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 if (bfqd
->queued
!= 0) {
465 bfq_log(bfqd
, "schedule dispatch");
466 blk_mq_run_hw_queues(bfqd
->queue
, true);
470 #define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
472 #define bfq_sample_valid(samples) ((samples) > 80)
475 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
476 * We choose the request that is closer to the head right now. Distance
477 * behind the head is penalized and only allowed to a certain extent.
479 static struct request
*bfq_choose_req(struct bfq_data
*bfqd
,
484 sector_t s1
, s2
, d1
= 0, d2
= 0;
485 unsigned long back_max
;
486 #define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
487 #define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
488 unsigned int wrap
= 0; /* bit mask: requests behind the disk head? */
490 if (!rq1
|| rq1
== rq2
)
495 if (rq_is_sync(rq1
) && !rq_is_sync(rq2
))
497 else if (rq_is_sync(rq2
) && !rq_is_sync(rq1
))
499 if ((rq1
->cmd_flags
& REQ_META
) && !(rq2
->cmd_flags
& REQ_META
))
501 else if ((rq2
->cmd_flags
& REQ_META
) && !(rq1
->cmd_flags
& REQ_META
))
504 s1
= blk_rq_pos(rq1
);
505 s2
= blk_rq_pos(rq2
);
508 * By definition, 1KiB is 2 sectors.
510 back_max
= bfqd
->bfq_back_max
* 2;
513 * Strict one way elevator _except_ in the case where we allow
514 * short backward seeks which are biased as twice the cost of a
515 * similar forward seek.
519 else if (s1
+ back_max
>= last
)
520 d1
= (last
- s1
) * bfqd
->bfq_back_penalty
;
522 wrap
|= BFQ_RQ1_WRAP
;
526 else if (s2
+ back_max
>= last
)
527 d2
= (last
- s2
) * bfqd
->bfq_back_penalty
;
529 wrap
|= BFQ_RQ2_WRAP
;
531 /* Found required data */
534 * By doing switch() on the bit mask "wrap" we avoid having to
535 * check two variables for all permutations: --> faster!
538 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
553 case BFQ_RQ1_WRAP
|BFQ_RQ2_WRAP
: /* both rqs wrapped */
556 * Since both rqs are wrapped,
557 * start with the one that's further behind head
558 * (--> only *one* back seek required),
559 * since back seek takes more time than forward.
569 * Async I/O can easily starve sync I/O (both sync reads and sync
570 * writes), by consuming all tags. Similarly, storms of sync writes,
571 * such as those that sync(2) may trigger, can starve sync reads.
572 * Limit depths of async I/O and sync writes so as to counter both
575 static void bfq_limit_depth(unsigned int op
, struct blk_mq_alloc_data
*data
)
577 struct bfq_data
*bfqd
= data
->q
->elevator
->elevator_data
;
579 if (op_is_sync(op
) && !op_is_write(op
))
582 data
->shallow_depth
=
583 bfqd
->word_depths
[!!bfqd
->wr_busy_queues
][op_is_sync(op
)];
585 bfq_log(bfqd
, "[%s] wr_busy %d sync %d depth %u",
586 __func__
, bfqd
->wr_busy_queues
, op_is_sync(op
),
587 data
->shallow_depth
);
590 static struct bfq_queue
*
591 bfq_rq_pos_tree_lookup(struct bfq_data
*bfqd
, struct rb_root
*root
,
592 sector_t sector
, struct rb_node
**ret_parent
,
593 struct rb_node
***rb_link
)
595 struct rb_node
**p
, *parent
;
596 struct bfq_queue
*bfqq
= NULL
;
604 bfqq
= rb_entry(parent
, struct bfq_queue
, pos_node
);
607 * Sort strictly based on sector. Smallest to the left,
608 * largest to the right.
610 if (sector
> blk_rq_pos(bfqq
->next_rq
))
612 else if (sector
< blk_rq_pos(bfqq
->next_rq
))
620 *ret_parent
= parent
;
624 bfq_log(bfqd
, "rq_pos_tree_lookup %llu: returning %d",
625 (unsigned long long)sector
,
626 bfqq
? bfqq
->pid
: 0);
631 static bool bfq_too_late_for_merging(struct bfq_queue
*bfqq
)
633 return bfqq
->service_from_backlogged
> 0 &&
634 time_is_before_jiffies(bfqq
->first_IO_time
+
635 bfq_merge_time_limit
);
639 * The following function is not marked as __cold because it is
640 * actually cold, but for the same performance goal described in the
641 * comments on the likely() at the beginning of
642 * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
643 * execution time for the case where this function is not invoked, we
644 * had to add an unlikely() in each involved if().
647 bfq_pos_tree_add_move(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
649 struct rb_node
**p
, *parent
;
650 struct bfq_queue
*__bfqq
;
652 if (bfqq
->pos_root
) {
653 rb_erase(&bfqq
->pos_node
, bfqq
->pos_root
);
654 bfqq
->pos_root
= NULL
;
657 /* oom_bfqq does not participate in queue merging */
658 if (bfqq
== &bfqd
->oom_bfqq
)
662 * bfqq cannot be merged any longer (see comments in
663 * bfq_setup_cooperator): no point in adding bfqq into the
666 if (bfq_too_late_for_merging(bfqq
))
669 if (bfq_class_idle(bfqq
))
674 bfqq
->pos_root
= &bfq_bfqq_to_bfqg(bfqq
)->rq_pos_tree
;
675 __bfqq
= bfq_rq_pos_tree_lookup(bfqd
, bfqq
->pos_root
,
676 blk_rq_pos(bfqq
->next_rq
), &parent
, &p
);
678 rb_link_node(&bfqq
->pos_node
, parent
, p
);
679 rb_insert_color(&bfqq
->pos_node
, bfqq
->pos_root
);
681 bfqq
->pos_root
= NULL
;
685 * The following function returns false either if every active queue
686 * must receive the same share of the throughput (symmetric scenario),
687 * or, as a special case, if bfqq must receive a share of the
688 * throughput lower than or equal to the share that every other active
689 * queue must receive. If bfqq does sync I/O, then these are the only
690 * two cases where bfqq happens to be guaranteed its share of the
691 * throughput even if I/O dispatching is not plugged when bfqq remains
692 * temporarily empty (for more details, see the comments in the
693 * function bfq_better_to_idle()). For this reason, the return value
694 * of this function is used to check whether I/O-dispatch plugging can
697 * The above first case (symmetric scenario) occurs when:
698 * 1) all active queues have the same weight,
699 * 2) all active queues belong to the same I/O-priority class,
700 * 3) all active groups at the same level in the groups tree have the same
702 * 4) all active groups at the same level in the groups tree have the same
703 * number of children.
705 * Unfortunately, keeping the necessary state for evaluating exactly
706 * the last two symmetry sub-conditions above would be quite complex
707 * and time consuming. Therefore this function evaluates, instead,
708 * only the following stronger three sub-conditions, for which it is
709 * much easier to maintain the needed state:
710 * 1) all active queues have the same weight,
711 * 2) all active queues belong to the same I/O-priority class,
712 * 3) there are no active groups.
713 * In particular, the last condition is always true if hierarchical
714 * support or the cgroups interface are not enabled, thus no state
715 * needs to be maintained in this case.
717 static bool bfq_asymmetric_scenario(struct bfq_data
*bfqd
,
718 struct bfq_queue
*bfqq
)
720 bool smallest_weight
= bfqq
&&
721 bfqq
->weight_counter
&&
722 bfqq
->weight_counter
==
724 rb_first_cached(&bfqd
->queue_weights_tree
),
725 struct bfq_weight_counter
,
729 * For queue weights to differ, queue_weights_tree must contain
730 * at least two nodes.
732 bool varied_queue_weights
= !smallest_weight
&&
733 !RB_EMPTY_ROOT(&bfqd
->queue_weights_tree
.rb_root
) &&
734 (bfqd
->queue_weights_tree
.rb_root
.rb_node
->rb_left
||
735 bfqd
->queue_weights_tree
.rb_root
.rb_node
->rb_right
);
737 bool multiple_classes_busy
=
738 (bfqd
->busy_queues
[0] && bfqd
->busy_queues
[1]) ||
739 (bfqd
->busy_queues
[0] && bfqd
->busy_queues
[2]) ||
740 (bfqd
->busy_queues
[1] && bfqd
->busy_queues
[2]);
742 return varied_queue_weights
|| multiple_classes_busy
743 #ifdef CONFIG_BFQ_GROUP_IOSCHED
744 || bfqd
->num_groups_with_pending_reqs
> 0
750 * If the weight-counter tree passed as input contains no counter for
751 * the weight of the input queue, then add that counter; otherwise just
752 * increment the existing counter.
754 * Note that weight-counter trees contain few nodes in mostly symmetric
755 * scenarios. For example, if all queues have the same weight, then the
756 * weight-counter tree for the queues may contain at most one node.
757 * This holds even if low_latency is on, because weight-raised queues
758 * are not inserted in the tree.
759 * In most scenarios, the rate at which nodes are created/destroyed
762 void bfq_weights_tree_add(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
763 struct rb_root_cached
*root
)
765 struct bfq_entity
*entity
= &bfqq
->entity
;
766 struct rb_node
**new = &(root
->rb_root
.rb_node
), *parent
= NULL
;
767 bool leftmost
= true;
770 * Do not insert if the queue is already associated with a
771 * counter, which happens if:
772 * 1) a request arrival has caused the queue to become both
773 * non-weight-raised, and hence change its weight, and
774 * backlogged; in this respect, each of the two events
775 * causes an invocation of this function,
776 * 2) this is the invocation of this function caused by the
777 * second event. This second invocation is actually useless,
778 * and we handle this fact by exiting immediately. More
779 * efficient or clearer solutions might possibly be adopted.
781 if (bfqq
->weight_counter
)
785 struct bfq_weight_counter
*__counter
= container_of(*new,
786 struct bfq_weight_counter
,
790 if (entity
->weight
== __counter
->weight
) {
791 bfqq
->weight_counter
= __counter
;
794 if (entity
->weight
< __counter
->weight
)
795 new = &((*new)->rb_left
);
797 new = &((*new)->rb_right
);
802 bfqq
->weight_counter
= kzalloc(sizeof(struct bfq_weight_counter
),
806 * In the unlucky event of an allocation failure, we just
807 * exit. This will cause the weight of queue to not be
808 * considered in bfq_asymmetric_scenario, which, in its turn,
809 * causes the scenario to be deemed wrongly symmetric in case
810 * bfqq's weight would have been the only weight making the
811 * scenario asymmetric. On the bright side, no unbalance will
812 * however occur when bfqq becomes inactive again (the
813 * invocation of this function is triggered by an activation
814 * of queue). In fact, bfq_weights_tree_remove does nothing
815 * if !bfqq->weight_counter.
817 if (unlikely(!bfqq
->weight_counter
))
820 bfqq
->weight_counter
->weight
= entity
->weight
;
821 rb_link_node(&bfqq
->weight_counter
->weights_node
, parent
, new);
822 rb_insert_color_cached(&bfqq
->weight_counter
->weights_node
, root
,
826 bfqq
->weight_counter
->num_active
++;
831 * Decrement the weight counter associated with the queue, and, if the
832 * counter reaches 0, remove the counter from the tree.
833 * See the comments to the function bfq_weights_tree_add() for considerations
836 void __bfq_weights_tree_remove(struct bfq_data
*bfqd
,
837 struct bfq_queue
*bfqq
,
838 struct rb_root_cached
*root
)
840 if (!bfqq
->weight_counter
)
843 bfqq
->weight_counter
->num_active
--;
844 if (bfqq
->weight_counter
->num_active
> 0)
845 goto reset_entity_pointer
;
847 rb_erase_cached(&bfqq
->weight_counter
->weights_node
, root
);
848 kfree(bfqq
->weight_counter
);
850 reset_entity_pointer
:
851 bfqq
->weight_counter
= NULL
;
856 * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
857 * of active groups for each queue's inactive parent entity.
859 void bfq_weights_tree_remove(struct bfq_data
*bfqd
,
860 struct bfq_queue
*bfqq
)
862 struct bfq_entity
*entity
= bfqq
->entity
.parent
;
864 for_each_entity(entity
) {
865 struct bfq_sched_data
*sd
= entity
->my_sched_data
;
867 if (sd
->next_in_service
|| sd
->in_service_entity
) {
869 * entity is still active, because either
870 * next_in_service or in_service_entity is not
871 * NULL (see the comments on the definition of
872 * next_in_service for details on why
873 * in_service_entity must be checked too).
875 * As a consequence, its parent entities are
876 * active as well, and thus this loop must
883 * The decrement of num_groups_with_pending_reqs is
884 * not performed immediately upon the deactivation of
885 * entity, but it is delayed to when it also happens
886 * that the first leaf descendant bfqq of entity gets
887 * all its pending requests completed. The following
888 * instructions perform this delayed decrement, if
889 * needed. See the comments on
890 * num_groups_with_pending_reqs for details.
892 if (entity
->in_groups_with_pending_reqs
) {
893 entity
->in_groups_with_pending_reqs
= false;
894 bfqd
->num_groups_with_pending_reqs
--;
899 * Next function is invoked last, because it causes bfqq to be
900 * freed if the following holds: bfqq is not in service and
901 * has no dispatched request. DO NOT use bfqq after the next
902 * function invocation.
904 __bfq_weights_tree_remove(bfqd
, bfqq
,
905 &bfqd
->queue_weights_tree
);
909 * Return expired entry, or NULL to just start from scratch in rbtree.
911 static struct request
*bfq_check_fifo(struct bfq_queue
*bfqq
,
912 struct request
*last
)
916 if (bfq_bfqq_fifo_expire(bfqq
))
919 bfq_mark_bfqq_fifo_expire(bfqq
);
921 rq
= rq_entry_fifo(bfqq
->fifo
.next
);
923 if (rq
== last
|| ktime_get_ns() < rq
->fifo_time
)
926 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "check_fifo: returned %p", rq
);
930 static struct request
*bfq_find_next_rq(struct bfq_data
*bfqd
,
931 struct bfq_queue
*bfqq
,
932 struct request
*last
)
934 struct rb_node
*rbnext
= rb_next(&last
->rb_node
);
935 struct rb_node
*rbprev
= rb_prev(&last
->rb_node
);
936 struct request
*next
, *prev
= NULL
;
938 /* Follow expired path, else get first next available. */
939 next
= bfq_check_fifo(bfqq
, last
);
944 prev
= rb_entry_rq(rbprev
);
947 next
= rb_entry_rq(rbnext
);
949 rbnext
= rb_first(&bfqq
->sort_list
);
950 if (rbnext
&& rbnext
!= &last
->rb_node
)
951 next
= rb_entry_rq(rbnext
);
954 return bfq_choose_req(bfqd
, next
, prev
, blk_rq_pos(last
));
957 /* see the definition of bfq_async_charge_factor for details */
958 static unsigned long bfq_serv_to_charge(struct request
*rq
,
959 struct bfq_queue
*bfqq
)
961 if (bfq_bfqq_sync(bfqq
) || bfqq
->wr_coeff
> 1 ||
962 bfq_asymmetric_scenario(bfqq
->bfqd
, bfqq
))
963 return blk_rq_sectors(rq
);
965 return blk_rq_sectors(rq
) * bfq_async_charge_factor
;
969 * bfq_updated_next_req - update the queue after a new next_rq selection.
970 * @bfqd: the device data the queue belongs to.
971 * @bfqq: the queue to update.
973 * If the first request of a queue changes we make sure that the queue
974 * has enough budget to serve at least its first request (if the
975 * request has grown). We do this because if the queue has not enough
976 * budget for its first request, it has to go through two dispatch
977 * rounds to actually get it dispatched.
979 static void bfq_updated_next_req(struct bfq_data
*bfqd
,
980 struct bfq_queue
*bfqq
)
982 struct bfq_entity
*entity
= &bfqq
->entity
;
983 struct request
*next_rq
= bfqq
->next_rq
;
984 unsigned long new_budget
;
989 if (bfqq
== bfqd
->in_service_queue
)
991 * In order not to break guarantees, budgets cannot be
992 * changed after an entity has been selected.
996 new_budget
= max_t(unsigned long,
997 max_t(unsigned long, bfqq
->max_budget
,
998 bfq_serv_to_charge(next_rq
, bfqq
)),
1000 if (entity
->budget
!= new_budget
) {
1001 entity
->budget
= new_budget
;
1002 bfq_log_bfqq(bfqd
, bfqq
, "updated next rq: new budget %lu",
1004 bfq_requeue_bfqq(bfqd
, bfqq
, false);
1008 static unsigned int bfq_wr_duration(struct bfq_data
*bfqd
)
1012 if (bfqd
->bfq_wr_max_time
> 0)
1013 return bfqd
->bfq_wr_max_time
;
1015 dur
= bfqd
->rate_dur_prod
;
1016 do_div(dur
, bfqd
->peak_rate
);
1019 * Limit duration between 3 and 25 seconds. The upper limit
1020 * has been conservatively set after the following worst case:
1021 * on a QEMU/KVM virtual machine
1022 * - running in a slow PC
1023 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
1024 * - serving a heavy I/O workload, such as the sequential reading
1026 * mplayer took 23 seconds to start, if constantly weight-raised.
1028 * As for higher values than that accommodating the above bad
1029 * scenario, tests show that higher values would often yield
1030 * the opposite of the desired result, i.e., would worsen
1031 * responsiveness by allowing non-interactive applications to
1032 * preserve weight raising for too long.
1034 * On the other end, lower values than 3 seconds make it
1035 * difficult for most interactive tasks to complete their jobs
1036 * before weight-raising finishes.
1038 return clamp_val(dur
, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
1041 /* switch back from soft real-time to interactive weight raising */
1042 static void switch_back_to_interactive_wr(struct bfq_queue
*bfqq
,
1043 struct bfq_data
*bfqd
)
1045 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1046 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1047 bfqq
->last_wr_start_finish
= bfqq
->wr_start_at_switch_to_srt
;
1051 bfq_bfqq_resume_state(struct bfq_queue
*bfqq
, struct bfq_data
*bfqd
,
1052 struct bfq_io_cq
*bic
, bool bfq_already_existing
)
1054 unsigned int old_wr_coeff
= 1;
1055 bool busy
= bfq_already_existing
&& bfq_bfqq_busy(bfqq
);
1057 if (bic
->saved_has_short_ttime
)
1058 bfq_mark_bfqq_has_short_ttime(bfqq
);
1060 bfq_clear_bfqq_has_short_ttime(bfqq
);
1062 if (bic
->saved_IO_bound
)
1063 bfq_mark_bfqq_IO_bound(bfqq
);
1065 bfq_clear_bfqq_IO_bound(bfqq
);
1067 bfqq
->last_serv_time_ns
= bic
->saved_last_serv_time_ns
;
1068 bfqq
->inject_limit
= bic
->saved_inject_limit
;
1069 bfqq
->decrease_time_jif
= bic
->saved_decrease_time_jif
;
1071 bfqq
->entity
.new_weight
= bic
->saved_weight
;
1072 bfqq
->ttime
= bic
->saved_ttime
;
1073 bfqq
->io_start_time
= bic
->saved_io_start_time
;
1074 bfqq
->tot_idle_time
= bic
->saved_tot_idle_time
;
1076 * Restore weight coefficient only if low_latency is on
1078 if (bfqd
->low_latency
) {
1079 old_wr_coeff
= bfqq
->wr_coeff
;
1080 bfqq
->wr_coeff
= bic
->saved_wr_coeff
;
1082 bfqq
->service_from_wr
= bic
->saved_service_from_wr
;
1083 bfqq
->wr_start_at_switch_to_srt
= bic
->saved_wr_start_at_switch_to_srt
;
1084 bfqq
->last_wr_start_finish
= bic
->saved_last_wr_start_finish
;
1085 bfqq
->wr_cur_max_time
= bic
->saved_wr_cur_max_time
;
1087 if (bfqq
->wr_coeff
> 1 && (bfq_bfqq_in_large_burst(bfqq
) ||
1088 time_is_before_jiffies(bfqq
->last_wr_start_finish
+
1089 bfqq
->wr_cur_max_time
))) {
1090 if (bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
&&
1091 !bfq_bfqq_in_large_burst(bfqq
) &&
1092 time_is_after_eq_jiffies(bfqq
->wr_start_at_switch_to_srt
+
1093 bfq_wr_duration(bfqd
))) {
1094 switch_back_to_interactive_wr(bfqq
, bfqd
);
1097 bfq_log_bfqq(bfqq
->bfqd
, bfqq
,
1098 "resume state: switching off wr");
1102 /* make sure weight will be updated, however we got here */
1103 bfqq
->entity
.prio_changed
= 1;
1108 if (old_wr_coeff
== 1 && bfqq
->wr_coeff
> 1)
1109 bfqd
->wr_busy_queues
++;
1110 else if (old_wr_coeff
> 1 && bfqq
->wr_coeff
== 1)
1111 bfqd
->wr_busy_queues
--;
1114 static int bfqq_process_refs(struct bfq_queue
*bfqq
)
1116 return bfqq
->ref
- bfqq
->allocated
- bfqq
->entity
.on_st_or_in_serv
-
1117 (bfqq
->weight_counter
!= NULL
) - bfqq
->stable_ref
;
1120 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1121 static void bfq_reset_burst_list(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
1123 struct bfq_queue
*item
;
1124 struct hlist_node
*n
;
1126 hlist_for_each_entry_safe(item
, n
, &bfqd
->burst_list
, burst_list_node
)
1127 hlist_del_init(&item
->burst_list_node
);
1130 * Start the creation of a new burst list only if there is no
1131 * active queue. See comments on the conditional invocation of
1132 * bfq_handle_burst().
1134 if (bfq_tot_busy_queues(bfqd
) == 0) {
1135 hlist_add_head(&bfqq
->burst_list_node
, &bfqd
->burst_list
);
1136 bfqd
->burst_size
= 1;
1138 bfqd
->burst_size
= 0;
1140 bfqd
->burst_parent_entity
= bfqq
->entity
.parent
;
1143 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1144 static void bfq_add_to_burst(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
1146 /* Increment burst size to take into account also bfqq */
1149 if (bfqd
->burst_size
== bfqd
->bfq_large_burst_thresh
) {
1150 struct bfq_queue
*pos
, *bfqq_item
;
1151 struct hlist_node
*n
;
1154 * Enough queues have been activated shortly after each
1155 * other to consider this burst as large.
1157 bfqd
->large_burst
= true;
1160 * We can now mark all queues in the burst list as
1161 * belonging to a large burst.
1163 hlist_for_each_entry(bfqq_item
, &bfqd
->burst_list
,
1165 bfq_mark_bfqq_in_large_burst(bfqq_item
);
1166 bfq_mark_bfqq_in_large_burst(bfqq
);
1169 * From now on, and until the current burst finishes, any
1170 * new queue being activated shortly after the last queue
1171 * was inserted in the burst can be immediately marked as
1172 * belonging to a large burst. So the burst list is not
1173 * needed any more. Remove it.
1175 hlist_for_each_entry_safe(pos
, n
, &bfqd
->burst_list
,
1177 hlist_del_init(&pos
->burst_list_node
);
1179 * Burst not yet large: add bfqq to the burst list. Do
1180 * not increment the ref counter for bfqq, because bfqq
1181 * is removed from the burst list before freeing bfqq
1184 hlist_add_head(&bfqq
->burst_list_node
, &bfqd
->burst_list
);
1188 * If many queues belonging to the same group happen to be created
1189 * shortly after each other, then the processes associated with these
1190 * queues have typically a common goal. In particular, bursts of queue
1191 * creations are usually caused by services or applications that spawn
1192 * many parallel threads/processes. Examples are systemd during boot,
1193 * or git grep. To help these processes get their job done as soon as
1194 * possible, it is usually better to not grant either weight-raising
1195 * or device idling to their queues, unless these queues must be
1196 * protected from the I/O flowing through other active queues.
1198 * In this comment we describe, firstly, the reasons why this fact
1199 * holds, and, secondly, the next function, which implements the main
1200 * steps needed to properly mark these queues so that they can then be
1201 * treated in a different way.
1203 * The above services or applications benefit mostly from a high
1204 * throughput: the quicker the requests of the activated queues are
1205 * cumulatively served, the sooner the target job of these queues gets
1206 * completed. As a consequence, weight-raising any of these queues,
1207 * which also implies idling the device for it, is almost always
1208 * counterproductive, unless there are other active queues to isolate
1209 * these new queues from. If there no other active queues, then
1210 * weight-raising these new queues just lowers throughput in most
1213 * On the other hand, a burst of queue creations may be caused also by
1214 * the start of an application that does not consist of a lot of
1215 * parallel I/O-bound threads. In fact, with a complex application,
1216 * several short processes may need to be executed to start-up the
1217 * application. In this respect, to start an application as quickly as
1218 * possible, the best thing to do is in any case to privilege the I/O
1219 * related to the application with respect to all other
1220 * I/O. Therefore, the best strategy to start as quickly as possible
1221 * an application that causes a burst of queue creations is to
1222 * weight-raise all the queues created during the burst. This is the
1223 * exact opposite of the best strategy for the other type of bursts.
1225 * In the end, to take the best action for each of the two cases, the
1226 * two types of bursts need to be distinguished. Fortunately, this
1227 * seems relatively easy, by looking at the sizes of the bursts. In
1228 * particular, we found a threshold such that only bursts with a
1229 * larger size than that threshold are apparently caused by
1230 * services or commands such as systemd or git grep. For brevity,
1231 * hereafter we call just 'large' these bursts. BFQ *does not*
1232 * weight-raise queues whose creation occurs in a large burst. In
1233 * addition, for each of these queues BFQ performs or does not perform
1234 * idling depending on which choice boosts the throughput more. The
1235 * exact choice depends on the device and request pattern at
1238 * Unfortunately, false positives may occur while an interactive task
1239 * is starting (e.g., an application is being started). The
1240 * consequence is that the queues associated with the task do not
1241 * enjoy weight raising as expected. Fortunately these false positives
1242 * are very rare. They typically occur if some service happens to
1243 * start doing I/O exactly when the interactive task starts.
1245 * Turning back to the next function, it is invoked only if there are
1246 * no active queues (apart from active queues that would belong to the
1247 * same, possible burst bfqq would belong to), and it implements all
1248 * the steps needed to detect the occurrence of a large burst and to
1249 * properly mark all the queues belonging to it (so that they can then
1250 * be treated in a different way). This goal is achieved by
1251 * maintaining a "burst list" that holds, temporarily, the queues that
1252 * belong to the burst in progress. The list is then used to mark
1253 * these queues as belonging to a large burst if the burst does become
1254 * large. The main steps are the following.
1256 * . when the very first queue is created, the queue is inserted into the
1257 * list (as it could be the first queue in a possible burst)
1259 * . if the current burst has not yet become large, and a queue Q that does
1260 * not yet belong to the burst is activated shortly after the last time
1261 * at which a new queue entered the burst list, then the function appends
1262 * Q to the burst list
1264 * . if, as a consequence of the previous step, the burst size reaches
1265 * the large-burst threshold, then
1267 * . all the queues in the burst list are marked as belonging to a
1270 * . the burst list is deleted; in fact, the burst list already served
1271 * its purpose (keeping temporarily track of the queues in a burst,
1272 * so as to be able to mark them as belonging to a large burst in the
1273 * previous sub-step), and now is not needed any more
1275 * . the device enters a large-burst mode
1277 * . if a queue Q that does not belong to the burst is created while
1278 * the device is in large-burst mode and shortly after the last time
1279 * at which a queue either entered the burst list or was marked as
1280 * belonging to the current large burst, then Q is immediately marked
1281 * as belonging to a large burst.
1283 * . if a queue Q that does not belong to the burst is created a while
1284 * later, i.e., not shortly after, than the last time at which a queue
1285 * either entered the burst list or was marked as belonging to the
1286 * current large burst, then the current burst is deemed as finished and:
1288 * . the large-burst mode is reset if set
1290 * . the burst list is emptied
1292 * . Q is inserted in the burst list, as Q may be the first queue
1293 * in a possible new burst (then the burst list contains just Q
1296 static void bfq_handle_burst(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
1299 * If bfqq is already in the burst list or is part of a large
1300 * burst, or finally has just been split, then there is
1301 * nothing else to do.
1303 if (!hlist_unhashed(&bfqq
->burst_list_node
) ||
1304 bfq_bfqq_in_large_burst(bfqq
) ||
1305 time_is_after_eq_jiffies(bfqq
->split_time
+
1306 msecs_to_jiffies(10)))
1310 * If bfqq's creation happens late enough, or bfqq belongs to
1311 * a different group than the burst group, then the current
1312 * burst is finished, and related data structures must be
1315 * In this respect, consider the special case where bfqq is
1316 * the very first queue created after BFQ is selected for this
1317 * device. In this case, last_ins_in_burst and
1318 * burst_parent_entity are not yet significant when we get
1319 * here. But it is easy to verify that, whether or not the
1320 * following condition is true, bfqq will end up being
1321 * inserted into the burst list. In particular the list will
1322 * happen to contain only bfqq. And this is exactly what has
1323 * to happen, as bfqq may be the first queue of the first
1326 if (time_is_before_jiffies(bfqd
->last_ins_in_burst
+
1327 bfqd
->bfq_burst_interval
) ||
1328 bfqq
->entity
.parent
!= bfqd
->burst_parent_entity
) {
1329 bfqd
->large_burst
= false;
1330 bfq_reset_burst_list(bfqd
, bfqq
);
1335 * If we get here, then bfqq is being activated shortly after the
1336 * last queue. So, if the current burst is also large, we can mark
1337 * bfqq as belonging to this large burst immediately.
1339 if (bfqd
->large_burst
) {
1340 bfq_mark_bfqq_in_large_burst(bfqq
);
1345 * If we get here, then a large-burst state has not yet been
1346 * reached, but bfqq is being activated shortly after the last
1347 * queue. Then we add bfqq to the burst.
1349 bfq_add_to_burst(bfqd
, bfqq
);
1352 * At this point, bfqq either has been added to the current
1353 * burst or has caused the current burst to terminate and a
1354 * possible new burst to start. In particular, in the second
1355 * case, bfqq has become the first queue in the possible new
1356 * burst. In both cases last_ins_in_burst needs to be moved
1359 bfqd
->last_ins_in_burst
= jiffies
;
1362 static int bfq_bfqq_budget_left(struct bfq_queue
*bfqq
)
1364 struct bfq_entity
*entity
= &bfqq
->entity
;
1366 return entity
->budget
- entity
->service
;
1370 * If enough samples have been computed, return the current max budget
1371 * stored in bfqd, which is dynamically updated according to the
1372 * estimated disk peak rate; otherwise return the default max budget
1374 static int bfq_max_budget(struct bfq_data
*bfqd
)
1376 if (bfqd
->budgets_assigned
< bfq_stats_min_budgets
)
1377 return bfq_default_max_budget
;
1379 return bfqd
->bfq_max_budget
;
1383 * Return min budget, which is a fraction of the current or default
1384 * max budget (trying with 1/32)
1386 static int bfq_min_budget(struct bfq_data
*bfqd
)
1388 if (bfqd
->budgets_assigned
< bfq_stats_min_budgets
)
1389 return bfq_default_max_budget
/ 32;
1391 return bfqd
->bfq_max_budget
/ 32;
1395 * The next function, invoked after the input queue bfqq switches from
1396 * idle to busy, updates the budget of bfqq. The function also tells
1397 * whether the in-service queue should be expired, by returning
1398 * true. The purpose of expiring the in-service queue is to give bfqq
1399 * the chance to possibly preempt the in-service queue, and the reason
1400 * for preempting the in-service queue is to achieve one of the two
1403 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1404 * expired because it has remained idle. In particular, bfqq may have
1405 * expired for one of the following two reasons:
1407 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1408 * and did not make it to issue a new request before its last
1409 * request was served;
1411 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1412 * a new request before the expiration of the idling-time.
