cfq-iosched.c revision 72e06c255181537d0b3e1f657a9ed81655d745b1
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#ifdef CONFIG_CFQ_GROUP_IOSCHED 1016static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg) 1017{ 1018 if (blkg) 1019 return container_of(blkg, struct cfq_group, blkg); 1020 return NULL; 1021} 1022 1023static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg, 1024 unsigned int weight) 1025{ 1026 struct cfq_group *cfqg = cfqg_of_blkg(blkg); 1027 cfqg->new_weight = weight; 1028 cfqg->needs_update = true; 1029} 1030 1031static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd, 1032 struct cfq_group *cfqg, struct blkio_cgroup *blkcg) 1033{ 1034 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info; 1035 unsigned int major, minor; 1036 1037 /* 1038 * Add group onto cgroup list. It might happen that bdi->dev is 1039 * not initialized yet. Initialize this new group without major 1040 * and minor info and this info will be filled in once a new thread 1041 * comes for IO. 1042 */ 1043 if (bdi->dev) { 1044 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); 1045 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, 1046 (void *)cfqd, MKDEV(major, minor)); 1047 } else 1048 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, 1049 (void *)cfqd, 0); 1050 1051 cfqd->nr_blkcg_linked_grps++; 1052 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev); 1053 1054 /* Add group on cfqd list */ 1055 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list); 1056} 1057 1058/* 1059 * Should be called from sleepable context. No request queue lock as per 1060 * cpu stats are allocated dynamically and alloc_percpu needs to be called 1061 * from sleepable context. 1062 */ 1063static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd) 1064{ 1065 struct cfq_group *cfqg = NULL; 1066 int i, j, ret; 1067 struct cfq_rb_root *st; 1068 1069 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node); 1070 if (!cfqg) 1071 return NULL; 1072 1073 for_each_cfqg_st(cfqg, i, j, st) 1074 *st = CFQ_RB_ROOT; 1075 RB_CLEAR_NODE(&cfqg->rb_node); 1076 1077 cfqg->ttime.last_end_request = jiffies; 1078 1079 /* 1080 * Take the initial reference that will be released on destroy 1081 * This can be thought of a joint reference by cgroup and 1082 * elevator which will be dropped by either elevator exit 1083 * or cgroup deletion path depending on who is exiting first. 1084 */ 1085 cfqg->ref = 1; 1086 1087 ret = blkio_alloc_blkg_stats(&cfqg->blkg); 1088 if (ret) { 1089 kfree(cfqg); 1090 return NULL; 1091 } 1092 1093 return cfqg; 1094} 1095 1096static struct cfq_group * 1097cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg) 1098{ 1099 struct cfq_group *cfqg = NULL; 1100 void *key = cfqd; 1101 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info; 1102 unsigned int major, minor; 1103 1104 /* 1105 * This is the common case when there are no blkio cgroups. 1106 * Avoid lookup in this case 1107 */ 1108 if (blkcg == &blkio_root_cgroup) 1109 cfqg = &cfqd->root_group; 1110 else 1111 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key)); 1112 1113 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) { 1114 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); 1115 cfqg->blkg.dev = MKDEV(major, minor); 1116 } 1117 1118 return cfqg; 1119} 1120 1121/* 1122 * Search for the cfq group current task belongs to. request_queue lock must 1123 * be held. 1124 */ 1125static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd) 1126{ 1127 struct blkio_cgroup *blkcg; 1128 struct cfq_group *cfqg = NULL, *__cfqg = NULL; 1129 struct request_queue *q = cfqd->queue; 1130 1131 rcu_read_lock(); 1132 blkcg = task_blkio_cgroup(current); 1133 cfqg = cfq_find_cfqg(cfqd, blkcg); 1134 if (cfqg) { 1135 rcu_read_unlock(); 1136 return cfqg; 1137 } 1138 1139 /* 1140 * Need to allocate a group. Allocation of group also needs allocation 1141 * of per cpu stats which in-turn takes a mutex() and can block. Hence 1142 * we need to drop rcu lock and queue_lock before we call alloc. 1143 * 1144 * Not taking any queue reference here and assuming that queue is 1145 * around by the time we return. CFQ queue allocation code does 1146 * the same. It might be racy though. 1147 */ 1148 1149 rcu_read_unlock(); 1150 spin_unlock_irq(q->queue_lock); 1151 1152 cfqg = cfq_alloc_cfqg(cfqd); 1153 1154 spin_lock_irq(q->queue_lock); 1155 1156 rcu_read_lock(); 1157 blkcg = task_blkio_cgroup(current); 1158 1159 /* 1160 * If some other thread already allocated the group while we were 1161 * not holding queue lock, free up the group 1162 */ 1163 __cfqg = cfq_find_cfqg(cfqd, blkcg); 1164 1165 if (__cfqg) { 1166 kfree(cfqg); 1167 rcu_read_unlock(); 1168 return __cfqg; 1169 } 1170 1171 if (!cfqg) 1172 cfqg = &cfqd->root_group; 1173 1174 cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg); 1175 rcu_read_unlock(); 1176 return cfqg; 1177} 1178 1179static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg) 1180{ 1181 cfqg->ref++; 1182 return cfqg; 1183} 1184 1185static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) 1186{ 1187 /* Currently, all async queues are mapped to root group */ 1188 if (!cfq_cfqq_sync(cfqq)) 1189 cfqg = &cfqq->cfqd->root_group; 1190 1191 cfqq->cfqg = cfqg; 1192 /* cfqq reference on cfqg */ 1193 cfqq->cfqg->ref++; 1194} 1195 1196static void cfq_put_cfqg(struct cfq_group *cfqg) 1197{ 1198 struct cfq_rb_root *st; 1199 int i, j; 1200 1201 BUG_ON(cfqg->ref <= 0); 1202 cfqg->ref--; 1203 if (cfqg->ref) 1204 return; 1205 for_each_cfqg_st(cfqg, i, j, st) 1206 BUG_ON(!RB_EMPTY_ROOT(&st->rb)); 1207 free_percpu(cfqg->blkg.stats_cpu); 1208 kfree(cfqg); 1209} 1210 1211static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg) 1212{ 1213 /* Something wrong if we are trying to remove same group twice */ 1214 BUG_ON(hlist_unhashed(&cfqg->cfqd_node)); 1215 1216 hlist_del_init(&cfqg->cfqd_node); 1217 1218 BUG_ON(cfqd->nr_blkcg_linked_grps <= 0); 1219 cfqd->nr_blkcg_linked_grps--; 1220 1221 /* 1222 * Put the reference taken at the time of creation so that when all 1223 * queues are gone, group can be destroyed. 1224 */ 1225 cfq_put_cfqg(cfqg); 1226} 1227 1228static bool cfq_release_cfq_groups(struct cfq_data *cfqd) 1229{ 1230 struct hlist_node *pos, *n; 1231 struct cfq_group *cfqg; 1232 bool empty = true; 1233 1234 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) { 1235 /* 1236 * If cgroup removal path got to blk_group first and removed 1237 * it from cgroup list, then it will take care of destroying 1238 * cfqg also. 1239 */ 1240 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg)) 1241 cfq_destroy_cfqg(cfqd, cfqg); 1242 else 1243 empty = false; 1244 } 1245 return empty; 1246} 1247 1248/* 1249 * Blk cgroup controller notification saying that blkio_group object is being 1250 * delinked as associated cgroup object is going away. That also means that 1251 * no new IO will come in this group. So get rid of this group as soon as 1252 * any pending IO in the group is finished. 1253 * 1254 * This function is called under rcu_read_lock(). key is the rcu protected 1255 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu 1256 * read lock. 1257 * 1258 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means 1259 * it should not be NULL as even if elevator was exiting, cgroup deltion 1260 * path got to it first. 1261 */ 1262static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg) 1263{ 1264 unsigned long flags; 1265 struct cfq_data *cfqd = key; 1266 1267 spin_lock_irqsave(cfqd->queue->queue_lock, flags); 1268 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg)); 1269 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); 1270} 1271 1272static struct elevator_type iosched_cfq; 1273 1274static bool cfq_clear_queue(struct request_queue *q) 1275{ 1276 lockdep_assert_held(q->queue_lock); 1277 1278 /* shoot down blkgs iff the current elevator is cfq */ 1279 if (!q->elevator || q->elevator->type != &iosched_cfq) 1280 return true; 1281 1282 return cfq_release_cfq_groups(q->elevator->elevator_data); 1283} 1284 1285#else /* GROUP_IOSCHED */ 1286static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd) 1287{ 1288 return &cfqd->root_group; 1289} 1290 1291static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg) 1292{ 1293 return cfqg; 1294} 1295 1296static inline void 1297cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) { 1298 cfqq->cfqg = cfqg; 1299} 1300 1301static void cfq_release_cfq_groups(struct cfq_data *cfqd) {} 1302static inline void cfq_put_cfqg(struct cfq_group *cfqg) {} 1303 1304#endif /* GROUP_IOSCHED */ 1305 1306/* 1307 * The cfqd->service_trees holds all pending cfq_queue's that have 1308 * requests waiting to be processed. It is sorted in the order that 1309 * we will service the queues. 1310 */ 1311static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1312 bool add_front) 1313{ 1314 struct rb_node **p, *parent; 1315 struct cfq_queue *__cfqq; 1316 unsigned long rb_key; 1317 struct cfq_rb_root *service_tree; 1318 int left; 1319 int new_cfqq = 1; 1320 1321 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq), 1322 cfqq_type(cfqq)); 1323 if (cfq_class_idle(cfqq)) { 1324 rb_key = CFQ_IDLE_DELAY; 1325 parent = rb_last(&service_tree->rb); 1326 if (parent && parent != &cfqq->rb_node) { 1327 __cfqq = rb_entry(parent, struct cfq_queue, rb_node); 1328 rb_key += __cfqq->rb_key; 1329 } else 1330 rb_key += jiffies; 1331 } else if (!add_front) { 1332 /* 1333 * Get our rb key offset. Subtract any residual slice 1334 * value carried from last service. A negative resid 1335 * count indicates slice overrun, and this should position 1336 * the next service time further away in the tree. 1337 */ 1338 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; 1339 rb_key -= cfqq->slice_resid; 1340 cfqq->slice_resid = 0; 1341 } else { 1342 rb_key = -HZ; 1343 __cfqq = cfq_rb_first(service_tree); 1344 rb_key += __cfqq ? __cfqq->rb_key : jiffies; 1345 } 1346 1347 if (!RB_EMPTY_NODE(&cfqq->rb_node)) { 1348 new_cfqq = 0; 1349 /* 1350 * same position, nothing more to do 1351 */ 1352 if (rb_key == cfqq->rb_key && 1353 cfqq->service_tree == service_tree) 1354 return; 1355 1356 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); 1357 cfqq->service_tree = NULL; 1358 } 1359 1360 left = 1; 1361 parent = NULL; 1362 cfqq->service_tree = service_tree; 1363 p = &service_tree->rb.rb_node; 1364 while (*p) { 1365 struct rb_node **n; 1366 1367 parent = *p; 1368 __cfqq = rb_entry(parent, struct cfq_queue, rb_node); 1369 1370 /* 1371 * sort by key, that represents service time. 1372 */ 1373 if (time_before(rb_key, __cfqq->rb_key)) 1374 n = &(*p)->rb_left; 1375 else { 1376 n = &(*p)->rb_right; 1377 left = 0; 1378 } 1379 1380 p = n; 1381 } 1382 1383 if (left) 1384 service_tree->left = &cfqq->rb_node; 1385 1386 cfqq->rb_key = rb_key; 1387 rb_link_node(&cfqq->rb_node, parent, p); 1388 rb_insert_color(&cfqq->rb_node, &service_tree->rb); 1389 service_tree->count++; 1390 if (add_front || !new_cfqq) 1391 return; 1392 cfq_group_notify_queue_add(cfqd, cfqq->cfqg); 1393} 1394 1395static struct cfq_queue * 1396cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, 1397 sector_t sector, struct rb_node **ret_parent, 1398 struct rb_node ***rb_link) 1399{ 1400 struct rb_node **p, *parent; 1401 struct cfq_queue *cfqq = NULL; 1402 1403 parent = NULL; 1404 p = &root->rb_node; 1405 while (*p) { 1406 struct rb_node **n; 1407 1408 parent = *p; 1409 cfqq = rb_entry(parent, struct cfq_queue, p_node); 1410 1411 /* 1412 * Sort strictly based on sector. Smallest to the left, 1413 * largest to the right. 