git.oblomov.eu Git - linux-2.6/blob - 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  * can we safely merge with this request?
  55  */
  56 inline int elv_rq_merge_ok(struct request *rq, struct bio *bio)
  57 {
  58         if (!rq_mergeable(rq))
  59                 return 0;
  60
  61         /*
  62          * different data direction or already started, don't merge
  63          */
  64         if (bio_data_dir(bio) != rq_data_dir(rq))
  65                 return 0;
  66
  67         /*
  68          * same device and no special stuff set, merge is ok
  69          */
  70         if (rq->rq_disk == bio->bi_bdev->bd_disk && !rq->special)
  71                 return 1;
  72
  73         return 0;
  74 }
  75 EXPORT_SYMBOL(elv_rq_merge_ok);
  76
  77 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
  78 {
  79         int ret = ELEVATOR_NO_MERGE;
  80
  81         /*
  82          * we can merge and sequence is ok, check if it's possible
  83          */
  84         if (elv_rq_merge_ok(__rq, bio)) {
  85                 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
  86                         ret = ELEVATOR_BACK_MERGE;
  87                 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
  88                         ret = ELEVATOR_FRONT_MERGE;
  89         }
  90
  91         return ret;
  92 }
  93
  94 static struct elevator_type *elevator_find(const char *name)
  95 {
  96         struct elevator_type *e;
  97         struct list_head *entry;
  98
  99         list_for_each(entry, &elv_list) {
 100
 101                 e = list_entry(entry, struct elevator_type, list);
 102
 103                 if (!strcmp(e->elevator_name, name))
 104                         return e;
 105         }
 106
 107         return NULL;
 108 }
 109
 110 static void elevator_put(struct elevator_type *e)
 111 {
 112         module_put(e->elevator_owner);
 113 }
 114
 115 static struct elevator_type *elevator_get(const char *name)
 116 {
 117         struct elevator_type *e;
 118
 119         spin_lock_irq(&elv_list_lock);
 120
 121         e = elevator_find(name);
 122         if (e && !try_module_get(e->elevator_owner))
 123                 e = NULL;
 124
 125         spin_unlock_irq(&elv_list_lock);
 126
 127         return e;
 128 }
 129
 130 static void *elevator_init_queue(request_queue_t *q, struct elevator_queue *eq)
 131 {
 132         return eq->ops->elevator_init_fn(q);
 133 }
 134
 135 static void elevator_attach(request_queue_t *q, struct elevator_queue *eq,
 136                            void *data)
 137 {
 138         q->elevator = eq;
 139         eq->elevator_data = data;
 140 }
 141
 142 static char chosen_elevator[16];
 143
 144 static int __init elevator_setup(char *str)
 145 {
 146         /*
 147          * Be backwards-compatible with previous kernels, so users
 148          * won't get the wrong elevator.
 149          */
 150         if (!strcmp(str, "as"))
 151                 strcpy(chosen_elevator, "anticipatory");
 152         else
 153                 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 154         return 1;
 155 }
 156
 157 __setup("elevator=", elevator_setup);
 158
 159 static struct kobj_type elv_ktype;
 160
 161 static elevator_t *elevator_alloc(request_queue_t *q, struct elevator_type *e)
 162 {
 163         elevator_t *eq;
 164         int i;
 165
 166         eq = kmalloc_node(sizeof(elevator_t), GFP_KERNEL, q->node);
 167         if (unlikely(!eq))
 168                 goto err;
 169
 170         memset(eq, 0, sizeof(*eq));
 171         eq->ops = &e->ops;
 172         eq->elevator_type = e;
 173         kobject_init(&eq->kobj);
 174         snprintf(eq->kobj.name, KOBJ_NAME_LEN, "%s", "iosched");
 175         eq->kobj.ktype = &elv_ktype;
 176         mutex_init(&eq->sysfs_lock);
