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;
368 * Maximum time between the creation of two queues, for stable merge
369 * to be activated (in ms)
371 static const unsigned long bfq_activation_stable_merging
= 600;
373 * Minimum time to be waited before evaluating delayed stable merge (in ms)
375 static const unsigned long bfq_late_stable_merging
= 600;
377 #define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
378 #define RQ_BFQQ(rq) ((rq)->elv.priv[1])
380 struct bfq_queue
*bic_to_bfqq(struct bfq_io_cq
*bic
, bool is_sync
)
382 return bic
->bfqq
[is_sync
];
385 static void bfq_put_stable_ref(struct bfq_queue
*bfqq
);
387 void bic_set_bfqq(struct bfq_io_cq
*bic
, struct bfq_queue
*bfqq
, bool is_sync
)
390 * If bfqq != NULL, then a non-stable queue merge between
391 * bic->bfqq and bfqq is happening here. This causes troubles
392 * in the following case: bic->bfqq has also been scheduled
393 * for a possible stable merge with bic->stable_merge_bfqq,
394 * and bic->stable_merge_bfqq == bfqq happens to
395 * hold. Troubles occur because bfqq may then undergo a split,
396 * thereby becoming eligible for a stable merge. Yet, if
397 * bic->stable_merge_bfqq points exactly to bfqq, then bfqq
398 * would be stably merged with itself. To avoid this anomaly,
399 * we cancel the stable merge if
400 * bic->stable_merge_bfqq == bfqq.
402 bic
->bfqq
[is_sync
] = bfqq
;
404 if (bfqq
&& bic
->stable_merge_bfqq
== bfqq
) {
406 * Actually, these same instructions are executed also
407 * in bfq_setup_cooperator, in case of abort or actual
408 * execution of a stable merge. We could avoid
409 * repeating these instructions there too, but if we
410 * did so, we would nest even more complexity in this
413 bfq_put_stable_ref(bic
->stable_merge_bfqq
);
415 bic
->stable_merge_bfqq
= NULL
;
419 struct bfq_data
*bic_to_bfqd(struct bfq_io_cq
*bic
)
421 return bic
->icq
.q
->elevator
->elevator_data
;
425 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
426 * @icq: the iocontext queue.
428 static struct bfq_io_cq
*icq_to_bic(struct io_cq
*icq
)
430 /* bic->icq is the first member, %NULL will convert to %NULL */
431 return container_of(icq
, struct bfq_io_cq
, icq
);
435 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
436 * @bfqd: the lookup key.
437 * @ioc: the io_context of the process doing I/O.
438 * @q: the request queue.
440 static struct bfq_io_cq
*bfq_bic_lookup(struct bfq_data
*bfqd
,
441 struct io_context
*ioc
,
442 struct request_queue
*q
)
446 struct bfq_io_cq
*icq
;
448 spin_lock_irqsave(&q
->queue_lock
, flags
);
449 icq
= icq_to_bic(ioc_lookup_icq(ioc
, q
));
450 spin_unlock_irqrestore(&q
->queue_lock
, flags
);
459 * Scheduler run of queue, if there are requests pending and no one in the
460 * driver that will restart queueing.
462 void bfq_schedule_dispatch(struct bfq_data
*bfqd
)
464 lockdep_assert_held(&bfqd
->lock
);
466 if (bfqd
->queued
!= 0) {
467 bfq_log(bfqd
, "schedule dispatch");
468 blk_mq_run_hw_queues(bfqd
->queue
, true);
472 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
474 #define bfq_sample_valid(samples) ((samples) > 80)
477 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
478 * We choose the request that is closer to the head right now. Distance
479 * behind the head is penalized and only allowed to a certain extent.
481 static struct request
*bfq_choose_req(struct bfq_data
*bfqd
,
486 sector_t s1
, s2
, d1
= 0, d2
= 0;
487 unsigned long back_max
;
488 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
489 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
490 unsigned int wrap
= 0; /* bit mask: requests behind the disk head? */
492 if (!rq1
|| rq1
== rq2
)
497 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
499 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
501 if ((rq1
->cmd_flags
& REQ_META
) && !(rq2
->cmd_flags
& REQ_META
))
503 else if ((rq2
->cmd_flags
& REQ_META
) && !(rq1
->cmd_flags
& REQ_META
))
506 s1
= blk_rq_pos(rq1
);
507 s2
= blk_rq_pos(rq2
);
510 * By definition, 1KiB is 2 sectors.
512 back_max
= bfqd
->bfq_back_max
* 2;
515 * Strict one way elevator _except_ in the case where we allow
516 * short backward seeks which are biased as twice the cost of a
517 * similar forward seek.
521 else if (s1
+ back_max
>= last
)
522 d1
= (last
- s1
) * bfqd
->bfq_back_penalty
;
524 wrap
|= BFQ_RQ1_WRAP
;
528 else if (s2
+ back_max
>= last
)
529 d2
= (last
- s2
) * bfqd
->bfq_back_penalty
;
531 wrap
|= BFQ_RQ2_WRAP
;
533 /* Found required data */
536 * By doing switch() on the bit mask "wrap" we avoid having to
537 * check two variables for all permutations: --> faster!
540 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
555 case BFQ_RQ1_WRAP
|BFQ_RQ2_WRAP
: /* both rqs wrapped */
558 * Since both rqs are wrapped,
559 * start with the one that's further behind head
560 * (--> only *one* back seek required),
561 * since back seek takes more time than forward.
571 * Async I/O can easily starve sync I/O (both sync reads and sync
572 * writes), by consuming all tags. Similarly, storms of sync writes,
573 * such as those that sync(2) may trigger, can starve sync reads.
574 * Limit depths of async I/O and sync writes so as to counter both
577 static void bfq_limit_depth(unsigned int op
, struct blk_mq_alloc_data
*data
)
579 struct bfq_data
*bfqd
= data
->q
->elevator
->elevator_data
;
581 if (op_is_sync(op
) && !op_is_write(op
))
584 data
->shallow_depth
=
585 bfqd
->word_depths
[!!bfqd
->wr_busy_queues
][op_is_sync(op
)];
587 bfq_log(bfqd
, "[%s] wr_busy %d sync %d depth %u",
588 __func__
, bfqd
->wr_busy_queues
, op_is_sync(op
),
589 data
->shallow_depth
);
592 static struct bfq_queue
*
593 bfq_rq_pos_tree_lookup(struct bfq_data
*bfqd
, struct rb_root
*root
,
594 sector_t sector
, struct rb_node
**ret_parent
,
595 struct rb_node
***rb_link
)
597 struct rb_node
**p
, *parent
;
598 struct bfq_queue
*bfqq
= NULL
;
606 bfqq
= rb_entry(parent
, struct bfq_queue
, pos_node
);
609 * Sort strictly based on sector. Smallest to the left,
610 * largest to the right.
612 if (sector
> blk_rq_pos(bfqq
->next_rq
))
614 else if (sector
< blk_rq_pos(bfqq
->next_rq
))
622 *ret_parent
= parent
;
626 bfq_log(bfqd
, "rq_pos_tree_lookup %llu: returning %d",
627 (unsigned long long)sector
,
628 bfqq
? bfqq
->pid
: 0);
633 static bool bfq_too_late_for_merging(struct bfq_queue
*bfqq
)
635 return bfqq
->service_from_backlogged
> 0 &&
636 time_is_before_jiffies(bfqq
->first_IO_time
+
637 bfq_merge_time_limit
);
641 * The following function is not marked as __cold because it is
642 * actually cold, but for the same performance goal described in the
643 * comments on the likely() at the beginning of
644 * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
645 * execution time for the case where this function is not invoked, we
646 * had to add an unlikely() in each involved if().
649 bfq_pos_tree_add_move(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
651 struct rb_node
**p
, *parent
;
652 struct bfq_queue
*__bfqq
;
654 if (bfqq
->pos_root
) {
655 rb_erase(&bfqq
->pos_node
, bfqq
->pos_root
);
656 bfqq
->pos_root
= NULL
;
659 /* oom_bfqq does not participate in queue merging */
660 if (bfqq
== &bfqd
->oom_bfqq
)
664 * bfqq cannot be merged any longer (see comments in
665 * bfq_setup_cooperator): no point in adding bfqq into the
668 if (bfq_too_late_for_merging(bfqq
))
671 if (bfq_class_idle(bfqq
))
676 bfqq
->pos_root
= &bfq_bfqq_to_bfqg(bfqq
)->rq_pos_tree
;
677 __bfqq
= bfq_rq_pos_tree_lookup(bfqd
, bfqq
->pos_root
,
678 blk_rq_pos(bfqq
->next_rq
), &parent
, &p
);
680 rb_link_node(&bfqq
->pos_node
, parent
, p
);
681 rb_insert_color(&bfqq
->pos_node
, bfqq
->pos_root
);
683 bfqq
->pos_root
= NULL
;
687 * The following function returns false either if every active queue
688 * must receive the same share of the throughput (symmetric scenario),
689 * or, as a special case, if bfqq must receive a share of the
690 * throughput lower than or equal to the share that every other active
691 * queue must receive. If bfqq does sync I/O, then these are the only
692 * two cases where bfqq happens to be guaranteed its share of the
693 * throughput even if I/O dispatching is not plugged when bfqq remains
694 * temporarily empty (for more details, see the comments in the
695 * function bfq_better_to_idle()). For this reason, the return value
696 * of this function is used to check whether I/O-dispatch plugging can
699 * The above first case (symmetric scenario) occurs when:
700 * 1) all active queues have the same weight,
701 * 2) all active queues belong to the same I/O-priority class,
702 * 3) all active groups at the same level in the groups tree have the same
704 * 4) all active groups at the same level in the groups tree have the same
705 * number of children.
707 * Unfortunately, keeping the necessary state for evaluating exactly
708 * the last two symmetry sub-conditions above would be quite complex
709 * and time consuming. Therefore this function evaluates, instead,
710 * only the following stronger three sub-conditions, for which it is
711 * much easier to maintain the needed state:
712 * 1) all active queues have the same weight,
713 * 2) all active queues belong to the same I/O-priority class,
714 * 3) there are no active groups.
715 * In particular, the last condition is always true if hierarchical
716 * support or the cgroups interface are not enabled, thus no state
717 * needs to be maintained in this case.
719 static bool bfq_asymmetric_scenario(struct bfq_data
*bfqd
,
720 struct bfq_queue
*bfqq
)
722 bool smallest_weight
= bfqq
&&
723 bfqq
->weight_counter
&&
724 bfqq
->weight_counter
==
726 rb_first_cached(&bfqd
->queue_weights_tree
),
727 struct bfq_weight_counter
,
731 * For queue weights to differ, queue_weights_tree must contain
732 * at least two nodes.
734 bool varied_queue_weights
= !smallest_weight
&&
735 !RB_EMPTY_ROOT(&bfqd
->queue_weights_tree
.rb_root
) &&
736 (bfqd
->queue_weights_tree
.rb_root
.rb_node
->rb_left
||
737 bfqd
->queue_weights_tree
.rb_root
.rb_node
->rb_right
);
739 bool multiple_classes_busy
=
740 (bfqd
->busy_queues
[0] && bfqd
->busy_queues
[1]) ||
741 (bfqd
->busy_queues
[0] && bfqd
->busy_queues
[2]) ||
742 (bfqd
->busy_queues
[1] && bfqd
->busy_queues
[2]);
744 return varied_queue_weights
|| multiple_classes_busy
745 #ifdef CONFIG_BFQ_GROUP_IOSCHED
746 || bfqd
->num_groups_with_pending_reqs
> 0
752 * If the weight-counter tree passed as input contains no counter for
753 * the weight of the input queue, then add that counter; otherwise just
754 * increment the existing counter.
756 * Note that weight-counter trees contain few nodes in mostly symmetric
757 * scenarios. For example, if all queues have the same weight, then the
758 * weight-counter tree for the queues may contain at most one node.
759 * This holds even if low_latency is on, because weight-raised queues
760 * are not inserted in the tree.
761 * In most scenarios, the rate at which nodes are created/destroyed
764 void bfq_weights_tree_add(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
765 struct rb_root_cached
*root
)
767 struct bfq_entity
*entity
= &bfqq
->entity
;
768 struct rb_node
**new = &(root
->rb_root
.rb_node
), *parent
= NULL
;
769 bool leftmost
= true;
772 * Do not insert if the queue is already associated with a
773 * counter, which happens if:
774 * 1) a request arrival has caused the queue to become both
775 * non-weight-raised, and hence change its weight, and
776 * backlogged; in this respect, each of the two events
777 * causes an invocation of this function,
778 * 2) this is the invocation of this function caused by the
779 * second event. This second invocation is actually useless,
780 * and we handle this fact by exiting immediately. More
781 * efficient or clearer solutions might possibly be adopted.
783 if (bfqq
->weight_counter
)
787 struct bfq_weight_counter
*__counter
= container_of(*new,
788 struct bfq_weight_counter
,
792 if (entity
->weight
== __counter
->weight
) {
793 bfqq
->weight_counter
= __counter
;
796 if (entity
->weight
< __counter
->weight
)
797 new = &((*new)->rb_left
);
799 new = &((*new)->rb_right
);
804 bfqq
->weight_counter
= kzalloc(sizeof(struct bfq_weight_counter
),
808 * In the unlucky event of an allocation failure, we just
809 * exit. This will cause the weight of queue to not be
810 * considered in bfq_asymmetric_scenario, which, in its turn,
811 * causes the scenario to be deemed wrongly symmetric in case
812 * bfqq's weight would have been the only weight making the
813 * scenario asymmetric. On the bright side, no unbalance will
814 * however occur when bfqq becomes inactive again (the
815 * invocation of this function is triggered by an activation
816 * of queue). In fact, bfq_weights_tree_remove does nothing
817 * if !bfqq->weight_counter.
819 if (unlikely(!bfqq
->weight_counter
))
822 bfqq
->weight_counter
->weight
= entity
->weight
;
823 rb_link_node(&bfqq
->weight_counter
->weights_node
, parent
, new);
824 rb_insert_color_cached(&bfqq
->weight_counter
->weights_node
, root
,
828 bfqq
->weight_counter
->num_active
++;
833 * Decrement the weight counter associated with the queue, and, if the
834 * counter reaches 0, remove the counter from the tree.
835 * See the comments to the function bfq_weights_tree_add() for considerations
838 void __bfq_weights_tree_remove(struct bfq_data
*bfqd
,
839 struct bfq_queue
*bfqq
,
840 struct rb_root_cached
*root
)
842 if (!bfqq
->weight_counter
)
845 bfqq
->weight_counter
->num_active
--;
846 if (bfqq
->weight_counter
->num_active
> 0)
847 goto reset_entity_pointer
;
849 rb_erase_cached(&bfqq
->weight_counter
->weights_node
, root
);
850 kfree(bfqq
->weight_counter
);
852 reset_entity_pointer
:
853 bfqq
->weight_counter
= NULL
;
858 * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
859 * of active groups for each queue's inactive parent entity.
861 void bfq_weights_tree_remove(struct bfq_data
*bfqd
,
862 struct bfq_queue
*bfqq
)
864 struct bfq_entity
*entity
= bfqq
->entity
.parent
;
866 for_each_entity(entity
) {
867 struct bfq_sched_data
*sd
= entity
->my_sched_data
;
869 if (sd
->next_in_service
|| sd
->in_service_entity
) {
871 * entity is still active, because either
872 * next_in_service or in_service_entity is not
873 * NULL (see the comments on the definition of
874 * next_in_service for details on why
875 * in_service_entity must be checked too).
877 * As a consequence, its parent entities are
878 * active as well, and thus this loop must
885 * The decrement of num_groups_with_pending_reqs is
886 * not performed immediately upon the deactivation of
887 * entity, but it is delayed to when it also happens
888 * that the first leaf descendant bfqq of entity gets
889 * all its pending requests completed. The following
890 * instructions perform this delayed decrement, if
891 * needed. See the comments on
892 * num_groups_with_pending_reqs for details.
894 if (entity
->in_groups_with_pending_reqs
) {
895 entity
->in_groups_with_pending_reqs
= false;
896 bfqd
->num_groups_with_pending_reqs
--;
901 * Next function is invoked last, because it causes bfqq to be
902 * freed if the following holds: bfqq is not in service and
903 * has no dispatched request. DO NOT use bfqq after the next
904 * function invocation.
906 __bfq_weights_tree_remove(bfqd
, bfqq
,
907 &bfqd
->queue_weights_tree
);
911 * Return expired entry, or NULL to just start from scratch in rbtree.
913 static struct request
*bfq_check_fifo(struct bfq_queue
*bfqq
,
914 struct request
*last
)
918 if (bfq_bfqq_fifo_expire(bfqq
))
921 bfq_mark_bfqq_fifo_expire(bfqq
);
923 rq
= rq_entry_fifo(bfqq
->fifo
.next
);
925 if (rq
== last
|| ktime_get_ns() < rq
->fifo_time
)
928 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "check_fifo: returned %p", rq
);
932 static struct request
*bfq_find_next_rq(struct bfq_data
*bfqd
,
933 struct bfq_queue
*bfqq
,
934 struct request
*last
)
936 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
937 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
938 struct request
*next
, *prev
= NULL
;
940 /* Follow expired path, else get first next available. */
941 next
= bfq_check_fifo(bfqq
, last
);
946 prev
= rb_entry_rq(rbprev
);
949 next
= rb_entry_rq(rbnext
);
951 rbnext
= rb_first(&bfqq
->sort_list
);
952 if (rbnext
&& rbnext
!= &last
->rb_node
)
953 next
= rb_entry_rq(rbnext
);
956 return bfq_choose_req(bfqd
, next
, prev
, blk_rq_pos(last
));
959 /* see the definition of bfq_async_charge_factor for details */
960 static unsigned long bfq_serv_to_charge(struct request
*rq
,
961 struct bfq_queue
*bfqq
)
963 if (bfq_bfqq_sync(bfqq
) || bfqq
->wr_coeff
> 1 ||
964 bfq_asymmetric_scenario(bfqq
->bfqd
, bfqq
))
965 return blk_rq_sectors(rq
);
967 return blk_rq_sectors(rq
) * bfq_async_charge_factor
;
971 * bfq_updated_next_req - update the queue after a new next_rq selection.
972 * @bfqd: the device data the queue belongs to.
973 * @bfqq: the queue to update.
975 * If the first request of a queue changes we make sure that the queue
976 * has enough budget to serve at least its first request (if the
977 * request has grown). We do this because if the queue has not enough
978 * budget for its first request, it has to go through two dispatch
979 * rounds to actually get it dispatched.
981 static void bfq_updated_next_req(struct bfq_data
*bfqd
,
982 struct bfq_queue
*bfqq
)
984 struct bfq_entity
*entity
= &bfqq
->entity
;
985 struct request
*next_rq
= bfqq
->next_rq
;
986 unsigned long new_budget
;
991 if (bfqq
== bfqd
->in_service_queue
)
993 * In order not to break guarantees, budgets cannot be
994 * changed after an entity has been selected.
998 new_budget
= max_t(unsigned long,
999 max_t(unsigned long, bfqq
->max_budget
,
1000 bfq_serv_to_charge(next_rq
, bfqq
)),
1002 if (entity
->budget
!= new_budget
) {
1003 entity
->budget
= new_budget
;
1004 bfq_log_bfqq(bfqd
, bfqq
, "updated next rq: new budget %lu",
1006 bfq_requeue_bfqq(bfqd
, bfqq
, false);
1010 static unsigned int bfq_wr_duration(struct bfq_data
*bfqd
)
1014 if (bfqd
->bfq_wr_max_time
> 0)
1015 return bfqd
->bfq_wr_max_time
;
1017 dur
= bfqd
->rate_dur_prod
;
1018 do_div(dur
, bfqd
->peak_rate
);
1021 * Limit duration between 3 and 25 seconds. The upper limit
1022 * has been conservatively set after the following worst case:
1023 * on a QEMU/KVM virtual machine
1024 * - running in a slow PC
1025 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
1026 * - serving a heavy I/O workload, such as the sequential reading
1028 * mplayer took 23 seconds to start, if constantly weight-raised.
1030 * As for higher values than that accommodating the above bad
1031 * scenario, tests show that higher values would often yield
1032 * the opposite of the desired result, i.e., would worsen
1033 * responsiveness by allowing non-interactive applications to
1034 * preserve weight raising for too long.
1036 * On the other end, lower values than 3 seconds make it
1037 * difficult for most interactive tasks to complete their jobs
1038 * before weight-raising finishes.
1040 return clamp_val(dur
, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
1043 /* switch back from soft real-time to interactive weight raising */
1044 static void switch_back_to_interactive_wr(struct bfq_queue
*bfqq
,
1045 struct bfq_data
*bfqd
)
1047 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1048 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1049 bfqq
->last_wr_start_finish
= bfqq
->wr_start_at_switch_to_srt
;
1053 bfq_bfqq_resume_state(struct bfq_queue
*bfqq
, struct bfq_data
*bfqd
,
1054 struct bfq_io_cq
*bic
, bool bfq_already_existing
)
1056 unsigned int old_wr_coeff
= 1;
1057 bool busy
= bfq_already_existing
&& bfq_bfqq_busy(bfqq
);
1059 if (bic
->saved_has_short_ttime
)
1060 bfq_mark_bfqq_has_short_ttime(bfqq
);
1062 bfq_clear_bfqq_has_short_ttime(bfqq
);
1064 if (bic
->saved_IO_bound
)
1065 bfq_mark_bfqq_IO_bound(bfqq
);
1067 bfq_clear_bfqq_IO_bound(bfqq
);
1069 bfqq
->last_serv_time_ns
= bic
->saved_last_serv_time_ns
;
1070 bfqq
->inject_limit
= bic
->saved_inject_limit
;
1071 bfqq
->decrease_time_jif
= bic
->saved_decrease_time_jif
;
1073 bfqq
->entity
.new_weight
= bic
->saved_weight
;
1074 bfqq
->ttime
= bic
->saved_ttime
;
1075 bfqq
->io_start_time
= bic
->saved_io_start_time
;
1076 bfqq
->tot_idle_time
= bic
->saved_tot_idle_time
;
1078 * Restore weight coefficient only if low_latency is on
1080 if (bfqd
->low_latency
) {
1081 old_wr_coeff
= bfqq
->wr_coeff
;
1082 bfqq
->wr_coeff
= bic
->saved_wr_coeff
;
1084 bfqq
->service_from_wr
= bic
->saved_service_from_wr
;
1085 bfqq
->wr_start_at_switch_to_srt
= bic
->saved_wr_start_at_switch_to_srt
;
1086 bfqq
->last_wr_start_finish
= bic
->saved_last_wr_start_finish
;
1087 bfqq
->wr_cur_max_time
= bic
->saved_wr_cur_max_time
;
1089 if (bfqq
->wr_coeff
> 1 && (bfq_bfqq_in_large_burst(bfqq
) ||
1090 time_is_before_jiffies(bfqq
->last_wr_start_finish
+
1091 bfqq
->wr_cur_max_time
))) {
1092 if (bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
&&
1093 !bfq_bfqq_in_large_burst(bfqq
) &&
1094 time_is_after_eq_jiffies(bfqq
->wr_start_at_switch_to_srt
+
1095 bfq_wr_duration(bfqd
))) {
1096 switch_back_to_interactive_wr(bfqq
, bfqd
);
1099 bfq_log_bfqq(bfqq
->bfqd
, bfqq
,
1100 "resume state: switching off wr");
1104 /* make sure weight will be updated, however we got here */
1105 bfqq
->entity
.prio_changed
= 1;
1110 if (old_wr_coeff
== 1 && bfqq
->wr_coeff
> 1)
1111 bfqd
->wr_busy_queues
++;
1112 else if (old_wr_coeff
> 1 && bfqq
->wr_coeff
== 1)
1113 bfqd
->wr_busy_queues
--;
1116 static int bfqq_process_refs(struct bfq_queue
*bfqq
)
1118 return bfqq
->ref
- bfqq
->allocated
- bfqq
->entity
.on_st_or_in_serv
-
1119 (bfqq
->weight_counter
!= NULL
) - bfqq
->stable_ref
;
1122 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1123 static void bfq_reset_burst_list(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
1125 struct bfq_queue
*item
;
1126 struct hlist_node
*n
;
1128 hlist_for_each_entry_safe(item
, n
, &bfqd
->burst_list
, burst_list_node
)
1129 hlist_del_init(&item
->burst_list_node
);
1132 * Start the creation of a new burst list only if there is no
1133 * active queue. See comments on the conditional invocation of
1134 * bfq_handle_burst().
1136 if (bfq_tot_busy_queues(bfqd
) == 0) {
1137 hlist_add_head(&bfqq
->burst_list_node
, &bfqd
->burst_list
);
1138 bfqd
->burst_size
= 1;
1140 bfqd
->burst_size
= 0;
1142 bfqd
->burst_parent_entity
= bfqq
->entity
.parent
;
1145 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1146 static void bfq_add_to_burst(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
1148 /* Increment burst size to take into account also bfqq */
1151 if (bfqd
->burst_size
== bfqd
->bfq_large_burst_thresh
) {
1152 struct bfq_queue
*pos
, *bfqq_item
;
1153 struct hlist_node
*n
;
1156 * Enough queues have been activated shortly after each
1157 * other to consider this burst as large.
1159 bfqd
->large_burst
= true;
1162 * We can now mark all queues in the burst list as
1163 * belonging to a large burst.
1165 hlist_for_each_entry(bfqq_item
, &bfqd
->burst_list
,
1167 bfq_mark_bfqq_in_large_burst(bfqq_item
);
1168 bfq_mark_bfqq_in_large_burst(bfqq
);
1171 * From now on, and until the current burst finishes, any
1172 * new queue being activated shortly after the last queue
1173 * was inserted in the burst can be immediately marked as
1174 * belonging to a large burst. So the burst list is not
1175 * needed any more. Remove it.
1177 hlist_for_each_entry_safe(pos
, n
, &bfqd
->burst_list
,
1179 hlist_del_init(&pos
->burst_list_node
);
1181 * Burst not yet large: add bfqq to the burst list. Do
1182 * not increment the ref counter for bfqq, because bfqq
1183 * is removed from the burst list before freeing bfqq
1186 hlist_add_head(&bfqq
->burst_list_node
, &bfqd
->burst_list
);
1190 * If many queues belonging to the same group happen to be created
1191 * shortly after each other, then the processes associated with these
1192 * queues have typically a common goal. In particular, bursts of queue
1193 * creations are usually caused by services or applications that spawn
1194 * many parallel threads/processes. Examples are systemd during boot,
1195 * or git grep. To help these processes get their job done as soon as
1196 * possible, it is usually better to not grant either weight-raising
1197 * or device idling to their queues, unless these queues must be
1198 * protected from the I/O flowing through other active queues.
1200 * In this comment we describe, firstly, the reasons why this fact
1201 * holds, and, secondly, the next function, which implements the main
1202 * steps needed to properly mark these queues so that they can then be
1203 * treated in a different way.
1205 * The above services or applications benefit mostly from a high
1206 * throughput: the quicker the requests of the activated queues are
1207 * cumulatively served, the sooner the target job of these queues gets
1208 * completed. As a consequence, weight-raising any of these queues,
1209 * which also implies idling the device for it, is almost always
1210 * counterproductive, unless there are other active queues to isolate
1211 * these new queues from. If there no other active queues, then
1212 * weight-raising these new queues just lowers throughput in most
1215 * On the other hand, a burst of queue creations may be caused also by
1216 * the start of an application that does not consist of a lot of
1217 * parallel I/O-bound threads. In fact, with a complex application,
1218 * several short processes may need to be executed to start-up the
1219 * application. In this respect, to start an application as quickly as
1220 * possible, the best thing to do is in any case to privilege the I/O
1221 * related to the application with respect to all other
1222 * I/O. Therefore, the best strategy to start as quickly as possible
1223 * an application that causes a burst of queue creations is to
1224 * weight-raise all the queues created during the burst. This is the
1225 * exact opposite of the best strategy for the other type of bursts.
1227 * In the end, to take the best action for each of the two cases, the
1228 * two types of bursts need to be distinguished. Fortunately, this
1229 * seems relatively easy, by looking at the sizes of the bursts. In
1230 * particular, we found a threshold such that only bursts with a
1231 * larger size than that threshold are apparently caused by
1232 * services or commands such as systemd or git grep. For brevity,
1233 * hereafter we call just 'large' these bursts. BFQ *does not*
1234 * weight-raise queues whose creation occurs in a large burst. In
1235 * addition, for each of these queues BFQ performs or does not perform
1236 * idling depending on which choice boosts the throughput more. The
1237 * exact choice depends on the device and request pattern at
1240 * Unfortunately, false positives may occur while an interactive task
1241 * is starting (e.g., an application is being started). The
1242 * consequence is that the queues associated with the task do not
1243 * enjoy weight raising as expected. Fortunately these false positives
1244 * are very rare. They typically occur if some service happens to
1245 * start doing I/O exactly when the interactive task starts.
1247 * Turning back to the next function, it is invoked only if there are
1248 * no active queues (apart from active queues that would belong to the
1249 * same, possible burst bfqq would belong to), and it implements all
1250 * the steps needed to detect the occurrence of a large burst and to
1251 * properly mark all the queues belonging to it (so that they can then
1252 * be treated in a different way). This goal is achieved by
1253 * maintaining a "burst list" that holds, temporarily, the queues that
1254 * belong to the burst in progress. The list is then used to mark
1255 * these queues as belonging to a large burst if the burst does become
1256 * large. The main steps are the following.
1258 * . when the very first queue is created, the queue is inserted into the
1259 * list (as it could be the first queue in a possible burst)
1261 * . if the current burst has not yet become large, and a queue Q that does
1262 * not yet belong to the burst is activated shortly after the last time
1263 * at which a new queue entered the burst list, then the function appends
1264 * Q to the burst list
1266 * . if, as a consequence of the previous step, the burst size reaches
1267 * the large-burst threshold, then
1269 * . all the queues in the burst list are marked as belonging to a
1272 * . the burst list is deleted; in fact, the burst list already served
1273 * its purpose (keeping temporarily track of the queues in a burst,
1274 * so as to be able to mark them as belonging to a large burst in the
1275 * previous sub-step), and now is not needed any more
1277 * . the device enters a large-burst mode
1279 * . if a queue Q that does not belong to the burst is created while
1280 * the device is in large-burst mode and shortly after the last time
1281 * at which a queue either entered the burst list or was marked as
1282 * belonging to the current large burst, then Q is immediately marked
1283 * as belonging to a large burst.
1285 * . if a queue Q that does not belong to the burst is created a while
1286 * later, i.e., not shortly after, than the last time at which a queue
1287 * either entered the burst list or was marked as belonging to the
1288 * current large burst, then the current burst is deemed as finished and:
1290 * . the large-burst mode is reset if set
1292 * . the burst list is emptied
1294 * . Q is inserted in the burst list, as Q may be the first queue
1295 * in a possible new burst (then the burst list contains just Q
1298 static void bfq_handle_burst(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
1301 * If bfqq is already in the burst list or is part of a large
1302 * burst, or finally has just been split, then there is
1303 * nothing else to do.
1305 if (!hlist_unhashed(&bfqq
->burst_list_node
) ||
1306 bfq_bfqq_in_large_burst(bfqq
) ||
1307 time_is_after_eq_jiffies(bfqq
->split_time
+
1308 msecs_to_jiffies(10)))
1312 * If bfqq's creation happens late enough, or bfqq belongs to
1313 * a different group than the burst group, then the current
1314 * burst is finished, and related data structures must be
1317 * In this respect, consider the special case where bfqq is
1318 * the very first queue created after BFQ is selected for this
1319 * device. In this case, last_ins_in_burst and
1320 * burst_parent_entity are not yet significant when we get
1321 * here. But it is easy to verify that, whether or not the
1322 * following condition is true, bfqq will end up being
1323 * inserted into the burst list. In particular the list will
1324 * happen to contain only bfqq. And this is exactly what has
1325 * to happen, as bfqq may be the first queue of the first
1328 if (time_is_before_jiffies(bfqd
->last_ins_in_burst
+
1329 bfqd
->bfq_burst_interval
) ||
1330 bfqq
->entity
.parent
!= bfqd
->burst_parent_entity
) {
1331 bfqd
->large_burst
= false;
1332 bfq_reset_burst_list(bfqd
, bfqq
);
1337 * If we get here, then bfqq is being activated shortly after the
1338 * last queue. So, if the current burst is also large, we can mark
1339 * bfqq as belonging to this large burst immediately.