1414 * Even if bfqq has expired for one of the above reasons, the process
1415 * associated with the queue may be however issuing requests greedily,
1416 * and thus be sensitive to the bandwidth it receives (bfqq may have
1417 * remained idle for other reasons: CPU high load, bfqq not enjoying
1418 * idling, I/O throttling somewhere in the path from the process to
1419 * the I/O scheduler, ...). But if, after every expiration for one of
1420 * the above two reasons, bfqq has to wait for the service of at least
1421 * one full budget of another queue before being served again, then
1422 * bfqq is likely to get a much lower bandwidth or resource time than
1423 * its reserved ones. To address this issue, two countermeasures need
1426 * First, the budget and the timestamps of bfqq need to be updated in
1427 * a special way on bfqq reactivation: they need to be updated as if
1428 * bfqq did not remain idle and did not expire. In fact, if they are
1429 * computed as if bfqq expired and remained idle until reactivation,
1430 * then the process associated with bfqq is treated as if, instead of
1431 * being greedy, it stopped issuing requests when bfqq remained idle,
1432 * and restarts issuing requests only on this reactivation. In other
1433 * words, the scheduler does not help the process recover the "service
1434 * hole" between bfqq expiration and reactivation. As a consequence,
1435 * the process receives a lower bandwidth than its reserved one. In
1436 * contrast, to recover this hole, the budget must be updated as if
1437 * bfqq was not expired at all before this reactivation, i.e., it must
1438 * be set to the value of the remaining budget when bfqq was
1439 * expired. Along the same line, timestamps need to be assigned the
1440 * value they had the last time bfqq was selected for service, i.e.,
1441 * before last expiration. Thus timestamps need to be back-shifted
1442 * with respect to their normal computation (see [1] for more details
1443 * on this tricky aspect).
1445 * Secondly, to allow the process to recover the hole, the in-service
1446 * queue must be expired too, to give bfqq the chance to preempt it
1447 * immediately. In fact, if bfqq has to wait for a full budget of the
1448 * in-service queue to be completed, then it may become impossible to
1449 * let the process recover the hole, even if the back-shifted
1450 * timestamps of bfqq are lower than those of the in-service queue. If
1451 * this happens for most or all of the holes, then the process may not
1452 * receive its reserved bandwidth. In this respect, it is worth noting
1453 * that, being the service of outstanding requests unpreemptible, a
1454 * little fraction of the holes may however be unrecoverable, thereby
1455 * causing a little loss of bandwidth.
1457 * The last important point is detecting whether bfqq does need this
1458 * bandwidth recovery. In this respect, the next function deems the
1459 * process associated with bfqq greedy, and thus allows it to recover
1460 * the hole, if: 1) the process is waiting for the arrival of a new
1461 * request (which implies that bfqq expired for one of the above two
1462 * reasons), and 2) such a request has arrived soon. The first
1463 * condition is controlled through the flag non_blocking_wait_rq,
1464 * while the second through the flag arrived_in_time. If both
1465 * conditions hold, then the function computes the budget in the
1466 * above-described special way, and signals that the in-service queue
1467 * should be expired. Timestamp back-shifting is done later in
1468 * __bfq_activate_entity.
1470 * 2. Reduce latency. Even if timestamps are not backshifted to let
1471 * the process associated with bfqq recover a service hole, bfqq may
1472 * however happen to have, after being (re)activated, a lower finish
1473 * timestamp than the in-service queue. That is, the next budget of
1474 * bfqq may have to be completed before the one of the in-service
1475 * queue. If this is the case, then preempting the in-service queue
1476 * allows this goal to be achieved, apart from the unpreemptible,
1477 * outstanding requests mentioned above.
1479 * Unfortunately, regardless of which of the above two goals one wants
1480 * to achieve, service trees need first to be updated to know whether
1481 * the in-service queue must be preempted. To have service trees
1482 * correctly updated, the in-service queue must be expired and
1483 * rescheduled, and bfqq must be scheduled too. This is one of the
1484 * most costly operations (in future versions, the scheduling
1485 * mechanism may be re-designed in such a way to make it possible to
1486 * know whether preemption is needed without needing to update service
1487 * trees). In addition, queue preemptions almost always cause random
1488 * I/O, which may in turn cause loss of throughput. Finally, there may
1489 * even be no in-service queue when the next function is invoked (so,
1490 * no queue to compare timestamps with). Because of these facts, the
1491 * next function adopts the following simple scheme to avoid costly
1492 * operations, too frequent preemptions and too many dependencies on
1493 * the state of the scheduler: it requests the expiration of the
1494 * in-service queue (unconditionally) only for queues that need to
1495 * recover a hole. Then it delegates to other parts of the code the
1496 * responsibility of handling the above case 2.
1498 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data
*bfqd
,
1499 struct bfq_queue
*bfqq
,
1500 bool arrived_in_time
)
1502 struct bfq_entity
*entity
= &bfqq
->entity
;
1505 * In the next compound condition, we check also whether there
1506 * is some budget left, because otherwise there is no point in
1507 * trying to go on serving bfqq with this same budget: bfqq
1508 * would be expired immediately after being selected for
1509 * service. This would only cause useless overhead.
1511 if (bfq_bfqq_non_blocking_wait_rq(bfqq
) && arrived_in_time
&&
1512 bfq_bfqq_budget_left(bfqq
) > 0) {
1514 * We do not clear the flag non_blocking_wait_rq here, as
1515 * the latter is used in bfq_activate_bfqq to signal
1516 * that timestamps need to be back-shifted (and is
1517 * cleared right after).
1521 * In next assignment we rely on that either
1522 * entity->service or entity->budget are not updated
1523 * on expiration if bfqq is empty (see
1524 * __bfq_bfqq_recalc_budget). Thus both quantities
1525 * remain unchanged after such an expiration, and the
1526 * following statement therefore assigns to
1527 * entity->budget the remaining budget on such an
1530 entity
->budget
= min_t(unsigned long,
1531 bfq_bfqq_budget_left(bfqq
),
1535 * At this point, we have used entity->service to get
1536 * the budget left (needed for updating
1537 * entity->budget). Thus we finally can, and have to,
1538 * reset entity->service. The latter must be reset
1539 * because bfqq would otherwise be charged again for
1540 * the service it has received during its previous
1543 entity
->service
= 0;
1549 * We can finally complete expiration, by setting service to 0.
1551 entity
->service
= 0;
1552 entity
->budget
= max_t(unsigned long, bfqq
->max_budget
,
1553 bfq_serv_to_charge(bfqq
->next_rq
, bfqq
));
1554 bfq_clear_bfqq_non_blocking_wait_rq(bfqq
);
1559 * Return the farthest past time instant according to jiffies
1562 static unsigned long bfq_smallest_from_now(void)
1564 return jiffies
- MAX_JIFFY_OFFSET
;
1567 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data
*bfqd
,
1568 struct bfq_queue
*bfqq
,
1569 unsigned int old_wr_coeff
,
1570 bool wr_or_deserves_wr
,
1575 if (old_wr_coeff
== 1 && wr_or_deserves_wr
) {
1576 /* start a weight-raising period */
1578 bfqq
->service_from_wr
= 0;
1579 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1580 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1583 * No interactive weight raising in progress
1584 * here: assign minus infinity to
1585 * wr_start_at_switch_to_srt, to make sure
1586 * that, at the end of the soft-real-time
1587 * weight raising periods that is starting
1588 * now, no interactive weight-raising period
1589 * may be wrongly considered as still in
1590 * progress (and thus actually started by
1593 bfqq
->wr_start_at_switch_to_srt
=
1594 bfq_smallest_from_now();
1595 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
*
1596 BFQ_SOFTRT_WEIGHT_FACTOR
;
1597 bfqq
->wr_cur_max_time
=
1598 bfqd
->bfq_wr_rt_max_time
;
1602 * If needed, further reduce budget to make sure it is
1603 * close to bfqq's backlog, so as to reduce the
1604 * scheduling-error component due to a too large
1605 * budget. Do not care about throughput consequences,
1606 * but only about latency. Finally, do not assign a
1607 * too small budget either, to avoid increasing
1608 * latency by causing too frequent expirations.
1610 bfqq
->entity
.budget
= min_t(unsigned long,
1611 bfqq
->entity
.budget
,
1612 2 * bfq_min_budget(bfqd
));
1613 } else if (old_wr_coeff
> 1) {
1614 if (interactive
) { /* update wr coeff and duration */
1615 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
;
1616 bfqq
->wr_cur_max_time
= bfq_wr_duration(bfqd
);
1617 } else if (in_burst
)
1621 * The application is now or still meeting the
1622 * requirements for being deemed soft rt. We
1623 * can then correctly and safely (re)charge
1624 * the weight-raising duration for the
1625 * application with the weight-raising
1626 * duration for soft rt applications.
1628 * In particular, doing this recharge now, i.e.,
1629 * before the weight-raising period for the
1630 * application finishes, reduces the probability
1631 * of the following negative scenario:
1632 * 1) the weight of a soft rt application is
1633 * raised at startup (as for any newly
1634 * created application),
1635 * 2) since the application is not interactive,
1636 * at a certain time weight-raising is
1637 * stopped for the application,
1638 * 3) at that time the application happens to
1639 * still have pending requests, and hence
1640 * is destined to not have a chance to be
1641 * deemed soft rt before these requests are
1642 * completed (see the comments to the
1643 * function bfq_bfqq_softrt_next_start()
1644 * for details on soft rt detection),
1645 * 4) these pending requests experience a high
1646 * latency because the application is not
1647 * weight-raised while they are pending.
1649 if (bfqq
->wr_cur_max_time
!=
1650 bfqd
->bfq_wr_rt_max_time
) {
1651 bfqq
->wr_start_at_switch_to_srt
=
1652 bfqq
->last_wr_start_finish
;
1654 bfqq
->wr_cur_max_time
=
1655 bfqd
->bfq_wr_rt_max_time
;
1656 bfqq
->wr_coeff
= bfqd
->bfq_wr_coeff
*
1657 BFQ_SOFTRT_WEIGHT_FACTOR
;
1659 bfqq
->last_wr_start_finish
= jiffies
;
1664 static bool bfq_bfqq_idle_for_long_time(struct bfq_data
*bfqd
,
1665 struct bfq_queue
*bfqq
)
1667 return bfqq
->dispatched
== 0 &&
1668 time_is_before_jiffies(
1669 bfqq
->budget_timeout
+
1670 bfqd
->bfq_wr_min_idle_time
);
1675 * Return true if bfqq is in a higher priority class, or has a higher
1676 * weight than the in-service queue.
1678 static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue
*bfqq
,
1679 struct bfq_queue
*in_serv_bfqq
)
1681 int bfqq_weight
, in_serv_weight
;
1683 if (bfqq
->ioprio_class
< in_serv_bfqq
->ioprio_class
)
1686 if (in_serv_bfqq
->entity
.parent
== bfqq
->entity
.parent
) {
1687 bfqq_weight
= bfqq
->entity
.weight
;
1688 in_serv_weight
= in_serv_bfqq
->entity
.weight
;
1690 if (bfqq
->entity
.parent
)
1691 bfqq_weight
= bfqq
->entity
.parent
->weight
;
1693 bfqq_weight
= bfqq
->entity
.weight
;
1694 if (in_serv_bfqq
->entity
.parent
)
1695 in_serv_weight
= in_serv_bfqq
->entity
.parent
->weight
;
1697 in_serv_weight
= in_serv_bfqq
->entity
.weight
;
1700 return bfqq_weight
> in_serv_weight
;
1703 static bool bfq_better_to_idle(struct bfq_queue
*bfqq
);
1705 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data
*bfqd
,
1706 struct bfq_queue
*bfqq
,
1711 bool soft_rt
, in_burst
, wr_or_deserves_wr
,
1712 bfqq_wants_to_preempt
,
1713 idle_for_long_time
= bfq_bfqq_idle_for_long_time(bfqd
, bfqq
),
1715 * See the comments on
1716 * bfq_bfqq_update_budg_for_activation for
1717 * details on the usage of the next variable.
1719 arrived_in_time
= ktime_get_ns() <=
1720 bfqq
->ttime
.last_end_request
+
1721 bfqd
->bfq_slice_idle
* 3;
1725 * bfqq deserves to be weight-raised if:
1727 * - it does not belong to a large burst,
1728 * - it has been idle for enough time or is soft real-time,
1729 * - is linked to a bfq_io_cq (it is not shared in any sense),
1730 * - has a default weight (otherwise we assume the user wanted
1731 * to control its weight explicitly)
1733 in_burst
= bfq_bfqq_in_large_burst(bfqq
);
1734 soft_rt
= bfqd
->bfq_wr_max_softrt_rate
> 0 &&
1735 !BFQQ_TOTALLY_SEEKY(bfqq
) &&
1737 time_is_before_jiffies(bfqq
->soft_rt_next_start
) &&
1738 bfqq
->dispatched
== 0 &&
1739 bfqq
->entity
.new_weight
== 40;
1740 *interactive
= !in_burst
&& idle_for_long_time
&&
1741 bfqq
->entity
.new_weight
== 40;
1743 * Merged bfq_queues are kept out of weight-raising
1744 * (low-latency) mechanisms. The reason is that these queues
1745 * are usually created for non-interactive and
1746 * non-soft-real-time tasks. Yet this is not the case for
1747 * stably-merged queues. These queues are merged just because
1748 * they are created shortly after each other. So they may
1749 * easily serve the I/O of an interactive or soft-real time
1750 * application, if the application happens to spawn multiple
1751 * processes. So let also stably-merged queued enjoy weight
1754 wr_or_deserves_wr
= bfqd
->low_latency
&&
1755 (bfqq
->wr_coeff
> 1 ||
1756 (bfq_bfqq_sync(bfqq
) &&
1757 (bfqq
->bic
|| RQ_BIC(rq
)->stably_merged
) &&
1758 (*interactive
|| soft_rt
)));
1761 * Using the last flag, update budget and check whether bfqq
1762 * may want to preempt the in-service queue.
1764 bfqq_wants_to_preempt
=
1765 bfq_bfqq_update_budg_for_activation(bfqd
, bfqq
,
1769 * If bfqq happened to be activated in a burst, but has been
1770 * idle for much more than an interactive queue, then we
1771 * assume that, in the overall I/O initiated in the burst, the
1772 * I/O associated with bfqq is finished. So bfqq does not need
1773 * to be treated as a queue belonging to a burst
1774 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1775 * if set, and remove bfqq from the burst list if it's
1776 * there. We do not decrement burst_size, because the fact
1777 * that bfqq does not need to belong to the burst list any
1778 * more does not invalidate the fact that bfqq was created in
1781 if (likely(!bfq_bfqq_just_created(bfqq
)) &&
1782 idle_for_long_time
&&
1783 time_is_before_jiffies(
1784 bfqq
->budget_timeout
+
1785 msecs_to_jiffies(10000))) {
1786 hlist_del_init(&bfqq
->burst_list_node
);
1787 bfq_clear_bfqq_in_large_burst(bfqq
);
1790 bfq_clear_bfqq_just_created(bfqq
);
1792 if (bfqd
->low_latency
) {
1793 if (unlikely(time_is_after_jiffies(bfqq
->split_time
)))
1796 jiffies
- bfqd
->bfq_wr_min_idle_time
- 1;
1798 if (time_is_before_jiffies(bfqq
->split_time
+
1799 bfqd
->bfq_wr_min_idle_time
)) {
1800 bfq_update_bfqq_wr_on_rq_arrival(bfqd
, bfqq
,
1807 if (old_wr_coeff
!= bfqq
->wr_coeff
)
1808 bfqq
->entity
.prio_changed
= 1;
1812 bfqq
->last_idle_bklogged
= jiffies
;
1813 bfqq
->service_from_backlogged
= 0;
1814 bfq_clear_bfqq_softrt_update(bfqq
);
1816 bfq_add_bfqq_busy(bfqd
, bfqq
);
1819 * Expire in-service queue if preemption may be needed for
1820 * guarantees or throughput. As for guarantees, we care
1821 * explicitly about two cases. The first is that bfqq has to
1822 * recover a service hole, as explained in the comments on
1823 * bfq_bfqq_update_budg_for_activation(), i.e., that
1824 * bfqq_wants_to_preempt is true. However, if bfqq does not
1825 * carry time-critical I/O, then bfqq's bandwidth is less
1826 * important than that of queues that carry time-critical I/O.
1827 * So, as a further constraint, we consider this case only if
1828 * bfqq is at least as weight-raised, i.e., at least as time
1829 * critical, as the in-service queue.
1831 * The second case is that bfqq is in a higher priority class,
1832 * or has a higher weight than the in-service queue. If this
1833 * condition does not hold, we don't care because, even if
1834 * bfqq does not start to be served immediately, the resulting
1835 * delay for bfqq's I/O is however lower or much lower than
1836 * the ideal completion time to be guaranteed to bfqq's I/O.
1838 * In both cases, preemption is needed only if, according to
1839 * the timestamps of both bfqq and of the in-service queue,
1840 * bfqq actually is the next queue to serve. So, to reduce
1841 * useless preemptions, the return value of
1842 * next_queue_may_preempt() is considered in the next compound
1843 * condition too. Yet next_queue_may_preempt() just checks a
1844 * simple, necessary condition for bfqq to be the next queue
1845 * to serve. In fact, to evaluate a sufficient condition, the
1846 * timestamps of the in-service queue would need to be
1847 * updated, and this operation is quite costly (see the
1848 * comments on bfq_bfqq_update_budg_for_activation()).
1850 * As for throughput, we ask bfq_better_to_idle() whether we
1851 * still need to plug I/O dispatching. If bfq_better_to_idle()
1852 * says no, then plugging is not needed any longer, either to
1853 * boost throughput or to perserve service guarantees. Then
1854 * the best option is to stop plugging I/O, as not doing so
1855 * would certainly lower throughput. We may end up in this
1856 * case if: (1) upon a dispatch attempt, we detected that it
1857 * was better to plug I/O dispatch, and to wait for a new
1858 * request to arrive for the currently in-service queue, but
1859 * (2) this switch of bfqq to busy changes the scenario.
1861 if (bfqd
->in_service_queue
&&
1862 ((bfqq_wants_to_preempt
&&
1863 bfqq
->wr_coeff
>= bfqd
->in_service_queue
->wr_coeff
) ||
1864 bfq_bfqq_higher_class_or_weight(bfqq
, bfqd
->in_service_queue
) ||
1865 !bfq_better_to_idle(bfqd
->in_service_queue
)) &&
1866 next_queue_may_preempt(bfqd
))
1867 bfq_bfqq_expire(bfqd
, bfqd
->in_service_queue
,
1868 false, BFQQE_PREEMPTED
);
1871 static void bfq_reset_inject_limit(struct bfq_data
*bfqd
,
1872 struct bfq_queue
*bfqq
)
1874 /* invalidate baseline total service time */
1875 bfqq
->last_serv_time_ns
= 0;
1878 * Reset pointer in case we are waiting for
1879 * some request completion.
1881 bfqd
->waited_rq
= NULL
;
1884 * If bfqq has a short think time, then start by setting the
1885 * inject limit to 0 prudentially, because the service time of
1886 * an injected I/O request may be higher than the think time
1887 * of bfqq, and therefore, if one request was injected when
1888 * bfqq remains empty, this injected request might delay the
1889 * service of the next I/O request for bfqq significantly. In
1890 * case bfqq can actually tolerate some injection, then the
1891 * adaptive update will however raise the limit soon. This
1892 * lucky circumstance holds exactly because bfqq has a short
1893 * think time, and thus, after remaining empty, is likely to
1894 * get new I/O enqueued---and then completed---before being
1895 * expired. This is the very pattern that gives the
1896 * limit-update algorithm the chance to measure the effect of
1897 * injection on request service times, and then to update the
1898 * limit accordingly.
1900 * However, in the following special case, the inject limit is
1901 * left to 1 even if the think time is short: bfqq's I/O is
1902 * synchronized with that of some other queue, i.e., bfqq may
1903 * receive new I/O only after the I/O of the other queue is
1904 * completed. Keeping the inject limit to 1 allows the
1905 * blocking I/O to be served while bfqq is in service. And
1906 * this is very convenient both for bfqq and for overall
1907 * throughput, as explained in detail in the comments in
1908 * bfq_update_has_short_ttime().
1910 * On the opposite end, if bfqq has a long think time, then
1911 * start directly by 1, because:
1912 * a) on the bright side, keeping at most one request in
1913 * service in the drive is unlikely to cause any harm to the
1914 * latency of bfqq's requests, as the service time of a single
1915 * request is likely to be lower than the think time of bfqq;
1916 * b) on the downside, after becoming empty, bfqq is likely to
1917 * expire before getting its next request. With this request
1918 * arrival pattern, it is very hard to sample total service
1919 * times and update the inject limit accordingly (see comments
1920 * on bfq_update_inject_limit()). So the limit is likely to be
1921 * never, or at least seldom, updated. As a consequence, by
1922 * setting the limit to 1, we avoid that no injection ever
1923 * occurs with bfqq. On the downside, this proactive step
1924 * further reduces chances to actually compute the baseline
1925 * total service time. Thus it reduces chances to execute the
1926 * limit-update algorithm and possibly raise the limit to more
1929 if (bfq_bfqq_has_short_ttime(bfqq
))
1930 bfqq
->inject_limit
= 0;
1932 bfqq
->inject_limit
= 1;
1934 bfqq
->decrease_time_jif
= jiffies
;
1937 static void bfq_update_io_intensity(struct bfq_queue
*bfqq
, u64 now_ns
)
1939 u64 tot_io_time
= now_ns
- bfqq
->io_start_time
;
1941 if (RB_EMPTY_ROOT(&bfqq
->sort_list
) && bfqq
->dispatched
== 0)
1942 bfqq
->tot_idle_time
+=
1943 now_ns
- bfqq
->ttime
.last_end_request
;
1945 if (unlikely(bfq_bfqq_just_created(bfqq
)))
1949 * Must be busy for at least about 80% of the time to be
1950 * considered I/O bound.
1952 if (bfqq
->tot_idle_time
* 5 > tot_io_time
)
1953 bfq_clear_bfqq_IO_bound(bfqq
);
1955 bfq_mark_bfqq_IO_bound(bfqq
);
1958 * Keep an observation window of at most 200 ms in the past
1961 if (tot_io_time
> 200 * NSEC_PER_MSEC
) {
1962 bfqq
->io_start_time
= now_ns
- (tot_io_time
>>1);
1963 bfqq
->tot_idle_time
>>= 1;
1968 * Detect whether bfqq's I/O seems synchronized with that of some
1969 * other queue, i.e., whether bfqq, after remaining empty, happens to
1970 * receive new I/O only right after some I/O request of the other
1971 * queue has been completed. We call waker queue the other queue, and
1972 * we assume, for simplicity, that bfqq may have at most one waker
1975 * A remarkable throughput boost can be reached by unconditionally
1976 * injecting the I/O of the waker queue, every time a new
1977 * bfq_dispatch_request happens to be invoked while I/O is being
1978 * plugged for bfqq. In addition to boosting throughput, this
1979 * unblocks bfqq's I/O, thereby improving bandwidth and latency for
1980 * bfqq. Note that these same results may be achieved with the general
1981 * injection mechanism, but less effectively. For details on this
1982 * aspect, see the comments on the choice of the queue for injection
1983 * in bfq_select_queue().
1985 * Turning back to the detection of a waker queue, a queue Q is deemed
1986 * as a waker queue for bfqq if, for three consecutive times, bfqq
1987 * happens to become non empty right after a request of Q has been
1988 * completed. In this respect, even if bfqq is empty, we do not check
1989 * for a waker if it still has some in-flight I/O. In fact, in this
1990 * case bfqq is actually still being served by the drive, and may
1991 * receive new I/O on the completion of some of the in-flight
1992 * requests. In particular, on the first time, Q is tentatively set as
1993 * a candidate waker queue, while on the third consecutive time that Q
1994 * is detected, the field waker_bfqq is set to Q, to confirm that Q is
1995 * a waker queue for bfqq. These detection steps are performed only if
1996 * bfqq has a long think time, so as to make it more likely that
1997 * bfqq's I/O is actually being blocked by a synchronization. This
1998 * last filter, plus the above three-times requirement, make false
1999 * positives less likely.
2003 * The sooner a waker queue is detected, the sooner throughput can be
2004 * boosted by injecting I/O from the waker queue. Fortunately,
2005 * detection is likely to be actually fast, for the following
2006 * reasons. While blocked by synchronization, bfqq has a long think
2007 * time. This implies that bfqq's inject limit is at least equal to 1
2008 * (see the comments in bfq_update_inject_limit()). So, thanks to
2009 * injection, the waker queue is likely to be served during the very
2010 * first I/O-plugging time interval for bfqq. This triggers the first
2011 * step of the detection mechanism. Thanks again to injection, the
2012 * candidate waker queue is then likely to be confirmed no later than
2013 * during the next I/O-plugging interval for bfqq.
2017 * On queue merging all waker information is lost.
2019 static void bfq_check_waker(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
2022 if (!bfqd
->last_completed_rq_bfqq
||
2023 bfqd
->last_completed_rq_bfqq
== bfqq
||
2024 bfq_bfqq_has_short_ttime(bfqq
) ||
2025 bfqq
->dispatched
> 0 ||
2026 now_ns
- bfqd
->last_completion
>= 4 * NSEC_PER_MSEC
||
2027 bfqd
->last_completed_rq_bfqq
== bfqq
->waker_bfqq
)
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 (RB_EMPTY_ROOT(&bfqq
->sort_list
) && bfq_bfqq_sync(bfqq
)) {
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
);
2341 bfqd
->bio_bfqq
= bic_to_bfqq(bic
, op_is_sync(bio
->bi_opf
));
2343 bfqd
->bio_bfqq
= NULL
;
2344 bfqd
->bio_bic
= bic
;
2346 ret
= blk_mq_sched_try_merge(q
, bio
, nr_segs
, &free
);
2348 spin_unlock_irq(&bfqd
->lock
);
2350 blk_mq_free_request(free
);
2355 static int bfq_request_merge(struct request_queue
*q
, struct request
**req
,
2358 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
2359 struct request
*__rq
;
2361 __rq
= bfq_find_rq_fmerge(bfqd
, bio
, q
);
2362 if (__rq
&& elv_bio_merge_ok(__rq
, bio
)) {
2364 return ELEVATOR_FRONT_MERGE
;
2367 return ELEVATOR_NO_MERGE
;
2370 static struct bfq_queue
*bfq_init_rq(struct request
*rq
);
2372 static void bfq_request_merged(struct request_queue
*q
, struct request
*req
,
2373 enum elv_merge type
)
2375 if (type
== ELEVATOR_FRONT_MERGE
&&
2376 rb_prev(&req
->rb_node
) &&
2378 blk_rq_pos(container_of(rb_prev(&req
->rb_node
),
2379 struct request
, rb_node
))) {
2380 struct bfq_queue
*bfqq
= bfq_init_rq(req
);
2381 struct bfq_data
*bfqd
;
2382 struct request
*prev
, *next_rq
;
2389 /* Reposition request in its sort_list */
2390 elv_rb_del(&bfqq
->sort_list
, req
);
2391 elv_rb_add(&bfqq
->sort_list
, req
);
2393 /* Choose next request to be served for bfqq */
2394 prev
= bfqq
->next_rq
;
2395 next_rq
= bfq_choose_req(bfqd
, bfqq
->next_rq
, req
,
2396 bfqd
->last_position
);
2397 bfqq
->next_rq
= next_rq
;
2399 * If next_rq changes, update both the queue's budget to
2400 * fit the new request and the queue's position in its
2403 if (prev
!= bfqq
->next_rq
) {
2404 bfq_updated_next_req(bfqd
, bfqq
);
2406 * See comments on bfq_pos_tree_add_move() for
2409 if (unlikely(!bfqd
->nonrot_with_queueing
))
2410 bfq_pos_tree_add_move(bfqd
, bfqq
);
2416 * This function is called to notify the scheduler that the requests
2417 * rq and 'next' have been merged, with 'next' going away. BFQ
2418 * exploits this hook to address the following issue: if 'next' has a
2419 * fifo_time lower that rq, then the fifo_time of rq must be set to
2420 * the value of 'next', to not forget the greater age of 'next'.
2422 * NOTE: in this function we assume that rq is in a bfq_queue, basing
2423 * on that rq is picked from the hash table q->elevator->hash, which,
2424 * in its turn, is filled only with I/O requests present in
2425 * bfq_queues, while BFQ is in use for the request queue q. In fact,
2426 * the function that fills this hash table (elv_rqhash_add) is called
2427 * only by bfq_insert_request.
2429 static void bfq_requests_merged(struct request_queue
*q
, struct request
*rq
,
2430 struct request
*next
)
2432 struct bfq_queue
*bfqq
= bfq_init_rq(rq
),
2433 *next_bfqq
= bfq_init_rq(next
);
2439 * If next and rq belong to the same bfq_queue and next is older
2440 * than rq, then reposition rq in the fifo (by substituting next
2441 * with rq). Otherwise, if next and rq belong to different
2442 * bfq_queues, never reposition rq: in fact, we would have to
2443 * reposition it with respect to next's position in its own fifo,
2444 * which would most certainly be too expensive with respect to
2447 if (bfqq
== next_bfqq
&&
2448 !list_empty(&rq
->queuelist
) && !list_empty(&next
->queuelist
) &&
2449 next
->fifo_time
< rq
->fifo_time
) {
2450 list_del_init(&rq
->queuelist
);
2451 list_replace_init(&next
->queuelist
, &rq
->queuelist
);
2452 rq
->fifo_time
= next
->fifo_time
;
2455 if (bfqq
->next_rq
== next
)
2458 bfqg_stats_update_io_merged(bfqq_group(bfqq
), next
->cmd_flags
);
2460 /* Merged request may be in the IO scheduler. Remove it. */
2461 if (!RB_EMPTY_NODE(&next
->rb_node
)) {
2462 bfq_remove_request(next
->q
, next
);
2464 bfqg_stats_update_io_remove(bfqq_group(next_bfqq
),
2469 /* Must be called with bfqq != NULL */
2470 static void bfq_bfqq_end_wr(struct bfq_queue
*bfqq
)
2473 * If bfqq has been enjoying interactive weight-raising, then
2474 * reset soft_rt_next_start. We do it for the following
2475 * reason. bfqq may have been conveying the I/O needed to load
2476 * a soft real-time application. Such an application actually
2477 * exhibits a soft real-time I/O pattern after it finishes
2478 * loading, and finally starts doing its job. But, if bfqq has
2479 * been receiving a lot of bandwidth so far (likely to happen
2480 * on a fast device), then soft_rt_next_start now contains a
2481 * high value that. So, without this reset, bfqq would be
2482 * prevented from being possibly considered as soft_rt for a
2486 if (bfqq
->wr_cur_max_time
!=
2487 bfqq
->bfqd
->bfq_wr_rt_max_time
)
2488 bfqq
->soft_rt_next_start
= jiffies
;
2490 if (bfq_bfqq_busy(bfqq
))
2491 bfqq
->bfqd
->wr_busy_queues
--;
2493 bfqq
->wr_cur_max_time
= 0;
2494 bfqq
->last_wr_start_finish
= jiffies
;
2496 * Trigger a weight change on the next invocation of
2497 * __bfq_entity_update_weight_prio.
