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@suse.de> :
   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/config.h>
  31 #include <linux/module.h>
  32 #include <linux/slab.h>
  33 #include <linux/init.h>
  34 #include <linux/compiler.h>
  35 #include <linux/delay.h>
  36
  37 #include <asm/uaccess.h>
  38
  39 static DEFINE_SPINLOCK(elv_list_lock);
  40 static LIST_HEAD(elv_list);
  41
  42 /*
  43  * can we safely merge with this request?
  44  */
  45 inline int elv_rq_merge_ok(struct request *rq, struct bio *bio)
  46 {
  47         if (!rq_mergeable(rq))
  48                 return 0;
  49
  50         /*
  51          * different data direction or already started, don't merge
  52          */
  53         if (bio_data_dir(bio) != rq_data_dir(rq))
  54                 return 0;
  55
  56         /*
  57          * same device and no special stuff set, merge is ok
  58          */
  59         if (rq->rq_disk == bio->bi_bdev->bd_disk &&
  60             !rq->waiting && !rq->special)
  61                 return 1;
  62
  63         return 0;
  64 }
  65 EXPORT_SYMBOL(elv_rq_merge_ok);
  66
  67 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
  68 {
  69         int ret = ELEVATOR_NO_MERGE;
  70
  71         /*
  72          * we can merge and sequence is ok, check if it's possible
  73          */
  74         if (elv_rq_merge_ok(__rq, bio)) {
  75                 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
  76                         ret = ELEVATOR_BACK_MERGE;
  77                 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
  78                         ret = ELEVATOR_FRONT_MERGE;
  79         }
  80
  81         return ret;
  82 }
  83
  84 static struct elevator_type *elevator_find(const char *name)
  85 {
  86         struct elevator_type *e = NULL;
  87         struct list_head *entry;
  88
  89         list_for_each(entry, &elv_list) {
  90                 struct elevator_type *__e;
  91
  92                 __e = list_entry(entry, struct elevator_type, list);
  93
  94                 if (!strcmp(__e->elevator_name, name)) {
  95                         e = __e;
  96                         break;
  97                 }
  98         }
  99
 100         return e;
 101 }
 102
 103 static void elevator_put(struct elevator_type *e)
 104 {
 105         module_put(e->elevator_owner);
 106 }
 107
 108 static struct elevator_type *elevator_get(const char *name)
 109 {
 110         struct elevator_type *e;
 111
 112         spin_lock_irq(&elv_list_lock);
 113
 114         e = elevator_find(name);
 115         if (e && !try_module_get(e->elevator_owner))
 116                 e = NULL;
 117
 118         spin_unlock_irq(&elv_list_lock);
 119
 120         return e;
 121 }
 122
 123 static int elevator_attach(request_queue_t *q, struct elevator_type *e,
 124                            struct elevator_queue *eq)
 125 {
 126         int ret = 0;
 127
 128         memset(eq, 0, sizeof(*eq));
 129         eq->ops = &e->ops;
 130         eq->elevator_type = e;
 131
 132         q->elevator = eq;
 133
 134         if (eq->ops->elevator_init_fn)
 135                 ret = eq->ops->elevator_init_fn(q, eq);
 136
 137         return ret;
 138 }
 139
 140 static char chosen_elevator[16];
 141
 142 static int __init elevator_setup(char *str)
 143 {
 144         /*
 145          * Be backwards-compatible with previous kernels, so users
 146          * won't get the wrong elevator.