1341 if (bfqd
->large_burst
) {
1342 bfq_mark_bfqq_in_large_burst(bfqq
);
1347 * If we get here, then a large-burst state has not yet been
1348 * reached, but bfqq is being activated shortly after the last
1349 * queue. Then we add bfqq to the burst.
1351 bfq_add_to_burst(bfqd
, bfqq
);
1354 * At this point, bfqq either has been added to the current
1355 * burst or has caused the current burst to terminate and a
1356 * possible new burst to start. In particular, in the second
1357 * case, bfqq has become the first queue in the possible new
1358 * burst. In both cases last_ins_in_burst needs to be moved
1361 bfqd
->last_ins_in_burst
= jiffies
;
1364 static int bfq_bfqq_budget_left(struct bfq_queue
*bfqq
)
1366 struct bfq_entity
*entity
= &bfqq
->entity
;
1368 return entity
->budget
- entity
->service
;
1372 * If enough samples have been computed, return the current max budget
1373 * stored in bfqd, which is dynamically updated according to the
1374 * estimated disk peak rate; otherwise return the default max budget
1376 static int bfq_max_budget(struct bfq_data
*bfqd
)
1378 if (bfqd
->budgets_assigned
< bfq_stats_min_budgets
)
1379 return bfq_default_max_budget
;
1381 return bfqd
->bfq_max_budget
;
1385 * Return min budget, which is a fraction of the current or default
1386 * max budget (trying with 1/32)
1388 static int bfq_min_budget(struct bfq_data
*bfqd
)
1390 if (bfqd
->budgets_assigned
< bfq_stats_min_budgets
)
1391 return bfq_default_max_budget
/ 32;
1393 return bfqd
->bfq_max_budget
/ 32;
1397 * The next function, invoked after the input queue bfqq switches from
1398 * idle to busy, updates the budget of bfqq. The function also tells
1399 * whether the in-service queue should be expired, by returning
1400 * true. The purpose of expiring the in-service queue is to give bfqq
1401 * the chance to possibly preempt the in-service queue, and the reason
1402 * for preempting the in-service queue is to achieve one of the two
1405 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1406 * expired because it has remained idle. In particular, bfqq may have
1407 * expired for one of the following two reasons:
1409 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1410 * and did not make it to issue a new request before its last
1411 * request was served;
1413 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1414 * a new request before the expiration of the idling-time.
1416 * Even if bfqq has expired for one of the above reasons, the process
1417 * associated with the queue may be however issuing requests greedily,
1418 * and thus be sensitive to the bandwidth it receives (bfqq may have
1419 * remained idle for other reasons: CPU high load, bfqq not enjoying
1420 * idling, I/O throttling somewhere in the path from the process to
1421 * the I/O scheduler, ...). But if, after every expiration for one of
1422 * the above two reasons, bfqq has to wait for the service of at least
1423 * one full budget of another queue before being served again, then
1424 * bfqq is likely to get a much lower bandwidth or resource time than
1425 * its reserved ones. To address this issue, two countermeasures need
1428 * First, the budget and the timestamps of bfqq need to be updated in
1429 * a special way on bfqq reactivation: they need to be updated as if
1430 * bfqq did not remain idle and did not expire. In fact, if they are
1431 * computed as if bfqq expired and remained idle until reactivation,
1432 * then the process associated with bfqq is treated as if, instead of
1433 * being greedy, it stopped issuing requests when bfqq remained idle,
1434 * and restarts issuing requests only on this reactivation. In other
1435 * words, the scheduler does not help the process recover the "service
1436 * hole" between bfqq expiration and reactivation. As a consequence,
1437 * the process receives a lower bandwidth than its reserved one. In
1438 * contrast, to recover this hole, the budget must be updated as if
1439 * bfqq was not expired at all before this reactivation, i.e., it must
1440 * be set to the value of the remaining budget when bfqq was
1441 * expired. Along the same line, timestamps need to be assigned the
1442 * value they had the last time bfqq was selected for service, i.e.,
1443 * before last expiration. Thus timestamps need to be back-shifted
1444 * with respect to their normal computation (see [1] for more details
1445 * on this tricky aspect).
1447 * Secondly, to allow the process to recover the hole, the in-service
1448 * queue must be expired too, to give bfqq the chance to preempt it
1449 * immediately. In fact, if bfqq has to wait for a full budget of the
1450 * in-service queue to be completed, then it may become impossible to
1451 * let the process recover the hole, even if the back-shifted
1452 * timestamps of bfqq are lower than those of the in-service queue. If
1453 * this happens for most or all of the holes, then the process may not
1454 * receive its reserved bandwidth. In this respect, it is worth noting
1455 * that, being the service of outstanding requests unpreemptible, a
1456 * little fraction of the holes may however be unrecoverable, thereby
1457 * causing a little loss of bandwidth.
1459 * The last important point is detecting whether bfqq does need this
1460 * bandwidth recovery. In this respect, the next function deems the
1461 * process associated with bfqq greedy, and thus allows it to recover
1462 * the hole, if: 1) the process is waiting for the arrival of a new
1463 * request (which implies that bfqq expired for one of the above two
1464 * reasons), and 2) such a request has arrived soon. The first
1465 * condition is controlled through the flag non_blocking_wait_rq,
1466 * while the second through the flag arrived_in_time. If both
1467 * conditions hold, then the function computes the budget in the
1468 * above-described special way, and signals that the in-service queue
1469 * should be expired. Timestamp back-shifting is done later in
1470 * __bfq_activate_entity.
1472 * 2. Reduce latency. Even if timestamps are not backshifted to let
1473 * the process associated with bfqq recover a service hole, bfqq may
1474 * however happen to have, after being (re)activated, a lower finish
1475 * timestamp than the in-service queue. That is, the next budget of
1476 * bfqq may have to be completed before the one of the in-service
1477 * queue. If this is the case, then preempting the in-service queue
1478 * allows this goal to be achieved, apart from the unpreemptible,
1479 * outstanding requests mentioned above.
1481 * Unfortunately, regardless of which of the above two goals one wants
1482 * to achieve, service trees need first to be updated to know whether
1483 * the in-service queue must be preempted. To have service trees
1484 * correctly updated, the in-service queue must be expired and
1485 * rescheduled, and bfqq must be scheduled too. This is one of the
1486 * most costly operations (in future versions, the scheduling
1487 * mechanism may be re-designed in such a way to make it possible to
1488 * know whether preemption is needed without needing to update service
1489 * trees). In addition, queue preemptions almost always cause random
1490 * I/O, which may in turn cause loss of throughput. Finally, there may
1491 * even be no in-service queue when the next function is invoked (so,
1492 * no queue to compare timestamps with). Because of these facts, the
1493 * next function adopts the following simple scheme to avoid costly
1494 * operations, too frequent preemptions and too many dependencies on
1495 * the state of the scheduler: it requests the expiration of the
1496 * in-service queue (unconditionally) only for queues that need to
1497 * recover a hole. Then it delegates to other parts of the code the
1498 * responsibility of handling the above case 2.
1500 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data
*bfqd
,
1501 struct bfq_queue
*bfqq
,
1502 bool arrived_in_time
)
1504 struct bfq_entity
*entity
= &bfqq
->entity
;
1507 * In the next compound condition, we check also whether there
1508 * is some budget left, because otherwise there is no point in
1509 * trying to go on serving bfqq with this same budget: bfqq
1510 * would be expired immediately after being selected for
1511 * service. This would only cause useless overhead.
1513 if (bfq_bfqq_non_blocking_wait_rq(bfqq
) && arrived_in_time
&&
1514 bfq_bfqq_budget_left(bfqq
) > 0) {
1516 * We do not clear the flag non_blocking_wait_rq here, as
1517 * the latter is used in bfq_activate_bfqq to signal
1518 * that timestamps need to be back-shifted (and is
1519 * cleared right after).
1523 * In next assignment we rely on that either
1524 * entity->service or entity->budget are not updated
1525 * on expiration if bfqq is empty (see
1526 * __bfq_bfqq_recalc_budget). Thus both quantities
1527 * remain unchanged after such an expiration, and the
1528 * following statement therefore assigns to
1529 * entity->budget the remaining budget on such an
1532 entity
->budget
= min_t(unsigned long,
1533 bfq_bfqq_budget_left(bfqq
),
1537 * At this point, we have used entity->service to get
1538 * the budget left (needed for updating
1539 * entity->budget). Thus we finally can, and have to,
1540 * reset entity->service. The latter must be reset
1541 * because bfqq would otherwise be charged again for
1542 * the service it has received during its previous
1545 entity
->service
= 0;
1551 * We can finally complete expiration, by setting service to 0.
1553 entity
->service
= 0;
1554 entity
->budget
= max_t(unsigned long, bfqq
->max_budget
,
1555 bfq_serv_to_charge(bfqq
->next_rq
, bfqq
));
1556 bfq_clear_bfqq_non_blocking_wait_rq(bfqq
);
1561 * Return the farthest past time instant according to jiffies
1564 static unsigned long bfq_smallest_from_now(void)
1566 return jiffies
- MAX_JIFFY_OFFSET
;
1569 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data
*bfqd
,
1570 struct bfq_queue
*bfqq
,
1571 unsigned int old_wr_coeff
,
1572 bool wr_or_deserves_wr
,
1577 if (old_wr_coeff
== 1 && wr_or_deserves_wr
) {
1578 /* start a weight-raising period */
1580 bfqq
->service_from_wr
= 0;
1581 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1582 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1585 * No interactive weight raising in progress
1586 * here: assign minus infinity to
1587 * wr_start_at_switch_to_srt, to make sure
1588 * that, at the end of the soft-real-time
1589 * weight raising periods that is starting
1590 * now, no interactive weight-raising period
1591 * may be wrongly considered as still in
1592 * progress (and thus actually started by
1595 bfqq
->wr_start_at_switch_to_srt
=
1596 bfq_smallest_from_now();
1597 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
*
1598 BFQ_SOFTRT_WEIGHT_FACTOR
;
1599 bfqq
->wr_cur_max_time
=
1600 bfqd
->bfq_wr_rt_max_time
;
1604 * If needed, further reduce budget to make sure it is
1605 * close to bfqq's backlog, so as to reduce the
1606 * scheduling-error component due to a too large
1607 * budget. Do not care about throughput consequences,
1608 * but only about latency. Finally, do not assign a
1609 * too small budget either, to avoid increasing
1610 * latency by causing too frequent expirations.
1612 bfqq
->entity
.budget
= min_t(unsigned long,
1613 bfqq
->entity
.budget
,
1614 2 * bfq_min_budget(bfqd
));
1615 } else if (old_wr_coeff
> 1) {
1616 if (interactive
) { /* update wr coeff and duration */
1617 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1618 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1619 } else if (in_burst
)
1623 * The application is now or still meeting the
1624 * requirements for being deemed soft rt. We
1625 * can then correctly and safely (re)charge
1626 * the weight-raising duration for the
1627 * application with the weight-raising
1628 * duration for soft rt applications.
1630 * In particular, doing this recharge now, i.e.,
1631 * before the weight-raising period for the
1632 * application finishes, reduces the probability
1633 * of the following negative scenario:
1634 * 1) the weight of a soft rt application is
1635 * raised at startup (as for any newly
1636 * created application),
1637 * 2) since the application is not interactive,
1638 * at a certain time weight-raising is
1639 * stopped for the application,
1640 * 3) at that time the application happens to
1641 * still have pending requests, and hence
1642 * is destined to not have a chance to be
1643 * deemed soft rt before these requests are
1644 * completed (see the comments to the
1645 * function bfq_bfqq_softrt_next_start()
1646 * for details on soft rt detection),
1647 * 4) these pending requests experience a high
1648 * latency because the application is not
1649 * weight-raised while they are pending.
1651 if (bfqq
->wr_cur_max_time
!=
1652 bfqd
->bfq_wr_rt_max_time
) {
1653 bfqq
->wr_start_at_switch_to_srt
=
1654 bfqq
->last_wr_start_finish
;
1656 bfqq
->wr_cur_max_time
=
1657 bfqd
->bfq_wr_rt_max_time
;
1658 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
*
1659 BFQ_SOFTRT_WEIGHT_FACTOR
;
1661 bfqq
->last_wr_start_finish
= jiffies
;
1666 static bool bfq_bfqq_idle_for_long_time(struct bfq_data
*bfqd
,
1667 struct bfq_queue
*bfqq
)
1669 return bfqq
->dispatched
== 0 &&
1670 time_is_before_jiffies(
1671 bfqq
->budget_timeout
+
1672 bfqd
->bfq_wr_min_idle_time
);
1677 * Return true if bfqq is in a higher priority class, or has a higher
1678 * weight than the in-service queue.
1680 static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue
*bfqq
,
1681 struct bfq_queue
*in_serv_bfqq
)
1683 int bfqq_weight
, in_serv_weight
;
1685 if (bfqq
->ioprio_class
< in_serv_bfqq
->ioprio_class
)
1688 if (in_serv_bfqq
->entity
.parent
== bfqq
->entity
.parent
) {
1689 bfqq_weight
= bfqq
->entity
.weight
;
1690 in_serv_weight
= in_serv_bfqq
->entity
.weight
;
1692 if (bfqq
->entity
.parent
)
1693 bfqq_weight
= bfqq
->entity
.parent
->weight
;
1695 bfqq_weight
= bfqq
->entity
.weight
;
1696 if (in_serv_bfqq
->entity
.parent
)
1697 in_serv_weight
= in_serv_bfqq
->entity
.parent
->weight
;
1699 in_serv_weight
= in_serv_bfqq
->entity
.weight
;
1702 return bfqq_weight
> in_serv_weight
;
1705 static bool bfq_better_to_idle(struct bfq_queue
*bfqq
);
1707 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data
*bfqd
,
1708 struct bfq_queue
*bfqq
,
1713 bool soft_rt
, in_burst
, wr_or_deserves_wr
,
1714 bfqq_wants_to_preempt
,
1715 idle_for_long_time
= bfq_bfqq_idle_for_long_time(bfqd
, bfqq
),
1717 * See the comments on
1718 * bfq_bfqq_update_budg_for_activation for
1719 * details on the usage of the next variable.
1721 arrived_in_time
= ktime_get_ns() <=
1722 bfqq
->ttime
.last_end_request
+
1723 bfqd
->bfq_slice_idle
* 3;
1727 * bfqq deserves to be weight-raised if:
1729 * - it does not belong to a large burst,
1730 * - it has been idle for enough time or is soft real-time,
1731 * - is linked to a bfq_io_cq (it is not shared in any sense),
1732 * - has a default weight (otherwise we assume the user wanted
1733 * to control its weight explicitly)
1735 in_burst
= bfq_bfqq_in_large_burst(bfqq
);
1736 soft_rt
= bfqd
->bfq_wr_max_softrt_rate
> 0 &&
1737 !BFQQ_TOTALLY_SEEKY(bfqq
) &&
1739 time_is_before_jiffies(bfqq
->soft_rt_next_start
) &&
1740 bfqq
->dispatched
== 0 &&
1741 bfqq
->entity
.new_weight
== 40;
1742 *interactive
= !in_burst
&& idle_for_long_time
&&
1743 bfqq
->entity
.new_weight
== 40;
1745 * Merged bfq_queues are kept out of weight-raising
1746 * (low-latency) mechanisms. The reason is that these queues
1747 * are usually created for non-interactive and
1748 * non-soft-real-time tasks. Yet this is not the case for
1749 * stably-merged queues. These queues are merged just because
1750 * they are created shortly after each other. So they may
1751 * easily serve the I/O of an interactive or soft-real time
1752 * application, if the application happens to spawn multiple
1753 * processes. So let also stably-merged queued enjoy weight
1756 wr_or_deserves_wr
= bfqd
->low_latency
&&
1757 (bfqq
->wr_coeff
> 1 ||
1758 (bfq_bfqq_sync(bfqq
) &&
1759 (bfqq
->bic
|| RQ_BIC(rq
)->stably_merged
) &&
1760 (*interactive
|| soft_rt
)));
1763 * Using the last flag, update budget and check whether bfqq
1764 * may want to preempt the in-service queue.
1766 bfqq_wants_to_preempt
=
1767 bfq_bfqq_update_budg_for_activation(bfqd
, bfqq
,
1771 * If bfqq happened to be activated in a burst, but has been
1772 * idle for much more than an interactive queue, then we
1773 * assume that, in the overall I/O initiated in the burst, the
1774 * I/O associated with bfqq is finished. So bfqq does not need
1775 * to be treated as a queue belonging to a burst
1776 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1777 * if set, and remove bfqq from the burst list if it's
1778 * there. We do not decrement burst_size, because the fact
1779 * that bfqq does not need to belong to the burst list any
1780 * more does not invalidate the fact that bfqq was created in
1783 if (likely(!bfq_bfqq_just_created(bfqq
)) &&
1784 idle_for_long_time
&&
1785 time_is_before_jiffies(
1786 bfqq
->budget_timeout
+
1787 msecs_to_jiffies(10000))) {
1788 hlist_del_init(&bfqq
->burst_list_node
);
1789 bfq_clear_bfqq_in_large_burst(bfqq
);
1792 bfq_clear_bfqq_just_created(bfqq
);
1794 if (bfqd
->low_latency
) {
1795 if (unlikely(time_is_after_jiffies(bfqq
->split_time
)))
1798 jiffies
- bfqd
->bfq_wr_min_idle_time
- 1;
1800 if (time_is_before_jiffies(bfqq
->split_time
+
1801 bfqd
->bfq_wr_min_idle_time
)) {
1802 bfq_update_bfqq_wr_on_rq_arrival(bfqd
, bfqq
,
1809 if (old_wr_coeff
!= bfqq
->wr_coeff
)
1810 bfqq
->entity
.prio_changed
= 1;
1814 bfqq
->last_idle_bklogged
= jiffies
;
1815 bfqq
->service_from_backlogged
= 0;
1816 bfq_clear_bfqq_softrt_update(bfqq
);
1818 bfq_add_bfqq_busy(bfqd
, bfqq
);
1821 * Expire in-service queue if preemption may be needed for
1822 * guarantees or throughput. As for guarantees, we care
1823 * explicitly about two cases. The first is that bfqq has to
1824 * recover a service hole, as explained in the comments on
1825 * bfq_bfqq_update_budg_for_activation(), i.e., that
1826 * bfqq_wants_to_preempt is true. However, if bfqq does not
1827 * carry time-critical I/O, then bfqq's bandwidth is less
1828 * important than that of queues that carry time-critical I/O.
1829 * So, as a further constraint, we consider this case only if
1830 * bfqq is at least as weight-raised, i.e., at least as time
1831 * critical, as the in-service queue.
1833 * The second case is that bfqq is in a higher priority class,
1834 * or has a higher weight than the in-service queue. If this
1835 * condition does not hold, we don't care because, even if
1836 * bfqq does not start to be served immediately, the resulting
1837 * delay for bfqq's I/O is however lower or much lower than
1838 * the ideal completion time to be guaranteed to bfqq's I/O.
1840 * In both cases, preemption is needed only if, according to
1841 * the timestamps of both bfqq and of the in-service queue,
1842 * bfqq actually is the next queue to serve. So, to reduce
1843 * useless preemptions, the return value of
1844 * next_queue_may_preempt() is considered in the next compound
1845 * condition too. Yet next_queue_may_preempt() just checks a
1846 * simple, necessary condition for bfqq to be the next queue
1847 * to serve. In fact, to evaluate a sufficient condition, the
1848 * timestamps of the in-service queue would need to be
1849 * updated, and this operation is quite costly (see the
1850 * comments on bfq_bfqq_update_budg_for_activation()).
1852 * As for throughput, we ask bfq_better_to_idle() whether we
1853 * still need to plug I/O dispatching. If bfq_better_to_idle()
1854 * says no, then plugging is not needed any longer, either to
1855 * boost throughput or to perserve service guarantees. Then
1856 * the best option is to stop plugging I/O, as not doing so
1857 * would certainly lower throughput. We may end up in this
1858 * case if: (1) upon a dispatch attempt, we detected that it
1859 * was better to plug I/O dispatch, and to wait for a new
1860 * request to arrive for the currently in-service queue, but
1861 * (2) this switch of bfqq to busy changes the scenario.
1863 if (bfqd
->in_service_queue
&&
1864 ((bfqq_wants_to_preempt
&&
1865 bfqq
->wr_coeff
>= bfqd
->in_service_queue
->wr_coeff
) ||
1866 bfq_bfqq_higher_class_or_weight(bfqq
, bfqd
->in_service_queue
) ||
1867 !bfq_better_to_idle(bfqd
->in_service_queue
)) &&
1868 next_queue_may_preempt(bfqd
))
1869 bfq_bfqq_expire(bfqd
, bfqd
->in_service_queue
,
1870 false, BFQQE_PREEMPTED
);
1873 static void bfq_reset_inject_limit(struct bfq_data
*bfqd
,
1874 struct bfq_queue
*bfqq
)
1876 /* invalidate baseline total service time */
1877 bfqq
->last_serv_time_ns
= 0;
1880 * Reset pointer in case we are waiting for
1881 * some request completion.
1883 bfqd
->waited_rq
= NULL
;
1886 * If bfqq has a short think time, then start by setting the
1887 * inject limit to 0 prudentially, because the service time of
1888 * an injected I/O request may be higher than the think time
1889 * of bfqq, and therefore, if one request was injected when
1890 * bfqq remains empty, this injected request might delay the
1891 * service of the next I/O request for bfqq significantly. In
1892 * case bfqq can actually tolerate some injection, then the
1893 * adaptive update will however raise the limit soon. This
1894 * lucky circumstance holds exactly because bfqq has a short
1895 * think time, and thus, after remaining empty, is likely to
1896 * get new I/O enqueued---and then completed---before being
1897 * expired. This is the very pattern that gives the
1898 * limit-update algorithm the chance to measure the effect of
1899 * injection on request service times, and then to update the
1900 * limit accordingly.
1902 * However, in the following special case, the inject limit is
1903 * left to 1 even if the think time is short: bfqq's I/O is
1904 * synchronized with that of some other queue, i.e., bfqq may
1905 * receive new I/O only after the I/O of the other queue is
1906 * completed. Keeping the inject limit to 1 allows the
1907 * blocking I/O to be served while bfqq is in service. And
1908 * this is very convenient both for bfqq and for overall
1909 * throughput, as explained in detail in the comments in
1910 * bfq_update_has_short_ttime().
1912 * On the opposite end, if bfqq has a long think time, then
1913 * start directly by 1, because:
1914 * a) on the bright side, keeping at most one request in
1915 * service in the drive is unlikely to cause any harm to the
1916 * latency of bfqq's requests, as the service time of a single
1917 * request is likely to be lower than the think time of bfqq;
1918 * b) on the downside, after becoming empty, bfqq is likely to
1919 * expire before getting its next request. With this request
1920 * arrival pattern, it is very hard to sample total service
1921 * times and update the inject limit accordingly (see comments
1922 * on bfq_update_inject_limit()). So the limit is likely to be
1923 * never, or at least seldom, updated. As a consequence, by
1924 * setting the limit to 1, we avoid that no injection ever
1925 * occurs with bfqq. On the downside, this proactive step
1926 * further reduces chances to actually compute the baseline
1927 * total service time. Thus it reduces chances to execute the
1928 * limit-update algorithm and possibly raise the limit to more
1931 if (bfq_bfqq_has_short_ttime(bfqq
))
1932 bfqq
->inject_limit
= 0;
1934 bfqq
->inject_limit
= 1;
1936 bfqq
->decrease_time_jif
= jiffies
;
1939 static void bfq_update_io_intensity(struct bfq_queue
*bfqq
, u64 now_ns
)
1941 u64 tot_io_time
= now_ns
- bfqq
->io_start_time
;
1943 if (RB_EMPTY_ROOT(&bfqq
->sort_list
) && bfqq
->dispatched
== 0)
1944 bfqq
->tot_idle_time
+=
1945 now_ns
- bfqq
->ttime
.last_end_request
;
1947 if (unlikely(bfq_bfqq_just_created(bfqq
)))
1951 * Must be busy for at least about 80% of the time to be
1952 * considered I/O bound.
1954 if (bfqq
->tot_idle_time
* 5 > tot_io_time
)
1955 bfq_clear_bfqq_IO_bound(bfqq
);
1957 bfq_mark_bfqq_IO_bound(bfqq
);
1960 * Keep an observation window of at most 200 ms in the past
1963 if (tot_io_time
> 200 * NSEC_PER_MSEC
) {
1964 bfqq
->io_start_time
= now_ns
- (tot_io_time
>>1);
1965 bfqq
->tot_idle_time
>>= 1;
1970 * Detect whether bfqq's I/O seems synchronized with that of some
1971 * other queue, i.e., whether bfqq, after remaining empty, happens to
1972 * receive new I/O only right after some I/O request of the other
1973 * queue has been completed. We call waker queue the other queue, and
1974 * we assume, for simplicity, that bfqq may have at most one waker
1977 * A remarkable throughput boost can be reached by unconditionally
1978 * injecting the I/O of the waker queue, every time a new
1979 * bfq_dispatch_request happens to be invoked while I/O is being
1980 * plugged for bfqq. In addition to boosting throughput, this
1981 * unblocks bfqq's I/O, thereby improving bandwidth and latency for
1982 * bfqq. Note that these same results may be achieved with the general
1983 * injection mechanism, but less effectively. For details on this
1984 * aspect, see the comments on the choice of the queue for injection
1985 * in bfq_select_queue().
1987 * Turning back to the detection of a waker queue, a queue Q is deemed
1988 * as a waker queue for bfqq if, for three consecutive times, bfqq
1989 * happens to become non empty right after a request of Q has been
1990 * completed. In this respect, even if bfqq is empty, we do not check
1991 * for a waker if it still has some in-flight I/O. In fact, in this
1992 * case bfqq is actually still being served by the drive, and may
1993 * receive new I/O on the completion of some of the in-flight
1994 * requests. In particular, on the first time, Q is tentatively set as
1995 * a candidate waker queue, while on the third consecutive time that Q
1996 * is detected, the field waker_bfqq is set to Q, to confirm that Q is
1997 * a waker queue for bfqq. These detection steps are performed only if
1998 * bfqq has a long think time, so as to make it more likely that
1999 * bfqq's I/O is actually being blocked by a synchronization. This
2000 * last filter, plus the above three-times requirement, make false
2001 * positives less likely.
2005 * The sooner a waker queue is detected, the sooner throughput can be
2006 * boosted by injecting I/O from the waker queue. Fortunately,
2007 * detection is likely to be actually fast, for the following
2008 * reasons. While blocked by synchronization, bfqq has a long think
2009 * time. This implies that bfqq's inject limit is at least equal to 1
2010 * (see the comments in bfq_update_inject_limit()). So, thanks to
2011 * injection, the waker queue is likely to be served during the very
2012 * first I/O-plugging time interval for bfqq. This triggers the first
2013 * step of the detection mechanism. Thanks again to injection, the
2014 * candidate waker queue is then likely to be confirmed no later than
2015 * during the next I/O-plugging interval for bfqq.
2019 * On queue merging all waker information is lost.
2021 static void bfq_check_waker(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
2024 if (!bfqd
->last_completed_rq_bfqq
||
2025 bfqd
->last_completed_rq_bfqq
== bfqq
||
2026 bfq_bfqq_has_short_ttime(bfqq
) ||
2027 now_ns
- bfqd
->last_completion
>= 4 * NSEC_PER_MSEC
)
2030 if (bfqd
->last_completed_rq_bfqq
!=
2031 bfqq
->tentative_waker_bfqq
) {
2033 * First synchronization detected with a
2034 * candidate waker queue, or with a different
2035 * candidate waker queue from the current one.
2037 bfqq
->tentative_waker_bfqq
=
2038 bfqd
->last_completed_rq_bfqq
;
2039 bfqq
->num_waker_detections
= 1;
2040 } else /* Same tentative waker queue detected again */
2041 bfqq
->num_waker_detections
++;
2043 if (bfqq
->num_waker_detections
== 3) {
2044 bfqq
->waker_bfqq
= bfqd
->last_completed_rq_bfqq
;
2045 bfqq
->tentative_waker_bfqq
= NULL
;
2048 * If the waker queue disappears, then
2049 * bfqq->waker_bfqq must be reset. To
2050 * this goal, we maintain in each
2051 * waker queue a list, woken_list, of
2052 * all the queues that reference the
2053 * waker queue through their
2054 * waker_bfqq pointer. When the waker
2055 * queue exits, the waker_bfqq pointer
2056 * of all the queues in the woken_list
2059 * In addition, if bfqq is already in
2060 * the woken_list of a waker queue,
2061 * then, before being inserted into
2062 * the woken_list of a new waker
2063 * queue, bfqq must be removed from
2064 * the woken_list of the old waker
2067 if (!hlist_unhashed(&bfqq
->woken_list_node
))
2068 hlist_del_init(&bfqq
->woken_list_node
);
2069 hlist_add_head(&bfqq
->woken_list_node
,
2070 &bfqd
->last_completed_rq_bfqq
->woken_list
);
2074 static void bfq_add_request(struct request
*rq
)
2076 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
2077 struct bfq_data
*bfqd
= bfqq
->bfqd
;
2078 struct request
*next_rq
, *prev
;
2079 unsigned int old_wr_coeff
= bfqq
->wr_coeff
;
2080 bool interactive
= false;
2081 u64 now_ns
= ktime_get_ns();
2083 bfq_log_bfqq(bfqd
, bfqq
, "add_request %d", rq_is_sync(rq
));
2084 bfqq
->queued
[rq_is_sync(rq
)]++;
2087 if (bfq_bfqq_sync(bfqq
) && RQ_BIC(rq
)->requests
<= 1) {
2088 bfq_check_waker(bfqd
, bfqq
, now_ns
);
2091 * Periodically reset inject limit, to make sure that
2092 * the latter eventually drops in case workload
2093 * changes, see step (3) in the comments on
2094 * bfq_update_inject_limit().
2096 if (time_is_before_eq_jiffies(bfqq
->decrease_time_jif
+
2097 msecs_to_jiffies(1000)))
2098 bfq_reset_inject_limit(bfqd
, bfqq
);
2101 * The following conditions must hold to setup a new
2102 * sampling of total service time, and then a new
2103 * update of the inject limit:
2104 * - bfqq is in service, because the total service
2105 * time is evaluated only for the I/O requests of
2106 * the queues in service;
2107 * - this is the right occasion to compute or to
2108 * lower the baseline total service time, because
2109 * there are actually no requests in the drive,
2111 * the baseline total service time is available, and
2112 * this is the right occasion to compute the other
2113 * quantity needed to update the inject limit, i.e.,
2114 * the total service time caused by the amount of
2115 * injection allowed by the current value of the
2116 * limit. It is the right occasion because injection
2117 * has actually been performed during the service
2118 * hole, and there are still in-flight requests,
2119 * which are very likely to be exactly the injected
2120 * requests, or part of them;
2121 * - the minimum interval for sampling the total
2122 * service time and updating the inject limit has
2125 if (bfqq
== bfqd
->in_service_queue
&&
2126 (bfqd
->rq_in_driver
== 0 ||
2127 (bfqq
->last_serv_time_ns
> 0 &&
2128 bfqd
->rqs_injected
&& bfqd
->rq_in_driver
> 0)) &&
2129 time_is_before_eq_jiffies(bfqq
->decrease_time_jif
+
2130 msecs_to_jiffies(10))) {
2131 bfqd
->last_empty_occupied_ns
= ktime_get_ns();
2133 * Start the state machine for measuring the
2134 * total service time of rq: setting
2135 * wait_dispatch will cause bfqd->waited_rq to
2136 * be set when rq will be dispatched.
2138 bfqd
->wait_dispatch
= true;
2140 * If there is no I/O in service in the drive,
2141 * then possible injection occurred before the
2142 * arrival of rq will not affect the total
2143 * service time of rq. So the injection limit
2144 * must not be updated as a function of such
2145 * total service time, unless new injection
2146 * occurs before rq is completed. To have the
2147 * injection limit updated only in the latter
2148 * case, reset rqs_injected here (rqs_injected
2149 * will be set in case injection is performed
2150 * on bfqq before rq is completed).
2152 if (bfqd
->rq_in_driver
== 0)
2153 bfqd
->rqs_injected
= false;
2157 if (bfq_bfqq_sync(bfqq
))
2158 bfq_update_io_intensity(bfqq
, now_ns
);
2160 elv_rb_add(&bfqq
->sort_list
, rq
);
2163 * Check if this request is a better next-serve candidate.
2165 prev
= bfqq
->next_rq
;
2166 next_rq
= bfq_choose_req(bfqd
, bfqq
->next_rq
, rq
, bfqd
->last_position
);
2167 bfqq
->next_rq
= next_rq
;
2170 * Adjust priority tree position, if next_rq changes.
