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 DEFINE_TRACE(block_rq_abort);
  46
  47 /*
  48  * Merge hash stuff.
  49  */
  50 static const int elv_hash_shift = 6;
  51 #define ELV_HASH_BLOCK(sec)     ((sec) >> 3)
  52 #define ELV_HASH_FN(sec)        \
  53                 (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
  54 #define ELV_HASH_ENTRIES        (1 << elv_hash_shift)
  55 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
  56 #define ELV_ON_HASH(rq)         (!hlist_unhashed(&(rq)->hash))
  57
  58 DEFINE_TRACE(block_rq_insert);
  59 DEFINE_TRACE(block_rq_issue);
  60
  61 /*
  62  * Query io scheduler to see if the current process issuing bio may be
  63  * merged with rq.
  64  */
  65 static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
  66 {
  67         struct request_queue *q = rq->q;
  68         struct elevator_queue *e = q->elevator;
  69
  70         if (e->ops->elevator_allow_merge_fn)
  71                 return e->ops->elevator_allow_merge_fn(q, rq, bio);
  72
  73         return 1;
  74 }
  75
  76 /*
  77  * can we safely merge with this request?
  78  */
  79 int elv_rq_merge_ok(struct request *rq, struct bio *bio)
  80 {
  81         if (!rq_mergeable(rq))
  82                 return 0;
  83
  84         /*
  85          * Don't merge file system requests and discard requests
  86          */
  87         if (bio_discard(bio) != bio_discard(rq->bio))
  88                 return 0;
  89
  90         /*
  91          * different data direction or already started, don't merge
  92          */
  93         if (bio_data_dir(bio) != rq_data_dir(rq))
  94                 return 0;
  95
  96         /*
  97          * must be same device and not a special request
  98          */
  99         if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
 100                 return 0;
 101
 102         /*
 103          * only merge integrity protected bio into ditto rq
 104          */
 105         if (bio_integrity(bio) != blk_integrity_rq(rq))
 106                 return 0;
 107
 108         if (!elv_iosched_allow_merge(rq, bio))
 109                 return 0;
 110
 111         return 1;
 112 }
 113 EXPORT_SYMBOL(elv_rq_merge_ok);
 114
 115 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
 116 {
 117         int ret = ELEVATOR_NO_MERGE;
 118
 119         /*
 120          * we can merge and sequence is ok, check if it's possible
 121          */
 122         if (elv_rq_merge_ok(__rq, bio)) {
 123                 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
 124                         ret = ELEVATOR_BACK_MERGE;
 125                 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
 126                         ret = ELEVATOR_FRONT_MERGE;
 127         }
 128
 129         return ret;
 130 }
 131
 132 static struct elevator_type *elevator_find(const char *name)
 133 {
 134         struct elevator_type *e;
 135
 136         list_for_each_entry(e, &elv_list, list) {
 137                 if (!strcmp(e->elevator_name, name))
 138                         return e;
 139         }
 140
 141         return NULL;
 142 }
 143
 144 static void elevator_put(struct elevator_type *e)
 145 {
 146         module_put(e->elevator_owner);
 147 }
 148
 149 static struct elevator_type *elevator_get(const char *name)
 150 {
 151         struct elevator_type *e;
 152
 153         spin_lock(&elv_list_lock);
 154
 155         e = elevator_find(name);
 156         if (!e) {
 157                 char elv[ELV_NAME_MAX + strlen("-iosched")];
 158
 159                 spin_unlock(&elv_list_lock);
 160
 161                 if (!strcmp(name, "anticipatory"))
 162                         sprintf(elv, "as-iosched");
 163                 else
 164                         sprintf(elv, "%s-iosched", name);
 165
 166                 request_module("%s", elv);
 167                 spin_lock(&elv_list_lock);
 168                 e = elevator_find(name);
 169         }
 170
 171         if (e && !try_module_get(e->elevator_owner))
 172                 e = NULL;
 173
 174         spin_unlock(&elv_list_lock);
 175
 176         return e;
 177 }
 178
 179 static void *elevator_init_queue(struct request_queue *q,
 180                                  struct elevator_queue *eq)
 181 {
 182         return eq->ops->elevator_init_fn(q);
 183 }
 184
 185 static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
 186                            void *data)
 187 {
 188         q->elevator = eq;
 189         eq->elevator_data = data;
 190 }
 191
 192 static char chosen_elevator[16];
 193
 194 static int __init elevator_setup(char *str)
 195 {
 196         /*
 197          * Be backwards-compatible with previous kernels, so users
 198          * won't get the wrong elevator.
