Root/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#include <linux/pm_runtime.h>
38
39#include <trace/events/block.h>
40
41#include "blk.h"
42#include "blk-cgroup.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_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_merge_fn)
62        return e->type->ops.elevator_allow_merge_fn(q, rq, bio);
63
64    return 1;
65}
66
67/*
68 * can we safely merge with this request?
69 */
70bool elv_rq_merge_ok(struct request *rq, struct bio *bio)
71{
72    if (!blk_rq_merge_ok(rq, bio))
73        return 0;
74
75    if (!elv_iosched_allow_merge(rq, bio))
76        return 0;
77
78    return 1;
79}
80EXPORT_SYMBOL(elv_rq_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        goto err;
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;
168err:
169    kfree(eq);
170    elevator_put(e);
171    return NULL;
172}
173EXPORT_SYMBOL(elevator_alloc);
174
175static void elevator_release(struct kobject *kobj)
176{
177    struct elevator_queue *e;
178
179    e = container_of(kobj, struct elevator_queue, kobj);
180    elevator_put(e->type);
181    kfree(e);
182}
183
184int elevator_init(struct request_queue *q, char *name)
185{
186    struct elevator_type *e = NULL;
187    int err;
188
189    if (unlikely(q->elevator))
190        return 0;
191
192    INIT_LIST_HEAD(&q->queue_head);
193    q->last_merge = NULL;
194    q->end_sector = 0;
195    q->boundary_rq = NULL;
196
197    if (name) {
198        e = elevator_get(name, true);
199        if (!e)
200            return -EINVAL;
201    }
202
203    /*
204     * Use the default elevator specified by config boot param or
205     * config option. Don't try to load modules as we could be running
206     * off async and request_module() isn't allowed from async.
207     */
208    if (!e && *chosen_elevator) {
209        e = elevator_get(chosen_elevator, false);
210        if (!e)
211            printk(KERN_ERR "I/O scheduler %s not found\n",
212                            chosen_elevator);
213    }
214
215    if (!e) {
216        e = elevator_get(CONFIG_DEFAULT_IOSCHED, false);
217        if (!e) {
218            printk(KERN_ERR
219                "Default I/O scheduler not found. " \
220                "Using noop.\n");
221            e = elevator_get("noop", false);
222        }
223    }
224
225    err = e->ops.elevator_init_fn(q, e);
226    return 0;
227}
228EXPORT_SYMBOL(elevator_init);
229
230void elevator_exit(struct elevator_queue *e)
231{
232    mutex_lock(&e->sysfs_lock);
233    if (e->type->ops.elevator_exit_fn)
234        e->type->ops.elevator_exit_fn(e);
235    mutex_unlock(&e->sysfs_lock);
236
237    kobject_put(&e->kobj);
238}
239EXPORT_SYMBOL(elevator_exit);
240
241static inline void __elv_rqhash_del(struct request *rq)
242{
243    hash_del(&rq->hash);
244}
245
246static void elv_rqhash_del(struct request_queue *q, struct request *rq)
247{
248    if (ELV_ON_HASH(rq))
249        __elv_rqhash_del(rq);
250}
251
252static void elv_rqhash_add(struct request_queue *q, struct request *rq)
253{
254    struct elevator_queue *e = q->elevator;
255
256    BUG_ON(ELV_ON_HASH(rq));
257    hash_add(e->hash, &rq->hash, rq_hash_key(rq));
258}
259
260static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
261{
262    __elv_rqhash_del(rq);
263    elv_rqhash_add(q, rq);
264}
265
266static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
267{
268    struct elevator_queue *e = q->elevator;
269    struct hlist_node *next;
270    struct request *rq;
271
272    hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
273        BUG_ON(!ELV_ON_HASH(rq));
274
275        if (unlikely(!rq_mergeable(rq))) {
276            __elv_rqhash_del(rq);
277            continue;
278        }
279
280        if (rq_hash_key(rq) == offset)
281            return rq;
282    }
283
284    return NULL;
285}
286
287/*
288 * RB-tree support functions for inserting/lookup/removal of requests
289 * in a sorted RB tree.
