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

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

Version: ~ [ linux-5.19-rc8 ] ~ [ linux-5.18.14 ] ~ [ linux-5.17.15 ] ~ [ linux-5.16.20 ] ~ [ linux-5.15.57 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.133 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.207 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.253 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.289 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.324 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.302 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.9 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 // SPDX-License-Identifier: GPL-2.0-or-later
  2 /*
  3  * Budget Fair Queueing (BFQ) I/O scheduler.
  4  *
  5  * Based on ideas and code from CFQ:
  6  * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
  7  *
  8  * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
  9  *                    Paolo Valente <paolo.valente@unimore.it>
 10  *
 11  * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
 12  *                    Arianna Avanzini <avanzini@google.com>
 13  *
 14  * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
 15  *
 16  * BFQ is a proportional-share I/O scheduler, with some extra
 17  * low-latency capabilities. BFQ also supports full hierarchical
 18  * scheduling through cgroups. Next paragraphs provide an introduction
 19  * on BFQ inner workings. Details on BFQ benefits, usage and
 20  * limitations can be found in Documentation/block/bfq-iosched.rst.
 21  *
 22  * BFQ is a proportional-share storage-I/O scheduling algorithm based
 23  * on the slice-by-slice service scheme of CFQ. But BFQ assigns
 24  * budgets, measured in number of sectors, to processes instead of
 25  * time slices. The device is not granted to the in-service process
 26  * for a given time slice, but until it has exhausted its assigned
 27  * budget. This change from the time to the service domain enables BFQ
 28  * to distribute the device throughput among processes as desired,
 29  * without any distortion due to throughput fluctuations, or to device
 30  * internal queueing. BFQ uses an ad hoc internal scheduler, called
 31  * B-WF2Q+, to schedule processes according to their budgets. More
 32  * precisely, BFQ schedules queues associated with processes. Each
 33  * process/queue is assigned a user-configurable weight, and B-WF2Q+
 34  * guarantees that each queue receives a fraction of the throughput
 35  * proportional to its weight. Thanks to the accurate policy of
 36  * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
 37  * processes issuing sequential requests (to boost the throughput),
 38  * and yet guarantee a low latency to interactive and soft real-time
 39  * applications.
 40  *
 41  * In particular, to provide these low-latency guarantees, BFQ
 42  * explicitly privileges the I/O of two classes of time-sensitive
 43  * applications: interactive and soft real-time. In more detail, BFQ
 44  * behaves this way if the low_latency parameter is set (default
 45  * configuration). This feature enables BFQ to provide applications in
 46  * these classes with a very low latency.
 47  *
 48  * To implement this feature, BFQ constantly tries to detect whether
 49  * the I/O requests in a bfq_queue come from an interactive or a soft
 50  * real-time application. For brevity, in these cases, the queue is
 51  * said to be interactive or soft real-time. In both cases, BFQ
 52  * privileges the service of the queue, over that of non-interactive
 53  * and non-soft-real-time queues. This privileging is performed,
 54  * mainly, by raising the weight of the queue. So, for brevity, we
 55  * call just weight-raising periods the time periods during which a
 56  * queue is privileged, because deemed interactive or soft real-time.
 57  *
 58  * The detection of soft real-time queues/applications is described in
 59  * detail in the comments on the function
 60  * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
 61  * interactive queue works as follows: a queue is deemed interactive
 62  * if it is constantly non empty only for a limited time interval,
 63  * after which it does become empty. The queue may be deemed
 64  * interactive again (for a limited time), if it restarts being
 65  * constantly non empty, provided that this happens only after the
 66  * queue has remained empty for a given minimum idle time.
 67  *
 68  * By default, BFQ computes automatically the above maximum time
 69  * interval, i.e., the time interval after which a constantly
 70  * non-empty queue stops being deemed interactive. Since a queue is
 71  * weight-raised while it is deemed interactive, this maximum time
 72  * interval happens to coincide with the (maximum) duration of the
 73  * weight-raising for interactive queues.
 74  *
 75  * Finally, BFQ also features additional heuristics for
 76  * preserving both a low latency and a high throughput on NCQ-capable,
 77  * rotational or flash-based devices, and to get the job done quickly
 78  * for applications consisting in many I/O-bound processes.
 79  *
 80  * NOTE: if the main or only goal, with a given device, is to achieve
 81  * the maximum-possible throughput at all times, then do switch off
 82  * all low-latency heuristics for that device, by setting low_latency
 83  * to 0.
 84  *
 85  * BFQ is described in [1], where also a reference to the initial,
 86  * more theoretical paper on BFQ can be found. The interested reader
 87  * can find in the latter paper full details on the main algorithm, as
 88  * well as formulas of the guarantees and formal proofs of all the
 89  * properties.  With respect to the version of BFQ presented in these
 90  * papers, this implementation adds a few more heuristics, such as the
 91  * ones that guarantee a low latency to interactive and soft real-time
 92  * applications, and a hierarchical extension based on H-WF2Q+.
 93  *
 94  * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
 95  * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
 96  * with O(log N) complexity derives from the one introduced with EEVDF
 97  * in [3].
 98  *
 99  * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
100  *     Scheduler", Proceedings of the First Workshop on Mobile System
101  *     Technologies (MST-2015), May 2015.
102  *     http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
103  *
104  * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
105  *     Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
106  *     Oct 1997.
107  *
108  * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
109  *
110  * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
111  *     First: A Flexible and Accurate Mechanism for Proportional Share
112  *     Resource Allocation", technical report.
113  *
114  * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
115  */
116 #include <linux/module.h>
117 #include <linux/slab.h>
118 #include <linux/blkdev.h>
119 #include <linux/cgroup.h>
120 #include <linux/ktime.h>
121 #include <linux/rbtree.h>
122 #include <linux/ioprio.h>
123 #include <linux/sbitmap.h>
124 #include <linux/delay.h>
125 #include <linux/backing-dev.h>
126 
127 #include <trace/events/block.h>
128 
129 #include "elevator.h"
130 #include "blk.h"
131 #include "blk-mq.h"
132 #include "blk-mq-tag.h"
133 #include "blk-mq-sched.h"
134 #include "bfq-iosched.h"
135 #include "blk-wbt.h"
136 
137 #define BFQ_BFQQ_FNS(name)                                              \
138 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq)                       \
139 {                                                                       \
140         __set_bit(BFQQF_##name, &(bfqq)->flags);                        \
141 }                                                                       \
142 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq)                      \
143 {                                                                       \
144         __clear_bit(BFQQF_##name, &(bfqq)->flags);              \
145 }                                                                       \
146 int bfq_bfqq_##name(const struct bfq_queue *bfqq)                       \
147 {                                                                       \
148         return test_bit(BFQQF_##name, &(bfqq)->flags);          \
149 }
150 
151 BFQ_BFQQ_FNS(just_created);
152 BFQ_BFQQ_FNS(busy);
153 BFQ_BFQQ_FNS(wait_request);
154 BFQ_BFQQ_FNS(non_blocking_wait_rq);
155 BFQ_BFQQ_FNS(fifo_expire);
156 BFQ_BFQQ_FNS(has_short_ttime);
157 BFQ_BFQQ_FNS(sync);
158 BFQ_BFQQ_FNS(IO_bound);
159 BFQ_BFQQ_FNS(in_large_burst);
160 BFQ_BFQQ_FNS(coop);
161 BFQ_BFQQ_FNS(split_coop);
162 BFQ_BFQQ_FNS(softrt_update);
163 #undef BFQ_BFQQ_FNS                                             \
164 
165 /* Expiration time of async (0) and sync (1) requests, in ns. */
166 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
167 
168 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
169 static const int bfq_back_max = 16 * 1024;
170 
171 /* Penalty of a backwards seek, in number of sectors. */
172 static const int bfq_back_penalty = 2;
173 
174 /* Idling period duration, in ns. */
175 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
176 
177 /* Minimum number of assigned budgets for which stats are safe to compute. */
178 static const int bfq_stats_min_budgets = 194;
179 
180 /* Default maximum budget values, in sectors and number of requests. */
181 static const int bfq_default_max_budget = 16 * 1024;
182 
183 /*
184  * When a sync request is dispatched, the queue that contains that
185  * request, and all the ancestor entities of that queue, are charged
186  * with the number of sectors of the request. In contrast, if the
187  * request is async, then the queue and its ancestor entities are
188  * charged with the number of sectors of the request, multiplied by
189  * the factor below. This throttles the bandwidth for async I/O,
190  * w.r.t. to sync I/O, and it is done to counter the tendency of async
191  * writes to steal I/O throughput to reads.
192  *
193  * The current value of this parameter is the result of a tuning with
194  * several hardware and software configurations. We tried to find the
195  * lowest value for which writes do not cause noticeable problems to
196  * reads. In fact, the lower this parameter, the stabler I/O control,
197  * in the following respect.  The lower this parameter is, the less
198  * the bandwidth enjoyed by a group decreases
199  * - when the group does writes, w.r.t. to when it does reads;
200  * - when other groups do reads, w.r.t. to when they do writes.
201  */
202 static const int bfq_async_charge_factor = 3;
203 
204 /* Default timeout values, in jiffies, approximating CFQ defaults. */
205 const int bfq_timeout = HZ / 8;
206 
207 /*
208  * Time limit for merging (see comments in bfq_setup_cooperator). Set
209  * to the slowest value that, in our tests, proved to be effective in
210  * removing false positives, while not causing true positives to miss
211  * queue merging.
212  *
213  * As can be deduced from the low time limit below, queue merging, if
214  * successful, happens at the very beginning of the I/O of the involved
215  * cooperating processes, as a consequence of the arrival of the very
216  * first requests from each cooperator.  After that, there is very
217  * little chance to find cooperators.
218  */
219 static const unsigned long bfq_merge_time_limit = HZ/10;
220 
221 static struct kmem_cache *bfq_pool;
222 
223 /* Below this threshold (in ns), we consider thinktime immediate. */
224 #define BFQ_MIN_TT              (2 * NSEC_PER_MSEC)
225 
226 /* hw_tag detection: parallel requests threshold and min samples needed. */
227 #define BFQ_HW_QUEUE_THRESHOLD  3
228 #define BFQ_HW_QUEUE_SAMPLES    32
229 
230 #define BFQQ_SEEK_THR           (sector_t)(8 * 100)
231 #define BFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
232 #define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \
233         (get_sdist(last_pos, rq) >                      \
234          BFQQ_SEEK_THR &&                               \
235          (!blk_queue_nonrot(bfqd->queue) ||             \
236           blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT))
237 #define BFQQ_CLOSE_THR          (sector_t)(8 * 1024)
238 #define BFQQ_SEEKY(bfqq)        (hweight32(bfqq->seek_history) > 19)
239 /*
240  * Sync random I/O is likely to be confused with soft real-time I/O,
241  * because it is characterized by limited throughput and apparently
242  * isochronous arrival pattern. To avoid false positives, queues
243  * containing only random (seeky) I/O are prevented from being tagged
244  * as soft real-time.
245  */
246 #define BFQQ_TOTALLY_SEEKY(bfqq)        (bfqq->seek_history == -1)
247 
248 /* Min number of samples required to perform peak-rate update */
249 #define BFQ_RATE_MIN_SAMPLES    32
250 /* Min observation time interval required to perform a peak-rate update (ns) */
251 #define BFQ_RATE_MIN_INTERVAL   (300*NSEC_PER_MSEC)
252 /* Target observation time interval for a peak-rate update (ns) */
253 #define BFQ_RATE_REF_INTERVAL   NSEC_PER_SEC
254 
255 /*
256  * Shift used for peak-rate fixed precision calculations.
257  * With
258  * - the current shift: 16 positions
259  * - the current type used to store rate: u32
260  * - the current unit of measure for rate: [sectors/usec], or, more precisely,
261  *   [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
262  * the range of rates that can be stored is
263  * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
264  * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
265  * [15, 65G] sectors/sec
266  * Which, assuming a sector size of 512B, corresponds to a range of
267  * [7.5K, 33T] B/sec
268  */
269 #define BFQ_RATE_SHIFT          16
270 
271 /*
272  * When configured for computing the duration of the weight-raising
273  * for interactive queues automatically (see the comments at the
274  * beginning of this file), BFQ does it using the following formula:
275  * duration = (ref_rate / r) * ref_wr_duration,
276  * where r is the peak rate of the device, and ref_rate and
277  * ref_wr_duration are two reference parameters.  In particular,
278  * ref_rate is the peak rate of the reference storage device (see
279  * below), and ref_wr_duration is about the maximum time needed, with
280  * BFQ and while reading two files in parallel, to load typical large
281  * applications on the reference device (see the comments on
282  * max_service_from_wr below, for more details on how ref_wr_duration
283  * is obtained).  In practice, the slower/faster the device at hand
284  * is, the more/less it takes to load applications with respect to the
285  * reference device.  Accordingly, the longer/shorter BFQ grants
286  * weight raising to interactive applications.
287  *
288  * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
289  * depending on whether the device is rotational or non-rotational.
290  *
291  * In the following definitions, ref_rate[0] and ref_wr_duration[0]
292  * are the reference values for a rotational device, whereas
293  * ref_rate[1] and ref_wr_duration[1] are the reference values for a
294  * non-rotational device. The reference rates are not the actual peak
295  * rates of the devices used as a reference, but slightly lower
296  * values. The reason for using slightly lower values is that the
297  * peak-rate estimator tends to yield slightly lower values than the
298  * actual peak rate (it can yield the actual peak rate only if there
299  * is only one process doing I/O, and the process does sequential
300  * I/O).
301  *
302  * The reference peak rates are measured in sectors/usec, left-shifted
303  * by BFQ_RATE_SHIFT.
304  */
305 static int ref_rate[2] = {14000, 33000};
306 /*
307  * To improve readability, a conversion function is used to initialize
308  * the following array, which entails that the array can be
309  * initialized only in a function.
310  */
311 static int ref_wr_duration[2];
312 
313 /*
314  * BFQ uses the above-detailed, time-based weight-raising mechanism to
315  * privilege interactive tasks. This mechanism is vulnerable to the
316  * following false positives: I/O-bound applications that will go on
317  * doing I/O for much longer than the duration of weight
318  * raising. These applications have basically no benefit from being
319  * weight-raised at the beginning of their I/O. On the opposite end,
320  * while being weight-raised, these applications
321  * a) unjustly steal throughput to applications that may actually need
322  * low latency;
323  * b) make BFQ uselessly perform device idling; device idling results
324  * in loss of device throughput with most flash-based storage, and may
325  * increase latencies when used purposelessly.
326  *
327  * BFQ tries to reduce these problems, by adopting the following
328  * countermeasure. To introduce this countermeasure, we need first to
329  * finish explaining how the duration of weight-raising for
330  * interactive tasks is computed.
331  *
332  * For a bfq_queue deemed as interactive, the duration of weight
333  * raising is dynamically adjusted, as a function of the estimated
334  * peak rate of the device, so as to be equal to the time needed to
335  * execute the 'largest' interactive task we benchmarked so far. By
336  * largest task, we mean the task for which each involved process has
337  * to do more I/O than for any of the other tasks we benchmarked. This
338  * reference interactive task is the start-up of LibreOffice Writer,
339  * and in this task each process/bfq_queue needs to have at most ~110K
340  * sectors transferred.
341  *
342  * This last piece of information enables BFQ to reduce the actual
343  * duration of weight-raising for at least one class of I/O-bound
344  * applications: those doing sequential or quasi-sequential I/O. An
345  * example is file copy. In fact, once started, the main I/O-bound
346  * processes of these applications usually consume the above 110K
347  * sectors in much less time than the processes of an application that
348  * is starting, because these I/O-bound processes will greedily devote
349  * almost all their CPU cycles only to their target,
350  * throughput-friendly I/O operations. This is even more true if BFQ
351  * happens to be underestimating the device peak rate, and thus
352  * overestimating the duration of weight raising. But, according to
353  * our measurements, once transferred 110K sectors, these processes
354  * have no right to be weight-raised any longer.
355  *
356  * Basing on the last consideration, BFQ ends weight-raising for a
357  * bfq_queue if the latter happens to have received an amount of
358  * service at least equal to the following constant. The constant is
359  * set to slightly more than 110K, to have a minimum safety margin.
360  *
361  * This early ending of weight-raising reduces the amount of time
362  * during which interactive false positives cause the two problems
363  * described at the beginning of these comments.
364  */
365 static const unsigned long max_service_from_wr = 120000;
366 
367 /*
368  * Maximum time between the creation of two queues, for stable merge
369  * to be activated (in ms)
370  */
371 static const unsigned long bfq_activation_stable_merging = 600;
372 /*
373  * Minimum time to be waited before evaluating delayed stable merge (in ms)
374  */
375 static const unsigned long bfq_late_stable_merging = 600;
376 
377 #define RQ_BIC(rq)              ((struct bfq_io_cq *)((rq)->elv.priv[0]))
378 #define RQ_BFQQ(rq)             ((rq)->elv.priv[1])
379 
380 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
381 {
382         return bic->bfqq[is_sync];
383 }
384 
385 static void bfq_put_stable_ref(struct bfq_queue *bfqq);
386 
387 void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
388 {
389         /*
390          * If bfqq != NULL, then a non-stable queue merge between
391          * bic->bfqq and bfqq is happening here. This causes troubles
392          * in the following case: bic->bfqq has also been scheduled
393          * for a possible stable merge with bic->stable_merge_bfqq,
394          * and bic->stable_merge_bfqq == bfqq happens to
395          * hold. Troubles occur because bfqq may then undergo a split,
396          * thereby becoming eligible for a stable merge. Yet, if
397          * bic->stable_merge_bfqq points exactly to bfqq, then bfqq
398          * would be stably merged with itself. To avoid this anomaly,
399          * we cancel the stable merge if
400          * bic->stable_merge_bfqq == bfqq.
401          */
402         bic->bfqq[is_sync] = bfqq;
403 
404         if (bfqq && bic->stable_merge_bfqq == bfqq) {
405                 /*
406                  * Actually, these same instructions are executed also
407                  * in bfq_setup_cooperator, in case of abort or actual
408                  * execution of a stable merge. We could avoid
409                  * repeating these instructions there too, but if we
410                  * did so, we would nest even more complexity in this
411                  * function.
412                  */
413                 bfq_put_stable_ref(bic->stable_merge_bfqq);
414 
415                 bic->stable_merge_bfqq = NULL;
416         }
417 }
418 
419 struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
420 {
421         return bic->icq.q->elevator->elevator_data;
422 }
423 
424 /**
425  * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
426  * @icq: the iocontext queue.
427  */
428 static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
429 {
430         /* bic->icq is the first member, %NULL will convert to %NULL */
431         return container_of(icq, struct bfq_io_cq, icq);
432 }
433 
434 /**
435  * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
436  * @q: the request queue.
437  */
438 static struct bfq_io_cq *bfq_bic_lookup(struct request_queue *q)
439 {
440         struct bfq_io_cq *icq;
441         unsigned long flags;
442 
443         if (!current->io_context)
444                 return NULL;
445 
446         spin_lock_irqsave(&q->queue_lock, flags);
447         icq = icq_to_bic(ioc_lookup_icq(q));
448         spin_unlock_irqrestore(&q->queue_lock, flags);
449 
450         return icq;
451 }
452 
453 /*
454  * Scheduler run of queue, if there are requests pending and no one in the
455  * driver that will restart queueing.
456  */
457 void bfq_schedule_dispatch(struct bfq_data *bfqd)
458 {
459         lockdep_assert_held(&bfqd->lock);
460 
461         if (bfqd->queued != 0) {
462                 bfq_log(bfqd, "schedule dispatch");
463                 blk_mq_run_hw_queues(bfqd->queue, true);
464         }
465 }
466 
467 #define bfq_class_idle(bfqq)    ((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
468 
469 #define bfq_sample_valid(samples)       ((samples) > 80)
470 
471 /*
472  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
473  * We choose the request that is closer to the head right now.  Distance
474  * behind the head is penalized and only allowed to a certain extent.
475  */
476 static struct request *bfq_choose_req(struct bfq_data *bfqd,
477                                       struct request *rq1,
478                                       struct request *rq2,
479                                       sector_t last)
480 {
481         sector_t s1, s2, d1 = 0, d2 = 0;
482         unsigned long back_max;
483 #define BFQ_RQ1_WRAP    0x01 /* request 1 wraps */
484 #define BFQ_RQ2_WRAP    0x02 /* request 2 wraps */
485         unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
486 
487         if (!rq1 || rq1 == rq2)
488                 return rq2;
489         if (!rq2)
490                 return rq1;
491 
492         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
493                 return rq1;
494         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
495                 return rq2;
496         if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
497                 return rq1;
498         else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
499                 return rq2;
500 
501         s1 = blk_rq_pos(rq1);
502         s2 = blk_rq_pos(rq2);
503 
504         /*
505          * By definition, 1KiB is 2 sectors.
506          */
507         back_max = bfqd->bfq_back_max * 2;
508 
509         /*
510          * Strict one way elevator _except_ in the case where we allow
511          * short backward seeks which are biased as twice the cost of a
512          * similar forward seek.
513          */
514         if (s1 >= last)
515                 d1 = s1 - last;
516         else if (s1 + back_max >= last)
517                 d1 = (last - s1) * bfqd->bfq_back_penalty;
518         else
519                 wrap |= BFQ_RQ1_WRAP;
520 
521         if (s2 >= last)
522                 d2 = s2 - last;
523         else if (s2 + back_max >= last)
524                 d2 = (last - s2) * bfqd->bfq_back_penalty;
525         else
526                 wrap |= BFQ_RQ2_WRAP;
527 
528         /* Found required data */
529 
530         /*
531          * By doing switch() on the bit mask "wrap" we avoid having to
532          * check two variables for all permutations: --> faster!
533          */
534         switch (wrap) {
535         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
536                 if (d1 < d2)
537                         return rq1;
538                 else if (d2 < d1)
539                         return rq2;
540 
541                 if (s1 >= s2)
542                         return rq1;
543                 else
544                         return rq2;
545 
546         case BFQ_RQ2_WRAP:
547                 return rq1;
548         case BFQ_RQ1_WRAP:
549                 return rq2;
550         case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
551         default:
552                 /*
553                  * Since both rqs are wrapped,
554                  * start with the one that's further behind head
555                  * (--> only *one* back seek required),
556                  * since back seek takes more time than forward.
557                  */
558                 if (s1 <= s2)
559                         return rq1;
560                 else
561                         return rq2;
562         }
563 }
564 
565 #define BFQ_LIMIT_INLINE_DEPTH 16
566 
567 #ifdef CONFIG_BFQ_GROUP_IOSCHED
568 static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
569 {
570         struct bfq_data *bfqd = bfqq->bfqd;
571         struct bfq_entity *entity = &bfqq->entity;
572         struct bfq_entity *inline_entities[BFQ_LIMIT_INLINE_DEPTH];
573         struct bfq_entity **entities = inline_entities;
574         int depth, level, alloc_depth = BFQ_LIMIT_INLINE_DEPTH;
575         int class_idx = bfqq->ioprio_class - 1;
576         struct bfq_sched_data *sched_data;
577         unsigned long wsum;
578         bool ret = false;
579 
580         if (!entity->on_st_or_in_serv)
581                 return false;
582 
583 retry:
584         spin_lock_irq(&bfqd->lock);
585         /* +1 for bfqq entity, root cgroup not included */
586         depth = bfqg_to_blkg(bfqq_group(bfqq))->blkcg->css.cgroup->level + 1;
587         if (depth > alloc_depth) {
588                 spin_unlock_irq(&bfqd->lock);
589                 if (entities != inline_entities)
590                         kfree(entities);
591                 entities = kmalloc_array(depth, sizeof(*entities), GFP_NOIO);
592                 if (!entities)
593                         return false;
594                 alloc_depth = depth;
595                 goto retry;
596         }
597 
598         sched_data = entity->sched_data;
599         /* Gather our ancestors as we need to traverse them in reverse order */
600         level = 0;
601         for_each_entity(entity) {
602                 /*
603                  * If at some level entity is not even active, allow request
604                  * queueing so that BFQ knows there's work to do and activate
605                  * entities.
606                  */
607                 if (!entity->on_st_or_in_serv)
608                         goto out;
609                 /* Uh, more parents than cgroup subsystem thinks? */
610                 if (WARN_ON_ONCE(level >= depth))
611                         break;
612                 entities[level++] = entity;
613         }
614         WARN_ON_ONCE(level != depth);
615         for (level--; level >= 0; level--) {
616                 entity = entities[level];
617                 if (level > 0) {
618                         wsum = bfq_entity_service_tree(entity)->wsum;
619                 } else {
620                         int i;
621                         /*
622                          * For bfqq itself we take into account service trees
623                          * of all higher priority classes and multiply their
624                          * weights so that low prio queue from higher class
625                          * gets more requests than high prio queue from lower
626                          * class.
627                          */
628                         wsum = 0;
629                         for (i = 0; i <= class_idx; i++) {
630                                 wsum = wsum * IOPRIO_BE_NR +
631                                         sched_data->service_tree[i].wsum;
632                         }
633                 }
634                 limit = DIV_ROUND_CLOSEST(limit * entity->weight, wsum);
635                 if (entity->allocated >= limit) {
636                         bfq_log_bfqq(bfqq->bfqd, bfqq,
637                                 "too many requests: allocated %d limit %d level %d",
638                                 entity->allocated, limit, level);
639                         ret = true;
640                         break;
641                 }
642         }
643 out:
644         spin_unlock_irq(&bfqd->lock);
645         if (entities != inline_entities)
646                 kfree(entities);
647         return ret;
648 }
649 #else
650 static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
651 {
652         return false;
653 }
654 #endif
655 
656 /*
657  * Async I/O can easily starve sync I/O (both sync reads and sync
658  * writes), by consuming all tags. Similarly, storms of sync writes,
659  * such as those that sync(2) may trigger, can starve sync reads.
660  * Limit depths of async I/O and sync writes so as to counter both
661  * problems.
662  *
663  * Also if a bfq queue or its parent cgroup consume more tags than would be
664  * appropriate for their weight, we trim the available tag depth to 1. This
665  * avoids a situation where one cgroup can starve another cgroup from tags and
666  * thus block service differentiation among cgroups. Note that because the
667  * queue / cgroup already has many requests allocated and queued, this does not
668  * significantly affect service guarantees coming from the BFQ scheduling
669  * algorithm.
670  */
671 static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
672 {
673         struct bfq_data *bfqd = data->q->elevator->elevator_data;
674         struct bfq_io_cq *bic = bfq_bic_lookup(data->q);
675         struct bfq_queue *bfqq = bic ? bic_to_bfqq(bic, op_is_sync(op)) : NULL;
676         int depth;
677         unsigned limit = data->q->nr_requests;
678 
679         /* Sync reads have full depth available */
680         if (op_is_sync(op) && !op_is_write(op)) {
681                 depth = 0;
682         } else {
683                 depth = bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)];
684                 limit = (limit * depth) >> bfqd->full_depth_shift;
685         }
686 
687         /*
688          * Does queue (or any parent entity) exceed number of requests that
689          * should be available to it? Heavily limit depth so that it cannot
690          * consume more available requests and thus starve other entities.
691          */
692         if (bfqq && bfqq_request_over_limit(bfqq, limit))
693                 depth = 1;
694 
695         bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
696                 __func__, bfqd->wr_busy_queues, op_is_sync(op), depth);
697         if (depth)
698                 data->shallow_depth = depth;
699 }
700 
701 static struct bfq_queue *
702 bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
703                      sector_t sector, struct rb_node **ret_parent,
704                      struct rb_node ***rb_link)
705 {
706         struct rb_node **p, *parent;
707         struct bfq_queue *bfqq = NULL;
708 
709         parent = NULL;
710         p = &root->rb_node;
711         while (*p) {
712                 struct rb_node **n;
713 
714                 parent = *p;
715                 bfqq = rb_entry(parent, struct bfq_queue, pos_node);
716 
717                 /*
718                  * Sort strictly based on sector. Smallest to the left,
719                  * largest to the right.
720                  */
721                 if (sector > blk_rq_pos(bfqq->next_rq))
722                         n = &(*p)->rb_right;
723                 else if (sector < blk_rq_pos(bfqq->next_rq))
724                         n = &(*p)->rb_left;
725                 else
726                         break;
727                 p = n;
728                 bfqq = NULL;
729         }
730 
731         *ret_parent = parent;
732         if (rb_link)
733                 *rb_link = p;
734 
735         bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
736                 (unsigned long long)sector,
737                 bfqq ? bfqq->pid : 0);
738 
739         return bfqq;
740 }
741 
742 static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
743 {
744         return bfqq->service_from_backlogged > 0 &&
745                 time_is_before_jiffies(bfqq->first_IO_time +
746                                        bfq_merge_time_limit);
747 }
748 
749 /*
750  * The following function is not marked as __cold because it is
751  * actually cold, but for the same performance goal described in the
752  * comments on the likely() at the beginning of
753  * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
754  * execution time for the case where this function is not invoked, we
755  * had to add an unlikely() in each involved if().
756  */
757 void __cold
758 bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
759 {
760         struct rb_node **p, *parent;
761         struct bfq_queue *__bfqq;
762 
763         if (bfqq->pos_root) {
764                 rb_erase(&bfqq->pos_node, bfqq->pos_root);
765                 bfqq->pos_root = NULL;
766         }
767 
768         /* oom_bfqq does not participate in queue merging */
769         if (bfqq == &bfqd->oom_bfqq)
770                 return;
771 
772         /*
773          * bfqq cannot be merged any longer (see comments in
774          * bfq_setup_cooperator): no point in adding bfqq into the
775          * position tree.
776          */
777         if (bfq_too_late_for_merging(bfqq))
778                 return;
779 
780         if (bfq_class_idle(bfqq))
781                 return;
782         if (!bfqq->next_rq)
783                 return;
784 
785         bfqq->pos_root = &bfqq_group(bfqq)->rq_pos_tree;
786         __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
787                         blk_rq_pos(bfqq->next_rq), &parent, &p);
788         if (!__bfqq) {
789                 rb_link_node(&bfqq->pos_node, parent, p);
790                 rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
791         } else
792                 bfqq->pos_root = NULL;
793 }
794 
795 /*
796  * The following function returns false either if every active queue
797  * must receive the same share of the throughput (symmetric scenario),
798  * or, as a special case, if bfqq must receive a share of the
799  * throughput lower than or equal to the share that every other active
800  * queue must receive.  If bfqq does sync I/O, then these are the only
801  * two cases where bfqq happens to be guaranteed its share of the
802  * throughput even if I/O dispatching is not plugged when bfqq remains
803  * temporarily empty (for more details, see the comments in the
804  * function bfq_better_to_idle()). For this reason, the return value
805  * of this function is used to check whether I/O-dispatch plugging can
806  * be avoided.
807  *
808  * The above first case (symmetric scenario) occurs when:
809  * 1) all active queues have the same weight,
810  * 2) all active queues belong to the same I/O-priority class,
811  * 3) all active groups at the same level in the groups tree have the same
812  *    weight,
813  * 4) all active groups at the same level in the groups tree have the same
814  *    number of children.
815  *
816  * Unfortunately, keeping the necessary state for evaluating exactly
817  * the last two symmetry sub-conditions above would be quite complex
818  * and time consuming. Therefore this function evaluates, instead,
819  * only the following stronger three sub-conditions, for which it is
820  * much easier to maintain the needed state:
821  * 1) all active queues have the same weight,
822  * 2) all active queues belong to the same I/O-priority class,
823  * 3) there are no active groups.
824  * In particular, the last condition is always true if hierarchical
825  * support or the cgroups interface are not enabled, thus no state
826  * needs to be maintained in this case.
827  */
828 static bool bfq_asymmetric_scenario(struct bfq_data *bfqd,
829                                    struct bfq_queue *bfqq)
830 {
831         bool smallest_weight = bfqq &&
832                 bfqq->weight_counter &&
833                 bfqq->weight_counter ==
834                 container_of(
835                         rb_first_cached(&bfqd->queue_weights_tree),
836                         struct bfq_weight_counter,
837                         weights_node);
838 
839         /*
840          * For queue weights to differ, queue_weights_tree must contain
841          * at least two nodes.
842          */
843         bool varied_queue_weights = !smallest_weight &&
844                 !RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) &&
845                 (bfqd->queue_weights_tree.rb_root.rb_node->rb_left ||
846                  bfqd->queue_weights_tree.rb_root.rb_node->rb_right);
847 
848         bool multiple_classes_busy =
849                 (bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
850                 (bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
851                 (bfqd->busy_queues[1] && bfqd->busy_queues[2]);
852 
853         return varied_queue_weights || multiple_classes_busy
854 #ifdef CONFIG_BFQ_GROUP_IOSCHED
855                || bfqd->num_groups_with_pending_reqs > 0
856 #endif
857                 ;
858 }
859 
860 /*
861  * If the weight-counter tree passed as input contains no counter for
862  * the weight of the input queue, then add that counter; otherwise just
863  * increment the existing counter.
864  *
865  * Note that weight-counter trees contain few nodes in mostly symmetric
866  * scenarios. For example, if all queues have the same weight, then the
867  * weight-counter tree for the queues may contain at most one node.
868  * This holds even if low_latency is on, because weight-raised queues
869  * are not inserted in the tree.
870  * In most scenarios, the rate at which nodes are created/destroyed
871  * should be low too.
872  */
873 void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq,
874                           struct rb_root_cached *root)
875 {
876         struct bfq_entity *entity = &bfqq->entity;
877         struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
878         bool leftmost = true;
879 
880         /*
881          * Do not insert if the queue is already associated with a
882          * counter, which happens if:
883          *   1) a request arrival has caused the queue to become both
884          *      non-weight-raised, and hence change its weight, and
885          *      backlogged; in this respect, each of the two events
886          *      causes an invocation of this function,
887          *   2) this is the invocation of this function caused by the
888          *      second event. This second invocation is actually useless,
889          *      and we handle this fact by exiting immediately. More
890          *      efficient or clearer solutions might possibly be adopted.
891          */
892         if (bfqq->weight_counter)
893                 return;
894 
895         while (*new) {
896                 struct bfq_weight_counter *__counter = container_of(*new,
897                                                 struct bfq_weight_counter,
898                                                 weights_node);
899                 parent = *new;
900 
901                 if (entity->weight == __counter->weight) {
902                         bfqq->weight_counter = __counter;
903                         goto inc_counter;
904                 }
905                 if (entity->weight < __counter->weight)
906                         new = &((*new)->rb_left);
907                 else {
908                         new = &((*new)->rb_right);
909                         leftmost = false;
910                 }
911         }
912 
913         bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
914                                        GFP_ATOMIC);
915 
916         /*
917          * In the unlucky event of an allocation failure, we just
918          * exit. This will cause the weight of queue to not be
919          * considered in bfq_asymmetric_scenario, which, in its turn,
920          * causes the scenario to be deemed wrongly symmetric in case
921          * bfqq's weight would have been the only weight making the
922          * scenario asymmetric.  On the bright side, no unbalance will
923          * however occur when bfqq becomes inactive again (the
924          * invocation of this function is triggered by an activation
925          * of queue).  In fact, bfq_weights_tree_remove does nothing
926          * if !bfqq->weight_counter.
927          */
928         if (unlikely(!bfqq->weight_counter))
929                 return;
930 
931         bfqq->weight_counter->weight = entity->weight;
932         rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
933         rb_insert_color_cached(&bfqq->weight_counter->weights_node, root,
934                                 leftmost);
935 
936 inc_counter:
937         bfqq->weight_counter->num_active++;
938         bfqq->ref++;
939 }
940 
941 /*
942  * Decrement the weight counter associated with the queue, and, if the
943  * counter reaches 0, remove the counter from the tree.
944  * See the comments to the function bfq_weights_tree_add() for considerations
945  * about overhead.
