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
  38 #include <asm/uaccess.h>
  39
  40 static DEFINE_SPINLOCK(elv_list_lock);
  41 static LIST_HEAD(elv_list);
  42
  43 /*
  44  * Merge hash stuff.
  45  */
  46 static const int elv_hash_shift = 6;
  47 #define ELV_HASH_BLOCK(sec)     ((sec) >> 3)
  48 #define ELV_HASH_FN(sec)        \
  49                 (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
  50 #define ELV_HASH_ENTRIES        (1 << elv_hash_shift)
  51 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
  52 #define ELV_ON_HASH(rq)         (!hlist_unhashed(&(rq)->hash))
  53
  54 /*
  55  * Query io scheduler to see if the current process issuing bio may be
  56  * merged with rq.
  57  */
  58 static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
  59 {
  60         struct request_queue *q = rq->q;
  61         elevator_t *e = q->elevator;
  62
  63         if (e->ops->elevator_allow_merge_fn)
  64                 return e->ops->elevator_allow_merge_fn(q, rq, bio);
  65
  66         return 1;
  67 }
  68
  69 /*
  70  * can we safely merge with this request?
  71  */
  72 inline int elv_rq_merge_ok(struct request *rq, struct bio *bio)
  73 {
  74         if (!rq_mergeable(rq))
  75                 return 0;
  76
  77         /*
  78          * different data direction or already started, don't merge
  79          */
  80         if (bio_data_dir(bio) != rq_data_dir(rq))
  81                 return 0;
  82
  83         /*
  84          * must be same device and not a special request
  85          */
  86         if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
  87                 return 0;
  88
  89         if (!elv_iosched_allow_merge(rq, bio))
  90                 return 0;
  91
  92         return 1;
  93 }
  94 EXPORT_SYMBOL(elv_rq_merge_ok);
  95
  96 static inline int elv_try_merge(struct request *__rq, struct bio *bio)
  97 {
  98         int ret = ELEVATOR_NO_MERGE;
  99
 100         /*
 101          * we can merge and sequence is ok, check if it's possible
 102          */
 103         if (elv_rq_merge_ok(__rq, bio)) {
 104                 if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
 105                         ret = ELEVATOR_BACK_MERGE;
 106                 else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
 107                         ret = ELEVATOR_FRONT_MERGE;
 108         }
 109
 110         return ret;
 111 }
 112
 113 static struct elevator_type *elevator_find(const char *name)
 114 {
 115         struct elevator_type *e;
 116
 117         list_for_each_entry(e, &elv_list, list) {
 118                 if (!strcmp(e->elevator_name, name))
 119                         return e;
 120         }
 121
 122         return NULL;
 123 }
 124
 125 static void elevator_put(struct elevator_type *e)
 126 {
 127         module_put(e->elevator_owner);
 128 }
 129
 130 static struct elevator_type *elevator_get(const char *name)
 131 {
 132         struct elevator_type *e;
 133
 134         spin_lock(&elv_list_lock);
 135
 136         e = elevator_find(name);
 137         if (e && !try_module_get(e->elevator_owner))
 138                 e = NULL;
 139
 140         spin_unlock(&elv_list_lock);
 141
 142         return e;
 143 }
 144
 145 static void *elevator_init_queue(struct request_queue *q,
 146                                  struct elevator_queue *eq)
 147 {
 148         return eq->ops->elevator_init_fn(q);
 149 }
 150
 151 static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
 152                            void *data)
 153 {
 154         q->elevator = eq;
 155         eq->elevator_data = data;
 156 }
 157
 158 static char chosen_elevator[16];
 159
 160 static int __init elevator_setup(char *str)
 161 {
 162         /*
 163          * Be backwards-compatible with previous kernels, so users
 164          * won't get the wrong elevator.
