block/elevator.c

   1 /*
   2  *  Block device elevator/IO-scheduler.
   3  *
   4  *  Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
   5  *
   6  * 30042000 Jens Axboe <axboe@kernel.dk> :
   7  *
   8  * Split the elevator a bit so that it is possible to choose a different
   9  * one or even write a new "plug in". There are three pieces:
  10  * - elevator_fn, inserts a new request in the queue list
  11  * - elevator_merge_fn, decides whether a new buffer can be merged with
  12  *   an existing request
  13  * - elevator_dequeue_fn, called when a request is taken off the active list
  14  *
  15  * 20082000 Dave Jones <davej@suse.de> :
  16  * Removed tests for max-bomb-segments, which was breaking elvtune
  17  *  when run without -bN
  18  *
  19  * Jens:
  20  * - Rework again to work with bio instead of buffer_heads
  21  * - loose bi_dev comparisons, partition handling is right now
  22  * - completely modularize elevator setup and teardown
  23  *
  24  */
  25 #include <linux/kernel.h>
  26 #include <linux/fs.h>
  27 #include <linux/blkdev.h>
  28 #include <linux/elevator.h>
  29 #include <linux/bio.h>
  30 #include <linux/module.h>
  31 #include <linux/slab.h>
  32 #include <linux/init.h>
  33 #include <linux/compiler.h>
  34 #include <linux/blktrace_api.h>
  35 #include <linux/hash.h>
  36 #include <linux/uaccess.h>
  37 #include <linux/pm_runtime.h>
  38 #include <linux/blk-cgroup.h>
  39
  40 #include <trace/events/block.h>
  41
  42 #include "blk.h"
  43 #include "blk-mq-sched.h"
  44 #include "blk-wbt.h"
  45
  46 static DEFINE_SPINLOCK(elv_list_lock);
  47 static LIST_HEAD(elv_list);
  48
  49 /*
  50  * Merge hash stuff.
  51  */
  52 #define rq_hash_key(rq)         (blk_rq_pos(rq) + blk_rq_sectors(rq))
  53
  54 /*
  55  * Query io scheduler to see if the current process issuing bio may be
  56  * merged with rq.
  57  */
  58 static int elv_iosched_allow_bio_merge(struct request *rq, struct bio *bio)
  59 {
  60         struct request_queue *q = rq->q;
  61         struct elevator_queue *e = q->elevator;
  62
  63         if (e->uses_mq && e->type->ops.mq.allow_merge)
  64                 return e->type->ops.mq.allow_merge(q, rq, bio);
  65         else if (!e->uses_mq && e->type->ops.sq.elevator_allow_bio_merge_fn)
  66                 return e->type->ops.sq.elevator_allow_bio_merge_fn(q, rq, bio);
  67
  68         return 1;
  69 }
  70
  71 /*
  72  * can we safely merge with this request?
  73  */
  74 bool elv_bio_merge_ok(struct request *rq, struct bio *bio)
  75 {
  76         if (!blk_rq_merge_ok(rq, bio))
  77                 return false;
  78
  79         if (!elv_iosched_allow_bio_merge(rq, bio))
  80                 return false;
  81
  82         return true;
  83 }
  84 EXPORT_SYMBOL(elv_bio_merge_ok);
  85
  86 static bool elevator_match(const struct elevator_type *e, const char *name)
  87 {
  88         if (!strcmp(e->elevator_name, name))
  89                 return true;
  90         if (e->elevator_alias && !strcmp(e->elevator_alias, name))
  91                 return true;
  92
  93         return false;
  94 }
  95
  96 /*
  97  * Return scheduler with name 'name' and with matching 'mq capability
  98  */
  99 static struct elevator_type *elevator_find(const char *name, bool mq)
 100 {
 101         struct elevator_type *e;
 102
 103         list_for_each_entry(e, &elv_list, list) {
 104                 if (elevator_match(e, name) && (mq == e->uses_mq))
 105                         return e;
 106         }
 107
 108         return NULL;
 109 }
 110
 111 static void elevator_put(struct elevator_type *e)
 112 {
 113         module_put(e->elevator_owner);
 114 }
 115
 116 static struct elevator_type *elevator_get(struct request_queue *q,
 117                                           const char *name, bool try_loading)
 118 {
 119         struct elevator_type *e;
 120
 121         spin_lock(&elv_list_lock);
 122
 123         e = elevator_find(name, q->mq_ops != NULL);
 124         if (!e && try_loading) {
 125                 spin_unlock(&elv_list_lock);
 126                 request_module("%s-iosched", name);
 127                 spin_lock(&elv_list_lock);
 128                 e = elevator_find(name, q->mq_ops != NULL);
 129         }
 130
 131         if (e && !try_module_get(e->elevator_owner))
 132                 e = NULL;
 133
 134         spin_unlock(&elv_list_lock);
 135         return e;
 136 }
 137
 138 static char chosen_elevator[ELV_NAME_MAX];
 139
 140 static int __init elevator_setup(char *str)
 141 {
 142         /*
 143          * Be backwards-compatible with previous kernels, so users
 144          * won't get the wrong elevator.
