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
  38 #include <trace/events/block.h>
  39
  40 #include "blk.h"
  41
  42 static DEFINE_SPINLOCK(elv_list_lock);
  43 static LIST_HEAD(elv_list);
  44
  45 /*
  46  * Merge hash stuff.
  47  */
  48 static const int elv_hash_shift = 6;
  49 #define ELV_HASH_BLOCK(sec)     ((sec) >> 3)
  50 #define ELV_HASH_FN(sec)        \
  51                 (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
  52 #define ELV_HASH_ENTRIES        (1 << elv_hash_shift)
  53 #define rq_hash_key(rq)         (blk_rq_pos(rq) + blk_rq_sectors(rq))
  54
  55 /*
  56  * Query io scheduler to see if the current process issuing bio may be
  57  * merged with rq.
  58  */
  59 static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
  60 {
  61         struct request_queue *q = rq->q;
  62         struct elevator_queue *e = q->elevator;
  63
  64         if (e->type->ops.elevator_allow_merge_fn)
  65                 return e->type->ops.elevator_allow_merge_fn(q, rq, bio);
  66
  67         return 1;
  68 }
  69
  70 /*
  71  * can we safely merge with this request?
  72  */
  73 bool elv_rq_merge_ok(struct request *rq, struct bio *bio)
  74 {
  75         if (!blk_rq_merge_ok(rq, bio))
  76                 return 0;
  77
  78         if (!elv_iosched_allow_merge(rq, bio))
  79                 return 0;
  80
  81         return 1;
  82 }
  83 EXPORT_SYMBOL(elv_rq_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)
 103 {
 104         struct elevator_type *e;
 105
 106         spin_lock(&elv_list_lock);
 107
 108         e = elevator_find(name);
 109         if (!e) {
 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 static struct kobj_type elv_ktype;
 139
 140 static struct elevator_queue *elevator_alloc(struct request_queue *q,
 141                                   struct elevator_type *e)
 142 {
 143         struct elevator_queue *eq;
 144         int i;
 145
 146         eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node);
 147         if (unlikely(!eq))
 148                 goto err;
 149
 150         eq->type = e;
 151         kobject_init(&eq->kobj, &elv_ktype);
 152         mutex_init(&eq->sysfs_lock);
 153
 154         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 155                                         GFP_KERNEL, q->node);
 156         if (!eq->hash)
 157                 goto err;
 158
 159         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 160                 INIT_HLIST_HEAD(&eq->hash[i]);
 161
 162         return eq;
 163 err:
 164         kfree(eq);
 165         elevator_put(e);
 166         return NULL;
 167 }
 168
 169 static void elevator_release(struct kobject *kobj)
 170 {
 171         struct elevator_queue *e;
 172
 173         e = container_of(kobj, struct elevator_queue, kobj);
 174         elevator_put(e->type);
 175         kfree(e->hash);
 176         kfree(e);
 177 }
 178
 179 int elevator_init(struct request_queue *q, char *name)
 180 {
 181         struct elevator_type *e = NULL;
 182         int err;
 183
 184         if (unlikely(q->elevator))
 185                 return 0;
 186
 187         INIT_LIST_HEAD(&q->queue_head);
 188         q->last_merge = NULL;
 189         q->end_sector = 0;
 190         q->boundary_rq = NULL;
 191
 192         if (name) {
 193                 e = elevator_get(name);
 194                 if (!e)
 195                         return -EINVAL;
 196         }
 197
 198         if (!e && *chosen_elevator) {
 199                 e = elevator_get(chosen_elevator);
 200                 if (!e)
 201                         printk(KERN_ERR "I/O scheduler %s not found\n",
 202                                                         chosen_elevator);
 203         }
 204
 205         if (!e) {
 206                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
 207                 if (!e) {
 208                         printk(KERN_ERR
 209                                 "Default I/O scheduler not found. " \
 210                                 "Using noop.\n");
