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/blktrace_api.h>
  35#include <linux/hash.h>
  36#include <linux/uaccess.h>
  37#include <linux/pm_runtime.h>
  38#include <linux/blk-cgroup.h>
  39
  40#include <trace/events/block.h>
  41
  42#include "blk.h"
  43
  44static DEFINE_SPINLOCK(elv_list_lock);
  45static LIST_HEAD(elv_list);
  46
  47/*
  48 * Merge hash stuff.
  49 */
  50#define rq_hash_key(rq)         (blk_rq_pos(rq) + blk_rq_sectors(rq))
  51
  52/*
  53 * Query io scheduler to see if the current process issuing bio may be
  54 * merged with rq.
  55 */
  56static int elv_iosched_allow_bio_merge(struct request *rq, struct bio *bio)
  57{
  58        struct request_queue *q = rq->q;
  59        struct elevator_queue *e = q->elevator;
  60
  61        if (e->type->ops.elevator_allow_bio_merge_fn)
  62                return e->type->ops.elevator_allow_bio_merge_fn(q, rq, bio);
  63
  64        return 1;
  65}
  66
  67/*
  68 * can we safely merge with this request?
  69 */
  70bool elv_bio_merge_ok(struct request *rq, struct bio *bio)
  71{
  72        if (!blk_rq_merge_ok(rq, bio))
  73                return false;
  74
  75        if (!elv_iosched_allow_bio_merge(rq, bio))
  76                return false;
  77
  78        return true;
  79}
  80EXPORT_SYMBOL(elv_bio_merge_ok);
  81
  82static struct elevator_type *elevator_find(const char *name)
  83{
  84        struct elevator_type *e;
  85
  86        list_for_each_entry(e, &elv_list, list) {
  87                if (!strcmp(e->elevator_name, name))
  88                        return e;
  89        }
  90
  91        return NULL;
  92}
  93
  94static void elevator_put(struct elevator_type *e)
  95{
  96        module_put(e->elevator_owner);
  97}
  98
  99static struct elevator_type *elevator_get(const char *name, bool try_loading)
 100{
 101        struct elevator_type *e;
 102
 103        spin_lock(&elv_list_lock);
 104
 105        e = elevator_find(name);
 106        if (!e && try_loading) {
 107                spin_unlock(&elv_list_lock);
 108                request_module("%s-iosched", name);
 109                spin_lock(&elv_list_lock);
 110                e = elevator_find(name);
 111        }
 112
 113        if (e && !try_module_get(e->elevator_owner))
 114                e = NULL;
 115
 116        spin_unlock(&elv_list_lock);
 117
 118        return e;
 119}
 120
 121static char chosen_elevator[ELV_NAME_MAX];
 122
 123static int __init elevator_setup(char *str)
 124{
 125        /*
 126         * Be backwards-compatible with previous kernels, so users
 127         * won't get the wrong elevator.
 128         */
 129        strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
 130        return 1;
 131}
 132
 133__setup("elevator=", elevator_setup);
 134
 135/* called during boot to load the elevator chosen by the elevator param */
 136void __init load_default_elevator_module(void)
 137{
 138        struct elevator_type *e;
 139
 140        if (!chosen_elevator[0])
 141                return;
 142
 143        spin_lock(&elv_list_lock);
 144        e = elevator_find(chosen_elevator);
 145        spin_unlock(&elv_list_lock);
 146
 147        if (!e)
 148                request_module("%s-iosched", chosen_elevator);
 149}
 150
 151static struct kobj_type elv_ktype;
 152
 153struct elevator_queue *elevator_alloc(struct request_queue *q,
 154                                  struct elevator_type *e)
 155{
 156        struct elevator_queue *eq;
 157
 158        eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node);
 159        if (unlikely(!eq))
 160                return NULL;
 161
 162        eq->type = e;
 163        kobject_init(&eq->kobj, &elv_ktype);
 164        mutex_init(&eq->sysfs_lock);
 165        hash_init(eq->hash);
 166
 167        return eq;
 168}
 169EXPORT_SYMBOL(elevator_alloc);
 170
 171static void elevator_release(struct kobject *kobj)
 172{
 173        struct elevator_queue *e;
 174
 175        e = container_of(kobj, struct elevator_queue, kobj);
 176        elevator_put(e->type);
 177        kfree(e);
 178}
 179
 180int elevator_init(struct request_queue *q, char *name)
 181{
 182        struct elevator_type *e = NULL;
 183        int err;
 184
 185        /*
 186         * q->sysfs_lock must be held to provide mutual exclusion between
 187         * elevator_switch() and here.
