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/blktrace_api.h>
  35 #include <linux/hash.h>
  36 #include <linux/uaccess.h>
  37
  38 #include <trace/events/block.h>
  39
  40 #include "blk.h"
  41
  42 static DEFINE_SPINLOCK(elv_list_lock);
  43 static LIST_HEAD(elv_list);
  44
  45 /*
  46  * Merge hash stuff.
  47  */
  48 static const int elv_hash_shift = 6;
  49 #define ELV_HASH_BLOCK(sec)     ((sec) >> 3)
  50 #define ELV_HASH_FN(sec)        \
  51                 (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
  52 #define ELV_HASH_ENTRIES        (1 << elv_hash_shift)
  53 #define rq_hash_key(rq)         (blk_rq_pos(rq) + blk_rq_sectors(rq))
  54
  55 /*
  56  * Query io scheduler to see if the current process issuing bio may be
  57  * merged with rq.
  58  */
  59 static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
  60 {
  61         struct request_queue *q = rq->q;
  62         struct elevator_queue *e = q->elevator;
  63
  64         if (e->type->ops.elevator_allow_merge_fn)
  65                 return e->type->ops.elevator_allow_merge_fn(q, rq, bio);
  66
  67         return 1;
  68 }
  69
  70 /*
  71  * can we safely merge with this request?
  72  */
  73 bool elv_rq_merge_ok(struct request *rq, struct bio *bio)
  74 {
  75         if (!blk_rq_merge_ok(rq, bio))
  76                 return 0;
  77
  78         if (!elv_iosched_allow_merge(rq, bio))
  79                 return 0;
  80
  81         return 1;
  82 }
  83 EXPORT_SYMBOL(elv_rq_merge_ok);
  84
  85 static struct elevator_type *elevator_find(const char *name)
  86 {
  87         struct elevator_type *e;
  88
  89         list_for_each_entry(e, &elv_list, list) {
  90                 if (!strcmp(e->elevator_name, name))
  91                         return e;
  92         }
  93
  94         return NULL;
  95 }
  96
  97 static void elevator_put(struct elevator_type *e)
  98 {
  99         module_put(e->elevator_owner);
 100 }
 101
 102 static struct elevator_type *elevator_get(const char *name)
 103 {
 104         struct elevator_type *e;
 105
 106         spin_lock(&elv_list_lock);
 107
 108         e = elevator_find(name);
 109         if (!e) {
 110                 spin_unlock(&elv_list_lock);
 111                 request_module("%s-iosched", name);
 112                 spin_lock(&elv_list_lock);
 113                 e = elevator_find(name);
 114         }
 115
 116         if (e && !try_module_get(e->elevator_owner))
 117                 e = NULL;
 118
 119         spin_unlock(&elv_list_lock);
 120
 121         return e;
 122 }
 123
 124 static int elevator_init_queue(struct request_queue *q)
 125 {
 126         q->elevator->elevator_data = q->elevator->type->ops.elevator_init_fn(q);
 127         if (q->elevator->elevator_data)
 128                 return 0;
 129         return -ENOMEM;
 130 }
 131
 132 static char chosen_elevator[ELV_NAME_MAX];
 133
 134 static int __init elevator_setup(char *str)
 135 {
 136         /*
 137          * Be backwards-compatible with previous kernels, so users
 138          * won't get the wrong elevator.