 211                         e = elevator_get("noop");
 212                 }
 213         }
 214
 215         q->elevator = elevator_alloc(q, e);
 216         if (!q->elevator)
 217                 return -ENOMEM;
 218
 219         err = e->ops.elevator_init_fn(q);
 220         if (err) {
 221                 kobject_put(&q->elevator->kobj);
 222                 return err;
 223         }
 224
 225         return 0;
 226 }
 227 EXPORT_SYMBOL(elevator_init);
 228
 229 void elevator_exit(struct elevator_queue *e)
 230 {
 231         mutex_lock(&e->sysfs_lock);
 232         if (e->type->ops.elevator_exit_fn)
 233                 e->type->ops.elevator_exit_fn(e);
 234         mutex_unlock(&e->sysfs_lock);
 235
 236         kobject_put(&e->kobj);
 237 }
 238 EXPORT_SYMBOL(elevator_exit);
 239
 240 static inline void __elv_rqhash_del(struct request *rq)
 241 {
 242         hlist_del_init(&rq->hash);
 243 }
 244
 245 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 246 {
 247         if (ELV_ON_HASH(rq))
 248                 __elv_rqhash_del(rq);
 249 }
 250
 251 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 252 {
 253         struct elevator_queue *e = q->elevator;
 254
 255         BUG_ON(ELV_ON_HASH(rq));
 256         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 257 }
 258
 259 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 260 {
 261         __elv_rqhash_del(rq);
 262         elv_rqhash_add(q, rq);
 263 }
 264
 265 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 266 {
 267         struct elevator_queue *e = q->elevator;
 268         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 269         struct hlist_node *entry, *next;
 270         struct request *rq;
 271
 272         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 273                 BUG_ON(!ELV_ON_HASH(rq));
 274
 275                 if (unlikely(!rq_mergeable(rq))) {
 276                         __elv_rqhash_del(rq);
 277                         continue;
 278                 }
 279
 280                 if (rq_hash_key(rq) == offset)
 281                         return rq;
 282         }
 283
 284         return NULL;
 285 }
 286
 287 /*
 288  * RB-tree support functions for inserting/lookup/removal of requests
 289  * in a sorted RB tree.
 290  */
 291 void elv_rb_add(struct rb_root *root, struct request *rq)
 292 {
 293         struct rb_node **p = &root->rb_node;
 294         struct rb_node *parent = NULL;
 295         struct request *__rq;
 296
 297         while (*p) {
 298                 parent = *p;
 299                 __rq = rb_entry(parent, struct request, rb_node);
 300
 301                 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
 302                         p = &(*p)->rb_left;
 303                 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
 304                         p = &(*p)->rb_right;
 305         }
 306
 307         rb_link_node(&rq->rb_node, parent, p);
 308         rb_insert_color(&rq->rb_node, root);
 309 }
 310 EXPORT_SYMBOL(elv_rb_add);
 311
 312 void elv_rb_del(struct rb_root *root, struct request *rq)
 313 {
 314         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 315         rb_erase(&rq->rb_node, root);
 316         RB_CLEAR_NODE(&rq->rb_node);
 317 }
 318 EXPORT_SYMBOL(elv_rb_del);
 319
 320 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 321 {
 322         struct rb_node *n = root->rb_node;
 323         struct request *rq;
 324
 325         while (n) {
 326                 rq = rb_entry(n, struct request, rb_node);
 327
 328                 if (sector < blk_rq_pos(rq))
 329                         n = n->rb_left;
 330                 else if (sector > blk_rq_pos(rq))
 331                         n = n->rb_right;
 332                 else
 333                         return rq;
 334         }
 335
 336         return NULL;
 337 }
 338 EXPORT_SYMBOL(elv_rb_find);
 339
 340 /*
 341  * Insert rq into dispatch queue of q.  Queue lock must be held on
 342  * entry.  rq is sort instead into the dispatch queue. To be used by
 343  * specific elevators.
