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 (blk_discard_rq(rq) != blk_discard_rq(pos))
 432                         break;
 433                 if (rq_data_dir(rq) != rq_data_dir(pos))
 434                         break;
 435                 if (pos->cmd_flags & stop_flags)
 436                         break;
 437                 if (blk_rq_pos(rq) >= boundary) {
 438                         if (blk_rq_pos(pos) < boundary)
 439                                 continue;
 440                 } else {
 441                         if (blk_rq_pos(pos) >= boundary)
 442                                 break;
 443                 }
 444                 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
 445                         break;
 446         }
 447
 448         list_add(&rq->queuelist, entry);
 449 }
 450 EXPORT_SYMBOL(elv_dispatch_sort);
 451
 452 /*
 453  * Insert rq into dispatch queue of q.  Queue lock must be held on
 454  * entry.  rq is added to the back of the dispatch queue. To be used by
 455  * specific elevators.
 456  */
 457 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 458 {
 459         if (q->last_merge == rq)
 460                 q->last_merge = NULL;
 461
 462         elv_rqhash_del(q, rq);
 463
 464         q->nr_sorted--;
 465
 466         q->end_sector = rq_end_sector(rq);
 467         q->boundary_rq = rq;
 468         list_add_tail(&rq->queuelist, &q->queue_head);
 469 }
 470 EXPORT_SYMBOL(elv_dispatch_add_tail);
 471
 472 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 473 {
 474         struct elevator_queue *e = q->elevator;
 475         struct request *__rq;
 476         int ret;
 477
 478         /*
 479          * Levels of merges:
 480          *      nomerges:  No merges at all attempted
 481          *      noxmerges: Only simple one-hit cache try
 482          *      merges:    All merge tries attempted
 483          */
 484         if (blk_queue_nomerges(q))
 485                 return ELEVATOR_NO_MERGE;
 486
 487         /*
 488          * First try one-hit cache.
 489          */
 490         if (q->last_merge) {
 491                 ret = elv_try_merge(q->last_merge, bio);
 492                 if (ret != ELEVATOR_NO_MERGE) {
 493                         *req = q->last_merge;
 494                         return ret;
 495                 }
 496         }
 497
 498         if (blk_queue_noxmerges(q))
 499                 return ELEVATOR_NO_MERGE;
 500
 501         /*
 502          * See if our hash lookup can find a potential backmerge.
 503          */
 504         __rq = elv_rqhash_find(q, bio->bi_sector);
 505         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 506                 *req = __rq;
 507                 return ELEVATOR_BACK_MERGE;
 508         }
 509
 510         if (e->ops->elevator_merge_fn)
 511                 return e->ops->elevator_merge_fn(q, req, bio);
 512
 513         return ELEVATOR_NO_MERGE;
 514 }
 515
 516 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 517 {
 518         struct elevator_queue *e = q->elevator;
 519
 520         if (e->ops->elevator_merged_fn)
 521                 e->ops->elevator_merged_fn(q, rq, type);
 522
 523         if (type == ELEVATOR_BACK_MERGE)
 524                 elv_rqhash_reposition(q, rq);
 525
 526         q->last_merge = rq;
 527 }
 528
 529 void elv_merge_requests(struct request_queue *q, struct request *rq,
 530                              struct request *next)
 531 {
 532         struct elevator_queue *e = q->elevator;
 533
 534         if (e->ops->elevator_merge_req_fn)
 535                 e->ops->elevator_merge_req_fn(q, rq, next);
 536
 537         elv_rqhash_reposition(q, rq);
 538         elv_rqhash_del(q, next);
 539
 540         q->nr_sorted--;
 541         q->last_merge = rq;
 542 }
 543
 544 void elv_bio_merged(struct request_queue *q, struct request *rq,
 545                         struct bio *bio)
 546 {
 547         struct elevator_queue *e = q->elevator;
 548
 549         if (e->ops->elevator_bio_merged_fn)
 550                 e->ops->elevator_bio_merged_fn(q, rq, bio);
 551 }
 552
 553 void elv_requeue_request(struct request_queue *q, struct request *rq)
 554 {
 555         /*
 556          * it already went through dequeue, we need to decrement the
 557          * in_flight count again
 558          */
