee holds for bfqq too.
4955 * In addition, the following queue lock guarantees that
4956 * bfqq_group(bfqq) exists as well.
4958 spin_lock_irq(&q
->queue_lock
);
4959 if (idle_timer_disabled
)
4961 * Since the idle timer has been disabled,
4962 * in_serv_queue contained some request when
4963 * __bfq_dispatch_request was invoked above, which
4964 * implies that rq was picked exactly from
4965 * in_serv_queue. Thus in_serv_queue == bfqq, and is
4966 * therefore guaranteed to exist because of the above
4969 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue
));
4971 struct bfq_group
*bfqg
= bfqq_group(bfqq
);
4973 bfqg_stats_update_avg_queue_size(bfqg
);
4974 bfqg_stats_set_start_empty_time(bfqg
);
4975 bfqg_stats_update_io_remove(bfqg
, rq
->cmd_flags
);
4977 spin_unlock_irq(&q
->queue_lock
);
4980 static inline void bfq_update_dispatch_stats(struct request_queue
*q
,
4982 struct bfq_queue
*in_serv_queue
,
4983 bool idle_timer_disabled
) {}
4984 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
4986 static struct request
*bfq_dispatch_request(struct blk_mq_hw_ctx
*hctx
)
4988 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
4990 struct bfq_queue
*in_serv_queue
;
4991 bool waiting_rq
, idle_timer_disabled
;
4993 spin_lock_irq(&bfqd
->lock
);
4995 in_serv_queue
= bfqd
->in_service_queue
;
4996 waiting_rq
= in_serv_queue
&& bfq_bfqq_wait_request(in_serv_queue
);
4998 rq
= __bfq_dispatch_request(hctx
);
5000 idle_timer_disabled
=
5001 waiting_rq
&& !bfq_bfqq_wait_request(in_serv_queue
);
5003 spin_unlock_irq(&bfqd
->lock
);
5005 bfq_update_dispatch_stats(hctx
->queue
, rq
, in_serv_queue
,
5006 idle_timer_disabled
);
5012 * Task holds one reference to the queue, dropped when task exits. Each rq
5013 * in-flight on this queue also holds a reference, dropped when rq is freed.
5015 * Scheduler lock must be held here. Recall not to use bfqq after calling
5016 * this function on it.
5018 void bfq_put_queue(struct bfq_queue
*bfqq
)
5020 struct bfq_queue
*item
;
5021 struct hlist_node
*n
;
5022 struct bfq_group
*bfqg
= bfqq_group(bfqq
);
5025 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "put_queue: %p %d",
5032 if (!hlist_unhashed(&bfqq
->burst_list_node
)) {
5033 hlist_del_init(&bfqq
->burst_list_node
);
5035 * Decrement also burst size after the removal, if the
5036 * process associated with bfqq is exiting, and thus
5037 * does not contribute to the burst any longer. This
5038 * decrement helps filter out false positives of large
5039 * bursts, when some short-lived process (often due to
5040 * the execution of commands by some service) happens
5041 * to start and exit while a complex application is
5042 * starting, and thus spawning several processes that
5043 * do I/O (and that *must not* be treated as a large
5044 * burst, see comments on bfq_handle_burst).
5046 * In particular, the decrement is performed only if:
5047 * 1) bfqq is not a merged queue, because, if it is,
5048 * then this free of bfqq is not triggered by the exit
5049 * of the process bfqq is associated with, but exactly
5050 * by the fact that bfqq has just been merged.
5051 * 2) burst_size is greater than 0, to handle
5052 * unbalanced decrements. Unbalanced decrements may
5053 * happen in te following case: bfqq is inserted into
5054 * the current burst list--without incrementing
5055 * bust_size--because of a split, but the current
5056 * burst list is not the burst list bfqq belonged to
5057 * (see comments on the case of a split in
5060 if (bfqq
->bic
&& bfqq
->bfqd
->burst_size
> 0)
5061 bfqq
->bfqd
->burst_size
--;
5065 * bfqq does not exist any longer, so it cannot be woken by
5066 * any other queue, and cannot wake any other queue. Then bfqq
5067 * must be removed from the woken list of its possible waker
5068 * queue, and all queues in the woken list of bfqq must stop
5069 * having a waker queue. Strictly speaking, these updates
5070 * should be performed when bfqq remains with no I/O source
5071 * attached to it, which happens before bfqq gets freed. In
5072 * particular, this happens when the last process associated
5073 * with bfqq exits or gets associated with a different
5074 * queue. However, both events lead to bfqq being freed soon,
5075 * and dangling references would come out only after bfqq gets
5076 * freed. So these updates are done here, as a simple and safe
5077 * way to handle all cases.
5079 /* remove bfqq from woken list */
5080 if (!hlist_unhashed(&bfqq
->woken_list_node
))
5081 hlist_del_init(&bfqq
->woken_list_node
);
5083 /* reset waker for all queues in woken list */
5084 hlist_for_each_entry_safe(item
, n
, &bfqq
->woken_list
,
5086 item
->waker_bfqq
= NULL
;
5087 hlist_del_init(&item
->woken_list_node
);
5090 if (bfqq
->bfqd
&& bfqq
->bfqd
->last_completed_rq_bfqq
== bfqq
)
5091 bfqq
->bfqd
->last_completed_rq_bfqq
= NULL
;
5093 kmem_cache_free(bfq_pool
, bfqq
);
5094 bfqg_and_blkg_put(bfqg
);
5097 static void bfq_put_stable_ref(struct bfq_queue
*bfqq
)
5100 bfq_put_queue(bfqq
);
5103 static void bfq_put_cooperator(struct bfq_queue
*bfqq
)
5105 struct bfq_queue
*__bfqq
, *next
;
5108 * If this queue was scheduled to merge with another queue, be
5109 * sure to drop the reference taken on that queue (and others in
5110 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
5112 __bfqq
= bfqq
->new_bfqq
;
5116 next
= __bfqq
->new_bfqq
;
5117 bfq_put_queue(__bfqq
);
5122 static void bfq_exit_bfqq(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
5124 if (bfqq
== bfqd
->in_service_queue
) {
5125 __bfq_bfqq_expire(bfqd
, bfqq
, BFQQE_BUDGET_TIMEOUT
);
5126 bfq_schedule_dispatch(bfqd
);
5129 bfq_log_bfqq(bfqd
, bfqq
, "exit_bfqq: %p, %d", bfqq
, bfqq
->ref
);
5131 bfq_put_cooperator(bfqq
);
5133 bfq_release_process_ref(bfqd
, bfqq
);
5136 static void bfq_exit_icq_bfqq(struct bfq_io_cq
*bic
, bool is_sync
)
5138 struct bfq_queue
*bfqq
= bic_to_bfqq(bic
, is_sync
);
5139 struct bfq_data
*bfqd
;
5142 bfqd
= bfqq
->bfqd
; /* NULL if scheduler already exited */
5145 unsigned long flags
;
5147 spin_lock_irqsave(&bfqd
->lock
, flags
);
5149 bfq_exit_bfqq(bfqd
, bfqq
);
5150 bic_set_bfqq(bic
, NULL
, is_sync
);
5151 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
5155 static void bfq_exit_icq(struct io_cq
*icq
)
5157 struct bfq_io_cq
*bic
= icq_to_bic(icq
);
5159 if (bic
->stable_merge_bfqq
) {
5160 struct bfq_data
*bfqd
= bic
->stable_merge_bfqq
->bfqd
;
5163 * bfqd is NULL if scheduler already exited, and in
5164 * that case this is the last time bfqq is accessed.
5167 unsigned long flags
;
5169 spin_lock_irqsave(&bfqd
->lock
, flags
);
5170 bfq_put_stable_ref(bic
->stable_merge_bfqq
);
5171 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
5173 bfq_put_stable_ref(bic
->stable_merge_bfqq
);
5177 bfq_exit_icq_bfqq(bic
, true);
5178 bfq_exit_icq_bfqq(bic
, false);
5182 * Update the entity prio values; note that the new values will not
5183 * be used until the next (re)activation.
5186 bfq_set_next_ioprio_data(struct bfq_queue
*bfqq
, struct bfq_io_cq
*bic
)
5188 struct task_struct
*tsk
= current
;
5190 struct bfq_data
*bfqd
= bfqq
->bfqd
;
5195 ioprio_class
= IOPRIO_PRIO_CLASS(bic
->ioprio
);
5196 switch (ioprio_class
) {
5198 pr_err("bdi %s: bfq: bad prio class %d\n",
5199 bdi_dev_name(bfqq
->bfqd
->queue
->backing_dev_info
),
5202 case IOPRIO_CLASS_NONE
:
5204 * No prio set, inherit CPU scheduling settings.
5206 bfqq
->new_ioprio
= task_nice_ioprio(tsk
);
5207 bfqq
->new_ioprio_class
= task_nice_ioclass(tsk
);
5209 case IOPRIO_CLASS_RT
:
5210 bfqq
->new_ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
5211 bfqq
->new_ioprio_class
= IOPRIO_CLASS_RT
;
5213 case IOPRIO_CLASS_BE
:
5214 bfqq
->new_ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
5215 bfqq
->new_ioprio_class
= IOPRIO_CLASS_BE
;
5217 case IOPRIO_CLASS_IDLE
:
5218 bfqq
->new_ioprio_class
= IOPRIO_CLASS_IDLE
;
5219 bfqq
->new_ioprio
= 7;
5223 if (bfqq
->new_ioprio
>= IOPRIO_BE_NR
) {
5224 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
5226 bfqq
->new_ioprio
= IOPRIO_BE_NR
;
5229 bfqq
->entity
.new_weight
= bfq_ioprio_to_weight(bfqq
->new_ioprio
);
5230 bfq_log_bfqq(bfqd
, bfqq
, "new_ioprio %d new_weight %d",
5231 bfqq
->new_ioprio
, bfqq
->entity
.new_weight
);
5232 bfqq
->entity
.prio_changed
= 1;
5235 static struct bfq_queue
*bfq_get_queue(struct bfq_data
*bfqd
,
5236 struct bio
*bio
, bool is_sync
,
5237 struct bfq_io_cq
*bic
,
5240 static void bfq_check_ioprio_change(struct bfq_io_cq
*bic
, struct bio
*bio
)
5242 struct bfq_data
*bfqd
= bic_to_bfqd(bic
);
5243 struct bfq_queue
*bfqq
;
5244 int ioprio
= bic
->icq
.ioc
->ioprio
;
5247 * This condition may trigger on a newly created bic, be sure to
5248 * drop the lock before returning.