 344  */
 345 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 346 {
 347         sector_t boundary;
 348         struct list_head *entry;
 349         int stop_flags;
 350
 351         if (q->last_merge == rq)
 352                 q->last_merge = NULL;
 353
 354         elv_rqhash_del(q, rq);
 355
 356         q->nr_sorted--;
 357
 358         boundary = q->end_sector;
 359         stop_flags = REQ_SOFTBARRIER | REQ_STARTED;
 360         list_for_each_prev(entry, &q->queue_head) {
 361                 struct request *pos = list_entry_rq(entry);
 362
 363                 if ((rq->cmd_flags & REQ_DISCARD) !=
 364                     (pos->cmd_flags & REQ_DISCARD))
 365                         break;
 366                 if (rq_data_dir(rq) != rq_data_dir(pos))
 367                         break;
 368                 if (pos->cmd_flags & stop_flags)
 369                         break;
 370                 if (blk_rq_pos(rq) >= boundary) {
 371                         if (blk_rq_pos(pos) < boundary)
 372                                 continue;
 373                 } else {
 374                         if (blk_rq_pos(pos) >= boundary)
 375                                 break;
 376                 }
 377                 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
 378                         break;
 379         }
 380
 381         list_add(&rq->queuelist, entry);
 382 }
 383 EXPORT_SYMBOL(elv_dispatch_sort);
 384
 385 /*
 386  * Insert rq into dispatch queue of q.  Queue lock must be held on
 387  * entry.  rq is added to the back of the dispatch queue. To be used by
 388  * specific elevators.
 389  */
 390 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 391 {
 392         if (q->last_merge == rq)
 393                 q->last_merge = NULL;
 394
 395         elv_rqhash_del(q, rq);
 396
 397         q->nr_sorted--;
 398
 399         q->end_sector = rq_end_sector(rq);
 400         q->boundary_rq = rq;
 401         list_add_tail(&rq->queuelist, &q->queue_head);
 402 }
 403 EXPORT_SYMBOL(elv_dispatch_add_tail);
 404
 405 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 406 {
 407         struct elevator_queue *e = q->elevator;
 408         struct request *__rq;
 409         int ret;
 410
 411         /*
 412          * Levels of merges:
 413          *      nomerges:  No merges at all attempted
 414          *      noxmerges: Only simple one-hit cache try
 415          *      merges:    All merge tries attempted
 416          */
 417         if (blk_queue_nomerges(q))
 418                 return ELEVATOR_NO_MERGE;
 419
 420         /*
 421          * First try one-hit cache.
 422          */
 423         if (q->last_merge && elv_rq_merge_ok(q->last_merge, bio)) {
 424                 ret = blk_try_merge(q->last_merge, bio);
 425                 if (ret != ELEVATOR_NO_MERGE) {
 426                         *req = q->last_merge;
 427                         return ret;
 428                 }
 429         }
 430
 431         if (blk_queue_noxmerges(q))
 432                 return ELEVATOR_NO_MERGE;
 433
 434         /*
 435          * See if our hash lookup can find a potential backmerge.
 436          */
 437         __rq = elv_rqhash_find(q, bio->bi_sector);
 438         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 439                 *req = __rq;
 440                 return ELEVATOR_BACK_MERGE;
 441         }
 442
 443         if (e->type->ops.elevator_merge_fn)
 444                 return e->type->ops.elevator_merge_fn(q, req, bio);
 445
 446         return ELEVATOR_NO_MERGE;
 447 }
 448
 449 /*
 450  * Attempt to do an insertion back merge. Only check for the case where
 451  * we can append 'rq' to an existing request, so we can throw 'rq' away
 452  * afterwards.
 453  *
 454  * Returns true if we merged, false otherwise
 455  */
 456 static bool elv_attempt_insert_merge(struct request_queue *q,
 457                                      struct request *rq)
 458 {
 459         struct request *__rq;
 460
 461         if (blk_queue_nomerges(q))
 462                 return false;
 463
 464         /*
 465          * First try one-hit cache.
