cfq-iosched.c revision e56da7e287967667474a58c4f60c286279e3f487
1/* 2 * CFQ, or complete fairness queueing, disk scheduler. 3 * 4 * Based on ideas from a previously unfinished io 5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli. 6 * 7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk> 8 */ 9#include <linux/module.h> 10#include <linux/slab.h> 11#include <linux/blkdev.h> 12#include <linux/elevator.h> 13#include <linux/jiffies.h> 14#include <linux/rbtree.h> 15#include <linux/ioprio.h> 16#include <linux/blktrace_api.h> 17#include "blk.h" 18#include "cfq.h" 19 20/* 21 * tunables 22 */ 23/* max queue in one round of service */ 24static const int cfq_quantum = 8; 25static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; 26/* maximum backwards seek, in KiB */ 27static const int cfq_back_max = 16 * 1024; 28/* penalty of a backwards seek */ 29static const int cfq_back_penalty = 2; 30static const int cfq_slice_sync = HZ / 10; 31static int cfq_slice_async = HZ / 25; 32static const int cfq_slice_async_rq = 2; 33static int cfq_slice_idle = HZ / 125; 34static int cfq_group_idle = HZ / 125; 35static const int cfq_target_latency = HZ * 3/10; /* 300 ms */ 36static const int cfq_hist_divisor = 4; 37 38/* 39 * offset from end of service tree 40 */ 41#define CFQ_IDLE_DELAY (HZ / 5) 42 43/* 44 * below this threshold, we consider thinktime immediate 45 */ 46#define CFQ_MIN_TT (2) 47 48#define CFQ_SLICE_SCALE (5) 49#define CFQ_HW_QUEUE_MIN (5) 50#define CFQ_SERVICE_SHIFT 12 51 52#define CFQQ_SEEK_THR (sector_t)(8 * 100) 53#define CFQQ_CLOSE_THR (sector_t)(8 * 1024) 54#define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32) 55#define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8) 56 57#define RQ_CIC(rq) icq_to_cic((rq)->elv.icq) 58#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0]) 59#define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1]) 60 61static struct kmem_cache *cfq_pool; 62 63#define CFQ_PRIO_LISTS IOPRIO_BE_NR 64#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) 65#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) 66 67#define sample_valid(samples) ((samples) > 80) 68#define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node) 69 70struct cfq_ttime { 71 unsigned long last_end_request; 72 73 unsigned long ttime_total; 74 unsigned long ttime_samples; 75 unsigned long ttime_mean; 76}; 77 78/* 79 * Most of our rbtree usage is for sorting with min extraction, so 80 * if we cache the leftmost node we don't have to walk down the tree 81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should 82 * move this into the elevator for the rq sorting as well. 83 */ 84struct cfq_rb_root { 85 struct rb_root rb; 86 struct rb_node *left; 87 unsigned count; 88 unsigned total_weight; 89 u64 min_vdisktime; 90 struct cfq_ttime ttime; 91}; 92#define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \ 93 .ttime = {.last_end_request = jiffies,},} 94 95/* 96 * Per process-grouping structure 97 */ 98struct cfq_queue { 99 /* reference count */ 100 int ref; 101 /* various state flags, see below */ 102 unsigned int flags; 103 /* parent cfq_data */ 104 struct cfq_data *cfqd; 105 /* service_tree member */ 106 struct rb_node rb_node; 107 /* service_tree key */ 108 unsigned long rb_key; 109 /* prio tree member */ 110 struct rb_node p_node; 111 /* prio tree root we belong to, if any */ 112 struct rb_root *p_root; 113 /* sorted list of pending requests */ 114 struct rb_root sort_list; 115 /* if fifo isn't expired, next request to serve */ 116 struct request *next_rq; 117 /* requests queued in sort_list */ 118 int queued[2]; 119 /* currently allocated requests */ 120 int allocated[2]; 121 /* fifo list of requests in sort_list */ 122 struct list_head fifo; 123 124 /* time when queue got scheduled in to dispatch first request. */ 125 unsigned long dispatch_start; 126 unsigned int allocated_slice; 127 unsigned int slice_dispatch; 128 /* time when first request from queue completed and slice started. */ 129 unsigned long slice_start; 130 unsigned long slice_end; 131 long slice_resid; 132 133 /* pending priority requests */ 134 int prio_pending; 135 /* number of requests that are on the dispatch list or inside driver */ 136 int dispatched; 137 138 /* io prio of this group */ 139 unsigned short ioprio, org_ioprio; 140 unsigned short ioprio_class; 141 142 pid_t pid; 143 144 u32 seek_history; 145 sector_t last_request_pos; 146 147 struct cfq_rb_root *service_tree; 148 struct cfq_queue *new_cfqq; 149 struct cfq_group *cfqg; 150 /* Number of sectors dispatched from queue in single dispatch round */ 151 unsigned long nr_sectors; 152}; 153 154/* 155 * First index in the service_trees. 156 * IDLE is handled separately, so it has negative index 157 */ 158enum wl_prio_t { 159 BE_WORKLOAD = 0, 160 RT_WORKLOAD = 1, 161 IDLE_WORKLOAD = 2, 162 CFQ_PRIO_NR, 163}; 164 165/* 166 * Second index in the service_trees. 167 */ 168enum wl_type_t { 169 ASYNC_WORKLOAD = 0, 170 SYNC_NOIDLE_WORKLOAD = 1, 171 SYNC_WORKLOAD = 2 172}; 173 174/* This is per cgroup per device grouping structure */ 175struct cfq_group { 176 /* group service_tree member */ 177 struct rb_node rb_node; 178 179 /* group service_tree key */ 180 u64 vdisktime; 181 unsigned int weight; 182 unsigned int new_weight; 183 bool needs_update; 184 185 /* number of cfqq currently on this group */ 186 int nr_cfqq; 187 188 /* 189 * Per group busy queues average. Useful for workload slice calc. We 190 * create the array for each prio class but at run time it is used 191 * only for RT and BE class and slot for IDLE class remains unused. 192 * This is primarily done to avoid confusion and a gcc warning. 193 */ 194 unsigned int busy_queues_avg[CFQ_PRIO_NR]; 195 /* 196 * rr lists of queues with requests. We maintain service trees for 197 * RT and BE classes. These trees are subdivided in subclasses 198 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE 199 * class there is no subclassification and all the cfq queues go on 200 * a single tree service_tree_idle. 201 * Counts are embedded in the cfq_rb_root 202 */ 203 struct cfq_rb_root service_trees[2][3]; 204 struct cfq_rb_root service_tree_idle; 205 206 unsigned long saved_workload_slice; 207 enum wl_type_t saved_workload; 208 enum wl_prio_t saved_serving_prio; 209 struct blkio_group blkg; 210#ifdef CONFIG_CFQ_GROUP_IOSCHED 211 struct hlist_node cfqd_node; 212 int ref; 213#endif 214 /* number of requests that are on the dispatch list or inside driver */ 215 int dispatched; 216 struct cfq_ttime ttime; 217}; 218 219struct cfq_io_cq { 220 struct io_cq icq; /* must be the first member */ 221 struct cfq_queue *cfqq[2]; 222 struct cfq_ttime ttime; 223}; 224 225/* 226 * Per block device queue structure 227 */ 228struct cfq_data { 229 struct request_queue *queue; 230 /* Root service tree for cfq_groups */ 231 struct cfq_rb_root grp_service_tree; 232 struct cfq_group *root_group; 233 234 /* 235 * The priority currently being served 236 */ 237 enum wl_prio_t serving_prio; 238 enum wl_type_t serving_type; 239 unsigned long workload_expires; 240 struct cfq_group *serving_group; 241 242 /* 243 * Each priority tree is sorted by next_request position. These 244 * trees are used when determining if two or more queues are 245 * interleaving requests (see cfq_close_cooperator). 246 */ 247 struct rb_root prio_trees[CFQ_PRIO_LISTS]; 248 249 unsigned int busy_queues; 250 unsigned int busy_sync_queues; 251 252 int rq_in_driver; 253 int rq_in_flight[2]; 254 255 /* 256 * queue-depth detection 257 */ 258 int rq_queued; 259 int hw_tag; 260 /* 261 * hw_tag can be 262 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection) 263 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth) 264 * 0 => no NCQ 265 */ 266 int hw_tag_est_depth; 267 unsigned int hw_tag_samples; 268 269 /* 270 * idle window management 271 */ 272 struct timer_list idle_slice_timer; 273 struct work_struct unplug_work; 274 275 struct cfq_queue *active_queue; 276 struct cfq_io_cq *active_cic; 277 278 /* 279 * async queue for each priority case 280 */ 281 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; 282 struct cfq_queue *async_idle_cfqq; 283 284 sector_t last_position; 285 286 /* 287 * tunables, see top of file 288 */ 289 unsigned int cfq_quantum; 290 unsigned int cfq_fifo_expire[2]; 291 unsigned int cfq_back_penalty; 292 unsigned int cfq_back_max; 293 unsigned int cfq_slice[2]; 294 unsigned int cfq_slice_async_rq; 295 unsigned int cfq_slice_idle; 296 unsigned int cfq_group_idle; 297 unsigned int cfq_latency; 298 299 /* 300 * Fallback dummy cfqq for extreme OOM conditions 301 */ 302 struct cfq_queue oom_cfqq; 303 304 unsigned long last_delayed_sync; 305 306 /* List of cfq groups being managed on this device*/ 307 struct hlist_head cfqg_list; 308 309 /* Number of groups which are on blkcg->blkg_list */ 310 unsigned int nr_blkcg_linked_grps; 311}; 312 313static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd); 314 315static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg, 316 enum wl_prio_t prio, 317 enum wl_type_t type) 318{ 319 if (!cfqg) 320 return NULL; 321 322 if (prio == IDLE_WORKLOAD) 323 return &cfqg->service_tree_idle; 324 325 return &cfqg->service_trees[prio][type]; 326} 327 328enum cfqq_state_flags { 329 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ 330 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ 331 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ 332 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ 333 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ 334 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ 335 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ 336 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ 337 CFQ_CFQQ_FLAG_sync, /* synchronous queue */ 338 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */ 339 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */ 340 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */ 341 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */ 342}; 343 344#define CFQ_CFQQ_FNS(name) \ 345static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ 346{ \ 347 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ 348} \ 349static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ 350{ \ 351 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ 352} \ 353static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ 354{ \ 355 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ 356} 357 358CFQ_CFQQ_FNS(on_rr); 359CFQ_CFQQ_FNS(wait_request); 360CFQ_CFQQ_FNS(must_dispatch); 361CFQ_CFQQ_FNS(must_alloc_slice); 362CFQ_CFQQ_FNS(fifo_expire); 363CFQ_CFQQ_FNS(idle_window); 364CFQ_CFQQ_FNS(prio_changed); 365CFQ_CFQQ_FNS(slice_new); 366CFQ_CFQQ_FNS(sync); 367CFQ_CFQQ_FNS(coop); 368CFQ_CFQQ_FNS(split_coop); 369CFQ_CFQQ_FNS(deep); 370CFQ_CFQQ_FNS(wait_busy); 371#undef CFQ_CFQQ_FNS 372 373#ifdef CONFIG_CFQ_GROUP_IOSCHED 374#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ 375 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \ 376 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \ 377 blkg_path(&(cfqq)->cfqg->blkg), ##args) 378 379#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \ 380 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \ 381 blkg_path(&(cfqg)->blkg), ##args) \ 382 383#else 384#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ 385 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) 386#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0) 387#endif 388#define cfq_log(cfqd, fmt, args...) \ 389 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) 390 391/* Traverses through cfq group service trees */ 392#define for_each_cfqg_st(cfqg, i, j, st) \ 393 for (i = 0; i <= IDLE_WORKLOAD; i++) \ 394 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\ 395 : &cfqg->service_tree_idle; \ 396 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \ 397 (i == IDLE_WORKLOAD && j == 0); \ 398 j++, st = i < IDLE_WORKLOAD ? \ 399 &cfqg->service_trees[i][j]: NULL) \ 400 401static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd, 402 struct cfq_ttime *ttime, bool group_idle) 403{ 404 unsigned long slice; 405 if (!sample_valid(ttime->ttime_samples)) 406 return false; 407 if (group_idle) 408 slice = cfqd->cfq_group_idle; 409 else 410 slice = cfqd->cfq_slice_idle; 411 return ttime->ttime_mean > slice; 412} 413 414static inline bool iops_mode(struct cfq_data *cfqd) 415{ 416 /* 417 * If we are not idling on queues and it is a NCQ drive, parallel 418 * execution of requests is on and measuring time is not possible 419 * in most of the cases until and unless we drive shallower queue 420 * depths and that becomes a performance bottleneck. In such cases 421 * switch to start providing fairness in terms of number of IOs. 422 */ 423 if (!cfqd->cfq_slice_idle && cfqd->hw_tag) 424 return true; 425 else 426 return false; 427} 428 429static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq) 430{ 431 if (cfq_class_idle(cfqq)) 432 return IDLE_WORKLOAD; 433 if (cfq_class_rt(cfqq)) 434 return RT_WORKLOAD; 435 return BE_WORKLOAD; 436} 437 438 439static enum wl_type_t cfqq_type(struct cfq_queue *cfqq) 440{ 441 if (!cfq_cfqq_sync(cfqq)) 442 return ASYNC_WORKLOAD; 443 if (!cfq_cfqq_idle_window(cfqq)) 444 return SYNC_NOIDLE_WORKLOAD; 445 return SYNC_WORKLOAD; 446} 447 448static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl, 449 struct cfq_data *cfqd, 450 struct cfq_group *cfqg) 451{ 452 if (wl == IDLE_WORKLOAD) 453 return cfqg->service_tree_idle.count; 454 455 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count 456 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count 457 + cfqg->service_trees[wl][SYNC_WORKLOAD].count; 458} 459 460static inline int cfqg_busy_async_queues(struct cfq_data *cfqd, 461 struct cfq_group *cfqg) 462{ 463 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count 464 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count; 465} 466 467static void cfq_dispatch_insert(struct request_queue *, struct request *); 468static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool, 469 struct io_context *, gfp_t); 470 471static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq) 472{ 473 /* cic->icq is the first member, %NULL will convert to %NULL */ 474 return container_of(icq, struct cfq_io_cq, icq); 475} 476 477static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd, 478 struct io_context *ioc) 479{ 480 if (ioc) 481 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue)); 482 return NULL; 483} 484 485static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync) 486{ 487 return cic->cfqq[is_sync]; 488} 489 490static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq, 491 bool is_sync) 492{ 493 cic->cfqq[is_sync] = cfqq; 494} 495 496static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic) 497{ 498 return cic->icq.q->elevator->elevator_data; 499} 500 501/* 502 * We regard a request as SYNC, if it's either a read or has the SYNC bit 503 * set (in which case it could also be direct WRITE). 