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  * 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         struct request_queue *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(struct request_queue *q,
 145                                  struct elevator_queue *eq)
 146 {
 147         return eq->ops->elevator_init_fn(q);
 148 }
 149
 150 static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
 151                            void *data)
 152 {
 153         q->elevator = eq;
 154         eq->elevator_data = data;
 155 }
 156
 157 static char chosen_elevator[16];
 158
 159 static int __init elevator_setup(char *str)
 160 {
 161         /*
 162          * Be backwards-compatible with previous kernels, so users
 163          * won't get the wrong elevator.
 164          */
 165         if (!strcmp(str, "as"))
 166                 strcpy(chosen_elevator, "anticipatory");
 167         else
 168                 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 169         return 1;
 170 }
 171
 172 __setup("elevator=", elevator_setup);
 173
 174 static struct kobj_type elv_ktype;
 175
 176 static elevator_t *elevator_alloc(struct request_queue *q,
 177                                   struct elevator_type *e)
 178 {
 179         elevator_t *eq;
 180         int i;
 181
 182         eq = kmalloc_node(sizeof(elevator_t), GFP_KERNEL | __GFP_ZERO, q->node);
 183         if (unlikely(!eq))
 184                 goto err;
 185
 186         eq->ops = &e->ops;
 187         eq->elevator_type = e;
 188         kobject_init(&eq->kobj);
 189         kobject_set_name(&eq->kobj, "%s", "iosched");
 190         eq->kobj.ktype = &elv_ktype;
 191         mutex_init(&eq->sysfs_lock);
 192
 193         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 194                                         GFP_KERNEL, q->node);
 195         if (!eq->hash)
 196                 goto err;
 197
 198         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 199                 INIT_HLIST_HEAD(&eq->hash[i]);
 200
 201         return eq;
 202 err:
 203         kfree(eq);
 204         elevator_put(e);
 205         return NULL;
 206 }
 207
 208 static void elevator_release(struct kobject *kobj)
 209 {
 210         elevator_t *e = container_of(kobj, elevator_t, kobj);
 211
 212         elevator_put(e->elevator_type);
 213         kfree(e->hash);
 214         kfree(e);
 215 }
 216
 217 int elevator_init(struct request_queue *q, char *name)
 218 {
 219         struct elevator_type *e = NULL;
 220         struct elevator_queue *eq;
 221         int ret = 0;
 222         void *data;
 223
 224         INIT_LIST_HEAD(&q->queue_head);
 225         q->last_merge = NULL;
 226         q->end_sector = 0;
 227         q->boundary_rq = NULL;
 228
 229         if (name && !(e = elevator_get(name)))
 230                 return -EINVAL;
 231
 232         if (!e && *chosen_elevator && !(e = elevator_get(chosen_elevator)))
 233                 printk("I/O scheduler %s not found\n", chosen_elevator);
 234
 235         if (!e && !(e = elevator_get(CONFIG_DEFAULT_IOSCHED))) {
 236                 printk("Default I/O scheduler not found, using no-op\n");
 237                 e = elevator_get("noop");
 238         }
 239
 240         eq = elevator_alloc(q, e);
 241         if (!eq)
 242                 return -ENOMEM;
 243
 244         data = elevator_init_queue(q, eq);
 245         if (!data) {
 246                 kobject_put(&eq->kobj);
 247                 return -ENOMEM;
 248         }
 249
 250         elevator_attach(q, eq, data);
 251         return ret;
 252 }
 253
 254 EXPORT_SYMBOL(elevator_init);
 255
 256 void elevator_exit(elevator_t *e)
 257 {
 258         mutex_lock(&e->sysfs_lock);
 259         if (e->ops->elevator_exit_fn)
 260                 e->ops->elevator_exit_fn(e);
 261         e->ops = NULL;
 262         mutex_unlock(&e->sysfs_lock);
 263
