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/delay.h>
  35 #include <linux/blktrace_api.h>
  36 #include <linux/hash.h>
  37 #include <linux/uaccess.h>
  38
  39 #include <trace/events/block.h>
  40
  41 #include "blk.h"
  42
  43 static DEFINE_SPINLOCK(elv_list_lock);
  44 static LIST_HEAD(elv_list);
  45
  46 /*
  47  * Merge hash stuff.
  48  */
  49 static const int elv_hash_shift = 6;
  50 #define ELV_HASH_BLOCK(sec)     ((sec) >> 3)
  51 #define ELV_HASH_FN(sec)        \
  52                 (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
  53 #define ELV_HASH_ENTRIES        (1 << elv_hash_shift)
  54 #define rq_hash_key(rq)         (blk_rq_pos(rq) + blk_rq_sectors(rq))
  55
  56 /*
  57  * Query io scheduler to see if the current process issuing bio may be
  58  * merged with rq.
  59  */
  60 static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
  61 {
  62         struct request_queue *q = rq->q;
  63         struct elevator_queue *e = q->elevator;
  64
  65         if (e->ops->elevator_allow_merge_fn)
  66                 return e->ops->elevator_allow_merge_fn(q, rq, bio);
  67
  68         return 1;
  69 }
  70
  71 /*
  72  * can we safely merge with this request?
  73  */
  74 int elv_rq_merge_ok(struct request *rq, struct bio *bio)
  75 {
  76         if (!rq_mergeable(rq))
  77                 return 0;
  78
  79         /*
  80          * Don't merge file system requests and discard requests
  81          */
  82         if (bio_rw_flagged(bio, BIO_RW_DISCARD) !=
  83             bio_rw_flagged(rq->bio, BIO_RW_DISCARD))
  84                 return 0;
  85
  86         /*
  87          * different data direction or already started, don't merge
  88          */
  89         if (bio_data_dir(bio) != rq_data_dir(rq))
  90                 return 0;
  91
  92         /*
  93          * must be same device and not a special request
  94          */
  95         if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
  96                 return 0;
  97
  98         /*
  99          * only merge integrity protected bio into ditto rq
 100          */
 101         if (bio_integrity(bio) != blk_integrity_rq(rq))
 102                 return 0;
 103
 104         if (!elv_iosched_allow_merge(rq, bio))
 105                 return 0;
 106
 107         return 1;
 108 }
 109 EXPORT_SYMBOL(elv_rq_merge_ok);
 110
 111 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
 112 {
 113         int ret = ELEVATOR_NO_MERGE;
 114
 115         /*
 116          * we can merge and sequence is ok, check if it's possible
 117          */
 118         if (elv_rq_merge_ok(__rq, bio)) {
 119                 if (blk_rq_pos(__rq) + blk_rq_sectors(__rq) == bio->bi_sector)
 120                         ret = ELEVATOR_BACK_MERGE;
 121                 else if (blk_rq_pos(__rq) - bio_sectors(bio) == bio->bi_sector)
 122                         ret = ELEVATOR_FRONT_MERGE;
 123         }
 124
 125         return ret;
 126 }
 127
 128 static struct elevator_type *elevator_find(const char *name)
 129 {
 130         struct elevator_type *e;
 131
 132         list_for_each_entry(e, &elv_list, list) {
 133                 if (!strcmp(e->elevator_name, name))
 134                         return e;
 135         }
 136
 137         return NULL;
 138 }
 139
 140 static void elevator_put(struct elevator_type *e)
 141 {
 142         module_put(e->elevator_owner);
 143 }
 144
 145 static struct elevator_type *elevator_get(const char *name)
 146 {
 147         struct elevator_type *e;
 148
 149         spin_lock(&elv_list_lock);
 150
 151         e = elevator_find(name);
 152         if (!e) {
 153                 char elv[ELV_NAME_MAX + strlen("-iosched")];
 154
 155                 spin_unlock(&elv_list_lock);
 156
 157                 snprintf(elv, sizeof(elv), "%s-iosched", name);
 158
 159                 request_module("%s", elv);
 160                 spin_lock(&elv_list_lock);
 161                 e = elevator_find(name);
 162         }
 163
 164         if (e && !try_module_get(e->elevator_owner))
 165                 e = NULL;
 166
 167         spin_unlock(&elv_list_lock);
 168
 169         return e;
 170 }
 171
 172 static void *elevator_init_queue(struct request_queue *q,
 173                                  struct elevator_queue *eq)
 174 {
 175         return eq->ops->elevator_init_fn(q);
 176 }
 177
 178 static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
 179                            void *data)
 180 {
 181         q->elevator = eq;
 182         eq->elevator_data = data;
 183 }
 184
 185 static char chosen_elevator[16];
 186
 187 static int __init elevator_setup(char *str)
 188 {
 189         /*
 190          * Be backwards-compatible with previous kernels, so users
 191          * won't get the wrong elevator.
