cfq-iosched.c revision e2d74ac0664c89757bde8fb18c98cd7bf53da61c
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@suse.de> 8 */ 9#include <linux/config.h> 10#include <linux/module.h> 11#include <linux/blkdev.h> 12#include <linux/elevator.h> 13#include <linux/hash.h> 14#include <linux/rbtree.h> 15#include <linux/ioprio.h> 16 17/* 18 * tunables 19 */ 20static const int cfq_quantum = 4; /* max queue in one round of service */ 21static const int cfq_queued = 8; /* minimum rq allocate limit per-queue*/ 22static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; 23static const int cfq_back_max = 16 * 1024; /* maximum backwards seek, in KiB */ 24static const int cfq_back_penalty = 2; /* penalty of a backwards seek */ 25 26static const int cfq_slice_sync = HZ / 10; 27static int cfq_slice_async = HZ / 25; 28static const int cfq_slice_async_rq = 2; 29static int cfq_slice_idle = HZ / 100; 30 31#define CFQ_IDLE_GRACE (HZ / 10) 32#define CFQ_SLICE_SCALE (5) 33 34#define CFQ_KEY_ASYNC (0) 35#define CFQ_KEY_ANY (0xffff) 36 37/* 38 * disable queueing at the driver/hardware level 39 */ 40static const int cfq_max_depth = 2; 41 42static DEFINE_RWLOCK(cfq_exit_lock); 43 44/* 45 * for the hash of cfqq inside the cfqd 46 */ 47#define CFQ_QHASH_SHIFT 6 48#define CFQ_QHASH_ENTRIES (1 << CFQ_QHASH_SHIFT) 49#define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash) 50 51/* 52 * for the hash of crq inside the cfqq 53 */ 54#define CFQ_MHASH_SHIFT 6 55#define CFQ_MHASH_BLOCK(sec) ((sec) >> 3) 56#define CFQ_MHASH_ENTRIES (1 << CFQ_MHASH_SHIFT) 57#define CFQ_MHASH_FN(sec) hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT) 58#define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors) 59#define list_entry_hash(ptr) hlist_entry((ptr), struct cfq_rq, hash) 60 61#define list_entry_cfqq(ptr) list_entry((ptr), struct cfq_queue, cfq_list) 62#define list_entry_fifo(ptr) list_entry((ptr), struct request, queuelist) 63 64#define RQ_DATA(rq) (rq)->elevator_private 65 66/* 67 * rb-tree defines 68 */ 69#define RB_NONE (2) 70#define RB_EMPTY(node) ((node)->rb_node == NULL) 71#define RB_CLEAR_COLOR(node) (node)->rb_color = RB_NONE 72#define RB_CLEAR(node) do { \ 73 (node)->rb_parent = NULL; \ 74 RB_CLEAR_COLOR((node)); \ 75 (node)->rb_right = NULL; \ 76 (node)->rb_left = NULL; \ 77} while (0) 78#define RB_CLEAR_ROOT(root) ((root)->rb_node = NULL) 79#define rb_entry_crq(node) rb_entry((node), struct cfq_rq, rb_node) 80#define rq_rb_key(rq) (rq)->sector 81 82static kmem_cache_t *crq_pool; 83static kmem_cache_t *cfq_pool; 84static kmem_cache_t *cfq_ioc_pool; 85 86static atomic_t ioc_count = ATOMIC_INIT(0); 87static struct completion *ioc_gone; 88 89#define CFQ_PRIO_LISTS IOPRIO_BE_NR 90#define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) 91#define cfq_class_be(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_BE) 92#define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) 93 94#define ASYNC (0) 95#define SYNC (1) 96 97#define cfq_cfqq_dispatched(cfqq) \ 98 ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC]) 99 100#define cfq_cfqq_class_sync(cfqq) ((cfqq)->key != CFQ_KEY_ASYNC) 101 102#define cfq_cfqq_sync(cfqq) \ 103 (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC]) 104 105/* 106 * Per block device queue structure 107 */ 108struct cfq_data { 109 request_queue_t *queue; 110 111 /* 112 * rr list of queues with requests and the count of them 113 */ 114 struct list_head rr_list[CFQ_PRIO_LISTS]; 115 struct list_head busy_rr; 116 struct list_head cur_rr; 117 struct list_head idle_rr; 118 unsigned int busy_queues; 119 120 /* 121 * non-ordered list of empty cfqq's 122 */ 123 struct list_head empty_list; 124 125 /* 126 * cfqq lookup hash 127 */ 128 struct hlist_head *cfq_hash; 129 130 /* 131 * global crq hash for all queues 132 */ 133 struct hlist_head *crq_hash; 134 135 unsigned int max_queued; 136 137 mempool_t *crq_pool; 138 139 int rq_in_driver; 140 141 /* 142 * schedule slice state info 143 */ 144 /* 145 * idle window management 146 */ 147 struct timer_list idle_slice_timer; 148 struct work_struct unplug_work; 149 150 struct cfq_queue *active_queue; 151 struct cfq_io_context *active_cic; 152 int cur_prio, cur_end_prio; 153 unsigned int dispatch_slice; 154 155 struct timer_list idle_class_timer; 156 157 sector_t last_sector; 158 unsigned long last_end_request; 159 160 unsigned int rq_starved; 161 162 /* 163 * tunables, see top of file 164 */ 165 unsigned int cfq_quantum; 166 unsigned int cfq_queued; 167 unsigned int cfq_fifo_expire[2]; 168 unsigned int cfq_back_penalty; 169 unsigned int cfq_back_max; 170 unsigned int cfq_slice[2]; 171 unsigned int cfq_slice_async_rq; 172 unsigned int cfq_slice_idle; 173 unsigned int cfq_max_depth; 174 175 struct list_head cic_list; 176}; 177 178/* 179 * Per process-grouping structure 180 */ 181struct cfq_queue { 182 /* reference count */ 183 atomic_t ref; 184 /* parent cfq_data */ 185 struct cfq_data *cfqd; 186 /* cfqq lookup hash */ 187 struct hlist_node cfq_hash; 188 /* hash key */ 189 unsigned int key; 190 /* on either rr or empty list of cfqd */ 191 struct list_head cfq_list; 192 /* sorted list of pending requests */ 193 struct rb_root sort_list; 194 /* if fifo isn't expired, next request to serve */ 195 struct cfq_rq *next_crq; 196 /* requests queued in sort_list */ 197 int queued[2]; 198 /* currently allocated requests */ 199 int allocated[2]; 200 /* fifo list of requests in sort_list */ 201 struct list_head fifo; 202 203 unsigned long slice_start; 204 unsigned long slice_end; 205 unsigned long slice_left; 206 unsigned long service_last; 207 208 /* number of requests that are on the dispatch list */ 209 int on_dispatch[2]; 210 211 /* io prio of this group */ 212 unsigned short ioprio, org_ioprio; 213 unsigned short ioprio_class, org_ioprio_class; 214 215 /* various state flags, see below */ 216 unsigned int flags; 217}; 218 219struct cfq_rq { 220 struct rb_node rb_node; 221 sector_t rb_key; 222 struct request *request; 223 struct hlist_node hash; 224 225 struct cfq_queue *cfq_queue; 226 struct cfq_io_context *io_context; 227 228 unsigned int crq_flags; 229}; 230 231enum cfqq_state_flags { 232 CFQ_CFQQ_FLAG_on_rr = 0, 233 CFQ_CFQQ_FLAG_wait_request, 234 CFQ_CFQQ_FLAG_must_alloc, 235 CFQ_CFQQ_FLAG_must_alloc_slice, 236 CFQ_CFQQ_FLAG_must_dispatch, 237 CFQ_CFQQ_FLAG_fifo_expire, 238 CFQ_CFQQ_FLAG_idle_window, 239 CFQ_CFQQ_FLAG_prio_changed, 240}; 241 242#define CFQ_CFQQ_FNS(name) \ 243static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ 244{ \ 245 cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ 246} \ 247static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ 248{ \ 249 cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ 250} \ 251static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ 252{ \ 253 return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ 254} 255 256CFQ_CFQQ_FNS(on_rr); 257CFQ_CFQQ_FNS(wait_request); 258CFQ_CFQQ_FNS(must_alloc); 259CFQ_CFQQ_FNS(must_alloc_slice); 260CFQ_CFQQ_FNS(must_dispatch); 261CFQ_CFQQ_FNS(fifo_expire); 262CFQ_CFQQ_FNS(idle_window); 263CFQ_CFQQ_FNS(prio_changed); 264#undef CFQ_CFQQ_FNS 265 266enum cfq_rq_state_flags { 267 CFQ_CRQ_FLAG_is_sync = 0, 268}; 269 270#define CFQ_CRQ_FNS(name) \ 271static inline void cfq_mark_crq_##name(struct cfq_rq *crq) \ 272{ \ 273 crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name); \ 274} \ 275static inline void cfq_clear_crq_##name(struct cfq_rq *crq) \ 276{ \ 277 crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name); \ 278} \ 279static inline int cfq_crq_##name(const struct cfq_rq *crq) \ 280{ \ 281 return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0; \ 282} 283 284CFQ_CRQ_FNS(is_sync); 285#undef CFQ_CRQ_FNS 286 287static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short); 288static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *); 289static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask); 290 291#define process_sync(tsk) ((tsk)->flags & PF_SYNCWRITE) 292 293/* 294 * lots of deadline iosched dupes, can be abstracted later... 