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