504 */ 505static inline bool cfq_bio_sync(struct bio *bio) 506{ 507 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC); 508} 509 510/* 511 * scheduler run of queue, if there are requests pending and no one in the 512 * driver that will restart queueing 513 */ 514static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) 515{ 516 if (cfqd->busy_queues) { 517 cfq_log(cfqd, "schedule dispatch"); 518 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); 519 } 520} 521 522/* 523 * Scale schedule slice based on io priority. Use the sync time slice only 524 * if a queue is marked sync and has sync io queued. A sync queue with async 525 * io only, should not get full sync slice length. 526 */ 527static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync, 528 unsigned short prio) 529{ 530 const int base_slice = cfqd->cfq_slice[sync]; 531 532 WARN_ON(prio >= IOPRIO_BE_NR); 533 534 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); 535} 536 537static inline int 538cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) 539{ 540 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); 541} 542 543static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg) 544{ 545 u64 d = delta << CFQ_SERVICE_SHIFT; 546 547 d = d * BLKIO_WEIGHT_DEFAULT; 548 do_div(d, cfqg->weight); 549 return d; 550} 551 552static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime) 553{ 554 s64 delta = (s64)(vdisktime - min_vdisktime); 555 if (delta > 0) 556 min_vdisktime = vdisktime; 557 558 return min_vdisktime; 559} 560 561static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime) 562{ 563 s64 delta = (s64)(vdisktime - min_vdisktime); 564 if (delta < 0) 565 min_vdisktime = vdisktime; 566 567 return min_vdisktime; 568} 569 570static void update_min_vdisktime(struct cfq_rb_root *st) 571{ 572 struct cfq_group *cfqg; 573 574 if (st->left) { 575 cfqg = rb_entry_cfqg(st->left); 576 st->min_vdisktime = max_vdisktime(st->min_vdisktime, 577 cfqg->vdisktime); 578 } 579} 580 581/* 582 * get averaged number of queues of RT/BE priority. 583 * average is updated, with a formula that gives more weight to higher numbers, 584 * to quickly follows sudden increases and decrease slowly 585 */ 586 587static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd, 588 struct cfq_group *cfqg, bool rt) 589{ 590 unsigned min_q, max_q; 591 unsigned mult = cfq_hist_divisor - 1; 592 unsigned round = cfq_hist_divisor / 2; 593 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg); 594 595 min_q = min(cfqg->busy_queues_avg[rt], busy); 596 max_q = max(cfqg->busy_queues_avg[rt], busy); 597 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) / 598 cfq_hist_divisor; 599 return cfqg->busy_queues_avg[rt]; 600} 601 602static inline unsigned 603cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg) 604{ 605 struct cfq_rb_root *st = &cfqd->grp_service_tree; 606 607 return cfq_target_latency * cfqg->weight / st->total_weight; 608} 609 610static inline unsigned 611cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) 612{ 613 unsigned slice = cfq_prio_to_slice(cfqd, cfqq); 614 if (cfqd->cfq_latency) { 615 /* 616 * interested queues (we consider only the ones with the same 617 * priority class in the cfq group) 618 */ 619 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg, 620 cfq_class_rt(cfqq)); 621 unsigned sync_slice = cfqd->cfq_slice[1]; 622 unsigned expect_latency = sync_slice * iq; 623 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg); 624 625 if (expect_latency > group_slice) { 626 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle; 627 /* scale low_slice according to IO priority 628 * and sync vs async */ 629 unsigned low_slice = 630 min(slice, base_low_slice * slice / sync_slice); 631 /* the adapted slice value is scaled to fit all iqs 632 * into the target latency */ 633 slice = max(slice * group_slice / expect_latency, 634 low_slice); 635 } 636 } 637 return slice; 638} 639 640static inline void 641cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) 642{ 643 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq); 644 645 cfqq->slice_start = jiffies; 646 cfqq->slice_end = jiffies + slice; 647 cfqq->allocated_slice = slice; 648 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); 649} 650 651/* 652 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end 653 * isn't valid until the first request from the dispatch is activated 654 * and the slice time set. 655 */ 656static inline bool cfq_slice_used(struct cfq_queue *cfqq) 657{ 658 if (cfq_cfqq_slice_new(cfqq)) 659 return false; 660 if (time_before(jiffies, cfqq->slice_end)) 661 return false; 662 663 return true; 664} 665 666/* 667 * Lifted from AS - choose which of rq1 and rq2 that is best served now. 668 * We choose the request that is closest to the head right now. Distance 669 * behind the head is penalized and only allowed to a certain extent. 670 */ 671static struct request * 672cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last) 673{ 674 sector_t s1, s2, d1 = 0, d2 = 0; 675 unsigned long back_max; 676#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ 677#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ 678 unsigned wrap = 0; /* bit mask: requests behind the disk head? */ 679 680 if (rq1 == NULL || rq1 == rq2) 681 return rq2; 682 if (rq2 == NULL) 683 return rq1; 684 685 if (rq_is_sync(rq1) != rq_is_sync(rq2)) 686 return rq_is_sync(rq1) ? rq1 : rq2; 687 688 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO) 689 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2; 690 691 s1 = blk_rq_pos(rq1); 692 s2 = blk_rq_pos(rq2); 693 694 /* 695 * by definition, 1KiB is 2 sectors 696 */ 697 back_max = cfqd->cfq_back_max * 2; 698 699 /* 700 * Strict one way elevator _except_ in the case where we allow 701 * short backward seeks which are biased as twice the cost of a 702 * similar forward seek. 703 */ 704 if (s1 >= last) 705 d1 = s1 - last; 706 else if (s1 + back_max >= last) 707 d1 = (last - s1) * cfqd->cfq_back_penalty; 708 else 709 wrap |= CFQ_RQ1_WRAP; 710 711 if (s2 >= last) 712 d2 = s2 - last; 713 else if (s2 + back_max >= last) 714 d2 = (last - s2) * cfqd->cfq_back_penalty; 715 else 716 wrap |= CFQ_RQ2_WRAP; 717 718 /* Found required data */ 719 720 /* 721 * By doing switch() on the bit mask "wrap" we avoid having to 722 * check two variables for all permutations: --> faster! 723 */ 724 switch (wrap) { 725 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ 726 if (d1 < d2) 727 return rq1; 728 else if (d2 < d1) 729 return rq2; 730 else { 731 if (s1 >= s2) 732 return rq1; 733 else 734 return rq2; 735 } 736 737 case CFQ_RQ2_WRAP: 738 return rq1; 739 case CFQ_RQ1_WRAP: 740 return rq2; 741 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ 742 default: 743 /* 744 * Since both rqs are wrapped, 745 * start with the one that's further behind head 746 * (--> only *one* back seek required), 747 * since back seek takes more time than forward. 748 */ 749 if (s1 <= s2) 750 return rq1; 751 else 752 return rq2; 753 } 754} 755 756/* 757 * The below is leftmost cache rbtree addon 758 */ 759static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) 760{ 761 /* Service tree is empty */ 762 if (!root->count) 763 return NULL; 764 765 if (!root->left) 766 root->left = rb_first(&root->rb); 767 768 if (root->left) 769 return rb_entry(root->left, struct cfq_queue, rb_node); 770 771 return NULL; 772} 773 774static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root) 775{ 776 if (!root->left) 777 root->left = rb_first(&root->rb); 778 779 if (root->left) 780 return rb_entry_cfqg(root->left); 781 782 return NULL; 783} 784 785static void rb_erase_init(struct rb_node *n, struct rb_root *root) 786{ 787 rb_erase(n, root); 788 RB_CLEAR_NODE(n); 789} 790 791static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) 792{ 793 if (root->left == n) 794 root->left = NULL; 795 rb_erase_init(n, &root->rb); 796 --root->count; 797} 798 799/* 800 * would be nice to take fifo expire time into account as well 801 */ 802static struct request * 803cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, 804 struct request *last) 805{ 806 struct rb_node *rbnext = rb_next(&last->rb_node); 807 struct rb_node *rbprev = rb_prev(&last->rb_node); 808 struct request *next = NULL, *prev = NULL; 809 810 BUG_ON(RB_EMPTY_NODE(&last->rb_node)); 811 812 if (rbprev) 813 prev = rb_entry_rq(rbprev); 814 815 if (rbnext) 816 next = rb_entry_rq(rbnext); 817 else { 818 rbnext = rb_first(&cfqq->sort_list); 819 if (rbnext && rbnext != &last->rb_node) 820 next = rb_entry_rq(rbnext); 821 } 822 823 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last)); 824} 825 826static unsigned long cfq_slice_offset(struct cfq_data *cfqd, 827 struct cfq_queue *cfqq) 828{ 829 /* 830 * just an approximation, should be ok. 831 */ 832 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) - 833 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); 834} 835 836static inline s64 837cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg) 838{ 839 return cfqg->vdisktime - st->min_vdisktime; 840} 841 842static void 843__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) 844{ 845 struct rb_node **node = &st->rb.rb_node; 846 struct rb_node *parent = NULL; 847 struct cfq_group *__cfqg; 848 s64 key = cfqg_key(st, cfqg); 849 int left = 1; 850 851 while (*node != NULL) { 852 parent = *node; 853 __cfqg = rb_entry_cfqg(parent); 854 855 if (key < cfqg_key(st, __cfqg)) 856 node = &parent->rb_left; 857 else { 858 node = &parent->rb_right; 859 left = 0; 860 } 861 } 862 863 if (left) 864 st->left = &cfqg->rb_node; 865 866 rb_link_node(&cfqg->rb_node, parent, node); 867 rb_insert_color(&cfqg->rb_node, &st->rb); 868} 869 870static void 871cfq_update_group_weight(struct cfq_group *cfqg) 872{ 873 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node)); 874 if (cfqg->needs_update) { 875 cfqg->weight = cfqg->new_weight; 876 cfqg->needs_update = false; 877 } 878} 879 880static void 881cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) 882{ 883 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node)); 884 885 cfq_update_group_weight(cfqg); 886 __cfq_group_service_tree_add(st, cfqg); 887 st->total_weight += cfqg->weight; 888} 889 890static void 891cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg) 892{ 893 struct cfq_rb_root *st = &cfqd->grp_service_tree; 894 struct cfq_group *__cfqg; 895 struct rb_node *n; 896 897 cfqg->nr_cfqq++; 898 if (!RB_EMPTY_NODE(&cfqg->rb_node)) 899 return; 900 901 /* 902 * Currently put the group at the end. Later implement something 903 * so that groups get lesser vtime based on their weights, so that 904 * if group does not loose all if it was not continuously backlogged. 905 */ 906 n = rb_last(&st->rb); 907 if (n) { 908 __cfqg = rb_entry_cfqg(n); 909 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY; 910 } else 911 cfqg->vdisktime = st->min_vdisktime; 912 cfq_group_service_tree_add(st, cfqg); 913} 914 915static void 916cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg) 917{ 918 st->total_weight -= cfqg->weight; 919 if (!RB_EMPTY_NODE(&cfqg->rb_node)) 920 cfq_rb_erase(&cfqg->rb_node, st); 921} 922 923static void 924cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg) 925{ 926 struct cfq_rb_root *st = &cfqd->grp_service_tree; 927 928 BUG_ON(cfqg->nr_cfqq < 1); 929 cfqg->nr_cfqq--; 930 931 /* If there are other cfq queues under this group, don't delete it */ 932 if (cfqg->nr_cfqq) 933 return; 934 935 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group"); 936 cfq_group_service_tree_del(st, cfqg); 937 cfqg->saved_workload_slice = 0; 938 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1); 939} 940 941static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq, 942 unsigned int *unaccounted_time) 943{ 944 unsigned int slice_used; 945 946 /* 947 * Queue got expired before even a single request completed or 948 * got expired immediately after first request completion. 