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 <trace/block.h>
  37 #include <linux/hash.h>
  38 #include <linux/uaccess.h>
  39
  40 #include "blk.h"
  41
  42 static DEFINE_SPINLOCK(elv_list_lock);
  43 static LIST_HEAD(elv_list);
  44
  45 /*
  46  * Merge hash stuff.
  47  */
  48 static const int elv_hash_shift = 6;
  49 #define ELV_HASH_BLOCK(sec)     ((sec) >> 3)
  50 #define ELV_HASH_FN(sec)        \
  51                 (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
  52 #define ELV_HASH_ENTRIES        (1 << elv_hash_shift)
  53 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
  54 #define ELV_ON_HASH(rq)         (!hlist_unhashed(&(rq)->hash))
  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         elevator_t *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_discard(bio) != bio_discard(rq->bio))
  83                 return 0;
  84
  85         /*
  86          * different data direction or already started, don't merge
  87          */
  88         if (bio_data_dir(bio) != rq_data_dir(rq))
  89                 return 0;
  90
  91         /*
  92          * must be same device and not a special request
  93          */
  94         if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
  95                 return 0;
  96
  97         /*
  98          * only merge integrity protected bio into ditto rq
  99          */
 100         if (bio_integrity(bio) != blk_integrity_rq(rq))
 101                 return 0;
 102
 103         if (!elv_iosched_allow_merge(rq, bio))
 104                 return 0;
 105
 106         return 1;
 107 }
 108 EXPORT_SYMBOL(elv_rq_merge_ok);
 109
 110 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
 111 {
 112         int ret = ELEVATOR_NO_MERGE;
 113
 114         /*
 115          * we can merge and sequence is ok, check if it's possible
 116          */
 117         if (elv_rq_merge_ok(__rq, bio)) {
 118                 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
 119                         ret = ELEVATOR_BACK_MERGE;
 120                 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
 121                         ret = ELEVATOR_FRONT_MERGE;
 122         }
 123
 124         return ret;
 125 }
 126
 127 static struct elevator_type *elevator_find(const char *name)
 128 {
 129         struct elevator_type *e;
 130
 131         list_for_each_entry(e, &elv_list, list) {
 132                 if (!strcmp(e->elevator_name, name))
 133                         return e;
 134         }
 135
 136         return NULL;
 137 }
 138
 139 static void elevator_put(struct elevator_type *e)
 140 {
 141         module_put(e->elevator_owner);
 142 }
 143
 144 static struct elevator_type *elevator_get(const char *name)
 145 {
 146         struct elevator_type *e;
 147
 148         spin_lock(&elv_list_lock);
 149
 150         e = elevator_find(name);
 151         if (!e) {
 152                 char elv[ELV_NAME_MAX + strlen("-iosched")];
 153
 154                 spin_unlock(&elv_list_lock);
 155
 156                 if (!strcmp(name, "anticipatory"))
 157                         sprintf(elv, "as-iosched");
 158                 else
 159                         sprintf(elv, "%s-iosched", name);
 160
 161                 request_module("%s", elv);
 162                 spin_lock(&elv_list_lock);
 163                 e = elevator_find(name);
 164         }
 165
 166         if (e && !try_module_get(e->elevator_owner))
 167                 e = NULL;
 168
 169         spin_unlock(&elv_list_lock);
 170
 171         return e;
 172 }
 173
 174 static void *elevator_init_queue(struct request_queue *q,
 175                                  struct elevator_queue *eq)
 176 {
 177         return eq->ops->elevator_init_fn(q);
 178 }
 179
 180 static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
 181                            void *data)
 182 {
 183         q->elevator = eq;
 184         eq->elevator_data = data;
 185 }
 186
 187 static char chosen_elevator[16];
 188
 189 static int __init elevator_setup(char *str)
 190 {
 191         /*
 192          * Be backwards-compatible with previous kernels, so users
 193          * won't get the wrong elevator.
