cfq-iosched.c revision ae30c286553c91c49af5cbc0265a05a6543d0c52
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/blkdev.h> 11#include <linux/elevator.h> 12#include <linux/jiffies.h> 13#include <linux/rbtree.h> 14#include <linux/ioprio.h> 15#include <linux/blktrace_api.h> 16#include "blk-cgroup.h" 17 18/* 19 * tunables 20 */ 21/* max queue in one round of service */ 22static const int cfq_quantum = 4; 23static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; 24/* maximum backwards seek, in KiB */ 25static const int cfq_back_max = 16 * 1024; 26/* penalty of a backwards seek */ 27static const int cfq_back_penalty = 2; 28static const int cfq_slice_sync = HZ / 10; 29static int cfq_slice_async = HZ / 25; 30static const int cfq_slice_async_rq = 2; 31static int cfq_slice_idle = HZ / 125; 32static const int cfq_target_latency = HZ * 3/10; /* 300 ms */ 33static const int cfq_hist_divisor = 4; 34 35/* 36 * offset from end of service tree 37 */ 38#define CFQ_IDLE_DELAY (HZ / 5) 39 40/* 41 * below this threshold, we consider thinktime immediate 42 */ 43#define CFQ_MIN_TT (2) 44 45/* 46 * Allow merged cfqqs to perform this amount of seeky I/O before 47 * deciding to break the queues up again. 48 */ 49#define CFQQ_COOP_TOUT (HZ) 50 51#define CFQ_SLICE_SCALE (5) 52#define CFQ_HW_QUEUE_MIN (5) 53#define CFQ_SERVICE_SHIFT 12 54 55#define RQ_CIC(rq) \ 56 ((struct cfq_io_context *) (rq)->elevator_private) 57#define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2) 58 59static struct kmem_cache *cfq_pool; 60static struct kmem_cache *cfq_ioc_pool; 61 62static DEFINE_PER_CPU(unsigned long, cfq_ioc_count); 63static struct completion *ioc_gone; 64static DEFINE_SPINLOCK(ioc_gone_lock); 65 66#define CFQ_PRIO_LISTS IOPRIO_BE_NR 67#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) 68#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) 69 70#define sample_valid(samples) ((samples) > 80) 71#define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node) 72 73/* 74 * Most of our rbtree usage is for sorting with min extraction, so 75 * if we cache the leftmost node we don't have to walk down the tree 76 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should 77 * move this into the elevator for the rq sorting as well. 78 */ 79struct cfq_rb_root { 80 struct rb_root rb; 81 struct rb_node *left; 82 unsigned count; 83 u64 min_vdisktime; 84 struct rb_node *active; 85 unsigned total_weight; 86}; 87#define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, } 88 89/* 90 * Per process-grouping structure 91 */ 92struct cfq_queue { 93 /* reference count */ 94 atomic_t ref; 95 /* various state flags, see below */ 96 unsigned int flags; 97 /* parent cfq_data */ 98 struct cfq_data *cfqd; 99 /* service_tree member */ 100 struct rb_node rb_node; 101 /* service_tree key */ 102 unsigned long rb_key; 103 /* prio tree member */ 104 struct rb_node p_node; 105 /* prio tree root we belong to, if any */ 106 struct rb_root *p_root; 107 /* sorted list of pending requests */ 108 struct rb_root sort_list; 109 /* if fifo isn't expired, next request to serve */ 110 struct request *next_rq; 111 /* requests queued in sort_list */ 112 int queued[2]; 113 /* currently allocated requests */ 114 int allocated[2]; 115 /* fifo list of requests in sort_list */ 116 struct list_head fifo; 117 118 /* time when queue got scheduled in to dispatch first request. */ 119 unsigned long dispatch_start; 120 unsigned int allocated_slice; 121 /* time when first request from queue completed and slice started. */ 122 unsigned long slice_start; 123 unsigned long slice_end; 124 long slice_resid; 125 unsigned int slice_dispatch; 126 127 /* pending metadata requests */ 128 int meta_pending; 129 /* number of requests that are on the dispatch list or inside driver */ 130 int dispatched; 131 132 /* io prio of this group */ 133 unsigned short ioprio, org_ioprio; 134 unsigned short ioprio_class, org_ioprio_class; 135 136 unsigned int seek_samples; 137 u64 seek_total; 138 sector_t seek_mean; 139 sector_t last_request_pos; 140 unsigned long seeky_start; 141 142 pid_t pid; 143 144 struct cfq_rb_root *service_tree; 145 struct cfq_queue *new_cfqq; 146 struct cfq_group *cfqg; 147 struct cfq_group *orig_cfqg; 148 /* Sectors dispatched in current 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}; 161 162/* 163 * Second index in the service_trees. 164 */ 165enum wl_type_t { 166 ASYNC_WORKLOAD = 0, 167 SYNC_NOIDLE_WORKLOAD = 1, 168 SYNC_WORKLOAD = 2 169}; 170 171/* This is per cgroup per device grouping structure */ 172struct cfq_group { 173 /* group service_tree member */ 174 struct rb_node rb_node; 175 176 /* group service_tree key */ 177 u64 vdisktime; 178 unsigned int weight; 179 bool on_st; 180 181 /* number of cfqq currently on this group */ 182 int nr_cfqq; 183 184 /* Per group busy queus average. Useful for workload slice calc. */ 185 unsigned int busy_queues_avg[2]; 186 /* 187 * rr lists of queues with requests, onle rr for each priority class. 188 * Counts are embedded in the cfq_rb_root 189 */ 190 struct cfq_rb_root service_trees[2][3]; 191 struct cfq_rb_root service_tree_idle; 192 193 unsigned long saved_workload_slice; 194 enum wl_type_t saved_workload; 195 enum wl_prio_t saved_serving_prio; 196 struct blkio_group blkg; 197#ifdef CONFIG_CFQ_GROUP_IOSCHED 198 struct hlist_node cfqd_node; 199 atomic_t ref; 200#endif 201}; 202 203/* 204 * Per block device queue structure 205 */ 206struct cfq_data { 207 struct request_queue *queue; 208 /* Root service tree for cfq_groups */ 209 struct cfq_rb_root grp_service_tree; 210 struct cfq_group root_group; 211 /* Number of active cfq groups on group service tree */ 212 int nr_groups; 213 214 /* 215 * The priority currently being served 216 */ 217 enum wl_prio_t serving_prio; 218 enum wl_type_t serving_type; 219 unsigned long workload_expires; 220 struct cfq_group *serving_group; 221 bool noidle_tree_requires_idle; 222 223 /* 224 * Each priority tree is sorted by next_request position. These 225 * trees are used when determining if two or more queues are 226 * interleaving requests (see cfq_close_cooperator). 227 */ 228 struct rb_root prio_trees[CFQ_PRIO_LISTS]; 229 230 unsigned int busy_queues; 231 232 int rq_in_driver[2]; 233 int sync_flight; 234 235 /* 236 * queue-depth detection 237 */ 238 int rq_queued; 239 int hw_tag; 240 /* 241 * hw_tag can be 242 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection) 243 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth) 244 * 0 => no NCQ 245 */ 246 int hw_tag_est_depth; 247 unsigned int hw_tag_samples; 248 249 /* 250 * idle window management 251 */ 252 struct timer_list idle_slice_timer; 253 struct work_struct unplug_work; 254 255 struct cfq_queue *active_queue; 256 struct cfq_io_context *active_cic; 257 258 /* 259 * async queue for each priority case 260 */ 261 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; 262 struct cfq_queue *async_idle_cfqq; 263 264 sector_t last_position; 265 266 /* 267 * tunables, see top of file 268 */ 269 unsigned int cfq_quantum; 270 unsigned int cfq_fifo_expire[2]; 271 unsigned int cfq_back_penalty; 272 unsigned int cfq_back_max; 273 unsigned int cfq_slice[2]; 274 unsigned int cfq_slice_async_rq; 275 unsigned int cfq_slice_idle; 276 unsigned int cfq_latency; 277 unsigned int cfq_group_isolation; 278 279 struct list_head cic_list; 280 281 /* 282 * Fallback dummy cfqq for extreme OOM conditions 283 */ 284 struct cfq_queue oom_cfqq; 285 286 unsigned long last_end_sync_rq; 287 288 /* List of cfq groups being managed on this device*/ 289 struct hlist_head cfqg_list; 290}; 291 292static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd); 293 294static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg, 295 enum wl_prio_t prio, 296 enum wl_type_t type, 297 struct cfq_data *cfqd) 298{ 299 if (!cfqg) 300 return NULL; 301 302 if (prio == IDLE_WORKLOAD) 303 return &cfqg->service_tree_idle; 304 305 return &cfqg->service_trees[prio][type]; 306} 307 308enum cfqq_state_flags { 309 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ 310 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ 311 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ 312 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ 313 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ 314 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ 315 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ 316 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ 317 CFQ_CFQQ_FLAG_sync, /* synchronous queue */ 318 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */ 319 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */ 320 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */ 321 CFQ_CFQQ_FLAG_wait_busy_done, /* Got new request. Expire the queue */ 322}; 323 324#define CFQ_CFQQ_FNS(name) \ 325static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ 326{ \ 327 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ 328} \ 329static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ 330{ \ 331 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ 332} \ 333static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ 334{ \ 335 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ 336} 337 338CFQ_CFQQ_FNS(on_rr); 339CFQ_CFQQ_FNS(wait_request); 340CFQ_CFQQ_FNS(must_dispatch); 341CFQ_CFQQ_FNS(must_alloc_slice); 342CFQ_CFQQ_FNS(fifo_expire); 343CFQ_CFQQ_FNS(idle_window); 344CFQ_CFQQ_FNS(prio_changed); 345CFQ_CFQQ_FNS(slice_new); 346CFQ_CFQQ_FNS(sync); 347CFQ_CFQQ_FNS(coop); 348CFQ_CFQQ_FNS(deep); 349CFQ_CFQQ_FNS(wait_busy); 350CFQ_CFQQ_FNS(wait_busy_done); 351#undef CFQ_CFQQ_FNS 352 353#ifdef CONFIG_DEBUG_CFQ_IOSCHED 354#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ 355 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \ 356 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \ 357 blkg_path(&(cfqq)->cfqg->blkg), ##args); 358 359#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \ 360 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \ 361 blkg_path(&(cfqg)->blkg), ##args); \ 362 363#else 364#define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ 365 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) 366#define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0); 367#endif 368#define cfq_log(cfqd, fmt, args...) \ 369 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) 370 371/* Traverses through cfq group service trees */ 372#define for_each_cfqg_st(cfqg, i, j, st) \ 373 for (i = 0; i <= IDLE_WORKLOAD; i++) \ 374 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\ 375 : &cfqg->service_tree_idle; \ 376 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \ 377 (i == IDLE_WORKLOAD && j == 0); \ 378 j++, st = i < IDLE_WORKLOAD ? \ 379 &cfqg->service_trees[i][j]: NULL) \ 380 381 382static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq) 383{ 384 if (cfq_class_idle(cfqq)) 385 return IDLE_WORKLOAD; 386 if (cfq_class_rt(cfqq)) 387 return RT_WORKLOAD; 388 return BE_WORKLOAD; 389} 390 391 392static enum wl_type_t cfqq_type(struct cfq_queue *cfqq) 393{ 394 if (!cfq_cfqq_sync(cfqq)) 395 return ASYNC_WORKLOAD; 396 if (!cfq_cfqq_idle_window(cfqq)) 397 return SYNC_NOIDLE_WORKLOAD; 398 return SYNC_WORKLOAD; 399} 400 401static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl, 402 struct cfq_data *cfqd, 403 struct cfq_group *cfqg) 404{ 405 if (wl == IDLE_WORKLOAD) 406 return cfqg->service_tree_idle.count; 407 408 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count 409 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count 410 + cfqg->service_trees[wl][SYNC_WORKLOAD].count; 411} 412 413static inline int cfqg_busy_async_queues(struct cfq_data *cfqd, 414 struct cfq_group *cfqg) 415{ 416 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count 417 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count; 418} 419 420static void cfq_dispatch_insert(struct request_queue *, struct request *); 421static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool, 422 struct io_context *, gfp_t); 423static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, 424 struct io_context *); 425 426static inline int rq_in_driver(struct cfq_data *cfqd) 427{ 428 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1]; 429} 430 431static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, 432 bool is_sync) 433{ 434 return cic->cfqq[is_sync]; 435} 436 437static inline void cic_set_cfqq(struct cfq_io_context *cic, 438 struct cfq_queue *cfqq, bool is_sync) 439{ 440 cic->cfqq[is_sync] = cfqq; 441} 442 443/* 444 * We regard a request as SYNC, if it's either a read or has the SYNC bit 445 * set (in which case it could also be direct WRITE). 446 */ 447static inline bool cfq_bio_sync(struct bio *bio) 448{ 449 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO); 450} 451 452/* 453 * scheduler run of queue, if there are requests pending and no one in the 454 * driver that will restart queueing 455 */ 456static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) 457{ 458 if (cfqd->busy_queues) { 459 cfq_log(cfqd, "schedule dispatch"); 460 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); 461 } 462} 463 464static int cfq_queue_empty(struct request_queue *q) 465{ 466 struct cfq_data *cfqd = q->elevator->elevator_data; 467 468 return !cfqd->rq_queued; 469} 470 471/* 472 * Scale schedule slice based on io priority. Use the sync time slice only 473 * if a queue is marked sync and has sync io queued. A sync queue with async 474 * io only, should not get full sync slice length. 