295 */ 296static inline void cfq_del_crq_hash(struct cfq_rq *crq) 297{ 298 hlist_del_init(&crq->hash); 299} 300 301static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq) 302{ 303 const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request)); 304 305 hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]); 306} 307 308static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset) 309{ 310 struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)]; 311 struct hlist_node *entry, *next; 312 313 hlist_for_each_safe(entry, next, hash_list) { 314 struct cfq_rq *crq = list_entry_hash(entry); 315 struct request *__rq = crq->request; 316 317 if (!rq_mergeable(__rq)) { 318 cfq_del_crq_hash(crq); 319 continue; 320 } 321 322 if (rq_hash_key(__rq) == offset) 323 return __rq; 324 } 325 326 return NULL; 327} 328 329/* 330 * scheduler run of queue, if there are requests pending and no one in the 331 * driver that will restart queueing 332 */ 333static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) 334{ 335 if (cfqd->busy_queues) 336 kblockd_schedule_work(&cfqd->unplug_work); 337} 338 339static int cfq_queue_empty(request_queue_t *q) 340{ 341 struct cfq_data *cfqd = q->elevator->elevator_data; 342 343 return !cfqd->busy_queues; 344} 345 346/* 347 * Lifted from AS - choose which of crq1 and crq2 that is best served now. 348 * We choose the request that is closest to the head right now. Distance 349 * behind the head are penalized and only allowed to a certain extent. 350 */ 351static struct cfq_rq * 352cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2) 353{ 354 sector_t last, s1, s2, d1 = 0, d2 = 0; 355 int r1_wrap = 0, r2_wrap = 0; /* requests are behind the disk head */ 356 unsigned long back_max; 357 358 if (crq1 == NULL || crq1 == crq2) 359 return crq2; 360 if (crq2 == NULL) 361 return crq1; 362 363 if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2)) 364 return crq1; 365 else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1)) 366 return crq2; 367 368 s1 = crq1->request->sector; 369 s2 = crq2->request->sector; 370 371 last = cfqd->last_sector; 372 373 /* 374 * by definition, 1KiB is 2 sectors 375 */ 376 back_max = cfqd->cfq_back_max * 2; 377 378 /* 379 * Strict one way elevator _except_ in the case where we allow 380 * short backward seeks which are biased as twice the cost of a 381 * similar forward seek. 382 */ 383 if (s1 >= last) 384 d1 = s1 - last; 385 else if (s1 + back_max >= last) 386 d1 = (last - s1) * cfqd->cfq_back_penalty; 387 else 388 r1_wrap = 1; 389 390 if (s2 >= last) 391 d2 = s2 - last; 392 else if (s2 + back_max >= last) 393 d2 = (last - s2) * cfqd->cfq_back_penalty; 394 else 395 r2_wrap = 1; 396 397 /* Found required data */ 398 if (!r1_wrap && r2_wrap) 399 return crq1; 400 else if (!r2_wrap && r1_wrap) 401 return crq2; 402 else if (r1_wrap && r2_wrap) { 403 /* both behind the head */ 404 if (s1 <= s2) 405 return crq1; 406 else 407 return crq2; 408 } 409 410 /* Both requests in front of the head */ 411 if (d1 < d2) 412 return crq1; 413 else if (d2 < d1) 414 return crq2; 415 else { 416 if (s1 >= s2) 417 return crq1; 418 else 419 return crq2; 420 } 421} 422 423/* 424 * would be nice to take fifo expire time into account as well 425 */ 426static struct cfq_rq * 427cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq, 428 struct cfq_rq *last) 429{ 430 struct cfq_rq *crq_next = NULL, *crq_prev = NULL; 431 struct rb_node *rbnext, *rbprev; 432 433 if (!(rbnext = rb_next(&last->rb_node))) { 434 rbnext = rb_first(&cfqq->sort_list); 435 if (rbnext == &last->rb_node) 436 rbnext = NULL; 437 } 438 439 rbprev = rb_prev(&last->rb_node); 440 441 if (rbprev) 442 crq_prev = rb_entry_crq(rbprev); 443 if (rbnext) 444 crq_next = rb_entry_crq(rbnext); 445 446 return cfq_choose_req(cfqd, crq_next, crq_prev); 447} 448 449static void cfq_update_next_crq(struct cfq_rq *crq) 450{ 451 struct cfq_queue *cfqq = crq->cfq_queue; 452 453 if (cfqq->next_crq == crq) 454 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq); 455} 456 457static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted) 458{ 459 struct cfq_data *cfqd = cfqq->cfqd; 460 struct list_head *list, *entry; 461 462 BUG_ON(!cfq_cfqq_on_rr(cfqq)); 463 464 list_del(&cfqq->cfq_list); 465 466 if (cfq_class_rt(cfqq)) 467 list = &cfqd->cur_rr; 468 else if (cfq_class_idle(cfqq)) 469 list = &cfqd->idle_rr; 470 else { 471 /* 472 * if cfqq has requests in flight, don't allow it to be 473 * found in cfq_set_active_queue before it has finished them. 474 * this is done to increase fairness between a process that 475 * has lots of io pending vs one that only generates one 476 * sporadically or synchronously 477 */ 478 if (cfq_cfqq_dispatched(cfqq)) 479 list = &cfqd->busy_rr; 480 else 481 list = &cfqd->rr_list[cfqq->ioprio]; 482 } 483 484 /* 485 * if queue was preempted, just add to front to be fair. busy_rr 486 * isn't sorted. 487 */ 488 if (preempted || list == &cfqd->busy_rr) { 489 list_add(&cfqq->cfq_list, list); 490 return; 491 } 492 493 /* 494 * sort by when queue was last serviced 495 */ 496 entry = list; 497 while ((entry = entry->prev) != list) { 498 struct cfq_queue *__cfqq = list_entry_cfqq(entry); 499 500 if (!__cfqq->service_last) 501 break; 502 if (time_before(__cfqq->service_last, cfqq->service_last)) 503 break; 504 } 505 506 list_add(&cfqq->cfq_list, entry); 507} 508 509/* 510 * add to busy list of queues for service, trying to be fair in ordering 511 * the pending list according to last request service 512 */ 513static inline void 514cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) 515{ 516 BUG_ON(cfq_cfqq_on_rr(cfqq)); 517 cfq_mark_cfqq_on_rr(cfqq); 518 cfqd->busy_queues++; 519 520 cfq_resort_rr_list(cfqq, 0); 521} 522 523static inline void 524cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) 525{ 526 BUG_ON(!cfq_cfqq_on_rr(cfqq)); 527 cfq_clear_cfqq_on_rr(cfqq); 528 list_move(&cfqq->cfq_list, &cfqd->empty_list); 529 530 BUG_ON(!cfqd->busy_queues); 531 cfqd->busy_queues--; 532} 533 534/* 535 * rb tree support functions 536 */ 537static inline void cfq_del_crq_rb(struct cfq_rq *crq) 538{ 539 struct cfq_queue *cfqq = crq->cfq_queue; 540 struct cfq_data *cfqd = cfqq->cfqd; 541 const int sync = cfq_crq_is_sync(crq); 542 543 BUG_ON(!cfqq->queued[sync]); 544 cfqq->queued[sync]--; 545 546 cfq_update_next_crq(crq); 547 548 rb_erase(&crq->rb_node, &cfqq->sort_list); 549 RB_CLEAR_COLOR(&crq->rb_node); 550 551 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list)) 552 cfq_del_cfqq_rr(cfqd, cfqq); 553} 554 555static struct cfq_rq * 556__cfq_add_crq_rb(struct cfq_rq *crq) 557{ 558 struct rb_node **p = &crq->cfq_queue->sort_list.rb_node; 559 struct rb_node *parent = NULL; 560 struct cfq_rq *__crq; 561 562 while (*p) { 563 parent = *p; 564 __crq = rb_entry_crq(parent); 565 566 if (crq->rb_key < __crq->rb_key) 567 p = &(*p)->rb_left; 568 else if (crq->rb_key > __crq->rb_key) 569 p = &(*p)->rb_right; 570 else 571 return __crq; 572 } 573 574 rb_link_node(&crq->rb_node, parent, p); 575 return NULL; 576} 577 578static void cfq_add_crq_rb(struct cfq_rq *crq) 579{ 580 struct cfq_queue *cfqq = crq->cfq_queue; 581 struct cfq_data *cfqd = cfqq->cfqd; 582 struct request *rq = crq->request; 583 struct cfq_rq *__alias; 584 585 crq->rb_key = rq_rb_key(rq); 586 cfqq->queued[cfq_crq_is_sync(crq)]++; 587 588 /* 589 * looks a little odd, but the first insert might return an alias. 