1414 */ 1415 if (sector > blk_rq_pos(cfqq->next_rq)) 1416 n = &(*p)->rb_right; 1417 else if (sector < blk_rq_pos(cfqq->next_rq)) 1418 n = &(*p)->rb_left; 1419 else 1420 break; 1421 p = n; 1422 cfqq = NULL; 1423 } 1424 1425 *ret_parent = parent; 1426 if (rb_link) 1427 *rb_link = p; 1428 return cfqq; 1429} 1430 1431static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1432{ 1433 struct rb_node **p, *parent; 1434 struct cfq_queue *__cfqq; 1435 1436 if (cfqq->p_root) { 1437 rb_erase(&cfqq->p_node, cfqq->p_root); 1438 cfqq->p_root = NULL; 1439 } 1440 1441 if (cfq_class_idle(cfqq)) 1442 return; 1443 if (!cfqq->next_rq) 1444 return; 1445 1446 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; 1447 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, 1448 blk_rq_pos(cfqq->next_rq), &parent, &p); 1449 if (!__cfqq) { 1450 rb_link_node(&cfqq->p_node, parent, p); 1451 rb_insert_color(&cfqq->p_node, cfqq->p_root); 1452 } else 1453 cfqq->p_root = NULL; 1454} 1455 1456/* 1457 * Update cfqq's position in the service tree. 1458 */ 1459static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1460{ 1461 /* 1462 * Resorting requires the cfqq to be on the RR list already. 1463 */ 1464 if (cfq_cfqq_on_rr(cfqq)) { 1465 cfq_service_tree_add(cfqd, cfqq, 0); 1466 cfq_prio_tree_add(cfqd, cfqq); 1467 } 1468} 1469 1470/* 1471 * add to busy list of queues for service, trying to be fair in ordering 1472 * the pending list according to last request service 1473 */ 1474static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1475{ 1476 cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); 1477 BUG_ON(cfq_cfqq_on_rr(cfqq)); 1478 cfq_mark_cfqq_on_rr(cfqq); 1479 cfqd->busy_queues++; 1480 if (cfq_cfqq_sync(cfqq)) 1481 cfqd->busy_sync_queues++; 1482 1483 cfq_resort_rr_list(cfqd, cfqq); 1484} 1485 1486/* 1487 * Called when the cfqq no longer has requests pending, remove it from 1488 * the service tree. 1489 */ 1490static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1491{ 1492 cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); 1493 BUG_ON(!cfq_cfqq_on_rr(cfqq)); 1494 cfq_clear_cfqq_on_rr(cfqq); 1495 1496 if (!RB_EMPTY_NODE(&cfqq->rb_node)) { 1497 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); 1498 cfqq->service_tree = NULL; 1499 } 1500 if (cfqq->p_root) { 1501 rb_erase(&cfqq->p_node, cfqq->p_root); 1502 cfqq->p_root = NULL; 1503 } 1504 1505 cfq_group_notify_queue_del(cfqd, cfqq->cfqg); 1506 BUG_ON(!cfqd->busy_queues); 1507 cfqd->busy_queues--; 1508 if (cfq_cfqq_sync(cfqq)) 1509 cfqd->busy_sync_queues--; 1510} 1511 1512/* 1513 * rb tree support functions 1514 */ 1515static void cfq_del_rq_rb(struct request *rq) 1516{ 1517 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1518 const int sync = rq_is_sync(rq); 1519 1520 BUG_ON(!cfqq->queued[sync]); 1521 cfqq->queued[sync]--; 1522 1523 elv_rb_del(&cfqq->sort_list, rq); 1524 1525 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) { 1526 /* 1527 * Queue will be deleted from service tree when we actually 1528 * expire it later. Right now just remove it from prio tree 1529 * as it is empty. 1530 */ 1531 if (cfqq->p_root) { 1532 rb_erase(&cfqq->p_node, cfqq->p_root); 1533 cfqq->p_root = NULL; 1534 } 1535 } 1536} 1537 1538static void cfq_add_rq_rb(struct request *rq) 1539{ 1540 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1541 struct cfq_data *cfqd = cfqq->cfqd; 1542 struct request *prev; 1543 1544 cfqq->queued[rq_is_sync(rq)]++; 1545 1546 elv_rb_add(&cfqq->sort_list, rq); 1547 1548 if (!cfq_cfqq_on_rr(cfqq)) 1549 cfq_add_cfqq_rr(cfqd, cfqq); 1550 1551 /* 1552 * check if this request is a better next-serve candidate 1553 */ 1554 prev = cfqq->next_rq; 1555 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position); 1556 1557 /* 1558 * adjust priority tree position, if ->next_rq changes 1559 */ 1560 if (prev != cfqq->next_rq) 1561 cfq_prio_tree_add(cfqd, cfqq); 1562 1563 BUG_ON(!cfqq->next_rq); 1564} 1565 1566static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) 1567{ 1568 elv_rb_del(&cfqq->sort_list, rq); 1569 cfqq->queued[rq_is_sync(rq)]--; 1570 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, 1571 rq_data_dir(rq), rq_is_sync(rq)); 1572 cfq_add_rq_rb(rq); 1573 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg, 1574 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq), 1575 rq_is_sync(rq)); 1576} 1577 1578static struct request * 1579cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) 1580{ 1581 struct task_struct *tsk = current; 1582 struct cfq_io_cq *cic; 1583 struct cfq_queue *cfqq; 1584 1585 cic = cfq_cic_lookup(cfqd, tsk->io_context); 1586 if (!cic) 1587 return NULL; 1588 1589 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); 1590 if (cfqq) { 1591 sector_t sector = bio->bi_sector + bio_sectors(bio); 1592 1593 return elv_rb_find(&cfqq->sort_list, sector); 1594 } 1595 1596 return NULL; 1597} 1598 1599static void cfq_activate_request(struct request_queue *q, struct request *rq) 1600{ 1601 struct cfq_data *cfqd = q->elevator->elevator_data; 1602 1603 cfqd->rq_in_driver++; 1604 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", 1605 cfqd->rq_in_driver); 1606 1607 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); 1608} 1609 1610static void cfq_deactivate_request(struct request_queue *q, struct request *rq) 1611{ 1612 struct cfq_data *cfqd = q->elevator->elevator_data; 1613 1614 WARN_ON(!cfqd->rq_in_driver); 1615 cfqd->rq_in_driver--; 1616 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", 1617 cfqd->rq_in_driver); 1618} 1619 1620static void cfq_remove_request(struct request *rq) 1621{ 1622 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1623 1624 if (cfqq->next_rq == rq) 1625 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); 1626 1627 list_del_init(&rq->queuelist); 1628 cfq_del_rq_rb(rq); 1629 1630 cfqq->cfqd->rq_queued--; 1631 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, 1632 rq_data_dir(rq), rq_is_sync(rq)); 1633 if (rq->cmd_flags & REQ_PRIO) { 1634 WARN_ON(!cfqq->prio_pending); 1635 cfqq->prio_pending--; 1636 } 1637} 1638 1639static int cfq_merge(struct request_queue *q, struct request **req, 1640 struct bio *bio) 1641{ 1642 struct cfq_data *cfqd = q->elevator->elevator_data; 1643 struct request *__rq; 1644 1645 __rq = cfq_find_rq_fmerge(cfqd, bio); 1646 if (__rq && elv_rq_merge_ok(__rq, bio)) { 1647 *req = __rq; 1648 return ELEVATOR_FRONT_MERGE; 1649 } 1650 1651 return ELEVATOR_NO_MERGE; 1652} 1653 1654static void cfq_merged_request(struct request_queue *q, struct request *req, 1655 int type) 1656{ 1657 if (type == ELEVATOR_FRONT_MERGE) { 1658 struct cfq_queue *cfqq = RQ_CFQQ(req); 1659 1660 cfq_reposition_rq_rb(cfqq, req); 1661 } 1662} 1663 1664static void cfq_bio_merged(struct request_queue *q, struct request *req, 1665 struct bio *bio) 1666{ 1667 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg, 1668 bio_data_dir(bio), cfq_bio_sync(bio)); 1669} 1670 1671static void 1672cfq_merged_requests(struct request_queue *q, struct request *rq, 1673 struct request *next) 1674{ 1675 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1676 struct cfq_data *cfqd = q->elevator->elevator_data; 1677 1678 /* 1679 * reposition in fifo if next is older than rq 1680 */ 1681 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && 1682 time_before(rq_fifo_time(next), rq_fifo_time(rq))) { 1683 list_move(&rq->queuelist, &next->queuelist); 1684 rq_set_fifo_time(rq, rq_fifo_time(next)); 1685 } 1686 1687 if (cfqq->next_rq == next) 1688 cfqq->next_rq = rq; 1689 cfq_remove_request(next); 1690 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg, 1691 rq_data_dir(next), rq_is_sync(next)); 1692 1693 cfqq = RQ_CFQQ(next); 1694 /* 1695 * all requests of this queue are merged to other queues, delete it 1696 * from the service tree. If it's the active_queue, 1697 * cfq_dispatch_requests() will choose to expire it or do idle 1698 */ 1699 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) && 1700 cfqq != cfqd->active_queue) 1701 cfq_del_cfqq_rr(cfqd, cfqq); 1702} 1703 1704static int cfq_allow_merge(struct request_queue *q, struct request *rq, 1705 struct bio *bio) 1706{ 1707 struct cfq_data *cfqd = q->elevator->elevator_data; 1708 struct cfq_io_cq *cic; 1709 struct cfq_queue *cfqq; 1710 1711 /* 1712 * Disallow merge of a sync bio into an async request. 1713 */ 1714 if (cfq_bio_sync(bio) && !rq_is_sync(rq)) 1715 return false; 1716 1717 /* 1718 * Lookup the cfqq that this bio will be queued with and allow 1719 * merge only if rq is queued there. 1720 */ 1721 cic = cfq_cic_lookup(cfqd, current->io_context); 1722 if (!cic) 1723 return false; 1724 1725 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); 1726 return cfqq == RQ_CFQQ(rq); 1727} 1728 1729static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1730{ 1731 del_timer(&cfqd->idle_slice_timer); 1732 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg); 1733} 1734 1735static void __cfq_set_active_queue(struct cfq_data *cfqd, 1736 struct cfq_queue *cfqq) 1737{ 1738 if (cfqq) { 1739 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d", 1740 cfqd->serving_prio, cfqd->serving_type); 1741 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg); 1742 cfqq->slice_start = 0; 1743 cfqq->dispatch_start = jiffies; 1744 cfqq->allocated_slice = 0; 1745 cfqq->slice_end = 0; 1746 cfqq->slice_dispatch = 0; 1747 cfqq->nr_sectors = 0; 1748 1749 cfq_clear_cfqq_wait_request(cfqq); 1750 cfq_clear_cfqq_must_dispatch(cfqq); 1751 cfq_clear_cfqq_must_alloc_slice(cfqq); 1752 cfq_clear_cfqq_fifo_expire(cfqq); 1753 cfq_mark_cfqq_slice_new(cfqq); 1754 1755 cfq_del_timer(cfqd, cfqq); 1756 } 1757 1758 cfqd->active_queue = cfqq; 1759} 1760 1761/* 1762 * current cfqq expired its slice (or was too idle), select new one 1763 */ 1764static void 1765__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1766 bool timed_out) 1767{ 1768 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); 1769 1770 if (cfq_cfqq_wait_request(cfqq)) 1771 cfq_del_timer(cfqd, cfqq); 1772 1773 cfq_clear_cfqq_wait_request(cfqq); 1774 cfq_clear_cfqq_wait_busy(cfqq); 1775 1776 /* 1777 * If this cfqq is shared between multiple processes, check to 1778 * make sure that those processes are still issuing I/Os within 1779 * the mean seek distance. If not, it may be time to break the 1780 * queues apart again. 