~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/block/bfq-iosched.c

Version: ~ [ linux-5.1-rc1 ] ~ [ linux-5.0.3 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.30 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.107 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.164 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.176 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.136 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.63 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.39.4 ] ~ [ linux-2.6.38.8 ] ~ [ linux-2.6.37.6 ] ~ [ linux-2.6.36.4 ] ~ [ linux-2.6.35.14 ] ~ [ linux-2.6.34.15 ] ~ [ linux-2.6.33.20 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * Budget Fair Queueing (BFQ) I/O scheduler.
  3  *
  4  * Based on ideas and code from CFQ:
  5  * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
  6  *
  7  * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
  8  *                    Paolo Valente <paolo.valente@unimore.it>
  9  *
 10  * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
 11  *                    Arianna Avanzini <avanzini@google.com>
 12  *
 13  * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
 14  *
 15  *  This program is free software; you can redistribute it and/or
 16  *  modify it under the terms of the GNU General Public License as
 17  *  published by the Free Software Foundation; either version 2 of the
 18  *  License, or (at your option) any later version.
 19  *
 20  *  This program is distributed in the hope that it will be useful,
 21  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
 22  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 23  *  General Public License for more details.
 24  *
 25  * BFQ is a proportional-share I/O scheduler, with some extra
 26  * low-latency capabilities. BFQ also supports full hierarchical
 27  * scheduling through cgroups. Next paragraphs provide an introduction
 28  * on BFQ inner workings. Details on BFQ benefits, usage and
 29  * limitations can be found in Documentation/block/bfq-iosched.txt.
 30  *
 31  * BFQ is a proportional-share storage-I/O scheduling algorithm based
 32  * on the slice-by-slice service scheme of CFQ. But BFQ assigns
 33  * budgets, measured in number of sectors, to processes instead of
 34  * time slices. The device is not granted to the in-service process
 35  * for a given time slice, but until it has exhausted its assigned
 36  * budget. This change from the time to the service domain enables BFQ
 37  * to distribute the device throughput among processes as desired,
 38  * without any distortion due to throughput fluctuations, or to device
 39  * internal queueing. BFQ uses an ad hoc internal scheduler, called
 40  * B-WF2Q+, to schedule processes according to their budgets. More
 41  * precisely, BFQ schedules queues associated with processes. Each
 42  * process/queue is assigned a user-configurable weight, and B-WF2Q+
 43  * guarantees that each queue receives a fraction of the throughput
 44  * proportional to its weight. Thanks to the accurate policy of
 45  * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
 46  * processes issuing sequential requests (to boost the throughput),
 47  * and yet guarantee a low latency to interactive and soft real-time
 48  * applications.
 49  *
 50  * In particular, to provide these low-latency guarantees, BFQ
 51  * explicitly privileges the I/O of two classes of time-sensitive
 52  * applications: interactive and soft real-time. In more detail, BFQ
 53  * behaves this way if the low_latency parameter is set (default
 54  * configuration). This feature enables BFQ to provide applications in
 55  * these classes with a very low latency.
 56  *
 57  * To implement this feature, BFQ constantly tries to detect whether
 58  * the I/O requests in a bfq_queue come from an interactive or a soft
 59  * real-time application. For brevity, in these cases, the queue is
 60  * said to be interactive or soft real-time. In both cases, BFQ
 61  * privileges the service of the queue, over that of non-interactive
 62  * and non-soft-real-time queues. This privileging is performed,
 63  * mainly, by raising the weight of the queue. So, for brevity, we
 64  * call just weight-raising periods the time periods during which a
 65  * queue is privileged, because deemed interactive or soft real-time.
 66  *
 67  * The detection of soft real-time queues/applications is described in
 68  * detail in the comments on the function
 69  * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
 70  * interactive queue works as follows: a queue is deemed interactive
 71  * if it is constantly non empty only for a limited time interval,
 72  * after which it does become empty. The queue may be deemed
 73  * interactive again (for a limited time), if it restarts being
 74  * constantly non empty, provided that this happens only after the
 75  * queue has remained empty for a given minimum idle time.
 76  *
 77  * By default, BFQ computes automatically the above maximum time
 78  * interval, i.e., the time interval after which a constantly
 79  * non-empty queue stops being deemed interactive. Since a queue is
 80  * weight-raised while it is deemed interactive, this maximum time
 81  * interval happens to coincide with the (maximum) duration of the
 82  * weight-raising for interactive queues.
 83  *
 84  * Finally, BFQ also features additional heuristics for
 85  * preserving both a low latency and a high throughput on NCQ-capable,
 86  * rotational or flash-based devices, and to get the job done quickly
 87  * for applications consisting in many I/O-bound processes.
 88  *
 89  * NOTE: if the main or only goal, with a given device, is to achieve
 90  * the maximum-possible throughput at all times, then do switch off
 91  * all low-latency heuristics for that device, by setting low_latency
 92  * to 0.
 93  *
 94  * BFQ is described in [1], where also a reference to the initial,
 95  * more theoretical paper on BFQ can be found. The interested reader
 96  * can find in the latter paper full details on the main algorithm, as
 97  * well as formulas of the guarantees and formal proofs of all the
 98  * properties.  With respect to the version of BFQ presented in these
 99  * papers, this implementation adds a few more heuristics, such as the
100  * ones that guarantee a low latency to interactive and soft real-time
101  * applications, and a hierarchical extension based on H-WF2Q+.
102  *
103  * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
104  * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
105  * with O(log N) complexity derives from the one introduced with EEVDF
106  * in [3].
107  *
108  * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
109  *     Scheduler", Proceedings of the First Workshop on Mobile System
110  *     Technologies (MST-2015), May 2015.
111  *     http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
112  *
113  * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
114  *     Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
115  *     Oct 1997.
116  *
117  * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
118  *
119  * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
120  *     First: A Flexible and Accurate Mechanism for Proportional Share
121  *     Resource Allocation", technical report.
122  *
123  * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
124  */
125 #include <linux/module.h>
126 #include <linux/slab.h>
127 #include <linux/blkdev.h>
128 #include <linux/cgroup.h>
129 #include <linux/elevator.h>
130 #include <linux/ktime.h>
131 #include <linux/rbtree.h>
132 #include <linux/ioprio.h>
133 #include <linux/sbitmap.h>
134 #include <linux/delay.h>
135 
136 #include "blk.h"
137 #include "blk-mq.h"
138 #include "blk-mq-tag.h"
139 #include "blk-mq-sched.h"
140 #include "bfq-iosched.h"
141 #include "blk-wbt.h"
142 
143 #define BFQ_BFQQ_FNS(name)                                              \
144 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq)                       \
145 {                                                                       \
146         __set_bit(BFQQF_##name, &(bfqq)->flags);                        \
147 }                                                                       \
148 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq)                      \
149 {                                                                       \
150         __clear_bit(BFQQF_##name, &(bfqq)->flags);              \
151 }                                                                       \
152 int bfq_bfqq_##name(const struct bfq_queue *bfqq)                       \
153 {                                                                       \
154         return test_bit(BFQQF_##name, &(bfqq)->flags);          \
155 }
156 
157 BFQ_BFQQ_FNS(just_created);
158 BFQ_BFQQ_FNS(busy);
159 BFQ_BFQQ_FNS(wait_request);
160 BFQ_BFQQ_FNS(non_blocking_wait_rq);
161 BFQ_BFQQ_FNS(fifo_expire);
162 BFQ_BFQQ_FNS(has_short_ttime);
163 BFQ_BFQQ_FNS(sync);
164 BFQ_BFQQ_FNS(IO_bound);
165 BFQ_BFQQ_FNS(in_large_burst);
166 BFQ_BFQQ_FNS(coop);
167 BFQ_BFQQ_FNS(split_coop);
168 BFQ_BFQQ_FNS(softrt_update);
169 #undef BFQ_BFQQ_FNS                                             \
170 
171 /* Expiration time of sync (0) and async (1) requests, in ns. */
172 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
173 
174 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
175 static const int bfq_back_max = 16 * 1024;
176 
177 /* Penalty of a backwards seek, in number of sectors. */
178 static const int bfq_back_penalty = 2;
179 
180 /* Idling period duration, in ns. */
181 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
182 
183 /* Minimum number of assigned budgets for which stats are safe to compute. */
184 static const int bfq_stats_min_budgets = 194;
185 
186 /* Default maximum budget values, in sectors and number of requests. */
187 static const int bfq_default_max_budget = 16 * 1024;
188 
189 /*
190  * When a sync request is dispatched, the queue that contains that
191  * request, and all the ancestor entities of that queue, are charged
192  * with the number of sectors of the request. In constrast, if the
193  * request is async, then the queue and its ancestor entities are
194  * charged with the number of sectors of the request, multiplied by
195  * the factor below. This throttles the bandwidth for async I/O,
196  * w.r.t. to sync I/O, and it is done to counter the tendency of async
197  * writes to steal I/O throughput to reads.
198  *
199  * The current value of this parameter is the result of a tuning with
200  * several hardware and software configurations. We tried to find the
201  * lowest value for which writes do not cause noticeable problems to
202  * reads. In fact, the lower this parameter, the stabler I/O control,
203  * in the following respect.  The lower this parameter is, the less
204  * the bandwidth enjoyed by a group decreases
205  * - when the group does writes, w.r.t. to when it does reads;
206  * - when other groups do reads, w.r.t. to when they do writes.
207  */
208 static const int bfq_async_charge_factor = 3;
209 
210 /* Default timeout values, in jiffies, approximating CFQ defaults. */
211 const int bfq_timeout = HZ / 8;
212 
213 /*
214  * Time limit for merging (see comments in bfq_setup_cooperator). Set
215  * to the slowest value that, in our tests, proved to be effective in
216  * removing false positives, while not causing true positives to miss
217  * queue merging.
218  *
219  * As can be deduced from the low time limit below, queue merging, if
220  * successful, happens at the very beggining of the I/O of the involved
221  * cooperating processes, as a consequence of the arrival of the very
222  * first requests from each cooperator.  After that, there is very
223  * little chance to find cooperators.
224  */
225 static const unsigned long bfq_merge_time_limit = HZ/10;
226 
227 static struct kmem_cache *bfq_pool;
228 
229 /* Below this threshold (in ns), we consider thinktime immediate. */
230 #define BFQ_MIN_TT              (2 * NSEC_PER_MSEC)
231 
232 /* hw_tag detection: parallel requests threshold and min samples needed. */
233 #define BFQ_HW_QUEUE_THRESHOLD  3
234 #define BFQ_HW_QUEUE_SAMPLES    32
235 
236 #define BFQQ_SEEK_THR           (sector_t)(8 * 100)
237 #define BFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
238 #define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \
239         (get_sdist(last_pos, rq) >                      \
240          BFQQ_SEEK_THR &&                               \
241          (!blk_queue_nonrot(bfqd->queue) ||             \
242           blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT))
243 #define BFQQ_CLOSE_THR          (sector_t)(8 * 1024)
244 #define BFQQ_SEEKY(bfqq)        (hweight32(bfqq->seek_history) > 19)
245 
246 /* Min number of samples required to perform peak-rate update */
247 #define BFQ_RATE_MIN_SAMPLES    32
248 /* Min observation time interval required to perform a peak-rate update (ns) */
249 #define BFQ_RATE_MIN_INTERVAL   (300*NSEC_PER_MSEC)
250 /* Target observation time interval for a peak-rate update (ns) */
251 #define BFQ_RATE_REF_INTERVAL   NSEC_PER_SEC
252 
253 /*
254  * Shift used for peak-rate fixed precision calculations.
255  * With
256  * - the current shift: 16 positions
257  * - the current type used to store rate: u32
258  * - the current unit of measure for rate: [sectors/usec], or, more precisely,
259  *   [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
260  * the range of rates that can be stored is
261  * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
262  * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
263  * [15, 65G] sectors/sec
264  * Which, assuming a sector size of 512B, corresponds to a range of
265  * [7.5K, 33T] B/sec
266  */
267 #define BFQ_RATE_SHIFT          16
268 
269 /*
270  * When configured for computing the duration of the weight-raising
271  * for interactive queues automatically (see the comments at the
272  * beginning of this file), BFQ does it using the following formula:
273  * duration = (ref_rate / r) * ref_wr_duration,
274  * where r is the peak rate of the device, and ref_rate and
275  * ref_wr_duration are two reference parameters.  In particular,
276  * ref_rate is the peak rate of the reference storage device (see
277  * below), and ref_wr_duration is about the maximum time needed, with
278  * BFQ and while reading two files in parallel, to load typical large
279  * applications on the reference device (see the comments on
280  * max_service_from_wr below, for more details on how ref_wr_duration
281  * is obtained).  In practice, the slower/faster the device at hand
282  * is, the more/less it takes to load applications with respect to the
283  * reference device.  Accordingly, the longer/shorter BFQ grants
284  * weight raising to interactive applications.
285  *
286  * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
287  * depending on whether the device is rotational or non-rotational.
288  *
289  * In the following definitions, ref_rate[0] and ref_wr_duration[0]
290  * are the reference values for a rotational device, whereas
291  * ref_rate[1] and ref_wr_duration[1] are the reference values for a
292  * non-rotational device. The reference rates are not the actual peak
293  * rates of the devices used as a reference, but slightly lower
294  * values. The reason for using slightly lower values is that the
295  * peak-rate estimator tends to yield slightly lower values than the
296  * actual peak rate (it can yield the actual peak rate only if there
297  * is only one process doing I/O, and the process does sequential
298  * I/O).
299  *
300  * The reference peak rates are measured in sectors/usec, left-shifted
301  * by BFQ_RATE_SHIFT.
302  */
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 closesr 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 void bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
599 {
600         struct rb_node **p, *parent;
601         struct bfq_queue *__bfqq;
602 
603         if (bfqq->pos_root) {
604                 rb_erase(&bfqq->pos_node, bfqq->pos_root);
605                 bfqq->pos_root = NULL;
606         }
607 
608         /*
609          * bfqq cannot be merged any longer (see comments in
610          * bfq_setup_cooperator): no point in adding bfqq into the
611          * position tree.
612          */
613         if (bfq_too_late_for_merging(bfqq))
614                 return;
615 
616         if (bfq_class_idle(bfqq))
617                 return;
618         if (!bfqq->next_rq)
619                 return;
620 
621         bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
622         __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
623                         blk_rq_pos(bfqq->next_rq), &parent, &p);
624         if (!__bfqq) {
625                 rb_link_node(&bfqq->pos_node, parent, p);
626                 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
627         } else
628                 bfqq->pos_root = NULL;
629 }
630 
631 /*
632  * The following function returns true if every queue must receive the
633  * same share of the throughput (this condition is used when deciding
634  * whether idling may be disabled, see the comments in the function
635  * bfq_better_to_idle()).
636  *
637  * Such a scenario occurs when:
638  * 1) all active queues have the same weight,
639  * 2) all active queues belong to the same I/O-priority class,
640  * 3) all active groups at the same level in the groups tree have the same
641  *    weight,
642  * 4) all active groups at the same level in the groups tree have the same
643  *    number of children.
644  *
645  * Unfortunately, keeping the necessary state for evaluating exactly
646  * the last two symmetry sub-conditions above would be quite complex
647  * and time consuming. Therefore this function evaluates, instead,
648  * only the following stronger three sub-conditions, for which it is
649  * much easier to maintain the needed state:
650  * 1) all active queues have the same weight,
651  * 2) all active queues belong to the same I/O-priority class,
652  * 3) there are no active groups.
653  * In particular, the last condition is always true if hierarchical
654  * support or the cgroups interface are not enabled, thus no state
655  * needs to be maintained in this case.
656  */
657 static bool bfq_symmetric_scenario(struct bfq_data *bfqd)
658 {
659         /*
660          * For queue weights to differ, queue_weights_tree must contain
661          * at least two nodes.
662          */
663         bool varied_queue_weights = !RB_EMPTY_ROOT(&bfqd->queue_weights_tree) &&
664                 (bfqd->queue_weights_tree.rb_node->rb_left ||
665                  bfqd->queue_weights_tree.rb_node->rb_right);
666 
667         bool multiple_classes_busy =
668                 (bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
669                 (bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
670                 (bfqd->busy_queues[1] && bfqd->busy_queues[2]);
671 
672         /*
673          * For queue weights to differ, queue_weights_tree must contain
674          * at least two nodes.
675          */
676         return !(varied_queue_weights || multiple_classes_busy
677 #ifdef BFQ_GROUP_IOSCHED_ENABLED
678                || bfqd->num_groups_with_pending_reqs > 0
679 #endif
680                 );
681 }
682 
683 /*
684  * If the weight-counter tree passed as input contains no counter for
685  * the weight of the input queue, then add that counter; otherwise just
686  * increment the existing counter.
687  *
688  * Note that weight-counter trees contain few nodes in mostly symmetric
689  * scenarios. For example, if all queues have the same weight, then the
690  * weight-counter tree for the queues may contain at most one node.
691  * This holds even if low_latency is on, because weight-raised queues
692  * are not inserted in the tree.
693  * In most scenarios, the rate at which nodes are created/destroyed
694  * should be low too.
695  */
696 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq,
697                           struct rb_root *root)
698 {
699         struct bfq_entity *entity = &bfqq->entity;
700         struct rb_node **new = &(root->rb_node), *parent = NULL;
701 
702         /*
703          * Do not insert if the queue is already associated with a
704          * counter, which happens if:
705          *   1) a request arrival has caused the queue to become both
706          *      non-weight-raised, and hence change its weight, and
707          *      backlogged; in this respect, each of the two events
708          *      causes an invocation of this function,
709          *   2) this is the invocation of this function caused by the
710          *      second event. This second invocation is actually useless,
711          *      and we handle this fact by exiting immediately. More
712          *      efficient or clearer solutions might possibly be adopted.
713          */
714         if (bfqq->weight_counter)
715                 return;
716 
717         while (*new) {
718                 struct bfq_weight_counter *__counter = container_of(*new,
719                                                 struct bfq_weight_counter,
720                                                 weights_node);
721                 parent = *new;
722 
723                 if (entity->weight == __counter->weight) {
724                         bfqq->weight_counter = __counter;
725                         goto inc_counter;
726                 }
727                 if (entity->weight < __counter->weight)
728                         new = &((*new)->rb_left);
729                 else
730                         new = &((*new)->rb_right);
731         }
732 
733         bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
734                                        GFP_ATOMIC);
735 
736         /*
737          * In the unlucky event of an allocation failure, we just
738          * exit. This will cause the weight of queue to not be
739          * considered in bfq_symmetric_scenario, which, in its turn,
740          * causes the scenario to be deemed wrongly symmetric in case
741          * bfqq's weight would have been the only weight making the
742          * scenario asymmetric.  On the bright side, no unbalance will
743          * however occur when bfqq becomes inactive again (the
744          * invocation of this function is triggered by an activation
745          * of queue).  In fact, bfq_weights_tree_remove does nothing
746          * if !bfqq->weight_counter.
747          */
748         if (unlikely(!bfqq->weight_counter))
749                 return;
750 
751         bfqq->weight_counter->weight = entity->weight;
752         rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
753         rb_insert_color(&bfqq->weight_counter->weights_node, root);
754 
755 inc_counter:
756         bfqq->weight_counter->num_active++;
757         bfqq->ref++;
758 }
759 
760 /*
761  * Decrement the weight counter associated with the queue, and, if the
762  * counter reaches 0, remove the counter from the tree.
763  * See the comments to the function bfq_weights_tree_add() for considerations
764  * about overhead.
765  */
766 void __bfq_weights_tree_remove(struct bfq_data *bfqd,
767                                struct bfq_queue *bfqq,
768                                struct rb_root *root)
769 {
770         if (!bfqq->weight_counter)
771                 return;
772 
773         bfqq->weight_counter->num_active--;
774         if (bfqq->weight_counter->num_active > 0)
775                 goto reset_entity_pointer;
776 
777         rb_erase(&bfqq->weight_counter->weights_node, root);
778         kfree(bfqq->weight_counter);
779 
780 reset_entity_pointer:
781         bfqq->weight_counter = NULL;
782         bfq_put_queue(bfqq);
783 }
784 
785 /*
786  * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
787  * of active groups for each queue's inactive parent entity.
788  */
789 void bfq_weights_tree_remove(struct bfq_data *bfqd,
790                              struct bfq_queue *bfqq)
791 {
792         struct bfq_entity *entity = bfqq->entity.parent;
793 
794         for_each_entity(entity) {
795                 struct bfq_sched_data *sd = entity->my_sched_data;
796 
797                 if (sd->next_in_service || sd->in_service_entity) {
798                         /*
799                          * entity is still active, because either
800                          * next_in_service or in_service_entity is not
801                          * NULL (see the comments on the definition of
802                          * next_in_service for details on why
803                          * in_service_entity must be checked too).
804                          *
805                          * As a consequence, its parent entities are
806                          * active as well, and thus this loop must
807                          * stop here.
808                          */
809                         break;
810                 }
811 
812                 /*
813                  * The decrement of num_groups_with_pending_reqs is
814                  * not performed immediately upon the deactivation of
815                  * entity, but it is delayed to when it also happens
816                  * that the first leaf descendant bfqq of entity gets
817                  * all its pending requests completed. The following
818                  * instructions perform this delayed decrement, if
819                  * needed. See the comments on
820                  * num_groups_with_pending_reqs for details.
821                  */
822                 if (entity->in_groups_with_pending_reqs) {
823                         entity->in_groups_with_pending_reqs = false;
824                         bfqd->num_groups_with_pending_reqs--;
825                 }
826         }
827 
828         /*
829          * Next function is invoked last, because it causes bfqq to be
830          * freed if the following holds: bfqq is not in service and
831          * has no dispatched request. DO NOT use bfqq after the next
832          * function invocation.
833          */
834         __bfq_weights_tree_remove(bfqd, bfqq,
835                                   &bfqd->queue_weights_tree);
836 }
837 
838 /*
839  * Return expired entry, or NULL to just start from scratch in rbtree.
840  */
841 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
842                                       struct request *last)
843 {
844         struct request *rq;
845 
846         if (bfq_bfqq_fifo_expire(bfqq))
847                 return NULL;
848 
849         bfq_mark_bfqq_fifo_expire(bfqq);
850 
851         rq = rq_entry_fifo(bfqq->fifo.next);
852 
853         if (rq == last || ktime_get_ns() < rq->fifo_time)
854                 return NULL;
855 
856         bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
857         return rq;
858 }
859 
860 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
861                                         struct bfq_queue *bfqq,
862                                         struct request *last)
863 {
864         struct rb_node *rbnext = rb_next(&last->rb_node);
865         struct rb_node *rbprev = rb_prev(&last->rb_node);
866         struct request *next, *prev = NULL;
867 
868         /* Follow expired path, else get first next available. */
869         next = bfq_check_fifo(bfqq, last);
870         if (next)
871                 return next;
872 
873         if (rbprev)
874                 prev = rb_entry_rq(rbprev);
875 
876         if (rbnext)
877                 next = rb_entry_rq(rbnext);
878         else {
879                 rbnext = rb_first(&bfqq->sort_list);
880                 if (rbnext && rbnext != &last->rb_node)
881                         next = rb_entry_rq(rbnext);
882         }
883 
884         return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
885 }
886 
887 /* see the definition of bfq_async_charge_factor for details */
888 static unsigned long bfq_serv_to_charge(struct request *rq,
889                                         struct bfq_queue *bfqq)
890 {
891         if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
892             !bfq_symmetric_scenario(bfqq->bfqd))
893                 return blk_rq_sectors(rq);
894 
895         return blk_rq_sectors(rq) * bfq_async_charge_factor;
896 }
897 
898 /**
899  * bfq_updated_next_req - update the queue after a new next_rq selection.
900  * @bfqd: the device data the queue belongs to.
901  * @bfqq: the queue to update.
902  *
903  * If the first request of a queue changes we make sure that the queue
904  * has enough budget to serve at least its first request (if the
905  * request has grown).  We do this because if the queue has not enough
906  * budget for its first request, it has to go through two dispatch
907  * rounds to actually get it dispatched.
908  */
909 static void bfq_updated_next_req(struct bfq_data *bfqd,
910                                  struct bfq_queue *bfqq)
911 {
912         struct bfq_entity *entity = &bfqq->entity;
913         struct request *next_rq = bfqq->next_rq;
914         unsigned long new_budget;
915 
916         if (!next_rq)
917                 return;
918 
919         if (bfqq == bfqd->in_service_queue)
920                 /*
921                  * In order not to break guarantees, budgets cannot be
922                  * changed after an entity has been selected.
923                  */
924                 return;
925 
926         new_budget = max_t(unsigned long,
927                            max_t(unsigned long, bfqq->max_budget,
928                                  bfq_serv_to_charge(next_rq, bfqq)),
929                            entity->service);
930         if (entity->budget != new_budget) {
931                 entity->budget = new_budget;
932                 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
933                                          new_budget);
934                 bfq_requeue_bfqq(bfqd, bfqq, false);
935         }
936 }
937 
938 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
939 {
940         u64 dur;
941 
942         if (bfqd->bfq_wr_max_time > 0)
943                 return bfqd->bfq_wr_max_time;
944 
945         dur = bfqd->rate_dur_prod;
946         do_div(dur, bfqd->peak_rate);
947 
948         /*
949          * Limit duration between 3 and 25 seconds. The upper limit
950          * has been conservatively set after the following worst case:
951          * on a QEMU/KVM virtual machine
952          * - running in a slow PC
953          * - with a virtual disk stacked on a slow low-end 5400rpm HDD
954          * - serving a heavy I/O workload, such as the sequential reading
955          *   of several files
956          * mplayer took 23 seconds to start, if constantly weight-raised.
