block/elevator.c

   1 /*
   2  *  Block device elevator/IO-scheduler.
   3  *
   4  *  Copyright (C) 2000 Andrea Arcangeli <[email protected]> SuSE
   5  *
   6  * 30042000 Jens Axboe <[email protected]> :
   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 <[email protected]> :
  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                 if (!strcmp(name, "anticipatory"))
 158                         sprintf(elv, "as-iosched");
 159                 else
 160                         sprintf(elv, "%s-iosched", name);
 161
 162                 request_module("%s", elv);
 163                 spin_lock(&elv_list_lock);
 164                 e = elevator_find(name);
 165         }
 166
 167         if (e && !try_module_get(e->elevator_owner))
 168                 e = NULL;
 169
 170         spin_unlock(&elv_list_lock);
 171
 172         return e;
 173 }
 174
 175 static void *elevator_init_queue(struct request_queue *q,
 176                                  struct elevator_queue *eq)
 177 {
 178         return eq->ops->elevator_init_fn(q);
 179 }
 180
 181 static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
 182                            void *data)
 183 {
 184         q->elevator = eq;
 185         eq->elevator_data = data;
 186 }
 187
 188 static char chosen_elevator[16];
 189
 190 static int __init elevator_setup(char *str)
 191 {
 192         /*
 193          * Be backwards-compatible with previous kernels, so users
 194          * won't get the wrong elevator.
 195          */
 196         if (!strcmp(str, "as"))
 197                 strcpy(chosen_elevator, "anticipatory");
 198         else
 199                 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 200         return 1;
 201 }
 202
 203 __setup("elevator=", elevator_setup);
 204
 205 static struct kobj_type elv_ktype;
 206
 207 static struct elevator_queue *elevator_alloc(struct request_queue *q,
 208                                   struct elevator_type *e)
 209 {
 210         struct elevator_queue *eq;
 211         int i;
 212
 213         eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node);
 214         if (unlikely(!eq))
 215                 goto err;
 216
 217         eq->ops = &e->ops;
 218         eq->elevator_type = e;
 219         kobject_init(&eq->kobj, &elv_ktype);
 220         mutex_init(&eq->sysfs_lock);
 221
 222         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 223                                         GFP_KERNEL, q->node);
 224         if (!eq->hash)
 225                 goto err;
 226
 227         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 228                 INIT_HLIST_HEAD(&eq->hash[i]);
 229
 230         return eq;
 231 err:
 232         kfree(eq);
 233         elevator_put(e);
 234         return NULL;
 235 }
 236
 237 static void elevator_release(struct kobject *kobj)
 238 {
 239         struct elevator_queue *e;
 240
 241         e = container_of(kobj, struct elevator_queue, kobj);
 242         elevator_put(e->elevator_type);
 243         kfree(e->hash);
 244         kfree(e);
 245 }
 246
 247 int elevator_init(struct request_queue *q, char *name)
 248 {
 249         struct elevator_type *e = NULL;
 250         struct elevator_queue *eq;
 251         int ret = 0;
 252         void *data;
 253
 254         INIT_LIST_HEAD(&q->queue_head);
 255         q->last_merge = NULL;
 256         q->end_sector = 0;
 257         q->boundary_rq = NULL;
 258
 259         if (name) {
 260                 e = elevator_get(name);
 261                 if (!e)
 262                         return -EINVAL;
 263         }
 264
 265         if (!e && *chosen_elevator) {
 266                 e = elevator_get(chosen_elevator);
 267                 if (!e)
 268                         printk(KERN_ERR "I/O scheduler %s not found\n",
 269                                                         chosen_elevator);
 270         }
 271
 272         if (!e) {
 273                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
 274                 if (!e) {
 275                         printk(KERN_ERR
 276                                 "Default I/O scheduler not found. " \
 277                                 "Using noop.\n");
 278                         e = elevator_get("noop");
 279                 }
 280         }
 281
 282         eq = elevator_alloc(q, e);
 283         if (!eq)
 284                 return -ENOMEM;
 285
 286         data = elevator_init_queue(q, eq);
 287         if (!data) {
 288                 kobject_put(&eq->kobj);
 289                 return -ENOMEM;
 290         }
 291
 292         elevator_attach(q, eq, data);
 293         return ret;
 294 }
 295 EXPORT_SYMBOL(elevator_init);
 296
 297 void elevator_exit(struct elevator_queue *e)
 298 {
 299         mutex_lock(&e->sysfs_lock);
 300         if (e->ops->elevator_exit_fn)
 301                 e->ops->elevator_exit_fn(e);
 302         e->ops = NULL;
 303         mutex_unlock(&e->sysfs_lock);
 304
 305         kobject_put(&e->kobj);