 177
 178         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 179                                         GFP_KERNEL, q->node);
 180         if (!eq->hash)
 181                 goto err;
 182
 183         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 184                 INIT_HLIST_HEAD(&eq->hash[i]);
 185
 186         return eq;
 187 err:
 188         kfree(eq);
 189         elevator_put(e);
 190         return NULL;
 191 }
 192
 193 static void elevator_release(struct kobject *kobj)
 194 {
 195         elevator_t *e = container_of(kobj, elevator_t, kobj);
 196
 197         elevator_put(e->elevator_type);
 198         kfree(e->hash);
 199         kfree(e);
 200 }
 201
 202 int elevator_init(request_queue_t *q, char *name)
 203 {
 204         struct elevator_type *e = NULL;
 205         struct elevator_queue *eq;
 206         int ret = 0;
 207         void *data;
 208
 209         INIT_LIST_HEAD(&q->queue_head);
 210         q->last_merge = NULL;
 211         q->end_sector = 0;
 212         q->boundary_rq = NULL;
 213
 214         if (name && !(e = elevator_get(name)))
 215                 return -EINVAL;
 216
 217         if (!e && *chosen_elevator && !(e = elevator_get(chosen_elevator)))
 218                 printk("I/O scheduler %s not found\n", chosen_elevator);
 219
 220         if (!e && !(e = elevator_get(CONFIG_DEFAULT_IOSCHED))) {
 221                 printk("Default I/O scheduler not found, using no-op\n");
 222                 e = elevator_get("noop");
 223         }
 224
 225         eq = elevator_alloc(q, e);
 226         if (!eq)
 227                 return -ENOMEM;
 228
 229         data = elevator_init_queue(q, eq);
 230         if (!data) {
 231                 kobject_put(&eq->kobj);
 232                 return -ENOMEM;
 233         }
 234
 235         elevator_attach(q, eq, data);
 236         return ret;
 237 }
 238
 239 EXPORT_SYMBOL(elevator_init);
 240
 241 void elevator_exit(elevator_t *e)
 242 {
 243         mutex_lock(&e->sysfs_lock);
 244         if (e->ops->elevator_exit_fn)
 245                 e->ops->elevator_exit_fn(e);
 246         e->ops = NULL;
 247         mutex_unlock(&e->sysfs_lock);
 248
 249         kobject_put(&e->kobj);
 250 }
 251
 252 EXPORT_SYMBOL(elevator_exit);
 253
 254 static inline void __elv_rqhash_del(struct request *rq)
 255 {
 256         hlist_del_init(&rq->hash);
 257 }
 258
 259 static void elv_rqhash_del(request_queue_t *q, struct request *rq)
 260 {
 261         if (ELV_ON_HASH(rq))
 262                 __elv_rqhash_del(rq);
 263 }
 264
 265 static void elv_rqhash_add(request_queue_t *q, struct request *rq)
 266 {
 267         elevator_t *e = q->elevator;
 268
 269         BUG_ON(ELV_ON_HASH(rq));
 270         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 271 }
 272
 273 static void elv_rqhash_reposition(request_queue_t *q, struct request *rq)
 274 {
 275         __elv_rqhash_del(rq);
 276         elv_rqhash_add(q, rq);
 277 }
 278
 279 static struct request *elv_rqhash_find(request_queue_t *q, sector_t offset)
 280 {
 281         elevator_t *e = q->elevator;
 282         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 283         struct hlist_node *entry, *next;
 284         struct request *rq;
 285
 286         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 287                 BUG_ON(!ELV_ON_HASH(rq));
 288
 289                 if (unlikely(!rq_mergeable(rq))) {
 290                         __elv_rqhash_del(rq);
 291                         continue;
 292                 }
 293
 294                 if (rq_hash_key(rq) == offset)
 295                         return rq;
 296         }
 297
 298         return NULL;
 299 }
 300
 301 /*
 302  * RB-tree support functions for inserting/lookup/removal of requests
 303  * in a sorted RB tree.