 145          */
 146         strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 147         return 1;
 148 }
 149
 150 __setup("elevator=", elevator_setup);
 151
 152 /* called during boot to load the elevator chosen by the elevator param */
 153 void __init load_default_elevator_module(void)
 154 {
 155         struct elevator_type *e;
 156
 157         if (!chosen_elevator[0])
 158                 return;
 159
 160         /*
 161          * Boot parameter is deprecated, we haven't supported that for MQ.
 162          * Only look for non-mq schedulers from here.
 163          */
 164         spin_lock(&elv_list_lock);
 165         e = elevator_find(chosen_elevator, false);
 166         spin_unlock(&elv_list_lock);
 167
 168         if (!e)
 169                 request_module("%s-iosched", chosen_elevator);
 170 }
 171
 172 static struct kobj_type elv_ktype;
 173
 174 struct elevator_queue *elevator_alloc(struct request_queue *q,
 175                                   struct elevator_type *e)
 176 {
 177         struct elevator_queue *eq;
 178
 179         eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node);
 180         if (unlikely(!eq))
 181                 return NULL;
 182
 183         eq->type = e;
 184         kobject_init(&eq->kobj, &elv_ktype);
 185         mutex_init(&eq->sysfs_lock);
 186         hash_init(eq->hash);
 187         eq->uses_mq = e->uses_mq;
 188
 189         return eq;
 190 }
 191 EXPORT_SYMBOL(elevator_alloc);
 192
 193 static void elevator_release(struct kobject *kobj)
 194 {
 195         struct elevator_queue *e;
 196
 197         e = container_of(kobj, struct elevator_queue, kobj);
 198         elevator_put(e->type);
 199         kfree(e);
 200 }
 201
 202 int elevator_init(struct request_queue *q, char *name)
 203 {
 204         struct elevator_type *e = NULL;
 205         int err;
 206
 207         /*
 208          * q->sysfs_lock must be held to provide mutual exclusion between
 209          * elevator_switch() and here.
 210          */
 211         lockdep_assert_held(&q->sysfs_lock);
 212
 213         if (unlikely(q->elevator))
 214                 return 0;
 215
 216         INIT_LIST_HEAD(&q->queue_head);
 217         q->last_merge = NULL;
 218         q->end_sector = 0;
 219         q->boundary_rq = NULL;
 220
 221         if (name) {
 222                 e = elevator_get(q, name, true);
 223                 if (!e)
 224                         return -EINVAL;
 225         }
 226
 227         /*
 228          * Use the default elevator specified by config boot param for
 229          * non-mq devices, or by config option. Don't try to load modules
 230          * as we could be running off async and request_module() isn't
 231          * allowed from async.
 232          */
 233         if (!e && !q->mq_ops && *chosen_elevator) {
 234                 e = elevator_get(q, chosen_elevator, false);
 235                 if (!e)
 236                         printk(KERN_ERR "I/O scheduler %s not found\n",
 237                                                         chosen_elevator);
 238         }
 239
 240         if (!e) {
 241                 /*
 242                  * For blk-mq devices, we default to using mq-deadline,
 243                  * if available, for single queue devices. If deadline
 244                  * isn't available OR we have multiple queues, default
 245                  * to "none".
 246                  */
 247                 if (q->mq_ops) {
 248                         if (q->nr_hw_queues == 1)
 249                                 e = elevator_get(q, "mq-deadline", false);
 250                         if (!e)
 251                                 return 0;
 252                 } else
 253                         e = elevator_get(q, CONFIG_DEFAULT_IOSCHED, false);
 254
 255                 if (!e) {
 256                         printk(KERN_ERR
 257                                 "Default I/O scheduler not found. " \
 258                                 "Using noop.\n");
 259                         e = elevator_get(q, "noop", false);
 260                 }
 261         }
 262
 263         if (e->uses_mq)
 264                 err = blk_mq_init_sched(q, e);
 265         else
 266                 err = e->ops.sq.elevator_init_fn(q, e);
 267         if (err)
 268                 elevator_put(e);
 269         return err;
 270 }
 271 EXPORT_SYMBOL(elevator_init);
 272
 273 void elevator_exit(struct request_queue *q, struct elevator_queue *e)
 274 {
 275         mutex_lock(&e->sysfs_lock);
 276         if (e->uses_mq && e->type->ops.mq.exit_sched)
 277                 blk_mq_exit_sched(q, e);
 278         else if (!e->uses_mq && e->type->ops.sq.elevator_exit_fn)
 279                 e->type->ops.sq.elevator_exit_fn(e);
 280         mutex_unlock(&e->sysfs_lock);
 281
 282         kobject_put(&e->kobj);
 283 }
 284 EXPORT_SYMBOL(elevator_exit);
 285
 286 static inline void __elv_rqhash_del(struct request *rq)
 287 {
 288         hash_del(&rq->hash);
 289         rq->rq_flags &= ~RQF_HASHED;
 290 }
 291
 292 void elv_rqhash_del(struct request_queue *q, struct request *rq)
 293 {
 294         if (ELV_ON_HASH(rq))
 295                 __elv_rqhash_del(rq);
 296 }
 297 EXPORT_SYMBOL_GPL(elv_rqhash_del);
 298
 299 void elv_rqhash_add(struct request_queue *q, struct request *rq)
 300 {
 301         struct elevator_queue *e = q->elevator;
 302
 303         BUG_ON(ELV_ON_HASH(rq));
 304         hash_add(e->hash, &rq->hash, rq_hash_key(rq));
 305         rq->rq_flags |= RQF_HASHED;
 306 }
 307 EXPORT_SYMBOL_GPL(elv_rqhash_add);
 308
 309 void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 310 {
 311         __elv_rqhash_del(rq);
 312         elv_rqhash_add(q, rq);
 313 }
 314
 315 struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 316 {
 317         struct elevator_queue *e = q->elevator;
 318         struct hlist_node *next;
 319         struct request *rq;
 320
 321         hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
 322                 BUG_ON(!ELV_ON_HASH(rq));
 323
 324                 if (unlikely(!rq_mergeable(rq))) {
 325                         __elv_rqhash_del(rq);
 326                         continue;
 327                 }
 328
 329                 if (rq_hash_key(rq) == offset)
 330                         return rq;
 331         }
 332
 333         return NULL;
 334 }
 335
 336 /*
 337  * RB-tree support functions for inserting/lookup/removal of requests
 338  * in a sorted RB tree.