2171 * See comments on bfq_pos_tree_add_move() for the unlikely().
2173 if (unlikely(!bfqd
->nonrot_with_queueing
&& prev
!= bfqq
->next_rq
))
2174 bfq_pos_tree_add_move(bfqd
, bfqq
);
2176 if (!bfq_bfqq_busy(bfqq
)) /* switching to busy ... */
2177 bfq_bfqq_handle_idle_busy_switch(bfqd
, bfqq
, old_wr_coeff
,
2180 if (bfqd
->low_latency
&& old_wr_coeff
== 1 && !rq_is_sync(rq
) &&
2181 time_is_before_jiffies(
2182 bfqq
->last_wr_start_finish
+
2183 bfqd
->bfq_wr_min_inter_arr_async
)) {
2184 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
2185 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
2187 bfqd
->wr_busy_queues
++;
2188 bfqq
->entity
.prio_changed
= 1;
2190 if (prev
!= bfqq
->next_rq
)
2191 bfq_updated_next_req(bfqd
, bfqq
);
2195 * Assign jiffies to last_wr_start_finish in the following
2198 * . if bfqq is not going to be weight-raised, because, for
2199 * non weight-raised queues, last_wr_start_finish stores the
2200 * arrival time of the last request; as of now, this piece
2201 * of information is used only for deciding whether to
2202 * weight-raise async queues
2204 * . if bfqq is not weight-raised, because, if bfqq is now
2205 * switching to weight-raised, then last_wr_start_finish
2206 * stores the time when weight-raising starts
2208 * . if bfqq is interactive, because, regardless of whether
2209 * bfqq is currently weight-raised, the weight-raising
2210 * period must start or restart (this case is considered
2211 * separately because it is not detected by the above
2212 * conditions, if bfqq is already weight-raised)
2214 * last_wr_start_finish has to be updated also if bfqq is soft
2215 * real-time, because the weight-raising period is constantly
2216 * restarted on idle-to-busy transitions for these queues, but
2217 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2220 if (bfqd
->low_latency
&&
2221 (old_wr_coeff
== 1 || bfqq
->wr_coeff
== 1 || interactive
))
2222 bfqq
->last_wr_start_finish
= jiffies
;
2225 static struct request
*bfq_find_rq_fmerge(struct bfq_data
*bfqd
,
2227 struct request_queue
*q
)
2229 struct bfq_queue
*bfqq
= bfqd
->bio_bfqq
;
2233 return elv_rb_find(&bfqq
->sort_list
, bio_end_sector(bio
));
2238 static sector_t
get_sdist(sector_t last_pos
, struct request
*rq
)
2241 return abs(blk_rq_pos(rq
) - last_pos
);
2246 #if 0 /* Still not clear if we can do without next two functions */
2247 static void bfq_activate_request(struct request_queue
*q
, struct request
*rq
)
2249 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
2251 bfqd
->rq_in_driver
++;
2254 static void bfq_deactivate_request(struct request_queue
*q
, struct request
*rq
)
2256 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
2258 bfqd
->rq_in_driver
--;
2262 static void bfq_remove_request(struct request_queue
*q
,
2265 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
2266 struct bfq_data
*bfqd
= bfqq
->bfqd
;
2267 const int sync
= rq_is_sync(rq
);
2269 if (bfqq
->next_rq
== rq
) {
2270 bfqq
->next_rq
= bfq_find_next_rq(bfqd
, bfqq
, rq
);
2271 bfq_updated_next_req(bfqd
, bfqq
);
2274 if (rq
->queuelist
.prev
!= &rq
->queuelist
)
2275 list_del_init(&rq
->queuelist
);
2276 bfqq
->queued
[sync
]--;
2278 elv_rb_del(&bfqq
->sort_list
, rq
);
2280 elv_rqhash_del(q
, rq
);
2281 if (q
->last_merge
== rq
)
2282 q
->last_merge
= NULL
;
2284 if (RB_EMPTY_ROOT(&bfqq
->sort_list
)) {
2285 bfqq
->next_rq
= NULL
;
2287 if (bfq_bfqq_busy(bfqq
) && bfqq
!= bfqd
->in_service_queue
) {
2288 bfq_del_bfqq_busy(bfqd
, bfqq
, false);
2290 * bfqq emptied. In normal operation, when
2291 * bfqq is empty, bfqq->entity.service and
2292 * bfqq->entity.budget must contain,
2293 * respectively, the service received and the
2294 * budget used last time bfqq emptied. These
2295 * facts do not hold in this case, as at least
2296 * this last removal occurred while bfqq is
2297 * not in service. To avoid inconsistencies,
2298 * reset both bfqq->entity.service and
2299 * bfqq->entity.budget, if bfqq has still a
2300 * process that may issue I/O requests to it.
2302 bfqq
->entity
.budget
= bfqq
->entity
.service
= 0;
2306 * Remove queue from request-position tree as it is empty.
2308 if (bfqq
->pos_root
) {
2309 rb_erase(&bfqq
->pos_node
, bfqq
->pos_root
);
2310 bfqq
->pos_root
= NULL
;
2313 /* see comments on bfq_pos_tree_add_move() for the unlikely() */
2314 if (unlikely(!bfqd
->nonrot_with_queueing
))
2315 bfq_pos_tree_add_move(bfqd
, bfqq
);
2318 if (rq
->cmd_flags
& REQ_META
)
2319 bfqq
->meta_pending
--;
2323 static bool bfq_bio_merge(struct request_queue
*q
, struct bio
*bio
,
2324 unsigned int nr_segs
)
2326 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
2327 struct request
*free
= NULL
;
2329 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2330 * store its return value for later use, to avoid nesting
2331 * queue_lock inside the bfqd->lock. We assume that the bic
2332 * returned by bfq_bic_lookup does not go away before
2333 * bfqd->lock is taken.
2335 struct bfq_io_cq
*bic
= bfq_bic_lookup(bfqd
, current
->io_context
, q
);
2338 spin_lock_irq(&bfqd
->lock
);
2342 * Make sure cgroup info is uptodate for current process before
2343 * considering the merge.
2345 bfq_bic_update_cgroup(bic
, bio
);
2347 bfqd
->bio_bfqq
= bic_to_bfqq(bic
, op_is_sync(bio
->bi_opf
));
2349 bfqd
->bio_bfqq
= NULL
;
2351 bfqd
->bio_bic
= bic
;
2353 ret
= blk_mq_sched_try_merge(q
, bio
, nr_segs
, &free
);
2355 spin_unlock_irq(&bfqd
->lock
);
2357 blk_mq_free_request(free
);
2362 static int bfq_request_merge(struct request_queue
*q
, struct request
**req
,
2365 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
2366 struct request
*__rq
;
2368 __rq
= bfq_find_rq_fmerge(bfqd
, bio
, q
);
2369 if (__rq
&& elv_bio_merge_ok(__rq
, bio
)) {
2372 if (blk_discard_mergable(__rq
))
2373 return ELEVATOR_DISCARD_MERGE
;
2374 return ELEVATOR_FRONT_MERGE
;
2377 return ELEVATOR_NO_MERGE
;
2380 static void bfq_request_merged(struct request_queue
*q
, struct request
*req
,
2381 enum elv_merge type
)
2383 if (type
== ELEVATOR_FRONT_MERGE
&&
2384 rb_prev(&req
->rb_node
) &&
2386 blk_rq_pos(container_of(rb_prev(&req
->rb_node
),
2387 struct request
, rb_node
))) {
2388 struct bfq_queue
*bfqq
= RQ_BFQQ(req
);
2389 struct bfq_data
*bfqd
;
2390 struct request
*prev
, *next_rq
;
2397 /* Reposition request in its sort_list */
2398 elv_rb_del(&bfqq
->sort_list
, req
);
2399 elv_rb_add(&bfqq
->sort_list
, req
);
2401 /* Choose next request to be served for bfqq */
2402 prev
= bfqq
->next_rq
;
2403 next_rq
= bfq_choose_req(bfqd
, bfqq
->next_rq
, req
,
2404 bfqd
->last_position
);
2405 bfqq
->next_rq
= next_rq
;
2407 * If next_rq changes, update both the queue's budget to
2408 * fit the new request and the queue's position in its
2411 if (prev
!= bfqq
->next_rq
) {
2412 bfq_updated_next_req(bfqd
, bfqq
);
2414 * See comments on bfq_pos_tree_add_move() for
2417 if (unlikely(!bfqd
->nonrot_with_queueing
))
2418 bfq_pos_tree_add_move(bfqd
, bfqq
);
2424 * This function is called to notify the scheduler that the requests
2425 * rq and 'next' have been merged, with 'next' going away. BFQ
2426 * exploits this hook to address the following issue: if 'next' has a
2427 * fifo_time lower that rq, then the fifo_time of rq must be set to
2428 * the value of 'next', to not forget the greater age of 'next'.
2430 * NOTE: in this function we assume that rq is in a bfq_queue, basing
2431 * on that rq is picked from the hash table q->elevator->hash, which,
2432 * in its turn, is filled only with I/O requests present in
2433 * bfq_queues, while BFQ is in use for the request queue q. In fact,
2434 * the function that fills this hash table (elv_rqhash_add) is called
2435 * only by bfq_insert_request.
2437 static void bfq_requests_merged(struct request_queue
*q
, struct request
*rq
,
2438 struct request
*next
)
2440 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
),
2441 *next_bfqq
= RQ_BFQQ(next
);
2447 * If next and rq belong to the same bfq_queue and next is older
2448 * than rq, then reposition rq in the fifo (by substituting next
2449 * with rq). Otherwise, if next and rq belong to different
2450 * bfq_queues, never reposition rq: in fact, we would have to
2451 * reposition it with respect to next's position in its own fifo,
2452 * which would most certainly be too expensive with respect to
2455 if (bfqq
== next_bfqq
&&
2456 !list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
2457 next
->fifo_time
< rq
->fifo_time
) {
2458 list_del_init(&rq
->queuelist
);
2459 list_replace_init(&next
->queuelist
, &rq
->queuelist
);
2460 rq
->fifo_time
= next
->fifo_time
;
2463 if (bfqq
->next_rq
== next
)
2466 bfqg_stats_update_io_merged(bfqq_group(bfqq
), next
->cmd_flags
);
2468 /* Merged request may be in the IO scheduler. Remove it. */
2469 if (!RB_EMPTY_NODE(&next
->rb_node
)) {
2470 bfq_remove_request(next
->q
, next
);
2472 bfqg_stats_update_io_remove(bfqq_group(next_bfqq
),
2477 /* Must be called with bfqq != NULL */
2478 static void bfq_bfqq_end_wr(struct bfq_queue
*bfqq
)
2481 * If bfqq has been enjoying interactive weight-raising, then
2482 * reset soft_rt_next_start. We do it for the following
2483 * reason. bfqq may have been conveying the I/O needed to load
2484 * a soft real-time application. Such an application actually
2485 * exhibits a soft real-time I/O pattern after it finishes
2486 * loading, and finally starts doing its job. But, if bfqq has
2487 * been receiving a lot of bandwidth so far (likely to happen
2488 * on a fast device), then soft_rt_next_start now contains a
2489 * high value that. So, without this reset, bfqq would be
2490 * prevented from being possibly considered as soft_rt for a
2494 if (bfqq
->wr_cur_max_time
!=
2495 bfqq
->bfqd
->bfq_wr_rt_max_time
)
2496 bfqq
->soft_rt_next_start
= jiffies
;
2498 if (bfq_bfqq_busy(bfqq
))
2499 bfqq
->bfqd
->wr_busy_queues
--;
2501 bfqq
->wr_cur_max_time
= 0;
2502 bfqq
->last_wr_start_finish
= jiffies
;
2504 * Trigger a weight change on the next invocation of
2505 * __bfq_entity_update_weight_prio.
2507 bfqq
->entity
.prio_changed
= 1;
2510 void bfq_end_wr_async_queues(struct bfq_data
*bfqd
,
2511 struct bfq_group
*bfqg
)
2515 for (i
= 0; i
< 2; i
++)
2516 for (j
= 0; j
< IOPRIO_NR_LEVELS
; j
++)
2517 if (bfqg
->async_bfqq
[i
][j
])
2518 bfq_bfqq_end_wr(bfqg
->async_bfqq
[i
][j
]);
2519 if (bfqg
->async_idle_bfqq
)
2520 bfq_bfqq_end_wr(bfqg
->async_idle_bfqq
);
2523 static void bfq_end_wr(struct bfq_data
*bfqd
)
2525 struct bfq_queue
*bfqq
;
2527 spin_lock_irq(&bfqd
->lock
);
2529 list_for_each_entry(bfqq
, &bfqd
->active_list
, bfqq_list
)
2530 bfq_bfqq_end_wr(bfqq
);
2531 list_for_each_entry(bfqq
, &bfqd
->idle_list
, bfqq_list
)
2532 bfq_bfqq_end_wr(bfqq
);
2533 bfq_end_wr_async(bfqd
);
2535 spin_unlock_irq(&bfqd
->lock
);
2538 static sector_t
bfq_io_struct_pos(void *io_struct
, bool request
)
2541 return blk_rq_pos(io_struct
);
2543 return ((struct bio
*)io_struct
)->bi_iter
.bi_sector
;
2546 static int bfq_rq_close_to_sector(void *io_struct
, bool request
,
2549 return abs(bfq_io_struct_pos(io_struct
, request
) - sector
) <=
2553 static struct bfq_queue
*bfqq_find_close(struct bfq_data
*bfqd
,
2554 struct bfq_queue
*bfqq
,
2557 struct rb_root
*root
= &bfq_bfqq_to_bfqg(bfqq
)->rq_pos_tree
;
2558 struct rb_node
*parent
, *node
;
2559 struct bfq_queue
*__bfqq
;
2561 if (RB_EMPTY_ROOT(root
))
2565 * First, if we find a request starting at the end of the last
2566 * request, choose it.
2568 __bfqq
= bfq_rq_pos_tree_lookup(bfqd
, root
, sector
, &parent
, NULL
);
2573 * If the exact sector wasn't found, the parent of the NULL leaf
2574 * will contain the closest sector (rq_pos_tree sorted by
2575 * next_request position).
2577 __bfqq
= rb_entry(parent
, struct bfq_queue
, pos_node
);
2578 if (bfq_rq_close_to_sector(__bfqq
->next_rq
, true, sector
))
2581 if (blk_rq_pos(__bfqq
->next_rq
) < sector
)
2582 node
= rb_next(&__bfqq
->pos_node
);
2584 node
= rb_prev(&__bfqq
->pos_node
);
2588 __bfqq
= rb_entry(node
, struct bfq_queue
, pos_node
);
2589 if (bfq_rq_close_to_sector(__bfqq
->next_rq
, true, sector
))
2595 static struct bfq_queue
*bfq_find_close_cooperator(struct bfq_data
*bfqd
,
2596 struct bfq_queue
*cur_bfqq
,
2599 struct bfq_queue
*bfqq
;
2602 * We shall notice if some of the queues are cooperating,
2603 * e.g., working closely on the same area of the device. In
2604 * that case, we can group them together and: 1) don't waste
2605 * time idling, and 2) serve the union of their requests in
2606 * the best possible order for throughput.
2608 bfqq
= bfqq_find_close(bfqd
, cur_bfqq
, sector
);
2609 if (!bfqq
|| bfqq
== cur_bfqq
)
2615 static struct bfq_queue
*
2616 bfq_setup_merge(struct bfq_queue
*bfqq
, struct bfq_queue
*new_bfqq
)
2618 int process_refs
, new_process_refs
;
2619 struct bfq_queue
*__bfqq
;
2622 * If there are no process references on the new_bfqq, then it is
2623 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2624 * may have dropped their last reference (not just their last process
2627 if (!bfqq_process_refs(new_bfqq
))
2630 /* Avoid a circular list and skip interim queue merges. */
2631 while ((__bfqq
= new_bfqq
->new_bfqq
)) {
2637 process_refs
= bfqq_process_refs(bfqq
);
2638 new_process_refs
= bfqq_process_refs(new_bfqq
);
2640 * If the process for the bfqq has gone away, there is no
2641 * sense in merging the queues.
2643 if (process_refs
== 0 || new_process_refs
== 0)
2647 * Make sure merged queues belong to the same parent. Parents could
2648 * have changed since the time we decided the two queues are suitable
2651 if (new_bfqq
->entity
.parent
!= bfqq
->entity
.parent
)
2654 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "scheduling merge with queue %d",
2658 * Merging is just a redirection: the requests of the process
2659 * owning one of the two queues are redirected to the other queue.
2660 * The latter queue, in its turn, is set as shared if this is the
2661 * first time that the requests of some process are redirected to
2664 * We redirect bfqq to new_bfqq and not the opposite, because
2665 * we are in the context of the process owning bfqq, thus we
2666 * have the io_cq of this process. So we can immediately
2667 * configure this io_cq to redirect the requests of the
2668 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2669 * not available any more (new_bfqq->bic == NULL).
2671 * Anyway, even in case new_bfqq coincides with the in-service
2672 * queue, redirecting requests the in-service queue is the
2673 * best option, as we feed the in-service queue with new
2674 * requests close to the last request served and, by doing so,
2675 * are likely to increase the throughput.
2677 bfqq
->new_bfqq
= new_bfqq
;
2679 * The above assignment schedules the following redirections:
2680 * each time some I/O for bfqq arrives, the process that
2681 * generated that I/O is disassociated from bfqq and
2682 * associated with new_bfqq. Here we increases new_bfqq->ref
2683 * in advance, adding the number of processes that are
2684 * expected to be associated with new_bfqq as they happen to
2687 new_bfqq
->ref
+= process_refs
;
2691 static bool bfq_may_be_close_cooperator(struct bfq_queue
*bfqq
,
2692 struct bfq_queue
*new_bfqq
)
2694 if (bfq_too_late_for_merging(new_bfqq
))
2697 if (bfq_class_idle(bfqq
) || bfq_class_idle(new_bfqq
) ||
2698 (bfqq
->ioprio_class
!= new_bfqq
->ioprio_class
))
2702 * If either of the queues has already been detected as seeky,
2703 * then merging it with the other queue is unlikely to lead to
2706 if (BFQQ_SEEKY(bfqq
) || BFQQ_SEEKY(new_bfqq
))
2710 * Interleaved I/O is known to be done by (some) applications
2711 * only for reads, so it does not make sense to merge async
2714 if (!bfq_bfqq_sync(bfqq
) || !bfq_bfqq_sync(new_bfqq
))
2720 static bool idling_boosts_thr_without_issues(struct bfq_data
*bfqd
,
2721 struct bfq_queue
*bfqq
);
2724 * Attempt to schedule a merge of bfqq with the currently in-service
2725 * queue or with a close queue among the scheduled queues. Return
2726 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2727 * structure otherwise.
2729 * The OOM queue is not allowed to participate to cooperation: in fact, since
2730 * the requests temporarily redirected to the OOM queue could be redirected
2731 * again to dedicated queues at any time, the state needed to correctly
2732 * handle merging with the OOM queue would be quite complex and expensive
2733 * to maintain. Besides, in such a critical condition as an out of memory,
2734 * the benefits of queue merging may be little relevant, or even negligible.
2736 * WARNING: queue merging may impair fairness among non-weight raised
2737 * queues, for at least two reasons: 1) the original weight of a
2738 * merged queue may change during the merged state, 2) even being the
2739 * weight the same, a merged queue may be bloated with many more
2740 * requests than the ones produced by its originally-associated
2743 static struct bfq_queue
*
2744 bfq_setup_cooperator(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
2745 void *io_struct
, bool request
, struct bfq_io_cq
*bic
)
2747 struct bfq_queue
*in_service_bfqq
, *new_bfqq
;
2749 /* if a merge has already been setup, then proceed with that first */
2751 return bfqq
->new_bfqq
;
2754 * Check delayed stable merge for rotational or non-queueing
2755 * devs. For this branch to be executed, bfqq must not be
2756 * currently merged with some other queue (i.e., bfqq->bic
2757 * must be non null). If we considered also merged queues,
2758 * then we should also check whether bfqq has already been
2759 * merged with bic->stable_merge_bfqq. But this would be
2760 * costly and complicated.
2762 if (unlikely(!bfqd
->nonrot_with_queueing
)) {
2764 * Make sure also that bfqq is sync, because
2765 * bic->stable_merge_bfqq may point to some queue (for
2766 * stable merging) also if bic is associated with a
2767 * sync queue, but this bfqq is async
2769 if (bfq_bfqq_sync(bfqq
) && bic
->stable_merge_bfqq
&&
2770 !bfq_bfqq_just_created(bfqq
) &&
2771 time_is_before_jiffies(bfqq
->split_time
+
2772 msecs_to_jiffies(bfq_late_stable_merging
)) &&
2773 time_is_before_jiffies(bfqq
->creation_time
+
2774 msecs_to_jiffies(bfq_late_stable_merging
))) {
2775 struct bfq_queue
*stable_merge_bfqq
=
2776 bic
->stable_merge_bfqq
;
2777 int proc_ref
= min(bfqq_process_refs(bfqq
),
2778 bfqq_process_refs(stable_merge_bfqq
));
2780 /* deschedule stable merge, because done or aborted here */
2781 bfq_put_stable_ref(stable_merge_bfqq
);
2783 bic
->stable_merge_bfqq
= NULL
;
2785 if (!idling_boosts_thr_without_issues(bfqd
, bfqq
) &&
2787 /* next function will take at least one ref */
2788 struct bfq_queue
*new_bfqq
=
2789 bfq_setup_merge(bfqq
, stable_merge_bfqq
);
2792 bic
->stably_merged
= true;
2794 new_bfqq
->bic
->stably_merged
=
2804 * Do not perform queue merging if the device is non
2805 * rotational and performs internal queueing. In fact, such a
2806 * device reaches a high speed through internal parallelism
2807 * and pipelining. This means that, to reach a high
2808 * throughput, it must have many requests enqueued at the same
2809 * time. But, in this configuration, the internal scheduling
2810 * algorithm of the device does exactly the job of queue
2811 * merging: it reorders requests so as to obtain as much as
2812 * possible a sequential I/O pattern. As a consequence, with
2813 * the workload generated by processes doing interleaved I/O,
2814 * the throughput reached by the device is likely to be the
2815 * same, with and without queue merging.
2817 * Disabling merging also provides a remarkable benefit in
2818 * terms of throughput. Merging tends to make many workloads
2819 * artificially more uneven, because of shared queues
2820 * remaining non empty for incomparably more time than
2821 * non-merged queues. This may accentuate workload
2822 * asymmetries. For example, if one of the queues in a set of
2823 * merged queues has a higher weight than a normal queue, then
2824 * the shared queue may inherit such a high weight and, by
2825 * staying almost always active, may force BFQ to perform I/O
2826 * plugging most of the time. This evidently makes it harder
2827 * for BFQ to let the device reach a high throughput.
2829 * Finally, the likely() macro below is not used because one
2830 * of the two branches is more likely than the other, but to
2831 * have the code path after the following if() executed as
2832 * fast as possible for the case of a non rotational device
2833 * with queueing. We want it because this is the fastest kind
2834 * of device. On the opposite end, the likely() may lengthen
2835 * the execution time of BFQ for the case of slower devices
2836 * (rotational or at least without queueing). But in this case
2837 * the execution time of BFQ matters very little, if not at
2840 if (likely(bfqd
->nonrot_with_queueing
))
2844 * Prevent bfqq from being merged if it has been created too
2845 * long ago. The idea is that true cooperating processes, and
2846 * thus their associated bfq_queues, are supposed to be
2847 * created shortly after each other. This is the case, e.g.,
2848 * for KVM/QEMU and dump I/O threads. Basing on this
2849 * assumption, the following filtering greatly reduces the
2850 * probability that two non-cooperating processes, which just
2851 * happen to do close I/O for some short time interval, have
2852 * their queues merged by mistake.
2854 if (bfq_too_late_for_merging(bfqq
))
2857 if (!io_struct
|| unlikely(bfqq
== &bfqd
->oom_bfqq
))
2860 /* If there is only one backlogged queue, don't search. */
2861 if (bfq_tot_busy_queues(bfqd
) == 1)
2864 in_service_bfqq
= bfqd
->in_service_queue
;
2866 if (in_service_bfqq
&& in_service_bfqq
!= bfqq
&&
2867 likely(in_service_bfqq
!= &bfqd
->oom_bfqq
) &&
2868 bfq_rq_close_to_sector(io_struct
, request
,
2869 bfqd
->in_serv_last_pos
) &&
2870 bfqq
->entity
.parent
== in_service_bfqq
->entity
.parent
&&
2871 bfq_may_be_close_cooperator(bfqq
, in_service_bfqq
)) {
2872 new_bfqq
= bfq_setup_merge(bfqq
, in_service_bfqq
);
2877 * Check whether there is a cooperator among currently scheduled
2878 * queues. The only thing we need is that the bio/request is not
2879 * NULL, as we need it to establish whether a cooperator exists.
2881 new_bfqq
= bfq_find_close_cooperator(bfqd
, bfqq
,
2882 bfq_io_struct_pos(io_struct
, request
));
2884 if (new_bfqq
&& likely(new_bfqq
!= &bfqd
->oom_bfqq
) &&
2885 bfq_may_be_close_cooperator(bfqq
, new_bfqq
))
2886 return bfq_setup_merge(bfqq
, new_bfqq
);
2891 static void bfq_bfqq_save_state(struct bfq_queue
*bfqq
)
2893 struct bfq_io_cq
*bic
= bfqq
->bic
;
2896 * If !bfqq->bic, the queue is already shared or its requests
2897 * have already been redirected to a shared queue; both idle window
2898 * and weight raising state have already been saved. Do nothing.
2903 bic
->saved_last_serv_time_ns
= bfqq
->last_serv_time_ns
;
2904 bic
->saved_inject_limit
= bfqq
->inject_limit
;
2905 bic
->saved_decrease_time_jif
= bfqq
->decrease_time_jif
;
2907 bic
->saved_weight
= bfqq
->entity
.orig_weight
;
2908 bic
->saved_ttime
= bfqq
->ttime
;
2909 bic
->saved_has_short_ttime
= bfq_bfqq_has_short_ttime(bfqq
);
2910 bic
->saved_IO_bound
= bfq_bfqq_IO_bound(bfqq
);
2911 bic
->saved_io_start_time
= bfqq
->io_start_time
;
2912 bic
->saved_tot_idle_time
= bfqq
->tot_idle_time
;
2913 bic
->saved_in_large_burst
= bfq_bfqq_in_large_burst(bfqq
);
2914 bic
->was_in_burst_list
= !hlist_unhashed(&bfqq
->burst_list_node
);
2915 if (unlikely(bfq_bfqq_just_created(bfqq
) &&
2916 !bfq_bfqq_in_large_burst(bfqq
) &&
2917 bfqq
->bfqd
->low_latency
)) {
2919 * bfqq being merged right after being created: bfqq
2920 * would have deserved interactive weight raising, but
2921 * did not make it to be set in a weight-raised state,
2922 * because of this early merge. Store directly the
2923 * weight-raising state that would have been assigned
2924 * to bfqq, so that to avoid that bfqq unjustly fails
2925 * to enjoy weight raising if split soon.
2927 bic
->saved_wr_coeff
= bfqq
->bfqd
->bfq_wr_coeff
;
2928 bic
->saved_wr_start_at_switch_to_srt
= bfq_smallest_from_now();
2929 bic
->saved_wr_cur_max_time
= bfq_wr_duration(bfqq
->bfqd
);
2930 bic
->saved_last_wr_start_finish
= jiffies
;
2932 bic
->saved_wr_coeff
= bfqq
->wr_coeff
;
2933 bic
->saved_wr_start_at_switch_to_srt
=
2934 bfqq
->wr_start_at_switch_to_srt
;
2935 bic
->saved_service_from_wr
= bfqq
->service_from_wr
;
2936 bic
->saved_last_wr_start_finish
= bfqq
->last_wr_start_finish
;
2937 bic
->saved_wr_cur_max_time
= bfqq
->wr_cur_max_time
;
2943 bfq_reassign_last_bfqq(struct bfq_queue
*cur_bfqq
, struct bfq_queue
*new_bfqq
)
2945 if (cur_bfqq
->entity
.parent
&&
2946 cur_bfqq
->entity
.parent
->last_bfqq_created
== cur_bfqq
)
2947 cur_bfqq
->entity
.parent
->last_bfqq_created
= new_bfqq
;
2948 else if (cur_bfqq
->bfqd
&& cur_bfqq
->bfqd
->last_bfqq_created
== cur_bfqq
)
2949 cur_bfqq
->bfqd
->last_bfqq_created
= new_bfqq
;
2952 void bfq_release_process_ref(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
2955 * To prevent bfqq's service guarantees from being violated,
2956 * bfqq may be left busy, i.e., queued for service, even if
2957 * empty (see comments in __bfq_bfqq_expire() for
2958 * details). But, if no process will send requests to bfqq any
2959 * longer, then there is no point in keeping bfqq queued for
2960 * service. In addition, keeping bfqq queued for service, but
2961 * with no process ref any longer, may have caused bfqq to be
2962 * freed when dequeued from service. But this is assumed to
2965 if (bfq_bfqq_busy(bfqq
) && RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
2966 bfqq
!= bfqd
->in_service_queue
)
2967 bfq_del_bfqq_busy(bfqd
, bfqq
, false);
2969 bfq_reassign_last_bfqq(bfqq
, NULL
);
2971 bfq_put_queue(bfqq
);
2975 bfq_merge_bfqqs(struct bfq_data
*bfqd
, struct bfq_io_cq
*bic
,
2976 struct bfq_queue
*bfqq
, struct bfq_queue
*new_bfqq
)
2978 bfq_log_bfqq(bfqd
, bfqq
, "merging with queue %lu",
2979 (unsigned long)new_bfqq
->pid
);
2980 /* Save weight raising and idle window of the merged queues */
2981 bfq_bfqq_save_state(bfqq
);
2982 bfq_bfqq_save_state(new_bfqq
);
2983 if (bfq_bfqq_IO_bound(bfqq
))
2984 bfq_mark_bfqq_IO_bound(new_bfqq
);
2985 bfq_clear_bfqq_IO_bound(bfqq
);
2988 * The processes associated with bfqq are cooperators of the
2989 * processes associated with new_bfqq. So, if bfqq has a
2990 * waker, then assume that all these processes will be happy
2991 * to let bfqq's waker freely inject I/O when they have no
2994 if (bfqq
->waker_bfqq
&& !new_bfqq
->waker_bfqq
&&
2995 bfqq
->waker_bfqq
!= new_bfqq
) {
2996 new_bfqq
->waker_bfqq
= bfqq
->waker_bfqq
;
2997 new_bfqq
->tentative_waker_bfqq
= NULL
;
3000 * If the waker queue disappears, then
3001 * new_bfqq->waker_bfqq must be reset. So insert
3002 * new_bfqq into the woken_list of the waker. See
3003 * bfq_check_waker for details.
3005 hlist_add_head(&new_bfqq
->woken_list_node
,
3006 &new_bfqq
->waker_bfqq
->woken_list
);
3011 * If bfqq is weight-raised, then let new_bfqq inherit
3012 * weight-raising. To reduce false positives, neglect the case
3013 * where bfqq has just been created, but has not yet made it
3014 * to be weight-raised (which may happen because EQM may merge
3015 * bfqq even before bfq_add_request is executed for the first
3016 * time for bfqq). Handling this case would however be very
3017 * easy, thanks to the flag just_created.
3019 if (new_bfqq
->wr_coeff
== 1 && bfqq
->wr_coeff
> 1) {
3020 new_bfqq
->wr_coeff
= bfqq
->wr_coeff
;
3021 new_bfqq
->wr_cur_max_time
= bfqq
->wr_cur_max_time
;
3022 new_bfqq
->last_wr_start_finish
= bfqq
->last_wr_start_finish
;
3023 new_bfqq
->wr_start_at_switch_to_srt
=
3024 bfqq
->wr_start_at_switch_to_srt
;
3025 if (bfq_bfqq_busy(new_bfqq
))
3026 bfqd
->wr_busy_queues
++;
3027 new_bfqq
->entity
.prio_changed
= 1;
3030 if (bfqq
->wr_coeff
> 1) { /* bfqq has given its wr to new_bfqq */
3032 bfqq
->entity
.prio_changed
= 1;
3033 if (bfq_bfqq_busy(bfqq
))
3034 bfqd
->wr_busy_queues
--;
3037 bfq_log_bfqq(bfqd
, new_bfqq
, "merge_bfqqs: wr_busy %d",
3038 bfqd
->wr_busy_queues
);
3041 * Merge queues (that is, let bic redirect its requests to new_bfqq)
3043 bic_set_bfqq(bic
, new_bfqq
, 1);
3044 bfq_mark_bfqq_coop(new_bfqq
);
3046 * new_bfqq now belongs to at least two bics (it is a shared queue):
3047 * set new_bfqq->bic to NULL. bfqq either:
3048 * - does not belong to any bic any more, and hence bfqq->bic must
3049 * be set to NULL, or
3050 * - is a queue whose owning bics have already been redirected to a
3051 * different queue, hence the queue is destined to not belong to
3052 * any bic soon and bfqq->bic is already NULL (therefore the next
3053 * assignment causes no harm).