1781 */ 1782 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq)) 1783 cfq_mark_cfqq_split_coop(cfqq); 1784 1785 /* 1786 * store what was left of this slice, if the queue idled/timed out 1787 */ 1788 if (timed_out) { 1789 if (cfq_cfqq_slice_new(cfqq)) 1790 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq); 1791 else 1792 cfqq->slice_resid = cfqq->slice_end - jiffies; 1793 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); 1794 } 1795 1796 cfq_group_served(cfqd, cfqq->cfqg, cfqq); 1797 1798 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) 1799 cfq_del_cfqq_rr(cfqd, cfqq); 1800 1801 cfq_resort_rr_list(cfqd, cfqq); 1802 1803 if (cfqq == cfqd->active_queue) 1804 cfqd->active_queue = NULL; 1805 1806 if (cfqd->active_cic) { 1807 put_io_context(cfqd->active_cic->icq.ioc); 1808 cfqd->active_cic = NULL; 1809 } 1810} 1811 1812static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) 1813{ 1814 struct cfq_queue *cfqq = cfqd->active_queue; 1815 1816 if (cfqq) 1817 __cfq_slice_expired(cfqd, cfqq, timed_out); 1818} 1819 1820/* 1821 * Get next queue for service. Unless we have a queue preemption, 1822 * we'll simply select the first cfqq in the service tree. 1823 */ 1824static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) 1825{ 1826 struct cfq_rb_root *service_tree = 1827 service_tree_for(cfqd->serving_group, cfqd->serving_prio, 1828 cfqd->serving_type); 1829 1830 if (!cfqd->rq_queued) 1831 return NULL; 1832 1833 /* There is nothing to dispatch */ 1834 if (!service_tree) 1835 return NULL; 1836 if (RB_EMPTY_ROOT(&service_tree->rb)) 1837 return NULL; 1838 return cfq_rb_first(service_tree); 1839} 1840 1841static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd) 1842{ 1843 struct cfq_group *cfqg; 1844 struct cfq_queue *cfqq; 1845 int i, j; 1846 struct cfq_rb_root *st; 1847 1848 if (!cfqd->rq_queued) 1849 return NULL; 1850 1851 cfqg = cfq_get_next_cfqg(cfqd); 1852 if (!cfqg) 1853 return NULL; 1854 1855 for_each_cfqg_st(cfqg, i, j, st) 1856 if ((cfqq = cfq_rb_first(st)) != NULL) 1857 return cfqq; 1858 return NULL; 1859} 1860 1861/* 1862 * Get and set a new active queue for service. 1863 */ 1864static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, 1865 struct cfq_queue *cfqq) 1866{ 1867 if (!cfqq) 1868 cfqq = cfq_get_next_queue(cfqd); 1869 1870 __cfq_set_active_queue(cfqd, cfqq); 1871 return cfqq; 1872} 1873 1874static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, 1875 struct request *rq) 1876{ 1877 if (blk_rq_pos(rq) >= cfqd->last_position) 1878 return blk_rq_pos(rq) - cfqd->last_position; 1879 else 1880 return cfqd->last_position - blk_rq_pos(rq); 1881} 1882 1883static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1884 struct request *rq) 1885{ 1886 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR; 1887} 1888 1889static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, 1890 struct cfq_queue *cur_cfqq) 1891{ 1892 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; 1893 struct rb_node *parent, *node; 1894 struct cfq_queue *__cfqq; 1895 sector_t sector = cfqd->last_position; 1896 1897 if (RB_EMPTY_ROOT(root)) 1898 return NULL; 1899 1900 /* 1901 * First, if we find a request starting at the end of the last 1902 * request, choose it. 1903 */ 1904 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); 1905 if (__cfqq) 1906 return __cfqq; 1907 1908 /* 1909 * If the exact sector wasn't found, the parent of the NULL leaf 1910 * will contain the closest sector. 1911 */ 1912 __cfqq = rb_entry(parent, struct cfq_queue, p_node); 1913 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) 1914 return __cfqq; 1915 1916 if (blk_rq_pos(__cfqq->next_rq) < sector) 1917 node = rb_next(&__cfqq->p_node); 1918 else 1919 node = rb_prev(&__cfqq->p_node); 1920 if (!node) 1921 return NULL; 1922 1923 __cfqq = rb_entry(node, struct cfq_queue, p_node); 1924 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) 1925 return __cfqq; 1926 1927 return NULL; 1928} 1929 1930/* 1931 * cfqd - obvious 1932 * cur_cfqq - passed in so that we don't decide that the current queue is 1933 * closely cooperating with itself. 1934 * 1935 * So, basically we're assuming that that cur_cfqq has dispatched at least 1936 * one request, and that cfqd->last_position reflects a position on the disk 1937 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid 1938 * assumption. 1939 */ 1940static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, 1941 struct cfq_queue *cur_cfqq) 1942{ 1943 struct cfq_queue *cfqq; 1944 1945 if (cfq_class_idle(cur_cfqq)) 1946 return NULL; 1947 if (!cfq_cfqq_sync(cur_cfqq)) 1948 return NULL; 1949 if (CFQQ_SEEKY(cur_cfqq)) 1950 return NULL; 1951 1952 /* 1953 * Don't search priority tree if it's the only queue in the group. 1954 */ 1955 if (cur_cfqq->cfqg->nr_cfqq == 1) 1956 return NULL; 1957 1958 /* 1959 * We should notice if some of the queues are cooperating, eg 1960 * working closely on the same area of the disk. In that case, 1961 * we can group them together and don't waste time idling. 1962 */ 1963 cfqq = cfqq_close(cfqd, cur_cfqq); 1964 if (!cfqq) 1965 return NULL; 1966 1967 /* If new queue belongs to different cfq_group, don't choose it */ 1968 if (cur_cfqq->cfqg != cfqq->cfqg) 1969 return NULL; 1970 1971 /* 1972 * It only makes sense to merge sync queues. 1973 */ 1974 if (!cfq_cfqq_sync(cfqq)) 1975 return NULL; 1976 if (CFQQ_SEEKY(cfqq)) 1977 return NULL; 1978 1979 /* 1980 * Do not merge queues of different priority classes 1981 */ 1982 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) 1983 return NULL; 1984 1985 return cfqq; 1986} 1987 1988/* 1989 * Determine whether we should enforce idle window for this queue. 1990 */ 1991 1992static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1993{ 1994 enum wl_prio_t prio = cfqq_prio(cfqq); 1995 struct cfq_rb_root *service_tree = cfqq->service_tree; 1996 1997 BUG_ON(!service_tree); 1998 BUG_ON(!service_tree->count); 1999 2000 if (!cfqd->cfq_slice_idle) 2001 return false; 2002 2003 /* We never do for idle class queues. */ 2004 if (prio == IDLE_WORKLOAD) 2005 return false; 2006 2007 /* We do for queues that were marked with idle window flag. */ 2008 if (cfq_cfqq_idle_window(cfqq) && 2009 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)) 2010 return true; 2011 2012 /* 2013 * Otherwise, we do only if they are the last ones 2014 * in their service tree. 2015 */ 2016 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) && 2017 !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false)) 2018 return true; 2019 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", 2020 service_tree->count); 2021 return false; 2022} 2023 2024static void cfq_arm_slice_timer(struct cfq_data *cfqd) 2025{ 2026 struct cfq_queue *cfqq = cfqd->active_queue; 2027 struct cfq_io_cq *cic; 2028 unsigned long sl, group_idle = 0; 2029 2030 /* 2031 * SSD device without seek penalty, disable idling. But only do so 2032 * for devices that support queuing, otherwise we still have a problem 2033 * with sync vs async workloads. 2034 */ 2035 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) 2036 return; 2037 2038 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); 2039 WARN_ON(cfq_cfqq_slice_new(cfqq)); 2040 2041 /* 2042 * idle is disabled, either manually or by past process history 2043 */ 2044 if (!cfq_should_idle(cfqd, cfqq)) { 2045 /* no queue idling. Check for group idling */ 2046 if (cfqd->cfq_group_idle) 2047 group_idle = cfqd->cfq_group_idle; 2048 else 2049 return; 2050 } 2051 2052 /* 2053 * still active requests from this queue, don't idle 2054 */ 2055 if (cfqq->dispatched) 2056 return; 2057 2058 /* 2059 * task has exited, don't wait 2060 */ 2061 cic = cfqd->active_cic; 2062 if (!cic || !atomic_read(&cic->icq.ioc->nr_tasks)) 2063 return; 2064 2065 /* 2066 * If our average think time is larger than the remaining time 2067 * slice, then don't idle. This avoids overrunning the allotted 2068 * time slice. 2069 */ 2070 if (sample_valid(cic->ttime.ttime_samples) && 2071 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) { 2072 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu", 2073 cic->ttime.ttime_mean); 2074 return; 2075 } 2076 2077 /* There are other queues in the group, don't do group idle */ 2078 if (group_idle && cfqq->cfqg->nr_cfqq > 1) 2079 return; 2080 2081 cfq_mark_cfqq_wait_request(cfqq); 2082 2083 if (group_idle) 2084 sl = cfqd->cfq_group_idle; 2085 else 2086 sl = cfqd->cfq_slice_idle; 2087 2088 mod_timer(&cfqd->idle_slice_timer, jiffies + sl); 2089 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg); 2090 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl, 2091 group_idle ? 1 : 0); 2092} 2093 2094/* 2095 * Move request from internal lists to the request queue dispatch list. 2096 */ 2097static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) 2098{ 2099 struct cfq_data *cfqd = q->elevator->elevator_data; 2100 struct cfq_queue *cfqq = RQ_CFQQ(rq); 2101 2102 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); 2103 2104 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); 2105 cfq_remove_request(rq); 2106 cfqq->dispatched++; 2107 (RQ_CFQG(rq))->dispatched++; 2108 elv_dispatch_sort(q, rq); 2109 2110 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++; 2111 cfqq->nr_sectors += blk_rq_sectors(rq); 2112 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq), 2113 rq_data_dir(rq), rq_is_sync(rq)); 2114} 2115 2116/* 2117 * return expired entry, or NULL to just start from scratch in rbtree 2118 */ 2119static struct request *cfq_check_fifo(struct cfq_queue *cfqq) 2120{ 2121 struct request *rq = NULL; 2122 2123 if (cfq_cfqq_fifo_expire(cfqq)) 2124 return NULL; 2125 2126 cfq_mark_cfqq_fifo_expire(cfqq); 2127 2128 if (list_empty(&cfqq->fifo)) 2129 return NULL; 2130 2131 rq = rq_entry_fifo(cfqq->fifo.next); 2132 if (time_before(jiffies, rq_fifo_time(rq))) 2133 rq = NULL; 2134 2135 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); 2136 return rq; 2137} 2138 2139static inline int 2140cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) 2141{ 2142 const int base_rq = cfqd->cfq_slice_async_rq; 2143 2144 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); 2145 2146 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio); 2147} 2148 2149/* 2150 * Must be called with the queue_lock held. 2151 */ 2152static int cfqq_process_refs(struct cfq_queue *cfqq) 2153{ 2154 int process_refs, io_refs; 2155 2156 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; 2157 process_refs = cfqq->ref - io_refs; 2158 BUG_ON(process_refs < 0); 2159 return process_refs; 2160} 2161 2162static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) 2163{ 2164 int process_refs, new_process_refs; 2165 struct cfq_queue *__cfqq; 2166 2167 /* 2168 * If there are no process references on the new_cfqq, then it is 2169 * unsafe to follow the ->new_cfqq chain as other cfqq's in the 2170 * chain may have dropped their last reference (not just their 2171 * last process reference). 