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         int unplug_it = 1;
 318
 319         rq->q = q;
 320
 321         switch (where) {
 322         case ELEVATOR_INSERT_FRONT:
 323                 rq->flags |= REQ_SOFTBARRIER;
 324
 325                 list_add(&rq->queuelist, &q->queue_head);
 326                 break;
 327
 328         case ELEVATOR_INSERT_BACK:
 329                 rq->flags |= REQ_SOFTBARRIER;
 330                 elv_drain_elevator(q);
 331                 list_add_tail(&rq->queuelist, &q->queue_head);
 332                 /*
 333                  * We kick the queue here for the following reasons.
 334                  * - The elevator might have returned NULL previously
 335                  *   to delay requests and returned them now.  As the
 336                  *   queue wasn't empty before this request, ll_rw_blk
 337                  *   won't run the queue on return, resulting in hang.
 338                  * - Usually, back inserted requests won't be merged
 339                  *   with anything.  There's no point in delaying queue
 340                  *   processing.
 341                  */
 342                 blk_remove_plug(q);
 343                 q->request_fn(q);
 344                 break;
 345
 346         case ELEVATOR_INSERT_SORT:
 347                 BUG_ON(!blk_fs_request(rq));
 348                 rq->flags |= REQ_SORTED;
 349                 q->nr_sorted++;
 350                 if (q->last_merge == NULL && rq_mergeable(rq))
 351                         q->last_merge = rq;
 352                 /*
 353                  * Some ioscheds (cfq) run q->request_fn directly, so
 354                  * rq cannot be accessed after calling
 355                  * elevator_add_req_fn.
 356                  */
 357                 q->elevator->ops->elevator_add_req_fn(q, rq);
 358                 break;
 359
 360         case ELEVATOR_INSERT_REQUEUE:
 361                 /*
 362                  * If ordered flush isn't in progress, we do front
 363                  * insertion; otherwise, requests should be requeued
 364                  * in ordseq order.
 365                  */
 366                 rq->flags |= REQ_SOFTBARRIER;
 367
 368                 if (q->ordseq == 0) {
 369                         list_add(&rq->queuelist, &q->queue_head);
 370                         break;
 371                 }
 372
 373                 ordseq = blk_ordered_req_seq(rq);
 374
 375                 list_for_each(pos, &q->queue_head) {
 376                         struct request *pos_rq = list_entry_rq(pos);
 377                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 378                                 break;
 379                 }
 380
 381                 list_add_tail(&rq->queuelist, pos);
 382                 /*
 383                  * most requeues happen because of a busy condition, don't
 384                  * force unplug of the queue for that case.
 385                  */
 386                 unplug_it = 0;
 387                 break;
 388
 389         default:
 390                 printk(KERN_ERR "%s: bad insertion point %d\n",
 391                        __FUNCTION__, where);
 392                 BUG();
 393         }
 394
 395         if (unplug_it && blk_queue_plugged(q)) {
 396                 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
 397                         - q->in_flight;
 398
 399                 if (nrq >= q->unplug_thresh)
 400                         __generic_unplug_device(q);
 401         }
 402 }
 403
 404 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
 405                        int plug)
 406 {
 407         if (q->ordcolor)
 408                 rq->flags |= REQ_ORDERED_COLOR;
 409
 410         if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 411                 /*
 412                  * toggle ordered color
 413                  */
 414                 if (blk_barrier_rq(rq))
 415                         q->ordcolor ^= 1;
 416
 417                 /*
 418                  * barriers implicitly indicate back insertion
 419                  */
 420                 if (where == ELEVATOR_INSERT_SORT)
 421                         where = ELEVATOR_INSERT_BACK;
 422
 423                 /*
 424                  * this request is scheduling boundary, update
 425                  * end_sector
 426                  */
 427                 if (blk_fs_request(rq)) {
 428                         q->end_sector = rq_end_sector(rq);
 429                         q->boundary_rq = rq;
 430                 }
 431         } else if (!(rq->flags & REQ_ELVPRIV) && where == ELEVATOR_INSERT_SORT)
 432                 where = ELEVATOR_INSERT_BACK;
 433
 434         if (plug)
 435                 blk_plug_device(q);
 436
 437         elv_insert(q, rq, where);
 438 }
 439
 440 void elv_add_request(request_queue_t *q, struct request *rq, int where,
 441                      int plug)
 442 {
 443         unsigned long flags;
 444
 445         spin_lock_irqsave(q->queue_lock, flags);
 446         __elv_add_request(q, rq, where, plug);
 447         spin_unlock_irqrestore(q->queue_lock, flags);
 448 }
 449
 450 static inline struct request *__elv_next_request(request_queue_t *q)
 451 {
 452         struct request *rq;
 453
 454         while (1) {
 455                 while (!list_empty(&q->queue_head)) {
 456                         rq = list_entry_rq(q->queue_head.next);
 457                         if (blk_do_ordered(q, &rq))
 458                                 return rq;
 459                 }
 460
 461                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
 462                         return NULL;
 463         }
 464 }
 465
 466 struct request *elv_next_request(request_queue_t *q)
 467 {
 468         struct request *rq;
 469         int ret;
 470
 471         while ((rq = __elv_next_request(q)) != NULL) {
 472                 if (!(rq->flags & REQ_STARTED)) {
 473                         elevator_t *e = q->elevator;
 474
 475                         /*
 476                          * This is the first time the device driver
 477                          * sees this request (possibly after
 478                          * requeueing).  Notify IO scheduler.
