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 #include <trace/events/block.h>
  40
  41 #include "blk.h"
  42
  43 static DEFINE_SPINLOCK(elv_list_lock);
  44 static LIST_HEAD(elv_list);
  45
  46 /*
  47  * Merge hash stuff.
  48  */
  49 static const int elv_hash_shift = 6;
  50 #define ELV_HASH_BLOCK(sec)     ((sec) >> 3)
  51 #define ELV_HASH_FN(sec)        \
  52                 (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
  53 #define ELV_HASH_ENTRIES        (1 << elv_hash_shift)
  54 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
  55 #define ELV_ON_HASH(rq)         (!hlist_unhashed(&(rq)->hash))
  56
  57 /*
  58  * Query io scheduler to see if the current process issuing bio may be
  59  * merged with rq.
  60  */
  61 static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
  62 {
  63         struct request_queue *q = rq->q;
  64         struct elevator_queue *e = q->elevator;
  65
  66         if (e->ops->elevator_allow_merge_fn)
  67                 return e->ops->elevator_allow_merge_fn(q, rq, bio);
  68
  69         return 1;
  70 }
  71
  72 /*
  73  * can we safely merge with this request?
  74  */
  75 int elv_rq_merge_ok(struct request *rq, struct bio *bio)
  76 {
  77         if (!rq_mergeable(rq))
  78                 return 0;
  79
  80         /*
  81          * Don't merge file system requests and discard requests
  82          */
  83         if (bio_discard(bio) != bio_discard(rq->bio))
  84                 return 0;
  85
  86         /*
  87          * different data direction or already started, don't merge
  88          */
  89         if (bio_data_dir(bio) != rq_data_dir(rq))
  90                 return 0;
  91
  92         /*
  93          * must be same device and not a special request
  94          */
  95         if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
  96                 return 0;
  97
  98         /*
  99          * only merge integrity protected bio into ditto rq
 100          */
 101         if (bio_integrity(bio) != blk_integrity_rq(rq))
 102                 return 0;
 103
 104         if (!elv_iosched_allow_merge(rq, bio))
 105                 return 0;
 106
 107         return 1;
 108 }
 109 EXPORT_SYMBOL(elv_rq_merge_ok);
 110
 111 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
 112 {
 113         int ret = ELEVATOR_NO_MERGE;
 114
 115         /*
 116          * we can merge and sequence is ok, check if it's possible
 117          */
 118         if (elv_rq_merge_ok(__rq, bio)) {
 119                 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
 120                         ret = ELEVATOR_BACK_MERGE;
 121                 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
 122                         ret = ELEVATOR_FRONT_MERGE;
 123         }
 124
 125         return ret;
 126 }
 127
 128 static struct elevator_type *elevator_find(const char *name)
 129 {
 130         struct elevator_type *e;
 131
 132         list_for_each_entry(e, &elv_list, list) {
 133                 if (!strcmp(e->elevator_name, name))
 134                         return e;
 135         }
 136
 137         return NULL;
 138 }
 139
 140 static void elevator_put(struct elevator_type *e)
 141 {
 142         module_put(e->elevator_owner);
 143 }
 144
 145 static struct elevator_type *elevator_get(const char *name)
 146 {
 147         struct elevator_type *e;
 148
 149         spin_lock(&elv_list_lock);
 150
 151         e = elevator_find(name);
 152         if (!e) {
 153                 char elv[ELV_NAME_MAX + strlen("-iosched")];
 154
 155                 spin_unlock(&elv_list_lock);
 156
 157                 if (!strcmp(name, "anticipatory"))
 158                         sprintf(elv, "as-iosched");
 159                 else
 160                         sprintf(elv, "%s-iosched", name);
 161
 162                 request_module("%s", elv);
 163                 spin_lock(&elv_list_lock);
 164                 e = elevator_find(name);
 165         }
 166
 167         if (e && !try_module_get(e->elevator_owner))
 168                 e = NULL;
 169
 170         spin_unlock(&elv_list_lock);
 171
 172         return e;
 173 }
 174
 175 static void *elevator_init_queue(struct request_queue *q,
 176                                  struct elevator_queue *eq)
 177 {
 178         return eq->ops->elevator_init_fn(q);
 179 }
 180
 181 static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
 182                            void *data)
 183 {
 184         q->elevator = eq;
 185         eq->elevator_data = data;
 186 }
 187
 188 static char chosen_elevator[16];
 189
 190 static int __init elevator_setup(char *str)
 191 {
 192         /*
 193          * Be backwards-compatible with previous kernels, so users
 194          * won't get the wrong elevator.