2172 */ 2173 if (!cfqq_process_refs(new_cfqq)) 2174 return; 2175 2176 /* Avoid a circular list and skip interim queue merges */ 2177 while ((__cfqq = new_cfqq->new_cfqq)) { 2178 if (__cfqq == cfqq) 2179 return; 2180 new_cfqq = __cfqq; 2181 } 2182 2183 process_refs = cfqq_process_refs(cfqq); 2184 new_process_refs = cfqq_process_refs(new_cfqq); 2185 /* 2186 * If the process for the cfqq has gone away, there is no 2187 * sense in merging the queues. 2188 */ 2189 if (process_refs == 0 || new_process_refs == 0) 2190 return; 2191 2192 /* 2193 * Merge in the direction of the lesser amount of work. 2194 */ 2195 if (new_process_refs >= process_refs) { 2196 cfqq->new_cfqq = new_cfqq; 2197 new_cfqq->ref += process_refs; 2198 } else { 2199 new_cfqq->new_cfqq = cfqq; 2200 cfqq->ref += new_process_refs; 2201 } 2202} 2203 2204static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, 2205 struct cfq_group *cfqg, enum wl_prio_t prio) 2206{ 2207 struct cfq_queue *queue; 2208 int i; 2209 bool key_valid = false; 2210 unsigned long lowest_key = 0; 2211 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD; 2212 2213 for (i = 0; i <= SYNC_WORKLOAD; ++i) { 2214 /* select the one with lowest rb_key */ 2215 queue = cfq_rb_first(service_tree_for(cfqg, prio, i)); 2216 if (queue && 2217 (!key_valid || time_before(queue->rb_key, lowest_key))) { 2218 lowest_key = queue->rb_key; 2219 cur_best = i; 2220 key_valid = true; 2221 } 2222 } 2223 2224 return cur_best; 2225} 2226 2227static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg) 2228{ 2229 unsigned slice; 2230 unsigned count; 2231 struct cfq_rb_root *st; 2232 unsigned group_slice; 2233 enum wl_prio_t original_prio = cfqd->serving_prio; 2234 2235 /* Choose next priority. RT > BE > IDLE */ 2236 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg)) 2237 cfqd->serving_prio = RT_WORKLOAD; 2238 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg)) 2239 cfqd->serving_prio = BE_WORKLOAD; 2240 else { 2241 cfqd->serving_prio = IDLE_WORKLOAD; 2242 cfqd->workload_expires = jiffies + 1; 2243 return; 2244 } 2245 2246 if (original_prio != cfqd->serving_prio) 2247 goto new_workload; 2248 2249 /* 2250 * For RT and BE, we have to choose also the type 2251 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload 2252 * expiration time 2253 */ 2254 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type); 2255 count = st->count; 2256 2257 /* 2258 * check workload expiration, and that we still have other queues ready 2259 */ 2260 if (count && !time_after(jiffies, cfqd->workload_expires)) 2261 return; 2262 2263new_workload: 2264 /* otherwise select new workload type */ 2265 cfqd->serving_type = 2266 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio); 2267 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type); 2268 count = st->count; 2269 2270 /* 2271 * the workload slice is computed as a fraction of target latency 2272 * proportional to the number of queues in that workload, over 2273 * all the queues in the same priority class 2274 */ 2275 group_slice = cfq_group_slice(cfqd, cfqg); 2276 2277 slice = group_slice * count / 2278 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio], 2279 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg)); 2280 2281 if (cfqd->serving_type == ASYNC_WORKLOAD) { 2282 unsigned int tmp; 2283 2284 /* 2285 * Async queues are currently system wide. Just taking 2286 * proportion of queues with-in same group will lead to higher 2287 * async ratio system wide as generally root group is going 2288 * to have higher weight. A more accurate thing would be to 2289 * calculate system wide asnc/sync ratio. 2290 */ 2291 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg); 2292 tmp = tmp/cfqd->busy_queues; 2293 slice = min_t(unsigned, slice, tmp); 2294 2295 /* async workload slice is scaled down according to 2296 * the sync/async slice ratio. */ 2297 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1]; 2298 } else 2299 /* sync workload slice is at least 2 * cfq_slice_idle */ 2300 slice = max(slice, 2 * cfqd->cfq_slice_idle); 2301 2302 slice = max_t(unsigned, slice, CFQ_MIN_TT); 2303 cfq_log(cfqd, "workload slice:%d", slice); 2304 cfqd->workload_expires = jiffies + slice; 2305} 2306 2307static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd) 2308{ 2309 struct cfq_rb_root *st = &cfqd->grp_service_tree; 2310 struct cfq_group *cfqg; 2311 2312 if (RB_EMPTY_ROOT(&st->rb)) 2313 return NULL; 2314 cfqg = cfq_rb_first_group(st); 2315 update_min_vdisktime(st); 2316 return cfqg; 2317} 2318 2319static void cfq_choose_cfqg(struct cfq_data *cfqd) 2320{ 2321 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd); 2322 2323 cfqd->serving_group = cfqg; 2324 2325 /* Restore the workload type data */ 2326 if (cfqg->saved_workload_slice) { 2327 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice; 2328 cfqd->serving_type = cfqg->saved_workload; 2329 cfqd->serving_prio = cfqg->saved_serving_prio; 2330 } else 2331 cfqd->workload_expires = jiffies - 1; 2332 2333 choose_service_tree(cfqd, cfqg); 2334} 2335 2336/* 2337 * Select a queue for service. If we have a current active queue, 2338 * check whether to continue servicing it, or retrieve and set a new one. 2339 */ 2340static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) 2341{ 2342 struct cfq_queue *cfqq, *new_cfqq = NULL; 2343 2344 cfqq = cfqd->active_queue; 2345 if (!cfqq) 2346 goto new_queue; 2347 2348 if (!cfqd->rq_queued) 2349 return NULL; 2350 2351 /* 2352 * We were waiting for group to get backlogged. Expire the queue 2353 */ 2354 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list)) 2355 goto expire; 2356 2357 /* 2358 * The active queue has run out of time, expire it and select new. 2359 */ 2360 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) { 2361 /* 2362 * If slice had not expired at the completion of last request 2363 * we might not have turned on wait_busy flag. Don't expire 2364 * the queue yet. Allow the group to get backlogged. 2365 * 2366 * The very fact that we have used the slice, that means we 2367 * have been idling all along on this queue and it should be 2368 * ok to wait for this request to complete. 2369 */ 2370 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list) 2371 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) { 2372 cfqq = NULL; 2373 goto keep_queue; 2374 } else 2375 goto check_group_idle; 2376 } 2377 2378 /* 2379 * The active queue has requests and isn't expired, allow it to 2380 * dispatch. 2381 */ 2382 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) 2383 goto keep_queue; 2384 2385 /* 2386 * If another queue has a request waiting within our mean seek 2387 * distance, let it run. The expire code will check for close 2388 * cooperators and put the close queue at the front of the service 2389 * tree. If possible, merge the expiring queue with the new cfqq. 2390 */ 2391 new_cfqq = cfq_close_cooperator(cfqd, cfqq); 2392 if (new_cfqq) { 2393 if (!cfqq->new_cfqq) 2394 cfq_setup_merge(cfqq, new_cfqq); 2395 goto expire; 2396 } 2397 2398 /* 2399 * No requests pending. If the active queue still has requests in 2400 * flight or is idling for a new request, allow either of these 2401 * conditions to happen (or time out) before selecting a new queue. 2402 */ 2403 if (timer_pending(&cfqd->idle_slice_timer)) { 2404 cfqq = NULL; 2405 goto keep_queue; 2406 } 2407 2408 /* 2409 * This is a deep seek queue, but the device is much faster than 2410 * the queue can deliver, don't idle 2411 **/ 2412 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) && 2413 (cfq_cfqq_slice_new(cfqq) || 2414 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) { 2415 cfq_clear_cfqq_deep(cfqq); 2416 cfq_clear_cfqq_idle_window(cfqq); 2417 } 2418 2419 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) { 2420 cfqq = NULL; 2421 goto keep_queue; 2422 } 2423 2424 /* 2425 * If group idle is enabled and there are requests dispatched from 2426 * this group, wait for requests to complete. 2427 */ 2428check_group_idle: 2429 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 && 2430 cfqq->cfqg->dispatched && 2431 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) { 2432 cfqq = NULL; 2433 goto keep_queue; 2434 } 2435 2436expire: 2437 cfq_slice_expired(cfqd, 0); 2438new_queue: 2439 /* 2440 * Current queue expired. Check if we have to switch to a new 2441 * service tree 2442 */ 2443 if (!new_cfqq) 2444 cfq_choose_cfqg(cfqd); 2445 2446 cfqq = cfq_set_active_queue(cfqd, new_cfqq); 2447keep_queue: 2448 return cfqq; 2449} 2450 2451static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) 2452{ 2453 int dispatched = 0; 2454 2455 while (cfqq->next_rq) { 2456 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); 2457 dispatched++; 2458 } 2459 2460 BUG_ON(!list_empty(&cfqq->fifo)); 2461 2462 /* By default cfqq is not expired if it is empty. Do it explicitly */ 2463 __cfq_slice_expired(cfqq->cfqd, cfqq, 0); 2464 return dispatched; 2465} 2466 2467/* 2468 * Drain our current requests. Used for barriers and when switching 2469 * io schedulers on-the-fly. 2470 */ 2471static int cfq_forced_dispatch(struct cfq_data *cfqd) 2472{ 2473 struct cfq_queue *cfqq; 2474 int dispatched = 0; 2475 2476 /* Expire the timeslice of the current active queue first */ 2477 cfq_slice_expired(cfqd, 0); 2478 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) { 2479 __cfq_set_active_queue(cfqd, cfqq); 2480 dispatched += __cfq_forced_dispatch_cfqq(cfqq); 2481 } 2482 2483 BUG_ON(cfqd->busy_queues); 2484 2485 cfq_log(cfqd, "forced_dispatch=%d", dispatched); 2486 return dispatched; 2487} 2488 2489static inline bool cfq_slice_used_soon(struct cfq_data *cfqd, 2490 struct cfq_queue *cfqq) 2491{ 2492 /* the queue hasn't finished any request, can't estimate */ 2493 if (cfq_cfqq_slice_new(cfqq)) 2494 return true; 2495 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched, 2496 cfqq->slice_end)) 2497 return true; 2498 2499 return false; 2500} 2501 2502static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) 2503{ 2504 unsigned int max_dispatch; 2505 2506 /* 2507 * Drain async requests before we start sync IO 2508 */ 2509 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC]) 2510 return false; 2511 2512 /* 2513 * If this is an async queue and we have sync IO in flight, let it wait 2514 */ 2515 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq)) 2516 return false; 2517 2518 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1); 2519 if (cfq_class_idle(cfqq)) 2520 max_dispatch = 1; 2521 2522 /* 2523 * Does this cfqq already have too much IO in flight? 