5250 if (unlikely(!bfqd
) || likely(bic
->ioprio
== ioprio
))
5253 bic
->ioprio
= ioprio
;
5255 bfqq
= bic_to_bfqq(bic
, false);
5257 bfq_release_process_ref(bfqd
, bfqq
);
5258 bfqq
= bfq_get_queue(bfqd
, bio
, BLK_RW_ASYNC
, bic
, true);
5259 bic_set_bfqq(bic
, bfqq
, false);
5262 bfqq
= bic_to_bfqq(bic
, true);
5264 bfq_set_next_ioprio_data(bfqq
, bic
);
5267 static void bfq_init_bfqq(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
5268 struct bfq_io_cq
*bic
, pid_t pid
, int is_sync
)
5270 u64 now_ns
= ktime_get_ns();
5272 RB_CLEAR_NODE(&bfqq
->entity
.rb_node
);
5273 INIT_LIST_HEAD(&bfqq
->fifo
);
5274 INIT_HLIST_NODE(&bfqq
->burst_list_node
);
5275 INIT_HLIST_NODE(&bfqq
->woken_list_node
);
5276 INIT_HLIST_HEAD(&bfqq
->woken_list
);
5282 bfq_set_next_ioprio_data(bfqq
, bic
);
5286 * No need to mark as has_short_ttime if in
5287 * idle_class, because no device idling is performed
5288 * for queues in idle class
5290 if (!bfq_class_idle(bfqq
))
5291 /* tentatively mark as has_short_ttime */
5292 bfq_mark_bfqq_has_short_ttime(bfqq
);
5293 bfq_mark_bfqq_sync(bfqq
);
5294 bfq_mark_bfqq_just_created(bfqq
);
5296 bfq_clear_bfqq_sync(bfqq
);
5298 /* set end request to minus infinity from now */
5299 bfqq
->ttime
.last_end_request
= now_ns
+ 1;
5301 bfqq
->creation_time
= jiffies
;
5303 bfqq
->io_start_time
= now_ns
;
5305 bfq_mark_bfqq_IO_bound(bfqq
);
5309 /* Tentative initial value to trade off between thr and lat */
5310 bfqq
->max_budget
= (2 * bfq_max_budget(bfqd
)) / 3;
5311 bfqq
->budget_timeout
= bfq_smallest_from_now();
5314 bfqq
->last_wr_start_finish
= jiffies
;
5315 bfqq
->wr_start_at_switch_to_srt
= bfq_smallest_from_now();
5316 bfqq
->split_time
= bfq_smallest_from_now();
5319 * To not forget the possibly high bandwidth consumed by a
5320 * process/queue in the recent past,
5321 * bfq_bfqq_softrt_next_start() returns a value at least equal
5322 * to the current value of bfqq->soft_rt_next_start (see
5323 * comments on bfq_bfqq_softrt_next_start). Set
5324 * soft_rt_next_start to now, to mean that bfqq has consumed
5325 * no bandwidth so far.
5327 bfqq
->soft_rt_next_start
= jiffies
;
5329 /* first request is almost certainly seeky */
5330 bfqq
->seek_history
= 1;
5333 static struct bfq_queue
**bfq_async_queue_prio(struct bfq_data
*bfqd
,
5334 struct bfq_group
*bfqg
,
5335 int ioprio_class
, int ioprio
)
5337 switch (ioprio_class
) {
5338 case IOPRIO_CLASS_RT
:
5339 return &bfqg
->async_bfqq
[0][ioprio
];
5340 case IOPRIO_CLASS_NONE
:
5341 ioprio
= IOPRIO_NORM
;
5343 case IOPRIO_CLASS_BE
:
5344 return &bfqg
->async_bfqq
[1][ioprio
];
5345 case IOPRIO_CLASS_IDLE
:
5346 return &bfqg
->async_idle_bfqq
;
5352 static struct bfq_queue
*
5353 bfq_do_early_stable_merge(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
5354 struct bfq_io_cq
*bic
,
5355 struct bfq_queue
*last_bfqq_created
)
5357 struct bfq_queue
*new_bfqq
=
5358 bfq_setup_merge(bfqq
, last_bfqq_created
);
5364 new_bfqq
->bic
->stably_merged
= true;
5365 bic
->stably_merged
= true;
5368 * Reusing merge functions. This implies that
5369 * bfqq->bic must be set too, for
5370 * bfq_merge_bfqqs to correctly save bfqq's
5371 * state before killing it.
5374 bfq_merge_bfqqs(bfqd
, bic
, bfqq
, new_bfqq
);
5380 * Many throughput-sensitive workloads are made of several parallel
5381 * I/O flows, with all flows generated by the same application, or
5382 * more generically by the same task (e.g., system boot). The most
5383 * counterproductive action with these workloads is plugging I/O
5384 * dispatch when one of the bfq_queues associated with these flows
5385 * remains temporarily empty.
5387 * To avoid this plugging, BFQ has been using a burst-handling
5388 * mechanism for years now. This mechanism has proven effective for
5389 * throughput, and not detrimental for service guarantees. The
5390 * following function pushes this mechanism a little bit further,
5391 * basing on the following two facts.
5393 * First, all the I/O flows of a the same application or task
5394 * contribute to the execution/completion of that common application
5395 * or task. So the performance figures that matter are total
5396 * throughput of the flows and task-wide I/O latency. In particular,
5397 * these flows do not need to be protected from each other, in terms
5398 * of individual bandwidth or latency.
5400 * Second, the above fact holds regardless of the number of flows.
5402 * Putting these two facts together, this commits merges stably the
5403 * bfq_queues associated with these I/O flows, i.e., with the
5404 * processes that generate these IO/ flows, regardless of how many the
5405 * involved processes are.
5407 * To decide whether a set of bfq_queues is actually associated with
5408 * the I/O flows of a common application or task, and to merge these
5409 * queues stably, this function operates as follows: given a bfq_queue,
5410 * say Q2, currently being created, and the last bfq_queue, say Q1,
5411 * created before Q2, Q2 is merged stably with Q1 if
5412 * - very little time has elapsed since when Q1 was created
5413 * - Q2 has the same ioprio as Q1
5414 * - Q2 belongs to the same group as Q1
5416 * Merging bfq_queues also reduces scheduling overhead. A fio test
5417 * with ten random readers on /dev/nullb shows a throughput boost of
5418 * 40%, with a quadcore. Since BFQ's execution time amounts to ~50% of
5419 * the total per-request processing time, the above throughput boost
5420 * implies that BFQ's overhead is reduced by more than 50%.
5422 * This new mechanism most certainly obsoletes the current
5423 * burst-handling heuristics. We keep those heuristics for the moment.
5425 static struct bfq_queue
*bfq_do_or_sched_stable_merge(struct bfq_data
*bfqd
,
5426 struct bfq_queue
*bfqq
,
5427 struct bfq_io_cq
*bic
)
5429 struct bfq_queue
**source_bfqq
= bfqq
->entity
.parent
?
5430 &bfqq
->entity
.parent
->last_bfqq_created
:
5431 &bfqd
->last_bfqq_created
;
5433 struct bfq_queue
*last_bfqq_created
= *source_bfqq
;
5436 * If last_bfqq_created has not been set yet, then init it. If
5437 * it has been set already, but too long ago, then move it
5438 * forward to bfqq. Finally, move also if bfqq belongs to a
5439 * different group than last_bfqq_created, or if bfqq has a
5440 * different ioprio or ioprio_class. If none of these
5441 * conditions holds true, then try an early stable merge or
5442 * schedule a delayed stable merge.
5444 * A delayed merge is scheduled (instead of performing an
5445 * early merge), in case bfqq might soon prove to be more
5446 * throughput-beneficial if not merged. Currently this is
5447 * possible only if bfqd is rotational with no queueing. For
5448 * such a drive, not merging bfqq is better for throughput if
5449 * bfqq happens to contain sequential I/O. So, we wait a
5450 * little bit for enough I/O to flow through bfqq. After that,
5451 * if such an I/O is sequential, then the merge is
5452 * canceled. Otherwise the merge is finally performed.
5454 if (!last_bfqq_created
||
5455 time_before(last_bfqq_created
->creation_time
+
5456 bfqd
->bfq_burst_interval
,
5457 bfqq
->creation_time
) ||
5458 bfqq
->entity
.parent
!= last_bfqq_created
->entity
.parent
||
5459 bfqq
->ioprio
!= last_bfqq_created
->ioprio
||
5460 bfqq
->ioprio_class
!= last_bfqq_created
->ioprio_class
)
5461 *source_bfqq
= bfqq
;
5462 else if (time_after_eq(last_bfqq_created
->creation_time
+
5463 bfqd
->bfq_burst_interval
,
5464 bfqq
->creation_time
)) {
5465 if (likely(bfqd
->nonrot_with_queueing
))
5467 * With this type of drive, leaving
5468 * bfqq alone may provide no
5469 * throughput benefits compared with
5470 * merging bfqq. So merge bfqq now.
5472 bfqq
= bfq_do_early_stable_merge(bfqd
, bfqq
,
5475 else { /* schedule tentative stable merge */
5477 * get reference on last_bfqq_created,
5478 * to prevent it from being freed,
5479 * until we decide whether to merge
5481 last_bfqq_created
->ref
++;
5483 * need to keep track of stable refs, to
5484 * compute process refs correctly
5486 last_bfqq_created
->stable_ref
++;
5488 * Record the bfqq to merge to.
5490 bic
->stable_merge_bfqq
= last_bfqq_created
;
5498 static struct bfq_queue
*bfq_get_queue(struct bfq_data
*bfqd
,
5499 struct bio
*bio
, bool is_sync
,
5500 struct bfq_io_cq
*bic
,
5503 const int ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
5504 const int ioprio_class
= IOPRIO_PRIO_CLASS(bic
->ioprio
);
5505 struct bfq_queue
**async_bfqq
= NULL
;
5506 struct bfq_queue
*bfqq
;
5507 struct bfq_group
*bfqg
;
5511 bfqg
= bfq_find_set_group(bfqd
, __bio_blkcg(bio
));
5513 bfqq
= &bfqd
->oom_bfqq
;
5518 async_bfqq
= bfq_async_queue_prio(bfqd
, bfqg
, ioprio_class
,
5525 bfqq
= kmem_cache_alloc_node(bfq_pool
,
5526 GFP_NOWAIT
| __GFP_ZERO
| __GFP_NOWARN
,
5530 bfq_init_bfqq(bfqd
, bfqq
, bic
, current
->pid
,
5532 bfq_init_entity(&bfqq
->entity
, bfqg
);
5533 bfq_log_bfqq(bfqd
, bfqq
, "allocated");
5535 bfqq
= &bfqd
->oom_bfqq
;
5536 bfq_log_bfqq(bfqd
, bfqq
, "using oom bfqq");
5541 * Pin the queue now that it's allocated, scheduler exit will
5546 * Extra group reference, w.r.t. sync
5547 * queue. This extra reference is removed
5548 * only if bfqq->bfqg disappears, to
5549 * guarantee that this queue is not freed
5550 * until its group goes away.
5552 bfq_log_bfqq(bfqd
, bfqq
, "get_queue, bfqq not in async: %p, %d",
5558 bfqq
->ref
++; /* get a process reference to this queue */
5560 if (bfqq
!= &bfqd
->oom_bfqq
&& is_sync
&& !respawn
)
5561 bfqq
= bfq_do_or_sched_stable_merge(bfqd
, bfqq
, bic
);
5567 static void bfq_update_io_thinktime(struct bfq_data
*bfqd
,
5568 struct bfq_queue
*bfqq
)
5570 struct bfq_ttime
*ttime
= &bfqq
->ttime
;
5574 * We are really interested in how long it takes for the queue to
5575 * become busy when there is no outstanding IO for this queue. So
5576 * ignore cases when the bfq queue has already IO queued.