590 * if that happens, put the alias on the dispatch list 591 */ 592 while ((__alias = __cfq_add_crq_rb(crq)) != NULL) 593 cfq_dispatch_insert(cfqd->queue, __alias); 594 595 rb_insert_color(&crq->rb_node, &cfqq->sort_list); 596 597 if (!cfq_cfqq_on_rr(cfqq)) 598 cfq_add_cfqq_rr(cfqd, cfqq); 599 600 /* 601 * check if this request is a better next-serve candidate 602 */ 603 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq); 604} 605 606static inline void 607cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq) 608{ 609 rb_erase(&crq->rb_node, &cfqq->sort_list); 610 cfqq->queued[cfq_crq_is_sync(crq)]--; 611 612 cfq_add_crq_rb(crq); 613} 614 615static struct request *cfq_find_rq_rb(struct cfq_data *cfqd, sector_t sector) 616 617{ 618 struct cfq_queue *cfqq = cfq_find_cfq_hash(cfqd, current->pid, CFQ_KEY_ANY); 619 struct rb_node *n; 620 621 if (!cfqq) 622 goto out; 623 624 n = cfqq->sort_list.rb_node; 625 while (n) { 626 struct cfq_rq *crq = rb_entry_crq(n); 627 628 if (sector < crq->rb_key) 629 n = n->rb_left; 630 else if (sector > crq->rb_key) 631 n = n->rb_right; 632 else 633 return crq->request; 634 } 635 636out: 637 return NULL; 638} 639 640static void cfq_activate_request(request_queue_t *q, struct request *rq) 641{ 642 struct cfq_data *cfqd = q->elevator->elevator_data; 643 644 cfqd->rq_in_driver++; 645} 646 647static void cfq_deactivate_request(request_queue_t *q, struct request *rq) 648{ 649 struct cfq_data *cfqd = q->elevator->elevator_data; 650 651 WARN_ON(!cfqd->rq_in_driver); 652 cfqd->rq_in_driver--; 653} 654 655static void cfq_remove_request(struct request *rq) 656{ 657 struct cfq_rq *crq = RQ_DATA(rq); 658 659 list_del_init(&rq->queuelist); 660 cfq_del_crq_rb(crq); 661 cfq_del_crq_hash(crq); 662} 663 664static int 665cfq_merge(request_queue_t *q, struct request **req, struct bio *bio) 666{ 667 struct cfq_data *cfqd = q->elevator->elevator_data; 668 struct request *__rq; 669 int ret; 670 671 __rq = cfq_find_rq_hash(cfqd, bio->bi_sector); 672 if (__rq && elv_rq_merge_ok(__rq, bio)) { 673 ret = ELEVATOR_BACK_MERGE; 674 goto out; 675 } 676 677 __rq = cfq_find_rq_rb(cfqd, bio->bi_sector + bio_sectors(bio)); 678 if (__rq && elv_rq_merge_ok(__rq, bio)) { 679 ret = ELEVATOR_FRONT_MERGE; 680 goto out; 681 } 682 683 return ELEVATOR_NO_MERGE; 684out: 685 *req = __rq; 686 return ret; 687} 688 689static void cfq_merged_request(request_queue_t *q, struct request *req) 690{ 691 struct cfq_data *cfqd = q->elevator->elevator_data; 692 struct cfq_rq *crq = RQ_DATA(req); 693 694 cfq_del_crq_hash(crq); 695 cfq_add_crq_hash(cfqd, crq); 696 697 if (rq_rb_key(req) != crq->rb_key) { 698 struct cfq_queue *cfqq = crq->cfq_queue; 699 700 cfq_update_next_crq(crq); 701 cfq_reposition_crq_rb(cfqq, crq); 702 } 703} 704 705static void 706cfq_merged_requests(request_queue_t *q, struct request *rq, 707 struct request *next) 708{ 709 cfq_merged_request(q, rq); 710 711 /* 712 * reposition in fifo if next is older than rq 713 */ 714 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && 715 time_before(next->start_time, rq->start_time)) 716 list_move(&rq->queuelist, &next->queuelist); 717 718 cfq_remove_request(next); 719} 720 721static inline void 722__cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) 723{ 724 if (cfqq) { 725 /* 726 * stop potential idle class queues waiting service 727 */ 728 del_timer(&cfqd->idle_class_timer); 729 730 cfqq->slice_start = jiffies; 731 cfqq->slice_end = 0; 732 cfqq->slice_left = 0; 733 cfq_clear_cfqq_must_alloc_slice(cfqq); 734 cfq_clear_cfqq_fifo_expire(cfqq); 735 } 736 737 cfqd->active_queue = cfqq; 738} 739 740/* 741 * current cfqq expired its slice (or was too idle), select new one 742 */ 743static void 744__cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, 745 int preempted) 746{ 747 unsigned long now = jiffies; 748 749 if (cfq_cfqq_wait_request(cfqq)) 750 del_timer(&cfqd->idle_slice_timer); 751 752 if (!preempted && !cfq_cfqq_dispatched(cfqq)) { 753 cfqq->service_last = now; 754 cfq_schedule_dispatch(cfqd); 755 } 756 757 cfq_clear_cfqq_must_dispatch(cfqq); 758 cfq_clear_cfqq_wait_request(cfqq); 759 760 /* 761 * store what was left of this slice, if the queue idled out 762 * or was preempted 763 */ 764 if (time_after(cfqq->slice_end, now)) 765 cfqq->slice_left = cfqq->slice_end - now; 766 else 767 cfqq->slice_left = 0; 768 769 if (cfq_cfqq_on_rr(cfqq)) 770 cfq_resort_rr_list(cfqq, preempted); 771 772 if (cfqq == cfqd->active_queue) 773 cfqd->active_queue = NULL; 774 775 if (cfqd->active_cic) { 776 put_io_context(cfqd->active_cic->ioc); 777 cfqd->active_cic = NULL; 778 } 779 780 cfqd->dispatch_slice = 0; 781} 782 783static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted) 784{ 785 struct cfq_queue *cfqq = cfqd->active_queue; 786 787 if (cfqq) 788 __cfq_slice_expired(cfqd, cfqq, preempted); 789} 790 791/* 792 * 0 793 * 0,1 794 * 0,1,2 795 * 0,1,2,3 796 * 0,1,2,3,4 797 * 0,1,2,3,4,5 798 * 0,1,2,3,4,5,6 799 * 0,1,2,3,4,5,6,7 800 */ 801static int cfq_get_next_prio_level(struct cfq_data *cfqd) 802{ 803 int prio, wrap; 804 805 prio = -1; 806 wrap = 0; 807 do { 808 int p; 809 810 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) { 811 if (!list_empty(&cfqd->rr_list[p])) { 812 prio = p; 813 break; 814 } 815 } 816 817 if (prio != -1) 818 break; 819 cfqd->cur_prio = 0; 820 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) { 821 cfqd->cur_end_prio = 0; 822 if (wrap) 823 break; 824 wrap = 1; 825 } 826 } while (1); 827 828 if (unlikely(prio == -1)) 829 return -1; 830 831 BUG_ON(prio >= CFQ_PRIO_LISTS); 832 833 list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr); 834 835 cfqd->cur_prio = prio + 1; 836 if (cfqd->cur_prio > cfqd->cur_end_prio) { 837 cfqd->cur_end_prio = cfqd->cur_prio; 838 cfqd->cur_prio = 0; 839 } 840 if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) { 841 cfqd->cur_prio = 0; 842 cfqd->cur_end_prio = 0; 843 } 844 845 return prio; 846} 847 848static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd) 849{ 850 struct cfq_queue *cfqq = NULL; 851 852 /* 853 * if current list is non-empty, grab first entry. if it is empty, 854 * get next prio level and grab first entry then if any are spliced 855 */ 856 if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) 857 cfqq = list_entry_cfqq(cfqd->cur_rr.next); 858 859 /* 860 * if we have idle queues and no rt or be queues had pending 861 * requests, either allow immediate service if the grace period 862 * has passed or arm the idle grace timer 863 */ 864 if (!cfqq && !list_empty(&cfqd->idle_rr)) { 865 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE; 866 867 if (time_after_eq(jiffies, end)) 868 cfqq = list_entry_cfqq(cfqd->idle_rr.next); 869 else 870 mod_timer(&cfqd->idle_class_timer, end); 871 } 872 873 __cfq_set_active_queue(cfqd, cfqq); 874 return cfqq; 875} 876 877static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq) 878 879{ 880 unsigned long sl; 881 882 WARN_ON(!RB_EMPTY(&cfqq->sort_list)); 883 WARN_ON(cfqq != cfqd->active_queue); 884 885 /* 886 * idle is disabled, either manually or by past process history 887 */ 888 if (!cfqd->cfq_slice_idle) 889 return 0; 890 if (!cfq_cfqq_idle_window(cfqq)) 891 return 0; 892 /* 893 * task has exited, don't wait 894 */ 895 if (cfqd->active_cic && !cfqd->active_cic->ioc->task) 896 return 0; 897 898 cfq_mark_cfqq_must_dispatch(cfqq); 899 cfq_mark_cfqq_wait_request(cfqq); 900 901 sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle); 902 mod_timer(&cfqd->idle_slice_timer, jiffies + sl); 903 return 1; 904} 905 906static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq) 907{ 908 struct cfq_data *cfqd = q->elevator->elevator_data; 909 struct cfq_queue *cfqq = crq->cfq_queue; 910 911 cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq); 912 cfq_remove_request(crq->request); 913 cfqq->on_dispatch[cfq_crq_is_sync(crq)]++; 914 elv_dispatch_sort(q, crq->request); 915} 916 917/* 918 * return expired entry, or NULL to just start from scratch in rbtree 919 */ 920static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq) 921{ 922 struct cfq_data *cfqd = cfqq->cfqd; 923 struct request *rq; 924 struct cfq_rq *crq; 925 926 if (cfq_cfqq_fifo_expire(cfqq)) 927 return NULL; 928 929 if (!list_empty(&cfqq->fifo)) { 930 int fifo = cfq_cfqq_class_sync(cfqq); 931 932 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next)); 933 rq = crq->request; 934 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) { 935 cfq_mark_cfqq_fifo_expire(cfqq); 936 return crq; 937 } 938 } 939 940 return NULL; 941} 942 943/* 944 * Scale schedule slice based on io priority. Use the sync time slice only 945 * if a queue is marked sync and has sync io queued. A sync queue with async 946 * io only, should not get full sync slice length. 