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 void elv_activate_rq(request_queue_t *q, struct request *rq)
 273 {
 274         elevator_t *e = q->elevator;
 275
 276         if (e->ops->elevator_activate_req_fn)
 277                 e->ops->elevator_activate_req_fn(q, rq);
 278 }
 279
 280 static void elv_deactivate_rq(request_queue_t *q, struct request *rq)
 281 {
 282         elevator_t *e = q->elevator;
 283
 284         if (e->ops->elevator_deactivate_req_fn)
 285                 e->ops->elevator_deactivate_req_fn(q, rq);
 286 }
 287
 288 static inline void __elv_rqhash_del(struct request *rq)
 289 {
 290         hlist_del_init(&rq->hash);
 291 }
 292
 293 static void elv_rqhash_del(request_queue_t *q, struct request *rq)
 294 {
 295         if (ELV_ON_HASH(rq))
 296                 __elv_rqhash_del(rq);
 297 }
 298
 299 static void elv_rqhash_add(request_queue_t *q, struct request *rq)
 300 {
 301         elevator_t *e = q->elevator;
 302
 303         BUG_ON(ELV_ON_HASH(rq));
 304         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 305 }
 306
 307 static void elv_rqhash_reposition(request_queue_t *q, struct request *rq)
 308 {
 309         __elv_rqhash_del(rq);
 310         elv_rqhash_add(q, rq);
 311 }
 312
 313 static struct request *elv_rqhash_find(request_queue_t *q, sector_t offset)
 314 {
 315         elevator_t *e = q->elevator;
 316         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 317         struct hlist_node *entry, *next;
 318         struct request *rq;
 319
 320         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 321                 BUG_ON(!ELV_ON_HASH(rq));
 322
 323                 if (unlikely(!rq_mergeable(rq))) {
 324                         __elv_rqhash_del(rq);
 325                         continue;
 326                 }
 327
 328                 if (rq_hash_key(rq) == offset)
 329                         return rq;
 330         }
 331
 332         return NULL;
 333 }
 334
 335 /*
 336  * RB-tree support functions for inserting/lookup/removal of requests
 337  * in a sorted RB tree.
 338  */
 339 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 340 {
 341         struct rb_node **p = &root->rb_node;
 342         struct rb_node *parent = NULL;
 343         struct request *__rq;
 344
 345         while (*p) {
 346                 parent = *p;
 347                 __rq = rb_entry(parent, struct request, rb_node);
 348
 349                 if (rq->sector < __rq->sector)
 350                         p = &(*p)->rb_left;
 351                 else if (rq->sector > __rq->sector)
 352                         p = &(*p)->rb_right;
 353                 else
 354                         return __rq;
 355         }
 356
 357         rb_link_node(&rq->rb_node, parent, p);
 358         rb_insert_color(&rq->rb_node, root);
 359         return NULL;
 360 }
 361
 362 EXPORT_SYMBOL(elv_rb_add);
 363
 364 void elv_rb_del(struct rb_root *root, struct request *rq)
 365 {
 366         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 367         rb_erase(&rq->rb_node, root);
 368         RB_CLEAR_NODE(&rq->rb_node);
 369 }
 370
 371 EXPORT_SYMBOL(elv_rb_del);
 372
 373 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 374 {
 375         struct rb_node *n = root->rb_node;
 376         struct request *rq;
 377
 378         while (n) {
 379                 rq = rb_entry(n, struct request, rb_node);
 380
 381                 if (sector < rq->sector)
 382                         n = n->rb_left;
 383                 else if (sector > rq->sector)
 384                         n = n->rb_right;
 385                 else
 386                         return rq;
 387         }
 388
 389         return NULL;
 390 }
 391
 392 EXPORT_SYMBOL(elv_rb_find);
 393
 394 /*
 395  * Insert rq into dispatch queue of q.  Queue lock must be held on
 396  * entry.  rq is sort insted into the dispatch queue. To be used by
 397  * specific elevators.