 199          */
 200         if (!strcmp(str, "as"))
 201                 strcpy(chosen_elevator, "anticipatory");
 202         else
 203                 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 204         return 1;
 205 }
 206
 207 __setup("elevator=", elevator_setup);
 208
 209 static struct kobj_type elv_ktype;
 210
 211 static struct elevator_queue *elevator_alloc(struct request_queue *q,
 212                                   struct elevator_type *e)
 213 {
 214         struct elevator_queue *eq;
 215         int i;
 216
 217         eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node);
 218         if (unlikely(!eq))
 219                 goto err;
 220
 221         eq->ops = &e->ops;
 222         eq->elevator_type = e;
 223         kobject_init(&eq->kobj, &elv_ktype);
 224         mutex_init(&eq->sysfs_lock);
 225
 226         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 227                                         GFP_KERNEL, q->node);
 228         if (!eq->hash)
 229                 goto err;
 230
 231         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 232                 INIT_HLIST_HEAD(&eq->hash[i]);
 233
 234         return eq;
 235 err:
 236         kfree(eq);
 237         elevator_put(e);
 238         return NULL;
 239 }
 240
 241 static void elevator_release(struct kobject *kobj)
 242 {
 243         struct elevator_queue *e;
 244
 245         e = container_of(kobj, struct elevator_queue, kobj);
 246         elevator_put(e->elevator_type);
 247         kfree(e->hash);
 248         kfree(e);
 249 }
 250
 251 int elevator_init(struct request_queue *q, char *name)
 252 {
 253         struct elevator_type *e = NULL;
 254         struct elevator_queue *eq;
 255         int ret = 0;
 256         void *data;
 257
 258         INIT_LIST_HEAD(&q->queue_head);
 259         q->last_merge = NULL;
 260         q->end_sector = 0;
 261         q->boundary_rq = NULL;
 262
 263         if (name) {
 264                 e = elevator_get(name);
 265                 if (!e)
 266                         return -EINVAL;
 267         }
 268
 269         if (!e && *chosen_elevator) {
 270                 e = elevator_get(chosen_elevator);
 271                 if (!e)
 272                         printk(KERN_ERR "I/O scheduler %s not found\n",
 273                                                         chosen_elevator);
 274         }
 275
 276         if (!e) {
 277                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
 278                 if (!e) {
 279                         printk(KERN_ERR
 280                                 "Default I/O scheduler not found. " \
 281                                 "Using noop.\n");
 282                         e = elevator_get("noop");
 283                 }
 284         }
 285
 286         eq = elevator_alloc(q, e);
 287         if (!eq)
 288                 return -ENOMEM;
 289
 290         data = elevator_init_queue(q, eq);
 291         if (!data) {
 292                 kobject_put(&eq->kobj);
 293                 return -ENOMEM;
 294         }
 295
 296         elevator_attach(q, eq, data);
 297         return ret;
 298 }
 299 EXPORT_SYMBOL(elevator_init);
 300
 301 void elevator_exit(struct elevator_queue *e)
 302 {
 303         mutex_lock(&e->sysfs_lock);
 304         if (e->ops->elevator_exit_fn)
 305                 e->ops->elevator_exit_fn(e);
 306         e->ops = NULL;
 307         mutex_unlock(&e->sysfs_lock);
 308
 309         kobject_put(&e->kobj);
 310 }
 311 EXPORT_SYMBOL(elevator_exit);
 312
 313 static void elv_activate_rq(struct request_queue *q, struct request *rq)
 314 {
 315         struct elevator_queue *e = q->elevator;
 316
 317         if (e->ops->elevator_activate_req_fn)
 318                 e->ops->elevator_activate_req_fn(q, rq);
 319 }
 320
 321 static void elv_deactivate_rq(struct request_queue *q, struct request *rq)
 322 {
 323         struct elevator_queue *e = q->elevator;
 324
 325         if (e->ops->elevator_deactivate_req_fn)
 326                 e->ops->elevator_deactivate_req_fn(q, rq);
 327 }
 328
 329 static inline void __elv_rqhash_del(struct request *rq)
 330 {
 331         hlist_del_init(&rq->hash);
 332 }
 333
 334 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 335 {
 336         if (ELV_ON_HASH(rq))
 337                 __elv_rqhash_del(rq);
 338 }
 339
 340 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 341 {
 342         struct elevator_queue *e = q->elevator;
 343
 344         BUG_ON(ELV_ON_HASH(rq));
 345         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 346 }
 347
 348 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 349 {
 350         __elv_rqhash_del(rq);
 351         elv_rqhash_add(q, rq);
 352 }
 353
 354 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 355 {
 356         struct elevator_queue *e = q->elevator;
 357         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 358         struct hlist_node *entry, *next;
 359         struct request *rq;
 360
 361         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 362                 BUG_ON(!ELV_ON_HASH(rq));
 363
 364                 if (unlikely(!rq_mergeable(rq))) {
 365                         __elv_rqhash_del(rq);
 366                         continue;
 367                 }
 368
 369                 if (rq_hash_key(rq) == offset)
 370                         return rq;
 371         }
 372
 373         return NULL;
 374 }
 375
 376 /*
 377  * RB-tree support functions for inserting/lookup/removal of requests
 378  * in a sorted RB tree.