957          *
958          * As for higher values than that accomodating the above bad
959          * scenario, tests show that higher values would often yield
960          * the opposite of the desired result, i.e., would worsen
961          * responsiveness by allowing non-interactive applications to
962          * preserve weight raising for too long.
963          *
964          * On the other end, lower values than 3 seconds make it
965          * difficult for most interactive tasks to complete their jobs
966          * before weight-raising finishes.
967          */
968         return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
969 }
970 
971 /* switch back from soft real-time to interactive weight raising */
972 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
973                                           struct bfq_data *bfqd)
974 {
975         bfqq->wr_coeff = bfqd->bfq_wr_coeff;
976         bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
977         bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
978 }
979 
980 static void
981 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
982                       struct bfq_io_cq *bic, bool bfq_already_existing)
983 {
984         unsigned int old_wr_coeff = bfqq->wr_coeff;
985         bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
986 
987         if (bic->saved_has_short_ttime)
988                 bfq_mark_bfqq_has_short_ttime(bfqq);
989         else
990                 bfq_clear_bfqq_has_short_ttime(bfqq);
991 
992         if (bic->saved_IO_bound)
993                 bfq_mark_bfqq_IO_bound(bfqq);
994         else
995                 bfq_clear_bfqq_IO_bound(bfqq);
996 
997         bfqq->ttime = bic->saved_ttime;
998         bfqq->wr_coeff = bic->saved_wr_coeff;
999         bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
1000         bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
1001         bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
1002 
1003         if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1004             time_is_before_jiffies(bfqq->last_wr_start_finish +
1005                                    bfqq->wr_cur_max_time))) {
1006                 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1007                     !bfq_bfqq_in_large_burst(bfqq) &&
1008                     time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1009                                              bfq_wr_duration(bfqd))) {
1010                         switch_back_to_interactive_wr(bfqq, bfqd);
1011                 } else {
1012                         bfqq->wr_coeff = 1;
1013                         bfq_log_bfqq(bfqq->bfqd, bfqq,
1014                                      "resume state: switching off wr");
1015                 }
1016         }
1017 
1018         /* make sure weight will be updated, however we got here */
1019         bfqq->entity.prio_changed = 1;
1020 
1021         if (likely(!busy))
1022                 return;
1023 
1024         if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1025                 bfqd->wr_busy_queues++;
1026         else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1027                 bfqd->wr_busy_queues--;
1028 }
1029 
1030 static int bfqq_process_refs(struct bfq_queue *bfqq)
1031 {
1032         return bfqq->ref - bfqq->allocated - bfqq->entity.on_st -
1033                 (bfqq->weight_counter != NULL);
1034 }
1035 
1036 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1037 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1038 {
1039         struct bfq_queue *item;
1040         struct hlist_node *n;
1041 
1042         hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1043                 hlist_del_init(&item->burst_list_node);
1044         hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1045         bfqd->burst_size = 1;
1046         bfqd->burst_parent_entity = bfqq->entity.parent;
1047 }
1048 
1049 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1050 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1051 {
1052         /* Increment burst size to take into account also bfqq */
1053         bfqd->burst_size++;
1054 
1055         if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1056                 struct bfq_queue *pos, *bfqq_item;
1057                 struct hlist_node *n;
1058 
1059                 /*
1060                  * Enough queues have been activated shortly after each
1061                  * other to consider this burst as large.
1062                  */
1063                 bfqd->large_burst = true;
1064 
1065                 /*
1066                  * We can now mark all queues in the burst list as
1067                  * belonging to a large burst.
1068                  */
1069                 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1070                                      burst_list_node)
1071                         bfq_mark_bfqq_in_large_burst(bfqq_item);
1072                 bfq_mark_bfqq_in_large_burst(bfqq);
1073 
1074                 /*
1075                  * From now on, and until the current burst finishes, any
1076                  * new queue being activated shortly after the last queue
1077                  * was inserted in the burst can be immediately marked as
1078                  * belonging to a large burst. So the burst list is not
1079                  * needed any more. Remove it.
1080                  */
1081                 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1082                                           burst_list_node)
1083                         hlist_del_init(&pos->burst_list_node);
1084         } else /*
1085                 * Burst not yet large: add bfqq to the burst list. Do
1086                 * not increment the ref counter for bfqq, because bfqq
1087                 * is removed from the burst list before freeing bfqq
1088                 * in put_queue.
1089                 */
1090                 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1091 }
1092 
1093 /*
1094  * If many queues belonging to the same group happen to be created
1095  * shortly after each other, then the processes associated with these
1096  * queues have typically a common goal. In particular, bursts of queue
1097  * creations are usually caused by services or applications that spawn
1098  * many parallel threads/processes. Examples are systemd during boot,
1099  * or git grep. To help these processes get their job done as soon as
1100  * possible, it is usually better to not grant either weight-raising
1101  * or device idling to their queues.
1102  *
1103  * In this comment we describe, firstly, the reasons why this fact
1104  * holds, and, secondly, the next function, which implements the main
1105  * steps needed to properly mark these queues so that they can then be
1106  * treated in a different way.
1107  *
1108  * The above services or applications benefit mostly from a high
1109  * throughput: the quicker the requests of the activated queues are
1110  * cumulatively served, the sooner the target job of these queues gets
1111  * completed. As a consequence, weight-raising any of these queues,
1112  * which also implies idling the device for it, is almost always
1113  * counterproductive. In most cases it just lowers throughput.
1114  *
1115  * On the other hand, a burst of queue creations may be caused also by
1116  * the start of an application that does not consist of a lot of
1117  * parallel I/O-bound threads. In fact, with a complex application,
1118  * several short processes may need to be executed to start-up the
1119  * application. In this respect, to start an application as quickly as
1120  * possible, the best thing to do is in any case to privilege the I/O
1121  * related to the application with respect to all other
1122  * I/O. Therefore, the best strategy to start as quickly as possible
1123  * an application that causes a burst of queue creations is to
1124  * weight-raise all the queues created during the burst. This is the
1125  * exact opposite of the best strategy for the other type of bursts.
1126  *
1127  * In the end, to take the best action for each of the two cases, the
1128  * two types of bursts need to be distinguished. Fortunately, this
1129  * seems relatively easy, by looking at the sizes of the bursts. In
1130  * particular, we found a threshold such that only bursts with a
1131  * larger size than that threshold are apparently caused by
1132  * services or commands such as systemd or git grep. For brevity,
1133  * hereafter we call just 'large' these bursts. BFQ *does not*
1134  * weight-raise queues whose creation occurs in a large burst. In
1135  * addition, for each of these queues BFQ performs or does not perform
1136  * idling depending on which choice boosts the throughput more. The
1137  * exact choice depends on the device and request pattern at
1138  * hand.
1139  *
1140  * Unfortunately, false positives may occur while an interactive task
1141  * is starting (e.g., an application is being started). The
1142  * consequence is that the queues associated with the task do not
1143  * enjoy weight raising as expected. Fortunately these false positives
1144  * are very rare. They typically occur if some service happens to
1145  * start doing I/O exactly when the interactive task starts.
1146  *
1147  * Turning back to the next function, it implements all the steps
1148  * needed to detect the occurrence of a large burst and to properly
1149  * mark all the queues belonging to it (so that they can then be
1150  * treated in a different way). This goal is achieved by maintaining a
1151  * "burst list" that holds, temporarily, the queues that belong to the
1152  * burst in progress. The list is then used to mark these queues as
1153  * belonging to a large burst if the burst does become large. The main
1154  * steps are the following.
1155  *
1156  * . when the very first queue is created, the queue is inserted into the
1157  *   list (as it could be the first queue in a possible burst)
1158  *
1159  * . if the current burst has not yet become large, and a queue Q that does
1160  *   not yet belong to the burst is activated shortly after the last time
1161  *   at which a new queue entered the burst list, then the function appends
1162  *   Q to the burst list
1163  *
1164  * . if, as a consequence of the previous step, the burst size reaches
1165  *   the large-burst threshold, then
1166  *
1167  *     . all the queues in the burst list are marked as belonging to a
1168  *       large burst
1169  *
1170  *     . the burst list is deleted; in fact, the burst list already served
1171  *       its purpose (keeping temporarily track of the queues in a burst,
1172  *       so as to be able to mark them as belonging to a large burst in the
1173  *       previous sub-step), and now is not needed any more
1174  *
1175  *     . the device enters a large-burst mode
1176  *
1177  * . if a queue Q that does not belong to the burst is created while
1178  *   the device is in large-burst mode and shortly after the last time
1179  *   at which a queue either entered the burst list or was marked as
1180  *   belonging to the current large burst, then Q is immediately marked
1181  *   as belonging to a large burst.
1182  *
1183  * . if a queue Q that does not belong to the burst is created a while
1184  *   later, i.e., not shortly after, than the last time at which a queue
1185  *   either entered the burst list or was marked as belonging to the
1186  *   current large burst, then the current burst is deemed as finished and:
1187  *
1188  *        . the large-burst mode is reset if set
1189  *
1190  *        . the burst list is emptied
1191  *
1192  *        . Q is inserted in the burst list, as Q may be the first queue
1193  *          in a possible new burst (then the burst list contains just Q
1194  *          after this step).
1195  */
1196 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1197 {
1198         /*
1199          * If bfqq is already in the burst list or is part of a large
1200          * burst, or finally has just been split, then there is
1201          * nothing else to do.
1202          */
1203         if (!hlist_unhashed(&bfqq->burst_list_node) ||
1204             bfq_bfqq_in_large_burst(bfqq) ||
1205             time_is_after_eq_jiffies(bfqq->split_time +
1206                                      msecs_to_jiffies(10)))
1207                 return;
1208 
1209         /*
1210          * If bfqq's creation happens late enough, or bfqq belongs to
1211          * a different group than the burst group, then the current
1212          * burst is finished, and related data structures must be
1213          * reset.
1214          *
1215          * In this respect, consider the special case where bfqq is
1216          * the very first queue created after BFQ is selected for this
1217          * device. In this case, last_ins_in_burst and
1218          * burst_parent_entity are not yet significant when we get
1219          * here. But it is easy to verify that, whether or not the
1220          * following condition is true, bfqq will end up being
1221          * inserted into the burst list. In particular the list will
1222          * happen to contain only bfqq. And this is exactly what has
1223          * to happen, as bfqq may be the first queue of the first
1224          * burst.
1225          */
1226         if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1227             bfqd->bfq_burst_interval) ||
1228             bfqq->entity.parent != bfqd->burst_parent_entity) {
1229                 bfqd->large_burst = false;
1230                 bfq_reset_burst_list(bfqd, bfqq);
1231                 goto end;
1232         }
1233 
1234         /*
1235          * If we get here, then bfqq is being activated shortly after the
1236          * last queue. So, if the current burst is also large, we can mark
1237          * bfqq as belonging to this large burst immediately.
1238          */
1239         if (bfqd->large_burst) {
1240                 bfq_mark_bfqq_in_large_burst(bfqq);
1241                 goto end;
1242         }
1243 
1244         /*
1245          * If we get here, then a large-burst state has not yet been
1246          * reached, but bfqq is being activated shortly after the last
1247          * queue. Then we add bfqq to the burst.
1248          */
1249         bfq_add_to_burst(bfqd, bfqq);
1250 end:
1251         /*
1252          * At this point, bfqq either has been added to the current
1253          * burst or has caused the current burst to terminate and a
1254          * possible new burst to start. In particular, in the second
1255          * case, bfqq has become the first queue in the possible new
1256          * burst.  In both cases last_ins_in_burst needs to be moved
1257          * forward.
1258          */
1259         bfqd->last_ins_in_burst = jiffies;
1260 }
1261 
1262 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1263 {
1264         struct bfq_entity *entity = &bfqq->entity;
1265 
1266         return entity->budget - entity->service;
1267 }
1268 
1269 /*
1270  * If enough samples have been computed, return the current max budget
1271  * stored in bfqd, which is dynamically updated according to the
1272  * estimated disk peak rate; otherwise return the default max budget
1273  */
1274 static int bfq_max_budget(struct bfq_data *bfqd)
1275 {
1276         if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1277                 return bfq_default_max_budget;
1278         else
1279                 return bfqd->bfq_max_budget;
1280 }
1281 
1282 /*
1283  * Return min budget, which is a fraction of the current or default
1284  * max budget (trying with 1/32)
1285  */
1286 static int bfq_min_budget(struct bfq_data *bfqd)
1287 {
1288         if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1289                 return bfq_default_max_budget / 32;
1290         else
1291                 return bfqd->bfq_max_budget / 32;
1292 }
1293 
1294 /*
1295  * The next function, invoked after the input queue bfqq switches from
1296  * idle to busy, updates the budget of bfqq. The function also tells
1297  * whether the in-service queue should be expired, by returning
1298  * true. The purpose of expiring the in-service queue is to give bfqq
1299  * the chance to possibly preempt the in-service queue, and the reason
1300  * for preempting the in-service queue is to achieve one of the two
1301  * goals below.
1302  *
1303  * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1304  * expired because it has remained idle. In particular, bfqq may have
1305  * expired for one of the following two reasons:
1306  *
1307  * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1308  *   and did not make it to issue a new request before its last
1309  *   request was served;
1310  *
1311  * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1312  *   a new request before the expiration of the idling-time.
1313  *
1314  * Even if bfqq has expired for one of the above reasons, the process
1315  * associated with the queue may be however issuing requests greedily,
1316  * and thus be sensitive to the bandwidth it receives (bfqq may have
1317  * remained idle for other reasons: CPU high load, bfqq not enjoying
1318  * idling, I/O throttling somewhere in the path from the process to
1319  * the I/O scheduler, ...). But if, after every expiration for one of
1320  * the above two reasons, bfqq has to wait for the service of at least
1321  * one full budget of another queue before being served again, then
1322  * bfqq is likely to get a much lower bandwidth or resource time than
1323  * its reserved ones. To address this issue, two countermeasures need
1324  * to be taken.
1325  *
1326  * First, the budget and the timestamps of bfqq need to be updated in
1327  * a special way on bfqq reactivation: they need to be updated as if
1328  * bfqq did not remain idle and did not expire. In fact, if they are
1329  * computed as if bfqq expired and remained idle until reactivation,
1330  * then the process associated with bfqq is treated as if, instead of
1331  * being greedy, it stopped issuing requests when bfqq remained idle,
1332  * and restarts issuing requests only on this reactivation. In other
1333  * words, the scheduler does not help the process recover the "service
1334  * hole" between bfqq expiration and reactivation. As a consequence,
1335  * the process receives a lower bandwidth than its reserved one. In
1336  * contrast, to recover this hole, the budget must be updated as if
1337  * bfqq was not expired at all before this reactivation, i.e., it must
1338  * be set to the value of the remaining budget when bfqq was
1339  * expired. Along the same line, timestamps need to be assigned the
1340  * value they had the last time bfqq was selected for service, i.e.,
1341  * before last expiration. Thus timestamps need to be back-shifted
1342  * with respect to their normal computation (see [1] for more details
1343  * on this tricky aspect).
1344  *
1345  * Secondly, to allow the process to recover the hole, the in-service
1346  * queue must be expired too, to give bfqq the chance to preempt it
1347  * immediately. In fact, if bfqq has to wait for a full budget of the
1348  * in-service queue to be completed, then it may become impossible to
1349  * let the process recover the hole, even if the back-shifted
1350  * timestamps of bfqq are lower than those of the in-service queue. If
1351  * this happens for most or all of the holes, then the process may not
1352  * receive its reserved bandwidth. In this respect, it is worth noting
1353  * that, being the service of outstanding requests unpreemptible, a
1354  * little fraction of the holes may however be unrecoverable, thereby
1355  * causing a little loss of bandwidth.
1356  *
1357  * The last important point is detecting whether bfqq does need this
1358  * bandwidth recovery. In this respect, the next function deems the
1359  * process associated with bfqq greedy, and thus allows it to recover
1360  * the hole, if: 1) the process is waiting for the arrival of a new
1361  * request (which implies that bfqq expired for one of the above two
1362  * reasons), and 2) such a request has arrived soon. The first
1363  * condition is controlled through the flag non_blocking_wait_rq,
1364  * while the second through the flag arrived_in_time. If both
1365  * conditions hold, then the function computes the budget in the
1366  * above-described special way, and signals that the in-service queue
1367  * should be expired. Timestamp back-shifting is done later in
1368  * __bfq_activate_entity.
1369  *
1370  * 2. Reduce latency. Even if timestamps are not backshifted to let
1371  * the process associated with bfqq recover a service hole, bfqq may
1372  * however happen to have, after being (re)activated, a lower finish
1373  * timestamp than the in-service queue.  That is, the next budget of
1374  * bfqq may have to be completed before the one of the in-service
1375  * queue. If this is the case, then preempting the in-service queue
1376  * allows this goal to be achieved, apart from the unpreemptible,
1377  * outstanding requests mentioned above.
1378  *
1379  * Unfortunately, regardless of which of the above two goals one wants
1380  * to achieve, service trees need first to be updated to know whether
1381  * the in-service queue must be preempted. To have service trees
1382  * correctly updated, the in-service queue must be expired and
1383  * rescheduled, and bfqq must be scheduled too. This is one of the
1384  * most costly operations (in future versions, the scheduling
1385  * mechanism may be re-designed in such a way to make it possible to
1386  * know whether preemption is needed without needing to update service
1387  * trees). In addition, queue preemptions almost always cause random
1388  * I/O, and thus loss of throughput. Because of these facts, the next
1389  * function adopts the following simple scheme to avoid both costly
1390  * operations and too frequent preemptions: it requests the expiration
1391  * of the in-service queue (unconditionally) only for queues that need
1392  * to recover a hole, or that either are weight-raised or deserve to
1393  * be weight-raised.
1394  */
1395 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1396                                                 struct bfq_queue *bfqq,
1397                                                 bool arrived_in_time,
1398                                                 bool wr_or_deserves_wr)
1399 {
1400         struct bfq_entity *entity = &bfqq->entity;
1401 
1402         /*
1403          * In the next compound condition, we check also whether there
1404          * is some budget left, because otherwise there is no point in
1405          * trying to go on serving bfqq with this same budget: bfqq
1406          * would be expired immediately after being selected for
1407          * service. This would only cause useless overhead.
1408          */
1409         if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1410             bfq_bfqq_budget_left(bfqq) > 0) {
1411                 /*
1412                  * We do not clear the flag non_blocking_wait_rq here, as
1413                  * the latter is used in bfq_activate_bfqq to signal
1414                  * that timestamps need to be back-shifted (and is
1415                  * cleared right after).
1416                  */
1417 
1418                 /*
1419                  * In next assignment we rely on that either
1420                  * entity->service or entity->budget are not updated
1421                  * on expiration if bfqq is empty (see
1422                  * __bfq_bfqq_recalc_budget). Thus both quantities
1423                  * remain unchanged after such an expiration, and the
1424                  * following statement therefore assigns to
1425                  * entity->budget the remaining budget on such an
1426                  * expiration.
1427                  */
1428                 entity->budget = min_t(unsigned long,
1429                                        bfq_bfqq_budget_left(bfqq),
1430                                        bfqq->max_budget);
1431 
1432                 /*
1433                  * At this point, we have used entity->service to get
1434                  * the budget left (needed for updating
1435                  * entity->budget). Thus we finally can, and have to,
1436                  * reset entity->service. The latter must be reset
1437                  * because bfqq would otherwise be charged again for
1438                  * the service it has received during its previous
1439                  * service slot(s).
1440                  */
1441                 entity->service = 0;
1442 
1443                 return true;
1444         }
1445 
1446         /*
1447          * We can finally complete expiration, by setting service to 0.
1448          */
1449         entity->service = 0;
1450         entity->budget = max_t(unsigned long, bfqq->max_budget,
1451                                bfq_serv_to_charge(bfqq->next_rq, bfqq));
1452         bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1453         return wr_or_deserves_wr;
1454 }
1455 
1456 /*
1457  * Return the farthest past time instant according to jiffies
1458  * macros.
1459  */
1460 static unsigned long bfq_smallest_from_now(void)
1461 {
1462         return jiffies - MAX_JIFFY_OFFSET;
1463 }
1464 
1465 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1466                                              struct bfq_queue *bfqq,
1467                                              unsigned int old_wr_coeff,
1468                                              bool wr_or_deserves_wr,
1469                                              bool interactive,
1470                                              bool in_burst,
1471                                              bool soft_rt)
1472 {
1473         if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1474                 /* start a weight-raising period */
1475                 if (interactive) {
1476                         bfqq->service_from_wr = 0;
1477                         bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1478                         bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1479                 } else {
1480                         /*
1481                          * No interactive weight raising in progress
1482                          * here: assign minus infinity to
1483                          * wr_start_at_switch_to_srt, to make sure
1484                          * that, at the end of the soft-real-time
1485                          * weight raising periods that is starting
1486                          * now, no interactive weight-raising period
1487                          * may be wrongly considered as still in
1488                          * progress (and thus actually started by
1489                          * mistake).
1490                          */
1491                         bfqq->wr_start_at_switch_to_srt =
1492                                 bfq_smallest_from_now();
1493                         bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1494                                 BFQ_SOFTRT_WEIGHT_FACTOR;
1495                         bfqq->wr_cur_max_time =
1496                                 bfqd->bfq_wr_rt_max_time;
1497                 }
1498 
1499                 /*
1500                  * If needed, further reduce budget to make sure it is
1501                  * close to bfqq's backlog, so as to reduce the
1502                  * scheduling-error component due to a too large
1503                  * budget. Do not care about throughput consequences,
1504                  * but only about latency. Finally, do not assign a
1505                  * too small budget either, to avoid increasing
1506                  * latency by causing too frequent expirations.
1507                  */
1508                 bfqq->entity.budget = min_t(unsigned long,
1509                                             bfqq->entity.budget,
1510                                             2 * bfq_min_budget(bfqd));
1511         } else if (old_wr_coeff > 1) {
1512                 if (interactive) { /* update wr coeff and duration */
1513                         bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1514                         bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1515                 } else if (in_burst)
1516                         bfqq->wr_coeff = 1;
1517                 else if (soft_rt) {
1518                         /*
1519                          * The application is now or still meeting the
1520                          * requirements for being deemed soft rt.  We
1521                          * can then correctly and safely (re)charge
1522                          * the weight-raising duration for the
1523                          * application with the weight-raising
1524                          * duration for soft rt applications.
1525                          *
1526                          * In particular, doing this recharge now, i.e.,
1527                          * before the weight-raising period for the
1528                          * application finishes, reduces the probability
1529                          * of the following negative scenario:
1530                          * 1) the weight of a soft rt application is
1531                          *    raised at startup (as for any newly
1532                          *    created application),
1533                          * 2) since the application is not interactive,
1534                          *    at a certain time weight-raising is
1535                          *    stopped for the application,
1536                          * 3) at that time the application happens to
1537                          *    still have pending requests, and hence
1538                          *    is destined to not have a chance to be
1539                          *    deemed soft rt before these requests are
1540                          *    completed (see the comments to the
1541                          *    function bfq_bfqq_softrt_next_start()
1542                          *    for details on soft rt detection),
1543                          * 4) these pending requests experience a high
1544                          *    latency because the application is not
1545                          *    weight-raised while they are pending.
1546                          */
1547                         if (bfqq->wr_cur_max_time !=
1548                                 bfqd->bfq_wr_rt_max_time) {
1549                                 bfqq->wr_start_at_switch_to_srt =
1550                                         bfqq->last_wr_start_finish;
1551 
1552                                 bfqq->wr_cur_max_time =
1553                                         bfqd->bfq_wr_rt_max_time;
1554                                 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1555                                         BFQ_SOFTRT_WEIGHT_FACTOR;
1556                         }
1557                         bfqq->last_wr_start_finish = jiffies;
1558                 }
1559         }
1560 }
1561 
1562 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1563                                         struct bfq_queue *bfqq)
1564 {
1565         return bfqq->dispatched == 0 &&
1566                 time_is_before_jiffies(
1567                         bfqq->budget_timeout +
1568                         bfqd->bfq_wr_min_idle_time);
1569 }
1570 
1571 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1572                                              struct bfq_queue *bfqq,
1573                                              int old_wr_coeff,
1574                                              struct request *rq,
1575                                              bool *interactive)
1576 {
1577         bool soft_rt, in_burst, wr_or_deserves_wr,
1578                 bfqq_wants_to_preempt,
1579                 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1580                 /*
1581                  * See the comments on
1582                  * bfq_bfqq_update_budg_for_activation for
1583                  * details on the usage of the next variable.
1584                  */
1585                 arrived_in_time =  ktime_get_ns() <=
1586                         bfqq->ttime.last_end_request +
1587                         bfqd->bfq_slice_idle * 3;
1588 
1589 
1590         /*
1591          * bfqq deserves to be weight-raised if:
1592          * - it is sync,
1593          * - it does not belong to a large burst,
1594          * - it has been idle for enough time or is soft real-time,
1595          * - is linked to a bfq_io_cq (it is not shared in any sense).
1596          */
1597         in_burst = bfq_bfqq_in_large_burst(bfqq);
1598         soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1599                 !in_burst &&
1600                 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1601                 bfqq->dispatched == 0;
1602         *interactive = !in_burst && idle_for_long_time;
1603         wr_or_deserves_wr = bfqd->low_latency &&
1604                 (bfqq->wr_coeff > 1 ||
1605                  (bfq_bfqq_sync(bfqq) &&
1606                   bfqq->bic && (*interactive || soft_rt)));
1607 
1608         /*
1609          * Using the last flag, update budget and check whether bfqq
1610          * may want to preempt the in-service queue.
