Linux Audio

Check our new training course

Loading...
v6.2
   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)						\
 138void bfq_mark_bfqq_##name(struct bfq_queue *bfqq)			\
 139{									\
 140	__set_bit(BFQQF_##name, &(bfqq)->flags);			\
 141}									\
 142void bfq_clear_bfqq_##name(struct bfq_queue *bfqq)			\
 143{									\
 144	__clear_bit(BFQQF_##name, &(bfqq)->flags);		\
 145}									\
 146int bfq_bfqq_##name(const struct bfq_queue *bfqq)			\
 147{									\
 148	return test_bit(BFQQF_##name, &(bfqq)->flags);		\
 149}
 150
 151BFQ_BFQQ_FNS(just_created);
 152BFQ_BFQQ_FNS(busy);
 153BFQ_BFQQ_FNS(wait_request);
 154BFQ_BFQQ_FNS(non_blocking_wait_rq);
 155BFQ_BFQQ_FNS(fifo_expire);
 156BFQ_BFQQ_FNS(has_short_ttime);
 157BFQ_BFQQ_FNS(sync);
 158BFQ_BFQQ_FNS(IO_bound);
 159BFQ_BFQQ_FNS(in_large_burst);
 160BFQ_BFQQ_FNS(coop);
 161BFQ_BFQQ_FNS(split_coop);
 162BFQ_BFQQ_FNS(softrt_update);
 163#undef BFQ_BFQQ_FNS						\
 164
 165/* Expiration time of async (0) and sync (1) requests, in ns. */
 166static 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. */
 169static const int bfq_back_max = 16 * 1024;
 170
 171/* Penalty of a backwards seek, in number of sectors. */
 172static const int bfq_back_penalty = 2;
 173
 174/* Idling period duration, in ns. */
 175static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
 176
 177/* Minimum number of assigned budgets for which stats are safe to compute. */
 178static const int bfq_stats_min_budgets = 194;
 179
 180/* Default maximum budget values, in sectors and number of requests. */
 181static 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 */
 202static const int bfq_async_charge_factor = 3;
 203
 204/* Default timeout values, in jiffies, approximating CFQ defaults. */
 205const 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 */
 219static const unsigned long bfq_merge_time_limit = HZ/10;
 220
 221static 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 */
 305static 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 */
 311static 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 */
 365static 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 */
 371static const unsigned long bfq_activation_stable_merging = 600;
 372/*
 373 * Minimum time to be waited before evaluating delayed stable merge (in ms)
 374 */
 375static 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
 380struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
 381{
 382	return bic->bfqq[is_sync];
 383}
 384
 385static void bfq_put_stable_ref(struct bfq_queue *bfqq);
 386
 387void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
 388{
 389	struct bfq_queue *old_bfqq = bic->bfqq[is_sync];
 390
 391	/* Clear bic pointer if bfqq is detached from this bic */
 392	if (old_bfqq && old_bfqq->bic == bic)
 393		old_bfqq->bic = NULL;
 394
 395	/*
 396	 * If bfqq != NULL, then a non-stable queue merge between
 397	 * bic->bfqq and bfqq is happening here. This causes troubles
 398	 * in the following case: bic->bfqq has also been scheduled
 399	 * for a possible stable merge with bic->stable_merge_bfqq,
 400	 * and bic->stable_merge_bfqq == bfqq happens to
 401	 * hold. Troubles occur because bfqq may then undergo a split,
 402	 * thereby becoming eligible for a stable merge. Yet, if
 403	 * bic->stable_merge_bfqq points exactly to bfqq, then bfqq
 404	 * would be stably merged with itself. To avoid this anomaly,
 405	 * we cancel the stable merge if
 406	 * bic->stable_merge_bfqq == bfqq.
 407	 */
 408	bic->bfqq[is_sync] = bfqq;
 409
 410	if (bfqq && bic->stable_merge_bfqq == bfqq) {
 411		/*
 412		 * Actually, these same instructions are executed also
 413		 * in bfq_setup_cooperator, in case of abort or actual
 414		 * execution of a stable merge. We could avoid
 415		 * repeating these instructions there too, but if we
 416		 * did so, we would nest even more complexity in this
 417		 * function.
 418		 */
 419		bfq_put_stable_ref(bic->stable_merge_bfqq);
 420
 421		bic->stable_merge_bfqq = NULL;
 422	}
 423}
 424
 425struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
 426{
 427	return bic->icq.q->elevator->elevator_data;
 428}
 429
 430/**
 431 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
 432 * @icq: the iocontext queue.
 433 */
 434static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
 435{
 436	/* bic->icq is the first member, %NULL will convert to %NULL */
 437	return container_of(icq, struct bfq_io_cq, icq);
 438}
 439
 440/**
 441 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
 
 
 442 * @q: the request queue.
 443 */
 444static struct bfq_io_cq *bfq_bic_lookup(struct request_queue *q)
 
 
 445{
 446	struct bfq_io_cq *icq;
 447	unsigned long flags;
 
 448
 449	if (!current->io_context)
 450		return NULL;
 
 451
 452	spin_lock_irqsave(&q->queue_lock, flags);
 453	icq = icq_to_bic(ioc_lookup_icq(q));
 454	spin_unlock_irqrestore(&q->queue_lock, flags);
 455
 456	return icq;
 457}
 458
 459/*
 460 * Scheduler run of queue, if there are requests pending and no one in the
 461 * driver that will restart queueing.
 462 */
 463void bfq_schedule_dispatch(struct bfq_data *bfqd)
 464{
 465	lockdep_assert_held(&bfqd->lock);
 466
 467	if (bfqd->queued != 0) {
 468		bfq_log(bfqd, "schedule dispatch");
 469		blk_mq_run_hw_queues(bfqd->queue, true);
 470	}
 471}
 472
 473#define bfq_class_idle(bfqq)	((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
 
 474
 475#define bfq_sample_valid(samples)	((samples) > 80)
 476
 477/*
 478 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 479 * We choose the request that is closer to the head right now.  Distance
 480 * behind the head is penalized and only allowed to a certain extent.
 481 */
 482static struct request *bfq_choose_req(struct bfq_data *bfqd,
 483				      struct request *rq1,
 484				      struct request *rq2,
 485				      sector_t last)
 486{
 487	sector_t s1, s2, d1 = 0, d2 = 0;
 488	unsigned long back_max;
 489#define BFQ_RQ1_WRAP	0x01 /* request 1 wraps */
 490#define BFQ_RQ2_WRAP	0x02 /* request 2 wraps */
 491	unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
 492
 493	if (!rq1 || rq1 == rq2)
 494		return rq2;
 495	if (!rq2)
 496		return rq1;
 497
 498	if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 499		return rq1;
 500	else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
 501		return rq2;
 502	if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
 503		return rq1;
 504	else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
 505		return rq2;
 506
 507	s1 = blk_rq_pos(rq1);
 508	s2 = blk_rq_pos(rq2);
 509
 510	/*
 511	 * By definition, 1KiB is 2 sectors.
 512	 */
 513	back_max = bfqd->bfq_back_max * 2;
 514
 515	/*
 516	 * Strict one way elevator _except_ in the case where we allow
 517	 * short backward seeks which are biased as twice the cost of a
 518	 * similar forward seek.
 519	 */
 520	if (s1 >= last)
 521		d1 = s1 - last;
 522	else if (s1 + back_max >= last)
 523		d1 = (last - s1) * bfqd->bfq_back_penalty;
 524	else
 525		wrap |= BFQ_RQ1_WRAP;
 526
 527	if (s2 >= last)
 528		d2 = s2 - last;
 529	else if (s2 + back_max >= last)
 530		d2 = (last - s2) * bfqd->bfq_back_penalty;
 531	else
 532		wrap |= BFQ_RQ2_WRAP;
 533
 534	/* Found required data */
 535
 536	/*
 537	 * By doing switch() on the bit mask "wrap" we avoid having to
 538	 * check two variables for all permutations: --> faster!
 539	 */
 540	switch (wrap) {
 541	case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
 542		if (d1 < d2)
 543			return rq1;
 544		else if (d2 < d1)
 545			return rq2;
 546
 547		if (s1 >= s2)
 548			return rq1;
 549		else
 550			return rq2;
 551
 552	case BFQ_RQ2_WRAP:
 553		return rq1;
 554	case BFQ_RQ1_WRAP:
 555		return rq2;
 556	case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
 557	default:
 558		/*
 559		 * Since both rqs are wrapped,
 560		 * start with the one that's further behind head
 561		 * (--> only *one* back seek required),
 562		 * since back seek takes more time than forward.
 563		 */
 564		if (s1 <= s2)
 565			return rq1;
 566		else
 567			return rq2;
 568	}
 569}
 570
 571#define BFQ_LIMIT_INLINE_DEPTH 16
 572
 573#ifdef CONFIG_BFQ_GROUP_IOSCHED
 574static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
 
 575{
 576	struct bfq_data *bfqd = bfqq->bfqd;
 577	struct bfq_entity *entity = &bfqq->entity;
 578	struct bfq_entity *inline_entities[BFQ_LIMIT_INLINE_DEPTH];
 579	struct bfq_entity **entities = inline_entities;
 580	int depth, level, alloc_depth = BFQ_LIMIT_INLINE_DEPTH;
 581	int class_idx = bfqq->ioprio_class - 1;
 582	struct bfq_sched_data *sched_data;
 583	unsigned long wsum;
 584	bool ret = false;
 585
 586	if (!entity->on_st_or_in_serv)
 587		return false;
 588
 589retry:
 590	spin_lock_irq(&bfqd->lock);
 591	/* +1 for bfqq entity, root cgroup not included */
 592	depth = bfqg_to_blkg(bfqq_group(bfqq))->blkcg->css.cgroup->level + 1;
 593	if (depth > alloc_depth) {
 594		spin_unlock_irq(&bfqd->lock);
 595		if (entities != inline_entities)
 596			kfree(entities);
 597		entities = kmalloc_array(depth, sizeof(*entities), GFP_NOIO);
 598		if (!entities)
 599			return false;
 600		alloc_depth = depth;
 601		goto retry;
 602	}
 
 
 
 
 603
 604	sched_data = entity->sched_data;
 605	/* Gather our ancestors as we need to traverse them in reverse order */
 606	level = 0;
 607	for_each_entity(entity) {
 608		/*
 609		 * If at some level entity is not even active, allow request
 610		 * queueing so that BFQ knows there's work to do and activate
 611		 * entities.
 612		 */
 613		if (!entity->on_st_or_in_serv)
 614			goto out;
 615		/* Uh, more parents than cgroup subsystem thinks? */
 616		if (WARN_ON_ONCE(level >= depth))
 617			break;
 618		entities[level++] = entity;
 619	}
 620	WARN_ON_ONCE(level != depth);
 621	for (level--; level >= 0; level--) {
 622		entity = entities[level];
 623		if (level > 0) {
 624			wsum = bfq_entity_service_tree(entity)->wsum;
 625		} else {
 626			int i;
 627			/*
 628			 * For bfqq itself we take into account service trees
 629			 * of all higher priority classes and multiply their
 630			 * weights so that low prio queue from higher class
 631			 * gets more requests than high prio queue from lower
 632			 * class.
 633			 */
 634			wsum = 0;
 635			for (i = 0; i <= class_idx; i++) {
 636				wsum = wsum * IOPRIO_BE_NR +
 637					sched_data->service_tree[i].wsum;
 638			}
 639		}
 640		limit = DIV_ROUND_CLOSEST(limit * entity->weight, wsum);
 641		if (entity->allocated >= limit) {
 642			bfq_log_bfqq(bfqq->bfqd, bfqq,
 643				"too many requests: allocated %d limit %d level %d",
 644				entity->allocated, limit, level);
 645			ret = true;
 646			break;
 647		}
 648	}
 649out:
 650	spin_unlock_irq(&bfqd->lock);
 651	if (entities != inline_entities)
 652		kfree(entities);
 653	return ret;
 654}
 655#else
 656static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
 657{
 658	return false;
 659}
 660#endif
 661
 662/*
 663 * Async I/O can easily starve sync I/O (both sync reads and sync
 664 * writes), by consuming all tags. Similarly, storms of sync writes,
 665 * such as those that sync(2) may trigger, can starve sync reads.
 666 * Limit depths of async I/O and sync writes so as to counter both
 667 * problems.
 668 *
 669 * Also if a bfq queue or its parent cgroup consume more tags than would be
 670 * appropriate for their weight, we trim the available tag depth to 1. This
 671 * avoids a situation where one cgroup can starve another cgroup from tags and
 672 * thus block service differentiation among cgroups. Note that because the
 673 * queue / cgroup already has many requests allocated and queued, this does not
 674 * significantly affect service guarantees coming from the BFQ scheduling
 675 * algorithm.
 676 */
 677static void bfq_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
 678{
 
 679	struct bfq_data *bfqd = data->q->elevator->elevator_data;
 680	struct bfq_io_cq *bic = bfq_bic_lookup(data->q);
 681	struct bfq_queue *bfqq = bic ? bic_to_bfqq(bic, op_is_sync(opf)) : NULL;
 682	int depth;
 683	unsigned limit = data->q->nr_requests;
 684
 685	/* Sync reads have full depth available */
 686	if (op_is_sync(opf) && !op_is_write(opf)) {
 687		depth = 0;
 688	} else {
 689		depth = bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(opf)];
 690		limit = (limit * depth) >> bfqd->full_depth_shift;
 691	}
 692
 693	/*
 694	 * Does queue (or any parent entity) exceed number of requests that
 695	 * should be available to it? Heavily limit depth so that it cannot
 696	 * consume more available requests and thus starve other entities.
 697	 */
 698	if (bfqq && bfqq_request_over_limit(bfqq, limit))
 699		depth = 1;
 
 
 
 
 
 
 
 
 
 
 700
 701	bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
 702		__func__, bfqd->wr_busy_queues, op_is_sync(opf), depth);
 703	if (depth)
 704		data->shallow_depth = depth;
 705}
 706
 707static struct bfq_queue *
 708bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
 709		     sector_t sector, struct rb_node **ret_parent,
 710		     struct rb_node ***rb_link)
 711{
 712	struct rb_node **p, *parent;
 713	struct bfq_queue *bfqq = NULL;
 714
 715	parent = NULL;
 716	p = &root->rb_node;
 717	while (*p) {
 718		struct rb_node **n;
 719
 720		parent = *p;
 721		bfqq = rb_entry(parent, struct bfq_queue, pos_node);
 722
 723		/*
 724		 * Sort strictly based on sector. Smallest to the left,
 725		 * largest to the right.
 726		 */
 727		if (sector > blk_rq_pos(bfqq->next_rq))
 728			n = &(*p)->rb_right;
 729		else if (sector < blk_rq_pos(bfqq->next_rq))
 730			n = &(*p)->rb_left;
 731		else
 732			break;
 733		p = n;
 734		bfqq = NULL;
 735	}
 736
 737	*ret_parent = parent;
 738	if (rb_link)
 739		*rb_link = p;
 740
 741	bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
 742		(unsigned long long)sector,
 743		bfqq ? bfqq->pid : 0);
 744
 745	return bfqq;
 746}
 747
 748static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
 749{
 750	return bfqq->service_from_backlogged > 0 &&
 751		time_is_before_jiffies(bfqq->first_IO_time +
 752				       bfq_merge_time_limit);
 753}
 754
 755/*
 756 * The following function is not marked as __cold because it is
 757 * actually cold, but for the same performance goal described in the
 758 * comments on the likely() at the beginning of
 759 * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
 760 * execution time for the case where this function is not invoked, we
 761 * had to add an unlikely() in each involved if().
 762 */
 763void __cold
 764bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
 765{
 766	struct rb_node **p, *parent;
 767	struct bfq_queue *__bfqq;
 768
 769	if (bfqq->pos_root) {
 770		rb_erase(&bfqq->pos_node, bfqq->pos_root);
 771		bfqq->pos_root = NULL;
 772	}
 773
 774	/* oom_bfqq does not participate in queue merging */
 775	if (bfqq == &bfqd->oom_bfqq)
 776		return;
 777
 778	/*
 779	 * bfqq cannot be merged any longer (see comments in
 780	 * bfq_setup_cooperator): no point in adding bfqq into the
 781	 * position tree.
 782	 */
 783	if (bfq_too_late_for_merging(bfqq))
 784		return;
 785
 786	if (bfq_class_idle(bfqq))
 787		return;
 788	if (!bfqq->next_rq)
 789		return;
 790
 791	bfqq->pos_root = &bfqq_group(bfqq)->rq_pos_tree;
 792	__bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
 793			blk_rq_pos(bfqq->next_rq), &parent, &p);
 794	if (!__bfqq) {
 795		rb_link_node(&bfqq->pos_node, parent, p);
 796		rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
 797	} else
 798		bfqq->pos_root = NULL;
 799}
 800
 801/*
 802 * The following function returns false either if every active queue
 803 * must receive the same share of the throughput (symmetric scenario),
 804 * or, as a special case, if bfqq must receive a share of the
 805 * throughput lower than or equal to the share that every other active
 806 * queue must receive.  If bfqq does sync I/O, then these are the only
 807 * two cases where bfqq happens to be guaranteed its share of the
 808 * throughput even if I/O dispatching is not plugged when bfqq remains
 809 * temporarily empty (for more details, see the comments in the
 810 * function bfq_better_to_idle()). For this reason, the return value
 811 * of this function is used to check whether I/O-dispatch plugging can
 812 * be avoided.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 813 *
 814 * The above first case (symmetric scenario) occurs when:
 815 * 1) all active queues have the same weight,
 816 * 2) all active queues belong to the same I/O-priority class,
 817 * 3) all active groups at the same level in the groups tree have the same
 818 *    weight,
 819 * 4) all active groups at the same level in the groups tree have the same
 820 *    number of children.
 821 *
 822 * Unfortunately, keeping the necessary state for evaluating exactly
 823 * the last two symmetry sub-conditions above would be quite complex
 824 * and time consuming. Therefore this function evaluates, instead,
 825 * only the following stronger three sub-conditions, for which it is
 826 * much easier to maintain the needed state:
 827 * 1) all active queues have the same weight,
 828 * 2) all active queues belong to the same I/O-priority class,
 829 * 3) there is at most one active group.
 830 * In particular, the last condition is always true if hierarchical
 831 * support or the cgroups interface are not enabled, thus no state
 832 * needs to be maintained in this case.
 833 */
 834static bool bfq_asymmetric_scenario(struct bfq_data *bfqd,
 835				   struct bfq_queue *bfqq)
 836{
 837	bool smallest_weight = bfqq &&
 838		bfqq->weight_counter &&
 839		bfqq->weight_counter ==
 840		container_of(
 841			rb_first_cached(&bfqd->queue_weights_tree),
 842			struct bfq_weight_counter,
 843			weights_node);
 844
 845	/*
 846	 * For queue weights to differ, queue_weights_tree must contain
 847	 * at least two nodes.
 848	 */
 849	bool varied_queue_weights = !smallest_weight &&
 850		!RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) &&
 851		(bfqd->queue_weights_tree.rb_root.rb_node->rb_left ||
 852		 bfqd->queue_weights_tree.rb_root.rb_node->rb_right);
 853
 854	bool multiple_classes_busy =
 855		(bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
 856		(bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
 857		(bfqd->busy_queues[1] && bfqd->busy_queues[2]);
 858
 859	return varied_queue_weights || multiple_classes_busy
 860#ifdef CONFIG_BFQ_GROUP_IOSCHED
 861	       || bfqd->num_groups_with_pending_reqs > 1
 862#endif
 863		;
 864}
 865
 866/*
 867 * If the weight-counter tree passed as input contains no counter for
 868 * the weight of the input queue, then add that counter; otherwise just
 869 * increment the existing counter.
 870 *
 871 * Note that weight-counter trees contain few nodes in mostly symmetric
 872 * scenarios. For example, if all queues have the same weight, then the
 873 * weight-counter tree for the queues may contain at most one node.
 874 * This holds even if low_latency is on, because weight-raised queues
 875 * are not inserted in the tree.
 876 * In most scenarios, the rate at which nodes are created/destroyed
 877 * should be low too.
 878 */
 879void bfq_weights_tree_add(struct bfq_queue *bfqq)
 
 880{
 881	struct rb_root_cached *root = &bfqq->bfqd->queue_weights_tree;
 882	struct bfq_entity *entity = &bfqq->entity;
 883	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
 884	bool leftmost = true;
 885
 886	/*
 887	 * Do not insert if the queue is already associated with a
 888	 * counter, which happens if:
 889	 *   1) a request arrival has caused the queue to become both
 
 890	 *      non-weight-raised, and hence change its weight, and
 891	 *      backlogged; in this respect, each of the two events
 892	 *      causes an invocation of this function,
 893	 *   2) this is the invocation of this function caused by the
 894	 *      second event. This second invocation is actually useless,
 895	 *      and we handle this fact by exiting immediately. More
 896	 *      efficient or clearer solutions might possibly be adopted.
 897	 */
 898	if (bfqq->weight_counter)
 899		return;
 900
 901	while (*new) {
 902		struct bfq_weight_counter *__counter = container_of(*new,
 903						struct bfq_weight_counter,
 904						weights_node);
 905		parent = *new;
 906
 907		if (entity->weight == __counter->weight) {
 908			bfqq->weight_counter = __counter;
 909			goto inc_counter;
 910		}
 911		if (entity->weight < __counter->weight)
 912			new = &((*new)->rb_left);
 913		else {
 914			new = &((*new)->rb_right);
 915			leftmost = false;
 916		}
 917	}
 918
 919	bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
 920				       GFP_ATOMIC);
 921
 922	/*
 923	 * In the unlucky event of an allocation failure, we just
 924	 * exit. This will cause the weight of queue to not be
 925	 * considered in bfq_asymmetric_scenario, which, in its turn,
 926	 * causes the scenario to be deemed wrongly symmetric in case
 927	 * bfqq's weight would have been the only weight making the
 928	 * scenario asymmetric.  On the bright side, no unbalance will
 929	 * however occur when bfqq becomes inactive again (the
 930	 * invocation of this function is triggered by an activation
 931	 * of queue).  In fact, bfq_weights_tree_remove does nothing
 932	 * if !bfqq->weight_counter.
 933	 */
 934	if (unlikely(!bfqq->weight_counter))
 935		return;
 936
 937	bfqq->weight_counter->weight = entity->weight;
 938	rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
 939	rb_insert_color_cached(&bfqq->weight_counter->weights_node, root,
 940				leftmost);
 941
 942inc_counter:
 943	bfqq->weight_counter->num_active++;
 944	bfqq->ref++;
 945}
 946
 947/*
 948 * Decrement the weight counter associated with the queue, and, if the
 949 * counter reaches 0, remove the counter from the tree.
 950 * See the comments to the function bfq_weights_tree_add() for considerations
 951 * about overhead.
 952 */
 953void bfq_weights_tree_remove(struct bfq_queue *bfqq)
 
 954{
 955	struct rb_root_cached *root;
 956
 957	if (!bfqq->weight_counter)
 958		return;
 959
 960	root = &bfqq->bfqd->queue_weights_tree;
 961	bfqq->weight_counter->num_active--;
 962	if (bfqq->weight_counter->num_active > 0)
 963		goto reset_entity_pointer;
 964
 965	rb_erase_cached(&bfqq->weight_counter->weights_node, root);
 966	kfree(bfqq->weight_counter);
 967
 968reset_entity_pointer:
 969	bfqq->weight_counter = NULL;
 970	bfq_put_queue(bfqq);
 971}
 972
 973/*
 974 * Return expired entry, or NULL to just start from scratch in rbtree.
 975 */
 976static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
 977				      struct request *last)
 978{
 979	struct request *rq;
 980
 981	if (bfq_bfqq_fifo_expire(bfqq))
 982		return NULL;
 983
 984	bfq_mark_bfqq_fifo_expire(bfqq);
 985
 986	rq = rq_entry_fifo(bfqq->fifo.next);
 987
 988	if (rq == last || ktime_get_ns() < rq->fifo_time)
 989		return NULL;
 990
 991	bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
 992	return rq;
 993}
 994
 995static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
 996					struct bfq_queue *bfqq,
 997					struct request *last)
 998{
 999	struct rb_node *rbnext = rb_next(&last->rb_node);
1000	struct rb_node *rbprev = rb_prev(&last->rb_node);
1001	struct request *next, *prev = NULL;
1002
1003	/* Follow expired path, else get first next available. */
1004	next = bfq_check_fifo(bfqq, last);
1005	if (next)
1006		return next;
1007
1008	if (rbprev)
1009		prev = rb_entry_rq(rbprev);
1010
1011	if (rbnext)
1012		next = rb_entry_rq(rbnext);
1013	else {
1014		rbnext = rb_first(&bfqq->sort_list);
1015		if (rbnext && rbnext != &last->rb_node)
1016			next = rb_entry_rq(rbnext);
1017	}
1018
1019	return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
1020}
1021
1022/* see the definition of bfq_async_charge_factor for details */
1023static unsigned long bfq_serv_to_charge(struct request *rq,
1024					struct bfq_queue *bfqq)
1025{
1026	if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
1027	    bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
1028		return blk_rq_sectors(rq);
1029
1030	return blk_rq_sectors(rq) * bfq_async_charge_factor;
 
 
 
 
 
 
 
 
 
1031}
1032
1033/**
1034 * bfq_updated_next_req - update the queue after a new next_rq selection.
1035 * @bfqd: the device data the queue belongs to.
1036 * @bfqq: the queue to update.
1037 *
1038 * If the first request of a queue changes we make sure that the queue
1039 * has enough budget to serve at least its first request (if the
1040 * request has grown).  We do this because if the queue has not enough
1041 * budget for its first request, it has to go through two dispatch
1042 * rounds to actually get it dispatched.
1043 */
1044static void bfq_updated_next_req(struct bfq_data *bfqd,
1045				 struct bfq_queue *bfqq)
1046{
1047	struct bfq_entity *entity = &bfqq->entity;
1048	struct request *next_rq = bfqq->next_rq;
1049	unsigned long new_budget;
1050
1051	if (!next_rq)
1052		return;
1053
1054	if (bfqq == bfqd->in_service_queue)
1055		/*
1056		 * In order not to break guarantees, budgets cannot be
1057		 * changed after an entity has been selected.
1058		 */
1059		return;
1060
1061	new_budget = max_t(unsigned long,
1062			   max_t(unsigned long, bfqq->max_budget,
1063				 bfq_serv_to_charge(next_rq, bfqq)),
1064			   entity->service);
1065	if (entity->budget != new_budget) {
1066		entity->budget = new_budget;
1067		bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
1068					 new_budget);
1069		bfq_requeue_bfqq(bfqd, bfqq, false);
1070	}
1071}
1072
1073static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
1074{
1075	u64 dur;
1076
1077	if (bfqd->bfq_wr_max_time > 0)
1078		return bfqd->bfq_wr_max_time;
1079
1080	dur = bfqd->rate_dur_prod;
1081	do_div(dur, bfqd->peak_rate);
1082
1083	/*
1084	 * Limit duration between 3 and 25 seconds. The upper limit
1085	 * has been conservatively set after the following worst case:
1086	 * on a QEMU/KVM virtual machine
1087	 * - running in a slow PC
1088	 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
1089	 * - serving a heavy I/O workload, such as the sequential reading
1090	 *   of several files
1091	 * mplayer took 23 seconds to start, if constantly weight-raised.
1092	 *
1093	 * As for higher values than that accommodating the above bad
1094	 * scenario, tests show that higher values would often yield
1095	 * the opposite of the desired result, i.e., would worsen
1096	 * responsiveness by allowing non-interactive applications to
1097	 * preserve weight raising for too long.
1098	 *
1099	 * On the other end, lower values than 3 seconds make it
1100	 * difficult for most interactive tasks to complete their jobs
1101	 * before weight-raising finishes.
1102	 */
1103	return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
 
 
 
 
 
1104}
1105
1106/* switch back from soft real-time to interactive weight raising */
1107static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
1108					  struct bfq_data *bfqd)
1109{
1110	bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1111	bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1112	bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
1113}
1114
1115static void
1116bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
1117		      struct bfq_io_cq *bic, bool bfq_already_existing)
1118{
1119	unsigned int old_wr_coeff = 1;
1120	bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
1121
1122	if (bic->saved_has_short_ttime)
1123		bfq_mark_bfqq_has_short_ttime(bfqq);
1124	else
1125		bfq_clear_bfqq_has_short_ttime(bfqq);
1126
1127	if (bic->saved_IO_bound)
1128		bfq_mark_bfqq_IO_bound(bfqq);
1129	else
1130		bfq_clear_bfqq_IO_bound(bfqq);
1131
1132	bfqq->last_serv_time_ns = bic->saved_last_serv_time_ns;
1133	bfqq->inject_limit = bic->saved_inject_limit;
1134	bfqq->decrease_time_jif = bic->saved_decrease_time_jif;
1135
1136	bfqq->entity.new_weight = bic->saved_weight;
1137	bfqq->ttime = bic->saved_ttime;
1138	bfqq->io_start_time = bic->saved_io_start_time;
1139	bfqq->tot_idle_time = bic->saved_tot_idle_time;
1140	/*
1141	 * Restore weight coefficient only if low_latency is on
1142	 */
1143	if (bfqd->low_latency) {
1144		old_wr_coeff = bfqq->wr_coeff;
1145		bfqq->wr_coeff = bic->saved_wr_coeff;
1146	}
1147	bfqq->service_from_wr = bic->saved_service_from_wr;
1148	bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
1149	bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
1150	bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
1151
1152	if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1153	    time_is_before_jiffies(bfqq->last_wr_start_finish +
1154				   bfqq->wr_cur_max_time))) {
1155		if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1156		    !bfq_bfqq_in_large_burst(bfqq) &&
1157		    time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1158					     bfq_wr_duration(bfqd))) {
1159			switch_back_to_interactive_wr(bfqq, bfqd);
1160		} else {
1161			bfqq->wr_coeff = 1;
1162			bfq_log_bfqq(bfqq->bfqd, bfqq,
1163				     "resume state: switching off wr");
1164		}
1165	}
1166
1167	/* make sure weight will be updated, however we got here */
1168	bfqq->entity.prio_changed = 1;
1169
1170	if (likely(!busy))
1171		return;
1172
1173	if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1174		bfqd->wr_busy_queues++;
1175	else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1176		bfqd->wr_busy_queues--;
1177}
1178
1179static int bfqq_process_refs(struct bfq_queue *bfqq)
1180{
1181	return bfqq->ref - bfqq->entity.allocated -
1182		bfqq->entity.on_st_or_in_serv -
1183		(bfqq->weight_counter != NULL) - bfqq->stable_ref;
1184}
1185
1186/* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1187static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1188{
1189	struct bfq_queue *item;
1190	struct hlist_node *n;
1191
1192	hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1193		hlist_del_init(&item->burst_list_node);
1194
1195	/*
1196	 * Start the creation of a new burst list only if there is no
1197	 * active queue. See comments on the conditional invocation of
1198	 * bfq_handle_burst().
1199	 */
1200	if (bfq_tot_busy_queues(bfqd) == 0) {
1201		hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1202		bfqd->burst_size = 1;
1203	} else
1204		bfqd->burst_size = 0;
1205
1206	bfqd->burst_parent_entity = bfqq->entity.parent;
1207}
1208
1209/* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1210static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1211{
1212	/* Increment burst size to take into account also bfqq */
1213	bfqd->burst_size++;
1214
1215	if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1216		struct bfq_queue *pos, *bfqq_item;
1217		struct hlist_node *n;
1218
1219		/*
1220		 * Enough queues have been activated shortly after each
1221		 * other to consider this burst as large.
1222		 */
1223		bfqd->large_burst = true;
1224
1225		/*
1226		 * We can now mark all queues in the burst list as
1227		 * belonging to a large burst.
1228		 */
1229		hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1230				     burst_list_node)
1231			bfq_mark_bfqq_in_large_burst(bfqq_item);
1232		bfq_mark_bfqq_in_large_burst(bfqq);
1233
1234		/*
1235		 * From now on, and until the current burst finishes, any
1236		 * new queue being activated shortly after the last queue
1237		 * was inserted in the burst can be immediately marked as
1238		 * belonging to a large burst. So the burst list is not
1239		 * needed any more. Remove it.
1240		 */
1241		hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1242					  burst_list_node)
1243			hlist_del_init(&pos->burst_list_node);
1244	} else /*
1245		* Burst not yet large: add bfqq to the burst list. Do
1246		* not increment the ref counter for bfqq, because bfqq
1247		* is removed from the burst list before freeing bfqq
1248		* in put_queue.
1249		*/
1250		hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1251}
1252
1253/*
1254 * If many queues belonging to the same group happen to be created
1255 * shortly after each other, then the processes associated with these
1256 * queues have typically a common goal. In particular, bursts of queue
1257 * creations are usually caused by services or applications that spawn
1258 * many parallel threads/processes. Examples are systemd during boot,
1259 * or git grep. To help these processes get their job done as soon as
1260 * possible, it is usually better to not grant either weight-raising
1261 * or device idling to their queues, unless these queues must be
1262 * protected from the I/O flowing through other active queues.
1263 *
1264 * In this comment we describe, firstly, the reasons why this fact
1265 * holds, and, secondly, the next function, which implements the main
1266 * steps needed to properly mark these queues so that they can then be
1267 * treated in a different way.
1268 *
1269 * The above services or applications benefit mostly from a high
1270 * throughput: the quicker the requests of the activated queues are
1271 * cumulatively served, the sooner the target job of these queues gets
1272 * completed. As a consequence, weight-raising any of these queues,
1273 * which also implies idling the device for it, is almost always
1274 * counterproductive, unless there are other active queues to isolate
1275 * these new queues from. If there no other active queues, then
1276 * weight-raising these new queues just lowers throughput in most
1277 * cases.
1278 *
1279 * On the other hand, a burst of queue creations may be caused also by
1280 * the start of an application that does not consist of a lot of
1281 * parallel I/O-bound threads. In fact, with a complex application,
1282 * several short processes may need to be executed to start-up the
1283 * application. In this respect, to start an application as quickly as
1284 * possible, the best thing to do is in any case to privilege the I/O
1285 * related to the application with respect to all other
1286 * I/O. Therefore, the best strategy to start as quickly as possible
1287 * an application that causes a burst of queue creations is to
1288 * weight-raise all the queues created during the burst. This is the
1289 * exact opposite of the best strategy for the other type of bursts.
1290 *
1291 * In the end, to take the best action for each of the two cases, the
1292 * two types of bursts need to be distinguished. Fortunately, this
1293 * seems relatively easy, by looking at the sizes of the bursts. In
1294 * particular, we found a threshold such that only bursts with a
1295 * larger size than that threshold are apparently caused by
1296 * services or commands such as systemd or git grep. For brevity,
1297 * hereafter we call just 'large' these bursts. BFQ *does not*
1298 * weight-raise queues whose creation occurs in a large burst. In
1299 * addition, for each of these queues BFQ performs or does not perform
1300 * idling depending on which choice boosts the throughput more. The
1301 * exact choice depends on the device and request pattern at
1302 * hand.
1303 *
1304 * Unfortunately, false positives may occur while an interactive task
1305 * is starting (e.g., an application is being started). The
1306 * consequence is that the queues associated with the task do not
1307 * enjoy weight raising as expected. Fortunately these false positives
1308 * are very rare. They typically occur if some service happens to
1309 * start doing I/O exactly when the interactive task starts.
1310 *
1311 * Turning back to the next function, it is invoked only if there are
1312 * no active queues (apart from active queues that would belong to the
1313 * same, possible burst bfqq would belong to), and it implements all
1314 * the steps needed to detect the occurrence of a large burst and to
1315 * properly mark all the queues belonging to it (so that they can then
1316 * be treated in a different way). This goal is achieved by
1317 * maintaining a "burst list" that holds, temporarily, the queues that
1318 * belong to the burst in progress. The list is then used to mark
1319 * these queues as belonging to a large burst if the burst does become
1320 * large. The main steps are the following.
1321 *
1322 * . when the very first queue is created, the queue is inserted into the
1323 *   list (as it could be the first queue in a possible burst)
1324 *
1325 * . if the current burst has not yet become large, and a queue Q that does
1326 *   not yet belong to the burst is activated shortly after the last time
1327 *   at which a new queue entered the burst list, then the function appends
1328 *   Q to the burst list
1329 *
1330 * . if, as a consequence of the previous step, the burst size reaches
1331 *   the large-burst threshold, then
1332 *
1333 *     . all the queues in the burst list are marked as belonging to a
1334 *       large burst
1335 *
1336 *     . the burst list is deleted; in fact, the burst list already served
1337 *       its purpose (keeping temporarily track of the queues in a burst,
1338 *       so as to be able to mark them as belonging to a large burst in the
1339 *       previous sub-step), and now is not needed any more
1340 *
1341 *     . the device enters a large-burst mode
1342 *
1343 * . if a queue Q that does not belong to the burst is created while
1344 *   the device is in large-burst mode and shortly after the last time
1345 *   at which a queue either entered the burst list or was marked as
1346 *   belonging to the current large burst, then Q is immediately marked
1347 *   as belonging to a large burst.
1348 *
1349 * . if a queue Q that does not belong to the burst is created a while
1350 *   later, i.e., not shortly after, than the last time at which a queue
1351 *   either entered the burst list or was marked as belonging to the
1352 *   current large burst, then the current burst is deemed as finished and:
1353 *
1354 *        . the large-burst mode is reset if set
1355 *
1356 *        . the burst list is emptied
1357 *
1358 *        . Q is inserted in the burst list, as Q may be the first queue
1359 *          in a possible new burst (then the burst list contains just Q
1360 *          after this step).
1361 */
1362static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1363{
1364	/*
1365	 * If bfqq is already in the burst list or is part of a large
1366	 * burst, or finally has just been split, then there is
1367	 * nothing else to do.
1368	 */
1369	if (!hlist_unhashed(&bfqq->burst_list_node) ||
1370	    bfq_bfqq_in_large_burst(bfqq) ||
1371	    time_is_after_eq_jiffies(bfqq->split_time +
1372				     msecs_to_jiffies(10)))
1373		return;
1374
1375	/*
1376	 * If bfqq's creation happens late enough, or bfqq belongs to
1377	 * a different group than the burst group, then the current
1378	 * burst is finished, and related data structures must be
1379	 * reset.
1380	 *
1381	 * In this respect, consider the special case where bfqq is
1382	 * the very first queue created after BFQ is selected for this
1383	 * device. In this case, last_ins_in_burst and
1384	 * burst_parent_entity are not yet significant when we get
1385	 * here. But it is easy to verify that, whether or not the
1386	 * following condition is true, bfqq will end up being
1387	 * inserted into the burst list. In particular the list will
1388	 * happen to contain only bfqq. And this is exactly what has
1389	 * to happen, as bfqq may be the first queue of the first
1390	 * burst.
1391	 */
1392	if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1393	    bfqd->bfq_burst_interval) ||
1394	    bfqq->entity.parent != bfqd->burst_parent_entity) {
1395		bfqd->large_burst = false;
1396		bfq_reset_burst_list(bfqd, bfqq);
1397		goto end;
1398	}
1399
1400	/*
1401	 * If we get here, then bfqq is being activated shortly after the
1402	 * last queue. So, if the current burst is also large, we can mark
1403	 * bfqq as belonging to this large burst immediately.
1404	 */
1405	if (bfqd->large_burst) {
1406		bfq_mark_bfqq_in_large_burst(bfqq);
1407		goto end;
1408	}
1409
1410	/*
1411	 * If we get here, then a large-burst state has not yet been
1412	 * reached, but bfqq is being activated shortly after the last
1413	 * queue. Then we add bfqq to the burst.
1414	 */
1415	bfq_add_to_burst(bfqd, bfqq);
1416end:
1417	/*
1418	 * At this point, bfqq either has been added to the current
1419	 * burst or has caused the current burst to terminate and a
1420	 * possible new burst to start. In particular, in the second
1421	 * case, bfqq has become the first queue in the possible new
1422	 * burst.  In both cases last_ins_in_burst needs to be moved
1423	 * forward.
1424	 */
1425	bfqd->last_ins_in_burst = jiffies;
1426}
1427
1428static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1429{
1430	struct bfq_entity *entity = &bfqq->entity;
1431
1432	return entity->budget - entity->service;
1433}
1434
1435/*
1436 * If enough samples have been computed, return the current max budget
1437 * stored in bfqd, which is dynamically updated according to the
1438 * estimated disk peak rate; otherwise return the default max budget
1439 */
1440static int bfq_max_budget(struct bfq_data *bfqd)
1441{
1442	if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1443		return bfq_default_max_budget;
1444	else
1445		return bfqd->bfq_max_budget;
1446}
1447
1448/*
1449 * Return min budget, which is a fraction of the current or default
1450 * max budget (trying with 1/32)
1451 */
1452static int bfq_min_budget(struct bfq_data *bfqd)
1453{
1454	if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1455		return bfq_default_max_budget / 32;
1456	else
1457		return bfqd->bfq_max_budget / 32;
1458}
1459
1460/*
1461 * The next function, invoked after the input queue bfqq switches from
1462 * idle to busy, updates the budget of bfqq. The function also tells
1463 * whether the in-service queue should be expired, by returning
1464 * true. The purpose of expiring the in-service queue is to give bfqq
1465 * the chance to possibly preempt the in-service queue, and the reason
1466 * for preempting the in-service queue is to achieve one of the two
1467 * goals below.
1468 *
1469 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1470 * expired because it has remained idle. In particular, bfqq may have
1471 * expired for one of the following two reasons:
1472 *
1473 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1474 *   and did not make it to issue a new request before its last
1475 *   request was served;
1476 *
1477 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1478 *   a new request before the expiration of the idling-time.
1479 *
1480 * Even if bfqq has expired for one of the above reasons, the process
1481 * associated with the queue may be however issuing requests greedily,
1482 * and thus be sensitive to the bandwidth it receives (bfqq may have
1483 * remained idle for other reasons: CPU high load, bfqq not enjoying
1484 * idling, I/O throttling somewhere in the path from the process to
1485 * the I/O scheduler, ...). But if, after every expiration for one of
1486 * the above two reasons, bfqq has to wait for the service of at least
1487 * one full budget of another queue before being served again, then
1488 * bfqq is likely to get a much lower bandwidth or resource time than
1489 * its reserved ones. To address this issue, two countermeasures need
1490 * to be taken.
1491 *
1492 * First, the budget and the timestamps of bfqq need to be updated in
1493 * a special way on bfqq reactivation: they need to be updated as if
1494 * bfqq did not remain idle and did not expire. In fact, if they are
1495 * computed as if bfqq expired and remained idle until reactivation,
1496 * then the process associated with bfqq is treated as if, instead of
1497 * being greedy, it stopped issuing requests when bfqq remained idle,
1498 * and restarts issuing requests only on this reactivation. In other
1499 * words, the scheduler does not help the process recover the "service
1500 * hole" between bfqq expiration and reactivation. As a consequence,
1501 * the process receives a lower bandwidth than its reserved one. In
1502 * contrast, to recover this hole, the budget must be updated as if
1503 * bfqq was not expired at all before this reactivation, i.e., it must
1504 * be set to the value of the remaining budget when bfqq was
1505 * expired. Along the same line, timestamps need to be assigned the
1506 * value they had the last time bfqq was selected for service, i.e.,
1507 * before last expiration. Thus timestamps need to be back-shifted
1508 * with respect to their normal computation (see [1] for more details
1509 * on this tricky aspect).
1510 *
1511 * Secondly, to allow the process to recover the hole, the in-service
1512 * queue must be expired too, to give bfqq the chance to preempt it
1513 * immediately. In fact, if bfqq has to wait for a full budget of the
1514 * in-service queue to be completed, then it may become impossible to
1515 * let the process recover the hole, even if the back-shifted
1516 * timestamps of bfqq are lower than those of the in-service queue. If
1517 * this happens for most or all of the holes, then the process may not
1518 * receive its reserved bandwidth. In this respect, it is worth noting
1519 * that, being the service of outstanding requests unpreemptible, a
1520 * little fraction of the holes may however be unrecoverable, thereby
1521 * causing a little loss of bandwidth.
1522 *
1523 * The last important point is detecting whether bfqq does need this
1524 * bandwidth recovery. In this respect, the next function deems the
1525 * process associated with bfqq greedy, and thus allows it to recover
1526 * the hole, if: 1) the process is waiting for the arrival of a new
1527 * request (which implies that bfqq expired for one of the above two
1528 * reasons), and 2) such a request has arrived soon. The first
1529 * condition is controlled through the flag non_blocking_wait_rq,
1530 * while the second through the flag arrived_in_time. If both
1531 * conditions hold, then the function computes the budget in the
1532 * above-described special way, and signals that the in-service queue
1533 * should be expired. Timestamp back-shifting is done later in
1534 * __bfq_activate_entity.
1535 *
1536 * 2. Reduce latency. Even if timestamps are not backshifted to let
1537 * the process associated with bfqq recover a service hole, bfqq may
1538 * however happen to have, after being (re)activated, a lower finish
1539 * timestamp than the in-service queue.	 That is, the next budget of
1540 * bfqq may have to be completed before the one of the in-service
1541 * queue. If this is the case, then preempting the in-service queue
1542 * allows this goal to be achieved, apart from the unpreemptible,
1543 * outstanding requests mentioned above.
1544 *
1545 * Unfortunately, regardless of which of the above two goals one wants
1546 * to achieve, service trees need first to be updated to know whether
1547 * the in-service queue must be preempted. To have service trees
1548 * correctly updated, the in-service queue must be expired and
1549 * rescheduled, and bfqq must be scheduled too. This is one of the
1550 * most costly operations (in future versions, the scheduling
1551 * mechanism may be re-designed in such a way to make it possible to
1552 * know whether preemption is needed without needing to update service
1553 * trees). In addition, queue preemptions almost always cause random
1554 * I/O, which may in turn cause loss of throughput. Finally, there may
1555 * even be no in-service queue when the next function is invoked (so,
1556 * no queue to compare timestamps with). Because of these facts, the
1557 * next function adopts the following simple scheme to avoid costly
1558 * operations, too frequent preemptions and too many dependencies on
1559 * the state of the scheduler: it requests the expiration of the
1560 * in-service queue (unconditionally) only for queues that need to
1561 * recover a hole. Then it delegates to other parts of the code the
1562 * responsibility of handling the above case 2.
1563 */
1564static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1565						struct bfq_queue *bfqq,
1566						bool arrived_in_time)
 
1567{
1568	struct bfq_entity *entity = &bfqq->entity;
1569
1570	/*
1571	 * In the next compound condition, we check also whether there
1572	 * is some budget left, because otherwise there is no point in
1573	 * trying to go on serving bfqq with this same budget: bfqq
1574	 * would be expired immediately after being selected for
1575	 * service. This would only cause useless overhead.
1576	 */
1577	if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1578	    bfq_bfqq_budget_left(bfqq) > 0) {
1579		/*
1580		 * We do not clear the flag non_blocking_wait_rq here, as
1581		 * the latter is used in bfq_activate_bfqq to signal
1582		 * that timestamps need to be back-shifted (and is
1583		 * cleared right after).
1584		 */
1585
1586		/*
1587		 * In next assignment we rely on that either
1588		 * entity->service or entity->budget are not updated
1589		 * on expiration if bfqq is empty (see
1590		 * __bfq_bfqq_recalc_budget). Thus both quantities
1591		 * remain unchanged after such an expiration, and the
1592		 * following statement therefore assigns to
1593		 * entity->budget the remaining budget on such an
1594		 * expiration.
 
 
 
1595		 */
1596		entity->budget = min_t(unsigned long,
1597				       bfq_bfqq_budget_left(bfqq),
1598				       bfqq->max_budget);
1599
1600		/*
1601		 * At this point, we have used entity->service to get
1602		 * the budget left (needed for updating
1603		 * entity->budget). Thus we finally can, and have to,
1604		 * reset entity->service. The latter must be reset
1605		 * because bfqq would otherwise be charged again for
1606		 * the service it has received during its previous
1607		 * service slot(s).
1608		 */
1609		entity->service = 0;
1610
1611		return true;
1612	}
1613
1614	/*
1615	 * We can finally complete expiration, by setting service to 0.
1616	 */
1617	entity->service = 0;
1618	entity->budget = max_t(unsigned long, bfqq->max_budget,
1619			       bfq_serv_to_charge(bfqq->next_rq, bfqq));
1620	bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1621	return false;
 
 
 
 
 
 
 
 
 
1622}
1623
1624/*
1625 * Return the farthest past time instant according to jiffies
1626 * macros.
1627 */
1628static unsigned long bfq_smallest_from_now(void)
1629{
1630	return jiffies - MAX_JIFFY_OFFSET;
1631}
1632
1633static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1634					     struct bfq_queue *bfqq,
1635					     unsigned int old_wr_coeff,
1636					     bool wr_or_deserves_wr,
1637					     bool interactive,
1638					     bool in_burst,
1639					     bool soft_rt)
1640{
1641	if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1642		/* start a weight-raising period */
1643		if (interactive) {
1644			bfqq->service_from_wr = 0;
1645			bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1646			bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1647		} else {
1648			/*
1649			 * No interactive weight raising in progress
1650			 * here: assign minus infinity to
1651			 * wr_start_at_switch_to_srt, to make sure
1652			 * that, at the end of the soft-real-time
1653			 * weight raising periods that is starting
1654			 * now, no interactive weight-raising period
1655			 * may be wrongly considered as still in
1656			 * progress (and thus actually started by
1657			 * mistake).
1658			 */
1659			bfqq->wr_start_at_switch_to_srt =
1660				bfq_smallest_from_now();
1661			bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1662				BFQ_SOFTRT_WEIGHT_FACTOR;
1663			bfqq->wr_cur_max_time =
1664				bfqd->bfq_wr_rt_max_time;
1665		}
1666
1667		/*
1668		 * If needed, further reduce budget to make sure it is
1669		 * close to bfqq's backlog, so as to reduce the
1670		 * scheduling-error component due to a too large
1671		 * budget. Do not care about throughput consequences,
1672		 * but only about latency. Finally, do not assign a
1673		 * too small budget either, to avoid increasing
1674		 * latency by causing too frequent expirations.
1675		 */
1676		bfqq->entity.budget = min_t(unsigned long,
1677					    bfqq->entity.budget,
1678					    2 * bfq_min_budget(bfqd));
1679	} else if (old_wr_coeff > 1) {
1680		if (interactive) { /* update wr coeff and duration */
1681			bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1682			bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1683		} else if (in_burst)
1684			bfqq->wr_coeff = 1;
1685		else if (soft_rt) {
1686			/*
1687			 * The application is now or still meeting the
1688			 * requirements for being deemed soft rt.  We
1689			 * can then correctly and safely (re)charge
1690			 * the weight-raising duration for the
1691			 * application with the weight-raising
1692			 * duration for soft rt applications.
1693			 *
1694			 * In particular, doing this recharge now, i.e.,
1695			 * before the weight-raising period for the
1696			 * application finishes, reduces the probability
1697			 * of the following negative scenario:
1698			 * 1) the weight of a soft rt application is
1699			 *    raised at startup (as for any newly
1700			 *    created application),
1701			 * 2) since the application is not interactive,
1702			 *    at a certain time weight-raising is
1703			 *    stopped for the application,
1704			 * 3) at that time the application happens to
1705			 *    still have pending requests, and hence
1706			 *    is destined to not have a chance to be
1707			 *    deemed soft rt before these requests are
1708			 *    completed (see the comments to the
1709			 *    function bfq_bfqq_softrt_next_start()
1710			 *    for details on soft rt detection),
1711			 * 4) these pending requests experience a high
1712			 *    latency because the application is not
1713			 *    weight-raised while they are pending.
1714			 */
1715			if (bfqq->wr_cur_max_time !=
1716				bfqd->bfq_wr_rt_max_time) {
1717				bfqq->wr_start_at_switch_to_srt =
1718					bfqq->last_wr_start_finish;
1719
1720				bfqq->wr_cur_max_time =
1721					bfqd->bfq_wr_rt_max_time;
1722				bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1723					BFQ_SOFTRT_WEIGHT_FACTOR;
1724			}
1725			bfqq->last_wr_start_finish = jiffies;
1726		}
1727	}
1728}
1729
1730static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1731					struct bfq_queue *bfqq)
1732{
1733	return bfqq->dispatched == 0 &&
1734		time_is_before_jiffies(
1735			bfqq->budget_timeout +
1736			bfqd->bfq_wr_min_idle_time);
1737}
1738
1739
1740/*
1741 * Return true if bfqq is in a higher priority class, or has a higher
1742 * weight than the in-service queue.
1743 */
1744static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq,
1745					    struct bfq_queue *in_serv_bfqq)
1746{
1747	int bfqq_weight, in_serv_weight;
1748
1749	if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class)
1750		return true;
1751
1752	if (in_serv_bfqq->entity.parent == bfqq->entity.parent) {
1753		bfqq_weight = bfqq->entity.weight;
1754		in_serv_weight = in_serv_bfqq->entity.weight;
1755	} else {
1756		if (bfqq->entity.parent)
1757			bfqq_weight = bfqq->entity.parent->weight;
1758		else
1759			bfqq_weight = bfqq->entity.weight;
1760		if (in_serv_bfqq->entity.parent)
1761			in_serv_weight = in_serv_bfqq->entity.parent->weight;
1762		else
1763			in_serv_weight = in_serv_bfqq->entity.weight;
1764	}
1765
1766	return bfqq_weight > in_serv_weight;
1767}
1768
1769static bool bfq_better_to_idle(struct bfq_queue *bfqq);
1770
1771static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1772					     struct bfq_queue *bfqq,
1773					     int old_wr_coeff,
1774					     struct request *rq,
1775					     bool *interactive)
1776{
1777	bool soft_rt, in_burst,	wr_or_deserves_wr,
1778		bfqq_wants_to_preempt,
1779		idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1780		/*
1781		 * See the comments on
1782		 * bfq_bfqq_update_budg_for_activation for
1783		 * details on the usage of the next variable.
1784		 */
1785		arrived_in_time =  ktime_get_ns() <=
1786			bfqq->ttime.last_end_request +
1787			bfqd->bfq_slice_idle * 3;
1788
1789
1790	/*
1791	 * bfqq deserves to be weight-raised if:
1792	 * - it is sync,
1793	 * - it does not belong to a large burst,
1794	 * - it has been idle for enough time or is soft real-time,
1795	 * - is linked to a bfq_io_cq (it is not shared in any sense),
1796	 * - has a default weight (otherwise we assume the user wanted
1797	 *   to control its weight explicitly)
1798	 */
1799	in_burst = bfq_bfqq_in_large_burst(bfqq);
1800	soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1801		!BFQQ_TOTALLY_SEEKY(bfqq) &&
1802		!in_burst &&
1803		time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1804		bfqq->dispatched == 0 &&
1805		bfqq->entity.new_weight == 40;
1806	*interactive = !in_burst && idle_for_long_time &&
1807		bfqq->entity.new_weight == 40;
1808	/*
1809	 * Merged bfq_queues are kept out of weight-raising
1810	 * (low-latency) mechanisms. The reason is that these queues
1811	 * are usually created for non-interactive and
1812	 * non-soft-real-time tasks. Yet this is not the case for
1813	 * stably-merged queues. These queues are merged just because
1814	 * they are created shortly after each other. So they may
1815	 * easily serve the I/O of an interactive or soft-real time
1816	 * application, if the application happens to spawn multiple
1817	 * processes. So let also stably-merged queued enjoy weight
1818	 * raising.
1819	 */
1820	wr_or_deserves_wr = bfqd->low_latency &&
1821		(bfqq->wr_coeff > 1 ||
1822		 (bfq_bfqq_sync(bfqq) &&
1823		  (bfqq->bic || RQ_BIC(rq)->stably_merged) &&
1824		   (*interactive || soft_rt)));
1825
1826	/*
1827	 * Using the last flag, update budget and check whether bfqq
1828	 * may want to preempt the in-service queue.
1829	 */
1830	bfqq_wants_to_preempt =
1831		bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1832						    arrived_in_time);
 
1833
1834	/*
1835	 * If bfqq happened to be activated in a burst, but has been
1836	 * idle for much more than an interactive queue, then we
1837	 * assume that, in the overall I/O initiated in the burst, the
1838	 * I/O associated with bfqq is finished. So bfqq does not need
1839	 * to be treated as a queue belonging to a burst
1840	 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1841	 * if set, and remove bfqq from the burst list if it's
1842	 * there. We do not decrement burst_size, because the fact
1843	 * that bfqq does not need to belong to the burst list any
1844	 * more does not invalidate the fact that bfqq was created in
1845	 * a burst.
1846	 */
1847	if (likely(!bfq_bfqq_just_created(bfqq)) &&
1848	    idle_for_long_time &&
1849	    time_is_before_jiffies(
1850		    bfqq->budget_timeout +
1851		    msecs_to_jiffies(10000))) {
1852		hlist_del_init(&bfqq->burst_list_node);
1853		bfq_clear_bfqq_in_large_burst(bfqq);
1854	}
1855
1856	bfq_clear_bfqq_just_created(bfqq);
1857
 
 
 
 
 
 
 
 
 
 
 
