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