1611          */
1612         bfqq_wants_to_preempt =
1613                 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1614                                                     arrived_in_time,
1615                                                     wr_or_deserves_wr);
1616 
1617         /*
1618          * If bfqq happened to be activated in a burst, but has been
1619          * idle for much more than an interactive queue, then we
1620          * assume that, in the overall I/O initiated in the burst, the
1621          * I/O associated with bfqq is finished. So bfqq does not need
1622          * to be treated as a queue belonging to a burst
1623          * anymore. Accordingly, we reset bfqq's in_large_burst flag
1624          * if set, and remove bfqq from the burst list if it's
1625          * there. We do not decrement burst_size, because the fact
1626          * that bfqq does not need to belong to the burst list any
1627          * more does not invalidate the fact that bfqq was created in
1628          * a burst.
1629          */
1630         if (likely(!bfq_bfqq_just_created(bfqq)) &&
1631             idle_for_long_time &&
1632             time_is_before_jiffies(
1633                     bfqq->budget_timeout +
1634                     msecs_to_jiffies(10000))) {
1635                 hlist_del_init(&bfqq->burst_list_node);
1636                 bfq_clear_bfqq_in_large_burst(bfqq);
1637         }
1638 
1639         bfq_clear_bfqq_just_created(bfqq);
1640 
1641 
1642         if (!bfq_bfqq_IO_bound(bfqq)) {
1643                 if (arrived_in_time) {
1644                         bfqq->requests_within_timer++;
1645                         if (bfqq->requests_within_timer >=
1646                             bfqd->bfq_requests_within_timer)
1647                                 bfq_mark_bfqq_IO_bound(bfqq);
1648                 } else
1649                         bfqq->requests_within_timer = 0;
1650         }
1651 
1652         if (bfqd->low_latency) {
1653                 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1654                         /* wraparound */
1655                         bfqq->split_time =
1656                                 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1657 
1658                 if (time_is_before_jiffies(bfqq->split_time +
1659                                            bfqd->bfq_wr_min_idle_time)) {
1660                         bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1661                                                          old_wr_coeff,
1662                                                          wr_or_deserves_wr,
1663                                                          *interactive,
1664                                                          in_burst,
1665                                                          soft_rt);
1666 
1667                         if (old_wr_coeff != bfqq->wr_coeff)
1668                                 bfqq->entity.prio_changed = 1;
1669                 }
1670         }
1671 
1672         bfqq->last_idle_bklogged = jiffies;
1673         bfqq->service_from_backlogged = 0;
1674         bfq_clear_bfqq_softrt_update(bfqq);
1675 
1676         bfq_add_bfqq_busy(bfqd, bfqq);
1677 
1678         /*
1679          * Expire in-service queue only if preemption may be needed
1680          * for guarantees. In this respect, the function
1681          * next_queue_may_preempt just checks a simple, necessary
1682          * condition, and not a sufficient condition based on
1683          * timestamps. In fact, for the latter condition to be
1684          * evaluated, timestamps would need first to be updated, and
1685          * this operation is quite costly (see the comments on the
1686          * function bfq_bfqq_update_budg_for_activation).
1687          */
1688         if (bfqd->in_service_queue && bfqq_wants_to_preempt &&
1689             bfqd->in_service_queue->wr_coeff < bfqq->wr_coeff &&
1690             next_queue_may_preempt(bfqd))
1691                 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1692                                 false, BFQQE_PREEMPTED);
1693 }
1694 
1695 static void bfq_add_request(struct request *rq)
1696 {
1697         struct bfq_queue *bfqq = RQ_BFQQ(rq);
1698         struct bfq_data *bfqd = bfqq->bfqd;
1699         struct request *next_rq, *prev;
1700         unsigned int old_wr_coeff = bfqq->wr_coeff;
1701         bool interactive = false;
1702 
1703         bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1704         bfqq->queued[rq_is_sync(rq)]++;
1705         bfqd->queued++;
1706 
1707         elv_rb_add(&bfqq->sort_list, rq);
1708 
1709         /*
1710          * Check if this request is a better next-serve candidate.
1711          */
1712         prev = bfqq->next_rq;
1713         next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
1714         bfqq->next_rq = next_rq;
1715 
1716         /*
1717          * Adjust priority tree position, if next_rq changes.
1718          */
1719         if (prev != bfqq->next_rq)
1720                 bfq_pos_tree_add_move(bfqd, bfqq);
1721 
1722         if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
1723                 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
1724                                                  rq, &interactive);
1725         else {
1726                 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
1727                     time_is_before_jiffies(
1728                                 bfqq->last_wr_start_finish +
1729                                 bfqd->bfq_wr_min_inter_arr_async)) {
1730                         bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1731                         bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1732 
1733                         bfqd->wr_busy_queues++;
1734                         bfqq->entity.prio_changed = 1;
1735                 }
1736                 if (prev != bfqq->next_rq)
1737                         bfq_updated_next_req(bfqd, bfqq);
1738         }
1739 
1740         /*
1741          * Assign jiffies to last_wr_start_finish in the following
1742          * cases:
1743          *
1744          * . if bfqq is not going to be weight-raised, because, for
1745          *   non weight-raised queues, last_wr_start_finish stores the
1746          *   arrival time of the last request; as of now, this piece
1747          *   of information is used only for deciding whether to
1748          *   weight-raise async queues
1749          *
1750          * . if bfqq is not weight-raised, because, if bfqq is now
1751          *   switching to weight-raised, then last_wr_start_finish
1752          *   stores the time when weight-raising starts
1753          *
1754          * . if bfqq is interactive, because, regardless of whether
1755          *   bfqq is currently weight-raised, the weight-raising
1756          *   period must start or restart (this case is considered
1757          *   separately because it is not detected by the above
1758          *   conditions, if bfqq is already weight-raised)
1759          *
1760          * last_wr_start_finish has to be updated also if bfqq is soft
1761          * real-time, because the weight-raising period is constantly
1762          * restarted on idle-to-busy transitions for these queues, but
1763          * this is already done in bfq_bfqq_handle_idle_busy_switch if
1764          * needed.
1765          */
1766         if (bfqd->low_latency &&
1767                 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
1768                 bfqq->last_wr_start_finish = jiffies;
1769 }
1770 
1771 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
1772                                           struct bio *bio,
1773                                           struct request_queue *q)
1774 {
1775         struct bfq_queue *bfqq = bfqd->bio_bfqq;
1776 
1777 
1778         if (bfqq)
1779                 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
1780 
1781         return NULL;
1782 }
1783 
1784 static sector_t get_sdist(sector_t last_pos, struct request *rq)
1785 {
1786         if (last_pos)
1787                 return abs(blk_rq_pos(rq) - last_pos);
1788 
1789         return 0;
1790 }
1791 
1792 #if 0 /* Still not clear if we can do without next two functions */
1793 static void bfq_activate_request(struct request_queue *q, struct request *rq)
1794 {
1795         struct bfq_data *bfqd = q->elevator->elevator_data;
1796 
1797         bfqd->rq_in_driver++;
1798 }
1799 
1800 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
1801 {
1802         struct bfq_data *bfqd = q->elevator->elevator_data;
1803 
1804         bfqd->rq_in_driver--;
1805 }
1806 #endif
1807 
1808 static void bfq_remove_request(struct request_queue *q,
1809                                struct request *rq)
1810 {
1811         struct bfq_queue *bfqq = RQ_BFQQ(rq);
1812         struct bfq_data *bfqd = bfqq->bfqd;
1813         const int sync = rq_is_sync(rq);
1814 
1815         if (bfqq->next_rq == rq) {
1816                 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
1817                 bfq_updated_next_req(bfqd, bfqq);
1818         }
1819 
1820         if (rq->queuelist.prev != &rq->queuelist)
1821                 list_del_init(&rq->queuelist);
1822         bfqq->queued[sync]--;
1823         bfqd->queued--;
1824         elv_rb_del(&bfqq->sort_list, rq);
1825 
1826         elv_rqhash_del(q, rq);
1827         if (q->last_merge == rq)
1828                 q->last_merge = NULL;
1829 
1830         if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
1831                 bfqq->next_rq = NULL;
1832 
1833                 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
1834                         bfq_del_bfqq_busy(bfqd, bfqq, false);
1835                         /*
1836                          * bfqq emptied. In normal operation, when
1837                          * bfqq is empty, bfqq->entity.service and
1838                          * bfqq->entity.budget must contain,
1839                          * respectively, the service received and the
1840                          * budget used last time bfqq emptied. These
1841                          * facts do not hold in this case, as at least
1842                          * this last removal occurred while bfqq is
1843                          * not in service. To avoid inconsistencies,
1844                          * reset both bfqq->entity.service and
1845                          * bfqq->entity.budget, if bfqq has still a
1846                          * process that may issue I/O requests to it.
1847                          */
1848                         bfqq->entity.budget = bfqq->entity.service = 0;
1849                 }
1850 
1851                 /*
1852                  * Remove queue from request-position tree as it is empty.
1853                  */
1854                 if (bfqq->pos_root) {
1855                         rb_erase(&bfqq->pos_node, bfqq->pos_root);
1856                         bfqq->pos_root = NULL;
1857                 }
1858         } else {
1859                 bfq_pos_tree_add_move(bfqd, bfqq);
1860         }
1861 
1862         if (rq->cmd_flags & REQ_META)
1863                 bfqq->meta_pending--;
1864 
1865 }
1866 
1867 static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio)
1868 {
1869         struct request_queue *q = hctx->queue;
1870         struct bfq_data *bfqd = q->elevator->elevator_data;
1871         struct request *free = NULL;
1872         /*
1873          * bfq_bic_lookup grabs the queue_lock: invoke it now and
1874          * store its return value for later use, to avoid nesting
1875          * queue_lock inside the bfqd->lock. We assume that the bic
1876          * returned by bfq_bic_lookup does not go away before
1877          * bfqd->lock is taken.
1878          */
1879         struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
1880         bool ret;
1881 
1882         spin_lock_irq(&bfqd->lock);
1883 
1884         if (bic)
1885                 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
1886         else
1887                 bfqd->bio_bfqq = NULL;
1888         bfqd->bio_bic = bic;
1889 
1890         ret = blk_mq_sched_try_merge(q, bio, &free);
1891 
1892         if (free)
1893                 blk_mq_free_request(free);
1894         spin_unlock_irq(&bfqd->lock);
1895 
1896         return ret;
1897 }
1898 
1899 static int bfq_request_merge(struct request_queue *q, struct request **req,
1900                              struct bio *bio)
1901 {
1902         struct bfq_data *bfqd = q->elevator->elevator_data;
1903         struct request *__rq;
1904 
1905         __rq = bfq_find_rq_fmerge(bfqd, bio, q);
1906         if (__rq && elv_bio_merge_ok(__rq, bio)) {
1907                 *req = __rq;
1908                 return ELEVATOR_FRONT_MERGE;
1909         }
1910 
1911         return ELEVATOR_NO_MERGE;
1912 }
1913 
1914 static struct bfq_queue *bfq_init_rq(struct request *rq);
1915 
1916 static void bfq_request_merged(struct request_queue *q, struct request *req,
1917                                enum elv_merge type)
1918 {
1919         if (type == ELEVATOR_FRONT_MERGE &&
1920             rb_prev(&req->rb_node) &&
1921             blk_rq_pos(req) <
1922             blk_rq_pos(container_of(rb_prev(&req->rb_node),
1923                                     struct request, rb_node))) {
1924                 struct bfq_queue *bfqq = bfq_init_rq(req);
1925                 struct bfq_data *bfqd = bfqq->bfqd;
1926                 struct request *prev, *next_rq;
1927 
1928                 /* Reposition request in its sort_list */
1929                 elv_rb_del(&bfqq->sort_list, req);
1930                 elv_rb_add(&bfqq->sort_list, req);
1931 
1932                 /* Choose next request to be served for bfqq */
1933                 prev = bfqq->next_rq;
1934                 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
1935                                          bfqd->last_position);
1936                 bfqq->next_rq = next_rq;
1937                 /*
1938                  * If next_rq changes, update both the queue's budget to
1939                  * fit the new request and the queue's position in its
1940                  * rq_pos_tree.
1941                  */
1942                 if (prev != bfqq->next_rq) {
1943                         bfq_updated_next_req(bfqd, bfqq);
1944                         bfq_pos_tree_add_move(bfqd, bfqq);
1945                 }
1946         }
1947 }
1948 
1949 /*
1950  * This function is called to notify the scheduler that the requests
1951  * rq and 'next' have been merged, with 'next' going away.  BFQ
1952  * exploits this hook to address the following issue: if 'next' has a
1953  * fifo_time lower that rq, then the fifo_time of rq must be set to
1954  * the value of 'next', to not forget the greater age of 'next'.
1955  *
1956  * NOTE: in this function we assume that rq is in a bfq_queue, basing
1957  * on that rq is picked from the hash table q->elevator->hash, which,
1958  * in its turn, is filled only with I/O requests present in
1959  * bfq_queues, while BFQ is in use for the request queue q. In fact,
1960  * the function that fills this hash table (elv_rqhash_add) is called
1961  * only by bfq_insert_request.
1962  */
1963 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
1964                                 struct request *next)
1965 {
1966         struct bfq_queue *bfqq = bfq_init_rq(rq),
1967                 *next_bfqq = bfq_init_rq(next);
1968 
1969         /*
1970          * If next and rq belong to the same bfq_queue and next is older
1971          * than rq, then reposition rq in the fifo (by substituting next
1972          * with rq). Otherwise, if next and rq belong to different
1973          * bfq_queues, never reposition rq: in fact, we would have to
1974          * reposition it with respect to next's position in its own fifo,
1975          * which would most certainly be too expensive with respect to
1976          * the benefits.
1977          */
1978         if (bfqq == next_bfqq &&
1979             !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1980             next->fifo_time < rq->fifo_time) {
1981                 list_del_init(&rq->queuelist);
1982                 list_replace_init(&next->queuelist, &rq->queuelist);
1983                 rq->fifo_time = next->fifo_time;
1984         }
1985 
1986         if (bfqq->next_rq == next)
1987                 bfqq->next_rq = rq;
1988 
1989         bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
1990 }
1991 
1992 /* Must be called with bfqq != NULL */
1993 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
1994 {
1995         if (bfq_bfqq_busy(bfqq))
1996                 bfqq->bfqd->wr_busy_queues--;
1997         bfqq->wr_coeff = 1;
1998         bfqq->wr_cur_max_time = 0;
1999         bfqq->last_wr_start_finish = jiffies;
2000         /*
2001          * Trigger a weight change on the next invocation of
2002          * __bfq_entity_update_weight_prio.
2003          */
2004         bfqq->entity.prio_changed = 1;
2005 }
2006 
2007 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2008                              struct bfq_group *bfqg)
2009 {
2010         int i, j;
2011 
2012         for (i = 0; i < 2; i++)
2013                 for (j = 0; j < IOPRIO_BE_NR; j++)
2014                         if (bfqg->async_bfqq[i][j])
2015                                 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2016         if (bfqg->async_idle_bfqq)
2017                 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2018 }
2019 
2020 static void bfq_end_wr(struct bfq_data *bfqd)
2021 {
2022         struct bfq_queue *bfqq;
2023 
2024         spin_lock_irq(&bfqd->lock);
2025 
2026         list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2027                 bfq_bfqq_end_wr(bfqq);
2028         list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2029                 bfq_bfqq_end_wr(bfqq);
2030         bfq_end_wr_async(bfqd);
2031 
2032         spin_unlock_irq(&bfqd->lock);
2033 }
2034 
2035 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2036 {
2037         if (request)
2038                 return blk_rq_pos(io_struct);
2039         else
2040                 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2041 }
2042 
2043 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2044                                   sector_t sector)
2045 {
2046         return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2047                BFQQ_CLOSE_THR;
2048 }
2049 
2050 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2051                                          struct bfq_queue *bfqq,
2052                                          sector_t sector)
2053 {
2054         struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
2055         struct rb_node *parent, *node;
2056         struct bfq_queue *__bfqq;
2057 
2058         if (RB_EMPTY_ROOT(root))
2059                 return NULL;
2060 
2061         /*
2062          * First, if we find a request starting at the end of the last
2063          * request, choose it.
2064          */
2065         __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2066         if (__bfqq)
2067                 return __bfqq;
2068 
2069         /*
2070          * If the exact sector wasn't found, the parent of the NULL leaf
2071          * will contain the closest sector (rq_pos_tree sorted by
2072          * next_request position).
2073          */
2074         __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2075         if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2076                 return __bfqq;
2077 
2078         if (blk_rq_pos(__bfqq->next_rq) < sector)
2079                 node = rb_next(&__bfqq->pos_node);
2080         else
2081                 node = rb_prev(&__bfqq->pos_node);
2082         if (!node)
2083                 return NULL;
2084 
2085         __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2086         if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2087                 return __bfqq;
2088 
2089         return NULL;
2090 }
2091 
2092 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2093                                                    struct bfq_queue *cur_bfqq,
2094                                                    sector_t sector)
2095 {
2096         struct bfq_queue *bfqq;
2097 
2098         /*
2099          * We shall notice if some of the queues are cooperating,
2100          * e.g., working closely on the same area of the device. In
2101          * that case, we can group them together and: 1) don't waste
2102          * time idling, and 2) serve the union of their requests in
2103          * the best possible order for throughput.
2104          */
2105         bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2106         if (!bfqq || bfqq == cur_bfqq)
2107                 return NULL;
2108 
2109         return bfqq;
2110 }
2111 
2112 static struct bfq_queue *
2113 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2114 {
2115         int process_refs, new_process_refs;
2116         struct bfq_queue *__bfqq;
2117 
2118         /*
2119          * If there are no process references on the new_bfqq, then it is
2120          * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2121          * may have dropped their last reference (not just their last process
2122          * reference).
2123          */
2124         if (!bfqq_process_refs(new_bfqq))
2125                 return NULL;
2126 
2127         /* Avoid a circular list and skip interim queue merges. */
2128         while ((__bfqq = new_bfqq->new_bfqq)) {
2129                 if (__bfqq == bfqq)
2130                         return NULL;
2131                 new_bfqq = __bfqq;
2132         }
2133 
2134         process_refs = bfqq_process_refs(bfqq);
2135         new_process_refs = bfqq_process_refs(new_bfqq);
2136         /*
2137          * If the process for the bfqq has gone away, there is no
2138          * sense in merging the queues.
2139          */
2140         if (process_refs == 0 || new_process_refs == 0)
2141                 return NULL;
2142 
2143         bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2144                 new_bfqq->pid);
2145 
2146         /*
2147          * Merging is just a redirection: the requests of the process
2148          * owning one of the two queues are redirected to the other queue.
2149          * The latter queue, in its turn, is set as shared if this is the
2150          * first time that the requests of some process are redirected to
2151          * it.
2152          *
2153          * We redirect bfqq to new_bfqq and not the opposite, because
2154          * we are in the context of the process owning bfqq, thus we
2155          * have the io_cq of this process. So we can immediately
2156          * configure this io_cq to redirect the requests of the
2157          * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2158          * not available any more (new_bfqq->bic == NULL).
2159          *
2160          * Anyway, even in case new_bfqq coincides with the in-service
2161          * queue, redirecting requests the in-service queue is the
2162          * best option, as we feed the in-service queue with new
2163          * requests close to the last request served and, by doing so,
2164          * are likely to increase the throughput.
2165          */
2166         bfqq->new_bfqq = new_bfqq;
2167         new_bfqq->ref += process_refs;
2168         return new_bfqq;
2169 }
2170 
2171 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2172                                         struct bfq_queue *new_bfqq)
2173 {
2174         if (bfq_too_late_for_merging(new_bfqq))
2175                 return false;
2176 
2177         if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2178             (bfqq->ioprio_class != new_bfqq->ioprio_class))
2179                 return false;
2180 
2181         /*
2182          * If either of the queues has already been detected as seeky,
2183          * then merging it with the other queue is unlikely to lead to
2184          * sequential I/O.
2185          */
2186         if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2187                 return false;
2188 
2189         /*
2190          * Interleaved I/O is known to be done by (some) applications
2191          * only for reads, so it does not make sense to merge async
2192          * queues.
2193          */
2194         if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2195                 return false;
2196 
2197         return true;
2198 }
2199 
2200 /*
2201  * Attempt to schedule a merge of bfqq with the currently in-service
2202  * queue or with a close queue among the scheduled queues.  Return
2203  * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2204  * structure otherwise.
2205  *
2206  * The OOM queue is not allowed to participate to cooperation: in fact, since
2207  * the requests temporarily redirected to the OOM queue could be redirected
2208  * again to dedicated queues at any time, the state needed to correctly
2209  * handle merging with the OOM queue would be quite complex and expensive
2210  * to maintain. Besides, in such a critical condition as an out of memory,
2211  * the benefits of queue merging may be little relevant, or even negligible.
2212  *
2213  * WARNING: queue merging may impair fairness among non-weight raised
2214  * queues, for at least two reasons: 1) the original weight of a
2215  * merged queue may change during the merged state, 2) even being the
2216  * weight the same, a merged queue may be bloated with many more
2217  * requests than the ones produced by its originally-associated
2218  * process.
2219  */
2220 static struct bfq_queue *
2221 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2222                      void *io_struct, bool request)
2223 {
2224         struct bfq_queue *in_service_bfqq, *new_bfqq;
2225 
2226         /*
2227          * Prevent bfqq from being merged if it has been created too
2228          * long ago. The idea is that true cooperating processes, and
2229          * thus their associated bfq_queues, are supposed to be
2230          * created shortly after each other. This is the case, e.g.,
2231          * for KVM/QEMU and dump I/O threads. Basing on this
2232          * assumption, the following filtering greatly reduces the
2233          * probability that two non-cooperating processes, which just
2234          * happen to do close I/O for some short time interval, have
2235          * their queues merged by mistake.
2236          */
2237         if (bfq_too_late_for_merging(bfqq))
2238                 return NULL;
2239 
2240         if (bfqq->new_bfqq)
2241                 return bfqq->new_bfqq;
2242 
2243         if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2244                 return NULL;
2245 
2246         /* If there is only one backlogged queue, don't search. */
2247         if (bfq_tot_busy_queues(bfqd) == 1)
2248                 return NULL;
2249 
2250         in_service_bfqq = bfqd->in_service_queue;
2251 
2252         if (in_service_bfqq && in_service_bfqq != bfqq &&
2253             likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2254             bfq_rq_close_to_sector(io_struct, request,
2255                                    bfqd->in_serv_last_pos) &&
2256             bfqq->entity.parent == in_service_bfqq->entity.parent &&
2257             bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2258                 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2259                 if (new_bfqq)
2260                         return new_bfqq;
2261         }
2262         /*
2263          * Check whether there is a cooperator among currently scheduled
2264          * queues. The only thing we need is that the bio/request is not
2265          * NULL, as we need it to establish whether a cooperator exists.
2266          */
2267         new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2268                         bfq_io_struct_pos(io_struct, request));
2269 
2270         if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2271             bfq_may_be_close_cooperator(bfqq, new_bfqq))
2272                 return bfq_setup_merge(bfqq, new_bfqq);
2273 
2274         return NULL;
2275 }
2276 
2277 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2278 {
2279         struct bfq_io_cq *bic = bfqq->bic;
2280 
2281         /*
2282          * If !bfqq->bic, the queue is already shared or its requests
2283          * have already been redirected to a shared queue; both idle window
2284          * and weight raising state have already been saved. Do nothing.
2285          */
2286         if (!bic)
2287                 return;
2288 
2289         bic->saved_ttime = bfqq->ttime;
2290         bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2291         bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2292         bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2293         bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2294         if (unlikely(bfq_bfqq_just_created(bfqq) &&
2295                      !bfq_bfqq_in_large_burst(bfqq) &&
2296                      bfqq->bfqd->low_latency)) {
2297                 /*
2298                  * bfqq being merged right after being created: bfqq
2299                  * would have deserved interactive weight raising, but
2300                  * did not make it to be set in a weight-raised state,
2301                  * because of this early merge. Store directly the
2302                  * weight-raising state that would have been assigned
2303                  * to bfqq, so that to avoid that bfqq unjustly fails
2304                  * to enjoy weight raising if split soon.
2305                  */
2306                 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
2307                 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2308                 bic->saved_last_wr_start_finish = jiffies;
2309         } else {
2310                 bic->saved_wr_coeff = bfqq->wr_coeff;
2311                 bic->saved_wr_start_at_switch_to_srt =
2312                         bfqq->wr_start_at_switch_to_srt;
2313                 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2314                 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2315         }
2316 }
2317 
2318 static void
2319 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2320                 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2321 {
2322         bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2323                 (unsigned long)new_bfqq->pid);
2324         /* Save weight raising and idle window of the merged queues */
2325         bfq_bfqq_save_state(bfqq);
2326         bfq_bfqq_save_state(new_bfqq);
2327         if (bfq_bfqq_IO_bound(bfqq))
2328                 bfq_mark_bfqq_IO_bound(new_bfqq);
2329         bfq_clear_bfqq_IO_bound(bfqq);
2330 
2331         /*
2332          * If bfqq is weight-raised, then let new_bfqq inherit
2333          * weight-raising. To reduce false positives, neglect the case
2334          * where bfqq has just been created, but has not yet made it
2335          * to be weight-raised (which may happen because EQM may merge
2336          * bfqq even before bfq_add_request is executed for the first
2337          * time for bfqq). Handling this case would however be very
2338          * easy, thanks to the flag just_created.
