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