 306 }
 307 EXPORT_SYMBOL(elevator_exit);
 308
 309 static inline void __elv_rqhash_del(struct request *rq)
 310 {
 311         hlist_del_init(&rq->hash);
 312 }
 313
 314 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 315 {
 316         if (ELV_ON_HASH(rq))
 317                 __elv_rqhash_del(rq);
 318 }
 319
 320 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 321 {
 322         struct elevator_queue *e = q->elevator;
 323
 324         BUG_ON(ELV_ON_HASH(rq));
 325         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 326 }
 327
 328 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 329 {
 330         __elv_rqhash_del(rq);
 331         elv_rqhash_add(q, rq);
 332 }
 333
 334 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 335 {
 336         struct elevator_queue *e = q->elevator;
 337         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 338         struct hlist_node *entry, *next;
 339         struct request *rq;
 340
 341         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 342                 BUG_ON(!ELV_ON_HASH(rq));
 343
 344                 if (unlikely(!rq_mergeable(rq))) {
 345                         __elv_rqhash_del(rq);
 346                         continue;
 347                 }
 348
 349                 if (rq_hash_key(rq) == offset)
 350                         return rq;
 351         }
 352
 353         return NULL;
 354 }
 355
 356 /*
 357  * RB-tree support functions for inserting/lookup/removal of requests
 358  * in a sorted RB tree.
 359  */
 360 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 361 {
 362         struct rb_node **p = &root->rb_node;
 363         struct rb_node *parent = NULL;
 364         struct request *__rq;
 365
 366         while (*p) {
 367                 parent = *p;
 368                 __rq = rb_entry(parent, struct request, rb_node);
 369
 370                 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
 371                         p = &(*p)->rb_left;
 372                 else if (blk_rq_pos(rq) > blk_rq_pos(__rq))
 373                         p = &(*p)->rb_right;
 374                 else
 375                         return __rq;
 376         }
 377
 378         rb_link_node(&rq->rb_node, parent, p);
 379         rb_insert_color(&rq->rb_node, root);
 380         return NULL;
 381 }
 382 EXPORT_SYMBOL(elv_rb_add);
 383
 384 void elv_rb_del(struct rb_root *root, struct request *rq)
 385 {
 386         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 387         rb_erase(&rq->rb_node, root);
 388         RB_CLEAR_NODE(&rq->rb_node);
 389 }
 390 EXPORT_SYMBOL(elv_rb_del);
 391
 392 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 393 {
 394         struct rb_node *n = root->rb_node;
 395         struct request *rq;
 396
 397         while (n) {
 398                 rq = rb_entry(n, struct request, rb_node);
 399
 400                 if (sector < blk_rq_pos(rq))
 401                         n = n->rb_left;
 402                 else if (sector > blk_rq_pos(rq))
 403                         n = n->rb_right;
 404                 else
 405                         return rq;
 406         }
 407
 408         return NULL;
 409 }
 410 EXPORT_SYMBOL(elv_rb_find);
 411
 412 /*
 413  * Insert rq into dispatch queue of q.  Queue lock must be held on
 414  * entry.  rq is sort instead into the dispatch queue. To be used by
 415  * specific elevators.
 416  */
 417 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 418 {
 419         sector_t boundary;
 420         struct list_head *entry;
 421         int stop_flags;
 422
 423         if (q->last_merge == rq)
 424                 q->last_merge = NULL;
 425
 426         elv_rqhash_del(q, rq);
 427
 428         q->nr_sorted--;
 429
 430         boundary = q->end_sector;
 431         stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
 432         list_for_each_prev(entry, &q->queue_head) {
 433                 struct request *pos = list_entry_rq(entry);
 434
 435                 if (blk_discard_rq(rq) != blk_discard_rq(pos))
 436                         break;
 437                 if (rq_data_dir(rq) != rq_data_dir(pos))
 438                         break;
 439                 if (pos->cmd_flags & stop_flags)
 440                         break;
 441                 if (blk_rq_pos(rq) >= boundary) {
 442                         if (blk_rq_pos(pos) < boundary)
 443                                 continue;
 444                 } else {
 445                         if (blk_rq_pos(pos) >= boundary)
 446                                 break;
 447                 }
 448                 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
 449                         break;
 450         }
 451
 452         list_add(&rq->queuelist, entry);
 453 }
 454 EXPORT_SYMBOL(elv_dispatch_sort);
 455
 456 /*
 457  * Insert rq into dispatch queue of q.  Queue lock must be held on
 458  * entry.  rq is added to the back of the dispatch queue. To be used by
 459  * specific elevators.