5578 if (bfqq
->dispatched
|| bfq_bfqq_busy(bfqq
))
5580 elapsed
= ktime_get_ns() - bfqq
->ttime
.last_end_request
;
5581 elapsed
= min_t(u64
, elapsed
, 2ULL * bfqd
->bfq_slice_idle
);
5583 ttime
->ttime_samples
= (7*ttime
->ttime_samples
+ 256) / 8;
5584 ttime
->ttime_total
= div_u64(7*ttime
->ttime_total
+ 256*elapsed
, 8);
5585 ttime
->ttime_mean
= div64_ul(ttime
->ttime_total
+ 128,
5586 ttime
->ttime_samples
);
5590 bfq_update_io_seektime(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
5593 bfqq
->seek_history
<<= 1;
5594 bfqq
->seek_history
|= BFQ_RQ_SEEKY(bfqd
, bfqq
->last_request_pos
, rq
);
5596 if (bfqq
->wr_coeff
> 1 &&
5597 bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
&&
5598 BFQQ_TOTALLY_SEEKY(bfqq
)) {
5599 if (time_is_before_jiffies(bfqq
->wr_start_at_switch_to_srt
+
5600 bfq_wr_duration(bfqd
))) {
5602 * In soft_rt weight raising with the
5603 * interactive-weight-raising period
5604 * elapsed (so no switch back to
5605 * interactive weight raising).
5607 bfq_bfqq_end_wr(bfqq
);
5609 * stopping soft_rt weight raising
5610 * while still in interactive period,
5611 * switch back to interactive weight
5614 switch_back_to_interactive_wr(bfqq
, bfqd
);
5615 bfqq
->entity
.prio_changed
= 1;
5620 static void bfq_update_has_short_ttime(struct bfq_data
*bfqd
,
5621 struct bfq_queue
*bfqq
,
5622 struct bfq_io_cq
*bic
)
5624 bool has_short_ttime
= true, state_changed
;
5627 * No need to update has_short_ttime if bfqq is async or in
5628 * idle io prio class, or if bfq_slice_idle is zero, because
5629 * no device idling is performed for bfqq in this case.
5631 if (!bfq_bfqq_sync(bfqq
) || bfq_class_idle(bfqq
) ||
5632 bfqd
->bfq_slice_idle
== 0)
5635 /* Idle window just restored, statistics are meaningless. */
5636 if (time_is_after_eq_jiffies(bfqq
->split_time
+
5637 bfqd
->bfq_wr_min_idle_time
))
5640 /* Think time is infinite if no process is linked to
5641 * bfqq. Otherwise check average think time to decide whether
5642 * to mark as has_short_ttime. To this goal, compare average
5643 * think time with half the I/O-plugging timeout.
5645 if (atomic_read(&bic
->icq
.ioc
->active_ref
) == 0 ||
5646 (bfq_sample_valid(bfqq
->ttime
.ttime_samples
) &&
5647 bfqq
->ttime
.ttime_mean
> bfqd
->bfq_slice_idle
>>1))
5648 has_short_ttime
= false;
5650 state_changed
= has_short_ttime
!= bfq_bfqq_has_short_ttime(bfqq
);
5652 if (has_short_ttime
)
5653 bfq_mark_bfqq_has_short_ttime(bfqq
);
5655 bfq_clear_bfqq_has_short_ttime(bfqq
);
5658 * Until the base value for the total service time gets
5659 * finally computed for bfqq, the inject limit does depend on
5660 * the think-time state (short|long). In particular, the limit
5661 * is 0 or 1 if the think time is deemed, respectively, as
5662 * short or long (details in the comments in
5663 * bfq_update_inject_limit()). Accordingly, the next
5664 * instructions reset the inject limit if the think-time state
5665 * has changed and the above base value is still to be
5668 * However, the reset is performed only if more than 100 ms
5669 * have elapsed since the last update of the inject limit, or
5670 * (inclusive) if the change is from short to long think
5671 * time. The reason for this waiting is as follows.
5673 * bfqq may have a long think time because of a
5674 * synchronization with some other queue, i.e., because the
5675 * I/O of some other queue may need to be completed for bfqq
5676 * to receive new I/O. Details in the comments on the choice
5677 * of the queue for injection in bfq_select_queue().
5679 * As stressed in those comments, if such a synchronization is
5680 * actually in place, then, without injection on bfqq, the
5681 * blocking I/O cannot happen to served while bfqq is in
5682 * service. As a consequence, if bfqq is granted
5683 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
5684 * is dispatched, until the idle timeout fires. This is likely
5685 * to result in lower bandwidth and higher latencies for bfqq,
5686 * and in a severe loss of total throughput.
5688 * On the opposite end, a non-zero inject limit may allow the
5689 * I/O that blocks bfqq to be executed soon, and therefore
5690 * bfqq to receive new I/O soon.
5692 * But, if the blocking gets actually eliminated, then the
5693 * next think-time sample for bfqq may be very low. This in
5694 * turn may cause bfqq's think time to be deemed
5695 * short. Without the 100 ms barrier, this new state change
5696 * would cause the body of the next if to be executed
5697 * immediately. But this would set to 0 the inject
5698 * limit. Without injection, the blocking I/O would cause the
5699 * think time of bfqq to become long again, and therefore the
5700 * inject limit to be raised again, and so on. The only effect
5701 * of such a steady oscillation between the two think-time
5702 * states would be to prevent effective injection on bfqq.
5704 * In contrast, if the inject limit is not reset during such a
5705 * long time interval as 100 ms, then the number of short
5706 * think time samples can grow significantly before the reset
5707 * is performed. As a consequence, the think time state can
5708 * become stable before the reset. Therefore there will be no
5709 * state change when the 100 ms elapse, and no reset of the
5710 * inject limit. The inject limit remains steadily equal to 1
5711 * both during and after the 100 ms. So injection can be
5712 * performed at all times, and throughput gets boosted.
5714 * An inject limit equal to 1 is however in conflict, in
5715 * general, with the fact that the think time of bfqq is
5716 * short, because injection may be likely to delay bfqq's I/O
5717 * (as explained in the comments in
5718 * bfq_update_inject_limit()). But this does not happen in
5719 * this special case, because bfqq's low think time is due to
5720 * an effective handling of a synchronization, through
5721 * injection. In this special case, bfqq's I/O does not get
5722 * delayed by injection; on the contrary, bfqq's I/O is
5723 * brought forward, because it is not blocked for
5726 * In addition, serving the blocking I/O much sooner, and much
5727 * more frequently than once per I/O-plugging timeout, makes
5728 * it much quicker to detect a waker queue (the concept of
5729 * waker queue is defined in the comments in
5730 * bfq_add_request()). This makes it possible to start sooner
5731 * to boost throughput more effectively, by injecting the I/O
5732 * of the waker queue unconditionally on every
5733 * bfq_dispatch_request().
5735 * One last, important benefit of not resetting the inject
5736 * limit before 100 ms is that, during this time interval, the
5737 * base value for the total service time is likely to get
5738 * finally computed for bfqq, freeing the inject limit from
5739 * its relation with the think time.
5741 if (state_changed
&& bfqq
->last_serv_time_ns
== 0 &&
5742 (time_is_before_eq_jiffies(bfqq
->decrease_time_jif
+
5743 msecs_to_jiffies(100)) ||
5745 bfq_reset_inject_limit(bfqd
, bfqq
);
5749 * Called when a new fs request (rq) is added to bfqq. Check if there's
5750 * something we should do about it.
5752 static void bfq_rq_enqueued(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
5755 if (rq
->cmd_flags
& REQ_META
)
5756 bfqq
->meta_pending
++;
5758 bfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
5760 if (bfqq
== bfqd
->in_service_queue
&& bfq_bfqq_wait_request(bfqq
)) {
5761 bool small_req
= bfqq
->queued
[rq_is_sync(rq
)] == 1 &&
5762 blk_rq_sectors(rq
) < 32;
5763 bool budget_timeout
= bfq_bfqq_budget_timeout(bfqq
);
5766 * There is just this request queued: if
5767 * - the request is small, and
5768 * - we are idling to boost throughput, and
5769 * - the queue is not to be expired,
5772 * In this way, if the device is being idled to wait
5773 * for a new request from the in-service queue, we
5774 * avoid unplugging the device and committing the
5775 * device to serve just a small request. In contrast
5776 * we wait for the block layer to decide when to
5777 * unplug the device: hopefully, new requests will be
5778 * merged to this one quickly, then the device will be
5779 * unplugged and larger requests will be dispatched.
5781 if (small_req
&& idling_boosts_thr_without_issues(bfqd
, bfqq
) &&
5786 * A large enough request arrived, or idling is being
5787 * performed to preserve service guarantees, or
5788 * finally the queue is to be expired: in all these
5789 * cases disk idling is to be stopped, so clear
5790 * wait_request flag and reset timer.
5792 bfq_clear_bfqq_wait_request(bfqq
);
5793 hrtimer_try_to_cancel(&bfqd
->idle_slice_timer
);
5796 * The queue is not empty, because a new request just
5797 * arrived. Hence we can safely expire the queue, in
5798 * case of budget timeout, without risking that the
5799 * timestamps of the queue are not updated correctly.
5800 * See [1] for more details.
5803 bfq_bfqq_expire(bfqd
, bfqq
, false,
5804 BFQQE_BUDGET_TIMEOUT
);
5808 /* returns true if it causes the idle timer to be disabled */
5809 static bool __bfq_insert_request(struct bfq_data
*bfqd
, struct request
*rq
)
5811 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
),
5812 *new_bfqq
= bfq_setup_cooperator(bfqd
, bfqq
, rq
, true,
5814 bool waiting
, idle_timer_disabled
= false;
5818 * Release the request's reference to the old bfqq
5819 * and make sure one is taken to the shared queue.