 339  */
 340 void elv_rb_add(struct rb_root *root, struct request *rq)
 341 {
 342         struct rb_node **p = &root->rb_node;
 343         struct rb_node *parent = NULL;
 344         struct request *__rq;
 345
 346         while (*p) {
 347                 parent = *p;
 348                 __rq = rb_entry(parent, struct request, rb_node);
 349
 350                 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
 351                         p = &(*p)->rb_left;
 352                 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
 353                         p = &(*p)->rb_right;
 354         }
 355
 356         rb_link_node(&rq->rb_node, parent, p);
 357         rb_insert_color(&rq->rb_node, root);
 358 }
 359 EXPORT_SYMBOL(elv_rb_add);
 360
 361 void elv_rb_del(struct rb_root *root, struct request *rq)
 362 {
 363         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 364         rb_erase(&rq->rb_node, root);
 365         RB_CLEAR_NODE(&rq->rb_node);
 366 }
 367 EXPORT_SYMBOL(elv_rb_del);
 368
 369 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 370 {
 371         struct rb_node *n = root->rb_node;
 372         struct request *rq;
 373
 374         while (n) {
 375                 rq = rb_entry(n, struct request, rb_node);
 376
 377                 if (sector < blk_rq_pos(rq))
 378                         n = n->rb_left;
 379                 else if (sector > blk_rq_pos(rq))
 380                         n = n->rb_right;
 381                 else
 382                         return rq;
 383         }
 384
 385         return NULL;
 386 }
 387 EXPORT_SYMBOL(elv_rb_find);
 388
 389 /*
 390  * Insert rq into dispatch queue of q.  Queue lock must be held on
 391  * entry.  rq is sort instead into the dispatch queue. To be used by
 392  * specific elevators.
 393  */
 394 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 395 {
 396         sector_t boundary;
 397         struct list_head *entry;
 398
 399         if (q->last_merge == rq)
 400                 q->last_merge = NULL;
 401
 402         elv_rqhash_del(q, rq);
 403
 404         q->nr_sorted--;
 405
 406         boundary = q->end_sector;
 407         list_for_each_prev(entry, &q->queue_head) {
 408                 struct request *pos = list_entry_rq(entry);
 409
 410                 if (req_op(rq) != req_op(pos))
 411                         break;
 412                 if (rq_data_dir(rq) != rq_data_dir(pos))
 413                         break;
 414                 if (pos->rq_flags & (RQF_STARTED | RQF_SOFTBARRIER))
 415                         break;
 416                 if (blk_rq_pos(rq) >= boundary) {
 417                         if (blk_rq_pos(pos) < boundary)
 418                                 continue;
 419                 } else {
 420                         if (blk_rq_pos(pos) >= boundary)
 421                                 break;
 422                 }
 423                 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
 424                         break;
 425         }
 426
 427         list_add(&rq->queuelist, entry);
 428 }
 429 EXPORT_SYMBOL(elv_dispatch_sort);
 430
 431 /*
 432  * Insert rq into dispatch queue of q.  Queue lock must be held on
 433  * entry.  rq is added to the back of the dispatch queue. To be used by
 434  * specific elevators.
 435  */
 436 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 437 {
 438         if (q->last_merge == rq)
 439                 q->last_merge = NULL;
 440
 441         elv_rqhash_del(q, rq);
 442
 443         q->nr_sorted--;
 444
 445         q->end_sector = rq_end_sector(rq);
 446         q->boundary_rq = rq;
 447         list_add_tail(&rq->queuelist, &q->queue_head);
 448 }
 449 EXPORT_SYMBOL(elv_dispatch_add_tail);
 450
 451 enum elv_merge elv_merge(struct request_queue *q, struct request **req,
 452                 struct bio *bio)
 453 {
 454         struct elevator_queue *e = q->elevator;
 455         struct request *__rq;
 456
 457         /*
 458          * Levels of merges:
 459          *      nomerges:  No merges at all attempted
 460          *      noxmerges: Only simple one-hit cache try
 461          *      merges:    All merge tries attempted
 462          */
 463         if (blk_queue_nomerges(q) || !bio_mergeable(bio))
 464                 return ELEVATOR_NO_MERGE;
 465
 466         /*
 467          * First try one-hit cache.