 139          */
 140         strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 141         return 1;
 142 }
 143
 144 __setup("elevator=", elevator_setup);
 145
 146 static struct kobj_type elv_ktype;
 147
 148 static struct elevator_queue *elevator_alloc(struct request_queue *q,
 149                                   struct elevator_type *e)
 150 {
 151         struct elevator_queue *eq;
 152         int i;
 153
 154         eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node);
 155         if (unlikely(!eq))
 156                 goto err;
 157
 158         eq->type = e;
 159         kobject_init(&eq->kobj, &elv_ktype);
 160         mutex_init(&eq->sysfs_lock);
 161
 162         eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
 163                                         GFP_KERNEL, q->node);
 164         if (!eq->hash)
 165                 goto err;
 166
 167         for (i = 0; i < ELV_HASH_ENTRIES; i++)
 168                 INIT_HLIST_HEAD(&eq->hash[i]);
 169
 170         return eq;
 171 err:
 172         kfree(eq);
 173         elevator_put(e);
 174         return NULL;
 175 }
 176
 177 static void elevator_release(struct kobject *kobj)
 178 {
 179         struct elevator_queue *e;
 180
 181         e = container_of(kobj, struct elevator_queue, kobj);
 182         elevator_put(e->type);
 183         kfree(e->hash);
 184         kfree(e);
 185 }
 186
 187 int elevator_init(struct request_queue *q, char *name)
 188 {
 189         struct elevator_type *e = NULL;
 190         int err;
 191
 192         if (unlikely(q->elevator))
 193                 return 0;
 194
 195         INIT_LIST_HEAD(&q->queue_head);
 196         q->last_merge = NULL;
 197         q->end_sector = 0;
 198         q->boundary_rq = NULL;
 199
 200         if (name) {
 201                 e = elevator_get(name);
 202                 if (!e)
 203                         return -EINVAL;
 204         }
 205
 206         if (!e && *chosen_elevator) {
 207                 e = elevator_get(chosen_elevator);
 208                 if (!e)
 209                         printk(KERN_ERR "I/O scheduler %s not found\n",
 210                                                         chosen_elevator);
 211         }
 212
 213         if (!e) {
 214                 e = elevator_get(CONFIG_DEFAULT_IOSCHED);
 215                 if (!e) {
 216                         printk(KERN_ERR
 217                                 "Default I/O scheduler not found. " \
 218                                 "Using noop.\n");
 219                         e = elevator_get("noop");
 220                 }
 221         }
 222
 223         q->elevator = elevator_alloc(q, e);
 224         if (!q->elevator)
 225                 return -ENOMEM;
 226
 227         err = elevator_init_queue(q);
 228         if (err) {
 229                 kobject_put(&q->elevator->kobj);
 230                 return err;
 231         }
 232
 233         return 0;
 234 }
 235 EXPORT_SYMBOL(elevator_init);
 236
 237 void elevator_exit(struct elevator_queue *e)
 238 {
 239         mutex_lock(&e->sysfs_lock);
 240         if (e->type->ops.elevator_exit_fn)
 241                 e->type->ops.elevator_exit_fn(e);
 242         mutex_unlock(&e->sysfs_lock);
 243
 244         kobject_put(&e->kobj);
 245 }
 246 EXPORT_SYMBOL(elevator_exit);
 247
 248 static inline void __elv_rqhash_del(struct request *rq)
 249 {
 250         hlist_del_init(&rq->hash);
 251 }
 252
 253 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 254 {
 255         if (ELV_ON_HASH(rq))
 256                 __elv_rqhash_del(rq);
 257 }
 258
 259 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 260 {
 261         struct elevator_queue *e = q->elevator;
 262
 263         BUG_ON(ELV_ON_HASH(rq));
 264         hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
 265 }
 266
 267 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 268 {
 269         __elv_rqhash_del(rq);
 270         elv_rqhash_add(q, rq);
 271 }
 272
 273 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 274 {
 275         struct elevator_queue *e = q->elevator;
 276         struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
 277         struct hlist_node *entry, *next;
 278         struct request *rq;
 279
 280         hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
 281                 BUG_ON(!ELV_ON_HASH(rq));
 282
 283                 if (unlikely(!rq_mergeable(rq))) {
 284                         __elv_rqhash_del(rq);
 285                         continue;
 286                 }
 287
 288                 if (rq_hash_key(rq) == offset)
 289                         return rq;
 290         }
 291
 292         return NULL;
 293 }
 294
 295 /*
 296  * RB-tree support functions for inserting/lookup/removal of requests
 297  * in a sorted RB tree.