 559         if (blk_account_rq(rq)) {
 560                 q->in_flight[rq_is_sync(rq)]--;
 561                 if (blk_sorted_rq(rq))
 562                         elv_deactivate_rq(q, rq);
 563         }
 564
 565         rq->cmd_flags &= ~REQ_STARTED;
 566
 567         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 568 }
 569
 570 void elv_drain_elevator(struct request_queue *q)
 571 {
 572         static int printed;
 573         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 574                 ;
 575         if (q->nr_sorted == 0)
 576                 return;
 577         if (printed++ < 10) {
 578                 printk(KERN_ERR "%s: forced dispatching is broken "
 579                        "(nr_sorted=%u), please report this\n",
 580                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 581         }
 582 }
 583
 584 /*
 585  * Call with queue lock held, interrupts disabled
 586  */
 587 void elv_quiesce_start(struct request_queue *q)
 588 {
 589         if (!q->elevator)
 590                 return;
 591
 592         queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
 593
 594         /*
 595          * make sure we don't have any requests in flight
 596          */
 597         elv_drain_elevator(q);
 598         while (q->rq.elvpriv) {
 599                 __blk_run_queue(q);
 600                 spin_unlock_irq(q->queue_lock);
 601                 msleep(10);
 602                 spin_lock_irq(q->queue_lock);
 603                 elv_drain_elevator(q);
 604         }
 605 }
 606
 607 void elv_quiesce_end(struct request_queue *q)
 608 {
 609         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
 610 }
 611
 612 void elv_insert(struct request_queue *q, struct request *rq, int where)
 613 {
 614         struct list_head *pos;
 615         unsigned ordseq;
 616         int unplug_it = 1;
 617
 618         trace_block_rq_insert(q, rq);
 619
 620         rq->q = q;
 621
 622         switch (where) {
 623         case ELEVATOR_INSERT_FRONT:
 624                 rq->cmd_flags |= REQ_SOFTBARRIER;
 625
 626                 list_add(&rq->queuelist, &q->queue_head);
 627                 break;
 628
 629         case ELEVATOR_INSERT_BACK:
 630                 rq->cmd_flags |= REQ_SOFTBARRIER;
 631                 elv_drain_elevator(q);
 632                 list_add_tail(&rq->queuelist, &q->queue_head);
 633                 /*
 634                  * We kick the queue here for the following reasons.
 635                  * - The elevator might have returned NULL previously
 636                  *   to delay requests and returned them now.  As the
 637                  *   queue wasn't empty before this request, ll_rw_blk
 638                  *   won't run the queue on return, resulting in hang.
 639                  * - Usually, back inserted requests won't be merged
 640                  *   with anything.  There's no point in delaying queue
 641                  *   processing.
 642                  */
 643                 __blk_run_queue(q);
 644                 break;
 645
 646         case ELEVATOR_INSERT_SORT:
 647                 BUG_ON(!blk_fs_request(rq) && !blk_discard_rq(rq));
 648                 rq->cmd_flags |= REQ_SORTED;
 649                 q->nr_sorted++;
 650                 if (rq_mergeable(rq)) {
 651                         elv_rqhash_add(q, rq);
 652                         if (!q->last_merge)
 653                                 q->last_merge = rq;
 654                 }
 655
 656                 /*
 657                  * Some ioscheds (cfq) run q->request_fn directly, so
 658                  * rq cannot be accessed after calling
 659                  * elevator_add_req_fn.
 660                  */
 661                 q->elevator->ops->elevator_add_req_fn(q, rq);
 662                 break;
 663
 664         case ELEVATOR_INSERT_REQUEUE:
 665                 /*
 666                  * If ordered flush isn't in progress, we do front
 667                  * insertion; otherwise, requests should be requeued
 668                  * in ordseq order.
 669                  */
 670                 rq->cmd_flags |= REQ_SOFTBARRIER;
 671
 672                 /*
 673                  * Most requeues happen because of a busy condition,
 674                  * don't force unplug of the queue for that case.