 468          */
 469         if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) {
 470                 enum elv_merge ret = blk_try_merge(q->last_merge, bio);
 471
 472                 if (ret != ELEVATOR_NO_MERGE) {
 473                         *req = q->last_merge;
 474                         return ret;
 475                 }
 476         }
 477
 478         if (blk_queue_noxmerges(q))
 479                 return ELEVATOR_NO_MERGE;
 480
 481         /*
 482          * See if our hash lookup can find a potential backmerge.
 483          */
 484         __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
 485         if (__rq && elv_bio_merge_ok(__rq, bio)) {
 486                 *req = __rq;
 487                 return ELEVATOR_BACK_MERGE;
 488         }
 489
 490         if (e->uses_mq && e->type->ops.mq.request_merge)
 491                 return e->type->ops.mq.request_merge(q, req, bio);
 492         else if (!e->uses_mq && e->type->ops.sq.elevator_merge_fn)
 493                 return e->type->ops.sq.elevator_merge_fn(q, req, bio);
 494
 495         return ELEVATOR_NO_MERGE;
 496 }
 497
 498 /*
 499  * Attempt to do an insertion back merge. Only check for the case where
 500  * we can append 'rq' to an existing request, so we can throw 'rq' away
 501  * afterwards.
 502  *
 503  * Returns true if we merged, false otherwise
 504  */
 505 bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq)
 506 {
 507         struct request *__rq;
 508         bool ret;
 509
 510         if (blk_queue_nomerges(q))
 511                 return false;
 512
 513         /*
 514          * First try one-hit cache.
 515          */
 516         if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
 517                 return true;
 518
 519         if (blk_queue_noxmerges(q))
 520                 return false;
 521
 522         ret = false;
 523         /*
 524          * See if our hash lookup can find a potential backmerge.
 525          */
 526         while (1) {
 527                 __rq = elv_rqhash_find(q, blk_rq_pos(rq));
 528                 if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
 529                         break;
 530
 531                 /* The merged request could be merged with others, try again */
 532                 ret = true;
 533                 rq = __rq;
 534         }
 535
 536         return ret;
 537 }
 538
 539 void elv_merged_request(struct request_queue *q, struct request *rq,
 540                 enum elv_merge type)
 541 {
 542         struct elevator_queue *e = q->elevator;
 543
 544         if (e->uses_mq && e->type->ops.mq.request_merged)
 545                 e->type->ops.mq.request_merged(q, rq, type);
 546         else if (!e->uses_mq && e->type->ops.sq.elevator_merged_fn)
 547                 e->type->ops.sq.elevator_merged_fn(q, rq, type);
 548
 549         if (type == ELEVATOR_BACK_MERGE)
 550                 elv_rqhash_reposition(q, rq);
 551
 552         q->last_merge = rq;
 553 }
 554
 555 void elv_merge_requests(struct request_queue *q, struct request *rq,
 556                              struct request *next)
 557 {
 558         struct elevator_queue *e = q->elevator;
 559         bool next_sorted = false;
 560
 561         if (e->uses_mq && e->type->ops.mq.requests_merged)
 562                 e->type->ops.mq.requests_merged(q, rq, next);
 563         else if (e->type->ops.sq.elevator_merge_req_fn) {
 564                 next_sorted = (__force bool)(next->rq_flags & RQF_SORTED);
 565                 if (next_sorted)
 566                         e->type->ops.sq.elevator_merge_req_fn(q, rq, next);
 567         }
 568
 569         elv_rqhash_reposition(q, rq);
 570
 571         if (next_sorted) {
 572                 elv_rqhash_del(q, next);
 573                 q->nr_sorted--;
 574         }
 575
 576         q->last_merge = rq;
 577 }
 578
 579 void elv_bio_merged(struct request_queue *q, struct request *rq,
 580                         struct bio *bio)
 581 {
 582         struct elevator_queue *e = q->elevator;
 583
 584         if (WARN_ON_ONCE(e->uses_mq))
 585                 return;
 586
 587         if (e->type->ops.sq.elevator_bio_merged_fn)
 588                 e->type->ops.sq.elevator_bio_merged_fn(q, rq, bio);
 589 }
 590
 591 #ifdef CONFIG_PM
 592 static void blk_pm_requeue_request(struct request *rq)
 593 {
 594         if (rq->q->dev && !(rq->rq_flags & RQF_PM))
 595                 rq->q->nr_pending--;
 596 }
 597
 598 static void blk_pm_add_request(struct request_queue *q, struct request *rq)
 599 {
 600         if (q->dev && !(rq->rq_flags & RQF_PM) && q->nr_pending++ == 0 &&