 195          */
 196         if (!strcmp(str, "as"))
 197                 strcpy(chosen_elevator, "anticipatory");
 198         else
 199                 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 200         return 1;
 201 }
 202
 203 __setup("elevator=", elevator_setup);
 204
 205 static struct kobj_type elv_ktype;
 206
 207 static struct elevator_queue *elevator_alloc(struct request_queue *q,
 208                                   struct elevator_type *e)
 209 {
 210         struct elevator_queue *eq;
 211         int i;
 212
 213         eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node);
 214         if (unlikely(!eq))
 215                 goto err;
 216
 217         eq->ops = &e->ops;
 218         eq->elevator_type = e;
 219         kobject_init(&eq->kobj, &elv_ktype);
 220         mutex_init(&eq->sysfs_lock);
 221
 222         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 223                                         GFP_KERNEL, q->node);
 224         if (!eq->hash)
 225                 goto err;
 226
 227         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 228                 INIT_HLIST_HEAD(&eq->hash[i]);
 229
 230         return eq;
 231 err:
 232         kfree(eq);
 233         elevator_put(e);
 234         return NULL;
 235 }
 236
 237 static void elevator_release(struct kobject *kobj)
 238 {
 239         struct elevator_queue *e;
 240
 241         e = container_of(kobj, struct elevator_queue, kobj);
 242         elevator_put(e->elevator_type);
 243         kfree(e->hash);
 244         kfree(e);
 245 }
 246
 247 int elevator_init(struct request_queue *q, char *name)
 248 {
 249         struct elevator_type *e = NULL;
 250         struct elevator_queue *eq;
 251         int ret = 0;
 252         void *data;
 253
 254         INIT_LIST_HEAD(&q->queue_head);
 255         q->last_merge = NULL;
 256         q->end_sector = 0;
 257         q->boundary_rq = NULL;
 258
 259         if (name) {
 260                 e = elevator_get(name);
 261                 if (!e)
 262                         return -EINVAL;
 263         }
 264
 265         if (!e && *chosen_elevator) {
 266                 e = elevator_get(chosen_elevator);
 267                 if (!e)
 268                         printk(KERN_ERR "I/O scheduler %s not found\n",
 269                                                         chosen_elevator);
 270         }
 271
 272         if (!e) {
 273                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
 274                 if (!e) {
 275                         printk(KERN_ERR
 276                                 "Default I/O scheduler not found. " \
 277                                 "Using noop.\n");
 278                         e = elevator_get("noop");
 279                 }
 280         }
 281
 282         eq = elevator_alloc(q, e);
 283         if (!eq)
 284                 return -ENOMEM;
 285
 286         data = elevator_init_queue(q, eq);
 287         if (!data) {
 288                 kobject_put(&eq->kobj);
 289                 return -ENOMEM;
 290         }
 291
 292         elevator_attach(q, eq, data);
 293         return ret;
 294 }
 295 EXPORT_SYMBOL(elevator_init);
 296
 297 void elevator_exit(struct elevator_queue *e)
 298 {
 299         mutex_lock(&e->sysfs_lock);
 300         if (e->ops->elevator_exit_fn)
 301                 e->ops->elevator_exit_fn(e);
 302         e->ops = NULL;
 303         mutex_unlock(&e->sysfs_lock);
 304
 305         kobject_put(&e->kobj);
 306 }
 307 EXPORT_SYMBOL(elevator_exit);
 308
 309 static void elv_activate_rq(struct request_queue *q, struct request *rq)
 310 {
 311         struct elevator_queue *e = q->elevator;
 312
 313         if (e->ops->elevator_activate_req_fn)
 314                 e->ops->elevator_activate_req_fn(q, rq);
 315 }
 316
 317 static void elv_deactivate_rq(struct request_queue *q, struct request *rq)
 318 {
 319         struct elevator_queue *e = q->elevator;
 320
 321         if (e->ops->elevator_deactivate_req_fn)
 322                 e->ops->elevator_deactivate_req_fn(q, rq);
 323 }
 324
 325 static inline void __elv_rqhash_del(struct request *rq)
 326 {
 327         hlist_del_init(&rq->hash);
 328 }
 329
 330 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 331 {
 332         if (ELV_ON_HASH(rq))
 333                 __elv_rqhash_del(rq);
 334 }
 335
 336 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 337 {
 338         struct elevator_queue *e = q->elevator;
 339
 340         BUG_ON(ELV_ON_HASH(rq));
 341         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 342 }
 343
 344 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 345 {
 346         __elv_rqhash_del(rq);
 347         elv_rqhash_add(q, rq);
 348 }
 349
 350 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 351 {
 352         struct elevator_queue *e = q->elevator;
 353         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 354         struct hlist_node *entry, *next;
 355         struct request *rq;
 356
 357         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 358                 BUG_ON(!ELV_ON_HASH(rq));
 359
 360                 if (unlikely(!rq_mergeable(rq))) {
 361                         __elv_rqhash_del(rq);
 362                         continue;
 363                 }
 364
 365                 if (rq_hash_key(rq) == offset)
 366                         return rq;
 367         }
 368
 369         return NULL;
 370 }
 371
 372 /*
 373  * RB-tree support functions for inserting/lookup/removal of requests
 374  * in a sorted RB tree.
