block/elevator.c

   1 /*
   2  *  Block device elevator/IO-scheduler.
   3  *
   4  *  Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
   5  *
   6  * 30042000 Jens Axboe <axboe@kernel.dk> :
   7  *
   8  * Split the elevator a bit so that it is possible to choose a different
   9  * one or even write a new "plug in". There are three pieces:
  10  * - elevator_fn, inserts a new request in the queue list
  11  * - elevator_merge_fn, decides whether a new buffer can be merged with
  12  *   an existing request
  13  * - elevator_dequeue_fn, called when a request is taken off the active list
  14  *
  15  * 20082000 Dave Jones <davej@suse.de> :
  16  * Removed tests for max-bomb-segments, which was breaking elvtune
  17  *  when run without -bN
  18  *
  19  * Jens:
  20  * - Rework again to work with bio instead of buffer_heads
  21  * - loose bi_dev comparisons, partition handling is right now
  22  * - completely modularize elevator setup and teardown
  23  *
  24  */
  25 #include <linux/kernel.h>
  26 #include <linux/fs.h>
  27 #include <linux/blkdev.h>
  28 #include <linux/elevator.h>
  29 #include <linux/bio.h>
  30 #include <linux/module.h>
  31 #include <linux/slab.h>
  32 #include <linux/init.h>
  33 #include <linux/compiler.h>
  34 #include <linux/delay.h>
  35 #include <linux/blktrace_api.h>
  36 #include <linux/hash.h>
  37
  38 #include <asm/uaccess.h>
  39
  40 static DEFINE_SPINLOCK(elv_list_lock);
  41 static LIST_HEAD(elv_list);
  42
  43 /*
  44  * Merge hash stuff.
  45  */
  46 static const int elv_hash_shift = 6;
  47 #define ELV_HASH_BLOCK(sec)     ((sec) >> 3)
  48 #define ELV_HASH_FN(sec)        (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
  49 #define ELV_HASH_ENTRIES        (1 << elv_hash_shift)
  50 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
  51 #define ELV_ON_HASH(rq)         (!hlist_unhashed(&(rq)->hash))
  52
  53 /*
  54  * Query io scheduler to see if the current process issuing bio may be
  55  * merged with rq.
  56  */
  57 static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
  58 {
  59         struct request_queue *q = rq->q;
  60         elevator_t *e = q->elevator;
  61
  62         if (e->ops->elevator_allow_merge_fn)
  63                 return e->ops->elevator_allow_merge_fn(q, rq, bio);
  64
  65         return 1;
  66 }
  67
  68 /*
  69  * can we safely merge with this request?
  70  */
  71 inline int elv_rq_merge_ok(struct request *rq, struct bio *bio)
  72 {
  73         if (!rq_mergeable(rq))
  74                 return 0;
  75
  76         /*
  77          * different data direction or already started, don't merge
  78          */
  79         if (bio_data_dir(bio) != rq_data_dir(rq))
  80                 return 0;
  81
  82         /*
  83          * must be same device and not a special request
  84          */
  85         if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
  86                 return 0;
  87
  88         if (!elv_iosched_allow_merge(rq, bio))
  89                 return 0;
  90
  91         return 1;
  92 }
  93 EXPORT_SYMBOL(elv_rq_merge_ok);
  94
  95 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
  96 {
  97         int ret = ELEVATOR_NO_MERGE;
  98
  99         /*
 100          * we can merge and sequence is ok, check if it's possible
 101          */
 102         if (elv_rq_merge_ok(__rq, bio)) {
 103                 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
 104                         ret = ELEVATOR_BACK_MERGE;
 105                 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
 106                         ret = ELEVATOR_FRONT_MERGE;
 107         }
 108
 109         return ret;
 110 }
 111
 112 static struct elevator_type *elevator_find(const char *name)
 113 {
 114         struct elevator_type *e;
 115
 116         list_for_each_entry(e, &elv_list, list) {
 117                 if (!strcmp(e->elevator_name, name))
 118                         return e;
 119         }
 120
 121         return NULL;
 122 }
 123
 124 static void elevator_put(struct elevator_type *e)
 125 {
 126         module_put(e->elevator_owner);
 127 }
 128
 129 static struct elevator_type *elevator_get(const char *name)
 130 {
 131         struct elevator_type *e;
 132
 133         spin_lock(&elv_list_lock);
 134
 135         e = elevator_find(name);
 136         if (e && !try_module_get(e->elevator_owner))
 137                 e = NULL;
 138
 139         spin_unlock(&elv_list_lock);
 140
 141         return e;
 142 }
 143
 144 static void *elevator_init_queue(struct request_queue *q,
 145                                  struct elevator_queue *eq)
 146 {
 147         return eq->ops->elevator_init_fn(q);
 148 }
 149
 150 static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
 151                            void *data)
 152 {
 153         q->elevator = eq;
 154         eq->elevator_data = data;
 155 }
 156
 157 static char chosen_elevator[16];
 158
 159 static int __init elevator_setup(char *str)
 160 {
 161         /*
 162          * Be backwards-compatible with previous kernels, so users
 163          * won't get the wrong elevator.