 194          */
 195         if (!strcmp(str, "as"))
 196                 strcpy(chosen_elevator, "anticipatory");
 197         else
 198                 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 199         return 1;
 200 }
 201
 202 __setup("elevator=", elevator_setup);
 203
 204 static struct kobj_type elv_ktype;
 205
 206 static elevator_t *elevator_alloc(struct request_queue *q,
 207                                   struct elevator_type *e)
 208 {
 209         elevator_t *eq;
 210         int i;
 211
 212         eq = kmalloc_node(sizeof(elevator_t), GFP_KERNEL | __GFP_ZERO, q->node);
 213         if (unlikely(!eq))
 214                 goto err;
 215
 216         eq->ops = &e->ops;
 217         eq->elevator_type = e;
 218         kobject_init(&eq->kobj, &elv_ktype);
 219         mutex_init(&eq->sysfs_lock);
 220
 221         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 222                                         GFP_KERNEL, q->node);
 223         if (!eq->hash)
 224                 goto err;
 225
 226         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 227                 INIT_HLIST_HEAD(&eq->hash[i]);
 228
 229         return eq;
 230 err:
 231         kfree(eq);
 232         elevator_put(e);
 233         return NULL;
 234 }
 235
 236 static void elevator_release(struct kobject *kobj)
 237 {
 238         elevator_t *e = container_of(kobj, elevator_t, kobj);
 239
 240         elevator_put(e->elevator_type);
 241         kfree(e->hash);
 242         kfree(e);
 243 }
 244
 245 int elevator_init(struct request_queue *q, char *name)
 246 {
 247         struct elevator_type *e = NULL;
 248         struct elevator_queue *eq;
 249         int ret = 0;
 250         void *data;
 251
 252         INIT_LIST_HEAD(&q->queue_head);
 253         q->last_merge = NULL;
 254         q->end_sector = 0;
 255         q->boundary_rq = NULL;
 256
 257         if (name) {
 258                 e = elevator_get(name);
 259                 if (!e)
 260                         return -EINVAL;
 261         }
 262
 263         if (!e && *chosen_elevator) {
 264                 e = elevator_get(chosen_elevator);
 265                 if (!e)
 266                         printk(KERN_ERR "I/O scheduler %s not found\n",
 267                                                         chosen_elevator);
 268         }
 269
 270         if (!e) {
 271                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
 272                 if (!e) {
 273                         printk(KERN_ERR
 274                                 "Default I/O scheduler not found. " \
 275                                 "Using noop.\n");
 276                         e = elevator_get("noop");
 277                 }
 278         }
 279
 280         eq = elevator_alloc(q, e);
 281         if (!eq)
 282                 return -ENOMEM;
 283
 284         data = elevator_init_queue(q, eq);
 285         if (!data) {
 286                 kobject_put(&eq->kobj);
 287                 return -ENOMEM;
 288         }
 289
 290         elevator_attach(q, eq, data);
 291         return ret;
 292 }
 293 EXPORT_SYMBOL(elevator_init);
 294
 295 void elevator_exit(elevator_t *e)
 296 {
 297         mutex_lock(&e->sysfs_lock);
 298         if (e->ops->elevator_exit_fn)
 299                 e->ops->elevator_exit_fn(e);
 300         e->ops = NULL;
 301         mutex_unlock(&e->sysfs_lock);
 302
 303         kobject_put(&e->kobj);
 304 }
 305 EXPORT_SYMBOL(elevator_exit);
 306
 307 static void elv_activate_rq(struct request_queue *q, struct request *rq)
 308 {
 309         elevator_t *e = q->elevator;
 310
 311         if (e->ops->elevator_activate_req_fn)
 312                 e->ops->elevator_activate_req_fn(q, rq);
 313 }
 314
 315 static void elv_deactivate_rq(struct request_queue *q, struct request *rq)
 316 {
 317         elevator_t *e = q->elevator;
 318
 319         if (e->ops->elevator_deactivate_req_fn)
 320                 e->ops->elevator_deactivate_req_fn(q, rq);
 321 }
 322
 323 static inline void __elv_rqhash_del(struct request *rq)
 324 {
 325         hlist_del_init(&rq->hash);
 326 }
 327
 328 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 329 {
 330         if (ELV_ON_HASH(rq))
 331                 __elv_rqhash_del(rq);
 332 }
 333
 334 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 335 {
 336         elevator_t *e = q->elevator;
 337
 338         BUG_ON(ELV_ON_HASH(rq));
 339         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 340 }
 341
 342 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 343 {
 344         __elv_rqhash_del(rq);
 345         elv_rqhash_add(q, rq);
 346 }
 347
 348 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 349 {
 350         elevator_t *e = q->elevator;
 351         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 352         struct hlist_node *entry, *next;
 353         struct request *rq;
 354
 355         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 356                 BUG_ON(!ELV_ON_HASH(rq));
 357
 358                 if (unlikely(!rq_mergeable(rq))) {
 359                         __elv_rqhash_del(rq);
 360                         continue;
 361                 }
 362
 363                 if (rq_hash_key(rq) == offset)
 364                         return rq;
 365         }
 366
 367         return NULL;
 368 }
 369
 370 /*
 371  * RB-tree support functions for inserting/lookup/removal of requests
 372  * in a sorted RB tree.