 304  */
 305 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 306 {
 307         struct rb_node **p = &root->rb_node;
 308         struct rb_node *parent = NULL;
 309         struct request *__rq;
 310
 311         while (*p) {
 312                 parent = *p;
 313                 __rq = rb_entry(parent, struct request, rb_node);
 314
 315                 if (rq->sector < __rq->sector)
 316                         p = &(*p)->rb_left;
 317                 else if (rq->sector > __rq->sector)
 318                         p = &(*p)->rb_right;
 319                 else
 320                         return __rq;
 321         }
 322
 323         rb_link_node(&rq->rb_node, parent, p);
 324         rb_insert_color(&rq->rb_node, root);
 325         return NULL;
 326 }
 327
 328 EXPORT_SYMBOL(elv_rb_add);
 329
 330 void elv_rb_del(struct rb_root *root, struct request *rq)
 331 {
 332         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 333         rb_erase(&rq->rb_node, root);
 334         RB_CLEAR_NODE(&rq->rb_node);
 335 }
 336
 337 EXPORT_SYMBOL(elv_rb_del);
 338
 339 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 340 {
 341         struct rb_node *n = root->rb_node;
 342         struct request *rq;
 343
 344         while (n) {
 345                 rq = rb_entry(n, struct request, rb_node);
 346
 347                 if (sector < rq->sector)
 348                         n = n->rb_left;
 349                 else if (sector > rq->sector)
 350                         n = n->rb_right;
 351                 else
 352                         return rq;
 353         }
 354
 355         return NULL;
 356 }
 357
 358 EXPORT_SYMBOL(elv_rb_find);
 359
 360 /*
 361  * Insert rq into dispatch queue of q.  Queue lock must be held on
 362  * entry.  rq is sort insted into the dispatch queue. To be used by
 363  * specific elevators.
 364  */
 365 void elv_dispatch_sort(request_queue_t *q, struct request *rq)
 366 {
 367         sector_t boundary;
 368         struct list_head *entry;
 369
 370         if (q->last_merge == rq)
 371                 q->last_merge = NULL;
 372
 373         elv_rqhash_del(q, rq);
 374
 375         q->nr_sorted--;
 376
 377         boundary = q->end_sector;
 378
 379         list_for_each_prev(entry, &q->queue_head) {
 380                 struct request *pos = list_entry_rq(entry);
 381
 382                 if (pos->cmd_flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
 383                         break;
 384                 if (rq->sector >= boundary) {
 385                         if (pos->sector < boundary)
 386                                 continue;
 387                 } else {
 388                         if (pos->sector >= boundary)
 389                                 break;
 390                 }
 391                 if (rq->sector >= pos->sector)
 392                         break;
 393         }
 394
 395         list_add(&rq->queuelist, entry);
 396 }
 397
 398 EXPORT_SYMBOL(elv_dispatch_sort);
 399
 400 /*
 401  * Insert rq into dispatch queue of q.  Queue lock must be held on
 402  * entry.  rq is added to the back of the dispatch queue. To be used by
 403  * specific elevators.
 404  */
 405 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 406 {
 407         if (q->last_merge == rq)
 408                 q->last_merge = NULL;
 409
 410         elv_rqhash_del(q, rq);
 411
 412         q->nr_sorted--;
 413
 414         q->end_sector = rq_end_sector(rq);
 415         q->boundary_rq = rq;
 416         list_add_tail(&rq->queuelist, &q->queue_head);
 417 }
 418
 419 EXPORT_SYMBOL(elv_dispatch_add_tail);
 420
 421 int elv_merge(request_queue_t *q, struct request **req, struct bio *bio)
 422 {
 423         elevator_t *e = q->elevator;
 424         struct request *__rq;
 425         int ret;
 426
 427         /*
 428          * First try one-hit cache.
 429          */
 430         if (q->last_merge) {
 431                 ret = elv_try_merge(q->last_merge, bio);
 432                 if (ret != ELEVATOR_NO_MERGE) {
 433                         *req = q->last_merge;
 434                         return ret;
 435                 }
 436         }
 437
 438         /*
 439          * See if our hash lookup can find a potential backmerge.
