cfq-iosched.c revision dc86900e0a8f665122de6faadd27fb4c6d2b3e4d
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) \ 58 ((struct cfq_io_context *) (rq)->elevator_private[0]) 59#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private[1]) 60#define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private[2]) 61 62static struct kmem_cache *cfq_pool; 63static struct kmem_cache *cfq_ioc_pool; 64 65static DEFINE_PER_CPU(unsigned long, cfq_ioc_count); 66static struct completion *ioc_gone; 67static DEFINE_SPINLOCK(ioc_gone_lock); 68 69#define CFQ_PRIO_LISTS IOPRIO_BE_NR 70#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) 71#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) 72 73#define sample_valid(samples) ((samples) > 80) 74#define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node) 75 76/* 77 * Most of our rbtree usage is for sorting with min extraction, so 78 * if we cache the leftmost node we don't have to walk down the tree 79 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should 80 * move this into the elevator for the rq sorting as well. 81 */ 82struct cfq_rb_root { 83 struct rb_root rb; 84 struct rb_node *left; 85 unsigned count; 86 unsigned total_weight; 87 u64 min_vdisktime; 88 struct cfq_ttime ttime; 89}; 90#define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \ 91 .ttime = {.last_end_request = jiffies,},} 92 93/* 94 * Per process-grouping structure 95 */ 96struct cfq_queue { 97 /* reference count */ 98 int ref; 99 /* various state flags, see below */ 100 unsigned int flags; 101 /* parent cfq_data */ 102 struct cfq_data *cfqd; 103 /* service_tree member */ 104 struct rb_node rb_node; 105 /* service_tree key */ 106 unsigned long rb_key; 107 /* prio tree member */ 108 struct rb_node p_node; 109 /* prio tree root we belong to, if any */ 110 struct rb_root *p_root; 111 /* sorted list of pending requests */ 112 struct rb_root sort_list; 113 /* if fifo isn't expired, next request to serve */ 114 struct request *next_rq; 115 /* requests queued in sort_list */ 116 int queued[2]; 117 /* currently allocated requests */ 118 int allocated[2]; 119 /* fifo list of requests in sort_list */ 120 struct list_head fifo; 121 122 /* time when queue got scheduled in to dispatch first request. */ 123 unsigned long dispatch_start; 124 unsigned int allocated_slice; 125 unsigned int slice_dispatch; 126 /* time when first request from queue completed and slice started. */ 127 unsigned long slice_start; 128 unsigned long slice_end; 129 long slice_resid; 130 131 /* pending priority requests */ 132 int prio_pending; 133 /* number of requests that are on the dispatch list or inside driver */ 134 int dispatched; 135 136 /* io prio of this group */ 137 unsigned short ioprio, org_ioprio; 138 unsigned short ioprio_class; 139 140 pid_t pid; 141 142 u32 seek_history; 143 sector_t last_request_pos; 144 145 struct cfq_rb_root *service_tree; 146 struct cfq_queue *new_cfqq; 147 struct cfq_group *cfqg; 148 /* Number of sectors dispatched from queue in single dispatch round */ 149 unsigned long nr_sectors; 150}; 151 152/* 153 * First index in the service_trees. 154 * IDLE is handled separately, so it has negative index 155 */ 156enum wl_prio_t { 157 BE_WORKLOAD = 0, 158 RT_WORKLOAD = 1, 159 IDLE_WORKLOAD = 2, 160 CFQ_PRIO_NR, 161}; 162 163/* 164 * Second index in the service_trees. 165 */ 166enum wl_type_t { 167 ASYNC_WORKLOAD = 0, 168 SYNC_NOIDLE_WORKLOAD = 1, 169 SYNC_WORKLOAD = 2 170}; 171 172/* This is per cgroup per device grouping structure */ 173struct cfq_group { 174 /* group service_tree member */ 175 struct rb_node rb_node; 176 177 /* group service_tree key */ 178 u64 vdisktime; 179 unsigned int weight; 180 unsigned int new_weight; 181 bool needs_update; 182 183 /* number of cfqq currently on this group */ 184 int nr_cfqq; 185 186 /* 187 * Per group busy queues average. Useful for workload slice calc. We 188 * create the array for each prio class but at run time it is used 189 * only for RT and BE class and slot for IDLE class remains unused. 190 * This is primarily done to avoid confusion and a gcc warning. 191 */ 192 unsigned int busy_queues_avg[CFQ_PRIO_NR]; 193 /* 194 * rr lists of queues with requests. We maintain service trees for 195 * RT and BE classes. These trees are subdivided in subclasses 196 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE 197 * class there is no subclassification and all the cfq queues go on 198 * a single tree service_tree_idle. 199 * Counts are embedded in the cfq_rb_root 200 */ 201 struct cfq_rb_root service_trees[2][3]; 202 struct cfq_rb_root service_tree_idle; 203 204 unsigned long saved_workload_slice; 205 enum wl_type_t saved_workload; 206 enum wl_prio_t saved_serving_prio; 207 struct blkio_group blkg; 208#ifdef CONFIG_CFQ_GROUP_IOSCHED 209 struct hlist_node cfqd_node; 210 int ref; 211#endif 212 /* number of requests that are on the dispatch list or inside driver */ 213 int dispatched; 214 struct cfq_ttime ttime; 215}; 216 217/* 218 * Per block device queue structure 219 */ 220struct cfq_data { 221 struct request_queue *queue; 222 /* Root service tree for cfq_groups */ 223 struct cfq_rb_root grp_service_tree; 224 struct cfq_group root_group; 225 226 /* 227 * The priority currently being served 228 */ 229 enum wl_prio_t serving_prio; 230 enum wl_type_t serving_type; 231 unsigned long workload_expires; 232 struct cfq_group *serving_group; 233 234 /* 235 * Each priority tree is sorted by next_request position. These 236 * trees are used when determining if two or more queues are 237 * interleaving requests (see cfq_close_cooperator). 238 */ 239 struct rb_root prio_trees[CFQ_PRIO_LISTS]; 240 241 unsigned int busy_queues; 242 unsigned int busy_sync_queues; 243 244 int rq_in_driver; 245 int rq_in_flight[2]; 246 247 /* 248 * queue-depth detection 249 */ 250 int rq_queued; 251 int hw_tag; 252 /* 253 * hw_tag can be 254 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection) 255 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth) 256 * 0 => no NCQ 257 */ 258 int hw_tag_est_depth; 259 unsigned int hw_tag_samples; 260 261 /* 262 * idle window management 263 */ 264 struct timer_list idle_slice_timer; 265 struct work_struct unplug_work; 266 267 struct cfq_queue *active_queue; 268 struct cfq_io_context *active_cic; 269 270 /* 271 * async queue for each priority case 272 */ 273 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; 274 struct cfq_queue *async_idle_cfqq; 275 276 sector_t last_position; 277 278 /* 279 * tunables, see top of file 280 */ 281 unsigned int cfq_quantum; 282 unsigned int cfq_fifo_expire[2]; 283 unsigned int cfq_back_penalty; 284 unsigned int cfq_back_max; 285 unsigned int cfq_slice[2]; 286 unsigned int cfq_slice_async_rq; 287 unsigned int cfq_slice_idle; 288 unsigned int cfq_group_idle; 289 unsigned int cfq_latency; 290 291 struct list_head cic_list; 292 293 /* 294 * Fallback dummy cfqq for extreme OOM conditions 295 */ 296 struct cfq_queue oom_cfqq; 297 298 unsigned long last_delayed_sync; 299 300 /* List of cfq groups being managed on this device*/ 301 struct hlist_head cfqg_list; 302 303 /* Number of groups which are on blkcg->blkg_list */ 304 unsigned int nr_blkcg_linked_grps; 305}; 306 307static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd); 308 309static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg, 310 enum wl_prio_t prio, 311 enum wl_type_t type) 312{ 313 if (!cfqg) 314 return NULL; 315 316 if (prio == IDLE_WORKLOAD) 317 return &cfqg->service_tree_idle; 318 319 return &cfqg->service_trees[prio][type]; 320} 321 322enum cfqq_state_flags { 323 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ 324 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ 325 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ 326 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ 327 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ 328 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ 329 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ 330 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ 331 CFQ_CFQQ_FLAG_sync, /* synchronous queue */ 332 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */ 333 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */ 334 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */ 335 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */ 336}; 337 338#define CFQ_CFQQ_FNS(name) \ 339static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ 340{ \ 341 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ 342} \ 343static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ 344{ \ 345 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ 346} \ 347static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ 348{ \ 349 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ 350} 351 352CFQ_CFQQ_FNS(on_rr); 353CFQ_CFQQ_FNS(wait_request); 354CFQ_CFQQ_FNS(must_dispatch); 355CFQ_CFQQ_FNS(must_alloc_slice); 356CFQ_CFQQ_FNS(fifo_expire); 357CFQ_CFQQ_FNS(idle_window); 358CFQ_CFQQ_FNS(prio_changed); 359CFQ_CFQQ_FNS(slice_new); 360CFQ_CFQQ_FNS(sync); 361CFQ_CFQQ_FNS(coop); 362CFQ_CFQQ_FNS(split_coop); 363CFQ_CFQQ_FNS(deep); 364CFQ_CFQQ_FNS(wait_busy); 365#undef CFQ_CFQQ_FNS 366 367#ifdef CONFIG_CFQ_GROUP_IOSCHED 368#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ 369 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \ 370 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \ 371 blkg_path(&(cfqq)->cfqg->blkg), ##args) 372 373#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \ 374 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \ 375 blkg_path(&(cfqg)->blkg), ##args) \ 376 377#else 378#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ 379 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) 380#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0) 381#endif 382#define cfq_log(cfqd, fmt, args...) \ 383 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) 384 385/* Traverses through cfq group service trees */ 386#define for_each_cfqg_st(cfqg, i, j, st) \ 387 for (i = 0; i <= IDLE_WORKLOAD; i++) \ 388 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\ 389 : &cfqg->service_tree_idle; \ 390 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \ 391 (i == IDLE_WORKLOAD && j == 0); \ 392 j++, st = i < IDLE_WORKLOAD ? \ 393 &cfqg->service_trees[i][j]: NULL) \ 394 395static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd, 396 struct cfq_ttime *ttime, bool group_idle) 397{ 398 unsigned long slice; 399 if (!sample_valid(ttime->ttime_samples)) 400 return false; 401 if (group_idle) 402 slice = cfqd->cfq_group_idle; 403 else 404 slice = cfqd->cfq_slice_idle; 405 return ttime->ttime_mean > slice; 406} 407 408static inline bool iops_mode(struct cfq_data *cfqd) 409{ 410 /* 411 * If we are not idling on queues and it is a NCQ drive, parallel 412 * execution of requests is on and measuring time is not possible 413 * in most of the cases until and unless we drive shallower queue 414 * depths and that becomes a performance bottleneck. In such cases 415 * switch to start providing fairness in terms of number of IOs. 416 */ 417 if (!cfqd->cfq_slice_idle && cfqd->hw_tag) 418 return true; 419 else 420 return false; 421} 422 423static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq) 424{ 425 if (cfq_class_idle(cfqq)) 426 return IDLE_WORKLOAD; 427 if (cfq_class_rt(cfqq)) 428 return RT_WORKLOAD; 429 return BE_WORKLOAD; 430} 431 432 433static enum wl_type_t cfqq_type(struct cfq_queue *cfqq) 434{ 435 if (!cfq_cfqq_sync(cfqq)) 436 return ASYNC_WORKLOAD; 437 if (!cfq_cfqq_idle_window(cfqq)) 438 return SYNC_NOIDLE_WORKLOAD; 439 return SYNC_WORKLOAD; 440} 441 442static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl, 443 struct cfq_data *cfqd, 444 struct cfq_group *cfqg) 445{ 446 if (wl == IDLE_WORKLOAD) 447 return cfqg->service_tree_idle.count; 448 449 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count 450 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count 451 + cfqg->service_trees[wl][SYNC_WORKLOAD].count; 452} 453 454static inline int cfqg_busy_async_queues(struct cfq_data *cfqd, 455 struct cfq_group *cfqg) 456{ 457 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count 458 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count; 459} 460 461static void cfq_dispatch_insert(struct request_queue *, struct request *); 462static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool, 463 struct io_context *, gfp_t); 464static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, 465 struct io_context *); 466 467static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, 468 bool is_sync) 469{ 470 return cic->cfqq[is_sync]; 471} 472 473static inline void cic_set_cfqq(struct cfq_io_context *cic, 474 struct cfq_queue *cfqq, bool is_sync) 475{ 476 cic->cfqq[is_sync] = cfqq; 477} 478 479#define CIC_DEAD_KEY 1ul 480#define CIC_DEAD_INDEX_SHIFT 1 481 482static inline void *cfqd_dead_key(struct cfq_data *cfqd) 483{ 484 return (void *)(cfqd->queue->id << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY); 485} 486 487static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic) 488{ 489 struct cfq_data *cfqd = cic->key; 490 491 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY)) 492 return NULL; 493 494 return cfqd; 495} 496 497/* 498 * We regard a request as SYNC, if it's either a read or has the SYNC bit 499 * set (in which case it could also be direct WRITE). 500 */ 501static inline bool cfq_bio_sync(struct bio *bio) 502{ 503 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC); 504} 505 506/* 507 * scheduler run of queue, if there are requests pending and no one in the 508 * driver that will restart queueing 509 */ 510static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) 511{ 512 if (cfqd->busy_queues) { 513 cfq_log(cfqd, "schedule dispatch"); 514 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); 515 } 516} 517 518/* 519 * Scale schedule slice based on io priority. Use the sync time slice only 520 * if a queue is marked sync and has sync io queued. A sync queue with async 521 * io only, should not get full sync slice length. 522 */ 523static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync, 524 unsigned short prio) 525{ 526 const int base_slice = cfqd->cfq_slice[sync]; 527 528 WARN_ON(prio >= IOPRIO_BE_NR); 529 530 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); 531} 532 533static inline int 534cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) 535{ 536 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); 537} 538 539static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg) 540{ 541 u64 d = delta << CFQ_SERVICE_SHIFT; 542 543 d = d * BLKIO_WEIGHT_DEFAULT; 544 do_div(d, cfqg->weight); 545 return d; 546} 547 548static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime) 549{ 550 s64 delta = (s64)(vdisktime - min_vdisktime); 551 if (delta > 0) 552 min_vdisktime = vdisktime; 553 554 return min_vdisktime; 555} 556 557static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime) 558{ 559 s64 delta = (s64)(vdisktime - min_vdisktime); 560 if (delta < 0) 561 min_vdisktime = vdisktime; 562 563 return min_vdisktime; 564} 565 566static void update_min_vdisktime(struct cfq_rb_root *st) 567{ 568 struct cfq_group *cfqg; 569 570 if (st->left) { 571 cfqg = rb_entry_cfqg(st->left); 572 st->min_vdisktime = max_vdisktime(st->min_vdisktime, 573 cfqg->vdisktime); 574 } 575} 576 577/* 578 * get averaged number of queues of RT/BE priority. 