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@kernel.dk> :
   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 #include <linux/hash.h>
  37
  38 #include <asm/uaccess.h>
  39
  40 static DEFINE_SPINLOCK(elv_list_lock);
  41 static LIST_HEAD(elv_list);
  42
  43 /*
  44  * Merge hash stuff.
  45  */
  46 static const int elv_hash_shift = 6;
  47 #define ELV_HASH_BLOCK(sec)     ((sec) >> 3)
  48 #define ELV_HASH_FN(sec)        \
  49                 (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
  50 #define ELV_HASH_ENTRIES        (1 << elv_hash_shift)
  51 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
  52 #define ELV_ON_HASH(rq)         (!hlist_unhashed(&(rq)->hash))
  53
  54 /*
  55  * Query io scheduler to see if the current process issuing bio may be
  56  * merged with rq.
  57  */
  58 static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
  59 {
  60         struct request_queue *q = rq->q;
  61         elevator_t *e = q->elevator;
  62
  63         if (e->ops->elevator_allow_merge_fn)
  64                 return e->ops->elevator_allow_merge_fn(q, rq, bio);
  65
  66         return 1;
  67 }
  68
  69 /*
  70  * can we safely merge with this request?
  71  */
  72 int elv_rq_merge_ok(struct request *rq, struct bio *bio)
  73 {
  74         if (!rq_mergeable(rq))
  75                 return 0;
  76
  77         /*
  78          * different data direction or already started, don't merge
  79          */
  80         if (bio_data_dir(bio) != rq_data_dir(rq))
  81                 return 0;
  82
  83         /*
  84          * must be same device and not a special request
  85          */
  86         if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
  87                 return 0;
  88
  89         /*
  90          * only merge integrity protected bio into ditto rq
  91          */
  92         if (bio_integrity(bio) != blk_integrity_rq(rq))
  93                 return 0;
  94
  95         if (!elv_iosched_allow_merge(rq, bio))
  96                 return 0;
  97
  98         return 1;
  99 }
 100 EXPORT_SYMBOL(elv_rq_merge_ok);
 101
 102 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
 103 {
 104         int ret = ELEVATOR_NO_MERGE;
 105
 106         /*
 107          * we can merge and sequence is ok, check if it's possible
 108          */
 109         if (elv_rq_merge_ok(__rq, bio)) {
 110                 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
 111                         ret = ELEVATOR_BACK_MERGE;
 112                 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
 113                         ret = ELEVATOR_FRONT_MERGE;
 114         }
 115
 116         return ret;
 117 }
 118
 119 static struct elevator_type *elevator_find(const char *name)
 120 {
 121         struct elevator_type *e;
 122
 123         list_for_each_entry(e, &elv_list, list) {
 124                 if (!strcmp(e->elevator_name, name))
 125                         return e;
 126         }
 127
 128         return NULL;
 129 }
 130
 131 static void elevator_put(struct elevator_type *e)
 132 {
 133         module_put(e->elevator_owner);
 134 }
 135
 136 static struct elevator_type *elevator_get(const char *name)
 137 {
 138         struct elevator_type *e;
 139
 140         spin_lock(&elv_list_lock);
 141
 142         e = elevator_find(name);
 143         if (!e) {
 144                 char elv[ELV_NAME_MAX + strlen("-iosched")];
 145
 146                 spin_unlock(&elv_list_lock);
 147
 148                 if (!strcmp(name, "anticipatory"))
 149                         sprintf(elv, "as-iosched");
 150                 else
 151                         sprintf(elv, "%s-iosched", name);
 152
 153                 request_module("%s", elv);
 154                 spin_lock(&elv_list_lock);
 155                 e = elevator_find(name);
 156         }
 157
 158         if (e && !try_module_get(e->elevator_owner))
 159                 e = NULL;
 160
 161         spin_unlock(&elv_list_lock);
 162
 163         return e;
 164 }
 165
 166 static void *elevator_init_queue(struct request_queue *q,
 167                                  struct elevator_queue *eq)
 168 {
 169         return eq->ops->elevator_init_fn(q);
 170 }
 171
 172 static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
 173                            void *data)
 174 {
 175         q->elevator = eq;
 176         eq->elevator_data = data;
 177 }
 178
 179 static char chosen_elevator[16];
 180
 181 static int __init elevator_setup(char *str)
 182 {
 183         /*
 184          * Be backwards-compatible with previous kernels, so users
 185          * won't get the wrong elevator.