 675                  */
 676                 unplug_it = 0;
 677
 678                 if (q->ordseq == 0) {
 679                         list_add(&rq->queuelist, &q->queue_head);
 680                         break;
 681                 }
 682
 683                 ordseq = blk_ordered_req_seq(rq);
 684
 685                 list_for_each(pos, &q->queue_head) {
 686                         struct request *pos_rq = list_entry_rq(pos);
 687                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 688                                 break;
 689                 }
 690
 691                 list_add_tail(&rq->queuelist, pos);
 692                 break;
 693
 694         default:
 695                 printk(KERN_ERR "%s: bad insertion point %d\n",
 696                        __func__, where);
 697                 BUG();
 698         }
 699
 700         if (unplug_it && blk_queue_plugged(q)) {
 701                 int nrq = q->rq.count[BLK_RW_SYNC] + q->rq.count[BLK_RW_ASYNC]
 702                                 - queue_in_flight(q);
 703
 704                 if (nrq >= q->unplug_thresh)
 705                         __generic_unplug_device(q);
 706         }
 707 }
 708
 709 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
 710                        int plug)
 711 {
 712         if (q->ordcolor)
 713                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 714
 715         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 716                 /*
 717                  * toggle ordered color
 718                  */
 719                 if (blk_barrier_rq(rq))
 720                         q->ordcolor ^= 1;
 721
 722                 /*
 723                  * barriers implicitly indicate back insertion
 724                  */
 725                 if (where == ELEVATOR_INSERT_SORT)
 726                         where = ELEVATOR_INSERT_BACK;
 727
 728                 /*
 729                  * this request is scheduling boundary, update
 730                  * end_sector
 731                  */
 732                 if (blk_fs_request(rq) || blk_discard_rq(rq)) {
 733                         q->end_sector = rq_end_sector(rq);
 734                         q->boundary_rq = rq;
 735                 }
 736         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 737                     where == ELEVATOR_INSERT_SORT)
 738                 where = ELEVATOR_INSERT_BACK;
 739
 740         if (plug)
 741                 blk_plug_device(q);
 742
 743         elv_insert(q, rq, where);
 744 }
 745 EXPORT_SYMBOL(__elv_add_request);
 746
 747 void elv_add_request(struct request_queue *q, struct request *rq, int where,
 748                      int plug)
 749 {
 750         unsigned long flags;
 751
 752         spin_lock_irqsave(q->queue_lock, flags);
 753         __elv_add_request(q, rq, where, plug);
 754         spin_unlock_irqrestore(q->queue_lock, flags);
 755 }
 756 EXPORT_SYMBOL(elv_add_request);
 757
 758 int elv_queue_empty(struct request_queue *q)
 759 {
 760         struct elevator_queue *e = q->elevator;
 761
 762         if (!list_empty(&q->queue_head))
 763                 return 0;
 764
 765         if (e->ops->elevator_queue_empty_fn)
 766                 return e->ops->elevator_queue_empty_fn(q);
 767
 768         return 1;
 769 }
 770 EXPORT_SYMBOL(elv_queue_empty);
 771
 772 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 773 {
 774         struct elevator_queue *e = q->elevator;
 775
 776         if (e->ops->elevator_latter_req_fn)
 777                 return e->ops->elevator_latter_req_fn(q, rq);
 778         return NULL;
 779 }
 780
 781 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 782 {
 783         struct elevator_queue *e = q->elevator;
 784
 785         if (e->ops->elevator_former_req_fn)
 786                 return e->ops->elevator_former_req_fn(q, rq);
 787         return NULL;
 788 }
 789
 790 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 791 {
 792         struct elevator_queue *e = q->elevator;
 793
 794         if (e->ops->elevator_set_req_fn)
 795                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 796
 797         rq->elevator_private = NULL;
 798         return 0;
 799 }
 800
 801 void elv_put_request(struct request_queue *q, struct request *rq)
 802 {
 803         struct elevator_queue *e = q->elevator;
 804
 805         if (e->ops->elevator_put_req_fn)
 806                 e->ops->elevator_put_req_fn(rq);
 807 }
 808
 809 int elv_may_queue(struct request_queue *q, int rw)
 810 {
 811         struct elevator_queue *e = q->elevator;
 812
 813         if (e->ops->elevator_may_queue_fn)
 814                 return e->ops->elevator_may_queue_fn(q, rw);
 815
 816         return ELV_MQUEUE_MAY;
 817 }
 818
 819 void elv_abort_queue(struct request_queue *q)
 820 {
 821         struct request *rq;
 822
 823         while (!list_empty(&q->queue_head)) {
 824                 rq = list_entry_rq(q->queue_head.next);
 825                 rq->cmd_flags |= REQ_QUIET;
 826                 trace_block_rq_abort(q, rq);
 827                 /*
 828                  * Mark this request as started so we don't trigger
 829                  * any debug logic in the end I/O path.
 830                  */
 831                 blk_start_request(rq);
 832                 __blk_end_request_all(rq, -EIO);
 833         }
 834 }
 835 EXPORT_SYMBOL(elv_abort_queue);
 836
 837 void elv_completed_request(struct request_queue *q, struct request *rq)
 838 {
 839         struct elevator_queue *e = q->elevator;
 840
 841         /*
 842          * request is released from the driver, io must be done
 843          */
 844         if (blk_account_rq(rq)) {
 845                 q->in_flight[rq_is_sync(rq)]--;
 846                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 847                         e->ops->elevator_completed_req_fn(q, rq);
 848         }
 849
 850         /*
 851          * Check if the queue is waiting for fs requests to be
 852          * drained for flush sequence.