 264         kobject_put(&e->kobj);
 265 }
 266
 267 EXPORT_SYMBOL(elevator_exit);
 268
 269 static void elv_activate_rq(struct request_queue *q, struct request *rq)
 270 {
 271         elevator_t *e = q->elevator;
 272
 273         if (e->ops->elevator_activate_req_fn)
 274                 e->ops->elevator_activate_req_fn(q, rq);
 275 }
 276
 277 static void elv_deactivate_rq(struct request_queue *q, struct request *rq)
 278 {
 279         elevator_t *e = q->elevator;
 280
 281         if (e->ops->elevator_deactivate_req_fn)
 282                 e->ops->elevator_deactivate_req_fn(q, rq);
 283 }
 284
 285 static inline void __elv_rqhash_del(struct request *rq)
 286 {
 287         hlist_del_init(&rq->hash);
 288 }
 289
 290 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 291 {
 292         if (ELV_ON_HASH(rq))
 293                 __elv_rqhash_del(rq);
 294 }
 295
 296 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 297 {
 298         elevator_t *e = q->elevator;
 299
 300         BUG_ON(ELV_ON_HASH(rq));
 301         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 302 }
 303
 304 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 305 {
 306         __elv_rqhash_del(rq);
 307         elv_rqhash_add(q, rq);
 308 }
 309
 310 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 311 {
 312         elevator_t *e = q->elevator;
 313         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 314         struct hlist_node *entry, *next;
 315         struct request *rq;
 316
 317         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 318                 BUG_ON(!ELV_ON_HASH(rq));
 319
 320                 if (unlikely(!rq_mergeable(rq))) {
 321                         __elv_rqhash_del(rq);
 322                         continue;
 323                 }
 324
 325                 if (rq_hash_key(rq) == offset)
 326                         return rq;
 327         }
 328
 329         return NULL;
 330 }
 331
 332 /*
 333  * RB-tree support functions for inserting/lookup/removal of requests
 334  * in a sorted RB tree.
 335  */
 336 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 337 {
 338         struct rb_node **p = &root->rb_node;
 339         struct rb_node *parent = NULL;
 340         struct request *__rq;
 341
 342         while (*p) {
 343                 parent = *p;
 344                 __rq = rb_entry(parent, struct request, rb_node);
 345
 346                 if (rq->sector < __rq->sector)
 347                         p = &(*p)->rb_left;
 348                 else if (rq->sector > __rq->sector)
 349                         p = &(*p)->rb_right;
 350                 else
 351                         return __rq;
 352         }
 353
 354         rb_link_node(&rq->rb_node, parent, p);
 355         rb_insert_color(&rq->rb_node, root);
 356         return NULL;
 357 }
 358
 359 EXPORT_SYMBOL(elv_rb_add);
 360
 361 void elv_rb_del(struct rb_root *root, struct request *rq)
 362 {
 363         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 364         rb_erase(&rq->rb_node, root);
 365         RB_CLEAR_NODE(&rq->rb_node);
 366 }
 367
 368 EXPORT_SYMBOL(elv_rb_del);
 369
 370 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 371 {
 372         struct rb_node *n = root->rb_node;
 373         struct request *rq;
 374
 375         while (n) {
 376                 rq = rb_entry(n, struct request, rb_node);
 377
 378                 if (sector < rq->sector)
 379                         n = n->rb_left;
 380                 else if (sector > rq->sector)
 381                         n = n->rb_right;
 382                 else
 383                         return rq;
 384         }
 385
 386         return NULL;
 387 }
 388
 389 EXPORT_SYMBOL(elv_rb_find);
 390
 391 /*
 392  * Insert rq into dispatch queue of q.  Queue lock must be held on
 393  * entry.  rq is sort instead into the dispatch queue. To be used by
 394  * specific elevators.