 192          */
 193         strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 194         return 1;
 195 }
 196
 197 __setup("elevator=", elevator_setup);
 198
 199 static struct kobj_type elv_ktype;
 200
 201 static struct elevator_queue *elevator_alloc(struct request_queue *q,
 202                                   struct elevator_type *e)
 203 {
 204         struct elevator_queue *eq;
 205         int i;
 206
 207         eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node);
 208         if (unlikely(!eq))
 209                 goto err;
 210
 211         eq->ops = &e->ops;
 212         eq->elevator_type = e;
 213         kobject_init(&eq->kobj, &elv_ktype);
 214         mutex_init(&eq->sysfs_lock);
 215
 216         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 217                                         GFP_KERNEL, q->node);
 218         if (!eq->hash)
 219                 goto err;
 220
 221         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 222                 INIT_HLIST_HEAD(&eq->hash[i]);
 223
 224         return eq;
 225 err:
 226         kfree(eq);
 227         elevator_put(e);
 228         return NULL;
 229 }
 230
 231 static void elevator_release(struct kobject *kobj)
 232 {
 233         struct elevator_queue *e;
 234
 235         e = container_of(kobj, struct elevator_queue, kobj);
 236         elevator_put(e->elevator_type);
 237         kfree(e->hash);
 238         kfree(e);
 239 }
 240
 241 int elevator_init(struct request_queue *q, char *name)
 242 {
 243         struct elevator_type *e = NULL;
 244         struct elevator_queue *eq;
 245         int ret = 0;
 246         void *data;
 247
 248         INIT_LIST_HEAD(&q->queue_head);
 249         q->last_merge = NULL;
 250         q->end_sector = 0;
 251         q->boundary_rq = NULL;
 252
 253         if (name) {
 254                 e = elevator_get(name);
 255                 if (!e)
 256                         return -EINVAL;
 257         }
 258
 259         if (!e && *chosen_elevator) {
 260                 e = elevator_get(chosen_elevator);
 261                 if (!e)
 262                         printk(KERN_ERR "I/O scheduler %s not found\n",
 263                                                         chosen_elevator);
 264         }
 265
 266         if (!e) {
 267                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
 268                 if (!e) {
 269                         printk(KERN_ERR
 270                                 "Default I/O scheduler not found. " \
 271                                 "Using noop.\n");
 272                         e = elevator_get("noop");
 273                 }
 274         }
 275
 276         eq = elevator_alloc(q, e);
 277         if (!eq)
 278                 return -ENOMEM;
 279
 280         data = elevator_init_queue(q, eq);
 281         if (!data) {
 282                 kobject_put(&eq->kobj);
 283                 return -ENOMEM;
 284         }
 285
 286         elevator_attach(q, eq, data);
 287         return ret;
 288 }
 289 EXPORT_SYMBOL(elevator_init);
 290
 291 void elevator_exit(struct elevator_queue *e)
 292 {
 293         mutex_lock(&e->sysfs_lock);
 294         if (e->ops->elevator_exit_fn)
 295                 e->ops->elevator_exit_fn(e);
 296         e->ops = NULL;
 297         mutex_unlock(&e->sysfs_lock);
 298
 299         kobject_put(&e->kobj);
 300 }
 301 EXPORT_SYMBOL(elevator_exit);
 302
 303 static inline void __elv_rqhash_del(struct request *rq)
 304 {
 305         hlist_del_init(&rq->hash);
 306 }
 307
 308 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 309 {
 310         if (ELV_ON_HASH(rq))
 311                 __elv_rqhash_del(rq);
 312 }
 313
 314 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 315 {
 316         struct elevator_queue *e = q->elevator;
 317
 318         BUG_ON(ELV_ON_HASH(rq));
 319         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 320 }
 321
 322 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 323 {
 324         __elv_rqhash_del(rq);
 325         elv_rqhash_add(q, rq);
 326 }
 327
 328 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 329 {
 330         struct elevator_queue *e = q->elevator;
 331         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 332         struct hlist_node *entry, *next;
 333         struct request *rq;
 334
 335         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 336                 BUG_ON(!ELV_ON_HASH(rq));
 337
 338                 if (unlikely(!rq_mergeable(rq))) {
 339                         __elv_rqhash_del(rq);
 340                         continue;
 341                 }
 342
 343                 if (rq_hash_key(rq) == offset)
 344                         return rq;
 345         }
 346
 347         return NULL;
 348 }
 349
 350 /*
 351  * RB-tree support functions for inserting/lookup/removal of requests
 352  * in a sorted RB tree.