 186          */
 187         if (!strcmp(str, "as"))
 188                 strcpy(chosen_elevator, "anticipatory");
 189         else
 190                 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 191         return 1;
 192 }
 193
 194 __setup("elevator=", elevator_setup);
 195
 196 static struct kobj_type elv_ktype;
 197
 198 static elevator_t *elevator_alloc(struct request_queue *q,
 199                                   struct elevator_type *e)
 200 {
 201         elevator_t *eq;
 202         int i;
 203
 204         eq = kmalloc_node(sizeof(elevator_t), GFP_KERNEL | __GFP_ZERO, q->node);
 205         if (unlikely(!eq))
 206                 goto err;
 207
 208         eq->ops = &e->ops;
 209         eq->elevator_type = e;
 210         kobject_init(&eq->kobj, &elv_ktype);
 211         mutex_init(&eq->sysfs_lock);
 212
 213         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 214                                         GFP_KERNEL, q->node);
 215         if (!eq->hash)
 216                 goto err;
 217
 218         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 219                 INIT_HLIST_HEAD(&eq->hash[i]);
 220
 221         return eq;
 222 err:
 223         kfree(eq);
 224         elevator_put(e);
 225         return NULL;
 226 }
 227
 228 static void elevator_release(struct kobject *kobj)
 229 {
 230         elevator_t *e = container_of(kobj, elevator_t, kobj);
 231
 232         elevator_put(e->elevator_type);
 233         kfree(e->hash);
 234         kfree(e);
 235 }
 236
 237 int elevator_init(struct request_queue *q, char *name)
 238 {
 239         struct elevator_type *e = NULL;
 240         struct elevator_queue *eq;
 241         int ret = 0;
 242         void *data;
 243
 244         INIT_LIST_HEAD(&q->queue_head);
 245         q->last_merge = NULL;
 246         q->end_sector = 0;
 247         q->boundary_rq = NULL;
 248
 249         if (name) {
 250                 e = elevator_get(name);
 251                 if (!e)
 252                         return -EINVAL;
 253         }
 254
 255         if (!e && *chosen_elevator) {
 256                 e = elevator_get(chosen_elevator);
 257                 if (!e)
 258                         printk(KERN_ERR "I/O scheduler %s not found\n",
 259                                                         chosen_elevator);
 260         }
 261
 262         if (!e) {
 263                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
 264                 if (!e) {
 265                         printk(KERN_ERR
 266                                 "Default I/O scheduler not found. " \
 267                                 "Using noop.\n");
 268                         e = elevator_get("noop");
 269                 }
 270         }
 271
 272         eq = elevator_alloc(q, e);
 273         if (!eq)
 274                 return -ENOMEM;
 275
 276         data = elevator_init_queue(q, eq);
 277         if (!data) {
 278                 kobject_put(&eq->kobj);
 279                 return -ENOMEM;
 280         }
 281
 282         elevator_attach(q, eq, data);
 283         return ret;
 284 }
 285 EXPORT_SYMBOL(elevator_init);
 286
 287 void elevator_exit(elevator_t *e)
 288 {
 289         mutex_lock(&e->sysfs_lock);
 290         if (e->ops->elevator_exit_fn)
 291                 e->ops->elevator_exit_fn(e);
 292         e->ops = NULL;
 293         mutex_unlock(&e->sysfs_lock);
 294
 295         kobject_put(&e->kobj);
 296 }
 297 EXPORT_SYMBOL(elevator_exit);
 298
 299 static void elv_activate_rq(struct request_queue *q, struct request *rq)
 300 {
 301         elevator_t *e = q->elevator;
 302
 303         if (e->ops->elevator_activate_req_fn)
 304                 e->ops->elevator_activate_req_fn(q, rq);
 305 }
 306
 307 static void elv_deactivate_rq(struct request_queue *q, struct request *rq)
 308 {
 309         elevator_t *e = q->elevator;
 310
 311         if (e->ops->elevator_deactivate_req_fn)
 312                 e->ops->elevator_deactivate_req_fn(q, rq);
 313 }
 314
 315 static inline void __elv_rqhash_del(struct request *rq)
 316 {
 317         hlist_del_init(&rq->hash);
 318 }
 319
 320 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 321 {
 322         if (ELV_ON_HASH(rq))
 323                 __elv_rqhash_del(rq);
 324 }
 325
 326 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 327 {
 328         elevator_t *e = q->elevator;
 329
 330         BUG_ON(ELV_ON_HASH(rq));
 331         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 332 }
 333
 334 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 335 {
 336         __elv_rqhash_del(rq);
 337         elv_rqhash_add(q, rq);
 338 }
 339
 340 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 341 {
 342         elevator_t *e = q->elevator;
 343         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 344         struct hlist_node *entry, *next;
 345         struct request *rq;
 346
 347         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 348                 BUG_ON(!ELV_ON_HASH(rq));
 349
 350                 if (unlikely(!rq_mergeable(rq))) {
 351                         __elv_rqhash_del(rq);
 352                         continue;
 353                 }
 354
 355                 if (rq_hash_key(rq) == offset)
 356                         return rq;
 357         }
 358
 359         return NULL;
 360 }
 361
 362 /*
 363  * RB-tree support functions for inserting/lookup/removal of requests
 364  * in a sorted RB tree.