 395  */
 396 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 397 {
 398         sector_t boundary;
 399         struct list_head *entry;
 400
 401         if (q->last_merge == rq)
 402                 q->last_merge = NULL;
 403
 404         elv_rqhash_del(q, rq);
 405
 406         q->nr_sorted--;
 407
 408         boundary = q->end_sector;
 409
 410         list_for_each_prev(entry, &q->queue_head) {
 411                 struct request *pos = list_entry_rq(entry);
 412
 413                 if (rq_data_dir(rq) != rq_data_dir(pos))
 414                         break;
 415                 if (pos->cmd_flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
 416                         break;
 417                 if (rq->sector >= boundary) {
 418                         if (pos->sector < boundary)
 419                                 continue;
 420                 } else {
 421                         if (pos->sector >= boundary)
 422                                 break;
 423                 }
 424                 if (rq->sector >= pos->sector)
 425                         break;
 426         }
 427
 428         list_add(&rq->queuelist, entry);
 429 }
 430
 431 EXPORT_SYMBOL(elv_dispatch_sort);
 432
 433 /*
 434  * Insert rq into dispatch queue of q.  Queue lock must be held on
 435  * entry.  rq is added to the back of the dispatch queue. To be used by
 436  * specific elevators.
 437  */
 438 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 439 {
 440         if (q->last_merge == rq)
 441                 q->last_merge = NULL;
 442
 443         elv_rqhash_del(q, rq);
 444
 445         q->nr_sorted--;
 446
 447         q->end_sector = rq_end_sector(rq);
 448         q->boundary_rq = rq;
 449         list_add_tail(&rq->queuelist, &q->queue_head);
 450 }
 451
 452 EXPORT_SYMBOL(elv_dispatch_add_tail);
 453
 454 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 455 {
 456         elevator_t *e = q->elevator;
 457         struct request *__rq;
 458         int ret;
 459
 460         /*
 461          * First try one-hit cache.
 462          */
 463         if (q->last_merge) {
 464                 ret = elv_try_merge(q->last_merge, bio);
 465                 if (ret != ELEVATOR_NO_MERGE) {
 466                         *req = q->last_merge;
 467                         return ret;
 468                 }
 469         }
 470
 471         /*
 472          * See if our hash lookup can find a potential backmerge.
 473          */
 474         __rq = elv_rqhash_find(q, bio->bi_sector);
 475         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 476                 *req = __rq;
 477                 return ELEVATOR_BACK_MERGE;
 478         }
 479
 480         if (e->ops->elevator_merge_fn)
 481                 return e->ops->elevator_merge_fn(q, req, bio);
 482
 483         return ELEVATOR_NO_MERGE;
 484 }
 485
 486 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 487 {
 488         elevator_t *e = q->elevator;
 489
 490         if (e->ops->elevator_merged_fn)
 491                 e->ops->elevator_merged_fn(q, rq, type);
 492
 493         if (type == ELEVATOR_BACK_MERGE)
 494                 elv_rqhash_reposition(q, rq);
 495
 496         q->last_merge = rq;
 497 }
 498
 499 void elv_merge_requests(struct request_queue *q, struct request *rq,
 500                              struct request *next)
 501 {
 502         elevator_t *e = q->elevator;
 503
 504         if (e->ops->elevator_merge_req_fn)
 505                 e->ops->elevator_merge_req_fn(q, rq, next);
 506
 507         elv_rqhash_reposition(q, rq);
 508         elv_rqhash_del(q, next);
 509
 510         q->nr_sorted--;
 511         q->last_merge = rq;
 512 }
 513
