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 #include <linux/blktrace_api.h>
  37
  38 #include <asm/uaccess.h>
  39
  40 static DEFINE_SPINLOCK(elv_list_lock);
  41 static LIST_HEAD(elv_list);
  42
  43 /*
  44  * can we safely merge with this request?
  45  */
  46 inline int elv_rq_merge_ok(struct request *rq, struct bio *bio)
  47 {
  48         if (!rq_mergeable(rq))
  49                 return 0;
  50
  51         /*
  52          * different data direction or already started, don't merge
  53          */
  54         if (bio_data_dir(bio) != rq_data_dir(rq))
  55                 return 0;
  56
  57         /*
  58          * same device and no special stuff set, merge is ok
  59          */
  60         if (rq->rq_disk == bio->bi_bdev->bd_disk &&
  61             !rq->waiting && !rq->special)
  62                 return 1;
  63
  64         return 0;
  65 }
  66 EXPORT_SYMBOL(elv_rq_merge_ok);
  67
  68 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
  69 {
  70         int ret = ELEVATOR_NO_MERGE;
  71
  72         /*
  73          * we can merge and sequence is ok, check if it's possible
  74          */
  75         if (elv_rq_merge_ok(__rq, bio)) {
  76                 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
  77                         ret = ELEVATOR_BACK_MERGE;
  78                 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
  79                         ret = ELEVATOR_FRONT_MERGE;
  80         }
  81
  82         return ret;
  83 }
  84
  85 static struct elevator_type *elevator_find(const char *name)
  86 {
  87         struct elevator_type *e = NULL;
  88         struct list_head *entry;
  89
  90         list_for_each(entry, &elv_list) {
  91                 struct elevator_type *__e;
  92
  93                 __e = list_entry(entry, struct elevator_type, list);
  94
  95                 if (!strcmp(__e->elevator_name, name)) {
  96                         e = __e;
  97                         break;
  98                 }
  99         }
 100
 101         return e;
 102 }
 103
 104 static void elevator_put(struct elevator_type *e)
 105 {
 106         module_put(e->elevator_owner);
 107 }
 108
 109 static struct elevator_type *elevator_get(const char *name)
 110 {
 111         struct elevator_type *e;
 112
 113         spin_lock_irq(&elv_list_lock);
 114
 115         e = elevator_find(name);
 116         if (e && !try_module_get(e->elevator_owner))
 117                 e = NULL;
 118
 119         spin_unlock_irq(&elv_list_lock);
 120
 121         return e;
 122 }
 123
 124 static void *elevator_init_queue(request_queue_t *q, struct elevator_queue *eq)
 125 {
 126         return eq->ops->elevator_init_fn(q, eq);
 127 }
 128
 129 static void elevator_attach(request_queue_t *q, struct elevator_queue *eq,
 130                            void *data)
 131 {
 132         q->elevator = eq;
 133         eq->elevator_data = data;
 134 }
 135
 136 static char chosen_elevator[16];
 137
 138 static int __init elevator_setup(char *str)
 139 {
 140         /*
 141          * Be backwards-compatible with previous kernels, so users
 142          * won't get the wrong elevator.