947 */ 948static inline int 949cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) 950{ 951 const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)]; 952 953 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); 954 955 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio)); 956} 957 958static inline void 959cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) 960{ 961 cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies; 962} 963 964static inline int 965cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) 966{ 967 const int base_rq = cfqd->cfq_slice_async_rq; 968 969 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); 970 971 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio)); 972} 973 974/* 975 * get next queue for service 976 */ 977static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) 978{ 979 unsigned long now = jiffies; 980 struct cfq_queue *cfqq; 981 982 cfqq = cfqd->active_queue; 983 if (!cfqq) 984 goto new_queue; 985 986 /* 987 * slice has expired 988 */ 989 if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end)) 990 goto expire; 991 992 /* 993 * if queue has requests, dispatch one. if not, check if 994 * enough slice is left to wait for one 995 */ 996 if (!RB_EMPTY(&cfqq->sort_list)) 997 goto keep_queue; 998 else if (cfq_cfqq_class_sync(cfqq) && 999 time_before(now, cfqq->slice_end)) { 1000 if (cfq_arm_slice_timer(cfqd, cfqq)) 1001 return NULL; 1002 } 1003 1004expire: 1005 cfq_slice_expired(cfqd, 0); 1006new_queue: 1007 cfqq = cfq_set_active_queue(cfqd); 1008keep_queue: 1009 return cfqq; 1010} 1011 1012static int 1013__cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1014 int max_dispatch) 1015{ 1016 int dispatched = 0; 1017 1018 BUG_ON(RB_EMPTY(&cfqq->sort_list)); 1019 1020 do { 1021 struct cfq_rq *crq; 1022 1023 /* 1024 * follow expired path, else get first next available 1025 */ 1026 if ((crq = cfq_check_fifo(cfqq)) == NULL) 1027 crq = cfqq->next_crq; 1028 1029 /* 1030 * finally, insert request into driver dispatch list 1031 */ 1032 cfq_dispatch_insert(cfqd->queue, crq); 1033 1034 cfqd->dispatch_slice++; 1035 dispatched++; 1036 1037 if (!cfqd->active_cic) { 1038 atomic_inc(&crq->io_context->ioc->refcount); 1039 cfqd->active_cic = crq->io_context; 1040 } 1041 1042 if (RB_EMPTY(&cfqq->sort_list)) 1043 break; 1044 1045 } while (dispatched < max_dispatch); 1046 1047 /* 1048 * if slice end isn't set yet, set it. if at least one request was 1049 * sync, use the sync time slice value 1050 */ 1051 if (!cfqq->slice_end) 1052 cfq_set_prio_slice(cfqd, cfqq); 1053 1054 /* 1055 * expire an async queue immediately if it has used up its slice. idle 1056 * queue always expire after 1 dispatch round. 1057 */ 1058 if ((!cfq_cfqq_sync(cfqq) && 1059 cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) || 1060 cfq_class_idle(cfqq)) 1061 cfq_slice_expired(cfqd, 0); 1062 1063 return dispatched; 1064} 1065 1066static int 1067cfq_forced_dispatch_cfqqs(struct list_head *list) 1068{ 1069 int dispatched = 0; 1070 struct cfq_queue *cfqq, *next; 1071 struct cfq_rq *crq; 1072 1073 list_for_each_entry_safe(cfqq, next, list, cfq_list) { 1074 while ((crq = cfqq->next_crq)) { 1075 cfq_dispatch_insert(cfqq->cfqd->queue, crq); 1076 dispatched++; 1077 } 1078 BUG_ON(!list_empty(&cfqq->fifo)); 1079 } 1080 return dispatched; 1081} 1082 1083static int 1084cfq_forced_dispatch(struct cfq_data *cfqd) 1085{ 1086 int i, dispatched = 0; 1087 1088 for (i = 0; i < CFQ_PRIO_LISTS; i++) 1089 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]); 1090 1091 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr); 1092 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr); 1093 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr); 1094 1095 cfq_slice_expired(cfqd, 0); 1096 1097 BUG_ON(cfqd->busy_queues); 1098 1099 return dispatched; 1100} 1101 1102static int 1103cfq_dispatch_requests(request_queue_t *q, int force) 1104{ 1105 struct cfq_data *cfqd = q->elevator->elevator_data; 1106 struct cfq_queue *cfqq; 1107 1108 if (!cfqd->busy_queues) 1109 return 0; 1110 1111 if (unlikely(force)) 1112 return cfq_forced_dispatch(cfqd); 1113 1114 cfqq = cfq_select_queue(cfqd); 1115 if (cfqq) { 1116 int max_dispatch; 1117 1118 /* 1119 * if idle window is disabled, allow queue buildup 1120 */ 1121 if (!cfq_cfqq_idle_window(cfqq) && 1122 cfqd->rq_in_driver >= cfqd->cfq_max_depth) 1123 return 0; 1124 1125 cfq_clear_cfqq_must_dispatch(cfqq); 1126 cfq_clear_cfqq_wait_request(cfqq); 1127 del_timer(&cfqd->idle_slice_timer); 1128 1129 max_dispatch = cfqd->cfq_quantum; 1130 if (cfq_class_idle(cfqq)) 1131 max_dispatch = 1; 1132 1133 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch); 1134 } 1135 1136 return 0; 1137} 1138 1139/* 1140 * task holds one reference to the queue, dropped when task exits. each crq 1141 * in-flight on this queue also holds a reference, dropped when crq is freed. 1142 * 1143 * queue lock must be held here. 1144 */ 1145static void cfq_put_queue(struct cfq_queue *cfqq) 1146{ 1147 struct cfq_data *cfqd = cfqq->cfqd; 1148 1149 BUG_ON(atomic_read(&cfqq->ref) <= 0); 1150 1151 if (!atomic_dec_and_test(&cfqq->ref)) 1152 return; 1153 1154 BUG_ON(rb_first(&cfqq->sort_list)); 1155 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); 1156 BUG_ON(cfq_cfqq_on_rr(cfqq)); 1157 1158 if (unlikely(cfqd->active_queue == cfqq)) 1159 __cfq_slice_expired(cfqd, cfqq, 0); 1160 1161 /* 1162 * it's on the empty list and still hashed 1163 */ 1164 list_del(&cfqq->cfq_list); 1165 hlist_del(&cfqq->cfq_hash); 1166 kmem_cache_free(cfq_pool, cfqq); 1167} 1168 1169static inline struct cfq_queue * 1170__cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio, 1171 const int hashval) 1172{ 1173 struct hlist_head *hash_list = &cfqd->cfq_hash[hashval]; 1174 struct hlist_node *entry, *next; 1175 1176 hlist_for_each_safe(entry, next, hash_list) { 1177 struct cfq_queue *__cfqq = list_entry_qhash(entry); 1178 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio); 1179 1180 if (__cfqq->key == key && (__p == prio || prio == CFQ_KEY_ANY)) 1181 return __cfqq; 1182 } 1183 1184 return NULL; 1185} 1186 1187static struct cfq_queue * 1188cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio) 1189{ 1190 return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT)); 1191} 1192 1193static void cfq_free_io_context(struct io_context *ioc) 1194{ 1195 struct cfq_io_context *__cic; 1196 struct rb_node *n; 1197 int freed = 0; 1198 1199 while ((n = rb_first(&ioc->cic_root)) != NULL) { 1200 __cic = rb_entry(n, struct cfq_io_context, rb_node); 1201 rb_erase(&__cic->rb_node, &ioc->cic_root); 1202 kmem_cache_free(cfq_ioc_pool, __cic); 1203 freed++; 1204 } 1205 1206 if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone) 1207 complete(ioc_gone); 1208} 1209 1210static void cfq_trim(struct io_context *ioc) 1211{ 1212 ioc->set_ioprio = NULL; 1213 cfq_free_io_context(ioc); 1214} 1215 1216/* 1217 * Called with interrupts disabled 1218 */ 1219static void cfq_exit_single_io_context(struct cfq_io_context *cic) 1220{ 1221 struct cfq_data *cfqd = cic->key; 1222 request_queue_t *q; 1223 1224 if (!cfqd) 1225 return; 1226 1227 q = cfqd->queue; 1228 1229 WARN_ON(!irqs_disabled()); 1230 1231 spin_lock(q->queue_lock); 1232 1233 if (cic->cfqq[ASYNC]) { 1234 if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue)) 1235 __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0); 1236 cfq_put_queue(cic->cfqq[ASYNC]); 1237 cic->cfqq[ASYNC] = NULL; 1238 } 1239 1240 if (cic->cfqq[SYNC]) { 1241 if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue)) 1242 __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0); 1243 cfq_put_queue(cic->cfqq[SYNC]); 1244 cic->cfqq[SYNC] = NULL; 1245 } 1246 1247 cic->key = NULL; 1248 list_del_init(&cic->queue_list); 1249 