2339          */
2340         if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2341                 new_bfqq->wr_coeff = bfqq->wr_coeff;
2342                 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2343                 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2344                 new_bfqq->wr_start_at_switch_to_srt =
2345                         bfqq->wr_start_at_switch_to_srt;
2346                 if (bfq_bfqq_busy(new_bfqq))
2347                         bfqd->wr_busy_queues++;
2348                 new_bfqq->entity.prio_changed = 1;
2349         }
2350 
2351         if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2352                 bfqq->wr_coeff = 1;
2353                 bfqq->entity.prio_changed = 1;
2354                 if (bfq_bfqq_busy(bfqq))
2355                         bfqd->wr_busy_queues--;
2356         }
2357 
2358         bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2359                      bfqd->wr_busy_queues);
2360 
2361         /*
2362          * Merge queues (that is, let bic redirect its requests to new_bfqq)
2363          */
2364         bic_set_bfqq(bic, new_bfqq, 1);
2365         bfq_mark_bfqq_coop(new_bfqq);
2366         /*
2367          * new_bfqq now belongs to at least two bics (it is a shared queue):
2368          * set new_bfqq->bic to NULL. bfqq either:
2369          * - does not belong to any bic any more, and hence bfqq->bic must
2370          *   be set to NULL, or
2371          * - is a queue whose owning bics have already been redirected to a
2372          *   different queue, hence the queue is destined to not belong to
2373          *   any bic soon and bfqq->bic is already NULL (therefore the next
2374          *   assignment causes no harm).
2375          */
2376         new_bfqq->bic = NULL;
2377         bfqq->bic = NULL;
2378         /* release process reference to bfqq */
2379         bfq_put_queue(bfqq);
2380 }
2381 
2382 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2383                                 struct bio *bio)
2384 {
2385         struct bfq_data *bfqd = q->elevator->elevator_data;
2386         bool is_sync = op_is_sync(bio->bi_opf);
2387         struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2388 
2389         /*
2390          * Disallow merge of a sync bio into an async request.
2391          */
2392         if (is_sync && !rq_is_sync(rq))
2393                 return false;
2394 
2395         /*
2396          * Lookup the bfqq that this bio will be queued with. Allow
2397          * merge only if rq is queued there.
2398          */
2399         if (!bfqq)
2400                 return false;
2401 
2402         /*
2403          * We take advantage of this function to perform an early merge
2404          * of the queues of possible cooperating processes.
2405          */
2406         new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2407         if (new_bfqq) {
2408                 /*
2409                  * bic still points to bfqq, then it has not yet been
2410                  * redirected to some other bfq_queue, and a queue
2411                  * merge beween bfqq and new_bfqq can be safely
2412                  * fulfillled, i.e., bic can be redirected to new_bfqq
2413                  * and bfqq can be put.
2414                  */
2415                 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2416                                 new_bfqq);
2417                 /*
2418                  * If we get here, bio will be queued into new_queue,
2419                  * so use new_bfqq to decide whether bio and rq can be
2420                  * merged.
2421                  */
2422                 bfqq = new_bfqq;
2423 
2424                 /*
2425                  * Change also bqfd->bio_bfqq, as
2426                  * bfqd->bio_bic now points to new_bfqq, and
2427                  * this function may be invoked again (and then may
2428                  * use again bqfd->bio_bfqq).
2429                  */
2430                 bfqd->bio_bfqq = bfqq;
2431         }
2432 
2433         return bfqq == RQ_BFQQ(rq);
2434 }
2435 
2436 /*
2437  * Set the maximum time for the in-service queue to consume its
2438  * budget. This prevents seeky processes from lowering the throughput.
2439  * In practice, a time-slice service scheme is used with seeky
2440  * processes.
2441  */
2442 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2443                                    struct bfq_queue *bfqq)
2444 {
2445         unsigned int timeout_coeff;
2446 
2447         if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2448                 timeout_coeff = 1;
2449         else
2450                 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2451 
2452         bfqd->last_budget_start = ktime_get();
2453 
2454         bfqq->budget_timeout = jiffies +
2455                 bfqd->bfq_timeout * timeout_coeff;
2456 }
2457 
2458 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2459                                        struct bfq_queue *bfqq)
2460 {
2461         if (bfqq) {
2462                 bfq_clear_bfqq_fifo_expire(bfqq);
2463 
2464                 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2465 
2466                 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2467                     bfqq->wr_coeff > 1 &&
2468                     bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2469                     time_is_before_jiffies(bfqq->budget_timeout)) {
2470                         /*
2471                          * For soft real-time queues, move the start
2472                          * of the weight-raising period forward by the
2473                          * time the queue has not received any
2474                          * service. Otherwise, a relatively long
2475                          * service delay is likely to cause the
2476                          * weight-raising period of the queue to end,
2477                          * because of the short duration of the
2478                          * weight-raising period of a soft real-time
2479                          * queue.  It is worth noting that this move
2480                          * is not so dangerous for the other queues,
2481                          * because soft real-time queues are not
2482                          * greedy.
2483                          *
2484                          * To not add a further variable, we use the
2485                          * overloaded field budget_timeout to
2486                          * determine for how long the queue has not
2487                          * received service, i.e., how much time has
2488                          * elapsed since the queue expired. However,
2489                          * this is a little imprecise, because
2490                          * budget_timeout is set to jiffies if bfqq
2491                          * not only expires, but also remains with no
2492                          * request.
2493                          */
2494                         if (time_after(bfqq->budget_timeout,
2495                                        bfqq->last_wr_start_finish))
2496                                 bfqq->last_wr_start_finish +=
2497                                         jiffies - bfqq->budget_timeout;
2498                         else
2499                                 bfqq->last_wr_start_finish = jiffies;
2500                 }
2501 
2502                 bfq_set_budget_timeout(bfqd, bfqq);
2503                 bfq_log_bfqq(bfqd, bfqq,
2504                              "set_in_service_queue, cur-budget = %d",
2505                              bfqq->entity.budget);
2506         }
2507 
2508         bfqd->in_service_queue = bfqq;
2509 }
2510 
2511 /*
2512  * Get and set a new queue for service.
2513  */
2514 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2515 {
2516         struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2517 
2518         __bfq_set_in_service_queue(bfqd, bfqq);
2519         return bfqq;
2520 }
2521 
2522 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2523 {
2524         struct bfq_queue *bfqq = bfqd->in_service_queue;
2525         u32 sl;
2526 
2527         bfq_mark_bfqq_wait_request(bfqq);
2528 
2529         /*
2530          * We don't want to idle for seeks, but we do want to allow
2531          * fair distribution of slice time for a process doing back-to-back
2532          * seeks. So allow a little bit of time for him to submit a new rq.
2533          */
2534         sl = bfqd->bfq_slice_idle;
2535         /*
2536          * Unless the queue is being weight-raised or the scenario is
2537          * asymmetric, grant only minimum idle time if the queue
2538          * is seeky. A long idling is preserved for a weight-raised
2539          * queue, or, more in general, in an asymmetric scenario,
2540          * because a long idling is needed for guaranteeing to a queue
2541          * its reserved share of the throughput (in particular, it is
2542          * needed if the queue has a higher weight than some other
2543          * queue).
2544          */
2545         if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
2546             bfq_symmetric_scenario(bfqd))
2547                 sl = min_t(u64, sl, BFQ_MIN_TT);
2548 
2549         bfqd->last_idling_start = ktime_get();
2550         hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
2551                       HRTIMER_MODE_REL);
2552         bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
2553 }
2554 
2555 /*
2556  * In autotuning mode, max_budget is dynamically recomputed as the
2557  * amount of sectors transferred in timeout at the estimated peak
2558  * rate. This enables BFQ to utilize a full timeslice with a full
2559  * budget, even if the in-service queue is served at peak rate. And
2560  * this maximises throughput with sequential workloads.
2561  */
2562 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
2563 {
2564         return (u64)bfqd->peak_rate * USEC_PER_MSEC *
2565                 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
2566 }
2567 
2568 /*
2569  * Update parameters related to throughput and responsiveness, as a
2570  * function of the estimated peak rate. See comments on
2571  * bfq_calc_max_budget(), and on the ref_wr_duration array.
2572  */
2573 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
2574 {
2575         if (bfqd->bfq_user_max_budget == 0) {
2576                 bfqd->bfq_max_budget =
2577                         bfq_calc_max_budget(bfqd);
2578                 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
2579         }
2580 }
2581 
2582 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
2583                                        struct request *rq)
2584 {
2585         if (rq != NULL) { /* new rq dispatch now, reset accordingly */
2586                 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
2587                 bfqd->peak_rate_samples = 1;
2588                 bfqd->sequential_samples = 0;
2589                 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
2590                         blk_rq_sectors(rq);
2591         } else /* no new rq dispatched, just reset the number of samples */
2592                 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
2593 
2594         bfq_log(bfqd,
2595                 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
2596                 bfqd->peak_rate_samples, bfqd->sequential_samples,
2597                 bfqd->tot_sectors_dispatched);
2598 }
2599 
2600 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
2601 {
2602         u32 rate, weight, divisor;
2603 
2604         /*
2605          * For the convergence property to hold (see comments on
2606          * bfq_update_peak_rate()) and for the assessment to be
2607          * reliable, a minimum number of samples must be present, and
2608          * a minimum amount of time must have elapsed. If not so, do
2609          * not compute new rate. Just reset parameters, to get ready
2610          * for a new evaluation attempt.
2611          */
2612         if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
2613             bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
2614                 goto reset_computation;
2615 
2616         /*
2617          * If a new request completion has occurred after last
2618          * dispatch, then, to approximate the rate at which requests
2619          * have been served by the device, it is more precise to
2620          * extend the observation interval to the last completion.
2621          */
2622         bfqd->delta_from_first =
2623                 max_t(u64, bfqd->delta_from_first,
2624                       bfqd->last_completion - bfqd->first_dispatch);
2625 
2626         /*
2627          * Rate computed in sects/usec, and not sects/nsec, for
2628          * precision issues.
2629          */
2630         rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
2631                         div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
2632 
2633         /*
2634          * Peak rate not updated if:
2635          * - the percentage of sequential dispatches is below 3/4 of the
2636          *   total, and rate is below the current estimated peak rate
2637          * - rate is unreasonably high (> 20M sectors/sec)
2638          */
2639         if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
2640              rate <= bfqd->peak_rate) ||
2641                 rate > 20<<BFQ_RATE_SHIFT)
2642                 goto reset_computation;
2643 
2644         /*
2645          * We have to update the peak rate, at last! To this purpose,
2646          * we use a low-pass filter. We compute the smoothing constant
2647          * of the filter as a function of the 'weight' of the new
2648          * measured rate.
2649          *
2650          * As can be seen in next formulas, we define this weight as a
2651          * quantity proportional to how sequential the workload is,
2652          * and to how long the observation time interval is.
2653          *
2654          * The weight runs from 0 to 8. The maximum value of the
2655          * weight, 8, yields the minimum value for the smoothing
2656          * constant. At this minimum value for the smoothing constant,
2657          * the measured rate contributes for half of the next value of
2658          * the estimated peak rate.
2659          *
2660          * So, the first step is to compute the weight as a function
2661          * of how sequential the workload is. Note that the weight
2662          * cannot reach 9, because bfqd->sequential_samples cannot
2663          * become equal to bfqd->peak_rate_samples, which, in its
2664          * turn, holds true because bfqd->sequential_samples is not
2665          * incremented for the first sample.
2666          */
2667         weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
2668 
2669         /*
2670          * Second step: further refine the weight as a function of the
2671          * duration of the observation interval.
2672          */
2673         weight = min_t(u32, 8,
2674                        div_u64(weight * bfqd->delta_from_first,
2675                                BFQ_RATE_REF_INTERVAL));
2676 
2677         /*
2678          * Divisor ranging from 10, for minimum weight, to 2, for
2679          * maximum weight.
2680          */
2681         divisor = 10 - weight;
2682 
2683         /*
2684          * Finally, update peak rate:
2685          *
2686          * peak_rate = peak_rate * (divisor-1) / divisor  +  rate / divisor
2687          */
2688         bfqd->peak_rate *= divisor-1;
2689         bfqd->peak_rate /= divisor;
2690         rate /= divisor; /* smoothing constant alpha = 1/divisor */
2691 
2692         bfqd->peak_rate += rate;
2693 
2694         /*
2695          * For a very slow device, bfqd->peak_rate can reach 0 (see
2696          * the minimum representable values reported in the comments
2697          * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
2698          * divisions by zero where bfqd->peak_rate is used as a
2699          * divisor.
2700          */
2701         bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
2702 
2703         update_thr_responsiveness_params(bfqd);
2704 
2705 reset_computation:
2706         bfq_reset_rate_computation(bfqd, rq);
2707 }
2708 
2709 /*
2710  * Update the read/write peak rate (the main quantity used for
2711  * auto-tuning, see update_thr_responsiveness_params()).
2712  *
2713  * It is not trivial to estimate the peak rate (correctly): because of
2714  * the presence of sw and hw queues between the scheduler and the
2715  * device components that finally serve I/O requests, it is hard to
2716  * say exactly when a given dispatched request is served inside the
2717  * device, and for how long. As a consequence, it is hard to know
2718  * precisely at what rate a given set of requests is actually served
2719  * by the device.
2720  *
2721  * On the opposite end, the dispatch time of any request is trivially
2722  * available, and, from this piece of information, the "dispatch rate"
2723  * of requests can be immediately computed. So, the idea in the next
2724  * function is to use what is known, namely request dispatch times
2725  * (plus, when useful, request completion times), to estimate what is
2726  * unknown, namely in-device request service rate.
2727  *
2728  * The main issue is that, because of the above facts, the rate at
2729  * which a certain set of requests is dispatched over a certain time
2730  * interval can vary greatly with respect to the rate at which the
2731  * same requests are then served. But, since the size of any
2732  * intermediate queue is limited, and the service scheme is lossless
2733  * (no request is silently dropped), the following obvious convergence
2734  * property holds: the number of requests dispatched MUST become
2735  * closer and closer to the number of requests completed as the
2736  * observation interval grows. This is the key property used in
2737  * the next function to estimate the peak service rate as a function
2738  * of the observed dispatch rate. The function assumes to be invoked
2739  * on every request dispatch.
2740  */
2741 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
2742 {
2743         u64 now_ns = ktime_get_ns();
2744 
2745         if (bfqd->peak_rate_samples == 0) { /* first dispatch */
2746                 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
2747                         bfqd->peak_rate_samples);
2748                 bfq_reset_rate_computation(bfqd, rq);
2749                 goto update_last_values; /* will add one sample */
2750         }
2751 
2752         /*
2753          * Device idle for very long: the observation interval lasting
2754          * up to this dispatch cannot be a valid observation interval
2755          * for computing a new peak rate (similarly to the late-
2756          * completion event in bfq_completed_request()). Go to
2757          * update_rate_and_reset to have the following three steps
2758          * taken:
2759          * - close the observation interval at the last (previous)
2760          *   request dispatch or completion
2761          * - compute rate, if possible, for that observation interval
2762          * - start a new observation interval with this dispatch
2763          */
2764         if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
2765             bfqd->rq_in_driver == 0)
2766                 goto update_rate_and_reset;
2767 
2768         /* Update sampling information */
2769         bfqd->peak_rate_samples++;
2770 
2771         if ((bfqd->rq_in_driver > 0 ||
2772                 now_ns - bfqd->last_completion < BFQ_MIN_TT)
2773             && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
2774                 bfqd->sequential_samples++;
2775 
2776         bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
2777 
2778         /* Reset max observed rq size every 32 dispatches */
2779         if (likely(bfqd->peak_rate_samples % 32))
2780                 bfqd->last_rq_max_size =
2781                         max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
2782         else
2783                 bfqd->last_rq_max_size = blk_rq_sectors(rq);
2784 
2785         bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
2786 
2787         /* Target observation interval not yet reached, go on sampling */
2788         if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
2789                 goto update_last_values;
2790 
2791 update_rate_and_reset:
2792         bfq_update_rate_reset(bfqd, rq);
2793 update_last_values:
2794         bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2795         if (RQ_BFQQ(rq) == bfqd->in_service_queue)
2796                 bfqd->in_serv_last_pos = bfqd->last_position;
2797         bfqd->last_dispatch = now_ns;
2798 }
2799 
2800 /*
2801  * Remove request from internal lists.
2802  */
2803 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
2804 {
2805         struct bfq_queue *bfqq = RQ_BFQQ(rq);
2806 
2807         /*
2808          * For consistency, the next instruction should have been
2809          * executed after removing the request from the queue and
2810          * dispatching it.  We execute instead this instruction before
2811          * bfq_remove_request() (and hence introduce a temporary
2812          * inconsistency), for efficiency.  In fact, should this
2813          * dispatch occur for a non in-service bfqq, this anticipated
2814          * increment prevents two counters related to bfqq->dispatched
2815          * from risking to be, first, uselessly decremented, and then
2816          * incremented again when the (new) value of bfqq->dispatched
2817          * happens to be taken into account.
2818          */
2819         bfqq->dispatched++;
2820         bfq_update_peak_rate(q->elevator->elevator_data, rq);
2821 
2822         bfq_remove_request(q, rq);
2823 }
2824 
2825 static void __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq)
2826 {
2827         /*
2828          * If this bfqq is shared between multiple processes, check
2829          * to make sure that those processes are still issuing I/Os
2830          * within the mean seek distance. If not, it may be time to
2831          * break the queues apart again.
2832          */
2833         if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
2834                 bfq_mark_bfqq_split_coop(bfqq);
2835 
2836         if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2837                 if (bfqq->dispatched == 0)
2838                         /*
2839                          * Overloading budget_timeout field to store
2840                          * the time at which the queue remains with no
2841                          * backlog and no outstanding request; used by
2842                          * the weight-raising mechanism.
2843                          */
2844                         bfqq->budget_timeout = jiffies;
2845 
2846                 bfq_del_bfqq_busy(bfqd, bfqq, true);
2847         } else {
2848                 bfq_requeue_bfqq(bfqd, bfqq, true);
2849                 /*
2850                  * Resort priority tree of potential close cooperators.
2851                  */
2852                 bfq_pos_tree_add_move(bfqd, bfqq);
2853         }
2854 
2855         /*
2856          * All in-service entities must have been properly deactivated
2857          * or requeued before executing the next function, which
2858          * resets all in-service entites as no more in service.
2859          */
2860         __bfq_bfqd_reset_in_service(bfqd);
2861 }
2862 
2863 /**
2864  * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
2865  * @bfqd: device data.
2866  * @bfqq: queue to update.
2867  * @reason: reason for expiration.
2868  *
2869  * Handle the feedback on @bfqq budget at queue expiration.
2870  * See the body for detailed comments.
2871  */
2872 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
2873                                      struct bfq_queue *bfqq,
2874                                      enum bfqq_expiration reason)
2875 {
2876         struct request *next_rq;
2877         int budget, min_budget;
2878 
2879         min_budget = bfq_min_budget(bfqd);
2880 
2881         if (bfqq->wr_coeff == 1)
2882                 budget = bfqq->max_budget;
2883         else /*
2884               * Use a constant, low budget for weight-raised queues,
2885               * to help achieve a low latency. Keep it slightly higher
2886               * than the minimum possible budget, to cause a little
2887               * bit fewer expirations.
2888               */
2889                 budget = 2 * min_budget;
2890 
2891         bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
2892                 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
2893         bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
2894                 budget, bfq_min_budget(bfqd));
2895         bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
2896                 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
2897 
2898         if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
2899                 switch (reason) {
2900                 /*
2901                  * Caveat: in all the following cases we trade latency
2902                  * for throughput.
2903                  */
2904                 case BFQQE_TOO_IDLE:
2905                         /*
2906                          * This is the only case where we may reduce
2907                          * the budget: if there is no request of the
2908                          * process still waiting for completion, then
2909                          * we assume (tentatively) that the timer has
2910                          * expired because the batch of requests of
2911                          * the process could have been served with a
2912                          * smaller budget.  Hence, betting that
2913                          * process will behave in the same way when it
2914                          * becomes backlogged again, we reduce its
2915                          * next budget.  As long as we guess right,
2916                          * this budget cut reduces the latency
2917                          * experienced by the process.
2918                          *
2919                          * However, if there are still outstanding
2920                          * requests, then the process may have not yet
2921                          * issued its next request just because it is
2922                          * still waiting for the completion of some of
2923                          * the still outstanding ones.  So in this
2924                          * subcase we do not reduce its budget, on the
2925                          * contrary we increase it to possibly boost
2926                          * the throughput, as discussed in the
2927                          * comments to the BUDGET_TIMEOUT case.
2928                          */
2929                         if (bfqq->dispatched > 0) /* still outstanding reqs */
2930                                 budget = min(budget * 2, bfqd->bfq_max_budget);
2931                         else {
2932                                 if (budget > 5 * min_budget)
2933                                         budget -= 4 * min_budget;
2934                                 else
2935                                         budget = min_budget;
2936                         }
2937                         break;
2938                 case BFQQE_BUDGET_TIMEOUT:
2939                         /*
2940                          * We double the budget here because it gives
2941                          * the chance to boost the throughput if this
2942                          * is not a seeky process (and has bumped into
2943                          * this timeout because of, e.g., ZBR).
2944                          */
2945                         budget = min(budget * 2, bfqd->bfq_max_budget);
2946                         break;
2947                 case BFQQE_BUDGET_EXHAUSTED:
2948                         /*
2949                          * The process still has backlog, and did not
2950                          * let either the budget timeout or the disk
2951                          * idling timeout expire. Hence it is not
2952                          * seeky, has a short thinktime and may be
2953                          * happy with a higher budget too. So
2954                          * definitely increase the budget of this good
2955                          * candidate to boost the disk throughput.
2956                          */
2957                         budget = min(budget * 4, bfqd->bfq_max_budget);
2958                         break;
2959                 case BFQQE_NO_MORE_REQUESTS:
2960                         /*
2961                          * For queues that expire for this reason, it
2962                          * is particularly important to keep the
2963                          * budget close to the actual service they
2964                          * need. Doing so reduces the timestamp
2965                          * misalignment problem described in the
2966                          * comments in the body of
2967                          * __bfq_activate_entity. In fact, suppose
2968                          * that a queue systematically expires for
2969                          * BFQQE_NO_MORE_REQUESTS and presents a
2970                          * new request in time to enjoy timestamp
2971                          * back-shifting. The larger the budget of the
2972                          * queue is with respect to the service the
2973                          * queue actually requests in each service
2974                          * slot, the more times the queue can be
2975                          * reactivated with the same virtual finish
2976                          * time. It follows that, even if this finish
2977                          * time is pushed to the system virtual time
2978                          * to reduce the consequent timestamp
2979                          * misalignment, the queue unjustly enjoys for
2980                          * many re-activations a lower finish time
2981                          * than all newly activated queues.
2982                          *
2983                          * The service needed by bfqq is measured
2984                          * quite precisely by bfqq->entity.service.
2985                          * Since bfqq does not enjoy device idling,
2986                          * bfqq->entity.service is equal to the number
2987                          * of sectors that the process associated with
2988                          * bfqq requested to read/write before waiting
2989                          * for request completions, or blocking for
2990                          * other reasons.
2991                          */
2992                         budget = max_t(int, bfqq->entity.service, min_budget);
2993                         break;
2994                 default:
2995                         return;
2996                 }
2997         } else if (!bfq_bfqq_sync(bfqq)) {
2998                 /*
2999                  * Async queues get always the maximum possible
3000                  * budget, as for them we do not care about latency
3001                  * (in addition, their ability to dispatch is limited
3002                  * by the charging factor).
3003                  */
3004                 budget = bfqd->bfq_max_budget;
3005         }
3006 
3007         bfqq->max_budget = budget;
3008 
3009         if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
3010             !bfqd->bfq_user_max_budget)
3011                 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
3012 
3013         /*
3014          * If there is still backlog, then assign a new budget, making
3015          * sure that it is large enough for the next request.  Since
3016          * the finish time of bfqq must be kept in sync with the
3017          * budget, be sure to call __bfq_bfqq_expire() *after* this
3018          * update.
3019          *
3020          * If there is no backlog, then no need to update the budget;
3021          * it will be updated on the arrival of a new request.
3022          */
3023         next_rq = bfqq->next_rq;
3024         if (next_rq)
3025                 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
3026                                             bfq_serv_to_charge(next_rq, bfqq));
3027 
3028         bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
3029                         next_rq ? blk_rq_sectors(next_rq) : 0,
3030                         bfqq->entity.budget);
3031 }
3032 
3033 /*
3034  * Return true if the process associated with bfqq is "slow". The slow
3035  * flag is used, in addition to the budget timeout, to reduce the
3036  * amount of service provided to seeky processes, and thus reduce
3037  * their chances to lower the throughput. More details in the comments
3038  * on the function bfq_bfqq_expire().
3039  *
3040  * An important observation is in order: as discussed in the comments
3041  * on the function bfq_update_peak_rate(), with devices with internal
3042  * queues, it is hard if ever possible to know when and for how long
3043  * an I/O request is processed by the device (apart from the trivial
3044  * I/O pattern where a new request is dispatched only after the
3045  * previous one has been completed). This makes it hard to evaluate
3046  * the real rate at which the I/O requests of each bfq_queue are
3047  * served.  In fact, for an I/O scheduler like BFQ, serving a
3048  * bfq_queue means just dispatching its requests during its service
3049  * slot (i.e., until the budget of the queue is exhausted, or the
3050  * queue remains idle, or, finally, a timeout fires). But, during the
3051  * service slot of a bfq_queue, around 100 ms at most, the device may
3052  * be even still processing requests of bfq_queues served in previous
3053  * service slots. On the opposite end, the requests of the in-service
3054  * bfq_queue may be completed after the service slot of the queue
3055  * finishes.
3056  *
3057  * Anyway, unless more sophisticated solutions are used
3058  * (where possible), the sum of the sizes of the requests dispatched
3059  * during the service slot of a bfq_queue is probably the only
3060  * approximation available for the service received by the bfq_queue
3061  * during its service slot. And this sum is the quantity used in this
3062  * function to evaluate the I/O speed of a process.