 375  */
 376 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 377 {
 378         struct rb_node **p = &root->rb_node;
 379         struct rb_node *parent = NULL;
 380         struct request *__rq;
 381
 382         while (*p) {
 383                 parent = *p;
 384                 __rq = rb_entry(parent, struct request, rb_node);
 385
 386                 if (rq->sector < __rq->sector)
 387                         p = &(*p)->rb_left;
 388                 else if (rq->sector > __rq->sector)
 389                         p = &(*p)->rb_right;
 390                 else
 391                         return __rq;
 392         }
 393
 394         rb_link_node(&rq->rb_node, parent, p);
 395         rb_insert_color(&rq->rb_node, root);
 396         return NULL;
 397 }
 398 EXPORT_SYMBOL(elv_rb_add);
 399
 400 void elv_rb_del(struct rb_root *root, struct request *rq)
 401 {
 402         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 403         rb_erase(&rq->rb_node, root);
 404         RB_CLEAR_NODE(&rq->rb_node);
 405 }
 406 EXPORT_SYMBOL(elv_rb_del);
 407
 408 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 409 {
 410         struct rb_node *n = root->rb_node;
 411         struct request *rq;
 412
 413         while (n) {
 414                 rq = rb_entry(n, struct request, rb_node);
 415
 416                 if (sector < rq->sector)
 417                         n = n->rb_left;
 418                 else if (sector > rq->sector)
 419                         n = n->rb_right;
 420                 else
 421                         return rq;
 422         }
 423
 424         return NULL;
 425 }
 426 EXPORT_SYMBOL(elv_rb_find);
 427
 428 /*
 429  * Insert rq into dispatch queue of q.  Queue lock must be held on
 430  * entry.  rq is sort instead into the dispatch queue. To be used by
 431  * specific elevators.
 432  */
 433 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 434 {
 435         sector_t boundary;
 436         struct list_head *entry;
 437         int stop_flags;
 438
 439         if (q->last_merge == rq)
 440                 q->last_merge = NULL;
 441
 442         elv_rqhash_del(q, rq);
 443
 444         q->nr_sorted--;
 445
 446         boundary = q->end_sector;
 447         stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
 448         list_for_each_prev(entry, &q->queue_head) {
 449                 struct request *pos = list_entry_rq(entry);
 450
 451                 if (blk_discard_rq(rq) != blk_discard_rq(pos))
 452                         break;
 453                 if (rq_data_dir(rq) != rq_data_dir(pos))
 454                         break;
 455                 if (pos->cmd_flags & stop_flags)
 456                         break;
 457                 if (rq->sector >= boundary) {
 458                         if (pos->sector < boundary)
 459                                 continue;
 460                 } else {
 461                         if (pos->sector >= boundary)
 462                                 break;
 463                 }
 464                 if (rq->sector >= pos->sector)
 465                         break;
 466         }
 467
 468         list_add(&rq->queuelist, entry);
 469 }
 470 EXPORT_SYMBOL(elv_dispatch_sort);
 471
 472 /*
 473  * Insert rq into dispatch queue of q.  Queue lock must be held on
 474  * entry.  rq is added to the back of the dispatch queue. To be used by
 475  * specific elevators.
 476  */
 477 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 478 {
 479         if (q->last_merge == rq)
 480                 q->last_merge = NULL;
 481
 482         elv_rqhash_del(q, rq);
 483
 484         q->nr_sorted--;
 485
 486         q->end_sector = rq_end_sector(rq);
 487         q->boundary_rq = rq;
 488         list_add_tail(&rq->queuelist, &q->queue_head);
 489 }
 490 EXPORT_SYMBOL(elv_dispatch_add_tail);
 491
 492 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 493 {
 494         struct elevator_queue *e = q->elevator;
 495         struct request *__rq;
 496         int ret;
 497
 498         /*
 499          * First try one-hit cache.
 500          */
 501         if (q->last_merge) {
 502                 ret = elv_try_merge(q->last_merge, bio);
 503                 if (ret != ELEVATOR_NO_MERGE) {
 504                         *req = q->last_merge;
 505                         return ret;
 506                 }
 507         }
 508
 509         if (blk_queue_nomerges(q))
 510                 return ELEVATOR_NO_MERGE;
 511
 512         /*
 513          * See if our hash lookup can find a potential backmerge.