475 */ 476static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync, 477 unsigned short prio) 478{ 479 const int base_slice = cfqd->cfq_slice[sync]; 480 481 WARN_ON(prio >= IOPRIO_BE_NR); 482 483 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); 484} 485 486static inline int 487cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) 488{ 489 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); 490} 491 492static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg) 493{ 494 u64 d = delta << CFQ_SERVICE_SHIFT; 495 496 d = d * BLKIO_WEIGHT_DEFAULT; 497 do_div(d, cfqg->weight); 498 return d; 499} 500 501static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime) 502{ 503 s64 delta = (s64)(vdisktime - min_vdisktime); 504 if (delta > 0) 505 min_vdisktime = vdisktime; 506 507 return min_vdisktime; 508} 509 510static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime) 511{ 512 s64 delta = (s64)(vdisktime - min_vdisktime); 513 if (delta < 0) 514 min_vdisktime = vdisktime; 515 516 return min_vdisktime; 517} 518 519static void update_min_vdisktime(struct cfq_rb_root *st) 520{ 521 u64 vdisktime = st->min_vdisktime; 522 struct cfq_group *cfqg; 523 524 if (st->active) { 525 cfqg = rb_entry_cfqg(st->active); 526 vdisktime = cfqg->vdisktime; 527 } 528 529 if (st->left) { 530 cfqg = rb_entry_cfqg(st->left); 531 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime); 532 } 533 534 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime); 535} 536 537/* 538 * get averaged number of queues of RT/BE priority. 539 * average is updated, with a formula that gives more weight to higher numbers, 540 * to quickly follows sudden increases and decrease slowly 541 */ 542 543static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd, 544 struct cfq_group *cfqg, bool rt) 545{ 546 unsigned min_q, max_q; 547 unsigned mult = cfq_hist_divisor - 1; 548 unsigned round = cfq_hist_divisor / 2; 549 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg); 550 551 min_q = min(cfqg->busy_queues_avg[rt], busy); 552 max_q = max(cfqg->busy_queues_avg[rt], busy); 553 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) / 554 cfq_hist_divisor; 555 return cfqg->busy_queues_avg[rt]; 556} 557 558static inline unsigned 559cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg) 560{ 561 struct cfq_rb_root *st = &cfqd->grp_service_tree; 562 563 return cfq_target_latency * cfqg->weight / st->total_weight; 564} 565 566static inline void 567cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) 568{ 569 unsigned slice = cfq_prio_to_slice(cfqd, cfqq); 570 if (cfqd->cfq_latency) { 571 /* 572 * interested queues (we consider only the ones with the same 573 * priority class in the cfq group) 574 */ 575 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg, 576 cfq_class_rt(cfqq)); 577 unsigned sync_slice = cfqd->cfq_slice[1]; 578 unsigned expect_latency = sync_slice * iq; 579 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg); 580 581 if (expect_latency > group_slice) { 582 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle; 583 /* scale low_slice according to IO priority 584 * and sync vs async */ 585 unsigned low_slice = 586 min(slice, base_low_slice * slice / sync_slice); 587 /* the adapted slice value is scaled to fit all iqs 588 * into the target latency */ 589 slice = max(slice * group_slice / expect_latency, 590 low_slice); 591 } 592 } 593 cfqq->slice_start = jiffies; 594 cfqq->slice_end = jiffies + slice; 595 cfqq->allocated_slice = slice; 596 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); 597} 598 599/* 600 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end 601 * isn't valid until the first request from the dispatch is activated 602 * and the slice time set. 603 */ 604static inline bool cfq_slice_used(struct cfq_queue *cfqq) 605{ 606 if (cfq_cfqq_slice_new(cfqq)) 607 return 0; 608 if (time_before(jiffies, cfqq->slice_end)) 609 return 0; 610 611 return 1; 612} 613 614/* 615 * Lifted from AS - choose which of rq1 and rq2 that is best served now. 616 * We choose the request that is closest to the head right now. Distance 617 * behind the head is penalized and only allowed to a certain extent. 618 */ 619static struct request * 620cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last) 621{ 622 sector_t s1, s2, d1 = 0, d2 = 0; 623 unsigned long back_max; 624#define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ 625#define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ 626 unsigned wrap = 0; /* bit mask: requests behind the disk head? */ 627 628 if (rq1 == NULL || rq1 == rq2) 629 return rq2; 630 if (rq2 == NULL) 631 return rq1; 632 633 if (rq_is_sync(rq1) && !rq_is_sync(rq2)) 634 return rq1; 635 else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) 636 return rq2; 637 if (rq_is_meta(rq1) && !rq_is_meta(rq2)) 638 return rq1; 639 else if (rq_is_meta(rq2) && !rq_is_meta(rq1)) 640 return rq2; 641 642 s1 = blk_rq_pos(rq1); 643 s2 = blk_rq_pos(rq2); 644 645 /* 646 * by definition, 1KiB is 2 sectors 647 */ 648 back_max = cfqd->cfq_back_max * 2; 649 650 /* 651 * Strict one way elevator _except_ in the case where we allow 652 * short backward seeks which are biased as twice the cost of a 653 * similar forward seek. 654 */ 655 if (s1 >= last) 656 d1 = s1 - last; 657 else if (s1 + back_max >= last) 658 d1 = (last - s1) * cfqd->cfq_back_penalty; 659 else 660 wrap |= CFQ_RQ1_WRAP; 661 662 if (s2 >= last) 663 d2 = s2 - last; 664 else if (s2 + back_max >= last) 665 d2 = (last - s2) * cfqd->cfq_back_penalty; 666 else 667 wrap |= CFQ_RQ2_WRAP; 668 669 /* Found required data */ 670 671 /* 672 * By doing switch() on the bit mask "wrap" we avoid having to 673 * check two variables for all permutations: --> faster! 674 */ 675 switch (wrap) { 676 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ 677 if (d1 < d2) 678 return rq1; 679 else if (d2 < d1) 680 return rq2; 681 else { 682 if (s1 >= s2) 683 return rq1; 684 else 685 return rq2; 686 } 687 688 case CFQ_RQ2_WRAP: 689 return rq1; 690 case CFQ_RQ1_WRAP: 691 return rq2; 692 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ 693 default: 694 /* 695 * Since both rqs are wrapped, 696 * start with the one that's further behind head 697 * (--> only *one* back seek required), 698 * since back seek takes more time than forward. 699 */ 700 if (s1 <= s2) 701 return rq1; 702 else 703 return rq2; 704 } 705} 706 707/* 708 * The below is leftmost cache rbtree addon 709 */ 710static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) 711{ 712 /* Service tree is empty */ 713 if (!root->count) 714 return NULL; 715 716 if (!root->left) 717 root->left = rb_first(&root->rb); 718 719 if (root->left) 720 return rb_entry(root->left, struct cfq_queue, rb_node); 721 722 return NULL; 723} 724 725static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root) 726{ 727 if (!root->left) 728 root->left = rb_first(&root->rb); 729 730 if (root->left) 731 return rb_entry_cfqg(root->left); 732 733 return NULL; 734} 735 736static void rb_erase_init(struct rb_node *n, struct rb_root *root) 737{ 738 rb_erase(n, root); 739 RB_CLEAR_NODE(n); 740} 741 742static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) 743{ 744 if (root->left == n) 745 root->left = NULL; 746 rb_erase_init(n, &root->rb); 747 --root->count; 748} 749 750/* 751 * would be nice to take fifo expire time into account as well 752 */ 753static struct request * 754cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, 755 struct request *last) 756{ 757 struct rb_node *rbnext = rb_next(&last->rb_node); 758 struct rb_node *rbprev = rb_prev(&last->rb_node); 759 struct request *next = NULL, *prev = NULL; 760 761 BUG_ON(RB_EMPTY_NODE(&last->rb_node)); 762 763 if (rbprev) 764 prev = rb_entry_rq(rbprev); 765 766 if (rbnext) 767 next = rb_entry_rq(rbnext); 768 else { 769 rbnext = rb_first(&cfqq->sort_list); 770 if (rbnext && rbnext != &last->rb_node) 771 next = rb_entry_rq(rbnext); 772 } 773 774 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last)); 775} 776 777static unsigned long cfq_slice_offset(struct cfq_data *cfqd, 778 struct cfq_queue *cfqq) 779{ 780 /* 781 * just an approximation, should be ok. 782 */ 783 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) - 784 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); 785} 786 787static inline s64 788cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg) 789{ 790 return cfqg->vdisktime - st->min_vdisktime; 791} 792 793static void 794__cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) 795{ 796 struct rb_node **node = &st->rb.rb_node; 797 struct rb_node *parent = NULL; 798 struct cfq_group *__cfqg; 799 s64 key = cfqg_key(st, cfqg); 800 int left = 1; 801 802 while (*node != NULL) { 803 parent = *node; 804 __cfqg = rb_entry_cfqg(parent); 805 806 if (key < cfqg_key(st, __cfqg)) 807 node = &parent->rb_left; 808 else { 809 node = &parent->rb_right; 810 left = 0; 811 } 812 } 813 814 if (left) 815 st->left = &cfqg->rb_node; 816 817 rb_link_node(&cfqg->rb_node, parent, node); 818 rb_insert_color(&cfqg->rb_node, &st->rb); 819} 820 821static void 822cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg) 823{ 824 struct cfq_rb_root *st = &cfqd->grp_service_tree; 825 struct cfq_group *__cfqg; 826 struct rb_node *n; 827 828 cfqg->nr_cfqq++; 829 if (cfqg->on_st) 830 return; 831 832 /* 833 * Currently put the group at the end. Later implement something 834 * so that groups get lesser vtime based on their weights, so that 835 * if group does not loose all if it was not continously backlogged. 836 */ 837 n = rb_last(&st->rb); 838 if (n) { 839 __cfqg = rb_entry_cfqg(n); 840 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY; 841 } else 842 cfqg->vdisktime = st->min_vdisktime; 843 844 __cfq_group_service_tree_add(st, cfqg); 845 cfqg->on_st = true; 846 cfqd->nr_groups++; 847 st->total_weight += cfqg->weight; 848} 849 850static void 851cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg) 852{ 853 struct cfq_rb_root *st = &cfqd->grp_service_tree; 854 855 if (st->active == &cfqg->rb_node) 856 st->active = NULL; 857 858 BUG_ON(cfqg->nr_cfqq < 1); 859 cfqg->nr_cfqq--; 860 861 /* If there are other cfq queues under this group, don't delete it */ 862 if (cfqg->nr_cfqq) 863 return; 864 865 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group"); 866 cfqg->on_st = false; 867 cfqd->nr_groups--; 868 st->total_weight -= cfqg->weight; 869 if (!RB_EMPTY_NODE(&cfqg->rb_node)) 870 cfq_rb_erase(&cfqg->rb_node, st); 871 cfqg->saved_workload_slice = 0; 872 blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1); 873} 874 875static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq) 876{ 877 unsigned int slice_used; 878 879 /* 880 * Queue got expired before even a single request completed or 881 * got expired immediately after first request completion. 882 */ 883 if (!cfqq->slice_start || cfqq->slice_start == jiffies) { 884 /* 885 * Also charge the seek time incurred to the group, otherwise 886 * if there are mutiple queues in the group, each can dispatch 887 * a single request on seeky media and cause lots of seek time 888 * and group will never know it. 889 */ 890 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start), 891 1); 892 } else { 893 slice_used = jiffies - cfqq->slice_start; 894 if (slice_used > cfqq->allocated_slice) 895 slice_used = cfqq->allocated_slice; 896 } 897 898 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used, 899 cfqq->nr_sectors); 900 return slice_used; 901} 902 903static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg, 904 struct cfq_queue *cfqq) 905{ 906 struct cfq_rb_root *st = &cfqd->grp_service_tree; 907 unsigned int used_sl, charge_sl; 908 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg) 909 - cfqg->service_tree_idle.count; 910 911 BUG_ON(nr_sync < 0); 912 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq); 913 914 if (!cfq_cfqq_sync(cfqq) && !nr_sync) 915 charge_sl = cfqq->allocated_slice; 916 917 /* Can't update vdisktime while group is on service tree */ 918 cfq_rb_erase(&cfqg->rb_node, st); 919 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg); 920 __cfq_group_service_tree_add(st, cfqg); 921 922 /* This group is being expired. Save the context */ 923 if (time_after(cfqd->workload_expires, jiffies)) { 924 cfqg->saved_workload_slice = cfqd->workload_expires 925 - jiffies; 926 cfqg->saved_workload = cfqd->serving_type; 927 cfqg->saved_serving_prio = cfqd->serving_prio; 928 } else 929 cfqg->saved_workload_slice = 0; 930 931 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime, 932 st->min_vdisktime); 933 blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl, 934 cfqq->nr_sectors); 935} 936 937#ifdef CONFIG_CFQ_GROUP_IOSCHED 938static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg) 939{ 940 if (blkg) 941 return container_of(blkg, struct cfq_group, blkg); 942 return NULL; 943} 944 945void 946cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight) 947{ 948 cfqg_of_blkg(blkg)->weight = weight; 949} 950 951static struct cfq_group * 952cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create) 953{ 954 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup); 955 struct cfq_group *cfqg = NULL; 956 void *key = cfqd; 957 int i, j; 958 struct cfq_rb_root *st; 959 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info; 960 unsigned int major, minor; 961 962 /* Do we need to take this reference */ 963 if (!css_tryget(&blkcg->css)) 964 return NULL;; 965 966 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key)); 967 if (cfqg || !create) 968 goto done; 969 970 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node); 971 if (!cfqg) 972 goto done; 973 974 cfqg->weight = blkcg->weight; 975 for_each_cfqg_st(cfqg, i, j, st) 976 *st = CFQ_RB_ROOT; 977 RB_CLEAR_NODE(&cfqg->rb_node); 978 979 /* 980 * Take the initial reference that will be released on destroy 981 * This can be thought of a joint reference by cgroup and 982 * elevator which will be dropped by either elevator exit 983 * or cgroup deletion path depending on who is exiting first. 