spin_unlock(q->queue_lock); 1250} 1251 1252static void cfq_exit_io_context(struct io_context *ioc) 1253{ 1254 struct cfq_io_context *__cic; 1255 unsigned long flags; 1256 struct rb_node *n; 1257 1258 /* 1259 * put the reference this task is holding to the various queues 1260 */ 1261 read_lock_irqsave(&cfq_exit_lock, flags); 1262 1263 n = rb_first(&ioc->cic_root); 1264 while (n != NULL) { 1265 __cic = rb_entry(n, struct cfq_io_context, rb_node); 1266 1267 cfq_exit_single_io_context(__cic); 1268 n = rb_next(n); 1269 } 1270 1271 read_unlock_irqrestore(&cfq_exit_lock, flags); 1272} 1273 1274static struct cfq_io_context * 1275cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) 1276{ 1277 struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask); 1278 1279 if (cic) { 1280 RB_CLEAR(&cic->rb_node); 1281 cic->key = NULL; 1282 cic->cfqq[ASYNC] = NULL; 1283 cic->cfqq[SYNC] = NULL; 1284 cic->last_end_request = jiffies; 1285 cic->ttime_total = 0; 1286 cic->ttime_samples = 0; 1287 cic->ttime_mean = 0; 1288 cic->dtor = cfq_free_io_context; 1289 cic->exit = cfq_exit_io_context; 1290 INIT_LIST_HEAD(&cic->queue_list); 1291 atomic_inc(&ioc_count); 1292 } 1293 1294 return cic; 1295} 1296 1297static void cfq_init_prio_data(struct cfq_queue *cfqq) 1298{ 1299 struct task_struct *tsk = current; 1300 int ioprio_class; 1301 1302 if (!cfq_cfqq_prio_changed(cfqq)) 1303 return; 1304 1305 ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio); 1306 switch (ioprio_class) { 1307 default: 1308 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); 1309 case IOPRIO_CLASS_NONE: 1310 /* 1311 * no prio set, place us in the middle of the BE classes 1312 */ 1313 cfqq->ioprio = task_nice_ioprio(tsk); 1314 cfqq->ioprio_class = IOPRIO_CLASS_BE; 1315 break; 1316 case IOPRIO_CLASS_RT: 1317 cfqq->ioprio = task_ioprio(tsk); 1318 cfqq->ioprio_class = IOPRIO_CLASS_RT; 1319 break; 1320 case IOPRIO_CLASS_BE: 1321 cfqq->ioprio = task_ioprio(tsk); 1322 cfqq->ioprio_class = IOPRIO_CLASS_BE; 1323 break; 1324 case IOPRIO_CLASS_IDLE: 1325 cfqq->ioprio_class = IOPRIO_CLASS_IDLE; 1326 cfqq->ioprio = 7; 1327 cfq_clear_cfqq_idle_window(cfqq); 1328 break; 1329 } 1330 1331 /* 1332 * keep track of original prio settings in case we have to temporarily 1333 * elevate the priority of this queue 1334 */ 1335 cfqq->org_ioprio = cfqq->ioprio; 1336 cfqq->org_ioprio_class = cfqq->ioprio_class; 1337 1338 if (cfq_cfqq_on_rr(cfqq)) 1339 cfq_resort_rr_list(cfqq, 0); 1340 1341 cfq_clear_cfqq_prio_changed(cfqq); 1342} 1343 1344static inline void changed_ioprio(struct cfq_io_context *cic) 1345{ 1346 struct cfq_data *cfqd = cic->key; 1347 struct cfq_queue *cfqq; 1348 if (cfqd) { 1349 spin_lock(cfqd->queue->queue_lock); 1350 cfqq = cic->cfqq[ASYNC]; 1351 if (cfqq) { 1352 struct cfq_queue *new_cfqq; 1353 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, 1354 cic->ioc->task, GFP_ATOMIC); 1355 if (new_cfqq) { 1356 cic->cfqq[ASYNC] = new_cfqq; 1357 cfq_put_queue(cfqq); 1358 } 1359 } 1360 cfqq = cic->cfqq[SYNC]; 1361 if (cfqq) { 1362 cfq_mark_cfqq_prio_changed(cfqq); 1363 cfq_init_prio_data(cfqq); 1364 } 1365 spin_unlock(cfqd->queue->queue_lock); 1366 } 1367} 1368 1369/* 1370 * callback from sys_ioprio_set, irqs are disabled 1371 */ 1372static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio) 1373{ 1374 struct cfq_io_context *cic; 1375 struct rb_node *n; 1376 1377 write_lock(&cfq_exit_lock); 1378 1379 n = rb_first(&ioc->cic_root); 1380 while (n != NULL) { 1381 cic = rb_entry(n, struct cfq_io_context, rb_node); 1382 1383 changed_ioprio(cic); 1384 n = rb_next(n); 1385 } 1386 1387 write_unlock(&cfq_exit_lock); 1388 1389 return 0; 1390} 1391 1392static struct cfq_queue * 1393cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, 1394 gfp_t gfp_mask) 1395{ 1396 const int hashval = hash_long(key, CFQ_QHASH_SHIFT); 1397 struct cfq_queue *cfqq, *new_cfqq = NULL; 1398 unsigned short ioprio; 1399 1400retry: 1401 ioprio = tsk->ioprio; 1402 cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval); 1403 1404 if (!cfqq) { 1405 if (new_cfqq) { 1406 cfqq = new_cfqq; 1407 new_cfqq = NULL; 1408 } else if (gfp_mask & __GFP_WAIT) { 1409 spin_unlock_irq(cfqd->queue->queue_lock); 1410 new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask); 1411 spin_lock_irq(cfqd->queue->queue_lock); 1412 goto retry; 1413 } else { 1414 cfqq = kmem_cache_alloc(cfq_pool, gfp_mask); 1415 if (!cfqq) 1416 goto out; 1417 } 1418 1419 memset(cfqq, 0, sizeof(*cfqq)); 1420 1421 INIT_HLIST_NODE(&cfqq->cfq_hash); 1422 INIT_LIST_HEAD(&cfqq->cfq_list); 1423 RB_CLEAR_ROOT(&cfqq->sort_list); 1424 INIT_LIST_HEAD(&cfqq->fifo); 1425 1426 cfqq->key = key; 1427 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]); 1428 atomic_set(&cfqq->ref, 0); 1429 cfqq->cfqd = cfqd; 1430 cfqq->service_last = 0; 1431 /* 1432 * set ->slice_left to allow preemption for a new process 1433 */ 1434 cfqq->slice_left = 2 * cfqd->cfq_slice_idle; 1435 cfq_mark_cfqq_idle_window(cfqq); 1436 cfq_mark_cfqq_prio_changed(cfqq); 1437 cfq_init_prio_data(cfqq); 1438 } 1439 1440 if (new_cfqq) 1441 kmem_cache_free(cfq_pool, new_cfqq); 1442 1443 atomic_inc(&cfqq->ref); 1444out: 1445 WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq); 1446 return cfqq; 1447} 1448 1449static struct cfq_io_context * 1450cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc) 1451{ 1452 struct rb_node *n = ioc->cic_root.rb_node; 1453 struct cfq_io_context *cic; 1454 void *key = cfqd; 1455 1456 while (n) { 1457 cic = rb_entry(n, struct cfq_io_context, rb_node); 1458 1459 if (key < cic->key) 1460 n = n->rb_left; 1461 else if (key > cic->key) 1462 n = n->rb_right; 1463 else 1464 return cic; 1465 } 1466 1467 return NULL; 1468} 1469 1470static inline void 1471cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, 1472 struct cfq_io_context *cic) 1473{ 1474 struct rb_node **p = &ioc->cic_root.rb_node; 1475 struct rb_node *parent = NULL; 1476 struct cfq_io_context *__cic; 1477 1478 read_lock(&cfq_exit_lock); 1479 1480 cic->ioc = ioc; 1481 cic->key = cfqd; 1482 1483 ioc->set_ioprio = cfq_ioc_set_ioprio; 1484 1485 while (*p) { 1486 parent = *p; 1487 __cic = rb_entry(parent, struct cfq_io_context, rb_node); 1488 1489 if (cic->key < __cic->key) 1490 p = &(*p)->rb_left; 1491 else if (cic->key > __cic->key) 1492 p = &(*p)->rb_right; 1493 else 1494 BUG(); 1495 } 1496 1497 rb_link_node(&cic->rb_node, parent, p); 1498 rb_insert_color(&cic->rb_node, &ioc->cic_root); 1499 list_add(&cic->queue_list, &cfqd->cic_list); 1500 read_unlock(&cfq_exit_lock); 1501} 1502 1503/* 1504 * Setup general io context and cfq io context. There can be several cfq 1505 * io contexts per general io context, if this process is doing io to more 1506 * than one device managed by cfq. 1507 */ 1508static struct cfq_io_context * 1509cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) 1510{ 1511 struct io_context *ioc = NULL; 1512 struct cfq_io_context *cic; 1513 1514 might_sleep_if(gfp_mask & __GFP_WAIT); 1515 1516 ioc = get_io_context(gfp_mask); 1517 if (!ioc) 1518 return NULL; 1519 1520 cic = cfq_cic_rb_lookup(cfqd, ioc); 1521 if (cic) 1522 goto out; 1523 1524 cic = cfq_alloc_io_context(cfqd, gfp_mask); 1525 if (cic == NULL) 1526 goto err; 1527 1528 cfq_cic_link(cfqd, ioc, cic); 1529out: 1530 return cic; 1531err: 1532 put_io_context(ioc); 1533 return NULL; 1534} 1535 1536static void 1537cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) 1538{ 1539 unsigned long elapsed, ttime; 1540 1541 /* 1542 * if this context already has stuff queued, thinktime is from 1543 * last queue not last end 1544 */ 1545#if 0 1546 if (time_after(cic->last_end_request, cic->last_queue)) 1547 elapsed = jiffies - cic->last_end_request; 1548 else 1549 elapsed = jiffies - cic->last_queue; 1550#else 1551 elapsed = jiffies - cic->last_end_request; 1552#endif 1553 1554 ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); 1555 1556 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; 1557 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; 1558 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; 1559} 1560 1561#define sample_valid(samples) ((samples) > 80) 1562 1563/* 1564 * Disable idle window if the process thinks too long or seeks so much that 1565 * it doesn't matter 1566 */ 1567static void 1568cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1569 struct cfq_io_context *cic) 1570{ 1571 int enable_idle = cfq_cfqq_idle_window(cfqq); 1572 1573 if (!cic->ioc->task || !cfqd->cfq_slice_idle) 1574 enable_idle = 0; 1575 else if (sample_valid(cic->ttime_samples)) { 1576 if (cic->ttime_mean > cfqd->cfq_slice_idle) 1577 enable_idle = 0; 1578 else 1579 enable_idle = 1; 1580 } 1581 1582 if (enable_idle) 1583 cfq_mark_cfqq_idle_window(cfqq); 1584 else 1585 cfq_clear_cfqq_idle_window(cfqq); 1586} 1587 1588 1589/* 1590 * Check if new_cfqq should preempt the currently active queue. Return 0 for 1591 * no or if we aren't sure, a 1 will cause a preempt. 