 365  */
 366 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 367 {
 368         struct rb_node **p = &root->rb_node;
 369         struct rb_node *parent = NULL;
 370         struct request *__rq;
 371
 372         while (*p) {
 373                 parent = *p;
 374                 __rq = rb_entry(parent, struct request, rb_node);
 375
 376                 if (rq->sector < __rq->sector)
 377                         p = &(*p)->rb_left;
 378                 else if (rq->sector > __rq->sector)
 379                         p = &(*p)->rb_right;
 380                 else
 381                         return __rq;
 382         }
 383
 384         rb_link_node(&rq->rb_node, parent, p);
 385         rb_insert_color(&rq->rb_node, root);
 386         return NULL;
 387 }
 388 EXPORT_SYMBOL(elv_rb_add);
 389
 390 void elv_rb_del(struct rb_root *root, struct request *rq)
 391 {
 392         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 393         rb_erase(&rq->rb_node, root);
 394         RB_CLEAR_NODE(&rq->rb_node);
 395 }
 396 EXPORT_SYMBOL(elv_rb_del);
 397
 398 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 399 {
 400         struct rb_node *n = root->rb_node;
 401         struct request *rq;
 402
 403         while (n) {
 404                 rq = rb_entry(n, struct request, rb_node);
 405
 406                 if (sector < rq->sector)
 407                         n = n->rb_left;
 408                 else if (sector > rq->sector)
 409                         n = n->rb_right;
 410                 else
 411                         return rq;
 412         }
 413
 414         return NULL;
 415 }
 416 EXPORT_SYMBOL(elv_rb_find);
 417
 418 /*
 419  * Insert rq into dispatch queue of q.  Queue lock must be held on
 420  * entry.  rq is sort instead into the dispatch queue. To be used by
 421  * specific elevators.
 422  */
 423 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 424 {
 425         sector_t boundary;
 426         struct list_head *entry;
 427         int stop_flags;
 428
 429         if (q->last_merge == rq)
 430                 q->last_merge = NULL;
 431
 432         elv_rqhash_del(q, rq);
 433
 434         q->nr_sorted--;
 435
 436         boundary = q->end_sector;
 437         stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
 438         list_for_each_prev(entry, &q->queue_head) {
 439                 struct request *pos = list_entry_rq(entry);
 440
 441                 if (rq_data_dir(rq) != rq_data_dir(pos))
 442                         break;
 443                 if (pos->cmd_flags & stop_flags)
 444                         break;
 445                 if (rq->sector >= boundary) {
 446                         if (pos->sector < boundary)
 447                                 continue;
 448                 } else {
 449                         if (pos->sector >= boundary)
 450                                 break;
 451                 }
 452                 if (rq->sector >= pos->sector)
 453                         break;
 454         }
 455
 456         list_add(&rq->queuelist, entry);
 457 }
 458 EXPORT_SYMBOL(elv_dispatch_sort);
 459
 460 /*
 461  * Insert rq into dispatch queue of q.  Queue lock must be held on
 462  * entry.  rq is added to the back of the dispatch queue. To be used by
 463  * specific elevators.
 464  */
 465 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 466 {
 467         if (q->last_merge == rq)
 468                 q->last_merge = NULL;
 469
 470         elv_rqhash_del(q, rq);
 471
 472         q->nr_sorted--;
 473
 474         q->end_sector = rq_end_sector(rq);
 475         q->boundary_rq = rq;
 476         list_add_tail(&rq->queuelist, &q->queue_head);
 477 }
 478 EXPORT_SYMBOL(elv_dispatch_add_tail);
 479
 480 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 481 {
 482         elevator_t *e = q->elevator;
 483         struct request *__rq;
 484         int ret;
 485
 486         /*
 487          * First try one-hit cache.