 164          */
 165         if (!strcmp(str, "as"))
 166                 strcpy(chosen_elevator, "anticipatory");
 167         else
 168                 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 169         return 1;
 170 }
 171
 172 __setup("elevator=", elevator_setup);
 173
 174 static struct kobj_type elv_ktype;
 175
 176 static elevator_t *elevator_alloc(struct request_queue *q,
 177                                   struct elevator_type *e)
 178 {
 179         elevator_t *eq;
 180         int i;
 181
 182         eq = kmalloc_node(sizeof(elevator_t), GFP_KERNEL | __GFP_ZERO, q->node);
 183         if (unlikely(!eq))
 184                 goto err;
 185
 186         eq->ops = &e->ops;
 187         eq->elevator_type = e;
 188         kobject_init(&eq->kobj, &elv_ktype);
 189         mutex_init(&eq->sysfs_lock);
 190
 191         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 192                                         GFP_KERNEL, q->node);
 193         if (!eq->hash)
 194                 goto err;
 195
 196         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 197                 INIT_HLIST_HEAD(&eq->hash[i]);
 198
 199         return eq;
 200 err:
 201         kfree(eq);
 202         elevator_put(e);
 203         return NULL;
 204 }
 205
 206 static void elevator_release(struct kobject *kobj)
 207 {
 208         elevator_t *e = container_of(kobj, elevator_t, kobj);
 209
 210         elevator_put(e->elevator_type);
 211         kfree(e->hash);
 212         kfree(e);
 213 }
 214
 215 int elevator_init(struct request_queue *q, char *name)
 216 {
 217         struct elevator_type *e = NULL;
 218         struct elevator_queue *eq;
 219         int ret = 0;
 220         void *data;
 221
 222         INIT_LIST_HEAD(&q->queue_head);
 223         q->last_merge = NULL;
 224         q->end_sector = 0;
 225         q->boundary_rq = NULL;
 226
 227         if (name && !(e = elevator_get(name)))
 228                 return -EINVAL;
 229
 230         if (!e && *chosen_elevator && !(e = elevator_get(chosen_elevator)))
 231                 printk("I/O scheduler %s not found\n", chosen_elevator);
 232
 233         if (!e && !(e = elevator_get(CONFIG_DEFAULT_IOSCHED))) {
 234                 printk("Default I/O scheduler not found, using no-op\n");
 235                 e = elevator_get("noop");
 236         }
 237
 238         eq = elevator_alloc(q, e);
 239         if (!eq)
 240                 return -ENOMEM;
 241
 242         data = elevator_init_queue(q, eq);
 243         if (!data) {
 244                 kobject_put(&eq->kobj);
 245                 return -ENOMEM;
 246         }
 247
 248         elevator_attach(q, eq, data);
 249         return ret;
 250 }
 251
 252 EXPORT_SYMBOL(elevator_init);
 253
 254 void elevator_exit(elevator_t *e)
 255 {
 256         mutex_lock(&e->sysfs_lock);
 257         if (e->ops->elevator_exit_fn)
 258                 e->ops->elevator_exit_fn(e);
 259         e->ops = NULL;
 260         mutex_unlock(&e->sysfs_lock);
 261
 262         kobject_put(&e->kobj);
 263 }
 264
 265 EXPORT_SYMBOL(elevator_exit);
 266
 267 static void elv_activate_rq(struct request_queue *q, struct request *rq)
 268 {
 269         elevator_t *e = q->elevator;
 270
 271         if (e->ops->elevator_activate_req_fn)
 272                 e->ops->elevator_activate_req_fn(q, rq);
 273 }
 274
 275 static void elv_deactivate_rq(struct request_queue *q, struct request *rq)
 276 {
 277         elevator_t *e = q->elevator;
 278
 279         if (e->ops->elevator_deactivate_req_fn)
 280                 e->ops->elevator_deactivate_req_fn(q, rq);
 281 }
 282
 283 static inline void __elv_rqhash_del(struct request *rq)
 284 {
 285         hlist_del_init(&rq->hash);
 286 }
 287
 288 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 289 {
 290         if (ELV_ON_HASH(rq))
 291                 __elv_rqhash_del(rq);
 292 }
 293
 294 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 295 {
 296         elevator_t *e = q->elevator;
 297
 298         BUG_ON(ELV_ON_HASH(rq));
 299         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 300 }
 301
 302 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 303 {
 304         __elv_rqhash_del(rq);
 305         elv_rqhash_add(q, rq);
 306 }
 307
 308 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 309 {
 310         elevator_t *e = q->elevator;
 311         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 312         struct hlist_node *entry, *next;
 313         struct request *rq;
 314
 315         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 316                 BUG_ON(!ELV_ON_HASH(rq));
 317
 318                 if (unlikely(!rq_mergeable(rq))) {
 319                         __elv_rqhash_del(rq);
 320                         continue;
 321                 }
 322
 323                 if (rq_hash_key(rq) == offset)
 324                         return rq;
 325         }
 326
 327         return NULL;
 328 }
 329
 330 /*
 331  * RB-tree support functions for inserting/lookup/removal of requests
 332  * in a sorted RB tree.
 333  */
 334 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 335 {
 336         struct rb_node **p = &root->rb_node;
 337         struct rb_node *parent = NULL;
 338         struct request *__rq;
 339
 340         while (*p) {
 341                 parent = *p;
 342                 __rq = rb_entry(parent, struct request, rb_node);
 343
 344                 if (rq->sector < __rq->sector)
 345                         p = &(*p)->rb_left;
 346                 else if (rq->sector > __rq->sector)
 347                         p = &(*p)->rb_right;
 348                 else
 349                         return __rq;
 350         }
 351
 352         rb_link_node(&rq->rb_node, parent, p);
 353         rb_insert_color(&rq->rb_node, root);
 354         return NULL;
 355 }
 356
 357 EXPORT_SYMBOL(elv_rb_add);
 358
 359 void elv_rb_del(struct rb_root *root, struct request *rq)
 360 {
 361         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 362         rb_erase(&rq->rb_node, root);
 363         RB_CLEAR_NODE(&rq->rb_node);
 364 }
 365
 366 EXPORT_SYMBOL(elv_rb_del);
 367
 368 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 369 {
 370         struct rb_node *n = root->rb_node;
 371         struct request *rq;
 372
 373         while (n) {
 374                 rq = rb_entry(n, struct request, rb_node);
 375
 376                 if (sector < rq->sector)
 377                         n = n->rb_left;
 378                 else if (sector > rq->sector)
 379                         n = n->rb_right;
 380                 else
 381                         return rq;
 382         }
 383
 384         return NULL;
 385 }
 386
 387 EXPORT_SYMBOL(elv_rb_find);
 388
 389 /*
 390  * Insert rq into dispatch queue of q.  Queue lock must be held on
 391  * entry.  rq is sort instead into the dispatch queue. To be used by
 392  * specific elevators.
 393  */
 394 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 395 {
 396         sector_t boundary;
 397         struct list_head *entry;
 398
 399         if (q->last_merge == rq)
 400                 q->last_merge = NULL;
 401
 402         elv_rqhash_del(q, rq);
 403
 404         q->nr_sorted--;
 405
 406         boundary = q->end_sector;
 407
 408         list_for_each_prev(entry, &q->queue_head) {
 409                 struct request *pos = list_entry_rq(entry);
 410
 411                 if (rq_data_dir(rq) != rq_data_dir(pos))
 412                         break;
 413                 if (pos->cmd_flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
 414                         break;
 415                 if (rq->sector >= boundary) {
 416                         if (pos->sector < boundary)
 417                                 continue;
 418                 } else {
 419                         if (pos->sector >= boundary)
 420                                 break;
 421                 }
 422                 if (rq->sector >= pos->sector)
 423                         break;
 424         }
 425
 426         list_add(&rq->queuelist, entry);
 427 }
 428
 429 EXPORT_SYMBOL(elv_dispatch_sort);
 430
 431 /*
 432  * Insert rq into dispatch queue of q.  Queue lock must be held on
 433  * entry.  rq is added to the back of the dispatch queue. To be used by
 434  * specific elevators.
 435  */
 436 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 437 {
 438         if (q->last_merge == rq)
 439                 q->last_merge = NULL;
 440
 441         elv_rqhash_del(q, rq);
 442
 443         q->nr_sorted--;
 444
 445         q->end_sector = rq_end_sector(rq);
 446         q->boundary_rq = rq;
 447         list_add_tail(&rq->queuelist, &q->queue_head);
 448 }
 449
 450 EXPORT_SYMBOL(elv_dispatch_add_tail);
 451
 452 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 453 {
 454         elevator_t *e = q->elevator;
 455         struct request *__rq;
 456         int ret;
 457
 458         /*
 459          * First try one-hit cache.
 460          */
 461         if (q->last_merge) {
 462                 ret = elv_try_merge(q->last_merge, bio);
 463                 if (ret != ELEVATOR_NO_MERGE) {
 464                         *req = q->last_merge;
 465                         return ret;
 466                 }
 467         }
 468
 469         /*
 470          * See if our hash lookup can find a potential backmerge.
 471          */
 472         __rq = elv_rqhash_find(q, bio->bi_sector);
 473         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 474                 *req = __rq;
 475                 return ELEVATOR_BACK_MERGE;
 476         }
 477
 478         if (e->ops->elevator_merge_fn)
 479                 return e->ops->elevator_merge_fn(q, req, bio);
 480
 481         return ELEVATOR_NO_MERGE;
 482 }
 483
 484 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 485 {
 486         elevator_t *e = q->elevator;
 487
 488         if (e->ops->elevator_merged_fn)
 489                 e->ops->elevator_merged_fn(q, rq, type);
 490
 491         if (type == ELEVATOR_BACK_MERGE)
 492                 elv_rqhash_reposition(q, rq);
 493
 494         q->last_merge = rq;
 495 }
 496
 497 void elv_merge_requests(struct request_queue *q, struct request *rq,
 498                              struct request *next)
 499 {
 500         elevator_t *e = q->elevator;
 501
 502         if (e->ops->elevator_merge_req_fn)
 503                 e->ops->elevator_merge_req_fn(q, rq, next);
 504
 505         elv_rqhash_reposition(q, rq);
 506         elv_rqhash_del(q, next);
 507
 508         q->nr_sorted--;
 509         q->last_merge = rq;
 510 }
 511
 512 void elv_requeue_request(struct request_queue *q, struct request *rq)
 513 {
 514         /*
 515          * it already went through dequeue, we need to decrement the
 516          * in_flight count again
 517          */
 518         if (blk_account_rq(rq)) {
 519                 q->in_flight--;
 520                 if (blk_sorted_rq(rq))
 521                         elv_deactivate_rq(q, rq);
 522         }
 523
 524         rq->cmd_flags &= ~REQ_STARTED;
 525
 526         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 527 }
 528
 529 static void elv_drain_elevator(struct request_queue *q)
 530 {
 531         static int printed;
 532         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 533                 ;
 534         if (q->nr_sorted == 0)
 535                 return;
 536         if (printed++ < 10) {
 537                 printk(KERN_ERR "%s: forced dispatching is broken "
 538                        "(nr_sorted=%u), please report this\n",
 539                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 540         }
 541 }
 542
 543 void elv_insert(struct request_queue *q, struct request *rq, int where)
 544 {
 545         struct list_head *pos;
 546         unsigned ordseq;
 547         int unplug_it = 1;
 548
 549         blk_add_trace_rq(q, rq, BLK_TA_INSERT);
 550
 551         rq->q = q;
 552
 553         switch (where) {
 554         case ELEVATOR_INSERT_FRONT:
 555                 rq->cmd_flags |= REQ_SOFTBARRIER;
 556
 557                 list_add(&rq->queuelist, &q->queue_head);
 558                 break;
 559
 560         case ELEVATOR_INSERT_BACK:
 561                 rq->cmd_flags |= REQ_SOFTBARRIER;
 562                 elv_drain_elevator(q);
 563                 list_add_tail(&rq->queuelist, &q->queue_head);
 564                 /*
 565                  * We kick the queue here for the following reasons.
 566                  * - The elevator might have returned NULL previously
 567                  *   to delay requests and returned them now.  As the
 568                  *   queue wasn't empty before this request, ll_rw_blk
 569                  *   won't run the queue on return, resulting in hang.
 570                  * - Usually, back inserted requests won't be merged
 571                  *   with anything.  There's no point in delaying queue
 572                  *   processing.
 573                  */
 574                 blk_remove_plug(q);
 575                 q->request_fn(q);
 576                 break;
 577
 578         case ELEVATOR_INSERT_SORT:
 579                 BUG_ON(!blk_fs_request(rq));
 580                 rq->cmd_flags |= REQ_SORTED;
 581                 q->nr_sorted++;
 582                 if (rq_mergeable(rq)) {
 583                         elv_rqhash_add(q, rq);
 584                         if (!q->last_merge)
 585                                 q->last_merge = rq;
 586                 }
 587
 588                 /*
 589                  * Some ioscheds (cfq) run q->request_fn directly, so
 590                  * rq cannot be accessed after calling
 591                  * elevator_add_req_fn.
 592                  */
 593                 q->elevator->ops->elevator_add_req_fn(q, rq);
 594                 break;
 595
 596         case ELEVATOR_INSERT_REQUEUE:
 597                 /*
 598                  * If ordered flush isn't in progress, we do front
 599                  * insertion; otherwise, requests should be requeued
 600                  * in ordseq order.
 601                  */
 602                 rq->cmd_flags |= REQ_SOFTBARRIER;
 603
 604                 /*
 605                  * Most requeues happen because of a busy condition,
 606                  * don't force unplug of the queue for that case.
 607                  */
 608                 unplug_it = 0;
 609
 610                 if (q->ordseq == 0) {
 611                         list_add(&rq->queuelist, &q->queue_head);
 612                         break;
 613                 }
 614
 615                 ordseq = blk_ordered_req_seq(rq);
 616
 617                 list_for_each(pos, &q->queue_head) {
 618                         struct request *pos_rq = list_entry_rq(pos);
 619                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 620                                 break;
 621                 }
 622
 623                 list_add_tail(&rq->queuelist, pos);
 624                 break;
 625
 626         default:
 627                 printk(KERN_ERR "%s: bad insertion point %d\n",
 628                        __FUNCTION__, where);
 629                 BUG();
 630         }
 631
 632         if (unplug_it && blk_queue_plugged(q)) {
 633                 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
 634                         - q->in_flight;
 635
 636                 if (nrq >= q->unplug_thresh)
 637                         __generic_unplug_device(q);
 638         }
 639 }
 640
 641 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
 642                        int plug)
 643 {
 644         if (q->ordcolor)
 645                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 646
 647         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 648                 /*
 649                  * toggle ordered color
 650                  */
 651                 if (blk_barrier_rq(rq))
 652                         q->ordcolor ^= 1;
 653
 654                 /*
 655                  * barriers implicitly indicate back insertion
 656                  */
 657                 if (where == ELEVATOR_INSERT_SORT)
 658                         where = ELEVATOR_INSERT_BACK;
 659
 660                 /*
 661                  * this request is scheduling boundary, update
 662                  * end_sector
 663                  */
 664                 if (blk_fs_request(rq)) {
 665                         q->end_sector = rq_end_sector(rq);
 666                         q->boundary_rq = rq;
 667                 }
 668         } else if (!(rq->cmd_flags & REQ_ELVPRIV) && where == ELEVATOR_INSERT_SORT)
 669                 where = ELEVATOR_INSERT_BACK;
 670
 671         if (plug)
 672                 blk_plug_device(q);
 673
 674         elv_insert(q, rq, where);
 675 }
 676
 677 EXPORT_SYMBOL(__elv_add_request);
 678
 679 void elv_add_request(struct request_queue *q, struct request *rq, int where,
 680                      int plug)
 681 {
 682         unsigned long flags;
 683
 684         spin_lock_irqsave(q->queue_lock, flags);
 685         __elv_add_request(q, rq, where, plug);
 686         spin_unlock_irqrestore(q->queue_lock, flags);
 687 }
 688
 689 EXPORT_SYMBOL(elv_add_request);
 690
 691 static inline struct request *__elv_next_request(struct request_queue *q)
 692 {
 693         struct request *rq;
 694
 695         while (1) {
 696                 while (!list_empty(&q->queue_head)) {
 697                         rq = list_entry_rq(q->queue_head.next);
 698                         if (blk_do_ordered(q, &rq))
 699                                 return rq;
 700                 }
 701
 702                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
 703                         return NULL;
 704         }
 705 }
 706
 707 struct request *elv_next_request(struct request_queue *q)
 708 {
 709         struct request *rq;
 710         int ret;
 711
 712         while ((rq = __elv_next_request(q)) != NULL) {
 713                 /*
 714                  * Kill the empty barrier place holder, the driver must
 715                  * not ever see it.
 716                  */
 717                 if (blk_empty_barrier(rq)) {
 718                         end_queued_request(rq, 1);
 719                         continue;
 720                 }
 721                 if (!(rq->cmd_flags & REQ_STARTED)) {
 722                         /*
 723                          * This is the first time the device driver
 724                          * sees this request (possibly after
 725                          * requeueing).  Notify IO scheduler.
 726                          */
 727                         if (blk_sorted_rq(rq))
 728                                 elv_activate_rq(q, rq);
 729
 730                         /*
 731                          * just mark as started even if we don't start
 732                          * it, a request that has been delayed should
 733                          * not be passed by new incoming requests
 734                          */
 735                         rq->cmd_flags |= REQ_STARTED;
 736                         blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
 737                 }
 738
 739                 if (!q->boundary_rq || q->boundary_rq == rq) {
 740                         q->end_sector = rq_end_sector(rq);
 741                         q->boundary_rq = NULL;
 742                 }
 743
 744                 if (rq->cmd_flags & REQ_DONTPREP)
 745                         break;
 746
 747                 if (q->dma_drain_size && rq->data_len) {
 748                         /*
 749                          * make sure space for the drain appears we
 750                          * know we can do this because max_hw_segments
 751                          * has been adjusted to be one fewer than the
 752                          * device can handle
 753                          */
 754                         rq->nr_phys_segments++;
 755                         rq->nr_hw_segments++;
 756                 }
 757
 758                 if (!q->prep_rq_fn)
 759                         break;
 760
 761                 ret = q->prep_rq_fn(q, rq);
 762                 if (ret == BLKPREP_OK) {
 763                         break;
 764                 } else if (ret == BLKPREP_DEFER) {
 765                         /*
 766                          * the request may have been (partially) prepped.
 767                          * we need to keep this request in the front to
 768                          * avoid resource deadlock.  REQ_STARTED will
 769                          * prevent other fs requests from passing this one.
 770                          */
 771                         if (q->dma_drain_size && rq->data_len &&
 772                             !(rq->cmd_flags & REQ_DONTPREP)) {
 773                                 /*
 774                                  * remove the space for the drain we added
 775                                  * so that we don't add it again
 776                                  */
 777                                 --rq->nr_phys_segments;
 778                                 --rq->nr_hw_segments;
 779                         }
 780
 781                         rq = NULL;
 782                         break;
 783                 } else if (ret == BLKPREP_KILL) {
 784                         rq->cmd_flags |= REQ_QUIET;
 785                         end_queued_request(rq, 0);
 786                 } else {
 787                         printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
 788                                                                 ret);
 789                         break;
 790                 }
 791         }
 792
 793         return rq;
 794 }
 795
 796 EXPORT_SYMBOL(elv_next_request);
 797
 798 void elv_dequeue_request(struct request_queue *q, struct request *rq)
 799 {
 800         BUG_ON(list_empty(&rq->queuelist));
 801         BUG_ON(ELV_ON_HASH(rq));
 802
 803         list_del_init(&rq->queuelist);
 804
 805         /*
 806          * the time frame between a request being removed from the lists
 807          * and to it is freed is accounted as io that is in progress at
 808          * the driver side.
 809          */
 810         if (blk_account_rq(rq))
 811                 q->in_flight++;
 812 }
 813
 814 EXPORT_SYMBOL(elv_dequeue_request);
 815
 816 int elv_queue_empty(struct request_queue *q)
 817 {
 818         elevator_t *e = q->elevator;
 819
 820         if (!list_empty(&q->queue_head))
 821                 return 0;
 822
 823         if (e->ops->elevator_queue_empty_fn)
 824                 return e->ops->elevator_queue_empty_fn(q);
 825
 826         return 1;
 827 }
 828
 829 EXPORT_SYMBOL(elv_queue_empty);
 830
 831 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 832 {
 833         elevator_t *e = q->elevator;
 834
 835         if (e->ops->elevator_latter_req_fn)
 836                 return e->ops->elevator_latter_req_fn(q, rq);
 837         return NULL;
 838 }
 839
 840 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 841 {
 842         elevator_t *e = q->elevator;
 843
 844         if (e->ops->elevator_former_req_fn)
 845                 return e->ops->elevator_former_req_fn(q, rq);
 846         return NULL;
 847 }
 848
 849 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 850 {
 851         elevator_t *e = q->elevator;
 852
 853         if (e->ops->elevator_set_req_fn)
 854                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 855
 856         rq->elevator_private = NULL;
 857         return 0;
 858 }
 859
 860 void elv_put_request(struct request_queue *q, struct request *rq)
 861 {
 862         elevator_t *e = q->elevator;
 863
 864         if (e->ops->elevator_put_req_fn)
 865                 e->ops->elevator_put_req_fn(rq);
 866 }
 867
 868 int elv_may_queue(struct request_queue *q, int rw)
 869 {
 870         elevator_t *e = q->elevator;
 871
 872         if (e->ops->elevator_may_queue_fn)
 873                 return e->ops->elevator_may_queue_fn(q, rw);
 874
 875         return ELV_MQUEUE_MAY;
 876 }
 877
 878 void elv_completed_request(struct request_queue *q, struct request *rq)
 879 {
 880         elevator_t *e = q->elevator;
 881
 882         /*
 883          * request is released from the driver, io must be done
 884          */
 885         if (blk_account_rq(rq)) {
 886                 q->in_flight--;
 887                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 888                         e->ops->elevator_completed_req_fn(q, rq);
 889         }
 890
 891         /*
 892          * Check if the queue is waiting for fs requests to be
 893          * drained for flush sequence.
 894          */
 895         if (unlikely(q->ordseq)) {
 896                 struct request *first_rq = list_entry_rq(q->queue_head.next);
 897                 if (q->in_flight == 0 &&
 898                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 899                     blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
 900                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 901                         q->request_fn(q);
 902                 }
 903         }
 904 }
 905
 906 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 907
 908 static ssize_t
 909 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 910 {
 911         elevator_t *e = container_of(kobj, elevator_t, kobj);
 912         struct elv_fs_entry *entry = to_elv(attr);
 913         ssize_t error;
 914
 915         if (!entry->show)
 916                 return -EIO;
 917
 918         mutex_lock(&e->sysfs_lock);
 919         error = e->ops ? entry->show(e, page) : -ENOENT;
 920         mutex_unlock(&e->sysfs_lock);
 921         return error;
 922 }
 923
 924 static ssize_t
 925 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 926                const char *page, size_t length)
 927 {
 928         elevator_t *e = container_of(kobj, elevator_t, kobj);
 929         struct elv_fs_entry *entry = to_elv(attr);
 930         ssize_t error;
 931
 932         if (!entry->store)
 933                 return -EIO;
 934
 935         mutex_lock(&e->sysfs_lock);
 936         error = e->ops ? entry->store(e, page, length) : -ENOENT;
 937         mutex_unlock(&e->sysfs_lock);
 938         return error;
 939 }
 940
 941 static struct sysfs_ops elv_sysfs_ops = {
 942         .show   = elv_attr_show,
 943         .store  = elv_attr_store,
 944 };
 945
 946 static struct kobj_type elv_ktype = {
 947         .sysfs_ops      = &elv_sysfs_ops,
 948         .release        = elevator_release,
 949 };
 950
 951 int elv_register_queue(struct request_queue *q)
 952 {
 953         elevator_t *e = q->elevator;
 954         int error;
 955
 956         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 957         if (!error) {
 958                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
 959                 if (attr) {
 960                         while (attr->attr.name) {
 961                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 962                                         break;
 963                                 attr++;
 964                         }
 965                 }
 966                 kobject_uevent(&e->kobj, KOBJ_ADD);
 967         }
 968         return error;
 969 }
 970
 971 static void __elv_unregister_queue(elevator_t *e)
 972 {
 973         kobject_uevent(&e->kobj, KOBJ_REMOVE);
 974         kobject_del(&e->kobj);
 975 }
 976
 977 void elv_unregister_queue(struct request_queue *q)
 978 {
 979         if (q)
 980                 __elv_unregister_queue(q->elevator);
 981 }
 982
 983 void elv_register(struct elevator_type *e)
 984 {
 985         char *def = "";
 986
 987         spin_lock(&elv_list_lock);
 988         BUG_ON(elevator_find(e->elevator_name));
 989         list_add_tail(&e->list, &elv_list);
 990         spin_unlock(&elv_list_lock);
 991
 992         if (!strcmp(e->elevator_name, chosen_elevator) ||
 993                         (!*chosen_elevator &&
 994                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 995                                 def = " (default)";
 996
 997         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name, def);
 998 }
 999 EXPORT_SYMBOL_GPL(elv_register);
1000
1001 void elv_unregister(struct elevator_type *e)
1002 {
1003         struct task_struct *g, *p;
1004
1005         /*
1006          * Iterate every thread in the process to remove the io contexts.