984 */ 985 atomic_set(&cfqg->ref, 1); 986 987 /* Add group onto cgroup list */ 988 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); 989 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd, 990 MKDEV(major, minor)); 991 992 /* Add group on cfqd list */ 993 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list); 994 995done: 996 css_put(&blkcg->css); 997 return cfqg; 998} 999 1000/* 1001 * Search for the cfq group current task belongs to. If create = 1, then also 1002 * create the cfq group if it does not exist. request_queue lock must be held. 1003 */ 1004static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create) 1005{ 1006 struct cgroup *cgroup; 1007 struct cfq_group *cfqg = NULL; 1008 1009 rcu_read_lock(); 1010 cgroup = task_cgroup(current, blkio_subsys_id); 1011 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create); 1012 if (!cfqg && create) 1013 cfqg = &cfqd->root_group; 1014 rcu_read_unlock(); 1015 return cfqg; 1016} 1017 1018static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) 1019{ 1020 /* Currently, all async queues are mapped to root group */ 1021 if (!cfq_cfqq_sync(cfqq)) 1022 cfqg = &cfqq->cfqd->root_group; 1023 1024 cfqq->cfqg = cfqg; 1025 /* cfqq reference on cfqg */ 1026 atomic_inc(&cfqq->cfqg->ref); 1027} 1028 1029static void cfq_put_cfqg(struct cfq_group *cfqg) 1030{ 1031 struct cfq_rb_root *st; 1032 int i, j; 1033 1034 BUG_ON(atomic_read(&cfqg->ref) <= 0); 1035 if (!atomic_dec_and_test(&cfqg->ref)) 1036 return; 1037 for_each_cfqg_st(cfqg, i, j, st) 1038 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL); 1039 kfree(cfqg); 1040} 1041 1042static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg) 1043{ 1044 /* Something wrong if we are trying to remove same group twice */ 1045 BUG_ON(hlist_unhashed(&cfqg->cfqd_node)); 1046 1047 hlist_del_init(&cfqg->cfqd_node); 1048 1049 /* 1050 * Put the reference taken at the time of creation so that when all 1051 * queues are gone, group can be destroyed. 1052 */ 1053 cfq_put_cfqg(cfqg); 1054} 1055 1056static void cfq_release_cfq_groups(struct cfq_data *cfqd) 1057{ 1058 struct hlist_node *pos, *n; 1059 struct cfq_group *cfqg; 1060 1061 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) { 1062 /* 1063 * If cgroup removal path got to blk_group first and removed 1064 * it from cgroup list, then it will take care of destroying 1065 * cfqg also. 1066 */ 1067 if (!blkiocg_del_blkio_group(&cfqg->blkg)) 1068 cfq_destroy_cfqg(cfqd, cfqg); 1069 } 1070} 1071 1072/* 1073 * Blk cgroup controller notification saying that blkio_group object is being 1074 * delinked as associated cgroup object is going away. That also means that 1075 * no new IO will come in this group. So get rid of this group as soon as 1076 * any pending IO in the group is finished. 1077 * 1078 * This function is called under rcu_read_lock(). key is the rcu protected 1079 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu 1080 * read lock. 1081 * 1082 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means 1083 * it should not be NULL as even if elevator was exiting, cgroup deltion 1084 * path got to it first. 1085 */ 1086void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg) 1087{ 1088 unsigned long flags; 1089 struct cfq_data *cfqd = key; 1090 1091 spin_lock_irqsave(cfqd->queue->queue_lock, flags); 1092 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg)); 1093 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); 1094} 1095 1096#else /* GROUP_IOSCHED */ 1097static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create) 1098{ 1099 return &cfqd->root_group; 1100} 1101static inline void 1102cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) { 1103 cfqq->cfqg = cfqg; 1104} 1105 1106static void cfq_release_cfq_groups(struct cfq_data *cfqd) {} 1107static inline void cfq_put_cfqg(struct cfq_group *cfqg) {} 1108 1109#endif /* GROUP_IOSCHED */ 1110 1111/* 1112 * The cfqd->service_trees holds all pending cfq_queue's that have 1113 * requests waiting to be processed. It is sorted in the order that 1114 * we will service the queues. 1115 */ 1116static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1117 bool add_front) 1118{ 1119 struct rb_node **p, *parent; 1120 struct cfq_queue *__cfqq; 1121 unsigned long rb_key; 1122 struct cfq_rb_root *service_tree; 1123 int left; 1124 int new_cfqq = 1; 1125 int group_changed = 0; 1126 1127#ifdef CONFIG_CFQ_GROUP_IOSCHED 1128 if (!cfqd->cfq_group_isolation 1129 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD 1130 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) { 1131 /* Move this cfq to root group */ 1132 cfq_log_cfqq(cfqd, cfqq, "moving to root group"); 1133 if (!RB_EMPTY_NODE(&cfqq->rb_node)) 1134 cfq_group_service_tree_del(cfqd, cfqq->cfqg); 1135 cfqq->orig_cfqg = cfqq->cfqg; 1136 cfqq->cfqg = &cfqd->root_group; 1137 atomic_inc(&cfqd->root_group.ref); 1138 group_changed = 1; 1139 } else if (!cfqd->cfq_group_isolation 1140 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) { 1141 /* cfqq is sequential now needs to go to its original group */ 1142 BUG_ON(cfqq->cfqg != &cfqd->root_group); 1143 if (!RB_EMPTY_NODE(&cfqq->rb_node)) 1144 cfq_group_service_tree_del(cfqd, cfqq->cfqg); 1145 cfq_put_cfqg(cfqq->cfqg); 1146 cfqq->cfqg = cfqq->orig_cfqg; 1147 cfqq->orig_cfqg = NULL; 1148 group_changed = 1; 1149 cfq_log_cfqq(cfqd, cfqq, "moved to origin group"); 1150 } 1151#endif 1152 1153 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq), 1154 cfqq_type(cfqq), cfqd); 1155 if (cfq_class_idle(cfqq)) { 1156 rb_key = CFQ_IDLE_DELAY; 1157 parent = rb_last(&service_tree->rb); 1158 if (parent && parent != &cfqq->rb_node) { 1159 __cfqq = rb_entry(parent, struct cfq_queue, rb_node); 1160 rb_key += __cfqq->rb_key; 1161 } else 1162 rb_key += jiffies; 1163 } else if (!add_front) { 1164 /* 1165 * Get our rb key offset. Subtract any residual slice 1166 * value carried from last service. A negative resid 1167 * count indicates slice overrun, and this should position 1168 * the next service time further away in the tree. 1169 */ 1170 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; 1171 rb_key -= cfqq->slice_resid; 1172 cfqq->slice_resid = 0; 1173 } else { 1174 rb_key = -HZ; 1175 __cfqq = cfq_rb_first(service_tree); 1176 rb_key += __cfqq ? __cfqq->rb_key : jiffies; 1177 } 1178 1179 if (!RB_EMPTY_NODE(&cfqq->rb_node)) { 1180 new_cfqq = 0; 1181 /* 1182 * same position, nothing more to do 1183 */ 1184 if (rb_key == cfqq->rb_key && 1185 cfqq->service_tree == service_tree) 1186 return; 1187 1188 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); 1189 cfqq->service_tree = NULL; 1190 } 1191 1192 left = 1; 1193 parent = NULL; 1194 cfqq->service_tree = service_tree; 1195 p = &service_tree->rb.rb_node; 1196 while (*p) { 1197 struct rb_node **n; 1198 1199 parent = *p; 1200 __cfqq = rb_entry(parent, struct cfq_queue, rb_node); 1201 1202 /* 1203 * sort by key, that represents service time. 1204 */ 1205 if (time_before(rb_key, __cfqq->rb_key)) 1206 n = &(*p)->rb_left; 1207 else { 1208 n = &(*p)->rb_right; 1209 left = 0; 1210 } 1211 1212 p = n; 1213 } 1214 1215 if (left) 1216 service_tree->left = &cfqq->rb_node; 1217 1218 cfqq->rb_key = rb_key; 1219 rb_link_node(&cfqq->rb_node, parent, p); 1220 rb_insert_color(&cfqq->rb_node, &service_tree->rb); 1221 service_tree->count++; 1222 if ((add_front || !new_cfqq) && !group_changed) 1223 return; 1224 cfq_group_service_tree_add(cfqd, cfqq->cfqg); 1225} 1226 1227static struct cfq_queue * 1228cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, 1229 sector_t sector, struct rb_node **ret_parent, 1230 struct rb_node ***rb_link) 1231{ 1232 struct rb_node **p, *parent; 1233 struct cfq_queue *cfqq = NULL; 1234 1235 parent = NULL; 1236 p = &root->rb_node; 1237 while (*p) { 1238 struct rb_node **n; 1239 1240 parent = *p; 1241 cfqq = rb_entry(parent, struct cfq_queue, p_node); 1242 1243 /* 1244 * Sort strictly based on sector. Smallest to the left, 1245 * largest to the right. 1246 */ 1247 if (sector > blk_rq_pos(cfqq->next_rq)) 1248 n = &(*p)->rb_right; 1249 else if (sector < blk_rq_pos(cfqq->next_rq)) 1250 n = &(*p)->rb_left; 1251 else 1252 break; 1253 p = n; 1254 cfqq = NULL; 1255 } 1256 1257 *ret_parent = parent; 1258 if (rb_link) 1259 *rb_link = p; 1260 return cfqq; 1261} 1262 1263static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1264{ 1265 struct rb_node **p, *parent; 1266 struct cfq_queue *__cfqq; 1267 1268 if (cfqq->p_root) { 1269 rb_erase(&cfqq->p_node, cfqq->p_root); 1270 cfqq->p_root = NULL; 1271 } 1272 1273 if (cfq_class_idle(cfqq)) 1274 return; 1275 if (!cfqq->next_rq) 1276 return; 1277 1278 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; 1279 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, 1280 blk_rq_pos(cfqq->next_rq), &parent, &p); 1281 if (!__cfqq) { 1282 rb_link_node(&cfqq->p_node, parent, p); 1283 rb_insert_color(&cfqq->p_node, cfqq->p_root); 1284 } else 1285 cfqq->p_root = NULL; 1286} 1287 1288/* 1289 * Update cfqq's position in the service tree. 1290 */ 1291static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1292{ 1293 /* 1294 * Resorting requires the cfqq to be on the RR list already. 1295 */ 1296 if (cfq_cfqq_on_rr(cfqq)) { 1297 cfq_service_tree_add(cfqd, cfqq, 0); 1298 cfq_prio_tree_add(cfqd, cfqq); 1299 } 1300} 1301 1302/* 1303 * add to busy list of queues for service, trying to be fair in ordering 1304 * the pending list according to last request service 1305 */ 1306static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1307{ 1308 cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); 1309 BUG_ON(cfq_cfqq_on_rr(cfqq)); 1310 cfq_mark_cfqq_on_rr(cfqq); 1311 cfqd->busy_queues++; 1312 1313 cfq_resort_rr_list(cfqd, cfqq); 1314} 1315 1316/* 1317 * Called when the cfqq no longer has requests pending, remove it from 1318 * the service tree. 1319 */ 1320static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1321{ 1322 cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); 1323 BUG_ON(!cfq_cfqq_on_rr(cfqq)); 1324 cfq_clear_cfqq_on_rr(cfqq); 1325 1326 if (!RB_EMPTY_NODE(&cfqq->rb_node)) { 1327 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); 1328 cfqq->service_tree = NULL; 1329 } 1330 if (cfqq->p_root) { 1331 rb_erase(&cfqq->p_node, cfqq->p_root); 1332 cfqq->p_root = NULL; 1333 } 1334 1335 cfq_group_service_tree_del(cfqd, cfqq->cfqg); 1336 BUG_ON(!cfqd->busy_queues); 1337 cfqd->busy_queues--; 1338} 1339 1340/* 1341 * rb tree support functions 1342 */ 1343static void cfq_del_rq_rb(struct request *rq) 1344{ 1345 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1346 const int sync = rq_is_sync(rq); 1347 1348 BUG_ON(!cfqq->queued[sync]); 1349 cfqq->queued[sync]--; 1350 1351 elv_rb_del(&cfqq->sort_list, rq); 1352 1353 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) { 1354 /* 1355 * Queue will be deleted from service tree when we actually 1356 * expire it later. Right now just remove it from prio tree 1357 * as it is empty. 1358 */ 1359 if (cfqq->p_root) { 1360 rb_erase(&cfqq->p_node, cfqq->p_root); 1361 cfqq->p_root = NULL; 1362 } 1363 } 1364} 1365 1366static void cfq_add_rq_rb(struct request *rq) 1367{ 1368 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1369 struct cfq_data *cfqd = cfqq->cfqd; 1370 struct request *__alias, *prev; 1371 1372 cfqq->queued[rq_is_sync(rq)]++; 1373 1374 /* 1375 * looks a little odd, but the first insert might return an alias. 1376 * if that happens, put the alias on the dispatch list 1377 */ 1378 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL) 1379 cfq_dispatch_insert(cfqd->queue, __alias); 1380 1381 if (!cfq_cfqq_on_rr(cfqq)) 1382 cfq_add_cfqq_rr(cfqd, cfqq); 1383 1384 /* 1385 * check if this request is a better next-serve candidate 1386 */ 1387 prev = cfqq->next_rq; 1388 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position); 1389 1390 /* 1391 * adjust priority tree position, if ->next_rq changes 1392 */ 1393 if (prev != cfqq->next_rq) 1394 cfq_prio_tree_add(cfqd, cfqq); 1395 1396 BUG_ON(!cfqq->next_rq); 1397} 1398 1399static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) 1400{ 1401 elv_rb_del(&cfqq->sort_list, rq); 1402 cfqq->queued[rq_is_sync(rq)]--; 1403 cfq_add_rq_rb(rq); 1404} 1405 1406static struct request * 1407cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) 1408{ 1409 struct task_struct *tsk = current; 1410 struct cfq_io_context *cic; 1411 struct cfq_queue *cfqq; 1412 1413 cic = cfq_cic_lookup(cfqd, tsk->io_context); 1414 if (!cic) 1415 return NULL; 1416 1417 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); 1418 if (cfqq) { 1419 sector_t sector = bio->bi_sector + bio_sectors(bio); 1420 1421 return elv_rb_find(&cfqq->sort_list, sector); 1422 } 1423 1424 return NULL; 1425} 1426 1427static void cfq_activate_request(struct request_queue *q, struct request *rq) 1428{ 1429 struct cfq_data *cfqd = q->elevator->elevator_data; 1430 1431 cfqd->rq_in_driver[rq_is_sync(rq)]++; 1432 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", 1433 rq_in_driver(cfqd)); 1434 1435 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); 1436} 1437 1438static void cfq_deactivate_request(struct request_queue *q, struct request *rq) 1439{ 1440 struct cfq_data *cfqd = q->elevator->elevator_data; 1441 const int sync = rq_is_sync(rq); 1442 1443 WARN_ON(!cfqd->rq_in_driver[sync]); 1444 cfqd->rq_in_driver[sync]--; 1445 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", 1446 rq_in_driver(cfqd)); 1447} 1448 1449static void cfq_remove_request(struct request *rq) 1450{ 1451 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1452 1453 if (cfqq->next_rq == rq) 1454 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); 1455 1456 list_del_init(&rq->queuelist); 1457 cfq_del_rq_rb(rq); 1458 1459 cfqq->cfqd->rq_queued--; 1460 if (rq_is_meta(rq)) { 1461 WARN_ON(!cfqq->meta_pending); 1462 cfqq->meta_pending--; 1463 } 1464} 1465 1466static int cfq_merge(struct request_queue *q, struct request **req, 1467 struct bio *bio) 1468{ 1469 struct cfq_data *cfqd = q->elevator->elevator_data; 1470 struct request *__rq; 1471 1472 __rq = cfq_find_rq_fmerge(cfqd, bio); 1473 if (__rq && elv_rq_merge_ok(__rq, bio)) { 1474 *req = __rq; 1475 return ELEVATOR_FRONT_MERGE; 1476 } 1477 1478 return ELEVATOR_NO_MERGE; 1479} 1480 1481static void cfq_merged_request(struct request_queue *q, struct request *req, 1482 int type) 1483{ 1484 if (type == ELEVATOR_FRONT_MERGE) { 1485 struct cfq_queue *cfqq = RQ_CFQQ(req); 1486 1487 cfq_reposition_rq_rb(cfqq, req); 1488 } 1489} 1490 1491static void 1492cfq_merged_requests(struct request_queue *q, struct request *rq, 1493 struct request *next) 1494{ 1495 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1496 /* 1497 * reposition in fifo if next is older than rq 1498 */ 1499 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && 1500 time_before(rq_fifo_time(next), rq_fifo_time(rq))) { 1501 list_move(&rq->queuelist, &next->queuelist); 1502 rq_set_fifo_time(rq, rq_fifo_time(next)); 1503 } 1504 1505 if (cfqq->next_rq == next) 1506 cfqq->next_rq = rq; 1507 cfq_remove_request(next); 1508} 1509 1510static int cfq_allow_merge(struct request_queue *q, struct request *rq, 1511 struct bio *bio) 1512{ 1513 struct cfq_data *cfqd = q->elevator->elevator_data; 1514 struct cfq_io_context *cic; 1515 struct cfq_queue *cfqq; 1516 1517 /* Deny merge if bio and rq don't belong to same cfq group */ 1518 if ((RQ_CFQQ(rq))->cfqg != cfq_get_cfqg(cfqd, 0)) 1519 return false; 1520 /* 1521 * Disallow merge of a sync bio into an async request. 