block/elevator.c

   1 /*
   2  *  Block device elevator/IO-scheduler.
   3  *
   4  *  Copyright (C) 2000 Andrea Arcangeli <[email protected]> SuSE
   5  *
   6  * 30042000 Jens Axboe <[email protected]> :
   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 <[email protected]> :
  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->bi_rw & REQ_DISCARD) != (rq->bio->bi_rw & REQ_DISCARD))
  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                 snprintf(elv, sizeof(elv), "%s-iosched", name);
 157
 158                 request_module("%s", elv);
 159                 spin_lock(&elv_list_lock);
 160                 e = elevator_find(name);
 161         }
 162
 163         if (e && !try_module_get(e->elevator_owner))
 164                 e = NULL;
 165
 166         spin_unlock(&elv_list_lock);
 167
 168         return e;
 169 }
 170
 171 static void *elevator_init_queue(struct request_queue *q,
 172                                  struct elevator_queue *eq)
 173 {
 174         return eq->ops->elevator_init_fn(q);
 175 }
 176
 177 static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
 178                            void *data)
 179 {
 180         q->elevator = eq;
 181         eq->elevator_data = data;
 182 }
 183
 184 static char chosen_elevator[16];
 185
 186 static int __init elevator_setup(char *str)
 187 {
 188         /*
 189          * Be backwards-compatible with previous kernels, so users
 190          * won't get the wrong elevator.
 191          */
 192         strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 193         return 1;
 194 }
 195
 196 __setup("elevator=", elevator_setup);
 197
 198 static struct kobj_type elv_ktype;
 199
 200 static struct elevator_queue *elevator_alloc(struct request_queue *q,
 201                                   struct elevator_type *e)
 202 {
 203         struct elevator_queue *eq;
 204         int i;
 205
 206         eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node);
 207         if (unlikely(!eq))
 208                 goto err;
 209
 210         eq->ops = &e->ops;
 211         eq->elevator_type = e;
 212         kobject_init(&eq->kobj, &elv_ktype);
 213         mutex_init(&eq->sysfs_lock);
 214
 215         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 216                                         GFP_KERNEL, q->node);
 217         if (!eq->hash)
 218                 goto err;
 219
 220         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 221                 INIT_HLIST_HEAD(&eq->hash[i]);
 222
 223         return eq;
 224 err:
 225         kfree(eq);
 226         elevator_put(e);
 227         return NULL;
 228 }
 229
 230 static void elevator_release(struct kobject *kobj)
 231 {
 232         struct elevator_queue *e;
 233
 234         e = container_of(kobj, struct elevator_queue, kobj);
 235         elevator_put(e->elevator_type);
 236         kfree(e->hash);
 237         kfree(e);
 238 }
 239
 240 int elevator_init(struct request_queue *q, char *name)
 241 {
 242         struct elevator_type *e = NULL;
 243         struct elevator_queue *eq;
 244         void *data;
 245
 246         if (unlikely(q->elevator))
 247                 return 0;
 248
 249         INIT_LIST_HEAD(&q->queue_head);
 250         q->last_merge = NULL;
 251         q->end_sector = 0;
 252         q->boundary_rq = NULL;
 253
 254         if (name) {
 255                 e = elevator_get(name);
 256                 if (!e)
 257                         return -EINVAL;
 258         }
 259
 260         if (!e && *chosen_elevator) {
 261                 e = elevator_get(chosen_elevator);
 262                 if (!e)
 263                         printk(KERN_ERR "I/O scheduler %s not found\n",
 264                                                         chosen_elevator);
 265         }
 266
 267         if (!e) {
 268                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
 269                 if (!e) {
 270                         printk(KERN_ERR
 271                                 "Default I/O scheduler not found. " \
 272                                 "Using noop.\n");
 273                         e = elevator_get("noop");
