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)         (blk_rq_pos(rq) + blk_rq_sectors(rq))
  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         struct elevator_queue *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 (blk_rq_pos(__rq) + blk_rq_sectors(__rq) == bio->bi_sector)
 119                         ret = ELEVATOR_BACK_MERGE;
 120                 else if (blk_rq_pos(__rq) - 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 struct elevator_queue *elevator_alloc(struct request_queue *q,
 207                                   struct elevator_type *e)
 208 {
 209         struct elevator_queue *eq;
 210         int i;
 211
 212         eq = kmalloc_node(sizeof(*eq), 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         struct elevator_queue *e;
 239
 240         e = container_of(kobj, struct elevator_queue, kobj);
 241         elevator_put(e->elevator_type);
 242         kfree(e->hash);
 243         kfree(e);
 244 }
 245
 246 int elevator_init(struct request_queue *q, char *name)
 247 {
 248         struct elevator_type *e = NULL;
 249         struct elevator_queue *eq;
 250         int ret = 0;
 251         void *data;
 252
 253         INIT_LIST_HEAD(&q->queue_head);
 254         q->last_merge = NULL;
 255         q->end_sector = 0;
 256         q->boundary_rq = NULL;
 257
 258         if (name) {
 259                 e = elevator_get(name);
 260                 if (!e)
 261                         return -EINVAL;
 262         }
 263
 264         if (!e && *chosen_elevator) {
 265                 e = elevator_get(chosen_elevator);
 266                 if (!e)
 267                         printk(KERN_ERR "I/O scheduler %s not found\n",
 268                                                         chosen_elevator);
 269         }
 270
 271         if (!e) {
 272                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
 273                 if (!e) {
 274                         printk(KERN_ERR
 275                                 "Default I/O scheduler not found. " \
 276                                 "Using noop.\n");
 277                         e = elevator_get("noop");
 278                 }
 279         }
 280
 281         eq = elevator_alloc(q, e);
 282         if (!eq)
 283                 return -ENOMEM;
 284
 285         data = elevator_init_queue(q, eq);
 286         if (!data) {
 287                 kobject_put(&eq->kobj);
 288                 return -ENOMEM;
 289         }
 290
 291         elevator_attach(q, eq, data);
 292         return ret;
 293 }
 294 EXPORT_SYMBOL(elevator_init);
 295
 296 void elevator_exit(struct elevator_queue *e)
 297 {
 298         mutex_lock(&e->sysfs_lock);
 299         if (e->ops->elevator_exit_fn)
 300                 e->ops->elevator_exit_fn(e);
 301         e->ops = NULL;
 302         mutex_unlock(&e->sysfs_lock);
 303
 304         kobject_put(&e->kobj);
 305 }
 306 EXPORT_SYMBOL(elevator_exit);
 307
 308 static inline void __elv_rqhash_del(struct request *rq)
 309 {
 310         hlist_del_init(&rq->hash);
 311 }
 312
 313 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 314 {
 315         if (ELV_ON_HASH(rq))
 316                 __elv_rqhash_del(rq);
 317 }
 318
 319 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 320 {
 321         struct elevator_queue *e = q->elevator;
 322
 323         BUG_ON(ELV_ON_HASH(rq));
 324         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 325 }
 326
 327 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 328 {
 329         __elv_rqhash_del(rq);
 330         elv_rqhash_add(q, rq);
 331 }
 332
 333 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 334 {
 335         struct elevator_queue *e = q->elevator;
 336         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 337         struct hlist_node *entry, *next;
 338         struct request *rq;
 339
 340         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 341                 BUG_ON(!ELV_ON_HASH(rq));
 342
 343                 if (unlikely(!rq_mergeable(rq))) {
 344                         __elv_rqhash_del(rq);
 345                         continue;
 346                 }
 347
 348                 if (rq_hash_key(rq) == offset)
 349                         return rq;
 350         }
 351
 352         return NULL;
 353 }
 354
 355 /*
 356  * RB-tree support functions for inserting/lookup/removal of requests
 357  * in a sorted RB tree.
