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
  38 #include <asm/uaccess.h>
  39
  40 static DEFINE_SPINLOCK(elv_list_lock);
  41 static LIST_HEAD(elv_list);
  42
  43 /*
  44  * Merge hash stuff.
  45  */
  46 static const int elv_hash_shift = 6;
  47 #define ELV_HASH_BLOCK(sec)     ((sec) >> 3)
  48 #define ELV_HASH_FN(sec)        (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
  49 #define ELV_HASH_ENTRIES        (1 << elv_hash_shift)
  50 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
  51 #define ELV_ON_HASH(rq)         (!hlist_unhashed(&(rq)->hash))
  52
  53 /*
  54  * Query io scheduler to see if the current process issuing bio may be
  55  * merged with rq.
  56  */
  57 static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
  58 {
  59         request_queue_t *q = rq->q;
  60         elevator_t *e = q->elevator;
  61
  62         if (e->ops->elevator_allow_merge_fn)
  63                 return e->ops->elevator_allow_merge_fn(q, rq, bio);
  64
  65         return 1;
  66 }
  67
  68 /*
  69  * can we safely merge with this request?
  70  */
  71 inline int elv_rq_merge_ok(struct request *rq, struct bio *bio)
  72 {
  73         if (!rq_mergeable(rq))
  74                 return 0;
  75
  76         /*
  77          * different data direction or already started, don't merge
  78          */
  79         if (bio_data_dir(bio) != rq_data_dir(rq))
  80                 return 0;
  81
  82         /*
  83          * must be same device and not a special request
  84          */
  85         if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
  86                 return 0;
  87
  88         if (!elv_iosched_allow_merge(rq, bio))
  89                 return 0;
  90
  91         return 1;
  92 }
  93 EXPORT_SYMBOL(elv_rq_merge_ok);
  94
  95 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
  96 {
  97         int ret = ELEVATOR_NO_MERGE;
  98
  99         /*
 100          * we can merge and sequence is ok, check if it's possible
 101          */
 102         if (elv_rq_merge_ok(__rq, bio)) {
 103                 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
 104                         ret = ELEVATOR_BACK_MERGE;
 105                 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
 106                         ret = ELEVATOR_FRONT_MERGE;
 107         }
 108
 109         return ret;
 110 }
 111
 112 static struct elevator_type *elevator_find(const char *name)
 113 {
 114         struct elevator_type *e;
 115
 116         list_for_each_entry(e, &elv_list, list) {
 117                 if (!strcmp(e->elevator_name, name))
 118                         return e;
 119         }
 120
 121         return NULL;
 122 }
 123
 124 static void elevator_put(struct elevator_type *e)
 125 {
 126         module_put(e->elevator_owner);
 127 }
 128
 129 static struct elevator_type *elevator_get(const char *name)
 130 {
 131         struct elevator_type *e;
 132
 133         spin_lock(&elv_list_lock);
 134
 135         e = elevator_find(name);
 136         if (e && !try_module_get(e->elevator_owner))
 137                 e = NULL;
 138
 139         spin_unlock(&elv_list_lock);
 140
 141         return e;
 142 }
 143
 144 static void *elevator_init_queue(request_queue_t *q, struct elevator_queue *eq)
 145 {
 146         return eq->ops->elevator_init_fn(q);
 147 }
 148
 149 static void elevator_attach(request_queue_t *q, struct elevator_queue *eq,
 150                            void *data)
 151 {
 152         q->elevator = eq;
 153         eq->elevator_data = data;
 154 }
 155
 156 static char chosen_elevator[16];
 157
 158 static int __init elevator_setup(char *str)
 159 {
 160         /*
 161          * Be backwards-compatible with previous kernels, so users
 162          * won't get the wrong elevator.