 373  */
 374 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 375 {
 376         struct rb_node **p = &root->rb_node;
 377         struct rb_node *parent = NULL;
 378         struct request *__rq;
 379
 380         while (*p) {
 381                 parent = *p;
 382                 __rq = rb_entry(parent, struct request, rb_node);
 383
 384                 if (rq->sector < __rq->sector)
 385                         p = &(*p)->rb_left;
 386                 else if (rq->sector > __rq->sector)
 387                         p = &(*p)->rb_right;
 388                 else
 389                         return __rq;
 390         }
 391
 392         rb_link_node(&rq->rb_node, parent, p);
 393         rb_insert_color(&rq->rb_node, root);
 394         return NULL;
 395 }
 396 EXPORT_SYMBOL(elv_rb_add);
 397
 398 void elv_rb_del(struct rb_root *root, struct request *rq)
 399 {
 400         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 401         rb_erase(&rq->rb_node, root);
 402         RB_CLEAR_NODE(&rq->rb_node);
 403 }
 404 EXPORT_SYMBOL(elv_rb_del);
 405
 406 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 407 {
 408         struct rb_node *n = root->rb_node;
 409         struct request *rq;
 410
 411         while (n) {
 412                 rq = rb_entry(n, struct request, rb_node);
 413
 414                 if (sector < rq->sector)
 415                         n = n->rb_left;
 416                 else if (sector > rq->sector)
 417                         n = n->rb_right;
 418                 else
 419                         return rq;
 420         }
 421
 422         return NULL;
 423 }
 424 EXPORT_SYMBOL(elv_rb_find);
 425
 426 /*
 427  * Insert rq into dispatch queue of q.  Queue lock must be held on
 428  * entry.  rq is sort instead into the dispatch queue. To be used by
 429  * specific elevators.
 430  */
 431 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 432 {
 433         sector_t boundary;
 434         struct list_head *entry;
 435         int stop_flags;
 436
 437         if (q->last_merge == rq)
 438                 q->last_merge = NULL;
 439
 440         elv_rqhash_del(q, rq);
 441
 442         q->nr_sorted--;
 443
 444         boundary = q->end_sector;
 445         stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
 446         list_for_each_prev(entry, &q->queue_head) {
 447                 struct request *pos = list_entry_rq(entry);
 448
 449                 if (blk_discard_rq(rq) != blk_discard_rq(pos))
 450                         break;
 451                 if (rq_data_dir(rq) != rq_data_dir(pos))
 452                         break;
 453                 if (pos->cmd_flags & stop_flags)
 454                         break;
 455                 if (rq->sector >= boundary) {
 456                         if (pos->sector < boundary)
 457                                 continue;
 458                 } else {
 459                         if (pos->sector >= boundary)
 460                                 break;
 461                 }
 462                 if (rq->sector >= pos->sector)
 463                         break;
 464         }
 465
 466         list_add(&rq->queuelist, entry);
 467 }
 468 EXPORT_SYMBOL(elv_dispatch_sort);
 469
 470 /*
 471  * Insert rq into dispatch queue of q.  Queue lock must be held on
 472  * entry.  rq is added to the back of the dispatch queue. To be used by
 473  * specific elevators.
 474  */
 475 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 476 {
 477         if (q->last_merge == rq)
 478                 q->last_merge = NULL;
 479
 480         elv_rqhash_del(q, rq);
 481
 482         q->nr_sorted--;
 483
 484         q->end_sector = rq_end_sector(rq);
 485         q->boundary_rq = rq;
 486         list_add_tail(&rq->queuelist, &q->queue_head);
 487 }
 488 EXPORT_SYMBOL(elv_dispatch_add_tail);
 489
 490 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 491 {
 492         elevator_t *e = q->elevator;
 493         struct request *__rq;
 494         int ret;
 495
 496         /*
 497          * First try one-hit cache.