5821 new_bfqq
->allocated
++;
5825 * If the bic associated with the process
5826 * issuing this request still points to bfqq
5827 * (and thus has not been already redirected
5828 * to new_bfqq or even some other bfq_queue),
5829 * then complete the merge and redirect it to
5832 if (bic_to_bfqq(RQ_BIC(rq
), 1) == bfqq
)
5833 bfq_merge_bfqqs(bfqd
, RQ_BIC(rq
),
5836 bfq_clear_bfqq_just_created(bfqq
);
5838 * rq is about to be enqueued into new_bfqq,
5839 * release rq reference on bfqq
5841 bfq_put_queue(bfqq
);
5842 rq
->elv
.priv
[1] = new_bfqq
;
5846 bfq_update_io_thinktime(bfqd
, bfqq
);
5847 bfq_update_has_short_ttime(bfqd
, bfqq
, RQ_BIC(rq
));
5848 bfq_update_io_seektime(bfqd
, bfqq
, rq
);
5850 waiting
= bfqq
&& bfq_bfqq_wait_request(bfqq
);
5851 bfq_add_request(rq
);
5852 idle_timer_disabled
= waiting
&& !bfq_bfqq_wait_request(bfqq
);
5854 rq
->fifo_time
= ktime_get_ns() + bfqd
->bfq_fifo_expire
[rq_is_sync(rq
)];
5855 list_add_tail(&rq
->queuelist
, &bfqq
->fifo
);
5857 bfq_rq_enqueued(bfqd
, bfqq
, rq
);
5859 return idle_timer_disabled
;
5862 #ifdef CONFIG_BFQ_CGROUP_DEBUG
5863 static void bfq_update_insert_stats(struct request_queue
*q
,
5864 struct bfq_queue
*bfqq
,
5865 bool idle_timer_disabled
,
5866 unsigned int cmd_flags
)
5872 * bfqq still exists, because it can disappear only after
5873 * either it is merged with another queue, or the process it
5874 * is associated with exits. But both actions must be taken by
5875 * the same process currently executing this flow of
5878 * In addition, the following queue lock guarantees that
5879 * bfqq_group(bfqq) exists as well.
5881 spin_lock_irq(&q
->queue_lock
);
5882 bfqg_stats_update_io_add(bfqq_group(bfqq
), bfqq
, cmd_flags
);
5883 if (idle_timer_disabled
)
5884 bfqg_stats_update_idle_time(bfqq_group(bfqq
));
5885 spin_unlock_irq(&q
->queue_lock
);
5888 static inline void bfq_update_insert_stats(struct request_queue
*q
,
5889 struct bfq_queue
*bfqq
,
5890 bool idle_timer_disabled
,
5891 unsigned int cmd_flags
) {}
5892 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
5894 static void bfq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
5897 struct request_queue
*q
= hctx
->queue
;
5898 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
5899 struct bfq_queue
*bfqq
;
5900 bool idle_timer_disabled
= false;
5901 unsigned int cmd_flags
;
5903 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5904 if (!cgroup_subsys_on_dfl(io_cgrp_subsys
) && rq
->bio
)
5905 bfqg_stats_update_legacy_io(q
, rq
);
5907 spin_lock_irq(&bfqd
->lock
);
5908 if (blk_mq_sched_try_insert_merge(q
, rq
)) {
5909 spin_unlock_irq(&bfqd
->lock
);
5913 spin_unlock_irq(&bfqd
->lock
);
5915 trace_block_rq_insert(rq
);
5917 spin_lock_irq(&bfqd
->lock
);
5918 bfqq
= bfq_init_rq(rq
);
5921 * Reqs with at_head or passthrough flags set are to be put
5922 * directly into dispatch list. Additional case for putting rq
5923 * directly into the dispatch queue: the only active
5924 * bfq_queues are bfqq and either its waker bfq_queue or one
5925 * of its woken bfq_queues. The rationale behind this
5926 * additional condition is as follows:
5927 * - consider a bfq_queue, say Q1, detected as a waker of
5928 * another bfq_queue, say Q2
5929 * - by definition of a waker, Q1 blocks the I/O of Q2, i.e.,
5930 * some I/O of Q1 needs to be completed for new I/O of Q2
5931 * to arrive. A notable example of waker is journald
5932 * - so, Q1 and Q2 are in any respect the queues of two
5933 * cooperating processes (or of two cooperating sets of
5934 * processes): the goal of Q1's I/O is doing what needs to
5935 * be done so that new Q2's I/O can finally be
5936 * issued. Therefore, if the service of Q1's I/O is delayed,
5937 * then Q2's I/O is delayed too. Conversely, if Q2's I/O is
5938 * delayed, the goal of Q1's I/O is hindered.
5939 * - as a consequence, if some I/O of Q1/Q2 arrives while
5940 * Q2/Q1 is the only queue in service, there is absolutely
5941 * no point in delaying the service of such an I/O. The
5942 * only possible result is a throughput loss
5943 * - so, when the above condition holds, the best option is to
5944 * have the new I/O dispatched as soon as possible
5945 * - the most effective and efficient way to attain the above
5946 * goal is to put the new I/O directly in the dispatch
5948 * - as an additional restriction, Q1 and Q2 must be the only
5949 * busy queues for this commit to put the I/O of Q2/Q1 in
5950 * the dispatch list. This is necessary, because, if also
5951 * other queues are waiting for service, then putting new
5952 * I/O directly in the dispatch list may evidently cause a
5953 * violation of service guarantees for the other queues
5956 (bfqq
!= bfqd
->in_service_queue
&&
5957 bfqd
->in_service_queue
!= NULL
&&
5958 bfq_tot_busy_queues(bfqd
) == 1 + bfq_bfqq_busy(bfqq
) &&
5959 (bfqq
->waker_bfqq
== bfqd
->in_service_queue
||
5960 bfqd
->in_service_queue
->waker_bfqq
== bfqq
)) || at_head
) {
5962 list_add(&rq
->queuelist
, &bfqd
->dispatch
);
5964 list_add_tail(&rq
->queuelist
, &bfqd
->dispatch
);
5966 idle_timer_disabled
= __bfq_insert_request(bfqd
, rq
);
5968 * Update bfqq, because, if a queue merge has occurred
5969 * in __bfq_insert_request, then rq has been
5970 * redirected into a new queue.
5974 if (rq_mergeable(rq
)) {
5975 elv_rqhash_add(q
, rq
);
5982 * Cache cmd_flags before releasing scheduler lock, because rq
5983 * may disappear afterwards (for example, because of a request
5986 cmd_flags
= rq
->cmd_flags
;
5988 spin_unlock_irq(&bfqd
->lock
);
5990 bfq_update_insert_stats(q
, bfqq
, idle_timer_disabled
,
5994 static void bfq_insert_requests(struct blk_mq_hw_ctx
*hctx
,
5995 struct list_head
*list
, bool at_head
)
5997 while (!list_empty(list
)) {
6000 rq
= list_first_entry(list
, struct request
, queuelist
);
6001 list_del_init(&rq
->queuelist
);
6002 bfq_insert_request(hctx
, rq
, at_head
);
6006 static void bfq_update_hw_tag(struct bfq_data
*bfqd
)
6008 struct bfq_queue
*bfqq
= bfqd
->in_service_queue
;
6010 bfqd
->max_rq_in_driver
= max_t(int, bfqd
->max_rq_in_driver
,
6011 bfqd
->rq_in_driver
);
6013 if (bfqd
->hw_tag
== 1)
6017 * This sample is valid if the number of outstanding requests
6018 * is large enough to allow a queueing behavior. Note that the
6019 * sum is not exact, as it's not taking into account deactivated
6022 if (bfqd
->rq_in_driver
+ bfqd
->queued
<= BFQ_HW_QUEUE_THRESHOLD
)
6026 * If active queue hasn't enough requests and can idle, bfq might not
6027 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
6030 if (bfqq
&& bfq_bfqq_has_short_ttime(bfqq
) &&
6031 bfqq
->dispatched
+ bfqq
->queued
[0] + bfqq
->queued
[1] <
6032 BFQ_HW_QUEUE_THRESHOLD
&&
6033 bfqd
->rq_in_driver
< BFQ_HW_QUEUE_THRESHOLD
)
6036 if (bfqd
->hw_tag_samples
++ < BFQ_HW_QUEUE_SAMPLES
)
6039 bfqd
->hw_tag
= bfqd
->max_rq_in_driver
> BFQ_HW_QUEUE_THRESHOLD
;
6040 bfqd
->max_rq_in_driver
= 0;
6041 bfqd
->hw_tag_samples
= 0;
6043 bfqd
->nonrot_with_queueing
=
6044 blk_queue_nonrot(bfqd
->queue
) && bfqd
->hw_tag
;
6047 static void bfq_completed_request(struct bfq_queue
*bfqq
, struct bfq_data
*bfqd
)
6052 bfq_update_hw_tag(bfqd
);
6054 bfqd
->rq_in_driver
--;
6057 if (!bfqq
->dispatched
&& !bfq_bfqq_busy(bfqq
)) {
6059 * Set budget_timeout (which we overload to store the
6060 * time at which the queue remains with no backlog and
6061 * no outstanding request; used by the weight-raising
6064 bfqq
->budget_timeout
= jiffies
;
6066 bfq_weights_tree_remove(bfqd
, bfqq
);
6069 now_ns
= ktime_get_ns();
6071 bfqq
->ttime
.last_end_request
= now_ns
;
6074 * Using us instead of ns, to get a reasonable precision in
6075 * computing rate in next check.
6077 delta_us
= div_u64(now_ns
- bfqd
->last_completion
, NSEC_PER_USEC
);
6080 * If the request took rather long to complete, and, according
6081 * to the maximum request size recorded, this completion latency
6082 * implies that the request was certainly served at a very low
6083 * rate (less than 1M sectors/sec), then the whole observation
6084 * interval that lasts up to this time instant cannot be a
6085 * valid time interval for computing a new peak rate. Invoke
6086 * bfq_update_rate_reset to have the following three steps
6088 * - close the observation interval at the last (previous)
6089 * request dispatch or completion
6090 * - compute rate, if possible, for that observation interval
6091 * - reset to zero samples, which will trigger a proper
6092 * re-initialization of the observation interval on next
6095 if (delta_us
> BFQ_MIN_TT
/NSEC_PER_USEC
&&
6096 (bfqd
->last_rq_max_size
<<BFQ_RATE_SHIFT
)/delta_us
<
6097 1UL<<(BFQ_RATE_SHIFT
- 10))
6098 bfq_update_rate_reset(bfqd
, NULL
);
6099 bfqd
->last_completion
= now_ns
;
6101 * Shared queues are likely to receive I/O at a high
6102 * rate. This may deceptively let them be considered as wakers
6103 * of other queues. But a false waker will unjustly steal
6104 * bandwidth to its supposedly woken queue. So considering
6105 * also shared queues in the waking mechanism may cause more
6106 * control troubles than throughput benefits. Then do not set
6107 * last_completed_rq_bfqq to bfqq if bfqq is a shared queue.
6109 if (!bfq_bfqq_coop(bfqq
))
6110 bfqd
->last_completed_rq_bfqq
= bfqq
;
6113 * If we are waiting to discover whether the request pattern
6114 * of the task associated with the queue is actually
6115 * isochronous, and both requisites for this condition to hold
6116 * are now satisfied, then compute soft_rt_next_start (see the
6117 * comments on the function bfq_bfqq_softrt_next_start()). We
6118 * do not compute soft_rt_next_start if bfqq is in interactive
6119 * weight raising (see the comments in bfq_bfqq_expire() for
6120 * an explanation). We schedule this delayed update when bfqq
6121 * expires, if it still has in-flight requests.
6123 if (bfq_bfqq_softrt_update(bfqq
) && bfqq
->dispatched
== 0 &&
6124 RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
6125 bfqq
->wr_coeff
!= bfqd
->bfq_wr_coeff
)
6126 bfqq
->soft_rt_next_start
=
6127 bfq_bfqq_softrt_next_start(bfqd
, bfqq
);
6130 * If this is the in-service queue, check if it needs to be expired,
6131 * or if we want to idle in case it has no pending requests.