 466          */
 467         if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
 468                 return true;
 469
 470         if (blk_queue_noxmerges(q))
 471                 return false;
 472
 473         /*
 474          * See if our hash lookup can find a potential backmerge.
 475          */
 476         __rq = elv_rqhash_find(q, blk_rq_pos(rq));
 477         if (__rq && blk_attempt_req_merge(q, __rq, rq))
 478                 return true;
 479
 480         return false;
 481 }
 482
 483 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 484 {
 485         struct elevator_queue *e = q->elevator;
 486
 487         if (e->type->ops.elevator_merged_fn)
 488                 e->type->ops.elevator_merged_fn(q, rq, type);
 489
 490         if (type == ELEVATOR_BACK_MERGE)
 491                 elv_rqhash_reposition(q, rq);
 492
 493         q->last_merge = rq;
 494 }
 495
 496 void elv_merge_requests(struct request_queue *q, struct request *rq,
 497                              struct request *next)
 498 {
 499         struct elevator_queue *e = q->elevator;
 500         const int next_sorted = next->cmd_flags & REQ_SORTED;
 501
 502         if (next_sorted && e->type->ops.elevator_merge_req_fn)
 503                 e->type->ops.elevator_merge_req_fn(q, rq, next);
 504
 505         elv_rqhash_reposition(q, rq);
 506
 507         if (next_sorted) {
 508                 elv_rqhash_del(q, next);
 509                 q->nr_sorted--;
 510         }
 511
 512         q->last_merge = rq;
 513 }
 514
 515 void elv_bio_merged(struct request_queue *q, struct request *rq,
 516                         struct bio *bio)
 517 {
 518         struct elevator_queue *e = q->elevator;
 519
 520         if (e->type->ops.elevator_bio_merged_fn)
 521                 e->type->ops.elevator_bio_merged_fn(q, rq, bio);
 522 }
 523
 524 void elv_requeue_request(struct request_queue *q, struct request *rq)
 525 {
 526         /*
 527          * it already went through dequeue, we need to decrement the
 528          * in_flight count again
 529          */
 530         if (blk_account_rq(rq)) {
 531                 q->in_flight[rq_is_sync(rq)]--;
 532                 if (rq->cmd_flags & REQ_SORTED)
 533                         elv_deactivate_rq(q, rq);
 534         }
 535
 536         rq->cmd_flags &= ~REQ_STARTED;
 537
 538         __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
 539 }
 540
 541 void elv_drain_elevator(struct request_queue *q)
 542 {
 543         static int printed;
 544
 545         lockdep_assert_held(q->queue_lock);
 546
 547         while (q->elevator->type->ops.elevator_dispatch_fn(q, 1))
 548                 ;
 549         if (q->nr_sorted && printed++ < 10) {
 550                 printk(KERN_ERR "%s: forced dispatching is broken "
 551                        "(nr_sorted=%u), please report this\n",
 552                        q->elevator->type->elevator_name, q->nr_sorted);
 553         }
 554 }
 555
 556 void elv_quiesce_start(struct request_queue *q)
 557 {
 558         if (!q->elevator)
 559                 return;
 560
 561         spin_lock_irq(q->queue_lock);
 562         queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
 563         spin_unlock_irq(q->queue_lock);
 564
 565         blk_drain_queue(q, false);
 566 }
 567
 568 void elv_quiesce_end(struct request_queue *q)
 569 {
 570         spin_lock_irq(q->queue_lock);
 571         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
 572         spin_unlock_irq(q->queue_lock);
 573 }
 574
 575 void __elv_add_request(struct request_queue *q, struct request *rq, int where)
 576 {
 577         trace_block_rq_insert(q, rq);
 578
 579         rq->q = q;
 580
 581         if (rq->cmd_flags & REQ_SOFTBARRIER) {
 582                 /* barriers are scheduling boundary, update end_sector */
 583                 if (rq->cmd_type == REQ_TYPE_FS ||
 584                     (rq->cmd_flags & REQ_DISCARD)) {
 585                         q->end_sector = rq_end_sector(rq);
 586                         q->boundary_rq = rq;
 587                 }
 588         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 589                     (where == ELEVATOR_INSERT_SORT ||
 590                      where == ELEVATOR_INSERT_SORT_MERGE))
 591                 where = ELEVATOR_INSERT_BACK;
 592
 593         switch (where) {
 594         case ELEVATOR_INSERT_REQUEUE:
 595         case ELEVATOR_INSERT_FRONT:
 596                 rq->cmd_flags |= REQ_SOFTBARRIER;
 597                 list_add(&rq->queuelist, &q->queue_head);
 598                 break;
 599
 600         case ELEVATOR_INSERT_BACK:
 601                 rq->cmd_flags |= REQ_SOFTBARRIER;
 602                 elv_drain_elevator(q);
 603                 list_add_tail(&rq->queuelist, &q->queue_head);
 604                 /*
 605                  * We kick the queue here for the following reasons.