 274                 }
 275         }
 276
 277         eq = elevator_alloc(q, e);
 278         if (!eq)
 279                 return -ENOMEM;
 280
 281         data = elevator_init_queue(q, eq);
 282         if (!data) {
 283                 kobject_put(&eq->kobj);
 284                 return -ENOMEM;
 285         }
 286
 287         elevator_attach(q, eq, data);
 288         return 0;
 289 }
 290 EXPORT_SYMBOL(elevator_init);
 291
 292 void elevator_exit(struct elevator_queue *e)
 293 {
 294         mutex_lock(&e->sysfs_lock);
 295         if (e->ops->elevator_exit_fn)
 296                 e->ops->elevator_exit_fn(e);
 297         e->ops = NULL;
 298         mutex_unlock(&e->sysfs_lock);
 299
 300         kobject_put(&e->kobj);
 301 }
 302 EXPORT_SYMBOL(elevator_exit);
 303
 304 static inline void __elv_rqhash_del(struct request *rq)
 305 {
 306         hlist_del_init(&rq->hash);
 307 }
 308
 309 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 310 {
 311         if (ELV_ON_HASH(rq))
 312                 __elv_rqhash_del(rq);
 313 }
 314
 315 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 316 {
 317         struct elevator_queue *e = q->elevator;
 318
 319         BUG_ON(ELV_ON_HASH(rq));
 320         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 321 }
 322
 323 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 324 {
 325         __elv_rqhash_del(rq);
 326         elv_rqhash_add(q, rq);
 327 }
 328
 329 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 330 {
 331         struct elevator_queue *e = q->elevator;
 332         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 333         struct hlist_node *entry, *next;
 334         struct request *rq;
 335
 336         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 337                 BUG_ON(!ELV_ON_HASH(rq));
 338
 339                 if (unlikely(!rq_mergeable(rq))) {
 340                         __elv_rqhash_del(rq);
 341                         continue;
 342                 }
 343
 344                 if (rq_hash_key(rq) == offset)
 345                         return rq;
 346         }
 347
 348         return NULL;
 349 }
 350
 351 /*
 352  * RB-tree support functions for inserting/lookup/removal of requests
 353  * in a sorted RB tree.
 354  */
 355 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 356 {
 357         struct rb_node **p = &root->rb_node;
 358         struct rb_node *parent = NULL;
 359         struct request *__rq;
 360
 361         while (*p) {
 362                 parent = *p;
 363                 __rq = rb_entry(parent, struct request, rb_node);
 364
 365                 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
 366                         p = &(*p)->rb_left;
 367                 else if (blk_rq_pos(rq) > blk_rq_pos(__rq))
 368                         p = &(*p)->rb_right;
 369                 else
 370                         return __rq;
 371         }
 372
 373         rb_link_node(&rq->rb_node, parent, p);
 374         rb_insert_color(&rq->rb_node, root);
 375         return NULL;
 376 }
 377 EXPORT_SYMBOL(elv_rb_add);
 378
 379 void elv_rb_del(struct rb_root *root, struct request *rq)
 380 {
 381         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 382         rb_erase(&rq->rb_node, root);
 383         RB_CLEAR_NODE(&rq->rb_node);
 384 }
 385 EXPORT_SYMBOL(elv_rb_del);
 386
 387 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 388 {
 389         struct rb_node *n = root->rb_node;
 390         struct request *rq;
 391
 392         while (n) {
 393                 rq = rb_entry(n, struct request, rb_node);
 394
 395                 if (sector < blk_rq_pos(rq))
 396                         n = n->rb_left;
 397                 else if (sector > blk_rq_pos(rq))
 398                         n = n->rb_right;
 399                 else
 400                         return rq;
 401         }
 402
 403         return NULL;
 404 }
 405 EXPORT_SYMBOL(elv_rb_find);
 406
 407 /*
 408  * Insert rq into dispatch queue of q.  Queue lock must be held on
 409  * entry.  rq is sort instead into the dispatch queue. To be used by
 410  * specific elevators.