2524 */ 2525 if (cfqq->dispatched >= max_dispatch) { 2526 bool promote_sync = false; 2527 /* 2528 * idle queue must always only have a single IO in flight 2529 */ 2530 if (cfq_class_idle(cfqq)) 2531 return false; 2532 2533 /* 2534 * If there is only one sync queue 2535 * we can ignore async queue here and give the sync 2536 * queue no dispatch limit. The reason is a sync queue can 2537 * preempt async queue, limiting the sync queue doesn't make 2538 * sense. This is useful for aiostress test. 2539 */ 2540 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1) 2541 promote_sync = true; 2542 2543 /* 2544 * We have other queues, don't allow more IO from this one 2545 */ 2546 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) && 2547 !promote_sync) 2548 return false; 2549 2550 /* 2551 * Sole queue user, no limit 2552 */ 2553 if (cfqd->busy_queues == 1 || promote_sync) 2554 max_dispatch = -1; 2555 else 2556 /* 2557 * Normally we start throttling cfqq when cfq_quantum/2 2558 * requests have been dispatched. But we can drive 2559 * deeper queue depths at the beginning of slice 2560 * subjected to upper limit of cfq_quantum. 2561 * */ 2562 max_dispatch = cfqd->cfq_quantum; 2563 } 2564 2565 /* 2566 * Async queues must wait a bit before being allowed dispatch. 2567 * We also ramp up the dispatch depth gradually for async IO, 2568 * based on the last sync IO we serviced 2569 */ 2570 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { 2571 unsigned long last_sync = jiffies - cfqd->last_delayed_sync; 2572 unsigned int depth; 2573 2574 depth = last_sync / cfqd->cfq_slice[1]; 2575 if (!depth && !cfqq->dispatched) 2576 depth = 1; 2577 if (depth < max_dispatch) 2578 max_dispatch = depth; 2579 } 2580 2581 /* 2582 * If we're below the current max, allow a dispatch 2583 */ 2584 return cfqq->dispatched < max_dispatch; 2585} 2586 2587/* 2588 * Dispatch a request from cfqq, moving them to the request queue 2589 * dispatch list. 2590 */ 2591static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) 2592{ 2593 struct request *rq; 2594 2595 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); 2596 2597 if (!cfq_may_dispatch(cfqd, cfqq)) 2598 return false; 2599 2600 /* 2601 * follow expired path, else get first next available 2602 */ 2603 rq = cfq_check_fifo(cfqq); 2604 if (!rq) 2605 rq = cfqq->next_rq; 2606 2607 /* 2608 * insert request into driver dispatch list 2609 */ 2610 cfq_dispatch_insert(cfqd->queue, rq); 2611 2612 if (!cfqd->active_cic) { 2613 struct cfq_io_cq *cic = RQ_CIC(rq); 2614 2615 atomic_long_inc(&cic->icq.ioc->refcount); 2616 cfqd->active_cic = cic; 2617 } 2618 2619 return true; 2620} 2621 2622/* 2623 * Find the cfqq that we need to service and move a request from that to the 2624 * dispatch list 2625 */ 2626static int cfq_dispatch_requests(struct request_queue *q, int force) 2627{ 2628 struct cfq_data *cfqd = q->elevator->elevator_data; 2629 struct cfq_queue *cfqq; 2630 2631 if (!cfqd->busy_queues) 2632 return 0; 2633 2634 if (unlikely(force)) 2635 return cfq_forced_dispatch(cfqd); 2636 2637 cfqq = cfq_select_queue(cfqd); 2638 if (!cfqq) 2639 return 0; 2640 2641 /* 2642 * Dispatch a request from this cfqq, if it is allowed 2643 */ 2644 if (!cfq_dispatch_request(cfqd, cfqq)) 2645 return 0; 2646 2647 cfqq->slice_dispatch++; 2648 cfq_clear_cfqq_must_dispatch(cfqq); 2649 2650 /* 2651 * expire an async queue immediately if it has used up its slice. idle 2652 * queue always expire after 1 dispatch round. 2653 */ 2654 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && 2655 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || 2656 cfq_class_idle(cfqq))) { 2657 cfqq->slice_end = jiffies + 1; 2658 cfq_slice_expired(cfqd, 0); 2659 } 2660 2661 cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); 2662 return 1; 2663} 2664 2665/* 2666 * task holds one reference to the queue, dropped when task exits. each rq 2667 * in-flight on this queue also holds a reference, dropped when rq is freed. 2668 * 2669 * Each cfq queue took a reference on the parent group. Drop it now. 2670 * queue lock must be held here. 2671 */ 2672static void cfq_put_queue(struct cfq_queue *cfqq) 2673{ 2674 struct cfq_data *cfqd = cfqq->cfqd; 2675 struct cfq_group *cfqg; 2676 2677 BUG_ON(cfqq->ref <= 0); 2678 2679 cfqq->ref--; 2680 if (cfqq->ref) 2681 return; 2682 2683 cfq_log_cfqq(cfqd, cfqq, "put_queue"); 2684 BUG_ON(rb_first(&cfqq->sort_list)); 2685 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); 2686 cfqg = cfqq->cfqg; 2687 2688 if (unlikely(cfqd->active_queue == cfqq)) { 2689 __cfq_slice_expired(cfqd, cfqq, 0); 2690 cfq_schedule_dispatch(cfqd); 2691 } 2692 2693 BUG_ON(cfq_cfqq_on_rr(cfqq)); 2694 kmem_cache_free(cfq_pool, cfqq); 2695 cfq_put_cfqg(cfqg); 2696} 2697 2698static void cfq_put_cooperator(struct cfq_queue *cfqq) 2699{ 2700 struct cfq_queue *__cfqq, *next; 2701 2702 /* 2703 * If this queue was scheduled to merge with another queue, be 2704 * sure to drop the reference taken on that queue (and others in 2705 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. 2706 */ 2707 __cfqq = cfqq->new_cfqq; 2708 while (__cfqq) { 2709 if (__cfqq == cfqq) { 2710 WARN(1, "cfqq->new_cfqq loop detected\n"); 2711 break; 2712 } 2713 next = __cfqq->new_cfqq; 2714 cfq_put_queue(__cfqq); 2715 __cfqq = next; 2716 } 2717} 2718 2719static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) 2720{ 2721 if (unlikely(cfqq == cfqd->active_queue)) { 2722 __cfq_slice_expired(cfqd, cfqq, 0); 2723 cfq_schedule_dispatch(cfqd); 2724 } 2725 2726 cfq_put_cooperator(cfqq); 2727 2728 cfq_put_queue(cfqq); 2729} 2730 2731static void cfq_init_icq(struct io_cq *icq) 2732{ 2733 struct cfq_io_cq *cic = icq_to_cic(icq); 2734 2735 cic->ttime.last_end_request = jiffies; 2736} 2737 2738static void cfq_exit_icq(struct io_cq *icq) 2739{ 2740 struct cfq_io_cq *cic = icq_to_cic(icq); 2741 struct cfq_data *cfqd = cic_to_cfqd(cic); 2742 2743 if (cic->cfqq[BLK_RW_ASYNC]) { 2744 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); 2745 cic->cfqq[BLK_RW_ASYNC] = NULL; 2746 } 2747 2748 if (cic->cfqq[BLK_RW_SYNC]) { 2749 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); 2750 cic->cfqq[BLK_RW_SYNC] = NULL; 2751 } 2752} 2753 2754static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) 2755{ 2756 struct task_struct *tsk = current; 2757 int ioprio_class; 2758 2759 if (!cfq_cfqq_prio_changed(cfqq)) 2760 return; 2761 2762 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); 2763 switch (ioprio_class) { 2764 default: 2765 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); 2766 case IOPRIO_CLASS_NONE: 2767 /* 2768 * no prio set, inherit CPU scheduling settings 2769 */ 2770 cfqq->ioprio = task_nice_ioprio(tsk); 2771 cfqq->ioprio_class = task_nice_ioclass(tsk); 2772 break; 2773 case IOPRIO_CLASS_RT: 2774 cfqq->ioprio = task_ioprio(ioc); 2775 cfqq->ioprio_class = IOPRIO_CLASS_RT; 2776 break; 2777 case IOPRIO_CLASS_BE: 2778 cfqq->ioprio = task_ioprio(ioc); 2779 cfqq->ioprio_class = IOPRIO_CLASS_BE; 2780 break; 2781 case IOPRIO_CLASS_IDLE: 2782 cfqq->ioprio_class = IOPRIO_CLASS_IDLE; 2783 cfqq->ioprio = 7; 2784 cfq_clear_cfqq_idle_window(cfqq); 2785 break; 2786 } 2787 2788 /* 2789 * keep track of original prio settings in case we have to temporarily 2790 * elevate the priority of this queue 2791 */ 2792 cfqq->org_ioprio = cfqq->ioprio; 2793 cfq_clear_cfqq_prio_changed(cfqq); 2794} 2795 2796static void changed_ioprio(struct cfq_io_cq *cic) 2797{ 2798 struct cfq_data *cfqd = cic_to_cfqd(cic); 2799 struct cfq_queue *cfqq; 2800 2801 if (unlikely(!cfqd)) 2802 return; 2803 2804 cfqq = cic->cfqq[BLK_RW_ASYNC]; 2805 if (cfqq) { 2806 struct cfq_queue *new_cfqq; 2807 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->icq.ioc, 2808 GFP_ATOMIC); 2809 if (new_cfqq) { 2810 cic->cfqq[BLK_RW_ASYNC] = new_cfqq; 2811 cfq_put_queue(cfqq); 2812 } 2813 } 2814 2815 cfqq = cic->cfqq[BLK_RW_SYNC]; 2816 if (cfqq) 2817 cfq_mark_cfqq_prio_changed(cfqq); 2818} 2819 2820static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, 2821 pid_t pid, bool is_sync) 2822{ 2823 RB_CLEAR_NODE(&cfqq->rb_node); 2824 RB_CLEAR_NODE(&cfqq->p_node); 2825 INIT_LIST_HEAD(&cfqq->fifo); 2826 2827 cfqq->ref = 0; 2828 cfqq->cfqd = cfqd; 2829 2830 cfq_mark_cfqq_prio_changed(cfqq); 2831 2832 if (is_sync) { 2833 if (!cfq_class_idle(cfqq)) 2834 cfq_mark_cfqq_idle_window(cfqq); 2835 cfq_mark_cfqq_sync(cfqq); 2836 } 2837 cfqq->pid = pid; 2838} 2839 2840#ifdef CONFIG_CFQ_GROUP_IOSCHED 2841static void changed_cgroup(struct cfq_io_cq *cic) 2842{ 2843 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1); 2844 struct cfq_data *cfqd = cic_to_cfqd(cic); 2845 struct request_queue *q; 2846 2847 if (unlikely(!cfqd)) 2848 return; 2849 2850 q = cfqd->queue; 2851 2852 if (sync_cfqq) { 2853 /* 2854 * Drop reference to sync queue. A new sync queue will be 2855 * assigned in new group upon arrival of a fresh request. 2856 */ 2857 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup"); 2858 cic_set_cfqq(cic, NULL, 1); 2859 cfq_put_queue(sync_cfqq); 2860 } 2861} 2862#endif /* CONFIG_CFQ_GROUP_IOSCHED */ 2863 2864static struct cfq_queue * 2865cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, 2866 struct io_context *ioc, gfp_t gfp_mask) 2867{ 2868 struct cfq_queue *cfqq, *new_cfqq = NULL; 2869 struct cfq_io_cq *cic; 2870 struct cfq_group *cfqg; 2871 2872retry: 2873 cfqg = cfq_get_cfqg(cfqd); 2874 cic = cfq_cic_lookup(cfqd, ioc); 2875 /* cic always exists here */ 2876 cfqq = cic_to_cfqq(cic, is_sync); 2877 2878 /* 2879 * Always try a new alloc if we fell back to the OOM cfqq 2880 * originally, since it should just be a temporary situation. 