290 */
291void elv_rb_add(struct rb_root *root, struct request *rq)
292{
293    struct rb_node **p = &root->rb_node;
294    struct rb_node *parent = NULL;
295    struct request *__rq;
296
297    while (*p) {
298        parent = *p;
299        __rq = rb_entry(parent, struct request, rb_node);
300
301        if (blk_rq_pos(rq) < blk_rq_pos(__rq))
302            p = &(*p)->rb_left;
303        else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
304            p = &(*p)->rb_right;
305    }
306
307    rb_link_node(&rq->rb_node, parent, p);
308    rb_insert_color(&rq->rb_node, root);
309}
310EXPORT_SYMBOL(elv_rb_add);
311
312void elv_rb_del(struct rb_root *root, struct request *rq)
313{
314    BUG_ON(RB_EMPTY_NODE(&rq->rb_node));
315    rb_erase(&rq->rb_node, root);
316    RB_CLEAR_NODE(&rq->rb_node);
317}
318EXPORT_SYMBOL(elv_rb_del);
319
320struct request *elv_rb_find(struct rb_root *root, sector_t sector)
321{
322    struct rb_node *n = root->rb_node;
323    struct request *rq;
324
325    while (n) {
326        rq = rb_entry(n, struct request, rb_node);
327
328        if (sector < blk_rq_pos(rq))
329            n = n->rb_left;
330        else if (sector > blk_rq_pos(rq))
331            n = n->rb_right;
332        else
333            return rq;
334    }
335
336    return NULL;
337}
338EXPORT_SYMBOL(elv_rb_find);
339
340/*
341 * Insert rq into dispatch queue of q. Queue lock must be held on
342 * entry. rq is sort instead into the dispatch queue. To be used by
343 * specific elevators.
344 */
345void elv_dispatch_sort(struct request_queue *q, struct request *rq)
346{
347    sector_t boundary;
348    struct list_head *entry;
349    int stop_flags;
350
351    if (q->last_merge == rq)
352        q->last_merge = NULL;
353
354    elv_rqhash_del(q, rq);
355
356    q->nr_sorted--;
357
358    boundary = q->end_sector;
359    stop_flags = REQ_SOFTBARRIER | REQ_STARTED;
360    list_for_each_prev(entry, &q->queue_head) {
361        struct request *pos = list_entry_rq(entry);
362
363        if ((rq->cmd_flags & REQ_DISCARD) !=
364            (pos->cmd_flags & REQ_DISCARD))
365            break;
366        if (rq_data_dir(rq) != rq_data_dir(pos))
367            break;
368        if (pos->cmd_flags & stop_flags)
369            break;
370        if (blk_rq_pos(rq) >= boundary) {
371            if (blk_rq_pos(pos) < boundary)
372                continue;
373        } else {
374            if (blk_rq_pos(pos) >= boundary)
375                break;
376        }
377        if (blk_rq_pos(rq) >= blk_rq_pos(pos))
378            break;
379    }
380
381    list_add(&rq->queuelist, entry);
382}
383EXPORT_SYMBOL(elv_dispatch_sort);
384
385/*
386 * Insert rq into dispatch queue of q. Queue lock must be held on
387 * entry. rq is added to the back of the dispatch queue. To be used by
388 * specific elevators.
389 */
390void elv_dispatch_add_tail(struct request_queue *q, struct request *rq)
391{
392    if (q->last_merge == rq)
393        q->last_merge = NULL;
394
395    elv_rqhash_del(q, rq);
396
397    q->nr_sorted--;
398
399    q->end_sector = rq_end_sector(rq);
400    q->boundary_rq = rq;
401    list_add_tail(&rq->queuelist, &q->queue_head);
402}
403EXPORT_SYMBOL(elv_dispatch_add_tail);
404
405int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
406{
407    struct elevator_queue *e = q->elevator;
408    struct request *__rq;
409    int ret;
410
411    /*
412     * Levels of merges:
413     * nomerges: No merges at all attempted
414     * noxmerges: Only simple one-hit cache try
415     * merges: All merge tries attempted
416     */
417    if (blk_queue_nomerges(q))
418        return ELEVATOR_NO_MERGE;
419
420    /*
421     * First try one-hit cache.
422     */
423    if (q->last_merge && elv_rq_merge_ok(q->last_merge, bio)) {
424        ret = blk_try_merge(q->last_merge, bio);
425        if (ret != ELEVATOR_NO_MERGE) {
426            *req = q->last_merge;
427            return ret;
428        }
429    }
430
431    if (blk_queue_noxmerges(q))
432        return ELEVATOR_NO_MERGE;
433
434    /*
435     * See if our hash lookup can find a potential backmerge.