3063  */
3064 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3065                                  bool compensate, enum bfqq_expiration reason,
3066                                  unsigned long *delta_ms)
3067 {
3068         ktime_t delta_ktime;
3069         u32 delta_usecs;
3070         bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
3071 
3072         if (!bfq_bfqq_sync(bfqq))
3073                 return false;
3074 
3075         if (compensate)
3076                 delta_ktime = bfqd->last_idling_start;
3077         else
3078                 delta_ktime = ktime_get();
3079         delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
3080         delta_usecs = ktime_to_us(delta_ktime);
3081 
3082         /* don't use too short time intervals */
3083         if (delta_usecs < 1000) {
3084                 if (blk_queue_nonrot(bfqd->queue))
3085                          /*
3086                           * give same worst-case guarantees as idling
3087                           * for seeky
3088                           */
3089                         *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
3090                 else /* charge at least one seek */
3091                         *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
3092 
3093                 return slow;
3094         }
3095 
3096         *delta_ms = delta_usecs / USEC_PER_MSEC;
3097 
3098         /*
3099          * Use only long (> 20ms) intervals to filter out excessive
3100          * spikes in service rate estimation.
3101          */
3102         if (delta_usecs > 20000) {
3103                 /*
3104                  * Caveat for rotational devices: processes doing I/O
3105                  * in the slower disk zones tend to be slow(er) even
3106                  * if not seeky. In this respect, the estimated peak
3107                  * rate is likely to be an average over the disk
3108                  * surface. Accordingly, to not be too harsh with
3109                  * unlucky processes, a process is deemed slow only if
3110                  * its rate has been lower than half of the estimated
3111                  * peak rate.
3112                  */
3113                 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
3114         }
3115 
3116         bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
3117 
3118         return slow;
3119 }
3120 
3121 /*
3122  * To be deemed as soft real-time, an application must meet two
3123  * requirements. First, the application must not require an average
3124  * bandwidth higher than the approximate bandwidth required to playback or
3125  * record a compressed high-definition video.
3126  * The next function is invoked on the completion of the last request of a
3127  * batch, to compute the next-start time instant, soft_rt_next_start, such
3128  * that, if the next request of the application does not arrive before
3129  * soft_rt_next_start, then the above requirement on the bandwidth is met.
3130  *
3131  * The second requirement is that the request pattern of the application is
3132  * isochronous, i.e., that, after issuing a request or a batch of requests,
3133  * the application stops issuing new requests until all its pending requests
3134  * have been completed. After that, the application may issue a new batch,
3135  * and so on.
3136  * For this reason the next function is invoked to compute
3137  * soft_rt_next_start only for applications that meet this requirement,
3138  * whereas soft_rt_next_start is set to infinity for applications that do
3139  * not.
3140  *
3141  * Unfortunately, even a greedy (i.e., I/O-bound) application may
3142  * happen to meet, occasionally or systematically, both the above
3143  * bandwidth and isochrony requirements. This may happen at least in
3144  * the following circumstances. First, if the CPU load is high. The
3145  * application may stop issuing requests while the CPUs are busy
3146  * serving other processes, then restart, then stop again for a while,
3147  * and so on. The other circumstances are related to the storage
3148  * device: the storage device is highly loaded or reaches a low-enough
3149  * throughput with the I/O of the application (e.g., because the I/O
3150  * is random and/or the device is slow). In all these cases, the
3151  * I/O of the application may be simply slowed down enough to meet
3152  * the bandwidth and isochrony requirements. To reduce the probability
3153  * that greedy applications are deemed as soft real-time in these
3154  * corner cases, a further rule is used in the computation of
3155  * soft_rt_next_start: the return value of this function is forced to
3156  * be higher than the maximum between the following two quantities.
3157  *
3158  * (a) Current time plus: (1) the maximum time for which the arrival
3159  *     of a request is waited for when a sync queue becomes idle,
3160  *     namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
3161  *     postpone for a moment the reason for adding a few extra
3162  *     jiffies; we get back to it after next item (b).  Lower-bounding
3163  *     the return value of this function with the current time plus
3164  *     bfqd->bfq_slice_idle tends to filter out greedy applications,
3165  *     because the latter issue their next request as soon as possible
3166  *     after the last one has been completed. In contrast, a soft
3167  *     real-time application spends some time processing data, after a
3168  *     batch of its requests has been completed.
3169  *
3170  * (b) Current value of bfqq->soft_rt_next_start. As pointed out
3171  *     above, greedy applications may happen to meet both the
3172  *     bandwidth and isochrony requirements under heavy CPU or
3173  *     storage-device load. In more detail, in these scenarios, these
3174  *     applications happen, only for limited time periods, to do I/O
3175  *     slowly enough to meet all the requirements described so far,
3176  *     including the filtering in above item (a). These slow-speed
3177  *     time intervals are usually interspersed between other time
3178  *     intervals during which these applications do I/O at a very high
3179  *     speed. Fortunately, exactly because of the high speed of the
3180  *     I/O in the high-speed intervals, the values returned by this
3181  *     function happen to be so high, near the end of any such
3182  *     high-speed interval, to be likely to fall *after* the end of
3183  *     the low-speed time interval that follows. These high values are
3184  *     stored in bfqq->soft_rt_next_start after each invocation of
3185  *     this function. As a consequence, if the last value of
3186  *     bfqq->soft_rt_next_start is constantly used to lower-bound the
3187  *     next value that this function may return, then, from the very
3188  *     beginning of a low-speed interval, bfqq->soft_rt_next_start is
3189  *     likely to be constantly kept so high that any I/O request
3190  *     issued during the low-speed interval is considered as arriving
3191  *     to soon for the application to be deemed as soft
3192  *     real-time. Then, in the high-speed interval that follows, the
3193  *     application will not be deemed as soft real-time, just because
3194  *     it will do I/O at a high speed. And so on.
3195  *
3196  * Getting back to the filtering in item (a), in the following two
3197  * cases this filtering might be easily passed by a greedy
3198  * application, if the reference quantity was just
3199  * bfqd->bfq_slice_idle:
3200  * 1) HZ is so low that the duration of a jiffy is comparable to or
3201  *    higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
3202  *    devices with HZ=100. The time granularity may be so coarse
3203  *    that the approximation, in jiffies, of bfqd->bfq_slice_idle
3204  *    is rather lower than the exact value.
3205  * 2) jiffies, instead of increasing at a constant rate, may stop increasing
3206  *    for a while, then suddenly 'jump' by several units to recover the lost
3207  *    increments. This seems to happen, e.g., inside virtual machines.
3208  * To address this issue, in the filtering in (a) we do not use as a
3209  * reference time interval just bfqd->bfq_slice_idle, but
3210  * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
3211  * minimum number of jiffies for which the filter seems to be quite
3212  * precise also in embedded systems and KVM/QEMU virtual machines.
3213  */
3214 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
3215                                                 struct bfq_queue *bfqq)
3216 {
3217         return max3(bfqq->soft_rt_next_start,
3218                     bfqq->last_idle_bklogged +
3219                     HZ * bfqq->service_from_backlogged /
3220                     bfqd->bfq_wr_max_softrt_rate,
3221                     jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
3222 }
3223 
3224 static bool bfq_bfqq_injectable(struct bfq_queue *bfqq)
3225 {
3226         return BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3227                 blk_queue_nonrot(bfqq->bfqd->queue) &&
3228                 bfqq->bfqd->hw_tag;
3229 }
3230 
3231 /**
3232  * bfq_bfqq_expire - expire a queue.
3233  * @bfqd: device owning the queue.
3234  * @bfqq: the queue to expire.
3235  * @compensate: if true, compensate for the time spent idling.
3236  * @reason: the reason causing the expiration.
3237  *
3238  * If the process associated with bfqq does slow I/O (e.g., because it
3239  * issues random requests), we charge bfqq with the time it has been
3240  * in service instead of the service it has received (see
3241  * bfq_bfqq_charge_time for details on how this goal is achieved). As
3242  * a consequence, bfqq will typically get higher timestamps upon
3243  * reactivation, and hence it will be rescheduled as if it had
3244  * received more service than what it has actually received. In the
3245  * end, bfqq receives less service in proportion to how slowly its
3246  * associated process consumes its budgets (and hence how seriously it
3247  * tends to lower the throughput). In addition, this time-charging
3248  * strategy guarantees time fairness among slow processes. In
3249  * contrast, if the process associated with bfqq is not slow, we
3250  * charge bfqq exactly with the service it has received.
3251  *
3252  * Charging time to the first type of queues and the exact service to
3253  * the other has the effect of using the WF2Q+ policy to schedule the
3254  * former on a timeslice basis, without violating service domain
3255  * guarantees among the latter.
3256  */
3257 void bfq_bfqq_expire(struct bfq_data *bfqd,
3258                      struct bfq_queue *bfqq,
3259                      bool compensate,
3260                      enum bfqq_expiration reason)
3261 {
3262         bool slow;
3263         unsigned long delta = 0;
3264         struct bfq_entity *entity = &bfqq->entity;
3265         int ref;
3266 
3267         /*
3268          * Check whether the process is slow (see bfq_bfqq_is_slow).
3269          */
3270         slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
3271 
3272         /*
3273          * As above explained, charge slow (typically seeky) and
3274          * timed-out queues with the time and not the service
3275          * received, to favor sequential workloads.
3276          *
3277          * Processes doing I/O in the slower disk zones will tend to
3278          * be slow(er) even if not seeky. Therefore, since the
3279          * estimated peak rate is actually an average over the disk
3280          * surface, these processes may timeout just for bad luck. To
3281          * avoid punishing them, do not charge time to processes that
3282          * succeeded in consuming at least 2/3 of their budget. This
3283          * allows BFQ to preserve enough elasticity to still perform
3284          * bandwidth, and not time, distribution with little unlucky
3285          * or quasi-sequential processes.
3286          */
3287         if (bfqq->wr_coeff == 1 &&
3288             (slow ||
3289              (reason == BFQQE_BUDGET_TIMEOUT &&
3290               bfq_bfqq_budget_left(bfqq) >=  entity->budget / 3)))
3291                 bfq_bfqq_charge_time(bfqd, bfqq, delta);
3292 
3293         if (reason == BFQQE_TOO_IDLE &&
3294             entity->service <= 2 * entity->budget / 10)
3295                 bfq_clear_bfqq_IO_bound(bfqq);
3296 
3297         if (bfqd->low_latency && bfqq->wr_coeff == 1)
3298                 bfqq->last_wr_start_finish = jiffies;
3299 
3300         if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
3301             RB_EMPTY_ROOT(&bfqq->sort_list)) {
3302                 /*
3303                  * If we get here, and there are no outstanding
3304                  * requests, then the request pattern is isochronous
3305                  * (see the comments on the function
3306                  * bfq_bfqq_softrt_next_start()). Thus we can compute
3307                  * soft_rt_next_start. And we do it, unless bfqq is in
3308                  * interactive weight raising. We do not do it in the
3309                  * latter subcase, for the following reason. bfqq may
3310                  * be conveying the I/O needed to load a soft
3311                  * real-time application. Such an application will
3312                  * actually exhibit a soft real-time I/O pattern after
3313                  * it finally starts doing its job. But, if
3314                  * soft_rt_next_start is computed here for an
3315                  * interactive bfqq, and bfqq had received a lot of
3316                  * service before remaining with no outstanding
3317                  * request (likely to happen on a fast device), then
3318                  * soft_rt_next_start would be assigned such a high
3319                  * value that, for a very long time, bfqq would be
3320                  * prevented from being possibly considered as soft
3321                  * real time.
3322                  *
3323                  * If, instead, the queue still has outstanding
3324                  * requests, then we have to wait for the completion
3325                  * of all the outstanding requests to discover whether
3326                  * the request pattern is actually isochronous.
3327                  */
3328                 if (bfqq->dispatched == 0 &&
3329                     bfqq->wr_coeff != bfqd->bfq_wr_coeff)
3330                         bfqq->soft_rt_next_start =
3331                                 bfq_bfqq_softrt_next_start(bfqd, bfqq);
3332                 else if (bfqq->dispatched > 0) {
3333                         /*
3334                          * Schedule an update of soft_rt_next_start to when
3335                          * the task may be discovered to be isochronous.
3336                          */
3337                         bfq_mark_bfqq_softrt_update(bfqq);
3338                 }
3339         }
3340 
3341         bfq_log_bfqq(bfqd, bfqq,
3342                 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
3343                 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
3344 
3345         /*
3346          * Increase, decrease or leave budget unchanged according to
3347          * reason.
3348          */
3349         __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
3350         ref = bfqq->ref;
3351         __bfq_bfqq_expire(bfqd, bfqq);
3352 
3353         if (ref == 1) /* bfqq is gone, no more actions on it */
3354                 return;
3355 
3356         bfqq->injected_service = 0;
3357 
3358         /* mark bfqq as waiting a request only if a bic still points to it */
3359         if (!bfq_bfqq_busy(bfqq) &&
3360             reason != BFQQE_BUDGET_TIMEOUT &&
3361             reason != BFQQE_BUDGET_EXHAUSTED) {
3362                 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
3363                 /*
3364                  * Not setting service to 0, because, if the next rq
3365                  * arrives in time, the queue will go on receiving
3366                  * service with this same budget (as if it never expired)
3367                  */
3368         } else
3369                 entity->service = 0;
3370 
3371         /*
3372          * Reset the received-service counter for every parent entity.
3373          * Differently from what happens with bfqq->entity.service,
3374          * the resetting of this counter never needs to be postponed
3375          * for parent entities. In fact, in case bfqq may have a
3376          * chance to go on being served using the last, partially
3377          * consumed budget, bfqq->entity.service needs to be kept,
3378          * because if bfqq then actually goes on being served using
3379          * the same budget, the last value of bfqq->entity.service is
3380          * needed to properly decrement bfqq->entity.budget by the
3381          * portion already consumed. In contrast, it is not necessary
3382          * to keep entity->service for parent entities too, because
3383          * the bubble up of the new value of bfqq->entity.budget will
3384          * make sure that the budgets of parent entities are correct,
3385          * even in case bfqq and thus parent entities go on receiving
3386          * service with the same budget.
3387          */
3388         entity = entity->parent;
3389         for_each_entity(entity)
3390                 entity->service = 0;
3391 }
3392 
3393 /*
3394  * Budget timeout is not implemented through a dedicated timer, but
3395  * just checked on request arrivals and completions, as well as on
3396  * idle timer expirations.
3397  */
3398 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
3399 {
3400         return time_is_before_eq_jiffies(bfqq->budget_timeout);
3401 }
3402 
3403 /*
3404  * If we expire a queue that is actively waiting (i.e., with the
3405  * device idled) for the arrival of a new request, then we may incur
3406  * the timestamp misalignment problem described in the body of the
3407  * function __bfq_activate_entity. Hence we return true only if this
3408  * condition does not hold, or if the queue is slow enough to deserve
3409  * only to be kicked off for preserving a high throughput.
3410  */
3411 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
3412 {
3413         bfq_log_bfqq(bfqq->bfqd, bfqq,
3414                 "may_budget_timeout: wait_request %d left %d timeout %d",
3415                 bfq_bfqq_wait_request(bfqq),
3416                         bfq_bfqq_budget_left(bfqq) >=  bfqq->entity.budget / 3,
3417                 bfq_bfqq_budget_timeout(bfqq));
3418 
3419         return (!bfq_bfqq_wait_request(bfqq) ||
3420                 bfq_bfqq_budget_left(bfqq) >=  bfqq->entity.budget / 3)
3421                 &&
3422                 bfq_bfqq_budget_timeout(bfqq);
3423 }
3424 
3425 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
3426                                              struct bfq_queue *bfqq)
3427 {
3428         bool rot_without_queueing =
3429                 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
3430                 bfqq_sequential_and_IO_bound,
3431                 idling_boosts_thr;
3432 
3433         bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
3434                 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
3435 
3436         /*
3437          * The next variable takes into account the cases where idling
3438          * boosts the throughput.
3439          *
3440          * The value of the variable is computed considering, first, that
3441          * idling is virtually always beneficial for the throughput if:
3442          * (a) the device is not NCQ-capable and rotational, or
3443          * (b) regardless of the presence of NCQ, the device is rotational and
3444          *     the request pattern for bfqq is I/O-bound and sequential, or
3445          * (c) regardless of whether it is rotational, the device is
3446          *     not NCQ-capable and the request pattern for bfqq is
3447          *     I/O-bound and sequential.
3448          *
3449          * Secondly, and in contrast to the above item (b), idling an
3450          * NCQ-capable flash-based device would not boost the
3451          * throughput even with sequential I/O; rather it would lower
3452          * the throughput in proportion to how fast the device
3453          * is. Accordingly, the next variable is true if any of the
3454          * above conditions (a), (b) or (c) is true, and, in
3455          * particular, happens to be false if bfqd is an NCQ-capable
3456          * flash-based device.
3457          */
3458         idling_boosts_thr = rot_without_queueing ||
3459                 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
3460                  bfqq_sequential_and_IO_bound);
3461 
3462         /*
3463          * The return value of this function is equal to that of
3464          * idling_boosts_thr, unless a special case holds. In this
3465          * special case, described below, idling may cause problems to
3466          * weight-raised queues.
3467          *
3468          * When the request pool is saturated (e.g., in the presence
3469          * of write hogs), if the processes associated with
3470          * non-weight-raised queues ask for requests at a lower rate,
3471          * then processes associated with weight-raised queues have a
3472          * higher probability to get a request from the pool
3473          * immediately (or at least soon) when they need one. Thus
3474          * they have a higher probability to actually get a fraction
3475          * of the device throughput proportional to their high
3476          * weight. This is especially true with NCQ-capable drives,
3477          * which enqueue several requests in advance, and further
3478          * reorder internally-queued requests.
3479          *
3480          * For this reason, we force to false the return value if
3481          * there are weight-raised busy queues. In this case, and if
3482          * bfqq is not weight-raised, this guarantees that the device
3483          * is not idled for bfqq (if, instead, bfqq is weight-raised,
3484          * then idling will be guaranteed by another variable, see
3485          * below). Combined with the timestamping rules of BFQ (see
3486          * [1] for details), this behavior causes bfqq, and hence any
3487          * sync non-weight-raised queue, to get a lower number of
3488          * requests served, and thus to ask for a lower number of
3489          * requests from the request pool, before the busy
3490          * weight-raised queues get served again. This often mitigates
3491          * starvation problems in the presence of heavy write
3492          * workloads and NCQ, thereby guaranteeing a higher
3493          * application and system responsiveness in these hostile
3494          * scenarios.
3495          */
3496         return idling_boosts_thr &&
3497                 bfqd->wr_busy_queues == 0;
3498 }
3499 
3500 /*
3501  * There is a case where idling must be performed not for
3502  * throughput concerns, but to preserve service guarantees.
3503  *
3504  * To introduce this case, we can note that allowing the drive
3505  * to enqueue more than one request at a time, and hence
3506  * delegating de facto final scheduling decisions to the
3507  * drive's internal scheduler, entails loss of control on the
3508  * actual request service order. In particular, the critical
3509  * situation is when requests from different processes happen
3510  * to be present, at the same time, in the internal queue(s)
3511  * of the drive. In such a situation, the drive, by deciding
3512  * the service order of the internally-queued requests, does
3513  * determine also the actual throughput distribution among
3514  * these processes. But the drive typically has no notion or
3515  * concern about per-process throughput distribution, and
3516  * makes its decisions only on a per-request basis. Therefore,
3517  * the service distribution enforced by the drive's internal
3518  * scheduler is likely to coincide with the desired
3519  * device-throughput distribution only in a completely
3520  * symmetric scenario where:
3521  * (i)  each of these processes must get the same throughput as
3522  *      the others;
3523  * (ii) the I/O of each process has the same properties, in
3524  *      terms of locality (sequential or random), direction
3525  *      (reads or writes), request sizes, greediness
3526  *      (from I/O-bound to sporadic), and so on.
3527  * In fact, in such a scenario, the drive tends to treat
3528  * the requests of each of these processes in about the same
3529  * way as the requests of the others, and thus to provide
3530  * each of these processes with about the same throughput
3531  * (which is exactly the desired throughput distribution). In
3532  * contrast, in any asymmetric scenario, device idling is
3533  * certainly needed to guarantee that bfqq receives its
3534  * assigned fraction of the device throughput (see [1] for
3535  * details).
3536  * The problem is that idling may significantly reduce
3537  * throughput with certain combinations of types of I/O and
3538  * devices. An important example is sync random I/O, on flash
3539  * storage with command queueing. So, unless bfqq falls in the
3540  * above cases where idling also boosts throughput, it would
3541  * be important to check conditions (i) and (ii) accurately,
3542  * so as to avoid idling when not strictly needed for service
3543  * guarantees.
3544  *
3545  * Unfortunately, it is extremely difficult to thoroughly
3546  * check condition (ii). And, in case there are active groups,
3547  * it becomes very difficult to check condition (i) too. In
3548  * fact, if there are active groups, then, for condition (i)
3549  * to become false, it is enough that an active group contains
3550  * more active processes or sub-groups than some other active
3551  * group. More precisely, for condition (i) to hold because of
3552  * such a group, it is not even necessary that the group is
3553  * (still) active: it is sufficient that, even if the group
3554  * has become inactive, some of its descendant processes still
3555  * have some request already dispatched but still waiting for
3556  * completion. In fact, requests have still to be guaranteed
3557  * their share of the throughput even after being
3558  * dispatched. In this respect, it is easy to show that, if a
3559  * group frequently becomes inactive while still having
3560  * in-flight requests, and if, when this happens, the group is
3561  * not considered in the calculation of whether the scenario
3562  * is asymmetric, then the group may fail to be guaranteed its
3563  * fair share of the throughput (basically because idling may
3564  * not be performed for the descendant processes of the group,
3565  * but it had to be).  We address this issue with the
3566  * following bi-modal behavior, implemented in the function
3567  * bfq_symmetric_scenario().
3568  *
3569  * If there are groups with requests waiting for completion
3570  * (as commented above, some of these groups may even be
3571  * already inactive), then the scenario is tagged as
3572  * asymmetric, conservatively, without checking any of the
3573  * conditions (i) and (ii). So the device is idled for bfqq.
3574  * This behavior matches also the fact that groups are created
3575  * exactly if controlling I/O is a primary concern (to
3576  * preserve bandwidth and latency guarantees).
3577  *
3578  * On the opposite end, if there are no groups with requests
3579  * waiting for completion, then only condition (i) is actually
3580  * controlled, i.e., provided that condition (i) holds, idling
3581  * is not performed, regardless of whether condition (ii)
3582  * holds. In other words, only if condition (i) does not hold,
3583  * then idling is allowed, and the device tends to be
3584  * prevented from queueing many requests, possibly of several
3585  * processes. Since there are no groups with requests waiting
3586  * for completion, then, to control condition (i) it is enough
3587  * to check just whether all the queues with requests waiting
3588  * for completion also have the same weight.
3589  *
3590  * Not checking condition (ii) evidently exposes bfqq to the
3591  * risk of getting less throughput than its fair share.
3592  * However, for queues with the same weight, a further
3593  * mechanism, preemption, mitigates or even eliminates this
3594  * problem. And it does so without consequences on overall
3595  * throughput. This mechanism and its benefits are explained
3596  * in the next three paragraphs.
3597  *
3598  * Even if a queue, say Q, is expired when it remains idle, Q
3599  * can still preempt the new in-service queue if the next
3600  * request of Q arrives soon (see the comments on
3601  * bfq_bfqq_update_budg_for_activation). If all queues and
3602  * groups have the same weight, this form of preemption,
3603  * combined with the hole-recovery heuristic described in the
3604  * comments on function bfq_bfqq_update_budg_for_activation,
3605  * are enough to preserve a correct bandwidth distribution in
3606  * the mid term, even without idling. In fact, even if not
3607  * idling allows the internal queues of the device to contain
3608  * many requests, and thus to reorder requests, we can rather
3609  * safely assume that the internal scheduler still preserves a
3610  * minimum of mid-term fairness.
3611  *
3612  * More precisely, this preemption-based, idleless approach
3613  * provides fairness in terms of IOPS, and not sectors per
3614  * second. This can be seen with a simple example. Suppose
3615  * that there are two queues with the same weight, but that
3616  * the first queue receives requests of 8 sectors, while the
3617  * second queue receives requests of 1024 sectors. In
3618  * addition, suppose that each of the two queues contains at
3619  * most one request at a time, which implies that each queue
3620  * always remains idle after it is served. Finally, after
3621  * remaining idle, each queue receives very quickly a new
3622  * request. It follows that the two queues are served
3623  * alternatively, preempting each other if needed. This
3624  * implies that, although both queues have the same weight,
3625  * the queue with large requests receives a service that is
3626  * 1024/8 times as high as the service received by the other
3627  * queue.
3628  *
3629  * The motivation for using preemption instead of idling (for
3630  * queues with the same weight) is that, by not idling,
3631  * service guarantees are preserved (completely or at least in
3632  * part) without minimally sacrificing throughput. And, if
3633  * there is no active group, then the primary expectation for
3634  * this device is probably a high throughput.
3635  *
3636  * We are now left only with explaining the additional
3637  * compound condition that is checked below for deciding
3638  * whether the scenario is asymmetric. To explain this
3639  * compound condition, we need to add that the function
3640  * bfq_symmetric_scenario checks the weights of only
3641  * non-weight-raised queues, for efficiency reasons (see
3642  * comments on bfq_weights_tree_add()). Then the fact that
3643  * bfqq is weight-raised is checked explicitly here. More
3644  * precisely, the compound condition below takes into account
3645  * also the fact that, even if bfqq is being weight-raised,
3646  * the scenario is still symmetric if all queues with requests
3647  * waiting for completion happen to be
3648  * weight-raised. Actually, we should be even more precise
3649  * here, and differentiate between interactive weight raising
3650  * and soft real-time weight raising.