 143          */
 144         if (!strcmp(str, "as"))
 145                 strcpy(chosen_elevator, "anticipatory");
 146         else
 147                 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 148         return 1;
 149 }
 150
 151 __setup("elevator=", elevator_setup);
 152
 153 static struct kobj_type elv_ktype;
 154
 155 static elevator_t *elevator_alloc(struct elevator_type *e)
 156 {
 157         elevator_t *eq = kmalloc(sizeof(elevator_t), GFP_KERNEL);
 158         if (eq) {
 159                 memset(eq, 0, sizeof(*eq));
 160                 eq->ops = &e->ops;
 161                 eq->elevator_type = e;
 162                 kobject_init(&eq->kobj);
 163                 snprintf(eq->kobj.name, KOBJ_NAME_LEN, "%s", "iosched");
 164                 eq->kobj.ktype = &elv_ktype;
 165                 mutex_init(&eq->sysfs_lock);
 166         } else {
 167                 elevator_put(e);
 168         }
 169         return eq;
 170 }
 171
 172 static void elevator_release(struct kobject *kobj)
 173 {
 174         elevator_t *e = container_of(kobj, elevator_t, kobj);
 175         elevator_put(e->elevator_type);
 176         kfree(e);
 177 }
 178
 179 int elevator_init(request_queue_t *q, char *name)
 180 {
 181         struct elevator_type *e = NULL;
 182         struct elevator_queue *eq;
 183         int ret = 0;
 184         void *data;
 185
 186         INIT_LIST_HEAD(&q->queue_head);
 187         q->last_merge = NULL;
 188         q->end_sector = 0;
 189         q->boundary_rq = NULL;
 190
 191         if (name && !(e = elevator_get(name)))
 192                 return -EINVAL;
 193
 194         if (!e && *chosen_elevator && !(e = elevator_get(chosen_elevator)))
 195                 printk("I/O scheduler %s not found\n", chosen_elevator);
 196
 197         if (!e && !(e = elevator_get(CONFIG_DEFAULT_IOSCHED))) {
 198                 printk("Default I/O scheduler not found, using no-op\n");
 199                 e = elevator_get("noop");
 200         }
 201
 202         eq = elevator_alloc(e);
 203         if (!eq)
 204                 return -ENOMEM;
 205
 206         data = elevator_init_queue(q, eq);
 207         if (!data) {
 208                 kobject_put(&eq->kobj);
 209                 return -ENOMEM;
 210         }
 211
 212         elevator_attach(q, eq, data);
 213         return ret;
 214 }
 215
 216 void elevator_exit(elevator_t *e)
 217 {
 218         mutex_lock(&e->sysfs_lock);
 219         if (e->ops->elevator_exit_fn)
 220                 e->ops->elevator_exit_fn(e);
 221         e->ops = NULL;
 222         mutex_unlock(&e->sysfs_lock);
 223
 224         kobject_put(&e->kobj);
 225 }
 226
 227 /*
 228  * Insert rq into dispatch queue of q.  Queue lock must be held on
 229  * entry.  If sort != 0, rq is sort-inserted; otherwise, rq will be
 230  * appended to the dispatch queue.  To be used by specific elevators.
 231  */
 232 void elv_dispatch_sort(request_queue_t *q, struct request *rq)
 233 {
 234         sector_t boundary;
 235         struct list_head *entry;
 236
 237         if (q->last_merge == rq)
 238                 q->last_merge = NULL;
 239         q->nr_sorted--;
 240
 241         boundary = q->end_sector;
 242
 243         list_for_each_prev(entry, &q->queue_head) {
 244                 struct request *pos = list_entry_rq(entry);
 245
 246                 if (pos->flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
 247                         break;
 248                 if (rq->sector >= boundary) {
 249                         if (pos->sector < boundary)
 250                                 continue;
 251                 } else {
 252                         if (pos->sector >= boundary)
 253                                 break;
 254                 }
 255                 if (rq->sector >= pos->sector)
 256                         break;
 257         }
 258
 259         list_add(&rq->queuelist, entry);
 260 }
 261
 262 int elv_merge(request_queue_t *q, struct request **req, struct bio *bio)
 263 {
 264         elevator_t *e = q->elevator;
 265         int ret;
 266
 267         if (q->last_merge) {
 268                 ret = elv_try_merge(q->last_merge, bio);
 269                 if (ret != ELEVATOR_NO_MERGE) {
 270                         *req = q->last_merge;
 271                         return ret;
 272                 }
 273         }
 274
 275         if (e->ops->elevator_merge_fn)
 276                 return e->ops->elevator_merge_fn(q, req, bio);
 277
 278         return ELEVATOR_NO_MERGE;
 279 }
 280
 281 void elv_merged_request(request_queue_t *q, struct request *rq)
 282 {
 283         elevator_t *e = q->elevator;
 284
 285         if (e->ops->elevator_merged_fn)