949 */ 950 if (!cfqq->slice_start || cfqq->slice_start == jiffies) { 951 /* 952 * Also charge the seek time incurred to the group, otherwise 953 * if there are mutiple queues in the group, each can dispatch 954 * a single request on seeky media and cause lots of seek time 955 * and group will never know it. 956 */ 957 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start), 958 1); 959 } else { 960 slice_used = jiffies - cfqq->slice_start; 961 if (slice_used > cfqq->allocated_slice) { 962 *unaccounted_time = slice_used - cfqq->allocated_slice; 963 slice_used = cfqq->allocated_slice; 964 } 965 if (time_after(cfqq->slice_start, cfqq->dispatch_start)) 966 *unaccounted_time += cfqq->slice_start - 967 cfqq->dispatch_start; 968 } 969 970 return slice_used; 971} 972 973static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg, 974 struct cfq_queue *cfqq) 975{ 976 struct cfq_rb_root *st = &cfqd->grp_service_tree; 977 unsigned int used_sl, charge, unaccounted_sl = 0; 978 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg) 979 - cfqg->service_tree_idle.count; 980 981 BUG_ON(nr_sync < 0); 982 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl); 983 984 if (iops_mode(cfqd)) 985 charge = cfqq->slice_dispatch; 986 else if (!cfq_cfqq_sync(cfqq) && !nr_sync) 987 charge = cfqq->allocated_slice; 988 989 /* Can't update vdisktime while group is on service tree */ 990 cfq_group_service_tree_del(st, cfqg); 991 cfqg->vdisktime += cfq_scale_slice(charge, cfqg); 992 /* If a new weight was requested, update now, off tree */ 993 cfq_group_service_tree_add(st, cfqg); 994 995 /* This group is being expired. Save the context */ 996 if (time_after(cfqd->workload_expires, jiffies)) { 997 cfqg->saved_workload_slice = cfqd->workload_expires 998 - jiffies; 999 cfqg->saved_workload = cfqd->serving_type; 1000 cfqg->saved_serving_prio = cfqd->serving_prio; 1001 } else 1002 cfqg->saved_workload_slice = 0; 1003 1004 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime, 1005 st->min_vdisktime); 1006 cfq_log_cfqq(cfqq->cfqd, cfqq, 1007 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu", 1008 used_sl, cfqq->slice_dispatch, charge, 1009 iops_mode(cfqd), cfqq->nr_sectors); 1010 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl, 1011 unaccounted_sl); 1012 cfq_blkiocg_set_start_empty_time(&cfqg->blkg); 1013} 1014 1015/** 1016 * cfq_init_cfqg_base - initialize base part of a cfq_group 1017 * @cfqg: cfq_group to initialize 1018 * 1019 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED 1020 * is enabled or not. 1021 */ 1022static void cfq_init_cfqg_base(struct cfq_group *cfqg) 1023{ 1024 struct cfq_rb_root *st; 1025 int i, j; 1026 1027 for_each_cfqg_st(cfqg, i, j, st) 1028 *st = CFQ_RB_ROOT; 1029 RB_CLEAR_NODE(&cfqg->rb_node); 1030 1031 cfqg->ttime.last_end_request = jiffies; 1032} 1033 1034#ifdef CONFIG_CFQ_GROUP_IOSCHED 1035static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg) 1036{ 1037 if (blkg) 1038 return container_of(blkg, struct cfq_group, blkg); 1039 return NULL; 1040} 1041 1042static void cfq_update_blkio_group_weight(struct request_queue *q, 1043 struct blkio_group *blkg, 1044 unsigned int weight) 1045{ 1046 struct cfq_group *cfqg = cfqg_of_blkg(blkg); 1047 cfqg->new_weight = weight; 1048 cfqg->needs_update = true; 1049} 1050 1051static void cfq_link_blkio_group(struct request_queue *q, 1052 struct blkio_group *blkg) 1053{ 1054 struct cfq_data *cfqd = q->elevator->elevator_data; 1055 struct backing_dev_info *bdi = &q->backing_dev_info; 1056 struct cfq_group *cfqg = cfqg_of_blkg(blkg); 1057 unsigned int major, minor; 1058 1059 /* 1060 * Add group onto cgroup list. It might happen that bdi->dev is 1061 * not initialized yet. Initialize this new group without major 1062 * and minor info and this info will be filled in once a new thread 1063 * comes for IO. 1064 */ 1065 if (bdi->dev) { 1066 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); 1067 blkg->dev = MKDEV(major, minor); 1068 } 1069 1070 cfqd->nr_blkcg_linked_grps++; 1071 1072 /* Add group on cfqd list */ 1073 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list); 1074} 1075 1076static struct blkio_group *cfq_alloc_blkio_group(struct request_queue *q, 1077 struct blkio_cgroup *blkcg) 1078{ 1079 struct cfq_group *cfqg; 1080 1081 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, q->node); 1082 if (!cfqg) 1083 return NULL; 1084 1085 cfq_init_cfqg_base(cfqg); 1086 cfqg->weight = blkcg->weight; 1087 1088 /* 1089 * Take the initial reference that will be released on destroy 1090 * This can be thought of a joint reference by cgroup and 1091 * elevator which will be dropped by either elevator exit 1092 * or cgroup deletion path depending on who is exiting first. 1093 */ 1094 cfqg->ref = 1; 1095 1096 return &cfqg->blkg; 1097} 1098 1099/* 1100 * Search for the cfq group current task belongs to. request_queue lock must 1101 * be held. 1102 */ 1103static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd, 1104 struct blkio_cgroup *blkcg) 1105{ 1106 struct request_queue *q = cfqd->queue; 1107 struct backing_dev_info *bdi = &q->backing_dev_info; 1108 struct cfq_group *cfqg = NULL; 1109 1110 /* avoid lookup for the common case where there's no blkio cgroup */ 1111 if (blkcg == &blkio_root_cgroup) { 1112 cfqg = cfqd->root_group; 1113 } else { 1114 struct blkio_group *blkg; 1115 1116 blkg = blkg_lookup_create(blkcg, q, BLKIO_POLICY_PROP, false); 1117 if (!IS_ERR(blkg)) 1118 cfqg = cfqg_of_blkg(blkg); 1119 } 1120 1121 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) { 1122 unsigned int major, minor; 1123 1124 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); 1125 cfqg->blkg.dev = MKDEV(major, minor); 1126 } 1127 1128 return cfqg; 1129} 1130 1131static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg) 1132{ 1133 cfqg->ref++; 1134 return cfqg; 1135} 1136 1137static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) 1138{ 1139 /* Currently, all async queues are mapped to root group */ 1140 if (!cfq_cfqq_sync(cfqq)) 1141 cfqg = cfqq->cfqd->root_group; 1142 1143 cfqq->cfqg = cfqg; 1144 /* cfqq reference on cfqg */ 1145 cfqq->cfqg->ref++; 1146} 1147 1148static void cfq_put_cfqg(struct cfq_group *cfqg) 1149{ 1150 struct cfq_rb_root *st; 1151 int i, j; 1152 1153 BUG_ON(cfqg->ref <= 0); 1154 cfqg->ref--; 1155 if (cfqg->ref) 1156 return; 1157 for_each_cfqg_st(cfqg, i, j, st) 1158 BUG_ON(!RB_EMPTY_ROOT(&st->rb)); 1159 free_percpu(cfqg->blkg.stats_cpu); 1160 kfree(cfqg); 1161} 1162 1163static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg) 1164{ 1165 /* Something wrong if we are trying to remove same group twice */ 1166 BUG_ON(hlist_unhashed(&cfqg->cfqd_node)); 1167 1168 hlist_del_init(&cfqg->cfqd_node); 1169 1170 BUG_ON(cfqd->nr_blkcg_linked_grps <= 0); 1171 cfqd->nr_blkcg_linked_grps--; 1172 1173 /* 1174 * Put the reference taken at the time of creation so that when all 1175 * queues are gone, group can be destroyed. 1176 */ 1177 cfq_put_cfqg(cfqg); 1178} 1179 1180static bool cfq_release_cfq_groups(struct cfq_data *cfqd) 1181{ 1182 struct hlist_node *pos, *n; 1183 struct cfq_group *cfqg; 1184 bool empty = true; 1185 1186 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) { 1187 /* 1188 * If cgroup removal path got to blk_group first and removed 1189 * it from cgroup list, then it will take care of destroying 1190 * cfqg also. 1191 */ 1192 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg)) 1193 cfq_destroy_cfqg(cfqd, cfqg); 1194 else 1195 empty = false; 1196 } 1197 return empty; 1198} 1199 1200/* 1201 * Blk cgroup controller notification saying that blkio_group object is being 1202 * delinked as associated cgroup object is going away. That also means that 1203 * no new IO will come in this group. So get rid of this group as soon as 1204 * any pending IO in the group is finished. 1205 * 1206 * This function is called under rcu_read_lock(). key is the rcu protected 1207 * pointer. That means @q is a valid request_queue pointer as long as we 1208 * are rcu read lock. 1209 * 1210 * @q was fetched from blkio_group under blkio_cgroup->lock. That means 1211 * it should not be NULL as even if elevator was exiting, cgroup deltion 1212 * path got to it first. 1213 */ 1214static void cfq_unlink_blkio_group(struct request_queue *q, 1215 struct blkio_group *blkg) 1216{ 1217 struct cfq_data *cfqd = q->elevator->elevator_data; 1218 unsigned long flags; 1219 1220 spin_lock_irqsave(q->queue_lock, flags); 1221 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg)); 1222 spin_unlock_irqrestore(q->queue_lock, flags); 1223} 1224 1225static struct elevator_type iosched_cfq; 1226 1227static bool cfq_clear_queue(struct request_queue *q) 1228{ 1229 lockdep_assert_held(q->queue_lock); 1230 1231 /* shoot down blkgs iff the current elevator is cfq */ 1232 if (!q->elevator || q->elevator->type != &iosched_cfq) 1233 return true; 1234 1235 return cfq_release_cfq_groups(q->elevator->elevator_data); 1236} 1237 1238#else /* GROUP_IOSCHED */ 1239static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd, 1240 struct blkio_cgroup *blkcg) 1241{ 1242 return cfqd->root_group; 1243} 1244 1245static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg) 1246{ 1247 return cfqg; 1248} 1249 1250static inline void 1251cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) { 1252 cfqq->cfqg = cfqg; 1253} 1254 1255static void cfq_release_cfq_groups(struct cfq_data *cfqd) {} 1256static inline void cfq_put_cfqg(struct cfq_group *cfqg) {} 1257 1258#endif /* GROUP_IOSCHED */ 1259 1260/* 1261 * The cfqd->service_trees holds all pending cfq_queue's that have 1262 * requests waiting to be processed. It is sorted in the order that 1263 * we will service the queues. 1264 */ 1265static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1266 bool add_front) 1267{ 1268 struct rb_node **p, *parent; 1269 struct cfq_queue *__cfqq; 1270 unsigned long rb_key; 1271 struct cfq_rb_root *service_tree; 1272 int left; 1273 int new_cfqq = 1; 1274 1275 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq), 1276 cfqq_type(cfqq)); 1277 if (cfq_class_idle(cfqq)) { 1278 rb_key = CFQ_IDLE_DELAY; 1279 parent = rb_last(&service_tree->rb); 1280 if (parent && parent != &cfqq->rb_node) { 1281 __cfqq = rb_entry(parent, struct cfq_queue, rb_node); 1282 rb_key += __cfqq->rb_key; 1283 } else 1284 rb_key += jiffies; 1285 } else if (!add_front) { 1286 /* 1287 * Get our rb key offset. Subtract any residual slice 1288 * value carried from last service. A negative resid 1289 * count indicates slice overrun, and this should position 1290 * the next service time further away in the tree. 1291 */ 1292 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; 1293 rb_key -= cfqq->slice_resid; 1294 cfqq->slice_resid = 0; 1295 } else { 1296 rb_key = -HZ; 1297 __cfqq = cfq_rb_first(service_tree); 1298 rb_key += __cfqq ? __cfqq->rb_key : jiffies; 1299 } 1300 1301 if (!RB_EMPTY_NODE(&cfqq->rb_node)) { 1302 new_cfqq = 0; 1303 /* 1304 * same position, nothing more to do 1305 */ 1306 if (rb_key == cfqq->rb_key && 1307 cfqq->service_tree == service_tree) 1308 return; 1309 1310 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); 1311 cfqq->service_tree = NULL; 1312 } 1313 1314 left = 1; 1315 parent = NULL; 1316 cfqq->service_tree = service_tree; 1317 p = &service_tree->rb.rb_node; 1318 while (*p) { 1319 struct rb_node **n; 1320 1321 parent = *p; 1322 __cfqq = rb_entry(parent, struct cfq_queue, rb_node); 1323 1324 /* 1325 * sort by key, that represents service time. 1326 */ 1327 if (time_before(rb_key, __cfqq->rb_key)) 1328 n = &(*p)->rb_left; 1329 else { 1330 n = &(*p)->rb_right; 1331 left = 0; 1332 } 1333 1334 p = n; 1335 } 1336 1337 if (left) 1338 service_tree->left = &cfqq->rb_node; 1339 1340 cfqq->rb_key = rb_key; 1341 rb_link_node(&cfqq->rb_node, parent, p); 1342 rb_insert_color(&cfqq->rb_node, &service_tree->rb); 1343 service_tree->count++; 1344 if (add_front || !new_cfqq) 1345 return; 1346 cfq_group_notify_queue_add(cfqd, cfqq->cfqg); 1347} 1348 1349static struct cfq_queue * 1350cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, 1351 sector_t sector, struct rb_node **ret_parent, 1352 struct rb_node ***rb_link) 1353{ 1354 struct rb_node **p, *parent; 1355 struct cfq_queue *cfqq = NULL; 1356 1357 parent = NULL; 1358 p = &root->rb_node; 1359 while (*p) { 1360 struct rb_node **n; 1361 1362 parent = *p; 1363 cfqq = rb_entry(parent, struct cfq_queue, p_node); 1364 1365 /* 1366 * Sort strictly based on sector. Smallest to the left, 1367 * largest to the right. 1368 */ 1369 if (sector > blk_rq_pos(cfqq->next_rq)) 1370 n = &(*p)->rb_right; 1371 else if (sector < blk_rq_pos(cfqq->next_rq)) 1372 n = &(*p)->rb_left; 1373 else 1374 break; 1375 p = n; 1376 cfqq = NULL; 1377 } 1378 1379 *ret_parent = parent; 1380 if (rb_link) 1381 *rb_link = p; 1382 return cfqq; 1383} 1384 1385static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1386{ 1387 struct rb_node **p, *parent; 1388 struct cfq_queue *__cfqq; 1389 1390 if (cfqq->p_root) { 1391 rb_erase(&cfqq->p_node, cfqq->p_root); 1392 cfqq->p_root = NULL; 1393 } 1394 1395 if (cfq_class_idle(cfqq)) 1396 return; 1397 if (!cfqq->next_rq) 1398 return; 1399 1400 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; 1401 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, 1402 blk_rq_pos(cfqq->next_rq), &parent, &p); 1403 if (!__cfqq) { 1404 rb_link_node(&cfqq->p_node, parent, p); 1405 rb_insert_color(&cfqq->p_node, cfqq->p_root); 1406 } else 1407 cfqq->p_root = NULL; 1408} 1409 1410/* 1411 * Update cfqq's position in the service tree. 1412 */ 1413static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1414{ 1415 /* 1416 * Resorting requires the cfqq to be on the RR list already. 1417 */ 1418 if (cfq_cfqq_on_rr(cfqq)) { 1419 cfq_service_tree_add(cfqd, cfqq, 0); 1420 cfq_prio_tree_add(cfqd, cfqq); 1421 } 1422} 1423 1424/* 1425 * add to busy list of queues for service, trying to be fair in ordering 1426 * the pending list according to last request service 1427 */ 1428static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1429{ 1430 cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); 1431 BUG_ON(cfq_cfqq_on_rr(cfqq)); 1432 cfq_mark_cfqq_on_rr(cfqq); 1433 cfqd->busy_queues++; 1434 if (cfq_cfqq_sync(cfqq)) 1435 cfqd->busy_sync_queues++; 1436 1437 cfq_resort_rr_list(cfqd, cfqq); 1438} 1439 1440/* 1441 * Called when the cfqq no longer has requests pending, remove it from 1442 * the service tree. 1443 */ 1444static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1445{ 1446 cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); 1447 BUG_ON(!cfq_cfqq_on_rr(cfqq)); 1448 cfq_clear_cfqq_on_rr(cfqq); 1449 1450 if (!RB_EMPTY_NODE(&cfqq->rb_node)) { 1451 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); 1452 cfqq->service_tree = NULL; 1453 } 1454 if (cfqq->p_root) { 1455 rb_erase(&cfqq->p_node, cfqq->p_root); 1456 cfqq->p_root = NULL; 1457 } 1458 1459 cfq_group_notify_queue_del(cfqd, cfqq->cfqg); 1460 BUG_ON(!cfqd->busy_queues); 1461 cfqd->busy_queues--; 1462 if (cfq_cfqq_sync(cfqq)) 1463 cfqd->busy_sync_queues--; 1464} 1465 1466/* 1467 * rb tree support functions 1468 */ 1469static void cfq_del_rq_rb(struct request *rq) 1470{ 1471 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1472 const int sync = rq_is_sync(rq); 1473 1474 BUG_ON(!cfqq->queued[sync]); 1475 cfqq->queued[sync]--; 1476 1477 elv_rb_del(&cfqq->sort_list, rq); 1478 1479 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) { 1480 /* 1481 * Queue will be deleted from service tree when we actually 1482 * expire it later. Right now just remove it from prio tree 1483 * as it is empty. 