 188         */
 189        lockdep_assert_held(&q->sysfs_lock);
 190
 191        if (unlikely(q->elevator))
 192                return 0;
 193
 194        INIT_LIST_HEAD(&q->queue_head);
 195        q->last_merge = NULL;
 196        q->end_sector = 0;
 197        q->boundary_rq = NULL;
 198
 199        if (name) {
 200                e = elevator_get(name, true);
 201                if (!e)
 202                        return -EINVAL;
 203        }
 204
 205        /*
 206         * Use the default elevator specified by config boot param or
 207         * config option.  Don't try to load modules as we could be running
 208         * off async and request_module() isn't allowed from async.
 209         */
 210        if (!e && *chosen_elevator) {
 211                e = elevator_get(chosen_elevator, false);
 212                if (!e)
 213                        printk(KERN_ERR "I/O scheduler %s not found\n",
 214                                                        chosen_elevator);
 215        }
 216
 217        if (!e) {
 218                e = elevator_get(CONFIG_DEFAULT_IOSCHED, false);
 219                if (!e) {
 220                        printk(KERN_ERR
 221                                "Default I/O scheduler not found. " \
 222                                "Using noop.\n");
 223                        e = elevator_get("noop", false);
 224                }
 225        }
 226
 227        err = e->ops.elevator_init_fn(q, e);
 228        if (err)
 229                elevator_put(e);
 230        return err;
 231}
 232EXPORT_SYMBOL(elevator_init);
 233
 234void elevator_exit(struct elevator_queue *e)
 235{
 236        mutex_lock(&e->sysfs_lock);
 237        if (e->type->ops.elevator_exit_fn)
 238                e->type->ops.elevator_exit_fn(e);
 239        mutex_unlock(&e->sysfs_lock);
 240
 241        kobject_put(&e->kobj);
 242}
 243EXPORT_SYMBOL(elevator_exit);
 244
 245static inline void __elv_rqhash_del(struct request *rq)
 246{
 247        hash_del(&rq->hash);
 248        rq->cmd_flags &= ~REQ_HASHED;
 249}
 250
 251static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 252{
 253        if (ELV_ON_HASH(rq))
 254                __elv_rqhash_del(rq);
 255}
 256
 257static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 258{
 259        struct elevator_queue *e = q->elevator;
 260
 261        BUG_ON(ELV_ON_HASH(rq));
 262        hash_add(e->hash, &rq->hash, rq_hash_key(rq));
 263        rq->cmd_flags |= REQ_HASHED;
 264}
 265
 266static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 267{
 268        __elv_rqhash_del(rq);
 269        elv_rqhash_add(q, rq);
 270}
 271
 272static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 273{
 274        struct elevator_queue *e = q->elevator;
 275        struct hlist_node *next;
 276        struct request *rq;
 277
 278        hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
 279                BUG_ON(!ELV_ON_HASH(rq));
 280
 281                if (unlikely(!rq_mergeable(rq))) {
 282                        __elv_rqhash_del(rq);
 283                        continue;
 284                }
 285
 286                if (rq_hash_key(rq) == offset)
 287                        return rq;
 288        }
 289
 290        return NULL;
 291}
 292
 293/*
 294 * RB-tree support functions for inserting/lookup/removal of requests
 295 * in a sorted RB tree.