 411  */
 412 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 413 {
 414         sector_t boundary;
 415         struct list_head *entry;
 416         int stop_flags;
 417
 418         if (q->last_merge == rq)
 419                 q->last_merge = NULL;
 420
 421         elv_rqhash_del(q, rq);
 422
 423         q->nr_sorted--;
 424
 425         boundary = q->end_sector;
 426         stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
 427         list_for_each_prev(entry, &q->queue_head) {
 428                 struct request *pos = list_entry_rq(entry);
 429
 430                 if ((rq->cmd_flags & REQ_DISCARD) !=
 431                     (pos->cmd_flags & REQ_DISCARD))
 432                         break;
 433                 if (rq_data_dir(rq) != rq_data_dir(pos))
 434                         break;
 435                 if (pos->cmd_flags & stop_flags)
 436                         break;
 437                 if (blk_rq_pos(rq) >= boundary) {
 438                         if (blk_rq_pos(pos) < boundary)
 439                                 continue;
 440                 } else {
 441                         if (blk_rq_pos(pos) >= boundary)
 442                                 break;
 443                 }
 444                 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
 445                         break;
 446         }
 447
 448         list_add(&rq->queuelist, entry);
 449 }
 450 EXPORT_SYMBOL(elv_dispatch_sort);
 451
 452 /*
 453  * Insert rq into dispatch queue of q.  Queue lock must be held on
 454  * entry.  rq is added to the back of the dispatch queue. To be used by
 455  * specific elevators.
 456  */
 457 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 458 {
 459         if (q->last_merge == rq)
 460                 q->last_merge = NULL;
 461
 462         elv_rqhash_del(q, rq);
 463
 464         q->nr_sorted--;
 465
 466         q->end_sector = rq_end_sector(rq);
 467         q->boundary_rq = rq;
 468         list_add_tail(&rq->queuelist, &q->queue_head);
 469 }
 470 EXPORT_SYMBOL(elv_dispatch_add_tail);
 471
 472 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 473 {
 474         struct elevator_queue *e = q->elevator;
 475         struct request *__rq;
 476         int ret;
 477
 478         /*
 479          * Levels of merges:
 480          *      nomerges:  No merges at all attempted
 481          *      noxmerges: Only simple one-hit cache try
 482          *      merges:    All merge tries attempted
 483          */
 484         if (blk_queue_nomerges(q))
 485                 return ELEVATOR_NO_MERGE;
 486
 487         /*
 488          * First try one-hit cache.
 489          */
 490         if (q->last_merge) {
 491                 ret = elv_try_merge(q->last_merge, bio);
 492                 if (ret != ELEVATOR_NO_MERGE) {
 493                         *req = q->last_merge;
 494                         return ret;
 495                 }
 496         }
 497
 498         if (blk_queue_noxmerges(q))
 499                 return ELEVATOR_NO_MERGE;
 500
 501         /*
 502          * See if our hash lookup can find a potential backmerge.