 379  */
 380 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 381 {
 382         struct rb_node **p = &root->rb_node;
 383         struct rb_node *parent = NULL;
 384         struct request *__rq;
 385
 386         while (*p) {
 387                 parent = *p;
 388                 __rq = rb_entry(parent, struct request, rb_node);
 389
 390                 if (rq->sector < __rq->sector)
 391                         p = &(*p)->rb_left;
 392                 else if (rq->sector > __rq->sector)
 393                         p = &(*p)->rb_right;
 394                 else
 395                         return __rq;
 396         }
 397
 398         rb_link_node(&rq->rb_node, parent, p);
 399         rb_insert_color(&rq->rb_node, root);
 400         return NULL;
 401 }
 402 EXPORT_SYMBOL(elv_rb_add);
 403
 404 void elv_rb_del(struct rb_root *root, struct request *rq)
 405 {
 406         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 407         rb_erase(&rq->rb_node, root);
 408         RB_CLEAR_NODE(&rq->rb_node);
 409 }
 410 EXPORT_SYMBOL(elv_rb_del);
 411
 412 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 413 {
 414         struct rb_node *n = root->rb_node;
 415         struct request *rq;
 416
 417         while (n) {
 418                 rq = rb_entry(n, struct request, rb_node);
 419
 420                 if (sector < rq->sector)
 421                         n = n->rb_left;
 422                 else if (sector > rq->sector)
 423                         n = n->rb_right;
 424                 else
 425                         return rq;
 426         }
 427
 428         return NULL;
 429 }
 430 EXPORT_SYMBOL(elv_rb_find);
 431
 432 /*
 433  * Insert rq into dispatch queue of q.  Queue lock must be held on
 434  * entry.  rq is sort instead into the dispatch queue. To be used by
 435  * specific elevators.
 436  */
 437 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 438 {
 439         sector_t boundary;
 440         struct list_head *entry;
 441         int stop_flags;
 442
 443         if (q->last_merge == rq)
 444                 q->last_merge = NULL;
 445
 446         elv_rqhash_del(q, rq);
 447
 448         q->nr_sorted--;
 449
 450         boundary = q->end_sector;
 451         stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
 452         list_for_each_prev(entry, &q->queue_head) {
 453                 struct request *pos = list_entry_rq(entry);
 454
 455                 if (blk_discard_rq(rq) != blk_discard_rq(pos))
 456                         break;
 457                 if (rq_data_dir(rq) != rq_data_dir(pos))
 458                         break;
 459                 if (pos->cmd_flags & stop_flags)
 460                         break;
 461                 if (rq->sector >= boundary) {
 462                         if (pos->sector < boundary)
 463                                 continue;
 464                 } else {
 465                         if (pos->sector >= boundary)
 466                                 break;
 467                 }
 468                 if (rq->sector >= pos->sector)
 469                         break;
 470         }
 471
 472         list_add(&rq->queuelist, entry);
 473 }
 474 EXPORT_SYMBOL(elv_dispatch_sort);
 475
 476 /*
 477  * Insert rq into dispatch queue of q.  Queue lock must be held on
 478  * entry.  rq is added to the back of the dispatch queue. To be used by
 479  * specific elevators.
 480  */
 481 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 482 {
 483         if (q->last_merge == rq)
 484                 q->last_merge = NULL;
 485
 486         elv_rqhash_del(q, rq);
 487
 488         q->nr_sorted--;
 489
 490         q->end_sector = rq_end_sector(rq);
 491         q->boundary_rq = rq;
 492         list_add_tail(&rq->queuelist, &q->queue_head);
 493 }
 494 EXPORT_SYMBOL(elv_dispatch_add_tail);
 495
 496 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 497 {
 498         struct elevator_queue *e = q->elevator;
 499         struct request *__rq;
 500         int ret;
 501
 502         /*
 503          * First try one-hit cache.
 504          */
 505         if (q->last_merge) {
 506                 ret = elv_try_merge(q->last_merge, bio);
 507                 if (ret != ELEVATOR_NO_MERGE) {
 508                         *req = q->last_merge;
 509                         return ret;
 510                 }
 511         }
 512
 513         if (blk_queue_nomerges(q))
 514                 return ELEVATOR_NO_MERGE;
 515
 516         /*
 517          * See if our hash lookup can find a potential backmerge.
