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