1522 */ 1523 if (cfq_bio_sync(bio) && !rq_is_sync(rq)) 1524 return false; 1525 1526 /* 1527 * Lookup the cfqq that this bio will be queued with. Allow 1528 * merge only if rq is queued there. 1529 */ 1530 cic = cfq_cic_lookup(cfqd, current->io_context); 1531 if (!cic) 1532 return false; 1533 1534 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); 1535 return cfqq == RQ_CFQQ(rq); 1536} 1537 1538static void __cfq_set_active_queue(struct cfq_data *cfqd, 1539 struct cfq_queue *cfqq) 1540{ 1541 if (cfqq) { 1542 cfq_log_cfqq(cfqd, cfqq, "set_active"); 1543 cfqq->slice_start = 0; 1544 cfqq->dispatch_start = jiffies; 1545 cfqq->allocated_slice = 0; 1546 cfqq->slice_end = 0; 1547 cfqq->slice_dispatch = 0; 1548 cfqq->nr_sectors = 0; 1549 1550 cfq_clear_cfqq_wait_request(cfqq); 1551 cfq_clear_cfqq_must_dispatch(cfqq); 1552 cfq_clear_cfqq_must_alloc_slice(cfqq); 1553 cfq_clear_cfqq_fifo_expire(cfqq); 1554 cfq_mark_cfqq_slice_new(cfqq); 1555 1556 del_timer(&cfqd->idle_slice_timer); 1557 } 1558 1559 cfqd->active_queue = cfqq; 1560} 1561 1562/* 1563 * current cfqq expired its slice (or was too idle), select new one 1564 */ 1565static void 1566__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1567 bool timed_out) 1568{ 1569 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); 1570 1571 if (cfq_cfqq_wait_request(cfqq)) 1572 del_timer(&cfqd->idle_slice_timer); 1573 1574 cfq_clear_cfqq_wait_request(cfqq); 1575 cfq_clear_cfqq_wait_busy(cfqq); 1576 cfq_clear_cfqq_wait_busy_done(cfqq); 1577 1578 /* 1579 * store what was left of this slice, if the queue idled/timed out 1580 */ 1581 if (timed_out && !cfq_cfqq_slice_new(cfqq)) { 1582 cfqq->slice_resid = cfqq->slice_end - jiffies; 1583 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); 1584 } 1585 1586 cfq_group_served(cfqd, cfqq->cfqg, cfqq); 1587 1588 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) 1589 cfq_del_cfqq_rr(cfqd, cfqq); 1590 1591 cfq_resort_rr_list(cfqd, cfqq); 1592 1593 if (cfqq == cfqd->active_queue) 1594 cfqd->active_queue = NULL; 1595 1596 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active) 1597 cfqd->grp_service_tree.active = NULL; 1598 1599 if (cfqd->active_cic) { 1600 put_io_context(cfqd->active_cic->ioc); 1601 cfqd->active_cic = NULL; 1602 } 1603} 1604 1605static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) 1606{ 1607 struct cfq_queue *cfqq = cfqd->active_queue; 1608 1609 if (cfqq) 1610 __cfq_slice_expired(cfqd, cfqq, timed_out); 1611} 1612 1613/* 1614 * Get next queue for service. Unless we have a queue preemption, 1615 * we'll simply select the first cfqq in the service tree. 1616 */ 1617static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) 1618{ 1619 struct cfq_rb_root *service_tree = 1620 service_tree_for(cfqd->serving_group, cfqd->serving_prio, 1621 cfqd->serving_type, cfqd); 1622 1623 if (!cfqd->rq_queued) 1624 return NULL; 1625 1626 /* There is nothing to dispatch */ 1627 if (!service_tree) 1628 return NULL; 1629 if (RB_EMPTY_ROOT(&service_tree->rb)) 1630 return NULL; 1631 return cfq_rb_first(service_tree); 1632} 1633 1634static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd) 1635{ 1636 struct cfq_group *cfqg; 1637 struct cfq_queue *cfqq; 1638 int i, j; 1639 struct cfq_rb_root *st; 1640 1641 if (!cfqd->rq_queued) 1642 return NULL; 1643 1644 cfqg = cfq_get_next_cfqg(cfqd); 1645 if (!cfqg) 1646 return NULL; 1647 1648 for_each_cfqg_st(cfqg, i, j, st) 1649 if ((cfqq = cfq_rb_first(st)) != NULL) 1650 return cfqq; 1651 return NULL; 1652} 1653 1654/* 1655 * Get and set a new active queue for service. 1656 */ 1657static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, 1658 struct cfq_queue *cfqq) 1659{ 1660 if (!cfqq) 1661 cfqq = cfq_get_next_queue(cfqd); 1662 1663 __cfq_set_active_queue(cfqd, cfqq); 1664 return cfqq; 1665} 1666 1667static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, 1668 struct request *rq) 1669{ 1670 if (blk_rq_pos(rq) >= cfqd->last_position) 1671 return blk_rq_pos(rq) - cfqd->last_position; 1672 else 1673 return cfqd->last_position - blk_rq_pos(rq); 1674} 1675 1676#define CFQQ_SEEK_THR 8 * 1024 1677#define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR) 1678 1679static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1680 struct request *rq) 1681{ 1682 sector_t sdist = cfqq->seek_mean; 1683 1684 if (!sample_valid(cfqq->seek_samples)) 1685 sdist = CFQQ_SEEK_THR; 1686 1687 return cfq_dist_from_last(cfqd, rq) <= sdist; 1688} 1689 1690static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, 1691 struct cfq_queue *cur_cfqq) 1692{ 1693 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; 1694 struct rb_node *parent, *node; 1695 struct cfq_queue *__cfqq; 1696 sector_t sector = cfqd->last_position; 1697 1698 if (RB_EMPTY_ROOT(root)) 1699 return NULL; 1700 1701 /* 1702 * First, if we find a request starting at the end of the last 1703 * request, choose it. 1704 */ 1705 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); 1706 if (__cfqq) 1707 return __cfqq; 1708 1709 /* 1710 * If the exact sector wasn't found, the parent of the NULL leaf 1711 * will contain the closest sector. 1712 */ 1713 __cfqq = rb_entry(parent, struct cfq_queue, p_node); 1714 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) 1715 return __cfqq; 1716 1717 if (blk_rq_pos(__cfqq->next_rq) < sector) 1718 node = rb_next(&__cfqq->p_node); 1719 else 1720 node = rb_prev(&__cfqq->p_node); 1721 if (!node) 1722 return NULL; 1723 1724 __cfqq = rb_entry(node, struct cfq_queue, p_node); 1725 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) 1726 return __cfqq; 1727 1728 return NULL; 1729} 1730 1731/* 1732 * cfqd - obvious 1733 * cur_cfqq - passed in so that we don't decide that the current queue is 1734 * closely cooperating with itself. 1735 * 1736 * So, basically we're assuming that that cur_cfqq has dispatched at least 1737 * one request, and that cfqd->last_position reflects a position on the disk 1738 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid 1739 * assumption. 1740 */ 1741static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, 1742 struct cfq_queue *cur_cfqq) 1743{ 1744 struct cfq_queue *cfqq; 1745 1746 if (!cfq_cfqq_sync(cur_cfqq)) 1747 return NULL; 1748 if (CFQQ_SEEKY(cur_cfqq)) 1749 return NULL; 1750 1751 /* 1752 * We should notice if some of the queues are cooperating, eg 1753 * working closely on the same area of the disk. In that case, 1754 * we can group them together and don't waste time idling. 1755 */ 1756 cfqq = cfqq_close(cfqd, cur_cfqq); 1757 if (!cfqq) 1758 return NULL; 1759 1760 /* If new queue belongs to different cfq_group, don't choose it */ 1761 if (cur_cfqq->cfqg != cfqq->cfqg) 1762 return NULL; 1763 1764 /* 1765 * It only makes sense to merge sync queues. 1766 */ 1767 if (!cfq_cfqq_sync(cfqq)) 1768 return NULL; 1769 if (CFQQ_SEEKY(cfqq)) 1770 return NULL; 1771 1772 /* 1773 * Do not merge queues of different priority classes 1774 */ 1775 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) 1776 return NULL; 1777 1778 return cfqq; 1779} 1780 1781/* 1782 * Determine whether we should enforce idle window for this queue. 1783 */ 1784 1785static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1786{ 1787 enum wl_prio_t prio = cfqq_prio(cfqq); 1788 struct cfq_rb_root *service_tree = cfqq->service_tree; 1789 1790 BUG_ON(!service_tree); 1791 BUG_ON(!service_tree->count); 1792 1793 /* We never do for idle class queues. */ 1794 if (prio == IDLE_WORKLOAD) 1795 return false; 1796 1797 /* We do for queues that were marked with idle window flag. */ 1798 if (cfq_cfqq_idle_window(cfqq)) 1799 return true; 1800 1801 /* 1802 * Otherwise, we do only if they are the last ones 1803 * in their service tree. 1804 */ 1805 return service_tree->count == 1; 1806} 1807 1808static void cfq_arm_slice_timer(struct cfq_data *cfqd) 1809{ 1810 struct cfq_queue *cfqq = cfqd->active_queue; 1811 struct cfq_io_context *cic; 1812 unsigned long sl; 1813 1814 /* 1815 * SSD device without seek penalty, disable idling. But only do so 1816 * for devices that support queuing, otherwise we still have a problem 1817 * with sync vs async workloads. 1818 */ 1819 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) 1820 return; 1821 1822 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); 1823 WARN_ON(cfq_cfqq_slice_new(cfqq)); 1824 1825 /* 1826 * idle is disabled, either manually or by past process history 1827 */ 1828 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq)) 1829 return; 1830 1831 /* 1832 * still active requests from this queue, don't idle 1833 */ 1834 if (cfqq->dispatched) 1835 return; 1836 1837 /* 1838 * task has exited, don't wait 1839 */ 1840 cic = cfqd->active_cic; 1841 if (!cic || !atomic_read(&cic->ioc->nr_tasks)) 1842 return; 1843 1844 /* 1845 * If our average think time is larger than the remaining time 1846 * slice, then don't idle. This avoids overrunning the allotted 1847 * time slice. 1848 */ 1849 if (sample_valid(cic->ttime_samples) && 1850 (cfqq->slice_end - jiffies < cic->ttime_mean)) 1851 return; 1852 1853 cfq_mark_cfqq_wait_request(cfqq); 1854 1855 sl = cfqd->cfq_slice_idle; 1856 1857 mod_timer(&cfqd->idle_slice_timer, jiffies + sl); 1858 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl); 1859} 1860 1861/* 1862 * Move request from internal lists to the request queue dispatch list. 1863 */ 1864static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) 1865{ 1866 struct cfq_data *cfqd = q->elevator->elevator_data; 1867 struct cfq_queue *cfqq = RQ_CFQQ(rq); 1868 1869 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); 1870 1871 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); 1872 cfq_remove_request(rq); 1873 cfqq->dispatched++; 1874 elv_dispatch_sort(q, rq); 1875 1876 if (cfq_cfqq_sync(cfqq)) 1877 cfqd->sync_flight++; 1878 cfqq->nr_sectors += blk_rq_sectors(rq); 1879} 1880 1881/* 1882 * return expired entry, or NULL to just start from scratch in rbtree 1883 */ 1884static struct request *cfq_check_fifo(struct cfq_queue *cfqq) 1885{ 1886 struct request *rq = NULL; 1887 1888 if (cfq_cfqq_fifo_expire(cfqq)) 1889 return NULL; 1890 1891 cfq_mark_cfqq_fifo_expire(cfqq); 1892 1893 if (list_empty(&cfqq->fifo)) 1894 return NULL; 1895 1896 rq = rq_entry_fifo(cfqq->fifo.next); 1897 if (time_before(jiffies, rq_fifo_time(rq))) 1898 rq = NULL; 1899 1900 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); 1901 return rq; 1902} 1903 1904static inline int 1905cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1906{ 1907 const int base_rq = cfqd->cfq_slice_async_rq; 1908 1909 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); 1910 1911 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio)); 1912} 1913 1914/* 1915 * Must be called with the queue_lock held. 1916 */ 1917static int cfqq_process_refs(struct cfq_queue *cfqq) 1918{ 1919 int process_refs, io_refs; 1920 1921 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; 1922 process_refs = atomic_read(&cfqq->ref) - io_refs; 1923 BUG_ON(process_refs < 0); 1924 return process_refs; 1925} 1926 1927static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) 1928{ 1929 int process_refs, new_process_refs; 1930 struct cfq_queue *__cfqq; 1931 1932 /* Avoid a circular list and skip interim queue merges */ 1933 while ((__cfqq = new_cfqq->new_cfqq)) { 1934 if (__cfqq == cfqq) 1935 return; 1936 new_cfqq = __cfqq; 1937 } 1938 1939 process_refs = cfqq_process_refs(cfqq); 1940 /* 1941 * If the process for the cfqq has gone away, there is no 1942 * sense in merging the queues. 