 514 void elv_requeue_request(struct request_queue *q, struct request *rq)
 515 {
 516         /*
 517          * it already went through dequeue, we need to decrement the
 518          * in_flight count again
 519          */
 520         if (blk_account_rq(rq)) {
 521                 q->in_flight--;
 522                 if (blk_sorted_rq(rq))
 523                         elv_deactivate_rq(q, rq);
 524         }
 525
 526         rq->cmd_flags &= ~REQ_STARTED;
 527
 528         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 529 }
 530
 531 static void elv_drain_elevator(struct request_queue *q)
 532 {
 533         static int printed;
 534         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 535                 ;
 536         if (q->nr_sorted == 0)
 537                 return;
 538         if (printed++ < 10) {
 539                 printk(KERN_ERR "%s: forced dispatching is broken "
 540                        "(nr_sorted=%u), please report this\n",
 541                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 542         }
 543 }
 544
 545 void elv_insert(struct request_queue *q, struct request *rq, int where)
 546 {
 547         struct list_head *pos;
 548         unsigned ordseq;
 549         int unplug_it = 1;
 550
 551         blk_add_trace_rq(q, rq, BLK_TA_INSERT);
 552
 553         rq->q = q;
 554
 555         switch (where) {
 556         case ELEVATOR_INSERT_FRONT:
 557                 rq->cmd_flags |= REQ_SOFTBARRIER;
 558
 559                 list_add(&rq->queuelist, &q->queue_head);
 560                 break;
 561
 562         case ELEVATOR_INSERT_BACK:
 563                 rq->cmd_flags |= REQ_SOFTBARRIER;
 564                 elv_drain_elevator(q);
 565                 list_add_tail(&rq->queuelist, &q->queue_head);
 566                 /*
 567                  * We kick the queue here for the following reasons.
 568                  * - The elevator might have returned NULL previously
 569                  *   to delay requests and returned them now.  As the
 570                  *   queue wasn't empty before this request, ll_rw_blk
 571                  *   won't run the queue on return, resulting in hang.
 572                  * - Usually, back inserted requests won't be merged
 573                  *   with anything.  There's no point in delaying queue
 574                  *   processing.
 575                  */
 576                 blk_remove_plug(q);
 577                 q->request_fn(q);
 578                 break;
 579
 580         case ELEVATOR_INSERT_SORT:
 581                 BUG_ON(!blk_fs_request(rq));
 582                 rq->cmd_flags |= REQ_SORTED;
 583                 q->nr_sorted++;
 584                 if (rq_mergeable(rq)) {
 585                         elv_rqhash_add(q, rq);
 586                         if (!q->last_merge)
 587                                 q->last_merge = rq;
 588                 }
 589
 590                 /*
 591                  * Some ioscheds (cfq) run q->request_fn directly, so
 592                  * rq cannot be accessed after calling
 593                  * elevator_add_req_fn.
 594                  */
 595                 q->elevator->ops->elevator_add_req_fn(q, rq);
 596                 break;
 597
 598         case ELEVATOR_INSERT_REQUEUE:
 599                 /*
 600                  * If ordered flush isn't in progress, we do front
 601                  * insertion; otherwise, requests should be requeued
 602                  * in ordseq order.
 603                  */
 604                 rq->cmd_flags |= REQ_SOFTBARRIER;
 605
 606                 /*
 607                  * Most requeues happen because of a busy condition,
 608                  * don't force unplug of the queue for that case.