 460  */
 461 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 462 {
 463         if (q->last_merge == rq)
 464                 q->last_merge = NULL;
 465
 466         elv_rqhash_del(q, rq);
 467
 468         q->nr_sorted--;
 469
 470         q->end_sector = rq_end_sector(rq);
 471         q->boundary_rq = rq;
 472         list_add_tail(&rq->queuelist, &q->queue_head);
 473 }
 474 EXPORT_SYMBOL(elv_dispatch_add_tail);
 475
 476 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 477 {
 478         struct elevator_queue *e = q->elevator;
 479         struct request *__rq;
 480         int ret;
 481
 482         /*
 483          * First try one-hit cache.
 484          */
 485         if (q->last_merge) {
 486                 ret = elv_try_merge(q->last_merge, bio);
 487                 if (ret != ELEVATOR_NO_MERGE) {
 488                         *req = q->last_merge;
 489                         return ret;
 490                 }
 491         }
 492
 493         if (blk_queue_nomerges(q))
 494                 return ELEVATOR_NO_MERGE;
 495
 496         /*
 497          * See if our hash lookup can find a potential backmerge.
 498          */
 499         __rq = elv_rqhash_find(q, bio->bi_sector);
 500         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 501                 *req = __rq;
 502                 return ELEVATOR_BACK_MERGE;
 503         }
 504
 505         if (e->ops->elevator_merge_fn)
 506                 return e->ops->elevator_merge_fn(q, req, bio);
 507
 508         return ELEVATOR_NO_MERGE;
 509 }
 510
 511 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 512 {
 513         struct elevator_queue *e = q->elevator;
 514
 515         if (e->ops->elevator_merged_fn)
 516                 e->ops->elevator_merged_fn(q, rq, type);
 517
 518         if (type == ELEVATOR_BACK_MERGE)
 519                 elv_rqhash_reposition(q, rq);
 520
 521         q->last_merge = rq;
 522 }
 523
 524 void elv_merge_requests(struct request_queue *q, struct request *rq,
 525                              struct request *next)
 526 {
 527         struct elevator_queue *e = q->elevator;
 528
 529         if (e->ops->elevator_merge_req_fn)
 530                 e->ops->elevator_merge_req_fn(q, rq, next);
 531
 532         elv_rqhash_reposition(q, rq);
 533         elv_rqhash_del(q, next);
 534
 535         q->nr_sorted--;
 536         q->last_merge = rq;
 537 }
 538
 539 void elv_requeue_request(struct request_queue *q, struct request *rq)
 540 {
 541         /*
 542          * it already went through dequeue, we need to decrement the
 543          * in_flight count again
 544          */
 545         if (blk_account_rq(rq)) {
 546                 q->in_flight[rq_is_sync(rq)]--;
 547                 if (blk_sorted_rq(rq))
 548                         elv_deactivate_rq(q, rq);
 549         }
 550
 551         rq->cmd_flags &= ~REQ_STARTED;
 552
 553         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 554 }
 555
 556 void elv_drain_elevator(struct request_queue *q)
 557 {
 558         static int printed;
 559         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 560                 ;
 561         if (q->nr_sorted == 0)
 562                 return;
 563         if (printed++ < 10) {
 564                 printk(KERN_ERR "%s: forced dispatching is broken "
 565                        "(nr_sorted=%u), please report this\n",
 566                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 567         }