1858	if (bfqd->low_latency) {
1859		if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1860			/* wraparound */
1861			bfqq->split_time =
1862				jiffies - bfqd->bfq_wr_min_idle_time - 1;
1863
1864		if (time_is_before_jiffies(bfqq->split_time +
1865					   bfqd->bfq_wr_min_idle_time)) {
1866			bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1867							 old_wr_coeff,
1868							 wr_or_deserves_wr,
1869							 *interactive,
1870							 in_burst,
1871							 soft_rt);
1872
1873			if (old_wr_coeff != bfqq->wr_coeff)
1874				bfqq->entity.prio_changed = 1;
1875		}
1876	}
1877
1878	bfqq->last_idle_bklogged = jiffies;
1879	bfqq->service_from_backlogged = 0;
1880	bfq_clear_bfqq_softrt_update(bfqq);
1881
1882	bfq_add_bfqq_busy(bfqq);
1883
1884	/*
1885	 * Expire in-service queue if preemption may be needed for
1886	 * guarantees or throughput. As for guarantees, we care
1887	 * explicitly about two cases. The first is that bfqq has to
1888	 * recover a service hole, as explained in the comments on
1889	 * bfq_bfqq_update_budg_for_activation(), i.e., that
1890	 * bfqq_wants_to_preempt is true. However, if bfqq does not
1891	 * carry time-critical I/O, then bfqq's bandwidth is less
1892	 * important than that of queues that carry time-critical I/O.
1893	 * So, as a further constraint, we consider this case only if
1894	 * bfqq is at least as weight-raised, i.e., at least as time
1895	 * critical, as the in-service queue.
1896	 *
1897	 * The second case is that bfqq is in a higher priority class,
1898	 * or has a higher weight than the in-service queue. If this
1899	 * condition does not hold, we don't care because, even if
1900	 * bfqq does not start to be served immediately, the resulting
1901	 * delay for bfqq's I/O is however lower or much lower than
1902	 * the ideal completion time to be guaranteed to bfqq's I/O.
1903	 *
1904	 * In both cases, preemption is needed only if, according to
1905	 * the timestamps of both bfqq and of the in-service queue,
1906	 * bfqq actually is the next queue to serve. So, to reduce
1907	 * useless preemptions, the return value of
1908	 * next_queue_may_preempt() is considered in the next compound
1909	 * condition too. Yet next_queue_may_preempt() just checks a
1910	 * simple, necessary condition for bfqq to be the next queue
1911	 * to serve. In fact, to evaluate a sufficient condition, the
1912	 * timestamps of the in-service queue would need to be
1913	 * updated, and this operation is quite costly (see the
1914	 * comments on bfq_bfqq_update_budg_for_activation()).
1915	 *
1916	 * As for throughput, we ask bfq_better_to_idle() whether we
1917	 * still need to plug I/O dispatching. If bfq_better_to_idle()
1918	 * says no, then plugging is not needed any longer, either to
1919	 * boost throughput or to perserve service guarantees. Then
1920	 * the best option is to stop plugging I/O, as not doing so
1921	 * would certainly lower throughput. We may end up in this
1922	 * case if: (1) upon a dispatch attempt, we detected that it
1923	 * was better to plug I/O dispatch, and to wait for a new
1924	 * request to arrive for the currently in-service queue, but
1925	 * (2) this switch of bfqq to busy changes the scenario.
1926	 */
1927	if (bfqd->in_service_queue &&
1928	    ((bfqq_wants_to_preempt &&
1929	      bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) ||
1930	     bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue) ||
1931	     !bfq_better_to_idle(bfqd->in_service_queue)) &&
1932	    next_queue_may_preempt(bfqd))
1933		bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1934				false, BFQQE_PREEMPTED);
1935}
1936
1937static void bfq_reset_inject_limit(struct bfq_data *bfqd,
1938				   struct bfq_queue *bfqq)
1939{
1940	/* invalidate baseline total service time */
1941	bfqq->last_serv_time_ns = 0;
1942
1943	/*
1944	 * Reset pointer in case we are waiting for
1945	 * some request completion.
1946	 */
1947	bfqd->waited_rq = NULL;
1948
1949	/*
1950	 * If bfqq has a short think time, then start by setting the
1951	 * inject limit to 0 prudentially, because the service time of
1952	 * an injected I/O request may be higher than the think time
1953	 * of bfqq, and therefore, if one request was injected when
1954	 * bfqq remains empty, this injected request might delay the
1955	 * service of the next I/O request for bfqq significantly. In
1956	 * case bfqq can actually tolerate some injection, then the
1957	 * adaptive update will however raise the limit soon. This
1958	 * lucky circumstance holds exactly because bfqq has a short
1959	 * think time, and thus, after remaining empty, is likely to
1960	 * get new I/O enqueued---and then completed---before being
1961	 * expired. This is the very pattern that gives the
1962	 * limit-update algorithm the chance to measure the effect of
1963	 * injection on request service times, and then to update the
1964	 * limit accordingly.
1965	 *
1966	 * However, in the following special case, the inject limit is
1967	 * left to 1 even if the think time is short: bfqq's I/O is
1968	 * synchronized with that of some other queue, i.e., bfqq may
1969	 * receive new I/O only after the I/O of the other queue is
1970	 * completed. Keeping the inject limit to 1 allows the
1971	 * blocking I/O to be served while bfqq is in service. And
1972	 * this is very convenient both for bfqq and for overall
1973	 * throughput, as explained in detail in the comments in
1974	 * bfq_update_has_short_ttime().
1975	 *
1976	 * On the opposite end, if bfqq has a long think time, then
1977	 * start directly by 1, because:
1978	 * a) on the bright side, keeping at most one request in
1979	 * service in the drive is unlikely to cause any harm to the
1980	 * latency of bfqq's requests, as the service time of a single
1981	 * request is likely to be lower than the think time of bfqq;
1982	 * b) on the downside, after becoming empty, bfqq is likely to
1983	 * expire before getting its next request. With this request
1984	 * arrival pattern, it is very hard to sample total service
1985	 * times and update the inject limit accordingly (see comments
1986	 * on bfq_update_inject_limit()). So the limit is likely to be
1987	 * never, or at least seldom, updated.  As a consequence, by
1988	 * setting the limit to 1, we avoid that no injection ever
1989	 * occurs with bfqq. On the downside, this proactive step
1990	 * further reduces chances to actually compute the baseline
1991	 * total service time. Thus it reduces chances to execute the
1992	 * limit-update algorithm and possibly raise the limit to more
1993	 * than 1.
1994	 */
1995	if (bfq_bfqq_has_short_ttime(bfqq))
1996		bfqq->inject_limit = 0;
1997	else
1998		bfqq->inject_limit = 1;
1999
2000	bfqq->decrease_time_jif = jiffies;
2001}
2002
2003static void bfq_update_io_intensity(struct bfq_queue *bfqq, u64 now_ns)
2004{
2005	u64 tot_io_time = now_ns - bfqq->io_start_time;
2006
2007	if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfqq->dispatched == 0)
2008		bfqq->tot_idle_time +=
2009			now_ns - bfqq->ttime.last_end_request;
2010
2011	if (unlikely(bfq_bfqq_just_created(bfqq)))
2012		return;
2013
2014	/*
2015	 * Must be busy for at least about 80% of the time to be
2016	 * considered I/O bound.
2017	 */
2018	if (bfqq->tot_idle_time * 5 > tot_io_time)
2019		bfq_clear_bfqq_IO_bound(bfqq);
2020	else
2021		bfq_mark_bfqq_IO_bound(bfqq);
2022
2023	/*
2024	 * Keep an observation window of at most 200 ms in the past
2025	 * from now.
2026	 */
2027	if (tot_io_time > 200 * NSEC_PER_MSEC) {
2028		bfqq->io_start_time = now_ns - (tot_io_time>>1);
2029		bfqq->tot_idle_time >>= 1;
2030	}
2031}
2032
2033/*
2034 * Detect whether bfqq's I/O seems synchronized with that of some
2035 * other queue, i.e., whether bfqq, after remaining empty, happens to
2036 * receive new I/O only right after some I/O request of the other
2037 * queue has been completed. We call waker queue the other queue, and
2038 * we assume, for simplicity, that bfqq may have at most one waker
2039 * queue.
2040 *
2041 * A remarkable throughput boost can be reached by unconditionally
2042 * injecting the I/O of the waker queue, every time a new
2043 * bfq_dispatch_request happens to be invoked while I/O is being
2044 * plugged for bfqq.  In addition to boosting throughput, this
2045 * unblocks bfqq's I/O, thereby improving bandwidth and latency for
2046 * bfqq. Note that these same results may be achieved with the general
2047 * injection mechanism, but less effectively. For details on this
2048 * aspect, see the comments on the choice of the queue for injection
2049 * in bfq_select_queue().
2050 *
2051 * Turning back to the detection of a waker queue, a queue Q is deemed as a
2052 * waker queue for bfqq if, for three consecutive times, bfqq happens to become
2053 * non empty right after a request of Q has been completed within given
2054 * timeout. In this respect, even if bfqq is empty, we do not check for a waker
2055 * if it still has some in-flight I/O. In fact, in this case bfqq is actually
2056 * still being served by the drive, and may receive new I/O on the completion
2057 * of some of the in-flight requests. In particular, on the first time, Q is
2058 * tentatively set as a candidate waker queue, while on the third consecutive
2059 * time that Q is detected, the field waker_bfqq is set to Q, to confirm that Q
2060 * is a waker queue for bfqq. These detection steps are performed only if bfqq
2061 * has a long think time, so as to make it more likely that bfqq's I/O is
2062 * actually being blocked by a synchronization. This last filter, plus the
2063 * above three-times requirement and time limit for detection, make false
2064 * positives less likely.
2065 *
2066 * NOTE
2067 *
2068 * The sooner a waker queue is detected, the sooner throughput can be
2069 * boosted by injecting I/O from the waker queue. Fortunately,
2070 * detection is likely to be actually fast, for the following
2071 * reasons. While blocked by synchronization, bfqq has a long think
2072 * time. This implies that bfqq's inject limit is at least equal to 1
2073 * (see the comments in bfq_update_inject_limit()). So, thanks to
2074 * injection, the waker queue is likely to be served during the very
2075 * first I/O-plugging time interval for bfqq. This triggers the first
2076 * step of the detection mechanism. Thanks again to injection, the
2077 * candidate waker queue is then likely to be confirmed no later than
2078 * during the next I/O-plugging interval for bfqq.
2079 *
2080 * ISSUE
2081 *
2082 * On queue merging all waker information is lost.
2083 */
2084static void bfq_check_waker(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2085			    u64 now_ns)
2086{
2087	char waker_name[MAX_BFQQ_NAME_LENGTH];
2088
2089	if (!bfqd->last_completed_rq_bfqq ||
2090	    bfqd->last_completed_rq_bfqq == bfqq ||
2091	    bfq_bfqq_has_short_ttime(bfqq) ||
2092	    now_ns - bfqd->last_completion >= 4 * NSEC_PER_MSEC ||
2093	    bfqd->last_completed_rq_bfqq == &bfqd->oom_bfqq ||
2094	    bfqq == &bfqd->oom_bfqq)
2095		return;
2096
2097	/*
2098	 * We reset waker detection logic also if too much time has passed
2099 	 * since the first detection. If wakeups are rare, pointless idling
2100	 * doesn't hurt throughput that much. The condition below makes sure
2101	 * we do not uselessly idle blocking waker in more than 1/64 cases. 
2102	 */
2103	if (bfqd->last_completed_rq_bfqq !=
2104	    bfqq->tentative_waker_bfqq ||
2105	    now_ns > bfqq->waker_detection_started +
2106					128 * (u64)bfqd->bfq_slice_idle) {
2107		/*
2108		 * First synchronization detected with a
2109		 * candidate waker queue, or with a different
2110		 * candidate waker queue from the current one.
2111		 */
2112		bfqq->tentative_waker_bfqq =
2113			bfqd->last_completed_rq_bfqq;
2114		bfqq->num_waker_detections = 1;
2115		bfqq->waker_detection_started = now_ns;
2116		bfq_bfqq_name(bfqq->tentative_waker_bfqq, waker_name,
2117			      MAX_BFQQ_NAME_LENGTH);
2118		bfq_log_bfqq(bfqd, bfqq, "set tentative waker %s", waker_name);
2119	} else /* Same tentative waker queue detected again */
2120		bfqq->num_waker_detections++;
2121
2122	if (bfqq->num_waker_detections == 3) {
2123		bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
2124		bfqq->tentative_waker_bfqq = NULL;
2125		bfq_bfqq_name(bfqq->waker_bfqq, waker_name,
2126			      MAX_BFQQ_NAME_LENGTH);
2127		bfq_log_bfqq(bfqd, bfqq, "set waker %s", waker_name);
2128
2129		/*
2130		 * If the waker queue disappears, then
2131		 * bfqq->waker_bfqq must be reset. To
2132		 * this goal, we maintain in each
2133		 * waker queue a list, woken_list, of
2134		 * all the queues that reference the
2135		 * waker queue through their
2136		 * waker_bfqq pointer. When the waker
2137		 * queue exits, the waker_bfqq pointer
2138		 * of all the queues in the woken_list
2139		 * is reset.
2140		 *
2141		 * In addition, if bfqq is already in
2142		 * the woken_list of a waker queue,
2143		 * then, before being inserted into
2144		 * the woken_list of a new waker
2145		 * queue, bfqq must be removed from
2146		 * the woken_list of the old waker
2147		 * queue.
2148		 */
2149		if (!hlist_unhashed(&bfqq->woken_list_node))
2150			hlist_del_init(&bfqq->woken_list_node);
2151		hlist_add_head(&bfqq->woken_list_node,
2152			       &bfqd->last_completed_rq_bfqq->woken_list);
2153	}
2154}
2155
2156static void bfq_add_request(struct request *rq)
2157{
2158	struct bfq_queue *bfqq = RQ_BFQQ(rq);
2159	struct bfq_data *bfqd = bfqq->bfqd;
2160	struct request *next_rq, *prev;
2161	unsigned int old_wr_coeff = bfqq->wr_coeff;
2162	bool interactive = false;
2163	u64 now_ns = ktime_get_ns();
2164
2165	bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
2166	bfqq->queued[rq_is_sync(rq)]++;
2167	/*
2168	 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2169	 * may be read without holding the lock in bfq_has_work().
2170	 */
2171	WRITE_ONCE(bfqd->queued, bfqd->queued + 1);
2172
2173	if (bfq_bfqq_sync(bfqq) && RQ_BIC(rq)->requests <= 1) {
2174		bfq_check_waker(bfqd, bfqq, now_ns);
2175
2176		/*
2177		 * Periodically reset inject limit, to make sure that
2178		 * the latter eventually drops in case workload
2179		 * changes, see step (3) in the comments on
2180		 * bfq_update_inject_limit().
2181		 */
2182		if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2183					     msecs_to_jiffies(1000)))
2184			bfq_reset_inject_limit(bfqd, bfqq);
2185
2186		/*
2187		 * The following conditions must hold to setup a new
2188		 * sampling of total service time, and then a new
2189		 * update of the inject limit:
2190		 * - bfqq is in service, because the total service
2191		 *   time is evaluated only for the I/O requests of
2192		 *   the queues in service;
2193		 * - this is the right occasion to compute or to
2194		 *   lower the baseline total service time, because
2195		 *   there are actually no requests in the drive,
2196		 *   or
2197		 *   the baseline total service time is available, and
2198		 *   this is the right occasion to compute the other
2199		 *   quantity needed to update the inject limit, i.e.,
2200		 *   the total service time caused by the amount of
2201		 *   injection allowed by the current value of the
2202		 *   limit. It is the right occasion because injection
2203		 *   has actually been performed during the service
2204		 *   hole, and there are still in-flight requests,
2205		 *   which are very likely to be exactly the injected
2206		 *   requests, or part of them;
2207		 * - the minimum interval for sampling the total
2208		 *   service time and updating the inject limit has
2209		 *   elapsed.
2210		 */
2211		if (bfqq == bfqd->in_service_queue &&
2212		    (bfqd->rq_in_driver == 0 ||
2213		     (bfqq->last_serv_time_ns > 0 &&
2214		      bfqd->rqs_injected && bfqd->rq_in_driver > 0)) &&
2215		    time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2216					      msecs_to_jiffies(10))) {
2217			bfqd->last_empty_occupied_ns = ktime_get_ns();
2218			/*
2219			 * Start the state machine for measuring the
2220			 * total service time of rq: setting
2221			 * wait_dispatch will cause bfqd->waited_rq to
2222			 * be set when rq will be dispatched.
2223			 */
2224			bfqd->wait_dispatch = true;
2225			/*
2226			 * If there is no I/O in service in the drive,
2227			 * then possible injection occurred before the
2228			 * arrival of rq will not affect the total
2229			 * service time of rq. So the injection limit
2230			 * must not be updated as a function of such
2231			 * total service time, unless new injection
2232			 * occurs before rq is completed. To have the
2233			 * injection limit updated only in the latter
2234			 * case, reset rqs_injected here (rqs_injected
2235			 * will be set in case injection is performed
2236			 * on bfqq before rq is completed).
2237			 */
2238			if (bfqd->rq_in_driver == 0)
2239				bfqd->rqs_injected = false;
2240		}
2241	}
2242
2243	if (bfq_bfqq_sync(bfqq))
2244		bfq_update_io_intensity(bfqq, now_ns);
2245
2246	elv_rb_add(&bfqq->sort_list, rq);
2247
2248	/*
2249	 * Check if this request is a better next-serve candidate.
2250	 */
2251	prev = bfqq->next_rq;
2252	next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
2253	bfqq->next_rq = next_rq;
2254
2255	/*
2256	 * Adjust priority tree position, if next_rq changes.
2257	 * See comments on bfq_pos_tree_add_move() for the unlikely().
2258	 */
2259	if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
2260		bfq_pos_tree_add_move(bfqd, bfqq);
2261
2262	if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
2263		bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
2264						 rq, &interactive);
2265	else {
2266		if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
2267		    time_is_before_jiffies(
2268				bfqq->last_wr_start_finish +
2269				bfqd->bfq_wr_min_inter_arr_async)) {
2270			bfqq->wr_coeff = bfqd->bfq_wr_coeff;
2271			bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
2272
2273			bfqd->wr_busy_queues++;
2274			bfqq->entity.prio_changed = 1;
2275		}
2276		if (prev != bfqq->next_rq)
2277			bfq_updated_next_req(bfqd, bfqq);
2278	}
2279
2280	/*
2281	 * Assign jiffies to last_wr_start_finish in the following
2282	 * cases:
2283	 *
2284	 * . if bfqq is not going to be weight-raised, because, for
2285	 *   non weight-raised queues, last_wr_start_finish stores the
2286	 *   arrival time of the last request; as of now, this piece
2287	 *   of information is used only for deciding whether to
2288	 *   weight-raise async queues
2289	 *
2290	 * . if bfqq is not weight-raised, because, if bfqq is now
2291	 *   switching to weight-raised, then last_wr_start_finish
2292	 *   stores the time when weight-raising starts
2293	 *
2294	 * . if bfqq is interactive, because, regardless of whether
2295	 *   bfqq is currently weight-raised, the weight-raising
2296	 *   period must start or restart (this case is considered
2297	 *   separately because it is not detected by the above
2298	 *   conditions, if bfqq is already weight-raised)
2299	 *
2300	 * last_wr_start_finish has to be updated also if bfqq is soft
2301	 * real-time, because the weight-raising period is constantly
2302	 * restarted on idle-to-busy transitions for these queues, but
2303	 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2304	 * needed.
2305	 */
2306	if (bfqd->low_latency &&
2307		(old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
2308		bfqq->last_wr_start_finish = jiffies;
2309}
2310
2311static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
2312					  struct bio *bio,
2313					  struct request_queue *q)
2314{
2315	struct bfq_queue *bfqq = bfqd->bio_bfqq;
2316
2317
2318	if (bfqq)
2319		return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
2320
2321	return NULL;
2322}
2323
2324static sector_t get_sdist(sector_t last_pos, struct request *rq)
2325{
2326	if (last_pos)
2327		return abs(blk_rq_pos(rq) - last_pos);
2328
2329	return 0;
2330}
2331
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2332static void bfq_remove_request(struct request_queue *q,
2333			       struct request *rq)
2334{
2335	struct bfq_queue *bfqq = RQ_BFQQ(rq);
2336	struct bfq_data *bfqd = bfqq->bfqd;
2337	const int sync = rq_is_sync(rq);
2338
2339	if (bfqq->next_rq == rq) {
2340		bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
2341		bfq_updated_next_req(bfqd, bfqq);
2342	}
2343
2344	if (rq->queuelist.prev != &rq->queuelist)
2345		list_del_init(&rq->queuelist);
2346	bfqq->queued[sync]--;
2347	/*
2348	 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2349	 * may be read without holding the lock in bfq_has_work().
2350	 */
2351	WRITE_ONCE(bfqd->queued, bfqd->queued - 1);
2352	elv_rb_del(&bfqq->sort_list, rq);
2353
2354	elv_rqhash_del(q, rq);
2355	if (q->last_merge == rq)
2356		q->last_merge = NULL;
2357
2358	if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2359		bfqq->next_rq = NULL;
2360
2361		if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
2362			bfq_del_bfqq_busy(bfqq, false);
2363			/*
2364			 * bfqq emptied. In normal operation, when
2365			 * bfqq is empty, bfqq->entity.service and
2366			 * bfqq->entity.budget must contain,
2367			 * respectively, the service received and the
2368			 * budget used last time bfqq emptied. These
2369			 * facts do not hold in this case, as at least
2370			 * this last removal occurred while bfqq is
2371			 * not in service. To avoid inconsistencies,
2372			 * reset both bfqq->entity.service and
2373			 * bfqq->entity.budget, if bfqq has still a
2374			 * process that may issue I/O requests to it.
2375			 */
2376			bfqq->entity.budget = bfqq->entity.service = 0;
2377		}
2378
2379		/*
2380		 * Remove queue from request-position tree as it is empty.
2381		 */
2382		if (bfqq->pos_root) {
2383			rb_erase(&bfqq->pos_node, bfqq->pos_root);
2384			bfqq->pos_root = NULL;
2385		}
2386	} else {
2387		/* see comments on bfq_pos_tree_add_move() for the unlikely() */
2388		if (unlikely(!bfqd->nonrot_with_queueing))
2389			bfq_pos_tree_add_move(bfqd, bfqq);
2390	}
2391
2392	if (rq->cmd_flags & REQ_META)
2393		bfqq->meta_pending--;
2394
2395}
2396
2397static bool bfq_bio_merge(struct request_queue *q, struct bio *bio,
2398		unsigned int nr_segs)
2399{
 
2400	struct bfq_data *bfqd = q->elevator->elevator_data;
2401	struct request *free = NULL;
2402	/*
2403	 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2404	 * store its return value for later use, to avoid nesting
2405	 * queue_lock inside the bfqd->lock. We assume that the bic
2406	 * returned by bfq_bic_lookup does not go away before
2407	 * bfqd->lock is taken.
2408	 */
2409	struct bfq_io_cq *bic = bfq_bic_lookup(q);
2410	bool ret;
2411
2412	spin_lock_irq(&bfqd->lock);
2413
2414	if (bic) {
2415		/*
2416		 * Make sure cgroup info is uptodate for current process before
2417		 * considering the merge.
2418		 */
2419		bfq_bic_update_cgroup(bic, bio);
2420
2421		bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
2422	} else {
2423		bfqd->bio_bfqq = NULL;
2424	}
2425	bfqd->bio_bic = bic;
2426
2427	ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
2428
2429	spin_unlock_irq(&bfqd->lock);
2430	if (free)
2431		blk_mq_free_request(free);
 
2432
2433	return ret;
2434}
2435
2436static int bfq_request_merge(struct request_queue *q, struct request **req,
2437			     struct bio *bio)
2438{
2439	struct bfq_data *bfqd = q->elevator->elevator_data;
2440	struct request *__rq;
2441
2442	__rq = bfq_find_rq_fmerge(bfqd, bio, q);
2443	if (__rq && elv_bio_merge_ok(__rq, bio)) {
2444		*req = __rq;
2445
2446		if (blk_discard_mergable(__rq))
2447			return ELEVATOR_DISCARD_MERGE;
2448		return ELEVATOR_FRONT_MERGE;
2449	}
2450
2451	return ELEVATOR_NO_MERGE;
2452}
2453
2454static void bfq_request_merged(struct request_queue *q, struct request *req,
2455			       enum elv_merge type)
2456{
2457	if (type == ELEVATOR_FRONT_MERGE &&
2458	    rb_prev(&req->rb_node) &&
2459	    blk_rq_pos(req) <
2460	    blk_rq_pos(container_of(rb_prev(&req->rb_node),
2461				    struct request, rb_node))) {
2462		struct bfq_queue *bfqq = RQ_BFQQ(req);
2463		struct bfq_data *bfqd;
2464		struct request *prev, *next_rq;
2465
2466		if (!bfqq)
2467			return;
2468
2469		bfqd = bfqq->bfqd;
2470
2471		/* Reposition request in its sort_list */
2472		elv_rb_del(&bfqq->sort_list, req);
2473		elv_rb_add(&bfqq->sort_list, req);
2474
2475		/* Choose next request to be served for bfqq */
2476		prev = bfqq->next_rq;
2477		next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
2478					 bfqd->last_position);
2479		bfqq->next_rq = next_rq;
2480		/*
2481		 * If next_rq changes, update both the queue's budget to
2482		 * fit the new request and the queue's position in its
2483		 * rq_pos_tree.
2484		 */
2485		if (prev != bfqq->next_rq) {
2486			bfq_updated_next_req(bfqd, bfqq);
2487			/*
2488			 * See comments on bfq_pos_tree_add_move() for
2489			 * the unlikely().
2490			 */
2491			if (unlikely(!bfqd->nonrot_with_queueing))
2492				bfq_pos_tree_add_move(bfqd, bfqq);
2493		}
2494	}
2495}
2496
2497/*
2498 * This function is called to notify the scheduler that the requests
2499 * rq and 'next' have been merged, with 'next' going away.  BFQ
2500 * exploits this hook to address the following issue: if 'next' has a
2501 * fifo_time lower that rq, then the fifo_time of rq must be set to
2502 * the value of 'next', to not forget the greater age of 'next'.
2503 *
2504 * NOTE: in this function we assume that rq is in a bfq_queue, basing
2505 * on that rq is picked from the hash table q->elevator->hash, which,
2506 * in its turn, is filled only with I/O requests present in
2507 * bfq_queues, while BFQ is in use for the request queue q. In fact,
2508 * the function that fills this hash table (elv_rqhash_add) is called
2509 * only by bfq_insert_request.
2510 */
2511static void bfq_requests_merged(struct request_queue *q, struct request *rq,
2512				struct request *next)
2513{
2514	struct bfq_queue *bfqq = RQ_BFQQ(rq),
2515		*next_bfqq = RQ_BFQQ(next);
2516
2517	if (!bfqq)
2518		goto remove;
 
2519
2520	/*
2521	 * If next and rq belong to the same bfq_queue and next is older
2522	 * than rq, then reposition rq in the fifo (by substituting next
2523	 * with rq). Otherwise, if next and rq belong to different
2524	 * bfq_queues, never reposition rq: in fact, we would have to
2525	 * reposition it with respect to next's position in its own fifo,
2526	 * which would most certainly be too expensive with respect to
2527	 * the benefits.
2528	 */
2529	if (bfqq == next_bfqq &&
2530	    !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2531	    next->fifo_time < rq->fifo_time) {
2532		list_del_init(&rq->queuelist);
2533		list_replace_init(&next->queuelist, &rq->queuelist);
2534		rq->fifo_time = next->fifo_time;
2535	}
2536
2537	if (bfqq->next_rq == next)
2538		bfqq->next_rq = rq;
2539
 
 
 
 
 
2540	bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
2541remove:
2542	/* Merged request may be in the IO scheduler. Remove it. */
2543	if (!RB_EMPTY_NODE(&next->rb_node)) {
2544		bfq_remove_request(next->q, next);
2545		if (next_bfqq)
2546			bfqg_stats_update_io_remove(bfqq_group(next_bfqq),
2547						    next->cmd_flags);
2548	}
2549}
2550
2551/* Must be called with bfqq != NULL */
2552static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
2553{
2554	/*
2555	 * If bfqq has been enjoying interactive weight-raising, then
2556	 * reset soft_rt_next_start. We do it for the following
2557	 * reason. bfqq may have been conveying the I/O needed to load
2558	 * a soft real-time application. Such an application actually
2559	 * exhibits a soft real-time I/O pattern after it finishes
2560	 * loading, and finally starts doing its job. But, if bfqq has
2561	 * been receiving a lot of bandwidth so far (likely to happen
2562	 * on a fast device), then soft_rt_next_start now contains a
2563	 * high value that. So, without this reset, bfqq would be
2564	 * prevented from being possibly considered as soft_rt for a
2565	 * very long time.
2566	 */
2567
2568	if (bfqq->wr_cur_max_time !=
2569	    bfqq->bfqd->bfq_wr_rt_max_time)
2570		bfqq->soft_rt_next_start = jiffies;
2571
2572	if (bfq_bfqq_busy(bfqq))
2573		bfqq->bfqd->wr_busy_queues--;
2574	bfqq->wr_coeff = 1;
2575	bfqq->wr_cur_max_time = 0;
2576	bfqq->last_wr_start_finish = jiffies;
2577	/*
2578	 * Trigger a weight change on the next invocation of
2579	 * __bfq_entity_update_weight_prio.
2580	 */
2581	bfqq->entity.prio_changed = 1;
2582}
2583
2584void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2585			     struct bfq_group *bfqg)
2586{
2587	int i, j;
2588
2589	for (i = 0; i < 2; i++)
2590		for (j = 0; j < IOPRIO_NR_LEVELS; j++)
2591			if (bfqg->async_bfqq[i][j])
2592				bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
2593	if (bfqg->async_idle_bfqq)
2594		bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
2595}
2596
2597static void bfq_end_wr(struct bfq_data *bfqd)
2598{
2599	struct bfq_queue *bfqq;
2600
2601	spin_lock_irq(&bfqd->lock);
2602
2603	list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
2604		bfq_bfqq_end_wr(bfqq);
2605	list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2606		bfq_bfqq_end_wr(bfqq);
2607	bfq_end_wr_async(bfqd);
2608
2609	spin_unlock_irq(&bfqd->lock);
2610}
2611
2612static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2613{
2614	if (request)
2615		return blk_rq_pos(io_struct);
2616	else
2617		return ((struct bio *)io_struct)->bi_iter.bi_sector;
2618}
2619
2620static int bfq_rq_close_to_sector(void *io_struct, bool request,
2621				  sector_t sector)
2622{
2623	return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2624	       BFQQ_CLOSE_THR;
2625}
2626
2627static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2628					 struct bfq_queue *bfqq,
2629					 sector_t sector)
2630{
2631	struct rb_root *root = &bfqq_group(bfqq)->rq_pos_tree;
2632	struct rb_node *parent, *node;
2633	struct bfq_queue *__bfqq;
2634
2635	if (RB_EMPTY_ROOT(root))
2636		return NULL;
2637
2638	/*
2639	 * First, if we find a request starting at the end of the last
2640	 * request, choose it.
2641	 */
2642	__bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2643	if (__bfqq)
2644		return __bfqq;
2645
2646	/*
2647	 * If the exact sector wasn't found, the parent of the NULL leaf
2648	 * will contain the closest sector (rq_pos_tree sorted by
2649	 * next_request position).
2650	 */
2651	__bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2652	if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2653		return __bfqq;
2654
2655	if (blk_rq_pos(__bfqq->next_rq) < sector)
2656		node = rb_next(&__bfqq->pos_node);
2657	else
2658		node = rb_prev(&__bfqq->pos_node);
2659	if (!node)
2660		return NULL;
2661
2662	__bfqq = rb_entry(node, struct bfq_queue, pos_node);
2663	if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2664		return __bfqq;
2665
2666	return NULL;
2667}
2668
2669static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2670						   struct bfq_queue *cur_bfqq,
2671						   sector_t sector)
2672{
2673	struct bfq_queue *bfqq;
2674
2675	/*
2676	 * We shall notice if some of the queues are cooperating,
2677	 * e.g., working closely on the same area of the device. In
2678	 * that case, we can group them together and: 1) don't waste
2679	 * time idling, and 2) serve the union of their requests in
2680	 * the best possible order for throughput.
2681	 */
2682	bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2683	if (!bfqq || bfqq == cur_bfqq)
2684		return NULL;
2685
2686	return bfqq;
2687}
2688
2689static struct bfq_queue *
2690bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2691{
2692	int process_refs, new_process_refs;
2693	struct bfq_queue *__bfqq;
2694
2695	/*
2696	 * If there are no process references on the new_bfqq, then it is
2697	 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2698	 * may have dropped their last reference (not just their last process
2699	 * reference).
2700	 */
2701	if (!bfqq_process_refs(new_bfqq))
2702		return NULL;
2703
2704	/* Avoid a circular list and skip interim queue merges. */
2705	while ((__bfqq = new_bfqq->new_bfqq)) {
2706		if (__bfqq == bfqq)
2707			return NULL;
2708		new_bfqq = __bfqq;
2709	}
2710
2711	process_refs = bfqq_process_refs(bfqq);
2712	new_process_refs = bfqq_process_refs(new_bfqq);
2713	/*
2714	 * If the process for the bfqq has gone away, there is no
2715	 * sense in merging the queues.
2716	 */
2717	if (process_refs == 0 || new_process_refs == 0)
2718		return NULL;
2719
2720	/*
2721	 * Make sure merged queues belong to the same parent. Parents could
2722	 * have changed since the time we decided the two queues are suitable
2723	 * for merging.
2724	 */
2725	if (new_bfqq->entity.parent != bfqq->entity.parent)
2726		return NULL;
2727
2728	bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2729		new_bfqq->pid);
2730
2731	/*
2732	 * Merging is just a redirection: the requests of the process
2733	 * owning one of the two queues are redirected to the other queue.
2734	 * The latter queue, in its turn, is set as shared if this is the
2735	 * first time that the requests of some process are redirected to
2736	 * it.
2737	 *
2738	 * We redirect bfqq to new_bfqq and not the opposite, because
2739	 * we are in the context of the process owning bfqq, thus we
2740	 * have the io_cq of this process. So we can immediately
2741	 * configure this io_cq to redirect the requests of the
2742	 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2743	 * not available any more (new_bfqq->bic == NULL).
2744	 *
2745	 * Anyway, even in case new_bfqq coincides with the in-service
2746	 * queue, redirecting requests the in-service queue is the
2747	 * best option, as we feed the in-service queue with new
2748	 * requests close to the last request served and, by doing so,
2749	 * are likely to increase the throughput.
2750	 */
2751	bfqq->new_bfqq = new_bfqq;
2752	/*
2753	 * The above assignment schedules the following redirections:
2754	 * each time some I/O for bfqq arrives, the process that
2755	 * generated that I/O is disassociated from bfqq and
2756	 * associated with new_bfqq. Here we increases new_bfqq->ref
2757	 * in advance, adding the number of processes that are
2758	 * expected to be associated with new_bfqq as they happen to
2759	 * issue I/O.
2760	 */
2761	new_bfqq->ref += process_refs;
2762	return new_bfqq;
2763}
2764
2765static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2766					struct bfq_queue *new_bfqq)
2767{
2768	if (bfq_too_late_for_merging(new_bfqq))
2769		return false;
2770
2771	if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2772	    (bfqq->ioprio_class != new_bfqq->ioprio_class))
2773		return false;
2774
2775	/*
2776	 * If either of the queues has already been detected as seeky,
2777	 * then merging it with the other queue is unlikely to lead to
2778	 * sequential I/O.
2779	 */
2780	if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2781		return false;
2782
2783	/*
2784	 * Interleaved I/O is known to be done by (some) applications
2785	 * only for reads, so it does not make sense to merge async
2786	 * queues.
2787	 */
2788	if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2789		return false;
2790
2791	return true;
2792}
2793
2794static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
2795					     struct bfq_queue *bfqq);
2796
2797/*
2798 * Attempt to schedule a merge of bfqq with the currently in-service
2799 * queue or with a close queue among the scheduled queues.  Return
2800 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2801 * structure otherwise.
2802 *
2803 * The OOM queue is not allowed to participate to cooperation: in fact, since
2804 * the requests temporarily redirected to the OOM queue could be redirected
2805 * again to dedicated queues at any time, the state needed to correctly
2806 * handle merging with the OOM queue would be quite complex and expensive
2807 * to maintain. Besides, in such a critical condition as an out of memory,
2808 * the benefits of queue merging may be little relevant, or even negligible.
2809 *
2810 * WARNING: queue merging may impair fairness among non-weight raised
2811 * queues, for at least two reasons: 1) the original weight of a
2812 * merged queue may change during the merged state, 2) even being the
2813 * weight the same, a merged queue may be bloated with many more
2814 * requests than the ones produced by its originally-associated
2815 * process.
2816 */
2817static struct bfq_queue *
2818bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2819		     void *io_struct, bool request, struct bfq_io_cq *bic)
2820{
2821	struct bfq_queue *in_service_bfqq, *new_bfqq;
2822
2823	/* if a merge has already been setup, then proceed with that first */
2824	if (bfqq->new_bfqq)
2825		return bfqq->new_bfqq;
2826
2827	/*
2828	 * Check delayed stable merge for rotational or non-queueing
2829	 * devs. For this branch to be executed, bfqq must not be
2830	 * currently merged with some other queue (i.e., bfqq->bic
2831	 * must be non null). If we considered also merged queues,
2832	 * then we should also check whether bfqq has already been
2833	 * merged with bic->stable_merge_bfqq. But this would be
2834	 * costly and complicated.
2835	 */
2836	if (unlikely(!bfqd->nonrot_with_queueing)) {
2837		/*
2838		 * Make sure also that bfqq is sync, because
2839		 * bic->stable_merge_bfqq may point to some queue (for
2840		 * stable merging) also if bic is associated with a
2841		 * sync queue, but this bfqq is async
2842		 */
2843		if (bfq_bfqq_sync(bfqq) && bic->stable_merge_bfqq &&
2844		    !bfq_bfqq_just_created(bfqq) &&
2845		    time_is_before_jiffies(bfqq->split_time +
2846					  msecs_to_jiffies(bfq_late_stable_merging)) &&
2847		    time_is_before_jiffies(bfqq->creation_time +
2848					   msecs_to_jiffies(bfq_late_stable_merging))) {
2849			struct bfq_queue *stable_merge_bfqq =
2850				bic->stable_merge_bfqq;
2851			int proc_ref = min(bfqq_process_refs(bfqq),
2852					   bfqq_process_refs(stable_merge_bfqq));
2853
2854			/* deschedule stable merge, because done or aborted here */
2855			bfq_put_stable_ref(stable_merge_bfqq);
2856
2857			bic->stable_merge_bfqq = NULL;
2858
2859			if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
2860			    proc_ref > 0) {
2861				/* next function will take at least one ref */
2862				struct bfq_queue *new_bfqq =
2863					bfq_setup_merge(bfqq, stable_merge_bfqq);
2864
2865				if (new_bfqq) {
2866					bic->stably_merged = true;
2867					if (new_bfqq->bic)
2868						new_bfqq->bic->stably_merged =
2869									true;
2870				}
2871				return new_bfqq;
2872			} else
2873				return NULL;
2874		}
2875	}
2876
2877	/*
2878	 * Do not perform queue merging if the device is non
2879	 * rotational and performs internal queueing. In fact, such a
2880	 * device reaches a high speed through internal parallelism
2881	 * and pipelining. This means that, to reach a high
2882	 * throughput, it must have many requests enqueued at the same
2883	 * time. But, in this configuration, the internal scheduling
2884	 * algorithm of the device does exactly the job of queue
2885	 * merging: it reorders requests so as to obtain as much as
2886	 * possible a sequential I/O pattern. As a consequence, with
2887	 * the workload generated by processes doing interleaved I/O,
2888	 * the throughput reached by the device is likely to be the
2889	 * same, with and without queue merging.
2890	 *
2891	 * Disabling merging also provides a remarkable benefit in
2892	 * terms of throughput. Merging tends to make many workloads
2893	 * artificially more uneven, because of shared queues
2894	 * remaining non empty for incomparably more time than
2895	 * non-merged queues. This may accentuate workload
2896	 * asymmetries. For example, if one of the queues in a set of
2897	 * merged queues has a higher weight than a normal queue, then
2898	 * the shared queue may inherit such a high weight and, by
2899	 * staying almost always active, may force BFQ to perform I/O
2900	 * plugging most of the time. This evidently makes it harder
2901	 * for BFQ to let the device reach a high throughput.
2902	 *
2903	 * Finally, the likely() macro below is not used because one
2904	 * of the two branches is more likely than the other, but to
2905	 * have the code path after the following if() executed as
2906	 * fast as possible for the case of a non rotational device
2907	 * with queueing. We want it because this is the fastest kind
2908	 * of device. On the opposite end, the likely() may lengthen
2909	 * the execution time of BFQ for the case of slower devices
2910	 * (rotational or at least without queueing). But in this case
2911	 * the execution time of BFQ matters very little, if not at
2912	 * all.
2913	 */
2914	if (likely(bfqd->nonrot_with_queueing))
2915		return NULL;
2916
2917	/*
2918	 * Prevent bfqq from being merged if it has been created too
2919	 * long ago. The idea is that true cooperating processes, and
2920	 * thus their associated bfq_queues, are supposed to be
2921	 * created shortly after each other. This is the case, e.g.,
2922	 * for KVM/QEMU and dump I/O threads. Basing on this
2923	 * assumption, the following filtering greatly reduces the
2924	 * probability that two non-cooperating processes, which just
2925	 * happen to do close I/O for some short time interval, have
2926	 * their queues merged by mistake.
2927	 */
2928	if (bfq_too_late_for_merging(bfqq))
2929		return NULL;
2930
 
 
 
2931	if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2932		return NULL;
2933
2934	/* If there is only one backlogged queue, don't search. */
2935	if (bfq_tot_busy_queues(bfqd) == 1)
2936		return NULL;
2937
2938	in_service_bfqq = bfqd->in_service_queue;
2939
2940	if (in_service_bfqq && in_service_bfqq != bfqq &&
2941	    likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2942	    bfq_rq_close_to_sector(io_struct, request,
2943				   bfqd->in_serv_last_pos) &&
2944	    bfqq->entity.parent == in_service_bfqq->entity.parent &&
2945	    bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2946		new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2947		if (new_bfqq)
2948			return new_bfqq;
2949	}
2950	/*
2951	 * Check whether there is a cooperator among currently scheduled
2952	 * queues. The only thing we need is that the bio/request is not
2953	 * NULL, as we need it to establish whether a cooperator exists.
2954	 */
2955	new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2956			bfq_io_struct_pos(io_struct, request));
2957
2958	if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2959	    bfq_may_be_close_cooperator(bfqq, new_bfqq))
2960		return bfq_setup_merge(bfqq, new_bfqq);
2961
2962	return NULL;
2963}
2964
2965static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2966{
2967	struct bfq_io_cq *bic = bfqq->bic;
2968
2969	/*
2970	 * If !bfqq->bic, the queue is already shared or its requests
2971	 * have already been redirected to a shared queue; both idle window
2972	 * and weight raising state have already been saved. Do nothing.
2973	 */
2974	if (!bic)
2975		return;
2976
2977	bic->saved_last_serv_time_ns = bfqq->last_serv_time_ns;
2978	bic->saved_inject_limit = bfqq->inject_limit;
2979	bic->saved_decrease_time_jif = bfqq->decrease_time_jif;
2980
2981	bic->saved_weight = bfqq->entity.orig_weight;
2982	bic->saved_ttime = bfqq->ttime;
2983	bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2984	bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
2985	bic->saved_io_start_time = bfqq->io_start_time;
2986	bic->saved_tot_idle_time = bfqq->tot_idle_time;
2987	bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2988	bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2989	if (unlikely(bfq_bfqq_just_created(bfqq) &&
2990		     !bfq_bfqq_in_large_burst(bfqq) &&
2991		     bfqq->bfqd->low_latency)) {
2992		/*
2993		 * bfqq being merged right after being created: bfqq
2994		 * would have deserved interactive weight raising, but
2995		 * did not make it to be set in a weight-raised state,
2996		 * because of this early merge.	Store directly the
2997		 * weight-raising state that would have been assigned
2998		 * to bfqq, so that to avoid that bfqq unjustly fails
2999		 * to enjoy weight raising if split soon.
3000		 */
3001		bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
3002		bic->saved_wr_start_at_switch_to_srt = bfq_smallest_from_now();
3003		bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
3004		bic->saved_last_wr_start_finish = jiffies;
3005	} else {
3006		bic->saved_wr_coeff = bfqq->wr_coeff;
3007		bic->saved_wr_start_at_switch_to_srt =
3008			bfqq->wr_start_at_switch_to_srt;
3009		bic->saved_service_from_wr = bfqq->service_from_wr;
3010		bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
3011		bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
3012	}
3013}
3014
3015
3016static void
3017bfq_reassign_last_bfqq(struct bfq_queue *cur_bfqq, struct bfq_queue *new_bfqq)
3018{
3019	if (cur_bfqq->entity.parent &&
3020	    cur_bfqq->entity.parent->last_bfqq_created == cur_bfqq)
3021		cur_bfqq->entity.parent->last_bfqq_created = new_bfqq;
3022	else if (cur_bfqq->bfqd && cur_bfqq->bfqd->last_bfqq_created == cur_bfqq)
3023		cur_bfqq->bfqd->last_bfqq_created = new_bfqq;
3024}
3025
3026void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3027{
3028	/*
3029	 * To prevent bfqq's service guarantees from being violated,
3030	 * bfqq may be left busy, i.e., queued for service, even if
3031	 * empty (see comments in __bfq_bfqq_expire() for
3032	 * details). But, if no process will send requests to bfqq any
3033	 * longer, then there is no point in keeping bfqq queued for
3034	 * service. In addition, keeping bfqq queued for service, but
3035	 * with no process ref any longer, may have caused bfqq to be
3036	 * freed when dequeued from service. But this is assumed to
3037	 * never happen.
3038	 */
3039	if (bfq_bfqq_busy(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) &&
3040	    bfqq != bfqd->in_service_queue)
3041		bfq_del_bfqq_busy(bfqq, false);
3042
3043	bfq_reassign_last_bfqq(bfqq, NULL);
3044
3045	bfq_put_queue(bfqq);
3046}
3047
3048static void
3049bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
3050		struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
3051{
3052	bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
3053		(unsigned long)new_bfqq->pid);
3054	/* Save weight raising and idle window of the merged queues */
3055	bfq_bfqq_save_state(bfqq);
3056	bfq_bfqq_save_state(new_bfqq);
3057	if (bfq_bfqq_IO_bound(bfqq))
3058		bfq_mark_bfqq_IO_bound(new_bfqq);
3059	bfq_clear_bfqq_IO_bound(bfqq);
3060
3061	/*
3062	 * The processes associated with bfqq are cooperators of the
3063	 * processes associated with new_bfqq. So, if bfqq has a
3064	 * waker, then assume that all these processes will be happy
3065	 * to let bfqq's waker freely inject I/O when they have no
3066	 * I/O.
3067	 */
3068	if (bfqq->waker_bfqq && !new_bfqq->waker_bfqq &&
3069	    bfqq->waker_bfqq != new_bfqq) {
3070		new_bfqq->waker_bfqq = bfqq->waker_bfqq;
3071		new_bfqq->tentative_waker_bfqq = NULL;
3072
3073		/*
3074		 * If the waker queue disappears, then
3075		 * new_bfqq->waker_bfqq must be reset. So insert
3076		 * new_bfqq into the woken_list of the waker. See
3077		 * bfq_check_waker for details.
3078		 */
3079		hlist_add_head(&new_bfqq->woken_list_node,
3080			       &new_bfqq->waker_bfqq->woken_list);
3081
3082	}
3083
3084	/*
3085	 * If bfqq is weight-raised, then let new_bfqq inherit
3086	 * weight-raising. To reduce false positives, neglect the case
3087	 * where bfqq has just been created, but has not yet made it
3088	 * to be weight-raised (which may happen because EQM may merge
3089	 * bfqq even before bfq_add_request is executed for the first
3090	 * time for bfqq). Handling this case would however be very
3091	 * easy, thanks to the flag just_created.
3092	 */
3093	if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
3094		new_bfqq->wr_coeff = bfqq->wr_coeff;
3095		new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
3096		new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
3097		new_bfqq->wr_start_at_switch_to_srt =
3098			bfqq->wr_start_at_switch_to_srt;
3099		if (bfq_bfqq_busy(new_bfqq))
3100			bfqd->wr_busy_queues++;
3101		new_bfqq->entity.prio_changed = 1;
3102	}
3103
3104	if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
3105		bfqq->wr_coeff = 1;
3106		bfqq->entity.prio_changed = 1;
3107		if (bfq_bfqq_busy(bfqq))
3108			bfqd->wr_busy_queues--;
3109	}
3110
3111	bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
3112		     bfqd->wr_busy_queues);
3113
3114	/*
3115	 * Merge queues (that is, let bic redirect its requests to new_bfqq)
3116	 */
3117	bic_set_bfqq(bic, new_bfqq, true);
3118	bfq_mark_bfqq_coop(new_bfqq);
3119	/*
3120	 * new_bfqq now belongs to at least two bics (it is a shared queue):
3121	 * set new_bfqq->bic to NULL. bfqq either:
3122	 * - does not belong to any bic any more, and hence bfqq->bic must
3123	 *   be set to NULL, or
3124	 * - is a queue whose owning bics have already been redirected to a
3125	 *   different queue, hence the queue is destined to not belong to
3126	 *   any bic soon and bfqq->bic is already NULL (therefore the next
3127	 *   assignment causes no harm).
3128	 */
3129	new_bfqq->bic = NULL;
3130	/*
3131	 * If the queue is shared, the pid is the pid of one of the associated
3132	 * processes. Which pid depends on the exact sequence of merge events
3133	 * the queue underwent. So printing such a pid is useless and confusing
3134	 * because it reports a random pid between those of the associated
3135	 * processes.
3136	 * We mark such a queue with a pid -1, and then print SHARED instead of
3137	 * a pid in logging messages.
3138	 */
3139	new_bfqq->pid = -1;
3140	bfqq->bic = NULL;
3141
3142	bfq_reassign_last_bfqq(bfqq, new_bfqq);
3143
3144	bfq_release_process_ref(bfqd, bfqq);
3145}
3146
3147static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
3148				struct bio *bio)
3149{
3150	struct bfq_data *bfqd = q->elevator->elevator_data;
3151	bool is_sync = op_is_sync(bio->bi_opf);
3152	struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
3153
3154	/*
3155	 * Disallow merge of a sync bio into an async request.
3156	 */
3157	if (is_sync && !rq_is_sync(rq))
3158		return false;
3159
3160	/*
3161	 * Lookup the bfqq that this bio will be queued with. Allow
3162	 * merge only if rq is queued there.
3163	 */
3164	if (!bfqq)
3165		return false;
3166
3167	/*
3168	 * We take advantage of this function to perform an early merge
3169	 * of the queues of possible cooperating processes.
3170	 */
3171	new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false, bfqd->bio_bic);
3172	if (new_bfqq) {
3173		/*
3174		 * bic still points to bfqq, then it has not yet been
3175		 * redirected to some other bfq_queue, and a queue
3176		 * merge between bfqq and new_bfqq can be safely
3177		 * fulfilled, i.e., bic can be redirected to new_bfqq
3178		 * and bfqq can be put.
3179		 */
3180		bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
3181				new_bfqq);
3182		/*
3183		 * If we get here, bio will be queued into new_queue,
3184		 * so use new_bfqq to decide whether bio and rq can be
3185		 * merged.
3186		 */
3187		bfqq = new_bfqq;
3188
3189		/*
3190		 * Change also bqfd->bio_bfqq, as
3191		 * bfqd->bio_bic now points to new_bfqq, and
3192		 * this function may be invoked again (and then may
3193		 * use again bqfd->bio_bfqq).
3194		 */
3195		bfqd->bio_bfqq = bfqq;
3196	}
3197
3198	return bfqq == RQ_BFQQ(rq);
3199}
3200
3201/*
3202 * Set the maximum time for the in-service queue to consume its
3203 * budget. This prevents seeky processes from lowering the throughput.
3204 * In practice, a time-slice service scheme is used with seeky
3205 * processes.
3206 */
3207static void bfq_set_budget_timeout(struct bfq_data *bfqd,
3208				   struct bfq_queue *bfqq)
3209{
3210	unsigned int timeout_coeff;
3211
3212	if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
3213		timeout_coeff = 1;
3214	else
3215		timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
3216
3217	bfqd->last_budget_start = ktime_get();
3218
3219	bfqq->budget_timeout = jiffies +
3220		bfqd->bfq_timeout * timeout_coeff;
3221}
3222
3223static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
3224				       struct bfq_queue *bfqq)
3225{
3226	if (bfqq) {
3227		bfq_clear_bfqq_fifo_expire(bfqq);
3228
3229		bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
3230
3231		if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
3232		    bfqq->wr_coeff > 1 &&
3233		    bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
3234		    time_is_before_jiffies(bfqq->budget_timeout)) {
3235			/*
3236			 * For soft real-time queues, move the start
3237			 * of the weight-raising period forward by the
3238			 * time the queue has not received any
3239			 * service. Otherwise, a relatively long
3240			 * service delay is likely to cause the
3241			 * weight-raising period of the queue to end,
3242			 * because of the short duration of the
3243			 * weight-raising period of a soft real-time
3244			 * queue.  It is worth noting that this move
3245			 * is not so dangerous for the other queues,
3246			 * because soft real-time queues are not
3247			 * greedy.
3248			 *
3249			 * To not add a further variable, we use the
3250			 * overloaded field budget_timeout to
3251			 * determine for how long the queue has not
3252			 * received service, i.e., how much time has
3253			 * elapsed since the queue expired. However,
3254			 * this is a little imprecise, because
3255			 * budget_timeout is set to jiffies if bfqq
3256			 * not only expires, but also remains with no
3257			 * request.
3258			 */
3259			if (time_after(bfqq->budget_timeout,
3260				       bfqq->last_wr_start_finish))
3261				bfqq->last_wr_start_finish +=
3262					jiffies - bfqq->budget_timeout;
3263			else
3264				bfqq->last_wr_start_finish = jiffies;
3265		}
3266
3267		bfq_set_budget_timeout(bfqd, bfqq);
3268		bfq_log_bfqq(bfqd, bfqq,
3269			     "set_in_service_queue, cur-budget = %d",
3270			     bfqq->entity.budget);
3271	}
3272
3273	bfqd->in_service_queue = bfqq;
3274	bfqd->in_serv_last_pos = 0;
3275}
3276
3277/*
3278 * Get and set a new queue for service.
3279 */
3280static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
3281{
3282	struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
3283
3284	__bfq_set_in_service_queue(bfqd, bfqq);
3285	return bfqq;
3286}
3287
3288static void bfq_arm_slice_timer(struct bfq_data *bfqd)
3289{
3290	struct bfq_queue *bfqq = bfqd->in_service_queue;
3291	u32 sl;
3292
3293	bfq_mark_bfqq_wait_request(bfqq);
3294
3295	/*
3296	 * We don't want to idle for seeks, but we do want to allow
3297	 * fair distribution of slice time for a process doing back-to-back
3298	 * seeks. So allow a little bit of time for him to submit a new rq.
3299	 */
3300	sl = bfqd->bfq_slice_idle;
3301	/*
3302	 * Unless the queue is being weight-raised or the scenario is
3303	 * asymmetric, grant only minimum idle time if the queue
3304	 * is seeky. A long idling is preserved for a weight-raised
3305	 * queue, or, more in general, in an asymmetric scenario,
3306	 * because a long idling is needed for guaranteeing to a queue
3307	 * its reserved share of the throughput (in particular, it is
3308	 * needed if the queue has a higher weight than some other
3309	 * queue).
3310	 */
3311	if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3312	    !bfq_asymmetric_scenario(bfqd, bfqq))
3313		sl = min_t(u64, sl, BFQ_MIN_TT);
3314	else if (bfqq->wr_coeff > 1)
3315		sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
3316
3317	bfqd->last_idling_start = ktime_get();
3318	bfqd->last_idling_start_jiffies = jiffies;
3319
3320	hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
3321		      HRTIMER_MODE_REL);
3322	bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
3323}
3324
3325/*
3326 * In autotuning mode, max_budget is dynamically recomputed as the
3327 * amount of sectors transferred in timeout at the estimated peak
3328 * rate. This enables BFQ to utilize a full timeslice with a full
3329 * budget, even if the in-service queue is served at peak rate. And
3330 * this maximises throughput with sequential workloads.
3331 */
3332static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
3333{
3334	return (u64)bfqd->peak_rate * USEC_PER_MSEC *
3335		jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
3336}
3337
3338/*
3339 * Update parameters related to throughput and responsiveness, as a
3340 * function of the estimated peak rate. See comments on
3341 * bfq_calc_max_budget(), and on the ref_wr_duration array.
3342 */
3343static void update_thr_responsiveness_params(struct bfq_data *bfqd)
3344{
3345	if (bfqd->bfq_user_max_budget == 0) {
 
 
3346		bfqd->bfq_max_budget =
3347			bfq_calc_max_budget(bfqd);
3348		bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
 
 
 
 
 
 
 
 
 
 
3349	}
 
 
 
 
 
 
 
 
 