 165          */
 166         if (!strcmp(str, "as"))
 167                 strcpy(chosen_elevator, "anticipatory");
 168         else
 169                 strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 170         return 1;
 171 }
 172
 173 __setup("elevator=", elevator_setup);
 174
 175 static struct kobj_type elv_ktype;
 176
 177 static elevator_t *elevator_alloc(struct request_queue *q,
 178                                   struct elevator_type *e)
 179 {
 180         elevator_t *eq;
 181         int i;
 182
 183         eq = kmalloc_node(sizeof(elevator_t), GFP_KERNEL | __GFP_ZERO, q->node);
 184         if (unlikely(!eq))
 185                 goto err;
 186
 187         eq->ops = &e->ops;
 188         eq->elevator_type = e;
 189         kobject_init(&eq->kobj, &elv_ktype);
 190         mutex_init(&eq->sysfs_lock);
 191
 192         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 193                                         GFP_KERNEL, q->node);
 194         if (!eq->hash)
 195                 goto err;
 196
 197         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 198                 INIT_HLIST_HEAD(&eq->hash[i]);
 199
 200         return eq;
 201 err:
 202         kfree(eq);
 203         elevator_put(e);
 204         return NULL;
 205 }
 206
 207 static void elevator_release(struct kobject *kobj)
 208 {
 209         elevator_t *e = container_of(kobj, elevator_t, kobj);
 210
 211         elevator_put(e->elevator_type);
 212         kfree(e->hash);
 213         kfree(e);
 214 }
 215
 216 int elevator_init(struct request_queue *q, char *name)
 217 {
 218         struct elevator_type *e = NULL;
 219         struct elevator_queue *eq;
 220         int ret = 0;
 221         void *data;
 222
 223         INIT_LIST_HEAD(&q->queue_head);
 224         q->last_merge = NULL;
 225         q->end_sector = 0;
 226         q->boundary_rq = NULL;
 227
 228         if (name) {
 229                 e = elevator_get(name);
 230                 if (!e)
 231                         return -EINVAL;
 232         }
 233
 234         if (!e && *chosen_elevator) {
 235                 e = elevator_get(chosen_elevator);
 236                 if (!e)
 237                         printk(KERN_ERR "I/O scheduler %s not found\n",
 238                                                         chosen_elevator);
 239         }
 240
 241         if (!e) {
 242                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
 243                 if (!e) {
 244                         printk(KERN_ERR
 245                                 "Default I/O scheduler not found. " \
 246                                 "Using noop.\n");
 247                         e = elevator_get("noop");
 248                 }
 249         }
 250
 251         eq = elevator_alloc(q, e);
 252         if (!eq)
 253                 return -ENOMEM;
 254
 255         data = elevator_init_queue(q, eq);
 256         if (!data) {
 257                 kobject_put(&eq->kobj);
 258                 return -ENOMEM;
 259         }
 260
 261         elevator_attach(q, eq, data);
 262         return ret;
 263 }
 264 EXPORT_SYMBOL(elevator_init);
 265
 266 void elevator_exit(elevator_t *e)
 267 {
 268         mutex_lock(&e->sysfs_lock);
 269         if (e->ops->elevator_exit_fn)
 270                 e->ops->elevator_exit_fn(e);
 271         e->ops = NULL;
 272         mutex_unlock(&e->sysfs_lock);
 273
 274         kobject_put(&e->kobj);
 275 }
 276 EXPORT_SYMBOL(elevator_exit);
 277
 278 static void elv_activate_rq(struct request_queue *q, struct request *rq)
 279 {
 280         elevator_t *e = q->elevator;
 281
 282         if (e->ops->elevator_activate_req_fn)
 283                 e->ops->elevator_activate_req_fn(q, rq);
 284 }
 285
 286 static void elv_deactivate_rq(struct request_queue *q, struct request *rq)
 287 {
 288         elevator_t *e = q->elevator;
 289
 290         if (e->ops->elevator_deactivate_req_fn)
 291                 e->ops->elevator_deactivate_req_fn(q, rq);
 292 }
 293
 294 static inline void __elv_rqhash_del(struct request *rq)
 295 {
 296         hlist_del_init(&rq->hash);
 297 }
 298
 299 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 300 {
 301         if (ELV_ON_HASH(rq))
 302                 __elv_rqhash_del(rq);
 303 }
 304
 305 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 306 {
 307         elevator_t *e = q->elevator;
 308
 309         BUG_ON(ELV_ON_HASH(rq));
 310         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 311 }
 312
 313 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 314 {
 315         __elv_rqhash_del(rq);
 316         elv_rqhash_add(q, rq);
 317 }
 318
 319 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 320 {
 321         elevator_t *e = q->elevator;
 322         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 323         struct hlist_node *entry, *next;
 324         struct request *rq;
 325
 326         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 327                 BUG_ON(!ELV_ON_HASH(rq));
 328
 329                 if (unlikely(!rq_mergeable(rq))) {
 330                         __elv_rqhash_del(rq);
 331                         continue;
 332                 }
 333
 334                 if (rq_hash_key(rq) == offset)
 335                         return rq;
 336         }
 337
 338         return NULL;
 339 }
 340
 341 /*
 342  * RB-tree support functions for inserting/lookup/removal of requests
 343  * in a sorted RB tree.