 503          */
 504         __rq = elv_rqhash_find(q, bio->bi_sector);
 505         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 506                 *req = __rq;
 507                 return ELEVATOR_BACK_MERGE;
 508         }
 509
 510         if (e->ops->elevator_merge_fn)
 511                 return e->ops->elevator_merge_fn(q, req, bio);
 512
 513         return ELEVATOR_NO_MERGE;
 514 }
 515
 516 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 517 {
 518         struct elevator_queue *e = q->elevator;
 519
 520         if (e->ops->elevator_merged_fn)
 521                 e->ops->elevator_merged_fn(q, rq, type);
 522
 523         if (type == ELEVATOR_BACK_MERGE)
 524                 elv_rqhash_reposition(q, rq);
 525
 526         q->last_merge = rq;
 527 }
 528
 529 void elv_merge_requests(struct request_queue *q, struct request *rq,
 530                              struct request *next)
 531 {
 532         struct elevator_queue *e = q->elevator;
 533
 534         if (e->ops->elevator_merge_req_fn)
 535                 e->ops->elevator_merge_req_fn(q, rq, next);
 536
 537         elv_rqhash_reposition(q, rq);
 538         elv_rqhash_del(q, next);
 539
 540         q->nr_sorted--;
 541         q->last_merge = rq;
 542 }
 543
 544 void elv_bio_merged(struct request_queue *q, struct request *rq,
 545                         struct bio *bio)
 546 {
 547         struct elevator_queue *e = q->elevator;
 548
 549         if (e->ops->elevator_bio_merged_fn)
 550                 e->ops->elevator_bio_merged_fn(q, rq, bio);
 551 }
 552
 553 void elv_requeue_request(struct request_queue *q, struct request *rq)
 554 {
 555         /*
 556          * it already went through dequeue, we need to decrement the
 557          * in_flight count again
 558          */
 559         if (blk_account_rq(rq)) {
 560                 q->in_flight[rq_is_sync(rq)]--;
 561                 if (rq->cmd_flags & REQ_SORTED)
 562                         elv_deactivate_rq(q, rq);
 563         }
 564
 565         rq->cmd_flags &= ~REQ_STARTED;
 566
 567         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 568 }
 569
 570 void elv_drain_elevator(struct request_queue *q)
 571 {
 572         static int printed;
 573         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 574                 ;
 575         if (q->nr_sorted == 0)
 576                 return;
 577         if (printed++ < 10) {
 578                 printk(KERN_ERR "%s: forced dispatching is broken "
 579                        "(nr_sorted=%u), please report this\n",
 580                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 581         }
 582 }
 583
 584 /*
 585  * Call with queue lock held, interrupts disabled
 586  */
 587 void elv_quiesce_start(struct request_queue *q)
 588 {
 589         if (!q->elevator)
 590                 return;
 591
 592         queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
 593
 594         /*
 595          * make sure we don't have any requests in flight
 596          */
 597         elv_drain_elevator(q);
 598         while (q->rq.elvpriv) {
 599                 __blk_run_queue(q);
 600                 spin_unlock_irq(q->queue_lock);
 601                 msleep(10);
 602                 spin_lock_irq(q->queue_lock);
 603                 elv_drain_elevator(q);
 604         }
 605 }
 606
 607 void elv_quiesce_end(struct request_queue *q)
 608 {
 609         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
 610 }
 611
 612 void elv_insert(struct request_queue *q, struct request *rq, int where)
 613 {
 614         struct list_head *pos;
 615         unsigned ordseq;
 616         int unplug_it = 1;
 617
 618         trace_block_rq_insert(q, rq);
 619
 620         rq->q = q;
 621
 622         switch (where) {
 623         case ELEVATOR_INSERT_FRONT:
 624                 rq->cmd_flags |= REQ_SOFTBARRIER;
 625
 626                 list_add(&rq->queuelist, &q->queue_head);
 627                 break;
 628
 629         case ELEVATOR_INSERT_BACK:
 630                 rq->cmd_flags |= REQ_SOFTBARRIER;
 631                 elv_drain_elevator(q);
 632                 list_add_tail(&rq->queuelist, &q->queue_head);
 633                 /*
 634                  * We kick the queue here for the following reasons.
 635                  * - The elevator might have returned NULL previously
 636                  *   to delay requests and returned them now.  As the
 637                  *   queue wasn't empty before this request, ll_rw_blk
 638                  *   won't run the queue on return, resulting in hang.
 639                  * - Usually, back inserted requests won't be merged
 640                  *   with anything.  There's no point in delaying queue
 641                  *   processing.