 298  */
 299 void elv_rb_add(struct rb_root *root, struct request *rq)
 300 {
 301         struct rb_node **p = &root->rb_node;
 302         struct rb_node *parent = NULL;
 303         struct request *__rq;
 304
 305         while (*p) {
 306                 parent = *p;
 307                 __rq = rb_entry(parent, struct request, rb_node);
 308
 309                 if (blk_rq_pos(rq) < blk_rq_pos(__rq))
 310                         p = &(*p)->rb_left;
 311                 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
 312                         p = &(*p)->rb_right;
 313         }
 314
 315         rb_link_node(&rq->rb_node, parent, p);
 316         rb_insert_color(&rq->rb_node, root);
 317 }
 318 EXPORT_SYMBOL(elv_rb_add);
 319
 320 void elv_rb_del(struct rb_root *root, struct request *rq)
 321 {
 322         BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 323         rb_erase(&rq->rb_node, root);
 324         RB_CLEAR_NODE(&rq->rb_node);
 325 }
 326 EXPORT_SYMBOL(elv_rb_del);
 327
 328 struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 329 {
 330         struct rb_node *n = root->rb_node;
 331         struct request *rq;
 332
 333         while (n) {
 334                 rq = rb_entry(n, struct request, rb_node);
 335
 336                 if (sector < blk_rq_pos(rq))
 337                         n = n->rb_left;
 338                 else if (sector > blk_rq_pos(rq))
 339                         n = n->rb_right;
 340                 else
 341                         return rq;
 342         }
 343
 344         return NULL;
 345 }
 346 EXPORT_SYMBOL(elv_rb_find);
 347
 348 /*
 349  * Insert rq into dispatch queue of q.  Queue lock must be held on
 350  * entry.  rq is sort instead into the dispatch queue. To be used by
 351  * specific elevators.
 352  */
 353 void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 354 {
 355         sector_t boundary;
 356         struct list_head *entry;
 357         int stop_flags;
 358
 359         if (q->last_merge == rq)
 360                 q->last_merge = NULL;
 361
 362         elv_rqhash_del(q, rq);
 363
 364         q->nr_sorted--;
 365
 366         boundary = q->end_sector;
 367         stop_flags = REQ_SOFTBARRIER | REQ_STARTED;
 368         list_for_each_prev(entry, &q->queue_head) {
 369                 struct request *pos = list_entry_rq(entry);
 370
 371                 if ((rq->cmd_flags & REQ_DISCARD) !=
 372                     (pos->cmd_flags & REQ_DISCARD))
 373                         break;
 374                 if (rq_data_dir(rq) != rq_data_dir(pos))
 375                         break;
 376                 if (pos->cmd_flags & stop_flags)
 377                         break;
 378                 if (blk_rq_pos(rq) >= boundary) {
 379                         if (blk_rq_pos(pos) < boundary)
 380                                 continue;
 381                 } else {
 382                         if (blk_rq_pos(pos) >= boundary)
 383                                 break;
 384                 }
 385                 if (blk_rq_pos(rq) >= blk_rq_pos(pos))
 386                         break;
 387         }
 388
 389         list_add(&rq->queuelist, entry);
 390 }
 391 EXPORT_SYMBOL(elv_dispatch_sort);
 392
 393 /*
 394  * Insert rq into dispatch queue of q.  Queue lock must be held on
 395  * entry.  rq is added to the back of the dispatch queue. To be used by
 396  * specific elevators.
 397  */
 398 void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 399 {
 400         if (q->last_merge == rq)
 401                 q->last_merge = NULL;
 402
 403         elv_rqhash_del(q, rq);
 404
 405         q->nr_sorted--;
 406
 407         q->end_sector = rq_end_sector(rq);
 408         q->boundary_rq = rq;
 409         list_add_tail(&rq->queuelist, &q->queue_head);
 410 }
 411 EXPORT_SYMBOL(elv_dispatch_add_tail);
 412
 413 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 414 {
 415         struct elevator_queue *e = q->elevator;
 416         struct request *__rq;
 417         int ret;
 418
 419         /*
 420          * Levels of merges:
 421          *      nomerges:  No merges at all attempted
 422          *      noxmerges: Only simple one-hit cache try
 423          *      merges:    All merge tries attempted
 424          */
 425         if (blk_queue_nomerges(q))
 426                 return ELEVATOR_NO_MERGE;
 427
 428         /*
 429          * First try one-hit cache.
