1/*
2 * Budget Fair Queueing (BFQ) I/O scheduler.
3 *
4 * Based on ideas and code from CFQ:
5 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
6 *
7 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
8 * Paolo Valente <paolo.valente@unimore.it>
9 *
10 * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
11 * Arianna Avanzini <avanzini@google.com>
12 *
13 * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
14 *
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License as
17 * published by the Free Software Foundation; either version 2 of the
18 * License, or (at your option) any later version.
19 *
20 * This program is distributed in the hope that it will be useful,
21 * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
23 * General Public License for more details.
24 *
25 * BFQ is a proportional-share I/O scheduler, with some extra
26 * low-latency capabilities. BFQ also supports full hierarchical
27 * scheduling through cgroups. Next paragraphs provide an introduction
28 * on BFQ inner workings. Details on BFQ benefits, usage and
29 * limitations can be found in Documentation/block/bfq-iosched.txt.
30 *
31 * BFQ is a proportional-share storage-I/O scheduling algorithm based
32 * on the slice-by-slice service scheme of CFQ. But BFQ assigns
33 * budgets, measured in number of sectors, to processes instead of
34 * time slices. The device is not granted to the in-service process
35 * for a given time slice, but until it has exhausted its assigned
36 * budget. This change from the time to the service domain enables BFQ
37 * to distribute the device throughput among processes as desired,
38 * without any distortion due to throughput fluctuations, or to device
39 * internal queueing. BFQ uses an ad hoc internal scheduler, called
40 * B-WF2Q+, to schedule processes according to their budgets. More
41 * precisely, BFQ schedules queues associated with processes. Each
42 * process/queue is assigned a user-configurable weight, and B-WF2Q+
43 * guarantees that each queue receives a fraction of the throughput
44 * proportional to its weight. Thanks to the accurate policy of
45 * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
46 * processes issuing sequential requests (to boost the throughput),
47 * and yet guarantee a low latency to interactive and soft real-time
48 * applications.
49 *
50 * In particular, to provide these low-latency guarantees, BFQ
51 * explicitly privileges the I/O of two classes of time-sensitive
52 * applications: interactive and soft real-time. In more detail, BFQ
53 * behaves this way if the low_latency parameter is set (default
54 * configuration). This feature enables BFQ to provide applications in
55 * these classes with a very low latency.
56 *
57 * To implement this feature, BFQ constantly tries to detect whether
58 * the I/O requests in a bfq_queue come from an interactive or a soft
59 * real-time application. For brevity, in these cases, the queue is
60 * said to be interactive or soft real-time. In both cases, BFQ
61 * privileges the service of the queue, over that of non-interactive
62 * and non-soft-real-time queues. This privileging is performed,
63 * mainly, by raising the weight of the queue. So, for brevity, we
64 * call just weight-raising periods the time periods during which a
65 * queue is privileged, because deemed interactive or soft real-time.
66 *
67 * The detection of soft real-time queues/applications is described in
68 * detail in the comments on the function
69 * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
70 * interactive queue works as follows: a queue is deemed interactive
71 * if it is constantly non empty only for a limited time interval,
72 * after which it does become empty. The queue may be deemed
73 * interactive again (for a limited time), if it restarts being
74 * constantly non empty, provided that this happens only after the
75 * queue has remained empty for a given minimum idle time.
76 *
77 * By default, BFQ computes automatically the above maximum time
78 * interval, i.e., the time interval after which a constantly
79 * non-empty queue stops being deemed interactive. Since a queue is
80 * weight-raised while it is deemed interactive, this maximum time
81 * interval happens to coincide with the (maximum) duration of the
82 * weight-raising for interactive queues.
83 *
84 * Finally, BFQ also features additional heuristics for
85 * preserving both a low latency and a high throughput on NCQ-capable,
86 * rotational or flash-based devices, and to get the job done quickly
87 * for applications consisting in many I/O-bound processes.
88 *
89 * NOTE: if the main or only goal, with a given device, is to achieve
90 * the maximum-possible throughput at all times, then do switch off
91 * all low-latency heuristics for that device, by setting low_latency
92 * to 0.
93 *
94 * BFQ is described in [1], where also a reference to the initial,
95 * more theoretical paper on BFQ can be found. The interested reader
96 * can find in the latter paper full details on the main algorithm, as
97 * well as formulas of the guarantees and formal proofs of all the
98 * properties. With respect to the version of BFQ presented in these
99 * papers, this implementation adds a few more heuristics, such as the
100 * ones that guarantee a low latency to interactive and soft real-time
101 * applications, and a hierarchical extension based on H-WF2Q+.
102 *
103 * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
104 * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
105 * with O(log N) complexity derives from the one introduced with EEVDF
106 * in [3].
107 *
108 * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
109 * Scheduler", Proceedings of the First Workshop on Mobile System
110 * Technologies (MST-2015), May 2015.
111 * http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
112 *
113 * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
114 * Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
115 * Oct 1997.
116 *
117 * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
118 *
119 * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
120 * First: A Flexible and Accurate Mechanism for Proportional Share
121 * Resource Allocation", technical report.
122 *
123 * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
124 */
125#include <linux/module.h>
126#include <linux/slab.h>
127#include <linux/blkdev.h>
128#include <linux/cgroup.h>
129#include <linux/elevator.h>
130#include <linux/ktime.h>
131#include <linux/rbtree.h>
132#include <linux/ioprio.h>
133#include <linux/sbitmap.h>
134#include <linux/delay.h>
135
136#include "blk.h"
137#include "blk-mq.h"
138#include "blk-mq-tag.h"
139#include "blk-mq-sched.h"
140#include "bfq-iosched.h"
141#include "blk-wbt.h"
142
143#define BFQ_BFQQ_FNS(name) \
144void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \
145{ \
146 __set_bit(BFQQF_##name, &(bfqq)->flags); \
147} \
148void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \
149{ \
150 __clear_bit(BFQQF_##name, &(bfqq)->flags); \
151} \
152int bfq_bfqq_##name(const struct bfq_queue *bfqq) \
153{ \
154 return test_bit(BFQQF_##name, &(bfqq)->flags); \
155}
156
157BFQ_BFQQ_FNS(just_created);
158BFQ_BFQQ_FNS(busy);
159BFQ_BFQQ_FNS(wait_request);
160BFQ_BFQQ_FNS(non_blocking_wait_rq);
161BFQ_BFQQ_FNS(fifo_expire);
162BFQ_BFQQ_FNS(has_short_ttime);
163BFQ_BFQQ_FNS(sync);
164BFQ_BFQQ_FNS(IO_bound);
165BFQ_BFQQ_FNS(in_large_burst);
166BFQ_BFQQ_FNS(coop);
167BFQ_BFQQ_FNS(split_coop);
168BFQ_BFQQ_FNS(softrt_update);
169#undef BFQ_BFQQ_FNS \
170
171/* Expiration time of sync (0) and async (1) requests, in ns. */
172static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
173
174/* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
175static const int bfq_back_max = 16 * 1024;
176
177/* Penalty of a backwards seek, in number of sectors. */
178static const int bfq_back_penalty = 2;
179
180/* Idling period duration, in ns. */
181static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
182
183/* Minimum number of assigned budgets for which stats are safe to compute. */
184static const int bfq_stats_min_budgets = 194;
185
186/* Default maximum budget values, in sectors and number of requests. */
187static const int bfq_default_max_budget = 16 * 1024;
188
189/*
190 * When a sync request is dispatched, the queue that contains that
191 * request, and all the ancestor entities of that queue, are charged
192 * with the number of sectors of the request. In constrast, if the
193 * request is async, then the queue and its ancestor entities are
194 * charged with the number of sectors of the request, multiplied by
195 * the factor below. This throttles the bandwidth for async I/O,
196 * w.r.t. to sync I/O, and it is done to counter the tendency of async
197 * writes to steal I/O throughput to reads.
198 *
199 * The current value of this parameter is the result of a tuning with
200 * several hardware and software configurations. We tried to find the
201 * lowest value for which writes do not cause noticeable problems to
202 * reads. In fact, the lower this parameter, the stabler I/O control,
203 * in the following respect. The lower this parameter is, the less
204 * the bandwidth enjoyed by a group decreases
205 * - when the group does writes, w.r.t. to when it does reads;
206 * - when other groups do reads, w.r.t. to when they do writes.
207 */
208static const int bfq_async_charge_factor = 3;
209
210/* Default timeout values, in jiffies, approximating CFQ defaults. */
211const int bfq_timeout = HZ / 8;
212
213/*
214 * Time limit for merging (see comments in bfq_setup_cooperator). Set
215 * to the slowest value that, in our tests, proved to be effective in
216 * removing false positives, while not causing true positives to miss
217 * queue merging.
218 *
219 * As can be deduced from the low time limit below, queue merging, if
220 * successful, happens at the very beggining of the I/O of the involved
221 * cooperating processes, as a consequence of the arrival of the very
222 * first requests from each cooperator. After that, there is very
223 * little chance to find cooperators.
224 */
225static const unsigned long bfq_merge_time_limit = HZ/10;
226
227static struct kmem_cache *bfq_pool;
228
229/* Below this threshold (in ns), we consider thinktime immediate. */
230#define BFQ_MIN_TT (2 * NSEC_PER_MSEC)
231
232/* hw_tag detection: parallel requests threshold and min samples needed. */
233#define BFQ_HW_QUEUE_THRESHOLD 3
234#define BFQ_HW_QUEUE_SAMPLES 32
235
236#define BFQQ_SEEK_THR (sector_t)(8 * 100)
237#define BFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
238#define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \
239 (get_sdist(last_pos, rq) > \
240 BFQQ_SEEK_THR && \
241 (!blk_queue_nonrot(bfqd->queue) || \
242 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT))
243#define BFQQ_CLOSE_THR (sector_t)(8 * 1024)
244#define BFQQ_SEEKY(bfqq) (hweight32(bfqq->seek_history) > 19)
245
246/* Min number of samples required to perform peak-rate update */
247#define BFQ_RATE_MIN_SAMPLES 32
248/* Min observation time interval required to perform a peak-rate update (ns) */
249#define BFQ_RATE_MIN_INTERVAL (300*NSEC_PER_MSEC)
250/* Target observation time interval for a peak-rate update (ns) */
251#define BFQ_RATE_REF_INTERVAL NSEC_PER_SEC
252
253/*
254 * Shift used for peak-rate fixed precision calculations.
255 * With
256 * - the current shift: 16 positions
257 * - the current type used to store rate: u32
258 * - the current unit of measure for rate: [sectors/usec], or, more precisely,
259 * [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
260 * the range of rates that can be stored is
261 * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
262 * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
263 * [15, 65G] sectors/sec
264 * Which, assuming a sector size of 512B, corresponds to a range of
265 * [7.5K, 33T] B/sec
266 */
267#define BFQ_RATE_SHIFT 16
268
269/*
270 * When configured for computing the duration of the weight-raising
271 * for interactive queues automatically (see the comments at the
272 * beginning of this file), BFQ does it using the following formula:
273 * duration = (ref_rate / r) * ref_wr_duration,
274 * where r is the peak rate of the device, and ref_rate and
275 * ref_wr_duration are two reference parameters. In particular,
276 * ref_rate is the peak rate of the reference storage device (see
277 * below), and ref_wr_duration is about the maximum time needed, with
278 * BFQ and while reading two files in parallel, to load typical large
279 * applications on the reference device (see the comments on
280 * max_service_from_wr below, for more details on how ref_wr_duration
281 * is obtained). In practice, the slower/faster the device at hand
282 * is, the more/less it takes to load applications with respect to the
283 * reference device. Accordingly, the longer/shorter BFQ grants
284 * weight raising to interactive applications.
285 *
286 * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
287 * depending on whether the device is rotational or non-rotational.
288 *
289 * In the following definitions, ref_rate[0] and ref_wr_duration[0]
290 * are the reference values for a rotational device, whereas
291 * ref_rate[1] and ref_wr_duration[1] are the reference values for a
292 * non-rotational device. The reference rates are not the actual peak
293 * rates of the devices used as a reference, but slightly lower
294 * values. The reason for using slightly lower values is that the
295 * peak-rate estimator tends to yield slightly lower values than the
296 * actual peak rate (it can yield the actual peak rate only if there
297 * is only one process doing I/O, and the process does sequential
298 * I/O).
299 *
300 * The reference peak rates are measured in sectors/usec, left-shifted
301 * by BFQ_RATE_SHIFT.
302 */
303static int ref_rate[2] = {14000, 33000};
304/*
305 * To improve readability, a conversion function is used to initialize
306 * the following array, which entails that the array can be
307 * initialized only in a function.
308 */
309static int ref_wr_duration[2];
310
311/*
312 * BFQ uses the above-detailed, time-based weight-raising mechanism to
313 * privilege interactive tasks. This mechanism is vulnerable to the
314 * following false positives: I/O-bound applications that will go on
315 * doing I/O for much longer than the duration of weight
316 * raising. These applications have basically no benefit from being
317 * weight-raised at the beginning of their I/O. On the opposite end,
318 * while being weight-raised, these applications
319 * a) unjustly steal throughput to applications that may actually need
320 * low latency;
321 * b) make BFQ uselessly perform device idling; device idling results
322 * in loss of device throughput with most flash-based storage, and may
323 * increase latencies when used purposelessly.
324 *
325 * BFQ tries to reduce these problems, by adopting the following
326 * countermeasure. To introduce this countermeasure, we need first to
327 * finish explaining how the duration of weight-raising for
328 * interactive tasks is computed.
329 *
330 * For a bfq_queue deemed as interactive, the duration of weight
331 * raising is dynamically adjusted, as a function of the estimated
332 * peak rate of the device, so as to be equal to the time needed to
333 * execute the 'largest' interactive task we benchmarked so far. By
334 * largest task, we mean the task for which each involved process has
335 * to do more I/O than for any of the other tasks we benchmarked. This
336 * reference interactive task is the start-up of LibreOffice Writer,
337 * and in this task each process/bfq_queue needs to have at most ~110K
338 * sectors transferred.
339 *
340 * This last piece of information enables BFQ to reduce the actual
341 * duration of weight-raising for at least one class of I/O-bound
342 * applications: those doing sequential or quasi-sequential I/O. An
343 * example is file copy. In fact, once started, the main I/O-bound
344 * processes of these applications usually consume the above 110K
345 * sectors in much less time than the processes of an application that
346 * is starting, because these I/O-bound processes will greedily devote
347 * almost all their CPU cycles only to their target,
348 * throughput-friendly I/O operations. This is even more true if BFQ
349 * happens to be underestimating the device peak rate, and thus
350 * overestimating the duration of weight raising. But, according to
351 * our measurements, once transferred 110K sectors, these processes
352 * have no right to be weight-raised any longer.
353 *
354 * Basing on the last consideration, BFQ ends weight-raising for a
355 * bfq_queue if the latter happens to have received an amount of
356 * service at least equal to the following constant. The constant is
357 * set to slightly more than 110K, to have a minimum safety margin.
358 *
359 * This early ending of weight-raising reduces the amount of time
360 * during which interactive false positives cause the two problems
361 * described at the beginning of these comments.
362 */
363static const unsigned long max_service_from_wr = 120000;
364
365#define RQ_BIC(rq) icq_to_bic((rq)->elv.priv[0])
366#define RQ_BFQQ(rq) ((rq)->elv.priv[1])
367
368struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
369{
370 return bic->bfqq[is_sync];
371}
372
373void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
374{
375 bic->bfqq[is_sync] = bfqq;
376}
377
378struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
379{
380 return bic->icq.q->elevator->elevator_data;
381}
382
383/**
384 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
385 * @icq: the iocontext queue.
386 */
387static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
388{
389 /* bic->icq is the first member, %NULL will convert to %NULL */
390 return container_of(icq, struct bfq_io_cq, icq);
391}
392
393/**
394 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
395 * @bfqd: the lookup key.
396 * @ioc: the io_context of the process doing I/O.
397 * @q: the request queue.
398 */
399static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd,
400 struct io_context *ioc,
401 struct request_queue *q)
402{
403 if (ioc) {
404 unsigned long flags;
405 struct bfq_io_cq *icq;
406
407 spin_lock_irqsave(&q->queue_lock, flags);
408 icq = icq_to_bic(ioc_lookup_icq(ioc, q));
409 spin_unlock_irqrestore(&q->queue_lock, flags);
410
411 return icq;
412 }
413
414 return NULL;
415}
416
417/*
418 * Scheduler run of queue, if there are requests pending and no one in the
419 * driver that will restart queueing.
420 */
421void bfq_schedule_dispatch(struct bfq_data *bfqd)
422{
423 if (bfqd->queued != 0) {
424 bfq_log(bfqd, "schedule dispatch");
425 blk_mq_run_hw_queues(bfqd->queue, true);
426 }
427}
428
429#define bfq_class_idle(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
430#define bfq_class_rt(bfqq) ((bfqq)->ioprio_class == IOPRIO_CLASS_RT)
431
432#define bfq_sample_valid(samples) ((samples) > 80)
433
434/*
435 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
436 * We choose the request that is closesr to the head right now. Distance
437 * behind the head is penalized and only allowed to a certain extent.
438 */
439static struct request *bfq_choose_req(struct bfq_data *bfqd,
440 struct request *rq1,
441 struct request *rq2,
442 sector_t last)
443{
444 sector_t s1, s2, d1 = 0, d2 = 0;
445 unsigned long back_max;
446#define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */
447#define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */
448 unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
449
450 if (!rq1 || rq1 == rq2)
451 return rq2;
452 if (!rq2)
453 return rq1;
454
455 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
456 return rq1;
457 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
458 return rq2;
459 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
460 return rq1;
461 else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
462 return rq2;
463
464 s1 = blk_rq_pos(rq1);
465 s2 = blk_rq_pos(rq2);
466
467 /*
468 * By definition, 1KiB is 2 sectors.
469 */
470 back_max = bfqd->bfq_back_max * 2;
471
472 /*
473 * Strict one way elevator _except_ in the case where we allow
474 * short backward seeks which are biased as twice the cost of a
475 * similar forward seek.
476 */
477 if (s1 >= last)
478 d1 = s1 - last;
479 else if (s1 + back_max >= last)
480 d1 = (last - s1) * bfqd->bfq_back_penalty;
481 else
482 wrap |= BFQ_RQ1_WRAP;
483
484 if (s2 >= last)
485 d2 = s2 - last;
486 else if (s2 + back_max >= last)
487 d2 = (last - s2) * bfqd->bfq_back_penalty;
488 else
489 wrap |= BFQ_RQ2_WRAP;
490
491 /* Found required data */
492
493 /*
494 * By doing switch() on the bit mask "wrap" we avoid having to
495 * check two variables for all permutations: --> faster!
496 */
497 switch (wrap) {
498 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
499 if (d1 < d2)
500 return rq1;
501 else if (d2 < d1)
502 return rq2;
503
504 if (s1 >= s2)
505 return rq1;
506 else
507 return rq2;
508
509 case BFQ_RQ2_WRAP:
510 return rq1;
511 case BFQ_RQ1_WRAP:
512 return rq2;
513 case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
514 default:
515 /*
516 * Since both rqs are wrapped,
517 * start with the one that's further behind head
518 * (--> only *one* back seek required),
519 * since back seek takes more time than forward.
520 */
521 if (s1 <= s2)
522 return rq1;
523 else
524 return rq2;
525 }
526}
527
528/*
529 * Async I/O can easily starve sync I/O (both sync reads and sync
530 * writes), by consuming all tags. Similarly, storms of sync writes,
531 * such as those that sync(2) may trigger, can starve sync reads.
532 * Limit depths of async I/O and sync writes so as to counter both
533 * problems.