436     */
437    __rq = elv_rqhash_find(q, bio->bi_sector);
438    if (__rq && elv_rq_merge_ok(__rq, bio)) {
439        *req = __rq;
440        return ELEVATOR_BACK_MERGE;
441    }
442
443    if (e->type->ops.elevator_merge_fn)
444        return e->type->ops.elevator_merge_fn(q, req, bio);
445
446    return ELEVATOR_NO_MERGE;
447}
448
449/*
450 * Attempt to do an insertion back merge. Only check for the case where
451 * we can append 'rq' to an existing request, so we can throw 'rq' away
452 * afterwards.
453 *
454 * Returns true if we merged, false otherwise
455 */
456static bool elv_attempt_insert_merge(struct request_queue *q,
457                     struct request *rq)
458{
459    struct request *__rq;
460    bool ret;
461
462    if (blk_queue_nomerges(q))
463        return false;
464
465    /*
466     * First try one-hit cache.
467     */
468    if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
469        return true;
470
471    if (blk_queue_noxmerges(q))
472        return false;
473
474    ret = false;
475    /*
476     * See if our hash lookup can find a potential backmerge.
477     */
478    while (1) {
479        __rq = elv_rqhash_find(q, blk_rq_pos(rq));
480        if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
481            break;
482
483        /* The merged request could be merged with others, try again */
484        ret = true;
485        rq = __rq;
486    }
487
488    return ret;
489}
490
491void 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
504void 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
523void 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#ifdef CONFIG_PM_RUNTIME
533static void blk_pm_requeue_request(struct request *rq)
534{
535    if (rq->q->dev && !(rq->cmd_flags & REQ_PM))
536        rq->q->nr_pending--;
537}
538
539static void blk_pm_add_request(struct request_queue *q, struct request *rq)
540{
541    if (q->dev && !(rq->cmd_flags & REQ_PM) && q->nr_pending++ == 0 &&
542        (q->rpm_status == RPM_SUSPENDED || q->rpm_status == RPM_SUSPENDING))
543        pm_request_resume(q->dev);
544}
545#else
546static inline void blk_pm_requeue_request(struct request *rq) {}
547static inline void blk_pm_add_request(struct request_queue *q,
548                      struct request *rq)
549{
550}
551#endif
552
553void elv_requeue_request(struct request_queue *q, struct request *rq)
554{
555    /*
556     * it already went through dequeue, we need to decrement the
557     * in_flight count again
558     */
559    if (blk_account_rq(rq)) {
560        q->in_flight[rq_is_sync(rq)]--;
561        if (rq->cmd_flags & REQ_SORTED)
562            elv_deactivate_rq(q, rq);
563    }
564
565    rq->cmd_flags &= ~REQ_STARTED;
566
567    blk_pm_requeue_request(rq);
568
569    __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
570}
571
572void elv_drain_elevator(struct request_queue *q)
573{
574    static int printed;
575
576    lockdep_assert_held(q->queue_lock);
577
578    while (q->elevator->type->ops.elevator_dispatch_fn(q, 1))
579        ;
580    if (q->nr_sorted && printed++ < 10) {
581        printk(KERN_ERR "%s: forced dispatching is broken "
582               "(nr_sorted=%u), please report this\n",
583               q->elevator->type->elevator_name, q->nr_sorted);
584    }
585}
586
587void __elv_add_request(struct request_queue *q, struct request *rq, int where)
588{
589    trace_block_rq_insert(q, rq);
590
591    blk_pm_add_request(q, rq);
592
593    rq->q = q;
594
595    if (rq->cmd_flags & REQ_SOFTBARRIER) {
596        /* barriers are scheduling boundary, update end_sector */
597        if (rq->cmd_type == REQ_TYPE_FS) {
598            q->end_sector = rq_end_sector(rq);
599            q->boundary_rq = rq;
600        }
601    } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
602            (where == ELEVATOR_INSERT_SORT ||
603             where == ELEVATOR_INSERT_SORT_MERGE))
604        where = ELEVATOR_INSERT_BACK;
605
606    switch (where) {
607    case ELEVATOR_INSERT_REQUEUE:
608    case ELEVATOR_INSERT_FRONT:
609        rq->cmd_flags |= REQ_SOFTBARRIER;
610        list_add(&rq->queuelist, &q->queue_head);
611        break;
612
613    case ELEVATOR_INSERT_BACK:
614        rq->cmd_flags |= REQ_SOFTBARRIER;
615        elv_drain_elevator(q);
616        list_add_tail(&rq->queuelist, &q->queue_head);
617        /*
618         * We kick the queue here for the following reasons.