 488          */
 489         if (q->last_merge) {
 490                 ret = elv_try_merge(q->last_merge, bio);
 491                 if (ret != ELEVATOR_NO_MERGE) {
 492                         *req = q->last_merge;
 493                         return ret;
 494                 }
 495         }
 496
 497         if (blk_queue_nomerges(q))
 498                 return ELEVATOR_NO_MERGE;
 499
 500         /*
 501          * See if our hash lookup can find a potential backmerge.
 502          */
 503         __rq = elv_rqhash_find(q, bio->bi_sector);
 504         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 505                 *req = __rq;
 506                 return ELEVATOR_BACK_MERGE;
 507         }
 508
 509         if (e->ops->elevator_merge_fn)
 510                 return e->ops->elevator_merge_fn(q, req, bio);
 511
 512         return ELEVATOR_NO_MERGE;
 513 }
 514
 515 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 516 {
 517         elevator_t *e = q->elevator;
 518
 519         if (e->ops->elevator_merged_fn)
 520                 e->ops->elevator_merged_fn(q, rq, type);
 521
 522         if (type == ELEVATOR_BACK_MERGE)
 523                 elv_rqhash_reposition(q, rq);
 524
 525         q->last_merge = rq;
 526 }
 527
 528 void elv_merge_requests(struct request_queue *q, struct request *rq,
 529                              struct request *next)
 530 {
 531         elevator_t *e = q->elevator;
 532
 533         if (e->ops->elevator_merge_req_fn)
 534                 e->ops->elevator_merge_req_fn(q, rq, next);
 535
 536         elv_rqhash_reposition(q, rq);
 537         elv_rqhash_del(q, next);
 538
 539         q->nr_sorted--;
 540         q->last_merge = rq;
 541 }
 542
 543 void elv_requeue_request(struct request_queue *q, struct request *rq)
 544 {
 545         /*
 546          * it already went through dequeue, we need to decrement the
 547          * in_flight count again
 548          */
 549         if (blk_account_rq(rq)) {
 550                 q->in_flight--;
 551                 if (blk_sorted_rq(rq))
 552                         elv_deactivate_rq(q, rq);
 553         }
 554
 555         rq->cmd_flags &= ~REQ_STARTED;
 556
 557         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 558 }
 559
 560 static void elv_drain_elevator(struct request_queue *q)
 561 {
 562         static int printed;
 563         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 564                 ;
 565         if (q->nr_sorted == 0)
 566                 return;
 567         if (printed++ < 10) {
 568                 printk(KERN_ERR "%s: forced dispatching is broken "
 569                        "(nr_sorted=%u), please report this\n",
 570                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 571         }
 572 }
 573
 574 void elv_insert(struct request_queue *q, struct request *rq, int where)
 575 {
 576         struct list_head *pos;
 577         unsigned ordseq;
 578         int unplug_it = 1;
 579
 580         blk_add_trace_rq(q, rq, BLK_TA_INSERT);
 581
 582         rq->q = q;
 583
 584         switch (where) {
 585         case ELEVATOR_INSERT_FRONT:
 586                 rq->cmd_flags |= REQ_SOFTBARRIER;
 587
 588                 list_add(&rq->queuelist, &q->queue_head);
 589                 break;
 590
 591         case ELEVATOR_INSERT_BACK:
 592                 rq->cmd_flags |= REQ_SOFTBARRIER;
 593                 elv_drain_elevator(q);
 594                 list_add_tail(&rq->queuelist, &q->queue_head);
 595                 /*
 596                  * We kick the queue here for the following reasons.
 597                  * - The elevator might have returned NULL previously
 598                  *   to delay requests and returned them now.  As the
 599                  *   queue wasn't empty before this request, ll_rw_blk
 600                  *   won't run the queue on return, resulting in hang.
 601                  * - Usually, back inserted requests won't be merged
 602                  *   with anything.  There's no point in delaying queue
 603                  *   processing.
 604                  */
 605                 blk_remove_plug(q);
 606                 q->request_fn(q);
 607                 break;
 608
 609         case ELEVATOR_INSERT_SORT:
 610                 BUG_ON(!blk_fs_request(rq));
 611                 rq->cmd_flags |= REQ_SORTED;
 612                 q->nr_sorted++;
 613                 if (rq_mergeable(rq)) {
 614                         elv_rqhash_add(q, rq);
 615                         if (!q->last_merge)
 616                                 q->last_merge = rq;
 617                 }
 618
 619                 /*
 620                  * Some ioscheds (cfq) run q->request_fn directly, so
 621                  * rq cannot be accessed after calling
 622                  * elevator_add_req_fn.