 568 }
 569
 570 /*
 571  * Call with queue lock held, interrupts disabled
 572  */
 573 void elv_quiesce_start(struct request_queue *q)
 574 {
 575         if (!q->elevator)
 576                 return;
 577
 578         queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
 579
 580         /*
 581          * make sure we don't have any requests in flight
 582          */
 583         elv_drain_elevator(q);
 584         while (q->rq.elvpriv) {
 585                 __blk_run_queue(q);
 586                 spin_unlock_irq(q->queue_lock);
 587                 msleep(10);
 588                 spin_lock_irq(q->queue_lock);
 589                 elv_drain_elevator(q);
 590         }
 591 }
 592
 593 void elv_quiesce_end(struct request_queue *q)
 594 {
 595         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
 596 }
 597
 598 void elv_insert(struct request_queue *q, struct request *rq, int where)
 599 {
 600         struct list_head *pos;
 601         unsigned ordseq;
 602         int unplug_it = 1;
 603
 604         trace_block_rq_insert(q, rq);
 605
 606         rq->q = q;
 607
 608         switch (where) {
 609         case ELEVATOR_INSERT_FRONT:
 610                 rq->cmd_flags |= REQ_SOFTBARRIER;
 611
 612                 list_add(&rq->queuelist, &q->queue_head);
 613                 break;
 614
 615         case ELEVATOR_INSERT_BACK:
 616                 rq->cmd_flags |= REQ_SOFTBARRIER;
 617                 elv_drain_elevator(q);
 618                 list_add_tail(&rq->queuelist, &q->queue_head);
 619                 /*
 620                  * We kick the queue here for the following reasons.
 621                  * - The elevator might have returned NULL previously
 622                  *   to delay requests and returned them now.  As the
 623                  *   queue wasn't empty before this request, ll_rw_blk
 624                  *   won't run the queue on return, resulting in hang.
 625                  * - Usually, back inserted requests won't be merged
 626                  *   with anything.  There's no point in delaying queue
 627                  *   processing.
 628                  */
 629                 __blk_run_queue(q);
 630                 break;
 631
 632         case ELEVATOR_INSERT_SORT:
 633                 BUG_ON(!blk_fs_request(rq) && !blk_discard_rq(rq));
 634                 rq->cmd_flags |= REQ_SORTED;
 635                 q->nr_sorted++;
 636                 if (rq_mergeable(rq)) {
 637                         elv_rqhash_add(q, rq);
 638                         if (!q->last_merge)
 639                                 q->last_merge = rq;
 640                 }
 641
 642                 /*
 643                  * Some ioscheds (cfq) run q->request_fn directly, so
 644                  * rq cannot be accessed after calling
 645                  * elevator_add_req_fn.
 646                  */
 647                 q->elevator->ops->elevator_add_req_fn(q, rq);
 648                 break;
 649
 650         case ELEVATOR_INSERT_REQUEUE:
 651                 /*
 652                  * If ordered flush isn't in progress, we do front
 653                  * insertion; otherwise, requests should be requeued
 654                  * in ordseq order.
 655                  */
 656                 rq->cmd_flags |= REQ_SOFTBARRIER;
 657
 658                 /*
 659                  * Most requeues happen because of a busy condition,
 660                  * don't force unplug of the queue for that case.