 514          */
 515         __rq = elv_rqhash_find(q, bio->bi_sector);
 516         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 517                 *req = __rq;
 518                 return ELEVATOR_BACK_MERGE;
 519         }
 520
 521         if (e->ops->elevator_merge_fn)
 522                 return e->ops->elevator_merge_fn(q, req, bio);
 523
 524         return ELEVATOR_NO_MERGE;
 525 }
 526
 527 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 528 {
 529         struct elevator_queue *e = q->elevator;
 530
 531         if (e->ops->elevator_merged_fn)
 532                 e->ops->elevator_merged_fn(q, rq, type);
 533
 534         if (type == ELEVATOR_BACK_MERGE)
 535                 elv_rqhash_reposition(q, rq);
 536
 537         q->last_merge = rq;
 538 }
 539
 540 void elv_merge_requests(struct request_queue *q, struct request *rq,
 541                              struct request *next)
 542 {
 543         struct elevator_queue *e = q->elevator;
 544
 545         if (e->ops->elevator_merge_req_fn)
 546                 e->ops->elevator_merge_req_fn(q, rq, next);
 547
 548         elv_rqhash_reposition(q, rq);
 549         elv_rqhash_del(q, next);
 550
 551         q->nr_sorted--;
 552         q->last_merge = rq;
 553 }
 554
 555 void elv_requeue_request(struct request_queue *q, struct request *rq)
 556 {
 557         /*
 558          * it already went through dequeue, we need to decrement the
 559          * in_flight count again
 560          */
 561         if (blk_account_rq(rq)) {
 562                 q->in_flight--;
 563                 if (blk_sorted_rq(rq))
 564                         elv_deactivate_rq(q, rq);
 565         }
 566
 567         rq->cmd_flags &= ~REQ_STARTED;
 568
 569         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 570 }
 571
 572 void elv_drain_elevator(struct request_queue *q)
 573 {
 574         static int printed;
 575         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 576                 ;
 577         if (q->nr_sorted == 0)
 578                 return;
 579         if (printed++ < 10) {
 580                 printk(KERN_ERR "%s: forced dispatching is broken "
 581                        "(nr_sorted=%u), please report this\n",
 582                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 583         }
 584 }
 585
 586 /*
 587  * Call with queue lock held, interrupts disabled
 588  */
 589 void elv_quiesce_start(struct request_queue *q)
 590 {
 591         queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
 592
 593         /*
 594          * make sure we don't have any requests in flight
 595          */
 596         elv_drain_elevator(q);
 597         while (q->rq.elvpriv) {
 598                 blk_start_queueing(q);
 599                 spin_unlock_irq(q->queue_lock);
 600                 msleep(10);
 601                 spin_lock_irq(q->queue_lock);
 602                 elv_drain_elevator(q);
 603         }
 604 }
 605
 606 void elv_quiesce_end(struct request_queue *q)
 607 {
 608         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
 609 }
 610
 611 void elv_insert(struct request_queue *q, struct request *rq, int where)
 612 {
 613         struct list_head *pos;
 614         unsigned ordseq;
 615         int unplug_it = 1;
 616
 617         trace_block_rq_insert(q, rq);
 618
 619         rq->q = q;
 620
 621         switch (where) {
 622         case ELEVATOR_INSERT_FRONT:
 623                 rq->cmd_flags |= REQ_SOFTBARRIER;
 624
 625                 list_add(&rq->queuelist, &q->queue_head);
 626                 break;
 627
 628         case ELEVATOR_INSERT_BACK:
 629                 rq->cmd_flags |= REQ_SOFTBARRIER;
 630                 elv_drain_elevator(q);
 631                 list_add_tail(&rq->queuelist, &q->queue_head);
 632                 /*
 633                  * We kick the queue here for the following reasons.
 634                  * - The elevator might have returned NULL previously
 635                  *   to delay requests and returned them now.  As the
 636                  *   queue wasn't empty before this request, ll_rw_blk
 637                  *   won't run the queue on return, resulting in hang.
 638                  * - Usually, back inserted requests won't be merged
 639                  *   with anything.  There's no point in delaying queue
 640                  *   processing.