 353  */
 354 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 355 {
 356         struct rb_node **p = &root->rb_node;
 357         struct rb_node *parent = NULL;
 358         struct request *__rq;
 359
 360         while (*p) {
 361                 parent = *p;
 362                 __rq = rb_entry(parent, struct request, rb_node);
 363
 364                 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
 365                         p = &(*p)->rb_left;
 366                 else if (blk_rq_pos(rq) > blk_rq_pos(__rq))
 367                         p = &(*p)->rb_right;
 368                 else
 369                         return __rq;
 370         }
 371
 372         rb_link_node(&rq->rb_node, parent, p);
 373         rb_insert_color(&rq->rb_node, root);
 374         return NULL;
 375 }
 376 EXPORT_SYMBOL(elv_rb_add);
 377
 378 void elv_rb_del(struct rb_root *root, struct request *rq)
 379 {
 380         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 381         rb_erase(&rq->rb_node, root);
 382         RB_CLEAR_NODE(&rq->rb_node);
 383 }
 384 EXPORT_SYMBOL(elv_rb_del);
 385
 386 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 387 {
 388         struct rb_node *n = root->rb_node;
 389         struct request *rq;
 390
 391         while (n) {
 392                 rq = rb_entry(n, struct request, rb_node);
 393
 394                 if (sector < blk_rq_pos(rq))
 395                         n = n->rb_left;
 396                 else if (sector > blk_rq_pos(rq))
 397                         n = n->rb_right;
 398                 else
 399                         return rq;
 400         }
 401
 402         return NULL;
 403 }
 404 EXPORT_SYMBOL(elv_rb_find);
 405
 406 /*
 407  * Insert rq into dispatch queue of q.  Queue lock must be held on
 408  * entry.  rq is sort instead into the dispatch queue. To be used by
 409  * specific elevators.
 410  */
 411 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 412 {
 413         sector_t boundary;
 414         struct list_head *entry;
 415         int stop_flags;
 416
 417         if (q->last_merge == rq)
 418                 q->last_merge = NULL;
 419
 420         elv_rqhash_del(q, rq);
 421
 422         q->nr_sorted--;
 423
 424         boundary = q->end_sector;
 425         stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
 426         list_for_each_prev(entry, &q->queue_head) {
 427                 struct request *pos = list_entry_rq(entry);
 428
 429                 if (blk_discard_rq(rq) != blk_discard_rq(pos))
 430                         break;
 431                 if (rq_data_dir(rq) != rq_data_dir(pos))
 432                         break;
 433                 if (pos->cmd_flags & stop_flags)
 434                         break;
 435                 if (blk_rq_pos(rq) >= boundary) {
 436                         if (blk_rq_pos(pos) < boundary)
 437                                 continue;
 438                 } else {
 439                         if (blk_rq_pos(pos) >= boundary)
 440                                 break;
 441                 }
 442                 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
 443                         break;
 444         }
 445
 446         list_add(&rq->queuelist, entry);
 447 }
 448 EXPORT_SYMBOL(elv_dispatch_sort);
 449
 450 /*
 451  * Insert rq into dispatch queue of q.  Queue lock must be held on
 452  * entry.  rq is added to the back of the dispatch queue. To be used by
 453  * specific elevators.