946  */
947 void __bfq_weights_tree_remove(struct bfq_data *bfqd,
948                                struct bfq_queue *bfqq,
949                                struct rb_root_cached *root)
950 {
951         if (!bfqq->weight_counter)
952                 return;
953 
954         bfqq->weight_counter->num_active--;
955         if (bfqq->weight_counter->num_active > 0)
956                 goto reset_entity_pointer;
957 
958         rb_erase_cached(&bfqq->weight_counter->weights_node, root);
959         kfree(bfqq->weight_counter);
960 
961 reset_entity_pointer:
962         bfqq->weight_counter = NULL;
963         bfq_put_queue(bfqq);
964 }
965 
966 /*
967  * Invoke __bfq_weights_tree_remove on bfqq and decrement the number
968  * of active groups for each queue's inactive parent entity.
969  */
970 void bfq_weights_tree_remove(struct bfq_data *bfqd,
971                              struct bfq_queue *bfqq)
972 {
973         struct bfq_entity *entity = bfqq->entity.parent;
974 
975         for_each_entity(entity) {
976                 struct bfq_sched_data *sd = entity->my_sched_data;
977 
978                 if (sd->next_in_service || sd->in_service_entity) {
979                         /*
980                          * entity is still active, because either
981                          * next_in_service or in_service_entity is not
982                          * NULL (see the comments on the definition of
983                          * next_in_service for details on why
984                          * in_service_entity must be checked too).
985                          *
986                          * As a consequence, its parent entities are
987                          * active as well, and thus this loop must
988                          * stop here.
989                          */
990                         break;
991                 }
992 
993                 /*
994                  * The decrement of num_groups_with_pending_reqs is
995                  * not performed immediately upon the deactivation of
996                  * entity, but it is delayed to when it also happens
997                  * that the first leaf descendant bfqq of entity gets
998                  * all its pending requests completed. The following
999                  * instructions perform this delayed decrement, if
1000                  * needed. See the comments on
1001                  * num_groups_with_pending_reqs for details.
1002                  */
1003                 if (entity->in_groups_with_pending_reqs) {
1004                         entity->in_groups_with_pending_reqs = false;
1005                         bfqd->num_groups_with_pending_reqs--;
1006                 }
1007         }
1008 
1009         /*
1010          * Next function is invoked last, because it causes bfqq to be
1011          * freed if the following holds: bfqq is not in service and
1012          * has no dispatched request. DO NOT use bfqq after the next
1013          * function invocation.
1014          */
1015         __bfq_weights_tree_remove(bfqd, bfqq,
1016                                   &bfqd->queue_weights_tree);
1017 }
1018 
1019 /*
1020  * Return expired entry, or NULL to just start from scratch in rbtree.
1021  */
1022 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
1023                                       struct request *last)
1024 {
1025         struct request *rq;
1026 
1027         if (bfq_bfqq_fifo_expire(bfqq))
1028                 return NULL;
1029 
1030         bfq_mark_bfqq_fifo_expire(bfqq);
1031 
1032         rq = rq_entry_fifo(bfqq->fifo.next);
1033 
1034         if (rq == last || ktime_get_ns() < rq->fifo_time)
1035                 return NULL;
1036 
1037         bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
1038         return rq;
1039 }
1040 
1041 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
1042                                         struct bfq_queue *bfqq,
1043                                         struct request *last)
1044 {
1045         struct rb_node *rbnext = rb_next(&last->rb_node);
1046         struct rb_node *rbprev = rb_prev(&last->rb_node);
1047         struct request *next, *prev = NULL;
1048 
1049         /* Follow expired path, else get first next available. */
1050         next = bfq_check_fifo(bfqq, last);
1051         if (next)
1052                 return next;
1053 
1054         if (rbprev)
1055                 prev = rb_entry_rq(rbprev);
1056 
1057         if (rbnext)
1058                 next = rb_entry_rq(rbnext);
1059         else {
1060                 rbnext = rb_first(&bfqq->sort_list);
1061                 if (rbnext && rbnext != &last->rb_node)
1062                         next = rb_entry_rq(rbnext);
1063         }
1064 
1065         return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
1066 }
1067 
1068 /* see the definition of bfq_async_charge_factor for details */
1069 static unsigned long bfq_serv_to_charge(struct request *rq,
1070                                         struct bfq_queue *bfqq)
1071 {
1072         if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
1073             bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
1074                 return blk_rq_sectors(rq);
1075 
1076         return blk_rq_sectors(rq) * bfq_async_charge_factor;
1077 }
1078 
1079 /**
1080  * bfq_updated_next_req - update the queue after a new next_rq selection.
1081  * @bfqd: the device data the queue belongs to.
1082  * @bfqq: the queue to update.
1083  *
1084  * If the first request of a queue changes we make sure that the queue
1085  * has enough budget to serve at least its first request (if the
1086  * request has grown).  We do this because if the queue has not enough
1087  * budget for its first request, it has to go through two dispatch
1088  * rounds to actually get it dispatched.
1089  */
1090 static void bfq_updated_next_req(struct bfq_data *bfqd,
1091                                  struct bfq_queue *bfqq)
1092 {
1093         struct bfq_entity *entity = &bfqq->entity;
1094         struct request *next_rq = bfqq->next_rq;
1095         unsigned long new_budget;
1096 
1097         if (!next_rq)
1098                 return;
1099 
1100         if (bfqq == bfqd->in_service_queue)
1101                 /*
1102                  * In order not to break guarantees, budgets cannot be
1103                  * changed after an entity has been selected.
1104                  */
1105                 return;
1106 
1107         new_budget = max_t(unsigned long,
1108                            max_t(unsigned long, bfqq->max_budget,
1109                                  bfq_serv_to_charge(next_rq, bfqq)),
1110                            entity->service);
1111         if (entity->budget != new_budget) {
1112                 entity->budget = new_budget;
1113                 bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
1114                                          new_budget);
1115                 bfq_requeue_bfqq(bfqd, bfqq, false);
1116         }
1117 }
1118 
1119 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
1120 {
1121         u64 dur;
1122 
1123         if (bfqd->bfq_wr_max_time > 0)
1124                 return bfqd->bfq_wr_max_time;
1125 
1126         dur = bfqd->rate_dur_prod;
1127         do_div(dur, bfqd->peak_rate);
1128 
1129         /*
1130          * Limit duration between 3 and 25 seconds. The upper limit
1131          * has been conservatively set after the following worst case:
1132          * on a QEMU/KVM virtual machine
1133          * - running in a slow PC
1134          * - with a virtual disk stacked on a slow low-end 5400rpm HDD
1135          * - serving a heavy I/O workload, such as the sequential reading
1136          *   of several files
1137          * mplayer took 23 seconds to start, if constantly weight-raised.
1138          *
1139          * As for higher values than that accommodating the above bad
1140          * scenario, tests show that higher values would often yield
1141          * the opposite of the desired result, i.e., would worsen
1142          * responsiveness by allowing non-interactive applications to
1143          * preserve weight raising for too long.
1144          *
1145          * On the other end, lower values than 3 seconds make it
1146          * difficult for most interactive tasks to complete their jobs
1147          * before weight-raising finishes.
1148          */
1149         return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
1150 }
1151 
1152 /* switch back from soft real-time to interactive weight raising */
1153 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
1154                                           struct bfq_data *bfqd)
1155 {
1156         bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1157         bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1158         bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
1159 }
1160 
1161 static void
1162 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
1163                       struct bfq_io_cq *bic, bool bfq_already_existing)
1164 {
1165         unsigned int old_wr_coeff = 1;
1166         bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
1167 
1168         if (bic->saved_has_short_ttime)
1169                 bfq_mark_bfqq_has_short_ttime(bfqq);
1170         else
1171                 bfq_clear_bfqq_has_short_ttime(bfqq);
1172 
1173         if (bic->saved_IO_bound)
1174                 bfq_mark_bfqq_IO_bound(bfqq);
1175         else
1176                 bfq_clear_bfqq_IO_bound(bfqq);
1177 
1178         bfqq->last_serv_time_ns = bic->saved_last_serv_time_ns;
1179         bfqq->inject_limit = bic->saved_inject_limit;
1180         bfqq->decrease_time_jif = bic->saved_decrease_time_jif;
1181 
1182         bfqq->entity.new_weight = bic->saved_weight;
1183         bfqq->ttime = bic->saved_ttime;
1184         bfqq->io_start_time = bic->saved_io_start_time;
1185         bfqq->tot_idle_time = bic->saved_tot_idle_time;
1186         /*
1187          * Restore weight coefficient only if low_latency is on
1188          */
1189         if (bfqd->low_latency) {
1190                 old_wr_coeff = bfqq->wr_coeff;
1191                 bfqq->wr_coeff = bic->saved_wr_coeff;
1192         }
1193         bfqq->service_from_wr = bic->saved_service_from_wr;
1194         bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
1195         bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
1196         bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
1197 
1198         if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1199             time_is_before_jiffies(bfqq->last_wr_start_finish +
1200                                    bfqq->wr_cur_max_time))) {
1201                 if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1202                     !bfq_bfqq_in_large_burst(bfqq) &&
1203                     time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1204                                              bfq_wr_duration(bfqd))) {
1205                         switch_back_to_interactive_wr(bfqq, bfqd);
1206                 } else {
1207                         bfqq->wr_coeff = 1;
1208                         bfq_log_bfqq(bfqq->bfqd, bfqq,
1209                                      "resume state: switching off wr");
1210                 }
1211         }
1212 
1213         /* make sure weight will be updated, however we got here */
1214         bfqq->entity.prio_changed = 1;
1215 
1216         if (likely(!busy))
1217                 return;
1218 
1219         if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1220                 bfqd->wr_busy_queues++;
1221         else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1222                 bfqd->wr_busy_queues--;
1223 }
1224 
1225 static int bfqq_process_refs(struct bfq_queue *bfqq)
1226 {
1227         return bfqq->ref - bfqq->entity.allocated -
1228                 bfqq->entity.on_st_or_in_serv -
1229                 (bfqq->weight_counter != NULL) - bfqq->stable_ref;
1230 }
1231 
1232 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1233 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1234 {
1235         struct bfq_queue *item;
1236         struct hlist_node *n;
1237 
1238         hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1239                 hlist_del_init(&item->burst_list_node);
1240 
1241         /*
1242          * Start the creation of a new burst list only if there is no
1243          * active queue. See comments on the conditional invocation of
1244          * bfq_handle_burst().
1245          */
1246         if (bfq_tot_busy_queues(bfqd) == 0) {
1247                 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1248                 bfqd->burst_size = 1;
1249         } else
1250                 bfqd->burst_size = 0;
1251 
1252         bfqd->burst_parent_entity = bfqq->entity.parent;
1253 }
1254 
1255 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1256 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1257 {
1258         /* Increment burst size to take into account also bfqq */
1259         bfqd->burst_size++;
1260 
1261         if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1262                 struct bfq_queue *pos, *bfqq_item;
1263                 struct hlist_node *n;
1264 
1265                 /*
1266                  * Enough queues have been activated shortly after each
1267                  * other to consider this burst as large.
1268                  */
1269                 bfqd->large_burst = true;
1270 
1271                 /*
1272                  * We can now mark all queues in the burst list as
1273                  * belonging to a large burst.
1274                  */
1275                 hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1276                                      burst_list_node)
1277                         bfq_mark_bfqq_in_large_burst(bfqq_item);
1278                 bfq_mark_bfqq_in_large_burst(bfqq);
1279 
1280                 /*
1281                  * From now on, and until the current burst finishes, any
1282                  * new queue being activated shortly after the last queue
1283                  * was inserted in the burst can be immediately marked as
1284                  * belonging to a large burst. So the burst list is not
1285                  * needed any more. Remove it.
1286                  */
1287                 hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1288                                           burst_list_node)
1289                         hlist_del_init(&pos->burst_list_node);
1290         } else /*
1291                 * Burst not yet large: add bfqq to the burst list. Do
1292                 * not increment the ref counter for bfqq, because bfqq
1293                 * is removed from the burst list before freeing bfqq
1294                 * in put_queue.
1295                 */
1296                 hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1297 }
1298 
1299 /*
1300  * If many queues belonging to the same group happen to be created
1301  * shortly after each other, then the processes associated with these
1302  * queues have typically a common goal. In particular, bursts of queue
1303  * creations are usually caused by services or applications that spawn
1304  * many parallel threads/processes. Examples are systemd during boot,
1305  * or git grep. To help these processes get their job done as soon as
1306  * possible, it is usually better to not grant either weight-raising
1307  * or device idling to their queues, unless these queues must be
1308  * protected from the I/O flowing through other active queues.
1309  *
1310  * In this comment we describe, firstly, the reasons why this fact
1311  * holds, and, secondly, the next function, which implements the main
1312  * steps needed to properly mark these queues so that they can then be
1313  * treated in a different way.
1314  *
1315  * The above services or applications benefit mostly from a high
1316  * throughput: the quicker the requests of the activated queues are
1317  * cumulatively served, the sooner the target job of these queues gets
1318  * completed. As a consequence, weight-raising any of these queues,
1319  * which also implies idling the device for it, is almost always
1320  * counterproductive, unless there are other active queues to isolate
1321  * these new queues from. If there no other active queues, then
1322  * weight-raising these new queues just lowers throughput in most
1323  * cases.
1324  *
1325  * On the other hand, a burst of queue creations may be caused also by
1326  * the start of an application that does not consist of a lot of
1327  * parallel I/O-bound threads. In fact, with a complex application,
1328  * several short processes may need to be executed to start-up the
1329  * application. In this respect, to start an application as quickly as
1330  * possible, the best thing to do is in any case to privilege the I/O
1331  * related to the application with respect to all other
1332  * I/O. Therefore, the best strategy to start as quickly as possible
1333  * an application that causes a burst of queue creations is to
1334  * weight-raise all the queues created during the burst. This is the
1335  * exact opposite of the best strategy for the other type of bursts.
1336  *
1337  * In the end, to take the best action for each of the two cases, the
1338  * two types of bursts need to be distinguished. Fortunately, this
1339  * seems relatively easy, by looking at the sizes of the bursts. In
1340  * particular, we found a threshold such that only bursts with a
1341  * larger size than that threshold are apparently caused by
1342  * services or commands such as systemd or git grep. For brevity,
1343  * hereafter we call just 'large' these bursts. BFQ *does not*
1344  * weight-raise queues whose creation occurs in a large burst. In
1345  * addition, for each of these queues BFQ performs or does not perform
1346  * idling depending on which choice boosts the throughput more. The
1347  * exact choice depends on the device and request pattern at
1348  * hand.
1349  *
1350  * Unfortunately, false positives may occur while an interactive task
1351  * is starting (e.g., an application is being started). The
1352  * consequence is that the queues associated with the task do not
1353  * enjoy weight raising as expected. Fortunately these false positives
1354  * are very rare. They typically occur if some service happens to
1355  * start doing I/O exactly when the interactive task starts.
1356  *
1357  * Turning back to the next function, it is invoked only if there are
1358  * no active queues (apart from active queues that would belong to the
1359  * same, possible burst bfqq would belong to), and it implements all
1360  * the steps needed to detect the occurrence of a large burst and to
1361  * properly mark all the queues belonging to it (so that they can then
1362  * be treated in a different way). This goal is achieved by
1363  * maintaining a "burst list" that holds, temporarily, the queues that
1364  * belong to the burst in progress. The list is then used to mark
1365  * these queues as belonging to a large burst if the burst does become
1366  * large. The main steps are the following.
1367  *
1368  * . when the very first queue is created, the queue is inserted into the
1369  *   list (as it could be the first queue in a possible burst)
1370  *
1371  * . if the current burst has not yet become large, and a queue Q that does
1372  *   not yet belong to the burst is activated shortly after the last time
1373  *   at which a new queue entered the burst list, then the function appends
1374  *   Q to the burst list
1375  *
1376  * . if, as a consequence of the previous step, the burst size reaches
1377  *   the large-burst threshold, then
1378  *
1379  *     . all the queues in the burst list are marked as belonging to a
1380  *       large burst
1381  *
1382  *     . the burst list is deleted; in fact, the burst list already served
1383  *       its purpose (keeping temporarily track of the queues in a burst,
1384  *       so as to be able to mark them as belonging to a large burst in the
1385  *       previous sub-step), and now is not needed any more
1386  *
1387  *     . the device enters a large-burst mode
1388  *
1389  * . if a queue Q that does not belong to the burst is created while
1390  *   the device is in large-burst mode and shortly after the last time
1391  *   at which a queue either entered the burst list or was marked as
1392  *   belonging to the current large burst, then Q is immediately marked
1393  *   as belonging to a large burst.
1394  *
1395  * . if a queue Q that does not belong to the burst is created a while
1396  *   later, i.e., not shortly after, than the last time at which a queue
1397  *   either entered the burst list or was marked as belonging to the
1398  *   current large burst, then the current burst is deemed as finished and:
1399  *
1400  *        . the large-burst mode is reset if set
1401  *
1402  *        . the burst list is emptied
1403  *
1404  *        . Q is inserted in the burst list, as Q may be the first queue
1405  *          in a possible new burst (then the burst list contains just Q
1406  *          after this step).
1407  */
1408 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1409 {
1410         /*
1411          * If bfqq is already in the burst list or is part of a large
1412          * burst, or finally has just been split, then there is
1413          * nothing else to do.
1414          */
1415         if (!hlist_unhashed(&bfqq->burst_list_node) ||
1416             bfq_bfqq_in_large_burst(bfqq) ||
1417             time_is_after_eq_jiffies(bfqq->split_time +
1418                                      msecs_to_jiffies(10)))
1419                 return;
1420 
1421         /*
1422          * If bfqq's creation happens late enough, or bfqq belongs to
1423          * a different group than the burst group, then the current
1424          * burst is finished, and related data structures must be
1425          * reset.
1426          *
1427          * In this respect, consider the special case where bfqq is
1428          * the very first queue created after BFQ is selected for this
1429          * device. In this case, last_ins_in_burst and
1430          * burst_parent_entity are not yet significant when we get
1431          * here. But it is easy to verify that, whether or not the
1432          * following condition is true, bfqq will end up being
1433          * inserted into the burst list. In particular the list will
1434          * happen to contain only bfqq. And this is exactly what has
1435          * to happen, as bfqq may be the first queue of the first
1436          * burst.
1437          */
1438         if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1439             bfqd->bfq_burst_interval) ||
1440             bfqq->entity.parent != bfqd->burst_parent_entity) {
1441                 bfqd->large_burst = false;
1442                 bfq_reset_burst_list(bfqd, bfqq);
1443                 goto end;
1444         }
1445 
1446         /*
1447          * If we get here, then bfqq is being activated shortly after the
1448          * last queue. So, if the current burst is also large, we can mark
1449          * bfqq as belonging to this large burst immediately.
1450          */
1451         if (bfqd->large_burst) {
1452                 bfq_mark_bfqq_in_large_burst(bfqq);
1453                 goto end;
1454         }
1455 
1456         /*
1457          * If we get here, then a large-burst state has not yet been
1458          * reached, but bfqq is being activated shortly after the last
1459          * queue. Then we add bfqq to the burst.
1460          */
1461         bfq_add_to_burst(bfqd, bfqq);
1462 end:
1463         /*
1464          * At this point, bfqq either has been added to the current
1465          * burst or has caused the current burst to terminate and a
1466          * possible new burst to start. In particular, in the second
1467          * case, bfqq has become the first queue in the possible new
1468          * burst.  In both cases last_ins_in_burst needs to be moved
1469          * forward.
1470          */
1471         bfqd->last_ins_in_burst = jiffies;
1472 }
1473 
1474 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1475 {
1476         struct bfq_entity *entity = &bfqq->entity;
1477 
1478         return entity->budget - entity->service;
1479 }
1480 
1481 /*
1482  * If enough samples have been computed, return the current max budget
1483  * stored in bfqd, which is dynamically updated according to the
1484  * estimated disk peak rate; otherwise return the default max budget
1485  */
1486 static int bfq_max_budget(struct bfq_data *bfqd)
1487 {
1488         if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1489                 return bfq_default_max_budget;
1490         else
1491                 return bfqd->bfq_max_budget;
1492 }
1493 
1494 /*
1495  * Return min budget, which is a fraction of the current or default
1496  * max budget (trying with 1/32)
1497  */
1498 static int bfq_min_budget(struct bfq_data *bfqd)
1499 {
1500         if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1501                 return bfq_default_max_budget / 32;
1502         else
1503                 return bfqd->bfq_max_budget / 32;
1504 }
1505 
1506 /*
1507  * The next function, invoked after the input queue bfqq switches from
1508  * idle to busy, updates the budget of bfqq. The function also tells
1509  * whether the in-service queue should be expired, by returning
1510  * true. The purpose of expiring the in-service queue is to give bfqq
1511  * the chance to possibly preempt the in-service queue, and the reason
1512  * for preempting the in-service queue is to achieve one of the two
1513  * goals below.
1514  *
1515  * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1516  * expired because it has remained idle. In particular, bfqq may have
1517  * expired for one of the following two reasons:
1518  *
1519  * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1520  *   and did not make it to issue a new request before its last
1521  *   request was served;
1522  *
1523  * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1524  *   a new request before the expiration of the idling-time.
1525  *
1526  * Even if bfqq has expired for one of the above reasons, the process
1527  * associated with the queue may be however issuing requests greedily,
1528  * and thus be sensitive to the bandwidth it receives (bfqq may have
1529  * remained idle for other reasons: CPU high load, bfqq not enjoying
1530  * idling, I/O throttling somewhere in the path from the process to
1531  * the I/O scheduler, ...). But if, after every expiration for one of
1532  * the above two reasons, bfqq has to wait for the service of at least
1533  * one full budget of another queue before being served again, then
1534  * bfqq is likely to get a much lower bandwidth or resource time than
1535  * its reserved ones. To address this issue, two countermeasures need
1536  * to be taken.
1537  *
1538  * First, the budget and the timestamps of bfqq need to be updated in
1539  * a special way on bfqq reactivation: they need to be updated as if
1540  * bfqq did not remain idle and did not expire. In fact, if they are
1541  * computed as if bfqq expired and remained idle until reactivation,
1542  * then the process associated with bfqq is treated as if, instead of
1543  * being greedy, it stopped issuing requests when bfqq remained idle,
1544  * and restarts issuing requests only on this reactivation. In other
1545  * words, the scheduler does not help the process recover the "service
1546  * hole" between bfqq expiration and reactivation. As a consequence,
1547  * the process receives a lower bandwidth than its reserved one. In
1548  * contrast, to recover this hole, the budget must be updated as if
1549  * bfqq was not expired at all before this reactivation, i.e., it must
1550  * be set to the value of the remaining budget when bfqq was
1551  * expired. Along the same line, timestamps need to be assigned the
1552  * value they had the last time bfqq was selected for service, i.e.,
1553  * before last expiration. Thus timestamps need to be back-shifted
1554  * with respect to their normal computation (see [1] for more details
1555  * on this tricky aspect).
1556  *
1557  * Secondly, to allow the process to recover the hole, the in-service
1558  * queue must be expired too, to give bfqq the chance to preempt it
1559  * immediately. In fact, if bfqq has to wait for a full budget of the
1560  * in-service queue to be completed, then it may become impossible to
1561  * let the process recover the hole, even if the back-shifted
1562  * timestamps of bfqq are lower than those of the in-service queue. If
1563  * this happens for most or all of the holes, then the process may not
1564  * receive its reserved bandwidth. In this respect, it is worth noting
1565  * that, being the service of outstanding requests unpreemptible, a
1566  * little fraction of the holes may however be unrecoverable, thereby
1567  * causing a little loss of bandwidth.
1568  *
1569  * The last important point is detecting whether bfqq does need this
1570  * bandwidth recovery. In this respect, the next function deems the
1571  * process associated with bfqq greedy, and thus allows it to recover
1572  * the hole, if: 1) the process is waiting for the arrival of a new
1573  * request (which implies that bfqq expired for one of the above two
1574  * reasons), and 2) such a request has arrived soon. The first
1575  * condition is controlled through the flag non_blocking_wait_rq,
1576  * while the second through the flag arrived_in_time. If both
1577  * conditions hold, then the function computes the budget in the
1578  * above-described special way, and signals that the in-service queue
1579  * should be expired. Timestamp back-shifting is done later in
1580  * __bfq_activate_entity.
1581  *
1582  * 2. Reduce latency. Even if timestamps are not backshifted to let
1583  * the process associated with bfqq recover a service hole, bfqq may
1584  * however happen to have, after being (re)activated, a lower finish
1585  * timestamp than the in-service queue.  That is, the next budget of
1586  * bfqq may have to be completed before the one of the in-service
1587  * queue. If this is the case, then preempting the in-service queue
1588  * allows this goal to be achieved, apart from the unpreemptible,
1589  * outstanding requests mentioned above.
1590  *
1591  * Unfortunately, regardless of which of the above two goals one wants
1592  * to achieve, service trees need first to be updated to know whether
1593  * the in-service queue must be preempted. To have service trees
1594  * correctly updated, the in-service queue must be expired and
1595  * rescheduled, and bfqq must be scheduled too. This is one of the
1596  * most costly operations (in future versions, the scheduling
1597  * mechanism may be re-designed in such a way to make it possible to
1598  * know whether preemption is needed without needing to update service
1599  * trees). In addition, queue preemptions almost always cause random
1600  * I/O, which may in turn cause loss of throughput. Finally, there may
1601  * even be no in-service queue when the next function is invoked (so,
1602  * no queue to compare timestamps with). Because of these facts, the
1603  * next function adopts the following simple scheme to avoid costly
1604  * operations, too frequent preemptions and too many dependencies on
1605  * the state of the scheduler: it requests the expiration of the
1606  * in-service queue (unconditionally) only for queues that need to
1607  * recover a hole. Then it delegates to other parts of the code the
1608  * responsibility of handling the above case 2.
1609  */
1610 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1611                                                 struct bfq_queue *bfqq,
1612                                                 bool arrived_in_time)
1613 {
1614         struct bfq_entity *entity = &bfqq->entity;
1615 
1616         /*
1617          * In the next compound condition, we check also whether there
1618          * is some budget left, because otherwise there is no point in
1619          * trying to go on serving bfqq with this same budget: bfqq
1620          * would be expired immediately after being selected for
1621          * service. This would only cause useless overhead.
1622          */
1623         if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1624             bfq_bfqq_budget_left(bfqq) > 0) {
1625                 /*
1626                  * We do not clear the flag non_blocking_wait_rq here, as
1627                  * the latter is used in bfq_activate_bfqq to signal
1628                  * that timestamps need to be back-shifted (and is
1629                  * cleared right after).
1630                  */
1631 
1632                 /*
1633                  * In next assignment we rely on that either
1634                  * entity->service or entity->budget are not updated
1635                  * on expiration if bfqq is empty (see
1636                  * __bfq_bfqq_recalc_budget). Thus both quantities
1637                  * remain unchanged after such an expiration, and the
1638                  * following statement therefore assigns to
1639                  * entity->budget the remaining budget on such an
1640                  * expiration.
1641                  */
1642                 entity->budget = min_t(unsigned long,
1643                                        bfq_bfqq_budget_left(bfqq),
1644                                        bfqq->max_budget);
1645 
1646                 /*
1647                  * At this point, we have used entity->service to get
1648                  * the budget left (needed for updating
1649                  * entity->budget). Thus we finally can, and have to,
1650                  * reset entity->service. The latter must be reset
1651                  * because bfqq would otherwise be charged again for
1652                  * the service it has received during its previous
1653                  * service slot(s).
1654                  */
1655                 entity->service = 0;
1656 
1657                 return true;
1658         }
1659 
1660         /*
1661          * We can finally complete expiration, by setting service to 0.
1662          */
1663         entity->service = 0;
1664         entity->budget = max_t(unsigned long, bfqq->max_budget,
1665                                bfq_serv_to_charge(bfqq->next_rq, bfqq));
1666         bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1667         return false;
1668 }
1669 
1670 /*
1671  * Return the farthest past time instant according to jiffies
1672  * macros.
1673  */
1674 static unsigned long bfq_smallest_from_now(void)
1675 {
1676         return jiffies - MAX_JIFFY_OFFSET;
1677 }
1678 
1679 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1680                                              struct bfq_queue *bfqq,
1681                                              unsigned int old_wr_coeff,
1682                                              bool wr_or_deserves_wr,
1683                                              bool interactive,
1684                                              bool in_burst,
1685                                              bool soft_rt)
1686 {
1687         if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1688                 /* start a weight-raising period */
1689                 if (interactive) {
1690                         bfqq->service_from_wr = 0;
1691                         bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1692                         bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1693                 } else {
1694                         /*
1695                          * No interactive weight raising in progress
1696                          * here: assign minus infinity to
1697                          * wr_start_at_switch_to_srt, to make sure
1698                          * that, at the end of the soft-real-time
1699                          * weight raising periods that is starting
1700                          * now, no interactive weight-raising period
1701                          * may be wrongly considered as still in
1702                          * progress (and thus actually started by
1703                          * mistake).
1704                          */
1705                         bfqq->wr_start_at_switch_to_srt =
1706                                 bfq_smallest_from_now();
1707                         bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1708                                 BFQ_SOFTRT_WEIGHT_FACTOR;
1709                         bfqq->wr_cur_max_time =
1710                                 bfqd->bfq_wr_rt_max_time;
1711                 }
1712 
1713                 /*
1714                  * If needed, further reduce budget to make sure it is
1715                  * close to bfqq's backlog, so as to reduce the
1716                  * scheduling-error component due to a too large
1717                  * budget. Do not care about throughput consequences,
1718                  * but only about latency. Finally, do not assign a
1719                  * too small budget either, to avoid increasing
1720                  * latency by causing too frequent expirations.
1721                  */
1722                 bfqq->entity.budget = min_t(unsigned long,
1723                                             bfqq->entity.budget,
1724                                             2 * bfq_min_budget(bfqd));
1725         } else if (old_wr_coeff > 1) {
1726                 if (interactive) { /* update wr coeff and duration */
1727                         bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1728                         bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1729                 } else if (in_burst)
1730                         bfqq->wr_coeff = 1;
1731                 else if (soft_rt) {
1732                         /*
1733                          * The application is now or still meeting the
1734                          * requirements for being deemed soft rt.  We
1735                          * can then correctly and safely (re)charge
1736                          * the weight-raising duration for the
1737                          * application with the weight-raising
1738                          * duration for soft rt applications.
1739                          *
1740                          * In particular, doing this recharge now, i.e.,
1741                          * before the weight-raising period for the
1742                          * application finishes, reduces the probability
1743                          * of the following negative scenario:
1744                          * 1) the weight of a soft rt application is
1745                          *    raised at startup (as for any newly
1746                          *    created application),
1747                          * 2) since the application is not interactive,
1748                          *    at a certain time weight-raising is
1749                          *    stopped for the application,
1750                          * 3) at that time the application happens to
1751                          *    still have pending requests, and hence
1752                          *    is destined to not have a chance to be
1753                          *    deemed soft rt before these requests are
1754                          *    completed (see the comments to the
1755                          *    function bfq_bfqq_softrt_next_start()
1756                          *    for details on soft rt detection),
1757                          * 4) these pending requests experience a high
1758                          *    latency because the application is not
1759                          *    weight-raised while they are pending.
1760                          */
1761                         if (bfqq->wr_cur_max_time !=
1762                                 bfqd->bfq_wr_rt_max_time) {
1763                                 bfqq->wr_start_at_switch_to_srt =
1764                                         bfqq->last_wr_start_finish;
1765 
1766                                 bfqq->wr_cur_max_time =
1767                                         bfqd->bfq_wr_rt_max_time;
1768                                 bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1769                                         BFQ_SOFTRT_WEIGHT_FACTOR;
1770                         }
1771                         bfqq->last_wr_start_finish = jiffies;
1772                 }
1773         }
1774 }
1775 
1776 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1777                                         struct bfq_queue *bfqq)
1778 {
1779         return bfqq->dispatched == 0 &&
1780                 time_is_before_jiffies(
1781                         bfqq->budget_timeout +
1782                         bfqd->bfq_wr_min_idle_time);
1783 }
1784 
1785 
1786 /*
1787  * Return true if bfqq is in a higher priority class, or has a higher
1788  * weight than the in-service queue.
1789  */
1790 static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq,
1791                                             struct bfq_queue *in_serv_bfqq)
1792 {
1793         int bfqq_weight, in_serv_weight;
1794 
1795         if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class)
1796                 return true;
1797 
1798         if (in_serv_bfqq->entity.parent == bfqq->entity.parent) {
1799                 bfqq_weight = bfqq->entity.weight;
1800                 in_serv_weight = in_serv_bfqq->entity.weight;
1801         } else {
1802                 if (bfqq->entity.parent)
1803                         bfqq_weight = bfqq->entity.parent->weight;
1804                 else
1805                         bfqq_weight = bfqq->entity.weight;
1806                 if (in_serv_bfqq->entity.parent)
1807                         in_serv_weight = in_serv_bfqq->entity.parent->weight;
1808                 else
1809                         in_serv_weight = in_serv_bfqq->entity.weight;
1810         }
1811 
1812         return bfqq_weight > in_serv_weight;
1813 }
1814 
1815 static bool bfq_better_to_idle(struct bfq_queue *bfqq);
1816 
1817 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1818                                              struct bfq_queue *bfqq,
1819                                              int old_wr_coeff,
1820                                              struct request *rq,
1821                                              bool *interactive)
1822 {
1823         bool soft_rt, in_burst, wr_or_deserves_wr,
1824                 bfqq_wants_to_preempt,
1825                 idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1826                 /*
1827                  * See the comments on
1828                  * bfq_bfqq_update_budg_for_activation for
1829                  * details on the usage of the next variable.
1830                  */
1831                 arrived_in_time =  ktime_get_ns() <=
1832                         bfqq->ttime.last_end_request +
1833                         bfqd->bfq_slice_idle * 3;
1834 
1835 
1836         /*
1837          * bfqq deserves to be weight-raised if:
1838          * - it is sync,
1839          * - it does not belong to a large burst,
1840          * - it has been idle for enough time or is soft real-time,
1841          * - is linked to a bfq_io_cq (it is not shared in any sense),
1842          * - has a default weight (otherwise we assume the user wanted
1843          *   to control its weight explicitly)
1844          */
1845         in_burst = bfq_bfqq_in_large_burst(bfqq);
1846         soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1847                 !BFQQ_TOTALLY_SEEKY(bfqq) &&
1848                 !in_burst &&
1849                 time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1850                 bfqq->dispatched == 0 &&
1851                 bfqq->entity.new_weight == 40;
1852         *interactive = !in_burst && idle_for_long_time &&
1853                 bfqq->entity.new_weight == 40;
1854         /*
1855          * Merged bfq_queues are kept out of weight-raising
1856          * (low-latency) mechanisms. The reason is that these queues
1857          * are usually created for non-interactive and
1858          * non-soft-real-time tasks. Yet this is not the case for
1859          * stably-merged queues. These queues are merged just because
1860          * they are created shortly after each other. So they may
1861          * easily serve the I/O of an interactive or soft-real time
1862          * application, if the application happens to spawn multiple
1863          * processes. So let also stably-merged queued enjoy weight
1864          * raising.
1865          */
1866         wr_or_deserves_wr = bfqd->low_latency &&
1867                 (bfqq->wr_coeff > 1 ||
1868                  (bfq_bfqq_sync(bfqq) &&
1869                   (bfqq->bic || RQ_BIC(rq)->stably_merged) &&
1870                    (*interactive || soft_rt)));
1871 
1872         /*
1873          * Using the last flag, update budget and check whether bfqq
1874          * may want to preempt the in-service queue.
1875          */
1876         bfqq_wants_to_preempt =
1877                 bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1878                                                     arrived_in_time);
1879 
1880         /*
1881          * If bfqq happened to be activated in a burst, but has been
1882          * idle for much more than an interactive queue, then we
1883          * assume that, in the overall I/O initiated in the burst, the
1884          * I/O associated with bfqq is finished. So bfqq does not need
1885          * to be treated as a queue belonging to a burst
1886          * anymore. Accordingly, we reset bfqq's in_large_burst flag
1887          * if set, and remove bfqq from the burst list if it's
1888          * there. We do not decrement burst_size, because the fact
1889          * that bfqq does not need to belong to the burst list any
1890          * more does not invalidate the fact that bfqq was created in
1891          * a burst.