 398  */
 399 void elv_dispatch_sort(request_queue_t *q, struct request *rq)
 400 {
 401         sector_t boundary;
 402         struct list_head *entry;
 403
 404         if (q->last_merge == rq)
 405                 q->last_merge = NULL;
 406
 407         elv_rqhash_del(q, rq);
 408
 409         q->nr_sorted--;
 410
 411         boundary = q->end_sector;
 412
 413         list_for_each_prev(entry, &q->queue_head) {
 414                 struct request *pos = list_entry_rq(entry);
 415
 416                 if (pos->cmd_flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
 417                         break;
 418                 if (rq->sector >= boundary) {
 419                         if (pos->sector < boundary)
 420                                 continue;
 421                 } else {
 422                         if (pos->sector >= boundary)
 423                                 break;
 424                 }
 425                 if (rq->sector >= pos->sector)
 426                         break;
 427         }
 428
 429         list_add(&rq->queuelist, entry);
 430 }
 431
 432 EXPORT_SYMBOL(elv_dispatch_sort);
 433
 434 /*
 435  * Insert rq into dispatch queue of q.  Queue lock must be held on
 436  * entry.  rq is added to the back of the dispatch queue. To be used by
 437  * specific elevators.
 438  */
 439 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 440 {
 441         if (q->last_merge == rq)
 442                 q->last_merge = NULL;
 443
 444         elv_rqhash_del(q, rq);
 445
 446         q->nr_sorted--;
 447
 448         q->end_sector = rq_end_sector(rq);
 449         q->boundary_rq = rq;
 450         list_add_tail(&rq->queuelist, &q->queue_head);
 451 }
 452
 453 EXPORT_SYMBOL(elv_dispatch_add_tail);
 454
 455 int elv_merge(request_queue_t *q, struct request **req, struct bio *bio)
 456 {
 457         elevator_t *e = q->elevator;
 458         struct request *__rq;
 459         int ret;
 460
 461         /*
 462          * First try one-hit cache.
 463          */
 464         if (q->last_merge) {
 465                 ret = elv_try_merge(q->last_merge, bio);
 466                 if (ret != ELEVATOR_NO_MERGE) {
 467                         *req = q->last_merge;
 468                         return ret;
 469                 }
 470         }
 471
 472         /*
 473          * See if our hash lookup can find a potential backmerge.
 474          */
 475         __rq = elv_rqhash_find(q, bio->bi_sector);
 476         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 477                 *req = __rq;
 478                 return ELEVATOR_BACK_MERGE;
 479         }
 480
 481         if (e->ops->elevator_merge_fn)
 482                 return e->ops->elevator_merge_fn(q, req, bio);
 483
 484         return ELEVATOR_NO_MERGE;
 485 }
 486
 487 void elv_merged_request(request_queue_t *q, struct request *rq, int type)
 488 {
 489         elevator_t *e = q->elevator;
 490
 491         if (e->ops->elevator_merged_fn)
 492                 e->ops->elevator_merged_fn(q, rq, type);
 493
 494         if (type == ELEVATOR_BACK_MERGE)
 495                 elv_rqhash_reposition(q, rq);
 496
 497         q->last_merge = rq;
 498 }
 499
 500 void elv_merge_requests(request_queue_t *q, struct request *rq,
 501                              struct request *next)
 502 {
 503         elevator_t *e = q->elevator;
 504
 505         if (e->ops->elevator_merge_req_fn)
 506                 e->ops->elevator_merge_req_fn(q, rq, next);
 507
 508         elv_rqhash_reposition(q, rq);
 509         elv_rqhash_del(q, next);
 510
 511         q->nr_sorted--;
 512         q->last_merge = rq;
 513 }
 514
 515 void elv_requeue_request(request_queue_t *q, struct request *rq)
 516 {
 517         /*
 518          * it already went through dequeue, we need to decrement the
 519          * in_flight count again
 520          */
 521         if (blk_account_rq(rq)) {
 522                 q->in_flight--;
 523                 if (blk_sorted_rq(rq))
 524                         elv_deactivate_rq(q, rq);
 525         }
 526
 527         rq->cmd_flags &= ~REQ_STARTED;
 528
 529         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 530 }
 531
 532 static void elv_drain_elevator(request_queue_t *q)
 533 {
 534         static int printed;
 535         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 536                 ;
 537         if (q->nr_sorted == 0)
 538                 return;
 539         if (printed++ < 10) {
 540                 printk(KERN_ERR "%s: forced dispatching is broken "
 541                        "(nr_sorted=%u), please report this\n",
 542                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 543         }
 544 }
 545
 546 void elv_insert(request_queue_t *q, struct request *rq, int where)
 547 {
 548         struct list_head *pos;
 549         unsigned ordseq;
 550         int unplug_it = 1;
 551
 552         blk_add_trace_rq(q, rq, BLK_TA_INSERT);
 553
 554         rq->q = q;
 555
 556         switch (where) {
 557         case ELEVATOR_INSERT_FRONT:
 558                 rq->cmd_flags |= REQ_SOFTBARRIER;
 559
 560                 list_add(&rq->queuelist, &q->queue_head);
 561                 break;
 562
 563         case ELEVATOR_INSERT_BACK:
 564                 rq->cmd_flags |= REQ_SOFTBARRIER;
 565                 elv_drain_elevator(q);
 566                 list_add_tail(&rq->queuelist, &q->queue_head);
 567                 /*
 568                  * We kick the queue here for the following reasons.
