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