 
3350}
3351
3352static void bfq_reset_rate_computation(struct bfq_data *bfqd,
3353				       struct request *rq)
3354{
3355	if (rq != NULL) { /* new rq dispatch now, reset accordingly */
3356		bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
3357		bfqd->peak_rate_samples = 1;
3358		bfqd->sequential_samples = 0;
3359		bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
3360			blk_rq_sectors(rq);
3361	} else /* no new rq dispatched, just reset the number of samples */
3362		bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
3363
3364	bfq_log(bfqd,
3365		"reset_rate_computation at end, sample %u/%u tot_sects %llu",
3366		bfqd->peak_rate_samples, bfqd->sequential_samples,
3367		bfqd->tot_sectors_dispatched);
3368}
3369
3370static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
3371{
3372	u32 rate, weight, divisor;
3373
3374	/*
3375	 * For the convergence property to hold (see comments on
3376	 * bfq_update_peak_rate()) and for the assessment to be
3377	 * reliable, a minimum number of samples must be present, and
3378	 * a minimum amount of time must have elapsed. If not so, do
3379	 * not compute new rate. Just reset parameters, to get ready
3380	 * for a new evaluation attempt.
3381	 */
3382	if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
3383	    bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
3384		goto reset_computation;
3385
3386	/*
3387	 * If a new request completion has occurred after last
3388	 * dispatch, then, to approximate the rate at which requests
3389	 * have been served by the device, it is more precise to
3390	 * extend the observation interval to the last completion.
3391	 */
3392	bfqd->delta_from_first =
3393		max_t(u64, bfqd->delta_from_first,
3394		      bfqd->last_completion - bfqd->first_dispatch);
3395
3396	/*
3397	 * Rate computed in sects/usec, and not sects/nsec, for
3398	 * precision issues.
3399	 */
3400	rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
3401			div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
3402
3403	/*
3404	 * Peak rate not updated if:
3405	 * - the percentage of sequential dispatches is below 3/4 of the
3406	 *   total, and rate is below the current estimated peak rate
3407	 * - rate is unreasonably high (> 20M sectors/sec)
3408	 */
3409	if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
3410	     rate <= bfqd->peak_rate) ||
3411		rate > 20<<BFQ_RATE_SHIFT)
3412		goto reset_computation;
3413
3414	/*
3415	 * We have to update the peak rate, at last! To this purpose,
3416	 * we use a low-pass filter. We compute the smoothing constant
3417	 * of the filter as a function of the 'weight' of the new
3418	 * measured rate.
3419	 *
3420	 * As can be seen in next formulas, we define this weight as a
3421	 * quantity proportional to how sequential the workload is,
3422	 * and to how long the observation time interval is.
3423	 *
3424	 * The weight runs from 0 to 8. The maximum value of the
3425	 * weight, 8, yields the minimum value for the smoothing
3426	 * constant. At this minimum value for the smoothing constant,
3427	 * the measured rate contributes for half of the next value of
3428	 * the estimated peak rate.
3429	 *
3430	 * So, the first step is to compute the weight as a function
3431	 * of how sequential the workload is. Note that the weight
3432	 * cannot reach 9, because bfqd->sequential_samples cannot
3433	 * become equal to bfqd->peak_rate_samples, which, in its
3434	 * turn, holds true because bfqd->sequential_samples is not
3435	 * incremented for the first sample.
3436	 */
3437	weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
3438
3439	/*
3440	 * Second step: further refine the weight as a function of the
3441	 * duration of the observation interval.
3442	 */
3443	weight = min_t(u32, 8,
3444		       div_u64(weight * bfqd->delta_from_first,
3445			       BFQ_RATE_REF_INTERVAL));
3446
3447	/*
3448	 * Divisor ranging from 10, for minimum weight, to 2, for
3449	 * maximum weight.
3450	 */
3451	divisor = 10 - weight;
3452
3453	/*
3454	 * Finally, update peak rate:
3455	 *
3456	 * peak_rate = peak_rate * (divisor-1) / divisor  +  rate / divisor
3457	 */
3458	bfqd->peak_rate *= divisor-1;
3459	bfqd->peak_rate /= divisor;
3460	rate /= divisor; /* smoothing constant alpha = 1/divisor */
3461
3462	bfqd->peak_rate += rate;
3463
3464	/*
3465	 * For a very slow device, bfqd->peak_rate can reach 0 (see
3466	 * the minimum representable values reported in the comments
3467	 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3468	 * divisions by zero where bfqd->peak_rate is used as a
3469	 * divisor.
3470	 */
3471	bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
3472
3473	update_thr_responsiveness_params(bfqd);
3474
3475reset_computation:
3476	bfq_reset_rate_computation(bfqd, rq);
3477}
3478
3479/*
3480 * Update the read/write peak rate (the main quantity used for
3481 * auto-tuning, see update_thr_responsiveness_params()).
3482 *
3483 * It is not trivial to estimate the peak rate (correctly): because of
3484 * the presence of sw and hw queues between the scheduler and the
3485 * device components that finally serve I/O requests, it is hard to
3486 * say exactly when a given dispatched request is served inside the
3487 * device, and for how long. As a consequence, it is hard to know
3488 * precisely at what rate a given set of requests is actually served
3489 * by the device.
3490 *
3491 * On the opposite end, the dispatch time of any request is trivially
3492 * available, and, from this piece of information, the "dispatch rate"
3493 * of requests can be immediately computed. So, the idea in the next
3494 * function is to use what is known, namely request dispatch times
3495 * (plus, when useful, request completion times), to estimate what is
3496 * unknown, namely in-device request service rate.
3497 *
3498 * The main issue is that, because of the above facts, the rate at
3499 * which a certain set of requests is dispatched over a certain time
3500 * interval can vary greatly with respect to the rate at which the
3501 * same requests are then served. But, since the size of any
3502 * intermediate queue is limited, and the service scheme is lossless
3503 * (no request is silently dropped), the following obvious convergence
3504 * property holds: the number of requests dispatched MUST become
3505 * closer and closer to the number of requests completed as the
3506 * observation interval grows. This is the key property used in
3507 * the next function to estimate the peak service rate as a function
3508 * of the observed dispatch rate. The function assumes to be invoked
3509 * on every request dispatch.
3510 */
3511static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
3512{
3513	u64 now_ns = ktime_get_ns();
3514
3515	if (bfqd->peak_rate_samples == 0) { /* first dispatch */
3516		bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
3517			bfqd->peak_rate_samples);
3518		bfq_reset_rate_computation(bfqd, rq);
3519		goto update_last_values; /* will add one sample */
3520	}
3521
3522	/*
3523	 * Device idle for very long: the observation interval lasting
3524	 * up to this dispatch cannot be a valid observation interval
3525	 * for computing a new peak rate (similarly to the late-
3526	 * completion event in bfq_completed_request()). Go to
3527	 * update_rate_and_reset to have the following three steps
3528	 * taken:
3529	 * - close the observation interval at the last (previous)
3530	 *   request dispatch or completion
3531	 * - compute rate, if possible, for that observation interval
3532	 * - start a new observation interval with this dispatch
3533	 */
3534	if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
3535	    bfqd->rq_in_driver == 0)
3536		goto update_rate_and_reset;
3537
3538	/* Update sampling information */
3539	bfqd->peak_rate_samples++;
3540
3541	if ((bfqd->rq_in_driver > 0 ||
3542		now_ns - bfqd->last_completion < BFQ_MIN_TT)
3543	    && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
3544		bfqd->sequential_samples++;
3545
3546	bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
3547
3548	/* Reset max observed rq size every 32 dispatches */
3549	if (likely(bfqd->peak_rate_samples % 32))
3550		bfqd->last_rq_max_size =
3551			max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
3552	else
3553		bfqd->last_rq_max_size = blk_rq_sectors(rq);
3554
3555	bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
3556
3557	/* Target observation interval not yet reached, go on sampling */
3558	if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
3559		goto update_last_values;
3560
3561update_rate_and_reset:
3562	bfq_update_rate_reset(bfqd, rq);
3563update_last_values:
3564	bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
3565	if (RQ_BFQQ(rq) == bfqd->in_service_queue)
3566		bfqd->in_serv_last_pos = bfqd->last_position;
3567	bfqd->last_dispatch = now_ns;
3568}
3569
3570/*
3571 * Remove request from internal lists.
3572 */
3573static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
3574{
3575	struct bfq_queue *bfqq = RQ_BFQQ(rq);
3576
3577	/*
3578	 * For consistency, the next instruction should have been
3579	 * executed after removing the request from the queue and
3580	 * dispatching it.  We execute instead this instruction before
3581	 * bfq_remove_request() (and hence introduce a temporary
3582	 * inconsistency), for efficiency.  In fact, should this
3583	 * dispatch occur for a non in-service bfqq, this anticipated
3584	 * increment prevents two counters related to bfqq->dispatched
3585	 * from risking to be, first, uselessly decremented, and then
3586	 * incremented again when the (new) value of bfqq->dispatched
3587	 * happens to be taken into account.
3588	 */
3589	bfqq->dispatched++;
3590	bfq_update_peak_rate(q->elevator->elevator_data, rq);
3591
3592	bfq_remove_request(q, rq);
3593}
3594
3595/*
3596 * There is a case where idling does not have to be performed for
3597 * throughput concerns, but to preserve the throughput share of
3598 * the process associated with bfqq.
3599 *
3600 * To introduce this case, we can note that allowing the drive
3601 * to enqueue more than one request at a time, and hence
3602 * delegating de facto final scheduling decisions to the
3603 * drive's internal scheduler, entails loss of control on the
3604 * actual request service order. In particular, the critical
3605 * situation is when requests from different processes happen
3606 * to be present, at the same time, in the internal queue(s)
3607 * of the drive. In such a situation, the drive, by deciding
3608 * the service order of the internally-queued requests, does
3609 * determine also the actual throughput distribution among
3610 * these processes. But the drive typically has no notion or
3611 * concern about per-process throughput distribution, and
3612 * makes its decisions only on a per-request basis. Therefore,
3613 * the service distribution enforced by the drive's internal
3614 * scheduler is likely to coincide with the desired throughput
3615 * distribution only in a completely symmetric, or favorably
3616 * skewed scenario where:
3617 * (i-a) each of these processes must get the same throughput as
3618 *	 the others,
3619 * (i-b) in case (i-a) does not hold, it holds that the process
3620 *       associated with bfqq must receive a lower or equal
3621 *	 throughput than any of the other processes;
3622 * (ii)  the I/O of each process has the same properties, in
3623 *       terms of locality (sequential or random), direction
3624 *       (reads or writes), request sizes, greediness
3625 *       (from I/O-bound to sporadic), and so on;
3626
3627 * In fact, in such a scenario, the drive tends to treat the requests
3628 * of each process in about the same way as the requests of the
3629 * others, and thus to provide each of these processes with about the
3630 * same throughput.  This is exactly the desired throughput
3631 * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3632 * even more convenient distribution for (the process associated with)
3633 * bfqq.
3634 *
3635 * In contrast, in any asymmetric or unfavorable scenario, device
3636 * idling (I/O-dispatch plugging) is certainly needed to guarantee
3637 * that bfqq receives its assigned fraction of the device throughput
3638 * (see [1] for details).
3639 *
3640 * The problem is that idling may significantly reduce throughput with
3641 * certain combinations of types of I/O and devices. An important
3642 * example is sync random I/O on flash storage with command
3643 * queueing. So, unless bfqq falls in cases where idling also boosts
3644 * throughput, it is important to check conditions (i-a), i(-b) and
3645 * (ii) accurately, so as to avoid idling when not strictly needed for
3646 * service guarantees.
3647 *
3648 * Unfortunately, it is extremely difficult to thoroughly check
3649 * condition (ii). And, in case there are active groups, it becomes
3650 * very difficult to check conditions (i-a) and (i-b) too.  In fact,
3651 * if there are active groups, then, for conditions (i-a) or (i-b) to
3652 * become false 'indirectly', it is enough that an active group
3653 * contains more active processes or sub-groups than some other active
3654 * group. More precisely, for conditions (i-a) or (i-b) to become
3655 * false because of such a group, it is not even necessary that the
3656 * group is (still) active: it is sufficient that, even if the group
3657 * has become inactive, some of its descendant processes still have
3658 * some request already dispatched but still waiting for
3659 * completion. In fact, requests have still to be guaranteed their
3660 * share of the throughput even after being dispatched. In this
3661 * respect, it is easy to show that, if a group frequently becomes
3662 * inactive while still having in-flight requests, and if, when this
3663 * happens, the group is not considered in the calculation of whether
3664 * the scenario is asymmetric, then the group may fail to be
3665 * guaranteed its fair share of the throughput (basically because
3666 * idling may not be performed for the descendant processes of the
3667 * group, but it had to be).  We address this issue with the following
3668 * bi-modal behavior, implemented in the function
3669 * bfq_asymmetric_scenario().
3670 *
3671 * If there are groups with requests waiting for completion
3672 * (as commented above, some of these groups may even be
3673 * already inactive), then the scenario is tagged as
3674 * asymmetric, conservatively, without checking any of the
3675 * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3676 * This behavior matches also the fact that groups are created
3677 * exactly if controlling I/O is a primary concern (to
3678 * preserve bandwidth and latency guarantees).
3679 *
3680 * On the opposite end, if there are no groups with requests waiting
3681 * for completion, then only conditions (i-a) and (i-b) are actually
3682 * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3683 * idling is not performed, regardless of whether condition (ii)
3684 * holds.  In other words, only if conditions (i-a) and (i-b) do not
3685 * hold, then idling is allowed, and the device tends to be prevented
3686 * from queueing many requests, possibly of several processes. Since
3687 * there are no groups with requests waiting for completion, then, to
3688 * control conditions (i-a) and (i-b) it is enough to check just
3689 * whether all the queues with requests waiting for completion also
3690 * have the same weight.
3691 *
3692 * Not checking condition (ii) evidently exposes bfqq to the
3693 * risk of getting less throughput than its fair share.
3694 * However, for queues with the same weight, a further
3695 * mechanism, preemption, mitigates or even eliminates this
3696 * problem. And it does so without consequences on overall
3697 * throughput. This mechanism and its benefits are explained
3698 * in the next three paragraphs.
3699 *
3700 * Even if a queue, say Q, is expired when it remains idle, Q
3701 * can still preempt the new in-service queue if the next
3702 * request of Q arrives soon (see the comments on
3703 * bfq_bfqq_update_budg_for_activation). If all queues and
3704 * groups have the same weight, this form of preemption,
3705 * combined with the hole-recovery heuristic described in the
3706 * comments on function bfq_bfqq_update_budg_for_activation,
3707 * are enough to preserve a correct bandwidth distribution in
3708 * the mid term, even without idling. In fact, even if not
3709 * idling allows the internal queues of the device to contain
3710 * many requests, and thus to reorder requests, we can rather
3711 * safely assume that the internal scheduler still preserves a
3712 * minimum of mid-term fairness.
3713 *
3714 * More precisely, this preemption-based, idleless approach
3715 * provides fairness in terms of IOPS, and not sectors per
3716 * second. This can be seen with a simple example. Suppose
3717 * that there are two queues with the same weight, but that
3718 * the first queue receives requests of 8 sectors, while the
3719 * second queue receives requests of 1024 sectors. In
3720 * addition, suppose that each of the two queues contains at
3721 * most one request at a time, which implies that each queue
3722 * always remains idle after it is served. Finally, after
3723 * remaining idle, each queue receives very quickly a new
3724 * request. It follows that the two queues are served
3725 * alternatively, preempting each other if needed. This
3726 * implies that, although both queues have the same weight,
3727 * the queue with large requests receives a service that is
3728 * 1024/8 times as high as the service received by the other
3729 * queue.
3730 *
3731 * The motivation for using preemption instead of idling (for
3732 * queues with the same weight) is that, by not idling,
3733 * service guarantees are preserved (completely or at least in
3734 * part) without minimally sacrificing throughput. And, if
3735 * there is no active group, then the primary expectation for
3736 * this device is probably a high throughput.
3737 *
3738 * We are now left only with explaining the two sub-conditions in the
3739 * additional compound condition that is checked below for deciding
3740 * whether the scenario is asymmetric. To explain the first
3741 * sub-condition, we need to add that the function
3742 * bfq_asymmetric_scenario checks the weights of only
3743 * non-weight-raised queues, for efficiency reasons (see comments on
3744 * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3745 * is checked explicitly here. More precisely, the compound condition
3746 * below takes into account also the fact that, even if bfqq is being
3747 * weight-raised, the scenario is still symmetric if all queues with
3748 * requests waiting for completion happen to be
3749 * weight-raised. Actually, we should be even more precise here, and
3750 * differentiate between interactive weight raising and soft real-time
3751 * weight raising.
3752 *
3753 * The second sub-condition checked in the compound condition is
3754 * whether there is a fair amount of already in-flight I/O not
3755 * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3756 * following reason. The drive may decide to serve in-flight
3757 * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3758 * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3759 * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3760 * basically uncontrolled amount of I/O from other queues may be
3761 * dispatched too, possibly causing the service of bfqq's I/O to be
3762 * delayed even longer in the drive. This problem gets more and more
3763 * serious as the speed and the queue depth of the drive grow,
3764 * because, as these two quantities grow, the probability to find no
3765 * queue busy but many requests in flight grows too. By contrast,
3766 * plugging I/O dispatching minimizes the delay induced by already
3767 * in-flight I/O, and enables bfqq to recover the bandwidth it may
3768 * lose because of this delay.
3769 *
3770 * As a side note, it is worth considering that the above
3771 * device-idling countermeasures may however fail in the following
3772 * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3773 * in a time period during which all symmetry sub-conditions hold, and
3774 * therefore the device is allowed to enqueue many requests, but at
3775 * some later point in time some sub-condition stops to hold, then it
3776 * may become impossible to make requests be served in the desired
3777 * order until all the requests already queued in the device have been
3778 * served. The last sub-condition commented above somewhat mitigates
3779 * this problem for weight-raised queues.
3780 *
3781 * However, as an additional mitigation for this problem, we preserve
3782 * plugging for a special symmetric case that may suddenly turn into
3783 * asymmetric: the case where only bfqq is busy. In this case, not
3784 * expiring bfqq does not cause any harm to any other queues in terms
3785 * of service guarantees. In contrast, it avoids the following unlucky
3786 * sequence of events: (1) bfqq is expired, (2) a new queue with a
3787 * lower weight than bfqq becomes busy (or more queues), (3) the new
3788 * queue is served until a new request arrives for bfqq, (4) when bfqq
3789 * is finally served, there are so many requests of the new queue in
3790 * the drive that the pending requests for bfqq take a lot of time to
3791 * be served. In particular, event (2) may case even already
3792 * dispatched requests of bfqq to be delayed, inside the drive. So, to
3793 * avoid this series of events, the scenario is preventively declared
3794 * as asymmetric also if bfqq is the only busy queues
3795 */
3796static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3797						 struct bfq_queue *bfqq)
3798{
3799	int tot_busy_queues = bfq_tot_busy_queues(bfqd);
3800
3801	/* No point in idling for bfqq if it won't get requests any longer */
3802	if (unlikely(!bfqq_process_refs(bfqq)))
3803		return false;
3804
3805	return (bfqq->wr_coeff > 1 &&
3806		(bfqd->wr_busy_queues <
3807		 tot_busy_queues ||
3808		 bfqd->rq_in_driver >=
3809		 bfqq->dispatched + 4)) ||
3810		bfq_asymmetric_scenario(bfqd, bfqq) ||
3811		tot_busy_queues == 1;
3812}
3813
3814static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3815			      enum bfqq_expiration reason)
3816{
3817	/*
3818	 * If this bfqq is shared between multiple processes, check
3819	 * to make sure that those processes are still issuing I/Os
3820	 * within the mean seek distance. If not, it may be time to
3821	 * break the queues apart again.
3822	 */
3823	if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
3824		bfq_mark_bfqq_split_coop(bfqq);
3825
3826	/*
3827	 * Consider queues with a higher finish virtual time than
3828	 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3829	 * true, then bfqq's bandwidth would be violated if an
3830	 * uncontrolled amount of I/O from these queues were
3831	 * dispatched while bfqq is waiting for its new I/O to
3832	 * arrive. This is exactly what may happen if this is a forced
3833	 * expiration caused by a preemption attempt, and if bfqq is
3834	 * not re-scheduled. To prevent this from happening, re-queue
3835	 * bfqq if it needs I/O-dispatch plugging, even if it is
3836	 * empty. By doing so, bfqq is granted to be served before the
3837	 * above queues (provided that bfqq is of course eligible).
3838	 */
3839	if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
3840	    !(reason == BFQQE_PREEMPTED &&
3841	      idling_needed_for_service_guarantees(bfqd, bfqq))) {
3842		if (bfqq->dispatched == 0)
3843			/*
3844			 * Overloading budget_timeout field to store
3845			 * the time at which the queue remains with no
3846			 * backlog and no outstanding request; used by
3847			 * the weight-raising mechanism.
3848			 */
3849			bfqq->budget_timeout = jiffies;
3850
3851		bfq_del_bfqq_busy(bfqq, true);
3852	} else {
3853		bfq_requeue_bfqq(bfqd, bfqq, true);
3854		/*
3855		 * Resort priority tree of potential close cooperators.
3856		 * See comments on bfq_pos_tree_add_move() for the unlikely().
3857		 */
3858		if (unlikely(!bfqd->nonrot_with_queueing &&
3859			     !RB_EMPTY_ROOT(&bfqq->sort_list)))
3860			bfq_pos_tree_add_move(bfqd, bfqq);
3861	}
3862
3863	/*
3864	 * All in-service entities must have been properly deactivated
3865	 * or requeued before executing the next function, which
3866	 * resets all in-service entities as no more in service. This
3867	 * may cause bfqq to be freed. If this happens, the next
3868	 * function returns true.
3869	 */
3870	return __bfq_bfqd_reset_in_service(bfqd);
3871}
3872
3873/**
3874 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3875 * @bfqd: device data.
3876 * @bfqq: queue to update.
3877 * @reason: reason for expiration.
3878 *
3879 * Handle the feedback on @bfqq budget at queue expiration.
3880 * See the body for detailed comments.
3881 */
3882static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
3883				     struct bfq_queue *bfqq,
3884				     enum bfqq_expiration reason)
3885{
3886	struct request *next_rq;
3887	int budget, min_budget;
3888
3889	min_budget = bfq_min_budget(bfqd);
3890
3891	if (bfqq->wr_coeff == 1)
3892		budget = bfqq->max_budget;
3893	else /*
3894	      * Use a constant, low budget for weight-raised queues,
3895	      * to help achieve a low latency. Keep it slightly higher
3896	      * than the minimum possible budget, to cause a little
3897	      * bit fewer expirations.
3898	      */
3899		budget = 2 * min_budget;
3900
3901	bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
3902		bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
3903	bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
3904		budget, bfq_min_budget(bfqd));
3905	bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
3906		bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
3907
3908	if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
3909		switch (reason) {
3910		/*
3911		 * Caveat: in all the following cases we trade latency
3912		 * for throughput.
3913		 */
3914		case BFQQE_TOO_IDLE:
3915			/*
3916			 * This is the only case where we may reduce
3917			 * the budget: if there is no request of the
3918			 * process still waiting for completion, then
3919			 * we assume (tentatively) that the timer has
3920			 * expired because the batch of requests of
3921			 * the process could have been served with a
3922			 * smaller budget.  Hence, betting that
3923			 * process will behave in the same way when it
3924			 * becomes backlogged again, we reduce its
3925			 * next budget.  As long as we guess right,
3926			 * this budget cut reduces the latency
3927			 * experienced by the process.
3928			 *
3929			 * However, if there are still outstanding
3930			 * requests, then the process may have not yet
3931			 * issued its next request just because it is
3932			 * still waiting for the completion of some of
3933			 * the still outstanding ones.  So in this
3934			 * subcase we do not reduce its budget, on the
3935			 * contrary we increase it to possibly boost
3936			 * the throughput, as discussed in the
3937			 * comments to the BUDGET_TIMEOUT case.
3938			 */
3939			if (bfqq->dispatched > 0) /* still outstanding reqs */
3940				budget = min(budget * 2, bfqd->bfq_max_budget);
3941			else {
3942				if (budget > 5 * min_budget)
3943					budget -= 4 * min_budget;
3944				else
3945					budget = min_budget;
3946			}
3947			break;
3948		case BFQQE_BUDGET_TIMEOUT:
3949			/*
3950			 * We double the budget here because it gives
3951			 * the chance to boost the throughput if this
3952			 * is not a seeky process (and has bumped into
3953			 * this timeout because of, e.g., ZBR).
3954			 */
3955			budget = min(budget * 2, bfqd->bfq_max_budget);
3956			break;
3957		case BFQQE_BUDGET_EXHAUSTED:
3958			/*
3959			 * The process still has backlog, and did not
3960			 * let either the budget timeout or the disk
3961			 * idling timeout expire. Hence it is not
3962			 * seeky, has a short thinktime and may be
3963			 * happy with a higher budget too. So
3964			 * definitely increase the budget of this good
3965			 * candidate to boost the disk throughput.
3966			 */
3967			budget = min(budget * 4, bfqd->bfq_max_budget);
3968			break;
3969		case BFQQE_NO_MORE_REQUESTS:
3970			/*
3971			 * For queues that expire for this reason, it
3972			 * is particularly important to keep the
3973			 * budget close to the actual service they
3974			 * need. Doing so reduces the timestamp
3975			 * misalignment problem described in the
3976			 * comments in the body of
3977			 * __bfq_activate_entity. In fact, suppose
3978			 * that a queue systematically expires for
3979			 * BFQQE_NO_MORE_REQUESTS and presents a
3980			 * new request in time to enjoy timestamp
3981			 * back-shifting. The larger the budget of the
3982			 * queue is with respect to the service the
3983			 * queue actually requests in each service
3984			 * slot, the more times the queue can be
3985			 * reactivated with the same virtual finish
3986			 * time. It follows that, even if this finish
3987			 * time is pushed to the system virtual time
3988			 * to reduce the consequent timestamp
3989			 * misalignment, the queue unjustly enjoys for
3990			 * many re-activations a lower finish time
3991			 * than all newly activated queues.
3992			 *
3993			 * The service needed by bfqq is measured
3994			 * quite precisely by bfqq->entity.service.
3995			 * Since bfqq does not enjoy device idling,
3996			 * bfqq->entity.service is equal to the number
3997			 * of sectors that the process associated with
3998			 * bfqq requested to read/write before waiting
3999			 * for request completions, or blocking for
4000			 * other reasons.
4001			 */
4002			budget = max_t(int, bfqq->entity.service, min_budget);
4003			break;
4004		default:
4005			return;
4006		}
4007	} else if (!bfq_bfqq_sync(bfqq)) {
4008		/*
4009		 * Async queues get always the maximum possible
4010		 * budget, as for them we do not care about latency
4011		 * (in addition, their ability to dispatch is limited
4012		 * by the charging factor).
4013		 */
4014		budget = bfqd->bfq_max_budget;
4015	}
4016
4017	bfqq->max_budget = budget;
4018
4019	if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
4020	    !bfqd->bfq_user_max_budget)
4021		bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
4022
4023	/*
4024	 * If there is still backlog, then assign a new budget, making
4025	 * sure that it is large enough for the next request.  Since
4026	 * the finish time of bfqq must be kept in sync with the
4027	 * budget, be sure to call __bfq_bfqq_expire() *after* this
4028	 * update.
4029	 *
4030	 * If there is no backlog, then no need to update the budget;
4031	 * it will be updated on the arrival of a new request.
4032	 */
4033	next_rq = bfqq->next_rq;
4034	if (next_rq)
4035		bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
4036					    bfq_serv_to_charge(next_rq, bfqq));
4037
4038	bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
4039			next_rq ? blk_rq_sectors(next_rq) : 0,
4040			bfqq->entity.budget);
4041}
4042
4043/*
4044 * Return true if the process associated with bfqq is "slow". The slow
4045 * flag is used, in addition to the budget timeout, to reduce the
4046 * amount of service provided to seeky processes, and thus reduce
4047 * their chances to lower the throughput. More details in the comments
4048 * on the function bfq_bfqq_expire().
4049 *
4050 * An important observation is in order: as discussed in the comments
4051 * on the function bfq_update_peak_rate(), with devices with internal
4052 * queues, it is hard if ever possible to know when and for how long
4053 * an I/O request is processed by the device (apart from the trivial
4054 * I/O pattern where a new request is dispatched only after the
4055 * previous one has been completed). This makes it hard to evaluate
4056 * the real rate at which the I/O requests of each bfq_queue are
4057 * served.  In fact, for an I/O scheduler like BFQ, serving a
4058 * bfq_queue means just dispatching its requests during its service
4059 * slot (i.e., until the budget of the queue is exhausted, or the
4060 * queue remains idle, or, finally, a timeout fires). But, during the
4061 * service slot of a bfq_queue, around 100 ms at most, the device may
4062 * be even still processing requests of bfq_queues served in previous
4063 * service slots. On the opposite end, the requests of the in-service
4064 * bfq_queue may be completed after the service slot of the queue
4065 * finishes.
4066 *
4067 * Anyway, unless more sophisticated solutions are used
4068 * (where possible), the sum of the sizes of the requests dispatched
4069 * during the service slot of a bfq_queue is probably the only
4070 * approximation available for the service received by the bfq_queue
4071 * during its service slot. And this sum is the quantity used in this
4072 * function to evaluate the I/O speed of a process.
4073 */
4074static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4075				 bool compensate, enum bfqq_expiration reason,
4076				 unsigned long *delta_ms)
4077{
4078	ktime_t delta_ktime;
4079	u32 delta_usecs;
4080	bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
4081
4082	if (!bfq_bfqq_sync(bfqq))
4083		return false;
4084
4085	if (compensate)
4086		delta_ktime = bfqd->last_idling_start;
4087	else
4088		delta_ktime = ktime_get();
4089	delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
4090	delta_usecs = ktime_to_us(delta_ktime);
4091
4092	/* don't use too short time intervals */
4093	if (delta_usecs < 1000) {
4094		if (blk_queue_nonrot(bfqd->queue))
4095			 /*
4096			  * give same worst-case guarantees as idling
4097			  * for seeky
4098			  */
4099			*delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
4100		else /* charge at least one seek */
4101			*delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
4102
4103		return slow;
4104	}
4105
4106	*delta_ms = delta_usecs / USEC_PER_MSEC;
4107
4108	/*
4109	 * Use only long (> 20ms) intervals to filter out excessive
4110	 * spikes in service rate estimation.
4111	 */
4112	if (delta_usecs > 20000) {
4113		/*
4114		 * Caveat for rotational devices: processes doing I/O
4115		 * in the slower disk zones tend to be slow(er) even
4116		 * if not seeky. In this respect, the estimated peak
4117		 * rate is likely to be an average over the disk
4118		 * surface. Accordingly, to not be too harsh with
4119		 * unlucky processes, a process is deemed slow only if
4120		 * its rate has been lower than half of the estimated
4121		 * peak rate.
4122		 */
4123		slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
4124	}
4125
4126	bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
4127
4128	return slow;
4129}
4130
4131/*
4132 * To be deemed as soft real-time, an application must meet two
4133 * requirements. First, the application must not require an average
4134 * bandwidth higher than the approximate bandwidth required to playback or
4135 * record a compressed high-definition video.
4136 * The next function is invoked on the completion of the last request of a
4137 * batch, to compute the next-start time instant, soft_rt_next_start, such
4138 * that, if the next request of the application does not arrive before
4139 * soft_rt_next_start, then the above requirement on the bandwidth is met.
4140 *
4141 * The second requirement is that the request pattern of the application is
4142 * isochronous, i.e., that, after issuing a request or a batch of requests,
4143 * the application stops issuing new requests until all its pending requests
4144 * have been completed. After that, the application may issue a new batch,
4145 * and so on.
4146 * For this reason the next function is invoked to compute
4147 * soft_rt_next_start only for applications that meet this requirement,
4148 * whereas soft_rt_next_start is set to infinity for applications that do
4149 * not.
4150 *
4151 * Unfortunately, even a greedy (i.e., I/O-bound) application may
4152 * happen to meet, occasionally or systematically, both the above
4153 * bandwidth and isochrony requirements. This may happen at least in
4154 * the following circumstances. First, if the CPU load is high. The
4155 * application may stop issuing requests while the CPUs are busy
4156 * serving other processes, then restart, then stop again for a while,
4157 * and so on. The other circumstances are related to the storage
4158 * device: the storage device is highly loaded or reaches a low-enough
4159 * throughput with the I/O of the application (e.g., because the I/O
4160 * is random and/or the device is slow). In all these cases, the
4161 * I/O of the application may be simply slowed down enough to meet
4162 * the bandwidth and isochrony requirements. To reduce the probability
4163 * that greedy applications are deemed as soft real-time in these
4164 * corner cases, a further rule is used in the computation of
4165 * soft_rt_next_start: the return value of this function is forced to
4166 * be higher than the maximum between the following two quantities.
4167 *
4168 * (a) Current time plus: (1) the maximum time for which the arrival
4169 *     of a request is waited for when a sync queue becomes idle,
4170 *     namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
4171 *     postpone for a moment the reason for adding a few extra
4172 *     jiffies; we get back to it after next item (b).  Lower-bounding
4173 *     the return value of this function with the current time plus
4174 *     bfqd->bfq_slice_idle tends to filter out greedy applications,
4175 *     because the latter issue their next request as soon as possible
4176 *     after the last one has been completed. In contrast, a soft
4177 *     real-time application spends some time processing data, after a
4178 *     batch of its requests has been completed.
4179 *
4180 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
4181 *     above, greedy applications may happen to meet both the
4182 *     bandwidth and isochrony requirements under heavy CPU or
4183 *     storage-device load. In more detail, in these scenarios, these
4184 *     applications happen, only for limited time periods, to do I/O
4185 *     slowly enough to meet all the requirements described so far,
4186 *     including the filtering in above item (a). These slow-speed
4187 *     time intervals are usually interspersed between other time
4188 *     intervals during which these applications do I/O at a very high
4189 *     speed. Fortunately, exactly because of the high speed of the
4190 *     I/O in the high-speed intervals, the values returned by this
4191 *     function happen to be so high, near the end of any such
4192 *     high-speed interval, to be likely to fall *after* the end of
4193 *     the low-speed time interval that follows. These high values are
4194 *     stored in bfqq->soft_rt_next_start after each invocation of
4195 *     this function. As a consequence, if the last value of
4196 *     bfqq->soft_rt_next_start is constantly used to lower-bound the
4197 *     next value that this function may return, then, from the very
4198 *     beginning of a low-speed interval, bfqq->soft_rt_next_start is
4199 *     likely to be constantly kept so high that any I/O request
4200 *     issued during the low-speed interval is considered as arriving
4201 *     to soon for the application to be deemed as soft
4202 *     real-time. Then, in the high-speed interval that follows, the
4203 *     application will not be deemed as soft real-time, just because
4204 *     it will do I/O at a high speed. And so on.
4205 *
4206 * Getting back to the filtering in item (a), in the following two
4207 * cases this filtering might be easily passed by a greedy
4208 * application, if the reference quantity was just
4209 * bfqd->bfq_slice_idle:
4210 * 1) HZ is so low that the duration of a jiffy is comparable to or
4211 *    higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
4212 *    devices with HZ=100. The time granularity may be so coarse
4213 *    that the approximation, in jiffies, of bfqd->bfq_slice_idle
4214 *    is rather lower than the exact value.
4215 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
4216 *    for a while, then suddenly 'jump' by several units to recover the lost
4217 *    increments. This seems to happen, e.g., inside virtual machines.
4218 * To address this issue, in the filtering in (a) we do not use as a
4219 * reference time interval just bfqd->bfq_slice_idle, but
4220 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
4221 * minimum number of jiffies for which the filter seems to be quite
4222 * precise also in embedded systems and KVM/QEMU virtual machines.
4223 */
4224static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
4225						struct bfq_queue *bfqq)
4226{
4227	return max3(bfqq->soft_rt_next_start,
4228		    bfqq->last_idle_bklogged +
4229		    HZ * bfqq->service_from_backlogged /
4230		    bfqd->bfq_wr_max_softrt_rate,
4231		    jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
4232}
4233
4234/**
4235 * bfq_bfqq_expire - expire a queue.
4236 * @bfqd: device owning the queue.
4237 * @bfqq: the queue to expire.
4238 * @compensate: if true, compensate for the time spent idling.
4239 * @reason: the reason causing the expiration.
4240 *
4241 * If the process associated with bfqq does slow I/O (e.g., because it
4242 * issues random requests), we charge bfqq with the time it has been
4243 * in service instead of the service it has received (see
4244 * bfq_bfqq_charge_time for details on how this goal is achieved). As
4245 * a consequence, bfqq will typically get higher timestamps upon
4246 * reactivation, and hence it will be rescheduled as if it had
4247 * received more service than what it has actually received. In the
4248 * end, bfqq receives less service in proportion to how slowly its
4249 * associated process consumes its budgets (and hence how seriously it
4250 * tends to lower the throughput). In addition, this time-charging
4251 * strategy guarantees time fairness among slow processes. In
4252 * contrast, if the process associated with bfqq is not slow, we
4253 * charge bfqq exactly with the service it has received.
4254 *
4255 * Charging time to the first type of queues and the exact service to
4256 * the other has the effect of using the WF2Q+ policy to schedule the
4257 * former on a timeslice basis, without violating service domain
4258 * guarantees among the latter.
4259 */
4260void bfq_bfqq_expire(struct bfq_data *bfqd,
4261		     struct bfq_queue *bfqq,
4262		     bool compensate,
4263		     enum bfqq_expiration reason)
4264{
4265	bool slow;
4266	unsigned long delta = 0;
4267	struct bfq_entity *entity = &bfqq->entity;
 
4268
4269	/*
4270	 * Check whether the process is slow (see bfq_bfqq_is_slow).
4271	 */
4272	slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
4273
4274	/*
4275	 * As above explained, charge slow (typically seeky) and
4276	 * timed-out queues with the time and not the service
4277	 * received, to favor sequential workloads.
4278	 *
4279	 * Processes doing I/O in the slower disk zones will tend to
4280	 * be slow(er) even if not seeky. Therefore, since the
4281	 * estimated peak rate is actually an average over the disk
4282	 * surface, these processes may timeout just for bad luck. To
4283	 * avoid punishing them, do not charge time to processes that
4284	 * succeeded in consuming at least 2/3 of their budget. This
4285	 * allows BFQ to preserve enough elasticity to still perform
4286	 * bandwidth, and not time, distribution with little unlucky
4287	 * or quasi-sequential processes.
4288	 */
4289	if (bfqq->wr_coeff == 1 &&
4290	    (slow ||
4291	     (reason == BFQQE_BUDGET_TIMEOUT &&
4292	      bfq_bfqq_budget_left(bfqq) >=  entity->budget / 3)))
4293		bfq_bfqq_charge_time(bfqd, bfqq, delta);
4294
 
 
 
 
4295	if (bfqd->low_latency && bfqq->wr_coeff == 1)
4296		bfqq->last_wr_start_finish = jiffies;
4297
4298	if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
4299	    RB_EMPTY_ROOT(&bfqq->sort_list)) {
4300		/*
4301		 * If we get here, and there are no outstanding
4302		 * requests, then the request pattern is isochronous
4303		 * (see the comments on the function
4304		 * bfq_bfqq_softrt_next_start()). Therefore we can
4305		 * compute soft_rt_next_start.
4306		 *
4307		 * If, instead, the queue still has outstanding
4308		 * requests, then we have to wait for the completion
4309		 * of all the outstanding requests to discover whether
4310		 * the request pattern is actually isochronous.
4311		 */
4312		if (bfqq->dispatched == 0)
4313			bfqq->soft_rt_next_start =
4314				bfq_bfqq_softrt_next_start(bfqd, bfqq);
4315		else if (bfqq->dispatched > 0) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4316			/*
4317			 * Schedule an update of soft_rt_next_start to when
4318			 * the task may be discovered to be isochronous.
4319			 */
4320			bfq_mark_bfqq_softrt_update(bfqq);
4321		}
4322	}
4323
4324	bfq_log_bfqq(bfqd, bfqq,
4325		"expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
4326		slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
4327
4328	/*
4329	 * bfqq expired, so no total service time needs to be computed
4330	 * any longer: reset state machine for measuring total service
4331	 * times.
4332	 */
4333	bfqd->rqs_injected = bfqd->wait_dispatch = false;
4334	bfqd->waited_rq = NULL;
4335
4336	/*
4337	 * Increase, decrease or leave budget unchanged according to
4338	 * reason.
4339	 */
4340	__bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
4341	if (__bfq_bfqq_expire(bfqd, bfqq, reason))
4342		/* bfqq is gone, no more actions on it */
4343		return;
4344
4345	/* mark bfqq as waiting a request only if a bic still points to it */
4346	if (!bfq_bfqq_busy(bfqq) &&
4347	    reason != BFQQE_BUDGET_TIMEOUT &&
4348	    reason != BFQQE_BUDGET_EXHAUSTED) {
4349		bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
4350		/*
4351		 * Not setting service to 0, because, if the next rq
4352		 * arrives in time, the queue will go on receiving
4353		 * service with this same budget (as if it never expired)
4354		 */
4355	} else
4356		entity->service = 0;
4357
4358	/*
4359	 * Reset the received-service counter for every parent entity.
4360	 * Differently from what happens with bfqq->entity.service,
4361	 * the resetting of this counter never needs to be postponed
4362	 * for parent entities. In fact, in case bfqq may have a
4363	 * chance to go on being served using the last, partially
4364	 * consumed budget, bfqq->entity.service needs to be kept,
4365	 * because if bfqq then actually goes on being served using
4366	 * the same budget, the last value of bfqq->entity.service is
4367	 * needed to properly decrement bfqq->entity.budget by the
4368	 * portion already consumed. In contrast, it is not necessary
4369	 * to keep entity->service for parent entities too, because
4370	 * the bubble up of the new value of bfqq->entity.budget will
4371	 * make sure that the budgets of parent entities are correct,
4372	 * even in case bfqq and thus parent entities go on receiving
4373	 * service with the same budget.
4374	 */
4375	entity = entity->parent;
4376	for_each_entity(entity)
4377		entity->service = 0;
4378}
4379
4380/*
4381 * Budget timeout is not implemented through a dedicated timer, but
4382 * just checked on request arrivals and completions, as well as on
4383 * idle timer expirations.
4384 */
4385static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
4386{
4387	return time_is_before_eq_jiffies(bfqq->budget_timeout);
4388}
4389
4390/*
4391 * If we expire a queue that is actively waiting (i.e., with the
4392 * device idled) for the arrival of a new request, then we may incur
4393 * the timestamp misalignment problem described in the body of the
4394 * function __bfq_activate_entity. Hence we return true only if this
4395 * condition does not hold, or if the queue is slow enough to deserve
4396 * only to be kicked off for preserving a high throughput.
4397 */
4398static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
4399{
4400	bfq_log_bfqq(bfqq->bfqd, bfqq,
4401		"may_budget_timeout: wait_request %d left %d timeout %d",
4402		bfq_bfqq_wait_request(bfqq),
4403			bfq_bfqq_budget_left(bfqq) >=  bfqq->entity.budget / 3,
4404		bfq_bfqq_budget_timeout(bfqq));
4405
4406	return (!bfq_bfqq_wait_request(bfqq) ||
4407		bfq_bfqq_budget_left(bfqq) >=  bfqq->entity.budget / 3)
4408		&&
4409		bfq_bfqq_budget_timeout(bfqq);
4410}
4411
4412static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
4413					     struct bfq_queue *bfqq)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4414{
 
4415	bool rot_without_queueing =
4416		!blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
4417		bfqq_sequential_and_IO_bound,
4418		idling_boosts_thr;
 
 
 
 
 
4419
4420	/* No point in idling for bfqq if it won't get requests any longer */
4421	if (unlikely(!bfqq_process_refs(bfqq)))
 
 
 
 
 
 
 
 
4422		return false;
4423
4424	bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
4425		bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
4426
4427	/*
4428	 * The next variable takes into account the cases where idling
4429	 * boosts the throughput.
4430	 *
4431	 * The value of the variable is computed considering, first, that
4432	 * idling is virtually always beneficial for the throughput if:
4433	 * (a) the device is not NCQ-capable and rotational, or
4434	 * (b) regardless of the presence of NCQ, the device is rotational and
4435	 *     the request pattern for bfqq is I/O-bound and sequential, or
4436	 * (c) regardless of whether it is rotational, the device is
4437	 *     not NCQ-capable and the request pattern for bfqq is
4438	 *     I/O-bound and sequential.
4439	 *
4440	 * Secondly, and in contrast to the above item (b), idling an
4441	 * NCQ-capable flash-based device would not boost the
4442	 * throughput even with sequential I/O; rather it would lower
4443	 * the throughput in proportion to how fast the device
4444	 * is. Accordingly, the next variable is true if any of the
4445	 * above conditions (a), (b) or (c) is true, and, in
4446	 * particular, happens to be false if bfqd is an NCQ-capable
4447	 * flash-based device.
4448	 */
4449	idling_boosts_thr = rot_without_queueing ||
4450		((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
4451		 bfqq_sequential_and_IO_bound);
4452
4453	/*
4454	 * The return value of this function is equal to that of
 
4455	 * idling_boosts_thr, unless a special case holds. In this
4456	 * special case, described below, idling may cause problems to
4457	 * weight-raised queues.
4458	 *
4459	 * When the request pool is saturated (e.g., in the presence
4460	 * of write hogs), if the processes associated with
4461	 * non-weight-raised queues ask for requests at a lower rate,
4462	 * then processes associated with weight-raised queues have a
4463	 * higher probability to get a request from the pool
4464	 * immediately (or at least soon) when they need one. Thus
4465	 * they have a higher probability to actually get a fraction
4466	 * of the device throughput proportional to their high
4467	 * weight. This is especially true with NCQ-capable drives,
4468	 * which enqueue several requests in advance, and further
4469	 * reorder internally-queued requests.
4470	 *
4471	 * For this reason, we force to false the return value if
4472	 * there are weight-raised busy queues. In this case, and if
4473	 * bfqq is not weight-raised, this guarantees that the device
4474	 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4475	 * then idling will be guaranteed by another variable, see
4476	 * below). Combined with the timestamping rules of BFQ (see
4477	 * [1] for details), this behavior causes bfqq, and hence any
4478	 * sync non-weight-raised queue, to get a lower number of
4479	 * requests served, and thus to ask for a lower number of
4480	 * requests from the request pool, before the busy
4481	 * weight-raised queues get served again. This often mitigates
4482	 * starvation problems in the presence of heavy write
4483	 * workloads and NCQ, thereby guaranteeing a higher
4484	 * application and system responsiveness in these hostile
4485	 * scenarios.
4486	 */
4487	return idling_boosts_thr &&
4488		bfqd->wr_busy_queues == 0;
4489}
4490
4491/*
4492 * For a queue that becomes empty, device idling is allowed only if
4493 * this function returns true for that queue. As a consequence, since
4494 * device idling plays a critical role for both throughput boosting
4495 * and service guarantees, the return value of this function plays a
4496 * critical role as well.
4497 *
4498 * In a nutshell, this function returns true only if idling is
4499 * beneficial for throughput or, even if detrimental for throughput,
4500 * idling is however necessary to preserve service guarantees (low
4501 * latency, desired throughput distribution, ...). In particular, on
4502 * NCQ-capable devices, this function tries to return false, so as to
4503 * help keep the drives' internal queues full, whenever this helps the
4504 * device boost the throughput without causing any service-guarantee
4505 * issue.
4506 *
4507 * Most of the issues taken into account to get the return value of
4508 * this function are not trivial. We discuss these issues in the two
4509 * functions providing the main pieces of information needed by this
4510 * function.
4511 */
4512static bool bfq_better_to_idle(struct bfq_queue *bfqq)
4513{
4514	struct bfq_data *bfqd = bfqq->bfqd;
4515	bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
4516
4517	/* No point in idling for bfqq if it won't get requests any longer */
4518	if (unlikely(!bfqq_process_refs(bfqq)))
4519		return false;
4520
4521	if (unlikely(bfqd->strict_guarantees))
4522		return true;
4523
4524	/*
4525	 * Idling is performed only if slice_idle > 0. In addition, we
4526	 * do not idle if
4527	 * (a) bfqq is async
4528	 * (b) bfqq is in the idle io prio class: in this case we do
4529	 * not idle because we want to minimize the bandwidth that
4530	 * queues in this class can steal to higher-priority queues
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4531	 */
4532	if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
4533	   bfq_class_idle(bfqq))
4534		return false;
4535
4536	idling_boosts_thr_with_no_issue =
4537		idling_boosts_thr_without_issues(bfqd, bfqq);
4538
4539	idling_needed_for_service_guar =
4540		idling_needed_for_service_guarantees(bfqd, bfqq);
4541
4542	/*
4543	 * We have now the two components we need to compute the
4544	 * return value of the function, which is true only if idling
4545	 * either boosts the throughput (without issues), or is
4546	 * necessary to preserve service guarantees.
4547	 */
4548	return idling_boosts_thr_with_no_issue ||
4549		idling_needed_for_service_guar;
4550}
4551
4552/*
4553 * If the in-service queue is empty but the function bfq_better_to_idle
4554 * returns true, then:
4555 * 1) the queue must remain in service and cannot be expired, and
4556 * 2) the device must be idled to wait for the possible arrival of a new
4557 *    request for the queue.
4558 * See the comments on the function bfq_better_to_idle for the reasons
4559 * why performing device idling is the best choice to boost the throughput
4560 * and preserve service guarantees when bfq_better_to_idle itself
4561 * returns true.
4562 */
4563static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
4564{
4565	return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
4566}
4567
4568/*
4569 * This function chooses the queue from which to pick the next extra
4570 * I/O request to inject, if it finds a compatible queue. See the
4571 * comments on bfq_update_inject_limit() for details on the injection
4572 * mechanism, and for the definitions of the quantities mentioned
4573 * below.
4574 */
4575static struct bfq_queue *
4576bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
4577{
4578	struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
4579	unsigned int limit = in_serv_bfqq->inject_limit;
4580	/*
4581	 * If
4582	 * - bfqq is not weight-raised and therefore does not carry
4583	 *   time-critical I/O,
4584	 * or
4585	 * - regardless of whether bfqq is weight-raised, bfqq has
4586	 *   however a long think time, during which it can absorb the
4587	 *   effect of an appropriate number of extra I/O requests
4588	 *   from other queues (see bfq_update_inject_limit for
4589	 *   details on the computation of this number);
4590	 * then injection can be performed without restrictions.
4591	 */
4592	bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
4593		!bfq_bfqq_has_short_ttime(in_serv_bfqq);
4594
4595	/*
4596	 * If
4597	 * - the baseline total service time could not be sampled yet,
4598	 *   so the inject limit happens to be still 0, and
4599	 * - a lot of time has elapsed since the plugging of I/O
4600	 *   dispatching started, so drive speed is being wasted
4601	 *   significantly;
4602	 * then temporarily raise inject limit to one request.
4603	 */
4604	if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
4605	    bfq_bfqq_wait_request(in_serv_bfqq) &&
4606	    time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
4607				      bfqd->bfq_slice_idle)
4608		)
4609		limit = 1;
4610
4611	if (bfqd->rq_in_driver >= limit)
4612		return NULL;
4613
4614	/*
4615	 * Linear search of the source queue for injection; but, with
4616	 * a high probability, very few steps are needed to find a
4617	 * candidate queue, i.e., a queue with enough budget left for
4618	 * its next request. In fact:
4619	 * - BFQ dynamically updates the budget of every queue so as
4620	 *   to accommodate the expected backlog of the queue;
4621	 * - if a queue gets all its requests dispatched as injected
4622	 *   service, then the queue is removed from the active list
4623	 *   (and re-added only if it gets new requests, but then it
4624	 *   is assigned again enough budget for its new backlog).
4625	 */
4626	list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
4627		if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4628		    (in_serv_always_inject || bfqq->wr_coeff > 1) &&
4629		    bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4630		    bfq_bfqq_budget_left(bfqq)) {
4631			/*
4632			 * Allow for only one large in-flight request
4633			 * on non-rotational devices, for the
4634			 * following reason. On non-rotationl drives,
4635			 * large requests take much longer than
4636			 * smaller requests to be served. In addition,
4637			 * the drive prefers to serve large requests
4638			 * w.r.t. to small ones, if it can choose. So,
4639			 * having more than one large requests queued
4640			 * in the drive may easily make the next first
4641			 * request of the in-service queue wait for so
4642			 * long to break bfqq's service guarantees. On
4643			 * the bright side, large requests let the
4644			 * drive reach a very high throughput, even if
4645			 * there is only one in-flight large request
4646			 * at a time.
4647			 */
4648			if (blk_queue_nonrot(bfqd->queue) &&
4649			    blk_rq_sectors(bfqq->next_rq) >=
4650			    BFQQ_SECT_THR_NONROT)
4651				limit = min_t(unsigned int, 1, limit);
4652			else
4653				limit = in_serv_bfqq->inject_limit;
4654
4655			if (bfqd->rq_in_driver < limit) {
4656				bfqd->rqs_injected = true;
4657				return bfqq;
4658			}
4659		}
4660
4661	return NULL;
4662}
4663
4664/*
4665 * Select a queue for service.  If we have a current queue in service,
4666 * check whether to continue servicing it, or retrieve and set a new one.
4667 */
4668static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
4669{
4670	struct bfq_queue *bfqq;
4671	struct request *next_rq;
4672	enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
4673
4674	bfqq = bfqd->in_service_queue;
4675	if (!bfqq)
4676		goto new_queue;
4677
4678	bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
4679
4680	/*
4681	 * Do not expire bfqq for budget timeout if bfqq may be about
4682	 * to enjoy device idling. The reason why, in this case, we
4683	 * prevent bfqq from expiring is the same as in the comments
4684	 * on the case where bfq_bfqq_must_idle() returns true, in
4685	 * bfq_completed_request().
4686	 */
4687	if (bfq_may_expire_for_budg_timeout(bfqq) &&
 
4688	    !bfq_bfqq_must_idle(bfqq))
4689		goto expire;
4690
4691check_queue:
4692	/*
4693	 * This loop is rarely executed more than once. Even when it
4694	 * happens, it is much more convenient to re-execute this loop
4695	 * than to return NULL and trigger a new dispatch to get a
4696	 * request served.
4697	 */
4698	next_rq = bfqq->next_rq;
4699	/*
4700	 * If bfqq has requests queued and it has enough budget left to
4701	 * serve them, keep the queue, otherwise expire it.
4702	 */
4703	if (next_rq) {
4704		if (bfq_serv_to_charge(next_rq, bfqq) >
4705			bfq_bfqq_budget_left(bfqq)) {
4706			/*
4707			 * Expire the queue for budget exhaustion,
4708			 * which makes sure that the next budget is
4709			 * enough to serve the next request, even if
4710			 * it comes from the fifo expired path.
4711			 */
4712			reason = BFQQE_BUDGET_EXHAUSTED;
4713			goto expire;
4714		} else {
4715			/*
4716			 * The idle timer may be pending because we may
4717			 * not disable disk idling even when a new request
4718			 * arrives.
4719			 */
4720			if (bfq_bfqq_wait_request(bfqq)) {
4721				/*
4722				 * If we get here: 1) at least a new request
4723				 * has arrived but we have not disabled the
4724				 * timer because the request was too small,
4725				 * 2) then the block layer has unplugged
4726				 * the device, causing the dispatch to be
4727				 * invoked.
4728				 *
4729				 * Since the device is unplugged, now the
4730				 * requests are probably large enough to
4731				 * provide a reasonable throughput.
4732				 * So we disable idling.
4733				 */
4734				bfq_clear_bfqq_wait_request(bfqq);
4735				hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4736			}
4737			goto keep_queue;
4738		}
4739	}
4740
4741	/*
4742	 * No requests pending. However, if the in-service queue is idling
4743	 * for a new request, or has requests waiting for a completion and
4744	 * may idle after their completion, then keep it anyway.
4745	 *
4746	 * Yet, inject service from other queues if it boosts
4747	 * throughput and is possible.
4748	 */
4749	if (bfq_bfqq_wait_request(bfqq) ||
4750	    (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
4751		struct bfq_queue *async_bfqq =
4752			bfqq->bic && bfqq->bic->bfqq[0] &&
4753			bfq_bfqq_busy(bfqq->bic->bfqq[0]) &&
4754			bfqq->bic->bfqq[0]->next_rq ?
4755			bfqq->bic->bfqq[0] : NULL;
4756		struct bfq_queue *blocked_bfqq =
4757			!hlist_empty(&bfqq->woken_list) ?
4758			container_of(bfqq->woken_list.first,
4759				     struct bfq_queue,
4760				     woken_list_node)
4761			: NULL;
4762
4763		/*
4764		 * The next four mutually-exclusive ifs decide
4765		 * whether to try injection, and choose the queue to
4766		 * pick an I/O request from.
4767		 *
4768		 * The first if checks whether the process associated
4769		 * with bfqq has also async I/O pending. If so, it
4770		 * injects such I/O unconditionally. Injecting async
4771		 * I/O from the same process can cause no harm to the
4772		 * process. On the contrary, it can only increase
4773		 * bandwidth and reduce latency for the process.
4774		 *
4775		 * The second if checks whether there happens to be a
4776		 * non-empty waker queue for bfqq, i.e., a queue whose
4777		 * I/O needs to be completed for bfqq to receive new
4778		 * I/O. This happens, e.g., if bfqq is associated with
4779		 * a process that does some sync. A sync generates
4780		 * extra blocking I/O, which must be completed before
4781		 * the process associated with bfqq can go on with its
4782		 * I/O. If the I/O of the waker queue is not served,
4783		 * then bfqq remains empty, and no I/O is dispatched,
4784		 * until the idle timeout fires for bfqq. This is
4785		 * likely to result in lower bandwidth and higher
4786		 * latencies for bfqq, and in a severe loss of total
4787		 * throughput. The best action to take is therefore to
4788		 * serve the waker queue as soon as possible. So do it
4789		 * (without relying on the third alternative below for
4790		 * eventually serving waker_bfqq's I/O; see the last
4791		 * paragraph for further details). This systematic
4792		 * injection of I/O from the waker queue does not
4793		 * cause any delay to bfqq's I/O. On the contrary,
4794		 * next bfqq's I/O is brought forward dramatically,
4795		 * for it is not blocked for milliseconds.
4796		 *
4797		 * The third if checks whether there is a queue woken
4798		 * by bfqq, and currently with pending I/O. Such a
4799		 * woken queue does not steal bandwidth from bfqq,
4800		 * because it remains soon without I/O if bfqq is not
4801		 * served. So there is virtually no risk of loss of
4802		 * bandwidth for bfqq if this woken queue has I/O
4803		 * dispatched while bfqq is waiting for new I/O.
4804		 *
4805		 * The fourth if checks whether bfqq is a queue for
4806		 * which it is better to avoid injection. It is so if
4807		 * bfqq delivers more throughput when served without
4808		 * any further I/O from other queues in the middle, or
4809		 * if the service times of bfqq's I/O requests both
4810		 * count more than overall throughput, and may be
4811		 * easily increased by injection (this happens if bfqq
4812		 * has a short think time). If none of these
4813		 * conditions holds, then a candidate queue for
4814		 * injection is looked for through
4815		 * bfq_choose_bfqq_for_injection(). Note that the
4816		 * latter may return NULL (for example if the inject
4817		 * limit for bfqq is currently 0).
4818		 *
4819		 * NOTE: motivation for the second alternative
4820		 *
4821		 * Thanks to the way the inject limit is updated in
4822		 * bfq_update_has_short_ttime(), it is rather likely
4823		 * that, if I/O is being plugged for bfqq and the
4824		 * waker queue has pending I/O requests that are
4825		 * blocking bfqq's I/O, then the fourth alternative
4826		 * above lets the waker queue get served before the
4827		 * I/O-plugging timeout fires. So one may deem the
4828		 * second alternative superfluous. It is not, because
4829		 * the fourth alternative may be way less effective in
4830		 * case of a synchronization. For two main
4831		 * reasons. First, throughput may be low because the
4832		 * inject limit may be too low to guarantee the same
4833		 * amount of injected I/O, from the waker queue or
4834		 * other queues, that the second alternative
4835		 * guarantees (the second alternative unconditionally
4836		 * injects a pending I/O request of the waker queue
4837		 * for each bfq_dispatch_request()). Second, with the
4838		 * fourth alternative, the duration of the plugging,
4839		 * i.e., the time before bfqq finally receives new I/O,
4840		 * may not be minimized, because the waker queue may
4841		 * happen to be served only after other queues.
4842		 */
4843		if (async_bfqq &&
4844		    icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
4845		    bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
4846		    bfq_bfqq_budget_left(async_bfqq))
4847			bfqq = bfqq->bic->bfqq[0];
4848		else if (bfqq->waker_bfqq &&
4849			   bfq_bfqq_busy(bfqq->waker_bfqq) &&
4850			   bfqq->waker_bfqq->next_rq &&
4851			   bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
4852					      bfqq->waker_bfqq) <=
4853			   bfq_bfqq_budget_left(bfqq->waker_bfqq)
4854			)
4855			bfqq = bfqq->waker_bfqq;
4856		else if (blocked_bfqq &&
4857			   bfq_bfqq_busy(blocked_bfqq) &&
4858			   blocked_bfqq->next_rq &&
4859			   bfq_serv_to_charge(blocked_bfqq->next_rq,
4860					      blocked_bfqq) <=
4861			   bfq_bfqq_budget_left(blocked_bfqq)
4862			)
4863			bfqq = blocked_bfqq;
4864		else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
4865			 (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
4866			  !bfq_bfqq_has_short_ttime(bfqq)))
4867			bfqq = bfq_choose_bfqq_for_injection(bfqd);
4868		else
4869			bfqq = NULL;
4870
4871		goto keep_queue;
4872	}
4873
4874	reason = BFQQE_NO_MORE_REQUESTS;
4875expire:
4876	bfq_bfqq_expire(bfqd, bfqq, false, reason);
4877new_queue:
4878	bfqq = bfq_set_in_service_queue(bfqd);
4879	if (bfqq) {
4880		bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
4881		goto check_queue;
4882	}
4883keep_queue:
4884	if (bfqq)
4885		bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
4886	else
4887		bfq_log(bfqd, "select_queue: no queue returned");
4888
4889	return bfqq;
4890}
4891
4892static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4893{
4894	struct bfq_entity *entity = &bfqq->entity;
4895
4896	if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
4897		bfq_log_bfqq(bfqd, bfqq,
4898			"raising period dur %u/%u msec, old coeff %u, w %d(%d)",
4899			jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
4900			jiffies_to_msecs(bfqq->wr_cur_max_time),
4901			bfqq->wr_coeff,
4902			bfqq->entity.weight, bfqq->entity.orig_weight);
4903
4904		if (entity->prio_changed)
4905			bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
4906
4907		/*
4908		 * If the queue was activated in a burst, or too much
4909		 * time has elapsed from the beginning of this
4910		 * weight-raising period, then end weight raising.
4911		 */
4912		if (bfq_bfqq_in_large_burst(bfqq))
4913			bfq_bfqq_end_wr(bfqq);
4914		else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
4915						bfqq->wr_cur_max_time)) {
4916			if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
4917			time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
4918					       bfq_wr_duration(bfqd))) {
4919				/*
4920				 * Either in interactive weight
4921				 * raising, or in soft_rt weight
4922				 * raising with the
4923				 * interactive-weight-raising period
4924				 * elapsed (so no switch back to
4925				 * interactive weight raising).
4926				 */
4927				bfq_bfqq_end_wr(bfqq);
4928			} else { /*
4929				  * soft_rt finishing while still in
4930				  * interactive period, switch back to
4931				  * interactive weight raising
4932				  */
4933				switch_back_to_interactive_wr(bfqq, bfqd);
4934				bfqq->entity.prio_changed = 1;
4935			}
4936		}
4937		if (bfqq->wr_coeff > 1 &&
4938		    bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
4939		    bfqq->service_from_wr > max_service_from_wr) {
4940			/* see comments on max_service_from_wr */
4941			bfq_bfqq_end_wr(bfqq);
4942		}
4943	}
4944	/*
4945	 * To improve latency (for this or other queues), immediately
4946	 * update weight both if it must be raised and if it must be
4947	 * lowered. Since, entity may be on some active tree here, and
4948	 * might have a pending change of its ioprio class, invoke
4949	 * next function with the last parameter unset (see the
4950	 * comments on the function).
4951	 */
4952	if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
4953		__bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
4954						entity, false);
4955}
4956
4957/*
4958 * Dispatch next request from bfqq.
4959 */
4960static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
4961						 struct bfq_queue *bfqq)
4962{
4963	struct request *rq = bfqq->next_rq;
4964	unsigned long service_to_charge;
4965
4966	service_to_charge = bfq_serv_to_charge(rq, bfqq);
4967
4968	bfq_bfqq_served(bfqq, service_to_charge);
4969
4970	if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
4971		bfqd->wait_dispatch = false;
4972		bfqd->waited_rq = rq;
4973	}
4974
4975	bfq_dispatch_remove(bfqd->queue, rq);
4976
4977	if (bfqq != bfqd->in_service_queue)
4978		goto return_rq;
4979
4980	/*
4981	 * If weight raising has to terminate for bfqq, then next
4982	 * function causes an immediate update of bfqq's weight,
4983	 * without waiting for next activation. As a consequence, on
4984	 * expiration, bfqq will be timestamped as if has never been
4985	 * weight-raised during this service slot, even if it has
4986	 * received part or even most of the service as a
4987	 * weight-raised queue. This inflates bfqq's timestamps, which
4988	 * is beneficial, as bfqq is then more willing to leave the
4989	 * device immediately to possible other weight-raised queues.
4990	 */
4991	bfq_update_wr_data(bfqd, bfqq);
4992
4993	/*
4994	 * Expire bfqq, pretending that its budget expired, if bfqq
4995	 * belongs to CLASS_IDLE and other queues are waiting for
4996	 * service.
4997	 */
4998	if (!(bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq)))
4999		goto return_rq;
5000
5001	bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
5002
5003return_rq:
 