 163          */
 164         if (!strcmp(str, "as"))
 165                 strcpy(chosen_elevator, "anticipatory");
 166         else
 167                 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 168         return 1;
 169 }
 170
 171 __setup("elevator=", elevator_setup);
 172
 173 static struct kobj_type elv_ktype;
 174
 175 static elevator_t *elevator_alloc(request_queue_t *q, struct elevator_type *e)
 176 {
 177         elevator_t *eq;
 178         int i;
 179
 180         eq = kmalloc_node(sizeof(elevator_t), GFP_KERNEL | __GFP_ZERO, q->node);
 181         if (unlikely(!eq))
 182                 goto err;
 183
 184         eq->ops = &e->ops;
 185         eq->elevator_type = e;
 186         kobject_init(&eq->kobj);
 187         snprintf(eq->kobj.name, KOBJ_NAME_LEN, "%s", "iosched");
 188         eq->kobj.ktype = &elv_ktype;
 189         mutex_init(&eq->sysfs_lock);
 190
 191         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 192                                         GFP_KERNEL, q->node);
 193         if (!eq->hash)
 194                 goto err;
 195
 196         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 197                 INIT_HLIST_HEAD(&eq->hash[i]);
 198
 199         return eq;
 200 err:
 201         kfree(eq);
 202         elevator_put(e);
 203         return NULL;
 204 }
 205
 206 static void elevator_release(struct kobject *kobj)
 207 {
 208         elevator_t *e = container_of(kobj, elevator_t, kobj);
 209
 210         elevator_put(e->elevator_type);
 211         kfree(e->hash);
 212         kfree(e);
 213 }
 214
 215 int elevator_init(request_queue_t *q, char *name)
 216 {
 217         struct elevator_type *e = NULL;
 218         struct elevator_queue *eq;
 219         int ret = 0;
 220         void *data;
 221
 222         INIT_LIST_HEAD(&q->queue_head);
 223         q->last_merge = NULL;
 224         q->end_sector = 0;
 225         q->boundary_rq = NULL;
 226
 227         if (name && !(e = elevator_get(name)))
 228                 return -EINVAL;
 229
 230         if (!e && *chosen_elevator && !(e = elevator_get(chosen_elevator)))
 231                 printk("I/O scheduler %s not found\n", chosen_elevator);
 232
 233         if (!e && !(e = elevator_get(CONFIG_DEFAULT_IOSCHED))) {
 234                 printk("Default I/O scheduler not found, using no-op\n");
 235                 e = elevator_get("noop");
 236         }
 237
 238         eq = elevator_alloc(q, e);
 239         if (!eq)
 240                 return -ENOMEM;
 241
 242         data = elevator_init_queue(q, eq);
 243         if (!data) {
 244                 kobject_put(&eq->kobj);
 245                 return -ENOMEM;
 246         }
 247
 248         elevator_attach(q, eq, data);
 249         return ret;
 250 }
 251
 252 EXPORT_SYMBOL(elevator_init);
 253
 254 void elevator_exit(elevator_t *e)
 255 {
 256         mutex_lock(&e->sysfs_lock);
 257         if (e->ops->elevator_exit_fn)
 258                 e->ops->elevator_exit_fn(e);
 259         e->ops = NULL;
 260         mutex_unlock(&e->sysfs_lock);
 261
 262         kobject_put(&e->kobj);
 263 }
 264
 265 EXPORT_SYMBOL(elevator_exit);
 266
 267 static void elv_activate_rq(request_queue_t *q, struct request *rq)
 268 {
 269         elevator_t *e = q->elevator;
 270
 271         if (e->ops->elevator_activate_req_fn)
 272                 e->ops->elevator_activate_req_fn(q, rq);
 273 }
 274
 275 static void elv_deactivate_rq(request_queue_t *q, struct request *rq)
 276 {
 277         elevator_t *e = q->elevator;
 278
 279         if (e->ops->elevator_deactivate_req_fn)
 280                 e->ops->elevator_deactivate_req_fn(q, rq);
 281 }
 282
 283 static inline void __elv_rqhash_del(struct request *rq)
 284 {
 285         hlist_del_init(&rq->hash);
 286 }
 287
 288 static void elv_rqhash_del(request_queue_t *q, struct request *rq)
 289 {
 290         if (ELV_ON_HASH(rq))
 291                 __elv_rqhash_del(rq);
 292 }
 293
 294 static void elv_rqhash_add(request_queue_t *q, struct request *rq)
 295 {
 296         elevator_t *e = q->elevator;
 297
 298         BUG_ON(ELV_ON_HASH(rq));
 299         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 300 }
 301
 302 static void elv_rqhash_reposition(request_queue_t *q, struct request *rq)
 303 {
 304         __elv_rqhash_del(rq);
 305         elv_rqhash_add(q, rq);
 306 }
 307
 308 static struct request *elv_rqhash_find(request_queue_t *q, sector_t offset)
 309 {
 310         elevator_t *e = q->elevator;
 311         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 312         struct hlist_node *entry, *next;
 313         struct request *rq;
 314
 315         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 316                 BUG_ON(!ELV_ON_HASH(rq));
 317
 318                 if (unlikely(!rq_mergeable(rq))) {
 319                         __elv_rqhash_del(rq);
 320                         continue;
 321                 }
 322
 323                 if (rq_hash_key(rq) == offset)
 324                         return rq;
 325         }
 326
 327         return NULL;
 328 }
 329
 330 /*
 331  * RB-tree support functions for inserting/lookup/removal of requests
 332  * in a sorted RB tree.