 454  */
 455 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 456 {
 457         if (q->last_merge == rq)
 458                 q->last_merge = NULL;
 459
 460         elv_rqhash_del(q, rq);
 461
 462         q->nr_sorted--;
 463
 464         q->end_sector = rq_end_sector(rq);
 465         q->boundary_rq = rq;
 466         list_add_tail(&rq->queuelist, &q->queue_head);
 467 }
 468 EXPORT_SYMBOL(elv_dispatch_add_tail);
 469
 470 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 471 {
 472         struct elevator_queue *e = q->elevator;
 473         struct request *__rq;
 474         int ret;
 475
 476         /*
 477          * Levels of merges:
 478          *      nomerges:  No merges at all attempted
 479          *      noxmerges: Only simple one-hit cache try
 480          *      merges:    All merge tries attempted
 481          */
 482         if (blk_queue_nomerges(q))
 483                 return ELEVATOR_NO_MERGE;
 484
 485         /*
 486          * First try one-hit cache.
 487          */
 488         if (q->last_merge) {
 489                 ret = elv_try_merge(q->last_merge, bio);
 490                 if (ret != ELEVATOR_NO_MERGE) {
 491                         *req = q->last_merge;
 492                         return ret;
 493                 }
 494         }
 495
 496         if (blk_queue_noxmerges(q))
 497                 return ELEVATOR_NO_MERGE;
 498
 499         /*
 500          * See if our hash lookup can find a potential backmerge.
 501          */
 502         __rq = elv_rqhash_find(q, bio->bi_sector);
 503         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 504                 *req = __rq;
 505                 return ELEVATOR_BACK_MERGE;
 506         }
 507
 508         if (e->ops->elevator_merge_fn)
 509                 return e->ops->elevator_merge_fn(q, req, bio);
 510
 511         return ELEVATOR_NO_MERGE;
 512 }
 513
 514 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 515 {
 516         struct elevator_queue *e = q->elevator;
 517
 518         if (e->ops->elevator_merged_fn)
 519                 e->ops->elevator_merged_fn(q, rq, type);
 520
 521         if (type == ELEVATOR_BACK_MERGE)
 522                 elv_rqhash_reposition(q, rq);
 523
 524         q->last_merge = rq;
 525 }
 526
 527 void elv_merge_requests(struct request_queue *q, struct request *rq,
 528                              struct request *next)
 529 {
 530         struct elevator_queue *e = q->elevator;
 531
 532         if (e->ops->elevator_merge_req_fn)
 533                 e->ops->elevator_merge_req_fn(q, rq, next);
 534
 535         elv_rqhash_reposition(q, rq);
 536         elv_rqhash_del(q, next);
 537
 538         q->nr_sorted--;
 539         q->last_merge = rq;
 540 }
 541
 542 void elv_bio_merged(struct request_queue *q, struct request *rq,
 543                         struct bio *bio)
 544 {
 545         struct elevator_queue *e = q->elevator;
 546
 547         if (e->ops->elevator_bio_merged_fn)
 548                 e->ops->elevator_bio_merged_fn(q, rq, bio);
 549 }
 550
 551 void elv_requeue_request(struct request_queue *q, struct request *rq)
 552 {
 553         /*
 554          * it already went through dequeue, we need to decrement the
 555          * in_flight count again
 556          */
 557         if (blk_account_rq(rq)) {
 558                 q->in_flight[rq_is_sync(rq)]--;
 559                 if (blk_sorted_rq(rq))
 560                         elv_deactivate_rq(q, rq);
 561         }
 562
 563         rq->cmd_flags &= ~REQ_STARTED;
 564
 565         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 566 }
 567
 568 void elv_drain_elevator(struct request_queue *q)
 569 {
 570         static int printed;
 571         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 572                 ;
 573         if (q->nr_sorted == 0)
 574                 return;
 575         if (printed++ < 10) {
 576                 printk(KERN_ERR "%s: forced dispatching is broken "
 577                        "(nr_sorted=%u), please report this\n",
 578                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 579         }
 580 }
 581
 582 /*
 583  * Call with queue lock held, interrupts disabled
 584  */
 585 void elv_quiesce_start(struct request_queue *q)
 586 {
 587         if (!q->elevator)
 588                 return;
 589
 590         queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
 591
 592         /*
 593          * make sure we don't have any requests in flight
 594          */
 595         elv_drain_elevator(q);
 596         while (q->rq.elvpriv) {
 597                 __blk_run_queue(q);
 598                 spin_unlock_irq(q->queue_lock);
 599                 msleep(10);
 600                 spin_lock_irq(q->queue_lock);
 601                 elv_drain_elevator(q);
 602         }
 603 }
 604
 605 void elv_quiesce_end(struct request_queue *q)
 606 {
 607         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
 608 }
 609
 610 void elv_insert(struct request_queue *q, struct request *rq, int where)
 611 {
 612         struct list_head *pos;
 613         unsigned ordseq;
 614         int unplug_it = 1;
 615
 616         trace_block_rq_insert(q, rq);
 617
 618         rq->q = q;
 619
 620         switch (where) {
 621         case ELEVATOR_INSERT_FRONT:
 622                 rq->cmd_flags |= REQ_SOFTBARRIER;
 623
 624                 list_add(&rq->queuelist, &q->queue_head);
 625                 break;
 626
 627         case ELEVATOR_INSERT_BACK:
 628                 rq->cmd_flags |= REQ_SOFTBARRIER;
 629                 elv_drain_elevator(q);
 630                 list_add_tail(&rq->queuelist, &q->queue_head);
 631                 /*
 632                  * We kick the queue here for the following reasons.