1484 */ 1485 if (cfqq->p_root) { 1486 rb_erase(&cfqq->p_node, cfqq->p_root); 1487 cfqq->p_root = NULL; 1488 } 1489 } 1490} 1491 1492static void cfq_add_rq_rb(struct request *rq) 1493{ 1494 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1495 struct cfq_data *cfqd = cfqq->cfqd; 1496 struct request *prev; 1497 1498 cfqq->queued[rq_is_sync(rq)]++; 1499 1500 elv_rb_add(&cfqq->sort_list, rq); 1501 1502 if (!cfq_cfqq_on_rr(cfqq)) 1503 cfq_add_cfqq_rr(cfqd, cfqq); 1504 1505 /* 1506 * check if this request is a better next-serve candidate 1507 */ 1508 prev = cfqq->next_rq; 1509 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position); 1510 1511 /* 1512 * adjust priority tree position, if ->next_rq changes 1513 */ 1514 if (prev != cfqq->next_rq) 1515 cfq_prio_tree_add(cfqd, cfqq); 1516 1517 BUG_ON(!cfqq->next_rq); 1518} 1519 1520static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) 1521{ 1522 elv_rb_del(&cfqq->sort_list, rq); 1523 cfqq->queued[rq_is_sync(rq)]--; 1524 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, 1525 rq_data_dir(rq), rq_is_sync(rq)); 1526 cfq_add_rq_rb(rq); 1527 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg, 1528 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq), 1529 rq_is_sync(rq)); 1530} 1531 1532static struct request * 1533cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) 1534{ 1535 struct task_struct *tsk = current; 1536 struct cfq_io_cq *cic; 1537 struct cfq_queue *cfqq; 1538 1539 cic = cfq_cic_lookup(cfqd, tsk->io_context); 1540 if (!cic) 1541 return NULL; 1542 1543 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); 1544 if (cfqq) { 1545 sector_t sector = bio->bi_sector + bio_sectors(bio); 1546 1547 return elv_rb_find(&cfqq->sort_list, sector); 1548 } 1549 1550 return NULL; 1551} 1552 1553static void cfq_activate_request(struct request_queue *q, struct request *rq) 1554{ 1555 struct cfq_data *cfqd = q->elevator->elevator_data; 1556 1557 cfqd->rq_in_driver++; 1558 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", 1559 cfqd->rq_in_driver); 1560 1561 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); 1562} 1563 1564static void cfq_deactivate_request(struct request_queue *q, struct request *rq) 1565{ 1566 struct cfq_data *cfqd = q->elevator->elevator_data; 1567 1568 WARN_ON(!cfqd->rq_in_driver); 1569 cfqd->rq_in_driver--; 1570 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", 1571 cfqd->rq_in_driver); 1572} 1573 1574static void cfq_remove_request(struct request *rq) 1575{ 1576 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1577 1578 if (cfqq->next_rq == rq) 1579 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); 1580 1581 list_del_init(&rq->queuelist); 1582 cfq_del_rq_rb(rq); 1583 1584 cfqq->cfqd->rq_queued--; 1585 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, 1586 rq_data_dir(rq), rq_is_sync(rq)); 1587 if (rq->cmd_flags & REQ_PRIO) { 1588 WARN_ON(!cfqq->prio_pending); 1589 cfqq->prio_pending--; 1590 } 1591} 1592 1593static int cfq_merge(struct request_queue *q, struct request **req, 1594 struct bio *bio) 1595{ 1596 struct cfq_data *cfqd = q->elevator->elevator_data; 1597 struct request *__rq; 1598 1599 __rq = cfq_find_rq_fmerge(cfqd, bio); 1600 if (__rq && elv_rq_merge_ok(__rq, bio)) { 1601 *req = __rq; 1602 return ELEVATOR_FRONT_MERGE; 1603 } 1604 1605 return ELEVATOR_NO_MERGE; 1606} 1607 1608static void cfq_merged_request(struct request_queue *q, struct request *req, 1609 int type) 1610{ 1611 if (type == ELEVATOR_FRONT_MERGE) { 1612 struct cfq_queue *cfqq = RQ_CFQQ(req); 1613 1614 cfq_reposition_rq_rb(cfqq, req); 1615 } 1616} 1617 1618static void cfq_bio_merged(struct request_queue *q, struct request *req, 1619 struct bio *bio) 1620{ 1621 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg, 1622 bio_data_dir(bio), cfq_bio_sync(bio)); 1623} 1624 1625static void 1626cfq_merged_requests(struct request_queue *q, struct request *rq, 1627 struct request *next) 1628{ 1629 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1630 struct cfq_data *cfqd = q->elevator->elevator_data; 1631 1632 /* 1633 * reposition in fifo if next is older than rq 1634 */ 1635 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && 1636 time_before(rq_fifo_time(next), rq_fifo_time(rq))) { 1637 list_move(&rq->queuelist, &next->queuelist); 1638 rq_set_fifo_time(rq, rq_fifo_time(next)); 1639 } 1640 1641 if (cfqq->next_rq == next) 1642 cfqq->next_rq = rq; 1643 cfq_remove_request(next); 1644 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg, 1645 rq_data_dir(next), rq_is_sync(next)); 1646 1647 cfqq = RQ_CFQQ(next); 1648 /* 1649 * all requests of this queue are merged to other queues, delete it 1650 * from the service tree. If it's the active_queue, 1651 * cfq_dispatch_requests() will choose to expire it or do idle 1652 */ 1653 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) && 1654 cfqq != cfqd->active_queue) 1655 cfq_del_cfqq_rr(cfqd, cfqq); 1656} 1657 1658static int cfq_allow_merge(struct request_queue *q, struct request *rq, 1659 struct bio *bio) 1660{ 1661 struct cfq_data *cfqd = q->elevator->elevator_data; 1662 struct cfq_io_cq *cic; 1663 struct cfq_queue *cfqq; 1664 1665 /* 1666 * Disallow merge of a sync bio into an async request. 1667 */ 1668 if (cfq_bio_sync(bio) && !rq_is_sync(rq)) 1669 return false; 1670 1671 /* 1672 * Lookup the cfqq that this bio will be queued with and allow 1673 * merge only if rq is queued there. 1674 */ 1675 cic = cfq_cic_lookup(cfqd, current->io_context); 1676 if (!cic) 1677 return false; 1678 1679 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); 1680 return cfqq == RQ_CFQQ(rq); 1681} 1682 1683static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1684{ 1685 del_timer(&cfqd->idle_slice_timer); 1686 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg); 1687} 1688 1689static void __cfq_set_active_queue(struct cfq_data *cfqd, 1690 struct cfq_queue *cfqq) 1691{ 1692 if (cfqq) { 1693 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d", 1694 cfqd->serving_prio, cfqd->serving_type); 1695 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg); 1696 cfqq->slice_start = 0; 1697 cfqq->dispatch_start = jiffies; 1698 cfqq->allocated_slice = 0; 1699 cfqq->slice_end = 0; 1700 cfqq->slice_dispatch = 0; 1701 cfqq->nr_sectors = 0; 1702 1703 cfq_clear_cfqq_wait_request(cfqq); 1704 cfq_clear_cfqq_must_dispatch(cfqq); 1705 cfq_clear_cfqq_must_alloc_slice(cfqq); 1706 cfq_clear_cfqq_fifo_expire(cfqq); 1707 cfq_mark_cfqq_slice_new(cfqq); 1708 1709 cfq_del_timer(cfqd, cfqq); 1710 } 1711 1712 cfqd->active_queue = cfqq; 1713} 1714 1715/* 1716 * current cfqq expired its slice (or was too idle), select new one 1717 */ 1718static void 1719__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1720 bool timed_out) 1721{ 1722 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); 1723 1724 if (cfq_cfqq_wait_request(cfqq)) 1725 cfq_del_timer(cfqd, cfqq); 1726 1727 cfq_clear_cfqq_wait_request(cfqq); 1728 cfq_clear_cfqq_wait_busy(cfqq); 1729 1730 /* 1731 * If this cfqq is shared between multiple processes, check to 1732 * make sure that those processes are still issuing I/Os within 1733 * the mean seek distance. If not, it may be time to break the 1734 * queues apart again. 1735 */ 1736 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq)) 1737 cfq_mark_cfqq_split_coop(cfqq); 1738 1739 /* 1740 * store what was left of this slice, if the queue idled/timed out 1741 */ 1742 if (timed_out) { 1743 if (cfq_cfqq_slice_new(cfqq)) 1744 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq); 1745 else 1746 cfqq->slice_resid = cfqq->slice_end - jiffies; 1747 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); 1748 } 1749 1750 cfq_group_served(cfqd, cfqq->cfqg, cfqq); 1751 1752 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) 1753 cfq_del_cfqq_rr(cfqd, cfqq); 1754 1755 cfq_resort_rr_list(cfqd, cfqq); 1756 1757 if (cfqq == cfqd->active_queue) 1758 cfqd->active_queue = NULL; 1759 1760 if (cfqd->active_cic) { 1761 put_io_context(cfqd->active_cic->icq.ioc); 1762 cfqd->active_cic = NULL; 1763 } 1764} 1765 1766static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) 1767{ 1768 struct cfq_queue *cfqq = cfqd->active_queue; 1769 1770 if (cfqq) 1771 __cfq_slice_expired(cfqd, cfqq, timed_out); 1772} 1773 1774/* 1775 * Get next queue for service. Unless we have a queue preemption, 1776 * we'll simply select the first cfqq in the service tree. 1777 */ 1778static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) 1779{ 1780 struct cfq_rb_root *service_tree = 1781 service_tree_for(cfqd->serving_group, cfqd->serving_prio, 1782 cfqd->serving_type); 1783 1784 if (!cfqd->rq_queued) 1785 return NULL; 1786 1787 /* There is nothing to dispatch */ 1788 if (!service_tree) 1789 return NULL; 1790 if (RB_EMPTY_ROOT(&service_tree->rb)) 1791 return NULL; 1792 return cfq_rb_first(service_tree); 1793} 1794 1795static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd) 1796{ 1797 struct cfq_group *cfqg; 1798 struct cfq_queue *cfqq; 1799 int i, j; 1800 struct cfq_rb_root *st; 1801 1802 if (!cfqd->rq_queued) 1803 return NULL; 1804 1805 cfqg = cfq_get_next_cfqg(cfqd); 1806 if (!cfqg) 1807 return NULL; 1808 1809 for_each_cfqg_st(cfqg, i, j, st) 1810 if ((cfqq = cfq_rb_first(st)) != NULL) 1811 return cfqq; 1812 return NULL; 1813} 1814 1815/* 1816 * Get and set a new active queue for service. 1817 */ 1818static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, 1819 struct cfq_queue *cfqq) 1820{ 1821 if (!cfqq) 1822 cfqq = cfq_get_next_queue(cfqd); 1823 1824 __cfq_set_active_queue(cfqd, cfqq); 1825 return cfqq; 1826} 1827 1828static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, 1829 struct request *rq) 1830{ 1831 if (blk_rq_pos(rq) >= cfqd->last_position) 1832 return blk_rq_pos(rq) - cfqd->last_position; 1833 else 1834 return cfqd->last_position - blk_rq_pos(rq); 1835} 1836 1837static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1838 struct request *rq) 1839{ 1840 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR; 1841} 1842 1843static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, 1844 struct cfq_queue *cur_cfqq) 1845{ 1846 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; 1847 struct rb_node *parent, *node; 1848 struct cfq_queue *__cfqq; 1849 sector_t sector = cfqd->last_position; 1850 1851 if (RB_EMPTY_ROOT(root)) 1852 return NULL; 1853 1854 /* 1855 * First, if we find a request starting at the end of the last 1856 * request, choose it. 1857 */ 1858 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); 1859 if (__cfqq) 1860 return __cfqq; 1861 1862 /* 1863 * If the exact sector wasn't found, the parent of the NULL leaf 1864 * will contain the closest sector. 1865 */ 1866 __cfqq = rb_entry(parent, struct cfq_queue, p_node); 1867 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) 1868 return __cfqq; 1869 1870 if (blk_rq_pos(__cfqq->next_rq) < sector) 1871 node = rb_next(&__cfqq->p_node); 1872 else 1873 node = rb_prev(&__cfqq->p_node); 1874 if (!node) 1875 return NULL; 1876 1877 __cfqq = rb_entry(node, struct cfq_queue, p_node); 1878 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) 1879 return __cfqq; 1880 1881 return NULL; 1882} 1883 1884/* 1885 * cfqd - obvious 1886 * cur_cfqq - passed in so that we don't decide that the current queue is 1887 * closely cooperating with itself. 1888 * 1889 * So, basically we're assuming that that cur_cfqq has dispatched at least 1890 * one request, and that cfqd->last_position reflects a position on the disk 1891 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid 1892 * assumption. 1893 */ 1894static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, 1895 struct cfq_queue *cur_cfqq) 1896{ 1897 struct cfq_queue *cfqq; 1898 1899 if (cfq_class_idle(cur_cfqq)) 1900 return NULL; 1901 if (!cfq_cfqq_sync(cur_cfqq)) 1902 return NULL; 1903 if (CFQQ_SEEKY(cur_cfqq)) 1904 return NULL; 1905 1906 /* 1907 * Don't search priority tree if it's the only queue in the group. 1908 */ 1909 if (cur_cfqq->cfqg->nr_cfqq == 1) 1910 return NULL; 1911 1912 /* 1913 * We should notice if some of the queues are cooperating, eg 1914 * working closely on the same area of the disk. In that case, 1915 * we can group them together and don't waste time idling. 1916 */ 1917 cfqq = cfqq_close(cfqd, cur_cfqq); 1918 if (!cfqq) 1919 return NULL; 1920 1921 /* If new queue belongs to different cfq_group, don't choose it */ 1922 if (cur_cfqq->cfqg != cfqq->cfqg) 1923 return NULL; 1924 1925 /* 1926 * It only makes sense to merge sync queues. 1927 */ 1928 if (!cfq_cfqq_sync(cfqq)) 1929 return NULL; 1930 if (CFQQ_SEEKY(cfqq)) 1931 return NULL; 1932 1933 /* 1934 * Do not merge queues of different priority classes 1935 */ 1936 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) 1937 return NULL; 1938 1939 return cfqq; 1940} 1941 1942/* 1943 * Determine whether we should enforce idle window for this queue. 1944 */ 1945 1946static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1947{ 1948 enum wl_prio_t prio = cfqq_prio(cfqq); 1949 struct cfq_rb_root *service_tree = cfqq->service_tree; 1950 1951 BUG_ON(!service_tree); 1952 BUG_ON(!service_tree->count); 1953 1954 if (!cfqd->cfq_slice_idle) 1955 return false; 1956 1957 /* We never do for idle class queues. */ 1958 if (prio == IDLE_WORKLOAD) 1959 return false; 1960 1961 /* We do for queues that were marked with idle window flag. */ 1962 if (cfq_cfqq_idle_window(cfqq) && 1963 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)) 1964 return true; 1965 1966 /* 1967 * Otherwise, we do only if they are the last ones 1968 * in their service tree. 1969 */ 1970 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) && 1971 !