2499 bfqq
->entity
.prio_changed
= 1;
2502 void bfq_end_wr_async_queues(struct bfq_data
*bfqd
,
2503 struct bfq_group
*bfqg
)
2507 for (i
= 0; i
< 2; i
++)
2508 for (j
= 0; j
< IOPRIO_BE_NR
; j
++)
2509 if (bfqg
->async_bfqq
[i
][j
])
2510 bfq_bfqq_end_wr(bfqg
->async_bfqq
[i
][j
]);
2511 if (bfqg
->async_idle_bfqq
)
2512 bfq_bfqq_end_wr(bfqg
->async_idle_bfqq
);
2515 static void bfq_end_wr(struct bfq_data
*bfqd
)
2517 struct bfq_queue
*bfqq
;
2519 spin_lock_irq(&bfqd
->lock
);
2521 list_for_each_entry(bfqq
, &bfqd
->active_list
, bfqq_list
)
2522 bfq_bfqq_end_wr(bfqq
);
2523 list_for_each_entry(bfqq
, &bfqd
->idle_list
, bfqq_list
)
2524 bfq_bfqq_end_wr(bfqq
);
2525 bfq_end_wr_async(bfqd
);
2527 spin_unlock_irq(&bfqd
->lock
);
2530 static sector_t
bfq_io_struct_pos(void *io_struct
, bool request
)
2533 return blk_rq_pos(io_struct
);
2535 return ((struct bio
*)io_struct
)->bi_iter
.bi_sector
;
2538 static int bfq_rq_close_to_sector(void *io_struct
, bool request
,
2541 return abs(bfq_io_struct_pos(io_struct
, request
) - sector
) <=
2545 static struct bfq_queue
*bfqq_find_close(struct bfq_data
*bfqd
,
2546 struct bfq_queue
*bfqq
,
2549 struct rb_root
*root
= &bfq_bfqq_to_bfqg(bfqq
)->rq_pos_tree
;
2550 struct rb_node
*parent
, *node
;
2551 struct bfq_queue
*__bfqq
;
2553 if (RB_EMPTY_ROOT(root
))
2557 * First, if we find a request starting at the end of the last
2558 * request, choose it.
2560 __bfqq
= bfq_rq_pos_tree_lookup(bfqd
, root
, sector
, &parent
, NULL
);
2565 * If the exact sector wasn't found, the parent of the NULL leaf
2566 * will contain the closest sector (rq_pos_tree sorted by
2567 * next_request position).
2569 __bfqq
= rb_entry(parent
, struct bfq_queue
, pos_node
);
2570 if (bfq_rq_close_to_sector(__bfqq
->next_rq
, true, sector
))
2573 if (blk_rq_pos(__bfqq
->next_rq
) < sector
)
2574 node
= rb_next(&__bfqq
->pos_node
);
2576 node
= rb_prev(&__bfqq
->pos_node
);
2580 __bfqq
= rb_entry(node
, struct bfq_queue
, pos_node
);
2581 if (bfq_rq_close_to_sector(__bfqq
->next_rq
, true, sector
))
2587 static struct bfq_queue
*bfq_find_close_cooperator(struct bfq_data
*bfqd
,
2588 struct bfq_queue
*cur_bfqq
,
2591 struct bfq_queue
*bfqq
;
2594 * We shall notice if some of the queues are cooperating,
2595 * e.g., working closely on the same area of the device. In
2596 * that case, we can group them together and: 1) don't waste
2597 * time idling, and 2) serve the union of their requests in
2598 * the best possible order for throughput.
2600 bfqq
= bfqq_find_close(bfqd
, cur_bfqq
, sector
);
2601 if (!bfqq
|| bfqq
== cur_bfqq
)
2607 static struct bfq_queue
*
2608 bfq_setup_merge(struct bfq_queue
*bfqq
, struct bfq_queue
*new_bfqq
)
2610 int process_refs
, new_process_refs
;
2611 struct bfq_queue
*__bfqq
;
2614 * If there are no process references on the new_bfqq, then it is
2615 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2616 * may have dropped their last reference (not just their last process
2619 if (!bfqq_process_refs(new_bfqq
))
2622 /* Avoid a circular list and skip interim queue merges. */
2623 while ((__bfqq
= new_bfqq
->new_bfqq
)) {
2629 process_refs
= bfqq_process_refs(bfqq
);
2630 new_process_refs
= bfqq_process_refs(new_bfqq
);
2632 * If the process for the bfqq has gone away, there is no
2633 * sense in merging the queues.
2635 if (process_refs
== 0 || new_process_refs
== 0)
2638 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "scheduling merge with queue %d",
2642 * Merging is just a redirection: the requests of the process
2643 * owning one of the two queues are redirected to the other queue.
2644 * The latter queue, in its turn, is set as shared if this is the
2645 * first time that the requests of some process are redirected to
2648 * We redirect bfqq to new_bfqq and not the opposite, because
2649 * we are in the context of the process owning bfqq, thus we
2650 * have the io_cq of this process. So we can immediately
2651 * configure this io_cq to redirect the requests of the
2652 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2653 * not available any more (new_bfqq->bic == NULL).
2655 * Anyway, even in case new_bfqq coincides with the in-service
2656 * queue, redirecting requests the in-service queue is the
2657 * best option, as we feed the in-service queue with new
2658 * requests close to the last request served and, by doing so,
2659 * are likely to increase the throughput.
2661 bfqq
->new_bfqq
= new_bfqq
;
2662 new_bfqq
->ref
+= process_refs
;
2666 static bool bfq_may_be_close_cooperator(struct bfq_queue
*bfqq
,
2667 struct bfq_queue
*new_bfqq
)
2669 if (bfq_too_late_for_merging(new_bfqq
))
2672 if (bfq_class_idle(bfqq
) || bfq_class_idle(new_bfqq
) ||
2673 (bfqq
->ioprio_class
!= new_bfqq
->ioprio_class
))
2677 * If either of the queues has already been detected as seeky,
2678 * then merging it with the other queue is unlikely to lead to
2681 if (BFQQ_SEEKY(bfqq
) || BFQQ_SEEKY(new_bfqq
))
2685 * Interleaved I/O is known to be done by (some) applications
2686 * only for reads, so it does not make sense to merge async
2689 if (!bfq_bfqq_sync(bfqq
) || !bfq_bfqq_sync(new_bfqq
))
2695 static bool idling_boosts_thr_without_issues(struct bfq_data
*bfqd
,
2696 struct bfq_queue
*bfqq
);
2699 * Attempt to schedule a merge of bfqq with the currently in-service
2700 * queue or with a close queue among the scheduled queues. Return
2701 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2702 * structure otherwise.
2704 * The OOM queue is not allowed to participate to cooperation: in fact, since
2705 * the requests temporarily redirected to the OOM queue could be redirected
2706 * again to dedicated queues at any time, the state needed to correctly
2707 * handle merging with the OOM queue would be quite complex and expensive
2708 * to maintain. Besides, in such a critical condition as an out of memory,
2709 * the benefits of queue merging may be little relevant, or even negligible.
2711 * WARNING: queue merging may impair fairness among non-weight raised
2712 * queues, for at least two reasons: 1) the original weight of a
2713 * merged queue may change during the merged state, 2) even being the
2714 * weight the same, a merged queue may be bloated with many more
2715 * requests than the ones produced by its originally-associated
2718 static struct bfq_queue
*
2719 bfq_setup_cooperator(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
2720 void *io_struct
, bool request
, struct bfq_io_cq
*bic
)
2722 struct bfq_queue
*in_service_bfqq
, *new_bfqq
;
2725 * Check delayed stable merge for rotational or non-queueing
2726 * devs. For this branch to be executed, bfqq must not be
2727 * currently merged with some other queue (i.e., bfqq->bic
2728 * must be non null). If we considered also merged queues,
2729 * then we should also check whether bfqq has already been
2730 * merged with bic->stable_merge_bfqq. But this would be
2731 * costly and complicated.
2733 if (unlikely(!bfqd
->nonrot_with_queueing
)) {
2735 * Make sure also that bfqq is sync, because
2736 * bic->stable_merge_bfqq may point to some queue (for
2737 * stable merging) also if bic is associated with a
2738 * sync queue, but this bfqq is async
2740 if (bfq_bfqq_sync(bfqq
) && bic
->stable_merge_bfqq
&&
2741 !bfq_bfqq_just_created(bfqq
) &&
2742 time_is_before_jiffies(bfqq
->split_time
+
2743 msecs_to_jiffies(bfq_late_stable_merging
)) &&
2744 time_is_before_jiffies(bfqq
->creation_time
+
2745 msecs_to_jiffies(bfq_late_stable_merging
))) {
2746 struct bfq_queue
*stable_merge_bfqq
=
2747 bic
->stable_merge_bfqq
;
2748 int proc_ref
= min(bfqq_process_refs(bfqq
),
2749 bfqq_process_refs(stable_merge_bfqq
));
2751 /* deschedule stable merge, because done or aborted here */
2752 bfq_put_stable_ref(stable_merge_bfqq
);
2754 bic
->stable_merge_bfqq
= NULL
;
2756 if (!idling_boosts_thr_without_issues(bfqd
, bfqq
) &&
2758 /* next function will take at least one ref */
2759 struct bfq_queue
*new_bfqq
=
2760 bfq_setup_merge(bfqq
, stable_merge_bfqq
);
2762 bic
->stably_merged
= true;
2763 if (new_bfqq
&& new_bfqq
->bic
)
2764 new_bfqq
->bic
->stably_merged
= true;
2772 * Do not perform queue merging if the device is non
2773 * rotational and performs internal queueing. In fact, such a
2774 * device reaches a high speed through internal parallelism
2775 * and pipelining. This means that, to reach a high
2776 * throughput, it must have many requests enqueued at the same
2777 * time. But, in this configuration, the internal scheduling
2778 * algorithm of the device does exactly the job of queue
2779 * merging: it reorders requests so as to obtain as much as
2780 * possible a sequential I/O pattern. As a consequence, with
2781 * the workload generated by processes doing interleaved I/O,
2782 * the throughput reached by the device is likely to be the
2783 * same, with and without queue merging.
2785 * Disabling merging also provides a remarkable benefit in
2786 * terms of throughput. Merging tends to make many workloads
2787 * artificially more uneven, because of shared queues
2788 * remaining non empty for incomparably more time than
2789 * non-merged queues. This may accentuate workload
2790 * asymmetries. For example, if one of the queues in a set of
2791 * merged queues has a higher weight than a normal queue, then
2792 * the shared queue may inherit such a high weight and, by
2793 * staying almost always active, may force BFQ to perform I/O
2794 * plugging most of the time. This evidently makes it harder
2795 * for BFQ to let the device reach a high throughput.
2797 * Finally, the likely() macro below is not used because one
2798 * of the two branches is more likely than the other, but to
2799 * have the code path after the following if() executed as
2800 * fast as possible for the case of a non rotational device
2801 * with queueing. We want it because this is the fastest kind
2802 * of device. On the opposite end, the likely() may lengthen
2803 * the execution time of BFQ for the case of slower devices
2804 * (rotational or at least without queueing). But in this case
2805 * the execution time of BFQ matters very little, if not at
2808 if (likely(bfqd
->nonrot_with_queueing
))
2812 * Prevent bfqq from being merged if it has been created too
2813 * long ago. The idea is that true cooperating processes, and
2814 * thus their associated bfq_queues, are supposed to be
2815 * created shortly after each other. This is the case, e.g.,
2816 * for KVM/QEMU and dump I/O threads. Basing on this
2817 * assumption, the following filtering greatly reduces the
2818 * probability that two non-cooperating processes, which just
2819 * happen to do close I/O for some short time interval, have
2820 * their queues merged by mistake.
2822 if (bfq_too_late_for_merging(bfqq
))
2826 return bfqq
->new_bfqq
;
2828 if (!io_struct
|| unlikely(bfqq
== &bfqd
->oom_bfqq
))
2831 /* If there is only one backlogged queue, don't search. */
2832 if (bfq_tot_busy_queues(bfqd
) == 1)
2835 in_service_bfqq
= bfqd
->in_service_queue
;
2837 if (in_service_bfqq
&& in_service_bfqq
!= bfqq
&&
2838 likely(in_service_bfqq
!= &bfqd
->oom_bfqq
) &&
2839 bfq_rq_close_to_sector(io_struct
, request
,
2840 bfqd
->in_serv_last_pos
) &&
2841 bfqq
->entity
.parent
== in_service_bfqq
->entity
.parent
&&
2842 bfq_may_be_close_cooperator(bfqq
, in_service_bfqq
)) {
2843 new_bfqq
= bfq_setup_merge(bfqq
, in_service_bfqq
);
2848 * Check whether there is a cooperator among currently scheduled
2849 * queues. The only thing we need is that the bio/request is not
2850 * NULL, as we need it to establish whether a cooperator exists.
2852 new_bfqq
= bfq_find_close_cooperator(bfqd
, bfqq
,
2853 bfq_io_struct_pos(io_struct
, request
));
2855 if (new_bfqq
&& likely(new_bfqq
!= &bfqd
->oom_bfqq
) &&
2856 bfq_may_be_close_cooperator(bfqq
, new_bfqq
))
2857 return bfq_setup_merge(bfqq
, new_bfqq
);
2862 static void bfq_bfqq_save_state(struct bfq_queue
*bfqq
)
2864 struct bfq_io_cq
*bic
= bfqq
->bic
;
2867 * If !bfqq->bic, the queue is already shared or its requests
2868 * have already been redirected to a shared queue; both idle window
2869 * and weight raising state have already been saved. Do nothing.
2874 bic
->saved_last_serv_time_ns
= bfqq
->last_serv_time_ns
;
2875 bic
->saved_inject_limit
= bfqq
->inject_limit
;
2876 bic
->saved_decrease_time_jif
= bfqq
->decrease_time_jif
;
2878 bic
->saved_weight
= bfqq
->entity
.orig_weight
;
2879 bic
->saved_ttime
= bfqq
->ttime
;
2880 bic
->saved_has_short_ttime
= bfq_bfqq_has_short_ttime(bfqq
);
2881 bic
->saved_IO_bound
= bfq_bfqq_IO_bound(bfqq
);
2882 bic
->saved_io_start_time
= bfqq
->io_start_time
;
2883 bic
->saved_tot_idle_time
= bfqq
->tot_idle_time
;
2884 bic
->saved_in_large_burst
= bfq_bfqq_in_large_burst(bfqq
);
2885 bic
->was_in_burst_list
= !hlist_unhashed(&bfqq
->burst_list_node
);
2886 if (unlikely(bfq_bfqq_just_created(bfqq
) &&
2887 !bfq_bfqq_in_large_burst(bfqq
) &&
2888 bfqq
->bfqd
->low_latency
)) {
2890 * bfqq being merged right after being created: bfqq
2891 * would have deserved interactive weight raising, but
2892 * did not make it to be set in a weight-raised state,
2893 * because of this early merge. Store directly the
2894 * weight-raising state that would have been assigned
2895 * to bfqq, so that to avoid that bfqq unjustly fails
2896 * to enjoy weight raising if split soon.
2898 bic
->saved_wr_coeff
= bfqq
->bfqd
->bfq_wr_coeff
;
2899 bic
->saved_wr_start_at_switch_to_srt
= bfq_smallest_from_now();
2900 bic
->saved_wr_cur_max_time
= bfq_wr_duration(bfqq
->bfqd
);
2901 bic
->saved_last_wr_start_finish
= jiffies
;
2903 bic
->saved_wr_coeff
= bfqq
->wr_coeff
;
2904 bic
->saved_wr_start_at_switch_to_srt
=
2905 bfqq
->wr_start_at_switch_to_srt
;
2906 bic
->saved_service_from_wr
= bfqq
->service_from_wr
;
2907 bic
->saved_last_wr_start_finish
= bfqq
->last_wr_start_finish
;
2908 bic
->saved_wr_cur_max_time
= bfqq
->wr_cur_max_time
;
2914 bfq_reassign_last_bfqq(struct bfq_queue
*cur_bfqq
, struct bfq_queue
*new_bfqq
)
2916 if (cur_bfqq
->entity
.parent
&&
2917 cur_bfqq
->entity
.parent
->last_bfqq_created
== cur_bfqq
)
2918 cur_bfqq
->entity
.parent
->last_bfqq_created
= new_bfqq
;
2919 else if (cur_bfqq
->bfqd
&& cur_bfqq
->bfqd
->last_bfqq_created
== cur_bfqq
)
2920 cur_bfqq
->bfqd
->last_bfqq_created
= new_bfqq
;
2923 void bfq_release_process_ref(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
2926 * To prevent bfqq's service guarantees from being violated,
2927 * bfqq may be left busy, i.e., queued for service, even if
2928 * empty (see comments in __bfq_bfqq_expire() for
2929 * details). But, if no process will send requests to bfqq any
2930 * longer, then there is no point in keeping bfqq queued for
2931 * service. In addition, keeping bfqq queued for service, but
2932 * with no process ref any longer, may have caused bfqq to be
2933 * freed when dequeued from service. But this is assumed to
2936 if (bfq_bfqq_busy(bfqq
) && RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
2937 bfqq
!= bfqd
->in_service_queue
)
2938 bfq_del_bfqq_busy(bfqd
, bfqq
, false);
2940 bfq_reassign_last_bfqq(bfqq
, NULL
);
2942 bfq_put_queue(bfqq
);
2946 bfq_merge_bfqqs(struct bfq_data
*bfqd
, struct bfq_io_cq
*bic
,
2947 struct bfq_queue
*bfqq
, struct bfq_queue
*new_bfqq
)
2949 bfq_log_bfqq(bfqd
, bfqq
, "merging with queue %lu",
2950 (unsigned long)new_bfqq
->pid
);
2951 /* Save weight raising and idle window of the merged queues */
2952 bfq_bfqq_save_state(bfqq
);
2953 bfq_bfqq_save_state(new_bfqq
);
2954 if (bfq_bfqq_IO_bound(bfqq
))
2955 bfq_mark_bfqq_IO_bound(new_bfqq
);
2956 bfq_clear_bfqq_IO_bound(bfqq
);
2959 * The processes associated with bfqq are cooperators of the
2960 * processes associated with new_bfqq. So, if bfqq has a
2961 * waker, then assume that all these processes will be happy
2962 * to let bfqq's waker freely inject I/O when they have no
2965 if (bfqq
->waker_bfqq
&& !new_bfqq
->waker_bfqq
&&
2966 bfqq
->waker_bfqq
!= new_bfqq
) {
2967 new_bfqq
->waker_bfqq
= bfqq
->waker_bfqq
;
2968 new_bfqq
->tentative_waker_bfqq
= NULL
;
2971 * If the waker queue disappears, then
2972 * new_bfqq->waker_bfqq must be reset. So insert
2973 * new_bfqq into the woken_list of the waker. See
2974 * bfq_check_waker for details.
2976 hlist_add_head(&new_bfqq
->woken_list_node
,
2977 &new_bfqq
->waker_bfqq
->woken_list
);
2982 * If bfqq is weight-raised, then let new_bfqq inherit
2983 * weight-raising. To reduce false positives, neglect the case
2984 * where bfqq has just been created, but has not yet made it
2985 * to be weight-raised (which may happen because EQM may merge
2986 * bfqq even before bfq_add_request is executed for the first
2987 * time for bfqq). Handling this case would however be very
2988 * easy, thanks to the flag just_created.
2990 if (new_bfqq
->wr_coeff
== 1 && bfqq
->wr_coeff
> 1) {
2991 new_bfqq
->wr_coeff
= bfqq
->wr_coeff
;
2992 new_bfqq
->wr_cur_max_time
= bfqq
->wr_cur_max_time
;
2993 new_bfqq
->last_wr_start_finish
= bfqq
->last_wr_start_finish
;
2994 new_bfqq
->wr_start_at_switch_to_srt
=
2995 bfqq
->wr_start_at_switch_to_srt
;
2996 if (bfq_bfqq_busy(new_bfqq
))
2997 bfqd
->wr_busy_queues
++;
2998 new_bfqq
->entity
.prio_changed
= 1;
3001 if (bfqq
->wr_coeff
> 1) { /* bfqq has given its wr to new_bfqq */
3003 bfqq
->entity
.prio_changed
= 1;
3004 if (bfq_bfqq_busy(bfqq
))
3005 bfqd
->wr_busy_queues
--;
3008 bfq_log_bfqq(bfqd
, new_bfqq
, "merge_bfqqs: wr_busy %d",
3009 bfqd
->wr_busy_queues
);
3012 * Merge queues (that is, let bic redirect its requests to new_bfqq)
3014 bic_set_bfqq(bic
, new_bfqq
, 1);
3015 bfq_mark_bfqq_coop(new_bfqq
);
3017 * new_bfqq now belongs to at least two bics (it is a shared queue):
3018 * set new_bfqq->bic to NULL. bfqq either:
3019 * - does not belong to any bic any more, and hence bfqq->bic must
3020 * be set to NULL, or
3021 * - is a queue whose owning bics have already been redirected to a
3022 * different queue, hence the queue is destined to not belong to
3023 * any bic soon and bfqq->bic is already NULL (therefore the next
3024 * assignment causes no harm).
3026 new_bfqq
->bic
= NULL
;
3028 * If the queue is shared, the pid is the pid of one of the associated
3029 * processes. Which pid depends on the exact sequence of merge events
3030 * the queue underwent. So printing such a pid is useless and confusing
3031 * because it reports a random pid between those of the associated
3033 * We mark such a queue with a pid -1, and then print SHARED instead of
3034 * a pid in logging messages.
3039 bfq_reassign_last_bfqq(bfqq
, new_bfqq
);
3041 bfq_release_process_ref(bfqd
, bfqq
);
3044 static bool bfq_allow_bio_merge(struct request_queue
*q
, struct request
*rq
,
3047 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
3048 bool is_sync
= op_is_sync(bio
->bi_opf
);
3049 struct bfq_queue
*bfqq
= bfqd
->bio_bfqq
, *new_bfqq
;
3052 * Disallow merge of a sync bio into an async request.
3054 if (is_sync
&& !rq_is_sync(rq
))
3058 * Lookup the bfqq that this bio will be queued with. Allow
3059 * merge only if rq is queued there.
3065 * We take advantage of this function to perform an early merge
3066 * of the queues of possible cooperating processes.
3068 new_bfqq
= bfq_setup_cooperator(bfqd
, bfqq
, bio
, false, bfqd
->bio_bic
);
3071 * bic still points to bfqq, then it has not yet been
3072 * redirected to some other bfq_queue, and a queue
3073 * merge between bfqq and new_bfqq can be safely
3074 * fulfilled, i.e., bic can be redirected to new_bfqq
3075 * and bfqq can be put.
3077 bfq_merge_bfqqs(bfqd
, bfqd
->bio_bic
, bfqq
,
3080 * If we get here, bio will be queued into new_queue,
3081 * so use new_bfqq to decide whether bio and rq can be
3087 * Change also bqfd->bio_bfqq, as
3088 * bfqd->bio_bic now points to new_bfqq, and
3089 * this function may be invoked again (and then may
3090 * use again bqfd->bio_bfqq).
3092 bfqd
->bio_bfqq
= bfqq
;
3095 return bfqq
== RQ_BFQQ(rq
);
3099 * Set the maximum time for the in-service queue to consume its
3100 * budget. This prevents seeky processes from lowering the throughput.
3101 * In practice, a time-slice service scheme is used with seeky
3104 static void bfq_set_budget_timeout(struct bfq_data
*bfqd
,
3105 struct bfq_queue
*bfqq
)
3107 unsigned int timeout_coeff
;
3109 if (bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
)
3112 timeout_coeff
= bfqq
->entity
.weight
/ bfqq
->entity
.orig_weight
;
3114 bfqd
->last_budget_start
= ktime_get();
3116 bfqq
->budget_timeout
= jiffies
+
3117 bfqd
->bfq_timeout
* timeout_coeff
;
3120 static void __bfq_set_in_service_queue(struct bfq_data
*bfqd
,
3121 struct bfq_queue
*bfqq
)
3124 bfq_clear_bfqq_fifo_expire(bfqq
);
3126 bfqd
->budgets_assigned
= (bfqd
->budgets_assigned
* 7 + 256) / 8;
3128 if (time_is_before_jiffies(bfqq
->last_wr_start_finish
) &&
3129 bfqq
->wr_coeff
> 1 &&
3130 bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
&&
3131 time_is_before_jiffies(bfqq
->budget_timeout
)) {
3133 * For soft real-time queues, move the start
3134 * of the weight-raising period forward by the
3135 * time the queue has not received any
3136 * service. Otherwise, a relatively long
3137 * service delay is likely to cause the
3138 * weight-raising period of the queue to end,
3139 * because of the short duration of the
3140 * weight-raising period of a soft real-time
3141 * queue. It is worth noting that this move
3142 * is not so dangerous for the other queues,
3143 * because soft real-time queues are not
3146 * To not add a further variable, we use the
3147 * overloaded field budget_timeout to
3148 * determine for how long the queue has not
3149 * received service, i.e., how much time has
3150 * elapsed since the queue expired. However,
3151 * this is a little imprecise, because
3152 * budget_timeout is set to jiffies if bfqq
3153 * not only expires, but also remains with no
3156 if (time_after(bfqq
->budget_timeout
,
3157 bfqq
->last_wr_start_finish
))
3158 bfqq
->last_wr_start_finish
+=
3159 jiffies
- bfqq
->budget_timeout
;
3161 bfqq
->last_wr_start_finish
= jiffies
;
3164 bfq_set_budget_timeout(bfqd
, bfqq
);
3165 bfq_log_bfqq(bfqd
, bfqq
,
3166 "set_in_service_queue, cur-budget = %d",
3167 bfqq
->entity
.budget
);
3170 bfqd
->in_service_queue
= bfqq
;
3171 bfqd
->in_serv_last_pos
= 0;
3175 * Get and set a new queue for service.
3177 static struct bfq_queue
*bfq_set_in_service_queue(struct bfq_data
*bfqd
)
3179 struct bfq_queue
*bfqq
= bfq_get_next_queue(bfqd
);
3181 __bfq_set_in_service_queue(bfqd
, bfqq
);
3185 static void bfq_arm_slice_timer(struct bfq_data
*bfqd
)
3187 struct bfq_queue
*bfqq
= bfqd
->in_service_queue
;
3190 bfq_mark_bfqq_wait_request(bfqq
);
3193 * We don't want to idle for seeks, but we do want to allow
3194 * fair distribution of slice time for a process doing back-to-back
3195 * seeks. So allow a little bit of time for him to submit a new rq.
3197 sl
= bfqd
->bfq_slice_idle
;
3199 * Unless the queue is being weight-raised or the scenario is
3200 * asymmetric, grant only minimum idle time if the queue
3201 * is seeky. A long idling is preserved for a weight-raised
3202 * queue, or, more in general, in an asymmetric scenario,
3203 * because a long idling is needed for guaranteeing to a queue
3204 * its reserved share of the throughput (in particular, it is
3205 * needed if the queue has a higher weight than some other
3208 if (BFQQ_SEEKY(bfqq
) && bfqq
->wr_coeff
== 1 &&
3209 !bfq_asymmetric_scenario(bfqd
, bfqq
))
3210 sl
= min_t(u64
, sl
, BFQ_MIN_TT
);
3211 else if (bfqq
->wr_coeff
> 1)
3212 sl
= max_t(u32
, sl
, 20ULL * NSEC_PER_MSEC
);
3214 bfqd
->last_idling_start
= ktime_get();
3215 bfqd
->last_idling_start_jiffies
= jiffies
;
3217 hrtimer_start(&bfqd
->idle_slice_timer
, ns_to_ktime(sl
),
3219 bfqg_stats_set_start_idle_time(bfqq_group(bfqq
));
3223 * In autotuning mode, max_budget is dynamically recomputed as the
3224 * amount of sectors transferred in timeout at the estimated peak
3225 * rate. This enables BFQ to utilize a full timeslice with a full
3226 * budget, even if the in-service queue is served at peak rate. And
3227 * this maximises throughput with sequential workloads.
3229 static unsigned long bfq_calc_max_budget(struct bfq_data
*bfqd
)
3231 return (u64
)bfqd
->peak_rate
* USEC_PER_MSEC
*
3232 jiffies_to_msecs(bfqd
->bfq_timeout
)>>BFQ_RATE_SHIFT
;
3236 * Update parameters related to throughput and responsiveness, as a
3237 * function of the estimated peak rate. See comments on
3238 * bfq_calc_max_budget(), and on the ref_wr_duration array.
3240 static void update_thr_responsiveness_params(struct bfq_data
*bfqd
)
3242 if (bfqd
->bfq_user_max_budget
== 0) {
3243 bfqd
->bfq_max_budget
=
3244 bfq_calc_max_budget(bfqd
);
3245 bfq_log(bfqd
, "new max_budget = %d", bfqd
->bfq_max_budget
);
3249 static void bfq_reset_rate_computation(struct bfq_data
*bfqd
,
3252 if (rq
!= NULL
) { /* new rq dispatch now, reset accordingly */
3253 bfqd
->last_dispatch
= bfqd
->first_dispatch
= ktime_get_ns();
3254 bfqd
->peak_rate_samples
= 1;
3255 bfqd
->sequential_samples
= 0;
3256 bfqd
->tot_sectors_dispatched
= bfqd
->last_rq_max_size
=
3258 } else /* no new rq dispatched, just reset the number of samples */
3259 bfqd
->peak_rate_samples
= 0; /* full re-init on next disp. */
3262 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
3263 bfqd
->peak_rate_samples
, bfqd
->sequential_samples
,
3264 bfqd
->tot_sectors_dispatched
);
3267 static void bfq_update_rate_reset(struct bfq_data
*bfqd
, struct request
*rq
)
3269 u32 rate
, weight
, divisor
;
3272 * For the convergence property to hold (see comments on
3273 * bfq_update_peak_rate()) and for the assessment to be
3274 * reliable, a minimum number of samples must be present, and
3275 * a minimum amount of time must have elapsed. If not so, do
3276 * not compute new rate. Just reset parameters, to get ready
3277 * for a new evaluation attempt.
3279 if (bfqd
->peak_rate_samples
< BFQ_RATE_MIN_SAMPLES
||
3280 bfqd
->delta_from_first
< BFQ_RATE_MIN_INTERVAL
)
3281 goto reset_computation
;
3284 * If a new request completion has occurred after last
3285 * dispatch, then, to approximate the rate at which requests
3286 * have been served by the device, it is more precise to
3287 * extend the observation interval to the last completion.
3289 bfqd
->delta_from_first
=
3290 max_t(u64
, bfqd
->delta_from_first
,
3291 bfqd
->last_completion
- bfqd
->first_dispatch
);
3294 * Rate computed in sects/usec, and not sects/nsec, for
3297 rate
= div64_ul(bfqd
->tot_sectors_dispatched
<<BFQ_RATE_SHIFT
,
3298 div_u64(bfqd
->delta_from_first
, NSEC_PER_USEC
));
3301 * Peak rate not updated if:
3302 * - the percentage of sequential dispatches is below 3/4 of the
3303 * total, and rate is below the current estimated peak rate
3304 * - rate is unreasonably high (> 20M sectors/sec)
3306 if ((bfqd
->sequential_samples
< (3 * bfqd
->peak_rate_samples
)>>2 &&
3307 rate
<= bfqd
->peak_rate
) ||
3308 rate
> 20<<BFQ_RATE_SHIFT
)
3309 goto reset_computation
;
3312 * We have to update the peak rate, at last! To this purpose,
3313 * we use a low-pass filter. We compute the smoothing constant
3314 * of the filter as a function of the 'weight' of the new
3317 * As can be seen in next formulas, we define this weight as a
3318 * quantity proportional to how sequential the workload is,
3319 * and to how long the observation time interval is.
3321 * The weight runs from 0 to 8. The maximum value of the
3322 * weight, 8, yields the minimum value for the smoothing
3323 * constant. At this minimum value for the smoothing constant,
3324 * the measured rate contributes for half of the next value of
3325 * the estimated peak rate.
3327 * So, the first step is to compute the weight as a function
3328 * of how sequential the workload is. Note that the weight
3329 * cannot reach 9, because bfqd->sequential_samples cannot
3330 * become equal to bfqd->peak_rate_samples, which, in its
3331 * turn, holds true because bfqd->sequential_samples is not
3332 * incremented for the first sample.
3334 weight
= (9 * bfqd
->sequential_samples
) / bfqd
->peak_rate_samples
;
3337 * Second step: further refine the weight as a function of the
3338 * duration of the observation interval.
3340 weight
= min_t(u32
, 8,
3341 div_u64(weight
* bfqd
->delta_from_first
,
3342 BFQ_RATE_REF_INTERVAL
));
3345 * Divisor ranging from 10, for minimum weight, to 2, for
3348 divisor
= 10 - weight
;
3351 * Finally, update peak rate:
3353 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
3355 bfqd
->peak_rate
*= divisor
-1;
3356 bfqd
->peak_rate
/= divisor
;
3357 rate
/= divisor
; /* smoothing constant alpha = 1/divisor */
3359 bfqd
->peak_rate
+= rate
;
3362 * For a very slow device, bfqd->peak_rate can reach 0 (see
3363 * the minimum representable values reported in the comments
3364 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3365 * divisions by zero where bfqd->peak_rate is used as a
3368 bfqd
->peak_rate
= max_t(u32
, 1, bfqd
->peak_rate
);
3370 update_thr_responsiveness_params(bfqd
);
3373 bfq_reset_rate_computation(bfqd
, rq
);
3377 * Update the read/write peak rate (the main quantity used for
3378 * auto-tuning, see update_thr_responsiveness_params()).
3380 * It is not trivial to estimate the peak rate (correctly): because of
3381 * the presence of sw and hw queues between the scheduler and the
3382 * device components that finally serve I/O requests, it is hard to
3383 * say exactly when a given dispatched request is served inside the
3384 * device, and for how long. As a consequence, it is hard to know
3385 * precisely at what rate a given set of requests is actually served
3388 * On the opposite end, the dispatch time of any request is trivially
3389 * available, and, from this piece of information, the "dispatch rate"
3390 * of requests can be immediately computed. So, the idea in the next
3391 * function is to use what is known, namely request dispatch times
3392 * (plus, when useful, request completion times), to estimate what is
3393 * unknown, namely in-device request service rate.
3395 * The main issue is that, because of the above facts, the rate at
3396 * which a certain set of requests is dispatched over a certain time
3397 * interval can vary greatly with respect to the rate at which the
3398 * same requests are then served. But, since the size of any
3399 * intermediate queue is limited, and the service scheme is lossless
3400 * (no request is silently dropped), the following obvious convergence
3401 * property holds: the number of requests dispatched MUST become
3402 * closer and closer to the number of requests completed as the
3403 * observation interval grows. This is the key property used in
3404 * the next function to estimate the peak service rate as a function
3405 * of the observed dispatch rate. The function assumes to be invoked
3406 * on every request dispatch.
3408 static void bfq_update_peak_rate(struct bfq_data
*bfqd
, struct request
*rq
)
3410 u64 now_ns
= ktime_get_ns();
3412 if (bfqd
->peak_rate_samples
== 0) { /* first dispatch */
3413 bfq_log(bfqd
, "update_peak_rate: goto reset, samples %d",
3414 bfqd
->peak_rate_samples
);
3415 bfq_reset_rate_computation(bfqd
, rq
);
3416 goto update_last_values
; /* will add one sample */
3420 * Device idle for very long: the observation interval lasting
3421 * up to this dispatch cannot be a valid observation interval
3422 * for computing a new peak rate (similarly to the late-
3423 * completion event in bfq_completed_request()). Go to
3424 * update_rate_and_reset to have the following three steps
3426 * - close the observation interval at the last (previous)
3427 * request dispatch or completion
3428 * - compute rate, if possible, for that observation interval
3429 * - start a new observation interval with this dispatch
3431 if (now_ns
- bfqd
->last_dispatch
> 100*NSEC_PER_MSEC
&&
3432 bfqd
->rq_in_driver
== 0)
3433 goto update_rate_and_reset
;
3435 /* Update sampling information */
3436 bfqd
->peak_rate_samples
++;
3438 if ((bfqd
->rq_in_driver
> 0 ||
3439 now_ns
- bfqd
->last_completion
< BFQ_MIN_TT
)
3440 && !BFQ_RQ_SEEKY(bfqd
, bfqd
->last_position
, rq
))
3441 bfqd
->sequential_samples
++;
3443 bfqd
->tot_sectors_dispatched
+= blk_rq_sectors(rq
);
3445 /* Reset max observed rq size every 32 dispatches */
3446 if (likely(bfqd
->peak_rate_samples
% 32))
3447 bfqd
->last_rq_max_size
=
3448 max_t(u32
, blk_rq_sectors(rq
), bfqd
->last_rq_max_size
);
3450 bfqd
->last_rq_max_size
= blk_rq_sectors(rq
);
3452 bfqd
->delta_from_first
= now_ns
- bfqd
->first_dispatch
;
3454 /* Target observation interval not yet reached, go on sampling */
3455 if (bfqd
->delta_from_first
< BFQ_RATE_REF_INTERVAL
)
3456 goto update_last_values
;
3458 update_rate_and_reset
:
3459 bfq_update_rate_reset(bfqd
, rq
);
3461 bfqd
->last_position
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
3462 if (RQ_BFQQ(rq
) == bfqd
->in_service_queue
)
3463 bfqd
->in_serv_last_pos
= bfqd
->last_position
;
3464 bfqd
->last_dispatch
= now_ns
;
3468 * Remove request from internal lists.
3470 static void bfq_dispatch_remove(struct request_queue
*q
, struct request
*rq
)
3472 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
3475 * For consistency, the next instruction should have been
3476 * executed after removing the request from the queue and
3477 * dispatching it. We execute instead this instruction before
3478 * bfq_remove_request() (and hence introduce a temporary
3479 * inconsistency), for efficiency. In fact, should this
3480 * dispatch occur for a non in-service bfqq, this anticipated
3481 * increment prevents two counters related to bfqq->dispatched
3482 * from risking to be, first, uselessly decremented, and then
3483 * incremented again when the (new) value of bfqq->dispatched
3484 * happens to be taken into account.
3487 bfq_update_peak_rate(q
->elevator
->elevator_data
, rq
);
3489 bfq_remove_request(q
, rq
);
3493 * There is a case where idling does not have to be performed for
3494 * throughput concerns, but to preserve the throughput share of
3495 * the process associated with bfqq.
3497 * To introduce this case, we can note that allowing the drive
3498 * to enqueue more than one request at a time, and hence
3499 * delegating de facto final scheduling decisions to the
3500 * drive's internal scheduler, entails loss of control on the
3501 * actual request service order. In particular, the critical
3502 * situation is when requests from different processes happen
3503 * to be present, at the same time, in the internal queue(s)
3504 * of the drive. In such a situation, the drive, by deciding
3505 * the service order of the internally-queued requests, does
3506 * determine also the actual throughput distribution among
3507 * these processes. But the drive typically has no notion or
3508 * concern about per-process throughput distribution, and
3509 * makes its decisions only on a per-request basis. Therefore,
3510 * the service distribution enforced by the drive's internal
3511 * scheduler is likely to coincide with the desired throughput
3512 * distribution only in a completely symmetric, or favorably
3513 * skewed scenario where:
3514 * (i-a) each of these processes must get the same throughput as
3516 * (i-b) in case (i-a) does not hold, it holds that the process
3517 * associated with bfqq must receive a lower or equal
3518 * throughput than any of the other processes;
3519 * (ii) the I/O of each process has the same properties, in
3520 * terms of locality (sequential or random), direction
3521 * (reads or writes), request sizes, greediness
3522 * (from I/O-bound to sporadic), and so on;
3524 * In fact, in such a scenario, the drive tends to treat the requests
3525 * of each process in about the same way as the requests of the
3526 * others, and thus to provide each of these processes with about the
3527 * same throughput. This is exactly the desired throughput
3528 * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3529 * even more convenient distribution for (the process associated with)
3532 * In contrast, in any asymmetric or unfavorable scenario, device
3533 * idling (I/O-dispatch plugging) is certainly needed to guarantee
3534 * that bfqq receives its assigned fraction of the device throughput
3535 * (see [1] for details).
3537 * The problem is that idling may significantly reduce throughput with
3538 * certain combinations of types of I/O and devices. An important
3539 * example is sync random I/O on flash storage with command
3540 * queueing. So, unless bfqq falls in cases where idling also boosts
3541 * throughput, it is important to check conditions (i-a), i(-b) and
3542 * (ii) accurately, so as to avoid idling when not strictly needed for
3543 * service guarantees.
3545 * Unfortunately, it is extremely difficult to thoroughly check
3546 * condition (ii). And, in case there are active groups, it becomes
3547 * very difficult to check conditions (i-a) and (i-b) too. In fact,
3548 * if there are active groups, then, for conditions (i-a) or (i-b) to
3549 * become false 'indirectly', it is enough that an active group
3550 * contains more active processes or sub-groups than some other active
3551 * group. More precisely, for conditions (i-a) or (i-b) to become
3552 * false because of such a group, it is not even necessary that the
3553 * group is (still) active: it is sufficient that, even if the group
3554 * has become inactive, some of its descendant processes still have
3555 * some request already dispatched but still waiting for
3556 * completion. In fact, requests have still to be guaranteed their
3557 * share of the throughput even after being dispatched. In this
3558 * respect, it is easy to show that, if a group frequently becomes
3559 * inactive while still having in-flight requests, and if, when this
3560 * happens, the group is not considered in the calculation of whether
3561 * the scenario is asymmetric, then the group may fail to be
3562 * guaranteed its fair share of the throughput (basically because
3563 * idling may not be performed for the descendant processes of the
3564 * group, but it had to be). We address this issue with the following
3565 * bi-modal behavior, implemented in the function
3566 * bfq_asymmetric_scenario().
3568 * If there are groups with requests waiting for completion
3569 * (as commented above, some of these groups may even be
3570 * already inactive), then the scenario is tagged as
3571 * asymmetric, conservatively, without checking any of the
3572 * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3573 * This behavior matches also the fact that groups are created
3574 * exactly if controlling I/O is a primary concern (to
3575 * preserve bandwidth and latency guarantees).
3577 * On the opposite end, if there are no groups with requests waiting
3578 * for completion, then only conditions (i-a) and (i-b) are actually
3579 * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3580 * idling is not performed, regardless of whether condition (ii)
3581 * holds. In other words, only if conditions (i-a) and (i-b) do not
3582 * hold, then idling is allowed, and the device tends to be prevented
3583 * from queueing many requests, possibly of several processes. Since
3584 * there are no groups with requests waiting for completion, then, to
3585 * control conditions (i-a) and (i-b) it is enough to check just
3586 * whether all the queues with requests waiting for completion also
3587 * have the same weight.
3589 * Not checking condition (ii) evidently exposes bfqq to the
3590 * risk of getting less throughput than its fair share.
3591 * However, for queues with the same weight, a further
3592 * mechanism, preemption, mitigates or even eliminates this
3593 * problem. And it does so without consequences on overall
3594 * throughput. This mechanism and its benefits are explained
3595 * in the next three paragraphs.
3597 * Even if a queue, say Q, is expired when it remains idle, Q
3598 * can still preempt the new in-service queue if the next
3599 * request of Q arrives soon (see the comments on
3600 * bfq_bfqq_update_budg_for_activation). If all queues and
3601 * groups have the same weight, this form of preemption,
3602 * combined with the hole-recovery heuristic described in the
3603 * comments on function bfq_bfqq_update_budg_for_activation,
3604 * are enough to preserve a correct bandwidth distribution in
3605 * the mid term, even without idling. In fact, even if not
3606 * idling allows the internal queues of the device to contain
3607 * many requests, and thus to reorder requests, we can rather
3608 * safely assume that the internal scheduler still preserves a
3609 * minimum of mid-term fairness.
3611 * More precisely, this preemption-based, idleless approach
3612 * provides fairness in terms of IOPS, and not sectors per
3613 * second. This can be seen with a simple example. Suppose
3614 * that there are two queues with the same weight, but that
3615 * the first queue receives requests of 8 sectors, while the
3616 * second queue receives requests of 1024 sectors. In
3617 * addition, suppose that each of the two queues contains at
3618 * most one request at a time, which implies that each queue
3619 * always remains idle after it is served. Finally, after
3620 * remaining idle, each queue receives very quickly a new
3621 * request. It follows that the two queues are served
3622 * alternatively, preempting each other if needed. This
3623 * implies that, although both queues have the same weight,
3624 * the queue with large requests receives a service that is
3625 * 1024/8 times as high as the service received by the other
3628 * The motivation for using preemption instead of idling (for
3629 * queues with the same weight) is that, by not idling,
3630 * service guarantees are preserved (completely or at least in
3631 * part) without minimally sacrificing throughput. And, if
3632 * there is no active group, then the primary expectation for
3633 * this device is probably a high throughput.
3635 * We are now left only with explaining the two sub-conditions in the
3636 * additional compound condition that is checked below for deciding
3637 * whether the scenario is asymmetric. To explain the first
3638 * sub-condition, we need to add that the function
3639 * bfq_asymmetric_scenario checks the weights of only
3640 * non-weight-raised queues, for efficiency reasons (see comments on
3641 * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3642 * is checked explicitly here. More precisely, the compound condition
3643 * below takes into account also the fact that, even if bfqq is being
3644 * weight-raised, the scenario is still symmetric if all queues with
3645 * requests waiting for completion happen to be
3646 * weight-raised. Actually, we should be even more precise here, and
3647 * differentiate between interactive weight raising and soft real-time
3650 * The second sub-condition checked in the compound condition is
3651 * whether there is a fair amount of already in-flight I/O not
3652 * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3653 * following reason. The drive may decide to serve in-flight
3654 * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3655 * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3656 * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3657 * basically uncontrolled amount of I/O from other queues may be
3658 * dispatched too, possibly causing the service of bfqq's I/O to be
3659 * delayed even longer in the drive. This problem gets more and more
3660 * serious as the speed and the queue depth of the drive grow,
3661 * because, as these two quantities grow, the probability to find no
3662 * queue busy but many requests in flight grows too. By contrast,
3663 * plugging I/O dispatching minimizes the delay induced by already
3664 * in-flight I/O, and enables bfqq to recover the bandwidth it may
3665 * lose because of this delay.
3667 * As a side note, it is worth considering that the above
3668 * device-idling countermeasures may however fail in the following
3669 * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3670 * in a time period during which all symmetry sub-conditions hold, and
3671 * therefore the device is allowed to enqueue many requests, but at
3672 * some later point in time some sub-condition stops to hold, then it
3673 * may become impossible to make requests be served in the desired
3674 * order until all the requests already queued in the device have been
3675 * served. The last sub-condition commented above somewhat mitigates
3676 * this problem for weight-raised queues.
3678 * However, as an additional mitigation for this problem, we preserve
3679 * plugging for a special symmetric case that may suddenly turn into
3680 * asymmetric: the case where only bfqq is busy. In this case, not
3681 * expiring bfqq does not cause any harm to any other queues in terms
3682 * of service guarantees. In contrast, it avoids the following unlucky
3683 * sequence of events: (1) bfqq is expired, (2) a new queue with a
3684 * lower weight than bfqq becomes busy (or more queues), (3) the new
3685 * queue is served until a new request arrives for bfqq, (4) when bfqq
3686 * is finally served, there are so many requests of the new queue in
3687 * the drive that the pending requests for bfqq take a lot of time to
3688 * be served. In particular, event (2) may case even already
3689 * dispatched requests of bfqq to be delayed, inside the drive. So, to
3690 * avoid this series of events, the scenario is preventively declared
3691 * as asymmetric also if bfqq is the only busy queues
3693 static bool idling_needed_for_service_guarantees(struct bfq_data
*bfqd
,
3694 struct bfq_queue
*bfqq
)
3696 int tot_busy_queues
= bfq_tot_busy_queues(bfqd
);
3698 /* No point in idling for bfqq if it won't get requests any longer */
3699 if (unlikely(!bfqq_process_refs(bfqq
)))
3702 return (bfqq
->wr_coeff
> 1 &&
3703 (bfqd
->wr_busy_queues
<
3705 bfqd
->rq_in_driver
>=
3706 bfqq
->dispatched
+ 4)) ||
3707 bfq_asymmetric_scenario(bfqd
, bfqq
) ||
3708 tot_busy_queues
== 1;
3711 static bool __bfq_bfqq_expire(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
3712 enum bfqq_expiration reason
)
3715 * If this bfqq is shared between multiple processes, check
3716 * to make sure that those processes are still issuing I/Os
3717 * within the mean seek distance. If not, it may be time to
3718 * break the queues apart again.
3720 if (bfq_bfqq_coop(bfqq
) && BFQQ_SEEKY(bfqq
))
3721 bfq_mark_bfqq_split_coop(bfqq
);
3724 * Consider queues with a higher finish virtual time than
3725 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3726 * true, then bfqq's bandwidth would be violated if an
3727 * uncontrolled amount of I/O from these queues were
3728 * dispatched while bfqq is waiting for its new I/O to
3729 * arrive. This is exactly what may happen if this is a forced
3730 * expiration caused by a preemption attempt, and if bfqq is
3731 * not re-scheduled. To prevent this from happening, re-queue
3732 * bfqq if it needs I/O-dispatch plugging, even if it is
3733 * empty. By doing so, bfqq is granted to be served before the
3734 * above queues (provided that bfqq is of course eligible).
3736 if (RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
3737 !(reason
== BFQQE_PREEMPTED
&&
3738 idling_needed_for_service_guarantees(bfqd
, bfqq
))) {
3739 if (bfqq
->dispatched
== 0)
3741 * Overloading budget_timeout field to store
3742 * the time at which the queue remains with no
3743 * backlog and no outstanding request; used by
3744 * the weight-raising mechanism.
3746 bfqq
->budget_timeout
= jiffies
;
3748 bfq_del_bfqq_busy(bfqd
, bfqq
, true);
3750 bfq_requeue_bfqq(bfqd
, bfqq
, true);
3752 * Resort priority tree of potential close cooperators.
3753 * See comments on bfq_pos_tree_add_move() for the unlikely().
3755 if (unlikely(!bfqd
->nonrot_with_queueing
&&
3756 !RB_EMPTY_ROOT(&bfqq
->sort_list
)))
3757 bfq_pos_tree_add_move(bfqd
, bfqq
);
3761 * All in-service entities must have been properly deactivated
3762 * or requeued before executing the next function, which
3763 * resets all in-service entities as no more in service. This
3764 * may cause bfqq to be freed. If this happens, the next
3765 * function returns true.
3767 return __bfq_bfqd_reset_in_service(bfqd
);
3771 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3772 * @bfqd: device data.
3773 * @bfqq: queue to update.
3774 * @reason: reason for expiration.
3776 * Handle the feedback on @bfqq budget at queue expiration.
3777 * See the body for detailed comments.
3779 static void __bfq_bfqq_recalc_budget(struct bfq_data
*bfqd
,
3780 struct bfq_queue
*bfqq
,
3781 enum bfqq_expiration reason
)
3783 struct request
*next_rq
;
3784 int budget
, min_budget
;
3786 min_budget
= bfq_min_budget(bfqd
);
3788 if (bfqq
->wr_coeff
== 1)
3789 budget
= bfqq
->max_budget
;
3791 * Use a constant, low budget for weight-raised queues,
3792 * to help achieve a low latency. Keep it slightly higher
3793 * than the minimum possible budget, to cause a little
3794 * bit fewer expirations.
3796 budget
= 2 * min_budget
;
3798 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: last budg %d, budg left %d",
3799 bfqq
->entity
.budget
, bfq_bfqq_budget_left(bfqq
));
3800 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: last max_budg %d, min budg %d",
3801 budget
, bfq_min_budget(bfqd
));
3802 bfq_log_bfqq(bfqd
, bfqq
, "recalc_budg: sync %d, seeky %d",
3803 bfq_bfqq_sync(bfqq
), BFQQ_SEEKY(bfqd
->in_service_queue
));
3805 if (bfq_bfqq_sync(bfqq
) && bfqq
->wr_coeff
== 1) {
3808 * Caveat: in all the following cases we trade latency
3811 case BFQQE_TOO_IDLE
:
3813 * This is the only case where we may reduce
3814 * the budget: if there is no request of the
3815 * process still waiting for completion, then
3816 * we assume (tentatively) that the timer has
3817 * expired because the batch of requests of
3818 * the process could have been served with a
3819 * smaller budget. Hence, betting that
3820 * process will behave in the same way when it
3821 * becomes backlogged again, we reduce its
3822 * next budget. As long as we guess right,
3823 * this budget cut reduces the latency
3824 * experienced by the process.
3826 * However, if there are still outstanding
3827 * requests, then the process may have not yet
3828 * issued its next request just because it is
3829 * still waiting for the completion of some of
3830 * the still outstanding ones. So in this
3831 * subcase we do not reduce its budget, on the
3832 * contrary we increase it to possibly boost
3833 * the throughput, as discussed in the
3834 * comments to the BUDGET_TIMEOUT case.
3836 if (bfqq
->dispatched
> 0) /* still outstanding reqs */
3837 budget
= min(budget
* 2, bfqd
->bfq_max_budget
);
3839 if (budget
> 5 * min_budget
)
3840 budget
-= 4 * min_budget
;
3842 budget
= min_budget
;
3845 case BFQQE_BUDGET_TIMEOUT
:
3847 * We double the budget here because it gives
3848 * the chance to boost the throughput if this
3849 * is not a seeky process (and has bumped into
3850 * this timeout because of, e.g., ZBR).
3852 budget
= min(budget
* 2, bfqd
->bfq_max_budget
);
3854 case BFQQE_BUDGET_EXHAUSTED
:
3856 * The process still has backlog, and did not
3857 * let either the budget timeout or the disk
3858 * idling timeout expire. Hence it is not
3859 * seeky, has a short thinktime and may be
3860 * happy with a higher budget too. So
3861 * definitely increase the budget of this good
3862 * candidate to boost the disk throughput.
3864 budget
= min(budget
* 4, bfqd
->bfq_max_budget
);
3866 case BFQQE_NO_MORE_REQUESTS
:
3868 * For queues that expire for this reason, it
3869 * is particularly important to keep the
3870 * budget close to the actual service they
3871 * need. Doing so reduces the timestamp
3872 * misalignment problem described in the
3873 * comments in the body of
3874 * __bfq_activate_entity. In fact, suppose
3875 * that a queue systematically expires for
3876 * BFQQE_NO_MORE_REQUESTS and presents a
3877 * new request in time to enjoy timestamp
3878 * back-shifting. The larger the budget of the
3879 * queue is with respect to the service the
3880 * queue actually requests in each service
3881 * slot, the more times the queue can be
3882 * reactivated with the same virtual finish
3883 * time. It follows that, even if this finish
3884 * time is pushed to the system virtual time
3885 * to reduce the consequent timestamp
3886 * misalignment, the queue unjustly enjoys for
3887 * many re-activations a lower finish time
3888 * than all newly activated queues.
3890 * The service needed by bfqq is measured
3891 * quite precisely by bfqq->entity.service.
3892 * Since bfqq does not enjoy device idling,
3893 * bfqq->entity.service is equal to the number
3894 * of sectors that the process associated with
3895 * bfqq requested to read/write before waiting
3896 * for request completions, or blocking for
3899 budget
= max_t(int, bfqq
->entity
.service
, min_budget
);
3904 } else if (!bfq_bfqq_sync(bfqq
)) {
3906 * Async queues get always the maximum possible
3907 * budget, as for them we do not care about latency
3908 * (in addition, their ability to dispatch is limited
3909 * by the charging factor).
3911 budget
= bfqd
->bfq_max_budget
;
3914 bfqq
->max_budget
= budget
;
3916 if (bfqd
->budgets_assigned
>= bfq_stats_min_budgets
&&
3917 !bfqd
->bfq_user_max_budget
)
3918 bfqq
->max_budget
= min(bfqq
->max_budget
, bfqd
->bfq_max_budget
);
3921 * If there is still backlog, then assign a new budget, making
3922 * sure that it is large enough for the next request. Since
3923 * the finish time of bfqq must be kept in sync with the
3924 * budget, be sure to call __bfq_bfqq_expire() *after* this
3927 * If there is no backlog, then no need to update the budget;
3928 * it will be updated on the arrival of a new request.
3930 next_rq
= bfqq
->next_rq
;
3932 bfqq
->entity
.budget
= max_t(unsigned long, bfqq
->max_budget
,
3933 bfq_serv_to_charge(next_rq
, bfqq
));
3935 bfq_log_bfqq(bfqd
, bfqq
, "head sect: %u, new budget %d",
3936 next_rq
? blk_rq_sectors(next_rq
) : 0,
3937 bfqq
->entity
.budget
);
3941 * Return true if the process associated with bfqq is "slow". The slow
3942 * flag is used, in addition to the budget timeout, to reduce the
3943 * amount of service provided to seeky processes, and thus reduce
3944 * their chances to lower the throughput. More details in the comments
3945 * on the function bfq_bfqq_expire().
3947 * An important observation is in order: as discussed in the comments
3948 * on the function bfq_update_peak_rate(), with devices with internal
3949 * queues, it is hard if ever possible to know when and for how long
3950 * an I/O request is processed by the device (apart from the trivial
3951 * I/O pattern where a new request is dispatched only after the
3952 * previous one has been completed). This makes it hard to evaluate
3953 * the real rate at which the I/O requests of each bfq_queue are
3954 * served. In fact, for an I/O scheduler like BFQ, serving a
3955 * bfq_queue means just dispatching its requests during its service
3956 * slot (i.e., until the budget of the queue is exhausted, or the
3957 * queue remains idle, or, finally, a timeout fires). But, during the
3958 * service slot of a bfq_queue, around 100 ms at most, the device may
3959 * be even still processing requests of bfq_queues served in previous
3960 * service slots. On the opposite end, the requests of the in-service
3961 * bfq_queue may be completed after the service slot of the queue
3964 * Anyway, unless more sophisticated solutions are used
3965 * (where possible), the sum of the sizes of the requests dispatched
3966 * during the service slot of a bfq_queue is probably the only
3967 * approximation available for the service received by the bfq_queue
3968 * during its service slot. And this sum is the quantity used in this
3969 * function to evaluate the I/O speed of a process.
3971 static bool bfq_bfqq_is_slow(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
3972 bool compensate
, enum bfqq_expiration reason
,
3973 unsigned long *delta_ms
)
3975 ktime_t delta_ktime
;
3977 bool slow
= BFQQ_SEEKY(bfqq
); /* if delta too short, use seekyness */
3979 if (!bfq_bfqq_sync(bfqq
))
3983 delta_ktime
= bfqd
->last_idling_start
;
3985 delta_ktime
= ktime_get();
3986 delta_ktime
= ktime_sub(delta_ktime
, bfqd
->last_budget_start
);
3987 delta_usecs
= ktime_to_us(delta_ktime
);
3989 /* don't use too short time intervals */
3990 if (delta_usecs
< 1000) {
3991 if (blk_queue_nonrot(bfqd
->queue
))
3993 * give same worst-case guarantees as idling
3996 *delta_ms
= BFQ_MIN_TT
/ NSEC_PER_MSEC
;
3997 else /* charge at least one seek */
3998 *delta_ms
= bfq_slice_idle
/ NSEC_PER_MSEC
;
4003 *delta_ms
= delta_usecs
/ USEC_PER_MSEC
;
4006 * Use only long (> 20ms) intervals to filter out excessive
4007 * spikes in service rate estimation.
4009 if (delta_usecs
> 20000) {
4011 * Caveat for rotational devices: processes doing I/O
4012 * in the slower disk zones tend to be slow(er) even
4013 * if not seeky. In this respect, the estimated peak
4014 * rate is likely to be an average over the disk
4015 * surface. Accordingly, to not be too harsh with
4016 * unlucky processes, a process is deemed slow only if
4017 * its rate has been lower than half of the estimated
4020 slow
= bfqq
->entity
.service
< bfqd
->bfq_max_budget
/ 2;
4023 bfq_log_bfqq(bfqd
, bfqq
, "bfq_bfqq_is_slow: slow %d", slow
);
4029 * To be deemed as soft real-time, an application must meet two
4030 * requirements. First, the application must not require an average
4031 * bandwidth higher than the approximate bandwidth required to playback or
4032 * record a compressed high-definition video.
4033 * The next function is invoked on the completion of the last request of a
4034 * batch, to compute the next-start time instant, soft_rt_next_start, such
4035 * that, if the next request of the application does not arrive before
4036 * soft_rt_next_start, then the above requirement on the bandwidth is met.
4038 * The second requirement is that the request pattern of the application is
4039 * isochronous, i.e., that, after issuing a request or a batch of requests,
4040 * the application stops issuing new requests until all its pending requests
4041 * have been completed. After that, the application may issue a new batch,
4043 * For this reason the next function is invoked to compute
4044 * soft_rt_next_start only for applications that meet this requirement,
4045 * whereas soft_rt_next_start is set to infinity for applications that do
4048 * Unfortunately, even a greedy (i.e., I/O-bound) application may
4049 * happen to meet, occasionally or systematically, both the above
4050 * bandwidth and isochrony requirements. This may happen at least in
4051 * the following circumstances. First, if the CPU load is high. The
4052 * application may stop issuing requests while the CPUs are busy
4053 * serving other processes, then restart, then stop again for a while,
4054 * and so on. The other circumstances are related to the storage
4055 * device: the storage device is highly loaded or reaches a low-enough
4056 * throughput with the I/O of the application (e.g., because the I/O
4057 * is random and/or the device is slow). In all these cases, the
4058 * I/O of the application may be simply slowed down enough to meet
4059 * the bandwidth and isochrony requirements. To reduce the probability
4060 * that greedy applications are deemed as soft real-time in these
4061 * corner cases, a further rule is used in the computation of
4062 * soft_rt_next_start: the return value of this function is forced to
4063 * be higher than the maximum between the following two quantities.
4065 * (a) Current time plus: (1) the maximum time for which the arrival
4066 * of a request is waited for when a sync queue becomes idle,
4067 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
4068 * postpone for a moment the reason for adding a few extra
4069 * jiffies; we get back to it after next item (b). Lower-bounding
4070 * the return value of this function with the current time plus
4071 * bfqd->bfq_slice_idle tends to filter out greedy applications,
4072 * because the latter issue their next request as soon as possible
4073 * after the last one has been completed. In contrast, a soft
4074 * real-time application spends some time processing data, after a
4075 * batch of its requests has been completed.
4077 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
4078 * above, greedy applications may happen to meet both the
4079 * bandwidth and isochrony requirements under heavy CPU or
4080 * storage-device load. In more detail, in these scenarios, these
4081 * applications happen, only for limited time periods, to do I/O
4082 * slowly enough to meet all the requirements described so far,
4083 * including the filtering in above item (a). These slow-speed
4084 * time intervals are usually interspersed between other time
4085 * intervals during which these applications do I/O at a very high
4086 * speed. Fortunately, exactly because of the high speed of the
4087 * I/O in the high-speed intervals, the values returned by this
4088 * function happen to be so high, near the end of any such
4089 * high-speed interval, to be likely to fall *after* the end of
4090 * the low-speed time interval that follows. These high values are
4091 * stored in bfqq->soft_rt_next_start after each invocation of
4092 * this function. As a consequence, if the last value of
4093 * bfqq->soft_rt_next_start is constantly used to lower-bound the
4094 * next value that this function may return, then, from the very
4095 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
4096 * likely to be constantly kept so high that any I/O request
4097 * issued during the low-speed interval is considered as arriving
4098 * to soon for the application to be deemed as soft
4099 * real-time. Then, in the high-speed interval that follows, the
4100 * application will not be deemed as soft real-time, just because
4101 * it will do I/O at a high speed. And so on.
4103 * Getting back to the filtering in item (a), in the following two
4104 * cases this filtering might be easily passed by a greedy
4105 * application, if the reference quantity was just
4106 * bfqd->bfq_slice_idle:
4107 * 1) HZ is so low that the duration of a jiffy is comparable to or
4108 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
4109 * devices with HZ=100. The time granularity may be so coarse
4110 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
4111 * is rather lower than the exact value.
4112 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
4113 * for a while, then suddenly 'jump' by several units to recover the lost
4114 * increments. This seems to happen, e.g., inside virtual machines.
4115 * To address this issue, in the filtering in (a) we do not use as a
4116 * reference time interval just bfqd->bfq_slice_idle, but
4117 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
4118 * minimum number of jiffies for which the filter seems to be quite
4119 * precise also in embedded systems and KVM/QEMU virtual machines.
4121 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data
*bfqd
,
4122 struct bfq_queue
*bfqq
)
4124 return max3(bfqq
->soft_rt_next_start
,
4125 bfqq
->last_idle_bklogged
+
4126 HZ
* bfqq
->service_from_backlogged
/
4127 bfqd
->bfq_wr_max_softrt_rate
,
4128 jiffies
+ nsecs_to_jiffies(bfqq
->bfqd
->bfq_slice_idle
) + 4);
4132 * bfq_bfqq_expire - expire a queue.
4133 * @bfqd: device owning the queue.
4134 * @bfqq: the queue to expire.
4135 * @compensate: if true, compensate for the time spent idling.
4136 * @reason: the reason causing the expiration.
4138 * If the process associated with bfqq does slow I/O (e.g., because it
4139 * issues random requests), we charge bfqq with the time it has been
4140 * in service instead of the service it has received (see
4141 * bfq_bfqq_charge_time for details on how this goal is achieved). As
4142 * a consequence, bfqq will typically get higher timestamps upon
4143 * reactivation, and hence it will be rescheduled as if it had
4144 * received more service than what it has actually received. In the
4145 * end, bfqq receives less service in proportion to how slowly its
4146 * associated process consumes its budgets (and hence how seriously it
4147 * tends to lower the throughput). In addition, this time-charging
4148 * strategy guarantees time fairness among slow processes. In
4149 * contrast, if the process associated with bfqq is not slow, we
4150 * charge bfqq exactly with the service it has received.
4152 * Charging time to the first type of queues and the exact service to
4153 * the other has the effect of using the WF2Q+ policy to schedule the
4154 * former on a timeslice basis, without violating service domain
4155 * guarantees among the latter.
4157 void bfq_bfqq_expire(struct bfq_data
*bfqd
,
4158 struct bfq_queue
*bfqq
,
4160 enum bfqq_expiration reason
)
4163 unsigned long delta
= 0;
4164 struct bfq_entity
*entity
= &bfqq
->entity
;
4167 * Check whether the process is slow (see bfq_bfqq_is_slow).
4169 slow
= bfq_bfqq_is_slow(bfqd
, bfqq
, compensate
, reason
, &delta
);
4172 * As above explained, charge slow (typically seeky) and
4173 * timed-out queues with the time and not the service
4174 * received, to favor sequential workloads.
4176 * Processes doing I/O in the slower disk zones will tend to
4177 * be slow(er) even if not seeky. Therefore, since the
4178 * estimated peak rate is actually an average over the disk
4179 * surface, these processes may timeout just for bad luck. To
4180 * avoid punishing them, do not charge time to processes that
4181 * succeeded in consuming at least 2/3 of their budget. This
4182 * allows BFQ to preserve enough elasticity to still perform
4183 * bandwidth, and not time, distribution with little unlucky
4184 * or quasi-sequential processes.
4186 if (bfqq
->wr_coeff
== 1 &&
4188 (reason
== BFQQE_BUDGET_TIMEOUT
&&
4189 bfq_bfqq_budget_left(bfqq
) >= entity
->budget
/ 3)))
4190 bfq_bfqq_charge_time(bfqd
, bfqq
, delta
);
4192 if (bfqd
->low_latency
&& bfqq
->wr_coeff
== 1)
4193 bfqq
->last_wr_start_finish
= jiffies
;
4195 if (bfqd
->low_latency
&& bfqd
->bfq_wr_max_softrt_rate
> 0 &&
4196 RB_EMPTY_ROOT(&bfqq
->sort_list
)) {
4198 * If we get here, and there are no outstanding
4199 * requests, then the request pattern is isochronous
4200 * (see the comments on the function
4201 * bfq_bfqq_softrt_next_start()). Therefore we can
4202 * compute soft_rt_next_start.
4204 * If, instead, the queue still has outstanding
4205 * requests, then we have to wait for the completion
4206 * of all the outstanding requests to discover whether
4207 * the request pattern is actually isochronous.
4209 if (bfqq
->dispatched
== 0)
4210 bfqq
->soft_rt_next_start
=
4211 bfq_bfqq_softrt_next_start(bfqd
, bfqq
);
4212 else if (bfqq
->dispatched
> 0) {
4214 * Schedule an update of soft_rt_next_start to when
4215 * the task may be discovered to be isochronous.
4217 bfq_mark_bfqq_softrt_update(bfqq
);
4221 bfq_log_bfqq(bfqd
, bfqq
,
4222 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason
,
4223 slow
, bfqq
->dispatched
, bfq_bfqq_has_short_ttime(bfqq
));
4226 * bfqq expired, so no total service time needs to be computed
4227 * any longer: reset state machine for measuring total service
4230 bfqd
->rqs_injected
= bfqd
->wait_dispatch
= false;
4231 bfqd
->waited_rq
= NULL
;
4234 * Increase, decrease or leave budget unchanged according to
4237 __bfq_bfqq_recalc_budget(bfqd
, bfqq
, reason
);
4238 if (__bfq_bfqq_expire(bfqd
, bfqq
, reason
))
4239 /* bfqq is gone, no more actions on it */
4242 /* mark bfqq as waiting a request only if a bic still points to it */
4243 if (!bfq_bfqq_busy(bfqq
) &&
4244 reason
!= BFQQE_BUDGET_TIMEOUT
&&
4245 reason
!= BFQQE_BUDGET_EXHAUSTED
) {
4246 bfq_mark_bfqq_non_blocking_wait_rq(bfqq
);
4248 * Not setting service to 0, because, if the next rq
4249 * arrives in time, the queue will go on receiving
4250 * service with this same budget (as if it never expired)
4253 entity
->service
= 0;
4256 * Reset the received-service counter for every parent entity.
4257 * Differently from what happens with bfqq->entity.service,
4258 * the resetting of this counter never needs to be postponed
4259 * for parent entities. In fact, in case bfqq may have a
4260 * chance to go on being served using the last, partially
4261 * consumed budget, bfqq->entity.service needs to be kept,
4262 * because if bfqq then actually goes on being served using
4263 * the same budget, the last value of bfqq->entity.service is
4264 * needed to properly decrement bfqq->entity.budget by the
4265 * portion already consumed. In contrast, it is not necessary
4266 * to keep entity->service for parent entities too, because
4267 * the bubble up of the new value of bfqq->entity.budget will
4268 * make sure that the budgets of parent entities are correct,
4269 * even in case bfqq and thus parent entities go on receiving
4270 * service with the same budget.
4272 entity
= entity
->parent
;
4273 for_each_entity(entity
)
4274 entity
->service
= 0;
4278 * Budget timeout is not implemented through a dedicated timer, but
4279 * just checked on request arrivals and completions, as well as on
4280 * idle timer expirations.
4282 static bool bfq_bfqq_budget_timeout(struct bfq_queue
*bfqq
)
4284 return time_is_before_eq_jiffies(bfqq
->budget_timeout
);
4288 * If we expire a queue that is actively waiting (i.e., with the
4289 * device idled) for the arrival of a new request, then we may incur
4290 * the timestamp misalignment problem described in the body of the
4291 * function __bfq_activate_entity. Hence we return true only if this
4292 * condition does not hold, or if the queue is slow enough to deserve
4293 * only to be kicked off for preserving a high throughput.
4295 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue
*bfqq
)
4297 bfq_log_bfqq(bfqq
->bfqd
, bfqq
,
4298 "may_budget_timeout: wait_request %d left %d timeout %d",
4299 bfq_bfqq_wait_request(bfqq
),
4300 bfq_bfqq_budget_left(bfqq
) >= bfqq
->entity
.budget
/ 3,
4301 bfq_bfqq_budget_timeout(bfqq
));
4303 return (!bfq_bfqq_wait_request(bfqq
) ||
4304 bfq_bfqq_budget_left(bfqq
) >= bfqq
->entity
.budget
/ 3)
4306 bfq_bfqq_budget_timeout(bfqq
);
4309 static bool idling_boosts_thr_without_issues(struct bfq_data
*bfqd
,
4310 struct bfq_queue
*bfqq
)
4312 bool rot_without_queueing
=
4313 !blk_queue_nonrot(bfqd
->queue
) && !bfqd
->hw_tag
,
4314 bfqq_sequential_and_IO_bound
,
4317 /* No point in idling for bfqq if it won't get requests any longer */
4318 if (unlikely(!bfqq_process_refs(bfqq
)))
4321 bfqq_sequential_and_IO_bound
= !BFQQ_SEEKY(bfqq
) &&
4322 bfq_bfqq_IO_bound(bfqq
) && bfq_bfqq_has_short_ttime(bfqq
);
4325 * The next variable takes into account the cases where idling
4326 * boosts the throughput.
4328 * The value of the variable is computed considering, first, that
4329 * idling is virtually always beneficial for the throughput if:
4330 * (a) the device is not NCQ-capable and rotational, or
4331 * (b) regardless of the presence of NCQ, the device is rotational and
4332 * the request pattern for bfqq is I/O-bound and sequential, or
4333 * (c) regardless of whether it is rotational, the device is
4334 * not NCQ-capable and the request pattern for bfqq is
4335 * I/O-bound and sequential.
4337 * Secondly, and in contrast to the above item (b), idling an
4338 * NCQ-capable flash-based device would not boost the
4339 * throughput even with sequential I/O; rather it would lower
4340 * the throughput in proportion to how fast the device
4341 * is. Accordingly, the next variable is true if any of the
4342 * above conditions (a), (b) or (c) is true, and, in
4343 * particular, happens to be false if bfqd is an NCQ-capable
4344 * flash-based device.
4346 idling_boosts_thr
= rot_without_queueing
||
4347 ((!blk_queue_nonrot(bfqd
->queue
) || !bfqd
->hw_tag
) &&
4348 bfqq_sequential_and_IO_bound
);
4351 * The return value of this function is equal to that of
4352 * idling_boosts_thr, unless a special case holds. In this
4353 * special case, described below, idling may cause problems to
4354 * weight-raised queues.
4356 * When the request pool is saturated (e.g., in the presence
4357 * of write hogs), if the processes associated with
4358 * non-weight-raised queues ask for requests at a lower rate,
4359 * then processes associated with weight-raised queues have a
4360 * higher probability to get a request from the pool
4361 * immediately (or at least soon) when they need one. Thus
4362 * they have a higher probability to actually get a fraction
4363 * of the device throughput proportional to their high
4364 * weight. This is especially true with NCQ-capable drives,
4365 * which enqueue several requests in advance, and further
4366 * reorder internally-queued requests.
4368 * For this reason, we force to false the return value if
4369 * there are weight-raised busy queues. In this case, and if
4370 * bfqq is not weight-raised, this guarantees that the device
4371 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4372 * then idling will be guaranteed by another variable, see
4373 * below). Combined with the timestamping rules of BFQ (see
4374 * [1] for details), this behavior causes bfqq, and hence any
4375 * sync non-weight-raised queue, to get a lower number of
4376 * requests served, and thus to ask for a lower number of
4377 * requests from the request pool, before the busy
4378 * weight-raised queues get served again. This often mitigates
4379 * starvation problems in the presence of heavy write
4380 * workloads and NCQ, thereby guaranteeing a higher
4381 * application and system responsiveness in these hostile
4384 return idling_boosts_thr
&&
4385 bfqd
->wr_busy_queues
== 0;
4389 * For a queue that becomes empty, device idling is allowed only if
4390 * this function returns true for that queue. As a consequence, since
4391 * device idling plays a critical role for both throughput boosting
4392 * and service guarantees, the return value of this function plays a
4393 * critical role as well.
4395 * In a nutshell, this function returns true only if idling is
4396 * beneficial for throughput or, even if detrimental for throughput,
4397 * idling is however necessary to preserve service guarantees (low
4398 * latency, desired throughput distribution, ...). In particular, on
4399 * NCQ-capable devices, this function tries to return false, so as to
4400 * help keep the drives' internal queues full, whenever this helps the
4401 * device boost the throughput without causing any service-guarantee
4404 * Most of the issues taken into account to get the return value of
4405 * this function are not trivial. We discuss these issues in the two
4406 * functions providing the main pieces of information needed by this
4409 static bool bfq_better_to_idle(struct bfq_queue
*bfqq
)
4411 struct bfq_data
*bfqd
= bfqq
->bfqd
;
4412 bool idling_boosts_thr_with_no_issue
, idling_needed_for_service_guar
;
4414 /* No point in idling for bfqq if it won't get requests any longer */
4415 if (unlikely(!bfqq_process_refs(bfqq
)))
4418 if (unlikely(bfqd
->strict_guarantees
))
4422 * Idling is performed only if slice_idle > 0. In addition, we
4425 * (b) bfqq is in the idle io prio class: in this case we do
4426 * not idle because we want to minimize the bandwidth that
4427 * queues in this class can steal to higher-priority queues
4429 if (bfqd
->bfq_slice_idle
== 0 || !bfq_bfqq_sync(bfqq
) ||
4430 bfq_class_idle(bfqq
))
4433 idling_boosts_thr_with_no_issue
=
4434 idling_boosts_thr_without_issues(bfqd
, bfqq
);
4436 idling_needed_for_service_guar
=
4437 idling_needed_for_service_guarantees(bfqd
, bfqq
);
4440 * We have now the two components we need to compute the
4441 * return value of the function, which is true only if idling
4442 * either boosts the throughput (without issues), or is
4443 * necessary to preserve service guarantees.
4445 return idling_boosts_thr_with_no_issue
||
4446 idling_needed_for_service_guar
;
4450 * If the in-service queue is empty but the function bfq_better_to_idle
4451 * returns true, then:
4452 * 1) the queue must remain in service and cannot be expired, and
4453 * 2) the device must be idled to wait for the possible arrival of a new
4454 * request for the queue.
4455 * See the comments on the function bfq_better_to_idle for the reasons
4456 * why performing device idling is the best choice to boost the throughput
4457 * and preserve service guarantees when bfq_better_to_idle itself
4460 static bool bfq_bfqq_must_idle(struct bfq_queue
*bfqq
)
4462 return RB_EMPTY_ROOT(&bfqq
->sort_list
) && bfq_better_to_idle(bfqq
);
4466 * This function chooses the queue from which to pick the next extra
4467 * I/O request to inject, if it finds a compatible queue. See the
4468 * comments on bfq_update_inject_limit() for details on the injection
4469 * mechanism, and for the definitions of the quantities mentioned
4472 static struct bfq_queue
*
4473 bfq_choose_bfqq_for_injection(struct bfq_data
*bfqd
)
4475 struct bfq_queue
*bfqq
, *in_serv_bfqq
= bfqd
->in_service_queue
;
4476 unsigned int limit
= in_serv_bfqq
->inject_limit
;
4479 * - bfqq is not weight-raised and therefore does not carry
4480 * time-critical I/O,
4482 * - regardless of whether bfqq is weight-raised, bfqq has
4483 * however a long think time, during which it can absorb the
4484 * effect of an appropriate number of extra I/O requests
4485 * from other queues (see bfq_update_inject_limit for
4486 * details on the computation of this number);
4487 * then injection can be performed without restrictions.
4489 bool in_serv_always_inject
= in_serv_bfqq
->wr_coeff
== 1 ||
4490 !bfq_bfqq_has_short_ttime(in_serv_bfqq
);
4494 * - the baseline total service time could not be sampled yet,
4495 * so the inject limit happens to be still 0, and
4496 * - a lot of time has elapsed since the plugging of I/O
4497 * dispatching started, so drive speed is being wasted
4499 * then temporarily raise inject limit to one request.
4501 if (limit
== 0 && in_serv_bfqq
->last_serv_time_ns
== 0 &&
4502 bfq_bfqq_wait_request(in_serv_bfqq
) &&
4503 time_is_before_eq_jiffies(bfqd
->last_idling_start_jiffies
+
4504 bfqd
->bfq_slice_idle
)
4508 if (bfqd
->rq_in_driver
>= limit
)
4512 * Linear search of the source queue for injection; but, with
4513 * a high probability, very few steps are needed to find a
4514 * candidate queue, i.e., a queue with enough budget left for
4515 * its next request. In fact:
4516 * - BFQ dynamically updates the budget of every queue so as
4517 * to accommodate the expected backlog of the queue;
4518 * - if a queue gets all its requests dispatched as injected
4519 * service, then the queue is removed from the active list
4520 * (and re-added only if it gets new requests, but then it
4521 * is assigned again enough budget for its new backlog).
4523 list_for_each_entry(bfqq
, &bfqd
->active_list
, bfqq_list
)
4524 if (!RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
4525 (in_serv_always_inject
|| bfqq
->wr_coeff
> 1) &&
4526 bfq_serv_to_charge(bfqq
->next_rq
, bfqq
) <=
4527 bfq_bfqq_budget_left(bfqq
)) {
4529 * Allow for only one large in-flight request
4530 * on non-rotational devices, for the
4531 * following reason. On non-rotationl drives,
4532 * large requests take much longer than
4533 * smaller requests to be served. In addition,
4534 * the drive prefers to serve large requests
4535 * w.r.t. to small ones, if it can choose. So,
4536 * having more than one large requests queued
4537 * in the drive may easily make the next first
4538 * request of the in-service queue wait for so
4539 * long to break bfqq's service guarantees. On
4540 * the bright side, large requests let the
4541 * drive reach a very high throughput, even if
4542 * there is only one in-flight large request
4545 if (blk_queue_nonrot(bfqd
->queue
) &&
4546 blk_rq_sectors(bfqq
->next_rq
) >=
4547 BFQQ_SECT_THR_NONROT
)
4548 limit
= min_t(unsigned int, 1, limit
);
4550 limit
= in_serv_bfqq
->inject_limit
;
4552 if (bfqd
->rq_in_driver
< limit
) {
4553 bfqd
->rqs_injected
= true;
4562 * Select a queue for service. If we have a current queue in service,
4563 * check whether to continue servicing it, or retrieve and set a new one.
4565 static struct bfq_queue
*bfq_select_queue(struct bfq_data
*bfqd
)
4567 struct bfq_queue
*bfqq
;
4568 struct request
*next_rq
;
4569 enum bfqq_expiration reason
= BFQQE_BUDGET_TIMEOUT
;
4571 bfqq
= bfqd
->in_service_queue
;
4575 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: already in-service queue");
4578 * Do not expire bfqq for budget timeout if bfqq may be about
4579 * to enjoy device idling. The reason why, in this case, we
4580 * prevent bfqq from expiring is the same as in the comments
4581 * on the case where bfq_bfqq_must_idle() returns true, in
4582 * bfq_completed_request().
4584 if (bfq_may_expire_for_budg_timeout(bfqq
) &&
4585 !bfq_bfqq_must_idle(bfqq
))
4590 * This loop is rarely executed more than once. Even when it
4591 * happens, it is much more convenient to re-execute this loop
4592 * than to return NULL and trigger a new dispatch to get a
4595 next_rq
= bfqq
->next_rq
;
4597 * If bfqq has requests queued and it has enough budget left to
4598 * serve them, keep the queue, otherwise expire it.
4601 if (bfq_serv_to_charge(next_rq
, bfqq
) >
4602 bfq_bfqq_budget_left(bfqq
)) {
4604 * Expire the queue for budget exhaustion,
4605 * which makes sure that the next budget is
4606 * enough to serve the next request, even if
4607 * it comes from the fifo expired path.
4609 reason
= BFQQE_BUDGET_EXHAUSTED
;
4613 * The idle timer may be pending because we may
4614 * not disable disk idling even when a new request
4617 if (bfq_bfqq_wait_request(bfqq
)) {
4619 * If we get here: 1) at least a new request
4620 * has arrived but we have not disabled the
4621 * timer because the request was too small,
4622 * 2) then the block layer has unplugged
4623 * the device, causing the dispatch to be
4626 * Since the device is unplugged, now the
4627 * requests are probably large enough to
4628 * provide a reasonable throughput.
4629 * So we disable idling.
4631 bfq_clear_bfqq_wait_request(bfqq
);
4632 hrtimer_try_to_cancel(&bfqd
->idle_slice_timer
);
4639 * No requests pending. However, if the in-service queue is idling
4640 * for a new request, or has requests waiting for a completion and
4641 * may idle after their completion, then keep it anyway.
4643 * Yet, inject service from other queues if it boosts
4644 * throughput and is possible.
4646 if (bfq_bfqq_wait_request(bfqq
) ||
4647 (bfqq
->dispatched
!= 0 && bfq_better_to_idle(bfqq
))) {
4648 struct bfq_queue
*async_bfqq
=
4649 bfqq
->bic
&& bfqq
->bic
->bfqq
[0] &&
4650 bfq_bfqq_busy(bfqq
->bic
->bfqq
[0]) &&
4651 bfqq
->bic
->bfqq
[0]->next_rq
?
4652 bfqq
->bic
->bfqq
[0] : NULL
;
4653 struct bfq_queue
*blocked_bfqq
=
4654 !hlist_empty(&bfqq
->woken_list
) ?
4655 container_of(bfqq
->woken_list
.first
,
4661 * The next four mutually-exclusive ifs decide
4662 * whether to try injection, and choose the queue to
4663 * pick an I/O request from.
4665 * The first if checks whether the process associated
4666 * with bfqq has also async I/O pending. If so, it
4667 * injects such I/O unconditionally. Injecting async
4668 * I/O from the same process can cause no harm to the
4669 * process. On the contrary, it can only increase
4670 * bandwidth and reduce latency for the process.
4672 * The second if checks whether there happens to be a
4673 * non-empty waker queue for bfqq, i.e., a queue whose
4674 * I/O needs to be completed for bfqq to receive new
4675 * I/O. This happens, e.g., if bfqq is associated with
4676 * a process that does some sync. A sync generates
4677 * extra blocking I/O, which must be completed before
4678 * the process associated with bfqq can go on with its
4679 * I/O. If the I/O of the waker queue is not served,
4680 * then bfqq remains empty, and no I/O is dispatched,
4681 * until the idle timeout fires for bfqq. This is
4682 * likely to result in lower bandwidth and higher
4683 * latencies for bfqq, and in a severe loss of total
4684 * throughput. The best action to take is therefore to
4685 * serve the waker queue as soon as possible. So do it
4686 * (without relying on the third alternative below for
4687 * eventually serving waker_bfqq's I/O; see the last
4688 * paragraph for further details). This systematic
4689 * injection of I/O from the waker queue does not
4690 * cause any delay to bfqq's I/O. On the contrary,
4691 * next bfqq's I/O is brought forward dramatically,
4692 * for it is not blocked for milliseconds.
4694 * The third if checks whether there is a queue woken
4695 * by bfqq, and currently with pending I/O. Such a
4696 * woken queue does not steal bandwidth from bfqq,
4697 * because it remains soon without I/O if bfqq is not
4698 * served. So there is virtually no risk of loss of
4699 * bandwidth for bfqq if this woken queue has I/O
4700 * dispatched while bfqq is waiting for new I/O.
4702 * The fourth if checks whether bfqq is a queue for
4703 * which it is better to avoid injection. It is so if
4704 * bfqq delivers more throughput when served without
4705 * any further I/O from other queues in the middle, or
4706 * if the service times of bfqq's I/O requests both
4707 * count more than overall throughput, and may be
4708 * easily increased by injection (this happens if bfqq
4709 * has a short think time). If none of these
4710 * conditions holds, then a candidate queue for
4711 * injection is looked for through
4712 * bfq_choose_bfqq_for_injection(). Note that the
4713 * latter may return NULL (for example if the inject
4714 * limit for bfqq is currently 0).
4716 * NOTE: motivation for the second alternative
4718 * Thanks to the way the inject limit is updated in
4719 * bfq_update_has_short_ttime(), it is rather likely
4720 * that, if I/O is being plugged for bfqq and the
4721 * waker queue has pending I/O requests that are
4722 * blocking bfqq's I/O, then the fourth alternative
4723 * above lets the waker queue get served before the
4724 * I/O-plugging timeout fires. So one may deem the
4725 * second alternative superfluous. It is not, because
4726 * the fourth alternative may be way less effective in
4727 * case of a synchronization. For two main
4728 * reasons. First, throughput may be low because the
4729 * inject limit may be too low to guarantee the same
4730 * amount of injected I/O, from the waker queue or
4731 * other queues, that the second alternative
4732 * guarantees (the second alternative unconditionally
4733 * injects a pending I/O request of the waker queue
4734 * for each bfq_dispatch_request()). Second, with the
4735 * fourth alternative, the duration of the plugging,
4736 * i.e., the time before bfqq finally receives new I/O,
4737 * may not be minimized, because the waker queue may
4738 * happen to be served only after other queues.
4741 icq_to_bic(async_bfqq
->next_rq
->elv
.icq
) == bfqq
->bic
&&
4742 bfq_serv_to_charge(async_bfqq
->next_rq
, async_bfqq
) <=
4743 bfq_bfqq_budget_left(async_bfqq
))
4744 bfqq
= bfqq
->bic
->bfqq
[0];
4745 else if (bfqq
->waker_bfqq
&&
4746 bfq_bfqq_busy(bfqq
->waker_bfqq
) &&
4747 bfqq
->waker_bfqq
->next_rq
&&
4748 bfq_serv_to_charge(bfqq
->waker_bfqq
->next_rq
,
4749 bfqq
->waker_bfqq
) <=
4750 bfq_bfqq_budget_left(bfqq
->waker_bfqq
)
4752 bfqq
= bfqq
->waker_bfqq
;
4753 else if (blocked_bfqq
&&
4754 bfq_bfqq_busy(blocked_bfqq
) &&
4755 blocked_bfqq
->next_rq
&&
4756 bfq_serv_to_charge(blocked_bfqq
->next_rq
,
4758 bfq_bfqq_budget_left(blocked_bfqq
)
4760 bfqq
= blocked_bfqq
;
4761 else if (!idling_boosts_thr_without_issues(bfqd
, bfqq
) &&
4762 (bfqq
->wr_coeff
== 1 || bfqd
->wr_busy_queues
> 1 ||
4763 !bfq_bfqq_has_short_ttime(bfqq
)))
4764 bfqq
= bfq_choose_bfqq_for_injection(bfqd
);
4771 reason
= BFQQE_NO_MORE_REQUESTS
;
4773 bfq_bfqq_expire(bfqd
, bfqq
, false, reason
);
4775 bfqq
= bfq_set_in_service_queue(bfqd
);
4777 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: checking new queue");
4782 bfq_log_bfqq(bfqd
, bfqq
, "select_queue: returned this queue");
4784 bfq_log(bfqd
, "select_queue: no queue returned");
4789 static void bfq_update_wr_data(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
4791 struct bfq_entity
*entity
= &bfqq
->entity
;
4793 if (bfqq
->wr_coeff
> 1) { /* queue is being weight-raised */
4794 bfq_log_bfqq(bfqd
, bfqq
,
4795 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
4796 jiffies_to_msecs(jiffies
- bfqq
->last_wr_start_finish
),
4797 jiffies_to_msecs(bfqq
->wr_cur_max_time
),
4799 bfqq
->entity
.weight
, bfqq
->entity
.orig_weight
);
4801 if (entity
->prio_changed
)
4802 bfq_log_bfqq(bfqd
, bfqq
, "WARN: pending prio change");
4805 * If the queue was activated in a burst, or too much
4806 * time has elapsed from the beginning of this
4807 * weight-raising period, then end weight raising.
4809 if (bfq_bfqq_in_large_burst(bfqq
))
4810 bfq_bfqq_end_wr(bfqq
);
4811 else if (time_is_before_jiffies(bfqq
->last_wr_start_finish
+
4812 bfqq
->wr_cur_max_time
)) {
4813 if (bfqq
->wr_cur_max_time
!= bfqd
->bfq_wr_rt_max_time
||
4814 time_is_before_jiffies(bfqq
->wr_start_at_switch_to_srt
+
4815 bfq_wr_duration(bfqd
))) {
4817 * Either in interactive weight
4818 * raising, or in soft_rt weight
4820 * interactive-weight-raising period
4821 * elapsed (so no switch back to
4822 * interactive weight raising).
4824 bfq_bfqq_end_wr(bfqq
);
4826 * soft_rt finishing while still in
4827 * interactive period, switch back to
4828 * interactive weight raising
4830 switch_back_to_interactive_wr(bfqq
, bfqd
);
4831 bfqq
->entity
.prio_changed
= 1;
4834 if (bfqq
->wr_coeff
> 1 &&
4835 bfqq
->wr_cur_max_time
!= bfqd
->bfq_wr_rt_max_time
&&
4836 bfqq
->service_from_wr
> max_service_from_wr
) {
4837 /* see comments on max_service_from_wr */
4838 bfq_bfqq_end_wr(bfqq
);
4842 * To improve latency (for this or other queues), immediately
4843 * update weight both if it must be raised and if it must be
4844 * lowered. Since, entity may be on some active tree here, and
4845 * might have a pending change of its ioprio class, invoke
4846 * next function with the last parameter unset (see the
4847 * comments on the function).
4849 if ((entity
->weight
> entity
->orig_weight
) != (bfqq
->wr_coeff
> 1))
4850 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity
),
4855 * Dispatch next request from bfqq.
4857 static struct request
*bfq_dispatch_rq_from_bfqq(struct bfq_data
*bfqd
,
4858 struct bfq_queue
*bfqq
)
4860 struct request
*rq
= bfqq
->next_rq
;
4861 unsigned long service_to_charge
;
4863 service_to_charge
= bfq_serv_to_charge(rq
, bfqq
);
4865 bfq_bfqq_served(bfqq
, service_to_charge
);
4867 if (bfqq
== bfqd
->in_service_queue
&& bfqd
->wait_dispatch
) {
4868 bfqd
->wait_dispatch
= false;
4869 bfqd
->waited_rq
= rq
;
4872 bfq_dispatch_remove(bfqd
->queue
, rq
);
4874 if (bfqq
!= bfqd
->in_service_queue
)
4878 * If weight raising has to terminate for bfqq, then next
4879 * function causes an immediate update of bfqq's weight,
4880 * without waiting for next activation. As a consequence, on
4881 * expiration, bfqq will be timestamped as if has never been
4882 * weight-raised during this service slot, even if it has
4883 * received part or even most of the service as a
4884 * weight-raised queue. This inflates bfqq's timestamps, which
4885 * is beneficial, as bfqq is then more willing to leave the
4886 * device immediately to possible other weight-raised queues.
4888 bfq_update_wr_data(bfqd
, bfqq
);
4891 * Expire bfqq, pretending that its budget expired, if bfqq
4892 * belongs to CLASS_IDLE and other queues are waiting for
4895 if (!(bfq_tot_busy_queues(bfqd
) > 1 && bfq_class_idle(bfqq
)))
4898 bfq_bfqq_expire(bfqd
, bfqq
, false, BFQQE_BUDGET_EXHAUSTED
);
4904 static bool bfq_has_work(struct blk_mq_hw_ctx
*hctx
)
4906 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
4909 * Avoiding lock: a race on bfqd->busy_queues should cause at
4910 * most a call to dispatch for nothing
4912 return !list_empty_careful(&bfqd
->dispatch
) ||
4913 bfq_tot_busy_queues(bfqd
) > 0;
4916 static struct request
*__bfq_dispatch_request(struct blk_mq_hw_ctx
*hctx
)
4918 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
4919 struct request
*rq
= NULL
;
4920 struct bfq_queue
*bfqq
= NULL
;
4922 if (!list_empty(&bfqd
->dispatch
)) {
4923 rq
= list_first_entry(&bfqd
->dispatch
, struct request
,
4925 list_del_init(&rq
->queuelist
);
4931 * Increment counters here, because this
4932 * dispatch does not follow the standard
4933 * dispatch flow (where counters are
4938 goto inc_in_driver_start_rq
;
4942 * We exploit the bfq_finish_requeue_request hook to
4943 * decrement rq_in_driver, but
4944 * bfq_finish_requeue_request will not be invoked on
4945 * this request. So, to avoid unbalance, just start
4946 * this request, without incrementing rq_in_driver. As
4947 * a negative consequence, rq_in_driver is deceptively
4948 * lower than it should be while this request is in
4949 * service. This may cause bfq_schedule_dispatch to be
4950 * invoked uselessly.
4952 * As for implementing an exact solution, the
4953 * bfq_finish_requeue_request hook, if defined, is
4954 * probably invoked also on this request. So, by
4955 * exploiting this hook, we could 1) increment
4956 * rq_in_driver here, and 2) decrement it in
4957 * bfq_finish_requeue_request. Such a solution would
4958 * let the value of the counter be always accurate,
4959 * but it would entail using an extra interface
4960 * function. This cost seems higher than the benefit,
4961 * being the frequency of non-elevator-private
4962 * requests very low.
4967 bfq_log(bfqd
, "dispatch requests: %d busy queues",
4968 bfq_tot_busy_queues(bfqd
));
4970 if (bfq_tot_busy_queues(bfqd
) == 0)
4974 * Force device to serve one request at a time if
4975 * strict_guarantees is true. Forcing this service scheme is
4976 * currently the ONLY way to guarantee that the request
4977 * service order enforced by the scheduler is respected by a
4978 * queueing device. Otherwise the device is free even to make
4979 * some unlucky request wait for as long as the device
4982 * Of course, serving one request at a time may cause loss of
4985 if (bfqd
->strict_guarantees
&& bfqd
->rq_in_driver
> 0)
4988 bfqq
= bfq_select_queue(bfqd
);
4992 rq
= bfq_dispatch_rq_from_bfqq(bfqd
, bfqq
);
4995 inc_in_driver_start_rq
:
4996 bfqd
->rq_in_driver
++;
4998 rq
->rq_flags
|= RQF_STARTED
;
5004 #ifdef CONFIG_BFQ_CGROUP_DEBUG
5005 static void bfq_update_dispatch_stats(struct request_queue
*q
,
5007 struct bfq_queue
*in_serv_queue
,
5008 bool idle_timer_disabled
)
5010 struct bfq_queue
*bfqq
= rq
? RQ_BFQQ(rq
) : NULL
;
5012 if (!idle_timer_disabled
&& !bfqq
)
5016 * rq and bfqq are guaranteed to exist until this function
5017 * ends, for the following reasons. First, rq can be
5018 * dispatched to the device, and then can be completed and
5019 * freed, only after this function ends. Second, rq cannot be
5020 * merged (and thus freed because of a merge) any longer,
5021 * because it has already started. Thus rq cannot be freed
5022 * before this function ends, and, since rq has a reference to
5023 * bfqq, the same guarantee holds for bfqq too.
5025 * In addition, the following queue lock guarantees that
5026 * bfqq_group(bfqq) exists as well.
5028 spin_lock_irq(&q
->queue_lock
);
5029 if (idle_timer_disabled
)
5031 * Since the idle timer has been disabled,
5032 * in_serv_queue contained some request when
5033 * __bfq_dispatch_request was invoked above, which
5034 * implies that rq was picked exactly from
5035 * in_serv_queue. Thus in_serv_queue == bfqq, and is
5036 * therefore guaranteed to exist because of the above
5039 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue
));
5041 struct bfq_group
*bfqg
= bfqq_group(bfqq
);
5043 bfqg_stats_update_avg_queue_size(bfqg
);
5044 bfqg_stats_set_start_empty_time(bfqg
);
5045 bfqg_stats_update_io_remove(bfqg
, rq
->cmd_flags
);
5047 spin_unlock_irq(&q
->queue_lock
);
5050 static inline void bfq_update_dispatch_stats(struct request_queue
*q
,
5052 struct bfq_queue
*in_serv_queue
,
5053 bool idle_timer_disabled
) {}
5054 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
5056 static struct request
*bfq_dispatch_request(struct blk_mq_hw_ctx
*hctx
)
5058 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
5060 struct bfq_queue
*in_serv_queue
;
5061 bool waiting_rq
, idle_timer_disabled
;
5063 spin_lock_irq(&bfqd
->lock
);
5065 in_serv_queue
= bfqd
->in_service_queue
;
5066 waiting_rq
= in_serv_queue
&& bfq_bfqq_wait_request(in_serv_queue
);
5068 rq
= __bfq_dispatch_request(hctx
);
5070 idle_timer_disabled
=
5071 waiting_rq
&& !bfq_bfqq_wait_request(in_serv_queue
);
5073 spin_unlock_irq(&bfqd
->lock
);
5075 bfq_update_dispatch_stats(hctx
->queue
, rq
, in_serv_queue
,
5076 idle_timer_disabled
);
5082 * Task holds one reference to the queue, dropped when task exits. Each rq
5083 * in-flight on this queue also holds a reference, dropped when rq is freed.
5085 * Scheduler lock must be held here. Recall not to use bfqq after calling
5086 * this function on it.
5088 void bfq_put_queue(struct bfq_queue
*bfqq
)
5090 struct bfq_queue
*item
;
5091 struct hlist_node
*n
;
5092 struct bfq_group
*bfqg
= bfqq_group(bfqq
);
5095 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "put_queue: %p %d",
5102 if (!hlist_unhashed(&bfqq
->burst_list_node
)) {
5103 hlist_del_init(&bfqq
->burst_list_node
);
5105 * Decrement also burst size after the removal, if the
5106 * process associated with bfqq is exiting, and thus
5107 * does not contribute to the burst any longer. This
5108 * decrement helps filter out false positives of large
5109 * bursts, when some short-lived process (often due to
5110 * the execution of commands by some service) happens
5111 * to start and exit while a complex application is
5112 * starting, and thus spawning several processes that
5113 * do I/O (and that *must not* be treated as a large
5114 * burst, see comments on bfq_handle_burst).
5116 * In particular, the decrement is performed only if:
5117 * 1) bfqq is not a merged queue, because, if it is,
5118 * then this free of bfqq is not triggered by the exit
5119 * of the process bfqq is associated with, but exactly
5120 * by the fact that bfqq has just been merged.
5121 * 2) burst_size is greater than 0, to handle
5122 * unbalanced decrements. Unbalanced decrements may
5123 * happen in te following case: bfqq is inserted into
5124 * the current burst list--without incrementing
5125 * bust_size--because of a split, but the current
5126 * burst list is not the burst list bfqq belonged to
5127 * (see comments on the case of a split in
5130 if (bfqq
->bic
&& bfqq
->bfqd
->burst_size
> 0)
5131 bfqq
->bfqd
->burst_size
--;
5135 * bfqq does not exist any longer, so it cannot be woken by
5136 * any other queue, and cannot wake any other queue. Then bfqq
5137 * must be removed from the woken list of its possible waker
5138 * queue, and all queues in the woken list of bfqq must stop
5139 * having a waker queue. Strictly speaking, these updates
5140 * should be performed when bfqq remains with no I/O source
5141 * attached to it, which happens before bfqq gets freed. In
5142 * particular, this happens when the last process associated
5143 * with bfqq exits or gets associated with a different
5144 * queue. However, both events lead to bfqq being freed soon,
5145 * and dangling references would come out only after bfqq gets
5146 * freed. So these updates are done here, as a simple and safe
5147 * way to handle all cases.
5149 /* remove bfqq from woken list */
5150 if (!hlist_unhashed(&bfqq
->woken_list_node
))
5151 hlist_del_init(&bfqq
->woken_list_node
);
5153 /* reset waker for all queues in woken list */
5154 hlist_for_each_entry_safe(item
, n
, &bfqq
->woken_list
,
5156 item
->waker_bfqq
= NULL
;
5157 hlist_del_init(&item
->woken_list_node
);
5160 if (bfqq
->bfqd
&& bfqq
->bfqd
->last_completed_rq_bfqq
== bfqq
)
5161 bfqq
->bfqd
->last_completed_rq_bfqq
= NULL
;
5163 kmem_cache_free(bfq_pool
, bfqq
);
5164 bfqg_and_blkg_put(bfqg
);
5167 static void bfq_put_stable_ref(struct bfq_queue
*bfqq
)
5170 bfq_put_queue(bfqq
);
5173 static void bfq_put_cooperator(struct bfq_queue
*bfqq
)
5175 struct bfq_queue
*__bfqq
, *next
;
5178 * If this queue was scheduled to merge with another queue, be
5179 * sure to drop the reference taken on that queue (and others in
5180 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
5182 __bfqq
= bfqq
->new_bfqq
;
5186 next
= __bfqq
->new_bfqq
;
5187 bfq_put_queue(__bfqq
);
5192 static void bfq_exit_bfqq(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
5194 if (bfqq
== bfqd
->in_service_queue
) {
5195 __bfq_bfqq_expire(bfqd
, bfqq
, BFQQE_BUDGET_TIMEOUT
);
5196 bfq_schedule_dispatch(bfqd
);
5199 bfq_log_bfqq(bfqd
, bfqq
, "exit_bfqq: %p, %d", bfqq
, bfqq
->ref
);
5201 bfq_put_cooperator(bfqq
);
5203 bfq_release_process_ref(bfqd
, bfqq
);
5206 static void bfq_exit_icq_bfqq(struct bfq_io_cq
*bic
, bool is_sync
)
5208 struct bfq_queue
*bfqq
= bic_to_bfqq(bic
, is_sync
);
5209 struct bfq_data
*bfqd
;
5212 bfqd
= bfqq
->bfqd
; /* NULL if scheduler already exited */
5215 unsigned long flags
;
5217 spin_lock_irqsave(&bfqd
->lock
, flags
);
5219 bfq_exit_bfqq(bfqd
, bfqq
);
5220 bic_set_bfqq(bic
, NULL
, is_sync
);
5221 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
5225 static void bfq_exit_icq(struct io_cq
*icq
)
5227 struct bfq_io_cq
*bic
= icq_to_bic(icq
);
5229 if (bic
->stable_merge_bfqq
) {
5230 struct bfq_data
*bfqd
= bic
->stable_merge_bfqq
->bfqd
;
5233 * bfqd is NULL if scheduler already exited, and in
5234 * that case this is the last time bfqq is accessed.
5237 unsigned long flags
;
5239 spin_lock_irqsave(&bfqd
->lock
, flags
);
5240 bfq_put_stable_ref(bic
->stable_merge_bfqq
);
5241 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
5243 bfq_put_stable_ref(bic
->stable_merge_bfqq
);
5247 bfq_exit_icq_bfqq(bic
, true);
5248 bfq_exit_icq_bfqq(bic
, false);
5252 * Update the entity prio values; note that the new values will not
5253 * be used until the next (re)activation.
5256 bfq_set_next_ioprio_data(struct bfq_queue
*bfqq
, struct bfq_io_cq
*bic
)
5258 struct task_struct
*tsk
= current
;
5260 struct bfq_data
*bfqd
= bfqq
->bfqd
;
5265 ioprio_class
= IOPRIO_PRIO_CLASS(bic
->ioprio
);
5266 switch (ioprio_class
) {
5268 pr_err("bdi %s: bfq: bad prio class %d\n",
5269 bdi_dev_name(bfqq
->bfqd
->queue
->backing_dev_info
),
5272 case IOPRIO_CLASS_NONE
:
5274 * No prio set, inherit CPU scheduling settings.
5276 bfqq
->new_ioprio
= task_nice_ioprio(tsk
);
5277 bfqq
->new_ioprio_class
= task_nice_ioclass(tsk
);
5279 case IOPRIO_CLASS_RT
:
5280 bfqq
->new_ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
5281 bfqq
->new_ioprio_class
= IOPRIO_CLASS_RT
;
5283 case IOPRIO_CLASS_BE
:
5284 bfqq
->new_ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
5285 bfqq
->new_ioprio_class
= IOPRIO_CLASS_BE
;
5287 case IOPRIO_CLASS_IDLE
:
5288 bfqq
->new_ioprio_class
= IOPRIO_CLASS_IDLE
;
5289 bfqq
->new_ioprio
= 7;
5293 if (bfqq
->new_ioprio
>= IOPRIO_BE_NR
) {
5294 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
5296 bfqq
->new_ioprio
= IOPRIO_BE_NR
;
5299 bfqq
->entity
.new_weight
= bfq_ioprio_to_weight(bfqq
->new_ioprio
);
5300 bfq_log_bfqq(bfqd
, bfqq
, "new_ioprio %d new_weight %d",
5301 bfqq
->new_ioprio
, bfqq
->entity
.new_weight
);
5302 bfqq
->entity
.prio_changed
= 1;
5305 static struct bfq_queue
*bfq_get_queue(struct bfq_data
*bfqd
,
5306 struct bio
*bio
, bool is_sync
,
5307 struct bfq_io_cq
*bic
,
5310 static void bfq_check_ioprio_change(struct bfq_io_cq
*bic
, struct bio
*bio
)
5312 struct bfq_data
*bfqd
= bic_to_bfqd(bic
);
5313 struct bfq_queue
*bfqq
;
5314 int ioprio
= bic
->icq
.ioc
->ioprio
;
5317 * This condition may trigger on a newly created bic, be sure to
5318 * drop the lock before returning.
5320 if (unlikely(!bfqd
) || likely(bic
->ioprio
== ioprio
))
5323 bic
->ioprio
= ioprio
;
5325 bfqq
= bic_to_bfqq(bic
, false);
5327 bfq_release_process_ref(bfqd
, bfqq
);
5328 bfqq
= bfq_get_queue(bfqd
, bio
, BLK_RW_ASYNC
, bic
, true);
5329 bic_set_bfqq(bic
, bfqq
, false);
5332 bfqq
= bic_to_bfqq(bic
, true);
5334 bfq_set_next_ioprio_data(bfqq
, bic
);
5337 static void bfq_init_bfqq(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
5338 struct bfq_io_cq
*bic
, pid_t pid
, int is_sync
)
5340 u64 now_ns
= ktime_get_ns();
5342 RB_CLEAR_NODE(&bfqq
->entity
.rb_node
);
5343 INIT_LIST_HEAD(&bfqq
->fifo
);
5344 INIT_HLIST_NODE(&bfqq
->burst_list_node
);
5345 INIT_HLIST_NODE(&bfqq
->woken_list_node
);
5346 INIT_HLIST_HEAD(&bfqq
->woken_list
);
5352 bfq_set_next_ioprio_data(bfqq
, bic
);
5356 * No need to mark as has_short_ttime if in
5357 * idle_class, because no device idling is performed
5358 * for queues in idle class
5360 if (!bfq_class_idle(bfqq
))
5361 /* tentatively mark as has_short_ttime */
5362 bfq_mark_bfqq_has_short_ttime(bfqq
);
5363 bfq_mark_bfqq_sync(bfqq
);
5364 bfq_mark_bfqq_just_created(bfqq
);
5366 bfq_clear_bfqq_sync(bfqq
);
5368 /* set end request to minus infinity from now */
5369 bfqq
->ttime
.last_end_request
= now_ns
+ 1;
5371 bfqq
->creation_time
= jiffies
;
5373 bfqq
->io_start_time
= now_ns
;
5375 bfq_mark_bfqq_IO_bound(bfqq
);
5379 /* Tentative initial value to trade off between thr and lat */
5380 bfqq
->max_budget
= (2 * bfq_max_budget(bfqd
)) / 3;
5381 bfqq
->budget_timeout
= bfq_smallest_from_now();
5384 bfqq
->last_wr_start_finish
= jiffies
;
5385 bfqq
->wr_start_at_switch_to_srt
= bfq_smallest_from_now();
5386 bfqq
->split_time
= bfq_smallest_from_now();
5389 * To not forget the possibly high bandwidth consumed by a
5390 * process/queue in the recent past,
5391 * bfq_bfqq_softrt_next_start() returns a value at least equal
5392 * to the current value of bfqq->soft_rt_next_start (see
5393 * comments on bfq_bfqq_softrt_next_start). Set
5394 * soft_rt_next_start to now, to mean that bfqq has consumed
5395 * no bandwidth so far.
5397 bfqq
->soft_rt_next_start
= jiffies
;
5399 /* first request is almost certainly seeky */
5400 bfqq
->seek_history
= 1;
5403 static struct bfq_queue
**bfq_async_queue_prio(struct bfq_data
*bfqd
,
5404 struct bfq_group
*bfqg
,
5405 int ioprio_class
, int ioprio
)
5407 switch (ioprio_class
) {
5408 case IOPRIO_CLASS_RT
:
5409 return &bfqg
->async_bfqq
[0][ioprio
];
5410 case IOPRIO_CLASS_NONE
:
5411 ioprio
= IOPRIO_NORM
;
5413 case IOPRIO_CLASS_BE
:
5414 return &bfqg
->async_bfqq
[1][ioprio
];
5415 case IOPRIO_CLASS_IDLE
:
5416 return &bfqg
->async_idle_bfqq
;
5422 static struct bfq_queue
*
5423 bfq_do_early_stable_merge(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
5424 struct bfq_io_cq
*bic
,
5425 struct bfq_queue
*last_bfqq_created
)
5427 struct bfq_queue
*new_bfqq
=
5428 bfq_setup_merge(bfqq
, last_bfqq_created
);
5434 new_bfqq
->bic
->stably_merged
= true;
5435 bic
->stably_merged
= true;
5438 * Reusing merge functions. This implies that
5439 * bfqq->bic must be set too, for
5440 * bfq_merge_bfqqs to correctly save bfqq's
5441 * state before killing it.
5444 bfq_merge_bfqqs(bfqd
, bic
, bfqq
, new_bfqq
);
5450 * Many throughput-sensitive workloads are made of several parallel
5451 * I/O flows, with all flows generated by the same application, or
5452 * more generically by the same task (e.g., system boot). The most
5453 * counterproductive action with these workloads is plugging I/O
5454 * dispatch when one of the bfq_queues associated with these flows
5455 * remains temporarily empty.
5457 * To avoid this plugging, BFQ has been using a burst-handling
5458 * mechanism for years now. This mechanism has proven effective for
5459 * throughput, and not detrimental for service guarantees. The
5460 * following function pushes this mechanism a little bit further,
5461 * basing on the following two facts.
5463 * First, all the I/O flows of a the same application or task
5464 * contribute to the execution/completion of that common application
5465 * or task. So the performance figures that matter are total
5466 * throughput of the flows and task-wide I/O latency. In particular,
5467 * these flows do not need to be protected from each other, in terms
5468 * of individual bandwidth or latency.
5470 * Second, the above fact holds regardless of the number of flows.
5472 * Putting these two facts together, this commits merges stably the
5473 * bfq_queues associated with these I/O flows, i.e., with the
5474 * processes that generate these IO/ flows, regardless of how many the
5475 * involved processes are.
5477 * To decide whether a set of bfq_queues is actually associated with
5478 * the I/O flows of a common application or task, and to merge these
5479 * queues stably, this function operates as follows: given a bfq_queue,
5480 * say Q2, currently being created, and the last bfq_queue, say Q1,
5481 * created before Q2, Q2 is merged stably with Q1 if
5482 * - very little time has elapsed since when Q1 was created
5483 * - Q2 has the same ioprio as Q1
5484 * - Q2 belongs to the same group as Q1
5486 * Merging bfq_queues also reduces scheduling overhead. A fio test
5487 * with ten random readers on /dev/nullb shows a throughput boost of
5488 * 40%, with a quadcore. Since BFQ's execution time amounts to ~50% of
5489 * the total per-request processing time, the above throughput boost
5490 * implies that BFQ's overhead is reduced by more than 50%.
5492 * This new mechanism most certainly obsoletes the current
5493 * burst-handling heuristics. We keep those heuristics for the moment.
5495 static struct bfq_queue
*bfq_do_or_sched_stable_merge(struct bfq_data
*bfqd
,
5496 struct bfq_queue
*bfqq
,
5497 struct bfq_io_cq
*bic
)
5499 struct bfq_queue
**source_bfqq
= bfqq
->entity
.parent
?
5500 &bfqq
->entity
.parent
->last_bfqq_created
:
5501 &bfqd
->last_bfqq_created
;
5503 struct bfq_queue
*last_bfqq_created
= *source_bfqq
;
5506 * If last_bfqq_created has not been set yet, then init it. If
5507 * it has been set already, but too long ago, then move it
5508 * forward to bfqq. Finally, move also if bfqq belongs to a
5509 * different group than last_bfqq_created, or if bfqq has a
5510 * different ioprio or ioprio_class. If none of these
5511 * conditions holds true, then try an early stable merge or
5512 * schedule a delayed stable merge.
5514 * A delayed merge is scheduled (instead of performing an
5515 * early merge), in case bfqq might soon prove to be more
5516 * throughput-beneficial if not merged. Currently this is
5517 * possible only if bfqd is rotational with no queueing. For
5518 * such a drive, not merging bfqq is better for throughput if
5519 * bfqq happens to contain sequential I/O. So, we wait a
5520 * little bit for enough I/O to flow through bfqq. After that,
5521 * if such an I/O is sequential, then the merge is
5522 * canceled. Otherwise the merge is finally performed.
5524 if (!last_bfqq_created
||
5525 time_before(last_bfqq_created
->creation_time
+
5526 msecs_to_jiffies(bfq_activation_stable_merging
),
5527 bfqq
->creation_time
) ||
5528 bfqq
->entity
.parent
!= last_bfqq_created
->entity
.parent
||
5529 bfqq
->ioprio
!= last_bfqq_created
->ioprio
||
5530 bfqq
->ioprio_class
!= last_bfqq_created
->ioprio_class
)
5531 *source_bfqq
= bfqq
;
5532 else if (time_after_eq(last_bfqq_created
->creation_time
+
5533 bfqd
->bfq_burst_interval
,
5534 bfqq
->creation_time
)) {
5535 if (likely(bfqd
->nonrot_with_queueing
))
5537 * With this type of drive, leaving
5538 * bfqq alone may provide no
5539 * throughput benefits compared with
5540 * merging bfqq. So merge bfqq now.
5542 bfqq
= bfq_do_early_stable_merge(bfqd
, bfqq
,
5545 else { /* schedule tentative stable merge */
5547 * get reference on last_bfqq_created,
5548 * to prevent it from being freed,
5549 * until we decide whether to merge
5551 last_bfqq_created
->ref
++;
5553 * need to keep track of stable refs, to
5554 * compute process refs correctly
5556 last_bfqq_created
->stable_ref
++;
5558 * Record the bfqq to merge to.
5560 bic
->stable_merge_bfqq
= last_bfqq_created
;
5568 static struct bfq_queue
*bfq_get_queue(struct bfq_data
*bfqd
,
5569 struct bio
*bio
, bool is_sync
,
5570 struct bfq_io_cq
*bic
,
5573 const int ioprio
= IOPRIO_PRIO_DATA(bic
->ioprio
);
5574 const int ioprio_class
= IOPRIO_PRIO_CLASS(bic
->ioprio
);
5575 struct bfq_queue
**async_bfqq
= NULL
;
5576 struct bfq_queue
*bfqq
;
5577 struct bfq_group
*bfqg
;
5581 bfqg
= bfq_find_set_group(bfqd
, __bio_blkcg(bio
));
5583 bfqq
= &bfqd
->oom_bfqq
;
5588 async_bfqq
= bfq_async_queue_prio(bfqd
, bfqg
, ioprio_class
,
5595 bfqq
= kmem_cache_alloc_node(bfq_pool
,
5596 GFP_NOWAIT
| __GFP_ZERO
| __GFP_NOWARN
,
5600 bfq_init_bfqq(bfqd
, bfqq
, bic
, current
->pid
,
5602 bfq_init_entity(&bfqq
->entity
, bfqg
);
5603 bfq_log_bfqq(bfqd
, bfqq
, "allocated");
5605 bfqq
= &bfqd
->oom_bfqq
;
5606 bfq_log_bfqq(bfqd
, bfqq
, "using oom bfqq");
5611 * Pin the queue now that it's allocated, scheduler exit will
5616 * Extra group reference, w.r.t. sync
5617 * queue. This extra reference is removed
5618 * only if bfqq->bfqg disappears, to
5619 * guarantee that this queue is not freed
5620 * until its group goes away.
5622 bfq_log_bfqq(bfqd
, bfqq
, "get_queue, bfqq not in async: %p, %d",
5628 bfqq
->ref
++; /* get a process reference to this queue */
5630 if (bfqq
!= &bfqd
->oom_bfqq
&& is_sync
&& !respawn
)
5631 bfqq
= bfq_do_or_sched_stable_merge(bfqd
, bfqq
, bic
);
5637 static void bfq_update_io_thinktime(struct bfq_data
*bfqd
,
5638 struct bfq_queue
*bfqq
)
5640 struct bfq_ttime
*ttime
= &bfqq
->ttime
;
5644 * We are really interested in how long it takes for the queue to
5645 * become busy when there is no outstanding IO for this queue. So
5646 * ignore cases when the bfq queue has already IO queued.
5648 if (bfqq
->dispatched
|| bfq_bfqq_busy(bfqq
))
5650 elapsed
= ktime_get_ns() - bfqq
->ttime
.last_end_request
;
5651 elapsed
= min_t(u64
, elapsed
, 2ULL * bfqd
->bfq_slice_idle
);
5653 ttime
->ttime_samples
= (7*ttime
->ttime_samples
+ 256) / 8;
5654 ttime
->ttime_total
= div_u64(7*ttime
->ttime_total
+ 256*elapsed
, 8);
5655 ttime
->ttime_mean
= div64_ul(ttime
->ttime_total
+ 128,
5656 ttime
->ttime_samples
);
5660 bfq_update_io_seektime(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
5663 bfqq
->seek_history
<<= 1;
5664 bfqq
->seek_history
|= BFQ_RQ_SEEKY(bfqd
, bfqq
->last_request_pos
, rq
);
5666 if (bfqq
->wr_coeff
> 1 &&
5667 bfqq
->wr_cur_max_time
== bfqd
->bfq_wr_rt_max_time
&&
5668 BFQQ_TOTALLY_SEEKY(bfqq
)) {
5669 if (time_is_before_jiffies(bfqq
->wr_start_at_switch_to_srt
+
5670 bfq_wr_duration(bfqd
))) {
5672 * In soft_rt weight raising with the
5673 * interactive-weight-raising period
5674 * elapsed (so no switch back to
5675 * interactive weight raising).
5677 bfq_bfqq_end_wr(bfqq
);
5679 * stopping soft_rt weight raising
5680 * while still in interactive period,
5681 * switch back to interactive weight
5684 switch_back_to_interactive_wr(bfqq
, bfqd
);
5685 bfqq
->entity
.prio_changed
= 1;
5690 static void bfq_update_has_short_ttime(struct bfq_data
*bfqd
,
5691 struct bfq_queue
*bfqq
,
5692 struct bfq_io_cq
*bic
)
5694 bool has_short_ttime
= true, state_changed
;
5697 * No need to update has_short_ttime if bfqq is async or in
5698 * idle io prio class, or if bfq_slice_idle is zero, because
5699 * no device idling is performed for bfqq in this case.
5701 if (!bfq_bfqq_sync(bfqq
) || bfq_class_idle(bfqq
) ||
5702 bfqd
->bfq_slice_idle
== 0)
5705 /* Idle window just restored, statistics are meaningless. */
5706 if (time_is_after_eq_jiffies(bfqq
->split_time
+
5707 bfqd
->bfq_wr_min_idle_time
))
5710 /* Think time is infinite if no process is linked to
5711 * bfqq. Otherwise check average think time to decide whether
5712 * to mark as has_short_ttime. To this goal, compare average
5713 * think time with half the I/O-plugging timeout.