3651  *
3652  * As a side note, it is worth considering that the above
3653  * device-idling countermeasures may however fail in the
3654  * following unlucky scenario: if idling is (correctly)
3655  * disabled in a time period during which all symmetry
3656  * sub-conditions hold, and hence the device is allowed to
3657  * enqueue many requests, but at some later point in time some
3658  * sub-condition stops to hold, then it may become impossible
3659  * to let requests be served in the desired order until all
3660  * the requests already queued in the device have been served.
3661  */
3662 static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3663                                                  struct bfq_queue *bfqq)
3664 {
3665         return (bfqq->wr_coeff > 1 &&
3666                 bfqd->wr_busy_queues <
3667                 bfq_tot_busy_queues(bfqd)) ||
3668                 !bfq_symmetric_scenario(bfqd);
3669 }
3670 
3671 /*
3672  * For a queue that becomes empty, device idling is allowed only if
3673  * this function returns true for that queue. As a consequence, since
3674  * device idling plays a critical role for both throughput boosting
3675  * and service guarantees, the return value of this function plays a
3676  * critical role as well.
3677  *
3678  * In a nutshell, this function returns true only if idling is
3679  * beneficial for throughput or, even if detrimental for throughput,
3680  * idling is however necessary to preserve service guarantees (low
3681  * latency, desired throughput distribution, ...). In particular, on
3682  * NCQ-capable devices, this function tries to return false, so as to
3683  * help keep the drives' internal queues full, whenever this helps the
3684  * device boost the throughput without causing any service-guarantee
3685  * issue.
3686  *
3687  * Most of the issues taken into account to get the return value of
3688  * this function are not trivial. We discuss these issues in the two
3689  * functions providing the main pieces of information needed by this
3690  * function.
3691  */
3692 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
3693 {
3694         struct bfq_data *bfqd = bfqq->bfqd;
3695         bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
3696 
3697         if (unlikely(bfqd->strict_guarantees))
3698                 return true;
3699 
3700         /*
3701          * Idling is performed only if slice_idle > 0. In addition, we
3702          * do not idle if
3703          * (a) bfqq is async
3704          * (b) bfqq is in the idle io prio class: in this case we do
3705          * not idle because we want to minimize the bandwidth that
3706          * queues in this class can steal to higher-priority queues
3707          */
3708         if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
3709            bfq_class_idle(bfqq))
3710                 return false;
3711 
3712         idling_boosts_thr_with_no_issue =
3713                 idling_boosts_thr_without_issues(bfqd, bfqq);
3714 
3715         idling_needed_for_service_guar =
3716                 idling_needed_for_service_guarantees(bfqd, bfqq);
3717 
3718         /*
3719          * We have now the two components we need to compute the
3720          * return value of the function, which is true only if idling
3721          * either boosts the throughput (without issues), or is
3722          * necessary to preserve service guarantees.
3723          */
3724         return idling_boosts_thr_with_no_issue ||
3725                 idling_needed_for_service_guar;
3726 }
3727 
3728 /*
3729  * If the in-service queue is empty but the function bfq_better_to_idle
3730  * returns true, then:
3731  * 1) the queue must remain in service and cannot be expired, and
3732  * 2) the device must be idled to wait for the possible arrival of a new
3733  *    request for the queue.
3734  * See the comments on the function bfq_better_to_idle for the reasons
3735  * why performing device idling is the best choice to boost the throughput
3736  * and preserve service guarantees when bfq_better_to_idle itself
3737  * returns true.
3738  */
3739 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
3740 {
3741         return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
3742 }
3743 
3744 static struct bfq_queue *bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
3745 {
3746         struct bfq_queue *bfqq;
3747 
3748         /*
3749          * A linear search; but, with a high probability, very few
3750          * steps are needed to find a candidate queue, i.e., a queue
3751          * with enough budget left for its next request. In fact:
3752          * - BFQ dynamically updates the budget of every queue so as
3753          *   to accommodate the expected backlog of the queue;
3754          * - if a queue gets all its requests dispatched as injected
3755          *   service, then the queue is removed from the active list
3756          *   (and re-added only if it gets new requests, but with
3757          *   enough budget for its new backlog).
3758          */
3759         list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
3760                 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
3761                     bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
3762                     bfq_bfqq_budget_left(bfqq))
3763                         return bfqq;
3764 
3765         return NULL;
3766 }
3767 
3768 /*
3769  * Select a queue for service.  If we have a current queue in service,
3770  * check whether to continue servicing it, or retrieve and set a new one.
3771  */
3772 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
3773 {
3774         struct bfq_queue *bfqq;
3775         struct request *next_rq;
3776         enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
3777 
3778         bfqq = bfqd->in_service_queue;
3779         if (!bfqq)
3780                 goto new_queue;
3781 
3782         bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
3783 
3784         /*
3785          * Do not expire bfqq for budget timeout if bfqq may be about
3786          * to enjoy device idling. The reason why, in this case, we
3787          * prevent bfqq from expiring is the same as in the comments
3788          * on the case where bfq_bfqq_must_idle() returns true, in
3789          * bfq_completed_request().
3790          */
3791         if (bfq_may_expire_for_budg_timeout(bfqq) &&
3792             !bfq_bfqq_must_idle(bfqq))
3793                 goto expire;
3794 
3795 check_queue:
3796         /*
3797          * This loop is rarely executed more than once. Even when it
3798          * happens, it is much more convenient to re-execute this loop
3799          * than to return NULL and trigger a new dispatch to get a
3800          * request served.
3801          */
3802         next_rq = bfqq->next_rq;
3803         /*
3804          * If bfqq has requests queued and it has enough budget left to
3805          * serve them, keep the queue, otherwise expire it.
3806          */
3807         if (next_rq) {
3808                 if (bfq_serv_to_charge(next_rq, bfqq) >
3809                         bfq_bfqq_budget_left(bfqq)) {
3810                         /*
3811                          * Expire the queue for budget exhaustion,
3812                          * which makes sure that the next budget is
3813                          * enough to serve the next request, even if
3814                          * it comes from the fifo expired path.
3815                          */
3816                         reason = BFQQE_BUDGET_EXHAUSTED;
3817                         goto expire;
3818                 } else {
3819                         /*
3820                          * The idle timer may be pending because we may
3821                          * not disable disk idling even when a new request
3822                          * arrives.
3823                          */
3824                         if (bfq_bfqq_wait_request(bfqq)) {
3825                                 /*
3826                                  * If we get here: 1) at least a new request
3827                                  * has arrived but we have not disabled the
3828                                  * timer because the request was too small,
3829                                  * 2) then the block layer has unplugged
3830                                  * the device, causing the dispatch to be
3831                                  * invoked.
3832                                  *
3833                                  * Since the device is unplugged, now the
3834                                  * requests are probably large enough to
3835                                  * provide a reasonable throughput.
3836                                  * So we disable idling.
3837                                  */
3838                                 bfq_clear_bfqq_wait_request(bfqq);
3839                                 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
3840                         }
3841                         goto keep_queue;
3842                 }
3843         }
3844 
3845         /*
3846          * No requests pending. However, if the in-service queue is idling
3847          * for a new request, or has requests waiting for a completion and
3848          * may idle after their completion, then keep it anyway.
3849          *
3850          * Yet, to boost throughput, inject service from other queues if
3851          * possible.
3852          */
3853         if (bfq_bfqq_wait_request(bfqq) ||
3854             (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
3855                 if (bfq_bfqq_injectable(bfqq) &&
3856                     bfqq->injected_service * bfqq->inject_coeff <
3857                     bfqq->entity.service * 10)
3858                         bfqq = bfq_choose_bfqq_for_injection(bfqd);
3859                 else
3860                         bfqq = NULL;
3861 
3862                 goto keep_queue;
3863         }
3864 
3865         reason = BFQQE_NO_MORE_REQUESTS;
3866 expire:
3867         bfq_bfqq_expire(bfqd, bfqq, false, reason);
3868 new_queue:
3869         bfqq = bfq_set_in_service_queue(bfqd);
3870         if (bfqq) {
3871                 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
3872                 goto check_queue;
3873         }
3874 keep_queue:
3875         if (bfqq)
3876                 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
3877         else
3878                 bfq_log(bfqd, "select_queue: no queue returned");
3879 
3880         return bfqq;
3881 }
3882 
3883 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3884 {
3885         struct bfq_entity *entity = &bfqq->entity;
3886 
3887         if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
3888                 bfq_log_bfqq(bfqd, bfqq,
3889                         "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
3890                         jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
3891                         jiffies_to_msecs(bfqq->wr_cur_max_time),
3892                         bfqq->wr_coeff,
3893                         bfqq->entity.weight, bfqq->entity.orig_weight);
3894 
3895                 if (entity->prio_changed)
3896                         bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
3897 
3898                 /*
3899                  * If the queue was activated in a burst, or too much
3900                  * time has elapsed from the beginning of this
3901                  * weight-raising period, then end weight raising.
3902                  */
3903                 if (bfq_bfqq_in_large_burst(bfqq))
3904                         bfq_bfqq_end_wr(bfqq);
3905                 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
3906                                                 bfqq->wr_cur_max_time)) {
3907                         if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
3908                         time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
3909                                                bfq_wr_duration(bfqd)))
3910                                 bfq_bfqq_end_wr(bfqq);
3911                         else {
3912                                 switch_back_to_interactive_wr(bfqq, bfqd);
3913                                 bfqq->entity.prio_changed = 1;
3914                         }
3915                 }
3916                 if (bfqq->wr_coeff > 1 &&
3917                     bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
3918                     bfqq->service_from_wr > max_service_from_wr) {
3919                         /* see comments on max_service_from_wr */
3920                         bfq_bfqq_end_wr(bfqq);
3921                 }
3922         }
3923         /*
3924          * To improve latency (for this or other queues), immediately
3925          * update weight both if it must be raised and if it must be
3926          * lowered. Since, entity may be on some active tree here, and
3927          * might have a pending change of its ioprio class, invoke
3928          * next function with the last parameter unset (see the
3929          * comments on the function).
3930          */
3931         if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
3932                 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
3933                                                 entity, false);
3934 }
3935 
3936 /*
3937  * Dispatch next request from bfqq.
3938  */
3939 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
3940                                                  struct bfq_queue *bfqq)
3941 {
3942         struct request *rq = bfqq->next_rq;
3943         unsigned long service_to_charge;
3944 
3945         service_to_charge = bfq_serv_to_charge(rq, bfqq);
3946 
3947         bfq_bfqq_served(bfqq, service_to_charge);
3948 
3949         bfq_dispatch_remove(bfqd->queue, rq);
3950 
3951         if (bfqq != bfqd->in_service_queue) {
3952                 if (likely(bfqd->in_service_queue))
3953                         bfqd->in_service_queue->injected_service +=
3954                                 bfq_serv_to_charge(rq, bfqq);
3955 
3956                 goto return_rq;
3957         }
3958 
3959         /*
3960          * If weight raising has to terminate for bfqq, then next
3961          * function causes an immediate update of bfqq's weight,
3962          * without waiting for next activation. As a consequence, on
3963          * expiration, bfqq will be timestamped as if has never been
3964          * weight-raised during this service slot, even if it has
3965          * received part or even most of the service as a
3966          * weight-raised queue. This inflates bfqq's timestamps, which
3967          * is beneficial, as bfqq is then more willing to leave the
3968          * device immediately to possible other weight-raised queues.
3969          */
3970         bfq_update_wr_data(bfqd, bfqq);
3971 
3972         /*
3973          * Expire bfqq, pretending that its budget expired, if bfqq
3974          * belongs to CLASS_IDLE and other queues are waiting for
3975          * service.
3976          */
3977         if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
3978                 goto return_rq;
3979 
3980         bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
3981 
3982 return_rq:
3983         return rq;
3984 }
3985 
3986 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
3987 {
3988         struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3989 
3990         /*
3991          * Avoiding lock: a race on bfqd->busy_queues should cause at
3992          * most a call to dispatch for nothing
3993          */
3994         return !list_empty_careful(&bfqd->dispatch) ||
3995                 bfq_tot_busy_queues(bfqd) > 0;
3996 }
3997 
3998 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3999 {
4000         struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4001         struct request *rq = NULL;
4002         struct bfq_queue *bfqq = NULL;
4003 
4004         if (!list_empty(&bfqd->dispatch)) {
4005                 rq = list_first_entry(&bfqd->dispatch, struct request,
4006                                       queuelist);
4007                 list_del_init(&rq->queuelist);
4008 
4009                 bfqq = RQ_BFQQ(rq);
4010 
4011                 if (bfqq) {
4012                         /*
4013                          * Increment counters here, because this
4014                          * dispatch does not follow the standard
4015                          * dispatch flow (where counters are
4016                          * incremented)
4017                          */
4018                         bfqq->dispatched++;
4019 
4020                         goto inc_in_driver_start_rq;
4021                 }
4022 
4023                 /*
4024                  * We exploit the bfq_finish_requeue_request hook to
4025                  * decrement rq_in_driver, but
4026                  * bfq_finish_requeue_request will not be invoked on
4027                  * this request. So, to avoid unbalance, just start
4028                  * this request, without incrementing rq_in_driver. As
4029                  * a negative consequence, rq_in_driver is deceptively
4030                  * lower than it should be while this request is in
4031                  * service. This may cause bfq_schedule_dispatch to be
4032                  * invoked uselessly.
4033                  *
4034                  * As for implementing an exact solution, the
4035                  * bfq_finish_requeue_request hook, if defined, is
4036                  * probably invoked also on this request. So, by
4037                  * exploiting this hook, we could 1) increment
4038                  * rq_in_driver here, and 2) decrement it in
4039                  * bfq_finish_requeue_request. Such a solution would
4040                  * let the value of the counter be always accurate,
4041                  * but it would entail using an extra interface
4042                  * function. This cost seems higher than the benefit,
4043                  * being the frequency of non-elevator-private
4044                  * requests very low.
4045                  */
4046                 goto start_rq;
4047         }
4048 
4049         bfq_log(bfqd, "dispatch requests: %d busy queues",
4050                 bfq_tot_busy_queues(bfqd));
4051 
4052         if (bfq_tot_busy_queues(bfqd) == 0)
4053                 goto exit;
4054 
4055         /*
4056          * Force device to serve one request at a time if
4057          * strict_guarantees is true. Forcing this service scheme is
4058          * currently the ONLY way to guarantee that the request
4059          * service order enforced by the scheduler is respected by a
4060          * queueing device. Otherwise the device is free even to make
4061          * some unlucky request wait for as long as the device
4062          * wishes.
4063          *
4064          * Of course, serving one request at at time may cause loss of
4065          * throughput.
4066          */
4067         if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
4068                 goto exit;
4069 
4070         bfqq = bfq_select_queue(bfqd);
4071         if (!bfqq)
4072                 goto exit;
4073 
4074         rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
4075 
4076         if (rq) {
4077 inc_in_driver_start_rq:
4078                 bfqd->rq_in_driver++;
4079 start_rq:
4080                 rq->rq_flags |= RQF_STARTED;
4081         }
4082 exit:
4083         return rq;
4084 }
4085 
4086 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
4087 static void bfq_update_dispatch_stats(struct request_queue *q,
4088                                       struct request *rq,
4089                                       struct bfq_queue *in_serv_queue,
4090                                       bool idle_timer_disabled)
4091 {
4092         struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
4093 
4094         if (!idle_timer_disabled && !bfqq)
4095                 return;
4096 
4097         /*
4098          * rq and bfqq are guaranteed to exist until this function
4099          * ends, for the following reasons. First, rq can be
4100          * dispatched to the device, and then can be completed and
4101          * freed, only after this function ends. Second, rq cannot be
4102          * merged (and thus freed because of a merge) any longer,
4103          * because it has already started. Thus rq cannot be freed
4104          * before this function ends, and, since rq has a reference to
4105          * bfqq, the same guarantee holds for bfqq too.
4106          *
4107          * In addition, the following queue lock guarantees that
4108          * bfqq_group(bfqq) exists as well.
4109          */
4110         spin_lock_irq(&q->queue_lock);
4111         if (idle_timer_disabled)
4112                 /*
4113                  * Since the idle timer has been disabled,
4114                  * in_serv_queue contained some request when
4115                  * __bfq_dispatch_request was invoked above, which
4116                  * implies that rq was picked exactly from
4117                  * in_serv_queue. Thus in_serv_queue == bfqq, and is
4118                  * therefore guaranteed to exist because of the above
4119                  * arguments.
4120                  */
4121                 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
4122         if (bfqq) {
4123                 struct bfq_group *bfqg = bfqq_group(bfqq);
4124 
4125                 bfqg_stats_update_avg_queue_size(bfqg);
4126                 bfqg_stats_set_start_empty_time(bfqg);
4127                 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
4128         }
4129         spin_unlock_irq(&q->queue_lock);
4130 }
4131 #else
4132 static inline void bfq_update_dispatch_stats(struct request_queue *q,
4133                                              struct request *rq,
4134                                              struct bfq_queue *in_serv_queue,
4135                                              bool idle_timer_disabled) {}
4136 #endif
4137 
4138 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
4139 {
4140         struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
4141         struct request *rq;
4142         struct bfq_queue *in_serv_queue;
4143         bool waiting_rq, idle_timer_disabled;
4144 
4145         spin_lock_irq(&bfqd->lock);
4146 
4147         in_serv_queue = bfqd->in_service_queue;
4148         waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
4149 
4150         rq = __bfq_dispatch_request(hctx);
4151 
4152         idle_timer_disabled =
4153                 waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
4154 
4155         spin_unlock_irq(&bfqd->lock);
4156 
4157         bfq_update_dispatch_stats(hctx->queue, rq, in_serv_queue,
4158                                   idle_timer_disabled);
4159 
4160         return rq;
4161 }
4162 
4163 /*
4164  * Task holds one reference to the queue, dropped when task exits.  Each rq
4165  * in-flight on this queue also holds a reference, dropped when rq is freed.
4166  *
4167  * Scheduler lock must be held here. Recall not to use bfqq after calling
4168  * this function on it.
4169  */
4170 void bfq_put_queue(struct bfq_queue *bfqq)
4171 {
4172 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4173         struct bfq_group *bfqg = bfqq_group(bfqq);
4174 #endif
4175 
4176         if (bfqq->bfqd)
4177                 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
4178                              bfqq, bfqq->ref);
4179 
4180         bfqq->ref--;
4181         if (bfqq->ref)
4182                 return;
4183 
4184         if (!hlist_unhashed(&bfqq->burst_list_node)) {
4185                 hlist_del_init(&bfqq->burst_list_node);
4186                 /*
4187                  * Decrement also burst size after the removal, if the
4188                  * process associated with bfqq is exiting, and thus
4189                  * does not contribute to the burst any longer. This
4190                  * decrement helps filter out false positives of large
4191                  * bursts, when some short-lived process (often due to
4192                  * the execution of commands by some service) happens
4193                  * to start and exit while a complex application is
4194                  * starting, and thus spawning several processes that
4195                  * do I/O (and that *must not* be treated as a large
4196                  * burst, see comments on bfq_handle_burst).
4197                  *
4198                  * In particular, the decrement is performed only if:
4199                  * 1) bfqq is not a merged queue, because, if it is,
4200                  * then this free of bfqq is not triggered by the exit
4201                  * of the process bfqq is associated with, but exactly
4202                  * by the fact that bfqq has just been merged.
4203                  * 2) burst_size is greater than 0, to handle
4204                  * unbalanced decrements. Unbalanced decrements may
4205                  * happen in te following case: bfqq is inserted into
4206                  * the current burst list--without incrementing
4207                  * bust_size--because of a split, but the current
4208                  * burst list is not the burst list bfqq belonged to
4209                  * (see comments on the case of a split in
4210                  * bfq_set_request).
4211                  */
4212                 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
4213                         bfqq->bfqd->burst_size--;
4214         }
4215 
4216         kmem_cache_free(bfq_pool, bfqq);
4217 #ifdef CONFIG_BFQ_GROUP_IOSCHED
4218         bfqg_and_blkg_put(bfqg);
4219 #endif
4220 }
4221 
4222 static void bfq_put_cooperator(struct bfq_queue *bfqq)
4223 {
4224         struct bfq_queue *__bfqq, *next;
4225 
4226         /*
4227          * If this queue was scheduled to merge with another queue, be
4228          * sure to drop the reference taken on that queue (and others in
4229          * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
4230          */
4231         __bfqq = bfqq->new_bfqq;
4232         while (__bfqq) {
4233                 if (__bfqq == bfqq)
4234                         break;
4235                 next = __bfqq->new_bfqq;
4236                 bfq_put_queue(__bfqq);
4237                 __bfqq = next;
4238         }
4239 }
4240 
4241 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4242 {
4243         if (bfqq == bfqd->in_service_queue) {
4244                 __bfq_bfqq_expire(bfqd, bfqq);
4245                 bfq_schedule_dispatch(bfqd);
4246         }
4247 
4248         bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
4249 
4250         bfq_put_cooperator(bfqq);
4251 
4252         bfq_put_queue(bfqq); /* release process reference */
4253 }
4254 
4255 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
4256 {
4257         struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4258         struct bfq_data *bfqd;
4259 
4260         if (bfqq)
4261                 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
4262 
4263         if (bfqq && bfqd) {
4264                 unsigned long flags;
4265 
4266                 spin_lock_irqsave(&bfqd->lock, flags);
4267                 bfq_exit_bfqq(bfqd, bfqq);
4268                 bic_set_bfqq(bic, NULL, is_sync);
4269                 spin_unlock_irqrestore(&bfqd->lock, flags);
4270         }
4271 }
4272 
4273 static void bfq_exit_icq(struct io_cq *icq)
4274 {
4275         struct bfq_io_cq *bic = icq_to_bic(icq);
4276 
4277         bfq_exit_icq_bfqq(bic, true);
4278         bfq_exit_icq_bfqq(bic, false);
4279 }
4280 
4281 /*
4282  * Update the entity prio values; note that the new values will not
4283  * be used until the next (re)activation.
4284  */
4285 static void
4286 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
4287 {
4288         struct task_struct *tsk = current;
4289         int ioprio_class;
4290         struct bfq_data *bfqd = bfqq->bfqd;
4291 
4292         if (!bfqd)
4293                 return;
4294 
4295         ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4296         switch (ioprio_class) {
4297         default:
4298                 dev_err(bfqq->bfqd->queue->backing_dev_info->dev,
4299                         "bfq: bad prio class %d\n", ioprio_class);
4300                 /* fall through */
4301         case IOPRIO_CLASS_NONE:
4302                 /*
4303                  * No prio set, inherit CPU scheduling settings.
4304                  */
4305                 bfqq->new_ioprio = task_nice_ioprio(tsk);
4306                 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
4307                 break;
4308         case IOPRIO_CLASS_RT:
4309                 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4310                 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
4311                 break;
4312         case IOPRIO_CLASS_BE:
4313                 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4314                 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
4315                 break;
4316         case IOPRIO_CLASS_IDLE:
4317                 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
4318                 bfqq->new_ioprio = 7;
4319                 break;
4320         }
4321 
4322         if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
4323                 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
4324                         bfqq->new_ioprio);
4325                 bfqq->new_ioprio = IOPRIO_BE_NR;
4326         }
4327 
4328         bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
4329         bfqq->entity.prio_changed = 1;
4330 }
4331 
4332 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4333                                        struct bio *bio, bool is_sync,
4334                                        struct bfq_io_cq *bic);
4335 
4336 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
4337 {
4338         struct bfq_data *bfqd = bic_to_bfqd(bic);
4339         struct bfq_queue *bfqq;
4340         int ioprio = bic->icq.ioc->ioprio;
4341 
4342         /*
4343          * This condition may trigger on a newly created bic, be sure to
4344          * drop the lock before returning.
4345          */
4346         if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
4347                 return;
4348 
4349         bic->ioprio = ioprio;
4350 
4351         bfqq = bic_to_bfqq(bic, false);
4352         if (bfqq) {
4353                 /* release process reference on this queue */
4354                 bfq_put_queue(bfqq);
4355                 bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
4356                 bic_set_bfqq(bic, bfqq, false);
4357         }
4358 
4359         bfqq = bic_to_bfqq(bic, true);
4360         if (bfqq)
4361                 bfq_set_next_ioprio_data(bfqq, bic);
4362 }
4363 
4364 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4365                           struct bfq_io_cq *bic, pid_t pid, int is_sync)
4366 {
4367         RB_CLEAR_NODE(&bfqq->entity.rb_node);
4368         INIT_LIST_HEAD(&bfqq->fifo);
4369         INIT_HLIST_NODE(&bfqq->burst_list_node);
4370 
4371         bfqq->ref = 0;
4372         bfqq->bfqd = bfqd;
4373 
4374         if (bic)
4375                 bfq_set_next_ioprio_data(bfqq, bic);
4376 
4377         if (is_sync) {
4378                 /*
4379                  * No need to mark as has_short_ttime if in
4380                  * idle_class, because no device idling is performed
4381                  * for queues in idle class
4382                  */
4383                 if (!bfq_class_idle(bfqq))
4384                         /* tentatively mark as has_short_ttime */
4385                         bfq_mark_bfqq_has_short_ttime(bfqq);
4386                 bfq_mark_bfqq_sync(bfqq);
4387                 bfq_mark_bfqq_just_created(bfqq);
4388                 /*
4389                  * Aggressively inject a lot of service: up to 90%.
4390                  * This coefficient remains constant during bfqq life,
4391                  * but this behavior might be changed, after enough
4392                  * testing and tuning.