 440          */
 441         __rq = elv_rqhash_find(q, bio->bi_sector);
 442         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 443                 *req = __rq;
 444                 return ELEVATOR_BACK_MERGE;
 445         }
 446
 447         if (e->ops->elevator_merge_fn)
 448                 return e->ops->elevator_merge_fn(q, req, bio);
 449
 450         return ELEVATOR_NO_MERGE;
 451 }
 452
 453 void elv_merged_request(request_queue_t *q, struct request *rq, int type)
 454 {
 455         elevator_t *e = q->elevator;
 456
 457         if (e->ops->elevator_merged_fn)
 458                 e->ops->elevator_merged_fn(q, rq, type);
 459
 460         if (type == ELEVATOR_BACK_MERGE)
 461                 elv_rqhash_reposition(q, rq);
 462
 463         q->last_merge = rq;
 464 }
 465
 466 void elv_merge_requests(request_queue_t *q, struct request *rq,
 467                              struct request *next)
 468 {
 469         elevator_t *e = q->elevator;
 470
 471         if (e->ops->elevator_merge_req_fn)
 472                 e->ops->elevator_merge_req_fn(q, rq, next);
 473
 474         elv_rqhash_reposition(q, rq);
 475         elv_rqhash_del(q, next);
 476
 477         q->nr_sorted--;
 478         q->last_merge = rq;
 479 }
 480
 481 void elv_requeue_request(request_queue_t *q, struct request *rq)
 482 {
 483         elevator_t *e = q->elevator;
 484
 485         /*
 486          * it already went through dequeue, we need to decrement the
 487          * in_flight count again
 488          */
 489         if (blk_account_rq(rq)) {
 490                 q->in_flight--;
 491                 if (blk_sorted_rq(rq) && e->ops->elevator_deactivate_req_fn)
 492                         e->ops->elevator_deactivate_req_fn(q, rq);
 493         }
 494
 495         rq->cmd_flags &= ~REQ_STARTED;
 496
 497         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 498 }
 499
 500 static void elv_drain_elevator(request_queue_t *q)
 501 {
 502         static int printed;
 503         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 504                 ;
 505         if (q->nr_sorted == 0)
 506                 return;
 507         if (printed++ < 10) {
 508                 printk(KERN_ERR "%s: forced dispatching is broken "
 509                        "(nr_sorted=%u), please report this\n",
 510                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 511         }
 512 }
 513
 514 void elv_insert(request_queue_t *q, struct request *rq, int where)
 515 {
 516         struct list_head *pos;
 517         unsigned ordseq;
 518         int unplug_it = 1;
 519
 520         blk_add_trace_rq(q, rq, BLK_TA_INSERT);
 521
 522         rq->q = q;
 523
 524         switch (where) {
 525         case ELEVATOR_INSERT_FRONT:
 526                 rq->cmd_flags |= REQ_SOFTBARRIER;
 527
 528                 list_add(&rq->queuelist, &q->queue_head);
 529                 break;
 530
 531         case ELEVATOR_INSERT_BACK:
 532                 rq->cmd_flags |= REQ_SOFTBARRIER;
 533                 elv_drain_elevator(q);
 534                 list_add_tail(&rq->queuelist, &q->queue_head);
 535                 /*
 536                  * We kick the queue here for the following reasons.
 537                  * - The elevator might have returned NULL previously
 538                  *   to delay requests and returned them now.  As the
 539                  *   queue wasn't empty before this request, ll_rw_blk
 540                  *   won't run the queue on return, resulting in hang.
 541                  * - Usually, back inserted requests won't be merged
 542                  *   with anything.  There's no point in delaying queue
 543                  *   processing.
 544                  */
 545                 blk_remove_plug(q);
 546                 q->request_fn(q);
 547                 break;
 548
 549         case ELEVATOR_INSERT_SORT:
 550                 BUG_ON(!blk_fs_request(rq));
 551                 rq->cmd_flags |= REQ_SORTED;
 552                 q->nr_sorted++;
 553                 if (rq_mergeable(rq)) {
 554                         elv_rqhash_add(q, rq);
 555                         if (!q->last_merge)
 556                                 q->last_merge = rq;
 557                 }
 558
 559                 /*
 560                  * Some ioscheds (cfq) run q->request_fn directly, so
 561                  * rq cannot be accessed after calling
 562                  * elevator_add_req_fn.
 563                  */
 564                 q->elevator->ops->elevator_add_req_fn(q, rq);
 565                 break;
 566
 567         case ELEVATOR_INSERT_REQUEUE:
 568                 /*
 569                  * If ordered flush isn't in progress, we do front
 570                  * insertion; otherwise, requests should be requeued
 571                  * in ordseq order.
 572                  */
 573                 rq->cmd_flags |= REQ_SOFTBARRIER;
 574
 575                 if (q->ordseq == 0) {
 576                         list_add(&rq->queuelist, &q->queue_head);
 577                         break;
 578                 }
 579
 580                 ordseq = blk_ordered_req_seq(rq);
 581
 582                 list_for_each(pos, &q->queue_head) {
 583                         struct request *pos_rq = list_entry_rq(pos);
 584                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 585                                 break;
 586                 }
 587
 588                 list_add_tail(&rq->queuelist, pos);
 589                 /*
 590                  * most requeues happen because of a busy condition, don't
 591                  * force unplug of the queue for that case.