1892          */
1893         if (likely(!bfq_bfqq_just_created(bfqq)) &&
1894             idle_for_long_time &&
1895             time_is_before_jiffies(
1896                     bfqq->budget_timeout +
1897                     msecs_to_jiffies(10000))) {
1898                 hlist_del_init(&bfqq->burst_list_node);
1899                 bfq_clear_bfqq_in_large_burst(bfqq);
1900         }
1901 
1902         bfq_clear_bfqq_just_created(bfqq);
1903 
1904         if (bfqd->low_latency) {
1905                 if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1906                         /* wraparound */
1907                         bfqq->split_time =
1908                                 jiffies - bfqd->bfq_wr_min_idle_time - 1;
1909 
1910                 if (time_is_before_jiffies(bfqq->split_time +
1911                                            bfqd->bfq_wr_min_idle_time)) {
1912                         bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1913                                                          old_wr_coeff,
1914                                                          wr_or_deserves_wr,
1915                                                          *interactive,
1916                                                          in_burst,
1917                                                          soft_rt);
1918 
1919                         if (old_wr_coeff != bfqq->wr_coeff)
1920                                 bfqq->entity.prio_changed = 1;
1921                 }
1922         }
1923 
1924         bfqq->last_idle_bklogged = jiffies;
1925         bfqq->service_from_backlogged = 0;
1926         bfq_clear_bfqq_softrt_update(bfqq);
1927 
1928         bfq_add_bfqq_busy(bfqd, bfqq);
1929 
1930         /*
1931          * Expire in-service queue if preemption may be needed for
1932          * guarantees or throughput. As for guarantees, we care
1933          * explicitly about two cases. The first is that bfqq has to
1934          * recover a service hole, as explained in the comments on
1935          * bfq_bfqq_update_budg_for_activation(), i.e., that
1936          * bfqq_wants_to_preempt is true. However, if bfqq does not
1937          * carry time-critical I/O, then bfqq's bandwidth is less
1938          * important than that of queues that carry time-critical I/O.
1939          * So, as a further constraint, we consider this case only if
1940          * bfqq is at least as weight-raised, i.e., at least as time
1941          * critical, as the in-service queue.
1942          *
1943          * The second case is that bfqq is in a higher priority class,
1944          * or has a higher weight than the in-service queue. If this
1945          * condition does not hold, we don't care because, even if
1946          * bfqq does not start to be served immediately, the resulting
1947          * delay for bfqq's I/O is however lower or much lower than
1948          * the ideal completion time to be guaranteed to bfqq's I/O.
1949          *
1950          * In both cases, preemption is needed only if, according to
1951          * the timestamps of both bfqq and of the in-service queue,
1952          * bfqq actually is the next queue to serve. So, to reduce
1953          * useless preemptions, the return value of
1954          * next_queue_may_preempt() is considered in the next compound
1955          * condition too. Yet next_queue_may_preempt() just checks a
1956          * simple, necessary condition for bfqq to be the next queue
1957          * to serve. In fact, to evaluate a sufficient condition, the
1958          * timestamps of the in-service queue would need to be
1959          * updated, and this operation is quite costly (see the
1960          * comments on bfq_bfqq_update_budg_for_activation()).
1961          *
1962          * As for throughput, we ask bfq_better_to_idle() whether we
1963          * still need to plug I/O dispatching. If bfq_better_to_idle()
1964          * says no, then plugging is not needed any longer, either to
1965          * boost throughput or to perserve service guarantees. Then
1966          * the best option is to stop plugging I/O, as not doing so
1967          * would certainly lower throughput. We may end up in this
1968          * case if: (1) upon a dispatch attempt, we detected that it
1969          * was better to plug I/O dispatch, and to wait for a new
1970          * request to arrive for the currently in-service queue, but
1971          * (2) this switch of bfqq to busy changes the scenario.
1972          */
1973         if (bfqd->in_service_queue &&
1974             ((bfqq_wants_to_preempt &&
1975               bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) ||
1976              bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue) ||
1977              !bfq_better_to_idle(bfqd->in_service_queue)) &&
1978             next_queue_may_preempt(bfqd))
1979                 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1980                                 false, BFQQE_PREEMPTED);
1981 }
1982 
1983 static void bfq_reset_inject_limit(struct bfq_data *bfqd,
1984                                    struct bfq_queue *bfqq)
1985 {
1986         /* invalidate baseline total service time */
1987         bfqq->last_serv_time_ns = 0;
1988 
1989         /*
1990          * Reset pointer in case we are waiting for
1991          * some request completion.
1992          */
1993         bfqd->waited_rq = NULL;
1994 
1995         /*
1996          * If bfqq has a short think time, then start by setting the
1997          * inject limit to 0 prudentially, because the service time of
1998          * an injected I/O request may be higher than the think time
1999          * of bfqq, and therefore, if one request was injected when
2000          * bfqq remains empty, this injected request might delay the
2001          * service of the next I/O request for bfqq significantly. In
2002          * case bfqq can actually tolerate some injection, then the
2003          * adaptive update will however raise the limit soon. This
2004          * lucky circumstance holds exactly because bfqq has a short
2005          * think time, and thus, after remaining empty, is likely to
2006          * get new I/O enqueued---and then completed---before being
2007          * expired. This is the very pattern that gives the
2008          * limit-update algorithm the chance to measure the effect of
2009          * injection on request service times, and then to update the
2010          * limit accordingly.
2011          *
2012          * However, in the following special case, the inject limit is
2013          * left to 1 even if the think time is short: bfqq's I/O is
2014          * synchronized with that of some other queue, i.e., bfqq may
2015          * receive new I/O only after the I/O of the other queue is
2016          * completed. Keeping the inject limit to 1 allows the
2017          * blocking I/O to be served while bfqq is in service. And
2018          * this is very convenient both for bfqq and for overall
2019          * throughput, as explained in detail in the comments in
2020          * bfq_update_has_short_ttime().
2021          *
2022          * On the opposite end, if bfqq has a long think time, then
2023          * start directly by 1, because:
2024          * a) on the bright side, keeping at most one request in
2025          * service in the drive is unlikely to cause any harm to the
2026          * latency of bfqq's requests, as the service time of a single
2027          * request is likely to be lower than the think time of bfqq;
2028          * b) on the downside, after becoming empty, bfqq is likely to
2029          * expire before getting its next request. With this request
2030          * arrival pattern, it is very hard to sample total service
2031          * times and update the inject limit accordingly (see comments
2032          * on bfq_update_inject_limit()). So the limit is likely to be
2033          * never, or at least seldom, updated.  As a consequence, by
2034          * setting the limit to 1, we avoid that no injection ever
2035          * occurs with bfqq. On the downside, this proactive step
2036          * further reduces chances to actually compute the baseline
2037          * total service time. Thus it reduces chances to execute the
2038          * limit-update algorithm and possibly raise the limit to more
2039          * than 1.
2040          */
2041         if (bfq_bfqq_has_short_ttime(bfqq))
2042                 bfqq->inject_limit = 0;
2043         else
2044                 bfqq->inject_limit = 1;
2045 
2046         bfqq->decrease_time_jif = jiffies;
2047 }
2048 
2049 static void bfq_update_io_intensity(struct bfq_queue *bfqq, u64 now_ns)
2050 {
2051         u64 tot_io_time = now_ns - bfqq->io_start_time;
2052 
2053         if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfqq->dispatched == 0)
2054                 bfqq->tot_idle_time +=
2055                         now_ns - bfqq->ttime.last_end_request;
2056 
2057         if (unlikely(bfq_bfqq_just_created(bfqq)))
2058                 return;
2059 
2060         /*
2061          * Must be busy for at least about 80% of the time to be
2062          * considered I/O bound.
2063          */
2064         if (bfqq->tot_idle_time * 5 > tot_io_time)
2065                 bfq_clear_bfqq_IO_bound(bfqq);
2066         else
2067                 bfq_mark_bfqq_IO_bound(bfqq);
2068 
2069         /*
2070          * Keep an observation window of at most 200 ms in the past
2071          * from now.
2072          */
2073         if (tot_io_time > 200 * NSEC_PER_MSEC) {
2074                 bfqq->io_start_time = now_ns - (tot_io_time>>1);
2075                 bfqq->tot_idle_time >>= 1;
2076         }
2077 }
2078 
2079 /*
2080  * Detect whether bfqq's I/O seems synchronized with that of some
2081  * other queue, i.e., whether bfqq, after remaining empty, happens to
2082  * receive new I/O only right after some I/O request of the other
2083  * queue has been completed. We call waker queue the other queue, and
2084  * we assume, for simplicity, that bfqq may have at most one waker
2085  * queue.
2086  *
2087  * A remarkable throughput boost can be reached by unconditionally
2088  * injecting the I/O of the waker queue, every time a new
2089  * bfq_dispatch_request happens to be invoked while I/O is being
2090  * plugged for bfqq.  In addition to boosting throughput, this
2091  * unblocks bfqq's I/O, thereby improving bandwidth and latency for
2092  * bfqq. Note that these same results may be achieved with the general
2093  * injection mechanism, but less effectively. For details on this
2094  * aspect, see the comments on the choice of the queue for injection
2095  * in bfq_select_queue().
2096  *
2097  * Turning back to the detection of a waker queue, a queue Q is deemed as a
2098  * waker queue for bfqq if, for three consecutive times, bfqq happens to become
2099  * non empty right after a request of Q has been completed within given
2100  * timeout. In this respect, even if bfqq is empty, we do not check for a waker
2101  * if it still has some in-flight I/O. In fact, in this case bfqq is actually
2102  * still being served by the drive, and may receive new I/O on the completion
2103  * of some of the in-flight requests. In particular, on the first time, Q is
2104  * tentatively set as a candidate waker queue, while on the third consecutive
2105  * time that Q is detected, the field waker_bfqq is set to Q, to confirm that Q
2106  * is a waker queue for bfqq. These detection steps are performed only if bfqq
2107  * has a long think time, so as to make it more likely that bfqq's I/O is
2108  * actually being blocked by a synchronization. This last filter, plus the
2109  * above three-times requirement and time limit for detection, make false
2110  * positives less likely.
2111  *
2112  * NOTE
2113  *
2114  * The sooner a waker queue is detected, the sooner throughput can be
2115  * boosted by injecting I/O from the waker queue. Fortunately,
2116  * detection is likely to be actually fast, for the following
2117  * reasons. While blocked by synchronization, bfqq has a long think
2118  * time. This implies that bfqq's inject limit is at least equal to 1
2119  * (see the comments in bfq_update_inject_limit()). So, thanks to
2120  * injection, the waker queue is likely to be served during the very
2121  * first I/O-plugging time interval for bfqq. This triggers the first
2122  * step of the detection mechanism. Thanks again to injection, the
2123  * candidate waker queue is then likely to be confirmed no later than
2124  * during the next I/O-plugging interval for bfqq.
2125  *
2126  * ISSUE
2127  *
2128  * On queue merging all waker information is lost.
2129  */
2130 static void bfq_check_waker(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2131                             u64 now_ns)
2132 {
2133         char waker_name[MAX_BFQQ_NAME_LENGTH];
2134 
2135         if (!bfqd->last_completed_rq_bfqq ||
2136             bfqd->last_completed_rq_bfqq == bfqq ||
2137             bfq_bfqq_has_short_ttime(bfqq) ||
2138             now_ns - bfqd->last_completion >= 4 * NSEC_PER_MSEC)
2139                 return;
2140 
2141         /*
2142          * We reset waker detection logic also if too much time has passed
2143          * since the first detection. If wakeups are rare, pointless idling
2144          * doesn't hurt throughput that much. The condition below makes sure
2145          * we do not uselessly idle blocking waker in more than 1/64 cases. 
2146          */
2147         if (bfqd->last_completed_rq_bfqq !=
2148             bfqq->tentative_waker_bfqq ||
2149             now_ns > bfqq->waker_detection_started +
2150                                         128 * (u64)bfqd->bfq_slice_idle) {
2151                 /*
2152                  * First synchronization detected with a
2153                  * candidate waker queue, or with a different
2154                  * candidate waker queue from the current one.
2155                  */
2156                 bfqq->tentative_waker_bfqq =
2157                         bfqd->last_completed_rq_bfqq;
2158                 bfqq->num_waker_detections = 1;
2159                 bfqq->waker_detection_started = now_ns;
2160                 bfq_bfqq_name(bfqq->tentative_waker_bfqq, waker_name,
2161                               MAX_BFQQ_NAME_LENGTH);
2162                 bfq_log_bfqq(bfqd, bfqq, "set tentative waker %s", waker_name);
2163         } else /* Same tentative waker queue detected again */
2164                 bfqq->num_waker_detections++;
2165 
2166         if (bfqq->num_waker_detections == 3) {
2167                 bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
2168                 bfqq->tentative_waker_bfqq = NULL;
2169                 bfq_bfqq_name(bfqq->waker_bfqq, waker_name,
2170                               MAX_BFQQ_NAME_LENGTH);
2171                 bfq_log_bfqq(bfqd, bfqq, "set waker %s", waker_name);
2172 
2173                 /*
2174                  * If the waker queue disappears, then
2175                  * bfqq->waker_bfqq must be reset. To
2176                  * this goal, we maintain in each
2177                  * waker queue a list, woken_list, of
2178                  * all the queues that reference the
2179                  * waker queue through their
2180                  * waker_bfqq pointer. When the waker
2181                  * queue exits, the waker_bfqq pointer
2182                  * of all the queues in the woken_list
2183                  * is reset.
2184                  *
2185                  * In addition, if bfqq is already in
2186                  * the woken_list of a waker queue,
2187                  * then, before being inserted into
2188                  * the woken_list of a new waker
2189                  * queue, bfqq must be removed from
2190                  * the woken_list of the old waker
2191                  * queue.
2192                  */
2193                 if (!hlist_unhashed(&bfqq->woken_list_node))
2194                         hlist_del_init(&bfqq->woken_list_node);
2195                 hlist_add_head(&bfqq->woken_list_node,
2196                                &bfqd->last_completed_rq_bfqq->woken_list);
2197         }
2198 }
2199 
2200 static void bfq_add_request(struct request *rq)
2201 {
2202         struct bfq_queue *bfqq = RQ_BFQQ(rq);
2203         struct bfq_data *bfqd = bfqq->bfqd;
2204         struct request *next_rq, *prev;
2205         unsigned int old_wr_coeff = bfqq->wr_coeff;
2206         bool interactive = false;
2207         u64 now_ns = ktime_get_ns();
2208 
2209         bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
2210         bfqq->queued[rq_is_sync(rq)]++;
2211         /*
2212          * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2213          * may be read without holding the lock in bfq_has_work().
2214          */
2215         WRITE_ONCE(bfqd->queued, bfqd->queued + 1);
2216 
2217         if (bfq_bfqq_sync(bfqq) && RQ_BIC(rq)->requests <= 1) {
2218                 bfq_check_waker(bfqd, bfqq, now_ns);
2219 
2220                 /*
2221                  * Periodically reset inject limit, to make sure that
2222                  * the latter eventually drops in case workload
2223                  * changes, see step (3) in the comments on
2224                  * bfq_update_inject_limit().
2225                  */
2226                 if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2227                                              msecs_to_jiffies(1000)))
2228                         bfq_reset_inject_limit(bfqd, bfqq);
2229 
2230                 /*
2231                  * The following conditions must hold to setup a new
2232                  * sampling of total service time, and then a new
2233                  * update of the inject limit:
2234                  * - bfqq is in service, because the total service
2235                  *   time is evaluated only for the I/O requests of
2236                  *   the queues in service;
2237                  * - this is the right occasion to compute or to
2238                  *   lower the baseline total service time, because
2239                  *   there are actually no requests in the drive,
2240                  *   or
2241                  *   the baseline total service time is available, and
2242                  *   this is the right occasion to compute the other
2243                  *   quantity needed to update the inject limit, i.e.,
2244                  *   the total service time caused by the amount of
2245                  *   injection allowed by the current value of the
2246                  *   limit. It is the right occasion because injection
2247                  *   has actually been performed during the service
2248                  *   hole, and there are still in-flight requests,
2249                  *   which are very likely to be exactly the injected
2250                  *   requests, or part of them;
2251                  * - the minimum interval for sampling the total
2252                  *   service time and updating the inject limit has
2253                  *   elapsed.
2254                  */
2255                 if (bfqq == bfqd->in_service_queue &&
2256                     (bfqd->rq_in_driver == 0 ||
2257                      (bfqq->last_serv_time_ns > 0 &&
2258                       bfqd->rqs_injected && bfqd->rq_in_driver > 0)) &&
2259                     time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2260                                               msecs_to_jiffies(10))) {
2261                         bfqd->last_empty_occupied_ns = ktime_get_ns();
2262                         /*
2263                          * Start the state machine for measuring the
2264                          * total service time of rq: setting
2265                          * wait_dispatch will cause bfqd->waited_rq to
2266                          * be set when rq will be dispatched.
2267                          */
2268                         bfqd->wait_dispatch = true;
2269                         /*
2270                          * If there is no I/O in service in the drive,
2271                          * then possible injection occurred before the
2272                          * arrival of rq will not affect the total
2273                          * service time of rq. So the injection limit
2274                          * must not be updated as a function of such
2275                          * total service time, unless new injection
2276                          * occurs before rq is completed. To have the
2277                          * injection limit updated only in the latter
2278                          * case, reset rqs_injected here (rqs_injected
2279                          * will be set in case injection is performed
2280                          * on bfqq before rq is completed).
2281                          */
2282                         if (bfqd->rq_in_driver == 0)
2283                                 bfqd->rqs_injected = false;
2284                 }
2285         }
2286 
2287         if (bfq_bfqq_sync(bfqq))
2288                 bfq_update_io_intensity(bfqq, now_ns);
2289 
2290         elv_rb_add(&bfqq->sort_list, rq);
2291 
2292         /*
2293          * Check if this request is a better next-serve candidate.
2294          */
2295         prev = bfqq->next_rq;
2296         next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
2297         bfqq->next_rq = next_rq;
2298 
2299         /*
2300          * Adjust priority tree position, if next_rq changes.
2301          * See comments on bfq_pos_tree_add_move() for the unlikely().
2302          */
2303         if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
2304                 bfq_pos_tree_add_move(bfqd, bfqq);
2305 
2306         if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
2307                 bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
2308                                                  rq, &interactive);
2309         else {
2310                 if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
2311                     time_is_before_jiffies(
2312                                 bfqq->last_wr_start_finish +
2313                                 bfqd->bfq_wr_min_inter_arr_async)) {
2314                         bfqq->wr_coeff = bfqd->bfq_wr_coeff;
2315                         bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
2316 
2317                         bfqd->wr_busy_queues++;
2318                         bfqq->entity.prio_changed = 1;
2319                 }
2320                 if (prev != bfqq->next_rq)
2321                         bfq_updated_next_req(bfqd, bfqq);
2322         }
2323 
2324         /*
2325          * Assign jiffies to last_wr_start_finish in the following
2326          * cases:
2327          *
2328          * . if bfqq is not going to be weight-raised, because, for
2329          *   non weight-raised queues, last_wr_start_finish stores the
2330          *   arrival time of the last request; as of now, this piece
2331          *   of information is used only for deciding whether to
2332          *   weight-raise async queues
2333          *
2334          * . if bfqq is not weight-raised, because, if bfqq is now
2335          *   switching to weight-raised, then last_wr_start_finish
2336          *   stores the time when weight-raising starts
2337          *
2338          * . if bfqq is interactive, because, regardless of whether
2339          *   bfqq is currently weight-raised, the weight-raising
2340          *   period must start or restart (this case is considered
2341          *   separately because it is not detected by the above
2342          *   conditions, if bfqq is already weight-raised)
2343          *
2344          * last_wr_start_finish has to be updated also if bfqq is soft
2345          * real-time, because the weight-raising period is constantly
2346          * restarted on idle-to-busy transitions for these queues, but
2347          * this is already done in bfq_bfqq_handle_idle_busy_switch if
2348          * needed.
2349          */
2350         if (bfqd->low_latency &&
2351                 (old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
2352                 bfqq->last_wr_start_finish = jiffies;
2353 }
2354 
2355 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
2356                                           struct bio *bio,
2357                                           struct request_queue *q)
2358 {
2359         struct bfq_queue *bfqq = bfqd->bio_bfqq;
2360 
2361 
2362         if (bfqq)
2363                 return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
2364 
2365         return NULL;
2366 }
2367 
2368 static sector_t get_sdist(sector_t last_pos, struct request *rq)
2369 {
2370         if (last_pos)
2371                 return abs(blk_rq_pos(rq) - last_pos);
2372 
2373         return 0;
2374 }
2375 
2376 #if 0 /* Still not clear if we can do without next two functions */
2377 static void bfq_activate_request(struct request_queue *q, struct request *rq)
2378 {
2379         struct bfq_data *bfqd = q->elevator->elevator_data;
2380 
2381         bfqd->rq_in_driver++;
2382 }
2383 
2384 static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
2385 {
2386         struct bfq_data *bfqd = q->elevator->elevator_data;
2387 
2388         bfqd->rq_in_driver--;
2389 }
2390 #endif
2391 
2392 static void bfq_remove_request(struct request_queue *q,
2393                                struct request *rq)
2394 {
2395         struct bfq_queue *bfqq = RQ_BFQQ(rq);
2396         struct bfq_data *bfqd = bfqq->bfqd;
2397         const int sync = rq_is_sync(rq);
2398 
2399         if (bfqq->next_rq == rq) {
2400                 bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
2401                 bfq_updated_next_req(bfqd, bfqq);
2402         }
2403 
2404         if (rq->queuelist.prev != &rq->queuelist)
2405                 list_del_init(&rq->queuelist);
2406         bfqq->queued[sync]--;
2407         /*
2408          * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2409          * may be read without holding the lock in bfq_has_work().
2410          */
2411         WRITE_ONCE(bfqd->queued, bfqd->queued - 1);
2412         elv_rb_del(&bfqq->sort_list, rq);
2413 
2414         elv_rqhash_del(q, rq);
2415         if (q->last_merge == rq)
2416                 q->last_merge = NULL;
2417 
2418         if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2419                 bfqq->next_rq = NULL;
2420 
2421                 if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
2422                         bfq_del_bfqq_busy(bfqd, bfqq, false);
2423                         /*
2424                          * bfqq emptied. In normal operation, when
2425                          * bfqq is empty, bfqq->entity.service and
2426                          * bfqq->entity.budget must contain,
2427                          * respectively, the service received and the
2428                          * budget used last time bfqq emptied. These
2429                          * facts do not hold in this case, as at least
2430                          * this last removal occurred while bfqq is
2431                          * not in service. To avoid inconsistencies,
2432                          * reset both bfqq->entity.service and
2433                          * bfqq->entity.budget, if bfqq has still a
2434                          * process that may issue I/O requests to it.
2435                          */
2436                         bfqq->entity.budget = bfqq->entity.service = 0;
2437                 }
2438 
2439                 /*
2440                  * Remove queue from request-position tree as it is empty.
2441                  */
2442                 if (bfqq->pos_root) {
2443                         rb_erase(&bfqq->pos_node, bfqq->pos_root);
2444                         bfqq->pos_root = NULL;
2445                 }
2446         } else {
2447                 /* see comments on bfq_pos_tree_add_move() for the unlikely() */
2448                 if (unlikely(!bfqd->nonrot_with_queueing))
2449                         bfq_pos_tree_add_move(bfqd, bfqq);
2450         }
2451 
2452         if (rq->cmd_flags & REQ_META)
2453                 bfqq->meta_pending--;
2454 
2455 }
2456 
2457 static bool bfq_bio_merge(struct request_queue *q, struct bio *bio,
2458                 unsigned int nr_segs)
2459 {
2460         struct bfq_data *bfqd = q->elevator->elevator_data;
2461         struct request *free = NULL;
2462         /*
2463          * bfq_bic_lookup grabs the queue_lock: invoke it now and
2464          * store its return value for later use, to avoid nesting
2465          * queue_lock inside the bfqd->lock. We assume that the bic
2466          * returned by bfq_bic_lookup does not go away before
2467          * bfqd->lock is taken.
2468          */
2469         struct bfq_io_cq *bic = bfq_bic_lookup(q);
2470         bool ret;
2471 
2472         spin_lock_irq(&bfqd->lock);
2473 
2474         if (bic) {
2475                 /*
2476                  * Make sure cgroup info is uptodate for current process before
2477                  * considering the merge.
2478                  */
2479                 bfq_bic_update_cgroup(bic, bio);
2480 
2481                 bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
2482         } else {
2483                 bfqd->bio_bfqq = NULL;
2484         }
2485         bfqd->bio_bic = bic;
2486 
2487         ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
2488 
2489         spin_unlock_irq(&bfqd->lock);
2490         if (free)
2491                 blk_mq_free_request(free);
2492 
2493         return ret;
2494 }
2495 
2496 static int bfq_request_merge(struct request_queue *q, struct request **req,
2497                              struct bio *bio)
2498 {
2499         struct bfq_data *bfqd = q->elevator->elevator_data;
2500         struct request *__rq;
2501 
2502         __rq = bfq_find_rq_fmerge(bfqd, bio, q);
2503         if (__rq && elv_bio_merge_ok(__rq, bio)) {
2504                 *req = __rq;
2505 
2506                 if (blk_discard_mergable(__rq))
2507                         return ELEVATOR_DISCARD_MERGE;
2508                 return ELEVATOR_FRONT_MERGE;
2509         }
2510 
2511         return ELEVATOR_NO_MERGE;
2512 }
2513 
2514 static void bfq_request_merged(struct request_queue *q, struct request *req,
2515                                enum elv_merge type)
2516 {
2517         if (type == ELEVATOR_FRONT_MERGE &&
2518             rb_prev(&req->rb_node) &&
2519             blk_rq_pos(req) <
2520             blk_rq_pos(container_of(rb_prev(&req->rb_node),
2521                                     struct request, rb_node))) {
2522                 struct bfq_queue *bfqq = RQ_BFQQ(req);
2523                 struct bfq_data *bfqd;
2524                 struct request *prev, *next_rq;
2525 
2526                 if (!bfqq)
2527                         return;
2528 
2529                 bfqd = bfqq->bfqd;
2530 
2531                 /* Reposition request in its sort_list */
2532                 elv_rb_del(&bfqq->sort_list, req);
2533                 elv_rb_add(&bfqq->sort_list, req);
2534 
2535                 /* Choose next request to be served for bfqq */
2536                 prev = bfqq->next_rq;
2537                 next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
2538                                          bfqd->last_position);
2539                 bfqq->next_rq = next_rq;
2540                 /*
2541                  * If next_rq changes, update both the queue's budget to
2542                  * fit the new request and the queue's position in its
2543                  * rq_pos_tree.
2544                  */
2545                 if (prev != bfqq->next_rq) {
2546                         bfq_updated_next_req(bfqd, bfqq);
2547                         /*
2548                          * See comments on bfq_pos_tree_add_move() for
2549                          * the unlikely().
2550                          */
2551                         if (unlikely(!bfqd->nonrot_with_queueing))
2552                                 bfq_pos_tree_add_move(bfqd, bfqq);
2553                 }
2554         }
2555 }
2556 
2557 /*
2558  * This function is called to notify the scheduler that the requests
2559  * rq and 'next' have been merged, with 'next' going away.  BFQ
2560  * exploits this hook to address the following issue: if 'next' has a
2561  * fifo_time lower that rq, then the fifo_time of rq must be set to
2562  * the value of 'next', to not forget the greater age of 'next'.
2563  *
2564  * NOTE: in this function we assume that rq is in a bfq_queue, basing
2565  * on that rq is picked from the hash table q->elevator->hash, which,
2566  * in its turn, is filled only with I/O requests present in
2567  * bfq_queues, while BFQ is in use for the request queue q. In fact,
2568  * the function that fills this hash table (elv_rqhash_add) is called
2569  * only by bfq_insert_request.
2570  */
2571 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
2572                                 struct request *next)
2573 {
2574         struct bfq_queue *bfqq = RQ_BFQQ(rq),
2575                 *next_bfqq = RQ_BFQQ(next);
2576 
2577         if (!bfqq)
2578                 goto remove;
2579 
2580         /*
2581          * If next and rq belong to the same bfq_queue and next is older
2582          * than rq, then reposition rq in the fifo (by substituting next
2583          * with rq). Otherwise, if next and rq belong to different
2584          * bfq_queues, never reposition rq: in fact, we would have to
2585          * reposition it with respect to next's position in its own fifo,
2586          * which would most certainly be too expensive with respect to
2587          * the benefits.
2588          */
2589         if (bfqq == next_bfqq &&
2590             !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2591             next->fifo_time < rq->fifo_time) {
2592                 list_del_init(&rq->queuelist);
2593                 list_replace_init(&next->queuelist, &rq->queuelist);
2594                 rq->fifo_time = next->fifo_time;
2595         }
2596 
2597         if (bfqq->next_rq == next)
2598                 bfqq->next_rq = rq;
2599 
2600         bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
2601 remove:
2602         /* Merged request may be in the IO scheduler. Remove it. */
2603         if (!RB_EMPTY_NODE(&next->rb_node)) {
2604                 bfq_remove_request(next->q, next);
2605                 if (next_bfqq)
2606                         bfqg_stats_update_io_remove(bfqq_group(next_bfqq),
2607                                                     next->cmd_flags);
2608         }
2609 }
2610 
2611 /* Must be called with bfqq != NULL */
2612 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
2613 {
2614         /*
2615          * If bfqq has been enjoying interactive weight-raising, then
2616          * reset soft_rt_next_start. We do it for the following
2617          * reason. bfqq may have been conveying the I/O needed to load
2618          * a soft real-time application. Such an application actually
2619          * exhibits a soft real-time I/O pattern after it finishes
2620          * loading, and finally starts doing its job. But, if bfqq has
2621          * been receiving a lot of bandwidth so far (likely to happen
2622          * on a fast device), then soft_rt_next_start now contains a
2623          * high value that. So, without this reset, bfqq would be
2624          * prevented from being possibly considered as soft_rt for a
2625          * very long time.
2626          */
2627 
2628         if (bfqq->wr_cur_max_time !=
2629             bfqq->bfqd->bfq_wr_rt_max_time)
2630                 bfqq->soft_rt_next_start = jiffies;
2631 
2632         if (bfq_bfqq_busy(bfqq))
2633                 bfqq->bfqd->wr_busy_queues--;
2634         bfqq->wr_coeff = 1;
2635         bfqq->wr_cur_max_time = 0;
2636         bfqq->last_wr_start_finish = jiffies;
2637         /*
2638          * Trigger a weight change on the next invocation of
2639          * __bfq_entity_update_weight_prio.
2640          */
2641         bfqq->entity.prio_changed = 1;
2642 }
2643 
2644 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2645                              struct bfq_group *bfqg)
2646 {
2647         int i, j;
2648 
2649         for (i = 0; i < 2; i++)
2650                 for (j = 0; j < IOPRIO_NR_LEVELS; j++)
2651                         if (bfqg->async_bfqq[i][j])
2652                                 bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2653         if (bfqg->async_idle_bfqq)
2654                 bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2655 }
2656 
2657 static void bfq_end_wr(struct bfq_data *bfqd)
2658 {
2659         struct bfq_queue *bfqq;
2660 
2661         spin_lock_irq(&bfqd->lock);
2662 
2663         list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2664                 bfq_bfqq_end_wr(bfqq);
2665         list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2666                 bfq_bfqq_end_wr(bfqq);
2667         bfq_end_wr_async(bfqd);
2668 
2669         spin_unlock_irq(&bfqd->lock);
2670 }
2671 
2672 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2673 {
2674         if (request)
2675                 return blk_rq_pos(io_struct);
2676         else
2677                 return ((struct bio *)io_struct)->bi_iter.bi_sector;
2678 }
2679 
2680 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2681                                   sector_t sector)
2682 {
2683         return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2684                BFQQ_CLOSE_THR;
2685 }
2686 
2687 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2688                                          struct bfq_queue *bfqq,
2689                                          sector_t sector)
2690 {
2691         struct rb_root *root = &bfqq_group(bfqq)->rq_pos_tree;
2692         struct rb_node *parent, *node;
2693         struct bfq_queue *__bfqq;
2694 
2695         if (RB_EMPTY_ROOT(root))
2696                 return NULL;
2697 
2698         /*
2699          * First, if we find a request starting at the end of the last
2700          * request, choose it.
2701          */
2702         __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2703         if (__bfqq)
2704                 return __bfqq;
2705 
2706         /*
2707          * If the exact sector wasn't found, the parent of the NULL leaf
2708          * will contain the closest sector (rq_pos_tree sorted by
2709          * next_request position).
2710          */
2711         __bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2712         if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2713                 return __bfqq;
2714 
2715         if (blk_rq_pos(__bfqq->next_rq) < sector)
2716                 node = rb_next(&__bfqq->pos_node);
2717         else
2718                 node = rb_prev(&__bfqq->pos_node);
2719         if (!node)
2720                 return NULL;
2721 
2722         __bfqq = rb_entry(node, struct bfq_queue, pos_node);
2723         if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2724                 return __bfqq;
2725 
2726         return NULL;
2727 }
2728 
2729 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2730                                                    struct bfq_queue *cur_bfqq,
2731                                                    sector_t sector)
2732 {
2733         struct bfq_queue *bfqq;
2734 
2735         /*
2736          * We shall notice if some of the queues are cooperating,
2737          * e.g., working closely on the same area of the device. In
2738          * that case, we can group them together and: 1) don't waste
2739          * time idling, and 2) serve the union of their requests in
2740          * the best possible order for throughput.
2741          */
2742         bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2743         if (!bfqq || bfqq == cur_bfqq)
2744                 return NULL;
2745 
2746         return bfqq;
2747 }
2748 
2749 static struct bfq_queue *
2750 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2751 {
2752         int process_refs, new_process_refs;
2753         struct bfq_queue *__bfqq;
2754 
2755         /*
2756          * If there are no process references on the new_bfqq, then it is
2757          * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2758          * may have dropped their last reference (not just their last process
2759          * reference).
2760          */
2761         if (!bfqq_process_refs(new_bfqq))
2762                 return NULL;
2763 
2764         /* Avoid a circular list and skip interim queue merges. */
2765         while ((__bfqq = new_bfqq->new_bfqq)) {
2766                 if (__bfqq == bfqq)
2767                         return NULL;
2768                 new_bfqq = __bfqq;
2769         }
2770 
2771         process_refs = bfqq_process_refs(bfqq);
2772         new_process_refs = bfqq_process_refs(new_bfqq);
2773         /*
2774          * If the process for the bfqq has gone away, there is no
2775          * sense in merging the queues.
2776          */
2777         if (process_refs == 0 || new_process_refs == 0)
2778                 return NULL;
2779 
2780         /*
2781          * Make sure merged queues belong to the same parent. Parents could
2782          * have changed since the time we decided the two queues are suitable
2783          * for merging.
2784          */
2785         if (new_bfqq->entity.parent != bfqq->entity.parent)
2786                 return NULL;
2787 
2788         bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2789                 new_bfqq->pid);
2790 
2791         /*
2792          * Merging is just a redirection: the requests of the process
2793          * owning one of the two queues are redirected to the other queue.
2794          * The latter queue, in its turn, is set as shared if this is the
2795          * first time that the requests of some process are redirected to
2796          * it.
2797          *
2798          * We redirect bfqq to new_bfqq and not the opposite, because
2799          * we are in the context of the process owning bfqq, thus we
2800          * have the io_cq of this process. So we can immediately
2801          * configure this io_cq to redirect the requests of the
2802          * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2803          * not available any more (new_bfqq->bic == NULL).