 344  */
 345 struct request *elv_rb_add(struct rb_root *root, struct request *rq)
 346 {
 347         struct rb_node **p = &root->rb_node;
 348         struct rb_node *parent = NULL;
 349         struct request *__rq;
 350
 351         while (*p) {
 352                 parent = *p;
 353                 __rq = rb_entry(parent, struct request, rb_node);
 354
 355                 if (rq->sector < __rq->sector)
 356                         p = &(*p)->rb_left;
 357                 else if (rq->sector > __rq->sector)
 358                         p = &(*p)->rb_right;
 359                 else
 360                         return __rq;
 361         }
 362
 363         rb_link_node(&rq->rb_node, parent, p);
 364         rb_insert_color(&rq->rb_node, root);
 365         return NULL;
 366 }
 367 EXPORT_SYMBOL(elv_rb_add);
 368
 369 void elv_rb_del(struct rb_root *root, struct request *rq)
 370 {
 371         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 372         rb_erase(&rq->rb_node, root);
 373         RB_CLEAR_NODE(&rq->rb_node);
 374 }
 375 EXPORT_SYMBOL(elv_rb_del);
 376
 377 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 378 {
 379         struct rb_node *n = root->rb_node;
 380         struct request *rq;
 381
 382         while (n) {
 383                 rq = rb_entry(n, struct request, rb_node);
 384
 385                 if (sector < rq->sector)
 386                         n = n->rb_left;
 387                 else if (sector > rq->sector)
 388                         n = n->rb_right;
 389                 else
 390                         return rq;
 391         }
 392
 393         return NULL;
 394 }
 395 EXPORT_SYMBOL(elv_rb_find);
 396
 397 /*
 398  * Insert rq into dispatch queue of q.  Queue lock must be held on
 399  * entry.  rq is sort instead into the dispatch queue. To be used by
 400  * specific elevators.
 401  */
 402 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 403 {
 404         sector_t boundary;
 405         struct list_head *entry;
 406         int stop_flags;
 407
 408         if (q->last_merge == rq)
 409                 q->last_merge = NULL;
 410
 411         elv_rqhash_del(q, rq);
 412
 413         q->nr_sorted--;
 414
 415         boundary = q->end_sector;
 416         stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
 417         list_for_each_prev(entry, &q->queue_head) {
 418                 struct request *pos = list_entry_rq(entry);
 419
 420                 if (rq_data_dir(rq) != rq_data_dir(pos))
 421                         break;
 422                 if (pos->cmd_flags & stop_flags)
 423                         break;
 424                 if (rq->sector >= boundary) {
 425                         if (pos->sector < boundary)
 426                                 continue;
 427                 } else {
 428                         if (pos->sector >= boundary)
 429                                 break;
 430                 }
 431                 if (rq->sector >= pos->sector)
 432                         break;
 433         }
 434
 435         list_add(&rq->queuelist, entry);
 436 }
 437 EXPORT_SYMBOL(elv_dispatch_sort);
 438
 439 /*
 440  * Insert rq into dispatch queue of q.  Queue lock must be held on
 441  * entry.  rq is added to the back of the dispatch queue. To be used by
 442  * specific elevators.
 443  */
 444 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 445 {
 446         if (q->last_merge == rq)
 447                 q->last_merge = NULL;
 448
 449         elv_rqhash_del(q, rq);
 450
 451         q->nr_sorted--;
 452
 453         q->end_sector = rq_end_sector(rq);
 454         q->boundary_rq = rq;
 455         list_add_tail(&rq->queuelist, &q->queue_head);
 456 }
 457 EXPORT_SYMBOL(elv_dispatch_add_tail);
 458
 459 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 460 {
 461         elevator_t *e = q->elevator;
 462         struct request *__rq;
 463         int ret;
 464
 465         /*
 466          * First try one-hit cache.
 467          */
 468         if (q->last_merge) {
 469                 ret = elv_try_merge(q->last_merge, bio);
 470                 if (ret != ELEVATOR_NO_MERGE) {
 471                         *req = q->last_merge;
 472                         return ret;
 473                 }
 474         }
 475
 476         /*
 477          * See if our hash lookup can find a potential backmerge.