 569                  * - The elevator might have returned NULL previously
 570                  *   to delay requests and returned them now.  As the
 571                  *   queue wasn't empty before this request, ll_rw_blk
 572                  *   won't run the queue on return, resulting in hang.
 573                  * - Usually, back inserted requests won't be merged
 574                  *   with anything.  There's no point in delaying queue
 575                  *   processing.
 576                  */
 577                 blk_remove_plug(q);
 578                 q->request_fn(q);
 579                 break;
 580
 581         case ELEVATOR_INSERT_SORT:
 582                 BUG_ON(!blk_fs_request(rq));
 583                 rq->cmd_flags |= REQ_SORTED;
 584                 q->nr_sorted++;
 585                 if (rq_mergeable(rq)) {
 586                         elv_rqhash_add(q, rq);
 587                         if (!q->last_merge)
 588                                 q->last_merge = rq;
 589                 }
 590
 591                 /*
 592                  * Some ioscheds (cfq) run q->request_fn directly, so
 593                  * rq cannot be accessed after calling
 594                  * elevator_add_req_fn.
 595                  */
 596                 q->elevator->ops->elevator_add_req_fn(q, rq);
 597                 break;
 598
 599         case ELEVATOR_INSERT_REQUEUE:
 600                 /*
 601                  * If ordered flush isn't in progress, we do front
 602                  * insertion; otherwise, requests should be requeued
 603                  * in ordseq order.
 604                  */
 605                 rq->cmd_flags |= REQ_SOFTBARRIER;
 606
 607                 /*
 608                  * Most requeues happen because of a busy condition,
 609                  * don't force unplug of the queue for that case.
 610                  */
 611                 unplug_it = 0;
 612
 613                 if (q->ordseq == 0) {
 614                         list_add(&rq->queuelist, &q->queue_head);
 615                         break;
 616                 }
 617
 618                 ordseq = blk_ordered_req_seq(rq);
 619
 620                 list_for_each(pos, &q->queue_head) {
 621                         struct request *pos_rq = list_entry_rq(pos);
 622                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 623                                 break;
 624                 }
 625
 626                 list_add_tail(&rq->queuelist, pos);
 627                 break;
 628
 629         default:
 630                 printk(KERN_ERR "%s: bad insertion point %d\n",
 631                        __FUNCTION__, where);
 632                 BUG();
 633         }
 634
 635         if (unplug_it && blk_queue_plugged(q)) {
 636                 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
 637                         - q->in_flight;
 638
 639                 if (nrq >= q->unplug_thresh)
 640                         __generic_unplug_device(q);
 641         }
 642 }
 643
 644 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
 645                        int plug)
 646 {
 647         if (q->ordcolor)
 648                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 649
 650         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 651                 /*
 652                  * toggle ordered color
 653                  */
 654                 if (blk_barrier_rq(rq))
 655                         q->ordcolor ^= 1;
 656
 657                 /*
 658                  * barriers implicitly indicate back insertion
 659                  */
 660                 if (where == ELEVATOR_INSERT_SORT)
 661                         where = ELEVATOR_INSERT_BACK;
 662
 663                 /*
 664                  * this request is scheduling boundary, update
 665                  * end_sector
 666                  */
 667                 if (blk_fs_request(rq)) {
 668                         q->end_sector = rq_end_sector(rq);
 669                         q->boundary_rq = rq;
 670                 }
 671         } else if (!(rq->cmd_flags & REQ_ELVPRIV) && where == ELEVATOR_INSERT_SORT)
 672                 where = ELEVATOR_INSERT_BACK;
 673
 674         if (plug)
 675                 blk_plug_device(q);
 676
 677         elv_insert(q, rq, where);
 678 }
 679
 680 EXPORT_SYMBOL(__elv_add_request);
 681
 682 void elv_add_request(request_queue_t *q, struct request *rq, int where,
 683                      int plug)
 684 {
 685         unsigned long flags;
 686
 687         spin_lock_irqsave(q->queue_lock, flags);
 688         __elv_add_request(q, rq, where, plug);
 689         spin_unlock_irqrestore(q->queue_lock, flags);
 690 }
 691
 692 EXPORT_SYMBOL(elv_add_request);
 693
 694 static inline struct request *__elv_next_request(request_queue_t *q)
 695 {
 696         struct request *rq;
 697
 698         while (1) {
 699                 while (!list_empty(&q->queue_head)) {
 700                         rq = list_entry_rq(q->queue_head.next);
 701                         if (blk_do_ordered(q, &rq))
 702                                 return rq;
 703                 }
 704
 705                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
 706                         return NULL;
 707         }
 708 }
 709
 710 struct request *elv_next_request(request_queue_t *q)
 711 {
 712         struct request *rq;
 713         int ret;
 714
 715         while ((rq = __elv_next_request(q)) != NULL) {
 716                 if (!(rq->cmd_flags & REQ_STARTED)) {
 717                         /*
 718                          * This is the first time the device driver
 719                          * sees this request (possibly after
 720                          * requeueing).  Notify IO scheduler.