 606                  * - The elevator might have returned NULL previously
 607                  *   to delay requests and returned them now.  As the
 608                  *   queue wasn't empty before this request, ll_rw_blk
 609                  *   won't run the queue on return, resulting in hang.
 610                  * - Usually, back inserted requests won't be merged
 611                  *   with anything.  There's no point in delaying queue
 612                  *   processing.
 613                  */
 614                 __blk_run_queue(q);
 615                 break;
 616
 617         case ELEVATOR_INSERT_SORT_MERGE:
 618                 /*
 619                  * If we succeed in merging this request with one in the
 620                  * queue already, we are done - rq has now been freed,
 621                  * so no need to do anything further.
 622                  */
 623                 if (elv_attempt_insert_merge(q, rq))
 624                         break;
 625         case ELEVATOR_INSERT_SORT:
 626                 BUG_ON(rq->cmd_type != REQ_TYPE_FS &&
 627                        !(rq->cmd_flags & REQ_DISCARD));
 628                 rq->cmd_flags |= REQ_SORTED;
 629                 q->nr_sorted++;
 630                 if (rq_mergeable(rq)) {
 631                         elv_rqhash_add(q, rq);
 632                         if (!q->last_merge)
 633                                 q->last_merge = rq;
 634                 }
 635
 636                 /*
 637                  * Some ioscheds (cfq) run q->request_fn directly, so
 638                  * rq cannot be accessed after calling
 639                  * elevator_add_req_fn.
 640                  */
 641                 q->elevator->type->ops.elevator_add_req_fn(q, rq);
 642                 break;
 643
 644         case ELEVATOR_INSERT_FLUSH:
 645                 rq->cmd_flags |= REQ_SOFTBARRIER;
 646                 blk_insert_flush(rq);
 647                 break;
 648         default:
 649                 printk(KERN_ERR "%s: bad insertion point %d\n",
 650                        __func__, where);
 651                 BUG();
 652         }
 653 }
 654 EXPORT_SYMBOL(__elv_add_request);
 655
 656 void elv_add_request(struct request_queue *q, struct request *rq, int where)
 657 {
 658         unsigned long flags;
 659
 660         spin_lock_irqsave(q->queue_lock, flags);
 661         __elv_add_request(q, rq, where);
 662         spin_unlock_irqrestore(q->queue_lock, flags);
 663 }
 664 EXPORT_SYMBOL(elv_add_request);
 665
 666 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 667 {
 668         struct elevator_queue *e = q->elevator;
 669
 670         if (e->type->ops.elevator_latter_req_fn)
 671                 return e->type->ops.elevator_latter_req_fn(q, rq);
 672         return NULL;
 673 }
 674
 675 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 676 {
 677         struct elevator_queue *e = q->elevator;
 678
 679         if (e->type->ops.elevator_former_req_fn)
 680                 return e->type->ops.elevator_former_req_fn(q, rq);
 681         return NULL;
 682 }
 683
 684 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 685 {
 686         struct elevator_queue *e = q->elevator;
 687
 688         if (e->type->ops.elevator_set_req_fn)
 689                 return e->type->ops.elevator_set_req_fn(q, rq, gfp_mask);
 690         return 0;
 691 }
 692
 693 void elv_put_request(struct request_queue *q, struct request *rq)
 694 {
 695         struct elevator_queue *e = q->elevator;
 696
 697         if (e->type->ops.elevator_put_req_fn)
 698                 e->type->ops.elevator_put_req_fn(rq);
 699 }
 700
 701 int elv_may_queue(struct request_queue *q, int rw)