 633                  * - The elevator might have returned NULL previously
 634                  *   to delay requests and returned them now.  As the
 635                  *   queue wasn't empty before this request, ll_rw_blk
 636                  *   won't run the queue on return, resulting in hang.
 637                  * - Usually, back inserted requests won't be merged
 638                  *   with anything.  There's no point in delaying queue
 639                  *   processing.
 640                  */
 641                 __blk_run_queue(q);
 642                 break;
 643
 644         case ELEVATOR_INSERT_SORT:
 645                 BUG_ON(!blk_fs_request(rq) && !blk_discard_rq(rq));
 646                 rq->cmd_flags |= REQ_SORTED;
 647                 q->nr_sorted++;
 648                 if (rq_mergeable(rq)) {
 649                         elv_rqhash_add(q, rq);
 650                         if (!q->last_merge)
 651                                 q->last_merge = rq;
 652                 }
 653
 654                 /*
 655                  * Some ioscheds (cfq) run q->request_fn directly, so
 656                  * rq cannot be accessed after calling
 657                  * elevator_add_req_fn.
 658                  */
 659                 q->elevator->ops->elevator_add_req_fn(q, rq);
 660                 break;
 661
 662         case ELEVATOR_INSERT_REQUEUE:
 663                 /*
 664                  * If ordered flush isn't in progress, we do front
 665                  * insertion; otherwise, requests should be requeued
 666                  * in ordseq order.
 667                  */
 668                 rq->cmd_flags |= REQ_SOFTBARRIER;
 669
 670                 /*
 671                  * Most requeues happen because of a busy condition,
 672                  * don't force unplug of the queue for that case.
 673                  */
 674                 unplug_it = 0;
 675
 676                 if (q->ordseq == 0) {
 677                         list_add(&rq->queuelist, &q->queue_head);
 678                         break;
 679                 }
 680
 681                 ordseq = blk_ordered_req_seq(rq);
 682
 683                 list_for_each(pos, &q->queue_head) {
 684                         struct request *pos_rq = list_entry_rq(pos);
 685                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 686                                 break;
 687                 }
 688
 689                 list_add_tail(&rq->queuelist, pos);
 690                 break;
 691
 692         default:
 693                 printk(KERN_ERR "%s: bad insertion point %d\n",
 694                        __func__, where);
 695                 BUG();
 696         }
 697
 698         if (unplug_it && blk_queue_plugged(q)) {
 699                 int nrq = q->rq.count[BLK_RW_SYNC] + q->rq.count[BLK_RW_ASYNC]
 700                                 - queue_in_flight(q);
 701
 702                 if (nrq >= q->unplug_thresh)
 703                         __generic_unplug_device(q);
 704         }
 705 }
 706
 707 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
 708                        int plug)
 709 {
 710         if (q->ordcolor)
 711                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 712
 713         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 714                 /*
 715                  * toggle ordered color
 716                  */
 717                 if (blk_barrier_rq(rq))
 718                         q->ordcolor ^= 1;
 719
 720                 /*
 721                  * barriers implicitly indicate back insertion
 722                  */
 723                 if (where == ELEVATOR_INSERT_SORT)
 724                         where = ELEVATOR_INSERT_BACK;
 725
 726                 /*
 727                  * this request is scheduling boundary, update
 728                  * end_sector
 729                  */
 730                 if (blk_fs_request(rq) || blk_discard_rq(rq)) {