 147          */
 148         if (!strcmp(str, "as"))
 149                 strcpy(chosen_elevator, "anticipatory");
 150         else
 151                 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 152         return 0;
 153 }
 154
 155 __setup("elevator=", elevator_setup);
 156
 157 int elevator_init(request_queue_t *q, char *name)
 158 {
 159         struct elevator_type *e = NULL;
 160         struct elevator_queue *eq;
 161         int ret = 0;
 162
 163         INIT_LIST_HEAD(&q->queue_head);
 164         q->last_merge = NULL;
 165         q->end_sector = 0;
 166         q->boundary_rq = NULL;
 167
 168         if (name && !(e = elevator_get(name)))
 169                 return -EINVAL;
 170
 171         if (!e && *chosen_elevator && !(e = elevator_get(chosen_elevator)))
 172                 printk("I/O scheduler %s not found\n", chosen_elevator);
 173
 174         if (!e && !(e = elevator_get(CONFIG_DEFAULT_IOSCHED))) {
 175                 printk("Default I/O scheduler not found, using no-op\n");
 176                 e = elevator_get("noop");
 177         }
 178
 179         eq = kmalloc(sizeof(struct elevator_queue), GFP_KERNEL);
 180         if (!eq) {
 181                 elevator_put(e);
 182                 return -ENOMEM;
 183         }
 184
 185         ret = elevator_attach(q, e, eq);
 186         if (ret) {
 187                 kfree(eq);
 188                 elevator_put(e);
 189         }
 190
 191         return ret;
 192 }
 193
 194 void elevator_exit(elevator_t *e)
 195 {
 196         if (e->ops->elevator_exit_fn)
 197                 e->ops->elevator_exit_fn(e);
 198
 199         elevator_put(e->elevator_type);
 200         e->elevator_type = NULL;
 201         kfree(e);
 202 }
 203
 204 /*
 205  * Insert rq into dispatch queue of q.  Queue lock must be held on
 206  * entry.  If sort != 0, rq is sort-inserted; otherwise, rq will be
 207  * appended to the dispatch queue.  To be used by specific elevators.
 208  */
 209 void elv_dispatch_sort(request_queue_t *q, struct request *rq)
 210 {
 211         sector_t boundary;
 212         struct list_head *entry;
 213
 214         if (q->last_merge == rq)
 215                 q->last_merge = NULL;
 216         q->nr_sorted--;
 217
 218         boundary = q->end_sector;
 219
 220         list_for_each_prev(entry, &q->queue_head) {
 221                 struct request *pos = list_entry_rq(entry);
 222
 223                 if (pos->flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
 224                         break;
 225                 if (rq->sector >= boundary) {
 226                         if (pos->sector < boundary)
 227                                 continue;
 228                 } else {
 229                         if (pos->sector >= boundary)
 230                                 break;
 231                 }
 232                 if (rq->sector >= pos->sector)
 233                         break;
 234         }
 235
 236         list_add(&rq->queuelist, entry);
 237 }
 238
 239 int elv_merge(request_queue_t *q, struct request **req, struct bio *bio)
 240 {
 241         elevator_t *e = q->elevator;
 242         int ret;
 243
 244         if (q->last_merge) {
 245                 ret = elv_try_merge(q->last_merge, bio);
 246                 if (ret != ELEVATOR_NO_MERGE) {
 247                         *req = q->last_merge;
 248                         return ret;
 249                 }
 250         }
 251
 252         if (e->ops->elevator_merge_fn)
 253                 return e->ops->elevator_merge_fn(q, req, bio);
 254
 255         return ELEVATOR_NO_MERGE;
 256 }
 257
 258 void elv_merged_request(request_queue_t *q, struct request *rq)
 259 {
 260         elevator_t *e = q->elevator;
 261
 262         if (e->ops->elevator_merged_fn)
 263                 e->ops->elevator_merged_fn(q, rq);
 264
 265         q->last_merge = rq;
 266 }
 267
 268 void elv_merge_requests(request_queue_t *q, struct request *rq,
 269                              struct request *next)
 270 {
 271         elevator_t *e = q->elevator;
 272
 273         if (e->ops->elevator_merge_req_fn)
 274                 e->ops->elevator_merge_req_fn(q, rq, next);
 275         q->nr_sorted--;
 276
 277         q->last_merge = rq;
 278 }
 279
 280 void elv_requeue_request(request_queue_t *q, struct request *rq)
 281 {
 282         elevator_t *e = q->elevator;
 283
 284         /*
 285          * it already went through dequeue, we need to decrement the
 286          * in_flight count again
 287          */
 288         if (blk_account_rq(rq)) {
 289                 q->in_flight--;
 290                 if (blk_sorted_rq(rq) && e->ops->elevator_deactivate_req_fn)
 291                         e->ops->elevator_deactivate_req_fn(q, rq);
 292         }
 293
 294         rq->flags &= ~REQ_STARTED;
 295
 296         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 297 }
 298
 299 static void elv_drain_elevator(request_queue_t *q)
 300 {
 301         static int printed;
 302         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 303                 ;
 304         if (q->nr_sorted == 0)
 305                 return;
 306         if (printed++ < 10) {
 307                 printk(KERN_ERR "%s: forced dispatching is broken "
 308                        "(nr_sorted=%u), please report this\n",
 309                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 310         }
 311 }
 312
 313 void elv_insert(request_queue_t *q, struct request *rq, int where)
 314 {
 315         struct list_head *pos;
 316         unsigned ordseq;
 317
 318         rq->q = q;
 319
 320         switch (where) {
 321         case ELEVATOR_INSERT_FRONT:
 322                 rq->flags |= REQ_SOFTBARRIER;
 323
 324                 list_add(&rq->queuelist, &q->queue_head);
 325                 break;
 326
 327         case ELEVATOR_INSERT_BACK:
 328                 rq->flags |= REQ_SOFTBARRIER;
 329                 elv_drain_elevator(q);
 330                 list_add_tail(&rq->queuelist, &q->queue_head);
 331                 /*
 332                  * We kick the queue here for the following reasons.