4393                  */
4394                 bfqq->inject_coeff = 1;
4395         } else
4396                 bfq_clear_bfqq_sync(bfqq);
4397 
4398         /* set end request to minus infinity from now */
4399         bfqq->ttime.last_end_request = ktime_get_ns() + 1;
4400 
4401         bfq_mark_bfqq_IO_bound(bfqq);
4402 
4403         bfqq->pid = pid;
4404 
4405         /* Tentative initial value to trade off between thr and lat */
4406         bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
4407         bfqq->budget_timeout = bfq_smallest_from_now();
4408 
4409         bfqq->wr_coeff = 1;
4410         bfqq->last_wr_start_finish = jiffies;
4411         bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
4412         bfqq->split_time = bfq_smallest_from_now();
4413 
4414         /*
4415          * To not forget the possibly high bandwidth consumed by a
4416          * process/queue in the recent past,
4417          * bfq_bfqq_softrt_next_start() returns a value at least equal
4418          * to the current value of bfqq->soft_rt_next_start (see
4419          * comments on bfq_bfqq_softrt_next_start).  Set
4420          * soft_rt_next_start to now, to mean that bfqq has consumed
4421          * no bandwidth so far.
4422          */
4423         bfqq->soft_rt_next_start = jiffies;
4424 
4425         /* first request is almost certainly seeky */
4426         bfqq->seek_history = 1;
4427 }
4428 
4429 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
4430                                                struct bfq_group *bfqg,
4431                                                int ioprio_class, int ioprio)
4432 {
4433         switch (ioprio_class) {
4434         case IOPRIO_CLASS_RT:
4435                 return &bfqg->async_bfqq[0][ioprio];
4436         case IOPRIO_CLASS_NONE:
4437                 ioprio = IOPRIO_NORM;
4438                 /* fall through */
4439         case IOPRIO_CLASS_BE:
4440                 return &bfqg->async_bfqq[1][ioprio];
4441         case IOPRIO_CLASS_IDLE:
4442                 return &bfqg->async_idle_bfqq;
4443         default:
4444                 return NULL;
4445         }
4446 }
4447 
4448 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4449                                        struct bio *bio, bool is_sync,
4450                                        struct bfq_io_cq *bic)
4451 {
4452         const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4453         const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4454         struct bfq_queue **async_bfqq = NULL;
4455         struct bfq_queue *bfqq;
4456         struct bfq_group *bfqg;
4457 
4458         rcu_read_lock();
4459 
4460         bfqg = bfq_find_set_group(bfqd, __bio_blkcg(bio));
4461         if (!bfqg) {
4462                 bfqq = &bfqd->oom_bfqq;
4463                 goto out;
4464         }
4465 
4466         if (!is_sync) {
4467                 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
4468                                                   ioprio);
4469                 bfqq = *async_bfqq;
4470                 if (bfqq)
4471                         goto out;
4472         }
4473 
4474         bfqq = kmem_cache_alloc_node(bfq_pool,
4475                                      GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
4476                                      bfqd->queue->node);
4477 
4478         if (bfqq) {
4479                 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
4480                               is_sync);
4481                 bfq_init_entity(&bfqq->entity, bfqg);
4482                 bfq_log_bfqq(bfqd, bfqq, "allocated");
4483         } else {
4484                 bfqq = &bfqd->oom_bfqq;
4485                 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
4486                 goto out;
4487         }
4488 
4489         /*
4490          * Pin the queue now that it's allocated, scheduler exit will
4491          * prune it.
4492          */
4493         if (async_bfqq) {
4494                 bfqq->ref++; /*
4495                               * Extra group reference, w.r.t. sync
4496                               * queue. This extra reference is removed
4497                               * only if bfqq->bfqg disappears, to
4498                               * guarantee that this queue is not freed
4499                               * until its group goes away.
4500                               */
4501                 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
4502                              bfqq, bfqq->ref);
4503                 *async_bfqq = bfqq;
4504         }
4505 
4506 out:
4507         bfqq->ref++; /* get a process reference to this queue */
4508         bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
4509         rcu_read_unlock();
4510         return bfqq;
4511 }
4512 
4513 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
4514                                     struct bfq_queue *bfqq)
4515 {
4516         struct bfq_ttime *ttime = &bfqq->ttime;
4517         u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
4518 
4519         elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
4520 
4521         ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
4522         ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed,  8);
4523         ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
4524                                      ttime->ttime_samples);
4525 }
4526 
4527 static void
4528 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4529                        struct request *rq)
4530 {
4531         bfqq->seek_history <<= 1;
4532         bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
4533 }
4534 
4535 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
4536                                        struct bfq_queue *bfqq,
4537                                        struct bfq_io_cq *bic)
4538 {
4539         bool has_short_ttime = true;
4540 
4541         /*
4542          * No need to update has_short_ttime if bfqq is async or in
4543          * idle io prio class, or if bfq_slice_idle is zero, because
4544          * no device idling is performed for bfqq in this case.
4545          */
4546         if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
4547             bfqd->bfq_slice_idle == 0)
4548                 return;
4549 
4550         /* Idle window just restored, statistics are meaningless. */
4551         if (time_is_after_eq_jiffies(bfqq->split_time +
4552                                      bfqd->bfq_wr_min_idle_time))
4553                 return;
4554 
4555         /* Think time is infinite if no process is linked to
4556          * bfqq. Otherwise check average think time to
4557          * decide whether to mark as has_short_ttime
4558          */
4559         if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
4560             (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
4561              bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle))
4562                 has_short_ttime = false;
4563 
4564         bfq_log_bfqq(bfqd, bfqq, "update_has_short_ttime: has_short_ttime %d",
4565                      has_short_ttime);
4566 
4567         if (has_short_ttime)
4568                 bfq_mark_bfqq_has_short_ttime(bfqq);
4569         else
4570                 bfq_clear_bfqq_has_short_ttime(bfqq);
4571 }
4572 
4573 /*
4574  * Called when a new fs request (rq) is added to bfqq.  Check if there's
4575  * something we should do about it.
4576  */
4577 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4578                             struct request *rq)
4579 {
4580         struct bfq_io_cq *bic = RQ_BIC(rq);
4581 
4582         if (rq->cmd_flags & REQ_META)
4583                 bfqq->meta_pending++;
4584 
4585         bfq_update_io_thinktime(bfqd, bfqq);
4586         bfq_update_has_short_ttime(bfqd, bfqq, bic);
4587         bfq_update_io_seektime(bfqd, bfqq, rq);
4588 
4589         bfq_log_bfqq(bfqd, bfqq,
4590                      "rq_enqueued: has_short_ttime=%d (seeky %d)",
4591                      bfq_bfqq_has_short_ttime(bfqq), BFQQ_SEEKY(bfqq));
4592 
4593         bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4594 
4595         if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
4596                 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
4597                                  blk_rq_sectors(rq) < 32;
4598                 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
4599 
4600                 /*
4601                  * There is just this request queued: if
4602                  * - the request is small, and
4603                  * - we are idling to boost throughput, and
4604                  * - the queue is not to be expired,
4605                  * then just exit.
4606                  *
4607                  * In this way, if the device is being idled to wait
4608                  * for a new request from the in-service queue, we
4609                  * avoid unplugging the device and committing the
4610                  * device to serve just a small request. In contrast
4611                  * we wait for the block layer to decide when to
4612                  * unplug the device: hopefully, new requests will be
4613                  * merged to this one quickly, then the device will be
4614                  * unplugged and larger requests will be dispatched.
4615                  */
4616                 if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
4617                     !budget_timeout)
4618                         return;
4619 
4620                 /*
4621                  * A large enough request arrived, or idling is being
4622                  * performed to preserve service guarantees, or
4623                  * finally the queue is to be expired: in all these
4624                  * cases disk idling is to be stopped, so clear
4625                  * wait_request flag and reset timer.
4626                  */
4627                 bfq_clear_bfqq_wait_request(bfqq);
4628                 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4629 
4630                 /*
4631                  * The queue is not empty, because a new request just
4632                  * arrived. Hence we can safely expire the queue, in
4633                  * case of budget timeout, without risking that the
4634                  * timestamps of the queue are not updated correctly.
4635                  * See [1] for more details.
4636                  */
4637                 if (budget_timeout)
4638                         bfq_bfqq_expire(bfqd, bfqq, false,
4639                                         BFQQE_BUDGET_TIMEOUT);
4640         }
4641 }
4642 
4643 /* returns true if it causes the idle timer to be disabled */
4644 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
4645 {
4646         struct bfq_queue *bfqq = RQ_BFQQ(rq),
4647                 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
4648         bool waiting, idle_timer_disabled = false;
4649 
4650         if (new_bfqq) {
4651                 /*
4652                  * Release the request's reference to the old bfqq
4653                  * and make sure one is taken to the shared queue.
4654                  */
4655                 new_bfqq->allocated++;
4656                 bfqq->allocated--;
4657                 new_bfqq->ref++;
4658                 /*
4659                  * If the bic associated with the process
4660                  * issuing this request still points to bfqq
4661                  * (and thus has not been already redirected
4662                  * to new_bfqq or even some other bfq_queue),
4663                  * then complete the merge and redirect it to
4664                  * new_bfqq.
4665                  */
4666                 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
4667                         bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
4668                                         bfqq, new_bfqq);
4669 
4670                 bfq_clear_bfqq_just_created(bfqq);
4671                 /*
4672                  * rq is about to be enqueued into new_bfqq,
4673                  * release rq reference on bfqq
4674                  */
4675                 bfq_put_queue(bfqq);
4676                 rq->elv.priv[1] = new_bfqq;
4677                 bfqq = new_bfqq;
4678         }
4679 
4680         waiting = bfqq && bfq_bfqq_wait_request(bfqq);
4681         bfq_add_request(rq);
4682         idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
4683 
4684         rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
4685         list_add_tail(&rq->queuelist, &bfqq->fifo);
4686 
4687         bfq_rq_enqueued(bfqd, bfqq, rq);
4688 
4689         return idle_timer_disabled;
4690 }
4691 
4692 #if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
4693 static void bfq_update_insert_stats(struct request_queue *q,
4694                                     struct bfq_queue *bfqq,
4695                                     bool idle_timer_disabled,
4696                                     unsigned int cmd_flags)
4697 {
4698         if (!bfqq)
4699                 return;
4700 
4701         /*
4702          * bfqq still exists, because it can disappear only after
4703          * either it is merged with another queue, or the process it
4704          * is associated with exits. But both actions must be taken by
4705          * the same process currently executing this flow of
4706          * instructions.
4707          *
4708          * In addition, the following queue lock guarantees that
4709          * bfqq_group(bfqq) exists as well.
4710          */
4711         spin_lock_irq(&q->queue_lock);
4712         bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
4713         if (idle_timer_disabled)
4714                 bfqg_stats_update_idle_time(bfqq_group(bfqq));
4715         spin_unlock_irq(&q->queue_lock);
4716 }
4717 #else
4718 static inline void bfq_update_insert_stats(struct request_queue *q,
4719                                            struct bfq_queue *bfqq,
4720                                            bool idle_timer_disabled,
4721                                            unsigned int cmd_flags) {}
4722 #endif
4723 
4724 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
4725                                bool at_head)
4726 {
4727         struct request_queue *q = hctx->queue;
4728         struct bfq_data *bfqd = q->elevator->elevator_data;
4729         struct bfq_queue *bfqq;
4730         bool idle_timer_disabled = false;
4731         unsigned int cmd_flags;
4732 
4733         spin_lock_irq(&bfqd->lock);
4734         if (blk_mq_sched_try_insert_merge(q, rq)) {
4735                 spin_unlock_irq(&bfqd->lock);
4736                 return;
4737         }
4738 
4739         spin_unlock_irq(&bfqd->lock);
4740 
4741         blk_mq_sched_request_inserted(rq);
4742 
4743         spin_lock_irq(&bfqd->lock);
4744         bfqq = bfq_init_rq(rq);
4745         if (at_head || blk_rq_is_passthrough(rq)) {
4746                 if (at_head)
4747                         list_add(&rq->queuelist, &bfqd->dispatch);
4748                 else
4749                         list_add_tail(&rq->queuelist, &bfqd->dispatch);
4750         } else { /* bfqq is assumed to be non null here */
4751                 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
4752                 /*
4753                  * Update bfqq, because, if a queue merge has occurred
4754                  * in __bfq_insert_request, then rq has been
4755                  * redirected into a new queue.
4756                  */
4757                 bfqq = RQ_BFQQ(rq);
4758 
4759                 if (rq_mergeable(rq)) {
4760                         elv_rqhash_add(q, rq);
4761                         if (!q->last_merge)
4762                                 q->last_merge = rq;
4763                 }
4764         }
4765 
4766         /*
4767          * Cache cmd_flags before releasing scheduler lock, because rq
4768          * may disappear afterwards (for example, because of a request
4769          * merge).
4770          */
4771         cmd_flags = rq->cmd_flags;
4772 
4773         spin_unlock_irq(&bfqd->lock);
4774 
4775         bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
4776                                 cmd_flags);
4777 }
4778 
4779 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
4780                                 struct list_head *list, bool at_head)
4781 {
4782         while (!list_empty(list)) {
4783                 struct request *rq;
4784 
4785                 rq = list_first_entry(list, struct request, queuelist);
4786                 list_del_init(&rq->queuelist);
4787                 bfq_insert_request(hctx, rq, at_head);
4788         }
4789 }
4790 
4791 static void bfq_update_hw_tag(struct bfq_data *bfqd)
4792 {
4793         struct bfq_queue *bfqq = bfqd->in_service_queue;
4794 
4795         bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
4796                                        bfqd->rq_in_driver);
4797 
4798         if (bfqd->hw_tag == 1)
4799                 return;
4800 
4801         /*
4802          * This sample is valid if the number of outstanding requests
4803          * is large enough to allow a queueing behavior.  Note that the
4804          * sum is not exact, as it's not taking into account deactivated
4805          * requests.
4806          */
4807         if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
4808                 return;
4809 
4810         /*
4811          * If active queue hasn't enough requests and can idle, bfq might not
4812          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4813          * case
4814          */
4815         if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
4816             bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
4817             BFQ_HW_QUEUE_THRESHOLD &&
4818             bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
4819                 return;
4820 
4821         if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
4822                 return;
4823 
4824         bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
4825         bfqd->max_rq_in_driver = 0;
4826         bfqd->hw_tag_samples = 0;
4827 }
4828 
4829 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
4830 {
4831         u64 now_ns;
4832         u32 delta_us;
4833 
4834         bfq_update_hw_tag(bfqd);
4835 
4836         bfqd->rq_in_driver--;
4837         bfqq->dispatched--;
4838 
4839         if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
4840                 /*
4841                  * Set budget_timeout (which we overload to store the
4842                  * time at which the queue remains with no backlog and
4843                  * no outstanding request; used by the weight-raising
4844                  * mechanism).
4845                  */
4846                 bfqq->budget_timeout = jiffies;
4847 
4848                 bfq_weights_tree_remove(bfqd, bfqq);
4849         }
4850 
4851         now_ns = ktime_get_ns();
4852 
4853         bfqq->ttime.last_end_request = now_ns;
4854 
4855         /*
4856          * Using us instead of ns, to get a reasonable precision in
4857          * computing rate in next check.
4858          */
4859         delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
4860 
4861         /*
4862          * If the request took rather long to complete, and, according
4863          * to the maximum request size recorded, this completion latency
4864          * implies that the request was certainly served at a very low
4865          * rate (less than 1M sectors/sec), then the whole observation
4866          * interval that lasts up to this time instant cannot be a
4867          * valid time interval for computing a new peak rate.  Invoke
4868          * bfq_update_rate_reset to have the following three steps
4869          * taken:
4870          * - close the observation interval at the last (previous)
4871          *   request dispatch or completion
4872          * - compute rate, if possible, for that observation interval
4873          * - reset to zero samples, which will trigger a proper
4874          *   re-initialization of the observation interval on next
4875          *   dispatch
4876          */
4877         if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
4878            (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
4879                         1UL<<(BFQ_RATE_SHIFT - 10))
4880                 bfq_update_rate_reset(bfqd, NULL);
4881         bfqd->last_completion = now_ns;
4882 
4883         /*
4884          * If we are waiting to discover whether the request pattern
4885          * of the task associated with the queue is actually
4886          * isochronous, and both requisites for this condition to hold
4887          * are now satisfied, then compute soft_rt_next_start (see the
4888          * comments on the function bfq_bfqq_softrt_next_start()). We
4889          * do not compute soft_rt_next_start if bfqq is in interactive
4890          * weight raising (see the comments in bfq_bfqq_expire() for
4891          * an explanation). We schedule this delayed update when bfqq
4892          * expires, if it still has in-flight requests.
4893          */
4894         if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
4895             RB_EMPTY_ROOT(&bfqq->sort_list) &&
4896             bfqq->wr_coeff != bfqd->bfq_wr_coeff)
4897                 bfqq->soft_rt_next_start =
4898                         bfq_bfqq_softrt_next_start(bfqd, bfqq);
4899 
4900         /*
4901          * If this is the in-service queue, check if it needs to be expired,
4902          * or if we want to idle in case it has no pending requests.
4903          */
4904         if (bfqd->in_service_queue == bfqq) {
4905                 if (bfq_bfqq_must_idle(bfqq)) {
4906                         if (bfqq->dispatched == 0)
4907                                 bfq_arm_slice_timer(bfqd);
4908                         /*
4909                          * If we get here, we do not expire bfqq, even
4910                          * if bfqq was in budget timeout or had no
4911                          * more requests (as controlled in the next
4912                          * conditional instructions). The reason for
4913                          * not expiring bfqq is as follows.
4914                          *
4915                          * Here bfqq->dispatched > 0 holds, but
4916                          * bfq_bfqq_must_idle() returned true. This
4917                          * implies that, even if no request arrives
4918                          * for bfqq before bfqq->dispatched reaches 0,
4919                          * bfqq will, however, not be expired on the
4920                          * completion event that causes bfqq->dispatch
4921                          * to reach zero. In contrast, on this event,
4922                          * bfqq will start enjoying device idling
4923                          * (I/O-dispatch plugging).
4924                          *
4925                          * But, if we expired bfqq here, bfqq would
4926                          * not have the chance to enjoy device idling
4927                          * when bfqq->dispatched finally reaches
4928                          * zero. This would expose bfqq to violation
4929                          * of its reserved service guarantees.
4930                          */
4931                         return;
4932                 } else if (bfq_may_expire_for_budg_timeout(bfqq))
4933                         bfq_bfqq_expire(bfqd, bfqq, false,
4934                                         BFQQE_BUDGET_TIMEOUT);
4935                 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
4936                          (bfqq->dispatched == 0 ||
4937                           !bfq_better_to_idle(bfqq)))
4938                         bfq_bfqq_expire(bfqd, bfqq, false,
4939                                         BFQQE_NO_MORE_REQUESTS);
4940         }
4941 
4942         if (!bfqd->rq_in_driver)
4943                 bfq_schedule_dispatch(bfqd);
4944 }
4945 
4946 static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq)
4947 {
4948         bfqq->allocated--;
4949 
4950         bfq_put_queue(bfqq);
4951 }
4952 
4953 /*
4954  * Handle either a requeue or a finish for rq. The things to do are
4955  * the same in both cases: all references to rq are to be dropped. In
4956  * particular, rq is considered completed from the point of view of
4957  * the scheduler.
4958  */
4959 static void bfq_finish_requeue_request(struct request *rq)
4960 {
4961         struct bfq_queue *bfqq = RQ_BFQQ(rq);
4962         struct bfq_data *bfqd;
4963 
4964         /*
4965          * Requeue and finish hooks are invoked in blk-mq without
4966          * checking whether the involved request is actually still
4967          * referenced in the scheduler. To handle this fact, the
4968          * following two checks make this function exit in case of
4969          * spurious invocations, for which there is nothing to do.
4970          *
4971          * First, check whether rq has nothing to do with an elevator.
4972          */
4973         if (unlikely(!(rq->rq_flags & RQF_ELVPRIV)))
4974                 return;
4975 
4976         /*
4977          * rq either is not associated with any icq, or is an already
4978          * requeued request that has not (yet) been re-inserted into
4979          * a bfq_queue.
4980          */
4981         if (!rq->elv.icq || !bfqq)
4982                 return;
4983 
4984         bfqd = bfqq->bfqd;
4985 
4986         if (rq->rq_flags & RQF_STARTED)
4987                 bfqg_stats_update_completion(bfqq_group(bfqq),
4988                                              rq->start_time_ns,
4989                                              rq->io_start_time_ns,
4990                                              rq->cmd_flags);
4991 
4992         if (likely(rq->rq_flags & RQF_STARTED)) {
4993                 unsigned long flags;
4994 
4995                 spin_lock_irqsave(&bfqd->lock, flags);
4996 
4997                 bfq_completed_request(bfqq, bfqd);
4998                 bfq_finish_requeue_request_body(bfqq);
4999 
5000                 spin_unlock_irqrestore(&bfqd->lock, flags);
5001         } else {
5002                 /*
5003                  * Request rq may be still/already in the scheduler,
5004                  * in which case we need to remove it (this should
5005                  * never happen in case of requeue). And we cannot
5006                  * defer such a check and removal, to avoid
5007                  * inconsistencies in the time interval from the end
5008                  * of this function to the start of the deferred work.
5009                  * This situation seems to occur only in process
5010                  * context, as a consequence of a merge. In the
5011                  * current version of the code, this implies that the
5012                  * lock is held.
5013                  */
5014 
5015                 if (!RB_EMPTY_NODE(&rq->rb_node)) {
5016                         bfq_remove_request(rq->q, rq);
5017                         bfqg_stats_update_io_remove(bfqq_group(bfqq),
5018                                                     rq->cmd_flags);
5019                 }
5020                 bfq_finish_requeue_request_body(bfqq);
5021         }
5022 
5023         /*
5024          * Reset private fields. In case of a requeue, this allows
5025          * this function to correctly do nothing if it is spuriously
5026          * invoked again on this same request (see the check at the
5027          * beginning of the function). Probably, a better general
5028          * design would be to prevent blk-mq from invoking the requeue
5029          * or finish hooks of an elevator, for a request that is not
5030          * referred by that elevator.
5031          *
5032          * Resetting the following fields would break the
5033          * request-insertion logic if rq is re-inserted into a bfq
5034          * internal queue, without a re-preparation. Here we assume
5035          * that re-insertions of requeued requests, without
5036          * re-preparation, can happen only for pass_through or at_head
5037          * requests (which are not re-inserted into bfq internal
5038          * queues).
5039          */
5040         rq->elv.priv[0] = NULL;
5041         rq->elv.priv[1] = NULL;
5042 }
5043 
5044 /*
5045  * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
5046  * was the last process referring to that bfqq.
5047  */
5048 static struct bfq_queue *
5049 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
5050 {
5051         bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
5052 
5053         if (bfqq_process_refs(bfqq) == 1) {
5054                 bfqq->pid = current->pid;
5055                 bfq_clear_bfqq_coop(bfqq);
5056                 bfq_clear_bfqq_split_coop(bfqq);
5057                 return bfqq;
5058         }
5059 
5060         bic_set_bfqq(bic, NULL, 1);
5061 
5062         bfq_put_cooperator(bfqq);
5063 
5064         bfq_put_queue(bfqq);
5065         return NULL;
5066 }
5067 
5068 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
5069                                                    struct bfq_io_cq *bic,
5070                                                    struct bio *bio,
5071                                                    bool split, bool is_sync,
5072                                                    bool *new_queue)
5073 {
5074         struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
5075 
5076         if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
5077                 return bfqq;
5078 
5079         if (new_queue)
5080                 *new_queue = true;
5081 
5082         if (bfqq)
5083                 bfq_put_queue(bfqq);
5084         bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
5085 
5086         bic_set_bfqq(bic, bfqq, is_sync);
5087         if (split && is_sync) {
5088                 if ((bic->was_in_burst_list && bfqd->large_burst) ||
5089                     bic->saved_in_large_burst)
5090                         bfq_mark_bfqq_in_large_burst(bfqq);
5091                 else {
5092                         bfq_clear_bfqq_in_large_burst(bfqq);
5093                         if (bic->was_in_burst_list)
5094                                 /*
5095                                  * If bfqq was in the current
5096                                  * burst list before being
5097                                  * merged, then we have to add
5098                                  * it back. And we do not need
5099                                  * to increase burst_size, as
5100                                  * we did not decrement
5101                                  * burst_size when we removed
5102                                  * bfqq from the burst list as
5103                                  * a consequence of a merge
5104                                  * (see comments in
5105                                  * bfq_put_queue). In this
5106                                  * respect, it would be rather
5107                                  * costly to know whether the
5108                                  * current burst list is still
5109                                  * the same burst list from
5110                                  * which bfqq was removed on
5111                                  * the merge. To avoid this
5112                                  * cost, if bfqq was in a
5113                                  * burst list, then we add
5114                                  * bfqq to the current burst
5115                                  * list without any further
5116                                  * check. This can cause
5117                                  * inappropriate insertions,
5118                                  * but rarely enough to not
5119                                  * harm the detection of large
5120                                  * bursts significantly.
5121                                  */
5122                                 hlist_add_head(&bfqq->burst_list_node,
5123                                                &bfqd->burst_list);
5124                 }
5125                 bfqq->split_time = jiffies;
5126         }
5127 
5128         return bfqq;
5129 }
5130 
5131 /*
5132  * Only reset private fields. The actual request preparation will be
5133  * performed by bfq_init_rq, when rq is either inserted or merged. See
5134  * comments on bfq_init_rq for the reason behind this delayed
5135  * preparation.
5136  */
5137 static void bfq_prepare_request(struct request *rq, struct bio *bio)
5138 {
5139         /*
5140          * Regardless of whether we have an icq attached, we have to
5141          * clear the scheduler pointers, as they might point to
5142          * previously allocated bic/bfqq structs.
5143          */
5144         rq->elv.priv[0] = rq->elv.priv[1] = NULL;
5145 }
5146 
5147 /*
5148  * If needed, init rq, allocate bfq data structures associated with
5149  * rq, and increment reference counters in the destination bfq_queue
5150  * for rq. Return the destination bfq_queue for rq, or NULL is rq is
5151  * not associated with any bfq_queue.