2881 */ 2882 if (!cfqq || cfqq == &cfqd->oom_cfqq) { 2883 cfqq = NULL; 2884 if (new_cfqq) { 2885 cfqq = new_cfqq; 2886 new_cfqq = NULL; 2887 } else if (gfp_mask & __GFP_WAIT) { 2888 spin_unlock_irq(cfqd->queue->queue_lock); 2889 new_cfqq = kmem_cache_alloc_node(cfq_pool, 2890 gfp_mask | __GFP_ZERO, 2891 cfqd->queue->node); 2892 spin_lock_irq(cfqd->queue->queue_lock); 2893 if (new_cfqq) 2894 goto retry; 2895 } else { 2896 cfqq = kmem_cache_alloc_node(cfq_pool, 2897 gfp_mask | __GFP_ZERO, 2898 cfqd->queue->node); 2899 } 2900 2901 if (cfqq) { 2902 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); 2903 cfq_init_prio_data(cfqq, ioc); 2904 cfq_link_cfqq_cfqg(cfqq, cfqg); 2905 cfq_log_cfqq(cfqd, cfqq, "alloced"); 2906 } else 2907 cfqq = &cfqd->oom_cfqq; 2908 } 2909 2910 if (new_cfqq) 2911 kmem_cache_free(cfq_pool, new_cfqq); 2912 2913 return cfqq; 2914} 2915 2916static struct cfq_queue ** 2917cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) 2918{ 2919 switch (ioprio_class) { 2920 case IOPRIO_CLASS_RT: 2921 return &cfqd->async_cfqq[0][ioprio]; 2922 case IOPRIO_CLASS_BE: 2923 return &cfqd->async_cfqq[1][ioprio]; 2924 case IOPRIO_CLASS_IDLE: 2925 return &cfqd->async_idle_cfqq; 2926 default: 2927 BUG(); 2928 } 2929} 2930 2931static struct cfq_queue * 2932cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc, 2933 gfp_t gfp_mask) 2934{ 2935 const int ioprio = task_ioprio(ioc); 2936 const int ioprio_class = task_ioprio_class(ioc); 2937 struct cfq_queue **async_cfqq = NULL; 2938 struct cfq_queue *cfqq = NULL; 2939 2940 if (!is_sync) { 2941 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); 2942 cfqq = *async_cfqq; 2943 } 2944 2945 if (!cfqq) 2946 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); 2947 2948 /* 2949 * pin the queue now that it's allocated, scheduler exit will prune it 2950 */ 2951 if (!is_sync && !(*async_cfqq)) { 2952 cfqq->ref++; 2953 *async_cfqq = cfqq; 2954 } 2955 2956 cfqq->ref++; 2957 return cfqq; 2958} 2959 2960static void 2961__cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle) 2962{ 2963 unsigned long elapsed = jiffies - ttime->last_end_request; 2964 elapsed = min(elapsed, 2UL * slice_idle); 2965 2966 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8; 2967 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8; 2968 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples; 2969} 2970 2971static void 2972cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq, 2973 struct cfq_io_cq *cic) 2974{ 2975 if (cfq_cfqq_sync(cfqq)) { 2976 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle); 2977 __cfq_update_io_thinktime(&cfqq->service_tree->ttime, 2978 cfqd->cfq_slice_idle); 2979 } 2980#ifdef CONFIG_CFQ_GROUP_IOSCHED 2981 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle); 2982#endif 2983} 2984 2985static void 2986cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, 2987 struct request *rq) 2988{ 2989 sector_t sdist = 0; 2990 sector_t n_sec = blk_rq_sectors(rq); 2991 if (cfqq->last_request_pos) { 2992 if (cfqq->last_request_pos < blk_rq_pos(rq)) 2993 sdist = blk_rq_pos(rq) - cfqq->last_request_pos; 2994 else 2995 sdist = cfqq->last_request_pos - blk_rq_pos(rq); 2996 } 2997 2998 cfqq->seek_history <<= 1; 2999 if (blk_queue_nonrot(cfqd->queue)) 3000 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT); 3001 else 3002 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR); 3003} 3004 3005/* 3006 * Disable idle window if the process thinks too long or seeks so much that 3007 * it doesn't matter 3008 */ 3009static void 3010cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, 3011 struct cfq_io_cq *cic) 3012{ 3013 int old_idle, enable_idle; 3014 3015 /* 3016 * Don't idle for async or idle io prio class 3017 */ 3018 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) 3019 return; 3020 3021 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); 3022 3023 if (cfqq->queued[0] + cfqq->queued[1] >= 4) 3024 cfq_mark_cfqq_deep(cfqq); 3025 3026 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE)) 3027 enable_idle = 0; 3028 else if (!atomic_read(&cic->icq.ioc->nr_tasks) || 3029 !cfqd->cfq_slice_idle || 3030 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq))) 3031 enable_idle = 0; 3032 else if (sample_valid(cic->ttime.ttime_samples)) { 3033 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle) 3034 enable_idle = 0; 3035 else 3036 enable_idle = 1; 3037 } 3038 3039 if (old_idle != enable_idle) { 3040 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); 3041 if (enable_idle) 3042 cfq_mark_cfqq_idle_window(cfqq); 3043 else 3044 cfq_clear_cfqq_idle_window(cfqq); 3045 } 3046} 3047 3048/* 3049 * Check if new_cfqq should preempt the currently active queue. Return 0 for 3050 * no or if we aren't sure, a 1 will cause a preempt. 3051 */ 3052static bool 3053cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, 3054 struct request *rq) 3055{ 3056 struct cfq_queue *cfqq; 3057 3058 cfqq = cfqd->active_queue; 3059 if (!cfqq) 3060 return false; 3061 3062 if (cfq_class_idle(new_cfqq)) 3063 return false; 3064 3065 if (cfq_class_idle(cfqq)) 3066 return true; 3067 3068 /* 3069 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice. 3070 */ 3071 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq)) 3072 return false; 3073 3074 /* 3075 * if the new request is sync, but the currently running queue is 3076 * not, let the sync request have priority. 3077 */ 3078 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) 3079 return true; 3080 3081 if (new_cfqq->cfqg != cfqq->cfqg) 3082 return false; 3083 3084 if (cfq_slice_used(cfqq)) 3085 return true; 3086 3087 /* Allow preemption only if we are idling on sync-noidle tree */ 3088 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD && 3089 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD && 3090 new_cfqq->service_tree->count == 2 && 3091 RB_EMPTY_ROOT(&cfqq->sort_list)) 3092 return true; 3093 3094 /* 3095 * So both queues are sync. Let the new request get disk time if 3096 * it's a metadata request and the current queue is doing regular IO. 3097 */ 3098 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending) 3099 return true; 3100 3101 /* 3102 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. 3103 */ 3104 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) 3105 return true; 3106 3107 /* An idle queue should not be idle now for some reason */ 3108 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq)) 3109 return true; 3110 3111 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) 3112 return false; 3113 3114 /* 3115 * if this request is as-good as one we would expect from the 3116 * current cfqq, let it preempt 3117 */ 3118 if (cfq_rq_close(cfqd, cfqq, rq)) 3119 return true; 3120 3121 return false; 3122} 3123 3124/* 3125 * cfqq preempts the active queue. if we allowed preempt with no slice left, 3126 * let it have half of its nominal slice. 3127 */ 3128static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) 3129{ 3130 enum wl_type_t old_type = cfqq_type(cfqd->active_queue); 3131 3132 cfq_log_cfqq(cfqd, cfqq, "preempt"); 3133 cfq_slice_expired(cfqd, 1); 3134 3135 /* 3136 * workload type is changed, don't save slice, otherwise preempt 3137 * doesn't happen 3138 */ 3139 if (old_type != cfqq_type(cfqq)) 3140 cfqq->cfqg->saved_workload_slice = 0; 3141 3142 /* 3143 * Put the new queue at the front of the of the current list, 3144 * so we know that it will be selected next. 3145 */ 3146 BUG_ON(!cfq_cfqq_on_rr(cfqq)); 3147 3148 cfq_service_tree_add(cfqd, cfqq, 1); 3149 3150 cfqq->slice_end = 0; 3151 cfq_mark_cfqq_slice_new(cfqq); 3152} 3153 3154/* 3155 * Called when a new fs request (rq) is added (to cfqq). Check if there's 3156 * something we should do about it 3157 */ 3158static void 3159cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, 3160 struct request *rq) 3161{ 3162 struct cfq_io_cq *cic = RQ_CIC(rq); 3163 3164 cfqd->rq_queued++; 3165 if (rq->cmd_flags & REQ_PRIO) 3166 cfqq->prio_pending++; 3167 3168 cfq_update_io_thinktime(cfqd, cfqq, cic); 3169 cfq_update_io_seektime(cfqd, cfqq, rq); 3170 cfq_update_idle_window(cfqd, cfqq, cic); 3171 3172 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); 3173 3174 if (cfqq == cfqd->active_queue) { 3175 /* 3176 * Remember that we saw a request from this process, but 3177 * don't start queuing just yet. Otherwise we risk seeing lots 3178 * of tiny requests, because we disrupt the normal plugging 3179 * and merging. If the request is already larger than a single 3180 * page, let it rip immediately. For that case we assume that 3181 * merging is already done. Ditto for a busy system that 3182 * has other work pending, don't risk delaying until the 3183 * idle timer unplug to continue working. 3184 */ 3185 if (cfq_cfqq_wait_request(cfqq)) { 3186 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || 3187 cfqd->busy_queues > 1) { 3188 cfq_del_timer(cfqd, cfqq); 3189 cfq_clear_cfqq_wait_request(cfqq); 3190 __blk_run_queue(cfqd->queue); 3191 } else { 3192 cfq_blkiocg_update_idle_time_stats( 3193 &cfqq->cfqg->blkg); 3194 cfq_mark_cfqq_must_dispatch(cfqq); 3195 } 3196 } 3197 } else if (cfq_should_preempt(cfqd, cfqq, rq)) { 3198 /* 3199 * not the active queue - expire current slice if it is 3200 * idle and has expired it's mean thinktime or this new queue 3201 * has some old slice time left and is of higher priority or 3202 * this new queue is RT and the current one is BE 3203 */ 3204 cfq_preempt_queue(cfqd, cfqq); 3205 __blk_run_queue(cfqd->queue); 3206 } 3207} 3208 3209static void cfq_insert_request(struct request_queue *q, struct request *rq) 3210{ 3211 struct cfq_data *cfqd = q->elevator->elevator_data; 3212 struct cfq_queue *cfqq = RQ_CFQQ(rq); 3213 3214 cfq_log_cfqq(cfqd, cfqq, "insert_request"); 3215 cfq_init_prio_data(cfqq, RQ_CIC(rq)->icq.ioc); 3216 3217 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); 3218 list_add_tail(&rq->queuelist, &cfqq->fifo); 3219 cfq_add_rq_rb(rq); 3220 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg, 3221 &cfqd->serving_group->blkg, rq_data_dir(rq), 3222 rq_is_sync(rq)); 3223 cfq_rq_enqueued(cfqd, cfqq, rq); 3224} 3225 3226/* 3227 * Update hw_tag based on peak queue depth over 50 samples under 3228 * sufficient load. 3229 */ 3230static void cfq_update_hw_tag(struct cfq_data *cfqd) 3231{ 3232 struct cfq_queue *cfqq = cfqd->active_queue; 3233 3234 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth) 3235 cfqd->hw_tag_est_depth = cfqd->rq_in_driver; 3236 3237 if (cfqd->hw_tag == 1) 3238 return; 3239 3240 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && 3241 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN) 