534 */
535static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
536{
537 struct bfq_data *bfqd = data->q->elevator->elevator_data;
538
539 if (op_is_sync(op) && !op_is_write(op))
540 return;
541
542 data->shallow_depth =
543 bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)];
544
545 bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
546 __func__, bfqd->wr_busy_queues, op_is_sync(op),
547 data->shallow_depth);
548}
549
550static struct bfq_queue *
551bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
552 sector_t sector, struct rb_node **ret_parent,
553 struct rb_node ***rb_link)
554{
555 struct rb_node **p, *parent;
556 struct bfq_queue *bfqq = NULL;
557
558 parent = NULL;
559 p = &root->rb_node;
560 while (*p) {
561 struct rb_node **n;
562
563 parent = *p;
564 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
565
566 /*
567 * Sort strictly based on sector. Smallest to the left,
568 * largest to the right.
569 */
570 if (sector > blk_rq_pos(bfqq->next_rq))
571 n = &(*p)->rb_right;
572 else if (sector < blk_rq_pos(bfqq->next_rq))
573 n = &(*p)->rb_left;
574 else
575 break;
576 p = n;
577 bfqq = NULL;
578 }
579
580 *ret_parent = parent;
581 if (rb_link)
582 *rb_link = p;
583
584 bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
585 (unsigned long long)sector,
586 bfqq ? bfqq->pid : 0);
587
588 return bfqq;
589}
590
591static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
592{
593 return bfqq->service_from_backlogged > 0 &&
594 time_is_before_jiffies(bfqq->first_IO_time +
595 bfq_merge_time_limit);
596}
597
598void bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
599{
600 struct rb_node **p, *parent;
601 struct bfq_queue *__bfqq;
602
603 if (bfqq->pos_root) {
604 rb_erase(&bfqq->pos_node, bfqq->pos_root);
605 bfqq->pos_root = NULL;
606 }
607
608 /*
609 * bfqq cannot be merged any longer (see comments in
610 * bfq_setup_cooperator): no point in adding bfqq into the
611 * position tree.
612 */
613 if (bfq_too_late_for_merging(bfqq))
614 return;
615
616 if (bfq_class_idle(bfqq))
617 return;
618 if (!bfqq->next_rq)
619 return;
620
621 bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
622 __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
623 blk_rq_pos(bfqq->next_rq), &parent, &p);
624 if (!__bfqq) {
625 rb_link_node(&bfqq->pos_node, parent, p);
626 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
627 } else
628 bfqq->pos_root = NULL;
629}
630
631/*
632 * The following function returns true if every queue must receive the
633 * same share of the throughput (this condition is used when deciding
634 * whether idling may be disabled, see the comments in the function
635 * bfq_better_to_idle()).
636 *
637 * Such a scenario occurs when:
638 * 1) all active queues have the same weight,
639 * 2) all active queues belong to the same I/O-priority class,
640 * 3) all active groups at the same level in the groups tree have the same
641 * weight,
642 * 4) all active groups at the same level in the groups tree have the same
643 * number of children.
644 *
645 * Unfortunately, keeping the necessary state for evaluating exactly
646 * the last two symmetry sub-conditions above would be quite complex
647 * and time consuming. Therefore this function evaluates, instead,
648 * only the following stronger three sub-conditions, for which it is
649 * much easier to maintain the needed state:
650 * 1) all active queues have the same weight,
651 * 2) all active queues belong to the same I/O-priority class,
652 * 3) there are no active groups.
653 * In particular, the last condition is always true if hierarchical
654 * support or the cgroups interface are not enabled, thus no state
655 * needs to be maintained in this case.
656 */
657static bool bfq_symmetric_scenario(struct bfq_data *bfqd)
658{
659 /*
660 * For queue weights to differ, queue_weights_tree must contain
661 * at least two nodes.
662 */
663 bool varied_queue_weights = !RB_EMPTY_ROOT(&bfqd->queue_weights_tree) &&
664 (bfqd->queue_weights_tree.rb_node->rb_left ||
665 bfqd->queue_weights_tree.rb_node->rb_right);
666
667 bool multiple_classes_busy =
668 (bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
669 (bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
670 (bfqd->busy_queues[1] && bfqd->busy_queues[2]);
671
672 /*
673 * For queue weights to differ, queue_weights_tree must contain
674 * at least two nodes.
675 */
676 return !(varied_queue_weights || multiple_classes_busy
677#ifdef BFQ_GROUP_IOSCHED_ENABLED
678 || bfqd->num_groups_with_pending_reqs > 0
679#endif
680 );
681}
682
683/*
684 * If the weight-counter tree passed as input contains no counter for
685 * the weight of the input queue, then add that counter; otherwise just
686 * increment the existing counter.
687 *
688 * Note that weight-counter trees contain few nodes in mostly symmetric
689 * scenarios. For example, if all queues have the same weight, then the
690 * weight-counter tree for the queues may contain at most one node.
691 * This holds even if low_latency is on, because weight-raised queues
692 * are not inserted in the tree.
693 * In most scenarios, the rate at which nodes are created/destroyed
694 * should be low too.
695 */
696void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq,
697 struct rb_root *root)
698{
699 struct bfq_entity *entity = &bfqq->entity;
700 struct rb_node **new = &(root->rb_node), *parent = NULL;
701
702 /*
703 * Do not insert if the queue is already associated with a
704 * counter, which happens if:
705 * 1) a request arrival has caused the queue to become both
706 * non-weight-raised, and hence change its weight, and
707 * backlogged; in this respect, each of the two events
708 * causes an invocation of this function,
709 * 2) this is the invocation of this function caused by the
710 * second event. This second invocation is actually useless,
711 * and we handle this fact by exiting immediately. More
712 * efficient or clearer solutions might possibly be adopted.
713 */
714 if (bfqq->weight_counter)
715 return;
716
717 while (*new) {
718 struct bfq_weight_counter *__counter = container_of(*new,
719 struct bfq_weight_counter,
720 weights_node);
721 parent = *new;
722
723 if (entity->weight == __counter->weight) {
724 bfqq->weight_counter = __counter;
725 goto inc_counter;
726 }
727 if (entity->weight < __counter->weight)
728 new = &((*new)->rb_left);
729 else
730 new = &((*new)->rb_right);
731 }
732
733 bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
734 GFP_ATOMIC);
735
736 /*
737 * In the unlucky event of an allocation failure, we just
738 * exit. This will cause the weight of queue to not be
739 * considered in bfq_symmetric_scenario, which, in its turn,
740 * causes the scenario to be deemed wrongly symmetric in case
741 * bfqq's weight would have been the only weight making the
742 * scenario asymmetric. On the bright side, no unbalance will
743 * however occur when bfqq becomes inactive again (the
744 * invocation of this function is triggered by an activation
745 * of queue). In fact, bfq_weights_tree_remove does nothing
746 * if !bfqq->weight_counter.
747 */
748 if (unlikely(!bfqq->weight_counter))
749 return;
750
751 bfqq->weight_counter->weight = entity->weight;
752 rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
753 rb_insert_color(&bfqq->weight_counter->weights_node, root);
754
755inc_counter:
756 bfqq->weight_counter->num_active++;
757 bfqq->ref++;
758}
759
760/*
761 * Decrement the weight counter associated with the queue, and, if the
762 * counter reaches 0, remove the counter from the tree.
763 * See the comments to the function bfq_weights_tree_add() for considerations
764 * about overhead.
765 */
766void __bfq_weights_tree_remove(struct bfq_data *bfqd,
767 struct bfq_queue *bfqq,
768 struct rb_root *root)
769{
770 if (!bfqq->weight_counter)
771 return;
772
773 bfqq->weight_counter->num_active--;
774 if (bfqq->weight_counter->num_active > 0)
775 goto reset_entity_pointer;
776
777 rb_erase(&bfqq->weight_counter->weights_node, root);
778 kfree(bfqq->weight_counter);
779
780reset_entity_pointer:
781 bfqq->weight_counter = NULL;
782 bfq_put_queue(bfqq);
783}
784
785/*
786 * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
787 * of active groups for each queue's inactive parent entity.
788 */
789void bfq_weights_tree_remove(struct bfq_data *bfqd,
790 struct bfq_queue *bfqq)
791{
792 struct bfq_entity *entity = bfqq->entity.parent;
793
794 for_each_entity(entity) {
795 struct bfq_sched_data *sd = entity->my_sched_data;
796
797 if (sd->next_in_service || sd->in_service_entity) {
798 /*
799 * entity is still active, because either
800 * next_in_service or in_service_entity is not
801 * NULL (see the comments on the definition of
802 * next_in_service for details on why
803 * in_service_entity must be checked too).
804 *
805 * As a consequence, its parent entities are
806 * active as well, and thus this loop must
807 * stop here.
808 */
809 break;
810 }
811
812 /*
813 * The decrement of num_groups_with_pending_reqs is
814 * not performed immediately upon the deactivation of
815 * entity, but it is delayed to when it also happens
816 * that the first leaf descendant bfqq of entity gets
817 * all its pending requests completed. The following
818 * instructions perform this delayed decrement, if
819 * needed. See the comments on
820 * num_groups_with_pending_reqs for details.
821 */
822 if (entity->in_groups_with_pending_reqs) {
823 entity->in_groups_with_pending_reqs = false;
824 bfqd->num_groups_with_pending_reqs--;
825 }
826 }
827
828 /*
829 * Next function is invoked last, because it causes bfqq to be
830 * freed if the following holds: bfqq is not in service and
831 * has no dispatched request. DO NOT use bfqq after the next
832 * function invocation.
833 */
834 __bfq_weights_tree_remove(bfqd, bfqq,
835 &bfqd->queue_weights_tree);
836}
837
838/*
839 * Return expired entry, or NULL to just start from scratch in rbtree.
840 */
841static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
842 struct request *last)
843{
844 struct request *rq;
845
846 if (bfq_bfqq_fifo_expire(bfqq))
847 return NULL;
848
849 bfq_mark_bfqq_fifo_expire(bfqq);
850
851 rq = rq_entry_fifo(bfqq->fifo.next);
852
853 if (rq == last || ktime_get_ns() < rq->fifo_time)
854 return NULL;
855
856 bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
857 return rq;
858}
859
860static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
861 struct bfq_queue *bfqq,
862 struct request *last)
863{
864 struct rb_node *rbnext = rb_next(&last->rb_node);
865 struct rb_node *rbprev = rb_prev(&last->rb_node);
866 struct request *next, *prev = NULL;
867
868 /* Follow expired path, else get first next available. */
869 next = bfq_check_fifo(bfqq, last);
870 if (next)
871 return next;
872
873 if (rbprev)
874 prev = rb_entry_rq(rbprev);
875
876 if (rbnext)
877 next = rb_entry_rq(rbnext);
878 else {
879 rbnext = rb_first(&bfqq->sort_list);
880 if (rbnext && rbnext != &last->rb_node)
881 next = rb_entry_rq(rbnext);
882 }
883
884 return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
885}
886
887/* see the definition of bfq_async_charge_factor for details */
888static unsigned long bfq_serv_to_charge(struct request *rq,
889 struct bfq_queue *bfqq)
890{
891 if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
892 !bfq_symmetric_scenario(bfqq->bfqd))
893 return blk_rq_sectors(rq);
894
895 return blk_rq_sectors(rq) * bfq_async_charge_factor;
896}
897
898/**
899 * bfq_updated_next_req - update the queue after a new next_rq selection.
900 * @bfqd: the device data the queue belongs to.
901 * @bfqq: the queue to update.
902 *
903 * If the first request of a queue changes we make sure that the queue
904 * has enough budget to serve at least its first request (if the
905 * request has grown). We do this because if the queue has not enough
906 * budget for its first request, it has to go through two dispatch
907 * rounds to actually get it dispatched.
908 */
909static void bfq_updated_next_req(struct bfq_data *bfqd,
910 struct bfq_queue *bfqq)
911{
912 struct bfq_entity *entity = &bfqq->entity;
913 struct request *next_rq = bfqq->next_rq;
914 unsigned long new_budget;
915
916 if (!next_rq)
917 return;
918
919 if (bfqq == bfqd->in_service_queue)
920 /*
921 * In order not to break guarantees, budgets cannot be
922 * changed after an entity has been selected.
923 */
924 return;
925
926 new_budget = max_t(unsigned long,
927 max_t(unsigned long, bfqq->max_budget,
928 bfq_serv_to_charge(next_rq, bfqq)),
929 entity->service);
930 if (entity->budget != new_budget) {
931 entity->budget = new_budget;
932 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
933 new_budget);
934 bfq_requeue_bfqq(bfqd, bfqq, false);
935 }
936}
937
938static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
939{
940 u64 dur;
941
942 if (bfqd->bfq_wr_max_time > 0)
943 return bfqd->bfq_wr_max_time;
944
945 dur = bfqd->rate_dur_prod;
946 do_div(dur, bfqd->peak_rate);
947
948 /*
949 * Limit duration between 3 and 25 seconds. The upper limit
950 * has been conservatively set after the following worst case:
951 * on a QEMU/KVM virtual machine
952 * - running in a slow PC
953 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
954 * - serving a heavy I/O workload, such as the sequential reading
955 * of several files
956 * mplayer took 23 seconds to start, if constantly weight-raised.
957 *
958 * As for higher values than that accomodating the above bad
959 * scenario, tests show that higher values would often yield
960 * the opposite of the desired result, i.e., would worsen
961 * responsiveness by allowing non-interactive applications to
962 * preserve weight raising for too long.
963 *
964 * On the other end, lower values than 3 seconds make it
965 * difficult for most interactive tasks to complete their jobs
966 * before weight-raising finishes.
967 */
968 return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
969}
970
971/* switch back from soft real-time to interactive weight raising */
972static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
973 struct bfq_data *bfqd)
974{
975 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
976 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
977 bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
978}
979
980static void
981bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
982 struct bfq_io_cq *bic, bool bfq_already_existing)
983{
984 unsigned int old_wr_coeff = bfqq->wr_coeff;
985 bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
986
987 if (bic->saved_has_short_ttime)
988 bfq_mark_bfqq_has_short_ttime(bfqq);
989 else
990 bfq_clear_bfqq_has_short_ttime(bfqq);
991
992 if (bic->saved_IO_bound)
993 bfq_mark_bfqq_IO_bound(bfqq);
994 else
995 bfq_clear_bfqq_IO_bound(bfqq);
996
997 bfqq->ttime = bic->saved_ttime;
998 bfqq->wr_coeff = bic->saved_wr_coeff;
999 bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
1000 bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
1001 bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
1002
1003 if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1004 time_is_before_jiffies(bfqq->last_wr_start_finish +
1005 bfqq->wr_cur_max_time))) {
1006 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1007 !bfq_bfqq_in_large_burst(bfqq) &&
1008 time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1009 bfq_wr_duration(bfqd))) {
1010 switch_back_to_interactive_wr(bfqq, bfqd);
1011 } else {
1012 bfqq->wr_coeff = 1;
1013 bfq_log_bfqq(bfqq->bfqd, bfqq,
1014 "resume state: switching off wr");
1015 }
1016 }
1017
1018 /* make sure weight will be updated, however we got here */
1019 bfqq->entity.prio_changed = 1;
1020
1021 if (likely(!busy))
1022 return;
1023
1024 if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1025 bfqd->wr_busy_queues++;
1026 else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1027 bfqd->wr_busy_queues--;
1028}
1029
1030static int bfqq_process_refs(struct bfq_queue *bfqq)
1031{
1032 return bfqq->ref - bfqq->allocated - bfqq->entity.on_st -
1033 (bfqq->weight_counter != NULL);
1034}
1035
1036/* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1037static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1038{
1039 struct bfq_queue *item;
1040 struct hlist_node *n;
1041
1042 hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1043 hlist_del_init(&item->burst_list_node);
1044 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1045 bfqd->burst_size = 1;
1046 bfqd->burst_parent_entity = bfqq->entity.parent;
1047}
1048
1049/* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1050static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1051{
1052 /* Increment burst size to take into account also bfqq */
1053 bfqd->burst_size++;
1054
1055 if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1056 struct bfq_queue *pos, *bfqq_item;
1057 struct hlist_node *n;
1058
1059 /*
1060 * Enough queues have been activated shortly after each
1061 * other to consider this burst as large.
1062 */
1063 bfqd->large_burst = true;
1064
1065 /*
1066 * We can now mark all queues in the burst list as
1067 * belonging to a large burst.
1068 */
1069 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1070 burst_list_node)
1071 bfq_mark_bfqq_in_large_burst(bfqq_item);
1072 bfq_mark_bfqq_in_large_burst(bfqq);
1073
1074 /*
1075 * From now on, and until the current burst finishes, any
1076 * new queue being activated shortly after the last queue
1077 * was inserted in the burst can be immediately marked as
1078 * belonging to a large burst. So the burst list is not
1079 * needed any more. Remove it.
1080 */
1081 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1082 burst_list_node)
1083 hlist_del_init(&pos->burst_list_node);
1084 } else /*
1085 * Burst not yet large: add bfqq to the burst list. Do
1086 * not increment the ref counter for bfqq, because bfqq
1087 * is removed from the burst list before freeing bfqq
1088 * in put_queue.
1089 */
1090 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1091}
1092
1093/*
1094 * If many queues belonging to the same group happen to be created
1095 * shortly after each other, then the processes associated with these
1096 * queues have typically a common goal. In particular, bursts of queue
1097 * creations are usually caused by services or applications that spawn
1098 * many parallel threads/processes. Examples are systemd during boot,
1099 * or git grep. To help these processes get their job done as soon as
1100 * possible, it is usually better to not grant either weight-raising
1101 * or device idling to their queues.
1102 *
1103 * In this comment we describe, firstly, the reasons why this fact
1104 * holds, and, secondly, the next function, which implements the main
1105 * steps needed to properly mark these queues so that they can then be
1106 * treated in a different way.
1107 *
1108 * The above services or applications benefit mostly from a high
1109 * throughput: the quicker the requests of the activated queues are
1110 * cumulatively served, the sooner the target job of these queues gets
1111 * completed. As a consequence, weight-raising any of these queues,
1112 * which also implies idling the device for it, is almost always
1113 * counterproductive. In most cases it just lowers throughput.
1114 *
1115 * On the other hand, a burst of queue creations may be caused also by
1116 * the start of an application that does not consist of a lot of
1117 * parallel I/O-bound threads. In fact, with a complex application,
1118 * several short processes may need to be executed to start-up the
1119 * application. In this respect, to start an application as quickly as
1120 * possible, the best thing to do is in any case to privilege the I/O
1121 * related to the application with respect to all other
1122 * I/O. Therefore, the best strategy to start as quickly as possible
1123 * an application that causes a burst of queue creations is to
1124 * weight-raise all the queues created during the burst. This is the
1125 * exact opposite of the best strategy for the other type of bursts.
1126 *
1127 * In the end, to take the best action for each of the two cases, the
1128 * two types of bursts need to be distinguished. Fortunately, this
1129 * seems relatively easy, by looking at the sizes of the bursts. In
1130 * particular, we found a threshold such that only bursts with a
1131 * larger size than that threshold are apparently caused by
1132 * services or commands such as systemd or git grep. For brevity,
1133 * hereafter we call just 'large' these bursts. BFQ *does not*
1134 * weight-raise queues whose creation occurs in a large burst. In
1135 * addition, for each of these queues BFQ performs or does not perform
1136 * idling depending on which choice boosts the throughput more. The
1137 * exact choice depends on the device and request pattern at
1138 * hand.
1139 *
1140 * Unfortunately, false positives may occur while an interactive task
1141 * is starting (e.g., an application is being started). The
1142 * consequence is that the queues associated with the task do not
1143 * enjoy weight raising as expected. Fortunately these false positives
1144 * are very rare. They typically occur if some service happens to
1145 * start doing I/O exactly when the interactive task starts.
1146 *
1147 * Turning back to the next function, it implements all the steps
1148 * needed to detect the occurrence of a large burst and to properly
1149 * mark all the queues belonging to it (so that they can then be
1150 * treated in a different way). This goal is achieved by maintaining a
1151 * "burst list" that holds, temporarily, the queues that belong to the
1152 * burst in progress. The list is then used to mark these queues as
1153 * belonging to a large burst if the burst does become large. The main
1154 * steps are the following.
