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