5004	return rq;
5005}
5006
5007static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
5008{
5009	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5010
5011	/*
5012	 * Avoiding lock: a race on bfqd->queued should cause at
5013	 * most a call to dispatch for nothing
5014	 */
5015	return !list_empty_careful(&bfqd->dispatch) ||
5016		READ_ONCE(bfqd->queued);
5017}
5018
5019static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5020{
5021	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5022	struct request *rq = NULL;
5023	struct bfq_queue *bfqq = NULL;
5024
5025	if (!list_empty(&bfqd->dispatch)) {
5026		rq = list_first_entry(&bfqd->dispatch, struct request,
5027				      queuelist);
5028		list_del_init(&rq->queuelist);
5029
5030		bfqq = RQ_BFQQ(rq);
5031
5032		if (bfqq) {
5033			/*
5034			 * Increment counters here, because this
5035			 * dispatch does not follow the standard
5036			 * dispatch flow (where counters are
5037			 * incremented)
5038			 */
5039			bfqq->dispatched++;
5040
5041			goto inc_in_driver_start_rq;
5042		}
5043
5044		/*
5045		 * We exploit the bfq_finish_requeue_request hook to
5046		 * decrement rq_in_driver, but
5047		 * bfq_finish_requeue_request will not be invoked on
5048		 * this request. So, to avoid unbalance, just start
5049		 * this request, without incrementing rq_in_driver. As
5050		 * a negative consequence, rq_in_driver is deceptively
5051		 * lower than it should be while this request is in
5052		 * service. This may cause bfq_schedule_dispatch to be
5053		 * invoked uselessly.
5054		 *
5055		 * As for implementing an exact solution, the
5056		 * bfq_finish_requeue_request hook, if defined, is
5057		 * probably invoked also on this request. So, by
5058		 * exploiting this hook, we could 1) increment
5059		 * rq_in_driver here, and 2) decrement it in
5060		 * bfq_finish_requeue_request. Such a solution would
5061		 * let the value of the counter be always accurate,
5062		 * but it would entail using an extra interface
5063		 * function. This cost seems higher than the benefit,
5064		 * being the frequency of non-elevator-private
5065		 * requests very low.
5066		 */
5067		goto start_rq;
5068	}
5069
5070	bfq_log(bfqd, "dispatch requests: %d busy queues",
5071		bfq_tot_busy_queues(bfqd));
5072
5073	if (bfq_tot_busy_queues(bfqd) == 0)
5074		goto exit;
5075
5076	/*
5077	 * Force device to serve one request at a time if
5078	 * strict_guarantees is true. Forcing this service scheme is
5079	 * currently the ONLY way to guarantee that the request
5080	 * service order enforced by the scheduler is respected by a
5081	 * queueing device. Otherwise the device is free even to make
5082	 * some unlucky request wait for as long as the device
5083	 * wishes.
5084	 *
5085	 * Of course, serving one request at a time may cause loss of
5086	 * throughput.
5087	 */
5088	if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
5089		goto exit;
5090
5091	bfqq = bfq_select_queue(bfqd);
5092	if (!bfqq)
5093		goto exit;
5094
5095	rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
5096
5097	if (rq) {
5098inc_in_driver_start_rq:
5099		bfqd->rq_in_driver++;
5100start_rq:
5101		rq->rq_flags |= RQF_STARTED;
5102	}
5103exit:
5104	return rq;
5105}
5106
5107#ifdef CONFIG_BFQ_CGROUP_DEBUG
5108static void bfq_update_dispatch_stats(struct request_queue *q,
5109				      struct request *rq,
5110				      struct bfq_queue *in_serv_queue,
5111				      bool idle_timer_disabled)
5112{
5113	struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
5114
5115	if (!idle_timer_disabled && !bfqq)
5116		return;
5117
5118	/*
5119	 * rq and bfqq are guaranteed to exist until this function
5120	 * ends, for the following reasons. First, rq can be
5121	 * dispatched to the device, and then can be completed and
5122	 * freed, only after this function ends. Second, rq cannot be
5123	 * merged (and thus freed because of a merge) any longer,
5124	 * because it has already started. Thus rq cannot be freed
5125	 * before this function ends, and, since rq has a reference to
5126	 * bfqq, the same guarantee holds for bfqq too.
5127	 *
5128	 * In addition, the following queue lock guarantees that
5129	 * bfqq_group(bfqq) exists as well.
5130	 */
5131	spin_lock_irq(&q->queue_lock);
5132	if (idle_timer_disabled)
5133		/*
5134		 * Since the idle timer has been disabled,
5135		 * in_serv_queue contained some request when
5136		 * __bfq_dispatch_request was invoked above, which
5137		 * implies that rq was picked exactly from
5138		 * in_serv_queue. Thus in_serv_queue == bfqq, and is
5139		 * therefore guaranteed to exist because of the above
5140		 * arguments.
5141		 */
5142		bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
5143	if (bfqq) {
5144		struct bfq_group *bfqg = bfqq_group(bfqq);
5145
5146		bfqg_stats_update_avg_queue_size(bfqg);
5147		bfqg_stats_set_start_empty_time(bfqg);
5148		bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
5149	}
5150	spin_unlock_irq(&q->queue_lock);
5151}
5152#else
5153static inline void bfq_update_dispatch_stats(struct request_queue *q,
5154					     struct request *rq,
5155					     struct bfq_queue *in_serv_queue,
5156					     bool idle_timer_disabled) {}
5157#endif /* CONFIG_BFQ_CGROUP_DEBUG */
5158
5159static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5160{
5161	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5162	struct request *rq;
5163	struct bfq_queue *in_serv_queue;
5164	bool waiting_rq, idle_timer_disabled = false;
5165
5166	spin_lock_irq(&bfqd->lock);
5167
5168	in_serv_queue = bfqd->in_service_queue;
5169	waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
5170
5171	rq = __bfq_dispatch_request(hctx);
5172	if (in_serv_queue == bfqd->in_service_queue) {
5173		idle_timer_disabled =
5174			waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
5175	}
5176
5177	spin_unlock_irq(&bfqd->lock);
5178	bfq_update_dispatch_stats(hctx->queue, rq,
5179			idle_timer_disabled ? in_serv_queue : NULL,
5180				idle_timer_disabled);
5181
5182	return rq;
5183}
5184
5185/*
5186 * Task holds one reference to the queue, dropped when task exits.  Each rq
5187 * in-flight on this queue also holds a reference, dropped when rq is freed.
5188 *
5189 * Scheduler lock must be held here. Recall not to use bfqq after calling
5190 * this function on it.
5191 */
5192void bfq_put_queue(struct bfq_queue *bfqq)
5193{
5194	struct bfq_queue *item;
5195	struct hlist_node *n;
5196	struct bfq_group *bfqg = bfqq_group(bfqq);
 
5197
5198	bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d", bfqq, bfqq->ref);
 
 
5199
5200	bfqq->ref--;
5201	if (bfqq->ref)
5202		return;
5203
5204	if (!hlist_unhashed(&bfqq->burst_list_node)) {
5205		hlist_del_init(&bfqq->burst_list_node);
5206		/*
5207		 * Decrement also burst size after the removal, if the
5208		 * process associated with bfqq is exiting, and thus
5209		 * does not contribute to the burst any longer. This
5210		 * decrement helps filter out false positives of large
5211		 * bursts, when some short-lived process (often due to
5212		 * the execution of commands by some service) happens
5213		 * to start and exit while a complex application is
5214		 * starting, and thus spawning several processes that
5215		 * do I/O (and that *must not* be treated as a large
5216		 * burst, see comments on bfq_handle_burst).
5217		 *
5218		 * In particular, the decrement is performed only if:
5219		 * 1) bfqq is not a merged queue, because, if it is,
5220		 * then this free of bfqq is not triggered by the exit
5221		 * of the process bfqq is associated with, but exactly
5222		 * by the fact that bfqq has just been merged.
5223		 * 2) burst_size is greater than 0, to handle
5224		 * unbalanced decrements. Unbalanced decrements may
5225		 * happen in te following case: bfqq is inserted into
5226		 * the current burst list--without incrementing
5227		 * bust_size--because of a split, but the current
5228		 * burst list is not the burst list bfqq belonged to
5229		 * (see comments on the case of a split in
5230		 * bfq_set_request).
5231		 */
5232		if (bfqq->bic && bfqq->bfqd->burst_size > 0)
5233			bfqq->bfqd->burst_size--;
5234	}
5235
5236	/*
5237	 * bfqq does not exist any longer, so it cannot be woken by
5238	 * any other queue, and cannot wake any other queue. Then bfqq
5239	 * must be removed from the woken list of its possible waker
5240	 * queue, and all queues in the woken list of bfqq must stop
5241	 * having a waker queue. Strictly speaking, these updates
5242	 * should be performed when bfqq remains with no I/O source
5243	 * attached to it, which happens before bfqq gets freed. In
5244	 * particular, this happens when the last process associated
5245	 * with bfqq exits or gets associated with a different
5246	 * queue. However, both events lead to bfqq being freed soon,
5247	 * and dangling references would come out only after bfqq gets
5248	 * freed. So these updates are done here, as a simple and safe
5249	 * way to handle all cases.
5250	 */
5251	/* remove bfqq from woken list */
5252	if (!hlist_unhashed(&bfqq->woken_list_node))
5253		hlist_del_init(&bfqq->woken_list_node);
5254
5255	/* reset waker for all queues in woken list */
5256	hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
5257				  woken_list_node) {
5258		item->waker_bfqq = NULL;
5259		hlist_del_init(&item->woken_list_node);
5260	}
5261
5262	if (bfqq->bfqd->last_completed_rq_bfqq == bfqq)
5263		bfqq->bfqd->last_completed_rq_bfqq = NULL;
5264
5265	kmem_cache_free(bfq_pool, bfqq);
 
5266	bfqg_and_blkg_put(bfqg);
 
5267}
5268
5269static void bfq_put_stable_ref(struct bfq_queue *bfqq)
5270{
5271	bfqq->stable_ref--;
5272	bfq_put_queue(bfqq);
5273}
5274
5275void bfq_put_cooperator(struct bfq_queue *bfqq)
5276{
5277	struct bfq_queue *__bfqq, *next;
5278
5279	/*
5280	 * If this queue was scheduled to merge with another queue, be
5281	 * sure to drop the reference taken on that queue (and others in
5282	 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
5283	 */
5284	__bfqq = bfqq->new_bfqq;
5285	while (__bfqq) {
5286		if (__bfqq == bfqq)
5287			break;
5288		next = __bfqq->new_bfqq;
5289		bfq_put_queue(__bfqq);
5290		__bfqq = next;
5291	}
5292}
5293
5294static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
5295{
5296	if (bfqq == bfqd->in_service_queue) {
5297		__bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT);
5298		bfq_schedule_dispatch(bfqd);
5299	}
5300
5301	bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
5302
5303	bfq_put_cooperator(bfqq);
5304
5305	bfq_release_process_ref(bfqd, bfqq);
5306}
5307
5308static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
5309{
5310	struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
5311	struct bfq_data *bfqd;
5312
5313	if (bfqq)
5314		bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
5315
5316	if (bfqq && bfqd) {
5317		unsigned long flags;
5318
5319		spin_lock_irqsave(&bfqd->lock, flags);
5320		bic_set_bfqq(bic, NULL, is_sync);
5321		bfq_exit_bfqq(bfqd, bfqq);
 
5322		spin_unlock_irqrestore(&bfqd->lock, flags);
5323	}
5324}
5325
5326static void bfq_exit_icq(struct io_cq *icq)
5327{
5328	struct bfq_io_cq *bic = icq_to_bic(icq);
5329
5330	if (bic->stable_merge_bfqq) {
5331		struct bfq_data *bfqd = bic->stable_merge_bfqq->bfqd;
5332
5333		/*
5334		 * bfqd is NULL if scheduler already exited, and in
5335		 * that case this is the last time bfqq is accessed.
5336		 */
5337		if (bfqd) {
5338			unsigned long flags;
5339
5340			spin_lock_irqsave(&bfqd->lock, flags);
5341			bfq_put_stable_ref(bic->stable_merge_bfqq);
5342			spin_unlock_irqrestore(&bfqd->lock, flags);
5343		} else {
5344			bfq_put_stable_ref(bic->stable_merge_bfqq);
5345		}
5346	}
5347
5348	bfq_exit_icq_bfqq(bic, true);
5349	bfq_exit_icq_bfqq(bic, false);
5350}
5351
5352/*
5353 * Update the entity prio values; note that the new values will not
5354 * be used until the next (re)activation.
5355 */
5356static void
5357bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
5358{
5359	struct task_struct *tsk = current;
5360	int ioprio_class;
5361	struct bfq_data *bfqd = bfqq->bfqd;
5362
5363	if (!bfqd)
5364		return;
5365
5366	ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5367	switch (ioprio_class) {
5368	default:
5369		pr_err("bdi %s: bfq: bad prio class %d\n",
5370			bdi_dev_name(bfqq->bfqd->queue->disk->bdi),
5371			ioprio_class);
5372		fallthrough;
5373	case IOPRIO_CLASS_NONE:
5374		/*
5375		 * No prio set, inherit CPU scheduling settings.
5376		 */
5377		bfqq->new_ioprio = task_nice_ioprio(tsk);
5378		bfqq->new_ioprio_class = task_nice_ioclass(tsk);
5379		break;
5380	case IOPRIO_CLASS_RT:
5381		bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5382		bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
5383		break;
5384	case IOPRIO_CLASS_BE:
5385		bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5386		bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
5387		break;
5388	case IOPRIO_CLASS_IDLE:
5389		bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
5390		bfqq->new_ioprio = 7;
5391		break;
5392	}
5393
5394	if (bfqq->new_ioprio >= IOPRIO_NR_LEVELS) {
5395		pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
5396			bfqq->new_ioprio);
5397		bfqq->new_ioprio = IOPRIO_NR_LEVELS - 1;
5398	}
5399
5400	bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
5401	bfq_log_bfqq(bfqd, bfqq, "new_ioprio %d new_weight %d",
5402		     bfqq->new_ioprio, bfqq->entity.new_weight);
5403	bfqq->entity.prio_changed = 1;
5404}
5405
5406static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5407				       struct bio *bio, bool is_sync,
5408				       struct bfq_io_cq *bic,
5409				       bool respawn);
5410
5411static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
5412{
5413	struct bfq_data *bfqd = bic_to_bfqd(bic);
5414	struct bfq_queue *bfqq;
5415	int ioprio = bic->icq.ioc->ioprio;
5416
5417	/*
5418	 * This condition may trigger on a newly created bic, be sure to
5419	 * drop the lock before returning.
5420	 */
5421	if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
5422		return;
5423
5424	bic->ioprio = ioprio;
5425
5426	bfqq = bic_to_bfqq(bic, false);
5427	if (bfqq) {
5428		struct bfq_queue *old_bfqq = bfqq;
5429
5430		bfqq = bfq_get_queue(bfqd, bio, false, bic, true);
5431		bic_set_bfqq(bic, bfqq, false);
5432		bfq_release_process_ref(bfqd, old_bfqq);
5433	}
5434
5435	bfqq = bic_to_bfqq(bic, true);
5436	if (bfqq)
5437		bfq_set_next_ioprio_data(bfqq, bic);
5438}
5439
5440static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5441			  struct bfq_io_cq *bic, pid_t pid, int is_sync)
5442{
5443	u64 now_ns = ktime_get_ns();
5444
5445	RB_CLEAR_NODE(&bfqq->entity.rb_node);
5446	INIT_LIST_HEAD(&bfqq->fifo);
5447	INIT_HLIST_NODE(&bfqq->burst_list_node);
5448	INIT_HLIST_NODE(&bfqq->woken_list_node);
5449	INIT_HLIST_HEAD(&bfqq->woken_list);
5450
5451	bfqq->ref = 0;
5452	bfqq->bfqd = bfqd;
5453
5454	if (bic)
5455		bfq_set_next_ioprio_data(bfqq, bic);
5456
5457	if (is_sync) {
5458		/*
5459		 * No need to mark as has_short_ttime if in
5460		 * idle_class, because no device idling is performed
5461		 * for queues in idle class
5462		 */
5463		if (!bfq_class_idle(bfqq))
5464			/* tentatively mark as has_short_ttime */
5465			bfq_mark_bfqq_has_short_ttime(bfqq);
5466		bfq_mark_bfqq_sync(bfqq);
5467		bfq_mark_bfqq_just_created(bfqq);
5468	} else
5469		bfq_clear_bfqq_sync(bfqq);
5470
5471	/* set end request to minus infinity from now */
5472	bfqq->ttime.last_end_request = now_ns + 1;
5473
5474	bfqq->creation_time = jiffies;
5475
5476	bfqq->io_start_time = now_ns;
5477
5478	bfq_mark_bfqq_IO_bound(bfqq);
5479
5480	bfqq->pid = pid;
5481
5482	/* Tentative initial value to trade off between thr and lat */
5483	bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
5484	bfqq->budget_timeout = bfq_smallest_from_now();
5485
5486	bfqq->wr_coeff = 1;
5487	bfqq->last_wr_start_finish = jiffies;
5488	bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
5489	bfqq->split_time = bfq_smallest_from_now();
5490
5491	/*
5492	 * To not forget the possibly high bandwidth consumed by a
5493	 * process/queue in the recent past,
5494	 * bfq_bfqq_softrt_next_start() returns a value at least equal
5495	 * to the current value of bfqq->soft_rt_next_start (see
5496	 * comments on bfq_bfqq_softrt_next_start).  Set
5497	 * soft_rt_next_start to now, to mean that bfqq has consumed
5498	 * no bandwidth so far.
5499	 */
5500	bfqq->soft_rt_next_start = jiffies;
5501
5502	/* first request is almost certainly seeky */
5503	bfqq->seek_history = 1;
5504}
5505
5506static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
5507					       struct bfq_group *bfqg,
5508					       int ioprio_class, int ioprio)
5509{
5510	switch (ioprio_class) {
5511	case IOPRIO_CLASS_RT:
5512		return &bfqg->async_bfqq[0][ioprio];
5513	case IOPRIO_CLASS_NONE:
5514		ioprio = IOPRIO_BE_NORM;
5515		fallthrough;
5516	case IOPRIO_CLASS_BE:
5517		return &bfqg->async_bfqq[1][ioprio];
5518	case IOPRIO_CLASS_IDLE:
5519		return &bfqg->async_idle_bfqq;
5520	default:
5521		return NULL;
5522	}
5523}
5524
5525static struct bfq_queue *
5526bfq_do_early_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5527			  struct bfq_io_cq *bic,
5528			  struct bfq_queue *last_bfqq_created)
5529{
5530	struct bfq_queue *new_bfqq =
5531		bfq_setup_merge(bfqq, last_bfqq_created);
5532
5533	if (!new_bfqq)
5534		return bfqq;
5535
5536	if (new_bfqq->bic)
5537		new_bfqq->bic->stably_merged = true;
5538	bic->stably_merged = true;
5539
5540	/*
5541	 * Reusing merge functions. This implies that
5542	 * bfqq->bic must be set too, for
5543	 * bfq_merge_bfqqs to correctly save bfqq's
5544	 * state before killing it.
5545	 */
5546	bfqq->bic = bic;
5547	bfq_merge_bfqqs(bfqd, bic, bfqq, new_bfqq);
5548
5549	return new_bfqq;
5550}
5551
5552/*
5553 * Many throughput-sensitive workloads are made of several parallel
5554 * I/O flows, with all flows generated by the same application, or
5555 * more generically by the same task (e.g., system boot). The most
5556 * counterproductive action with these workloads is plugging I/O
5557 * dispatch when one of the bfq_queues associated with these flows
5558 * remains temporarily empty.
5559 *
5560 * To avoid this plugging, BFQ has been using a burst-handling
5561 * mechanism for years now. This mechanism has proven effective for
5562 * throughput, and not detrimental for service guarantees. The
5563 * following function pushes this mechanism a little bit further,
5564 * basing on the following two facts.
5565 *
5566 * First, all the I/O flows of a the same application or task
5567 * contribute to the execution/completion of that common application
5568 * or task. So the performance figures that matter are total
5569 * throughput of the flows and task-wide I/O latency.  In particular,
5570 * these flows do not need to be protected from each other, in terms
5571 * of individual bandwidth or latency.
5572 *
5573 * Second, the above fact holds regardless of the number of flows.
5574 *
5575 * Putting these two facts together, this commits merges stably the
5576 * bfq_queues associated with these I/O flows, i.e., with the
5577 * processes that generate these IO/ flows, regardless of how many the
5578 * involved processes are.
5579 *
5580 * To decide whether a set of bfq_queues is actually associated with
5581 * the I/O flows of a common application or task, and to merge these
5582 * queues stably, this function operates as follows: given a bfq_queue,
5583 * say Q2, currently being created, and the last bfq_queue, say Q1,
5584 * created before Q2, Q2 is merged stably with Q1 if
5585 * - very little time has elapsed since when Q1 was created
5586 * - Q2 has the same ioprio as Q1
5587 * - Q2 belongs to the same group as Q1
5588 *
5589 * Merging bfq_queues also reduces scheduling overhead. A fio test
5590 * with ten random readers on /dev/nullb shows a throughput boost of
5591 * 40%, with a quadcore. Since BFQ's execution time amounts to ~50% of
5592 * the total per-request processing time, the above throughput boost
5593 * implies that BFQ's overhead is reduced by more than 50%.
5594 *
5595 * This new mechanism most certainly obsoletes the current
5596 * burst-handling heuristics. We keep those heuristics for the moment.
5597 */
5598static struct bfq_queue *bfq_do_or_sched_stable_merge(struct bfq_data *bfqd,
5599						      struct bfq_queue *bfqq,
5600						      struct bfq_io_cq *bic)
5601{
5602	struct bfq_queue **source_bfqq = bfqq->entity.parent ?
5603		&bfqq->entity.parent->last_bfqq_created :
5604		&bfqd->last_bfqq_created;
5605
5606	struct bfq_queue *last_bfqq_created = *source_bfqq;
5607
5608	/*
5609	 * If last_bfqq_created has not been set yet, then init it. If
5610	 * it has been set already, but too long ago, then move it
5611	 * forward to bfqq. Finally, move also if bfqq belongs to a
5612	 * different group than last_bfqq_created, or if bfqq has a
5613	 * different ioprio or ioprio_class. If none of these
5614	 * conditions holds true, then try an early stable merge or
5615	 * schedule a delayed stable merge.
5616	 *
5617	 * A delayed merge is scheduled (instead of performing an
5618	 * early merge), in case bfqq might soon prove to be more
5619	 * throughput-beneficial if not merged. Currently this is
5620	 * possible only if bfqd is rotational with no queueing. For
5621	 * such a drive, not merging bfqq is better for throughput if
5622	 * bfqq happens to contain sequential I/O. So, we wait a
5623	 * little bit for enough I/O to flow through bfqq. After that,
5624	 * if such an I/O is sequential, then the merge is
5625	 * canceled. Otherwise the merge is finally performed.
5626	 */
5627	if (!last_bfqq_created ||
5628	    time_before(last_bfqq_created->creation_time +
5629			msecs_to_jiffies(bfq_activation_stable_merging),
5630			bfqq->creation_time) ||
5631		bfqq->entity.parent != last_bfqq_created->entity.parent ||
5632		bfqq->ioprio != last_bfqq_created->ioprio ||
5633		bfqq->ioprio_class != last_bfqq_created->ioprio_class)
5634		*source_bfqq = bfqq;
5635	else if (time_after_eq(last_bfqq_created->creation_time +
5636				 bfqd->bfq_burst_interval,
5637				 bfqq->creation_time)) {
5638		if (likely(bfqd->nonrot_with_queueing))
5639			/*
5640			 * With this type of drive, leaving
5641			 * bfqq alone may provide no
5642			 * throughput benefits compared with
5643			 * merging bfqq. So merge bfqq now.
5644			 */
5645			bfqq = bfq_do_early_stable_merge(bfqd, bfqq,
5646							 bic,
5647							 last_bfqq_created);
5648		else { /* schedule tentative stable merge */
5649			/*
5650			 * get reference on last_bfqq_created,
5651			 * to prevent it from being freed,
5652			 * until we decide whether to merge
5653			 */
5654			last_bfqq_created->ref++;
5655			/*
5656			 * need to keep track of stable refs, to
5657			 * compute process refs correctly
5658			 */
5659			last_bfqq_created->stable_ref++;
5660			/*
5661			 * Record the bfqq to merge to.
5662			 */
5663			bic->stable_merge_bfqq = last_bfqq_created;
5664		}
5665	}
5666
5667	return bfqq;
5668}
5669
5670
5671static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5672				       struct bio *bio, bool is_sync,
5673				       struct bfq_io_cq *bic,
5674				       bool respawn)
5675{
5676	const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
5677	const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5678	struct bfq_queue **async_bfqq = NULL;
5679	struct bfq_queue *bfqq;
5680	struct bfq_group *bfqg;
5681
5682	bfqg = bfq_bio_bfqg(bfqd, bio);
 
 
 
 
 
 
 
5683	if (!is_sync) {
5684		async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
5685						  ioprio);
5686		bfqq = *async_bfqq;
5687		if (bfqq)
5688			goto out;
5689	}
5690
5691	bfqq = kmem_cache_alloc_node(bfq_pool,
5692				     GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
5693				     bfqd->queue->node);
5694
5695	if (bfqq) {
5696		bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
5697			      is_sync);
5698		bfq_init_entity(&bfqq->entity, bfqg);
5699		bfq_log_bfqq(bfqd, bfqq, "allocated");
5700	} else {
5701		bfqq = &bfqd->oom_bfqq;
5702		bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
5703		goto out;
5704	}
5705
5706	/*
5707	 * Pin the queue now that it's allocated, scheduler exit will
5708	 * prune it.
5709	 */
5710	if (async_bfqq) {
5711		bfqq->ref++; /*
5712			      * Extra group reference, w.r.t. sync
5713			      * queue. This extra reference is removed
5714			      * only if bfqq->bfqg disappears, to
5715			      * guarantee that this queue is not freed
5716			      * until its group goes away.
5717			      */
5718		bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
5719			     bfqq, bfqq->ref);
5720		*async_bfqq = bfqq;
5721	}
5722
5723out:
5724	bfqq->ref++; /* get a process reference to this queue */
5725
5726	if (bfqq != &bfqd->oom_bfqq && is_sync && !respawn)
5727		bfqq = bfq_do_or_sched_stable_merge(bfqd, bfqq, bic);
5728	return bfqq;
5729}
5730
5731static void bfq_update_io_thinktime(struct bfq_data *bfqd,
5732				    struct bfq_queue *bfqq)
5733{
5734	struct bfq_ttime *ttime = &bfqq->ttime;
5735	u64 elapsed;
5736
5737	/*
5738	 * We are really interested in how long it takes for the queue to
5739	 * become busy when there is no outstanding IO for this queue. So
5740	 * ignore cases when the bfq queue has already IO queued.
5741	 */
5742	if (bfqq->dispatched || bfq_bfqq_busy(bfqq))
5743		return;
5744	elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
5745	elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
5746
5747	ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
5748	ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed,  8);
5749	ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
5750				     ttime->ttime_samples);
5751}
5752
5753static void
5754bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5755		       struct request *rq)
5756{
5757	bfqq->seek_history <<= 1;
5758	bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
5759
5760	if (bfqq->wr_coeff > 1 &&
5761	    bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
5762	    BFQQ_TOTALLY_SEEKY(bfqq)) {
5763		if (time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
5764					   bfq_wr_duration(bfqd))) {
5765			/*
5766			 * In soft_rt weight raising with the
5767			 * interactive-weight-raising period
5768			 * elapsed (so no switch back to
5769			 * interactive weight raising).
5770			 */
5771			bfq_bfqq_end_wr(bfqq);
5772		} else { /*
5773			  * stopping soft_rt weight raising
5774			  * while still in interactive period,
5775			  * switch back to interactive weight
5776			  * raising
5777			  */
5778			switch_back_to_interactive_wr(bfqq, bfqd);
5779			bfqq->entity.prio_changed = 1;
5780		}
5781	}
5782}
5783
5784static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
5785				       struct bfq_queue *bfqq,
5786				       struct bfq_io_cq *bic)
5787{
5788	bool has_short_ttime = true, state_changed;
5789
5790	/*
5791	 * No need to update has_short_ttime if bfqq is async or in
5792	 * idle io prio class, or if bfq_slice_idle is zero, because
5793	 * no device idling is performed for bfqq in this case.
5794	 */
5795	if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
5796	    bfqd->bfq_slice_idle == 0)
5797		return;
5798
5799	/* Idle window just restored, statistics are meaningless. */
5800	if (time_is_after_eq_jiffies(bfqq->split_time +
5801				     bfqd->bfq_wr_min_idle_time))
5802		return;
5803
5804	/* Think time is infinite if no process is linked to
5805	 * bfqq. Otherwise check average think time to decide whether
5806	 * to mark as has_short_ttime. To this goal, compare average
5807	 * think time with half the I/O-plugging timeout.
5808	 */
5809	if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
5810	    (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
5811	     bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle>>1))
5812		has_short_ttime = false;
5813
5814	state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
 
5815
5816	if (has_short_ttime)
5817		bfq_mark_bfqq_has_short_ttime(bfqq);
5818	else
5819		bfq_clear_bfqq_has_short_ttime(bfqq);
5820
5821	/*
5822	 * Until the base value for the total service time gets
5823	 * finally computed for bfqq, the inject limit does depend on
5824	 * the think-time state (short|long). In particular, the limit
5825	 * is 0 or 1 if the think time is deemed, respectively, as
5826	 * short or long (details in the comments in
5827	 * bfq_update_inject_limit()). Accordingly, the next
5828	 * instructions reset the inject limit if the think-time state
5829	 * has changed and the above base value is still to be
5830	 * computed.
5831	 *
5832	 * However, the reset is performed only if more than 100 ms
5833	 * have elapsed since the last update of the inject limit, or
5834	 * (inclusive) if the change is from short to long think
5835	 * time. The reason for this waiting is as follows.
5836	 *
5837	 * bfqq may have a long think time because of a
5838	 * synchronization with some other queue, i.e., because the
5839	 * I/O of some other queue may need to be completed for bfqq
5840	 * to receive new I/O. Details in the comments on the choice
5841	 * of the queue for injection in bfq_select_queue().
5842	 *
5843	 * As stressed in those comments, if such a synchronization is
5844	 * actually in place, then, without injection on bfqq, the
5845	 * blocking I/O cannot happen to served while bfqq is in
5846	 * service. As a consequence, if bfqq is granted
5847	 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
5848	 * is dispatched, until the idle timeout fires. This is likely
5849	 * to result in lower bandwidth and higher latencies for bfqq,
5850	 * and in a severe loss of total throughput.
5851	 *
5852	 * On the opposite end, a non-zero inject limit may allow the
5853	 * I/O that blocks bfqq to be executed soon, and therefore
5854	 * bfqq to receive new I/O soon.
5855	 *
5856	 * But, if the blocking gets actually eliminated, then the
5857	 * next think-time sample for bfqq may be very low. This in
5858	 * turn may cause bfqq's think time to be deemed
5859	 * short. Without the 100 ms barrier, this new state change
5860	 * would cause the body of the next if to be executed
5861	 * immediately. But this would set to 0 the inject
5862	 * limit. Without injection, the blocking I/O would cause the
5863	 * think time of bfqq to become long again, and therefore the
5864	 * inject limit to be raised again, and so on. The only effect
5865	 * of such a steady oscillation between the two think-time
5866	 * states would be to prevent effective injection on bfqq.
5867	 *
5868	 * In contrast, if the inject limit is not reset during such a
5869	 * long time interval as 100 ms, then the number of short
5870	 * think time samples can grow significantly before the reset
5871	 * is performed. As a consequence, the think time state can
5872	 * become stable before the reset. Therefore there will be no
5873	 * state change when the 100 ms elapse, and no reset of the
5874	 * inject limit. The inject limit remains steadily equal to 1
5875	 * both during and after the 100 ms. So injection can be
5876	 * performed at all times, and throughput gets boosted.
5877	 *
5878	 * An inject limit equal to 1 is however in conflict, in
5879	 * general, with the fact that the think time of bfqq is
5880	 * short, because injection may be likely to delay bfqq's I/O
5881	 * (as explained in the comments in
5882	 * bfq_update_inject_limit()). But this does not happen in
5883	 * this special case, because bfqq's low think time is due to
5884	 * an effective handling of a synchronization, through
5885	 * injection. In this special case, bfqq's I/O does not get
5886	 * delayed by injection; on the contrary, bfqq's I/O is
5887	 * brought forward, because it is not blocked for
5888	 * milliseconds.
5889	 *
5890	 * In addition, serving the blocking I/O much sooner, and much
5891	 * more frequently than once per I/O-plugging timeout, makes
5892	 * it much quicker to detect a waker queue (the concept of
5893	 * waker queue is defined in the comments in
5894	 * bfq_add_request()). This makes it possible to start sooner
5895	 * to boost throughput more effectively, by injecting the I/O
5896	 * of the waker queue unconditionally on every
5897	 * bfq_dispatch_request().
5898	 *
5899	 * One last, important benefit of not resetting the inject
5900	 * limit before 100 ms is that, during this time interval, the
5901	 * base value for the total service time is likely to get
5902	 * finally computed for bfqq, freeing the inject limit from
5903	 * its relation with the think time.
5904	 */
5905	if (state_changed && bfqq->last_serv_time_ns == 0 &&
5906	    (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
5907				      msecs_to_jiffies(100)) ||
5908	     !has_short_ttime))
5909		bfq_reset_inject_limit(bfqd, bfqq);
5910}
5911
5912/*
5913 * Called when a new fs request (rq) is added to bfqq.  Check if there's
5914 * something we should do about it.
5915 */
5916static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5917			    struct request *rq)
5918{
 
 
5919	if (rq->cmd_flags & REQ_META)
5920		bfqq->meta_pending++;
5921
 
 
 
 
 
 
 
 
5922	bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
5923
5924	if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
5925		bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
5926				 blk_rq_sectors(rq) < 32;
5927		bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
5928
5929		/*
5930		 * There is just this request queued: if
5931		 * - the request is small, and
5932		 * - we are idling to boost throughput, and
5933		 * - the queue is not to be expired,
5934		 * then just exit.
5935		 *
5936		 * In this way, if the device is being idled to wait
5937		 * for a new request from the in-service queue, we
5938		 * avoid unplugging the device and committing the
5939		 * device to serve just a small request. In contrast
5940		 * we wait for the block layer to decide when to
5941		 * unplug the device: hopefully, new requests will be
5942		 * merged to this one quickly, then the device will be
5943		 * unplugged and larger requests will be dispatched.
 
5944		 */
5945		if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
5946		    !budget_timeout)
5947			return;
5948
5949		/*
5950		 * A large enough request arrived, or idling is being
5951		 * performed to preserve service guarantees, or
5952		 * finally the queue is to be expired: in all these
5953		 * cases disk idling is to be stopped, so clear
5954		 * wait_request flag and reset timer.
5955		 */
5956		bfq_clear_bfqq_wait_request(bfqq);
5957		hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
5958
5959		/*
5960		 * The queue is not empty, because a new request just
5961		 * arrived. Hence we can safely expire the queue, in
5962		 * case of budget timeout, without risking that the
5963		 * timestamps of the queue are not updated correctly.
5964		 * See [1] for more details.
5965		 */
5966		if (budget_timeout)
5967			bfq_bfqq_expire(bfqd, bfqq, false,
5968					BFQQE_BUDGET_TIMEOUT);
5969	}
5970}
5971
5972static void bfqq_request_allocated(struct bfq_queue *bfqq)
5973{
5974	struct bfq_entity *entity = &bfqq->entity;
5975
5976	for_each_entity(entity)
5977		entity->allocated++;
5978}
5979
5980static void bfqq_request_freed(struct bfq_queue *bfqq)
5981{
5982	struct bfq_entity *entity = &bfqq->entity;
5983
5984	for_each_entity(entity)
5985		entity->allocated--;
5986}
5987
5988/* returns true if it causes the idle timer to be disabled */
5989static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
5990{
5991	struct bfq_queue *bfqq = RQ_BFQQ(rq),
5992		*new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true,
5993						 RQ_BIC(rq));
5994	bool waiting, idle_timer_disabled = false;
5995
5996	if (new_bfqq) {
 
 
5997		/*
5998		 * Release the request's reference to the old bfqq
5999		 * and make sure one is taken to the shared queue.
6000		 */
6001		bfqq_request_allocated(new_bfqq);
6002		bfqq_request_freed(bfqq);
6003		new_bfqq->ref++;
6004		/*
6005		 * If the bic associated with the process
6006		 * issuing this request still points to bfqq
6007		 * (and thus has not been already redirected
6008		 * to new_bfqq or even some other bfq_queue),
6009		 * then complete the merge and redirect it to
6010		 * new_bfqq.
6011		 */
6012		if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
6013			bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
6014					bfqq, new_bfqq);
6015
6016		bfq_clear_bfqq_just_created(bfqq);
6017		/*
6018		 * rq is about to be enqueued into new_bfqq,
6019		 * release rq reference on bfqq
6020		 */
6021		bfq_put_queue(bfqq);
6022		rq->elv.priv[1] = new_bfqq;
6023		bfqq = new_bfqq;
6024	}
6025
6026	bfq_update_io_thinktime(bfqd, bfqq);
6027	bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
6028	bfq_update_io_seektime(bfqd, bfqq, rq);
6029
6030	waiting = bfqq && bfq_bfqq_wait_request(bfqq);
6031	bfq_add_request(rq);
6032	idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
6033
6034	rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
6035	list_add_tail(&rq->queuelist, &bfqq->fifo);
6036
6037	bfq_rq_enqueued(bfqd, bfqq, rq);
6038
6039	return idle_timer_disabled;
6040}
6041
6042#ifdef CONFIG_BFQ_CGROUP_DEBUG
6043static void bfq_update_insert_stats(struct request_queue *q,
6044				    struct bfq_queue *bfqq,
6045				    bool idle_timer_disabled,
6046				    blk_opf_t cmd_flags)
6047{
6048	if (!bfqq)
6049		return;
6050
6051	/*
6052	 * bfqq still exists, because it can disappear only after
6053	 * either it is merged with another queue, or the process it
6054	 * is associated with exits. But both actions must be taken by
6055	 * the same process currently executing this flow of
6056	 * instructions.
6057	 *
6058	 * In addition, the following queue lock guarantees that
6059	 * bfqq_group(bfqq) exists as well.
6060	 */
6061	spin_lock_irq(&q->queue_lock);
6062	bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
6063	if (idle_timer_disabled)
6064		bfqg_stats_update_idle_time(bfqq_group(bfqq));
6065	spin_unlock_irq(&q->queue_lock);
6066}
6067#else
6068static inline void bfq_update_insert_stats(struct request_queue *q,
6069					   struct bfq_queue *bfqq,
6070					   bool idle_timer_disabled,
6071					   blk_opf_t cmd_flags) {}
6072#endif /* CONFIG_BFQ_CGROUP_DEBUG */
6073
6074static struct bfq_queue *bfq_init_rq(struct request *rq);
6075
6076static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
6077			       bool at_head)
6078{
6079	struct request_queue *q = hctx->queue;
6080	struct bfq_data *bfqd = q->elevator->elevator_data;
6081	struct bfq_queue *bfqq;
6082	bool idle_timer_disabled = false;
6083	blk_opf_t cmd_flags;
6084	LIST_HEAD(free);
6085
6086#ifdef CONFIG_BFQ_GROUP_IOSCHED
6087	if (!cgroup_subsys_on_dfl(io_cgrp_subsys) && rq->bio)
6088		bfqg_stats_update_legacy_io(q, rq);
6089#endif
6090	spin_lock_irq(&bfqd->lock);
6091	bfqq = bfq_init_rq(rq);
6092	if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
6093		spin_unlock_irq(&bfqd->lock);
6094		blk_mq_free_requests(&free);
6095		return;
6096	}
6097
6098	trace_block_rq_insert(rq);
 
 
6099
6100	if (!bfqq || at_head) {
 
6101		if (at_head)
6102			list_add(&rq->queuelist, &bfqd->dispatch);
6103		else
6104			list_add_tail(&rq->queuelist, &bfqd->dispatch);
6105	} else {
 
 
 
 
 
 
 
 
 
 
 
 
6106		idle_timer_disabled = __bfq_insert_request(bfqd, rq);
6107		/*
6108		 * Update bfqq, because, if a queue merge has occurred
6109		 * in __bfq_insert_request, then rq has been
6110		 * redirected into a new queue.
6111		 */
6112		bfqq = RQ_BFQQ(rq);
6113
6114		if (rq_mergeable(rq)) {
6115			elv_rqhash_add(q, rq);
6116			if (!q->last_merge)
6117				q->last_merge = rq;
6118		}
6119	}
6120
6121	/*
6122	 * Cache cmd_flags before releasing scheduler lock, because rq
6123	 * may disappear afterwards (for example, because of a request
6124	 * merge).
6125	 */
6126	cmd_flags = rq->cmd_flags;
 
6127	spin_unlock_irq(&bfqd->lock);
6128
6129	bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
6130				cmd_flags);
6131}
6132
6133static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
6134				struct list_head *list, bool at_head)
6135{
6136	while (!list_empty(list)) {
6137		struct request *rq;
6138
6139		rq = list_first_entry(list, struct request, queuelist);
6140		list_del_init(&rq->queuelist);
6141		bfq_insert_request(hctx, rq, at_head);
6142	}
6143}
6144
6145static void bfq_update_hw_tag(struct bfq_data *bfqd)
6146{
6147	struct bfq_queue *bfqq = bfqd->in_service_queue;
6148
6149	bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
6150				       bfqd->rq_in_driver);
6151
6152	if (bfqd->hw_tag == 1)
6153		return;
6154
6155	/*
6156	 * This sample is valid if the number of outstanding requests
6157	 * is large enough to allow a queueing behavior.  Note that the
6158	 * sum is not exact, as it's not taking into account deactivated
6159	 * requests.
6160	 */
6161	if (bfqd->rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
6162		return;
6163
6164	/*
6165	 * If active queue hasn't enough requests and can idle, bfq might not
6166	 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
6167	 * case
6168	 */
6169	if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
6170	    bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
6171	    BFQ_HW_QUEUE_THRESHOLD &&
6172	    bfqd->rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
6173		return;
6174
6175	if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
6176		return;
6177
6178	bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
6179	bfqd->max_rq_in_driver = 0;
6180	bfqd->hw_tag_samples = 0;
6181
6182	bfqd->nonrot_with_queueing =
6183		blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
6184}
6185
6186static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
6187{
6188	u64 now_ns;
6189	u32 delta_us;
6190
6191	bfq_update_hw_tag(bfqd);
6192
6193	bfqd->rq_in_driver--;
6194	bfqq->dispatched--;
6195
6196	if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
6197		/*
6198		 * Set budget_timeout (which we overload to store the
6199		 * time at which the queue remains with no backlog and
6200		 * no outstanding request; used by the weight-raising
6201		 * mechanism).
6202		 */
6203		bfqq->budget_timeout = jiffies;
6204
6205		bfq_del_bfqq_in_groups_with_pending_reqs(bfqq);
6206		bfq_weights_tree_remove(bfqq);
6207	}
6208
6209	now_ns = ktime_get_ns();
6210
6211	bfqq->ttime.last_end_request = now_ns;
6212
6213	/*
6214	 * Using us instead of ns, to get a reasonable precision in
6215	 * computing rate in next check.
6216	 */
6217	delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
6218
6219	/*
6220	 * If the request took rather long to complete, and, according
6221	 * to the maximum request size recorded, this completion latency
6222	 * implies that the request was certainly served at a very low
6223	 * rate (less than 1M sectors/sec), then the whole observation
6224	 * interval that lasts up to this time instant cannot be a
6225	 * valid time interval for computing a new peak rate.  Invoke
6226	 * bfq_update_rate_reset to have the following three steps
6227	 * taken:
6228	 * - close the observation interval at the last (previous)
6229	 *   request dispatch or completion
6230	 * - compute rate, if possible, for that observation interval
6231	 * - reset to zero samples, which will trigger a proper
6232	 *   re-initialization of the observation interval on next
6233	 *   dispatch
6234	 */
6235	if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
6236	   (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
6237			1UL<<(BFQ_RATE_SHIFT - 10))
6238		bfq_update_rate_reset(bfqd, NULL);
6239	bfqd->last_completion = now_ns;
6240	/*
6241	 * Shared queues are likely to receive I/O at a high
6242	 * rate. This may deceptively let them be considered as wakers
6243	 * of other queues. But a false waker will unjustly steal
6244	 * bandwidth to its supposedly woken queue. So considering
6245	 * also shared queues in the waking mechanism may cause more
6246	 * control troubles than throughput benefits. Then reset
6247	 * last_completed_rq_bfqq if bfqq is a shared queue.
6248	 */
6249	if (!bfq_bfqq_coop(bfqq))
6250		bfqd->last_completed_rq_bfqq = bfqq;
6251	else
6252		bfqd->last_completed_rq_bfqq = NULL;
6253
6254	/*
6255	 * If we are waiting to discover whether the request pattern
6256	 * of the task associated with the queue is actually
6257	 * isochronous, and both requisites for this condition to hold
6258	 * are now satisfied, then compute soft_rt_next_start (see the
6259	 * comments on the function bfq_bfqq_softrt_next_start()). We
6260	 * do not compute soft_rt_next_start if bfqq is in interactive
6261	 * weight raising (see the comments in bfq_bfqq_expire() for
6262	 * an explanation). We schedule this delayed update when bfqq
6263	 * expires, if it still has in-flight requests.
6264	 */
6265	if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
6266	    RB_EMPTY_ROOT(&bfqq->sort_list) &&
6267	    bfqq->wr_coeff != bfqd->bfq_wr_coeff)
6268		bfqq->soft_rt_next_start =
6269			bfq_bfqq_softrt_next_start(bfqd, bfqq);
6270
6271	/*
6272	 * If this is the in-service queue, check if it needs to be expired,
6273	 * or if we want to idle in case it has no pending requests.
6274	 */
6275	if (bfqd->in_service_queue == bfqq) {
6276		if (bfq_bfqq_must_idle(bfqq)) {
6277			if (bfqq->dispatched == 0)
6278				bfq_arm_slice_timer(bfqd);
6279			/*
6280			 * If we get here, we do not expire bfqq, even
6281			 * if bfqq was in budget timeout or had no
6282			 * more requests (as controlled in the next
6283			 * conditional instructions). The reason for
6284			 * not expiring bfqq is as follows.
6285			 *
6286			 * Here bfqq->dispatched > 0 holds, but
6287			 * bfq_bfqq_must_idle() returned true. This
6288			 * implies that, even if no request arrives
6289			 * for bfqq before bfqq->dispatched reaches 0,
6290			 * bfqq will, however, not be expired on the
6291			 * completion event that causes bfqq->dispatch
6292			 * to reach zero. In contrast, on this event,
6293			 * bfqq will start enjoying device idling
6294			 * (I/O-dispatch plugging).
6295			 *
6296			 * But, if we expired bfqq here, bfqq would
6297			 * not have the chance to enjoy device idling
6298			 * when bfqq->dispatched finally reaches
6299			 * zero. This would expose bfqq to violation
6300			 * of its reserved service guarantees.
6301			 */
6302			return;
6303		} else if (bfq_may_expire_for_budg_timeout(bfqq))
6304			bfq_bfqq_expire(bfqd, bfqq, false,
6305					BFQQE_BUDGET_TIMEOUT);
6306		else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
6307			 (bfqq->dispatched == 0 ||
6308			  !bfq_better_to_idle(bfqq)))
6309			bfq_bfqq_expire(bfqd, bfqq, false,
6310					BFQQE_NO_MORE_REQUESTS);
6311	}
6312
6313	if (!bfqd->rq_in_driver)
6314		bfq_schedule_dispatch(bfqd);
6315}
6316
6317/*
6318 * The processes associated with bfqq may happen to generate their
6319 * cumulative I/O at a lower rate than the rate at which the device
6320 * could serve the same I/O. This is rather probable, e.g., if only
6321 * one process is associated with bfqq and the device is an SSD. It
6322 * results in bfqq becoming often empty while in service. In this
6323 * respect, if BFQ is allowed to switch to another queue when bfqq
6324 * remains empty, then the device goes on being fed with I/O requests,
6325 * and the throughput is not affected. In contrast, if BFQ is not
6326 * allowed to switch to another queue---because bfqq is sync and
6327 * I/O-dispatch needs to be plugged while bfqq is temporarily
6328 * empty---then, during the service of bfqq, there will be frequent
6329 * "service holes", i.e., time intervals during which bfqq gets empty
6330 * and the device can only consume the I/O already queued in its
6331 * hardware queues. During service holes, the device may even get to
6332 * remaining idle. In the end, during the service of bfqq, the device
6333 * is driven at a lower speed than the one it can reach with the kind
6334 * of I/O flowing through bfqq.
6335 *
6336 * To counter this loss of throughput, BFQ implements a "request
6337 * injection mechanism", which tries to fill the above service holes
6338 * with I/O requests taken from other queues. The hard part in this
6339 * mechanism is finding the right amount of I/O to inject, so as to
6340 * both boost throughput and not break bfqq's bandwidth and latency
6341 * guarantees. In this respect, the mechanism maintains a per-queue
6342 * inject limit, computed as below. While bfqq is empty, the injection
6343 * mechanism dispatches extra I/O requests only until the total number
6344 * of I/O requests in flight---i.e., already dispatched but not yet
6345 * completed---remains lower than this limit.
6346 *
6347 * A first definition comes in handy to introduce the algorithm by
6348 * which the inject limit is computed.  We define as first request for
6349 * bfqq, an I/O request for bfqq that arrives while bfqq is in
6350 * service, and causes bfqq to switch from empty to non-empty. The
6351 * algorithm updates the limit as a function of the effect of
6352 * injection on the service times of only the first requests of
6353 * bfqq. The reason for this restriction is that these are the
6354 * requests whose service time is affected most, because they are the
6355 * first to arrive after injection possibly occurred.
6356 *
6357 * To evaluate the effect of injection, the algorithm measures the
6358 * "total service time" of first requests. We define as total service
6359 * time of an I/O request, the time that elapses since when the
6360 * request is enqueued into bfqq, to when it is completed. This
6361 * quantity allows the whole effect of injection to be measured. It is
6362 * easy to see why. Suppose that some requests of other queues are
6363 * actually injected while bfqq is empty, and that a new request R
6364 * then arrives for bfqq. If the device does start to serve all or
6365 * part of the injected requests during the service hole, then,
6366 * because of this extra service, it may delay the next invocation of
6367 * the dispatch hook of BFQ. Then, even after R gets eventually
6368 * dispatched, the device may delay the actual service of R if it is
6369 * still busy serving the extra requests, or if it decides to serve,
6370 * before R, some extra request still present in its queues. As a
6371 * conclusion, the cumulative extra delay caused by injection can be
6372 * easily evaluated by just comparing the total service time of first
6373 * requests with and without injection.
6374 *
6375 * The limit-update algorithm works as follows. On the arrival of a
6376 * first request of bfqq, the algorithm measures the total time of the
6377 * request only if one of the three cases below holds, and, for each
6378 * case, it updates the limit as described below:
6379 *
6380 * (1) If there is no in-flight request. This gives a baseline for the
6381 *     total service time of the requests of bfqq. If the baseline has
6382 *     not been computed yet, then, after computing it, the limit is
6383 *     set to 1, to start boosting throughput, and to prepare the
6384 *     ground for the next case. If the baseline has already been
6385 *     computed, then it is updated, in case it results to be lower
6386 *     than the previous value.
6387 *
6388 * (2) If the limit is higher than 0 and there are in-flight
6389 *     requests. By comparing the total service time in this case with
6390 *     the above baseline, it is possible to know at which extent the
6391 *     current value of the limit is inflating the total service
6392 *     time. If the inflation is below a certain threshold, then bfqq
6393 *     is assumed to be suffering from no perceivable loss of its
6394 *     service guarantees, and the limit is even tentatively
6395 *     increased. If the inflation is above the threshold, then the
6396 *     limit is decreased. Due to the lack of any hysteresis, this
6397 *     logic makes the limit oscillate even in steady workload
6398 *     conditions. Yet we opted for it, because it is fast in reaching
6399 *     the best value for the limit, as a function of the current I/O
6400 *     workload. To reduce oscillations, this step is disabled for a
6401 *     short time interval after the limit happens to be decreased.
6402 *
6403 * (3) Periodically, after resetting the limit, to make sure that the
6404 *     limit eventually drops in case the workload changes. This is
6405 *     needed because, after the limit has gone safely up for a
6406 *     certain workload, it is impossible to guess whether the
6407 *     baseline total service time may have changed, without measuring
6408 *     it again without injection. A more effective version of this
6409 *     step might be to just sample the baseline, by interrupting
6410 *     injection only once, and then to reset/lower the limit only if
6411 *     the total service time with the current limit does happen to be
6412 *     too large.
6413 *
6414 * More details on each step are provided in the comments on the
6415 * pieces of code that implement these steps: the branch handling the
6416 * transition from empty to non empty in bfq_add_request(), the branch
6417 * handling injection in bfq_select_queue(), and the function
6418 * bfq_choose_bfqq_for_injection(). These comments also explain some
6419 * exceptions, made by the injection mechanism in some special cases.
6420 */
6421static void bfq_update_inject_limit(struct bfq_data *bfqd,
6422				    struct bfq_queue *bfqq)
6423{
6424	u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns;
6425	unsigned int old_limit = bfqq->inject_limit;
6426
6427	if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) {
6428		u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
6429
6430		if (tot_time_ns >= threshold && old_limit > 0) {
6431			bfqq->inject_limit--;
6432			bfqq->decrease_time_jif = jiffies;
6433		} else if (tot_time_ns < threshold &&
6434			   old_limit <= bfqd->max_rq_in_driver)
6435			bfqq->inject_limit++;
6436	}
6437
6438	/*
6439	 * Either we still have to compute the base value for the
6440	 * total service time, and there seem to be the right
6441	 * conditions to do it, or we can lower the last base value
6442	 * computed.
6443	 *
6444	 * NOTE: (bfqd->rq_in_driver == 1) means that there is no I/O
6445	 * request in flight, because this function is in the code
6446	 * path that handles the completion of a request of bfqq, and,
6447	 * in particular, this function is executed before
6448	 * bfqd->rq_in_driver is decremented in such a code path.
6449	 */
6450	if ((bfqq->last_serv_time_ns == 0 && bfqd->rq_in_driver == 1) ||
6451	    tot_time_ns < bfqq->last_serv_time_ns) {
6452		if (bfqq->last_serv_time_ns == 0) {
6453			/*
6454			 * Now we certainly have a base value: make sure we
6455			 * start trying injection.
6456			 */
6457			bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
6458		}
6459		bfqq->last_serv_time_ns = tot_time_ns;
6460	} else if (!bfqd->rqs_injected && bfqd->rq_in_driver == 1)
6461		/*
6462		 * No I/O injected and no request still in service in
6463		 * the drive: these are the exact conditions for
6464		 * computing the base value of the total service time
6465		 * for bfqq. So let's update this value, because it is
6466		 * rather variable. For example, it varies if the size
6467		 * or the spatial locality of the I/O requests in bfqq
6468		 * change.
6469		 */
6470		bfqq->last_serv_time_ns = tot_time_ns;
6471
6472
6473	/* update complete, not waiting for any request completion any longer */
6474	bfqd->waited_rq = NULL;
6475	bfqd->rqs_injected = false;
6476}
6477
6478/*
6479 * Handle either a requeue or a finish for rq. The things to do are
6480 * the same in both cases: all references to rq are to be dropped. In
6481 * particular, rq is considered completed from the point of view of
6482 * the scheduler.
6483 */
6484static void bfq_finish_requeue_request(struct request *rq)
6485{
6486	struct bfq_queue *bfqq = RQ_BFQQ(rq);
6487	struct bfq_data *bfqd;
6488	unsigned long flags;
 
 
 
 
 
 
 
 
 
 
 
6489
6490	/*
6491	 * rq either is not associated with any icq, or is an already
6492	 * requeued request that has not (yet) been re-inserted into
6493	 * a bfq_queue.
6494	 */
6495	if (!rq->elv.icq || !bfqq)
6496		return;
6497
6498	bfqd = bfqq->bfqd;
6499
6500	if (rq->rq_flags & RQF_STARTED)
6501		bfqg_stats_update_completion(bfqq_group(bfqq),
6502					     rq->start_time_ns,
6503					     rq->io_start_time_ns,
6504					     rq->cmd_flags);
6505
6506	spin_lock_irqsave(&bfqd->lock, flags);
6507	if (likely(rq->rq_flags & RQF_STARTED)) {
6508		if (rq == bfqd->waited_rq)
6509			bfq_update_inject_limit(bfqd, bfqq);
 