1592 */ 1593static int 1594cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, 1595 struct cfq_rq *crq) 1596{ 1597 struct cfq_queue *cfqq = cfqd->active_queue; 1598 1599 if (cfq_class_idle(new_cfqq)) 1600 return 0; 1601 1602 if (!cfqq) 1603 return 1; 1604 1605 if (cfq_class_idle(cfqq)) 1606 return 1; 1607 if (!cfq_cfqq_wait_request(new_cfqq)) 1608 return 0; 1609 /* 1610 * if it doesn't have slice left, forget it 1611 */ 1612 if (new_cfqq->slice_left < cfqd->cfq_slice_idle) 1613 return 0; 1614 if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq)) 1615 return 1; 1616 1617 return 0; 1618} 1619 1620/* 1621 * cfqq preempts the active queue. if we allowed preempt with no slice left, 1622 * let it have half of its nominal slice. 1623 */ 1624static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1625{ 1626 struct cfq_queue *__cfqq, *next; 1627 1628 list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list) 1629 cfq_resort_rr_list(__cfqq, 1); 1630 1631 if (!cfqq->slice_left) 1632 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2; 1633 1634 cfqq->slice_end = cfqq->slice_left + jiffies; 1635 __cfq_slice_expired(cfqd, cfqq, 1); 1636 __cfq_set_active_queue(cfqd, cfqq); 1637} 1638 1639/* 1640 * should really be a ll_rw_blk.c helper 1641 */ 1642static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq) 1643{ 1644 request_queue_t *q = cfqd->queue; 1645 1646 if (!blk_queue_plugged(q)) 1647 q->request_fn(q); 1648 else 1649 __generic_unplug_device(q); 1650} 1651 1652/* 1653 * Called when a new fs request (crq) is added (to cfqq). Check if there's 1654 * something we should do about it 1655 */ 1656static void 1657cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1658 struct cfq_rq *crq) 1659{ 1660 struct cfq_io_context *cic; 1661 1662 cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq); 1663 1664 /* 1665 * we never wait for an async request and we don't allow preemption 1666 * of an async request. so just return early 1667 */ 1668 if (!cfq_crq_is_sync(crq)) 1669 return; 1670 1671 cic = crq->io_context; 1672 1673 cfq_update_io_thinktime(cfqd, cic); 1674 cfq_update_idle_window(cfqd, cfqq, cic); 1675 1676 cic->last_queue = jiffies; 1677 1678 if (cfqq == cfqd->active_queue) { 1679 /* 1680 * if we are waiting for a request for this queue, let it rip 1681 * immediately and flag that we must not expire this queue 1682 * just now 1683 */ 1684 if (cfq_cfqq_wait_request(cfqq)) { 1685 cfq_mark_cfqq_must_dispatch(cfqq); 1686 del_timer(&cfqd->idle_slice_timer); 1687 cfq_start_queueing(cfqd, cfqq); 1688 } 1689 } else if (cfq_should_preempt(cfqd, cfqq, crq)) { 1690 /* 1691 * not the active queue - expire current slice if it is 1692 * idle and has expired it's mean thinktime or this new queue 1693 * has some old slice time left and is of higher priority 1694 */ 1695 cfq_preempt_queue(cfqd, cfqq); 1696 cfq_mark_cfqq_must_dispatch(cfqq); 1697 cfq_start_queueing(cfqd, cfqq); 1698 } 1699} 1700 1701static void cfq_insert_request(request_queue_t *q, struct request *rq) 1702{ 1703 struct cfq_data *cfqd = q->elevator->elevator_data; 1704 struct cfq_rq *crq = RQ_DATA(rq); 1705 struct cfq_queue *cfqq = crq->cfq_queue; 1706 1707 cfq_init_prio_data(cfqq); 1708 1709 cfq_add_crq_rb(crq); 1710 1711 list_add_tail(&rq->queuelist, &cfqq->fifo); 1712 1713 if (rq_mergeable(rq)) 1714 cfq_add_crq_hash(cfqd, crq); 1715 1716 cfq_crq_enqueued(cfqd, cfqq, crq); 1717} 1718 1719static void cfq_completed_request(request_queue_t *q, struct request *rq) 1720{ 1721 struct cfq_rq *crq = RQ_DATA(rq); 1722 struct cfq_queue *cfqq = crq->cfq_queue; 1723 struct cfq_data *cfqd = cfqq->cfqd; 1724 const int sync = cfq_crq_is_sync(crq); 1725 unsigned long now; 1726 1727 now = jiffies; 1728 1729 WARN_ON(!cfqd->rq_in_driver); 1730 WARN_ON(!cfqq->on_dispatch[sync]); 1731 cfqd->rq_in_driver--; 1732 cfqq->on_dispatch[sync]--; 1733 1734 if (!cfq_class_idle(cfqq)) 1735 cfqd->last_end_request = now; 1736 1737 if (!cfq_cfqq_dispatched(cfqq)) { 1738 if (cfq_cfqq_on_rr(cfqq)) { 1739 cfqq->service_last = now; 1740 cfq_resort_rr_list(cfqq, 0); 1741 } 1742 cfq_schedule_dispatch(cfqd); 1743 } 1744 1745 if (cfq_crq_is_sync(crq)) 1746 crq->io_context->last_end_request = now; 1747} 1748 1749static struct request * 1750cfq_former_request(request_queue_t *q, struct request *rq) 1751{ 1752 struct cfq_rq *crq = RQ_DATA(rq); 1753 struct rb_node *rbprev = rb_prev(&crq->rb_node); 1754 1755 if (rbprev) 1756 return rb_entry_crq(rbprev)->request; 1757 1758 return NULL; 1759} 1760 1761static struct request * 1762cfq_latter_request(request_queue_t *q, struct request *rq) 1763{ 1764 struct cfq_rq *crq = RQ_DATA(rq); 1765 struct rb_node *rbnext = rb_next(&crq->rb_node); 1766 1767 if (rbnext) 1768 return rb_entry_crq(rbnext)->request; 1769 1770 return NULL; 1771} 1772 1773/* 1774 * we temporarily boost lower priority queues if they are holding fs exclusive 1775 * resources. they are boosted to normal prio (CLASS_BE/4) 1776 */ 1777static void cfq_prio_boost(struct cfq_queue *cfqq) 1778{ 1779 const int ioprio_class = cfqq->ioprio_class; 1780 const int ioprio = cfqq->ioprio; 1781 1782 if (has_fs_excl()) { 1783 /* 1784 * boost idle prio on transactions that would lock out other 1785 * users of the filesystem 1786 */ 1787 if (cfq_class_idle(cfqq)) 1788 cfqq->ioprio_class = IOPRIO_CLASS_BE; 1789 if (cfqq->ioprio > IOPRIO_NORM) 1790 cfqq->ioprio = IOPRIO_NORM; 1791 } else { 1792 /* 1793 * check if we need to unboost the queue 1794 */ 1795 if (cfqq->ioprio_class != cfqq->org_ioprio_class) 1796 cfqq->ioprio_class = cfqq->org_ioprio_class; 1797 if (cfqq->ioprio != cfqq->org_ioprio) 1798 cfqq->ioprio = cfqq->org_ioprio; 1799 } 1800 1801 /* 1802 * refile between round-robin lists if we moved the priority class 1803 */ 1804 if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) && 1805 cfq_cfqq_on_rr(cfqq)) 1806 cfq_resort_rr_list(cfqq, 0); 1807} 1808 1809static inline pid_t cfq_queue_pid(struct task_struct *task, int rw) 1810{ 1811 if (rw == READ || process_sync(task)) 1812 return task->pid; 1813 1814 return CFQ_KEY_ASYNC; 1815} 1816 1817static inline int 1818__cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq, 1819 struct task_struct *task, int rw) 1820{ 1821#if 1 1822 if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) && 1823 !cfq_cfqq_must_alloc_slice(cfqq)) { 1824 cfq_mark_cfqq_must_alloc_slice(cfqq); 1825 return ELV_MQUEUE_MUST; 1826 } 1827 1828 return ELV_MQUEUE_MAY; 1829#else 1830 if (!cfqq || task->flags & PF_MEMALLOC) 1831 return ELV_MQUEUE_MAY; 1832 if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) { 1833 if (cfq_cfqq_wait_request(cfqq)) 1834 return ELV_MQUEUE_MUST; 1835 1836 /* 1837 * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we 1838 * can quickly flood the queue with writes from a single task 1839 */ 1840 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) { 1841 cfq_mark_cfqq_must_alloc_slice(cfqq); 1842 return ELV_MQUEUE_MUST; 1843 } 1844 1845 return ELV_MQUEUE_MAY; 1846 } 1847 if (cfq_class_idle(cfqq)) 1848 return ELV_MQUEUE_NO; 1849 if (cfqq->allocated[rw] >= cfqd->max_queued) { 1850 struct io_context *ioc = get_io_context(GFP_ATOMIC); 1851 int ret = ELV_MQUEUE_NO; 1852 1853 if (ioc && ioc->nr_batch_requests) 1854 ret = ELV_MQUEUE_MAY; 1855 1856 put_io_context(ioc); 1857 return ret; 1858 } 1859 1860 return ELV_MQUEUE_MAY; 1861#endif 1862} 1863 1864static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio) 1865{ 1866 struct cfq_data *cfqd = q->elevator->elevator_data; 1867 struct task_struct *tsk = current; 1868 struct cfq_queue *cfqq; 1869 1870 /* 1871 * don't force setup of a queue from here, as a call to may_queue 1872 * does not necessarily imply that a request actually will be queued. 