 601             (q->rpm_status == RPM_SUSPENDED || q->rpm_status == RPM_SUSPENDING))
 602                 pm_request_resume(q->dev);
 603 }
 604 #else
 605 static inline void blk_pm_requeue_request(struct request *rq) {}
 606 static inline void blk_pm_add_request(struct request_queue *q,
 607                                       struct request *rq)
 608 {
 609 }
 610 #endif
 611
 612 void elv_requeue_request(struct request_queue *q, struct request *rq)
 613 {
 614         /*
 615          * it already went through dequeue, we need to decrement the
 616          * in_flight count again
 617          */
 618         if (blk_account_rq(rq)) {
 619                 q->in_flight[rq_is_sync(rq)]--;
 620                 if (rq->rq_flags & RQF_SORTED)
 621                         elv_deactivate_rq(q, rq);
 622         }
 623
 624         rq->rq_flags &= ~RQF_STARTED;
 625
 626         blk_pm_requeue_request(rq);
 627
 628         __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
 629 }
 630
 631 void elv_drain_elevator(struct request_queue *q)
 632 {
 633         struct elevator_queue *e = q->elevator;
 634         static int printed;
 635
 636         if (WARN_ON_ONCE(e->uses_mq))
 637                 return;
 638
 639         lockdep_assert_held(q->queue_lock);
 640
 641         while (e->type->ops.sq.elevator_dispatch_fn(q, 1))
 642                 ;
 643         if (q->nr_sorted && printed++ < 10) {
 644                 printk(KERN_ERR "%s: forced dispatching is broken "
 645                        "(nr_sorted=%u), please report this\n",
 646                        q->elevator->type->elevator_name, q->nr_sorted);
 647         }
 648 }
 649
 650 void __elv_add_request(struct request_queue *q, struct request *rq, int where)
 651 {
 652         trace_block_rq_insert(q, rq);
 653
 654         blk_pm_add_request(q, rq);
 655
 656         rq->q = q;
 657
 658         if (rq->rq_flags & RQF_SOFTBARRIER) {
 659                 /* barriers are scheduling boundary, update end_sector */
 660                 if (!blk_rq_is_passthrough(rq)) {
 661                         q->end_sector = rq_end_sector(rq);
 662                         q->boundary_rq = rq;
 663                 }
 664         } else if (!(rq->rq_flags & RQF_ELVPRIV) &&
 665                     (where == ELEVATOR_INSERT_SORT ||
 666                      where == ELEVATOR_INSERT_SORT_MERGE))
 667                 where = ELEVATOR_INSERT_BACK;
 668
 669         switch (where) {
 670         case ELEVATOR_INSERT_REQUEUE:
 671         case ELEVATOR_INSERT_FRONT:
 672                 rq->rq_flags |= RQF_SOFTBARRIER;
 673                 list_add(&rq->queuelist, &q->queue_head);
 674                 break;
 675
 676         case ELEVATOR_INSERT_BACK:
 677                 rq->rq_flags |= RQF_SOFTBARRIER;
 678                 elv_drain_elevator(q);
 679                 list_add_tail(&rq->queuelist, &q->queue_head);
 680                 /*
 681                  * We kick the queue here for the following reasons.
 682                  * - The elevator might have returned NULL previously
 683                  *   to delay requests and returned them now.  As the
 684                  *   queue wasn't empty before this request, ll_rw_blk
 685                  *   won't run the queue on return, resulting in hang.
 686                  * - Usually, back inserted requests won't be merged
 687                  *   with anything.  There's no point in delaying queue
 688                  *   processing.
 689                  */
 690                 __blk_run_queue(q);
 691                 break;
 692
 693         case ELEVATOR_INSERT_SORT_MERGE:
 694                 /*
 695                  * If we succeed in merging this request with one in the
 696                  * queue already, we are done - rq has now been freed,
 697                  * so no need to do anything further.
 698                  */
 699                 if (elv_attempt_insert_merge(q, rq))
 700                         break;
 701                 /* fall through */
 702         case ELEVATOR_INSERT_SORT:
 703                 BUG_ON(blk_rq_is_passthrough(rq));
 704                 rq->rq_flags |= RQF_SORTED;
 705                 q->nr_sorted++;
 706                 if (rq_mergeable(rq)) {
 707                         elv_rqhash_add(q, rq);
 708                         if (!q->last_merge)
 709                                 q->last_merge = rq;
 710                 }
 711
 712                 /*
 713                  * Some ioscheds (cfq) run q->request_fn directly, so
 714                  * rq cannot be accessed after calling
 715                  * elevator_add_req_fn.