619         * - The elevator might have returned NULL previously
620         * to delay requests and returned them now. As the
621         * queue wasn't empty before this request, ll_rw_blk
622         * won't run the queue on return, resulting in hang.
623         * - Usually, back inserted requests won't be merged
624         * with anything. There's no point in delaying queue
625         * processing.
626         */
627        __blk_run_queue(q);
628        break;
629
630    case ELEVATOR_INSERT_SORT_MERGE:
631        /*
632         * If we succeed in merging this request with one in the
633         * queue already, we are done - rq has now been freed,
634         * so no need to do anything further.
635         */
636        if (elv_attempt_insert_merge(q, rq))
637            break;
638    case ELEVATOR_INSERT_SORT:
639        BUG_ON(rq->cmd_type != REQ_TYPE_FS);
640        rq->cmd_flags |= REQ_SORTED;
641        q->nr_sorted++;
642        if (rq_mergeable(rq)) {
643            elv_rqhash_add(q, rq);
644            if (!q->last_merge)
645                q->last_merge = rq;
646        }
647
648        /*
649         * Some ioscheds (cfq) run q->request_fn directly, so
650         * rq cannot be accessed after calling
651         * elevator_add_req_fn.
652         */
653        q->elevator->type->ops.elevator_add_req_fn(q, rq);
654        break;
655
656    case ELEVATOR_INSERT_FLUSH:
657        rq->cmd_flags |= REQ_SOFTBARRIER;
658        blk_insert_flush(rq);
659        break;
660    default:
661        printk(KERN_ERR "%s: bad insertion point %d\n",
662               __func__, where);
663        BUG();
664    }
665}
666EXPORT_SYMBOL(__elv_add_request);
667
668void elv_add_request(struct request_queue *q, struct request *rq, int where)
669{
670    unsigned long flags;
671
672    spin_lock_irqsave(q->queue_lock, flags);
673    __elv_add_request(q, rq, where);
674    spin_unlock_irqrestore(q->queue_lock, flags);
675}
676EXPORT_SYMBOL(elv_add_request);
677
678struct request *elv_latter_request(struct request_queue *q, struct request *rq)
679{
680    struct elevator_queue *e = q->elevator;
681
682    if (e->type->ops.elevator_latter_req_fn)
683        return e->type->ops.elevator_latter_req_fn(q, rq);
684    return NULL;
685}
686
687struct request *elv_former_request(struct request_queue *q, struct request *rq)
688{
689    struct elevator_queue *e = q->elevator;
690
691    if (e->type->ops.elevator_former_req_fn)
692        return e->type->ops.elevator_former_req_fn(q, rq);
693    return NULL;
694}
695
696int elv_set_request(struct request_queue *q, struct request *rq,
697            struct bio *bio, gfp_t gfp_mask)
698{
699    struct elevator_queue *e = q->elevator;
700
701    if (e->type->ops.elevator_set_req_fn)
702        return e->type->ops.elevator_set_req_fn(q, rq, bio, gfp_mask);
703    return 0;
704}
705
706void elv_put_request(struct request_queue *q, struct request *rq)
707{
708    struct elevator_queue *e = q->elevator;
709
710    if (e->type->ops.elevator_put_req_fn)
711        e->type->ops.elevator_put_req_fn(rq);
712}
713
714int elv_may_queue(struct request_queue *q, int rw)
715{
716    struct elevator_queue *e = q->elevator;
717
718    if (e->type->ops.elevator_may_queue_fn)
719        return e->type->ops.elevator_may_queue_fn(q, rw);
720
721    return ELV_MQUEUE_MAY;
722}
723
724void elv_abort_queue(struct request_queue *q)
725{
726    struct request *rq;
727
728    blk_abort_flushes(q);
729
730    while (!list_empty(&q->queue_head)) {
731        rq = list_entry_rq(q->queue_head.next);
732        rq->cmd_flags |= REQ_QUIET;
733        trace_block_rq_abort(q, rq);
734        /*
735         * Mark this request as started so we don't trigger
736         * any debug logic in the end I/O path.
