xref: /openbmc/linux/block/bfq-iosched.c (revision b9df3997)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Budget Fair Queueing (BFQ) I/O scheduler.
4  *
5  * Based on ideas and code from CFQ:
6  * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7  *
8  * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9  *		      Paolo Valente <paolo.valente@unimore.it>
10  *
11  * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
12  *                    Arianna Avanzini <avanzini@google.com>
13  *
14  * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
15  *
16  * BFQ is a proportional-share I/O scheduler, with some extra
17  * low-latency capabilities. BFQ also supports full hierarchical
18  * scheduling through cgroups. Next paragraphs provide an introduction
19  * on BFQ inner workings. Details on BFQ benefits, usage and
20  * limitations can be found in Documentation/block/bfq-iosched.rst.
21  *
22  * BFQ is a proportional-share storage-I/O scheduling algorithm based
23  * on the slice-by-slice service scheme of CFQ. But BFQ assigns
24  * budgets, measured in number of sectors, to processes instead of
25  * time slices. The device is not granted to the in-service process
26  * for a given time slice, but until it has exhausted its assigned
27  * budget. This change from the time to the service domain enables BFQ
28  * to distribute the device throughput among processes as desired,
29  * without any distortion due to throughput fluctuations, or to device
30  * internal queueing. BFQ uses an ad hoc internal scheduler, called
31  * B-WF2Q+, to schedule processes according to their budgets. More
32  * precisely, BFQ schedules queues associated with processes. Each
33  * process/queue is assigned a user-configurable weight, and B-WF2Q+
34  * guarantees that each queue receives a fraction of the throughput
35  * proportional to its weight. Thanks to the accurate policy of
36  * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
37  * processes issuing sequential requests (to boost the throughput),
38  * and yet guarantee a low latency to interactive and soft real-time
39  * applications.
40  *
41  * In particular, to provide these low-latency guarantees, BFQ
42  * explicitly privileges the I/O of two classes of time-sensitive
43  * applications: interactive and soft real-time. In more detail, BFQ
44  * behaves this way if the low_latency parameter is set (default
45  * configuration). This feature enables BFQ to provide applications in
46  * these classes with a very low latency.
47  *
48  * To implement this feature, BFQ constantly tries to detect whether
49  * the I/O requests in a bfq_queue come from an interactive or a soft
50  * real-time application. For brevity, in these cases, the queue is
51  * said to be interactive or soft real-time. In both cases, BFQ
52  * privileges the service of the queue, over that of non-interactive
53  * and non-soft-real-time queues. This privileging is performed,
54  * mainly, by raising the weight of the queue. So, for brevity, we
55  * call just weight-raising periods the time periods during which a
56  * queue is privileged, because deemed interactive or soft real-time.
57  *
58  * The detection of soft real-time queues/applications is described in
59  * detail in the comments on the function
60  * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
61  * interactive queue works as follows: a queue is deemed interactive
62  * if it is constantly non empty only for a limited time interval,
63  * after which it does become empty. The queue may be deemed
64  * interactive again (for a limited time), if it restarts being
65  * constantly non empty, provided that this happens only after the
66  * queue has remained empty for a given minimum idle time.
67  *
68  * By default, BFQ computes automatically the above maximum time
69  * interval, i.e., the time interval after which a constantly
70  * non-empty queue stops being deemed interactive. Since a queue is
71  * weight-raised while it is deemed interactive, this maximum time
72  * interval happens to coincide with the (maximum) duration of the
73  * weight-raising for interactive queues.
74  *
75  * Finally, BFQ also features additional heuristics for
76  * preserving both a low latency and a high throughput on NCQ-capable,
77  * rotational or flash-based devices, and to get the job done quickly
78  * for applications consisting in many I/O-bound processes.
79  *
80  * NOTE: if the main or only goal, with a given device, is to achieve
81  * the maximum-possible throughput at all times, then do switch off
82  * all low-latency heuristics for that device, by setting low_latency
83  * to 0.
84  *
85  * BFQ is described in [1], where also a reference to the initial,
86  * more theoretical paper on BFQ can be found. The interested reader
87  * can find in the latter paper full details on the main algorithm, as
88  * well as formulas of the guarantees and formal proofs of all the
89  * properties.  With respect to the version of BFQ presented in these
90  * papers, this implementation adds a few more heuristics, such as the
91  * ones that guarantee a low latency to interactive and soft real-time
92  * applications, and a hierarchical extension based on H-WF2Q+.
93  *
94  * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
95  * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
96  * with O(log N) complexity derives from the one introduced with EEVDF
97  * in [3].
98  *
99  * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
100  *     Scheduler", Proceedings of the First Workshop on Mobile System
101  *     Technologies (MST-2015), May 2015.
102  *     http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
103  *
104  * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
105  *     Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
106  *     Oct 1997.
107  *
108  * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
109  *
110  * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
111  *     First: A Flexible and Accurate Mechanism for Proportional Share
112  *     Resource Allocation", technical report.
113  *
114  * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
115  */
116 #include <linux/module.h>
117 #include <linux/slab.h>
118 #include <linux/blkdev.h>
119 #include <linux/cgroup.h>
120 #include <linux/elevator.h>
121 #include <linux/ktime.h>
122 #include <linux/rbtree.h>
123 #include <linux/ioprio.h>
124 #include <linux/sbitmap.h>
125 #include <linux/delay.h>
126 
127 #include "blk.h"
128 #include "blk-mq.h"
129 #include "blk-mq-tag.h"
130 #include "blk-mq-sched.h"
131 #include "bfq-iosched.h"
132 #include "blk-wbt.h"
133 
134 #define BFQ_BFQQ_FNS(name)						\
135 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq)			\
136 {									\
137 	__set_bit(BFQQF_##name, &(bfqq)->flags);			\
138 }									\
139 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq)			\
140 {									\
141 	__clear_bit(BFQQF_##name, &(bfqq)->flags);		\
142 }									\
143 int bfq_bfqq_##name(const struct bfq_queue *bfqq)			\
144 {									\
145 	return test_bit(BFQQF_##name, &(bfqq)->flags);		\
146 }
147 
148 BFQ_BFQQ_FNS(just_created);
149 BFQ_BFQQ_FNS(busy);
150 BFQ_BFQQ_FNS(wait_request);
151 BFQ_BFQQ_FNS(non_blocking_wait_rq);
152 BFQ_BFQQ_FNS(fifo_expire);
153 BFQ_BFQQ_FNS(has_short_ttime);
154 BFQ_BFQQ_FNS(sync);
155 BFQ_BFQQ_FNS(IO_bound);
156 BFQ_BFQQ_FNS(in_large_burst);
157 BFQ_BFQQ_FNS(coop);
158 BFQ_BFQQ_FNS(split_coop);
159 BFQ_BFQQ_FNS(softrt_update);
160 BFQ_BFQQ_FNS(has_waker);
161 #undef BFQ_BFQQ_FNS						\
162 
163 /* Expiration time of sync (0) and async (1) requests, in ns. */
164 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
165 
166 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
167 static const int bfq_back_max = 16 * 1024;
168 
169 /* Penalty of a backwards seek, in number of sectors. */
170 static const int bfq_back_penalty = 2;
171 
172 /* Idling period duration, in ns. */
173 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
174 
175 /* Minimum number of assigned budgets for which stats are safe to compute. */
176 static const int bfq_stats_min_budgets = 194;
177 
178 /* Default maximum budget values, in sectors and number of requests. */
179 static const int bfq_default_max_budget = 16 * 1024;
180 
181 /*
182  * When a sync request is dispatched, the queue that contains that
183  * request, and all the ancestor entities of that queue, are charged
184  * with the number of sectors of the request. In contrast, if the
185  * request is async, then the queue and its ancestor entities are
186  * charged with the number of sectors of the request, multiplied by
187  * the factor below. This throttles the bandwidth for async I/O,
188  * w.r.t. to sync I/O, and it is done to counter the tendency of async
189  * writes to steal I/O throughput to reads.
190  *
191  * The current value of this parameter is the result of a tuning with
192  * several hardware and software configurations. We tried to find the
193  * lowest value for which writes do not cause noticeable problems to
194  * reads. In fact, the lower this parameter, the stabler I/O control,
195  * in the following respect.  The lower this parameter is, the less
196  * the bandwidth enjoyed by a group decreases
197  * - when the group does writes, w.r.t. to when it does reads;
198  * - when other groups do reads, w.r.t. to when they do writes.
199  */
200 static const int bfq_async_charge_factor = 3;
201 
202 /* Default timeout values, in jiffies, approximating CFQ defaults. */
203 const int bfq_timeout = HZ / 8;
204 
205 /*
206  * Time limit for merging (see comments in bfq_setup_cooperator). Set
207  * to the slowest value that, in our tests, proved to be effective in
208  * removing false positives, while not causing true positives to miss
209  * queue merging.
210  *
211  * As can be deduced from the low time limit below, queue merging, if
212  * successful, happens at the very beginning of the I/O of the involved
213  * cooperating processes, as a consequence of the arrival of the very
214  * first requests from each cooperator.  After that, there is very
215  * little chance to find cooperators.
216  */
217 static const unsigned long bfq_merge_time_limit = HZ/10;
218 
219 static struct kmem_cache *bfq_pool;
220 
221 /* Below this threshold (in ns), we consider thinktime immediate. */
222 #define BFQ_MIN_TT		(2 * NSEC_PER_MSEC)
223 
224 /* hw_tag detection: parallel requests threshold and min samples needed. */
225 #define BFQ_HW_QUEUE_THRESHOLD	3
226 #define BFQ_HW_QUEUE_SAMPLES	32
227 
228 #define BFQQ_SEEK_THR		(sector_t)(8 * 100)
229 #define BFQQ_SECT_THR_NONROT	(sector_t)(2 * 32)
230 #define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \
231 	(get_sdist(last_pos, rq) >			\
232 	 BFQQ_SEEK_THR &&				\
233 	 (!blk_queue_nonrot(bfqd->queue) ||		\
234 	  blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT))
235 #define BFQQ_CLOSE_THR		(sector_t)(8 * 1024)
236 #define BFQQ_SEEKY(bfqq)	(hweight32(bfqq->seek_history) > 19)
237 /*
238  * Sync random I/O is likely to be confused with soft real-time I/O,
239  * because it is characterized by limited throughput and apparently
240  * isochronous arrival pattern. To avoid false positives, queues
241  * containing only random (seeky) I/O are prevented from being tagged
242  * as soft real-time.
243  */
244 #define BFQQ_TOTALLY_SEEKY(bfqq)	(bfqq->seek_history == -1)
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  */
303 static 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  */
309 static 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  */
363 static 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 
368 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
369 {
370 	return bic->bfqq[is_sync];
371 }
372 
373 void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
374 {
375 	bic->bfqq[is_sync] = bfqq;
376 }
377 
378 struct 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  */
387 static 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  */
399 static 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  */
421 void 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 closer to the head right now.  Distance
437  * behind the head is penalized and only allowed to a certain extent.
438  */
439 static 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  */
535 static 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 
550 static struct bfq_queue *
551 bfq_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 
591 static 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 
598 /*
599  * The following function is not marked as __cold because it is
600  * actually cold, but for the same performance goal described in the
601  * comments on the likely() at the beginning of
602  * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
603  * execution time for the case where this function is not invoked, we
604  * had to add an unlikely() in each involved if().
605  */
606 void __cold
607 bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
608 {
609 	struct rb_node **p, *parent;
610 	struct bfq_queue *__bfqq;
611 
612 	if (bfqq->pos_root) {
613 		rb_erase(&bfqq->pos_node, bfqq->pos_root);
614 		bfqq->pos_root = NULL;
615 	}
616 
617 	/*
618 	 * bfqq cannot be merged any longer (see comments in
619 	 * bfq_setup_cooperator): no point in adding bfqq into the
620 	 * position tree.
621 	 */
622 	if (bfq_too_late_for_merging(bfqq))
623 		return;
624 
625 	if (bfq_class_idle(bfqq))
626 		return;
627 	if (!bfqq->next_rq)
628 		return;
629 
630 	bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
631 	__bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
632 			blk_rq_pos(bfqq->next_rq), &parent, &p);
633 	if (!__bfqq) {
634 		rb_link_node(&bfqq->pos_node, parent, p);
635 		rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
636 	} else
637 		bfqq->pos_root = NULL;
638 }
639 
640 /*
641  * The following function returns false either if every active queue
642  * must receive the same share of the throughput (symmetric scenario),
643  * or, as a special case, if bfqq must receive a share of the
644  * throughput lower than or equal to the share that every other active
645  * queue must receive.  If bfqq does sync I/O, then these are the only
646  * two cases where bfqq happens to be guaranteed its share of the
647  * throughput even if I/O dispatching is not plugged when bfqq remains
648  * temporarily empty (for more details, see the comments in the
649  * function bfq_better_to_idle()). For this reason, the return value
650  * of this function is used to check whether I/O-dispatch plugging can
651  * be avoided.
652  *
653  * The above first case (symmetric scenario) occurs when:
654  * 1) all active queues have the same weight,
655  * 2) all active queues belong to the same I/O-priority class,
656  * 3) all active groups at the same level in the groups tree have the same
657  *    weight,
658  * 4) all active groups at the same level in the groups tree have the same
659  *    number of children.
660  *
661  * Unfortunately, keeping the necessary state for evaluating exactly
662  * the last two symmetry sub-conditions above would be quite complex
663  * and time consuming. Therefore this function evaluates, instead,
664  * only the following stronger three sub-conditions, for which it is
665  * much easier to maintain the needed state:
666  * 1) all active queues have the same weight,
667  * 2) all active queues belong to the same I/O-priority class,
668  * 3) there are no active groups.
669  * In particular, the last condition is always true if hierarchical
670  * support or the cgroups interface are not enabled, thus no state
671  * needs to be maintained in this case.
672  */
673 static bool bfq_asymmetric_scenario(struct bfq_data *bfqd,
674 				   struct bfq_queue *bfqq)
675 {
676 	bool smallest_weight = bfqq &&
677 		bfqq->weight_counter &&
678 		bfqq->weight_counter ==
679 		container_of(
680 			rb_first_cached(&bfqd->queue_weights_tree),
681 			struct bfq_weight_counter,
682 			weights_node);
683 
684 	/*
685 	 * For queue weights to differ, queue_weights_tree must contain
686 	 * at least two nodes.
687 	 */
688 	bool varied_queue_weights = !smallest_weight &&
689 		!RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) &&
690 		(bfqd->queue_weights_tree.rb_root.rb_node->rb_left ||
691 		 bfqd->queue_weights_tree.rb_root.rb_node->rb_right);
692 
693 	bool multiple_classes_busy =
694 		(bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
695 		(bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
696 		(bfqd->busy_queues[1] && bfqd->busy_queues[2]);
697 
698 	return varied_queue_weights || multiple_classes_busy
699 #ifdef CONFIG_BFQ_GROUP_IOSCHED
700 	       || bfqd->num_groups_with_pending_reqs > 0
701 #endif
702 		;
703 }
704 
705 /*
706  * If the weight-counter tree passed as input contains no counter for
707  * the weight of the input queue, then add that counter; otherwise just
708  * increment the existing counter.
709  *
710  * Note that weight-counter trees contain few nodes in mostly symmetric
711  * scenarios. For example, if all queues have the same weight, then the
712  * weight-counter tree for the queues may contain at most one node.
713  * This holds even if low_latency is on, because weight-raised queues
714  * are not inserted in the tree.
715  * In most scenarios, the rate at which nodes are created/destroyed
716  * should be low too.
717  */
718 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq,
719 			  struct rb_root_cached *root)
720 {
721 	struct bfq_entity *entity = &bfqq->entity;
722 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
723 	bool leftmost = true;
724 
725 	/*
726 	 * Do not insert if the queue is already associated with a
727 	 * counter, which happens if:
728 	 *   1) a request arrival has caused the queue to become both
729 	 *      non-weight-raised, and hence change its weight, and
730 	 *      backlogged; in this respect, each of the two events
731 	 *      causes an invocation of this function,
732 	 *   2) this is the invocation of this function caused by the
733 	 *      second event. This second invocation is actually useless,
734 	 *      and we handle this fact by exiting immediately. More
735 	 *      efficient or clearer solutions might possibly be adopted.
736 	 */
737 	if (bfqq->weight_counter)
738 		return;
739 
740 	while (*new) {
741 		struct bfq_weight_counter *__counter = container_of(*new,
742 						struct bfq_weight_counter,
743 						weights_node);
744 		parent = *new;
745 
746 		if (entity->weight == __counter->weight) {
747 			bfqq->weight_counter = __counter;
748 			goto inc_counter;
749 		}
750 		if (entity->weight < __counter->weight)
751 			new = &((*new)->rb_left);
752 		else {
753 			new = &((*new)->rb_right);
754 			leftmost = false;
755 		}
756 	}
757 
758 	bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
759 				       GFP_ATOMIC);
760 
761 	/*
762 	 * In the unlucky event of an allocation failure, we just
763 	 * exit. This will cause the weight of queue to not be
764 	 * considered in bfq_asymmetric_scenario, which, in its turn,
765 	 * causes the scenario to be deemed wrongly symmetric in case
766 	 * bfqq's weight would have been the only weight making the
767 	 * scenario asymmetric.  On the bright side, no unbalance will
768 	 * however occur when bfqq becomes inactive again (the
769 	 * invocation of this function is triggered by an activation
770 	 * of queue).  In fact, bfq_weights_tree_remove does nothing
771 	 * if !bfqq->weight_counter.
772 	 */
773 	if (unlikely(!bfqq->weight_counter))
774 		return;
775 
776 	bfqq->weight_counter->weight = entity->weight;
777 	rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
778 	rb_insert_color_cached(&bfqq->weight_counter->weights_node, root,
779 				leftmost);
780 
781 inc_counter:
782 	bfqq->weight_counter->num_active++;
783 	bfqq->ref++;
784 }
785 
786 /*
787  * Decrement the weight counter associated with the queue, and, if the
788  * counter reaches 0, remove the counter from the tree.
789  * See the comments to the function bfq_weights_tree_add() for considerations
790  * about overhead.
791  */
792 void __bfq_weights_tree_remove(struct bfq_data *bfqd,
793 			       struct bfq_queue *bfqq,
794 			       struct rb_root_cached *root)
795 {
796 	if (!bfqq->weight_counter)
797 		return;
798 
799 	bfqq->weight_counter->num_active--;
800 	if (bfqq->weight_counter->num_active > 0)
801 		goto reset_entity_pointer;
802 
803 	rb_erase_cached(&bfqq->weight_counter->weights_node, root);
804 	kfree(bfqq->weight_counter);
805 
806 reset_entity_pointer:
807 	bfqq->weight_counter = NULL;
808 	bfq_put_queue(bfqq);
809 }
810 
811 /*
812  * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
813  * of active groups for each queue's inactive parent entity.
814  */
815 void bfq_weights_tree_remove(struct bfq_data *bfqd,
816 			     struct bfq_queue *bfqq)
817 {
818 	struct bfq_entity *entity = bfqq->entity.parent;
819 
820 	for_each_entity(entity) {
821 		struct bfq_sched_data *sd = entity->my_sched_data;
822 
823 		if (sd->next_in_service || sd->in_service_entity) {
824 			/*
825 			 * entity is still active, because either
826 			 * next_in_service or in_service_entity is not
827 			 * NULL (see the comments on the definition of
828 			 * next_in_service for details on why
829 			 * in_service_entity must be checked too).
830 			 *
831 			 * As a consequence, its parent entities are
832 			 * active as well, and thus this loop must
833 			 * stop here.
834 			 */
835 			break;
836 		}
837 
838 		/*
839 		 * The decrement of num_groups_with_pending_reqs is
840 		 * not performed immediately upon the deactivation of
841 		 * entity, but it is delayed to when it also happens
842 		 * that the first leaf descendant bfqq of entity gets
843 		 * all its pending requests completed. The following
844 		 * instructions perform this delayed decrement, if
845 		 * needed. See the comments on
846 		 * num_groups_with_pending_reqs for details.
847 		 */
848 		if (entity->in_groups_with_pending_reqs) {
849 			entity->in_groups_with_pending_reqs = false;
850 			bfqd->num_groups_with_pending_reqs--;
851 		}
852 	}
853 
854 	/*
855 	 * Next function is invoked last, because it causes bfqq to be
856 	 * freed if the following holds: bfqq is not in service and
857 	 * has no dispatched request. DO NOT use bfqq after the next
858 	 * function invocation.
859 	 */
860 	__bfq_weights_tree_remove(bfqd, bfqq,
861 				  &bfqd->queue_weights_tree);
862 }
863 
864 /*
865  * Return expired entry, or NULL to just start from scratch in rbtree.
866  */
867 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
868 				      struct request *last)
869 {
870 	struct request *rq;
871 
872 	if (bfq_bfqq_fifo_expire(bfqq))
873 		return NULL;
874 
875 	bfq_mark_bfqq_fifo_expire(bfqq);
876 
877 	rq = rq_entry_fifo(bfqq->fifo.next);
878 
879 	if (rq == last || ktime_get_ns() < rq->fifo_time)
880 		return NULL;
881 
882 	bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
883 	return rq;
884 }
885 
886 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
887 					struct bfq_queue *bfqq,
888 					struct request *last)
889 {
890 	struct rb_node *rbnext = rb_next(&last->rb_node);
891 	struct rb_node *rbprev = rb_prev(&last->rb_node);
892 	struct request *next, *prev = NULL;
893 
894 	/* Follow expired path, else get first next available. */
895 	next = bfq_check_fifo(bfqq, last);
896 	if (next)
897 		return next;
898 
899 	if (rbprev)
900 		prev = rb_entry_rq(rbprev);
901 
902 	if (rbnext)
903 		next = rb_entry_rq(rbnext);
904 	else {
905 		rbnext = rb_first(&bfqq->sort_list);
906 		if (rbnext && rbnext != &last->rb_node)
907 			next = rb_entry_rq(rbnext);
908 	}
909 
910 	return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
911 }
912 
913 /* see the definition of bfq_async_charge_factor for details */
914 static unsigned long bfq_serv_to_charge(struct request *rq,
915 					struct bfq_queue *bfqq)
916 {
917 	if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
918 	    bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
919 		return blk_rq_sectors(rq);
920 
921 	return blk_rq_sectors(rq) * bfq_async_charge_factor;
922 }
923 
924 /**
925  * bfq_updated_next_req - update the queue after a new next_rq selection.
926  * @bfqd: the device data the queue belongs to.
927  * @bfqq: the queue to update.
928  *
929  * If the first request of a queue changes we make sure that the queue
930  * has enough budget to serve at least its first request (if the
931  * request has grown).  We do this because if the queue has not enough
932  * budget for its first request, it has to go through two dispatch
933  * rounds to actually get it dispatched.
934  */
935 static void bfq_updated_next_req(struct bfq_data *bfqd,
936 				 struct bfq_queue *bfqq)
937 {
938 	struct bfq_entity *entity = &bfqq->entity;
939 	struct request *next_rq = bfqq->next_rq;
940 	unsigned long new_budget;
941 
942 	if (!next_rq)
943 		return;
944 
945 	if (bfqq == bfqd->in_service_queue)
946 		/*
947 		 * In order not to break guarantees, budgets cannot be
948 		 * changed after an entity has been selected.
949 		 */
950 		return;
951 
952 	new_budget = max_t(unsigned long,
953 			   max_t(unsigned long, bfqq->max_budget,
954 				 bfq_serv_to_charge(next_rq, bfqq)),
955 			   entity->service);
956 	if (entity->budget != new_budget) {
957 		entity->budget = new_budget;
958 		bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
959 					 new_budget);
960 		bfq_requeue_bfqq(bfqd, bfqq, false);
961 	}
962 }
963 
964 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
965 {
966 	u64 dur;
967 
968 	if (bfqd->bfq_wr_max_time > 0)
969 		return bfqd->bfq_wr_max_time;
970 
971 	dur = bfqd->rate_dur_prod;
972 	do_div(dur, bfqd->peak_rate);
973 
974 	/*
975 	 * Limit duration between 3 and 25 seconds. The upper limit
976 	 * has been conservatively set after the following worst case:
977 	 * on a QEMU/KVM virtual machine
978 	 * - running in a slow PC
979 	 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
980 	 * - serving a heavy I/O workload, such as the sequential reading
981 	 *   of several files
982 	 * mplayer took 23 seconds to start, if constantly weight-raised.
983 	 *
984 	 * As for higher values than that accommodating the above bad
985 	 * scenario, tests show that higher values would often yield
986 	 * the opposite of the desired result, i.e., would worsen
987 	 * responsiveness by allowing non-interactive applications to
988 	 * preserve weight raising for too long.
989 	 *
990 	 * On the other end, lower values than 3 seconds make it
991 	 * difficult for most interactive tasks to complete their jobs
992 	 * before weight-raising finishes.
993 	 */
994 	return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
995 }
996 
997 /* switch back from soft real-time to interactive weight raising */
998 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
999 					  struct bfq_data *bfqd)
1000 {
1001 	bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1002 	bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1003 	bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
1004 }
1005 
1006 static void
1007 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
1008 		      struct bfq_io_cq *bic, bool bfq_already_existing)
1009 {
1010 	unsigned int old_wr_coeff = bfqq->wr_coeff;
1011 	bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
1012 
1013 	if (bic->saved_has_short_ttime)
1014 		bfq_mark_bfqq_has_short_ttime(bfqq);
1015 	else
1016 		bfq_clear_bfqq_has_short_ttime(bfqq);
1017 
1018 	if (bic->saved_IO_bound)
1019 		bfq_mark_bfqq_IO_bound(bfqq);
1020 	else
1021 		bfq_clear_bfqq_IO_bound(bfqq);
1022 
1023 	bfqq->entity.new_weight = bic->saved_weight;
1024 	bfqq->ttime = bic->saved_ttime;
1025 	bfqq->wr_coeff = bic->saved_wr_coeff;
1026 	bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
1027 	bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
1028 	bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
1029 
1030 	if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1031 	    time_is_before_jiffies(bfqq->last_wr_start_finish +
1032 				   bfqq->wr_cur_max_time))) {
1033 		if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1034 		    !bfq_bfqq_in_large_burst(bfqq) &&
1035 		    time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1036 					     bfq_wr_duration(bfqd))) {
1037 			switch_back_to_interactive_wr(bfqq, bfqd);
1038 		} else {
1039 			bfqq->wr_coeff = 1;
1040 			bfq_log_bfqq(bfqq->bfqd, bfqq,
1041 				     "resume state: switching off wr");
1042 		}
1043 	}
1044 
1045 	/* make sure weight will be updated, however we got here */
1046 	bfqq->entity.prio_changed = 1;
1047 
1048 	if (likely(!busy))
1049 		return;
1050 
1051 	if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1052 		bfqd->wr_busy_queues++;
1053 	else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1054 		bfqd->wr_busy_queues--;
1055 }
1056 
1057 static int bfqq_process_refs(struct bfq_queue *bfqq)
1058 {
1059 	return bfqq->ref - bfqq->allocated - bfqq->entity.on_st -
1060 		(bfqq->weight_counter != NULL);
1061 }
1062 
1063 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1064 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1065 {
1066 	struct bfq_queue *item;
1067 	struct hlist_node *n;
1068 
1069 	hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1070 		hlist_del_init(&item->burst_list_node);
1071 
1072 	/*
1073 	 * Start the creation of a new burst list only if there is no
1074 	 * active queue. See comments on the conditional invocation of
1075 	 * bfq_handle_burst().
1076 	 */
1077 	if (bfq_tot_busy_queues(bfqd) == 0) {
1078 		hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1079 		bfqd->burst_size = 1;
1080 	} else
1081 		bfqd->burst_size = 0;
1082 
1083 	bfqd->burst_parent_entity = bfqq->entity.parent;
1084 }
1085 
1086 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1087 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1088 {
1089 	/* Increment burst size to take into account also bfqq */
1090 	bfqd->burst_size++;
1091 
1092 	if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1093 		struct bfq_queue *pos, *bfqq_item;
1094 		struct hlist_node *n;
1095 
1096 		/*
1097 		 * Enough queues have been activated shortly after each
1098 		 * other to consider this burst as large.
1099 		 */
1100 		bfqd->large_burst = true;
1101 
1102 		/*
1103 		 * We can now mark all queues in the burst list as
1104 		 * belonging to a large burst.
1105 		 */
1106 		hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1107 				     burst_list_node)
1108 			bfq_mark_bfqq_in_large_burst(bfqq_item);
1109 		bfq_mark_bfqq_in_large_burst(bfqq);
1110 
1111 		/*
1112 		 * From now on, and until the current burst finishes, any
1113 		 * new queue being activated shortly after the last queue
1114 		 * was inserted in the burst can be immediately marked as
1115 		 * belonging to a large burst. So the burst list is not
1116 		 * needed any more. Remove it.
1117 		 */
1118 		hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1119 					  burst_list_node)
1120 			hlist_del_init(&pos->burst_list_node);
1121 	} else /*
1122 		* Burst not yet large: add bfqq to the burst list. Do
1123 		* not increment the ref counter for bfqq, because bfqq
1124 		* is removed from the burst list before freeing bfqq
1125 		* in put_queue.
1126 		*/
1127 		hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1128 }
1129 
1130 /*
1131  * If many queues belonging to the same group happen to be created
1132  * shortly after each other, then the processes associated with these
1133  * queues have typically a common goal. In particular, bursts of queue
1134  * creations are usually caused by services or applications that spawn
1135  * many parallel threads/processes. Examples are systemd during boot,
1136  * or git grep. To help these processes get their job done as soon as
1137  * possible, it is usually better to not grant either weight-raising
1138  * or device idling to their queues, unless these queues must be
1139  * protected from the I/O flowing through other active queues.
1140  *
1141  * In this comment we describe, firstly, the reasons why this fact
1142  * holds, and, secondly, the next function, which implements the main
1143  * steps needed to properly mark these queues so that they can then be
1144  * treated in a different way.
1145  *
1146  * The above services or applications benefit mostly from a high
1147  * throughput: the quicker the requests of the activated queues are
1148  * cumulatively served, the sooner the target job of these queues gets
1149  * completed. As a consequence, weight-raising any of these queues,
1150  * which also implies idling the device for it, is almost always
1151  * counterproductive, unless there are other active queues to isolate
1152  * these new queues from. If there no other active queues, then
1153  * weight-raising these new queues just lowers throughput in most
1154  * cases.
1155  *
1156  * On the other hand, a burst of queue creations may be caused also by
1157  * the start of an application that does not consist of a lot of
1158  * parallel I/O-bound threads. In fact, with a complex application,
1159  * several short processes may need to be executed to start-up the
1160  * application. In this respect, to start an application as quickly as
1161  * possible, the best thing to do is in any case to privilege the I/O
1162  * related to the application with respect to all other
1163  * I/O. Therefore, the best strategy to start as quickly as possible
1164  * an application that causes a burst of queue creations is to
1165  * weight-raise all the queues created during the burst. This is the
1166  * exact opposite of the best strategy for the other type of bursts.
1167  *
1168  * In the end, to take the best action for each of the two cases, the
1169  * two types of bursts need to be distinguished. Fortunately, this
1170  * seems relatively easy, by looking at the sizes of the bursts. In
1171  * particular, we found a threshold such that only bursts with a
1172  * larger size than that threshold are apparently caused by
1173  * services or commands such as systemd or git grep. For brevity,
1174  * hereafter we call just 'large' these bursts. BFQ *does not*
1175  * weight-raise queues whose creation occurs in a large burst. In
1176  * addition, for each of these queues BFQ performs or does not perform
1177  * idling depending on which choice boosts the throughput more. The
1178  * exact choice depends on the device and request pattern at
1179  * hand.