6510
6511		bfq_completed_request(bfqq, bfqd);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6512	}
6513	bfqq_request_freed(bfqq);
6514	bfq_put_queue(bfqq);
6515	RQ_BIC(rq)->requests--;
6516	spin_unlock_irqrestore(&bfqd->lock, flags);
6517
6518	/*
6519	 * Reset private fields. In case of a requeue, this allows
6520	 * this function to correctly do nothing if it is spuriously
6521	 * invoked again on this same request (see the check at the
6522	 * beginning of the function). Probably, a better general
6523	 * design would be to prevent blk-mq from invoking the requeue
6524	 * or finish hooks of an elevator, for a request that is not
6525	 * referred by that elevator.
6526	 *
6527	 * Resetting the following fields would break the
6528	 * request-insertion logic if rq is re-inserted into a bfq
6529	 * internal queue, without a re-preparation. Here we assume
6530	 * that re-insertions of requeued requests, without
6531	 * re-preparation, can happen only for pass_through or at_head
6532	 * requests (which are not re-inserted into bfq internal
6533	 * queues).
6534	 */
6535	rq->elv.priv[0] = NULL;
6536	rq->elv.priv[1] = NULL;
6537}
6538
6539static void bfq_finish_request(struct request *rq)
6540{
6541	bfq_finish_requeue_request(rq);
6542
6543	if (rq->elv.icq) {
6544		put_io_context(rq->elv.icq->ioc);
6545		rq->elv.icq = NULL;
6546	}
6547}
6548
6549/*
6550 * Removes the association between the current task and bfqq, assuming
6551 * that bic points to the bfq iocontext of the task.
6552 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
6553 * was the last process referring to that bfqq.
6554 */
6555static struct bfq_queue *
6556bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
6557{
6558	bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
6559
6560	if (bfqq_process_refs(bfqq) == 1) {
6561		bfqq->pid = current->pid;
6562		bfq_clear_bfqq_coop(bfqq);
6563		bfq_clear_bfqq_split_coop(bfqq);
6564		return bfqq;
6565	}
6566
6567	bic_set_bfqq(bic, NULL, true);
6568
6569	bfq_put_cooperator(bfqq);
6570
6571	bfq_release_process_ref(bfqq->bfqd, bfqq);
6572	return NULL;
6573}
6574
6575static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
6576						   struct bfq_io_cq *bic,
6577						   struct bio *bio,
6578						   bool split, bool is_sync,
6579						   bool *new_queue)
6580{
6581	struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
6582
6583	if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
6584		return bfqq;
6585
6586	if (new_queue)
6587		*new_queue = true;
6588
6589	if (bfqq)
6590		bfq_put_queue(bfqq);
6591	bfqq = bfq_get_queue(bfqd, bio, is_sync, bic, split);
6592
6593	bic_set_bfqq(bic, bfqq, is_sync);
6594	if (split && is_sync) {
6595		if ((bic->was_in_burst_list && bfqd->large_burst) ||
6596		    bic->saved_in_large_burst)
6597			bfq_mark_bfqq_in_large_burst(bfqq);
6598		else {
6599			bfq_clear_bfqq_in_large_burst(bfqq);
6600			if (bic->was_in_burst_list)
6601				/*
6602				 * If bfqq was in the current
6603				 * burst list before being
6604				 * merged, then we have to add
6605				 * it back. And we do not need
6606				 * to increase burst_size, as
6607				 * we did not decrement
6608				 * burst_size when we removed
6609				 * bfqq from the burst list as
6610				 * a consequence of a merge
6611				 * (see comments in
6612				 * bfq_put_queue). In this
6613				 * respect, it would be rather
6614				 * costly to know whether the
6615				 * current burst list is still
6616				 * the same burst list from
6617				 * which bfqq was removed on
6618				 * the merge. To avoid this
6619				 * cost, if bfqq was in a
6620				 * burst list, then we add
6621				 * bfqq to the current burst
6622				 * list without any further
6623				 * check. This can cause
6624				 * inappropriate insertions,
6625				 * but rarely enough to not
6626				 * harm the detection of large
6627				 * bursts significantly.
6628				 */
6629				hlist_add_head(&bfqq->burst_list_node,
6630					       &bfqd->burst_list);
6631		}
6632		bfqq->split_time = jiffies;
6633	}
6634
6635	return bfqq;
6636}
6637
6638/*
6639 * Only reset private fields. The actual request preparation will be
6640 * performed by bfq_init_rq, when rq is either inserted or merged. See
6641 * comments on bfq_init_rq for the reason behind this delayed
6642 * preparation.
6643 */
6644static void bfq_prepare_request(struct request *rq)
6645{
6646	rq->elv.icq = ioc_find_get_icq(rq->q);
6647
6648	/*
6649	 * Regardless of whether we have an icq attached, we have to
6650	 * clear the scheduler pointers, as they might point to
6651	 * previously allocated bic/bfqq structs.
6652	 */
6653	rq->elv.priv[0] = rq->elv.priv[1] = NULL;
6654}
6655
6656/*
6657 * If needed, init rq, allocate bfq data structures associated with
6658 * rq, and increment reference counters in the destination bfq_queue
6659 * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6660 * not associated with any bfq_queue.
6661 *
6662 * This function is invoked by the functions that perform rq insertion
6663 * or merging. One may have expected the above preparation operations
6664 * to be performed in bfq_prepare_request, and not delayed to when rq
6665 * is inserted or merged. The rationale behind this delayed
6666 * preparation is that, after the prepare_request hook is invoked for
6667 * rq, rq may still be transformed into a request with no icq, i.e., a
6668 * request not associated with any queue. No bfq hook is invoked to
6669 * signal this transformation. As a consequence, should these
6670 * preparation operations be performed when the prepare_request hook
6671 * is invoked, and should rq be transformed one moment later, bfq
6672 * would end up in an inconsistent state, because it would have
6673 * incremented some queue counters for an rq destined to
6674 * transformation, without any chance to correctly lower these
6675 * counters back. In contrast, no transformation can still happen for
6676 * rq after rq has been inserted or merged. So, it is safe to execute
6677 * these preparation operations when rq is finally inserted or merged.
6678 */
6679static struct bfq_queue *bfq_init_rq(struct request *rq)
6680{
6681	struct request_queue *q = rq->q;
6682	struct bio *bio = rq->bio;
6683	struct bfq_data *bfqd = q->elevator->elevator_data;
6684	struct bfq_io_cq *bic;
6685	const int is_sync = rq_is_sync(rq);
6686	struct bfq_queue *bfqq;
6687	bool new_queue = false;
6688	bool bfqq_already_existing = false, split = false;
6689
6690	if (unlikely(!rq->elv.icq))
6691		return NULL;
6692
6693	/*
6694	 * Assuming that elv.priv[1] is set only if everything is set
6695	 * for this rq. This holds true, because this function is
6696	 * invoked only for insertion or merging, and, after such
6697	 * events, a request cannot be manipulated any longer before
6698	 * being removed from bfq.
6699	 */
6700	if (rq->elv.priv[1])
6701		return rq->elv.priv[1];
 
 
6702
6703	bic = icq_to_bic(rq->elv.icq);
6704
 
 
6705	bfq_check_ioprio_change(bic, bio);
6706
6707	bfq_bic_update_cgroup(bic, bio);
6708
6709	bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
6710					 &new_queue);
6711
6712	if (likely(!new_queue)) {
6713		/* If the queue was seeky for too long, break it apart. */
6714		if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq) &&
6715			!bic->stably_merged) {
6716			struct bfq_queue *old_bfqq = bfqq;
6717
6718			/* Update bic before losing reference to bfqq */
6719			if (bfq_bfqq_in_large_burst(bfqq))
6720				bic->saved_in_large_burst = true;
6721
6722			bfqq = bfq_split_bfqq(bic, bfqq);
6723			split = true;
6724
6725			if (!bfqq) {
6726				bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
6727								 true, is_sync,
6728								 NULL);
6729				if (unlikely(bfqq == &bfqd->oom_bfqq))
6730					bfqq_already_existing = true;
6731			} else
6732				bfqq_already_existing = true;
6733
6734			if (!bfqq_already_existing) {
6735				bfqq->waker_bfqq = old_bfqq->waker_bfqq;
6736				bfqq->tentative_waker_bfqq = NULL;
6737
6738				/*
6739				 * If the waker queue disappears, then
6740				 * new_bfqq->waker_bfqq must be
6741				 * reset. So insert new_bfqq into the
6742				 * woken_list of the waker. See
6743				 * bfq_check_waker for details.
6744				 */
6745				if (bfqq->waker_bfqq)
6746					hlist_add_head(&bfqq->woken_list_node,
6747						       &bfqq->waker_bfqq->woken_list);
6748			}
6749		}
6750	}
6751
6752	bfqq_request_allocated(bfqq);
6753	bfqq->ref++;
6754	bic->requests++;
6755	bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
6756		     rq, bfqq, bfqq->ref);
6757
6758	rq->elv.priv[0] = bic;
6759	rq->elv.priv[1] = bfqq;
6760
6761	/*
6762	 * If a bfq_queue has only one process reference, it is owned
6763	 * by only this bic: we can then set bfqq->bic = bic. in
6764	 * addition, if the queue has also just been split, we have to
6765	 * resume its state.
6766	 */
6767	if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
6768		bfqq->bic = bic;
6769		if (split) {
6770			/*
6771			 * The queue has just been split from a shared
6772			 * queue: restore the idle window and the
6773			 * possible weight raising period.
6774			 */
6775			bfq_bfqq_resume_state(bfqq, bfqd, bic,
6776					      bfqq_already_existing);
6777		}
6778	}
6779
6780	/*
6781	 * Consider bfqq as possibly belonging to a burst of newly
6782	 * created queues only if:
6783	 * 1) A burst is actually happening (bfqd->burst_size > 0)
6784	 * or
6785	 * 2) There is no other active queue. In fact, if, in
6786	 *    contrast, there are active queues not belonging to the
6787	 *    possible burst bfqq may belong to, then there is no gain
6788	 *    in considering bfqq as belonging to a burst, and
6789	 *    therefore in not weight-raising bfqq. See comments on
6790	 *    bfq_handle_burst().
6791	 *
6792	 * This filtering also helps eliminating false positives,
6793	 * occurring when bfqq does not belong to an actual large
6794	 * burst, but some background task (e.g., a service) happens
6795	 * to trigger the creation of new queues very close to when
6796	 * bfqq and its possible companion queues are created. See
6797	 * comments on bfq_handle_burst() for further details also on
6798	 * this issue.
6799	 */
6800	if (unlikely(bfq_bfqq_just_created(bfqq) &&
6801		     (bfqd->burst_size > 0 ||
6802		      bfq_tot_busy_queues(bfqd) == 0)))
6803		bfq_handle_burst(bfqd, bfqq);
6804
6805	return bfqq;
6806}
6807
6808static void
6809bfq_idle_slice_timer_body(struct bfq_data *bfqd, struct bfq_queue *bfqq)
6810{
 
6811	enum bfqq_expiration reason;
6812	unsigned long flags;
6813
6814	spin_lock_irqsave(&bfqd->lock, flags);
 
6815
6816	/*
6817	 * Considering that bfqq may be in race, we should firstly check
6818	 * whether bfqq is in service before doing something on it. If
6819	 * the bfqq in race is not in service, it has already been expired
6820	 * through __bfq_bfqq_expire func and its wait_request flags has
6821	 * been cleared in __bfq_bfqd_reset_in_service func.
6822	 */
6823	if (bfqq != bfqd->in_service_queue) {
6824		spin_unlock_irqrestore(&bfqd->lock, flags);
6825		return;
6826	}
6827
6828	bfq_clear_bfqq_wait_request(bfqq);
6829
6830	if (bfq_bfqq_budget_timeout(bfqq))
6831		/*
6832		 * Also here the queue can be safely expired
6833		 * for budget timeout without wasting
6834		 * guarantees
6835		 */
6836		reason = BFQQE_BUDGET_TIMEOUT;
6837	else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
6838		/*
6839		 * The queue may not be empty upon timer expiration,
6840		 * because we may not disable the timer when the
6841		 * first request of the in-service queue arrives
6842		 * during disk idling.
6843		 */
6844		reason = BFQQE_TOO_IDLE;
6845	else
6846		goto schedule_dispatch;
6847
6848	bfq_bfqq_expire(bfqd, bfqq, true, reason);
6849
6850schedule_dispatch:
6851	bfq_schedule_dispatch(bfqd);
6852	spin_unlock_irqrestore(&bfqd->lock, flags);
 
6853}
6854
6855/*
6856 * Handler of the expiration of the timer running if the in-service queue
6857 * is idling inside its time slice.
6858 */
6859static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
6860{
6861	struct bfq_data *bfqd = container_of(timer, struct bfq_data,
6862					     idle_slice_timer);
6863	struct bfq_queue *bfqq = bfqd->in_service_queue;
6864
6865	/*
6866	 * Theoretical race here: the in-service queue can be NULL or
6867	 * different from the queue that was idling if a new request
6868	 * arrives for the current queue and there is a full dispatch
6869	 * cycle that changes the in-service queue.  This can hardly
6870	 * happen, but in the worst case we just expire a queue too
6871	 * early.
6872	 */
6873	if (bfqq)
6874		bfq_idle_slice_timer_body(bfqd, bfqq);
6875
6876	return HRTIMER_NORESTART;
6877}
6878
6879static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
6880				 struct bfq_queue **bfqq_ptr)
6881{
6882	struct bfq_queue *bfqq = *bfqq_ptr;
6883
6884	bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
6885	if (bfqq) {
6886		bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
6887
6888		bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
6889			     bfqq, bfqq->ref);
6890		bfq_put_queue(bfqq);
6891		*bfqq_ptr = NULL;
6892	}
6893}
6894
6895/*
6896 * Release all the bfqg references to its async queues.  If we are
6897 * deallocating the group these queues may still contain requests, so
6898 * we reparent them to the root cgroup (i.e., the only one that will
6899 * exist for sure until all the requests on a device are gone).
6900 */
6901void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
6902{
6903	int i, j;
6904
6905	for (i = 0; i < 2; i++)
6906		for (j = 0; j < IOPRIO_NR_LEVELS; j++)
6907			__bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
6908
6909	__bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
6910}
6911
6912/*
6913 * See the comments on bfq_limit_depth for the purpose of
6914 * the depths set in the function. Return minimum shallow depth we'll use.
6915 */
6916static void bfq_update_depths(struct bfq_data *bfqd, struct sbitmap_queue *bt)
6917{
6918	unsigned int depth = 1U << bt->sb.shift;
6919
6920	bfqd->full_depth_shift = bt->sb.shift;
6921	/*
6922	 * In-word depths if no bfq_queue is being weight-raised:
6923	 * leaving 25% of tags only for sync reads.
6924	 *
6925	 * In next formulas, right-shift the value
6926	 * (1U<<bt->sb.shift), instead of computing directly
6927	 * (1U<<(bt->sb.shift - something)), to be robust against
6928	 * any possible value of bt->sb.shift, without having to
6929	 * limit 'something'.
6930	 */
6931	/* no more than 50% of tags for async I/O */
6932	bfqd->word_depths[0][0] = max(depth >> 1, 1U);
6933	/*
6934	 * no more than 75% of tags for sync writes (25% extra tags
6935	 * w.r.t. async I/O, to prevent async I/O from starving sync
6936	 * writes)
6937	 */
6938	bfqd->word_depths[0][1] = max((depth * 3) >> 2, 1U);
6939
6940	/*
6941	 * In-word depths in case some bfq_queue is being weight-
6942	 * raised: leaving ~63% of tags for sync reads. This is the
6943	 * highest percentage for which, in our tests, application
6944	 * start-up times didn't suffer from any regression due to tag
6945	 * shortage.
6946	 */
6947	/* no more than ~18% of tags for async I/O */
6948	bfqd->word_depths[1][0] = max((depth * 3) >> 4, 1U);
6949	/* no more than ~37% of tags for sync writes (~20% extra tags) */
6950	bfqd->word_depths[1][1] = max((depth * 6) >> 4, 1U);
6951}
6952
6953static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
6954{
6955	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
6956	struct blk_mq_tags *tags = hctx->sched_tags;
6957
6958	bfq_update_depths(bfqd, &tags->bitmap_tags);
6959	sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, 1);
6960}
6961
6962static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
6963{
6964	bfq_depth_updated(hctx);
6965	return 0;
6966}
6967
6968static void bfq_exit_queue(struct elevator_queue *e)
6969{
6970	struct bfq_data *bfqd = e->elevator_data;
6971	struct bfq_queue *bfqq, *n;
6972
6973	hrtimer_cancel(&bfqd->idle_slice_timer);
6974
6975	spin_lock_irq(&bfqd->lock);
6976	list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
6977		bfq_deactivate_bfqq(bfqd, bfqq, false, false);
6978	spin_unlock_irq(&bfqd->lock);
6979
6980	hrtimer_cancel(&bfqd->idle_slice_timer);
6981
 
6982	/* release oom-queue reference to root group */
6983	bfqg_and_blkg_put(bfqd->root_group);
6984
6985#ifdef CONFIG_BFQ_GROUP_IOSCHED
6986	blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
6987#else
6988	spin_lock_irq(&bfqd->lock);
6989	bfq_put_async_queues(bfqd, bfqd->root_group);
6990	kfree(bfqd->root_group);
6991	spin_unlock_irq(&bfqd->lock);
6992#endif
6993
6994	blk_stat_disable_accounting(bfqd->queue);
6995	clear_bit(ELEVATOR_FLAG_DISABLE_WBT, &e->flags);
6996	wbt_enable_default(bfqd->queue);
6997
6998	kfree(bfqd);
6999}
7000
7001static void bfq_init_root_group(struct bfq_group *root_group,
7002				struct bfq_data *bfqd)
7003{
7004	int i;
7005
7006#ifdef CONFIG_BFQ_GROUP_IOSCHED
7007	root_group->entity.parent = NULL;
7008	root_group->my_entity = NULL;
7009	root_group->bfqd = bfqd;
7010#endif
7011	root_group->rq_pos_tree = RB_ROOT;
7012	for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
7013		root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
7014	root_group->sched_data.bfq_class_idle_last_service = jiffies;
7015}
7016
7017static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
7018{
7019	struct bfq_data *bfqd;
7020	struct elevator_queue *eq;
7021
7022	eq = elevator_alloc(q, e);
7023	if (!eq)
7024		return -ENOMEM;
7025
7026	bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
7027	if (!bfqd) {
7028		kobject_put(&eq->kobj);
7029		return -ENOMEM;
7030	}
7031	eq->elevator_data = bfqd;
7032
7033	spin_lock_irq(&q->queue_lock);
7034	q->elevator = eq;
7035	spin_unlock_irq(&q->queue_lock);
7036
7037	/*
7038	 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
7039	 * Grab a permanent reference to it, so that the normal code flow
7040	 * will not attempt to free it.
7041	 */
7042	bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
7043	bfqd->oom_bfqq.ref++;
7044	bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
7045	bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
7046	bfqd->oom_bfqq.entity.new_weight =
7047		bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
7048
7049	/* oom_bfqq does not participate to bursts */
7050	bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
7051
7052	/*
7053	 * Trigger weight initialization, according to ioprio, at the
7054	 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
7055	 * class won't be changed any more.
7056	 */
7057	bfqd->oom_bfqq.entity.prio_changed = 1;
7058
7059	bfqd->queue = q;
7060
7061	INIT_LIST_HEAD(&bfqd->dispatch);
7062
7063	hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
7064		     HRTIMER_MODE_REL);
7065	bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
7066
7067	bfqd->queue_weights_tree = RB_ROOT_CACHED;
7068#ifdef CONFIG_BFQ_GROUP_IOSCHED
7069	bfqd->num_groups_with_pending_reqs = 0;
7070#endif
7071
7072	INIT_LIST_HEAD(&bfqd->active_list);
7073	INIT_LIST_HEAD(&bfqd->idle_list);
7074	INIT_HLIST_HEAD(&bfqd->burst_list);
7075
7076	bfqd->hw_tag = -1;
7077	bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
7078
7079	bfqd->bfq_max_budget = bfq_default_max_budget;
7080
7081	bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
7082	bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
7083	bfqd->bfq_back_max = bfq_back_max;
7084	bfqd->bfq_back_penalty = bfq_back_penalty;
7085	bfqd->bfq_slice_idle = bfq_slice_idle;
7086	bfqd->bfq_timeout = bfq_timeout;
7087
 
 
7088	bfqd->bfq_large_burst_thresh = 8;
7089	bfqd->bfq_burst_interval = msecs_to_jiffies(180);
7090
7091	bfqd->low_latency = true;
7092
7093	/*
7094	 * Trade-off between responsiveness and fairness.
7095	 */
7096	bfqd->bfq_wr_coeff = 30;
7097	bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
7098	bfqd->bfq_wr_max_time = 0;
7099	bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
7100	bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
7101	bfqd->bfq_wr_max_softrt_rate = 7000; /*
7102					      * Approximate rate required
7103					      * to playback or record a
7104					      * high-definition compressed
7105					      * video.
7106					      */
7107	bfqd->wr_busy_queues = 0;
7108
7109	/*
7110	 * Begin by assuming, optimistically, that the device peak
7111	 * rate is equal to 2/3 of the highest reference rate.
7112	 */
7113	bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
7114		ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
7115	bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
 
 
7116
7117	spin_lock_init(&bfqd->lock);
7118
7119	/*
7120	 * The invocation of the next bfq_create_group_hierarchy
7121	 * function is the head of a chain of function calls
7122	 * (bfq_create_group_hierarchy->blkcg_activate_policy->
7123	 * blk_mq_freeze_queue) that may lead to the invocation of the
7124	 * has_work hook function. For this reason,
7125	 * bfq_create_group_hierarchy is invoked only after all
7126	 * scheduler data has been initialized, apart from the fields
7127	 * that can be initialized only after invoking
7128	 * bfq_create_group_hierarchy. This, in particular, enables
7129	 * has_work to correctly return false. Of course, to avoid
7130	 * other inconsistencies, the blk-mq stack must then refrain
7131	 * from invoking further scheduler hooks before this init
7132	 * function is finished.
7133	 */
7134	bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
7135	if (!bfqd->root_group)
7136		goto out_free;
7137	bfq_init_root_group(bfqd->root_group, bfqd);
7138	bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
7139
7140	/* We dispatch from request queue wide instead of hw queue */
7141	blk_queue_flag_set(QUEUE_FLAG_SQ_SCHED, q);
7142
7143	set_bit(ELEVATOR_FLAG_DISABLE_WBT, &eq->flags);
7144	wbt_disable_default(q);
7145	blk_stat_enable_accounting(q);
7146
7147	return 0;
7148
7149out_free:
7150	kfree(bfqd);
7151	kobject_put(&eq->kobj);
7152	return -ENOMEM;
7153}
7154
7155static void bfq_slab_kill(void)
7156{
7157	kmem_cache_destroy(bfq_pool);
7158}
7159
7160static int __init bfq_slab_setup(void)
7161{
7162	bfq_pool = KMEM_CACHE(bfq_queue, 0);
7163	if (!bfq_pool)
7164		return -ENOMEM;
7165	return 0;
7166}
7167
7168static ssize_t bfq_var_show(unsigned int var, char *page)
7169{
7170	return sprintf(page, "%u\n", var);
7171}
7172
7173static int bfq_var_store(unsigned long *var, const char *page)
7174{
7175	unsigned long new_val;
7176	int ret = kstrtoul(page, 10, &new_val);
7177
7178	if (ret)
7179		return ret;
7180	*var = new_val;
7181	return 0;
7182}
7183
7184#define SHOW_FUNCTION(__FUNC, __VAR, __CONV)				\
7185static ssize_t __FUNC(struct elevator_queue *e, char *page)		\
7186{									\
7187	struct bfq_data *bfqd = e->elevator_data;			\
7188	u64 __data = __VAR;						\
7189	if (__CONV == 1)						\
7190		__data = jiffies_to_msecs(__data);			\
7191	else if (__CONV == 2)						\
7192		__data = div_u64(__data, NSEC_PER_MSEC);		\
7193	return bfq_var_show(__data, (page));				\
7194}
7195SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
7196SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
7197SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
7198SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
7199SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
7200SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
7201SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
7202SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
7203SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
7204#undef SHOW_FUNCTION
7205
7206#define USEC_SHOW_FUNCTION(__FUNC, __VAR)				\
7207static ssize_t __FUNC(struct elevator_queue *e, char *page)		\
7208{									\
7209	struct bfq_data *bfqd = e->elevator_data;			\
7210	u64 __data = __VAR;						\
7211	__data = div_u64(__data, NSEC_PER_USEC);			\
7212	return bfq_var_show(__data, (page));				\
7213}
7214USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
7215#undef USEC_SHOW_FUNCTION
7216
7217#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)			\
7218static ssize_t								\
7219__FUNC(struct elevator_queue *e, const char *page, size_t count)	\
7220{									\
7221	struct bfq_data *bfqd = e->elevator_data;			\
7222	unsigned long __data, __min = (MIN), __max = (MAX);		\
7223	int ret;							\
7224									\
7225	ret = bfq_var_store(&__data, (page));				\
7226	if (ret)							\
7227		return ret;						\
7228	if (__data < __min)						\
7229		__data = __min;						\
7230	else if (__data > __max)					\
7231		__data = __max;						\
7232	if (__CONV == 1)						\
7233		*(__PTR) = msecs_to_jiffies(__data);			\
7234	else if (__CONV == 2)						\
7235		*(__PTR) = (u64)__data * NSEC_PER_MSEC;			\
7236	else								\
7237		*(__PTR) = __data;					\
7238	return count;							\
7239}
7240STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
7241		INT_MAX, 2);
7242STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
7243		INT_MAX, 2);
7244STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
7245STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
7246		INT_MAX, 0);
7247STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
7248#undef STORE_FUNCTION
7249
7250#define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX)			\
7251static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
7252{									\
7253	struct bfq_data *bfqd = e->elevator_data;			\
7254	unsigned long __data, __min = (MIN), __max = (MAX);		\
7255	int ret;							\
7256									\
7257	ret = bfq_var_store(&__data, (page));				\
7258	if (ret)							\
7259		return ret;						\
7260	if (__data < __min)						\
7261		__data = __min;						\
7262	else if (__data > __max)					\
7263		__data = __max;						\
7264	*(__PTR) = (u64)__data * NSEC_PER_USEC;				\
7265	return count;							\
7266}
7267USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
7268		    UINT_MAX);
7269#undef USEC_STORE_FUNCTION
7270
7271static ssize_t bfq_max_budget_store(struct elevator_queue *e,
7272				    const char *page, size_t count)
7273{
7274	struct bfq_data *bfqd = e->elevator_data;
7275	unsigned long __data;
7276	int ret;
7277
7278	ret = bfq_var_store(&__data, (page));
7279	if (ret)
7280		return ret;
7281
7282	if (__data == 0)
7283		bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7284	else {
7285		if (__data > INT_MAX)
7286			__data = INT_MAX;
7287		bfqd->bfq_max_budget = __data;
7288	}
7289
7290	bfqd->bfq_user_max_budget = __data;
7291
7292	return count;
7293}
7294
7295/*
7296 * Leaving this name to preserve name compatibility with cfq
7297 * parameters, but this timeout is used for both sync and async.
7298 */
7299static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
7300				      const char *page, size_t count)
7301{
7302	struct bfq_data *bfqd = e->elevator_data;
7303	unsigned long __data;
7304	int ret;
7305
7306	ret = bfq_var_store(&__data, (page));
7307	if (ret)
7308		return ret;
7309
7310	if (__data < 1)
7311		__data = 1;
7312	else if (__data > INT_MAX)
7313		__data = INT_MAX;
7314
7315	bfqd->bfq_timeout = msecs_to_jiffies(__data);
7316	if (bfqd->bfq_user_max_budget == 0)
7317		bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7318
7319	return count;
7320}
7321
7322static ssize_t bfq_strict_guarantees_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 > 1)
7334		__data = 1;
7335	if (!bfqd->strict_guarantees && __data == 1
7336	    && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
7337		bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
7338
7339	bfqd->strict_guarantees = __data;
7340
7341	return count;
7342}
7343
7344static ssize_t bfq_low_latency_store(struct elevator_queue *e,
7345				     const char *page, size_t count)
7346{
7347	struct bfq_data *bfqd = e->elevator_data;
7348	unsigned long __data;
7349	int ret;
7350
7351	ret = bfq_var_store(&__data, (page));
7352	if (ret)
7353		return ret;
7354
7355	if (__data > 1)
7356		__data = 1;
7357	if (__data == 0 && bfqd->low_latency != 0)
7358		bfq_end_wr(bfqd);
7359	bfqd->low_latency = __data;
7360
7361	return count;
7362}
7363
7364#define BFQ_ATTR(name) \
7365	__ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
7366
7367static struct elv_fs_entry bfq_attrs[] = {
7368	BFQ_ATTR(fifo_expire_sync),
7369	BFQ_ATTR(fifo_expire_async),
7370	BFQ_ATTR(back_seek_max),
7371	BFQ_ATTR(back_seek_penalty),
7372	BFQ_ATTR(slice_idle),
7373	BFQ_ATTR(slice_idle_us),
7374	BFQ_ATTR(max_budget),
7375	BFQ_ATTR(timeout_sync),
7376	BFQ_ATTR(strict_guarantees),
7377	BFQ_ATTR(low_latency),
7378	__ATTR_NULL
7379};
7380
7381static struct elevator_type iosched_bfq_mq = {
7382	.ops = {
7383		.limit_depth		= bfq_limit_depth,
7384		.prepare_request	= bfq_prepare_request,
7385		.requeue_request        = bfq_finish_requeue_request,
7386		.finish_request		= bfq_finish_request,
7387		.exit_icq		= bfq_exit_icq,
7388		.insert_requests	= bfq_insert_requests,
7389		.dispatch_request	= bfq_dispatch_request,
7390		.next_request		= elv_rb_latter_request,
7391		.former_request		= elv_rb_former_request,
7392		.allow_merge		= bfq_allow_bio_merge,
7393		.bio_merge		= bfq_bio_merge,
7394		.request_merge		= bfq_request_merge,
7395		.requests_merged	= bfq_requests_merged,
7396		.request_merged		= bfq_request_merged,
7397		.has_work		= bfq_has_work,
7398		.depth_updated		= bfq_depth_updated,
7399		.init_hctx		= bfq_init_hctx,
7400		.init_sched		= bfq_init_queue,
7401		.exit_sched		= bfq_exit_queue,
7402	},
7403
 
7404	.icq_size =		sizeof(struct bfq_io_cq),
7405	.icq_align =		__alignof__(struct bfq_io_cq),
7406	.elevator_attrs =	bfq_attrs,
7407	.elevator_name =	"bfq",
7408	.elevator_owner =	THIS_MODULE,
7409};
7410MODULE_ALIAS("bfq-iosched");
7411
7412static int __init bfq_init(void)
7413{
7414	int ret;
7415
7416#ifdef CONFIG_BFQ_GROUP_IOSCHED
7417	ret = blkcg_policy_register(&blkcg_policy_bfq);
7418	if (ret)
7419		return ret;
7420#endif
7421
7422	ret = -ENOMEM;
7423	if (bfq_slab_setup())
7424		goto err_pol_unreg;
7425
7426	/*
7427	 * Times to load large popular applications for the typical
7428	 * systems installed on the reference devices (see the
7429	 * comments before the definition of the next
7430	 * array). Actually, we use slightly lower values, as the
7431	 * estimated peak rate tends to be smaller than the actual
7432	 * peak rate.  The reason for this last fact is that estimates
7433	 * are computed over much shorter time intervals than the long
7434	 * intervals typically used for benchmarking. Why? First, to
7435	 * adapt more quickly to variations. Second, because an I/O
7436	 * scheduler cannot rely on a peak-rate-evaluation workload to
7437	 * be run for a long time.
7438	 */
7439	ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
7440	ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7441
7442	ret = elv_register(&iosched_bfq_mq);
7443	if (ret)
7444		goto slab_kill;
7445
7446	return 0;
7447
7448slab_kill:
7449	bfq_slab_kill();
7450err_pol_unreg:
7451#ifdef CONFIG_BFQ_GROUP_IOSCHED
7452	blkcg_policy_unregister(&blkcg_policy_bfq);
7453#endif
7454	return ret;
7455}
7456
7457static void __exit bfq_exit(void)
7458{
7459	elv_unregister(&iosched_bfq_mq);
7460#ifdef CONFIG_BFQ_GROUP_IOSCHED
7461	blkcg_policy_unregister(&blkcg_policy_bfq);
7462#endif
7463	bfq_slab_kill();
7464}
7465
7466module_init(bfq_init);
7467module_exit(bfq_exit);
7468
7469MODULE_AUTHOR("Paolo Valente");
7470MODULE_LICENSE("GPL");
7471MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");
v4.17
 
   1/*
   2 * Budget Fair Queueing (BFQ) I/O scheduler.
   3 *
   4 * Based on ideas and code from CFQ:
   5 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
   6 *
   7 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
   8 *		      Paolo Valente <paolo.valente@unimore.it>
   9 *
  10 * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
  11 *                    Arianna Avanzini <avanzini@google.com>
  12 *
  13 * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
  14 *
  15 *  This program is free software; you can redistribute it and/or
  16 *  modify it under the terms of the GNU General Public License as
  17 *  published by the Free Software Foundation; either version 2 of the
  18 *  License, or (at your option) any later version.
  19 *
  20 *  This program is distributed in the hope that it will be useful,
  21 *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  22 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  23 *  General Public License for more details.
  24 *
  25 * BFQ is a proportional-share I/O scheduler, with some extra
  26 * low-latency capabilities. BFQ also supports full hierarchical
  27 * scheduling through cgroups. Next paragraphs provide an introduction
  28 * on BFQ inner workings. Details on BFQ benefits, usage and
  29 * limitations can be found in Documentation/block/bfq-iosched.txt.
  30 *
  31 * BFQ is a proportional-share storage-I/O scheduling algorithm based
  32 * on the slice-by-slice service scheme of CFQ. But BFQ assigns
  33 * budgets, measured in number of sectors, to processes instead of
  34 * time slices. The device is not granted to the in-service process
  35 * for a given time slice, but until it has exhausted its assigned
  36 * budget. This change from the time to the service domain enables BFQ
  37 * to distribute the device throughput among processes as desired,
  38 * without any distortion due to throughput fluctuations, or to device
  39 * internal queueing. BFQ uses an ad hoc internal scheduler, called
  40 * B-WF2Q+, to schedule processes according to their budgets. More
  41 * precisely, BFQ schedules queues associated with processes. Each
  42 * process/queue is assigned a user-configurable weight, and B-WF2Q+
  43 * guarantees that each queue receives a fraction of the throughput
  44 * proportional to its weight. Thanks to the accurate policy of
  45 * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
  46 * processes issuing sequential requests (to boost the throughput),
  47 * and yet guarantee a low latency to interactive and soft real-time
  48 * applications.
  49 *
  50 * In particular, to provide these low-latency guarantees, BFQ
  51 * explicitly privileges the I/O of two classes of time-sensitive
  52 * applications: interactive and soft real-time. This feature enables
  53 * BFQ to provide applications in these classes with a very low
  54 * latency. Finally, BFQ also features additional heuristics for
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  55 * preserving both a low latency and a high throughput on NCQ-capable,
  56 * rotational or flash-based devices, and to get the job done quickly
  57 * for applications consisting in many I/O-bound processes.
  58 *
  59 * NOTE: if the main or only goal, with a given device, is to achieve
  60 * the maximum-possible throughput at all times, then do switch off
  61 * all low-latency heuristics for that device, by setting low_latency
  62 * to 0.
  63 *
  64 * BFQ is described in [1], where also a reference to the initial, more
  65 * theoretical paper on BFQ can be found. The interested reader can find
  66 * in the latter paper full details on the main algorithm, as well as
  67 * formulas of the guarantees and formal proofs of all the properties.
  68 * With respect to the version of BFQ presented in these papers, this
  69 * implementation adds a few more heuristics, such as the one that
  70 * guarantees a low latency to soft real-time applications, and a
  71 * hierarchical extension based on H-WF2Q+.
  72 *
  73 * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
  74 * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
  75 * with O(log N) complexity derives from the one introduced with EEVDF
  76 * in [3].
  77 *
  78 * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
  79 *     Scheduler", Proceedings of the First Workshop on Mobile System
  80 *     Technologies (MST-2015), May 2015.
  81 *     http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
  82 *
  83 * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
  84 *     Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
  85 *     Oct 1997.
  86 *
  87 * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
  88 *
  89 * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
  90 *     First: A Flexible and Accurate Mechanism for Proportional Share
  91 *     Resource Allocation", technical report.
  92 *
  93 * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
  94 */
  95#include <linux/module.h>
  96#include <linux/slab.h>
  97#include <linux/blkdev.h>
  98#include <linux/cgroup.h>
  99#include <linux/elevator.h>
 100#include <linux/ktime.h>
 101#include <linux/rbtree.h>
 102#include <linux/ioprio.h>
 103#include <linux/sbitmap.h>
 104#include <linux/delay.h>
 
 
 
 105
 
 106#include "blk.h"
 107#include "blk-mq.h"
 108#include "blk-mq-tag.h"
 109#include "blk-mq-sched.h"
 110#include "bfq-iosched.h"
 111#include "blk-wbt.h"
 112
 113#define BFQ_BFQQ_FNS(name)						\
 114void bfq_mark_bfqq_##name(struct bfq_queue *bfqq)			\
 115{									\
 116	__set_bit(BFQQF_##name, &(bfqq)->flags);			\
 117}									\
 118void bfq_clear_bfqq_##name(struct bfq_queue *bfqq)			\
 119{									\
 120	__clear_bit(BFQQF_##name, &(bfqq)->flags);		\
 121}									\
 122int bfq_bfqq_##name(const struct bfq_queue *bfqq)			\
 123{									\
 124	return test_bit(BFQQF_##name, &(bfqq)->flags);		\
 125}
 126
 127BFQ_BFQQ_FNS(just_created);
 128BFQ_BFQQ_FNS(busy);
 129BFQ_BFQQ_FNS(wait_request);
 130BFQ_BFQQ_FNS(non_blocking_wait_rq);
 131BFQ_BFQQ_FNS(fifo_expire);
 132BFQ_BFQQ_FNS(has_short_ttime);
 133BFQ_BFQQ_FNS(sync);
 134BFQ_BFQQ_FNS(IO_bound);
 135BFQ_BFQQ_FNS(in_large_burst);
 136BFQ_BFQQ_FNS(coop);
 137BFQ_BFQQ_FNS(split_coop);
 138BFQ_BFQQ_FNS(softrt_update);
 139#undef BFQ_BFQQ_FNS						\
 140
 141/* Expiration time of sync (0) and async (1) requests, in ns. */
 142static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
 143
 144/* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
 145static const int bfq_back_max = 16 * 1024;
 146
 147/* Penalty of a backwards seek, in number of sectors. */
 148static const int bfq_back_penalty = 2;
 149
 150/* Idling period duration, in ns. */
 151static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
 152
 153/* Minimum number of assigned budgets for which stats are safe to compute. */
 154static const int bfq_stats_min_budgets = 194;
 155
 156/* Default maximum budget values, in sectors and number of requests. */
 157static const int bfq_default_max_budget = 16 * 1024;
 158
 159/*
 160 * Async to sync throughput distribution is controlled as follows:
 161 * when an async request is served, the entity is charged the number
 162 * of sectors of the request, multiplied by the factor below
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 163 */
 164static const int bfq_async_charge_factor = 10;
 165
 166/* Default timeout values, in jiffies, approximating CFQ defaults. */
 167const int bfq_timeout = HZ / 8;
 168
 169/*
 170 * Time limit for merging (see comments in bfq_setup_cooperator). Set
 171 * to the slowest value that, in our tests, proved to be effective in
 172 * removing false positives, while not causing true positives to miss
 173 * queue merging.
 174 *
 175 * As can be deduced from the low time limit below, queue merging, if
 176 * successful, happens at the very beggining of the I/O of the involved
 177 * cooperating processes, as a consequence of the arrival of the very
 178 * first requests from each cooperator.  After that, there is very
 179 * little chance to find cooperators.
 180 */
 181static const unsigned long bfq_merge_time_limit = HZ/10;
 182
 183static struct kmem_cache *bfq_pool;
 184
 185/* Below this threshold (in ns), we consider thinktime immediate. */
 186#define BFQ_MIN_TT		(2 * NSEC_PER_MSEC)
 187
 188/* hw_tag detection: parallel requests threshold and min samples needed. */
 189#define BFQ_HW_QUEUE_THRESHOLD	4
 190#define BFQ_HW_QUEUE_SAMPLES	32
 191
 192#define BFQQ_SEEK_THR		(sector_t)(8 * 100)
 193#define BFQQ_SECT_THR_NONROT	(sector_t)(2 * 32)
 
 
 
 
 
 194#define BFQQ_CLOSE_THR		(sector_t)(8 * 1024)
 195#define BFQQ_SEEKY(bfqq)	(hweight32(bfqq->seek_history) > 19)
 
 
 
 
 
 
 
 
 196
 197/* Min number of samples required to perform peak-rate update */
 198#define BFQ_RATE_MIN_SAMPLES	32
 199/* Min observation time interval required to perform a peak-rate update (ns) */
 200#define BFQ_RATE_MIN_INTERVAL	(300*NSEC_PER_MSEC)
 201/* Target observation time interval for a peak-rate update (ns) */
 202#define BFQ_RATE_REF_INTERVAL	NSEC_PER_SEC
 203
 204/*
 205 * Shift used for peak-rate fixed precision calculations.
 206 * With
 207 * - the current shift: 16 positions
 208 * - the current type used to store rate: u32
 209 * - the current unit of measure for rate: [sectors/usec], or, more precisely,
 210 *   [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
 211 * the range of rates that can be stored is
 212 * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
 213 * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
 214 * [15, 65G] sectors/sec
 215 * Which, assuming a sector size of 512B, corresponds to a range of
 216 * [7.5K, 33T] B/sec
 217 */
 218#define BFQ_RATE_SHIFT		16
 219
 220/*
 221 * By default, BFQ computes the duration of the weight raising for
 222 * interactive applications automatically, using the following formula:
 223 * duration = (R / r) * T, where r is the peak rate of the device, and
 224 * R and T are two reference parameters.
 225 * In particular, R is the peak rate of the reference device (see
 226 * below), and T is a reference time: given the systems that are
 227 * likely to be installed on the reference device according to its
 228 * speed class, T is about the maximum time needed, under BFQ and
 229 * while reading two files in parallel, to load typical large
 230 * applications on these systems (see the comments on
 231 * max_service_from_wr below, for more details on how T is obtained).
 232 * In practice, the slower/faster the device at hand is, the more/less
 233 * it takes to load applications with respect to the reference device.
 234 * Accordingly, the longer/shorter BFQ grants weight raising to
 235 * interactive applications.
 236 *
 237 * BFQ uses four different reference pairs (R, T), depending on:
 238 * . whether the device is rotational or non-rotational;
 239 * . whether the device is slow, such as old or portable HDDs, as well as
 240 *   SD cards, or fast, such as newer HDDs and SSDs.
 241 *
 242 * The device's speed class is dynamically (re)detected in
 243 * bfq_update_peak_rate() every time the estimated peak rate is updated.
 244 *
 245 * In the following definitions, R_slow[0]/R_fast[0] and
 246 * T_slow[0]/T_fast[0] are the reference values for a slow/fast
 247 * rotational device, whereas R_slow[1]/R_fast[1] and
 248 * T_slow[1]/T_fast[1] are the reference values for a slow/fast
 249 * non-rotational device. Finally, device_speed_thresh are the
 250 * thresholds used to switch between speed classes. The reference
 251 * rates are not the actual peak rates of the devices used as a
 252 * reference, but slightly lower values. The reason for using these
 253 * slightly lower values is that the peak-rate estimator tends to
 254 * yield slightly lower values than the actual peak rate (it can yield
 255 * the actual peak rate only if there is only one process doing I/O,
 256 * and the process does sequential I/O).
 257 *
 258 * Both the reference peak rates and the thresholds are measured in
 259 * sectors/usec, left-shifted by BFQ_RATE_SHIFT.
 260 */
 261static int R_slow[2] = {1000, 10700};
 262static int R_fast[2] = {14000, 33000};
 263/*
 264 * To improve readability, a conversion function is used to initialize the
 265 * following arrays, which entails that they can be initialized only in a
 266 * function.
 267 */
 268static int T_slow[2];
 269static int T_fast[2];
 270static int device_speed_thresh[2];
 271
 272/*
 273 * BFQ uses the above-detailed, time-based weight-raising mechanism to
 274 * privilege interactive tasks. This mechanism is vulnerable to the
 275 * following false positives: I/O-bound applications that will go on
 276 * doing I/O for much longer than the duration of weight
 277 * raising. These applications have basically no benefit from being
 278 * weight-raised at the beginning of their I/O. On the opposite end,
 279 * while being weight-raised, these applications
 280 * a) unjustly steal throughput to applications that may actually need
 281 * low latency;
 282 * b) make BFQ uselessly perform device idling; device idling results
 283 * in loss of device throughput with most flash-based storage, and may
 284 * increase latencies when used purposelessly.
 285 *
 286 * BFQ tries to reduce these problems, by adopting the following
 287 * countermeasure. To introduce this countermeasure, we need first to
 288 * finish explaining how the duration of weight-raising for
 289 * interactive tasks is computed.
 290 *
 291 * For a bfq_queue deemed as interactive, the duration of weight
 292 * raising is dynamically adjusted, as a function of the estimated
 293 * peak rate of the device, so as to be equal to the time needed to
 294 * execute the 'largest' interactive task we benchmarked so far. By
 295 * largest task, we mean the task for which each involved process has
 296 * to do more I/O than for any of the other tasks we benchmarked. This
 297 * reference interactive task is the start-up of LibreOffice Writer,
 298 * and in this task each process/bfq_queue needs to have at most ~110K
 299 * sectors transferred.
 300 *
 301 * This last piece of information enables BFQ to reduce the actual
 302 * duration of weight-raising for at least one class of I/O-bound
 303 * applications: those doing sequential or quasi-sequential I/O. An
 304 * example is file copy. In fact, once started, the main I/O-bound
 305 * processes of these applications usually consume the above 110K
 306 * sectors in much less time than the processes of an application that
 307 * is starting, because these I/O-bound processes will greedily devote
 308 * almost all their CPU cycles only to their target,
 309 * throughput-friendly I/O operations. This is even more true if BFQ
 310 * happens to be underestimating the device peak rate, and thus
 311 * overestimating the duration of weight raising. But, according to
 312 * our measurements, once transferred 110K sectors, these processes
 313 * have no right to be weight-raised any longer.
 314 *
 315 * Basing on the last consideration, BFQ ends weight-raising for a
 316 * bfq_queue if the latter happens to have received an amount of
 317 * service at least equal to the following constant. The constant is
 318 * set to slightly more than 110K, to have a minimum safety margin.
 319 *
 320 * This early ending of weight-raising reduces the amount of time
 321 * during which interactive false positives cause the two problems
 322 * described at the beginning of these comments.
 323 */
 324static const unsigned long max_service_from_wr = 120000;
 325
 326#define RQ_BIC(rq)		icq_to_bic((rq)->elv.priv[0])
 
 
 
 
 
 
 
 
 
 
 327#define RQ_BFQQ(rq)		((rq)->elv.priv[1])
 328
 329struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync)
 330{
 331	return bic->bfqq[is_sync];
 332}
 333
 
 
 334void bic_set_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq, bool is_sync)
 335{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 336	bic->bfqq[is_sync] = bfqq;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 337}
 338
 339struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
 340{
 341	return bic->icq.q->elevator->elevator_data;
 342}
 343
 344/**
 345 * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
 346 * @icq: the iocontext queue.
 347 */
 348static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
 349{
 350	/* bic->icq is the first member, %NULL will convert to %NULL */
 351	return container_of(icq, struct bfq_io_cq, icq);
 352}
 353
 354/**
 355 * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
 356 * @bfqd: the lookup key.
 357 * @ioc: the io_context of the process doing I/O.
 358 * @q: the request queue.
 359 */
 360static struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd,
 361					struct io_context *ioc,
 362					struct request_queue *q)
 363{
 364	if (ioc) {
 365		unsigned long flags;
 366		struct bfq_io_cq *icq;
 367
 368		spin_lock_irqsave(q->queue_lock, flags);
 369		icq = icq_to_bic(ioc_lookup_icq(ioc, q));
 370		spin_unlock_irqrestore(q->queue_lock, flags);
 371
 372		return icq;
 373	}
 
 374
 375	return NULL;
 376}
 377
 378/*
 379 * Scheduler run of queue, if there are requests pending and no one in the
 380 * driver that will restart queueing.
 381 */
 382void bfq_schedule_dispatch(struct bfq_data *bfqd)
 383{
 
 
 384	if (bfqd->queued != 0) {
 385		bfq_log(bfqd, "schedule dispatch");
 386		blk_mq_run_hw_queues(bfqd->queue, true);
 387	}
 388}
 389
 390#define bfq_class_idle(bfqq)	((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
 391#define bfq_class_rt(bfqq)	((bfqq)->ioprio_class == IOPRIO_CLASS_RT)
 392
 393#define bfq_sample_valid(samples)	((samples) > 80)
 394
 395/*
 396 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 397 * We choose the request that is closesr to the head right now.  Distance
 398 * behind the head is penalized and only allowed to a certain extent.
 399 */
 400static struct request *bfq_choose_req(struct bfq_data *bfqd,
 401				      struct request *rq1,
 402				      struct request *rq2,
 403				      sector_t last)
 404{
 405	sector_t s1, s2, d1 = 0, d2 = 0;
 406	unsigned long back_max;
 407#define BFQ_RQ1_WRAP	0x01 /* request 1 wraps */
 408#define BFQ_RQ2_WRAP	0x02 /* request 2 wraps */
 409	unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
 410
 411	if (!rq1 || rq1 == rq2)
 412		return rq2;
 413	if (!rq2)
 414		return rq1;
 415
 416	if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 417		return rq1;
 418	else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
 419		return rq2;
 420	if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
 421		return rq1;
 422	else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
 423		return rq2;
 424
 425	s1 = blk_rq_pos(rq1);
 426	s2 = blk_rq_pos(rq2);
 427
 428	/*
 429	 * By definition, 1KiB is 2 sectors.
 430	 */
 431	back_max = bfqd->bfq_back_max * 2;
 432
 433	/*
 434	 * Strict one way elevator _except_ in the case where we allow
 435	 * short backward seeks which are biased as twice the cost of a
 436	 * similar forward seek.
 437	 */
 438	if (s1 >= last)
 439		d1 = s1 - last;
 440	else if (s1 + back_max >= last)
 441		d1 = (last - s1) * bfqd->bfq_back_penalty;
 442	else
 443		wrap |= BFQ_RQ1_WRAP;
 444
 445	if (s2 >= last)
 446		d2 = s2 - last;
 447	else if (s2 + back_max >= last)
 448		d2 = (last - s2) * bfqd->bfq_back_penalty;
 449	else
 450		wrap |= BFQ_RQ2_WRAP;
 451
 452	/* Found required data */
 453
 454	/*
 455	 * By doing switch() on the bit mask "wrap" we avoid having to
 456	 * check two variables for all permutations: --> faster!
 457	 */
 458	switch (wrap) {
 459	case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
 460		if (d1 < d2)
 461			return rq1;
 462		else if (d2 < d1)
 463			return rq2;
 464
 465		if (s1 >= s2)
 466			return rq1;
 467		else
 468			return rq2;
 469
 470	case BFQ_RQ2_WRAP:
 471		return rq1;
 472	case BFQ_RQ1_WRAP:
 473		return rq2;
 474	case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
 475	default:
 476		/*
 477		 * Since both rqs are wrapped,
 478		 * start with the one that's further behind head
 479		 * (--> only *one* back seek required),
 480		 * since back seek takes more time than forward.
 481		 */
 482		if (s1 <= s2)
 483			return rq1;
 484		else
 485			return rq2;
 486	}
 487}
 488
 489/*
 490 * See the comments on bfq_limit_depth for the purpose of
 491 * the depths set in the function.
 492 */
 493static void bfq_update_depths(struct bfq_data *bfqd, struct sbitmap_queue *bt)
 494{
 495	bfqd->sb_shift = bt->sb.shift;
 
 
 
 
 
 
 
 
 
 
 
 496
 497	/*
 498	 * In-word depths if no bfq_queue is being weight-raised:
 499	 * leaving 25% of tags only for sync reads.
 500	 *
 501	 * In next formulas, right-shift the value
 502	 * (1U<<bfqd->sb_shift), instead of computing directly
 503	 * (1U<<(bfqd->sb_shift - something)), to be robust against
 504	 * any possible value of bfqd->sb_shift, without having to
 505	 * limit 'something'.
 506	 */
 507	/* no more than 50% of tags for async I/O */
 508	bfqd->word_depths[0][0] = max((1U<<bfqd->sb_shift)>>1, 1U);
 509	/*
 510	 * no more than 75% of tags for sync writes (25% extra tags
 511	 * w.r.t. async I/O, to prevent async I/O from starving sync
 512	 * writes)
 513	 */
 514	bfqd->word_depths[0][1] = max(((1U<<bfqd->sb_shift) * 3)>>2, 1U);
 515
 516	/*
 517	 * In-word depths in case some bfq_queue is being weight-
 518	 * raised: leaving ~63% of tags for sync reads. This is the
 519	 * highest percentage for which, in our tests, application
 520	 * start-up times didn't suffer from any regression due to tag
 521	 * shortage.
 522	 */
 523	/* no more than ~18% of tags for async I/O */
 524	bfqd->word_depths[1][0] = max(((1U<<bfqd->sb_shift) * 3)>>4, 1U);
 525	/* no more than ~37% of tags for sync writes (~20% extra tags) */
 526	bfqd->word_depths[1][1] = max(((1U<<bfqd->sb_shift) * 6)>>4, 1U);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 527}
 
 528
 529/*
 530 * Async I/O can easily starve sync I/O (both sync reads and sync
 531 * writes), by consuming all tags. Similarly, storms of sync writes,
 532 * such as those that sync(2) may trigger, can starve sync reads.
 533 * Limit depths of async I/O and sync writes so as to counter both
 534 * problems.
 
 
 
 
 
 
 
 
 535 */
 536static void bfq_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
 537{
 538	struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
 539	struct bfq_data *bfqd = data->q->elevator->elevator_data;
 540	struct sbitmap_queue *bt;
 
 
 
 
 
 
 
 
 
 
 
 541
 542	if (op_is_sync(op) && !op_is_write(op))
 543		return;
 544
 545	if (data->flags & BLK_MQ_REQ_RESERVED) {
 546		if (unlikely(!tags->nr_reserved_tags)) {
 547			WARN_ON_ONCE(1);
 548			return;
 549		}
 550		bt = &tags->breserved_tags;
 551	} else
 552		bt = &tags->bitmap_tags;
 553
 554	if (unlikely(bfqd->sb_shift != bt->sb.shift))
 555		bfq_update_depths(bfqd, bt);
 556
 557	data->shallow_depth =
 558		bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(op)];
 559
 560	bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
 561			__func__, bfqd->wr_busy_queues, op_is_sync(op),
 562			data->shallow_depth);
 
 563}
 564
 565static struct bfq_queue *
 566bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
 567		     sector_t sector, struct rb_node **ret_parent,
 568		     struct rb_node ***rb_link)
 569{
 570	struct rb_node **p, *parent;
 571	struct bfq_queue *bfqq = NULL;
 572
 573	parent = NULL;
 574	p = &root->rb_node;
 575	while (*p) {
 576		struct rb_node **n;
 577
 578		parent = *p;
 579		bfqq = rb_entry(parent, struct bfq_queue, pos_node);
 580
 581		/*
 582		 * Sort strictly based on sector. Smallest to the left,
 583		 * largest to the right.
 584		 */
 585		if (sector > blk_rq_pos(bfqq->next_rq))
 586			n = &(*p)->rb_right;
 587		else if (sector < blk_rq_pos(bfqq->next_rq))
 588			n = &(*p)->rb_left;
 589		else
 590			break;
 591		p = n;
 592		bfqq = NULL;
 593	}
 594
 595	*ret_parent = parent;
 596	if (rb_link)
 597		*rb_link = p;
 598
 599	bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
 600		(unsigned long long)sector,
 601		bfqq ? bfqq->pid : 0);
 602
 603	return bfqq;
 604}
 605
 606static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
 607{
 608	return bfqq->service_from_backlogged > 0 &&
 609		time_is_before_jiffies(bfqq->first_IO_time +
 610				       bfq_merge_time_limit);
 611}
 612
 613void bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
 
 
 
 
 
 
 
 
 
 614{
 615	struct rb_node **p, *parent;
 616	struct bfq_queue *__bfqq;
 617
 618	if (bfqq->pos_root) {
 619		rb_erase(&bfqq->pos_node, bfqq->pos_root);
 620		bfqq->pos_root = NULL;
 621	}
 622
 
 
 
 
 623	/*
 624	 * bfqq cannot be merged any longer (see comments in
 625	 * bfq_setup_cooperator): no point in adding bfqq into the
 626	 * position tree.
 627	 */
 628	if (bfq_too_late_for_merging(bfqq))
 629		return;
 630
 631	if (bfq_class_idle(bfqq))
 632		return;
 633	if (!bfqq->next_rq)
 634		return;
 635
 636	bfqq->pos_root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
 637	__bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
 638			blk_rq_pos(bfqq->next_rq), &parent, &p);
 639	if (!__bfqq) {
 640		rb_link_node(&bfqq->pos_node, parent, p);
 641		rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
 642	} else
 643		bfqq->pos_root = NULL;
 644}
 645
 646/*
 647 * Tell whether there are active queues or groups with differentiated weights.
 648 */
 649static bool bfq_differentiated_weights(struct bfq_data *bfqd)
 650{
 651	/*
 652	 * For weights to differ, at least one of the trees must contain
 653	 * at least two nodes.
 654	 */
 655	return (!RB_EMPTY_ROOT(&bfqd->queue_weights_tree) &&
 656		(bfqd->queue_weights_tree.rb_node->rb_left ||
 657		 bfqd->queue_weights_tree.rb_node->rb_right)
 658#ifdef CONFIG_BFQ_GROUP_IOSCHED
 659	       ) ||
 660	       (!RB_EMPTY_ROOT(&bfqd->group_weights_tree) &&
 661		(bfqd->group_weights_tree.rb_node->rb_left ||
 662		 bfqd->group_weights_tree.rb_node->rb_right)
 663#endif
 664	       );
 665}
 666
 667/*
 668 * The following function returns true if every queue must receive the
 669 * same share of the throughput (this condition is used when deciding
 670 * whether idling may be disabled, see the comments in the function
 671 * bfq_bfqq_may_idle()).
 672 *
 673 * Such a scenario occurs when:
 674 * 1) all active queues have the same weight,
 675 * 2) all active groups at the same level in the groups tree have the same
 
 676 *    weight,
 677 * 3) all active groups at the same level in the groups tree have the same
 678 *    number of children.
 679 *
 680 * Unfortunately, keeping the necessary state for evaluating exactly the
 681 * above symmetry conditions would be quite complex and time-consuming.
 682 * Therefore this function evaluates, instead, the following stronger
 683 * sub-conditions, for which it is much easier to maintain the needed
 684 * state:
 685 * 1) all active queues have the same weight,
 686 * 2) all active groups have the same weight,
 687 * 3) all active groups have at most one active child each.
 688 * In particular, the last two conditions are always true if hierarchical
 689 * support and the cgroups interface are not enabled, thus no state needs
 690 * to be maintained in this case.
 691 */
 692static bool bfq_symmetric_scenario(struct bfq_data *bfqd)
 
 693{
 694	return !bfq_differentiated_weights(bfqd);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 695}
 696
 697/*
 698 * If the weight-counter tree passed as input contains no counter for
 699 * the weight of the input entity, then add that counter; otherwise just
 700 * increment the existing counter.
 701 *
 702 * Note that weight-counter trees contain few nodes in mostly symmetric
 703 * scenarios. For example, if all queues have the same weight, then the
 704 * weight-counter tree for the queues may contain at most one node.
 705 * This holds even if low_latency is on, because weight-raised queues
 706 * are not inserted in the tree.
 707 * In most scenarios, the rate at which nodes are created/destroyed
 708 * should be low too.
 709 */
 710void bfq_weights_tree_add(struct bfq_data *bfqd, struct bfq_entity *entity,
 711			  struct rb_root *root)
 712{
 713	struct rb_node **new = &(root->rb_node), *parent = NULL;
 
 
 
 714
 715	/*
 716	 * Do not insert if the entity is already associated with a
 717	 * counter, which happens if:
 718	 *   1) the entity is associated with a queue,
 719	 *   2) a request arrival has caused the queue to become both
 720	 *      non-weight-raised, and hence change its weight, and
 721	 *      backlogged; in this respect, each of the two events
 722	 *      causes an invocation of this function,
 723	 *   3) this is the invocation of this function caused by the
 724	 *      second event. This second invocation is actually useless,
 725	 *      and we handle this fact by exiting immediately. More
 726	 *      efficient or clearer solutions might possibly be adopted.
 727	 */
 728	if (entity->weight_counter)
 729		return;
 730
 731	while (*new) {
 732		struct bfq_weight_counter *__counter = container_of(*new,
 733						struct bfq_weight_counter,
 734						weights_node);
 735		parent = *new;
 736
 737		if (entity->weight == __counter->weight) {
 738			entity->weight_counter = __counter;
 739			goto inc_counter;
 740		}
 741		if (entity->weight < __counter->weight)
 742			new = &((*new)->rb_left);
 743		else
 744			new = &((*new)->rb_right);
 
 
 745	}
 746
 747	entity->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
 748					 GFP_ATOMIC);
 749
 750	/*
 751	 * In the unlucky event of an allocation failure, we just
 752	 * exit. This will cause the weight of entity to not be
 753	 * considered in bfq_differentiated_weights, which, in its
 754	 * turn, causes the scenario to be deemed wrongly symmetric in
 755	 * case entity's weight would have been the only weight making
 756	 * the scenario asymmetric. On the bright side, no unbalance
 757	 * will however occur when entity becomes inactive again (the
 758	 * invocation of this function is triggered by an activation
 759	 * of entity). In fact, bfq_weights_tree_remove does nothing
 760	 * if !entity->weight_counter.
 761	 */
 762	if (unlikely(!entity->weight_counter))
 763		return;
 764
 765	entity->weight_counter->weight = entity->weight;
 766	rb_link_node(&entity->weight_counter->weights_node, parent, new);
 767	rb_insert_color(&entity->weight_counter->weights_node, root);
 
 768
 769inc_counter:
 770	entity->weight_counter->num_active++;
 
 771}
 772
 773/*
 774 * Decrement the weight counter associated with the entity, and, if the
 775 * counter reaches 0, remove the counter from the tree.
 776 * See the comments to the function bfq_weights_tree_add() for considerations
 777 * about overhead.
 778 */
 779void bfq_weights_tree_remove(struct bfq_data *bfqd, struct bfq_entity *entity,
 780			     struct rb_root *root)
 781{
 782	if (!entity->weight_counter)
 
 
 783		return;
 784
 785	entity->weight_counter->num_active--;
 786	if (entity->weight_counter->num_active > 0)
 
 787		goto reset_entity_pointer;
 788
 789	rb_erase(&entity->weight_counter->weights_node, root);
 790	kfree(entity->weight_counter);
 791
 792reset_entity_pointer:
 793	entity->weight_counter = NULL;
 
 794}
 795
 796/*
 797 * Return expired entry, or NULL to just start from scratch in rbtree.
 798 */
 799static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
 800				      struct request *last)
 801{
 802	struct request *rq;
 803
 804	if (bfq_bfqq_fifo_expire(bfqq))
 805		return NULL;
 806
 807	bfq_mark_bfqq_fifo_expire(bfqq);
 808
 809	rq = rq_entry_fifo(bfqq->fifo.next);
 810
 811	if (rq == last || ktime_get_ns() < rq->fifo_time)
 812		return NULL;
 813
 814	bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
 815	return rq;
 816}
 817
 818static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
 819					struct bfq_queue *bfqq,
 820					struct request *last)
 821{
 822	struct rb_node *rbnext = rb_next(&last->rb_node);
 823	struct rb_node *rbprev = rb_prev(&last->rb_node);
 824	struct request *next, *prev = NULL;
 825
 826	/* Follow expired path, else get first next available. */
 827	next = bfq_check_fifo(bfqq, last);
 828	if (next)
 829		return next;
 830
 831	if (rbprev)
 832		prev = rb_entry_rq(rbprev);
 833
 834	if (rbnext)
 835		next = rb_entry_rq(rbnext);
 836	else {
 837		rbnext = rb_first(&bfqq->sort_list);
 838		if (rbnext && rbnext != &last->rb_node)
 839			next = rb_entry_rq(rbnext);
 840	}
 841
 842	return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
 843}
 844
 845/* see the definition of bfq_async_charge_factor for details */
 846static unsigned long bfq_serv_to_charge(struct request *rq,
 847					struct bfq_queue *bfqq)
 848{
 849	if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1)
 
 850		return blk_rq_sectors(rq);
 851
 852	/*
 853	 * If there are no weight-raised queues, then amplify service
 854	 * by just the async charge factor; otherwise amplify service
 855	 * by twice the async charge factor, to further reduce latency
 856	 * for weight-raised queues.
 857	 */
 858	if (bfqq->bfqd->wr_busy_queues == 0)
 859		return blk_rq_sectors(rq) * bfq_async_charge_factor;
 860
 861	return blk_rq_sectors(rq) * 2 * bfq_async_charge_factor;
 862}
 863
 864/**
 865 * bfq_updated_next_req - update the queue after a new next_rq selection.
 866 * @bfqd: the device data the queue belongs to.
 867 * @bfqq: the queue to update.
 868 *
 869 * If the first request of a queue changes we make sure that the queue
 870 * has enough budget to serve at least its first request (if the
 871 * request has grown).  We do this because if the queue has not enough
 872 * budget for its first request, it has to go through two dispatch
 873 * rounds to actually get it dispatched.
 874 */
 875static void bfq_updated_next_req(struct bfq_data *bfqd,
 876				 struct bfq_queue *bfqq)
 877{
 878	struct bfq_entity *entity = &bfqq->entity;
 879	struct request *next_rq = bfqq->next_rq;
 880	unsigned long new_budget;
 881
 882	if (!next_rq)
 883		return;
 884
 885	if (bfqq == bfqd->in_service_queue)
 886		/*
 887		 * In order not to break guarantees, budgets cannot be
 888		 * changed after an entity has been selected.
 889		 */
 890		return;
 891
 892	new_budget = max_t(unsigned long, bfqq->max_budget,
 893			   bfq_serv_to_charge(next_rq, bfqq));
 
 
 894	if (entity->budget != new_budget) {
 895		entity->budget = new_budget;
 896		bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
 897					 new_budget);
 898		bfq_requeue_bfqq(bfqd, bfqq, false);
 899	}
 900}
 901
 902static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
 903{
 904	u64 dur;
 905
 906	if (bfqd->bfq_wr_max_time > 0)
 907		return bfqd->bfq_wr_max_time;
 908
 909	dur = bfqd->RT_prod;
 910	do_div(dur, bfqd->peak_rate);
 911
 912	/*
 913	 * Limit duration between 3 and 13 seconds. Tests show that
 914	 * higher values than 13 seconds often yield the opposite of
 915	 * the desired result, i.e., worsen responsiveness by letting
 916	 * non-interactive and non-soft-real-time applications
 917	 * preserve weight raising for a too long time interval.
 