 479                          */
 480                         if (blk_sorted_rq(rq) &&
 481                             e->ops->elevator_activate_req_fn)
 482                                 e->ops->elevator_activate_req_fn(q, rq);
 483
 484                         /*
 485                          * just mark as started even if we don't start
 486                          * it, a request that has been delayed should
 487                          * not be passed by new incoming requests
 488                          */
 489                         rq->flags |= REQ_STARTED;
 490                 }
 491
 492                 if (!q->boundary_rq || q->boundary_rq == rq) {
 493                         q->end_sector = rq_end_sector(rq);
 494                         q->boundary_rq = NULL;
 495                 }
 496
 497                 if ((rq->flags & REQ_DONTPREP) || !q->prep_rq_fn)
 498                         break;
 499
 500                 ret = q->prep_rq_fn(q, rq);
 501                 if (ret == BLKPREP_OK) {
 502                         break;
 503                 } else if (ret == BLKPREP_DEFER) {
 504                         /*
 505                          * the request may have been (partially) prepped.
 506                          * we need to keep this request in the front to
 507                          * avoid resource deadlock.  REQ_STARTED will
 508                          * prevent other fs requests from passing this one.
 509                          */
 510                         rq = NULL;
 511                         break;
 512                 } else if (ret == BLKPREP_KILL) {
 513                         int nr_bytes = rq->hard_nr_sectors << 9;
 514
 515                         if (!nr_bytes)
 516                                 nr_bytes = rq->data_len;
 517
 518                         blkdev_dequeue_request(rq);
 519                         rq->flags |= REQ_QUIET;
 520                         end_that_request_chunk(rq, 0, nr_bytes);
 521                         end_that_request_last(rq, 0);
 522                 } else {
 523                         printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
 524                                                                 ret);
 525                         break;
 526                 }
 527         }
 528
 529         return rq;
 530 }
 531
 532 void elv_dequeue_request(request_queue_t *q, struct request *rq)
 533 {
 534         BUG_ON(list_empty(&rq->queuelist));
 535
 536         list_del_init(&rq->queuelist);
 537
 538         /*
 539          * the time frame between a request being removed from the lists
 540          * and to it is freed is accounted as io that is in progress at
 541          * the driver side.