 358  */
 359 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 360 {
 361         struct rb_node **p = &root->rb_node;
 362         struct rb_node *parent = NULL;
 363         struct request *__rq;
 364
 365         while (*p) {
 366                 parent = *p;
 367                 __rq = rb_entry(parent, struct request, rb_node);
 368
 369                 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
 370                         p = &(*p)->rb_left;
 371                 else if (blk_rq_pos(rq) > blk_rq_pos(__rq))
 372                         p = &(*p)->rb_right;
 373                 else
 374                         return __rq;
 375         }
 376
 377         rb_link_node(&rq->rb_node, parent, p);
 378         rb_insert_color(&rq->rb_node, root);
 379         return NULL;
 380 }
 381 EXPORT_SYMBOL(elv_rb_add);
 382
 383 void elv_rb_del(struct rb_root *root, struct request *rq)
 384 {
 385         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 386         rb_erase(&rq->rb_node, root);
 387         RB_CLEAR_NODE(&rq->rb_node);
 388 }
 389 EXPORT_SYMBOL(elv_rb_del);
 390
 391 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 392 {
 393         struct rb_node *n = root->rb_node;
 394         struct request *rq;
 395
 396         while (n) {
 397                 rq = rb_entry(n, struct request, rb_node);
 398
 399                 if (sector < blk_rq_pos(rq))
 400                         n = n->rb_left;
 401                 else if (sector > blk_rq_pos(rq))
 402                         n = n->rb_right;
 403                 else
 404                         return rq;
 405         }
 406
 407         return NULL;
 408 }
 409 EXPORT_SYMBOL(elv_rb_find);
 410
 411 /*
 412  * Insert rq into dispatch queue of q.  Queue lock must be held on
 413  * entry.  rq is sort instead into the dispatch queue. To be used by
 414  * specific elevators.
 415  */
 416 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 417 {
 418         sector_t boundary;
 419         struct list_head *entry;
 420         int stop_flags;
 421
 422         if (q->last_merge == rq)
 423                 q->last_merge = NULL;
 424
 425         elv_rqhash_del(q, rq);
 426
 427         q->nr_sorted--;
 428
 429         boundary = q->end_sector;
 430         stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
 431         list_for_each_prev(entry, &q->queue_head) {
 432                 struct request *pos = list_entry_rq(entry);
 433
 434                 if (blk_discard_rq(rq) != blk_discard_rq(pos))
 435                         break;
 436                 if (rq_data_dir(rq) != rq_data_dir(pos))
 437                         break;
 438                 if (pos->cmd_flags & stop_flags)
 439                         break;
 440                 if (blk_rq_pos(rq) >= boundary) {
 441                         if (blk_rq_pos(pos) < boundary)
 442                                 continue;
 443                 } else {
 444                         if (blk_rq_pos(pos) >= boundary)
 445                                 break;
 446                 }
 447                 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
 448                         break;
 449         }
 450
 451         list_add(&rq->queuelist, entry);
 452 }
 453 EXPORT_SYMBOL(elv_dispatch_sort);
 454
 455 /*
 456  * Insert rq into dispatch queue of q.  Queue lock must be held on
 457  * entry.  rq is added to the back of the dispatch queue. To be used by
 458  * specific elevators.
 459  */
 460 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 461 {
 462         if (q->last_merge == rq)
 463                 q->last_merge = NULL;
 464
 465         elv_rqhash_del(q, rq);
 466
 467         q->nr_sorted--;
 468
 469         q->end_sector = rq_end_sector(rq);
 470         q->boundary_rq = rq;
 471         list_add_tail(&rq->queuelist, &q->queue_head);
 472 }
 473 EXPORT_SYMBOL(elv_dispatch_add_tail);
 474
 475 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 476 {
 477         struct elevator_queue *e = q->elevator;
 478         struct request *__rq;
 479         int ret;
 480
 481         /*
 482          * First try one-hit cache.