3055 new_bfqq
->bic
= NULL
;
3057 * If the queue is shared, the pid is the pid of one of the associated
3058 * processes. Which pid depends on the exact sequence of merge events
3059 * the queue underwent. So printing such a pid is useless and confusing
3060 * because it reports a random pid between those of the associated
3062 * We mark such a queue with a pid -1, and then print SHARED instead of
3063 * a pid in logging messages.
3068 bfq_reassign_last_bfqq(bfqq
, new_bfqq
);
3070 bfq_release_process_ref(bfqd
, bfqq
);
3073 static bool bfq_allow_bio_merge(struct request_queue
*q
, struct request
*rq
,
3076 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
3077 bool is_sync
= op_is_sync(bio
->bi_opf
);
3078 struct bfq_queue
*bfqq
= bfqd
->bio_bfqq
, *new_bfqq
;
3081 * Disallow merge of a sync bio into an async request.
3083 if (is_sync
&& !rq_is_sync(rq
))
3087 * Lookup the bfqq that this bio will be queued with. Allow
3088 * merge only if rq is queued there.
3094 * We take advantage of this function to perform an early merge
3095 * of the queues of possible cooperating processes.
3097 new_bfqq
= bfq_setup_cooperator(bfqd
, bfqq
, bio
, false, bfqd
->bio_bic
);
3100 * bic still points to bfqq, then it has not yet been
3101 * redirected to some other bfq_queue, and a queue
3102 * merge between bfqq and new_bfqq can be safely
3103 * fulfilled, i.e., bic can be redirected to new_bfqq
3104 * and bfqq can be put.
3106 bfq_merge_bfqqs(bfqd
, bfqd
->bio_bic
, bfqq
,
3109 * If we get here, bio will be queued into new_queue,
3110 * so use new_bfqq to decide whether bio and rq can be
3116 * Change also bqfd->bio_bfqq, as
3117 * bfqd->bio_bic now points to new_bfqq, and
3118 * this function may be invoked again (and then may
3119 * use again bqfd->bio_bfqq).
3121 bfqd
->bio_bfqq
= bfqq
;
3124 return bfqq
== RQ_BFQQ(rq
);
3128 * Set the maximum time for the in-service queue to consume its
3129 * budget. This prevents seeky processes from lowering the throughput.
3130 * In practice, a time-slice service scheme is used with seeky
3133 static void bfq_set_budget_timeout(struct bfq_data
*bfqd
,
3134 struct bfq_queue
*bfqq
)
3136 unsigned int timeout_coeff
;
3138 if (bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
)
3141 timeout_coeff
= bfqq
->entity
.weight
/ bfqq
->entity
.orig_weight
;
3143 bfqd
->last_budget_start
= ktime_get();
3145 bfqq
->budget_timeout
= jiffies
+
3146 bfqd
->bfq_timeout
* timeout_coeff
;
3149 static void __bfq_set_in_service_queue(struct bfq_data
*bfqd
,
3150 struct bfq_queue
*bfqq
)
3153 bfq_clear_bfqq_fifo_expire(bfqq
);
3155 bfqd
->budgets_assigned
= (bfqd
->budgets_assigned
* 7 + 256) / 8;
3157 if (time_is_before_jiffies(bfqq
->last_wr_start_finish
) &&
3158 bfqq
->wr_coeff
> 1 &&
3159 bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
&&
3160 time_is_before_jiffies(bfqq
->budget_timeout
)) {
3162 * For soft real-time queues, move the start
3163 * of the weight-raising period forward by the
3164 * time the queue has not received any
3165 * service. Otherwise, a relatively long
3166 * service delay is likely to cause the
3167 * weight-raising period of the queue to end,
3168 * because of the short duration of the
3169 * weight-raising period of a soft real-time
3170 * queue. It is worth noting that this move
3171 * is not so dangerous for the other queues,
3172 * because soft real-time queues are not
3175 * To not add a further variable, we use the
3176 * overloaded field budget_timeout to
3177 * determine for how long the queue has not
3178 * received service, i.e., how much time has
3179 * elapsed since the queue expired. However,
3180 * this is a little imprecise, because
3181 * budget_timeout is set to jiffies if bfqq
3182 * not only expires, but also remains with no
3185 if (time_after(bfqq
->budget_timeout
,
3186 bfqq
->last_wr_start_finish
))
3187 bfqq
->last_wr_start_finish
+=
3188 jiffies
- bfqq
->budget_timeout
;
3190 bfqq
->last_wr_start_finish
= jiffies
;
3193 bfq_set_budget_timeout(bfqd
, bfqq
);
3194 bfq_log_bfqq(bfqd
, bfqq
,
3195 "set_in_service_queue, cur-budget = %d",
3196 bfqq
->entity
.budget
);
3199 bfqd
->in_service_queue
= bfqq
;
3200 bfqd
->in_serv_last_pos
= 0;
3204 * Get and set a new queue for service.
3206 static struct bfq_queue
*bfq_set_in_service_queue(struct bfq_data
*bfqd
)
3208 struct bfq_queue
*bfqq
= bfq_get_next_queue(bfqd
);
3210 __bfq_set_in_service_queue(bfqd
, bfqq
);
3214 static void bfq_arm_slice_timer(struct bfq_data
*bfqd
)
3216 struct bfq_queue
*bfqq
= bfqd
->in_service_queue
;
3219 bfq_mark_bfqq_wait_request(bfqq
);
3222 * We don't want to idle for seeks, but we do want to allow
3223 * fair distribution of slice time for a process doing back-to-back
3224 * seeks. So allow a little bit of time for him to submit a new rq.
3226 sl
= bfqd
->bfq_slice_idle
;
3228 * Unless the queue is being weight-raised or the scenario is
3229 * asymmetric, grant only minimum idle time if the queue
3230 * is seeky. A long idling is preserved for a weight-raised
3231 * queue, or, more in general, in an asymmetric scenario,
3232 * because a long idling is needed for guaranteeing to a queue
3233 * its reserved share of the throughput (in particular, it is
3234 * needed if the queue has a higher weight than some other
3237 if (BFQQ_SEEKY(bfqq
) && bfqq
->wr_coeff
== 1 &&
3238 !bfq_asymmetric_scenario(bfqd
, bfqq
))
3239 sl
= min_t(u64
, sl
, BFQ_MIN_TT
);
3240 else if (bfqq
->wr_coeff
> 1)
3241 sl
= max_t(u32
, sl
, 20ULL * NSEC_PER_MSEC
);
3243 bfqd
->last_idling_start
= ktime_get();
3244 bfqd
->last_idling_start_jiffies
= jiffies
;
3246 hrtimer_start(&bfqd
->idle_slice_timer
, ns_to_ktime(sl
),
3248 bfqg_stats_set_start_idle_time(bfqq_group(bfqq
));
3252 * In autotuning mode, max_budget is dynamically recomputed as the
3253 * amount of sectors transferred in timeout at the estimated peak
3254 * rate. This enables BFQ to utilize a full timeslice with a full
3255 * budget, even if the in-service queue is served at peak rate. And
3256 * this maximises throughput with sequential workloads.
3258 static unsigned long bfq_calc_max_budget(struct bfq_data
*bfqd
)
3260 return (u64
)bfqd
->peak_rate
* USEC_PER_MSEC
*
3261 jiffies_to_msecs(bfqd
->bfq_timeout
)>>BFQ_RATE_SHIFT
;
3265 * Update parameters related to throughput and responsiveness, as a
3266 * function of the estimated peak rate. See comments on
3267 * bfq_calc_max_budget(), and on the ref_wr_duration array.
3269 static void update_thr_responsiveness_params(struct bfq_data
*bfqd
)
3271 if (bfqd
->bfq_user_max_budget
== 0) {
3272 bfqd
->bfq_max_budget
=
3273 bfq_calc_max_budget(bfqd
);
3274 bfq_log(bfqd
, "new max_budget = %d", bfqd
->bfq_max_budget
);
3278 static void bfq_reset_rate_computation(struct bfq_data
*bfqd
,
3281 if (rq
!= NULL
) { /* new rq dispatch now, reset accordingly */
3282 bfqd
->last_dispatch
= bfqd
->first_dispatch
= ktime_get_ns();
3283 bfqd
->peak_rate_samples
= 1;
3284 bfqd
->sequential_samples
= 0;
3285 bfqd
->tot_sectors_dispatched
= bfqd
->last_rq_max_size
=
3287 } else /* no new rq dispatched, just reset the number of samples */
3288 bfqd
->peak_rate_samples
= 0; /* full re-init on next disp. */
3291 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
3292 bfqd
->peak_rate_samples
, bfqd
->sequential_samples
,
3293 bfqd
->tot_sectors_dispatched
);
3296 static void bfq_update_rate_reset(struct bfq_data
*bfqd
, struct request
*rq
)
3298 u32 rate
, weight
, divisor
;
3301 * For the convergence property to hold (see comments on
3302 * bfq_update_peak_rate()) and for the assessment to be
3303 * reliable, a minimum number of samples must be present, and
3304 * a minimum amount of time must have elapsed. If not so, do
3305 * not compute new rate. Just reset parameters, to get ready
3306 * for a new evaluation attempt.
3308 if (bfqd
->peak_rate_samples
< BFQ_RATE_MIN_SAMPLES
||
3309 bfqd
->delta_from_first
< BFQ_RATE_MIN_INTERVAL
)
3310 goto reset_computation
;
3313 * If a new request completion has occurred after last
3314 * dispatch, then, to approximate the rate at which requests
3315 * have been served by the device, it is more precise to
3316 * extend the observation interval to the last completion.
3318 bfqd
->delta_from_first
=
3319 max_t(u64
, bfqd
->delta_from_first
,
3320 bfqd
->last_completion
- bfqd
->first_dispatch
);
3323 * Rate computed in sects/usec, and not sects/nsec, for
3326 rate
= div64_ul(bfqd
->tot_sectors_dispatched
<<BFQ_RATE_SHIFT
,
3327 div_u64(bfqd
->delta_from_first
, NSEC_PER_USEC
));
3330 * Peak rate not updated if:
3331 * - the percentage of sequential dispatches is below 3/4 of the
3332 * total, and rate is below the current estimated peak rate
3333 * - rate is unreasonably high (> 20M sectors/sec)
3335 if ((bfqd
->sequential_samples
< (3 * bfqd
->peak_rate_samples
)>>2 &&
3336 rate
<= bfqd
->peak_rate
) ||
3337 rate
> 20<<BFQ_RATE_SHIFT
)
3338 goto reset_computation
;
3341 * We have to update the peak rate, at last! To this purpose,
3342 * we use a low-pass filter. We compute the smoothing constant
3343 * of the filter as a function of the 'weight' of the new
3346 * As can be seen in next formulas, we define this weight as a
3347 * quantity proportional to how sequential the workload is,
3348 * and to how long the observation time interval is.
3350 * The weight runs from 0 to 8. The maximum value of the
3351 * weight, 8, yields the minimum value for the smoothing
3352 * constant. At this minimum value for the smoothing constant,
3353 * the measured rate contributes for half of the next value of
3354 * the estimated peak rate.
3356 * So, the first step is to compute the weight as a function
3357 * of how sequential the workload is. Note that the weight
3358 * cannot reach 9, because bfqd->sequential_samples cannot
3359 * become equal to bfqd->peak_rate_samples, which, in its
3360 * turn, holds true because bfqd->sequential_samples is not
3361 * incremented for the first sample.
3363 weight
= (9 * bfqd
->sequential_samples
) / bfqd
->peak_rate_samples
;
3366 * Second step: further refine the weight as a function of the
3367 * duration of the observation interval.
3369 weight
= min_t(u32
, 8,
3370 div_u64(weight
* bfqd
->delta_from_first
,
3371 BFQ_RATE_REF_INTERVAL
));
3374 * Divisor ranging from 10, for minimum weight, to 2, for
3377 divisor
= 10 - weight
;
3380 * Finally, update peak rate:
3382 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
3384 bfqd
->peak_rate
*= divisor
-1;
3385 bfqd
->peak_rate
/= divisor
;
3386 rate
/= divisor
; /* smoothing constant alpha = 1/divisor */
3388 bfqd
->peak_rate
+= rate
;
3391 * For a very slow device, bfqd->peak_rate can reach 0 (see
3392 * the minimum representable values reported in the comments
3393 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3394 * divisions by zero where bfqd->peak_rate is used as a
3397 bfqd
->peak_rate
= max_t(u32
, 1, bfqd
->peak_rate
);
3399 update_thr_responsiveness_params(bfqd
);
3402 bfq_reset_rate_computation(bfqd
, rq
);
3406 * Update the read/write peak rate (the main quantity used for
3407 * auto-tuning, see update_thr_responsiveness_params()).
3409 * It is not trivial to estimate the peak rate (correctly): because of
3410 * the presence of sw and hw queues between the scheduler and the
3411 * device components that finally serve I/O requests, it is hard to
3412 * say exactly when a given dispatched request is served inside the
3413 * device, and for how long. As a consequence, it is hard to know
3414 * precisely at what rate a given set of requests is actually served
3417 * On the opposite end, the dispatch time of any request is trivially
3418 * available, and, from this piece of information, the "dispatch rate"
3419 * of requests can be immediately computed. So, the idea in the next
3420 * function is to use what is known, namely request dispatch times
3421 * (plus, when useful, request completion times), to estimate what is
3422 * unknown, namely in-device request service rate.
3424 * The main issue is that, because of the above facts, the rate at
3425 * which a certain set of requests is dispatched over a certain time
3426 * interval can vary greatly with respect to the rate at which the
3427 * same requests are then served. But, since the size of any
3428 * intermediate queue is limited, and the service scheme is lossless
3429 * (no request is silently dropped), the following obvious convergence
3430 * property holds: the number of requests dispatched MUST become
3431 * closer and closer to the number of requests completed as the
3432 * observation interval grows. This is the key property used in
3433 * the next function to estimate the peak service rate as a function
3434 * of the observed dispatch rate. The function assumes to be invoked
3435 * on every request dispatch.
3437 static void bfq_update_peak_rate(struct bfq_data
*bfqd
, struct request
*rq
)
3439 u64 now_ns
= ktime_get_ns();
3441 if (bfqd
->peak_rate_samples
== 0) { /* first dispatch */
3442 bfq_log(bfqd
, "update_peak_rate: goto reset, samples %d",
3443 bfqd
->peak_rate_samples
);
3444 bfq_reset_rate_computation(bfqd
, rq
);
3445 goto update_last_values
; /* will add one sample */
3449 * Device idle for very long: the observation interval lasting
3450 * up to this dispatch cannot be a valid observation interval
3451 * for computing a new peak rate (similarly to the late-
3452 * completion event in bfq_completed_request()). Go to
3453 * update_rate_and_reset to have the following three steps
3455 * - close the observation interval at the last (previous)
3456 * request dispatch or completion
3457 * - compute rate, if possible, for that observation interval
3458 * - start a new observation interval with this dispatch
3460 if (now_ns
- bfqd
->last_dispatch
> 100*NSEC_PER_MSEC
&&
3461 bfqd
->rq_in_driver
== 0)
3462 goto update_rate_and_reset
;
3464 /* Update sampling information */
3465 bfqd
->peak_rate_samples
++;
3467 if ((bfqd
->rq_in_driver
> 0 ||
3468 now_ns
- bfqd
->last_completion
< BFQ_MIN_TT
)
3469 && !BFQ_RQ_SEEKY(bfqd
, bfqd
->last_position
, rq
))
3470 bfqd
->sequential_samples
++;
3472 bfqd
->tot_sectors_dispatched
+= blk_rq_sectors(rq
);
3474 /* Reset max observed rq size every 32 dispatches */
3475 if (likely(bfqd
->peak_rate_samples
% 32))
3476 bfqd
->last_rq_max_size
=
3477 max_t(u32
, blk_rq_sectors(rq
), bfqd
->last_rq_max_size
);
3479 bfqd
->last_rq_max_size
= blk_rq_sectors(rq
);
3481 bfqd
->delta_from_first
= now_ns
- bfqd
->first_dispatch
;
3483 /* Target observation interval not yet reached, go on sampling */
3484 if (bfqd
->delta_from_first
< BFQ_RATE_REF_INTERVAL
)
3485 goto update_last_values
;
3487 update_rate_and_reset
:
3488 bfq_update_rate_reset(bfqd
, rq
);
3490 bfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
3491 if (RQ_BFQQ(rq
) == bfqd
->in_service_queue
)
3492 bfqd
->in_serv_last_pos
= bfqd
->last_position
;
3493 bfqd
->last_dispatch
= now_ns
;
3497 * Remove request from internal lists.
3499 static void bfq_dispatch_remove(struct request_queue
*q
, struct request
*rq
)
3501 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
3504 * For consistency, the next instruction should have been
3505 * executed after removing the request from the queue and
3506 * dispatching it. We execute instead this instruction before
3507 * bfq_remove_request() (and hence introduce a temporary
3508 * inconsistency), for efficiency. In fact, should this
3509 * dispatch occur for a non in-service bfqq, this anticipated
3510 * increment prevents two counters related to bfqq->dispatched
3511 * from risking to be, first, uselessly decremented, and then
3512 * incremented again when the (new) value of bfqq->dispatched
3513 * happens to be taken into account.
3516 bfq_update_peak_rate(q
->elevator
->elevator_data
, rq
);
3518 bfq_remove_request(q
, rq
);
3522 * There is a case where idling does not have to be performed for
3523 * throughput concerns, but to preserve the throughput share of
3524 * the process associated with bfqq.
3526 * To introduce this case, we can note that allowing the drive
3527 * to enqueue more than one request at a time, and hence
3528 * delegating de facto final scheduling decisions to the
3529 * drive's internal scheduler, entails loss of control on the
3530 * actual request service order. In particular, the critical
3531 * situation is when requests from different processes happen
3532 * to be present, at the same time, in the internal queue(s)
3533 * of the drive. In such a situation, the drive, by deciding
3534 * the service order of the internally-queued requests, does
3535 * determine also the actual throughput distribution among
3536 * these processes. But the drive typically has no notion or
3537 * concern about per-process throughput distribution, and
3538 * makes its decisions only on a per-request basis. Therefore,
3539 * the service distribution enforced by the drive's internal
3540 * scheduler is likely to coincide with the desired throughput
3541 * distribution only in a completely symmetric, or favorably
3542 * skewed scenario where:
3543 * (i-a) each of these processes must get the same throughput as
3545 * (i-b) in case (i-a) does not hold, it holds that the process
3546 * associated with bfqq must receive a lower or equal
3547 * throughput than any of the other processes;
3548 * (ii) the I/O of each process has the same properties, in
3549 * terms of locality (sequential or random), direction
3550 * (reads or writes), request sizes, greediness
3551 * (from I/O-bound to sporadic), and so on;
3553 * In fact, in such a scenario, the drive tends to treat the requests
3554 * of each process in about the same way as the requests of the
3555 * others, and thus to provide each of these processes with about the
3556 * same throughput. This is exactly the desired throughput
3557 * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3558 * even more convenient distribution for (the process associated with)
3561 * In contrast, in any asymmetric or unfavorable scenario, device
3562 * idling (I/O-dispatch plugging) is certainly needed to guarantee
3563 * that bfqq receives its assigned fraction of the device throughput
3564 * (see [1] for details).
3566 * The problem is that idling may significantly reduce throughput with
3567 * certain combinations of types of I/O and devices. An important
3568 * example is sync random I/O on flash storage with command
3569 * queueing. So, unless bfqq falls in cases where idling also boosts
3570 * throughput, it is important to check conditions (i-a), i(-b) and
3571 * (ii) accurately, so as to avoid idling when not strictly needed for
3572 * service guarantees.
3574 * Unfortunately, it is extremely difficult to thoroughly check
3575 * condition (ii). And, in case there are active groups, it becomes
3576 * very difficult to check conditions (i-a) and (i-b) too. In fact,
3577 * if there are active groups, then, for conditions (i-a) or (i-b) to
3578 * become false 'indirectly', it is enough that an active group
3579 * contains more active processes or sub-groups than some other active
3580 * group. More precisely, for conditions (i-a) or (i-b) to become
3581 * false because of such a group, it is not even necessary that the
3582 * group is (still) active: it is sufficient that, even if the group
3583 * has become inactive, some of its descendant processes still have
3584 * some request already dispatched but still waiting for
3585 * completion. In fact, requests have still to be guaranteed their
3586 * share of the throughput even after being dispatched. In this
3587 * respect, it is easy to show that, if a group frequently becomes
3588 * inactive while still having in-flight requests, and if, when this
3589 * happens, the group is not considered in the calculation of whether
3590 * the scenario is asymmetric, then the group may fail to be
3591 * guaranteed its fair share of the throughput (basically because
3592 * idling may not be performed for the descendant processes of the
3593 * group, but it had to be). We address this issue with the following
3594 * bi-modal behavior, implemented in the function
3595 * bfq_asymmetric_scenario().
3597 * If there are groups with requests waiting for completion
3598 * (as commented above, some of these groups may even be
3599 * already inactive), then the scenario is tagged as
3600 * asymmetric, conservatively, without checking any of the
3601 * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3602 * This behavior matches also the fact that groups are created
3603 * exactly if controlling I/O is a primary concern (to
3604 * preserve bandwidth and latency guarantees).
3606 * On the opposite end, if there are no groups with requests waiting
3607 * for completion, then only conditions (i-a) and (i-b) are actually
3608 * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3609 * idling is not performed, regardless of whether condition (ii)
3610 * holds. In other words, only if conditions (i-a) and (i-b) do not
3611 * hold, then idling is allowed, and the device tends to be prevented
3612 * from queueing many requests, possibly of several processes. Since
3613 * there are no groups with requests waiting for completion, then, to
3614 * control conditions (i-a) and (i-b) it is enough to check just
3615 * whether all the queues with requests waiting for completion also
3616 * have the same weight.
3618 * Not checking condition (ii) evidently exposes bfqq to the
3619 * risk of getting less throughput than its fair share.
3620 * However, for queues with the same weight, a further
3621 * mechanism, preemption, mitigates or even eliminates this
3622 * problem. And it does so without consequences on overall
3623 * throughput. This mechanism and its benefits are explained
3624 * in the next three paragraphs.
3626 * Even if a queue, say Q, is expired when it remains idle, Q
3627 * can still preempt the new in-service queue if the next
3628 * request of Q arrives soon (see the comments on
3629 * bfq_bfqq_update_budg_for_activation). If all queues and
3630 * groups have the same weight, this form of preemption,
3631 * combined with the hole-recovery heuristic described in the
3632 * comments on function bfq_bfqq_update_budg_for_activation,
3633 * are enough to preserve a correct bandwidth distribution in
3634 * the mid term, even without idling. In fact, even if not
3635 * idling allows the internal queues of the device to contain
3636 * many requests, and thus to reorder requests, we can rather
3637 * safely assume that the internal scheduler still preserves a
3638 * minimum of mid-term fairness.
3640 * More precisely, this preemption-based, idleless approach
3641 * provides fairness in terms of IOPS, and not sectors per
3642 * second. This can be seen with a simple example. Suppose
3643 * that there are two queues with the same weight, but that
3644 * the first queue receives requests of 8 sectors, while the
3645 * second queue receives requests of 1024 sectors. In
3646 * addition, suppose that each of the two queues contains at
3647 * most one request at a time, which implies that each queue
3648 * always remains idle after it is served. Finally, after
3649 * remaining idle, each queue receives very quickly a new
3650 * request. It follows that the two queues are served
3651 * alternatively, preempting each other if needed. This
3652 * implies that, although both queues have the same weight,
3653 * the queue with large requests receives a service that is
3654 * 1024/8 times as high as the service received by the other
3657 * The motivation for using preemption instead of idling (for
3658 * queues with the same weight) is that, by not idling,
3659 * service guarantees are preserved (completely or at least in
3660 * part) without minimally sacrificing throughput. And, if
3661 * there is no active group, then the primary expectation for
3662 * this device is probably a high throughput.
3664 * We are now left only with explaining the two sub-conditions in the
3665 * additional compound condition that is checked below for deciding
3666 * whether the scenario is asymmetric. To explain the first
3667 * sub-condition, we need to add that the function
3668 * bfq_asymmetric_scenario checks the weights of only
3669 * non-weight-raised queues, for efficiency reasons (see comments on
3670 * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3671 * is checked explicitly here. More precisely, the compound condition
3672 * below takes into account also the fact that, even if bfqq is being
3673 * weight-raised, the scenario is still symmetric if all queues with
3674 * requests waiting for completion happen to be
3675 * weight-raised. Actually, we should be even more precise here, and
3676 * differentiate between interactive weight raising and soft real-time
3679 * The second sub-condition checked in the compound condition is
3680 * whether there is a fair amount of already in-flight I/O not
3681 * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3682 * following reason. The drive may decide to serve in-flight
3683 * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3684 * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3685 * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3686 * basically uncontrolled amount of I/O from other queues may be
3687 * dispatched too, possibly causing the service of bfqq's I/O to be
3688 * delayed even longer in the drive. This problem gets more and more
3689 * serious as the speed and the queue depth of the drive grow,
3690 * because, as these two quantities grow, the probability to find no
3691 * queue busy but many requests in flight grows too. By contrast,
3692 * plugging I/O dispatching minimizes the delay induced by already
3693 * in-flight I/O, and enables bfqq to recover the bandwidth it may
3694 * lose because of this delay.
3696 * As a side note, it is worth considering that the above
3697 * device-idling countermeasures may however fail in the following
3698 * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3699 * in a time period during which all symmetry sub-conditions hold, and
3700 * therefore the device is allowed to enqueue many requests, but at
3701 * some later point in time some sub-condition stops to hold, then it
3702 * may become impossible to make requests be served in the desired
3703 * order until all the requests already queued in the device have been
3704 * served. The last sub-condition commented above somewhat mitigates
3705 * this problem for weight-raised queues.
3707 * However, as an additional mitigation for this problem, we preserve
3708 * plugging for a special symmetric case that may suddenly turn into
3709 * asymmetric: the case where only bfqq is busy. In this case, not
3710 * expiring bfqq does not cause any harm to any other queues in terms
3711 * of service guarantees. In contrast, it avoids the following unlucky
3712 * sequence of events: (1) bfqq is expired, (2) a new queue with a
3713 * lower weight than bfqq becomes busy (or more queues), (3) the new
3714 * queue is served until a new request arrives for bfqq, (4) when bfqq
3715 * is finally served, there are so many requests of the new queue in
3716 * the drive that the pending requests for bfqq take a lot of time to
3717 * be served. In particular, event (2) may case even already
3718 * dispatched requests of bfqq to be delayed, inside the drive. So, to
3719 * avoid this series of events, the scenario is preventively declared
3720 * as asymmetric also if bfqq is the only busy queues
3722 static bool idling_needed_for_service_guarantees(struct bfq_data
*bfqd
,
3723 struct bfq_queue
*bfqq
)
3725 int tot_busy_queues
= bfq_tot_busy_queues(bfqd
);
3727 /* No point in idling for bfqq if it won't get requests any longer */
3728 if (unlikely(!bfqq_process_refs(bfqq
)))
3731 return (bfqq
->wr_coeff
> 1 &&
3732 (bfqd
->wr_busy_queues
<
3734 bfqd
->rq_in_driver
>=
3735 bfqq
->dispatched
+ 4)) ||
3736 bfq_asymmetric_scenario(bfqd
, bfqq
) ||
3737 tot_busy_queues
== 1;
3740 static bool __bfq_bfqq_expire(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
3741 enum bfqq_expiration reason
)
3744 * If this bfqq is shared between multiple processes, check
3745 * to make sure that those processes are still issuing I/Os
3746 * within the mean seek distance. If not, it may be time to
3747 * break the queues apart again.
3749 if (bfq_bfqq_coop(bfqq
) && BFQQ_SEEKY(bfqq
))
3750 bfq_mark_bfqq_split_coop(bfqq
);
3753 * Consider queues with a higher finish virtual time than
3754 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3755 * true, then bfqq's bandwidth would be violated if an
3756 * uncontrolled amount of I/O from these queues were
3757 * dispatched while bfqq is waiting for its new I/O to
3758 * arrive. This is exactly what may happen if this is a forced
3759 * expiration caused by a preemption attempt, and if bfqq is
3760 * not re-scheduled. To prevent this from happening, re-queue
3761 * bfqq if it needs I/O-dispatch plugging, even if it is
3762 * empty. By doing so, bfqq is granted to be served before the
3763 * above queues (provided that bfqq is of course eligible).
3765 if (RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
3766 !(reason
== BFQQE_PREEMPTED
&&
3767 idling_needed_for_service_guarantees(bfqd
, bfqq
))) {
3768 if (bfqq
->dispatched
== 0)
3770 * Overloading budget_timeout field to store
3771 * the time at which the queue remains with no
3772 * backlog and no outstanding request; used by
3773 * the weight-raising mechanism.
3775 bfqq
->budget_timeout
= jiffies
;
3777 bfq_del_bfqq_busy(bfqd
, bfqq
, true);
3779 bfq_requeue_bfqq(bfqd
, bfqq
, true);
3781 * Resort priority tree of potential close cooperators.
3782 * See comments on bfq_pos_tree_add_move() for the unlikely().
3784 if (unlikely(!bfqd
->nonrot_with_queueing
&&
3785 !RB_EMPTY_ROOT(&bfqq
->sort_list
)))
3786 bfq_pos_tree_add_move(bfqd
, bfqq
);
3790 * All in-service entities must have been properly deactivated
3791 * or requeued before executing the next function, which
3792 * resets all in-service entities as no more in service. This
3793 * may cause bfqq to be freed. If this happens, the next
3794 * function returns true.
3796 return __bfq_bfqd_reset_in_service(bfqd
);
3800 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3801 * @bfqd: device data.
3802 * @bfqq: queue to update.
3803 * @reason: reason for expiration.
3805 * Handle the feedback on @bfqq budget at queue expiration.
3806 * See the body for detailed comments.
3808 static void __bfq_bfqq_recalc_budget(struct bfq_data
*bfqd
,
3809 struct bfq_queue
*bfqq
,
3810 enum bfqq_expiration reason
)
3812 struct request
*next_rq
;
3813 int budget
, min_budget
;
3815 min_budget
= bfq_min_budget(bfqd
);
3817 if (bfqq
->wr_coeff
== 1)
3818 budget
= bfqq
->max_budget
;
3820 * Use a constant, low budget for weight-raised queues,
3821 * to help achieve a low latency. Keep it slightly higher
3822 * than the minimum possible budget, to cause a little
3823 * bit fewer expirations.
3825 budget
= 2 * min_budget
;
3827 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: last budg %d, budg left %d",
3828 bfqq
->entity
.budget
, bfq_bfqq_budget_left(bfqq
));
3829 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: last max_budg %d, min budg %d",
3830 budget
, bfq_min_budget(bfqd
));
3831 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: sync %d, seeky %d",
3832 bfq_bfqq_sync(bfqq
), BFQQ_SEEKY(bfqd
->in_service_queue
));
3834 if (bfq_bfqq_sync(bfqq
) && bfqq
->wr_coeff
== 1) {
3837 * Caveat: in all the following cases we trade latency
3840 case BFQQE_TOO_IDLE
:
3842 * This is the only case where we may reduce
3843 * the budget: if there is no request of the
3844 * process still waiting for completion, then
3845 * we assume (tentatively) that the timer has
3846 * expired because the batch of requests of
3847 * the process could have been served with a
3848 * smaller budget. Hence, betting that
3849 * process will behave in the same way when it
3850 * becomes backlogged again, we reduce its
3851 * next budget. As long as we guess right,
3852 * this budget cut reduces the latency
3853 * experienced by the process.
3855 * However, if there are still outstanding
3856 * requests, then the process may have not yet
3857 * issued its next request just because it is
3858 * still waiting for the completion of some of
3859 * the still outstanding ones. So in this
3860 * subcase we do not reduce its budget, on the
3861 * contrary we increase it to possibly boost
3862 * the throughput, as discussed in the
3863 * comments to the BUDGET_TIMEOUT case.