1943 */ 1944 if (process_refs == 0) 1945 return; 1946 1947 /* 1948 * Merge in the direction of the lesser amount of work. 1949 */ 1950 new_process_refs = cfqq_process_refs(new_cfqq); 1951 if (new_process_refs >= process_refs) { 1952 cfqq->new_cfqq = new_cfqq; 1953 atomic_add(process_refs, &new_cfqq->ref); 1954 } else { 1955 new_cfqq->new_cfqq = cfqq; 1956 atomic_add(new_process_refs, &cfqq->ref); 1957 } 1958} 1959 1960static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, 1961 struct cfq_group *cfqg, enum wl_prio_t prio, 1962 bool prio_changed) 1963{ 1964 struct cfq_queue *queue; 1965 int i; 1966 bool key_valid = false; 1967 unsigned long lowest_key = 0; 1968 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD; 1969 1970 if (prio_changed) { 1971 /* 1972 * When priorities switched, we prefer starting 1973 * from SYNC_NOIDLE (first choice), or just SYNC 1974 * over ASYNC 1975 */ 1976 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count) 1977 return cur_best; 1978 cur_best = SYNC_WORKLOAD; 1979 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count) 1980 return cur_best; 1981 1982 return ASYNC_WORKLOAD; 1983 } 1984 1985 for (i = 0; i < 3; ++i) { 1986 /* otherwise, select the one with lowest rb_key */ 1987 queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd)); 1988 if (queue && 1989 (!key_valid || time_before(queue->rb_key, lowest_key))) { 1990 lowest_key = queue->rb_key; 1991 cur_best = i; 1992 key_valid = true; 1993 } 1994 } 1995 1996 return cur_best; 1997} 1998 1999static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg) 2000{ 2001 enum wl_prio_t previous_prio = cfqd->serving_prio; 2002 bool prio_changed; 2003 unsigned slice; 2004 unsigned count; 2005 struct cfq_rb_root *st; 2006 unsigned group_slice; 2007 2008 if (!cfqg) { 2009 cfqd->serving_prio = IDLE_WORKLOAD; 2010 cfqd->workload_expires = jiffies + 1; 2011 return; 2012 } 2013 2014 /* Choose next priority. RT > BE > IDLE */ 2015 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg)) 2016 cfqd->serving_prio = RT_WORKLOAD; 2017 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg)) 2018 cfqd->serving_prio = BE_WORKLOAD; 2019 else { 2020 cfqd->serving_prio = IDLE_WORKLOAD; 2021 cfqd->workload_expires = jiffies + 1; 2022 return; 2023 } 2024 2025 /* 2026 * For RT and BE, we have to choose also the type 2027 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload 2028 * expiration time 2029 */ 2030 prio_changed = (cfqd->serving_prio != previous_prio); 2031 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type, 2032 cfqd); 2033 count = st->count; 2034 2035 /* 2036 * If priority didn't change, check workload expiration, 2037 * and that we still have other queues ready 2038 */ 2039 if (!prio_changed && count && 2040 !time_after(jiffies, cfqd->workload_expires)) 2041 return; 2042 2043 /* otherwise select new workload type */ 2044 cfqd->serving_type = 2045 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed); 2046 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type, 2047 cfqd); 2048 count = st->count; 2049 2050 /* 2051 * the workload slice is computed as a fraction of target latency 2052 * proportional to the number of queues in that workload, over 2053 * all the queues in the same priority class 2054 */ 2055 group_slice = cfq_group_slice(cfqd, cfqg); 2056 2057 slice = group_slice * count / 2058 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio], 2059 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg)); 2060 2061 if (cfqd->serving_type == ASYNC_WORKLOAD) { 2062 unsigned int tmp; 2063 2064 /* 2065 * Async queues are currently system wide. Just taking 2066 * proportion of queues with-in same group will lead to higher 2067 * async ratio system wide as generally root group is going 2068 * to have higher weight. A more accurate thing would be to 2069 * calculate system wide asnc/sync ratio. 2070 */ 2071 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg); 2072 tmp = tmp/cfqd->busy_queues; 2073 slice = min_t(unsigned, slice, tmp); 2074 2075 /* async workload slice is scaled down according to 2076 * the sync/async slice ratio. */ 2077 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1]; 2078 } else 2079 /* sync workload slice is at least 2 * cfq_slice_idle */ 2080 slice = max(slice, 2 * cfqd->cfq_slice_idle); 2081 2082 slice = max_t(unsigned, slice, CFQ_MIN_TT); 2083 cfqd->workload_expires = jiffies + slice; 2084 cfqd->noidle_tree_requires_idle = false; 2085} 2086 2087static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd) 2088{ 2089 struct cfq_rb_root *st = &cfqd->grp_service_tree; 2090 struct cfq_group *cfqg; 2091 2092 if (RB_EMPTY_ROOT(&st->rb)) 2093 return NULL; 2094 cfqg = cfq_rb_first_group(st); 2095 st->active = &cfqg->rb_node; 2096 update_min_vdisktime(st); 2097 return cfqg; 2098} 2099 2100static void cfq_choose_cfqg(struct cfq_data *cfqd) 2101{ 2102 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd); 2103 2104 cfqd->serving_group = cfqg; 2105 2106 /* Restore the workload type data */ 2107 if (cfqg->saved_workload_slice) { 2108 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice; 2109 cfqd->serving_type = cfqg->saved_workload; 2110 cfqd->serving_prio = cfqg->saved_serving_prio; 2111 } 2112 choose_service_tree(cfqd, cfqg); 2113} 2114 2115/* 2116 * Select a queue for service. If we have a current active queue, 2117 * check whether to continue servicing it, or retrieve and set a new one. 2118 */ 2119static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) 2120{ 2121 struct cfq_queue *cfqq, *new_cfqq = NULL; 2122 2123 cfqq = cfqd->active_queue; 2124 if (!cfqq) 2125 goto new_queue; 2126 2127 if (!cfqd->rq_queued) 2128 return NULL; 2129 /* 2130 * The active queue has run out of time, expire it and select new. 2131 */ 2132 if ((cfq_slice_used(cfqq) || cfq_cfqq_wait_busy_done(cfqq)) 2133 && !cfq_cfqq_must_dispatch(cfqq)) 2134 goto expire; 2135 2136 /* 2137 * The active queue has requests and isn't expired, allow it to 2138 * dispatch. 2139 */ 2140 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) 2141 goto keep_queue; 2142 2143 /* 2144 * If another queue has a request waiting within our mean seek 2145 * distance, let it run. The expire code will check for close 2146 * cooperators and put the close queue at the front of the service 2147 * tree. If possible, merge the expiring queue with the new cfqq. 2148 */ 2149 new_cfqq = cfq_close_cooperator(cfqd, cfqq); 2150 if (new_cfqq) { 2151 if (!cfqq->new_cfqq) 2152 cfq_setup_merge(cfqq, new_cfqq); 2153 goto expire; 2154 } 2155 2156 /* 2157 * No requests pending. If the active queue still has requests in 2158 * flight or is idling for a new request, allow either of these 2159 * conditions to happen (or time out) before selecting a new queue. 2160 */ 2161 if (timer_pending(&cfqd->idle_slice_timer) || 2162 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) { 2163 cfqq = NULL; 2164 goto keep_queue; 2165 } 2166 2167expire: 2168 cfq_slice_expired(cfqd, 0); 2169new_queue: 2170 /* 2171 * Current queue expired. Check if we have to switch to a new 2172 * service tree 2173 */ 2174 if (!new_cfqq) 2175 cfq_choose_cfqg(cfqd); 2176 2177 cfqq = cfq_set_active_queue(cfqd, new_cfqq); 2178keep_queue: 2179 return cfqq; 2180} 2181 2182static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) 2183{ 2184 int dispatched = 0; 2185 2186 while (cfqq->next_rq) { 2187 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); 2188 dispatched++; 2189 } 2190 2191 BUG_ON(!list_empty(&cfqq->fifo)); 2192 2193 /* By default cfqq is not expired if it is empty. Do it explicitly */ 2194 __cfq_slice_expired(cfqq->cfqd, cfqq, 0); 2195 return dispatched; 2196} 2197 2198/* 2199 * Drain our current requests. Used for barriers and when switching 2200 * io schedulers on-the-fly. 2201 */ 2202static int cfq_forced_dispatch(struct cfq_data *cfqd) 2203{ 2204 struct cfq_queue *cfqq; 2205 int dispatched = 0; 2206 2207 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) 2208 dispatched += __cfq_forced_dispatch_cfqq(cfqq); 2209 2210 cfq_slice_expired(cfqd, 0); 2211 BUG_ON(cfqd->busy_queues); 2212 2213 cfq_log(cfqd, "forced_dispatch=%d", dispatched); 2214 return dispatched; 2215} 2216 2217static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) 2218{ 2219 unsigned int max_dispatch; 2220 2221 /* 2222 * Drain async requests before we start sync IO 2223 */ 2224 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC]) 2225 return false; 2226 2227 /* 2228 * If this is an async queue and we have sync IO in flight, let it wait 2229 */ 2230 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) 2231 return false; 2232 2233 max_dispatch = cfqd->cfq_quantum; 2234 if (cfq_class_idle(cfqq)) 2235 max_dispatch = 1; 2236 2237 /* 2238 * Does this cfqq already have too much IO in flight? 2239 */ 2240 if (cfqq->dispatched >= max_dispatch) { 2241 /* 2242 * idle queue must always only have a single IO in flight 2243 */ 2244 if (cfq_class_idle(cfqq)) 2245 return false; 2246 2247 /* 2248 * We have other queues, don't allow more IO from this one 2249 */ 2250 if (cfqd->busy_queues > 1) 2251 return false; 2252 2253 /* 2254 * Sole queue user, no limit 2255 */ 2256 max_dispatch = -1; 2257 } 2258 2259 /* 2260 * Async queues must wait a bit before being allowed dispatch. 2261 * We also ramp up the dispatch depth gradually for async IO, 2262 * based on the last sync IO we serviced 2263 */ 2264 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { 2265 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq; 2266 unsigned int depth; 2267 2268 depth = last_sync / cfqd->cfq_slice[1]; 2269 if (!depth && !cfqq->dispatched) 2270 depth = 1; 2271 if (depth < max_dispatch) 2272 max_dispatch = depth; 2273 } 2274 2275 /* 2276 * If we're below the current max, allow a dispatch 2277 */ 2278 return cfqq->dispatched < max_dispatch; 2279} 2280 2281/* 2282 * Dispatch a request from cfqq, moving them to the request queue 2283 * dispatch list. 2284 */ 2285static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) 2286{ 2287 struct request *rq; 2288 2289 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); 2290 2291 if (!cfq_may_dispatch(cfqd, cfqq)) 2292 return false; 2293 2294 /* 2295 * follow expired path, else get first next available 2296 */ 2297 rq = cfq_check_fifo(cfqq); 2298 if (!rq) 2299 rq = cfqq->next_rq; 2300 2301 /* 2302 * insert request into driver dispatch list 2303 */ 2304 cfq_dispatch_insert(cfqd->queue, rq); 2305 2306 if (!cfqd->active_cic) { 2307 struct cfq_io_context *cic = RQ_CIC(rq); 2308 2309 atomic_long_inc(&cic->ioc->refcount); 2310 cfqd->active_cic = cic; 2311 } 2312 2313 return true; 2314} 2315 2316/* 2317 * Find the cfqq that we need to service and move a request from that to the 2318 * dispatch list 2319 */ 2320static int cfq_dispatch_requests(struct request_queue *q, int force) 2321{ 2322 struct cfq_data *cfqd = q->elevator->elevator_data; 2323 struct cfq_queue *cfqq; 2324 2325 if (!cfqd->busy_queues) 2326 return 0; 2327 2328 if (unlikely(force)) 2329 return cfq_forced_dispatch(cfqd); 2330 2331 cfqq = cfq_select_queue(cfqd); 2332 if (!cfqq) 2333 return 0; 2334 2335 /* 2336 * Dispatch a request from this cfqq, if it is allowed 2337 */ 2338 if (!cfq_dispatch_request(cfqd, cfqq)) 2339 return 0; 2340 2341 cfqq->slice_dispatch++; 2342 cfq_clear_cfqq_must_dispatch(cfqq); 2343 2344 /* 2345 * expire an async queue immediately if it has used up its slice. idle 2346 * queue always expire after 1 dispatch round. 2347 */ 2348 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && 2349 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || 2350 cfq_class_idle(cfqq))) { 2351 cfqq->slice_end = jiffies + 1; 2352 cfq_slice_expired(cfqd, 0); 2353 } 2354 2355 cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); 2356 return 1; 2357} 2358 2359/* 2360 * task holds one reference to the queue, dropped when task exits. each rq 2361 * in-flight on this queue also holds a reference, dropped when rq is freed. 2362 * 2363 * Each cfq queue took a reference on the parent group. Drop it now. 2364 * queue lock must be held here. 2365 */ 2366static void cfq_put_queue(struct cfq_queue *cfqq) 2367{ 2368 struct cfq_data *cfqd = cfqq->cfqd; 2369 struct cfq_group *cfqg; 2370 2371 BUG_ON(atomic_read(&cfqq->ref) <= 0); 2372 2373 if (!atomic_dec_and_test(&cfqq->ref)) 2374 return; 2375 2376 cfq_log_cfqq(cfqd, cfqq, "put_queue"); 2377 BUG_ON(rb_first(&cfqq->sort_list)); 2378 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); 2379 cfqg = cfqq->cfqg; 2380 2381 if (unlikely(cfqd->active_queue == cfqq)) { 2382 __cfq_slice_expired(cfqd, cfqq, 0); 2383 cfq_schedule_dispatch(cfqd); 2384 } 2385 2386 BUG_ON(cfq_cfqq_on_rr(cfqq)); 2387 kmem_cache_free(cfq_pool, cfqq); 2388 cfq_put_cfqg(cfqg); 2389 if (cfqq->orig_cfqg) 2390 cfq_put_cfqg(cfqq->orig_cfqg); 2391} 2392 2393/* 2394 * Must always be called with the rcu_read_lock() held 2395 */ 2396static void 2397__call_for_each_cic(struct io_context *ioc, 2398 void (*func)(struct io_context *, struct cfq_io_context *)) 2399{ 2400 struct cfq_io_context *cic; 2401 struct hlist_node *n; 2402 2403 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list) 2404 func(ioc, cic); 2405} 2406 2407/* 2408 * Call func for each cic attached to this ioc. 2409 */ 2410static void 2411call_for_each_cic(struct io_context *ioc, 2412 void (*func)(struct io_context *, struct cfq_io_context *)) 2413{ 2414 rcu_read_lock(); 2415 __call_for_each_cic(ioc, func); 2416 rcu_read_unlock(); 2417} 2418 2419static void cfq_cic_free_rcu(struct rcu_head *head) 2420{ 2421 struct cfq_io_context *cic; 2422 2423 cic = container_of(head, struct cfq_io_context, rcu_head); 2424 2425 kmem_cache_free(cfq_ioc_pool, cic); 2426 elv_ioc_count_dec(cfq_ioc_count); 2427 2428 if (ioc_gone) { 2429 /* 2430 * CFQ scheduler is exiting, grab exit lock and check 2431 * the pending io context count. If it hits zero, 2432 * complete ioc_gone and set it back to NULL 2433 */ 2434 spin_lock(&ioc_gone_lock); 2435 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) { 2436 complete(ioc_gone); 2437 ioc_gone = NULL; 2438 } 2439 spin_unlock(&ioc_gone_lock); 2440 } 2441} 2442 2443static void cfq_cic_free(struct cfq_io_context *cic) 2444{ 2445 call_rcu(&cic->rcu_head, cfq_cic_free_rcu); 2446} 2447 2448static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) 2449{ 2450 unsigned long flags; 2451 2452 BUG_ON(!cic->dead_key); 2453 2454 spin_lock_irqsave(&ioc->lock, flags); 2455 radix_tree_delete(&ioc->radix_root, cic->dead_key); 2456 hlist_del_rcu(&cic->cic_list); 2457 spin_unlock_irqrestore(&ioc->lock, flags); 2458 2459 cfq_cic_free(cic); 2460} 2461 2462/* 2463 * Must be called with rcu_read_lock() held or preemption otherwise disabled. 