 609                  */
 610                 unplug_it = 0;
 611
 612                 if (q->ordseq == 0) {
 613                         list_add(&rq->queuelist, &q->queue_head);
 614                         break;
 615                 }
 616
 617                 ordseq = blk_ordered_req_seq(rq);
 618
 619                 list_for_each(pos, &q->queue_head) {
 620                         struct request *pos_rq = list_entry_rq(pos);
 621                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 622                                 break;
 623                 }
 624
 625                 list_add_tail(&rq->queuelist, pos);
 626                 break;
 627
 628         default:
 629                 printk(KERN_ERR "%s: bad insertion point %d\n",
 630                        __FUNCTION__, where);
 631                 BUG();
 632         }
 633
 634         if (unplug_it && blk_queue_plugged(q)) {
 635                 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
 636                         - q->in_flight;
 637
 638                 if (nrq >= q->unplug_thresh)
 639                         __generic_unplug_device(q);
 640         }
 641 }
 642
 643 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
 644                        int plug)
 645 {
 646         if (q->ordcolor)
 647                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 648
 649         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 650                 /*
 651                  * toggle ordered color
 652                  */
 653                 if (blk_barrier_rq(rq))
 654                         q->ordcolor ^= 1;
 655
 656                 /*
 657                  * barriers implicitly indicate back insertion
 658                  */
 659                 if (where == ELEVATOR_INSERT_SORT)
 660                         where = ELEVATOR_INSERT_BACK;
 661
 662                 /*
 663                  * this request is scheduling boundary, update
 664                  * end_sector
 665                  */
 666                 if (blk_fs_request(rq)) {
 667                         q->end_sector = rq_end_sector(rq);
 668                         q->boundary_rq = rq;
 669                 }
 670         } else if (!(rq->cmd_flags & REQ_ELVPRIV) && where == ELEVATOR_INSERT_SORT)
 671                 where = ELEVATOR_INSERT_BACK;
 672
 673         if (plug)
 674                 blk_plug_device(q);
 675
 676         elv_insert(q, rq, where);
 677 }
 678
 679 EXPORT_SYMBOL(__elv_add_request);
 680
 681 void elv_add_request(struct request_queue *q, struct request *rq, int where,
 682                      int plug)
 683 {
 684         unsigned long flags;
 685
 686         spin_lock_irqsave(q->queue_lock, flags);
 687         __elv_add_request(q, rq, where, plug);
 688         spin_unlock_irqrestore(q->queue_lock, flags);
 689 }
 690
 691 EXPORT_SYMBOL(elv_add_request);
 692
 693 static inline struct request *__elv_next_request(struct request_queue *q)
 694 {
 695         struct request *rq;
 696
 697         while (1) {
 698                 while (!list_empty(&q->queue_head)) {
 699                         rq = list_entry_rq(q->queue_head.next);
 700                         if (blk_do_ordered(q, &rq))
 701                                 return rq;
 702                 }
 703
 704                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
 705                         return NULL;
 706         }
 707 }
 708
 709 struct request *elv_next_request(struct request_queue *q)
 710 {
 711         struct request *rq;
 712         int ret;
 713
 714         while ((rq = __elv_next_request(q)) != NULL) {
 715                 /*
 716                  * Kill the empty barrier place holder, the driver must
 717                  * not ever see it.
 718                  */
 719                 if (blk_empty_barrier(rq)) {
 720                         end_queued_request(rq, 1);
 721                         continue;
 722                 }
 723                 if (!(rq->cmd_flags & REQ_STARTED)) {
 724                         /*
 725                          * This is the first time the device driver
 726                          * sees this request (possibly after
 727                          * requeueing).  Notify IO scheduler.
 728                          */
 729                         if (blk_sorted_rq(rq))
 730                                 elv_activate_rq(q, rq);
 731
 732                         /*
 733                          * just mark as started even if we don't start
 734                          * it, a request that has been delayed should
 735                          * not be passed by new incoming requests
 736                          */
 737                         rq->cmd_flags |= REQ_STARTED;
 738                         blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
 739                 }
 740
 741                 if (!q->boundary_rq || q->boundary_rq == rq) {
 742                         q->end_sector = rq_end_sector(rq);
 743                         q->boundary_rq = NULL;
 744                 }
 745
 746                 if ((rq->cmd_flags & REQ_DONTPREP) || !q->prep_rq_fn)
 747                         break;
 748
 749                 ret = q->prep_rq_fn(q, rq);
 750                 if (ret == BLKPREP_OK) {
 751                         break;
 752                 } else if (ret == BLKPREP_DEFER) {
 753                         /*
 754                          * the request may have been (partially) prepped.
 755                          * we need to keep this request in the front to
 756                          * avoid resource deadlock.  REQ_STARTED will
 757                          * prevent other fs requests from passing this one.