1155 *
1156 * . when the very first queue is created, the queue is inserted into the
1157 * list (as it could be the first queue in a possible burst)
1158 *
1159 * . if the current burst has not yet become large, and a queue Q that does
1160 * not yet belong to the burst is activated shortly after the last time
1161 * at which a new queue entered the burst list, then the function appends
1162 * Q to the burst list
1163 *
1164 * . if, as a consequence of the previous step, the burst size reaches
1165 * the large-burst threshold, then
1166 *
1167 * . all the queues in the burst list are marked as belonging to a
1168 * large burst
1169 *
1170 * . the burst list is deleted; in fact, the burst list already served
1171 * its purpose (keeping temporarily track of the queues in a burst,
1172 * so as to be able to mark them as belonging to a large burst in the
1173 * previous sub-step), and now is not needed any more
1174 *
1175 * . the device enters a large-burst mode
1176 *
1177 * . if a queue Q that does not belong to the burst is created while
1178 * the device is in large-burst mode and shortly after the last time
1179 * at which a queue either entered the burst list or was marked as
1180 * belonging to the current large burst, then Q is immediately marked
1181 * as belonging to a large burst.
1182 *
1183 * . if a queue Q that does not belong to the burst is created a while
1184 * later, i.e., not shortly after, than the last time at which a queue
1185 * either entered the burst list or was marked as belonging to the
1186 * current large burst, then the current burst is deemed as finished and:
1187 *
1188 * . the large-burst mode is reset if set
1189 *
1190 * . the burst list is emptied
1191 *
1192 * . Q is inserted in the burst list, as Q may be the first queue
1193 * in a possible new burst (then the burst list contains just Q
1194 * after this step).
1195 */
1196static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1197{
1198 /*
1199 * If bfqq is already in the burst list or is part of a large
1200 * burst, or finally has just been split, then there is
1201 * nothing else to do.
1202 */
1203 if (!hlist_unhashed(&bfqq->burst_list_node) ||
1204 bfq_bfqq_in_large_burst(bfqq) ||
1205 time_is_after_eq_jiffies(bfqq->split_time +
1206 msecs_to_jiffies(10)))
1207 return;
1208
1209 /*
1210 * If bfqq's creation happens late enough, or bfqq belongs to
1211 * a different group than the burst group, then the current
1212 * burst is finished, and related data structures must be
1213 * reset.
1214 *
1215 * In this respect, consider the special case where bfqq is
1216 * the very first queue created after BFQ is selected for this
1217 * device. In this case, last_ins_in_burst and
1218 * burst_parent_entity are not yet significant when we get
1219 * here. But it is easy to verify that, whether or not the
1220 * following condition is true, bfqq will end up being
1221 * inserted into the burst list. In particular the list will
1222 * happen to contain only bfqq. And this is exactly what has
1223 * to happen, as bfqq may be the first queue of the first
1224 * burst.
1225 */
1226 if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1227 bfqd->bfq_burst_interval) ||
1228 bfqq->entity.parent != bfqd->burst_parent_entity) {
1229 bfqd->large_burst = false;
1230 bfq_reset_burst_list(bfqd, bfqq);
1231 goto end;
1232 }
1233
1234 /*
1235 * If we get here, then bfqq is being activated shortly after the
1236 * last queue. So, if the current burst is also large, we can mark
1237 * bfqq as belonging to this large burst immediately.
1238 */
1239 if (bfqd->large_burst) {
1240 bfq_mark_bfqq_in_large_burst(bfqq);
1241 goto end;
1242 }
1243
1244 /*
1245 * If we get here, then a large-burst state has not yet been
1246 * reached, but bfqq is being activated shortly after the last
1247 * queue. Then we add bfqq to the burst.
1248 */
1249 bfq_add_to_burst(bfqd, bfqq);
1250end:
1251 /*
1252 * At this point, bfqq either has been added to the current
1253 * burst or has caused the current burst to terminate and a
1254 * possible new burst to start. In particular, in the second
1255 * case, bfqq has become the first queue in the possible new
1256 * burst. In both cases last_ins_in_burst needs to be moved
1257 * forward.
1258 */
1259 bfqd->last_ins_in_burst = jiffies;
1260}
1261
1262static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1263{
1264 struct bfq_entity *entity = &bfqq->entity;
1265
1266 return entity->budget - entity->service;
1267}
1268
1269/*
1270 * If enough samples have been computed, return the current max budget
1271 * stored in bfqd, which is dynamically updated according to the
1272 * estimated disk peak rate; otherwise return the default max budget
1273 */
1274static int bfq_max_budget(struct bfq_data *bfqd)
1275{
1276 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1277 return bfq_default_max_budget;
1278 else
1279 return bfqd->bfq_max_budget;
1280}
1281
1282/*
1283 * Return min budget, which is a fraction of the current or default
1284 * max budget (trying with 1/32)
1285 */
1286static int bfq_min_budget(struct bfq_data *bfqd)
1287{
1288 if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1289 return bfq_default_max_budget / 32;
1290 else
1291 return bfqd->bfq_max_budget / 32;
1292}
1293
1294/*
1295 * The next function, invoked after the input queue bfqq switches from
1296 * idle to busy, updates the budget of bfqq. The function also tells
1297 * whether the in-service queue should be expired, by returning
1298 * true. The purpose of expiring the in-service queue is to give bfqq
1299 * the chance to possibly preempt the in-service queue, and the reason
1300 * for preempting the in-service queue is to achieve one of the two
1301 * goals below.
1302 *
1303 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1304 * expired because it has remained idle. In particular, bfqq may have
1305 * expired for one of the following two reasons:
1306 *
1307 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1308 * and did not make it to issue a new request before its last
1309 * request was served;
1310 *
1311 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1312 * a new request before the expiration of the idling-time.
1313 *
1314 * Even if bfqq has expired for one of the above reasons, the process
1315 * associated with the queue may be however issuing requests greedily,
1316 * and thus be sensitive to the bandwidth it receives (bfqq may have
1317 * remained idle for other reasons: CPU high load, bfqq not enjoying
1318 * idling, I/O throttling somewhere in the path from the process to
1319 * the I/O scheduler, ...). But if, after every expiration for one of
1320 * the above two reasons, bfqq has to wait for the service of at least
1321 * one full budget of another queue before being served again, then
1322 * bfqq is likely to get a much lower bandwidth or resource time than
1323 * its reserved ones. To address this issue, two countermeasures need
1324 * to be taken.
1325 *
1326 * First, the budget and the timestamps of bfqq need to be updated in
1327 * a special way on bfqq reactivation: they need to be updated as if
1328 * bfqq did not remain idle and did not expire. In fact, if they are
1329 * computed as if bfqq expired and remained idle until reactivation,
1330 * then the process associated with bfqq is treated as if, instead of
1331 * being greedy, it stopped issuing requests when bfqq remained idle,
1332 * and restarts issuing requests only on this reactivation. In other
1333 * words, the scheduler does not help the process recover the "service
1334 * hole" between bfqq expiration and reactivation. As a consequence,
1335 * the process receives a lower bandwidth than its reserved one. In
1336 * contrast, to recover this hole, the budget must be updated as if
1337 * bfqq was not expired at all before this reactivation, i.e., it must
1338 * be set to the value of the remaining budget when bfqq was
1339 * expired. Along the same line, timestamps need to be assigned the
1340 * value they had the last time bfqq was selected for service, i.e.,
1341 * before last expiration. Thus timestamps need to be back-shifted
1342 * with respect to their normal computation (see [1] for more details
1343 * on this tricky aspect).
1344 *
1345 * Secondly, to allow the process to recover the hole, the in-service
1346 * queue must be expired too, to give bfqq the chance to preempt it
1347 * immediately. In fact, if bfqq has to wait for a full budget of the
1348 * in-service queue to be completed, then it may become impossible to
1349 * let the process recover the hole, even if the back-shifted
1350 * timestamps of bfqq are lower than those of the in-service queue. If
1351 * this happens for most or all of the holes, then the process may not
1352 * receive its reserved bandwidth. In this respect, it is worth noting
1353 * that, being the service of outstanding requests unpreemptible, a
1354 * little fraction of the holes may however be unrecoverable, thereby
1355 * causing a little loss of bandwidth.
1356 *
1357 * The last important point is detecting whether bfqq does need this
1358 * bandwidth recovery. In this respect, the next function deems the
1359 * process associated with bfqq greedy, and thus allows it to recover
1360 * the hole, if: 1) the process is waiting for the arrival of a new
1361 * request (which implies that bfqq expired for one of the above two
1362 * reasons), and 2) such a request has arrived soon. The first
1363 * condition is controlled through the flag non_blocking_wait_rq,
1364 * while the second through the flag arrived_in_time. If both
1365 * conditions hold, then the function computes the budget in the
1366 * above-described special way, and signals that the in-service queue
1367 * should be expired. Timestamp back-shifting is done later in
1368 * __bfq_activate_entity.
1369 *
1370 * 2. Reduce latency. Even if timestamps are not backshifted to let
1371 * the process associated with bfqq recover a service hole, bfqq may
1372 * however happen to have, after being (re)activated, a lower finish
1373 * timestamp than the in-service queue. That is, the next budget of
1374 * bfqq may have to be completed before the one of the in-service
1375 * queue. If this is the case, then preempting the in-service queue
1376 * allows this goal to be achieved, apart from the unpreemptible,
1377 * outstanding requests mentioned above.
1378 *
1379 * Unfortunately, regardless of which of the above two goals one wants
1380 * to achieve, service trees need first to be updated to know whether
1381 * the in-service queue must be preempted. To have service trees
1382 * correctly updated, the in-service queue must be expired and
1383 * rescheduled, and bfqq must be scheduled too. This is one of the
1384 * most costly operations (in future versions, the scheduling
1385 * mechanism may be re-designed in such a way to make it possible to
1386 * know whether preemption is needed without needing to update service
1387 * trees). In addition, queue preemptions almost always cause random
1388 * I/O, and thus loss of throughput. Because of these facts, the next
1389 * function adopts the following simple scheme to avoid both costly
1390 * operations and too frequent preemptions: it requests the expiration
1391 * of the in-service queue (unconditionally) only for queues that need
1392 * to recover a hole, or that either are weight-raised or deserve to
1393 * be weight-raised.
1394 */
1395static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1396 struct bfq_queue *bfqq,
1397 bool arrived_in_time,
1398 bool wr_or_deserves_wr)
1399{
1400 struct bfq_entity *entity = &bfqq->entity;
1401
1402 /*
1403 * In the next compound condition, we check also whether there
1404 * is some budget left, because otherwise there is no point in
1405 * trying to go on serving bfqq with this same budget: bfqq
1406 * would be expired immediately after being selected for
1407 * service. This would only cause useless overhead.
1408 */
1409 if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1410 bfq_bfqq_budget_left(bfqq) > 0) {
1411 /*
1412 * We do not clear the flag non_blocking_wait_rq here, as
1413 * the latter is used in bfq_activate_bfqq to signal
1414 * that timestamps need to be back-shifted (and is
1415 * cleared right after).
1416 */
1417
1418 /*
1419 * In next assignment we rely on that either
1420 * entity->service or entity->budget are not updated
1421 * on expiration if bfqq is empty (see
1422 * __bfq_bfqq_recalc_budget). Thus both quantities
1423 * remain unchanged after such an expiration, and the
1424 * following statement therefore assigns to
1425 * entity->budget the remaining budget on such an
1426 * expiration.
1427 */
1428 entity->budget = min_t(unsigned long,
1429 bfq_bfqq_budget_left(bfqq),
1430 bfqq->max_budget);
1431
1432 /*
1433 * At this point, we have used entity->service to get
1434 * the budget left (needed for updating
1435 * entity->budget). Thus we finally can, and have to,
1436 * reset entity->service. The latter must be reset
1437 * because bfqq would otherwise be charged again for
1438 * the service it has received during its previous
1439 * service slot(s).
1440 */
1441 entity->service = 0;
1442
1443 return true;
1444 }
1445
1446 /*
1447 * We can finally complete expiration, by setting service to 0.
1448 */
1449 entity->service = 0;
1450 entity->budget = max_t(unsigned long, bfqq->max_budget,
1451 bfq_serv_to_charge(bfqq->next_rq, bfqq));
1452 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1453 return wr_or_deserves_wr;
1454}
1455
1456/*
1457 * Return the farthest past time instant according to jiffies
1458 * macros.
1459 */
1460static unsigned long bfq_smallest_from_now(void)
1461{
1462 return jiffies - MAX_JIFFY_OFFSET;
1463}
1464
1465static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1466 struct bfq_queue *bfqq,
1467 unsigned int old_wr_coeff,
1468 bool wr_or_deserves_wr,
1469 bool interactive,
1470 bool in_burst,
1471 bool soft_rt)
1472{
1473 if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1474 /* start a weight-raising period */
1475 if (interactive) {
1476 bfqq->service_from_wr = 0;
1477 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1478 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1479 } else {
1480 /*
1481 * No interactive weight raising in progress
1482 * here: assign minus infinity to
1483 * wr_start_at_switch_to_srt, to make sure
1484 * that, at the end of the soft-real-time
1485 * weight raising periods that is starting
1486 * now, no interactive weight-raising period
1487 * may be wrongly considered as still in
1488 * progress (and thus actually started by
1489 * mistake).
1490 */
1491 bfqq->wr_start_at_switch_to_srt =
1492 bfq_smallest_from_now();
1493 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1494 BFQ_SOFTRT_WEIGHT_FACTOR;
1495 bfqq->wr_cur_max_time =
1496 bfqd->bfq_wr_rt_max_time;
1497 }
1498
1499 /*
1500 * If needed, further reduce budget to make sure it is
1501 * close to bfqq's backlog, so as to reduce the
1502 * scheduling-error component due to a too large
1503 * budget. Do not care about throughput consequences,
1504 * but only about latency. Finally, do not assign a
1505 * too small budget either, to avoid increasing
1506 * latency by causing too frequent expirations.
1507 */
1508 bfqq->entity.budget = min_t(unsigned long,
1509 bfqq->entity.budget,
1510 2 * bfq_min_budget(bfqd));
1511 } else if (old_wr_coeff > 1) {
1512 if (interactive) { /* update wr coeff and duration */
1513 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1514 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1515 } else if (in_burst)
1516 bfqq->wr_coeff = 1;
1517 else if (soft_rt) {
1518 /*
1519 * The application is now or still meeting the
1520 * requirements for being deemed soft rt. We
1521 * can then correctly and safely (re)charge
1522 * the weight-raising duration for the
1523 * application with the weight-raising
1524 * duration for soft rt applications.
1525 *
1526 * In particular, doing this recharge now, i.e.,
1527 * before the weight-raising period for the
1528 * application finishes, reduces the probability
1529 * of the following negative scenario:
1530 * 1) the weight of a soft rt application is
1531 * raised at startup (as for any newly
1532 * created application),
1533 * 2) since the application is not interactive,
1534 * at a certain time weight-raising is
1535 * stopped for the application,
1536 * 3) at that time the application happens to
1537 * still have pending requests, and hence
1538 * is destined to not have a chance to be
1539 * deemed soft rt before these requests are
1540 * completed (see the comments to the
1541 * function bfq_bfqq_softrt_next_start()
1542 * for details on soft rt detection),
1543 * 4) these pending requests experience a high
1544 * latency because the application is not
1545 * weight-raised while they are pending.
1546 */
1547 if (bfqq->wr_cur_max_time !=
1548 bfqd->bfq_wr_rt_max_time) {
1549 bfqq->wr_start_at_switch_to_srt =
1550 bfqq->last_wr_start_finish;
1551
1552 bfqq->wr_cur_max_time =
1553 bfqd->bfq_wr_rt_max_time;
1554 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1555 BFQ_SOFTRT_WEIGHT_FACTOR;
1556 }
1557 bfqq->last_wr_start_finish = jiffies;
1558 }
1559 }
1560}
1561
1562static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1563 struct bfq_queue *bfqq)
1564{
1565 return bfqq->dispatched == 0 &&
1566 time_is_before_jiffies(
1567 bfqq->budget_timeout +
1568 bfqd->bfq_wr_min_idle_time);
1569}
1570
1571static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1572 struct bfq_queue *bfqq,
1573 int old_wr_coeff,
1574 struct request *rq,
1575 bool *interactive)
1576{
1577 bool soft_rt, in_burst, wr_or_deserves_wr,
1578 bfqq_wants_to_preempt,
1579 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1580 /*
1581 * See the comments on
1582 * bfq_bfqq_update_budg_for_activation for
1583 * details on the usage of the next variable.
1584 */
1585 arrived_in_time = ktime_get_ns() <=
1586 bfqq->ttime.last_end_request +
1587 bfqd->bfq_slice_idle * 3;
1588
1589
1590 /*
1591 * bfqq deserves to be weight-raised if:
1592 * - it is sync,
1593 * - it does not belong to a large burst,
1594 * - it has been idle for enough time or is soft real-time,
1595 * - is linked to a bfq_io_cq (it is not shared in any sense).
1596 */
1597 in_burst = bfq_bfqq_in_large_burst(bfqq);
1598 soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1599 !in_burst &&
1600 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1601 bfqq->dispatched == 0;
1602 *interactive = !in_burst && idle_for_long_time;
1603 wr_or_deserves_wr = bfqd->low_latency &&
1604 (bfqq->wr_coeff > 1 ||
1605 (bfq_bfqq_sync(bfqq) &&
1606 bfqq->bic && (*interactive || soft_rt)));
1607
1608 /*
1609 * Using the last flag, update budget and check whether bfqq
1610 * may want to preempt the in-service queue.
1611 */
1612 bfqq_wants_to_preempt =
1613 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1614 arrived_in_time,
1615 wr_or_deserves_wr);
1616
1617 /*
1618 * If bfqq happened to be activated in a burst, but has been
1619 * idle for much more than an interactive queue, then we
1620 * assume that, in the overall I/O initiated in the burst, the
1621 * I/O associated with bfqq is finished. So bfqq does not need
1622 * to be treated as a queue belonging to a burst
1623 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1624 * if set, and remove bfqq from the burst list if it's
1625 * there. We do not decrement burst_size, because the fact
1626 * that bfqq does not need to belong to the burst list any
1627 * more does not invalidate the fact that bfqq was created in
1628 * a burst.
1629 */
1630 if (likely(!bfq_bfqq_just_created(bfqq)) &&
1631 idle_for_long_time &&
1632 time_is_before_jiffies(
1633 bfqq->budget_timeout +
1634 msecs_to_jiffies(10000))) {
1635 hlist_del_init(&bfqq->burst_list_node);
1636 bfq_clear_bfqq_in_large_burst(bfqq);
1637 }
1638
1639 bfq_clear_bfqq_just_created(bfqq);
1640
1641
1642 if (!bfq_bfqq_IO_bound(bfqq)) {
1643 if (arrived_in_time) {
1644 bfqq->requests_within_timer++;
1645 if (bfqq->requests_within_timer >=
1646 bfqd->bfq_requests_within_timer)
1647 bfq_mark_bfqq_IO_bound(bfqq);
1648 } else
1649 bfqq->requests_within_timer = 0;
1650 }
1651
1652 if (bfqd->low_latency) {
1653 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1654 /* wraparound */
1655 bfqq->split_time =
1656 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1657
1658 if (time_is_before_jiffies(bfqq->split_time +
1659 bfqd->bfq_wr_min_idle_time)) {
1660 bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1661 old_wr_coeff,
1662 wr_or_deserves_wr,
1663 *interactive,
1664 in_burst,
1665 soft_rt);
1666
1667 if (old_wr_coeff != bfqq->wr_coeff)
1668 bfqq->entity.prio_changed = 1;
1669 }
1670 }
1671
1672 bfqq->last_idle_bklogged = jiffies;
1673 bfqq->service_from_backlogged = 0;
1674 bfq_clear_bfqq_softrt_update(bfqq);
1675
1676 bfq_add_bfqq_busy(bfqd, bfqq);
1677
1678 /*
1679 * Expire in-service queue only if preemption may be needed
1680 * for guarantees. In this respect, the function
1681 * next_queue_may_preempt just checks a simple, necessary
1682 * condition, and not a sufficient condition based on
1683 * timestamps. In fact, for the latter condition to be
1684 * evaluated, timestamps would need first to be updated, and
1685 * this operation is quite costly (see the comments on the
1686 * function bfq_bfqq_update_budg_for_activation).