 641                  */
 642                 blk_remove_plug(q);
 643                 blk_start_queueing(q);
 644                 break;
 645
 646         case ELEVATOR_INSERT_SORT:
 647                 BUG_ON(!blk_fs_request(rq) && !blk_discard_rq(rq));
 648                 rq->cmd_flags |= REQ_SORTED;
 649                 q->nr_sorted++;
 650                 if (rq_mergeable(rq)) {
 651                         elv_rqhash_add(q, rq);
 652                         if (!q->last_merge)
 653                                 q->last_merge = rq;
 654                 }
 655
 656                 /*
 657                  * Some ioscheds (cfq) run q->request_fn directly, so
 658                  * rq cannot be accessed after calling
 659                  * elevator_add_req_fn.
 660                  */
 661                 q->elevator->ops->elevator_add_req_fn(q, rq);
 662                 break;
 663
 664         case ELEVATOR_INSERT_REQUEUE:
 665                 /*
 666                  * If ordered flush isn't in progress, we do front
 667                  * insertion; otherwise, requests should be requeued
 668                  * in ordseq order.
 669                  */
 670                 rq->cmd_flags |= REQ_SOFTBARRIER;
 671
 672                 /*
 673                  * Most requeues happen because of a busy condition,
 674                  * don't force unplug of the queue for that case.
 675                  */
 676                 unplug_it = 0;
 677
 678                 if (q->ordseq == 0) {
 679                         list_add(&rq->queuelist, &q->queue_head);
 680                         break;
 681                 }
 682
 683                 ordseq = blk_ordered_req_seq(rq);
 684
 685                 list_for_each(pos, &q->queue_head) {
 686                         struct request *pos_rq = list_entry_rq(pos);
 687                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 688                                 break;
 689                 }
 690
 691                 list_add_tail(&rq->queuelist, pos);
 692                 break;
 693
 694         default:
 695                 printk(KERN_ERR "%s: bad insertion point %d\n",
 696                        __func__, where);
 697                 BUG();
 698         }
 699
 700         if (unplug_it && blk_queue_plugged(q)) {
 701                 int nrq = q->rq.count[BLK_RW_SYNC] + q->rq.count[BLK_RW_ASYNC]
 702                         - q->in_flight;
 703
 704                 if (nrq >= q->unplug_thresh)
 705                         __generic_unplug_device(q);
 706         }
 707 }
 708
 709 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
 710                        int plug)
 711 {
 712         if (q->ordcolor)
 713                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 714
 715         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 716                 /*
 717                  * toggle ordered color
 718                  */
 719                 if (blk_barrier_rq(rq))
 720                         q->ordcolor ^= 1;
 721
 722                 /*
 723                  * barriers implicitly indicate back insertion
 724                  */
 725                 if (where == ELEVATOR_INSERT_SORT)
 726                         where = ELEVATOR_INSERT_BACK;
 727
 728                 /*
 729                  * this request is scheduling boundary, update
 730                  * end_sector
 731                  */
 732                 if (blk_fs_request(rq) || blk_discard_rq(rq)) {
 733                         q->end_sector = rq_end_sector(rq);
 734                         q->boundary_rq = rq;
 735                 }
 736         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 737                     where == ELEVATOR_INSERT_SORT)
 738                 where = ELEVATOR_INSERT_BACK;
 739
 740         if (plug)
 741                 blk_plug_device(q);
 742
 743         elv_insert(q, rq, where);
 744 }
 745 EXPORT_SYMBOL(__elv_add_request);
 746
 747 void elv_add_request(struct request_queue *q, struct request *rq, int where,
 748                      int plug)
 749 {
 750         unsigned long flags;
 751
 752         spin_lock_irqsave(q->queue_lock, flags);
 753         __elv_add_request(q, rq, where, plug);
 754         spin_unlock_irqrestore(q->queue_lock, flags);
 755 }
 756 EXPORT_SYMBOL(elv_add_request);
 757
 758 static inline struct request *__elv_next_request(struct request_queue *q)
 759 {
 760         struct request *rq;
 761
 762         while (1) {
 763                 while (!list_empty(&q->queue_head)) {
 764                         rq = list_entry_rq(q->queue_head.next);
 765                         if (blk_do_ordered(q, &rq))
 766                                 return rq;
 767                 }
 768
 769                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
 770                         return NULL;
 771         }
 772 }
 773
 774 struct request *elv_next_request(struct request_queue *q)
 775 {
 776         struct request *rq;
 777         int ret;
 778
 779         while ((rq = __elv_next_request(q)) != NULL) {
 780                 if (!(rq->cmd_flags & REQ_STARTED)) {
 781                         /*
 782                          * This is the first time the device driver
 783                          * sees this request (possibly after
 784                          * requeueing).  Notify IO scheduler.