 702 {
 703         struct elevator_queue *e = q->elevator;
 704
 705         if (e->type->ops.elevator_may_queue_fn)
 706                 return e->type->ops.elevator_may_queue_fn(q, rw);
 707
 708         return ELV_MQUEUE_MAY;
 709 }
 710
 711 void elv_abort_queue(struct request_queue *q)
 712 {
 713         struct request *rq;
 714
 715         blk_abort_flushes(q);
 716
 717         while (!list_empty(&q->queue_head)) {
 718                 rq = list_entry_rq(q->queue_head.next);
 719                 rq->cmd_flags |= REQ_QUIET;
 720                 trace_block_rq_abort(q, rq);
 721                 /*
 722                  * Mark this request as started so we don't trigger
 723                  * any debug logic in the end I/O path.
 724                  */
 725                 blk_start_request(rq);
 726                 __blk_end_request_all(rq, -EIO);
 727         }
 728 }
 729 EXPORT_SYMBOL(elv_abort_queue);
 730
 731 void elv_completed_request(struct request_queue *q, struct request *rq)
 732 {
 733         struct elevator_queue *e = q->elevator;
 734
 735         /*
 736          * request is released from the driver, io must be done
 737          */
 738         if (blk_account_rq(rq)) {
 739                 q->in_flight[rq_is_sync(rq)]--;
 740                 if ((rq->cmd_flags & REQ_SORTED) &&
 741                     e->type->ops.elevator_completed_req_fn)
 742                         e->type->ops.elevator_completed_req_fn(q, rq);
 743         }
 744 }
 745
 746 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 747
 748 static ssize_t
 749 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 750 {
 751         struct elv_fs_entry *entry = to_elv(attr);
 752         struct elevator_queue *e;
 753         ssize_t error;
 754
 755         if (!entry->show)
 756                 return -EIO;
 757
 758         e = container_of(kobj, struct elevator_queue, kobj);
 759         mutex_lock(&e->sysfs_lock);
 760         error = e->type ? entry->show(e, page) : -ENOENT;
 761         mutex_unlock(&e->sysfs_lock);
 762         return error;
 763 }
 764
 765 static ssize_t
 766 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 767                const char *page, size_t length)
 768 {
 769         struct elv_fs_entry *entry = to_elv(attr);
 770         struct elevator_queue *e;
 771         ssize_t error;
 772
 773         if (!entry->store)
 774                 return -EIO;
 775
 776         e = container_of(kobj, struct elevator_queue, kobj);
 777         mutex_lock(&e->sysfs_lock);
 778         error = e->type ? entry->store(e, page, length) : -ENOENT;
 779         mutex_unlock(&e->sysfs_lock);
 780         return error;
 781 }
 782
 783 static const struct sysfs_ops elv_sysfs_ops = {
 784         .show   = elv_attr_show,
 785         .store  = elv_attr_store,
 786 };
 787
 788 static struct kobj_type elv_ktype = {
 789         .sysfs_ops      = &elv_sysfs_ops,
 790         .release        = elevator_release,
 791 };
 792
 793 int elv_register_queue(struct request_queue *q)
 794 {
 795         struct elevator_queue *e = q->elevator;
 796         int error;
 797
 798         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 799         if (!error) {
 800                 struct elv_fs_entry *attr = e->type->elevator_attrs;
 801                 if (attr) {
 802                         while (attr->attr.name) {
 803                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 804                                         break;
 805                                 attr++;
 806                         }
 807                 }
 808                 kobject_uevent(&e->kobj, KOBJ_ADD);