3865 if (bfqq
->dispatched
> 0) /* still outstanding reqs */
3866 budget
= min(budget
* 2, bfqd
->bfq_max_budget
);
3868 if (budget
> 5 * min_budget
)
3869 budget
-= 4 * min_budget
;
3871 budget
= min_budget
;
3874 case BFQQE_BUDGET_TIMEOUT
:
3876 * We double the budget here because it gives
3877 * the chance to boost the throughput if this
3878 * is not a seeky process (and has bumped into
3879 * this timeout because of, e.g., ZBR).
3881 budget
= min(budget
* 2, bfqd
->bfq_max_budget
);
3883 case BFQQE_BUDGET_EXHAUSTED
:
3885 * The process still has backlog, and did not
3886 * let either the budget timeout or the disk
3887 * idling timeout expire. Hence it is not
3888 * seeky, has a short thinktime and may be
3889 * happy with a higher budget too. So
3890 * definitely increase the budget of this good
3891 * candidate to boost the disk throughput.
3893 budget
= min(budget
* 4, bfqd
->bfq_max_budget
);
3895 case BFQQE_NO_MORE_REQUESTS
:
3897 * For queues that expire for this reason, it
3898 * is particularly important to keep the
3899 * budget close to the actual service they
3900 * need. Doing so reduces the timestamp
3901 * misalignment problem described in the
3902 * comments in the body of
3903 * __bfq_activate_entity. In fact, suppose
3904 * that a queue systematically expires for
3905 * BFQQE_NO_MORE_REQUESTS and presents a
3906 * new request in time to enjoy timestamp
3907 * back-shifting. The larger the budget of the
3908 * queue is with respect to the service the
3909 * queue actually requests in each service
3910 * slot, the more times the queue can be
3911 * reactivated with the same virtual finish
3912 * time. It follows that, even if this finish
3913 * time is pushed to the system virtual time
3914 * to reduce the consequent timestamp
3915 * misalignment, the queue unjustly enjoys for
3916 * many re-activations a lower finish time
3917 * than all newly activated queues.
3919 * The service needed by bfqq is measured
3920 * quite precisely by bfqq->entity.service.
3921 * Since bfqq does not enjoy device idling,
3922 * bfqq->entity.service is equal to the number
3923 * of sectors that the process associated with
3924 * bfqq requested to read/write before waiting
3925 * for request completions, or blocking for
3928 budget
= max_t(int, bfqq
->entity
.service
, min_budget
);
3933 } else if (!bfq_bfqq_sync(bfqq
)) {
3935 * Async queues get always the maximum possible
3936 * budget, as for them we do not care about latency
3937 * (in addition, their ability to dispatch is limited
3938 * by the charging factor).
3940 budget
= bfqd
->bfq_max_budget
;
3943 bfqq
->max_budget
= budget
;
3945 if (bfqd
->budgets_assigned
>= bfq_stats_min_budgets
&&
3946 !bfqd
->bfq_user_max_budget
)
3947 bfqq
->max_budget
= min(bfqq
->max_budget
, bfqd
->bfq_max_budget
);
3950 * If there is still backlog, then assign a new budget, making
3951 * sure that it is large enough for the next request. Since
3952 * the finish time of bfqq must be kept in sync with the
3953 * budget, be sure to call __bfq_bfqq_expire() *after* this
3956 * If there is no backlog, then no need to update the budget;
3957 * it will be updated on the arrival of a new request.
3959 next_rq
= bfqq
->next_rq
;
3961 bfqq
->entity
.budget
= max_t(unsigned long, bfqq
->max_budget
,
3962 bfq_serv_to_charge(next_rq
, bfqq
));
3964 bfq_log_bfqq(bfqd
, bfqq
, "head sect: %u, new budget %d",
3965 next_rq
? blk_rq_sectors(next_rq
) : 0,
3966 bfqq
->entity
.budget
);
3970 * Return true if the process associated with bfqq is "slow". The slow
3971 * flag is used, in addition to the budget timeout, to reduce the
3972 * amount of service provided to seeky processes, and thus reduce
3973 * their chances to lower the throughput. More details in the comments
3974 * on the function bfq_bfqq_expire().
3976 * An important observation is in order: as discussed in the comments
3977 * on the function bfq_update_peak_rate(), with devices with internal
3978 * queues, it is hard if ever possible to know when and for how long
3979 * an I/O request is processed by the device (apart from the trivial
3980 * I/O pattern where a new request is dispatched only after the
3981 * previous one has been completed). This makes it hard to evaluate
3982 * the real rate at which the I/O requests of each bfq_queue are
3983 * served. In fact, for an I/O scheduler like BFQ, serving a
3984 * bfq_queue means just dispatching its requests during its service
3985 * slot (i.e., until the budget of the queue is exhausted, or the
3986 * queue remains idle, or, finally, a timeout fires). But, during the
3987 * service slot of a bfq_queue, around 100 ms at most, the device may
3988 * be even still processing requests of bfq_queues served in previous
3989 * service slots. On the opposite end, the requests of the in-service
3990 * bfq_queue may be completed after the service slot of the queue
3993 * Anyway, unless more sophisticated solutions are used
3994 * (where possible), the sum of the sizes of the requests dispatched
3995 * during the service slot of a bfq_queue is probably the only
3996 * approximation available for the service received by the bfq_queue
3997 * during its service slot. And this sum is the quantity used in this
3998 * function to evaluate the I/O speed of a process.
4000 static bool bfq_bfqq_is_slow(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
4001 bool compensate
, enum bfqq_expiration reason
,
4002 unsigned long *delta_ms
)
4004 ktime_t delta_ktime
;
4006 bool slow
= BFQQ_SEEKY(bfqq
); /* if delta too short, use seekyness */
4008 if (!bfq_bfqq_sync(bfqq
))
4012 delta_ktime
= bfqd
->last_idling_start
;
4014 delta_ktime
= ktime_get();
4015 delta_ktime
= ktime_sub(delta_ktime
, bfqd
->last_budget_start
);
4016 delta_usecs
= ktime_to_us(delta_ktime
);
4018 /* don't use too short time intervals */
4019 if (delta_usecs
< 1000) {
4020 if (blk_queue_nonrot(bfqd
->queue
))
4022 * give same worst-case guarantees as idling
4025 *delta_ms
= BFQ_MIN_TT
/ NSEC_PER_MSEC
;
4026 else /* charge at least one seek */
4027 *delta_ms
= bfq_slice_idle
/ NSEC_PER_MSEC
;
4032 *delta_ms
= delta_usecs
/ USEC_PER_MSEC
;
4035 * Use only long (> 20ms) intervals to filter out excessive
4036 * spikes in service rate estimation.
4038 if (delta_usecs
> 20000) {
4040 * Caveat for rotational devices: processes doing I/O
4041 * in the slower disk zones tend to be slow(er) even
4042 * if not seeky. In this respect, the estimated peak
4043 * rate is likely to be an average over the disk
4044 * surface. Accordingly, to not be too harsh with
4045 * unlucky processes, a process is deemed slow only if
4046 * its rate has been lower than half of the estimated
4049 slow
= bfqq
->entity
.service
< bfqd
->bfq_max_budget
/ 2;
4052 bfq_log_bfqq(bfqd
, bfqq
, "bfq_bfqq_is_slow: slow %d", slow
);
4058 * To be deemed as soft real-time, an application must meet two
4059 * requirements. First, the application must not require an average
4060 * bandwidth higher than the approximate bandwidth required to playback or
4061 * record a compressed high-definition video.
4062 * The next function is invoked on the completion of the last request of a
4063 * batch, to compute the next-start time instant, soft_rt_next_start, such
4064 * that, if the next request of the application does not arrive before
4065 * soft_rt_next_start, then the above requirement on the bandwidth is met.
4067 * The second requirement is that the request pattern of the application is
4068 * isochronous, i.e., that, after issuing a request or a batch of requests,
4069 * the application stops issuing new requests until all its pending requests
4070 * have been completed. After that, the application may issue a new batch,
4072 * For this reason the next function is invoked to compute
4073 * soft_rt_next_start only for applications that meet this requirement,
4074 * whereas soft_rt_next_start is set to infinity for applications that do
4077 * Unfortunately, even a greedy (i.e., I/O-bound) application may
4078 * happen to meet, occasionally or systematically, both the above
4079 * bandwidth and isochrony requirements. This may happen at least in
4080 * the following circumstances. First, if the CPU load is high. The
4081 * application may stop issuing requests while the CPUs are busy
4082 * serving other processes, then restart, then stop again for a while,
4083 * and so on. The other circumstances are related to the storage
4084 * device: the storage device is highly loaded or reaches a low-enough
4085 * throughput with the I/O of the application (e.g., because the I/O
4086 * is random and/or the device is slow). In all these cases, the
4087 * I/O of the application may be simply slowed down enough to meet
4088 * the bandwidth and isochrony requirements. To reduce the probability
4089 * that greedy applications are deemed as soft real-time in these
4090 * corner cases, a further rule is used in the computation of
4091 * soft_rt_next_start: the return value of this function is forced to
4092 * be higher than the maximum between the following two quantities.
4094 * (a) Current time plus: (1) the maximum time for which the arrival
4095 * of a request is waited for when a sync queue becomes idle,
4096 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
4097 * postpone for a moment the reason for adding a few extra
4098 * jiffies; we get back to it after next item (b). Lower-bounding
4099 * the return value of this function with the current time plus
4100 * bfqd->bfq_slice_idle tends to filter out greedy applications,
4101 * because the latter issue their next request as soon as possible
4102 * after the last one has been completed. In contrast, a soft
4103 * real-time application spends some time processing data, after a
4104 * batch of its requests has been completed.
4106 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
4107 * above, greedy applications may happen to meet both the
4108 * bandwidth and isochrony requirements under heavy CPU or
4109 * storage-device load. In more detail, in these scenarios, these
4110 * applications happen, only for limited time periods, to do I/O
4111 * slowly enough to meet all the requirements described so far,
4112 * including the filtering in above item (a). These slow-speed
4113 * time intervals are usually interspersed between other time
4114 * intervals during which these applications do I/O at a very high
4115 * speed. Fortunately, exactly because of the high speed of the
4116 * I/O in the high-speed intervals, the values returned by this
4117 * function happen to be so high, near the end of any such
4118 * high-speed interval, to be likely to fall *after* the end of
4119 * the low-speed time interval that follows. These high values are
4120 * stored in bfqq->soft_rt_next_start after each invocation of
4121 * this function. As a consequence, if the last value of
4122 * bfqq->soft_rt_next_start is constantly used to lower-bound the
4123 * next value that this function may return, then, from the very
4124 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
4125 * likely to be constantly kept so high that any I/O request
4126 * issued during the low-speed interval is considered as arriving
4127 * to soon for the application to be deemed as soft
4128 * real-time. Then, in the high-speed interval that follows, the
4129 * application will not be deemed as soft real-time, just because
4130 * it will do I/O at a high speed. And so on.
4132 * Getting back to the filtering in item (a), in the following two
4133 * cases this filtering might be easily passed by a greedy
4134 * application, if the reference quantity was just
4135 * bfqd->bfq_slice_idle:
4136 * 1) HZ is so low that the duration of a jiffy is comparable to or
4137 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
4138 * devices with HZ=100. The time granularity may be so coarse
4139 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
4140 * is rather lower than the exact value.
4141 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
4142 * for a while, then suddenly 'jump' by several units to recover the lost
4143 * increments. This seems to happen, e.g., inside virtual machines.
4144 * To address this issue, in the filtering in (a) we do not use as a
4145 * reference time interval just bfqd->bfq_slice_idle, but
4146 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
4147 * minimum number of jiffies for which the filter seems to be quite
4148 * precise also in embedded systems and KVM/QEMU virtual machines.
4150 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data
*bfqd
,
4151 struct bfq_queue
*bfqq
)
4153 return max3(bfqq
->soft_rt_next_start
,
4154 bfqq
->last_idle_bklogged
+
4155 HZ
* bfqq
->service_from_backlogged
/
4156 bfqd
->bfq_wr_max_softrt_rate
,
4157 jiffies
+ nsecs_to_jiffies(bfqq
->bfqd
->bfq_slice_idle
) + 4);
4161 * bfq_bfqq_expire - expire a queue.
4162 * @bfqd: device owning the queue.
4163 * @bfqq: the queue to expire.
4164 * @compensate: if true, compensate for the time spent idling.
4165 * @reason: the reason causing the expiration.
4167 * If the process associated with bfqq does slow I/O (e.g., because it
4168 * issues random requests), we charge bfqq with the time it has been
4169 * in service instead of the service it has received (see
4170 * bfq_bfqq_charge_time for details on how this goal is achieved). As
4171 * a consequence, bfqq will typically get higher timestamps upon
4172 * reactivation, and hence it will be rescheduled as if it had
4173 * received more service than what it has actually received. In the
4174 * end, bfqq receives less service in proportion to how slowly its
4175 * associated process consumes its budgets (and hence how seriously it
4176 * tends to lower the throughput). In addition, this time-charging
4177 * strategy guarantees time fairness among slow processes. In
4178 * contrast, if the process associated with bfqq is not slow, we
4179 * charge bfqq exactly with the service it has received.
4181 * Charging time to the first type of queues and the exact service to
4182 * the other has the effect of using the WF2Q+ policy to schedule the
4183 * former on a timeslice basis, without violating service domain
4184 * guarantees among the latter.
4186 void bfq_bfqq_expire(struct bfq_data
*bfqd
,
4187 struct bfq_queue
*bfqq
,
4189 enum bfqq_expiration reason
)
4192 unsigned long delta
= 0;
4193 struct bfq_entity
*entity
= &bfqq
->entity
;
4196 * Check whether the process is slow (see bfq_bfqq_is_slow).
4198 slow
= bfq_bfqq_is_slow(bfqd
, bfqq
, compensate
, reason
, &delta
);
4201 * As above explained, charge slow (typically seeky) and
4202 * timed-out queues with the time and not the service
4203 * received, to favor sequential workloads.
4205 * Processes doing I/O in the slower disk zones will tend to
4206 * be slow(er) even if not seeky. Therefore, since the
4207 * estimated peak rate is actually an average over the disk
4208 * surface, these processes may timeout just for bad luck. To
4209 * avoid punishing them, do not charge time to processes that
4210 * succeeded in consuming at least 2/3 of their budget. This
4211 * allows BFQ to preserve enough elasticity to still perform
4212 * bandwidth, and not time, distribution with little unlucky
4213 * or quasi-sequential processes.
4215 if (bfqq
->wr_coeff
== 1 &&
4217 (reason
== BFQQE_BUDGET_TIMEOUT
&&
4218 bfq_bfqq_budget_left(bfqq
) >= entity
->budget
/ 3)))
4219 bfq_bfqq_charge_time(bfqd
, bfqq
, delta
);
4221 if (bfqd
->low_latency
&& bfqq
->wr_coeff
== 1)
4222 bfqq
->last_wr_start_finish
= jiffies
;
4224 if (bfqd
->low_latency
&& bfqd
->bfq_wr_max_softrt_rate
> 0 &&
4225 RB_EMPTY_ROOT(&bfqq
->sort_list
)) {
4227 * If we get here, and there are no outstanding
4228 * requests, then the request pattern is isochronous
4229 * (see the comments on the function
4230 * bfq_bfqq_softrt_next_start()). Therefore we can
4231 * compute soft_rt_next_start.
4233 * If, instead, the queue still has outstanding
4234 * requests, then we have to wait for the completion
4235 * of all the outstanding requests to discover whether
4236 * the request pattern is actually isochronous.
4238 if (bfqq
->dispatched
== 0)
4239 bfqq
->soft_rt_next_start
=
4240 bfq_bfqq_softrt_next_start(bfqd
, bfqq
);
4241 else if (bfqq
->dispatched
> 0) {
4243 * Schedule an update of soft_rt_next_start to when
4244 * the task may be discovered to be isochronous.
4246 bfq_mark_bfqq_softrt_update(bfqq
);
4250 bfq_log_bfqq(bfqd
, bfqq
,
4251 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason
,
4252 slow
, bfqq
->dispatched
, bfq_bfqq_has_short_ttime(bfqq
));
4255 * bfqq expired, so no total service time needs to be computed
4256 * any longer: reset state machine for measuring total service
4259 bfqd
->rqs_injected
= bfqd
->wait_dispatch
= false;
4260 bfqd
->waited_rq
= NULL
;
4263 * Increase, decrease or leave budget unchanged according to
4266 __bfq_bfqq_recalc_budget(bfqd
, bfqq
, reason
);
4267 if (__bfq_bfqq_expire(bfqd
, bfqq
, reason
))
4268 /* bfqq is gone, no more actions on it */
4271 /* mark bfqq as waiting a request only if a bic still points to it */
4272 if (!bfq_bfqq_busy(bfqq
) &&
4273 reason
!= BFQQE_BUDGET_TIMEOUT
&&
4274 reason
!= BFQQE_BUDGET_EXHAUSTED
) {
4275 bfq_mark_bfqq_non_blocking_wait_rq(bfqq
);
4277 * Not setting service to 0, because, if the next rq
4278 * arrives in time, the queue will go on receiving
4279 * service with this same budget (as if it never expired)
4282 entity
->service
= 0;
4285 * Reset the received-service counter for every parent entity.
4286 * Differently from what happens with bfqq->entity.service,
4287 * the resetting of this counter never needs to be postponed
4288 * for parent entities. In fact, in case bfqq may have a
4289 * chance to go on being served using the last, partially
4290 * consumed budget, bfqq->entity.service needs to be kept,
4291 * because if bfqq then actually goes on being served using
4292 * the same budget, the last value of bfqq->entity.service is
4293 * needed to properly decrement bfqq->entity.budget by the
4294 * portion already consumed. In contrast, it is not necessary
4295 * to keep entity->service for parent entities too, because
4296 * the bubble up of the new value of bfqq->entity.budget will
4297 * make sure that the budgets of parent entities are correct,
4298 * even in case bfqq and thus parent entities go on receiving
4299 * service with the same budget.
4301 entity
= entity
->parent
;
4302 for_each_entity(entity
)
4303 entity
->service
= 0;
4307 * Budget timeout is not implemented through a dedicated timer, but
4308 * just checked on request arrivals and completions, as well as on
4309 * idle timer expirations.
4311 static bool bfq_bfqq_budget_timeout(struct bfq_queue
*bfqq
)
4313 return time_is_before_eq_jiffies(bfqq
->budget_timeout
);
4317 * If we expire a queue that is actively waiting (i.e., with the
4318 * device idled) for the arrival of a new request, then we may incur
4319 * the timestamp misalignment problem described in the body of the
4320 * function __bfq_activate_entity. Hence we return true only if this
4321 * condition does not hold, or if the queue is slow enough to deserve
4322 * only to be kicked off for preserving a high throughput.
4324 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue
*bfqq
)
4326 bfq_log_bfqq(bfqq
->bfqd
, bfqq
,
4327 "may_budget_timeout: wait_request %d left %d timeout %d",
4328 bfq_bfqq_wait_request(bfqq
),
4329 bfq_bfqq_budget_left(bfqq
) >= bfqq
->entity
.budget
/ 3,
4330 bfq_bfqq_budget_timeout(bfqq
));
4332 return (!bfq_bfqq_wait_request(bfqq
) ||
4333 bfq_bfqq_budget_left(bfqq
) >= bfqq
->entity
.budget
/ 3)
4335 bfq_bfqq_budget_timeout(bfqq
);
4338 static bool idling_boosts_thr_without_issues(struct bfq_data
*bfqd
,
4339 struct bfq_queue
*bfqq
)
4341 bool rot_without_queueing
=
4342 !blk_queue_nonrot(bfqd
->queue
) && !bfqd
->hw_tag
,
4343 bfqq_sequential_and_IO_bound
,
4346 /* No point in idling for bfqq if it won't get requests any longer */
4347 if (unlikely(!bfqq_process_refs(bfqq
)))
4350 bfqq_sequential_and_IO_bound
= !BFQQ_SEEKY(bfqq
) &&
4351 bfq_bfqq_IO_bound(bfqq
) && bfq_bfqq_has_short_ttime(bfqq
);
4354 * The next variable takes into account the cases where idling
4355 * boosts the throughput.
4357 * The value of the variable is computed considering, first, that
4358 * idling is virtually always beneficial for the throughput if:
4359 * (a) the device is not NCQ-capable and rotational, or
4360 * (b) regardless of the presence of NCQ, the device is rotational and
4361 * the request pattern for bfqq is I/O-bound and sequential, or
4362 * (c) regardless of whether it is rotational, the device is
4363 * not NCQ-capable and the request pattern for bfqq is
4364 * I/O-bound and sequential.
4366 * Secondly, and in contrast to the above item (b), idling an
4367 * NCQ-capable flash-based device would not boost the
4368 * throughput even with sequential I/O; rather it would lower
4369 * the throughput in proportion to how fast the device
4370 * is. Accordingly, the next variable is true if any of the
4371 * above conditions (a), (b) or (c) is true, and, in
4372 * particular, happens to be false if bfqd is an NCQ-capable
4373 * flash-based device.
4375 idling_boosts_thr
= rot_without_queueing
||
4376 ((!blk_queue_nonrot(bfqd
->queue
) || !bfqd
->hw_tag
) &&
4377 bfqq_sequential_and_IO_bound
);
4380 * The return value of this function is equal to that of
4381 * idling_boosts_thr, unless a special case holds. In this
4382 * special case, described below, idling may cause problems to
4383 * weight-raised queues.
4385 * When the request pool is saturated (e.g., in the presence
4386 * of write hogs), if the processes associated with
4387 * non-weight-raised queues ask for requests at a lower rate,
4388 * then processes associated with weight-raised queues have a
4389 * higher probability to get a request from the pool
4390 * immediately (or at least soon) when they need one. Thus
4391 * they have a higher probability to actually get a fraction
4392 * of the device throughput proportional to their high
4393 * weight. This is especially true with NCQ-capable drives,
4394 * which enqueue several requests in advance, and further
4395 * reorder internally-queued requests.
4397 * For this reason, we force to false the return value if
4398 * there are weight-raised busy queues. In this case, and if
4399 * bfqq is not weight-raised, this guarantees that the device
4400 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4401 * then idling will be guaranteed by another variable, see
4402 * below). Combined with the timestamping rules of BFQ (see
4403 * [1] for details), this behavior causes bfqq, and hence any
4404 * sync non-weight-raised queue, to get a lower number of
4405 * requests served, and thus to ask for a lower number of
4406 * requests from the request pool, before the busy
4407 * weight-raised queues get served again. This often mitigates
4408 * starvation problems in the presence of heavy write
4409 * workloads and NCQ, thereby guaranteeing a higher
4410 * application and system responsiveness in these hostile
4413 return idling_boosts_thr
&&
4414 bfqd
->wr_busy_queues
== 0;
4418 * For a queue that becomes empty, device idling is allowed only if
4419 * this function returns true for that queue. As a consequence, since
4420 * device idling plays a critical role for both throughput boosting
4421 * and service guarantees, the return value of this function plays a
4422 * critical role as well.
4424 * In a nutshell, this function returns true only if idling is
4425 * beneficial for throughput or, even if detrimental for throughput,
4426 * idling is however necessary to preserve service guarantees (low
4427 * latency, desired throughput distribution, ...). In particular, on
4428 * NCQ-capable devices, this function tries to return false, so as to
4429 * help keep the drives' internal queues full, whenever this helps the
4430 * device boost the throughput without causing any service-guarantee
4433 * Most of the issues taken into account to get the return value of
4434 * this function are not trivial. We discuss these issues in the two
4435 * functions providing the main pieces of information needed by this
4438 static bool bfq_better_to_idle(struct bfq_queue
*bfqq
)
4440 struct bfq_data
*bfqd
= bfqq
->bfqd
;
4441 bool idling_boosts_thr_with_no_issue
, idling_needed_for_service_guar
;
4443 /* No point in idling for bfqq if it won't get requests any longer */
4444 if (unlikely(!bfqq_process_refs(bfqq
)))
4447 if (unlikely(bfqd
->strict_guarantees
))
4451 * Idling is performed only if slice_idle > 0. In addition, we
4454 * (b) bfqq is in the idle io prio class: in this case we do
4455 * not idle because we want to minimize the bandwidth that
4456 * queues in this class can steal to higher-priority queues
4458 if (bfqd
->bfq_slice_idle
== 0 || !bfq_bfqq_sync(bfqq
) ||
4459 bfq_class_idle(bfqq
))
4462 idling_boosts_thr_with_no_issue
=
4463 idling_boosts_thr_without_issues(bfqd
, bfqq
);
4465 idling_needed_for_service_guar
=
4466 idling_needed_for_service_guarantees(bfqd
, bfqq
);
4469 * We have now the two components we need to compute the
4470 * return value of the function, which is true only if idling
4471 * either boosts the throughput (without issues), or is
4472 * necessary to preserve service guarantees.
4474 return idling_boosts_thr_with_no_issue
||
4475 idling_needed_for_service_guar
;
4479 * If the in-service queue is empty but the function bfq_better_to_idle
4480 * returns true, then:
4481 * 1) the queue must remain in service and cannot be expired, and
4482 * 2) the device must be idled to wait for the possible arrival of a new
4483 * request for the queue.
4484 * See the comments on the function bfq_better_to_idle for the reasons
4485 * why performing device idling is the best choice to boost the throughput
4486 * and preserve service guarantees when bfq_better_to_idle itself
4489 static bool bfq_bfqq_must_idle(struct bfq_queue
*bfqq
)
4491 return RB_EMPTY_ROOT(&bfqq
->sort_list
) && bfq_better_to_idle(bfqq
);
4495 * This function chooses the queue from which to pick the next extra
4496 * I/O request to inject, if it finds a compatible queue. See the
4497 * comments on bfq_update_inject_limit() for details on the injection
4498 * mechanism, and for the definitions of the quantities mentioned
4501 static struct bfq_queue
*
4502 bfq_choose_bfqq_for_injection(struct bfq_data
*bfqd
)
4504 struct bfq_queue
*bfqq
, *in_serv_bfqq
= bfqd
->in_service_queue
;
4505 unsigned int limit
= in_serv_bfqq
->inject_limit
;
4508 * - bfqq is not weight-raised and therefore does not carry
4509 * time-critical I/O,
4511 * - regardless of whether bfqq is weight-raised, bfqq has
4512 * however a long think time, during which it can absorb the
4513 * effect of an appropriate number of extra I/O requests
4514 * from other queues (see bfq_update_inject_limit for
4515 * details on the computation of this number);
4516 * then injection can be performed without restrictions.
4518 bool in_serv_always_inject
= in_serv_bfqq
->wr_coeff
== 1 ||
4519 !bfq_bfqq_has_short_ttime(in_serv_bfqq
);
4523 * - the baseline total service time could not be sampled yet,
4524 * so the inject limit happens to be still 0, and
4525 * - a lot of time has elapsed since the plugging of I/O
4526 * dispatching started, so drive speed is being wasted
4528 * then temporarily raise inject limit to one request.
4530 if (limit
== 0 && in_serv_bfqq
->last_serv_time_ns
== 0 &&
4531 bfq_bfqq_wait_request(in_serv_bfqq
) &&
4532 time_is_before_eq_jiffies(bfqd
->last_idling_start_jiffies
+
4533 bfqd
->bfq_slice_idle
)
4537 if (bfqd
->rq_in_driver
>= limit
)
4541 * Linear search of the source queue for injection; but, with
4542 * a high probability, very few steps are needed to find a
4543 * candidate queue, i.e., a queue with enough budget left for
4544 * its next request. In fact:
4545 * - BFQ dynamically updates the budget of every queue so as
4546 * to accommodate the expected backlog of the queue;
4547 * - if a queue gets all its requests dispatched as injected
4548 * service, then the queue is removed from the active list
4549 * (and re-added only if it gets new requests, but then it
4550 * is assigned again enough budget for its new backlog).
4552 list_for_each_entry(bfqq
, &bfqd
->active_list
, bfqq_list
)
4553 if (!RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
4554 (in_serv_always_inject
|| bfqq
->wr_coeff
> 1) &&
4555 bfq_serv_to_charge(bfqq
->next_rq
, bfqq
) <=
4556 bfq_bfqq_budget_left(bfqq
)) {
4558 * Allow for only one large in-flight request
4559 * on non-rotational devices, for the
4560 * following reason. On non-rotationl drives,
4561 * large requests take much longer than
4562 * smaller requests to be served. In addition,
4563 * the drive prefers to serve large requests
4564 * w.r.t. to small ones, if it can choose. So,
4565 * having more than one large requests queued
4566 * in the drive may easily make the next first
4567 * request of the in-service queue wait for so
4568 * long to break bfqq's service guarantees. On
4569 * the bright side, large requests let the
4570 * drive reach a very high throughput, even if
4571 * there is only one in-flight large request
4574 if (blk_queue_nonrot(bfqd
->queue
) &&
4575 blk_rq_sectors(bfqq
->next_rq
) >=
4576 BFQQ_SECT_THR_NONROT
)
4577 limit
= min_t(unsigned int, 1, limit
);
4579 limit
= in_serv_bfqq
->inject_limit
;
4581 if (bfqd
->rq_in_driver
< limit
) {
4582 bfqd
->rqs_injected
= true;
4591 * Select a queue for service. If we have a current queue in service,
4592 * check whether to continue servicing it, or retrieve and set a new one.
4594 static struct bfq_queue
*bfq_select_queue(struct bfq_data
*bfqd
)
4596 struct bfq_queue
*bfqq
;
4597 struct request
*next_rq
;
4598 enum bfqq_expiration reason
= BFQQE_BUDGET_TIMEOUT
;
4600 bfqq
= bfqd
->in_service_queue
;
4604 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: already in-service queue");
4607 * Do not expire bfqq for budget timeout if bfqq may be about
4608 * to enjoy device idling. The reason why, in this case, we
4609 * prevent bfqq from expiring is the same as in the comments
4610 * on the case where bfq_bfqq_must_idle() returns true, in
4611 * bfq_completed_request().
4613 if (bfq_may_expire_for_budg_timeout(bfqq
) &&
4614 !bfq_bfqq_must_idle(bfqq
))
4619 * This loop is rarely executed more than once. Even when it
4620 * happens, it is much more convenient to re-execute this loop
4621 * than to return NULL and trigger a new dispatch to get a
4624 next_rq
= bfqq
->next_rq
;
4626 * If bfqq has requests queued and it has enough budget left to
4627 * serve them, keep the queue, otherwise expire it.
4630 if (bfq_serv_to_charge(next_rq
, bfqq
) >
4631 bfq_bfqq_budget_left(bfqq
)) {
4633 * Expire the queue for budget exhaustion,
4634 * which makes sure that the next budget is
4635 * enough to serve the next request, even if
4636 * it comes from the fifo expired path.
4638 reason
= BFQQE_BUDGET_EXHAUSTED
;
4642 * The idle timer may be pending because we may
4643 * not disable disk idling even when a new request
4646 if (bfq_bfqq_wait_request(bfqq
)) {
4648 * If we get here: 1) at least a new request
4649 * has arrived but we have not disabled the
4650 * timer because the request was too small,
4651 * 2) then the block layer has unplugged
4652 * the device, causing the dispatch to be
4655 * Since the device is unplugged, now the
4656 * requests are probably large enough to
4657 * provide a reasonable throughput.
4658 * So we disable idling.
4660 bfq_clear_bfqq_wait_request(bfqq
);
4661 hrtimer_try_to_cancel(&bfqd
->idle_slice_timer
);
4668 * No requests pending. However, if the in-service queue is idling
4669 * for a new request, or has requests waiting for a completion and
4670 * may idle after their completion, then keep it anyway.
4672 * Yet, inject service from other queues if it boosts
4673 * throughput and is possible.
4675 if (bfq_bfqq_wait_request(bfqq
) ||
4676 (bfqq
->dispatched
!= 0 && bfq_better_to_idle(bfqq
))) {
4677 struct bfq_queue
*async_bfqq
=
4678 bfqq
->bic
&& bfqq
->bic
->bfqq
[0] &&
4679 bfq_bfqq_busy(bfqq
->bic
->bfqq
[0]) &&
4680 bfqq
->bic
->bfqq
[0]->next_rq
?
4681 bfqq
->bic
->bfqq
[0] : NULL
;
4682 struct bfq_queue
*blocked_bfqq
=
4683 !hlist_empty(&bfqq
->woken_list
) ?
4684 container_of(bfqq
->woken_list
.first
,
4690 * The next four mutually-exclusive ifs decide
4691 * whether to try injection, and choose the queue to
4692 * pick an I/O request from.