 642                  */
 643                 __blk_run_queue(q);
 644                 break;
 645
 646         case ELEVATOR_INSERT_SORT:
 647                 BUG_ON(rq->cmd_type != REQ_TYPE_FS &&
 648                        !(rq->cmd_flags & REQ_DISCARD));
 649                 rq->cmd_flags |= REQ_SORTED;
 650                 q->nr_sorted++;
 651                 if (rq_mergeable(rq)) {
 652                         elv_rqhash_add(q, rq);
 653                         if (!q->last_merge)
 654                                 q->last_merge = rq;
 655                 }
 656
 657                 /*
 658                  * Some ioscheds (cfq) run q->request_fn directly, so
 659                  * rq cannot be accessed after calling
 660                  * elevator_add_req_fn.
 661                  */
 662                 q->elevator->ops->elevator_add_req_fn(q, rq);
 663                 break;
 664
 665         case ELEVATOR_INSERT_REQUEUE:
 666                 /*
 667                  * If ordered flush isn't in progress, we do front
 668                  * insertion; otherwise, requests should be requeued
 669                  * in ordseq order.
 670                  */
 671                 rq->cmd_flags |= REQ_SOFTBARRIER;
 672
 673                 /*
 674                  * Most requeues happen because of a busy condition,
 675                  * don't force unplug of the queue for that case.
 676                  */
 677                 unplug_it = 0;
 678
 679                 if (q->ordseq == 0) {
 680                         list_add(&rq->queuelist, &q->queue_head);
 681                         break;
 682                 }
 683
 684                 ordseq = blk_ordered_req_seq(rq);
 685
 686                 list_for_each(pos, &q->queue_head) {
 687                         struct request *pos_rq = list_entry_rq(pos);
 688                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 689                                 break;
 690                 }
 691
 692                 list_add_tail(&rq->queuelist, pos);
 693                 break;
 694
 695         default:
 696                 printk(KERN_ERR "%s: bad insertion point %d\n",
 697                        __func__, where);
 698                 BUG();
 699         }
 700
 701         if (unplug_it && blk_queue_plugged(q)) {
 702                 int nrq = q->rq.count[BLK_RW_SYNC] + q->rq.count[BLK_RW_ASYNC]
 703                                 - queue_in_flight(q);
 704
 705                 if (nrq >= q->unplug_thresh)
 706                         __generic_unplug_device(q);
 707         }
 708 }
 709
 710 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
 711                        int plug)
 712 {
 713         if (q->ordcolor)
 714                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 715
 716         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 717                 /*
 718                  * toggle ordered color
 719                  */
 720                 if (rq->cmd_flags & REQ_HARDBARRIER)
 721                         q->ordcolor ^= 1;
 722
 723                 /*
 724                  * barriers implicitly indicate back insertion
 725                  */
 726                 if (where == ELEVATOR_INSERT_SORT)
 727                         where = ELEVATOR_INSERT_BACK;
 728
 729                 /*
 730                  * this request is scheduling boundary, update
 731                  * end_sector
 732                  */
 733                 if (rq->cmd_type == REQ_TYPE_FS ||
 734                     (rq->cmd_flags & REQ_DISCARD)) {
 735                         q->end_sector = rq_end_sector(rq);
 736                         q->boundary_rq = rq;
 737                 }