 430          */
 431         if (q->last_merge && elv_rq_merge_ok(q->last_merge, bio)) {
 432                 ret = blk_try_merge(q->last_merge, bio);
 433                 if (ret != ELEVATOR_NO_MERGE) {
 434                         *req = q->last_merge;
 435                         return ret;
 436                 }
 437         }
 438
 439         if (blk_queue_noxmerges(q))
 440                 return ELEVATOR_NO_MERGE;
 441
 442         /*
 443          * See if our hash lookup can find a potential backmerge.
 444          */
 445         __rq = elv_rqhash_find(q, bio->bi_sector);
 446         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 447                 *req = __rq;
 448                 return ELEVATOR_BACK_MERGE;
 449         }
 450
 451         if (e->type->ops.elevator_merge_fn)
 452                 return e->type->ops.elevator_merge_fn(q, req, bio);
 453
 454         return ELEVATOR_NO_MERGE;
 455 }
 456
 457 /*
 458  * Attempt to do an insertion back merge. Only check for the case where
 459  * we can append 'rq' to an existing request, so we can throw 'rq' away
 460  * afterwards.
 461  *
 462  * Returns true if we merged, false otherwise
 463  */
 464 static bool elv_attempt_insert_merge(struct request_queue *q,
 465                                      struct request *rq)
 466 {
 467         struct request *__rq;
 468
 469         if (blk_queue_nomerges(q))
 470                 return false;
 471
 472         /*
 473          * First try one-hit cache.
 474          */
 475         if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
 476                 return true;
 477
 478         if (blk_queue_noxmerges(q))
 479                 return false;
 480
 481         /*
 482          * See if our hash lookup can find a potential backmerge.
 483          */
 484         __rq = elv_rqhash_find(q, blk_rq_pos(rq));
 485         if (__rq && blk_attempt_req_merge(q, __rq, rq))
 486                 return true;
 487
 488         return false;
 489 }
 490
 491 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 492 {
 493         struct elevator_queue *e = q->elevator;
 494
 495         if (e->type->ops.elevator_merged_fn)
 496                 e->type->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         struct elevator_queue *e = q->elevator;
 508         const int next_sorted = next->cmd_flags & REQ_SORTED;
 509
 510         if (next_sorted && e->type->ops.elevator_merge_req_fn)
 511                 e->type->ops.elevator_merge_req_fn(q, rq, next);
 512
 513         elv_rqhash_reposition(q, rq);
 514
 515         if (next_sorted) {
 516                 elv_rqhash_del(q, next);
 517                 q->nr_sorted--;
 518         }
 519
 520         q->last_merge = rq;
 521 }
 522
 523 void elv_bio_merged(struct request_queue *q, struct request *rq,
 524                         struct bio *bio)
 525 {
 526         struct elevator_queue *e = q->elevator;
 527
 528         if (e->type->ops.elevator_bio_merged_fn)
 529                 e->type->ops.elevator_bio_merged_fn(q, rq, bio);
 530 }
 531
 532 void elv_requeue_request(struct request_queue *q, struct request *rq)
 533 {
 534         /*
 535          * it already went through dequeue, we need to decrement the
 536          * in_flight count again
 537          */
 538         if (blk_account_rq(rq)) {
 539                 q->in_flight[rq_is_sync(rq)]--;
 540                 if (rq->cmd_flags & REQ_SORTED)
 541                         elv_deactivate_rq(q, rq);
 542         }
 543
 544         rq->cmd_flags &= ~REQ_STARTED;
 545
 546         __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
 547 }
 548
 549 void elv_drain_elevator(struct request_queue *q)
 550 {
 551         static int printed;
 552
 553         lockdep_assert_held(q->queue_lock);
 554
 555         while (q->elevator->type->ops.elevator_dispatch_fn(q, 1))
 556                 ;