 286                 e->ops->elevator_merged_fn(q, rq);
 287
 288         q->last_merge = rq;
 289 }
 290
 291 void elv_merge_requests(request_queue_t *q, struct request *rq,
 292                              struct request *next)
 293 {
 294         elevator_t *e = q->elevator;
 295
 296         if (e->ops->elevator_merge_req_fn)
 297                 e->ops->elevator_merge_req_fn(q, rq, next);
 298         q->nr_sorted--;
 299
 300         q->last_merge = rq;
 301 }
 302
 303 void elv_requeue_request(request_queue_t *q, struct request *rq)
 304 {
 305         elevator_t *e = q->elevator;
 306
 307         /*
 308          * it already went through dequeue, we need to decrement the
 309          * in_flight count again
 310          */
 311         if (blk_account_rq(rq)) {
 312                 q->in_flight--;
 313                 if (blk_sorted_rq(rq) && e->ops->elevator_deactivate_req_fn)
 314                         e->ops->elevator_deactivate_req_fn(q, rq);
 315         }
 316
 317         rq->flags &= ~REQ_STARTED;
 318
 319         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 320 }
 321
 322 static void elv_drain_elevator(request_queue_t *q)
 323 {
 324         static int printed;
 325         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 326                 ;
 327         if (q->nr_sorted == 0)
 328                 return;
 329         if (printed++ < 10) {
 330                 printk(KERN_ERR "%s: forced dispatching is broken "
 331                        "(nr_sorted=%u), please report this\n",
 332                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 333         }
 334 }
 335
 336 void elv_insert(request_queue_t *q, struct request *rq, int where)
 337 {
 338         struct list_head *pos;
 339         unsigned ordseq;
 340         int unplug_it = 1;
 341
 342         blk_add_trace_rq(q, rq, BLK_TA_INSERT);
 343
 344         rq->q = q;
 345
 346         switch (where) {
 347         case ELEVATOR_INSERT_FRONT:
 348                 rq->flags |= REQ_SOFTBARRIER;
 349
 350                 list_add(&rq->queuelist, &q->queue_head);
 351                 break;
 352
 353         case ELEVATOR_INSERT_BACK:
 354                 rq->flags |= REQ_SOFTBARRIER;
 355                 elv_drain_elevator(q);
 356                 list_add_tail(&rq->queuelist, &q->queue_head);
 357                 /*
 358                  * We kick the queue here for the following reasons.
 359                  * - The elevator might have returned NULL previously
 360                  *   to delay requests and returned them now.  As the
 361                  *   queue wasn't empty before this request, ll_rw_blk
 362                  *   won't run the queue on return, resulting in hang.
 363                  * - Usually, back inserted requests won't be merged
 364                  *   with anything.  There's no point in delaying queue
 365                  *   processing.
 366                  */
 367                 blk_remove_plug(q);
 368                 q->request_fn(q);
 369                 break;
 370
 371         case ELEVATOR_INSERT_SORT:
 372                 BUG_ON(!blk_fs_request(rq));
 373                 rq->flags |= REQ_SORTED;
 374                 q->nr_sorted++;
 375                 if (q->last_merge == NULL && rq_mergeable(rq))
 376                         q->last_merge = rq;
 377                 /*
 378                  * Some ioscheds (cfq) run q->request_fn directly, so
 379                  * rq cannot be accessed after calling
 380                  * elevator_add_req_fn.
 381                  */
 382                 q->elevator->ops->elevator_add_req_fn(q, rq);
 383                 break;
 384
 385         case ELEVATOR_INSERT_REQUEUE:
 386                 /*
 387                  * If ordered flush isn't in progress, we do front
 388                  * insertion; otherwise, requests should be requeued
 389                  * in ordseq order.
 390                  */
 391                 rq->flags |= REQ_SOFTBARRIER;
 392
 393                 if (q->ordseq == 0) {
 394                         list_add(&rq->queuelist, &q->queue_head);
 395                         break;
 396                 }
 397
 398                 ordseq = blk_ordered_req_seq(rq);
 399
 400                 list_for_each(pos, &q->queue_head) {
 401                         struct request *pos_rq = list_entry_rq(pos);
 402                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 403                                 break;
 404                 }
 405
 406                 list_add_tail(&rq->queuelist, pos);
 407                 /*
 408                  * most requeues happen because of a busy condition, don't
 409                  * force unplug of the queue for that case.