 661                  */
 662                 unplug_it = 0;
 663
 664                 if (q->ordseq == 0) {
 665                         list_add(&rq->queuelist, &q->queue_head);
 666                         break;
 667                 }
 668
 669                 ordseq = blk_ordered_req_seq(rq);
 670
 671                 list_for_each(pos, &q->queue_head) {
 672                         struct request *pos_rq = list_entry_rq(pos);
 673                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 674                                 break;
 675                 }
 676
 677                 list_add_tail(&rq->queuelist, pos);
 678                 break;
 679
 680         default:
 681                 printk(KERN_ERR "%s: bad insertion point %d\n",
 682                        __func__, where);
 683                 BUG();
 684         }
 685
 686         if (unplug_it && blk_queue_plugged(q)) {
 687                 int nrq = q->rq.count[BLK_RW_SYNC] + q->rq.count[BLK_RW_ASYNC]
 688                                 - queue_in_flight(q);
 689
 690                 if (nrq >= q->unplug_thresh)
 691                         __generic_unplug_device(q);
 692         }
 693 }
 694
 695 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
 696                        int plug)
 697 {
 698         if (q->ordcolor)
 699                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 700
 701         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 702                 /*
 703                  * toggle ordered color
 704                  */
 705                 if (blk_barrier_rq(rq))
 706                         q->ordcolor ^= 1;
 707
 708                 /*
 709                  * barriers implicitly indicate back insertion
 710                  */
 711                 if (where == ELEVATOR_INSERT_SORT)
 712                         where = ELEVATOR_INSERT_BACK;
 713
 714                 /*
 715                  * this request is scheduling boundary, update
 716                  * end_sector
 717                  */
 718                 if (blk_fs_request(rq) || blk_discard_rq(rq)) {
 719                         q->end_sector = rq_end_sector(rq);
 720                         q->boundary_rq = rq;
 721                 }
 722         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 723                     where == ELEVATOR_INSERT_SORT)
 724                 where = ELEVATOR_INSERT_BACK;
 725
 726         if (plug)
 727                 blk_plug_device(q);
 728
 729         elv_insert(q, rq, where);
 730 }
 731 EXPORT_SYMBOL(__elv_add_request);
 732
 733 void elv_add_request(struct request_queue *q, struct request *rq, int where,
 734                      int plug)
 735 {
 736         unsigned long flags;
 737
 738         spin_lock_irqsave(q->queue_lock, flags);
 739         __elv_add_request(q, rq, where, plug);
 740         spin_unlock_irqrestore(q->queue_lock, flags);
 741 }
 742 EXPORT_SYMBOL(elv_add_request);
 743
 744 int elv_queue_empty(struct request_queue *q)
 745 {
 746         struct elevator_queue *e = q->elevator;
 747
 748         if (!list_empty(&q->queue_head))
 749                 return 0;
 750
 751         if (e->ops->elevator_queue_empty_fn)
 752                 return e->ops->elevator_queue_empty_fn(q);
 753
 754         return 1;
 755 }
 756 EXPORT_SYMBOL(elv_queue_empty);
 757
 758 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 759 {
 760         struct elevator_queue *e = q->elevator;
 761
 762         if (e->ops->elevator_latter_req_fn)
 763                 return e->ops->elevator_latter_req_fn(q, rq);
 764         return NULL;
 765 }
 766
 767 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 768 {
 769         struct elevator_queue *e = q->elevator;
 770
 771         if (e->ops->elevator_former_req_fn)
 772                 return e->ops->elevator_former_req_fn(q, rq);
 773         return NULL;
 774 }
 775
 776 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 777 {
 778         struct elevator_queue *e = q->elevator;
 779
 780         if (e->ops->elevator_set_req_fn)
 781                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 782
 783         rq->elevator_private = NULL;
 784         return 0;
 785 }
 786
 787 void elv_put_request(struct request_queue *q, struct request *rq)
 788 {
 789         struct elevator_queue *e = q->elevator;
 790
 791         if (e->ops->elevator_put_req_fn)
 792                 e->ops->elevator_put_req_fn(rq);
 793 }
 794
 795 int elv_may_queue(struct request_queue *q, int rw)
 796 {
 797         struct elevator_queue *e = q->elevator;
 798
 799         if (e->ops->elevator_may_queue_fn)
 800                 return e->ops->elevator_may_queue_fn(q, rw);
 801
 802         return ELV_MQUEUE_MAY;
 803 }
 804
 805 void elv_abort_queue(struct request_queue *q)
 806 {
 807         struct request *rq;
 808
 809         while (!list_empty(&q->queue_head)) {
 810                 rq = list_entry_rq(q->queue_head.next);
 811                 rq->cmd_flags |= REQ_QUIET;
 812                 trace_block_rq_abort(q, rq);
 813                 /*
 814                  * Mark this request as started so we don't trigger
 815                  * any debug logic in the end I/O path.
 816                  */
 817                 blk_start_request(rq);
 818                 __blk_end_request_all(rq, -EIO);
 819         }
 820 }
 821 EXPORT_SYMBOL(elv_abort_queue);
 822
 823 void elv_completed_request(struct request_queue *q, struct request *rq)
 824 {
 825         struct elevator_queue *e = q->elevator;
 826
 827         /*
 828          * request is released from the driver, io must be done
 829          */
 830         if (blk_account_rq(rq)) {
 831                 q->in_flight[rq_is_sync(rq)]--;
 832                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 833                         e->ops->elevator_completed_req_fn(q, rq);
 834         }
 835
 836         /*
 837          * Check if the queue is waiting for fs requests to be
 838          * drained for flush sequence.