1687 */
1688 if (bfqd->in_service_queue && bfqq_wants_to_preempt &&
1689 bfqd->in_service_queue->wr_coeff < bfqq->wr_coeff &&
1690 next_queue_may_preempt(bfqd))
1691 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1692 false, BFQQE_PREEMPTED);
1693}
1694
1695static void bfq_add_request(struct request *rq)
1696{
1697 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1698 struct bfq_data *bfqd = bfqq->bfqd;
1699 struct request *next_rq, *prev;
1700 unsigned int old_wr_coeff = bfqq->wr_coeff;
1701 bool interactive = false;
1702
1703 bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1704 bfqq->queued[rq_is_sync(rq)]++;
1705 bfqd->queued++;
1706
1707 elv_rb_add(&bfqq->sort_list, rq);
1708
1709 /*
1710 * Check if this request is a better next-serve candidate.
1711 */
1712 prev = bfqq->next_rq;
1713 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
1714 bfqq->next_rq = next_rq;
1715
1716 /*
1717 * Adjust priority tree position, if next_rq changes.
1718 */
1719 if (prev != bfqq->next_rq)
1720 bfq_pos_tree_add_move(bfqd, bfqq);
1721
1722 if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
1723 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
1724 rq, &interactive);
1725 else {
1726 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
1727 time_is_before_jiffies(
1728 bfqq->last_wr_start_finish +
1729 bfqd->bfq_wr_min_inter_arr_async)) {
1730 bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1731 bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1732
1733 bfqd->wr_busy_queues++;
1734 bfqq->entity.prio_changed = 1;
1735 }
1736 if (prev != bfqq->next_rq)
1737 bfq_updated_next_req(bfqd, bfqq);
1738 }
1739
1740 /*
1741 * Assign jiffies to last_wr_start_finish in the following
1742 * cases:
1743 *
1744 * . if bfqq is not going to be weight-raised, because, for
1745 * non weight-raised queues, last_wr_start_finish stores the
1746 * arrival time of the last request; as of now, this piece
1747 * of information is used only for deciding whether to
1748 * weight-raise async queues
1749 *
1750 * . if bfqq is not weight-raised, because, if bfqq is now
1751 * switching to weight-raised, then last_wr_start_finish
1752 * stores the time when weight-raising starts
1753 *
1754 * . if bfqq is interactive, because, regardless of whether
1755 * bfqq is currently weight-raised, the weight-raising
1756 * period must start or restart (this case is considered
1757 * separately because it is not detected by the above
1758 * conditions, if bfqq is already weight-raised)
1759 *
1760 * last_wr_start_finish has to be updated also if bfqq is soft
1761 * real-time, because the weight-raising period is constantly
1762 * restarted on idle-to-busy transitions for these queues, but
1763 * this is already done in bfq_bfqq_handle_idle_busy_switch if
1764 * needed.
1765 */
1766 if (bfqd->low_latency &&
1767 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
1768 bfqq->last_wr_start_finish = jiffies;
1769}
1770
1771static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
1772 struct bio *bio,
1773 struct request_queue *q)
1774{
1775 struct bfq_queue *bfqq = bfqd->bio_bfqq;
1776
1777
1778 if (bfqq)
1779 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
1780
1781 return NULL;
1782}
1783
1784static sector_t get_sdist(sector_t last_pos, struct request *rq)
1785{
1786 if (last_pos)
1787 return abs(blk_rq_pos(rq) - last_pos);
1788
1789 return 0;
1790}
1791
1792#if 0 /* Still not clear if we can do without next two functions */
1793static void bfq_activate_request(struct request_queue *q, struct request *rq)
1794{
1795 struct bfq_data *bfqd = q->elevator->elevator_data;
1796
1797 bfqd->rq_in_driver++;
1798}
1799
1800static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
1801{
1802 struct bfq_data *bfqd = q->elevator->elevator_data;
1803
1804 bfqd->rq_in_driver--;
1805}
1806#endif
1807
1808static void bfq_remove_request(struct request_queue *q,
1809 struct request *rq)
1810{
1811 struct bfq_queue *bfqq = RQ_BFQQ(rq);
1812 struct bfq_data *bfqd = bfqq->bfqd;
1813 const int sync = rq_is_sync(rq);
1814
1815 if (bfqq->next_rq == rq) {
1816 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
1817 bfq_updated_next_req(bfqd, bfqq);
1818 }
1819
1820 if (rq->queuelist.prev != &rq->queuelist)
1821 list_del_init(&rq->queuelist);
1822 bfqq->queued[sync]--;
1823 bfqd->queued--;
1824 elv_rb_del(&bfqq->sort_list, rq);
1825
1826 elv_rqhash_del(q, rq);
1827 if (q->last_merge == rq)
1828 q->last_merge = NULL;
1829
1830 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
1831 bfqq->next_rq = NULL;
1832
1833 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
1834 bfq_del_bfqq_busy(bfqd, bfqq, false);
1835 /*
1836 * bfqq emptied. In normal operation, when
1837 * bfqq is empty, bfqq->entity.service and
1838 * bfqq->entity.budget must contain,
1839 * respectively, the service received and the
1840 * budget used last time bfqq emptied. These
1841 * facts do not hold in this case, as at least
1842 * this last removal occurred while bfqq is
1843 * not in service. To avoid inconsistencies,
1844 * reset both bfqq->entity.service and
1845 * bfqq->entity.budget, if bfqq has still a
1846 * process that may issue I/O requests to it.
1847 */
1848 bfqq->entity.budget = bfqq->entity.service = 0;
1849 }
1850
1851 /*
1852 * Remove queue from request-position tree as it is empty.
1853 */
1854 if (bfqq->pos_root) {
1855 rb_erase(&bfqq->pos_node, bfqq->pos_root);
1856 bfqq->pos_root = NULL;
1857 }
1858 } else {
1859 bfq_pos_tree_add_move(bfqd, bfqq);
1860 }
1861
1862 if (rq->cmd_flags & REQ_META)
1863 bfqq->meta_pending--;
1864
1865}
1866
1867static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio)
1868{
1869 struct request_queue *q = hctx->queue;
1870 struct bfq_data *bfqd = q->elevator->elevator_data;
1871 struct request *free = NULL;
1872 /*
1873 * bfq_bic_lookup grabs the queue_lock: invoke it now and
1874 * store its return value for later use, to avoid nesting
1875 * queue_lock inside the bfqd->lock. We assume that the bic
1876 * returned by bfq_bic_lookup does not go away before
1877 * bfqd->lock is taken.
1878 */
1879 struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
1880 bool ret;
1881
1882 spin_lock_irq(&bfqd->lock);
1883
1884 if (bic)
1885 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
1886 else
1887 bfqd->bio_bfqq = NULL;
1888 bfqd->bio_bic = bic;
1889
1890 ret = blk_mq_sched_try_merge(q, bio, &free);
1891
1892 if (free)
1893 blk_mq_free_request(free);
1894 spin_unlock_irq(&bfqd->lock);
1895
1896 return ret;
1897}
1898
1899static int bfq_request_merge(struct request_queue *q, struct request **req,
1900 struct bio *bio)
1901{
1902 struct bfq_data *bfqd = q->elevator->elevator_data;
1903 struct request *__rq;
1904
1905 __rq = bfq_find_rq_fmerge(bfqd, bio, q);
1906 if (__rq && elv_bio_merge_ok(__rq, bio)) {
1907 *req = __rq;
1908 return ELEVATOR_FRONT_MERGE;
1909 }
1910
1911 return ELEVATOR_NO_MERGE;
1912}
1913
1914static struct bfq_queue *bfq_init_rq(struct request *rq);
1915
1916static void bfq_request_merged(struct request_queue *q, struct request *req,
1917 enum elv_merge type)
1918{
1919 if (type == ELEVATOR_FRONT_MERGE &&
1920 rb_prev(&req->rb_node) &&
1921 blk_rq_pos(req) <
1922 blk_rq_pos(container_of(rb_prev(&req->rb_node),
1923 struct request, rb_node))) {
1924 struct bfq_queue *bfqq = bfq_init_rq(req);
1925 struct bfq_data *bfqd = bfqq->bfqd;
1926 struct request *prev, *next_rq;
1927
1928 /* Reposition request in its sort_list */
1929 elv_rb_del(&bfqq->sort_list, req);
1930 elv_rb_add(&bfqq->sort_list, req);
1931
1932 /* Choose next request to be served for bfqq */
1933 prev = bfqq->next_rq;
1934 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
1935 bfqd->last_position);
1936 bfqq->next_rq = next_rq;
1937 /*
1938 * If next_rq changes, update both the queue's budget to
1939 * fit the new request and the queue's position in its
1940 * rq_pos_tree.
1941 */
1942 if (prev != bfqq->next_rq) {
1943 bfq_updated_next_req(bfqd, bfqq);
1944 bfq_pos_tree_add_move(bfqd, bfqq);
1945 }
1946 }
1947}
1948
1949/*
1950 * This function is called to notify the scheduler that the requests
1951 * rq and 'next' have been merged, with 'next' going away. BFQ
1952 * exploits this hook to address the following issue: if 'next' has a
1953 * fifo_time lower that rq, then the fifo_time of rq must be set to
1954 * the value of 'next', to not forget the greater age of 'next'.
1955 *
1956 * NOTE: in this function we assume that rq is in a bfq_queue, basing
1957 * on that rq is picked from the hash table q->elevator->hash, which,
1958 * in its turn, is filled only with I/O requests present in
1959 * bfq_queues, while BFQ is in use for the request queue q. In fact,
1960 * the function that fills this hash table (elv_rqhash_add) is called
1961 * only by bfq_insert_request.
1962 */
1963static void bfq_requests_merged(struct request_queue *q, struct request *rq,
1964 struct request *next)
1965{
1966 struct bfq_queue *bfqq = bfq_init_rq(rq),
1967 *next_bfqq = bfq_init_rq(next);
1968
1969 /*
1970 * If next and rq belong to the same bfq_queue and next is older
1971 * than rq, then reposition rq in the fifo (by substituting next
1972 * with rq). Otherwise, if next and rq belong to different
1973 * bfq_queues, never reposition rq: in fact, we would have to
1974 * reposition it with respect to next's position in its own fifo,
1975 * which would most certainly be too expensive with respect to
1976 * the benefits.
1977 */
1978 if (bfqq == next_bfqq &&
1979 !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1980 next->fifo_time < rq->fifo_time) {
1981 list_del_init(&rq->queuelist);
1982 list_replace_init(&next->queuelist, &rq->queuelist);
1983 rq->fifo_time = next->fifo_time;
1984 }
1985
1986 if (bfqq->next_rq == next)
1987 bfqq->next_rq = rq;
1988
1989 bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
1990}
1991
1992/* Must be called with bfqq != NULL */
1993static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
1994{
1995 if (bfq_bfqq_busy(bfqq))
1996 bfqq->bfqd->wr_busy_queues--;
1997 bfqq->wr_coeff = 1;
1998 bfqq->wr_cur_max_time = 0;
1999 bfqq->last_wr_start_finish = jiffies;
2000 /*
2001 * Trigger a weight change on the next invocation of
2002 * __bfq_entity_update_weight_prio.
2003 */
2004 bfqq->entity.prio_changed = 1;
2005}
2006
2007void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2008 struct bfq_group *bfqg)
2009{
2010 int i, j;
2011
2012 for (i = 0; i < 2; i++)
2013 for (j = 0; j < IOPRIO_BE_NR; j++)
2014 if (bfqg->async_bfqq[i][j])
2015 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2016 if (bfqg->async_idle_bfqq)
2017 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2018}
2019
2020static void bfq_end_wr(struct bfq_data *bfqd)
2021{
2022 struct bfq_queue *bfqq;
2023
2024 spin_lock_irq(&bfqd->lock);
2025
2026 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2027 bfq_bfqq_end_wr(bfqq);
2028 list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2029 bfq_bfqq_end_wr(bfqq);
2030 bfq_end_wr_async(bfqd);
2031
2032 spin_unlock_irq(&bfqd->lock);
2033}
2034
2035static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2036{
2037 if (request)
2038 return blk_rq_pos(io_struct);
2039 else
2040 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2041}
2042
2043static int bfq_rq_close_to_sector(void *io_struct, bool request,
2044 sector_t sector)
2045{
2046 return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2047 BFQQ_CLOSE_THR;
2048}
2049
2050static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2051 struct bfq_queue *bfqq,
2052 sector_t sector)
2053{
2054 struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
2055 struct rb_node *parent, *node;
2056 struct bfq_queue *__bfqq;
2057
2058 if (RB_EMPTY_ROOT(root))
2059 return NULL;
2060
2061 /*
2062 * First, if we find a request starting at the end of the last
2063 * request, choose it.
2064 */
2065 __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2066 if (__bfqq)
2067 return __bfqq;
2068
2069 /*
2070 * If the exact sector wasn't found, the parent of the NULL leaf
2071 * will contain the closest sector (rq_pos_tree sorted by
2072 * next_request position).
2073 */
2074 __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2075 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2076 return __bfqq;
2077
2078 if (blk_rq_pos(__bfqq->next_rq) < sector)
2079 node = rb_next(&__bfqq->pos_node);
2080 else
2081 node = rb_prev(&__bfqq->pos_node);
2082 if (!node)
2083 return NULL;
2084
2085 __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2086 if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2087 return __bfqq;
2088
2089 return NULL;
2090}
2091
2092static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2093 struct bfq_queue *cur_bfqq,
2094 sector_t sector)
2095{
2096 struct bfq_queue *bfqq;
2097
2098 /*
2099 * We shall notice if some of the queues are cooperating,
2100 * e.g., working closely on the same area of the device. In
2101 * that case, we can group them together and: 1) don't waste
2102 * time idling, and 2) serve the union of their requests in
2103 * the best possible order for throughput.
2104 */
2105 bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2106 if (!bfqq || bfqq == cur_bfqq)
2107 return NULL;
2108
2109 return bfqq;
2110}
2111
2112static struct bfq_queue *
2113bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2114{
2115 int process_refs, new_process_refs;
2116 struct bfq_queue *__bfqq;
2117
2118 /*
2119 * If there are no process references on the new_bfqq, then it is
2120 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2121 * may have dropped their last reference (not just their last process
2122 * reference).
2123 */
2124 if (!bfqq_process_refs(new_bfqq))
2125 return NULL;
2126
2127 /* Avoid a circular list and skip interim queue merges. */
2128 while ((__bfqq = new_bfqq->new_bfqq)) {
2129 if (__bfqq == bfqq)
2130 return NULL;
2131 new_bfqq = __bfqq;
2132 }
2133
2134 process_refs = bfqq_process_refs(bfqq);
2135 new_process_refs = bfqq_process_refs(new_bfqq);
2136 /*
2137 * If the process for the bfqq has gone away, there is no
2138 * sense in merging the queues.
2139 */
2140 if (process_refs == 0 || new_process_refs == 0)
2141 return NULL;
2142
2143 bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2144 new_bfqq->pid);
2145
2146 /*
2147 * Merging is just a redirection: the requests of the process
2148 * owning one of the two queues are redirected to the other queue.
2149 * The latter queue, in its turn, is set as shared if this is the
2150 * first time that the requests of some process are redirected to
2151 * it.
2152 *
2153 * We redirect bfqq to new_bfqq and not the opposite, because
2154 * we are in the context of the process owning bfqq, thus we
2155 * have the io_cq of this process. So we can immediately
2156 * configure this io_cq to redirect the requests of the
2157 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2158 * not available any more (new_bfqq->bic == NULL).
2159 *
2160 * Anyway, even in case new_bfqq coincides with the in-service
2161 * queue, redirecting requests the in-service queue is the
2162 * best option, as we feed the in-service queue with new
2163 * requests close to the last request served and, by doing so,
2164 * are likely to increase the throughput.
2165 */
2166 bfqq->new_bfqq = new_bfqq;
2167 new_bfqq->ref += process_refs;
2168 return new_bfqq;
2169}
2170
2171static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2172 struct bfq_queue *new_bfqq)
2173{
2174 if (bfq_too_late_for_merging(new_bfqq))
2175 return false;
2176
2177 if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2178 (bfqq->ioprio_class != new_bfqq->ioprio_class))
2179 return false;
2180
2181 /*
2182 * If either of the queues has already been detected as seeky,
2183 * then merging it with the other queue is unlikely to lead to
2184 * sequential I/O.
2185 */
2186 if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2187 return false;
2188
2189 /*
2190 * Interleaved I/O is known to be done by (some) applications
2191 * only for reads, so it does not make sense to merge async
2192 * queues.
2193 */
2194 if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2195 return false;
2196
2197 return true;
2198}
2199
2200/*
2201 * Attempt to schedule a merge of bfqq with the currently in-service
2202 * queue or with a close queue among the scheduled queues. Return
2203 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2204 * structure otherwise.
2205 *
2206 * The OOM queue is not allowed to participate to cooperation: in fact, since
2207 * the requests temporarily redirected to the OOM queue could be redirected
2208 * again to dedicated queues at any time, the state needed to correctly
2209 * handle merging with the OOM queue would be quite complex and expensive
2210 * to maintain. Besides, in such a critical condition as an out of memory,
2211 * the benefits of queue merging may be little relevant, or even negligible.
2212 *
2213 * WARNING: queue merging may impair fairness among non-weight raised
2214 * queues, for at least two reasons: 1) the original weight of a
2215 * merged queue may change during the merged state, 2) even being the
2216 * weight the same, a merged queue may be bloated with many more
2217 * requests than the ones produced by its originally-associated
2218 * process.
2219 */
2220static struct bfq_queue *
2221bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2222 void *io_struct, bool request)
2223{
2224 struct bfq_queue *in_service_bfqq, *new_bfqq;
2225
2226 /*
2227 * Prevent bfqq from being merged if it has been created too
2228 * long ago. The idea is that true cooperating processes, and
2229 * thus their associated bfq_queues, are supposed to be
2230 * created shortly after each other. This is the case, e.g.,
2231 * for KVM/QEMU and dump I/O threads. Basing on this
2232 * assumption, the following filtering greatly reduces the
2233 * probability that two non-cooperating processes, which just
2234 * happen to do close I/O for some short time interval, have
2235 * their queues merged by mistake.
2236 */
2237 if (bfq_too_late_for_merging(bfqq))
2238 return NULL;
2239
2240 if (bfqq->new_bfqq)
2241 return bfqq->new_bfqq;
2242
2243 if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2244 return NULL;
2245
2246 /* If there is only one backlogged queue, don't search. */
2247 if (bfq_tot_busy_queues(bfqd) == 1)
2248 return NULL;
2249
2250 in_service_bfqq = bfqd->in_service_queue;
2251
2252 if (in_service_bfqq && in_service_bfqq != bfqq &&
2253 likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2254 bfq_rq_close_to_sector(io_struct, request,
2255 bfqd->in_serv_last_pos) &&
2256 bfqq->entity.parent == in_service_bfqq->entity.parent &&
2257 bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2258 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2259 if (new_bfqq)
2260 return new_bfqq;
2261 }
2262 /*
2263 * Check whether there is a cooperator among currently scheduled
2264 * queues. The only thing we need is that the bio/request is not
2265 * NULL, as we need it to establish whether a cooperator exists.
2266 */
2267 new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2268 bfq_io_struct_pos(io_struct, request));
2269
2270 if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2271 bfq_may_be_close_cooperator(bfqq, new_bfqq))
2272 return bfq_setup_merge(bfqq, new_bfqq);
2273
2274 return NULL;
2275}
2276
2277static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2278{
2279 struct bfq_io_cq *bic = bfqq->bic;
2280
2281 /*
2282 * If !bfqq->bic, the queue is already shared or its requests
2283 * have already been redirected to a shared queue; both idle window
2284 * and weight raising state have already been saved. Do nothing.