 410                  */
 411                 unplug_it = 0;
 412                 break;
 413
 414         default:
 415                 printk(KERN_ERR "%s: bad insertion point %d\n",
 416                        __FUNCTION__, where);
 417                 BUG();
 418         }
 419
 420         if (unplug_it && blk_queue_plugged(q)) {
 421                 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
 422                         - q->in_flight;
 423
 424                 if (nrq >= q->unplug_thresh)
 425                         __generic_unplug_device(q);
 426         }
 427 }
 428
 429 void __elv_add_request(request_queue_t *q, struct request *rq, int where,
 430                        int plug)
 431 {
 432         if (q->ordcolor)
 433                 rq->flags |= REQ_ORDERED_COLOR;
 434
 435         if (rq->flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 436                 /*
 437                  * toggle ordered color
 438                  */
 439                 if (blk_barrier_rq(rq))
 440                         q->ordcolor ^= 1;
 441
 442                 /*
 443                  * barriers implicitly indicate back insertion
 444                  */
 445                 if (where == ELEVATOR_INSERT_SORT)
 446                         where = ELEVATOR_INSERT_BACK;
 447
 448                 /*
 449                  * this request is scheduling boundary, update
 450                  * end_sector
 451                  */
 452                 if (blk_fs_request(rq)) {
 453                         q->end_sector = rq_end_sector(rq);
 454                         q->boundary_rq = rq;
 455                 }
 456         } else if (!(rq->flags & REQ_ELVPRIV) && where == ELEVATOR_INSERT_SORT)
 457                 where = ELEVATOR_INSERT_BACK;
 458
 459         if (plug)
 460                 blk_plug_device(q);
 461
 462         elv_insert(q, rq, where);
 463 }
 464
 465 void elv_add_request(request_queue_t *q, struct request *rq, int where,
 466                      int plug)
 467 {
 468         unsigned long flags;
 469
 470         spin_lock_irqsave(q->queue_lock, flags);
 471         __elv_add_request(q, rq, where, plug);
 472         spin_unlock_irqrestore(q->queue_lock, flags);
 473 }
 474
 475 static inline struct request *__elv_next_request(request_queue_t *q)
 476 {
 477         struct request *rq;
 478
 479         while (1) {
 480                 while (!list_empty(&q->queue_head)) {
 481                         rq = list_entry_rq(q->queue_head.next);
 482                         if (blk_do_ordered(q, &rq))
 483                                 return rq;
 484                 }
 485
 486                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
 487                         return NULL;
 488         }
 489 }
 490
 491 struct request *elv_next_request(request_queue_t *q)
 492 {
 493         struct request *rq;
 494         int ret;
 495
 496         while ((rq = __elv_next_request(q)) != NULL) {
 497                 if (!(rq->flags & REQ_STARTED)) {
 498                         elevator_t *e = q->elevator;
 499
 500                         /*
 501                          * This is the first time the device driver
 502                          * sees this request (possibly after
 503                          * requeueing).  Notify IO scheduler.
 504                          */
 505                         if (blk_sorted_rq(rq) &&
 506                             e->ops->elevator_activate_req_fn)
 507                                 e->ops->elevator_activate_req_fn(q, rq);
 508
 509                         /*
 510                          * just mark as started even if we don't start
 511                          * it, a request that has been delayed should
 512                          * not be passed by new incoming requests
 513                          */
 514                         rq->flags |= REQ_STARTED;
 515                         blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
 516                 }
 517
 518                 if (!q->boundary_rq || q->boundary_rq == rq) {
 519                         q->end_sector = rq_end_sector(rq);
 520                         q->boundary_rq = NULL;
 521                 }
 522
 523                 if ((rq->flags & REQ_DONTPREP) || !q->prep_rq_fn)
 524                         break;
 525
 526                 ret = q->prep_rq_fn(q, rq);
 527                 if (ret == BLKPREP_OK) {
 528                         break;
 529                 } else if (ret == BLKPREP_DEFER) {
 530                         /*
 531                          * the request may have been (partially) prepped.
 532                          * we need to keep this request in the front to
 533                          * avoid resource deadlock.  REQ_STARTED will
 534                          * prevent other fs requests from passing this one.
 535                          */
 536                         rq = NULL;
 537                         break;
 538                 } else if (ret == BLKPREP_KILL) {
 539                         int nr_bytes = rq->hard_nr_sectors << 9;
 540
 541                         if (!nr_bytes)
 542                                 nr_bytes = rq->data_len;
 543
 544                         blkdev_dequeue_request(rq);
 545                         rq->flags |= REQ_QUIET;
 546                         end_that_request_chunk(rq, 0, nr_bytes);
 547                         end_that_request_last(rq, 0);
 548                 } else {
 549                         printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
 550                                                                 ret);
 551                         break;
 552                 }
 553         }
 554
 555         return rq;
 556 }
 557
 558 void elv_dequeue_request(request_queue_t *q, struct request *rq)
 559 {
 560         BUG_ON(list_empty(&rq->queuelist));
 561
 562         list_del_init(&rq->queuelist);
 563
 564         /*
 565          * the time frame between a request being removed from the lists
 566          * and to it is freed is accounted as io that is in progress at
 567          * the driver side.
