SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 static DEFINE_SPINLOCK(elv_list_lock);
  40 static LIST_HEAD(elv_list);
  41
  42 /*
  43  * Merge hash stuff.
  44  */
  45 static const int elv_hash_shift = 6;
  46 #define ELV_HASH_BLOCK(sec)     ((sec) >> 3)
  47 #define ELV_HASH_FN(sec)        \
  48                 (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 int elv_rq_merge_ok(struct request *rq, struct bio *bio)
  72 {
  73         if (!rq_mergeable(rq))
  74                 return 0;
  75
  76         /*
  77          * Don't merge file system requests and discard requests
  78          */
  79         if (bio_discard(bio) != bio_discard(rq->bio))
  80                 return 0;
  81
  82         /*
  83          * different data direction or already started, don't merge
  84          */
  85         if (bio_data_dir(bio) != rq_data_dir(rq))
  86                 return 0;
  87
  88         /*
  89          * must be same device and not a special request
  90          */
  91         if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
  92                 return 0;
  93
  94         /*
  95          * only merge integrity protected bio into ditto rq
  96          */
  97         if (bio_integrity(bio) != blk_integrity_rq(rq))
  98                 return 0;
  99
 100         if (!elv_iosched_allow_merge(rq, bio))
 101                 return 0;
 102
 103         return 1;
 104 }
 105 EXPORT_SYMBOL(elv_rq_merge_ok);
 106
 107 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
 108 {
 109         int ret = ELEVATOR_NO_MERGE;
 110
 111         /*
 112          * we can merge and sequence is ok, check if it's possible
 113          */
 114         if (elv_rq_merge_ok(__rq, bio)) {
 115                 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
 116                         ret = ELEVATOR_BACK_MERGE;
 117                 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
 118                         ret = ELEVATOR_FRONT_MERGE;
 119         }
 120
 121         return ret;
 122 }
 123
 124 static struct elevator_type *elevator_find(const char *name)
 125 {
 126         struct elevator_type *e;
 127
 128         list_for_each_entry(e, &elv_list, list) {
 129                 if (!strcmp(e->elevator_name, name))
 130                         return e;
 131         }
 132
 133         return NULL;
 134 }
 135
 136 static void elevator_put(struct elevator_type *e)
 137 {
 138         module_put(e->elevator_owner);
 139 }
 140
 141 static struct elevator_type *elevator_get(const char *name)
 142 {
 143         struct elevator_type *e;
 144
 145         spin_lock(&elv_list_lock);
 146
 147         e = elevator_find(name);
 148         if (!e) {
 149                 char elv[ELV_NAME_MAX + strlen("-iosched")];
 150
 151                 spin_unlock(&elv_list_lock);
 152
 153                 if (!strcmp(name, "anticipatory"))
 154                         sprintf(elv, "as-iosched");
 155                 else
 156                         sprintf(elv, "%s-iosched", name);
 157
 158                 request_module("%s", elv);
 159                 spin_lock(&elv_list_lock);
 160                 e = elevator_find(name);
 161         }
 162
 163         if (e && !try_module_get(e->elevator_owner))
 164                 e = NULL;
 165
 166         spin_unlock(&elv_list_lock);
 167
 168         return e;
 169 }
 170
 171 static void *elevator_init_queue(struct request_queue *q,
 172                                  struct elevator_queue *eq)
 173 {
 174         return eq->ops->elevator_init_fn(q);
 175 }
 176
 177 static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
 178                            void *data)
 179 {
 180         q->elevator = eq;
 181         eq->elevator_data = data;
 182 }
 183
 184 static char chosen_elevator[16];
 185
 186 static int __init elevator_setup(char *str)
 187 {
 188         /*
 189          * Be backwards-compatible with previous kernels, so users
 190          * won't get the wrong elevator.
