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