 296 */
 297void elv_rb_add(struct rb_root *root, struct request *rq)
 298{
 299        struct rb_node **p = &root->rb_node;
 300        struct rb_node *parent = NULL;
 301        struct request *__rq;
 302
 303        while (*p) {
 304                parent = *p;
 305                __rq = rb_entry(parent, struct request, rb_node);
 306
 307                if (blk_rq_pos(rq) < blk_rq_pos(__rq))
 308                        p = &(*p)->rb_left;
 309                else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
 310                        p = &(*p)->rb_right;
 311        }
 312
 313        rb_link_node(&rq->rb_node, parent, p);
 314        rb_insert_color(&rq->rb_node, root);
 315}
 316EXPORT_SYMBOL(elv_rb_add);
 317
 318void elv_rb_del(struct rb_root *root, struct request *rq)
 319{
 320        BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
 321        rb_erase(&rq->rb_node, root);
 322        RB_CLEAR_NODE(&rq->rb_node);
 323}
 324EXPORT_SYMBOL(elv_rb_del);
 325
 326struct request *elv_rb_find(struct rb_root *root, sector_t sector)
 327{
 328        struct rb_node *n = root->rb_node;
 329        struct request *rq;
 330
 331        while (n) {
 332                rq = rb_entry(n, struct request, rb_node);
 333
 334                if (sector < blk_rq_pos(rq))
 335                        n = n->rb_left;
 336                else if (sector > blk_rq_pos(rq))
 337                        n = n->rb_right;
 338                else
 339                        return rq;
 340        }
 341
 342        return NULL;
 343}
 344EXPORT_SYMBOL(elv_rb_find);
 345
 346/*
 347 * Insert rq into dispatch queue of q.  Queue lock must be held on
 348 * entry.  rq is sort instead into the dispatch queue. To be used by
 349 * specific elevators.
 350 */
 351void elv_dispatch_sort(struct request_queue *q, struct request *rq)
 352{
 353        sector_t boundary;
 354        struct list_head *entry;
 355        int stop_flags;
 356
 357        if (q->last_merge == rq)
 358                q->last_merge = NULL;
 359
 360        elv_rqhash_del(q, rq);
 361
 362        q->nr_sorted--;
 363
 364        boundary = q->end_sector;
 365        stop_flags = REQ_SOFTBARRIER | REQ_STARTED;
 366        list_for_each_prev(entry, &q->queue_head) {
 367                struct request *pos = list_entry_rq(entry);
 368
 369                if (req_op(rq) != req_op(pos))
 370                        break;
 371                if (rq_data_dir(rq) != rq_data_dir(pos))
 372                        break;
 373                if (pos->cmd_flags & stop_flags)
 374                        break;
 375                if (blk_rq_pos(rq) >= boundary) {
 376                        if (blk_rq_pos(pos) < boundary)
 377                                continue;
 378                } else {
 379                        if (blk_rq_pos(pos) >= boundary)
 380                                break;
 381                }
 382                if (blk_rq_pos(rq) >= blk_rq_pos(pos))
 383                        break;
 384        }
 385
 386        list_add(&rq->queuelist, entry);
 387}
 388EXPORT_SYMBOL(elv_dispatch_sort);
 389
 390/*
 391 * Insert rq into dispatch queue of q.  Queue lock must be held on
 392 * entry.  rq is added to the back of the dispatch queue. To be used by
 393 * specific elevators.
 394 */
 395void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
 396{
 397        if (q->last_merge == rq)
 398                q->last_merge = NULL;
 399
 400        elv_rqhash_del(q, rq);
 401
 402        q->nr_sorted--;
 403
 404        q->end_sector = rq_end_sector(rq);
 405        q->boundary_rq = rq;
 406        list_add_tail(&rq->queuelist, &q->queue_head);
 407}
 408EXPORT_SYMBOL(elv_dispatch_add_tail);
 409
 410int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 411{
 412        struct elevator_queue *e = q->elevator;
 413        struct request *__rq;
 414        int ret;
 415
 416        /*
 417         * Levels of merges:
 418         *      nomerges:  No merges at all attempted
 419         *      noxmerges: Only simple one-hit cache try
 420         *      merges:    All merge tries attempted
 421         */
 422        if (blk_queue_nomerges(q) || !bio_mergeable(bio))
 423                return ELEVATOR_NO_MERGE;
 424
 425        /*
 426         * First try one-hit cache.