 478          */
 479         __rq = elv_rqhash_find(q, bio->bi_sector);
 480         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 481                 *req = __rq;
 482                 return ELEVATOR_BACK_MERGE;
 483         }
 484
 485         if (e->ops->elevator_merge_fn)
 486                 return e->ops->elevator_merge_fn(q, req, bio);
 487
 488         return ELEVATOR_NO_MERGE;
 489 }
 490
 491 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 492 {
 493         elevator_t *e = q->elevator;
 494
 495         if (e->ops->elevator_merged_fn)
 496                 e->ops->elevator_merged_fn(q, rq, type);
 497
 498         if (type == ELEVATOR_BACK_MERGE)
 499                 elv_rqhash_reposition(q, rq);
 500
 501         q->last_merge = rq;
 502 }
 503
 504 void elv_merge_requests(struct request_queue *q, struct request *rq,
 505                              struct request *next)
 506 {
 507         elevator_t *e = q->elevator;
 508
 509         if (e->ops->elevator_merge_req_fn)
 510                 e->ops->elevator_merge_req_fn(q, rq, next);
 511
 512         elv_rqhash_reposition(q, rq);
 513         elv_rqhash_del(q, next);
 514
 515         q->nr_sorted--;
 516         q->last_merge = rq;
 517 }
 518
 519 void elv_requeue_request(struct request_queue *q, struct request *rq)
 520 {
 521         /*
 522          * it already went through dequeue, we need to decrement the
 523          * in_flight count again
 524          */
 525         if (blk_account_rq(rq)) {
 526                 q->in_flight--;
 527                 if (blk_sorted_rq(rq))
 528                         elv_deactivate_rq(q, rq);
 529         }
 530
 531         rq->cmd_flags &= ~REQ_STARTED;
 532
 533         elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
 534 }
 535
 536 static void elv_drain_elevator(struct request_queue *q)
 537 {
 538         static int printed;
 539         while (q->elevator->ops->elevator_dispatch_fn(q, 1))
 540                 ;
 541         if (q->nr_sorted == 0)
 542                 return;
 543         if (printed++ < 10) {
 544                 printk(KERN_ERR "%s: forced dispatching is broken "
 545                        "(nr_sorted=%u), please report this\n",
 546                        q->elevator->elevator_type->elevator_name, q->nr_sorted);
 547         }
 548 }
 549
 550 void elv_insert(struct request_queue *q, struct request *rq, int where)
 551 {
 552         struct list_head *pos;
 553         unsigned ordseq;
 554         int unplug_it = 1;
 555
 556         blk_add_trace_rq(q, rq, BLK_TA_INSERT);
 557
 558         rq->q = q;
 559
 560         switch (where) {
 561         case ELEVATOR_INSERT_FRONT:
 562                 rq->cmd_flags |= REQ_SOFTBARRIER;
 563
 564                 list_add(&rq->queuelist, &q->queue_head);
 565                 break;
 566
 567         case ELEVATOR_INSERT_BACK:
 568                 rq->cmd_flags |= REQ_SOFTBARRIER;
 569                 elv_drain_elevator(q);
 570                 list_add_tail(&rq->queuelist, &q->queue_head);
 571                 /*
 572                  * We kick the queue here for the following reasons.
 573                  * - The elevator might have returned NULL previously
 574                  *   to delay requests and returned them now.  As the
 575                  *   queue wasn't empty before this request, ll_rw_blk
 576                  *   won't run the queue on return, resulting in hang.
 577                  * - Usually, back inserted requests won't be merged
 578                  *   with anything.  There's no point in delaying queue
 579                  *   processing.
 580                  */
 581                 blk_remove_plug(q);
 582                 q->request_fn(q);
 583                 break;
 584
 585         case ELEVATOR_INSERT_SORT:
 586                 BUG_ON(!blk_fs_request(rq));
 587                 rq->cmd_flags |= REQ_SORTED;
 588                 q->nr_sorted++;
 589                 if (rq_mergeable(rq)) {
 590                         elv_rqhash_add(q, rq);
 591                         if (!q->last_merge)
 592                                 q->last_merge = rq;
 593                 }
 594
 595                 /*
 596                  * Some ioscheds (cfq) run q->request_fn directly, so
 597                  * rq cannot be accessed after calling
 598                  * elevator_add_req_fn.
 599                  */
 600                 q->elevator->ops->elevator_add_req_fn(q, rq);
 601                 break;
 602
 603         case ELEVATOR_INSERT_REQUEUE:
 604                 /*
 605                  * If ordered flush isn't in progress, we do front
 606                  * insertion; otherwise, requests should be requeued
 607                  * in ordseq order.