 623                  */
 624                 q->elevator->ops->elevator_add_req_fn(q, rq);
 625                 break;
 626
 627         case ELEVATOR_INSERT_REQUEUE:
 628                 /*
 629                  * If ordered flush isn't in progress, we do front
 630                  * insertion; otherwise, requests should be requeued
 631                  * in ordseq order.
 632                  */
 633                 rq->cmd_flags |= REQ_SOFTBARRIER;
 634
 635                 /*
 636                  * Most requeues happen because of a busy condition,
 637                  * don't force unplug of the queue for that case.
 638                  */
 639                 unplug_it = 0;
 640
 641                 if (q->ordseq == 0) {
 642                         list_add(&rq->queuelist, &q->queue_head);
 643                         break;
 644                 }
 645
 646                 ordseq = blk_ordered_req_seq(rq);
 647
 648                 list_for_each(pos, &q->queue_head) {
 649                         struct request *pos_rq = list_entry_rq(pos);
 650                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 651                                 break;
 652                 }
 653
 654                 list_add_tail(&rq->queuelist, pos);
 655                 break;
 656
 657         default:
 658                 printk(KERN_ERR "%s: bad insertion point %d\n",
 659                        __func__, where);
 660                 BUG();
 661         }
 662
 663         if (unplug_it && blk_queue_plugged(q)) {
 664                 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
 665                         - q->in_flight;
 666
 667                 if (nrq >= q->unplug_thresh)
 668                         __generic_unplug_device(q);
 669         }
 670 }
 671
 672 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
 673                        int plug)
 674 {
 675         if (q->ordcolor)
 676                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 677
 678         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 679                 /*
 680                  * toggle ordered color
 681                  */
 682                 if (blk_barrier_rq(rq))
 683                         q->ordcolor ^= 1;
 684
 685                 /*
 686                  * barriers implicitly indicate back insertion
 687                  */
 688                 if (where == ELEVATOR_INSERT_SORT)
 689                         where = ELEVATOR_INSERT_BACK;
 690
 691                 /*
 692                  * this request is scheduling boundary, update
 693                  * end_sector
 694                  */
 695                 if (blk_fs_request(rq)) {
 696                         q->end_sector = rq_end_sector(rq);
 697                         q->boundary_rq = rq;
 698                 }
 699         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 700                     where == ELEVATOR_INSERT_SORT)
 701                 where = ELEVATOR_INSERT_BACK;
 702
 703         if (plug)
 704                 blk_plug_device(q);
 705
 706         elv_insert(q, rq, where);
 707 }
 708 EXPORT_SYMBOL(__elv_add_request);
 709
 710 void elv_add_request(struct request_queue *q, struct request *rq, int where,
 711                      int plug)
 712 {
 713         unsigned long flags;
 714
 715         spin_lock_irqsave(q->queue_lock, flags);
 716         __elv_add_request(q, rq, where, plug);
 717         spin_unlock_irqrestore(q->queue_lock, flags);
 718 }
 719 EXPORT_SYMBOL(elv_add_request);
 720
 721 static inline struct request *__elv_next_request(struct request_queue *q)
 722 {
 723         struct request *rq;
 724
 725         while (1) {
 726                 while (!list_empty(&q->queue_head)) {
 727                         rq = list_entry_rq(q->queue_head.next);
 728                         if (blk_do_ordered(q, &rq))
 729                                 return rq;
 730                 }
 731
 732                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
 733                         return NULL;
 734         }
 735 }
 736
 737 struct request *elv_next_request(struct request_queue *q)
 738 {
 739         struct request *rq;
 740         int ret;
 741
 742         while ((rq = __elv_next_request(q)) != NULL) {
 743                 /*
 744                  * Kill the empty barrier place holder, the driver must
 745                  * not ever see it.
 746                  */
 747                 if (blk_empty_barrier(rq)) {
 748                         end_queued_request(rq, 1);
 749                         continue;
 750                 }
 751                 if (!(rq->cmd_flags & REQ_STARTED)) {
 752                         /*
 753                          * This is the first time the device driver
 754                          * sees this request (possibly after
 755                          * requeueing).  Notify IO scheduler.