 557         if (q->nr_sorted && printed++ < 10) {
 558                 printk(KERN_ERR "%s: forced dispatching is broken "
 559                        "(nr_sorted=%u), please report this\n",
 560                        q->elevator->type->elevator_name, q->nr_sorted);
 561         }
 562 }
 563
 564 void elv_quiesce_start(struct request_queue *q)
 565 {
 566         if (!q->elevator)
 567                 return;
 568
 569         spin_lock_irq(q->queue_lock);
 570         queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
 571         spin_unlock_irq(q->queue_lock);
 572
 573         blk_drain_queue(q, false);
 574 }
 575
 576 void elv_quiesce_end(struct request_queue *q)
 577 {
 578         spin_lock_irq(q->queue_lock);
 579         queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
 580         spin_unlock_irq(q->queue_lock);
 581 }
 582
 583 void __elv_add_request(struct request_queue *q, struct request *rq, int where)
 584 {
 585         trace_block_rq_insert(q, rq);
 586
 587         rq->q = q;
 588
 589         if (rq->cmd_flags & REQ_SOFTBARRIER) {
 590                 /* barriers are scheduling boundary, update end_sector */
 591                 if (rq->cmd_type == REQ_TYPE_FS ||
 592                     (rq->cmd_flags & REQ_DISCARD)) {
 593                         q->end_sector = rq_end_sector(rq);
 594                         q->boundary_rq = rq;
 595                 }
 596         } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 597                     (where == ELEVATOR_INSERT_SORT ||
 598                      where == ELEVATOR_INSERT_SORT_MERGE))
 599                 where = ELEVATOR_INSERT_BACK;
 600
 601         switch (where) {
 602         case ELEVATOR_INSERT_REQUEUE:
 603         case ELEVATOR_INSERT_FRONT:
 604                 rq->cmd_flags |= REQ_SOFTBARRIER;
 605                 list_add(&rq->queuelist, &q->queue_head);
 606                 break;
 607
 608         case ELEVATOR_INSERT_BACK:
 609                 rq->cmd_flags |= REQ_SOFTBARRIER;
 610                 elv_drain_elevator(q);
 611                 list_add_tail(&rq->queuelist, &q->queue_head);
 612                 /*
 613                  * We kick the queue here for the following reasons.
 614                  * - The elevator might have returned NULL previously
 615                  *   to delay requests and returned them now.  As the
 616                  *   queue wasn't empty before this request, ll_rw_blk
 617                  *   won't run the queue on return, resulting in hang.
 618                  * - Usually, back inserted requests won't be merged
 619                  *   with anything.  There's no point in delaying queue
 620                  *   processing.
 621                  */
 622                 __blk_run_queue(q);
 623                 break;
 624
 625         case ELEVATOR_INSERT_SORT_MERGE:
 626                 /*
 627                  * If we succeed in merging this request with one in the
 628                  * queue already, we are done - rq has now been freed,
 629                  * so no need to do anything further.
 630                  */
 631                 if (elv_attempt_insert_merge(q, rq))
 632                         break;
 633         case ELEVATOR_INSERT_SORT:
 634                 BUG_ON(rq->cmd_type != REQ_TYPE_FS &&
 635                        !(rq->cmd_flags & REQ_DISCARD));
 636                 rq->cmd_flags |= REQ_SORTED;
 637                 q->nr_sorted++;
 638                 if (rq_mergeable(rq)) {
 639                         elv_rqhash_add(q, rq);
 640                         if (!q->last_merge)
 641                                 q->last_merge = rq;
 642                 }
 643
 644                 /*
 645                  * Some ioscheds (cfq) run q->request_fn directly, so
 646                  * rq cannot be accessed after calling
 647                  * elevator_add_req_fn.