 785                          */
 786                         if (blk_sorted_rq(rq))
 787                                 elv_activate_rq(q, rq);
 788
 789                         /*
 790                          * just mark as started even if we don't start
 791                          * it, a request that has been delayed should
 792                          * not be passed by new incoming requests
 793                          */
 794                         rq->cmd_flags |= REQ_STARTED;
 795                         trace_block_rq_issue(q, rq);
 796                 }
 797
 798                 if (!q->boundary_rq || q->boundary_rq == rq) {
 799                         q->end_sector = rq_end_sector(rq);
 800                         q->boundary_rq = NULL;
 801                 }
 802
 803                 if (rq->cmd_flags & REQ_DONTPREP)
 804                         break;
 805
 806                 if (q->dma_drain_size && rq->data_len) {
 807                         /*
 808                          * make sure space for the drain appears we
 809                          * know we can do this because max_hw_segments
 810                          * has been adjusted to be one fewer than the
 811                          * device can handle
 812                          */
 813                         rq->nr_phys_segments++;
 814                 }
 815
 816                 if (!q->prep_rq_fn)
 817                         break;
 818
 819                 ret = q->prep_rq_fn(q, rq);
 820                 if (ret == BLKPREP_OK) {
 821                         break;
 822                 } else if (ret == BLKPREP_DEFER) {
 823                         /*
 824                          * the request may have been (partially) prepped.
 825                          * we need to keep this request in the front to
 826                          * avoid resource deadlock.  REQ_STARTED will
 827                          * prevent other fs requests from passing this one.
 828                          */
 829                         if (q->dma_drain_size && rq->data_len &&
 830                             !(rq->cmd_flags & REQ_DONTPREP)) {
 831                                 /*
 832                                  * remove the space for the drain we added
 833                                  * so that we don't add it again
 834                                  */
 835                                 --rq->nr_phys_segments;
 836                         }
 837
 838                         rq = NULL;
 839                         break;
 840                 } else if (ret == BLKPREP_KILL) {
 841                         rq->cmd_flags |= REQ_QUIET;
 842                         __blk_end_request(rq, -EIO, blk_rq_bytes(rq));
 843                 } else {
 844                         printk(KERN_ERR "%s: bad return=%d\n", __func__, ret);
 845                         break;
 846                 }
 847         }
 848
 849         return rq;
 850 }
 851 EXPORT_SYMBOL(elv_next_request);
 852
 853 void elv_dequeue_request(struct request_queue *q, struct request *rq)
 854 {
 855         BUG_ON(list_empty(&rq->queuelist));
 856         BUG_ON(ELV_ON_HASH(rq));
 857
 858         list_del_init(&rq->queuelist);
 859
 860         /*
 861          * the time frame between a request being removed from the lists
 862          * and to it is freed is accounted as io that is in progress at
 863          * the driver side.
 864          */
 865         if (blk_account_rq(rq))
 866                 q->in_flight++;
 867 }
 868
 869 int elv_queue_empty(struct request_queue *q)
 870 {
 871         struct elevator_queue *e = q->elevator;
 872
 873         if (!list_empty(&q->queue_head))
 874                 return 0;
 875
 876         if (e->ops->elevator_queue_empty_fn)
 877                 return e->ops->elevator_queue_empty_fn(q);
 878
 879         return 1;
 880 }
 881 EXPORT_SYMBOL(elv_queue_empty);
 882
 883 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 884 {
 885         struct elevator_queue *e = q->elevator;
 886
 887         if (e->ops->elevator_latter_req_fn)
 888                 return e->ops->elevator_latter_req_fn(q, rq);
 889         return NULL;
 890 }
 891
 892 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 893 {
 894         struct elevator_queue *e = q->elevator;
 895
 896         if (e->ops->elevator_former_req_fn)
 897                 return e->ops->elevator_former_req_fn(q, rq);
 898         return NULL;
 899 }
 900
 901 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 902 {
 903         struct elevator_queue *e = q->elevator;
 904
 905         if (e->ops->elevator_set_req_fn)
 906                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 907
 908         rq->elevator_private = NULL;
 909         return 0;
 910 }
 911
 912 void elv_put_request(struct request_queue *q, struct request *rq)
 913 {
 914         struct elevator_queue *e = q->elevator;
 915
 916         if (e->ops->elevator_put_req_fn)
 917                 e->ops->elevator_put_req_fn(rq);
 918 }
 919
 920 int elv_may_queue(struct request_queue *q, int rw)
 921 {
 922         struct elevator_queue *e = q->elevator;
 923
 924         if (e->ops->elevator_may_queue_fn)
 925                 return e->ops->elevator_may_queue_fn(q, rw);
 926
 927         return ELV_MQUEUE_MAY;
 928 }
 929
 930 void elv_abort_queue(struct request_queue *q)
 931 {
 932         struct request *rq;
 933
 934         while (!list_empty(&q->queue_head)) {
 935                 rq = list_entry_rq(q->queue_head.next);
 936                 rq->cmd_flags |= REQ_QUIET;
 937                 trace_block_rq_abort(q, rq);
 938                 __blk_end_request(rq, -EIO, blk_rq_bytes(rq));
 939         }
 940 }
 941 EXPORT_SYMBOL(elv_abort_queue);
 942
 943 void elv_completed_request(struct request_queue *q, struct request *rq)
 944 {
 945         struct elevator_queue *e = q->elevator;
 946
 947         /*
 948          * request is released from the driver, io must be done
 949          */
 950         if (blk_account_rq(rq)) {
 951                 q->in_flight--;
 952                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 953                         e->ops->elevator_completed_req_fn(q, rq);
 954         }
 955
 956         /*
 957          * Check if the queue is waiting for fs requests to be
 958          * drained for flush sequence.