579 * average is updated, with a formula that gives more weight to higher numbers, 580 * to quickly follows sudden increases and decrease slowly 581 */ 582 583static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd, 584 struct cfq_group *cfqg, bool rt) 585{ 586 unsigned min_q, max_q; 587 unsigned mult = cfq_hist_divisor - 1; 588 unsigned round = cfq_hist_divisor / 2; 589 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg); 590 591 min_q = min(cfqg->busy_queues_avg[rt], busy); 592 max_q = max(cfqg->busy_queues_avg[rt], busy); 593 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) / 594 cfq_hist_divisor; 595 return cfqg->busy_queues_avg[rt]; 596} 597 598static inline unsigned 599cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg) 600{ 601 struct cfq_rb_root *st = &cfqd->grp_service_tree; 602 603 return cfq_target_latency * cfqg->weight / st->total_weight; 604} 605 606static inline unsigned 607cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) 608{ 609 unsigned slice = cfq_prio_to_slice(cfqd, cfqq); 610 if (cfqd->cfq_latency) { 611 /* 612 * interested queues (we consider only the ones with the same 613 * priority class in the cfq group) 614 */ 615 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg, 616 cfq_class_rt(cfqq)); 617 unsigned sync_slice = cfqd->cfq_slice[1]; 618 unsigned expect_latency = sync_slice * iq; 619 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg); 620 621 if (expect_latency > group_slice) { 622 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle; 623 /* scale low_slice according to IO priority 624 * and sync vs async */ 625 unsigned low_slice = 626 min(slice, base_low_slice * slice / sync_slice); 627 /* the adapted slice value is scaled to fit all iqs 628 * into the target latency */ 629 slice = max(slice * group_slice / expect_latency, 630 low_slice); 631 } 632 } 633 return slice; 634} 635 636static inline void 637cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) 638{ 639 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq); 640 641 cfqq->slice_start = jiffies; 642 cfqq->slice_end = jiffies + slice; 643 cfqq->allocated_slice = slice; 644 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); 645} 646 647/* 648 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end 649 * isn't valid until the first request from the dispatch is activated 650 * and the slice time set. 651 */ 652static inline bool cfq_slice_used(struct cfq_queue *cfqq) 653{ 654 if (cfq_cfqq_slice_new(cfqq)) 655 return false; 656 if (time_before(jiffies, cfqq->slice_end)) 657 return false; 658 659 return true; 660} 661 662/* 663 * Lifted from AS - choose which of rq1 and rq2 that is best served now. 664 * We choose the request that is closest to the head right now. Distance 665 * behind the head is penalized and only allowed to a certain extent. 666 */ 667static struct request * 668cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last) 669{ 670 sector_t s1, s2, d1 = 0, d2 = 0; 671 unsigned long back_max; 672#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ 673#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ 674 unsigned wrap = 0; /* bit mask: requests behind the disk head? */ 675 676 if (rq1 == NULL || rq1 == rq2) 677 return rq2; 678 if (rq2 == NULL) 679 return rq1; 680 681 if (rq_is_sync(rq1) != rq_is_sync(rq2)) 682 return rq_is_sync(rq1) ? rq1 : rq2; 683 684 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO) 685 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2; 686 687 s1 = blk_rq_pos(rq1); 688 s2 = blk_rq_pos(rq2); 689 690 /* 691 * by definition, 1KiB is 2 sectors 692 */ 693 back_max = cfqd->cfq_back_max * 2; 694 695 /* 696 * Strict one way elevator _except_ in the case where we allow 697 * short backward seeks which are biased as twice the cost of a 698 * similar forward seek. 699 */ 700 if (s1 >= last) 701 d1 = s1 - last; 702 else if (s1 + back_max >= last) 703 d1 = (last - s1) * cfqd->cfq_back_penalty; 704 else 705 wrap |= CFQ_RQ1_WRAP; 706 707 if (s2 >= last) 708 d2 = s2 - last; 709 else if (s2 + back_max >= last) 710 d2 = (last - s2) * cfqd->cfq_back_penalty; 711 else 712 wrap |= CFQ_RQ2_WRAP; 713 714 /* Found required data */ 715 716 /* 717 * By doing switch() on the bit mask "wrap" we avoid having to 718 * check two variables for all permutations: --> faster! 719 */ 720 switch (wrap) { 721 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ 722 if (d1 < d2) 723 return rq1; 724 else if (d2 < d1) 725 return rq2; 726 else { 727 if (s1 >= s2) 728 return rq1; 729 else 730 return rq2; 731 } 732 733 case CFQ_RQ2_WRAP: 734 return rq1; 735 case CFQ_RQ1_WRAP: 736 return rq2; 737 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ 738 default: 739 /* 740 * Since both rqs are wrapped, 741 * start with the one that's further behind head 742 * (--> only *one* back seek required), 743 * since back seek takes more time than forward. 744 */ 745 if (s1 <= s2) 746 return rq1; 747 else 748 return rq2; 749 } 750} 751 752/* 753 * The below is leftmost cache rbtree addon 754 */ 755static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) 756{ 757 /* Service tree is empty */ 758 if (!root->count) 759 return NULL; 760 761 if (!root->left) 762 root->left = rb_first(&root->rb); 763 764 if (root->left) 765 return rb_entry(root->left, struct cfq_queue, rb_node); 766 767 return NULL; 768} 769 770static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root) 771{ 772 if (!root->left) 773 root->left = rb_first(&root->rb); 774 775 if (root->left) 776 return rb_entry_cfqg(root->left); 777 778 return NULL; 779} 780 781static void rb_erase_init(struct rb_node *n, struct rb_root *root) 782{ 783 rb_erase(n, root); 784 RB_CLEAR_NODE(n); 785} 786 787static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) 788{ 789 if (root->left == n) 790 root->left = NULL; 791 rb_erase_init(n, &root->rb); 792 --root->count; 793} 794 795/* 796 * would be nice to take fifo expire time into account as well 797 */ 798static struct request * 799cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, 800 struct request *last) 801{ 802 struct rb_node *rbnext = rb_next(&last->rb_node); 803 struct rb_node *rbprev = rb_prev(&last->rb_node); 804 struct request *next = NULL, *prev = NULL; 805 806 BUG_ON(RB_EMPTY_NODE(&last->rb_node)); 807 808 if (rbprev) 809 prev = rb_entry_rq(rbprev); 810 811 if (rbnext) 812 next = rb_entry_rq(rbnext); 813 else { 814 rbnext = rb_first(&cfqq->sort_list); 815 if (rbnext && rbnext != &last->rb_node) 816 next = rb_entry_rq(rbnext); 817 } 818 819 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last)); 820} 821 822static unsigned long cfq_slice_offset(struct cfq_data *cfqd, 823 struct cfq_queue *cfqq) 824{ 825 /* 826 * just an approximation, should be ok. 827 */ 828 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) - 829 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); 830} 831 832static inline s64 833cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg) 834{ 835 return cfqg->vdisktime - st->min_vdisktime; 836} 837 838static void 839__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) 840{ 841 struct rb_node **node = &st->rb.rb_node; 842 struct rb_node *parent = NULL; 843 struct cfq_group *__cfqg; 844 s64 key = cfqg_key(st, cfqg); 845 int left = 1; 846 847 while (*node != NULL) { 848 parent = *node; 849 __cfqg = rb_entry_cfqg(parent); 850 851 if (key < cfqg_key(st, __cfqg)) 852 node = &parent->rb_left; 853 else { 854 node = &parent->rb_right; 855 left = 0; 856 } 857 } 858 859 if (left) 860 st->left = &cfqg->rb_node; 861 862 rb_link_node(&cfqg->rb_node, parent, node); 863 rb_insert_color(&cfqg->rb_node, &st->rb); 864} 865 866static void 867cfq_update_group_weight(struct cfq_group *cfqg) 868{ 869 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node)); 870 if (cfqg->needs_update) { 871 cfqg->weight = cfqg->new_weight; 872 cfqg->needs_update = false; 873 } 874} 875 876static void 877cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) 878{ 879 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node)); 880 881 cfq_update_group_weight(cfqg); 882 __cfq_group_service_tree_add(st, cfqg); 883 st->total_weight += cfqg->weight; 884} 885 886static void 887cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg) 888{ 889 struct cfq_rb_root *st = &cfqd->grp_service_tree; 890 struct cfq_group *__cfqg; 891 struct rb_node *n; 892 893 cfqg->nr_cfqq++; 894 if (!RB_EMPTY_NODE(&cfqg->rb_node)) 895 return; 896 897 /* 898 * Currently put the group at the end. Later implement something 899 * so that groups get lesser vtime based on their weights, so that 900 * if group does not loose all if it was not continuously backlogged. 901 */ 902 n = rb_last(&st->rb); 903 if (n) { 904 __cfqg = rb_entry_cfqg(n); 905 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY; 906 } else 907 cfqg->vdisktime = st->min_vdisktime; 908 cfq_group_service_tree_add(st, cfqg); 909} 910 911static void 912cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg) 913{ 914 st->total_weight -= cfqg->weight; 915 if (!RB_EMPTY_NODE(&cfqg->rb_node)) 916 cfq_rb_erase(&cfqg->rb_node, st); 917} 918 919static void 920cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg) 921{ 922 struct cfq_rb_root *st = &cfqd->grp_service_tree; 923 924 BUG_ON(cfqg->nr_cfqq < 1); 925 cfqg->nr_cfqq--; 926 927 /* If there are other cfq queues under this group, don't delete it */ 928 if (cfqg->nr_cfqq) 929 return; 930 931 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group"); 932 cfq_group_service_tree_del(st, cfqg); 933 cfqg->saved_workload_slice = 0; 934 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1); 935} 936 937static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq, 938 unsigned int *unaccounted_time) 939{ 940 unsigned int slice_used; 941 942 /* 943 * Queue got expired before even a single request completed or 944 * got expired immediately after first request completion. 945 */ 946 if (!cfqq->slice_start || cfqq->slice_start == jiffies) { 947 /* 948 * Also charge the seek time incurred to the group, otherwise 949 * if there are mutiple queues in the group, each can dispatch 950 * a single request on seeky media and cause lots of seek time 951 * and group will never know it. 952 */ 953 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start), 954 1); 955 } else { 956 slice_used = jiffies - cfqq->slice_start; 957 if (slice_used > cfqq->allocated_slice) { 958 *unaccounted_time = slice_used - cfqq->allocated_slice; 959 slice_used = cfqq->allocated_slice; 960 } 961 if (time_after(cfqq->slice_start, cfqq->dispatch_start)) 962 *unaccounted_time += cfqq->slice_start - 963 cfqq->dispatch_start; 964 } 965 966 return slice_used; 967} 968 969static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg, 970 struct cfq_queue *cfqq) 971{ 972 struct cfq_rb_root *st = &cfqd->grp_service_tree; 973 unsigned int used_sl, charge, unaccounted_sl = 0; 974 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg) 975 - cfqg->service_tree_idle.count; 976 977 BUG_ON(nr_sync < 0); 978 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl); 979 980 if (iops_mode(cfqd)) 981 charge = cfqq->slice_dispatch; 982 else if (!cfq_cfqq_sync(cfqq) && !nr_sync) 983 charge = cfqq->allocated_slice; 984 985 /* Can't update vdisktime while group is on service tree */ 986 cfq_group_service_tree_del(st, cfqg); 987 cfqg->vdisktime += cfq_scale_slice(charge, cfqg); 988 /* If a new weight was requested, update now, off tree */ 989 cfq_group_service_tree_add(st, cfqg); 990 991 /* This group is being expired. Save the context */ 992 if (time_after(cfqd->workload_expires, jiffies)) { 993 cfqg->saved_workload_slice = cfqd->workload_expires 994 - jiffies; 995 cfqg->saved_workload = cfqd->serving_type; 996 cfqg->saved_serving_prio = cfqd->serving_prio; 997 } else 998 cfqg->saved_workload_slice = 0; 999 1000 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime, 1001 st->min_vdisktime); 1002 cfq_log_cfqq(cfqq->cfqd, cfqq, 1003 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu", 1004 used_sl, cfqq->slice_dispatch, charge, 1005 iops_mode(cfqd), cfqq->nr_sectors); 1006 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl, 1007 unaccounted_sl); 1008 cfq_blkiocg_set_start_empty_time(&cfqg->blkg); 1009} 1010 1011#ifdef CONFIG_CFQ_GROUP_IOSCHED 1012static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg) 1013{ 1014 if (blkg) 1015 return container_of(blkg, struct cfq_group, blkg); 1016 return NULL; 1017} 1018 1019static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg, 1020 unsigned int weight) 1021{ 1022 struct cfq_group *cfqg = cfqg_of_blkg(blkg); 1023 cfqg->new_weight = weight; 1024 cfqg->needs_update = true; 1025} 1026 1027static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd, 1028 struct cfq_group *cfqg, struct blkio_cgroup *blkcg) 1029{ 1030 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info; 1031 unsigned int major, minor; 1032 1033 /* 1034 * Add group onto cgroup list. It might happen that bdi->dev is 1035 * not initialized yet. Initialize this new group without major 1036 * and minor info and this info will be filled in once a new thread 1037 * comes for IO. 1038 */ 1039 if (bdi->dev) { 1040 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); 1041 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, 1042 (void *)cfqd, MKDEV(major, minor)); 1043 } else 1044 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, 1045 (void *)cfqd, 0); 1046 1047 cfqd->nr_blkcg_linked_grps++; 1048 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev); 1049 1050 /* Add group on cfqd list */ 1051 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list); 1052} 1053 1054/* 1055 * Should be called from sleepable context. No request queue lock as per 1056 * cpu stats are allocated dynamically and alloc_percpu needs to be called 1057 * from sleepable context. 1058 */ 1059static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd) 1060{ 1061 struct cfq_group *cfqg = NULL; 1062 int i, j, ret; 1063 struct cfq_rb_root *st; 1064 1065 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node); 1066 if (!cfqg) 1067 return NULL; 1068 1069 for_each_cfqg_st(cfqg, i, j, st) 1070 *st = CFQ_RB_ROOT; 1071 RB_CLEAR_NODE(&cfqg->rb_node); 1072 1073 cfqg->ttime.last_end_request = jiffies; 1074 1075 /* 1076 * Take the initial reference that will be released on destroy 1077 * This can be thought of a joint reference by cgroup and 1078 * elevator which will be dropped by either elevator exit 1079 * or cgroup deletion path depending on who is exiting first. 