 568          */
 569         if (blk_account_rq(rq))
 570                 q->in_flight++;
 571 }
 572
 573 int elv_queue_empty(request_queue_t *q)
 574 {
 575         elevator_t *e = q->elevator;
 576
 577         if (!list_empty(&q->queue_head))
 578                 return 0;
 579
 580         if (e->ops->elevator_queue_empty_fn)
 581                 return e->ops->elevator_queue_empty_fn(q);
 582
 583         return 1;
 584 }
 585
 586 struct request *elv_latter_request(request_queue_t *q, struct request *rq)
 587 {
 588         elevator_t *e = q->elevator;
 589
 590         if (e->ops->elevator_latter_req_fn)
 591                 return e->ops->elevator_latter_req_fn(q, rq);
 592         return NULL;
 593 }
 594
 595 struct request *elv_former_request(request_queue_t *q, struct request *rq)
 596 {
 597         elevator_t *e = q->elevator;
 598
 599         if (e->ops->elevator_former_req_fn)
 600                 return e->ops->elevator_former_req_fn(q, rq);
 601         return NULL;
 602 }
 603
 604 int elv_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
 605                     gfp_t gfp_mask)
 606 {
 607         elevator_t *e = q->elevator;
 608
 609         if (e->ops->elevator_set_req_fn)
 610                 return e->ops->elevator_set_req_fn(q, rq, bio, gfp_mask);
 611
 612         rq->elevator_private = NULL;
 613         return 0;
 614 }
 615
 616 void elv_put_request(request_queue_t *q, struct request *rq)
 617 {
 618         elevator_t *e = q->elevator;
 619
 620         if (e->ops->elevator_put_req_fn)
 621                 e->ops->elevator_put_req_fn(q, rq);
 622 }
 623
 624 int elv_may_queue(request_queue_t *q, int rw, struct bio *bio)
 625 {
 626         elevator_t *e = q->elevator;
 627
 628         if (e->ops->elevator_may_queue_fn)
 629                 return e->ops->elevator_may_queue_fn(q, rw, bio);
 630
 631         return ELV_MQUEUE_MAY;
 632 }
 633
 634 void elv_completed_request(request_queue_t *q, struct request *rq)
 635 {
 636         elevator_t *e = q->elevator;
 637
 638         /*
 639          * request is released from the driver, io must be done
 640          */
 641         if (blk_account_rq(rq)) {
 642                 q->in_flight--;
 643                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 644                         e->ops->elevator_completed_req_fn(q, rq);
 645         }
 646
 647         /*
 648          * Check if the queue is waiting for fs requests to be
 649          * drained for flush sequence.
 650          */
 651         if (unlikely(q->ordseq)) {
 652                 struct request *first_rq = list_entry_rq(q->queue_head.next);
 653                 if (q->in_flight == 0 &&
 654                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 655                     blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
 656                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 657                         q->request_fn(q);
 658                 }
 659         }
 660 }
 661
 662 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 663
 664 static ssize_t
 665 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 666 {
 667         elevator_t *e = container_of(kobj, elevator_t, kobj);
 668         struct elv_fs_entry *entry = to_elv(attr);
 669         ssize_t error;
 670
 671         if (!entry->show)
 672                 return -EIO;
 673
 674         mutex_lock(&e->sysfs_lock);
 675         error = e->ops ? entry->show(e, page) : -ENOENT;
 676         mutex_unlock(&e->sysfs_lock);
 677         return error;
 678 }
 679
 680 static ssize_t
 681 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 682                const char *page, size_t length)
 683 {
 684         elevator_t *e = container_of(kobj, elevator_t, kobj);
 685         struct elv_fs_entry *entry = to_elv(attr);
 686         ssize_t error;
 687
 688         if (!entry->store)
 689                 return -EIO;
 690
 691         mutex_lock(&e->sysfs_lock);
 692         error = e->ops ? entry->store(e, page, length) : -ENOENT;
 693         mutex_unlock(&e->sysfs_lock);
 694         return error;
 695 }
 696
 697 static struct sysfs_ops elv_sysfs_ops = {
 698         .show   = elv_attr_show,
 699         .store  = elv_attr_store,
 700 };
 701
 702 static struct kobj_type elv_ktype = {
 703         .sysfs_ops      = &elv_sysfs_ops,
 704         .release        = elevator_release,
 705 };
 706
 707 int elv_register_queue(struct request_queue *q)
 708 {
 709         elevator_t *e = q->elevator;
