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