 853          */
 854         if (unlikely(q->ordseq)) {
 855                 struct request *next = NULL;
 856
 857                 if (!list_empty(&q->queue_head))
 858                         next = list_entry_rq(q->queue_head.next);
 859
 860                 if (!queue_in_flight(q) &&
 861                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 862                     (!next || blk_ordered_req_seq(next) > QUEUE_ORDSEQ_DRAIN)) {
 863                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 864                         __blk_run_queue(q);
 865                 }
 866         }
 867 }
 868
 869 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 870
 871 static ssize_t
 872 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 873 {
 874         struct elv_fs_entry *entry = to_elv(attr);
 875         struct elevator_queue *e;
 876         ssize_t error;
 877
 878         if (!entry->show)
 879                 return -EIO;
 880
 881         e = container_of(kobj, struct elevator_queue, kobj);
 882         mutex_lock(&e->sysfs_lock);
 883         error = e->ops ? entry->show(e, page) : -ENOENT;
 884         mutex_unlock(&e->sysfs_lock);
 885         return error;
 886 }
 887
 888 static ssize_t
 889 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 890                const char *page, size_t length)
 891 {
 892         struct elv_fs_entry *entry = to_elv(attr);
 893         struct elevator_queue *e;
 894         ssize_t error;
 895
 896         if (!entry->store)
 897                 return -EIO;
 898
 899         e = container_of(kobj, struct elevator_queue, kobj);
 900         mutex_lock(&e->sysfs_lock);
 901         error = e->ops ? entry->store(e, page, length) : -ENOENT;
 902         mutex_unlock(&e->sysfs_lock);
 903         return error;
 904 }
 905
 906 static const struct sysfs_ops elv_sysfs_ops = {
 907         .show   = elv_attr_show,
 908         .store  = elv_attr_store,
 909 };
 910
 911 static struct kobj_type elv_ktype = {
 912         .sysfs_ops      = &elv_sysfs_ops,
 913         .release        = elevator_release,
 914 };
 915
 916 int elv_register_queue(struct request_queue *q)
 917 {
 918         struct elevator_queue *e = q->elevator;
 919         int error;
 920
 921         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 922         if (!error) {
 923                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
 924                 if (attr) {
 925                         while (attr->attr.name) {
 926                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 927                                         break;
 928                                 attr++;
 929                         }
 930                 }
 931                 kobject_uevent(&e->kobj, KOBJ_ADD);
 932         }
 933         return error;
 934 }
 935 EXPORT_SYMBOL(elv_register_queue);
 936
 937 static void __elv_unregister_queue(struct elevator_queue *e)
 938 {
 939         kobject_uevent(&e->kobj, KOBJ_REMOVE);
 940         kobject_del(&e->kobj);
 941 }
 942
 943 void elv_unregister_queue(struct request_queue *q)
 944 {
 945         if (q)
 946                 __elv_unregister_queue(q->elevator);
 947 }
 948 EXPORT_SYMBOL(elv_unregister_queue);
 949
 950 void elv_register(struct elevator_type *e)
 951 {
 952         char *def = "";
 953
 954         spin_lock(&elv_list_lock);
 955         BUG_ON(elevator_find(e->elevator_name));
 956         list_add_tail(&e->list, &elv_list);
 957         spin_unlock(&elv_list_lock);
 958
 959         if (!strcmp(e->elevator_name, chosen_elevator) ||
 960                         (!*chosen_elevator &&
 961                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 962                                 def = " (default)";
 963
 964         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
 965                                                                 def);
 966 }
 967 EXPORT_SYMBOL_GPL(elv_register);
 968
 969 void elv_unregister(struct elevator_type *e)
 970 {
 971         struct task_struct *g, *p;
 972
 973         /*
 974          * Iterate every thread in the process to remove the io contexts.