 333                  * - The elevator might have returned NULL previously
 334                  *   to delay requests and returned them now.  As the
 335                  *   queue wasn't empty before this request, ll_rw_blk
 336                  *   won't run the queue on return, resulting in hang.
 337                  * - Usually, back inserted requests won't be merged
 338                  *   with anything.  There's no point in delaying queue
 339                  *   processing.
 340                  */
 341                 blk_remove_plug(q);
 342                 q->request_fn(q);
 343                 break;
 344
 345         case ELEVATOR_INSERT_SORT:
 346                 BUG_ON(!blk_fs_request(rq));
 347                 rq->flags |= REQ_SORTED;
 348                 q->nr_sorted++;
 349                 if (q->last_merge == NULL && rq_mergeable(rq))
 350                         q->last_merge = rq;
 351                 /*
 352                  * Some ioscheds (cfq) run q->request_fn directly, so
 353                  * rq cannot be accessed after calling
 354                  * elevator_add_req_fn.
 355                  */
 356                 q->elevator->ops->elevator_add_req_fn(q, rq);
 357                 break;
 358
 359         case ELEVATOR_INSERT_REQUEUE:
 360                 /*
 361                  * If ordered flush isn't in progress, we do front
 362                  * insertion; otherwise, requests should be requeued
 363                  * in ordseq order.
 364                  */
 365                 rq->flags |= REQ_SOFTBARRIER;
 366
 367                 if (q->ordseq == 0) {
 368                         list_add(&rq->queuelist, &q->queue_head);
 369                         break;
 370                 }
 371
 372                 ordseq = blk_ordered_req_seq(rq);
 373
 374                 list_for_each(pos, &q->queue_head) {
 375                         struct request *pos_rq = list_entry_rq(pos);
 376                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 377                                 break;
 378                 }
 379
 380                 list_add_tail(&rq->queuelist, pos);
 381                 break;
 382
 383         default:
 384                 printk(KERN_ERR "%s: bad insertion point %d\n",
 385                        __FUNCTION__, where);
 386                 BUG();
 387         }
 388
 389         if (blk_queue_plugged(q)) {
 390                 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
 391                         - q->in_flight;
 392
 393                 if (nrq >= q->unplug_thresh)
 394                         __generic_unplug_device(q);
 395         }
 396 }
 397
 398 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
 399                        int plug)
 400 {
 401         if (q->ordcolor)
 402                 rq->flags |= REQ_ORDERED_COLOR;
 403
 404         if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 405                 /*
 406                  * toggle ordered color
 407                  */
 408                 if (blk_barrier_rq(rq))
 409                         q->ordcolor ^= 1;
 410
 411                 /*
 412                  * barriers implicitly indicate back insertion
 413                  */
 414                 if (where == ELEVATOR_INSERT_SORT)
 415                         where = ELEVATOR_INSERT_BACK;
 416
 417                 /*
 418                  * this request is scheduling boundary, update
 419                  * end_sector
 420                  */
 421                 if (blk_fs_request(rq)) {
 422                         q->end_sector = rq_end_sector(rq);
 423                         q->boundary_rq = rq;
 424                 }
 425         } else if (!(rq->flags & REQ_ELVPRIV) && where == ELEVATOR_INSERT_SORT)
 426                 where = ELEVATOR_INSERT_BACK;
 427
 428         if (plug)
 429                 blk_plug_device(q);
 430
 431         elv_insert(q, rq, where);
 432 }
 433
 434 void elv_add_request(request_queue_t *q, struct request *rq, int where,
 435                      int plug)
 436 {
 437         unsigned long flags;
 438
 439         spin_lock_irqsave(q->queue_lock, flags);
 440         __elv_add_request(q, rq, where, plug);
 441         spin_unlock_irqrestore(q->queue_lock, flags);
 442 }
 443
 444 static inline struct request *__elv_next_request(request_queue_t *q)
 445 {
 446         struct request *rq;
 447
 448         while (1) {
 449                 while (!list_empty(&q->queue_head)) {
 450                         rq = list_entry_rq(q->queue_head.next);
 451                         if (blk_do_ordered(q, &rq))
 452                                 return rq;
 453                 }
 454
 455                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
 456                         return NULL;
 457         }
 458 }
 459
 460 struct request *elv_next_request(request_queue_t *q)
 461 {
 462         struct request *rq;
 463         int ret;
 464
 465         while ((rq = __elv_next_request(q)) != NULL) {
 466                 if (!(rq->flags & REQ_STARTED)) {
 467                         elevator_t *e = q->elevator;
 468
 469                         /*
 470                          * This is the first time the device driver
 471                          * sees this request (possibly after
 472                          * requeueing).  Notify IO scheduler.