 809                 e->registered = 1;
 810         }
 811         return error;
 812 }
 813 EXPORT_SYMBOL(elv_register_queue);
 814
 815 void elv_unregister_queue(struct request_queue *q)
 816 {
 817         if (q) {
 818                 struct elevator_queue *e = q->elevator;
 819
 820                 kobject_uevent(&e->kobj, KOBJ_REMOVE);
 821                 kobject_del(&e->kobj);
 822                 e->registered = 0;
 823         }
 824 }
 825 EXPORT_SYMBOL(elv_unregister_queue);
 826
 827 int elv_register(struct elevator_type *e)
 828 {
 829         char *def = "";
 830
 831         /* create icq_cache if requested */
 832         if (e->icq_size) {
 833                 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
 834                     WARN_ON(e->icq_align < __alignof__(struct io_cq)))
 835                         return -EINVAL;
 836
 837                 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
 838                          "%s_io_cq", e->elevator_name);
 839                 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
 840                                                  e->icq_align, 0, NULL);
 841                 if (!e->icq_cache)
 842                         return -ENOMEM;
 843         }
 844
 845         /* register, don't allow duplicate names */
 846         spin_lock(&elv_list_lock);
 847         if (elevator_find(e->elevator_name)) {
 848                 spin_unlock(&elv_list_lock);
 849                 if (e->icq_cache)
 850                         kmem_cache_destroy(e->icq_cache);
 851                 return -EBUSY;
 852         }
 853         list_add_tail(&e->list, &elv_list);
 854         spin_unlock(&elv_list_lock);
 855
 856         /* print pretty message */
 857         if (!strcmp(e->elevator_name, chosen_elevator) ||
 858                         (!*chosen_elevator &&
 859                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 860                                 def = " (default)";
 861
 862         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
 863                                                                 def);
 864         return 0;
 865 }
 866 EXPORT_SYMBOL_GPL(elv_register);
 867
 868 void elv_unregister(struct elevator_type *e)
 869 {
 870         /* unregister */
 871         spin_lock(&elv_list_lock);
 872         list_del_init(&e->list);
 873         spin_unlock(&elv_list_lock);
 874
 875         /*
 876          * Destroy icq_cache if it exists.  icq's are RCU managed.  Make
 877          * sure all RCU operations are complete before proceeding.
 878          */
 879         if (e->icq_cache) {
 880                 rcu_barrier();
 881                 kmem_cache_destroy(e->icq_cache);
 882                 e->icq_cache = NULL;
 883         }
 884 }
 885 EXPORT_SYMBOL_GPL(elv_unregister);
 886
 887 /*
 888  * switch to new_e io scheduler. be careful not to introduce deadlocks -
 889  * we don't free the old io scheduler, before we have allocated what we
 890  * need for the new one. this way we have a chance of going back to the old
 891  * one, if the new one fails init for some reason.
 892  */
 893 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
 894 {
 895         struct elevator_queue *old = q->elevator;
 896         bool registered = old->registered;
 897         int err;
 898
 899         /*
 900          * Turn on BYPASS and drain all requests w/ elevator private data.
 901          * Block layer doesn't call into a quiesced elevator - all requests
 902          * are directly put on the dispatch list without elevator data
 903          * using INSERT_BACK.  All requests have SOFTBARRIER set and no
 904          * merge happens either.