 975          */
 976         if (e->ops.trim) {
 977                 read_lock(&tasklist_lock);
 978                 do_each_thread(g, p) {
 979                         task_lock(p);
 980                         if (p->io_context)
 981                                 e->ops.trim(p->io_context);
 982                         task_unlock(p);
 983                 } while_each_thread(g, p);
 984                 read_unlock(&tasklist_lock);
 985         }
 986
 987         spin_lock(&elv_list_lock);
 988         list_del_init(&e->list);
 989         spin_unlock(&elv_list_lock);
 990 }
 991 EXPORT_SYMBOL_GPL(elv_unregister);
 992
 993 /*
 994  * switch to new_e io scheduler. be careful not to introduce deadlocks -
 995  * we don't free the old io scheduler, before we have allocated what we
 996  * need for the new one. this way we have a chance of going back to the old
 997  * one, if the new one fails init for some reason.
 998  */
 999 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1000 {
1001         struct elevator_queue *old_elevator, *e;
1002         void *data;
1003
1004         /*
1005          * Allocate new elevator
1006          */
1007         e = elevator_alloc(q, new_e);
1008         if (!e)
1009                 return 0;
1010
1011         data = elevator_init_queue(q, e);
1012         if (!data) {
1013                 kobject_put(&e->kobj);
1014                 return 0;
1015         }
1016
1017         /*
1018          * Turn on BYPASS and drain all requests w/ elevator private data
1019          */
1020         spin_lock_irq(q->queue_lock);
1021         elv_quiesce_start(q);
1022
1023         /*
1024          * Remember old elevator.
1025          */
1026         old_elevator = q->elevator;
1027
1028         /*
1029          * attach and start new elevator
1030          */
1031         elevator_attach(q, e, data);
1032
1033         spin_unlock_irq(q->queue_lock);
1034
1035         __elv_unregister_queue(old_elevator);
1036
1037         if (elv_register_queue(q))
1038                 goto fail_register;
1039
1040         /*
1041          * finally exit old elevator and turn off BYPASS.
1042          */
1043         elevator_exit(old_elevator);
1044         spin_lock_irq(q->queue_lock);
1045         elv_quiesce_end(q);
1046         spin_unlock_irq(q->queue_lock);
1047
1048         blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
1049
1050         return 1;
1051
1052 fail_register:
1053         /*
1054          * switch failed, exit the new io scheduler and reattach the old
1055          * one again (along with re-adding the sysfs dir)
1056          */
1057         elevator_exit(e);
1058         q->elevator = old_elevator;
1059         elv_register_queue(q);
1060
1061         spin_lock_irq(q->queue_lock);
1062         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1063         spin_unlock_irq(q->queue_lock);
1064
1065         return 0;
1066 }
1067
1068 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1069                           size_t count)
1070 {
1071         char elevator_name[ELV_NAME_MAX];
1072         struct elevator_type *e;
1073
1074         if (!q->elevator)
1075                 return count;
1076
1077         strlcpy(elevator_name, name, sizeof(elevator_name));
1078         e = elevator_get(strstrip(elevator_name));
1079         if (!e) {
1080                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1081                 return -EINVAL;
1082         }
1083
1084         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1085                 elevator_put(e);
1086                 return count;
1087         }
1088
1089         if (!elevator_switch(q, e))
1090                 printk(KERN_ERR "elevator: switch to %s failed\n",
1091                                                         elevator_name);
1092         return count;
1093 }
1094
1095 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1096 {
1097         struct elevator_queue *e = q->elevator;
1098         struct elevator_type *elv;
1099         struct elevator_type *__e;
1100         int len = 0;
1101
1102         if (!q->elevator || !blk_queue_stackable(q))
1103                 return sprintf(name, "none\n");
1104
1105         elv = e->elevator_type;
1106
1107         spin_lock(&elv_list_lock);
1108         list_for_each_entry(__e, &elv_list, list) {
1109                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1110                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1111                 else
1112                         len += sprintf(name+len, "%s ", __e->elevator_name);
1113         }
1114         spin_unlock(&elv_list_lock);
1115
1116         len += sprintf(len+name, "\n");
1117         return len;
1118 }
1119
1120 struct request *elv_rb_former_request(struct request_queue *q,
1121                                       struct request *rq)
1122 {
1123         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1124
1125         if (rbprev)
1126                 return rb_entry_rq(rbprev);
1127
1128         return NULL;
1129 }
1130 EXPORT_SYMBOL(elv_rb_former_request);
1131
1132 struct request *elv_rb_latter_request(struct request_queue *q,
1133                                       struct request *rq)
1134 {
1135         struct rb_node *rbnext = rb_next(&rq->rb_node);
1136
1137         if (rbnext)
1138                 return rb_entry_rq(rbnext);
1139
1140         return NULL;
1141 }
1142 EXPORT_SYMBOL(elv_rb_latter_request);