1080 */ 1081 cfqg->ref = 1; 1082 1083 ret = blkio_alloc_blkg_stats(&cfqg->blkg); 1084 if (ret) { 1085 kfree(cfqg); 1086 return NULL; 1087 } 1088 1089 return cfqg; 1090} 1091 1092static struct cfq_group * 1093cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg) 1094{ 1095 struct cfq_group *cfqg = NULL; 1096 void *key = cfqd; 1097 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info; 1098 unsigned int major, minor; 1099 1100 /* 1101 * This is the common case when there are no blkio cgroups. 1102 * Avoid lookup in this case 1103 */ 1104 if (blkcg == &blkio_root_cgroup) 1105 cfqg = &cfqd->root_group; 1106 else 1107 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key)); 1108 1109 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) { 1110 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); 1111 cfqg->blkg.dev = MKDEV(major, minor); 1112 } 1113 1114 return cfqg; 1115} 1116 1117/* 1118 * Search for the cfq group current task belongs to. request_queue lock must 1119 * be held. 1120 */ 1121static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd) 1122{ 1123 struct blkio_cgroup *blkcg; 1124 struct cfq_group *cfqg = NULL, *__cfqg = NULL; 1125 struct request_queue *q = cfqd->queue; 1126 1127 rcu_read_lock(); 1128 blkcg = task_blkio_cgroup(current); 1129 cfqg = cfq_find_cfqg(cfqd, blkcg); 1130 if (cfqg) { 1131 rcu_read_unlock(); 1132 return cfqg; 1133 } 1134 1135 /* 1136 * Need to allocate a group. Allocation of group also needs allocation 1137 * of per cpu stats which in-turn takes a mutex() and can block. Hence 1138 * we need to drop rcu lock and queue_lock before we call alloc. 1139 * 1140 * Not taking any queue reference here and assuming that queue is 1141 * around by the time we return. CFQ queue allocation code does 1142 * the same. It might be racy though. 1143 */ 1144 1145 rcu_read_unlock(); 1146 spin_unlock_irq(q->queue_lock); 1147 1148 cfqg = cfq_alloc_cfqg(cfqd); 1149 1150 spin_lock_irq(q->queue_lock); 1151 1152 rcu_read_lock(); 1153 blkcg = task_blkio_cgroup(current); 1154 1155 /* 1156 * If some other thread already allocated the group while we were 1157 * not holding queue lock, free up the group 1158 */ 1159 __cfqg = cfq_find_cfqg(cfqd, blkcg); 1160 1161 if (__cfqg) { 1162 kfree(cfqg); 1163 rcu_read_unlock(); 1164 return __cfqg; 1165 } 1166 1167 if (!cfqg) 1168 cfqg = &cfqd->root_group; 1169 1170 cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg); 1171 rcu_read_unlock(); 1172 return cfqg; 1173} 1174 1175static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg) 1176{ 1177 cfqg->ref++; 1178 return cfqg; 1179} 1180 1181static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) 1182{ 1183 /* Currently, all async queues are mapped to root group */ 1184 if (!cfq_cfqq_sync(cfqq)) 1185 cfqg = &cfqq->cfqd->root_group; 1186 1187 cfqq->cfqg = cfqg; 1188 /* cfqq reference on cfqg */ 1189 cfqq->cfqg->ref++; 1190} 1191 1192static void cfq_put_cfqg(struct cfq_group *cfqg) 1193{ 1194 struct cfq_rb_root *st; 1195 int i, j; 1196 1197 BUG_ON(cfqg->ref <= 0); 1198 cfqg->ref--; 1199 if (cfqg->ref) 1200 return; 1201 for_each_cfqg_st(cfqg, i, j, st) 1202 BUG_ON(!RB_EMPTY_ROOT(&st->rb)); 1203 free_percpu(cfqg->blkg.stats_cpu); 1204 kfree(cfqg); 1205} 1206 1207static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg) 1208{ 1209 /* Something wrong if we are trying to remove same group twice */ 1210 BUG_ON(hlist_unhashed(&cfqg->cfqd_node)); 1211 1212 hlist_del_init(&cfqg->cfqd_node); 1213 1214 BUG_ON(cfqd->nr_blkcg_linked_grps <= 0); 1215 cfqd->nr_blkcg_linked_grps--; 1216 1217 /* 1218 * Put the reference taken at the time of creation so that when all 1219 * queues are gone, group can be destroyed. 1220 */ 1221 cfq_put_cfqg(cfqg); 1222} 1223 1224static void cfq_release_cfq_groups(struct cfq_data *cfqd) 1225{ 1226 struct hlist_node *pos, *n; 1227 struct cfq_group *cfqg; 1228 1229 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) { 1230 /* 1231 * If cgroup removal path got to blk_group first and removed 1232 * it from cgroup list, then it will take care of destroying 1233 * cfqg also. 1234 */ 1235 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg)) 1236 cfq_destroy_cfqg(cfqd, cfqg); 1237 } 1238} 1239 1240/* 1241 * Blk cgroup controller notification saying that blkio_group object is being 1242 * delinked as associated cgroup object is going away. That also means that 1243 * no new IO will come in this group. So get rid of this group as soon as 1244 * any pending IO in the group is finished. 1245 * 1246 * This function is called under rcu_read_lock(). key is the rcu protected 1247 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu 1248 * read lock. 1249 * 1250 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means 1251 * it should not be NULL as even if elevator was exiting, cgroup deltion 1252 * path got to it first. 1253 */ 1254static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg) 1255{ 1256 unsigned long flags; 1257 struct cfq_data *cfqd = key; 1258 1259 spin_lock_irqsave(cfqd->queue->queue_lock, flags); 1260 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg)); 1261 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); 1262} 1263 1264#else /* GROUP_IOSCHED */ 1265static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd) 1266{ 1267 return &cfqd->root_group; 1268} 1269 1270static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg) 1271{ 1272 return cfqg; 1273} 1274 1275static inline void 1276cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) { 1277 cfqq->cfqg = cfqg; 1278} 1279 1280static void cfq_release_cfq_groups(struct cfq_data *cfqd) {} 1281static inline void cfq_put_cfqg(struct cfq_group *cfqg) {} 1282 1283#endif /* GROUP_IOSCHED */ 1284 1285/* 1286 * The cfqd->service_trees holds all pending cfq_queue's that have 1287 * requests waiting to be processed. It is sorted in the order that 1288 * we will service the queues. 1289 */ 1290static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1291 bool add_front) 1292{ 1293 struct rb_node **p, *parent; 1294 struct cfq_queue *__cfqq; 1295 unsigned long rb_key; 1296 struct cfq_rb_root *service_tree; 1297 int left; 1298 int new_cfqq = 1; 1299 1300 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq), 1301 cfqq_type(cfqq)); 1302 if (cfq_class_idle(cfqq)) { 1303 rb_key = CFQ_IDLE_DELAY; 1304 parent = rb_last(&service_tree->rb); 1305 if (parent && parent != &cfqq->rb_node) { 1306 __cfqq = rb_entry(parent, struct cfq_queue, rb_node); 1307 rb_key += __cfqq->rb_key; 1308 } else 1309 rb_key += jiffies; 1310 } else if (!add_front) { 1311 /* 1312 * Get our rb key offset. Subtract any residual slice 1313 * value carried from last service. A negative resid 1314 * count indicates slice overrun, and this should position 1315 * the next service time further away in the tree. 1316 */ 1317 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; 1318 rb_key -= cfqq->slice_resid; 1319 cfqq->slice_resid = 0; 1320 } else { 1321 rb_key = -HZ; 1322 __cfqq = cfq_rb_first(service_tree); 1323 rb_key += __cfqq ? __cfqq->rb_key : jiffies; 1324 } 1325 1326 if (!RB_EMPTY_NODE(&cfqq->rb_node)) { 1327 new_cfqq = 0; 1328 /* 1329 * same position, nothing more to do 1330 */ 1331 if (rb_key == cfqq->rb_key && 1332 cfqq->service_tree == service_tree) 1333 return; 1334 1335 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); 1336 cfqq->service_tree = NULL; 1337 } 1338 1339 left = 1; 1340 parent = NULL; 1341 cfqq->service_tree = service_tree; 1342 p = &service_tree->rb.rb_node; 1343 while (*p) { 1344 struct rb_node **n; 1345 1346 parent = *p; 1347 __cfqq = rb_entry(parent, struct cfq_queue, rb_node); 1348 1349 /* 1350 * sort by key, that represents service time. 1351 */ 1352 if (time_before(rb_key, __cfqq->rb_key)) 1353 n = &(*p)->rb_left; 1354 else { 1355 n = &(*p)->rb_right; 1356 left = 0; 1357 } 1358 1359 p = n; 1360 } 1361 1362 if (left) 1363 service_tree->left = &cfqq->rb_node; 1364 1365 cfqq->rb_key = rb_key; 1366 rb_link_node(&cfqq->rb_node, parent, p); 1367 rb_insert_color(&cfqq->rb_node, &service_tree->rb); 1368 service_tree->count++; 1369 if (add_front || !new_cfqq) 1370 return; 1371 cfq_group_notify_queue_add(cfqd, cfqq->cfqg); 1372} 1373 1374static struct cfq_queue * 1375cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, 1376 sector_t sector, struct rb_node **ret_parent, 1377 struct rb_node ***rb_link) 1378{ 1379 struct rb_node **p, *parent; 1380 struct cfq_queue *cfqq = NULL; 1381 1382 parent = NULL; 1383 p = &root->rb_node; 1384 while (*p) { 1385 struct rb_node **n; 1386 1387 parent = *p; 1388 cfqq = rb_entry(parent, struct cfq_queue, p_node); 1389 1390 /* 1391 * Sort strictly based on sector. Smallest to the left, 1392 * largest to the right. 1393 */ 1394 if (sector > blk_rq_pos(cfqq->next_rq)) 1395 n = &(*p)->rb_right; 1396 else if (sector < blk_rq_pos(cfqq->next_rq)) 1397 n = &(*p)->rb_left; 1398 else 1399 break; 1400 p = n; 1401 cfqq = NULL; 1402 } 1403 1404 *ret_parent = parent; 1405 if (rb_link) 1406 *rb_link = p; 1407 return cfqq; 1408} 1409 1410static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1411{ 1412 struct rb_node **p, *parent; 1413 struct cfq_queue *__cfqq; 1414 1415 if (cfqq->p_root) { 1416 rb_erase(&cfqq->p_node, cfqq->p_root); 1417 cfqq->p_root = NULL; 1418 } 1419 1420 if (cfq_class_idle(cfqq)) 1421 return; 1422 if (!cfqq->next_rq) 1423 return; 1424 1425 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; 1426 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, 1427 blk_rq_pos(cfqq->next_rq), &parent, &p); 1428 if (!__cfqq) { 1429 rb_link_node(&cfqq->p_node, parent, p); 1430 rb_insert_color(&cfqq->p_node, cfqq->p_root); 1431 } else 1432 cfqq->p_root = NULL; 1433} 1434 1435/* 1436 * Update cfqq's position in the service tree. 1437 */ 1438static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1439{ 1440 /* 1441 * Resorting requires the cfqq to be on the RR list already. 1442 */ 1443 if (cfq_cfqq_on_rr(cfqq)) { 1444 cfq_service_tree_add(cfqd, cfqq, 0); 1445 cfq_prio_tree_add(cfqd, cfqq); 1446 } 1447} 1448 1449/* 1450 * add to busy list of queues for service, trying to be fair in ordering 1451 * the pending list according to last request service 1452 */ 1453static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1454{ 1455 cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); 1456 BUG_ON(cfq_cfqq_on_rr(cfqq)); 1457 cfq_mark_cfqq_on_rr(cfqq); 1458 cfqd->busy_queues++; 1459 if (cfq_cfqq_sync(cfqq)) 1460 cfqd->busy_sync_queues++; 1461 1462 cfq_resort_rr_list(cfqd, cfqq); 1463} 1464 1465/* 1466 * Called when the cfqq no longer has requests pending, remove it from 1467 * the service tree. 1468 */ 1469static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1470{ 1471 cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); 1472 BUG_ON(!cfq_cfqq_on_rr(cfqq)); 1473 cfq_clear_cfqq_on_rr(cfqq); 1474 1475 if (!RB_EMPTY_NODE(&cfqq->rb_node)) { 1476 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); 1477 cfqq->service_tree = NULL; 1478 } 1479 if (cfqq->p_root) { 1480 rb_erase(&cfqq->p_node, cfqq->p_root); 1481 cfqq->p_root = NULL; 1482 } 1483 1484 cfq_group_notify_queue_del(cfqd, cfqq->cfqg); 1485 BUG_ON(!cfqd->busy_queues); 1486 cfqd->busy_queues--; 1487 if (cfq_cfqq_sync(cfqq)) 1488 cfqd->busy_sync_queues--; 1489} 1490 1491/* 1492 * rb tree support functions 1493 */ 1494static void cfq_del_rq_rb(struct request *rq) 1495{ 1496 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1497 const int sync = rq_is_sync(rq); 1498 1499 BUG_ON(!cfqq->queued[sync]); 1500 cfqq->queued[sync]--; 1501 1502 elv_rb_del(&cfqq->sort_list, rq); 1503 1504 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) { 1505 /* 1506 * Queue will be deleted from service tree when we actually 1507 * expire it later. Right now just remove it from prio tree 1508 * as it is empty. 1509 */ 1510 if (cfqq->p_root) { 1511 rb_erase(&cfqq->p_node, cfqq->p_root); 1512 cfqq->p_root = NULL; 1513 } 1514 } 1515} 1516 1517static void cfq_add_rq_rb(struct request *rq) 1518{ 1519 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1520 struct cfq_data *cfqd = cfqq->cfqd; 1521 struct request *prev; 1522 1523 cfqq->queued[rq_is_sync(rq)]++; 1524 1525 elv_rb_add(&cfqq->sort_list, rq); 1526 1527 if (!cfq_cfqq_on_rr(cfqq)) 1528 cfq_add_cfqq_rr(cfqd, cfqq); 1529 1530 /* 1531 * check if this request is a better next-serve candidate 1532 */ 1533 prev = cfqq->next_rq; 1534 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position); 1535 1536 /* 1537 * adjust priority tree position, if ->next_rq changes 1538 */ 1539 if (prev != cfqq->next_rq) 1540 cfq_prio_tree_add(cfqd, cfqq); 1541 1542 BUG_ON(!cfqq->next_rq); 1543} 1544 1545static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) 1546{ 1547 elv_rb_del(&cfqq->sort_list, rq); 1548 cfqq->queued[rq_is_sync(rq)]--; 1549 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, 1550 rq_data_dir(rq), rq_is_sync(rq)); 1551 cfq_add_rq_rb(rq); 1552 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg, 1553 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq), 1554 rq_is_sync(rq)); 1555} 1556 1557static struct request * 1558cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) 1559{ 1560 struct task_struct *tsk = current; 1561 struct cfq_io_context *cic; 1562 struct cfq_queue *cfqq; 1563 1564 cic = cfq_cic_lookup(cfqd, tsk->io_context); 1565 if (!cic) 1566 return NULL; 1567 1568 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); 1569 if (cfqq) { 1570 sector_t sector = bio->bi_sector + bio_sectors(bio); 1571 1572 return elv_rb_find(&cfqq->sort_list, sector); 1573 } 1574 1575 return NULL; 1576} 1577 1578static void cfq_activate_request(struct request_queue *q, struct request *rq) 1579{ 1580 struct cfq_data *cfqd = q->elevator->elevator_data; 1581 1582 cfqd->rq_in_driver++; 1583 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", 1584 cfqd->rq_in_driver); 1585 1586 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); 1587} 1588 1589static void cfq_deactivate_request(struct request_queue *q, struct request *rq) 1590{ 1591 struct cfq_data *cfqd = q->elevator->elevator_data; 1592 1593 WARN_ON(!cfqd->rq_in_driver); 1594 cfqd->rq_in_driver--; 1595 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", 1596 cfqd->rq_in_driver); 1597} 1598 1599static void cfq_remove_request(struct request *rq) 1600{ 1601 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1602 1603 if (cfqq->next_rq == rq) 1604 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); 1605 1606 list_del_init(&rq->queuelist); 1607 cfq_del_rq_rb(rq); 1608 1609 cfqq->cfqd->rq_queued--; 1610 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, 1611 rq_data_dir(rq), rq_is_sync(rq)); 1612 if (rq->cmd_flags & REQ_PRIO) { 1613 WARN_ON(!cfqq->prio_pending); 1614 cfqq->prio_pending--; 1615 } 1616} 1617 1618static int cfq_merge(struct request_queue *q, struct request **req, 1619 struct bio *bio) 1620{ 1621 struct cfq_data *cfqd = q->elevator->elevator_data; 1622 struct request *__rq; 1623 1624 __rq = cfq_find_rq_fmerge(cfqd, bio); 1625 if (__rq && elv_rq_merge_ok(__rq, bio)) { 1626 *req = __rq; 1627 return ELEVATOR_FRONT_MERGE; 1628 } 1629 1630 return ELEVATOR_NO_MERGE; 1631} 1632 1633static void cfq_merged_request(struct request_queue *q, struct request *req, 1634 int type) 1635{ 1636 if (type == ELEVATOR_FRONT_MERGE) { 1637 struct cfq_queue *cfqq = RQ_CFQQ(req); 1638 1639 cfq_reposition_rq_rb(cfqq, req); 1640 } 1641} 1642 1643static void cfq_bio_merged(struct request_queue *q, struct request *req, 1644 struct bio *bio) 1645{ 1646 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg, 1647 bio_data_dir(bio), cfq_bio_sync(bio)); 1648} 1649 1650static void 1651cfq_merged_requests(struct request_queue *q, struct request *rq, 1652 struct request *next) 1653{ 1654 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1655 /* 1656 * reposition in fifo if next is older than rq 1657 */ 1658 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && 1659 time_before(rq_fifo_time(next), rq_fifo_time(rq))) { 1660 list_move(&rq->queuelist, &next->queuelist); 1661 rq_set_fifo_time(rq, rq_fifo_time(next)); 1662 } 1663 1664 if (cfqq->next_rq == next) 1665 cfqq->next_rq = rq; 1666 cfq_remove_request(next); 1667 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg, 1668 rq_data_dir(next), rq_is_sync(next)); 1669} 1670 1671static int cfq_allow_merge(struct request_queue *q, struct request *rq, 1672 struct bio *bio) 1673{ 1674 struct cfq_data *cfqd = q->elevator->elevator_data; 1675 struct cfq_io_context *cic; 1676 struct cfq_queue *cfqq; 1677 1678 /* 1679 * Disallow merge of a sync bio into an async request. 