1180  *
1181  * Unfortunately, false positives may occur while an interactive task
1182  * is starting (e.g., an application is being started). The
1183  * consequence is that the queues associated with the task do not
1184  * enjoy weight raising as expected. Fortunately these false positives
1185  * are very rare. They typically occur if some service happens to
1186  * start doing I/O exactly when the interactive task starts.
1187  *
1188  * Turning back to the next function, it is invoked only if there are
1189  * no active queues (apart from active queues that would belong to the
1190  * same, possible burst bfqq would belong to), and it implements all
1191  * the steps needed to detect the occurrence of a large burst and to
1192  * properly mark all the queues belonging to it (so that they can then
1193  * be treated in a different way). This goal is achieved by
1194  * maintaining a "burst list" that holds, temporarily, the queues that
1195  * belong to the burst in progress. The list is then used to mark
1196  * these queues as belonging to a large burst if the burst does become
1197  * large. The main steps are the following.
1198  *
1199  * . when the very first queue is created, the queue is inserted into the
1200  *   list (as it could be the first queue in a possible burst)
1201  *
1202  * . if the current burst has not yet become large, and a queue Q that does
1203  *   not yet belong to the burst is activated shortly after the last time
1204  *   at which a new queue entered the burst list, then the function appends
1205  *   Q to the burst list
1206  *
1207  * . if, as a consequence of the previous step, the burst size reaches
1208  *   the large-burst threshold, then
1209  *
1210  *     . all the queues in the burst list are marked as belonging to a
1211  *       large burst
1212  *
1213  *     . the burst list is deleted; in fact, the burst list already served
1214  *       its purpose (keeping temporarily track of the queues in a burst,
1215  *       so as to be able to mark them as belonging to a large burst in the
1216  *       previous sub-step), and now is not needed any more
1217  *
1218  *     . the device enters a large-burst mode
1219  *
1220  * . if a queue Q that does not belong to the burst is created while
1221  *   the device is in large-burst mode and shortly after the last time
1222  *   at which a queue either entered the burst list or was marked as
1223  *   belonging to the current large burst, then Q is immediately marked
1224  *   as belonging to a large burst.
1225  *
1226  * . if a queue Q that does not belong to the burst is created a while
1227  *   later, i.e., not shortly after, than the last time at which a queue
1228  *   either entered the burst list or was marked as belonging to the
1229  *   current large burst, then the current burst is deemed as finished and:
1230  *
1231  *        . the large-burst mode is reset if set
1232  *
1233  *        . the burst list is emptied
1234  *
1235  *        . Q is inserted in the burst list, as Q may be the first queue
1236  *          in a possible new burst (then the burst list contains just Q
1237  *          after this step).
1238  */
1239 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1240 {
1241 	/*
1242 	 * If bfqq is already in the burst list or is part of a large
1243 	 * burst, or finally has just been split, then there is
1244 	 * nothing else to do.
1245 	 */
1246 	if (!hlist_unhashed(&bfqq->burst_list_node) ||
1247 	    bfq_bfqq_in_large_burst(bfqq) ||
1248 	    time_is_after_eq_jiffies(bfqq->split_time +
1249 				     msecs_to_jiffies(10)))
1250 		return;
1251 
1252 	/*
1253 	 * If bfqq's creation happens late enough, or bfqq belongs to
1254 	 * a different group than the burst group, then the current
1255 	 * burst is finished, and related data structures must be
1256 	 * reset.
1257 	 *
1258 	 * In this respect, consider the special case where bfqq is
1259 	 * the very first queue created after BFQ is selected for this
1260 	 * device. In this case, last_ins_in_burst and
1261 	 * burst_parent_entity are not yet significant when we get
1262 	 * here. But it is easy to verify that, whether or not the
1263 	 * following condition is true, bfqq will end up being
1264 	 * inserted into the burst list. In particular the list will
1265 	 * happen to contain only bfqq. And this is exactly what has
1266 	 * to happen, as bfqq may be the first queue of the first
1267 	 * burst.
1268 	 */
1269 	if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1270 	    bfqd->bfq_burst_interval) ||
1271 	    bfqq->entity.parent != bfqd->burst_parent_entity) {
1272 		bfqd->large_burst = false;
1273 		bfq_reset_burst_list(bfqd, bfqq);
1274 		goto end;
1275 	}
1276 
1277 	/*
1278 	 * If we get here, then bfqq is being activated shortly after the
1279 	 * last queue. So, if the current burst is also large, we can mark
1280 	 * bfqq as belonging to this large burst immediately.
1281 	 */
1282 	if (bfqd->large_burst) {
1283 		bfq_mark_bfqq_in_large_burst(bfqq);
1284 		goto end;
1285 	}
1286 
1287 	/*
1288 	 * If we get here, then a large-burst state has not yet been
1289 	 * reached, but bfqq is being activated shortly after the last
1290 	 * queue. Then we add bfqq to the burst.
1291 	 */
1292 	bfq_add_to_burst(bfqd, bfqq);
1293 end:
1294 	/*
1295 	 * At this point, bfqq either has been added to the current
1296 	 * burst or has caused the current burst to terminate and a
1297 	 * possible new burst to start. In particular, in the second
1298 	 * case, bfqq has become the first queue in the possible new
1299 	 * burst.  In both cases last_ins_in_burst needs to be moved
1300 	 * forward.
1301 	 */
1302 	bfqd->last_ins_in_burst = jiffies;
1303 }
1304 
1305 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1306 {
1307 	struct bfq_entity *entity = &bfqq->entity;
1308 
1309 	return entity->budget - entity->service;
1310 }
1311 
1312 /*
1313  * If enough samples have been computed, return the current max budget
1314  * stored in bfqd, which is dynamically updated according to the
1315  * estimated disk peak rate; otherwise return the default max budget
1316  */
1317 static int bfq_max_budget(struct bfq_data *bfqd)
1318 {
1319 	if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1320 		return bfq_default_max_budget;
1321 	else
1322 		return bfqd->bfq_max_budget;
1323 }
1324 
1325 /*
1326  * Return min budget, which is a fraction of the current or default
1327  * max budget (trying with 1/32)
1328  */
1329 static int bfq_min_budget(struct bfq_data *bfqd)
1330 {
1331 	if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1332 		return bfq_default_max_budget / 32;
1333 	else
1334 		return bfqd->bfq_max_budget / 32;
1335 }
1336 
1337 /*
1338  * The next function, invoked after the input queue bfqq switches from
1339  * idle to busy, updates the budget of bfqq. The function also tells
1340  * whether the in-service queue should be expired, by returning
1341  * true. The purpose of expiring the in-service queue is to give bfqq
1342  * the chance to possibly preempt the in-service queue, and the reason
1343  * for preempting the in-service queue is to achieve one of the two
1344  * goals below.
1345  *
1346  * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1347  * expired because it has remained idle. In particular, bfqq may have
1348  * expired for one of the following two reasons:
1349  *
1350  * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1351  *   and did not make it to issue a new request before its last
1352  *   request was served;
1353  *
1354  * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1355  *   a new request before the expiration of the idling-time.
1356  *
1357  * Even if bfqq has expired for one of the above reasons, the process
1358  * associated with the queue may be however issuing requests greedily,
1359  * and thus be sensitive to the bandwidth it receives (bfqq may have
1360  * remained idle for other reasons: CPU high load, bfqq not enjoying
1361  * idling, I/O throttling somewhere in the path from the process to
1362  * the I/O scheduler, ...). But if, after every expiration for one of
1363  * the above two reasons, bfqq has to wait for the service of at least
1364  * one full budget of another queue before being served again, then
1365  * bfqq is likely to get a much lower bandwidth or resource time than
1366  * its reserved ones. To address this issue, two countermeasures need
1367  * to be taken.
1368  *
1369  * First, the budget and the timestamps of bfqq need to be updated in
1370  * a special way on bfqq reactivation: they need to be updated as if
1371  * bfqq did not remain idle and did not expire. In fact, if they are
1372  * computed as if bfqq expired and remained idle until reactivation,
1373  * then the process associated with bfqq is treated as if, instead of
1374  * being greedy, it stopped issuing requests when bfqq remained idle,
1375  * and restarts issuing requests only on this reactivation. In other
1376  * words, the scheduler does not help the process recover the "service
1377  * hole" between bfqq expiration and reactivation. As a consequence,
1378  * the process receives a lower bandwidth than its reserved one. In
1379  * contrast, to recover this hole, the budget must be updated as if
1380  * bfqq was not expired at all before this reactivation, i.e., it must
1381  * be set to the value of the remaining budget when bfqq was
1382  * expired. Along the same line, timestamps need to be assigned the
1383  * value they had the last time bfqq was selected for service, i.e.,
1384  * before last expiration. Thus timestamps need to be back-shifted
1385  * with respect to their normal computation (see [1] for more details
1386  * on this tricky aspect).
1387  *
1388  * Secondly, to allow the process to recover the hole, the in-service
1389  * queue must be expired too, to give bfqq the chance to preempt it
1390  * immediately. In fact, if bfqq has to wait for a full budget of the
1391  * in-service queue to be completed, then it may become impossible to
1392  * let the process recover the hole, even if the back-shifted
1393  * timestamps of bfqq are lower than those of the in-service queue. If
1394  * this happens for most or all of the holes, then the process may not
1395  * receive its reserved bandwidth. In this respect, it is worth noting
1396  * that, being the service of outstanding requests unpreemptible, a
1397  * little fraction of the holes may however be unrecoverable, thereby
1398  * causing a little loss of bandwidth.
1399  *
1400  * The last important point is detecting whether bfqq does need this
1401  * bandwidth recovery. In this respect, the next function deems the
1402  * process associated with bfqq greedy, and thus allows it to recover
1403  * the hole, if: 1) the process is waiting for the arrival of a new
1404  * request (which implies that bfqq expired for one of the above two
1405  * reasons), and 2) such a request has arrived soon. The first
1406  * condition is controlled through the flag non_blocking_wait_rq,
1407  * while the second through the flag arrived_in_time. If both
1408  * conditions hold, then the function computes the budget in the
1409  * above-described special way, and signals that the in-service queue
1410  * should be expired. Timestamp back-shifting is done later in
1411  * __bfq_activate_entity.
1412  *
1413  * 2. Reduce latency. Even if timestamps are not backshifted to let
1414  * the process associated with bfqq recover a service hole, bfqq may
1415  * however happen to have, after being (re)activated, a lower finish
1416  * timestamp than the in-service queue.	 That is, the next budget of
1417  * bfqq may have to be completed before the one of the in-service
1418  * queue. If this is the case, then preempting the in-service queue
1419  * allows this goal to be achieved, apart from the unpreemptible,
1420  * outstanding requests mentioned above.
1421  *
1422  * Unfortunately, regardless of which of the above two goals one wants
1423  * to achieve, service trees need first to be updated to know whether
1424  * the in-service queue must be preempted. To have service trees
1425  * correctly updated, the in-service queue must be expired and
1426  * rescheduled, and bfqq must be scheduled too. This is one of the
1427  * most costly operations (in future versions, the scheduling
1428  * mechanism may be re-designed in such a way to make it possible to
1429  * know whether preemption is needed without needing to update service
1430  * trees). In addition, queue preemptions almost always cause random
1431  * I/O, which may in turn cause loss of throughput. Finally, there may
1432  * even be no in-service queue when the next function is invoked (so,
1433  * no queue to compare timestamps with). Because of these facts, the
1434  * next function adopts the following simple scheme to avoid costly
1435  * operations, too frequent preemptions and too many dependencies on
1436  * the state of the scheduler: it requests the expiration of the
1437  * in-service queue (unconditionally) only for queues that need to
1438  * recover a hole. Then it delegates to other parts of the code the
1439  * responsibility of handling the above case 2.
1440  */
1441 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1442 						struct bfq_queue *bfqq,
1443 						bool arrived_in_time)
1444 {
1445 	struct bfq_entity *entity = &bfqq->entity;
1446 
1447 	/*
1448 	 * In the next compound condition, we check also whether there
1449 	 * is some budget left, because otherwise there is no point in
1450 	 * trying to go on serving bfqq with this same budget: bfqq
1451 	 * would be expired immediately after being selected for
1452 	 * service. This would only cause useless overhead.
1453 	 */
1454 	if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1455 	    bfq_bfqq_budget_left(bfqq) > 0) {
1456 		/*
1457 		 * We do not clear the flag non_blocking_wait_rq here, as
1458 		 * the latter is used in bfq_activate_bfqq to signal
1459 		 * that timestamps need to be back-shifted (and is
1460 		 * cleared right after).
1461 		 */
1462 
1463 		/*
1464 		 * In next assignment we rely on that either
1465 		 * entity->service or entity->budget are not updated
1466 		 * on expiration if bfqq is empty (see
1467 		 * __bfq_bfqq_recalc_budget). Thus both quantities
1468 		 * remain unchanged after such an expiration, and the
1469 		 * following statement therefore assigns to
1470 		 * entity->budget the remaining budget on such an
1471 		 * expiration.
1472 		 */
1473 		entity->budget = min_t(unsigned long,
1474 				       bfq_bfqq_budget_left(bfqq),
1475 				       bfqq->max_budget);
1476 
1477 		/*
1478 		 * At this point, we have used entity->service to get
1479 		 * the budget left (needed for updating
1480 		 * entity->budget). Thus we finally can, and have to,
1481 		 * reset entity->service. The latter must be reset
1482 		 * because bfqq would otherwise be charged again for
1483 		 * the service it has received during its previous
1484 		 * service slot(s).
1485 		 */
1486 		entity->service = 0;
1487 
1488 		return true;
1489 	}
1490 
1491 	/*
1492 	 * We can finally complete expiration, by setting service to 0.
1493 	 */
1494 	entity->service = 0;
1495 	entity->budget = max_t(unsigned long, bfqq->max_budget,
1496 			       bfq_serv_to_charge(bfqq->next_rq, bfqq));
1497 	bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1498 	return false;
1499 }
1500 
1501 /*
1502  * Return the farthest past time instant according to jiffies
1503  * macros.
1504  */
1505 static unsigned long bfq_smallest_from_now(void)
1506 {
1507 	return jiffies - MAX_JIFFY_OFFSET;
1508 }
1509 
1510 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1511 					     struct bfq_queue *bfqq,
1512 					     unsigned int old_wr_coeff,
1513 					     bool wr_or_deserves_wr,
1514 					     bool interactive,
1515 					     bool in_burst,
1516 					     bool soft_rt)
1517 {
1518 	if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1519 		/* start a weight-raising period */
1520 		if (interactive) {
1521 			bfqq->service_from_wr = 0;
1522 			bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1523 			bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1524 		} else {
1525 			/*
1526 			 * No interactive weight raising in progress
1527 			 * here: assign minus infinity to
1528 			 * wr_start_at_switch_to_srt, to make sure
1529 			 * that, at the end of the soft-real-time
1530 			 * weight raising periods that is starting
1531 			 * now, no interactive weight-raising period
1532 			 * may be wrongly considered as still in
1533 			 * progress (and thus actually started by
1534 			 * mistake).
1535 			 */
1536 			bfqq->wr_start_at_switch_to_srt =
1537 				bfq_smallest_from_now();
1538 			bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1539 				BFQ_SOFTRT_WEIGHT_FACTOR;
1540 			bfqq->wr_cur_max_time =
1541 				bfqd->bfq_wr_rt_max_time;
1542 		}
1543 
1544 		/*
1545 		 * If needed, further reduce budget to make sure it is
1546 		 * close to bfqq's backlog, so as to reduce the
1547 		 * scheduling-error component due to a too large
1548 		 * budget. Do not care about throughput consequences,
1549 		 * but only about latency. Finally, do not assign a
1550 		 * too small budget either, to avoid increasing
1551 		 * latency by causing too frequent expirations.
1552 		 */
1553 		bfqq->entity.budget = min_t(unsigned long,
1554 					    bfqq->entity.budget,
1555 					    2 * bfq_min_budget(bfqd));
1556 	} else if (old_wr_coeff > 1) {
1557 		if (interactive) { /* update wr coeff and duration */
1558 			bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1559 			bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1560 		} else if (in_burst)
1561 			bfqq->wr_coeff = 1;
1562 		else if (soft_rt) {
1563 			/*
1564 			 * The application is now or still meeting the
1565 			 * requirements for being deemed soft rt.  We
1566 			 * can then correctly and safely (re)charge
1567 			 * the weight-raising duration for the
1568 			 * application with the weight-raising
1569 			 * duration for soft rt applications.
1570 			 *
1571 			 * In particular, doing this recharge now, i.e.,
1572 			 * before the weight-raising period for the
1573 			 * application finishes, reduces the probability
1574 			 * of the following negative scenario:
1575 			 * 1) the weight of a soft rt application is
1576 			 *    raised at startup (as for any newly
1577 			 *    created application),
1578 			 * 2) since the application is not interactive,
1579 			 *    at a certain time weight-raising is
1580 			 *    stopped for the application,
1581 			 * 3) at that time the application happens to
1582 			 *    still have pending requests, and hence
1583 			 *    is destined to not have a chance to be
1584 			 *    deemed soft rt before these requests are
1585 			 *    completed (see the comments to the
1586 			 *    function bfq_bfqq_softrt_next_start()
1587 			 *    for details on soft rt detection),
1588 			 * 4) these pending requests experience a high
1589 			 *    latency because the application is not
1590 			 *    weight-raised while they are pending.
1591 			 */
1592 			if (bfqq->wr_cur_max_time !=
1593 				bfqd->bfq_wr_rt_max_time) {
1594 				bfqq->wr_start_at_switch_to_srt =
1595 					bfqq->last_wr_start_finish;
1596 
1597 				bfqq->wr_cur_max_time =
1598 					bfqd->bfq_wr_rt_max_time;
1599 				bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1600 					BFQ_SOFTRT_WEIGHT_FACTOR;
1601 			}
1602 			bfqq->last_wr_start_finish = jiffies;
1603 		}
1604 	}
1605 }
1606 
1607 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1608 					struct bfq_queue *bfqq)
1609 {
1610 	return bfqq->dispatched == 0 &&
1611 		time_is_before_jiffies(
1612 			bfqq->budget_timeout +
1613 			bfqd->bfq_wr_min_idle_time);
1614 }
1615 
1616 
1617 /*
1618  * Return true if bfqq is in a higher priority class, or has a higher
1619  * weight than the in-service queue.
1620  */
1621 static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq,
1622 					    struct bfq_queue *in_serv_bfqq)
1623 {
1624 	int bfqq_weight, in_serv_weight;
1625 
1626 	if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class)
1627 		return true;
1628 
1629 	if (in_serv_bfqq->entity.parent == bfqq->entity.parent) {
1630 		bfqq_weight = bfqq->entity.weight;
1631 		in_serv_weight = in_serv_bfqq->entity.weight;
1632 	} else {
1633 		if (bfqq->entity.parent)
1634 			bfqq_weight = bfqq->entity.parent->weight;
1635 		else
1636 			bfqq_weight = bfqq->entity.weight;
1637 		if (in_serv_bfqq->entity.parent)
1638 			in_serv_weight = in_serv_bfqq->entity.parent->weight;
1639 		else
1640 			in_serv_weight = in_serv_bfqq->entity.weight;
1641 	}
1642 
1643 	return bfqq_weight > in_serv_weight;
1644 }
1645 
1646 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1647 					     struct bfq_queue *bfqq,
1648 					     int old_wr_coeff,
1649 					     struct request *rq,
1650 					     bool *interactive)
1651 {
1652 	bool soft_rt, in_burst,	wr_or_deserves_wr,
1653 		bfqq_wants_to_preempt,
1654 		idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1655 		/*
1656 		 * See the comments on
1657 		 * bfq_bfqq_update_budg_for_activation for
1658 		 * details on the usage of the next variable.
1659 		 */
1660 		arrived_in_time =  ktime_get_ns() <=
1661 			bfqq->ttime.last_end_request +
1662 			bfqd->bfq_slice_idle * 3;
1663 
1664 
1665 	/*
1666 	 * bfqq deserves to be weight-raised if:
1667 	 * - it is sync,
1668 	 * - it does not belong to a large burst,
1669 	 * - it has been idle for enough time or is soft real-time,
1670 	 * - is linked to a bfq_io_cq (it is not shared in any sense).
1671 	 */
1672 	in_burst = bfq_bfqq_in_large_burst(bfqq);
1673 	soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1674 		!BFQQ_TOTALLY_SEEKY(bfqq) &&
1675 		!in_burst &&
1676 		time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1677 		bfqq->dispatched == 0;
1678 	*interactive = !in_burst && idle_for_long_time;
1679 	wr_or_deserves_wr = bfqd->low_latency &&
1680 		(bfqq->wr_coeff > 1 ||
1681 		 (bfq_bfqq_sync(bfqq) &&
1682 		  bfqq->bic && (*interactive || soft_rt)));
1683 
1684 	/*
1685 	 * Using the last flag, update budget and check whether bfqq
1686 	 * may want to preempt the in-service queue.
1687 	 */
1688 	bfqq_wants_to_preempt =
1689 		bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1690 						    arrived_in_time);
1691 
1692 	/*
1693 	 * If bfqq happened to be activated in a burst, but has been
1694 	 * idle for much more than an interactive queue, then we
1695 	 * assume that, in the overall I/O initiated in the burst, the
1696 	 * I/O associated with bfqq is finished. So bfqq does not need
1697 	 * to be treated as a queue belonging to a burst
1698 	 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1699 	 * if set, and remove bfqq from the burst list if it's
1700 	 * there. We do not decrement burst_size, because the fact
1701 	 * that bfqq does not need to belong to the burst list any
1702 	 * more does not invalidate the fact that bfqq was created in
1703 	 * a burst.
1704 	 */
1705 	if (likely(!bfq_bfqq_just_created(bfqq)) &&
1706 	    idle_for_long_time &&
1707 	    time_is_before_jiffies(
1708 		    bfqq->budget_timeout +
1709 		    msecs_to_jiffies(10000))) {
1710 		hlist_del_init(&bfqq->burst_list_node);
1711 		bfq_clear_bfqq_in_large_burst(bfqq);
1712 	}
1713 
1714 	bfq_clear_bfqq_just_created(bfqq);
1715 
1716 
1717 	if (!bfq_bfqq_IO_bound(bfqq)) {
1718 		if (arrived_in_time) {
1719 			bfqq->requests_within_timer++;
1720 			if (bfqq->requests_within_timer >=
1721 			    bfqd->bfq_requests_within_timer)
1722 				bfq_mark_bfqq_IO_bound(bfqq);
1723 		} else
1724 			bfqq->requests_within_timer = 0;
1725 	}
1726 
1727 	if (bfqd->low_latency) {
1728 		if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1729 			/* wraparound */
1730 			bfqq->split_time =
1731 				jiffies - bfqd->bfq_wr_min_idle_time - 1;
1732 
1733 		if (time_is_before_jiffies(bfqq->split_time +
1734 					   bfqd->bfq_wr_min_idle_time)) {
1735 			bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1736 							 old_wr_coeff,
1737 							 wr_or_deserves_wr,
1738 							 *interactive,
1739 							 in_burst,
1740 							 soft_rt);
1741 
1742 			if (old_wr_coeff != bfqq->wr_coeff)
1743 				bfqq->entity.prio_changed = 1;
1744 		}
1745 	}
1746 
1747 	bfqq->last_idle_bklogged = jiffies;
1748 	bfqq->service_from_backlogged = 0;
1749 	bfq_clear_bfqq_softrt_update(bfqq);
1750 
1751 	bfq_add_bfqq_busy(bfqd, bfqq);
1752 
1753 	/*
1754 	 * Expire in-service queue only if preemption may be needed
1755 	 * for guarantees. In particular, we care only about two
1756 	 * cases. The first is that bfqq has to recover a service
1757 	 * hole, as explained in the comments on
1758 	 * bfq_bfqq_update_budg_for_activation(), i.e., that
1759 	 * bfqq_wants_to_preempt is true. However, if bfqq does not
1760 	 * carry time-critical I/O, then bfqq's bandwidth is less
1761 	 * important than that of queues that carry time-critical I/O.
1762 	 * So, as a further constraint, we consider this case only if
1763 	 * bfqq is at least as weight-raised, i.e., at least as time
1764 	 * critical, as the in-service queue.
1765 	 *
1766 	 * The second case is that bfqq is in a higher priority class,
1767 	 * or has a higher weight than the in-service queue. If this
1768 	 * condition does not hold, we don't care because, even if
1769 	 * bfqq does not start to be served immediately, the resulting
1770 	 * delay for bfqq's I/O is however lower or much lower than
1771 	 * the ideal completion time to be guaranteed to bfqq's I/O.
1772 	 *
1773 	 * In both cases, preemption is needed only if, according to
1774 	 * the timestamps of both bfqq and of the in-service queue,
1775 	 * bfqq actually is the next queue to serve. So, to reduce
1776 	 * useless preemptions, the return value of
1777 	 * next_queue_may_preempt() is considered in the next compound
1778 	 * condition too. Yet next_queue_may_preempt() just checks a
1779 	 * simple, necessary condition for bfqq to be the next queue
1780 	 * to serve. In fact, to evaluate a sufficient condition, the
1781 	 * timestamps of the in-service queue would need to be
1782 	 * updated, and this operation is quite costly (see the
1783 	 * comments on bfq_bfqq_update_budg_for_activation()).
1784 	 */
1785 	if (bfqd->in_service_queue &&
1786 	    ((bfqq_wants_to_preempt &&
1787 	      bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) ||
1788 	     bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue)) &&
1789 	    next_queue_may_preempt(bfqd))
1790 		bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1791 				false, BFQQE_PREEMPTED);
1792 }
1793 
1794 static void bfq_reset_inject_limit(struct bfq_data *bfqd,
1795 				   struct bfq_queue *bfqq)
1796 {
1797 	/* invalidate baseline total service time */
1798 	bfqq->last_serv_time_ns = 0;
1799 
1800 	/*
1801 	 * Reset pointer in case we are waiting for
1802 	 * some request completion.
1803 	 */
1804 	bfqd->waited_rq = NULL;
1805 
1806 	/*
1807 	 * If bfqq has a short think time, then start by setting the
1808 	 * inject limit to 0 prudentially, because the service time of
1809 	 * an injected I/O request may be higher than the think time
1810 	 * of bfqq, and therefore, if one request was injected when
1811 	 * bfqq remains empty, this injected request might delay the
1812 	 * service of the next I/O request for bfqq significantly. In
1813 	 * case bfqq can actually tolerate some injection, then the
1814 	 * adaptive update will however raise the limit soon. This
1815 	 * lucky circumstance holds exactly because bfqq has a short
1816 	 * think time, and thus, after remaining empty, is likely to
1817 	 * get new I/O enqueued---and then completed---before being
1818 	 * expired. This is the very pattern that gives the
1819 	 * limit-update algorithm the chance to measure the effect of
1820 	 * injection on request service times, and then to update the
1821 	 * limit accordingly.
1822 	 *
1823 	 * However, in the following special case, the inject limit is
1824 	 * left to 1 even if the think time is short: bfqq's I/O is
1825 	 * synchronized with that of some other queue, i.e., bfqq may
1826 	 * receive new I/O only after the I/O of the other queue is
1827 	 * completed. Keeping the inject limit to 1 allows the
1828 	 * blocking I/O to be served while bfqq is in service. And
1829 	 * this is very convenient both for bfqq and for overall
1830 	 * throughput, as explained in detail in the comments in
1831 	 * bfq_update_has_short_ttime().
1832 	 *
1833 	 * On the opposite end, if bfqq has a long think time, then
1834 	 * start directly by 1, because:
1835 	 * a) on the bright side, keeping at most one request in
1836 	 * service in the drive is unlikely to cause any harm to the
1837 	 * latency of bfqq's requests, as the service time of a single
1838 	 * request is likely to be lower than the think time of bfqq;
1839 	 * b) on the downside, after becoming empty, bfqq is likely to
1840 	 * expire before getting its next request. With this request
1841 	 * arrival pattern, it is very hard to sample total service
1842 	 * times and update the inject limit accordingly (see comments
1843 	 * on bfq_update_inject_limit()). So the limit is likely to be
1844 	 * never, or at least seldom, updated.  As a consequence, by
1845 	 * setting the limit to 1, we avoid that no injection ever
1846 	 * occurs with bfqq. On the downside, this proactive step
1847 	 * further reduces chances to actually compute the baseline
1848 	 * total service time. Thus it reduces chances to execute the
1849 	 * limit-update algorithm and possibly raise the limit to more
1850 	 * than 1.
1851 	 */
1852 	if (bfq_bfqq_has_short_ttime(bfqq))
1853 		bfqq->inject_limit = 0;
1854 	else
1855 		bfqq->inject_limit = 1;
1856 
1857 	bfqq->decrease_time_jif = jiffies;
1858 }
1859 
1860 static void bfq_add_request(struct request *rq)
1861 {
1862 	struct bfq_queue *bfqq = RQ_BFQQ(rq);
1863 	struct bfq_data *bfqd = bfqq->bfqd;
1864 	struct request *next_rq, *prev;
1865 	unsigned int old_wr_coeff = bfqq->wr_coeff;
1866 	bool interactive = false;
1867 
1868 	bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1869 	bfqq->queued[rq_is_sync(rq)]++;
1870 	bfqd->queued++;
1871 
1872 	if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_bfqq_sync(bfqq)) {
1873 		/*
1874 		 * Detect whether bfqq's I/O seems synchronized with
1875 		 * that of some other queue, i.e., whether bfqq, after
1876 		 * remaining empty, happens to receive new I/O only
1877 		 * right after some I/O request of the other queue has
1878 		 * been completed. We call waker queue the other
1879 		 * queue, and we assume, for simplicity, that bfqq may
1880 		 * have at most one waker queue.
1881 		 *
1882 		 * A remarkable throughput boost can be reached by
1883 		 * unconditionally injecting the I/O of the waker
1884 		 * queue, every time a new bfq_dispatch_request
1885 		 * happens to be invoked while I/O is being plugged
1886 		 * for bfqq.  In addition to boosting throughput, this
1887 		 * unblocks bfqq's I/O, thereby improving bandwidth
1888 		 * and latency for bfqq. Note that these same results
1889 		 * may be achieved with the general injection
1890 		 * mechanism, but less effectively. For details on
1891 		 * this aspect, see the comments on the choice of the
1892 		 * queue for injection in bfq_select_queue().
1893 		 *
1894 		 * Turning back to the detection of a waker queue, a
1895 		 * queue Q is deemed as a waker queue for bfqq if, for
1896 		 * two consecutive times, bfqq happens to become non
1897 		 * empty right after a request of Q has been
1898 		 * completed. In particular, on the first time, Q is
1899 		 * tentatively set as a candidate waker queue, while
1900 		 * on the second time, the flag
1901 		 * bfq_bfqq_has_waker(bfqq) is set to confirm that Q
1902 		 * is a waker queue for bfqq. These detection steps
1903 		 * are performed only if bfqq has a long think time,
1904 		 * so as to make it more likely that bfqq's I/O is
1905 		 * actually being blocked by a synchronization. This
1906 		 * last filter, plus the above two-times requirement,
1907 		 * make false positives less likely.
1908 		 *
1909 		 * NOTE
1910 		 *
1911 		 * The sooner a waker queue is detected, the sooner
1912 		 * throughput can be boosted by injecting I/O from the
1913 		 * waker queue. Fortunately, detection is likely to be
1914 		 * actually fast, for the following reasons. While
1915 		 * blocked by synchronization, bfqq has a long think
1916 		 * time. This implies that bfqq's inject limit is at
1917 		 * least equal to 1 (see the comments in
1918 		 * bfq_update_inject_limit()). So, thanks to
1919 		 * injection, the waker queue is likely to be served
1920 		 * during the very first I/O-plugging time interval
1921 		 * for bfqq. This triggers the first step of the
1922 		 * detection mechanism. Thanks again to injection, the
1923 		 * candidate waker queue is then likely to be
1924 		 * confirmed no later than during the next
1925 		 * I/O-plugging interval for bfqq.