 608                  */
 609                 rq->cmd_flags |= REQ_SOFTBARRIER;
 610
 611                 /*
 612                  * Most requeues happen because of a busy condition,
 613                  * don't force unplug of the queue for that case.
 614                  */
 615                 unplug_it = 0;
 616
 617                 if (q->ordseq == 0) {
 618                         list_add(&rq->queuelist, &q->queue_head);
 619                         break;
 620                 }
 621
 622                 ordseq = blk_ordered_req_seq(rq);
 623
 624                 list_for_each(pos, &q->queue_head) {
 625                         struct request *pos_rq = list_entry_rq(pos);
 626                         if (ordseq <= blk_ordered_req_seq(pos_rq))
 627                                 break;
 628                 }
 629
 630                 list_add_tail(&rq->queuelist, pos);
 631                 break;
 632
 633         default:
 634                 printk(KERN_ERR "%s: bad insertion point %d\n",
 635                        __FUNCTION__, where);
 636                 BUG();
 637         }
 638
 639         if (unplug_it && blk_queue_plugged(q)) {
 640                 int nrq = q->rq.count[READ] + q->rq.count[WRITE]
 641                         - q->in_flight;
 642
 643                 if (nrq >= q->unplug_thresh)
 644                         __generic_unplug_device(q);
 645         }
 646 }
 647
 648 void __elv_add_request(struct request_queue *q, struct request *rq, int where,
 649                        int plug)
 650 {
 651         if (q->ordcolor)
 652                 rq->cmd_flags |= REQ_ORDERED_COLOR;
 653
 654         if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
 655                 /*
 656                  * toggle ordered color
 657                  */
 658                 if (blk_barrier_rq(rq))
 659                         q->ordcolor ^= 1;
 660
 661                 /*
 662                  * barriers implicitly indicate back insertion
 663                  */
 664                 if (where == ELEVATOR_INSERT_SORT)
 665                         where = ELEVATOR_INSERT_BACK;
 666
 667                 /*
 668                  * this request is scheduling boundary, update
 669                  * end_sector
 670                  */
 671                 if (blk_fs_request(rq)) {
 672                         q->end_sector = rq_end_sector(rq);
 673                         q->boundary_rq = rq;
 674                 }
 675         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 676                     where == ELEVATOR_INSERT_SORT)
 677                 where = ELEVATOR_INSERT_BACK;
 678
 679         if (plug)
 680                 blk_plug_device(q);
 681
 682         elv_insert(q, rq, where);
 683 }
 684 EXPORT_SYMBOL(__elv_add_request);
 685
 686 void elv_add_request(struct request_queue *q, struct request *rq, int where,
 687                      int plug)
 688 {
 689         unsigned long flags;
 690
 691         spin_lock_irqsave(q->queue_lock, flags);
 692         __elv_add_request(q, rq, where, plug);
 693         spin_unlock_irqrestore(q->queue_lock, flags);
 694 }
 695 EXPORT_SYMBOL(elv_add_request);
 696
 697 static inline struct request *__elv_next_request(struct request_queue *q)
 698 {
 699         struct request *rq;
 700
 701         while (1) {
 702                 while (!list_empty(&q->queue_head)) {
 703                         rq = list_entry_rq(q->queue_head.next);
 704                         if (blk_do_ordered(q, &rq))
 705                                 return rq;
 706                 }
 707
 708                 if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
 709                         return NULL;
 710         }
 711 }
 712
 713 struct request *elv_next_request(struct request_queue *q)
 714 {
 715         struct request *rq;
 716         int ret;
 717
 718         while ((rq = __elv_next_request(q)) != NULL) {
 719                 /*
 720                  * Kill the empty barrier place holder, the driver must
 721                  * not ever see it.
 722                  */
 723                 if (blk_empty_barrier(rq)) {
 724                         end_queued_request(rq, 1);
 725                         continue;
 726                 }
 727                 if (!(rq->cmd_flags & REQ_STARTED)) {
 728                         /*
 729                          * This is the first time the device driver
 730                          * sees this request (possibly after
 731                          * requeueing).  Notify IO scheduler.