 333  */
 334 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 335 {
 336         struct rb_node **p = &root->rb_node;
 337         struct rb_node *parent = NULL;
 338         struct request *__rq;
 339
 340         while (*p) {
 341                 parent = *p;
 342                 __rq = rb_entry(parent, struct request, rb_node);
 343
 344                 if (rq->sector < __rq->sector)
 345                         p = &(*p)->rb_left;
 346                 else if (rq->sector > __rq->sector)
 347                         p = &(*p)->rb_right;
 348                 else
 349                         return __rq;
 350         }
 351
 352         rb_link_node(&rq->rb_node, parent, p);
 353         rb_insert_color(&rq->rb_node, root);
 354         return NULL;
 355 }
 356
 357 EXPORT_SYMBOL(elv_rb_add);
 358
 359 void elv_rb_del(struct rb_root *root, struct request *rq)
 360 {
 361         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 362         rb_erase(&rq->rb_node, root);
 363         RB_CLEAR_NODE(&rq->rb_node);
 364 }
 365
 366 EXPORT_SYMBOL(elv_rb_del);
 367
 368 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 369 {
 370         struct rb_node *n = root->rb_node;
 371         struct request *rq;
 372
 373         while (n) {
 374                 rq = rb_entry(n, struct request, rb_node);
 375
 376                 if (sector < rq->sector)
 377                         n = n->rb_left;
 378                 else if (sector > rq->sector)
 379                         n = n->rb_right;
 380                 else
 381                         return rq;
 382         }
 383
 384         return NULL;
 385 }
 386
 387 EXPORT_SYMBOL(elv_rb_find);
 388
 389 /*
 390  * Insert rq into dispatch queue of q.  Queue lock must be held on
 391  * entry.  rq is sort insted into the dispatch queue. To be used by
 392  * specific elevators.
 393  */
 394 void elv_dispatch_sort(request_queue_t *q, struct request *rq)
 395 {
 396         sector_t boundary;
 397         struct list_head *entry;
 398
 399         if (q->last_merge == rq)
 400                 q->last_merge = NULL;
 401
 402         elv_rqhash_del(q, rq);
 403
 404         q->nr_sorted--;
 405
 406         boundary = q->end_sector;
 407
 408         list_for_each_prev(entry, &q->queue_head) {
 409                 struct request *pos = list_entry_rq(entry);
 410
 411                 if (rq_data_dir(rq) != rq_data_dir(pos))
 412                         break;
 413                 if (pos->cmd_flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
 414                         break;
 415                 if (rq->sector >= boundary) {
 416                         if (pos->sector < boundary)
 417                                 continue;
 418                 } else {
 419                         if (pos->sector >= boundary)
 420                                 break;
 421                 }
 422                 if (rq->sector >= pos->sector)
 423                         break;
 424         }
 425
 426         list_add(&rq->queuelist, entry);
 427 }
 428
 429 EXPORT_SYMBOL(elv_dispatch_sort);
 430
 431 /*
 432  * Insert rq into dispatch queue of q.  Queue lock must be held on
 433  * entry.  rq is added to the back of the dispatch queue. To be used by
 434  * specific elevators.
 435  */
 436 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 437 {
 438         if (q->last_merge == rq)
 439                 q->last_merge = NULL;
 440
 441         elv_rqhash_del(q, rq);
 442
 443         q->nr_sorted--;
 444
 445         q->end_sector = rq_end_sector(rq);
 446         q->boundary_rq = rq;
 447         list_add_tail(&rq->queuelist, &q->queue_head);
 448 }
 449
 450 EXPORT_SYMBOL(elv_dispatch_add_tail);
 451
 452 int elv_merge(request_queue_t *q, struct request **req, struct bio *bio)
 453 {
 454         elevator_t *e = q->elevator;
 455         struct request *__rq;
 456         int ret;
 457
 458         /*
 459          * First try one-hit cache.
