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