 518          */
 519         __rq = elv_rqhash_find(q, bio->bi_sector);
 520         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 521                 *req = __rq;
 522                 return ELEVATOR_BACK_MERGE;
 523         }
 524
 525         if (e->ops->elevator_merge_fn)
 526                 return e->ops->elevator_merge_fn(q, req, bio);
 527
 528         return ELEVATOR_NO_MERGE;
 529 }
 530
 531 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 532 {
 533         struct elevator_queue *e = q->elevator;
 534
 535         if (e->ops->elevator_merged_fn)
 536                 e->ops->elevator_merged_fn(q, rq, type);
 537
 538         if (type == ELEVATOR_BACK_MERGE)
 539                 elv_rqhash_reposition(q, rq);
 540
 541         q->last_merge = rq;
 542 }
 543
 544 void elv_merge_requests(struct request_queue *q, struct request *rq,
 545                              struct request *next)
 546 {
 547         struct elevator_queue *e = q->elevator;
 548
 549         if (e->ops->elevator_merge_req_fn)
 550                 e->ops->elevator_merge_req_fn(q, rq, next);
 551
 552         elv_rqhash_reposition(q, rq);
 553         elv_rqhash_del(q, next);
 554
 555         q->nr_sorted--;
 556         q->last_merge = rq;
 557 }
 558
 559 void elv_requeue_request(struct request_queue *q, struct request *rq)
 560 {
 561         /*
 562          * it already went through dequeue, we need to decrement the
 563          * in_flight count again
 564          */
 565         if (blk_account_rq(rq)) {
 566                 q->in_flight--;
 567                 if (blk_sorted_rq(rq))
 568                         elv_deactivate_rq(q, rq);
 569         }
 570
 571         rq->cmd_flags &= ~REQ_STARTED;
 572
 573         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 574 }
 575
 576 void elv_drain_elevator(struct request_queue *q)
 577 {
 578         static int printed;
 579         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 580                 ;
 581         if (q->nr_sorted == 0)
 582                 return;
 583         if (printed++ < 10) {
 584                 printk(KERN_ERR "%s: forced dispatching is broken "
 585                        "(nr_sorted=%u), please report this\n",
 586                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 587         }
 588 }
 589
 590 /*
 591  * Call with queue lock held, interrupts disabled
 592  */
 593 void elv_quisce_start(struct request_queue *q)
 594 {
 595         queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
 596
 597         /*
 598          * make sure we don't have any requests in flight
 599          */
 600         elv_drain_elevator(q);
 601         while (q->rq.elvpriv) {
 602                 blk_start_queueing(q);
 603                 spin_unlock_irq(q->queue_lock);
 604                 msleep(10);
 605                 spin_lock_irq(q->queue_lock);
 606                 elv_drain_elevator(q);
 607         }
 608 }
 609
 610 void elv_quisce_end(struct request_queue *q)
 611 {
 612         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
 613 }
 614
 615 void elv_insert(struct request_queue *q, struct request *rq, int where)
 616 {
 617         struct list_head *pos;
 618         unsigned ordseq;
 619         int unplug_it = 1;
 620
 621         trace_block_rq_insert(q, rq);
 622
 623         rq->q = q;
 624
 625         switch (where) {
 626         case ELEVATOR_INSERT_FRONT:
 627                 rq->cmd_flags |= REQ_SOFTBARRIER;
 628
 629                 list_add(&rq->queuelist, &q->queue_head);
 630                 break;
 631
 632         case ELEVATOR_INSERT_BACK:
 633                 rq->cmd_flags |= REQ_SOFTBARRIER;
 634                 elv_drain_elevator(q);
 635                 list_add_tail(&rq->queuelist, &q->queue_head);
 636                 /*
 637                  * We kick the queue here for the following reasons.
 638                  * - The elevator might have returned NULL previously
 639                  *   to delay requests and returned them now.  As the
 640                  *   queue wasn't empty before this request, ll_rw_blk
 641                  *   won't run the queue on return, resulting in hang.
 642                  * - Usually, back inserted requests won't be merged
 643                  *   with anything.  There's no point in delaying queue
 644                  *   processing.
 645                  */
 646                 blk_remove_plug(q);
 647                 blk_start_queueing(q);
 648                 break;
 649
 650         case ELEVATOR_INSERT_SORT:
 651                 BUG_ON(!blk_fs_request(rq) && !blk_discard_rq(rq));
 652                 rq->cmd_flags |= REQ_SORTED;
 653                 q->nr_sorted++;
 654                 if (rq_mergeable(rq)) {
 655                         elv_rqhash_add(q, rq);
 656                         if (!q->last_merge)
 657                                 q->last_merge = rq;
 658                 }
 659
 660                 /*
 661                  * Some ioscheds (cfq) run q->request_fn directly, so
 662                  * rq cannot be accessed after calling
 663                  * elevator_add_req_fn.