2804          *
2805          * Anyway, even in case new_bfqq coincides with the in-service
2806          * queue, redirecting requests the in-service queue is the
2807          * best option, as we feed the in-service queue with new
2808          * requests close to the last request served and, by doing so,
2809          * are likely to increase the throughput.
2810          */
2811         bfqq->new_bfqq = new_bfqq;
2812         /*
2813          * The above assignment schedules the following redirections:
2814          * each time some I/O for bfqq arrives, the process that
2815          * generated that I/O is disassociated from bfqq and
2816          * associated with new_bfqq. Here we increases new_bfqq->ref
2817          * in advance, adding the number of processes that are
2818          * expected to be associated with new_bfqq as they happen to
2819          * issue I/O.
2820          */
2821         new_bfqq->ref += process_refs;
2822         return new_bfqq;
2823 }
2824 
2825 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2826                                         struct bfq_queue *new_bfqq)
2827 {
2828         if (bfq_too_late_for_merging(new_bfqq))
2829                 return false;
2830 
2831         if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2832             (bfqq->ioprio_class != new_bfqq->ioprio_class))
2833                 return false;
2834 
2835         /*
2836          * If either of the queues has already been detected as seeky,
2837          * then merging it with the other queue is unlikely to lead to
2838          * sequential I/O.
2839          */
2840         if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2841                 return false;
2842 
2843         /*
2844          * Interleaved I/O is known to be done by (some) applications
2845          * only for reads, so it does not make sense to merge async
2846          * queues.
2847          */
2848         if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2849                 return false;
2850 
2851         return true;
2852 }
2853 
2854 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
2855                                              struct bfq_queue *bfqq);
2856 
2857 /*
2858  * Attempt to schedule a merge of bfqq with the currently in-service
2859  * queue or with a close queue among the scheduled queues.  Return
2860  * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2861  * structure otherwise.
2862  *
2863  * The OOM queue is not allowed to participate to cooperation: in fact, since
2864  * the requests temporarily redirected to the OOM queue could be redirected
2865  * again to dedicated queues at any time, the state needed to correctly
2866  * handle merging with the OOM queue would be quite complex and expensive
2867  * to maintain. Besides, in such a critical condition as an out of memory,
2868  * the benefits of queue merging may be little relevant, or even negligible.
2869  *
2870  * WARNING: queue merging may impair fairness among non-weight raised
2871  * queues, for at least two reasons: 1) the original weight of a
2872  * merged queue may change during the merged state, 2) even being the
2873  * weight the same, a merged queue may be bloated with many more
2874  * requests than the ones produced by its originally-associated
2875  * process.
2876  */
2877 static struct bfq_queue *
2878 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2879                      void *io_struct, bool request, struct bfq_io_cq *bic)
2880 {
2881         struct bfq_queue *in_service_bfqq, *new_bfqq;
2882 
2883         /* if a merge has already been setup, then proceed with that first */
2884         if (bfqq->new_bfqq)
2885                 return bfqq->new_bfqq;
2886 
2887         /*
2888          * Check delayed stable merge for rotational or non-queueing
2889          * devs. For this branch to be executed, bfqq must not be
2890          * currently merged with some other queue (i.e., bfqq->bic
2891          * must be non null). If we considered also merged queues,
2892          * then we should also check whether bfqq has already been
2893          * merged with bic->stable_merge_bfqq. But this would be
2894          * costly and complicated.
2895          */
2896         if (unlikely(!bfqd->nonrot_with_queueing)) {
2897                 /*
2898                  * Make sure also that bfqq is sync, because
2899                  * bic->stable_merge_bfqq may point to some queue (for
2900                  * stable merging) also if bic is associated with a
2901                  * sync queue, but this bfqq is async
2902                  */
2903                 if (bfq_bfqq_sync(bfqq) && bic->stable_merge_bfqq &&
2904                     !bfq_bfqq_just_created(bfqq) &&
2905                     time_is_before_jiffies(bfqq->split_time +
2906                                           msecs_to_jiffies(bfq_late_stable_merging)) &&
2907                     time_is_before_jiffies(bfqq->creation_time +
2908                                            msecs_to_jiffies(bfq_late_stable_merging))) {
2909                         struct bfq_queue *stable_merge_bfqq =
2910                                 bic->stable_merge_bfqq;
2911                         int proc_ref = min(bfqq_process_refs(bfqq),
2912                                            bfqq_process_refs(stable_merge_bfqq));
2913 
2914                         /* deschedule stable merge, because done or aborted here */
2915                         bfq_put_stable_ref(stable_merge_bfqq);
2916 
2917                         bic->stable_merge_bfqq = NULL;
2918 
2919                         if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
2920                             proc_ref > 0) {
2921                                 /* next function will take at least one ref */
2922                                 struct bfq_queue *new_bfqq =
2923                                         bfq_setup_merge(bfqq, stable_merge_bfqq);
2924 
2925                                 if (new_bfqq) {
2926                                         bic->stably_merged = true;
2927                                         if (new_bfqq->bic)
2928                                                 new_bfqq->bic->stably_merged =
2929                                                                         true;
2930                                 }
2931                                 return new_bfqq;
2932                         } else
2933                                 return NULL;
2934                 }
2935         }
2936 
2937         /*
2938          * Do not perform queue merging if the device is non
2939          * rotational and performs internal queueing. In fact, such a
2940          * device reaches a high speed through internal parallelism
2941          * and pipelining. This means that, to reach a high
2942          * throughput, it must have many requests enqueued at the same
2943          * time. But, in this configuration, the internal scheduling
2944          * algorithm of the device does exactly the job of queue
2945          * merging: it reorders requests so as to obtain as much as
2946          * possible a sequential I/O pattern. As a consequence, with
2947          * the workload generated by processes doing interleaved I/O,
2948          * the throughput reached by the device is likely to be the
2949          * same, with and without queue merging.
2950          *
2951          * Disabling merging also provides a remarkable benefit in
2952          * terms of throughput. Merging tends to make many workloads
2953          * artificially more uneven, because of shared queues
2954          * remaining non empty for incomparably more time than
2955          * non-merged queues. This may accentuate workload
2956          * asymmetries. For example, if one of the queues in a set of
2957          * merged queues has a higher weight than a normal queue, then
2958          * the shared queue may inherit such a high weight and, by
2959          * staying almost always active, may force BFQ to perform I/O
2960          * plugging most of the time. This evidently makes it harder
2961          * for BFQ to let the device reach a high throughput.
2962          *
2963          * Finally, the likely() macro below is not used because one
2964          * of the two branches is more likely than the other, but to
2965          * have the code path after the following if() executed as
2966          * fast as possible for the case of a non rotational device
2967          * with queueing. We want it because this is the fastest kind
2968          * of device. On the opposite end, the likely() may lengthen
2969          * the execution time of BFQ for the case of slower devices
2970          * (rotational or at least without queueing). But in this case
2971          * the execution time of BFQ matters very little, if not at
2972          * all.
2973          */
2974         if (likely(bfqd->nonrot_with_queueing))
2975                 return NULL;
2976 
2977         /*
2978          * Prevent bfqq from being merged if it has been created too
2979          * long ago. The idea is that true cooperating processes, and
2980          * thus their associated bfq_queues, are supposed to be
2981          * created shortly after each other. This is the case, e.g.,
2982          * for KVM/QEMU and dump I/O threads. Basing on this
2983          * assumption, the following filtering greatly reduces the
2984          * probability that two non-cooperating processes, which just
2985          * happen to do close I/O for some short time interval, have
2986          * their queues merged by mistake.
2987          */
2988         if (bfq_too_late_for_merging(bfqq))
2989                 return NULL;
2990 
2991         if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2992                 return NULL;
2993 
2994         /* If there is only one backlogged queue, don't search. */
2995         if (bfq_tot_busy_queues(bfqd) == 1)
2996                 return NULL;
2997 
2998         in_service_bfqq = bfqd->in_service_queue;
2999 
3000         if (in_service_bfqq && in_service_bfqq != bfqq &&
3001             likely(in_service_bfqq != &bfqd->oom_bfqq) &&
3002             bfq_rq_close_to_sector(io_struct, request,
3003                                    bfqd->in_serv_last_pos) &&
3004             bfqq->entity.parent == in_service_bfqq->entity.parent &&
3005             bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
3006                 new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
3007                 if (new_bfqq)
3008                         return new_bfqq;
3009         }
3010         /*
3011          * Check whether there is a cooperator among currently scheduled
3012          * queues. The only thing we need is that the bio/request is not
3013          * NULL, as we need it to establish whether a cooperator exists.
3014          */
3015         new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
3016                         bfq_io_struct_pos(io_struct, request));
3017 
3018         if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
3019             bfq_may_be_close_cooperator(bfqq, new_bfqq))
3020                 return bfq_setup_merge(bfqq, new_bfqq);
3021 
3022         return NULL;
3023 }
3024 
3025 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
3026 {
3027         struct bfq_io_cq *bic = bfqq->bic;
3028 
3029         /*
3030          * If !bfqq->bic, the queue is already shared or its requests
3031          * have already been redirected to a shared queue; both idle window
3032          * and weight raising state have already been saved. Do nothing.
3033          */
3034         if (!bic)
3035                 return;
3036 
3037         bic->saved_last_serv_time_ns = bfqq->last_serv_time_ns;
3038         bic->saved_inject_limit = bfqq->inject_limit;
3039         bic->saved_decrease_time_jif = bfqq->decrease_time_jif;
3040 
3041         bic->saved_weight = bfqq->entity.orig_weight;
3042         bic->saved_ttime = bfqq->ttime;
3043         bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
3044         bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
3045         bic->saved_io_start_time = bfqq->io_start_time;
3046         bic->saved_tot_idle_time = bfqq->tot_idle_time;
3047         bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
3048         bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
3049         if (unlikely(bfq_bfqq_just_created(bfqq) &&
3050                      !bfq_bfqq_in_large_burst(bfqq) &&
3051                      bfqq->bfqd->low_latency)) {
3052                 /*
3053                  * bfqq being merged right after being created: bfqq
3054                  * would have deserved interactive weight raising, but
3055                  * did not make it to be set in a weight-raised state,
3056                  * because of this early merge. Store directly the
3057                  * weight-raising state that would have been assigned
3058                  * to bfqq, so that to avoid that bfqq unjustly fails
3059                  * to enjoy weight raising if split soon.
3060                  */
3061                 bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
3062                 bic->saved_wr_start_at_switch_to_srt = bfq_smallest_from_now();
3063                 bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
3064                 bic->saved_last_wr_start_finish = jiffies;
3065         } else {
3066                 bic->saved_wr_coeff = bfqq->wr_coeff;
3067                 bic->saved_wr_start_at_switch_to_srt =
3068                         bfqq->wr_start_at_switch_to_srt;
3069                 bic->saved_service_from_wr = bfqq->service_from_wr;
3070                 bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
3071                 bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
3072         }
3073 }
3074 
3075 
3076 static void
3077 bfq_reassign_last_bfqq(struct bfq_queue *cur_bfqq, struct bfq_queue *new_bfqq)
3078 {
3079         if (cur_bfqq->entity.parent &&
3080             cur_bfqq->entity.parent->last_bfqq_created == cur_bfqq)
3081                 cur_bfqq->entity.parent->last_bfqq_created = new_bfqq;
3082         else if (cur_bfqq->bfqd && cur_bfqq->bfqd->last_bfqq_created == cur_bfqq)
3083                 cur_bfqq->bfqd->last_bfqq_created = new_bfqq;
3084 }
3085 
3086 void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3087 {
3088         /*
3089          * To prevent bfqq's service guarantees from being violated,
3090          * bfqq may be left busy, i.e., queued for service, even if
3091          * empty (see comments in __bfq_bfqq_expire() for
3092          * details). But, if no process will send requests to bfqq any
3093          * longer, then there is no point in keeping bfqq queued for
3094          * service. In addition, keeping bfqq queued for service, but
3095          * with no process ref any longer, may have caused bfqq to be
3096          * freed when dequeued from service. But this is assumed to
3097          * never happen.
3098          */
3099         if (bfq_bfqq_busy(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) &&
3100             bfqq != bfqd->in_service_queue)
3101                 bfq_del_bfqq_busy(bfqd, bfqq, false);
3102 
3103         bfq_reassign_last_bfqq(bfqq, NULL);
3104 
3105         bfq_put_queue(bfqq);
3106 }
3107 
3108 static void
3109 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
3110                 struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
3111 {
3112         bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
3113                 (unsigned long)new_bfqq->pid);
3114         /* Save weight raising and idle window of the merged queues */
3115         bfq_bfqq_save_state(bfqq);
3116         bfq_bfqq_save_state(new_bfqq);
3117         if (bfq_bfqq_IO_bound(bfqq))
3118                 bfq_mark_bfqq_IO_bound(new_bfqq);
3119         bfq_clear_bfqq_IO_bound(bfqq);
3120 
3121         /*
3122          * The processes associated with bfqq are cooperators of the
3123          * processes associated with new_bfqq. So, if bfqq has a
3124          * waker, then assume that all these processes will be happy
3125          * to let bfqq's waker freely inject I/O when they have no
3126          * I/O.
3127          */
3128         if (bfqq->waker_bfqq && !new_bfqq->waker_bfqq &&
3129             bfqq->waker_bfqq != new_bfqq) {
3130                 new_bfqq->waker_bfqq = bfqq->waker_bfqq;
3131                 new_bfqq->tentative_waker_bfqq = NULL;
3132 
3133                 /*
3134                  * If the waker queue disappears, then
3135                  * new_bfqq->waker_bfqq must be reset. So insert
3136                  * new_bfqq into the woken_list of the waker. See
3137                  * bfq_check_waker for details.
3138                  */
3139                 hlist_add_head(&new_bfqq->woken_list_node,
3140                                &new_bfqq->waker_bfqq->woken_list);
3141 
3142         }
3143 
3144         /*
3145          * If bfqq is weight-raised, then let new_bfqq inherit
3146          * weight-raising. To reduce false positives, neglect the case
3147          * where bfqq has just been created, but has not yet made it
3148          * to be weight-raised (which may happen because EQM may merge
3149          * bfqq even before bfq_add_request is executed for the first
3150          * time for bfqq). Handling this case would however be very
3151          * easy, thanks to the flag just_created.
3152          */
3153         if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
3154                 new_bfqq->wr_coeff = bfqq->wr_coeff;
3155                 new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
3156                 new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
3157                 new_bfqq->wr_start_at_switch_to_srt =
3158                         bfqq->wr_start_at_switch_to_srt;
3159                 if (bfq_bfqq_busy(new_bfqq))
3160                         bfqd->wr_busy_queues++;
3161                 new_bfqq->entity.prio_changed = 1;
3162         }
3163 
3164         if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
3165                 bfqq->wr_coeff = 1;
3166                 bfqq->entity.prio_changed = 1;
3167                 if (bfq_bfqq_busy(bfqq))
3168                         bfqd->wr_busy_queues--;
3169         }
3170 
3171         bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
3172                      bfqd->wr_busy_queues);
3173 
3174         /*
3175          * Merge queues (that is, let bic redirect its requests to new_bfqq)
3176          */
3177         bic_set_bfqq(bic, new_bfqq, 1);
3178         bfq_mark_bfqq_coop(new_bfqq);
3179         /*
3180          * new_bfqq now belongs to at least two bics (it is a shared queue):
3181          * set new_bfqq->bic to NULL. bfqq either:
3182          * - does not belong to any bic any more, and hence bfqq->bic must
3183          *   be set to NULL, or
3184          * - is a queue whose owning bics have already been redirected to a
3185          *   different queue, hence the queue is destined to not belong to
3186          *   any bic soon and bfqq->bic is already NULL (therefore the next
3187          *   assignment causes no harm).
3188          */
3189         new_bfqq->bic = NULL;
3190         /*
3191          * If the queue is shared, the pid is the pid of one of the associated
3192          * processes. Which pid depends on the exact sequence of merge events
3193          * the queue underwent. So printing such a pid is useless and confusing
3194          * because it reports a random pid between those of the associated
3195          * processes.
3196          * We mark such a queue with a pid -1, and then print SHARED instead of
3197          * a pid in logging messages.
3198          */
3199         new_bfqq->pid = -1;
3200         bfqq->bic = NULL;
3201 
3202         bfq_reassign_last_bfqq(bfqq, new_bfqq);
3203 
3204         bfq_release_process_ref(bfqd, bfqq);
3205 }
3206 
3207 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
3208                                 struct bio *bio)
3209 {
3210         struct bfq_data *bfqd = q->elevator->elevator_data;
3211         bool is_sync = op_is_sync(bio->bi_opf);
3212         struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
3213 
3214         /*
3215          * Disallow merge of a sync bio into an async request.
3216          */
3217         if (is_sync && !rq_is_sync(rq))
3218                 return false;
3219 
3220         /*
3221          * Lookup the bfqq that this bio will be queued with. Allow
3222          * merge only if rq is queued there.
3223          */
3224         if (!bfqq)
3225                 return false;
3226 
3227         /*
3228          * We take advantage of this function to perform an early merge
3229          * of the queues of possible cooperating processes.
3230          */
3231         new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false, bfqd->bio_bic);
3232         if (new_bfqq) {
3233                 /*
3234                  * bic still points to bfqq, then it has not yet been
3235                  * redirected to some other bfq_queue, and a queue
3236                  * merge between bfqq and new_bfqq can be safely
3237                  * fulfilled, i.e., bic can be redirected to new_bfqq
3238                  * and bfqq can be put.
3239                  */
3240                 bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
3241                                 new_bfqq);
3242                 /*
3243                  * If we get here, bio will be queued into new_queue,
3244                  * so use new_bfqq to decide whether bio and rq can be
3245                  * merged.
3246                  */
3247                 bfqq = new_bfqq;
3248 
3249                 /*
3250                  * Change also bqfd->bio_bfqq, as
3251                  * bfqd->bio_bic now points to new_bfqq, and
3252                  * this function may be invoked again (and then may
3253                  * use again bqfd->bio_bfqq).
3254                  */
3255                 bfqd->bio_bfqq = bfqq;
3256         }
3257 
3258         return bfqq == RQ_BFQQ(rq);
3259 }
3260 
3261 /*
3262  * Set the maximum time for the in-service queue to consume its
3263  * budget. This prevents seeky processes from lowering the throughput.
3264  * In practice, a time-slice service scheme is used with seeky
3265  * processes.
3266  */
3267 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
3268                                    struct bfq_queue *bfqq)
3269 {
3270         unsigned int timeout_coeff;
3271 
3272         if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
3273                 timeout_coeff = 1;
3274         else
3275                 timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
3276 
3277         bfqd->last_budget_start = ktime_get();
3278 
3279         bfqq->budget_timeout = jiffies +
3280                 bfqd->bfq_timeout * timeout_coeff;
3281 }
3282 
3283 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
3284                                        struct bfq_queue *bfqq)
3285 {
3286         if (bfqq) {
3287                 bfq_clear_bfqq_fifo_expire(bfqq);
3288 
3289                 bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
3290 
3291                 if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
3292                     bfqq->wr_coeff > 1 &&
3293                     bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
3294                     time_is_before_jiffies(bfqq->budget_timeout)) {
3295                         /*
3296                          * For soft real-time queues, move the start
3297                          * of the weight-raising period forward by the
3298                          * time the queue has not received any
3299                          * service. Otherwise, a relatively long
3300                          * service delay is likely to cause the
3301                          * weight-raising period of the queue to end,
3302                          * because of the short duration of the
3303                          * weight-raising period of a soft real-time
3304                          * queue.  It is worth noting that this move
3305                          * is not so dangerous for the other queues,
3306                          * because soft real-time queues are not
3307                          * greedy.
3308                          *
3309                          * To not add a further variable, we use the
3310                          * overloaded field budget_timeout to
3311                          * determine for how long the queue has not
3312                          * received service, i.e., how much time has
3313                          * elapsed since the queue expired. However,
3314                          * this is a little imprecise, because
3315                          * budget_timeout is set to jiffies if bfqq
3316                          * not only expires, but also remains with no
3317                          * request.
3318                          */
3319                         if (time_after(bfqq->budget_timeout,
3320                                        bfqq->last_wr_start_finish))
3321                                 bfqq->last_wr_start_finish +=
3322                                         jiffies - bfqq->budget_timeout;
3323                         else
3324                                 bfqq->last_wr_start_finish = jiffies;
3325                 }
3326 
3327                 bfq_set_budget_timeout(bfqd, bfqq);
3328                 bfq_log_bfqq(bfqd, bfqq,
3329                              "set_in_service_queue, cur-budget = %d",
3330                              bfqq->entity.budget);
3331         }
3332 
3333         bfqd->in_service_queue = bfqq;
3334         bfqd->in_serv_last_pos = 0;
3335 }
3336 
3337 /*
3338  * Get and set a new queue for service.
3339  */
3340 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
3341 {
3342         struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
3343 
3344         __bfq_set_in_service_queue(bfqd, bfqq);
3345         return bfqq;
3346 }
3347 
3348 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
3349 {
3350         struct bfq_queue *bfqq = bfqd->in_service_queue;
3351         u32 sl;
3352 
3353         bfq_mark_bfqq_wait_request(bfqq);
3354 
3355         /*
3356          * We don't want to idle for seeks, but we do want to allow
3357          * fair distribution of slice time for a process doing back-to-back
3358          * seeks. So allow a little bit of time for him to submit a new rq.
3359          */
3360         sl = bfqd->bfq_slice_idle;
3361         /*
3362          * Unless the queue is being weight-raised or the scenario is
3363          * asymmetric, grant only minimum idle time if the queue
3364          * is seeky. A long idling is preserved for a weight-raised
3365          * queue, or, more in general, in an asymmetric scenario,
3366          * because a long idling is needed for guaranteeing to a queue
3367          * its reserved share of the throughput (in particular, it is
3368          * needed if the queue has a higher weight than some other
3369          * queue).
3370          */
3371         if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3372             !bfq_asymmetric_scenario(bfqd, bfqq))
3373                 sl = min_t(u64, sl, BFQ_MIN_TT);
3374         else if (bfqq->wr_coeff > 1)
3375                 sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
3376 
3377         bfqd->last_idling_start = ktime_get();
3378         bfqd->last_idling_start_jiffies = jiffies;
3379 
3380         hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
3381                       HRTIMER_MODE_REL);
3382         bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
3383 }
3384 
3385 /*
3386  * In autotuning mode, max_budget is dynamically recomputed as the
3387  * amount of sectors transferred in timeout at the estimated peak
3388  * rate. This enables BFQ to utilize a full timeslice with a full
3389  * budget, even if the in-service queue is served at peak rate. And
3390  * this maximises throughput with sequential workloads.
3391  */
3392 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
3393 {
3394         return (u64)bfqd->peak_rate * USEC_PER_MSEC *
3395                 jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
3396 }
3397 
3398 /*
3399  * Update parameters related to throughput and responsiveness, as a
3400  * function of the estimated peak rate. See comments on
3401  * bfq_calc_max_budget(), and on the ref_wr_duration array.
3402  */
3403 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
3404 {
3405         if (bfqd->bfq_user_max_budget == 0) {
3406                 bfqd->bfq_max_budget =
3407                         bfq_calc_max_budget(bfqd);
3408                 bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
3409         }
3410 }
3411 
3412 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
3413                                        struct request *rq)
3414 {
3415         if (rq != NULL) { /* new rq dispatch now, reset accordingly */
3416                 bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
3417                 bfqd->peak_rate_samples = 1;
3418                 bfqd->sequential_samples = 0;
3419                 bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
3420                         blk_rq_sectors(rq);
3421         } else /* no new rq dispatched, just reset the number of samples */
3422                 bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
3423 
3424         bfq_log(bfqd,
3425                 "reset_rate_computation at end, sample %u/%u tot_sects %llu",
3426                 bfqd->peak_rate_samples, bfqd->sequential_samples,
3427                 bfqd->tot_sectors_dispatched);
3428 }
3429 
3430 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
3431 {
3432         u32 rate, weight, divisor;
3433 
3434         /*
3435          * For the convergence property to hold (see comments on
3436          * bfq_update_peak_rate()) and for the assessment to be
3437          * reliable, a minimum number of samples must be present, and
3438          * a minimum amount of time must have elapsed. If not so, do
3439          * not compute new rate. Just reset parameters, to get ready
3440          * for a new evaluation attempt.
3441          */
3442         if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
3443             bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
3444                 goto reset_computation;
3445 
3446         /*
3447          * If a new request completion has occurred after last
3448          * dispatch, then, to approximate the rate at which requests
3449          * have been served by the device, it is more precise to
3450          * extend the observation interval to the last completion.
3451          */
3452         bfqd->delta_from_first =
3453                 max_t(u64, bfqd->delta_from_first,
3454                       bfqd->last_completion - bfqd->first_dispatch);
3455 
3456         /*
3457          * Rate computed in sects/usec, and not sects/nsec, for
3458          * precision issues.
3459          */
3460         rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
3461                         div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
3462 
3463         /*
3464          * Peak rate not updated if:
3465          * - the percentage of sequential dispatches is below 3/4 of the
3466          *   total, and rate is below the current estimated peak rate
3467          * - rate is unreasonably high (> 20M sectors/sec)
3468          */
3469         if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
3470              rate <= bfqd->peak_rate) ||
3471                 rate > 20<<BFQ_RATE_SHIFT)
3472                 goto reset_computation;
3473 
3474         /*
3475          * We have to update the peak rate, at last! To this purpose,
3476          * we use a low-pass filter. We compute the smoothing constant
3477          * of the filter as a function of the 'weight' of the new
3478          * measured rate.
3479          *
3480          * As can be seen in next formulas, we define this weight as a
3481          * quantity proportional to how sequential the workload is,
3482          * and to how long the observation time interval is.
3483          *
3484          * The weight runs from 0 to 8. The maximum value of the
3485          * weight, 8, yields the minimum value for the smoothing
3486          * constant. At this minimum value for the smoothing constant,
3487          * the measured rate contributes for half of the next value of
3488          * the estimated peak rate.
3489          *
3490          * So, the first step is to compute the weight as a function
3491          * of how sequential the workload is. Note that the weight
3492          * cannot reach 9, because bfqd->sequential_samples cannot
3493          * become equal to bfqd->peak_rate_samples, which, in its
3494          * turn, holds true because bfqd->sequential_samples is not
3495          * incremented for the first sample.
3496          */
3497         weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
3498 
3499         /*
3500          * Second step: further refine the weight as a function of the
3501          * duration of the observation interval.
3502          */
3503         weight = min_t(u32, 8,
3504                        div_u64(weight * bfqd->delta_from_first,
3505                                BFQ_RATE_REF_INTERVAL));
3506 
3507         /*
3508          * Divisor ranging from 10, for minimum weight, to 2, for
3509          * maximum weight.
3510          */
3511         divisor = 10 - weight;
3512 
3513         /*
3514          * Finally, update peak rate:
3515          *
3516          * peak_rate = peak_rate * (divisor-1) / divisor  +  rate / divisor
3517          */
3518         bfqd->peak_rate *= divisor-1;
3519         bfqd->peak_rate /= divisor;
3520         rate /= divisor; /* smoothing constant alpha = 1/divisor */
3521 
3522         bfqd->peak_rate += rate;
3523 
3524         /*
3525          * For a very slow device, bfqd->peak_rate can reach 0 (see
3526          * the minimum representable values reported in the comments
3527          * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3528          * divisions by zero where bfqd->peak_rate is used as a
3529          * divisor.
3530          */
3531         bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
3532 
3533         update_thr_responsiveness_params(bfqd);
3534 
3535 reset_computation:
3536         bfq_reset_rate_computation(bfqd, rq);
3537 }
3538 
3539 /*
3540  * Update the read/write peak rate (the main quantity used for
3541  * auto-tuning, see update_thr_responsiveness_params()).
3542  *
3543  * It is not trivial to estimate the peak rate (correctly): because of
3544  * the presence of sw and hw queues between the scheduler and the
3545  * device components that finally serve I/O requests, it is hard to
3546  * say exactly when a given dispatched request is served inside the
3547  * device, and for how long. As a consequence, it is hard to know
3548  * precisely at what rate a given set of requests is actually served
3549  * by the device.
3550  *
3551  * On the opposite end, the dispatch time of any request is trivially
3552  * available, and, from this piece of information, the "dispatch rate"
3553  * of requests can be immediately computed. So, the idea in the next
3554  * function is to use what is known, namely request dispatch times
3555  * (plus, when useful, request completion times), to estimate what is
3556  * unknown, namely in-device request service rate.
3557  *
3558  * The main issue is that, because of the above facts, the rate at
3559  * which a certain set of requests is dispatched over a certain time
3560  * interval can vary greatly with respect to the rate at which the
3561  * same requests are then served. But, since the size of any
3562  * intermediate queue is limited, and the service scheme is lossless
3563  * (no request is silently dropped), the following obvious convergence
3564  * property holds: the number of requests dispatched MUST become
3565  * closer and closer to the number of requests completed as the
3566  * observation interval grows. This is the key property used in
3567  * the next function to estimate the peak service rate as a function
3568  * of the observed dispatch rate. The function assumes to be invoked
3569  * on every request dispatch.
3570  */
3571 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
3572 {
3573         u64 now_ns = ktime_get_ns();
3574 
3575         if (bfqd->peak_rate_samples == 0) { /* first dispatch */
3576                 bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
3577                         bfqd->peak_rate_samples);
3578                 bfq_reset_rate_computation(bfqd, rq);
3579                 goto update_last_values; /* will add one sample */
3580         }
3581 
3582         /*
3583          * Device idle for very long: the observation interval lasting
3584          * up to this dispatch cannot be a valid observation interval
3585          * for computing a new peak rate (similarly to the late-
3586          * completion event in bfq_completed_request()). Go to
3587          * update_rate_and_reset to have the following three steps
3588          * taken:
3589          * - close the observation interval at the last (previous)
3590          *   request dispatch or completion
3591          * - compute rate, if possible, for that observation interval
3592          * - start a new observation interval with this dispatch
3593          */
3594         if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
3595             bfqd->rq_in_driver == 0)
3596                 goto update_rate_and_reset;
3597 
3598         /* Update sampling information */
3599         bfqd->peak_rate_samples++;
3600 
3601         if ((bfqd->rq_in_driver > 0 ||
3602                 now_ns - bfqd->last_completion < BFQ_MIN_TT)
3603             && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
3604                 bfqd->sequential_samples++;
3605 
3606         bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
3607 
3608         /* Reset max observed rq size every 32 dispatches */
3609         if (likely(bfqd->peak_rate_samples % 32))
3610                 bfqd->last_rq_max_size =
3611                         max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
3612         else
3613                 bfqd->last_rq_max_size = blk_rq_sectors(rq);
3614 
3615         bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
3616 
3617         /* Target observation interval not yet reached, go on sampling */
3618         if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
3619                 goto update_last_values;
3620 
3621 update_rate_and_reset:
3622         bfq_update_rate_reset(bfqd, rq);
3623 update_last_values:
3624         bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
3625         if (RQ_BFQQ(rq) == bfqd->in_service_queue)
3626                 bfqd->in_serv_last_pos = bfqd->last_position;
3627         bfqd->last_dispatch = now_ns;
3628 }
3629 
3630 /*
3631  * Remove request from internal lists.
3632  */
3633 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
3634 {
3635         struct bfq_queue *bfqq = RQ_BFQQ(rq);
3636 
3637         /*
3638          * For consistency, the next instruction should have been
3639          * executed after removing the request from the queue and
3640          * dispatching it.  We execute instead this instruction before
3641          * bfq_remove_request() (and hence introduce a temporary
3642          * inconsistency), for efficiency.  In fact, should this
3643          * dispatch occur for a non in-service bfqq, this anticipated
3644          * increment prevents two counters related to bfqq->dispatched
3645          * from risking to be, first, uselessly decremented, and then
3646          * incremented again when the (new) value of bfqq->dispatched
3647          * happens to be taken into account.
3648          */
3649         bfqq->dispatched++;
3650         bfq_update_peak_rate(q->elevator->elevator_data, rq);
3651 
3652         bfq_remove_request(q, rq);
3653 }
3654 
3655 /*
3656  * There is a case where idling does not have to be performed for
3657  * throughput concerns, but to preserve the throughput share of
3658  * the process associated with bfqq.
3659  *
3660  * To introduce this case, we can note that allowing the drive
3661  * to enqueue more than one request at a time, and hence
3662  * delegating de facto final scheduling decisions to the
3663  * drive's internal scheduler, entails loss of control on the
3664  * actual request service order. In particular, the critical
3665  * situation is when requests from different processes happen
3666  * to be present, at the same time, in the internal queue(s)
3667  * of the drive. In such a situation, the drive, by deciding
3668  * the service order of the internally-queued requests, does
3669  * determine also the actual throughput distribution among
3670  * these processes. But the drive typically has no notion or
3671  * concern about per-process throughput distribution, and
3672  * makes its decisions only on a per-request basis. Therefore,
3673  * the service distribution enforced by the drive's internal
3674  * scheduler is likely to coincide with the desired throughput
3675  * distribution only in a completely symmetric, or favorably
3676  * skewed scenario where:
3677  * (i-a) each of these processes must get the same throughput as
3678  *       the others,
3679  * (i-b) in case (i-a) does not hold, it holds that the process
3680  *       associated with bfqq must receive a lower or equal
3681  *       throughput than any of the other processes;
3682  * (ii)  the I/O of each process has the same properties, in
3683  *       terms of locality (sequential or random), direction
3684  *       (reads or writes), request sizes, greediness
3685  *       (from I/O-bound to sporadic), and so on;
3686 
3687  * In fact, in such a scenario, the drive tends to treat the requests
3688  * of each process in about the same way as the requests of the
3689  * others, and thus to provide each of these processes with about the
3690  * same throughput.  This is exactly the desired throughput
3691  * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3692  * even more convenient distribution for (the process associated with)
3693  * bfqq.
3694  *
3695  * In contrast, in any asymmetric or unfavorable scenario, device
3696  * idling (I/O-dispatch plugging) is certainly needed to guarantee
3697  * that bfqq receives its assigned fraction of the device throughput
3698  * (see [1] for details).
3699  *
3700  * The problem is that idling may significantly reduce throughput with
3701  * certain combinations of types of I/O and devices. An important
3702  * example is sync random I/O on flash storage with command
3703  * queueing. So, unless bfqq falls in cases where idling also boosts
3704  * throughput, it is important to check conditions (i-a), i(-b) and
3705  * (ii) accurately, so as to avoid idling when not strictly needed for
3706  * service guarantees.
3707  *
3708  * Unfortunately, it is extremely difficult to thoroughly check
3709  * condition (ii). And, in case there are active groups, it becomes
3710  * very difficult to check conditions (i-a) and (i-b) too.  In fact,
3711  * if there are active groups, then, for conditions (i-a) or (i-b) to
3712  * become false 'indirectly', it is enough that an active group
3713  * contains more active processes or sub-groups than some other active
3714  * group. More precisely, for conditions (i-a) or (i-b) to become
3715  * false because of such a group, it is not even necessary that the
3716  * group is (still) active: it is sufficient that, even if the group
3717  * has become inactive, some of its descendant processes still have
3718  * some request already dispatched but still waiting for
3719  * completion. In fact, requests have still to be guaranteed their
3720  * share of the throughput even after being dispatched. In this
3721  * respect, it is easy to show that, if a group frequently becomes
3722  * inactive while still having in-flight requests, and if, when this
3723  * happens, the group is not considered in the calculation of whether
3724  * the scenario is asymmetric, then the group may fail to be
3725  * guaranteed its fair share of the throughput (basically because
3726  * idling may not be performed for the descendant processes of the
3727  * group, but it had to be).  We address this issue with the following
3728  * bi-modal behavior, implemented in the function
3729  * bfq_asymmetric_scenario().
3730  *
3731  * If there are groups with requests waiting for completion
3732  * (as commented above, some of these groups may even be
3733  * already inactive), then the scenario is tagged as
3734  * asymmetric, conservatively, without checking any of the
3735  * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3736  * This behavior matches also the fact that groups are created
3737  * exactly if controlling I/O is a primary concern (to
3738  * preserve bandwidth and latency guarantees).