6133 if (bfqd
->in_service_queue
== bfqq
) {
6134 if (bfq_bfqq_must_idle(bfqq
)) {
6135 if (bfqq
->dispatched
== 0)
6136 bfq_arm_slice_timer(bfqd
);
6138 * If we get here, we do not expire bfqq, even
6139 * if bfqq was in budget timeout or had no
6140 * more requests (as controlled in the next
6141 * conditional instructions). The reason for
6142 * not expiring bfqq is as follows.
6144 * Here bfqq->dispatched > 0 holds, but
6145 * bfq_bfqq_must_idle() returned true. This
6146 * implies that, even if no request arrives
6147 * for bfqq before bfqq->dispatched reaches 0,
6148 * bfqq will, however, not be expired on the
6149 * completion event that causes bfqq->dispatch
6150 * to reach zero. In contrast, on this event,
6151 * bfqq will start enjoying device idling
6152 * (I/O-dispatch plugging).
6154 * But, if we expired bfqq here, bfqq would
6155 * not have the chance to enjoy device idling
6156 * when bfqq->dispatched finally reaches
6157 * zero. This would expose bfqq to violation
6158 * of its reserved service guarantees.
6161 } else if (bfq_may_expire_for_budg_timeout(bfqq
))
6162 bfq_bfqq_expire(bfqd
, bfqq
, false,
6163 BFQQE_BUDGET_TIMEOUT
);
6164 else if (RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
6165 (bfqq
->dispatched
== 0 ||
6166 !bfq_better_to_idle(bfqq
)))
6167 bfq_bfqq_expire(bfqd
, bfqq
, false,
6168 BFQQE_NO_MORE_REQUESTS
);
6171 if (!bfqd
->rq_in_driver
)
6172 bfq_schedule_dispatch(bfqd
);
6175 static void bfq_finish_requeue_request_body(struct bfq_queue
*bfqq
)
6179 bfq_put_queue(bfqq
);
6183 * The processes associated with bfqq may happen to generate their
6184 * cumulative I/O at a lower rate than the rate at which the device
6185 * could serve the same I/O. This is rather probable, e.g., if only
6186 * one process is associated with bfqq and the device is an SSD. It
6187 * results in bfqq becoming often empty while in service. In this
6188 * respect, if BFQ is allowed to switch to another queue when bfqq
6189 * remains empty, then the device goes on being fed with I/O requests,
6190 * and the throughput is not affected. In contrast, if BFQ is not
6191 * allowed to switch to another queue---because bfqq is sync and
6192 * I/O-dispatch needs to be plugged while bfqq is temporarily
6193 * empty---then, during the service of bfqq, there will be frequent
6194 * "service holes", i.e., time intervals during which bfqq gets empty
6195 * and the device can only consume the I/O already queued in its
6196 * hardware queues. During service holes, the device may even get to
6197 * remaining idle. In the end, during the service of bfqq, the device
6198 * is driven at a lower speed than the one it can reach with the kind
6199 * of I/O flowing through bfqq.
6201 * To counter this loss of throughput, BFQ implements a "request
6202 * injection mechanism", which tries to fill the above service holes
6203 * with I/O requests taken from other queues. The hard part in this
6204 * mechanism is finding the right amount of I/O to inject, so as to
6205 * both boost throughput and not break bfqq's bandwidth and latency
6206 * guarantees. In this respect, the mechanism maintains a per-queue
6207 * inject limit, computed as below. While bfqq is empty, the injection
6208 * mechanism dispatches extra I/O requests only until the total number
6209 * of I/O requests in flight---i.e., already dispatched but not yet
6210 * completed---remains lower than this limit.
6212 * A first definition comes in handy to introduce the algorithm by
6213 * which the inject limit is computed. We define as first request for
6214 * bfqq, an I/O request for bfqq that arrives while bfqq is in
6215 * service, and causes bfqq to switch from empty to non-empty. The
6216 * algorithm updates the limit as a function of the effect of
6217 * injection on the service times of only the first requests of
6218 * bfqq. The reason for this restriction is that these are the
6219 * requests whose service time is affected most, because they are the
6220 * first to arrive after injection possibly occurred.
6222 * To evaluate the effect of injection, the algorithm measures the
6223 * "total service time" of first requests. We define as total service
6224 * time of an I/O request, the time that elapses since when the
6225 * request is enqueued into bfqq, to when it is completed. This
6226 * quantity allows the whole effect of injection to be measured. It is
6227 * easy to see why. Suppose that some requests of other queues are
6228 * actually injected while bfqq is empty, and that a new request R
6229 * then arrives for bfqq. If the device does start to serve all or
6230 * part of the injected requests during the service hole, then,
6231 * because of this extra service, it may delay the next invocation of
6232 * the dispatch hook of BFQ. Then, even after R gets eventually
6233 * dispatched, the device may delay the actual service of R if it is
6234 * still busy serving the extra requests, or if it decides to serve,
6235 * before R, some extra request still present in its queues. As a
6236 * conclusion, the cumulative extra delay caused by injection can be
6237 * easily evaluated by just comparing the total service time of first
6238 * requests with and without injection.
6240 * The limit-update algorithm works as follows. On the arrival of a
6241 * first request of bfqq, the algorithm measures the total time of the
6242 * request only if one of the three cases below holds, and, for each
6243 * case, it updates the limit as described below:
6245 * (1) If there is no in-flight request. This gives a baseline for the
6246 * total service time of the requests of bfqq. If the baseline has
6247 * not been computed yet, then, after computing it, the limit is
6248 * set to 1, to start boosting throughput, and to prepare the
6249 * ground for the next case. If the baseline has already been
6250 * computed, then it is updated, in case it results to be lower
6251 * than the previous value.
6253 * (2) If the limit is higher than 0 and there are in-flight
6254 * requests. By comparing the total service time in this case with
6255 * the above baseline, it is possible to know at which extent the
6256 * current value of the limit is inflating the total service
6257 * time. If the inflation is below a certain threshold, then bfqq
6258 * is assumed to be suffering from no perceivable loss of its
6259 * service guarantees, and the limit is even tentatively
6260 * increased. If the inflation is above the threshold, then the
6261 * limit is decreased. Due to the lack of any hysteresis, this
6262 * logic makes the limit oscillate even in steady workload
6263 * conditions. Yet we opted for it, because it is fast in reaching
6264 * the best value for the limit, as a function of the current I/O
6265 * workload. To reduce oscillations, this step is disabled for a
6266 * short time interval after the limit happens to be decreased.
6268 * (3) Periodically, after resetting the limit, to make sure that the
6269 * limit eventually drops in case the workload changes. This is
6270 * needed because, after the limit has gone safely up for a
6271 * certain workload, it is impossible to guess whether the
6272 * baseline total service time may have changed, without measuring
6273 * it again without injection. A more effective version of this
6274 * step might be to just sample the baseline, by interrupting
6275 * injection only once, and then to reset/lower the limit only if
6276 * the total service time with the current limit does happen to be
6279 * More details on each step are provided in the comments on the
6280 * pieces of code that implement these steps: the branch handling the
6281 * transition from empty to non empty in bfq_add_request(), the branch
6282 * handling injection in bfq_select_queue(), and the function
6283 * bfq_choose_bfqq_for_injection(). These comments also explain some
6284 * exceptions, made by the injection mechanism in some special cases.
6286 static void bfq_update_inject_limit(struct bfq_data
*bfqd
,
6287 struct bfq_queue
*bfqq
)
6289 u64 tot_time_ns
= ktime_get_ns() - bfqd
->last_empty_occupied_ns
;
6290 unsigned int old_limit
= bfqq
->inject_limit
;
6292 if (bfqq
->last_serv_time_ns
> 0 && bfqd
->rqs_injected
) {
6293 u64 threshold
= (bfqq
->last_serv_time_ns
* 3)>>1;
6295 if (tot_time_ns
>= threshold
&& old_limit
> 0) {
6296 bfqq
->inject_limit
--;
6297 bfqq
->decrease_time_jif
= jiffies
;
6298 } else if (tot_time_ns
< threshold
&&
6299 old_limit
<= bfqd
->max_rq_in_driver
)
6300 bfqq
->inject_limit
++;
6304 * Either we still have to compute the base value for the
6305 * total service time, and there seem to be the right
6306 * conditions to do it, or we can lower the last base value
6309 * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
6310 * request in flight, because this function is in the code
6311 * path that handles the completion of a request of bfqq, and,
6312 * in particular, this function is executed before
6313 * bfqd->rq_in_driver is decremented in such a code path.
6315 if ((bfqq
->last_serv_time_ns
== 0 && bfqd
->rq_in_driver
== 1) ||
6316 tot_time_ns
< bfqq
->last_serv_time_ns
) {
6317 if (bfqq
->last_serv_time_ns
== 0) {
6319 * Now we certainly have a base value: make sure we
6320 * start trying injection.
6322 bfqq
->inject_limit
= max_t(unsigned int, 1, old_limit
);
6324 bfqq
->last_serv_time_ns
= tot_time_ns
;
6325 } else if (!bfqd
->rqs_injected
&& bfqd
->rq_in_driver
== 1)
6327 * No I/O injected and no request still in service in
6328 * the drive: these are the exact conditions for
6329 * computing the base value of the total service time
6330 * for bfqq. So let's update this value, because it is
6331 * rather variable. For example, it varies if the size
6332 * or the spatial locality of the I/O requests in bfqq
6335 bfqq
->last_serv_time_ns
= tot_time_ns
;
6338 /* update complete, not waiting for any request completion any longer */
6339 bfqd
->waited_rq
= NULL
;
6340 bfqd
->rqs_injected
= false;
6344 * Handle either a requeue or a finish for rq. The things to do are
6345 * the same in both cases: all references to rq are to be dropped. In
6346 * particular, rq is considered completed from the point of view of
6349 static void bfq_finish_requeue_request(struct request
*rq
)
6351 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
6352 struct bfq_data
*bfqd
;
6355 * rq either is not associated with any icq, or is an already
6356 * requeued request that has not (yet) been re-inserted into
6359 if (!rq
->elv
.icq
|| !bfqq
)
6364 if (rq
->rq_flags
& RQF_STARTED
)
6365 bfqg_stats_update_completion(bfqq_group(bfqq
),
6367 rq
->io_start_time_ns
,
6370 if (likely(rq
->rq_flags
& RQF_STARTED
)) {
6371 unsigned long flags
;
6373 spin_lock_irqsave(&bfqd
->lock
, flags
);
6375 if (rq
== bfqd
->waited_rq
)
6376 bfq_update_inject_limit(bfqd
, bfqq
);
6378 bfq_completed_request(bfqq
, bfqd
);
6379 bfq_finish_requeue_request_body(bfqq
);
6381 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
6384 * Request rq may be still/already in the scheduler,
6385 * in which case we need to remove it (this should
6386 * never happen in case of requeue). And we cannot
6387 * defer such a check and removal, to avoid
6388 * inconsistencies in the time interval from the end
6389 * of this function to the start of the deferred work.