2464 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(), 2465 * and ->trim() which is called with the task lock held 2466 */ 2467static void cfq_free_io_context(struct io_context *ioc) 2468{ 2469 /* 2470 * ioc->refcount is zero here, or we are called from elv_unregister(), 2471 * so no more cic's are allowed to be linked into this ioc. So it 2472 * should be ok to iterate over the known list, we will see all cic's 2473 * since no new ones are added. 2474 */ 2475 __call_for_each_cic(ioc, cic_free_func); 2476} 2477 2478static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) 2479{ 2480 struct cfq_queue *__cfqq, *next; 2481 2482 if (unlikely(cfqq == cfqd->active_queue)) { 2483 __cfq_slice_expired(cfqd, cfqq, 0); 2484 cfq_schedule_dispatch(cfqd); 2485 } 2486 2487 /* 2488 * If this queue was scheduled to merge with another queue, be 2489 * sure to drop the reference taken on that queue (and others in 2490 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. 2491 */ 2492 __cfqq = cfqq->new_cfqq; 2493 while (__cfqq) { 2494 if (__cfqq == cfqq) { 2495 WARN(1, "cfqq->new_cfqq loop detected\n"); 2496 break; 2497 } 2498 next = __cfqq->new_cfqq; 2499 cfq_put_queue(__cfqq); 2500 __cfqq = next; 2501 } 2502 2503 cfq_put_queue(cfqq); 2504} 2505 2506static void __cfq_exit_single_io_context(struct cfq_data *cfqd, 2507 struct cfq_io_context *cic) 2508{ 2509 struct io_context *ioc = cic->ioc; 2510 2511 list_del_init(&cic->queue_list); 2512 2513 /* 2514 * Make sure key == NULL is seen for dead queues 2515 */ 2516 smp_wmb(); 2517 cic->dead_key = (unsigned long) cic->key; 2518 cic->key = NULL; 2519 2520 if (ioc->ioc_data == cic) 2521 rcu_assign_pointer(ioc->ioc_data, NULL); 2522 2523 if (cic->cfqq[BLK_RW_ASYNC]) { 2524 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); 2525 cic->cfqq[BLK_RW_ASYNC] = NULL; 2526 } 2527 2528 if (cic->cfqq[BLK_RW_SYNC]) { 2529 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); 2530 cic->cfqq[BLK_RW_SYNC] = NULL; 2531 } 2532} 2533 2534static void cfq_exit_single_io_context(struct io_context *ioc, 2535 struct cfq_io_context *cic) 2536{ 2537 struct cfq_data *cfqd = cic->key; 2538 2539 if (cfqd) { 2540 struct request_queue *q = cfqd->queue; 2541 unsigned long flags; 2542 2543 spin_lock_irqsave(q->queue_lock, flags); 2544 2545 /* 2546 * Ensure we get a fresh copy of the ->key to prevent 2547 * race between exiting task and queue 2548 */ 2549 smp_read_barrier_depends(); 2550 if (cic->key) 2551 __cfq_exit_single_io_context(cfqd, cic); 2552 2553 spin_unlock_irqrestore(q->queue_lock, flags); 2554 } 2555} 2556 2557/* 2558 * The process that ioc belongs to has exited, we need to clean up 2559 * and put the internal structures we have that belongs to that process. 2560 */ 2561static void cfq_exit_io_context(struct io_context *ioc) 2562{ 2563 call_for_each_cic(ioc, cfq_exit_single_io_context); 2564} 2565 2566static struct cfq_io_context * 2567cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) 2568{ 2569 struct cfq_io_context *cic; 2570 2571 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, 2572 cfqd->queue->node); 2573 if (cic) { 2574 cic->last_end_request = jiffies; 2575 INIT_LIST_HEAD(&cic->queue_list); 2576 INIT_HLIST_NODE(&cic->cic_list); 2577 cic->dtor = cfq_free_io_context; 2578 cic->exit = cfq_exit_io_context; 2579 elv_ioc_count_inc(cfq_ioc_count); 2580 } 2581 2582 return cic; 2583} 2584 2585static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) 2586{ 2587 struct task_struct *tsk = current; 2588 int ioprio_class; 2589 2590 if (!cfq_cfqq_prio_changed(cfqq)) 2591 return; 2592 2593 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); 2594 switch (ioprio_class) { 2595 default: 2596 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); 2597 case IOPRIO_CLASS_NONE: 2598 /* 2599 * no prio set, inherit CPU scheduling settings 2600 */ 2601 cfqq->ioprio = task_nice_ioprio(tsk); 2602 cfqq->ioprio_class = task_nice_ioclass(tsk); 2603 break; 2604 case IOPRIO_CLASS_RT: 2605 cfqq->ioprio = task_ioprio(ioc); 2606 cfqq->ioprio_class = IOPRIO_CLASS_RT; 2607 break; 2608 case IOPRIO_CLASS_BE: 2609 cfqq->ioprio = task_ioprio(ioc); 2610 cfqq->ioprio_class = IOPRIO_CLASS_BE; 2611 break; 2612 case IOPRIO_CLASS_IDLE: 2613 cfqq->ioprio_class = IOPRIO_CLASS_IDLE; 2614 cfqq->ioprio = 7; 2615 cfq_clear_cfqq_idle_window(cfqq); 2616 break; 2617 } 2618 2619 /* 2620 * keep track of original prio settings in case we have to temporarily 2621 * elevate the priority of this queue 2622 */ 2623 cfqq->org_ioprio = cfqq->ioprio; 2624 cfqq->org_ioprio_class = cfqq->ioprio_class; 2625 cfq_clear_cfqq_prio_changed(cfqq); 2626} 2627 2628static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic) 2629{ 2630 struct cfq_data *cfqd = cic->key; 2631 struct cfq_queue *cfqq; 2632 unsigned long flags; 2633 2634 if (unlikely(!cfqd)) 2635 return; 2636 2637 spin_lock_irqsave(cfqd->queue->queue_lock, flags); 2638 2639 cfqq = cic->cfqq[BLK_RW_ASYNC]; 2640 if (cfqq) { 2641 struct cfq_queue *new_cfqq; 2642 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc, 2643 GFP_ATOMIC); 2644 if (new_cfqq) { 2645 cic->cfqq[BLK_RW_ASYNC] = new_cfqq; 2646 cfq_put_queue(cfqq); 2647 } 2648 } 2649 2650 cfqq = cic->cfqq[BLK_RW_SYNC]; 2651 if (cfqq) 2652 cfq_mark_cfqq_prio_changed(cfqq); 2653 2654 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); 2655} 2656 2657static void cfq_ioc_set_ioprio(struct io_context *ioc) 2658{ 2659 call_for_each_cic(ioc, changed_ioprio); 2660 ioc->ioprio_changed = 0; 2661} 2662 2663static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, 2664 pid_t pid, bool is_sync) 2665{ 2666 RB_CLEAR_NODE(&cfqq->rb_node); 2667 RB_CLEAR_NODE(&cfqq->p_node); 2668 INIT_LIST_HEAD(&cfqq->fifo); 2669 2670 atomic_set(&cfqq->ref, 0); 2671 cfqq->cfqd = cfqd; 2672 2673 cfq_mark_cfqq_prio_changed(cfqq); 2674 2675 if (is_sync) { 2676 if (!cfq_class_idle(cfqq)) 2677 cfq_mark_cfqq_idle_window(cfqq); 2678 cfq_mark_cfqq_sync(cfqq); 2679 } 2680 cfqq->pid = pid; 2681} 2682 2683#ifdef CONFIG_CFQ_GROUP_IOSCHED 2684static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic) 2685{ 2686 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1); 2687 struct cfq_data *cfqd = cic->key; 2688 unsigned long flags; 2689 struct request_queue *q; 2690 2691 if (unlikely(!cfqd)) 2692 return; 2693 2694 q = cfqd->queue; 2695 2696 spin_lock_irqsave(q->queue_lock, flags); 2697 2698 if (sync_cfqq) { 2699 /* 2700 * Drop reference to sync queue. A new sync queue will be 2701 * assigned in new group upon arrival of a fresh request. 2702 */ 2703 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup"); 2704 cic_set_cfqq(cic, NULL, 1); 2705 cfq_put_queue(sync_cfqq); 2706 } 2707 2708 spin_unlock_irqrestore(q->queue_lock, flags); 2709} 2710 2711static void cfq_ioc_set_cgroup(struct io_context *ioc) 2712{ 2713 call_for_each_cic(ioc, changed_cgroup); 2714 ioc->cgroup_changed = 0; 2715} 2716#endif /* CONFIG_CFQ_GROUP_IOSCHED */ 2717 2718static struct cfq_queue * 2719cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, 2720 struct io_context *ioc, gfp_t gfp_mask) 2721{ 2722 struct cfq_queue *cfqq, *new_cfqq = NULL; 2723 struct cfq_io_context *cic; 2724 struct cfq_group *cfqg; 2725 2726retry: 2727 cfqg = cfq_get_cfqg(cfqd, 1); 2728 cic = cfq_cic_lookup(cfqd, ioc); 2729 /* cic always exists here */ 2730 cfqq = cic_to_cfqq(cic, is_sync); 2731 2732 /* 2733 * Always try a new alloc if we fell back to the OOM cfqq 2734 * originally, since it should just be a temporary situation. 2735 */ 2736 if (!cfqq || cfqq == &cfqd->oom_cfqq) { 2737 cfqq = NULL; 2738 if (new_cfqq) { 2739 cfqq = new_cfqq; 2740 new_cfqq = NULL; 2741 } else if (gfp_mask & __GFP_WAIT) { 2742 spin_unlock_irq(cfqd->queue->queue_lock); 2743 new_cfqq = kmem_cache_alloc_node(cfq_pool, 2744 gfp_mask | __GFP_ZERO, 2745 cfqd->queue->node); 2746 spin_lock_irq(cfqd->queue->queue_lock); 2747 if (new_cfqq) 2748 goto retry; 2749 } else { 2750 cfqq = kmem_cache_alloc_node(cfq_pool, 2751 gfp_mask | __GFP_ZERO, 2752 cfqd->queue->node); 2753 } 2754 2755 if (cfqq) { 2756 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); 2757 cfq_init_prio_data(cfqq, ioc); 2758 cfq_link_cfqq_cfqg(cfqq, cfqg); 2759 cfq_log_cfqq(cfqd, cfqq, "alloced"); 2760 } else 2761 cfqq = &cfqd->oom_cfqq; 2762 } 2763 2764 if (new_cfqq) 2765 kmem_cache_free(cfq_pool, new_cfqq); 2766 2767 return cfqq; 2768} 2769 2770static struct cfq_queue ** 2771cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) 2772{ 2773 switch (ioprio_class) { 2774 case IOPRIO_CLASS_RT: 2775 return &cfqd->async_cfqq[0][ioprio]; 2776 case IOPRIO_CLASS_BE: 2777 return &cfqd->async_cfqq[1][ioprio]; 2778 case IOPRIO_CLASS_IDLE: 2779 return &cfqd->async_idle_cfqq; 2780 default: 2781 BUG(); 2782 } 2783} 2784 2785static struct cfq_queue * 2786cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc, 2787 gfp_t gfp_mask) 2788{ 2789 const int ioprio = task_ioprio(ioc); 2790 const int ioprio_class = task_ioprio_class(ioc); 2791 struct cfq_queue **async_cfqq = NULL; 2792 struct cfq_queue *cfqq = NULL; 2793 2794 if (!is_sync) { 2795 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); 2796 cfqq = *async_cfqq; 2797 } 2798 2799 if (!cfqq) 2800 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); 2801 2802 /* 2803 * pin the queue now that it's allocated, scheduler exit will prune it 2804 */ 2805 if (!is_sync && !(*async_cfqq)) { 2806 atomic_inc(&cfqq->ref); 2807 *async_cfqq = cfqq; 2808 } 2809 2810 atomic_inc(&cfqq->ref); 2811 return cfqq; 2812} 2813 2814/* 2815 * We drop cfq io contexts lazily, so we may find a dead one. 2816 */ 2817static void 2818cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, 2819 struct cfq_io_context *cic) 2820{ 2821 unsigned long flags; 2822 2823 WARN_ON(!list_empty(&cic->queue_list)); 2824 2825 spin_lock_irqsave(&ioc->lock, flags); 2826 2827 BUG_ON(ioc->ioc_data == cic); 2828 2829 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd); 2830 hlist_del_rcu(&cic->cic_list); 2831 spin_unlock_irqrestore(&ioc->lock, flags); 2832 2833 cfq_cic_free(cic); 2834} 2835 2836static struct cfq_io_context * 2837cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) 2838{ 2839 struct cfq_io_context *cic; 2840 unsigned long flags; 2841 void *k; 2842 2843 if (unlikely(!ioc)) 2844 return NULL; 2845 2846 rcu_read_lock(); 2847 2848 /* 2849 * we maintain a last-hit cache, to avoid browsing over the tree 2850 */ 2851 cic = rcu_dereference(ioc->ioc_data); 2852 if (cic && cic->key == cfqd) { 2853 rcu_read_unlock(); 2854 return cic; 2855 } 2856 2857 do { 2858 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd); 2859 rcu_read_unlock(); 2860 if (!cic) 2861 break; 2862 /* ->key must be copied to avoid race with cfq_exit_queue() */ 2863 k = cic->key; 2864 if (unlikely(!k)) { 2865 cfq_drop_dead_cic(cfqd, ioc, cic); 2866 rcu_read_lock(); 2867 continue; 2868 } 2869 2870 spin_lock_irqsave(&ioc->lock, flags); 2871 rcu_assign_pointer(ioc->ioc_data, cic); 2872 spin_unlock_irqrestore(&ioc->lock, flags); 2873 break; 2874 } while (1); 2875 2876 return cic; 2877} 2878 2879/* 2880 * Add cic into ioc, using cfqd as the search key. This enables us to lookup 2881 * the process specific cfq io context when entered from the block layer. 2882 * Also adds the cic to a per-cfqd list, used when this queue is removed. 2883 */ 2884static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, 2885 struct cfq_io_context *cic, gfp_t gfp_mask) 2886{ 2887 unsigned long flags; 2888 int ret; 2889 2890 ret = radix_tree_preload(gfp_mask); 2891 if (!ret) { 2892 cic->ioc = ioc; 2893 cic->key = cfqd; 2894 2895 spin_lock_irqsave(&ioc->lock, flags); 2896 ret = radix_tree_insert(&ioc->radix_root, 2897 (unsigned long) cfqd, cic); 2898 if (!ret) 2899 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list); 2900 spin_unlock_irqrestore(&ioc->lock, flags); 2901 2902 radix_tree_preload_end(); 2903 2904 if (!ret) { 2905 spin_lock_irqsave(cfqd->queue->queue_lock, flags); 2906 list_add(&cic->queue_list, &cfqd->cic_list); 2907 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); 2908 } 2909 } 2910 2911 if (ret) 2912 printk(KERN_ERR "cfq: cic link failed!