 716                  */
 717                 q->elevator->type->ops.sq.elevator_add_req_fn(q, rq);
 718                 break;
 719
 720         case ELEVATOR_INSERT_FLUSH:
 721                 rq->rq_flags |= RQF_SOFTBARRIER;
 722                 blk_insert_flush(rq);
 723                 break;
 724         default:
 725                 printk(KERN_ERR "%s: bad insertion point %d\n",
 726                        __func__, where);
 727                 BUG();
 728         }
 729 }
 730 EXPORT_SYMBOL(__elv_add_request);
 731
 732 void elv_add_request(struct request_queue *q, struct request *rq, int where)
 733 {
 734         unsigned long flags;
 735
 736         spin_lock_irqsave(q->queue_lock, flags);
 737         __elv_add_request(q, rq, where);
 738         spin_unlock_irqrestore(q->queue_lock, flags);
 739 }
 740 EXPORT_SYMBOL(elv_add_request);
 741
 742 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 743 {
 744         struct elevator_queue *e = q->elevator;
 745
 746         if (e->uses_mq && e->type->ops.mq.next_request)
 747                 return e->type->ops.mq.next_request(q, rq);
 748         else if (!e->uses_mq && e->type->ops.sq.elevator_latter_req_fn)
 749                 return e->type->ops.sq.elevator_latter_req_fn(q, rq);
 750
 751         return NULL;
 752 }
 753
 754 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 755 {
 756         struct elevator_queue *e = q->elevator;
 757
 758         if (e->uses_mq && e->type->ops.mq.former_request)
 759                 return e->type->ops.mq.former_request(q, rq);
 760         if (!e->uses_mq && e->type->ops.sq.elevator_former_req_fn)
 761                 return e->type->ops.sq.elevator_former_req_fn(q, rq);
 762         return NULL;
 763 }
 764
 765 int elv_set_request(struct request_queue *q, struct request *rq,
 766                     struct bio *bio, gfp_t gfp_mask)
 767 {
 768         struct elevator_queue *e = q->elevator;
 769
 770         if (WARN_ON_ONCE(e->uses_mq))
 771                 return 0;
 772
 773         if (e->type->ops.sq.elevator_set_req_fn)
 774                 return e->type->ops.sq.elevator_set_req_fn(q, rq, bio, gfp_mask);
 775         return 0;
 776 }
 777
 778 void elv_put_request(struct request_queue *q, struct request *rq)
 779 {
 780         struct elevator_queue *e = q->elevator;
 781
 782         if (WARN_ON_ONCE(e->uses_mq))
 783                 return;
 784
 785         if (e->type->ops.sq.elevator_put_req_fn)
 786                 e->type->ops.sq.elevator_put_req_fn(rq);
 787 }
 788
 789 int elv_may_queue(struct request_queue *q, unsigned int op)
 790 {
 791         struct elevator_queue *e = q->elevator;
 792
 793         if (WARN_ON_ONCE(e->uses_mq))
 794                 return 0;
 795
 796         if (e->type->ops.sq.elevator_may_queue_fn)
 797                 return e->type->ops.sq.elevator_may_queue_fn(q, op);
 798
 799         return ELV_MQUEUE_MAY;
 800 }
 801
 802 void elv_completed_request(struct request_queue *q, struct request *rq)
 803 {
 804         struct elevator_queue *e = q->elevator;
 805
 806         if (WARN_ON_ONCE(e->uses_mq))
 807                 return;
 808
 809         /*
 810          * request is released from the driver, io must be done
 811          */
 812         if (blk_account_rq(rq)) {
 813                 q->in_flight[rq_is_sync(rq)]--;
 814                 if ((rq->rq_flags & RQF_SORTED) &&
 815                     e->type->ops.sq.elevator_completed_req_fn)
 816                         e->type->ops.sq.elevator_completed_req_fn(q, rq);
 817         }
 818 }
 819
 820 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 821
 822 static ssize_t
 823 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 824 {
 825         struct elv_fs_entry *entry = to_elv(attr);
 826         struct elevator_queue *e;
 827         ssize_t error;
 828
 829         if (!entry->show)
 830                 return -EIO;
 831
 832         e = container_of(kobj, struct elevator_queue, kobj);
 833         mutex_lock(&e->sysfs_lock);
 834         error = e->type ? entry->show(e, page) : -ENOENT;
 835         mutex_unlock(&e->sysfs_lock);
 836         return error;
 837 }
 838
 839 static ssize_t
 840 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 841                const char *page, size_t length)
 842 {
 843         struct elv_fs_entry *entry = to_elv(attr);
 844         struct elevator_queue *e;
 845         ssize_t error;
 846
 847         if (!entry->store)
 848                 return -EIO;
 849
 850         e = container_of(kobj, struct elevator_queue, kobj);
 851         mutex_lock(&e->sysfs_lock);
 852         error = e->type ? entry->store(e, page, length) : -ENOENT;
 853         mutex_unlock(&e->sysfs_lock);
 854         return error;
 855 }
 856
 857 static const struct sysfs_ops elv_sysfs_ops = {
 858         .show   = elv_attr_show,
 859         .store  = elv_attr_store,
 860 };
 861
 862 static struct kobj_type elv_ktype = {
 863         .sysfs_ops      = &elv_sysfs_ops,
 864         .release        = elevator_release,
 865 };
 866
 867 int elv_register_queue(struct request_queue *q)
 868 {
 869         struct elevator_queue *e = q->elevator;
 870         int error;
 871
 872         lockdep_assert_held(&q->sysfs_lock);
 873
 874         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 875         if (!error) {
 876                 struct elv_fs_entry *attr = e->type->elevator_attrs;