 427         */
 428        if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) {
 429                ret = blk_try_merge(q->last_merge, bio);
 430                if (ret != ELEVATOR_NO_MERGE) {
 431                        *req = q->last_merge;
 432                        return ret;
 433                }
 434        }
 435
 436        if (blk_queue_noxmerges(q))
 437                return ELEVATOR_NO_MERGE;
 438
 439        /*
 440         * See if our hash lookup can find a potential backmerge.
 441         */
 442        __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);
 443        if (__rq && elv_bio_merge_ok(__rq, bio)) {
 444                *req = __rq;
 445                return ELEVATOR_BACK_MERGE;
 446        }
 447
 448        if (e->type->ops.elevator_merge_fn)
 449                return e->type->ops.elevator_merge_fn(q, req, bio);
 450
 451        return ELEVATOR_NO_MERGE;
 452}
 453
 454/*
 455 * Attempt to do an insertion back merge. Only check for the case where
 456 * we can append 'rq' to an existing request, so we can throw 'rq' away
 457 * afterwards.
 458 *
 459 * Returns true if we merged, false otherwise
 460 */
 461static bool elv_attempt_insert_merge(struct request_queue *q,
 462                                     struct request *rq)
 463{
 464        struct request *__rq;
 465        bool ret;
 466
 467        if (blk_queue_nomerges(q))
 468                return false;
 469
 470        /*
 471         * First try one-hit cache.
 472         */
 473        if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
 474                return true;
 475
 476        if (blk_queue_noxmerges(q))
 477                return false;
 478
 479        ret = false;
 480        /*
 481         * See if our hash lookup can find a potential backmerge.
 482         */
 483        while (1) {
 484                __rq = elv_rqhash_find(q, blk_rq_pos(rq));
 485                if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
 486                        break;
 487
 488                /* The merged request could be merged with others, try again */
 489                ret = true;
 490                rq = __rq;
 491        }
 492
 493        return ret;
 494}
 495
 496void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 497{
 498        struct elevator_queue *e = q->elevator;
 499
 500        if (e->type->ops.elevator_merged_fn)
 501                e->type->ops.elevator_merged_fn(q, rq, type);
 502
 503        if (type == ELEVATOR_BACK_MERGE)
 504                elv_rqhash_reposition(q, rq);
 505
 506        q->last_merge = rq;
 507}
 508
 509void elv_merge_requests(struct request_queue *q, struct request *rq,
 510                             struct request *next)
 511{
 512        struct elevator_queue *e = q->elevator;
 513        const int next_sorted = next->cmd_flags & REQ_SORTED;
 514
 515        if (next_sorted && e->type->ops.elevator_merge_req_fn)
 516                e->type->ops.elevator_merge_req_fn(q, rq, next);
 517
 518        elv_rqhash_reposition(q, rq);
 519
 520        if (next_sorted) {
 521                elv_rqhash_del(q, next);
 522                q->nr_sorted--;
 523        }
 524
 525        q->last_merge = rq;
 526}
 527
 528void elv_bio_merged(struct request_queue *q, struct request *rq,
 529                        struct bio *bio)
 530{
 531        struct elevator_queue *e = q->elevator;
 532
 533        if (e->type->ops.elevator_bio_merged_fn)
 534                e->type->ops.elevator_bio_merged_fn(q, rq, bio);
 535}
 536
 537#ifdef CONFIG_PM
 538static void blk_pm_requeue_request(struct request *rq)
 539{
 540        if (rq->q->dev && !(rq->cmd_flags & REQ_PM))
 541                rq->q->nr_pending--;
 542}
 543
 544static void blk_pm_add_request(struct request_queue *q, struct request *rq)
 545{
 546        if (q->dev && !(rq->cmd_flags & REQ_PM) && q->nr_pending++ == 0 &&
 547            (q->rpm_status == RPM_SUSPENDED || q->rpm_status == RPM_SUSPENDING))
 548                pm_request_resume(q->dev);
 549}
 550#else
 551static inline void blk_pm_requeue_request(struct request *rq) {}
 552static inline void blk_pm_add_request(struct request_queue *q,