1007          */
1008         if (e->ops.trim) {
1009                 read_lock(&tasklist_lock);
1010                 do_each_thread(g, p) {
1011                         task_lock(p);
1012                         if (p->io_context)
1013                                 e->ops.trim(p->io_context);
1014                         task_unlock(p);
1015                 } while_each_thread(g, p);
1016                 read_unlock(&tasklist_lock);
1017         }
1018
1019         spin_lock(&elv_list_lock);
1020         list_del_init(&e->list);
1021         spin_unlock(&elv_list_lock);
1022 }
1023 EXPORT_SYMBOL_GPL(elv_unregister);
1024
1025 /*
1026  * switch to new_e io scheduler. be careful not to introduce deadlocks -
1027  * we don't free the old io scheduler, before we have allocated what we
1028  * need for the new one. this way we have a chance of going back to the old
1029  * one, if the new one fails init for some reason.
1030  */
1031 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1032 {
1033         elevator_t *old_elevator, *e;
1034         void *data;
1035
1036         /*
1037          * Allocate new elevator
1038          */
1039         e = elevator_alloc(q, new_e);
1040         if (!e)
1041                 return 0;
1042
1043         data = elevator_init_queue(q, e);
1044         if (!data) {
1045                 kobject_put(&e->kobj);
1046                 return 0;
1047         }
1048
1049         /*
1050          * Turn on BYPASS and drain all requests w/ elevator private data
1051          */
1052         spin_lock_irq(q->queue_lock);
1053
1054         set_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1055
1056         elv_drain_elevator(q);
1057
1058         while (q->rq.elvpriv) {
1059                 blk_remove_plug(q);
1060                 q->request_fn(q);
1061                 spin_unlock_irq(q->queue_lock);
1062                 msleep(10);
1063                 spin_lock_irq(q->queue_lock);
1064                 elv_drain_elevator(q);
1065         }
1066
1067         /*
1068          * Remember old elevator.
1069          */
1070         old_elevator = q->elevator;
1071
1072         /*
1073          * attach and start new elevator
1074          */
1075         elevator_attach(q, e, data);
1076
1077         spin_unlock_irq(q->queue_lock);
1078
1079         __elv_unregister_queue(old_elevator);
1080
1081         if (elv_register_queue(q))
1082                 goto fail_register;
1083
1084         /*
1085          * finally exit old elevator and turn off BYPASS.
1086          */
1087         elevator_exit(old_elevator);
1088         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1089         return 1;
1090
1091 fail_register:
1092         /*
1093          * switch failed, exit the new io scheduler and reattach the old
1094          * one again (along with re-adding the sysfs dir)
1095          */
1096         elevator_exit(e);
1097         q->elevator = old_elevator;
1098         elv_register_queue(q);
1099         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1100         return 0;
1101 }
1102
1103 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1104                           size_t count)
1105 {
1106         char elevator_name[ELV_NAME_MAX];
1107         size_t len;
1108         struct elevator_type *e;
1109
1110         elevator_name[sizeof(elevator_name) - 1] = '\0';
1111         strncpy(elevator_name, name, sizeof(elevator_name) - 1);
1112         len = strlen(elevator_name);
1113
1114         if (len && elevator_name[len - 1] == '\n')
1115                 elevator_name[len - 1] = '\0';
1116
1117         e = elevator_get(elevator_name);
1118         if (!e) {
1119                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1120                 return -EINVAL;
1121         }
1122
1123         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1124                 elevator_put(e);
1125                 return count;
1126         }
1127
1128         if (!elevator_switch(q, e))
1129                 printk(KERN_ERR "elevator: switch to %s failed\n",elevator_name);
1130         return count;
1131 }
1132
1133 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1134 {
1135         elevator_t *e = q->elevator;
1136         struct elevator_type *elv = e->elevator_type;
1137         struct elevator_type *__e;
1138         int len = 0;
1139
1140         spin_lock(&elv_list_lock);
1141         list_for_each_entry(__e, &elv_list, list) {
1142                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1143                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1144                 else
1145                         len += sprintf(name+len, "%s ", __e->elevator_name);
1146         }
1147         spin_unlock(&elv_list_lock);
1148
1149         len += sprintf(len+name, "\n");
1150         return len;
1151 }
1152
1153 struct request *elv_rb_former_request(struct request_queue *q,
1154                                       struct request *rq)
1155 {
1156         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1157
1158         if (rbprev)
1159                 return rb_entry_rq(rbprev);
1160
1161         return NULL;
1162 }
1163
1164 EXPORT_SYMBOL(elv_rb_former_request);
1165
1166 struct request *elv_rb_latter_request(struct request_queue *q,
1167                                       struct request *rq)
1168 {
1169         struct rb_node *rbnext = rb_next(&rq->rb_node);
1170
1171         if (rbnext)
1172                 return rb_entry_rq(rbnext);
1173
1174         return NULL;
1175 }
1176
1177 EXPORT_SYMBOL(elv_rb_latter_request);