 758                          */
 759                         rq = NULL;
 760                         break;
 761                 } else if (ret == BLKPREP_KILL) {
 762                         rq->cmd_flags |= REQ_QUIET;
 763                         end_queued_request(rq, 0);
 764                 } else {
 765                         printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
 766                                                                 ret);
 767                         break;
 768                 }
 769         }
 770
 771         return rq;
 772 }
 773
 774 EXPORT_SYMBOL(elv_next_request);
 775
 776 void elv_dequeue_request(struct request_queue *q, struct request *rq)
 777 {
 778         BUG_ON(list_empty(&rq->queuelist));
 779         BUG_ON(ELV_ON_HASH(rq));
 780
 781         list_del_init(&rq->queuelist);
 782
 783         /*
 784          * the time frame between a request being removed from the lists
 785          * and to it is freed is accounted as io that is in progress at
 786          * the driver side.
 787          */
 788         if (blk_account_rq(rq))
 789                 q->in_flight++;
 790 }
 791
 792 EXPORT_SYMBOL(elv_dequeue_request);
 793
 794 int elv_queue_empty(struct request_queue *q)
 795 {
 796         elevator_t *e = q->elevator;
 797
 798         if (!list_empty(&q->queue_head))
 799                 return 0;
 800
 801         if (e->ops->elevator_queue_empty_fn)
 802                 return e->ops->elevator_queue_empty_fn(q);
 803
 804         return 1;
 805 }
 806
 807 EXPORT_SYMBOL(elv_queue_empty);
 808
 809 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 810 {
 811         elevator_t *e = q->elevator;
 812
 813         if (e->ops->elevator_latter_req_fn)
 814                 return e->ops->elevator_latter_req_fn(q, rq);
 815         return NULL;
 816 }
 817
 818 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 819 {
 820         elevator_t *e = q->elevator;
 821
 822         if (e->ops->elevator_former_req_fn)
 823                 return e->ops->elevator_former_req_fn(q, rq);
 824         return NULL;
 825 }
 826
 827 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 828 {
 829         elevator_t *e = q->elevator;
 830
 831         if (e->ops->elevator_set_req_fn)
 832                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 833
 834         rq->elevator_private = NULL;
 835         return 0;
 836 }
 837
 838 void elv_put_request(struct request_queue *q, struct request *rq)
 839 {
 840         elevator_t *e = q->elevator;
 841
 842         if (e->ops->elevator_put_req_fn)
 843                 e->ops->elevator_put_req_fn(rq);
 844 }
 845
 846 int elv_may_queue(struct request_queue *q, int rw)
 847 {
 848         elevator_t *e = q->elevator;
 849
 850         if (e->ops->elevator_may_queue_fn)
 851                 return e->ops->elevator_may_queue_fn(q, rw);
 852
 853         return ELV_MQUEUE_MAY;
 854 }
 855
 856 void elv_completed_request(struct request_queue *q, struct request *rq)
 857 {
 858         elevator_t *e = q->elevator;
 859
 860         /*
 861          * request is released from the driver, io must be done
 862          */
 863         if (blk_account_rq(rq)) {
 864                 q->in_flight--;
 865                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 866                         e->ops->elevator_completed_req_fn(q, rq);
 867         }
 868
 869         /*
 870          * Check if the queue is waiting for fs requests to be
 871          * drained for flush sequence.