 553                                      struct request *rq)
 554{
 555}
 556#endif
 557
 558void elv_requeue_request(struct request_queue *q, struct request *rq)
 559{
 560        /*
 561         * it already went through dequeue, we need to decrement the
 562         * in_flight count again
 563         */
 564        if (blk_account_rq(rq)) {
 565                q->in_flight[rq_is_sync(rq)]--;
 566                if (rq->cmd_flags & REQ_SORTED)
 567                        elv_deactivate_rq(q, rq);
 568        }
 569
 570        rq->cmd_flags &= ~REQ_STARTED;
 571
 572        blk_pm_requeue_request(rq);
 573
 574        __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
 575}
 576
 577void elv_drain_elevator(struct request_queue *q)
 578{
 579        static int printed;
 580
 581        lockdep_assert_held(q->queue_lock);
 582
 583        while (q->elevator->type->ops.elevator_dispatch_fn(q, 1))
 584                ;
 585        if (q->nr_sorted && printed++ < 10) {
 586                printk(KERN_ERR "%s: forced dispatching is broken "
 587                       "(nr_sorted=%u), please report this\n",
 588                       q->elevator->type->elevator_name, q->nr_sorted);
 589        }
 590}
 591
 592void __elv_add_request(struct request_queue *q, struct request *rq, int where)
 593{
 594        trace_block_rq_insert(q, rq);
 595
 596        blk_pm_add_request(q, rq);
 597
 598        rq->q = q;
 599
 600        if (rq->cmd_flags & REQ_SOFTBARRIER) {
 601                /* barriers are scheduling boundary, update end_sector */
 602                if (rq->cmd_type == REQ_TYPE_FS) {
 603                        q->end_sector = rq_end_sector(rq);
 604                        q->boundary_rq = rq;
 605                }
 606        } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
 607                    (where == ELEVATOR_INSERT_SORT ||
 608                     where == ELEVATOR_INSERT_SORT_MERGE))
 609                where = ELEVATOR_INSERT_BACK;
 610
 611        switch (where) {
 612        case ELEVATOR_INSERT_REQUEUE:
 613        case ELEVATOR_INSERT_FRONT:
 614                rq->cmd_flags |= REQ_SOFTBARRIER;
 615                list_add(&rq->queuelist, &q->queue_head);
 616                break;
 617
 618        case ELEVATOR_INSERT_BACK:
 619                rq->cmd_flags |= REQ_SOFTBARRIER;
 620                elv_drain_elevator(q);
 621                list_add_tail(&rq->queuelist, &q->queue_head);
 622                /*
 623                 * We kick the queue here for the following reasons.
 624                 * - The elevator might have returned NULL previously
 625                 *   to delay requests and returned them now.  As the
 626                 *   queue wasn't empty before this request, ll_rw_blk
 627                 *   won't run the queue on return, resulting in hang.
 628                 * - Usually, back inserted requests won't be merged
 629                 *   with anything.  There's no point in delaying queue
 630                 *   processing.
 631                 */
 632                __blk_run_queue(q);
 633                break;
 634
 635        case ELEVATOR_INSERT_SORT_MERGE:
 636                /*
 637                 * If we succeed in merging this request with one in the
 638                 * queue already, we are done - rq has now been freed,
 639                 * so no need to do anything further.
 640                 */
 641                if (elv_attempt_insert_merge(q, rq))
 642                        break;
 643        case ELEVATOR_INSERT_SORT:
 644                BUG_ON(rq->cmd_type != REQ_TYPE_FS);
 645                rq->cmd_flags |= REQ_SORTED;
 646                q->nr_sorted++;
 647                if (rq_mergeable(rq)) {
 648                        elv_rqhash_add(q, rq);
 649                        if (!q->last_merge)
 650                                q->last_merge = rq;
 651                }
 652
 653                /*
 654                 * Some ioscheds (cfq) run q->request_fn directly, so
 655                 * rq cannot be accessed after calling
 656                 * elevator_add_req_fn.