1926 		 */
1927 		if (bfqd->last_completed_rq_bfqq &&
1928 		    !bfq_bfqq_has_short_ttime(bfqq) &&
1929 		    ktime_get_ns() - bfqd->last_completion <
1930 		    200 * NSEC_PER_USEC) {
1931 			if (bfqd->last_completed_rq_bfqq != bfqq &&
1932 			    bfqd->last_completed_rq_bfqq !=
1933 			    bfqq->waker_bfqq) {
1934 				/*
1935 				 * First synchronization detected with
1936 				 * a candidate waker queue, or with a
1937 				 * different candidate waker queue
1938 				 * from the current one.
1939 				 */
1940 				bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
1941 
1942 				/*
1943 				 * If the waker queue disappears, then
1944 				 * bfqq->waker_bfqq must be reset. To
1945 				 * this goal, we maintain in each
1946 				 * waker queue a list, woken_list, of
1947 				 * all the queues that reference the
1948 				 * waker queue through their
1949 				 * waker_bfqq pointer. When the waker
1950 				 * queue exits, the waker_bfqq pointer
1951 				 * of all the queues in the woken_list
1952 				 * is reset.
1953 				 *
1954 				 * In addition, if bfqq is already in
1955 				 * the woken_list of a waker queue,
1956 				 * then, before being inserted into
1957 				 * the woken_list of a new waker
1958 				 * queue, bfqq must be removed from
1959 				 * the woken_list of the old waker
1960 				 * queue.
1961 				 */
1962 				if (!hlist_unhashed(&bfqq->woken_list_node))
1963 					hlist_del_init(&bfqq->woken_list_node);
1964 				hlist_add_head(&bfqq->woken_list_node,
1965 				    &bfqd->last_completed_rq_bfqq->woken_list);
1966 
1967 				bfq_clear_bfqq_has_waker(bfqq);
1968 			} else if (bfqd->last_completed_rq_bfqq ==
1969 				   bfqq->waker_bfqq &&
1970 				   !bfq_bfqq_has_waker(bfqq)) {
1971 				/*
1972 				 * synchronization with waker_bfqq
1973 				 * seen for the second time
1974 				 */
1975 				bfq_mark_bfqq_has_waker(bfqq);
1976 			}
1977 		}
1978 
1979 		/*
1980 		 * Periodically reset inject limit, to make sure that
1981 		 * the latter eventually drops in case workload
1982 		 * changes, see step (3) in the comments on
1983 		 * bfq_update_inject_limit().
1984 		 */
1985 		if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
1986 					     msecs_to_jiffies(1000)))
1987 			bfq_reset_inject_limit(bfqd, bfqq);
1988 
1989 		/*
1990 		 * The following conditions must hold to setup a new
1991 		 * sampling of total service time, and then a new
1992 		 * update of the inject limit:
1993 		 * - bfqq is in service, because the total service
1994 		 *   time is evaluated only for the I/O requests of
1995 		 *   the queues in service;
1996 		 * - this is the right occasion to compute or to
1997 		 *   lower the baseline total service time, because
1998 		 *   there are actually no requests in the drive,
1999 		 *   or
2000 		 *   the baseline total service time is available, and
2001 		 *   this is the right occasion to compute the other
2002 		 *   quantity needed to update the inject limit, i.e.,
2003 		 *   the total service time caused by the amount of
2004 		 *   injection allowed by the current value of the
2005 		 *   limit. It is the right occasion because injection
2006 		 *   has actually been performed during the service
2007 		 *   hole, and there are still in-flight requests,
2008 		 *   which are very likely to be exactly the injected
2009 		 *   requests, or part of them;
2010 		 * - the minimum interval for sampling the total
2011 		 *   service time and updating the inject limit has
2012 		 *   elapsed.
2013 		 */
2014 		if (bfqq == bfqd->in_service_queue &&
2015 		    (bfqd->rq_in_driver == 0 ||
2016 		     (bfqq->last_serv_time_ns > 0 &&
2017 		      bfqd->rqs_injected && bfqd->rq_in_driver > 0)) &&
2018 		    time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2019 					      msecs_to_jiffies(10))) {
2020 			bfqd->last_empty_occupied_ns = ktime_get_ns();
2021 			/*
2022 			 * Start the state machine for measuring the
2023 			 * total service time of rq: setting
2024 			 * wait_dispatch will cause bfqd->waited_rq to
2025 			 * be set when rq will be dispatched.
2026 			 */
2027 			bfqd->wait_dispatch = true;
2028 			/*
2029 			 * If there is no I/O in service in the drive,
2030 			 * then possible injection occurred before the
2031 			 * arrival of rq will not affect the total
2032 			 * service time of rq. So the injection limit
2033 			 * must not be updated as a function of such
2034 			 * total service time, unless new injection
2035 			 * occurs before rq is completed. To have the
2036 			 * injection limit updated only in the latter
2037 			 * case, reset rqs_injected here (rqs_injected
2038 			 * will be set in case injection is performed
2039 			 * on bfqq before rq is completed).
2040 			 */
2041 			if (bfqd->rq_in_driver == 0)
2042 				bfqd->rqs_injected = false;
2043 		}
2044 	}
2045 
2046 	elv_rb_add(&bfqq->sort_list, rq);
2047 
2048 	/*
2049 	 * Check if this request is a better next-serve candidate.
2050 	 */
2051 	prev = bfqq->next_rq;
2052 	next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
2053 	bfqq->next_rq = next_rq;
2054 
2055 	/*
2056 	 * Adjust priority tree position, if next_rq changes.
2057 	 * See comments on bfq_pos_tree_add_move() for the unlikely().
2058 	 */
2059 	if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
2060 		bfq_pos_tree_add_move(bfqd, bfqq);
2061 
2062 	if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
2063 		bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
2064 						 rq, &interactive);
2065 	else {
2066 		if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
2067 		    time_is_before_jiffies(
2068 				bfqq->last_wr_start_finish +
2069 				bfqd->bfq_wr_min_inter_arr_async)) {
2070 			bfqq->wr_coeff = bfqd->bfq_wr_coeff;
2071 			bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
2072 
2073 			bfqd->wr_busy_queues++;
2074 			bfqq->entity.prio_changed = 1;
2075 		}
2076 		if (prev != bfqq->next_rq)
2077 			bfq_updated_next_req(bfqd, bfqq);
2078 	}
2079 
2080 	/*
2081 	 * Assign jiffies to last_wr_start_finish in the following
2082 	 * cases:
2083 	 *
2084 	 * . if bfqq is not going to be weight-raised, because, for
2085 	 *   non weight-raised queues, last_wr_start_finish stores the
2086 	 *   arrival time of the last request; as of now, this piece
2087 	 *   of information is used only for deciding whether to
2088 	 *   weight-raise async queues
2089 	 *
2090 	 * . if bfqq is not weight-raised, because, if bfqq is now
2091 	 *   switching to weight-raised, then last_wr_start_finish
2092 	 *   stores the time when weight-raising starts
2093 	 *
2094 	 * . if bfqq is interactive, because, regardless of whether
2095 	 *   bfqq is currently weight-raised, the weight-raising
2096 	 *   period must start or restart (this case is considered
2097 	 *   separately because it is not detected by the above
2098 	 *   conditions, if bfqq is already weight-raised)
2099 	 *
2100 	 * last_wr_start_finish has to be updated also if bfqq is soft
2101 	 * real-time, because the weight-raising period is constantly
2102 	 * restarted on idle-to-busy transitions for these queues, but
2103 	 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2104 	 * needed.
2105 	 */
2106 	if (bfqd->low_latency &&
2107 		(old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
2108 		bfqq->last_wr_start_finish = jiffies;
2109 }
2110 
2111 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
2112 					  struct bio *bio,
2113 					  struct request_queue *q)
2114 {
2115 	struct bfq_queue *bfqq = bfqd->bio_bfqq;
2116 
2117 
2118 	if (bfqq)
2119 		return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
2120 
2121 	return NULL;
2122 }
2123 
2124 static sector_t get_sdist(sector_t last_pos, struct request *rq)
2125 {
2126 	if (last_pos)
2127 		return abs(blk_rq_pos(rq) - last_pos);
2128 
2129 	return 0;
2130 }
2131 
2132 #if 0 /* Still not clear if we can do without next two functions */
2133 static void bfq_activate_request(struct request_queue *q, struct request *rq)
2134 {
2135 	struct bfq_data *bfqd = q->elevator->elevator_data;
2136 
2137 	bfqd->rq_in_driver++;
2138 }
2139 
2140 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
2141 {
2142 	struct bfq_data *bfqd = q->elevator->elevator_data;
2143 
2144 	bfqd->rq_in_driver--;
2145 }
2146 #endif
2147 
2148 static void bfq_remove_request(struct request_queue *q,
2149 			       struct request *rq)
2150 {
2151 	struct bfq_queue *bfqq = RQ_BFQQ(rq);
2152 	struct bfq_data *bfqd = bfqq->bfqd;
2153 	const int sync = rq_is_sync(rq);
2154 
2155 	if (bfqq->next_rq == rq) {
2156 		bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
2157 		bfq_updated_next_req(bfqd, bfqq);
2158 	}
2159 
2160 	if (rq->queuelist.prev != &rq->queuelist)
2161 		list_del_init(&rq->queuelist);
2162 	bfqq->queued[sync]--;
2163 	bfqd->queued--;
2164 	elv_rb_del(&bfqq->sort_list, rq);
2165 
2166 	elv_rqhash_del(q, rq);
2167 	if (q->last_merge == rq)
2168 		q->last_merge = NULL;
2169 
2170 	if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2171 		bfqq->next_rq = NULL;
2172 
2173 		if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
2174 			bfq_del_bfqq_busy(bfqd, bfqq, false);
2175 			/*
2176 			 * bfqq emptied. In normal operation, when
2177 			 * bfqq is empty, bfqq->entity.service and
2178 			 * bfqq->entity.budget must contain,
2179 			 * respectively, the service received and the
2180 			 * budget used last time bfqq emptied. These
2181 			 * facts do not hold in this case, as at least
2182 			 * this last removal occurred while bfqq is
2183 			 * not in service. To avoid inconsistencies,
2184 			 * reset both bfqq->entity.service and
2185 			 * bfqq->entity.budget, if bfqq has still a
2186 			 * process that may issue I/O requests to it.
2187 			 */
2188 			bfqq->entity.budget = bfqq->entity.service = 0;
2189 		}
2190 
2191 		/*
2192 		 * Remove queue from request-position tree as it is empty.
2193 		 */
2194 		if (bfqq->pos_root) {
2195 			rb_erase(&bfqq->pos_node, bfqq->pos_root);
2196 			bfqq->pos_root = NULL;
2197 		}
2198 	} else {
2199 		/* see comments on bfq_pos_tree_add_move() for the unlikely() */
2200 		if (unlikely(!bfqd->nonrot_with_queueing))
2201 			bfq_pos_tree_add_move(bfqd, bfqq);
2202 	}
2203 
2204 	if (rq->cmd_flags & REQ_META)
2205 		bfqq->meta_pending--;
2206 
2207 }
2208 
2209 static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio,
2210 		unsigned int nr_segs)
2211 {
2212 	struct request_queue *q = hctx->queue;
2213 	struct bfq_data *bfqd = q->elevator->elevator_data;
2214 	struct request *free = NULL;
2215 	/*
2216 	 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2217 	 * store its return value for later use, to avoid nesting
2218 	 * queue_lock inside the bfqd->lock. We assume that the bic
2219 	 * returned by bfq_bic_lookup does not go away before
2220 	 * bfqd->lock is taken.
2221 	 */
2222 	struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
2223 	bool ret;
2224 
2225 	spin_lock_irq(&bfqd->lock);
2226 
2227 	if (bic)
2228 		bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
2229 	else
2230 		bfqd->bio_bfqq = NULL;
2231 	bfqd->bio_bic = bic;
2232 
2233 	ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
2234 
2235 	if (free)
2236 		blk_mq_free_request(free);
2237 	spin_unlock_irq(&bfqd->lock);
2238 
2239 	return ret;
2240 }
2241 
2242 static int bfq_request_merge(struct request_queue *q, struct request **req,
2243 			     struct bio *bio)
2244 {
2245 	struct bfq_data *bfqd = q->elevator->elevator_data;
2246 	struct request *__rq;
2247 
2248 	__rq = bfq_find_rq_fmerge(bfqd, bio, q);
2249 	if (__rq && elv_bio_merge_ok(__rq, bio)) {
2250 		*req = __rq;
2251 		return ELEVATOR_FRONT_MERGE;
2252 	}
2253 
2254 	return ELEVATOR_NO_MERGE;
2255 }
2256 
2257 static struct bfq_queue *bfq_init_rq(struct request *rq);
2258 
2259 static void bfq_request_merged(struct request_queue *q, struct request *req,
2260 			       enum elv_merge type)
2261 {
2262 	if (type == ELEVATOR_FRONT_MERGE &&
2263 	    rb_prev(&req->rb_node) &&
2264 	    blk_rq_pos(req) <
2265 	    blk_rq_pos(container_of(rb_prev(&req->rb_node),
2266 				    struct request, rb_node))) {
2267 		struct bfq_queue *bfqq = bfq_init_rq(req);
2268 		struct bfq_data *bfqd;
2269 		struct request *prev, *next_rq;
2270 
2271 		if (!bfqq)
2272 			return;
2273 
2274 		bfqd = bfqq->bfqd;
2275 
2276 		/* Reposition request in its sort_list */
2277 		elv_rb_del(&bfqq->sort_list, req);
2278 		elv_rb_add(&bfqq->sort_list, req);
2279 
2280 		/* Choose next request to be served for bfqq */
2281 		prev = bfqq->next_rq;
2282 		next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
2283 					 bfqd->last_position);
2284 		bfqq->next_rq = next_rq;
2285 		/*
2286 		 * If next_rq changes, update both the queue's budget to
2287 		 * fit the new request and the queue's position in its
2288 		 * rq_pos_tree.
2289 		 */
2290 		if (prev != bfqq->next_rq) {
2291 			bfq_updated_next_req(bfqd, bfqq);
2292 			/*
2293 			 * See comments on bfq_pos_tree_add_move() for
2294 			 * the unlikely().
2295 			 */
2296 			if (unlikely(!bfqd->nonrot_with_queueing))
2297 				bfq_pos_tree_add_move(bfqd, bfqq);
2298 		}
2299 	}
2300 }
2301 
2302 /*
2303  * This function is called to notify the scheduler that the requests
2304  * rq and 'next' have been merged, with 'next' going away.  BFQ
2305  * exploits this hook to address the following issue: if 'next' has a
2306  * fifo_time lower that rq, then the fifo_time of rq must be set to
2307  * the value of 'next', to not forget the greater age of 'next'.
2308  *
2309  * NOTE: in this function we assume that rq is in a bfq_queue, basing
2310  * on that rq is picked from the hash table q->elevator->hash, which,
2311  * in its turn, is filled only with I/O requests present in
2312  * bfq_queues, while BFQ is in use for the request queue q. In fact,
2313  * the function that fills this hash table (elv_rqhash_add) is called
2314  * only by bfq_insert_request.
2315  */
2316 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
2317 				struct request *next)
2318 {
2319 	struct bfq_queue *bfqq = bfq_init_rq(rq),
2320 		*next_bfqq = bfq_init_rq(next);
2321 
2322 	if (!bfqq)
2323 		return;
2324 
2325 	/*
2326 	 * If next and rq belong to the same bfq_queue and next is older
2327 	 * than rq, then reposition rq in the fifo (by substituting next
2328 	 * with rq). Otherwise, if next and rq belong to different
2329 	 * bfq_queues, never reposition rq: in fact, we would have to
2330 	 * reposition it with respect to next's position in its own fifo,
2331 	 * which would most certainly be too expensive with respect to
2332 	 * the benefits.
2333 	 */
2334 	if (bfqq == next_bfqq &&
2335 	    !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2336 	    next->fifo_time < rq->fifo_time) {
2337 		list_del_init(&rq->queuelist);
2338 		list_replace_init(&next->queuelist, &rq->queuelist);
2339 		rq->fifo_time = next->fifo_time;
2340 	}
2341 
2342 	if (bfqq->next_rq == next)
2343 		bfqq->next_rq = rq;
2344 
2345 	bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
2346 }
2347 
2348 /* Must be called with bfqq != NULL */
2349 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
2350 {
2351 	if (bfq_bfqq_busy(bfqq))
2352 		bfqq->bfqd->wr_busy_queues--;
2353 	bfqq->wr_coeff = 1;
2354 	bfqq->wr_cur_max_time = 0;
2355 	bfqq->last_wr_start_finish = jiffies;
2356 	/*
2357 	 * Trigger a weight change on the next invocation of
2358 	 * __bfq_entity_update_weight_prio.
2359 	 */
2360 	bfqq->entity.prio_changed = 1;
2361 }
2362 
2363 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2364 			     struct bfq_group *bfqg)
2365 {
2366 	int i, j;
2367 
2368 	for (i = 0; i < 2; i++)
2369 		for (j = 0; j < IOPRIO_BE_NR; j++)
2370 			if (bfqg->async_bfqq[i][j])
2371 				bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2372 	if (bfqg->async_idle_bfqq)
2373 		bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2374 }
2375 
2376 static void bfq_end_wr(struct bfq_data *bfqd)
2377 {
2378 	struct bfq_queue *bfqq;
2379 
2380 	spin_lock_irq(&bfqd->lock);
2381 
2382 	list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2383 		bfq_bfqq_end_wr(bfqq);
2384 	list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2385 		bfq_bfqq_end_wr(bfqq);
2386 	bfq_end_wr_async(bfqd);
2387 
2388 	spin_unlock_irq(&bfqd->lock);
2389 }
2390 
2391 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2392 {
2393 	if (request)
2394 		return blk_rq_pos(io_struct);
2395 	else
2396 		return ((struct bio *)io_struct)->bi_iter.bi_sector;
2397 }
2398 
2399 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2400 				  sector_t sector)
2401 {
2402 	return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2403 	       BFQQ_CLOSE_THR;
2404 }
2405 
2406 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2407 					 struct bfq_queue *bfqq,
2408 					 sector_t sector)
2409 {
2410 	struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
2411 	struct rb_node *parent, *node;
2412 	struct bfq_queue *__bfqq;
2413 
2414 	if (RB_EMPTY_ROOT(root))
2415 		return NULL;
2416 
2417 	/*
2418 	 * First, if we find a request starting at the end of the last
2419 	 * request, choose it.
2420 	 */
2421 	__bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2422 	if (__bfqq)
2423 		return __bfqq;
2424 
2425 	/*
2426 	 * If the exact sector wasn't found, the parent of the NULL leaf
2427 	 * will contain the closest sector (rq_pos_tree sorted by
2428 	 * next_request position).
2429 	 */
2430 	__bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2431 	if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2432 		return __bfqq;
2433 
2434 	if (blk_rq_pos(__bfqq->next_rq) < sector)
2435 		node = rb_next(&__bfqq->pos_node);
2436 	else
2437 		node = rb_prev(&__bfqq->pos_node);
2438 	if (!node)
2439 		return NULL;
2440 
2441 	__bfqq = rb_entry(node, struct bfq_queue, pos_node);
2442 	if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2443 		return __bfqq;
2444 
2445 	return NULL;
2446 }
2447 
2448 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2449 						   struct bfq_queue *cur_bfqq,
2450 						   sector_t sector)
2451 {
2452 	struct bfq_queue *bfqq;
2453 
2454 	/*
2455 	 * We shall notice if some of the queues are cooperating,
2456 	 * e.g., working closely on the same area of the device. In
2457 	 * that case, we can group them together and: 1) don't waste
2458 	 * time idling, and 2) serve the union of their requests in
2459 	 * the best possible order for throughput.
2460 	 */
2461 	bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2462 	if (!bfqq || bfqq == cur_bfqq)
2463 		return NULL;
2464 
2465 	return bfqq;
2466 }
2467 
2468 static struct bfq_queue *
2469 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2470 {
2471 	int process_refs, new_process_refs;
2472 	struct bfq_queue *__bfqq;
2473 
2474 	/*
2475 	 * If there are no process references on the new_bfqq, then it is
2476 	 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2477 	 * may have dropped their last reference (not just their last process
2478 	 * reference).
2479 	 */
2480 	if (!bfqq_process_refs(new_bfqq))
2481 		return NULL;
2482 
2483 	/* Avoid a circular list and skip interim queue merges. */
2484 	while ((__bfqq = new_bfqq->new_bfqq)) {
2485 		if (__bfqq == bfqq)
2486 			return NULL;
2487 		new_bfqq = __bfqq;
2488 	}
2489 
2490 	process_refs = bfqq_process_refs(bfqq);
2491 	new_process_refs = bfqq_process_refs(new_bfqq);
2492 	/*
2493 	 * If the process for the bfqq has gone away, there is no
2494 	 * sense in merging the queues.
2495 	 */
2496 	if (process_refs == 0 || new_process_refs == 0)
2497 		return NULL;
2498 
2499 	bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2500 		new_bfqq->pid);
2501 
2502 	/*
2503 	 * Merging is just a redirection: the requests of the process
2504 	 * owning one of the two queues are redirected to the other queue.
2505 	 * The latter queue, in its turn, is set as shared if this is the
2506 	 * first time that the requests of some process are redirected to
2507 	 * it.
2508 	 *
2509 	 * We redirect bfqq to new_bfqq and not the opposite, because
2510 	 * we are in the context of the process owning bfqq, thus we
2511 	 * have the io_cq of this process. So we can immediately
2512 	 * configure this io_cq to redirect the requests of the
2513 	 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2514 	 * not available any more (new_bfqq->bic == NULL).
2515 	 *
2516 	 * Anyway, even in case new_bfqq coincides with the in-service
2517 	 * queue, redirecting requests the in-service queue is the
2518 	 * best option, as we feed the in-service queue with new
2519 	 * requests close to the last request served and, by doing so,
2520 	 * are likely to increase the throughput.
2521 	 */
2522 	bfqq->new_bfqq = new_bfqq;
2523 	new_bfqq->ref += process_refs;
2524 	return new_bfqq;
2525 }
2526 
2527 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2528 					struct bfq_queue *new_bfqq)
2529 {
2530 	if (bfq_too_late_for_merging(new_bfqq))
2531 		return false;
2532 
2533 	if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2534 	    (bfqq->ioprio_class != new_bfqq->ioprio_class))
2535 		return false;
2536 
2537 	/*
2538 	 * If either of the queues has already been detected as seeky,
2539 	 * then merging it with the other queue is unlikely to lead to
2540 	 * sequential I/O.
2541 	 */
2542 	if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2543 		return false;
2544 
2545 	/*
2546 	 * Interleaved I/O is known to be done by (some) applications
2547 	 * only for reads, so it does not make sense to merge async
2548 	 * queues.
2549 	 */
2550 	if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2551 		return false;
2552 
2553 	return true;
2554 }
2555 
2556 /*
2557  * Attempt to schedule a merge of bfqq with the currently in-service
2558  * queue or with a close queue among the scheduled queues.  Return
2559  * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2560  * structure otherwise.
2561  *
2562  * The OOM queue is not allowed to participate to cooperation: in fact, since
2563  * the requests temporarily redirected to the OOM queue could be redirected
2564  * again to dedicated queues at any time, the state needed to correctly
2565  * handle merging with the OOM queue would be quite complex and expensive
2566  * to maintain. Besides, in such a critical condition as an out of memory,
2567  * the benefits of queue merging may be little relevant, or even negligible.
2568  *
2569  * WARNING: queue merging may impair fairness among non-weight raised
2570  * queues, for at least two reasons: 1) the original weight of a
2571  * merged queue may change during the merged state, 2) even being the
2572  * weight the same, a merged queue may be bloated with many more
2573  * requests than the ones produced by its originally-associated
2574  * process.
2575  */
2576 static struct bfq_queue *
2577 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2578 		     void *io_struct, bool request)
2579 {
2580 	struct bfq_queue *in_service_bfqq, *new_bfqq;
2581 
2582 	/*
2583 	 * Do not perform queue merging if the device is non
2584 	 * rotational and performs internal queueing. In fact, such a
2585 	 * device reaches a high speed through internal parallelism
2586 	 * and pipelining. This means that, to reach a high
2587 	 * throughput, it must have many requests enqueued at the same
2588 	 * time. But, in this configuration, the internal scheduling
2589 	 * algorithm of the device does exactly the job of queue
2590 	 * merging: it reorders requests so as to obtain as much as
2591 	 * possible a sequential I/O pattern. As a consequence, with
2592 	 * the workload generated by processes doing interleaved I/O,
2593 	 * the throughput reached by the device is likely to be the
2594 	 * same, with and without queue merging.
2595 	 *
2596 	 * Disabling merging also provides a remarkable benefit in
2597 	 * terms of throughput. Merging tends to make many workloads
2598 	 * artificially more uneven, because of shared queues
2599 	 * remaining non empty for incomparably more time than
2600 	 * non-merged queues. This may accentuate workload
2601 	 * asymmetries. For example, if one of the queues in a set of
2602 	 * merged queues has a higher weight than a normal queue, then
2603 	 * the shared queue may inherit such a high weight and, by
2604 	 * staying almost always active, may force BFQ to perform I/O
2605 	 * plugging most of the time. This evidently makes it harder
2606 	 * for BFQ to let the device reach a high throughput.
2607 	 *
2608 	 * Finally, the likely() macro below is not used because one
2609 	 * of the two branches is more likely than the other, but to
2610 	 * have the code path after the following if() executed as
2611 	 * fast as possible for the case of a non rotational device
2612 	 * with queueing. We want it because this is the fastest kind
2613 	 * of device. On the opposite end, the likely() may lengthen
2614 	 * the execution time of BFQ for the case of slower devices
2615 	 * (rotational or at least without queueing). But in this case
2616 	 * the execution time of BFQ matters very little, if not at
2617 	 * all.
2618 	 */
2619 	if (likely(bfqd->nonrot_with_queueing))
2620 		return NULL;
2621 
2622 	/*
2623 	 * Prevent bfqq from being merged if it has been created too
2624 	 * long ago. The idea is that true cooperating processes, and
2625 	 * thus their associated bfq_queues, are supposed to be
2626 	 * created shortly after each other. This is the case, e.g.,
2627 	 * for KVM/QEMU and dump I/O threads. Basing on this
2628 	 * assumption, the following filtering greatly reduces the
2629 	 * probability that two non-cooperating processes, which just
2630 	 * happen to do close I/O for some short time interval, have
2631 	 * their queues merged by mistake.
2632 	 */
2633 	if (bfq_too_late_for_merging(bfqq))
2634 		return NULL;
2635 
2636 	if (bfqq->new_bfqq)
2637 		return bfqq->new_bfqq;
2638 
2639 	if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2640 		return NULL;
2641 
2642 	/* If there is only one backlogged queue, don't search. */
2643 	if (bfq_tot_busy_queues(bfqd) == 1)
2644 		return NULL;
2645 
2646 	in_service_bfqq = bfqd->in_service_queue;
2647 
2648 	if (in_service_bfqq && in_service_bfqq != bfqq &&
2649 	    likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2650 	    bfq_rq_close_to_sector(io_struct, request,
2651 				   bfqd->in_serv_last_pos) &&
2652 	    bfqq->entity.parent == in_service_bfqq->entity.parent &&
2653 	    bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2654 		new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2655 		if (new_bfqq)
2656 			return new_bfqq;
2657 	}
2658 	/*
2659 	 * Check whether there is a cooperator among currently scheduled
2660 	 * queues. The only thing we need is that the bio/request is not
2661 	 * NULL, as we need it to establish whether a cooperator exists.
2662 	 */
2663 	new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2664 			bfq_io_struct_pos(io_struct, request));
2665 
2666 	if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2667 	    bfq_may_be_close_cooperator(bfqq, new_bfqq))
2668 		return bfq_setup_merge(bfqq, new_bfqq);
2669 
2670 	return NULL;
2671 }
2672 
2673 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2674 {
2675 	struct bfq_io_cq *bic = bfqq->bic;
2676 
2677 	/*
2678 	 * If !bfqq->bic, the queue is already shared or its requests
2679 	 * have already been redirected to a shared queue; both idle window
2680 	 * and weight raising state have already been saved. Do nothing.
2681 	 */
2682 	if (!bic)
2683 		return;
2684 
2685 	bic->saved_weight = bfqq->entity.orig_weight;
2686 	bic->saved_ttime = bfqq->ttime;
2687 	bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2688 	bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2689 	bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2690 	bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2691 	if (unlikely(bfq_bfqq_just_created(bfqq) &&
2692 		     !bfq_bfqq_in_large_burst(bfqq) &&
2693 		     bfqq->bfqd->low_latency)) {
2694 		/*
2695 		 * bfqq being merged right after being created: bfqq
2696 		 * would have deserved interactive weight raising, but
2697 		 * did not make it to be set in a weight-raised state,
2698 		 * because of this early merge.	Store directly the
2699 		 * weight-raising state that would have been assigned
2700 		 * to bfqq, so that to avoid that bfqq unjustly fails
2701 		 * to enjoy weight raising if split soon.
2702 		 */
2703 		bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
2704 		bic->saved_wr_start_at_switch_to_srt = bfq_smallest_from_now();
2705 		bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2706 		bic->saved_last_wr_start_finish = jiffies;
2707 	} else {
2708 		bic->saved_wr_coeff = bfqq->wr_coeff;
2709 		bic->saved_wr_start_at_switch_to_srt =
2710 			bfqq->wr_start_at_switch_to_srt;
2711 		bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2712 		bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2713 	}
2714 }
2715 
2716 static void
2717 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2718 		struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2719 {
2720 	bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2721 		(unsigned long)new_bfqq->pid);
2722 	/* Save weight raising and idle window of the merged queues */
2723 	bfq_bfqq_save_state(bfqq);
2724 	bfq_bfqq_save_state(new_bfqq);
2725 	if (bfq_bfqq_IO_bound(bfqq))
2726 		bfq_mark_bfqq_IO_bound(new_bfqq);
2727 	bfq_clear_bfqq_IO_bound(bfqq);
2728 
2729 	/*
2730 	 * If bfqq is weight-raised, then let new_bfqq inherit
2731 	 * weight-raising. To reduce false positives, neglect the case
2732 	 * where bfqq has just been created, but has not yet made it
2733 	 * to be weight-raised (which may happen because EQM may merge
2734 	 * bfqq even before bfq_add_request is executed for the first
2735 	 * time for bfqq). Handling this case would however be very
2736 	 * easy, thanks to the flag just_created.
2737 	 */
2738 	if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2739 		new_bfqq->wr_coeff = bfqq->wr_coeff;
2740 		new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2741 		new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2742 		new_bfqq->wr_start_at_switch_to_srt =
2743 			bfqq->wr_start_at_switch_to_srt;
2744 		if (bfq_bfqq_busy(new_bfqq))
2745 			bfqd->wr_busy_queues++;
2746 		new_bfqq->entity.prio_changed = 1;
2747 	}
2748 
2749 	if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2750 		bfqq->wr_coeff = 1;
2751 		bfqq->entity.prio_changed = 1;
2752 		if (bfq_bfqq_busy(bfqq))
2753 			bfqd->wr_busy_queues--;
2754 	}
2755 
2756 	bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2757 		     bfqd->wr_busy_queues);
2758 
2759 	/*
2760 	 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2761 	 */
2762 	bic_set_bfqq(bic, new_bfqq, 1);
2763 	bfq_mark_bfqq_coop(new_bfqq);
2764 	/*
2765 	 * new_bfqq now belongs to at least two bics (it is a shared queue):
2766 	 * set new_bfqq->bic to NULL. bfqq either:
2767 	 * - does not belong to any bic any more, and hence bfqq->bic must
2768 	 *   be set to NULL, or
2769 	 * - is a queue whose owning bics have already been redirected to a
2770 	 *   different queue, hence the queue is destined to not belong to
2771 	 *   any bic soon and bfqq->bic is already NULL (therefore the next
2772 	 *   assignment causes no harm).