 732                          */
 733                         if (blk_sorted_rq(rq))
 734                                 elv_activate_rq(q, rq);
 735
 736                         /*
 737                          * just mark as started even if we don't start
 738                          * it, a request that has been delayed should
 739                          * not be passed by new incoming requests
 740                          */
 741                         rq->cmd_flags |= REQ_STARTED;
 742                         blk_add_trace_rq(q, rq, BLK_TA_ISSUE);
 743                 }
 744
 745                 if (!q->boundary_rq || q->boundary_rq == rq) {
 746                         q->end_sector = rq_end_sector(rq);
 747                         q->boundary_rq = NULL;
 748                 }
 749
 750                 if (rq->cmd_flags & REQ_DONTPREP)
 751                         break;
 752
 753                 if (q->dma_drain_size && rq->data_len) {
 754                         /*
 755                          * make sure space for the drain appears we
 756                          * know we can do this because max_hw_segments
 757                          * has been adjusted to be one fewer than the
 758                          * device can handle
 759                          */
 760                         rq->nr_phys_segments++;
 761                         rq->nr_hw_segments++;
 762                 }
 763
 764                 if (!q->prep_rq_fn)
 765                         break;
 766
 767                 ret = q->prep_rq_fn(q, rq);
 768                 if (ret == BLKPREP_OK) {
 769                         break;
 770                 } else if (ret == BLKPREP_DEFER) {
 771                         /*
 772                          * the request may have been (partially) prepped.
 773                          * we need to keep this request in the front to
 774                          * avoid resource deadlock.  REQ_STARTED will
 775                          * prevent other fs requests from passing this one.
 776                          */
 777                         if (q->dma_drain_size && rq->data_len &&
 778                             !(rq->cmd_flags & REQ_DONTPREP)) {
 779                                 /*
 780                                  * remove the space for the drain we added
 781                                  * so that we don't add it again
 782                                  */
 783                                 --rq->nr_phys_segments;
 784                                 --rq->nr_hw_segments;
 785                         }
 786
 787                         rq = NULL;
 788                         break;
 789                 } else if (ret == BLKPREP_KILL) {
 790                         rq->cmd_flags |= REQ_QUIET;
 791                         end_queued_request(rq, 0);
 792                 } else {
 793                         printk(KERN_ERR "%s: bad return=%d\n", __FUNCTION__,
 794                                                                 ret);
 795                         break;
 796                 }
 797         }
 798
 799         return rq;
 800 }
 801 EXPORT_SYMBOL(elv_next_request);
 802
 803 void elv_dequeue_request(struct request_queue *q, struct request *rq)
 804 {
 805         BUG_ON(list_empty(&rq->queuelist));
 806         BUG_ON(ELV_ON_HASH(rq));
 807
 808         list_del_init(&rq->queuelist);
 809
 810         /*
 811          * the time frame between a request being removed from the lists
 812          * and to it is freed is accounted as io that is in progress at
 813          * the driver side.
 814          */
 815         if (blk_account_rq(rq))
 816                 q->in_flight++;
 817 }
 818 EXPORT_SYMBOL(elv_dequeue_request);
 819
 820 int elv_queue_empty(struct request_queue *q)
 821 {
 822         elevator_t *e = q->elevator;
 823
 824         if (!list_empty(&q->queue_head))
 825                 return 0;
 826
 827         if (e->ops->elevator_queue_empty_fn)
 828                 return e->ops->elevator_queue_empty_fn(q);
 829
 830         return 1;
 831 }
 832 EXPORT_SYMBOL(elv_queue_empty);
 833
 834 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 835 {
 836         elevator_t *e = q->elevator;
 837
 838         if (e->ops->elevator_latter_req_fn)
 839                 return e->ops->elevator_latter_req_fn(q, rq);
 840         return NULL;
 841 }
 842
 843 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 844 {
 845         elevator_t *e = q->elevator;
 846
 847         if (e->ops->elevator_former_req_fn)
 848                 return e->ops->elevator_former_req_fn(q, rq);
 849         return NULL;
 850 }
 851
 852 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 853 {
 854         elevator_t *e = q->elevator;
 855
 856         if (e->ops->elevator_set_req_fn)
 857                 return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
 858
 859         rq->elevator_private = NULL;
 860         return 0;
 861 }
 862
 863 void elv_put_request(struct request_queue *q, struct request *rq)
 864 {
 865         elevator_t *e = q->elevator;
 866
 867         if (e->ops->elevator_put_req_fn)
 868                 e->ops->elevator_put_req_fn(rq);
 869 }
 870
 871 int elv_may_queue(struct request_queue *q, int rw)
 872 {
 873         elevator_t *e = q->elevator;
 874
 875         if (e->ops->elevator_may_queue_fn)
 876                 return e->ops->elevator_may_queue_fn(q, rw);
 877
 878         return ELV_MQUEUE_MAY;
 879 }
 880
 881 void elv_completed_request(struct request_queue *q, struct request *rq)
 882 {
 883         elevator_t *e = q->elevator;
 884
 885         /*
 886          * request is released from the driver, io must be done
 887          */
 888         if (blk_account_rq(rq)) {
 889                 q->in_flight--;
 890                 if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
 891                         e->ops->elevator_completed_req_fn(q, rq);
 892         }
 893
 894         /*
 895          * Check if the queue is waiting for fs requests to be
 896          * drained for flush sequence.