 731                         q->end_sector = rq_end_sector(rq);
 732                         q->boundary_rq = rq;
 733                 }
 734         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 735                     where == ELEVATOR_INSERT_SORT)
 736                 where = ELEVATOR_INSERT_BACK;
 737
 738         if (plug)
 739                 blk_plug_device(q);
 740
 741         elv_insert(q, rq, where);
 742 }
 743 EXPORT_SYMBOL(__elv_add_request);
 744
 745 void elv_add_request(struct request_queue *q, struct request *rq, int where,
 746                      int plug)
 747 {
 748         unsigned long flags;
 749
 750         spin_lock_irqsave(q->queue_lock, flags);
 751         __elv_add_request(q, rq, where, plug);
 752         spin_unlock_irqrestore(q->queue_lock, flags);
 753 }
 754 EXPORT_SYMBOL(elv_add_request);
 755
 756 int elv_queue_empty(struct request_queue *q)
 757 {
 758         struct elevator_queue *e = q->elevator;
 759
 760         if (!list_empty(&q->queue_head))
 761                 return 0;
 762
 763         if (e->ops->elevator_queue_empty_fn)
 764                 return e->ops->elevator_queue_empty_fn(q);
 765
 766         return 1;
 767 }
 768 EXPORT_SYMBOL(elv_queue_empty);
 769
 770 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 771 {
 772         struct elevator_queue *e = q->elevator;
 773
 774         if (e->ops->elevator_latter_req_fn)
 775                 return e->ops->elevator_latter_req_fn(q, rq);
 776         return NULL;
 777 }
 778
 779 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 780 {
 781         struct elevator_queue *e = q->elevator;
 782
 783         if (e->ops->elevator_former_req_fn)
 784                 return e->ops->elevator_former_req_fn(q, rq);
 785         return NULL;
 786 }
 787
 788 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 789 {
 790         struct elevator_queue *e = q->elevator;
 791
 792         if (e->ops->elevator_set_req_fn)
 793                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 794
 795         rq->elevator_private = NULL;
 796         return 0;
 797 }
 798
 799 void elv_put_request(struct request_queue *q, struct request *rq)
 800 {
 801         struct elevator_queue *e = q->elevator;
 802
 803         if (e->ops->elevator_put_req_fn)
 804                 e->ops->elevator_put_req_fn(rq);
 805 }
 806
 807 int elv_may_queue(struct request_queue *q, int rw)
 808 {
 809         struct elevator_queue *e = q->elevator;
 810
 811         if (e->ops->elevator_may_queue_fn)
 812                 return e->ops->elevator_may_queue_fn(q, rw);
 813
 814         return ELV_MQUEUE_MAY;
 815 }
 816
 817 void elv_abort_queue(struct request_queue *q)
 818 {
 819         struct request *rq;
 820
 821         while (!list_empty(&q->queue_head)) {
 822                 rq = list_entry_rq(q->queue_head.next);
 823                 rq->cmd_flags |= REQ_QUIET;
 824                 trace_block_rq_abort(q, rq);
 825                 /*
 826                  * Mark this request as started so we don't trigger
 827                  * any debug logic in the end I/O path.
 828                  */
 829                 blk_start_request(rq);
 830                 __blk_end_request_all(rq, -EIO);
 831         }
 832 }
 833 EXPORT_SYMBOL(elv_abort_queue);
 834
 835 void elv_completed_request(struct request_queue *q, struct request *rq)
 836 {
 837         struct elevator_queue *e = q->elevator;
 838
 839         /*
 840          * request is released from the driver, io must be done
 841          */
 842         if (blk_account_rq(rq)) {
 843                 q->in_flight[rq_is_sync(rq)]--;
 844                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 845                         e->ops->elevator_completed_req_fn(q, rq);
 846         }
 847
 848         /*
 849          * Check if the queue is waiting for fs requests to be
 850          * drained for flush sequence.