 483          */
 484         if (q->last_merge) {
 485                 ret = elv_try_merge(q->last_merge, bio);
 486                 if (ret != ELEVATOR_NO_MERGE) {
 487                         *req = q->last_merge;
 488                         return ret;
 489                 }
 490         }
 491
 492         if (blk_queue_nomerges(q))
 493                 return ELEVATOR_NO_MERGE;
 494
 495         /*
 496          * See if our hash lookup can find a potential backmerge.
 497          */
 498         __rq = elv_rqhash_find(q, bio->bi_sector);
 499         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 500                 *req = __rq;
 501                 return ELEVATOR_BACK_MERGE;
 502         }
 503
 504         if (e->ops->elevator_merge_fn)
 505                 return e->ops->elevator_merge_fn(q, req, bio);
 506
 507         return ELEVATOR_NO_MERGE;
 508 }
 509
 510 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 511 {
 512         struct elevator_queue *e = q->elevator;
 513
 514         if (e->ops->elevator_merged_fn)
 515                 e->ops->elevator_merged_fn(q, rq, type);
 516
 517         if (type == ELEVATOR_BACK_MERGE)
 518                 elv_rqhash_reposition(q, rq);
 519
 520         q->last_merge = rq;
 521 }
 522
 523 void elv_merge_requests(struct request_queue *q, struct request *rq,
 524                              struct request *next)
 525 {
 526         struct elevator_queue *e = q->elevator;
 527
 528         if (e->ops->elevator_merge_req_fn)
 529                 e->ops->elevator_merge_req_fn(q, rq, next);
 530
 531         elv_rqhash_reposition(q, rq);
 532         elv_rqhash_del(q, next);
 533
 534         q->nr_sorted--;
 535         q->last_merge = rq;
 536 }
 537
 538 void elv_requeue_request(struct request_queue *q, struct request *rq)
 539 {
 540         /*
 541          * it already went through dequeue, we need to decrement the
 542          * in_flight count again
 543          */
 544         if (blk_account_rq(rq)) {
 545                 q->in_flight[rq_is_sync(rq)]--;
 546                 if (blk_sorted_rq(rq))
 547                         elv_deactivate_rq(q, rq);
 548         }
 549
 550         rq->cmd_flags &= ~REQ_STARTED;
 551
 552         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 553 }
 554
 555 void elv_drain_elevator(struct request_queue *q)
 556 {
 557         static int printed;
 558         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 559                 ;
 560         if (q->nr_sorted == 0)
 561                 return;
 562         if (printed++ < 10) {
 563                 printk(KERN_ERR "%s: forced dispatching is broken "
 564                        "(nr_sorted=%u), please report this\n",
 565                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 566         }
 567 }
 568
 569 /*
 570  * Call with queue lock held, interrupts disabled
 571  */
 572 void elv_quiesce_start(struct request_queue *q)
 573 {
 574         if (!q->elevator)
 575                 return;
 576
 577         queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
 578
 579         /*
 580          * make sure we don't have any requests in flight
 581          */
 582         elv_drain_elevator(q);
 583         while (q->rq.elvpriv) {
 584                 __blk_run_queue(q);
 585                 spin_unlock_irq(q->queue_lock);
 586                 msleep(10);
 587                 spin_lock_irq(q->queue_lock);
 588                 elv_drain_elevator(q);
 589         }
 590 }
 591
 592 void elv_quiesce_end(struct request_queue *q)
 593 {
 594         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
 595 }
 596
 597 void elv_insert(struct request_queue *q, struct request *rq, int where)
 598 {
 599         struct list_head *pos;
 600         unsigned ordseq;
 601         int unplug_it = 1;
 602
 603         trace_block_rq_insert(q, rq);
 604
 605         rq->q = q;
 606
 607         switch (where) {
 608         case ELEVATOR_INSERT_FRONT:
 609                 rq->cmd_flags |= REQ_SOFTBARRIER;
 610
 611                 list_add(&rq->queuelist, &q->queue_head);
 612                 break;
 613
 614         case ELEVATOR_INSERT_BACK:
 615                 rq->cmd_flags |= REQ_SOFTBARRIER;
 616                 elv_drain_elevator(q);
 617                 list_add_tail(&rq->queuelist, &q->queue_head);
 618                 /*
 619                  * We kick the queue here for the following reasons.