 498          */
 499         if (q->last_merge) {
 500                 ret = elv_try_merge(q->last_merge, bio);
 501                 if (ret != ELEVATOR_NO_MERGE) {
 502                         *req = q->last_merge;
 503                         return ret;
 504                 }
 505         }
 506
 507         if (blk_queue_nomerges(q))
 508                 return ELEVATOR_NO_MERGE;
 509
 510         /*
 511          * See if our hash lookup can find a potential backmerge.
 512          */
 513         __rq = elv_rqhash_find(q, bio->bi_sector);
 514         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 515                 *req = __rq;
 516                 return ELEVATOR_BACK_MERGE;
 517         }
 518
 519         if (e->ops->elevator_merge_fn)
 520                 return e->ops->elevator_merge_fn(q, req, bio);
 521
 522         return ELEVATOR_NO_MERGE;
 523 }
 524
 525 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 526 {
 527         elevator_t *e = q->elevator;
 528
 529         if (e->ops->elevator_merged_fn)
 530                 e->ops->elevator_merged_fn(q, rq, type);
 531
 532         if (type == ELEVATOR_BACK_MERGE)
 533                 elv_rqhash_reposition(q, rq);
 534
 535         q->last_merge = rq;
 536 }
 537
 538 void elv_merge_requests(struct request_queue *q, struct request *rq,
 539                              struct request *next)
 540 {
 541         elevator_t *e = q->elevator;
 542
 543         if (e->ops->elevator_merge_req_fn)
 544                 e->ops->elevator_merge_req_fn(q, rq, next);
 545
 546         elv_rqhash_reposition(q, rq);
 547         elv_rqhash_del(q, next);
 548
 549         q->nr_sorted--;
 550         q->last_merge = rq;
 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--;
 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 static 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 void elv_insert(struct request_queue *q, struct request *rq, int where)
 585 {
 586         struct list_head *pos;
 587         unsigned ordseq;
 588         int unplug_it = 1;
 589
 590         trace_block_rq_insert(q, rq);
 591
 592         rq->q = q;
 593
 594         switch (where) {
 595         case ELEVATOR_INSERT_FRONT:
 596                 rq->cmd_flags |= REQ_SOFTBARRIER;
 597
 598                 list_add(&rq->queuelist, &q->queue_head);
 599                 break;
 600
 601         case ELEVATOR_INSERT_BACK:
 602                 rq->cmd_flags |= REQ_SOFTBARRIER;
 603                 elv_drain_elevator(q);
 604                 list_add_tail(&rq->queuelist, &q->queue_head);
 605                 /*
 606                  * We kick the queue here for the following reasons.
 607                  * - The elevator might have returned NULL previously
 608                  *   to delay requests and returned them now.  As the
 609                  *   queue wasn't empty before this request, ll_rw_blk
 610                  *   won't run the queue on return, resulting in hang.
 611                  * - Usually, back inserted requests won't be merged
 612                  *   with anything.  There's no point in delaying queue
 613                  *   processing.
 614                  */
 615                 blk_remove_plug(q);
 616                 blk_start_queueing(q);
 617                 break;
 618
 619         case ELEVATOR_INSERT_SORT:
 620                 BUG_ON(!blk_fs_request(rq) && !blk_discard_rq(rq));
 621                 rq->cmd_flags |= REQ_SORTED;
 622                 q->nr_sorted++;
 623                 if (rq_mergeable(rq)) {
 624                         elv_rqhash_add(q, rq);
 625                         if (!q->last_merge)
 626                                 q->last_merge = rq;
 627                 }
 628
 629                 /*
 630                  * Some ioscheds (cfq) run q->request_fn directly, so
 631                  * rq cannot be accessed after calling
 632                  * elevator_add_req_fn.
 633                  */
 634                 q->elevator->ops->elevator_add_req_fn(q, rq);
 635                 break;
 636
 637         case ELEVATOR_INSERT_REQUEUE:
 638                 /*
 639                  * If ordered flush isn't in progress, we do front
 640                  * insertion; otherwise, requests should be requeued
 641                  * in ordseq order.
 642                  */
 643                 rq->cmd_flags |= REQ_SOFTBARRIER;
 644
 645                 /*
 646                  * Most requeues happen because of a busy condition,
 647                  * don't force unplug of the queue for that case.