 542          */
 543         if (blk_account_rq(rq))
 544                 q->in_flight++;
 545 }
 546
 547 int elv_queue_empty(request_queue_t *q)
 548 {
 549         elevator_t *e = q->elevator;
 550
 551         if (!list_empty(&q->queue_head))
 552                 return 0;
 553
 554         if (e->ops->elevator_queue_empty_fn)
 555                 return e->ops->elevator_queue_empty_fn(q);
 556
 557         return 1;
 558 }
 559
 560 struct request *elv_latter_request(request_queue_t *q, struct request *rq)
 561 {
 562         elevator_t *e = q->elevator;
 563
 564         if (e->ops->elevator_latter_req_fn)
 565                 return e->ops->elevator_latter_req_fn(q, rq);
 566         return NULL;
 567 }
 568
 569 struct request *elv_former_request(request_queue_t *q, struct request *rq)
 570 {
 571         elevator_t *e = q->elevator;
 572
 573         if (e->ops->elevator_former_req_fn)
 574                 return e->ops->elevator_former_req_fn(q, rq);
 575         return NULL;
 576 }
 577
 578 int elv_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
 579                     gfp_t gfp_mask)
 580 {
 581         elevator_t *e = q->elevator;
 582
 583         if (e->ops->elevator_set_req_fn)
 584                 return e->ops->elevator_set_req_fn(q, rq, bio, gfp_mask);
 585
 586         rq->elevator_private = NULL;
 587         return 0;
 588 }
 589
 590 void elv_put_request(request_queue_t *q, struct request *rq)
 591 {
 592         elevator_t *e = q->elevator;
 593
 594         if (e->ops->elevator_put_req_fn)
 595                 e->ops->elevator_put_req_fn(q, rq);
 596 }
 597
 598 int elv_may_queue(request_queue_t *q, int rw, struct bio *bio)
 599 {
 600         elevator_t *e = q->elevator;
 601
 602         if (e->ops->elevator_may_queue_fn)
 603                 return e->ops->elevator_may_queue_fn(q, rw, bio);
 604
 605         return ELV_MQUEUE_MAY;
 606 }
 607
 608 void elv_completed_request(request_queue_t *q, struct request *rq)
 609 {
 610         elevator_t *e = q->elevator;
 611
 612         /*
 613          * request is released from the driver, io must be done
 614          */
 615         if (blk_account_rq(rq)) {
 616                 q->in_flight--;
 617                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 618                         e->ops->elevator_completed_req_fn(q, rq);
 619         }
 620
 621         /*
 622          * Check if the queue is waiting for fs requests to be
 623          * drained for flush sequence.
 624          */
 625         if (unlikely(q->ordseq)) {
 626                 struct request *first_rq = list_entry_rq(q->queue_head.next);
 627                 if (q->in_flight == 0 &&
 628                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 629                     blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
 630                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 631                         q->request_fn(q);
 632                 }
 633         }
 634 }
 635
 636 int elv_register_queue(struct request_queue *q)
 637 {
 638         elevator_t *e = q->elevator;
 639
 640         e->kobj.parent = kobject_get(&q->kobj);
 641         if (!e->kobj.parent)
 642                 return -EBUSY;
 643
 644         snprintf(e->kobj.name, KOBJ_NAME_LEN, "%s", "iosched");
 645         e->kobj.ktype = e->elevator_type->elevator_ktype;
 646
 647         return kobject_register(&e->kobj);
 648 }
 649
 650 void elv_unregister_queue(struct request_queue *q)
 651 {
 652         if (q) {
 653                 elevator_t *e = q->elevator;
 654                 kobject_unregister(&e->kobj);
 655                 kobject_put(&q->kobj);
 656         }
 657 }
 658
 659 int elv_register(struct elevator_type *e)
 660 {
 661         spin_lock_irq(&elv_list_lock);
 662         if (elevator_find(e->elevator_name))
 663                 BUG();
 664         list_add_tail(&e->list, &elv_list);
 665         spin_unlock_irq(&elv_list_lock);
 666
 667         printk(KERN_INFO "io scheduler %s registered", e->elevator_name);
 668         if (!strcmp(e->elevator_name, chosen_elevator) ||
 669                         (!*chosen_elevator &&
 670                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 671                                 printk(" (default)");
 672         printk("\n");
 673         return 0;
 674 }
 675 EXPORT_SYMBOL_GPL(elv_register);
 676
 677 void elv_unregister(struct elevator_type *e)
 678 {
 679         struct task_struct *g, *p;
 680
 681         /*
 682          * Iterate every thread in the process to remove the io contexts.
 683          */
 684         read_lock(&tasklist_lock);
 685         do_each_thread(g, p) {
 686                 struct io_context *ioc = p->io_context;
 687                 if (ioc && ioc->cic) {
 688                         ioc->cic->exit(ioc->cic);
 689                         ioc->cic->dtor(ioc->cic);
 690                         ioc->cic = NULL;
 691                 }
 692                 if (ioc && ioc->aic) {
 693                         ioc->aic->exit(ioc->aic);
 694                         ioc->aic->dtor(ioc->aic);
 695                         ioc->aic = NULL;
 696                 }
 697         } while_each_thread(g, p);
 698         read_unlock(&tasklist_lock);
 699
 700         spin_lock_irq(&elv_list_lock);
 701         list_del_init(&e->list);
 702         spin_unlock_irq(&elv_list_lock);
 703 }
 704 EXPORT_SYMBOL_GPL(elv_unregister);
 705
 706 /*
 707  * switch to new_e io scheduler. be careful not to introduce deadlocks -
 708  * we don't free the old io scheduler, before we have allocated what we
 709  * need for the new one. this way we have a chance of going back to the old
 710  * one, if the new one fails init for some reason.