 756                          */
 757                         if (blk_sorted_rq(rq))
 758                                 elv_activate_rq(q, rq);
 759
 760                         /*
 761                          * just mark as started even if we don't start
 762                          * it, a request that has been delayed should
 763                          * not be passed by new incoming requests
 764                          */
 765                         rq->cmd_flags |= REQ_STARTED;
 766                         blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
 767                 }
 768
 769                 if (!q->boundary_rq || q->boundary_rq == rq) {
 770                         q->end_sector = rq_end_sector(rq);
 771                         q->boundary_rq = NULL;
 772                 }
 773
 774                 if (rq->cmd_flags & REQ_DONTPREP)
 775                         break;
 776
 777                 if (q->dma_drain_size && rq->data_len) {
 778                         /*
 779                          * make sure space for the drain appears we
 780                          * know we can do this because max_hw_segments
 781                          * has been adjusted to be one fewer than the
 782                          * device can handle
 783                          */
 784                         rq->nr_phys_segments++;
 785                         rq->nr_hw_segments++;
 786                 }
 787
 788                 if (!q->prep_rq_fn)
 789                         break;
 790
 791                 ret = q->prep_rq_fn(q, rq);
 792                 if (ret == BLKPREP_OK) {
 793                         break;
 794                 } else if (ret == BLKPREP_DEFER) {
 795                         /*
 796                          * the request may have been (partially) prepped.
 797                          * we need to keep this request in the front to
 798                          * avoid resource deadlock.  REQ_STARTED will
 799                          * prevent other fs requests from passing this one.
 800                          */
 801                         if (q->dma_drain_size && rq->data_len &&
 802                             !(rq->cmd_flags & REQ_DONTPREP)) {
 803                                 /*
 804                                  * remove the space for the drain we added
 805                                  * so that we don't add it again
 806                                  */
 807                                 --rq->nr_phys_segments;
 808                                 --rq->nr_hw_segments;
 809                         }
 810
 811                         rq = NULL;
 812                         break;
 813                 } else if (ret == BLKPREP_KILL) {
 814                         rq->cmd_flags |= REQ_QUIET;
 815                         end_queued_request(rq, 0);
 816                 } else {
 817                         printk(KERN_ERR "%s: bad return=%d\n", __func__, ret);
 818                         break;
 819                 }
 820         }
 821
 822         return rq;
 823 }
 824 EXPORT_SYMBOL(elv_next_request);
 825
 826 void elv_dequeue_request(struct request_queue *q, struct request *rq)
 827 {
 828         BUG_ON(list_empty(&rq->queuelist));
 829         BUG_ON(ELV_ON_HASH(rq));
 830
 831         list_del_init(&rq->queuelist);
 832
 833         /*
 834          * the time frame between a request being removed from the lists
 835          * and to it is freed is accounted as io that is in progress at
 836          * the driver side.
 837          */
 838         if (blk_account_rq(rq))
 839                 q->in_flight++;
 840 }
 841 EXPORT_SYMBOL(elv_dequeue_request);
 842
 843 int elv_queue_empty(struct request_queue *q)
 844 {
 845         elevator_t *e = q->elevator;
 846
 847         if (!list_empty(&q->queue_head))
 848                 return 0;
 849
 850         if (e->ops->elevator_queue_empty_fn)
 851                 return e->ops->elevator_queue_empty_fn(q);
 852
 853         return 1;
 854 }
 855 EXPORT_SYMBOL(elv_queue_empty);
 856
 857 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 858 {
 859         elevator_t *e = q->elevator;
 860
 861         if (e->ops->elevator_latter_req_fn)
 862                 return e->ops->elevator_latter_req_fn(q, rq);
 863         return NULL;
 864 }
 865
 866 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 867 {
 868         elevator_t *e = q->elevator;
 869
 870         if (e->ops->elevator_former_req_fn)
 871                 return e->ops->elevator_former_req_fn(q, rq);
 872         return NULL;
 873 }
 874
 875 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 876 {
 877         elevator_t *e = q->elevator;
 878
 879         if (e->ops->elevator_set_req_fn)
 880                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 881
 882         rq->elevator_private = NULL;
 883         return 0;
 884 }
 885
 886 void elv_put_request(struct request_queue *q, struct request *rq)
 887 {
 888         elevator_t *e = q->elevator;
 889
 890         if (e->ops->elevator_put_req_fn)
 891                 e->ops->elevator_put_req_fn(rq);
 892 }
 893
 894 int elv_may_queue(struct request_queue *q, int rw)
 895 {
 896         elevator_t *e = q->elevator;
 897
 898         if (e->ops->elevator_may_queue_fn)
 899                 return e->ops->elevator_may_queue_fn(q, rw);
 900
 901         return ELV_MQUEUE_MAY;
 902 }
 903
 904 void elv_completed_request(struct request_queue *q, struct request *rq)
 905 {
 906         elevator_t *e = q->elevator;
 907
 908         /*
 909          * request is released from the driver, io must be done
 910          */
 911         if (blk_account_rq(rq)) {
 912                 q->in_flight--;
 913                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 914                         e->ops->elevator_completed_req_fn(q, rq);
 915         }
 916
 917         /*
 918          * Check if the queue is waiting for fs requests to be
 919          * drained for flush sequence.