 738         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 739                     where == ELEVATOR_INSERT_SORT)
 740                 where = ELEVATOR_INSERT_BACK;
 741
 742         if (plug)
 743                 blk_plug_device(q);
 744
 745         elv_insert(q, rq, where);
 746 }
 747 EXPORT_SYMBOL(__elv_add_request);
 748
 749 void elv_add_request(struct request_queue *q, struct request *rq, int where,
 750                      int plug)
 751 {
 752         unsigned long flags;
 753
 754         spin_lock_irqsave(q->queue_lock, flags);
 755         __elv_add_request(q, rq, where, plug);
 756         spin_unlock_irqrestore(q->queue_lock, flags);
 757 }
 758 EXPORT_SYMBOL(elv_add_request);
 759
 760 int elv_queue_empty(struct request_queue *q)
 761 {
 762         struct elevator_queue *e = q->elevator;
 763
 764         if (!list_empty(&q->queue_head))
 765                 return 0;
 766
 767         if (e->ops->elevator_queue_empty_fn)
 768                 return e->ops->elevator_queue_empty_fn(q);
 769
 770         return 1;
 771 }
 772 EXPORT_SYMBOL(elv_queue_empty);
 773
 774 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 775 {
 776         struct elevator_queue *e = q->elevator;
 777
 778         if (e->ops->elevator_latter_req_fn)
 779                 return e->ops->elevator_latter_req_fn(q, rq);
 780         return NULL;
 781 }
 782
 783 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 784 {
 785         struct elevator_queue *e = q->elevator;
 786
 787         if (e->ops->elevator_former_req_fn)
 788                 return e->ops->elevator_former_req_fn(q, rq);
 789         return NULL;
 790 }
 791
 792 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 793 {
 794         struct elevator_queue *e = q->elevator;
 795
 796         if (e->ops->elevator_set_req_fn)
 797                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 798
 799         rq->elevator_private = NULL;
 800         return 0;
 801 }
 802
 803 void elv_put_request(struct request_queue *q, struct request *rq)
 804 {
 805         struct elevator_queue *e = q->elevator;
 806
 807         if (e->ops->elevator_put_req_fn)
 808                 e->ops->elevator_put_req_fn(rq);
 809 }
 810
 811 int elv_may_queue(struct request_queue *q, int rw)
 812 {
 813         struct elevator_queue *e = q->elevator;
 814
 815         if (e->ops->elevator_may_queue_fn)
 816                 return e->ops->elevator_may_queue_fn(q, rw);
 817
 818         return ELV_MQUEUE_MAY;
 819 }
 820
 821 void elv_abort_queue(struct request_queue *q)
 822 {
 823         struct request *rq;
 824
 825         while (!list_empty(&q->queue_head)) {
 826                 rq = list_entry_rq(q->queue_head.next);
 827                 rq->cmd_flags |= REQ_QUIET;
 828                 trace_block_rq_abort(q, rq);
 829                 /*
 830                  * Mark this request as started so we don't trigger
 831                  * any debug logic in the end I/O path.
 832                  */
 833                 blk_start_request(rq);
 834                 __blk_end_request_all(rq, -EIO);
 835         }
 836 }
 837 EXPORT_SYMBOL(elv_abort_queue);
 838
 839 void elv_completed_request(struct request_queue *q, struct request *rq)
 840 {
 841         struct elevator_queue *e = q->elevator;
 842
 843         /*
 844          * request is released from the driver, io must be done
 845          */
 846         if (blk_account_rq(rq)) {
 847                 q->in_flight[rq_is_sync(rq)]--;
 848                 if ((rq->cmd_flags & REQ_SORTED) &&
 849                     e->ops->elevator_completed_req_fn)
 850                         e->ops->elevator_completed_req_fn(q, rq);
 851         }
 852
 853         /*
 854          * Check if the queue is waiting for fs requests to be
 855          * drained for flush sequence.