5152  *
5153  * This function is invoked by the functions that perform rq insertion
5154  * or merging. One may have expected the above preparation operations
5155  * to be performed in bfq_prepare_request, and not delayed to when rq
5156  * is inserted or merged. The rationale behind this delayed
5157  * preparation is that, after the prepare_request hook is invoked for
5158  * rq, rq may still be transformed into a request with no icq, i.e., a
5159  * request not associated with any queue. No bfq hook is invoked to
5160  * signal this tranformation. As a consequence, should these
5161  * preparation operations be performed when the prepare_request hook
5162  * is invoked, and should rq be transformed one moment later, bfq
5163  * would end up in an inconsistent state, because it would have
5164  * incremented some queue counters for an rq destined to
5165  * transformation, without any chance to correctly lower these
5166  * counters back. In contrast, no transformation can still happen for
5167  * rq after rq has been inserted or merged. So, it is safe to execute
5168  * these preparation operations when rq is finally inserted or merged.
5169  */
5170 static struct bfq_queue *bfq_init_rq(struct request *rq)
5171 {
5172         struct request_queue *q = rq->q;
5173         struct bio *bio = rq->bio;
5174         struct bfq_data *bfqd = q->elevator->elevator_data;
5175         struct bfq_io_cq *bic;
5176         const int is_sync = rq_is_sync(rq);
5177         struct bfq_queue *bfqq;
5178         bool new_queue = false;
5179         bool bfqq_already_existing = false, split = false;
5180 
5181         if (unlikely(!rq->elv.icq))
5182                 return NULL;
5183 
5184         /*
5185          * Assuming that elv.priv[1] is set only if everything is set
5186          * for this rq. This holds true, because this function is
5187          * invoked only for insertion or merging, and, after such
5188          * events, a request cannot be manipulated any longer before
5189          * being removed from bfq.
5190          */
5191         if (rq->elv.priv[1])
5192                 return rq->elv.priv[1];
5193 
5194         bic = icq_to_bic(rq->elv.icq);
5195 
5196         bfq_check_ioprio_change(bic, bio);
5197 
5198         bfq_bic_update_cgroup(bic, bio);
5199 
5200         bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
5201                                          &new_queue);
5202 
5203         if (likely(!new_queue)) {
5204                 /* If the queue was seeky for too long, break it apart. */
5205                 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
5206                         bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
5207 
5208                         /* Update bic before losing reference to bfqq */
5209                         if (bfq_bfqq_in_large_burst(bfqq))
5210                                 bic->saved_in_large_burst = true;
5211 
5212                         bfqq = bfq_split_bfqq(bic, bfqq);
5213                         split = true;
5214 
5215                         if (!bfqq)
5216                                 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
5217                                                                  true, is_sync,
5218                                                                  NULL);
5219                         else
5220                                 bfqq_already_existing = true;
5221                 }
5222         }
5223 
5224         bfqq->allocated++;
5225         bfqq->ref++;
5226         bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
5227                      rq, bfqq, bfqq->ref);
5228 
5229         rq->elv.priv[0] = bic;
5230         rq->elv.priv[1] = bfqq;
5231 
5232         /*
5233          * If a bfq_queue has only one process reference, it is owned
5234          * by only this bic: we can then set bfqq->bic = bic. in
5235          * addition, if the queue has also just been split, we have to
5236          * resume its state.
5237          */
5238         if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
5239                 bfqq->bic = bic;
5240                 if (split) {
5241                         /*
5242                          * The queue has just been split from a shared
5243                          * queue: restore the idle window and the
5244                          * possible weight raising period.
5245                          */
5246                         bfq_bfqq_resume_state(bfqq, bfqd, bic,
5247                                               bfqq_already_existing);
5248                 }
5249         }
5250 
5251         if (unlikely(bfq_bfqq_just_created(bfqq)))
5252                 bfq_handle_burst(bfqd, bfqq);
5253 
5254         return bfqq;
5255 }
5256 
5257 static void bfq_idle_slice_timer_body(struct bfq_queue *bfqq)
5258 {
5259         struct bfq_data *bfqd = bfqq->bfqd;
5260         enum bfqq_expiration reason;
5261         unsigned long flags;
5262 
5263         spin_lock_irqsave(&bfqd->lock, flags);
5264         bfq_clear_bfqq_wait_request(bfqq);
5265 
5266         if (bfqq != bfqd->in_service_queue) {
5267                 spin_unlock_irqrestore(&bfqd->lock, flags);
5268                 return;
5269         }
5270 
5271         if (bfq_bfqq_budget_timeout(bfqq))
5272                 /*
5273                  * Also here the queue can be safely expired
5274                  * for budget timeout without wasting
5275                  * guarantees
5276                  */
5277                 reason = BFQQE_BUDGET_TIMEOUT;
5278         else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
5279                 /*
5280                  * The queue may not be empty upon timer expiration,
5281                  * because we may not disable the timer when the
5282                  * first request of the in-service queue arrives
5283                  * during disk idling.
5284                  */
5285                 reason = BFQQE_TOO_IDLE;
5286         else
5287                 goto schedule_dispatch;
5288 
5289         bfq_bfqq_expire(bfqd, bfqq, true, reason);
5290 
5291 schedule_dispatch:
5292         spin_unlock_irqrestore(&bfqd->lock, flags);
5293         bfq_schedule_dispatch(bfqd);
5294 }
5295 
5296 /*
5297  * Handler of the expiration of the timer running if the in-service queue
5298  * is idling inside its time slice.
5299  */
5300 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
5301 {
5302         struct bfq_data *bfqd = container_of(timer, struct bfq_data,
5303                                              idle_slice_timer);
5304         struct bfq_queue *bfqq = bfqd->in_service_queue;
5305 
5306         /*
5307          * Theoretical race here: the in-service queue can be NULL or
5308          * different from the queue that was idling if a new request
5309          * arrives for the current queue and there is a full dispatch
5310          * cycle that changes the in-service queue.  This can hardly
5311          * happen, but in the worst case we just expire a queue too
5312          * early.
5313          */
5314         if (bfqq)
5315                 bfq_idle_slice_timer_body(bfqq);
5316 
5317         return HRTIMER_NORESTART;
5318 }
5319 
5320 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
5321                                  struct bfq_queue **bfqq_ptr)
5322 {
5323         struct bfq_queue *bfqq = *bfqq_ptr;
5324 
5325         bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
5326         if (bfqq) {
5327                 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
5328 
5329                 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
5330                              bfqq, bfqq->ref);
5331                 bfq_put_queue(bfqq);
5332                 *bfqq_ptr = NULL;
5333         }
5334 }
5335 
5336 /*
5337  * Release all the bfqg references to its async queues.  If we are
5338  * deallocating the group these queues may still contain requests, so
5339  * we reparent them to the root cgroup (i.e., the only one that will
5340  * exist for sure until all the requests on a device are gone).
5341  */
5342 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
5343 {
5344         int i, j;
5345 
5346         for (i = 0; i < 2; i++)
5347                 for (j = 0; j < IOPRIO_BE_NR; j++)
5348                         __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
5349 
5350         __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
5351 }
5352 
5353 /*
5354  * See the comments on bfq_limit_depth for the purpose of
5355  * the depths set in the function. Return minimum shallow depth we'll use.
5356  */
5357 static unsigned int bfq_update_depths(struct bfq_data *bfqd,
5358                                       struct sbitmap_queue *bt)
5359 {
5360         unsigned int i, j, min_shallow = UINT_MAX;
5361 
5362         /*
5363          * In-word depths if no bfq_queue is being weight-raised:
5364          * leaving 25% of tags only for sync reads.
5365          *
5366          * In next formulas, right-shift the value
5367          * (1U<<bt->sb.shift), instead of computing directly
5368          * (1U<<(bt->sb.shift - something)), to be robust against
5369          * any possible value of bt->sb.shift, without having to
5370          * limit 'something'.
5371          */
5372         /* no more than 50% of tags for async I/O */
5373         bfqd->word_depths[0][0] = max((1U << bt->sb.shift) >> 1, 1U);
5374         /*
5375          * no more than 75% of tags for sync writes (25% extra tags
5376          * w.r.t. async I/O, to prevent async I/O from starving sync
5377          * writes)
5378          */
5379         bfqd->word_depths[0][1] = max(((1U << bt->sb.shift) * 3) >> 2, 1U);
5380 
5381         /*
5382          * In-word depths in case some bfq_queue is being weight-
5383          * raised: leaving ~63% of tags for sync reads. This is the
5384          * highest percentage for which, in our tests, application
5385          * start-up times didn't suffer from any regression due to tag
5386          * shortage.
5387          */
5388         /* no more than ~18% of tags for async I/O */
5389         bfqd->word_depths[1][0] = max(((1U << bt->sb.shift) * 3) >> 4, 1U);
5390         /* no more than ~37% of tags for sync writes (~20% extra tags) */
5391         bfqd->word_depths[1][1] = max(((1U << bt->sb.shift) * 6) >> 4, 1U);
5392 
5393         for (i = 0; i < 2; i++)
5394                 for (j = 0; j < 2; j++)
5395                         min_shallow = min(min_shallow, bfqd->word_depths[i][j]);
5396 
5397         return min_shallow;
5398 }
5399 
5400 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
5401 {
5402         struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5403         struct blk_mq_tags *tags = hctx->sched_tags;
5404         unsigned int min_shallow;
5405 
5406         min_shallow = bfq_update_depths(bfqd, &tags->bitmap_tags);
5407         sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, min_shallow);
5408         return 0;
5409 }
5410 
5411 static void bfq_exit_queue(struct elevator_queue *e)
5412 {
5413         struct bfq_data *bfqd = e->elevator_data;
5414         struct bfq_queue *bfqq, *n;
5415 
5416         hrtimer_cancel(&bfqd->idle_slice_timer);
5417 
5418         spin_lock_irq(&bfqd->lock);
5419         list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
5420                 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
5421         spin_unlock_irq(&bfqd->lock);
5422 
5423         hrtimer_cancel(&bfqd->idle_slice_timer);
5424 
5425 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5426         /* release oom-queue reference to root group */
5427         bfqg_and_blkg_put(bfqd->root_group);
5428 
5429         blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
5430 #else
5431         spin_lock_irq(&bfqd->lock);
5432         bfq_put_async_queues(bfqd, bfqd->root_group);
5433         kfree(bfqd->root_group);
5434         spin_unlock_irq(&bfqd->lock);
5435 #endif
5436 
5437         kfree(bfqd);
5438 }
5439 
5440 static void bfq_init_root_group(struct bfq_group *root_group,
5441                                 struct bfq_data *bfqd)
5442 {
5443         int i;
5444 
5445 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5446         root_group->entity.parent = NULL;
5447         root_group->my_entity = NULL;
5448         root_group->bfqd = bfqd;
5449 #endif
5450         root_group->rq_pos_tree = RB_ROOT;
5451         for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
5452                 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
5453         root_group->sched_data.bfq_class_idle_last_service = jiffies;
5454 }
5455 
5456 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
5457 {
5458         struct bfq_data *bfqd;
5459         struct elevator_queue *eq;
5460 
5461         eq = elevator_alloc(q, e);
5462         if (!eq)
5463                 return -ENOMEM;
5464 
5465         bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
5466         if (!bfqd) {
5467                 kobject_put(&eq->kobj);
5468                 return -ENOMEM;
5469         }
5470         eq->elevator_data = bfqd;
5471 
5472         spin_lock_irq(&q->queue_lock);
5473         q->elevator = eq;
5474         spin_unlock_irq(&q->queue_lock);
5475 
5476         /*
5477          * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
5478          * Grab a permanent reference to it, so that the normal code flow
5479          * will not attempt to free it.
5480          */
5481         bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
5482         bfqd->oom_bfqq.ref++;
5483         bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
5484         bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
5485         bfqd->oom_bfqq.entity.new_weight =
5486                 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
5487 
5488         /* oom_bfqq does not participate to bursts */
5489         bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
5490 
5491         /*
5492          * Trigger weight initialization, according to ioprio, at the
5493          * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
5494          * class won't be changed any more.
5495          */
5496         bfqd->oom_bfqq.entity.prio_changed = 1;
5497 
5498         bfqd->queue = q;
5499 
5500         INIT_LIST_HEAD(&bfqd->dispatch);
5501 
5502         hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
5503                      HRTIMER_MODE_REL);
5504         bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
5505 
5506         bfqd->queue_weights_tree = RB_ROOT;
5507         bfqd->num_groups_with_pending_reqs = 0;
5508 
5509         INIT_LIST_HEAD(&bfqd->active_list);
5510         INIT_LIST_HEAD(&bfqd->idle_list);
5511         INIT_HLIST_HEAD(&bfqd->burst_list);
5512 
5513         bfqd->hw_tag = -1;
5514 
5515         bfqd->bfq_max_budget = bfq_default_max_budget;
5516 
5517         bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
5518         bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
5519         bfqd->bfq_back_max = bfq_back_max;
5520         bfqd->bfq_back_penalty = bfq_back_penalty;
5521         bfqd->bfq_slice_idle = bfq_slice_idle;
5522         bfqd->bfq_timeout = bfq_timeout;
5523 
5524         bfqd->bfq_requests_within_timer = 120;
5525 
5526         bfqd->bfq_large_burst_thresh = 8;
5527         bfqd->bfq_burst_interval = msecs_to_jiffies(180);
5528 
5529         bfqd->low_latency = true;
5530 
5531         /*
5532          * Trade-off between responsiveness and fairness.
5533          */
5534         bfqd->bfq_wr_coeff = 30;
5535         bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
5536         bfqd->bfq_wr_max_time = 0;
5537         bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
5538         bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
5539         bfqd->bfq_wr_max_softrt_rate = 7000; /*
5540                                               * Approximate rate required
5541                                               * to playback or record a
5542                                               * high-definition compressed
5543                                               * video.
5544                                               */
5545         bfqd->wr_busy_queues = 0;
5546 
5547         /*
5548          * Begin by assuming, optimistically, that the device peak
5549          * rate is equal to 2/3 of the highest reference rate.
5550          */
5551         bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
5552                 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
5553         bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
5554 
5555         spin_lock_init(&bfqd->lock);
5556 
5557         /*
5558          * The invocation of the next bfq_create_group_hierarchy
5559          * function is the head of a chain of function calls
5560          * (bfq_create_group_hierarchy->blkcg_activate_policy->
5561          * blk_mq_freeze_queue) that may lead to the invocation of the
5562          * has_work hook function. For this reason,
5563          * bfq_create_group_hierarchy is invoked only after all
5564          * scheduler data has been initialized, apart from the fields
5565          * that can be initialized only after invoking
5566          * bfq_create_group_hierarchy. This, in particular, enables
5567          * has_work to correctly return false. Of course, to avoid
5568          * other inconsistencies, the blk-mq stack must then refrain
5569          * from invoking further scheduler hooks before this init
5570          * function is finished.
5571          */
5572         bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
5573         if (!bfqd->root_group)
5574                 goto out_free;
5575         bfq_init_root_group(bfqd->root_group, bfqd);
5576         bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
5577 
5578         wbt_disable_default(q);
5579         return 0;
5580 
5581 out_free:
5582         kfree(bfqd);
5583         kobject_put(&eq->kobj);
5584         return -ENOMEM;
5585 }
5586 
5587 static void bfq_slab_kill(void)
5588 {
5589         kmem_cache_destroy(bfq_pool);
5590 }
5591 
5592 static int __init bfq_slab_setup(void)
5593 {
5594         bfq_pool = KMEM_CACHE(bfq_queue, 0);
5595         if (!bfq_pool)
5596                 return -ENOMEM;
5597         return 0;
5598 }
5599 
5600 static ssize_t bfq_var_show(unsigned int var, char *page)
5601 {
5602         return sprintf(page, "%u\n", var);
5603 }
5604 
5605 static int bfq_var_store(unsigned long *var, const char *page)
5606 {
5607         unsigned long new_val;
5608         int ret = kstrtoul(page, 10, &new_val);
5609 
5610         if (ret)
5611                 return ret;
5612         *var = new_val;
5613         return 0;
5614 }
5615 
5616 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
5617 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
5618 {                                                                       \
5619         struct bfq_data *bfqd = e->elevator_data;                       \
5620         u64 __data = __VAR;                                             \
5621         if (__CONV == 1)                                                \
5622                 __data = jiffies_to_msecs(__data);                      \
5623         else if (__CONV == 2)                                           \
5624                 __data = div_u64(__data, NSEC_PER_MSEC);                \
5625         return bfq_var_show(__data, (page));                            \
5626 }
5627 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
5628 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
5629 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
5630 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
5631 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
5632 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
5633 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
5634 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
5635 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
5636 #undef SHOW_FUNCTION
5637 
5638 #define USEC_SHOW_FUNCTION(__FUNC, __VAR)                               \
5639 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
5640 {                                                                       \
5641         struct bfq_data *bfqd = e->elevator_data;                       \
5642         u64 __data = __VAR;                                             \
5643         __data = div_u64(__data, NSEC_PER_USEC);                        \
5644         return bfq_var_show(__data, (page));                            \
5645 }
5646 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
5647 #undef USEC_SHOW_FUNCTION
5648 
5649 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
5650 static ssize_t                                                          \
5651 __FUNC(struct elevator_queue *e, const char *page, size_t count)        \
5652 {                                                                       \
5653         struct bfq_data *bfqd = e->elevator_data;                       \
5654         unsigned long __data, __min = (MIN), __max = (MAX);             \
5655         int ret;                                                        \
5656                                                                         \
5657         ret = bfq_var_store(&__data, (page));                           \
5658         if (ret)                                                        \
5659                 return ret;                                             \
5660         if (__data < __min)                                             \
5661                 __data = __min;                                         \
5662         else if (__data > __max)                                        \
5663                 __data = __max;                                         \
5664         if (__CONV == 1)                                                \
5665                 *(__PTR) = msecs_to_jiffies(__data);                    \
5666         else if (__CONV == 2)                                           \
5667                 *(__PTR) = (u64)__data * NSEC_PER_MSEC;                 \
5668         else                                                            \
5669                 *(__PTR) = __data;                                      \
5670         return count;                                                   \
5671 }
5672 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
5673                 INT_MAX, 2);
5674 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
5675                 INT_MAX, 2);
5676 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
5677 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
5678                 INT_MAX, 0);
5679 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
5680 #undef STORE_FUNCTION
5681 
5682 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX)                    \
5683 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
5684 {                                                                       \
5685         struct bfq_data *bfqd = e->elevator_data;                       \
5686         unsigned long __data, __min = (MIN), __max = (MAX);             \
5687         int ret;                                                        \
5688                                                                         \
5689         ret = bfq_var_store(&__data, (page));                           \
5690         if (ret)                                                        \
5691                 return ret;                                             \
5692         if (__data < __min)                                             \
5693                 __data = __min;                                         \
5694         else if (__data > __max)                                        \
5695                 __data = __max;                                         \
5696         *(__PTR) = (u64)__data * NSEC_PER_USEC;                         \
5697         return count;                                                   \
5698 }
5699 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
5700                     UINT_MAX);
5701 #undef USEC_STORE_FUNCTION
5702 
5703 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
5704                                     const char *page, size_t count)
5705 {
5706         struct bfq_data *bfqd = e->elevator_data;
5707         unsigned long __data;
5708         int ret;
5709 
5710         ret = bfq_var_store(&__data, (page));
5711         if (ret)
5712                 return ret;
5713 
5714         if (__data == 0)
5715                 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5716         else {
5717                 if (__data > INT_MAX)
5718                         __data = INT_MAX;
5719                 bfqd->bfq_max_budget = __data;
5720         }
5721 
5722         bfqd->bfq_user_max_budget = __data;
5723 
5724         return count;
5725 }
5726 
5727 /*
5728  * Leaving this name to preserve name compatibility with cfq
5729  * parameters, but this timeout is used for both sync and async.
5730  */
5731 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
5732                                       const char *page, size_t count)
5733 {
5734         struct bfq_data *bfqd = e->elevator_data;
5735         unsigned long __data;
5736         int ret;
5737 
5738         ret = bfq_var_store(&__data, (page));
5739         if (ret)
5740                 return ret;
5741 
5742         if (__data < 1)
5743                 __data = 1;
5744         else if (__data > INT_MAX)
5745                 __data = INT_MAX;
5746 
5747         bfqd->bfq_timeout = msecs_to_jiffies(__data);
5748         if (bfqd->bfq_user_max_budget == 0)
5749                 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5750 
5751         return count;
5752 }
5753 
5754 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
5755                                      const char *page, size_t count)
5756 {
5757         struct bfq_data *bfqd = e->elevator_data;
5758         unsigned long __data;
5759         int ret;
5760 
5761         ret = bfq_var_store(&__data, (page));
5762         if (ret)
5763                 return ret;
5764 
5765         if (__data > 1)
5766                 __data = 1;
5767         if (!bfqd->strict_guarantees && __data == 1
5768             && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
5769                 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
5770 
5771         bfqd->strict_guarantees = __data;
5772 
5773         return count;
5774 }
5775 
5776 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
5777                                      const char *page, size_t count)
5778 {
5779         struct bfq_data *bfqd = e->elevator_data;
5780         unsigned long __data;
5781         int ret;
5782 
5783         ret = bfq_var_store(&__data, (page));
5784         if (ret)
5785                 return ret;
5786 
5787         if (__data > 1)
5788                 __data = 1;
5789         if (__data == 0 && bfqd->low_latency != 0)
5790                 bfq_end_wr(bfqd);
5791         bfqd->low_latency = __data;
5792 
5793         return count;
5794 }
5795 
5796 #define BFQ_ATTR(name) \
5797         __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
5798 
5799 static struct elv_fs_entry bfq_attrs[] = {
5800         BFQ_ATTR(fifo_expire_sync),
5801         BFQ_ATTR(fifo_expire_async),
5802         BFQ_ATTR(back_seek_max),
5803         BFQ_ATTR(back_seek_penalty),
5804         BFQ_ATTR(slice_idle),
5805         BFQ_ATTR(slice_idle_us),
5806         BFQ_ATTR(max_budget),
5807         BFQ_ATTR(timeout_sync),
5808         BFQ_ATTR(strict_guarantees),
5809         BFQ_ATTR(low_latency),
5810         __ATTR_NULL
5811 };
5812 
5813 static struct elevator_type iosched_bfq_mq = {
5814         .ops = {
5815                 .limit_depth            = bfq_limit_depth,
5816                 .prepare_request        = bfq_prepare_request,
5817                 .requeue_request        = bfq_finish_requeue_request,
5818                 .finish_request         = bfq_finish_requeue_request,
5819                 .exit_icq               = bfq_exit_icq,
5820                 .insert_requests        = bfq_insert_requests,
5821                 .dispatch_request       = bfq_dispatch_request,
5822                 .next_request           = elv_rb_latter_request,
5823                 .former_request         = elv_rb_former_request,
5824                 .allow_merge            = bfq_allow_bio_merge,
5825                 .bio_merge              = bfq_bio_merge,
5826                 .request_merge          = bfq_request_merge,
5827                 .requests_merged        = bfq_requests_merged,
5828                 .request_merged         = bfq_request_merged,
5829                 .has_work               = bfq_has_work,
5830                 .init_hctx              = bfq_init_hctx,
5831                 .init_sched             = bfq_init_queue,
5832                 .exit_sched             = bfq_exit_queue,
5833         },
5834 
5835         .icq_size =             sizeof(struct bfq_io_cq),
5836         .icq_align =            __alignof__(struct bfq_io_cq),
5837         .elevator_attrs =       bfq_attrs,
5838         .elevator_name =        "bfq",
5839         .elevator_owner =       THIS_MODULE,
5840 };
5841 MODULE_ALIAS("bfq-iosched");
5842 
5843 static int __init bfq_init(void)
5844 {
5845         int ret;
5846 
5847 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5848         ret = blkcg_policy_register(&blkcg_policy_bfq);
5849         if (ret)
5850                 return ret;
5851 #endif
5852 
5853         ret = -ENOMEM;
5854         if (bfq_slab_setup())
5855                 goto err_pol_unreg;
5856 
5857         /*
5858          * Times to load large popular applications for the typical
5859          * systems installed on the reference devices (see the
5860          * comments before the definition of the next
5861          * array). Actually, we use slightly lower values, as the
5862          * estimated peak rate tends to be smaller than the actual
5863          * peak rate.  The reason for this last fact is that estimates
5864          * are computed over much shorter time intervals than the long
5865          * intervals typically used for benchmarking. Why? First, to
5866          * adapt more quickly to variations. Second, because an I/O
5867          * scheduler cannot rely on a peak-rate-evaluation workload to
5868          * be run for a long time.
5869          */
5870         ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
5871         ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
5872 
5873         ret = elv_register(&iosched_bfq_mq);
5874         if (ret)
5875                 goto slab_kill;
5876 
5877         return 0;
5878 
5879 slab_kill:
5880         bfq_slab_kill();
5881 err_pol_unreg:
5882 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5883         blkcg_policy_unregister(&blkcg_policy_bfq);
5884 #endif
5885         return ret;
5886 }
5887 
5888 static void __exit bfq_exit(void)
5889 {
5890         elv_unregister(&iosched_bfq_mq);
5891 #ifdef CONFIG_BFQ_GROUP_IOSCHED
5892         blkcg_policy_unregister(&blkcg_policy_bfq);
5893 #endif
5894         bfq_slab_kill();
5895 }
5896 
5897 module_init(bfq_init);
5898 module_exit(bfq_exit);
5899 
5900 MODULE_AUTHOR("Paolo Valente");
5901 MODULE_LICENSE("GPL");
5902 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");
5903 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp