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