 856          */
 857         if (unlikely(q->ordseq)) {
 858                 struct request *next = NULL;
 859
 860                 if (!list_empty(&q->queue_head))
 861                         next = list_entry_rq(q->queue_head.next);
 862
 863                 if (!queue_in_flight(q) &&
 864                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 865                     (!next || blk_ordered_req_seq(next) > QUEUE_ORDSEQ_DRAIN)) {
 866                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 867                         __blk_run_queue(q);
 868                 }
 869         }
 870 }
 871
 872 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 873
 874 static ssize_t
 875 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 876 {
 877         struct elv_fs_entry *entry = to_elv(attr);
 878         struct elevator_queue *e;
 879         ssize_t error;
 880
 881         if (!entry->show)
 882                 return -EIO;
 883
 884         e = container_of(kobj, struct elevator_queue, kobj);
 885         mutex_lock(&e->sysfs_lock);
 886         error = e->ops ? entry->show(e, page) : -ENOENT;
 887         mutex_unlock(&e->sysfs_lock);
 888         return error;
 889 }
 890
 891 static ssize_t
 892 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 893                const char *page, size_t length)
 894 {
 895         struct elv_fs_entry *entry = to_elv(attr);
 896         struct elevator_queue *e;
 897         ssize_t error;
 898
 899         if (!entry->store)
 900                 return -EIO;
 901
 902         e = container_of(kobj, struct elevator_queue, kobj);
 903         mutex_lock(&e->sysfs_lock);
 904         error = e->ops ? entry->store(e, page, length) : -ENOENT;
 905         mutex_unlock(&e->sysfs_lock);
 906         return error;
 907 }
 908
 909 static const struct sysfs_ops elv_sysfs_ops = {
 910         .show   = elv_attr_show,
 911         .store  = elv_attr_store,
 912 };
 913
 914 static struct kobj_type elv_ktype = {
 915         .sysfs_ops      = &elv_sysfs_ops,
 916         .release        = elevator_release,
 917 };
 918
 919 int elv_register_queue(struct request_queue *q)
 920 {
 921         struct elevator_queue *e = q->elevator;
 922         int error;
 923
 924         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 925         if (!error) {
 926                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
 927                 if (attr) {
 928                         while (attr->attr.name) {
 929                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 930                                         break;
 931                                 attr++;
 932                         }
 933                 }
 934                 kobject_uevent(&e->kobj, KOBJ_ADD);
 935         }
 936         return error;
 937 }
 938 EXPORT_SYMBOL(elv_register_queue);
 939
 940 static void __elv_unregister_queue(struct elevator_queue *e)
 941 {
 942         kobject_uevent(&e->kobj, KOBJ_REMOVE);
 943         kobject_del(&e->kobj);
 944 }
 945
 946 void elv_unregister_queue(struct request_queue *q)
 947 {
 948         if (q)
 949                 __elv_unregister_queue(q->elevator);
 950 }
 951 EXPORT_SYMBOL(elv_unregister_queue);
 952
 953 void elv_register(struct elevator_type *e)
 954 {
 955         char *def = "";
 956
 957         spin_lock(&elv_list_lock);
 958         BUG_ON(elevator_find(e->elevator_name));
 959         list_add_tail(&e->list, &elv_list);
 960         spin_unlock(&elv_list_lock);
 961
 962         if (!strcmp(e->elevator_name, chosen_elevator) ||
 963                         (!*chosen_elevator &&
 964                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 965                                 def = " (default)";
 966
 967         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
 968                                                                 def);
 969 }
 970 EXPORT_SYMBOL_GPL(elv_register);
 971
 972 void elv_unregister(struct elevator_type *e)
 973 {
 974         struct task_struct *g, *p;
 975
 976         /*
 977          * Iterate every thread in the process to remove the io contexts.
 978          */
 979         if (e->ops.trim) {
 980                 read_lock(&tasklist_lock);
 981                 do_each_thread(g, p) {
 982                         task_lock(p);
 983                         if (p->io_context)
 984                                 e->ops.trim(p->io_context);
 985                         task_unlock(p);
 986                 } while_each_thread(g, p);
 987                 read_unlock(&tasklist_lock);
 988         }
 989
 990         spin_lock(&elv_list_lock);
 991         list_del_init(&e->list);
 992         spin_unlock(&elv_list_lock);
 993 }
 994 EXPORT_SYMBOL_GPL(elv_unregister);
 995
 996 /*
 997  * switch to new_e io scheduler. be careful not to introduce deadlocks -
 998  * we don't free the old io scheduler, before we have allocated what we
 999  * need for the new one. this way we have a chance of going back to the old
1000  * one, if the new one fails init for some reason.