3739  *
3740  * On the opposite end, if there are no groups with requests waiting
3741  * for completion, then only conditions (i-a) and (i-b) are actually
3742  * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3743  * idling is not performed, regardless of whether condition (ii)
3744  * holds.  In other words, only if conditions (i-a) and (i-b) do not
3745  * hold, then idling is allowed, and the device tends to be prevented
3746  * from queueing many requests, possibly of several processes. Since
3747  * there are no groups with requests waiting for completion, then, to
3748  * control conditions (i-a) and (i-b) it is enough to check just
3749  * whether all the queues with requests waiting for completion also
3750  * have the same weight.
3751  *
3752  * Not checking condition (ii) evidently exposes bfqq to the
3753  * risk of getting less throughput than its fair share.
3754  * However, for queues with the same weight, a further
3755  * mechanism, preemption, mitigates or even eliminates this
3756  * problem. And it does so without consequences on overall
3757  * throughput. This mechanism and its benefits are explained
3758  * in the next three paragraphs.
3759  *
3760  * Even if a queue, say Q, is expired when it remains idle, Q
3761  * can still preempt the new in-service queue if the next
3762  * request of Q arrives soon (see the comments on
3763  * bfq_bfqq_update_budg_for_activation). If all queues and
3764  * groups have the same weight, this form of preemption,
3765  * combined with the hole-recovery heuristic described in the
3766  * comments on function bfq_bfqq_update_budg_for_activation,
3767  * are enough to preserve a correct bandwidth distribution in
3768  * the mid term, even without idling. In fact, even if not
3769  * idling allows the internal queues of the device to contain
3770  * many requests, and thus to reorder requests, we can rather
3771  * safely assume that the internal scheduler still preserves a
3772  * minimum of mid-term fairness.
3773  *
3774  * More precisely, this preemption-based, idleless approach
3775  * provides fairness in terms of IOPS, and not sectors per
3776  * second. This can be seen with a simple example. Suppose
3777  * that there are two queues with the same weight, but that
3778  * the first queue receives requests of 8 sectors, while the
3779  * second queue receives requests of 1024 sectors. In
3780  * addition, suppose that each of the two queues contains at
3781  * most one request at a time, which implies that each queue
3782  * always remains idle after it is served. Finally, after
3783  * remaining idle, each queue receives very quickly a new
3784  * request. It follows that the two queues are served
3785  * alternatively, preempting each other if needed. This
3786  * implies that, although both queues have the same weight,
3787  * the queue with large requests receives a service that is
3788  * 1024/8 times as high as the service received by the other
3789  * queue.
3790  *
3791  * The motivation for using preemption instead of idling (for
3792  * queues with the same weight) is that, by not idling,
3793  * service guarantees are preserved (completely or at least in
3794  * part) without minimally sacrificing throughput. And, if
3795  * there is no active group, then the primary expectation for
3796  * this device is probably a high throughput.
3797  *
3798  * We are now left only with explaining the two sub-conditions in the
3799  * additional compound condition that is checked below for deciding
3800  * whether the scenario is asymmetric. To explain the first
3801  * sub-condition, we need to add that the function
3802  * bfq_asymmetric_scenario checks the weights of only
3803  * non-weight-raised queues, for efficiency reasons (see comments on
3804  * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3805  * is checked explicitly here. More precisely, the compound condition
3806  * below takes into account also the fact that, even if bfqq is being
3807  * weight-raised, the scenario is still symmetric if all queues with
3808  * requests waiting for completion happen to be
3809  * weight-raised. Actually, we should be even more precise here, and
3810  * differentiate between interactive weight raising and soft real-time
3811  * weight raising.
3812  *
3813  * The second sub-condition checked in the compound condition is
3814  * whether there is a fair amount of already in-flight I/O not
3815  * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3816  * following reason. The drive may decide to serve in-flight
3817  * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3818  * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3819  * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3820  * basically uncontrolled amount of I/O from other queues may be
3821  * dispatched too, possibly causing the service of bfqq's I/O to be
3822  * delayed even longer in the drive. This problem gets more and more
3823  * serious as the speed and the queue depth of the drive grow,
3824  * because, as these two quantities grow, the probability to find no
3825  * queue busy but many requests in flight grows too. By contrast,
3826  * plugging I/O dispatching minimizes the delay induced by already
3827  * in-flight I/O, and enables bfqq to recover the bandwidth it may
3828  * lose because of this delay.
3829  *
3830  * As a side note, it is worth considering that the above
3831  * device-idling countermeasures may however fail in the following
3832  * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3833  * in a time period during which all symmetry sub-conditions hold, and
3834  * therefore the device is allowed to enqueue many requests, but at
3835  * some later point in time some sub-condition stops to hold, then it
3836  * may become impossible to make requests be served in the desired
3837  * order until all the requests already queued in the device have been
3838  * served. The last sub-condition commented above somewhat mitigates
3839  * this problem for weight-raised queues.
3840  *
3841  * However, as an additional mitigation for this problem, we preserve
3842  * plugging for a special symmetric case that may suddenly turn into
3843  * asymmetric: the case where only bfqq is busy. In this case, not
3844  * expiring bfqq does not cause any harm to any other queues in terms
3845  * of service guarantees. In contrast, it avoids the following unlucky
3846  * sequence of events: (1) bfqq is expired, (2) a new queue with a
3847  * lower weight than bfqq becomes busy (or more queues), (3) the new
3848  * queue is served until a new request arrives for bfqq, (4) when bfqq
3849  * is finally served, there are so many requests of the new queue in
3850  * the drive that the pending requests for bfqq take a lot of time to
3851  * be served. In particular, event (2) may case even already
3852  * dispatched requests of bfqq to be delayed, inside the drive. So, to
3853  * avoid this series of events, the scenario is preventively declared
3854  * as asymmetric also if bfqq is the only busy queues
3855  */
3856 static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3857                                                  struct bfq_queue *bfqq)
3858 {
3859         int tot_busy_queues = bfq_tot_busy_queues(bfqd);
3860 
3861         /* No point in idling for bfqq if it won't get requests any longer */
3862         if (unlikely(!bfqq_process_refs(bfqq)))
3863                 return false;
3864 
3865         return (bfqq->wr_coeff > 1 &&
3866                 (bfqd->wr_busy_queues <
3867                  tot_busy_queues ||
3868                  bfqd->rq_in_driver >=
3869                  bfqq->dispatched + 4)) ||
3870                 bfq_asymmetric_scenario(bfqd, bfqq) ||
3871                 tot_busy_queues == 1;
3872 }
3873 
3874 static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3875                               enum bfqq_expiration reason)
3876 {
3877         /*
3878          * If this bfqq is shared between multiple processes, check
3879          * to make sure that those processes are still issuing I/Os
3880          * within the mean seek distance. If not, it may be time to
3881          * break the queues apart again.
3882          */
3883         if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
3884                 bfq_mark_bfqq_split_coop(bfqq);
3885 
3886         /*
3887          * Consider queues with a higher finish virtual time than
3888          * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3889          * true, then bfqq's bandwidth would be violated if an
3890          * uncontrolled amount of I/O from these queues were
3891          * dispatched while bfqq is waiting for its new I/O to
3892          * arrive. This is exactly what may happen if this is a forced
3893          * expiration caused by a preemption attempt, and if bfqq is
3894          * not re-scheduled. To prevent this from happening, re-queue
3895          * bfqq if it needs I/O-dispatch plugging, even if it is
3896          * empty. By doing so, bfqq is granted to be served before the
3897          * above queues (provided that bfqq is of course eligible).
3898          */
3899         if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
3900             !(reason == BFQQE_PREEMPTED &&
3901               idling_needed_for_service_guarantees(bfqd, bfqq))) {
3902                 if (bfqq->dispatched == 0)
3903                         /*
3904                          * Overloading budget_timeout field to store
3905                          * the time at which the queue remains with no
3906                          * backlog and no outstanding request; used by
3907                          * the weight-raising mechanism.
3908                          */
3909                         bfqq->budget_timeout = jiffies;
3910 
3911                 bfq_del_bfqq_busy(bfqd, bfqq, true);
3912         } else {
3913                 bfq_requeue_bfqq(bfqd, bfqq, true);
3914                 /*
3915                  * Resort priority tree of potential close cooperators.
3916                  * See comments on bfq_pos_tree_add_move() for the unlikely().
3917                  */
3918                 if (unlikely(!bfqd->nonrot_with_queueing &&
3919                              !RB_EMPTY_ROOT(&bfqq->sort_list)))
3920                         bfq_pos_tree_add_move(bfqd, bfqq);
3921         }
3922 
3923         /*
3924          * All in-service entities must have been properly deactivated
3925          * or requeued before executing the next function, which
3926          * resets all in-service entities as no more in service. This
3927          * may cause bfqq to be freed. If this happens, the next
3928          * function returns true.
3929          */
3930         return __bfq_bfqd_reset_in_service(bfqd);
3931 }
3932 
3933 /**
3934  * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3935  * @bfqd: device data.
3936  * @bfqq: queue to update.
3937  * @reason: reason for expiration.
3938  *
3939  * Handle the feedback on @bfqq budget at queue expiration.
3940  * See the body for detailed comments.
3941  */
3942 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
3943                                      struct bfq_queue *bfqq,
3944                                      enum bfqq_expiration reason)
3945 {
3946         struct request *next_rq;
3947         int budget, min_budget;
3948 
3949         min_budget = bfq_min_budget(bfqd);
3950 
3951         if (bfqq->wr_coeff == 1)
3952                 budget = bfqq->max_budget;
3953         else /*
3954               * Use a constant, low budget for weight-raised queues,
3955               * to help achieve a low latency. Keep it slightly higher
3956               * than the minimum possible budget, to cause a little
3957               * bit fewer expirations.
3958               */
3959                 budget = 2 * min_budget;
3960 
3961         bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
3962                 bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
3963         bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
3964                 budget, bfq_min_budget(bfqd));
3965         bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
3966                 bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
3967 
3968         if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
3969                 switch (reason) {
3970                 /*
3971                  * Caveat: in all the following cases we trade latency
3972                  * for throughput.
3973                  */
3974                 case BFQQE_TOO_IDLE:
3975                         /*
3976                          * This is the only case where we may reduce
3977                          * the budget: if there is no request of the
3978                          * process still waiting for completion, then
3979                          * we assume (tentatively) that the timer has
3980                          * expired because the batch of requests of
3981                          * the process could have been served with a
3982                          * smaller budget.  Hence, betting that
3983                          * process will behave in the same way when it
3984                          * becomes backlogged again, we reduce its
3985                          * next budget.  As long as we guess right,
3986                          * this budget cut reduces the latency
3987                          * experienced by the process.
3988                          *
3989                          * However, if there are still outstanding
3990                          * requests, then the process may have not yet
3991                          * issued its next request just because it is
3992                          * still waiting for the completion of some of
3993                          * the still outstanding ones.  So in this
3994                          * subcase we do not reduce its budget, on the
3995                          * contrary we increase it to possibly boost
3996                          * the throughput, as discussed in the
3997                          * comments to the BUDGET_TIMEOUT case.
3998                          */
3999                         if (bfqq->dispatched > 0) /* still outstanding reqs */
4000                                 budget = min(budget * 2, bfqd->bfq_max_budget);
4001                         else {
4002                                 if (budget > 5 * min_budget)
4003                                         budget -= 4 * min_budget;
4004                                 else
4005                                         budget = min_budget;
4006                         }
4007                         break;
4008                 case BFQQE_BUDGET_TIMEOUT:
4009                         /*
4010                          * We double the budget here because it gives
4011                          * the chance to boost the throughput if this
4012                          * is not a seeky process (and has bumped into
4013                          * this timeout because of, e.g., ZBR).
4014                          */
4015                         budget = min(budget * 2, bfqd->bfq_max_budget);
4016                         break;
4017                 case BFQQE_BUDGET_EXHAUSTED:
4018                         /*
4019                          * The process still has backlog, and did not
4020                          * let either the budget timeout or the disk
4021                          * idling timeout expire. Hence it is not
4022                          * seeky, has a short thinktime and may be
4023                          * happy with a higher budget too. So
4024                          * definitely increase the budget of this good
4025                          * candidate to boost the disk throughput.
4026                          */
4027                         budget = min(budget * 4, bfqd->bfq_max_budget);
4028                         break;
4029                 case BFQQE_NO_MORE_REQUESTS:
4030                         /*
4031                          * For queues that expire for this reason, it
4032                          * is particularly important to keep the
4033                          * budget close to the actual service they
4034                          * need. Doing so reduces the timestamp
4035                          * misalignment problem described in the
4036                          * comments in the body of
4037                          * __bfq_activate_entity. In fact, suppose
4038                          * that a queue systematically expires for
4039                          * BFQQE_NO_MORE_REQUESTS and presents a
4040                          * new request in time to enjoy timestamp
4041                          * back-shifting. The larger the budget of the
4042                          * queue is with respect to the service the
4043                          * queue actually requests in each service
4044                          * slot, the more times the queue can be
4045                          * reactivated with the same virtual finish
4046                          * time. It follows that, even if this finish
4047                          * time is pushed to the system virtual time
4048                          * to reduce the consequent timestamp
4049                          * misalignment, the queue unjustly enjoys for
4050                          * many re-activations a lower finish time
4051                          * than all newly activated queues.
4052                          *
4053                          * The service needed by bfqq is measured
4054                          * quite precisely by bfqq->entity.service.
4055                          * Since bfqq does not enjoy device idling,
4056                          * bfqq->entity.service is equal to the number
4057                          * of sectors that the process associated with
4058                          * bfqq requested to read/write before waiting
4059                          * for request completions, or blocking for
4060                          * other reasons.
4061                          */
4062                         budget = max_t(int, bfqq->entity.service, min_budget);
4063                         break;
4064                 default:
4065                         return;
4066                 }
4067         } else if (!bfq_bfqq_sync(bfqq)) {
4068                 /*
4069                  * Async queues get always the maximum possible
4070                  * budget, as for them we do not care about latency
4071                  * (in addition, their ability to dispatch is limited
4072                  * by the charging factor).
4073                  */
4074                 budget = bfqd->bfq_max_budget;
4075         }
4076 
4077         bfqq->max_budget = budget;
4078 
4079         if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
4080             !bfqd->bfq_user_max_budget)
4081                 bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
4082 
4083         /*
4084          * If there is still backlog, then assign a new budget, making
4085          * sure that it is large enough for the next request.  Since
4086          * the finish time of bfqq must be kept in sync with the
4087          * budget, be sure to call __bfq_bfqq_expire() *after* this
4088          * update.
4089          *
4090          * If there is no backlog, then no need to update the budget;
4091          * it will be updated on the arrival of a new request.
4092          */
4093         next_rq = bfqq->next_rq;
4094         if (next_rq)
4095                 bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
4096                                             bfq_serv_to_charge(next_rq, bfqq));
4097 
4098         bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
4099                         next_rq ? blk_rq_sectors(next_rq) : 0,
4100                         bfqq->entity.budget);
4101 }
4102 
4103 /*
4104  * Return true if the process associated with bfqq is "slow". The slow
4105  * flag is used, in addition to the budget timeout, to reduce the
4106  * amount of service provided to seeky processes, and thus reduce
4107  * their chances to lower the throughput. More details in the comments
4108  * on the function bfq_bfqq_expire().
4109  *
4110  * An important observation is in order: as discussed in the comments
4111  * on the function bfq_update_peak_rate(), with devices with internal
4112  * queues, it is hard if ever possible to know when and for how long
4113  * an I/O request is processed by the device (apart from the trivial
4114  * I/O pattern where a new request is dispatched only after the
4115  * previous one has been completed). This makes it hard to evaluate
4116  * the real rate at which the I/O requests of each bfq_queue are
4117  * served.  In fact, for an I/O scheduler like BFQ, serving a
4118  * bfq_queue means just dispatching its requests during its service
4119  * slot (i.e., until the budget of the queue is exhausted, or the
4120  * queue remains idle, or, finally, a timeout fires). But, during the
4121  * service slot of a bfq_queue, around 100 ms at most, the device may
4122  * be even still processing requests of bfq_queues served in previous
4123  * service slots. On the opposite end, the requests of the in-service
4124  * bfq_queue may be completed after the service slot of the queue
4125  * finishes.
4126  *
4127  * Anyway, unless more sophisticated solutions are used
4128  * (where possible), the sum of the sizes of the requests dispatched
4129  * during the service slot of a bfq_queue is probably the only
4130  * approximation available for the service received by the bfq_queue
4131  * during its service slot. And this sum is the quantity used in this
4132  * function to evaluate the I/O speed of a process.
4133  */
4134 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4135                                  bool compensate, enum bfqq_expiration reason,
4136                                  unsigned long *delta_ms)
4137 {
4138         ktime_t delta_ktime;
4139         u32 delta_usecs;
4140         bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
4141 
4142         if (!bfq_bfqq_sync(bfqq))
4143                 return false;
4144 
4145         if (compensate)
4146                 delta_ktime = bfqd->last_idling_start;
4147         else
4148                 delta_ktime = ktime_get();
4149         delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
4150         delta_usecs = ktime_to_us(delta_ktime);
4151 
4152         /* don't use too short time intervals */
4153         if (delta_usecs < 1000) {
4154                 if (blk_queue_nonrot(bfqd->queue))
4155                          /*
4156                           * give same worst-case guarantees as idling
4157                           * for seeky
4158                           */
4159                         *delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
4160                 else /* charge at least one seek */
4161                         *delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
4162 
4163                 return slow;
4164         }
4165 
4166         *delta_ms = delta_usecs / USEC_PER_MSEC;
4167 
4168         /*
4169          * Use only long (> 20ms) intervals to filter out excessive
4170          * spikes in service rate estimation.
4171          */
4172         if (delta_usecs > 20000) {
4173                 /*
4174                  * Caveat for rotational devices: processes doing I/O
4175                  * in the slower disk zones tend to be slow(er) even
4176                  * if not seeky. In this respect, the estimated peak
4177                  * rate is likely to be an average over the disk
4178                  * surface. Accordingly, to not be too harsh with
4179                  * unlucky processes, a process is deemed slow only if
4180                  * its rate has been lower than half of the estimated
4181                  * peak rate.
4182                  */
4183                 slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
4184         }
4185 
4186         bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
4187 
4188         return slow;
4189 }
4190 
4191 /*
4192  * To be deemed as soft real-time, an application must meet two
4193  * requirements. First, the application must not require an average
4194  * bandwidth higher than the approximate bandwidth required to playback or
4195  * record a compressed high-definition video.
4196  * The next function is invoked on the completion of the last request of a
4197  * batch, to compute the next-start time instant, soft_rt_next_start, such
4198  * that, if the next request of the application does not arrive before
4199  * soft_rt_next_start, then the above requirement on the bandwidth is met.
4200  *
4201  * The second requirement is that the request pattern of the application is
4202  * isochronous, i.e., that, after issuing a request or a batch of requests,
4203  * the application stops issuing new requests until all its pending requests
4204  * have been completed. After that, the application may issue a new batch,
4205  * and so on.
4206  * For this reason the next function is invoked to compute
4207  * soft_rt_next_start only for applications that meet this requirement,
4208  * whereas soft_rt_next_start is set to infinity for applications that do
4209  * not.
4210  *
4211  * Unfortunately, even a greedy (i.e., I/O-bound) application may
4212  * happen to meet, occasionally or systematically, both the above
4213  * bandwidth and isochrony requirements. This may happen at least in
4214  * the following circumstances. First, if the CPU load is high. The
4215  * application may stop issuing requests while the CPUs are busy
4216  * serving other processes, then restart, then stop again for a while,
4217  * and so on. The other circumstances are related to the storage
4218  * device: the storage device is highly loaded or reaches a low-enough
4219  * throughput with the I/O of the application (e.g., because the I/O
4220  * is random and/or the device is slow). In all these cases, the
4221  * I/O of the application may be simply slowed down enough to meet
4222  * the bandwidth and isochrony requirements. To reduce the probability
4223  * that greedy applications are deemed as soft real-time in these
4224  * corner cases, a further rule is used in the computation of
4225  * soft_rt_next_start: the return value of this function is forced to
4226  * be higher than the maximum between the following two quantities.
4227  *
4228  * (a) Current time plus: (1) the maximum time for which the arrival
4229  *     of a request is waited for when a sync queue becomes idle,
4230  *     namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
4231  *     postpone for a moment the reason for adding a few extra
4232  *     jiffies; we get back to it after next item (b).  Lower-bounding
4233  *     the return value of this function with the current time plus
4234  *     bfqd->bfq_slice_idle tends to filter out greedy applications,
4235  *     because the latter issue their next request as soon as possible
4236  *     after the last one has been completed. In contrast, a soft
4237  *     real-time application spends some time processing data, after a
4238  *     batch of its requests has been completed.
4239  *
4240  * (b) Current value of bfqq->soft_rt_next_start. As pointed out
4241  *     above, greedy applications may happen to meet both the
4242  *     bandwidth and isochrony requirements under heavy CPU or
4243  *     storage-device load. In more detail, in these scenarios, these
4244  *     applications happen, only for limited time periods, to do I/O
4245  *     slowly enough to meet all the requirements described so far,
4246  *     including the filtering in above item (a). These slow-speed
4247  *     time intervals are usually interspersed between other time
4248  *     intervals during which these applications do I/O at a very high
4249  *     speed. Fortunately, exactly because of the high speed of the
4250  *     I/O in the high-speed intervals, the values returned by this
4251  *     function happen to be so high, near the end of any such
4252  *     high-speed interval, to be likely to fall *after* the end of
4253  *     the low-speed time interval that follows. These high values are
4254  *     stored in bfqq->soft_rt_next_start after each invocation of
4255  *     this function. As a consequence, if the last value of
4256  *     bfqq->soft_rt_next_start is constantly used to lower-bound the
4257  *     next value that this function may return, then, from the very
4258  *     beginning of a low-speed interval, bfqq->soft_rt_next_start is
4259  *     likely to be constantly kept so high that any I/O request
4260  *     issued during the low-speed interval is considered as arriving
4261  *     to soon for the application to be deemed as soft
4262  *     real-time. Then, in the high-speed interval that follows, the
4263  *     application will not be deemed as soft real-time, just because
4264  *     it will do I/O at a high speed. And so on.
4265  *
4266  * Getting back to the filtering in item (a), in the following two
4267  * cases this filtering might be easily passed by a greedy
4268  * application, if the reference quantity was just
4269  * bfqd->bfq_slice_idle:
4270  * 1) HZ is so low that the duration of a jiffy is comparable to or
4271  *    higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
4272  *    devices with HZ=100. The time granularity may be so coarse
4273  *    that the approximation, in jiffies, of bfqd->bfq_slice_idle
4274  *    is rather lower than the exact value.
4275  * 2) jiffies, instead of increasing at a constant rate, may stop increasing
4276  *    for a while, then suddenly 'jump' by several units to recover the lost
4277  *    increments. This seems to happen, e.g., inside virtual machines.
4278  * To address this issue, in the filtering in (a) we do not use as a
4279  * reference time interval just bfqd->bfq_slice_idle, but
4280  * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
4281  * minimum number of jiffies for which the filter seems to be quite
4282  * precise also in embedded systems and KVM/QEMU virtual machines.
4283  */
4284 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
4285                                                 struct bfq_queue *bfqq)
4286 {
4287         return max3(bfqq->soft_rt_next_start,
4288                     bfqq->last_idle_bklogged +
4289                     HZ * bfqq->service_from_backlogged /
4290                     bfqd->bfq_wr_max_softrt_rate,
4291                     jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
4292 }
4293 
4294 /**
4295  * bfq_bfqq_expire - expire a queue.
4296  * @bfqd: device owning the queue.
4297  * @bfqq: the queue to expire.
4298  * @compensate: if true, compensate for the time spent idling.
4299  * @reason: the reason causing the expiration.
4300  *
4301  * If the process associated with bfqq does slow I/O (e.g., because it
4302  * issues random requests), we charge bfqq with the time it has been
4303  * in service instead of the service it has received (see
4304  * bfq_bfqq_charge_time for details on how this goal is achieved). As
4305  * a consequence, bfqq will typically get higher timestamps upon
4306  * reactivation, and hence it will be rescheduled as if it had
4307  * received more service than what it has actually received. In the
4308  * end, bfqq receives less service in proportion to how slowly its
4309  * associated process consumes its budgets (and hence how seriously it
4310  * tends to lower the throughput). In addition, this time-charging
4311  * strategy guarantees time fairness among slow processes. In
4312  * contrast, if the process associated with bfqq is not slow, we
4313  * charge bfqq exactly with the service it has received.
4314  *
4315  * Charging time to the first type of queues and the exact service to
4316  * the other has the effect of using the WF2Q+ policy to schedule the
4317  * former on a timeslice basis, without violating service domain
4318  * guarantees among the latter.
4319  */
4320 void bfq_bfqq_expire(struct bfq_data *bfqd,
4321                      struct bfq_queue *bfqq,
4322                      bool compensate,
4323                      enum bfqq_expiration reason)
4324 {
4325         bool slow;
4326         unsigned long delta = 0;
4327         struct bfq_entity *entity = &bfqq->entity;
4328 
4329         /*
4330          * Check whether the process is slow (see bfq_bfqq_is_slow).
4331          */
4332         slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
4333 
4334         /*
4335          * As above explained, charge slow (typically seeky) and
4336          * timed-out queues with the time and not the service
4337          * received, to favor sequential workloads.
4338          *
4339          * Processes doing I/O in the slower disk zones will tend to
4340          * be slow(er) even if not seeky. Therefore, since the
4341          * estimated peak rate is actually an average over the disk
4342          * surface, these processes may timeout just for bad luck. To
4343          * avoid punishing them, do not charge time to processes that
4344          * succeeded in consuming at least 2/3 of their budget. This
4345          * allows BFQ to preserve enough elasticity to still perform
4346          * bandwidth, and not time, distribution with little unlucky
4347          * or quasi-sequential processes.
4348          */
4349         if (bfqq->wr_coeff == 1 &&
4350             (slow ||
4351              (reason == BFQQE_BUDGET_TIMEOUT &&
4352               bfq_bfqq_budget_left(bfqq) >=  entity->budget / 3)))
4353                 bfq_bfqq_charge_time(bfqd, bfqq, delta);
4354 
4355         if (bfqd->low_latency && bfqq->wr_coeff == 1)
4356                 bfqq->last_wr_start_finish = jiffies;
4357 
4358         if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
4359             RB_EMPTY_ROOT(&bfqq->sort_list)) {
4360                 /*
4361                  * If we get here, and there are no outstanding
4362                  * requests, then the request pattern is isochronous
4363                  * (see the comments on the function
4364                  * bfq_bfqq_softrt_next_start()). Therefore we can
4365                  * compute soft_rt_next_start.
4366                  *
4367                  * If, instead, the queue still has outstanding
4368                  * requests, then we have to wait for the completion
4369                  * of all the outstanding requests to discover whether
4370                  * the request pattern is actually isochronous.
4371                  */
4372                 if (bfqq->dispatched == 0)
4373                         bfqq->soft_rt_next_start =
4374                                 bfq_bfqq_softrt_next_start(bfqd, bfqq);
4375                 else if (bfqq->dispatched > 0) {
4376                         /*
4377                          * Schedule an update of soft_rt_next_start to when
4378                          * the task may be discovered to be isochronous.
4379                          */
4380                         bfq_mark_bfqq_softrt_update(bfqq);
4381                 }
4382         }
4383 
4384         bfq_log_bfqq(bfqd, bfqq,
4385                 "expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
4386                 slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
4387 
4388         /*
4389          * bfqq expired, so no total service time needs to be computed
4390          * any longer: reset state machine for measuring total service
4391          * times.
4392          */
4393         bfqd->rqs_injected = bfqd->wait_dispatch = false;
4394         bfqd->waited_rq = NULL;
4395 
4396         /*
4397          * Increase, decrease or leave budget unchanged according to
4398          * reason.
4399          */
4400         __bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
4401         if (__bfq_bfqq_expire(bfqd, bfqq, reason))
4402                 /* bfqq is gone, no more actions on it */
4403                 return;
4404 
4405         /* mark bfqq as waiting a request only if a bic still points to it */
4406         if (!bfq_bfqq_busy(bfqq) &&
4407             reason != BFQQE_BUDGET_TIMEOUT &&
4408             reason != BFQQE_BUDGET_EXHAUSTED) {
4409                 bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
4410                 /*
4411                  * Not setting service to 0, because, if the next rq
4412                  * arrives in time, the queue will go on receiving
4413                  * service with this same budget (as if it never expired)
4414                  */
4415         } else
4416                 entity->service = 0;
4417 
4418         /*
4419          * Reset the received-service counter for every parent entity.
4420          * Differently from what happens with bfqq->entity.service,
4421          * the resetting of this counter never needs to be postponed
4422          * for parent entities. In fact, in case bfqq may have a
4423          * chance to go on being served using the last, partially
4424          * consumed budget, bfqq->entity.service needs to be kept,
4425          * because if bfqq then actually goes on being served using
4426          * the same budget, the last value of bfqq->entity.service is
4427          * needed to properly decrement bfqq->entity.budget by the
4428          * portion already consumed. In contrast, it is not necessary
4429          * to keep entity->service for parent entities too, because
4430          * the bubble up of the new value of bfqq->entity.budget will
4431          * make sure that the budgets of parent entities are correct,
4432          * even in case bfqq and thus parent entities go on receiving
4433          * service with the same budget.
4434          */
4435         entity = entity->parent;
4436         for_each_entity(entity)
4437                 entity->service = 0;
4438 }
4439 
4440 /*
4441  * Budget timeout is not implemented through a dedicated timer, but
4442  * just checked on request arrivals and completions, as well as on
4443  * idle timer expirations.
4444  */
4445 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
4446 {
4447         return time_is_before_eq_jiffies(bfqq->budget_timeout);
4448 }
4449 
4450 /*
4451  * If we expire a queue that is actively waiting (i.e., with the
4452  * device idled) for the arrival of a new request, then we may incur
4453  * the timestamp misalignment problem described in the body of the
4454  * function __bfq_activate_entity. Hence we return true only if this
4455  * condition does not hold, or if the queue is slow enough to deserve
4456  * only to be kicked off for preserving a high throughput.
4457  */
4458 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
4459 {
4460         bfq_log_bfqq(bfqq->bfqd, bfqq,
4461                 "may_budget_timeout: wait_request %d left %d timeout %d",
4462                 bfq_bfqq_wait_request(bfqq),
4463                         bfq_bfqq_budget_left(bfqq) >=  bfqq->entity.budget / 3,
4464                 bfq_bfqq_budget_timeout(bfqq));
4465 
4466         return (!bfq_bfqq_wait_request(bfqq) ||
4467                 bfq_bfqq_budget_left(bfqq) >=  bfqq->entity.budget / 3)
4468                 &&
4469                 bfq_bfqq_budget_timeout(bfqq);
4470 }
4471 
4472 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
4473                                              struct bfq_queue *bfqq)
4474 {
4475         bool rot_without_queueing =
4476                 !blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
4477                 bfqq_sequential_and_IO_bound,
4478                 idling_boosts_thr;
4479 
4480         /* No point in idling for bfqq if it won't get requests any longer */
4481         if (unlikely(!bfqq_process_refs(bfqq)))
4482                 return false;
4483 
4484         bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
4485                 bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
4486 
4487         /*
4488          * The next variable takes into account the cases where idling
4489          * boosts the throughput.
4490          *
4491          * The value of the variable is computed considering, first, that
4492          * idling is virtually always beneficial for the throughput if:
4493          * (a) the device is not NCQ-capable and rotational, or
4494          * (b) regardless of the presence of NCQ, the device is rotational and
4495          *     the request pattern for bfqq is I/O-bound and sequential, or
4496          * (c) regardless of whether it is rotational, the device is
4497          *     not NCQ-capable and the request pattern for bfqq is
4498          *     I/O-bound and sequential.
4499          *
4500          * Secondly, and in contrast to the above item (b), idling an
4501          * NCQ-capable flash-based device would not boost the
4502          * throughput even with sequential I/O; rather it would lower
4503          * the throughput in proportion to how fast the device
4504          * is. Accordingly, the next variable is true if any of the
4505          * above conditions (a), (b) or (c) is true, and, in
4506          * particular, happens to be false if bfqd is an NCQ-capable
4507          * flash-based device.
4508          */
4509         idling_boosts_thr = rot_without_queueing ||
4510                 ((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
4511                  bfqq_sequential_and_IO_bound);
4512 
4513         /*
4514          * The return value of this function is equal to that of
4515          * idling_boosts_thr, unless a special case holds. In this
4516          * special case, described below, idling may cause problems to
4517          * weight-raised queues.
4518          *
4519          * When the request pool is saturated (e.g., in the presence
4520          * of write hogs), if the processes associated with
4521          * non-weight-raised queues ask for requests at a lower rate,
4522          * then processes associated with weight-raised queues have a
4523          * higher probability to get a request from the pool
4524          * immediately (or at least soon) when they need one. Thus
4525          * they have a higher probability to actually get a fraction
4526          * of the device throughput proportional to their high
4527          * weight. This is especially true with NCQ-capable drives,
4528          * which enqueue several requests in advance, and further
4529          * reorder internally-queued requests.
4530          *
4531          * For this reason, we force to false the return value if
4532          * there are weight-raised busy queues. In this case, and if
4533          * bfqq is not weight-raised, this guarantees that the device
4534          * is not idled for bfqq (if, instead, bfqq is weight-raised,
4535          * then idling will be guaranteed by another variable, see
4536          * below). Combined with the timestamping rules of BFQ (see
4537          * [1] for details), this behavior causes bfqq, and hence any
4538          * sync non-weight-raised queue, to get a lower number of
4539          * requests served, and thus to ask for a lower number of
4540          * requests from the request pool, before the busy
4541          * weight-raised queues get served again. This often mitigates
4542          * starvation problems in the presence of heavy write
4543          * workloads and NCQ, thereby guaranteeing a higher
4544          * application and system responsiveness in these hostile
4545          * scenarios.
4546          */
4547         return idling_boosts_thr &&
4548                 bfqd->wr_busy_queues == 0;
4549 }
4550 
4551 /*
4552  * For a queue that becomes empty, device idling is allowed only if
4553  * this function returns true for that queue. As a consequence, since
4554  * device idling plays a critical role for both throughput boosting
4555  * and service guarantees, the return value of this function plays a
4556  * critical role as well.
4557  *
4558  * In a nutshell, this function returns true only if idling is
4559  * beneficial for throughput or, even if detrimental for throughput,
4560  * idling is however necessary to preserve service guarantees (low
4561  * latency, desired throughput distribution, ...). In particular, on
4562  * NCQ-capable devices, this function tries to return false, so as to
4563  * help keep the drives' internal queues full, whenever this helps the
4564  * device boost the throughput without causing any service-guarantee
4565  * issue.
4566  *
4567  * Most of the issues taken into account to get the return value of
4568  * this function are not trivial. We discuss these issues in the two
4569  * functions providing the main pieces of information needed by this
4570  * function.
4571  */
4572 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
4573 {
4574         struct bfq_data *bfqd = bfqq->bfqd;
4575         bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
4576 
4577         /* No point in idling for bfqq if it won't get requests any longer */
4578         if (unlikely(!bfqq_process_refs(bfqq)))
4579                 return false;
4580 
4581         if (unlikely(bfqd->strict_guarantees))
4582                 return true;
4583 
4584         /*
4585          * Idling is performed only if slice_idle > 0. In addition, we
4586          * do not idle if
4587          * (a) bfqq is async
4588          * (b) bfqq is in the idle io prio class: in this case we do
4589          * not idle because we want to minimize the bandwidth that
4590          * queues in this class can steal to higher-priority queues
4591          */
4592         if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
4593            bfq_class_idle(bfqq))
4594                 return false;
4595 
4596         idling_boosts_thr_with_no_issue =
4597                 idling_boosts_thr_without_issues(bfqd, bfqq);
4598 
4599         idling_needed_for_service_guar =
4600                 idling_needed_for_service_guarantees(bfqd, bfqq);
4601 
4602         /*
4603          * We have now the two components we need to compute the
4604          * return value of the function, which is true only if idling
4605          * either boosts the throughput (without issues), or is
4606          * necessary to preserve service guarantees.