6390 * This situation seems to occur only in process
6391 * context, as a consequence of a merge. In the
6392 * current version of the code, this implies that the
6396 if (!RB_EMPTY_NODE(&rq
->rb_node
)) {
6397 bfq_remove_request(rq
->q
, rq
);
6398 bfqg_stats_update_io_remove(bfqq_group(bfqq
),
6401 bfq_finish_requeue_request_body(bfqq
);
6405 * Reset private fields. In case of a requeue, this allows
6406 * this function to correctly do nothing if it is spuriously
6407 * invoked again on this same request (see the check at the
6408 * beginning of the function). Probably, a better general
6409 * design would be to prevent blk-mq from invoking the requeue
6410 * or finish hooks of an elevator, for a request that is not
6411 * referred by that elevator.
6413 * Resetting the following fields would break the
6414 * request-insertion logic if rq is re-inserted into a bfq
6415 * internal queue, without a re-preparation. Here we assume
6416 * that re-insertions of requeued requests, without
6417 * re-preparation, can happen only for pass_through or at_head
6418 * requests (which are not re-inserted into bfq internal
6421 rq
->elv
.priv
[0] = NULL
;
6422 rq
->elv
.priv
[1] = NULL
;
6426 * Removes the association between the current task and bfqq, assuming
6427 * that bic points to the bfq iocontext of the task.
6428 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
6429 * was the last process referring to that bfqq.
6431 static struct bfq_queue
*
6432 bfq_split_bfqq(struct bfq_io_cq
*bic
, struct bfq_queue
*bfqq
)
6434 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "splitting queue");
6436 if (bfqq_process_refs(bfqq
) == 1) {
6437 bfqq
->pid
= current
->pid
;
6438 bfq_clear_bfqq_coop(bfqq
);
6439 bfq_clear_bfqq_split_coop(bfqq
);
6443 bic_set_bfqq(bic
, NULL
, 1);
6445 bfq_put_cooperator(bfqq
);
6447 bfq_release_process_ref(bfqq
->bfqd
, bfqq
);
6451 static struct bfq_queue
*bfq_get_bfqq_handle_split(struct bfq_data
*bfqd
,
6452 struct bfq_io_cq
*bic
,
6454 bool split
, bool is_sync
,
6457 struct bfq_queue
*bfqq
= bic_to_bfqq(bic
, is_sync
);
6459 if (likely(bfqq
&& bfqq
!= &bfqd
->oom_bfqq
))
6466 bfq_put_queue(bfqq
);
6467 bfqq
= bfq_get_queue(bfqd
, bio
, is_sync
, bic
, split
);
6469 bic_set_bfqq(bic
, bfqq
, is_sync
);
6470 if (split
&& is_sync
) {
6471 if ((bic
->was_in_burst_list
&& bfqd
->large_burst
) ||
6472 bic
->saved_in_large_burst
)
6473 bfq_mark_bfqq_in_large_burst(bfqq
);
6475 bfq_clear_bfqq_in_large_burst(bfqq
);
6476 if (bic
->was_in_burst_list
)
6478 * If bfqq was in the current
6479 * burst list before being
6480 * merged, then we have to add
6481 * it back. And we do not need
6482 * to increase burst_size, as
6483 * we did not decrement
6484 * burst_size when we removed
6485 * bfqq from the burst list as
6486 * a consequence of a merge
6488 * bfq_put_queue). In this
6489 * respect, it would be rather
6490 * costly to know whether the
6491 * current burst list is still
6492 * the same burst list from
6493 * which bfqq was removed on
6494 * the merge. To avoid this
6495 * cost, if bfqq was in a
6496 * burst list, then we add
6497 * bfqq to the current burst
6498 * list without any further
6499 * check. This can cause
6500 * inappropriate insertions,
6501 * but rarely enough to not
6502 * harm the detection of large
6503 * bursts significantly.
6505 hlist_add_head(&bfqq
->burst_list_node
,
6508 bfqq
->split_time
= jiffies
;
6515 * Only reset private fields. The actual request preparation will be
6516 * performed by bfq_init_rq, when rq is either inserted or merged. See
6517 * comments on bfq_init_rq for the reason behind this delayed
6520 static void bfq_prepare_request(struct request
*rq
)
6523 * Regardless of whether we have an icq attached, we have to
6524 * clear the scheduler pointers, as they might point to
6525 * previously allocated bic/bfqq structs.
6527 rq
->elv
.priv
[0] = rq
->elv
.priv
[1] = NULL
;
6531 * If needed, init rq, allocate bfq data structures associated with
6532 * rq, and increment reference counters in the destination bfq_queue
6533 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6534 * not associated with any bfq_queue.
6536 * This function is invoked by the functions that perform rq insertion
6537 * or merging. One may have expected the above preparation operations
6538 * to be performed in bfq_prepare_request, and not delayed to when rq
6539 * is inserted or merged. The rationale behind this delayed
6540 * preparation is that, after the prepare_request hook is invoked for
6541 * rq, rq may still be transformed into a request with no icq, i.e., a
6542 * request not associated with any queue. No bfq hook is invoked to
6543 * signal this transformation. As a consequence, should these
6544 * preparation operations be performed when the prepare_request hook
6545 * is invoked, and should rq be transformed one moment later, bfq
6546 * would end up in an inconsistent state, because it would have
6547 * incremented some queue counters for an rq destined to
6548 * transformation, without any chance to correctly lower these
6549 * counters back. In contrast, no transformation can still happen for
6550 * rq after rq has been inserted or merged. So, it is safe to execute
6551 * these preparation operations when rq is finally inserted or merged.
6553 static struct bfq_queue
*bfq_init_rq(struct request
*rq
)
6555 struct request_queue
*q
= rq
->q
;
6556 struct bio
*bio
= rq
->bio
;
6557 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
6558 struct bfq_io_cq
*bic
;
6559 const int is_sync
= rq_is_sync(rq
);
6560 struct bfq_queue
*bfqq
;
6561 bool new_queue
= false;
6562 bool bfqq_already_existing
= false, split
= false;
6564 if (unlikely(!rq
->elv
.icq
))
6568 * Assuming that elv.priv[1] is set only if everything is set
6569 * for this rq. This holds true, because this function is
6570 * invoked only for insertion or merging, and, after such
6571 * events, a request cannot be manipulated any longer before
6572 * being removed from bfq.
6574 if (rq
->elv
.priv
[1])
6575 return rq
->elv
.priv
[1];
6577 bic
= icq_to_bic(rq
->elv
.icq
);
6579 bfq_check_ioprio_change(bic
, bio
);
6581 bfq_bic_update_cgroup(bic
, bio
);
6583 bfqq
= bfq_get_bfqq_handle_split(bfqd
, bic
, bio
, false, is_sync
,
6586 if (likely(!new_queue
)) {
6587 /* If the queue was seeky for too long, break it apart. */
6588 if (bfq_bfqq_coop(bfqq
) && bfq_bfqq_split_coop(bfqq
) &&
6589 !bic
->stably_merged
) {
6590 struct bfq_queue
*old_bfqq
= bfqq
;
6592 /* Update bic before losing reference to bfqq */
6593 if (bfq_bfqq_in_large_burst(bfqq
))
6594 bic
->saved_in_large_burst
= true;
6596 bfqq
= bfq_split_bfqq(bic
, bfqq
);
6600 bfqq
= bfq_get_bfqq_handle_split(bfqd
, bic
, bio
,
6603 bfqq
->waker_bfqq
= old_bfqq
->waker_bfqq
;
6604 bfqq
->tentative_waker_bfqq
= NULL
;
6607 * If the waker queue disappears, then
6608 * new_bfqq->waker_bfqq must be
6609 * reset. So insert new_bfqq into the
6610 * woken_list of the waker. See
6611 * bfq_check_waker for details.
6613 if (bfqq
->waker_bfqq
)
6614 hlist_add_head(&bfqq
->woken_list_node
,
6615 &bfqq
->waker_bfqq
->woken_list
);
6617 bfqq_already_existing
= true;
6623 bfq_log_bfqq(bfqd
, bfqq
, "get_request %p: bfqq %p, %d",
6624 rq
, bfqq
, bfqq
->ref
);
6626 rq
->elv
.priv
[0] = bic
;
6627 rq
->elv
.priv
[1] = bfqq
;
6630 * If a bfq_queue has only one process reference, it is owned
6631 * by only this bic: we can then set bfqq->bic = bic. in
6632 * addition, if the queue has also just been split, we have to
6635 if (likely(bfqq
!= &bfqd
->oom_bfqq
) && bfqq_process_refs(bfqq
) == 1) {
6639 * The queue has just been split from a shared
6640 * queue: restore the idle window and the
6641 * possible weight raising period.
6643 bfq_bfqq_resume_state(bfqq
, bfqd
, bic
,
6644 bfqq_already_existing
);
6649 * Consider bfqq as possibly belonging to a burst of newly
6650 * created queues only if:
6651 * 1) A burst is actually happening (bfqd->burst_size > 0)
6653 * 2) There is no other active queue. In fact, if, in
6654 * contrast, there are active queues not belonging to the
6655 * possible burst bfqq may belong to, then there is no gain
6656 * in considering bfqq as belonging to a burst, and
6657 * therefore in not weight-raising bfqq. See comments on
6658 * bfq_handle_burst().
6660 * This filtering also helps eliminating false positives,
6661 * occurring when bfqq does not belong to an actual large
6662 * burst, but some background task (e.g., a service) happens
6663 * to trigger the creation of new queues very close to when
6664 * bfqq and its possible companion queues are created. See
6665 * comments on bfq_handle_burst() for further details also on
6668 if (unlikely(bfq_bfqq_just_created(bfqq
) &&
6669 (bfqd
->burst_size
> 0 ||
6670 bfq_tot_busy_queues(bfqd
) == 0)))
6671 bfq_handle_burst(bfqd
, bfqq
);
6677 bfq_idle_slice_timer_body(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
6679 enum bfqq_expiration reason
;
6680 unsigned long flags
;
6682 spin_lock_irqsave(&bfqd
->lock
, flags
);
6685 * Considering that bfqq may be in race, we should firstly check
6686 * whether bfqq is in service before doing something on it. If
6687 * the bfqq in race is not in service, it has already been expired
6688 * through __bfq_bfqq_expire func and its wait_request flags has
6689 * been cleared in __bfq_bfqd_reset_in_service func.
6691 if (bfqq
!= bfqd
->in_service_queue
) {
6692 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
6696 bfq_clear_bfqq_wait_request(bfqq
);
6698 if (bfq_bfqq_budget_timeout(bfqq
))
6700 * Also here the queue can be safely expired
6701 * for budget timeout without wasting
6704 reason
= BFQQE_BUDGET_TIMEOUT
;
6705 else if (bfqq
->queued
[0] == 0 && bfqq
->queued
[1] == 0)
6707 * The queue may not be empty upon timer expiration,
6708 * because we may not disable the timer when the
6709 * first request of the in-service queue arrives
6710 * during disk idling.