1001  */
1002 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1003 {
1004         struct elevator_queue *old_elevator, *e;
1005         void *data;
1006
1007         /*
1008          * Allocate new elevator
1009          */
1010         e = elevator_alloc(q, new_e);
1011         if (!e)
1012                 return 0;
1013
1014         data = elevator_init_queue(q, e);
1015         if (!data) {
1016                 kobject_put(&e->kobj);
1017                 return 0;
1018         }
1019
1020         /*
1021          * Turn on BYPASS and drain all requests w/ elevator private data
1022          */
1023         spin_lock_irq(q->queue_lock);
1024         elv_quiesce_start(q);
1025
1026         /*
1027          * Remember old elevator.
1028          */
1029         old_elevator = q->elevator;
1030
1031         /*
1032          * attach and start new elevator
1033          */
1034         elevator_attach(q, e, data);
1035
1036         spin_unlock_irq(q->queue_lock);
1037
1038         __elv_unregister_queue(old_elevator);
1039
1040         if (elv_register_queue(q))
1041                 goto fail_register;
1042
1043         /*
1044          * finally exit old elevator and turn off BYPASS.
1045          */
1046         elevator_exit(old_elevator);
1047         spin_lock_irq(q->queue_lock);
1048         elv_quiesce_end(q);
1049         spin_unlock_irq(q->queue_lock);
1050
1051         blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
1052
1053         return 1;
1054
1055 fail_register:
1056         /*
1057          * switch failed, exit the new io scheduler and reattach the old
1058          * one again (along with re-adding the sysfs dir)
1059          */
1060         elevator_exit(e);
1061         q->elevator = old_elevator;
1062         elv_register_queue(q);
1063
1064         spin_lock_irq(q->queue_lock);
1065         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
1066         spin_unlock_irq(q->queue_lock);
1067
1068         return 0;
1069 }
1070
1071 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1072                           size_t count)
1073 {
1074         char elevator_name[ELV_NAME_MAX];
1075         struct elevator_type *e;
1076
1077         if (!q->elevator)
1078                 return count;
1079
1080         strlcpy(elevator_name, name, sizeof(elevator_name));
1081         e = elevator_get(strstrip(elevator_name));
1082         if (!e) {
1083                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1084                 return -EINVAL;
1085         }
1086
1087         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1088                 elevator_put(e);
1089                 return count;
1090         }
1091
1092         if (!elevator_switch(q, e))
1093                 printk(KERN_ERR "elevator: switch to %s failed\n",
1094                                                         elevator_name);
1095         return count;
1096 }
1097
1098 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1099 {
1100         struct elevator_queue *e = q->elevator;
1101         struct elevator_type *elv;
1102         struct elevator_type *__e;
1103         int len = 0;
1104
1105         if (!q->elevator || !blk_queue_stackable(q))
1106                 return sprintf(name, "none\n");
1107
1108         elv = e->elevator_type;
1109
1110         spin_lock(&elv_list_lock);
1111         list_for_each_entry(__e, &elv_list, list) {
1112                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1113                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1114                 else
1115                         len += sprintf(name+len, "%s ", __e->elevator_name);
1116         }
1117         spin_unlock(&elv_list_lock);
1118
1119         len += sprintf(len+name, "\n");
1120         return len;
1121 }
1122
1123 struct request *elv_rb_former_request(struct request_queue *q,
1124                                       struct request *rq)
1125 {
1126         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1127
1128         if (rbprev)
1129                 return rb_entry_rq(rbprev);
1130
1131         return NULL;
1132 }
1133 EXPORT_SYMBOL(elv_rb_former_request);
1134
1135 struct request *elv_rb_latter_request(struct request_queue *q,
1136                                       struct request *rq)
1137 {
1138         struct rb_node *rbnext = rb_next(&rq->rb_node);
1139
1140         if (rbnext)
1141                 return rb_entry_rq(rbnext);
1142
1143         return NULL;
1144 }
1145 EXPORT_SYMBOL(elv_rb_latter_request);