 592                  */
 593                 unplug_it = 0;
 594                 break;
 595
 596         default:
 597                 printk(KERN_ERR "%s: bad insertion point %d\n",
 598                        __FUNCTION__, where);
 599                 BUG();
 600         }
 601
 602         if (unplug_it && blk_queue_plugged(q)) {
 603                 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
 604                         - q->in_flight;
 605
 606                 if (nrq >= q->unplug_thresh)
 607                         __generic_unplug_device(q);
 608         }
 609 }
 610
 611 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
 612                        int plug)
 613 {
 614         if (q->ordcolor)
 615                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 616
 617         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 618                 /*
 619                  * toggle ordered color
 620                  */
 621                 if (blk_barrier_rq(rq))
 622                         q->ordcolor ^= 1;
 623
 624                 /*
 625                  * barriers implicitly indicate back insertion
 626                  */
 627                 if (where == ELEVATOR_INSERT_SORT)
 628                         where = ELEVATOR_INSERT_BACK;
 629
 630                 /*
 631                  * this request is scheduling boundary, update
 632                  * end_sector
 633                  */
 634                 if (blk_fs_request(rq)) {
 635                         q->end_sector = rq_end_sector(rq);
 636                         q->boundary_rq = rq;
 637                 }
 638         } else if (!(rq->cmd_flags & REQ_ELVPRIV) && where == ELEVATOR_INSERT_SORT)
 639                 where = ELEVATOR_INSERT_BACK;
 640
 641         if (plug)
 642                 blk_plug_device(q);
 643
 644         elv_insert(q, rq, where);
 645 }
 646
 647 EXPORT_SYMBOL(__elv_add_request);
 648
 649 void elv_add_request(request_queue_t *q, struct request *rq, int where,
 650                      int plug)
 651 {
 652         unsigned long flags;
 653
 654         spin_lock_irqsave(q->queue_lock, flags);
 655         __elv_add_request(q, rq, where, plug);
 656         spin_unlock_irqrestore(q->queue_lock, flags);
 657 }
 658
 659 EXPORT_SYMBOL(elv_add_request);
 660
 661 static inline struct request *__elv_next_request(request_queue_t *q)
 662 {
 663         struct request *rq;
 664
 665         while (1) {
 666                 while (!list_empty(&q->queue_head)) {
 667                         rq = list_entry_rq(q->queue_head.next);
 668                         if (blk_do_ordered(q, &rq))
 669                                 return rq;
 670                 }
 671
 672                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
 673                         return NULL;
 674         }
 675 }
 676
 677 struct request *elv_next_request(request_queue_t *q)
 678 {
 679         struct request *rq;
 680         int ret;
 681
 682         while ((rq = __elv_next_request(q)) != NULL) {
 683                 if (!(rq->cmd_flags & REQ_STARTED)) {
 684                         elevator_t *e = q->elevator;
 685
 686                         /*
 687                          * This is the first time the device driver
 688                          * sees this request (possibly after
 689                          * requeueing).  Notify IO scheduler.
 690                          */
 691                         if (blk_sorted_rq(rq) &&
 692                             e->ops->elevator_activate_req_fn)
 693                                 e->ops->elevator_activate_req_fn(q, rq);
 694
 695                         /*
 696                          * just mark as started even if we don't start
 697                          * it, a request that has been delayed should
 698                          * not be passed by new incoming requests
 699                          */
 700                         rq->cmd_flags |= REQ_STARTED;
 701                         blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
 702                 }
 703
 704                 if (!q->boundary_rq || q->boundary_rq == rq) {
 705                         q->end_sector = rq_end_sector(rq);
 706                         q->boundary_rq = NULL;
 707                 }
 708
 709                 if ((rq->cmd_flags & REQ_DONTPREP) || !q->prep_rq_fn)
 710                         break;
 711
 712                 ret = q->prep_rq_fn(q, rq);
 713                 if (ret == BLKPREP_OK) {
 714                         break;
 715                 } else if (ret == BLKPREP_DEFER) {
 716                         /*
 717                          * the request may have been (partially) prepped.