 897          */
 898         if (unlikely(q->ordseq)) {
 899                 struct request *first_rq = list_entry_rq(q->queue_head.next);
 900                 if (q->in_flight == 0 &&
 901                     blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
 902                     blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
 903                         blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
 904                         q->request_fn(q);
 905                 }
 906         }
 907 }
 908
 909 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 910
 911 static ssize_t
 912 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 913 {
 914         elevator_t *e = container_of(kobj, elevator_t, kobj);
 915         struct elv_fs_entry *entry = to_elv(attr);
 916         ssize_t error;
 917
 918         if (!entry->show)
 919                 return -EIO;
 920
 921         mutex_lock(&e->sysfs_lock);
 922         error = e->ops ? entry->show(e, page) : -ENOENT;
 923         mutex_unlock(&e->sysfs_lock);
 924         return error;
 925 }
 926
 927 static ssize_t
 928 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 929                const char *page, size_t length)
 930 {
 931         elevator_t *e = container_of(kobj, elevator_t, kobj);
 932         struct elv_fs_entry *entry = to_elv(attr);
 933         ssize_t error;
 934
 935         if (!entry->store)
 936                 return -EIO;
 937
 938         mutex_lock(&e->sysfs_lock);
 939         error = e->ops ? entry->store(e, page, length) : -ENOENT;
 940         mutex_unlock(&e->sysfs_lock);
 941         return error;
 942 }
 943
 944 static struct sysfs_ops elv_sysfs_ops = {
 945         .show   = elv_attr_show,
 946         .store  = elv_attr_store,
 947 };
 948
 949 static struct kobj_type elv_ktype = {
 950         .sysfs_ops      = &elv_sysfs_ops,
 951         .release        = elevator_release,
 952 };
 953
 954 int elv_register_queue(struct request_queue *q)
 955 {
 956         elevator_t *e = q->elevator;
 957         int error;
 958
 959         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 960         if (!error) {
 961                 struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
 962                 if (attr) {
 963                         while (attr->attr.name) {
 964                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 965                                         break;
 966                                 attr++;
 967                         }
 968                 }
 969                 kobject_uevent(&e->kobj, KOBJ_ADD);
 970         }
 971         return error;
 972 }
 973
 974 static void __elv_unregister_queue(elevator_t *e)
 975 {
 976         kobject_uevent(&e->kobj, KOBJ_REMOVE);
 977         kobject_del(&e->kobj);
 978 }
 979
 980 void elv_unregister_queue(struct request_queue *q)
 981 {
 982         if (q)
 983                 __elv_unregister_queue(q->elevator);
 984 }
 985
 986 void elv_register(struct elevator_type *e)
 987 {
 988         char *def = "";
 989
 990         spin_lock(&elv_list_lock);
 991         BUG_ON(elevator_find(e->elevator_name));
 992         list_add_tail(&e->list, &elv_list);
 993         spin_unlock(&elv_list_lock);
 994
 995         if (!strcmp(e->elevator_name, chosen_elevator) ||
 996                         (!*chosen_elevator &&
 997                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 998                                 def = " (default)";
 999
1000         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
1001                                                                 def);
1002 }
1003 EXPORT_SYMBOL_GPL(elv_register);
1004
1005 void elv_unregister(struct elevator_type *e)
1006 {
1007         struct task_struct *g, *p;
1008
1009         /*
1010          * Iterate every thread in the process to remove the io contexts.
1011          */
1012         if (e->ops.trim) {
1013                 read_lock(&tasklist_lock);
1014                 do_each_thread(g, p) {
1015                         task_lock(p);
1016                         if (p->io_context)
1017                                 e->ops.trim(p->io_context);
1018                         task_unlock(p);
1019                 } while_each_thread(g, p);
1020                 read_unlock(&tasklist_lock);
1021         }
1022
1023         spin_lock(&elv_list_lock);
1024         list_del_init(&e->list);
1025         spin_unlock(&elv_list_lock);
1026 }
1027 EXPORT_SYMBOL_GPL(elv_unregister);
1028
1029 /*
1030  * switch to new_e io scheduler. be careful not to introduce deadlocks -
1031  * we don't free the old io scheduler, before we have allocated what we
1032  * need for the new one. this way we have a chance of going back to the old
1033  * one, if the new one fails init for some reason.