 648                  */
 649                 unplug_it = 0;
 650
 651                 if (q->ordseq == 0) {
 652                         list_add(&rq->queuelist, &q->queue_head);
 653                         break;
 654                 }
 655
 656                 ordseq = blk_ordered_req_seq(rq);
 657
 658                 list_for_each(pos, &q->queue_head) {
 659                         struct request *pos_rq = list_entry_rq(pos);
 660                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 661                                 break;
 662                 }
 663
 664                 list_add_tail(&rq->queuelist, pos);
 665                 break;
 666
 667         default:
 668                 printk(KERN_ERR "%s: bad insertion point %d\n",
 669                        __func__, where);
 670                 BUG();
 671         }
 672
 673         if (unplug_it && blk_queue_plugged(q)) {
 674                 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
 675                         - q->in_flight;
 676
 677                 if (nrq >= q->unplug_thresh)
 678                         __generic_unplug_device(q);
 679         }
 680 }
 681
 682 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
 683                        int plug)
 684 {
 685         if (q->ordcolor)
 686                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 687
 688         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 689                 /*
 690                  * toggle ordered color
 691                  */
 692                 if (blk_barrier_rq(rq))
 693                         q->ordcolor ^= 1;
 694
 695                 /*
 696                  * barriers implicitly indicate back insertion
 697                  */
 698                 if (where == ELEVATOR_INSERT_SORT)
 699                         where = ELEVATOR_INSERT_BACK;
 700
 701                 /*
 702                  * this request is scheduling boundary, update
 703                  * end_sector
 704                  */
 705                 if (blk_fs_request(rq) || blk_discard_rq(rq)) {
 706                         q->end_sector = rq_end_sector(rq);
 707                         q->boundary_rq = rq;
 708                 }
 709         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 710                     where == ELEVATOR_INSERT_SORT)
 711                 where = ELEVATOR_INSERT_BACK;
 712
 713         if (plug)
 714                 blk_plug_device(q);
 715
 716         elv_insert(q, rq, where);
 717 }
 718 EXPORT_SYMBOL(__elv_add_request);
 719
 720 void elv_add_request(struct request_queue *q, struct request *rq, int where,
 721                      int plug)
 722 {
 723         unsigned long flags;
 724
 725         spin_lock_irqsave(q->queue_lock, flags);
 726         __elv_add_request(q, rq, where, plug);
 727         spin_unlock_irqrestore(q->queue_lock, flags);
 728 }
 729 EXPORT_SYMBOL(elv_add_request);
 730
 731 static inline struct request *__elv_next_request(struct request_queue *q)
 732 {
 733         struct request *rq;
 734
 735         while (1) {
 736                 while (!list_empty(&q->queue_head)) {
 737                         rq = list_entry_rq(q->queue_head.next);
 738                         if (blk_do_ordered(q, &rq))
 739                                 return rq;
 740                 }
 741
 742                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
 743                         return NULL;
 744         }
 745 }
 746
 747 struct request *elv_next_request(struct request_queue *q)
 748 {
 749         struct request *rq;
 750         int ret;
 751
 752         while ((rq = __elv_next_request(q)) != NULL) {
 753                 /*
 754                  * Kill the empty barrier place holder, the driver must
 755                  * not ever see it.
 756                  */
 757                 if (blk_empty_barrier(rq)) {
 758                         __blk_end_request(rq, 0, blk_rq_bytes(rq));
 759                         continue;
 760                 }
 761                 if (!(rq->cmd_flags & REQ_STARTED)) {
 762                         /*
 763                          * This is the first time the device driver
 764                          * sees this request (possibly after
 765                          * requeueing).  Notify IO scheduler.
 766                          */
 767                         if (blk_sorted_rq(rq))
 768                                 elv_activate_rq(q, rq);
 769
 770                         /*
 771                          * just mark as started even if we don't start
 772                          * it, a request that has been delayed should
 773                          * not be passed by new incoming requests
 774                          */
 775                         rq->cmd_flags |= REQ_STARTED;
 776                         trace_block_rq_issue(q, rq);
 777                 }
 778
 779                 if (!q->boundary_rq || q->boundary_rq == rq) {
 780                         q->end_sector = rq_end_sector(rq);
 781                         q->boundary_rq = NULL;
 782                 }
 783
 784                 if (rq->cmd_flags & REQ_DONTPREP)
 785                         break;
 786
 787                 if (q->dma_drain_size && rq->data_len) {
 788                         /*
 789                          * make sure space for the drain appears we
 790                          * know we can do this because max_hw_segments
 791                          * has been adjusted to be one fewer than the
 792                          * device can handle
 793                          */
 794                         rq->nr_phys_segments++;
 795                 }
 796
 797                 if (!q->prep_rq_fn)
 798                         break;
 799
 800                 ret = q->prep_rq_fn(q, rq);
 801                 if (ret == BLKPREP_OK) {
 802                         break;
 803                 } else if (ret == BLKPREP_DEFER) {
 804                         /*
 805                          * the request may have been (partially) prepped.