 872          */
 873         if (unlikely(q->ordseq)) {
 874                 struct request *first_rq = list_entry_rq(q->queue_head.next);
 875                 if (q->in_flight == 0 &&
 876                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 877                     blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
 878                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 879                         q->request_fn(q);
 880                 }
 881         }
 882 }
 883
 884 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 885
 886 static ssize_t
 887 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 888 {
 889         elevator_t *e = container_of(kobj, elevator_t, kobj);
 890         struct elv_fs_entry *entry = to_elv(attr);
 891         ssize_t error;
 892
 893         if (!entry->show)
 894                 return -EIO;
 895
 896         mutex_lock(&e->sysfs_lock);
 897         error = e->ops ? entry->show(e, page) : -ENOENT;
 898         mutex_unlock(&e->sysfs_lock);
 899         return error;
 900 }
 901
 902 static ssize_t
 903 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 904                const char *page, size_t length)
 905 {
 906         elevator_t *e = container_of(kobj, elevator_t, kobj);
 907         struct elv_fs_entry *entry = to_elv(attr);
 908         ssize_t error;
 909
 910         if (!entry->store)
 911                 return -EIO;
 912
 913         mutex_lock(&e->sysfs_lock);
 914         error = e->ops ? entry->store(e, page, length) : -ENOENT;
 915         mutex_unlock(&e->sysfs_lock);
 916         return error;
 917 }
 918
 919 static struct sysfs_ops elv_sysfs_ops = {
 920         .show   = elv_attr_show,
 921         .store  = elv_attr_store,
 922 };
 923
 924 static struct kobj_type elv_ktype = {
 925         .sysfs_ops      = &elv_sysfs_ops,
 926         .release        = elevator_release,
 927 };
 928
 929 int elv_register_queue(struct request_queue *q)
 930 {
 931         elevator_t *e = q->elevator;
 932         int error;
 933
 934         e->kobj.parent = &q->kobj;
 935
 936         error = kobject_add(&e->kobj);
 937         if (!error) {
 938                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
 939                 if (attr) {
 940                         while (attr->attr.name) {
 941                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 942                                         break;
 943                                 attr++;
 944                         }
 945                 }
 946                 kobject_uevent(&e->kobj, KOBJ_ADD);
 947         }
 948         return error;
 949 }
 950
 951 static void __elv_unregister_queue(elevator_t *e)
 952 {
 953         kobject_uevent(&e->kobj, KOBJ_REMOVE);
 954         kobject_del(&e->kobj);
 955 }
 956
 957 void elv_unregister_queue(struct request_queue *q)
 958 {
 959         if (q)
 960                 __elv_unregister_queue(q->elevator);
 961 }
 962
 963 void elv_register(struct elevator_type *e)
 964 {
 965         char *def = "";
 966
 967         spin_lock(&elv_list_lock);
 968         BUG_ON(elevator_find(e->elevator_name));
 969         list_add_tail(&e->list, &elv_list);
 970         spin_unlock(&elv_list_lock);
 971
 972         if (!strcmp(e->elevator_name, chosen_elevator) ||
 973                         (!*chosen_elevator &&
 974                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 975                                 def = " (default)";
 976
 977         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name, def);
 978 }
 979 EXPORT_SYMBOL_GPL(elv_register);
 980
 981 void elv_unregister(struct elevator_type *e)
 982 {
 983         struct task_struct *g, *p;
 984
 985         /*
 986          * Iterate every thread in the process to remove the io contexts.
 987          */
 988         if (e->ops.trim) {
 989                 read_lock(&tasklist_lock);
 990                 do_each_thread(g, p) {
 991                         task_lock(p);
 992                         if (p->io_context)
 993                                 e->ops.trim(p->io_context);
 994                         task_unlock(p);
 995                 } while_each_thread(g, p);
 996                 read_unlock(&tasklist_lock);
 997         }
 998
 999         spin_lock(&elv_list_lock);
1000         list_del_init(&e->list);
1001         spin_unlock(&elv_list_lock);
1002 }
1003 EXPORT_SYMBOL_GPL(elv_unregister);
1004
1005 /*
1006  * switch to new_e io scheduler. be careful not to introduce deadlocks -
1007  * we don't free the old io scheduler, before we have allocated what we
1008  * need for the new one. this way we have a chance of going back to the old
1009  * one, if the new one fails init for some reason.