\n"); 2913 2914 return ret; 2915} 2916 2917/* 2918 * Setup general io context and cfq io context. There can be several cfq 2919 * io contexts per general io context, if this process is doing io to more 2920 * than one device managed by cfq. 2921 */ 2922static struct cfq_io_context * 2923cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) 2924{ 2925 struct io_context *ioc = NULL; 2926 struct cfq_io_context *cic; 2927 2928 might_sleep_if(gfp_mask & __GFP_WAIT); 2929 2930 ioc = get_io_context(gfp_mask, cfqd->queue->node); 2931 if (!ioc) 2932 return NULL; 2933 2934 cic = cfq_cic_lookup(cfqd, ioc); 2935 if (cic) 2936 goto out; 2937 2938 cic = cfq_alloc_io_context(cfqd, gfp_mask); 2939 if (cic == NULL) 2940 goto err; 2941 2942 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) 2943 goto err_free; 2944 2945out: 2946 smp_read_barrier_depends(); 2947 if (unlikely(ioc->ioprio_changed)) 2948 cfq_ioc_set_ioprio(ioc); 2949 2950#ifdef CONFIG_CFQ_GROUP_IOSCHED 2951 if (unlikely(ioc->cgroup_changed)) 2952 cfq_ioc_set_cgroup(ioc); 2953#endif 2954 return cic; 2955err_free: 2956 cfq_cic_free(cic); 2957err: 2958 put_io_context(ioc); 2959 return NULL; 2960} 2961 2962static void 2963cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) 2964{ 2965 unsigned long elapsed = jiffies - cic->last_end_request; 2966 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); 2967 2968 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; 2969 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; 2970 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; 2971} 2972 2973static void 2974cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, 2975 struct request *rq) 2976{ 2977 sector_t sdist; 2978 u64 total; 2979 2980 if (!cfqq->last_request_pos) 2981 sdist = 0; 2982 else if (cfqq->last_request_pos < blk_rq_pos(rq)) 2983 sdist = blk_rq_pos(rq) - cfqq->last_request_pos; 2984 else 2985 sdist = cfqq->last_request_pos - blk_rq_pos(rq); 2986 2987 /* 2988 * Don't allow the seek distance to get too large from the 2989 * odd fragment, pagein, etc 2990 */ 2991 if (cfqq->seek_samples <= 60) /* second&third seek */ 2992 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024); 2993 else 2994 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64); 2995 2996 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8; 2997 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8; 2998 total = cfqq->seek_total + (cfqq->seek_samples/2); 2999 do_div(total, cfqq->seek_samples); 3000 cfqq->seek_mean = (sector_t)total; 3001 3002 /* 3003 * If this cfqq is shared between multiple processes, check to 3004 * make sure that those processes are still issuing I/Os within 3005 * the mean seek distance. If not, it may be time to break the 3006 * queues apart again. 3007 */ 3008 if (cfq_cfqq_coop(cfqq)) { 3009 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start) 3010 cfqq->seeky_start = jiffies; 3011 else if (!CFQQ_SEEKY(cfqq)) 3012 cfqq->seeky_start = 0; 3013 } 3014} 3015 3016/* 3017 * Disable idle window if the process thinks too long or seeks so much that 3018 * it doesn't matter 3019 */ 3020static void 3021cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, 3022 struct cfq_io_context *cic) 3023{ 3024 int old_idle, enable_idle; 3025 3026 /* 3027 * Don't idle for async or idle io prio class 3028 */ 3029 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) 3030 return; 3031 3032 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); 3033 3034 if (cfqq->queued[0] + cfqq->queued[1] >= 4) 3035 cfq_mark_cfqq_deep(cfqq); 3036 3037 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || 3038 (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples) 3039 && CFQQ_SEEKY(cfqq))) 3040 enable_idle = 0; 3041 else if (sample_valid(cic->ttime_samples)) { 3042 if (cic->ttime_mean > cfqd->cfq_slice_idle) 3043 enable_idle = 0; 3044 else 3045 enable_idle = 1; 3046 } 3047 3048 if (old_idle != enable_idle) { 3049 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); 3050 if (enable_idle) 3051 cfq_mark_cfqq_idle_window(cfqq); 3052 else 3053 cfq_clear_cfqq_idle_window(cfqq); 3054 } 3055} 3056 3057/* 3058 * Check if new_cfqq should preempt the currently active queue. Return 0 for 3059 * no or if we aren't sure, a 1 will cause a preempt. 3060 */ 3061static bool 3062cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, 3063 struct request *rq) 3064{ 3065 struct cfq_queue *cfqq; 3066 3067 cfqq = cfqd->active_queue; 3068 if (!cfqq) 3069 return false; 3070 3071 if (cfq_class_idle(new_cfqq)) 3072 return false; 3073 3074 if (cfq_class_idle(cfqq)) 3075 return true; 3076 3077 /* 3078 * if the new request is sync, but the currently running queue is 3079 * not, let the sync request have priority. 3080 */ 3081 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) 3082 return true; 3083 3084 if (new_cfqq->cfqg != cfqq->cfqg) 3085 return false; 3086 3087 if (cfq_slice_used(cfqq)) 3088 return true; 3089 3090 /* Allow preemption only if we are idling on sync-noidle tree */ 3091 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD && 3092 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD && 3093 new_cfqq->service_tree->count == 2 && 3094 RB_EMPTY_ROOT(&cfqq->sort_list)) 3095 return true; 3096 3097 /* 3098 * So both queues are sync. Let the new request get disk time if 3099 * it's a metadata request and the current queue is doing regular IO. 3100 */ 3101 if (rq_is_meta(rq) && !cfqq->meta_pending) 3102 return true; 3103 3104 /* 3105 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. 3106 */ 3107 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) 3108 return true; 3109 3110 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) 3111 return false; 3112 3113 /* 3114 * if this request is as-good as one we would expect from the 3115 * current cfqq, let it preempt 3116 */ 3117 if (cfq_rq_close(cfqd, cfqq, rq)) 3118 return true; 3119 3120 return false; 3121} 3122 3123/* 3124 * cfqq preempts the active queue. if we allowed preempt with no slice left, 3125 * let it have half of its nominal slice. 3126 */ 3127static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) 3128{ 3129 cfq_log_cfqq(cfqd, cfqq, "preempt"); 3130 cfq_slice_expired(cfqd, 1); 3131 3132 /* 3133 * Put the new queue at the front of the of the current list, 3134 * so we know that it will be selected next. 3135 */ 3136 BUG_ON(!cfq_cfqq_on_rr(cfqq)); 3137 3138 cfq_service_tree_add(cfqd, cfqq, 1); 3139 3140 cfqq->slice_end = 0; 3141 cfq_mark_cfqq_slice_new(cfqq); 3142} 3143 3144/* 3145 * Called when a new fs request (rq) is added (to cfqq). Check if there's 3146 * something we should do about it 3147 */ 3148static void 3149cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, 3150 struct request *rq) 3151{ 3152 struct cfq_io_context *cic = RQ_CIC(rq); 3153 3154 cfqd->rq_queued++; 3155 if (rq_is_meta(rq)) 3156 cfqq->meta_pending++; 3157 3158 cfq_update_io_thinktime(cfqd, cic); 3159 cfq_update_io_seektime(cfqd, cfqq, rq); 3160 cfq_update_idle_window(cfqd, cfqq, cic); 3161 3162 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); 3163 3164 if (cfqq == cfqd->active_queue) { 3165 if (cfq_cfqq_wait_busy(cfqq)) { 3166 cfq_clear_cfqq_wait_busy(cfqq); 3167 cfq_mark_cfqq_wait_busy_done(cfqq); 3168 } 3169 /* 3170 * Remember that we saw a request from this process, but 3171 * don't start queuing just yet. Otherwise we risk seeing lots 3172 * of tiny requests, because we disrupt the normal plugging 3173 * and merging. If the request is already larger than a single 3174 * page, let it rip immediately. For that case we assume that 3175 * merging is already done. Ditto for a busy system that 3176 * has other work pending, don't risk delaying until the 3177 * idle timer unplug to continue working. 3178 */ 3179 if (cfq_cfqq_wait_request(cfqq)) { 3180 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || 3181 cfqd->busy_queues > 1) { 3182 del_timer(&cfqd->idle_slice_timer); 3183 __blk_run_queue(cfqd->queue); 3184 } else 3185 cfq_mark_cfqq_must_dispatch(cfqq); 3186 } 3187 } else if (cfq_should_preempt(cfqd, cfqq, rq)) { 3188 /* 3189 * not the active queue - expire current slice if it is 3190 * idle and has expired it's mean thinktime or this new queue 3191 * has some old slice time left and is of higher priority or 3192 * this new queue is RT and the current one is BE 3193 */ 3194 cfq_preempt_queue(cfqd, cfqq); 3195 __blk_run_queue(cfqd->queue); 3196 } 3197} 3198 3199static void cfq_insert_request(struct request_queue *q, struct request *rq) 3200{ 3201 struct cfq_data *cfqd = q->elevator->elevator_data; 3202 struct cfq_queue *cfqq = RQ_CFQQ(rq); 3203 3204 cfq_log_cfqq(cfqd, cfqq, "insert_request"); 3205 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); 3206 3207 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); 3208 list_add_tail(&rq->queuelist, &cfqq->fifo); 3209 cfq_add_rq_rb(rq); 3210 3211 cfq_rq_enqueued(cfqd, cfqq, rq); 3212} 3213 3214/* 3215 * Update hw_tag based on peak queue depth over 50 samples under 3216 * sufficient load. 3217 */ 3218static void cfq_update_hw_tag(struct cfq_data *cfqd) 3219{ 3220 struct cfq_queue *cfqq = cfqd->active_queue; 3221 3222 if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth) 3223 cfqd->hw_tag_est_depth = rq_in_driver(cfqd); 3224 3225 if (cfqd->hw_tag == 1) 3226 return; 3227 3228 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && 3229 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN) 3230 return; 3231 3232 /* 3233 * If active queue hasn't enough requests and can idle, cfq might not 3234 * dispatch sufficient requests to hardware. Don't zero hw_tag in this 3235 * case 3236 */ 3237 if (cfqq && cfq_cfqq_idle_window(cfqq) && 3238 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < 3239 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN) 3240 return; 3241 3242 if (cfqd->hw_tag_samples++ < 50) 3243 return; 3244 3245 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN) 3246 cfqd->hw_tag = 1; 3247 else 3248 cfqd->hw_tag = 0; 3249} 3250 3251static void cfq_completed_request(struct request_queue *q, struct request *rq) 3252{ 3253 struct cfq_queue *cfqq = RQ_CFQQ(rq); 3254 struct cfq_data *cfqd = cfqq->cfqd; 3255 const int sync = rq_is_sync(rq); 3256 unsigned long now; 3257 3258 now = jiffies; 3259 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq)); 3260 3261 cfq_update_hw_tag(cfqd); 3262 3263 WARN_ON(!cfqd->rq_in_driver[sync]); 3264 WARN_ON(!cfqq->dispatched); 3265 cfqd->rq_in_driver[sync]--; 3266 cfqq->dispatched--; 3267 3268 if (cfq_cfqq_sync(cfqq)) 3269 cfqd->sync_flight--; 3270 3271 if (sync) { 3272 RQ_CIC(rq)->last_end_request = now; 3273 cfqd->last_end_sync_rq = now; 3274 } 3275 3276 /* 3277 * If this is the active queue, check if it needs to be expired, 3278 * or if we want to idle in case it has no pending requests. 3279 */ 3280 if (cfqd->active_queue == cfqq) { 3281 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); 3282 3283 if (cfq_cfqq_slice_new(cfqq)) { 3284 cfq_set_prio_slice(cfqd, cfqq); 3285 cfq_clear_cfqq_slice_new(cfqq); 3286 } 3287 3288 /* 3289 * If this queue consumed its slice and this is last queue 3290 * in the group, wait for next request before we expire 3291 * the queue 3292 */ 3293 if (cfq_slice_used(cfqq) && cfqq->cfqg->nr_cfqq == 1) { 3294 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle; 3295 cfq_mark_cfqq_wait_busy(cfqq); 3296 } 3297 3298 /* 3299 * Idling is not enabled on: 3300 * - expired queues 3301 * - idle-priority queues 3302 * - async queues 3303 * - queues with still some requests queued 3304 * - when there is a close cooperator 3305 */ 3306 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) 3307 cfq_slice_expired(cfqd, 1); 3308 else if (sync && cfqq_empty && 3309 !cfq_close_cooperator(cfqd, cfqq)) { 3310 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq); 3311 /* 3312 * Idling is enabled for SYNC_WORKLOAD. 3313 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree 3314 * only if we processed at least one !rq_noidle request 3315 */ 3316 if (cfqd->serving_type == SYNC_WORKLOAD 3317 || cfqd->noidle_tree_requires_idle) 3318 cfq_arm_slice_timer(cfqd); 3319 } 3320 } 3321 3322 if (!rq_in_driver(cfqd)) 3323 cfq_schedule_dispatch(cfqd); 3324} 3325 3326/* 3327 * we temporarily boost lower priority queues if they are holding fs exclusive 3328 * resources. they are boosted to normal prio (CLASS_BE/4) 3329 */ 3330static void cfq_prio_boost(struct cfq_queue *cfqq) 3331{ 3332 if (has_fs_excl()) { 3333 /* 3334 * boost idle prio on transactions that would lock out other 3335 * users of the filesystem 3336 */ 3337 if (cfq_class_idle(cfqq)) 3338 cfqq->ioprio_class = IOPRIO_CLASS_BE; 3339 if (cfqq->ioprio > IOPRIO_NORM) 3340 cfqq->ioprio = IOPRIO_NORM; 3341 } else { 3342 /* 3343 * unboost the queue (if needed) 3344 */ 3345 cfqq->ioprio_class = cfqq->org_ioprio_class; 3346 cfqq->ioprio = cfqq->org_ioprio; 3347 } 3348} 3349 3350static inline int __cfq_may_queue(struct cfq_queue *cfqq) 3351{ 3352 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { 3353 cfq_mark_cfqq_must_alloc_slice(cfqq); 3354 return ELV_MQUEUE_MUST; 3355 } 3356 3357 return ELV_MQUEUE_MAY; 3358} 3359 3360static int cfq_may_queue(struct request_queue *q, int rw) 3361{ 3362 struct cfq_data *cfqd = q->elevator->elevator_data; 3363 struct task_struct *tsk = current; 3364 struct cfq_io_context *cic; 3365 struct cfq_queue *cfqq; 3366 3367 /* 3368 * don't force setup of a queue from here, as a call to may_queue 3369 * does not necessarily imply that a request actually will be queued. 3370 * so just lookup a possibly existing queue, or return 'may queue' 3371 * if that fails 3372 */ 3373 cic = cfq_cic_lookup(cfqd, tsk->io_context); 3374 if (!cic) 3375 return ELV_MQUEUE_MAY; 3376 3377 cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); 3378 if (cfqq) { 3379 cfq_init_prio_data(cfqq, cic->ioc); 3380 cfq_prio_boost(cfqq); 3381 3382 return __cfq_may_queue(cfqq); 3383 } 3384 3385 return ELV_MQUEUE_MAY; 3386} 3387 3388/* 3389 * queue lock held here 3390 */ 3391static void cfq_put_request(struct request *rq) 3392{ 3393 struct cfq_queue *cfqq = RQ_CFQQ(rq); 3394 3395 if (cfqq) { 3396 const int rw = rq_data_dir(rq); 3397 3398 BUG_ON(!cfqq->allocated[rw]); 3399 cfqq->allocated[rw]--; 3400 3401 put_io_context(RQ_CIC(rq)->ioc); 3402 3403 rq->elevator_private = NULL; 3404 rq->elevator_private2 = NULL; 3405 3406 cfq_put_queue(cfqq); 3407 } 3408} 3409 3410static struct cfq_queue * 3411cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic, 3412 struct cfq_queue *cfqq) 3413{ 3414 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); 3415 cic_set_cfqq(cic, cfqq->new_cfqq, 1); 3416 cfq_mark_cfqq_coop(cfqq->new_cfqq); 3417 cfq_put_queue(cfqq); 3418 return cic_to_cfqq(cic, 1); 3419} 3420 3421static int should_split_cfqq(struct cfq_queue *cfqq) 3422{ 3423 if (cfqq->seeky_start && 3424 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT)) 3425 return 1; 3426 return 0; 3427} 3428 3429/* 3430 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this 3431 * was the last process referring to said cfqq. 3432 */ 3433static struct cfq_queue * 3434split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq) 3435{ 3436 if (cfqq_process_refs(cfqq) == 1) { 3437 cfqq->seeky_start = 0; 3438 cfqq->pid = current->pid; 3439 cfq_clear_cfqq_coop(cfqq); 3440 return cfqq; 3441 } 3442 3443 cic_set_cfqq(cic, NULL, 1); 3444 cfq_put_queue(cfqq); 3445 return NULL; 3446} 3447/* 3448 * Allocate cfq data structures associated with this request. 