2773 	 */
2774 	new_bfqq->bic = NULL;
2775 	/*
2776 	 * If the queue is shared, the pid is the pid of one of the associated
2777 	 * processes. Which pid depends on the exact sequence of merge events
2778 	 * the queue underwent. So printing such a pid is useless and confusing
2779 	 * because it reports a random pid between those of the associated
2780 	 * processes.
2781 	 * We mark such a queue with a pid -1, and then print SHARED instead of
2782 	 * a pid in logging messages.
2783 	 */
2784 	new_bfqq->pid = -1;
2785 	bfqq->bic = NULL;
2786 	/* release process reference to bfqq */
2787 	bfq_put_queue(bfqq);
2788 }
2789 
2790 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2791 				struct bio *bio)
2792 {
2793 	struct bfq_data *bfqd = q->elevator->elevator_data;
2794 	bool is_sync = op_is_sync(bio->bi_opf);
2795 	struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2796 
2797 	/*
2798 	 * Disallow merge of a sync bio into an async request.
2799 	 */
2800 	if (is_sync && !rq_is_sync(rq))
2801 		return false;
2802 
2803 	/*
2804 	 * Lookup the bfqq that this bio will be queued with. Allow
2805 	 * merge only if rq is queued there.
2806 	 */
2807 	if (!bfqq)
2808 		return false;
2809 
2810 	/*
2811 	 * We take advantage of this function to perform an early merge
2812 	 * of the queues of possible cooperating processes.
2813 	 */
2814 	new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2815 	if (new_bfqq) {
2816 		/*
2817 		 * bic still points to bfqq, then it has not yet been
2818 		 * redirected to some other bfq_queue, and a queue
2819 		 * merge between bfqq and new_bfqq can be safely
2820 		 * fulfilled, i.e., bic can be redirected to new_bfqq
2821 		 * and bfqq can be put.
2822 		 */
2823 		bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2824 				new_bfqq);
2825 		/*
2826 		 * If we get here, bio will be queued into new_queue,
2827 		 * so use new_bfqq to decide whether bio and rq can be
2828 		 * merged.
2829 		 */
2830 		bfqq = new_bfqq;
2831 
2832 		/*
2833 		 * Change also bqfd->bio_bfqq, as
2834 		 * bfqd->bio_bic now points to new_bfqq, and
2835 		 * this function may be invoked again (and then may
2836 		 * use again bqfd->bio_bfqq).
2837 		 */
2838 		bfqd->bio_bfqq = bfqq;
2839 	}
2840 
2841 	return bfqq == RQ_BFQQ(rq);
2842 }
2843 
2844 /*
2845  * Set the maximum time for the in-service queue to consume its
2846  * budget. This prevents seeky processes from lowering the throughput.
2847  * In practice, a time-slice service scheme is used with seeky
2848  * processes.
2849  */
2850 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2851 				   struct bfq_queue *bfqq)
2852 {
2853 	unsigned int timeout_coeff;
2854 
2855 	if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2856 		timeout_coeff = 1;
2857 	else
2858 		timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2859 
2860 	bfqd->last_budget_start = ktime_get();
2861 
2862 	bfqq->budget_timeout = jiffies +
2863 		bfqd->bfq_timeout * timeout_coeff;
2864 }
2865 
2866 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2867 				       struct bfq_queue *bfqq)
2868 {
2869 	if (bfqq) {
2870 		bfq_clear_bfqq_fifo_expire(bfqq);
2871 
2872 		bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2873 
2874 		if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2875 		    bfqq->wr_coeff > 1 &&
2876 		    bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2877 		    time_is_before_jiffies(bfqq->budget_timeout)) {
2878 			/*
2879 			 * For soft real-time queues, move the start
2880 			 * of the weight-raising period forward by the
2881 			 * time the queue has not received any
2882 			 * service. Otherwise, a relatively long
2883 			 * service delay is likely to cause the
2884 			 * weight-raising period of the queue to end,
2885 			 * because of the short duration of the
2886 			 * weight-raising period of a soft real-time
2887 			 * queue.  It is worth noting that this move
2888 			 * is not so dangerous for the other queues,
2889 			 * because soft real-time queues are not
2890 			 * greedy.
2891 			 *
2892 			 * To not add a further variable, we use the
2893 			 * overloaded field budget_timeout to
2894 			 * determine for how long the queue has not
2895 			 * received service, i.e., how much time has
2896 			 * elapsed since the queue expired. However,
2897 			 * this is a little imprecise, because
2898 			 * budget_timeout is set to jiffies if bfqq
2899 			 * not only expires, but also remains with no
2900 			 * request.
2901 			 */
2902 			if (time_after(bfqq->budget_timeout,
2903 				       bfqq->last_wr_start_finish))
2904 				bfqq->last_wr_start_finish +=
2905 					jiffies - bfqq->budget_timeout;
2906 			else
2907 				bfqq->last_wr_start_finish = jiffies;
2908 		}
2909 
2910 		bfq_set_budget_timeout(bfqd, bfqq);
2911 		bfq_log_bfqq(bfqd, bfqq,
2912 			     "set_in_service_queue, cur-budget = %d",
2913 			     bfqq->entity.budget);
2914 	}
2915 
2916 	bfqd->in_service_queue = bfqq;
2917 }
2918 
2919 /*
2920  * Get and set a new queue for service.
2921  */
2922 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2923 {
2924 	struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2925 
2926 	__bfq_set_in_service_queue(bfqd, bfqq);
2927 	return bfqq;
2928 }
2929 
2930 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2931 {
2932 	struct bfq_queue *bfqq = bfqd->in_service_queue;
2933 	u32 sl;
2934 
2935 	bfq_mark_bfqq_wait_request(bfqq);
2936 
2937 	/*
2938 	 * We don't want to idle for seeks, but we do want to allow
2939 	 * fair distribution of slice time for a process doing back-to-back
2940 	 * seeks. So allow a little bit of time for him to submit a new rq.
2941 	 */
2942 	sl = bfqd->bfq_slice_idle;
2943 	/*
2944 	 * Unless the queue is being weight-raised or the scenario is
2945 	 * asymmetric, grant only minimum idle time if the queue
2946 	 * is seeky. A long idling is preserved for a weight-raised
2947 	 * queue, or, more in general, in an asymmetric scenario,
2948 	 * because a long idling is needed for guaranteeing to a queue
2949 	 * its reserved share of the throughput (in particular, it is
2950 	 * needed if the queue has a higher weight than some other
2951 	 * queue).
2952 	 */
2953 	if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
2954 	    !bfq_asymmetric_scenario(bfqd, bfqq))
2955 		sl = min_t(u64, sl, BFQ_MIN_TT);
2956 	else if (bfqq->wr_coeff > 1)
2957 		sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
2958 
2959 	bfqd->last_idling_start = ktime_get();
2960 	bfqd->last_idling_start_jiffies = jiffies;
2961 
2962 	hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
2963 		      HRTIMER_MODE_REL);
2964 	bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
2965 }
2966 
2967 /*
2968  * In autotuning mode, max_budget is dynamically recomputed as the
2969  * amount of sectors transferred in timeout at the estimated peak
2970  * rate. This enables BFQ to utilize a full timeslice with a full
2971  * budget, even if the in-service queue is served at peak rate. And
2972  * this maximises throughput with sequential workloads.
2973  */
2974 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
2975 {
2976 	return (u64)bfqd->peak_rate * USEC_PER_MSEC *
2977 		jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
2978 }
2979 
2980 /*
2981  * Update parameters related to throughput and responsiveness, as a
2982  * function of the estimated peak rate. See comments on
2983  * bfq_calc_max_budget(), and on the ref_wr_duration array.
2984  */
2985 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
2986 {
2987 	if (bfqd->bfq_user_max_budget == 0) {
2988 		bfqd->bfq_max_budget =
2989 			bfq_calc_max_budget(bfqd);
2990 		bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
2991 	}
2992 }
2993 
2994 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
2995 				       struct request *rq)
2996 {
2997 	if (rq != NULL) { /* new rq dispatch now, reset accordingly */
2998 		bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
2999 		bfqd->peak_rate_samples = 1;
3000 		bfqd->sequential_samples = 0;
3001 		bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
3002 			blk_rq_sectors(rq);
3003 	} else /* no new rq dispatched, just reset the number of samples */
3004 		bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
3005 
3006 	bfq_log(bfqd,
3007 		"reset_rate_computation at end, sample %u/%u tot_sects %llu",
3008 		bfqd->peak_rate_samples, bfqd->sequential_samples,
3009 		bfqd->tot_sectors_dispatched);
3010 }
3011 
3012 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
3013 {
3014 	u32 rate, weight, divisor;
3015 
3016 	/*
3017 	 * For the convergence property to hold (see comments on
3018 	 * bfq_update_peak_rate()) and for the assessment to be
3019 	 * reliable, a minimum number of samples must be present, and
3020 	 * a minimum amount of time must have elapsed. If not so, do
3021 	 * not compute new rate. Just reset parameters, to get ready
3022 	 * for a new evaluation attempt.
3023 	 */
3024 	if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
3025 	    bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
3026 		goto reset_computation;
3027 
3028 	/*
3029 	 * If a new request completion has occurred after last
3030 	 * dispatch, then, to approximate the rate at which requests
3031 	 * have been served by the device, it is more precise to
3032 	 * extend the observation interval to the last completion.
3033 	 */
3034 	bfqd->delta_from_first =
3035 		max_t(u64, bfqd->delta_from_first,
3036 		      bfqd->last_completion - bfqd->first_dispatch);
3037 
3038 	/*
3039 	 * Rate computed in sects/usec, and not sects/nsec, for
3040 	 * precision issues.
3041 	 */
3042 	rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
3043 			div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
3044 
3045 	/*
3046 	 * Peak rate not updated if:
3047 	 * - the percentage of sequential dispatches is below 3/4 of the
3048 	 *   total, and rate is below the current estimated peak rate
3049 	 * - rate is unreasonably high (> 20M sectors/sec)
3050 	 */
3051 	if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
3052 	     rate <= bfqd->peak_rate) ||
3053 		rate > 20<<BFQ_RATE_SHIFT)
3054 		goto reset_computation;
3055 
3056 	/*
3057 	 * We have to update the peak rate, at last! To this purpose,
3058 	 * we use a low-pass filter. We compute the smoothing constant
3059 	 * of the filter as a function of the 'weight' of the new
3060 	 * measured rate.
3061 	 *
3062 	 * As can be seen in next formulas, we define this weight as a
3063 	 * quantity proportional to how sequential the workload is,
3064 	 * and to how long the observation time interval is.
3065 	 *
3066 	 * The weight runs from 0 to 8. The maximum value of the
3067 	 * weight, 8, yields the minimum value for the smoothing
3068 	 * constant. At this minimum value for the smoothing constant,
3069 	 * the measured rate contributes for half of the next value of
3070 	 * the estimated peak rate.
3071 	 *
3072 	 * So, the first step is to compute the weight as a function
3073 	 * of how sequential the workload is. Note that the weight
3074 	 * cannot reach 9, because bfqd->sequential_samples cannot
3075 	 * become equal to bfqd->peak_rate_samples, which, in its
3076 	 * turn, holds true because bfqd->sequential_samples is not
3077 	 * incremented for the first sample.
3078 	 */
3079 	weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
3080 
3081 	/*
3082 	 * Second step: further refine the weight as a function of the
3083 	 * duration of the observation interval.
3084 	 */
3085 	weight = min_t(u32, 8,
3086 		       div_u64(weight * bfqd->delta_from_first,
3087 			       BFQ_RATE_REF_INTERVAL));
3088 
3089 	/*
3090 	 * Divisor ranging from 10, for minimum weight, to 2, for
3091 	 * maximum weight.
3092 	 */
3093 	divisor = 10 - weight;
3094 
3095 	/*
3096 	 * Finally, update peak rate:
3097 	 *
3098 	 * peak_rate = peak_rate * (divisor-1) / divisor  +  rate / divisor
3099 	 */
3100 	bfqd->peak_rate *= divisor-1;
3101 	bfqd->peak_rate /= divisor;
3102 	rate /= divisor; /* smoothing constant alpha = 1/divisor */
3103 
3104 	bfqd->peak_rate += rate;
3105 
3106 	/*
3107 	 * For a very slow device, bfqd->peak_rate can reach 0 (see
3108 	 * the minimum representable values reported in the comments
3109 	 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3110 	 * divisions by zero where bfqd->peak_rate is used as a
3111 	 * divisor.
3112 	 */
3113 	bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
3114 
3115 	update_thr_responsiveness_params(bfqd);
3116 
3117 reset_computation:
3118 	bfq_reset_rate_computation(bfqd, rq);
3119 }
3120 
3121 /*
3122  * Update the read/write peak rate (the main quantity used for
3123  * auto-tuning, see update_thr_responsiveness_params()).
3124  *
3125  * It is not trivial to estimate the peak rate (correctly): because of
3126  * the presence of sw and hw queues between the scheduler and the
3127  * device components that finally serve I/O requests, it is hard to
3128  * say exactly when a given dispatched request is served inside the
3129  * device, and for how long. As a consequence, it is hard to know
3130  * precisely at what rate a given set of requests is actually served
3131  * by the device.
3132  *
3133  * On the opposite end, the dispatch time of any request is trivially
3134  * available, and, from this piece of information, the "dispatch rate"
3135  * of requests can be immediately computed. So, the idea in the next
3136  * function is to use what is known, namely request dispatch times
3137  * (plus, when useful, request completion times), to estimate what is
3138  * unknown, namely in-device request service rate.
3139  *
3140  * The main issue is that, because of the above facts, the rate at
3141  * which a certain set of requests is dispatched over a certain time
3142  * interval can vary greatly with respect to the rate at which the
3143  * same requests are then served. But, since the size of any
3144  * intermediate queue is limited, and the service scheme is lossless
3145  * (no request is silently dropped), the following obvious convergence
3146  * property holds: the number of requests dispatched MUST become
3147  * closer and closer to the number of requests completed as the
3148  * observation interval grows. This is the key property used in
3149  * the next function to estimate the peak service rate as a function
3150  * of the observed dispatch rate. The function assumes to be invoked
3151  * on every request dispatch.
3152  */
3153 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
3154 {
3155 	u64 now_ns = ktime_get_ns();
3156 
3157 	if (bfqd->peak_rate_samples == 0) { /* first dispatch */
3158 		bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
3159 			bfqd->peak_rate_samples);
3160 		bfq_reset_rate_computation(bfqd, rq);
3161 		goto update_last_values; /* will add one sample */
3162 	}
3163 
3164 	/*
3165 	 * Device idle for very long: the observation interval lasting
3166 	 * up to this dispatch cannot be a valid observation interval
3167 	 * for computing a new peak rate (similarly to the late-
3168 	 * completion event in bfq_completed_request()). Go to
3169 	 * update_rate_and_reset to have the following three steps
3170 	 * taken:
3171 	 * - close the observation interval at the last (previous)
3172 	 *   request dispatch or completion
3173 	 * - compute rate, if possible, for that observation interval
3174 	 * - start a new observation interval with this dispatch
3175 	 */
3176 	if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
3177 	    bfqd->rq_in_driver == 0)
3178 		goto update_rate_and_reset;
3179 
3180 	/* Update sampling information */
3181 	bfqd->peak_rate_samples++;
3182 
3183 	if ((bfqd->rq_in_driver > 0 ||
3184 		now_ns - bfqd->last_completion < BFQ_MIN_TT)
3185 	    && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
3186 		bfqd->sequential_samples++;
3187 
3188 	bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
3189 
3190 	/* Reset max observed rq size every 32 dispatches */
3191 	if (likely(bfqd->peak_rate_samples % 32))
3192 		bfqd->last_rq_max_size =
3193 			max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
3194 	else
3195 		bfqd->last_rq_max_size = blk_rq_sectors(rq);
3196 
3197 	bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
3198 
3199 	/* Target observation interval not yet reached, go on sampling */
3200 	if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
3201 		goto update_last_values;
3202 
3203 update_rate_and_reset:
3204 	bfq_update_rate_reset(bfqd, rq);
3205 update_last_values:
3206 	bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
3207 	if (RQ_BFQQ(rq) == bfqd->in_service_queue)
3208 		bfqd->in_serv_last_pos = bfqd->last_position;
3209 	bfqd->last_dispatch = now_ns;
3210 }
3211 
3212 /*
3213  * Remove request from internal lists.
3214  */
3215 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
3216 {
3217 	struct bfq_queue *bfqq = RQ_BFQQ(rq);
3218 
3219 	/*
3220 	 * For consistency, the next instruction should have been
3221 	 * executed after removing the request from the queue and
3222 	 * dispatching it.  We execute instead this instruction before
3223 	 * bfq_remove_request() (and hence introduce a temporary
3224 	 * inconsistency), for efficiency.  In fact, should this
3225 	 * dispatch occur for a non in-service bfqq, this anticipated
3226 	 * increment prevents two counters related to bfqq->dispatched
3227 	 * from risking to be, first, uselessly decremented, and then
3228 	 * incremented again when the (new) value of bfqq->dispatched
3229 	 * happens to be taken into account.
3230 	 */
3231 	bfqq->dispatched++;
3232 	bfq_update_peak_rate(q->elevator->elevator_data, rq);
3233 
3234 	bfq_remove_request(q, rq);
3235 }
3236 
3237 /*
3238  * There is a case where idling does not have to be performed for
3239  * throughput concerns, but to preserve the throughput share of
3240  * the process associated with bfqq.
3241  *
3242  * To introduce this case, we can note that allowing the drive
3243  * to enqueue more than one request at a time, and hence
3244  * delegating de facto final scheduling decisions to the
3245  * drive's internal scheduler, entails loss of control on the
3246  * actual request service order. In particular, the critical
3247  * situation is when requests from different processes happen
3248  * to be present, at the same time, in the internal queue(s)
3249  * of the drive. In such a situation, the drive, by deciding
3250  * the service order of the internally-queued requests, does
3251  * determine also the actual throughput distribution among
3252  * these processes. But the drive typically has no notion or
3253  * concern about per-process throughput distribution, and
3254  * makes its decisions only on a per-request basis. Therefore,
3255  * the service distribution enforced by the drive's internal
3256  * scheduler is likely to coincide with the desired throughput
3257  * distribution only in a completely symmetric, or favorably
3258  * skewed scenario where:
3259  * (i-a) each of these processes must get the same throughput as
3260  *	 the others,
3261  * (i-b) in case (i-a) does not hold, it holds that the process
3262  *       associated with bfqq must receive a lower or equal
3263  *	 throughput than any of the other processes;
3264  * (ii)  the I/O of each process has the same properties, in
3265  *       terms of locality (sequential or random), direction
3266  *       (reads or writes), request sizes, greediness
3267  *       (from I/O-bound to sporadic), and so on;
3268 
3269  * In fact, in such a scenario, the drive tends to treat the requests
3270  * of each process in about the same way as the requests of the
3271  * others, and thus to provide each of these processes with about the
3272  * same throughput.  This is exactly the desired throughput
3273  * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3274  * even more convenient distribution for (the process associated with)
3275  * bfqq.
3276  *
3277  * In contrast, in any asymmetric or unfavorable scenario, device
3278  * idling (I/O-dispatch plugging) is certainly needed to guarantee
3279  * that bfqq receives its assigned fraction of the device throughput
3280  * (see [1] for details).
3281  *
3282  * The problem is that idling may significantly reduce throughput with
3283  * certain combinations of types of I/O and devices. An important
3284  * example is sync random I/O on flash storage with command
3285  * queueing. So, unless bfqq falls in cases where idling also boosts
3286  * throughput, it is important to check conditions (i-a), i(-b) and
3287  * (ii) accurately, so as to avoid idling when not strictly needed for
3288  * service guarantees.
3289  *
3290  * Unfortunately, it is extremely difficult to thoroughly check
3291  * condition (ii). And, in case there are active groups, it becomes
3292  * very difficult to check conditions (i-a) and (i-b) too.  In fact,
3293  * if there are active groups, then, for conditions (i-a) or (i-b) to
3294  * become false 'indirectly', it is enough that an active group
3295  * contains more active processes or sub-groups than some other active
3296  * group. More precisely, for conditions (i-a) or (i-b) to become
3297  * false because of such a group, it is not even necessary that the
3298  * group is (still) active: it is sufficient that, even if the group
3299  * has become inactive, some of its descendant processes still have
3300  * some request already dispatched but still waiting for
3301  * completion. In fact, requests have still to be guaranteed their
3302  * share of the throughput even after being dispatched. In this
3303  * respect, it is easy to show that, if a group frequently becomes
3304  * inactive while still having in-flight requests, and if, when this
3305  * happens, the group is not considered in the calculation of whether
3306  * the scenario is asymmetric, then the group may fail to be
3307  * guaranteed its fair share of the throughput (basically because
3308  * idling may not be performed for the descendant processes of the
3309  * group, but it had to be).  We address this issue with the following
3310  * bi-modal behavior, implemented in the function
3311  * bfq_asymmetric_scenario().
3312  *
3313  * If there are groups with requests waiting for completion
3314  * (as commented above, some of these groups may even be
3315  * already inactive), then the scenario is tagged as
3316  * asymmetric, conservatively, without checking any of the
3317  * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3318  * This behavior matches also the fact that groups are created
3319  * exactly if controlling I/O is a primary concern (to
3320  * preserve bandwidth and latency guarantees).
3321  *
3322  * On the opposite end, if there are no groups with requests waiting
3323  * for completion, then only conditions (i-a) and (i-b) are actually
3324  * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3325  * idling is not performed, regardless of whether condition (ii)
3326  * holds.  In other words, only if conditions (i-a) and (i-b) do not
3327  * hold, then idling is allowed, and the device tends to be prevented
3328  * from queueing many requests, possibly of several processes. Since
3329  * there are no groups with requests waiting for completion, then, to
3330  * control conditions (i-a) and (i-b) it is enough to check just
3331  * whether all the queues with requests waiting for completion also
3332  * have the same weight.
3333  *
3334  * Not checking condition (ii) evidently exposes bfqq to the
3335  * risk of getting less throughput than its fair share.
3336  * However, for queues with the same weight, a further
3337  * mechanism, preemption, mitigates or even eliminates this
3338  * problem. And it does so without consequences on overall
3339  * throughput. This mechanism and its benefits are explained
3340  * in the next three paragraphs.
3341  *
3342  * Even if a queue, say Q, is expired when it remains idle, Q
3343  * can still preempt the new in-service queue if the next
3344  * request of Q arrives soon (see the comments on
3345  * bfq_bfqq_update_budg_for_activation). If all queues and
3346  * groups have the same weight, this form of preemption,
3347  * combined with the hole-recovery heuristic described in the
3348  * comments on function bfq_bfqq_update_budg_for_activation,
3349  * are enough to preserve a correct bandwidth distribution in
3350  * the mid term, even without idling. In fact, even if not
3351  * idling allows the internal queues of the device to contain
3352  * many requests, and thus to reorder requests, we can rather
3353  * safely assume that the internal scheduler still preserves a
3354  * minimum of mid-term fairness.
3355  *
3356  * More precisely, this preemption-based, idleless approach
3357  * provides fairness in terms of IOPS, and not sectors per
3358  * second. This can be seen with a simple example. Suppose
3359  * that there are two queues with the same weight, but that
3360  * the first queue receives requests of 8 sectors, while the
3361  * second queue receives requests of 1024 sectors. In
3362  * addition, suppose that each of the two queues contains at
3363  * most one request at a time, which implies that each queue
3364  * always remains idle after it is served. Finally, after
3365  * remaining idle, each queue receives very quickly a new
3366  * request. It follows that the two queues are served
3367  * alternatively, preempting each other if needed. This
3368  * implies that, although both queues have the same weight,
3369  * the queue with large requests receives a service that is
3370  * 1024/8 times as high as the service received by the other
3371  * queue.
3372  *
3373  * The motivation for using preemption instead of idling (for
3374  * queues with the same weight) is that, by not idling,
3375  * service guarantees are preserved (completely or at least in
3376  * part) without minimally sacrificing throughput. And, if
3377  * there is no active group, then the primary expectation for
3378  * this device is probably a high throughput.
3379  *
3380  * We are now left only with explaining the two sub-conditions in the
3381  * additional compound condition that is checked below for deciding
3382  * whether the scenario is asymmetric. To explain the first
3383  * sub-condition, we need to add that the function
3384  * bfq_asymmetric_scenario checks the weights of only
3385  * non-weight-raised queues, for efficiency reasons (see comments on
3386  * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3387  * is checked explicitly here. More precisely, the compound condition
3388  * below takes into account also the fact that, even if bfqq is being
3389  * weight-raised, the scenario is still symmetric if all queues with
3390  * requests waiting for completion happen to be
3391  * weight-raised. Actually, we should be even more precise here, and
3392  * differentiate between interactive weight raising and soft real-time
3393  * weight raising.
3394  *
3395  * The second sub-condition checked in the compound condition is
3396  * whether there is a fair amount of already in-flight I/O not
3397  * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3398  * following reason. The drive may decide to serve in-flight
3399  * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3400  * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3401  * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3402  * basically uncontrolled amount of I/O from other queues may be
3403  * dispatched too, possibly causing the service of bfqq's I/O to be
3404  * delayed even longer in the drive. This problem gets more and more
3405  * serious as the speed and the queue depth of the drive grow,
3406  * because, as these two quantities grow, the probability to find no
3407  * queue busy but many requests in flight grows too. By contrast,
3408  * plugging I/O dispatching minimizes the delay induced by already
3409  * in-flight I/O, and enables bfqq to recover the bandwidth it may
3410  * lose because of this delay.
3411  *
3412  * As a side note, it is worth considering that the above
3413  * device-idling countermeasures may however fail in the following
3414  * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3415  * in a time period during which all symmetry sub-conditions hold, and
3416  * therefore the device is allowed to enqueue many requests, but at
3417  * some later point in time some sub-condition stops to hold, then it
3418  * may become impossible to make requests be served in the desired
3419  * order until all the requests already queued in the device have been
3420  * served. The last sub-condition commented above somewhat mitigates
3421  * this problem for weight-raised queues.
3422  */
3423 static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3424 						 struct bfq_queue *bfqq)
3425 {
3426 	return (bfqq->wr_coeff > 1 &&
3427 		(bfqd->wr_busy_queues <
3428 		 bfq_tot_busy_queues(bfqd) ||
3429 		 bfqd->rq_in_driver >=
3430 		 bfqq->dispatched + 4)) ||
3431 		bfq_asymmetric_scenario(bfqd, bfqq);
3432 }
3433 
3434 static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3435 			      enum bfqq_expiration reason)
3436 {
3437 	/*
3438 	 * If this bfqq is shared between multiple processes, check
3439 	 * to make sure that those processes are still issuing I/Os
3440 	 * within the mean seek distance. If not, it may be time to
3441 	 * break the queues apart again.
3442 	 */
3443 	if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
3444 		bfq_mark_bfqq_split_coop(bfqq);
3445 
3446 	/*
3447 	 * Consider queues with a higher finish virtual time than
3448 	 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3449 	 * true, then bfqq's bandwidth would be violated if an
3450 	 * uncontrolled amount of I/O from these queues were
3451 	 * dispatched while bfqq is waiting for its new I/O to
3452 	 * arrive. This is exactly what may happen if this is a forced
3453 	 * expiration caused by a preemption attempt, and if bfqq is
3454 	 * not re-scheduled. To prevent this from happening, re-queue
3455 	 * bfqq if it needs I/O-dispatch plugging, even if it is
3456 	 * empty. By doing so, bfqq is granted to be served before the
3457 	 * above queues (provided that bfqq is of course eligible).
3458 	 */
3459 	if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
3460 	    !(reason == BFQQE_PREEMPTED &&
3461 	      idling_needed_for_service_guarantees(bfqd, bfqq))) {
3462 		if (bfqq->dispatched == 0)
3463 			/*
3464 			 * Overloading budget_timeout field to store
3465 			 * the time at which the queue remains with no
3466 			 * backlog and no outstanding request; used by
3467 			 * the weight-raising mechanism.
3468 			 */
3469 			bfqq->budget_timeout = jiffies;
3470 
3471 		bfq_del_bfqq_busy(bfqd, bfqq, true);
3472 	} else {
3473 		bfq_requeue_bfqq(bfqd, bfqq, true);
3474 		/*
3475 		 * Resort priority tree of potential close cooperators.
3476 		 * See comments on bfq_pos_tree_add_move() for the unlikely().
3477 		 */
3478 		if (unlikely(!bfqd->nonrot_with_queueing &&
3479 			     !RB_EMPTY_ROOT(&bfqq->sort_list)))
3480 			bfq_pos_tree_add_move(bfqd, bfqq);
3481 	}
3482 
3483 	/*
3484 	 * All in-service entities must have been properly deactivated
3485 	 * or requeued before executing the next function, which
3486 	 * resets all in-service entities as no more in service. This
3487 	 * may cause bfqq to be freed. If this happens, the next
3488 	 * function returns true.
3489 	 */
3490 	return __bfq_bfqd_reset_in_service(bfqd);
3491 }
3492 
3493 /**
3494  * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3495  * @bfqd: device data.
3496  * @bfqq: queue to update.
3497  * @reason: reason for expiration.
3498  *
3499  * Handle the feedback on @bfqq budget at queue expiration.
3500  * See the body for detailed comments.
3501  */
3502 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
3503 				     struct bfq_queue *bfqq,
3504 				     enum bfqq_expiration reason)
3505 {
3506 	struct request *next_rq;
3507 	int budget, min_budget;
3508 
3509 	min_budget = bfq_min_budget(bfqd);
3510 
3511 	if (bfqq->wr_coeff == 1)
3512 		budget = bfqq->max_budget;
3513 	else /*
3514 	      * Use a constant, low budget for weight-raised queues,
3515 	      * to help achieve a low latency. Keep it slightly higher
3516 	      * than the minimum possible budget, to cause a little
3517 	      * bit fewer expirations.
3518 	      */
3519 		budget = 2 * min_budget;
3520 
3521 	bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
3522 		bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
3523 	bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
3524 		budget, bfq_min_budget(bfqd));
3525 	bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
3526 		bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
3527 
3528 	if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
3529 		switch (reason) {
3530 		/*
3531 		 * Caveat: in all the following cases we trade latency
3532 		 * for throughput.
3533 		 */
3534 		case BFQQE_TOO_IDLE:
3535 			/*
3536 			 * This is the only case where we may reduce
3537 			 * the budget: if there is no request of the
3538 			 * process still waiting for completion, then
3539 			 * we assume (tentatively) that the timer has
3540 			 * expired because the batch of requests of
3541 			 * the process could have been served with a
3542 			 * smaller budget.  Hence, betting that
3543 			 * process will behave in the same way when it
3544 			 * becomes backlogged again, we reduce its
3545 			 * next budget.  As long as we guess right,
3546 			 * this budget cut reduces the latency
3547 			 * experienced by the process.
3548 			 *
3549 			 * However, if there are still outstanding
3550 			 * requests, then the process may have not yet
3551 			 * issued its next request just because it is
3552 			 * still waiting for the completion of some of
3553 			 * the still outstanding ones.  So in this
3554 			 * subcase we do not reduce its budget, on the
3555 			 * contrary we increase it to possibly boost
3556 			 * the throughput, as discussed in the
3557 			 * comments to the BUDGET_TIMEOUT case.