2285 */
2286 if (!bic)
2287 return;
2288
2289 bic->saved_ttime = bfqq->ttime;
2290 bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2291 bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2292 bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2293 bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2294 if (unlikely(bfq_bfqq_just_created(bfqq) &&
2295 !bfq_bfqq_in_large_burst(bfqq) &&
2296 bfqq->bfqd->low_latency)) {
2297 /*
2298 * bfqq being merged right after being created: bfqq
2299 * would have deserved interactive weight raising, but
2300 * did not make it to be set in a weight-raised state,
2301 * because of this early merge. Store directly the
2302 * weight-raising state that would have been assigned
2303 * to bfqq, so that to avoid that bfqq unjustly fails
2304 * to enjoy weight raising if split soon.
2305 */
2306 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
2307 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2308 bic->saved_last_wr_start_finish = jiffies;
2309 } else {
2310 bic->saved_wr_coeff = bfqq->wr_coeff;
2311 bic->saved_wr_start_at_switch_to_srt =
2312 bfqq->wr_start_at_switch_to_srt;
2313 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2314 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2315 }
2316}
2317
2318static void
2319bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2320 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2321{
2322 bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2323 (unsigned long)new_bfqq->pid);
2324 /* Save weight raising and idle window of the merged queues */
2325 bfq_bfqq_save_state(bfqq);
2326 bfq_bfqq_save_state(new_bfqq);
2327 if (bfq_bfqq_IO_bound(bfqq))
2328 bfq_mark_bfqq_IO_bound(new_bfqq);
2329 bfq_clear_bfqq_IO_bound(bfqq);
2330
2331 /*
2332 * If bfqq is weight-raised, then let new_bfqq inherit
2333 * weight-raising. To reduce false positives, neglect the case
2334 * where bfqq has just been created, but has not yet made it
2335 * to be weight-raised (which may happen because EQM may merge
2336 * bfqq even before bfq_add_request is executed for the first
2337 * time for bfqq). Handling this case would however be very
2338 * easy, thanks to the flag just_created.
2339 */
2340 if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2341 new_bfqq->wr_coeff = bfqq->wr_coeff;
2342 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2343 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2344 new_bfqq->wr_start_at_switch_to_srt =
2345 bfqq->wr_start_at_switch_to_srt;
2346 if (bfq_bfqq_busy(new_bfqq))
2347 bfqd->wr_busy_queues++;
2348 new_bfqq->entity.prio_changed = 1;
2349 }
2350
2351 if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2352 bfqq->wr_coeff = 1;
2353 bfqq->entity.prio_changed = 1;
2354 if (bfq_bfqq_busy(bfqq))
2355 bfqd->wr_busy_queues--;
2356 }
2357
2358 bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2359 bfqd->wr_busy_queues);
2360
2361 /*
2362 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2363 */
2364 bic_set_bfqq(bic, new_bfqq, 1);
2365 bfq_mark_bfqq_coop(new_bfqq);
2366 /*
2367 * new_bfqq now belongs to at least two bics (it is a shared queue):
2368 * set new_bfqq->bic to NULL. bfqq either:
2369 * - does not belong to any bic any more, and hence bfqq->bic must
2370 * be set to NULL, or
2371 * - is a queue whose owning bics have already been redirected to a
2372 * different queue, hence the queue is destined to not belong to
2373 * any bic soon and bfqq->bic is already NULL (therefore the next
2374 * assignment causes no harm).
2375 */
2376 new_bfqq->bic = NULL;
2377 bfqq->bic = NULL;
2378 /* release process reference to bfqq */
2379 bfq_put_queue(bfqq);
2380}
2381
2382static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2383 struct bio *bio)
2384{
2385 struct bfq_data *bfqd = q->elevator->elevator_data;
2386 bool is_sync = op_is_sync(bio->bi_opf);
2387 struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2388
2389 /*
2390 * Disallow merge of a sync bio into an async request.
2391 */
2392 if (is_sync && !rq_is_sync(rq))
2393 return false;
2394
2395 /*
2396 * Lookup the bfqq that this bio will be queued with. Allow
2397 * merge only if rq is queued there.
2398 */
2399 if (!bfqq)
2400 return false;
2401
2402 /*
2403 * We take advantage of this function to perform an early merge
2404 * of the queues of possible cooperating processes.
2405 */
2406 new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2407 if (new_bfqq) {
2408 /*
2409 * bic still points to bfqq, then it has not yet been
2410 * redirected to some other bfq_queue, and a queue
2411 * merge beween bfqq and new_bfqq can be safely
2412 * fulfillled, i.e., bic can be redirected to new_bfqq
2413 * and bfqq can be put.
2414 */
2415 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2416 new_bfqq);
2417 /*
2418 * If we get here, bio will be queued into new_queue,
2419 * so use new_bfqq to decide whether bio and rq can be
2420 * merged.
2421 */
2422 bfqq = new_bfqq;
2423
2424 /*
2425 * Change also bqfd->bio_bfqq, as
2426 * bfqd->bio_bic now points to new_bfqq, and
2427 * this function may be invoked again (and then may
2428 * use again bqfd->bio_bfqq).
2429 */
2430 bfqd->bio_bfqq = bfqq;
2431 }
2432
2433 return bfqq == RQ_BFQQ(rq);
2434}
2435
2436/*
2437 * Set the maximum time for the in-service queue to consume its
2438 * budget. This prevents seeky processes from lowering the throughput.
2439 * In practice, a time-slice service scheme is used with seeky
2440 * processes.
2441 */
2442static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2443 struct bfq_queue *bfqq)
2444{
2445 unsigned int timeout_coeff;
2446
2447 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2448 timeout_coeff = 1;
2449 else
2450 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2451
2452 bfqd->last_budget_start = ktime_get();
2453
2454 bfqq->budget_timeout = jiffies +
2455 bfqd->bfq_timeout * timeout_coeff;
2456}
2457
2458static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2459 struct bfq_queue *bfqq)
2460{
2461 if (bfqq) {
2462 bfq_clear_bfqq_fifo_expire(bfqq);
2463
2464 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2465
2466 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2467 bfqq->wr_coeff > 1 &&
2468 bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2469 time_is_before_jiffies(bfqq->budget_timeout)) {
2470 /*
2471 * For soft real-time queues, move the start
2472 * of the weight-raising period forward by the
2473 * time the queue has not received any
2474 * service. Otherwise, a relatively long
2475 * service delay is likely to cause the
2476 * weight-raising period of the queue to end,
2477 * because of the short duration of the
2478 * weight-raising period of a soft real-time
2479 * queue. It is worth noting that this move
2480 * is not so dangerous for the other queues,
2481 * because soft real-time queues are not
2482 * greedy.
2483 *
2484 * To not add a further variable, we use the
2485 * overloaded field budget_timeout to
2486 * determine for how long the queue has not
2487 * received service, i.e., how much time has
2488 * elapsed since the queue expired. However,
2489 * this is a little imprecise, because
2490 * budget_timeout is set to jiffies if bfqq
2491 * not only expires, but also remains with no
2492 * request.
2493 */
2494 if (time_after(bfqq->budget_timeout,
2495 bfqq->last_wr_start_finish))
2496 bfqq->last_wr_start_finish +=
2497 jiffies - bfqq->budget_timeout;
2498 else
2499 bfqq->last_wr_start_finish = jiffies;
2500 }
2501
2502 bfq_set_budget_timeout(bfqd, bfqq);
2503 bfq_log_bfqq(bfqd, bfqq,
2504 "set_in_service_queue, cur-budget = %d",
2505 bfqq->entity.budget);
2506 }
2507
2508 bfqd->in_service_queue = bfqq;
2509}
2510
2511/*
2512 * Get and set a new queue for service.
2513 */
2514static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2515{
2516 struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2517
2518 __bfq_set_in_service_queue(bfqd, bfqq);
2519 return bfqq;
2520}
2521
2522static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2523{
2524 struct bfq_queue *bfqq = bfqd->in_service_queue;
2525 u32 sl;
2526
2527 bfq_mark_bfqq_wait_request(bfqq);
2528
2529 /*
2530 * We don't want to idle for seeks, but we do want to allow
2531 * fair distribution of slice time for a process doing back-to-back
2532 * seeks. So allow a little bit of time for him to submit a new rq.
2533 */
2534 sl = bfqd->bfq_slice_idle;
2535 /*
2536 * Unless the queue is being weight-raised or the scenario is
2537 * asymmetric, grant only minimum idle time if the queue
2538 * is seeky. A long idling is preserved for a weight-raised
2539 * queue, or, more in general, in an asymmetric scenario,
2540 * because a long idling is needed for guaranteeing to a queue
2541 * its reserved share of the throughput (in particular, it is
2542 * needed if the queue has a higher weight than some other
2543 * queue).
2544 */
2545 if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
2546 bfq_symmetric_scenario(bfqd))
2547 sl = min_t(u64, sl, BFQ_MIN_TT);
2548
2549 bfqd->last_idling_start = ktime_get();
2550 hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
2551 HRTIMER_MODE_REL);
2552 bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
2553}
2554
2555/*
2556 * In autotuning mode, max_budget is dynamically recomputed as the
2557 * amount of sectors transferred in timeout at the estimated peak
2558 * rate. This enables BFQ to utilize a full timeslice with a full
2559 * budget, even if the in-service queue is served at peak rate. And
2560 * this maximises throughput with sequential workloads.
2561 */
2562static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
2563{
2564 return (u64)bfqd->peak_rate * USEC_PER_MSEC *
2565 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
2566}
2567
2568/*
2569 * Update parameters related to throughput and responsiveness, as a
2570 * function of the estimated peak rate. See comments on
2571 * bfq_calc_max_budget(), and on the ref_wr_duration array.
2572 */
2573static void update_thr_responsiveness_params(struct bfq_data *bfqd)
2574{
2575 if (bfqd->bfq_user_max_budget == 0) {
2576 bfqd->bfq_max_budget =
2577 bfq_calc_max_budget(bfqd);
2578 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
2579 }
2580}
2581
2582static void bfq_reset_rate_computation(struct bfq_data *bfqd,
2583 struct request *rq)
2584{
2585 if (rq != NULL) { /* new rq dispatch now, reset accordingly */
2586 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
2587 bfqd->peak_rate_samples = 1;
2588 bfqd->sequential_samples = 0;
2589 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
2590 blk_rq_sectors(rq);
2591 } else /* no new rq dispatched, just reset the number of samples */
2592 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
2593
2594 bfq_log(bfqd,
2595 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
2596 bfqd->peak_rate_samples, bfqd->sequential_samples,
2597 bfqd->tot_sectors_dispatched);
2598}
2599
2600static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
2601{
2602 u32 rate, weight, divisor;
2603
2604 /*
2605 * For the convergence property to hold (see comments on
2606 * bfq_update_peak_rate()) and for the assessment to be
2607 * reliable, a minimum number of samples must be present, and
2608 * a minimum amount of time must have elapsed. If not so, do
2609 * not compute new rate. Just reset parameters, to get ready
2610 * for a new evaluation attempt.
2611 */
2612 if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
2613 bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
2614 goto reset_computation;
2615
2616 /*
2617 * If a new request completion has occurred after last
2618 * dispatch, then, to approximate the rate at which requests
2619 * have been served by the device, it is more precise to
2620 * extend the observation interval to the last completion.
2621 */
2622 bfqd->delta_from_first =
2623 max_t(u64, bfqd->delta_from_first,
2624 bfqd->last_completion - bfqd->first_dispatch);
2625
2626 /*
2627 * Rate computed in sects/usec, and not sects/nsec, for
2628 * precision issues.
2629 */
2630 rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
2631 div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
2632
2633 /*
2634 * Peak rate not updated if:
2635 * - the percentage of sequential dispatches is below 3/4 of the
2636 * total, and rate is below the current estimated peak rate
2637 * - rate is unreasonably high (> 20M sectors/sec)
2638 */
2639 if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
2640 rate <= bfqd->peak_rate) ||
2641 rate > 20<<BFQ_RATE_SHIFT)
2642 goto reset_computation;
2643
2644 /*
2645 * We have to update the peak rate, at last! To this purpose,
2646 * we use a low-pass filter. We compute the smoothing constant
2647 * of the filter as a function of the 'weight' of the new
2648 * measured rate.
2649 *
2650 * As can be seen in next formulas, we define this weight as a
2651 * quantity proportional to how sequential the workload is,
2652 * and to how long the observation time interval is.
2653 *
2654 * The weight runs from 0 to 8. The maximum value of the
2655 * weight, 8, yields the minimum value for the smoothing
2656 * constant. At this minimum value for the smoothing constant,
2657 * the measured rate contributes for half of the next value of
2658 * the estimated peak rate.
2659 *
2660 * So, the first step is to compute the weight as a function
2661 * of how sequential the workload is. Note that the weight
2662 * cannot reach 9, because bfqd->sequential_samples cannot
2663 * become equal to bfqd->peak_rate_samples, which, in its
2664 * turn, holds true because bfqd->sequential_samples is not
2665 * incremented for the first sample.
2666 */
2667 weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
2668
2669 /*
2670 * Second step: further refine the weight as a function of the
2671 * duration of the observation interval.
2672 */
2673 weight = min_t(u32, 8,
2674 div_u64(weight * bfqd->delta_from_first,
2675 BFQ_RATE_REF_INTERVAL));
2676
2677 /*
2678 * Divisor ranging from 10, for minimum weight, to 2, for
2679 * maximum weight.
2680 */
2681 divisor = 10 - weight;
2682
2683 /*
2684 * Finally, update peak rate:
2685 *
2686 * peak_rate = peak_rate * (divisor-1) / divisor + rate / divisor
2687 */
2688 bfqd->peak_rate *= divisor-1;
2689 bfqd->peak_rate /= divisor;
2690 rate /= divisor; /* smoothing constant alpha = 1/divisor */
2691
2692 bfqd->peak_rate += rate;
2693
2694 /*
2695 * For a very slow device, bfqd->peak_rate can reach 0 (see
2696 * the minimum representable values reported in the comments
2697 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
2698 * divisions by zero where bfqd->peak_rate is used as a
2699 * divisor.
2700 */
2701 bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
2702
2703 update_thr_responsiveness_params(bfqd);
2704
2705reset_computation:
2706 bfq_reset_rate_computation(bfqd, rq);
2707}
2708
2709/*
2710 * Update the read/write peak rate (the main quantity used for
2711 * auto-tuning, see update_thr_responsiveness_params()).
2712 *
2713 * It is not trivial to estimate the peak rate (correctly): because of
2714 * the presence of sw and hw queues between the scheduler and the
2715 * device components that finally serve I/O requests, it is hard to
2716 * say exactly when a given dispatched request is served inside the
2717 * device, and for how long. As a consequence, it is hard to know
2718 * precisely at what rate a given set of requests is actually served
2719 * by the device.
2720 *
2721 * On the opposite end, the dispatch time of any request is trivially
2722 * available, and, from this piece of information, the "dispatch rate"
2723 * of requests can be immediately computed. So, the idea in the next
2724 * function is to use what is known, namely request dispatch times
2725 * (plus, when useful, request completion times), to estimate what is
2726 * unknown, namely in-device request service rate.
2727 *
2728 * The main issue is that, because of the above facts, the rate at
2729 * which a certain set of requests is dispatched over a certain time
2730 * interval can vary greatly with respect to the rate at which the
2731 * same requests are then served. But, since the size of any
2732 * intermediate queue is limited, and the service scheme is lossless
2733 * (no request is silently dropped), the following obvious convergence
2734 * property holds: the number of requests dispatched MUST become
2735 * closer and closer to the number of requests completed as the
2736 * observation interval grows. This is the key property used in
2737 * the next function to estimate the peak service rate as a function
2738 * of the observed dispatch rate. The function assumes to be invoked
2739 * on every request dispatch.
2740 */
2741static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
2742{
2743 u64 now_ns = ktime_get_ns();
2744
2745 if (bfqd->peak_rate_samples == 0) { /* first dispatch */
2746 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
2747 bfqd->peak_rate_samples);
2748 bfq_reset_rate_computation(bfqd, rq);
2749 goto update_last_values; /* will add one sample */
2750 }
2751
2752 /*
2753 * Device idle for very long: the observation interval lasting
2754 * up to this dispatch cannot be a valid observation interval
2755 * for computing a new peak rate (similarly to the late-
2756 * completion event in bfq_completed_request()). Go to
2757 * update_rate_and_reset to have the following three steps
2758 * taken:
2759 * - close the observation interval at the last (previous)
2760 * request dispatch or completion
2761 * - compute rate, if possible, for that observation interval
2762 * - start a new observation interval with this dispatch
2763 */
2764 if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
2765 bfqd->rq_in_driver == 0)
2766 goto update_rate_and_reset;
2767
2768 /* Update sampling information */
2769 bfqd->peak_rate_samples++;
2770
2771 if ((bfqd->rq_in_driver > 0 ||
2772 now_ns - bfqd->last_completion < BFQ_MIN_TT)
2773 && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
2774 bfqd->sequential_samples++;
2775
2776 bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
2777
2778 /* Reset max observed rq size every 32 dispatches */
2779 if (likely(bfqd->peak_rate_samples % 32))
2780 bfqd->last_rq_max_size =
2781 max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
2782 else
2783 bfqd->last_rq_max_size = blk_rq_sectors(rq);
2784
2785 bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
2786
2787 /* Target observation interval not yet reached, go on sampling */
2788 if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
2789 goto update_last_values;
2790
2791update_rate_and_reset:
2792 bfq_update_rate_reset(bfqd, rq);
2793update_last_values:
2794 bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2795 if (RQ_BFQQ(rq) == bfqd->in_service_queue)
2796 bfqd->in_serv_last_pos = bfqd->last_position;
2797 bfqd->last_dispatch = now_ns;
2798}
2799
2800/*
2801 * Remove request from internal lists.
2802 */
2803static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
2804{
2805 struct bfq_queue *bfqq = RQ_BFQQ(rq);
2806
2807 /*
2808 * For consistency, the next instruction should have been
2809 * executed after removing the request from the queue and
2810 * dispatching it. We execute instead this instruction before
2811 * bfq_remove_request() (and hence introduce a temporary
2812 * inconsistency), for efficiency. In fact, should this
2813 * dispatch occur for a non in-service bfqq, this anticipated
2814 * increment prevents two counters related to bfqq->dispatched
2815 * from risking to be, first, uselessly decremented, and then
2816 * incremented again when the (new) value of bfqq->dispatched
2817 * happens to be taken into account.
2818 */
2819 bfqq->dispatched++;
2820 bfq_update_peak_rate(q->elevator->elevator_data, rq);
2821
2822 bfq_remove_request(q, rq);
2823}
2824
2825static void __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq)
2826{
2827 /*
2828 * If this bfqq is shared between multiple processes, check
2829 * to make sure that those processes are still issuing I/Os
2830 * within the mean seek distance. If not, it may be time to
2831 * break the queues apart again.
2832 */
2833 if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
2834 bfq_mark_bfqq_split_coop(bfqq);
2835
2836 if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2837 if (bfqq->dispatched == 0)
2838 /*
2839 * Overloading budget_timeout field to store
2840 * the time at which the queue remains with no
2841 * backlog and no outstanding request; used by
2842 * the weight-raising mechanism.
2843 */
2844 bfqq->budget_timeout = jiffies;
2845
2846 bfq_del_bfqq_busy(bfqd, bfqq, true);
2847 } else {
2848 bfq_requeue_bfqq(bfqd, bfqq, true);
2849 /*
2850 * Resort priority tree of potential close cooperators.
2851 */
2852 bfq_pos_tree_add_move(bfqd, bfqq);
2853 }
2854
2855 /*
2856 * All in-service entities must have been properly deactivated
2857 * or requeued before executing the next function, which
2858 * resets all in-service entites as no more in service.
2859 */
2860 __bfq_bfqd_reset_in_service(bfqd);
2861}
2862
2863/**
2864 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
2865 * @bfqd: device data.
2866 * @bfqq: queue to update.
2867 * @reason: reason for expiration.
2868 *
2869 * Handle the feedback on @bfqq budget at queue expiration.
2870 * See the body for detailed comments.
2871 */
2872static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
2873 struct bfq_queue *bfqq,
2874 enum bfqq_expiration reason)
2875{
2876 struct request *next_rq;
2877 int budget, min_budget;
2878
2879 min_budget = bfq_min_budget(bfqd);
2880
2881 if (bfqq->wr_coeff == 1)
2882 budget = bfqq->max_budget;
2883 else /*
2884 * Use a constant, low budget for weight-raised queues,
2885 * to help achieve a low latency. Keep it slightly higher
2886 * than the minimum possible budget, to cause a little
2887 * bit fewer expirations.
2888 */
2889 budget = 2 * min_budget;
2890
2891 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
2892 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
2893 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
2894 budget, bfq_min_budget(bfqd));
2895 bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
2896 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
2897
2898 if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
2899 switch (reason) {
2900 /*
2901 * Caveat: in all the following cases we trade latency
2902 * for throughput.
2903 */
2904 case BFQQE_TOO_IDLE:
2905 /*
2906 * This is the only case where we may reduce
2907 * the budget: if there is no request of the
2908 * process still waiting for completion, then
2909 * we assume (tentatively) that the timer has
2910 * expired because the batch of requests of
2911 * the process could have been served with a
2912 * smaller budget. Hence, betting that
2913 * process will behave in the same way when it
2914 * becomes backlogged again, we reduce its
2915 * next budget. As long as we guess right,
2916 * this budget cut reduces the latency
2917 * experienced by the process.
2918 *
2919 * However, if there are still outstanding
2920 * requests, then the process may have not yet
2921 * issued its next request just because it is
2922 * still waiting for the completion of some of
2923 * the still outstanding ones. So in this
2924 * subcase we do not reduce its budget, on the
2925 * contrary we increase it to possibly boost
2926 * the throughput, as discussed in the
2927 * comments to the BUDGET_TIMEOUT case.
2928 */
2929 if (bfqq->dispatched > 0) /* still outstanding reqs */
2930 budget = min(budget * 2, bfqd->bfq_max_budget);
2931 else {
2932 if (budget > 5 * min_budget)
2933 budget -= 4 * min_budget;
2934 else
2935 budget = min_budget;
2936 }
2937 break;
2938 case BFQQE_BUDGET_TIMEOUT:
2939 /*
2940 * We double the budget here because it gives
2941 * the chance to boost the throughput if this
2942 * is not a seeky process (and has bumped into
2943 * this timeout because of, e.g., ZBR).
2944 */
2945 budget = min(budget * 2, bfqd->bfq_max_budget);
2946 break;
2947 case BFQQE_BUDGET_EXHAUSTED:
2948 /*
2949 * The process still has backlog, and did not
2950 * let either the budget timeout or the disk
2951 * idling timeout expire. Hence it is not
2952 * seeky, has a short thinktime and may be
2953 * happy with a higher budget too. So
2954 * definitely increase the budget of this good
2955 * candidate to boost the disk throughput.
2956 */
2957 budget = min(budget * 4, bfqd->bfq_max_budget);
2958 break;
2959 case BFQQE_NO_MORE_REQUESTS:
2960 /*
2961 * For queues that expire for this reason, it
2962 * is particularly important to keep the
2963 * budget close to the actual service they
2964 * need. Doing so reduces the timestamp
2965 * misalignment problem described in the
2966 * comments in the body of
2967 * __bfq_activate_entity. In fact, suppose
2968 * that a queue systematically expires for
2969 * BFQQE_NO_MORE_REQUESTS and presents a
2970 * new request in time to enjoy timestamp
2971 * back-shifting. The larger the budget of the
2972 * queue is with respect to the service the
2973 * queue actually requests in each service
2974 * slot, the more times the queue can be
2975 * reactivated with the same virtual finish
2976 * time. It follows that, even if this finish
2977 * time is pushed to the system virtual time
2978 * to reduce the consequent timestamp
2979 * misalignment, the queue unjustly enjoys for
2980 * many re-activations a lower finish time
2981 * than all newly activated queues.
2982 *
2983 * The service needed by bfqq is measured
2984 * quite precisely by bfqq->entity.service.
2985 * Since bfqq does not enjoy device idling,
2986 * bfqq->entity.service is equal to the number
2987 * of sectors that the process associated with
2988 * bfqq requested to read/write before waiting
2989 * for request completions, or blocking for
2990 * other reasons.
2991 */
2992 budget = max_t(int, bfqq->entity.service, min_budget);
2993 break;
2994 default:
2995 return;
2996 }
2997 } else if (!bfq_bfqq_sync(bfqq)) {
2998 /*
2999 * Async queues get always the maximum possible
3000 * budget, as for them we do not care about latency
3001 * (in addition, their ability to dispatch is limited
3002 * by the charging factor).
3003 */
3004 budget = bfqd->bfq_max_budget;
3005 }
3006
3007 bfqq->max_budget = budget;
3008
3009 if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
3010 !bfqd->bfq_user_max_budget)
3011 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
3012
3013 /*
3014 * If there is still backlog, then assign a new budget, making
3015 * sure that it is large enough for the next request. Since
3016 * the finish time of bfqq must be kept in sync with the
3017 * budget, be sure to call __bfq_bfqq_expire() *after* this
3018 * update.
3019 *
3020 * If there is no backlog, then no need to update the budget;
3021 * it will be updated on the arrival of a new request.
3022 */
3023 next_rq = bfqq->next_rq;
3024 if (next_rq)
3025 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
3026 bfq_serv_to_charge(next_rq, bfqq));
3027
3028 bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
3029 next_rq ? blk_rq_sectors(next_rq) : 0,
3030 bfqq->entity.budget);
3031}
3032
3033/*
3034 * Return true if the process associated with bfqq is "slow". The slow
3035 * flag is used, in addition to the budget timeout, to reduce the
3036 * amount of service provided to seeky processes, and thus reduce
3037 * their chances to lower the throughput. More details in the comments
3038 * on the function bfq_bfqq_expire().
3039 *
3040 * An important observation is in order: as discussed in the comments
3041 * on the function bfq_update_peak_rate(), with devices with internal
3042 * queues, it is hard if ever possible to know when and for how long
3043 * an I/O request is processed by the device (apart from the trivial
3044 * I/O pattern where a new request is dispatched only after the
3045 * previous one has been completed). This makes it hard to evaluate
3046 * the real rate at which the I/O requests of each bfq_queue are
3047 * served. In fact, for an I/O scheduler like BFQ, serving a
3048 * bfq_queue means just dispatching its requests during its service
3049 * slot (i.e., until the budget of the queue is exhausted, or the
3050 * queue remains idle, or, finally, a timeout fires). But, during the
3051 * service slot of a bfq_queue, around 100 ms at most, the device may
3052 * be even still processing requests of bfq_queues served in previous
3053 * service slots. On the opposite end, the requests of the in-service
3054 * bfq_queue may be completed after the service slot of the queue
3055 * finishes.
3056 *
3057 * Anyway, unless more sophisticated solutions are used
3058 * (where possible), the sum of the sizes of the requests dispatched
3059 * during the service slot of a bfq_queue is probably the only
3060 * approximation available for the service received by the bfq_queue
3061 * during its service slot. And this sum is the quantity used in this
3062 * function to evaluate the I/O speed of a process.
3063 */
3064static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3065 bool compensate, enum bfqq_expiration reason,
3066 unsigned long *delta_ms)
3067{
3068 ktime_t delta_ktime;
3069 u32 delta_usecs;
3070 bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
3071
3072 if (!bfq_bfqq_sync(bfqq))
3073 return false;
3074
3075 if (compensate)
3076 delta_ktime = bfqd->last_idling_start;
3077 else
3078 delta_ktime = ktime_get();
3079 delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
3080 delta_usecs = ktime_to_us(delta_ktime);
3081
3082 /* don't use too short time intervals */
3083 if (delta_usecs < 1000) {
3084 if (blk_queue_nonrot(bfqd->queue))
3085 /*
3086 * give same worst-case guarantees as idling
3087 * for seeky
3088 */
3089 *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
3090 else /* charge at least one seek */
3091 *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
3092
3093 return slow;
3094 }
3095
3096 *delta_ms = delta_usecs / USEC_PER_MSEC;
3097
3098 /*
3099 * Use only long (> 20ms) intervals to filter out excessive
3100 * spikes in service rate estimation.
3101 */
3102 if (delta_usecs > 20000) {
3103 /*
3104 * Caveat for rotational devices: processes doing I/O
3105 * in the slower disk zones tend to be slow(er) even
3106 * if not seeky. In this respect, the estimated peak
3107 * rate is likely to be an average over the disk
3108 * surface. Accordingly, to not be too harsh with
3109 * unlucky processes, a process is deemed slow only if
3110 * its rate has been lower than half of the estimated
3111 * peak rate.
3112 */
3113 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
3114 }
3115
3116 bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
3117
3118 return slow;
3119}
3120
3121/*
3122 * To be deemed as soft real-time, an application must meet two
3123 * requirements. First, the application must not require an average
3124 * bandwidth higher than the approximate bandwidth required to playback or
3125 * record a compressed high-definition video.
3126 * The next function is invoked on the completion of the last request of a
3127 * batch, to compute the next-start time instant, soft_rt_next_start, such
3128 * that, if the next request of the application does not arrive before
3129 * soft_rt_next_start, then the above requirement on the bandwidth is met.
3130 *
3131 * The second requirement is that the request pattern of the application is
3132 * isochronous, i.e., that, after issuing a request or a batch of requests,
3133 * the application stops issuing new requests until all its pending requests
3134 * have been completed. After that, the application may issue a new batch,
3135 * and so on.
3136 * For this reason the next function is invoked to compute
3137 * soft_rt_next_start only for applications that meet this requirement,
3138 * whereas soft_rt_next_start is set to infinity for applications that do
3139 * not.
3140 *
3141 * Unfortunately, even a greedy (i.e., I/O-bound) application may
3142 * happen to meet, occasionally or systematically, both the above
3143 * bandwidth and isochrony requirements. This may happen at least in
3144 * the following circumstances. First, if the CPU load is high. The
3145 * application may stop issuing requests while the CPUs are busy
3146 * serving other processes, then restart, then stop again for a while,
3147 * and so on. The other circumstances are related to the storage
3148 * device: the storage device is highly loaded or reaches a low-enough
3149 * throughput with the I/O of the application (e.g., because the I/O
3150 * is random and/or the device is slow). In all these cases, the
3151 * I/O of the application may be simply slowed down enough to meet
3152 * the bandwidth and isochrony requirements. To reduce the probability
3153 * that greedy applications are deemed as soft real-time in these
3154 * corner cases, a further rule is used in the computation of
3155 * soft_rt_next_start: the return value of this function is forced to
3156 * be higher than the maximum between the following two quantities.
3157 *
3158 * (a) Current time plus: (1) the maximum time for which the arrival
3159 * of a request is waited for when a sync queue becomes idle,
3160 * namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
3161 * postpone for a moment the reason for adding a few extra
3162 * jiffies; we get back to it after next item (b). Lower-bounding
3163 * the return value of this function with the current time plus
3164 * bfqd->bfq_slice_idle tends to filter out greedy applications,
3165 * because the latter issue their next request as soon as possible
3166 * after the last one has been completed. In contrast, a soft
3167 * real-time application spends some time processing data, after a
3168 * batch of its requests has been completed.
3169 *
3170 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
3171 * above, greedy applications may happen to meet both the
3172 * bandwidth and isochrony requirements under heavy CPU or
3173 * storage-device load. In more detail, in these scenarios, these
3174 * applications happen, only for limited time periods, to do I/O
3175 * slowly enough to meet all the requirements described so far,
3176 * including the filtering in above item (a). These slow-speed
3177 * time intervals are usually interspersed between other time
3178 * intervals during which these applications do I/O at a very high
3179 * speed. Fortunately, exactly because of the high speed of the
3180 * I/O in the high-speed intervals, the values returned by this
3181 * function happen to be so high, near the end of any such
3182 * high-speed interval, to be likely to fall *after* the end of
3183 * the low-speed time interval that follows. These high values are
3184 * stored in bfqq->soft_rt_next_start after each invocation of
3185 * this function. As a consequence, if the last value of
3186 * bfqq->soft_rt_next_start is constantly used to lower-bound the
3187 * next value that this function may return, then, from the very
3188 * beginning of a low-speed interval, bfqq->soft_rt_next_start is
3189 * likely to be constantly kept so high that any I/O request
3190 * issued during the low-speed interval is considered as arriving
3191 * to soon for the application to be deemed as soft
3192 * real-time. Then, in the high-speed interval that follows, the
3193 * application will not be deemed as soft real-time, just because
3194 * it will do I/O at a high speed. And so on.
3195 *
3196 * Getting back to the filtering in item (a), in the following two
3197 * cases this filtering might be easily passed by a greedy
3198 * application, if the reference quantity was just
3199 * bfqd->bfq_slice_idle:
3200 * 1) HZ is so low that the duration of a jiffy is comparable to or
3201 * higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
3202 * devices with HZ=100. The time granularity may be so coarse
3203 * that the approximation, in jiffies, of bfqd->bfq_slice_idle
3204 * is rather lower than the exact value.
3205 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
3206 * for a while, then suddenly 'jump' by several units to recover the lost
3207 * increments. This seems to happen, e.g., inside virtual machines.
3208 * To address this issue, in the filtering in (a) we do not use as a
3209 * reference time interval just bfqd->bfq_slice_idle, but
3210 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
3211 * minimum number of jiffies for which the filter seems to be quite
3212 * precise also in embedded systems and KVM/QEMU virtual machines.
3213 */
3214static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
3215 struct bfq_queue *bfqq)
3216{
3217 return max3(bfqq->soft_rt_next_start,
3218 bfqq->last_idle_bklogged +
3219 HZ * bfqq->service_from_backlogged /
3220 bfqd->bfq_wr_max_softrt_rate,
3221 jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
3222}
3223
3224static bool bfq_bfqq_injectable(struct bfq_queue *bfqq)
3225{
3226 return BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3227 blk_queue_nonrot(bfqq->bfqd->queue) &&
3228 bfqq->bfqd->hw_tag;
3229}
3230
3231/**
3232 * bfq_bfqq_expire - expire a queue.
3233 * @bfqd: device owning the queue.
3234 * @bfqq: the queue to expire.
3235 * @compensate: if true, compensate for the time spent idling.
3236 * @reason: the reason causing the expiration.
3237 *
3238 * If the process associated with bfqq does slow I/O (e.g., because it
3239 * issues random requests), we charge bfqq with the time it has been
3240 * in service instead of the service it has received (see
3241 * bfq_bfqq_charge_time for details on how this goal is achieved). As
3242 * a consequence, bfqq will typically get higher timestamps upon
3243 * reactivation, and hence it will be rescheduled as if it had
3244 * received more service than what it has actually received. In the
3245 * end, bfqq receives less service in proportion to how slowly its
3246 * associated process consumes its budgets (and hence how seriously it
3247 * tends to lower the throughput). In addition, this time-charging
3248 * strategy guarantees time fairness among slow processes. In
3249 * contrast, if the process associated with bfqq is not slow, we
3250 * charge bfqq exactly with the service it has received.
3251 *
3252 * Charging time to the first type of queues and the exact service to
3253 * the other has the effect of using the WF2Q+ policy to schedule the
3254 * former on a timeslice basis, without violating service domain
3255 * guarantees among the latter.
3256 */
3257void bfq_bfqq_expire(struct bfq_data *bfqd,
3258 struct bfq_queue *bfqq,
3259 bool compensate,
3260 enum bfqq_expiration reason)
3261{
3262 bool slow;
3263 unsigned long delta = 0;
3264 struct bfq_entity *entity = &bfqq->entity;
3265 int ref;
3266
3267 /*
3268 * Check whether the process is slow (see bfq_bfqq_is_slow).
3269 */
3270 slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
3271
3272 /*
3273 * As above explained, charge slow (typically seeky) and
3274 * timed-out queues with the time and not the service
3275 * received, to favor sequential workloads.
3276 *
3277 * Processes doing I/O in the slower disk zones will tend to
3278 * be slow(er) even if not seeky. Therefore, since the
3279 * estimated peak rate is actually an average over the disk
3280 * surface, these processes may timeout just for bad luck. To
3281 * avoid punishing them, do not charge time to processes that
3282 * succeeded in consuming at least 2/3 of their budget. This
3283 * allows BFQ to preserve enough elasticity to still perform
3284 * bandwidth, and not time, distribution with little unlucky
3285 * or quasi-sequential processes.
3286 */
3287 if (bfqq->wr_coeff == 1 &&
3288 (slow ||
3289 (reason == BFQQE_BUDGET_TIMEOUT &&
3290 bfq_bfqq_budget_left(bfqq) >= entity->budget / 3)))
3291 bfq_bfqq_charge_time(bfqd, bfqq, delta);
3292
3293 if (reason == BFQQE_TOO_IDLE &&
3294 entity->service <= 2 * entity->budget / 10)
3295 bfq_clear_bfqq_IO_bound(bfqq);
3296
3297 if (bfqd->low_latency && bfqq->wr_coeff == 1)
3298 bfqq->last_wr_start_finish = jiffies;
3299
3300 if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
3301 RB_EMPTY_ROOT(&bfqq->sort_list)) {
3302 /*
3303 * If we get here, and there are no outstanding
3304 * requests, then the request pattern is isochronous
3305 * (see the comments on the function
3306 * bfq_bfqq_softrt_next_start()). Thus we can compute
3307 * soft_rt_next_start. And we do it, unless bfqq is in
3308 * interactive weight raising. We do not do it in the
3309 * latter subcase, for the following reason. bfqq may
3310 * be conveying the I/O needed to load a soft
3311 * real-time application. Such an application will
3312 * actually exhibit a soft real-time I/O pattern after
3313 * it finally starts doing its job. But, if
3314 * soft_rt_next_start is computed here for an
3315 * interactive bfqq, and bfqq had received a lot of
3316 * service before remaining with no outstanding
3317 * request (likely to happen on a fast device), then
3318 * soft_rt_next_start would be assigned such a high
3319 * value that, for a very long time, bfqq would be
3320 * prevented from being possibly considered as soft
3321 * real time.
3322 *
3323 * If, instead, the queue still has outstanding
3324 * requests, then we have to wait for the completion
3325 * of all the outstanding requests to discover whether
3326 * the request pattern is actually isochronous.
3327 */
3328 if (bfqq->dispatched == 0 &&
3329 bfqq->wr_coeff != bfqd->bfq_wr_coeff)
3330 bfqq->soft_rt_next_start =
3331 bfq_bfqq_softrt_next_start(bfqd, bfqq);
3332 else if (bfqq->dispatched > 0) {
3333 /*
3334 * Schedule an update of soft_rt_next_start to when
3335 * the task may be discovered to be isochronous.
3336 */
3337 bfq_mark_bfqq_softrt_update(bfqq);
3338 }
3339 }
3340
3341 bfq_log_bfqq(bfqd, bfqq,
3342 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
3343 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
3344
3345 /*
3346 * Increase, decrease or leave budget unchanged according to
3347 * reason.
3348 */
3349 __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
3350 ref = bfqq->ref;
3351 __bfq_bfqq_expire(bfqd, bfqq);
3352
3353 if (ref == 1) /* bfqq is gone, no more actions on it */
3354 return;
3355
3356 bfqq->injected_service = 0;
3357
3358 /* mark bfqq as waiting a request only if a bic still points to it */
3359 if (!bfq_bfqq_busy(bfqq) &&
3360 reason != BFQQE_BUDGET_TIMEOUT &&
3361 reason != BFQQE_BUDGET_EXHAUSTED) {
3362 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
3363 /*
3364 * Not setting service to 0, because, if the next rq
3365 * arrives in time, the queue will go on receiving
3366 * service with this same budget (as if it never expired)
3367 */
3368 } else
3369 entity->service = 0;
3370
3371 /*
3372 * Reset the received-service counter for every parent entity.
3373 * Differently from what happens with bfqq->entity.service,
3374 * the resetting of this counter never needs to be postponed
3375 * for parent entities. In fact, in case bfqq may have a
3376 * chance to go on being served using the last, partially
3377 * consumed budget, bfqq->entity.service needs to be kept,
3378 * because if bfqq then actually goes on being served using
3379 * the same budget, the last value of bfqq->entity.service is
3380 * needed to properly decrement bfqq->entity.budget by the
3381 * portion already consumed. In contrast, it is not necessary
3382 * to keep entity->service for parent entities too, because
3383 * the bubble up of the new value of bfqq->entity.budget will
3384 * make sure that the budgets of parent entities are correct,
3385 * even in case bfqq and thus parent entities go on receiving
3386 * service with the same budget.
3387 */
3388 entity = entity->parent;
3389 for_each_entity(entity)
3390 entity->service = 0;
3391}
3392
3393/*
3394 * Budget timeout is not implemented through a dedicated timer, but
3395 * just checked on request arrivals and completions, as well as on
3396 * idle timer expirations.
3397 */
3398static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
3399{
3400 return time_is_before_eq_jiffies(bfqq->budget_timeout);
3401}
3402
3403/*
3404 * If we expire a queue that is actively waiting (i.e., with the
3405 * device idled) for the arrival of a new request, then we may incur
3406 * the timestamp misalignment problem described in the body of the
3407 * function __bfq_activate_entity. Hence we return true only if this
3408 * condition does not hold, or if the queue is slow enough to deserve
3409 * only to be kicked off for preserving a high throughput.
3410 */
3411static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
3412{
3413 bfq_log_bfqq(bfqq->bfqd, bfqq,
3414 "may_budget_timeout: wait_request %d left %d timeout %d",
3415 bfq_bfqq_wait_request(bfqq),
3416 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3,
3417 bfq_bfqq_budget_timeout(bfqq));
3418
3419 return (!bfq_bfqq_wait_request(bfqq) ||
3420 bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)
3421 &&
3422 bfq_bfqq_budget_timeout(bfqq);
3423}
3424
3425static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
3426 struct bfq_queue *bfqq)
3427{
3428 bool rot_without_queueing =
3429 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
3430 bfqq_sequential_and_IO_bound,
3431 idling_boosts_thr;
3432
3433 bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
3434 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
3435
3436 /*
3437 * The next variable takes into account the cases where idling
3438 * boosts the throughput.
3439 *
3440 * The value of the variable is computed considering, first, that
3441 * idling is virtually always beneficial for the throughput if:
3442 * (a) the device is not NCQ-capable and rotational, or
3443 * (b) regardless of the presence of NCQ, the device is rotational and
3444 * the request pattern for bfqq is I/O-bound and sequential, or
3445 * (c) regardless of whether it is rotational, the device is
3446 * not NCQ-capable and the request pattern for bfqq is
3447 * I/O-bound and sequential.
3448 *
3449 * Secondly, and in contrast to the above item (b), idling an
3450 * NCQ-capable flash-based device would not boost the
3451 * throughput even with sequential I/O; rather it would lower
3452 * the throughput in proportion to how fast the device
3453 * is. Accordingly, the next variable is true if any of the
3454 * above conditions (a), (b) or (c) is true, and, in
3455 * particular, happens to be false if bfqd is an NCQ-capable
3456 * flash-based device.
3457 */
3458 idling_boosts_thr = rot_without_queueing ||
3459 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
3460 bfqq_sequential_and_IO_bound);
3461
3462 /*
3463 * The return value of this function is equal to that of
3464 * idling_boosts_thr, unless a special case holds. In this
3465 * special case, described below, idling may cause problems to
3466 * weight-raised queues.
3467 *
3468 * When the request pool is saturated (e.g., in the presence
3469 * of write hogs), if the processes associated with
3470 * non-weight-raised queues ask for requests at a lower rate,
3471 * then processes associated with weight-raised queues have a
3472 * higher probability to get a request from the pool
3473 * immediately (or at least soon) when they need one. Thus
3474 * they have a higher probability to actually get a fraction
3475 * of the device throughput proportional to their high
3476 * weight. This is especially true with NCQ-capable drives,
3477 * which enqueue several requests in advance, and further
3478 * reorder internally-queued requests.
3479 *
3480 * For this reason, we force to false the return value if
3481 * there are weight-raised busy queues. In this case, and if
3482 * bfqq is not weight-raised, this guarantees that the device
3483 * is not idled for bfqq (if, instead, bfqq is weight-raised,
3484 * then idling will be guaranteed by another variable, see
3485 * below). Combined with the timestamping rules of BFQ (see
3486 * [1] for details), this behavior causes bfqq, and hence any
3487 * sync non-weight-raised queue, to get a lower number of
3488 * requests served, and thus to ask for a lower number of
3489 * requests from the request pool, before the busy
3490 * weight-raised queues get served again. This often mitigates
3491 * starvation problems in the presence of heavy write
3492 * workloads and NCQ, thereby guaranteeing a higher
3493 * application and system responsiveness in these hostile
3494 * scenarios.
3495 */
3496 return idling_boosts_thr &&
3497 bfqd->wr_busy_queues == 0;
3498}
3499
3500/*
3501 * There is a case where idling must be performed not for
3502 * throughput concerns, but to preserve service guarantees.
3503 *
3504 * To introduce this case, we can note that allowing the drive
3505 * to enqueue more than one request at a time, and hence
3506 * delegating de facto final scheduling decisions to the
3507 * drive's internal scheduler, entails loss of control on the
3508 * actual request service order. In particular, the critical
3509 * situation is when requests from different processes happen
3510 * to be present, at the same time, in the internal queue(s)
3511 * of the drive. In such a situation, the drive, by deciding
3512 * the service order of the internally-queued requests, does
3513 * determine also the actual throughput distribution among
3514 * these processes. But the drive typically has no notion or
3515 * concern about per-process throughput distribution, and
3516 * makes its decisions only on a per-request basis. Therefore,
3517 * the service distribution enforced by the drive's internal
3518 * scheduler is likely to coincide with the desired
3519 * device-throughput distribution only in a completely
3520 * symmetric scenario where:
3521 * (i) each of these processes must get the same throughput as
3522 * the others;
3523 * (ii) the I/O of each process has the same properties, in
3524 * terms of locality (sequential or random), direction
3525 * (reads or writes), request sizes, greediness
3526 * (from I/O-bound to sporadic), and so on.
3527 * In fact, in such a scenario, the drive tends to treat
3528 * the requests of each of these processes in about the same
3529 * way as the requests of the others, and thus to provide
3530 * each of these processes with about the same throughput
3531 * (which is exactly the desired throughput distribution). In
3532 * contrast, in any asymmetric scenario, device idling is
3533 * certainly needed to guarantee that bfqq receives its
3534 * assigned fraction of the device throughput (see [1] for
3535 * details).
3536 * The problem is that idling may significantly reduce
3537 * throughput with certain combinations of types of I/O and
3538 * devices. An important example is sync random I/O, on flash
3539 * storage with command queueing. So, unless bfqq falls in the
3540 * above cases where idling also boosts throughput, it would
3541 * be important to check conditions (i) and (ii) accurately,
3542 * so as to avoid idling when not strictly needed for service
3543 * guarantees.
3544 *
3545 * Unfortunately, it is extremely difficult to thoroughly
3546 * check condition (ii). And, in case there are active groups,
3547 * it becomes very difficult to check condition (i) too. In
3548 * fact, if there are active groups, then, for condition (i)
3549 * to become false, it is enough that an active group contains
3550 * more active processes or sub-groups than some other active
3551 * group. More precisely, for condition (i) to hold because of
3552 * such a group, it is not even necessary that the group is
3553 * (still) active: it is sufficient that, even if the group
3554 * has become inactive, some of its descendant processes still
3555 * have some request already dispatched but still waiting for
3556 * completion. In fact, requests have still to be guaranteed
3557 * their share of the throughput even after being
3558 * dispatched. In this respect, it is easy to show that, if a
3559 * group frequently becomes inactive while still having
3560 * in-flight requests, and if, when this happens, the group is
3561 * not considered in the calculation of whether the scenario
3562 * is asymmetric, then the group may fail to be guaranteed its
3563 * fair share of the throughput (basically because idling may
3564 * not be performed for the descendant processes of the group,
3565 * but it had to be). We address this issue with the
3566 * following bi-modal behavior, implemented in the function
3567 * bfq_symmetric_scenario().
3568 *
3569 * If there are groups with requests waiting for completion
3570 * (as commented above, some of these groups may even be
3571 * already inactive), then the scenario is tagged as
3572 * asymmetric, conservatively, without checking any of the
3573 * conditions (i) and (ii). So the device is idled for bfqq.
3574 * This behavior matches also the fact that groups are created
3575 * exactly if controlling I/O is a primary concern (to
3576 * preserve bandwidth and latency guarantees).
3577 *
3578 * On the opposite end, if there are no groups with requests
3579 * waiting for completion, then only condition (i) is actually
3580 * controlled, i.e., provided that condition (i) holds, idling
3581 * is not performed, regardless of whether condition (ii)
3582 * holds. In other words, only if condition (i) does not hold,
3583 * then idling is allowed, and the device tends to be
3584 * prevented from queueing many requests, possibly of several
3585 * processes. Since there are no groups with requests waiting
3586 * for completion, then, to control condition (i) it is enough
3587 * to check just whether all the queues with requests waiting
3588 * for completion also have the same weight.
3589 *
3590 * Not checking condition (ii) evidently exposes bfqq to the
3591 * risk of getting less throughput than its fair share.
3592 * However, for queues with the same weight, a further
3593 * mechanism, preemption, mitigates or even eliminates this
3594 * problem. And it does so without consequences on overall
3595 * throughput. This mechanism and its benefits are explained
3596 * in the next three paragraphs.
3597 *
3598 * Even if a queue, say Q, is expired when it remains idle, Q
3599 * can still preempt the new in-service queue if the next
3600 * request of Q arrives soon (see the comments on
3601 * bfq_bfqq_update_budg_for_activation). If all queues and
3602 * groups have the same weight, this form of preemption,
3603 * combined with the hole-recovery heuristic described in the
3604 * comments on function bfq_bfqq_update_budg_for_activation,
3605 * are enough to preserve a correct bandwidth distribution in
3606 * the mid term, even without idling. In fact, even if not
3607 * idling allows the internal queues of the device to contain
3608 * many requests, and thus to reorder requests, we can rather
3609 * safely assume that the internal scheduler still preserves a
3610 * minimum of mid-term fairness.
3611 *
3612 * More precisely, this preemption-based, idleless approach
3613 * provides fairness in terms of IOPS, and not sectors per
3614 * second. This can be seen with a simple example. Suppose
3615 * that there are two queues with the same weight, but that
3616 * the first queue receives requests of 8 sectors, while the
3617 * second queue receives requests of 1024 sectors. In
3618 * addition, suppose that each of the two queues contains at
3619 * most one request at a time, which implies that each queue
3620 * always remains idle after it is served. Finally, after
3621 * remaining idle, each queue receives very quickly a new
3622 * request. It follows that the two queues are served
3623 * alternatively, preempting each other if needed. This
3624 * implies that, although both queues have the same weight,
3625 * the queue with large requests receives a service that is
3626 * 1024/8 times as high as the service received by the other
3627 * queue.
3628 *
3629 * The motivation for using preemption instead of idling (for
3630 * queues with the same weight) is that, by not idling,
3631 * service guarantees are preserved (completely or at least in
3632 * part) without minimally sacrificing throughput. And, if
3633 * there is no active group, then the primary expectation for
3634 * this device is probably a high throughput.
3635 *
3636 * We are now left only with explaining the additional
3637 * compound condition that is checked below for deciding
3638 * whether the scenario is asymmetric. To explain this
3639 * compound condition, we need to add that the function
3640 * bfq_symmetric_scenario checks the weights of only
3641 * non-weight-raised queues, for efficiency reasons (see
3642 * comments on bfq_weights_tree_add()). Then the fact that
3643 * bfqq is weight-raised is checked explicitly here. More
3644 * precisely, the compound condition below takes into account
3645 * also the fact that, even if bfqq is being weight-raised,
3646 * the scenario is still symmetric if all queues with requests
3647 * waiting for completion happen to be
3648 * weight-raised. Actually, we should be even more precise
3649 * here, and differentiate between interactive weight raising
3650 * and soft real-time weight raising.
3651 *
3652 * As a side note, it is worth considering that the above
3653 * device-idling countermeasures may however fail in the
3654 * following unlucky scenario: if idling is (correctly)
3655 * disabled in a time period during which all symmetry
3656 * sub-conditions hold, and hence the device is allowed to
3657 * enqueue many requests, but at some later point in time some
3658 * sub-condition stops to hold, then it may become impossible
3659 * to let requests be served in the desired order until all
3660 * the requests already queued in the device have been served.
3661 */
3662static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3663 struct bfq_queue *bfqq)
3664{
3665 return (bfqq->wr_coeff > 1 &&
3666 bfqd->wr_busy_queues <
3667 bfq_tot_busy_queues(bfqd)) ||
3668 !bfq_symmetric_scenario(bfqd);
3669}
3670
3671/*
3672 * For a queue that becomes empty, device idling is allowed only if
3673 * this function returns true for that queue. As a consequence, since
3674 * device idling plays a critical role for both throughput boosting
3675 * and service guarantees, the return value of this function plays a
3676 * critical role as well.
3677 *
3678 * In a nutshell, this function returns true only if idling is
3679 * beneficial for throughput or, even if detrimental for throughput,
3680 * idling is however necessary to preserve service guarantees (low
3681 * latency, desired throughput distribution, ...). In particular, on
3682 * NCQ-capable devices, this function tries to return false, so as to
3683 * help keep the drives' internal queues full, whenever this helps the
3684 * device boost the throughput without causing any service-guarantee
3685 * issue.
3686 *
3687 * Most of the issues taken into account to get the return value of
3688 * this function are not trivial. We discuss these issues in the two
3689 * functions providing the main pieces of information needed by this
3690 * function.
3691 */
3692static bool bfq_better_to_idle(struct bfq_queue *bfqq)
3693{
3694 struct bfq_data *bfqd = bfqq->bfqd;
3695 bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
3696
3697 if (unlikely(bfqd->strict_guarantees))
3698 return true;
3699
3700 /*
3701 * Idling is performed only if slice_idle > 0. In addition, we
3702 * do not idle if
3703 * (a) bfqq is async
3704 * (b) bfqq is in the idle io prio class: in this case we do
3705 * not idle because we want to minimize the bandwidth that
3706 * queues in this class can steal to higher-priority queues
3707 */
3708 if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
3709 bfq_class_idle(bfqq))
3710 return false;
3711
3712 idling_boosts_thr_with_no_issue =
3713 idling_boosts_thr_without_issues(bfqd, bfqq);
3714
3715 idling_needed_for_service_guar =
3716 idling_needed_for_service_guarantees(bfqd, bfqq);
3717
3718 /*
3719 * We have now the two components we need to compute the
3720 * return value of the function, which is true only if idling
3721 * either boosts the throughput (without issues), or is
3722 * necessary to preserve service guarantees.
3723 */
3724 return idling_boosts_thr_with_no_issue ||
3725 idling_needed_for_service_guar;
3726}
3727
3728/*
3729 * If the in-service queue is empty but the function bfq_better_to_idle
3730 * returns true, then:
3731 * 1) the queue must remain in service and cannot be expired, and
3732 * 2) the device must be idled to wait for the possible arrival of a new
3733 * request for the queue.
3734 * See the comments on the function bfq_better_to_idle for the reasons
3735 * why performing device idling is the best choice to boost the throughput
3736 * and preserve service guarantees when bfq_better_to_idle itself
3737 * returns true.
3738 */
3739static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
3740{
3741 return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
3742}
3743
3744static struct bfq_queue *bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
3745{
3746 struct bfq_queue *bfqq;
3747
3748 /*
3749 * A linear search; but, with a high probability, very few
3750 * steps are needed to find a candidate queue, i.e., a queue
3751 * with enough budget left for its next request. In fact:
3752 * - BFQ dynamically updates the budget of every queue so as
3753 * to accommodate the expected backlog of the queue;
3754 * - if a queue gets all its requests dispatched as injected
3755 * service, then the queue is removed from the active list
3756 * (and re-added only if it gets new requests, but with
3757 * enough budget for its new backlog).
3758 */
3759 list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
3760 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
3761 bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
3762 bfq_bfqq_budget_left(bfqq))
3763 return bfqq;
3764
3765 return NULL;
3766}
3767
3768/*
3769 * Select a queue for service. If we have a current queue in service,
3770 * check whether to continue servicing it, or retrieve and set a new one.
3771 */
3772static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
3773{
3774 struct bfq_queue *bfqq;
3775 struct request *next_rq;
3776 enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
3777
3778 bfqq = bfqd->in_service_queue;
3779 if (!bfqq)
3780 goto new_queue;
3781
3782 bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
3783
3784 /*
3785 * Do not expire bfqq for budget timeout if bfqq may be about
3786 * to enjoy device idling. The reason why, in this case, we
3787 * prevent bfqq from expiring is the same as in the comments
3788 * on the case where bfq_bfqq_must_idle() returns true, in
3789 * bfq_completed_request().
3790 */
3791 if (bfq_may_expire_for_budg_timeout(bfqq) &&
3792 !bfq_bfqq_must_idle(bfqq))
3793 goto expire;
3794
3795check_queue:
3796 /*
3797 * This loop is rarely executed more than once. Even when it
3798 * happens, it is much more convenient to re-execute this loop
3799 * than to return NULL and trigger a new dispatch to get a
3800 * request served.
3801 */
3802 next_rq = bfqq->next_rq;
3803 /*
3804 * If bfqq has requests queued and it has enough budget left to
3805 * serve them, keep the queue, otherwise expire it.
3806 */
3807 if (next_rq) {
3808 if (bfq_serv_to_charge(next_rq, bfqq) >
3809 bfq_bfqq_budget_left(bfqq)) {
3810 /*
3811 * Expire the queue for budget exhaustion,
3812 * which makes sure that the next budget is
3813 * enough to serve the next request, even if
3814 * it comes from the fifo expired path.
3815 */
3816 reason = BFQQE_BUDGET_EXHAUSTED;
3817 goto expire;
3818 } else {
3819 /*
3820 * The idle timer may be pending because we may
3821 * not disable disk idling even when a new request
3822 * arrives.
3823 */
3824 if (bfq_bfqq_wait_request(bfqq)) {
3825 /*
3826 * If we get here: 1) at least a new request
3827 * has arrived but we have not disabled the
3828 * timer because the request was too small,
3829 * 2) then the block layer has unplugged
3830 * the device, causing the dispatch to be
3831 * invoked.
3832 *
3833 * Since the device is unplugged, now the
3834 * requests are probably large enough to
3835 * provide a reasonable throughput.
3836 * So we disable idling.
3837 */
3838 bfq_clear_bfqq_wait_request(bfqq);
3839 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
3840 }
3841 goto keep_queue;
3842 }
3843 }
3844
3845 /*
3846 * No requests pending. However, if the in-service queue is idling
3847 * for a new request, or has requests waiting for a completion and
3848 * may idle after their completion, then keep it anyway.
3849 *
3850 * Yet, to boost throughput, inject service from other queues if
3851 * possible.
3852 */
3853 if (bfq_bfqq_wait_request(bfqq) ||
3854 (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
3855 if (bfq_bfqq_injectable(bfqq) &&
3856 bfqq->injected_service * bfqq->inject_coeff <
3857 bfqq->entity.service * 10)
3858 bfqq = bfq_choose_bfqq_for_injection(bfqd);
3859 else
3860 bfqq = NULL;
3861
3862 goto keep_queue;
3863 }
3864
3865 reason = BFQQE_NO_MORE_REQUESTS;
3866expire:
3867 bfq_bfqq_expire(bfqd, bfqq, false, reason);
3868new_queue:
3869 bfqq = bfq_set_in_service_queue(bfqd);
3870 if (bfqq) {
3871 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
3872 goto check_queue;
3873 }
3874keep_queue:
3875 if (bfqq)
3876 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
3877 else
3878 bfq_log(bfqd, "select_queue: no queue returned");
3879
3880 return bfqq;
3881}
3882
3883static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3884{
3885 struct bfq_entity *entity = &bfqq->entity;
3886
3887 if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
3888 bfq_log_bfqq(bfqd, bfqq,
3889 "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
3890 jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
3891 jiffies_to_msecs(bfqq->wr_cur_max_time),
3892 bfqq->wr_coeff,
3893 bfqq->entity.weight, bfqq->entity.orig_weight);
3894
3895 if (entity->prio_changed)
3896 bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
3897
3898 /*
3899 * If the queue was activated in a burst, or too much
3900 * time has elapsed from the beginning of this
3901 * weight-raising period, then end weight raising.
3902 */
3903 if (bfq_bfqq_in_large_burst(bfqq))
3904 bfq_bfqq_end_wr(bfqq);
3905 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
3906 bfqq->wr_cur_max_time)) {
3907 if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
3908 time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
3909 bfq_wr_duration(bfqd)))
3910 bfq_bfqq_end_wr(bfqq);
3911 else {
3912 switch_back_to_interactive_wr(bfqq, bfqd);
3913 bfqq->entity.prio_changed = 1;
3914 }
3915 }
3916 if (bfqq->wr_coeff > 1 &&
3917 bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
3918 bfqq->service_from_wr > max_service_from_wr) {
3919 /* see comments on max_service_from_wr */
3920 bfq_bfqq_end_wr(bfqq);
3921 }
3922 }
3923 /*
3924 * To improve latency (for this or other queues), immediately
3925 * update weight both if it must be raised and if it must be
3926 * lowered. Since, entity may be on some active tree here, and
3927 * might have a pending change of its ioprio class, invoke
3928 * next function with the last parameter unset (see the
3929 * comments on the function).
3930 */
3931 if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
3932 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
3933 entity, false);
3934}
3935
3936/*
3937 * Dispatch next request from bfqq.
3938 */
3939static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
3940 struct bfq_queue *bfqq)
3941{
3942 struct request *rq = bfqq->next_rq;
3943 unsigned long service_to_charge;
3944
3945 service_to_charge = bfq_serv_to_charge(rq, bfqq);
3946
3947 bfq_bfqq_served(bfqq, service_to_charge);
3948
3949 bfq_dispatch_remove(bfqd->queue, rq);
3950
3951 if (bfqq != bfqd->in_service_queue) {
3952 if (likely(bfqd->in_service_queue))
3953 bfqd->in_service_queue->injected_service +=
3954 bfq_serv_to_charge(rq, bfqq);
3955
3956 goto return_rq;
3957 }
3958
3959 /*
3960 * If weight raising has to terminate for bfqq, then next
3961 * function causes an immediate update of bfqq's weight,
3962 * without waiting for next activation. As a consequence, on
3963 * expiration, bfqq will be timestamped as if has never been
3964 * weight-raised during this service slot, even if it has
3965 * received part or even most of the service as a
3966 * weight-raised queue. This inflates bfqq's timestamps, which
3967 * is beneficial, as bfqq is then more willing to leave the
3968 * device immediately to possible other weight-raised queues.
3969 */
3970 bfq_update_wr_data(bfqd, bfqq);
3971
3972 /*
3973 * Expire bfqq, pretending that its budget expired, if bfqq
3974 * belongs to CLASS_IDLE and other queues are waiting for
3975 * service.
3976 */
3977 if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
3978 goto return_rq;
3979
3980 bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
3981
3982return_rq:
3983 return rq;
3984}
3985
3986static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
3987{
3988 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3989
3990 /*
3991 * Avoiding lock: a race on bfqd->busy_queues should cause at
3992 * most a call to dispatch for nothing
3993 */
3994 return !list_empty_careful(&bfqd->dispatch) ||
3995 bfq_tot_busy_queues(bfqd) > 0;
3996}
3997
3998static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3999{
4000 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4001 struct request *rq = NULL;
4002 struct bfq_queue *bfqq = NULL;
4003
4004 if (!list_empty(&bfqd->dispatch)) {
4005 rq = list_first_entry(&bfqd->dispatch, struct request,
4006 queuelist);
4007 list_del_init(&rq->queuelist);
4008
4009 bfqq = RQ_BFQQ(rq);
4010
4011 if (bfqq) {
4012 /*
4013 * Increment counters here, because this
4014 * dispatch does not follow the standard
4015 * dispatch flow (where counters are
4016 * incremented)
4017 */
4018 bfqq->dispatched++;
4019
4020 goto inc_in_driver_start_rq;
4021 }
4022
4023 /*
4024 * We exploit the bfq_finish_requeue_request hook to
4025 * decrement rq_in_driver, but
4026 * bfq_finish_requeue_request will not be invoked on
4027 * this request. So, to avoid unbalance, just start
4028 * this request, without incrementing rq_in_driver. As
4029 * a negative consequence, rq_in_driver is deceptively
4030 * lower than it should be while this request is in
4031 * service. This may cause bfq_schedule_dispatch to be
4032 * invoked uselessly.
4033 *
4034 * As for implementing an exact solution, the
4035 * bfq_finish_requeue_request hook, if defined, is
4036 * probably invoked also on this request. So, by
4037 * exploiting this hook, we could 1) increment
4038 * rq_in_driver here, and 2) decrement it in
4039 * bfq_finish_requeue_request. Such a solution would
4040 * let the value of the counter be always accurate,
4041 * but it would entail using an extra interface
4042 * function. This cost seems higher than the benefit,
4043 * being the frequency of non-elevator-private
4044 * requests very low.
4045 */
4046 goto start_rq;
4047 }
4048
4049 bfq_log(bfqd, "dispatch requests: %d busy queues",
4050 bfq_tot_busy_queues(bfqd));
4051
4052 if (bfq_tot_busy_queues(bfqd) == 0)
4053 goto exit;
4054
4055 /*
4056 * Force device to serve one request at a time if
4057 * strict_guarantees is true. Forcing this service scheme is
4058 * currently the ONLY way to guarantee that the request
4059 * service order enforced by the scheduler is respected by a
4060 * queueing device. Otherwise the device is free even to make
4061 * some unlucky request wait for as long as the device
4062 * wishes.
4063 *
4064 * Of course, serving one request at at time may cause loss of
4065 * throughput.
4066 */
4067 if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
4068 goto exit;
4069
4070 bfqq = bfq_select_queue(bfqd);
4071 if (!bfqq)
4072 goto exit;
4073
4074 rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
4075
4076 if (rq) {
4077inc_in_driver_start_rq:
4078 bfqd->rq_in_driver++;
4079start_rq:
4080 rq->rq_flags |= RQF_STARTED;
4081 }
4082exit:
4083 return rq;
4084}
4085
4086#if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
4087static void bfq_update_dispatch_stats(struct request_queue *q,
4088 struct request *rq,
4089 struct bfq_queue *in_serv_queue,
4090 bool idle_timer_disabled)
4091{
4092 struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
4093
4094 if (!idle_timer_disabled && !bfqq)
4095 return;
4096
4097 /*
4098 * rq and bfqq are guaranteed to exist until this function
4099 * ends, for the following reasons. First, rq can be
4100 * dispatched to the device, and then can be completed and
4101 * freed, only after this function ends. Second, rq cannot be
4102 * merged (and thus freed because of a merge) any longer,
4103 * because it has already started. Thus rq cannot be freed
4104 * before this function ends, and, since rq has a reference to
4105 * bfqq, the same guarantee holds for bfqq too.
4106 *
4107 * In addition, the following queue lock guarantees that
4108 * bfqq_group(bfqq) exists as well.
4109 */
4110 spin_lock_irq(&q->queue_lock);
4111 if (idle_timer_disabled)
4112 /*
4113 * Since the idle timer has been disabled,
4114 * in_serv_queue contained some request when
4115 * __bfq_dispatch_request was invoked above, which
4116 * implies that rq was picked exactly from
4117 * in_serv_queue. Thus in_serv_queue == bfqq, and is
4118 * therefore guaranteed to exist because of the above
4119 * arguments.
4120 */
4121 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
4122 if (bfqq) {
4123 struct bfq_group *bfqg = bfqq_group(bfqq);
4124
4125 bfqg_stats_update_avg_queue_size(bfqg);
4126 bfqg_stats_set_start_empty_time(bfqg);
4127 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
4128 }
4129 spin_unlock_irq(&q->queue_lock);
4130}
4131#else
4132static inline void bfq_update_dispatch_stats(struct request_queue *q,
4133 struct request *rq,
4134 struct bfq_queue *in_serv_queue,
4135 bool idle_timer_disabled) {}
4136#endif
4137
4138static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
4139{
4140 struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4141 struct request *rq;
4142 struct bfq_queue *in_serv_queue;
4143 bool waiting_rq, idle_timer_disabled;
4144
4145 spin_lock_irq(&bfqd->lock);
4146
4147 in_serv_queue = bfqd->in_service_queue;
4148 waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
4149
4150 rq = __bfq_dispatch_request(hctx);
4151
4152 idle_timer_disabled =
4153 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
4154
4155 spin_unlock_irq(&bfqd->lock);
4156
4157 bfq_update_dispatch_stats(hctx->queue, rq, in_serv_queue,
4158 idle_timer_disabled);
4159
4160 return rq;
4161}
4162
4163/*
4164 * Task holds one reference to the queue, dropped when task exits. Each rq
4165 * in-flight on this queue also holds a reference, dropped when rq is freed.
4166 *
4167 * Scheduler lock must be held here. Recall not to use bfqq after calling
4168 * this function on it.
4169 */
4170void bfq_put_queue(struct bfq_queue *bfqq)
4171{
4172#ifdef CONFIG_BFQ_GROUP_IOSCHED
4173 struct bfq_group *bfqg = bfqq_group(bfqq);
4174#endif
4175
4176 if (bfqq->bfqd)
4177 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
4178 bfqq, bfqq->ref);
4179
4180 bfqq->ref--;
4181 if (bfqq->ref)
4182 return;
4183
4184 if (!hlist_unhashed(&bfqq->burst_list_node)) {
4185 hlist_del_init(&bfqq->burst_list_node);
4186 /*
4187 * Decrement also burst size after the removal, if the
4188 * process associated with bfqq is exiting, and thus
4189 * does not contribute to the burst any longer. This
4190 * decrement helps filter out false positives of large
4191 * bursts, when some short-lived process (often due to
4192 * the execution of commands by some service) happens
4193 * to start and exit while a complex application is
4194 * starting, and thus spawning several processes that
4195 * do I/O (and that *must not* be treated as a large
4196 * burst, see comments on bfq_handle_burst).
4197 *
4198 * In particular, the decrement is performed only if:
4199 * 1) bfqq is not a merged queue, because, if it is,
4200 * then this free of bfqq is not triggered by the exit
4201 * of the process bfqq is associated with, but exactly
4202 * by the fact that bfqq has just been merged.
4203 * 2) burst_size is greater than 0, to handle
4204 * unbalanced decrements. Unbalanced decrements may
4205 * happen in te following case: bfqq is inserted into
4206 * the current burst list--without incrementing
4207 * bust_size--because of a split, but the current
4208 * burst list is not the burst list bfqq belonged to
4209 * (see comments on the case of a split in
4210 * bfq_set_request).
4211 */
4212 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
4213 bfqq->bfqd->burst_size--;
4214 }
4215
4216 kmem_cache_free(bfq_pool, bfqq);
4217#ifdef CONFIG_BFQ_GROUP_IOSCHED
4218 bfqg_and_blkg_put(bfqg);
4219#endif
4220}
4221
4222static void bfq_put_cooperator(struct bfq_queue *bfqq)
4223{
4224 struct bfq_queue *__bfqq, *next;
4225
4226 /*
4227 * If this queue was scheduled to merge with another queue, be
4228 * sure to drop the reference taken on that queue (and others in
4229 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
4230 */
4231 __bfqq = bfqq->new_bfqq;
4232 while (__bfqq) {
4233 if (__bfqq == bfqq)
4234 break;
4235 next = __bfqq->new_bfqq;
4236 bfq_put_queue(__bfqq);
4237 __bfqq = next;
4238 }
4239}
4240
4241static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4242{
4243 if (bfqq == bfqd->in_service_queue) {
4244 __bfq_bfqq_expire(bfqd, bfqq);
4245 bfq_schedule_dispatch(bfqd);
4246 }
4247
4248 bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
4249
4250 bfq_put_cooperator(bfqq);
4251
4252 bfq_put_queue(bfqq); /* release process reference */
4253}
4254
4255static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
4256{
4257 struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4258 struct bfq_data *bfqd;
4259
4260 if (bfqq)
4261 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
4262
4263 if (bfqq && bfqd) {
4264 unsigned long flags;
4265
4266 spin_lock_irqsave(&bfqd->lock, flags);
4267 bfq_exit_bfqq(bfqd, bfqq);
4268 bic_set_bfqq(bic, NULL, is_sync);
4269 spin_unlock_irqrestore(&bfqd->lock, flags);
4270 }
4271}
4272
4273static void bfq_exit_icq(struct io_cq *icq)
4274{
4275 struct bfq_io_cq *bic = icq_to_bic(icq);
4276
4277 bfq_exit_icq_bfqq(bic, true);
4278 bfq_exit_icq_bfqq(bic, false);
4279}
4280
4281/*
4282 * Update the entity prio values; note that the new values will not
4283 * be used until the next (re)activation.
4284 */
4285static void
4286bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
4287{
4288 struct task_struct *tsk = current;
4289 int ioprio_class;
4290 struct bfq_data *bfqd = bfqq->bfqd;
4291
4292 if (!bfqd)
4293 return;
4294
4295 ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4296 switch (ioprio_class) {
4297 default:
4298 dev_err(bfqq->bfqd->queue->backing_dev_info->dev,
4299 "bfq: bad prio class %d\n", ioprio_class);
4300 /* fall through */
4301 case IOPRIO_CLASS_NONE:
4302 /*
4303 * No prio set, inherit CPU scheduling settings.
4304 */
4305 bfqq->new_ioprio = task_nice_ioprio(tsk);
4306 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
4307 break;
4308 case IOPRIO_CLASS_RT:
4309 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4310 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
4311 break;
4312 case IOPRIO_CLASS_BE:
4313 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4314 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
4315 break;
4316 case IOPRIO_CLASS_IDLE:
4317 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
4318 bfqq->new_ioprio = 7;
4319 break;
4320 }
4321
4322 if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
4323 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
4324 bfqq->new_ioprio);
4325 bfqq->new_ioprio = IOPRIO_BE_NR;
4326 }
4327
4328 bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
4329 bfqq->entity.prio_changed = 1;
4330}
4331
4332static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4333 struct bio *bio, bool is_sync,
4334 struct bfq_io_cq *bic);
4335
4336static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
4337{
4338 struct bfq_data *bfqd = bic_to_bfqd(bic);
4339 struct bfq_queue *bfqq;
4340 int ioprio = bic->icq.ioc->ioprio;
4341
4342 /*
4343 * This condition may trigger on a newly created bic, be sure to
4344 * drop the lock before returning.
4345 */
4346 if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
4347 return;
4348
4349 bic->ioprio = ioprio;
4350
4351 bfqq = bic_to_bfqq(bic, false);
4352 if (bfqq) {
4353 /* release process reference on this queue */
4354 bfq_put_queue(bfqq);
4355 bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
4356 bic_set_bfqq(bic, bfqq, false);
4357 }
4358
4359 bfqq = bic_to_bfqq(bic, true);
4360 if (bfqq)
4361 bfq_set_next_ioprio_data(bfqq, bic);
4362}
4363
4364static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4365 struct bfq_io_cq *bic, pid_t pid, int is_sync)
4366{
4367 RB_CLEAR_NODE(&bfqq->entity.rb_node);
4368 INIT_LIST_HEAD(&bfqq->fifo);
4369 INIT_HLIST_NODE(&bfqq->burst_list_node);
4370
4371 bfqq->ref = 0;
4372 bfqq->bfqd = bfqd;
4373
4374 if (bic)
4375 bfq_set_next_ioprio_data(bfqq, bic);
4376
4377 if (is_sync) {
4378 /*
4379 * No need to mark as has_short_ttime if in
4380 * idle_class, because no device idling is performed
4381 * for queues in idle class
4382 */
4383 if (!bfq_class_idle(bfqq))
4384 /* tentatively mark as has_short_ttime */
4385 bfq_mark_bfqq_has_short_ttime(bfqq);
4386 bfq_mark_bfqq_sync(bfqq);
4387 bfq_mark_bfqq_just_created(bfqq);
4388 /*
4389 * Aggressively inject a lot of service: up to 90%.
4390 * This coefficient remains constant during bfqq life,
4391 * but this behavior might be changed, after enough
4392 * testing and tuning.
4393 */
4394 bfqq->inject_coeff = 1;
4395 } else
4396 bfq_clear_bfqq_sync(bfqq);
4397
4398 /* set end request to minus infinity from now */
4399 bfqq->ttime.last_end_request = ktime_get_ns() + 1;
4400
4401 bfq_mark_bfqq_IO_bound(bfqq);
4402
4403 bfqq->pid = pid;
4404
4405 /* Tentative initial value to trade off between thr and lat */
4406 bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
4407 bfqq->budget_timeout = bfq_smallest_from_now();
4408
4409 bfqq->wr_coeff = 1;
4410 bfqq->last_wr_start_finish = jiffies;
4411 bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
4412 bfqq->split_time = bfq_smallest_from_now();
4413
4414 /*
4415 * To not forget the possibly high bandwidth consumed by a
4416 * process/queue in the recent past,
4417 * bfq_bfqq_softrt_next_start() returns a value at least equal
4418 * to the current value of bfqq->soft_rt_next_start (see
4419 * comments on bfq_bfqq_softrt_next_start). Set
4420 * soft_rt_next_start to now, to mean that bfqq has consumed
4421 * no bandwidth so far.
4422 */
4423 bfqq->soft_rt_next_start = jiffies;
4424
4425 /* first request is almost certainly seeky */
4426 bfqq->seek_history = 1;
4427}
4428
4429static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
4430 struct bfq_group *bfqg,
4431 int ioprio_class, int ioprio)
4432{
4433 switch (ioprio_class) {
4434 case IOPRIO_CLASS_RT:
4435 return &bfqg->async_bfqq[0][ioprio];
4436 case IOPRIO_CLASS_NONE:
4437 ioprio = IOPRIO_NORM;
4438 /* fall through */
4439 case IOPRIO_CLASS_BE:
4440 return &bfqg->async_bfqq[1][ioprio];
4441 case IOPRIO_CLASS_IDLE:
4442 return &bfqg->async_idle_bfqq;
4443 default:
4444 return NULL;
4445 }
4446}
4447
4448static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4449 struct bio *bio, bool is_sync,
4450 struct bfq_io_cq *bic)
4451{
4452 const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4453 const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4454 struct bfq_queue **async_bfqq = NULL;
4455 struct bfq_queue *bfqq;
4456 struct bfq_group *bfqg;
4457
4458 rcu_read_lock();
4459
4460 bfqg = bfq_find_set_group(bfqd, __bio_blkcg(bio));
4461 if (!bfqg) {
4462 bfqq = &bfqd->oom_bfqq;
4463 goto out;
4464 }
4465
4466 if (!is_sync) {
4467 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
4468 ioprio);
4469 bfqq = *async_bfqq;
4470 if (bfqq)
4471 goto out;
4472 }
4473
4474 bfqq = kmem_cache_alloc_node(bfq_pool,
4475 GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
4476 bfqd->queue->node);
4477
4478 if (bfqq) {
4479 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
4480 is_sync);
4481 bfq_init_entity(&bfqq->entity, bfqg);
4482 bfq_log_bfqq(bfqd, bfqq, "allocated");
4483 } else {
4484 bfqq = &bfqd->oom_bfqq;
4485 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
4486 goto out;
4487 }
4488
4489 /*
4490 * Pin the queue now that it's allocated, scheduler exit will
4491 * prune it.
4492 */
4493 if (