 920          */
 921         if (unlikely(q->ordseq)) {
 922                 struct request *first_rq = list_entry_rq(q->queue_head.next);
 923                 if (q->in_flight == 0 &&
 924                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 925                     blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
 926                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 927                         q->request_fn(q);
 928                 }
 929         }
 930 }
 931
 932 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 933
 934 static ssize_t
 935 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 936 {
 937         elevator_t *e = container_of(kobj, elevator_t, kobj);
 938         struct elv_fs_entry *entry = to_elv(attr);
 939         ssize_t error;
 940
 941         if (!entry->show)
 942                 return -EIO;
 943
 944         mutex_lock(&e->sysfs_lock);
 945         error = e->ops ? entry->show(e, page) : -ENOENT;
 946         mutex_unlock(&e->sysfs_lock);
 947         return error;
 948 }
 949
 950 static ssize_t
 951 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 952                const char *page, size_t length)
 953 {
 954         elevator_t *e = container_of(kobj, elevator_t, kobj);
 955         struct elv_fs_entry *entry = to_elv(attr);
 956         ssize_t error;
 957
 958         if (!entry->store)
 959                 return -EIO;
 960
 961         mutex_lock(&e->sysfs_lock);
 962         error = e->ops ? entry->store(e, page, length) : -ENOENT;
 963         mutex_unlock(&e->sysfs_lock);
 964         return error;
 965 }
 966
 967 static struct sysfs_ops elv_sysfs_ops = {
 968         .show   = elv_attr_show,
 969         .store  = elv_attr_store,
 970 };
 971
 972 static struct kobj_type elv_ktype = {
 973         .sysfs_ops      = &elv_sysfs_ops,
 974         .release        = elevator_release,
 975 };
 976
 977 int elv_register_queue(struct request_queue *q)
 978 {
 979         elevator_t *e = q->elevator;
 980         int error;
 981
 982         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 983         if (!error) {
 984                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
 985                 if (attr) {
 986                         while (attr->attr.name) {
 987                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 988                                         break;
 989                                 attr++;
 990                         }
 991                 }
 992                 kobject_uevent(&e->kobj, KOBJ_ADD);
 993         }
 994         return error;
 995 }
 996
 997 static void __elv_unregister_queue(elevator_t *e)
 998 {
 999         kobject_uevent(&e->kobj, KOBJ_REMOVE);
1000         kobject_del(&e->kobj);
1001 }
1002
1003 void elv_unregister_queue(struct request_queue *q)
1004 {
1005         if (q)
1006                 __elv_unregister_queue(q->elevator);
1007 }
1008
1009 void elv_register(struct elevator_type *e)
1010 {
1011         char *def = "";
1012
1013         spin_lock(&elv_list_lock);
1014         BUG_ON(elevator_find(e->elevator_name));
1015         list_add_tail(&e->list, &elv_list);
1016         spin_unlock(&elv_list_lock);
1017
1018         if (!strcmp(e->elevator_name, chosen_elevator) ||
1019                         (!*chosen_elevator &&
1020                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
1021                                 def = " (default)";
1022
1023         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
1024                                                                 def);
1025 }
1026 EXPORT_SYMBOL_GPL(elv_register);
1027
1028 void elv_unregister(struct elevator_type *e)
1029 {
1030         struct task_struct *g, *p;
1031
1032         /*
1033          * Iterate every thread in the process to remove the io contexts.
1034          */
1035         if (e->ops.trim) {
1036                 read_lock(&tasklist_lock);
1037                 do_each_thread(g, p) {
1038                         task_lock(p);
1039                         if (p->io_context)
1040                                 e->ops.trim(p->io_context);
1041                         task_unlock(p);
1042                 } while_each_thread(g, p);
1043                 read_unlock(&tasklist_lock);
1044         }
1045
1046         spin_lock(&elv_list_lock);
1047         list_del_init(&e->list);
1048         spin_unlock(&elv_list_lock);
1049 }
1050 EXPORT_SYMBOL_GPL(elv_unregister);
1051
1052 /*
1053  * switch to new_e io scheduler. be careful not to introduce deadlocks -
1054  * we don't free the old io scheduler, before we have allocated what we
1055  * need for the new one. this way we have a chance of going back to the old
1056  * one, if the new one fails init for some reason.