 460          */
 461         if (q->last_merge) {
 462                 ret = elv_try_merge(q->last_merge, bio);
 463                 if (ret != ELEVATOR_NO_MERGE) {
 464                         *req = q->last_merge;
 465                         return ret;
 466                 }
 467         }
 468
 469         /*
 470          * See if our hash lookup can find a potential backmerge.
 471          */
 472         __rq = elv_rqhash_find(q, bio->bi_sector);
 473         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 474                 *req = __rq;
 475                 return ELEVATOR_BACK_MERGE;
 476         }
 477
 478         if (e->ops->elevator_merge_fn)
 479                 return e->ops->elevator_merge_fn(q, req, bio);
 480
 481         return ELEVATOR_NO_MERGE;
 482 }
 483
 484 void elv_merged_request(request_queue_t *q, struct request *rq, int type)
 485 {
 486         elevator_t *e = q->elevator;
 487
 488         if (e->ops->elevator_merged_fn)
 489                 e->ops->elevator_merged_fn(q, rq, type);
 490
 491         if (type == ELEVATOR_BACK_MERGE)
 492                 elv_rqhash_reposition(q, rq);
 493
 494         q->last_merge = rq;
 495 }
 496
 497 void elv_merge_requests(request_queue_t *q, struct request *rq,
 498                              struct request *next)
 499 {
 500         elevator_t *e = q->elevator;
 501
 502         if (e->ops->elevator_merge_req_fn)
 503                 e->ops->elevator_merge_req_fn(q, rq, next);
 504
 505         elv_rqhash_reposition(q, rq);
 506         elv_rqhash_del(q, next);
 507
 508         q->nr_sorted--;
 509         q->last_merge = rq;
 510 }
 511
 512 void elv_requeue_request(request_queue_t *q, struct request *rq)
 513 {
 514         /*
 515          * it already went through dequeue, we need to decrement the
 516          * in_flight count again
 517          */
 518         if (blk_account_rq(rq)) {
 519                 q->in_flight--;
 520                 if (blk_sorted_rq(rq))
 521                         elv_deactivate_rq(q, rq);
 522         }
 523
 524         rq->cmd_flags &= ~REQ_STARTED;
 525
 526         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 527 }
 528
 529 static void elv_drain_elevator(request_queue_t *q)
 530 {
 531         static int printed;
 532         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 533                 ;
 534         if (q->nr_sorted == 0)
 535                 return;
 536         if (printed++ < 10) {
 537                 printk(KERN_ERR "%s: forced dispatching is broken "
 538                        "(nr_sorted=%u), please report this\n",
 539                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 540         }
 541 }
 542
 543 void elv_insert(request_queue_t *q, struct request *rq, int where)
 544 {
 545         struct list_head *pos;
 546         unsigned ordseq;
 547         int unplug_it = 1;
 548
 549         blk_add_trace_rq(q, rq, BLK_TA_INSERT);
 550
 551         rq->q = q;
 552
 553         switch (where) {
 554         case ELEVATOR_INSERT_FRONT:
 555                 rq->cmd_flags |= REQ_SOFTBARRIER;
 556
 557                 list_add(&rq->queuelist, &q->queue_head);
 558                 break;
 559
 560         case ELEVATOR_INSERT_BACK:
 561                 rq->cmd_flags |= REQ_SOFTBARRIER;
 562                 elv_drain_elevator(q);
 563                 list_add_tail(&rq->queuelist, &q->queue_head);
 564                 /*
 565                  * We kick the queue here for the following reasons.
 566                  * - The elevator might have returned NULL previously
 567                  *   to delay requests and returned them now.  As the
 568                  *   queue wasn't empty before this request, ll_rw_blk
 569                  *   won't run the queue on return, resulting in hang.
 570                  * - Usually, back inserted requests won't be merged
 571                  *   with anything.  There's no point in delaying queue
 572                  *   processing.
 573                  */
 574                 blk_remove_plug(q);
 575                 q->request_fn(q);
 576                 break;
 577
 578         case ELEVATOR_INSERT_SORT:
 579                 BUG_ON(!blk_fs_request(rq));
 580                 rq->cmd_flags |= REQ_SORTED;
 581                 q->nr_sorted++;
 582                 if (rq_mergeable(rq)) {
 583                         elv_rqhash_add(q, rq);
 584                         if (!q->last_merge)
 585                                 q->last_merge = rq;
 586                 }
 587
 588                 /*
 589                  * Some ioscheds (cfq) run q->request_fn directly, so
 590                  * rq cannot be accessed after calling
 591                  * elevator_add_req_fn.