1680 */ 1681 if (cfq_bio_sync(bio) && !rq_is_sync(rq)) 1682 return false; 1683 1684 /* 1685 * Lookup the cfqq that this bio will be queued with. Allow 1686 * merge only if rq is queued there. 1687 */ 1688 cic = cfq_cic_lookup(cfqd, current->io_context); 1689 if (!cic) 1690 return false; 1691 1692 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); 1693 return cfqq == RQ_CFQQ(rq); 1694} 1695 1696static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1697{ 1698 del_timer(&cfqd->idle_slice_timer); 1699 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg); 1700} 1701 1702static void __cfq_set_active_queue(struct cfq_data *cfqd, 1703 struct cfq_queue *cfqq) 1704{ 1705 if (cfqq) { 1706 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d", 1707 cfqd->serving_prio, cfqd->serving_type); 1708 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg); 1709 cfqq->slice_start = 0; 1710 cfqq->dispatch_start = jiffies; 1711 cfqq->allocated_slice = 0; 1712 cfqq->slice_end = 0; 1713 cfqq->slice_dispatch = 0; 1714 cfqq->nr_sectors = 0; 1715 1716 cfq_clear_cfqq_wait_request(cfqq); 1717 cfq_clear_cfqq_must_dispatch(cfqq); 1718 cfq_clear_cfqq_must_alloc_slice(cfqq); 1719 cfq_clear_cfqq_fifo_expire(cfqq); 1720 cfq_mark_cfqq_slice_new(cfqq); 1721 1722 cfq_del_timer(cfqd, cfqq); 1723 } 1724 1725 cfqd->active_queue = cfqq; 1726} 1727 1728/* 1729 * current cfqq expired its slice (or was too idle), select new one 1730 */ 1731static void 1732__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1733 bool timed_out) 1734{ 1735 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); 1736 1737 if (cfq_cfqq_wait_request(cfqq)) 1738 cfq_del_timer(cfqd, cfqq); 1739 1740 cfq_clear_cfqq_wait_request(cfqq); 1741 cfq_clear_cfqq_wait_busy(cfqq); 1742 1743 /* 1744 * If this cfqq is shared between multiple processes, check to 1745 * make sure that those processes are still issuing I/Os within 1746 * the mean seek distance. If not, it may be time to break the 1747 * queues apart again. 1748 */ 1749 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq)) 1750 cfq_mark_cfqq_split_coop(cfqq); 1751 1752 /* 1753 * store what was left of this slice, if the queue idled/timed out 1754 */ 1755 if (timed_out) { 1756 if (cfq_cfqq_slice_new(cfqq)) 1757 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq); 1758 else 1759 cfqq->slice_resid = cfqq->slice_end - jiffies; 1760 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); 1761 } 1762 1763 cfq_group_served(cfqd, cfqq->cfqg, cfqq); 1764 1765 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) 1766 cfq_del_cfqq_rr(cfqd, cfqq); 1767 1768 cfq_resort_rr_list(cfqd, cfqq); 1769 1770 if (cfqq == cfqd->active_queue) 1771 cfqd->active_queue = NULL; 1772 1773 if (cfqd->active_cic) { 1774 put_io_context(cfqd->active_cic->ioc); 1775 cfqd->active_cic = NULL; 1776 } 1777} 1778 1779static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) 1780{ 1781 struct cfq_queue *cfqq = cfqd->active_queue; 1782 1783 if (cfqq) 1784 __cfq_slice_expired(cfqd, cfqq, timed_out); 1785} 1786 1787/* 1788 * Get next queue for service. Unless we have a queue preemption, 1789 * we'll simply select the first cfqq in the service tree. 1790 */ 1791static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) 1792{ 1793 struct cfq_rb_root *service_tree = 1794 service_tree_for(cfqd->serving_group, cfqd->serving_prio, 1795 cfqd->serving_type); 1796 1797 if (!cfqd->rq_queued) 1798 return NULL; 1799 1800 /* There is nothing to dispatch */ 1801 if (!service_tree) 1802 return NULL; 1803 if (RB_EMPTY_ROOT(&service_tree->rb)) 1804 return NULL; 1805 return cfq_rb_first(service_tree); 1806} 1807 1808static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd) 1809{ 1810 struct cfq_group *cfqg; 1811 struct cfq_queue *cfqq; 1812 int i, j; 1813 struct cfq_rb_root *st; 1814 1815 if (!cfqd->rq_queued) 1816 return NULL; 1817 1818 cfqg = cfq_get_next_cfqg(cfqd); 1819 if (!cfqg) 1820 return NULL; 1821 1822 for_each_cfqg_st(cfqg, i, j, st) 1823 if ((cfqq = cfq_rb_first(st)) != NULL) 1824 return cfqq; 1825 return NULL; 1826} 1827 1828/* 1829 * Get and set a new active queue for service. 1830 */ 1831static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, 1832 struct cfq_queue *cfqq) 1833{ 1834 if (!cfqq) 1835 cfqq = cfq_get_next_queue(cfqd); 1836 1837 __cfq_set_active_queue(cfqd, cfqq); 1838 return cfqq; 1839} 1840 1841static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, 1842 struct request *rq) 1843{ 1844 if (blk_rq_pos(rq) >= cfqd->last_position) 1845 return blk_rq_pos(rq) - cfqd->last_position; 1846 else 1847 return cfqd->last_position - blk_rq_pos(rq); 1848} 1849 1850static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1851 struct request *rq) 1852{ 1853 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR; 1854} 1855 1856static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, 1857 struct cfq_queue *cur_cfqq) 1858{ 1859 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; 1860 struct rb_node *parent, *node; 1861 struct cfq_queue *__cfqq; 1862 sector_t sector = cfqd->last_position; 1863 1864 if (RB_EMPTY_ROOT(root)) 1865 return NULL; 1866 1867 /* 1868 * First, if we find a request starting at the end of the last 1869 * request, choose it. 1870 */ 1871 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); 1872 if (__cfqq) 1873 return __cfqq; 1874 1875 /* 1876 * If the exact sector wasn't found, the parent of the NULL leaf 1877 * will contain the closest sector. 1878 */ 1879 __cfqq = rb_entry(parent, struct cfq_queue, p_node); 1880 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) 1881 return __cfqq; 1882 1883 if (blk_rq_pos(__cfqq->next_rq) < sector) 1884 node = rb_next(&__cfqq->p_node); 1885 else 1886 node = rb_prev(&__cfqq->p_node); 1887 if (!node) 1888 return NULL; 1889 1890 __cfqq = rb_entry(node, struct cfq_queue, p_node); 1891 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) 1892 return __cfqq; 1893 1894 return NULL; 1895} 1896 1897/* 1898 * cfqd - obvious 1899 * cur_cfqq - passed in so that we don't decide that the current queue is 1900 * closely cooperating with itself. 1901 * 1902 * So, basically we're assuming that that cur_cfqq has dispatched at least 1903 * one request, and that cfqd->last_position reflects a position on the disk 1904 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid 1905 * assumption. 1906 */ 1907static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, 1908 struct cfq_queue *cur_cfqq) 1909{ 1910 struct cfq_queue *cfqq; 1911 1912 if (cfq_class_idle(cur_cfqq)) 1913 return NULL; 1914 if (!cfq_cfqq_sync(cur_cfqq)) 1915 return NULL; 1916 if (CFQQ_SEEKY(cur_cfqq)) 1917 return NULL; 1918 1919 /* 1920 * Don't search priority tree if it's the only queue in the group. 1921 */ 1922 if (cur_cfqq->cfqg->nr_cfqq == 1) 1923 return NULL; 1924 1925 /* 1926 * We should notice if some of the queues are cooperating, eg 1927 * working closely on the same area of the disk. In that case, 1928 * we can group them together and don't waste time idling. 1929 */ 1930 cfqq = cfqq_close(cfqd, cur_cfqq); 1931 if (!cfqq) 1932 return NULL; 1933 1934 /* If new queue belongs to different cfq_group, don't choose it */ 1935 if (cur_cfqq->cfqg != cfqq->cfqg) 1936 return NULL; 1937 1938 /* 1939 * It only makes sense to merge sync queues. 1940 */ 1941 if (!cfq_cfqq_sync(cfqq)) 1942 return NULL; 1943 if (CFQQ_SEEKY(cfqq)) 1944 return NULL; 1945 1946 /* 1947 * Do not merge queues of different priority classes 1948 */ 1949 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) 1950 return NULL; 1951 1952 return cfqq; 1953} 1954 1955/* 1956 * Determine whether we should enforce idle window for this queue. 1957 */ 1958 1959static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1960{ 1961 enum wl_prio_t prio = cfqq_prio(cfqq); 1962 struct cfq_rb_root *service_tree = cfqq->service_tree; 1963 1964 BUG_ON(!service_tree); 1965 BUG_ON(!service_tree->count); 1966 1967 if (!cfqd->cfq_slice_idle) 1968 return false; 1969 1970 /* We never do for idle class queues. */ 1971 if (prio == IDLE_WORKLOAD) 1972 return false; 1973 1974 /* We do for queues that were marked with idle window flag. */ 1975 if (cfq_cfqq_idle_window(cfqq) && 1976 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)) 1977 return true; 1978 1979 /* 1980 * Otherwise, we do only if they are the last ones 1981 * in their service tree. 1982 */ 1983 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) && 1984 !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false)) 1985 return true; 1986 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", 1987 service_tree->count); 1988 return false; 1989} 1990 1991static void cfq_arm_slice_timer(struct cfq_data *cfqd) 1992{ 1993 struct cfq_queue *cfqq = cfqd->active_queue; 1994 struct cfq_io_context *cic; 1995 unsigned long sl, group_idle = 0; 1996 1997 /* 1998 * SSD device without seek penalty, disable idling. But only do so 1999 * for devices that support queuing, otherwise we still have a problem 2000 * with sync vs async workloads. 2001 */ 2002 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) 2003 return; 2004 2005 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); 2006 WARN_ON(cfq_cfqq_slice_new(cfqq)); 2007 2008 /* 2009 * idle is disabled, either manually or by past process history 2010 */ 2011 if (!cfq_should_idle(cfqd, cfqq)) { 2012 /* no queue idling. Check for group idling */ 2013 if (cfqd->cfq_group_idle) 2014 group_idle = cfqd->cfq_group_idle; 2015 else 2016 return; 2017 } 2018 2019 /* 2020 * still active requests from this queue, don't idle 2021 */ 2022 if (cfqq->dispatched) 2023 return; 2024 2025 /* 2026 * task has exited, don't wait 2027 */ 2028 cic = cfqd->active_cic; 2029 if (!cic || !atomic_read(&cic->ioc->nr_tasks)) 2030 return; 2031 2032 /* 2033 * If our average think time is larger than the remaining time 2034 * slice, then don't idle. This avoids overrunning the allotted 2035 * time slice. 2036 */ 2037 if (sample_valid(cic->ttime.ttime_samples) && 2038 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) { 2039 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu", 2040 cic->ttime.ttime_mean); 2041 return; 2042 } 2043 2044 /* There are other queues in the group, don't do group idle */ 2045 if (group_idle && cfqq->cfqg->nr_cfqq > 1) 2046 return; 2047 2048 cfq_mark_cfqq_wait_request(cfqq); 2049 2050 if (group_idle) 2051 sl = cfqd->cfq_group_idle; 2052 else 2053 sl = cfqd->cfq_slice_idle; 2054 2055 mod_timer(&cfqd->idle_slice_timer, jiffies + sl); 2056 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg); 2057 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl, 2058 group_idle ? 1 : 0); 2059} 2060 2061/* 2062 * Move request from internal lists to the request queue dispatch list. 2063 */ 2064static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) 2065{ 2066 struct cfq_data *cfqd = q->elevator->elevator_data; 2067 struct cfq_queue *cfqq = RQ_CFQQ(rq); 2068 2069 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); 2070 2071 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); 2072 cfq_remove_request(rq); 2073 cfqq->dispatched++; 2074 (RQ_CFQG(rq))->dispatched++; 2075 elv_dispatch_sort(q, rq); 2076 2077 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++; 2078 cfqq->nr_sectors += blk_rq_sectors(rq); 2079 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq), 2080 rq_data_dir(rq), rq_is_sync(rq)); 2081} 2082 2083/* 2084 * return expired entry, or NULL to just start from scratch in rbtree 2085 */ 2086static struct request *cfq_check_fifo(struct cfq_queue *cfqq) 2087{ 2088 struct request *rq = NULL; 2089 2090 if (cfq_cfqq_fifo_expire(cfqq)) 2091 return NULL; 2092 2093 cfq_mark_cfqq_fifo_expire(cfqq); 2094 2095 if (list_empty(&cfqq->fifo)) 2096 return NULL; 2097 2098 rq = rq_entry_fifo(cfqq->fifo.next); 2099 if (time_before(jiffies, rq_fifo_time(rq))) 2100 rq = NULL; 2101 2102 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); 2103 return rq; 2104} 2105 2106static inline int 2107cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) 2108{ 2109 const int base_rq = cfqd->cfq_slice_async_rq; 2110 2111 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); 2112 2113 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio); 2114} 2115 2116/* 2117 * Must be called with the queue_lock held. 2118 */ 2119static int cfqq_process_refs(struct cfq_queue *cfqq) 2120{ 2121 int process_refs, io_refs; 2122 2123 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; 2124 process_refs = cfqq->ref - io_refs; 2125 BUG_ON(process_refs < 0); 2126 return process_refs; 2127} 2128 2129static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) 2130{ 2131 int process_refs, new_process_refs; 2132 struct cfq_queue *__cfqq; 2133 2134 /* 2135 * If there are no process references on the new_cfqq, then it is 2136 * unsafe to follow the ->new_cfqq chain as other cfqq's in the 2137 * chain may have dropped their last reference (not just their 2138 * last process reference). 2139 */ 2140 if (!cfqq_process_refs(new_cfqq)) 2141 return; 2142 2143 /* Avoid a circular list and skip interim queue merges */ 2144 while ((__cfqq = new_cfqq->new_cfqq)) { 2145 if (__cfqq == cfqq) 2146 return; 2147 new_cfqq = __cfqq; 2148 } 2149 2150 process_refs = cfqq_process_refs(cfqq); 2151 new_process_refs = cfqq_process_refs(new_cfqq); 2152 /* 2153 * If the process for the cfqq has gone away, there is no 2154 * sense in merging the queues. 2155 */ 2156 if (process_refs == 0 || new_process_refs == 0) 2157 return; 2158 2159 /* 2160 * Merge in the direction of the lesser amount of work. 2161 */ 2162 if (new_process_refs >= process_refs) { 2163 cfqq->new_cfqq = new_cfqq; 2164 new_cfqq->ref += process_refs; 2165 } else { 2166 new_cfqq->new_cfqq = cfqq; 2167 cfqq->ref += new_process_refs; 2168 } 2169} 2170 2171static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, 2172 struct cfq_group *cfqg, enum wl_prio_t prio) 2173{ 2174 struct cfq_queue *queue; 2175 int i; 2176 bool key_valid = false; 2177 unsigned long lowest_key = 0; 2178 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD; 2179 2180 for (i = 0; i <= SYNC_WORKLOAD; ++i) { 2181 /* select the one with lowest rb_key */ 2182 queue = cfq_rb_first(service_tree_for(cfqg, prio, i)); 2183 if (queue && 2184 (!key_valid || time_before(queue->rb_key, lowest_key))) { 2185 lowest_key = queue->rb_key; 2186 cur_best = i; 2187 key_valid = true; 2188 } 2189 } 2190 2191 return cur_best; 2192} 2193 2194static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg) 2195{ 2196 unsigned slice; 2197 unsigned count; 2198 struct cfq_rb_root *st; 2199 unsigned group_slice; 2200 enum wl_prio_t original_prio = cfqd->serving_prio; 2201 2202 /* Choose next priority. RT > BE > IDLE */ 2203 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg)) 2204 cfqd->serving_prio = RT_WORKLOAD; 2205 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg)) 2206 cfqd->serving_prio = BE_WORKLOAD; 2207 else { 2208 cfqd->serving_prio = IDLE_WORKLOAD; 2209 cfqd->workload_expires = jiffies + 1; 2210 return; 2211 } 2212 2213 if (original_prio != cfqd->serving_prio) 2214 goto new_workload; 2215 2216 /* 2217 * For RT and BE, we have to choose also the type 2218 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload 2219 * expiration time 2220 */ 2221 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type); 2222 count = st->count; 2223 2224 /* 2225 * check workload expiration, and that we still have other queues ready 2226 */ 2227 if (count && !time_after(jiffies, cfqd->workload_expires)) 2228 return; 2229 2230new_workload: 2231 /* otherwise select new workload type */ 2232 cfqd->serving_type = 2233 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio); 2234 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type); 2235 count = st->count; 2236 2237 /* 2238 * the workload slice is computed as a fraction of target latency 2239 * proportional to the number of queues in that workload, over 2240 * all the queues in the same priority class 2241 */ 2242 group_slice = cfq_group_slice(cfqd, cfqg); 2243 2244 slice = group_slice * count / 2245 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio], 2246 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg)); 2247 2248 if (cfqd->serving_type == ASYNC_WORKLOAD) { 2249 unsigned int tmp; 2250 2251 /* 2252 * Async queues are currently system wide. Just taking 2253 * proportion of queues with-in same group will lead to higher 2254 * async ratio system wide as generally root group is going 2255 * to have higher weight. A more accurate thing would be to 2256 * calculate system wide asnc/sync ratio. 2257 */ 2258 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg); 2259 tmp = tmp/cfqd->busy_queues; 2260 slice = min_t(unsigned, slice, tmp); 2261 2262 /* async workload slice is scaled down according to 2263 * the sync/async slice ratio. */ 2264 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1]; 2265 } else 2266 /* sync workload slice is at least 2 * cfq_slice_idle */ 2267 slice = max(slice, 2 * cfqd->cfq_slice_idle); 2268 2269 slice = max_t(unsigned, slice, CFQ_MIN_TT); 2270 cfq_log(cfqd, "workload slice:%d", slice); 2271 cfqd->workload_expires = jiffies + slice; 2272} 2273 2274static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd) 2275{ 2276 struct cfq_rb_root *st = &cfqd->grp_service_tree; 2277 struct cfq_group *cfqg; 2278 2279 if (RB_EMPTY_ROOT(&st->rb)) 2280 return NULL; 2281 cfqg = cfq_rb_first_group(st); 2282 update_min_vdisktime(st); 2283 return cfqg; 2284} 2285 2286static void cfq_choose_cfqg(struct cfq_data *cfqd) 2287{ 2288 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd); 2289 2290 cfqd->serving_group = cfqg; 2291 2292 /* Restore the workload type data */ 2293 if (cfqg->saved_workload_slice) { 2294 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice; 2295 cfqd->serving_type = cfqg->saved_workload; 2296 cfqd->serving_prio = cfqg->saved_serving_prio; 2297 } else 2298 cfqd->workload_expires = jiffies - 1; 2299 2300 choose_service_tree(cfqd, cfqg); 2301} 2302 2303/* 2304 * Select a queue for service. If we have a current active queue, 2305 * check whether to continue servicing it, or retrieve and set a new one. 2306 */ 2307static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) 2308{ 2309 struct cfq_queue *cfqq, *new_cfqq = NULL; 2310 2311 cfqq = cfqd->active_queue; 2312 if (!cfqq) 2313 goto new_queue; 2314 2315 if (!cfqd->rq_queued) 2316 return NULL; 2317 2318 /* 2319 * We were waiting for group to get backlogged. Expire the queue 2320 */ 2321 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list)) 2322 goto expire; 2323 2324 /* 2325 * The active queue has run out of time, expire it and select new. 2326 */ 2327 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) { 2328 /* 2329 * If slice had not expired at the completion of last request 2330 * we might not have turned on wait_busy flag. Don't expire 2331 * the queue yet. Allow the group to get backlogged. 2332 * 2333 * The very fact that we have used the slice, that means we 2334 * have been idling all along on this queue and it should be 2335 * ok to wait for this request to complete. 2336 */ 2337 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list) 2338 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) { 2339 cfqq = NULL; 2340 goto keep_queue; 2341 } else 2342 goto check_group_idle; 2343 } 2344 2345 /* 2346 * The active queue has requests and isn't expired, allow it to 2347 * dispatch. 2348 */ 2349 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) 2350 goto keep_queue; 2351 2352 /* 2353 * If another queue has a request waiting within our mean seek 2354 * distance, let it run. The expire code will check for close 2355 * cooperators and put the close queue at the front of the service 2356 * tree. If possible, merge the expiring queue with the new cfqq. 2357 */ 2358 new_cfqq = cfq_close_cooperator(cfqd, cfqq); 2359 if (new_cfqq) { 2360 if (!cfqq->new_cfqq) 2361 cfq_setup_merge(cfqq, new_cfqq); 2362 goto expire; 2363 } 2364 2365 /* 2366 * No requests pending. If the active queue still has requests in 2367 * flight or is idling for a new request, allow either of these 2368 * conditions to happen (or time out) before selecting a new queue. 2369 */ 2370 if (timer_pending(&cfqd->idle_slice_timer)) { 2371 cfqq = NULL; 2372 goto keep_queue; 2373 } 2374 2375 /* 2376 * This is a deep seek queue, but the device is much faster than 2377 * the queue can deliver, don't idle 2378 **/ 2379 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) && 2380 (cfq_cfqq_slice_new(cfqq) || 2381 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) { 2382 cfq_clear_cfqq_deep(cfqq); 2383 cfq_clear_cfqq_idle_window(cfqq); 2384 } 2385 2386 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) { 2387 cfqq = NULL; 2388 goto keep_queue; 2389 } 2390 2391 /* 2392 * If group idle is enabled and there are requests dispatched from 2393 * this group, wait for requests to complete. 2394 */ 2395check_group_idle: 2396 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 && 2397 cfqq->cfqg->dispatched && 2398 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) { 2399 cfqq = NULL; 2400 goto keep_queue; 2401 } 2402 2403expire: 2404 cfq_slice_expired(cfqd, 0); 2405new_queue: 2406 /* 2407 * Current queue expired. Check if we have to switch to a new 2408 * service tree 2409 */ 2410 if (!new_cfqq) 2411 cfq_choose_cfqg(cfqd); 2412 2413 cfqq = cfq_set_active_queue(cfqd, new_cfqq); 2414keep_queue: 2415 return cfqq; 2416} 2417 2418static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) 2419{ 2420 int dispatched = 0; 2421 2422 while (cfqq->next_rq) { 2423 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); 2424 dispatched++; 2425 } 2426 2427 BUG_ON(!list_empty(&cfqq->fifo)); 2428 2429 /* By default cfqq is not expired if it is empty. Do it explicitly */ 2430 __cfq_slice_expired(cfqq->cfqd, cfqq, 0); 2431 return dispatched; 2432} 2433 2434/* 2435 * Drain our current requests. Used for barriers and when switching 2436 * io schedulers on-the-fly. 2437 */ 2438static int cfq_forced_dispatch(struct cfq_data *cfqd) 2439{ 2440 struct cfq_queue *cfqq; 2441 int dispatched = 0; 2442 2443 /* Expire the timeslice of the current active queue first */ 2444 cfq_slice_expired(cfqd, 0); 2445 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) { 2446 __cfq_set_active_queue(cfqd, cfqq); 2447 dispatched += __cfq_forced_dispatch_cfqq(cfqq); 2448 } 2449 2450 BUG_ON(cfqd->busy_queues); 2451 2452 cfq_log(cfqd, "forced_dispatch=%d", dispatched); 2453 return dispatched; 2454} 2455 2456static inline bool cfq_slice_used_soon(struct cfq_data *cfqd, 2457 struct cfq_queue *cfqq) 2458{ 2459 /* the queue hasn't finished any request, can't estimate */ 2460 if (cfq_cfqq_slice_new(cfqq)) 2461 return true; 2462 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched, 2463 cfqq->slice_end)) 2464 return true; 2465 2466 return false; 2467} 2468 2469static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) 2470{ 2471 unsigned int max_dispatch; 2472 2473 /* 2474 * Drain async requests before we start sync IO 2475 */ 2476 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC]) 2477 return false; 2478 2479 /* 2480 * If this is an async queue and we have sync IO in flight, let it wait 2481 */ 2482 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq)) 2483 return false; 2484 2485 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1); 2486 if (cfq_class_idle(cfqq)) 2487 max_dispatch = 1; 2488 2489 /* 2490 * Does this cfqq already have too much IO in flight? 2491 */ 2492 if (cfqq->dispatched >= max_dispatch) { 2493 bool promote_sync = false; 2494 /* 2495 * idle queue must always only have a single IO in flight 2496 */ 2497 if (cfq_class_idle(cfqq)) 2498 return false; 2499 2500 /* 2501 * If there is only one sync queue 2502 * we can ignore async queue here and give the sync 2503 * queue no dispatch limit. The reason is a sync queue can 2504 * preempt async queue, limiting the sync queue doesn't make 2505 * sense. This is useful for aiostress test. 2506 */ 2507 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1) 2508 promote_sync = true; 2509 2510 /* 2511 * We have other queues, don't allow more IO from this one 2512 */ 2513 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) && 2514 !promote_sync) 2515 return false; 2516 2517 /* 2518 * Sole queue user, no limit 2519 */ 2520 if (cfqd->busy_queues == 1 || promote_sync) 2521 max_dispatch = -1; 2522 else 2523 /* 2524 * Normally we start throttling cfqq when cfq_quantum/2 2525 * requests have been dispatched. But we can drive 2526 * deeper queue depths at the beginning of slice 2527 * subjected to upper limit of cfq_quantum. 2528 * */ 2529 max_dispatch = cfqd->cfq_quantum; 2530 } 2531 2532 /* 2533 * Async queues must wait a bit before being allowed dispatch. 2534 * We also ramp up the dispatch depth gradually for async IO, 2535 * based on the last sync IO we serviced 2536 */ 2537 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { 2538 unsigned long last_sync = jiffies - cfqd->last_delayed_sync; 2539 unsigned int depth; 2540 2541 depth = last_sync / cfqd->cfq_slice[1]; 2542 if (!depth && !cfqq->dispatched) 2543 depth = 1; 2544 if (depth < max_dispatch) 2545 max_dispatch = depth; 2546 } 2547 2548 /* 2549 * If we're below the current max, allow a dispatch 2550 */ 2551 return cfqq->dispatched < max_dispatch; 2552} 2553 2554/* 2555 * Dispatch a request from cfqq, moving them to the request queue 2556 * dispatch list. 2557 */ 2558static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) 2559{ 2560 struct request *rq; 2561 2562 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); 2563 2564 if (!cfq_may_dispatch(cfqd, cfqq)) 2565 return false; 2566 2567 /* 2568 * follow expired path, else get first next available 2569 */ 2570 rq = cfq_check_fifo(cfqq); 2571 if (!rq) 2572 rq = cfqq->next_rq; 2573 2574 /* 2575 * insert request into driver dispatch list 2576 */ 2577 cfq_dispatch_insert(cfqd->queue, rq); 2578 2579 if (!cfqd->active_cic) { 2580 struct cfq_io_context *cic = RQ_CIC(rq); 2581 2582 atomic_long_inc(&cic->ioc->refcount); 2583 cfqd->active_cic = cic; 2584 } 2585 2586 return true; 2587} 2588 2589/* 2590 * Find the cfqq that we need to service and move a request from that to the 2591 * dispatch list 2592 */ 2593static int cfq_dispatch_requests(struct request_queue *q, int force) 2594{ 2595 struct cfq_data *cfqd = q->elevator->elevator_data; 2596 struct cfq_queue *cfqq; 2597 2598 if (!cfqd->busy_queues) 2599 return 0; 2600 2601 if (unlikely(force)) 2602 return cfq_forced_dispatch(cfqd); 2603 2604 cfqq = cfq_select_queue(cfqd); 2605 if (!cfqq) 2606 return 0; 2607 2608 /* 2609 * Dispatch a request from this cfqq, if it is allowed 2610 */ 2611 if (!cfq_dispatch_request(cfqd, cfqq)) 2612 return 0; 2613 2614 cfqq->slice_dispatch++; 2615 cfq_clear_cfqq_must_dispatch(cfqq); 2616 2617 /* 2618 * expire an async queue immediately if it has used up its slice. idle 2619 * queue always expire after 1 dispatch round. 2620 */ 2621 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && 2622 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || 2623 cfq_class_idle(cfqq))) { 2624 cfqq->slice_end = jiffies + 1; 2625 cfq_slice_expired(cfqd, 0); 2626 } 2627 2628 cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); 2629 return 1; 2630} 2631 2632/* 2633 * task holds one reference to the queue, dropped when task exits. each rq 2634 * in-flight on this queue also holds a reference, dropped when rq is freed. 2635 * 2636 * Each cfq queue took a reference on the parent group. Drop it now. 2637 * queue lock must be held here. 2638 */ 2639static void cfq_put_queue(struct cfq_queue *cfqq) 2640{ 2641 struct cfq_data *cfqd = cfqq->cfqd; 2642 struct cfq_group *cfqg; 2643 2644 BUG_ON(cfqq->ref <= 0); 2645 2646 cfqq->ref--; 2647 if (cfqq->ref) 2648 return; 2649 2650 cfq_log_cfqq(cfqd, cfqq, "put_queue"); 2651 BUG_ON(rb_first(&cfqq->sort_list)); 2652 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); 2653 cfqg = cfqq->cfqg; 2654 2655 if (unlikely(cfqd->active_queue == cfqq)) { 2656 __cfq_slice_expired(cfqd, cfqq, 0); 2657 cfq_schedule_dispatch(cfqd); 2658 } 2659 2660 BUG_ON(cfq_cfqq_on_rr(cfqq)); 2661 kmem_cache_free(cfq_pool, cfqq); 2662 cfq_put_cfqg(cfqg); 2663} 2664 2665/* 2666 * Call func for each cic attached to this ioc. 2667 */ 2668static void 2669call_for_each_cic(struct io_context *ioc, 2670 void (*func)(struct io_context *, struct cfq_io_context *)) 2671{ 2672 struct cfq_io_context *cic; 2673 struct hlist_node *n; 2674 2675 rcu_read_lock(); 2676 2677 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list) 2678 func(ioc, cic); 2679 2680 rcu_read_unlock(); 2681} 2682 2683static void cfq_cic_free_rcu(struct rcu_head *head) 2684{ 2685 struct cfq_io_context *cic; 2686 2687 cic = container_of(head, struct cfq_io_context, rcu_head); 2688 2689 kmem_cache_free(cfq_ioc_pool, cic); 2690 elv_ioc_count_dec(cfq_ioc_count); 2691 2692 if (ioc_gone) { 2693 /* 2694 * CFQ scheduler is exiting, grab exit lock and check 2695 * the pending io context count. If it hits zero, 2696 * complete ioc_gone and set it back to NULL 2697 */ 2698 spin_lock(&ioc_gone_lock); 2699 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) { 2700 complete(ioc_gone); 2701 ioc_gone = NULL; 2702 } 2703 spin_unlock(&ioc_gone_lock); 2704 } 2705} 2706 2707static void cfq_cic_free(struct cfq_io_context *cic) 2708{ 2709 call_rcu(&cic->rcu_head, cfq_cic_free_rcu); 2710} 2711 2712static void cfq_release_cic(struct cfq_io_context *cic) 2713{ 2714 struct io_context *ioc = cic->ioc; 2715 unsigned long dead_key = (unsigned long) cic->key; 2716 2717 BUG_ON(!