 877                 if (attr) {
 878                         while (attr->attr.name) {
 879                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 880                                         break;
 881                                 attr++;
 882                         }
 883                 }
 884                 kobject_uevent(&e->kobj, KOBJ_ADD);
 885                 e->registered = 1;
 886                 if (!e->uses_mq && e->type->ops.sq.elevator_registered_fn)
 887                         e->type->ops.sq.elevator_registered_fn(q);
 888         }
 889         return error;
 890 }
 891
 892 void elv_unregister_queue(struct request_queue *q)
 893 {
 894         lockdep_assert_held(&q->sysfs_lock);
 895
 896         if (q) {
 897                 struct elevator_queue *e = q->elevator;
 898
 899                 kobject_uevent(&e->kobj, KOBJ_REMOVE);
 900                 kobject_del(&e->kobj);
 901                 e->registered = 0;
 902                 /* Re-enable throttling in case elevator disabled it */
 903                 wbt_enable_default(q);
 904         }
 905 }
 906
 907 int elv_register(struct elevator_type *e)
 908 {
 909         char *def = "";
 910
 911         /* create icq_cache if requested */
 912         if (e->icq_size) {
 913                 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
 914                     WARN_ON(e->icq_align < __alignof__(struct io_cq)))
 915                         return -EINVAL;
 916
 917                 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
 918                          "%s_io_cq", e->elevator_name);
 919                 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
 920                                                  e->icq_align, 0, NULL);
 921                 if (!e->icq_cache)
 922                         return -ENOMEM;
 923         }
 924
 925         /* register, don't allow duplicate names */
 926         spin_lock(&elv_list_lock);
 927         if (elevator_find(e->elevator_name, e->uses_mq)) {
 928                 spin_unlock(&elv_list_lock);
 929                 if (e->icq_cache)
 930                         kmem_cache_destroy(e->icq_cache);
 931                 return -EBUSY;
 932         }
 933         list_add_tail(&e->list, &elv_list);
 934         spin_unlock(&elv_list_lock);
 935
 936         /* print pretty message */
 937         if (elevator_match(e, chosen_elevator) ||
 938                         (!*chosen_elevator &&
 939                          elevator_match(e, CONFIG_DEFAULT_IOSCHED)))
 940                                 def = " (default)";
 941
 942         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
 943                                                                 def);
 944         return 0;
 945 }
 946 EXPORT_SYMBOL_GPL(elv_register);
 947
 948 void elv_unregister(struct elevator_type *e)
 949 {
 950         /* unregister */
 951         spin_lock(&elv_list_lock);
 952         list_del_init(&e->list);
 953         spin_unlock(&elv_list_lock);
 954
 955         /*
 956          * Destroy icq_cache if it exists.  icq's are RCU managed.  Make
 957          * sure all RCU operations are complete before proceeding.
 958          */
 959         if (e->icq_cache) {
 960                 rcu_barrier();
 961                 kmem_cache_destroy(e->icq_cache);
 962                 e->icq_cache = NULL;
 963         }
 964 }
 965 EXPORT_SYMBOL_GPL(elv_unregister);
 966
 967 static int elevator_switch_mq(struct request_queue *q,
 968                               struct elevator_type *new_e)
 969 {
 970         int ret;
 971
 972         lockdep_assert_held(&q->sysfs_lock);
 973
 974         blk_mq_freeze_queue(q);
 975         blk_mq_quiesce_queue(q);
 976
 977         if (q->elevator) {
 978                 if (q->elevator->registered)
 979                         elv_unregister_queue(q);
 980                 ioc_clear_queue(q);
 981                 elevator_exit(q, q->elevator);
 982         }
 983
 984         ret = blk_mq_init_sched(q, new_e);
 985         if (ret)
 986                 goto out;
 987
 988         if (new_e) {
 989                 ret = elv_register_queue(q);
 990                 if (ret) {
 991                         elevator_exit(q, q->elevator);
 992                         goto out;
 993                 }
 994         }
 995
 996         if (new_e)
 997                 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
 998         else
 999                 blk_add_trace_msg(q, "elv switch: none");
1000
1001 out:
1002         blk_mq_unquiesce_queue(q);
1003         blk_mq_unfreeze_queue(q);
1004         return ret;
1005 }
1006
1007 /*
1008  * switch to new_e io scheduler. be careful not to introduce deadlocks -
1009  * we don't free the old io scheduler, before we have allocated what we
1010  * need for the new one. this way we have a chance of going back to the old
1011  * one, if the new one fails init for some reason.
1012  */
1013 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1014 {
1015         struct elevator_queue *old = q->elevator;
1016         bool old_registered = false;
1017         int err;
1018
1019         lockdep_assert_held(&q->sysfs_lock);
1020
1021         if (q->mq_ops)
1022                 return elevator_switch_mq(q, new_e);
1023
1024         /*
1025          * Turn on BYPASS and drain all requests w/ elevator private data.