 592                  */
 593                 q->elevator->ops->elevator_add_req_fn(q, rq);
 594                 break;
 595
 596         case ELEVATOR_INSERT_REQUEUE:
 597                 /*
 598                  * If ordered flush isn't in progress, we do front
 599                  * insertion; otherwise, requests should be requeued
 600                  * in ordseq order.
 601                  */
 602                 rq->cmd_flags |= REQ_SOFTBARRIER;
 603
 604                 /*
 605                  * Most requeues happen because of a busy condition,
 606                  * don't force unplug of the queue for that case.
 607                  */
 608                 unplug_it = 0;
 609
 610                 if (q->ordseq == 0) {
 611                         list_add(&rq->queuelist, &q->queue_head);
 612                         break;
 613                 }
 614
 615                 ordseq = blk_ordered_req_seq(rq);
 616
 617                 list_for_each(pos, &q->queue_head) {
 618                         struct request *pos_rq = list_entry_rq(pos);
 619                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 620                                 break;
 621                 }
 622
 623                 list_add_tail(&rq->queuelist, pos);
 624                 break;
 625
 626         default:
 627                 printk(KERN_ERR "%s: bad insertion point %d\n",
 628                        __FUNCTION__, where);
 629                 BUG();
 630         }
 631
 632         if (unplug_it && blk_queue_plugged(q)) {
 633                 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
 634                         - q->in_flight;
 635
 636                 if (nrq >= q->unplug_thresh)
 637                         __generic_unplug_device(q);
 638         }
 639 }
 640
 641 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
 642                        int plug)
 643 {
 644         if (q->ordcolor)
 645                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 646
 647         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 648                 /*
 649                  * toggle ordered color
 650                  */
 651                 if (blk_barrier_rq(rq))
 652                         q->ordcolor ^= 1;
 653
 654                 /*
 655                  * barriers implicitly indicate back insertion
 656                  */
 657                 if (where == ELEVATOR_INSERT_SORT)
 658                         where = ELEVATOR_INSERT_BACK;
 659
 660                 /*
 661                  * this request is scheduling boundary, update
 662                  * end_sector
 663                  */
 664                 if (blk_fs_request(rq)) {
 665                         q->end_sector = rq_end_sector(rq);
 666                         q->boundary_rq = rq;
 667                 }
 668         } else if (!(rq->cmd_flags & REQ_ELVPRIV) && where == ELEVATOR_INSERT_SORT)
 669                 where = ELEVATOR_INSERT_BACK;
 670
 671         if (plug)
 672                 blk_plug_device(q);
 673
 674         elv_insert(q, rq, where);
 675 }
 676
 677 EXPORT_SYMBOL(__elv_add_request);
 678
 679 void elv_add_request(request_queue_t *q, struct request *rq, int where,
 680                      int plug)
 681 {
 682         unsigned long flags;
 683
 684         spin_lock_irqsave(q->queue_lock, flags);
 685         __elv_add_request(q, rq, where, plug);
 686         spin_unlock_irqrestore(q->queue_lock, flags);
 687 }
 688
 689 EXPORT_SYMBOL(elv_add_request);
 690
 691 static inline struct request *__elv_next_request(request_queue_t *q)
 692 {
 693         struct request *rq;
 694
 695         while (1) {
 696                 while (!list_empty(&q->queue_head)) {
 697                         rq = list_entry_rq(q->queue_head.next);
 698                         if (blk_do_ordered(q, &rq))
 699                                 return rq;
 700                 }
 701
 702                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
 703                         return NULL;
 704         }
 705 }
 706
 707 struct request *elv_next_request(request_queue_t *q)
 708 {
 709         struct request *rq;
 710         int ret;
 711
 712         while ((rq = __elv_next_request(q)) != NULL) {
 713                 if (!(rq->cmd_flags & REQ_STARTED)) {
 714                         /*
 715                          * This is the first time the device driver
 716                          * sees this request (possibly after
 717                          * requeueing).  Notify IO scheduler.