 657                 */
 658                q->elevator->type->ops.elevator_add_req_fn(q, rq);
 659                break;
 660
 661        case ELEVATOR_INSERT_FLUSH:
 662                rq->cmd_flags |= REQ_SOFTBARRIER;
 663                blk_insert_flush(rq);
 664                break;
 665        default:
 666                printk(KERN_ERR "%s: bad insertion point %d\n",
 667                       __func__, where);
 668                BUG();
 669        }
 670}
 671EXPORT_SYMBOL(__elv_add_request);
 672
 673void elv_add_request(struct request_queue *q, struct request *rq, int where)
 674{
 675        unsigned long flags;
 676
 677        spin_lock_irqsave(q->queue_lock, flags);
 678        __elv_add_request(q, rq, where);
 679        spin_unlock_irqrestore(q->queue_lock, flags);
 680}
 681EXPORT_SYMBOL(elv_add_request);
 682
 683struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 684{
 685        struct elevator_queue *e = q->elevator;
 686
 687        if (e->type->ops.elevator_latter_req_fn)
 688                return e->type->ops.elevator_latter_req_fn(q, rq);
 689        return NULL;
 690}
 691
 692struct request *elv_former_request(struct request_queue *q, struct request *rq)
 693{
 694        struct elevator_queue *e = q->elevator;
 695
 696        if (e->type->ops.elevator_former_req_fn)
 697                return e->type->ops.elevator_former_req_fn(q, rq);
 698        return NULL;
 699}
 700
 701int elv_set_request(struct request_queue *q, struct request *rq,
 702                    struct bio *bio, gfp_t gfp_mask)
 703{
 704        struct elevator_queue *e = q->elevator;
 705
 706        if (e->type->ops.elevator_set_req_fn)
 707                return e->type->ops.elevator_set_req_fn(q, rq, bio, gfp_mask);
 708        return 0;
 709}
 710
 711void elv_put_request(struct request_queue *q, struct request *rq)
 712{
 713        struct elevator_queue *e = q->elevator;
 714
 715        if (e->type->ops.elevator_put_req_fn)
 716                e->type->ops.elevator_put_req_fn(rq);
 717}
 718
 719int elv_may_queue(struct request_queue *q, int op, int op_flags)
 720{
 721        struct elevator_queue *e = q->elevator;
 722
 723        if (e->type->ops.elevator_may_queue_fn)
 724                return e->type->ops.elevator_may_queue_fn(q, op, op_flags);
 725
 726        return ELV_MQUEUE_MAY;
 727}
 728
 729void elv_completed_request(struct request_queue *q, struct request *rq)
 730{
 731        struct elevator_queue *e = q->elevator;
 732
 733        /*
 734         * request is released from the driver, io must be done
 735         */
 736        if (blk_account_rq(rq)) {
 737                q->in_flight[rq_is_sync(rq)]--;
 738                if ((rq->cmd_flags & REQ_SORTED) &&
 739                    e->type->ops.elevator_completed_req_fn)
 740                        e->type->ops.elevator_completed_req_fn(q, rq);
 741        }
 742}
 743
 744#define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
 745
 746static ssize_t
 747elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 748{
 749        struct elv_fs_entry *entry = to_elv(attr);
 750        struct elevator_queue *e;
 751        ssize_t error;
 752
 753        if (!entry->show)
 754                return -EIO;
 755
 756        e = container_of(kobj, struct elevator_queue, kobj);
 757        mutex_lock(&e->sysfs_lock);
 758        error = e->type ? entry->show(e, page) : -ENOENT;
 759        mutex_unlock(&e->sysfs_lock);
 760        return error;
 761}
 762
 763static ssize_t
 764elv_attr_store(struct kobject *kobj, struct attribute *attr,
 765               const char *page, size_t length)
 766{
 767        struct elv_fs_entry *entry = to_elv(attr);
 768        struct elevator_queue *e;
 769        ssize_t error;
 770
 771        if (!entry->store)
 772                return -EIO;
 773
 774        e = container_of(kobj, struct elevator_queue, kobj);
 775        mutex_lock(&e->sysfs_lock);
 776        error = e->type ? entry->store(e, page, length) : -ENOENT;
 777        mutex_unlock(&e->sysfs_lock);
 778        return error;
 779}
 780
 781static const struct sysfs_ops elv_sysfs_ops = {
 782        .show   = elv_attr_show,
 783        .store  = elv_attr_store,
 784};
 785
 786static struct kobj_type elv_ktype = {
 787        .sysfs_ops      = &elv_sysfs_ops,
 788        .release        = elevator_release,
 789};
 790
 791int elv_register_queue(struct request_queue *q)