 959          */
 960         if (unlikely(q->ordseq)) {
 961                 struct request *next = NULL;
 962
 963                 if (!list_empty(&q->queue_head))
 964                         next = list_entry_rq(q->queue_head.next);
 965
 966                 if (!q->in_flight &&
 967                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 968                     (!next || blk_ordered_req_seq(next) > QUEUE_ORDSEQ_DRAIN)) {
 969                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 970                         blk_start_queueing(q);
 971                 }
 972         }
 973 }
 974
 975 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 976
 977 static ssize_t
 978 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 979 {
 980         struct elv_fs_entry *entry = to_elv(attr);
 981         struct elevator_queue *e;
 982         ssize_t error;
 983
 984         if (!entry->show)
 985                 return -EIO;
 986
 987         e = container_of(kobj, struct elevator_queue, kobj);
 988         mutex_lock(&e->sysfs_lock);
 989         error = e->ops ? entry->show(e, page) : -ENOENT;
 990         mutex_unlock(&e->sysfs_lock);
 991         return error;
 992 }
 993
 994 static ssize_t
 995 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 996                const char *page, size_t length)
 997 {
 998         struct elv_fs_entry *entry = to_elv(attr);
 999         struct elevator_queue *e;
1000         ssize_t error;
1001
1002         if (!entry->store)
1003                 return -EIO;
1004
1005         e = container_of(kobj, struct elevator_queue, kobj);
1006         mutex_lock(&e->sysfs_lock);
1007         error = e->ops ? entry->store(e, page, length) : -ENOENT;
1008         mutex_unlock(&e->sysfs_lock);
1009         return error;
1010 }
1011
1012 static struct sysfs_ops elv_sysfs_ops = {
1013         .show   = elv_attr_show,
1014         .store  = elv_attr_store,
1015 };
1016
1017 static struct kobj_type elv_ktype = {
1018         .sysfs_ops      = &elv_sysfs_ops,
1019         .release        = elevator_release,
1020 };
1021
1022 int elv_register_queue(struct request_queue *q)
1023 {
1024         struct elevator_queue *e = q->elevator;
1025         int error;
1026
1027         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
1028         if (!error) {
1029                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
1030                 if (attr) {
1031                         while (attr->attr.name) {
1032                                 if (sysfs_create_file(&e->kobj, &attr->attr))
1033                                         break;
1034                                 attr++;
1035                         }
1036                 }
1037                 kobject_uevent(&e->kobj, KOBJ_ADD);
1038         }
1039         return error;
1040 }
1041
1042 static void __elv_unregister_queue(struct elevator_queue *e)
1043 {
1044         kobject_uevent(&e->kobj, KOBJ_REMOVE);
1045         kobject_del(&e->kobj);
1046 }
1047
1048 void elv_unregister_queue(struct request_queue *q)
1049 {
1050         if (q)
1051                 __elv_unregister_queue(q->elevator);
1052 }
1053
1054 void elv_register(struct elevator_type *e)
1055 {
1056         char *def = "";
1057
1058         spin_lock(&elv_list_lock);
1059         BUG_ON(elevator_find(e->elevator_name));
1060         list_add_tail(&e->list, &elv_list);
1061         spin_unlock(&elv_list_lock);
1062
1063         if (!strcmp(e->elevator_name, chosen_elevator) ||
1064                         (!*chosen_elevator &&
1065                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
1066                                 def = " (default)";
1067
1068         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
1069                                                                 def);
1070 }
1071 EXPORT_SYMBOL_GPL(elv_register);
1072
1073 void elv_unregister(struct elevator_type *e)
1074 {
1075         struct task_struct *g, *p;
1076
1077         /*
1078          * Iterate every thread in the process to remove the io contexts.