(dead_key & CIC_DEAD_KEY)); 2718 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT); 2719 hlist_del_rcu(&cic->cic_list); 2720 cfq_cic_free(cic); 2721} 2722 2723static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) 2724{ 2725 unsigned long flags; 2726 2727 spin_lock_irqsave(&ioc->lock, flags); 2728 cfq_release_cic(cic); 2729 spin_unlock_irqrestore(&ioc->lock, flags); 2730} 2731 2732/* 2733 * Must be called with rcu_read_lock() held or preemption otherwise disabled. 2734 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(), 2735 * and ->trim() which is called with the task lock held 2736 */ 2737static void cfq_free_io_context(struct io_context *ioc) 2738{ 2739 /* 2740 * ioc->refcount is zero here, or we are called from elv_unregister(), 2741 * so no more cic's are allowed to be linked into this ioc. So it 2742 * should be ok to iterate over the known list, we will see all cic's 2743 * since no new ones are added. 2744 */ 2745 call_for_each_cic(ioc, cic_free_func); 2746} 2747 2748static void cfq_put_cooperator(struct cfq_queue *cfqq) 2749{ 2750 struct cfq_queue *__cfqq, *next; 2751 2752 /* 2753 * If this queue was scheduled to merge with another queue, be 2754 * sure to drop the reference taken on that queue (and others in 2755 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. 2756 */ 2757 __cfqq = cfqq->new_cfqq; 2758 while (__cfqq) { 2759 if (__cfqq == cfqq) { 2760 WARN(1, "cfqq->new_cfqq loop detected\n"); 2761 break; 2762 } 2763 next = __cfqq->new_cfqq; 2764 cfq_put_queue(__cfqq); 2765 __cfqq = next; 2766 } 2767} 2768 2769static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) 2770{ 2771 if (unlikely(cfqq == cfqd->active_queue)) { 2772 __cfq_slice_expired(cfqd, cfqq, 0); 2773 cfq_schedule_dispatch(cfqd); 2774 } 2775 2776 cfq_put_cooperator(cfqq); 2777 2778 cfq_put_queue(cfqq); 2779} 2780 2781static void cfq_exit_cic(struct cfq_io_context *cic) 2782{ 2783 struct cfq_data *cfqd = cic_to_cfqd(cic); 2784 struct io_context *ioc = cic->ioc; 2785 2786 list_del_init(&cic->queue_list); 2787 2788 /* 2789 * Make sure dead mark is seen for dead queues 2790 */ 2791 smp_wmb(); 2792 cic->key = cfqd_dead_key(cfqd); 2793 2794 rcu_read_lock(); 2795 if (rcu_dereference(ioc->ioc_data) == cic) { 2796 rcu_read_unlock(); 2797 spin_lock(&ioc->lock); 2798 rcu_assign_pointer(ioc->ioc_data, NULL); 2799 spin_unlock(&ioc->lock); 2800 } else 2801 rcu_read_unlock(); 2802 2803 if (cic->cfqq[BLK_RW_ASYNC]) { 2804 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); 2805 cic->cfqq[BLK_RW_ASYNC] = NULL; 2806 } 2807 2808 if (cic->cfqq[BLK_RW_SYNC]) { 2809 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); 2810 cic->cfqq[BLK_RW_SYNC] = NULL; 2811 } 2812} 2813 2814static void cfq_exit_single_io_context(struct io_context *ioc, 2815 struct cfq_io_context *cic) 2816{ 2817 struct cfq_data *cfqd = cic_to_cfqd(cic); 2818 2819 if (cfqd) { 2820 struct request_queue *q = cfqd->queue; 2821 unsigned long flags; 2822 2823 spin_lock_irqsave(q->queue_lock, flags); 2824 2825 /* 2826 * Ensure we get a fresh copy of the ->key to prevent 2827 * race between exiting task and queue 2828 */ 2829 smp_read_barrier_depends(); 2830 if (cic->key == cfqd) 2831 cfq_exit_cic(cic); 2832 2833 spin_unlock_irqrestore(q->queue_lock, flags); 2834 } 2835} 2836 2837/* 2838 * The process that ioc belongs to has exited, we need to clean up 2839 * and put the internal structures we have that belongs to that process. 2840 */ 2841static void cfq_exit_io_context(struct io_context *ioc) 2842{ 2843 call_for_each_cic(ioc, cfq_exit_single_io_context); 2844} 2845 2846static struct cfq_io_context * 2847cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) 2848{ 2849 struct cfq_io_context *cic; 2850 2851 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, 2852 cfqd->queue->node); 2853 if (cic) { 2854 cic->ttime.last_end_request = jiffies; 2855 INIT_LIST_HEAD(&cic->queue_list); 2856 INIT_HLIST_NODE(&cic->cic_list); 2857 cic->dtor = cfq_free_io_context; 2858 cic->exit = cfq_exit_io_context; 2859 elv_ioc_count_inc(cfq_ioc_count); 2860 } 2861 2862 return cic; 2863} 2864 2865static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) 2866{ 2867 struct task_struct *tsk = current; 2868 int ioprio_class; 2869 2870 if (!cfq_cfqq_prio_changed(cfqq)) 2871 return; 2872 2873 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); 2874 switch (ioprio_class) { 2875 default: 2876 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); 2877 case IOPRIO_CLASS_NONE: 2878 /* 2879 * no prio set, inherit CPU scheduling settings 2880 */ 2881 cfqq->ioprio = task_nice_ioprio(tsk); 2882 cfqq->ioprio_class = task_nice_ioclass(tsk); 2883 break; 2884 case IOPRIO_CLASS_RT: 2885 cfqq->ioprio = task_ioprio(ioc); 2886 cfqq->ioprio_class = IOPRIO_CLASS_RT; 2887 break; 2888 case IOPRIO_CLASS_BE: 2889 cfqq->ioprio = task_ioprio(ioc); 2890 cfqq->ioprio_class = IOPRIO_CLASS_BE; 2891 break; 2892 case IOPRIO_CLASS_IDLE: 2893 cfqq->ioprio_class = IOPRIO_CLASS_IDLE; 2894 cfqq->ioprio = 7; 2895 cfq_clear_cfqq_idle_window(cfqq); 2896 break; 2897 } 2898 2899 /* 2900 * keep track of original prio settings in case we have to temporarily 2901 * elevate the priority of this queue 2902 */ 2903 cfqq->org_ioprio = cfqq->ioprio; 2904 cfq_clear_cfqq_prio_changed(cfqq); 2905} 2906 2907static void changed_ioprio(struct cfq_io_context *cic) 2908{ 2909 struct cfq_data *cfqd = cic_to_cfqd(cic); 2910 struct cfq_queue *cfqq; 2911 unsigned long flags; 2912 2913 if (unlikely(!cfqd)) 2914 return; 2915 2916 spin_lock_irqsave(cfqd->queue->queue_lock, flags); 2917 2918 cfqq = cic->cfqq[BLK_RW_ASYNC]; 2919 if (cfqq) { 2920 struct cfq_queue *new_cfqq; 2921 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc, 2922 GFP_ATOMIC); 2923 if (new_cfqq) { 2924 cic->cfqq[BLK_RW_ASYNC] = new_cfqq; 2925 cfq_put_queue(cfqq); 2926 } 2927 } 2928 2929 cfqq = cic->cfqq[BLK_RW_SYNC]; 2930 if (cfqq) 2931 cfq_mark_cfqq_prio_changed(cfqq); 2932 2933 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); 2934} 2935 2936static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, 2937 pid_t pid, bool is_sync) 2938{ 2939 RB_CLEAR_NODE(&cfqq->rb_node); 2940 RB_CLEAR_NODE(&cfqq->p_node); 2941 INIT_LIST_HEAD(&cfqq->fifo); 2942 2943 cfqq->ref = 0; 2944 cfqq->cfqd = cfqd; 2945 2946 cfq_mark_cfqq_prio_changed(cfqq); 2947 2948 if (is_sync) { 2949 if (!cfq_class_idle(cfqq)) 2950 cfq_mark_cfqq_idle_window(cfqq); 2951 cfq_mark_cfqq_sync(cfqq); 2952 } 2953 cfqq->pid = pid; 2954} 2955 2956#ifdef CONFIG_CFQ_GROUP_IOSCHED 2957static void changed_cgroup(struct cfq_io_context *cic) 2958{ 2959 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1); 2960 struct cfq_data *cfqd = cic_to_cfqd(cic); 2961 unsigned long flags; 2962 struct request_queue *q; 2963 2964 if (unlikely(!cfqd)) 2965 return; 2966 2967 q = cfqd->queue; 2968 2969 spin_lock_irqsave(q->queue_lock, flags); 2970 2971 if (sync_cfqq) { 2972 /* 2973 * Drop reference to sync queue. A new sync queue will be 2974 * assigned in new group upon arrival of a fresh request. 2975 */ 2976 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup"); 2977 cic_set_cfqq(cic, NULL, 1); 2978 cfq_put_queue(sync_cfqq); 2979 } 2980 2981 spin_unlock_irqrestore(q->queue_lock, flags); 2982} 2983#endif /* CONFIG_CFQ_GROUP_IOSCHED */ 2984 2985static struct cfq_queue * 2986cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, 2987 struct io_context *ioc, gfp_t gfp_mask) 2988{ 2989 struct cfq_queue *cfqq, *new_cfqq = NULL; 2990 struct cfq_io_context *cic; 2991 struct cfq_group *cfqg; 2992 2993retry: 2994 cfqg = cfq_get_cfqg(cfqd); 2995 cic = cfq_cic_lookup(cfqd, ioc); 2996 /* cic always exists here */ 2997 cfqq = cic_to_cfqq(cic, is_sync); 2998 2999 /* 3000 * Always try a new alloc if we fell back to the OOM cfqq 3001 * originally, since it should just be a temporary situation. 3002 */ 3003 if (!cfqq || cfqq == &cfqd->oom_cfqq) { 3004 cfqq = NULL; 3005 if (new_cfqq) { 3006 cfqq = new_cfqq; 3007 new_cfqq = NULL; 3008 } else if (gfp_mask & __GFP_WAIT) { 3009 spin_unlock_irq(cfqd->queue->queue_lock); 3010 new_cfqq = kmem_cache_alloc_node(cfq_pool, 3011 gfp_mask | __GFP_ZERO, 3012 cfqd->queue->node); 3013 spin_lock_irq(cfqd->queue->queue_lock); 3014 if (new_cfqq) 3015 goto retry; 3016 } else { 3017 cfqq = kmem_cache_alloc_node(cfq_pool, 3018 gfp_mask | __GFP_ZERO, 3019 cfqd->queue->node); 3020 } 3021 3022 if (cfqq) { 3023 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); 3024 cfq_init_prio_data(cfqq, ioc); 3025 cfq_link_cfqq_cfqg(cfqq, cfqg); 3026 cfq_log_cfqq(cfqd, cfqq, "alloced"); 3027 } else 3028 cfqq = &cfqd->oom_cfqq; 3029 } 3030 3031 if (new_cfqq) 3032 kmem_cache_free(cfq_pool, new_cfqq); 3033 3034 return cfqq; 3035} 3036 3037static struct cfq_queue ** 3038cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) 3039{ 3040 switch (ioprio_class) { 3041 case IOPRIO_CLASS_RT: 3042 return &cfqd->async_cfqq[0][ioprio]; 3043 case IOPRIO_CLASS_BE: 3044 return &cfqd->async_cfqq[1][ioprio]; 3045 case IOPRIO_CLASS_IDLE: 3046 return &cfqd->async_idle_cfqq; 3047 default: 3048 BUG(); 3049 } 3050} 3051 3052static struct cfq_queue * 3053cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc, 3054 gfp_t gfp_mask) 3055{ 3056 const int ioprio = task_ioprio(ioc); 3057 const int ioprio_class = task_ioprio_class(ioc); 3058 struct cfq_queue **async_cfqq = NULL; 3059 struct cfq_queue *cfqq = NULL; 3060 3061 if (!is_sync) { 3062 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); 3063 cfqq = *async_cfqq; 3064 } 3065 3066 if (!cfqq) 3067 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); 3068 3069 /* 3070 * pin the queue now that it's allocated, scheduler exit will prune it 3071 */ 3072 if (!is_sync && !(*async_cfqq)) { 3073 cfqq->ref++; 3074 *async_cfqq = cfqq; 3075 } 3076 3077 cfqq->ref++; 3078 return cfqq; 3079} 3080 3081/* 3082 * We drop cfq io contexts lazily, so we may find a dead one. 3083 */ 3084static void 3085cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, 3086 struct cfq_io_context *cic) 3087{ 3088 unsigned long flags; 3089 3090 WARN_ON(!list_empty(&cic->queue_list)); 3091 BUG_ON(cic->key != cfqd_dead_key(cfqd)); 3092 3093 spin_lock_irqsave(&ioc->lock, flags); 3094 3095 BUG_ON(rcu_dereference_check(ioc->ioc_data, 3096 lockdep_is_held(&ioc->lock)) == cic); 3097 3098 radix_tree_delete(&ioc->radix_root, cfqd->queue->id); 3099 hlist_del_rcu(&cic->cic_list); 3100 spin_unlock_irqrestore(&ioc->lock, flags); 3101 3102 cfq_cic_free(cic); 3103} 3104 3105static struct cfq_io_context * 3106cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) 3107{ 3108 struct cfq_io_context *cic; 3109 unsigned long flags; 3110 3111 if (unlikely(!ioc)) 3112 return NULL; 3113 3114 rcu_read_lock(); 3115 3116 /* 3117 * we maintain a last-hit cache, to avoid browsing over the tree 3118 */ 3119 cic = rcu_dereference(ioc->ioc_data); 3120 if (cic && cic->key == cfqd) { 3121 rcu_read_unlock(); 3122 return cic; 3123 } 3124 3125 do { 3126 cic = radix_tree_lookup(&ioc->radix_root, cfqd->queue->id); 3127 rcu_read_unlock(); 3128 if (!cic) 3129 break; 3130 if (unlikely(cic->key != cfqd)) { 3131 cfq_drop_dead_cic(cfqd, ioc, cic); 3132 rcu_read_lock(); 3133 continue; 3134 } 3135 3136 spin_lock_irqsave(&ioc->lock, flags); 3137 rcu_assign_pointer(ioc->ioc_data, cic); 3138 spin_unlock_irqrestore(&ioc->lock, flags); 3139 break; 3140 } while (1); 3141 3142 return cic; 3143} 3144 3145/* 3146 * Add cic into ioc, using cfqd as the search key. This enables us to lookup 3147 * the process specific cfq io context when entered from the block layer. 3148 * Also adds the cic to a per-cfqd list, used when this queue is removed. 3149 */ 3150static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, 3151 struct cfq_io_context *cic, gfp_t gfp_mask) 3152{ 3153 unsigned long flags; 3154 int ret; 3155 3156 ret = radix_tree_preload(gfp_mask); 3157 if (ret) 3158 goto out; 3159 3160 cic->ioc = ioc; 3161 cic->key = cfqd; 3162 cic->q = cfqd->queue; 3163 3164 spin_lock_irqsave(&ioc->lock, flags); 3165 ret = radix_tree_insert(&ioc->radix_root, cfqd->queue->id, cic); 3166 if (!ret) 3167 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list); 3168 spin_unlock_irqrestore(&ioc->lock, flags); 3169 3170 radix_tree_preload_end(); 3171 3172 if (!ret) { 3173 spin_lock_irqsave(cfqd->queue->queue_lock, flags); 3174 list_add(&cic->queue_list, &cfqd->cic_list); 3175 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); 3176 } 3177out: 3178 if (ret) 3179 printk(KERN_ERR "cfq: cic link failed!