 851          */
 852         if (unlikely(q->ordseq)) {
 853                 struct request *next = NULL;
 854
 855                 if (!list_empty(&q->queue_head))
 856                         next = list_entry_rq(q->queue_head.next);
 857
 858                 if (!queue_in_flight(q) &&
 859                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 860                     (!next || blk_ordered_req_seq(next) > QUEUE_ORDSEQ_DRAIN)) {
 861                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 862                         __blk_run_queue(q);
 863                 }
 864         }
 865 }
 866
 867 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 868
 869 static ssize_t
 870 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 871 {
 872         struct elv_fs_entry *entry = to_elv(attr);
 873         struct elevator_queue *e;
 874         ssize_t error;
 875
 876         if (!entry->show)
 877                 return -EIO;
 878
 879         e = container_of(kobj, struct elevator_queue, kobj);
 880         mutex_lock(&e->sysfs_lock);
 881         error = e->ops ? entry->show(e, page) : -ENOENT;
 882         mutex_unlock(&e->sysfs_lock);
 883         return error;
 884 }
 885
 886 static ssize_t
 887 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 888                const char *page, size_t length)
 889 {
 890         struct elv_fs_entry *entry = to_elv(attr);
 891         struct elevator_queue *e;
 892         ssize_t error;
 893
 894         if (!entry->store)
 895                 return -EIO;
 896
 897         e = container_of(kobj, struct elevator_queue, kobj);
 898         mutex_lock(&e->sysfs_lock);
 899         error = e->ops ? entry->store(e, page, length) : -ENOENT;
 900         mutex_unlock(&e->sysfs_lock);
 901         return error;
 902 }
 903
 904 static const struct sysfs_ops elv_sysfs_ops = {
 905         .show   = elv_attr_show,
 906         .store  = elv_attr_store,
 907 };
 908
 909 static struct kobj_type elv_ktype = {
 910         .sysfs_ops      = &elv_sysfs_ops,
 911         .release        = elevator_release,
 912 };
 913
 914 int elv_register_queue(struct request_queue *q)
 915 {
 916         struct elevator_queue *e = q->elevator;
 917         int error;
 918
 919         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 920         if (!error) {
 921                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
 922                 if (attr) {
 923                         while (attr->attr.name) {
 924                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 925                                         break;
 926                                 attr++;
 927                         }
 928                 }
 929                 kobject_uevent(&e->kobj, KOBJ_ADD);
 930         }
 931         return error;
 932 }
 933 EXPORT_SYMBOL(elv_register_queue);
 934
 935 static void __elv_unregister_queue(struct elevator_queue *e)
 936 {
 937         kobject_uevent(&e->kobj, KOBJ_REMOVE);
 938         kobject_del(&e->kobj);
 939 }
 940
 941 void elv_unregister_queue(struct request_queue *q)
 942 {
 943         if (q)
 944                 __elv_unregister_queue(q->elevator);
 945 }
 946 EXPORT_SYMBOL(elv_unregister_queue);
 947
 948 void elv_register(struct elevator_type *e)
 949 {
 950         char *def = "";
 951
 952         spin_lock(&elv_list_lock);
 953         BUG_ON(elevator_find(e->elevator_name));
 954         list_add_tail(&e->list, &elv_list);
 955         spin_unlock(&elv_list_lock);
 956
 957         if (!strcmp(e->elevator_name, chosen_elevator) ||
 958                         (!*chosen_elevator &&
 959                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 960                                 def = " (default)";
 961
 962         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
 963                                                                 def);
 964 }
 965 EXPORT_SYMBOL_GPL(elv_register);
 966
 967 void elv_unregister(struct elevator_type *e)
 968 {
 969         struct task_struct *g, *p;
 970
 971         /*
 972          * Iterate every thread in the process to remove the io contexts.
 973          */
 974         if (e->ops.trim) {
 975                 read_lock(&tasklist_lock);
 976                 do_each_thread(g, p) {
 977                         task_lock(p);
 978                         if (p->io_context)
 979                                 e->ops.trim(p->io_context);
 980                         task_unlock(p);
 981                 } while_each_thread(g, p);
 982                 read_unlock(&tasklist_lock);
 983         }
 984
 985         spin_lock(&elv_list_lock);
 986         list_del_init(&e->list);
 987         spin_unlock(&elv_list_lock);
 988 }
 989 EXPORT_SYMBOL_GPL(elv_unregister);
 990
 991 /*
 992  * switch to new_e io scheduler. be careful not to introduce deadlocks -
 993  * we don't free the old io scheduler, before we have allocated what we
 994  * need for the new one. this way we have a chance of going back to the old
 995  * one, if the new one fails init for some reason.