1079          */
1080         if (e->ops.trim) {
1081                 read_lock(&tasklist_lock);
1082                 do_each_thread(g, p) {
1083                         task_lock(p);
1084                         if (p->io_context)
1085                                 e->ops.trim(p->io_context);
1086                         task_unlock(p);
1087                 } while_each_thread(g, p);
1088                 read_unlock(&tasklist_lock);
1089         }
1090
1091         spin_lock(&elv_list_lock);
1092         list_del_init(&e->list);
1093         spin_unlock(&elv_list_lock);
1094 }
1095 EXPORT_SYMBOL_GPL(elv_unregister);
1096
1097 /*
1098  * switch to new_e io scheduler. be careful not to introduce deadlocks -
1099  * we don't free the old io scheduler, before we have allocated what we
1100  * need for the new one. this way we have a chance of going back to the old
1101  * one, if the new one fails init for some reason.
1102  */
1103 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1104 {
1105         struct elevator_queue *old_elevator, *e;
1106         void *data;
1107
1108         /*
1109          * Allocate new elevator
1110          */
1111         e = elevator_alloc(q, new_e);
1112         if (!e)
1113                 return 0;
1114
1115         data = elevator_init_queue(q, e);
1116         if (!data) {
1117                 kobject_put(&e->kobj);
1118                 return 0;
1119         }
1120
1121         /*
1122          * Turn on BYPASS and drain all requests w/ elevator private data
1123          */
1124         spin_lock_irq(q->queue_lock);
1125         elv_quiesce_start(q);
1126
1127         /*
1128          * Remember old elevator.
1129          */
1130         old_elevator = q->elevator;
1131
1132         /*
1133          * attach and start new elevator
1134          */
1135         elevator_attach(q, e, data);
1136
1137         spin_unlock_irq(q->queue_lock);
1138
1139         __elv_unregister_queue(old_elevator);
1140
1141         if (elv_register_queue(q))
1142                 goto fail_register;
1143
1144         /*
1145          * finally exit old elevator and turn off BYPASS.
1146          */
1147         elevator_exit(old_elevator);
1148         spin_lock_irq(q->queue_lock);
1149         elv_quiesce_end(q);
1150         spin_unlock_irq(q->queue_lock);
1151
1152         blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
1153
1154         return 1;
1155
1156 fail_register:
1157         /*
1158          * switch failed, exit the new io scheduler and reattach the old
1159          * one again (along with re-adding the sysfs dir)
1160          */
1161         elevator_exit(e);
1162         q->elevator = old_elevator;
1163         elv_register_queue(q);
1164
1165         spin_lock_irq(q->queue_lock);
1166         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1167         spin_unlock_irq(q->queue_lock);
1168
1169         return 0;
1170 }
1171
1172 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1173                           size_t count)
1174 {
1175         char elevator_name[ELV_NAME_MAX];
1176         struct elevator_type *e;
1177
1178         strlcpy(elevator_name, name, sizeof(elevator_name));
1179         strstrip(elevator_name);
1180
1181         e = elevator_get(elevator_name);
1182         if (!e) {
1183                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1184                 return -EINVAL;
1185         }
1186
1187         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1188                 elevator_put(e);
1189                 return count;
1190         }
1191
1192         if (!elevator_switch(q, e))
1193                 printk(KERN_ERR "elevator: switch to %s failed\n",
1194                                                         elevator_name);
1195         return count;
1196 }
1197
1198 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1199 {
1200         struct elevator_queue *e = q->elevator;
1201         struct elevator_type *elv = e->elevator_type;
1202         struct elevator_type *__e;
1203         int len = 0;
1204
1205         spin_lock(&elv_list_lock);
1206         list_for_each_entry(__e, &elv_list, list) {
1207                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1208                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1209                 else
1210                         len += sprintf(name+len, "%s ", __e->elevator_name);
1211         }
1212         spin_unlock(&elv_list_lock);
1213
1214         len += sprintf(len+name, "\n");
1215         return len;
1216 }
1217
1218 struct request *elv_rb_former_request(struct request_queue *q,
1219                                       struct request *rq)
1220 {
1221         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1222
1223         if (rbprev)
1224                 return rb_entry_rq(rbprev);
1225
1226         return NULL;
1227 }
1228 EXPORT_SYMBOL(elv_rb_former_request);
1229
1230 struct request *elv_rb_latter_request(struct request_queue *q,
1231                                       struct request *rq)
1232 {
1233         struct rb_node *rbnext = rb_next(&rq->rb_node);
1234
1235         if (rbnext)
1236                 return rb_entry_rq(rbnext);
1237
1238         return NULL;
1239 }
1240 EXPORT_SYMBOL(elv_rb_latter_request);