 620                  * - The elevator might have returned NULL previously
 621                  *   to delay requests and returned them now.  As the
 622                  *   queue wasn't empty before this request, ll_rw_blk
 623                  *   won't run the queue on return, resulting in hang.
 624                  * - Usually, back inserted requests won't be merged
 625                  *   with anything.  There's no point in delaying queue
 626                  *   processing.
 627                  */
 628                 __blk_run_queue(q);
 629                 break;
 630
 631         case ELEVATOR_INSERT_SORT:
 632                 BUG_ON(!blk_fs_request(rq) && !blk_discard_rq(rq));
 633                 rq->cmd_flags |= REQ_SORTED;
 634                 q->nr_sorted++;
 635                 if (rq_mergeable(rq)) {
 636                         elv_rqhash_add(q, rq);
 637                         if (!q->last_merge)
 638                                 q->last_merge = rq;
 639                 }
 640
 641                 /*
 642                  * Some ioscheds (cfq) run q->request_fn directly, so
 643                  * rq cannot be accessed after calling
 644                  * elevator_add_req_fn.
 645                  */
 646                 q->elevator->ops->elevator_add_req_fn(q, rq);
 647                 break;
 648
 649         case ELEVATOR_INSERT_REQUEUE:
 650                 /*
 651                  * If ordered flush isn't in progress, we do front
 652                  * insertion; otherwise, requests should be requeued
 653                  * in ordseq order.
 654                  */
 655                 rq->cmd_flags |= REQ_SOFTBARRIER;
 656
 657                 /*
 658                  * Most requeues happen because of a busy condition,
 659                  * don't force unplug of the queue for that case.
 660                  */
 661                 unplug_it = 0;
 662
 663                 if (q->ordseq == 0) {
 664                         list_add(&rq->queuelist, &q->queue_head);
 665                         break;
 666                 }
 667
 668                 ordseq = blk_ordered_req_seq(rq);
 669
 670                 list_for_each(pos, &q->queue_head) {
 671                         struct request *pos_rq = list_entry_rq(pos);
 672                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 673                                 break;
 674                 }
 675
 676                 list_add_tail(&rq->queuelist, pos);
 677                 break;
 678
 679         default:
 680                 printk(KERN_ERR "%s: bad insertion point %d\n",
 681                        __func__, where);
 682                 BUG();
 683         }
 684
 685         if (unplug_it && blk_queue_plugged(q)) {
 686                 int nrq = q->rq.count[BLK_RW_SYNC] + q->rq.count[BLK_RW_ASYNC]
 687                                 - queue_in_flight(q);
 688
 689                 if (nrq >= q->unplug_thresh)
 690                         __generic_unplug_device(q);
 691         }
 692 }
 693
 694 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
 695                        int plug)
 696 {
 697         if (q->ordcolor)
 698                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 699
 700         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 701                 /*
 702                  * toggle ordered color
 703                  */
 704                 if (blk_barrier_rq(rq))
 705                         q->ordcolor ^= 1;
 706
 707                 /*
 708                  * barriers implicitly indicate back insertion
 709                  */
 710                 if (where == ELEVATOR_INSERT_SORT)
 711                         where = ELEVATOR_INSERT_BACK;
 712
 713                 /*
 714                  * this request is scheduling boundary, update
 715                  * end_sector
 716                  */
 717                 if (blk_fs_request(rq) || blk_discard_rq(rq)) {