6712 reason
= BFQQE_TOO_IDLE
;
6714 goto schedule_dispatch
;
6716 bfq_bfqq_expire(bfqd
, bfqq
, true, reason
);
6719 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
6720 bfq_schedule_dispatch(bfqd
);
6724 * Handler of the expiration of the timer running if the in-service queue
6725 * is idling inside its time slice.
6727 static enum hrtimer_restart
bfq_idle_slice_timer(struct hrtimer
*timer
)
6729 struct bfq_data
*bfqd
= container_of(timer
, struct bfq_data
,
6731 struct bfq_queue
*bfqq
= bfqd
->in_service_queue
;
6734 * Theoretical race here: the in-service queue can be NULL or
6735 * different from the queue that was idling if a new request
6736 * arrives for the current queue and there is a full dispatch
6737 * cycle that changes the in-service queue. This can hardly
6738 * happen, but in the worst case we just expire a queue too
6742 bfq_idle_slice_timer_body(bfqd
, bfqq
);
6744 return HRTIMER_NORESTART
;
6747 static void __bfq_put_async_bfqq(struct bfq_data
*bfqd
,
6748 struct bfq_queue
**bfqq_ptr
)
6750 struct bfq_queue
*bfqq
= *bfqq_ptr
;
6752 bfq_log(bfqd
, "put_async_bfqq: %p", bfqq
);
6754 bfq_bfqq_move(bfqd
, bfqq
, bfqd
->root_group
);
6756 bfq_log_bfqq(bfqd
, bfqq
, "put_async_bfqq: putting %p, %d",
6758 bfq_put_queue(bfqq
);
6764 * Release all the bfqg references to its async queues. If we are
6765 * deallocating the group these queues may still contain requests, so
6766 * we reparent them to the root cgroup (i.e., the only one that will
6767 * exist for sure until all the requests on a device are gone).
6769 void bfq_put_async_queues(struct bfq_data
*bfqd
, struct bfq_group
*bfqg
)
6773 for (i
= 0; i
< 2; i
++)
6774 for (j
= 0; j
< IOPRIO_BE_NR
; j
++)
6775 __bfq_put_async_bfqq(bfqd
, &bfqg
->async_bfqq
[i
][j
]);
6777 __bfq_put_async_bfqq(bfqd
, &bfqg
->async_idle_bfqq
);
6781 * See the comments on bfq_limit_depth for the purpose of
6782 * the depths set in the function. Return minimum shallow depth we'll use.
6784 static unsigned int bfq_update_depths(struct bfq_data
*bfqd
,
6785 struct sbitmap_queue
*bt
)
6787 unsigned int i
, j
, min_shallow
= UINT_MAX
;
6790 * In-word depths if no bfq_queue is being weight-raised:
6791 * leaving 25% of tags only for sync reads.
6793 * In next formulas, right-shift the value
6794 * (1U<<bt->sb.shift), instead of computing directly
6795 * (1U<<(bt->sb.shift - something)), to be robust against
6796 * any possible value of bt->sb.shift, without having to
6797 * limit 'something'.
6799 /* no more than 50% of tags for async I/O */
6800 bfqd
->word_depths
[0][0] = max((1U << bt
->sb
.shift
) >> 1, 1U);
6802 * no more than 75% of tags for sync writes (25% extra tags
6803 * w.r.t. async I/O, to prevent async I/O from starving sync
6806 bfqd
->word_depths
[0][1] = max(((1U << bt
->sb
.shift
) * 3) >> 2, 1U);
6809 * In-word depths in case some bfq_queue is being weight-
6810 * raised: leaving ~63% of tags for sync reads. This is the
6811 * highest percentage for which, in our tests, application
6812 * start-up times didn't suffer from any regression due to tag
6815 /* no more than ~18% of tags for async I/O */
6816 bfqd
->word_depths
[1][0] = max(((1U << bt
->sb
.shift
) * 3) >> 4, 1U);
6817 /* no more than ~37% of tags for sync writes (~20% extra tags) */
6818 bfqd
->word_depths
[1][1] = max(((1U << bt
->sb
.shift
) * 6) >> 4, 1U);
6820 for (i
= 0; i
< 2; i
++)
6821 for (j
= 0; j
< 2; j
++)
6822 min_shallow
= min(min_shallow
, bfqd
->word_depths
[i
][j
]);
6827 static void bfq_depth_updated(struct blk_mq_hw_ctx
*hctx
)
6829 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
6830 struct blk_mq_tags
*tags
= hctx
->sched_tags
;
6831 unsigned int min_shallow
;
6833 min_shallow
= bfq_update_depths(bfqd
, tags
->bitmap_tags
);
6834 sbitmap_queue_min_shallow_depth(tags
->bitmap_tags
, min_shallow
);
6837 static int bfq_init_hctx(struct blk_mq_hw_ctx
*hctx
, unsigned int index
)
6839 bfq_depth_updated(hctx
);
6843 static void bfq_exit_queue(struct elevator_queue
*e
)
6845 struct bfq_data
*bfqd
= e
->elevator_data
;
6846 struct bfq_queue
*bfqq
, *n
;
6848 hrtimer_cancel(&bfqd
->idle_slice_timer
);
6850 spin_lock_irq(&bfqd
->lock
);
6851 list_for_each_entry_safe(bfqq
, n
, &bfqd
->idle_list
, bfqq_list
)
6852 bfq_deactivate_bfqq(bfqd
, bfqq
, false, false);
6853 spin_unlock_irq(&bfqd
->lock
);
6855 hrtimer_cancel(&bfqd
->idle_slice_timer
);
6857 /* release oom-queue reference to root group */
6858 bfqg_and_blkg_put(bfqd
->root_group
);
6860 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6861 blkcg_deactivate_policy(bfqd
->queue
, &blkcg_policy_bfq
);
6863 spin_lock_irq(&bfqd
->lock
);
6864 bfq_put_async_queues(bfqd
, bfqd
->root_group
);
6865 kfree(bfqd
->root_group
);
6866 spin_unlock_irq(&bfqd
->lock
);
6872 static void bfq_init_root_group(struct bfq_group
*root_group
,
6873 struct bfq_data
*bfqd
)
6877 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6878 root_group
->entity
.parent
= NULL
;
6879 root_group
->my_entity
= NULL
;
6880 root_group
->bfqd
= bfqd
;
6882 root_group
->rq_pos_tree
= RB_ROOT
;
6883 for (i
= 0; i
< BFQ_IOPRIO_CLASSES
; i
++)
6884 root_group
->sched_data
.service_tree
[i
] = BFQ_SERVICE_TREE_INIT
;
6885 root_group
->sched_data
.bfq_class_idle_last_service
= jiffies
;
6888 static int bfq_init_queue(struct request_queue
*q
, struct elevator_type
*e
)
6890 struct bfq_data
*bfqd
;
6891 struct elevator_queue
*eq
;
6893 eq
= elevator_alloc(q
, e
);
6897 bfqd
= kzalloc_node(sizeof(*bfqd
), GFP_KERNEL
, q
->node
);
6899 kobject_put(&eq
->kobj
);
6902 eq
->elevator_data
= bfqd
;
6904 spin_lock_irq(&q
->queue_lock
);
6906 spin_unlock_irq(&q
->queue_lock
);
6909 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
6910 * Grab a permanent reference to it, so that the normal code flow
6911 * will not attempt to free it.
6913 bfq_init_bfqq(bfqd
, &bfqd
->oom_bfqq
, NULL
, 1, 0);
6914 bfqd
->oom_bfqq
.ref
++;
6915 bfqd
->oom_bfqq
.new_ioprio
= BFQ_DEFAULT_QUEUE_IOPRIO
;
6916 bfqd
->oom_bfqq
.new_ioprio_class
= IOPRIO_CLASS_BE
;
6917 bfqd
->oom_bfqq
.entity
.new_weight
=
6918 bfq_ioprio_to_weight(bfqd
->oom_bfqq
.new_ioprio
);
6920 /* oom_bfqq does not participate to bursts */
6921 bfq_clear_bfqq_just_created(&bfqd
->oom_bfqq
);
6924 * Trigger weight initialization, according to ioprio, at the
6925 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
6926 * class won't be changed any more.
6928 bfqd
->oom_bfqq
.entity
.prio_changed
= 1;
6932 INIT_LIST_HEAD(&bfqd
->dispatch
);
6934 hrtimer_init(&bfqd
->idle_slice_timer
, CLOCK_MONOTONIC
,
6936 bfqd
->idle_slice_timer
.function
= bfq_idle_slice_timer
;
6938 bfqd
->queue_weights_tree
= RB_ROOT_CACHED
;
6939 bfqd
->num_groups_with_pending_reqs
= 0;
6941 INIT_LIST_HEAD(&bfqd
->active_list
);
6942 INIT_LIST_HEAD(&bfqd
->idle_list
);
6943 INIT_HLIST_HEAD(&bfqd
->burst_list
);
6946 bfqd
->nonrot_with_queueing
= blk_queue_nonrot(bfqd
->queue
);
6948 bfqd
->bfq_max_budget
= bfq_default_max_budget
;
6950 bfqd
->bfq_fifo_expire
[0] = bfq_fifo_expire
[0];
6951 bfqd
->bfq_fifo_expire
[1] = bfq_fifo_expire
[1];
6952 bfqd
->bfq_back_max
= bfq_back_max
;
6953 bfqd
->bfq_back_penalty
= bfq_back_penalty
;
6954 bfqd
->bfq_slice_idle
= bfq_slice_idle
;
6955 bfqd
->bfq_timeout
= bfq_timeout
;
6957 bfqd
->bfq_large_burst_thresh
= 8;
6958 bfqd
->bfq_burst_interval
= msecs_to_jiffies(180);
6960 bfqd
->low_latency
= true;
6963 * Trade-off between responsiveness and fairness.
6965 bfqd
->bfq_wr_coeff
= 30;
6966 bfqd
->bfq_wr_rt_max_time
= msecs_to_jiffies(300);
6967 bfqd
->bfq_wr_max_time
= 0;
6968 bfqd
->bfq_wr_min_idle_time
= msecs_to_jiffies(2000);
6969 bfqd
->bfq_wr_min_inter_arr_async
= msecs_to_jiffies(500);
6970 bfqd
->bfq_wr_max_softrt_rate
= 7000; /*
6971 * Approximate rate required
6972 * to playback or record a
6973 * high-definition compressed
6976 bfqd
->wr_busy_queues
= 0;
6979 * Begin by assuming, optimistically, that the device peak
6980 * rate is equal to 2/3 of the highest reference rate.