 648                  */
 649                 q->elevator->type->ops.elevator_add_req_fn(q, rq);
 650                 break;
 651
 652         case ELEVATOR_INSERT_FLUSH:
 653                 rq->cmd_flags |= REQ_SOFTBARRIER;
 654                 blk_insert_flush(rq);
 655                 break;
 656         default:
 657                 printk(KERN_ERR "%s: bad insertion point %d\n",
 658                        __func__, where);
 659                 BUG();
 660         }
 661 }
 662 EXPORT_SYMBOL(__elv_add_request);
 663
 664 void elv_add_request(struct request_queue *q, struct request *rq, int where)
 665 {
 666         unsigned long flags;
 667
 668         spin_lock_irqsave(q->queue_lock, flags);
 669         __elv_add_request(q, rq, where);
 670         spin_unlock_irqrestore(q->queue_lock, flags);
 671 }
 672 EXPORT_SYMBOL(elv_add_request);
 673
 674 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 675 {
 676         struct elevator_queue *e = q->elevator;
 677
 678         if (e->type->ops.elevator_latter_req_fn)
 679                 return e->type->ops.elevator_latter_req_fn(q, rq);
 680         return NULL;
 681 }
 682
 683 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 684 {
 685         struct elevator_queue *e = q->elevator;
 686
 687         if (e->type->ops.elevator_former_req_fn)
 688                 return e->type->ops.elevator_former_req_fn(q, rq);
 689         return NULL;
 690 }
 691
 692 int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
 693 {
 694         struct elevator_queue *e = q->elevator;
 695
 696         if (e->type->ops.elevator_set_req_fn)
 697                 return e->type->ops.elevator_set_req_fn(q, rq, gfp_mask);
 698         return 0;
 699 }
 700
 701 void elv_put_request(struct request_queue *q, struct request *rq)
 702 {
 703         struct elevator_queue *e = q->elevator;
 704
 705         if (e->type->ops.elevator_put_req_fn)
 706                 e->type->ops.elevator_put_req_fn(rq);
 707 }
 708
 709 int elv_may_queue(struct request_queue *q, int rw)
 710 {
 711         struct elevator_queue *e = q->elevator;
 712
 713         if (e->type->ops.elevator_may_queue_fn)
 714                 return e->type->ops.elevator_may_queue_fn(q, rw);
 715
 716         return ELV_MQUEUE_MAY;
 717 }
 718
 719 void elv_abort_queue(struct request_queue *q)
 720 {
 721         struct request *rq;
 722
 723         blk_abort_flushes(q);
 724
 725         while (!list_empty(&q->queue_head)) {
 726                 rq = list_entry_rq(q->queue_head.next);
 727                 rq->cmd_flags |= REQ_QUIET;
 728                 trace_block_rq_abort(q, rq);
 729                 /*
 730                  * Mark this request as started so we don't trigger
 731                  * any debug logic in the end I/O path.
 732                  */
 733                 blk_start_request(rq);
 734                 __blk_end_request_all(rq, -EIO);
 735         }
 736 }
 737 EXPORT_SYMBOL(elv_abort_queue);
 738
 739 void elv_completed_request(struct request_queue *q, struct request *rq)
 740 {
 741         struct elevator_queue *e = q->elevator;
 742
 743         /*
 744          * request is released from the driver, io must be done
 745          */
 746         if (blk_account_rq(rq)) {
 747                 q->in_flight[rq_is_sync(rq)]--;
 748                 if ((rq->cmd_flags & REQ_SORTED) &&
 749                     e->type->ops.elevator_completed_req_fn)
 750                         e->type->ops.elevator_completed_req_fn(q, rq);
 751         }
 752 }
 753
 754 #define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 755
 756 static ssize_t
 757 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 758 {
 759         struct elv_fs_entry *entry = to_elv(attr);
 760         struct elevator_queue *e;
 761         ssize_t error;
 762
 763         if (!entry->show)
 764                 return -EIO;
 765
 766         e = container_of(kobj, struct elevator_queue, kobj);
 767         mutex_lock(&e->sysfs_lock);
 768         error = e->type ? entry->show(e, page) : -ENOENT;
 769         mutex_unlock(&e->sysfs_lock);
 770         return error;
 771 }
 772
 773 static ssize_t
 774 elv_attr_store(struct kobject *kobj, struct attribute *attr,
 775                const char *page, size_t length)
 776 {
 777         struct elv_fs_entry *entry = to_elv(attr);
 778         struct elevator_queue *e;
 779         ssize_t error;
 780
 781         if (!entry->store)
 782                 return -EIO;
 783
 784         e = container_of(kobj, struct elevator_queue, kobj);
 785         mutex_lock(&e->sysfs_lock);
 786         error = e->type ? entry->store(e, page, length) : -ENOENT;
 787         mutex_unlock(&e->sysfs_lock);
 788         return error;