 792{
 793        struct elevator_queue *e = q->elevator;
 794        int error;
 795
 796        error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
 797        if (!error) {
 798                struct elv_fs_entry *attr = e->type->elevator_attrs;
 799                if (attr) {
 800                        while (attr->attr.name) {
 801                                if (sysfs_create_file(&e->kobj, &attr->attr))
 802                                        break;
 803                                attr++;
 804                        }
 805                }
 806                kobject_uevent(&e->kobj, KOBJ_ADD);
 807                e->registered = 1;
 808                if (e->type->ops.elevator_registered_fn)
 809                        e->type->ops.elevator_registered_fn(q);
 810        }
 811        return error;
 812}
 813EXPORT_SYMBOL(elv_register_queue);
 814
 815void elv_unregister_queue(struct request_queue *q)
 816{
 817        if (q) {
 818                struct elevator_queue *e = q->elevator;
 819
 820                kobject_uevent(&e->kobj, KOBJ_REMOVE);
 821                kobject_del(&e->kobj);
 822                e->registered = 0;
 823        }
 824}
 825EXPORT_SYMBOL(elv_unregister_queue);
 826
 827int elv_register(struct elevator_type *e)
 828{
 829        char *def = "";
 830
 831        /* create icq_cache if requested */
 832        if (e->icq_size) {
 833                if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
 834                    WARN_ON(e->icq_align < __alignof__(struct io_cq)))
 835                        return -EINVAL;
 836
 837                snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
 838                         "%s_io_cq", e->elevator_name);
 839                e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
 840                                                 e->icq_align, 0, NULL);
 841                if (!e->icq_cache)
 842                        return -ENOMEM;
 843        }
 844
 845        /* register, don't allow duplicate names */
 846        spin_lock(&elv_list_lock);
 847        if (elevator_find(e->elevator_name)) {
 848                spin_unlock(&elv_list_lock);
 849                if (e->icq_cache)
 850                        kmem_cache_destroy(e->icq_cache);
 851                return -EBUSY;
 852        }
 853        list_add_tail(&e->list, &elv_list);
 854        spin_unlock(&elv_list_lock);
 855
 856        /* print pretty message */
 857        if (!strcmp(e->elevator_name, chosen_elevator) ||
 858                        (!*chosen_elevator &&
 859                         !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
 860                                def = " (default)";
 861
 862        printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
 863                                                                def);
 864        return 0;
 865}
 866EXPORT_SYMBOL_GPL(elv_register);
 867
 868void elv_unregister(struct elevator_type *e)
 869{
 870        /* unregister */
 871        spin_lock(&elv_list_lock);
 872        list_del_init(&e->list);
 873        spin_unlock(&elv_list_lock);
 874
 875        /*
 876         * Destroy icq_cache if it exists.  icq's are RCU managed.  Make
 877         * sure all RCU operations are complete before proceeding.
 878         */
 879        if (e->icq_cache) {
 880                rcu_barrier();
 881                kmem_cache_destroy(e->icq_cache);
 882                e->icq_cache = NULL;
 883        }
 884}
 885EXPORT_SYMBOL_GPL(elv_unregister);
 886
 887/*
 888 * switch to new_e io scheduler. be careful not to introduce deadlocks -
 889 * we don't free the old io scheduler, before we have allocated what we
 890 * need for the new one. this way we have a chance of going back to the old
 891 * one, if the new one fails init for some reason.
 892 */
 893static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
 894{
 895        struct elevator_queue *old = q->elevator;
 896        bool registered = old->registered;
 897        int err;
 898
 899        /*
 900         * Turn on BYPASS and drain all requests w/ elevator private data.
 901         * Block layer doesn't call into a quiesced elevator - all requests
 902         * are directly put on the dispatch list without elevator data
 903         * using INSERT_BACK.  All requests have SOFTBARRIER set and no
 904         * merge happens either.