737         */
738        blk_start_request(rq);
739        __blk_end_request_all(rq, -EIO);
740    }
741}
742EXPORT_SYMBOL(elv_abort_queue);
743
744void elv_completed_request(struct request_queue *q, struct request *rq)
745{
746    struct elevator_queue *e = q->elevator;
747
748    /*
749     * request is released from the driver, io must be done
750     */
751    if (blk_account_rq(rq)) {
752        q->in_flight[rq_is_sync(rq)]--;
753        if ((rq->cmd_flags & REQ_SORTED) &&
754            e->type->ops.elevator_completed_req_fn)
755            e->type->ops.elevator_completed_req_fn(q, rq);
756    }
757}
758
759#define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)
760
761static ssize_t
762elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
763{
764    struct elv_fs_entry *entry = to_elv(attr);
765    struct elevator_queue *e;
766    ssize_t error;
767
768    if (!entry->show)
769        return -EIO;
770
771    e = container_of(kobj, struct elevator_queue, kobj);
772    mutex_lock(&e->sysfs_lock);
773    error = e->type ? entry->show(e, page) : -ENOENT;
774    mutex_unlock(&e->sysfs_lock);
775    return error;
776}
777
778static ssize_t
779elv_attr_store(struct kobject *kobj, struct attribute *attr,
780           const char *page, size_t length)
781{
782    struct elv_fs_entry *entry = to_elv(attr);
783    struct elevator_queue *e;
784    ssize_t error;
785
786    if (!entry->store)
787        return -EIO;
788
789    e = container_of(kobj, struct elevator_queue, kobj);
790    mutex_lock(&e->sysfs_lock);
791    error = e->type ? entry->store(e, page, length) : -ENOENT;
792    mutex_unlock(&e->sysfs_lock);
793    return error;
794}
795
796static const struct sysfs_ops elv_sysfs_ops = {
797    .show = elv_attr_show,
798    .store = elv_attr_store,
799};
800
801static struct kobj_type elv_ktype = {
802    .sysfs_ops = &elv_sysfs_ops,
803    .release = elevator_release,
804};
805
806int elv_register_queue(struct request_queue *q)
807{
808    struct elevator_queue *e = q->elevator;
809    int error;
810
811    error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
812    if (!error) {
813        struct elv_fs_entry *attr = e->type->elevator_attrs;
814        if (attr) {
815            while (attr->attr.name) {
816                if (sysfs_create_file(&e->kobj, &attr->attr))
817                    break;
818                attr++;
819            }
820        }
821        kobject_uevent(&e->kobj, KOBJ_ADD);
822        e->registered = 1;
823    }
824    return error;
825}
826EXPORT_SYMBOL(elv_register_queue);
827
828void elv_unregister_queue(struct request_queue *q)
829{
830    if (q) {
831        struct elevator_queue *e = q->elevator;
832
833        kobject_uevent(&e->kobj, KOBJ_REMOVE);
834        kobject_del(&e->kobj);
835        e->registered = 0;
836    }
837}
838EXPORT_SYMBOL(elv_unregister_queue);
839
840int elv_register(struct elevator_type *e)
841{
842    char *def = "";
843
844    /* create icq_cache if requested */
845    if (e->icq_size) {
846        if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
847            WARN_ON(e->icq_align < __alignof__(struct io_cq)))
848            return -EINVAL;
849
850        snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
851             "%s_io_cq", e->elevator_name);
852        e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
853                         e->icq_align, 0, NULL);
854        if (!e->icq_cache)
855            return -ENOMEM;
856    }
857
858    /* register, don't allow duplicate names */
859    spin_lock(&elv_list_lock);
860    if (elevator_find(e->elevator_name)) {
861        spin_unlock(&elv_list_lock);
862        if (e->icq_cache)
863            kmem_cache_destroy(e->icq_cache);
864        return -EBUSY;
865    }
866    list_add_tail(&e->list, &elv_list);
867    spin_unlock(&elv_list_lock);
868
869    /* print pretty message */
870    if (!strcmp(e->elevator_name, chosen_elevator) ||
871            (!*chosen_elevator &&
872             !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
873                def = " (default)";
874
875    printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
876                                def);
877    return 0;
878}
879EXPORT_SYMBOL_GPL(elv_register);
880
881void elv_unregister(struct elevator_type *e)
882{
883    /* unregister */
884    spin_lock(&elv_list_lock);
885    list_del_init(&e->list);
886    spin_unlock(&elv_list_lock);
887
888    /*
889     * Destroy icq_cache if it exists. icq's are RCU managed. Make
890     * sure all RCU operations are complete before proceeding.
891     */
892    if (e->icq_cache) {
893        rcu_barrier();
894        kmem_cache_destroy(e->icq_cache);
895        e->icq_cache = NULL;
896    }
897}
898EXPORT_SYMBOL_GPL(elv_unregister);
899
900/*
901 * switch to new_e io scheduler. be careful not to introduce deadlocks -
902 * we don't free the old io scheduler, before we have allocated what we
903 * need for the new one. this way we have a chance of going back to the old
904 * one, if the new one fails init for some reason.