 
 
 
 
 
 
 
 
 918	 *
 919	 * On the other end, lower values than 3 seconds make it
 920	 * difficult for most interactive tasks to complete their jobs
 921	 * before weight-raising finishes.
 922	 */
 923	if (dur > msecs_to_jiffies(13000))
 924		dur = msecs_to_jiffies(13000);
 925	else if (dur < msecs_to_jiffies(3000))
 926		dur = msecs_to_jiffies(3000);
 927
 928	return dur;
 929}
 930
 931/* switch back from soft real-time to interactive weight raising */
 932static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
 933					  struct bfq_data *bfqd)
 934{
 935	bfqq->wr_coeff = bfqd->bfq_wr_coeff;
 936	bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
 937	bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
 938}
 939
 940static void
 941bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
 942		      struct bfq_io_cq *bic, bool bfq_already_existing)
 943{
 944	unsigned int old_wr_coeff = bfqq->wr_coeff;
 945	bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
 946
 947	if (bic->saved_has_short_ttime)
 948		bfq_mark_bfqq_has_short_ttime(bfqq);
 949	else
 950		bfq_clear_bfqq_has_short_ttime(bfqq);
 951
 952	if (bic->saved_IO_bound)
 953		bfq_mark_bfqq_IO_bound(bfqq);
 954	else
 955		bfq_clear_bfqq_IO_bound(bfqq);
 956
 
 
 
 
 
 957	bfqq->ttime = bic->saved_ttime;
 958	bfqq->wr_coeff = bic->saved_wr_coeff;
 
 
 
 
 
 
 
 
 
 959	bfqq->wr_start_at_switch_to_srt = bic->saved_wr_start_at_switch_to_srt;
 960	bfqq->last_wr_start_finish = bic->saved_last_wr_start_finish;
 961	bfqq->wr_cur_max_time = bic->saved_wr_cur_max_time;
 962
 963	if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
 964	    time_is_before_jiffies(bfqq->last_wr_start_finish +
 965				   bfqq->wr_cur_max_time))) {
 966		if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
 967		    !bfq_bfqq_in_large_burst(bfqq) &&
 968		    time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
 969					     bfq_wr_duration(bfqd))) {
 970			switch_back_to_interactive_wr(bfqq, bfqd);
 971		} else {
 972			bfqq->wr_coeff = 1;
 973			bfq_log_bfqq(bfqq->bfqd, bfqq,
 974				     "resume state: switching off wr");
 975		}
 976	}
 977
 978	/* make sure weight will be updated, however we got here */
 979	bfqq->entity.prio_changed = 1;
 980
 981	if (likely(!busy))
 982		return;
 983
 984	if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
 985		bfqd->wr_busy_queues++;
 986	else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
 987		bfqd->wr_busy_queues--;
 988}
 989
 990static int bfqq_process_refs(struct bfq_queue *bfqq)
 991{
 992	return bfqq->ref - bfqq->allocated - bfqq->entity.on_st;
 
 
 993}
 994
 995/* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
 996static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
 997{
 998	struct bfq_queue *item;
 999	struct hlist_node *n;
1000
1001	hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1002		hlist_del_init(&item->burst_list_node);
1003	hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1004	bfqd->burst_size = 1;
 
 
 
 
 
 
 
 
 
 
1005	bfqd->burst_parent_entity = bfqq->entity.parent;
1006}
1007
1008/* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1009static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1010{
1011	/* Increment burst size to take into account also bfqq */
1012	bfqd->burst_size++;
1013
1014	if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1015		struct bfq_queue *pos, *bfqq_item;
1016		struct hlist_node *n;
1017
1018		/*
1019		 * Enough queues have been activated shortly after each
1020		 * other to consider this burst as large.
1021		 */
1022		bfqd->large_burst = true;
1023
1024		/*
1025		 * We can now mark all queues in the burst list as
1026		 * belonging to a large burst.
1027		 */
1028		hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1029				     burst_list_node)
1030			bfq_mark_bfqq_in_large_burst(bfqq_item);
1031		bfq_mark_bfqq_in_large_burst(bfqq);
1032
1033		/*
1034		 * From now on, and until the current burst finishes, any
1035		 * new queue being activated shortly after the last queue
1036		 * was inserted in the burst can be immediately marked as
1037		 * belonging to a large burst. So the burst list is not
1038		 * needed any more. Remove it.
1039		 */
1040		hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1041					  burst_list_node)
1042			hlist_del_init(&pos->burst_list_node);
1043	} else /*
1044		* Burst not yet large: add bfqq to the burst list. Do
1045		* not increment the ref counter for bfqq, because bfqq
1046		* is removed from the burst list before freeing bfqq
1047		* in put_queue.
1048		*/
1049		hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1050}
1051
1052/*
1053 * If many queues belonging to the same group happen to be created
1054 * shortly after each other, then the processes associated with these
1055 * queues have typically a common goal. In particular, bursts of queue
1056 * creations are usually caused by services or applications that spawn
1057 * many parallel threads/processes. Examples are systemd during boot,
1058 * or git grep. To help these processes get their job done as soon as
1059 * possible, it is usually better to not grant either weight-raising
1060 * or device idling to their queues.
 
1061 *
1062 * In this comment we describe, firstly, the reasons why this fact
1063 * holds, and, secondly, the next function, which implements the main
1064 * steps needed to properly mark these queues so that they can then be
1065 * treated in a different way.
1066 *
1067 * The above services or applications benefit mostly from a high
1068 * throughput: the quicker the requests of the activated queues are
1069 * cumulatively served, the sooner the target job of these queues gets
1070 * completed. As a consequence, weight-raising any of these queues,
1071 * which also implies idling the device for it, is almost always
1072 * counterproductive. In most cases it just lowers throughput.
 
 
 
1073 *
1074 * On the other hand, a burst of queue creations may be caused also by
1075 * the start of an application that does not consist of a lot of
1076 * parallel I/O-bound threads. In fact, with a complex application,
1077 * several short processes may need to be executed to start-up the
1078 * application. In this respect, to start an application as quickly as
1079 * possible, the best thing to do is in any case to privilege the I/O
1080 * related to the application with respect to all other
1081 * I/O. Therefore, the best strategy to start as quickly as possible
1082 * an application that causes a burst of queue creations is to
1083 * weight-raise all the queues created during the burst. This is the
1084 * exact opposite of the best strategy for the other type of bursts.
1085 *
1086 * In the end, to take the best action for each of the two cases, the
1087 * two types of bursts need to be distinguished. Fortunately, this
1088 * seems relatively easy, by looking at the sizes of the bursts. In
1089 * particular, we found a threshold such that only bursts with a
1090 * larger size than that threshold are apparently caused by
1091 * services or commands such as systemd or git grep. For brevity,
1092 * hereafter we call just 'large' these bursts. BFQ *does not*
1093 * weight-raise queues whose creation occurs in a large burst. In
1094 * addition, for each of these queues BFQ performs or does not perform
1095 * idling depending on which choice boosts the throughput more. The
1096 * exact choice depends on the device and request pattern at
1097 * hand.
1098 *
1099 * Unfortunately, false positives may occur while an interactive task
1100 * is starting (e.g., an application is being started). The
1101 * consequence is that the queues associated with the task do not
1102 * enjoy weight raising as expected. Fortunately these false positives
1103 * are very rare. They typically occur if some service happens to
1104 * start doing I/O exactly when the interactive task starts.
1105 *
1106 * Turning back to the next function, it implements all the steps
1107 * needed to detect the occurrence of a large burst and to properly
1108 * mark all the queues belonging to it (so that they can then be
1109 * treated in a different way). This goal is achieved by maintaining a
1110 * "burst list" that holds, temporarily, the queues that belong to the
1111 * burst in progress. The list is then used to mark these queues as
1112 * belonging to a large burst if the burst does become large. The main
1113 * steps are the following.
 
 
1114 *
1115 * . when the very first queue is created, the queue is inserted into the
1116 *   list (as it could be the first queue in a possible burst)
1117 *
1118 * . if the current burst has not yet become large, and a queue Q that does
1119 *   not yet belong to the burst is activated shortly after the last time
1120 *   at which a new queue entered the burst list, then the function appends
1121 *   Q to the burst list
1122 *
1123 * . if, as a consequence of the previous step, the burst size reaches
1124 *   the large-burst threshold, then
1125 *
1126 *     . all the queues in the burst list are marked as belonging to a
1127 *       large burst
1128 *
1129 *     . the burst list is deleted; in fact, the burst list already served
1130 *       its purpose (keeping temporarily track of the queues in a burst,
1131 *       so as to be able to mark them as belonging to a large burst in the
1132 *       previous sub-step), and now is not needed any more
1133 *
1134 *     . the device enters a large-burst mode
1135 *
1136 * . if a queue Q that does not belong to the burst is created while
1137 *   the device is in large-burst mode and shortly after the last time
1138 *   at which a queue either entered the burst list or was marked as
1139 *   belonging to the current large burst, then Q is immediately marked
1140 *   as belonging to a large burst.
1141 *
1142 * . if a queue Q that does not belong to the burst is created a while
1143 *   later, i.e., not shortly after, than the last time at which a queue
1144 *   either entered the burst list or was marked as belonging to the
1145 *   current large burst, then the current burst is deemed as finished and:
1146 *
1147 *        . the large-burst mode is reset if set
1148 *
1149 *        . the burst list is emptied
1150 *
1151 *        . Q is inserted in the burst list, as Q may be the first queue
1152 *          in a possible new burst (then the burst list contains just Q
1153 *          after this step).
1154 */
1155static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1156{
1157	/*
1158	 * If bfqq is already in the burst list or is part of a large
1159	 * burst, or finally has just been split, then there is
1160	 * nothing else to do.
1161	 */
1162	if (!hlist_unhashed(&bfqq->burst_list_node) ||
1163	    bfq_bfqq_in_large_burst(bfqq) ||
1164	    time_is_after_eq_jiffies(bfqq->split_time +
1165				     msecs_to_jiffies(10)))
1166		return;
1167
1168	/*
1169	 * If bfqq's creation happens late enough, or bfqq belongs to
1170	 * a different group than the burst group, then the current
1171	 * burst is finished, and related data structures must be
1172	 * reset.
1173	 *
1174	 * In this respect, consider the special case where bfqq is
1175	 * the very first queue created after BFQ is selected for this
1176	 * device. In this case, last_ins_in_burst and
1177	 * burst_parent_entity are not yet significant when we get
1178	 * here. But it is easy to verify that, whether or not the
1179	 * following condition is true, bfqq will end up being
1180	 * inserted into the burst list. In particular the list will
1181	 * happen to contain only bfqq. And this is exactly what has
1182	 * to happen, as bfqq may be the first queue of the first
1183	 * burst.
1184	 */
1185	if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1186	    bfqd->bfq_burst_interval) ||
1187	    bfqq->entity.parent != bfqd->burst_parent_entity) {
1188		bfqd->large_burst = false;
1189		bfq_reset_burst_list(bfqd, bfqq);
1190		goto end;
1191	}
1192
1193	/*
1194	 * If we get here, then bfqq is being activated shortly after the
1195	 * last queue. So, if the current burst is also large, we can mark
1196	 * bfqq as belonging to this large burst immediately.
1197	 */
1198	if (bfqd->large_burst) {
1199		bfq_mark_bfqq_in_large_burst(bfqq);
1200		goto end;
1201	}
1202
1203	/*
1204	 * If we get here, then a large-burst state has not yet been
1205	 * reached, but bfqq is being activated shortly after the last
1206	 * queue. Then we add bfqq to the burst.
1207	 */
1208	bfq_add_to_burst(bfqd, bfqq);
1209end:
1210	/*
1211	 * At this point, bfqq either has been added to the current
1212	 * burst or has caused the current burst to terminate and a
1213	 * possible new burst to start. In particular, in the second
1214	 * case, bfqq has become the first queue in the possible new
1215	 * burst.  In both cases last_ins_in_burst needs to be moved
1216	 * forward.
1217	 */
1218	bfqd->last_ins_in_burst = jiffies;
1219}
1220
1221static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1222{
1223	struct bfq_entity *entity = &bfqq->entity;
1224
1225	return entity->budget - entity->service;
1226}
1227
1228/*
1229 * If enough samples have been computed, return the current max budget
1230 * stored in bfqd, which is dynamically updated according to the
1231 * estimated disk peak rate; otherwise return the default max budget
1232 */
1233static int bfq_max_budget(struct bfq_data *bfqd)
1234{
1235	if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1236		return bfq_default_max_budget;
1237	else
1238		return bfqd->bfq_max_budget;
1239}
1240
1241/*
1242 * Return min budget, which is a fraction of the current or default
1243 * max budget (trying with 1/32)
1244 */
1245static int bfq_min_budget(struct bfq_data *bfqd)
1246{
1247	if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1248		return bfq_default_max_budget / 32;
1249	else
1250		return bfqd->bfq_max_budget / 32;
1251}
1252
1253/*
1254 * The next function, invoked after the input queue bfqq switches from
1255 * idle to busy, updates the budget of bfqq. The function also tells
1256 * whether the in-service queue should be expired, by returning
1257 * true. The purpose of expiring the in-service queue is to give bfqq
1258 * the chance to possibly preempt the in-service queue, and the reason
1259 * for preempting the in-service queue is to achieve one of the two
1260 * goals below.
1261 *
1262 * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1263 * expired because it has remained idle. In particular, bfqq may have
1264 * expired for one of the following two reasons:
1265 *
1266 * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1267 *   and did not make it to issue a new request before its last
1268 *   request was served;
1269 *
1270 * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1271 *   a new request before the expiration of the idling-time.
1272 *
1273 * Even if bfqq has expired for one of the above reasons, the process
1274 * associated with the queue may be however issuing requests greedily,
1275 * and thus be sensitive to the bandwidth it receives (bfqq may have
1276 * remained idle for other reasons: CPU high load, bfqq not enjoying
1277 * idling, I/O throttling somewhere in the path from the process to
1278 * the I/O scheduler, ...). But if, after every expiration for one of
1279 * the above two reasons, bfqq has to wait for the service of at least
1280 * one full budget of another queue before being served again, then
1281 * bfqq is likely to get a much lower bandwidth or resource time than
1282 * its reserved ones. To address this issue, two countermeasures need
1283 * to be taken.
1284 *
1285 * First, the budget and the timestamps of bfqq need to be updated in
1286 * a special way on bfqq reactivation: they need to be updated as if
1287 * bfqq did not remain idle and did not expire. In fact, if they are
1288 * computed as if bfqq expired and remained idle until reactivation,
1289 * then the process associated with bfqq is treated as if, instead of
1290 * being greedy, it stopped issuing requests when bfqq remained idle,
1291 * and restarts issuing requests only on this reactivation. In other
1292 * words, the scheduler does not help the process recover the "service
1293 * hole" between bfqq expiration and reactivation. As a consequence,
1294 * the process receives a lower bandwidth than its reserved one. In
1295 * contrast, to recover this hole, the budget must be updated as if
1296 * bfqq was not expired at all before this reactivation, i.e., it must
1297 * be set to the value of the remaining budget when bfqq was
1298 * expired. Along the same line, timestamps need to be assigned the
1299 * value they had the last time bfqq was selected for service, i.e.,
1300 * before last expiration. Thus timestamps need to be back-shifted
1301 * with respect to their normal computation (see [1] for more details
1302 * on this tricky aspect).
1303 *
1304 * Secondly, to allow the process to recover the hole, the in-service
1305 * queue must be expired too, to give bfqq the chance to preempt it
1306 * immediately. In fact, if bfqq has to wait for a full budget of the
1307 * in-service queue to be completed, then it may become impossible to
1308 * let the process recover the hole, even if the back-shifted
1309 * timestamps of bfqq are lower than those of the in-service queue. If
1310 * this happens for most or all of the holes, then the process may not
1311 * receive its reserved bandwidth. In this respect, it is worth noting
1312 * that, being the service of outstanding requests unpreemptible, a
1313 * little fraction of the holes may however be unrecoverable, thereby
1314 * causing a little loss of bandwidth.
1315 *
1316 * The last important point is detecting whether bfqq does need this
1317 * bandwidth recovery. In this respect, the next function deems the
1318 * process associated with bfqq greedy, and thus allows it to recover
1319 * the hole, if: 1) the process is waiting for the arrival of a new
1320 * request (which implies that bfqq expired for one of the above two
1321 * reasons), and 2) such a request has arrived soon. The first
1322 * condition is controlled through the flag non_blocking_wait_rq,
1323 * while the second through the flag arrived_in_time. If both
1324 * conditions hold, then the function computes the budget in the
1325 * above-described special way, and signals that the in-service queue
1326 * should be expired. Timestamp back-shifting is done later in
1327 * __bfq_activate_entity.
1328 *
1329 * 2. Reduce latency. Even if timestamps are not backshifted to let
1330 * the process associated with bfqq recover a service hole, bfqq may
1331 * however happen to have, after being (re)activated, a lower finish
1332 * timestamp than the in-service queue.	 That is, the next budget of
1333 * bfqq may have to be completed before the one of the in-service
1334 * queue. If this is the case, then preempting the in-service queue
1335 * allows this goal to be achieved, apart from the unpreemptible,
1336 * outstanding requests mentioned above.
1337 *
1338 * Unfortunately, regardless of which of the above two goals one wants
1339 * to achieve, service trees need first to be updated to know whether
1340 * the in-service queue must be preempted. To have service trees
1341 * correctly updated, the in-service queue must be expired and
1342 * rescheduled, and bfqq must be scheduled too. This is one of the
1343 * most costly operations (in future versions, the scheduling
1344 * mechanism may be re-designed in such a way to make it possible to
1345 * know whether preemption is needed without needing to update service
1346 * trees). In addition, queue preemptions almost always cause random
1347 * I/O, and thus loss of throughput. Because of these facts, the next
1348 * function adopts the following simple scheme to avoid both costly
1349 * operations and too frequent preemptions: it requests the expiration
1350 * of the in-service queue (unconditionally) only for queues that need
1351 * to recover a hole, or that either are weight-raised or deserve to
1352 * be weight-raised.
 
 
 
1353 */
1354static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1355						struct bfq_queue *bfqq,
1356						bool arrived_in_time,
1357						bool wr_or_deserves_wr)
1358{
1359	struct bfq_entity *entity = &bfqq->entity;
1360
1361	if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time) {
 
 
 
 
 
 
 
 
1362		/*
1363		 * We do not clear the flag non_blocking_wait_rq here, as
1364		 * the latter is used in bfq_activate_bfqq to signal
1365		 * that timestamps need to be back-shifted (and is
1366		 * cleared right after).
1367		 */
1368
1369		/*
1370		 * In next assignment we rely on that either
1371		 * entity->service or entity->budget are not updated
1372		 * on expiration if bfqq is empty (see
1373		 * __bfq_bfqq_recalc_budget). Thus both quantities
1374		 * remain unchanged after such an expiration, and the
1375		 * following statement therefore assigns to
1376		 * entity->budget the remaining budget on such an
1377		 * expiration. For clarity, entity->service is not
1378		 * updated on expiration in any case, and, in normal
1379		 * operation, is reset only when bfqq is selected for
1380		 * service (see bfq_get_next_queue).
1381		 */
1382		entity->budget = min_t(unsigned long,
1383				       bfq_bfqq_budget_left(bfqq),
1384				       bfqq->max_budget);
1385
 
 
 
 
 
 
 
 
 
 
 
1386		return true;
1387	}
1388
 
 
 
 
1389	entity->budget = max_t(unsigned long, bfqq->max_budget,
1390			       bfq_serv_to_charge(bfqq->next_rq, bfqq));
1391	bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1392	return wr_or_deserves_wr;
1393}
1394
1395/*
1396 * Return the farthest future time instant according to jiffies
1397 * macros.
1398 */
1399static unsigned long bfq_greatest_from_now(void)
1400{
1401	return jiffies + MAX_JIFFY_OFFSET;
1402}
1403
1404/*
1405 * Return the farthest past time instant according to jiffies
1406 * macros.
1407 */
1408static unsigned long bfq_smallest_from_now(void)
1409{
1410	return jiffies - MAX_JIFFY_OFFSET;
1411}
1412
1413static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1414					     struct bfq_queue *bfqq,
1415					     unsigned int old_wr_coeff,
1416					     bool wr_or_deserves_wr,
1417					     bool interactive,
1418					     bool in_burst,
1419					     bool soft_rt)
1420{
1421	if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1422		/* start a weight-raising period */
1423		if (interactive) {
1424			bfqq->service_from_wr = 0;
1425			bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1426			bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1427		} else {
1428			/*
1429			 * No interactive weight raising in progress
1430			 * here: assign minus infinity to
1431			 * wr_start_at_switch_to_srt, to make sure
1432			 * that, at the end of the soft-real-time
1433			 * weight raising periods that is starting
1434			 * now, no interactive weight-raising period
1435			 * may be wrongly considered as still in
1436			 * progress (and thus actually started by
1437			 * mistake).
1438			 */
1439			bfqq->wr_start_at_switch_to_srt =
1440				bfq_smallest_from_now();
1441			bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1442				BFQ_SOFTRT_WEIGHT_FACTOR;
1443			bfqq->wr_cur_max_time =
1444				bfqd->bfq_wr_rt_max_time;
1445		}
1446
1447		/*
1448		 * If needed, further reduce budget to make sure it is
1449		 * close to bfqq's backlog, so as to reduce the
1450		 * scheduling-error component due to a too large
1451		 * budget. Do not care about throughput consequences,
1452		 * but only about latency. Finally, do not assign a
1453		 * too small budget either, to avoid increasing
1454		 * latency by causing too frequent expirations.
1455		 */
1456		bfqq->entity.budget = min_t(unsigned long,
1457					    bfqq->entity.budget,
1458					    2 * bfq_min_budget(bfqd));
1459	} else if (old_wr_coeff > 1) {
1460		if (interactive) { /* update wr coeff and duration */
1461			bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1462			bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1463		} else if (in_burst)
1464			bfqq->wr_coeff = 1;
1465		else if (soft_rt) {
1466			/*
1467			 * The application is now or still meeting the
1468			 * requirements for being deemed soft rt.  We
1469			 * can then correctly and safely (re)charge
1470			 * the weight-raising duration for the
1471			 * application with the weight-raising
1472			 * duration for soft rt applications.
1473			 *
1474			 * In particular, doing this recharge now, i.e.,
1475			 * before the weight-raising period for the
1476			 * application finishes, reduces the probability
1477			 * of the following negative scenario:
1478			 * 1) the weight of a soft rt application is
1479			 *    raised at startup (as for any newly
1480			 *    created application),
1481			 * 2) since the application is not interactive,
1482			 *    at a certain time weight-raising is
1483			 *    stopped for the application,
1484			 * 3) at that time the application happens to
1485			 *    still have pending requests, and hence
1486			 *    is destined to not have a chance to be
1487			 *    deemed soft rt before these requests are
1488			 *    completed (see the comments to the
1489			 *    function bfq_bfqq_softrt_next_start()
1490			 *    for details on soft rt detection),
1491			 * 4) these pending requests experience a high
1492			 *    latency because the application is not
1493			 *    weight-raised while they are pending.
1494			 */
1495			if (bfqq->wr_cur_max_time !=
1496				bfqd->bfq_wr_rt_max_time) {
1497				bfqq->wr_start_at_switch_to_srt =
1498					bfqq->last_wr_start_finish;
1499
1500				bfqq->wr_cur_max_time =
1501					bfqd->bfq_wr_rt_max_time;
1502				bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1503					BFQ_SOFTRT_WEIGHT_FACTOR;
1504			}
1505			bfqq->last_wr_start_finish = jiffies;
1506		}
1507	}
1508}
1509
1510static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1511					struct bfq_queue *bfqq)
1512{
1513	return bfqq->dispatched == 0 &&
1514		time_is_before_jiffies(
1515			bfqq->budget_timeout +
1516			bfqd->bfq_wr_min_idle_time);
1517}
1518
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1519static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1520					     struct bfq_queue *bfqq,
1521					     int old_wr_coeff,
1522					     struct request *rq,
1523					     bool *interactive)
1524{
1525	bool soft_rt, in_burst,	wr_or_deserves_wr,
1526		bfqq_wants_to_preempt,
1527		idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1528		/*
1529		 * See the comments on
1530		 * bfq_bfqq_update_budg_for_activation for
1531		 * details on the usage of the next variable.
1532		 */
1533		arrived_in_time =  ktime_get_ns() <=
1534			bfqq->ttime.last_end_request +
1535			bfqd->bfq_slice_idle * 3;
1536
1537
1538	/*
1539	 * bfqq deserves to be weight-raised if:
1540	 * - it is sync,
1541	 * - it does not belong to a large burst,
1542	 * - it has been idle for enough time or is soft real-time,
1543	 * - is linked to a bfq_io_cq (it is not shared in any sense).
 
 
1544	 */
1545	in_burst = bfq_bfqq_in_large_burst(bfqq);
1546	soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
 
1547		!in_burst &&
1548		time_is_before_jiffies(bfqq->soft_rt_next_start);
1549	*interactive = !in_burst && idle_for_long_time;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1550	wr_or_deserves_wr = bfqd->low_latency &&
1551		(bfqq->wr_coeff > 1 ||
1552		 (bfq_bfqq_sync(bfqq) &&
1553		  bfqq->bic && (*interactive || soft_rt)));
 
1554
1555	/*
1556	 * Using the last flag, update budget and check whether bfqq
1557	 * may want to preempt the in-service queue.
1558	 */
1559	bfqq_wants_to_preempt =
1560		bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1561						    arrived_in_time,
1562						    wr_or_deserves_wr);
1563
1564	/*
1565	 * If bfqq happened to be activated in a burst, but has been
1566	 * idle for much more than an interactive queue, then we
1567	 * assume that, in the overall I/O initiated in the burst, the
1568	 * I/O associated with bfqq is finished. So bfqq does not need
1569	 * to be treated as a queue belonging to a burst
1570	 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1571	 * if set, and remove bfqq from the burst list if it's
1572	 * there. We do not decrement burst_size, because the fact
1573	 * that bfqq does not need to belong to the burst list any
1574	 * more does not invalidate the fact that bfqq was created in
1575	 * a burst.
1576	 */
1577	if (likely(!bfq_bfqq_just_created(bfqq)) &&
1578	    idle_for_long_time &&
1579	    time_is_before_jiffies(
1580		    bfqq->budget_timeout +
1581		    msecs_to_jiffies(10000))) {
1582		hlist_del_init(&bfqq->burst_list_node);
1583		bfq_clear_bfqq_in_large_burst(bfqq);
1584	}
1585
1586	bfq_clear_bfqq_just_created(bfqq);
1587
1588
1589	if (!bfq_bfqq_IO_bound(bfqq)) {
1590		if (arrived_in_time) {
1591			bfqq->requests_within_timer++;
1592			if (bfqq->requests_within_timer >=
1593			    bfqd->bfq_requests_within_timer)
1594				bfq_mark_bfqq_IO_bound(bfqq);
1595		} else
1596			bfqq->requests_within_timer = 0;
1597	}
1598
1599	if (bfqd->low_latency) {
1600		if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1601			/* wraparound */
1602			bfqq->split_time =
1603				jiffies - bfqd->bfq_wr_min_idle_time - 1;
1604
1605		if (time_is_before_jiffies(bfqq->split_time +
1606					   bfqd->bfq_wr_min_idle_time)) {
1607			bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1608							 old_wr_coeff,
1609							 wr_or_deserves_wr,
1610							 *interactive,
1611							 in_burst,
1612							 soft_rt);
1613
1614			if (old_wr_coeff != bfqq->wr_coeff)
1615				bfqq->entity.prio_changed = 1;
1616		}
1617	}
1618
1619	bfqq->last_idle_bklogged = jiffies;
1620	bfqq->service_from_backlogged = 0;
1621	bfq_clear_bfqq_softrt_update(bfqq);
1622
1623	bfq_add_bfqq_busy(bfqd, bfqq);
1624
1625	/*
1626	 * Expire in-service queue only if preemption may be needed
1627	 * for guarantees. In this respect, the function
1628	 * next_queue_may_preempt just checks a simple, necessary
1629	 * condition, and not a sufficient condition based on
1630	 * timestamps. In fact, for the latter condition to be
1631	 * evaluated, timestamps would need first to be updated, and
1632	 * this operation is quite costly (see the comments on the
1633	 * function bfq_bfqq_update_budg_for_activation).
1634	 */
1635	if (bfqd->in_service_queue && bfqq_wants_to_preempt &&
1636	    bfqd->in_service_queue->wr_coeff < bfqq->wr_coeff &&
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1637	    next_queue_may_preempt(bfqd))
1638		bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1639				false, BFQQE_PREEMPTED);
1640}
1641
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1642static void bfq_add_request(struct request *rq)
1643{
1644	struct bfq_queue *bfqq = RQ_BFQQ(rq);
1645	struct bfq_data *bfqd = bfqq->bfqd;
1646	struct request *next_rq, *prev;
1647	unsigned int old_wr_coeff = bfqq->wr_coeff;
1648	bool interactive = false;
 
1649
1650	bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
1651	bfqq->queued[rq_is_sync(rq)]++;
1652	bfqd->queued++;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1653
1654	elv_rb_add(&bfqq->sort_list, rq);
1655
1656	/*
1657	 * Check if this request is a better next-serve candidate.
1658	 */
1659	prev = bfqq->next_rq;
1660	next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
1661	bfqq->next_rq = next_rq;
1662
1663	/*
1664	 * Adjust priority tree position, if next_rq changes.
 
1665	 */
1666	if (prev != bfqq->next_rq)
1667		bfq_pos_tree_add_move(bfqd, bfqq);
1668
1669	if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
1670		bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
1671						 rq, &interactive);
1672	else {
1673		if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
1674		    time_is_before_jiffies(
1675				bfqq->last_wr_start_finish +
1676				bfqd->bfq_wr_min_inter_arr_async)) {
1677			bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1678			bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1679
1680			bfqd->wr_busy_queues++;
1681			bfqq->entity.prio_changed = 1;
1682		}
1683		if (prev != bfqq->next_rq)
1684			bfq_updated_next_req(bfqd, bfqq);
1685	}
1686
1687	/*
1688	 * Assign jiffies to last_wr_start_finish in the following
1689	 * cases:
1690	 *
1691	 * . if bfqq is not going to be weight-raised, because, for
1692	 *   non weight-raised queues, last_wr_start_finish stores the
1693	 *   arrival time of the last request; as of now, this piece
1694	 *   of information is used only for deciding whether to
1695	 *   weight-raise async queues
1696	 *
1697	 * . if bfqq is not weight-raised, because, if bfqq is now
1698	 *   switching to weight-raised, then last_wr_start_finish
1699	 *   stores the time when weight-raising starts
1700	 *
1701	 * . if bfqq is interactive, because, regardless of whether
1702	 *   bfqq is currently weight-raised, the weight-raising
1703	 *   period must start or restart (this case is considered
1704	 *   separately because it is not detected by the above
1705	 *   conditions, if bfqq is already weight-raised)
1706	 *
1707	 * last_wr_start_finish has to be updated also if bfqq is soft
1708	 * real-time, because the weight-raising period is constantly
1709	 * restarted on idle-to-busy transitions for these queues, but
1710	 * this is already done in bfq_bfqq_handle_idle_busy_switch if
1711	 * needed.
1712	 */
1713	if (bfqd->low_latency &&
1714		(old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
1715		bfqq->last_wr_start_finish = jiffies;
1716}
1717
1718static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
1719					  struct bio *bio,
1720					  struct request_queue *q)
1721{
1722	struct bfq_queue *bfqq = bfqd->bio_bfqq;
1723
1724
1725	if (bfqq)
1726		return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
1727
1728	return NULL;
1729}
1730
1731static sector_t get_sdist(sector_t last_pos, struct request *rq)
1732{
1733	if (last_pos)
1734		return abs(blk_rq_pos(rq) - last_pos);
1735
1736	return 0;
1737}
1738
1739#if 0 /* Still not clear if we can do without next two functions */
1740static void bfq_activate_request(struct request_queue *q, struct request *rq)
1741{
1742	struct bfq_data *bfqd = q->elevator->elevator_data;
1743
1744	bfqd->rq_in_driver++;
1745}
1746
1747static void bfq_deactivate_request(struct request_queue *q, struct request *rq)
1748{
1749	struct bfq_data *bfqd = q->elevator->elevator_data;
1750
1751	bfqd->rq_in_driver--;
1752}
1753#endif
1754
1755static void bfq_remove_request(struct request_queue *q,
1756			       struct request *rq)
1757{
1758	struct bfq_queue *bfqq = RQ_BFQQ(rq);
1759	struct bfq_data *bfqd = bfqq->bfqd;
1760	const int sync = rq_is_sync(rq);
1761
1762	if (bfqq->next_rq == rq) {
1763		bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
1764		bfq_updated_next_req(bfqd, bfqq);
1765	}
1766
1767	if (rq->queuelist.prev != &rq->queuelist)
1768		list_del_init(&rq->queuelist);
1769	bfqq->queued[sync]--;
1770	bfqd->queued--;
 
 
 
 
1771	elv_rb_del(&bfqq->sort_list, rq);
1772
1773	elv_rqhash_del(q, rq);
1774	if (q->last_merge == rq)
1775		q->last_merge = NULL;
1776
1777	if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
1778		bfqq->next_rq = NULL;
1779
1780		if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
1781			bfq_del_bfqq_busy(bfqd, bfqq, false);
1782			/*
1783			 * bfqq emptied. In normal operation, when
1784			 * bfqq is empty, bfqq->entity.service and
1785			 * bfqq->entity.budget must contain,
1786			 * respectively, the service received and the
1787			 * budget used last time bfqq emptied. These
1788			 * facts do not hold in this case, as at least
1789			 * this last removal occurred while bfqq is
1790			 * not in service. To avoid inconsistencies,
1791			 * reset both bfqq->entity.service and
1792			 * bfqq->entity.budget, if bfqq has still a
1793			 * process that may issue I/O requests to it.
1794			 */
1795			bfqq->entity.budget = bfqq->entity.service = 0;
1796		}
1797
1798		/*
1799		 * Remove queue from request-position tree as it is empty.
1800		 */
1801		if (bfqq->pos_root) {
1802			rb_erase(&bfqq->pos_node, bfqq->pos_root);
1803			bfqq->pos_root = NULL;
1804		}
1805	} else {
1806		bfq_pos_tree_add_move(bfqd, bfqq);
 
 
1807	}
1808
1809	if (rq->cmd_flags & REQ_META)
1810		bfqq->meta_pending--;
1811
1812}
1813
1814static bool bfq_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio)
 
1815{
1816	struct request_queue *q = hctx->queue;
1817	struct bfq_data *bfqd = q->elevator->elevator_data;
1818	struct request *free = NULL;
1819	/*
1820	 * bfq_bic_lookup grabs the queue_lock: invoke it now and
1821	 * store its return value for later use, to avoid nesting
1822	 * queue_lock inside the bfqd->lock. We assume that the bic
1823	 * returned by bfq_bic_lookup does not go away before
1824	 * bfqd->lock is taken.
1825	 */
1826	struct bfq_io_cq *bic = bfq_bic_lookup(bfqd, current->io_context, q);
1827	bool ret;
1828
1829	spin_lock_irq(&bfqd->lock);
1830
1831	if (bic)
 
 
 
 
 
 
1832		bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf));
1833	else
1834		bfqd->bio_bfqq = NULL;
 
1835	bfqd->bio_bic = bic;
1836
1837	ret = blk_mq_sched_try_merge(q, bio, &free);
1838
 
1839	if (free)
1840		blk_mq_free_request(free);
1841	spin_unlock_irq(&bfqd->lock);
1842
1843	return ret;
1844}
1845
1846static int bfq_request_merge(struct request_queue *q, struct request **req,
1847			     struct bio *bio)
1848{
1849	struct bfq_data *bfqd = q->elevator->elevator_data;
1850	struct request *__rq;
1851
1852	__rq = bfq_find_rq_fmerge(bfqd, bio, q);
1853	if (__rq && elv_bio_merge_ok(__rq, bio)) {
1854		*req = __rq;
 
 
 
1855		return ELEVATOR_FRONT_MERGE;
1856	}
1857
1858	return ELEVATOR_NO_MERGE;
1859}
1860
1861static void bfq_request_merged(struct request_queue *q, struct request *req,
1862			       enum elv_merge type)
1863{
1864	if (type == ELEVATOR_FRONT_MERGE &&
1865	    rb_prev(&req->rb_node) &&
1866	    blk_rq_pos(req) <
1867	    blk_rq_pos(container_of(rb_prev(&req->rb_node),
1868				    struct request, rb_node))) {
1869		struct bfq_queue *bfqq = RQ_BFQQ(req);
1870		struct bfq_data *bfqd = bfqq->bfqd;
1871		struct request *prev, *next_rq;
1872
 
 
 
 
 
1873		/* Reposition request in its sort_list */
1874		elv_rb_del(&bfqq->sort_list, req);
1875		elv_rb_add(&bfqq->sort_list, req);
1876
1877		/* Choose next request to be served for bfqq */
1878		prev = bfqq->next_rq;
1879		next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
1880					 bfqd->last_position);
1881		bfqq->next_rq = next_rq;
1882		/*
1883		 * If next_rq changes, update both the queue's budget to
1884		 * fit the new request and the queue's position in its
1885		 * rq_pos_tree.
1886		 */
1887		if (prev != bfqq->next_rq) {
1888			bfq_updated_next_req(bfqd, bfqq);
1889			bfq_pos_tree_add_move(bfqd, bfqq);
 
 
 
 
 
1890		}
1891	}
1892}
1893
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1894static void bfq_requests_merged(struct request_queue *q, struct request *rq,
1895				struct request *next)
1896{
1897	struct bfq_queue *bfqq = RQ_BFQQ(rq), *next_bfqq = RQ_BFQQ(next);
 
1898
1899	if (!RB_EMPTY_NODE(&rq->rb_node))
1900		goto end;
1901	spin_lock_irq(&bfqq->bfqd->lock);
1902
1903	/*
1904	 * If next and rq belong to the same bfq_queue and next is older
1905	 * than rq, then reposition rq in the fifo (by substituting next
1906	 * with rq). Otherwise, if next and rq belong to different
1907	 * bfq_queues, never reposition rq: in fact, we would have to
1908	 * reposition it with respect to next's position in its own fifo,
1909	 * which would most certainly be too expensive with respect to
1910	 * the benefits.
1911	 */
1912	if (bfqq == next_bfqq &&
1913	    !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1914	    next->fifo_time < rq->fifo_time) {
1915		list_del_init(&rq->queuelist);
1916		list_replace_init(&next->queuelist, &rq->queuelist);
1917		rq->fifo_time = next->fifo_time;
1918	}
1919
1920	if (bfqq->next_rq == next)
1921		bfqq->next_rq = rq;
1922
1923	bfq_remove_request(q, next);
1924	bfqg_stats_update_io_remove(bfqq_group(bfqq), next->cmd_flags);
1925
1926	spin_unlock_irq(&bfqq->bfqd->lock);
1927end:
1928	bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
 
 
 
 
 
 
 
 
1929}
1930
1931/* Must be called with bfqq != NULL */
1932static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
1933{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1934	if (bfq_bfqq_busy(bfqq))
1935		bfqq->bfqd->wr_busy_queues--;
1936	bfqq->wr_coeff = 1;
1937	bfqq->wr_cur_max_time = 0;
1938	bfqq->last_wr_start_finish = jiffies;
1939	/*
1940	 * Trigger a weight change on the next invocation of
1941	 * __bfq_entity_update_weight_prio.
1942	 */
1943	bfqq->entity.prio_changed = 1;
1944}
1945
1946void bfq_end_wr_async_queues(struct bfq_data *bfqd,
1947			     struct bfq_group *bfqg)
1948{
1949	int i, j;
1950
1951	for (i = 0; i < 2; i++)
1952		for (j = 0; j < IOPRIO_BE_NR; j++)
1953			if (bfqg->async_bfqq[i][j])
1954				bfq_bfqq_end_wr(bfqg->async_bfqq[i][j]);
1955	if (bfqg->async_idle_bfqq)
1956		bfq_bfqq_end_wr(bfqg->async_idle_bfqq);
1957}
1958
1959static void bfq_end_wr(struct bfq_data *bfqd)
1960{
1961	struct bfq_queue *bfqq;
1962
1963	spin_lock_irq(&bfqd->lock);
1964
1965	list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list)
1966		bfq_bfqq_end_wr(bfqq);
1967	list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
1968		bfq_bfqq_end_wr(bfqq);
1969	bfq_end_wr_async(bfqd);
1970
1971	spin_unlock_irq(&bfqd->lock);
1972}
1973
1974static sector_t bfq_io_struct_pos(void *io_struct, bool request)
1975{
1976	if (request)
1977		return blk_rq_pos(io_struct);
1978	else
1979		return ((struct bio *)io_struct)->bi_iter.bi_sector;
1980}
1981
1982static int bfq_rq_close_to_sector(void *io_struct, bool request,
1983				  sector_t sector)
1984{
1985	return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
1986	       BFQQ_CLOSE_THR;
1987}
1988
1989static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
1990					 struct bfq_queue *bfqq,
1991					 sector_t sector)
1992{
1993	struct rb_root *root = &bfq_bfqq_to_bfqg(bfqq)->rq_pos_tree;
1994	struct rb_node *parent, *node;
1995	struct bfq_queue *__bfqq;
1996
1997	if (RB_EMPTY_ROOT(root))
1998		return NULL;
1999
2000	/*
2001	 * First, if we find a request starting at the end of the last
2002	 * request, choose it.
2003	 */
2004	__bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2005	if (__bfqq)
2006		return __bfqq;
2007
2008	/*
2009	 * If the exact sector wasn't found, the parent of the NULL leaf
2010	 * will contain the closest sector (rq_pos_tree sorted by
2011	 * next_request position).
2012	 */
2013	__bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2014	if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2015		return __bfqq;
2016
2017	if (blk_rq_pos(__bfqq->next_rq) < sector)
2018		node = rb_next(&__bfqq->pos_node);
2019	else
2020		node = rb_prev(&__bfqq->pos_node);
2021	if (!node)
2022		return NULL;
2023
2024	__bfqq = rb_entry(node, struct bfq_queue, pos_node);
2025	if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2026		return __bfqq;
2027
2028	return NULL;
2029}
2030
2031static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2032						   struct bfq_queue *cur_bfqq,
2033						   sector_t sector)
2034{
2035	struct bfq_queue *bfqq;
2036
2037	/*
2038	 * We shall notice if some of the queues are cooperating,
2039	 * e.g., working closely on the same area of the device. In
2040	 * that case, we can group them together and: 1) don't waste
2041	 * time idling, and 2) serve the union of their requests in
2042	 * the best possible order for throughput.
2043	 */
2044	bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2045	if (!bfqq || bfqq == cur_bfqq)
2046		return NULL;
2047
2048	return bfqq;
2049}
2050
2051static struct bfq_queue *
2052bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2053{
2054	int process_refs, new_process_refs;
2055	struct bfq_queue *__bfqq;
2056
2057	/*
2058	 * If there are no process references on the new_bfqq, then it is
2059	 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2060	 * may have dropped their last reference (not just their last process
2061	 * reference).
2062	 */
2063	if (!bfqq_process_refs(new_bfqq))
2064		return NULL;
2065
2066	/* Avoid a circular list and skip interim queue merges. */
2067	while ((__bfqq = new_bfqq->new_bfqq)) {
2068		if (__bfqq == bfqq)
2069			return NULL;
2070		new_bfqq = __bfqq;
2071	}
2072
2073	process_refs = bfqq_process_refs(bfqq);
2074	new_process_refs = bfqq_process_refs(new_bfqq);
2075	/*
2076	 * If the process for the bfqq has gone away, there is no
2077	 * sense in merging the queues.
2078	 */
2079	if (process_refs == 0 || new_process_refs == 0)
2080		return NULL;
2081
 
 
 
 
 
 
 
 
2082	bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2083		new_bfqq->pid);
2084
2085	/*
2086	 * Merging is just a redirection: the requests of the process
2087	 * owning one of the two queues are redirected to the other queue.
2088	 * The latter queue, in its turn, is set as shared if this is the
2089	 * first time that the requests of some process are redirected to
2090	 * it.
2091	 *
2092	 * We redirect bfqq to new_bfqq and not the opposite, because
2093	 * we are in the context of the process owning bfqq, thus we
2094	 * have the io_cq of this process. So we can immediately
2095	 * configure this io_cq to redirect the requests of the
2096	 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2097	 * not available any more (new_bfqq->bic == NULL).
2098	 *
2099	 * Anyway, even in case new_bfqq coincides with the in-service
2100	 * queue, redirecting requests the in-service queue is the
2101	 * best option, as we feed the in-service queue with new
2102	 * requests close to the last request served and, by doing so,
2103	 * are likely to increase the throughput.
2104	 */
2105	bfqq->new_bfqq = new_bfqq;
 
 
 
 
 
 
 
 
 
2106	new_bfqq->ref += process_refs;
2107	return new_bfqq;
2108}
2109
2110static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2111					struct bfq_queue *new_bfqq)
2112{
2113	if (bfq_too_late_for_merging(new_bfqq))
2114		return false;
2115
2116	if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2117	    (bfqq->ioprio_class != new_bfqq->ioprio_class))
2118		return false;
2119
2120	/*
2121	 * If either of the queues has already been detected as seeky,
2122	 * then merging it with the other queue is unlikely to lead to
2123	 * sequential I/O.
2124	 */
2125	if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2126		return false;
2127
2128	/*
2129	 * Interleaved I/O is known to be done by (some) applications
2130	 * only for reads, so it does not make sense to merge async
2131	 * queues.
2132	 */
2133	if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2134		return false;
2135
2136	return true;
2137}
2138
 
 
 
2139/*
2140 * Attempt to schedule a merge of bfqq with the currently in-service
2141 * queue or with a close queue among the scheduled queues.  Return
2142 * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2143 * structure otherwise.
2144 *
2145 * The OOM queue is not allowed to participate to cooperation: in fact, since
2146 * the requests temporarily redirected to the OOM queue could be redirected
2147 * again to dedicated queues at any time, the state needed to correctly
2148 * handle merging with the OOM queue would be quite complex and expensive
2149 * to maintain. Besides, in such a critical condition as an out of memory,
2150 * the benefits of queue merging may be little relevant, or even negligible.
2151 *
2152 * WARNING: queue merging may impair fairness among non-weight raised
2153 * queues, for at least two reasons: 1) the original weight of a
2154 * merged queue may change during the merged state, 2) even being the
2155 * weight the same, a merged queue may be bloated with many more
2156 * requests than the ones produced by its originally-associated
2157 * process.
2158 */
2159static struct bfq_queue *
2160bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2161		     void *io_struct, bool request)
2162{
2163	struct bfq_queue *in_service_bfqq, *new_bfqq;
2164
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2165	/*
2166	 * Prevent bfqq from being merged if it has been created too
2167	 * long ago. The idea is that true cooperating processes, and
2168	 * thus their associated bfq_queues, are supposed to be
2169	 * created shortly after each other. This is the case, e.g.,
2170	 * for KVM/QEMU and dump I/O threads. Basing on this
2171	 * assumption, the following filtering greatly reduces the
2172	 * probability that two non-cooperating processes, which just
2173	 * happen to do close I/O for some short time interval, have
2174	 * their queues merged by mistake.
2175	 */
2176	if (bfq_too_late_for_merging(bfqq))
2177		return NULL;
2178
2179	if (bfqq->new_bfqq)
2180		return bfqq->new_bfqq;
2181
2182	if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
2183		return NULL;
2184
2185	/* If there is only one backlogged queue, don't search. */
2186	if (bfqd->busy_queues == 1)
2187		return NULL;
2188
2189	in_service_bfqq = bfqd->in_service_queue;
2190
2191	if (in_service_bfqq && in_service_bfqq != bfqq &&
2192	    likely(in_service_bfqq != &bfqd->oom_bfqq) &&
2193	    bfq_rq_close_to_sector(io_struct, request, bfqd->last_position) &&
 
2194	    bfqq->entity.parent == in_service_bfqq->entity.parent &&
2195	    bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
2196		new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
2197		if (new_bfqq)
2198			return new_bfqq;
2199	}
2200	/*
2201	 * Check whether there is a cooperator among currently scheduled
2202	 * queues. The only thing we need is that the bio/request is not
2203	 * NULL, as we need it to establish whether a cooperator exists.
2204	 */
2205	new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
2206			bfq_io_struct_pos(io_struct, request));
2207
2208	if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
2209	    bfq_may_be_close_cooperator(bfqq, new_bfqq))
2210		return bfq_setup_merge(bfqq, new_bfqq);
2211
2212	return NULL;
2213}
2214
2215static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
2216{
2217	struct bfq_io_cq *bic = bfqq->bic;
2218
2219	/*
2220	 * If !bfqq->bic, the queue is already shared or its requests
2221	 * have already been redirected to a shared queue; both idle window
2222	 * and weight raising state have already been saved. Do nothing.
2223	 */
2224	if (!bic)
2225		return;
2226
 
 
 
 
 
2227	bic->saved_ttime = bfqq->ttime;
2228	bic->saved_has_short_ttime = bfq_bfqq_has_short_ttime(bfqq);
2229	bic->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
 
 
2230	bic->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
2231	bic->was_in_burst_list = !hlist_unhashed(&bfqq->burst_list_node);
2232	if (unlikely(bfq_bfqq_just_created(bfqq) &&
2233		     !bfq_bfqq_in_large_burst(bfqq) &&
2234		     bfqq->bfqd->low_latency)) {
2235		/*
2236		 * bfqq being merged right after being created: bfqq
2237		 * would have deserved interactive weight raising, but
2238		 * did not make it to be set in a weight-raised state,
2239		 * because of this early merge.	Store directly the
2240		 * weight-raising state that would have been assigned
2241		 * to bfqq, so that to avoid that bfqq unjustly fails
2242		 * to enjoy weight raising if split soon.
2243		 */
2244		bic->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
 
2245		bic->saved_wr_cur_max_time = bfq_wr_duration(bfqq->bfqd);
2246		bic->saved_last_wr_start_finish = jiffies;
2247	} else {
2248		bic->saved_wr_coeff = bfqq->wr_coeff;
2249		bic->saved_wr_start_at_switch_to_srt =
2250			bfqq->wr_start_at_switch_to_srt;
 
2251		bic->saved_last_wr_start_finish = bfqq->last_wr_start_finish;
2252		bic->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
2253	}
2254}
2255
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2256static void
2257bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
2258		struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2259{
2260	bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
2261		(unsigned long)new_bfqq->pid);
2262	/* Save weight raising and idle window of the merged queues */
2263	bfq_bfqq_save_state(bfqq);
2264	bfq_bfqq_save_state(new_bfqq);
2265	if (bfq_bfqq_IO_bound(bfqq))
2266		bfq_mark_bfqq_IO_bound(new_bfqq);
2267	bfq_clear_bfqq_IO_bound(bfqq);
2268
2269	/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2270	 * If bfqq is weight-raised, then let new_bfqq inherit
2271	 * weight-raising. To reduce false positives, neglect the case
2272	 * where bfqq has just been created, but has not yet made it
2273	 * to be weight-raised (which may happen because EQM may merge
2274	 * bfqq even before bfq_add_request is executed for the first
2275	 * time for bfqq). Handling this case would however be very
2276	 * easy, thanks to the flag just_created.
2277	 */
2278	if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
2279		new_bfqq->wr_coeff = bfqq->wr_coeff;
2280		new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
2281		new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
2282		new_bfqq->wr_start_at_switch_to_srt =
2283			bfqq->wr_start_at_switch_to_srt;
2284		if (bfq_bfqq_busy(new_bfqq))
2285			bfqd->wr_busy_queues++;
2286		new_bfqq->entity.prio_changed = 1;
2287	}
2288
2289	if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
2290		bfqq->wr_coeff = 1;
2291		bfqq->entity.prio_changed = 1;
2292		if (bfq_bfqq_busy(bfqq))
2293			bfqd->wr_busy_queues--;
2294	}
2295
2296	bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
2297		     bfqd->wr_busy_queues);
2298
2299	/*
2300	 * Merge queues (that is, let bic redirect its requests to new_bfqq)
2301	 */
2302	bic_set_bfqq(bic, new_bfqq, 1);
2303	bfq_mark_bfqq_coop(new_bfqq);
2304	/*
2305	 * new_bfqq now belongs to at least two bics (it is a shared queue):
2306	 * set new_bfqq->bic to NULL. bfqq either:
2307	 * - does not belong to any bic any more, and hence bfqq->bic must
2308	 *   be set to NULL, or
2309	 * - is a queue whose owning bics have already been redirected to a
2310	 *   different queue, hence the queue is destined to not belong to
2311	 *   any bic soon and bfqq->bic is already NULL (therefore the next
2312	 *   assignment causes no harm).
2313	 */
2314	new_bfqq->bic = NULL;
 
 
 
 
 
 
 
 
 
 
2315	bfqq->bic = NULL;
2316	/* release process reference to bfqq */
2317	bfq_put_queue(bfqq);
 