 839          */
 840         if (unlikely(q->ordseq)) {
 841                 struct request *next = NULL;
 842
 843                 if (!list_empty(&q->queue_head))
 844                         next = list_entry_rq(q->queue_head.next);
 845
 846                 if (!queue_in_flight(q) &&
 847                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 848                     (!next || blk_ordered_req_seq(next) > QUEUE_ORDSEQ_DRAIN)) {
 849                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 850                         __blk_run_queue(q);
 851                 }
 852         }
 853 }
 854
 855 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 856
 857 static ssize_t
 858 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 859 {
 860         struct elv_fs_entry *entry = to_elv(attr);
 861         struct elevator_queue *e;
 862         ssize_t error;
 863
 864         if (!entry->show)
 865                 return -EIO;
 866
 867         e = container_of(kobj, struct elevator_queue, kobj);
 868         mutex_lock(&e->sysfs_lock);
 869         error = e->ops ? entry->show(e, page) : -ENOENT;
 870         mutex_unlock(&e->sysfs_lock);
 871         return error;
 872 }
 873
 874 static ssize_t
 875 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 876                const char *page, size_t length)
 877 {
 878         struct elv_fs_entry *entry = to_elv(attr);
 879         struct elevator_queue *e;
 880         ssize_t error;
 881
 882         if (!entry->store)
 883                 return -EIO;
 884
 885         e = container_of(kobj, struct elevator_queue, kobj);
 886         mutex_lock(&e->sysfs_lock);
 887         error = e->ops ? entry->store(e, page, length) : -ENOENT;
 888         mutex_unlock(&e->sysfs_lock);
 889         return error;
 890 }
 891
 892 static struct sysfs_ops elv_sysfs_ops = {
 893         .show   = elv_attr_show,
 894         .store  = elv_attr_store,
 895 };
 896
 897 static struct kobj_type elv_ktype = {
 898         .sysfs_ops      = &elv_sysfs_ops,
 899         .release        = elevator_release,
 900 };
 901
 902 int elv_register_queue(struct request_queue *q)
 903 {
 904         struct elevator_queue *e = q->elevator;
 905         int error;
 906
 907         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 908         if (!error) {
 909                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
 910                 if (attr) {
 911                         while (attr->attr.name) {
 912                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 913                                         break;
 914                                 attr++;
 915                         }
 916                 }
 917                 kobject_uevent(&e->kobj, KOBJ_ADD);
 918         }
 919         return error;
 920 }
 921
 922 static void __elv_unregister_queue(struct elevator_queue *e)
 923 {
 924         kobject_uevent(&e->kobj, KOBJ_REMOVE);
 925         kobject_del(&e->kobj);
 926 }
 927
 928 void elv_unregister_queue(struct request_queue *q)
 929 {
 930         if (q)
 931                 __elv_unregister_queue(q->elevator);
 932 }
 933
 934 void elv_register(struct elevator_type *e)
 935 {
 936         char *def = "";
 937
 938         spin_lock(&elv_list_lock);
 939         BUG_ON(elevator_find(e->elevator_name));
 940         list_add_tail(&e->list, &elv_list);
 941         spin_unlock(&elv_list_lock);
 942
 943         if (!strcmp(e->elevator_name, chosen_elevator) ||
 944                         (!*chosen_elevator &&
 945                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 946                                 def = " (default)";
 947
 948         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
 949                                                                 def);
 950 }
 951 EXPORT_SYMBOL_GPL(elv_register);
 952
 953 void elv_unregister(struct elevator_type *e)
 954 {
 955         struct task_struct *g, *p;
 956
 957         /*
 958          * Iterate every thread in the process to remove the io contexts.
 959          */
 960         if (e->ops.trim) {
 961                 read_lock(&tasklist_lock);
 962                 do_each_thread(g, p) {
 963                         task_lock(p);
 964                         if (p->io_context)
 965                                 e->ops.trim(p->io_context);
 966                         task_unlock(p);
 967                 } while_each_thread(g, p);
 968                 read_unlock(&tasklist_lock);
 969         }
 970
 971         spin_lock(&elv_list_lock);
 972         list_del_init(&e->list);
 973         spin_unlock(&elv_list_lock);
 974 }
 975 EXPORT_SYMBOL_GPL(elv_unregister);
 976
 977 /*
 978  * switch to new_e io scheduler. be careful not to introduce deadlocks -
 979  * we don't free the old io scheduler, before we have allocated what we
 980  * need for the new one. this way we have a chance of going back to the old
 981  * one, if the new one fails init for some reason.