 718                         q->end_sector = rq_end_sector(rq);
 719                         q->boundary_rq = rq;
 720                 }
 721         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 722                     where == ELEVATOR_INSERT_SORT)
 723                 where = ELEVATOR_INSERT_BACK;
 724
 725         if (plug)
 726                 blk_plug_device(q);
 727
 728         elv_insert(q, rq, where);
 729 }
 730 EXPORT_SYMBOL(__elv_add_request);
 731
 732 void elv_add_request(struct request_queue *q, struct request *rq, int where,
 733                      int plug)
 734 {
 735         unsigned long flags;
 736
 737         spin_lock_irqsave(q->queue_lock, flags);
 738         __elv_add_request(q, rq, where, plug);
 739         spin_unlock_irqrestore(q->queue_lock, flags);
 740 }
 741 EXPORT_SYMBOL(elv_add_request);
 742
 743 int elv_queue_empty(struct request_queue *q)
 744 {
 745         struct elevator_queue *e = q->elevator;
 746
 747         if (!list_empty(&q->queue_head))
 748                 return 0;
 749
 750         if (e->ops->elevator_queue_empty_fn)
 751                 return e->ops->elevator_queue_empty_fn(q);
 752
 753         return 1;
 754 }
 755 EXPORT_SYMBOL(elv_queue_empty);
 756
 757 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 758 {
 759         struct elevator_queue *e = q->elevator;
 760
 761         if (e->ops->elevator_latter_req_fn)
 762                 return e->ops->elevator_latter_req_fn(q, rq);
 763         return NULL;
 764 }
 765
 766 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 767 {
 768         struct elevator_queue *e = q->elevator;
 769
 770         if (e->ops->elevator_former_req_fn)
 771                 return e->ops->elevator_former_req_fn(q, rq);
 772         return NULL;
 773 }
 774
 775 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 776 {
 777         struct elevator_queue *e = q->elevator;
 778
 779         if (e->ops->elevator_set_req_fn)
 780                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 781
 782         rq->elevator_private = NULL;
 783         return 0;
 784 }
 785
 786 void elv_put_request(struct request_queue *q, struct request *rq)
 787 {
 788         struct elevator_queue *e = q->elevator;
 789
 790         if (e->ops->elevator_put_req_fn)
 791                 e->ops->elevator_put_req_fn(rq);
 792 }
 793
 794 int elv_may_queue(struct request_queue *q, int rw)
 795 {
 796         struct elevator_queue *e = q->elevator;
 797
 798         if (e->ops->elevator_may_queue_fn)
 799                 return e->ops->elevator_may_queue_fn(q, rw);
 800
 801         return ELV_MQUEUE_MAY;
 802 }
 803
 804 void elv_abort_queue(struct request_queue *q)
 805 {
 806         struct request *rq;
 807
 808         while (!list_empty(&q->queue_head)) {
 809                 rq = list_entry_rq(q->queue_head.next);
 810                 rq->cmd_flags |= REQ_QUIET;
 811                 trace_block_rq_abort(q, rq);
 812                 /*
 813                  * Mark this request as started so we don't trigger
 814                  * any debug logic in the end I/O path.
 815                  */
 816                 blk_start_request(rq);
 817                 __blk_end_request_all(rq, -EIO);
 818         }
 819 }
 820 EXPORT_SYMBOL(elv_abort_queue);
 821
 822 void elv_completed_request(struct request_queue *q, struct request *rq)
 823 {
 824         struct elevator_queue *e = q->elevator;
 825
 826         /*
 827          * request is released from the driver, io must be done
 828          */
 829         if (blk_account_rq(rq)) {
 830                 q->in_flight[rq_is_sync(rq)]--;
 831                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 832                         e->ops->elevator_completed_req_fn(q, rq);
 833         }
 834
 835         /*
 836          * Check if the queue is waiting for fs requests to be
 837          * drained for flush sequence.