6982 bfqd
->rate_dur_prod
= ref_rate
[blk_queue_nonrot(bfqd
->queue
)] *
6983 ref_wr_duration
[blk_queue_nonrot(bfqd
->queue
)];
6984 bfqd
->peak_rate
= ref_rate
[blk_queue_nonrot(bfqd
->queue
)] * 2 / 3;
6986 spin_lock_init(&bfqd
->lock
);
6989 * The invocation of the next bfq_create_group_hierarchy
6990 * function is the head of a chain of function calls
6991 * (bfq_create_group_hierarchy->blkcg_activate_policy->
6992 * blk_mq_freeze_queue) that may lead to the invocation of the
6993 * has_work hook function. For this reason,
6994 * bfq_create_group_hierarchy is invoked only after all
6995 * scheduler data has been initialized, apart from the fields
6996 * that can be initialized only after invoking
6997 * bfq_create_group_hierarchy. This, in particular, enables
6998 * has_work to correctly return false. Of course, to avoid
6999 * other inconsistencies, the blk-mq stack must then refrain
7000 * from invoking further scheduler hooks before this init
7001 * function is finished.
7003 bfqd
->root_group
= bfq_create_group_hierarchy(bfqd
, q
->node
);
7004 if (!bfqd
->root_group
)
7006 bfq_init_root_group(bfqd
->root_group
, bfqd
);
7007 bfq_init_entity(&bfqd
->oom_bfqq
.entity
, bfqd
->root_group
);
7009 wbt_disable_default(q
);
7014 kobject_put(&eq
->kobj
);
7018 static void bfq_slab_kill(void)
7020 kmem_cache_destroy(bfq_pool
);
7023 static int __init
bfq_slab_setup(void)
7025 bfq_pool
= KMEM_CACHE(bfq_queue
, 0);
7031 static ssize_t
bfq_var_show(unsigned int var
, char *page
)
7033 return sprintf(page
, "%u\n", var
);
7036 static int bfq_var_store(unsigned long *var
, const char *page
)
7038 unsigned long new_val
;
7039 int ret
= kstrtoul(page
, 10, &new_val
);
7047 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
7048 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7050 struct bfq_data *bfqd = e->elevator_data; \
7051 u64 __data = __VAR; \
7053 __data = jiffies_to_msecs(__data); \
7054 else if (__CONV == 2) \
7055 __data = div_u64(__data, NSEC_PER_MSEC); \
7056 return bfq_var_show(__data, (page)); \
7058 SHOW_FUNCTION(bfq_fifo_expire_sync_show
, bfqd
->bfq_fifo_expire
[1], 2);
7059 SHOW_FUNCTION(bfq_fifo_expire_async_show
, bfqd
->bfq_fifo_expire
[0], 2);
7060 SHOW_FUNCTION(bfq_back_seek_max_show
, bfqd
->bfq_back_max
, 0);
7061 SHOW_FUNCTION(bfq_back_seek_penalty_show
, bfqd
->bfq_back_penalty
, 0);
7062 SHOW_FUNCTION(bfq_slice_idle_show
, bfqd
->bfq_slice_idle
, 2);
7063 SHOW_FUNCTION(bfq_max_budget_show
, bfqd
->bfq_user_max_budget
, 0);
7064 SHOW_FUNCTION(bfq_timeout_sync_show
, bfqd
->bfq_timeout
, 1);
7065 SHOW_FUNCTION(bfq_strict_guarantees_show
, bfqd
->strict_guarantees
, 0);
7066 SHOW_FUNCTION(bfq_low_latency_show
, bfqd
->low_latency
, 0);
7067 #undef SHOW_FUNCTION
7069 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
7070 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7072 struct bfq_data *bfqd = e->elevator_data; \
7073 u64 __data = __VAR; \
7074 __data = div_u64(__data, NSEC_PER_USEC); \
7075 return bfq_var_show(__data, (page)); \
7077 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show
, bfqd
->bfq_slice_idle
);
7078 #undef USEC_SHOW_FUNCTION
7080 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
7082 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
7084 struct bfq_data *bfqd = e->elevator_data; \
7085 unsigned long __data, __min = (MIN), __max = (MAX); \
7088 ret = bfq_var_store(&__data, (page)); \
7091 if (__data < __min) \
7093 else if (__data > __max) \
7096 *(__PTR) = msecs_to_jiffies(__data); \
7097 else if (__CONV == 2) \
7098 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
7100 *(__PTR) = __data; \
7103 STORE_FUNCTION(bfq_fifo_expire_sync_store
, &bfqd
->bfq_fifo_expire
[1], 1,
7105 STORE_FUNCTION(bfq_fifo_expire_async_store
, &bfqd
->bfq_fifo_expire
[0], 1,
7107 STORE_FUNCTION(bfq_back_seek_max_store
, &bfqd
->bfq_back_max
, 0, INT_MAX
, 0);
7108 STORE_FUNCTION(bfq_back_seek_penalty_store
, &bfqd
->bfq_back_penalty
, 1,
7110 STORE_FUNCTION(bfq_slice_idle_store
, &bfqd
->bfq_slice_idle
, 0, INT_MAX
, 2);
7111 #undef STORE_FUNCTION
7113 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
7114 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
7116 struct bfq_data *bfqd = e->elevator_data; \
7117 unsigned long __data, __min = (MIN), __max = (MAX); \
7120 ret = bfq_var_store(&__data, (page)); \
7123 if (__data < __min) \
7125 else if (__data > __max) \
7127 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
7130 USEC_STORE_FUNCTION(bfq_slice_idle_us_store
, &bfqd
->bfq_slice_idle
, 0,
7132 #undef USEC_STORE_FUNCTION
7134 static ssize_t
bfq_max_budget_store(struct elevator_queue
*e
,
7135 const char *page
, size_t count
)
7137 struct bfq_data
*bfqd
= e
->elevator_data
;
7138 unsigned long __data
;
7141 ret
= bfq_var_store(&__data
, (page
));
7146 bfqd
->bfq_max_budget
= bfq_calc_max_budget(bfqd
);
7148 if (__data
> INT_MAX
)
7150 bfqd
->bfq_max_budget
= __data
;
7153 bfqd
->bfq_user_max_budget
= __data
;
7159 * Leaving this name to preserve name compatibility with cfq
7160 * parameters, but this timeout is used for both sync and async.
7162 static ssize_t
bfq_timeout_sync_store(struct elevator_queue
*e
,
7163 const char *page
, size_t count
)
7165 struct bfq_data
*bfqd
= e
->elevator_data
;
7166 unsigned long __data
;
7169 ret
= bfq_var_store(&__data
, (page
));
7175 else if (__data
> INT_MAX
)
7178 bfqd
->bfq_timeout
= msecs_to_jiffies(__data
);
7179 if (bfqd
->bfq_user_max_budget
== 0)
7180 bfqd
->bfq_max_budget
= bfq_calc_max_budget(bfqd
);
7185 static ssize_t
bfq_strict_guarantees_store(struct elevator_queue
*e
,
7186 const char *page
, size_t count
)
7188 struct bfq_data
*bfqd
= e
->elevator_data
;
7189 unsigned long __data
;
7192 ret
= bfq_var_store(&__data
, (page
));
7198 if (!bfqd
->strict_guarantees
&& __data
== 1
7199 && bfqd
->bfq_slice_idle
< 8 * NSEC_PER_MSEC
)
7200 bfqd
->bfq_slice_idle
= 8 * NSEC_PER_MSEC
;
7202 bfqd
->strict_guarantees
= __data
;
7207 static ssize_t
bfq_low_latency_store(struct elevator_queue
*e
,
7208 const char *page
, size_t count
)
7210 struct bfq_data
*bfqd
= e
->elevator_data
;
7211 unsigned long __data
;
7214 ret
= bfq_var_store(&__data
, (page
));
7220 if (__data
== 0 && bfqd
->low_latency
!= 0)
7222 bfqd
->low_latency
= __data
;
7227 #define BFQ_ATTR(name) \
7228 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
7230 static struct elv_fs_entry bfq_attrs
[] = {
7231 BFQ_ATTR(fifo_expire_sync
),
7232 BFQ_ATTR(fifo_expire_async
),
7233 BFQ_ATTR(back_seek_max
),
7234 BFQ_ATTR(back_seek_penalty
),
7235 BFQ_ATTR(slice_idle
),
7236 BFQ_ATTR(slice_idle_us
),
7237 BFQ_ATTR(max_budget
),
7238 BFQ_ATTR(timeout_sync
),
7239 BFQ_ATTR(strict_guarantees
),
7240 BFQ_ATTR(low_latency
),
7244 static struct elevator_type iosched_bfq_mq
= {
7246 .limit_depth
= bfq_limit_depth
,
7247 .prepare_request
= bfq_prepare_request
,
7248 .requeue_request
= bfq_finish_requeue_request
,
7249 .finish_request
= bfq_finish_requeue_request
,
7250 .exit_icq
= bfq_exit_icq
,
7251 .insert_requests
= bfq_insert_requests
,
7252 .dispatch_request
= bfq_dispatch_request
,
7253 .next_request
= elv_rb_latter_request
,
7254 .former_request
= elv_rb_former_request
,
7255 .allow_merge
= bfq_allow_bio_merge
,
7256 .bio_merge
= bfq_bio_merge
,
7257 .request_merge
= bfq_request_merge
,
7258 .requests_merged
= bfq_requests_merged
,
7259 .request_merged
= bfq_request_merged
,
7260 .has_work
= bfq_has_work
,
7261 .depth_updated
= bfq_depth_updated
,
7262 .init_hctx
= bfq_init_hctx
,
7263 .init_sched
= bfq_init_queue
,
7264 .exit_sched
= bfq_exit_queue
,
7267 .icq_size
= sizeof(struct bfq_io_cq
),
7268 .icq_align
= __alignof__(struct bfq_io_cq
),
7269 .elevator_attrs
= bfq_attrs
,
7270 .elevator_name
= "bfq",
7271 .elevator_owner
= THIS_MODULE
,
7273 MODULE_ALIAS("bfq-iosched");
7275 static int __init
bfq_init(void)
7279 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7280 ret
= blkcg_policy_register(&blkcg_policy_bfq
);
7286 if (bfq_slab_setup())
7290 * Times to load large popular applications for the typical
7291 * systems installed on the reference devices (see the
7292 * comments before the definition of the next
7293 * array). Actually, we use slightly lower values, as the
7294 * estimated peak rate tends to be smaller than the actual
7295 * peak rate. The reason for this last fact is that estimates
7296 * are computed over much shorter time intervals than the long
7297 * intervals typically used for benchmarking. Why? First, to
7298 * adapt more quickly to variations. Second, because an I/O
7299 * scheduler cannot rely on a peak-rate-evaluation workload to
7300 * be run for a long time.
7302 ref_wr_duration
[0] = msecs_to_jiffies(7000); /* actually 8 sec */
7303 ref_wr_duration
[1] = msecs_to_jiffies(2500); /* actually 3 sec */
7305 ret
= elv_register(&iosched_bfq_mq
);
7314 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7315 blkcg_policy_unregister(&blkcg_policy_bfq
);
7320 static void __exit
bfq_exit(void)
7322 elv_unregister(&iosched_bfq_mq
);
7323 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7324 blkcg_policy_unregister(&blkcg_policy_bfq
);
7329 module_init(bfq_init
);
7330 module_exit(bfq_exit
);
7332 MODULE_AUTHOR("Paolo Valente");
7333 MODULE_LICENSE("GPL");
7334 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");