 191          */
 192         if (!strcmp(str, "as"))
 193                 strcpy(chosen_elevator, "anticipatory");
 194         else
 195                 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 196         return 1;
 197 }
 198
 199 __setup("elevator=", elevator_setup);
 200
 201 static struct kobj_type elv_ktype;
 202
 203 static elevator_t *elevator_alloc(struct request_queue *q,
 204                                   struct elevator_type *e)
 205 {
 206         elevator_t *eq;
 207         int i;
 208
 209         eq = kmalloc_node(sizeof(elevator_t), GFP_KERNEL | __GFP_ZERO, q->node);
 210         if (unlikely(!eq))
 211                 goto err;
 212
 213         eq->ops = &e->ops;
 214         eq->elevator_type = e;
 215         kobject_init(&eq->kobj, &elv_ktype);
 216         mutex_init(&eq->sysfs_lock);
 217
 218         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 219                                         GFP_KERNEL, q->node);
 220         if (!eq->hash)
 221                 goto err;
 222
 223         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 224                 INIT_HLIST_HEAD(&eq->hash[i]);
 225
 226         return eq;
 227 err:
 228         kfree(eq);
 229         elevator_put(e);
 230         return NULL;
 231 }
 232
 233 static void elevator_release(struct kobject *kobj)
 234 {
 235         elevator_t *e = container_of(kobj, elevator_t, kobj);
 236
 237         elevator_put(e->elevator_type);
 238         kfree(e->hash);
 239         kfree(e);
 240 }
 241
 242 int elevator_init(struct request_queue *q, char *name)
 243 {
 244         struct elevator_type *e = NULL;
 245         struct elevator_queue *eq;
 246         int ret = 0;
 247         void *data;
 248
 249         INIT_LIST_HEAD(&q->queue_head);
 250         q->last_merge = NULL;
 251         q->end_sector = 0;
 252         q->boundary_rq = NULL;
 253
 254         if (name) {
 255                 e = elevator_get(name);
 256                 if (!e)
 257                         return -EINVAL;
 258         }
 259
 260         if (!e && *chosen_elevator) {
 261                 e = elevator_get(chosen_elevator);
 262                 if (!e)
 263                         printk(KERN_ERR "I/O scheduler %s not found\n",
 264                                                         chosen_elevator);
 265         }
 266
 267         if (!e) {
 268                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
 269                 if (!e) {
 270                         printk(KERN_ERR
 271                                 "Default I/O scheduler not found. " \
 272                                 "Using noop.\n");
 273                         e = elevator_get("noop");
 274                 }
 275         }
 276
 277         eq = elevator_alloc(q, e);
 278         if (!eq)
 279                 return -ENOMEM;
 280
 281         data = elevator_init_queue(q, eq);
 282         if (!data) {
 283                 kobject_put(&eq->kobj);
 284                 return -ENOMEM;
 285         }
 286
 287         elevator_attach(q, eq, data);
 288         return ret;
 289 }
 290 EXPORT_SYMBOL(elevator_init);
 291
 292 void elevator_exit(elevator_t *e)
 293 {
 294         mutex_lock(&e->sysfs_lock);
 295         if (e->ops->elevator_exit_fn)
 296                 e->ops->elevator_exit_fn(e);
 297         e->ops = NULL;
 298         mutex_unlock(&e->sysfs_lock);
 299
 300         kobject_put(&e->kobj);
 301 }
 302 EXPORT_SYMBOL(elevator_exit);
 303
 304 static void elv_activate_rq(struct request_queue *q, struct request *rq)
 305 {
 306         elevator_t *e = q->elevator;
 307
 308         if (e->ops->elevator_activate_req_fn)
 309                 e->ops->elevator_activate_req_fn(q, rq);
 310 }
 311
 312 static void elv_deactivate_rq(struct request_queue *q, struct request *rq)
 313 {
 314         elevator_t *e = q->elevator;
 315
 316         if (e->ops->elevator_deactivate_req_fn)
 317                 e->ops->elevator_deactivate_req_fn(q, rq);
 318 }
 319
 320 static inline void __elv_rqhash_del(struct request *rq)
 321 {
 322         hlist_del_init(&rq->hash);
 323 }
 324
 325 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 326 {
 327         if (ELV_ON_HASH(rq))
 328                 __elv_rqhash_del(rq);
 329 }
 330
 331 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 332 {
 333         elevator_t *e = q->elevator;
 334
 335         BUG_ON(ELV_ON_HASH(rq));
 336         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 337 }
 338
 339 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 340 {
 341         __elv_rqhash_del(rq);
 342         elv_rqhash_add(q, rq);
 343 }
 344
 345 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 346 {
 347         elevator_t *e = q->elevator;
 348         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 349         struct hlist_node *entry, *next;
 350         struct request *rq;
 351
 352         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 353                 BUG_ON(!ELV_ON_HASH(rq));
 354
 355                 if (unlikely(!rq_mergeable(rq))) {
 356                         __elv_rqhash_del(rq);
 357                         continue;
 358                 }
 359
 360                 if (rq_hash_key(rq) == offset)
 361                         return rq;
 362         }
 363
 364         return NULL;
 365 }
 366
 367 /*
 368  * RB-tree support functions for inserting/lookup/removal of requests
 369  * in a sorted RB tree.