 664                  */
 665                 q->elevator->ops->elevator_add_req_fn(q, rq);
 666                 break;
 667
 668         case ELEVATOR_INSERT_REQUEUE:
 669                 /*
 670                  * If ordered flush isn't in progress, we do front
 671                  * insertion; otherwise, requests should be requeued
 672                  * in ordseq order.
 673                  */
 674                 rq->cmd_flags |= REQ_SOFTBARRIER;
 675
 676                 /*
 677                  * Most requeues happen because of a busy condition,
 678                  * don't force unplug of the queue for that case.
 679                  */
 680                 unplug_it = 0;
 681
 682                 if (q->ordseq == 0) {
 683                         list_add(&rq->queuelist, &q->queue_head);
 684                         break;
 685                 }
 686
 687                 ordseq = blk_ordered_req_seq(rq);
 688
 689                 list_for_each(pos, &q->queue_head) {
 690                         struct request *pos_rq = list_entry_rq(pos);
 691                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 692                                 break;
 693                 }
 694
 695                 list_add_tail(&rq->queuelist, pos);
 696                 break;
 697
 698         default:
 699                 printk(KERN_ERR "%s: bad insertion point %d\n",
 700                        __func__, where);
 701                 BUG();
 702         }
 703
 704         if (unplug_it && blk_queue_plugged(q)) {
 705                 int nrq = q->rq.count[BLK_RW_SYNC] + q->rq.count[BLK_RW_ASYNC]
 706                         - q->in_flight;
 707
 708                 if (nrq >= q->unplug_thresh)
 709                         __generic_unplug_device(q);
 710         }
 711 }
 712
 713 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
 714                        int plug)
 715 {
 716         if (q->ordcolor)
 717                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 718
 719         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 720                 /*
 721                  * toggle ordered color
 722                  */
 723                 if (blk_barrier_rq(rq))
 724                         q->ordcolor ^= 1;
 725
 726                 /*
 727                  * barriers implicitly indicate back insertion
 728                  */
 729                 if (where == ELEVATOR_INSERT_SORT)
 730                         where = ELEVATOR_INSERT_BACK;
 731
 732                 /*
 733                  * this request is scheduling boundary, update
 734                  * end_sector
 735                  */
 736                 if (blk_fs_request(rq) || blk_discard_rq(rq)) {
 737                         q->end_sector = rq_end_sector(rq);
 738                         q->boundary_rq = rq;
 739                 }
 740         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 741                     where == ELEVATOR_INSERT_SORT)
 742                 where = ELEVATOR_INSERT_BACK;
 743
 744         if (plug)
 745                 blk_plug_device(q);
 746
 747         elv_insert(q, rq, where);
 748 }
 749 EXPORT_SYMBOL(__elv_add_request);
 750
 751 void elv_add_request(struct request_queue *q, struct request *rq, int where,
 752                      int plug)
 753 {
 754         unsigned long flags;
 755
 756         spin_lock_irqsave(q->queue_lock, flags);
 757         __elv_add_request(q, rq, where, plug);
 758         spin_unlock_irqrestore(q->queue_lock, flags);
 759 }
 760 EXPORT_SYMBOL(elv_add_request);
 761
 762 static inline struct request *__elv_next_request(struct request_queue *q)
 763 {
 764         struct request *rq;
 765
 766         while (1) {
 767                 while (!list_empty(&q->queue_head)) {
 768                         rq = list_entry_rq(q->queue_head.next);
 769                         if (blk_do_ordered(q, &rq))
 770                                 return rq;
 771                 }
 772
 773                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
 774                         return NULL;
 775         }
 776 }
 777
 778 struct request *elv_next_request(struct request_queue *q)
 779 {
 780         struct request *rq;
 781         int ret;
 782
 783         while ((rq = __elv_next_request(q)) != NULL) {
 784                 if (!(rq->cmd_flags & REQ_STARTED)) {
 785                         /*
 786                          * This is the first time the device driver
 787                          * sees this request (possibly after
 788                          * requeueing).  Notify IO scheduler.