 996  */
 997 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
 998 {
 999         struct elevator_queue *old_elevator, *e;
1000         void *data;
1001
1002         /*
1003          * Allocate new elevator
1004          */
1005         e = elevator_alloc(q, new_e);
1006         if (!e)
1007                 return 0;
1008
1009         data = elevator_init_queue(q, e);
1010         if (!data) {
1011                 kobject_put(&e->kobj);
1012                 return 0;
1013         }
1014
1015         /*
1016          * Turn on BYPASS and drain all requests w/ elevator private data
1017          */
1018         spin_lock_irq(q->queue_lock);
1019         elv_quiesce_start(q);
1020
1021         /*
1022          * Remember old elevator.
1023          */
1024         old_elevator = q->elevator;
1025
1026         /*
1027          * attach and start new elevator
1028          */
1029         elevator_attach(q, e, data);
1030
1031         spin_unlock_irq(q->queue_lock);
1032
1033         __elv_unregister_queue(old_elevator);
1034
1035         if (elv_register_queue(q))
1036                 goto fail_register;
1037
1038         /*
1039          * finally exit old elevator and turn off BYPASS.
1040          */
1041         elevator_exit(old_elevator);
1042         spin_lock_irq(q->queue_lock);
1043         elv_quiesce_end(q);
1044         spin_unlock_irq(q->queue_lock);
1045
1046         blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
1047
1048         return 1;
1049
1050 fail_register:
1051         /*
1052          * switch failed, exit the new io scheduler and reattach the old
1053          * one again (along with re-adding the sysfs dir)
1054          */
1055         elevator_exit(e);
1056         q->elevator = old_elevator;
1057         elv_register_queue(q);
1058
1059         spin_lock_irq(q->queue_lock);
1060         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1061         spin_unlock_irq(q->queue_lock);
1062
1063         return 0;
1064 }
1065
1066 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1067                           size_t count)
1068 {
1069         char elevator_name[ELV_NAME_MAX];
1070         struct elevator_type *e;
1071
1072         if (!q->elevator)
1073                 return count;
1074
1075         strlcpy(elevator_name, name, sizeof(elevator_name));
1076         e = elevator_get(strstrip(elevator_name));
1077         if (!e) {
1078                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1079                 return -EINVAL;
1080         }
1081
1082         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1083                 elevator_put(e);
1084                 return count;
1085         }
1086
1087         if (!elevator_switch(q, e))
1088                 printk(KERN_ERR "elevator: switch to %s failed\n",
1089                                                         elevator_name);
1090         return count;
1091 }
1092
1093 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1094 {
1095         struct elevator_queue *e = q->elevator;
1096         struct elevator_type *elv;
1097         struct elevator_type *__e;
1098         int len = 0;
1099
1100         if (!q->elevator)
1101                 return sprintf(name, "none\n");
1102
1103         elv = e->elevator_type;
1104
1105         spin_lock(&elv_list_lock);
1106         list_for_each_entry(__e, &elv_list, list) {
1107                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1108                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1109                 else
1110                         len += sprintf(name+len, "%s ", __e->elevator_name);
1111         }
1112         spin_unlock(&elv_list_lock);
1113
1114         len += sprintf(len+name, "\n");
1115         return len;
1116 }
1117
1118 struct request *elv_rb_former_request(struct request_queue *q,
1119                                       struct request *rq)
1120 {
1121         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1122
1123         if (rbprev)
1124                 return rb_entry_rq(rbprev);
1125
1126         return NULL;
1127 }
1128 EXPORT_SYMBOL(elv_rb_former_request);
1129
1130 struct request *elv_rb_latter_request(struct request_queue *q,
1131                                       struct request *rq)
1132 {
1133         struct rb_node *rbnext = rb_next(&rq->rb_node);
1134
1135         if (rbnext)
1136                 return rb_entry_rq(rbnext);
1137
1138         return NULL;
1139 }
1140 EXPORT_SYMBOL(elv_rb_latter_request);