 838          */
 839         if (unlikely(q->ordseq)) {
 840                 struct request *next = NULL;
 841
 842                 if (!list_empty(&q->queue_head))
 843                         next = list_entry_rq(q->queue_head.next);
 844
 845                 if (!queue_in_flight(q) &&
 846                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 847                     (!next || blk_ordered_req_seq(next) > QUEUE_ORDSEQ_DRAIN)) {
 848                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 849                         __blk_run_queue(q);
 850                 }
 851         }
 852 }
 853
 854 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 855
 856 static ssize_t
 857 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 858 {
 859         struct elv_fs_entry *entry = to_elv(attr);
 860         struct elevator_queue *e;
 861         ssize_t error;
 862
 863         if (!entry->show)
 864                 return -EIO;
 865
 866         e = container_of(kobj, struct elevator_queue, kobj);
 867         mutex_lock(&e->sysfs_lock);
 868         error = e->ops ? entry->show(e, page) : -ENOENT;
 869         mutex_unlock(&e->sysfs_lock);
 870         return error;
 871 }
 872
 873 static ssize_t
 874 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 875                const char *page, size_t length)
 876 {
 877         struct elv_fs_entry *entry = to_elv(attr);
 878         struct elevator_queue *e;
 879         ssize_t error;
 880
 881         if (!entry->store)
 882                 return -EIO;
 883
 884         e = container_of(kobj, struct elevator_queue, kobj);
 885         mutex_lock(&e->sysfs_lock);
 886         error = e->ops ? entry->store(e, page, length) : -ENOENT;
 887         mutex_unlock(&e->sysfs_lock);
 888         return error;
 889 }
 890
 891 static struct sysfs_ops elv_sysfs_ops = {
 892         .show   = elv_attr_show,
 893         .store  = elv_attr_store,
 894 };
 895
 896 static struct kobj_type elv_ktype = {
 897         .sysfs_ops      = &elv_sysfs_ops,
 898         .release        = elevator_release,
 899 };
 900
 901 int elv_register_queue(struct request_queue *q)
 902 {
 903         struct elevator_queue *e = q->elevator;
 904         int error;
 905
 906         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 907         if (!error) {
 908                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
 909                 if (attr) {
 910                         while (attr->attr.name) {
 911                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 912                                         break;
 913                                 attr++;
 914                         }
 915                 }
 916                 kobject_uevent(&e->kobj, KOBJ_ADD);
 917         }
 918         return error;
 919 }
 920
 921 static void __elv_unregister_queue(struct elevator_queue *e)
 922 {
 923         kobject_uevent(&e->kobj, KOBJ_REMOVE);
 924         kobject_del(&e->kobj);
 925 }
 926
 927 void elv_unregister_queue(struct request_queue *q)
 928 {
 929         if (q)
 930                 __elv_unregister_queue(q->elevator);
 931 }
 932
 933 void elv_register(struct elevator_type *e)
 934 {
 935         char *def = "";
 936
 937         spin_lock(&elv_list_lock);
 938         BUG_ON(elevator_find(e->elevator_name));
 939         list_add_tail(&e->list, &elv_list);
 940         spin_unlock(&elv_list_lock);
 941
 942         if (!strcmp(e->elevator_name, chosen_elevator) ||
 943                         (!*chosen_elevator &&
 944                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 945                                 def = " (default)";
 946
 947         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
 948                                                                 def);
 949 }
 950 EXPORT_SYMBOL_GPL(elv_register);
 951
 952 void elv_unregister(struct elevator_type *e)
 953 {
 954         struct task_struct *g, *p;
 955
 956         /*
 957          * Iterate every thread in the process to remove the io contexts.
 958          */
 959         if (e->ops.trim) {
 960                 read_lock(&tasklist_lock);
 961                 do_each_thread(g, p) {
 962                         task_lock(p);
 963                         if (p->io_context)
 964                                 e->ops.trim(p->io_context);
 965                         task_unlock(p);
 966                 } while_each_thread(g, p);
 967                 read_unlock(&tasklist_lock);
 968         }
 969
 970         spin_lock(&elv_list_lock);
 971         list_del_init(&e->list);
 972         spin_unlock(&elv_list_lock);
 973 }
 974 EXPORT_SYMBOL_GPL(elv_unregister);
 975
 976 /*
 977  * switch to new_e io scheduler. be careful not to introduce deadlocks -
 978  * we don't free the old io scheduler, before we have allocated what we
 979  * need for the new one. this way we have a chance of going back to the old
 980  * one, if the new one fails init for some reason.