4607          */
4608         return idling_boosts_thr_with_no_issue ||
4609                 idling_needed_for_service_guar;
4610 }
4611 
4612 /*
4613  * If the in-service queue is empty but the function bfq_better_to_idle
4614  * returns true, then:
4615  * 1) the queue must remain in service and cannot be expired, and
4616  * 2) the device must be idled to wait for the possible arrival of a new
4617  *    request for the queue.
4618  * See the comments on the function bfq_better_to_idle for the reasons
4619  * why performing device idling is the best choice to boost the throughput
4620  * and preserve service guarantees when bfq_better_to_idle itself
4621  * returns true.
4622  */
4623 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
4624 {
4625         return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
4626 }
4627 
4628 /*
4629  * This function chooses the queue from which to pick the next extra
4630  * I/O request to inject, if it finds a compatible queue. See the
4631  * comments on bfq_update_inject_limit() for details on the injection
4632  * mechanism, and for the definitions of the quantities mentioned
4633  * below.
4634  */
4635 static struct bfq_queue *
4636 bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
4637 {
4638         struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
4639         unsigned int limit = in_serv_bfqq->inject_limit;
4640         /*
4641          * If
4642          * - bfqq is not weight-raised and therefore does not carry
4643          *   time-critical I/O,
4644          * or
4645          * - regardless of whether bfqq is weight-raised, bfqq has
4646          *   however a long think time, during which it can absorb the
4647          *   effect of an appropriate number of extra I/O requests
4648          *   from other queues (see bfq_update_inject_limit for
4649          *   details on the computation of this number);
4650          * then injection can be performed without restrictions.
4651          */
4652         bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
4653                 !bfq_bfqq_has_short_ttime(in_serv_bfqq);
4654 
4655         /*
4656          * If
4657          * - the baseline total service time could not be sampled yet,
4658          *   so the inject limit happens to be still 0, and
4659          * - a lot of time has elapsed since the plugging of I/O
4660          *   dispatching started, so drive speed is being wasted
4661          *   significantly;
4662          * then temporarily raise inject limit to one request.
4663          */
4664         if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
4665             bfq_bfqq_wait_request(in_serv_bfqq) &&
4666             time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
4667                                       bfqd->bfq_slice_idle)
4668                 )
4669                 limit = 1;
4670 
4671         if (bfqd->rq_in_driver >= limit)
4672                 return NULL;
4673 
4674         /*
4675          * Linear search of the source queue for injection; but, with
4676          * a high probability, very few steps are needed to find a
4677          * candidate queue, i.e., a queue with enough budget left for
4678          * its next request. In fact:
4679          * - BFQ dynamically updates the budget of every queue so as
4680          *   to accommodate the expected backlog of the queue;
4681          * - if a queue gets all its requests dispatched as injected
4682          *   service, then the queue is removed from the active list
4683          *   (and re-added only if it gets new requests, but then it
4684          *   is assigned again enough budget for its new backlog).
4685          */
4686         list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
4687                 if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4688                     (in_serv_always_inject || bfqq->wr_coeff > 1) &&
4689                     bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4690                     bfq_bfqq_budget_left(bfqq)) {
4691                         /*
4692                          * Allow for only one large in-flight request
4693                          * on non-rotational devices, for the
4694                          * following reason. On non-rotationl drives,
4695                          * large requests take much longer than
4696                          * smaller requests to be served. In addition,
4697                          * the drive prefers to serve large requests
4698                          * w.r.t. to small ones, if it can choose. So,
4699                          * having more than one large requests queued
4700                          * in the drive may easily make the next first
4701                          * request of the in-service queue wait for so
4702                          * long to break bfqq's service guarantees. On
4703                          * the bright side, large requests let the
4704                          * drive reach a very high throughput, even if
4705                          * there is only one in-flight large request
4706                          * at a time.
4707                          */
4708                         if (blk_queue_nonrot(bfqd->queue) &&
4709                             blk_rq_sectors(bfqq->next_rq) >=
4710                             BFQQ_SECT_THR_NONROT)
4711                                 limit = min_t(unsigned int, 1, limit);
4712                         else
4713                                 limit = in_serv_bfqq->inject_limit;
4714 
4715                         if (bfqd->rq_in_driver < limit) {
4716                                 bfqd->rqs_injected = true;
4717                                 return bfqq;
4718                         }
4719                 }
4720 
4721         return NULL;
4722 }
4723 
4724 /*
4725  * Select a queue for service.  If we have a current queue in service,
4726  * check whether to continue servicing it, or retrieve and set a new one.
4727  */
4728 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
4729 {
4730         struct bfq_queue *bfqq;
4731         struct request *next_rq;
4732         enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
4733 
4734         bfqq = bfqd->in_service_queue;
4735         if (!bfqq)
4736                 goto new_queue;
4737 
4738         bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
4739 
4740         /*
4741          * Do not expire bfqq for budget timeout if bfqq may be about
4742          * to enjoy device idling. The reason why, in this case, we
4743          * prevent bfqq from expiring is the same as in the comments
4744          * on the case where bfq_bfqq_must_idle() returns true, in
4745          * bfq_completed_request().
4746          */
4747         if (bfq_may_expire_for_budg_timeout(bfqq) &&
4748             !bfq_bfqq_must_idle(bfqq))
4749                 goto expire;
4750 
4751 check_queue:
4752         /*
4753          * This loop is rarely executed more than once. Even when it
4754          * happens, it is much more convenient to re-execute this loop
4755          * than to return NULL and trigger a new dispatch to get a
4756          * request served.
4757          */
4758         next_rq = bfqq->next_rq;
4759         /*
4760          * If bfqq has requests queued and it has enough budget left to
4761          * serve them, keep the queue, otherwise expire it.
4762          */
4763         if (next_rq) {
4764                 if (bfq_serv_to_charge(next_rq, bfqq) >
4765                         bfq_bfqq_budget_left(bfqq)) {
4766                         /*
4767                          * Expire the queue for budget exhaustion,
4768                          * which makes sure that the next budget is
4769                          * enough to serve the next request, even if
4770                          * it comes from the fifo expired path.
4771                          */
4772                         reason = BFQQE_BUDGET_EXHAUSTED;
4773                         goto expire;
4774                 } else {
4775                         /*
4776                          * The idle timer may be pending because we may
4777                          * not disable disk idling even when a new request
4778                          * arrives.
4779                          */
4780                         if (bfq_bfqq_wait_request(bfqq)) {
4781                                 /*
4782                                  * If we get here: 1) at least a new request
4783                                  * has arrived but we have not disabled the
4784                                  * timer because the request was too small,
4785                                  * 2) then the block layer has unplugged
4786                                  * the device, causing the dispatch to be
4787                                  * invoked.
4788                                  *
4789                                  * Since the device is unplugged, now the
4790                                  * requests are probably large enough to
4791                                  * provide a reasonable throughput.
4792                                  * So we disable idling.
4793                                  */
4794                                 bfq_clear_bfqq_wait_request(bfqq);
4795                                 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4796                         }
4797                         goto keep_queue;
4798                 }
4799         }
4800 
4801         /*
4802          * No requests pending. However, if the in-service queue is idling
4803          * for a new request, or has requests waiting for a completion and
4804          * may idle after their completion, then keep it anyway.
4805          *
4806          * Yet, inject service from other queues if it boosts
4807          * throughput and is possible.
4808          */
4809         if (bfq_bfqq_wait_request(bfqq) ||
4810             (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
4811                 struct bfq_queue *async_bfqq =
4812                         bfqq->bic && bfqq->bic->bfqq[0] &&
4813                         bfq_bfqq_busy(bfqq->bic->bfqq[0]) &&
4814                         bfqq->bic->bfqq[0]->next_rq ?
4815                         bfqq->bic->bfqq[0] : NULL;
4816                 struct bfq_queue *blocked_bfqq =
4817                         !hlist_empty(&bfqq->woken_list) ?
4818                         container_of(bfqq->woken_list.first,
4819                                      struct bfq_queue,
4820                                      woken_list_node)
4821                         : NULL;
4822 
4823                 /*
4824                  * The next four mutually-exclusive ifs decide
4825                  * whether to try injection, and choose the queue to
4826                  * pick an I/O request from.
4827                  *
4828                  * The first if checks whether the process associated
4829                  * with bfqq has also async I/O pending. If so, it
4830                  * injects such I/O unconditionally. Injecting async
4831                  * I/O from the same process can cause no harm to the
4832                  * process. On the contrary, it can only increase
4833                  * bandwidth and reduce latency for the process.
4834                  *
4835                  * The second if checks whether there happens to be a
4836                  * non-empty waker queue for bfqq, i.e., a queue whose
4837                  * I/O needs to be completed for bfqq to receive new
4838                  * I/O. This happens, e.g., if bfqq is associated with
4839                  * a process that does some sync. A sync generates
4840                  * extra blocking I/O, which must be completed before
4841                  * the process associated with bfqq can go on with its
4842                  * I/O. If the I/O of the waker queue is not served,
4843                  * then bfqq remains empty, and no I/O is dispatched,
4844                  * until the idle timeout fires for bfqq. This is
4845                  * likely to result in lower bandwidth and higher
4846                  * latencies for bfqq, and in a severe loss of total
4847                  * throughput. The best action to take is therefore to
4848                  * serve the waker queue as soon as possible. So do it
4849                  * (without relying on the third alternative below for
4850                  * eventually serving waker_bfqq's I/O; see the last
4851                  * paragraph for further details). This systematic
4852                  * injection of I/O from the waker queue does not
4853                  * cause any delay to bfqq's I/O. On the contrary,
4854                  * next bfqq's I/O is brought forward dramatically,
4855                  * for it is not blocked for milliseconds.
4856                  *
4857                  * The third if checks whether there is a queue woken
4858                  * by bfqq, and currently with pending I/O. Such a
4859                  * woken queue does not steal bandwidth from bfqq,
4860                  * because it remains soon without I/O if bfqq is not
4861                  * served. So there is virtually no risk of loss of
4862                  * bandwidth for bfqq if this woken queue has I/O
4863                  * dispatched while bfqq is waiting for new I/O.
4864                  *
4865                  * The fourth if checks whether bfqq is a queue for
4866                  * which it is better to avoid injection. It is so if
4867                  * bfqq delivers more throughput when served without
4868                  * any further I/O from other queues in the middle, or
4869                  * if the service times of bfqq's I/O requests both
4870                  * count more than overall throughput, and may be
4871                  * easily increased by injection (this happens if bfqq
4872                  * has a short think time). If none of these
4873                  * conditions holds, then a candidate queue for
4874                  * injection is looked for through
4875                  * bfq_choose_bfqq_for_injection(). Note that the
4876                  * latter may return NULL (for example if the inject
4877                  * limit for bfqq is currently 0).
4878                  *
4879                  * NOTE: motivation for the second alternative
4880                  *
4881                  * Thanks to the way the inject limit is updated in
4882                  * bfq_update_has_short_ttime(), it is rather likely
4883                  * that, if I/O is being plugged for bfqq and the
4884                  * waker queue has pending I/O requests that are
4885                  * blocking bfqq's I/O, then the fourth alternative
4886                  * above lets the waker queue get served before the
4887                  * I/O-plugging timeout fires. So one may deem the
4888                  * second alternative superfluous. It is not, because
4889                  * the fourth alternative may be way less effective in
4890                  * case of a synchronization. For two main
4891                  * reasons. First, throughput may be low because the
4892                  * inject limit may be too low to guarantee the same
4893                  * amount of injected I/O, from the waker queue or
4894                  * other queues, that the second alternative
4895                  * guarantees (the second alternative unconditionally
4896                  * injects a pending I/O request of the waker queue
4897                  * for each bfq_dispatch_request()). Second, with the
4898                  * fourth alternative, the duration of the plugging,
4899                  * i.e., the time before bfqq finally receives new I/O,
4900                  * may not be minimized, because the waker queue may
4901                  * happen to be served only after other queues.
4902                  */
4903                 if (async_bfqq &&
4904                     icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
4905                     bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
4906                     bfq_bfqq_budget_left(async_bfqq))
4907                         bfqq = bfqq->bic->bfqq[0];
4908                 else if (bfqq->waker_bfqq &&
4909                            bfq_bfqq_busy(bfqq->waker_bfqq) &&
4910                            bfqq->waker_bfqq->next_rq &&
4911                            bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
4912                                               bfqq->waker_bfqq) <=
4913                            bfq_bfqq_budget_left(bfqq->waker_bfqq)
4914                         )
4915                         bfqq = bfqq->waker_bfqq;
4916                 else if (blocked_bfqq &&
4917                            bfq_bfqq_busy(blocked_bfqq) &&
4918                            blocked_bfqq->next_rq &&
4919                            bfq_serv_to_charge(blocked_bfqq->next_rq,
4920                                               blocked_bfqq) <=
4921                            bfq_bfqq_budget_left(blocked_bfqq)
4922                         )
4923                         bfqq = blocked_bfqq;
4924                 else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
4925                          (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
4926                           !bfq_bfqq_has_short_ttime(bfqq)))
4927                         bfqq = bfq_choose_bfqq_for_injection(bfqd);
4928                 else
4929                         bfqq = NULL;
4930 
4931                 goto keep_queue;
4932         }
4933 
4934         reason = BFQQE_NO_MORE_REQUESTS;
4935 expire:
4936         bfq_bfqq_expire(bfqd, bfqq, false, reason);
4937 new_queue:
4938         bfqq = bfq_set_in_service_queue(bfqd);
4939         if (bfqq) {
4940                 bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
4941                 goto check_queue;
4942         }
4943 keep_queue:
4944         if (bfqq)
4945                 bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
4946         else
4947                 bfq_log(bfqd, "select_queue: no queue returned");
4948 
4949         return bfqq;
4950 }
4951 
4952 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4953 {
4954         struct bfq_entity *entity = &bfqq->entity;
4955 
4956         if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
4957                 bfq_log_bfqq(bfqd, bfqq,
4958                         "raising period dur %u/%u msec, old coeff %u, w %d(%d)",
4959                         jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
4960                         jiffies_to_msecs(bfqq->wr_cur_max_time),
4961                         bfqq->wr_coeff,
4962                         bfqq->entity.weight, bfqq->entity.orig_weight);
4963 
4964                 if (entity->prio_changed)
4965                         bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
4966 
4967                 /*
4968                  * If the queue was activated in a burst, or too much
4969                  * time has elapsed from the beginning of this
4970                  * weight-raising period, then end weight raising.
4971                  */
4972                 if (bfq_bfqq_in_large_burst(bfqq))
4973                         bfq_bfqq_end_wr(bfqq);
4974                 else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
4975                                                 bfqq->wr_cur_max_time)) {
4976                         if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
4977                         time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
4978                                                bfq_wr_duration(bfqd))) {
4979                                 /*
4980                                  * Either in interactive weight
4981                                  * raising, or in soft_rt weight
4982                                  * raising with the
4983                                  * interactive-weight-raising period
4984                                  * elapsed (so no switch back to
4985                                  * interactive weight raising).
4986                                  */
4987                                 bfq_bfqq_end_wr(bfqq);
4988                         } else { /*
4989                                   * soft_rt finishing while still in
4990                                   * interactive period, switch back to
4991                                   * interactive weight raising
4992                                   */
4993                                 switch_back_to_interactive_wr(bfqq, bfqd);
4994                                 bfqq->entity.prio_changed = 1;
4995                         }
4996                 }
4997                 if (bfqq->wr_coeff > 1 &&
4998                     bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
4999                     bfqq->service_from_wr > max_service_from_wr) {
5000                         /* see comments on max_service_from_wr */
5001                         bfq_bfqq_end_wr(bfqq);
5002                 }
5003         }
5004         /*
5005          * To improve latency (for this or other queues), immediately
5006          * update weight both if it must be raised and if it must be
5007          * lowered. Since, entity may be on some active tree here, and
5008          * might have a pending change of its ioprio class, invoke
5009          * next function with the last parameter unset (see the
5010          * comments on the function).
5011          */
5012         if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
5013                 __bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
5014                                                 entity, false);
5015 }
5016 
5017 /*
5018  * Dispatch next request from bfqq.
5019  */
5020 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
5021                                                  struct bfq_queue *bfqq)
5022 {
5023         struct request *rq = bfqq->next_rq;
5024         unsigned long service_to_charge;
5025 
5026         service_to_charge = bfq_serv_to_charge(rq, bfqq);
5027 
5028         bfq_bfqq_served(bfqq, service_to_charge);
5029 
5030         if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
5031                 bfqd->wait_dispatch = false;
5032                 bfqd->waited_rq = rq;
5033         }
5034 
5035         bfq_dispatch_remove(bfqd->queue, rq);
5036 
5037         if (bfqq != bfqd->in_service_queue)
5038                 goto return_rq;
5039 
5040         /*
5041          * If weight raising has to terminate for bfqq, then next
5042          * function causes an immediate update of bfqq's weight,
5043          * without waiting for next activation. As a consequence, on
5044          * expiration, bfqq will be timestamped as if has never been
5045          * weight-raised during this service slot, even if it has
5046          * received part or even most of the service as a
5047          * weight-raised queue. This inflates bfqq's timestamps, which
5048          * is beneficial, as bfqq is then more willing to leave the
5049          * device immediately to possible other weight-raised queues.
5050          */
5051         bfq_update_wr_data(bfqd, bfqq);
5052 
5053         /*
5054          * Expire bfqq, pretending that its budget expired, if bfqq
5055          * belongs to CLASS_IDLE and other queues are waiting for
5056          * service.
5057          */
5058         if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
5059                 goto return_rq;
5060 
5061         bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
5062 
5063 return_rq:
5064         return rq;
5065 }
5066 
5067 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
5068 {
5069         struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5070 
5071         /*
5072          * Avoiding lock: a race on bfqd->queued should cause at
5073          * most a call to dispatch for nothing
5074          */
5075         return !list_empty_careful(&bfqd->dispatch) ||
5076                 READ_ONCE(bfqd->queued);
5077 }
5078 
5079 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5080 {
5081         struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5082         struct request *rq = NULL;
5083         struct bfq_queue *bfqq = NULL;
5084 
5085         if (!list_empty(&bfqd->dispatch)) {
5086                 rq = list_first_entry(&bfqd->dispatch, struct request,
5087                                       queuelist);
5088                 list_del_init(&rq->queuelist);
5089 
5090                 bfqq = RQ_BFQQ(rq);
5091 
5092                 if (bfqq) {
5093                         /*
5094                          * Increment counters here, because this
5095                          * dispatch does not follow the standard
5096                          * dispatch flow (where counters are
5097                          * incremented)
5098                          */
5099                         bfqq->dispatched++;
5100 
5101                         goto inc_in_driver_start_rq;
5102                 }
5103 
5104                 /*
5105                  * We exploit the bfq_finish_requeue_request hook to
5106                  * decrement rq_in_driver, but
5107                  * bfq_finish_requeue_request will not be invoked on
5108                  * this request. So, to avoid unbalance, just start
5109                  * this request, without incrementing rq_in_driver. As
5110                  * a negative consequence, rq_in_driver is deceptively
5111                  * lower than it should be while this request is in
5112                  * service. This may cause bfq_schedule_dispatch to be
5113                  * invoked uselessly.
5114                  *
5115                  * As for implementing an exact solution, the
5116                  * bfq_finish_requeue_request hook, if defined, is
5117                  * probably invoked also on this request. So, by
5118                  * exploiting this hook, we could 1) increment
5119                  * rq_in_driver here, and 2) decrement it in
5120                  * bfq_finish_requeue_request. Such a solution would
5121                  * let the value of the counter be always accurate,
5122                  * but it would entail using an extra interface
5123                  * function. This cost seems higher than the benefit,
5124                  * being the frequency of non-elevator-private
5125                  * requests very low.
5126                  */
5127                 goto start_rq;
5128         }
5129 
5130         bfq_log(bfqd, "dispatch requests: %d busy queues",
5131                 bfq_tot_busy_queues(bfqd));
5132 
5133         if (bfq_tot_busy_queues(bfqd) == 0)
5134                 goto exit;
5135 
5136         /*
5137          * Force device to serve one request at a time if
5138          * strict_guarantees is true. Forcing this service scheme is
5139          * currently the ONLY way to guarantee that the request
5140          * service order enforced by the scheduler is respected by a
5141          * queueing device. Otherwise the device is free even to make
5142          * some unlucky request wait for as long as the device
5143          * wishes.
5144          *
5145          * Of course, serving one request at a time may cause loss of
5146          * throughput.
5147          */
5148         if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
5149                 goto exit;
5150 
5151         bfqq = bfq_select_queue(bfqd);
5152         if (!bfqq)
5153                 goto exit;
5154 
5155         rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
5156 
5157         if (rq) {
5158 inc_in_driver_start_rq:
5159                 bfqd->rq_in_driver++;
5160 start_rq:
5161                 rq->rq_flags |= RQF_STARTED;
5162         }
5163 exit:
5164         return rq;
5165 }
5166 
5167 #ifdef CONFIG_BFQ_CGROUP_DEBUG
5168 static void bfq_update_dispatch_stats(struct request_queue *q,
5169                                       struct request *rq,
5170                                       struct bfq_queue *in_serv_queue,
5171                                       bool idle_timer_disabled)
5172 {
5173         struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
5174 
5175         if (!idle_timer_disabled && !bfqq)
5176                 return;
5177 
5178         /*
5179          * rq and bfqq are guaranteed to exist until this function
5180          * ends, for the following reasons. First, rq can be
5181          * dispatched to the device, and then can be completed and
5182          * freed, only after this function ends. Second, rq cannot be
5183          * merged (and thus freed because of a merge) any longer,
5184          * because it has already started. Thus rq cannot be freed
5185          * before this function ends, and, since rq has a reference to
5186          * bfqq, the same guarantee holds for bfqq too.
5187          *
5188          * In addition, the following queue lock guarantees that
5189          * bfqq_group(bfqq) exists as well.
5190          */
5191         spin_lock_irq(&q->queue_lock);
5192         if (idle_timer_disabled)
5193                 /*
5194                  * Since the idle timer has been disabled,
5195                  * in_serv_queue contained some request when
5196                  * __bfq_dispatch_request was invoked above, which
5197                  * implies that rq was picked exactly from
5198                  * in_serv_queue. Thus in_serv_queue == bfqq, and is
5199                  * therefore guaranteed to exist because of the above
5200                  * arguments.
5201                  */
5202                 bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
5203         if (bfqq) {
5204                 struct bfq_group *bfqg = bfqq_group(bfqq);
5205 
5206                 bfqg_stats_update_avg_queue_size(bfqg);
5207                 bfqg_stats_set_start_empty_time(bfqg);
5208                 bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
5209         }
5210         spin_unlock_irq(&q->queue_lock);
5211 }
5212 #else
5213 static inline void bfq_update_dispatch_stats(struct request_queue *q,
5214                                              struct request *rq,
5215                                              struct bfq_queue *in_serv_queue,
5216                                              bool idle_timer_disabled) {}
5217 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
5218 
5219 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5220 {
5221         struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5222         struct request *rq;
5223         struct bfq_queue *in_serv_queue;
5224         bool waiting_rq, idle_timer_disabled = false;
5225 
5226         spin_lock_irq(&bfqd->lock);
5227 
5228         in_serv_queue = bfqd->in_service_queue;
5229         waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
5230 
5231         rq = __bfq_dispatch_request(hctx);
5232         if (in_serv_queue == bfqd->in_service_queue) {
5233                 idle_timer_disabled =
5234                         waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
5235         }
5236 
5237         spin_unlock_irq(&bfqd->lock);
5238         bfq_update_dispatch_stats(hctx->queue, rq,
5239                         idle_timer_disabled ? in_serv_queue : NULL,
5240                                 idle_timer_disabled);
5241 
5242         return rq;
5243 }
5244 
5245 /*
5246  * Task holds one reference to the queue, dropped when task exits.  Each rq
5247  * in-flight on this queue also holds a reference, dropped when rq is freed.
5248  *
5249  * Scheduler lock must be held here. Recall not to use bfqq after calling
5250  * this function on it.
5251  */
5252 void bfq_put_queue(struct bfq_queue *bfqq)
5253 {
5254         struct bfq_queue *item;
5255         struct hlist_node *n;
5256         struct bfq_group *bfqg = bfqq_group(bfqq);
5257 
5258         if (bfqq->bfqd)
5259                 bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
5260                              bfqq, bfqq->ref);
5261 
5262         bfqq->ref--;
5263         if (bfqq->ref)
5264                 return;
5265 
5266         if (!hlist_unhashed(&bfqq->burst_list_node)) {
5267                 hlist_del_init(&bfqq->burst_list_node);
5268                 /*
5269                  * Decrement also burst size after the removal, if the
5270                  * process associated with bfqq is exiting, and thus
5271                  * does not contribute to the burst any longer. This
5272                  * decrement helps filter out false positives of large
5273                  * bursts, when some short-lived process (often due to
5274                  * the execution of commands by some service) happens
5275                  * to start and exit while a complex application is
5276                  * starting, and thus spawning several processes that
5277                  * do I/O (and that *must not* be treated as a large
5278                  * burst, see comments on bfq_handle_burst).
5279                  *
5280                  * In particular, the decrement is performed only if:
5281                  * 1) bfqq is not a merged queue, because, if it is,
5282                  * then this free of bfqq is not triggered by the exit
5283                  * of the process bfqq is associated with, but exactly
5284                  * by the fact that bfqq has just been merged.
5285                  * 2) burst_size is greater than 0, to handle
5286                  * unbalanced decrements. Unbalanced decrements may
5287                  * happen in te following case: bfqq is inserted into
5288                  * the current burst list--without incrementing
5289                  * bust_size--because of a split, but the current
5290                  * burst list is not the burst list bfqq belonged to
5291                  * (see comments on the case of a split in
5292                  * bfq_set_request).
5293                  */
5294                 if (bfqq->bic && bfqq->bfqd->burst_size > 0)
5295                         bfqq->bfqd->burst_size--;
5296         }
5297 
5298         /*
5299          * bfqq does not exist any longer, so it cannot be woken by
5300          * any other queue, and cannot wake any other queue. Then bfqq
5301          * must be removed from the woken list of its possible waker
5302          * queue, and all queues in the woken list of bfqq must stop
5303          * having a waker queue. Strictly speaking, these updates
5304          * should be performed when bfqq remains with no I/O source
5305          * attached to it, which happens before bfqq gets freed. In
5306          * particular, this happens when the last process associated
5307          * with bfqq exits or gets associated with a different
5308          * queue. However, both events lead to bfqq being freed soon,
5309          * and dangling references would come out only after bfqq gets
5310          * freed. So these updates are done here, as a simple and safe
5311          * way to handle all cases.
5312          */
5313         /* remove bfqq from woken list */
5314         if (!hlist_unhashed(&bfqq->woken_list_node))
5315                 hlist_del_init(&bfqq->woken_list_node);
5316 
5317         /* reset waker for all queues in woken list */
5318         hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
5319                                   woken_list_node) {
5320                 item->waker_bfqq = NULL;
5321                 hlist_del_init(&item->woken_list_node);
5322         }
5323 
5324         if (bfqq->bfqd && bfqq->bfqd->last_completed_rq_bfqq == bfqq)
5325                 bfqq->bfqd->last_completed_rq_bfqq = NULL;
5326 
5327         kmem_cache_free(bfq_pool, bfqq);
5328         bfqg_and_blkg_put(bfqg);
5329 }
5330 
5331 static void bfq_put_stable_ref(struct bfq_queue *bfqq)
5332 {
5333         bfqq->stable_ref--;
5334         bfq_put_queue(bfqq);
5335 }
5336 
5337 void bfq_put_cooperator(struct bfq_queue *bfqq)
5338 {
5339         struct bfq_queue *__bfqq, *next;
5340 
5341         /*
5342          * If this queue was scheduled to merge with another queue, be
5343          * sure to drop the reference taken on that queue (and others in
5344          * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
5345          */
5346         __bfqq = bfqq->new_bfqq;
5347         while (__bfqq) {
5348                 if (__bfqq == bfqq)
5349                         break;
5350                 next = __bfqq->new_bfqq;
5351                 bfq_put_queue(__bfqq);
5352                 __bfqq = next;
5353         }
5354 }
5355 
5356 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
5357 {
5358         if (bfqq == bfqd->in_service_queue) {
5359                 __bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT);
5360                 bfq_schedule_dispatch(bfqd);
5361         }
5362 
5363         bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
5364 
5365         bfq_put_cooperator(bfqq);
5366 
5367         bfq_release_process_ref(bfqd, bfqq);
5368 }
5369 
5370 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
5371 {
5372         struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
5373         struct bfq_data *bfqd;
5374 
5375         if (bfqq)
5376                 bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
5377 
5378         if (bfqq && bfqd) {
5379                 unsigned long flags;
5380 
5381                 spin_lock_irqsave(&bfqd->lock, flags);
5382                 bfqq->bic = NULL;
5383                 bfq_exit_bfqq(bfqd, bfqq);
5384                 bic_set_bfqq(bic, NULL, is_sync);
5385                 spin_unlock_irqrestore(&bfqd->lock, flags);
5386         }
5387 }
5388 
5389 static void bfq_exit_icq(struct io_cq *icq)
5390 {
5391         struct bfq_io_cq *bic = icq_to_bic(icq);
5392 
5393         if (bic->stable_merge_bfqq) {
5394                 struct bfq_data *bfqd = bic->stable_merge_bfqq->bfqd;
5395 
5396                 /*
5397                  * bfqd is NULL if scheduler already exited, and in
5398                  * that case this is the last time bfqq is accessed.
5399                  */
5400                 if (bfqd) {
5401                         unsigned long flags;
5402 
5403                         spin_lock_irqsave(&bfqd->lock, flags);
5404                         bfq_put_stable_ref(bic->stable_merge_bfqq);
5405                         spin_unlock_irqrestore(&bfqd->lock, flags);
5406                 } else {
5407                         bfq_put_stable_ref(bic->stable_merge_bfqq);
5408                 }
5409         }
5410 
5411         bfq_exit_icq_bfqq(bic, true);
5412         bfq_exit_icq_bfqq(bic, false);
5413 }
5414 
5415 /*
5416  * Update the entity prio values; note that the new values will not
5417  * be used until the next (re)activation.
5418  */
5419 static void
5420 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
5421 {
5422         struct task_struct *tsk = current;
5423         int ioprio_class;
5424         struct bfq_data *bfqd = bfqq->bfqd;
5425 
5426         if (!bfqd)
5427                 return;
5428 
5429         ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5430         switch (ioprio_class) {
5431         default:
5432                 pr_err("bdi %s: bfq: bad prio class %d\n",
5433                         bdi_dev_name(bfqq->bfqd->queue->disk->bdi),
5434                         ioprio_class);
5435                 fallthrough;
5436         case IOPRIO_CLASS_NONE:
5437                 /*
5438                  * No prio set, inherit CPU scheduling settings.
5439                  */
5440                 bfqq->new_ioprio = task_nice_ioprio(tsk);
5441                 bfqq->new_ioprio_class = task_nice_ioclass(tsk);
5442                 break;
5443         case IOPRIO_CLASS_RT:
5444                 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5445                 bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
5446                 break;
5447         case IOPRIO_CLASS_BE:
5448                 bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5449                 bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
5450                 break;
5451         case IOPRIO_CLASS_IDLE:
5452                 bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
5453                 bfqq->new_ioprio = 7;
5454                 break;
5455         }
5456 
5457         if (bfqq->new_ioprio >= IOPRIO_NR_LEVELS) {
5458                 pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
5459                         bfqq->new_ioprio);
5460                 bfqq->new_ioprio = IOPRIO_NR_LEVELS - 1;
5461         }
5462 
5463         bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
5464         bfq_log_bfqq(bfqd, bfqq, "new_ioprio %d new_weight %d",
5465                      bfqq->new_ioprio, bfqq->entity.new_weight);
5466         bfqq->entity.prio_changed = 1;
5467 }
5468 
5469 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5470                                        struct bio *bio, bool is_sync,
5471                                        struct bfq_io_cq *bic,
5472                                        bool respawn);
5473 
5474 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
5475 {
5476         struct bfq_data *bfqd = bic_to_bfqd(bic);
5477         struct bfq_queue *bfqq;
5478         int ioprio = bic->icq.ioc->ioprio;
5479 
5480         /*
5481          * This condition may trigger on a newly created bic, be sure to
5482          * drop the lock before returning.
5483          */
5484         if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
5485                 return;
5486 
5487         bic->ioprio = ioprio;
5488 
5489         bfqq = bic_to_bfqq(bic, false);
5490         if (bfqq) {
5491                 bfq_release_process_ref(bfqd, bfqq);
5492                 bfqq = bfq_get_queue(bfqd, bio, false, bic, true);
5493                 bic_set_bfqq(bic, bfqq, false);
5494         }
5495 
5496         bfqq = bic_to_bfqq(bic, true);
5497         if (bfqq)
5498                 bfq_set_next_ioprio_data(bfqq, bic);
5499 }
5500 
5501 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5502                           struct bfq_io_cq *bic, pid_t pid, int is_sync)
5503 {
5504         u64 now_ns = ktime_get_ns();
5505 
5506         RB_CLEAR_NODE(&bfqq->entity.rb_node);
5507         INIT_LIST_HEAD(&bfqq->fifo);
5508         INIT_HLIST_NODE(&bfqq->burst_list_node);
5509         INIT_HLIST_NODE(&bfqq->woken_list_node);
5510         INIT_HLIST_HEAD(&bfqq->woken_list);
5511 
5512         bfqq->ref = 0;
5513         bfqq->bfqd = bfqd;
5514 
5515         if (bic)
5516                 bfq_set_next_ioprio_data(bfqq, bic);
5517 
5518         if (is_sync) {
5519                 /*
5520                  * No need to mark as has_short_ttime if in
5521                  * idle_class, because no device idling is performed
5522                  * for queues in idle class
5523                  */
5524                 if (!bfq_class_idle(bfqq))
5525                         /* tentatively mark as has_short_ttime */
5526                         bfq_mark_bfqq_has_short_ttime(bfqq);
5527                 bfq_mark_bfqq_sync(bfqq);
5528                 bfq_mark_bfqq_just_created(bfqq);
5529         } else
5530                 bfq_clear_bfqq_sync(bfqq);
5531 
5532         /* set end request to minus infinity from now */
5533         bfqq->ttime.last_end_request = now_ns + 1;
5534 
5535         bfqq->creation_time = jiffies;
5536 
5537         bfqq->io_start_time = now_ns;
5538 
5539         bfq_mark_bfqq_IO_bound(bfqq);
5540 
5541         bfqq->pid = pid;
5542 
5543         /* Tentative initial value to trade off between thr and lat */
5544         bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
5545         bfqq->budget_timeout = bfq_smallest_from_now();
5546 
5547         bfqq->wr_coeff = 1;
5548         bfqq->last_wr_start_finish = jiffies;
5549         bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
5550         bfqq->split_time = bfq_smallest_from_now();
5551 
5552         /*
5553          * To not forget the possibly high bandwidth consumed by a
5554          * process/queue in the recent past,
5555          * bfq_bfqq_softrt_next_start() returns a value at least equal
5556          * to the current value of bfqq->soft_rt_next_start (see
5557          * comments on bfq_bfqq_softrt_next_start).  Set
5558          * soft_rt_next_start to now, to mean that bfqq has consumed
5559          * no bandwidth so far.