 806                          * we need to keep this request in the front to
 807                          * avoid resource deadlock.  REQ_STARTED will
 808                          * prevent other fs requests from passing this one.
 809                          */
 810                         if (q->dma_drain_size && rq->data_len &&
 811                             !(rq->cmd_flags & REQ_DONTPREP)) {
 812                                 /*
 813                                  * remove the space for the drain we added
 814                                  * so that we don't add it again
 815                                  */
 816                                 --rq->nr_phys_segments;
 817                         }
 818
 819                         rq = NULL;
 820                         break;
 821                 } else if (ret == BLKPREP_KILL) {
 822                         rq->cmd_flags |= REQ_QUIET;
 823                         __blk_end_request(rq, -EIO, blk_rq_bytes(rq));
 824                 } else {
 825                         printk(KERN_ERR "%s: bad return=%d\n", __func__, ret);
 826                         break;
 827                 }
 828         }
 829
 830         return rq;
 831 }
 832 EXPORT_SYMBOL(elv_next_request);
 833
 834 void elv_dequeue_request(struct request_queue *q, struct request *rq)
 835 {
 836         BUG_ON(list_empty(&rq->queuelist));
 837         BUG_ON(ELV_ON_HASH(rq));
 838
 839         list_del_init(&rq->queuelist);
 840
 841         /*
 842          * the time frame between a request being removed from the lists
 843          * and to it is freed is accounted as io that is in progress at
 844          * the driver side.
 845          */
 846         if (blk_account_rq(rq))
 847                 q->in_flight++;
 848
 849         /*
 850          * We are now handing the request to the hardware, add the
 851          * timeout handler.
 852          */
 853         blk_add_timer(rq);
 854 }
 855 EXPORT_SYMBOL(elv_dequeue_request);
 856
 857 int elv_queue_empty(struct request_queue *q)
 858 {
 859         elevator_t *e = q->elevator;
 860
 861         if (!list_empty(&q->queue_head))
 862                 return 0;
 863
 864         if (e->ops->elevator_queue_empty_fn)
 865                 return e->ops->elevator_queue_empty_fn(q);
 866
 867         return 1;
 868 }
 869 EXPORT_SYMBOL(elv_queue_empty);
 870
 871 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 872 {
 873         elevator_t *e = q->elevator;
 874
 875         if (e->ops->elevator_latter_req_fn)
 876                 return e->ops->elevator_latter_req_fn(q, rq);
 877         return NULL;
 878 }
 879
 880 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 881 {
 882         elevator_t *e = q->elevator;
 883
 884         if (e->ops->elevator_former_req_fn)
 885                 return e->ops->elevator_former_req_fn(q, rq);
 886         return NULL;
 887 }
 888
 889 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 890 {
 891         elevator_t *e = q->elevator;
 892
 893         if (e->ops->elevator_set_req_fn)
 894                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 895
 896         rq->elevator_private = NULL;
 897         return 0;
 898 }
 899
 900 void elv_put_request(struct request_queue *q, struct request *rq)
 901 {
 902         elevator_t *e = q->elevator;
 903
 904         if (e->ops->elevator_put_req_fn)
 905                 e->ops->elevator_put_req_fn(rq);
 906 }
 907
 908 int elv_may_queue(struct request_queue *q, int rw)
 909 {
 910         elevator_t *e = q->elevator;
 911
 912         if (e->ops->elevator_may_queue_fn)
 913                 return e->ops->elevator_may_queue_fn(q, rw);
 914
 915         return ELV_MQUEUE_MAY;
 916 }
 917
 918 void elv_abort_queue(struct request_queue *q)
 919 {
 920         struct request *rq;
 921
 922         while (!list_empty(&q->queue_head)) {
 923                 rq = list_entry_rq(q->queue_head.next);
 924                 rq->cmd_flags |= REQ_QUIET;
 925                 trace_block_rq_abort(q, rq);
 926                 __blk_end_request(rq, -EIO, blk_rq_bytes(rq));
 927         }
 928 }
 929 EXPORT_SYMBOL(elv_abort_queue);
 930
 931 void elv_completed_request(struct request_queue *q, struct request *rq)
 932 {
 933         elevator_t *e = q->elevator;
 934
 935         /*
 936          * request is released from the driver, io must be done
 937          */
 938         if (blk_account_rq(rq)) {
 939                 q->in_flight--;
 940                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 941                         e->ops->elevator_completed_req_fn(q, rq);
 942         }
 943
 944         /*
 945          * Check if the queue is waiting for fs requests to be
 946          * drained for flush sequence.