 473                          */
 474                         if (blk_sorted_rq(rq) &&
 475                             e->ops->elevator_activate_req_fn)
 476                                 e->ops->elevator_activate_req_fn(q, rq);
 477
 478                         /*
 479                          * just mark as started even if we don't start
 480                          * it, a request that has been delayed should
 481                          * not be passed by new incoming requests
 482                          */
 483                         rq->flags |= REQ_STARTED;
 484                 }
 485
 486                 if (!q->boundary_rq || q->boundary_rq == rq) {
 487                         q->end_sector = rq_end_sector(rq);
 488                         q->boundary_rq = NULL;
 489                 }
 490
 491                 if ((rq->flags & REQ_DONTPREP) || !q->prep_rq_fn)
 492                         break;
 493
 494                 ret = q->prep_rq_fn(q, rq);
 495                 if (ret == BLKPREP_OK) {
 496                         break;
 497                 } else if (ret == BLKPREP_DEFER) {
 498                         /*
 499                          * the request may have been (partially) prepped.
 500                          * we need to keep this request in the front to
 501                          * avoid resource deadlock.  REQ_STARTED will
 502                          * prevent other fs requests from passing this one.
 503                          */
 504                         rq = NULL;
 505                         break;
 506                 } else if (ret == BLKPREP_KILL) {
 507                         int nr_bytes = rq->hard_nr_sectors << 9;
 508
 509                         if (!nr_bytes)
 510                                 nr_bytes = rq->data_len;
 511
 512                         blkdev_dequeue_request(rq);
 513                         rq->flags |= REQ_QUIET;
 514                         end_that_request_chunk(rq, 0, nr_bytes);
 515                         end_that_request_last(rq, 0);
 516                 } else {
 517                         printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
 518                                                                 ret);
 519                         break;
 520                 }
 521         }
 522
 523         return rq;
 524 }
 525
 526 void elv_dequeue_request(request_queue_t *q, struct request *rq)
 527 {
 528         BUG_ON(list_empty(&rq->queuelist));
 529
 530         list_del_init(&rq->queuelist);
 531
 532         /*
 533          * the time frame between a request being removed from the lists
 534          * and to it is freed is accounted as io that is in progress at
 535          * the driver side.
