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