3558 			 */
3559 			if (bfqq->dispatched > 0) /* still outstanding reqs */
3560 				budget = min(budget * 2, bfqd->bfq_max_budget);
3561 			else {
3562 				if (budget > 5 * min_budget)
3563 					budget -= 4 * min_budget;
3564 				else
3565 					budget = min_budget;
3566 			}
3567 			break;
3568 		case BFQQE_BUDGET_TIMEOUT:
3569 			/*
3570 			 * We double the budget here because it gives
3571 			 * the chance to boost the throughput if this
3572 			 * is not a seeky process (and has bumped into
3573 			 * this timeout because of, e.g., ZBR).
3574 			 */
3575 			budget = min(budget * 2, bfqd->bfq_max_budget);
3576 			break;
3577 		case BFQQE_BUDGET_EXHAUSTED:
3578 			/*
3579 			 * The process still has backlog, and did not
3580 			 * let either the budget timeout or the disk
3581 			 * idling timeout expire. Hence it is not
3582 			 * seeky, has a short thinktime and may be
3583 			 * happy with a higher budget too. So
3584 			 * definitely increase the budget of this good
3585 			 * candidate to boost the disk throughput.
3586 			 */
3587 			budget = min(budget * 4, bfqd->bfq_max_budget);
3588 			break;
3589 		case BFQQE_NO_MORE_REQUESTS:
3590 			/*
3591 			 * For queues that expire for this reason, it
3592 			 * is particularly important to keep the
3593 			 * budget close to the actual service they
3594 			 * need. Doing so reduces the timestamp
3595 			 * misalignment problem described in the
3596 			 * comments in the body of
3597 			 * __bfq_activate_entity. In fact, suppose
3598 			 * that a queue systematically expires for
3599 			 * BFQQE_NO_MORE_REQUESTS and presents a
3600 			 * new request in time to enjoy timestamp
3601 			 * back-shifting. The larger the budget of the
3602 			 * queue is with respect to the service the
3603 			 * queue actually requests in each service
3604 			 * slot, the more times the queue can be
3605 			 * reactivated with the same virtual finish
3606 			 * time. It follows that, even if this finish
3607 			 * time is pushed to the system virtual time
3608 			 * to reduce the consequent timestamp
3609 			 * misalignment, the queue unjustly enjoys for
3610 			 * many re-activations a lower finish time
3611 			 * than all newly activated queues.
3612 			 *
3613 			 * The service needed by bfqq is measured
3614 			 * quite precisely by bfqq->entity.service.
3615 			 * Since bfqq does not enjoy device idling,
3616 			 * bfqq->entity.service is equal to the number
3617 			 * of sectors that the process associated with
3618 			 * bfqq requested to read/write before waiting
3619 			 * for request completions, or blocking for
3620 			 * other reasons.
3621 			 */
3622 			budget = max_t(int, bfqq->entity.service, min_budget);
3623 			break;
3624 		default:
3625 			return;
3626 		}
3627 	} else if (!bfq_bfqq_sync(bfqq)) {
3628 		/*
3629 		 * Async queues get always the maximum possible
3630 		 * budget, as for them we do not care about latency
3631 		 * (in addition, their ability to dispatch is limited
3632 		 * by the charging factor).
3633 		 */
3634 		budget = bfqd->bfq_max_budget;
3635 	}
3636 
3637 	bfqq->max_budget = budget;
3638 
3639 	if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
3640 	    !bfqd->bfq_user_max_budget)
3641 		bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
3642 
3643 	/*
3644 	 * If there is still backlog, then assign a new budget, making
3645 	 * sure that it is large enough for the next request.  Since
3646 	 * the finish time of bfqq must be kept in sync with the
3647 	 * budget, be sure to call __bfq_bfqq_expire() *after* this
3648 	 * update.
3649 	 *
3650 	 * If there is no backlog, then no need to update the budget;
3651 	 * it will be updated on the arrival of a new request.
3652 	 */
3653 	next_rq = bfqq->next_rq;
3654 	if (next_rq)
3655 		bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
3656 					    bfq_serv_to_charge(next_rq, bfqq));
3657 
3658 	bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
3659 			next_rq ? blk_rq_sectors(next_rq) : 0,
3660 			bfqq->entity.budget);
3661 }
3662 
3663 /*
3664  * Return true if the process associated with bfqq is "slow". The slow
3665  * flag is used, in addition to the budget timeout, to reduce the
3666  * amount of service provided to seeky processes, and thus reduce
3667  * their chances to lower the throughput. More details in the comments
3668  * on the function bfq_bfqq_expire().
3669  *
3670  * An important observation is in order: as discussed in the comments
3671  * on the function bfq_update_peak_rate(), with devices with internal
3672  * queues, it is hard if ever possible to know when and for how long
3673  * an I/O request is processed by the device (apart from the trivial
3674  * I/O pattern where a new request is dispatched only after the
3675  * previous one has been completed). This makes it hard to evaluate
3676  * the real rate at which the I/O requests of each bfq_queue are
3677  * served.  In fact, for an I/O scheduler like BFQ, serving a
3678  * bfq_queue means just dispatching its requests during its service
3679  * slot (i.e., until the budget of the queue is exhausted, or the
3680  * queue remains idle, or, finally, a timeout fires). But, during the
3681  * service slot of a bfq_queue, around 100 ms at most, the device may
3682  * be even still processing requests of bfq_queues served in previous
3683  * service slots. On the opposite end, the requests of the in-service
3684  * bfq_queue may be completed after the service slot of the queue
3685  * finishes.
3686  *
3687  * Anyway, unless more sophisticated solutions are used
3688  * (where possible), the sum of the sizes of the requests dispatched
3689  * during the service slot of a bfq_queue is probably the only
3690  * approximation available for the service received by the bfq_queue
3691  * during its service slot. And this sum is the quantity used in this
3692  * function to evaluate the I/O speed of a process.
3693  */
3694 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3695 				 bool compensate, enum bfqq_expiration reason,
3696 				 unsigned long *delta_ms)
3697 {
3698 	ktime_t delta_ktime;
3699 	u32 delta_usecs;
3700 	bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
3701 
3702 	if (!bfq_bfqq_sync(bfqq))
3703 		return false;
3704 
3705 	if (compensate)
3706 		delta_ktime = bfqd->last_idling_start;
3707 	else
3708 		delta_ktime = ktime_get();
3709 	delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
3710 	delta_usecs = ktime_to_us(delta_ktime);
3711 
3712 	/* don't use too short time intervals */
3713 	if (delta_usecs < 1000) {
3714 		if (blk_queue_nonrot(bfqd->queue))
3715 			 /*
3716 			  * give same worst-case guarantees as idling
3717 			  * for seeky
3718 			  */
3719 			*delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
3720 		else /* charge at least one seek */
3721 			*delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
3722 
3723 		return slow;
3724 	}
3725 
3726 	*delta_ms = delta_usecs / USEC_PER_MSEC;
3727 
3728 	/*
3729 	 * Use only long (> 20ms) intervals to filter out excessive
3730 	 * spikes in service rate estimation.
3731 	 */
3732 	if (delta_usecs > 20000) {
3733 		/*
3734 		 * Caveat for rotational devices: processes doing I/O
3735 		 * in the slower disk zones tend to be slow(er) even
3736 		 * if not seeky. In this respect, the estimated peak
3737 		 * rate is likely to be an average over the disk
3738 		 * surface. Accordingly, to not be too harsh with
3739 		 * unlucky processes, a process is deemed slow only if
3740 		 * its rate has been lower than half of the estimated
3741 		 * peak rate.
3742 		 */
3743 		slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
3744 	}
3745 
3746 	bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
3747 
3748 	return slow;
3749 }
3750 
3751 /*
3752  * To be deemed as soft real-time, an application must meet two
3753  * requirements. First, the application must not require an average
3754  * bandwidth higher than the approximate bandwidth required to playback or
3755  * record a compressed high-definition video.
3756  * The next function is invoked on the completion of the last request of a
3757  * batch, to compute the next-start time instant, soft_rt_next_start, such
3758  * that, if the next request of the application does not arrive before
3759  * soft_rt_next_start, then the above requirement on the bandwidth is met.
3760  *
3761  * The second requirement is that the request pattern of the application is
3762  * isochronous, i.e., that, after issuing a request or a batch of requests,
3763  * the application stops issuing new requests until all its pending requests
3764  * have been completed. After that, the application may issue a new batch,
3765  * and so on.
3766  * For this reason the next function is invoked to compute
3767  * soft_rt_next_start only for applications that meet this requirement,
3768  * whereas soft_rt_next_start is set to infinity for applications that do
3769  * not.
3770  *
3771  * Unfortunately, even a greedy (i.e., I/O-bound) application may
3772  * happen to meet, occasionally or systematically, both the above
3773  * bandwidth and isochrony requirements. This may happen at least in
3774  * the following circumstances. First, if the CPU load is high. The
3775  * application may stop issuing requests while the CPUs are busy
3776  * serving other processes, then restart, then stop again for a while,
3777  * and so on. The other circumstances are related to the storage
3778  * device: the storage device is highly loaded or reaches a low-enough
3779  * throughput with the I/O of the application (e.g., because the I/O
3780  * is random and/or the device is slow). In all these cases, the
3781  * I/O of the application may be simply slowed down enough to meet
3782  * the bandwidth and isochrony requirements. To reduce the probability
3783  * that greedy applications are deemed as soft real-time in these
3784  * corner cases, a further rule is used in the computation of
3785  * soft_rt_next_start: the return value of this function is forced to
3786  * be higher than the maximum between the following two quantities.
3787  *
3788  * (a) Current time plus: (1) the maximum time for which the arrival
3789  *     of a request is waited for when a sync queue becomes idle,
3790  *     namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
3791  *     postpone for a moment the reason for adding a few extra
3792  *     jiffies; we get back to it after next item (b).  Lower-bounding
3793  *     the return value of this function with the current time plus
3794  *     bfqd->bfq_slice_idle tends to filter out greedy applications,
3795  *     because the latter issue their next request as soon as possible
3796  *     after the last one has been completed. In contrast, a soft
3797  *     real-time application spends some time processing data, after a
3798  *     batch of its requests has been completed.
3799  *
3800  * (b) Current value of bfqq->soft_rt_next_start. As pointed out
3801  *     above, greedy applications may happen to meet both the
3802  *     bandwidth and isochrony requirements under heavy CPU or
3803  *     storage-device load. In more detail, in these scenarios, these
3804  *     applications happen, only for limited time periods, to do I/O
3805  *     slowly enough to meet all the requirements described so far,
3806  *     including the filtering in above item (a). These slow-speed
3807  *     time intervals are usually interspersed between other time
3808  *     intervals during which these applications do I/O at a very high
3809  *     speed. Fortunately, exactly because of the high speed of the
3810  *     I/O in the high-speed intervals, the values returned by this
3811  *     function happen to be so high, near the end of any such
3812  *     high-speed interval, to be likely to fall *after* the end of
3813  *     the low-speed time interval that follows. These high values are
3814  *     stored in bfqq->soft_rt_next_start after each invocation of
3815  *     this function. As a consequence, if the last value of
3816  *     bfqq->soft_rt_next_start is constantly used to lower-bound the
3817  *     next value that this function may return, then, from the very
3818  *     beginning of a low-speed interval, bfqq->soft_rt_next_start is
3819  *     likely to be constantly kept so high that any I/O request
3820  *     issued during the low-speed interval is considered as arriving
3821  *     to soon for the application to be deemed as soft
3822  *     real-time. Then, in the high-speed interval that follows, the
3823  *     application will not be deemed as soft real-time, just because
3824  *     it will do I/O at a high speed. And so on.
3825  *
3826  * Getting back to the filtering in item (a), in the following two
3827  * cases this filtering might be easily passed by a greedy
3828  * application, if the reference quantity was just
3829  * bfqd->bfq_slice_idle:
3830  * 1) HZ is so low that the duration of a jiffy is comparable to or
3831  *    higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
3832  *    devices with HZ=100. The time granularity may be so coarse
3833  *    that the approximation, in jiffies, of bfqd->bfq_slice_idle
3834  *    is rather lower than the exact value.
3835  * 2) jiffies, instead of increasing at a constant rate, may stop increasing
3836  *    for a while, then suddenly 'jump' by several units to recover the lost
3837  *    increments. This seems to happen, e.g., inside virtual machines.
3838  * To address this issue, in the filtering in (a) we do not use as a
3839  * reference time interval just bfqd->bfq_slice_idle, but
3840  * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
3841  * minimum number of jiffies for which the filter seems to be quite
3842  * precise also in embedded systems and KVM/QEMU virtual machines.
3843  */
3844 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
3845 						struct bfq_queue *bfqq)
3846 {
3847 	return max3(bfqq->soft_rt_next_start,
3848 		    bfqq->last_idle_bklogged +
3849 		    HZ * bfqq->service_from_backlogged /
3850 		    bfqd->bfq_wr_max_softrt_rate,
3851 		    jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
3852 }
3853 
3854 /**
3855  * bfq_bfqq_expire - expire a queue.
3856  * @bfqd: device owning the queue.
3857  * @bfqq: the queue to expire.
3858  * @compensate: if true, compensate for the time spent idling.
3859  * @reason: the reason causing the expiration.
3860  *
3861  * If the process associated with bfqq does slow I/O (e.g., because it
3862  * issues random requests), we charge bfqq with the time it has been
3863  * in service instead of the service it has received (see
3864  * bfq_bfqq_charge_time for details on how this goal is achieved). As
3865  * a consequence, bfqq will typically get higher timestamps upon
3866  * reactivation, and hence it will be rescheduled as if it had
3867  * received more service than what it has actually received. In the
3868  * end, bfqq receives less service in proportion to how slowly its
3869  * associated process consumes its budgets (and hence how seriously it
3870  * tends to lower the throughput). In addition, this time-charging
3871  * strategy guarantees time fairness among slow processes. In
3872  * contrast, if the process associated with bfqq is not slow, we
3873  * charge bfqq exactly with the service it has received.
3874  *
3875  * Charging time to the first type of queues and the exact service to
3876  * the other has the effect of using the WF2Q+ policy to schedule the
3877  * former on a timeslice basis, without violating service domain
3878  * guarantees among the latter.
3879  */
3880 void bfq_bfqq_expire(struct bfq_data *bfqd,
3881 		     struct bfq_queue *bfqq,
3882 		     bool compensate,
3883 		     enum bfqq_expiration reason)
3884 {
3885 	bool slow;
3886 	unsigned long delta = 0;
3887 	struct bfq_entity *entity = &bfqq->entity;
3888 
3889 	/*
3890 	 * Check whether the process is slow (see bfq_bfqq_is_slow).
3891 	 */
3892 	slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
3893 
3894 	/*
3895 	 * As above explained, charge slow (typically seeky) and
3896 	 * timed-out queues with the time and not the service
3897 	 * received, to favor sequential workloads.
3898 	 *
3899 	 * Processes doing I/O in the slower disk zones will tend to
3900 	 * be slow(er) even if not seeky. Therefore, since the
3901 	 * estimated peak rate is actually an average over the disk
3902 	 * surface, these processes may timeout just for bad luck. To
3903 	 * avoid punishing them, do not charge time to processes that
3904 	 * succeeded in consuming at least 2/3 of their budget. This
3905 	 * allows BFQ to preserve enough elasticity to still perform
3906 	 * bandwidth, and not time, distribution with little unlucky
3907 	 * or quasi-sequential processes.
3908 	 */
3909 	if (bfqq->wr_coeff == 1 &&
3910 	    (slow ||
3911 	     (reason == BFQQE_BUDGET_TIMEOUT &&
3912 	      bfq_bfqq_budget_left(bfqq) >=  entity->budget / 3)))
3913 		bfq_bfqq_charge_time(bfqd, bfqq, delta);
3914 
3915 	if (reason == BFQQE_TOO_IDLE &&
3916 	    entity->service <= 2 * entity->budget / 10)
3917 		bfq_clear_bfqq_IO_bound(bfqq);
3918 
3919 	if (bfqd->low_latency && bfqq->wr_coeff == 1)
3920 		bfqq->last_wr_start_finish = jiffies;
3921 
3922 	if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
3923 	    RB_EMPTY_ROOT(&bfqq->sort_list)) {
3924 		/*
3925 		 * If we get here, and there are no outstanding
3926 		 * requests, then the request pattern is isochronous
3927 		 * (see the comments on the function
3928 		 * bfq_bfqq_softrt_next_start()). Thus we can compute
3929 		 * soft_rt_next_start. And we do it, unless bfqq is in
3930 		 * interactive weight raising. We do not do it in the
3931 		 * latter subcase, for the following reason. bfqq may
3932 		 * be conveying the I/O needed to load a soft
3933 		 * real-time application. Such an application will
3934 		 * actually exhibit a soft real-time I/O pattern after
3935 		 * it finally starts doing its job. But, if
3936 		 * soft_rt_next_start is computed here for an
3937 		 * interactive bfqq, and bfqq had received a lot of
3938 		 * service before remaining with no outstanding
3939 		 * request (likely to happen on a fast device), then
3940 		 * soft_rt_next_start would be assigned such a high
3941 		 * value that, for a very long time, bfqq would be
3942 		 * prevented from being possibly considered as soft
3943 		 * real time.
3944 		 *
3945 		 * If, instead, the queue still has outstanding
3946 		 * requests, then we have to wait for the completion
3947 		 * of all the outstanding requests to discover whether
3948 		 * the request pattern is actually isochronous.
3949 		 */
3950 		if (bfqq->dispatched == 0 &&
3951 		    bfqq->wr_coeff != bfqd->bfq_wr_coeff)
3952 			bfqq->soft_rt_next_start =
3953 				bfq_bfqq_softrt_next_start(bfqd, bfqq);
3954 		else if (bfqq->dispatched > 0) {
3955 			/*
3956 			 * Schedule an update of soft_rt_next_start to when
3957 			 * the task may be discovered to be isochronous.
3958 			 */
3959 			bfq_mark_bfqq_softrt_update(bfqq);
3960 		}
3961 	}
3962 
3963 	bfq_log_bfqq(bfqd, bfqq,
3964 		"expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
3965 		slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
3966 
3967 	/*
3968 	 * bfqq expired, so no total service time needs to be computed
3969 	 * any longer: reset state machine for measuring total service
3970 	 * times.
3971 	 */
3972 	bfqd->rqs_injected = bfqd->wait_dispatch = false;
3973 	bfqd->waited_rq = NULL;
3974 
3975 	/*
3976 	 * Increase, decrease or leave budget unchanged according to
3977 	 * reason.
3978 	 */
3979 	__bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
3980 	if (__bfq_bfqq_expire(bfqd, bfqq, reason))
3981 		/* bfqq is gone, no more actions on it */
3982 		return;
3983 
3984 	/* mark bfqq as waiting a request only if a bic still points to it */
3985 	if (!bfq_bfqq_busy(bfqq) &&
3986 	    reason != BFQQE_BUDGET_TIMEOUT &&
3987 	    reason != BFQQE_BUDGET_EXHAUSTED) {
3988 		bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
3989 		/*
3990 		 * Not setting service to 0, because, if the next rq
3991 		 * arrives in time, the queue will go on receiving
3992 		 * service with this same budget (as if it never expired)
3993 		 */
3994 	} else
3995 		entity->service = 0;
3996 
3997 	/*
3998 	 * Reset the received-service counter for every parent entity.
3999 	 * Differently from what happens with bfqq->entity.service,
4000 	 * the resetting of this counter never needs to be postponed
4001 	 * for parent entities. In fact, in case bfqq may have a
4002 	 * chance to go on being served using the last, partially
4003 	 * consumed budget, bfqq->entity.service needs to be kept,
4004 	 * because if bfqq then actually goes on being served using
4005 	 * the same budget, the last value of bfqq->entity.service is
4006 	 * needed to properly decrement bfqq->entity.budget by the
4007 	 * portion already consumed. In contrast, it is not necessary
4008 	 * to keep entity->service for parent entities too, because
4009 	 * the bubble up of the new value of bfqq->entity.budget will
4010 	 * make sure that the budgets of parent entities are correct,
4011 	 * even in case bfqq and thus parent entities go on receiving
4012 	 * service with the same budget.
4013 	 */
4014 	entity = entity->parent;
4015 	for_each_entity(entity)
4016 		entity->service = 0;
4017 }
4018 
4019 /*
4020  * Budget timeout is not implemented through a dedicated timer, but
4021  * just checked on request arrivals and completions, as well as on
4022  * idle timer expirations.
4023  */
4024 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
4025 {
4026 	return time_is_before_eq_jiffies(bfqq->budget_timeout);
4027 }
4028 
4029 /*
4030  * If we expire a queue that is actively waiting (i.e., with the
4031  * device idled) for the arrival of a new request, then we may incur
4032  * the timestamp misalignment problem described in the body of the
4033  * function __bfq_activate_entity. Hence we return true only if this
4034  * condition does not hold, or if the queue is slow enough to deserve
4035  * only to be kicked off for preserving a high throughput.
4036  */
4037 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
4038 {
4039 	bfq_log_bfqq(bfqq->bfqd, bfqq,
4040 		"may_budget_timeout: wait_request %d left %d timeout %d",
4041 		bfq_bfqq_wait_request(bfqq),
4042 			bfq_bfqq_budget_left(bfqq) >=  bfqq->entity.budget / 3,
4043 		bfq_bfqq_budget_timeout(bfqq));
4044 
4045 	return (!bfq_bfqq_wait_request(bfqq) ||
4046 		bfq_bfqq_budget_left(bfqq) >=  bfqq->entity.budget / 3)
4047 		&&
4048 		bfq_bfqq_budget_timeout(bfqq);
4049 }
4050 
4051 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
4052 					     struct bfq_queue *bfqq)
4053 {
4054 	bool rot_without_queueing =
4055 		!blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
4056 		bfqq_sequential_and_IO_bound,
4057 		idling_boosts_thr;
4058 
4059 	bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
4060 		bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
4061 
4062 	/*
4063 	 * The next variable takes into account the cases where idling
4064 	 * boosts the throughput.
4065 	 *
4066 	 * The value of the variable is computed considering, first, that
4067 	 * idling is virtually always beneficial for the throughput if:
4068 	 * (a) the device is not NCQ-capable and rotational, or
4069 	 * (b) regardless of the presence of NCQ, the device is rotational and
4070 	 *     the request pattern for bfqq is I/O-bound and sequential, or
4071 	 * (c) regardless of whether it is rotational, the device is
4072 	 *     not NCQ-capable and the request pattern for bfqq is
4073 	 *     I/O-bound and sequential.
4074 	 *
4075 	 * Secondly, and in contrast to the above item (b), idling an
4076 	 * NCQ-capable flash-based device would not boost the
4077 	 * throughput even with sequential I/O; rather it would lower
4078 	 * the throughput in proportion to how fast the device
4079 	 * is. Accordingly, the next variable is true if any of the
4080 	 * above conditions (a), (b) or (c) is true, and, in
4081 	 * particular, happens to be false if bfqd is an NCQ-capable
4082 	 * flash-based device.
4083 	 */
4084 	idling_boosts_thr = rot_without_queueing ||
4085 		((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
4086 		 bfqq_sequential_and_IO_bound);
4087 
4088 	/*
4089 	 * The return value of this function is equal to that of
4090 	 * idling_boosts_thr, unless a special case holds. In this
4091 	 * special case, described below, idling may cause problems to
4092 	 * weight-raised queues.
4093 	 *
4094 	 * When the request pool is saturated (e.g., in the presence
4095 	 * of write hogs), if the processes associated with
4096 	 * non-weight-raised queues ask for requests at a lower rate,
4097 	 * then processes associated with weight-raised queues have a
4098 	 * higher probability to get a request from the pool
4099 	 * immediately (or at least soon) when they need one. Thus
4100 	 * they have a higher probability to actually get a fraction
4101 	 * of the device throughput proportional to their high
4102 	 * weight. This is especially true with NCQ-capable drives,
4103 	 * which enqueue several requests in advance, and further
4104 	 * reorder internally-queued requests.
4105 	 *
4106 	 * For this reason, we force to false the return value if
4107 	 * there are weight-raised busy queues. In this case, and if
4108 	 * bfqq is not weight-raised, this guarantees that the device
4109 	 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4110 	 * then idling will be guaranteed by another variable, see
4111 	 * below). Combined with the timestamping rules of BFQ (see
4112 	 * [1] for details), this behavior causes bfqq, and hence any
4113 	 * sync non-weight-raised queue, to get a lower number of
4114 	 * requests served, and thus to ask for a lower number of
4115 	 * requests from the request pool, before the busy
4116 	 * weight-raised queues get served again. This often mitigates
4117 	 * starvation problems in the presence of heavy write
4118 	 * workloads and NCQ, thereby guaranteeing a higher
4119 	 * application and system responsiveness in these hostile
4120 	 * scenarios.
4121 	 */
4122 	return idling_boosts_thr &&
4123 		bfqd->wr_busy_queues == 0;
4124 }
4125 
4126 /*
4127  * For a queue that becomes empty, device idling is allowed only if
4128  * this function returns true for that queue. As a consequence, since
4129  * device idling plays a critical role for both throughput boosting
4130  * and service guarantees, the return value of this function plays a
4131  * critical role as well.
4132  *
4133  * In a nutshell, this function returns true only if idling is
4134  * beneficial for throughput or, even if detrimental for throughput,
4135  * idling is however necessary to preserve service guarantees (low
4136  * latency, desired throughput distribution, ...). In particular, on
4137  * NCQ-capable devices, this function tries to return false, so as to
4138  * help keep the drives' internal queues full, whenever this helps the
4139  * device boost the throughput without causing any service-guarantee
4140  * issue.
4141  *
4142  * Most of the issues taken into account to get the return value of
4143  * this function are not trivial. We discuss these issues in the two
4144  * functions providing the main pieces of information needed by this
4145  * function.
4146  */
4147 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
4148 {
4149 	struct bfq_data *bfqd = bfqq->bfqd;
4150 	bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
4151 
4152 	if (unlikely(bfqd->strict_guarantees))
4153 		return true;
4154 
4155 	/*
4156 	 * Idling is performed only if slice_idle > 0. In addition, we
4157 	 * do not idle if
4158 	 * (a) bfqq is async
4159 	 * (b) bfqq is in the idle io prio class: in this case we do
4160 	 * not idle because we want to minimize the bandwidth that
4161 	 * queues in this class can steal to higher-priority queues
4162 	 */
4163 	if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
4164 	   bfq_class_idle(bfqq))
4165 		return false;
4166 
4167 	idling_boosts_thr_with_no_issue =
4168 		idling_boosts_thr_without_issues(bfqd, bfqq);
4169 
4170 	idling_needed_for_service_guar =
4171 		idling_needed_for_service_guarantees(bfqd, bfqq);
4172 
4173 	/*
4174 	 * We have now the two components we need to compute the
4175 	 * return value of the function, which is true only if idling
4176 	 * either boosts the throughput (without issues), or is
4177 	 * necessary to preserve service guarantees.
4178 	 */
4179 	return idling_boosts_thr_with_no_issue ||
4180 		idling_needed_for_service_guar;
4181 }
4182 
4183 /*
4184  * If the in-service queue is empty but the function bfq_better_to_idle
4185  * returns true, then:
4186  * 1) the queue must remain in service and cannot be expired, and
4187  * 2) the device must be idled to wait for the possible arrival of a new
4188  *    request for the queue.
4189  * See the comments on the function bfq_better_to_idle for the reasons
4190  * why performing device idling is the best choice to boost the throughput
4191  * and preserve service guarantees when bfq_better_to_idle itself
4192  * returns true.
4193  */
4194 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
4195 {
4196 	return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
4197 }
4198 
4199 /*
4200  * This function chooses the queue from which to pick the next extra
4201  * I/O request to inject, if it finds a compatible queue. See the
4202  * comments on bfq_update_inject_limit() for details on the injection
4203  * mechanism, and for the definitions of the quantities mentioned
4204  * below.
4205  */
4206 static struct bfq_queue *
4207 bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
4208 {
4209 	struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
4210 	unsigned int limit = in_serv_bfqq->inject_limit;
4211 	/*
4212 	 * If
4213 	 * - bfqq is not weight-raised and therefore does not carry
4214 	 *   time-critical I/O,
4215 	 * or
4216 	 * - regardless of whether bfqq is weight-raised, bfqq has
4217 	 *   however a long think time, during which it can absorb the
4218 	 *   effect of an appropriate number of extra I/O requests
4219 	 *   from other queues (see bfq_update_inject_limit for
4220 	 *   details on the computation of this number);
4221 	 * then injection can be performed without restrictions.
4222 	 */
4223 	bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
4224 		!bfq_bfqq_has_short_ttime(in_serv_bfqq);
4225 
4226 	/*
4227 	 * If
4228 	 * - the baseline total service time could not be sampled yet,
4229 	 *   so the inject limit happens to be still 0, and
4230 	 * - a lot of time has elapsed since the plugging of I/O
4231 	 *   dispatching started, so drive speed is being wasted
4232 	 *   significantly;
4233 	 * then temporarily raise inject limit to one request.
4234 	 */
4235 	if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
4236 	    bfq_bfqq_wait_request(in_serv_bfqq) &&
4237 	    time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
4238 				      bfqd->bfq_slice_idle)
4239 		)
4240 		limit = 1;
4241 
4242 	if (bfqd->rq_in_driver >= limit)
4243 		return NULL;
4244 
4245 	/*
4246 	 * Linear search of the source queue for injection; but, with
4247 	 * a high probability, very few steps are needed to find a
4248 	 * candidate queue, i.e., a queue with enough budget left for
4249 	 * its next request. In fact:
4250 	 * - BFQ dynamically updates the budget of every queue so as
4251 	 *   to accommodate the expected backlog of the queue;
4252 	 * - if a queue gets all its requests dispatched as injected
4253 	 *   service, then the queue is removed from the active list
4254 	 *   (and re-added only if it gets new requests, but then it
4255 	 *   is assigned again enough budget for its new backlog).
4256 	 */
4257 	list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
4258 		if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4259 		    (in_serv_always_inject || bfqq->wr_coeff > 1) &&
4260 		    bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4261 		    bfq_bfqq_budget_left(bfqq)) {
4262 			/*
4263 			 * Allow for only one large in-flight request
4264 			 * on non-rotational devices, for the
4265 			 * following reason. On non-rotationl drives,
4266 			 * large requests take much longer than
4267 			 * smaller requests to be served. In addition,
4268 			 * the drive prefers to serve large requests
4269 			 * w.r.t. to small ones, if it can choose. So,
4270 			 * having more than one large requests queued
4271 			 * in the drive may easily make the next first
4272 			 * request of the in-service queue wait for so
4273 			 * long to break bfqq's service guarantees. On
4274 			 * the bright side, large requests let the
4275 			 * drive reach a very high throughput, even if
4276 			 * there is only one in-flight large request
4277 			 * at a time.
4278 			 */
4279 			if (blk_queue_nonrot(bfqd->queue) &&
4280 			    blk_rq_sectors(bfqq->next_rq) >=
4281 			    BFQQ_SECT_THR_NONROT)
4282 				limit = min_t(unsigned int, 1, limit);
4283 			else
4284 				limit = in_serv_bfqq->inject_limit;
4285 
4286 			if (bfqd->rq_in_driver < limit) {
4287 				bfqd->rqs_injected = true;
4288 				return bfqq;
4289 			}
4290 		}
4291 
4292 	return NULL;
4293 }
4294 
4295 /*
4296  * Select a queue for service.  If we have a current queue in service,
4297  * check whether to continue servicing it, or retrieve and set a new one.