3242 return; 3243 3244 /* 3245 * If active queue hasn't enough requests and can idle, cfq might not 3246 * dispatch sufficient requests to hardware. Don't zero hw_tag in this 3247 * case 3248 */ 3249 if (cfqq && cfq_cfqq_idle_window(cfqq) && 3250 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < 3251 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN) 3252 return; 3253 3254 if (cfqd->hw_tag_samples++ < 50) 3255 return; 3256 3257 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN) 3258 cfqd->hw_tag = 1; 3259 else 3260 cfqd->hw_tag = 0; 3261} 3262 3263static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq) 3264{ 3265 struct cfq_io_cq *cic = cfqd->active_cic; 3266 3267 /* If the queue already has requests, don't wait */ 3268 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) 3269 return false; 3270 3271 /* If there are other queues in the group, don't wait */ 3272 if (cfqq->cfqg->nr_cfqq > 1) 3273 return false; 3274 3275 /* the only queue in the group, but think time is big */ 3276 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) 3277 return false; 3278 3279 if (cfq_slice_used(cfqq)) 3280 return true; 3281 3282 /* if slice left is less than think time, wait busy */ 3283 if (cic && sample_valid(cic->ttime.ttime_samples) 3284 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) 3285 return true; 3286 3287 /* 3288 * If think times is less than a jiffy than ttime_mean=0 and above 3289 * will not be true. It might happen that slice has not expired yet 3290 * but will expire soon (4-5 ns) during select_queue(). To cover the 3291 * case where think time is less than a jiffy, mark the queue wait 3292 * busy if only 1 jiffy is left in the slice. 3293 */ 3294 if (cfqq->slice_end - jiffies == 1) 3295 return true; 3296 3297 return false; 3298} 3299 3300static void cfq_completed_request(struct request_queue *q, struct request *rq) 3301{ 3302 struct cfq_queue *cfqq = RQ_CFQQ(rq); 3303 struct cfq_data *cfqd = cfqq->cfqd; 3304 const int sync = rq_is_sync(rq); 3305 unsigned long now; 3306 3307 now = jiffies; 3308 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", 3309 !!(rq->cmd_flags & REQ_NOIDLE)); 3310 3311 cfq_update_hw_tag(cfqd); 3312 3313 WARN_ON(!cfqd->rq_in_driver); 3314 WARN_ON(!cfqq->dispatched); 3315 cfqd->rq_in_driver--; 3316 cfqq->dispatched--; 3317 (RQ_CFQG(rq))->dispatched--; 3318 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg, 3319 rq_start_time_ns(rq), rq_io_start_time_ns(rq), 3320 rq_data_dir(rq), rq_is_sync(rq)); 3321 3322 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--; 3323 3324 if (sync) { 3325 struct cfq_rb_root *service_tree; 3326 3327 RQ_CIC(rq)->ttime.last_end_request = now; 3328 3329 if (cfq_cfqq_on_rr(cfqq)) 3330 service_tree = cfqq->service_tree; 3331 else 3332 service_tree = service_tree_for(cfqq->cfqg, 3333 cfqq_prio(cfqq), cfqq_type(cfqq)); 3334 service_tree->ttime.last_end_request = now; 3335 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now)) 3336 cfqd->last_delayed_sync = now; 3337 } 3338 3339#ifdef CONFIG_CFQ_GROUP_IOSCHED 3340 cfqq->cfqg->ttime.last_end_request = now; 3341#endif 3342 3343 /* 3344 * If this is the active queue, check if it needs to be expired, 3345 * or if we want to idle in case it has no pending requests. 3346 */ 3347 if (cfqd->active_queue == cfqq) { 3348 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); 3349 3350 if (cfq_cfqq_slice_new(cfqq)) { 3351 cfq_set_prio_slice(cfqd, cfqq); 3352 cfq_clear_cfqq_slice_new(cfqq); 3353 } 3354 3355 /* 3356 * Should we wait for next request to come in before we expire 3357 * the queue. 3358 */ 3359 if (cfq_should_wait_busy(cfqd, cfqq)) { 3360 unsigned long extend_sl = cfqd->cfq_slice_idle; 3361 if (!cfqd->cfq_slice_idle) 3362 extend_sl = cfqd->cfq_group_idle; 3363 cfqq->slice_end = jiffies + extend_sl; 3364 cfq_mark_cfqq_wait_busy(cfqq); 3365 cfq_log_cfqq(cfqd, cfqq, "will busy wait"); 3366 } 3367 3368 /* 3369 * Idling is not enabled on: 3370 * - expired queues 3371 * - idle-priority queues 3372 * - async queues 3373 * - queues with still some requests queued 3374 * - when there is a close cooperator 3375 */ 3376 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) 3377 cfq_slice_expired(cfqd, 1); 3378 else if (sync && cfqq_empty && 3379 !cfq_close_cooperator(cfqd, cfqq)) { 3380 cfq_arm_slice_timer(cfqd); 3381 } 3382 } 3383 3384 if (!cfqd->rq_in_driver) 3385 cfq_schedule_dispatch(cfqd); 3386} 3387 3388static inline int __cfq_may_queue(struct cfq_queue *cfqq) 3389{ 3390 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { 3391 cfq_mark_cfqq_must_alloc_slice(cfqq); 3392 return ELV_MQUEUE_MUST; 3393 } 3394 3395 return ELV_MQUEUE_MAY; 3396} 3397 3398static int cfq_may_queue(struct request_queue *q, int rw) 3399{ 3400 struct cfq_data *cfqd = q->elevator->elevator_data; 3401 struct task_struct *tsk = current; 3402 struct cfq_io_cq *cic; 3403 struct cfq_queue *cfqq; 3404 3405 /* 3406 * don't force setup of a queue from here, as a call to may_queue 3407 * does not necessarily imply that a request actually will be queued. 3408 * so just lookup a possibly existing queue, or return 'may queue' 3409 * if that fails 3410 */ 3411 cic = cfq_cic_lookup(cfqd, tsk->io_context); 3412 if (!cic) 3413 return ELV_MQUEUE_MAY; 3414 3415 cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); 3416 if (cfqq) { 3417 cfq_init_prio_data(cfqq, cic->icq.ioc); 3418 3419 return __cfq_may_queue(cfqq); 3420 } 3421 3422 return ELV_MQUEUE_MAY; 3423} 3424 3425/* 3426 * queue lock held here 3427 */ 3428static void cfq_put_request(struct request *rq) 3429{ 3430 struct cfq_queue *cfqq = RQ_CFQQ(rq); 3431 3432 if (cfqq) { 3433 const int rw = rq_data_dir(rq); 3434 3435 BUG_ON(!cfqq->allocated[rw]); 3436 cfqq->allocated[rw]--; 3437 3438 /* Put down rq reference on cfqg */ 3439 cfq_put_cfqg(RQ_CFQG(rq)); 3440 rq->elv.priv[0] = NULL; 3441 rq->elv.priv[1] = NULL; 3442 3443 cfq_put_queue(cfqq); 3444 } 3445} 3446 3447static struct cfq_queue * 3448cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic, 3449 struct cfq_queue *cfqq) 3450{ 3451 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); 3452 cic_set_cfqq(cic, cfqq->new_cfqq, 1); 3453 cfq_mark_cfqq_coop(cfqq->new_cfqq); 3454 cfq_put_queue(cfqq); 3455 return cic_to_cfqq(cic, 1); 3456} 3457 3458/* 3459 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this 3460 * was the last process referring to said cfqq. 3461 */ 3462static struct cfq_queue * 3463split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq) 3464{ 3465 if (cfqq_process_refs(cfqq) == 1) { 3466 cfqq->pid = current->pid; 3467 cfq_clear_cfqq_coop(cfqq); 3468 cfq_clear_cfqq_split_coop(cfqq); 3469 return cfqq; 3470 } 3471 3472 cic_set_cfqq(cic, NULL, 1); 3473 3474 cfq_put_cooperator(cfqq); 3475 3476 cfq_put_queue(cfqq); 3477 return NULL; 3478} 3479/* 3480 * Allocate cfq data structures associated with this request. 3481 */ 3482static int 3483cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) 3484{ 3485 struct cfq_data *cfqd = q->elevator->elevator_data; 3486 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq); 3487 const int rw = rq_data_dir(rq); 3488 const bool is_sync = rq_is_sync(rq); 3489 struct cfq_queue *cfqq; 3490 unsigned int changed; 3491 3492 might_sleep_if(gfp_mask & __GFP_WAIT); 3493 3494 spin_lock_irq(q->queue_lock); 3495 3496 /* handle changed notifications */ 3497 changed = icq_get_changed(&cic->icq); 3498 if (unlikely(changed & ICQ_IOPRIO_CHANGED)) 3499 changed_ioprio(cic); 3500#ifdef CONFIG_CFQ_GROUP_IOSCHED 3501 if (unlikely(changed & ICQ_CGROUP_CHANGED)) 3502 changed_cgroup(cic); 3503#endif 3504 3505new_queue: 3506 cfqq = cic_to_cfqq(cic, is_sync); 3507 if (!cfqq || cfqq == &cfqd->oom_cfqq) { 3508 cfqq = cfq_get_queue(cfqd, is_sync, cic->icq.ioc, gfp_mask); 3509 cic_set_cfqq(cic, cfqq, is_sync); 3510 } else { 3511 /* 3512 * If the queue was seeky for too long, break it apart. 3513 */ 3514 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) { 3515 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); 3516 cfqq = split_cfqq(cic, cfqq); 3517 if (!cfqq) 3518 goto new_queue; 3519 } 3520 3521 /* 3522 * Check to see if this queue is scheduled to merge with 3523 * another, closely cooperating queue. The merging of 3524 * queues happens here as it must be done in process context. 3525 * The reference on new_cfqq was taken in merge_cfqqs. 3526 */ 3527 if (cfqq->new_cfqq) 3528 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); 3529 } 3530 3531 cfqq->allocated[rw]++; 3532 3533 cfqq->ref++; 3534 rq->elv.priv[0] = cfqq; 3535 rq->elv.priv[1] = cfq_ref_get_cfqg(cfqq->cfqg); 3536 spin_unlock_irq(q->queue_lock); 3537 return 0; 3538} 3539 3540static void cfq_kick_queue(struct work_struct *work) 3541{ 3542 struct cfq_data *cfqd = 3543 container_of(work, struct cfq_data, unplug_work); 3544 struct request_queue *q = cfqd->queue; 3545 3546 spin_lock_irq(q->queue_lock); 3547 __blk_run_queue(cfqd->queue); 3548 spin_unlock_irq(q->queue_lock); 3549} 3550 3551/* 3552 * Timer running if the active_queue is currently idling inside its time slice 3553 */ 3554static void cfq_idle_slice_timer(unsigned long data) 3555{ 3556 struct cfq_data *cfqd = (struct cfq_data *) data; 3557 struct cfq_queue *cfqq; 3558 unsigned long flags; 3559 int timed_out = 1; 3560 3561 cfq_log(cfqd, "idle timer fired"); 3562 3563 spin_lock_irqsave(cfqd->queue->queue_lock, flags); 3564 3565 cfqq = cfqd->active_queue; 3566 if (cfqq) { 3567 timed_out = 0; 3568 3569 /* 3570 * We saw a request before the queue expired, let it through 3571 */ 3572 if (cfq_cfqq_must_dispatch(cfqq)) 3573 goto out_kick; 3574 3575 /* 3576 * expired 3577 */ 3578 if (cfq_slice_used(cfqq)) 3579 goto expire; 3580 3581 /* 3582 * only expire and reinvoke request handler, if there are 3583 * other queues with pending requests 3584 */ 3585 if (!cfqd->busy_queues) 3586 goto out_cont; 3587 3588 /* 3589 * not expired and it has a request pending, let it dispatch 3590 */ 3591 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) 3592 goto out_kick; 3593 3594 /* 3595 * Queue depth flag is reset only when the idle didn't succeed 3596 */ 3597 cfq_clear_cfqq_deep(cfqq); 3598 } 3599expire: 3600 cfq_slice_expired(cfqd, timed_out); 3601out_kick: 3602 cfq_schedule_dispatch(cfqd); 3603out_cont: 3604 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); 3605} 3606 3607static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) 3608{ 3609 del_timer_sync(&cfqd->idle_slice_timer); 3610 cancel_work_sync(&cfqd->unplug_work); 3611} 3612 3613static void cfq_put_async_queues(struct cfq_data *cfqd) 3614{ 3615 int i; 3616 3617 for (i = 0; i < IOPRIO_BE_NR; i++) { 3618 if (cfqd->async_cfqq[0][i]) 3619 cfq_put_queue(cfqd->async_cfqq[0][i]); 3620 if (cfqd->async_cfqq[1][i]) 3621 cfq_put_queue(cfqd->async_cfqq[1][i]); 3622 } 3623 3624 if (cfqd->async_idle_cfqq) 3625 cfq_put_queue(cfqd->async_idle_cfqq); 3626} 3627 3628static void cfq_exit_queue(struct elevator_queue *e) 3629{ 3630 struct cfq_data *cfqd = e->elevator_data; 3631 struct request_queue *q = cfqd->queue; 3632 bool wait = false; 3633 3634 cfq_shutdown_timer_wq(cfqd); 3635 3636 spin_lock_irq(q->queue_lock); 3637 3638 if (cfqd->active_queue) 3639 __cfq_slice_expired(cfqd, cfqd->active_queue, 0); 3640 3641 cfq_put_async_queues(cfqd); 3642 cfq_release_cfq_groups(cfqd); 3643 3644 /* 3645 * If there are groups which we could not unlink from blkcg list, 3646 * wait for a rcu period for them to be freed. 