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false)) 1972 return true; 1973 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", 1974 service_tree->count); 1975 return false; 1976} 1977 1978static void cfq_arm_slice_timer(struct cfq_data *cfqd) 1979{ 1980 struct cfq_queue *cfqq = cfqd->active_queue; 1981 struct cfq_io_cq *cic; 1982 unsigned long sl, group_idle = 0; 1983 1984 /* 1985 * SSD device without seek penalty, disable idling. But only do so 1986 * for devices that support queuing, otherwise we still have a problem 1987 * with sync vs async workloads. 1988 */ 1989 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) 1990 return; 1991 1992 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); 1993 WARN_ON(cfq_cfqq_slice_new(cfqq)); 1994 1995 /* 1996 * idle is disabled, either manually or by past process history 1997 */ 1998 if (!cfq_should_idle(cfqd, cfqq)) { 1999 /* no queue idling. Check for group idling */ 2000 if (cfqd->cfq_group_idle) 2001 group_idle = cfqd->cfq_group_idle; 2002 else 2003 return; 2004 } 2005 2006 /* 2007 * still active requests from this queue, don't idle 2008 */ 2009 if (cfqq->dispatched) 2010 return; 2011 2012 /* 2013 * task has exited, don't wait 2014 */ 2015 cic = cfqd->active_cic; 2016 if (!cic || !atomic_read(&cic->icq.ioc->nr_tasks)) 2017 return; 2018 2019 /* 2020 * If our average think time is larger than the remaining time 2021 * slice, then don't idle. This avoids overrunning the allotted 2022 * time slice. 2023 */ 2024 if (sample_valid(cic->ttime.ttime_samples) && 2025 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) { 2026 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu", 2027 cic->ttime.ttime_mean); 2028 return; 2029 } 2030 2031 /* There are other queues in the group, don't do group idle */ 2032 if (group_idle && cfqq->cfqg->nr_cfqq > 1) 2033 return; 2034 2035 cfq_mark_cfqq_wait_request(cfqq); 2036 2037 if (group_idle) 2038 sl = cfqd->cfq_group_idle; 2039 else 2040 sl = cfqd->cfq_slice_idle; 2041 2042 mod_timer(&cfqd->idle_slice_timer, jiffies + sl); 2043 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg); 2044 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl, 2045 group_idle ? 1 : 0); 2046} 2047 2048/* 2049 * Move request from internal lists to the request queue dispatch list. 2050 */ 2051static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) 2052{ 2053 struct cfq_data *cfqd = q->elevator->elevator_data; 2054 struct cfq_queue *cfqq = RQ_CFQQ(rq); 2055 2056 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); 2057 2058 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); 2059 cfq_remove_request(rq); 2060 cfqq->dispatched++; 2061 (RQ_CFQG(rq))->dispatched++; 2062 elv_dispatch_sort(q, rq); 2063 2064 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++; 2065 cfqq->nr_sectors += blk_rq_sectors(rq); 2066 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq), 2067 rq_data_dir(rq), rq_is_sync(rq)); 2068} 2069 2070/* 2071 * return expired entry, or NULL to just start from scratch in rbtree 2072 */ 2073static struct request *cfq_check_fifo(struct cfq_queue *cfqq) 2074{ 2075 struct request *rq = NULL; 2076 2077 if (cfq_cfqq_fifo_expire(cfqq)) 2078 return NULL; 2079 2080 cfq_mark_cfqq_fifo_expire(cfqq); 2081 2082 if (list_empty(&cfqq->fifo)) 2083 return NULL; 2084 2085 rq = rq_entry_fifo(cfqq->fifo.next); 2086 if (time_before(jiffies, rq_fifo_time(rq))) 2087 rq = NULL; 2088 2089 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); 2090 return rq; 2091} 2092 2093static inline int 2094cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) 2095{ 2096 const int base_rq = cfqd->cfq_slice_async_rq; 2097 2098 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); 2099 2100 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio); 2101} 2102 2103/* 2104 * Must be called with the queue_lock held. 2105 */ 2106static int cfqq_process_refs(struct cfq_queue *cfqq) 2107{ 2108 int process_refs, io_refs; 2109 2110 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; 2111 process_refs = cfqq->ref - io_refs; 2112 BUG_ON(process_refs < 0); 2113 return process_refs; 2114} 2115 2116static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) 2117{ 2118 int process_refs, new_process_refs; 2119 struct cfq_queue *__cfqq; 2120 2121 /* 2122 * If there are no process references on the new_cfqq, then it is 2123 * unsafe to follow the ->new_cfqq chain as other cfqq's in the 2124 * chain may have dropped their last reference (not just their 2125 * last process reference). 2126 */ 2127 if (!cfqq_process_refs(new_cfqq)) 2128 return; 2129 2130 /* Avoid a circular list and skip interim queue merges */ 2131 while ((__cfqq = new_cfqq->new_cfqq)) { 2132 if (__cfqq == cfqq) 2133 return; 2134 new_cfqq = __cfqq; 2135 } 2136 2137 process_refs = cfqq_process_refs(cfqq); 2138 new_process_refs = cfqq_process_refs(new_cfqq); 2139 /* 2140 * If the process for the cfqq has gone away, there is no 2141 * sense in merging the queues. 2142 */ 2143 if (process_refs == 0 || new_process_refs == 0) 2144 return; 2145 2146 /* 2147 * Merge in the direction of the lesser amount of work. 2148 */ 2149 if (new_process_refs >= process_refs) { 2150 cfqq->new_cfqq = new_cfqq; 2151 new_cfqq->ref += process_refs; 2152 } else { 2153 new_cfqq->new_cfqq = cfqq; 2154 cfqq->ref += new_process_refs; 2155 } 2156} 2157 2158static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, 2159 struct cfq_group *cfqg, enum wl_prio_t prio) 2160{ 2161 struct cfq_queue *queue; 2162 int i; 2163 bool key_valid = false; 2164 unsigned long lowest_key = 0; 2165 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD; 2166 2167 for (i = 0; i <= SYNC_WORKLOAD; ++i) { 2168 /* select the one with lowest rb_key */ 2169 queue = cfq_rb_first(service_tree_for(cfqg, prio, i)); 2170 if (queue && 2171 (!key_valid || time_before(queue->rb_key, lowest_key))) { 2172 lowest_key = queue->rb_key; 2173 cur_best = i; 2174 key_valid = true; 2175 } 2176 } 2177 2178 return cur_best; 2179} 2180 2181static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg) 2182{ 2183 unsigned slice; 2184 unsigned count; 2185 struct cfq_rb_root *st; 2186 unsigned group_slice; 2187 enum wl_prio_t original_prio = cfqd->serving_prio; 2188 2189 /* Choose next priority. RT > BE > IDLE */ 2190 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg)) 2191 cfqd->serving_prio = RT_WORKLOAD; 2192 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg)) 2193 cfqd->serving_prio = BE_WORKLOAD; 2194 else { 2195 cfqd->serving_prio = IDLE_WORKLOAD; 2196 cfqd->workload_expires = jiffies + 1; 2197 return; 2198 } 2199 2200 if (original_prio != cfqd->serving_prio) 2201 goto new_workload; 2202 2203 /* 2204 * For RT and BE, we have to choose also the type 2205 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload 2206 * expiration time 2207 */ 2208 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type); 2209 count = st->count; 2210 2211 /* 2212 * check workload expiration, and that we still have other queues ready 2213 */ 2214 if (count && !time_after(jiffies, cfqd->workload_expires)) 2215 return; 2216 2217new_workload: 2218 /* otherwise select new workload type */ 2219 cfqd->serving_type = 2220 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio); 2221 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type); 2222 count = st->count; 2223 2224 /* 2225 * the workload slice is computed as a fraction of target latency 2226 * proportional to the number of queues in that workload, over 2227 * all the queues in the same priority class 2228 */ 2229 group_slice = cfq_group_slice(cfqd, cfqg); 2230 2231 slice = group_slice * count / 2232 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio], 2233 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg)); 2234 2235 if (cfqd->serving_type == ASYNC_WORKLOAD) { 2236 unsigned int tmp; 2237 2238 /* 2239 * Async queues are currently system wide. Just taking 2240 * proportion of queues with-in same group will lead to higher 2241 * async ratio system wide as generally root group is going 2242 * to have higher weight. A more accurate thing would be to 2243 * calculate system wide asnc/sync ratio. 2244 */ 2245 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg); 2246 tmp = tmp/cfqd->busy_queues; 2247 slice = min_t(unsigned, slice, tmp); 2248 2249 /* async workload slice is scaled down according to 2250 * the sync/async slice ratio. */ 2251 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1]; 2252 } else 2253 /* sync workload slice is at least 2 * cfq_slice_idle */ 2254 slice = max(slice, 2 * cfqd->cfq_slice_idle); 2255 2256 slice = max_t(unsigned, slice, CFQ_MIN_TT); 2257 cfq_log(cfqd, "workload slice:%d", slice); 2258 cfqd->workload_expires = jiffies + slice; 2259} 2260 2261static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd) 2262{ 2263 struct cfq_rb_root *st = &cfqd->grp_service_tree; 2264 struct cfq_group *cfqg; 2265 2266 if (RB_EMPTY_ROOT(&st->rb)) 2267 return NULL; 2268 cfqg = cfq_rb_first_group(st); 2269 update_min_vdisktime(st); 2270 return cfqg; 2271} 2272 2273static void cfq_choose_cfqg(struct cfq_data *cfqd) 2274{ 2275 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd); 2276 2277 cfqd->serving_group = cfqg; 2278 2279 /* Restore the workload type data */ 2280 if (cfqg->saved_workload_slice) { 2281 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice; 2282 cfqd->serving_type = cfqg->saved_workload; 2283 cfqd->serving_prio = cfqg->saved_serving_prio; 2284 } else 2285 cfqd->workload_expires = jiffies - 1; 2286 2287 choose_service_tree(cfqd, cfqg); 2288} 2289 2290/* 2291 * Select a queue for service. If we have a current active queue, 2292 * check whether to continue servicing it, or retrieve and set a new one. 2293 */ 2294static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) 2295{ 2296 struct cfq_queue *cfqq, *new_cfqq = NULL; 2297 2298 cfqq = cfqd->active_queue; 2299 if (!cfqq) 2300 goto new_queue; 2301 2302 if (!cfqd->rq_queued) 2303 return NULL; 2304 2305 /* 2306 * We were waiting for group to get backlogged. Expire the queue 2307 */ 2308 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list)) 2309 goto expire; 2310 2311 /* 2312 * The active queue has run out of time, expire it and select new. 2313 */ 2314 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) { 2315 /* 2316 * If slice had not expired at the completion of last request 2317 * we might not have turned on wait_busy flag. Don't expire 2318 * the queue yet. Allow the group to get backlogged. 2319 * 2320 * The very fact that we have used the slice, that means we 2321 * have been idling all along on this queue and it should be 2322 * ok to wait for this request to complete. 2323 */ 2324 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list) 2325 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) { 2326 cfqq = NULL; 2327 goto keep_queue; 2328 } else 2329 goto check_group_idle; 2330 } 2331 2332 /* 2333 * The active queue has requests and isn't expired, allow it to 2334 * dispatch. 2335 */ 2336 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) 2337 goto keep_queue; 2338 2339 /* 2340 * If another queue has a request waiting within our mean seek 2341 * distance, let it run. The expire code will check for close 2342 * cooperators and put the close queue at the front of the service 2343 * tree. If possible, merge the expiring queue with the new cfqq. 2344 */ 2345 new_cfqq = cfq_close_cooperator(cfqd, cfqq); 2346 if (new_cfqq) { 2347 if (!cfqq->new_cfqq) 2348 cfq_setup_merge(cfqq, new_cfqq); 2349 goto expire; 2350 } 2351 2352 /* 2353 * No requests pending. If the active queue still has requests in 2354 * flight or is idling for a new request, allow either of these 2355 * conditions to happen (or time out) before selecting a new queue. 2356 */ 2357 if (timer_pending(&cfqd->idle_slice_timer)) { 2358 cfqq = NULL; 2359 goto keep_queue; 2360 } 2361 2362 /* 2363 * This is a deep seek queue, but the device is much faster than 2364 * the queue can deliver, don't idle 2365 **/ 2366 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) && 2367 (cfq_cfqq_slice_new(cfqq) || 2368 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) { 2369 cfq_clear_cfqq_deep(cfqq); 2370 cfq_clear_cfqq_idle_window(cfqq); 2371 } 2372 2373 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) { 2374 cfqq = NULL; 2375 goto keep_queue; 2376 } 2377 2378 /* 2379 * If group idle is enabled and there are requests dispatched from 2380 * this group, wait for requests to complete. 2381 */ 2382check_group_idle: 2383 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 && 2384 cfqq->cfqg->dispatched && 2385 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) { 2386 cfqq = NULL; 2387 goto keep_queue; 2388 } 2389 2390expire: 2391 cfq_slice_expired(cfqd, 0); 2392new_queue: 2393 /* 2394 * Current queue expired. Check if we have to switch to a new 2395 * service tree 2396 */ 2397 if (!new_cfqq) 2398 cfq_choose_cfqg(cfqd); 2399 2400 cfqq = cfq_set_active_queue(cfqd, new_cfqq); 2401keep_queue: 2402 return cfqq; 2403} 2404 2405static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) 2406{ 2407 int dispatched = 0; 2408 2409 while (cfqq->next_rq) { 2410 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); 2411 dispatched++; 2412 } 2413 2414 BUG_ON(!list_empty(&cfqq->fifo)); 2415 2416 /* By default cfqq is not expired if it is empty. Do it explicitly */ 2417 __cfq_slice_expired(cfqq->cfqd, cfqq, 0); 2418 return dispatched; 2419} 2420 2421/* 2422 * Drain our current requests. Used for barriers and when switching 2423 * io schedulers on-the-fly. 2424 */ 2425static int cfq_forced_dispatch(struct cfq_data *cfqd) 2426{ 2427 struct cfq_queue *cfqq; 2428 int dispatched = 0; 2429 2430 /* Expire the timeslice of the current active queue first */ 2431 cfq_slice_expired(cfqd, 0); 2432 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) { 2433 __cfq_set_active_queue(cfqd, cfqq); 2434 dispatched += __cfq_forced_dispatch_cfqq(cfqq); 2435 } 2436 2437 BUG_ON(cfqd->busy_queues); 2438 2439 cfq_log(cfqd, "forced_dispatch=%d", dispatched); 2440 return dispatched; 2441} 2442 2443static inline bool cfq_slice_used_soon(struct cfq_data *cfqd, 2444 struct cfq_queue *cfqq) 2445{ 2446 /* the queue hasn't finished any request, can't estimate */ 2447 if (cfq_cfqq_slice_new(cfqq)) 2448 return true; 2449 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched, 2450 cfqq->slice_end)) 2451 return true; 2452 2453 return false; 2454} 2455 2456static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) 2457{ 2458 unsigned int max_dispatch; 2459 2460 /* 2461 * Drain async requests before we start sync IO 2462 */ 2463 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC]) 2464 return false; 2465 2466 /* 2467 * If this is an async queue and we have sync IO in flight, let it wait 2468 */ 2469 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq)) 2470 return false; 2471 2472 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1); 2473 if (cfq_class_idle(cfqq)) 2474 max_dispatch = 1; 2475 2476 /* 2477 * Does this cfqq already have too much IO in flight? 