 
2318}
2319
2320static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2321				struct bio *bio)
2322{
2323	struct bfq_data *bfqd = q->elevator->elevator_data;
2324	bool is_sync = op_is_sync(bio->bi_opf);
2325	struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
2326
2327	/*
2328	 * Disallow merge of a sync bio into an async request.
2329	 */
2330	if (is_sync && !rq_is_sync(rq))
2331		return false;
2332
2333	/*
2334	 * Lookup the bfqq that this bio will be queued with. Allow
2335	 * merge only if rq is queued there.
2336	 */
2337	if (!bfqq)
2338		return false;
2339
2340	/*
2341	 * We take advantage of this function to perform an early merge
2342	 * of the queues of possible cooperating processes.
2343	 */
2344	new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
2345	if (new_bfqq) {
2346		/*
2347		 * bic still points to bfqq, then it has not yet been
2348		 * redirected to some other bfq_queue, and a queue
2349		 * merge beween bfqq and new_bfqq can be safely
2350		 * fulfillled, i.e., bic can be redirected to new_bfqq
2351		 * and bfqq can be put.
2352		 */
2353		bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
2354				new_bfqq);
2355		/*
2356		 * If we get here, bio will be queued into new_queue,
2357		 * so use new_bfqq to decide whether bio and rq can be
2358		 * merged.
2359		 */
2360		bfqq = new_bfqq;
2361
2362		/*
2363		 * Change also bqfd->bio_bfqq, as
2364		 * bfqd->bio_bic now points to new_bfqq, and
2365		 * this function may be invoked again (and then may
2366		 * use again bqfd->bio_bfqq).
2367		 */
2368		bfqd->bio_bfqq = bfqq;
2369	}
2370
2371	return bfqq == RQ_BFQQ(rq);
2372}
2373
2374/*
2375 * Set the maximum time for the in-service queue to consume its
2376 * budget. This prevents seeky processes from lowering the throughput.
2377 * In practice, a time-slice service scheme is used with seeky
2378 * processes.
2379 */
2380static void bfq_set_budget_timeout(struct bfq_data *bfqd,
2381				   struct bfq_queue *bfqq)
2382{
2383	unsigned int timeout_coeff;
2384
2385	if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
2386		timeout_coeff = 1;
2387	else
2388		timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
2389
2390	bfqd->last_budget_start = ktime_get();
2391
2392	bfqq->budget_timeout = jiffies +
2393		bfqd->bfq_timeout * timeout_coeff;
2394}
2395
2396static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
2397				       struct bfq_queue *bfqq)
2398{
2399	if (bfqq) {
2400		bfq_clear_bfqq_fifo_expire(bfqq);
2401
2402		bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
2403
2404		if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
2405		    bfqq->wr_coeff > 1 &&
2406		    bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
2407		    time_is_before_jiffies(bfqq->budget_timeout)) {
2408			/*
2409			 * For soft real-time queues, move the start
2410			 * of the weight-raising period forward by the
2411			 * time the queue has not received any
2412			 * service. Otherwise, a relatively long
2413			 * service delay is likely to cause the
2414			 * weight-raising period of the queue to end,
2415			 * because of the short duration of the
2416			 * weight-raising period of a soft real-time
2417			 * queue.  It is worth noting that this move
2418			 * is not so dangerous for the other queues,
2419			 * because soft real-time queues are not
2420			 * greedy.
2421			 *
2422			 * To not add a further variable, we use the
2423			 * overloaded field budget_timeout to
2424			 * determine for how long the queue has not
2425			 * received service, i.e., how much time has
2426			 * elapsed since the queue expired. However,
2427			 * this is a little imprecise, because
2428			 * budget_timeout is set to jiffies if bfqq
2429			 * not only expires, but also remains with no
2430			 * request.
2431			 */
2432			if (time_after(bfqq->budget_timeout,
2433				       bfqq->last_wr_start_finish))
2434				bfqq->last_wr_start_finish +=
2435					jiffies - bfqq->budget_timeout;
2436			else
2437				bfqq->last_wr_start_finish = jiffies;
2438		}
2439
2440		bfq_set_budget_timeout(bfqd, bfqq);
2441		bfq_log_bfqq(bfqd, bfqq,
2442			     "set_in_service_queue, cur-budget = %d",
2443			     bfqq->entity.budget);
2444	}
2445
2446	bfqd->in_service_queue = bfqq;
 
2447}
2448
2449/*
2450 * Get and set a new queue for service.
2451 */
2452static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
2453{
2454	struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
2455
2456	__bfq_set_in_service_queue(bfqd, bfqq);
2457	return bfqq;
2458}
2459
2460static void bfq_arm_slice_timer(struct bfq_data *bfqd)
2461{
2462	struct bfq_queue *bfqq = bfqd->in_service_queue;
2463	u32 sl;
2464
2465	bfq_mark_bfqq_wait_request(bfqq);
2466
2467	/*
2468	 * We don't want to idle for seeks, but we do want to allow
2469	 * fair distribution of slice time for a process doing back-to-back
2470	 * seeks. So allow a little bit of time for him to submit a new rq.
2471	 */
2472	sl = bfqd->bfq_slice_idle;
2473	/*
2474	 * Unless the queue is being weight-raised or the scenario is
2475	 * asymmetric, grant only minimum idle time if the queue
2476	 * is seeky. A long idling is preserved for a weight-raised
2477	 * queue, or, more in general, in an asymmetric scenario,
2478	 * because a long idling is needed for guaranteeing to a queue
2479	 * its reserved share of the throughput (in particular, it is
2480	 * needed if the queue has a higher weight than some other
2481	 * queue).
2482	 */
2483	if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
2484	    bfq_symmetric_scenario(bfqd))
2485		sl = min_t(u64, sl, BFQ_MIN_TT);
 
 
2486
2487	bfqd->last_idling_start = ktime_get();
 
 
2488	hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
2489		      HRTIMER_MODE_REL);
2490	bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
2491}
2492
2493/*
2494 * In autotuning mode, max_budget is dynamically recomputed as the
2495 * amount of sectors transferred in timeout at the estimated peak
2496 * rate. This enables BFQ to utilize a full timeslice with a full
2497 * budget, even if the in-service queue is served at peak rate. And
2498 * this maximises throughput with sequential workloads.
2499 */
2500static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
2501{
2502	return (u64)bfqd->peak_rate * USEC_PER_MSEC *
2503		jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
2504}
2505
2506/*
2507 * Update parameters related to throughput and responsiveness, as a
2508 * function of the estimated peak rate. See comments on
2509 * bfq_calc_max_budget(), and on T_slow and T_fast arrays.
2510 */
2511static void update_thr_responsiveness_params(struct bfq_data *bfqd)
2512{
2513	int dev_type = blk_queue_nonrot(bfqd->queue);
2514
2515	if (bfqd->bfq_user_max_budget == 0)
2516		bfqd->bfq_max_budget =
2517			bfq_calc_max_budget(bfqd);
2518
2519	if (bfqd->device_speed == BFQ_BFQD_FAST &&
2520	    bfqd->peak_rate < device_speed_thresh[dev_type]) {
2521		bfqd->device_speed = BFQ_BFQD_SLOW;
2522		bfqd->RT_prod = R_slow[dev_type] *
2523			T_slow[dev_type];
2524	} else if (bfqd->device_speed == BFQ_BFQD_SLOW &&
2525		   bfqd->peak_rate > device_speed_thresh[dev_type]) {
2526		bfqd->device_speed = BFQ_BFQD_FAST;
2527		bfqd->RT_prod = R_fast[dev_type] *
2528			T_fast[dev_type];
2529	}
2530
2531	bfq_log(bfqd,
2532"dev_type %s dev_speed_class = %s (%llu sects/sec), thresh %llu setcs/sec",
2533		dev_type == 0 ? "ROT" : "NONROT",
2534		bfqd->device_speed == BFQ_BFQD_FAST ? "FAST" : "SLOW",
2535		bfqd->device_speed == BFQ_BFQD_FAST ?
2536		(USEC_PER_SEC*(u64)R_fast[dev_type])>>BFQ_RATE_SHIFT :
2537		(USEC_PER_SEC*(u64)R_slow[dev_type])>>BFQ_RATE_SHIFT,
2538		(USEC_PER_SEC*(u64)device_speed_thresh[dev_type])>>
2539		BFQ_RATE_SHIFT);
2540}
2541
2542static void bfq_reset_rate_computation(struct bfq_data *bfqd,
2543				       struct request *rq)
2544{
2545	if (rq != NULL) { /* new rq dispatch now, reset accordingly */
2546		bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
2547		bfqd->peak_rate_samples = 1;
2548		bfqd->sequential_samples = 0;
2549		bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
2550			blk_rq_sectors(rq);
2551	} else /* no new rq dispatched, just reset the number of samples */
2552		bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
2553
2554	bfq_log(bfqd,
2555		"reset_rate_computation at end, sample %u/%u tot_sects %llu",
2556		bfqd->peak_rate_samples, bfqd->sequential_samples,
2557		bfqd->tot_sectors_dispatched);
2558}
2559
2560static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
2561{
2562	u32 rate, weight, divisor;
2563
2564	/*
2565	 * For the convergence property to hold (see comments on
2566	 * bfq_update_peak_rate()) and for the assessment to be
2567	 * reliable, a minimum number of samples must be present, and
2568	 * a minimum amount of time must have elapsed. If not so, do
2569	 * not compute new rate. Just reset parameters, to get ready
2570	 * for a new evaluation attempt.
2571	 */
2572	if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
2573	    bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
2574		goto reset_computation;
2575
2576	/*
2577	 * If a new request completion has occurred after last
2578	 * dispatch, then, to approximate the rate at which requests
2579	 * have been served by the device, it is more precise to
2580	 * extend the observation interval to the last completion.
2581	 */
2582	bfqd->delta_from_first =
2583		max_t(u64, bfqd->delta_from_first,
2584		      bfqd->last_completion - bfqd->first_dispatch);
2585
2586	/*
2587	 * Rate computed in sects/usec, and not sects/nsec, for
2588	 * precision issues.
2589	 */
2590	rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
2591			div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
2592
2593	/*
2594	 * Peak rate not updated if:
2595	 * - the percentage of sequential dispatches is below 3/4 of the
2596	 *   total, and rate is below the current estimated peak rate
2597	 * - rate is unreasonably high (> 20M sectors/sec)
2598	 */
2599	if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
2600	     rate <= bfqd->peak_rate) ||
2601		rate > 20<<BFQ_RATE_SHIFT)
2602		goto reset_computation;
2603
2604	/*
2605	 * We have to update the peak rate, at last! To this purpose,
2606	 * we use a low-pass filter. We compute the smoothing constant
2607	 * of the filter as a function of the 'weight' of the new
2608	 * measured rate.
2609	 *
2610	 * As can be seen in next formulas, we define this weight as a
2611	 * quantity proportional to how sequential the workload is,
2612	 * and to how long the observation time interval is.
2613	 *
2614	 * The weight runs from 0 to 8. The maximum value of the
2615	 * weight, 8, yields the minimum value for the smoothing
2616	 * constant. At this minimum value for the smoothing constant,
2617	 * the measured rate contributes for half of the next value of
2618	 * the estimated peak rate.
2619	 *
2620	 * So, the first step is to compute the weight as a function
2621	 * of how sequential the workload is. Note that the weight
2622	 * cannot reach 9, because bfqd->sequential_samples cannot
2623	 * become equal to bfqd->peak_rate_samples, which, in its
2624	 * turn, holds true because bfqd->sequential_samples is not
2625	 * incremented for the first sample.
2626	 */
2627	weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
2628
2629	/*
2630	 * Second step: further refine the weight as a function of the
2631	 * duration of the observation interval.
2632	 */
2633	weight = min_t(u32, 8,
2634		       div_u64(weight * bfqd->delta_from_first,
2635			       BFQ_RATE_REF_INTERVAL));
2636
2637	/*
2638	 * Divisor ranging from 10, for minimum weight, to 2, for
2639	 * maximum weight.
2640	 */
2641	divisor = 10 - weight;
2642
2643	/*
2644	 * Finally, update peak rate:
2645	 *
2646	 * peak_rate = peak_rate * (divisor-1) / divisor  +  rate / divisor
2647	 */
2648	bfqd->peak_rate *= divisor-1;
2649	bfqd->peak_rate /= divisor;
2650	rate /= divisor; /* smoothing constant alpha = 1/divisor */
2651
2652	bfqd->peak_rate += rate;
2653
2654	/*
2655	 * For a very slow device, bfqd->peak_rate can reach 0 (see
2656	 * the minimum representable values reported in the comments
2657	 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
2658	 * divisions by zero where bfqd->peak_rate is used as a
2659	 * divisor.
2660	 */
2661	bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
2662
2663	update_thr_responsiveness_params(bfqd);
2664
2665reset_computation:
2666	bfq_reset_rate_computation(bfqd, rq);
2667}
2668
2669/*
2670 * Update the read/write peak rate (the main quantity used for
2671 * auto-tuning, see update_thr_responsiveness_params()).
2672 *
2673 * It is not trivial to estimate the peak rate (correctly): because of
2674 * the presence of sw and hw queues between the scheduler and the
2675 * device components that finally serve I/O requests, it is hard to
2676 * say exactly when a given dispatched request is served inside the
2677 * device, and for how long. As a consequence, it is hard to know
2678 * precisely at what rate a given set of requests is actually served
2679 * by the device.
2680 *
2681 * On the opposite end, the dispatch time of any request is trivially
2682 * available, and, from this piece of information, the "dispatch rate"
2683 * of requests can be immediately computed. So, the idea in the next
2684 * function is to use what is known, namely request dispatch times
2685 * (plus, when useful, request completion times), to estimate what is
2686 * unknown, namely in-device request service rate.
2687 *
2688 * The main issue is that, because of the above facts, the rate at
2689 * which a certain set of requests is dispatched over a certain time
2690 * interval can vary greatly with respect to the rate at which the
2691 * same requests are then served. But, since the size of any
2692 * intermediate queue is limited, and the service scheme is lossless
2693 * (no request is silently dropped), the following obvious convergence
2694 * property holds: the number of requests dispatched MUST become
2695 * closer and closer to the number of requests completed as the
2696 * observation interval grows. This is the key property used in
2697 * the next function to estimate the peak service rate as a function
2698 * of the observed dispatch rate. The function assumes to be invoked
2699 * on every request dispatch.
2700 */
2701static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
2702{
2703	u64 now_ns = ktime_get_ns();
2704
2705	if (bfqd->peak_rate_samples == 0) { /* first dispatch */
2706		bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
2707			bfqd->peak_rate_samples);
2708		bfq_reset_rate_computation(bfqd, rq);
2709		goto update_last_values; /* will add one sample */
2710	}
2711
2712	/*
2713	 * Device idle for very long: the observation interval lasting
2714	 * up to this dispatch cannot be a valid observation interval
2715	 * for computing a new peak rate (similarly to the late-
2716	 * completion event in bfq_completed_request()). Go to
2717	 * update_rate_and_reset to have the following three steps
2718	 * taken:
2719	 * - close the observation interval at the last (previous)
2720	 *   request dispatch or completion
2721	 * - compute rate, if possible, for that observation interval
2722	 * - start a new observation interval with this dispatch
2723	 */
2724	if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
2725	    bfqd->rq_in_driver == 0)
2726		goto update_rate_and_reset;
2727
2728	/* Update sampling information */
2729	bfqd->peak_rate_samples++;
2730
2731	if ((bfqd->rq_in_driver > 0 ||
2732		now_ns - bfqd->last_completion < BFQ_MIN_TT)
2733	     && get_sdist(bfqd->last_position, rq) < BFQQ_SEEK_THR)
2734		bfqd->sequential_samples++;
2735
2736	bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
2737
2738	/* Reset max observed rq size every 32 dispatches */
2739	if (likely(bfqd->peak_rate_samples % 32))
2740		bfqd->last_rq_max_size =
2741			max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
2742	else
2743		bfqd->last_rq_max_size = blk_rq_sectors(rq);
2744
2745	bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
2746
2747	/* Target observation interval not yet reached, go on sampling */
2748	if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
2749		goto update_last_values;
2750
2751update_rate_and_reset:
2752	bfq_update_rate_reset(bfqd, rq);
2753update_last_values:
2754	bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
 
 
2755	bfqd->last_dispatch = now_ns;
2756}
2757
2758/*
2759 * Remove request from internal lists.
2760 */
2761static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
2762{
2763	struct bfq_queue *bfqq = RQ_BFQQ(rq);
2764
2765	/*
2766	 * For consistency, the next instruction should have been
2767	 * executed after removing the request from the queue and
2768	 * dispatching it.  We execute instead this instruction before
2769	 * bfq_remove_request() (and hence introduce a temporary
2770	 * inconsistency), for efficiency.  In fact, should this
2771	 * dispatch occur for a non in-service bfqq, this anticipated
2772	 * increment prevents two counters related to bfqq->dispatched
2773	 * from risking to be, first, uselessly decremented, and then
2774	 * incremented again when the (new) value of bfqq->dispatched
2775	 * happens to be taken into account.
2776	 */
2777	bfqq->dispatched++;
2778	bfq_update_peak_rate(q->elevator->elevator_data, rq);
2779
2780	bfq_remove_request(q, rq);
2781}
2782
2783static void __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2784{
2785	/*
2786	 * If this bfqq is shared between multiple processes, check
2787	 * to make sure that those processes are still issuing I/Os
2788	 * within the mean seek distance. If not, it may be time to
2789	 * break the queues apart again.
2790	 */
2791	if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
2792		bfq_mark_bfqq_split_coop(bfqq);
2793
2794	if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2795		if (bfqq->dispatched == 0)
2796			/*
2797			 * Overloading budget_timeout field to store
2798			 * the time at which the queue remains with no
2799			 * backlog and no outstanding request; used by
2800			 * the weight-raising mechanism.
2801			 */
2802			bfqq->budget_timeout = jiffies;
2803
2804		bfq_del_bfqq_busy(bfqd, bfqq, true);
2805	} else {
2806		bfq_requeue_bfqq(bfqd, bfqq, true);
2807		/*
2808		 * Resort priority tree of potential close cooperators.
 
2809		 */
2810		bfq_pos_tree_add_move(bfqd, bfqq);
 
 
2811	}
2812
2813	/*
2814	 * All in-service entities must have been properly deactivated
2815	 * or requeued before executing the next function, which
2816	 * resets all in-service entites as no more in service.
 
 
2817	 */
2818	__bfq_bfqd_reset_in_service(bfqd);
2819}
2820
2821/**
2822 * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
2823 * @bfqd: device data.
2824 * @bfqq: queue to update.
2825 * @reason: reason for expiration.
2826 *
2827 * Handle the feedback on @bfqq budget at queue expiration.
2828 * See the body for detailed comments.
2829 */
2830static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
2831				     struct bfq_queue *bfqq,
2832				     enum bfqq_expiration reason)
2833{
2834	struct request *next_rq;
2835	int budget, min_budget;
2836
2837	min_budget = bfq_min_budget(bfqd);
2838
2839	if (bfqq->wr_coeff == 1)
2840		budget = bfqq->max_budget;
2841	else /*
2842	      * Use a constant, low budget for weight-raised queues,
2843	      * to help achieve a low latency. Keep it slightly higher
2844	      * than the minimum possible budget, to cause a little
2845	      * bit fewer expirations.
2846	      */
2847		budget = 2 * min_budget;
2848
2849	bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
2850		bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
2851	bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
2852		budget, bfq_min_budget(bfqd));
2853	bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
2854		bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
2855
2856	if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
2857		switch (reason) {
2858		/*
2859		 * Caveat: in all the following cases we trade latency
2860		 * for throughput.
2861		 */
2862		case BFQQE_TOO_IDLE:
2863			/*
2864			 * This is the only case where we may reduce
2865			 * the budget: if there is no request of the
2866			 * process still waiting for completion, then
2867			 * we assume (tentatively) that the timer has
2868			 * expired because the batch of requests of
2869			 * the process could have been served with a
2870			 * smaller budget.  Hence, betting that
2871			 * process will behave in the same way when it
2872			 * becomes backlogged again, we reduce its
2873			 * next budget.  As long as we guess right,
2874			 * this budget cut reduces the latency
2875			 * experienced by the process.
2876			 *
2877			 * However, if there are still outstanding
2878			 * requests, then the process may have not yet
2879			 * issued its next request just because it is
2880			 * still waiting for the completion of some of
2881			 * the still outstanding ones.  So in this
2882			 * subcase we do not reduce its budget, on the
2883			 * contrary we increase it to possibly boost
2884			 * the throughput, as discussed in the
2885			 * comments to the BUDGET_TIMEOUT case.
2886			 */
2887			if (bfqq->dispatched > 0) /* still outstanding reqs */
2888				budget = min(budget * 2, bfqd->bfq_max_budget);
2889			else {
2890				if (budget > 5 * min_budget)
2891					budget -= 4 * min_budget;
2892				else
2893					budget = min_budget;
2894			}
2895			break;
2896		case BFQQE_BUDGET_TIMEOUT:
2897			/*
2898			 * We double the budget here because it gives
2899			 * the chance to boost the throughput if this
2900			 * is not a seeky process (and has bumped into
2901			 * this timeout because of, e.g., ZBR).
2902			 */
2903			budget = min(budget * 2, bfqd->bfq_max_budget);
2904			break;
2905		case BFQQE_BUDGET_EXHAUSTED:
2906			/*
2907			 * The process still has backlog, and did not
2908			 * let either the budget timeout or the disk
2909			 * idling timeout expire. Hence it is not
2910			 * seeky, has a short thinktime and may be
2911			 * happy with a higher budget too. So
2912			 * definitely increase the budget of this good
2913			 * candidate to boost the disk throughput.
2914			 */
2915			budget = min(budget * 4, bfqd->bfq_max_budget);
2916			break;
2917		case BFQQE_NO_MORE_REQUESTS:
2918			/*
2919			 * For queues that expire for this reason, it
2920			 * is particularly important to keep the
2921			 * budget close to the actual service they
2922			 * need. Doing so reduces the timestamp
2923			 * misalignment problem described in the
2924			 * comments in the body of
2925			 * __bfq_activate_entity. In fact, suppose
2926			 * that a queue systematically expires for
2927			 * BFQQE_NO_MORE_REQUESTS and presents a
2928			 * new request in time to enjoy timestamp
2929			 * back-shifting. The larger the budget of the
2930			 * queue is with respect to the service the
2931			 * queue actually requests in each service
2932			 * slot, the more times the queue can be
2933			 * reactivated with the same virtual finish
2934			 * time. It follows that, even if this finish
2935			 * time is pushed to the system virtual time
2936			 * to reduce the consequent timestamp
2937			 * misalignment, the queue unjustly enjoys for
2938			 * many re-activations a lower finish time
2939			 * than all newly activated queues.
2940			 *
2941			 * The service needed by bfqq is measured
2942			 * quite precisely by bfqq->entity.service.
2943			 * Since bfqq does not enjoy device idling,
2944			 * bfqq->entity.service is equal to the number
2945			 * of sectors that the process associated with
2946			 * bfqq requested to read/write before waiting
2947			 * for request completions, or blocking for
2948			 * other reasons.
2949			 */
2950			budget = max_t(int, bfqq->entity.service, min_budget);
2951			break;
2952		default:
2953			return;
2954		}
2955	} else if (!bfq_bfqq_sync(bfqq)) {
2956		/*
2957		 * Async queues get always the maximum possible
2958		 * budget, as for them we do not care about latency
2959		 * (in addition, their ability to dispatch is limited
2960		 * by the charging factor).
2961		 */
2962		budget = bfqd->bfq_max_budget;
2963	}
2964
2965	bfqq->max_budget = budget;
2966
2967	if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
2968	    !bfqd->bfq_user_max_budget)
2969		bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
2970
2971	/*
2972	 * If there is still backlog, then assign a new budget, making
2973	 * sure that it is large enough for the next request.  Since
2974	 * the finish time of bfqq must be kept in sync with the
2975	 * budget, be sure to call __bfq_bfqq_expire() *after* this
2976	 * update.
2977	 *
2978	 * If there is no backlog, then no need to update the budget;
2979	 * it will be updated on the arrival of a new request.
2980	 */
2981	next_rq = bfqq->next_rq;
2982	if (next_rq)
2983		bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
2984					    bfq_serv_to_charge(next_rq, bfqq));
2985
2986	bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
2987			next_rq ? blk_rq_sectors(next_rq) : 0,
2988			bfqq->entity.budget);
2989}
2990
2991/*
2992 * Return true if the process associated with bfqq is "slow". The slow
2993 * flag is used, in addition to the budget timeout, to reduce the
2994 * amount of service provided to seeky processes, and thus reduce
2995 * their chances to lower the throughput. More details in the comments
2996 * on the function bfq_bfqq_expire().
2997 *
2998 * An important observation is in order: as discussed in the comments
2999 * on the function bfq_update_peak_rate(), with devices with internal
3000 * queues, it is hard if ever possible to know when and for how long
3001 * an I/O request is processed by the device (apart from the trivial
3002 * I/O pattern where a new request is dispatched only after the
3003 * previous one has been completed). This makes it hard to evaluate
3004 * the real rate at which the I/O requests of each bfq_queue are
3005 * served.  In fact, for an I/O scheduler like BFQ, serving a
3006 * bfq_queue means just dispatching its requests during its service
3007 * slot (i.e., until the budget of the queue is exhausted, or the
3008 * queue remains idle, or, finally, a timeout fires). But, during the
3009 * service slot of a bfq_queue, around 100 ms at most, the device may
3010 * be even still processing requests of bfq_queues served in previous
3011 * service slots. On the opposite end, the requests of the in-service
3012 * bfq_queue may be completed after the service slot of the queue
3013 * finishes.
3014 *
3015 * Anyway, unless more sophisticated solutions are used
3016 * (where possible), the sum of the sizes of the requests dispatched
3017 * during the service slot of a bfq_queue is probably the only
3018 * approximation available for the service received by the bfq_queue
3019 * during its service slot. And this sum is the quantity used in this
3020 * function to evaluate the I/O speed of a process.
3021 */
3022static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3023				 bool compensate, enum bfqq_expiration reason,
3024				 unsigned long *delta_ms)
3025{
3026	ktime_t delta_ktime;
3027	u32 delta_usecs;
3028	bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
3029
3030	if (!bfq_bfqq_sync(bfqq))
3031		return false;
3032
3033	if (compensate)
3034		delta_ktime = bfqd->last_idling_start;
3035	else
3036		delta_ktime = ktime_get();
3037	delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
3038	delta_usecs = ktime_to_us(delta_ktime);
3039
3040	/* don't use too short time intervals */
3041	if (delta_usecs < 1000) {
3042		if (blk_queue_nonrot(bfqd->queue))
3043			 /*
3044			  * give same worst-case guarantees as idling
3045			  * for seeky
3046			  */
3047			*delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
3048		else /* charge at least one seek */
3049			*delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
3050
3051		return slow;
3052	}
3053
3054	*delta_ms = delta_usecs / USEC_PER_MSEC;
3055
3056	/*
3057	 * Use only long (> 20ms) intervals to filter out excessive
3058	 * spikes in service rate estimation.
3059	 */
3060	if (delta_usecs > 20000) {
3061		/*
3062		 * Caveat for rotational devices: processes doing I/O
3063		 * in the slower disk zones tend to be slow(er) even
3064		 * if not seeky. In this respect, the estimated peak
3065		 * rate is likely to be an average over the disk
3066		 * surface. Accordingly, to not be too harsh with
3067		 * unlucky processes, a process is deemed slow only if
3068		 * its rate has been lower than half of the estimated
3069		 * peak rate.
3070		 */
3071		slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
3072	}
3073
3074	bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
3075
3076	return slow;
3077}
3078
3079/*
3080 * To be deemed as soft real-time, an application must meet two
3081 * requirements. First, the application must not require an average
3082 * bandwidth higher than the approximate bandwidth required to playback or
3083 * record a compressed high-definition video.
3084 * The next function is invoked on the completion of the last request of a
3085 * batch, to compute the next-start time instant, soft_rt_next_start, such
3086 * that, if the next request of the application does not arrive before
3087 * soft_rt_next_start, then the above requirement on the bandwidth is met.
3088 *
3089 * The second requirement is that the request pattern of the application is
3090 * isochronous, i.e., that, after issuing a request or a batch of requests,
3091 * the application stops issuing new requests until all its pending requests
3092 * have been completed. After that, the application may issue a new batch,
3093 * and so on.
3094 * For this reason the next function is invoked to compute
3095 * soft_rt_next_start only for applications that meet this requirement,
3096 * whereas soft_rt_next_start is set to infinity for applications that do
3097 * not.
3098 *
3099 * Unfortunately, even a greedy (i.e., I/O-bound) application may
3100 * happen to meet, occasionally or systematically, both the above
3101 * bandwidth and isochrony requirements. This may happen at least in
3102 * the following circumstances. First, if the CPU load is high. The
3103 * application may stop issuing requests while the CPUs are busy
3104 * serving other processes, then restart, then stop again for a while,
3105 * and so on. The other circumstances are related to the storage
3106 * device: the storage device is highly loaded or reaches a low-enough
3107 * throughput with the I/O of the application (e.g., because the I/O
3108 * is random and/or the device is slow). In all these cases, the
3109 * I/O of the application may be simply slowed down enough to meet
3110 * the bandwidth and isochrony requirements. To reduce the probability
3111 * that greedy applications are deemed as soft real-time in these
3112 * corner cases, a further rule is used in the computation of
3113 * soft_rt_next_start: the return value of this function is forced to
3114 * be higher than the maximum between the following two quantities.
3115 *
3116 * (a) Current time plus: (1) the maximum time for which the arrival
3117 *     of a request is waited for when a sync queue becomes idle,
3118 *     namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
3119 *     postpone for a moment the reason for adding a few extra
3120 *     jiffies; we get back to it after next item (b).  Lower-bounding
3121 *     the return value of this function with the current time plus
3122 *     bfqd->bfq_slice_idle tends to filter out greedy applications,
3123 *     because the latter issue their next request as soon as possible
3124 *     after the last one has been completed. In contrast, a soft
3125 *     real-time application spends some time processing data, after a
3126 *     batch of its requests has been completed.
3127 *
3128 * (b) Current value of bfqq->soft_rt_next_start. As pointed out
3129 *     above, greedy applications may happen to meet both the
3130 *     bandwidth and isochrony requirements under heavy CPU or
3131 *     storage-device load. In more detail, in these scenarios, these
3132 *     applications happen, only for limited time periods, to do I/O
3133 *     slowly enough to meet all the requirements described so far,
3134 *     including the filtering in above item (a). These slow-speed
3135 *     time intervals are usually interspersed between other time
3136 *     intervals during which these applications do I/O at a very high
3137 *     speed. Fortunately, exactly because of the high speed of the
3138 *     I/O in the high-speed intervals, the values returned by this
3139 *     function happen to be so high, near the end of any such
3140 *     high-speed interval, to be likely to fall *after* the end of
3141 *     the low-speed time interval that follows. These high values are
3142 *     stored in bfqq->soft_rt_next_start after each invocation of
3143 *     this function. As a consequence, if the last value of
3144 *     bfqq->soft_rt_next_start is constantly used to lower-bound the
3145 *     next value that this function may return, then, from the very
3146 *     beginning of a low-speed interval, bfqq->soft_rt_next_start is
3147 *     likely to be constantly kept so high that any I/O request
3148 *     issued during the low-speed interval is considered as arriving
3149 *     to soon for the application to be deemed as soft
3150 *     real-time. Then, in the high-speed interval that follows, the
3151 *     application will not be deemed as soft real-time, just because
3152 *     it will do I/O at a high speed. And so on.
3153 *
3154 * Getting back to the filtering in item (a), in the following two
3155 * cases this filtering might be easily passed by a greedy
3156 * application, if the reference quantity was just
3157 * bfqd->bfq_slice_idle:
3158 * 1) HZ is so low that the duration of a jiffy is comparable to or
3159 *    higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
3160 *    devices with HZ=100. The time granularity may be so coarse
3161 *    that the approximation, in jiffies, of bfqd->bfq_slice_idle
3162 *    is rather lower than the exact value.
3163 * 2) jiffies, instead of increasing at a constant rate, may stop increasing
3164 *    for a while, then suddenly 'jump' by several units to recover the lost
3165 *    increments. This seems to happen, e.g., inside virtual machines.
3166 * To address this issue, in the filtering in (a) we do not use as a
3167 * reference time interval just bfqd->bfq_slice_idle, but
3168 * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
3169 * minimum number of jiffies for which the filter seems to be quite
3170 * precise also in embedded systems and KVM/QEMU virtual machines.
3171 */
3172static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
3173						struct bfq_queue *bfqq)
3174{
3175	return max3(bfqq->soft_rt_next_start,
3176		    bfqq->last_idle_bklogged +
3177		    HZ * bfqq->service_from_backlogged /
3178		    bfqd->bfq_wr_max_softrt_rate,
3179		    jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
3180}
3181
3182/**
3183 * bfq_bfqq_expire - expire a queue.
3184 * @bfqd: device owning the queue.
3185 * @bfqq: the queue to expire.
3186 * @compensate: if true, compensate for the time spent idling.
3187 * @reason: the reason causing the expiration.
3188 *
3189 * If the process associated with bfqq does slow I/O (e.g., because it
3190 * issues random requests), we charge bfqq with the time it has been
3191 * in service instead of the service it has received (see
3192 * bfq_bfqq_charge_time for details on how this goal is achieved). As
3193 * a consequence, bfqq will typically get higher timestamps upon
3194 * reactivation, and hence it will be rescheduled as if it had
3195 * received more service than what it has actually received. In the
3196 * end, bfqq receives less service in proportion to how slowly its
3197 * associated process consumes its budgets (and hence how seriously it
3198 * tends to lower the throughput). In addition, this time-charging
3199 * strategy guarantees time fairness among slow processes. In
3200 * contrast, if the process associated with bfqq is not slow, we
3201 * charge bfqq exactly with the service it has received.
3202 *
3203 * Charging time to the first type of queues and the exact service to
3204 * the other has the effect of using the WF2Q+ policy to schedule the
3205 * former on a timeslice basis, without violating service domain
3206 * guarantees among the latter.
3207 */
3208void bfq_bfqq_expire(struct bfq_data *bfqd,
3209		     struct bfq_queue *bfqq,
3210		     bool compensate,
3211		     enum bfqq_expiration reason)
3212{
3213	bool slow;
3214	unsigned long delta = 0;
3215	struct bfq_entity *entity = &bfqq->entity;
3216	int ref;
3217
3218	/*
3219	 * Check whether the process is slow (see bfq_bfqq_is_slow).
3220	 */
3221	slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, reason, &delta);
3222
3223	/*
3224	 * As above explained, charge slow (typically seeky) and
3225	 * timed-out queues with the time and not the service
3226	 * received, to favor sequential workloads.
3227	 *
3228	 * Processes doing I/O in the slower disk zones will tend to
3229	 * be slow(er) even if not seeky. Therefore, since the
3230	 * estimated peak rate is actually an average over the disk
3231	 * surface, these processes may timeout just for bad luck. To
3232	 * avoid punishing them, do not charge time to processes that
3233	 * succeeded in consuming at least 2/3 of their budget. This
3234	 * allows BFQ to preserve enough elasticity to still perform
3235	 * bandwidth, and not time, distribution with little unlucky
3236	 * or quasi-sequential processes.
3237	 */
3238	if (bfqq->wr_coeff == 1 &&
3239	    (slow ||
3240	     (reason == BFQQE_BUDGET_TIMEOUT &&
3241	      bfq_bfqq_budget_left(bfqq) >=  entity->budget / 3)))
3242		bfq_bfqq_charge_time(bfqd, bfqq, delta);
3243
3244	if (reason == BFQQE_TOO_IDLE &&
3245	    entity->service <= 2 * entity->budget / 10)
3246		bfq_clear_bfqq_IO_bound(bfqq);
3247
3248	if (bfqd->low_latency && bfqq->wr_coeff == 1)
3249		bfqq->last_wr_start_finish = jiffies;
3250
3251	if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
3252	    RB_EMPTY_ROOT(&bfqq->sort_list)) {
3253		/*
3254		 * If we get here, and there are no outstanding
3255		 * requests, then the request pattern is isochronous
3256		 * (see the comments on the function
3257		 * bfq_bfqq_softrt_next_start()). Thus we can compute
3258		 * soft_rt_next_start. If, instead, the queue still
3259		 * has outstanding requests, then we have to wait for
3260		 * the completion of all the outstanding requests to
3261		 * discover whether the request pattern is actually
3262		 * isochronous.
 
3263		 */
3264		if (bfqq->dispatched == 0)
3265			bfqq->soft_rt_next_start =
3266				bfq_bfqq_softrt_next_start(bfqd, bfqq);
3267		else {
3268			/*
3269			 * The application is still waiting for the
3270			 * completion of one or more requests:
3271			 * prevent it from possibly being incorrectly
3272			 * deemed as soft real-time by setting its
3273			 * soft_rt_next_start to infinity. In fact,
3274			 * without this assignment, the application
3275			 * would be incorrectly deemed as soft
3276			 * real-time if:
3277			 * 1) it issued a new request before the
3278			 *    completion of all its in-flight
3279			 *    requests, and
3280			 * 2) at that time, its soft_rt_next_start
3281			 *    happened to be in the past.
3282			 */
3283			bfqq->soft_rt_next_start =
3284				bfq_greatest_from_now();
3285			/*
3286			 * Schedule an update of soft_rt_next_start to when
3287			 * the task may be discovered to be isochronous.
3288			 */
3289			bfq_mark_bfqq_softrt_update(bfqq);
3290		}
3291	}
3292
3293	bfq_log_bfqq(bfqd, bfqq,
3294		"expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
3295		slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
3296
3297	/*
 
 
 
 
 
 
 
 
3298	 * Increase, decrease or leave budget unchanged according to
3299	 * reason.
3300	 */
3301	__bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
3302	ref = bfqq->ref;
3303	__bfq_bfqq_expire(bfqd, bfqq);
 
3304
3305	/* mark bfqq as waiting a request only if a bic still points to it */
3306	if (ref > 1 && !bfq_bfqq_busy(bfqq) &&
3307	    reason != BFQQE_BUDGET_TIMEOUT &&
3308	    reason != BFQQE_BUDGET_EXHAUSTED)
3309		bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3310}
3311
3312/*
3313 * Budget timeout is not implemented through a dedicated timer, but
3314 * just checked on request arrivals and completions, as well as on
3315 * idle timer expirations.
3316 */
3317static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
3318{
3319	return time_is_before_eq_jiffies(bfqq->budget_timeout);
3320}
3321
3322/*
3323 * If we expire a queue that is actively waiting (i.e., with the
3324 * device idled) for the arrival of a new request, then we may incur
3325 * the timestamp misalignment problem described in the body of the
3326 * function __bfq_activate_entity. Hence we return true only if this
3327 * condition does not hold, or if the queue is slow enough to deserve
3328 * only to be kicked off for preserving a high throughput.
3329 */
3330static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
3331{
3332	bfq_log_bfqq(bfqq->bfqd, bfqq,
3333		"may_budget_timeout: wait_request %d left %d timeout %d",
3334		bfq_bfqq_wait_request(bfqq),
3335			bfq_bfqq_budget_left(bfqq) >=  bfqq->entity.budget / 3,
3336		bfq_bfqq_budget_timeout(bfqq));
3337
3338	return (!bfq_bfqq_wait_request(bfqq) ||
3339		bfq_bfqq_budget_left(bfqq) >=  bfqq->entity.budget / 3)
3340		&&
3341		bfq_bfqq_budget_timeout(bfqq);
3342}
3343
3344/*
3345 * For a queue that becomes empty, device idling is allowed only if
3346 * this function returns true for the queue. As a consequence, since
3347 * device idling plays a critical role in both throughput boosting and
3348 * service guarantees, the return value of this function plays a
3349 * critical role in both these aspects as well.
3350 *
3351 * In a nutshell, this function returns true only if idling is
3352 * beneficial for throughput or, even if detrimental for throughput,
3353 * idling is however necessary to preserve service guarantees (low
3354 * latency, desired throughput distribution, ...). In particular, on
3355 * NCQ-capable devices, this function tries to return false, so as to
3356 * help keep the drives' internal queues full, whenever this helps the
3357 * device boost the throughput without causing any service-guarantee
3358 * issue.
3359 *
3360 * In more detail, the return value of this function is obtained by,
3361 * first, computing a number of boolean variables that take into
3362 * account throughput and service-guarantee issues, and, then,
3363 * combining these variables in a logical expression. Most of the
3364 * issues taken into account are not trivial. We discuss these issues
3365 * individually while introducing the variables.
3366 */
3367static bool bfq_bfqq_may_idle(struct bfq_queue *bfqq)
3368{
3369	struct bfq_data *bfqd = bfqq->bfqd;
3370	bool rot_without_queueing =
3371		!blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
3372		bfqq_sequential_and_IO_bound,
3373		idling_boosts_thr, idling_boosts_thr_without_issues,
3374		idling_needed_for_service_guarantees,
3375		asymmetric_scenario;
3376
3377	if (bfqd->strict_guarantees)
3378		return true;
3379
3380	/*
3381	 * Idling is performed only if slice_idle > 0. In addition, we
3382	 * do not idle if
3383	 * (a) bfqq is async
3384	 * (b) bfqq is in the idle io prio class: in this case we do
3385	 * not idle because we want to minimize the bandwidth that
3386	 * queues in this class can steal to higher-priority queues
3387	 */
3388	if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
3389	    bfq_class_idle(bfqq))
3390		return false;
3391
3392	bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
3393		bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
3394
3395	/*
3396	 * The next variable takes into account the cases where idling
3397	 * boosts the throughput.
3398	 *
3399	 * The value of the variable is computed considering, first, that
3400	 * idling is virtually always beneficial for the throughput if:
3401	 * (a) the device is not NCQ-capable and rotational, or
3402	 * (b) regardless of the presence of NCQ, the device is rotational and
3403	 *     the request pattern for bfqq is I/O-bound and sequential, or
3404	 * (c) regardless of whether it is rotational, the device is
3405	 *     not NCQ-capable and the request pattern for bfqq is
3406	 *     I/O-bound and sequential.
3407	 *
3408	 * Secondly, and in contrast to the above item (b), idling an
3409	 * NCQ-capable flash-based device would not boost the
3410	 * throughput even with sequential I/O; rather it would lower
3411	 * the throughput in proportion to how fast the device
3412	 * is. Accordingly, the next variable is true if any of the
3413	 * above conditions (a), (b) or (c) is true, and, in
3414	 * particular, happens to be false if bfqd is an NCQ-capable
3415	 * flash-based device.
3416	 */
3417	idling_boosts_thr = rot_without_queueing ||
3418		((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
3419		 bfqq_sequential_and_IO_bound);
3420
3421	/*
3422	 * The value of the next variable,
3423	 * idling_boosts_thr_without_issues, is equal to that of
3424	 * idling_boosts_thr, unless a special case holds. In this
3425	 * special case, described below, idling may cause problems to
3426	 * weight-raised queues.
3427	 *
3428	 * When the request pool is saturated (e.g., in the presence
3429	 * of write hogs), if the processes associated with
3430	 * non-weight-raised queues ask for requests at a lower rate,
3431	 * then processes associated with weight-raised queues have a
3432	 * higher probability to get a request from the pool
3433	 * immediately (or at least soon) when they need one. Thus
3434	 * they have a higher probability to actually get a fraction
3435	 * of the device throughput proportional to their high
3436	 * weight. This is especially true with NCQ-capable drives,
3437	 * which enqueue several requests in advance, and further
3438	 * reorder internally-queued requests.
3439	 *
3440	 * For this reason, we force to false the value of
3441	 * idling_boosts_thr_without_issues if there are weight-raised
3442	 * busy queues. In this case, and if bfqq is not weight-raised,
3443	 * this guarantees that the device is not idled for bfqq (if,
3444	 * instead, bfqq is weight-raised, then idling will be
3445	 * guaranteed by another variable, see below). Combined with
3446	 * the timestamping rules of BFQ (see [1] for details), this
3447	 * behavior causes bfqq, and hence any sync non-weight-raised
3448	 * queue, to get a lower number of requests served, and thus
3449	 * to ask for a lower number of requests from the request
3450	 * pool, before the busy weight-raised queues get served
3451	 * again. This often mitigates starvation problems in the
3452	 * presence of heavy write workloads and NCQ, thereby
3453	 * guaranteeing a higher application and system responsiveness
3454	 * in these hostile scenarios.
3455	 */
3456	idling_boosts_thr_without_issues = idling_boosts_thr &&
3457		bfqd->wr_busy_queues == 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3458
3459	/*
3460	 * There is then a case where idling must be performed not
3461	 * for throughput concerns, but to preserve service
3462	 * guarantees.
3463	 *
3464	 * To introduce this case, we can note that allowing the drive
3465	 * to enqueue more than one request at a time, and hence
3466	 * delegating de facto final scheduling decisions to the
3467	 * drive's internal scheduler, entails loss of control on the
3468	 * actual request service order. In particular, the critical
3469	 * situation is when requests from different processes happen
3470	 * to be present, at the same time, in the internal queue(s)
3471	 * of the drive. In such a situation, the drive, by deciding
3472	 * the service order of the internally-queued requests, does
3473	 * determine also the actual throughput distribution among
3474	 * these processes. But the drive typically has no notion or
3475	 * concern about per-process throughput distribution, and
3476	 * makes its decisions only on a per-request basis. Therefore,
3477	 * the service distribution enforced by the drive's internal
3478	 * scheduler is likely to coincide with the desired
3479	 * device-throughput distribution only in a completely
3480	 * symmetric scenario where:
3481	 * (i)  each of these processes must get the same throughput as
3482	 *      the others;
3483	 * (ii) all these processes have the same I/O pattern
3484		(either sequential or random).
3485	 * In fact, in such a scenario, the drive will tend to treat
3486	 * the requests of each of these processes in about the same
3487	 * way as the requests of the others, and thus to provide
3488	 * each of these processes with about the same throughput
3489	 * (which is exactly the desired throughput distribution). In
3490	 * contrast, in any asymmetric scenario, device idling is
3491	 * certainly needed to guarantee that bfqq receives its
3492	 * assigned fraction of the device throughput (see [1] for
3493	 * details).
3494	 *
3495	 * We address this issue by controlling, actually, only the
3496	 * symmetry sub-condition (i), i.e., provided that
3497	 * sub-condition (i) holds, idling is not performed,
3498	 * regardless of whether sub-condition (ii) holds. In other
3499	 * words, only if sub-condition (i) holds, then idling is
3500	 * allowed, and the device tends to be prevented from queueing
3501	 * many requests, possibly of several processes. The reason
3502	 * for not controlling also sub-condition (ii) is that we
3503	 * exploit preemption to preserve guarantees in case of
3504	 * symmetric scenarios, even if (ii) does not hold, as
3505	 * explained in the next two paragraphs.
3506	 *
3507	 * Even if a queue, say Q, is expired when it remains idle, Q
3508	 * can still preempt the new in-service queue if the next
3509	 * request of Q arrives soon (see the comments on
3510	 * bfq_bfqq_update_budg_for_activation). If all queues and
3511	 * groups have the same weight, this form of preemption,
3512	 * combined with the hole-recovery heuristic described in the
3513	 * comments on function bfq_bfqq_update_budg_for_activation,
3514	 * are enough to preserve a correct bandwidth distribution in
3515	 * the mid term, even without idling. In fact, even if not
3516	 * idling allows the internal queues of the device to contain
3517	 * many requests, and thus to reorder requests, we can rather
3518	 * safely assume that the internal scheduler still preserves a
3519	 * minimum of mid-term fairness. The motivation for using
3520	 * preemption instead of idling is that, by not idling,
3521	 * service guarantees are preserved without minimally
3522	 * sacrificing throughput. In other words, both a high
3523	 * throughput and its desired distribution are obtained.
3524	 *
3525	 * More precisely, this preemption-based, idleless approach
3526	 * provides fairness in terms of IOPS, and not sectors per
3527	 * second. This can be seen with a simple example. Suppose
3528	 * that there are two queues with the same weight, but that
3529	 * the first queue receives requests of 8 sectors, while the
3530	 * second queue receives requests of 1024 sectors. In
3531	 * addition, suppose that each of the two queues contains at
3532	 * most one request at a time, which implies that each queue
3533	 * always remains idle after it is served. Finally, after
3534	 * remaining idle, each queue receives very quickly a new
3535	 * request. It follows that the two queues are served
3536	 * alternatively, preempting each other if needed. This
3537	 * implies that, although both queues have the same weight,
3538	 * the queue with large requests receives a service that is
3539	 * 1024/8 times as high as the service received by the other
3540	 * queue.
3541	 *
3542	 * On the other hand, device idling is performed, and thus
3543	 * pure sector-domain guarantees are provided, for the
3544	 * following queues, which are likely to need stronger
3545	 * throughput guarantees: weight-raised queues, and queues
3546	 * with a higher weight than other queues. When such queues
3547	 * are active, sub-condition (i) is false, which triggers
3548	 * device idling.
3549	 *
3550	 * According to the above considerations, the next variable is
3551	 * true (only) if sub-condition (i) holds. To compute the
3552	 * value of this variable, we not only use the return value of
3553	 * the function bfq_symmetric_scenario(), but also check
3554	 * whether bfqq is being weight-raised, because
3555	 * bfq_symmetric_scenario() does not take into account also
3556	 * weight-raised queues (see comments on
3557	 * bfq_weights_tree_add()).
3558	 *
3559	 * As a side note, it is worth considering that the above
3560	 * device-idling countermeasures may however fail in the
3561	 * following unlucky scenario: if idling is (correctly)
3562	 * disabled in a time period during which all symmetry
3563	 * sub-conditions hold, and hence the device is allowed to
3564	 * enqueue many requests, but at some later point in time some
3565	 * sub-condition stops to hold, then it may become impossible
3566	 * to let requests be served in the desired order until all
3567	 * the requests already queued in the device have been served.
3568	 */
3569	asymmetric_scenario = bfqq->wr_coeff > 1 ||
3570		!bfq_symmetric_scenario(bfqd);
3571
3572	/*
3573	 * Finally, there is a case where maximizing throughput is the
3574	 * best choice even if it may cause unfairness toward
3575	 * bfqq. Such a case is when bfqq became active in a burst of
3576	 * queue activations. Queues that became active during a large
3577	 * burst benefit only from throughput, as discussed in the
3578	 * comments on bfq_handle_burst. Thus, if bfqq became active
3579	 * in a burst and not idling the device maximizes throughput,
3580	 * then the device must no be idled, because not idling the
3581	 * device provides bfqq and all other queues in the burst with
3582	 * maximum benefit. Combining this and the above case, we can
3583	 * now establish when idling is actually needed to preserve
3584	 * service guarantees.
3585	 */
3586	idling_needed_for_service_guarantees =
3587		asymmetric_scenario && !bfq_bfqq_in_large_burst(bfqq);
 
 
 
 
 
 
 
3588
3589	/*
3590	 * We have now all the components we need to compute the
3591	 * return value of the function, which is true only if idling
3592	 * either boosts the throughput (without issues), or is
3593	 * necessary to preserve service guarantees.
3594	 */
3595	return idling_boosts_thr_without_issues ||
3596		idling_needed_for_service_guarantees;
3597}
3598
3599/*
3600 * If the in-service queue is empty but the function bfq_bfqq_may_idle
3601 * returns true, then:
3602 * 1) the queue must remain in service and cannot be expired, and
3603 * 2) the device must be idled to wait for the possible arrival of a new
3604 *    request for the queue.
3605 * See the comments on the function bfq_bfqq_may_idle for the reasons
3606 * why performing device idling is the best choice to boost the throughput
3607 * and preserve service guarantees when bfq_bfqq_may_idle itself
3608 * returns true.
3609 */
3610static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
3611{
3612	return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_bfqq_may_idle(bfqq);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3613}
3614
3615/*
3616 * Select a queue for service.  If we have a current queue in service,
3617 * check whether to continue servicing it, or retrieve and set a new one.
3618 */
3619static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
3620{
3621	struct bfq_queue *bfqq;
3622	struct request *next_rq;
3623	enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
3624
3625	bfqq = bfqd->in_service_queue;
3626	if (!bfqq)
3627		goto new_queue;
3628
3629	bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
3630
 
 
 
 
 
 
 
3631	if (bfq_may_expire_for_budg_timeout(bfqq) &&
3632	    !bfq_bfqq_wait_request(bfqq) &&
3633	    !bfq_bfqq_must_idle(bfqq))
3634		goto expire;
3635
3636check_queue:
3637	/*
3638	 * This loop is rarely executed more than once. Even when it
3639	 * happens, it is much more convenient to re-execute this loop
3640	 * than to return NULL and trigger a new dispatch to get a
3641	 * request served.
3642	 */
3643	next_rq = bfqq->next_rq;
3644	/*
3645	 * If bfqq has requests queued and it has enough budget left to
3646	 * serve them, keep the queue, otherwise expire it.
3647	 */
3648	if (next_rq) {
3649		if (bfq_serv_to_charge(next_rq, bfqq) >
3650			bfq_bfqq_budget_left(bfqq)) {
3651			/*
3652			 * Expire the queue for budget exhaustion,
3653			 * which makes sure that the next budget is
3654			 * enough to serve the next request, even if
3655			 * it comes from the fifo expired path.
3656			 */
3657			reason = BFQQE_BUDGET_EXHAUSTED;
3658			goto expire;
3659		} else {
3660			/*
3661			 * The idle timer may be pending because we may
3662			 * not disable disk idling even when a new request
3663			 * arrives.
3664			 */
3665			if (bfq_bfqq_wait_request(bfqq)) {
3666				/*
3667				 * If we get here: 1) at least a new request
3668				 * has arrived but we have not disabled the
3669				 * timer because the request was too small,
3670				 * 2) then the block layer has unplugged
3671				 * the device, causing the dispatch to be
3672				 * invoked.
3673				 *
3674				 * Since the device is unplugged, now the
3675				 * requests are probably large enough to
3676				 * provide a reasonable throughput.
3677				 * So we disable idling.
3678				 */
3679				bfq_clear_bfqq_wait_request(bfqq);
3680				hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
3681			}
3682			goto keep_queue;
3683		}
3684	}
3685
3686	/*
3687	 * No requests pending. However, if the in-service queue is idling
3688	 * for a new request, or has requests waiting for a completion and
3689	 * may idle after their completion, then keep it anyway.
 
 
 
3690	 */
3691	if (bfq_bfqq_wait_request(bfqq) ||
3692	    (bfqq->dispatched != 0 && bfq_bfqq_may_idle(bfqq))) {
3693		bfqq = NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3694		goto keep_queue;
3695	}
3696
3697	reason = BFQQE_NO_MORE_REQUESTS;
3698expire:
3699	bfq_bfqq_expire(bfqd, bfqq, false, reason);
3700new_queue:
3701	bfqq = bfq_set_in_service_queue(bfqd);
3702	if (bfqq) {
3703		bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
3704		goto check_queue;
3705	}
3706keep_queue:
3707	if (bfqq)
3708		bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
3709	else
3710		bfq_log(bfqd, "select_queue: no queue returned");
3711
3712	return bfqq;
3713}
3714
3715static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3716{
3717	struct bfq_entity *entity = &bfqq->entity;
3718
3719	if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
3720		bfq_log_bfqq(bfqd, bfqq,
3721			"raising period dur %u/%u msec, old coeff %u, w %d(%d)",
3722			jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
3723			jiffies_to_msecs(bfqq->wr_cur_max_time),
3724			bfqq->wr_coeff,
3725			bfqq->entity.weight, bfqq->entity.orig_weight);
3726
3727		if (entity->prio_changed)
3728			bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
3729
3730		/*
3731		 * If the queue was activated in a burst, or too much
3732		 * time has elapsed from the beginning of this
3733		 * weight-raising period, then end weight raising.
3734		 */
3735		if (bfq_bfqq_in_large_burst(bfqq))
3736			bfq_bfqq_end_wr(bfqq);
3737		else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
3738						bfqq->wr_cur_max_time)) {
3739			if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
3740			time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
3741					       bfq_wr_duration(bfqd)))
 
 
 
 
 
 
 
 
3742				bfq_bfqq_end_wr(bfqq);
3743			else {
 
 
 
 
3744				switch_back_to_interactive_wr(bfqq, bfqd);
3745				bfqq->entity.prio_changed = 1;
3746			}
3747		}
3748		if (bfqq->wr_coeff > 1 &&
3749		    bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
3750		    bfqq->service_from_wr > max_service_from_wr) {
3751			/* see comments on max_service_from_wr */
3752			bfq_bfqq_end_wr(bfqq);
3753		}
3754	}
3755	/*
3756	 * To improve latency (for this or other queues), immediately
3757	 * update weight both if it must be raised and if it must be
3758	 * lowered. Since, entity may be on some active tree here, and
3759	 * might have a pending change of its ioprio class, invoke
3760	 * next function with the last parameter unset (see the
3761	 * comments on the function).
3762	 */
3763	if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
3764		__bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
3765						entity, false);
3766}
3767
3768/*
3769 * Dispatch next request from bfqq.
3770 */
3771static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
3772						 struct bfq_queue *bfqq)
3773{
3774	struct request *rq = bfqq->next_rq;
3775	unsigned long service_to_charge;
3776
3777	service_to_charge = bfq_serv_to_charge(rq, bfqq);
3778
3779	bfq_bfqq_served(bfqq, service_to_charge);
3780
 
 
 
 
 
3781	bfq_dispatch_remove(bfqd->queue, rq);
3782
 
 
 
3783	/*
3784	 * If weight raising has to terminate for bfqq, then next
3785	 * function causes an immediate update of bfqq's weight,
3786	 * without waiting for next activation. As a consequence, on
3787	 * expiration, bfqq will be timestamped as if has never been
3788	 * weight-raised during this service slot, even if it has
3789	 * received part or even most of the service as a
3790	 * weight-raised queue. This inflates bfqq's timestamps, which
3791	 * is beneficial, as bfqq is then more willing to leave the
3792	 * device immediately to possible other weight-raised queues.
3793	 */
3794	bfq_update_wr_data(bfqd, bfqq);
3795
3796	/*
3797	 * Expire bfqq, pretending that its budget expired, if bfqq
3798	 * belongs to CLASS_IDLE and other queues are waiting for
3799	 * service.
3800	 */
3801	if (bfqd->busy_queues > 1 && bfq_class_idle(bfqq))
3802		goto expire;
3803
3804	return rq;
3805
3806expire:
3807	bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
3808	return rq;
3809}
3810
3811static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
3812{
3813	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3814
3815	/*
3816	 * Avoiding lock: a race on bfqd->busy_queues should cause at
3817	 * most a call to dispatch for nothing
3818	 */
3819	return !list_empty_careful(&bfqd->dispatch) ||
3820		bfqd->busy_queues > 0;
3821}
3822
3823static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3824{
3825	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3826	struct request *rq = NULL;
3827	struct bfq_queue *bfqq = NULL;
3828
3829	if (!list_empty(&bfqd->dispatch)) {
3830		rq = list_first_entry(&bfqd->dispatch, struct request,
3831				      queuelist);
3832		list_del_init(&rq->queuelist);
3833
3834		bfqq = RQ_BFQQ(rq);
3835
3836		if (bfqq) {
3837			/*
3838			 * Increment counters here, because this
3839			 * dispatch does not follow the standard
3840			 * dispatch flow (where counters are
3841			 * incremented)
3842			 */
3843			bfqq->dispatched++;
3844
3845			goto inc_in_driver_start_rq;
3846		}
3847
3848		/*
3849		 * We exploit the bfq_finish_requeue_request hook to
3850		 * decrement rq_in_driver, but
3851		 * bfq_finish_requeue_request will not be invoked on
3852		 * this request. So, to avoid unbalance, just start
3853		 * this request, without incrementing rq_in_driver. As
3854		 * a negative consequence, rq_in_driver is deceptively
3855		 * lower than it should be while this request is in
3856		 * service. This may cause bfq_schedule_dispatch to be
3857		 * invoked uselessly.
3858		 *
3859		 * As for implementing an exact solution, the
3860		 * bfq_finish_requeue_request hook, if defined, is
3861		 * probably invoked also on this request. So, by
3862		 * exploiting this hook, we could 1) increment
3863		 * rq_in_driver here, and 2) decrement it in
3864		 * bfq_finish_requeue_request. Such a solution would
3865		 * let the value of the counter be always accurate,
3866		 * but it would entail using an extra interface
3867		 * function. This cost seems higher than the benefit,
3868		 * being the frequency of non-elevator-private
3869		 * requests very low.
3870		 */
3871		goto start_rq;
3872	}
3873
3874	bfq_log(bfqd, "dispatch requests: %d busy queues", bfqd->busy_queues);
 
3875
3876	if (bfqd->busy_queues == 0)
3877		goto exit;
3878
3879	/*
3880	 * Force device to serve one request at a time if
3881	 * strict_guarantees is true. Forcing this service scheme is
3882	 * currently the ONLY way to guarantee that the request
3883	 * service order enforced by the scheduler is respected by a
3884	 * queueing device. Otherwise the device is free even to make
3885	 * some unlucky request wait for as long as the device
3886	 * wishes.
3887	 *
3888	 * Of course, serving one request at at time may cause loss of
3889	 * throughput.
3890	 */
3891	if (bfqd->strict_guarantees && bfqd->rq_in_driver > 0)
3892		goto exit;
3893
3894	bfqq = bfq_select_queue(bfqd);
3895	if (!bfqq)
3896		goto exit;
3897
3898	rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
3899
3900	if (rq) {
3901inc_in_driver_start_rq:
3902		bfqd->rq_in_driver++;
3903start_rq:
3904		rq->rq_flags |= RQF_STARTED;
3905	}
3906exit:
3907	return rq;
3908}
3909
3910#if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
3911static void bfq_update_dispatch_stats(struct request_queue *q,
3912				      struct request *rq,
3913				      struct bfq_queue *in_serv_queue,
3914				      bool idle_timer_disabled)
3915{
3916	struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
3917
3918	if (!idle_timer_disabled && !bfqq)
3919		return;
3920
3921	/*
3922	 * rq and bfqq are guaranteed to exist until this function
3923	 * ends, for the following reasons. First, rq can be
3924	 * dispatched to the device, and then can be completed and
3925	 * freed, only after this function ends. Second, rq cannot be
3926	 * merged (and thus freed because of a merge) any longer,
3927	 * because it has already started. Thus rq cannot be freed
3928	 * before this function ends, and, since rq has a reference to
3929	 * bfqq, the same guarantee holds for bfqq too.
3930	 *
3931	 * In addition, the following queue lock guarantees that
3932	 * bfqq_group(bfqq) exists as well.
3933	 */
3934	spin_lock_irq(q->queue_lock);
3935	if (idle_timer_disabled)
3936		/*
3937		 * Since the idle timer has been disabled,
3938		 * in_serv_queue contained some request when
3939		 * __bfq_dispatch_request was invoked above, which
3940		 * implies that rq was picked exactly from
3941		 * in_serv_queue. Thus in_serv_queue == bfqq, and is
3942		 * therefore guaranteed to exist because of the above
3943		 * arguments.
3944		 */
3945		bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
3946	if (bfqq) {
3947		struct bfq_group *bfqg = bfqq_group(bfqq);
3948
3949		bfqg_stats_update_avg_queue_size(bfqg);
3950		bfqg_stats_set_start_empty_time(bfqg);
3951		bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
3952	}
3953	spin_unlock_irq(q->queue_lock);
3954}
3955#else
3956static inline void bfq_update_dispatch_stats(struct request_queue *q,
3957					     struct request *rq,
3958					     struct bfq_queue *in_serv_queue,
3959					     bool idle_timer_disabled) {}
3960#endif
3961
3962static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
3963{
3964	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
3965	struct request *rq;
3966	struct bfq_queue *in_serv_queue;
3967	bool waiting_rq, idle_timer_disabled;
3968
3969	spin_lock_irq(&bfqd->lock);
3970
3971	in_serv_queue = bfqd->in_service_queue;
3972	waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
3973
3974	rq = __bfq_dispatch_request(hctx);
3975
3976	idle_timer_disabled =
3977		waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
 
3978
3979	spin_unlock_irq(&bfqd->lock);
3980
3981	bfq_update_dispatch_stats(hctx->queue, rq, in_serv_queue,
3982				  idle_timer_disabled);
3983
3984	return rq;
3985}
3986
3987/*
3988 * Task holds one reference to the queue, dropped when task exits.  Each rq
3989 * in-flight on this queue also holds a reference, dropped when rq is freed.
3990 *
3991 * Scheduler lock must be held here. Recall not to use bfqq after calling
3992 * this function on it.
3993 */
3994void bfq_put_queue(struct bfq_queue *bfqq)
3995{
3996#ifdef CONFIG_BFQ_GROUP_IOSCHED
 
3997	struct bfq_group *bfqg = bfqq_group(bfqq);
3998#endif
3999
4000	if (bfqq->bfqd)
4001		bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d",
4002			     bfqq, bfqq->ref);
4003
4004	bfqq->ref--;
4005	if (bfqq->ref)
4006		return;
4007
4008	if (!hlist_unhashed(&bfqq->burst_list_node)) {
4009		hlist_del_init(&bfqq->burst_list_node);
4010		/*
4011		 * Decrement also burst size after the removal, if the
4012		 * process associated with bfqq is exiting, and thus
4013		 * does not contribute to the burst any longer. This
4014		 * decrement helps filter out false positives of large
4015		 * bursts, when some short-lived process (often due to
4016		 * the execution of commands by some service) happens
4017		 * to start and exit while a complex application is
4018		 * starting, and thus spawning several processes that
4019		 * do I/O (and that *must not* be treated as a large
4020		 * burst, see comments on bfq_handle_burst).
4021		 *
4022		 * In particular, the decrement is performed only if:
4023		 * 1) bfqq is not a merged queue, because, if it is,
4024		 * then this free of bfqq is not triggered by the exit
4025		 * of the process bfqq is associated with, but exactly
4026		 * by the fact that bfqq has just been merged.
4027		 * 2) burst_size is greater than 0, to handle
4028		 * unbalanced decrements. Unbalanced decrements may
4029		 * happen in te following case: bfqq is inserted into
4030		 * the current burst list--without incrementing
4031		 * bust_size--because of a split, but the current
4032		 * burst list is not the burst list bfqq belonged to
4033		 * (see comments on the case of a split in
4034		 * bfq_set_request).
4035		 */
4036		if (bfqq->bic && bfqq->bfqd->burst_size > 0)
4037			bfqq->bfqd->burst_size--;
4038	}
4039
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4040	kmem_cache_free(bfq_pool, bfqq);
4041#ifdef CONFIG_BFQ_GROUP_IOSCHED
4042	bfqg_and_blkg_put(bfqg);
4043#endif
4044}
4045
4046static void bfq_put_cooperator(struct bfq_queue *bfqq)
 
 
 
 
 
 
4047{
4048	struct bfq_queue *__bfqq, *next;
4049
4050	/*
4051	 * If this queue was scheduled to merge with another queue, be
4052	 * sure to drop the reference taken on that queue (and others in
4053	 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
4054	 */
4055	__bfqq = bfqq->new_bfqq;
4056	while (__bfqq) {
4057		if (__bfqq == bfqq)
4058			break;
4059		next = __bfqq->new_bfqq;
4060		bfq_put_queue(__bfqq);
4061		__bfqq = next;
4062	}
4063}
4064
4065static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
4066{
4067	if (bfqq == bfqd->in_service_queue) {
4068		__bfq_bfqq_expire(bfqd, bfqq);
4069		bfq_schedule_dispatch(bfqd);
4070	}
4071
4072	bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
4073
4074	bfq_put_cooperator(bfqq);
4075
4076	bfq_put_queue(bfqq); /* release process reference */
4077}
4078
4079static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync)
4080{
4081	struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4082	struct bfq_data *bfqd;
4083
4084	if (bfqq)
4085		bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
4086
4087	if (bfqq && bfqd) {
4088		unsigned long flags;
4089
4090		spin_lock_irqsave(&bfqd->lock, flags);
 
4091		bfq_exit_bfqq(bfqd, bfqq);
4092		bic_set_bfqq(bic, NULL, is_sync);
4093		spin_unlock_irqrestore(&bfqd->lock, flags);
4094	}
4095}
4096
4097static void bfq_exit_icq(struct io_cq *icq)
4098{
4099	struct bfq_io_cq *bic = icq_to_bic(icq);
4100
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4101	bfq_exit_icq_bfqq(bic, true);
4102	bfq_exit_icq_bfqq(bic, false);
4103}
4104
4105/*
4106 * Update the entity prio values; note that the new values will not
4107 * be used until the next (re)activation.
4108 */
4109static void
4110bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
4111{
4112	struct task_struct *tsk = current;
4113	int ioprio_class;
4114	struct bfq_data *bfqd = bfqq->bfqd;
4115
4116	if (!bfqd)
4117		return;
4118
4119	ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4120	switch (ioprio_class) {
4121	default:
4122		dev_err(bfqq->bfqd->queue->backing_dev_info->dev,
4123			"bfq: bad prio class %d\n", ioprio_class);
4124		/* fall through */
 
4125	case IOPRIO_CLASS_NONE:
4126		/*
4127		 * No prio set, inherit CPU scheduling settings.
4128		 */
4129		bfqq->new_ioprio = task_nice_ioprio(tsk);
4130		bfqq->new_ioprio_class = task_nice_ioclass(tsk);
4131		break;
4132	case IOPRIO_CLASS_RT:
4133		bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4134		bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
4135		break;
4136	case IOPRIO_CLASS_BE:
4137		bfqq->new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4138		bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
4139		break;
4140	case IOPRIO_CLASS_IDLE:
4141		bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
4142		bfqq->new_ioprio = 7;
4143		break;
4144	}
4145
4146	if (bfqq->new_ioprio >= IOPRIO_BE_NR) {
4147		pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
4148			bfqq->new_ioprio);
4149		bfqq->new_ioprio = IOPRIO_BE_NR;
4150	}
4151
4152	bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
 
 
4153	bfqq->entity.prio_changed = 1;
4154}
4155
4156static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4157				       struct bio *bio, bool is_sync,
4158				       struct bfq_io_cq *bic);
 
4159
4160static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
4161{
4162	struct bfq_data *bfqd = bic_to_bfqd(bic);
4163	struct bfq_queue *bfqq;
4164	int ioprio = bic->icq.ioc->ioprio;
4165
4166	/*
4167	 * This condition may trigger on a newly created bic, be sure to
4168	 * drop the lock before returning.
4169	 */
4170	if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
4171		return;
4172
4173	bic->ioprio = ioprio;
4174
4175	bfqq = bic_to_bfqq(bic, false);
4176	if (bfqq) {
4177		/* release process reference on this queue */
4178		bfq_put_queue(bfqq);
4179		bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
4180		bic_set_bfqq(bic, bfqq, false);
 
4181	}
4182
4183	bfqq = bic_to_bfqq(bic, true);
4184	if (bfqq)
4185		bfq_set_next_ioprio_data(bfqq, bic);
4186}
4187
4188static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4189			  struct bfq_io_cq *bic, pid_t pid, int is_sync)
4190{
 
 
4191	RB_CLEAR_NODE(&bfqq->entity.rb_node);
4192	INIT_LIST_HEAD(&bfqq->fifo);
4193	INIT_HLIST_NODE(&bfqq->burst_list_node);
 
 
4194
4195	bfqq->ref = 0;
4196	bfqq->bfqd = bfqd;
4197
4198	if (bic)
4199		bfq_set_next_ioprio_data(bfqq, bic);
4200
4201	if (is_sync) {
4202		/*
4203		 * No need to mark as has_short_ttime if in
4204		 * idle_class, because no device idling is performed
4205		 * for queues in idle class
4206		 */
4207		if (!bfq_class_idle(bfqq))
4208			/* tentatively mark as has_short_ttime */
4209			bfq_mark_bfqq_has_short_ttime(bfqq);
4210		bfq_mark_bfqq_sync(bfqq);
4211		bfq_mark_bfqq_just_created(bfqq);
4212	} else
4213		bfq_clear_bfqq_sync(bfqq);
4214
4215	/* set end request to minus infinity from now */
4216	bfqq->ttime.last_end_request = ktime_get_ns() + 1;
 
 
 
 
4217
4218	bfq_mark_bfqq_IO_bound(bfqq);
4219
4220	bfqq->pid = pid;
4221
4222	/* Tentative initial value to trade off between thr and lat */
4223	bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
4224	bfqq->budget_timeout = bfq_smallest_from_now();
4225
4226	bfqq->wr_coeff = 1;
4227	bfqq->last_wr_start_finish = jiffies;
4228	bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
4229	bfqq->split_time = bfq_smallest_from_now();
4230
4231	/*
4232	 * To not forget the possibly high bandwidth consumed by a
4233	 * process/queue in the recent past,
4234	 * bfq_bfqq_softrt_next_start() returns a value at least equal
4235	 * to the current value of bfqq->soft_rt_next_start (see
4236	 * comments on bfq_bfqq_softrt_next_start).  Set
4237	 * soft_rt_next_start to now, to mean that bfqq has consumed
4238	 * no bandwidth so far.
4239	 */
4240	bfqq->soft_rt_next_start = jiffies;
4241
4242	/* first request is almost certainly seeky */
4243	bfqq->seek_history = 1;
4244}
4245
4246static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
4247					       struct bfq_group *bfqg,
4248					       int ioprio_class, int ioprio)
4249{
4250	switch (ioprio_class) {
4251	case IOPRIO_CLASS_RT:
4252		return &bfqg->async_bfqq[0][ioprio];
4253	case IOPRIO_CLASS_NONE:
4254		ioprio = IOPRIO_NORM;
4255		/* fall through */
4256	case IOPRIO_CLASS_BE:
4257		return &bfqg->async_bfqq[1][ioprio];
4258	case IOPRIO_CLASS_IDLE:
4259		return &bfqg->async_idle_bfqq;
4260	default:
4261		return NULL;
4262	}
4263}
4264
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4265static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
4266				       struct bio *bio, bool is_sync,
4267				       struct bfq_io_cq *bic)
 
4268{
4269	const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
4270	const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
4271	struct bfq_queue **async_bfqq = NULL;
4272	struct bfq_queue *bfqq;
4273	struct bfq_group *bfqg;
4274
4275	rcu_read_lock();
4276
4277	bfqg = bfq_find_set_group(bfqd, bio_blkcg(bio));
4278	if (!bfqg) {
4279		bfqq = &bfqd->oom_bfqq;
4280		goto out;
4281	}
4282
4283	if (!is_sync) {
4284		async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
4285						  ioprio);
4286		bfqq = *async_bfqq;
4287		if (bfqq)
4288			goto out;
4289	}
4290
4291	bfqq = kmem_cache_alloc_node(bfq_pool,
4292				     GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
4293				     bfqd->queue->node);
4294
4295	if (bfqq) {
4296		bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
4297			      is_sync);
4298		bfq_init_entity(&bfqq->entity, bfqg);
4299		bfq_log_bfqq(bfqd, bfqq, "allocated");
4300	} else {
4301		bfqq = &bfqd->oom_bfqq;
4302		bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
4303		goto out;
4304	}
4305
4306	/*
4307	 * Pin the queue now that it's allocated, scheduler exit will
4308	 * prune it.
4309	 */
4310	if (async_bfqq) {
4311		bfqq->ref++; /*
4312			      * Extra group reference, w.r.t. sync
4313			      * queue. This extra reference is removed
4314			      * only if bfqq->bfqg disappears, to
4315			      * guarantee that this queue is not freed
4316			      * until its group goes away.
4317			      */
4318		bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
4319			     bfqq, bfqq->ref);
4320		*async_bfqq = bfqq;
4321	}
4322
4323out:
4324	bfqq->ref++; /* get a process reference to this queue */
4325	bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
4326	rcu_read_unlock();
 
4327	return bfqq;
4328}
4329
4330static void bfq_update_io_thinktime(struct bfq_data *bfqd,
4331				    struct bfq_queue *bfqq)
4332{
4333	struct bfq_ttime *ttime = &bfqq->ttime;
4334	u64 elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
4335
 
 
 
 
 
 
 
 
4336	elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
4337
4338	ttime->ttime_samples = (7*bfqq->ttime.ttime_samples + 256) / 8;
4339	ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed,  8);
4340	ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
4341				     ttime->ttime_samples);
4342}
4343
4344static void
4345bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4346		       struct request *rq)
4347{
4348	bfqq->seek_history <<= 1;
4349	bfqq->seek_history |=
4350		get_sdist(bfqq->last_request_pos, rq) > BFQQ_SEEK_THR &&
4351		(!blk_queue_nonrot(bfqd->queue) ||
4352		 blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4353}
4354
4355static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
4356				       struct bfq_queue *bfqq,
4357				       struct bfq_io_cq *bic)
4358{
4359	bool has_short_ttime = true;
4360
4361	/*
4362	 * No need to update has_short_ttime if bfqq is async or in
4363	 * idle io prio class, or if bfq_slice_idle is zero, because
4364	 * no device idling is performed for bfqq in this case.
4365	 */
4366	if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
4367	    bfqd->bfq_slice_idle == 0)
4368		return;
4369
4370	/* Idle window just restored, statistics are meaningless. */
4371	if (time_is_after_eq_jiffies(bfqq->split_time +
4372				     bfqd->bfq_wr_min_idle_time))
4373		return;
4374
4375	/* Think time is infinite if no process is linked to
4376	 * bfqq. Otherwise check average think time to
4377	 * decide whether to mark as has_short_ttime
 
4378	 */
4379	if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
4380	    (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
4381	     bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle))
4382		has_short_ttime = false;
4383
4384	bfq_log_bfqq(bfqd, bfqq, "update_has_short_ttime: has_short_ttime %d",
4385		     has_short_ttime);
4386
4387	if (has_short_ttime)
4388		bfq_mark_bfqq_has_short_ttime(bfqq);
4389	else
4390		bfq_clear_bfqq_has_short_ttime(bfqq);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4391}
4392
4393/*
4394 * Called when a new fs request (rq) is added to bfqq.  Check if there's
4395 * something we should do about it.
4396 */
4397static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4398			    struct request *rq)
4399{
4400	struct bfq_io_cq *bic = RQ_BIC(rq);
4401
4402	if (rq->cmd_flags & REQ_META)
4403		bfqq->meta_pending++;
4404
4405	bfq_update_io_thinktime(bfqd, bfqq);
4406	bfq_update_has_short_ttime(bfqd, bfqq, bic);
4407	bfq_update_io_seektime(bfqd, bfqq, rq);
4408
4409	bfq_log_bfqq(bfqd, bfqq,
4410		     "rq_enqueued: has_short_ttime=%d (seeky %d)",
4411		     bfq_bfqq_has_short_ttime(bfqq), BFQQ_SEEKY(bfqq));
4412
4413	bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4414
4415	if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
4416		bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
4417				 blk_rq_sectors(rq) < 32;
4418		bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
4419
4420		/*
4421		 * There is just this request queued: if the request
4422		 * is small and the queue is not to be expired, then
4423		 * just exit.
 
 
4424		 *
4425		 * In this way, if the device is being idled to wait
4426		 * for a new request from the in-service queue, we
4427		 * avoid unplugging the device and committing the
4428		 * device to serve just a small request. On the
4429		 * contrary, we wait for the block layer to decide
4430		 * when to unplug the device: hopefully, new requests
4431		 * will be merged to this one quickly, then the device
4432		 * will be unplugged and larger requests will be
4433		 * dispatched.
4434		 */
4435		if (small_req && !budget_timeout)
 
4436			return;
4437
4438		/*
4439		 * A large enough request arrived, or the queue is to
4440		 * be expired: in both cases disk idling is to be
4441		 * stopped, so clear wait_request flag and reset
4442		 * timer.
 
4443		 */
4444		bfq_clear_bfqq_wait_request(bfqq);
4445		hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4446
4447		/*
4448		 * The queue is not empty, because a new request just
4449		 * arrived. Hence we can safely expire the queue, in
4450		 * case of budget timeout, without risking that the
4451		 * timestamps of the queue are not updated correctly.
4452		 * See [1] for more details.
4453		 */
4454		if (budget_timeout)
4455			bfq_bfqq_expire(bfqd, bfqq, false,
4456					BFQQE_BUDGET_TIMEOUT);
4457	}
4458}
4459
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4460/* returns true if it causes the idle timer to be disabled */
4461static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
4462{
4463	struct bfq_queue *bfqq = RQ_BFQQ(rq),
4464		*new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
 
4465	bool waiting, idle_timer_disabled = false;
4466
4467	if (new_bfqq) {
4468		if (bic_to_bfqq(RQ_BIC(rq), 1) != bfqq)
4469			new_bfqq = bic_to_bfqq(RQ_BIC(rq), 1);
4470		/*
4471		 * Release the request's reference to the old bfqq
4472		 * and make sure one is taken to the shared queue.
4473		 */
4474		new_bfqq->allocated++;
4475		bfqq->allocated--;
4476		new_bfqq->ref++;
4477		/*
4478		 * If the bic associated with the process
4479		 * issuing this request still points to bfqq
4480		 * (and thus has not been already redirected
4481		 * to new_bfqq or even some other bfq_queue),
4482		 * then complete the merge and redirect it to
4483		 * new_bfqq.
4484		 */
4485		if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq)
4486			bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
4487					bfqq, new_bfqq);
4488
4489		bfq_clear_bfqq_just_created(bfqq);
4490		/*
4491		 * rq is about to be enqueued into new_bfqq,
4492		 * release rq reference on bfqq
4493		 */
4494		bfq_put_queue(bfqq);
4495		rq->elv.priv[1] = new_bfqq;
4496		bfqq = new_bfqq;
4497	}
4498
 
 
 
 
4499	waiting = bfqq && bfq_bfqq_wait_request(bfqq);
4500	bfq_add_request(rq);
4501	idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
4502
4503	rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
4504	list_add_tail(&rq->queuelist, &bfqq->fifo);
4505
4506	bfq_rq_enqueued(bfqd, bfqq, rq);
4507
4508	return idle_timer_disabled;
4509}
4510
4511#if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
4512static void bfq_update_insert_stats(struct request_queue *q,
4513				    struct bfq_queue *bfqq,
4514				    bool idle_timer_disabled,
4515				    unsigned int cmd_flags)
4516{
4517	if (!bfqq)
4518		return;
4519
4520	/*
4521	 * bfqq still exists, because it can disappear only after
4522	 * either it is merged with another queue, or the process it
4523	 * is associated with exits. But both actions must be taken by
4524	 * the same process currently executing this flow of
4525	 * instructions.
4526	 *
4527	 * In addition, the following queue lock guarantees that
4528	 * bfqq_group(bfqq) exists as well.
4529	 */
4530	spin_lock_irq(q->queue_lock);
4531	bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
4532	if (idle_timer_disabled)
4533		bfqg_stats_update_idle_time(bfqq_group(bfqq));
4534	spin_unlock_irq(q->queue_lock);
4535}
4536#else
4537static inline void bfq_update_insert_stats(struct request_queue *q,
4538					   struct bfq_queue *bfqq,
4539					   bool idle_timer_disabled,
4540					   unsigned int cmd_flags) {}
4541#endif
4542
4543static void bfq_prepare_request(struct request *rq, struct bio *bio);
4544
4545static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
4546			       bool at_head)
4547{
4548	struct request_queue *q = hctx->queue;
4549	struct bfq_data *bfqd = q->elevator->elevator_data;
4550	struct bfq_queue *bfqq = RQ_BFQQ(rq);
4551	bool idle_timer_disabled = false;
4552	unsigned int cmd_flags;
 
4553
 
 
 
 
4554	spin_lock_irq(&bfqd->lock);
4555	if (blk_mq_sched_try_insert_merge(q, rq)) {
 
4556		spin_unlock_irq(&bfqd->lock);
 
4557		return;
4558	}
4559
4560	spin_unlock_irq(&bfqd->lock);
4561
4562	blk_mq_sched_request_inserted(rq);
4563
4564	spin_lock_irq(&bfqd->lock);
4565	if (at_head || blk_rq_is_passthrough(rq)) {
4566		if (at_head)
4567			list_add(&rq->queuelist, &bfqd->dispatch);
4568		else
4569			list_add_tail(&rq->queuelist, &bfqd->dispatch);
4570	} else {
4571		if (WARN_ON_ONCE(!bfqq)) {
4572			/*
4573			 * This should never happen. Most likely rq is
4574			 * a requeued regular request, being
4575			 * re-inserted without being first
4576			 * re-prepared. Do a prepare, to avoid
4577			 * failure.
4578			 */
4579			bfq_prepare_request(rq, rq->bio);
4580			bfqq = RQ_BFQQ(rq);
4581		}
4582
4583		idle_timer_disabled = __bfq_insert_request(bfqd, rq);
4584		/*
4585		 * Update bfqq, because, if a queue merge has occurred
4586		 * in __bfq_insert_request, then rq has been
4587		 * redirected into a new queue.
4588		 */
4589		bfqq = RQ_BFQQ(rq);
4590
4591		if (rq_mergeable(rq)) {
4592			elv_rqhash_add(q, rq);
4593			if (!q->last_merge)
4594				q->last_merge = rq;
4595		}
4596	}
4597
4598	/*
4599	 * Cache cmd_flags before releasing scheduler lock, because rq
4600	 * may disappear afterwards (for example, because of a request
4601	 * merge).
4602	 */
4603	cmd_flags = rq->cmd_flags;
4604
4605	spin_unlock_irq(&bfqd->lock);
4606
4607	bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
4608				cmd_flags);
4609}
4610
4611static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
4612				struct list_head *list, bool at_head)
4613{
4614	while (!list_empty(list)) {
4615		struct request *rq;
4616
4617		rq = list_first_entry(list, struct request, queuelist);
4618		list_del_init(&rq->queuelist);
4619		bfq_insert_request(hctx, rq, at_head);
4620	}
4621}
4622
4623static void bfq_update_hw_tag(struct bfq_data *bfqd)
4624{
 
 
4625	bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
4626				       bfqd->rq_in_driver);
4627
4628	if (bfqd->hw_tag == 1)
4629		return;
4630
4631	/*
4632	 * This sample is valid if the number of outstanding requests
4633	 * is large enough to allow a queueing behavior.  Note that the
4634	 * sum is not exact, as it's not taking into account deactivated
4635	 * requests.
4636	 */
4637	if (bfqd->rq_in_driver + bfqd->queued < BFQ_HW_QUEUE_THRESHOLD)
 
 
 
 
 
 
 
 
 
 
 
4638		return;
4639
4640	if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
4641		return;
4642
4643	bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
4644	bfqd->max_rq_in_driver = 0;
4645	bfqd->hw_tag_samples = 0;
 
 
 
4646}
4647
4648static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
4649{
4650	u64 now_ns;
4651	u32 delta_us;
4652
4653	bfq_update_hw_tag(bfqd);
4654
4655	bfqd->rq_in_driver--;
4656	bfqq->dispatched--;
4657
4658	if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
4659		/*
4660		 * Set budget_timeout (which we overload to store the
4661		 * time at which the queue remains with no backlog and
4662		 * no outstanding request; used by the weight-raising
4663		 * mechanism).
4664		 */
4665		bfqq->budget_timeout = jiffies;
4666
4667		bfq_weights_tree_remove(bfqd, &bfqq->entity,
4668					&bfqd->queue_weights_tree);
4669	}
4670
4671	now_ns = ktime_get_ns();
4672
4673	bfqq->ttime.last_end_request = now_ns;
4674
4675	/*
4676	 * Using us instead of ns, to get a reasonable precision in
4677	 * computing rate in next check.
4678	 */
4679	delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
4680
4681	/*
4682	 * If the request took rather long to complete, and, according
4683	 * to the maximum request size recorded, this completion latency
4684	 * implies that the request was certainly served at a very low
4685	 * rate (less than 1M sectors/sec), then the whole observation
4686	 * interval that lasts up to this time instant cannot be a
4687	 * valid time interval for computing a new peak rate.  Invoke
4688	 * bfq_update_rate_reset to have the following three steps
4689	 * taken:
4690	 * - close the observation interval at the last (previous)
4691	 *   request dispatch or completion
4692	 * - compute rate, if possible, for that observation interval
4693	 * - reset to zero samples, which will trigger a proper
4694	 *   re-initialization of the observation interval on next
4695	 *   dispatch
4696	 */
4697	if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
4698	   (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
4699			1UL<<(BFQ_RATE_SHIFT - 10))
4700		bfq_update_rate_reset(bfqd, NULL);
4701	bfqd->last_completion = now_ns;
 
 
 
 
 
 
 
 
 
 
 
 
 
4702
4703	/*
4704	 * If we are waiting to discover whether the request pattern
4705	 * of the task associated with the queue is actually
4706	 * isochronous, and both requisites for this condition to hold
4707	 * are now satisfied, then compute soft_rt_next_start (see the
4708	 * comments on the function bfq_bfqq_softrt_next_start()). We
4709	 * schedule this delayed check when bfqq expires, if it still
4710	 * has in-flight requests.
 
 
4711	 */
4712	if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
4713	    RB_EMPTY_ROOT(&bfqq->sort_list))
 
4714		bfqq->soft_rt_next_start =
4715			bfq_bfqq_softrt_next_start(bfqd, bfqq);
4716
4717	/*
4718	 * If this is the in-service queue, check if it needs to be expired,
4719	 * or if we want to idle in case it has no pending requests.
4720	 */
4721	if (bfqd->in_service_queue == bfqq) {
4722		if (bfqq->dispatched == 0 && bfq_bfqq_must_idle(bfqq)) {
4723			bfq_arm_slice_timer(bfqd);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4724			return;
4725		} else if (bfq_may_expire_for_budg_timeout(bfqq))
4726			bfq_bfqq_expire(bfqd, bfqq, false,
4727					BFQQE_BUDGET_TIMEOUT);
4728		else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
4729			 (bfqq->dispatched == 0 ||
4730			  !bfq_bfqq_may_idle(bfqq)))
4731			bfq_bfqq_expire(bfqd, bfqq, false,
4732					BFQQE_NO_MORE_REQUESTS);
4733	}
4734
4735	if (!bfqd->rq_in_driver)
4736		bfq_schedule_dispatch(bfqd);
4737}
4738
4739static void bfq_finish_requeue_request_body(struct bfq_queue *bfqq)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4740{
4741	bfqq->allocated--;
 
 
 
 
4742
4743	bfq_put_queue(bfqq);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4744}
4745
4746/*
4747 * Handle either a requeue or a finish for rq. The things to do are
4748 * the same in both cases: all references to rq are to be dropped. In
4749 * particular, rq is considered completed from the point of view of
4750 * the scheduler.
4751 */
4752static void bfq_finish_requeue_request(struct request *rq)
4753{
4754	struct bfq_queue *bfqq = RQ_BFQQ(rq);
4755	struct bfq_data *bfqd;
4756
4757	/*
4758	 * Requeue and finish hooks are invoked in blk-mq without
4759	 * checking whether the involved request is actually still
4760	 * referenced in the scheduler. To handle this fact, the
4761	 * following two checks make this function exit in case of
4762	 * spurious invocations, for which there is nothing to do.
4763	 *
4764	 * First, check whether rq has nothing to do with an elevator.
4765	 */
4766	if (unlikely(!(rq->rq_flags & RQF_ELVPRIV)))
4767		return;
4768
4769	/*
4770	 * rq either is not associated with any icq, or is an already
4771	 * requeued request that has not (yet) been re-inserted into
4772	 * a bfq_queue.
4773	 */
4774	if (!rq->elv.icq || !bfqq)
4775		return;
4776
4777	bfqd = bfqq->bfqd;
4778
4779	if (rq->rq_flags & RQF_STARTED)
4780		bfqg_stats_update_completion(bfqq_group(bfqq),
4781					     rq_start_time_ns(rq),
4782					     rq_io_start_time_ns(rq),
4783					     rq->cmd_flags);
4784
 
4785	if (likely(rq->rq_flags & RQF_STARTED)) {
4786		unsigned long flags;
4787
4788		spin_lock_irqsave(&bfqd->lock, flags);
4789
4790		bfq_completed_request(bfqq, bfqd);
4791		bfq_finish_requeue_request_body(bfqq);
4792
4793		spin_unlock_irqrestore(&bfqd->lock, flags);
4794	} else {
4795		/*
4796		 * Request rq may be still/already in the scheduler,
4797		 * in which case we need to remove it (this should
4798		 * never happen in case of requeue). And we cannot
4799		 * defer such a check and removal, to avoid
4800		 * inconsistencies in the time interval from the end
4801		 * of this function to the start of the deferred work.
4802		 * This situation seems to occur only in process
4803		 * context, as a consequence of a merge. In the
4804		 * current version of the code, this implies that the
4805		 * lock is held.
4806		 */
4807
4808		if (!RB_EMPTY_NODE(&rq->rb_node)) {
4809			bfq_remove_request(rq->q, rq);
4810			bfqg_stats_update_io_remove(bfqq_group(bfqq),
4811						    rq->cmd_flags);
4812		}
4813		bfq_finish_requeue_request_body(bfqq);
4814	}
 
 
 
 
4815
4816	/*
4817	 * Reset private fields. In case of a requeue, this allows
4818	 * this function to correctly do nothing if it is spuriously
4819	 * invoked again on this same request (see the check at the
4820	 * beginning of the function). Probably, a better general
4821	 * design would be to prevent blk-mq from invoking the requeue
4822	 * or finish hooks of an elevator, for a request that is not
4823	 * referred by that elevator.
4824	 *
4825	 * Resetting the following fields would break the
4826	 * request-insertion logic if rq is re-inserted into a bfq
4827	 * internal queue, without a re-preparation. Here we assume
4828	 * that re-insertions of requeued requests, without
4829	 * re-preparation, can happen only for pass_through or at_head
4830	 * requests (which are not re-inserted into bfq internal
4831	 * queues).
4832	 */
4833	rq->elv.priv[0] = NULL;
4834	rq->elv.priv[1] = NULL;
4835}
4836
 
 
 
 
 
 
 
 
 
 
4837/*
 
 
4838 * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
4839 * was the last process referring to that bfqq.
4840 */
4841static struct bfq_queue *
4842bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
4843{
4844	bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
4845
4846	if (bfqq_process_refs(bfqq) == 1) {
4847		bfqq->pid = current->pid;
4848		bfq_clear_bfqq_coop(bfqq);
4849		bfq_clear_bfqq_split_coop(bfqq);
4850		return bfqq;
4851	}
4852
4853	bic_set_bfqq(bic, NULL, 1);
4854
4855	bfq_put_cooperator(bfqq);
4856
4857	bfq_put_queue(bfqq);
4858	return NULL;
4859}
4860
4861static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
4862						   struct bfq_io_cq *bic,
4863						   struct bio *bio,
4864						   bool split, bool is_sync,
4865						   bool *new_queue)
4866{
4867	struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync);
4868
4869	if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
4870		return bfqq;
4871
4872	if (new_queue)
4873		*new_queue = true;
4874
4875	if (bfqq)
4876		bfq_put_queue(bfqq);
4877	bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
4878
4879	bic_set_bfqq(bic, bfqq, is_sync);
4880	if (split && is_sync) {
4881		if ((bic->was_in_burst_list && bfqd->large_burst) ||
4882		    bic->saved_in_large_burst)
4883			bfq_mark_bfqq_in_large_burst(bfqq);
4884		else {
4885			bfq_clear_bfqq_in_large_burst(bfqq);
4886			if (bic->was_in_burst_list)
4887				/*
4888				 * If bfqq was in the current
4889				 * burst list before being
4890				 * merged, then we have to add
4891				 * it back. And we do not need
4892				 * to increase burst_size, as
4893				 * we did not decrement
4894				 * burst_size when we removed
4895				 * bfqq from the burst list as
4896				 * a consequence of a merge
4897				 * (see comments in
4898				 * bfq_put_queue). In this
4899				 * respect, it would be rather
4900				 * costly to know whether the
4901				 * current burst list is still
4902				 * the same burst list from
4903				 * which bfqq was removed on
4904				 * the merge. To avoid this
4905				 * cost, if bfqq was in a
4906				 * burst list, then we add
4907				 * bfqq to the current burst
4908				 * list without any further
4909				 * check. This can cause
4910				 * inappropriate insertions,
4911				 * but rarely enough to not
4912				 * harm the detection of large
4913				 * bursts significantly.
4914				 */
4915				hlist_add_head(&bfqq->burst_list_node,
4916					       &bfqd->burst_list);
4917		}
4918		bfqq->split_time = jiffies;
4919	}
4920
4921	return bfqq;
4922}
4923
4924/*
4925 * Allocate bfq data structures associated with this request.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4926 */
4927static void bfq_prepare_request(struct request *rq, struct bio *bio)
4928{
4929	struct request_queue *q = rq->q;
 
4930	struct bfq_data *bfqd = q->elevator->elevator_data;
4931	struct bfq_io_cq *bic;
4932	const int is_sync = rq_is_sync(rq);
4933	struct bfq_queue *bfqq;
4934	bool new_queue = false;
4935	bool bfqq_already_existing = false, split = false;
4936
 
 
 
4937	/*
4938	 * Even if we don't have an icq attached, we should still clear
4939	 * the scheduler pointers, as they might point to previously
4940	 * allocated bic/bfqq structs.
 
 
4941	 */
4942	if (!rq->elv.icq) {
4943		rq->elv.priv[0] = rq->elv.priv[1] = NULL;
4944		return;
4945	}
4946
4947	bic = icq_to_bic(rq->elv.icq);
4948
4949	spin_lock_irq(&bfqd->lock);
4950
4951	bfq_check_ioprio_change(bic, bio);
4952
4953	bfq_bic_update_cgroup(bic, bio);
4954
4955	bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
4956					 &new_queue);
4957
4958	if (likely(!new_queue)) {
4959		/* If the queue was seeky for too long, break it apart. */
4960		if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
4961			bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq");
 
4962
4963			/* Update bic before losing reference to bfqq */
4964			if (bfq_bfqq_in_large_burst(bfqq))
4965				bic->saved_in_large_burst = true;
4966
4967			bfqq = bfq_split_bfqq(bic, bfqq);
4968			split = true;
4969
4970			if (!bfqq)
4971				bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
4972								 true, is_sync,
4973								 NULL);
4974			else
 
 
4975				bfqq_already_existing = true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4976		}
4977	}
4978
4979	bfqq->allocated++;
4980	bfqq->ref++;
 
4981	bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
4982		     rq, bfqq, bfqq->ref);
4983
4984	rq->elv.priv[0] = bic;
4985	rq->elv.priv[1] = bfqq;
4986
4987	/*
4988	 * If a bfq_queue has only one process reference, it is owned
4989	 * by only this bic: we can then set bfqq->bic = bic. in
4990	 * addition, if the queue has also just been split, we have to
4991	 * resume its state.
4992	 */
4993	if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
4994		bfqq->bic = bic;
4995		if (split) {
4996			/*
4997			 * The queue has just been split from a shared
4998			 * queue: restore the idle window and the
4999			 * possible weight raising period.
5000			 */
5001			bfq_bfqq_resume_state(bfqq, bfqd, bic,
5002					      bfqq_already_existing);
5003		}
5004	}
5005
5006	if (unlikely(bfq_bfqq_just_created(bfqq)))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5007		bfq_handle_burst(bfqd, bfqq);
5008
5009	spin_unlock_irq(&bfqd->lock);
5010}
5011
5012static void bfq_idle_slice_timer_body(struct bfq_queue *bfqq)
 
5013{
5014	struct bfq_data *bfqd = bfqq->bfqd;
5015	enum bfqq_expiration reason;
5016	unsigned long flags;
5017
5018	spin_lock_irqsave(&bfqd->lock, flags);
5019	bfq_clear_bfqq_wait_request(bfqq);
5020
 
 
 
 
 
 
 
5021	if (bfqq != bfqd->in_service_queue) {
5022		spin_unlock_irqrestore(&bfqd->lock, flags);
5023		return;
5024	}
5025
 
 
5026	if (bfq_bfqq_budget_timeout(bfqq))
5027		/*
5028		 * Also here the queue can be safely expired
5029		 * for budget timeout without wasting
5030		 * guarantees
5031		 */
5032		reason = BFQQE_BUDGET_TIMEOUT;
5033	else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
5034		/*
5035		 * The queue may not be empty upon timer expiration,
5036		 * because we may not disable the timer when the
5037		 * first request of the in-service queue arrives
5038		 * during disk idling.
5039		 */
5040		reason = BFQQE_TOO_IDLE;
5041	else
5042		goto schedule_dispatch;
5043
5044	bfq_bfqq_expire(bfqd, bfqq, true, reason);
5045
5046schedule_dispatch:
 
5047	spin_unlock_irqrestore(&bfqd->lock, flags);
5048	bfq_schedule_dispatch(bfqd);
5049}
5050
5051/*
5052 * Handler of the expiration of the timer running if the in-service queue
5053 * is idling inside its time slice.
5054 */
5055static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
5056{
5057	struct bfq_data *bfqd = container_of(timer, struct bfq_data,
5058					     idle_slice_timer);
5059	struct bfq_queue *bfqq = bfqd->in_service_queue;
5060
5061	/*
5062	 * Theoretical race here: the in-service queue can be NULL or
5063	 * different from the queue that was idling if a new request
5064	 * arrives for the current queue and there is a full dispatch
5065	 * cycle that changes the in-service queue.  This can hardly
5066	 * happen, but in the worst case we just expire a queue too
5067	 * early.
5068	 */
5069	if (bfqq)
5070		bfq_idle_slice_timer_body(bfqq);
5071
5072	return HRTIMER_NORESTART;
5073}
5074
5075static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
5076				 struct bfq_queue **bfqq_ptr)
5077{
5078	struct bfq_queue *bfqq = *bfqq_ptr;
5079
5080	bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
5081	if (bfqq) {
5082		bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
5083
5084		bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
5085			     bfqq, bfqq->ref);
5086		bfq_put_queue(bfqq);
5087		*bfqq_ptr = NULL;
5088	}
5089}
5090
5091/*
5092 * Release all the bfqg references to its async queues.  If we are
5093 * deallocating the group these queues may still contain requests, so
5094 * we reparent them to the root cgroup (i.e., the only one that will
5095 * exist for sure until all the requests on a device are gone).
5096 */
5097void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
5098{
5099	int i, j;
5100
5101	for (i = 0; i < 2; i++)
5102		for (j = 0; j < IOPRIO_BE_NR; j++)
5103			__bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]);
5104
5105	__bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq);
5106}
5107
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5108static void bfq_exit_queue(struct elevator_queue *e)
5109{
5110	struct bfq_data *bfqd = e->elevator_data;
5111	struct bfq_queue *bfqq, *n;
5112
5113	hrtimer_cancel(&bfqd->idle_slice_timer);
5114
5115	spin_lock_irq(&bfqd->lock);
5116	list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
5117		bfq_deactivate_bfqq(bfqd, bfqq, false, false);
5118	spin_unlock_irq(&bfqd->lock);
5119
5120	hrtimer_cancel(&bfqd->idle_slice_timer);
5121
5122#ifdef CONFIG_BFQ_GROUP_IOSCHED
5123	/* release oom-queue reference to root group */
5124	bfqg_and_blkg_put(bfqd->root_group);
5125
 
5126	blkcg_deactivate_policy(bfqd->queue, &blkcg_policy_bfq);
5127#else
5128	spin_lock_irq(&bfqd->lock);
5129	bfq_put_async_queues(bfqd, bfqd->root_group);
5130	kfree(bfqd->root_group);
5131	spin_unlock_irq(&bfqd->lock);
5132#endif
5133
 
 
 
 
5134	kfree(bfqd);
5135}
5136
5137static void bfq_init_root_group(struct bfq_group *root_group,
5138				struct bfq_data *bfqd)
5139{
5140	int i;
5141
5142#ifdef CONFIG_BFQ_GROUP_IOSCHED
5143	root_group->entity.parent = NULL;
5144	root_group->my_entity = NULL;
5145	root_group->bfqd = bfqd;
5146#endif
5147	root_group->rq_pos_tree = RB_ROOT;
5148	for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
5149		root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
5150	root_group->sched_data.bfq_class_idle_last_service = jiffies;
5151}
5152
5153static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
5154{
5155	struct bfq_data *bfqd;
5156	struct elevator_queue *eq;
5157
5158	eq = elevator_alloc(q, e);
5159	if (!eq)
5160		return -ENOMEM;
5161
5162	bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
5163	if (!bfqd) {
5164		kobject_put(&eq->kobj);
5165		return -ENOMEM;
5166	}
5167	eq->elevator_data = bfqd;
5168
5169	spin_lock_irq(q->queue_lock);
5170	q->elevator = eq;
5171	spin_unlock_irq(q->queue_lock);
5172
5173	/*
5174	 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
5175	 * Grab a permanent reference to it, so that the normal code flow
5176	 * will not attempt to free it.
5177	 */
5178	bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0);
5179	bfqd->oom_bfqq.ref++;
5180	bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
5181	bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
5182	bfqd->oom_bfqq.entity.new_weight =
5183		bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
5184
5185	/* oom_bfqq does not participate to bursts */
5186	bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
5187
5188	/*
5189	 * Trigger weight initialization, according to ioprio, at the
5190	 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
5191	 * class won't be changed any more.
5192	 */
5193	bfqd->oom_bfqq.entity.prio_changed = 1;
5194
5195	bfqd->queue = q;
5196
5197	INIT_LIST_HEAD(&bfqd->dispatch);
5198
5199	hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
5200		     HRTIMER_MODE_REL);
5201	bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
5202
5203	bfqd->queue_weights_tree = RB_ROOT;
5204	bfqd->group_weights_tree = RB_ROOT;
 
 
5205
5206	INIT_LIST_HEAD(&bfqd->active_list);
5207	INIT_LIST_HEAD(&bfqd->idle_list);
5208	INIT_HLIST_HEAD(&bfqd->burst_list);
5209
5210	bfqd->hw_tag = -1;
 
5211
5212	bfqd->bfq_max_budget = bfq_default_max_budget;
5213
5214	bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
5215	bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
5216	bfqd->bfq_back_max = bfq_back_max;
5217	bfqd->bfq_back_penalty = bfq_back_penalty;
5218	bfqd->bfq_slice_idle = bfq_slice_idle;
5219	bfqd->bfq_timeout = bfq_timeout;
5220
5221	bfqd->bfq_requests_within_timer = 120;
5222
5223	bfqd->bfq_large_burst_thresh = 8;
5224	bfqd->bfq_burst_interval = msecs_to_jiffies(180);
5225
5226	bfqd->low_latency = true;
5227
5228	/*
5229	 * Trade-off between responsiveness and fairness.
5230	 */
5231	bfqd->bfq_wr_coeff = 30;
5232	bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
5233	bfqd->bfq_wr_max_time = 0;
5234	bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
5235	bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
5236	bfqd->bfq_wr_max_softrt_rate = 7000; /*
5237					      * Approximate rate required
5238					      * to playback or record a
5239					      * high-definition compressed
5240					      * video.
5241					      */
5242	bfqd->wr_busy_queues = 0;
5243
5244	/*
5245	 * Begin by assuming, optimistically, that the device is a
5246	 * high-speed one, and that its peak rate is equal to 2/3 of
5247	 * the highest reference rate.
5248	 */
5249	bfqd->RT_prod = R_fast[blk_queue_nonrot(bfqd->queue)] *
5250			T_fast[blk_queue_nonrot(bfqd->queue)];
5251	bfqd->peak_rate = R_fast[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
5252	bfqd->device_speed = BFQ_BFQD_FAST;
5253
5254	spin_lock_init(&bfqd->lock);
5255
5256	/*
5257	 * The invocation of the next bfq_create_group_hierarchy
5258	 * function is the head of a chain of function calls
5259	 * (bfq_create_group_hierarchy->blkcg_activate_policy->
5260	 * blk_mq_freeze_queue) that may lead to the invocation of the
5261	 * has_work hook function. For this reason,
5262	 * bfq_create_group_hierarchy is invoked only after all
5263	 * scheduler data has been initialized, apart from the fields
5264	 * that can be initialized only after invoking
5265	 * bfq_create_group_hierarchy. This, in particular, enables
5266	 * has_work to correctly return false. Of course, to avoid
5267	 * other inconsistencies, the blk-mq stack must then refrain
5268	 * from invoking further scheduler hooks before this init
5269	 * function is finished.
5270	 */
5271	bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
5272	if (!bfqd->root_group)
5273		goto out_free;
5274	bfq_init_root_group(bfqd->root_group, bfqd);
5275	bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
5276
 
 
 
 
5277	wbt_disable_default(q);
 
 
5278	return 0;
5279
5280out_free:
5281	kfree(bfqd);
5282	kobject_put(&eq->kobj);
5283	return -ENOMEM;
5284}
5285
5286static void bfq_slab_kill(void)
5287{
5288	kmem_cache_destroy(bfq_pool);
5289}
5290
5291static int __init bfq_slab_setup(void)
5292{
5293	bfq_pool = KMEM_CACHE(bfq_queue, 0);
5294	if (!bfq_pool)
5295		return -ENOMEM;
5296	return 0;
5297}
5298
5299static ssize_t bfq_var_show(unsigned int var, char *page)
5300{
5301	return sprintf(page, "%u\n", var);
5302}
5303
5304static int bfq_var_store(unsigned long *var, const char *page)
5305{
5306	unsigned long new_val;
5307	int ret = kstrtoul(page, 10, &new_val);
5308
5309	if (ret)
5310		return ret;
5311	*var = new_val;
5312	return 0;
5313}
5314
5315#define SHOW_FUNCTION(__FUNC, __VAR, __CONV)				\
5316static ssize_t __FUNC(struct elevator_queue *e, char *page)		\
5317{									\
5318	struct bfq_data *bfqd = e->elevator_data;			\
5319	u64 __data = __VAR;						\
5320	if (__CONV == 1)						\
5321		__data = jiffies_to_msecs(__data);			\
5322	else if (__CONV == 2)						\
5323		__data = div_u64(__data, NSEC_PER_MSEC);		\
5324	return bfq_var_show(__data, (page));				\
5325}
5326SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
5327SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
5328SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
5329SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
5330SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
5331SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
5332SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
5333SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
5334SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
5335#undef SHOW_FUNCTION
5336
5337#define USEC_SHOW_FUNCTION(__FUNC, __VAR)				\
5338static ssize_t __FUNC(struct elevator_queue *e, char *page)		\
5339{									\
5340	struct bfq_data *bfqd = e->elevator_data;			\
5341	u64 __data = __VAR;						\
5342	__data = div_u64(__data, NSEC_PER_USEC);			\
5343	return bfq_var_show(__data, (page));				\
5344}
5345USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
5346#undef USEC_SHOW_FUNCTION
5347
5348#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)			\
5349static ssize_t								\
5350__FUNC(struct elevator_queue *e, const char *page, size_t count)	\
5351{									\
5352	struct bfq_data *bfqd = e->elevator_data;			\
5353	unsigned long __data, __min = (MIN), __max = (MAX);		\
5354	int ret;							\
5355									\
5356	ret = bfq_var_store(&__data, (page));				\
5357	if (ret)							\
5358		return ret;						\
5359	if (__data < __min)						\
5360		__data = __min;						\
5361	else if (__data > __max)					\
5362		__data = __max;						\
5363	if (__CONV == 1)						\
5364		*(__PTR) = msecs_to_jiffies(__data);			\
5365	else if (__CONV == 2)						\
5366		*(__PTR) = (u64)__data * NSEC_PER_MSEC;			\
5367	else								\
5368		*(__PTR) = __data;					\
5369	return count;							\
5370}
5371STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
5372		INT_MAX, 2);
5373STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
5374		INT_MAX, 2);
5375STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
5376STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
5377		INT_MAX, 0);
5378STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
5379#undef STORE_FUNCTION
5380
5381#define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX)			\
5382static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
5383{									\
5384	struct bfq_data *bfqd = e->elevator_data;			\
5385	unsigned long __data, __min = (MIN), __max = (MAX);		\
5386	int ret;							\
5387									\
5388	ret = bfq_var_store(&__data, (page));				\
5389	if (ret)							\
5390		return ret;						\
5391	if (__data < __min)						\
5392		__data = __min;						\
5393	else if (__data > __max)					\
5394		__data = __max;						\
5395	*(__PTR) = (u64)__data * NSEC_PER_USEC;				\
5396	return count;							\
5397}
5398USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
5399		    UINT_MAX);
5400#undef USEC_STORE_FUNCTION
5401
5402static ssize_t bfq_max_budget_store(struct elevator_queue *e,
5403				    const char *page, size_t count)
5404{
5405	struct bfq_data *bfqd = e->elevator_data;
5406	unsigned long __data;
5407	int ret;
5408
5409	ret = bfq_var_store(&__data, (page));
5410	if (ret)
5411		return ret;
5412
5413	if (__data == 0)
5414		bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5415	else {
5416		if (__data > INT_MAX)
5417			__data = INT_MAX;
5418		bfqd->bfq_max_budget = __data;
5419	}
5420
5421	bfqd->bfq_user_max_budget = __data;
5422
5423	return count;
5424}
5425
5426/*
5427 * Leaving this name to preserve name compatibility with cfq
5428 * parameters, but this timeout is used for both sync and async.
5429 */
5430static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
5431				      const char *page, size_t count)
5432{
5433	struct bfq_data *bfqd = e->elevator_data;
5434	unsigned long __data;
5435	int ret;
5436
5437	ret = bfq_var_store(&__data, (page));
5438	if (ret)
5439		return ret;
5440
5441	if (__data < 1)
5442		__data = 1;
5443	else if (__data > INT_MAX)
5444		__data = INT_MAX;
5445
5446	bfqd->bfq_timeout = msecs_to_jiffies(__data);
5447	if (bfqd->bfq_user_max_budget == 0)
5448		bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
5449
5450	return count;
5451}
5452
5453static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
5454				     const char *page, size_t count)
5455{
5456	struct bfq_data *bfqd = e->elevator_data;
5457	unsigned long __data;
5458	int ret;
5459
5460	ret = bfq_var_store(&__data, (page));
5461	if (ret)
5462		return ret;
5463
5464	if (__data > 1)
5465		__data = 1;
5466	if (!bfqd->strict_guarantees && __data == 1
5467	    && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
5468		bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
5469
5470	bfqd->strict_guarantees = __data;
5471
5472	return count;
5473}
5474
5475static ssize_t bfq_low_latency_store(struct elevator_queue *e,
5476				     const char *page, size_t count)
5477{
5478	struct bfq_data *bfqd = e->elevator_data;
5479	unsigned long __data;
5480	int ret;
5481
5482	ret = bfq_var_store(&__data, (page));
5483	if (ret)
5484		return ret;
5485
5486	if (__data > 1)
5487		__data = 1;
5488	if (__data == 0 && bfqd->low_latency != 0)
5489		bfq_end_wr(bfqd);
5490	bfqd->low_latency = __data;
5491
5492	return count;
5493}
5494
5495#define BFQ_ATTR(name) \
5496	__ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
5497
5498static struct elv_fs_entry bfq_attrs[] = {
5499	BFQ_ATTR(fifo_expire_sync),
5500	BFQ_ATTR(fifo_expire_async),
5501	BFQ_ATTR(back_seek_max),
5502	BFQ_ATTR(back_seek_penalty),
5503	BFQ_ATTR(slice_idle),
5504	BFQ_ATTR(slice_idle_us),
5505	BFQ_ATTR(max_budget),
5506	BFQ_ATTR(timeout_sync),
5507	BFQ_ATTR(strict_guarantees),
5508	BFQ_ATTR(low_latency),
5509	__ATTR_NULL
5510};
5511
5512static struct elevator_type iosched_bfq_mq = {
5513	.ops.mq = {
5514		.limit_depth		= bfq_limit_depth,
5515		.prepare_request	= bfq_prepare_request,
5516		.requeue_request        = bfq_finish_requeue_request,
5517		.finish_request		= bfq_finish_requeue_request,
5518		.exit_icq		= bfq_exit_icq,
5519		.insert_requests	= bfq_insert_requests,
5520		.dispatch_request	= bfq_dispatch_request,
5521		.next_request		= elv_rb_latter_request,
5522		.former_request		= elv_rb_former_request,
5523		.allow_merge		= bfq_allow_bio_merge,
5524		.bio_merge		= bfq_bio_merge,
5525		.request_merge		= bfq_request_merge,
5526		.requests_merged	= bfq_requests_merged,
5527		.request_merged		= bfq_request_merged,
5528		.has_work		= bfq_has_work,
 
 
5529		.init_sched		= bfq_init_queue,
5530		.exit_sched		= bfq_exit_queue,
5531	},
5532
5533	.uses_mq =		true,
5534	.icq_size =		sizeof(struct bfq_io_cq),
5535	.icq_align =		__alignof__(struct bfq_io_cq),
5536	.elevator_attrs =	bfq_attrs,
5537	.elevator_name =	"bfq",
5538	.elevator_owner =	THIS_MODULE,
5539};
5540MODULE_ALIAS("bfq-iosched");
5541
5542static int __init bfq_init(void)
5543{
5544	int ret;
5545
5546#ifdef CONFIG_BFQ_GROUP_IOSCHED
5547	ret = blkcg_policy_register(&blkcg_policy_bfq);
5548	if (ret)
5549		return ret;
5550#endif
5551
5552	ret = -ENOMEM;
5553	if (bfq_slab_setup())
5554		goto err_pol_unreg;
5555
5556	/*
5557	 * Times to load large popular applications for the typical
5558	 * systems installed on the reference devices (see the
5559	 * comments before the definitions of the next two
5560	 * arrays). Actually, we use slightly slower values, as the
5561	 * estimated peak rate tends to be smaller than the actual
5562	 * peak rate.  The reason for this last fact is that estimates
5563	 * are computed over much shorter time intervals than the long
5564	 * intervals typically used for benchmarking. Why? First, to
5565	 * adapt more quickly to variations. Second, because an I/O
5566	 * scheduler cannot rely on a peak-rate-evaluation workload to
5567	 * be run for a long time.
5568	 */
5569	T_slow[0] = msecs_to_jiffies(3500); /* actually 4 sec */
5570	T_slow[1] = msecs_to_jiffies(6000); /* actually 6.5 sec */
5571	T_fast[0] = msecs_to_jiffies(7000); /* actually 8 sec */
5572	T_fast[1] = msecs_to_jiffies(2500); /* actually 3 sec */
5573
5574	/*
5575	 * Thresholds that determine the switch between speed classes
5576	 * (see the comments before the definition of the array
5577	 * device_speed_thresh). These thresholds are biased towards
5578	 * transitions to the fast class. This is safer than the
5579	 * opposite bias. In fact, a wrong transition to the slow
5580	 * class results in short weight-raising periods, because the
5581	 * speed of the device then tends to be higher that the
5582	 * reference peak rate. On the opposite end, a wrong
5583	 * transition to the fast class tends to increase
5584	 * weight-raising periods, because of the opposite reason.
5585	 */
5586	device_speed_thresh[0] = (4 * R_slow[0]) / 3;
5587	device_speed_thresh[1] = (4 * R_slow[1]) / 3;
5588
5589	ret = elv_register(&iosched_bfq_mq);
5590	if (ret)
5591		goto slab_kill;
5592
5593	return 0;
5594
5595slab_kill:
5596	bfq_slab_kill();
5597err_pol_unreg:
5598#ifdef CONFIG_BFQ_GROUP_IOSCHED
5599	blkcg_policy_unregister(&blkcg_policy_bfq);
5600#endif
5601	return ret;
5602}
5603
5604static void __exit bfq_exit(void)
5605{
5606	elv_unregister(&iosched_bfq_mq);
5607#ifdef CONFIG_BFQ_GROUP_IOSCHED
5608	blkcg_policy_unregister(&blkcg_policy_bfq);
5609#endif
5610	bfq_slab_kill();
5611}
5612
5613module_init(bfq_init);
5614module_exit(bfq_exit);
5615
5616MODULE_AUTHOR("Paolo Valente");
5617MODULE_LICENSE("GPL");
5618MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");