 789                          */
 790                         if (blk_sorted_rq(rq))
 791                                 elv_activate_rq(q, rq);
 792
 793                         /*
 794                          * just mark as started even if we don't start
 795                          * it, a request that has been delayed should
 796                          * not be passed by new incoming requests
 797                          */
 798                         rq->cmd_flags |= REQ_STARTED;
 799                         trace_block_rq_issue(q, rq);
 800                 }
 801
 802                 if (!q->boundary_rq || q->boundary_rq == rq) {
 803                         q->end_sector = rq_end_sector(rq);
 804                         q->boundary_rq = NULL;
 805                 }
 806
 807                 if (rq->cmd_flags & REQ_DONTPREP)
 808                         break;
 809
 810                 if (q->dma_drain_size && rq->data_len) {
 811                         /*
 812                          * make sure space for the drain appears we
 813                          * know we can do this because max_hw_segments
 814                          * has been adjusted to be one fewer than the
 815                          * device can handle
 816                          */
 817                         rq->nr_phys_segments++;
 818                 }
 819
 820                 if (!q->prep_rq_fn)
 821                         break;
 822
 823                 ret = q->prep_rq_fn(q, rq);
 824                 if (ret == BLKPREP_OK) {
 825                         break;
 826                 } else if (ret == BLKPREP_DEFER) {
 827                         /*
 828                          * the request may have been (partially) prepped.
 829                          * we need to keep this request in the front to
 830                          * avoid resource deadlock.  REQ_STARTED will
 831                          * prevent other fs requests from passing this one.
 832                          */
 833                         if (q->dma_drain_size && rq->data_len &&
 834                             !(rq->cmd_flags & REQ_DONTPREP)) {
 835                                 /*
 836                                  * remove the space for the drain we added
 837                                  * so that we don't add it again
 838                                  */
 839                                 --rq->nr_phys_segments;
 840                         }
 841
 842                         rq = NULL;
 843                         break;
 844                 } else if (ret == BLKPREP_KILL) {
 845                         rq->cmd_flags |= REQ_QUIET;
 846                         __blk_end_request(rq, -EIO, blk_rq_bytes(rq));
 847                 } else {
 848                         printk(KERN_ERR "%s: bad return=%d\n", __func__, ret);
 849                         break;
 850                 }
 851         }
 852
 853         return rq;
 854 }
 855 EXPORT_SYMBOL(elv_next_request);
 856
 857 void elv_dequeue_request(struct request_queue *q, struct request *rq)
 858 {
 859         BUG_ON(list_empty(&rq->queuelist));
 860         BUG_ON(ELV_ON_HASH(rq));
 861
 862         list_del_init(&rq->queuelist);
 863
 864         /*
 865          * the time frame between a request being removed from the lists
 866          * and to it is freed is accounted as io that is in progress at
 867          * the driver side.
 868          */
 869         if (blk_account_rq(rq))
 870                 q->in_flight++;
 871 }
 872
 873 int elv_queue_empty(struct request_queue *q)
 874 {
 875         struct elevator_queue *e = q->elevator;
 876
 877         if (!list_empty(&q->queue_head))
 878                 return 0;
 879
 880         if (e->ops->elevator_queue_empty_fn)
 881                 return e->ops->elevator_queue_empty_fn(q);
 882
 883         return 1;
 884 }
 885 EXPORT_SYMBOL(elv_queue_empty);
 886
 887 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 888 {
 889         struct elevator_queue *e = q->elevator;
 890
 891         if (e->ops->elevator_latter_req_fn)
 892                 return e->ops->elevator_latter_req_fn(q, rq);
 893         return NULL;
 894 }
 895
 896 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 897 {
 898         struct elevator_queue *e = q->elevator;
 899
 900         if (e->ops->elevator_former_req_fn)
 901                 return e->ops->elevator_former_req_fn(q, rq);
 902         return NULL;
 903 }
 904
 905 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 906 {
 907         struct elevator_queue *e = q->elevator;
 908
 909         if (e->ops->elevator_set_req_fn)
 910                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 911
 912         rq->elevator_private = NULL;
 913         return 0;
 914 }
 915
 916 void elv_put_request(struct request_queue *q, struct request *rq)
 917 {
 918         struct elevator_queue *e = q->elevator;
 919
 920         if (e->ops->elevator_put_req_fn)
 921                 e->ops->elevator_put_req_fn(rq);
 922 }
 923
 924 int elv_may_queue(struct request_queue *q, int rw)
 925 {
 926         struct elevator_queue *e = q->elevator;
 927
 928         if (e->ops->elevator_may_queue_fn)
 929                 return e->ops->elevator_may_queue_fn(q, rw);
 930
 931         return ELV_MQUEUE_MAY;
 932 }
 933
 934 void elv_abort_queue(struct request_queue *q)
 935 {
 936         struct request *rq;
 937
 938         while (!list_empty(&q->queue_head)) {
 939                 rq = list_entry_rq(q->queue_head.next);
 940                 rq->cmd_flags |= REQ_QUIET;
 941                 trace_block_rq_abort(q, rq);
 942                 __blk_end_request(rq, -EIO, blk_rq_bytes(rq));
 943         }
 944 }
 945 EXPORT_SYMBOL(elv_abort_queue);
 946
 947 void elv_completed_request(struct request_queue *q, struct request *rq)
 948 {
 949         struct elevator_queue *e = q->elevator;
 950
 951         /*
 952          * request is released from the driver, io must be done
 953          */
 954         if (blk_account_rq(rq)) {
 955                 q->in_flight--;
 956                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 957                         e->ops->elevator_completed_req_fn(q, rq);
 958         }
 959
 960         /*
 961          * Check if the queue is waiting for fs requests to be
 962          * drained for flush sequence.