 718                          * we need to keep this request in the front to
 719                          * avoid resource deadlock.  REQ_STARTED will
 720                          * prevent other fs requests from passing this one.
 721                          */
 722                         rq = NULL;
 723                         break;
 724                 } else if (ret == BLKPREP_KILL) {
 725                         int nr_bytes = rq->hard_nr_sectors << 9;
 726
 727                         if (!nr_bytes)
 728                                 nr_bytes = rq->data_len;
 729
 730                         blkdev_dequeue_request(rq);
 731                         rq->cmd_flags |= REQ_QUIET;
 732                         end_that_request_chunk(rq, 0, nr_bytes);
 733                         end_that_request_last(rq, 0);
 734                 } else {
 735                         printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
 736                                                                 ret);
 737                         break;
 738                 }
 739         }
 740
 741         return rq;
 742 }
 743
 744 EXPORT_SYMBOL(elv_next_request);
 745
 746 void elv_dequeue_request(request_queue_t *q, struct request *rq)
 747 {
 748         BUG_ON(list_empty(&rq->queuelist));
 749         BUG_ON(ELV_ON_HASH(rq));
 750
 751         list_del_init(&rq->queuelist);
 752
 753         /*
 754          * the time frame between a request being removed from the lists
 755          * and to it is freed is accounted as io that is in progress at
 756          * the driver side.
 757          */
 758         if (blk_account_rq(rq))
 759                 q->in_flight++;
 760 }
 761
 762 EXPORT_SYMBOL(elv_dequeue_request);
 763
 764 int elv_queue_empty(request_queue_t *q)
 765 {
 766         elevator_t *e = q->elevator;
 767
 768         if (!list_empty(&q->queue_head))
 769                 return 0;
 770
 771         if (e->ops->elevator_queue_empty_fn)
 772                 return e->ops->elevator_queue_empty_fn(q);
 773
 774         return 1;
 775 }
 776
 777 EXPORT_SYMBOL(elv_queue_empty);
 778
 779 struct request *elv_latter_request(request_queue_t *q, struct request *rq)
 780 {
 781         elevator_t *e = q->elevator;
 782
 783         if (e->ops->elevator_latter_req_fn)
 784                 return e->ops->elevator_latter_req_fn(q, rq);
 785         return NULL;
 786 }
 787
 788 struct request *elv_former_request(request_queue_t *q, struct request *rq)
 789 {
 790         elevator_t *e = q->elevator;
 791
 792         if (e->ops->elevator_former_req_fn)
 793                 return e->ops->elevator_former_req_fn(q, rq);
 794         return NULL;
 795 }
 796
 797 int elv_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
 798 {
 799         elevator_t *e = q->elevator;
 800
 801         if (e->ops->elevator_set_req_fn)
 802                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 803
 804         rq->elevator_private = NULL;
 805         return 0;
 806 }
 807
 808 void elv_put_request(request_queue_t *q, struct request *rq)
 809 {
 810         elevator_t *e = q->elevator;
 811
 812         if (e->ops->elevator_put_req_fn)
 813                 e->ops->elevator_put_req_fn(rq);
 814 }
 815
 816 int elv_may_queue(request_queue_t *q, int rw)
 817 {
 818         elevator_t *e = q->elevator;
 819
 820         if (e->ops->elevator_may_queue_fn)
 821                 return e->ops->elevator_may_queue_fn(q, rw);
 822
 823         return ELV_MQUEUE_MAY;
 824 }
 825
 826 void elv_completed_request(request_queue_t *q, struct request *rq)
 827 {
 828         elevator_t *e = q->elevator;
 829
 830         /*
 831          * request is released from the driver, io must be done
 832          */
 833         if (blk_account_rq(rq)) {
 834                 q->in_flight--;
 835                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 836                         e->ops->elevator_completed_req_fn(q, rq);
 837         }
 838
 839         /*
 840          * Check if the queue is waiting for fs requests to be
 841          * drained for flush sequence.