 711  */
 712 static void elevator_switch(request_queue_t *q, struct elevator_type *new_e)
 713 {
 714         elevator_t *old_elevator, *e;
 715
 716         /*
 717          * Allocate new elevator
 718          */
 719         e = kmalloc(sizeof(elevator_t), GFP_KERNEL);
 720         if (!e)
 721                 goto error;
 722
 723         /*
 724          * Turn on BYPASS and drain all requests w/ elevator private data
 725          */
 726         spin_lock_irq(q->queue_lock);
 727
 728         set_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
 729
 730         elv_drain_elevator(q);
 731
 732         while (q->rq.elvpriv) {
 733                 blk_remove_plug(q);
 734                 q->request_fn(q);
 735                 spin_unlock_irq(q->queue_lock);
 736                 msleep(10);
 737                 spin_lock_irq(q->queue_lock);
 738                 elv_drain_elevator(q);
 739         }
 740
 741         spin_unlock_irq(q->queue_lock);
 742
 743         /*
 744          * unregister old elevator data
 745          */
 746         elv_unregister_queue(q);
 747         old_elevator = q->elevator;
 748
 749         /*
 750          * attach and start new elevator
 751          */
 752         if (elevator_attach(q, new_e, e))
 753                 goto fail;
 754
 755         if (elv_register_queue(q))
 756                 goto fail_register;
 757
 758         /*
 759          * finally exit old elevator and turn off BYPASS.
 760          */
 761         elevator_exit(old_elevator);
 762         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
 763         return;
 764
 765 fail_register:
 766         /*
 767          * switch failed, exit the new io scheduler and reattach the old
 768          * one again (along with re-adding the sysfs dir)
 769          */
 770         elevator_exit(e);
 771         e = NULL;
 772 fail:
 773         q->elevator = old_elevator;
 774         elv_register_queue(q);
 775         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
 776         kfree(e);
 777 error:
 778         elevator_put(new_e);
 779         printk(KERN_ERR "elevator: switch to %s failed\n",new_e->elevator_name);
 780 }
 781
 782 ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count)
 783 {
 784         char elevator_name[ELV_NAME_MAX];
 785         size_t len;
 786         struct elevator_type *e;
 787
 788         elevator_name[sizeof(elevator_name) - 1] = '\0';
 789         strncpy(elevator_name, name, sizeof(elevator_name) - 1);
 790         len = strlen(elevator_name);
 791
 792         if (len && elevator_name[len - 1] == '\n')
 793                 elevator_name[len - 1] = '\0';
 794
 795         e = elevator_get(elevator_name);
 796         if (!e) {
 797                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
 798                 return -EINVAL;
 799         }
 800
 801         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
 802                 elevator_put(e);
 803                 return count;
 804         }
 805
 806         elevator_switch(q, e);
 807         return count;
 808 }
 809
 810 ssize_t elv_iosched_show(request_queue_t *q, char *name)
 811 {
 812         elevator_t *e = q->elevator;
 813         struct elevator_type *elv = e->elevator_type;
 814         struct list_head *entry;
 815         int len = 0;
 816
 817         spin_lock_irq(&elv_list_lock);
 818         list_for_each(entry, &elv_list) {
 819                 struct elevator_type *__e;
 820
 821                 __e = list_entry(entry, struct elevator_type, list);
 822                 if (!strcmp(elv->elevator_name, __e->elevator_name))
 823                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
 824                 else
 825                         len += sprintf(name+len, "%s ", __e->elevator_name);
 826         }
 827         spin_unlock_irq(&elv_list_lock);
 828
 829         len += sprintf(len+name, "\n");
 830         return len;
 831 }
 832
 833 EXPORT_SYMBOL(elv_dispatch_sort);
 834 EXPORT_SYMBOL(elv_add_request);
 835 EXPORT_SYMBOL(__elv_add_request);
 836 EXPORT_SYMBOL(elv_requeue_request);
 837 EXPORT_SYMBOL(elv_next_request);
 838 EXPORT_SYMBOL(elv_dequeue_request);
 839 EXPORT_SYMBOL(elv_queue_empty);
 840 EXPORT_SYMBOL(elv_completed_request);
 841 EXPORT_SYMBOL(elevator_exit);
 842 EXPORT_SYMBOL(elevator_init);