 710         int error;
 711
 712         e->kobj.parent = &q->kobj;
 713
 714         error = kobject_add(&e->kobj);
 715         if (!error) {
 716                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
 717                 if (attr) {
 718                         while (attr->attr.name) {
 719                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 720                                         break;
 721                                 attr++;
 722                         }
 723                 }
 724                 kobject_uevent(&e->kobj, KOBJ_ADD);
 725         }
 726         return error;
 727 }
 728
 729 static void __elv_unregister_queue(elevator_t *e)
 730 {
 731         kobject_uevent(&e->kobj, KOBJ_REMOVE);
 732         kobject_del(&e->kobj);
 733 }
 734
 735 void elv_unregister_queue(struct request_queue *q)
 736 {
 737         if (q)
 738                 __elv_unregister_queue(q->elevator);
 739 }
 740
 741 int elv_register(struct elevator_type *e)
 742 {
 743         spin_lock_irq(&elv_list_lock);
 744         BUG_ON(elevator_find(e->elevator_name));
 745         list_add_tail(&e->list, &elv_list);
 746         spin_unlock_irq(&elv_list_lock);
 747
 748         printk(KERN_INFO "io scheduler %s registered", e->elevator_name);
 749         if (!strcmp(e->elevator_name, chosen_elevator) ||
 750                         (!*chosen_elevator &&
 751                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 752                                 printk(" (default)");
 753         printk("\n");
 754         return 0;
 755 }
 756 EXPORT_SYMBOL_GPL(elv_register);
 757
 758 void elv_unregister(struct elevator_type *e)
 759 {
 760         struct task_struct *g, *p;
 761
 762         /*
 763          * Iterate every thread in the process to remove the io contexts.
 764          */
 765         if (e->ops.trim) {
 766                 read_lock(&tasklist_lock);
 767                 do_each_thread(g, p) {
 768                         task_lock(p);
 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         e = NULL;
 854         q->elevator = old_elevator;
 855         elv_register_queue(q);
 856         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
 857         if (e)
 858                 kobject_put(&e->kobj);
 859         return 0;
 860 }
 861
 862 ssize_t elv_iosched_store(request_queue_t *q, const char *name, size_t count)
 863 {
 864         char elevator_name[ELV_NAME_MAX];
 865         size_t len;
 866         struct elevator_type *e;
 867
 868         elevator_name[sizeof(elevator_name) - 1] = '\0';
 869         strncpy(elevator_name, name, sizeof(elevator_name) - 1);
 870         len = strlen(elevator_name);
 871
 872         if (len && elevator_name[len - 1] == '\n')
 873                 elevator_name[len - 1] = '\0';
 874
 875         e = elevator_get(elevator_name);
 876         if (!e) {
 877                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
 878                 return -EINVAL;
 879         }
 880
 881         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
 882                 elevator_put(e);
 883                 return count;
 884         }
 885
 886         if (!elevator_switch(q, e))
 887                 printk(KERN_ERR "elevator: switch to %s failed\n",elevator_name);
 888         return count;
 889 }
 890
 891 ssize_t elv_iosched_show(request_queue_t *q, char *name)
 892 {
 893         elevator_t *e = q->elevator;
 894         struct elevator_type *elv = e->elevator_type;
 895         struct list_head *entry;
 896         int len = 0;
 897
 898         spin_lock_irq(q->queue_lock);
 899         list_for_each(entry, &elv_list) {
 900                 struct elevator_type *__e;
 901
 902                 __e = list_entry(entry, struct elevator_type, list);
 903                 if (!strcmp(elv->elevator_name, __e->elevator_name))
 904                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
 905                 else
 906                         len += sprintf(name+len, "%s ", __e->elevator_name);
 907         }
 908         spin_unlock_irq(q->queue_lock);
 909
 910         len += sprintf(len+name, "\n");
 911         return len;
 912 }
 913
 914 EXPORT_SYMBOL(elv_dispatch_sort);
 915 EXPORT_SYMBOL(elv_add_request);
 916 EXPORT_SYMBOL(__elv_add_request);
 917 EXPORT_SYMBOL(elv_next_request);
 918 EXPORT_SYMBOL(elv_dequeue_request);
 919 EXPORT_SYMBOL(elv_queue_empty);
 920 EXPORT_SYMBOL(elevator_exit);
 921 EXPORT_SYMBOL(elevator_init);