905 */
906static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
907{
908    struct elevator_queue *old = q->elevator;
909    bool registered = old->registered;
910    int err;
911
912    /*
913     * Turn on BYPASS and drain all requests w/ elevator private data.
914     * Block layer doesn't call into a quiesced elevator - all requests
915     * are directly put on the dispatch list without elevator data
916     * using INSERT_BACK. All requests have SOFTBARRIER set and no
917     * merge happens either.
918     */
919    blk_queue_bypass_start(q);
920
921    /* unregister and clear all auxiliary data of the old elevator */
922    if (registered)
923        elv_unregister_queue(q);
924
925    spin_lock_irq(q->queue_lock);
926    ioc_clear_queue(q);
927    spin_unlock_irq(q->queue_lock);
928
929    /* allocate, init and register new elevator */
930    err = new_e->ops.elevator_init_fn(q, new_e);
931    if (err)
932        goto fail_init;
933
934    if (registered) {
935        err = elv_register_queue(q);
936        if (err)
937            goto fail_register;
938    }
939
940    /* done, kill the old one and finish */
941    elevator_exit(old);
942    blk_queue_bypass_end(q);
943
944    blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
945
946    return 0;
947
948fail_register:
949    elevator_exit(q->elevator);
950fail_init:
951    /* switch failed, restore and re-register old elevator */
952    q->elevator = old;
953    elv_register_queue(q);
954    blk_queue_bypass_end(q);
955
956    return err;
957}
958
959/*
960 * Switch this queue to the given IO scheduler.
961 */
962int elevator_change(struct request_queue *q, const char *name)
963{
964    char elevator_name[ELV_NAME_MAX];
965    struct elevator_type *e;
966
967    if (!q->elevator)
968        return -ENXIO;
969
970    strlcpy(elevator_name, name, sizeof(elevator_name));
971    e = elevator_get(strstrip(elevator_name), true);
972    if (!e) {
973        printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
974        return -EINVAL;
975    }
976
977    if (!strcmp(elevator_name, q->elevator->type->elevator_name)) {
978        elevator_put(e);
979        return 0;
980    }
981
982    return elevator_switch(q, e);
983}
984EXPORT_SYMBOL(elevator_change);
985
986ssize_t elv_iosched_store(struct request_queue *q, const char *name,
987              size_t count)
988{
989    int ret;
990
991    if (!q->elevator)
992        return count;
993
994    ret = elevator_change(q, name);
995    if (!ret)
996        return count;
997
998    printk(KERN_ERR "elevator: switch to %s failed\n", name);
999    return ret;
1000}
1001
1002ssize_t elv_iosched_show(struct request_queue *q, char *name)
1003{
1004    struct elevator_queue *e = q->elevator;
1005    struct elevator_type *elv;
1006    struct elevator_type *__e;
1007    int len = 0;
1008
1009    if (!q->elevator || !blk_queue_stackable(q))
1010        return sprintf(name, "none\n");
1011
1012    elv = e->type;
1013
1014    spin_lock(&elv_list_lock);
1015    list_for_each_entry(__e, &elv_list, list) {
1016        if (!strcmp(elv->elevator_name, __e->elevator_name))
1017            len += sprintf(name+len, "[%s] ", elv->elevator_name);
1018        else
1019            len += sprintf(name+len, "%s ", __e->elevator_name);
1020    }
1021    spin_unlock(&elv_list_lock);
1022
1023    len += sprintf(len+name, "\n");
1024    return len;
1025}
1026
1027struct request *elv_rb_former_request(struct request_queue *q,
1028                      struct request *rq)
1029{
1030    struct rb_node *rbprev = rb_prev(&rq->rb_node);
1031
1032    if (rbprev)
1033        return rb_entry_rq(rbprev);
1034
1035    return NULL;
1036}
1037EXPORT_SYMBOL(elv_rb_former_request);
1038
1039struct request *elv_rb_latter_request(struct request_queue *q,
1040                      struct request *rq)
1041{
1042    struct rb_node *rbnext = rb_next(&rq->rb_node);
1043
1044    if (rbnext)
1045        return rb_entry_rq(rbnext);
1046
1047    return NULL;
1048}
1049EXPORT_SYMBOL(elv_rb_latter_request);
1050

Archive Download this file



interactive