4694 * The first if checks whether the process associated
4695 * with bfqq has also async I/O pending. If so, it
4696 * injects such I/O unconditionally. Injecting async
4697 * I/O from the same process can cause no harm to the
4698 * process. On the contrary, it can only increase
4699 * bandwidth and reduce latency for the process.
4701 * The second if checks whether there happens to be a
4702 * non-empty waker queue for bfqq, i.e., a queue whose
4703 * I/O needs to be completed for bfqq to receive new
4704 * I/O. This happens, e.g., if bfqq is associated with
4705 * a process that does some sync. A sync generates
4706 * extra blocking I/O, which must be completed before
4707 * the process associated with bfqq can go on with its
4708 * I/O. If the I/O of the waker queue is not served,
4709 * then bfqq remains empty, and no I/O is dispatched,
4710 * until the idle timeout fires for bfqq. This is
4711 * likely to result in lower bandwidth and higher
4712 * latencies for bfqq, and in a severe loss of total
4713 * throughput. The best action to take is therefore to
4714 * serve the waker queue as soon as possible. So do it
4715 * (without relying on the third alternative below for
4716 * eventually serving waker_bfqq's I/O; see the last
4717 * paragraph for further details). This systematic
4718 * injection of I/O from the waker queue does not
4719 * cause any delay to bfqq's I/O. On the contrary,
4720 * next bfqq's I/O is brought forward dramatically,
4721 * for it is not blocked for milliseconds.
4723 * The third if checks whether there is a queue woken
4724 * by bfqq, and currently with pending I/O. Such a
4725 * woken queue does not steal bandwidth from bfqq,
4726 * because it remains soon without I/O if bfqq is not
4727 * served. So there is virtually no risk of loss of
4728 * bandwidth for bfqq if this woken queue has I/O
4729 * dispatched while bfqq is waiting for new I/O.
4731 * The fourth if checks whether bfqq is a queue for
4732 * which it is better to avoid injection. It is so if
4733 * bfqq delivers more throughput when served without
4734 * any further I/O from other queues in the middle, or
4735 * if the service times of bfqq's I/O requests both
4736 * count more than overall throughput, and may be
4737 * easily increased by injection (this happens if bfqq
4738 * has a short think time). If none of these
4739 * conditions holds, then a candidate queue for
4740 * injection is looked for through
4741 * bfq_choose_bfqq_for_injection(). Note that the
4742 * latter may return NULL (for example if the inject
4743 * limit for bfqq is currently 0).
4745 * NOTE: motivation for the second alternative
4747 * Thanks to the way the inject limit is updated in
4748 * bfq_update_has_short_ttime(), it is rather likely
4749 * that, if I/O is being plugged for bfqq and the
4750 * waker queue has pending I/O requests that are
4751 * blocking bfqq's I/O, then the fourth alternative
4752 * above lets the waker queue get served before the
4753 * I/O-plugging timeout fires. So one may deem the
4754 * second alternative superfluous. It is not, because
4755 * the fourth alternative may be way less effective in
4756 * case of a synchronization. For two main
4757 * reasons. First, throughput may be low because the
4758 * inject limit may be too low to guarantee the same
4759 * amount of injected I/O, from the waker queue or
4760 * other queues, that the second alternative
4761 * guarantees (the second alternative unconditionally
4762 * injects a pending I/O request of the waker queue
4763 * for each bfq_dispatch_request()). Second, with the
4764 * fourth alternative, the duration of the plugging,
4765 * i.e., the time before bfqq finally receives new I/O,
4766 * may not be minimized, because the waker queue may
4767 * happen to be served only after other queues.
4770 icq_to_bic(async_bfqq
->next_rq
->elv
.icq
) == bfqq
->bic
&&
4771 bfq_serv_to_charge(async_bfqq
->next_rq
, async_bfqq
) <=
4772 bfq_bfqq_budget_left(async_bfqq
))
4773 bfqq
= bfqq
->bic
->bfqq
[0];
4774 else if (bfqq
->waker_bfqq
&&
4775 bfq_bfqq_busy(bfqq
->waker_bfqq
) &&
4776 bfqq
->waker_bfqq
->next_rq
&&
4777 bfq_serv_to_charge(bfqq
->waker_bfqq
->next_rq
,
4778 bfqq
->waker_bfqq
) <=
4779 bfq_bfqq_budget_left(bfqq
->waker_bfqq
)
4781 bfqq
= bfqq
->waker_bfqq
;
4782 else if (blocked_bfqq
&&
4783 bfq_bfqq_busy(blocked_bfqq
) &&
4784 blocked_bfqq
->next_rq
&&
4785 bfq_serv_to_charge(blocked_bfqq
->next_rq
,
4787 bfq_bfqq_budget_left(blocked_bfqq
)
4789 bfqq
= blocked_bfqq
;
4790 else if (!idling_boosts_thr_without_issues(bfqd
, bfqq
) &&
4791 (bfqq
->wr_coeff
== 1 || bfqd
->wr_busy_queues
> 1 ||
4792 !bfq_bfqq_has_short_ttime(bfqq
)))
4793 bfqq
= bfq_choose_bfqq_for_injection(bfqd
);
4800 reason
= BFQQE_NO_MORE_REQUESTS
;
4802 bfq_bfqq_expire(bfqd
, bfqq
, false, reason
);
4804 bfqq
= bfq_set_in_service_queue(bfqd
);
4806 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: checking new queue");
4811 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: returned this queue");
4813 bfq_log(bfqd
, "select_queue: no queue returned");
4818 static void bfq_update_wr_data(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
4820 struct bfq_entity
*entity
= &bfqq
->entity
;
4822 if (bfqq
->wr_coeff
> 1) { /* queue is being weight-raised */
4823 bfq_log_bfqq(bfqd
, bfqq
,
4824 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
4825 jiffies_to_msecs(jiffies
- bfqq
->last_wr_start_finish
),
4826 jiffies_to_msecs(bfqq
->wr_cur_max_time
),
4828 bfqq
->entity
.weight
, bfqq
->entity
.orig_weight
);
4830 if (entity
->prio_changed
)
4831 bfq_log_bfqq(bfqd
, bfqq
, "WARN: pending prio change");
4834 * If the queue was activated in a burst, or too much
4835 * time has elapsed from the beginning of this
4836 * weight-raising period, then end weight raising.
4838 if (bfq_bfqq_in_large_burst(bfqq
))
4839 bfq_bfqq_end_wr(bfqq
);
4840 else if (time_is_before_jiffies(bfqq
->last_wr_start_finish
+
4841 bfqq
->wr_cur_max_time
)) {
4842 if (bfqq
->wr_cur_max_time
!= bfqd
->bfq_wr_rt_max_time
||
4843 time_is_before_jiffies(bfqq
->wr_start_at_switch_to_srt
+
4844 bfq_wr_duration(bfqd
))) {
4846 * Either in interactive weight
4847 * raising, or in soft_rt weight
4849 * interactive-weight-raising period
4850 * elapsed (so no switch back to
4851 * interactive weight raising).
4853 bfq_bfqq_end_wr(bfqq
);
4855 * soft_rt finishing while still in
4856 * interactive period, switch back to
4857 * interactive weight raising
4859 switch_back_to_interactive_wr(bfqq
, bfqd
);
4860 bfqq
->entity
.prio_changed
= 1;
4863 if (bfqq
->wr_coeff
> 1 &&
4864 bfqq
->wr_cur_max_time
!= bfqd
->bfq_wr_rt_max_time
&&
4865 bfqq
->service_from_wr
> max_service_from_wr
) {
4866 /* see comments on max_service_from_wr */
4867 bfq_bfqq_end_wr(bfqq
);
4871 * To improve latency (for this or other queues), immediately
4872 * update weight both if it must be raised and if it must be
4873 * lowered. Since, entity may be on some active tree here, and
4874 * might have a pending change of its ioprio class, invoke
4875 * next function with the last parameter unset (see the
4876 * comments on the function).
4878 if ((entity
->weight
> entity
->orig_weight
) != (bfqq
->wr_coeff
> 1))
4879 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity
),
4884 * Dispatch next request from bfqq.
4886 static struct request
*bfq_dispatch_rq_from_bfqq(struct bfq_data
*bfqd
,
4887 struct bfq_queue
*bfqq
)
4889 struct request
*rq
= bfqq
->next_rq
;
4890 unsigned long service_to_charge
;
4892 service_to_charge
= bfq_serv_to_charge(rq
, bfqq
);
4894 bfq_bfqq_served(bfqq
, service_to_charge
);
4896 if (bfqq
== bfqd
->in_service_queue
&& bfqd
->wait_dispatch
) {
4897 bfqd
->wait_dispatch
= false;
4898 bfqd
->waited_rq
= rq
;
4901 bfq_dispatch_remove(bfqd
->queue
, rq
);
4903 if (bfqq
!= bfqd
->in_service_queue
)
4907 * If weight raising has to terminate for bfqq, then next
4908 * function causes an immediate update of bfqq's weight,
4909 * without waiting for next activation. As a consequence, on
4910 * expiration, bfqq will be timestamped as if has never been
4911 * weight-raised during this service slot, even if it has
4912 * received part or even most of the service as a
4913 * weight-raised queue. This inflates bfqq's timestamps, which
4914 * is beneficial, as bfqq is then more willing to leave the
4915 * device immediately to possible other weight-raised queues.
4917 bfq_update_wr_data(bfqd
, bfqq
);
4920 * Expire bfqq, pretending that its budget expired, if bfqq
4921 * belongs to CLASS_IDLE and other queues are waiting for
4924 if (!(bfq_tot_busy_queues(bfqd
) > 1 && bfq_class_idle(bfqq
)))
4927 bfq_bfqq_expire(bfqd
, bfqq
, false, BFQQE_BUDGET_EXHAUSTED
);
4933 static bool bfq_has_work(struct blk_mq_hw_ctx
*hctx
)
4935 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
4938 * Avoiding lock: a race on bfqd->busy_queues should cause at
4939 * most a call to dispatch for nothing
4941 return !list_empty_careful(&bfqd
->dispatch
) ||
4942 bfq_tot_busy_queues(bfqd
) > 0;
4945 static struct request
*__bfq_dispatch_request(struct blk_mq_hw_ctx
*hctx
)
4947 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
4948 struct request
*rq
= NULL
;
4949 struct bfq_queue
*bfqq
= NULL
;
4951 if (!list_empty(&bfqd
->dispatch
)) {
4952 rq
= list_first_entry(&bfqd
->dispatch
, struct request
,
4954 list_del_init(&rq
->queuelist
);
4960 * Increment counters here, because this
4961 * dispatch does not follow the standard
4962 * dispatch flow (where counters are
4967 goto inc_in_driver_start_rq
;
4971 * We exploit the bfq_finish_requeue_request hook to
4972 * decrement rq_in_driver, but
4973 * bfq_finish_requeue_request will not be invoked on
4974 * this request. So, to avoid unbalance, just start
4975 * this request, without incrementing rq_in_driver. As
4976 * a negative consequence, rq_in_driver is deceptively
4977 * lower than it should be while this request is in
4978 * service. This may cause bfq_schedule_dispatch to be
4979 * invoked uselessly.
4981 * As for implementing an exact solution, the
4982 * bfq_finish_requeue_request hook, if defined, is
4983 * probably invoked also on this request. So, by
4984 * exploiting this hook, we could 1) increment
4985 * rq_in_driver here, and 2) decrement it in
4986 * bfq_finish_requeue_request. Such a solution would
4987 * let the value of the counter be always accurate,
4988 * but it would entail using an extra interface
4989 * function. This cost seems higher than the benefit,
4990 * being the frequency of non-elevator-private
4991 * requests very low.
4996 bfq_log(bfqd
, "dispatch requests: %d busy queues",
4997 bfq_tot_busy_queues(bfqd
));
4999 if (bfq_tot_busy_queues(bfqd
) == 0)
5003 * Force device to serve one request at a time if
5004 * strict_guarantees is true. Forcing this service scheme is
5005 * currently the ONLY way to guarantee that the request
5006 * service order enforced by the scheduler is respected by a
5007 * queueing device. Otherwise the device is free even to make
5008 * some unlucky request wait for as long as the device
5011 * Of course, serving one request at a time may cause loss of
5014 if (bfqd
->strict_guarantees
&& bfqd
->rq_in_driver
> 0)
5017 bfqq
= bfq_select_queue(bfqd
);
5021 rq
= bfq_dispatch_rq_from_bfqq(bfqd
, bfqq
);
5024 inc_in_driver_start_rq
:
5025 bfqd
->rq_in_driver
++;
5027 rq
->rq_flags
|= RQF_STARTED
;
5033 #ifdef CONFIG_BFQ_CGROUP_DEBUG
5034 static void bfq_update_dispatch_stats(struct request_queue
*q
,
5036 struct bfq_queue
*in_serv_queue
,
5037 bool idle_timer_disabled
)
5039 struct bfq_queue
*bfqq
= rq
? RQ_BFQQ(rq
) : NULL
;
5041 if (!idle_timer_disabled
&& !bfqq
)
5045 * rq and bfqq are guaranteed to exist until this function
5046 * ends, for the following reasons. First, rq can be
5047 * dispatched to the device, and then can be completed and
5048 * freed, only after this function ends. Second, rq cannot be
5049 * merged (and thus freed because of a merge) any longer,
5050 * because it has already started. Thus rq cannot be freed
5051 * before this function ends, and, since rq has a reference to
5052 * bfqq, the same guarantee holds for bfqq too.
5054 * In addition, the following queue lock guarantees that
5055 * bfqq_group(bfqq) exists as well.
5057 spin_lock_irq(&q
->queue_lock
);
5058 if (idle_timer_disabled
)
5060 * Since the idle timer has been disabled,
5061 * in_serv_queue contained some request when
5062 * __bfq_dispatch_request was invoked above, which
5063 * implies that rq was picked exactly from
5064 * in_serv_queue. Thus in_serv_queue == bfqq, and is
5065 * therefore guaranteed to exist because of the above
5068 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue
));
5070 struct bfq_group
*bfqg
= bfqq_group(bfqq
);
5072 bfqg_stats_update_avg_queue_size(bfqg
);
5073 bfqg_stats_set_start_empty_time(bfqg
);
5074 bfqg_stats_update_io_remove(bfqg
, rq
->cmd_flags
);
5076 spin_unlock_irq(&q
->queue_lock
);
5079 static inline void bfq_update_dispatch_stats(struct request_queue
*q
,
5081 struct bfq_queue
*in_serv_queue
,
5082 bool idle_timer_disabled
) {}
5083 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
5085 static struct request
*bfq_dispatch_request(struct blk_mq_hw_ctx
*hctx
)
5087 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
5089 struct bfq_queue
*in_serv_queue
;
5090 bool waiting_rq
, idle_timer_disabled
= false;
5092 spin_lock_irq(&bfqd
->lock
);
5094 in_serv_queue
= bfqd
->in_service_queue
;
5095 waiting_rq
= in_serv_queue
&& bfq_bfqq_wait_request(in_serv_queue
);
5097 rq
= __bfq_dispatch_request(hctx
);
5098 if (in_serv_queue
== bfqd
->in_service_queue
) {
5099 idle_timer_disabled
=
5100 waiting_rq
&& !bfq_bfqq_wait_request(in_serv_queue
);
5103 spin_unlock_irq(&bfqd
->lock
);
5104 bfq_update_dispatch_stats(hctx
->queue
, rq
,
5105 idle_timer_disabled
? in_serv_queue
: NULL
,
5106 idle_timer_disabled
);
5112 * Task holds one reference to the queue, dropped when task exits. Each rq
5113 * in-flight on this queue also holds a reference, dropped when rq is freed.
5115 * Scheduler lock must be held here. Recall not to use bfqq after calling
5116 * this function on it.
5118 void bfq_put_queue(struct bfq_queue
*bfqq
)
5120 struct bfq_queue
*item
;
5121 struct hlist_node
*n
;
5122 struct bfq_group
*bfqg
= bfqq_group(bfqq
);
5125 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "put_queue: %p %d",
5132 if (!hlist_unhashed(&bfqq
->burst_list_node
)) {
5133 hlist_del_init(&bfqq
->burst_list_node
);
5135 * Decrement also burst size after the removal, if the
5136 * process associated with bfqq is exiting, and thus
5137 * does not contribute to the burst any longer. This
5138 * decrement helps filter out false positives of large
5139 * bursts, when some short-lived process (often due to
5140 * the execution of commands by some service) happens
5141 * to start and exit while a complex application is
5142 * starting, and thus spawning several processes that
5143 * do I/O (and that *must not* be treated as a large
5144 * burst, see comments on bfq_handle_burst).
5146 * In particular, the decrement is performed only if:
5147 * 1) bfqq is not a merged queue, because, if it is,
5148 * then this free of bfqq is not triggered by the exit
5149 * of the process bfqq is associated with, but exactly
5150 * by the fact that bfqq has just been merged.
5151 * 2) burst_size is greater than 0, to handle
5152 * unbalanced decrements. Unbalanced decrements may
5153 * happen in te following case: bfqq is inserted into
5154 * the current burst list--without incrementing
5155 * bust_size--because of a split, but the current
5156 * burst list is not the burst list bfqq belonged to
5157 * (see comments on the case of a split in
5160 if (bfqq
->bic
&& bfqq
->bfqd
->burst_size
> 0)
5161 bfqq
->bfqd
->burst_size
--;
5165 * bfqq does not exist any longer, so it cannot be woken by
5166 * any other queue, and cannot wake any other queue. Then bfqq
5167 * must be removed from the woken list of its possible waker
5168 * queue, and all queues in the woken list of bfqq must stop
5169 * having a waker queue. Strictly speaking, these updates
5170 * should be performed when bfqq remains with no I/O source
5171 * attached to it, which happens before bfqq gets freed. In
5172 * particular, this happens when the last process associated
5173 * with bfqq exits or gets associated with a different
5174 * queue. However, both events lead to bfqq being freed soon,
5175 * and dangling references would come out only after bfqq gets
5176 * freed. So these updates are done here, as a simple and safe
5177 * way to handle all cases.
5179 /* remove bfqq from woken list */
5180 if (!hlist_unhashed(&bfqq
->woken_list_node
))
5181 hlist_del_init(&bfqq
->woken_list_node
);
5183 /* reset waker for all queues in woken list */
5184 hlist_for_each_entry_safe(item
, n
, &bfqq
->woken_list
,
5186 item
->waker_bfqq
= NULL
;
5187 hlist_del_init(&item
->woken_list_node
);
5190 if (bfqq
->bfqd
&& bfqq
->bfqd
->last_completed_rq_bfqq
== bfqq
)
5191 bfqq
->bfqd
->last_completed_rq_bfqq
= NULL
;
5193 kmem_cache_free(bfq_pool
, bfqq
);
5194 bfqg_and_blkg_put(bfqg
);
5197 static void bfq_put_stable_ref(struct bfq_queue
*bfqq
)
5200 bfq_put_queue(bfqq
);
5203 void bfq_put_cooperator(struct bfq_queue
*bfqq
)
5205 struct bfq_queue
*__bfqq
, *next
;
5208 * If this queue was scheduled to merge with another queue, be
5209 * sure to drop the reference taken on that queue (and others in
5210 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
5212 __bfqq
= bfqq
->new_bfqq
;
5216 next
= __bfqq
->new_bfqq
;
5217 bfq_put_queue(__bfqq
);
5222 static void bfq_exit_bfqq(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
5224 if (bfqq
== bfqd
->in_service_queue
) {
5225 __bfq_bfqq_expire(bfqd
, bfqq
, BFQQE_BUDGET_TIMEOUT
);
5226 bfq_schedule_dispatch(bfqd
);
5229 bfq_log_bfqq(bfqd
, bfqq
, "exit_bfqq: %p, %d", bfqq
, bfqq
->ref
);
5231 bfq_put_cooperator(bfqq
);
5233 bfq_release_process_ref(bfqd
, bfqq
);
5236 static void bfq_exit_icq_bfqq(struct bfq_io_cq
*bic
, bool is_sync
)
5238 struct bfq_queue
*bfqq
= bic_to_bfqq(bic
, is_sync
);
5239 struct bfq_data
*bfqd
;
5242 bfqd
= bfqq
->bfqd
; /* NULL if scheduler already exited */
5245 unsigned long flags
;
5247 spin_lock_irqsave(&bfqd
->lock
, flags
);
5249 bfq_exit_bfqq(bfqd
, bfqq
);
5250 bic_set_bfqq(bic
, NULL
, is_sync
);
5251 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
5255 static void bfq_exit_icq(struct io_cq
*icq
)
5257 struct bfq_io_cq
*bic
= icq_to_bic(icq
);
5259 if (bic
->stable_merge_bfqq
) {
5260 struct bfq_data
*bfqd
= bic
->stable_merge_bfqq
->bfqd
;
5263 * bfqd is NULL if scheduler already exited, and in
5264 * that case this is the last time bfqq is accessed.
5267 unsigned long flags
;
5269 spin_lock_irqsave(&bfqd
->lock
, flags
);
5270 bfq_put_stable_ref(bic
->stable_merge_bfqq
);
5271 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
5273 bfq_put_stable_ref(bic
->stable_merge_bfqq
);
5277 bfq_exit_icq_bfqq(bic
, true);
5278 bfq_exit_icq_bfqq(bic
, false);
5282 * Update the entity prio values; note that the new values will not
5283 * be used until the next (re)activation.
5286 bfq_set_next_ioprio_data(struct bfq_queue
*bfqq
, struct bfq_io_cq
*bic
)
5288 struct task_struct
*tsk
= current
;
5290 struct bfq_data
*bfqd
= bfqq
->bfqd
;
5295 ioprio_class
= IOPRIO_PRIO_CLASS(bic
->ioprio
);
5296 switch (ioprio_class
) {
5298 pr_err("bdi %s: bfq: bad prio class %d\n",
5299 bdi_dev_name(bfqq
->bfqd
->queue
->disk
->bdi
),
5302 case IOPRIO_CLASS_NONE
:
5304 * No prio set, inherit CPU scheduling settings.
5306 bfqq
->new_ioprio
= task_nice_ioprio(tsk
);
5307 bfqq
->new_ioprio_class
= task_nice_ioclass(tsk
);
5309 case IOPRIO_CLASS_RT
:
5310 bfqq
->new_ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
5311 bfqq
->new_ioprio_class
= IOPRIO_CLASS_RT
;
5313 case IOPRIO_CLASS_BE
:
5314 bfqq
->new_ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
5315 bfqq
->new_ioprio_class
= IOPRIO_CLASS_BE
;
5317 case IOPRIO_CLASS_IDLE
:
5318 bfqq
->new_ioprio_class
= IOPRIO_CLASS_IDLE
;
5319 bfqq
->new_ioprio
= 7;
5323 if (bfqq
->new_ioprio
>= IOPRIO_NR_LEVELS
) {
5324 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
5326 bfqq
->new_ioprio
= IOPRIO_NR_LEVELS
- 1;
5329 bfqq
->entity
.new_weight
= bfq_ioprio_to_weight(bfqq
->new_ioprio
);
5330 bfq_log_bfqq(bfqd
, bfqq
, "new_ioprio %d new_weight %d",
5331 bfqq
->new_ioprio
, bfqq
->entity
.new_weight
);
5332 bfqq
->entity
.prio_changed
= 1;
5335 static struct bfq_queue
*bfq_get_queue(struct bfq_data
*bfqd
,
5336 struct bio
*bio
, bool is_sync
,
5337 struct bfq_io_cq
*bic
,
5340 static void bfq_check_ioprio_change(struct bfq_io_cq
*bic
, struct bio
*bio
)
5342 struct bfq_data
*bfqd
= bic_to_bfqd(bic
);
5343 struct bfq_queue
*bfqq
;
5344 int ioprio
= bic
->icq
.ioc
->ioprio
;
5347 * This condition may trigger on a newly created bic, be sure to
5348 * drop the lock before returning.
5350 if (unlikely(!bfqd
) || likely(bic
->ioprio
== ioprio
))
5353 bic
->ioprio
= ioprio
;
5355 bfqq
= bic_to_bfqq(bic
, false);
5357 bfq_release_process_ref(bfqd
, bfqq
);
5358 bfqq
= bfq_get_queue(bfqd
, bio
, BLK_RW_ASYNC
, bic
, true);
5359 bic_set_bfqq(bic
, bfqq
, false);
5362 bfqq
= bic_to_bfqq(bic
, true);
5364 bfq_set_next_ioprio_data(bfqq
, bic
);
5367 static void bfq_init_bfqq(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
5368 struct bfq_io_cq
*bic
, pid_t pid
, int is_sync
)
5370 u64 now_ns
= ktime_get_ns();
5372 RB_CLEAR_NODE(&bfqq
->entity
.rb_node
);
5373 INIT_LIST_HEAD(&bfqq
->fifo
);
5374 INIT_HLIST_NODE(&bfqq
->burst_list_node
);
5375 INIT_HLIST_NODE(&bfqq
->woken_list_node
);
5376 INIT_HLIST_HEAD(&bfqq
->woken_list
);
5382 bfq_set_next_ioprio_data(bfqq
, bic
);
5386 * No need to mark as has_short_ttime if in
5387 * idle_class, because no device idling is performed
5388 * for queues in idle class
5390 if (!bfq_class_idle(bfqq
))
5391 /* tentatively mark as has_short_ttime */
5392 bfq_mark_bfqq_has_short_ttime(bfqq
);
5393 bfq_mark_bfqq_sync(bfqq
);
5394 bfq_mark_bfqq_just_created(bfqq
);
5396 bfq_clear_bfqq_sync(bfqq
);
5398 /* set end request to minus infinity from now */
5399 bfqq
->ttime
.last_end_request
= now_ns
+ 1;
5401 bfqq
->creation_time
= jiffies
;
5403 bfqq
->io_start_time
= now_ns
;
5405 bfq_mark_bfqq_IO_bound(bfqq
);
5409 /* Tentative initial value to trade off between thr and lat */
5410 bfqq
->max_budget
= (2 * bfq_max_budget(bfqd
)) / 3;
5411 bfqq
->budget_timeout
= bfq_smallest_from_now();
5414 bfqq
->last_wr_start_finish
= jiffies
;
5415 bfqq
->wr_start_at_switch_to_srt
= bfq_smallest_from_now();
5416 bfqq
->split_time
= bfq_smallest_from_now();
5419 * To not forget the possibly high bandwidth consumed by a
5420 * process/queue in the recent past,
5421 * bfq_bfqq_softrt_next_start() returns a value at least equal
5422 * to the current value of bfqq->soft_rt_next_start (see
5423 * comments on bfq_bfqq_softrt_next_start). Set
5424 * soft_rt_next_start to now, to mean that bfqq has consumed
5425 * no bandwidth so far.
5427 bfqq
->soft_rt_next_start
= jiffies
;
5429 /* first request is almost certainly seeky */
5430 bfqq
->seek_history
= 1;
5433 static struct bfq_queue
**bfq_async_queue_prio(struct bfq_data
*bfqd
,
5434 struct bfq_group
*bfqg
,
5435 int ioprio_class
, int ioprio
)
5437 switch (ioprio_class
) {
5438 case IOPRIO_CLASS_RT
:
5439 return &bfqg
->async_bfqq
[0][ioprio
];
5440 case IOPRIO_CLASS_NONE
:
5441 ioprio
= IOPRIO_BE_NORM
;
5443 case IOPRIO_CLASS_BE
:
5444 return &bfqg
->async_bfqq
[1][ioprio
];
5445 case IOPRIO_CLASS_IDLE
:
5446 return &bfqg
->async_idle_bfqq
;
5452 static struct bfq_queue
*
5453 bfq_do_early_stable_merge(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
5454 struct bfq_io_cq
*bic
,
5455 struct bfq_queue
*last_bfqq_created
)
5457 struct bfq_queue
*new_bfqq
=
5458 bfq_setup_merge(bfqq
, last_bfqq_created
);
5464 new_bfqq
->bic
->stably_merged
= true;
5465 bic
->stably_merged
= true;
5468 * Reusing merge functions. This implies that
5469 * bfqq->bic must be set too, for
5470 * bfq_merge_bfqqs to correctly save bfqq's
5471 * state before killing it.
5474 bfq_merge_bfqqs(bfqd
, bic
, bfqq
, new_bfqq
);
5480 * Many throughput-sensitive workloads are made of several parallel
5481 * I/O flows, with all flows generated by the same application, or
5482 * more generically by the same task (e.g., system boot). The most
5483 * counterproductive action with these workloads is plugging I/O
5484 * dispatch when one of the bfq_queues associated with these flows
5485 * remains temporarily empty.
5487 * To avoid this plugging, BFQ has been using a burst-handling
5488 * mechanism for years now. This mechanism has proven effective for
5489 * throughput, and not detrimental for service guarantees. The
5490 * following function pushes this mechanism a little bit further,
5491 * basing on the following two facts.
5493 * First, all the I/O flows of a the same application or task
5494 * contribute to the execution/completion of that common application
5495 * or task. So the performance figures that matter are total
5496 * throughput of the flows and task-wide I/O latency. In particular,
5497 * these flows do not need to be protected from each other, in terms
5498 * of individual bandwidth or latency.
5500 * Second, the above fact holds regardless of the number of flows.
5502 * Putting these two facts together, this commits merges stably the
5503 * bfq_queues associated with these I/O flows, i.e., with the
5504 * processes that generate these IO/ flows, regardless of how many the
5505 * involved processes are.
5507 * To decide whether a set of bfq_queues is actually associated with
5508 * the I/O flows of a common application or task, and to merge these
5509 * queues stably, this function operates as follows: given a bfq_queue,
5510 * say Q2, currently being created, and the last bfq_queue, say Q1,
5511 * created before Q2, Q2 is merged stably with Q1 if
5512 * - very little time has elapsed since when Q1 was created
5513 * - Q2 has the same ioprio as Q1
5514 * - Q2 belongs to the same group as Q1
5516 * Merging bfq_queues also reduces scheduling overhead. A fio test
5517 * with ten random readers on /dev/nullb shows a throughput boost of
5518 * 40%, with a quadcore. Since BFQ's execution time amounts to ~50% of
5519 * the total per-request processing time, the above throughput boost
5520 * implies that BFQ's overhead is reduced by more than 50%.
5522 * This new mechanism most certainly obsoletes the current
5523 * burst-handling heuristics. We keep those heuristics for the moment.
5525 static struct bfq_queue
*bfq_do_or_sched_stable_merge(struct bfq_data
*bfqd
,
5526 struct bfq_queue
*bfqq
,
5527 struct bfq_io_cq
*bic
)
5529 struct bfq_queue
**source_bfqq
= bfqq
->entity
.parent
?
5530 &bfqq
->entity
.parent
->last_bfqq_created
:
5531 &bfqd
->last_bfqq_created
;
5533 struct bfq_queue
*last_bfqq_created
= *source_bfqq
;
5536 * If last_bfqq_created has not been set yet, then init it. If
5537 * it has been set already, but too long ago, then move it
5538 * forward to bfqq. Finally, move also if bfqq belongs to a
5539 * different group than last_bfqq_created, or if bfqq has a
5540 * different ioprio or ioprio_class. If none of these
5541 * conditions holds true, then try an early stable merge or
5542 * schedule a delayed stable merge.
5544 * A delayed merge is scheduled (instead of performing an
5545 * early merge), in case bfqq might soon prove to be more
5546 * throughput-beneficial if not merged. Currently this is
5547 * possible only if bfqd is rotational with no queueing. For
5548 * such a drive, not merging bfqq is better for throughput if
5549 * bfqq happens to contain sequential I/O. So, we wait a
5550 * little bit for enough I/O to flow through bfqq. After that,
5551 * if such an I/O is sequential, then the merge is
5552 * canceled. Otherwise the merge is finally performed.
5554 if (!last_bfqq_created
||
5555 time_before(last_bfqq_created
->creation_time
+
5556 msecs_to_jiffies(bfq_activation_stable_merging
),
5557 bfqq
->creation_time
) ||
5558 bfqq
->entity
.parent
!= last_bfqq_created
->entity
.parent
||
5559 bfqq
->ioprio
!= last_bfqq_created
->ioprio
||
5560 bfqq
->ioprio_class
!= last_bfqq_created
->ioprio_class
)
5561 *source_bfqq
= bfqq
;
5562 else if (time_after_eq(last_bfqq_created
->creation_time
+
5563 bfqd
->bfq_burst_interval
,
5564 bfqq
->creation_time
)) {
5565 if (likely(bfqd
->nonrot_with_queueing
))
5567 * With this type of drive, leaving
5568 * bfqq alone may provide no
5569 * throughput benefits compared with
5570 * merging bfqq. So merge bfqq now.