2478 */ 2479 if (cfqq->dispatched >= max_dispatch) { 2480 bool promote_sync = false; 2481 /* 2482 * idle queue must always only have a single IO in flight 2483 */ 2484 if (cfq_class_idle(cfqq)) 2485 return false; 2486 2487 /* 2488 * If there is only one sync queue 2489 * we can ignore async queue here and give the sync 2490 * queue no dispatch limit. The reason is a sync queue can 2491 * preempt async queue, limiting the sync queue doesn't make 2492 * sense. This is useful for aiostress test. 2493 */ 2494 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1) 2495 promote_sync = true; 2496 2497 /* 2498 * We have other queues, don't allow more IO from this one 2499 */ 2500 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) && 2501 !promote_sync) 2502 return false; 2503 2504 /* 2505 * Sole queue user, no limit 2506 */ 2507 if (cfqd->busy_queues == 1 || promote_sync) 2508 max_dispatch = -1; 2509 else 2510 /* 2511 * Normally we start throttling cfqq when cfq_quantum/2 2512 * requests have been dispatched. But we can drive 2513 * deeper queue depths at the beginning of slice 2514 * subjected to upper limit of cfq_quantum. 2515 * */ 2516 max_dispatch = cfqd->cfq_quantum; 2517 } 2518 2519 /* 2520 * Async queues must wait a bit before being allowed dispatch. 2521 * We also ramp up the dispatch depth gradually for async IO, 2522 * based on the last sync IO we serviced 2523 */ 2524 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { 2525 unsigned long last_sync = jiffies - cfqd->last_delayed_sync; 2526 unsigned int depth; 2527 2528 depth = last_sync / cfqd->cfq_slice[1]; 2529 if (!depth && !cfqq->dispatched) 2530 depth = 1; 2531 if (depth < max_dispatch) 2532 max_dispatch = depth; 2533 } 2534 2535 /* 2536 * If we're below the current max, allow a dispatch 2537 */ 2538 return cfqq->dispatched < max_dispatch; 2539} 2540 2541/* 2542 * Dispatch a request from cfqq, moving them to the request queue 2543 * dispatch list. 2544 */ 2545static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) 2546{ 2547 struct request *rq; 2548 2549 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); 2550 2551 if (!cfq_may_dispatch(cfqd, cfqq)) 2552 return false; 2553 2554 /* 2555 * follow expired path, else get first next available 2556 */ 2557 rq = cfq_check_fifo(cfqq); 2558 if (!rq) 2559 rq = cfqq->next_rq; 2560 2561 /* 2562 * insert request into driver dispatch list 2563 */ 2564 cfq_dispatch_insert(cfqd->queue, rq); 2565 2566 if (!cfqd->active_cic) { 2567 struct cfq_io_cq *cic = RQ_CIC(rq); 2568 2569 atomic_long_inc(&cic->icq.ioc->refcount); 2570 cfqd->active_cic = cic; 2571 } 2572 2573 return true; 2574} 2575 2576/* 2577 * Find the cfqq that we need to service and move a request from that to the 2578 * dispatch list 2579 */ 2580static int cfq_dispatch_requests(struct request_queue *q, int force) 2581{ 2582 struct cfq_data *cfqd = q->elevator->elevator_data; 2583 struct cfq_queue *cfqq; 2584 2585 if (!cfqd->busy_queues) 2586 return 0; 2587 2588 if (unlikely(force)) 2589 return cfq_forced_dispatch(cfqd); 2590 2591 cfqq = cfq_select_queue(cfqd); 2592 if (!cfqq) 2593 return 0; 2594 2595 /* 2596 * Dispatch a request from this cfqq, if it is allowed 2597 */ 2598 if (!cfq_dispatch_request(cfqd, cfqq)) 2599 return 0; 2600 2601 cfqq->slice_dispatch++; 2602 cfq_clear_cfqq_must_dispatch(cfqq); 2603 2604 /* 2605 * expire an async queue immediately if it has used up its slice. idle 2606 * queue always expire after 1 dispatch round. 2607 */ 2608 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && 2609 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || 2610 cfq_class_idle(cfqq))) { 2611 cfqq->slice_end = jiffies + 1; 2612 cfq_slice_expired(cfqd, 0); 2613 } 2614 2615 cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); 2616 return 1; 2617} 2618 2619/* 2620 * task holds one reference to the queue, dropped when task exits. each rq 2621 * in-flight on this queue also holds a reference, dropped when rq is freed. 2622 * 2623 * Each cfq queue took a reference on the parent group. Drop it now. 2624 * queue lock must be held here. 2625 */ 2626static void cfq_put_queue(struct cfq_queue *cfqq) 2627{ 2628 struct cfq_data *cfqd = cfqq->cfqd; 2629 struct cfq_group *cfqg; 2630 2631 BUG_ON(cfqq->ref <= 0); 2632 2633 cfqq->ref--; 2634 if (cfqq->ref) 2635 return; 2636 2637 cfq_log_cfqq(cfqd, cfqq, "put_queue"); 2638 BUG_ON(rb_first(&cfqq->sort_list)); 2639 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); 2640 cfqg = cfqq->cfqg; 2641 2642 if (unlikely(cfqd->active_queue == cfqq)) { 2643 __cfq_slice_expired(cfqd, cfqq, 0); 2644 cfq_schedule_dispatch(cfqd); 2645 } 2646 2647 BUG_ON(cfq_cfqq_on_rr(cfqq)); 2648 kmem_cache_free(cfq_pool, cfqq); 2649 cfq_put_cfqg(cfqg); 2650} 2651 2652static void cfq_put_cooperator(struct cfq_queue *cfqq) 2653{ 2654 struct cfq_queue *__cfqq, *next; 2655 2656 /* 2657 * If this queue was scheduled to merge with another queue, be 2658 * sure to drop the reference taken on that queue (and others in 2659 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. 2660 */ 2661 __cfqq = cfqq->new_cfqq; 2662 while (__cfqq) { 2663 if (__cfqq == cfqq) { 2664 WARN(1, "cfqq->new_cfqq loop detected\n"); 2665 break; 2666 } 2667 next = __cfqq->new_cfqq; 2668 cfq_put_queue(__cfqq); 2669 __cfqq = next; 2670 } 2671} 2672 2673static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) 2674{ 2675 if (unlikely(cfqq == cfqd->active_queue)) { 2676 __cfq_slice_expired(cfqd, cfqq, 0); 2677 cfq_schedule_dispatch(cfqd); 2678 } 2679 2680 cfq_put_cooperator(cfqq); 2681 2682 cfq_put_queue(cfqq); 2683} 2684 2685static void cfq_init_icq(struct io_cq *icq) 2686{ 2687 struct cfq_io_cq *cic = icq_to_cic(icq); 2688 2689 cic->ttime.last_end_request = jiffies; 2690} 2691 2692static void cfq_exit_icq(struct io_cq *icq) 2693{ 2694 struct cfq_io_cq *cic = icq_to_cic(icq); 2695 struct cfq_data *cfqd = cic_to_cfqd(cic); 2696 2697 if (cic->cfqq[BLK_RW_ASYNC]) { 2698 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); 2699 cic->cfqq[BLK_RW_ASYNC] = NULL; 2700 } 2701 2702 if (cic->cfqq[BLK_RW_SYNC]) { 2703 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); 2704 cic->cfqq[BLK_RW_SYNC] = NULL; 2705 } 2706} 2707 2708static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) 2709{ 2710 struct task_struct *tsk = current; 2711 int ioprio_class; 2712 2713 if (!cfq_cfqq_prio_changed(cfqq)) 2714 return; 2715 2716 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); 2717 switch (ioprio_class) { 2718 default: 2719 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); 2720 case IOPRIO_CLASS_NONE: 2721 /* 2722 * no prio set, inherit CPU scheduling settings 2723 */ 2724 cfqq->ioprio = task_nice_ioprio(tsk); 2725 cfqq->ioprio_class = task_nice_ioclass(tsk); 2726 break; 2727 case IOPRIO_CLASS_RT: 2728 cfqq->ioprio = task_ioprio(ioc); 2729 cfqq->ioprio_class = IOPRIO_CLASS_RT; 2730 break; 2731 case IOPRIO_CLASS_BE: 2732 cfqq->ioprio = task_ioprio(ioc); 2733 cfqq->ioprio_class = IOPRIO_CLASS_BE; 2734 break; 2735 case IOPRIO_CLASS_IDLE: 2736 cfqq->ioprio_class = IOPRIO_CLASS_IDLE; 2737 cfqq->ioprio = 7; 2738 cfq_clear_cfqq_idle_window(cfqq); 2739 break; 2740 } 2741 2742 /* 2743 * keep track of original prio settings in case we have to temporarily 2744 * elevate the priority of this queue 2745 */ 2746 cfqq->org_ioprio = cfqq->ioprio; 2747 cfq_clear_cfqq_prio_changed(cfqq); 2748} 2749 2750static void changed_ioprio(struct cfq_io_cq *cic) 2751{ 2752 struct cfq_data *cfqd = cic_to_cfqd(cic); 2753 struct cfq_queue *cfqq; 2754 2755 if (unlikely(!cfqd)) 2756 return; 2757 2758 cfqq = cic->cfqq[BLK_RW_ASYNC]; 2759 if (cfqq) { 2760 struct cfq_queue *new_cfqq; 2761 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->icq.ioc, 2762 GFP_ATOMIC); 2763 if (new_cfqq) { 2764 cic->cfqq[BLK_RW_ASYNC] = new_cfqq; 2765 cfq_put_queue(cfqq); 2766 } 2767 } 2768 2769 cfqq = cic->cfqq[BLK_RW_SYNC]; 2770 if (cfqq) 2771 cfq_mark_cfqq_prio_changed(cfqq); 2772} 2773 2774static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, 2775 pid_t pid, bool is_sync) 2776{ 2777 RB_CLEAR_NODE(&cfqq->rb_node); 2778 RB_CLEAR_NODE(&cfqq->p_node); 2779 INIT_LIST_HEAD(&cfqq->fifo); 2780 2781 cfqq->ref = 0; 2782 cfqq->cfqd = cfqd; 2783 2784 cfq_mark_cfqq_prio_changed(cfqq); 2785 2786 if (is_sync) { 2787 if (!cfq_class_idle(cfqq)) 2788 cfq_mark_cfqq_idle_window(cfqq); 2789 cfq_mark_cfqq_sync(cfqq); 2790 } 2791 cfqq->pid = pid; 2792} 2793 2794#ifdef CONFIG_CFQ_GROUP_IOSCHED 2795static void changed_cgroup(struct cfq_io_cq *cic) 2796{ 2797 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1); 2798 struct cfq_data *cfqd = cic_to_cfqd(cic); 2799 struct request_queue *q; 2800 2801 if (unlikely(!cfqd)) 2802 return; 2803 2804 q = cfqd->queue; 2805 2806 if (sync_cfqq) { 2807 /* 2808 * Drop reference to sync queue. A new sync queue will be 2809 * assigned in new group upon arrival of a fresh request. 2810 */ 2811 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup"); 2812 cic_set_cfqq(cic, NULL, 1); 2813 cfq_put_queue(sync_cfqq); 2814 } 2815} 2816#endif /* CONFIG_CFQ_GROUP_IOSCHED */ 2817 2818static struct cfq_queue * 2819cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, 2820 struct io_context *ioc, gfp_t gfp_mask) 2821{ 2822 struct blkio_cgroup *blkcg; 2823 struct cfq_queue *cfqq, *new_cfqq = NULL; 2824 struct cfq_io_cq *cic; 2825 struct cfq_group *cfqg; 2826 2827retry: 2828 rcu_read_lock(); 2829 2830 blkcg = task_blkio_cgroup(current); 2831 2832 cfqg = cfq_lookup_create_cfqg(cfqd, blkcg); 2833 2834 cic = cfq_cic_lookup(cfqd, ioc); 2835 /* cic always exists here */ 2836 cfqq = cic_to_cfqq(cic, is_sync); 2837 2838 /* 2839 * Always try a new alloc if we fell back to the OOM cfqq 2840 * originally, since it should just be a temporary situation. 2841 */ 2842 if (!cfqq || cfqq == &cfqd->oom_cfqq) { 2843 cfqq = NULL; 2844 if (new_cfqq) { 2845 cfqq = new_cfqq; 2846 new_cfqq = NULL; 2847 } else if (gfp_mask & __GFP_WAIT) { 2848 rcu_read_unlock(); 2849 spin_unlock_irq(cfqd->queue->queue_lock); 2850 new_cfqq = kmem_cache_alloc_node(cfq_pool, 2851 gfp_mask | __GFP_ZERO, 2852 cfqd->queue->node); 2853 spin_lock_irq(cfqd->queue->queue_lock); 2854 if (new_cfqq) 2855 goto retry; 2856 } else { 2857 cfqq = kmem_cache_alloc_node(cfq_pool, 2858 gfp_mask | __GFP_ZERO, 2859 cfqd->queue->node); 2860 } 2861 2862 if (cfqq) { 2863 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); 2864 cfq_init_prio_data(cfqq, ioc); 2865 cfq_link_cfqq_cfqg(cfqq, cfqg); 2866 cfq_log_cfqq(cfqd, cfqq, "alloced"); 2867 } else 2868 cfqq = &cfqd->oom_cfqq; 2869 } 2870 2871 if (new_cfqq) 2872 kmem_cache_free(cfq_pool, new_cfqq); 2873 2874 rcu_read_unlock(); 2875 return cfqq; 2876} 2877 2878static struct cfq_queue ** 2879cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) 2880{ 2881 switch (ioprio_class) { 2882 case IOPRIO_CLASS_RT: 2883 return &cfqd->async_cfqq[0][ioprio]; 2884 case IOPRIO_CLASS_BE: 2885 return &cfqd->async_cfqq[1][ioprio]; 2886 case IOPRIO_CLASS_IDLE: 2887 return &cfqd->async_idle_cfqq; 2888 default: 2889 BUG(); 2890 } 2891} 2892 2893static struct cfq_queue * 2894cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc, 2895 gfp_t gfp_mask) 2896{ 2897 const int ioprio = task_ioprio(ioc); 2898 const int ioprio_class = task_ioprio_class(ioc); 2899 struct cfq_queue **async_cfqq = NULL; 2900 struct cfq_queue *cfqq = NULL; 2901 2902 if (!is_sync) { 2903 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); 2904 cfqq = *async_cfqq; 2905 } 2906 2907 if (!cfqq) 2908 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); 2909 2910 /* 2911 * pin the queue now that it's allocated, scheduler exit will prune it 2912 */ 2913 if (!is_sync && !(*async_cfqq)) { 2914 cfqq->ref++; 2915 *async_cfqq = cfqq; 2916 } 2917 2918 cfqq->ref++; 2919 return cfqq; 2920} 2921 2922static void 2923__cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle) 2924{ 2925 unsigned long elapsed = jiffies - ttime->last_end_request; 2926 elapsed = min(elapsed, 2UL * slice_idle); 2927 2928 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8; 2929 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8; 2930 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples; 2931} 2932 2933static void 2934cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq, 2935 struct cfq_io_cq *cic) 2936{ 2937 if (cfq_cfqq_sync(cfqq)) { 2938 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle); 2939 __cfq_update_io_thinktime(&cfqq->service_tree->ttime, 2940 cfqd->cfq_slice_idle); 2941 } 2942#ifdef CONFIG_CFQ_GROUP_IOSCHED 2943 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle); 2944#endif 2945} 2946 2947static void 2948cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, 2949 struct request *rq) 2950{ 2951 sector_t sdist = 0; 2952 sector_t n_sec = blk_rq_sectors(rq); 2953 if (cfqq->last_request_pos) { 2954 if (cfqq->last_request_pos < blk_rq_pos(rq)) 2955 sdist = blk_rq_pos(rq) - cfqq->last_request_pos; 2956 else 2957 sdist = cfqq->last_request_pos - blk_rq_pos(rq); 2958 } 2959 2960 cfqq->seek_history <<= 1; 2961 if (blk_queue_nonrot(cfqd->queue)) 2962 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT); 2963 else 2964 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR); 2965} 2966 2967/* 2968 * Disable idle window if the process thinks too long or seeks so much that 2969 * it doesn't matter 2970 */ 2971static void 2972cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, 2973 struct cfq_io_cq *cic) 2974{ 2975 int old_idle, enable_idle; 2976 2977 /* 2978 * Don't idle for async or idle io prio class 2979 */ 2980 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) 2981 return; 2982 2983 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); 2984 2985 if (cfqq->queued[0] + cfqq->queued[1] >= 4) 2986 cfq_mark_cfqq_deep(cfqq); 2987 2988 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE)) 2989 enable_idle = 0; 2990 else if (!atomic_read(&cic->icq.ioc->nr_tasks) || 2991 !cfqd->cfq_slice_idle || 2992 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq))) 2993 enable_idle = 0; 2994 else if (sample_valid(cic->ttime.ttime_samples)) { 2995 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle) 2996 enable_idle = 0; 2997 else 2998 enable_idle = 1; 2999 } 3000 3001 if (old_idle != enable_idle) { 3002 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); 3003 if (enable_idle) 3004 cfq_mark_cfqq_idle_window(cfqq); 3005 else 3006 cfq_clear_cfqq_idle_window(cfqq); 3007 } 3008} 3009 3010/* 3011 * Check if new_cfqq should preempt the currently active queue. Return 0 for 3012 * no or if we aren't sure, a 1 will cause a preempt. 