1057  */
1058 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1059 {
1060         elevator_t *old_elevator, *e;
1061         void *data;
1062
1063         /*
1064          * Allocate new elevator
1065          */
1066         e = elevator_alloc(q, new_e);
1067         if (!e)
1068                 return 0;
1069
1070         data = elevator_init_queue(q, e);
1071         if (!data) {
1072                 kobject_put(&e->kobj);
1073                 return 0;
1074         }
1075
1076         /*
1077          * Turn on BYPASS and drain all requests w/ elevator private data
1078          */
1079         spin_lock_irq(q->queue_lock);
1080
1081         queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
1082
1083         elv_drain_elevator(q);
1084
1085         while (q->rq.elvpriv) {
1086                 blk_remove_plug(q);
1087                 q->request_fn(q);
1088                 spin_unlock_irq(q->queue_lock);
1089                 msleep(10);
1090                 spin_lock_irq(q->queue_lock);
1091                 elv_drain_elevator(q);
1092         }
1093
1094         /*
1095          * Remember old elevator.
1096          */
1097         old_elevator = q->elevator;
1098
1099         /*
1100          * attach and start new elevator
1101          */
1102         elevator_attach(q, e, data);
1103
1104         spin_unlock_irq(q->queue_lock);
1105
1106         __elv_unregister_queue(old_elevator);
1107
1108         if (elv_register_queue(q))
1109                 goto fail_register;
1110
1111         /*
1112          * finally exit old elevator and turn off BYPASS.
1113          */
1114         elevator_exit(old_elevator);
1115         spin_lock_irq(q->queue_lock);
1116         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1117         spin_unlock_irq(q->queue_lock);
1118
1119         blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
1120
1121         return 1;
1122
1123 fail_register:
1124         /*
1125          * switch failed, exit the new io scheduler and reattach the old
1126          * one again (along with re-adding the sysfs dir)
1127          */
1128         elevator_exit(e);
1129         q->elevator = old_elevator;
1130         elv_register_queue(q);
1131
1132         spin_lock_irq(q->queue_lock);
1133         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1134         spin_unlock_irq(q->queue_lock);
1135
1136         return 0;
1137 }
1138
1139 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1140                           size_t count)
1141 {
1142         char elevator_name[ELV_NAME_MAX];
1143         size_t len;
1144         struct elevator_type *e;
1145
1146         elevator_name[sizeof(elevator_name) - 1] = '\0';
1147         strncpy(elevator_name, name, sizeof(elevator_name) - 1);
1148         len = strlen(elevator_name);
1149
1150         if (len && elevator_name[len - 1] == '\n')
1151                 elevator_name[len - 1] = '\0';
1152
1153         e = elevator_get(elevator_name);
1154         if (!e) {
1155                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1156                 return -EINVAL;
1157         }
1158
1159         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1160                 elevator_put(e);
1161                 return count;
1162         }
1163
1164         if (!elevator_switch(q, e))
1165                 printk(KERN_ERR "elevator: switch to %s failed\n",
1166                                                         elevator_name);
1167         return count;
1168 }
1169
1170 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1171 {
1172         elevator_t *e = q->elevator;
1173         struct elevator_type *elv = e->elevator_type;
1174         struct elevator_type *__e;
1175         int len = 0;
1176
1177         spin_lock(&elv_list_lock);
1178         list_for_each_entry(__e, &elv_list, list) {
1179                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1180                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1181                 else
1182                         len += sprintf(name+len, "%s ", __e->elevator_name);
1183         }
1184         spin_unlock(&elv_list_lock);
1185
1186         len += sprintf(len+name, "\n");
1187         return len;
1188 }
1189
1190 struct request *elv_rb_former_request(struct request_queue *q,
1191                                       struct request *rq)
1192 {
1193         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1194
1195         if (rbprev)
1196                 return rb_entry_rq(rbprev);
1197
1198         return NULL;
1199 }
1200 EXPORT_SYMBOL(elv_rb_former_request);
1201
1202 struct request *elv_rb_latter_request(struct request_queue *q,
1203                                       struct request *rq)
1204 {
1205         struct rb_node *rbnext = rb_next(&rq->rb_node);
1206
1207         if (rbnext)
1208                 return rb_entry_rq(rbnext);
1209
1210         return NULL;
1211 }
1212 EXPORT_SYMBOL(elv_rb_latter_request);