 842          */
 843         if (unlikely(q->ordseq)) {
 844                 struct request *first_rq = list_entry_rq(q->queue_head.next);
 845                 if (q->in_flight == 0 &&
 846                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 847                     blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
 848                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 849                         q->request_fn(q);
 850                 }
 851         }
 852 }
 853
 854 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 855
 856 static ssize_t
 857 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 858 {
 859         elevator_t *e = container_of(kobj, elevator_t, kobj);
 860         struct elv_fs_entry *entry = to_elv(attr);
 861         ssize_t error;
 862
 863         if (!entry->show)
 864                 return -EIO;
 865
 866         mutex_lock(&e->sysfs_lock);
 867         error = e->ops ? entry->show(e, page) : -ENOENT;
 868         mutex_unlock(&e->sysfs_lock);
 869         return error;
 870 }
 871
 872 static ssize_t
 873 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 874                const char *page, size_t length)
 875 {
 876         elevator_t *e = container_of(kobj, elevator_t, kobj);
 877         struct elv_fs_entry *entry = to_elv(attr);
 878         ssize_t error;
 879
 880         if (!entry->store)
 881                 return -EIO;
 882
 883         mutex_lock(&e->sysfs_lock);
 884         error = e->ops ? entry->store(e, page, length) : -ENOENT;
 885         mutex_unlock(&e->sysfs_lock);
 886         return error;
 887 }
 888
 889 static struct sysfs_ops elv_sysfs_ops = {
 890         .show   = elv_attr_show,
 891         .store  = elv_attr_store,
 892 };
 893
 894 static struct kobj_type elv_ktype = {
 895         .sysfs_ops      = &elv_sysfs_ops,
 896         .release        = elevator_release,
 897 };
 898
 899 int elv_register_queue(struct request_queue *q)
 900 {
 901         elevator_t *e = q->elevator;
 902         int error;
 903
 904         e->kobj.parent = &q->kobj;
 905
 906         error = kobject_add(&e->kobj);
 907         if (!error) {
 908                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
 909                 if (attr) {
 910                         while (attr->attr.name) {
 911                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 912                                         break;
 913                                 attr++;
 914                         }
 915                 }
 916                 kobject_uevent(&e->kobj, KOBJ_ADD);
 917         }
 918         return error;
 919 }
 920
 921 static void __elv_unregister_queue(elevator_t *e)
 922 {
 923         kobject_uevent(&e->kobj, KOBJ_REMOVE);
 924         kobject_del(&e->kobj);
 925 }
 926
 927 void elv_unregister_queue(struct request_queue *q)
 928 {
 929         if (q)
 930                 __elv_unregister_queue(q->elevator);
 931 }
 932
 933 int elv_register(struct elevator_type *e)
 934 {
 935         spin_lock_irq(&elv_list_lock);
 936         BUG_ON(elevator_find(e->elevator_name));
 937         list_add_tail(&e->list, &elv_list);
 938         spin_unlock_irq(&elv_list_lock);
 939
 940         printk(KERN_INFO "io scheduler %s registered", e->elevator_name);
 941         if (!strcmp(e->elevator_name, chosen_elevator) ||
 942                         (!*chosen_elevator &&
 943                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 944                                 printk(" (default)");
 945         printk("\n");
 946         return 0;
 947 }
 948 EXPORT_SYMBOL_GPL(elv_register);
 949
 950 void elv_unregister(struct elevator_type *e)
 951 {
 952         struct task_struct *g, *p;
 953
 954         /*
 955          * Iterate every thread in the process to remove the io contexts.
 956          */
 957         if (e->ops.trim) {
 958                 read_lock(&tasklist_lock);
 959                 do_each_thread(g, p) {
 960                         task_lock(p);
 961                         if (p->io_context)
 962                                 e->ops.trim(p->io_context);
 963                         task_unlock(p);
 964                 } while_each_thread(g, p);
 965                 read_unlock(&tasklist_lock);
 966         }
 967
 968         spin_lock_irq(&elv_list_lock);
 969         list_del_init(&e->list);
 970         spin_unlock_irq(&elv_list_lock);
 971 }
 972 EXPORT_SYMBOL_GPL(elv_unregister);
 973
 974 /*
 975  * switch to new_e io scheduler. be careful not to introduce deadlocks -
 976  * we don't free the old io scheduler, before we have allocated what we
 977  * need for the new one. this way we have a chance of going back to the old
 978  * one, if the new one fails init for some reason.