5572 bfqq
= bfq_do_early_stable_merge(bfqd
, bfqq
,
5575 else { /* schedule tentative stable merge */
5577 * get reference on last_bfqq_created,
5578 * to prevent it from being freed,
5579 * until we decide whether to merge
5581 last_bfqq_created
->ref
++;
5583 * need to keep track of stable refs, to
5584 * compute process refs correctly
5586 last_bfqq_created
->stable_ref
++;
5588 * Record the bfqq to merge to.
5590 bic
->stable_merge_bfqq
= last_bfqq_created
;
5598 static struct bfq_queue
*bfq_get_queue(struct bfq_data
*bfqd
,
5599 struct bio
*bio
, bool is_sync
,
5600 struct bfq_io_cq
*bic
,
5603 const int ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
5604 const int ioprio_class
= IOPRIO_PRIO_CLASS(bic
->ioprio
);
5605 struct bfq_queue
**async_bfqq
= NULL
;
5606 struct bfq_queue
*bfqq
;
5607 struct bfq_group
*bfqg
;
5609 bfqg
= bfq_bio_bfqg(bfqd
, bio
);
5611 async_bfqq
= bfq_async_queue_prio(bfqd
, bfqg
, ioprio_class
,
5618 bfqq
= kmem_cache_alloc_node(bfq_pool
,
5619 GFP_NOWAIT
| __GFP_ZERO
| __GFP_NOWARN
,
5623 bfq_init_bfqq(bfqd
, bfqq
, bic
, current
->pid
,
5625 bfq_init_entity(&bfqq
->entity
, bfqg
);
5626 bfq_log_bfqq(bfqd
, bfqq
, "allocated");
5628 bfqq
= &bfqd
->oom_bfqq
;
5629 bfq_log_bfqq(bfqd
, bfqq
, "using oom bfqq");
5634 * Pin the queue now that it's allocated, scheduler exit will
5639 * Extra group reference, w.r.t. sync
5640 * queue. This extra reference is removed
5641 * only if bfqq->bfqg disappears, to
5642 * guarantee that this queue is not freed
5643 * until its group goes away.
5645 bfq_log_bfqq(bfqd
, bfqq
, "get_queue, bfqq not in async: %p, %d",
5651 bfqq
->ref
++; /* get a process reference to this queue */
5653 if (bfqq
!= &bfqd
->oom_bfqq
&& is_sync
&& !respawn
)
5654 bfqq
= bfq_do_or_sched_stable_merge(bfqd
, bfqq
, bic
);
5658 static void bfq_update_io_thinktime(struct bfq_data
*bfqd
,
5659 struct bfq_queue
*bfqq
)
5661 struct bfq_ttime
*ttime
= &bfqq
->ttime
;
5665 * We are really interested in how long it takes for the queue to
5666 * become busy when there is no outstanding IO for this queue. So
5667 * ignore cases when the bfq queue has already IO queued.
5669 if (bfqq
->dispatched
|| bfq_bfqq_busy(bfqq
))
5671 elapsed
= ktime_get_ns() - bfqq
->ttime
.last_end_request
;
5672 elapsed
= min_t(u64
, elapsed
, 2ULL * bfqd
->bfq_slice_idle
);
5674 ttime
->ttime_samples
= (7*ttime
->ttime_samples
+ 256) / 8;
5675 ttime
->ttime_total
= div_u64(7*ttime
->ttime_total
+ 256*elapsed
, 8);
5676 ttime
->ttime_mean
= div64_ul(ttime
->ttime_total
+ 128,
5677 ttime
->ttime_samples
);
5681 bfq_update_io_seektime(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
5684 bfqq
->seek_history
<<= 1;
5685 bfqq
->seek_history
|= BFQ_RQ_SEEKY(bfqd
, bfqq
->last_request_pos
, rq
);
5687 if (bfqq
->wr_coeff
> 1 &&
5688 bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
&&
5689 BFQQ_TOTALLY_SEEKY(bfqq
)) {
5690 if (time_is_before_jiffies(bfqq
->wr_start_at_switch_to_srt
+
5691 bfq_wr_duration(bfqd
))) {
5693 * In soft_rt weight raising with the
5694 * interactive-weight-raising period
5695 * elapsed (so no switch back to
5696 * interactive weight raising).
5698 bfq_bfqq_end_wr(bfqq
);
5700 * stopping soft_rt weight raising
5701 * while still in interactive period,
5702 * switch back to interactive weight
5705 switch_back_to_interactive_wr(bfqq
, bfqd
);
5706 bfqq
->entity
.prio_changed
= 1;
5711 static void bfq_update_has_short_ttime(struct bfq_data
*bfqd
,
5712 struct bfq_queue
*bfqq
,
5713 struct bfq_io_cq
*bic
)
5715 bool has_short_ttime
= true, state_changed
;
5718 * No need to update has_short_ttime if bfqq is async or in
5719 * idle io prio class, or if bfq_slice_idle is zero, because
5720 * no device idling is performed for bfqq in this case.
5722 if (!bfq_bfqq_sync(bfqq
) || bfq_class_idle(bfqq
) ||
5723 bfqd
->bfq_slice_idle
== 0)
5726 /* Idle window just restored, statistics are meaningless. */
5727 if (time_is_after_eq_jiffies(bfqq
->split_time
+
5728 bfqd
->bfq_wr_min_idle_time
))
5731 /* Think time is infinite if no process is linked to
5732 * bfqq. Otherwise check average think time to decide whether
5733 * to mark as has_short_ttime. To this goal, compare average
5734 * think time with half the I/O-plugging timeout.
5736 if (atomic_read(&bic
->icq
.ioc
->active_ref
) == 0 ||
5737 (bfq_sample_valid(bfqq
->ttime
.ttime_samples
) &&
5738 bfqq
->ttime
.ttime_mean
> bfqd
->bfq_slice_idle
>>1))
5739 has_short_ttime
= false;
5741 state_changed
= has_short_ttime
!= bfq_bfqq_has_short_ttime(bfqq
);
5743 if (has_short_ttime
)
5744 bfq_mark_bfqq_has_short_ttime(bfqq
);
5746 bfq_clear_bfqq_has_short_ttime(bfqq
);
5749 * Until the base value for the total service time gets
5750 * finally computed for bfqq, the inject limit does depend on
5751 * the think-time state (short|long). In particular, the limit
5752 * is 0 or 1 if the think time is deemed, respectively, as
5753 * short or long (details in the comments in
5754 * bfq_update_inject_limit()). Accordingly, the next
5755 * instructions reset the inject limit if the think-time state
5756 * has changed and the above base value is still to be
5759 * However, the reset is performed only if more than 100 ms
5760 * have elapsed since the last update of the inject limit, or
5761 * (inclusive) if the change is from short to long think
5762 * time. The reason for this waiting is as follows.
5764 * bfqq may have a long think time because of a
5765 * synchronization with some other queue, i.e., because the
5766 * I/O of some other queue may need to be completed for bfqq
5767 * to receive new I/O. Details in the comments on the choice
5768 * of the queue for injection in bfq_select_queue().
5770 * As stressed in those comments, if such a synchronization is
5771 * actually in place, then, without injection on bfqq, the
5772 * blocking I/O cannot happen to served while bfqq is in
5773 * service. As a consequence, if bfqq is granted
5774 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
5775 * is dispatched, until the idle timeout fires. This is likely
5776 * to result in lower bandwidth and higher latencies for bfqq,
5777 * and in a severe loss of total throughput.
5779 * On the opposite end, a non-zero inject limit may allow the
5780 * I/O that blocks bfqq to be executed soon, and therefore
5781 * bfqq to receive new I/O soon.
5783 * But, if the blocking gets actually eliminated, then the
5784 * next think-time sample for bfqq may be very low. This in
5785 * turn may cause bfqq's think time to be deemed
5786 * short. Without the 100 ms barrier, this new state change
5787 * would cause the body of the next if to be executed
5788 * immediately. But this would set to 0 the inject
5789 * limit. Without injection, the blocking I/O would cause the
5790 * think time of bfqq to become long again, and therefore the
5791 * inject limit to be raised again, and so on. The only effect
5792 * of such a steady oscillation between the two think-time
5793 * states would be to prevent effective injection on bfqq.
5795 * In contrast, if the inject limit is not reset during such a
5796 * long time interval as 100 ms, then the number of short
5797 * think time samples can grow significantly before the reset
5798 * is performed. As a consequence, the think time state can
5799 * become stable before the reset. Therefore there will be no
5800 * state change when the 100 ms elapse, and no reset of the
5801 * inject limit. The inject limit remains steadily equal to 1
5802 * both during and after the 100 ms. So injection can be
5803 * performed at all times, and throughput gets boosted.
5805 * An inject limit equal to 1 is however in conflict, in
5806 * general, with the fact that the think time of bfqq is
5807 * short, because injection may be likely to delay bfqq's I/O
5808 * (as explained in the comments in
5809 * bfq_update_inject_limit()). But this does not happen in
5810 * this special case, because bfqq's low think time is due to
5811 * an effective handling of a synchronization, through
5812 * injection. In this special case, bfqq's I/O does not get
5813 * delayed by injection; on the contrary, bfqq's I/O is
5814 * brought forward, because it is not blocked for
5817 * In addition, serving the blocking I/O much sooner, and much
5818 * more frequently than once per I/O-plugging timeout, makes
5819 * it much quicker to detect a waker queue (the concept of
5820 * waker queue is defined in the comments in
5821 * bfq_add_request()). This makes it possible to start sooner
5822 * to boost throughput more effectively, by injecting the I/O
5823 * of the waker queue unconditionally on every
5824 * bfq_dispatch_request().
5826 * One last, important benefit of not resetting the inject
5827 * limit before 100 ms is that, during this time interval, the
5828 * base value for the total service time is likely to get
5829 * finally computed for bfqq, freeing the inject limit from
5830 * its relation with the think time.
5832 if (state_changed
&& bfqq
->last_serv_time_ns
== 0 &&
5833 (time_is_before_eq_jiffies(bfqq
->decrease_time_jif
+
5834 msecs_to_jiffies(100)) ||
5836 bfq_reset_inject_limit(bfqd
, bfqq
);
5840 * Called when a new fs request (rq) is added to bfqq. Check if there's
5841 * something we should do about it.
5843 static void bfq_rq_enqueued(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
5846 if (rq
->cmd_flags
& REQ_META
)
5847 bfqq
->meta_pending
++;
5849 bfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
5851 if (bfqq
== bfqd
->in_service_queue
&& bfq_bfqq_wait_request(bfqq
)) {
5852 bool small_req
= bfqq
->queued
[rq_is_sync(rq
)] == 1 &&
5853 blk_rq_sectors(rq
) < 32;
5854 bool budget_timeout
= bfq_bfqq_budget_timeout(bfqq
);
5857 * There is just this request queued: if
5858 * - the request is small, and
5859 * - we are idling to boost throughput, and
5860 * - the queue is not to be expired,
5863 * In this way, if the device is being idled to wait
5864 * for a new request from the in-service queue, we
5865 * avoid unplugging the device and committing the
5866 * device to serve just a small request. In contrast
5867 * we wait for the block layer to decide when to
5868 * unplug the device: hopefully, new requests will be
5869 * merged to this one quickly, then the device will be
5870 * unplugged and larger requests will be dispatched.
5872 if (small_req
&& idling_boosts_thr_without_issues(bfqd
, bfqq
) &&
5877 * A large enough request arrived, or idling is being
5878 * performed to preserve service guarantees, or
5879 * finally the queue is to be expired: in all these
5880 * cases disk idling is to be stopped, so clear
5881 * wait_request flag and reset timer.
5883 bfq_clear_bfqq_wait_request(bfqq
);
5884 hrtimer_try_to_cancel(&bfqd
->idle_slice_timer
);
5887 * The queue is not empty, because a new request just
5888 * arrived. Hence we can safely expire the queue, in
5889 * case of budget timeout, without risking that the
5890 * timestamps of the queue are not updated correctly.
5891 * See [1] for more details.
5894 bfq_bfqq_expire(bfqd
, bfqq
, false,
5895 BFQQE_BUDGET_TIMEOUT
);
5899 /* returns true if it causes the idle timer to be disabled */
5900 static bool __bfq_insert_request(struct bfq_data
*bfqd
, struct request
*rq
)
5902 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
),
5903 *new_bfqq
= bfq_setup_cooperator(bfqd
, bfqq
, rq
, true,
5905 bool waiting
, idle_timer_disabled
= false;
5909 * Release the request's reference to the old bfqq
5910 * and make sure one is taken to the shared queue.
5912 new_bfqq
->allocated
++;
5916 * If the bic associated with the process
5917 * issuing this request still points to bfqq
5918 * (and thus has not been already redirected
5919 * to new_bfqq or even some other bfq_queue),
5920 * then complete the merge and redirect it to
5923 if (bic_to_bfqq(RQ_BIC(rq
), 1) == bfqq
)
5924 bfq_merge_bfqqs(bfqd
, RQ_BIC(rq
),
5927 bfq_clear_bfqq_just_created(bfqq
);
5929 * rq is about to be enqueued into new_bfqq,
5930 * release rq reference on bfqq
5932 bfq_put_queue(bfqq
);
5933 rq
->elv
.priv
[1] = new_bfqq
;
5937 bfq_update_io_thinktime(bfqd
, bfqq
);
5938 bfq_update_has_short_ttime(bfqd
, bfqq
, RQ_BIC(rq
));
5939 bfq_update_io_seektime(bfqd
, bfqq
, rq
);
5941 waiting
= bfqq
&& bfq_bfqq_wait_request(bfqq
);
5942 bfq_add_request(rq
);
5943 idle_timer_disabled
= waiting
&& !bfq_bfqq_wait_request(bfqq
);
5945 rq
->fifo_time
= ktime_get_ns() + bfqd
->bfq_fifo_expire
[rq_is_sync(rq
)];
5946 list_add_tail(&rq
->queuelist
, &bfqq
->fifo
);
5948 bfq_rq_enqueued(bfqd
, bfqq
, rq
);
5950 return idle_timer_disabled
;
5953 #ifdef CONFIG_BFQ_CGROUP_DEBUG
5954 static void bfq_update_insert_stats(struct request_queue
*q
,
5955 struct bfq_queue
*bfqq
,
5956 bool idle_timer_disabled
,
5957 unsigned int cmd_flags
)
5963 * bfqq still exists, because it can disappear only after
5964 * either it is merged with another queue, or the process it
5965 * is associated with exits. But both actions must be taken by
5966 * the same process currently executing this flow of
5969 * In addition, the following queue lock guarantees that
5970 * bfqq_group(bfqq) exists as well.
5972 spin_lock_irq(&q
->queue_lock
);
5973 bfqg_stats_update_io_add(bfqq_group(bfqq
), bfqq
, cmd_flags
);
5974 if (idle_timer_disabled
)
5975 bfqg_stats_update_idle_time(bfqq_group(bfqq
));
5976 spin_unlock_irq(&q
->queue_lock
);
5979 static inline void bfq_update_insert_stats(struct request_queue
*q
,
5980 struct bfq_queue
*bfqq
,
5981 bool idle_timer_disabled
,
5982 unsigned int cmd_flags
) {}
5983 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
5985 static struct bfq_queue
*bfq_init_rq(struct request
*rq
);
5987 static void bfq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
5990 struct request_queue
*q
= hctx
->queue
;
5991 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
5992 struct bfq_queue
*bfqq
;
5993 bool idle_timer_disabled
= false;
5994 unsigned int cmd_flags
;
5997 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5998 if (!cgroup_subsys_on_dfl(io_cgrp_subsys
) && rq
->bio
)
5999 bfqg_stats_update_legacy_io(q
, rq
);
6001 spin_lock_irq(&bfqd
->lock
);
6002 bfqq
= bfq_init_rq(rq
);
6003 if (blk_mq_sched_try_insert_merge(q
, rq
, &free
)) {
6004 spin_unlock_irq(&bfqd
->lock
);
6005 blk_mq_free_requests(&free
);
6009 trace_block_rq_insert(rq
);
6011 if (!bfqq
|| at_head
) {
6013 list_add(&rq
->queuelist
, &bfqd
->dispatch
);
6015 list_add_tail(&rq
->queuelist
, &bfqd
->dispatch
);
6017 idle_timer_disabled
= __bfq_insert_request(bfqd
, rq
);
6019 * Update bfqq, because, if a queue merge has occurred
6020 * in __bfq_insert_request, then rq has been
6021 * redirected into a new queue.
6025 if (rq_mergeable(rq
)) {
6026 elv_rqhash_add(q
, rq
);
6033 * Cache cmd_flags before releasing scheduler lock, because rq
6034 * may disappear afterwards (for example, because of a request
6037 cmd_flags
= rq
->cmd_flags
;
6038 spin_unlock_irq(&bfqd
->lock
);
6040 bfq_update_insert_stats(q
, bfqq
, idle_timer_disabled
,
6044 static void bfq_insert_requests(struct blk_mq_hw_ctx
*hctx
,
6045 struct list_head
*list
, bool at_head
)
6047 while (!list_empty(list
)) {
6050 rq
= list_first_entry(list
, struct request
, queuelist
);
6051 list_del_init(&rq
->queuelist
);
6052 bfq_insert_request(hctx
, rq
, at_head
);
6056 static void bfq_update_hw_tag(struct bfq_data
*bfqd
)
6058 struct bfq_queue
*bfqq
= bfqd
->in_service_queue
;
6060 bfqd
->max_rq_in_driver
= max_t(int, bfqd
->max_rq_in_driver
,
6061 bfqd
->rq_in_driver
);
6063 if (bfqd
->hw_tag
== 1)
6067 * This sample is valid if the number of outstanding requests
6068 * is large enough to allow a queueing behavior. Note that the
6069 * sum is not exact, as it's not taking into account deactivated
6072 if (bfqd
->rq_in_driver
+ bfqd
->queued
<= BFQ_HW_QUEUE_THRESHOLD
)
6076 * If active queue hasn't enough requests and can idle, bfq might not
6077 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
6080 if (bfqq
&& bfq_bfqq_has_short_ttime(bfqq
) &&
6081 bfqq
->dispatched
+ bfqq
->queued
[0] + bfqq
->queued
[1] <
6082 BFQ_HW_QUEUE_THRESHOLD
&&
6083 bfqd
->rq_in_driver
< BFQ_HW_QUEUE_THRESHOLD
)
6086 if (bfqd
->hw_tag_samples
++ < BFQ_HW_QUEUE_SAMPLES
)
6089 bfqd
->hw_tag
= bfqd
->max_rq_in_driver
> BFQ_HW_QUEUE_THRESHOLD
;
6090 bfqd
->max_rq_in_driver
= 0;
6091 bfqd
->hw_tag_samples
= 0;
6093 bfqd
->nonrot_with_queueing
=
6094 blk_queue_nonrot(bfqd
->queue
) && bfqd
->hw_tag
;
6097 static void bfq_completed_request(struct bfq_queue
*bfqq
, struct bfq_data
*bfqd
)
6102 bfq_update_hw_tag(bfqd
);
6104 bfqd
->rq_in_driver
--;
6107 if (!bfqq
->dispatched
&& !bfq_bfqq_busy(bfqq
)) {
6109 * Set budget_timeout (which we overload to store the
6110 * time at which the queue remains with no backlog and
6111 * no outstanding request; used by the weight-raising
6114 bfqq
->budget_timeout
= jiffies
;
6116 bfq_weights_tree_remove(bfqd
, bfqq
);
6119 now_ns
= ktime_get_ns();
6121 bfqq
->ttime
.last_end_request
= now_ns
;
6124 * Using us instead of ns, to get a reasonable precision in
6125 * computing rate in next check.
6127 delta_us
= div_u64(now_ns
- bfqd
->last_completion
, NSEC_PER_USEC
);
6130 * If the request took rather long to complete, and, according
6131 * to the maximum request size recorded, this completion latency
6132 * implies that the request was certainly served at a very low
6133 * rate (less than 1M sectors/sec), then the whole observation
6134 * interval that lasts up to this time instant cannot be a
6135 * valid time interval for computing a new peak rate. Invoke
6136 * bfq_update_rate_reset to have the following three steps
6138 * - close the observation interval at the last (previous)
6139 * request dispatch or completion
6140 * - compute rate, if possible, for that observation interval
6141 * - reset to zero samples, which will trigger a proper
6142 * re-initialization of the observation interval on next
6145 if (delta_us
> BFQ_MIN_TT
/NSEC_PER_USEC
&&
6146 (bfqd
->last_rq_max_size
<<BFQ_RATE_SHIFT
)/delta_us
<
6147 1UL<<(BFQ_RATE_SHIFT
- 10))
6148 bfq_update_rate_reset(bfqd
, NULL
);
6149 bfqd
->last_completion
= now_ns
;
6151 * Shared queues are likely to receive I/O at a high
6152 * rate. This may deceptively let them be considered as wakers
6153 * of other queues. But a false waker will unjustly steal
6154 * bandwidth to its supposedly woken queue. So considering
6155 * also shared queues in the waking mechanism may cause more
6156 * control troubles than throughput benefits. Then reset
6157 * last_completed_rq_bfqq if bfqq is a shared queue.
6159 if (!bfq_bfqq_coop(bfqq
))
6160 bfqd
->last_completed_rq_bfqq
= bfqq
;
6162 bfqd
->last_completed_rq_bfqq
= NULL
;
6165 * If we are waiting to discover whether the request pattern
6166 * of the task associated with the queue is actually
6167 * isochronous, and both requisites for this condition to hold
6168 * are now satisfied, then compute soft_rt_next_start (see the
6169 * comments on the function bfq_bfqq_softrt_next_start()). We
6170 * do not compute soft_rt_next_start if bfqq is in interactive
6171 * weight raising (see the comments in bfq_bfqq_expire() for
6172 * an explanation). We schedule this delayed update when bfqq
6173 * expires, if it still has in-flight requests.
6175 if (bfq_bfqq_softrt_update(bfqq
) && bfqq
->dispatched
== 0 &&
6176 RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
6177 bfqq
->wr_coeff
!= bfqd
->bfq_wr_coeff
)
6178 bfqq
->soft_rt_next_start
=
6179 bfq_bfqq_softrt_next_start(bfqd
, bfqq
);
6182 * If this is the in-service queue, check if it needs to be expired,
6183 * or if we want to idle in case it has no pending requests.
6185 if (bfqd
->in_service_queue
== bfqq
) {
6186 if (bfq_bfqq_must_idle(bfqq
)) {
6187 if (bfqq
->dispatched
== 0)
6188 bfq_arm_slice_timer(bfqd
);
6190 * If we get here, we do not expire bfqq, even
6191 * if bfqq was in budget timeout or had no
6192 * more requests (as controlled in the next
6193 * conditional instructions). The reason for
6194 * not expiring bfqq is as follows.
6196 * Here bfqq->dispatched > 0 holds, but
6197 * bfq_bfqq_must_idle() returned true. This
6198 * implies that, even if no request arrives
6199 * for bfqq before bfqq->dispatched reaches 0,
6200 * bfqq will, however, not be expired on the
6201 * completion event that causes bfqq->dispatch
6202 * to reach zero. In contrast, on this event,
6203 * bfqq will start enjoying device idling
6204 * (I/O-dispatch plugging).
6206 * But, if we expired bfqq here, bfqq would
6207 * not have the chance to enjoy device idling
6208 * when bfqq->dispatched finally reaches
6209 * zero. This would expose bfqq to violation
6210 * of its reserved service guarantees.
6213 } else if (bfq_may_expire_for_budg_timeout(bfqq
))
6214 bfq_bfqq_expire(bfqd
, bfqq
, false,
6215 BFQQE_BUDGET_TIMEOUT
);
6216 else if (RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
6217 (bfqq
->dispatched
== 0 ||
6218 !bfq_better_to_idle(bfqq
)))
6219 bfq_bfqq_expire(bfqd
, bfqq
, false,
6220 BFQQE_NO_MORE_REQUESTS
);
6223 if (!bfqd
->rq_in_driver
)
6224 bfq_schedule_dispatch(bfqd
);
6227 static void bfq_finish_requeue_request_body(struct bfq_queue
*bfqq
)
6231 bfq_put_queue(bfqq
);
6235 * The processes associated with bfqq may happen to generate their
6236 * cumulative I/O at a lower rate than the rate at which the device
6237 * could serve the same I/O. This is rather probable, e.g., if only
6238 * one process is associated with bfqq and the device is an SSD. It
6239 * results in bfqq becoming often empty while in service. In this
6240 * respect, if BFQ is allowed to switch to another queue when bfqq
6241 * remains empty, then the device goes on being fed with I/O requests,
6242 * and the throughput is not affected. In contrast, if BFQ is not
6243 * allowed to switch to another queue---because bfqq is sync and
6244 * I/O-dispatch needs to be plugged while bfqq is temporarily
6245 * empty---then, during the service of bfqq, there will be frequent
6246 * "service holes", i.e., time intervals during which bfqq gets empty
6247 * and the device can only consume the I/O already queued in its
6248 * hardware queues. During service holes, the device may even get to
6249 * remaining idle. In the end, during the service of bfqq, the device
6250 * is driven at a lower speed than the one it can reach with the kind
6251 * of I/O flowing through bfqq.
6253 * To counter this loss of throughput, BFQ implements a "request
6254 * injection mechanism", which tries to fill the above service holes
6255 * with I/O requests taken from other queues. The hard part in this
6256 * mechanism is finding the right amount of I/O to inject, so as to
6257 * both boost throughput and not break bfqq's bandwidth and latency
6258 * guarantees. In this respect, the mechanism maintains a per-queue
6259 * inject limit, computed as below. While bfqq is empty, the injection
6260 * mechanism dispatches extra I/O requests only until the total number
6261 * of I/O requests in flight---i.e., already dispatched but not yet
6262 * completed---remains lower than this limit.
6264 * A first definition comes in handy to introduce the algorithm by
6265 * which the inject limit is computed. We define as first request for
6266 * bfqq, an I/O request for bfqq that arrives while bfqq is in
6267 * service, and causes bfqq to switch from empty to non-empty. The
6268 * algorithm updates the limit as a function of the effect of
6269 * injection on the service times of only the first requests of
6270 * bfqq. The reason for this restriction is that these are the
6271 * requests whose service time is affected most, because they are the
6272 * first to arrive after injection possibly occurred.
6274 * To evaluate the effect of injection, the algorithm measures the
6275 * "total service time" of first requests. We define as total service
6276 * time of an I/O request, the time that elapses since when the
6277 * request is enqueued into bfqq, to when it is completed. This
6278 * quantity allows the whole effect of injection to be measured. It is
6279 * easy to see why. Suppose that some requests of other queues are
6280 * actually injected while bfqq is empty, and that a new request R
6281 * then arrives for bfqq. If the device does start to serve all or
6282 * part of the injected requests during the service hole, then,
6283 * because of this extra service, it may delay the next invocation of
6284 * the dispatch hook of BFQ. Then, even after R gets eventually
6285 * dispatched, the device may delay the actual service of R if it is
6286 * still busy serving the extra requests, or if it decides to serve,
6287 * before R, some extra request still present in its queues. As a
6288 * conclusion, the cumulative extra delay caused by injection can be
6289 * easily evaluated by just comparing the total service time of first
6290 * requests with and without injection.
6292 * The limit-update algorithm works as follows. On the arrival of a
6293 * first request of bfqq, the algorithm measures the total time of the
6294 * request only if one of the three cases below holds, and, for each
6295 * case, it updates the limit as described below:
6297 * (1) If there is no in-flight request. This gives a baseline for the
6298 * total service time of the requests of bfqq. If the baseline has
6299 * not been computed yet, then, after computing it, the limit is
6300 * set to 1, to start boosting throughput, and to prepare the
6301 * ground for the next case. If the baseline has already been
6302 * computed, then it is updated, in case it results to be lower
6303 * than the previous value.
6305 * (2) If the limit is higher than 0 and there are in-flight
6306 * requests. By comparing the total service time in this case with
6307 * the above baseline, it is possible to know at which extent the
6308 * current value of the limit is inflating the total service
6309 * time. If the inflation is below a certain threshold, then bfqq
6310 * is assumed to be suffering from no perceivable loss of its
6311 * service guarantees, and the limit is even tentatively
6312 * increased. If the inflation is above the threshold, then the
6313 * limit is decreased. Due to the lack of any hysteresis, this
6314 * logic makes the limit oscillate even in steady workload
6315 * conditions. Yet we opted for it, because it is fast in reaching
6316 * the best value for the limit, as a function of the current I/O
6317 * workload. To reduce oscillations, this step is disabled for a
6318 * short time interval after the limit happens to be decreased.
6320 * (3) Periodically, after resetting the limit, to make sure that the
6321 * limit eventually drops in case the workload changes. This is
6322 * needed because, after the limit has gone safely up for a
6323 * certain workload, it is impossible to guess whether the
6324 * baseline total service time may have changed, without measuring
6325 * it again without injection. A more effective version of this
6326 * step might be to just sample the baseline, by interrupting
6327 * injection only once, and then to reset/lower the limit only if
6328 * the total service time with the current limit does happen to be
6331 * More details on each step are provided in the comments on the
6332 * pieces of code that implement these steps: the branch handling the
6333 * transition from empty to non empty in bfq_add_request(), the branch
6334 * handling injection in bfq_select_queue(), and the function
6335 * bfq_choose_bfqq_for_injection(). These comments also explain some
6336 * exceptions, made by the injection mechanism in some special cases.
6338 static void bfq_update_inject_limit(struct bfq_data
*bfqd
,
6339 struct bfq_queue
*bfqq
)
6341 u64 tot_time_ns
= ktime_get_ns() - bfqd
->last_empty_occupied_ns
;
6342 unsigned int old_limit
= bfqq
->inject_limit
;
6344 if (bfqq
->last_serv_time_ns
> 0 && bfqd
->rqs_injected
) {
6345 u64 threshold
= (bfqq
->last_serv_time_ns
* 3)>>1;
6347 if (tot_time_ns
>= threshold
&& old_limit
> 0) {
6348 bfqq
->inject_limit
--;
6349 bfqq
->decrease_time_jif
= jiffies
;
6350 } else if (tot_time_ns
< threshold
&&
6351 old_limit
<= bfqd
->max_rq_in_driver
)
6352 bfqq
->inject_limit
++;
6356 * Either we still have to compute the base value for the
6357 * total service time, and there seem to be the right
6358 * conditions to do it, or we can lower the last base value
6361 * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
6362 * request in flight, because this function is in the code
6363 * path that handles the completion of a request of bfqq, and,
6364 * in particular, this function is executed before
6365 * bfqd->rq_in_driver is decremented in such a code path.
6367 if ((bfqq
->last_serv_time_ns
== 0 && bfqd
->rq_in_driver
== 1) ||
6368 tot_time_ns
< bfqq
->last_serv_time_ns
) {
6369 if (bfqq
->last_serv_time_ns
== 0) {
6371 * Now we certainly have a base value: make sure we
6372 * start trying injection.
6374 bfqq
->inject_limit
= max_t(unsigned int, 1, old_limit
);
6376 bfqq
->last_serv_time_ns
= tot_time_ns
;
6377 } else if (!bfqd
->rqs_injected
&& bfqd
->rq_in_driver
== 1)
6379 * No I/O injected and no request still in service in
6380 * the drive: these are the exact conditions for
6381 * computing the base value of the total service time
6382 * for bfqq. So let's update this value, because it is
6383 * rather variable. For example, it varies if the size
6384 * or the spatial locality of the I/O requests in bfqq
6387 bfqq
->last_serv_time_ns
= tot_time_ns
;
6390 /* update complete, not waiting for any request completion any longer */
6391 bfqd
->waited_rq
= NULL
;
6392 bfqd
->rqs_injected
= false;
6396 * Handle either a requeue or a finish for rq. The things to do are
6397 * the same in both cases: all references to rq are to be dropped. In
6398 * particular, rq is considered completed from the point of view of
6401 static void bfq_finish_requeue_request(struct request
*rq
)
6403 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
6404 struct bfq_data
*bfqd
;
6405 unsigned long flags
;
6408 * rq either is not associated with any icq, or is an already
6409 * requeued request that has not (yet) been re-inserted into
6412 if (!rq
->elv
.icq
|| !bfqq
)
6417 if (rq
->rq_flags
& RQF_STARTED
)
6418 bfqg_stats_update_completion(bfqq_group(bfqq
),
6420 rq
->io_start_time_ns
,
6423 spin_lock_irqsave(&bfqd
->lock
, flags
);
6424 if (likely(rq
->rq_flags
& RQF_STARTED
)) {
6425 if (rq
== bfqd
->waited_rq
)
6426 bfq_update_inject_limit(bfqd
, bfqq
);
6428 bfq_completed_request(bfqq
, bfqd
);
6430 bfq_finish_requeue_request_body(bfqq
);
6431 RQ_BIC(rq
)->requests
--;
6432 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
6435 * Reset private fields. In case of a requeue, this allows
6436 * this function to correctly do nothing if it is spuriously
6437 * invoked again on this same request (see the check at the
6438 * beginning of the function). Probably, a better general
6439 * design would be to prevent blk-mq from invoking the requeue
6440 * or finish hooks of an elevator, for a request that is not
6441 * referred by that elevator.