3013 */ 3014static bool 3015cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, 3016 struct request *rq) 3017{ 3018 struct cfq_queue *cfqq; 3019 3020 cfqq = cfqd->active_queue; 3021 if (!cfqq) 3022 return false; 3023 3024 if (cfq_class_idle(new_cfqq)) 3025 return false; 3026 3027 if (cfq_class_idle(cfqq)) 3028 return true; 3029 3030 /* 3031 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice. 3032 */ 3033 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq)) 3034 return false; 3035 3036 /* 3037 * if the new request is sync, but the currently running queue is 3038 * not, let the sync request have priority. 3039 */ 3040 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) 3041 return true; 3042 3043 if (new_cfqq->cfqg != cfqq->cfqg) 3044 return false; 3045 3046 if (cfq_slice_used(cfqq)) 3047 return true; 3048 3049 /* Allow preemption only if we are idling on sync-noidle tree */ 3050 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD && 3051 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD && 3052 new_cfqq->service_tree->count == 2 && 3053 RB_EMPTY_ROOT(&cfqq->sort_list)) 3054 return true; 3055 3056 /* 3057 * So both queues are sync. Let the new request get disk time if 3058 * it's a metadata request and the current queue is doing regular IO. 3059 */ 3060 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending) 3061 return true; 3062 3063 /* 3064 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. 3065 */ 3066 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) 3067 return true; 3068 3069 /* An idle queue should not be idle now for some reason */ 3070 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq)) 3071 return true; 3072 3073 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) 3074 return false; 3075 3076 /* 3077 * if this request is as-good as one we would expect from the 3078 * current cfqq, let it preempt 3079 */ 3080 if (cfq_rq_close(cfqd, cfqq, rq)) 3081 return true; 3082 3083 return false; 3084} 3085 3086/* 3087 * cfqq preempts the active queue. if we allowed preempt with no slice left, 3088 * let it have half of its nominal slice. 3089 */ 3090static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) 3091{ 3092 enum wl_type_t old_type = cfqq_type(cfqd->active_queue); 3093 3094 cfq_log_cfqq(cfqd, cfqq, "preempt"); 3095 cfq_slice_expired(cfqd, 1); 3096 3097 /* 3098 * workload type is changed, don't save slice, otherwise preempt 3099 * doesn't happen 3100 */ 3101 if (old_type != cfqq_type(cfqq)) 3102 cfqq->cfqg->saved_workload_slice = 0; 3103 3104 /* 3105 * Put the new queue at the front of the of the current list, 3106 * so we know that it will be selected next. 3107 */ 3108 BUG_ON(!cfq_cfqq_on_rr(cfqq)); 3109 3110 cfq_service_tree_add(cfqd, cfqq, 1); 3111 3112 cfqq->slice_end = 0; 3113 cfq_mark_cfqq_slice_new(cfqq); 3114} 3115 3116/* 3117 * Called when a new fs request (rq) is added (to cfqq). Check if there's 3118 * something we should do about it 3119 */ 3120static void 3121cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, 3122 struct request *rq) 3123{ 3124 struct cfq_io_cq *cic = RQ_CIC(rq); 3125 3126 cfqd->rq_queued++; 3127 if (rq->cmd_flags & REQ_PRIO) 3128 cfqq->prio_pending++; 3129 3130 cfq_update_io_thinktime(cfqd, cfqq, cic); 3131 cfq_update_io_seektime(cfqd, cfqq, rq); 3132 cfq_update_idle_window(cfqd, cfqq, cic); 3133 3134 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); 3135 3136 if (cfqq == cfqd->active_queue) { 3137 /* 3138 * Remember that we saw a request from this process, but 3139 * don't start queuing just yet. Otherwise we risk seeing lots 3140 * of tiny requests, because we disrupt the normal plugging 3141 * and merging. If the request is already larger than a single 3142 * page, let it rip immediately. For that case we assume that 3143 * merging is already done. Ditto for a busy system that 3144 * has other work pending, don't risk delaying until the 3145 * idle timer unplug to continue working. 3146 */ 3147 if (cfq_cfqq_wait_request(cfqq)) { 3148 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || 3149 cfqd->busy_queues > 1) { 3150 cfq_del_timer(cfqd, cfqq); 3151 cfq_clear_cfqq_wait_request(cfqq); 3152 __blk_run_queue(cfqd->queue); 3153 } else { 3154 cfq_blkiocg_update_idle_time_stats( 3155 &cfqq->cfqg->blkg); 3156 cfq_mark_cfqq_must_dispatch(cfqq); 3157 } 3158 } 3159 } else if (cfq_should_preempt(cfqd, cfqq, rq)) { 3160 /* 3161 * not the active queue - expire current slice if it is 3162 * idle and has expired it's mean thinktime or this new queue 3163 * has some old slice time left and is of higher priority or 3164 * this new queue is RT and the current one is BE 3165 */ 3166 cfq_preempt_queue(cfqd, cfqq); 3167 __blk_run_queue(cfqd->queue); 3168 } 3169} 3170 3171static void cfq_insert_request(struct request_queue *q, struct request *rq) 3172{ 3173 struct cfq_data *cfqd = q->elevator->elevator_data; 3174 struct cfq_queue *cfqq = RQ_CFQQ(rq); 3175 3176 cfq_log_cfqq(cfqd, cfqq, "insert_request"); 3177 cfq_init_prio_data(cfqq, RQ_CIC(rq)->icq.ioc); 3178 3179 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); 3180 list_add_tail(&rq->queuelist, &cfqq->fifo); 3181 cfq_add_rq_rb(rq); 3182 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg, 3183 &cfqd->serving_group->blkg, rq_data_dir(rq), 3184 rq_is_sync(rq)); 3185 cfq_rq_enqueued(cfqd, cfqq, rq); 3186} 3187 3188/* 3189 * Update hw_tag based on peak queue depth over 50 samples under 3190 * sufficient load. 3191 */ 3192static void cfq_update_hw_tag(struct cfq_data *cfqd) 3193{ 3194 struct cfq_queue *cfqq = cfqd->active_queue; 3195 3196 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth) 3197 cfqd->hw_tag_est_depth = cfqd->rq_in_driver; 3198 3199 if (cfqd->hw_tag == 1) 3200 return; 3201 3202 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && 3203 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN) 3204 return; 3205 3206 /* 3207 * If active queue hasn't enough requests and can idle, cfq might not 3208 * dispatch sufficient requests to hardware. Don't zero hw_tag in this 3209 * case 3210 */ 3211 if (cfqq && cfq_cfqq_idle_window(cfqq) && 3212 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < 3213 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN) 3214 return; 3215 3216 if (cfqd->hw_tag_samples++ < 50) 3217 return; 3218 3219 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN) 3220 cfqd->hw_tag = 1; 3221 else 3222 cfqd->hw_tag = 0; 3223} 3224 3225static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq) 3226{ 3227 struct cfq_io_cq *cic = cfqd->active_cic; 3228 3229 /* If the queue already has requests, don't wait */ 3230 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) 3231 return false; 3232 3233 /* If there are other queues in the group, don't wait */ 3234 if (cfqq->cfqg->nr_cfqq > 1) 3235 return false; 3236 3237 /* the only queue in the group, but think time is big */ 3238 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) 3239 return false; 3240 3241 if (cfq_slice_used(cfqq)) 3242 return true; 3243 3244 /* if slice left is less than think time, wait busy */ 3245 if (cic && sample_valid(cic->ttime.ttime_samples) 3246 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) 3247 return true; 3248 3249 /* 3250 * If think times is less than a jiffy than ttime_mean=0 and above 3251 * will not be true. It might happen that slice has not expired yet 3252 * but will expire soon (4-5 ns) during select_queue(). To cover the 3253 * case where think time is less than a jiffy, mark the queue wait 3254 * busy if only 1 jiffy is left in the slice. 3255 */ 3256 if (cfqq->slice_end - jiffies == 1) 3257 return true; 3258 3259 return false; 3260} 3261 3262static void cfq_completed_request(struct request_queue *q, struct request *rq) 3263{ 3264 struct cfq_queue *cfqq = RQ_CFQQ(rq); 3265 struct cfq_data *cfqd = cfqq->cfqd; 3266 const int sync = rq_is_sync(rq); 3267 unsigned long now; 3268 3269 now = jiffies; 3270 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", 3271 !!(rq->cmd_flags & REQ_NOIDLE)); 3272 3273 cfq_update_hw_tag(cfqd); 3274 3275 WARN_ON(!cfqd->rq_in_driver); 3276 WARN_ON(!cfqq->dispatched); 3277 cfqd->rq_in_driver--; 3278 cfqq->dispatched--; 3279 (RQ_CFQG(rq))->dispatched--; 3280 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg, 3281 rq_start_time_ns(rq), rq_io_start_time_ns(rq), 3282 rq_data_dir(rq), rq_is_sync(rq)); 3283 3284 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--; 3285 3286 if (sync) { 3287 struct cfq_rb_root *service_tree; 3288 3289 RQ_CIC(rq)->ttime.last_end_request = now; 3290 3291 if (cfq_cfqq_on_rr(cfqq)) 3292 service_tree = cfqq->service_tree; 3293 else 3294 service_tree = service_tree_for(cfqq->cfqg, 3295 cfqq_prio(cfqq), cfqq_type(cfqq)); 3296 service_tree->ttime.last_end_request = now; 3297 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now)) 3298 cfqd->last_delayed_sync = now; 3299 } 3300 3301#ifdef CONFIG_CFQ_GROUP_IOSCHED 3302 cfqq->cfqg->ttime.last_end_request = now; 3303#endif 3304 3305 /* 3306 * If this is the active queue, check if it needs to be expired, 3307 * or if we want to idle in case it has no pending requests. 3308 */ 3309 if (cfqd->active_queue == cfqq) { 3310 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); 3311 3312 if (cfq_cfqq_slice_new(cfqq)) { 3313 cfq_set_prio_slice(cfqd, cfqq); 3314 cfq_clear_cfqq_slice_new(cfqq); 3315 } 3316 3317 /* 3318 * Should we wait for next request to come in before we expire 3319 * the queue. 3320 */ 3321 if (cfq_should_wait_busy(cfqd, cfqq)) { 3322 unsigned long extend_sl = cfqd->cfq_slice_idle; 3323 if (!cfqd->cfq_slice_idle) 3324 extend_sl = cfqd->cfq_group_idle; 3325 cfqq->slice_end = jiffies + extend_sl; 3326 cfq_mark_cfqq_wait_busy(cfqq); 3327 cfq_log_cfqq(cfqd, cfqq, "will busy wait"); 3328 } 3329 3330 /* 3331 * Idling is not enabled on: 3332 * - expired queues 3333 * - idle-priority queues 3334 * - async queues 3335 * - queues with still some requests queued 3336 * - when there is a close cooperator 3337 */ 3338 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) 3339 cfq_slice_expired(cfqd, 1); 3340 else if (sync && cfqq_empty && 3341 !cfq_close_cooperator(cfqd, cfqq)) { 3342 cfq_arm_slice_timer(cfqd); 3343 } 3344 } 3345 3346 if (!cfqd->rq_in_driver) 3347 cfq_schedule_dispatch(cfqd); 3348} 3349 3350static inline int __cfq_may_queue(struct cfq_queue *cfqq) 3351{ 3352 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { 3353 cfq_mark_cfqq_must_alloc_slice(cfqq); 3354 return ELV_MQUEUE_MUST; 3355 } 3356 3357 return ELV_MQUEUE_MAY; 3358} 3359 3360static int cfq_may_queue(struct request_queue *q, int rw) 3361{ 3362 struct cfq_data *cfqd = q->elevator->elevator_data; 3363 struct task_struct *tsk = current; 3364 struct cfq_io_cq *cic; 3365 struct cfq_queue *cfqq; 3366 3367 /* 3368 * don't force setup of a queue from here, as a call to may_queue 3369 * does not necessarily imply that a request actually will be queued. 3370 * so just lookup a possibly existing queue, or return 'may queue' 3371 * if that fails 3372 */ 3373 cic = cfq_cic_lookup(cfqd, tsk->io_context); 3374 if (!cic) 3375 return ELV_MQUEUE_MAY; 3376 3377 cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); 3378 if (cfqq) { 3379 cfq_init_prio_data(cfqq, cic->icq.ioc); 3380 3381 return __cfq_may_queue(cfqq); 3382 } 3383 3384 return ELV_MQUEUE_MAY; 3385} 3386 3387/* 3388 * queue lock held here 3389 */ 3390static void cfq_put_request(struct request *rq) 3391{ 3392 struct cfq_queue *cfqq = RQ_CFQQ(rq); 3393 3394 if (cfqq) { 3395 const int rw = rq_data_dir(rq); 3396 3397 BUG_ON(!cfqq->allocated[rw]); 3398 cfqq->allocated[rw]--; 3399 3400 /* Put down rq reference on cfqg */ 3401 cfq_put_cfqg(RQ_CFQG(rq)); 3402 rq->elv.priv[0] = NULL; 3403 rq->elv.priv[1] = NULL; 3404 3405 cfq_put_queue(cfqq); 3406 } 3407} 3408 3409static struct cfq_queue * 3410cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic, 3411 struct cfq_queue *cfqq) 3412{ 3413 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); 3414 cic_set_cfqq(cic, cfqq->new_cfqq, 1); 3415 cfq_mark_cfqq_coop(cfqq->new_cfqq); 3416 cfq_put_queue(cfqq); 3417 return cic_to_cfqq(cic, 1); 3418} 3419 3420/* 3421 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this 3422 * was the last process referring to said cfqq. 3423 */ 3424static struct cfq_queue * 3425split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq) 3426{ 3427 if (cfqq_process_refs(cfqq) == 1) { 3428 cfqq->pid = current->pid; 3429 cfq_clear_cfqq_coop(cfqq); 3430 cfq_clear_cfqq_split_coop(cfqq); 3431 return cfqq; 3432 } 3433 3434 cic_set_cfqq(cic, NULL, 1); 3435 3436 cfq_put_cooperator(cfqq); 3437 3438 cfq_put_queue(cfqq); 3439 return NULL; 3440} 3441/* 3442 * Allocate cfq data structures associated with this request. 