 789 }
 790
 791 static const struct sysfs_ops elv_sysfs_ops = {
 792         .show   = elv_attr_show,
 793         .store  = elv_attr_store,
 794 };
 795
 796 static struct kobj_type elv_ktype = {
 797         .sysfs_ops      = &elv_sysfs_ops,
 798         .release        = elevator_release,
 799 };
 800
 801 int elv_register_queue(struct request_queue *q)
 802 {
 803         struct elevator_queue *e = q->elevator;
 804         int error;
 805
 806         error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 807         if (!error) {
 808                 struct elv_fs_entry *attr = e->type->elevator_attrs;
 809                 if (attr) {
 810                         while (attr->attr.name) {
 811                                 if (sysfs_create_file(&e->kobj, &attr->attr))
 812                                         break;
 813                                 attr++;
 814                         }
 815                 }
 816                 kobject_uevent(&e->kobj, KOBJ_ADD);
 817                 e->registered = 1;
 818         }
 819         return error;
 820 }
 821 EXPORT_SYMBOL(elv_register_queue);
 822
 823 void elv_unregister_queue(struct request_queue *q)
 824 {
 825         if (q) {
 826                 struct elevator_queue *e = q->elevator;
 827
 828                 kobject_uevent(&e->kobj, KOBJ_REMOVE);
 829                 kobject_del(&e->kobj);
 830                 e->registered = 0;
 831         }
 832 }
 833 EXPORT_SYMBOL(elv_unregister_queue);
 834
 835 int elv_register(struct elevator_type *e)
 836 {
 837         char *def = "";
 838
 839         /* create icq_cache if requested */
 840         if (e->icq_size) {
 841                 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
 842                     WARN_ON(e->icq_align < __alignof__(struct io_cq)))
 843                         return -EINVAL;
 844
 845                 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
 846                          "%s_io_cq", e->elevator_name);
 847                 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
 848                                                  e->icq_align, 0, NULL);
 849                 if (!e->icq_cache)
 850                         return -ENOMEM;
 851         }
 852
 853         /* register, don't allow duplicate names */
 854         spin_lock(&elv_list_lock);
 855         if (elevator_find(e->elevator_name)) {
 856                 spin_unlock(&elv_list_lock);
 857                 if (e->icq_cache)
 858                         kmem_cache_destroy(e->icq_cache);
 859                 return -EBUSY;
 860         }
 861         list_add_tail(&e->list, &elv_list);
 862         spin_unlock(&elv_list_lock);
 863
 864         /* print pretty message */
 865         if (!strcmp(e->elevator_name, chosen_elevator) ||
 866                         (!*chosen_elevator &&
 867                          !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 868                                 def = " (default)";
 869
 870         printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
 871                                                                 def);
 872         return 0;
 873 }
 874 EXPORT_SYMBOL_GPL(elv_register);
 875
 876 void elv_unregister(struct elevator_type *e)
 877 {
 878         /* unregister */
 879         spin_lock(&elv_list_lock);
 880         list_del_init(&e->list);
 881         spin_unlock(&elv_list_lock);
 882
 883         /*
 884          * Destroy icq_cache if it exists.  icq's are RCU managed.  Make
 885          * sure all RCU operations are complete before proceeding.
 886          */
 887         if (e->icq_cache) {
 888                 rcu_barrier();
 889                 kmem_cache_destroy(e->icq_cache);
 890                 e->icq_cache = NULL;
 891         }
 892 }
 893 EXPORT_SYMBOL_GPL(elv_unregister);
 894
 895 /*
 896  * switch to new_e io scheduler. be careful not to introduce deadlocks -
 897  * we don't free the old io scheduler, before we have allocated what we
 898  * need for the new one. this way we have a chance of going back to the old
 899  * one, if the new one fails init for some reason.
 900  */
 901 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
 902 {
 903         struct elevator_queue *old = q->elevator;
 904         bool registered = old->registered;
 905         int err;
 906
 907         /*
 908          * Turn on BYPASS and drain all requests w/ elevator private data.
 909          * Block layer doesn't call into a quiesced elevator - all requests
 910          * are directly put on the dispatch list without elevator data
 911          * using INSERT_BACK.  All requests have SOFTBARRIER set and no
 912          * merge happens either.