 905          */
 906         elv_quiesce_start(q);
 907
 908         /* unregister and clear all auxiliary data of the old elevator */
 909         if (registered)
 910                 elv_unregister_queue(q);
 911
 912         spin_lock_irq(q->queue_lock);
 913         ioc_clear_queue(q);
 914         spin_unlock_irq(q->queue_lock);
 915
 916         /* allocate, init and register new elevator */
 917         err = -ENOMEM;
 918         q->elevator = elevator_alloc(q, new_e);
 919         if (!q->elevator)
 920                 goto fail_init;
 921
 922         err = new_e->ops.elevator_init_fn(q);
 923         if (err) {
 924                 kobject_put(&q->elevator->kobj);
 925                 goto fail_init;
 926         }
 927
 928         if (registered) {
 929                 err = elv_register_queue(q);
 930                 if (err)
 931                         goto fail_register;
 932         }
 933
 934         /* done, kill the old one and finish */
 935         elevator_exit(old);
 936         elv_quiesce_end(q);
 937
 938         blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
 939
 940         return 0;
 941
 942 fail_register:
 943         elevator_exit(q->elevator);
 944 fail_init:
 945         /* switch failed, restore and re-register old elevator */
 946         q->elevator = old;
 947         elv_register_queue(q);
 948         elv_quiesce_end(q);
 949
 950         return err;
 951 }
 952
 953 /*
 954  * Switch this queue to the given IO scheduler.
 955  */
 956 int elevator_change(struct request_queue *q, const char *name)
 957 {
 958         char elevator_name[ELV_NAME_MAX];
 959         struct elevator_type *e;
 960
 961         if (!q->elevator)
 962                 return -ENXIO;
 963
 964         strlcpy(elevator_name, name, sizeof(elevator_name));
 965         e = elevator_get(strstrip(elevator_name));
 966         if (!e) {
 967                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
 968                 return -EINVAL;
 969         }
 970
 971         if (!strcmp(elevator_name, q->elevator->type->elevator_name)) {
 972                 elevator_put(e);
 973                 return 0;
 974         }
 975
 976         return elevator_switch(q, e);
 977 }
 978 EXPORT_SYMBOL(elevator_change);
 979
 980 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
 981                           size_t count)
 982 {
 983         int ret;
 984
 985         if (!q->elevator)
 986                 return count;
 987
 988         ret = elevator_change(q, name);
 989         if (!ret)
 990                 return count;
 991
 992         printk(KERN_ERR "elevator: switch to %s failed\n", name);
 993         return ret;
 994 }
 995
 996 ssize_t elv_iosched_show(struct request_queue *q, char *name)
 997 {
 998         struct elevator_queue *e = q->elevator;
 999         struct elevator_type *elv;
1000         struct elevator_type *__e;
1001         int len = 0;
1002
1003         if (!q->elevator || !blk_queue_stackable(q))
1004                 return sprintf(name, "none\n");
1005
1006         elv = e->type;
1007
1008         spin_lock(&elv_list_lock);
1009         list_for_each_entry(__e, &elv_list, list) {
1010                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1011                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1012                 else
1013                         len += sprintf(name+len, "%s ", __e->elevator_name);
1014         }
1015         spin_unlock(&elv_list_lock);
1016
1017         len += sprintf(len+name, "\n");
1018         return len;
1019 }
1020
1021 struct request *elv_rb_former_request(struct request_queue *q,
1022                                       struct request *rq)
1023 {
1024         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1025
1026         if (rbprev)
1027                 return rb_entry_rq(rbprev);
1028
1029         return NULL;
1030 }
1031 EXPORT_SYMBOL(elv_rb_former_request);
1032
1033 struct request *elv_rb_latter_request(struct request_queue *q,
1034                                       struct request *rq)
1035 {
1036         struct rb_node *rbnext = rb_next(&rq->rb_node);
1037
1038         if (rbnext)
1039                 return rb_entry_rq(rbnext);
1040
1041         return NULL;
1042 }
1043 EXPORT_SYMBOL(elv_rb_latter_request);