 536          */
 537         if (blk_account_rq(rq))
 538                 q->in_flight++;
 539 }
 540
 541 int elv_queue_empty(request_queue_t *q)
 542 {
 543         elevator_t *e = q->elevator;
 544
 545         if (!list_empty(&q->queue_head))
 546                 return 0;
 547
 548         if (e->ops->elevator_queue_empty_fn)
 549                 return e->ops->elevator_queue_empty_fn(q);
 550
 551         return 1;
 552 }
 553
 554 struct request *elv_latter_request(request_queue_t *q, struct request *rq)
 555 {
 556         elevator_t *e = q->elevator;
 557
 558         if (e->ops->elevator_latter_req_fn)
 559                 return e->ops->elevator_latter_req_fn(q, rq);
 560         return NULL;
 561 }
 562
 563 struct request *elv_former_request(request_queue_t *q, struct request *rq)
 564 {
 565         elevator_t *e = q->elevator;
 566
 567         if (e->ops->elevator_former_req_fn)
 568                 return e->ops->elevator_former_req_fn(q, rq);
 569         return NULL;
 570 }
 571
 572 int elv_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
 573                     gfp_t gfp_mask)
 574 {
 575         elevator_t *e = q->elevator;
 576
 577         if (e->ops->elevator_set_req_fn)
 578                 return e->ops->elevator_set_req_fn(q, rq, bio, gfp_mask);
 579
 580         rq->elevator_private = NULL;
 581         return 0;
 582 }
 583
 584 void elv_put_request(request_queue_t *q, struct request *rq)
 585 {
 586         elevator_t *e = q->elevator;
 587
 588         if (e->ops->elevator_put_req_fn)
 589                 e->ops->elevator_put_req_fn(q, rq);
 590 }
 591
 592 int elv_may_queue(request_queue_t *q, int rw, struct bio *bio)
 593 {
 594         elevator_t *e = q->elevator;
 595
 596         if (e->ops->elevator_may_queue_fn)
 597                 return e->ops->elevator_may_queue_fn(q, rw, bio);
 598
 599         return ELV_MQUEUE_MAY;
 600 }
 601
 602 void elv_completed_request(request_queue_t *q, struct request *rq)
 603 {
 604         elevator_t *e = q->elevator;
 605
 606         /*
 607          * request is released from the driver, io must be done
 608          */
 609         if (blk_account_rq(rq)) {
 610                 q->in_flight--;
 611                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 612                         e->ops->elevator_completed_req_fn(q, rq);
 613         }
 614
 615         /*
 616          * Check if the queue is waiting for fs requests to be
 617          * drained for flush sequence.
 618          */
 619         if (unlikely(q->ordseq)) {
 620                 struct request *first_rq = list_entry_rq(q->queue_head.next);
 621                 if (q->in_flight == 0 &&
 622                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 623                     blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
 624                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 625                         q->request_fn(q);
 626                 }
 627         }
 628 }
 629
 630 int elv_register_queue(struct request_queue *q)
 631 {
 632         elevator_t *e = q->elevator;
 633
 634         e->kobj.parent = kobject_get(&q->kobj);
 635         if (!e->kobj.parent)
 636                 return -EBUSY;
 637
 638         snprintf(e->kobj.name, KOBJ_NAME_LEN, "%s", "iosched");
 639         e->kobj.ktype = e->elevator_type->elevator_ktype;
 640
 641         return kobject_register(&e->kobj);
 642 }
 643
 644 void elv_unregister_queue(struct request_queue *q)
 645 {
 646         if (q) {
 647                 elevator_t *e = q->elevator;
 648                 kobject_unregister(&e->kobj);
 649                 kobject_put(&q->kobj);
 650         }
 651 }
 652
 653 int elv_register(struct elevator_type *e)
 654 {
 655         spin_lock_irq(&elv_list_lock);
 656         if (elevator_find(e->elevator_name))
 657                 BUG();
 658         list_add_tail(&e->list, &elv_list);
 659         spin_unlock_irq(&elv_list_lock);
 660
 661         printk(KERN_INFO "io scheduler %s registered", e->elevator_name);
 662         if (!strcmp(e->elevator_name, chosen_elevator) ||
 663                         (!*chosen_elevator &&
 664                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 665                                 printk(" (default)");
 666         printk("\n");
 667         return 0;
 668 }
 669 EXPORT_SYMBOL_GPL(elv_register);
 670
 671 void elv_unregister(struct elevator_type *e)
 672 {
 673         struct task_struct *g, *p;
 674
 675         /*
 676          * Iterate every thread in the process to remove the io contexts.
 677          */
 678         read_lock(&tasklist_lock);
 679         do_each_thread(g, p) {
 680                 struct io_context *ioc = p->io_context;
 681                 if (ioc && ioc->cic) {
 682                         ioc->cic->exit(ioc->cic);
 683                         ioc->cic->dtor(ioc->cic);
 684                         ioc->cic = NULL;
 685                 }
 686                 if (ioc && ioc->aic) {
 687                         ioc->aic->exit(ioc->aic);
 688                         ioc->aic->dtor(ioc->aic);
 689                         ioc->aic = NULL;
 690                 }
 691         } while_each_thread(g, p);
 692         read_unlock(&tasklist_lock);
 693
 694         spin_lock_irq(&elv_list_lock);
 695         list_del_init(&e->list);
 696         spin_unlock_irq(&elv_list_lock);
 697 }
 698 EXPORT_SYMBOL_GPL(elv_unregister);
 699
 700 /*
 701  * switch to new_e io scheduler. be careful not to introduce deadlocks -
 702  * we don't free the old io scheduler, before we have allocated what we
 703  * need for the new one. this way we have a chance of going back to the old
 704  * one, if the new one fails init for some reason.