6443 * Resetting the following fields would break the
6444 * request-insertion logic if rq is re-inserted into a bfq
6445 * internal queue, without a re-preparation. Here we assume
6446 * that re-insertions of requeued requests, without
6447 * re-preparation, can happen only for pass_through or at_head
6448 * requests (which are not re-inserted into bfq internal
6451 rq
->elv
.priv
[0] = NULL
;
6452 rq
->elv
.priv
[1] = NULL
;
6456 * Removes the association between the current task and bfqq, assuming
6457 * that bic points to the bfq iocontext of the task.
6458 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
6459 * was the last process referring to that bfqq.
6461 static struct bfq_queue
*
6462 bfq_split_bfqq(struct bfq_io_cq
*bic
, struct bfq_queue
*bfqq
)
6464 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "splitting queue");
6466 if (bfqq_process_refs(bfqq
) == 1) {
6467 bfqq
->pid
= current
->pid
;
6468 bfq_clear_bfqq_coop(bfqq
);
6469 bfq_clear_bfqq_split_coop(bfqq
);
6473 bic_set_bfqq(bic
, NULL
, 1);
6475 bfq_put_cooperator(bfqq
);
6477 bfq_release_process_ref(bfqq
->bfqd
, bfqq
);
6481 static struct bfq_queue
*bfq_get_bfqq_handle_split(struct bfq_data
*bfqd
,
6482 struct bfq_io_cq
*bic
,
6484 bool split
, bool is_sync
,
6487 struct bfq_queue
*bfqq
= bic_to_bfqq(bic
, is_sync
);
6489 if (likely(bfqq
&& bfqq
!= &bfqd
->oom_bfqq
))
6496 bfq_put_queue(bfqq
);
6497 bfqq
= bfq_get_queue(bfqd
, bio
, is_sync
, bic
, split
);
6499 bic_set_bfqq(bic
, bfqq
, is_sync
);
6500 if (split
&& is_sync
) {
6501 if ((bic
->was_in_burst_list
&& bfqd
->large_burst
) ||
6502 bic
->saved_in_large_burst
)
6503 bfq_mark_bfqq_in_large_burst(bfqq
);
6505 bfq_clear_bfqq_in_large_burst(bfqq
);
6506 if (bic
->was_in_burst_list
)
6508 * If bfqq was in the current
6509 * burst list before being
6510 * merged, then we have to add
6511 * it back. And we do not need
6512 * to increase burst_size, as
6513 * we did not decrement
6514 * burst_size when we removed
6515 * bfqq from the burst list as
6516 * a consequence of a merge
6518 * bfq_put_queue). In this
6519 * respect, it would be rather
6520 * costly to know whether the
6521 * current burst list is still
6522 * the same burst list from
6523 * which bfqq was removed on
6524 * the merge. To avoid this
6525 * cost, if bfqq was in a
6526 * burst list, then we add
6527 * bfqq to the current burst
6528 * list without any further
6529 * check. This can cause
6530 * inappropriate insertions,
6531 * but rarely enough to not
6532 * harm the detection of large
6533 * bursts significantly.
6535 hlist_add_head(&bfqq
->burst_list_node
,
6538 bfqq
->split_time
= jiffies
;
6545 * Only reset private fields. The actual request preparation will be
6546 * performed by bfq_init_rq, when rq is either inserted or merged. See
6547 * comments on bfq_init_rq for the reason behind this delayed
6550 static void bfq_prepare_request(struct request
*rq
)
6553 * Regardless of whether we have an icq attached, we have to
6554 * clear the scheduler pointers, as they might point to
6555 * previously allocated bic/bfqq structs.
6557 rq
->elv
.priv
[0] = rq
->elv
.priv
[1] = NULL
;
6561 * If needed, init rq, allocate bfq data structures associated with
6562 * rq, and increment reference counters in the destination bfq_queue
6563 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6564 * not associated with any bfq_queue.
6566 * This function is invoked by the functions that perform rq insertion
6567 * or merging. One may have expected the above preparation operations
6568 * to be performed in bfq_prepare_request, and not delayed to when rq
6569 * is inserted or merged. The rationale behind this delayed
6570 * preparation is that, after the prepare_request hook is invoked for
6571 * rq, rq may still be transformed into a request with no icq, i.e., a
6572 * request not associated with any queue. No bfq hook is invoked to
6573 * signal this transformation. As a consequence, should these
6574 * preparation operations be performed when the prepare_request hook
6575 * is invoked, and should rq be transformed one moment later, bfq
6576 * would end up in an inconsistent state, because it would have
6577 * incremented some queue counters for an rq destined to
6578 * transformation, without any chance to correctly lower these
6579 * counters back. In contrast, no transformation can still happen for
6580 * rq after rq has been inserted or merged. So, it is safe to execute
6581 * these preparation operations when rq is finally inserted or merged.
6583 static struct bfq_queue
*bfq_init_rq(struct request
*rq
)
6585 struct request_queue
*q
= rq
->q
;
6586 struct bio
*bio
= rq
->bio
;
6587 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
6588 struct bfq_io_cq
*bic
;
6589 const int is_sync
= rq_is_sync(rq
);
6590 struct bfq_queue
*bfqq
;
6591 bool new_queue
= false;
6592 bool bfqq_already_existing
= false, split
= false;
6594 if (unlikely(!rq
->elv
.icq
))
6598 * Assuming that elv.priv[1] is set only if everything is set
6599 * for this rq. This holds true, because this function is
6600 * invoked only for insertion or merging, and, after such
6601 * events, a request cannot be manipulated any longer before
6602 * being removed from bfq.
6604 if (rq
->elv
.priv
[1])
6605 return rq
->elv
.priv
[1];
6607 bic
= icq_to_bic(rq
->elv
.icq
);
6609 bfq_check_ioprio_change(bic
, bio
);
6611 bfq_bic_update_cgroup(bic
, bio
);
6613 bfqq
= bfq_get_bfqq_handle_split(bfqd
, bic
, bio
, false, is_sync
,
6616 if (likely(!new_queue
)) {
6617 /* If the queue was seeky for too long, break it apart. */
6618 if (bfq_bfqq_coop(bfqq
) && bfq_bfqq_split_coop(bfqq
) &&
6619 !bic
->stably_merged
) {
6620 struct bfq_queue
*old_bfqq
= bfqq
;
6622 /* Update bic before losing reference to bfqq */
6623 if (bfq_bfqq_in_large_burst(bfqq
))
6624 bic
->saved_in_large_burst
= true;
6626 bfqq
= bfq_split_bfqq(bic
, bfqq
);
6630 bfqq
= bfq_get_bfqq_handle_split(bfqd
, bic
, bio
,
6633 bfqq
->waker_bfqq
= old_bfqq
->waker_bfqq
;
6634 bfqq
->tentative_waker_bfqq
= NULL
;
6637 * If the waker queue disappears, then
6638 * new_bfqq->waker_bfqq must be
6639 * reset. So insert new_bfqq into the
6640 * woken_list of the waker. See
6641 * bfq_check_waker for details.
6643 if (bfqq
->waker_bfqq
)
6644 hlist_add_head(&bfqq
->woken_list_node
,
6645 &bfqq
->waker_bfqq
->woken_list
);
6647 bfqq_already_existing
= true;
6654 bfq_log_bfqq(bfqd
, bfqq
, "get_request %p: bfqq %p, %d",
6655 rq
, bfqq
, bfqq
->ref
);
6657 rq
->elv
.priv
[0] = bic
;
6658 rq
->elv
.priv
[1] = bfqq
;
6661 * If a bfq_queue has only one process reference, it is owned
6662 * by only this bic: we can then set bfqq->bic = bic. in
6663 * addition, if the queue has also just been split, we have to
6666 if (likely(bfqq
!= &bfqd
->oom_bfqq
) && bfqq_process_refs(bfqq
) == 1) {
6670 * The queue has just been split from a shared
6671 * queue: restore the idle window and the
6672 * possible weight raising period.
6674 bfq_bfqq_resume_state(bfqq
, bfqd
, bic
,
6675 bfqq_already_existing
);
6680 * Consider bfqq as possibly belonging to a burst of newly
6681 * created queues only if:
6682 * 1) A burst is actually happening (bfqd->burst_size > 0)
6684 * 2) There is no other active queue. In fact, if, in
6685 * contrast, there are active queues not belonging to the
6686 * possible burst bfqq may belong to, then there is no gain
6687 * in considering bfqq as belonging to a burst, and
6688 * therefore in not weight-raising bfqq. See comments on
6689 * bfq_handle_burst().
6691 * This filtering also helps eliminating false positives,
6692 * occurring when bfqq does not belong to an actual large
6693 * burst, but some background task (e.g., a service) happens
6694 * to trigger the creation of new queues very close to when
6695 * bfqq and its possible companion queues are created. See
6696 * comments on bfq_handle_burst() for further details also on
6699 if (unlikely(bfq_bfqq_just_created(bfqq
) &&
6700 (bfqd
->burst_size
> 0 ||
6701 bfq_tot_busy_queues(bfqd
) == 0)))
6702 bfq_handle_burst(bfqd
, bfqq
);
6708 bfq_idle_slice_timer_body(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
6710 enum bfqq_expiration reason
;
6711 unsigned long flags
;
6713 spin_lock_irqsave(&bfqd
->lock
, flags
);
6716 * Considering that bfqq may be in race, we should firstly check
6717 * whether bfqq is in service before doing something on it. If
6718 * the bfqq in race is not in service, it has already been expired
6719 * through __bfq_bfqq_expire func and its wait_request flags has
6720 * been cleared in __bfq_bfqd_reset_in_service func.
6722 if (bfqq
!= bfqd
->in_service_queue
) {
6723 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
6727 bfq_clear_bfqq_wait_request(bfqq
);
6729 if (bfq_bfqq_budget_timeout(bfqq
))
6731 * Also here the queue can be safely expired
6732 * for budget timeout without wasting
6735 reason
= BFQQE_BUDGET_TIMEOUT
;
6736 else if (bfqq
->queued
[0] == 0 && bfqq
->queued
[1] == 0)
6738 * The queue may not be empty upon timer expiration,
6739 * because we may not disable the timer when the
6740 * first request of the in-service queue arrives
6741 * during disk idling.
6743 reason
= BFQQE_TOO_IDLE
;
6745 goto schedule_dispatch
;
6747 bfq_bfqq_expire(bfqd
, bfqq
, true, reason
);
6750 bfq_schedule_dispatch(bfqd
);
6751 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
6755 * Handler of the expiration of the timer running if the in-service queue
6756 * is idling inside its time slice.
6758 static enum hrtimer_restart
bfq_idle_slice_timer(struct hrtimer
*timer
)
6760 struct bfq_data
*bfqd
= container_of(timer
, struct bfq_data
,
6762 struct bfq_queue
*bfqq
= bfqd
->in_service_queue
;
6765 * Theoretical race here: the in-service queue can be NULL or
6766 * different from the queue that was idling if a new request
6767 * arrives for the current queue and there is a full dispatch
6768 * cycle that changes the in-service queue. This can hardly
6769 * happen, but in the worst case we just expire a queue too
6773 bfq_idle_slice_timer_body(bfqd
, bfqq
);
6775 return HRTIMER_NORESTART
;
6778 static void __bfq_put_async_bfqq(struct bfq_data
*bfqd
,
6779 struct bfq_queue
**bfqq_ptr
)
6781 struct bfq_queue
*bfqq
= *bfqq_ptr
;
6783 bfq_log(bfqd
, "put_async_bfqq: %p", bfqq
);
6785 bfq_bfqq_move(bfqd
, bfqq
, bfqd
->root_group
);
6787 bfq_log_bfqq(bfqd
, bfqq
, "put_async_bfqq: putting %p, %d",
6789 bfq_put_queue(bfqq
);
6795 * Release all the bfqg references to its async queues. If we are
6796 * deallocating the group these queues may still contain requests, so
6797 * we reparent them to the root cgroup (i.e., the only one that will
6798 * exist for sure until all the requests on a device are gone).
6800 void bfq_put_async_queues(struct bfq_data
*bfqd
, struct bfq_group
*bfqg
)
6804 for (i
= 0; i
< 2; i
++)
6805 for (j
= 0; j
< IOPRIO_NR_LEVELS
; j
++)
6806 __bfq_put_async_bfqq(bfqd
, &bfqg
->async_bfqq
[i
][j
]);
6808 __bfq_put_async_bfqq(bfqd
, &bfqg
->async_idle_bfqq
);
6812 * See the comments on bfq_limit_depth for the purpose of
6813 * the depths set in the function. Return minimum shallow depth we'll use.
6815 static unsigned int bfq_update_depths(struct bfq_data
*bfqd
,
6816 struct sbitmap_queue
*bt
)
6818 unsigned int i
, j
, min_shallow
= UINT_MAX
;
6821 * In-word depths if no bfq_queue is being weight-raised:
6822 * leaving 25% of tags only for sync reads.
6824 * In next formulas, right-shift the value
6825 * (1U<<bt->sb.shift), instead of computing directly
6826 * (1U<<(bt->sb.shift - something)), to be robust against
6827 * any possible value of bt->sb.shift, without having to
6828 * limit 'something'.
6830 /* no more than 50% of tags for async I/O */
6831 bfqd
->word_depths
[0][0] = max((1U << bt
->sb
.shift
) >> 1, 1U);
6833 * no more than 75% of tags for sync writes (25% extra tags
6834 * w.r.t. async I/O, to prevent async I/O from starving sync
6837 bfqd
->word_depths
[0][1] = max(((1U << bt
->sb
.shift
) * 3) >> 2, 1U);
6840 * In-word depths in case some bfq_queue is being weight-
6841 * raised: leaving ~63% of tags for sync reads. This is the
6842 * highest percentage for which, in our tests, application
6843 * start-up times didn't suffer from any regression due to tag
6846 /* no more than ~18% of tags for async I/O */
6847 bfqd
->word_depths
[1][0] = max(((1U << bt
->sb
.shift
) * 3) >> 4, 1U);
6848 /* no more than ~37% of tags for sync writes (~20% extra tags) */
6849 bfqd
->word_depths
[1][1] = max(((1U << bt
->sb
.shift
) * 6) >> 4, 1U);
6851 for (i
= 0; i
< 2; i
++)
6852 for (j
= 0; j
< 2; j
++)
6853 min_shallow
= min(min_shallow
, bfqd
->word_depths
[i
][j
]);
6858 static void bfq_depth_updated(struct blk_mq_hw_ctx
*hctx
)
6860 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
6861 struct blk_mq_tags
*tags
= hctx
->sched_tags
;
6862 unsigned int min_shallow
;
6864 min_shallow
= bfq_update_depths(bfqd
, tags
->bitmap_tags
);
6865 sbitmap_queue_min_shallow_depth(tags
->bitmap_tags
, min_shallow
);
6868 static int bfq_init_hctx(struct blk_mq_hw_ctx
*hctx
, unsigned int index
)
6870 bfq_depth_updated(hctx
);
6874 static void bfq_exit_queue(struct elevator_queue
*e
)
6876 struct bfq_data
*bfqd
= e
->elevator_data
;
6877 struct bfq_queue
*bfqq
, *n
;
6879 hrtimer_cancel(&bfqd
->idle_slice_timer
);
6881 spin_lock_irq(&bfqd
->lock
);
6882 list_for_each_entry_safe(bfqq
, n
, &bfqd
->idle_list
, bfqq_list
)
6883 bfq_deactivate_bfqq(bfqd
, bfqq
, false, false);
6884 spin_unlock_irq(&bfqd
->lock
);
6886 hrtimer_cancel(&bfqd
->idle_slice_timer
);
6888 /* release oom-queue reference to root group */
6889 bfqg_and_blkg_put(bfqd
->root_group
);
6891 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6892 blkcg_deactivate_policy(bfqd
->queue
, &blkcg_policy_bfq
);
6894 spin_lock_irq(&bfqd
->lock
);
6895 bfq_put_async_queues(bfqd
, bfqd
->root_group
);
6896 kfree(bfqd
->root_group
);
6897 spin_unlock_irq(&bfqd
->lock
);
6900 wbt_enable_default(bfqd
->queue
);
6905 static void bfq_init_root_group(struct bfq_group
*root_group
,
6906 struct bfq_data
*bfqd
)
6910 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6911 root_group
->entity
.parent
= NULL
;
6912 root_group
->my_entity
= NULL
;
6913 root_group
->bfqd
= bfqd
;
6915 root_group
->rq_pos_tree
= RB_ROOT
;
6916 for (i
= 0; i
< BFQ_IOPRIO_CLASSES
; i
++)
6917 root_group
->sched_data
.service_tree
[i
] = BFQ_SERVICE_TREE_INIT
;
6918 root_group
->sched_data
.bfq_class_idle_last_service
= jiffies
;
6921 static int bfq_init_queue(struct request_queue
*q
, struct elevator_type
*e
)
6923 struct bfq_data
*bfqd
;
6924 struct elevator_queue
*eq
;
6926 eq
= elevator_alloc(q
, e
);
6930 bfqd
= kzalloc_node(sizeof(*bfqd
), GFP_KERNEL
, q
->node
);
6932 kobject_put(&eq
->kobj
);
6935 eq
->elevator_data
= bfqd
;
6937 spin_lock_irq(&q
->queue_lock
);
6939 spin_unlock_irq(&q
->queue_lock
);
6942 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
6943 * Grab a permanent reference to it, so that the normal code flow
6944 * will not attempt to free it.
6946 bfq_init_bfqq(bfqd
, &bfqd
->oom_bfqq
, NULL
, 1, 0);
6947 bfqd
->oom_bfqq
.ref
++;
6948 bfqd
->oom_bfqq
.new_ioprio
= BFQ_DEFAULT_QUEUE_IOPRIO
;
6949 bfqd
->oom_bfqq
.new_ioprio_class
= IOPRIO_CLASS_BE
;
6950 bfqd
->oom_bfqq
.entity
.new_weight
=
6951 bfq_ioprio_to_weight(bfqd
->oom_bfqq
.new_ioprio
);
6953 /* oom_bfqq does not participate to bursts */
6954 bfq_clear_bfqq_just_created(&bfqd
->oom_bfqq
);
6957 * Trigger weight initialization, according to ioprio, at the
6958 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
6959 * class won't be changed any more.
6961 bfqd
->oom_bfqq
.entity
.prio_changed
= 1;
6965 INIT_LIST_HEAD(&bfqd
->dispatch
);
6967 hrtimer_init(&bfqd
->idle_slice_timer
, CLOCK_MONOTONIC
,
6969 bfqd
->idle_slice_timer
.function
= bfq_idle_slice_timer
;
6971 bfqd
->queue_weights_tree
= RB_ROOT_CACHED
;
6972 bfqd
->num_groups_with_pending_reqs
= 0;
6974 INIT_LIST_HEAD(&bfqd
->active_list
);
6975 INIT_LIST_HEAD(&bfqd
->idle_list
);
6976 INIT_HLIST_HEAD(&bfqd
->burst_list
);
6979 bfqd
->nonrot_with_queueing
= blk_queue_nonrot(bfqd
->queue
);
6981 bfqd
->bfq_max_budget
= bfq_default_max_budget
;
6983 bfqd
->bfq_fifo_expire
[0] = bfq_fifo_expire
[0];
6984 bfqd
->bfq_fifo_expire
[1] = bfq_fifo_expire
[1];
6985 bfqd
->bfq_back_max
= bfq_back_max
;
6986 bfqd
->bfq_back_penalty
= bfq_back_penalty
;
6987 bfqd
->bfq_slice_idle
= bfq_slice_idle
;
6988 bfqd
->bfq_timeout
= bfq_timeout
;
6990 bfqd
->bfq_large_burst_thresh
= 8;
6991 bfqd
->bfq_burst_interval
= msecs_to_jiffies(180);
6993 bfqd
->low_latency
= true;
6996 * Trade-off between responsiveness and fairness.
6998 bfqd
->bfq_wr_coeff
= 30;
6999 bfqd
->bfq_wr_rt_max_time
= msecs_to_jiffies(300);
7000 bfqd
->bfq_wr_max_time
= 0;
7001 bfqd
->bfq_wr_min_idle_time
= msecs_to_jiffies(2000);
7002 bfqd
->bfq_wr_min_inter_arr_async
= msecs_to_jiffies(500);
7003 bfqd
->bfq_wr_max_softrt_rate
= 7000; /*
7004 * Approximate rate required
7005 * to playback or record a
7006 * high-definition compressed
7009 bfqd
->wr_busy_queues
= 0;
7012 * Begin by assuming, optimistically, that the device peak
7013 * rate is equal to 2/3 of the highest reference rate.
7015 bfqd
->rate_dur_prod
= ref_rate
[blk_queue_nonrot(bfqd
->queue
)] *
7016 ref_wr_duration
[blk_queue_nonrot(bfqd
->queue
)];
7017 bfqd
->peak_rate
= ref_rate
[blk_queue_nonrot(bfqd
->queue
)] * 2 / 3;
7019 spin_lock_init(&bfqd
->lock
);
7022 * The invocation of the next bfq_create_group_hierarchy
7023 * function is the head of a chain of function calls
7024 * (bfq_create_group_hierarchy->blkcg_activate_policy->
7025 * blk_mq_freeze_queue) that may lead to the invocation of the
7026 * has_work hook function. For this reason,
7027 * bfq_create_group_hierarchy is invoked only after all
7028 * scheduler data has been initialized, apart from the fields
7029 * that can be initialized only after invoking
7030 * bfq_create_group_hierarchy. This, in particular, enables
7031 * has_work to correctly return false. Of course, to avoid
7032 * other inconsistencies, the blk-mq stack must then refrain
7033 * from invoking further scheduler hooks before this init
7034 * function is finished.
7036 bfqd
->root_group
= bfq_create_group_hierarchy(bfqd
, q
->node
);
7037 if (!bfqd
->root_group
)
7039 bfq_init_root_group(bfqd
->root_group
, bfqd
);
7040 bfq_init_entity(&bfqd
->oom_bfqq
.entity
, bfqd
->root_group
);
7042 wbt_disable_default(q
);
7047 kobject_put(&eq
->kobj
);
7051 static void bfq_slab_kill(void)
7053 kmem_cache_destroy(bfq_pool
);
7056 static int __init
bfq_slab_setup(void)
7058 bfq_pool
= KMEM_CACHE(bfq_queue
, 0);
7064 static ssize_t
bfq_var_show(unsigned int var
, char *page
)
7066 return sprintf(page
, "%u\n", var
);
7069 static int bfq_var_store(unsigned long *var
, const char *page
)
7071 unsigned long new_val
;
7072 int ret
= kstrtoul(page
, 10, &new_val
);
7080 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
7081 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7083 struct bfq_data *bfqd = e->elevator_data; \
7084 u64 __data = __VAR; \
7086 __data = jiffies_to_msecs(__data); \
7087 else if (__CONV == 2) \
7088 __data = div_u64(__data, NSEC_PER_MSEC); \
7089 return bfq_var_show(__data, (page)); \
7091 SHOW_FUNCTION(bfq_fifo_expire_sync_show
, bfqd
->bfq_fifo_expire
[1], 2);
7092 SHOW_FUNCTION(bfq_fifo_expire_async_show
, bfqd
->bfq_fifo_expire
[0], 2);
7093 SHOW_FUNCTION(bfq_back_seek_max_show
, bfqd
->bfq_back_max
, 0);
7094 SHOW_FUNCTION(bfq_back_seek_penalty_show
, bfqd
->bfq_back_penalty
, 0);
7095 SHOW_FUNCTION(bfq_slice_idle_show
, bfqd
->bfq_slice_idle
, 2);
7096 SHOW_FUNCTION(bfq_max_budget_show
, bfqd
->bfq_user_max_budget
, 0);
7097 SHOW_FUNCTION(bfq_timeout_sync_show
, bfqd
->bfq_timeout
, 1);
7098 SHOW_FUNCTION(bfq_strict_guarantees_show
, bfqd
->strict_guarantees
, 0);
7099 SHOW_FUNCTION(bfq_low_latency_show
, bfqd
->low_latency
, 0);
7100 #undef SHOW_FUNCTION
7102 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
7103 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7105 struct bfq_data *bfqd = e->elevator_data; \
7106 u64 __data = __VAR; \
7107 __data = div_u64(__data, NSEC_PER_USEC); \
7108 return bfq_var_show(__data, (page)); \
7110 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show
, bfqd
->bfq_slice_idle
);
7111 #undef USEC_SHOW_FUNCTION
7113 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
7115 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
7117 struct bfq_data *bfqd = e->elevator_data; \
7118 unsigned long __data, __min = (MIN), __max = (MAX); \
7121 ret = bfq_var_store(&__data, (page)); \
7124 if (__data < __min) \
7126 else if (__data > __max) \
7129 *(__PTR) = msecs_to_jiffies(__data); \
7130 else if (__CONV == 2) \
7131 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
7133 *(__PTR) = __data; \
7136 STORE_FUNCTION(bfq_fifo_expire_sync_store
, &bfqd
->bfq_fifo_expire
[1], 1,
7138 STORE_FUNCTION(bfq_fifo_expire_async_store
, &bfqd
->bfq_fifo_expire
[0], 1,
7140 STORE_FUNCTION(bfq_back_seek_max_store
, &bfqd
->bfq_back_max
, 0, INT_MAX
, 0);
7141 STORE_FUNCTION(bfq_back_seek_penalty_store
, &bfqd
->bfq_back_penalty
, 1,
7143 STORE_FUNCTION(bfq_slice_idle_store
, &bfqd
->bfq_slice_idle
, 0, INT_MAX
, 2);
7144 #undef STORE_FUNCTION
7146 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
7147 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
7149 struct bfq_data *bfqd = e->elevator_data; \
7150 unsigned long __data, __min = (MIN), __max = (MAX); \
7153 ret = bfq_var_store(&__data, (page)); \
7156 if (__data < __min) \
7158 else if (__data > __max) \
7160 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
7163 USEC_STORE_FUNCTION(bfq_slice_idle_us_store
, &bfqd
->bfq_slice_idle
, 0,
7165 #undef USEC_STORE_FUNCTION
7167 static ssize_t
bfq_max_budget_store(struct elevator_queue
*e
,
7168 const char *page
, size_t count
)
7170 struct bfq_data
*bfqd
= e
->elevator_data
;
7171 unsigned long __data
;
7174 ret
= bfq_var_store(&__data
, (page
));
7179 bfqd
->bfq_max_budget
= bfq_calc_max_budget(bfqd
);
7181 if (__data
> INT_MAX
)
7183 bfqd
->bfq_max_budget
= __data
;
7186 bfqd
->bfq_user_max_budget
= __data
;
7192 * Leaving this name to preserve name compatibility with cfq
7193 * parameters, but this timeout is used for both sync and async.
7195 static ssize_t
bfq_timeout_sync_store(struct elevator_queue
*e
,
7196 const char *page
, size_t count
)
7198 struct bfq_data
*bfqd
= e
->elevator_data
;
7199 unsigned long __data
;
7202 ret
= bfq_var_store(&__data
, (page
));
7208 else if (__data
> INT_MAX
)
7211 bfqd
->bfq_timeout
= msecs_to_jiffies(__data
);
7212 if (bfqd
->bfq_user_max_budget
== 0)
7213 bfqd
->bfq_max_budget
= bfq_calc_max_budget(bfqd
);
7218 static ssize_t
bfq_strict_guarantees_store(struct elevator_queue
*e
,
7219 const char *page
, size_t count
)
7221 struct bfq_data
*bfqd
= e
->elevator_data
;
7222 unsigned long __data
;
7225 ret
= bfq_var_store(&__data
, (page
));
7231 if (!bfqd
->strict_guarantees
&& __data
== 1
7232 && bfqd
->bfq_slice_idle
< 8 * NSEC_PER_MSEC
)
7233 bfqd
->bfq_slice_idle
= 8 * NSEC_PER_MSEC
;
7235 bfqd
->strict_guarantees
= __data
;
7240 static ssize_t
bfq_low_latency_store(struct elevator_queue
*e
,
7241 const char *page
, size_t count
)
7243 struct bfq_data
*bfqd
= e
->elevator_data
;
7244 unsigned long __data
;
7247 ret
= bfq_var_store(&__data
, (page
));
7253 if (__data
== 0 && bfqd
->low_latency
!= 0)
7255 bfqd
->low_latency
= __data
;
7260 #define BFQ_ATTR(name) \
7261 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
7263 static struct elv_fs_entry bfq_attrs
[] = {
7264 BFQ_ATTR(fifo_expire_sync
),
7265 BFQ_ATTR(fifo_expire_async
),
7266 BFQ_ATTR(back_seek_max
),
7267 BFQ_ATTR(back_seek_penalty
),
7268 BFQ_ATTR(slice_idle
),
7269 BFQ_ATTR(slice_idle_us
),
7270 BFQ_ATTR(max_budget
),
7271 BFQ_ATTR(timeout_sync
),
7272 BFQ_ATTR(strict_guarantees
),
7273 BFQ_ATTR(low_latency
),
7277 static struct elevator_type iosched_bfq_mq
= {
7279 .limit_depth
= bfq_limit_depth
,
7280 .prepare_request
= bfq_prepare_request
,
7281 .requeue_request
= bfq_finish_requeue_request
,
7282 .finish_request
= bfq_finish_requeue_request
,
7283 .exit_icq
= bfq_exit_icq
,
7284 .insert_requests
= bfq_insert_requests
,
7285 .dispatch_request
= bfq_dispatch_request
,
7286 .next_request
= elv_rb_latter_request
,
7287 .former_request
= elv_rb_former_request
,
7288 .allow_merge
= bfq_allow_bio_merge
,
7289 .bio_merge
= bfq_bio_merge
,
7290 .request_merge
= bfq_request_merge
,
7291 .requests_merged
= bfq_requests_merged
,
7292 .request_merged
= bfq_request_merged
,
7293 .has_work
= bfq_has_work
,
7294 .depth_updated
= bfq_depth_updated
,
7295 .init_hctx
= bfq_init_hctx
,
7296 .init_sched
= bfq_init_queue
,
7297 .exit_sched
= bfq_exit_queue
,
7300 .icq_size
= sizeof(struct bfq_io_cq
),
7301 .icq_align
= __alignof__(struct bfq_io_cq
),
7302 .elevator_attrs
= bfq_attrs
,
7303 .elevator_name
= "bfq",
7304 .elevator_owner
= THIS_MODULE
,
7306 MODULE_ALIAS("bfq-iosched");
7308 static int __init
bfq_init(void)
7312 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7313 ret
= blkcg_policy_register(&blkcg_policy_bfq
);
7319 if (bfq_slab_setup())
7323 * Times to load large popular applications for the typical
7324 * systems installed on the reference devices (see the
7325 * comments before the definition of the next
7326 * array). Actually, we use slightly lower values, as the
7327 * estimated peak rate tends to be smaller than the actual
7328 * peak rate. The reason for this last fact is that estimates
7329 * are computed over much shorter time intervals than the long
7330 * intervals typically used for benchmarking. Why? First, to
7331 * adapt more quickly to variations. Second, because an I/O
7332 * scheduler cannot rely on a peak-rate-evaluation workload to
7333 * be run for a long time.
7335 ref_wr_duration
[0] = msecs_to_jiffies(7000); /* actually 8 sec */
7336 ref_wr_duration
[1] = msecs_to_jiffies(2500); /* actually 3 sec */
7338 ret
= elv_register(&iosched_bfq_mq
);
7347 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7348 blkcg_policy_unregister(&blkcg_policy_bfq
);
7353 static void __exit
bfq_exit(void)
7355 elv_unregister(&iosched_bfq_mq
);
7356 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7357 blkcg_policy_unregister(&blkcg_policy_bfq
);
7362 module_init(bfq_init
);
7363 module_exit(bfq_exit
);
7365 MODULE_AUTHOR("Paolo Valente");
7366 MODULE_LICENSE("GPL");
7367 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");