 718                          */
 719                         if (blk_sorted_rq(rq))
 720                                 elv_activate_rq(q, rq);
 721
 722                         /*
 723                          * just mark as started even if we don't start
 724                          * it, a request that has been delayed should
 725                          * not be passed by new incoming requests
 726                          */
 727                         rq->cmd_flags |= REQ_STARTED;
 728                         blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
 729                 }
 730
 731                 if (!q->boundary_rq || q->boundary_rq == rq) {
 732                         q->end_sector = rq_end_sector(rq);
 733                         q->boundary_rq = NULL;
 734                 }
 735
 736                 if ((rq->cmd_flags & REQ_DONTPREP) || !q->prep_rq_fn)
 737                         break;
 738
 739                 ret = q->prep_rq_fn(q, rq);
 740                 if (ret == BLKPREP_OK) {
 741                         break;
 742                 } else if (ret == BLKPREP_DEFER) {
 743                         /*
 744                          * the request may have been (partially) prepped.
 745                          * we need to keep this request in the front to
 746                          * avoid resource deadlock.  REQ_STARTED will
 747                          * prevent other fs requests from passing this one.
 748                          */
 749                         rq = NULL;
 750                         break;
 751                 } else if (ret == BLKPREP_KILL) {
 752                         int nr_bytes = rq->hard_nr_sectors << 9;
 753
 754                         if (!nr_bytes)
 755                                 nr_bytes = rq->data_len;
 756
 757                         blkdev_dequeue_request(rq);
 758                         rq->cmd_flags |= REQ_QUIET;
 759                         end_that_request_chunk(rq, 0, nr_bytes);
 760                         end_that_request_last(rq, 0);
 761                 } else {
 762                         printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
 763                                                                 ret);
 764                         break;
 765                 }
 766         }
 767
 768         return rq;
 769 }
 770
 771 EXPORT_SYMBOL(elv_next_request);
 772
 773 void elv_dequeue_request(request_queue_t *q, struct request *rq)
 774 {
 775         BUG_ON(list_empty(&rq->queuelist));
 776         BUG_ON(ELV_ON_HASH(rq));
 777
 778         list_del_init(&rq->queuelist);
 779
 780         /*
 781          * the time frame between a request being removed from the lists
 782          * and to it is freed is accounted as io that is in progress at
 783          * the driver side.
 784          */
 785         if (blk_account_rq(rq))
 786                 q->in_flight++;
 787 }
 788
 789 EXPORT_SYMBOL(elv_dequeue_request);
 790
 791 int elv_queue_empty(request_queue_t *q)
 792 {
 793         elevator_t *e = q->elevator;
 794
 795         if (!list_empty(&q->queue_head))
 796                 return 0;
 797
 798         if (e->ops->elevator_queue_empty_fn)
 799                 return e->ops->elevator_queue_empty_fn(q);
 800
 801         return 1;
 802 }
 803
 804 EXPORT_SYMBOL(elv_queue_empty);
 805
 806 struct request *elv_latter_request(request_queue_t *q, struct request *rq)
 807 {
 808         elevator_t *e = q->elevator;
 809
 810         if (e->ops->elevator_latter_req_fn)
 811                 return e->ops->elevator_latter_req_fn(q, rq);
 812         return NULL;
 813 }
 814
 815 struct request *elv_former_request(request_queue_t *q, struct request *rq)
 816 {
 817         elevator_t *e = q->elevator;
 818
 819         if (e->ops->elevator_former_req_fn)
 820                 return e->ops->elevator_former_req_fn(q, rq);
 821         return NULL;
 822 }
 823
 824 int elv_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
 825 {
 826         elevator_t *e = q->elevator;
 827
 828         if (e->ops->elevator_set_req_fn)
 829                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 830
 831         rq->elevator_private = NULL;
 832         return 0;
 833 }
 834
 835 void elv_put_request(request_queue_t *q, struct request *rq)
 836 {
 837         elevator_t *e = q->elevator;
 838
 839         if (e->ops->elevator_put_req_fn)
 840                 e->ops->elevator_put_req_fn(rq);
 841 }
 842
 843 int elv_may_queue(request_queue_t *q, int rw)
 844 {
 845         elevator_t *e = q->elevator;
 846
 847         if (e->ops->elevator_may_queue_fn)
 848                 return e->ops->elevator_may_queue_fn(q, rw);
 849
 850         return ELV_MQUEUE_MAY;
 851 }
 852
 853 void elv_completed_request(request_queue_t *q, struct request *rq)
 854 {
 855         elevator_t *e = q->elevator;
 856
 857         /*
 858          * request is released from the driver, io must be done
 859          */
 860         if (blk_account_rq(rq)) {
 861                 q->in_flight--;
 862                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 863                         e->ops->elevator_completed_req_fn(q, rq);
 864         }
 865
 866         /*
 867          * Check if the queue is waiting for fs requests to be
 868          * drained for flush sequence.