 370  */
 371 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 372 {
 373         struct rb_node **p = &root->rb_node;
 374         struct rb_node *parent = NULL;
 375         struct request *__rq;
 376
 377         while (*p) {
 378                 parent = *p;
 379                 __rq = rb_entry(parent, struct request, rb_node);
 380
 381                 if (rq->sector < __rq->sector)
 382                         p = &(*p)->rb_left;
 383                 else if (rq->sector > __rq->sector)
 384                         p = &(*p)->rb_right;
 385                 else
 386                         return __rq;
 387         }
 388
 389         rb_link_node(&rq->rb_node, parent, p);
 390         rb_insert_color(&rq->rb_node, root);
 391         return NULL;
 392 }
 393 EXPORT_SYMBOL(elv_rb_add);
 394
 395 void elv_rb_del(struct rb_root *root, struct request *rq)
 396 {
 397         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 398         rb_erase(&rq->rb_node, root);
 399         RB_CLEAR_NODE(&rq->rb_node);
 400 }
 401 EXPORT_SYMBOL(elv_rb_del);
 402
 403 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 404 {
 405         struct rb_node *n = root->rb_node;
 406         struct request *rq;
 407
 408         while (n) {
 409                 rq = rb_entry(n, struct request, rb_node);
 410
 411                 if (sector < rq->sector)
 412                         n = n->rb_left;
 413                 else if (sector > rq->sector)
 414                         n = n->rb_right;
 415                 else
 416                         return rq;
 417         }
 418
 419         return NULL;
 420 }
 421 EXPORT_SYMBOL(elv_rb_find);
 422
 423 /*
 424  * Insert rq into dispatch queue of q.  Queue lock must be held on
 425  * entry.  rq is sort instead into the dispatch queue. To be used by
 426  * specific elevators.
 427  */
 428 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 429 {
 430         sector_t boundary;
 431         struct list_head *entry;
 432         int stop_flags;
 433
 434         if (q->last_merge == rq)
 435                 q->last_merge = NULL;
 436
 437         elv_rqhash_del(q, rq);
 438
 439         q->nr_sorted--;
 440
 441         boundary = q->end_sector;
 442         stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
 443         list_for_each_prev(entry, &q->queue_head) {
 444                 struct request *pos = list_entry_rq(entry);
 445
 446                 if (blk_discard_rq(rq) != blk_discard_rq(pos))
 447                         break;
 448                 if (rq_data_dir(rq) != rq_data_dir(pos))
 449                         break;
 450                 if (pos->cmd_flags & stop_flags)
 451                         break;
 452                 if (rq->sector >= boundary) {
 453                         if (pos->sector < boundary)
 454                                 continue;
 455                 } else {
 456                         if (pos->sector >= boundary)
 457                                 break;
 458                 }
 459                 if (rq->sector >= pos->sector)
 460                         break;
 461         }
 462
 463         list_add(&rq->queuelist, entry);
 464 }
 465 EXPORT_SYMBOL(elv_dispatch_sort);
 466
 467 /*
 468  * Insert rq into dispatch queue of q.  Queue lock must be held on
 469  * entry.  rq is added to the back of the dispatch queue. To be used by
 470  * specific elevators.
 471  */
 472 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 473 {
 474         if (q->last_merge == rq)
 475                 q->last_merge = NULL;
 476
 477         elv_rqhash_del(q, rq);
 478
 479         q->nr_sorted--;
 480
 481         q->end_sector = rq_end_sector(rq);
 482         q->boundary_rq = rq;
 483         list_add_tail(&rq->queuelist, &q->queue_head);
 484 }
 485 EXPORT_SYMBOL(elv_dispatch_add_tail);
 486
 487 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 488 {
 489         elevator_t *e = q->elevator;
 490         struct request *__rq;
 491         int ret;
 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_nomerges(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         elevator_t *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         elevator_t *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_requeue_request(struct request_queue *q, struct request *rq)
 551 {
 552         /*
 553          * it already went through dequeue, we need to decrement the
 554          * in_flight count again
 555          */
 556         if (blk_account_rq(rq)) {
 557                 q->in_flight--;
 558                 if (blk_sorted_rq(rq))
 559                         elv_deactivate_rq(q, rq);
 560         }
 561
 562         rq->cmd_flags &= ~REQ_STARTED;
 563
 564         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 565 }
 566
 567 static void elv_drain_elevator(struct request_queue *q)
 568 {
 569         static int printed;
 570         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 571                 ;
 572         if (q->nr_sorted == 0)
 573                 return;
 574         if (printed++ < 10) {
 575                 printk(KERN_ERR "%s: forced dispatching is broken "
 576                        "(nr_sorted=%u), please report this\n",
 577                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 578         }
 579 }
 580
 581 void elv_insert(struct request_queue *q, struct request *rq, int where)
 582 {
 583         struct list_head *pos;
 584         unsigned ordseq;
 585         int unplug_it = 1;
 586
 587         blk_add_trace_rq(q, rq, BLK_TA_INSERT);
 588
 589         rq->q = q;
 590
 591         switch (where) {
 592         case ELEVATOR_INSERT_FRONT:
 593                 rq->cmd_flags |= REQ_SOFTBARRIER;
 594
 595                 list_add(&rq->queuelist, &q->queue_head);
 596                 break;
 597
 598         case ELEVATOR_INSERT_BACK:
 599                 rq->cmd_flags |= REQ_SOFTBARRIER;
 600                 elv_drain_elevator(q);
 601                 list_add_tail(&rq->queuelist, &q->queue_head);
 602                 /*
 603                  * We kick the queue here for the following reasons.
 604                  * - The elevator might have returned NULL previously
 605                  *   to delay requests and returned them now.  As the
 606                  *   queue wasn't empty before this request, ll_rw_blk
 607                  *   won't run the queue on return, resulting in hang.
 608                  * - Usually, back inserted requests won't be merged
 609                  *   with anything.  There's no point in delaying queue
 610                  *   processing.
 611                  */
 612                 blk_remove_plug(q);
 613                 q->request_fn(q);
 614                 break;
 615
 616         case ELEVATOR_INSERT_SORT:
 617                 BUG_ON(!blk_fs_request(rq) && !blk_discard_rq(rq));
 618                 rq->cmd_flags |= REQ_SORTED;
 619                 q->nr_sorted++;
 620                 if (rq_mergeable(rq)) {
 621                         elv_rqhash_add(q, rq);
 622                         if (!q->last_merge)
 623                                 q->last_merge = rq;
 624                 }
 625
 626                 /*
 627                  * Some ioscheds (cfq) run q->request_fn directly, so
 628                  * rq cannot be accessed after calling
 629                  * elevator_add_req_fn.
 630                  */
 631                 q->elevator->ops->elevator_add_req_fn(q, rq);
 632                 break;
 633
 634         case ELEVATOR_INSERT_REQUEUE:
 635                 /*
 636                  * If ordered flush isn't in progress, we do front
 637                  * insertion; otherwise, requests should be requeued
 638                  * in ordseq order.
 639                  */
 640                 rq->cmd_flags |= REQ_SOFTBARRIER;
 641
 642                 /*
 643                  * Most requeues happen because of a busy condition,
 644                  * don't force unplug of the queue for that case.
 645                  */
 646                 unplug_it = 0;
 647
 648                 if (q->ordseq == 0) {
 649                         list_add(&rq->queuelist, &q->queue_head);
 650                         break;
 651                 }
 652
 653                 ordseq = blk_ordered_req_seq(rq);
 654
 655                 list_for_each(pos, &q->queue_head) {
 656                         struct request *pos_rq = list_entry_rq(pos);
 657                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 658                                 break;
 659                 }
 660
 661                 list_add_tail(&rq->queuelist, pos);
 662                 break;
 663
 664         default:
 665                 printk(KERN_ERR "%s: bad insertion point %d\n",
 666                        __func__, where);
 667                 BUG();
 668         }
 669
 670         if (unplug_it && blk_queue_plugged(q)) {
 671                 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
 672                         - q->in_flight;
 673
 674                 if (nrq >= q->unplug_thresh)
 675                         __generic_unplug_device(q);
 676         }
 677 }
 678
 679 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
 680                        int plug)
 681 {
 682         if (q->ordcolor)
 683                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 684
 685         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 686                 /*
 687                  * toggle ordered color
 688                  */
 689                 if (blk_barrier_rq(rq))
 690                         q->ordcolor ^= 1;
 691
 692                 /*
 693                  * barriers implicitly indicate back insertion
 694                  */
 695                 if (where == ELEVATOR_INSERT_SORT)
 696                         where = ELEVATOR_INSERT_BACK;
 697
 698                 /*
 699                  * this request is scheduling boundary, update
 700                  * end_sector
 701                  */
 702                 if (blk_fs_request(rq) || blk_discard_rq(rq)) {
 703                         q->end_sector = rq_end_sector(rq);
 704                         q->boundary_rq = rq;
 705                 }
 706         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 707                     where == ELEVATOR_INSERT_SORT)
 708                 where = ELEVATOR_INSERT_BACK;
 709
 710         if (plug)
 711                 blk_plug_device(q);
 712
 713         elv_insert(q, rq, where);
 714 }
 715 EXPORT_SYMBOL(__elv_add_request);
 716
 717 void elv_add_request(struct request_queue *q, struct request *rq, int where,
 718                      int plug)
 719 {
 720         unsigned long flags;
 721
 722         spin_lock_irqsave(q->queue_lock, flags);
 723         __elv_add_request(q, rq, where, plug);
 724         spin_unlock_irqrestore(q->queue_lock, flags);
 725 }
 726 EXPORT_SYMBOL(elv_add_request);
 727
 728 static inline struct request *__elv_next_request(struct request_queue *q)
 729 {
 730         struct request *rq;
 731
 732         while (1) {
 733                 while (!list_empty(&q->queue_head)) {
 734                         rq = list_entry_rq(q->queue_head.next);
 735                         if (blk_do_ordered(q, &rq))
 736                                 return rq;
 737                 }
 738
 739                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
 740                         return NULL;
 741         }
 742 }
 743
 744 struct request *elv_next_request(struct request_queue *q)
 745 {
 746         struct request *rq;
 747         int ret;
 748
 749         while ((rq = __elv_next_request(q)) != NULL) {
 750                 /*
 751                  * Kill the empty barrier place holder, the driver must
 752                  * not ever see it.
 753                  */
 754                 if (blk_empty_barrier(rq)) {
 755                         end_queued_request(rq, 1);
 756                         continue;
 757                 }
 758                 if (!(rq->cmd_flags & REQ_STARTED)) {
 759                         /*
 760                          * This is the first time the device driver
 761                          * sees this request (possibly after
 762                          * requeueing).  Notify IO scheduler.
 763                          */
 764                         if (blk_sorted_rq(rq))
 765                                 elv_activate_rq(q, rq);
 766
 767                         /*
 768                          * just mark as started even if we don't start
 769                          * it, a request that has been delayed should
 770                          * not be passed by new incoming requests
 771                          */
 772                         rq->cmd_flags |= REQ_STARTED;
 773                         blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
 774                 }
 775
 776                 if (!q->boundary_rq || q->boundary_rq == rq) {
 777                         q->end_sector = rq_end_sector(rq);
 778                         q->boundary_rq = NULL;
 779                 }
 780
 781                 if (rq->cmd_flags & REQ_DONTPREP)
 782                         break;
 783
 784                 if (q->dma_drain_size && rq->data_len) {
 785                         /*
 786                          * make sure space for the drain appears we
 787                          * know we can do this because max_hw_segments
 788                          * has been adjusted to be one fewer than the
 789                          * device can handle
 790                          */
 791                         rq->nr_phys_segments++;
 792                 }
 793
 794                 if (!q->prep_rq_fn)
 795                         break;
 796
 797                 ret = q->prep_rq_fn(q, rq);
 798                 if (ret == BLKPREP_OK) {
 799                         break;
 800                 } else if (ret == BLKPREP_DEFER) {
 801                         /*
 802                          * the request may have been (partially) prepped.
 803                          * we need to keep this request in the front to
 804                          * avoid resource deadlock.  REQ_STARTED will
 805                          * prevent other fs requests from passing this one.
 806                          */
 807                         if (q->dma_drain_size && rq->data_len &&
 808                             !(rq->cmd_flags & REQ_DONTPREP)) {
 809                                 /*
 810                                  * remove the space for the drain we added
 811                                  * so that we don't add it again
 812                                  */
 813                                 --rq->nr_phys_segments;
 814                         }
 815
 816                         rq = NULL;
 817                         break;
 818                 } else if (ret == BLKPREP_KILL) {
 819                         rq->cmd_flags |= REQ_QUIET;
 820                         end_queued_request(rq, 0);
 821                 } else {
 822                         printk(KERN_ERR "%s: bad return=%d\n", __func__, ret);
 823                         break;
 824                 }
 825         }
 826
 827         return rq;
 828 }
 829 EXPORT_SYMBOL(elv_next_request);
 830
 831 void elv_dequeue_request(struct request_queue *q, struct request *rq)
 832 {
 833         BUG_ON(list_empty(&rq->queuelist));
 834         BUG_ON(ELV_ON_HASH(rq));
 835
 836         list_del_init(&rq->queuelist);
 837
 838         /*
 839          * the time frame between a request being removed from the lists
 840          * and to it is freed is accounted as io that is in progress at
 841          * the driver side.
 842          */
 843         if (blk_account_rq(rq))
 844                 q->in_flight++;
 845 }
 846 EXPORT_SYMBOL(elv_dequeue_request);
 847
 848 int elv_queue_empty(struct request_queue *q)
 849 {
 850         elevator_t *e = q->elevator;
 851
 852         if (!list_empty(&q->queue_head))
 853                 return 0;
 854
 855         if (e->ops->elevator_queue_empty_fn)
 856                 return e->ops->elevator_queue_empty_fn(q);
 857
 858         return 1;
 859 }
 860 EXPORT_SYMBOL(elv_queue_empty);
 861
 862 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 863 {
 864         elevator_t *e = q->elevator;
 865
 866         if (e->ops->elevator_latter_req_fn)
 867                 return e->ops->elevator_latter_req_fn(q, rq);
 868         return NULL;
 869 }
 870
 871 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 872 {
 873         elevator_t *e = q->elevator;
 874
 875         if (e->ops->elevator_former_req_fn)
 876                 return e->ops->elevator_former_req_fn(q, rq);
 877         return NULL;
 878 }
 879
 880 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 881 {
 882         elevator_t *e = q->elevator;
 883
 884         if (e->ops->elevator_set_req_fn)
 885                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 886
 887         rq->elevator_private = NULL;
 888         return 0;
 889 }
 890
 891 void elv_put_request(struct request_queue *q, struct request *rq)
 892 {
 893         elevator_t *e = q->elevator;
 894
 895         if (e->ops->elevator_put_req_fn)
 896                 e->ops->elevator_put_req_fn(rq);
 897 }
 898
 899 int elv_may_queue(struct request_queue *q, int rw)
 900 {
 901         elevator_t *e = q->elevator;
 902
 903         if (e->ops->elevator_may_queue_fn)
 904                 return e->ops->elevator_may_queue_fn(q, rw);
 905
 906         return ELV_MQUEUE_MAY;
 907 }
 908
 909 void elv_completed_request(struct request_queue *q, struct request *rq)
 910 {
 911         elevator_t *e = q->elevator;
 912
 913         /*
 914          * request is released from the driver, io must be done
 915          */
 916         if (blk_account_rq(rq)) {
 917                 q->in_flight--;
 918                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 919                         e->ops->elevator_completed_req_fn(q, rq);
 920         }
 921
 922         /*
 923          * Check if the queue is waiting for fs requests to be
 924          * drained for flush sequence.
 925          */
 926         if (unlikely(q->ordseq)) {
 927                 struct request *first_rq = list_entry_rq(q->queue_head.next);
 928                 if (q->in_flight == 0 &&
 929                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 930                     blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
 931                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 932                         q->request_fn(q);
 933                 }
 934         }
 935 }
 936
 937 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 938
 939 static ssize_t
 940 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 941 {
 942         elevator_t *e = container_of(kobj, elevator_t, kobj);
 943         struct elv_fs_entry *entry = to_elv(attr);
 944         ssize_t error;
 945
 946         if (!entry->show)
 947                 return -EIO;
 948
 949         mutex_lock(&e->sysfs_lock);
 950         error = e->ops ? entry->show(e, page) : -ENOENT;
 951         mutex_unlock(&e->sysfs_lock);
 952         return error;
 953 }
 954
 955 static ssize_t
 956 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 957                const char *page, size_t length)
 958 {
 959         elevator_t *e = container_of(kobj, elevator_t, kobj);
 960         struct elv_fs_entry *entry = to_elv(attr);
 961         ssize_t error;
 962
 963         if (!entry->store)
 964                 return -EIO;
 965
 966         mutex_lock(&e->sysfs_lock);
 967         error = e->ops ? entry->store(e, page, length) : -ENOENT;
 968         mutex_unlock(&e->sysfs_lock);
 969         return error;
 970 }
 971
 972 static struct sysfs_ops elv_sysfs_ops = {
 973         .show   = elv_attr_show,
 974         .store  = elv_attr_store,
 975 };
 976
 977 static struct kobj_type elv_ktype = {
 978         .sysfs_ops      = &elv_sysfs_ops,
 979         .release        = elevator_release,
 980 };
 981
 982 int elv_register_queue(struct request_queue *q)
 983 {
 984         elevator_t *e = q->elevator;
 985         int error;
 986
 987         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 988         if (!error) {
 989                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
 990                 if (attr) {
 991                         while (attr->attr.name) {
 992                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 993                                         break;
 994                                 attr++;
 995                         }
 996                 }
 997                 kobject_uevent(&e->kobj, KOBJ_ADD);
 998         }
 999         return error;
1000 }
1001
1002 static void __elv_unregister_queue(elevator_t *e)
1003 {
1004         kobject_uevent(&e->kobj, KOBJ_REMOVE);
1005         kobject_del(&e->kobj);
1006 }
1007
1008 void elv_unregister_queue(struct request_queue *q)
1009 {
1010         if (q)
1011                 __elv_unregister_queue(q->elevator);
1012 }
1013
1014 void elv_register(struct elevator_type *e)
1015 {
1016         char *def = "";
1017
1018         spin_lock(&elv_list_lock);
1019         BUG_ON(elevator_find(e->elevator_name));
1020         list_add_tail(&e->list, &elv_list);
1021         spin_unlock(&elv_list_lock);
1022
1023         if (!strcmp(e->elevator_name, chosen_elevator) ||
1024                         (!*chosen_elevator &&
1025                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
1026                                 def = " (default)";
1027
1028         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
1029                                                                 def);
1030 }
1031 EXPORT_SYMBOL_GPL(elv_register);
1032
1033 void elv_unregister(struct elevator_type *e)
1034 {
1035         struct task_struct *g, *p;
1036
1037         /*
1038          * Iterate every thread in the process to remove the io contexts.
1039          */
1040         if (e->ops.trim) {
1041                 read_lock(&tasklist_lock);
1042                 do_each_thread(g, p) {
1043                         task_lock(p);
1044                         if (p->io_context)
1045                                 e->ops.trim(p->io_context);
1046                         task_unlock(p);
1047                 } while_each_thread(g, p);
1048                 read_unlock(&tasklist_lock);
1049         }
1050
1051         spin_lock(&elv_list_lock);
1052         list_del_init(&e->list);
1053         spin_unlock(&elv_list_lock);
1054 }
1055 EXPORT_SYMBOL_GPL(elv_unregister);
1056
1057 /*
1058  * switch to new_e io scheduler. be careful not to introduce deadlocks -
1059  * we don't free the old io scheduler, before we have allocated what we
1060  * need for the new one. this way we have a chance of going back to the old
1061  * one, if the new one fails init for some reason.
1062  */
1063 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1064 {
1065         elevator_t *old_elevator, *e;
1066         void *data;
1067
1068         /*
1069          * Allocate new elevator
1070          */
1071         e = elevator_alloc(q, new_e);
1072         if (!e)
1073                 return 0;
1074
1075         data = elevator_init_queue(q, e);
1076         if (!data) {
1077                 kobject_put(&e->kobj);
1078                 return 0;
1079         }
1080
1081         /*
1082          * Turn on BYPASS and drain all requests w/ elevator private data
1083          */
1084         spin_lock_irq(q->queue_lock);
1085
1086         queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
1087
1088         elv_drain_elevator(q);
1089
1090         while (q->rq.elvpriv) {
1091                 blk_remove_plug(q);
1092                 q->request_fn(q);
1093                 spin_unlock_irq(q->queue_lock);
1094                 msleep(10);
1095                 spin_lock_irq(q->queue_lock);
1096                 elv_drain_elevator(q);
1097         }
1098
1099         /*
1100          * Remember old elevator.
1101          */
1102         old_elevator = q->elevator;
1103
1104         /*
1105          * attach and start new elevator
1106          */
1107         elevator_attach(q, e, data);
1108
1109         spin_unlock_irq(q->queue_lock);
1110
1111         __elv_unregister_queue(old_elevator);
1112
1113         if (elv_register_queue(q))
1114                 goto fail_register;
1115
1116         /*
1117          * finally exit old elevator and turn off BYPASS.
1118          */
1119         elevator_exit(old_elevator);
1120         spin_lock_irq(q->queue_lock);
1121         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1122         spin_unlock_irq(q->queue_lock);
1123
1124         blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
1125
1126         return 1;
1127
1128 fail_register:
1129         /*
1130          * switch failed, exit the new io scheduler and reattach the old
1131          * one again (along with re-adding the sysfs dir)
1132          */
1133         elevator_exit(e);
1134         q->elevator = old_elevator;
1135         elv_register_queue(q);
1136
1137         spin_lock_irq(q->queue_lock);
1138         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1139         spin_unlock_irq(q->queue_lock);
1140
1141         return 0;
1142 }
1143
1144 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1145                           size_t count)
1146 {
1147         char elevator_name[ELV_NAME_MAX];
1148         size_t len;
1149         struct elevator_type *e;
1150
1151         elevator_name[sizeof(elevator_name) - 1] = '\0';
1152         strncpy(elevator_name, name, sizeof(elevator_name) - 1);
1153         len = strlen(elevator_name);
1154
1155         if (len && elevator_name[len - 1] == '\n')
1156                 elevator_name[len - 1] = '\0';
1157
1158         e = elevator_get(elevator_name);
1159         if (!e) {
1160                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1161                 return -EINVAL;
1162         }
1163
1164         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1165                 elevator_put(e);
1166                 return count;
1167         }
1168
1169         if (!elevator_switch(q, e))
1170                 printk(KERN_ERR "elevator: switch to %s failed\n",
1171                                                         elevator_name);
1172         return count;
1173 }
1174
1175 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1176 {
1177         elevator_t *e = q->elevator;
1178         struct elevator_type *elv = e->elevator_type;
1179         struct elevator_type *__e;
1180         int len = 0;
1181
1182         spin_lock(&elv_list_lock);
1183         list_for_each_entry(__e, &elv_list, list) {
1184                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1185                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1186                 else
1187                         len += sprintf(name+len, "%s ", __e->elevator_name);
1188         }
1189         spin_unlock(&elv_list_lock);
1190
1191         len += sprintf(len+name, "\n");
1192         return len;
1193 }
1194
1195 struct request *elv_rb_former_request(struct request_queue *q,
1196                                       struct request *rq)
1197 {
1198         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1199
1200         if (rbprev)
1201                 return rb_entry_rq(rbprev);
1202
1203         return NULL;
1204 }
1205 EXPORT_SYMBOL(elv_rb_former_request);
1206
1207 struct request *elv_rb_latter_request(struct request_queue *q,
1208                                       struct request *rq)
1209 {
1210         struct rb_node *rbnext = rb_next(&rq->rb_node);
1211
1212         if (rbnext)
1213                 return rb_entry_rq(rbnext);
1214
1215         return NULL;
1216 }
1217 EXPORT_SYMBOL(elv_rb_latter_request);