1034  */
1035 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
1036 {
1037         elevator_t *old_elevator, *e;
1038         void *data;
1039
1040         /*
1041          * Allocate new elevator
1042          */
1043         e = elevator_alloc(q, new_e);
1044         if (!e)
1045                 return 0;
1046
1047         data = elevator_init_queue(q, e);
1048         if (!data) {
1049                 kobject_put(&e->kobj);
1050                 return 0;
1051         }
1052
1053         /*
1054          * Turn on BYPASS and drain all requests w/ elevator private data
1055          */
1056         spin_lock_irq(q->queue_lock);
1057
1058         set_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1059
1060         elv_drain_elevator(q);
1061
1062         while (q->rq.elvpriv) {
1063                 blk_remove_plug(q);
1064                 q->request_fn(q);
1065                 spin_unlock_irq(q->queue_lock);
1066                 msleep(10);
1067                 spin_lock_irq(q->queue_lock);
1068                 elv_drain_elevator(q);
1069         }
1070
1071         /*
1072          * Remember old elevator.
1073          */
1074         old_elevator = q->elevator;
1075
1076         /*
1077          * attach and start new elevator
1078          */
1079         elevator_attach(q, e, data);
1080
1081         spin_unlock_irq(q->queue_lock);
1082
1083         __elv_unregister_queue(old_elevator);
1084
1085         if (elv_register_queue(q))
1086                 goto fail_register;
1087
1088         /*
1089          * finally exit old elevator and turn off BYPASS.
1090          */
1091         elevator_exit(old_elevator);
1092         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1093         return 1;
1094
1095 fail_register:
1096         /*
1097          * switch failed, exit the new io scheduler and reattach the old
1098          * one again (along with re-adding the sysfs dir)
1099          */
1100         elevator_exit(e);
1101         q->elevator = old_elevator;
1102         elv_register_queue(q);
1103         clear_bit(QUEUE_FLAG_ELVSWITCH, &q->queue_flags);
1104         return 0;
1105 }
1106
1107 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
1108                           size_t count)
1109 {
1110         char elevator_name[ELV_NAME_MAX];
1111         size_t len;
1112         struct elevator_type *e;
1113
1114         elevator_name[sizeof(elevator_name) - 1] = '\0';
1115         strncpy(elevator_name, name, sizeof(elevator_name) - 1);
1116         len = strlen(elevator_name);
1117
1118         if (len && elevator_name[len - 1] == '\n')
1119                 elevator_name[len - 1] = '\0';
1120
1121         e = elevator_get(elevator_name);
1122         if (!e) {
1123                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
1124                 return -EINVAL;
1125         }
1126
1127         if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
1128                 elevator_put(e);
1129                 return count;
1130         }
1131
1132         if (!elevator_switch(q, e))
1133                 printk(KERN_ERR "elevator: switch to %s failed\n",
1134                                                         elevator_name);
1135         return count;
1136 }
1137
1138 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1139 {
1140         elevator_t *e = q->elevator;
1141         struct elevator_type *elv = e->elevator_type;
1142         struct elevator_type *__e;
1143         int len = 0;
1144
1145         spin_lock(&elv_list_lock);
1146         list_for_each_entry(__e, &elv_list, list) {
1147                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1148                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1149                 else
1150                         len += sprintf(name+len, "%s ", __e->elevator_name);
1151         }
1152         spin_unlock(&elv_list_lock);
1153
1154         len += sprintf(len+name, "\n");
1155         return len;
1156 }
1157
1158 struct request *elv_rb_former_request(struct request_queue *q,
1159                                       struct request *rq)
1160 {
1161         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1162
1163         if (rbprev)
1164                 return rb_entry_rq(rbprev);
1165
1166         return NULL;
1167 }
1168 EXPORT_SYMBOL(elv_rb_former_request);
1169
1170 struct request *elv_rb_latter_request(struct request_queue *q,
1171                                       struct request *rq)
1172 {
1173         struct rb_node *rbnext = rb_next(&rq->rb_node);
1174
1175         if (rbnext)
1176                 return rb_entry_rq(rbnext);
1177
1178         return NULL;
1179 }
1180 EXPORT_SYMBOL(elv_rb_latter_request);