block/elevator.c

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