 705  */
 706 static void elevator_switch(request_queue_t *q, struct elevator_type *new_e)
 707 {
 708         elevator_t *old_elevator, *e;
 709
 710         /*
 711          * Allocate new elevator
 712          */
 713         e = kmalloc(sizeof(elevator_t), GFP_KERNEL);
 714         if (!e)
 715                 goto error;
 716
 717         /*
 718          * Turn on BYPASS and drain all requests w/ elevator private data
 719          */
 720         spin_lock_irq(q->queue_lock);
 721
 722         set_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
 723
 724         elv_drain_elevator(q);
 725
 726         while (q->rq.elvpriv) {
 727                 blk_remove_plug(q);
 728                 q->request_fn(q);
 729                 spin_unlock_irq(q->queue_lock);
 730                 msleep(10);
 731                 spin_lock_irq(q->queue_lock);
 732                 elv_drain_elevator(q);
 733         }
 734
 735         spin_unlock_irq(q->queue_lock);
 736
 737         /*
 738          * unregister old elevator data
 739          */
 740         elv_unregister_queue(q);
 741         old_elevator = q->elevator;
 742
 743         /*
 744          * attach and start new elevator
 745          */
 746         if (elevator_attach(q, new_e, e))
 747                 goto fail;
 748
 749         if (elv_register_queue(q))
 750                 goto fail_register;
 751
 752         /*
 753          * finally exit old elevator and turn off BYPASS.
 754          */
 755         elevator_exit(old_elevator);
 756         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
 757         return;
 758
 759 fail_register:
 760         /*
 761          * switch failed, exit the new io scheduler and reattach the old
 762          * one again (along with re-adding the sysfs dir)
 763          */
 764         elevator_exit(e);
 765         e = NULL;
 766 fail:
 767         q->elevator = old_elevator;
 768         elv_register_queue(q);
 769         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
 770         kfree(e);
 771 error:
 772         elevator_put(new_e);
 773         printk(KERN_ERR "elevator: switch to %s failed\n",new_e->elevator_name);
 774 }
 775
 776 ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count)
 777 {
 778         char elevator_name[ELV_NAME_MAX];
 779         size_t len;
 780         struct elevator_type *e;
 781
 782         elevator_name[sizeof(elevator_name) - 1] = '\0';
 783         strncpy(elevator_name, name, sizeof(elevator_name) - 1);
 784         len = strlen(elevator_name);
 785
 786         if (len && elevator_name[len - 1] == '\n')
 787                 elevator_name[len - 1] = '\0';
 788
 789         e = elevator_get(elevator_name);
 790         if (!e) {
 791                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
 792                 return -EINVAL;
 793         }
 794
 795         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
 796                 elevator_put(e);
 797                 return count;
 798         }
 799
 800         elevator_switch(q, e);
 801         return count;
 802 }
 803
 804 ssize_t elv_iosched_show(request_queue_t *q, char *name)
 805 {
 806         elevator_t *e = q->elevator;
 807         struct elevator_type *elv = e->elevator_type;
 808         struct list_head *entry;
 809         int len = 0;
 810
 811         spin_lock_irq(q->queue_lock);
 812         list_for_each(entry, &elv_list) {
 813                 struct elevator_type *__e;
 814
 815                 __e = list_entry(entry, struct elevator_type, list);
 816                 if (!strcmp(elv->elevator_name, __e->elevator_name))
 817                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
 818                 else
 819                         len += sprintf(name+len, "%s ", __e->elevator_name);
 820         }
 821         spin_unlock_irq(q->queue_lock);
 822
 823         len += sprintf(len+name, "\n");
 824         return len;
 825 }
 826
 827 EXPORT_SYMBOL(elv_dispatch_sort);
 828 EXPORT_SYMBOL(elv_add_request);
 829 EXPORT_SYMBOL(__elv_add_request);
 830 EXPORT_SYMBOL(elv_requeue_request);
 831 EXPORT_SYMBOL(elv_next_request);
 832 EXPORT_SYMBOL(elv_dequeue_request);
 833 EXPORT_SYMBOL(elv_queue_empty);
 834 EXPORT_SYMBOL(elv_completed_request);
 835 EXPORT_SYMBOL(elevator_exit);
 836 EXPORT_SYMBOL(elevator_init);