5560          */
5561         bfqq->soft_rt_next_start = jiffies;
5562 
5563         /* first request is almost certainly seeky */
5564         bfqq->seek_history = 1;
5565 }
5566 
5567 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
5568                                                struct bfq_group *bfqg,
5569                                                int ioprio_class, int ioprio)
5570 {
5571         switch (ioprio_class) {
5572         case IOPRIO_CLASS_RT:
5573                 return &bfqg->async_bfqq[0][ioprio];
5574         case IOPRIO_CLASS_NONE:
5575                 ioprio = IOPRIO_BE_NORM;
5576                 fallthrough;
5577         case IOPRIO_CLASS_BE:
5578                 return &bfqg->async_bfqq[1][ioprio];
5579         case IOPRIO_CLASS_IDLE:
5580                 return &bfqg->async_idle_bfqq;
5581         default:
5582                 return NULL;
5583         }
5584 }
5585 
5586 static struct bfq_queue *
5587 bfq_do_early_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5588                           struct bfq_io_cq *bic,
5589                           struct bfq_queue *last_bfqq_created)
5590 {
5591         struct bfq_queue *new_bfqq =
5592                 bfq_setup_merge(bfqq, last_bfqq_created);
5593 
5594         if (!new_bfqq)
5595                 return bfqq;
5596 
5597         if (new_bfqq->bic)
5598                 new_bfqq->bic->stably_merged = true;
5599         bic->stably_merged = true;
5600 
5601         /*
5602          * Reusing merge functions. This implies that
5603          * bfqq->bic must be set too, for
5604          * bfq_merge_bfqqs to correctly save bfqq's
5605          * state before killing it.
5606          */
5607         bfqq->bic = bic;
5608         bfq_merge_bfqqs(bfqd, bic, bfqq, new_bfqq);
5609 
5610         return new_bfqq;
5611 }
5612 
5613 /*
5614  * Many throughput-sensitive workloads are made of several parallel
5615  * I/O flows, with all flows generated by the same application, or
5616  * more generically by the same task (e.g., system boot). The most
5617  * counterproductive action with these workloads is plugging I/O
5618  * dispatch when one of the bfq_queues associated with these flows
5619  * remains temporarily empty.
5620  *
5621  * To avoid this plugging, BFQ has been using a burst-handling
5622  * mechanism for years now. This mechanism has proven effective for
5623  * throughput, and not detrimental for service guarantees. The
5624  * following function pushes this mechanism a little bit further,
5625  * basing on the following two facts.
5626  *
5627  * First, all the I/O flows of a the same application or task
5628  * contribute to the execution/completion of that common application
5629  * or task. So the performance figures that matter are total
5630  * throughput of the flows and task-wide I/O latency.  In particular,
5631  * these flows do not need to be protected from each other, in terms
5632  * of individual bandwidth or latency.
5633  *
5634  * Second, the above fact holds regardless of the number of flows.
5635  *
5636  * Putting these two facts together, this commits merges stably the
5637  * bfq_queues associated with these I/O flows, i.e., with the
5638  * processes that generate these IO/ flows, regardless of how many the
5639  * involved processes are.
5640  *
5641  * To decide whether a set of bfq_queues is actually associated with
5642  * the I/O flows of a common application or task, and to merge these
5643  * queues stably, this function operates as follows: given a bfq_queue,
5644  * say Q2, currently being created, and the last bfq_queue, say Q1,
5645  * created before Q2, Q2 is merged stably with Q1 if
5646  * - very little time has elapsed since when Q1 was created
5647  * - Q2 has the same ioprio as Q1
5648  * - Q2 belongs to the same group as Q1
5649  *
5650  * Merging bfq_queues also reduces scheduling overhead. A fio test
5651  * with ten random readers on /dev/nullb shows a throughput boost of
5652  * 40%, with a quadcore. Since BFQ's execution time amounts to ~50% of
5653  * the total per-request processing time, the above throughput boost
5654  * implies that BFQ's overhead is reduced by more than 50%.
5655  *
5656  * This new mechanism most certainly obsoletes the current
5657  * burst-handling heuristics. We keep those heuristics for the moment.
5658  */
5659 static struct bfq_queue *bfq_do_or_sched_stable_merge(struct bfq_data *bfqd,
5660                                                       struct bfq_queue *bfqq,
5661                                                       struct bfq_io_cq *bic)
5662 {
5663         struct bfq_queue **source_bfqq = bfqq->entity.parent ?
5664                 &bfqq->entity.parent->last_bfqq_created :
5665                 &bfqd->last_bfqq_created;
5666 
5667         struct bfq_queue *last_bfqq_created = *source_bfqq;
5668 
5669         /*
5670          * If last_bfqq_created has not been set yet, then init it. If
5671          * it has been set already, but too long ago, then move it
5672          * forward to bfqq. Finally, move also if bfqq belongs to a
5673          * different group than last_bfqq_created, or if bfqq has a
5674          * different ioprio or ioprio_class. If none of these
5675          * conditions holds true, then try an early stable merge or
5676          * schedule a delayed stable merge.
5677          *
5678          * A delayed merge is scheduled (instead of performing an
5679          * early merge), in case bfqq might soon prove to be more
5680          * throughput-beneficial if not merged. Currently this is
5681          * possible only if bfqd is rotational with no queueing. For
5682          * such a drive, not merging bfqq is better for throughput if
5683          * bfqq happens to contain sequential I/O. So, we wait a
5684          * little bit for enough I/O to flow through bfqq. After that,
5685          * if such an I/O is sequential, then the merge is
5686          * canceled. Otherwise the merge is finally performed.
5687          */
5688         if (!last_bfqq_created ||
5689             time_before(last_bfqq_created->creation_time +
5690                         msecs_to_jiffies(bfq_activation_stable_merging),
5691                         bfqq->creation_time) ||
5692                 bfqq->entity.parent != last_bfqq_created->entity.parent ||
5693                 bfqq->ioprio != last_bfqq_created->ioprio ||
5694                 bfqq->ioprio_class != last_bfqq_created->ioprio_class)
5695                 *source_bfqq = bfqq;
5696         else if (time_after_eq(last_bfqq_created->creation_time +
5697                                  bfqd->bfq_burst_interval,
5698                                  bfqq->creation_time)) {
5699                 if (likely(bfqd->nonrot_with_queueing))
5700                         /*
5701                          * With this type of drive, leaving
5702                          * bfqq alone may provide no
5703                          * throughput benefits compared with
5704                          * merging bfqq. So merge bfqq now.
5705                          */
5706                         bfqq = bfq_do_early_stable_merge(bfqd, bfqq,
5707                                                          bic,
5708                                                          last_bfqq_created);
5709                 else { /* schedule tentative stable merge */
5710                         /*
5711                          * get reference on last_bfqq_created,
5712                          * to prevent it from being freed,
5713                          * until we decide whether to merge
5714                          */
5715                         last_bfqq_created->ref++;
5716                         /*
5717                          * need to keep track of stable refs, to
5718                          * compute process refs correctly
5719                          */
5720                         last_bfqq_created->stable_ref++;
5721                         /*
5722                          * Record the bfqq to merge to.
5723                          */
5724                         bic->stable_merge_bfqq = last_bfqq_created;
5725                 }
5726         }
5727 
5728         return bfqq;
5729 }
5730 
5731 
5732 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5733                                        struct bio *bio, bool is_sync,
5734                                        struct bfq_io_cq *bic,
5735                                        bool respawn)
5736 {
5737         const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5738         const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5739         struct bfq_queue **async_bfqq = NULL;
5740         struct bfq_queue *bfqq;
5741         struct bfq_group *bfqg;
5742 
5743         bfqg = bfq_bio_bfqg(bfqd, bio);
5744         if (!is_sync) {
5745                 async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
5746                                                   ioprio);
5747                 bfqq = *async_bfqq;
5748                 if (bfqq)
5749                         goto out;
5750         }
5751 
5752         bfqq = kmem_cache_alloc_node(bfq_pool,
5753                                      GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
5754                                      bfqd->queue->node);
5755 
5756         if (bfqq) {
5757                 bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
5758                               is_sync);
5759                 bfq_init_entity(&bfqq->entity, bfqg);
5760                 bfq_log_bfqq(bfqd, bfqq, "allocated");
5761         } else {
5762                 bfqq = &bfqd->oom_bfqq;
5763                 bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
5764                 goto out;
5765         }
5766 
5767         /*
5768          * Pin the queue now that it's allocated, scheduler exit will
5769          * prune it.
5770          */
5771         if (async_bfqq) {
5772                 bfqq->ref++; /*
5773                               * Extra group reference, w.r.t. sync
5774                               * queue. This extra reference is removed
5775                               * only if bfqq->bfqg disappears, to
5776                               * guarantee that this queue is not freed
5777                               * until its group goes away.
5778                               */
5779                 bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
5780                              bfqq, bfqq->ref);
5781                 *async_bfqq = bfqq;
5782         }
5783 
5784 out:
5785         bfqq->ref++; /* get a process reference to this queue */
5786 
5787         if (bfqq != &bfqd->oom_bfqq && is_sync && !respawn)
5788                 bfqq = bfq_do_or_sched_stable_merge(bfqd, bfqq, bic);
5789         return bfqq;
5790 }
5791 
5792 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
5793                                     struct bfq_queue *bfqq)
5794 {
5795         struct bfq_ttime *ttime = &bfqq->ttime;
5796         u64 elapsed;
5797 
5798         /*
5799          * We are really interested in how long it takes for the queue to
5800          * become busy when there is no outstanding IO for this queue. So
5801          * ignore cases when the bfq queue has already IO queued.
5802          */
5803         if (bfqq->dispatched || bfq_bfqq_busy(bfqq))
5804                 return;
5805         elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
5806         elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
5807 
5808         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
5809         ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed,  8);
5810         ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
5811                                      ttime->ttime_samples);
5812 }
5813 
5814 static void
5815 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5816                        struct request *rq)
5817 {
5818         bfqq->seek_history <<= 1;
5819         bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
5820 
5821         if (bfqq->wr_coeff > 1 &&
5822             bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
5823             BFQQ_TOTALLY_SEEKY(bfqq)) {
5824                 if (time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
5825                                            bfq_wr_duration(bfqd))) {
5826                         /*
5827                          * In soft_rt weight raising with the
5828                          * interactive-weight-raising period
5829                          * elapsed (so no switch back to
5830                          * interactive weight raising).
5831                          */
5832                         bfq_bfqq_end_wr(bfqq);
5833                 } else { /*
5834                           * stopping soft_rt weight raising
5835                           * while still in interactive period,
5836                           * switch back to interactive weight
5837                           * raising
5838                           */
5839                         switch_back_to_interactive_wr(bfqq, bfqd);
5840                         bfqq->entity.prio_changed = 1;
5841                 }
5842         }
5843 }
5844 
5845 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
5846                                        struct bfq_queue *bfqq,
5847                                        struct bfq_io_cq *bic)
5848 {
5849         bool has_short_ttime = true, state_changed;
5850 
5851         /*
5852          * No need to update has_short_ttime if bfqq is async or in
5853          * idle io prio class, or if bfq_slice_idle is zero, because
5854          * no device idling is performed for bfqq in this case.
5855          */
5856         if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
5857             bfqd->bfq_slice_idle == 0)
5858                 return;
5859 
5860         /* Idle window just restored, statistics are meaningless. */
5861         if (time_is_after_eq_jiffies(bfqq->split_time +
5862                                      bfqd->bfq_wr_min_idle_time))
5863                 return;
5864 
5865         /* Think time is infinite if no process is linked to
5866          * bfqq. Otherwise check average think time to decide whether
5867          * to mark as has_short_ttime. To this goal, compare average
5868          * think time with half the I/O-plugging timeout.
5869          */
5870         if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
5871             (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
5872              bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle>>1))
5873                 has_short_ttime = false;
5874 
5875         state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
5876 
5877         if (has_short_ttime)
5878                 bfq_mark_bfqq_has_short_ttime(bfqq);
5879         else
5880                 bfq_clear_bfqq_has_short_ttime(bfqq);
5881 
5882         /*
5883          * Until the base value for the total service time gets
5884          * finally computed for bfqq, the inject limit does depend on
5885          * the think-time state (short|long). In particular, the limit
5886          * is 0 or 1 if the think time is deemed, respectively, as
5887          * short or long (details in the comments in
5888          * bfq_update_inject_limit()). Accordingly, the next
5889          * instructions reset the inject limit if the think-time state
5890          * has changed and the above base value is still to be
5891          * computed.
5892          *
5893          * However, the reset is performed only if more than 100 ms
5894          * have elapsed since the last update of the inject limit, or
5895          * (inclusive) if the change is from short to long think
5896          * time. The reason for this waiting is as follows.
5897          *
5898          * bfqq may have a long think time because of a
5899          * synchronization with some other queue, i.e., because the
5900          * I/O of some other queue may need to be completed for bfqq
5901          * to receive new I/O. Details in the comments on the choice
5902          * of the queue for injection in bfq_select_queue().
5903          *
5904          * As stressed in those comments, if such a synchronization is
5905          * actually in place, then, without injection on bfqq, the
5906          * blocking I/O cannot happen to served while bfqq is in
5907          * service. As a consequence, if bfqq is granted
5908          * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
5909          * is dispatched, until the idle timeout fires. This is likely
5910          * to result in lower bandwidth and higher latencies for bfqq,
5911          * and in a severe loss of total throughput.
5912          *
5913          * On the opposite end, a non-zero inject limit may allow the
5914          * I/O that blocks bfqq to be executed soon, and therefore
5915          * bfqq to receive new I/O soon.
5916          *
5917          * But, if the blocking gets actually eliminated, then the
5918          * next think-time sample for bfqq may be very low. This in
5919          * turn may cause bfqq's think time to be deemed
5920          * short. Without the 100 ms barrier, this new state change
5921          * would cause the body of the next if to be executed
5922          * immediately. But this would set to 0 the inject
5923          * limit. Without injection, the blocking I/O would cause the
5924          * think time of bfqq to become long again, and therefore the
5925          * inject limit to be raised again, and so on. The only effect
5926          * of such a steady oscillation between the two think-time
5927          * states would be to prevent effective injection on bfqq.
5928          *
5929          * In contrast, if the inject limit is not reset during such a
5930          * long time interval as 100 ms, then the number of short
5931          * think time samples can grow significantly before the reset
5932          * is performed. As a consequence, the think time state can
5933          * become stable before the reset. Therefore there will be no
5934          * state change when the 100 ms elapse, and no reset of the
5935          * inject limit. The inject limit remains steadily equal to 1
5936          * both during and after the 100 ms. So injection can be
5937          * performed at all times, and throughput gets boosted.
5938          *
5939          * An inject limit equal to 1 is however in conflict, in
5940          * general, with the fact that the think time of bfqq is
5941          * short, because injection may be likely to delay bfqq's I/O
5942          * (as explained in the comments in
5943          * bfq_update_inject_limit()). But this does not happen in
5944          * this special case, because bfqq's low think time is due to
5945          * an effective handling of a synchronization, through
5946          * injection. In this special case, bfqq's I/O does not get
5947          * delayed by injection; on the contrary, bfqq's I/O is
5948          * brought forward, because it is not blocked for
5949          * milliseconds.
5950          *
5951          * In addition, serving the blocking I/O much sooner, and much
5952          * more frequently than once per I/O-plugging timeout, makes
5953          * it much quicker to detect a waker queue (the concept of
5954          * waker queue is defined in the comments in
5955          * bfq_add_request()). This makes it possible to start sooner
5956          * to boost throughput more effectively, by injecting the I/O
5957          * of the waker queue unconditionally on every
5958          * bfq_dispatch_request().
5959          *
5960          * One last, important benefit of not resetting the inject
5961          * limit before 100 ms is that, during this time interval, the
5962          * base value for the total service time is likely to get
5963          * finally computed for bfqq, freeing the inject limit from
5964          * its relation with the think time.
5965          */
5966         if (state_changed && bfqq->last_serv_time_ns == 0 &&
5967             (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
5968                                       msecs_to_jiffies(100)) ||
5969              !has_short_ttime))
5970                 bfq_reset_inject_limit(bfqd, bfqq);
5971 }
5972 
5973 /*
5974  * Called when a new fs request (rq) is added to bfqq.  Check if there's
5975  * something we should do about it.
5976  */
5977 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5978                             struct request *rq)
5979 {
5980         if (rq->cmd_flags & REQ_META)
5981                 bfqq->meta_pending++;
5982 
5983         bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
5984 
5985         if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
5986                 bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
5987                                  blk_rq_sectors(rq) < 32;
5988                 bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
5989 
5990                 /*
5991                  * There is just this request queued: if
5992                  * - the request is small, and
5993                  * - we are idling to boost throughput, and
5994                  * - the queue is not to be expired,
5995                  * then just exit.
5996                  *
5997                  * In this way, if the device is being idled to wait
5998                  * for a new request from the in-service queue, we
5999                  * avoid unplugging the device and committing the
6000                  * device to serve just a small request. In contrast
6001                  * we wait for the block layer to decide when to
6002                  * unplug the device: hopefully, new requests will be
6003                  * merged to this one quickly, then the device will be
6004                  * unplugged and larger requests will be dispatched.
6005                  */
6006                 if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
6007                     !budget_timeout)
6008                         return;
6009 
6010                 /*
6011                  * A large enough request arrived, or idling is being
6012                  * performed to preserve service guarantees, or
6013                  * finally the queue is to be expired: in all these
6014                  * cases disk idling is to be stopped, so clear
6015                  * wait_request flag and reset timer.
6016                  */
6017                 bfq_clear_bfqq_wait_request(bfqq);
6018                 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
6019 
6020                 /*
6021                  * The queue is not empty, because a new request just
6022                  * arrived. Hence we can safely expire the queue, in
6023                  * case of budget timeout, without risking that the
6024                  * timestamps of the queue are not updated correctly.
6025                  * See [1] for more details.
6026                  */
6027                 if (budget_timeout)
6028                         bfq_bfqq_expire(bfqd, bfqq, false,
6029                                         BFQQE_BUDGET_TIMEOUT);
6030         }
6031 }
6032 
6033 static void bfqq_request_allocated(struct bfq_queue *bfqq)
6034 {
6035         struct bfq_entity *entity = &bfqq->entity;
6036 
6037         for_each_entity(entity)
6038                 entity->allocated++;
6039 }
6040 
6041 static void bfqq_request_freed(struct bfq_queue *bfqq)
6042 {
6043         struct bfq_entity *entity = &bfqq->entity;
6044 
6045         for_each_entity(entity)
6046                 entity->allocated--;
6047 }
6048 
6049 /* returns true if it causes the idle timer to be disabled */
6050 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
6051 {
6052         struct bfq_queue *bfqq = RQ_BFQQ(rq),
6053                 *new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true,
6054                                                  RQ_BIC(rq));
6055         bool waiting, idle_timer_disabled = false;
6056 
6057         if (new_bfqq) {
6058                 /*
6059                  * Release the request's reference to the old bfqq
6060                  * and make sure one is taken to the shared queue.
6061                  */
6062                 bfqq_request_allocated(new_bfqq);
6063                 bfqq_request_freed(bfqq);
6064                 new_bfqq->ref++;
6065                 /*
6066                  * If the bic associated with the process
6067                  * issuing this request still points to bfqq
6068                  * (and thus has not been already redirected
6069                  * to new_bfqq or even some other bfq_queue),
6070                  * then complete the merge and redirect it to
6071                  * new_bfqq.
6072                  */
6073                 if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
6074                         bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
6075                                         bfqq, new_bfqq);
6076 
6077                 bfq_clear_bfqq_just_created(bfqq);
6078                 /*
6079                  * rq is about to be enqueued into new_bfqq,
6080                  * release rq reference on bfqq
6081                  */
6082                 bfq_put_queue(bfqq);
6083                 rq->elv.priv[1] = new_bfqq;
6084                 bfqq = new_bfqq;
6085         }
6086 
6087         bfq_update_io_thinktime(bfqd, bfqq);
6088         bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
6089         bfq_update_io_seektime(bfqd, bfqq, rq);
6090 
6091         waiting = bfqq && bfq_bfqq_wait_request(bfqq);
6092         bfq_add_request(rq);
6093         idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
6094 
6095         rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
6096         list_add_tail(&rq->queuelist, &bfqq->fifo);
6097 
6098         bfq_rq_enqueued(bfqd, bfqq, rq);
6099 
6100         return idle_timer_disabled;
6101 }
6102 
6103 #ifdef CONFIG_BFQ_CGROUP_DEBUG
6104 static void bfq_update_insert_stats(struct request_queue *q,
6105                                     struct bfq_queue *bfqq,
6106                                     bool idle_timer_disabled,
6107                                     unsigned int cmd_flags)
6108 {
6109         if (!bfqq)
6110                 return;
6111 
6112         /*
6113          * bfqq still exists, because it can disappear only after
6114          * either it is merged with another queue, or the process it
6115          * is associated with exits. But both actions must be taken by
6116          * the same process currently executing this flow of
6117          * instructions.
6118          *
6119          * In addition, the following queue lock guarantees that
6120          * bfqq_group(bfqq) exists as well.
6121          */
6122         spin_lock_irq(&q->queue_lock);
6123         bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
6124         if (idle_timer_disabled)
6125                 bfqg_stats_update_idle_time(bfqq_group(bfqq));
6126         spin_unlock_irq(&q->queue_lock);
6127 }
6128 #else
6129 static inline void bfq_update_insert_stats(struct request_queue *q,
6130                                            struct bfq_queue *bfqq,
6131                                            bool idle_timer_disabled,
6132                                            unsigned int cmd_flags) {}
6133 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
6134 
6135 static struct bfq_queue *bfq_init_rq(struct request *rq);
6136 
6137 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
6138                                bool at_head)
6139 {
6140         struct request_queue *q = hctx->queue;
6141         struct bfq_data *bfqd = q->elevator->elevator_data;
6142         struct bfq_queue *bfqq;
6143         bool idle_timer_disabled = false;
6144         unsigned int cmd_flags;
6145         LIST_HEAD(free);
6146 
6147 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6148         if (!cgroup_subsys_on_dfl(io_cgrp_subsys) && rq->bio)
6149                 bfqg_stats_update_legacy_io(q, rq);
6150 #endif
6151         spin_lock_irq(&bfqd->lock);
6152         bfqq = bfq_init_rq(rq);
6153         if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
6154                 spin_unlock_irq(&bfqd->lock);
6155                 blk_mq_free_requests(&free);
6156                 return;
6157         }
6158 
6159         trace_block_rq_insert(rq);
6160 
6161         if (!bfqq || at_head) {
6162                 if (at_head)
6163                         list_add(&rq->queuelist, &bfqd->dispatch);
6164                 else
6165                         list_add_tail(&rq->queuelist, &bfqd->dispatch);
6166         } else {
6167                 idle_timer_disabled = __bfq_insert_request(bfqd, rq);
6168                 /*
6169                  * Update bfqq, because, if a queue merge has occurred
6170                  * in __bfq_insert_request, then rq has been
6171                  * redirected into a new queue.
6172                  */
6173                 bfqq = RQ_BFQQ(rq);
6174 
6175                 if (rq_mergeable(rq)) {
6176                         elv_rqhash_add(q, rq);
6177                         if (!q->last_merge)
6178                                 q->last_merge = rq;
6179                 }
6180         }
6181 
6182         /*
6183          * Cache cmd_flags before releasing scheduler lock, because rq
6184          * may disappear afterwards (for example, because of a request
6185          * merge).
6186          */
6187         cmd_flags = rq->cmd_flags;
6188         spin_unlock_irq(&bfqd->lock);
6189 
6190         bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
6191                                 cmd_flags);
6192 }
6193 
6194 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
6195                                 struct list_head *list, bool at_head)
6196 {
6197         while (!list_empty(list)) {
6198                 struct request *rq;
6199 
6200                 rq = list_first_entry(list, struct request, queuelist);
6201                 list_del_init(&rq->queuelist);
6202                 bfq_insert_request(hctx, rq, at_head);
6203         }
6204 }
6205 
6206 static void bfq_update_hw_tag(struct bfq_data *bfqd)
6207 {
6208         struct bfq_queue *bfqq = bfqd->in_service_queue;
6209 
6210         bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
6211                                        bfqd->rq_in_driver);
6212 
6213         if (bfqd->hw_tag == 1)
6214                 return;
6215 
6216         /*
6217          * This sample is valid if the number of outstanding requests
6218          * is large enough to allow a queueing behavior.  Note that the
6219          * sum is not exact, as it's not taking into account deactivated
6220          * requests.
6221          */
6222         if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
6223                 return;
6224 
6225         /*
6226          * If active queue hasn't enough requests and can idle, bfq might not
6227          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
6228          * case
6229          */
6230         if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
6231             bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
6232             BFQ_HW_QUEUE_THRESHOLD &&
6233             bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
6234                 return;
6235 
6236         if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
6237                 return;
6238 
6239         bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
6240         bfqd->max_rq_in_driver = 0;
6241         bfqd->hw_tag_samples = 0;
6242 
6243         bfqd->nonrot_with_queueing =
6244                 blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
6245 }
6246 
6247 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
6248 {
6249         u64 now_ns;
6250         u32 delta_us;
6251 
6252         bfq_update_hw_tag(bfqd);
6253 
6254         bfqd->rq_in_driver--;
6255         bfqq->dispatched--;
6256 
6257         if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
6258                 /*
6259                  * Set budget_timeout (which we overload to store the
6260                  * time at which the queue remains with no backlog and
6261                  * no outstanding request; used by the weight-raising
6262                  * mechanism).
6263                  */
6264                 bfqq->budget_timeout = jiffies;
6265 
6266                 bfq_weights_tree_remove(bfqd, bfqq);
6267         }
6268 
6269         now_ns = ktime_get_ns();
6270 
6271         bfqq->ttime.last_end_request = now_ns;
6272 
6273         /*
6274          * Using us instead of ns, to get a reasonable precision in
6275          * computing rate in next check.
6276          */
6277         delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
6278 
6279         /*
6280          * If the request took rather long to complete, and, according
6281          * to the maximum request size recorded, this completion latency
6282          * implies that the request was certainly served at a very low
6283          * rate (less than 1M sectors/sec), then the whole observation
6284          * interval that lasts up to this time instant cannot be a
6285          * valid time interval for computing a new peak rate.  Invoke
6286          * bfq_update_rate_reset to have the following three steps
6287          * taken:
6288          * - close the observation interval at the last (previous)
6289          *   request dispatch or completion
6290          * - compute rate, if possible, for that observation interval
6291          * - reset to zero samples, which will trigger a proper
6292          *   re-initialization of the observation interval on next
6293          *   dispatch
6294          */
6295         if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
6296            (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
6297                         1UL<<(BFQ_RATE_SHIFT - 10))
6298                 bfq_update_rate_reset(bfqd, NULL);
6299         bfqd->last_completion = now_ns;
6300         /*
6301          * Shared queues are likely to receive I/O at a high
6302          * rate. This may deceptively let them be considered as wakers
6303          * of other queues. But a false waker will unjustly steal
6304          * bandwidth to its supposedly woken queue. So considering
6305          * also shared queues in the waking mechanism may cause more
6306          * control troubles than throughput benefits. Then reset
6307          * last_completed_rq_bfqq if bfqq is a shared queue.
6308          */
6309         if (!bfq_bfqq_coop(bfqq))
6310                 bfqd->last_completed_rq_bfqq = bfqq;
6311         else
6312                 bfqd->last_completed_rq_bfqq = NULL;
6313 
6314         /*
6315          * If we are waiting to discover whether the request pattern
6316          * of the task associated with the queue is actually
6317          * isochronous, and both requisites for this condition to hold
6318          * are now satisfied, then compute soft_rt_next_start (see the
6319          * comments on the function bfq_bfqq_softrt_next_start()). We
6320          * do not compute soft_rt_next_start if bfqq is in interactive
6321          * weight raising (see the comments in bfq_bfqq_expire() for
6322          * an explanation). We schedule this delayed update when bfqq
6323          * expires, if it still has in-flight requests.
6324          */
6325         if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
6326             RB_EMPTY_ROOT(&bfqq->sort_list) &&
6327             bfqq->wr_coeff != bfqd->bfq_wr_coeff)
6328                 bfqq->soft_rt_next_start =
6329                         bfq_bfqq_softrt_next_start(bfqd, bfqq);
6330 
6331         /*
6332          * If this is the in-service queue, check if it needs to be expired,
6333          * or if we want to idle in case it has no pending requests.
6334          */
6335         if (bfqd->in_service_queue == bfqq) {
6336                 if (bfq_bfqq_must_idle(bfqq)) {
6337                         if (bfqq->dispatched == 0)
6338                                 bfq_arm_slice_timer(bfqd);
6339                         /*
6340                          * If we get here, we do not expire bfqq, even
6341                          * if bfqq was in budget timeout or had no
6342                          * more requests (as controlled in the next
6343                          * conditional instructions). The reason for
6344                          * not expiring bfqq is as follows.
6345                          *
6346                          * Here bfqq->dispatched > 0 holds, but
6347                          * bfq_bfqq_must_idle() returned true. This
6348                          * implies that, even if no request arrives
6349                          * for bfqq before bfqq->dispatched reaches 0,
6350                          * bfqq will, however, not be expired on the
6351                          * completion event that causes bfqq->dispatch
6352                          * to reach zero. In contrast, on this event,
6353                          * bfqq will start enjoying device idling
6354                          * (I/O-dispatch plugging).
6355                          *
6356                          * But, if we expired bfqq here, bfqq would
6357                          * not have the chance to enjoy device idling
6358                          * when bfqq->dispatched finally reaches
6359                          * zero. This would expose bfqq to violation
6360                          * of its reserved service guarantees.
6361                          */
6362                         return;
6363                 } else if (bfq_may_expire_for_budg_timeout(bfqq))
6364                         bfq_bfqq_expire(bfqd, bfqq, false,
6365                                         BFQQE_BUDGET_TIMEOUT);
6366                 else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
6367                          (bfqq->dispatched == 0 ||
6368                           !bfq_better_to_idle(bfqq)))
6369                         bfq_bfqq_expire(bfqd, bfqq, false,
6370                                         BFQQE_NO_MORE_REQUESTS);
6371         }
6372 
6373         if (!bfqd->rq_in_driver)
6374                 bfq_schedule_dispatch(bfqd);
6375 }
6376 
6377 /*
6378  * The processes associated with bfqq may happen to generate their
6379  * cumulative I/O at a lower rate than the rate at which the device
6380  * could serve the same I/O. This is rather probable, e.g., if only
6381  * one process is associated with bfqq and the device is an SSD. It
6382  * results in bfqq becoming often empty while in service. In this
6383  * respect, if BFQ is allowed to switch to another queue when bfqq
6384  * remains empty, then the device goes on being fed with I/O requests,
6385  * and the throughput is not affected. In contrast, if BFQ is not
6386  * allowed to switch to another queue---because bfqq is sync and
6387  * I/O-dispatch needs to be plugged while bfqq is temporarily
6388  * empty---then, during the service of bfqq, there will be frequent
6389  * "service holes", i.e., time intervals during which bfqq gets empty
6390  * and the device can only consume the I/O already queued in its
6391  * hardware queues. During service holes, the device may even get to
6392  * remaining idle. In the end, during the service of bfqq, the device
6393  * is driven at a lower speed than the one it can reach with the kind
6394  * of I/O flowing through bfqq.
6395  *
6396  * To counter this loss of throughput, BFQ implements a "request
6397  * injection mechanism", which tries to fill the above service holes
6398  * with I/O requests taken from other queues. The hard part in this
6399  * mechanism is finding the right amount of I/O to inject, so as to
6400  * both boost throughput and not break bfqq's bandwidth and latency
6401  * guarantees. In this respect, the mechanism maintains a per-queue
6402  * inject limit, computed as below. While bfqq is empty, the injection
6403  * mechanism dispatches extra I/O requests only until the total number
6404  * of I/O requests in flight---i.e., already dispatched but not yet
6405  * completed---remains lower than this limit.
6406  *
6407  * A first definition comes in handy to introduce the algorithm by
6408  * which the inject limit is computed.  We define as first request for
6409  * bfqq, an I/O request for bfqq that arrives while bfqq is in
6410  * service, and causes bfqq to switch from empty to non-empty. The
6411  * algorithm updates the limit as a function of the effect of
6412  * injection on the service times of only the first requests of
6413  * bfqq. The reason for this restriction is that these are the
6414  * requests whose service time is affected most, because they are the
6415  * first to arrive after injection possibly occurred.
6416  *
6417  * To evaluate the effect of injection, the algorithm measures the
6418  * "total service time" of first requests. We define as total service
6419  * time of an I/O request, the time that elapses since when the
6420  * request is enqueued into bfqq, to when it is completed. This
6421  * quantity allows the whole effect of injection to be measured. It is
6422  * easy to see why. Suppose that some requests of other queues are
6423  * actually injected while bfqq is empty, and that a new request R
6424  * then arrives for bfqq. If the device does start to serve all or
6425  * part of the injected requests during the service hole, then,
6426  * because of this extra service, it may delay the next invocation of
6427  * the dispatch hook of BFQ. Then, even after R gets eventually
6428  * dispatched, the device may delay the actual service of R if it is
6429  * still busy serving the extra requests, or if it decides to serve,
6430  * before R, some extra request still present in its queues. As a
6431  * conclusion, the cumulative extra delay caused by injection can be
6432  * easily evaluated by just comparing the total service time of first
6433  * requests with and without injection.
6434  *
6435  * The limit-update algorithm works as follows. On the arrival of a
6436  * first request of bfqq, the algorithm measures the total time of the
6437  * request only if one of the three cases below holds, and, for each
6438  * case, it updates the limit as described below:
6439  *
6440  * (1) If there is no in-flight request. This gives a baseline for the
6441  *     total service time of the requests of bfqq. If the baseline has
6442  *     not been computed yet, then, after computing it, the limit is
6443  *     set to 1, to start boosting throughput, and to prepare the
6444  *     ground for the next case. If the baseline has already been
6445  *     computed, then it is updated, in case it results to be lower
6446  *     than the previous value.
6447  *
6448  * (2) If the limit is higher than 0 and there are in-flight
6449  *     requests. By comparing the total service time in this case with
6450  *     the above baseline, it is possible to know at which extent the
6451  *     current value of the limit is inflating the total service
6452  *     time. If the inflation is below a certain threshold, then bfqq
6453  *     is assumed to be suffering from no perceivable loss of its
6454  *     service guarantees, and the limit is even tentatively
6455  *     increased. If the inflation is above the threshold, then the
6456  *     limit is decreased. Due to the lack of any hysteresis, this
6457  *     logic makes the limit oscillate even in steady workload
6458  *     conditions. Yet we opted for it, because it is fast in reaching
6459  *     the best value for the limit, as a function of the current I/O
6460  *     workload. To reduce oscillations, this step is disabled for a
6461  *     short time interval after the limit happens to be decreased.
6462  *
6463  * (3) Periodically, after resetting the limit, to make sure that the
6464  *     limit eventually drops in case the workload changes. This is
6465  *     needed because, after the limit has gone safely up for a
6466  *     certain workload, it is impossible to guess whether the
6467  *     baseline total service time may have changed, without measuring
6468  *     it again without injection. A more effective version of this
6469  *     step might be to just sample the baseline, by interrupting
6470  *     injection only once, and then to reset/lower the limit only if
6471  *     the total service time with the current limit does happen to be
6472  *     too large.
6473  *
6474  * More details on each step are provided in the comments on the
6475  * pieces of code that implement these steps: the branch handling the
6476  * transition from empty to non empty in bfq_add_request(), the branch
6477  * handling injection in bfq_select_queue(), and the function
6478  * bfq_choose_bfqq_for_injection(). These comments also explain some
6479  * exceptions, made by the injection mechanism in some special cases.
6480  */
6481 static void bfq_update_inject_limit(struct bfq_data *bfqd,
6482                                     struct bfq_queue *bfqq)
6483 {
6484         u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns;
6485         unsigned int old_limit = bfqq->inject_limit;
6486 
6487         if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) {
6488                 u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
6489 
6490                 if (tot_time_ns >= threshold && old_limit > 0) {
6491                         bfqq->inject_limit--;
6492                         bfqq->decrease_time_jif = jiffies;
6493                 } else if (tot_time_ns < threshold &&
6494                            old_limit <= bfqd->max_rq_in_driver)
6495                         bfqq->inject_limit++;
6496         }
6497 
6498         /*
6499          * Either we still have to compute the base value for the
6500          * total service time, and there seem to be the right
6501          * conditions to do it, or we can lower the last base value
6502          * computed.
6503          *
6504          * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
6505          * request in flight, because this function is in the code
6506          * path that handles the completion of a request of bfqq, and,
6507          * in particular, this function is executed before
6508          * bfqd->rq_in_driver is decremented in such a code path.
6509          */
6510         if ((bfqq->last_serv_time_ns == 0 && bfqd->rq_in_driver == 1) ||
6511             tot_time_ns < bfqq->last_serv_time_ns) {
6512                 if (bfqq->last_serv_time_ns == 0) {
6513                         /*
6514                          * Now we certainly have a base value: make sure we
6515                          * start trying injection.
6516                          */
6517                         bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
6518                 }
6519                 bfqq->last_serv_time_ns = tot_time_ns;
6520         } else if (!bfqd->rqs_injected && bfqd->rq_in_driver == 1)
6521                 /*
6522                  * No I/O injected and no request still in service in
6523                  * the drive: these are the exact conditions for
6524                  * computing the base value of the total service time
6525                  * for bfqq. So let's update this value, because it is
6526                  * rather variable. For example, it varies if the size
6527                  * or the spatial locality of the I/O requests in bfqq
6528                  * change.
6529                  */
6530                 bfqq->last_serv_time_ns = tot_time_ns;
6531 
6532 
6533         /* update complete, not waiting for any request completion any longer */
6534         bfqd->waited_rq = NULL;
6535         bfqd->rqs_injected = false;
6536 }
6537 
6538 /*
6539  * Handle either a requeue or a finish for rq. The things to do are
6540  * the same in both cases: all references to rq are to be dropped. In
6541  * particular, rq is considered completed from the point of view of
6542  * the scheduler.
6543  */
6544 static void bfq_finish_requeue_request(struct request *rq)
6545 {
6546         struct bfq_queue *bfqq = RQ_BFQQ(rq);
6547         struct bfq_data *bfqd;
6548         unsigned long flags;
6549 
6550         /*
6551          * rq either is not associated with any icq, or is an already
6552          * requeued request that has not (yet) been re-inserted into
6553          * a bfq_queue.
6554          */
6555         if (!rq->elv.icq || !bfqq)
6556                 return;
6557 
6558         bfqd = bfqq->bfqd;
6559 
6560         if (rq->rq_flags & RQF_STARTED)
6561                 bfqg_stats_update_completion(bfqq_group(bfqq),
6562                                              rq->start_time_ns,
6563                                              rq->io_start_time_ns,
6564                                              rq->cmd_flags);
6565 
6566         spin_lock_irqsave(&bfqd->lock, flags);
6567         if (likely(rq->rq_flags & RQF_STARTED)) {
6568                 if (rq == bfqd->waited_rq)
6569                         bfq_update_inject_limit(bfqd, bfqq);
6570 
6571                 bfq_completed_request(bfqq, bfqd);
6572         }
6573         bfqq_request_freed(bfqq);
6574         bfq_put_queue(bfqq);
6575         RQ_BIC(rq)->requests--;
6576         spin_unlock_irqrestore(&bfqd->lock, flags);
6577 
6578         /*
6579          * Reset private fields. In case of a requeue, this allows
6580          * this function to correctly do nothing if it is spuriously
6581          * invoked again on this same request (see the check at the
6582          * beginning of the function). Probably, a better general
6583          * design would be to prevent blk-mq from invoking the requeue
6584          * or finish hooks of an elevator, for a request that is not
6585          * referred by that elevator.
6586          *
6587          * Resetting the following fields would break the
6588          * request-insertion logic if rq is re-inserted into a bfq
6589          * internal queue, without a re-preparation. Here we assume
6590          * that re-insertions of requeued requests, without
6591          * re-preparation, can happen only for pass_through or at_head
6592          * requests (which are not re-inserted into bfq internal
6593          * queues).
6594          */
6595         rq->elv.priv[0] = NULL;
6596         rq->elv.priv[1] = NULL;
6597 }
6598 
6599 static void bfq_finish_request(struct request *rq)
6600 {
6601         bfq_finish_requeue_request(rq);
6602 
6603         if (rq->elv.icq) {
6604                 put_io_context(rq->elv.icq->ioc);
6605                 rq->elv.icq = NULL;
6606         }
6607 }
6608 
6609 /*
6610  * Removes the association between the current task and bfqq, assuming
6611  * that bic points to the bfq iocontext of the task.
6612  * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
6613  * was the last process referring to that bfqq.
6614  */
6615 static struct bfq_queue *
6616 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
6617 {
6618         bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
6619 
6620         if (bfqq_process_refs(bfqq) == 1) {
6621                 bfqq->pid = current->pid;
6622                 bfq_clear_bfqq_coop(bfqq);
6623                 bfq_clear_bfqq_split_coop(bfqq);
6624                 return bfqq;
6625         }
6626 
6627         bic_set_bfqq(bic, NULL, 1);
6628 
6629         bfq_put_cooperator(bfqq);
6630 
6631         bfq_release_process_ref(bfqq->bfqd, bfqq);
6632         return NULL;
6633 }
6634 
6635 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
6636                                                    struct bfq_io_cq *bic,
6637                                                    struct bio *bio,
6638                                                    bool split, bool is_sync,
6639                                                    bool *new_queue)
6640 {
6641         struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
6642 
6643         if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
6644                 return bfqq;
6645 
6646         if (new_queue)
6647                 *new_queue = true;
6648 
6649         if (bfqq)
6650                 bfq_put_queue(bfqq);
6651         bfqq = bfq_get_queue(bfqd, bio, is_sync, bic, split);
6652 
6653         bic_set_bfqq(bic, bfqq, is_sync);
6654         if (split && is_sync) {
6655                 if ((bic->was_in_burst_list && bfqd->large_burst) ||
6656                     bic->saved_in_large_burst)
6657                         bfq_mark_bfqq_in_large_burst(bfqq);
6658                 else {
6659                         bfq_clear_bfqq_in_large_burst(bfqq);
6660                         if (bic->was_in_burst_list)
6661                                 /*
6662                                  * If bfqq was in the current
6663                                  * burst list before being
6664                                  * merged, then we have to add
6665                                  * it back. And we do not need
6666                                  * to increase burst_size, as
6667                                  * we did not decrement
6668                                  * burst_size when we removed
6669                                  * bfqq from the burst list as
6670                                  * a consequence of a merge
6671                                  * (see comments in
6672                                  * bfq_put_queue). In this
6673                                  * respect, it would be rather
6674                                  * costly to know whether the
6675                                  * current burst list is still
6676                                  * the same burst list from
6677                                  * which bfqq was removed on
6678                                  * the merge. To avoid this
6679                                  * cost, if bfqq was in a
6680                                  * burst list, then we add
6681                                  * bfqq to the current burst
6682                                  * list without any further
6683                                  * check. This can cause
6684                                  * inappropriate insertions,
6685                                  * but rarely enough to not
6686                                  * harm the detection of large
6687                                  * bursts significantly.
6688                                  */
6689                                 hlist_add_head(&bfqq->burst_list_node,
6690                                                &bfqd->burst_list);
6691                 }
6692                 bfqq->split_time = jiffies;
6693         }
6694 
6695         return bfqq;
6696 }
6697 
6698 /*
6699  * Only reset private fields. The actual request preparation will be
6700  * performed by bfq_init_rq, when rq is either inserted or merged. See
6701  * comments on bfq_init_rq for the reason behind this delayed
6702  * preparation.
6703  */
6704 static void bfq_prepare_request(struct request *rq)
6705 {
6706         rq->elv.icq = ioc_find_get_icq(rq->q);
6707 
6708         /*
6709          * Regardless of whether we have an icq attached, we have to
6710          * clear the scheduler pointers, as they might point to
6711          * previously allocated bic/bfqq structs.
6712          */
6713         rq->elv.priv[0] = rq->elv.priv[1] = NULL;
6714 }
6715 
6716 /*
6717  * If needed, init rq, allocate bfq data structures associated with
6718  * rq, and increment reference counters in the destination bfq_queue
6719  * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6720  * not associated with any bfq_queue.
6721  *
6722  * This function is invoked by the functions that perform rq insertion
6723  * or merging. One may have expected the above preparation operations
6724  * to be performed in bfq_prepare_request, and not delayed to when rq
6725  * is inserted or merged. The rationale behind this delayed
6726  * preparation is that, after the prepare_request hook is invoked for
6727  * rq, rq may still be transformed into a request with no icq, i.e., a
6728  * request not associated with any queue. No bfq hook is invoked to
6729  * signal this transformation. As a consequence, should these
6730  * preparation operations be performed when the prepare_request hook
6731  * is invoked, and should rq be transformed one moment later, bfq
6732  * would end up in an inconsistent state, because it would have
6733  * incremented some queue counters for an rq destined to
6734  * transformation, without any chance to correctly lower these
6735  * counters back. In contrast, no transformation can still happen for
6736  * rq after rq has been inserted or merged. So, it is safe to execute
6737  * these preparation operations when rq is finally inserted or merged.
6738  */
6739 static struct bfq_queue *bfq_init_rq(struct request *rq)
6740 {
6741         struct request_queue *q = rq->q;
6742         struct bio *bio = rq->bio;
6743         struct bfq_data *bfqd = q->elevator->elevator_data;
6744         struct bfq_io_cq *bic;
6745         const int is_sync = rq_is_sync(rq);
6746         struct bfq_queue *bfqq;
6747         bool new_queue = false;
6748         bool bfqq_already_existing = false, split = false;
6749 
6750         if (unlikely(!rq->elv.icq))
6751                 return NULL;
6752 
6753         /*
6754          * Assuming that elv.priv[1] is set only if everything is set
6755          * for this rq. This holds true, because this function is
6756          * invoked only for insertion or merging, and, after such
6757          * events, a request cannot be manipulated any longer before
6758          * being removed from bfq.
6759          */
6760         if (rq->elv.priv[1])
6761                 return rq->elv.priv[1];
6762 
6763         bic = icq_to_bic(rq->elv.icq);
6764 
6765         bfq_check_ioprio_change(bic, bio);
6766 
6767         bfq_bic_update_cgroup(bic, bio);
6768 
6769         bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
6770                                          &new_queue);
6771 
6772         if (likely(!new_queue)) {
6773                 /* If the queue was seeky for too long, break it apart. */
6774                 if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq) &&
6775                         !bic->stably_merged) {
6776                         struct bfq_queue *old_bfqq = bfqq;
6777 
6778                         /* Update bic before losing reference to bfqq */
6779                         if (bfq_bfqq_in_large_burst(bfqq))
6780                                 bic->saved_in_large_burst = true;
6781 
6782                         bfqq = bfq_split_bfqq(bic, bfqq);
6783                         split = true;
6784 
6785                         if (!bfqq) {
6786                                 bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
6787                                                                  true, is_sync,
6788                                                                  NULL);
6789                                 bfqq->waker_bfqq = old_bfqq->waker_bfqq;
6790                                 bfqq->tentative_waker_bfqq = NULL;
6791 
6792                                 /*
6793                                  * If the waker queue disappears, then
6794                                  * new_bfqq->waker_bfqq must be
6795                                  * reset. So insert new_bfqq into the
6796                                  * woken_list of the waker. See
6797                                  * bfq_check_waker for details.
6798                                  */
6799                                 if (bfqq->waker_bfqq)
6800                                         hlist_add_head(&bfqq->woken_list_node,
6801                                                        &bfqq->waker_bfqq->woken_list);
6802                         } else
6803                                 bfqq_already_existing = true;
6804                 }
6805         }
6806 
6807         bfqq_request_allocated(bfqq);
6808         bfqq->ref++;
6809         bic->requests++;
6810         bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
6811                      rq, bfqq, bfqq->ref);
6812 
6813         rq->elv.priv[0] = bic;
6814         rq->elv.priv[1] = bfqq;
6815 
6816         /*
6817          * If a bfq_queue has only one process reference, it is owned
6818          * by only this bic: we can then set bfqq->bic = bic. in
6819          * addition, if the queue has also just been split, we have to
6820          * resume its state.
6821          */
6822         if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
6823                 bfqq->bic = bic;
6824                 if (split) {
6825                         /*
6826                          * The queue has just been split from a shared
6827                          * queue: restore the idle window and the
6828                          * possible weight raising period.
6829                          */
6830                         bfq_bfqq_resume_state(bfqq, bfqd, bic,
6831                                               bfqq_already_existing);
6832                 }
6833         }
6834 
6835         /*
6836          * Consider bfqq as possibly belonging to a burst of newly
6837          * created queues only if:
6838          * 1) A burst is actually happening (bfqd->burst_size > 0)
6839          * or
6840          * 2) There is no other active queue. In fact, if, in
6841          *    contrast, there are active queues not belonging to the
6842          *    possible burst bfqq may belong to, then there is no gain
6843          *    in considering bfqq as belonging to a burst, and
6844          *    therefore in not weight-raising bfqq. See comments on
6845          *    bfq_handle_burst().
6846          *
6847          * This filtering also helps eliminating false positives,
6848          * occurring when bfqq does not belong to an actual large
6849          * burst, but some background task (e.g., a service) happens
6850          * to trigger the creation of new queues very close to when
6851          * bfqq and its possible companion queues are created. See
6852          * comments on bfq_handle_burst() for further details also on
6853          * this issue.
6854          */
6855         if (unlikely(bfq_bfqq_just_created(bfqq) &&
6856                      (bfqd->burst_size > 0 ||
6857                       bfq_tot_busy_queues(bfqd) == 0)))
6858                 bfq_handle_burst(bfqd, bfqq);
6859 
6860         return bfqq;
6861 }
6862 
6863 static void
6864 bfq_idle_slice_timer_body(struct bfq_data *bfqd, struct bfq_queue *bfqq)
6865 {
6866         enum bfqq_expiration reason;
6867         unsigned long flags;
6868 
6869         spin_lock_irqsave(&bfqd->lock, flags);
6870 
6871         /*
6872          * Considering that bfqq may be in race, we should firstly check
6873          * whether bfqq is in service before doing something on it. If
6874          * the bfqq in race is not in service, it has already been expired
6875          * through __bfq_bfqq_expire func and its wait_request flags has
6876          * been cleared in __bfq_bfqd_reset_in_service func.
6877          */
6878         if (bfqq != bfqd->in_service_queue) {
6879                 spin_unlock_irqrestore(&bfqd->lock, flags);
6880                 return;
6881         }
6882 
6883         bfq_clear_bfqq_wait_request(bfqq);
6884 
6885         if (bfq_bfqq_budget_timeout(bfqq))
6886                 /*
6887                  * Also here the queue can be safely expired
6888                  * for budget timeout without wasting
6889                  * guarantees
6890                  */
6891                 reason = BFQQE_BUDGET_TIMEOUT;
6892         else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
6893                 /*
6894                  * The queue may not be empty upon timer expiration,
6895                  * because we may not disable the timer when the
6896                  * first request of the in-service queue arrives
6897                  * during disk idling.
6898                  */
6899                 reason = BFQQE_TOO_IDLE;
6900         else
6901                 goto schedule_dispatch;
6902 
6903         bfq_bfqq_expire(bfqd, bfqq, true, reason);
6904 
6905 schedule_dispatch:
6906         bfq_schedule_dispatch(bfqd);
6907         spin_unlock_irqrestore(&bfqd->lock, flags);
6908 }
6909 
6910 /*
6911  * Handler of the expiration of the timer running if the in-service queue
6912  * is idling inside its time slice.
6913  */
6914 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
6915 {
6916         struct bfq_data *bfqd = container_of(timer, struct bfq_data,
6917                                              idle_slice_timer);
6918         struct bfq_queue *bfqq = bfqd->in_service_queue;
6919 
6920         /*
6921          * Theoretical race here: the in-service queue can be NULL or
6922          * different from the queue that was idling if a new request
6923          * arrives for the current queue and there is a full dispatch
6924          * cycle that changes the in-service queue.  This can hardly
6925          * happen, but in the worst case we just expire a queue too
6926          * early.
6927          */
6928         if (bfqq)
6929                 bfq_idle_slice_timer_body(bfqd, bfqq);
6930 
6931         return HRTIMER_NORESTART;
6932 }
6933 
6934 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
6935                                  struct bfq_queue **bfqq_ptr)
6936 {
6937         struct bfq_queue *bfqq = *bfqq_ptr;
6938 
6939         bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
6940         if (bfqq) {
6941                 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
6942 
6943                 bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
6944                              bfqq, bfqq->ref);
6945                 bfq_put_queue(bfqq);
6946                 *bfqq_ptr = NULL;
6947         }
6948 }
6949 
6950 /*
6951  * Release all the bfqg references to its async queues.  If we are
6952  * deallocating the group these queues may still contain requests, so
6953  * we reparent them to the root cgroup (i.e., the only one that will
6954  * exist for sure until all the requests on a device are gone).
6955  */
6956 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
6957 {
6958         int i, j;
6959 
6960         for (i = 0; i < 2; i++)
6961                 for (j = 0; j < IOPRIO_NR_LEVELS; j++)
6962                         __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
6963 
6964         __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
6965 }
6966 
6967 /*
6968  * See the comments on bfq_limit_depth for the purpose of
6969  * the depths set in the function. Return minimum shallow depth we'll use.
6970  */
6971 static void bfq_update_depths(struct bfq_data *bfqd, struct sbitmap_queue *bt)
6972 {
6973         unsigned int depth = 1U << bt->sb.shift;
6974 
6975         bfqd->full_depth_shift = bt->sb.shift;
6976         /*
6977          * In-word depths if no bfq_queue is being weight-raised:
6978          * leaving 25% of tags only for sync reads.
6979          *
6980          * In next formulas, right-shift the value
6981          * (1U<<bt->sb.shift), instead of computing directly
6982          * (1U<<(bt->sb.shift - something)), to be robust against
6983          * any possible value of bt->sb.shift, without having to
6984          * limit 'something'.
6985          */
6986         /* no more than 50% of tags for async I/O */
6987         bfqd->word_depths[0][0] = max(depth >> 1, 1U);
6988         /*
6989          * no more than 75% of tags for sync writes (25% extra tags
6990          * w.r.t. async I/O, to prevent async I/O from starving sync
6991          * writes)
6992          */
6993         bfqd->word_depths[0][1] = max((depth * 3) >> 2, 1U);
6994 
6995         /*
6996          * In-word depths in case some bfq_queue is being weight-
6997          * raised: leaving ~63% of tags for sync reads. This is the
6998          * highest percentage for which, in our tests, application
6999          * start-up times didn't suffer from any regression due to tag
7000          * shortage.
7001          */
7002         /* no more than ~18% of tags for async I/O */
7003         bfqd->word_depths[1][0] = max((depth * 3) >> 4, 1U);
7004         /* no more than ~37% of tags for sync writes (~20% extra tags) */
7005         bfqd->word_depths[1][1] = max((depth * 6) >> 4, 1U);
7006 }
7007 
7008 static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
7009 {
7010         struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
7011         struct blk_mq_tags *tags = hctx->sched_tags;
7012 
7013         bfq_update_depths(bfqd, &tags->bitmap_tags);
7014         sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, 1);
7015 }
7016 
7017 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
7018 {
7019         bfq_depth_updated(hctx);
7020         return 0;
7021 }
7022 
7023 static void bfq_exit_queue(struct elevator_queue *e)
7024 {
7025         struct bfq_data *bfqd = e->elevator_data;
7026         struct bfq_queue *bfqq, *n;
7027 
7028         hrtimer_cancel(&bfqd->idle_slice_timer);
7029 
7030         spin_lock_irq(&bfqd->lock);
7031         list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
7032                 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
7033         spin_unlock_irq(&bfqd->lock);
7034 
7035         hrtimer_cancel(&bfqd->idle_slice_timer);
7036 
7037         /* release oom-queue reference to root group */
7038         bfqg_and_blkg_put(bfqd->root_group);
7039 
7040 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7041         blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
7042 #else
7043         spin_lock_irq(&bfqd->lock);
7044         bfq_put_async_queues(bfqd, bfqd->root_group);
7045         kfree(bfqd->root_group);
7046         spin_unlock_irq(&bfqd->lock);
7047 #endif
7048 
7049         blk_stat_disable_accounting(bfqd->queue);
7050         wbt_enable_default(bfqd->queue);
7051 
7052         kfree(bfqd);
7053 }
7054 
7055 static void bfq_init_root_group(struct bfq_group *root_group,
7056                                 struct bfq_data *bfqd)
7057 {
7058         int i;
7059 
7060 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7061         root_group->entity.parent = NULL;
7062         root_group->my_entity = NULL;
7063         root_group->bfqd = bfqd;
7064 #endif
7065         root_group->rq_pos_tree = RB_ROOT;
7066         for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
7067                 root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
7068         root_group->sched_data.bfq_class_idle_last_service = jiffies;
7069 }
7070 
7071 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
7072 {
7073         struct bfq_data *bfqd;
7074         struct elevator_queue *eq;
7075 
7076         eq = elevator_alloc(q, e);
7077         if (!eq)
7078                 return -ENOMEM;
7079 
7080         bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
7081         if (!bfqd) {
7082                 kobject_put(&eq->kobj);
7083                 return -ENOMEM;
7084         }
7085         eq->elevator_data = bfqd;
7086 
7087         spin_lock_irq(&q->queue_lock);
7088         q->elevator = eq;
7089         spin_unlock_irq(&q->queue_lock);
7090 
7091         /*
7092          * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
7093          * Grab a permanent reference to it, so that the normal code flow
7094          * will not attempt to free it.
7095          */
7096         bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
7097         bfqd->oom_bfqq.ref++;
7098         bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
7099         bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
7100         bfqd->oom_bfqq.entity.new_weight =
7101                 bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
7102 
7103         /* oom_bfqq does not participate to bursts */
7104         bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
7105 
7106         /*
7107          * Trigger weight initialization, according to ioprio, at the
7108          * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
7109          * class won't be changed any more.
7110          */
7111         bfqd->oom_bfqq.entity.prio_changed = 1;
7112 
7113         bfqd->queue = q;
7114 
7115         INIT_LIST_HEAD(&bfqd->dispatch);
7116 
7117         hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
7118                      HRTIMER_MODE_REL);
7119         bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
7120 
7121         bfqd->queue_weights_tree = RB_ROOT_CACHED;
7122         bfqd->num_groups_with_pending_reqs = 0;
7123 
7124         INIT_LIST_HEAD(&bfqd->active_list);
7125         INIT_LIST_HEAD(&bfqd->idle_list);
7126         INIT_HLIST_HEAD(&bfqd->burst_list);
7127 
7128         bfqd->hw_tag = -1;
7129         bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
7130 
7131         bfqd->bfq_max_budget = bfq_default_max_budget;
7132 
7133         bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
7134         bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
7135         bfqd->bfq_back_max = bfq_back_max;
7136         bfqd->bfq_back_penalty = bfq_back_penalty;
7137         bfqd->bfq_slice_idle = bfq_slice_idle;
7138         bfqd->bfq_timeout = bfq_timeout;
7139 
7140         bfqd->bfq_large_burst_thresh = 8;
7141         bfqd->bfq_burst_interval = msecs_to_jiffies(180);
7142 
7143         bfqd->low_latency = true;
7144 
7145         /*
7146          * Trade-off between responsiveness and fairness.
7147          */
7148         bfqd->bfq_wr_coeff = 30;
7149         bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
7150         bfqd->bfq_wr_max_time = 0;
7151         bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
7152         bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
7153         bfqd->bfq_wr_max_softrt_rate = 7000; /*
7154                                               * Approximate rate required
7155                                               * to playback or record a
7156                                               * high-definition compressed
7157                                               * video.
7158                                               */
7159         bfqd->wr_busy_queues = 0;
7160 
7161         /*
7162          * Begin by assuming, optimistically, that the device peak
7163          * rate is equal to 2/3 of the highest reference rate.
7164          */
7165         bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
7166                 ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
7167         bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
7168 
7169         spin_lock_init(&bfqd->lock);
7170 
7171         /*
7172          * The invocation of the next bfq_create_group_hierarchy
7173          * function is the head of a chain of function calls
7174          * (bfq_create_group_hierarchy->blkcg_activate_policy->
7175          * blk_mq_freeze_queue) that may lead to the invocation of the
7176          * has_work hook function. For this reason,
7177          * bfq_create_group_hierarchy is invoked only after all
7178          * scheduler data has been initialized, apart from the fields
7179          * that can be initialized only after invoking
7180          * bfq_create_group_hierarchy. This, in particular, enables
7181          * has_work to correctly return false. Of course, to avoid
7182          * other inconsistencies, the blk-mq stack must then refrain
7183          * from invoking further scheduler hooks before this init
7184          * function is finished.
7185          */
7186         bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
7187         if (!bfqd->root_group)
7188                 goto out_free;
7189         bfq_init_root_group(bfqd->root_group, bfqd);
7190         bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
7191 
7192         /* We dispatch from request queue wide instead of hw queue */
7193         blk_queue_flag_set(QUEUE_FLAG_SQ_SCHED, q);
7194 
7195         wbt_disable_default(q);
7196         blk_stat_enable_accounting(q);
7197 
7198         return 0;
7199 
7200 out_free:
7201         kfree(bfqd);
7202         kobject_put(&eq->kobj);
7203         return -ENOMEM;
7204 }
7205 
7206 static void bfq_slab_kill(void)
7207 {
7208         kmem_cache_destroy(bfq_pool);
7209 }
7210 
7211 static int __init bfq_slab_setup(void)
7212 {
7213         bfq_pool = KMEM_CACHE(bfq_queue, 0);
7214         if (!bfq_pool)
7215                 return -ENOMEM;
7216         return 0;
7217 }
7218 
7219 static ssize_t bfq_var_show(unsigned int var, char *page)
7220 {
7221         return sprintf(page, "%u\n", var);
7222 }
7223 
7224 static int bfq_var_store(unsigned long *var, const char *page)
7225 {
7226         unsigned long new_val;
7227         int ret = kstrtoul(page, 10, &new_val);
7228 
7229         if (ret)
7230                 return ret;
7231         *var = new_val;
7232         return 0;
7233 }
7234 
7235 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
7236 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
7237 {                                                                       \
7238         struct bfq_data *bfqd = e->elevator_data;                       \
7239         u64 __data = __VAR;                                             \
7240         if (__CONV == 1)                                                \
7241                 __data = jiffies_to_msecs(__data);                      \
7242         else if (__CONV == 2)                                           \
7243                 __data = div_u64(__data, NSEC_PER_MSEC);                \
7244         return bfq_var_show(__data, (page));                            \
7245 }
7246 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
7247 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
7248 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
7249 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
7250 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
7251 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
7252 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
7253 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
7254 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
7255 #undef SHOW_FUNCTION
7256 
7257 #define USEC_SHOW_FUNCTION(__FUNC, __VAR)                               \
7258 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
7259 {                                                                       \
7260         struct bfq_data *bfqd = e->elevator_data;                       \
7261         u64 __data = __VAR;                                             \
7262         __data = div_u64(__data, NSEC_PER_USEC);                        \
7263         return bfq_var_show(__data, (page));                            \
7264 }
7265 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
7266 #undef USEC_SHOW_FUNCTION
7267 
7268 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
7269 static ssize_t                                                          \
7270 __FUNC(struct elevator_queue *e, const char *page, size_t count)        \
7271 {                                                                       \
7272         struct bfq_data *bfqd = e->elevator_data;                       \
7273         unsigned long __data, __min = (MIN), __max = (MAX);             \
7274         int ret;                                                        \
7275                                                                         \
7276         ret = bfq_var_store(&__data, (page));                           \
7277         if (ret)                                                        \
7278                 return ret;                                             \
7279         if (__data < __min)                                             \
7280                 __data = __min;                                         \
7281         else if (__data > __max)                                        \
7282                 __data = __max;                                         \
7283         if (__CONV == 1)                                                \
7284                 *(__PTR) = msecs_to_jiffies(__data);                    \
7285         else if (__CONV == 2)                                           \
7286                 *(__PTR) = (u64)__data * NSEC_PER_MSEC;                 \
7287         else                                                            \
7288                 *(__PTR) = __data;                                      \
7289         return count;                                                   \
7290 }
7291 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
7292                 INT_MAX, 2);
7293 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
7294                 INT_MAX, 2);
7295 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
7296 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
7297                 INT_MAX, 0);
7298 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
7299 #undef STORE_FUNCTION
7300 
7301 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX)                    \
7302 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
7303 {                                                                       \
7304         struct bfq_data *bfqd = e->elevator_data;                       \
7305         unsigned long __data, __min = (MIN), __max = (MAX);             \
7306         int ret;                                                        \
7307                                                                         \
7308         ret = bfq_var_store(&__data, (page));                           \
7309         if (ret)                                                        \
7310                 return ret;                                             \
7311         if (__data < __min)                                             \
7312                 __data = __min;                                         \
7313         else if (__data > __max)                                        \
7314                 __data = __max;                                         \
7315         *(__PTR) = (u64)__data * NSEC_PER_USEC;                         \
7316         return count;                                                   \
7317 }
7318 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
7319                     UINT_MAX);
7320 #undef USEC_STORE_FUNCTION
7321 
7322 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
7323                                     const char *page, size_t count)
7324 {
7325         struct bfq_data *bfqd = e->elevator_data;
7326         unsigned long __data;
7327         int ret;
7328 
7329         ret = bfq_var_store(&__data, (page));
7330         if (ret)
7331                 return ret;
7332 
7333         if (__data == 0)
7334                 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7335         else {
7336                 if (__data > INT_MAX)
7337                         __data = INT_MAX;
7338                 bfqd->bfq_max_budget = __data;
7339         }
7340 
7341         bfqd->bfq_user_max_budget = __data;
7342 
7343         return count;
7344 }
7345 
7346 /*
7347  * Leaving this name to preserve name compatibility with cfq
7348  * parameters, but this timeout is used for both sync and async.
7349  */
7350 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
7351                                       const char *page, size_t count)
7352 {
7353         struct bfq_data *bfqd = e->elevator_data;
7354         unsigned long __data;
7355         int ret;
7356 
7357         ret = bfq_var_store(&__data, (page));
7358         if (ret)
7359                 return ret;
7360 
7361         if (__data < 1)
7362                 __data = 1;
7363         else if (__data > INT_MAX)
7364                 __data = INT_MAX;
7365 
7366         bfqd->bfq_timeout = msecs_to_jiffies(__data);
7367         if (bfqd->bfq_user_max_budget == 0)
7368                 bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7369 
7370         return count;
7371 }
7372 
7373 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
7374                                      const char *page, size_t count)
7375 {
7376         struct bfq_data *bfqd = e->elevator_data;
7377         unsigned long __data;
7378         int ret;
7379 
7380         ret = bfq_var_store(&__data, (page));
7381         if (ret)
7382                 return ret;
7383 
7384         if (__data > 1)
7385                 __data = 1;
7386         if (!bfqd->strict_guarantees && __data == 1
7387             && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
7388                 bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
7389 
7390         bfqd->strict_guarantees = __data;
7391 
7392         return count;
7393 }
7394 
7395 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
7396                                      const char *page, size_t count)
7397 {
7398         struct bfq_data *bfqd = e->elevator_data;
7399         unsigned long __data;
7400         int ret;
7401 
7402         ret = bfq_var_store(&__data, (page));
7403         if (ret)
7404                 return ret;
7405 
7406         if (__data > 1)
7407                 __data = 1;
7408         if (__data == 0 && bfqd->low_latency != 0)
7409                 bfq_end_wr(bfqd);
7410         bfqd->low_latency = __data;
7411 
7412         return count;
7413 }
7414 
7415 #define BFQ_ATTR(name) \
7416         __ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
7417 
7418 static struct elv_fs_entry bfq_attrs[] = {
7419         BFQ_ATTR(fifo_expire_sync),
7420         BFQ_ATTR(fifo_expire_async),
7421         BFQ_ATTR(back_seek_max),
7422         BFQ_ATTR(back_seek_penalty),
7423         BFQ_ATTR(slice_idle),
7424         BFQ_ATTR(slice_idle_us),
7425         BFQ_ATTR(max_budget),
7426         BFQ_ATTR(timeout_sync),
7427         BFQ_ATTR(strict_guarantees),
7428         BFQ_ATTR(low_latency),
7429         __ATTR_NULL
7430 };
7431 
7432 static struct elevator_type iosched_bfq_mq = {
7433         .ops = {
7434                 .limit_depth            = bfq_limit_depth,
7435                 .prepare_request        = bfq_prepare_request,
7436                 .requeue_request        = bfq_finish_requeue_request,
7437                 .finish_request         = bfq_finish_request,
7438                 .exit_icq               = bfq_exit_icq,
7439                 .insert_requests        = bfq_insert_requests,
7440                 .dispatch_request       = bfq_dispatch_request,
7441                 .next_request           = elv_rb_latter_request,
7442                 .former_request         = elv_rb_former_request,
7443                 .allow_merge            = bfq_allow_bio_merge,
7444                 .bio_merge              = bfq_bio_merge,
7445                 .request_merge          = bfq_request_merge,
7446                 .requests_merged        = bfq_requests_merged,
7447                 .request_merged         = bfq_request_merged,
7448                 .has_work               = bfq_has_work,
7449                 .depth_updated          = bfq_depth_updated,
7450                 .init_hctx              = bfq_init_hctx,
7451                 .init_sched             = bfq_init_queue,
7452                 .exit_sched             = bfq_exit_queue,
7453         },
7454 
7455         .icq_size =             sizeof(struct bfq_io_cq),
7456         .icq_align =            __alignof__(struct bfq_io_cq),
7457         .elevator_attrs =       bfq_attrs,
7458         .elevator_name =        "bfq",
7459         .elevator_owner =       THIS_MODULE,
7460 };
7461 MODULE_ALIAS("bfq-iosched");
7462 
7463 static int __init bfq_init(void)
7464 {
7465         int ret;
7466 
7467 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7468         ret = blkcg_policy_register(&blkcg_policy_bfq);
7469         if (ret)
7470                 return ret;
7471 #endif
7472 
7473         ret = -ENOMEM;
7474         if (bfq_slab_setup())
7475                 goto err_pol_unreg;
7476 
7477         /*
7478          * Times to load large popular applications for the typical
7479          * systems installed on the reference devices (see the
7480          * comments before the definition of the next
7481          * array). Actually, we use slightly lower values, as the
7482          * estimated peak rate tends to be smaller than the actual
7483          * peak rate.  The reason for this last fact is that estimates
7484          * are computed over much shorter time intervals than the long
7485          * intervals typically used for benchmarking. Why? First, to
7486          * adapt more quickly to variations. Second, because an I/O
7487          * scheduler cannot rely on a peak-rate-evaluation workload to
7488          * be run for a long time.
7489          */
7490         ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
7491         ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
7492 
7493         ret = elv_register(&iosched_bfq_mq);
7494         if (ret)
7495                 goto slab_kill;
7496 
7497         return 0;
7498 
7499 slab_kill:
7500         bfq_slab_kill();
7501 err_pol_unreg:
7502 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7503         blkcg_policy_unregister(&blkcg_policy_bfq);
7504 #endif
7505         return ret;
7506 }
7507 
7508 static void __exit bfq_exit(void)
7509 {
7510         elv_unregister(&iosched_bfq_mq);
7511 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7512         blkcg_policy_unregister(&blkcg_policy_bfq);
7513 #endif
7514         bfq_slab_kill();
7515 }
7516 
7517 module_init(bfq_init);
7518 module_exit(bfq_exit);
7519 
7520 MODULE_AUTHOR("Paolo Valente");
7521 MODULE_LICENSE("GPL");
7522 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");
7523 

~ [ 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