4298  */
4299 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
4300 {
4301 	struct bfq_queue *bfqq;
4302 	struct request *next_rq;
4303 	enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
4304 
4305 	bfqq = bfqd->in_service_queue;
4306 	if (!bfqq)
4307 		goto new_queue;
4308 
4309 	bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
4310 
4311 	/*
4312 	 * Do not expire bfqq for budget timeout if bfqq may be about
4313 	 * to enjoy device idling. The reason why, in this case, we
4314 	 * prevent bfqq from expiring is the same as in the comments
4315 	 * on the case where bfq_bfqq_must_idle() returns true, in
4316 	 * bfq_completed_request().
4317 	 */
4318 	if (bfq_may_expire_for_budg_timeout(bfqq) &&
4319 	    !bfq_bfqq_must_idle(bfqq))
4320 		goto expire;
4321 
4322 check_queue:
4323 	/*
4324 	 * This loop is rarely executed more than once. Even when it
4325 	 * happens, it is much more convenient to re-execute this loop
4326 	 * than to return NULL and trigger a new dispatch to get a
4327 	 * request served.
4328 	 */
4329 	next_rq = bfqq->next_rq;
4330 	/*
4331 	 * If bfqq has requests queued and it has enough budget left to
4332 	 * serve them, keep the queue, otherwise expire it.
4333 	 */
4334 	if (next_rq) {
4335 		if (bfq_serv_to_charge(next_rq, bfqq) >
4336 			bfq_bfqq_budget_left(bfqq)) {
4337 			/*
4338 			 * Expire the queue for budget exhaustion,
4339 			 * which makes sure that the next budget is
4340 			 * enough to serve the next request, even if
4341 			 * it comes from the fifo expired path.
4342 			 */
4343 			reason = BFQQE_BUDGET_EXHAUSTED;
4344 			goto expire;
4345 		} else {
4346 			/*
4347 			 * The idle timer may be pending because we may
4348 			 * not disable disk idling even when a new request
4349 			 * arrives.
4350 			 */
4351 			if (bfq_bfqq_wait_request(bfqq)) {
4352 				/*
4353 				 * If we get here: 1) at least a new request
4354 				 * has arrived but we have not disabled the
4355 				 * timer because the request was too small,
4356 				 * 2) then the block layer has unplugged
4357 				 * the device, causing the dispatch to be
4358 				 * invoked.
4359 				 *
4360 				 * Since the device is unplugged, now the
4361 				 * requests are probably large enough to
4362 				 * provide a reasonable throughput.
4363 				 * So we disable idling.
4364 				 */
4365 				bfq_clear_bfqq_wait_request(bfqq);
4366 				hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4367 			}
4368 			goto keep_queue;
4369 		}
4370 	}
4371 
4372 	/*
4373 	 * No requests pending. However, if the in-service queue is idling
4374 	 * for a new request, or has requests waiting for a completion and
4375 	 * may idle after their completion, then keep it anyway.
4376 	 *
4377 	 * Yet, inject service from other queues if it boosts
4378 	 * throughput and is possible.
4379 	 */
4380 	if (bfq_bfqq_wait_request(bfqq) ||
4381 	    (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
4382 		struct bfq_queue *async_bfqq =
4383 			bfqq->bic && bfqq->bic->bfqq[0] &&
4384 			bfq_bfqq_busy(bfqq->bic->bfqq[0]) &&
4385 			bfqq->bic->bfqq[0]->next_rq ?
4386 			bfqq->bic->bfqq[0] : NULL;
4387 
4388 		/*
4389 		 * The next three mutually-exclusive ifs decide
4390 		 * whether to try injection, and choose the queue to
4391 		 * pick an I/O request from.
4392 		 *
4393 		 * The first if checks whether the process associated
4394 		 * with bfqq has also async I/O pending. If so, it
4395 		 * injects such I/O unconditionally. Injecting async
4396 		 * I/O from the same process can cause no harm to the
4397 		 * process. On the contrary, it can only increase
4398 		 * bandwidth and reduce latency for the process.
4399 		 *
4400 		 * The second if checks whether there happens to be a
4401 		 * non-empty waker queue for bfqq, i.e., a queue whose
4402 		 * I/O needs to be completed for bfqq to receive new
4403 		 * I/O. This happens, e.g., if bfqq is associated with
4404 		 * a process that does some sync. A sync generates
4405 		 * extra blocking I/O, which must be completed before
4406 		 * the process associated with bfqq can go on with its
4407 		 * I/O. If the I/O of the waker queue is not served,
4408 		 * then bfqq remains empty, and no I/O is dispatched,
4409 		 * until the idle timeout fires for bfqq. This is
4410 		 * likely to result in lower bandwidth and higher
4411 		 * latencies for bfqq, and in a severe loss of total
4412 		 * throughput. The best action to take is therefore to
4413 		 * serve the waker queue as soon as possible. So do it
4414 		 * (without relying on the third alternative below for
4415 		 * eventually serving waker_bfqq's I/O; see the last
4416 		 * paragraph for further details). This systematic
4417 		 * injection of I/O from the waker queue does not
4418 		 * cause any delay to bfqq's I/O. On the contrary,
4419 		 * next bfqq's I/O is brought forward dramatically,
4420 		 * for it is not blocked for milliseconds.
4421 		 *
4422 		 * The third if checks whether bfqq is a queue for
4423 		 * which it is better to avoid injection. It is so if
4424 		 * bfqq delivers more throughput when served without
4425 		 * any further I/O from other queues in the middle, or
4426 		 * if the service times of bfqq's I/O requests both
4427 		 * count more than overall throughput, and may be
4428 		 * easily increased by injection (this happens if bfqq
4429 		 * has a short think time). If none of these
4430 		 * conditions holds, then a candidate queue for
4431 		 * injection is looked for through
4432 		 * bfq_choose_bfqq_for_injection(). Note that the
4433 		 * latter may return NULL (for example if the inject
4434 		 * limit for bfqq is currently 0).
4435 		 *
4436 		 * NOTE: motivation for the second alternative
4437 		 *
4438 		 * Thanks to the way the inject limit is updated in
4439 		 * bfq_update_has_short_ttime(), it is rather likely
4440 		 * that, if I/O is being plugged for bfqq and the
4441 		 * waker queue has pending I/O requests that are
4442 		 * blocking bfqq's I/O, then the third alternative
4443 		 * above lets the waker queue get served before the
4444 		 * I/O-plugging timeout fires. So one may deem the
4445 		 * second alternative superfluous. It is not, because
4446 		 * the third alternative may be way less effective in
4447 		 * case of a synchronization. For two main
4448 		 * reasons. First, throughput may be low because the
4449 		 * inject limit may be too low to guarantee the same
4450 		 * amount of injected I/O, from the waker queue or
4451 		 * other queues, that the second alternative
4452 		 * guarantees (the second alternative unconditionally
4453 		 * injects a pending I/O request of the waker queue
4454 		 * for each bfq_dispatch_request()). Second, with the
4455 		 * third alternative, the duration of the plugging,
4456 		 * i.e., the time before bfqq finally receives new I/O,
4457 		 * may not be minimized, because the waker queue may
4458 		 * happen to be served only after other queues.
4459 		 */
4460 		if (async_bfqq &&
4461 		    icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
4462 		    bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
4463 		    bfq_bfqq_budget_left(async_bfqq))
4464 			bfqq = bfqq->bic->bfqq[0];
4465 		else if (bfq_bfqq_has_waker(bfqq) &&
4466 			   bfq_bfqq_busy(bfqq->waker_bfqq) &&
4467 			   bfqq->next_rq &&
4468 			   bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
4469 					      bfqq->waker_bfqq) <=
4470 			   bfq_bfqq_budget_left(bfqq->waker_bfqq)
4471 			)
4472 			bfqq = bfqq->waker_bfqq;
4473 		else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
4474 			 (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
4475 			  !bfq_bfqq_has_short_ttime(bfqq)))
4476 			bfqq = bfq_choose_bfqq_for_injection(bfqd);
4477 		else
4478 			bfqq = NULL;
4479 
4480 		goto keep_queue;
4481 	}
4482 
4483 	reason = BFQQE_NO_MORE_REQUESTS;
4484 expire:
4485 	bfq_bfqq_expire(bfqd, bfqq, false, reason);
4486 new_queue:
4487 	bfqq = bfq_set_in_service_queue(bfqd);
4488 	if (bfqq) {
4489 		bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
4490 		goto check_queue;
4491 	}
4492 keep_queue:
4493 	if (bfqq)
4494 		bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
4495 	else
4496 		bfq_log(bfqd, "select_queue: no queue returned");
4497 
4498 	return bfqq;
4499 }
4500 
4501 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4502 {
4503 	struct bfq_entity *entity = &bfqq->entity;
4504 
4505 	if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
4506 		bfq_log_bfqq(bfqd, bfqq,
4507 			"raising period dur %u/%u msec, old coeff %u, w %d(%d)",
4508 			jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
4509 			jiffies_to_msecs(bfqq->wr_cur_max_time),
4510 			bfqq->wr_coeff,
4511 			bfqq->entity.weight, bfqq->entity.orig_weight);
4512 
4513 		if (entity->prio_changed)
4514 			bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
4515 
4516 		/*
4517 		 * If the queue was activated in a burst, or too much
4518 		 * time has elapsed from the beginning of this
4519 		 * weight-raising period, then end weight raising.
4520 		 */
4521 		if (bfq_bfqq_in_large_burst(bfqq))
4522 			bfq_bfqq_end_wr(bfqq);
4523 		else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
4524 						bfqq->wr_cur_max_time)) {
4525 			if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
4526 			time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
4527 					       bfq_wr_duration(bfqd)))
4528 				bfq_bfqq_end_wr(bfqq);
4529 			else {
4530 				switch_back_to_interactive_wr(bfqq, bfqd);
4531 				bfqq->entity.prio_changed = 1;
4532 			}
4533 		}
4534 		if (bfqq->wr_coeff > 1 &&
4535 		    bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
4536 		    bfqq->service_from_wr > max_service_from_wr) {
4537 			/* see comments on max_service_from_wr */
4538 			bfq_bfqq_end_wr(bfqq);
4539 		}
4540 	}
4541 	/*
4542 	 * To improve latency (for this or other queues), immediately
4543 	 * update weight both if it must be raised and if it must be
4544 	 * lowered. Since, entity may be on some active tree here, and
4545 	 * might have a pending change of its ioprio class, invoke
4546 	 * next function with the last parameter unset (see the
4547 	 * comments on the function).
4548 	 */
4549 	if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
4550 		__bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
4551 						entity, false);
4552 }
4553 
4554 /*
4555  * Dispatch next request from bfqq.
4556  */
4557 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
4558 						 struct bfq_queue *bfqq)
4559 {
4560 	struct request *rq = bfqq->next_rq;
4561 	unsigned long service_to_charge;
4562 
4563 	service_to_charge = bfq_serv_to_charge(rq, bfqq);
4564 
4565 	bfq_bfqq_served(bfqq, service_to_charge);
4566 
4567 	if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
4568 		bfqd->wait_dispatch = false;
4569 		bfqd->waited_rq = rq;
4570 	}
4571 
4572 	bfq_dispatch_remove(bfqd->queue, rq);
4573 
4574 	if (bfqq != bfqd->in_service_queue)
4575 		goto return_rq;
4576 
4577 	/*
4578 	 * If weight raising has to terminate for bfqq, then next
4579 	 * function causes an immediate update of bfqq's weight,
4580 	 * without waiting for next activation. As a consequence, on
4581 	 * expiration, bfqq will be timestamped as if has never been
4582 	 * weight-raised during this service slot, even if it has
4583 	 * received part or even most of the service as a
4584 	 * weight-raised queue. This inflates bfqq's timestamps, which
4585 	 * is beneficial, as bfqq is then more willing to leave the
4586 	 * device immediately to possible other weight-raised queues.
4587 	 */
4588 	bfq_update_wr_data(bfqd, bfqq);
4589 
4590 	/*
4591 	 * Expire bfqq, pretending that its budget expired, if bfqq
4592 	 * belongs to CLASS_IDLE and other queues are waiting for
4593 	 * service.
4594 	 */
4595 	if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
4596 		goto return_rq;
4597 
4598 	bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
4599 
4600 return_rq:
4601 	return rq;
4602 }
4603 
4604 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
4605 {
4606 	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4607 
4608 	/*
4609 	 * Avoiding lock: a race on bfqd->busy_queues should cause at
4610 	 * most a call to dispatch for nothing
4611 	 */
4612 	return !list_empty_careful(&bfqd->dispatch) ||
4613 		bfq_tot_busy_queues(bfqd) > 0;
4614 }
4615 
4616 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
4617 {
4618 	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4619 	struct request *rq = NULL;
4620 	struct bfq_queue *bfqq = NULL;
4621 
4622 	if (!list_empty(&bfqd->dispatch)) {
4623 		rq = list_first_entry(&bfqd->dispatch, struct request,
4624 				      queuelist);
4625 		list_del_init(&rq->queuelist);
4626 
4627 		bfqq = RQ_BFQQ(rq);
4628 
4629 		if (bfqq) {
4630 			/*
4631 			 * Increment counters here, because this
4632 			 * dispatch does not follow the standard
4633 			 * dispatch flow (where counters are
4634 			 * incremented)
4635 			 */
4636 			bfqq->dispatched++;
4637 
4638 			goto inc_in_driver_start_rq;
4639 		}
4640 
4641 		/*
4642 		 * We exploit the bfq_finish_requeue_request hook to
4643 		 * decrement rq_in_driver, but
4644 		 * bfq_finish_requeue_request will not be invoked on
4645 		 * this request. So, to avoid unbalance, just start
4646 		 * this request, without incrementing rq_in_driver. As
4647 		 * a negative consequence, rq_in_driver is deceptively
4648 		 * lower than it should be while this request is in
4649 		 * service. This may cause bfq_schedule_dispatch to be
4650 		 * invoked uselessly.
4651 		 *
4652 		 * As for implementing an exact solution, the
4653 		 * bfq_finish_requeue_request hook, if defined, is
4654 		 * probably invoked also on this request. So, by
4655 		 * exploiting this hook, we could 1) increment
4656 		 * rq_in_driver here, and 2) decrement it in
4657 		 * bfq_finish_requeue_request. Such a solution would
4658 		 * let the value of the counter be always accurate,
4659 		 * but it would entail using an extra interface
4660 		 * function. This cost seems higher than the benefit,
4661 		 * being the frequency of non-elevator-private
4662 		 * requests very low.
4663 		 */
4664 		goto start_rq;
4665 	}
4666 
4667 	bfq_log(bfqd, "dispatch requests: %d busy queues",
4668 		bfq_tot_busy_queues(bfqd));
4669 
4670 	if (bfq_tot_busy_queues(bfqd) == 0)
4671 		goto exit;
4672 
4673 	/*
4674 	 * Force device to serve one request at a time if
4675 	 * strict_guarantees is true. Forcing this service scheme is
4676 	 * currently the ONLY way to guarantee that the request
4677 	 * service order enforced by the scheduler is respected by a
4678 	 * queueing device. Otherwise the device is free even to make
4679 	 * some unlucky request wait for as long as the device
4680 	 * wishes.
4681 	 *
4682 	 * Of course, serving one request at at time may cause loss of
4683 	 * throughput.
4684 	 */
4685 	if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
4686 		goto exit;
4687 
4688 	bfqq = bfq_select_queue(bfqd);
4689 	if (!bfqq)
4690 		goto exit;
4691 
4692 	rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
4693 
4694 	if (rq) {
4695 inc_in_driver_start_rq:
4696 		bfqd->rq_in_driver++;
4697 start_rq:
4698 		rq->rq_flags |= RQF_STARTED;
4699 	}
4700 exit:
4701 	return rq;
4702 }
4703 
4704 #ifdef CONFIG_BFQ_CGROUP_DEBUG
4705 static void bfq_update_dispatch_stats(struct request_queue *q,
4706 				      struct request *rq,
4707 				      struct bfq_queue *in_serv_queue,
4708 				      bool idle_timer_disabled)
4709 {
4710 	struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
4711 
4712 	if (!idle_timer_disabled && !bfqq)
4713 		return;
4714 
4715 	/*
4716 	 * rq and bfqq are guaranteed to exist until this function
4717 	 * ends, for the following reasons. First, rq can be
4718 	 * dispatched to the device, and then can be completed and
4719 	 * freed, only after this function ends. Second, rq cannot be
4720 	 * merged (and thus freed because of a merge) any longer,
4721 	 * because it has already started. Thus rq cannot be freed
4722 	 * before this function ends, and, since rq has a reference to
4723 	 * bfqq, the same guarantee holds for bfqq too.
4724 	 *
4725 	 * In addition, the following queue lock guarantees that
4726 	 * bfqq_group(bfqq) exists as well.
4727 	 */
4728 	spin_lock_irq(&q->queue_lock);
4729 	if (idle_timer_disabled)
4730 		/*
4731 		 * Since the idle timer has been disabled,
4732 		 * in_serv_queue contained some request when
4733 		 * __bfq_dispatch_request was invoked above, which
4734 		 * implies that rq was picked exactly from
4735 		 * in_serv_queue. Thus in_serv_queue == bfqq, and is
4736 		 * therefore guaranteed to exist because of the above
4737 		 * arguments.
4738 		 */
4739 		bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
4740 	if (bfqq) {
4741 		struct bfq_group *bfqg = bfqq_group(bfqq);
4742 
4743 		bfqg_stats_update_avg_queue_size(bfqg);
4744 		bfqg_stats_set_start_empty_time(bfqg);
4745 		bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
4746 	}
4747 	spin_unlock_irq(&q->queue_lock);
4748 }
4749 #else
4750 static inline void bfq_update_dispatch_stats(struct request_queue *q,
4751 					     struct request *rq,
4752 					     struct bfq_queue *in_serv_queue,
4753 					     bool idle_timer_disabled) {}
4754 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
4755 
4756 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
4757 {
4758 	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4759 	struct request *rq;
4760 	struct bfq_queue *in_serv_queue;
4761 	bool waiting_rq, idle_timer_disabled;
4762 
4763 	spin_lock_irq(&bfqd->lock);
4764 
4765 	in_serv_queue = bfqd->in_service_queue;
4766 	waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
4767 
4768 	rq = __bfq_dispatch_request(hctx);
4769 
4770 	idle_timer_disabled =
4771 		waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
4772 
4773 	spin_unlock_irq(&bfqd->lock);
4774 
4775 	bfq_update_dispatch_stats(hctx->queue, rq, in_serv_queue,
4776 				  idle_timer_disabled);
4777 
4778 	return rq;
4779 }
4780 
4781 /*
4782  * Task holds one reference to the queue, dropped when task exits.  Each rq
4783  * in-flight on this queue also holds a reference, dropped when rq is freed.
4784  *
4785  * Scheduler lock must be held here. Recall not to use bfqq after calling
4786  * this function on it.
4787  */
4788 void bfq_put_queue(struct bfq_queue *bfqq)
4789 {
4790 	struct bfq_queue *item;
4791 	struct hlist_node *n;
4792 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4793 	struct bfq_group *bfqg = bfqq_group(bfqq);
4794 #endif
4795 
4796 	if (bfqq->bfqd)
4797 		bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
4798 			     bfqq, bfqq->ref);
4799 
4800 	bfqq->ref--;
4801 	if (bfqq->ref)
4802 		return;
4803 
4804 	if (!hlist_unhashed(&bfqq->burst_list_node)) {
4805 		hlist_del_init(&bfqq->burst_list_node);
4806 		/*
4807 		 * Decrement also burst size after the removal, if the
4808 		 * process associated with bfqq is exiting, and thus
4809 		 * does not contribute to the burst any longer. This
4810 		 * decrement helps filter out false positives of large
4811 		 * bursts, when some short-lived process (often due to
4812 		 * the execution of commands by some service) happens
4813 		 * to start and exit while a complex application is
4814 		 * starting, and thus spawning several processes that
4815 		 * do I/O (and that *must not* be treated as a large
4816 		 * burst, see comments on bfq_handle_burst).
4817 		 *
4818 		 * In particular, the decrement is performed only if:
4819 		 * 1) bfqq is not a merged queue, because, if it is,
4820 		 * then this free of bfqq is not triggered by the exit
4821 		 * of the process bfqq is associated with, but exactly
4822 		 * by the fact that bfqq has just been merged.
4823 		 * 2) burst_size is greater than 0, to handle
4824 		 * unbalanced decrements. Unbalanced decrements may
4825 		 * happen in te following case: bfqq is inserted into
4826 		 * the current burst list--without incrementing
4827 		 * bust_size--because of a split, but the current
4828 		 * burst list is not the burst list bfqq belonged to
4829 		 * (see comments on the case of a split in
4830 		 * bfq_set_request).
4831 		 */
4832 		if (bfqq->bic && bfqq->bfqd->burst_size > 0)
4833 			bfqq->bfqd->burst_size--;
4834 	}
4835 
4836 	/*
4837 	 * bfqq does not exist any longer, so it cannot be woken by
4838 	 * any other queue, and cannot wake any other queue. Then bfqq
4839 	 * must be removed from the woken list of its possible waker
4840 	 * queue, and all queues in the woken list of bfqq must stop
4841 	 * having a waker queue. Strictly speaking, these updates
4842 	 * should be performed when bfqq remains with no I/O source
4843 	 * attached to it, which happens before bfqq gets freed. In
4844 	 * particular, this happens when the last process associated
4845 	 * with bfqq exits or gets associated with a different
4846 	 * queue. However, both events lead to bfqq being freed soon,
4847 	 * and dangling references would come out only after bfqq gets
4848 	 * freed. So these updates are done here, as a simple and safe
4849 	 * way to handle all cases.
4850 	 */
4851 	/* remove bfqq from woken list */
4852 	if (!hlist_unhashed(&bfqq->woken_list_node))
4853 		hlist_del_init(&bfqq->woken_list_node);
4854 
4855 	/* reset waker for all queues in woken list */
4856 	hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
4857 				  woken_list_node) {
4858 		item->waker_bfqq = NULL;
4859 		bfq_clear_bfqq_has_waker(item);
4860 		hlist_del_init(&item->woken_list_node);
4861 	}
4862 
4863 	if (bfqq->bfqd && bfqq->bfqd->last_completed_rq_bfqq == bfqq)
4864 		bfqq->bfqd->last_completed_rq_bfqq = NULL;
4865 
4866 	kmem_cache_free(bfq_pool, bfqq);
4867 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4868 	bfqg_and_blkg_put(bfqg);
4869 #endif
4870 }
4871 
4872 static void bfq_put_cooperator(struct bfq_queue *bfqq)
4873 {
4874 	struct bfq_queue *__bfqq, *next;
4875 
4876 	/*
4877 	 * If this queue was scheduled to merge with another queue, be
4878 	 * sure to drop the reference taken on that queue (and others in
4879 	 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
4880 	 */
4881 	__bfqq = bfqq->new_bfqq;
4882 	while (__bfqq) {
4883 		if (__bfqq == bfqq)
4884 			break;
4885 		next = __bfqq->new_bfqq;
4886 		bfq_put_queue(__bfqq);
4887 		__bfqq = next;
4888 	}
4889 }
4890 
4891 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4892 {
4893 	if (bfqq == bfqd->in_service_queue) {
4894 		__bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT);
4895 		bfq_schedule_dispatch(bfqd);
4896 	}
4897 
4898 	bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
4899 
4900 	bfq_put_cooperator(bfqq);
4901 
4902 	bfq_put_queue(bfqq); /* release process reference */
4903 }
4904 
4905 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
4906 {
4907 	struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4908 	struct bfq_data *bfqd;
4909 
4910 	if (bfqq)
4911 		bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
4912 
4913 	if (bfqq && bfqd) {
4914 		unsigned long flags;
4915 
4916 		spin_lock_irqsave(&bfqd->lock, flags);
4917 		bfqq->bic = NULL;
4918 		bfq_exit_bfqq(bfqd, bfqq);
4919 		bic_set_bfqq(bic, NULL, is_sync);
4920 		spin_unlock_irqrestore(&bfqd->lock, flags);
4921 	}
4922 }
4923 
4924 static void bfq_exit_icq(struct io_cq *icq)
4925 {
4926 	struct bfq_io_cq *bic = icq_to_bic(icq);
4927 
4928 	bfq_exit_icq_bfqq(bic, true);
4929 	bfq_exit_icq_bfqq(bic, false);
4930 }
4931 
4932 /*
4933  * Update the entity prio values; note that the new values will not
4934  * be used until the next (re)activation.
4935  */
4936 static void
4937 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
4938 {
4939 	struct task_struct *tsk = current;
4940 	int ioprio_class;
4941 	struct bfq_data *bfqd = bfqq->bfqd;
4942 
4943 	if (!bfqd)
4944 		return;
4945 
4946 	ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4947 	switch (ioprio_class) {
4948 	default:
4949 		dev_err(bfqq->bfqd->queue->backing_dev_info->dev,
4950 			"bfq: bad prio class %d\n", ioprio_class);
4951 		/* fall through */
4952 	case IOPRIO_CLASS_NONE:
4953 		/*
4954 		 * No prio set, inherit CPU scheduling settings.
4955 		 */
4956 		bfqq->new_ioprio = task_nice_ioprio(tsk);
4957 		bfqq->new_ioprio_class = task_nice_ioclass(tsk);
4958 		break;
4959 	case IOPRIO_CLASS_RT:
4960 		bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4961 		bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
4962 		break;
4963 	case IOPRIO_CLASS_BE:
4964 		bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4965 		bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
4966 		break;
4967 	case IOPRIO_CLASS_IDLE:
4968 		bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
4969 		bfqq->new_ioprio = 7;
4970 		break;
4971 	}
4972 
4973 	if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
4974 		pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
4975 			bfqq->new_ioprio);
4976 		bfqq->new_ioprio = IOPRIO_BE_NR;
4977 	}
4978 
4979 	bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
4980 	bfqq->entity.prio_changed = 1;
4981 }
4982 
4983 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4984 				       struct bio *bio, bool is_sync,
4985 				       struct bfq_io_cq *bic);
4986 
4987 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
4988 {
4989 	struct bfq_data *bfqd = bic_to_bfqd(bic);
4990 	struct bfq_queue *bfqq;
4991 	int ioprio = bic->icq.ioc->ioprio;
4992 
4993 	/*
4994 	 * This condition may trigger on a newly created bic, be sure to
4995 	 * drop the lock before returning.
4996 	 */
4997 	if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
4998 		return;
4999 
5000 	bic->ioprio = ioprio;
5001 
5002 	bfqq = bic_to_bfqq(bic, false);
5003 	if (bfqq) {
5004 		/* release process reference on this queue */
5005 		bfq_put_queue(bfqq);
5006 		bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
5007 		bic_set_bfqq(bic, bfqq, false);
5008 	}
5009 
5010 	bfqq = bic_to_bfqq(bic, true);
5011 	if (bfqq)
5012 		bfq_set_next_ioprio_data(bfqq, bic);
5013 }
5014 
5015 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5016 			  struct bfq_io_cq *bic, pid_t pid, int is_sync)
5017 {
5018 	RB_CLEAR_NODE(&bfqq->entity.rb_node);
5019 	INIT_LIST_HEAD(&bfqq->fifo);
5020 	INIT_HLIST_NODE(&bfqq->burst_list_node);
5021 	INIT_HLIST_NODE(&bfqq->woken_list_node);
5022 	INIT_HLIST_HEAD(&bfqq->woken_list);
5023 
5024 	bfqq->ref = 0;
5025 	bfqq->bfqd = bfqd;
5026 
5027 	if (bic)
5028 		bfq_set_next_ioprio_data(bfqq, bic);
5029 
5030 	if (is_sync) {
5031 		/*
5032 		 * No need to mark as has_short_ttime if in
5033 		 * idle_class, because no device idling is performed
5034 		 * for queues in idle class
5035 		 */
5036 		if (!bfq_class_idle(bfqq))
5037 			/* tentatively mark as has_short_ttime */
5038 			bfq_mark_bfqq_has_short_ttime(bfqq);
5039 		bfq_mark_bfqq_sync(bfqq);
5040 		bfq_mark_bfqq_just_created(bfqq);
5041 	} else
5042 		bfq_clear_bfqq_sync(bfqq);
5043 
5044 	/* set end request to minus infinity from now */
5045 	bfqq->ttime.last_end_request = ktime_get_ns() + 1;
5046 
5047 	bfq_mark_bfqq_IO_bound(bfqq);
5048 
5049 	bfqq->pid = pid;
5050 
5051 	/* Tentative initial value to trade off between thr and lat */
5052 	bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
5053 	bfqq->budget_timeout = bfq_smallest_from_now();
5054 
5055 	bfqq->wr_coeff = 1;
5056 	bfqq->last_wr_start_finish = jiffies;
5057 	bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
5058 	bfqq->split_time = bfq_smallest_from_now();
5059 
5060 	/*
5061 	 * To not forget the possibly high bandwidth consumed by a
5062 	 * process/queue in the recent past,
5063 	 * bfq_bfqq_softrt_next_start() returns a value at least equal
5064 	 * to the current value of bfqq->soft_rt_next_start (see
5065 	 * comments on bfq_bfqq_softrt_next_start).  Set
5066 	 * soft_rt_next_start to now, to mean that bfqq has consumed
5067 	 * no bandwidth so far.
5068 	 */
5069 	bfqq->soft_rt_next_start = jiffies;
5070 
5071 	/* first request is almost certainly seeky */
5072 	bfqq->seek_history = 1;
5073 }
5074 
5075 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
5076 					       struct bfq_group *bfqg,
5077 					       int ioprio_class, int ioprio)
5078 {
5079 	switch (ioprio_class) {
5080 	case IOPRIO_CLASS_RT:
5081 		return &bfqg->async_bfqq[0][ioprio];
5082 	case IOPRIO_CLASS_NONE:
5083 		ioprio = IOPRIO_NORM;
5084 		/* fall through */
5085 	case IOPRIO_CLASS_BE:
5086 		return &bfqg->async_bfqq[1][ioprio];
5087 	case IOPRIO_CLASS_IDLE:
5088 		return &bfqg->async_idle_bfqq;
5089 	default:
5090 		return NULL;
5091 	}
5092 }
5093 
5094 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5095 				       struct bio *bio, bool is_sync,
5096 				       struct bfq_io_cq *bic)
5097 {
5098 	const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5099 	const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5100 	struct bfq_queue **async_bfqq = NULL;
5101 	struct bfq_queue *bfqq;
5102 	struct bfq_group *bfqg;
5103 
5104 	rcu_read_lock();
5105 
5106 	bfqg = bfq_find_set_group(bfqd, __bio_blkcg(bio));
5107 	if (!bfqg) {
5108 		bfqq = &bfqd->oom_bfqq;
5109 		goto out;
5110 	}
5111 
5112 	if (!is_sync) {
5113 		async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
5114 						  ioprio);
5115 		bfqq = *async_bfqq;
5116 		if (bfqq)
5117 			goto out;
5118 	}
5119 
5120 	bfqq = kmem_cache_alloc_node(bfq_pool,
5121 				     GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
5122 				     bfqd->queue->node);
5123 
5124 	if (bfqq) {
5125 		bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
5126 			      is_sync);
5127 		bfq_init_entity(&bfqq->entity, bfqg);
5128 		bfq_log_bfqq(bfqd, bfqq, "allocated");
5129 	} else {
5130 		bfqq = &bfqd->oom_bfqq;
5131 		bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
5132 		goto out;
5133 	}
5134 
5135 	/*
5136 	 * Pin the queue now that it's allocated, scheduler exit will
5137 	 * prune it.
5138 	 */
5139 	if (async_bfqq) {
5140 		bfqq->ref++; /*
5141 			      * Extra group reference, w.r.t. sync
5142 			      * queue. This extra reference is removed
5143 			      * only if bfqq->bfqg disappears, to
5144 			      * guarantee that this queue is not freed
5145 			      * until its group goes away.
5146 			      */
5147 		bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
5148 			     bfqq, bfqq->ref);
5149 		*async_bfqq = bfqq;
5150 	}
5151 
5152 out:
5153 	bfqq->ref++; /* get a process reference to this queue */
5154 	bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
5155 	rcu_read_unlock();
5156 	return bfqq;
5157 }
5158 
5159 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
5160 				    struct bfq_queue *bfqq)
5161 {
5162 	struct bfq_ttime *ttime = &bfqq->ttime;
5163 	u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
5164 
5165 	elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
5166 
5167 	ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
5168 	ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed,  8);
5169 	ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
5170 				     ttime->ttime_samples);
5171 }
5172 
5173 static void
5174 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5175 		       struct request *rq)
5176 {
5177 	bfqq->seek_history <<= 1;
5178 	bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
5179 
5180 	if (bfqq->wr_coeff > 1 &&
5181 	    bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
5182 	    BFQQ_TOTALLY_SEEKY(bfqq))
5183 		bfq_bfqq_end_wr(bfqq);
5184 }
5185 
5186 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
5187 				       struct bfq_queue *bfqq,
5188 				       struct bfq_io_cq *bic)
5189 {
5190 	bool has_short_ttime = true, state_changed;
5191 
5192 	/*
5193 	 * No need to update has_short_ttime if bfqq is async or in
5194 	 * idle io prio class, or if bfq_slice_idle is zero, because
5195 	 * no device idling is performed for bfqq in this case.
5196 	 */
5197 	if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
5198 	    bfqd->bfq_slice_idle == 0)
5199 		return;
5200 
5201 	/* Idle window just restored, statistics are meaningless. */
5202 	if (time_is_after_eq_jiffies(bfqq->split_time +
5203 				     bfqd->bfq_wr_min_idle_time))
5204 		return;
5205 
5206 	/* Think time is infinite if no process is linked to
5207 	 * bfqq. Otherwise check average think time to
5208 	 * decide whether to mark as has_short_ttime
5209 	 */
5210 	if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
5211 	    (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
5212 	     bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle))
5213 		has_short_ttime = false;
5214 
5215 	state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
5216 
5217 	if (has_short_ttime)
5218 		bfq_mark_bfqq_has_short_ttime(bfqq);
5219 	else
5220 		bfq_clear_bfqq_has_short_ttime(bfqq);
5221 
5222 	/*
5223 	 * Until the base value for the total service time gets
5224 	 * finally computed for bfqq, the inject limit does depend on
5225 	 * the think-time state (short|long). In particular, the limit
5226 	 * is 0 or 1 if the think time is deemed, respectively, as
5227 	 * short or long (details in the comments in
5228 	 * bfq_update_inject_limit()). Accordingly, the next
5229 	 * instructions reset the inject limit if the think-time state
5230 	 * has changed and the above base value is still to be
5231 	 * computed.
5232 	 *
5233 	 * However, the reset is performed only if more than 100 ms
5234 	 * have elapsed since the last update of the inject limit, or
5235 	 * (inclusive) if the change is from short to long think
5236 	 * time. The reason for this waiting is as follows.
5237 	 *
5238 	 * bfqq may have a long think time because of a
5239 	 * synchronization with some other queue, i.e., because the
5240 	 * I/O of some other queue may need to be completed for bfqq
5241 	 * to receive new I/O. Details in the comments on the choice
5242 	 * of the queue for injection in bfq_select_queue().
5243 	 *
5244 	 * As stressed in those comments, if such a synchronization is
5245 	 * actually in place, then, without injection on bfqq, the
5246 	 * blocking I/O cannot happen to served while bfqq is in
5247 	 * service. As a consequence, if bfqq is granted
5248 	 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
5249 	 * is dispatched, until the idle timeout fires. This is likely
5250 	 * to result in lower bandwidth and higher latencies for bfqq,
5251 	 * and in a severe loss of total throughput.
5252 	 *
5253 	 * On the opposite end, a non-zero inject limit may allow the
5254 	 * I/O that blocks bfqq to be executed soon, and therefore
5255 	 * bfqq to receive new I/O soon.
5256 	 *
5257 	 * But, if the blocking gets actually eliminated, then the
5258 	 * next think-time sample for bfqq may be very low. This in
5259 	 * turn may cause bfqq's think time to be deemed
5260 	 * short. Without the 100 ms barrier, this new state change
5261 	 * would cause the body of the next if to be executed
5262 	 * immediately. But this would set to 0 the inject
5263 	 * limit. Without injection, the blocking I/O would cause the
5264 	 * think time of bfqq to become long again, and therefore the
5265 	 * inject limit to be raised again, and so on. The only effect
5266 	 * of such a steady oscillation between the two think-time
5267 	 * states would be to prevent effective injection on bfqq.
5268 	 *
5269 	 * In contrast, if the inject limit is not reset during such a
5270 	 * long time interval as 100 ms, then the number of short
5271 	 * think time samples can grow significantly before the reset
5272 	 * is performed. As a consequence, the think time state can
5273 	 * become stable before the reset. Therefore there will be no
5274 	 * state change when the 100 ms elapse, and no reset of the
5275 	 * inject limit. The inject limit remains steadily equal to 1
5276 	 * both during and after the 100 ms. So injection can be
5277 	 * performed at all times, and throughput gets boosted.
5278 	 *
5279 	 * An inject limit equal to 1 is however in conflict, in
5280 	 * general, with the fact that the think time of bfqq is
5281 	 * short, because injection may be likely to delay bfqq's I/O
5282 	 * (as explained in the comments in
5283 	 * bfq_update_inject_limit()). But this does not happen in
5284 	 * this special case, because bfqq's low think time is due to
5285 	 * an effective handling of a synchronization, through
5286 	 * injection. In this special case, bfqq's I/O does not get
5287 	 * delayed by injection; on the contrary, bfqq's I/O is
5288 	 * brought forward, because it is not blocked for
5289 	 * milliseconds.
5290 	 *
5291 	 * In addition, serving the blocking I/O much sooner, and much
5292 	 * more frequently than once per I/O-plugging timeout, makes
5293 	 * it much quicker to detect a waker queue (the concept of
5294 	 * waker queue is defined in the comments in
5295 	 * bfq_add_request()). This makes it possible to start sooner
5296 	 * to boost throughput more effectively, by injecting the I/O
5297 	 * of the waker queue unconditionally on every
5298 	 * bfq_dispatch_request().
5299 	 *
5300 	 * One last, important benefit of not resetting the inject
5301 	 * limit before 100 ms is that, during this time interval, the
5302 	 * base value for the total service time is likely to get
5303 	 * finally computed for bfqq, freeing the inject limit from
5304 	 * its relation with the think time.
5305 	 */
5306 	if (state_changed && bfqq->last_serv_time_ns == 0 &&
5307 	    (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
5308 				      msecs_to_jiffies(100)) ||
5309 	     !has_short_ttime))
5310 		bfq_reset_inject_limit(bfqd, bfqq);
5311 }
5312 
5313 /*
5314  * Called when a new fs request (rq) is added to bfqq.  Check if there's
5315  * something we should do about it.
5316  */
5317 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5318 			    struct request *rq)
5319 {
5320 	if (rq->cmd_flags & REQ_META)
5321 		bfqq->meta_pending++;
5322 
5323 	bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
5324 
5325 	if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
5326 		bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
5327 				 blk_rq_sectors(rq) < 32;
5328 		bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
5329 
5330 		/*
5331 		 * There is just this request queued: if
5332 		 * - the request is small, and
5333 		 * - we are idling to boost throughput, and
5334 		 * - the queue is not to be expired,
5335 		 * then just exit.
5336 		 *
5337 		 * In this way, if the device is being idled to wait
5338 		 * for a new request from the in-service queue, we
5339 		 * avoid unplugging the device and committing the
5340 		 * device to serve just a small request. In contrast
5341 		 * we wait for the block layer to decide when to
5342 		 * unplug the device: hopefully, new requests will be
5343 		 * merged to this one quickly, then the device will be
5344 		 * unplugged and larger requests will be dispatched.
5345 		 */
5346 		if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
5347 		    !budget_timeout)
5348 			return;
5349 
5350 		/*
5351 		 * A large enough request arrived, or idling is being
5352 		 * performed to preserve service guarantees, or
5353 		 * finally the queue is to be expired: in all these
5354 		 * cases disk idling is to be stopped, so clear
5355 		 * wait_request flag and reset timer.
5356 		 */
5357 		bfq_clear_bfqq_wait_request(bfqq);
5358 		hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
5359 
5360 		/*
5361 		 * The queue is not empty, because a new request just
5362 		 * arrived. Hence we can safely expire the queue, in
5363 		 * case of budget timeout, without risking that the
5364 		 * timestamps of the queue are not updated correctly.
5365 		 * See [1] for more details.
5366 		 */
5367 		if (budget_timeout)
5368 			bfq_bfqq_expire(bfqd, bfqq, false,
5369 					BFQQE_BUDGET_TIMEOUT);
5370 	}
5371 }
5372 
5373 /* returns true if it causes the idle timer to be disabled */
5374 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
5375 {
5376 	struct bfq_queue *bfqq = RQ_BFQQ(rq),
5377 		*new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
5378 	bool waiting, idle_timer_disabled = false;
5379 
5380 	if (new_bfqq) {
5381 		/*
5382 		 * Release the request's reference to the old bfqq
5383 		 * and make sure one is taken to the shared queue.
5384 		 */
5385 		new_bfqq->allocated++;
5386 		bfqq->allocated--;
5387 		new_bfqq->ref++;
5388 		/*
5389 		 * If the bic associated with the process
5390 		 * issuing this request still points to bfqq
5391 		 * (and thus has not been already redirected
5392 		 * to new_bfqq or even some other bfq_queue),
5393 		 * then complete the merge and redirect it to
5394 		 * new_bfqq.
5395 		 */
5396 		if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
5397 			bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
5398 					bfqq, new_bfqq);
5399 
5400 		bfq_clear_bfqq_just_created(bfqq);
5401 		/*
5402 		 * rq is about to be enqueued into new_bfqq,
5403 		 * release rq reference on bfqq
5404 		 */
5405 		bfq_put_queue(bfqq);
5406 		rq->elv.priv[1] = new_bfqq;
5407 		bfqq = new_bfqq;
5408 	}
5409 
5410 	bfq_update_io_thinktime(bfqd, bfqq);
5411 	bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
5412 	bfq_update_io_seektime(bfqd, bfqq, rq);
5413 
5414 	waiting = bfqq && bfq_bfqq_wait_request(bfqq);
5415 	bfq_add_request(rq);
5416 	idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
5417 
5418 	rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
5419 	list_add_tail(&rq->queuelist, &bfqq->fifo);
5420 
5421 	bfq_rq_enqueued(bfqd, bfqq, rq);
5422 
5423 	return idle_timer_disabled;
5424 }
5425 
5426 #ifdef CONFIG_BFQ_CGROUP_DEBUG
5427 static void bfq_update_insert_stats(struct request_queue *q,
5428 				    struct bfq_queue *bfqq,
5429 				    bool idle_timer_disabled,
5430 				    unsigned int cmd_flags)
5431 {
5432 	if (!bfqq)
5433 		return;
5434 
5435 	/*
5436 	 * bfqq still exists, because it can disappear only after
5437 	 * either it is merged with another queue, or the process it
5438 	 * is associated with exits. But both actions must be taken by
5439 	 * the same process currently executing this flow of
5440 	 * instructions.
5441 	 *
5442 	 * In addition, the following queue lock guarantees that
5443 	 * bfqq_group(bfqq) exists as well.
5444 	 */
5445 	spin_lock_irq(&q->queue_lock);
5446 	bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
5447 	if (idle_timer_disabled)
5448 		bfqg_stats_update_idle_time(bfqq_group(bfqq));
5449 	spin_unlock_irq(&q->queue_lock);
5450 }
5451 #else
5452 static inline void bfq_update_insert_stats(struct request_queue *q,
5453 					   struct bfq_queue *bfqq,
5454 					   bool idle_timer_disabled,
5455 					   unsigned int cmd_flags) {}
5456 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
5457 
5458 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
5459 			       bool at_head)
5460 {
5461 	struct request_queue *q = hctx->queue;
5462 	struct bfq_data *bfqd = q->elevator->elevator_data;
5463 	struct bfq_queue *bfqq;
5464 	bool idle_timer_disabled = false;
5465 	unsigned int cmd_flags;
5466 
5467 	spin_lock_irq(&bfqd->lock);
5468 	if (blk_mq_sched_try_insert_merge(q, rq)) {
5469 		spin_unlock_irq(&bfqd->lock);
5470 		return;
5471 	}
5472 
5473 	spin_unlock_irq(&bfqd->lock);
5474 
5475 	blk_mq_sched_request_inserted(rq);
5476 
5477 	spin_lock_irq(&bfqd->lock);
5478 	bfqq = bfq_init_rq(rq);
5479 	if (!bfqq || at_head || blk_rq_is_passthrough(rq)) {
5480 		if (at_head)
5481 			list_add(&rq->queuelist, &bfqd->dispatch);
5482 		else
5483 			list_add_tail(&rq->queuelist, &bfqd->dispatch);
5484 	} else {
5485 		idle_timer_disabled = __bfq_insert_request(bfqd, rq);
5486 		/*
5487 		 * Update bfqq, because, if a queue merge has occurred
5488 		 * in __bfq_insert_request, then rq has been
5489 		 * redirected into a new queue.
5490 		 */
5491 		bfqq = RQ_BFQQ(rq);
5492 
5493 		if (rq_mergeable(rq)) {
5494 			elv_rqhash_add(q, rq);
5495 			if (!q->last_merge)
5496 				q->last_merge = rq;
5497 		}
5498 	}
5499 
5500 	/*
5501 	 * Cache cmd_flags before releasing scheduler lock, because rq
5502 	 * may disappear afterwards (for example, because of a request
5503 	 * merge).
5504 	 */
5505 	cmd_flags = rq->cmd_flags;
5506 
5507 	spin_unlock_irq(&bfqd->lock);
5508 
5509 	bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
5510 				cmd_flags);
5511 }
5512 
5513 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
5514 				struct list_head *list, bool at_head)
5515 {
5516 	while (!list_empty(list)) {
5517 		struct request *rq;
5518 
5519 		rq = list_first_entry(list, struct request, queuelist);
5520 		list_del_init(&rq->queuelist);
5521 		bfq_insert_request(hctx, rq, at_head);
5522 	}
5523 }
5524 
5525 static void bfq_update_hw_tag(struct bfq_data *bfqd)
5526 {
5527 	struct bfq_queue *bfqq = bfqd->in_service_queue;
5528 
5529 	bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
5530 				       bfqd->rq_in_driver);
5531 
5532 	if (bfqd->hw_tag == 1)
5533 		return;
5534 
5535 	/*
5536 	 * This sample is valid if the number of outstanding requests
5537 	 * is large enough to allow a queueing behavior.  Note that the
5538 	 * sum is not exact, as it's not taking into account deactivated
5539 	 * requests.
5540 	 */
5541 	if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
5542 		return;
5543 
5544 	/*
5545 	 * If active queue hasn't enough requests and can idle, bfq might not
5546 	 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
5547 	 * case
5548 	 */
5549 	if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
5550 	    bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
5551 	    BFQ_HW_QUEUE_THRESHOLD &&
5552 	    bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
5553 		return;
5554 
5555 	if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
5556 		return;
5557 
5558 	bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
5559 	bfqd->max_rq_in_driver = 0;
5560 	bfqd->hw_tag_samples = 0;
5561 
5562 	bfqd->nonrot_with_queueing =
5563 		blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
5564 }
5565 
5566 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
5567 {
5568 	u64 now_ns;
5569 	u32 delta_us;
5570 
5571 	bfq_update_hw_tag(bfqd);
5572 
5573 	bfqd->rq_in_driver--;
5574 	bfqq->dispatched--;
5575 
5576 	if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
5577 		/*
5578 		 * Set budget_timeout (which we overload to store the
5579 		 * time at which the queue remains with no backlog and
5580 		 * no outstanding request; used by the weight-raising
5581 		 * mechanism).
5582 		 */
5583 		bfqq->budget_timeout = jiffies;
5584 
5585 		bfq_weights_tree_remove(bfqd, bfqq);
5586 	}
5587 
5588 	now_ns = ktime_get_ns();
5589 
5590 	bfqq->ttime.last_end_request = now_ns;
5591 
5592 	/*
5593 	 * Using us instead of ns, to get a reasonable precision in
5594 	 * computing rate in next check.
5595 	 */
5596 	delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
5597 
5598 	/*
5599 	 * If the request took rather long to complete, and, according
5600 	 * to the maximum request size recorded, this completion latency
5601 	 * implies that the request was certainly served at a very low
5602 	 * rate (less than 1M sectors/sec), then the whole observation
5603 	 * interval that lasts up to this time instant cannot be a
5604 	 * valid time interval for computing a new peak rate.  Invoke
5605 	 * bfq_update_rate_reset to have the following three steps
5606 	 * taken:
5607 	 * - close the observation interval at the last (previous)
5608 	 *   request dispatch or completion
5609 	 * - compute rate, if possible, for that observation interval
5610 	 * - reset to zero samples, which will trigger a proper
5611 	 *   re-initialization of the observation interval on next
5612 	 *   dispatch
5613 	 */
5614 	if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
5615 	   (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
5616 			1UL<<(BFQ_RATE_SHIFT - 10))
5617 		bfq_update_rate_reset(bfqd, NULL);
5618 	bfqd->last_completion = now_ns;
5619 	bfqd->last_completed_rq_bfqq = bfqq;
5620 
5621 	/*
5622 	 * If we are waiting to discover whether the request pattern
5623 	 * of the task associated with the queue is actually
5624 	 * isochronous, and both requisites for this condition to hold
5625 	 * are now satisfied, then compute soft_rt_next_start (see the
5626 	 * comments on the function bfq_bfqq_softrt_next_start()). We
5627 	 * do not compute soft_rt_next_start if bfqq is in interactive
5628 	 * weight raising (see the comments in bfq_bfqq_expire() for
5629 	 * an explanation). We schedule this delayed update when bfqq
5630 	 * expires, if it still has in-flight requests.
5631 	 */
5632 	if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
5633 	    RB_EMPTY_ROOT(&bfqq->sort_list) &&
5634 	    bfqq->wr_coeff != bfqd->bfq_wr_coeff)
5635 		bfqq->soft_rt_next_start =
5636 			bfq_bfqq_softrt_next_start(bfqd, bfqq);
5637 
5638 	/*
5639 	 * If this is the in-service queue, check if it needs to be expired,
5640 	 * or if we want to idle in case it has no pending requests.
5641 	 */
5642 	if (bfqd->in_service_queue == bfqq) {
5643 		if (bfq_bfqq_must_idle(bfqq)) {
5644 			if (bfqq->dispatched == 0)
5645 				bfq_arm_slice_timer(bfqd);
5646 			/*
5647 			 * If we get here, we do not expire bfqq, even
5648 			 * if bfqq was in budget timeout or had no
5649 			 * more requests (as controlled in the next
5650 			 * conditional instructions). The reason for
5651 			 * not expiring bfqq is as follows.
5652 			 *
5653 			 * Here bfqq->dispatched > 0 holds, but
5654 			 * bfq_bfqq_must_idle() returned true. This
5655 			 * implies that, even if no request arrives
5656 			 * for bfqq before bfqq->dispatched reaches 0,
5657 			 * bfqq will, however, not be expired on the
5658 			 * completion event that causes bfqq->dispatch
5659 			 * to reach zero. In contrast, on this event,
5660 			 * bfqq will start enjoying device idling
5661 			 * (I/O-dispatch plugging).
5662 			 *
5663 			 * But, if we expired bfqq here, bfqq would
5664 			 * not have the chance to enjoy device idling
5665 			 * when bfqq->dispatched finally reaches
5666 			 * zero. This would expose bfqq to violation
5667 			 * of its reserved service guarantees.
5668 			 */
5669 			return;
5670 		} else if (bfq_may_expire_for_budg_timeout(bfqq))
5671 			bfq_bfqq_expire(bfqd, bfqq, false,
5672 					BFQQE_BUDGET_TIMEOUT);
5673 		else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
5674 			 (bfqq->dispatched == 0 ||
5675 			  !bfq_better_to_idle(bfqq)))
5676 			bfq_bfqq_expire(bfqd, bfqq, false,
5677 					BFQQE_NO_MORE_REQUESTS);
5678 	}
5679 
5680 	if (!bfqd->rq_in_driver)
5681 		bfq_schedule_dispatch(bfqd);
5682 }
5683 
5684 static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq)
5685 {
5686 	bfqq->allocated--;
5687 
5688 	bfq_put_queue(bfqq);
5689 }
5690 
5691 /*
5692  * The processes associated with bfqq may happen to generate their
5693  * cumulative I/O at a lower rate than the rate at which the device
5694  * could serve the same I/O. This is rather probable, e.g., if only
5695  * one process is associated with bfqq and the device is an SSD. It
5696  * results in bfqq becoming often empty while in service. In this
5697  * respect, if BFQ is allowed to switch to another queue when bfqq
5698  * remains empty, then the device goes on being fed with I/O requests,
5699  * and the throughput is not affected. In contrast, if BFQ is not
5700  * allowed to switch to another queue---because bfqq is sync and
5701  * I/O-dispatch needs to be plugged while bfqq is temporarily
5702  * empty---then, during the service of bfqq, there will be frequent
5703  * "service holes", i.e., time intervals during which bfqq gets empty
5704  * and the device can only consume the I/O already queued in its
5705  * hardware queues. During service holes, the device may even get to
5706  * remaining idle. In the end, during the service of bfqq, the device
5707  * is driven at a lower speed than the one it can reach with the kind
5708  * of I/O flowing through bfqq.
5709  *
5710  * To counter this loss of throughput, BFQ implements a "request
5711  * injection mechanism", which tries to fill the above service holes
5712  * with I/O requests taken from other queues. The hard part in this
5713  * mechanism is finding the right amount of I/O to inject, so as to
5714  * both boost throughput and not break bfqq's bandwidth and latency
5715  * guarantees. In this respect, the mechanism maintains a per-queue
5716  * inject limit, computed as below. While bfqq is empty, the injection
5717  * mechanism dispatches extra I/O requests only until the total number
5718  * of I/O requests in flight---i.e., already dispatched but not yet
5719  * completed---remains lower than this limit.
5720  *
5721  * A first definition comes in handy to introduce the algorithm by
5722  * which the inject limit is computed.  We define as first request for
5723  * bfqq, an I/O request for bfqq that arrives while bfqq is in
5724  * service, and causes bfqq to switch from empty to non-empty. The
5725  * algorithm updates the limit as a function of the effect of
5726  * injection on the service times of only the first requests of
5727  * bfqq. The reason for this restriction is that these are the
5728  * requests whose service time is affected most, because they are the
5729  * first to arrive after injection possibly occurred.
5730  *
5731  * To evaluate the effect of injection, the algorithm measures the
5732  * "total service time" of first requests. We define as total service
5733  * time of an I/O request, the time that elapses since when the
5734  * request is enqueued into bfqq, to when it is completed. This
5735  * quantity allows the whole effect of injection to be measured. It is
5736  * easy to see why. Suppose that some requests of other queues are
5737  * actually injected while bfqq is empty, and that a new request R
5738  * then arrives for bfqq. If the device does start to serve all or
5739  * part of the injected requests during the service hole, then,
5740  * because of this extra service, it may delay the next invocation of
5741  * the dispatch hook of BFQ. Then, even after R gets eventually
5742  * dispatched, the device may delay the actual service of R if it is
5743  * still busy serving the extra requests, or if it decides to serve,
5744  * before R, some extra request still present in its queues. As a
5745  * conclusion, the cumulative extra delay caused by injection can be
5746  * easily evaluated by just comparing the total service time of first
5747  * requests with and without injection.
5748  *
5749  * The limit-update algorithm works as follows. On the arrival of a
5750  * first request of bfqq, the algorithm measures the total time of the
5751  * request only if one of the three cases below holds, and, for each
5752  * case, it updates the limit as described below:
5753  *
5754  * (1) If there is no in-flight request. This gives a baseline for the
5755  *     total service time of the requests of bfqq. If the baseline has
5756  *     not been computed yet, then, after computing it, the limit is
5757  *     set to 1, to start boosting throughput, and to prepare the
5758  *     ground for the next case. If the baseline has already been
5759  *     computed, then it is updated, in case it results to be lower
5760  *     than the previous value.
5761  *
5762  * (2) If the limit is higher than 0 and there are in-flight
5763  *     requests. By comparing the total service time in this case with
5764  *     the above baseline, it is possible to know at which extent the
5765  *     current value of the limit is inflating the total service
5766  *     time. If the inflation is below a certain threshold, then bfqq
5767  *     is assumed to be suffering from no perceivable loss of its
5768  *     service guarantees, and the limit is even tentatively
5769  *     increased. If the inflation is above the threshold, then the
5770  *     limit is decreased. Due to the lack of any hysteresis, this
5771  *     logic makes the limit oscillate even in steady workload
5772  *     conditions. Yet we opted for it, because it is fast in reaching
5773  *     the best value for the limit, as a function of the current I/O
5774  *     workload. To reduce oscillations, this step is disabled for a
5775  *     short time interval after the limit happens to be decreased.
5776  *
5777  * (3) Periodically, after resetting the limit, to make sure that the
5778  *     limit eventually drops in case the workload changes. This is
5779  *     needed because, after the limit has gone safely up for a
5780  *     certain workload, it is impossible to guess whether the
5781  *     baseline total service time may have changed, without measuring
5782  *     it again without injection. A more effective version of this
5783  *     step might be to just sample the baseline, by interrupting
5784  *     injection only once, and then to reset/lower the limit only if
5785  *     the total service time with the current limit does happen to be
5786  *     too large.
5787  *
5788  * More details on each step are provided in the comments on the
5789  * pieces of code that implement these steps: the branch handling the
5790  * transition from empty to non empty in bfq_add_request(), the branch
5791  * handling injection in bfq_select_queue(), and the function
5792  * bfq_choose_bfqq_for_injection(). These comments also explain some
5793  * exceptions, made by the injection mechanism in some special cases.
5794  */
5795 static void bfq_update_inject_limit(struct bfq_data *bfqd,
5796 				    struct bfq_queue *bfqq)
5797 {
5798 	u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns;
5799 	unsigned int old_limit = bfqq->inject_limit;
5800 
5801 	if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) {
5802 		u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
5803 
5804 		if (tot_time_ns >= threshold && old_limit > 0) {
5805 			bfqq->inject_limit--;
5806 			bfqq->decrease_time_jif = jiffies;
5807 		} else if (tot_time_ns < threshold &&
5808 			   old_limit <= bfqd->max_rq_in_driver)
5809 			bfqq->inject_limit++;
5810 	}
5811 
5812 	/*
5813 	 * Either we still have to compute the base value for the
5814 	 * total service time, and there seem to be the right
5815 	 * conditions to do it, or we can lower the last base value
5816 	 * computed.
5817 	 *
5818 	 * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
5819 	 * request in flight, because this function is in the code
5820 	 * path that handles the completion of a request of bfqq, and,
5821 	 * in particular, this function is executed before
5822 	 * bfqd->rq_in_driver is decremented in such a code path.
5823 	 */
5824 	if ((bfqq->last_serv_time_ns == 0 && bfqd->rq_in_driver == 1) ||
5825 	    tot_time_ns < bfqq->last_serv_time_ns) {
5826 		if (bfqq->last_serv_time_ns == 0) {
5827 			/*
5828 			 * Now we certainly have a base value: make sure we
5829 			 * start trying injection.
5830 			 */
5831 			bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
5832 		}
5833 		bfqq->last_serv_time_ns = tot_time_ns;
5834 	} else if (!bfqd->rqs_injected && bfqd->rq_in_driver == 1)
5835 		/*
5836 		 * No I/O injected and no request still in service in
5837 		 * the drive: these are the exact conditions for
5838 		 * computing the base value of the total service time
5839 		 * for bfqq. So let's update this value, because it is
5840 		 * rather variable. For example, it varies if the size
5841 		 * or the spatial locality of the I/O requests in bfqq
5842 		 * change.
5843 		 */
5844 		bfqq->last_serv_time_ns = tot_time_ns;
5845 
5846 
5847 	/* update complete, not waiting for any request completion any longer */
5848 	bfqd->waited_rq = NULL;
5849 	bfqd->rqs_injected = false;
5850 }
5851 
5852 /*
5853  * Handle either a requeue or a finish for rq. The things to do are
5854  * the same in both cases: all references to rq are to be dropped. In
5855  * particular, rq is considered completed from the point of view of
5856  * the scheduler.
5857  */
5858 static void bfq_finish_requeue_request(struct request *rq)
5859 {
5860 	struct bfq_queue *bfqq = RQ_BFQQ(rq);
5861 	struct bfq_data *bfqd;
5862 
5863 	/*
5864 	 * Requeue and finish hooks are invoked in blk-mq without
5865 	 * checking whether the involved request is actually still
5866 	 * referenced in the scheduler. To handle this fact, the
5867 	 * following two checks make this function exit in case of
5868 	 * spurious invocations, for which there is nothing to do.
5869 	 *
5870 	 * First, check whether rq has nothing to do with an elevator.
5871 	 */
5872 	if (unlikely(!(rq->rq_flags & RQF_ELVPRIV)))
5873 		return;
5874 
5875 	/*
5876 	 * rq either is not associated with any icq, or is an already
5877 	 * requeued request that has not (yet) been re-inserted into
5878 	 * a bfq_queue.
5879 	 */
5880 	if (!rq->elv.icq || !bfqq)
5881 		return;
5882 
5883 	bfqd = bfqq->bfqd;
5884 
5885 	if (rq->rq_flags & RQF_STARTED)
5886 		bfqg_stats_update_completion(bfqq_group(bfqq),
5887 					     rq->start_time_ns,
5888 					     rq->io_start_time_ns,
5889 					     rq->cmd_flags);
5890 
5891 	if (likely(rq->rq_flags & RQF_STARTED)) {
5892 		unsigned long flags;
5893 
5894 		spin_lock_irqsave(&bfqd->lock, flags);
5895 
5896 		if (rq == bfqd->waited_rq)
5897 			bfq_update_inject_limit(bfqd, bfqq);
5898 
5899 		bfq_completed_request(bfqq, bfqd);
5900 		bfq_finish_requeue_request_body(bfqq);
5901 
5902 		spin_unlock_irqrestore(&bfqd->lock, flags);
5903 	} else {
5904 		/*
5905 		 * Request rq may be still/already in the scheduler,
5906 		 * in which case we need to remove it (this should
5907 		 * never happen in case of requeue). And we cannot
5908 		 * defer such a check and removal, to avoid
5909 		 * inconsistencies in the time interval from the end
5910 		 * of this function to the start of the deferred work.
5911 		 * This situation seems to occur only in process
5912 		 * context, as a consequence of a merge. In the
5913 		 * current version of the code, this implies that the
5914 		 * lock is held.
5915 		 */
5916 
5917 		if (!RB_EMPTY_NODE(&rq->rb_node)) {
5918 			bfq_remove_request(rq->q, rq);
5919 			bfqg_stats_update_io_remove(bfqq_group(bfqq),
5920 						    rq->cmd_flags);
5921 		}
5922 		bfq_finish_requeue_request_body(bfqq);
5923 	}
5924 
5925 	/*
5926 	 * Reset private fields. In case of a requeue, this allows
5927 	 * this function to correctly do nothing if it is spuriously
5928 	 * invoked again on this same request (see the check at the
5929 	 * beginning of the function). Probably, a better general
5930 	 * design would be to prevent blk-mq from invoking the requeue
5931 	 * or finish hooks of an elevator, for a request that is not
5932 	 * referred by that elevator.
5933 	 *
5934 	 * Resetting the following fields would break the
5935 	 * request-insertion logic if rq is re-inserted into a bfq
5936 	 * internal queue, without a re-preparation. Here we assume
5937 	 * that re-insertions of requeued requests, without
5938 	 * re-preparation, can happen only for pass_through or at_head
5939 	 * requests (which are not re-inserted into bfq internal
5940 	 * queues).
5941 	 */
5942 	rq->elv.priv[0] = NULL;
5943 	rq->elv.priv[1] = NULL;
5944 }
5945 
5946 /*
5947  * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
5948  * was the last process referring to that bfqq.
5949  */
5950 static struct bfq_queue *
5951 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
5952 {
5953 	bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
5954 
5955 	if (bfqq_process_refs(bfqq) == 1) {
5956 		bfqq->pid = current->pid;
5957 		bfq_clear_bfqq_coop(bfqq);
5958 		bfq_clear_bfqq_split_coop(bfqq);
5959 		return bfqq;
5960 	}
5961 
5962 	bic_set_bfqq(bic, NULL, 1);
5963 
5964 	bfq_put_cooperator(bfqq);
5965 
5966 	bfq_put_queue(bfqq);
5967 	return NULL;
5968 }
5969 
5970 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
5971 						   struct bfq_io_cq *bic,
5972 						   struct bio *bio,
5973 						   bool split, bool is_sync,
5974 						   bool *new_queue)
5975 {
5976 	struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
5977 
5978 	if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
5979 		return bfqq;
5980 
5981 	if (new_queue)
5982 		*new_queue = true;
5983 
5984 	if (bfqq)
5985 		bfq_put_queue(bfqq);
5986 	bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
5987 
5988 	bic_set_bfqq(bic, bfqq, is_sync);
5989 	if (split && is_sync) {
5990 		if ((bic->was_in_burst_list && bfqd->large_burst) ||
5991 		    bic->saved_in_large_burst)
5992 			bfq_mark_bfqq_in_large_burst(bfqq);
5993 		else {
5994 			bfq_clear_bfqq_in_large_burst(bfqq);
5995 			if (bic->was_in_burst_list)
5996 				/*
5997 				 * If bfqq was in the current
5998 				 * burst list before being
5999 				 * merged, then we have to add
6000 				 * it back. And we do not need
6001 				 * to increase burst_size, as
6002 				 * we did not decrement
6003 				 * burst_size when we removed
6004 				 * bfqq from the burst list as
6005 				 * a consequence of a merge
6006 				 * (see comments in
6007 				 * bfq_put_queue). In this
6008 				 * respect, it would be rather
6009 				 * costly to know whether the
6010 				 * current burst list is still
6011 				 * the same burst list from
6012 				 * which bfqq was removed on
6013 				 * the merge. To avoid this
6014 				 * cost, if bfqq was in a
6015 				 * burst list, then we add
6016 				 * bfqq to the current burst
6017 				 * list without any further
6018 				 * check. This can cause
6019 				 * inappropriate insertions,
6020 				 * but rarely enough to not
6021 				 * harm the detection of large
6022 				 * bursts significantly.
6023 				 */
6024 				hlist_add_head(&bfqq->burst_list_node,
6025 					       &bfqd->burst_list);
6026 		}
6027 		bfqq->split_time = jiffies;
6028 	}
6029 
6030 	return bfqq;
6031 }
6032 
6033 /*
6034  * Only reset private fields. The actual request preparation will be
6035  * performed by bfq_init_rq, when rq is either inserted or merged. See
6036  * comments on bfq_init_rq for the reason behind this delayed
6037  * preparation.
6038  */
6039 static void bfq_prepare_request(struct request *rq, struct bio *bio)
6040 {
6041 	/*
6042 	 * Regardless of whether we have an icq attached, we have to
6043 	 * clear the scheduler pointers, as they might point to
6044 	 * previously allocated bic/bfqq structs.
6045 	 */
6046 	rq->elv.priv[0] = rq->elv.priv[1] = NULL;
6047 }
6048 
6049 /*
6050  * If needed, init rq, allocate bfq data structures associated with
6051  * rq, and increment reference counters in the destination bfq_queue
6052  * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6053  * not associated with any bfq_queue.
6054  *
6055  * This function is invoked by the functions that perform rq insertion
6056  * or merging. One may have expected the above preparation operations
6057  * to be performed in bfq_prepare_request, and not delayed to when rq
6058  * is inserted or merged. The rationale behind this delayed
6059  * preparation is that, after the prepare_request hook is invoked for
6060  * rq, rq may still be transformed into a request with no icq, i.e., a
6061  * request not associated with any queue. No bfq hook is invoked to
6062  * signal this transformation. As a consequence, should these
6063  * preparation operations be performed when the prepare_request hook
6064  * is invoked, and should rq be transformed one moment later, bfq
6065  * would end up in an inconsistent state, because it would have
6066  * incremented some queue counters for an rq destined to
6067  * transformation, without any chance to correctly lower these
6068  * counters back. In contrast, no transformation can still happen for
6069  * rq after rq has been inserted or merged. So, it is safe to execute
6070  * these preparation operations when rq is finally inserted or merged.
6071  */
6072 static struct bfq_queue *bfq_init_rq(struct request *rq)
6073 {
6074 	struct request_queue *q = rq->q;
6075 	struct bio *bio = rq->bio;
6076 	struct bfq_data *bfqd = q->elevator->elevator_data;
6077 	struct bfq_io_cq *bic;
6078 	const int is_sync = rq_is_sync(rq);
6079 	struct bfq_queue *bfqq;
6080 	bool new_queue = false;
6081 	bool bfqq_already_existing = false, split = false;
6082 
6083 	if (unlikely(!rq->elv.icq))
6084 		return NULL;
6085 
6086 	/*
6087 	 * Assuming that elv.priv[1] is set only if everything is set
6088 	 * for this rq. This holds true, because this function is
6089 	 * invoked only for insertion or merging, and, after such
6090 	 * events, a request cannot be manipulated any longer before
6091 	 * being removed from bfq.
6092 	 */
6093 	if (rq->elv.priv[1])
6094 		return rq->elv.priv[1];
6095 
6096 	bic = icq_to_bic(rq->elv.icq);
6097 
6098 	bfq_check_ioprio_change(bic, bio);
6099 
6100 	bfq_bic_update_cgroup(bic, bio);
6101 
6102 	bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
6103 					 &new_queue);
6104 
6105 	if (likely(!new_queue)) {
6106 		/* If the queue was seeky for too long, break it apart. */
6107 		if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
6108 			bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
6109 
6110 			/* Update bic before losing reference to bfqq */
6111 			if (bfq_bfqq_in_large_burst(bfqq))
6112 				bic->saved_in_large_burst = true;
6113 
6114 			bfqq = bfq_split_bfqq(bic, bfqq);
6115 			split = true;
6116 
6117 			if (!bfqq)
6118 				bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
6119 								 true, is_sync,
6120 								 NULL);
6121 			else
6122 				bfqq_already_existing = true;
6123 		}
6124 	}
6125 
6126 	bfqq->allocated++;
6127 	bfqq->ref++;
6128 	bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
6129 		     rq, bfqq, bfqq->ref);
6130 
6131 	rq->elv.priv[0] = bic;
6132 	rq->elv.priv[1] = bfqq;
6133 
6134 	/*
6135 	 * If a bfq_queue has only one process reference, it is owned
6136 	 * by only this bic: we can then set bfqq->bic = bic. in
6137 	 * addition, if the queue has also just been split, we have to
6138 	 * resume its state.
6139 	 */
6140 	if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
6141 		bfqq->bic = bic;
6142 		if (split) {
6143 			/*
6144 			 * The queue has just been split from a shared
6145 			 * queue: restore the idle window and the
6146 			 * possible weight raising period.
6147 			 */
6148 			bfq_bfqq_resume_state(bfqq, bfqd, bic,
6149 					      bfqq_already_existing);
6150 		}
6151 	}
6152 
6153 	/*
6154 	 * Consider bfqq as possibly belonging to a burst of newly
6155 	 * created queues only if:
6156 	 * 1) A burst is actually happening (bfqd->burst_size > 0)
6157 	 * or
6158 	 * 2) There is no other active queue. In fact, if, in
6159 	 *    contrast, there are active queues not belonging to the
6160 	 *    possible burst bfqq may belong to, then there is no gain
6161 	 *    in considering bfqq as belonging to a burst, and
6162 	 *    therefore in not weight-raising bfqq. See comments on
6163 	 *    bfq_handle_burst().
6164 	 *
6165 	 * This filtering also helps eliminating false positives,
6166 	 * occurring when bfqq does not belong to an actual large
6167 	 * burst, but some background task (e.g., a service) happens
6168 	 * to trigger the creation of new queues very close to when
6169 	 * bfqq and its possible companion queues are created. See
6170 	 * comments on bfq_handle_burst() for further details also on
6171 	 * this issue.
6172 	 */
6173 	if (unlikely(bfq_bfqq_just_created(bfqq) &&
6174 		     (bfqd->burst_size > 0 ||
6175 		      bfq_tot_busy_queues(bfqd) == 0)))
6176 		bfq_handle_burst(bfqd, bfqq);
6177 
6178 	return bfqq;
6179 }
6180 
6181 static void bfq_idle_slice_timer_body(struct bfq_queue *bfqq)
6182 {
6183 	struct bfq_data *bfqd = bfqq->bfqd;
6184 	enum bfqq_expiration reason;
6185 	unsigned long flags;
6186 
6187 	spin_lock_irqsave(&bfqd->lock, flags);
6188 	bfq_clear_bfqq_wait_request(bfqq);
6189 
6190 	if (bfqq != bfqd->in_service_queue) {
6191 		spin_unlock_irqrestore(&bfqd->lock, flags);
6192 		return;
6193 	}
6194 
6195 	if (bfq_bfqq_budget_timeout(bfqq))
6196 		/*
6197 		 * Also here the queue can be safely expired
6198 		 * for budget timeout without wasting
6199 		 * guarantees
6200 		 */
6201 		reason = BFQQE_BUDGET_TIMEOUT;
6202 	else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
6203 		/*
6204 		 * The queue may not be empty upon timer expiration,
6205 		 * because we may not disable the timer when the
6206 		 * first request of the in-service queue arrives
6207 		 * during disk idling.
6208 		 */
6209 		reason = BFQQE_TOO_IDLE;
6210 	else
6211 		goto schedule_dispatch;
6212 
6213 	bfq_bfqq_expire(bfqd, bfqq, true, reason);
6214 
6215 schedule_dispatch:
6216 	spin_unlock_irqrestore(&bfqd->lock, flags);
6217 	bfq_schedule_dispatch(bfqd);
6218 }
6219 
6220 /*
6221  * Handler of the expiration of the timer running if the in-service queue
6222  * is idling inside its time slice.
6223  */
6224 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
6225 {
6226 	struct bfq_data *bfqd = container_of(timer, struct bfq_data,
6227 					     idle_slice_timer);
6228 	struct bfq_queue *bfqq = bfqd->in_service_queue;
6229 
6230 	/*
6231 	 * Theoretical race here: the in-service queue can be NULL or
6232 	 * different from the queue that was idling if a new request
6233 	 * arrives for the current queue and there is a full dispatch
6234 	 * cycle that changes the in-service queue.  This can hardly
6235 	 * happen, but in the worst case we just expire a queue too
6236 	 * early.
6237 	 */
6238 	if (bfqq)
6239 		bfq_idle_slice_timer_body(bfqq);
6240 
6241 	return HRTIMER_NORESTART;
6242 }
6243 
6244 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
6245 				 struct bfq_queue **bfqq_ptr)
6246 {
6247 	struct bfq_queue *bfqq = *bfqq_ptr;
6248 
6249 	bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
6250 	if (bfqq) {
6251 		bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
6252 
6253 		bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
6254 			     bfqq, bfqq->ref);
6255 		bfq_put_queue(bfqq);
6256 		*bfqq_ptr = NULL;
6257 	}
6258 }
6259 
6260 /*
6261  * Release all the bfqg references to its async queues.  If we are
6262  * deallocating the group these queues may still contain requests, so
6263  * we reparent them to the root cgroup (i.e., the only one that will
6264  * exist for sure until all the requests on a device are gone).
6265  */
6266 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
6267 {
6268 	int i, j;
6269 
6270 	for (i = 0; i < 2; i++)
6271 		for (j = 0; j < IOPRIO_BE_NR; j++)
6272 			__bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
6273 
6274 	__bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
6275 }
6276 
6277 /*
6278  * See the comments on bfq_limit_depth for the purpose of
6279  * the depths set in the function. Return minimum shallow depth we'll use.
6280  */
6281 static unsigned int bfq_update_depths(struct bfq_data *bfqd,
6282 				      struct sbitmap_queue *bt)
6283 {
6284 	unsigned int i, j, min_shallow = UINT_MAX;
6285 
6286 	/*
6287 	 * In-word depths if no bfq_queue is being weight-raised:
6288 	 * leaving 25% of tags only for sync reads.
6289 	 *
6290 	 * In next formulas, right-shift the value
6291 	 * (1U<<bt->sb.shift), instead of computing directly
6292 	 * (1U<<(bt->sb.shift - something)), to be robust against
6293 	 * any possible value of bt->sb.shift, without having to
6294 	 * limit 'something'.
6295 	 */
6296 	/* no more than 50% of tags for async I/O */
6297 	bfqd->word_depths[0][0] = max((1U << bt->sb.shift) >> 1, 1U);
6298 	/*
6299 	 * no more than 75% of tags for sync writes (25% extra tags
6300 	 * w.r.t. async I/O, to prevent async I/O from starving sync
6301 	 * writes)
6302 	 */
6303 	bfqd->word_depths[0][1] = max(((1U << bt->sb.shift) * 3) >> 2, 1U);
6304 
6305 	/*
6306 	 * In-word depths in case some bfq_queue is being weight-
6307 	 * raised: leaving ~63% of tags for sync reads. This is the
6308 	 * highest percentage for which, in our tests, application
6309 	 * start-up times didn't suffer from any regression due to tag
6310 	 * shortage.
6311 	 */
6312 	/* no more than ~18% of tags for async I/O */
6313 	bfqd->word_depths[1][0] = max(((1U << bt->sb.shift) * 3) >> 4, 1U);
6314 	/* no more than ~37% of tags for sync writes (~20% extra tags) */
6315 	bfqd->word_depths[1][1] = max(((1U << bt->sb.shift) * 6) >> 4, 1U);
6316 
6317 	for (i = 0; i < 2; i++)
6318 		for (j = 0; j < 2; j++)
6319 			min_shallow = min(min_shallow, bfqd->word_depths[i][j]);
6320 
6321 	return min_shallow;
6322 }
6323 
6324 static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
6325 {
6326 	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
6327 	struct blk_mq_tags *tags = hctx->sched_tags;
6328 	unsigned int min_shallow;
6329 
6330 	min_shallow = bfq_update_depths(bfqd, &tags->bitmap_tags);
6331 	sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, min_shallow);
6332 }
6333 
6334 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
6335 {
6336 	bfq_depth_updated(hctx);
6337 	return 0;
6338 }
6339 
6340 static void bfq_exit_queue(struct elevator_queue *e)
6341 {
6342 	struct bfq_data *bfqd = e->elevator_data;
6343 	struct bfq_queue *bfqq, *n;
6344 
6345 	hrtimer_cancel(&bfqd->idle_slice_timer);
6346 
6347 	spin_lock_irq(&bfqd->lock);
6348 	list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
6349 		bfq_deactivate_bfqq(bfqd, bfqq, false, false);
6350 	spin_unlock_irq(&bfqd->lock);
6351 
6352 	hrtimer_cancel(&bfqd->idle_slice_timer);
6353 
6354 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6355 	/* release oom-queue reference to root group */
6356 	bfqg_and_blkg_put(bfqd->root_group);
6357 
6358 	blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
6359 #else
6360 	spin_lock_irq(&bfqd->lock);
6361 	bfq_put_async_queues(bfqd, bfqd->root_group);
6362 	kfree(bfqd->root_group);
6363 	spin_unlock_irq(&bfqd->lock);
6364 #endif
6365 
6366 	kfree(bfqd);
6367 }
6368 
6369 static void bfq_init_root_group(struct bfq_group *root_group,
6370 				struct bfq_data *bfqd)
6371 {
6372 	int i;
6373 
6374 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6375 	root_group->entity.parent = NULL;
6376 	root_group->my_entity = NULL;
6377 	root_group->bfqd = bfqd;
6378 #endif
6379 	root_group->rq_pos_tree = RB_ROOT;
6380 	for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
6381 		root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
6382 	root_group->sched_data.bfq_class_idle_last_service = jiffies;
6383 }
6384 
6385 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
6386 {
6387 	struct bfq_data *bfqd;
6388 	struct elevator_queue *eq;
6389 
6390 	eq = elevator_alloc(q, e);
6391 	if (!eq)
6392 		return -ENOMEM;
6393 
6394 	bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
6395 	if (!bfqd) {
6396 		kobject_put(&eq->kobj);
6397 		return -ENOMEM;
6398 	}
6399 	eq->elevator_data = bfqd;
6400 
6401 	spin_lock_irq(&q->queue_lock);
6402 	q->elevator = eq;
6403 	spin_unlock_irq(&q->queue_lock);
6404 
6405 	/*
6406 	 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
6407 	 * Grab a permanent reference to it, so that the normal code flow
6408 	 * will not attempt to free it.
6409 	 */
6410 	bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
6411 	bfqd->oom_bfqq.ref++;
6412 	bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
6413 	bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
6414 	bfqd->oom_bfqq.entity.new_weight =
6415 		bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
6416 
6417 	/* oom_bfqq does not participate to bursts */
6418 	bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
6419 
6420 	/*
6421 	 * Trigger weight initialization, according to ioprio, at the
6422 	 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
6423 	 * class won't be changed any more.
6424 	 */
6425 	bfqd->oom_bfqq.entity.prio_changed = 1;
6426 
6427 	bfqd->queue = q;
6428 
6429 	INIT_LIST_HEAD(&bfqd->dispatch);
6430 
6431 	hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
6432 		     HRTIMER_MODE_REL);
6433 	bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
6434 
6435 	bfqd->queue_weights_tree = RB_ROOT_CACHED;
6436 	bfqd->num_groups_with_pending_reqs = 0;
6437 
6438 	INIT_LIST_HEAD(&bfqd->active_list);
6439 	INIT_LIST_HEAD(&bfqd->idle_list);
6440 	INIT_HLIST_HEAD(&bfqd->burst_list);
6441 
6442 	bfqd->hw_tag = -1;
6443 	bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
6444 
6445 	bfqd->bfq_max_budget = bfq_default_max_budget;
6446 
6447 	bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
6448 	bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
6449 	bfqd->bfq_back_max = bfq_back_max;
6450 	bfqd->bfq_back_penalty = bfq_back_penalty;
6451 	bfqd->bfq_slice_idle = bfq_slice_idle;
6452 	bfqd->bfq_timeout = bfq_timeout;
6453 
6454 	bfqd->bfq_requests_within_timer = 120;
6455 
6456 	bfqd->bfq_large_burst_thresh = 8;
6457 	bfqd->bfq_burst_interval = msecs_to_jiffies(180);
6458 
6459 	bfqd->low_latency = true;
6460 
6461 	/*
6462 	 * Trade-off between responsiveness and fairness.
6463 	 */
6464 	bfqd->bfq_wr_coeff = 30;
6465 	bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
6466 	bfqd->bfq_wr_max_time = 0;
6467 	bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
6468 	bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
6469 	bfqd->bfq_wr_max_softrt_rate = 7000; /*
6470 					      * Approximate rate required
6471 					      * to playback or record a
6472 					      * high-definition compressed
6473 					      * video.
6474 					      */
6475 	bfqd->wr_busy_queues = 0;
6476 
6477 	/*
6478 	 * Begin by assuming, optimistically, that the device peak
6479 	 * rate is equal to 2/3 of the highest reference rate.
6480 	 */
6481 	bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
6482 		ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
6483 	bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
6484 
6485 	spin_lock_init(&bfqd->lock);
6486 
6487 	/*
6488 	 * The invocation of the next bfq_create_group_hierarchy
6489 	 * function is the head of a chain of function calls
6490 	 * (bfq_create_group_hierarchy->blkcg_activate_policy->
6491 	 * blk_mq_freeze_queue) that may lead to the invocation of the
6492 	 * has_work hook function. For this reason,
6493 	 * bfq_create_group_hierarchy is invoked only after all
6494 	 * scheduler data has been initialized, apart from the fields
6495 	 * that can be initialized only after invoking
6496 	 * bfq_create_group_hierarchy. This, in particular, enables
6497 	 * has_work to correctly return false. Of course, to avoid
6498 	 * other inconsistencies, the blk-mq stack must then refrain
6499 	 * from invoking further scheduler hooks before this init
6500 	 * function is finished.
6501 	 */
6502 	bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
6503 	if (!bfqd->root_group)
6504 		goto out_free;
6505 	bfq_init_root_group(bfqd->root_group, bfqd);
6506 	bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
6507 
6508 	wbt_disable_default(q);
6509 	return 0;
6510 
6511 out_free:
6512 	kfree(bfqd);
6513 	kobject_put(&eq->kobj);
6514 	return -ENOMEM;
6515 }
6516 
6517 static void bfq_slab_kill(void)
6518 {
6519 	kmem_cache_destroy(bfq_pool);
6520 }
6521 
6522 static int __init bfq_slab_setup(void)
6523 {
6524 	bfq_pool = KMEM_CACHE(bfq_queue, 0);
6525 	if (!bfq_pool)
6526 		return -ENOMEM;
6527 	return 0;
6528 }
6529 
6530 static ssize_t bfq_var_show(unsigned int var, char *page)
6531 {
6532 	return sprintf(page, "%u\n", var);
6533 }
6534 
6535 static int bfq_var_store(unsigned long *var, const char *page)
6536 {
6537 	unsigned long new_val;
6538 	int ret = kstrtoul(page, 10, &new_val);
6539 
6540 	if (ret)
6541 		return ret;
6542 	*var = new_val;
6543 	return 0;
6544 }
6545 
6546 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)				\
6547 static ssize_t __FUNC(struct elevator_queue *e, char *page)		\
6548 {									\
6549 	struct bfq_data *bfqd = e->elevator_data;			\
6550 	u64 __data = __VAR;						\
6551 	if (__CONV == 1)						\
6552 		__data = jiffies_to_msecs(__data);			\
6553 	else if (__CONV == 2)						\
6554 		__data = div_u64(__data, NSEC_PER_MSEC);		\
6555 	return bfq_var_show(__data, (page));				\
6556 }
6557 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
6558 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
6559 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
6560 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
6561 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
6562 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
6563 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
6564 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
6565 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
6566 #undef SHOW_FUNCTION
6567 
6568 #define USEC_SHOW_FUNCTION(__FUNC, __VAR)				\
6569 static ssize_t __FUNC(struct elevator_queue *e, char *page)		\
6570 {									\
6571 	struct bfq_data *bfqd = e->elevator_data;			\
6572 	u64 __data = __VAR;						\
6573 	__data = div_u64(__data, NSEC_PER_USEC);			\
6574 	return bfq_var_show(__data, (page));				\
6575 }
6576 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
6577 #undef USEC_SHOW_FUNCTION
6578 
6579 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)			\
6580 static ssize_t								\
6581 __FUNC(struct elevator_queue *e, const char *page, size_t count)	\
6582 {									\
6583 	struct bfq_data *bfqd = e->elevator_data;			\
6584 	unsigned long __data, __min = (MIN), __max = (MAX);		\
6585 	int ret;							\
6586 									\
6587 	ret = bfq_var_store(&__data, (page));				\
6588 	if (ret)							\
6589 		return ret;						\
6590 	if (__data < __min)						\
6591 		__data = __min;						\
6592 	else if (__data > __max)					\
6593 		__data = __max;						\
6594 	if (__CONV == 1)						\
6595 		*(__PTR) = msecs_to_jiffies(__data);			\
6596 	else if (__CONV == 2)						\
6597 		*(__PTR) = (u64)__data * NSEC_PER_MSEC;			\
6598 	else								\
6599 		*(__PTR) = __data;					\
6600 	return count;							\
6601 }
6602 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
6603 		INT_MAX, 2);
6604 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
6605 		INT_MAX, 2);
6606 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
6607 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
6608 		INT_MAX, 0);
6609 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
6610 #undef STORE_FUNCTION
6611 
6612 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX)			\
6613 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
6614 {									\
6615 	struct bfq_data *bfqd = e->elevator_data;			\
6616 	unsigned long __data, __min = (MIN), __max = (MAX);		\
6617 	int ret;							\
6618 									\
6619 	ret = bfq_var_store(&__data, (page));				\
6620 	if (ret)							\
6621 		return ret;						\
6622 	if (__data < __min)						\
6623 		__data = __min;						\
6624 	else if (__data > __max)					\
6625 		__data = __max;						\
6626 	*(__PTR) = (u64)__data * NSEC_PER_USEC;				\
6627 	return count;							\
6628 }
6629 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
6630 		    UINT_MAX);
6631 #undef USEC_STORE_FUNCTION
6632 
6633 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
6634 				    const char *page, size_t count)
6635 {
6636 	struct bfq_data *bfqd = e->elevator_data;
6637 	unsigned long __data;
6638 	int ret;
6639 
6640 	ret = bfq_var_store(&__data, (page));
6641 	if (ret)
6642 		return ret;
6643 
6644 	if (__data == 0)
6645 		bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
6646 	else {
6647 		if (__data > INT_MAX)
6648 			__data = INT_MAX;
6649 		bfqd->bfq_max_budget = __data;
6650 	}
6651 
6652 	bfqd->bfq_user_max_budget = __data;
6653 
6654 	return count;
6655 }
6656 
6657 /*
6658  * Leaving this name to preserve name compatibility with cfq
6659  * parameters, but this timeout is used for both sync and async.
6660  */
6661 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
6662 				      const char *page, size_t count)
6663 {
6664 	struct bfq_data *bfqd = e->elevator_data;
6665 	unsigned long __data;
6666 	int ret;
6667 
6668 	ret = bfq_var_store(&__data, (page));
6669 	if (ret)
6670 		return ret;
6671 
6672 	if (__data < 1)
6673 		__data = 1;
6674 	else if (__data > INT_MAX)
6675 		__data = INT_MAX;
6676 
6677 	bfqd->bfq_timeout = msecs_to_jiffies(__data);
6678 	if (bfqd->bfq_user_max_budget == 0)
6679 		bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
6680 
6681 	return count;
6682 }
6683 
6684 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
6685 				     const char *page, size_t count)
6686 {
6687 	struct bfq_data *bfqd = e->elevator_data;
6688 	unsigned long __data;
6689 	int ret;
6690 
6691 	ret = bfq_var_store(&__data, (page));
6692 	if (ret)
6693 		return ret;
6694 
6695 	if (__data > 1)
6696 		__data = 1;
6697 	if (!bfqd->strict_guarantees && __data == 1
6698 	    && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
6699 		bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
6700 
6701 	bfqd->strict_guarantees = __data;
6702 
6703 	return count;
6704 }
6705 
6706 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
6707 				     const char *page, size_t count)
6708 {
6709 	struct bfq_data *bfqd = e->elevator_data;
6710 	unsigned long __data;
6711 	int ret;
6712 
6713 	ret = bfq_var_store(&__data, (page));
6714 	if (ret)
6715 		return ret;
6716 
6717 	if (__data > 1)
6718 		__data = 1;
6719 	if (__data == 0 && bfqd->low_latency != 0)
6720 		bfq_end_wr(bfqd);
6721 	bfqd->low_latency = __data;
6722 
6723 	return count;
6724 }
6725 
6726 #define BFQ_ATTR(name) \
6727 	__ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
6728 
6729 static struct elv_fs_entry bfq_attrs[] = {
6730 	BFQ_ATTR(fifo_expire_sync),
6731 	BFQ_ATTR(fifo_expire_async),
6732 	BFQ_ATTR(back_seek_max),
6733 	BFQ_ATTR(back_seek_penalty),
6734 	BFQ_ATTR(slice_idle),
6735 	BFQ_ATTR(slice_idle_us),
6736 	BFQ_ATTR(max_budget),
6737 	BFQ_ATTR(timeout_sync),
6738 	BFQ_ATTR(strict_guarantees),
6739 	BFQ_ATTR(low_latency),
6740 	__ATTR_NULL
6741 };
6742 
6743 static struct elevator_type iosched_bfq_mq = {
6744 	.ops = {
6745 		.limit_depth		= bfq_limit_depth,
6746 		.prepare_request	= bfq_prepare_request,
6747 		.requeue_request        = bfq_finish_requeue_request,
6748 		.finish_request		= bfq_finish_requeue_request,
6749 		.exit_icq		= bfq_exit_icq,
6750 		.insert_requests	= bfq_insert_requests,
6751 		.dispatch_request	= bfq_dispatch_request,
6752 		.next_request		= elv_rb_latter_request,
6753 		.former_request		= elv_rb_former_request,
6754 		.allow_merge		= bfq_allow_bio_merge,
6755 		.bio_merge		= bfq_bio_merge,
6756 		.request_merge		= bfq_request_merge,
6757 		.requests_merged	= bfq_requests_merged,
6758 		.request_merged		= bfq_request_merged,
6759 		.has_work		= bfq_has_work,
6760 		.depth_updated		= bfq_depth_updated,
6761 		.init_hctx		= bfq_init_hctx,
6762 		.init_sched		= bfq_init_queue,
6763 		.exit_sched		= bfq_exit_queue,
6764 	},
6765 
6766 	.icq_size =		sizeof(struct bfq_io_cq),
6767 	.icq_align =		__alignof__(struct bfq_io_cq),
6768 	.elevator_attrs =	bfq_attrs,
6769 	.elevator_name =	"bfq",
6770 	.elevator_owner =	THIS_MODULE,
6771 };
6772 MODULE_ALIAS("bfq-iosched");
6773 
6774 static int __init bfq_init(void)
6775 {
6776 	int ret;
6777 
6778 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6779 	ret = blkcg_policy_register(&blkcg_policy_bfq);
6780 	if (ret)
6781 		return ret;
6782 #endif
6783 
6784 	ret = -ENOMEM;
6785 	if (bfq_slab_setup())
6786 		goto err_pol_unreg;
6787 
6788 	/*
6789 	 * Times to load large popular applications for the typical
6790 	 * systems installed on the reference devices (see the
6791 	 * comments before the definition of the next
6792 	 * array). Actually, we use slightly lower values, as the
6793 	 * estimated peak rate tends to be smaller than the actual
6794 	 * peak rate.  The reason for this last fact is that estimates
6795 	 * are computed over much shorter time intervals than the long
6796 	 * intervals typically used for benchmarking. Why? First, to
6797 	 * adapt more quickly to variations. Second, because an I/O
6798 	 * scheduler cannot rely on a peak-rate-evaluation workload to
6799 	 * be run for a long time.
6800 	 */
6801 	ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
6802 	ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
6803 
6804 	ret = elv_register(&iosched_bfq_mq);
6805 	if (ret)
6806 		goto slab_kill;
6807 
6808 	return 0;
6809 
6810 slab_kill:
6811 	bfq_slab_kill();
6812 err_pol_unreg:
6813 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6814 	blkcg_policy_unregister(&blkcg_policy_bfq);
6815 #endif
6816 	return ret;
6817 }
6818 
6819 static void __exit bfq_exit(void)
6820 {
6821 	elv_unregister(&iosched_bfq_mq);
6822 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6823 	blkcg_policy_unregister(&blkcg_policy_bfq);
6824 #endif
6825 	bfq_slab_kill();
6826 }
6827 
6828 module_init(bfq_init);
6829 module_exit(bfq_exit);
6830 
6831 MODULE_AUTHOR("Paolo Valente");
6832 MODULE_LICENSE("GPL");
6833 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");
6834