1026          * Block layer doesn't call into a quiesced elevator - all requests
1027          * are directly put on the dispatch list without elevator data
1028          * using INSERT_BACK.  All requests have SOFTBARRIER set and no
1029          * merge happens either.
1030          */
1031         if (old) {
1032                 old_registered = old->registered;
1033
1034                 blk_queue_bypass_start(q);
1035
1036                 /* unregister and clear all auxiliary data of the old elevator */
1037                 if (old_registered)
1038                         elv_unregister_queue(q);
1039
1040                 ioc_clear_queue(q);
1041         }
1042
1043         /* allocate, init and register new elevator */
1044         err = new_e->ops.sq.elevator_init_fn(q, new_e);
1045         if (err)
1046                 goto fail_init;
1047
1048         err = elv_register_queue(q);
1049         if (err)
1050                 goto fail_register;
1051
1052         /* done, kill the old one and finish */
1053         if (old) {
1054                 elevator_exit(q, old);
1055                 blk_queue_bypass_end(q);
1056         }
1057
1058         blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
1059
1060         return 0;
1061
1062 fail_register:
1063         elevator_exit(q, q->elevator);
1064 fail_init:
1065         /* switch failed, restore and re-register old elevator */
1066         if (old) {
1067                 q->elevator = old;
1068                 elv_register_queue(q);
1069                 blk_queue_bypass_end(q);
1070         }
1071
1072         return err;
1073 }
1074
1075 /*
1076  * Switch this queue to the given IO scheduler.
1077  */
1078 static int __elevator_change(struct request_queue *q, const char *name)
1079 {
1080         char elevator_name[ELV_NAME_MAX];
1081         struct elevator_type *e;
1082
1083         /* Make sure queue is not in the middle of being removed */
1084         if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags))
1085                 return -ENOENT;
1086
1087         /*
1088          * Special case for mq, turn off scheduling
1089          */
1090         if (q->mq_ops && !strncmp(name, "none", 4))
1091                 return elevator_switch(q, NULL);
1092
1093         strlcpy(elevator_name, name, sizeof(elevator_name));
1094         e = elevator_get(q, strstrip(elevator_name), true);
1095         if (!e)
1096                 return -EINVAL;
1097
1098         if (q->elevator && elevator_match(q->elevator->type, elevator_name)) {
1099                 elevator_put(e);
1100                 return 0;
1101         }
1102
1103         return elevator_switch(q, e);
1104 }
1105
1106 static inline bool elv_support_iosched(struct request_queue *q)
1107 {
1108         if (q->mq_ops && q->tag_set && (q->tag_set->flags &
1109                                 BLK_MQ_F_NO_SCHED))
1110                 return false;
1111         return true;
1112 }
1113
1114 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1115                           size_t count)
1116 {
1117         int ret;
1118
1119         if (!(q->mq_ops || q->request_fn) || !elv_support_iosched(q))
1120                 return count;
1121
1122         ret = __elevator_change(q, name);
1123         if (!ret)
1124                 return count;
1125
1126         return ret;
1127 }
1128
1129 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1130 {
1131         struct elevator_queue *e = q->elevator;
1132         struct elevator_type *elv = NULL;
1133         struct elevator_type *__e;
1134         bool uses_mq = q->mq_ops != NULL;
1135         int len = 0;
1136
1137         if (!queue_is_rq_based(q))
1138                 return sprintf(name, "none\n");
1139
1140         if (!q->elevator)
1141                 len += sprintf(name+len, "[none] ");
1142         else
1143                 elv = e->type;
1144
1145         spin_lock(&elv_list_lock);
1146         list_for_each_entry(__e, &elv_list, list) {
1147                 if (elv && elevator_match(elv, __e->elevator_name) &&
1148                     (__e->uses_mq == uses_mq)) {
1149                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1150                         continue;
1151                 }
1152                 if (__e->uses_mq && q->mq_ops && elv_support_iosched(q))
1153                         len += sprintf(name+len, "%s ", __e->elevator_name);
1154                 else if (!__e->uses_mq && !q->mq_ops)
1155                         len += sprintf(name+len, "%s ", __e->elevator_name);
1156         }
1157         spin_unlock(&elv_list_lock);
1158
1159         if (q->mq_ops && q->elevator)
1160                 len += sprintf(name+len, "none");
1161
1162         len += sprintf(len+name, "\n");
1163         return len;
1164 }
1165
1166 struct request *elv_rb_former_request(struct request_queue *q,
1167                                       struct request *rq)
1168 {
1169         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1170
1171         if (rbprev)
1172                 return rb_entry_rq(rbprev);
1173
1174         return NULL;
1175 }
1176 EXPORT_SYMBOL(elv_rb_former_request);
1177
1178 struct request *elv_rb_latter_request(struct request_queue *q,
1179                                       struct request *rq)
1180 {
1181         struct rb_node *rbnext = rb_next(&rq->rb_node);
1182
1183         if (rbnext)
1184                 return rb_entry_rq(rbnext);
1185
1186         return NULL;
1187 }
1188 EXPORT_SYMBOL(elv_rb_latter_request);