3443 */ 3444static int 3445cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) 3446{ 3447 struct cfq_data *cfqd = q->elevator->elevator_data; 3448 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq); 3449 const int rw = rq_data_dir(rq); 3450 const bool is_sync = rq_is_sync(rq); 3451 struct cfq_queue *cfqq; 3452 unsigned int changed; 3453 3454 might_sleep_if(gfp_mask & __GFP_WAIT); 3455 3456 spin_lock_irq(q->queue_lock); 3457 3458 /* handle changed notifications */ 3459 changed = icq_get_changed(&cic->icq); 3460 if (unlikely(changed & ICQ_IOPRIO_CHANGED)) 3461 changed_ioprio(cic); 3462#ifdef CONFIG_CFQ_GROUP_IOSCHED 3463 if (unlikely(changed & ICQ_CGROUP_CHANGED)) 3464 changed_cgroup(cic); 3465#endif 3466 3467new_queue: 3468 cfqq = cic_to_cfqq(cic, is_sync); 3469 if (!cfqq || cfqq == &cfqd->oom_cfqq) { 3470 cfqq = cfq_get_queue(cfqd, is_sync, cic->icq.ioc, gfp_mask); 3471 cic_set_cfqq(cic, cfqq, is_sync); 3472 } else { 3473 /* 3474 * If the queue was seeky for too long, break it apart. 3475 */ 3476 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) { 3477 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); 3478 cfqq = split_cfqq(cic, cfqq); 3479 if (!cfqq) 3480 goto new_queue; 3481 } 3482 3483 /* 3484 * Check to see if this queue is scheduled to merge with 3485 * another, closely cooperating queue. The merging of 3486 * queues happens here as it must be done in process context. 3487 * The reference on new_cfqq was taken in merge_cfqqs. 3488 */ 3489 if (cfqq->new_cfqq) 3490 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); 3491 } 3492 3493 cfqq->allocated[rw]++; 3494 3495 cfqq->ref++; 3496 rq->elv.priv[0] = cfqq; 3497 rq->elv.priv[1] = cfq_ref_get_cfqg(cfqq->cfqg); 3498 spin_unlock_irq(q->queue_lock); 3499 return 0; 3500} 3501 3502static void cfq_kick_queue(struct work_struct *work) 3503{ 3504 struct cfq_data *cfqd = 3505 container_of(work, struct cfq_data, unplug_work); 3506 struct request_queue *q = cfqd->queue; 3507 3508 spin_lock_irq(q->queue_lock); 3509 __blk_run_queue(cfqd->queue); 3510 spin_unlock_irq(q->queue_lock); 3511} 3512 3513/* 3514 * Timer running if the active_queue is currently idling inside its time slice 3515 */ 3516static void cfq_idle_slice_timer(unsigned long data) 3517{ 3518 struct cfq_data *cfqd = (struct cfq_data *) data; 3519 struct cfq_queue *cfqq; 3520 unsigned long flags; 3521 int timed_out = 1; 3522 3523 cfq_log(cfqd, "idle timer fired"); 3524 3525 spin_lock_irqsave(cfqd->queue->queue_lock, flags); 3526 3527 cfqq = cfqd->active_queue; 3528 if (cfqq) { 3529 timed_out = 0; 3530 3531 /* 3532 * We saw a request before the queue expired, let it through 3533 */ 3534 if (cfq_cfqq_must_dispatch(cfqq)) 3535 goto out_kick; 3536 3537 /* 3538 * expired 3539 */ 3540 if (cfq_slice_used(cfqq)) 3541 goto expire; 3542 3543 /* 3544 * only expire and reinvoke request handler, if there are 3545 * other queues with pending requests 3546 */ 3547 if (!cfqd->busy_queues) 3548 goto out_cont; 3549 3550 /* 3551 * not expired and it has a request pending, let it dispatch 3552 */ 3553 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) 3554 goto out_kick; 3555 3556 /* 3557 * Queue depth flag is reset only when the idle didn't succeed 3558 */ 3559 cfq_clear_cfqq_deep(cfqq); 3560 } 3561expire: 3562 cfq_slice_expired(cfqd, timed_out); 3563out_kick: 3564 cfq_schedule_dispatch(cfqd); 3565out_cont: 3566 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); 3567} 3568 3569static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) 3570{ 3571 del_timer_sync(&cfqd->idle_slice_timer); 3572 cancel_work_sync(&cfqd->unplug_work); 3573} 3574 3575static void cfq_put_async_queues(struct cfq_data *cfqd) 3576{ 3577 int i; 3578 3579 for (i = 0; i < IOPRIO_BE_NR; i++) { 3580 if (cfqd->async_cfqq[0][i]) 3581 cfq_put_queue(cfqd->async_cfqq[0][i]); 3582 if (cfqd->async_cfqq[1][i]) 3583 cfq_put_queue(cfqd->async_cfqq[1][i]); 3584 } 3585 3586 if (cfqd->async_idle_cfqq) 3587 cfq_put_queue(cfqd->async_idle_cfqq); 3588} 3589 3590static void cfq_exit_queue(struct elevator_queue *e) 3591{ 3592 struct cfq_data *cfqd = e->elevator_data; 3593 struct request_queue *q = cfqd->queue; 3594 bool wait = false; 3595 3596 cfq_shutdown_timer_wq(cfqd); 3597 3598 spin_lock_irq(q->queue_lock); 3599 3600 if (cfqd->active_queue) 3601 __cfq_slice_expired(cfqd, cfqd->active_queue, 0); 3602 3603 cfq_put_async_queues(cfqd); 3604 cfq_release_cfq_groups(cfqd); 3605 3606 /* 3607 * If there are groups which we could not unlink from blkcg list, 3608 * wait for a rcu period for them to be freed. 3609 */ 3610 if (cfqd->nr_blkcg_linked_grps) 3611 wait = true; 3612 3613 spin_unlock_irq(q->queue_lock); 3614 3615 cfq_shutdown_timer_wq(cfqd); 3616 3617 /* 3618 * Wait for cfqg->blkg->key accessors to exit their grace periods. 3619 * Do this wait only if there are other unlinked groups out 3620 * there. This can happen if cgroup deletion path claimed the 3621 * responsibility of cleaning up a group before queue cleanup code 3622 * get to the group. 3623 * 3624 * Do not call synchronize_rcu() unconditionally as there are drivers 3625 * which create/delete request queue hundreds of times during scan/boot 3626 * and synchronize_rcu() can take significant time and slow down boot. 3627 */ 3628 if (wait) 3629 synchronize_rcu(); 3630 3631#ifndef CONFIG_CFQ_GROUP_IOSCHED 3632 kfree(cfqd->root_group); 3633#endif 3634 kfree(cfqd); 3635} 3636 3637static int cfq_init_queue(struct request_queue *q) 3638{ 3639 struct cfq_data *cfqd; 3640 struct blkio_group *blkg __maybe_unused; 3641 int i; 3642 3643 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); 3644 if (!cfqd) 3645 return -ENOMEM; 3646 3647 cfqd->queue = q; 3648 q->elevator->elevator_data = cfqd; 3649 3650 /* Init root service tree */ 3651 cfqd->grp_service_tree = CFQ_RB_ROOT; 3652 3653 /* Init root group and prefer root group over other groups by default */ 3654#ifdef CONFIG_CFQ_GROUP_IOSCHED 3655 rcu_read_lock(); 3656 spin_lock_irq(q->queue_lock); 3657 3658 blkg = blkg_lookup_create(&blkio_root_cgroup, q, BLKIO_POLICY_PROP, 3659 true); 3660 if (!IS_ERR(blkg)) 3661 cfqd->root_group = cfqg_of_blkg(blkg); 3662 3663 spin_unlock_irq(q->queue_lock); 3664 rcu_read_unlock(); 3665#else 3666 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group), 3667 GFP_KERNEL, cfqd->queue->node); 3668 if (cfqd->root_group) 3669 cfq_init_cfqg_base(cfqd->root_group); 3670#endif 3671 if (!cfqd->root_group) { 3672 kfree(cfqd); 3673 return -ENOMEM; 3674 } 3675 3676 cfqd->root_group->weight = 2*BLKIO_WEIGHT_DEFAULT; 3677 3678 /* 3679 * Not strictly needed (since RB_ROOT just clears the node and we 3680 * zeroed cfqd on alloc), but better be safe in case someone decides 3681 * to add magic to the rb code 3682 */ 3683 for (i = 0; i < CFQ_PRIO_LISTS; i++) 3684 cfqd->prio_trees[i] = RB_ROOT; 3685 3686 /* 3687 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. 3688 * Grab a permanent reference to it, so that the normal code flow 3689 * will not attempt to free it. oom_cfqq is linked to root_group 3690 * but shouldn't hold a reference as it'll never be unlinked. Lose 3691 * the reference from linking right away. 3692 */ 3693 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); 3694 cfqd->oom_cfqq.ref++; 3695 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group); 3696 cfq_put_cfqg(cfqd->root_group); 3697 3698 init_timer(&cfqd->idle_slice_timer); 3699 cfqd->idle_slice_timer.function = cfq_idle_slice_timer; 3700 cfqd->idle_slice_timer.data = (unsigned long) cfqd; 3701 3702 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); 3703 3704 cfqd->cfq_quantum = cfq_quantum; 3705 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; 3706 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; 3707 cfqd->cfq_back_max = cfq_back_max; 3708 cfqd->cfq_back_penalty = cfq_back_penalty; 3709 cfqd->cfq_slice[0] = cfq_slice_async; 3710 cfqd->cfq_slice[1] = cfq_slice_sync; 3711 cfqd->cfq_slice_async_rq = cfq_slice_async_rq; 3712 cfqd->cfq_slice_idle = cfq_slice_idle; 3713 cfqd->cfq_group_idle = cfq_group_idle; 3714 cfqd->cfq_latency = 1; 3715 cfqd->hw_tag = -1; 3716 /* 3717 * we optimistically start assuming sync ops weren't delayed in last 3718 * second, in order to have larger depth for async operations. 3719 */ 3720 cfqd->last_delayed_sync = jiffies - HZ; 3721 return 0; 3722} 3723 3724/* 3725 * sysfs parts below --> 3726 */ 3727static ssize_t 3728cfq_var_show(unsigned int var, char *page) 3729{ 3730 return sprintf(page, "%d\n", var); 3731} 3732 3733static ssize_t 3734cfq_var_store(unsigned int *var, const char *page, size_t count) 3735{ 3736 char *p = (char *) page; 3737 3738 *var = simple_strtoul(p, &p, 10); 3739 return count; 3740} 3741 3742#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ 3743static ssize_t __FUNC(struct elevator_queue *e, char *page) \ 3744{ \ 3745 struct cfq_data *cfqd = e->elevator_data; \ 3746 unsigned int __data = __VAR; \ 3747 if (__CONV) \ 3748 __data = jiffies_to_msecs(__data); \ 3749 return cfq_var_show(__data, (page)); \ 3750} 3751SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); 3752SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); 3753SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); 3754SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); 3755SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); 3756SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); 3757SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1); 3758SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); 3759SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); 3760SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); 3761SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); 3762#undef SHOW_FUNCTION 3763 3764#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ 3765static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ 3766{ \ 3767 struct cfq_data *cfqd = e->elevator_data; \ 3768 unsigned int __data; \ 3769 int ret = cfq_var_store(&__data, (page), count); \ 3770 if (__data < (MIN)) \ 3771 __data = (MIN); \ 3772 else if (__data > (MAX)) \ 3773 __data = (MAX); \ 3774 if (__CONV) \ 3775 *(__PTR) = msecs_to_jiffies(__data); \ 3776 else \ 3777 *(__PTR) = __data; \ 3778 return ret; \ 3779} 3780STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); 3781STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, 3782 UINT_MAX, 1); 3783STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, 3784 UINT_MAX, 1); 3785STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); 3786STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, 3787 UINT_MAX, 0); 3788STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); 3789STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1); 3790STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); 3791STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); 3792STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, 3793 UINT_MAX, 0); 3794STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); 3795#undef STORE_FUNCTION 3796 3797#define CFQ_ATTR(name) \ 3798 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) 3799 3800static struct elv_fs_entry cfq_attrs[] = { 3801 CFQ_ATTR(quantum), 3802 CFQ_ATTR(fifo_expire_sync), 3803 CFQ_ATTR(fifo_expire_async), 3804 CFQ_ATTR(back_seek_max), 3805 CFQ_ATTR(back_seek_penalty), 3806 CFQ_ATTR(slice_sync), 3807 CFQ_ATTR(slice_async), 3808 CFQ_ATTR(slice_async_rq), 3809 CFQ_ATTR(slice_idle), 3810 CFQ_ATTR(group_idle), 3811 CFQ_ATTR(low_latency), 3812 __ATTR_NULL 3813}; 3814 3815static struct elevator_type iosched_cfq = { 3816 .ops = { 3817 .elevator_merge_fn = cfq_merge, 3818 .elevator_merged_fn = cfq_merged_request, 3819 .elevator_merge_req_fn = cfq_merged_requests, 3820 .elevator_allow_merge_fn = cfq_allow_merge, 3821 .elevator_bio_merged_fn = cfq_bio_merged, 3822 .elevator_dispatch_fn = cfq_dispatch_requests, 3823 .elevator_add_req_fn = cfq_insert_request, 3824 .elevator_activate_req_fn = cfq_activate_request, 3825 .elevator_deactivate_req_fn = cfq_deactivate_request, 3826 .elevator_completed_req_fn = cfq_completed_request, 3827 .elevator_former_req_fn = elv_rb_former_request, 3828 .elevator_latter_req_fn = elv_rb_latter_request, 3829 .elevator_init_icq_fn = cfq_init_icq, 3830 .elevator_exit_icq_fn = cfq_exit_icq, 3831 .elevator_set_req_fn = cfq_set_request, 3832 .elevator_put_req_fn = cfq_put_request, 3833 .elevator_may_queue_fn = cfq_may_queue, 3834 .elevator_init_fn = cfq_init_queue, 3835 .elevator_exit_fn = cfq_exit_queue, 3836 }, 3837 .icq_size = sizeof(struct cfq_io_cq), 3838 .icq_align = __alignof__(struct cfq_io_cq), 3839 .elevator_attrs = cfq_attrs, 3840 .elevator_name = "cfq", 3841 .elevator_owner = THIS_MODULE, 3842}; 3843 3844#ifdef CONFIG_CFQ_GROUP_IOSCHED 3845static struct blkio_policy_type blkio_policy_cfq = { 3846 .ops = { 3847 .blkio_alloc_group_fn = cfq_alloc_blkio_group, 3848 .blkio_link_group_fn = cfq_link_blkio_group, 3849 .blkio_unlink_group_fn = cfq_unlink_blkio_group, 3850 .blkio_clear_queue_fn = cfq_clear_queue, 3851 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight, 3852 }, 3853 .plid = BLKIO_POLICY_PROP, 3854}; 3855#endif 3856 3857static int __init cfq_init(void) 3858{ 3859 int ret; 3860 3861 /* 3862 * could be 0 on HZ < 1000 setups 3863 */ 3864 if (!cfq_slice_async) 3865 cfq_slice_async = 1; 3866 if (!cfq_slice_idle) 3867 cfq_slice_idle = 1; 3868 3869#ifdef CONFIG_CFQ_GROUP_IOSCHED 3870 if (!cfq_group_idle) 3871 cfq_group_idle = 1; 3872#else 3873 cfq_group_idle = 0; 3874#endif 3875 cfq_pool = KMEM_CACHE(cfq_queue, 0); 3876 if (!cfq_pool) 3877 return -ENOMEM; 3878 3879 ret = elv_register(&iosched_cfq); 3880 if (ret) { 3881 kmem_cache_destroy(cfq_pool); 3882 return ret; 3883 } 3884 3885#ifdef CONFIG_CFQ_GROUP_IOSCHED 3886 blkio_policy_register(&blkio_policy_cfq); 3887#endif 3888 return 0; 3889} 3890 3891static void __exit cfq_exit(void) 3892{ 3893#ifdef CONFIG_CFQ_GROUP_IOSCHED 3894 blkio_policy_unregister(&blkio_policy_cfq); 3895#endif 3896 elv_unregister(&iosched_cfq); 3897 kmem_cache_destroy(cfq_pool); 3898} 3899 3900module_init(cfq_init); 3901module_exit(cfq_exit); 3902 3903MODULE_AUTHOR("Jens Axboe"); 3904MODULE_LICENSE("GPL"); 3905MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); 3906