 981  */
 982 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
 983 {
 984         struct elevator_queue *old_elevator, *e;
 985         void *data;
 986
 987         /*
 988          * Allocate new elevator
 989          */
 990         e = elevator_alloc(q, new_e);
 991         if (!e)
 992                 return 0;
 993
 994         data = elevator_init_queue(q, e);
 995         if (!data) {
 996                 kobject_put(&e->kobj);
 997                 return 0;
 998         }
 999
1000         /*
1001          * Turn on BYPASS and drain all requests w/ elevator private data
1002          */
1003         spin_lock_irq(q->queue_lock);
1004         elv_quiesce_start(q);
1005
1006         /*
1007          * Remember old elevator.
1008          */
1009         old_elevator = q->elevator;
1010
1011         /*
1012          * attach and start new elevator
1013          */
1014         elevator_attach(q, e, data);
1015
1016         spin_unlock_irq(q->queue_lock);
1017
1018         __elv_unregister_queue(old_elevator);
1019
1020         if (elv_register_queue(q))
1021                 goto fail_register;
1022
1023         /*
1024          * finally exit old elevator and turn off BYPASS.
1025          */
1026         elevator_exit(old_elevator);
1027         spin_lock_irq(q->queue_lock);
1028         elv_quiesce_end(q);
1029         spin_unlock_irq(q->queue_lock);
1030
1031         blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
1032
1033         return 1;
1034
1035 fail_register:
1036         /*
1037          * switch failed, exit the new io scheduler and reattach the old
1038          * one again (along with re-adding the sysfs dir)
1039          */
1040         elevator_exit(e);
1041         q->elevator = old_elevator;
1042         elv_register_queue(q);
1043
1044         spin_lock_irq(q->queue_lock);
1045         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1046         spin_unlock_irq(q->queue_lock);
1047
1048         return 0;
1049 }
1050
1051 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1052                           size_t count)
1053 {
1054         char elevator_name[ELV_NAME_MAX];
1055         struct elevator_type *e;
1056
1057         if (!q->elevator)
1058                 return count;
1059
1060         strlcpy(elevator_name, name, sizeof(elevator_name));
1061         strstrip(elevator_name);
1062
1063         e = elevator_get(elevator_name);
1064         if (!e) {
1065                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1066                 return -EINVAL;
1067         }
1068
1069         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1070                 elevator_put(e);
1071                 return count;
1072         }
1073
1074         if (!elevator_switch(q, e))
1075                 printk(KERN_ERR "elevator: switch to %s failed\n",
1076                                                         elevator_name);
1077         return count;
1078 }
1079
1080 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1081 {
1082         struct elevator_queue *e = q->elevator;
1083         struct elevator_type *elv;
1084         struct elevator_type *__e;
1085         int len = 0;
1086
1087         if (!q->elevator)
1088                 return sprintf(name, "none\n");
1089
1090         elv = e->elevator_type;
1091
1092         spin_lock(&elv_list_lock);
1093         list_for_each_entry(__e, &elv_list, list) {
1094                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1095                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1096                 else
1097                         len += sprintf(name+len, "%s ", __e->elevator_name);
1098         }
1099         spin_unlock(&elv_list_lock);
1100
1101         len += sprintf(len+name, "\n");
1102         return len;
1103 }
1104
1105 struct request *elv_rb_former_request(struct request_queue *q,
1106                                       struct request *rq)
1107 {
1108         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1109
1110         if (rbprev)
1111                 return rb_entry_rq(rbprev);
1112
1113         return NULL;
1114 }
1115 EXPORT_SYMBOL(elv_rb_former_request);
1116
1117 struct request *elv_rb_latter_request(struct request_queue *q,
1118                                       struct request *rq)
1119 {
1120         struct rb_node *rbnext = rb_next(&rq->rb_node);
1121
1122         if (rbnext)
1123                 return rb_entry_rq(rbnext);
1124
1125         return NULL;
1126 }
1127 EXPORT_SYMBOL(elv_rb_latter_request);