 963          */
 964         if (unlikely(q->ordseq)) {
 965                 struct request *next = NULL;
 966
 967                 if (!list_empty(&q->queue_head))
 968                         next = list_entry_rq(q->queue_head.next);
 969
 970                 if (!q->in_flight &&
 971                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 972                     (!next || blk_ordered_req_seq(next) > QUEUE_ORDSEQ_DRAIN)) {
 973                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 974                         blk_start_queueing(q);
 975                 }
 976         }
 977 }
 978
 979 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 980
 981 static ssize_t
 982 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 983 {
 984         struct elv_fs_entry *entry = to_elv(attr);
 985         struct elevator_queue *e;
 986         ssize_t error;
 987
 988         if (!entry->show)
 989                 return -EIO;
 990
 991         e = container_of(kobj, struct elevator_queue, kobj);
 992         mutex_lock(&e->sysfs_lock);
 993         error = e->ops ? entry->show(e, page) : -ENOENT;
 994         mutex_unlock(&e->sysfs_lock);
 995         return error;
 996 }
 997
 998 static ssize_t
 999 elv_attr_store(struct kobject *kobj, struct attribute *attr,
1000                const char *page, size_t length)
1001 {
1002         struct elv_fs_entry *entry = to_elv(attr);
1003         struct elevator_queue *e;
1004         ssize_t error;
1005
1006         if (!entry->store)
1007                 return -EIO;
1008
1009         e = container_of(kobj, struct elevator_queue, kobj);
1010         mutex_lock(&e->sysfs_lock);
1011         error = e->ops ? entry->store(e, page, length) : -ENOENT;
1012         mutex_unlock(&e->sysfs_lock);
1013         return error;
1014 }
1015
1016 static struct sysfs_ops elv_sysfs_ops = {
1017         .show   = elv_attr_show,
1018         .store  = elv_attr_store,
1019 };
1020
1021 static struct kobj_type elv_ktype = {
1022         .sysfs_ops      = &elv_sysfs_ops,
1023         .release        = elevator_release,
1024 };
1025
1026 int elv_register_queue(struct request_queue *q)
1027 {
1028         struct elevator_queue *e = q->elevator;
1029         int error;
1030
1031         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
1032         if (!error) {
1033                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
1034                 if (attr) {
1035                         while (attr->attr.name) {
1036                                 if (sysfs_create_file(&e->kobj, &attr->attr))
1037                                         break;
1038                                 attr++;
1039                         }
1040                 }
1041                 kobject_uevent(&e->kobj, KOBJ_ADD);
1042         }
1043         return error;
1044 }
1045
1046 static void __elv_unregister_queue(struct elevator_queue *e)
1047 {
1048         kobject_uevent(&e->kobj, KOBJ_REMOVE);
1049         kobject_del(&e->kobj);
1050 }
1051
1052 void elv_unregister_queue(struct request_queue *q)
1053 {
1054         if (q)
1055                 __elv_unregister_queue(q->elevator);
1056 }
1057
1058 void elv_register(struct elevator_type *e)
1059 {
1060         char *def = "";
1061
1062         spin_lock(&elv_list_lock);
1063         BUG_ON(elevator_find(e->elevator_name));
1064         list_add_tail(&e->list, &elv_list);
1065         spin_unlock(&elv_list_lock);
1066
1067         if (!strcmp(e->elevator_name, chosen_elevator) ||
1068                         (!*chosen_elevator &&
1069                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
1070                                 def = " (default)";
1071
1072         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
1073                                                                 def);
1074 }
1075 EXPORT_SYMBOL_GPL(elv_register);
1076
1077 void elv_unregister(struct elevator_type *e)
1078 {
1079         struct task_struct *g, *p;
1080
1081         /*
1082          * Iterate every thread in the process to remove the io contexts.