 721                          */
 722                         if (blk_sorted_rq(rq))
 723                                 elv_activate_rq(q, rq);
 724
 725                         /*
 726                          * just mark as started even if we don't start
 727                          * it, a request that has been delayed should
 728                          * not be passed by new incoming requests
 729                          */
 730                         rq->cmd_flags |= REQ_STARTED;
 731                         blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
 732                 }
 733
 734                 if (!q->boundary_rq || q->boundary_rq == rq) {
 735                         q->end_sector = rq_end_sector(rq);
 736                         q->boundary_rq = NULL;
 737                 }
 738
 739                 if ((rq->cmd_flags & REQ_DONTPREP) || !q->prep_rq_fn)
 740                         break;
 741
 742                 ret = q->prep_rq_fn(q, rq);
 743                 if (ret == BLKPREP_OK) {
 744                         break;
 745                 } else if (ret == BLKPREP_DEFER) {
 746                         /*
 747                          * the request may have been (partially) prepped.
 748                          * we need to keep this request in the front to
 749                          * avoid resource deadlock.  REQ_STARTED will
 750                          * prevent other fs requests from passing this one.
 751                          */
 752                         rq = NULL;
 753                         break;
 754                 } else if (ret == BLKPREP_KILL) {
 755                         int nr_bytes = rq->hard_nr_sectors << 9;
 756
 757                         if (!nr_bytes)
 758                                 nr_bytes = rq->data_len;
 759
 760                         blkdev_dequeue_request(rq);
 761                         rq->cmd_flags |= REQ_QUIET;
 762                         end_that_request_chunk(rq, 0, nr_bytes);
 763                         end_that_request_last(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(request_queue_t *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(request_queue_t *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(request_queue_t *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(request_queue_t *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(request_queue_t *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(request_queue_t *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(request_queue_t *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(request_queue_t *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 int elv_register(struct elevator_type *e)
 964 {
 965         spin_lock_irq(&elv_list_lock);
 966         BUG_ON(elevator_find(e->elevator_name));
 967         list_add_tail(&e->list, &elv_list);
 968         spin_unlock_irq(&elv_list_lock);
 969
 970         printk(KERN_INFO "io scheduler %s registered", e->elevator_name);
 971         if (!strcmp(e->elevator_name, chosen_elevator) ||
 972                         (!*chosen_elevator &&
 973                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 974                                 printk(" (default)");
 975         printk("\n");
 976         return 0;
 977 }
 978 EXPORT_SYMBOL_GPL(elv_register);
 979
 980 void elv_unregister(struct elevator_type *e)
 981 {
 982         struct task_struct *g, *p;
 983
 984         /*
 985          * Iterate every thread in the process to remove the io contexts.
 986          */
 987         if (e->ops.trim) {
 988                 read_lock(&tasklist_lock);
 989                 do_each_thread(g, p) {
 990                         task_lock(p);
 991                         if (p->io_context)
 992                                 e->ops.trim(p->io_context);
 993                         task_unlock(p);
 994                 } while_each_thread(g, p);
 995                 read_unlock(&tasklist_lock);
 996         }
 997
 998         spin_lock_irq(&elv_list_lock);
 999         list_del_init(&e->list);
1000         spin_unlock_irq(&elv_list_lock);
1001 }
1002 EXPORT_SYMBOL_GPL(elv_unregister);
1003
1004 /*
1005  * switch to new_e io scheduler. be careful not to introduce deadlocks -
1006  * we don't free the old io scheduler, before we have allocated what we
1007  * need for the new one. this way we have a chance of going back to the old
1008  * one, if the new one fails init for some reason.