3647 */ 3648 if (cfqd->nr_blkcg_linked_grps) 3649 wait = true; 3650 3651 spin_unlock_irq(q->queue_lock); 3652 3653 cfq_shutdown_timer_wq(cfqd); 3654 3655 /* 3656 * Wait for cfqg->blkg->key accessors to exit their grace periods. 3657 * Do this wait only if there are other unlinked groups out 3658 * there. This can happen if cgroup deletion path claimed the 3659 * responsibility of cleaning up a group before queue cleanup code 3660 * get to the group. 3661 * 3662 * Do not call synchronize_rcu() unconditionally as there are drivers 3663 * which create/delete request queue hundreds of times during scan/boot 3664 * and synchronize_rcu() can take significant time and slow down boot. 3665 */ 3666 if (wait) 3667 synchronize_rcu(); 3668 3669#ifdef CONFIG_CFQ_GROUP_IOSCHED 3670 /* Free up per cpu stats for root group */ 3671 free_percpu(cfqd->root_group.blkg.stats_cpu); 3672#endif 3673 kfree(cfqd); 3674} 3675 3676static int cfq_init_queue(struct request_queue *q) 3677{ 3678 struct cfq_data *cfqd; 3679 int i, j; 3680 struct cfq_group *cfqg; 3681 struct cfq_rb_root *st; 3682 3683 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); 3684 if (!cfqd) 3685 return -ENOMEM; 3686 3687 /* Init root service tree */ 3688 cfqd->grp_service_tree = CFQ_RB_ROOT; 3689 3690 /* Init root group */ 3691 cfqg = &cfqd->root_group; 3692 for_each_cfqg_st(cfqg, i, j, st) 3693 *st = CFQ_RB_ROOT; 3694 RB_CLEAR_NODE(&cfqg->rb_node); 3695 3696 /* Give preference to root group over other groups */ 3697 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT; 3698 3699#ifdef CONFIG_CFQ_GROUP_IOSCHED 3700 /* 3701 * Set root group reference to 2. One reference will be dropped when 3702 * all groups on cfqd->cfqg_list are being deleted during queue exit. 3703 * Other reference will remain there as we don't want to delete this 3704 * group as it is statically allocated and gets destroyed when 3705 * throtl_data goes away. 3706 */ 3707 cfqg->ref = 2; 3708 3709 if (blkio_alloc_blkg_stats(&cfqg->blkg)) { 3710 kfree(cfqg); 3711 kfree(cfqd); 3712 return -ENOMEM; 3713 } 3714 3715 rcu_read_lock(); 3716 3717 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, 3718 (void *)cfqd, 0); 3719 rcu_read_unlock(); 3720 cfqd->nr_blkcg_linked_grps++; 3721 3722 /* Add group on cfqd->cfqg_list */ 3723 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list); 3724#endif 3725 /* 3726 * Not strictly needed (since RB_ROOT just clears the node and we 3727 * zeroed cfqd on alloc), but better be safe in case someone decides 3728 * to add magic to the rb code 3729 */ 3730 for (i = 0; i < CFQ_PRIO_LISTS; i++) 3731 cfqd->prio_trees[i] = RB_ROOT; 3732 3733 /* 3734 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. 3735 * Grab a permanent reference to it, so that the normal code flow 3736 * will not attempt to free it. 3737 */ 3738 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); 3739 cfqd->oom_cfqq.ref++; 3740 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group); 3741 3742 cfqd->queue = q; 3743 q->elevator->elevator_data = cfqd; 3744 3745 init_timer(&cfqd->idle_slice_timer); 3746 cfqd->idle_slice_timer.function = cfq_idle_slice_timer; 3747 cfqd->idle_slice_timer.data = (unsigned long) cfqd; 3748 3749 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); 3750 3751 cfqd->cfq_quantum = cfq_quantum; 3752 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; 3753 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; 3754 cfqd->cfq_back_max = cfq_back_max; 3755 cfqd->cfq_back_penalty = cfq_back_penalty; 3756 cfqd->cfq_slice[0] = cfq_slice_async; 3757 cfqd->cfq_slice[1] = cfq_slice_sync; 3758 cfqd->cfq_slice_async_rq = cfq_slice_async_rq; 3759 cfqd->cfq_slice_idle = cfq_slice_idle; 3760 cfqd->cfq_group_idle = cfq_group_idle; 3761 cfqd->cfq_latency = 1; 3762 cfqd->hw_tag = -1; 3763 /* 3764 * we optimistically start assuming sync ops weren't delayed in last 3765 * second, in order to have larger depth for async operations. 3766 */ 3767 cfqd->last_delayed_sync = jiffies - HZ; 3768 return 0; 3769} 3770 3771/* 3772 * sysfs parts below --> 3773 */ 3774static ssize_t 3775cfq_var_show(unsigned int var, char *page) 3776{ 3777 return sprintf(page, "%d\n", var); 3778} 3779 3780static ssize_t 3781cfq_var_store(unsigned int *var, const char *page, size_t count) 3782{ 3783 char *p = (char *) page; 3784 3785 *var = simple_strtoul(p, &p, 10); 3786 return count; 3787} 3788 3789#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ 3790static ssize_t __FUNC(struct elevator_queue *e, char *page) \ 3791{ \ 3792 struct cfq_data *cfqd = e->elevator_data; \ 3793 unsigned int __data = __VAR; \ 3794 if (__CONV) \ 3795 __data = jiffies_to_msecs(__data); \ 3796 return cfq_var_show(__data, (page)); \ 3797} 3798SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); 3799SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); 3800SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); 3801SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); 3802SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); 3803SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); 3804SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1); 3805SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); 3806SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); 3807SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); 3808SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); 3809#undef SHOW_FUNCTION 3810 3811#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ 3812static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ 3813{ \ 3814 struct cfq_data *cfqd = e->elevator_data; \ 3815 unsigned int __data; \ 3816 int ret = cfq_var_store(&__data, (page), count); \ 3817 if (__data < (MIN)) \ 3818 __data = (MIN); \ 3819 else if (__data > (MAX)) \ 3820 __data = (MAX); \ 3821 if (__CONV) \ 3822 *(__PTR) = msecs_to_jiffies(__data); \ 3823 else \ 3824 *(__PTR) = __data; \ 3825 return ret; \ 3826} 3827STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); 3828STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, 3829 UINT_MAX, 1); 3830STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, 3831 UINT_MAX, 1); 3832STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); 3833STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, 3834 UINT_MAX, 0); 3835STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); 3836STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1); 3837STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); 3838STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); 3839STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, 3840 UINT_MAX, 0); 3841STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); 3842#undef STORE_FUNCTION 3843 3844#define CFQ_ATTR(name) \ 3845 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) 3846 3847static struct elv_fs_entry cfq_attrs[] = { 3848 CFQ_ATTR(quantum), 3849 CFQ_ATTR(fifo_expire_sync), 3850 CFQ_ATTR(fifo_expire_async), 3851 CFQ_ATTR(back_seek_max), 3852 CFQ_ATTR(back_seek_penalty), 3853 CFQ_ATTR(slice_sync), 3854 CFQ_ATTR(slice_async), 3855 CFQ_ATTR(slice_async_rq), 3856 CFQ_ATTR(slice_idle), 3857 CFQ_ATTR(group_idle), 3858 CFQ_ATTR(low_latency), 3859 __ATTR_NULL 3860}; 3861 3862static struct elevator_type iosched_cfq = { 3863 .ops = { 3864 .elevator_merge_fn = cfq_merge, 3865 .elevator_merged_fn = cfq_merged_request, 3866 .elevator_merge_req_fn = cfq_merged_requests, 3867 .elevator_allow_merge_fn = cfq_allow_merge, 3868 .elevator_bio_merged_fn = cfq_bio_merged, 3869 .elevator_dispatch_fn = cfq_dispatch_requests, 3870 .elevator_add_req_fn = cfq_insert_request, 3871 .elevator_activate_req_fn = cfq_activate_request, 3872 .elevator_deactivate_req_fn = cfq_deactivate_request, 3873 .elevator_completed_req_fn = cfq_completed_request, 3874 .elevator_former_req_fn = elv_rb_former_request, 3875 .elevator_latter_req_fn = elv_rb_latter_request, 3876 .elevator_init_icq_fn = cfq_init_icq, 3877 .elevator_exit_icq_fn = cfq_exit_icq, 3878 .elevator_set_req_fn = cfq_set_request, 3879 .elevator_put_req_fn = cfq_put_request, 3880 .elevator_may_queue_fn = cfq_may_queue, 3881 .elevator_init_fn = cfq_init_queue, 3882 .elevator_exit_fn = cfq_exit_queue, 3883 }, 3884 .icq_size = sizeof(struct cfq_io_cq), 3885 .icq_align = __alignof__(struct cfq_io_cq), 3886 .elevator_attrs = cfq_attrs, 3887 .elevator_name = "cfq", 3888 .elevator_owner = THIS_MODULE, 3889}; 3890 3891#ifdef CONFIG_CFQ_GROUP_IOSCHED 3892static struct blkio_policy_type blkio_policy_cfq = { 3893 .ops = { 3894 .blkio_unlink_group_fn = cfq_unlink_blkio_group, 3895 .blkio_clear_queue_fn = cfq_clear_queue, 3896 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight, 3897 }, 3898 .plid = BLKIO_POLICY_PROP, 3899}; 3900#endif 3901 3902static int __init cfq_init(void) 3903{ 3904 int ret; 3905 3906 /* 3907 * could be 0 on HZ < 1000 setups 3908 */ 3909 if (!cfq_slice_async) 3910 cfq_slice_async = 1; 3911 if (!cfq_slice_idle) 3912 cfq_slice_idle = 1; 3913 3914#ifdef CONFIG_CFQ_GROUP_IOSCHED 3915 if (!cfq_group_idle) 3916 cfq_group_idle = 1; 3917#else 3918 cfq_group_idle = 0; 3919#endif 3920 cfq_pool = KMEM_CACHE(cfq_queue, 0); 3921 if (!cfq_pool) 3922 return -ENOMEM; 3923 3924 ret = elv_register(&iosched_cfq); 3925 if (ret) { 3926 kmem_cache_destroy(cfq_pool); 3927 return ret; 3928 } 3929 3930#ifdef CONFIG_CFQ_GROUP_IOSCHED 3931 blkio_policy_register(&blkio_policy_cfq); 3932#endif 3933 return 0; 3934} 3935 3936static void __exit cfq_exit(void) 3937{ 3938#ifdef CONFIG_CFQ_GROUP_IOSCHED 3939 blkio_policy_unregister(&blkio_policy_cfq); 3940#endif 3941 elv_unregister(&iosched_cfq); 3942 kmem_cache_destroy(cfq_pool); 3943} 3944 3945module_init(cfq_init); 3946module_exit(cfq_exit); 3947 3948MODULE_AUTHOR("Jens Axboe"); 3949MODULE_LICENSE("GPL"); 3950MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); 3951