1083          */
1084         if (e->ops.trim) {
1085                 read_lock(&tasklist_lock);
1086                 do_each_thread(g, p) {
1087                         task_lock(p);
1088                         if (p->io_context)
1089                                 e->ops.trim(p->io_context);
1090                         task_unlock(p);
1091                 } while_each_thread(g, p);
1092                 read_unlock(&tasklist_lock);
1093         }
1094
1095         spin_lock(&elv_list_lock);
1096         list_del_init(&e->list);
1097         spin_unlock(&elv_list_lock);
1098 }
1099 EXPORT_SYMBOL_GPL(elv_unregister);
1100
1101 /*
1102  * switch to new_e io scheduler. be careful not to introduce deadlocks -
1103  * we don't free the old io scheduler, before we have allocated what we
1104  * need for the new one. this way we have a chance of going back to the old
1105  * one, if the new one fails init for some reason.
1106  */
1107 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1108 {
1109         struct elevator_queue *old_elevator, *e;
1110         void *data;
1111
1112         /*
1113          * Allocate new elevator
1114          */
1115         e = elevator_alloc(q, new_e);
1116         if (!e)
1117                 return 0;
1118
1119         data = elevator_init_queue(q, e);
1120         if (!data) {
1121                 kobject_put(&e->kobj);
1122                 return 0;
1123         }
1124
1125         /*
1126          * Turn on BYPASS and drain all requests w/ elevator private data
1127          */
1128         spin_lock_irq(q->queue_lock);
1129         elv_quisce_start(q);
1130
1131         /*
1132          * Remember old elevator.
1133          */
1134         old_elevator = q->elevator;
1135
1136         /*
1137          * attach and start new elevator
1138          */
1139         elevator_attach(q, e, data);
1140
1141         spin_unlock_irq(q->queue_lock);
1142
1143         __elv_unregister_queue(old_elevator);
1144
1145         if (elv_register_queue(q))
1146                 goto fail_register;
1147
1148         /*
1149          * finally exit old elevator and turn off BYPASS.
1150          */
1151         elevator_exit(old_elevator);
1152         spin_lock_irq(q->queue_lock);
1153         elv_quisce_end(q);
1154         spin_unlock_irq(q->queue_lock);
1155
1156         blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
1157
1158         return 1;
1159
1160 fail_register:
1161         /*
1162          * switch failed, exit the new io scheduler and reattach the old
1163          * one again (along with re-adding the sysfs dir)
1164          */
1165         elevator_exit(e);
1166         q->elevator = old_elevator;
1167         elv_register_queue(q);
1168
1169         spin_lock_irq(q->queue_lock);
1170         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1171         spin_unlock_irq(q->queue_lock);
1172
1173         return 0;
1174 }
1175
1176 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1177                           size_t count)
1178 {
1179         char elevator_name[ELV_NAME_MAX];
1180         struct elevator_type *e;
1181
1182         strlcpy(elevator_name, name, sizeof(elevator_name));
1183         strstrip(elevator_name);
1184
1185         e = elevator_get(elevator_name);
1186         if (!e) {
1187                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1188                 return -EINVAL;
1189         }
1190
1191         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1192                 elevator_put(e);
1193                 return count;
1194         }
1195
1196         if (!elevator_switch(q, e))
1197                 printk(KERN_ERR "elevator: switch to %s failed\n",
1198                                                         elevator_name);
1199         return count;
1200 }
1201
1202 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1203 {
1204         struct elevator_queue *e = q->elevator;
1205         struct elevator_type *elv = e->elevator_type;
1206         struct elevator_type *__e;
1207         int len = 0;
1208
1209         spin_lock(&elv_list_lock);
1210         list_for_each_entry(__e, &elv_list, list) {
1211                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1212                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1213                 else
1214                         len += sprintf(name+len, "%s ", __e->elevator_name);
1215         }
1216         spin_unlock(&elv_list_lock);
1217
1218         len += sprintf(len+name, "\n");
1219         return len;
1220 }
1221
1222 struct request *elv_rb_former_request(struct request_queue *q,
1223                                       struct request *rq)
1224 {
1225         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1226
1227         if (rbprev)
1228                 return rb_entry_rq(rbprev);
1229
1230         return NULL;
1231 }
1232 EXPORT_SYMBOL(elv_rb_former_request);
1233
1234 struct request *elv_rb_latter_request(struct request_queue *q,
1235                                       struct request *rq)
1236 {
1237         struct rb_node *rbnext = rb_next(&rq->rb_node);
1238
1239         if (rbnext)
1240                 return rb_entry_rq(rbnext);
1241
1242         return NULL;
1243 }
1244 EXPORT_SYMBOL(elv_rb_latter_request);