1873 * so just lookup a possibly existing queue, or return 'may queue' 1874 * if that fails 1875 */ 1876 cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio); 1877 if (cfqq) { 1878 cfq_init_prio_data(cfqq); 1879 cfq_prio_boost(cfqq); 1880 1881 return __cfq_may_queue(cfqd, cfqq, tsk, rw); 1882 } 1883 1884 return ELV_MQUEUE_MAY; 1885} 1886 1887static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq) 1888{ 1889 struct cfq_data *cfqd = q->elevator->elevator_data; 1890 struct request_list *rl = &q->rq; 1891 1892 if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) { 1893 smp_mb(); 1894 if (waitqueue_active(&rl->wait[READ])) 1895 wake_up(&rl->wait[READ]); 1896 } 1897 1898 if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) { 1899 smp_mb(); 1900 if (waitqueue_active(&rl->wait[WRITE])) 1901 wake_up(&rl->wait[WRITE]); 1902 } 1903} 1904 1905/* 1906 * queue lock held here 1907 */ 1908static void cfq_put_request(request_queue_t *q, struct request *rq) 1909{ 1910 struct cfq_data *cfqd = q->elevator->elevator_data; 1911 struct cfq_rq *crq = RQ_DATA(rq); 1912 1913 if (crq) { 1914 struct cfq_queue *cfqq = crq->cfq_queue; 1915 const int rw = rq_data_dir(rq); 1916 1917 BUG_ON(!cfqq->allocated[rw]); 1918 cfqq->allocated[rw]--; 1919 1920 put_io_context(crq->io_context->ioc); 1921 1922 mempool_free(crq, cfqd->crq_pool); 1923 rq->elevator_private = NULL; 1924 1925 cfq_check_waiters(q, cfqq); 1926 cfq_put_queue(cfqq); 1927 } 1928} 1929 1930/* 1931 * Allocate cfq data structures associated with this request. 1932 */ 1933static int 1934cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio, 1935 gfp_t gfp_mask) 1936{ 1937 struct cfq_data *cfqd = q->elevator->elevator_data; 1938 struct task_struct *tsk = current; 1939 struct cfq_io_context *cic; 1940 const int rw = rq_data_dir(rq); 1941 pid_t key = cfq_queue_pid(tsk, rw); 1942 struct cfq_queue *cfqq; 1943 struct cfq_rq *crq; 1944 unsigned long flags; 1945 int is_sync = key != CFQ_KEY_ASYNC; 1946 1947 might_sleep_if(gfp_mask & __GFP_WAIT); 1948 1949 cic = cfq_get_io_context(cfqd, gfp_mask); 1950 1951 spin_lock_irqsave(q->queue_lock, flags); 1952 1953 if (!cic) 1954 goto queue_fail; 1955 1956 if (!cic->cfqq[is_sync]) { 1957 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask); 1958 if (!cfqq) 1959 goto queue_fail; 1960 1961 cic->cfqq[is_sync] = cfqq; 1962 } else 1963 cfqq = cic->cfqq[is_sync]; 1964 1965 cfqq->allocated[rw]++; 1966 cfq_clear_cfqq_must_alloc(cfqq); 1967 cfqd->rq_starved = 0; 1968 atomic_inc(&cfqq->ref); 1969 spin_unlock_irqrestore(q->queue_lock, flags); 1970 1971 crq = mempool_alloc(cfqd->crq_pool, gfp_mask); 1972 if (crq) { 1973 RB_CLEAR(&crq->rb_node); 1974 crq->rb_key = 0; 1975 crq->request = rq; 1976 INIT_HLIST_NODE(&crq->hash); 1977 crq->cfq_queue = cfqq; 1978 crq->io_context = cic; 1979 1980 if (is_sync) 1981 cfq_mark_crq_is_sync(crq); 1982 else 1983 cfq_clear_crq_is_sync(crq); 1984 1985 rq->elevator_private = crq; 1986 return 0; 1987 } 1988 1989 spin_lock_irqsave(q->queue_lock, flags); 1990 cfqq->allocated[rw]--; 1991 if (!(cfqq->allocated[0] + cfqq->allocated[1])) 1992 cfq_mark_cfqq_must_alloc(cfqq); 1993 cfq_put_queue(cfqq); 1994queue_fail: 1995 if (cic) 1996 put_io_context(cic->ioc); 1997 /* 1998 * mark us rq allocation starved. we need to kickstart the process 1999 * ourselves if there are no pending requests that can do it for us. 2000 * that would be an extremely rare OOM situation 2001 */ 2002 cfqd->rq_starved = 1; 2003 cfq_schedule_dispatch(cfqd); 2004 spin_unlock_irqrestore(q->queue_lock, flags); 2005 return 1; 2006} 2007 2008static void cfq_kick_queue(void *data) 2009{ 2010 request_queue_t *q = data; 2011 struct cfq_data *cfqd = q->elevator->elevator_data; 2012 unsigned long flags; 2013 2014 spin_lock_irqsave(q->queue_lock, flags); 2015 2016 if (cfqd->rq_starved) { 2017 struct request_list *rl = &q->rq; 2018 2019 /* 2020 * we aren't guaranteed to get a request after this, but we 2021 * have to be opportunistic 2022 */ 2023 smp_mb(); 2024 if (waitqueue_active(&rl->wait[READ])) 2025 wake_up(&rl->wait[READ]); 2026 if (waitqueue_active(&rl->wait[WRITE])) 2027 wake_up(&rl->wait[WRITE]); 2028 } 2029 2030 blk_remove_plug(q); 2031 q->request_fn(q); 2032 spin_unlock_irqrestore(q->queue_lock, flags); 2033} 2034 2035/* 2036 * Timer running if the active_queue is currently idling inside its time slice 2037 */ 2038static void cfq_idle_slice_timer(unsigned long data) 2039{ 2040 struct cfq_data *cfqd = (struct cfq_data *) data; 2041 struct cfq_queue *cfqq; 2042 unsigned long flags; 2043 2044 spin_lock_irqsave(cfqd->queue->queue_lock, flags); 2045 2046 if ((cfqq = cfqd->active_queue) != NULL) { 2047 unsigned long now = jiffies; 2048 2049 /* 2050 * expired 2051 */ 2052 if (time_after(now, cfqq->slice_end)) 2053 goto expire; 2054 2055 /* 2056 * only expire and reinvoke request handler, if there are 2057 * other queues with pending requests 2058 */ 2059 if (!cfqd->busy_queues) { 2060 cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end); 2061 add_timer(&cfqd->idle_slice_timer); 2062 goto out_cont; 2063 } 2064 2065 /* 2066 * not expired and it has a request pending, let it dispatch 2067 */ 2068 if (!RB_EMPTY(&cfqq->sort_list)) { 2069 cfq_mark_cfqq_must_dispatch(cfqq); 2070 goto out_kick; 2071 } 2072 } 2073expire: 2074 cfq_slice_expired(cfqd, 0); 2075out_kick: 2076 cfq_schedule_dispatch(cfqd); 2077out_cont: 2078 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); 2079} 2080 2081/* 2082 * Timer running if an idle class queue is waiting for service 2083 */ 2084static void cfq_idle_class_timer(unsigned long data) 2085{ 2086 struct cfq_data *cfqd = (struct cfq_data *) data; 2087 unsigned long flags, end; 2088 2089 spin_lock_irqsave(cfqd->queue->queue_lock, flags); 2090 2091 /* 2092 * race with a non-idle queue, reset timer 2093 */ 2094 end = cfqd->last_end_request + CFQ_IDLE_GRACE; 2095 if (!time_after_eq(jiffies, end)) { 2096 cfqd->idle_class_timer.expires = end; 2097 add_timer(&cfqd->idle_class_timer); 2098 } else 2099 cfq_schedule_dispatch(cfqd); 2100 2101 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); 2102} 2103 2104static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) 2105{ 2106 del_timer_sync(&cfqd->idle_slice_timer); 2107 del_timer_sync(&cfqd->idle_class_timer); 2108 blk_sync_queue(cfqd->queue); 2109} 2110 2111static void cfq_exit_queue(elevator_t *e) 2112{ 2113 struct cfq_data *cfqd = e->elevator_data; 2114 request_queue_t *q = cfqd->queue; 2115 2116 cfq_shutdown_timer_wq(cfqd); 2117 2118 write_lock(&cfq_exit_lock); 2119 spin_lock_irq(q->queue_lock); 2120 2121 if (cfqd->active_queue) 2122 __cfq_slice_expired(cfqd, cfqd->active_queue, 0); 2123 2124 while (!list_empty(&cfqd->cic_list)) { 2125 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, 2126 struct cfq_io_context, 2127 queue_list); 2128 if (cic->cfqq[ASYNC]) { 2129 cfq_put_queue(cic->cfqq[ASYNC]); 2130 cic->cfqq[ASYNC] = NULL; 2131 } 2132 if (cic->cfqq[SYNC]) { 2133 cfq_put_queue(cic->cfqq[SYNC]); 2134 cic->cfqq[SYNC] = NULL; 2135 } 2136 cic->key = NULL; 2137 list_del_init(&cic->queue_list); 2138 } 2139 2140 spin_unlock_irq(q->queue_lock); 2141 write_unlock(&cfq_exit_lock); 2142 2143 cfq_shutdown_timer_wq(cfqd); 2144 2145 mempool_destroy(cfqd->crq_pool); 2146 kfree(cfqd->crq_hash); 2147 kfree(cfqd->cfq_hash); 2148 kfree(cfqd); 2149} 2150 2151static int cfq_init_queue(request_queue_t *q, elevator_t *e) 2152{ 2153 struct cfq_data *cfqd; 2154 int i; 2155 2156 cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL); 2157 if (!cfqd) 2158 return -ENOMEM; 2159 2160 memset(cfqd, 0, sizeof(*cfqd)); 2161 2162 for (i = 0; i < CFQ_PRIO_LISTS; i++) 2163 INIT_LIST_HEAD(&cfqd->rr_list[i]); 2164 2165 INIT_LIST_HEAD(&cfqd->busy_rr); 2166 INIT_LIST_HEAD(&cfqd->cur_rr); 2167 INIT_LIST_HEAD(&cfqd->idle_rr); 2168 INIT_LIST_HEAD(&cfqd->empty_list); 2169 INIT_LIST_HEAD(&cfqd->cic_list); 2170 2171 cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL); 2172 if (!cfqd->crq_hash) 2173 goto out_crqhash; 2174 2175 cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL); 2176 if (!cfqd->cfq_hash) 2177 goto out_cfqhash; 2178 2179 cfqd->crq_pool = mempool_create_slab_pool(BLKDEV_MIN_RQ, crq_pool); 2180 if (!cfqd->crq_pool) 2181 goto out_crqpool; 2182 2183 for (i = 0; i < CFQ_MHASH_ENTRIES; i++) 2184 INIT_HLIST_HEAD(&cfqd->crq_hash[i]); 2185 for (i = 0; i < CFQ_QHASH_ENTRIES; i++) 2186 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]); 2187 2188 e->elevator_data = cfqd; 2189 2190 cfqd->queue = q; 2191 2192 cfqd->max_queued = q->nr_requests / 4; 2193 q->nr_batching = cfq_queued; 2194 2195 init_timer(&cfqd->idle_slice_timer); 2196 cfqd->idle_slice_timer.function = cfq_idle_slice_timer; 2197 cfqd->idle_slice_timer.data = (unsigned long) cfqd; 2198 2199 init_timer(&cfqd->idle_class_timer); 2200 cfqd->idle_class_timer.function = cfq_idle_class_timer; 2201 cfqd->idle_class_timer.data = (unsigned long) cfqd; 2202 2203 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q); 2204 2205 cfqd->cfq_queued = cfq_queued; 2206 cfqd->cfq_quantum = cfq_quantum; 2207 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; 2208 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; 2209 cfqd->cfq_back_max = cfq_back_max; 2210 cfqd->cfq_back_penalty = cfq_back_penalty; 2211 cfqd->cfq_slice[0] = cfq_slice_async; 2212 cfqd->cfq_slice[1] = cfq_slice_sync; 2213 cfqd->cfq_slice_async_rq = cfq_slice_async_rq; 2214 cfqd->cfq_slice_idle = cfq_slice_idle; 2215 cfqd->cfq_max_depth = cfq_max_depth; 2216 2217 return 0; 2218out_crqpool: 2219 kfree(cfqd->cfq_hash); 2220out_cfqhash: 2221 kfree(cfqd->crq_hash); 2222out_crqhash: 2223 kfree(cfqd); 2224 return -ENOMEM; 2225} 2226 2227static void cfq_slab_kill(void) 2228{ 2229 if (crq_pool) 2230 kmem_cache_destroy(crq_pool); 2231 if (cfq_pool) 2232 kmem_cache_destroy(cfq_pool); 2233 if (cfq_ioc_pool) 2234 kmem_cache_destroy(cfq_ioc_pool); 2235} 2236 2237static int __init cfq_slab_setup(void) 2238{ 2239 crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0, 2240 NULL, NULL); 2241 if (!crq_pool) 2242 goto fail; 2243 2244 cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0, 2245 NULL, NULL); 2246 if (!cfq_pool) 2247 goto fail; 2248 2249 cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool", 2250 sizeof(struct cfq_io_context), 0, 0, NULL, NULL); 2251 if (!cfq_ioc_pool) 2252 goto fail; 2253 2254 return 0; 2255fail: 2256 cfq_slab_kill(); 2257 return -ENOMEM; 2258} 2259 2260/* 2261 * sysfs parts below --> 2262 */ 2263 2264static ssize_t 2265cfq_var_show(unsigned int var, char *page) 2266{ 2267 return sprintf(page, "%d\n", var); 2268} 2269 2270static ssize_t 2271cfq_var_store(unsigned int *var, const char *page, size_t count) 2272{ 2273 char *p = (char *) page; 2274 2275 *var = simple_strtoul(p, &p, 10); 2276 return count; 2277} 2278 2279#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ 2280static ssize_t __FUNC(elevator_t *e, char *page) \ 2281{ \ 2282 struct cfq_data *cfqd = e->elevator_data; \ 2283 unsigned int __data = __VAR; \ 2284 if (__CONV) \ 2285 __data = jiffies_to_msecs(__data); \ 2286 return cfq_var_show(__data, (page)); \ 2287} 2288SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); 2289SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0); 2290SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); 2291SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); 2292SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); 2293SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); 2294SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); 2295SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); 2296SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); 2297SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); 2298SHOW_FUNCTION(cfq_max_depth_show, cfqd->cfq_max_depth, 0); 2299#undef SHOW_FUNCTION 2300 2301#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ 2302static ssize_t __FUNC(elevator_t *e, const char *page, size_t count) \ 2303{ \ 2304 struct cfq_data *cfqd = e->elevator_data; \ 2305 unsigned int __data; \ 2306 int ret = cfq_var_store(&__data, (page), count); \ 2307 if (__data < (MIN)) \ 2308 __data = (MIN); \ 2309 else if (__data > (MAX)) \ 2310 __data = (MAX); \ 2311 if (__CONV) \ 2312 *(__PTR) = msecs_to_jiffies(__data); \ 2313 else \ 2314 *(__PTR) = __data; \ 2315 return ret; \ 2316} 2317STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); 2318STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0); 2319STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1); 2320STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1); 2321STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); 2322STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0); 2323STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); 2324STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); 2325STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); 2326STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0); 2327STORE_FUNCTION(cfq_max_depth_store, &cfqd->cfq_max_depth, 1, UINT_MAX, 0); 2328#undef STORE_FUNCTION 2329 2330#define CFQ_ATTR(name) \ 2331 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) 2332 2333static struct elv_fs_entry cfq_attrs[] = { 2334 CFQ_ATTR(quantum), 2335 CFQ_ATTR(queued), 2336 CFQ_ATTR(fifo_expire_sync), 2337 CFQ_ATTR(fifo_expire_async), 2338 CFQ_ATTR(back_seek_max), 2339 CFQ_ATTR(back_seek_penalty), 2340 CFQ_ATTR(slice_sync), 2341 CFQ_ATTR(slice_async), 2342 CFQ_ATTR(slice_async_rq), 2343 CFQ_ATTR(slice_idle), 2344 CFQ_ATTR(max_depth), 2345 __ATTR_NULL 2346}; 2347 2348static struct elevator_type iosched_cfq = { 2349 .ops = { 2350 .elevator_merge_fn = cfq_merge, 2351 .elevator_merged_fn = cfq_merged_request, 2352 .elevator_merge_req_fn = cfq_merged_requests, 2353 .elevator_dispatch_fn = cfq_dispatch_requests, 2354 .elevator_add_req_fn = cfq_insert_request, 2355 .elevator_activate_req_fn = cfq_activate_request, 2356 .elevator_deactivate_req_fn = cfq_deactivate_request, 2357 .elevator_queue_empty_fn = cfq_queue_empty, 2358 .elevator_completed_req_fn = cfq_completed_request, 2359 .elevator_former_req_fn = cfq_former_request, 2360 .elevator_latter_req_fn = cfq_latter_request, 2361 .elevator_set_req_fn = cfq_set_request, 2362 .elevator_put_req_fn = cfq_put_request, 2363 .elevator_may_queue_fn = cfq_may_queue, 2364 .elevator_init_fn = cfq_init_queue, 2365 .elevator_exit_fn = cfq_exit_queue, 2366 .trim = cfq_trim, 2367 }, 2368 .elevator_attrs = cfq_attrs, 2369 .elevator_name = "cfq", 2370 .elevator_owner = THIS_MODULE, 2371}; 2372 2373static int __init cfq_init(void) 2374{ 2375 int ret; 2376 2377 /* 2378 * could be 0 on HZ < 1000 setups 2379 */ 2380 if (!cfq_slice_async) 2381 cfq_slice_async = 1; 2382 if (!cfq_slice_idle) 2383 cfq_slice_idle = 1; 2384 2385 if (cfq_slab_setup()) 2386 return -ENOMEM; 2387 2388 ret = elv_register(&iosched_cfq); 2389 if (ret) 2390 cfq_slab_kill(); 2391 2392 return ret; 2393} 2394 2395static void __exit cfq_exit(void) 2396{ 2397 DECLARE_COMPLETION(all_gone); 2398 elv_unregister(&iosched_cfq); 2399 ioc_gone = &all_gone; 2400 barrier(); 2401 if (atomic_read(&ioc_count)) 2402 complete(ioc_gone); 2403 synchronize_rcu(); 2404 cfq_slab_kill(); 2405} 2406 2407module_init(cfq_init); 2408module_exit(cfq_exit); 2409 2410MODULE_AUTHOR("Jens Axboe"); 2411MODULE_LICENSE("GPL"); 2412MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); 2413