\n"); 3180 return ret; 3181} 3182 3183/* 3184 * Setup general io context and cfq io context. There can be several cfq 3185 * io contexts per general io context, if this process is doing io to more 3186 * than one device managed by cfq. 3187 */ 3188static struct cfq_io_context * 3189cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) 3190{ 3191 struct io_context *ioc = NULL; 3192 struct cfq_io_context *cic = NULL; 3193 3194 might_sleep_if(gfp_mask & __GFP_WAIT); 3195 3196 ioc = current_io_context(gfp_mask, cfqd->queue->node); 3197 if (!ioc) 3198 goto err; 3199 3200 cic = cfq_cic_lookup(cfqd, ioc); 3201 if (cic) 3202 goto out; 3203 3204 cic = cfq_alloc_io_context(cfqd, gfp_mask); 3205 if (cic == NULL) 3206 goto err; 3207 3208 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) 3209 goto err; 3210out: 3211 get_io_context(ioc); 3212 3213 if (unlikely(cic->changed)) { 3214 if (test_and_clear_bit(CIC_IOPRIO_CHANGED, &cic->changed)) 3215 changed_ioprio(cic); 3216#ifdef CONFIG_CFQ_GROUP_IOSCHED 3217 if (test_and_clear_bit(CIC_CGROUP_CHANGED, &cic->changed)) 3218 changed_cgroup(cic); 3219#endif 3220 } 3221 3222 return cic; 3223err: 3224 if (cic) 3225 cfq_cic_free(cic); 3226 return NULL; 3227} 3228 3229static void 3230__cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle) 3231{ 3232 unsigned long elapsed = jiffies - ttime->last_end_request; 3233 elapsed = min(elapsed, 2UL * slice_idle); 3234 3235 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8; 3236 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8; 3237 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples; 3238} 3239 3240static void 3241cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq, 3242 struct cfq_io_context *cic) 3243{ 3244 if (cfq_cfqq_sync(cfqq)) { 3245 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle); 3246 __cfq_update_io_thinktime(&cfqq->service_tree->ttime, 3247 cfqd->cfq_slice_idle); 3248 } 3249#ifdef CONFIG_CFQ_GROUP_IOSCHED 3250 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle); 3251#endif 3252} 3253 3254static void 3255cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, 3256 struct request *rq) 3257{ 3258 sector_t sdist = 0; 3259 sector_t n_sec = blk_rq_sectors(rq); 3260 if (cfqq->last_request_pos) { 3261 if (cfqq->last_request_pos < blk_rq_pos(rq)) 3262 sdist = blk_rq_pos(rq) - cfqq->last_request_pos; 3263 else 3264 sdist = cfqq->last_request_pos - blk_rq_pos(rq); 3265 } 3266 3267 cfqq->seek_history <<= 1; 3268 if (blk_queue_nonrot(cfqd->queue)) 3269 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT); 3270 else 3271 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR); 3272} 3273 3274/* 3275 * Disable idle window if the process thinks too long or seeks so much that 3276 * it doesn't matter 3277 */ 3278static void 3279cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, 3280 struct cfq_io_context *cic) 3281{ 3282 int old_idle, enable_idle; 3283 3284 /* 3285 * Don't idle for async or idle io prio class 3286 */ 3287 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) 3288 return; 3289 3290 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); 3291 3292 if (cfqq->queued[0] + cfqq->queued[1] >= 4) 3293 cfq_mark_cfqq_deep(cfqq); 3294 3295 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE)) 3296 enable_idle = 0; 3297 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || 3298 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq))) 3299 enable_idle = 0; 3300 else if (sample_valid(cic->ttime.ttime_samples)) { 3301 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle) 3302 enable_idle = 0; 3303 else 3304 enable_idle = 1; 3305 } 3306 3307 if (old_idle != enable_idle) { 3308 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); 3309 if (enable_idle) 3310 cfq_mark_cfqq_idle_window(cfqq); 3311 else 3312 cfq_clear_cfqq_idle_window(cfqq); 3313 } 3314} 3315 3316/* 3317 * Check if new_cfqq should preempt the currently active queue. Return 0 for 3318 * no or if we aren't sure, a 1 will cause a preempt. 3319 */ 3320static bool 3321cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, 3322 struct request *rq) 3323{ 3324 struct cfq_queue *cfqq; 3325 3326 cfqq = cfqd->active_queue; 3327 if (!cfqq) 3328 return false; 3329 3330 if (cfq_class_idle(new_cfqq)) 3331 return false; 3332 3333 if (cfq_class_idle(cfqq)) 3334 return true; 3335 3336 /* 3337 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice. 3338 */ 3339 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq)) 3340 return false; 3341 3342 /* 3343 * if the new request is sync, but the currently running queue is 3344 * not, let the sync request have priority. 3345 */ 3346 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) 3347 return true; 3348 3349 if (new_cfqq->cfqg != cfqq->cfqg) 3350 return false; 3351 3352 if (cfq_slice_used(cfqq)) 3353 return true; 3354 3355 /* Allow preemption only if we are idling on sync-noidle tree */ 3356 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD && 3357 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD && 3358 new_cfqq->service_tree->count == 2 && 3359 RB_EMPTY_ROOT(&cfqq->sort_list)) 3360 return true; 3361 3362 /* 3363 * So both queues are sync. Let the new request get disk time if 3364 * it's a metadata request and the current queue is doing regular IO. 3365 */ 3366 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending) 3367 return true; 3368 3369 /* 3370 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. 3371 */ 3372 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) 3373 return true; 3374 3375 /* An idle queue should not be idle now for some reason */ 3376 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq)) 3377 return true; 3378 3379 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) 3380 return false; 3381 3382 /* 3383 * if this request is as-good as one we would expect from the 3384 * current cfqq, let it preempt 3385 */ 3386 if (cfq_rq_close(cfqd, cfqq, rq)) 3387 return true; 3388 3389 return false; 3390} 3391 3392/* 3393 * cfqq preempts the active queue. if we allowed preempt with no slice left, 3394 * let it have half of its nominal slice. 3395 */ 3396static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) 3397{ 3398 struct cfq_queue *old_cfqq = cfqd->active_queue; 3399 3400 cfq_log_cfqq(cfqd, cfqq, "preempt"); 3401 cfq_slice_expired(cfqd, 1); 3402 3403 /* 3404 * workload type is changed, don't save slice, otherwise preempt 3405 * doesn't happen 3406 */ 3407 if (cfqq_type(old_cfqq) != cfqq_type(cfqq)) 3408 cfqq->cfqg->saved_workload_slice = 0; 3409 3410 /* 3411 * Put the new queue at the front of the of the current list, 3412 * so we know that it will be selected next. 3413 */ 3414 BUG_ON(!cfq_cfqq_on_rr(cfqq)); 3415 3416 cfq_service_tree_add(cfqd, cfqq, 1); 3417 3418 cfqq->slice_end = 0; 3419 cfq_mark_cfqq_slice_new(cfqq); 3420} 3421 3422/* 3423 * Called when a new fs request (rq) is added (to cfqq). Check if there's 3424 * something we should do about it 3425 */ 3426static void 3427cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, 3428 struct request *rq) 3429{ 3430 struct cfq_io_context *cic = RQ_CIC(rq); 3431 3432 cfqd->rq_queued++; 3433 if (rq->cmd_flags & REQ_PRIO) 3434 cfqq->prio_pending++; 3435 3436 cfq_update_io_thinktime(cfqd, cfqq, cic); 3437 cfq_update_io_seektime(cfqd, cfqq, rq); 3438 cfq_update_idle_window(cfqd, cfqq, cic); 3439 3440 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); 3441 3442 if (cfqq == cfqd->active_queue) { 3443 /* 3444 * Remember that we saw a request from this process, but 3445 * don't start queuing just yet. Otherwise we risk seeing lots 3446 * of tiny requests, because we disrupt the normal plugging 3447 * and merging. If the request is already larger than a single 3448 * page, let it rip immediately. For that case we assume that 3449 * merging is already done. Ditto for a busy system that 3450 * has other work pending, don't risk delaying until the 3451 * idle timer unplug to continue working. 3452 */ 3453 if (cfq_cfqq_wait_request(cfqq)) { 3454 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || 3455 cfqd->busy_queues > 1) { 3456 cfq_del_timer(cfqd, cfqq); 3457 cfq_clear_cfqq_wait_request(cfqq); 3458 __blk_run_queue(cfqd->queue); 3459 } else { 3460 cfq_blkiocg_update_idle_time_stats( 3461 &cfqq->cfqg->blkg); 3462 cfq_mark_cfqq_must_dispatch(cfqq); 3463 } 3464 } 3465 } else if (cfq_should_preempt(cfqd, cfqq, rq)) { 3466 /* 3467 * not the active queue - expire current slice if it is 3468 * idle and has expired it's mean thinktime or this new queue 3469 * has some old slice time left and is of higher priority or 3470 * this new queue is RT and the current one is BE 3471 */ 3472 cfq_preempt_queue(cfqd, cfqq); 3473 __blk_run_queue(cfqd->queue); 3474 } 3475} 3476 3477static void cfq_insert_request(struct request_queue *q, struct request *rq) 3478{ 3479 struct cfq_data *cfqd = q->elevator->elevator_data; 3480 struct cfq_queue *cfqq = RQ_CFQQ(rq); 3481 3482 cfq_log_cfqq(cfqd, cfqq, "insert_request"); 3483 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); 3484 3485 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); 3486 list_add_tail(&rq->queuelist, &cfqq->fifo); 3487 cfq_add_rq_rb(rq); 3488 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg, 3489 &cfqd->serving_group->blkg, rq_data_dir(rq), 3490 rq_is_sync(rq)); 3491 cfq_rq_enqueued(cfqd, cfqq, rq); 3492} 3493 3494/* 3495 * Update hw_tag based on peak queue depth over 50 samples under 3496 * sufficient load. 3497 */ 3498static void cfq_update_hw_tag(struct cfq_data *cfqd) 3499{ 3500 struct cfq_queue *cfqq = cfqd->active_queue; 3501 3502 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth) 3503 cfqd->hw_tag_est_depth = cfqd->rq_in_driver; 3504 3505 if (cfqd->hw_tag == 1) 3506 return; 3507 3508 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && 3509 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN) 3510 return; 3511 3512 /* 3513 * If active queue hasn't enough requests and can idle, cfq might not 3514 * dispatch sufficient requests to hardware. Don't zero hw_tag in this 3515 * case 3516 */ 3517 if (cfqq && cfq_cfqq_idle_window(cfqq) && 3518 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < 3519 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN) 3520 return; 3521 3522 if (cfqd->hw_tag_samples++ < 50) 3523 return; 3524 3525 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN) 3526 cfqd->hw_tag = 1; 3527 else 3528 cfqd->hw_tag = 0; 3529} 3530 3531static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq) 3532{ 3533 struct cfq_io_context *cic = cfqd->active_cic; 3534 3535 /* If the queue already has requests, don't wait */ 3536 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) 3537 return false; 3538 3539 /* If there are other queues in the group, don't wait */ 3540 if (cfqq->cfqg->nr_cfqq > 1) 3541 return false; 3542 3543 /* the only queue in the group, but think time is big */ 3544 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) 3545 return false; 3546 3547 if (cfq_slice_used(cfqq)) 3548 return true; 3549 3550 /* if slice left is less than think time, wait busy */ 3551 if (cic && sample_valid(cic->ttime.ttime_samples) 3552 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) 3553 return true; 3554 3555 /* 3556 * If think times is less than a jiffy than ttime_mean=0 and above 3557 * will not be true. It might happen that slice has not expired yet 3558 * but will expire soon (4-5 ns) during select_queue(). To cover the 3559 * case where think time is less than a jiffy, mark the queue wait 3560 * busy if only 1 jiffy is left in the slice. 3561 */ 3562 if (cfqq->slice_end - jiffies == 1) 3563 return true; 3564 3565 return false; 3566} 3567 3568static void cfq_completed_request(struct request_queue *q, struct request *rq) 3569{ 3570 struct cfq_queue *cfqq = RQ_CFQQ(rq); 3571 struct cfq_data *cfqd = cfqq->cfqd; 3572 const int sync = rq_is_sync(rq); 3573 unsigned long now; 3574 3575 now = jiffies; 3576 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", 3577 !!(rq->cmd_flags & REQ_NOIDLE)); 3578 3579 cfq_update_hw_tag(cfqd); 3580 3581 WARN_ON(!cfqd->rq_in_driver); 3582 WARN_ON(!cfqq->dispatched); 3583 cfqd->rq_in_driver--; 3584 cfqq->dispatched--; 3585 (RQ_CFQG(rq))->dispatched--; 3586 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg, 3587 rq_start_time_ns(rq), rq_io_start_time_ns(rq), 3588 rq_data_dir(rq), rq_is_sync(rq)); 3589 3590 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--; 3591 3592 if (sync) { 3593 struct cfq_rb_root *service_tree; 3594 3595 RQ_CIC(rq)->ttime.last_end_request = now; 3596 3597 if (cfq_cfqq_on_rr(cfqq)) 3598 service_tree = cfqq->service_tree; 3599 else 3600 service_tree = service_tree_for(cfqq->cfqg, 3601 cfqq_prio(cfqq), cfqq_type(cfqq)); 3602 service_tree->ttime.last_end_request = now; 3603 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now)) 3604 cfqd->last_delayed_sync = now; 3605 } 3606 3607#ifdef CONFIG_CFQ_GROUP_IOSCHED 3608 cfqq->cfqg->ttime.last_end_request = now; 3609#endif 3610 3611 /* 3612 * If this is the active queue, check if it needs to be expired, 3613 * or if we want to idle in case it has no pending requests. 3614 */ 3615 if (cfqd->active_queue == cfqq) { 3616 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); 3617 3618 if (cfq_cfqq_slice_new(cfqq)) { 3619 cfq_set_prio_slice(cfqd, cfqq); 3620 cfq_clear_cfqq_slice_new(cfqq); 3621 } 3622 3623 /* 3624 * Should we wait for next request to come in before we expire 3625 * the queue. 3626 */ 3627 if (cfq_should_wait_busy(cfqd, cfqq)) { 3628 unsigned long extend_sl = cfqd->cfq_slice_idle; 3629 if (!cfqd->cfq_slice_idle) 3630 extend_sl = cfqd->cfq_group_idle; 3631 cfqq->slice_end = jiffies + extend_sl; 3632 cfq_mark_cfqq_wait_busy(cfqq); 3633 cfq_log_cfqq(cfqd, cfqq, "will busy wait"); 3634 } 3635 3636 /* 3637 * Idling is not enabled on: 3638 * - expired queues 3639 * - idle-priority queues 3640 * - async queues 3641 * - queues with still some requests queued 3642 * - when there is a close cooperator 3643 */ 3644 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) 3645 cfq_slice_expired(cfqd, 1); 3646 else if (sync && cfqq_empty && 3647 !cfq_close_cooperator(cfqd, cfqq)) { 3648 cfq_arm_slice_timer(cfqd); 3649 } 3650 } 3651 3652 if (!cfqd->rq_in_driver) 3653 cfq_schedule_dispatch(cfqd); 3654} 3655 3656static inline int __cfq_may_queue(struct cfq_queue *cfqq) 3657{ 3658 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { 3659 cfq_mark_cfqq_must_alloc_slice(cfqq); 3660 return ELV_MQUEUE_MUST; 3661 } 3662 3663 return ELV_MQUEUE_MAY; 3664} 3665 3666static int cfq_may_queue(struct request_queue *q, int rw) 3667{ 3668 struct cfq_data *cfqd = q->elevator->elevator_data; 3669 struct task_struct *tsk = current; 3670 struct cfq_io_context *cic; 3671 struct cfq_queue *cfqq; 3672 3673 /* 3674 * don't force setup of a queue from here, as a call to may_queue 3675 * does not necessarily imply that a request actually will be queued. 3676 * so just lookup a possibly existing queue, or return 'may queue' 3677 * if that fails 3678 */ 3679 cic = cfq_cic_lookup(cfqd, tsk->io_context); 3680 if (!cic) 3681 return ELV_MQUEUE_MAY; 3682 3683 cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); 3684 if (cfqq) { 3685 cfq_init_prio_data(cfqq, cic->ioc); 3686 3687 return __cfq_may_queue(cfqq); 3688 } 3689 3690 return ELV_MQUEUE_MAY; 3691} 3692 3693/* 3694 * queue lock held here 3695 */ 3696static void cfq_put_request(struct request *rq) 3697{ 3698 struct cfq_queue *cfqq = RQ_CFQQ(rq); 3699 3700 if (cfqq) { 3701 const int rw = rq_data_dir(rq); 3702 3703 BUG_ON(!cfqq->allocated[rw]); 3704 cfqq->allocated[rw]--; 3705 3706 put_io_context(RQ_CIC(rq)->ioc); 3707 3708 rq->elevator_private[0] = NULL; 3709 rq->elevator_private[1] = NULL; 3710 3711 /* Put down rq reference on cfqg */ 3712 cfq_put_cfqg(RQ_CFQG(rq)); 3713 rq->elevator_private[2] = NULL; 3714 3715 cfq_put_queue(cfqq); 3716 } 3717} 3718 3719static struct cfq_queue * 3720cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic, 3721 struct cfq_queue *cfqq) 3722{ 3723 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); 3724 cic_set_cfqq(cic, cfqq->new_cfqq, 1); 3725 cfq_mark_cfqq_coop(cfqq->new_cfqq); 3726 cfq_put_queue(cfqq); 3727 return cic_to_cfqq(cic, 1); 3728} 3729 3730/* 3731 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this 3732 * was the last process referring to said cfqq. 3733 */ 3734static struct cfq_queue * 3735split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq) 3736{ 3737 if (cfqq_process_refs(cfqq) == 1) { 3738 cfqq->pid = current->pid; 3739 cfq_clear_cfqq_coop(cfqq); 3740 cfq_clear_cfqq_split_coop(cfqq); 3741 return cfqq; 3742 } 3743 3744 cic_set_cfqq(cic, NULL, 1); 3745 3746 cfq_put_cooperator(cfqq); 3747 3748 cfq_put_queue(cfqq); 3749 return NULL; 3750} 3751/* 3752 * Allocate cfq data structures associated with this request. 