 979  */
 980 static int elevator_switch(request_queue_t *q, struct elevator_type *new_e)
 981 {
 982         elevator_t *old_elevator, *e;
 983         void *data;
 984
 985         /*
 986          * Allocate new elevator
 987          */
 988         e = elevator_alloc(q, new_e);
 989         if (!e)
 990                 return 0;
 991
 992         data = elevator_init_queue(q, e);
 993         if (!data) {
 994                 kobject_put(&e->kobj);
 995                 return 0;
 996         }
 997
 998         /*
 999          * Turn on BYPASS and drain all requests w/ elevator private data
1000          */
1001         spin_lock_irq(q->queue_lock);
1002
1003         set_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1004
1005         elv_drain_elevator(q);
1006
1007         while (q->rq.elvpriv) {
1008                 blk_remove_plug(q);
1009                 q->request_fn(q);
1010                 spin_unlock_irq(q->queue_lock);
1011                 msleep(10);
1012                 spin_lock_irq(q->queue_lock);
1013                 elv_drain_elevator(q);
1014         }
1015
1016         /*
1017          * Remember old elevator.
1018          */
1019         old_elevator = q->elevator;
1020
1021         /*
1022          * attach and start new elevator
1023          */
1024         elevator_attach(q, e, data);
1025
1026         spin_unlock_irq(q->queue_lock);
1027
1028         __elv_unregister_queue(old_elevator);
1029
1030         if (elv_register_queue(q))
1031                 goto fail_register;
1032
1033         /*
1034          * finally exit old elevator and turn off BYPASS.
1035          */
1036         elevator_exit(old_elevator);
1037         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1038         return 1;
1039
1040 fail_register:
1041         /*
1042          * switch failed, exit the new io scheduler and reattach the old
1043          * one again (along with re-adding the sysfs dir)
1044          */
1045         elevator_exit(e);
1046         q->elevator = old_elevator;
1047         elv_register_queue(q);
1048         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1049         return 0;
1050 }
1051
1052 ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count)
1053 {
1054         char elevator_name[ELV_NAME_MAX];
1055         size_t len;
1056         struct elevator_type *e;
1057
1058         elevator_name[sizeof(elevator_name) - 1] = '\0';
1059         strncpy(elevator_name, name, sizeof(elevator_name) - 1);
1060         len = strlen(elevator_name);
1061
1062         if (len && elevator_name[len - 1] == '\n')
1063                 elevator_name[len - 1] = '\0';
1064
1065         e = elevator_get(elevator_name);
1066         if (!e) {
1067                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1068                 return -EINVAL;
1069         }
1070
1071         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1072                 elevator_put(e);
1073                 return count;
1074         }
1075
1076         if (!elevator_switch(q, e))
1077                 printk(KERN_ERR "elevator: switch to %s failed\n",elevator_name);
1078         return count;
1079 }
1080
1081 ssize_t elv_iosched_show(request_queue_t *q, char *name)
1082 {
1083         elevator_t *e = q->elevator;
1084         struct elevator_type *elv = e->elevator_type;
1085         struct list_head *entry;
1086         int len = 0;
1087
1088         spin_lock_irq(&elv_list_lock);
1089         list_for_each(entry, &elv_list) {
1090                 struct elevator_type *__e;
1091
1092                 __e = list_entry(entry, struct elevator_type, list);
1093                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1094                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1095                 else
1096                         len += sprintf(name+len, "%s ", __e->elevator_name);
1097         }
1098         spin_unlock_irq(&elv_list_lock);
1099
1100         len += sprintf(len+name, "\n");
1101         return len;
1102 }
1103
1104 struct request *elv_rb_former_request(request_queue_t *q, struct request *rq)
1105 {
1106         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1107
1108         if (rbprev)
1109                 return rb_entry_rq(rbprev);
1110
1111         return NULL;
1112 }
1113
1114 EXPORT_SYMBOL(elv_rb_former_request);
1115
1116 struct request *elv_rb_latter_request(request_queue_t *q, struct request *rq)
1117 {
1118         struct rb_node *rbnext = rb_next(&rq->rb_node);
1119
1120         if (rbnext)
1121                 return rb_entry_rq(rbnext);
1122
1123         return NULL;
1124 }
1125
1126 EXPORT_SYMBOL(elv_rb_latter_request);