1009  */
1010 static int elevator_switch(request_queue_t *q, struct elevator_type *new_e)
1011 {
1012         elevator_t *old_elevator, *e;
1013         void *data;
1014
1015         /*
1016          * Allocate new elevator
1017          */
1018         e = elevator_alloc(q, new_e);
1019         if (!e)
1020                 return 0;
1021
1022         data = elevator_init_queue(q, e);
1023         if (!data) {
1024                 kobject_put(&e->kobj);
1025                 return 0;
1026         }
1027
1028         /*
1029          * Turn on BYPASS and drain all requests w/ elevator private data
1030          */
1031         spin_lock_irq(q->queue_lock);
1032
1033         set_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1034
1035         elv_drain_elevator(q);
1036
1037         while (q->rq.elvpriv) {
1038                 blk_remove_plug(q);
1039                 q->request_fn(q);
1040                 spin_unlock_irq(q->queue_lock);
1041                 msleep(10);
1042                 spin_lock_irq(q->queue_lock);
1043                 elv_drain_elevator(q);
1044         }
1045
1046         /*
1047          * Remember old elevator.
1048          */
1049         old_elevator = q->elevator;
1050
1051         /*
1052          * attach and start new elevator
1053          */
1054         elevator_attach(q, e, data);
1055
1056         spin_unlock_irq(q->queue_lock);
1057
1058         __elv_unregister_queue(old_elevator);
1059
1060         if (elv_register_queue(q))
1061                 goto fail_register;
1062
1063         /*
1064          * finally exit old elevator and turn off BYPASS.
1065          */
1066         elevator_exit(old_elevator);
1067         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1068         return 1;
1069
1070 fail_register:
1071         /*
1072          * switch failed, exit the new io scheduler and reattach the old
1073          * one again (along with re-adding the sysfs dir)
1074          */
1075         elevator_exit(e);
1076         q->elevator = old_elevator;
1077         elv_register_queue(q);
1078         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1079         return 0;
1080 }
1081
1082 ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count)
1083 {
1084         char elevator_name[ELV_NAME_MAX];
1085         size_t len;
1086         struct elevator_type *e;
1087
1088         elevator_name[sizeof(elevator_name) - 1] = '\0';
1089         strncpy(elevator_name, name, sizeof(elevator_name) - 1);
1090         len = strlen(elevator_name);
1091
1092         if (len && elevator_name[len - 1] == '\n')
1093                 elevator_name[len - 1] = '\0';
1094
1095         e = elevator_get(elevator_name);
1096         if (!e) {
1097                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1098                 return -EINVAL;
1099         }
1100
1101         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1102                 elevator_put(e);
1103                 return count;
1104         }
1105
1106         if (!elevator_switch(q, e))
1107                 printk(KERN_ERR "elevator: switch to %s failed\n",elevator_name);
1108         return count;
1109 }
1110
1111 ssize_t elv_iosched_show(request_queue_t *q, char *name)
1112 {
1113         elevator_t *e = q->elevator;
1114         struct elevator_type *elv = e->elevator_type;
1115         struct list_head *entry;
1116         int len = 0;
1117
1118         spin_lock_irq(&elv_list_lock);
1119         list_for_each(entry, &elv_list) {
1120                 struct elevator_type *__e;
1121
1122                 __e = list_entry(entry, struct elevator_type, list);
1123                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1124                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1125                 else
1126                         len += sprintf(name+len, "%s ", __e->elevator_name);
1127         }
1128         spin_unlock_irq(&elv_list_lock);
1129
1130         len += sprintf(len+name, "\n");
1131         return len;
1132 }
1133
1134 struct request *elv_rb_former_request(request_queue_t *q, 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(request_queue_t *q, struct request *rq)
1147 {
1148         struct rb_node *rbnext = rb_next(&rq->rb_node);
1149
1150         if (rbnext)
1151                 return rb_entry_rq(rbnext);
1152
1153         return NULL;
1154 }
1155
1156 EXPORT_SYMBOL(elv_rb_latter_request);