3753 */ 3754static int 3755cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) 3756{ 3757 struct cfq_data *cfqd = q->elevator->elevator_data; 3758 struct cfq_io_context *cic; 3759 const int rw = rq_data_dir(rq); 3760 const bool is_sync = rq_is_sync(rq); 3761 struct cfq_queue *cfqq; 3762 unsigned long flags; 3763 3764 might_sleep_if(gfp_mask & __GFP_WAIT); 3765 3766 cic = cfq_get_io_context(cfqd, gfp_mask); 3767 3768 spin_lock_irqsave(q->queue_lock, flags); 3769 3770 if (!cic) 3771 goto queue_fail; 3772 3773new_queue: 3774 cfqq = cic_to_cfqq(cic, is_sync); 3775 if (!cfqq || cfqq == &cfqd->oom_cfqq) { 3776 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); 3777 cic_set_cfqq(cic, cfqq, is_sync); 3778 } else { 3779 /* 3780 * If the queue was seeky for too long, break it apart. 3781 */ 3782 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) { 3783 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); 3784 cfqq = split_cfqq(cic, cfqq); 3785 if (!cfqq) 3786 goto new_queue; 3787 } 3788 3789 /* 3790 * Check to see if this queue is scheduled to merge with 3791 * another, closely cooperating queue. The merging of 3792 * queues happens here as it must be done in process context. 3793 * The reference on new_cfqq was taken in merge_cfqqs. 3794 */ 3795 if (cfqq->new_cfqq) 3796 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); 3797 } 3798 3799 cfqq->allocated[rw]++; 3800 3801 cfqq->ref++; 3802 rq->elevator_private[0] = cic; 3803 rq->elevator_private[1] = cfqq; 3804 rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg); 3805 spin_unlock_irqrestore(q->queue_lock, flags); 3806 return 0; 3807 3808queue_fail: 3809 cfq_schedule_dispatch(cfqd); 3810 spin_unlock_irqrestore(q->queue_lock, flags); 3811 cfq_log(cfqd, "set_request fail"); 3812 return 1; 3813} 3814 3815static void cfq_kick_queue(struct work_struct *work) 3816{ 3817 struct cfq_data *cfqd = 3818 container_of(work, struct cfq_data, unplug_work); 3819 struct request_queue *q = cfqd->queue; 3820 3821 spin_lock_irq(q->queue_lock); 3822 __blk_run_queue(cfqd->queue); 3823 spin_unlock_irq(q->queue_lock); 3824} 3825 3826/* 3827 * Timer running if the active_queue is currently idling inside its time slice 3828 */ 3829static void cfq_idle_slice_timer(unsigned long data) 3830{ 3831 struct cfq_data *cfqd = (struct cfq_data *) data; 3832 struct cfq_queue *cfqq; 3833 unsigned long flags; 3834 int timed_out = 1; 3835 3836 cfq_log(cfqd, "idle timer fired"); 3837 3838 spin_lock_irqsave(cfqd->queue->queue_lock, flags); 3839 3840 cfqq = cfqd->active_queue; 3841 if (cfqq) { 3842 timed_out = 0; 3843 3844 /* 3845 * We saw a request before the queue expired, let it through 3846 */ 3847 if (cfq_cfqq_must_dispatch(cfqq)) 3848 goto out_kick; 3849 3850 /* 3851 * expired 3852 */ 3853 if (cfq_slice_used(cfqq)) 3854 goto expire; 3855 3856 /* 3857 * only expire and reinvoke request handler, if there are 3858 * other queues with pending requests 3859 */ 3860 if (!cfqd->busy_queues) 3861 goto out_cont; 3862 3863 /* 3864 * not expired and it has a request pending, let it dispatch 3865 */ 3866 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) 3867 goto out_kick; 3868 3869 /* 3870 * Queue depth flag is reset only when the idle didn't succeed 3871 */ 3872 cfq_clear_cfqq_deep(cfqq); 3873 } 3874expire: 3875 cfq_slice_expired(cfqd, timed_out); 3876out_kick: 3877 cfq_schedule_dispatch(cfqd); 3878out_cont: 3879 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); 3880} 3881 3882static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) 3883{ 3884 del_timer_sync(&cfqd->idle_slice_timer); 3885 cancel_work_sync(&cfqd->unplug_work); 3886} 3887 3888static void cfq_put_async_queues(struct cfq_data *cfqd) 3889{ 3890 int i; 3891 3892 for (i = 0; i < IOPRIO_BE_NR; i++) { 3893 if (cfqd->async_cfqq[0][i]) 3894 cfq_put_queue(cfqd->async_cfqq[0][i]); 3895 if (cfqd->async_cfqq[1][i]) 3896 cfq_put_queue(cfqd->async_cfqq[1][i]); 3897 } 3898 3899 if (cfqd->async_idle_cfqq) 3900 cfq_put_queue(cfqd->async_idle_cfqq); 3901} 3902 3903static void cfq_exit_queue(struct elevator_queue *e) 3904{ 3905 struct cfq_data *cfqd = e->elevator_data; 3906 struct request_queue *q = cfqd->queue; 3907 bool wait = false; 3908 3909 cfq_shutdown_timer_wq(cfqd); 3910 3911 spin_lock_irq(q->queue_lock); 3912 3913 if (cfqd->active_queue) 3914 __cfq_slice_expired(cfqd, cfqd->active_queue, 0); 3915 3916 while (!list_empty(&cfqd->cic_list)) { 3917 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, 3918 struct cfq_io_context, 3919 queue_list); 3920 3921 cfq_exit_cic(cic); 3922 } 3923 3924 cfq_put_async_queues(cfqd); 3925 cfq_release_cfq_groups(cfqd); 3926 3927 /* 3928 * If there are groups which we could not unlink from blkcg list, 3929 * wait for a rcu period for them to be freed. 3930 */ 3931 if (cfqd->nr_blkcg_linked_grps) 3932 wait = true; 3933 3934 spin_unlock_irq(q->queue_lock); 3935 3936 cfq_shutdown_timer_wq(cfqd); 3937 3938 /* 3939 * Wait for cfqg->blkg->key accessors to exit their grace periods. 3940 * Do this wait only if there are other unlinked groups out 3941 * there. This can happen if cgroup deletion path claimed the 3942 * responsibility of cleaning up a group before queue cleanup code 3943 * get to the group. 3944 * 3945 * Do not call synchronize_rcu() unconditionally as there are drivers 3946 * which create/delete request queue hundreds of times during scan/boot 3947 * and synchronize_rcu() can take significant time and slow down boot. 3948 */ 3949 if (wait) 3950 synchronize_rcu(); 3951 3952#ifdef CONFIG_CFQ_GROUP_IOSCHED 3953 /* Free up per cpu stats for root group */ 3954 free_percpu(cfqd->root_group.blkg.stats_cpu); 3955#endif 3956 kfree(cfqd); 3957} 3958 3959static void *cfq_init_queue(struct request_queue *q) 3960{ 3961 struct cfq_data *cfqd; 3962 int i, j; 3963 struct cfq_group *cfqg; 3964 struct cfq_rb_root *st; 3965 3966 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); 3967 if (!cfqd) 3968 return NULL; 3969 3970 /* Init root service tree */ 3971 cfqd->grp_service_tree = CFQ_RB_ROOT; 3972 3973 /* Init root group */ 3974 cfqg = &cfqd->root_group; 3975 for_each_cfqg_st(cfqg, i, j, st) 3976 *st = CFQ_RB_ROOT; 3977 RB_CLEAR_NODE(&cfqg->rb_node); 3978 3979 /* Give preference to root group over other groups */ 3980 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT; 3981 3982#ifdef CONFIG_CFQ_GROUP_IOSCHED 3983 /* 3984 * Set root group reference to 2. One reference will be dropped when 3985 * all groups on cfqd->cfqg_list are being deleted during queue exit. 3986 * Other reference will remain there as we don't want to delete this 3987 * group as it is statically allocated and gets destroyed when 3988 * throtl_data goes away. 3989 */ 3990 cfqg->ref = 2; 3991 3992 if (blkio_alloc_blkg_stats(&cfqg->blkg)) { 3993 kfree(cfqg); 3994 kfree(cfqd); 3995 return NULL; 3996 } 3997 3998 rcu_read_lock(); 3999 4000 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, 4001 (void *)cfqd, 0); 4002 rcu_read_unlock(); 4003 cfqd->nr_blkcg_linked_grps++; 4004 4005 /* Add group on cfqd->cfqg_list */ 4006 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list); 4007#endif 4008 /* 4009 * Not strictly needed (since RB_ROOT just clears the node and we 4010 * zeroed cfqd on alloc), but better be safe in case someone decides 4011 * to add magic to the rb code 4012 */ 4013 for (i = 0; i < CFQ_PRIO_LISTS; i++) 4014 cfqd->prio_trees[i] = RB_ROOT; 4015 4016 /* 4017 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. 4018 * Grab a permanent reference to it, so that the normal code flow 4019 * will not attempt to free it. 4020 */ 4021 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); 4022 cfqd->oom_cfqq.ref++; 4023 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group); 4024 4025 INIT_LIST_HEAD(&cfqd->cic_list); 4026 4027 cfqd->queue = q; 4028 4029 init_timer(&cfqd->idle_slice_timer); 4030 cfqd->idle_slice_timer.function = cfq_idle_slice_timer; 4031 cfqd->idle_slice_timer.data = (unsigned long) cfqd; 4032 4033 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); 4034 4035 cfqd->cfq_quantum = cfq_quantum; 4036 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; 4037 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; 4038 cfqd->cfq_back_max = cfq_back_max; 4039 cfqd->cfq_back_penalty = cfq_back_penalty; 4040 cfqd->cfq_slice[0] = cfq_slice_async; 4041 cfqd->cfq_slice[1] = cfq_slice_sync; 4042 cfqd->cfq_slice_async_rq = cfq_slice_async_rq; 4043 cfqd->cfq_slice_idle = cfq_slice_idle; 4044 cfqd->cfq_group_idle = cfq_group_idle; 4045 cfqd->cfq_latency = 1; 4046 cfqd->hw_tag = -1; 4047 /* 4048 * we optimistically start assuming sync ops weren't delayed in last 4049 * second, in order to have larger depth for async operations. 4050 */ 4051 cfqd->last_delayed_sync = jiffies - HZ; 4052 return cfqd; 4053} 4054 4055static void cfq_slab_kill(void) 4056{ 4057 /* 4058 * Caller already ensured that pending RCU callbacks are completed, 4059 * so we should have no busy allocations at this point. 4060 */ 4061 if (cfq_pool) 4062 kmem_cache_destroy(cfq_pool); 4063 if (cfq_ioc_pool) 4064 kmem_cache_destroy(cfq_ioc_pool); 4065} 4066 4067static int __init cfq_slab_setup(void) 4068{ 4069 cfq_pool = KMEM_CACHE(cfq_queue, 0); 4070 if (!cfq_pool) 4071 goto fail; 4072 4073 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0); 4074 if (!cfq_ioc_pool) 4075 goto fail; 4076 4077 return 0; 4078fail: 4079 cfq_slab_kill(); 4080 return -ENOMEM; 4081} 4082 4083/* 4084 * sysfs parts below --> 4085 */ 4086static ssize_t 4087cfq_var_show(unsigned int var, char *page) 4088{ 4089 return sprintf(page, "%d\n", var); 4090} 4091 4092static ssize_t 4093cfq_var_store(unsigned int *var, const char *page, size_t count) 4094{ 4095 char *p = (char *) page; 4096 4097 *var = simple_strtoul(p, &p, 10); 4098 return count; 4099} 4100 4101#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ 4102static ssize_t __FUNC(struct elevator_queue *e, char *page) \ 4103{ \ 4104 struct cfq_data *cfqd = e->elevator_data; \ 4105 unsigned int __data = __VAR; \ 4106 if (__CONV) \ 4107 __data = jiffies_to_msecs(__data); \ 4108 return cfq_var_show(__data, (page)); \ 4109} 4110SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); 4111SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); 4112SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); 4113SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); 4114SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); 4115SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); 4116SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1); 4117SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); 4118SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); 4119SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); 4120SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); 4121#undef SHOW_FUNCTION 4122 4123#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ 4124static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ 4125{ \ 4126 struct cfq_data *cfqd = e->elevator_data; \ 4127 unsigned int __data; \ 4128 int ret = cfq_var_store(&__data, (page), count); \ 4129 if (__data < (MIN)) \ 4130 __data = (MIN); \ 4131 else if (__data > (MAX)) \ 4132 __data = (MAX); \ 4133 if (__CONV) \ 4134 *(__PTR) = msecs_to_jiffies(__data); \ 4135 else \ 4136 *(__PTR) = __data; \ 4137 return ret; \ 4138} 4139STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); 4140STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, 4141 UINT_MAX, 1); 4142STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, 4143 UINT_MAX, 1); 4144STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); 4145STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, 4146 UINT_MAX, 0); 4147STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); 4148STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1); 4149STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); 4150STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); 4151STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, 4152 UINT_MAX, 0); 4153STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); 4154#undef STORE_FUNCTION 4155 4156#define CFQ_ATTR(name) \ 4157 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) 4158 4159static struct elv_fs_entry cfq_attrs[] = { 4160 CFQ_ATTR(quantum), 4161 CFQ_ATTR(fifo_expire_sync), 4162 CFQ_ATTR(fifo_expire_async), 4163 CFQ_ATTR(back_seek_max), 4164 CFQ_ATTR(back_seek_penalty), 4165 CFQ_ATTR(slice_sync), 4166 CFQ_ATTR(slice_async), 4167 CFQ_ATTR(slice_async_rq), 4168 CFQ_ATTR(slice_idle), 4169 CFQ_ATTR(group_idle), 4170 CFQ_ATTR(low_latency), 4171 __ATTR_NULL 4172}; 4173 4174static struct elevator_type iosched_cfq = { 4175 .ops = { 4176 .elevator_merge_fn = cfq_merge, 4177 .elevator_merged_fn = cfq_merged_request, 4178 .elevator_merge_req_fn = cfq_merged_requests, 4179 .elevator_allow_merge_fn = cfq_allow_merge, 4180 .elevator_bio_merged_fn = cfq_bio_merged, 4181 .elevator_dispatch_fn = cfq_dispatch_requests, 4182 .elevator_add_req_fn = cfq_insert_request, 4183 .elevator_activate_req_fn = cfq_activate_request, 4184 .elevator_deactivate_req_fn = cfq_deactivate_request, 4185 .elevator_completed_req_fn = cfq_completed_request, 4186 .elevator_former_req_fn = elv_rb_former_request, 4187 .elevator_latter_req_fn = elv_rb_latter_request, 4188 .elevator_set_req_fn = cfq_set_request, 4189 .elevator_put_req_fn = cfq_put_request, 4190 .elevator_may_queue_fn = cfq_may_queue, 4191 .elevator_init_fn = cfq_init_queue, 4192 .elevator_exit_fn = cfq_exit_queue, 4193 .trim = cfq_free_io_context, 4194 }, 4195 .elevator_attrs = cfq_attrs, 4196 .elevator_name = "cfq", 4197 .elevator_owner = THIS_MODULE, 4198}; 4199 4200#ifdef CONFIG_CFQ_GROUP_IOSCHED 4201static struct blkio_policy_type blkio_policy_cfq = { 4202 .ops = { 4203 .blkio_unlink_group_fn = cfq_unlink_blkio_group, 4204 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight, 4205 }, 4206 .plid = BLKIO_POLICY_PROP, 4207}; 4208#else 4209static struct blkio_policy_type blkio_policy_cfq; 4210#endif 4211 4212static int __init cfq_init(void) 4213{ 4214 /* 4215 * could be 0 on HZ < 1000 setups 4216 */ 4217 if (!cfq_slice_async) 4218 cfq_slice_async = 1; 4219 if (!cfq_slice_idle) 4220 cfq_slice_idle = 1; 4221 4222#ifdef CONFIG_CFQ_GROUP_IOSCHED 4223 if (!cfq_group_idle) 4224 cfq_group_idle = 1; 4225#else 4226 cfq_group_idle = 0; 4227#endif 4228 if (cfq_slab_setup()) 4229 return -ENOMEM; 4230 4231 elv_register(&iosched_cfq); 4232 blkio_policy_register(&blkio_policy_cfq); 4233 4234 return 0; 4235} 4236 4237static void __exit cfq_exit(void) 4238{ 4239 DECLARE_COMPLETION_ONSTACK(all_gone); 4240 blkio_policy_unregister(&blkio_policy_cfq); 4241 elv_unregister(&iosched_cfq); 4242 ioc_gone = &all_gone; 4243 /* ioc_gone's update must be visible before reading ioc_count */ 4244 smp_wmb(); 4245 4246 /* 4247 * this also protects us from entering cfq_slab_kill() with 4248 * pending RCU callbacks 4249 */ 4250 if (elv_ioc_count_read(cfq_ioc_count)) 4251 wait_for_completion(&all_gone); 4252 cfq_slab_kill(); 4253} 4254 4255module_init(cfq_init); 4256module_exit(cfq_exit); 4257 4258MODULE_AUTHOR("Jens Axboe"); 4259MODULE_LICENSE("GPL"); 4260MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); 4261