 913          */
 914         elv_quiesce_start(q);
 915
 916         /* unregister and clear all auxiliary data of the old elevator */
 917         if (registered)
 918                 elv_unregister_queue(q);
 919
 920         spin_lock_irq(q->queue_lock);
 921         ioc_clear_queue(q);
 922         spin_unlock_irq(q->queue_lock);
 923
 924         /* allocate, init and register new elevator */
 925         err = -ENOMEM;
 926         q->elevator = elevator_alloc(q, new_e);
 927         if (!q->elevator)
 928                 goto fail_init;
 929
 930         err = elevator_init_queue(q);
 931         if (err) {
 932                 kobject_put(&q->elevator->kobj);
 933                 goto fail_init;
 934         }
 935
 936         if (registered) {
 937                 err = elv_register_queue(q);
 938                 if (err)
 939                         goto fail_register;
 940         }
 941
 942         /* done, kill the old one and finish */
 943         elevator_exit(old);
 944         elv_quiesce_end(q);
 945
 946         blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
 947
 948         return 0;
 949
 950 fail_register:
 951         elevator_exit(q->elevator);
 952 fail_init:
 953         /* switch failed, restore and re-register old elevator */
 954         q->elevator = old;
 955         elv_register_queue(q);
 956         elv_quiesce_end(q);
 957
 958         return err;
 959 }
 960
 961 /*
 962  * Switch this queue to the given IO scheduler.
 963  */
 964 int elevator_change(struct request_queue *q, const char *name)
 965 {
 966         char elevator_name[ELV_NAME_MAX];
 967         struct elevator_type *e;
 968
 969         if (!q->elevator)
 970                 return -ENXIO;
 971
 972         strlcpy(elevator_name, name, sizeof(elevator_name));
 973         e = elevator_get(strstrip(elevator_name));
 974         if (!e) {
 975                 printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
 976                 return -EINVAL;
 977         }
 978
 979         if (!strcmp(elevator_name, q->elevator->type->elevator_name)) {
 980                 elevator_put(e);
 981                 return 0;
 982         }
 983
 984         return elevator_switch(q, e);
 985 }
 986 EXPORT_SYMBOL(elevator_change);
 987
 988 ssize_t elv_iosched_store(struct request_queue *q, const char *name,
 989                           size_t count)
 990 {
 991         int ret;
 992
 993         if (!q->elevator)
 994                 return count;
 995
 996         ret = elevator_change(q, name);
 997         if (!ret)
 998                 return count;
 999
1000         printk(KERN_ERR "elevator: switch to %s failed\n", name);
1001         return ret;
1002 }
1003
1004 ssize_t elv_iosched_show(struct request_queue *q, char *name)
1005 {
1006         struct elevator_queue *e = q->elevator;
1007         struct elevator_type *elv;
1008         struct elevator_type *__e;
1009         int len = 0;
1010
1011         if (!q->elevator || !blk_queue_stackable(q))
1012                 return sprintf(name, "none\n");
1013
1014         elv = e->type;
1015
1016         spin_lock(&elv_list_lock);
1017         list_for_each_entry(__e, &elv_list, list) {
1018                 if (!strcmp(elv->elevator_name, __e->elevator_name))
1019                         len += sprintf(name+len, "[%s] ", elv->elevator_name);
1020                 else
1021                         len += sprintf(name+len, "%s ", __e->elevator_name);
1022         }
1023         spin_unlock(&elv_list_lock);
1024
1025         len += sprintf(len+name, "\n");
1026         return len;
1027 }
1028
1029 struct request *elv_rb_former_request(struct request_queue *q,
1030                                       struct request *rq)
1031 {
1032         struct rb_node *rbprev = rb_prev(&rq->rb_node);
1033
1034         if (rbprev)
1035                 return rb_entry_rq(rbprev);
1036
1037         return NULL;
1038 }
1039 EXPORT_SYMBOL(elv_rb_former_request);
1040
1041 struct request *elv_rb_latter_request(struct request_queue *q,
1042                                       struct request *rq)
1043 {
1044         struct rb_node *rbnext = rb_next(&rq->rb_node);
1045
1046         if (rbnext)
1047                 return rb_entry_rq(rbnext);
1048
1049         return NULL;
1050 }
1051 EXPORT_SYMBOL(elv_rb_latter_request);