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