 869          */
 870         if (unlikely(q->ordseq)) {
 871                 struct request *first_rq = list_entry_rq(q->queue_head.next);
 872                 if (q->in_flight == 0 &&
 873                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 874                     blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
 875                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 876                         q->request_fn(q);
 877                 }
 878         }
 879 }
 880
 881 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 882
 883 static ssize_t
 884 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 885 {
 886         elevator_t *e = container_of(kobj, elevator_t, kobj);
 887         struct elv_fs_entry *entry = to_elv(attr);
 888         ssize_t error;
 889
 890         if (!entry->show)
 891                 return -EIO;
 892
 893         mutex_lock(&e->sysfs_lock);
 894         error = e->ops ? entry->show(e, page) : -ENOENT;
 895         mutex_unlock(&e->sysfs_lock);
 896         return error;
 897 }
 898
 899 static ssize_t
 900 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 901                const char *page, size_t length)
 902 {
 903         elevator_t *e = container_of(kobj, elevator_t, kobj);
 904         struct elv_fs_entry *entry = to_elv(attr);
 905         ssize_t error;
 906
 907         if (!entry->store)
 908                 return -EIO;
 909
 910         mutex_lock(&e->sysfs_lock);
 911         error = e->ops ? entry->store(e, page, length) : -ENOENT;
 912         mutex_unlock(&e->sysfs_lock);
 913         return error;
 914 }
 915
 916 static struct sysfs_ops elv_sysfs_ops = {
 917         .show   = elv_attr_show,
 918         .store  = elv_attr_store,
 919 };
 920
 921 static struct kobj_type elv_ktype = {
 922         .sysfs_ops      = &elv_sysfs_ops,
 923         .release        = elevator_release,
 924 };
 925
 926 int elv_register_queue(struct request_queue *q)
 927 {
 928         elevator_t *e = q->elevator;
 929         int error;
 930
 931         e->kobj.parent = &q->kobj;
 932
 933         error = kobject_add(&e->kobj);
 934         if (!error) {
 935                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
 936                 if (attr) {
 937                         while (attr->attr.name) {
 938                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 939                                         break;
 940                                 attr++;
 941                         }
 942                 }
 943                 kobject_uevent(&e->kobj, KOBJ_ADD);
 944         }
 945         return error;
 946 }
 947
 948 static void __elv_unregister_queue(elevator_t *e)
 949 {
 950         kobject_uevent(&e->kobj, KOBJ_REMOVE);
 951         kobject_del(&e->kobj);
 952 }
 953
 954 void elv_unregister_queue(struct request_queue *q)
 955 {
 956         if (q)
 957                 __elv_unregister_queue(q->elevator);
 958 }
 959
 960 int elv_register(struct elevator_type *e)
 961 {
 962         char *def = "";
 963
 964         spin_lock(&elv_list_lock);
 965         BUG_ON(elevator_find(e->elevator_name));
 966         list_add_tail(&e->list, &elv_list);
 967         spin_unlock(&elv_list_lock);
 968
 969         if (!strcmp(e->elevator_name, chosen_elevator) ||
 970                         (!*chosen_elevator &&
 971                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 972                                 def = " (default)";
 973
 974         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name, def);
 975         return 0;
 976 }
 977 EXPORT_SYMBOL_GPL(elv_register);
 978
 979 void elv_unregister(struct elevator_type *e)
 980 {
 981         struct task_struct *g, *p;
 982
 983         /*
 984          * Iterate every thread in the process to remove the io contexts.
 985          */
 986         if (e->ops.trim) {
 987                 read_lock(&tasklist_lock);
 988                 do_each_thread(g, p) {
 989                         task_lock(p);
 990                         if (p->io_context)
 991                                 e->ops.trim(p->io_context);
 992                         task_unlock(p);
 993                 } while_each_thread(g, p);
 994                 read_unlock(&tasklist_lock);
 995         }
 996
 997         spin_lock(&elv_list_lock);
 998         list_del_init(&e->list);
 999         spin_unlock(&elv_list_lock);
1000 }
1001 EXPORT_SYMBOL_GPL(elv_unregister);
1002
1003 /*
1004  * switch to new_e io scheduler. be careful not to introduce deadlocks -
1005  * we don't free the old io scheduler, before we have allocated what we
1006  * need for the new one. this way we have a chance of going back to the old
1007  * one, if the new one fails init for some reason.