 982  */
 983 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
 984 {
 985         struct elevator_queue *old_elevator, *e;
 986         void *data;
 987
 988         /*
 989          * Allocate new elevator
 990          */
 991         e = elevator_alloc(q, new_e);
 992         if (!e)
 993                 return 0;
 994
 995         data = elevator_init_queue(q, e);
 996         if (!data) {
 997                 kobject_put(&e->kobj);
 998                 return 0;
 999         }
1000
1001         /*
1002          * Turn on BYPASS and drain all requests w/ elevator private data
1003          */
1004         spin_lock_irq(q->queue_lock);
1005         elv_quiesce_start(q);
1006
1007         /*
1008          * Remember old elevator.
1009          */
1010         old_elevator = q->elevator;
1011
1012         /*
1013          * attach and start new elevator
1014          */
1015         elevator_attach(q, e, data);
1016
1017         spin_unlock_irq(q->queue_lock);
1018
1019         __elv_unregister_queue(old_elevator);
1020
1021         if (elv_register_queue(q))
1022                 goto fail_register;
1023
1024         /*
1025          * finally exit old elevator and turn off BYPASS.
1026          */
1027         elevator_exit(old_elevator);
1028         spin_lock_irq(q->queue_lock);
1029         elv_quiesce_end(q);
1030         spin_unlock_irq(q->queue_lock);
1031
1032         blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
1033
1034         return 1;
1035
1036 fail_register:
1037         /*
1038          * switch failed, exit the new io scheduler and reattach the old
1039          * one again (along with re-adding the sysfs dir)
1040          */
1041         elevator_exit(e);
1042         q->elevator = old_elevator;
1043         elv_register_queue(q);
1044
1045         spin_lock_irq(q->queue_lock);
1046         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1047         spin_unlock_irq(q->queue_lock);
1048
1049         return 0;
1050 }
1051
1052 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1053                           size_t count)
1054 {
1055         char elevator_name[ELV_NAME_MAX];
1056         struct elevator_type *e;
1057
1058         if (!q->elevator)
1059                 return count;
1060
1061         strlcpy(elevator_name, name, sizeof(elevator_name));
1062         strstrip(elevator_name);
1063
1064         e = elevator_get(elevator_name);
1065         if (!e) {
1066                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1067                 return -EINVAL;
1068         }
1069
1070         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1071                 elevator_put(e);
1072                 return count;
1073         }
1074
1075         if (!elevator_switch(q, e))
1076                 printk(KERN_ERR "elevator: switch to %s failed\n",
1077                                                         elevator_name);
1078         return count;
1079 }
1080
1081 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1082 {
1083         struct elevator_queue *e = q->elevator;
1084         struct elevator_type *elv;
1085         struct elevator_type *__e;
1086         int len = 0;
1087
1088         if (!q->elevator)
1089                 return sprintf(name, "none\n");
1090
1091         elv = e->elevator_type;
1092
1093         spin_lock(&elv_list_lock);
1094         list_for_each_entry(__e, &elv_list, list) {
1095                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1096                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1097                 else
1098                         len += sprintf(name+len, "%s ", __e->elevator_name);
1099         }
1100         spin_unlock(&elv_list_lock);
1101
1102         len += sprintf(len+name, "\n");
1103         return len;
1104 }
1105
1106 struct request *elv_rb_former_request(struct request_queue *q,
1107                                       struct request *rq)
1108 {
1109         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1110
1111         if (rbprev)
1112                 return rb_entry_rq(rbprev);
1113
1114         return NULL;
1115 }
1116 EXPORT_SYMBOL(elv_rb_former_request);
1117
1118 struct request *elv_rb_latter_request(struct request_queue *q,
1119                                       struct request *rq)
1120 {
1121         struct rb_node *rbnext = rb_next(&rq->rb_node);
1122
1123         if (rbnext)
1124                 return rb_entry_rq(rbnext);
1125
1126         return NULL;
1127 }
1128 EXPORT_SYMBOL(elv_rb_latter_request);