 947          */
 948         if (unlikely(q->ordseq)) {
 949                 struct request *first_rq = list_entry_rq(q->queue_head.next);
 950                 if (q->in_flight == 0 &&
 951                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 952                     blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
 953                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 954                         blk_start_queueing(q);
 955                 }
 956         }
 957 }
 958
 959 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 960
 961 static ssize_t
 962 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 963 {
 964         elevator_t *e = container_of(kobj, elevator_t, kobj);
 965         struct elv_fs_entry *entry = to_elv(attr);
 966         ssize_t error;
 967
 968         if (!entry->show)
 969                 return -EIO;
 970
 971         mutex_lock(&e->sysfs_lock);
 972         error = e->ops ? entry->show(e, page) : -ENOENT;
 973         mutex_unlock(&e->sysfs_lock);
 974         return error;
 975 }
 976
 977 static ssize_t
 978 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 979                const char *page, size_t length)
 980 {
 981         elevator_t *e = container_of(kobj, elevator_t, kobj);
 982         struct elv_fs_entry *entry = to_elv(attr);
 983         ssize_t error;
 984
 985         if (!entry->store)
 986                 return -EIO;
 987
 988         mutex_lock(&e->sysfs_lock);
 989         error = e->ops ? entry->store(e, page, length) : -ENOENT;
 990         mutex_unlock(&e->sysfs_lock);
 991         return error;
 992 }
 993
 994 static struct sysfs_ops elv_sysfs_ops = {
 995         .show   = elv_attr_show,
 996         .store  = elv_attr_store,
 997 };
 998
 999 static struct kobj_type elv_ktype = {
1000         .sysfs_ops      = &elv_sysfs_ops,
1001         .release        = elevator_release,
1002 };
1003
1004 int elv_register_queue(struct request_queue *q)
1005 {
1006         elevator_t *e = q->elevator;
1007         int error;
1008
1009         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
1010         if (!error) {
1011                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
1012                 if (attr) {
1013                         while (attr->attr.name) {
1014                                 if (sysfs_create_file(&e->kobj, &attr->attr))
1015                                         break;
1016                                 attr++;
1017                         }
1018                 }
1019                 kobject_uevent(&e->kobj, KOBJ_ADD);
1020         }
1021         return error;
1022 }
1023
1024 static void __elv_unregister_queue(elevator_t *e)
1025 {
1026         kobject_uevent(&e->kobj, KOBJ_REMOVE);
1027         kobject_del(&e->kobj);
1028 }
1029
1030 void elv_unregister_queue(struct request_queue *q)
1031 {
1032         if (q)
1033                 __elv_unregister_queue(q->elevator);
1034 }
1035
1036 void elv_register(struct elevator_type *e)
1037 {
1038         char *def = "";
1039
1040         spin_lock(&elv_list_lock);
1041         BUG_ON(elevator_find(e->elevator_name));
1042         list_add_tail(&e->list, &elv_list);
1043         spin_unlock(&elv_list_lock);
1044
1045         if (!strcmp(e->elevator_name, chosen_elevator) ||
1046                         (!*chosen_elevator &&
1047                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
1048                                 def = " (default)";
1049
1050         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
1051                                                                 def);
1052 }
1053 EXPORT_SYMBOL_GPL(elv_register);
1054
1055 void elv_unregister(struct elevator_type *e)
1056 {
1057         struct task_struct *g, *p;
1058
1059         /*
1060          * Iterate every thread in the process to remove the io contexts.
1061          */
1062         if (e->ops.trim) {
1063                 read_lock(&tasklist_lock);
1064                 do_each_thread(g, p) {
1065                         task_lock(p);
1066                         if (p->io_context)
1067                                 e->ops.trim(p->io_context);
1068                         task_unlock(p);
1069                 } while_each_thread(g, p);
1070                 read_unlock(&tasklist_lock);
1071         }
1072
1073         spin_lock(&elv_list_lock);
1074         list_del_init(&e->list);
1075         spin_unlock(&elv_list_lock);
1076 }
1077 EXPORT_SYMBOL_GPL(elv_unregister);
1078
1079 /*
1080  * switch to new_e io scheduler. be careful not to introduce deadlocks -
1081  * we don't free the old io scheduler, before we have allocated what we
1082  * need for the new one. this way we have a chance of going back to the old
1083  * one, if the new one fails init for some reason.