1008  */
1009 static int elevator_switch(request_queue_t *q, struct elevator_type *new_e)
1010 {
1011         elevator_t *old_elevator, *e;
1012         void *data;
1013
1014         /*
1015          * Allocate new elevator
1016          */
1017         e = elevator_alloc(q, new_e);
1018         if (!e)
1019                 return 0;
1020
1021         data = elevator_init_queue(q, e);
1022         if (!data) {
1023                 kobject_put(&e->kobj);
1024                 return 0;
1025         }
1026
1027         /*
1028          * Turn on BYPASS and drain all requests w/ elevator private data
1029          */
1030         spin_lock_irq(q->queue_lock);
1031
1032         set_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1033
1034         elv_drain_elevator(q);
1035
1036         while (q->rq.elvpriv) {
1037                 blk_remove_plug(q);
1038                 q->request_fn(q);
1039                 spin_unlock_irq(q->queue_lock);
1040                 msleep(10);
1041                 spin_lock_irq(q->queue_lock);
1042                 elv_drain_elevator(q);
1043         }
1044
1045         /*
1046          * Remember old elevator.
1047          */
1048         old_elevator = q->elevator;
1049
1050         /*
1051          * attach and start new elevator
1052          */
1053         elevator_attach(q, e, data);
1054
1055         spin_unlock_irq(q->queue_lock);
1056
1057         __elv_unregister_queue(old_elevator);
1058
1059         if (elv_register_queue(q))
1060                 goto fail_register;
1061
1062         /*
1063          * finally exit old elevator and turn off BYPASS.
1064          */
1065         elevator_exit(old_elevator);
1066         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1067         return 1;
1068
1069 fail_register:
1070         /*
1071          * switch failed, exit the new io scheduler and reattach the old
1072          * one again (along with re-adding the sysfs dir)
1073          */
1074         elevator_exit(e);
1075         q->elevator = old_elevator;
1076         elv_register_queue(q);
1077         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1078         return 0;
1079 }
1080
1081 ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count)
1082 {
1083         char elevator_name[ELV_NAME_MAX];
1084         size_t len;
1085         struct elevator_type *e;
1086
1087         elevator_name[sizeof(elevator_name) - 1] = '\0';
1088         strncpy(elevator_name, name, sizeof(elevator_name) - 1);
1089         len = strlen(elevator_name);
1090
1091         if (len && elevator_name[len - 1] == '\n')
1092                 elevator_name[len - 1] = '\0';
1093
1094         e = elevator_get(elevator_name);
1095         if (!e) {
1096                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1097                 return -EINVAL;
1098         }
1099
1100         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1101                 elevator_put(e);
1102                 return count;
1103         }
1104
1105         if (!elevator_switch(q, e))
1106                 printk(KERN_ERR "elevator: switch to %s failed\n",elevator_name);
1107         return count;
1108 }
1109
1110 ssize_t elv_iosched_show(request_queue_t *q, char *name)
1111 {
1112         elevator_t *e = q->elevator;
1113         struct elevator_type *elv = e->elevator_type;
1114         struct elevator_type *__e;
1115         int len = 0;
1116
1117         spin_lock(&elv_list_lock);
1118         list_for_each_entry(__e, &elv_list, list) {
1119                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1120                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1121                 else
1122                         len += sprintf(name+len, "%s ", __e->elevator_name);
1123         }
1124         spin_unlock(&elv_list_lock);
1125
1126         len += sprintf(len+name, "\n");
1127         return len;
1128 }
1129
1130 struct request *elv_rb_former_request(request_queue_t *q, struct request *rq)
1131 {
1132         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1133
1134         if (rbprev)
1135                 return rb_entry_rq(rbprev);
1136
1137         return NULL;
1138 }
1139
1140 EXPORT_SYMBOL(elv_rb_former_request);
1141
1142 struct request *elv_rb_latter_request(request_queue_t *q, struct request *rq)
1143 {
1144         struct rb_node *rbnext = rb_next(&rq->rb_node);
1145
1146         if (rbnext)
1147                 return rb_entry_rq(rbnext);
1148
1149         return NULL;
1150 }
1151
1152 EXPORT_SYMBOL(elv_rb_latter_request);