 905         */
 906        blk_queue_bypass_start(q);
 907
 908        /* unregister and clear all auxiliary data of the old elevator */
 909        if (registered)
 910                elv_unregister_queue(q);
 911
 912        spin_lock_irq(q->queue_lock);
 913        ioc_clear_queue(q);
 914        spin_unlock_irq(q->queue_lock);
 915
 916        /* allocate, init and register new elevator */
 917        err = new_e->ops.elevator_init_fn(q, new_e);
 918        if (err)
 919                goto fail_init;
 920
 921        if (registered) {
 922                err = elv_register_queue(q);
 923                if (err)
 924                        goto fail_register;
 925        }
 926
 927        /* done, kill the old one and finish */
 928        elevator_exit(old);
 929        blk_queue_bypass_end(q);
 930
 931        blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
 932
 933        return 0;
 934
 935fail_register:
 936        elevator_exit(q->elevator);
 937fail_init:
 938        /* switch failed, restore and re-register old elevator */
 939        q->elevator = old;
 940        elv_register_queue(q);
 941        blk_queue_bypass_end(q);
 942
 943        return err;
 944}
 945
 946/*
 947 * Switch this queue to the given IO scheduler.
 948 */
 949static int __elevator_change(struct request_queue *q, const char *name)
 950{
 951        char elevator_name[ELV_NAME_MAX];
 952        struct elevator_type *e;
 953
 954        if (!q->elevator)
 955                return -ENXIO;
 956
 957        strlcpy(elevator_name, name, sizeof(elevator_name));
 958        e = elevator_get(strstrip(elevator_name), true);
 959        if (!e) {
 960                printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
 961                return -EINVAL;
 962        }
 963
 964        if (!strcmp(elevator_name, q->elevator->type->elevator_name)) {
 965                elevator_put(e);
 966                return 0;
 967        }
 968
 969        return elevator_switch(q, e);
 970}
 971
 972int elevator_change(struct request_queue *q, const char *name)
 973{
 974        int ret;
 975
 976        /* Protect q->elevator from elevator_init() */
 977        mutex_lock(&q->sysfs_lock);
 978        ret = __elevator_change(q, name);
 979        mutex_unlock(&q->sysfs_lock);
 980
 981        return ret;
 982}
 983EXPORT_SYMBOL(elevator_change);
 984
 985ssize_t elv_iosched_store(struct request_queue *q, const char *name,
 986                          size_t count)
 987{
 988        int ret;
 989
 990        if (!q->elevator)
 991                return count;
 992
 993        ret = __elevator_change(q, name);
 994        if (!ret)
 995                return count;
 996
 997        printk(KERN_ERR "elevator: switch to %s failed\n", name);
 998        return ret;
 999}
1000
1001ssize_t elv_iosched_show(struct request_queue *q, char *name)
1002{
1003        struct elevator_queue *e = q->elevator;
1004        struct elevator_type *elv;
1005        struct elevator_type *__e;
1006        int len = 0;
1007
1008        if (!q->elevator || !blk_queue_stackable(q))
1009                return sprintf(name, "none\n");
1010
1011        elv = e->type;
1012
1013        spin_lock(&elv_list_lock);
1014        list_for_each_entry(__e, &elv_list, list) {
1015                if (!strcmp(elv->elevator_name, __e->elevator_name))
1016                        len += sprintf(name+len, "[%s] ", elv->elevator_name);
1017                else
1018                        len += sprintf(name+len, "%s ", __e->elevator_name);
1019        }
1020        spin_unlock(&elv_list_lock);
1021
1022        len += sprintf(len+name, "\n");
1023        return len;
1024}
1025
1026struct request *elv_rb_former_request(struct request_queue *q,
1027                                      struct request *rq)
1028{
1029        struct rb_node *rbprev = rb_prev(&rq->rb_node);
1030
1031        if (rbprev)
1032                return rb_entry_rq(rbprev);
1033
1034        return NULL;
1035}
1036EXPORT_SYMBOL(elv_rb_former_request);
1037
1038struct request *elv_rb_latter_request(struct request_queue *q,
1039                                      struct request *rq)
1040{
1041        struct rb_node *rbnext = rb_next(&rq->rb_node);
1042
1043        if (rbnext)
1044                return rb_entry_rq(rbnext);
1045
1046        return NULL;
1047}
1048EXPORT_SYMBOL(elv_rb_latter_request);
1049