1084  */
1085 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1086 {
1087         elevator_t *old_elevator, *e;
1088         void *data;
1089
1090         /*
1091          * Allocate new elevator
1092          */
1093         e = elevator_alloc(q, new_e);
1094         if (!e)
1095                 return 0;
1096
1097         data = elevator_init_queue(q, e);
1098         if (!data) {
1099                 kobject_put(&e->kobj);
1100                 return 0;
1101         }
1102
1103         /*
1104          * Turn on BYPASS and drain all requests w/ elevator private data
1105          */
1106         spin_lock_irq(q->queue_lock);
1107
1108         queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
1109
1110         elv_drain_elevator(q);
1111
1112         while (q->rq.elvpriv) {
1113                 blk_start_queueing(q);
1114                 spin_unlock_irq(q->queue_lock);
1115                 msleep(10);
1116                 spin_lock_irq(q->queue_lock);
1117                 elv_drain_elevator(q);
1118         }
1119
1120         /*
1121          * Remember old elevator.
1122          */
1123         old_elevator = q->elevator;
1124
1125         /*
1126          * attach and start new elevator
1127          */
1128         elevator_attach(q, e, data);
1129
1130         spin_unlock_irq(q->queue_lock);
1131
1132         __elv_unregister_queue(old_elevator);
1133
1134         if (elv_register_queue(q))
1135                 goto fail_register;
1136
1137         /*
1138          * finally exit old elevator and turn off BYPASS.
1139          */
1140         elevator_exit(old_elevator);
1141         spin_lock_irq(q->queue_lock);
1142         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1143         spin_unlock_irq(q->queue_lock);
1144
1145         blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
1146
1147         return 1;
1148
1149 fail_register:
1150         /*
1151          * switch failed, exit the new io scheduler and reattach the old
1152          * one again (along with re-adding the sysfs dir)
1153          */
1154         elevator_exit(e);
1155         q->elevator = old_elevator;
1156         elv_register_queue(q);
1157
1158         spin_lock_irq(q->queue_lock);
1159         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1160         spin_unlock_irq(q->queue_lock);
1161
1162         return 0;
1163 }
1164
1165 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1166                           size_t count)
1167 {
1168         char elevator_name[ELV_NAME_MAX];
1169         struct elevator_type *e;
1170
1171         strlcpy(elevator_name, name, sizeof(elevator_name));
1172         strstrip(elevator_name);
1173
1174         e = elevator_get(elevator_name);
1175         if (!e) {
1176                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1177                 return -EINVAL;
1178         }
1179
1180         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1181                 elevator_put(e);
1182                 return count;
1183         }
1184
1185         if (!elevator_switch(q, e))
1186                 printk(KERN_ERR "elevator: switch to %s failed\n",
1187                                                         elevator_name);
1188         return count;
1189 }
1190
1191 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1192 {
1193         elevator_t *e = q->elevator;
1194         struct elevator_type *elv = e->elevator_type;
1195         struct elevator_type *__e;
1196         int len = 0;
1197
1198         spin_lock(&elv_list_lock);
1199         list_for_each_entry(__e, &elv_list, list) {
1200                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1201                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1202                 else
1203                         len += sprintf(name+len, "%s ", __e->elevator_name);
1204         }
1205         spin_unlock(&elv_list_lock);
1206
1207         len += sprintf(len+name, "\n");
1208         return len;
1209 }
1210
1211 struct request *elv_rb_former_request(struct request_queue *q,
1212                                       struct request *rq)
1213 {
1214         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1215
1216         if (rbprev)
1217                 return rb_entry_rq(rbprev);
1218
1219         return NULL;
1220 }
1221 EXPORT_SYMBOL(elv_rb_former_request);
1222
1223 struct request *elv_rb_latter_request(struct request_queue *q,
1224                                       struct request *rq)
1225 {
1226         struct rb_node *rbnext = rb_next(&rq->rb_node);
1227
1228         if (rbnext)
1229                 return rb_entry_rq(rbnext);
1230
1231         return NULL;
1232 }
1233 EXPORT_SYMBOL(elv_rb_latter_request);