5715 if (atomic_read(&bic
->icq
.ioc
->active_ref
) == 0 ||
5716 (bfq_sample_valid(bfqq
->ttime
.ttime_samples
) &&
5717 bfqq
->ttime
.ttime_mean
> bfqd
->bfq_slice_idle
>>1))
5718 has_short_ttime
= false;
5720 state_changed
= has_short_ttime
!= bfq_bfqq_has_short_ttime(bfqq
);
5722 if (has_short_ttime
)
5723 bfq_mark_bfqq_has_short_ttime(bfqq
);
5725 bfq_clear_bfqq_has_short_ttime(bfqq
);
5728 * Until the base value for the total service time gets
5729 * finally computed for bfqq, the inject limit does depend on
5730 * the think-time state (short|long). In particular, the limit
5731 * is 0 or 1 if the think time is deemed, respectively, as
5732 * short or long (details in the comments in
5733 * bfq_update_inject_limit()). Accordingly, the next
5734 * instructions reset the inject limit if the think-time state
5735 * has changed and the above base value is still to be
5738 * However, the reset is performed only if more than 100 ms
5739 * have elapsed since the last update of the inject limit, or
5740 * (inclusive) if the change is from short to long think
5741 * time. The reason for this waiting is as follows.
5743 * bfqq may have a long think time because of a
5744 * synchronization with some other queue, i.e., because the
5745 * I/O of some other queue may need to be completed for bfqq
5746 * to receive new I/O. Details in the comments on the choice
5747 * of the queue for injection in bfq_select_queue().
5749 * As stressed in those comments, if such a synchronization is
5750 * actually in place, then, without injection on bfqq, the
5751 * blocking I/O cannot happen to served while bfqq is in
5752 * service. As a consequence, if bfqq is granted
5753 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
5754 * is dispatched, until the idle timeout fires. This is likely
5755 * to result in lower bandwidth and higher latencies for bfqq,
5756 * and in a severe loss of total throughput.
5758 * On the opposite end, a non-zero inject limit may allow the
5759 * I/O that blocks bfqq to be executed soon, and therefore
5760 * bfqq to receive new I/O soon.
5762 * But, if the blocking gets actually eliminated, then the
5763 * next think-time sample for bfqq may be very low. This in
5764 * turn may cause bfqq's think time to be deemed
5765 * short. Without the 100 ms barrier, this new state change
5766 * would cause the body of the next if to be executed
5767 * immediately. But this would set to 0 the inject
5768 * limit. Without injection, the blocking I/O would cause the
5769 * think time of bfqq to become long again, and therefore the
5770 * inject limit to be raised again, and so on. The only effect
5771 * of such a steady oscillation between the two think-time
5772 * states would be to prevent effective injection on bfqq.
5774 * In contrast, if the inject limit is not reset during such a
5775 * long time interval as 100 ms, then the number of short
5776 * think time samples can grow significantly before the reset
5777 * is performed. As a consequence, the think time state can
5778 * become stable before the reset. Therefore there will be no
5779 * state change when the 100 ms elapse, and no reset of the
5780 * inject limit. The inject limit remains steadily equal to 1
5781 * both during and after the 100 ms. So injection can be
5782 * performed at all times, and throughput gets boosted.
5784 * An inject limit equal to 1 is however in conflict, in
5785 * general, with the fact that the think time of bfqq is
5786 * short, because injection may be likely to delay bfqq's I/O
5787 * (as explained in the comments in
5788 * bfq_update_inject_limit()). But this does not happen in
5789 * this special case, because bfqq's low think time is due to
5790 * an effective handling of a synchronization, through
5791 * injection. In this special case, bfqq's I/O does not get
5792 * delayed by injection; on the contrary, bfqq's I/O is
5793 * brought forward, because it is not blocked for
5796 * In addition, serving the blocking I/O much sooner, and much
5797 * more frequently than once per I/O-plugging timeout, makes
5798 * it much quicker to detect a waker queue (the concept of
5799 * waker queue is defined in the comments in
5800 * bfq_add_request()). This makes it possible to start sooner
5801 * to boost throughput more effectively, by injecting the I/O
5802 * of the waker queue unconditionally on every
5803 * bfq_dispatch_request().
5805 * One last, important benefit of not resetting the inject
5806 * limit before 100 ms is that, during this time interval, the
5807 * base value for the total service time is likely to get
5808 * finally computed for bfqq, freeing the inject limit from
5809 * its relation with the think time.
5811 if (state_changed
&& bfqq
->last_serv_time_ns
== 0 &&
5812 (time_is_before_eq_jiffies(bfqq
->decrease_time_jif
+
5813 msecs_to_jiffies(100)) ||
5815 bfq_reset_inject_limit(bfqd
, bfqq
);
5819 * Called when a new fs request (rq) is added to bfqq. Check if there's
5820 * something we should do about it.
5822 static void bfq_rq_enqueued(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
,
5825 if (rq
->cmd_flags
& REQ_META
)
5826 bfqq
->meta_pending
++;
5828 bfqq
->last_request_pos
= blk_rq_pos(rq
) + blk_rq_sectors(rq
);
5830 if (bfqq
== bfqd
->in_service_queue
&& bfq_bfqq_wait_request(bfqq
)) {
5831 bool small_req
= bfqq
->queued
[rq_is_sync(rq
)] == 1 &&
5832 blk_rq_sectors(rq
) < 32;
5833 bool budget_timeout
= bfq_bfqq_budget_timeout(bfqq
);
5836 * There is just this request queued: if
5837 * - the request is small, and
5838 * - we are idling to boost throughput, and
5839 * - the queue is not to be expired,
5842 * In this way, if the device is being idled to wait
5843 * for a new request from the in-service queue, we
5844 * avoid unplugging the device and committing the
5845 * device to serve just a small request. In contrast
5846 * we wait for the block layer to decide when to
5847 * unplug the device: hopefully, new requests will be
5848 * merged to this one quickly, then the device will be
5849 * unplugged and larger requests will be dispatched.
5851 if (small_req
&& idling_boosts_thr_without_issues(bfqd
, bfqq
) &&
5856 * A large enough request arrived, or idling is being
5857 * performed to preserve service guarantees, or
5858 * finally the queue is to be expired: in all these
5859 * cases disk idling is to be stopped, so clear
5860 * wait_request flag and reset timer.
5862 bfq_clear_bfqq_wait_request(bfqq
);
5863 hrtimer_try_to_cancel(&bfqd
->idle_slice_timer
);
5866 * The queue is not empty, because a new request just
5867 * arrived. Hence we can safely expire the queue, in
5868 * case of budget timeout, without risking that the
5869 * timestamps of the queue are not updated correctly.
5870 * See [1] for more details.
5873 bfq_bfqq_expire(bfqd
, bfqq
, false,
5874 BFQQE_BUDGET_TIMEOUT
);
5878 /* returns true if it causes the idle timer to be disabled */
5879 static bool __bfq_insert_request(struct bfq_data
*bfqd
, struct request
*rq
)
5881 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
),
5882 *new_bfqq
= bfq_setup_cooperator(bfqd
, bfqq
, rq
, true,
5884 bool waiting
, idle_timer_disabled
= false;
5888 * Release the request's reference to the old bfqq
5889 * and make sure one is taken to the shared queue.
5891 new_bfqq
->allocated
++;
5895 * If the bic associated with the process
5896 * issuing this request still points to bfqq
5897 * (and thus has not been already redirected
5898 * to new_bfqq or even some other bfq_queue),
5899 * then complete the merge and redirect it to
5902 if (bic_to_bfqq(RQ_BIC(rq
), 1) == bfqq
)
5903 bfq_merge_bfqqs(bfqd
, RQ_BIC(rq
),
5906 bfq_clear_bfqq_just_created(bfqq
);
5908 * rq is about to be enqueued into new_bfqq,
5909 * release rq reference on bfqq
5911 bfq_put_queue(bfqq
);
5912 rq
->elv
.priv
[1] = new_bfqq
;
5916 bfq_update_io_thinktime(bfqd
, bfqq
);
5917 bfq_update_has_short_ttime(bfqd
, bfqq
, RQ_BIC(rq
));
5918 bfq_update_io_seektime(bfqd
, bfqq
, rq
);
5920 waiting
= bfqq
&& bfq_bfqq_wait_request(bfqq
);
5921 bfq_add_request(rq
);
5922 idle_timer_disabled
= waiting
&& !bfq_bfqq_wait_request(bfqq
);
5924 rq
->fifo_time
= ktime_get_ns() + bfqd
->bfq_fifo_expire
[rq_is_sync(rq
)];
5925 list_add_tail(&rq
->queuelist
, &bfqq
->fifo
);
5927 bfq_rq_enqueued(bfqd
, bfqq
, rq
);
5929 return idle_timer_disabled
;
5932 #ifdef CONFIG_BFQ_CGROUP_DEBUG
5933 static void bfq_update_insert_stats(struct request_queue
*q
,
5934 struct bfq_queue
*bfqq
,
5935 bool idle_timer_disabled
,
5936 unsigned int cmd_flags
)
5942 * bfqq still exists, because it can disappear only after
5943 * either it is merged with another queue, or the process it
5944 * is associated with exits. But both actions must be taken by
5945 * the same process currently executing this flow of
5948 * In addition, the following queue lock guarantees that
5949 * bfqq_group(bfqq) exists as well.
5951 spin_lock_irq(&q
->queue_lock
);
5952 bfqg_stats_update_io_add(bfqq_group(bfqq
), bfqq
, cmd_flags
);
5953 if (idle_timer_disabled
)
5954 bfqg_stats_update_idle_time(bfqq_group(bfqq
));
5955 spin_unlock_irq(&q
->queue_lock
);
5958 static inline void bfq_update_insert_stats(struct request_queue
*q
,
5959 struct bfq_queue
*bfqq
,
5960 bool idle_timer_disabled
,
5961 unsigned int cmd_flags
) {}
5962 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
5964 static void bfq_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
5967 struct request_queue
*q
= hctx
->queue
;
5968 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
5969 struct bfq_queue
*bfqq
;
5970 bool idle_timer_disabled
= false;
5971 unsigned int cmd_flags
;
5974 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5975 if (!cgroup_subsys_on_dfl(io_cgrp_subsys
) && rq
->bio
)
5976 bfqg_stats_update_legacy_io(q
, rq
);
5978 spin_lock_irq(&bfqd
->lock
);
5979 if (blk_mq_sched_try_insert_merge(q
, rq
, &free
)) {
5980 spin_unlock_irq(&bfqd
->lock
);
5981 blk_mq_free_requests(&free
);
5985 spin_unlock_irq(&bfqd
->lock
);
5987 trace_block_rq_insert(rq
);
5989 spin_lock_irq(&bfqd
->lock
);
5990 bfqq
= bfq_init_rq(rq
);
5993 * Reqs with at_head or passthrough flags set are to be put
5994 * directly into dispatch list. Additional case for putting rq
5995 * directly into the dispatch queue: the only active
5996 * bfq_queues are bfqq and either its waker bfq_queue or one
5997 * of its woken bfq_queues. The rationale behind this
5998 * additional condition is as follows:
5999 * - consider a bfq_queue, say Q1, detected as a waker of
6000 * another bfq_queue, say Q2
6001 * - by definition of a waker, Q1 blocks the I/O of Q2, i.e.,
6002 * some I/O of Q1 needs to be completed for new I/O of Q2
6003 * to arrive. A notable example of waker is journald
6004 * - so, Q1 and Q2 are in any respect the queues of two
6005 * cooperating processes (or of two cooperating sets of
6006 * processes): the goal of Q1's I/O is doing what needs to
6007 * be done so that new Q2's I/O can finally be
6008 * issued. Therefore, if the service of Q1's I/O is delayed,
6009 * then Q2's I/O is delayed too. Conversely, if Q2's I/O is
6010 * delayed, the goal of Q1's I/O is hindered.
6011 * - as a consequence, if some I/O of Q1/Q2 arrives while
6012 * Q2/Q1 is the only queue in service, there is absolutely
6013 * no point in delaying the service of such an I/O. The
6014 * only possible result is a throughput loss
6015 * - so, when the above condition holds, the best option is to
6016 * have the new I/O dispatched as soon as possible
6017 * - the most effective and efficient way to attain the above
6018 * goal is to put the new I/O directly in the dispatch
6020 * - as an additional restriction, Q1 and Q2 must be the only
6021 * busy queues for this commit to put the I/O of Q2/Q1 in
6022 * the dispatch list. This is necessary, because, if also
6023 * other queues are waiting for service, then putting new
6024 * I/O directly in the dispatch list may evidently cause a
6025 * violation of service guarantees for the other queues
6028 (bfqq
!= bfqd
->in_service_queue
&&
6029 bfqd
->in_service_queue
!= NULL
&&
6030 bfq_tot_busy_queues(bfqd
) == 1 + bfq_bfqq_busy(bfqq
) &&
6031 (bfqq
->waker_bfqq
== bfqd
->in_service_queue
||
6032 bfqd
->in_service_queue
->waker_bfqq
== bfqq
)) || at_head
) {
6034 list_add(&rq
->queuelist
, &bfqd
->dispatch
);
6036 list_add_tail(&rq
->queuelist
, &bfqd
->dispatch
);
6038 idle_timer_disabled
= __bfq_insert_request(bfqd
, rq
);
6040 * Update bfqq, because, if a queue merge has occurred
6041 * in __bfq_insert_request, then rq has been
6042 * redirected into a new queue.
6046 if (rq_mergeable(rq
)) {
6047 elv_rqhash_add(q
, rq
);
6054 * Cache cmd_flags before releasing scheduler lock, because rq
6055 * may disappear afterwards (for example, because of a request
6058 cmd_flags
= rq
->cmd_flags
;
6060 spin_unlock_irq(&bfqd
->lock
);
6062 bfq_update_insert_stats(q
, bfqq
, idle_timer_disabled
,
6066 static void bfq_insert_requests(struct blk_mq_hw_ctx
*hctx
,
6067 struct list_head
*list
, bool at_head
)
6069 while (!list_empty(list
)) {
6072 rq
= list_first_entry(list
, struct request
, queuelist
);
6073 list_del_init(&rq
->queuelist
);
6074 bfq_insert_request(hctx
, rq
, at_head
);
6078 static void bfq_update_hw_tag(struct bfq_data
*bfqd
)
6080 struct bfq_queue
*bfqq
= bfqd
->in_service_queue
;
6082 bfqd
->max_rq_in_driver
= max_t(int, bfqd
->max_rq_in_driver
,
6083 bfqd
->rq_in_driver
);
6085 if (bfqd
->hw_tag
== 1)
6089 * This sample is valid if the number of outstanding requests
6090 * is large enough to allow a queueing behavior. Note that the
6091 * sum is not exact, as it's not taking into account deactivated
6094 if (bfqd
->rq_in_driver
+ bfqd
->queued
<= BFQ_HW_QUEUE_THRESHOLD
)
6098 * If active queue hasn't enough requests and can idle, bfq might not
6099 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
6102 if (bfqq
&& bfq_bfqq_has_short_ttime(bfqq
) &&
6103 bfqq
->dispatched
+ bfqq
->queued
[0] + bfqq
->queued
[1] <
6104 BFQ_HW_QUEUE_THRESHOLD
&&
6105 bfqd
->rq_in_driver
< BFQ_HW_QUEUE_THRESHOLD
)
6108 if (bfqd
->hw_tag_samples
++ < BFQ_HW_QUEUE_SAMPLES
)
6111 bfqd
->hw_tag
= bfqd
->max_rq_in_driver
> BFQ_HW_QUEUE_THRESHOLD
;
6112 bfqd
->max_rq_in_driver
= 0;
6113 bfqd
->hw_tag_samples
= 0;
6115 bfqd
->nonrot_with_queueing
=
6116 blk_queue_nonrot(bfqd
->queue
) && bfqd
->hw_tag
;
6119 static void bfq_completed_request(struct bfq_queue
*bfqq
, struct bfq_data
*bfqd
)
6124 bfq_update_hw_tag(bfqd
);
6126 bfqd
->rq_in_driver
--;
6129 if (!bfqq
->dispatched
&& !bfq_bfqq_busy(bfqq
)) {
6131 * Set budget_timeout (which we overload to store the
6132 * time at which the queue remains with no backlog and
6133 * no outstanding request; used by the weight-raising
6136 bfqq
->budget_timeout
= jiffies
;
6138 bfq_weights_tree_remove(bfqd
, bfqq
);
6141 now_ns
= ktime_get_ns();
6143 bfqq
->ttime
.last_end_request
= now_ns
;
6146 * Using us instead of ns, to get a reasonable precision in
6147 * computing rate in next check.
6149 delta_us
= div_u64(now_ns
- bfqd
->last_completion
, NSEC_PER_USEC
);
6152 * If the request took rather long to complete, and, according
6153 * to the maximum request size recorded, this completion latency
6154 * implies that the request was certainly served at a very low
6155 * rate (less than 1M sectors/sec), then the whole observation
6156 * interval that lasts up to this time instant cannot be a
6157 * valid time interval for computing a new peak rate. Invoke
6158 * bfq_update_rate_reset to have the following three steps
6160 * - close the observation interval at the last (previous)
6161 * request dispatch or completion
6162 * - compute rate, if possible, for that observation interval
6163 * - reset to zero samples, which will trigger a proper
6164 * re-initialization of the observation interval on next
6167 if (delta_us
> BFQ_MIN_TT
/NSEC_PER_USEC
&&
6168 (bfqd
->last_rq_max_size
<<BFQ_RATE_SHIFT
)/delta_us
<
6169 1UL<<(BFQ_RATE_SHIFT
- 10))
6170 bfq_update_rate_reset(bfqd
, NULL
);
6171 bfqd
->last_completion
= now_ns
;
6173 * Shared queues are likely to receive I/O at a high
6174 * rate. This may deceptively let them be considered as wakers
6175 * of other queues. But a false waker will unjustly steal
6176 * bandwidth to its supposedly woken queue. So considering
6177 * also shared queues in the waking mechanism may cause more
6178 * control troubles than throughput benefits. Then reset
6179 * last_completed_rq_bfqq if bfqq is a shared queue.
6181 if (!bfq_bfqq_coop(bfqq
))
6182 bfqd
->last_completed_rq_bfqq
= bfqq
;
6184 bfqd
->last_completed_rq_bfqq
= NULL
;
6187 * If we are waiting to discover whether the request pattern
6188 * of the task associated with the queue is actually
6189 * isochronous, and both requisites for this condition to hold
6190 * are now satisfied, then compute soft_rt_next_start (see the
6191 * comments on the function bfq_bfqq_softrt_next_start()). We
6192 * do not compute soft_rt_next_start if bfqq is in interactive
6193 * weight raising (see the comments in bfq_bfqq_expire() for
6194 * an explanation). We schedule this delayed update when bfqq
6195 * expires, if it still has in-flight requests.
6197 if (bfq_bfqq_softrt_update(bfqq
) && bfqq
->dispatched
== 0 &&
6198 RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
6199 bfqq
->wr_coeff
!= bfqd
->bfq_wr_coeff
)
6200 bfqq
->soft_rt_next_start
=
6201 bfq_bfqq_softrt_next_start(bfqd
, bfqq
);
6204 * If this is the in-service queue, check if it needs to be expired,
6205 * or if we want to idle in case it has no pending requests.
6207 if (bfqd
->in_service_queue
== bfqq
) {
6208 if (bfq_bfqq_must_idle(bfqq
)) {
6209 if (bfqq
->dispatched
== 0)
6210 bfq_arm_slice_timer(bfqd
);
6212 * If we get here, we do not expire bfqq, even
6213 * if bfqq was in budget timeout or had no
6214 * more requests (as controlled in the next
6215 * conditional instructions). The reason for
6216 * not expiring bfqq is as follows.
6218 * Here bfqq->dispatched > 0 holds, but
6219 * bfq_bfqq_must_idle() returned true. This
6220 * implies that, even if no request arrives
6221 * for bfqq before bfqq->dispatched reaches 0,
6222 * bfqq will, however, not be expired on the
6223 * completion event that causes bfqq->dispatch
6224 * to reach zero. In contrast, on this event,
6225 * bfqq will start enjoying device idling
6226 * (I/O-dispatch plugging).
6228 * But, if we expired bfqq here, bfqq would
6229 * not have the chance to enjoy device idling
6230 * when bfqq->dispatched finally reaches
6231 * zero. This would expose bfqq to violation
6232 * of its reserved service guarantees.
6235 } else if (bfq_may_expire_for_budg_timeout(bfqq
))
6236 bfq_bfqq_expire(bfqd
, bfqq
, false,
6237 BFQQE_BUDGET_TIMEOUT
);
6238 else if (RB_EMPTY_ROOT(&bfqq
->sort_list
) &&
6239 (bfqq
->dispatched
== 0 ||
6240 !bfq_better_to_idle(bfqq
)))
6241 bfq_bfqq_expire(bfqd
, bfqq
, false,
6242 BFQQE_NO_MORE_REQUESTS
);
6245 if (!bfqd
->rq_in_driver
)
6246 bfq_schedule_dispatch(bfqd
);
6249 static void bfq_finish_requeue_request_body(struct bfq_queue
*bfqq
)
6253 bfq_put_queue(bfqq
);
6257 * The processes associated with bfqq may happen to generate their
6258 * cumulative I/O at a lower rate than the rate at which the device
6259 * could serve the same I/O. This is rather probable, e.g., if only
6260 * one process is associated with bfqq and the device is an SSD. It
6261 * results in bfqq becoming often empty while in service. In this
6262 * respect, if BFQ is allowed to switch to another queue when bfqq
6263 * remains empty, then the device goes on being fed with I/O requests,
6264 * and the throughput is not affected. In contrast, if BFQ is not
6265 * allowed to switch to another queue---because bfqq is sync and
6266 * I/O-dispatch needs to be plugged while bfqq is temporarily
6267 * empty---then, during the service of bfqq, there will be frequent
6268 * "service holes", i.e., time intervals during which bfqq gets empty
6269 * and the device can only consume the I/O already queued in its
6270 * hardware queues. During service holes, the device may even get to
6271 * remaining idle. In the end, during the service of bfqq, the device
6272 * is driven at a lower speed than the one it can reach with the kind
6273 * of I/O flowing through bfqq.
6275 * To counter this loss of throughput, BFQ implements a "request
6276 * injection mechanism", which tries to fill the above service holes
6277 * with I/O requests taken from other queues. The hard part in this
6278 * mechanism is finding the right amount of I/O to inject, so as to
6279 * both boost throughput and not break bfqq's bandwidth and latency
6280 * guarantees. In this respect, the mechanism maintains a per-queue
6281 * inject limit, computed as below. While bfqq is empty, the injection
6282 * mechanism dispatches extra I/O requests only until the total number
6283 * of I/O requests in flight---i.e., already dispatched but not yet
6284 * completed---remains lower than this limit.
6286 * A first definition comes in handy to introduce the algorithm by
6287 * which the inject limit is computed. We define as first request for
6288 * bfqq, an I/O request for bfqq that arrives while bfqq is in
6289 * service, and causes bfqq to switch from empty to non-empty. The
6290 * algorithm updates the limit as a function of the effect of
6291 * injection on the service times of only the first requests of
6292 * bfqq. The reason for this restriction is that these are the
6293 * requests whose service time is affected most, because they are the
6294 * first to arrive after injection possibly occurred.
6296 * To evaluate the effect of injection, the algorithm measures the
6297 * "total service time" of first requests. We define as total service
6298 * time of an I/O request, the time that elapses since when the
6299 * request is enqueued into bfqq, to when it is completed. This
6300 * quantity allows the whole effect of injection to be measured. It is
6301 * easy to see why. Suppose that some requests of other queues are
6302 * actually injected while bfqq is empty, and that a new request R
6303 * then arrives for bfqq. If the device does start to serve all or
6304 * part of the injected requests during the service hole, then,
6305 * because of this extra service, it may delay the next invocation of
6306 * the dispatch hook of BFQ. Then, even after R gets eventually
6307 * dispatched, the device may delay the actual service of R if it is
6308 * still busy serving the extra requests, or if it decides to serve,
6309 * before R, some extra request still present in its queues. As a
6310 * conclusion, the cumulative extra delay caused by injection can be
6311 * easily evaluated by just comparing the total service time of first
6312 * requests with and without injection.
6314 * The limit-update algorithm works as follows. On the arrival of a
6315 * first request of bfqq, the algorithm measures the total time of the
6316 * request only if one of the three cases below holds, and, for each
6317 * case, it updates the limit as described below:
6319 * (1) If there is no in-flight request. This gives a baseline for the
6320 * total service time of the requests of bfqq. If the baseline has
6321 * not been computed yet, then, after computing it, the limit is
6322 * set to 1, to start boosting throughput, and to prepare the
6323 * ground for the next case. If the baseline has already been
6324 * computed, then it is updated, in case it results to be lower
6325 * than the previous value.
6327 * (2) If the limit is higher than 0 and there are in-flight
6328 * requests. By comparing the total service time in this case with
6329 * the above baseline, it is possible to know at which extent the
6330 * current value of the limit is inflating the total service
6331 * time. If the inflation is below a certain threshold, then bfqq
6332 * is assumed to be suffering from no perceivable loss of its
6333 * service guarantees, and the limit is even tentatively
6334 * increased. If the inflation is above the threshold, then the
6335 * limit is decreased. Due to the lack of any hysteresis, this
6336 * logic makes the limit oscillate even in steady workload
6337 * conditions. Yet we opted for it, because it is fast in reaching
6338 * the best value for the limit, as a function of the current I/O
6339 * workload. To reduce oscillations, this step is disabled for a
6340 * short time interval after the limit happens to be decreased.
6342 * (3) Periodically, after resetting the limit, to make sure that the
6343 * limit eventually drops in case the workload changes. This is
6344 * needed because, after the limit has gone safely up for a
6345 * certain workload, it is impossible to guess whether the
6346 * baseline total service time may have changed, without measuring
6347 * it again without injection. A more effective version of this
6348 * step might be to just sample the baseline, by interrupting
6349 * injection only once, and then to reset/lower the limit only if
6350 * the total service time with the current limit does happen to be
6353 * More details on each step are provided in the comments on the
6354 * pieces of code that implement these steps: the branch handling the
6355 * transition from empty to non empty in bfq_add_request(), the branch
6356 * handling injection in bfq_select_queue(), and the function
6357 * bfq_choose_bfqq_for_injection(). These comments also explain some
6358 * exceptions, made by the injection mechanism in some special cases.
6360 static void bfq_update_inject_limit(struct bfq_data
*bfqd
,
6361 struct bfq_queue
*bfqq
)
6363 u64 tot_time_ns
= ktime_get_ns() - bfqd
->last_empty_occupied_ns
;
6364 unsigned int old_limit
= bfqq
->inject_limit
;
6366 if (bfqq
->last_serv_time_ns
> 0 && bfqd
->rqs_injected
) {
6367 u64 threshold
= (bfqq
->last_serv_time_ns
* 3)>>1;
6369 if (tot_time_ns
>= threshold
&& old_limit
> 0) {
6370 bfqq
->inject_limit
--;
6371 bfqq
->decrease_time_jif
= jiffies
;
6372 } else if (tot_time_ns
< threshold
&&
6373 old_limit
<= bfqd
->max_rq_in_driver
)
6374 bfqq
->inject_limit
++;
6378 * Either we still have to compute the base value for the
6379 * total service time, and there seem to be the right
6380 * conditions to do it, or we can lower the last base value
6383 * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
6384 * request in flight, because this function is in the code
6385 * path that handles the completion of a request of bfqq, and,
6386 * in particular, this function is executed before
6387 * bfqd->rq_in_driver is decremented in such a code path.
6389 if ((bfqq
->last_serv_time_ns
== 0 && bfqd
->rq_in_driver
== 1) ||
6390 tot_time_ns
< bfqq
->last_serv_time_ns
) {
6391 if (bfqq
->last_serv_time_ns
== 0) {
6393 * Now we certainly have a base value: make sure we
6394 * start trying injection.
6396 bfqq
->inject_limit
= max_t(unsigned int, 1, old_limit
);
6398 bfqq
->last_serv_time_ns
= tot_time_ns
;
6399 } else if (!bfqd
->rqs_injected
&& bfqd
->rq_in_driver
== 1)
6401 * No I/O injected and no request still in service in
6402 * the drive: these are the exact conditions for
6403 * computing the base value of the total service time
6404 * for bfqq. So let's update this value, because it is
6405 * rather variable. For example, it varies if the size
6406 * or the spatial locality of the I/O requests in bfqq
6409 bfqq
->last_serv_time_ns
= tot_time_ns
;
6412 /* update complete, not waiting for any request completion any longer */
6413 bfqd
->waited_rq
= NULL
;
6414 bfqd
->rqs_injected
= false;
6418 * Handle either a requeue or a finish for rq. The things to do are
6419 * the same in both cases: all references to rq are to be dropped. In
6420 * particular, rq is considered completed from the point of view of
6423 static void bfq_finish_requeue_request(struct request
*rq
)
6425 struct bfq_queue
*bfqq
= RQ_BFQQ(rq
);
6426 struct bfq_data
*bfqd
;
6427 unsigned long flags
;
6430 * rq either is not associated with any icq, or is an already
6431 * requeued request that has not (yet) been re-inserted into
6434 if (!rq
->elv
.icq
|| !bfqq
)
6439 if (rq
->rq_flags
& RQF_STARTED
)
6440 bfqg_stats_update_completion(bfqq_group(bfqq
),
6442 rq
->io_start_time_ns
,
6445 spin_lock_irqsave(&bfqd
->lock
, flags
);
6446 if (likely(rq
->rq_flags
& RQF_STARTED
)) {
6447 if (rq
== bfqd
->waited_rq
)
6448 bfq_update_inject_limit(bfqd
, bfqq
);
6450 bfq_completed_request(bfqq
, bfqd
);
6452 bfq_finish_requeue_request_body(bfqq
);
6453 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
6456 * Reset private fields. In case of a requeue, this allows
6457 * this function to correctly do nothing if it is spuriously
6458 * invoked again on this same request (see the check at the
6459 * beginning of the function). Probably, a better general
6460 * design would be to prevent blk-mq from invoking the requeue
6461 * or finish hooks of an elevator, for a request that is not
6462 * referred by that elevator.
6464 * Resetting the following fields would break the
6465 * request-insertion logic if rq is re-inserted into a bfq
6466 * internal queue, without a re-preparation. Here we assume
6467 * that re-insertions of requeued requests, without
6468 * re-preparation, can happen only for pass_through or at_head
6469 * requests (which are not re-inserted into bfq internal
6472 rq
->elv
.priv
[0] = NULL
;
6473 rq
->elv
.priv
[1] = NULL
;
6477 * Removes the association between the current task and bfqq, assuming
6478 * that bic points to the bfq iocontext of the task.
6479 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
6480 * was the last process referring to that bfqq.