3449 */ 3450static int 3451cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) 3452{ 3453 struct cfq_data *cfqd = q->elevator->elevator_data; 3454 struct cfq_io_context *cic; 3455 const int rw = rq_data_dir(rq); 3456 const bool is_sync = rq_is_sync(rq); 3457 struct cfq_queue *cfqq; 3458 unsigned long flags; 3459 3460 might_sleep_if(gfp_mask & __GFP_WAIT); 3461 3462 cic = cfq_get_io_context(cfqd, gfp_mask); 3463 3464 spin_lock_irqsave(q->queue_lock, flags); 3465 3466 if (!cic) 3467 goto queue_fail; 3468 3469new_queue: 3470 cfqq = cic_to_cfqq(cic, is_sync); 3471 if (!cfqq || cfqq == &cfqd->oom_cfqq) { 3472 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); 3473 cic_set_cfqq(cic, cfqq, is_sync); 3474 } else { 3475 /* 3476 * If the queue was seeky for too long, break it apart. 3477 */ 3478 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) { 3479 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); 3480 cfqq = split_cfqq(cic, cfqq); 3481 if (!cfqq) 3482 goto new_queue; 3483 } 3484 3485 /* 3486 * Check to see if this queue is scheduled to merge with 3487 * another, closely cooperating queue. The merging of 3488 * queues happens here as it must be done in process context. 3489 * The reference on new_cfqq was taken in merge_cfqqs. 3490 */ 3491 if (cfqq->new_cfqq) 3492 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); 3493 } 3494 3495 cfqq->allocated[rw]++; 3496 atomic_inc(&cfqq->ref); 3497 3498 spin_unlock_irqrestore(q->queue_lock, flags); 3499 3500 rq->elevator_private = cic; 3501 rq->elevator_private2 = cfqq; 3502 return 0; 3503 3504queue_fail: 3505 if (cic) 3506 put_io_context(cic->ioc); 3507 3508 cfq_schedule_dispatch(cfqd); 3509 spin_unlock_irqrestore(q->queue_lock, flags); 3510 cfq_log(cfqd, "set_request fail"); 3511 return 1; 3512} 3513 3514static void cfq_kick_queue(struct work_struct *work) 3515{ 3516 struct cfq_data *cfqd = 3517 container_of(work, struct cfq_data, unplug_work); 3518 struct request_queue *q = cfqd->queue; 3519 3520 spin_lock_irq(q->queue_lock); 3521 __blk_run_queue(cfqd->queue); 3522 spin_unlock_irq(q->queue_lock); 3523} 3524 3525/* 3526 * Timer running if the active_queue is currently idling inside its time slice 3527 */ 3528static void cfq_idle_slice_timer(unsigned long data) 3529{ 3530 struct cfq_data *cfqd = (struct cfq_data *) data; 3531 struct cfq_queue *cfqq; 3532 unsigned long flags; 3533 int timed_out = 1; 3534 3535 cfq_log(cfqd, "idle timer fired"); 3536 3537 spin_lock_irqsave(cfqd->queue->queue_lock, flags); 3538 3539 cfqq = cfqd->active_queue; 3540 if (cfqq) { 3541 timed_out = 0; 3542 3543 /* 3544 * We saw a request before the queue expired, let it through 3545 */ 3546 if (cfq_cfqq_must_dispatch(cfqq)) 3547 goto out_kick; 3548 3549 /* 3550 * expired 3551 */ 3552 if (cfq_slice_used(cfqq)) 3553 goto expire; 3554 3555 /* 3556 * only expire and reinvoke request handler, if there are 3557 * other queues with pending requests 3558 */ 3559 if (!cfqd->busy_queues) 3560 goto out_cont; 3561 3562 /* 3563 * not expired and it has a request pending, let it dispatch 3564 */ 3565 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) 3566 goto out_kick; 3567 3568 /* 3569 * Queue depth flag is reset only when the idle didn't succeed 3570 */ 3571 cfq_clear_cfqq_deep(cfqq); 3572 } 3573expire: 3574 cfq_slice_expired(cfqd, timed_out); 3575out_kick: 3576 cfq_schedule_dispatch(cfqd); 3577out_cont: 3578 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); 3579} 3580 3581static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) 3582{ 3583 del_timer_sync(&cfqd->idle_slice_timer); 3584 cancel_work_sync(&cfqd->unplug_work); 3585} 3586 3587static void cfq_put_async_queues(struct cfq_data *cfqd) 3588{ 3589 int i; 3590 3591 for (i = 0; i < IOPRIO_BE_NR; i++) { 3592 if (cfqd->async_cfqq[0][i]) 3593 cfq_put_queue(cfqd->async_cfqq[0][i]); 3594 if (cfqd->async_cfqq[1][i]) 3595 cfq_put_queue(cfqd->async_cfqq[1][i]); 3596 } 3597 3598 if (cfqd->async_idle_cfqq) 3599 cfq_put_queue(cfqd->async_idle_cfqq); 3600} 3601 3602static void cfq_exit_queue(struct elevator_queue *e) 3603{ 3604 struct cfq_data *cfqd = e->elevator_data; 3605 struct request_queue *q = cfqd->queue; 3606 3607 cfq_shutdown_timer_wq(cfqd); 3608 3609 spin_lock_irq(q->queue_lock); 3610 3611 if (cfqd->active_queue) 3612 __cfq_slice_expired(cfqd, cfqd->active_queue, 0); 3613 3614 while (!list_empty(&cfqd->cic_list)) { 3615 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, 3616 struct cfq_io_context, 3617 queue_list); 3618 3619 __cfq_exit_single_io_context(cfqd, cic); 3620 } 3621 3622 cfq_put_async_queues(cfqd); 3623 cfq_release_cfq_groups(cfqd); 3624 blkiocg_del_blkio_group(&cfqd->root_group.blkg); 3625 3626 spin_unlock_irq(q->queue_lock); 3627 3628 cfq_shutdown_timer_wq(cfqd); 3629 3630 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */ 3631 synchronize_rcu(); 3632 kfree(cfqd); 3633} 3634 3635static void *cfq_init_queue(struct request_queue *q) 3636{ 3637 struct cfq_data *cfqd; 3638 int i, j; 3639 struct cfq_group *cfqg; 3640 struct cfq_rb_root *st; 3641 3642 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); 3643 if (!cfqd) 3644 return NULL; 3645 3646 /* Init root service tree */ 3647 cfqd->grp_service_tree = CFQ_RB_ROOT; 3648 3649 /* Init root group */ 3650 cfqg = &cfqd->root_group; 3651 for_each_cfqg_st(cfqg, i, j, st) 3652 *st = CFQ_RB_ROOT; 3653 RB_CLEAR_NODE(&cfqg->rb_node); 3654 3655 /* Give preference to root group over other groups */ 3656 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT; 3657 3658#ifdef CONFIG_CFQ_GROUP_IOSCHED 3659 /* 3660 * Take a reference to root group which we never drop. This is just 3661 * to make sure that cfq_put_cfqg() does not try to kfree root group 3662 */ 3663 atomic_set(&cfqg->ref, 1); 3664 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd, 3665 0); 3666#endif 3667 /* 3668 * Not strictly needed (since RB_ROOT just clears the node and we 3669 * zeroed cfqd on alloc), but better be safe in case someone decides 3670 * to add magic to the rb code 3671 */ 3672 for (i = 0; i < CFQ_PRIO_LISTS; i++) 3673 cfqd->prio_trees[i] = RB_ROOT; 3674 3675 /* 3676 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. 3677 * Grab a permanent reference to it, so that the normal code flow 3678 * will not attempt to free it. 3679 */ 3680 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); 3681 atomic_inc(&cfqd->oom_cfqq.ref); 3682 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group); 3683 3684 INIT_LIST_HEAD(&cfqd->cic_list); 3685 3686 cfqd->queue = q; 3687 3688 init_timer(&cfqd->idle_slice_timer); 3689 cfqd->idle_slice_timer.function = cfq_idle_slice_timer; 3690 cfqd->idle_slice_timer.data = (unsigned long) cfqd; 3691 3692 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); 3693 3694 cfqd->cfq_quantum = cfq_quantum; 3695 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; 3696 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; 3697 cfqd->cfq_back_max = cfq_back_max; 3698 cfqd->cfq_back_penalty = cfq_back_penalty; 3699 cfqd->cfq_slice[0] = cfq_slice_async; 3700 cfqd->cfq_slice[1] = cfq_slice_sync; 3701 cfqd->cfq_slice_async_rq = cfq_slice_async_rq; 3702 cfqd->cfq_slice_idle = cfq_slice_idle; 3703 cfqd->cfq_latency = 1; 3704 cfqd->cfq_group_isolation = 0; 3705 cfqd->hw_tag = -1; 3706 cfqd->last_end_sync_rq = jiffies; 3707 return cfqd; 3708} 3709 3710static void cfq_slab_kill(void) 3711{ 3712 /* 3713 * Caller already ensured that pending RCU callbacks are completed, 3714 * so we should have no busy allocations at this point. 3715 */ 3716 if (cfq_pool) 3717 kmem_cache_destroy(cfq_pool); 3718 if (cfq_ioc_pool) 3719 kmem_cache_destroy(cfq_ioc_pool); 3720} 3721 3722static int __init cfq_slab_setup(void) 3723{ 3724 cfq_pool = KMEM_CACHE(cfq_queue, 0); 3725 if (!cfq_pool) 3726 goto fail; 3727 3728 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0); 3729 if (!cfq_ioc_pool) 3730 goto fail; 3731 3732 return 0; 3733fail: 3734 cfq_slab_kill(); 3735 return -ENOMEM; 3736} 3737 3738/* 3739 * sysfs parts below --> 3740 */ 3741static ssize_t 3742cfq_var_show(unsigned int var, char *page) 3743{ 3744 return sprintf(page, "%d\n", var); 3745} 3746 3747static ssize_t 3748cfq_var_store(unsigned int *var, const char *page, size_t count) 3749{ 3750 char *p = (char *) page; 3751 3752 *var = simple_strtoul(p, &p, 10); 3753 return count; 3754} 3755 3756#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ 3757static ssize_t __FUNC(struct elevator_queue *e, char *page) \ 3758{ \ 3759 struct cfq_data *cfqd = e->elevator_data; \ 3760 unsigned int __data = __VAR; \ 3761 if (__CONV) \ 3762 __data = jiffies_to_msecs(__data); \ 3763 return cfq_var_show(__data, (page)); \ 3764} 3765SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); 3766SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); 3767SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); 3768SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); 3769SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); 3770SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); 3771SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); 3772SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); 3773SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); 3774SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); 3775SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0); 3776#undef SHOW_FUNCTION 3777 3778#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ 3779static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ 3780{ \ 3781 struct cfq_data *cfqd = e->elevator_data; \ 3782 unsigned int __data; \ 3783 int ret = cfq_var_store(&__data, (page), count); \ 3784 if (__data < (MIN)) \ 3785 __data = (MIN); \ 3786 else if (__data > (MAX)) \ 3787 __data = (MAX); \ 3788 if (__CONV) \ 3789 *(__PTR) = msecs_to_jiffies(__data); \ 3790 else \ 3791 *(__PTR) = __data; \ 3792 return ret; \ 3793} 3794STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); 3795STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, 3796 UINT_MAX, 1); 3797STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, 3798 UINT_MAX, 1); 3799STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); 3800STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, 3801 UINT_MAX, 0); 3802STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); 3803STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); 3804STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); 3805STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, 3806 UINT_MAX, 0); 3807STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); 3808STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0); 3809#undef STORE_FUNCTION 3810 3811#define CFQ_ATTR(name) \ 3812 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) 3813 3814static struct elv_fs_entry cfq_attrs[] = { 3815 CFQ_ATTR(quantum), 3816 CFQ_ATTR(fifo_expire_sync), 3817 CFQ_ATTR(fifo_expire_async), 3818 CFQ_ATTR(back_seek_max), 3819 CFQ_ATTR(back_seek_penalty), 3820 CFQ_ATTR(slice_sync), 3821 CFQ_ATTR(slice_async), 3822 CFQ_ATTR(slice_async_rq), 3823 CFQ_ATTR(slice_idle), 3824 CFQ_ATTR(low_latency), 3825 CFQ_ATTR(group_isolation), 3826 __ATTR_NULL 3827}; 3828 3829static struct elevator_type iosched_cfq = { 3830 .ops = { 3831 .elevator_merge_fn = cfq_merge, 3832 .elevator_merged_fn = cfq_merged_request, 3833 .elevator_merge_req_fn = cfq_merged_requests, 3834 .elevator_allow_merge_fn = cfq_allow_merge, 3835 .elevator_dispatch_fn = cfq_dispatch_requests, 3836 .elevator_add_req_fn = cfq_insert_request, 3837 .elevator_activate_req_fn = cfq_activate_request, 3838 .elevator_deactivate_req_fn = cfq_deactivate_request, 3839 .elevator_queue_empty_fn = cfq_queue_empty, 3840 .elevator_completed_req_fn = cfq_completed_request, 3841 .elevator_former_req_fn = elv_rb_former_request, 3842 .elevator_latter_req_fn = elv_rb_latter_request, 3843 .elevator_set_req_fn = cfq_set_request, 3844 .elevator_put_req_fn = cfq_put_request, 3845 .elevator_may_queue_fn = cfq_may_queue, 3846 .elevator_init_fn = cfq_init_queue, 3847 .elevator_exit_fn = cfq_exit_queue, 3848 .trim = cfq_free_io_context, 3849 }, 3850 .elevator_attrs = cfq_attrs, 3851 .elevator_name = "cfq", 3852 .elevator_owner = THIS_MODULE, 3853}; 3854 3855static int __init cfq_init(void) 3856{ 3857 /* 3858 * could be 0 on HZ < 1000 setups 3859 */ 3860 if (!cfq_slice_async) 3861 cfq_slice_async = 1; 3862 if (!cfq_slice_idle) 3863 cfq_slice_idle = 1; 3864 3865 if (cfq_slab_setup()) 3866 return -ENOMEM; 3867 3868 elv_register(&iosched_cfq); 3869 3870 return 0; 3871} 3872 3873static void __exit cfq_exit(void) 3874{ 3875 DECLARE_COMPLETION_ONSTACK(all_gone); 3876 elv_unregister(&iosched_cfq); 3877 ioc_gone = &all_gone; 3878 /* ioc_gone's update must be visible before reading ioc_count */ 3879 smp_wmb(); 3880 3881 /* 3882 * this also protects us from entering cfq_slab_kill() with 3883 * pending RCU callbacks 3884 */ 3885 if (elv_ioc_count_read(cfq_ioc_count)) 3886 wait_for_completion(&all_gone); 3887 cfq_slab_kill(); 3888} 3889 3890module_init(cfq_init); 3891module_exit(cfq_exit); 3892 3893MODULE_AUTHOR("Jens Axboe"); 3894MODULE_LICENSE("GPL"); 3895MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); 3896