1010  */
1011 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1012 {
1013         elevator_t *old_elevator, *e;
1014         void *data;
1015
1016         /*
1017          * Allocate new elevator
1018          */
1019         e = elevator_alloc(q, new_e);
1020         if (!e)
1021                 return 0;
1022
1023         data = elevator_init_queue(q, e);
1024         if (!data) {
1025                 kobject_put(&e->kobj);
1026                 return 0;
1027         }
1028
1029         /*
1030          * Turn on BYPASS and drain all requests w/ elevator private data
1031          */
1032         spin_lock_irq(q->queue_lock);
1033
1034         set_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1035
1036         elv_drain_elevator(q);
1037
1038         while (q->rq.elvpriv) {
1039                 blk_remove_plug(q);
1040                 q->request_fn(q);
1041                 spin_unlock_irq(q->queue_lock);
1042                 msleep(10);
1043                 spin_lock_irq(q->queue_lock);
1044                 elv_drain_elevator(q);
1045         }
1046
1047         /*
1048          * Remember old elevator.
1049          */
1050         old_elevator = q->elevator;
1051
1052         /*
1053          * attach and start new elevator
1054          */
1055         elevator_attach(q, e, data);
1056
1057         spin_unlock_irq(q->queue_lock);
1058
1059         __elv_unregister_queue(old_elevator);
1060
1061         if (elv_register_queue(q))
1062                 goto fail_register;
1063
1064         /*
1065          * finally exit old elevator and turn off BYPASS.
1066          */
1067         elevator_exit(old_elevator);
1068         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1069         return 1;
1070
1071 fail_register:
1072         /*
1073          * switch failed, exit the new io scheduler and reattach the old
1074          * one again (along with re-adding the sysfs dir)
1075          */
1076         elevator_exit(e);
1077         q->elevator = old_elevator;
1078         elv_register_queue(q);
1079         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1080         return 0;
1081 }
1082
1083 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1084                           size_t count)
1085 {
1086         char elevator_name[ELV_NAME_MAX];
1087         size_t len;
1088         struct elevator_type *e;
1089
1090         elevator_name[sizeof(elevator_name) - 1] = '\0';
1091         strncpy(elevator_name, name, sizeof(elevator_name) - 1);
1092         len = strlen(elevator_name);
1093
1094         if (len && elevator_name[len - 1] == '\n')
1095                 elevator_name[len - 1] = '\0';
1096
1097         e = elevator_get(elevator_name);
1098         if (!e) {
1099                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1100                 return -EINVAL;
1101         }
1102
1103         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1104                 elevator_put(e);
1105                 return count;
1106         }
1107
1108         if (!elevator_switch(q, e))
1109                 printk(KERN_ERR "elevator: switch to %s failed\n",elevator_name);
1110         return count;
1111 }
1112
1113 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1114 {
1115         elevator_t *e = q->elevator;
1116         struct elevator_type *elv = e->elevator_type;
1117         struct elevator_type *__e;
1118         int len = 0;
1119
1120         spin_lock(&elv_list_lock);
1121         list_for_each_entry(__e, &elv_list, list) {
1122                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1123                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1124                 else
1125                         len += sprintf(name+len, "%s ", __e->elevator_name);
1126         }
1127         spin_unlock(&elv_list_lock);
1128
1129         len += sprintf(len+name, "\n");
1130         return len;
1131 }
1132
1133 struct request *elv_rb_former_request(struct request_queue *q,
1134                                       struct request *rq)
1135 {
1136         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1137
1138         if (rbprev)
1139                 return rb_entry_rq(rbprev);
1140
1141         return NULL;
1142 }
1143
1144 EXPORT_SYMBOL(elv_rb_former_request);
1145
1146 struct request *elv_rb_latter_request(struct request_queue *q,
1147                                       struct request *rq)
1148 {
1149         struct rb_node *rbnext = rb_next(&rq->rb_node);
1150
1151         if (rbnext)
1152                 return rb_entry_rq(rbnext);
1153
1154         return NULL;
1155 }
1156
1157 EXPORT_SYMBOL(elv_rb_latter_request);