6482 static struct bfq_queue
*
6483 bfq_split_bfqq(struct bfq_io_cq
*bic
, struct bfq_queue
*bfqq
)
6485 bfq_log_bfqq(bfqq
->bfqd
, bfqq
, "splitting queue");
6487 if (bfqq_process_refs(bfqq
) == 1) {
6488 bfqq
->pid
= current
->pid
;
6489 bfq_clear_bfqq_coop(bfqq
);
6490 bfq_clear_bfqq_split_coop(bfqq
);
6494 bic_set_bfqq(bic
, NULL
, 1);
6496 bfq_put_cooperator(bfqq
);
6498 bfq_release_process_ref(bfqq
->bfqd
, bfqq
);
6502 static struct bfq_queue
*bfq_get_bfqq_handle_split(struct bfq_data
*bfqd
,
6503 struct bfq_io_cq
*bic
,
6505 bool split
, bool is_sync
,
6508 struct bfq_queue
*bfqq
= bic_to_bfqq(bic
, is_sync
);
6510 if (likely(bfqq
&& bfqq
!= &bfqd
->oom_bfqq
))
6517 bfq_put_queue(bfqq
);
6518 bfqq
= bfq_get_queue(bfqd
, bio
, is_sync
, bic
, split
);
6520 bic_set_bfqq(bic
, bfqq
, is_sync
);
6521 if (split
&& is_sync
) {
6522 if ((bic
->was_in_burst_list
&& bfqd
->large_burst
) ||
6523 bic
->saved_in_large_burst
)
6524 bfq_mark_bfqq_in_large_burst(bfqq
);
6526 bfq_clear_bfqq_in_large_burst(bfqq
);
6527 if (bic
->was_in_burst_list
)
6529 * If bfqq was in the current
6530 * burst list before being
6531 * merged, then we have to add
6532 * it back. And we do not need
6533 * to increase burst_size, as
6534 * we did not decrement
6535 * burst_size when we removed
6536 * bfqq from the burst list as
6537 * a consequence of a merge
6539 * bfq_put_queue). In this
6540 * respect, it would be rather
6541 * costly to know whether the
6542 * current burst list is still
6543 * the same burst list from
6544 * which bfqq was removed on
6545 * the merge. To avoid this
6546 * cost, if bfqq was in a
6547 * burst list, then we add
6548 * bfqq to the current burst
6549 * list without any further
6550 * check. This can cause
6551 * inappropriate insertions,
6552 * but rarely enough to not
6553 * harm the detection of large
6554 * bursts significantly.
6556 hlist_add_head(&bfqq
->burst_list_node
,
6559 bfqq
->split_time
= jiffies
;
6566 * Only reset private fields. The actual request preparation will be
6567 * performed by bfq_init_rq, when rq is either inserted or merged. See
6568 * comments on bfq_init_rq for the reason behind this delayed
6571 static void bfq_prepare_request(struct request
*rq
)
6574 * Regardless of whether we have an icq attached, we have to
6575 * clear the scheduler pointers, as they might point to
6576 * previously allocated bic/bfqq structs.
6578 rq
->elv
.priv
[0] = rq
->elv
.priv
[1] = NULL
;
6582 * If needed, init rq, allocate bfq data structures associated with
6583 * rq, and increment reference counters in the destination bfq_queue
6584 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6585 * not associated with any bfq_queue.
6587 * This function is invoked by the functions that perform rq insertion
6588 * or merging. One may have expected the above preparation operations
6589 * to be performed in bfq_prepare_request, and not delayed to when rq
6590 * is inserted or merged. The rationale behind this delayed
6591 * preparation is that, after the prepare_request hook is invoked for
6592 * rq, rq may still be transformed into a request with no icq, i.e., a
6593 * request not associated with any queue. No bfq hook is invoked to
6594 * signal this transformation. As a consequence, should these
6595 * preparation operations be performed when the prepare_request hook
6596 * is invoked, and should rq be transformed one moment later, bfq
6597 * would end up in an inconsistent state, because it would have
6598 * incremented some queue counters for an rq destined to
6599 * transformation, without any chance to correctly lower these
6600 * counters back. In contrast, no transformation can still happen for
6601 * rq after rq has been inserted or merged. So, it is safe to execute
6602 * these preparation operations when rq is finally inserted or merged.
6604 static struct bfq_queue
*bfq_init_rq(struct request
*rq
)
6606 struct request_queue
*q
= rq
->q
;
6607 struct bio
*bio
= rq
->bio
;
6608 struct bfq_data
*bfqd
= q
->elevator
->elevator_data
;
6609 struct bfq_io_cq
*bic
;
6610 const int is_sync
= rq_is_sync(rq
);
6611 struct bfq_queue
*bfqq
;
6612 bool new_queue
= false;
6613 bool bfqq_already_existing
= false, split
= false;
6615 if (unlikely(!rq
->elv
.icq
))
6619 * Assuming that elv.priv[1] is set only if everything is set
6620 * for this rq. This holds true, because this function is
6621 * invoked only for insertion or merging, and, after such
6622 * events, a request cannot be manipulated any longer before
6623 * being removed from bfq.
6625 if (rq
->elv
.priv
[1])
6626 return rq
->elv
.priv
[1];
6628 bic
= icq_to_bic(rq
->elv
.icq
);
6630 bfq_check_ioprio_change(bic
, bio
);
6632 bfq_bic_update_cgroup(bic
, bio
);
6634 bfqq
= bfq_get_bfqq_handle_split(bfqd
, bic
, bio
, false, is_sync
,
6637 if (likely(!new_queue
)) {
6638 /* If the queue was seeky for too long, break it apart. */
6639 if (bfq_bfqq_coop(bfqq
) && bfq_bfqq_split_coop(bfqq
) &&
6640 !bic
->stably_merged
) {
6641 struct bfq_queue
*old_bfqq
= bfqq
;
6643 /* Update bic before losing reference to bfqq */
6644 if (bfq_bfqq_in_large_burst(bfqq
))
6645 bic
->saved_in_large_burst
= true;
6647 bfqq
= bfq_split_bfqq(bic
, bfqq
);
6651 bfqq
= bfq_get_bfqq_handle_split(bfqd
, bic
, bio
,
6654 bfqq
->waker_bfqq
= old_bfqq
->waker_bfqq
;
6655 bfqq
->tentative_waker_bfqq
= NULL
;
6658 * If the waker queue disappears, then
6659 * new_bfqq->waker_bfqq must be
6660 * reset. So insert new_bfqq into the
6661 * woken_list of the waker. See
6662 * bfq_check_waker for details.
6664 if (bfqq
->waker_bfqq
)
6665 hlist_add_head(&bfqq
->woken_list_node
,
6666 &bfqq
->waker_bfqq
->woken_list
);
6668 bfqq_already_existing
= true;
6674 bfq_log_bfqq(bfqd
, bfqq
, "get_request %p: bfqq %p, %d",
6675 rq
, bfqq
, bfqq
->ref
);
6677 rq
->elv
.priv
[0] = bic
;
6678 rq
->elv
.priv
[1] = bfqq
;
6681 * If a bfq_queue has only one process reference, it is owned
6682 * by only this bic: we can then set bfqq->bic = bic. in
6683 * addition, if the queue has also just been split, we have to
6686 if (likely(bfqq
!= &bfqd
->oom_bfqq
) && bfqq_process_refs(bfqq
) == 1) {
6690 * The queue has just been split from a shared
6691 * queue: restore the idle window and the
6692 * possible weight raising period.
6694 bfq_bfqq_resume_state(bfqq
, bfqd
, bic
,
6695 bfqq_already_existing
);
6700 * Consider bfqq as possibly belonging to a burst of newly
6701 * created queues only if:
6702 * 1) A burst is actually happening (bfqd->burst_size > 0)
6704 * 2) There is no other active queue. In fact, if, in
6705 * contrast, there are active queues not belonging to the
6706 * possible burst bfqq may belong to, then there is no gain
6707 * in considering bfqq as belonging to a burst, and
6708 * therefore in not weight-raising bfqq. See comments on
6709 * bfq_handle_burst().
6711 * This filtering also helps eliminating false positives,
6712 * occurring when bfqq does not belong to an actual large
6713 * burst, but some background task (e.g., a service) happens
6714 * to trigger the creation of new queues very close to when
6715 * bfqq and its possible companion queues are created. See
6716 * comments on bfq_handle_burst() for further details also on
6719 if (unlikely(bfq_bfqq_just_created(bfqq
) &&
6720 (bfqd
->burst_size
> 0 ||
6721 bfq_tot_busy_queues(bfqd
) == 0)))
6722 bfq_handle_burst(bfqd
, bfqq
);
6728 bfq_idle_slice_timer_body(struct bfq_data
*bfqd
, struct bfq_queue
*bfqq
)
6730 enum bfqq_expiration reason
;
6731 unsigned long flags
;
6733 spin_lock_irqsave(&bfqd
->lock
, flags
);
6736 * Considering that bfqq may be in race, we should firstly check
6737 * whether bfqq is in service before doing something on it. If
6738 * the bfqq in race is not in service, it has already been expired
6739 * through __bfq_bfqq_expire func and its wait_request flags has
6740 * been cleared in __bfq_bfqd_reset_in_service func.
6742 if (bfqq
!= bfqd
->in_service_queue
) {
6743 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
6747 bfq_clear_bfqq_wait_request(bfqq
);
6749 if (bfq_bfqq_budget_timeout(bfqq
))
6751 * Also here the queue can be safely expired
6752 * for budget timeout without wasting
6755 reason
= BFQQE_BUDGET_TIMEOUT
;
6756 else if (bfqq
->queued
[0] == 0 && bfqq
->queued
[1] == 0)
6758 * The queue may not be empty upon timer expiration,
6759 * because we may not disable the timer when the
6760 * first request of the in-service queue arrives
6761 * during disk idling.
6763 reason
= BFQQE_TOO_IDLE
;
6765 goto schedule_dispatch
;
6767 bfq_bfqq_expire(bfqd
, bfqq
, true, reason
);
6770 spin_unlock_irqrestore(&bfqd
->lock
, flags
);
6771 bfq_schedule_dispatch(bfqd
);
6775 * Handler of the expiration of the timer running if the in-service queue
6776 * is idling inside its time slice.
6778 static enum hrtimer_restart
bfq_idle_slice_timer(struct hrtimer
*timer
)
6780 struct bfq_data
*bfqd
= container_of(timer
, struct bfq_data
,
6782 struct bfq_queue
*bfqq
= bfqd
->in_service_queue
;
6785 * Theoretical race here: the in-service queue can be NULL or
6786 * different from the queue that was idling if a new request
6787 * arrives for the current queue and there is a full dispatch
6788 * cycle that changes the in-service queue. This can hardly
6789 * happen, but in the worst case we just expire a queue too
6793 bfq_idle_slice_timer_body(bfqd
, bfqq
);
6795 return HRTIMER_NORESTART
;
6798 static void __bfq_put_async_bfqq(struct bfq_data
*bfqd
,
6799 struct bfq_queue
**bfqq_ptr
)
6801 struct bfq_queue
*bfqq
= *bfqq_ptr
;
6803 bfq_log(bfqd
, "put_async_bfqq: %p", bfqq
);
6805 bfq_bfqq_move(bfqd
, bfqq
, bfqd
->root_group
);
6807 bfq_log_bfqq(bfqd
, bfqq
, "put_async_bfqq: putting %p, %d",
6809 bfq_put_queue(bfqq
);
6815 * Release all the bfqg references to its async queues. If we are
6816 * deallocating the group these queues may still contain requests, so
6817 * we reparent them to the root cgroup (i.e., the only one that will
6818 * exist for sure until all the requests on a device are gone).
6820 void bfq_put_async_queues(struct bfq_data
*bfqd
, struct bfq_group
*bfqg
)
6824 for (i
= 0; i
< 2; i
++)
6825 for (j
= 0; j
< IOPRIO_BE_NR
; j
++)
6826 __bfq_put_async_bfqq(bfqd
, &bfqg
->async_bfqq
[i
][j
]);
6828 __bfq_put_async_bfqq(bfqd
, &bfqg
->async_idle_bfqq
);
6832 * See the comments on bfq_limit_depth for the purpose of
6833 * the depths set in the function. Return minimum shallow depth we'll use.
6835 static unsigned int bfq_update_depths(struct bfq_data
*bfqd
,
6836 struct sbitmap_queue
*bt
)
6838 unsigned int i
, j
, min_shallow
= UINT_MAX
;
6841 * In-word depths if no bfq_queue is being weight-raised:
6842 * leaving 25% of tags only for sync reads.
6844 * In next formulas, right-shift the value
6845 * (1U<<bt->sb.shift), instead of computing directly
6846 * (1U<<(bt->sb.shift - something)), to be robust against
6847 * any possible value of bt->sb.shift, without having to
6848 * limit 'something'.
6850 /* no more than 50% of tags for async I/O */
6851 bfqd
->word_depths
[0][0] = max((1U << bt
->sb
.shift
) >> 1, 1U);
6853 * no more than 75% of tags for sync writes (25% extra tags
6854 * w.r.t. async I/O, to prevent async I/O from starving sync
6857 bfqd
->word_depths
[0][1] = max(((1U << bt
->sb
.shift
) * 3) >> 2, 1U);
6860 * In-word depths in case some bfq_queue is being weight-
6861 * raised: leaving ~63% of tags for sync reads. This is the
6862 * highest percentage for which, in our tests, application
6863 * start-up times didn't suffer from any regression due to tag
6866 /* no more than ~18% of tags for async I/O */
6867 bfqd
->word_depths
[1][0] = max(((1U << bt
->sb
.shift
) * 3) >> 4, 1U);
6868 /* no more than ~37% of tags for sync writes (~20% extra tags) */
6869 bfqd
->word_depths
[1][1] = max(((1U << bt
->sb
.shift
) * 6) >> 4, 1U);
6871 for (i
= 0; i
< 2; i
++)
6872 for (j
= 0; j
< 2; j
++)
6873 min_shallow
= min(min_shallow
, bfqd
->word_depths
[i
][j
]);
6878 static void bfq_depth_updated(struct blk_mq_hw_ctx
*hctx
)
6880 struct bfq_data
*bfqd
= hctx
->queue
->elevator
->elevator_data
;
6881 struct blk_mq_tags
*tags
= hctx
->sched_tags
;
6882 unsigned int min_shallow
;
6884 min_shallow
= bfq_update_depths(bfqd
, tags
->bitmap_tags
);
6885 sbitmap_queue_min_shallow_depth(tags
->bitmap_tags
, min_shallow
);
6888 static int bfq_init_hctx(struct blk_mq_hw_ctx
*hctx
, unsigned int index
)
6890 bfq_depth_updated(hctx
);
6894 static void bfq_exit_queue(struct elevator_queue
*e
)
6896 struct bfq_data
*bfqd
= e
->elevator_data
;
6897 struct bfq_queue
*bfqq
, *n
;
6899 hrtimer_cancel(&bfqd
->idle_slice_timer
);
6901 spin_lock_irq(&bfqd
->lock
);
6902 list_for_each_entry_safe(bfqq
, n
, &bfqd
->idle_list
, bfqq_list
)
6903 bfq_deactivate_bfqq(bfqd
, bfqq
, false, false);
6904 spin_unlock_irq(&bfqd
->lock
);
6906 hrtimer_cancel(&bfqd
->idle_slice_timer
);
6908 /* release oom-queue reference to root group */
6909 bfqg_and_blkg_put(bfqd
->root_group
);
6911 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6912 blkcg_deactivate_policy(bfqd
->queue
, &blkcg_policy_bfq
);
6914 spin_lock_irq(&bfqd
->lock
);
6915 bfq_put_async_queues(bfqd
, bfqd
->root_group
);
6916 kfree(bfqd
->root_group
);
6917 spin_unlock_irq(&bfqd
->lock
);
6923 static void bfq_init_root_group(struct bfq_group
*root_group
,
6924 struct bfq_data
*bfqd
)
6928 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6929 root_group
->entity
.parent
= NULL
;
6930 root_group
->my_entity
= NULL
;
6931 root_group
->bfqd
= bfqd
;
6933 root_group
->rq_pos_tree
= RB_ROOT
;
6934 for (i
= 0; i
< BFQ_IOPRIO_CLASSES
; i
++)
6935 root_group
->sched_data
.service_tree
[i
] = BFQ_SERVICE_TREE_INIT
;
6936 root_group
->sched_data
.bfq_class_idle_last_service
= jiffies
;
6939 static int bfq_init_queue(struct request_queue
*q
, struct elevator_type
*e
)
6941 struct bfq_data
*bfqd
;
6942 struct elevator_queue
*eq
;
6944 eq
= elevator_alloc(q
, e
);
6948 bfqd
= kzalloc_node(sizeof(*bfqd
), GFP_KERNEL
, q
->node
);
6950 kobject_put(&eq
->kobj
);
6953 eq
->elevator_data
= bfqd
;
6955 spin_lock_irq(&q
->queue_lock
);
6957 spin_unlock_irq(&q
->queue_lock
);
6960 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
6961 * Grab a permanent reference to it, so that the normal code flow
6962 * will not attempt to free it.
6964 bfq_init_bfqq(bfqd
, &bfqd
->oom_bfqq
, NULL
, 1, 0);
6965 bfqd
->oom_bfqq
.ref
++;
6966 bfqd
->oom_bfqq
.new_ioprio
= BFQ_DEFAULT_QUEUE_IOPRIO
;
6967 bfqd
->oom_bfqq
.new_ioprio_class
= IOPRIO_CLASS_BE
;
6968 bfqd
->oom_bfqq
.entity
.new_weight
=
6969 bfq_ioprio_to_weight(bfqd
->oom_bfqq
.new_ioprio
);
6971 /* oom_bfqq does not participate to bursts */
6972 bfq_clear_bfqq_just_created(&bfqd
->oom_bfqq
);
6975 * Trigger weight initialization, according to ioprio, at the
6976 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
6977 * class won't be changed any more.
6979 bfqd
->oom_bfqq
.entity
.prio_changed
= 1;
6983 INIT_LIST_HEAD(&bfqd
->dispatch
);
6985 hrtimer_init(&bfqd
->idle_slice_timer
, CLOCK_MONOTONIC
,
6987 bfqd
->idle_slice_timer
.function
= bfq_idle_slice_timer
;
6989 bfqd
->queue_weights_tree
= RB_ROOT_CACHED
;
6990 bfqd
->num_groups_with_pending_reqs
= 0;
6992 INIT_LIST_HEAD(&bfqd
->active_list
);
6993 INIT_LIST_HEAD(&bfqd
->idle_list
);
6994 INIT_HLIST_HEAD(&bfqd
->burst_list
);
6997 bfqd
->nonrot_with_queueing
= blk_queue_nonrot(bfqd
->queue
);
6999 bfqd
->bfq_max_budget
= bfq_default_max_budget
;
7001 bfqd
->bfq_fifo_expire
[0] = bfq_fifo_expire
[0];
7002 bfqd
->bfq_fifo_expire
[1] = bfq_fifo_expire
[1];
7003 bfqd
->bfq_back_max
= bfq_back_max
;
7004 bfqd
->bfq_back_penalty
= bfq_back_penalty
;
7005 bfqd
->bfq_slice_idle
= bfq_slice_idle
;
7006 bfqd
->bfq_timeout
= bfq_timeout
;
7008 bfqd
->bfq_large_burst_thresh
= 8;
7009 bfqd
->bfq_burst_interval
= msecs_to_jiffies(180);
7011 bfqd
->low_latency
= true;
7014 * Trade-off between responsiveness and fairness.
7016 bfqd
->bfq_wr_coeff
= 30;
7017 bfqd
->bfq_wr_rt_max_time
= msecs_to_jiffies(300);
7018 bfqd
->bfq_wr_max_time
= 0;
7019 bfqd
->bfq_wr_min_idle_time
= msecs_to_jiffies(2000);
7020 bfqd
->bfq_wr_min_inter_arr_async
= msecs_to_jiffies(500);
7021 bfqd
->bfq_wr_max_softrt_rate
= 7000; /*
7022 * Approximate rate required
7023 * to playback or record a
7024 * high-definition compressed
7027 bfqd
->wr_busy_queues
= 0;
7030 * Begin by assuming, optimistically, that the device peak
7031 * rate is equal to 2/3 of the highest reference rate.
7033 bfqd
->rate_dur_prod
= ref_rate
[blk_queue_nonrot(bfqd
->queue
)] *
7034 ref_wr_duration
[blk_queue_nonrot(bfqd
->queue
)];
7035 bfqd
->peak_rate
= ref_rate
[blk_queue_nonrot(bfqd
->queue
)] * 2 / 3;
7037 spin_lock_init(&bfqd
->lock
);
7040 * The invocation of the next bfq_create_group_hierarchy
7041 * function is the head of a chain of function calls
7042 * (bfq_create_group_hierarchy->blkcg_activate_policy->
7043 * blk_mq_freeze_queue) that may lead to the invocation of the
7044 * has_work hook function. For this reason,
7045 * bfq_create_group_hierarchy is invoked only after all
7046 * scheduler data has been initialized, apart from the fields
7047 * that can be initialized only after invoking
7048 * bfq_create_group_hierarchy. This, in particular, enables
7049 * has_work to correctly return false. Of course, to avoid
7050 * other inconsistencies, the blk-mq stack must then refrain
7051 * from invoking further scheduler hooks before this init
7052 * function is finished.
7054 bfqd
->root_group
= bfq_create_group_hierarchy(bfqd
, q
->node
);
7055 if (!bfqd
->root_group
)
7057 bfq_init_root_group(bfqd
->root_group
, bfqd
);
7058 bfq_init_entity(&bfqd
->oom_bfqq
.entity
, bfqd
->root_group
);
7060 wbt_disable_default(q
);
7065 kobject_put(&eq
->kobj
);
7069 static void bfq_slab_kill(void)
7071 kmem_cache_destroy(bfq_pool
);
7074 static int __init
bfq_slab_setup(void)
7076 bfq_pool
= KMEM_CACHE(bfq_queue
, 0);
7082 static ssize_t
bfq_var_show(unsigned int var
, char *page
)
7084 return sprintf(page
, "%u\n", var
);
7087 static int bfq_var_store(unsigned long *var
, const char *page
)
7089 unsigned long new_val
;
7090 int ret
= kstrtoul(page
, 10, &new_val
);
7098 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
7099 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7101 struct bfq_data *bfqd = e->elevator_data; \
7102 u64 __data = __VAR; \
7104 __data = jiffies_to_msecs(__data); \
7105 else if (__CONV == 2) \
7106 __data = div_u64(__data, NSEC_PER_MSEC); \
7107 return bfq_var_show(__data, (page)); \
7109 SHOW_FUNCTION(bfq_fifo_expire_sync_show
, bfqd
->bfq_fifo_expire
[1], 2);
7110 SHOW_FUNCTION(bfq_fifo_expire_async_show
, bfqd
->bfq_fifo_expire
[0], 2);
7111 SHOW_FUNCTION(bfq_back_seek_max_show
, bfqd
->bfq_back_max
, 0);
7112 SHOW_FUNCTION(bfq_back_seek_penalty_show
, bfqd
->bfq_back_penalty
, 0);
7113 SHOW_FUNCTION(bfq_slice_idle_show
, bfqd
->bfq_slice_idle
, 2);
7114 SHOW_FUNCTION(bfq_max_budget_show
, bfqd
->bfq_user_max_budget
, 0);
7115 SHOW_FUNCTION(bfq_timeout_sync_show
, bfqd
->bfq_timeout
, 1);
7116 SHOW_FUNCTION(bfq_strict_guarantees_show
, bfqd
->strict_guarantees
, 0);
7117 SHOW_FUNCTION(bfq_low_latency_show
, bfqd
->low_latency
, 0);
7118 #undef SHOW_FUNCTION
7120 #define USEC_SHOW_FUNCTION(__FUNC, __VAR) \
7121 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
7123 struct bfq_data *bfqd = e->elevator_data; \
7124 u64 __data = __VAR; \
7125 __data = div_u64(__data, NSEC_PER_USEC); \
7126 return bfq_var_show(__data, (page)); \
7128 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show
, bfqd
->bfq_slice_idle
);
7129 #undef USEC_SHOW_FUNCTION
7131 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
7133 __FUNC(struct elevator_queue *e, const char *page, size_t count) \
7135 struct bfq_data *bfqd = e->elevator_data; \
7136 unsigned long __data, __min = (MIN), __max = (MAX); \
7139 ret = bfq_var_store(&__data, (page)); \
7142 if (__data < __min) \
7144 else if (__data > __max) \
7147 *(__PTR) = msecs_to_jiffies(__data); \
7148 else if (__CONV == 2) \
7149 *(__PTR) = (u64)__data * NSEC_PER_MSEC; \
7151 *(__PTR) = __data; \
7154 STORE_FUNCTION(bfq_fifo_expire_sync_store
, &bfqd
->bfq_fifo_expire
[1], 1,
7156 STORE_FUNCTION(bfq_fifo_expire_async_store
, &bfqd
->bfq_fifo_expire
[0], 1,
7158 STORE_FUNCTION(bfq_back_seek_max_store
, &bfqd
->bfq_back_max
, 0, INT_MAX
, 0);
7159 STORE_FUNCTION(bfq_back_seek_penalty_store
, &bfqd
->bfq_back_penalty
, 1,
7161 STORE_FUNCTION(bfq_slice_idle_store
, &bfqd
->bfq_slice_idle
, 0, INT_MAX
, 2);
7162 #undef STORE_FUNCTION
7164 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \
7165 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
7167 struct bfq_data *bfqd = e->elevator_data; \
7168 unsigned long __data, __min = (MIN), __max = (MAX); \
7171 ret = bfq_var_store(&__data, (page)); \
7174 if (__data < __min) \
7176 else if (__data > __max) \
7178 *(__PTR) = (u64)__data * NSEC_PER_USEC; \
7181 USEC_STORE_FUNCTION(bfq_slice_idle_us_store
, &bfqd
->bfq_slice_idle
, 0,
7183 #undef USEC_STORE_FUNCTION
7185 static ssize_t
bfq_max_budget_store(struct elevator_queue
*e
,
7186 const char *page
, size_t count
)
7188 struct bfq_data
*bfqd
= e
->elevator_data
;
7189 unsigned long __data
;
7192 ret
= bfq_var_store(&__data
, (page
));
7197 bfqd
->bfq_max_budget
= bfq_calc_max_budget(bfqd
);
7199 if (__data
> INT_MAX
)
7201 bfqd
->bfq_max_budget
= __data
;
7204 bfqd
->bfq_user_max_budget
= __data
;
7210 * Leaving this name to preserve name compatibility with cfq
7211 * parameters, but this timeout is used for both sync and async.
7213 static ssize_t
bfq_timeout_sync_store(struct elevator_queue
*e
,
7214 const char *page
, size_t count
)
7216 struct bfq_data
*bfqd
= e
->elevator_data
;
7217 unsigned long __data
;
7220 ret
= bfq_var_store(&__data
, (page
));
7226 else if (__data
> INT_MAX
)
7229 bfqd
->bfq_timeout
= msecs_to_jiffies(__data
);
7230 if (bfqd
->bfq_user_max_budget
== 0)
7231 bfqd
->bfq_max_budget
= bfq_calc_max_budget(bfqd
);
7236 static ssize_t
bfq_strict_guarantees_store(struct elevator_queue
*e
,
7237 const char *page
, size_t count
)
7239 struct bfq_data
*bfqd
= e
->elevator_data
;
7240 unsigned long __data
;
7243 ret
= bfq_var_store(&__data
, (page
));
7249 if (!bfqd
->strict_guarantees
&& __data
== 1
7250 && bfqd
->bfq_slice_idle
< 8 * NSEC_PER_MSEC
)
7251 bfqd
->bfq_slice_idle
= 8 * NSEC_PER_MSEC
;
7253 bfqd
->strict_guarantees
= __data
;
7258 static ssize_t
bfq_low_latency_store(struct elevator_queue
*e
,
7259 const char *page
, size_t count
)
7261 struct bfq_data
*bfqd
= e
->elevator_data
;
7262 unsigned long __data
;
7265 ret
= bfq_var_store(&__data
, (page
));
7271 if (__data
== 0 && bfqd
->low_latency
!= 0)
7273 bfqd
->low_latency
= __data
;
7278 #define BFQ_ATTR(name) \
7279 __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
7281 static struct elv_fs_entry bfq_attrs
[] = {
7282 BFQ_ATTR(fifo_expire_sync
),
7283 BFQ_ATTR(fifo_expire_async
),
7284 BFQ_ATTR(back_seek_max
),
7285 BFQ_ATTR(back_seek_penalty
),
7286 BFQ_ATTR(slice_idle
),
7287 BFQ_ATTR(slice_idle_us
),
7288 BFQ_ATTR(max_budget
),
7289 BFQ_ATTR(timeout_sync
),
7290 BFQ_ATTR(strict_guarantees
),
7291 BFQ_ATTR(low_latency
),
7295 static struct elevator_type iosched_bfq_mq
= {
7297 .limit_depth
= bfq_limit_depth
,
7298 .prepare_request
= bfq_prepare_request
,
7299 .requeue_request
= bfq_finish_requeue_request
,
7300 .finish_request
= bfq_finish_requeue_request
,
7301 .exit_icq
= bfq_exit_icq
,
7302 .insert_requests
= bfq_insert_requests
,
7303 .dispatch_request
= bfq_dispatch_request
,
7304 .next_request
= elv_rb_latter_request
,
7305 .former_request
= elv_rb_former_request
,
7306 .allow_merge
= bfq_allow_bio_merge
,
7307 .bio_merge
= bfq_bio_merge
,
7308 .request_merge
= bfq_request_merge
,
7309 .requests_merged
= bfq_requests_merged
,
7310 .request_merged
= bfq_request_merged
,
7311 .has_work
= bfq_has_work
,
7312 .depth_updated
= bfq_depth_updated
,
7313 .init_hctx
= bfq_init_hctx
,
7314 .init_sched
= bfq_init_queue
,
7315 .exit_sched
= bfq_exit_queue
,
7318 .icq_size
= sizeof(struct bfq_io_cq
),
7319 .icq_align
= __alignof__(struct bfq_io_cq
),
7320 .elevator_attrs
= bfq_attrs
,
7321 .elevator_name
= "bfq",
7322 .elevator_owner
= THIS_MODULE
,
7324 MODULE_ALIAS("bfq-iosched");
7326 static int __init
bfq_init(void)
7330 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7331 ret
= blkcg_policy_register(&blkcg_policy_bfq
);
7337 if (bfq_slab_setup())
7341 * Times to load large popular applications for the typical
7342 * systems installed on the reference devices (see the
7343 * comments before the definition of the next
7344 * array). Actually, we use slightly lower values, as the
7345 * estimated peak rate tends to be smaller than the actual
7346 * peak rate. The reason for this last fact is that estimates
7347 * are computed over much shorter time intervals than the long
7348 * intervals typically used for benchmarking. Why? First, to
7349 * adapt more quickly to variations. Second, because an I/O
7350 * scheduler cannot rely on a peak-rate-evaluation workload to
7351 * be run for a long time.
7353 ref_wr_duration
[0] = msecs_to_jiffies(7000); /* actually 8 sec */
7354 ref_wr_duration
[1] = msecs_to_jiffies(2500); /* actually 3 sec */
7356 ret
= elv_register(&iosched_bfq_mq
);
7365 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7366 blkcg_policy_unregister(&blkcg_policy_bfq
);
7371 static void __exit
bfq_exit(void)
7373 elv_unregister(&iosched_bfq_mq
);
7374 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7375 blkcg_policy_unregister(&blkcg_policy_bfq
);
7380 module_init(bfq_init
);
7381 module_exit(bfq_exit
);
7383 MODULE_AUTHOR("Paolo Valente");
7384 MODULE_LICENSE("GPL");
7385 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");