808 lines · c
1// SPDX-License-Identifier: GPL-2.02/*3 * Block device elevator/IO-scheduler.4 *5 * Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE6 *7 * 30042000 Jens Axboe <axboe@kernel.dk> :8 *9 * Split the elevator a bit so that it is possible to choose a different10 * one or even write a new "plug in". There are three pieces:11 * - elevator_fn, inserts a new request in the queue list12 * - elevator_merge_fn, decides whether a new buffer can be merged with13 * an existing request14 * - elevator_dequeue_fn, called when a request is taken off the active list15 *16 * 20082000 Dave Jones <davej@suse.de> :17 * Removed tests for max-bomb-segments, which was breaking elvtune18 * when run without -bN19 *20 * Jens:21 * - Rework again to work with bio instead of buffer_heads22 * - loose bi_dev comparisons, partition handling is right now23 * - completely modularize elevator setup and teardown24 *25 */26#include <linux/kernel.h>27#include <linux/fs.h>28#include <linux/blkdev.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 "elevator.h"42#include "blk.h"43#include "blk-mq-sched.h"44#include "blk-pm.h"45#include "blk-wbt.h"46#include "blk-cgroup.h"47 48static DEFINE_SPINLOCK(elv_list_lock);49static LIST_HEAD(elv_list);50 51/*52 * Merge hash stuff.53 */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 be58 * merged with rq.59 */60static bool elv_iosched_allow_bio_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->type->ops.allow_merge)66 return e->type->ops.allow_merge(q, rq, bio);67 68 return true;69}70 71/*72 * can we safely merge with this request?73 */74bool elv_bio_merge_ok(struct request *rq, struct bio *bio)75{76 if (!blk_rq_merge_ok(rq, bio))77 return false;78 79 if (!elv_iosched_allow_bio_merge(rq, bio))80 return false;81 82 return true;83}84EXPORT_SYMBOL(elv_bio_merge_ok);85 86/**87 * elevator_match - Check whether @e's name or alias matches @name88 * @e: Scheduler to test89 * @name: Elevator name to test90 *91 * Return true if the elevator @e's name or alias matches @name.92 */93static bool elevator_match(const struct elevator_type *e, const char *name)94{95 return !strcmp(e->elevator_name, name) ||96 (e->elevator_alias && !strcmp(e->elevator_alias, name));97}98 99static struct elevator_type *__elevator_find(const char *name)100{101 struct elevator_type *e;102 103 list_for_each_entry(e, &elv_list, list)104 if (elevator_match(e, name))105 return e;106 return NULL;107}108 109static struct elevator_type *elevator_find_get(const char *name)110{111 struct elevator_type *e;112 113 spin_lock(&elv_list_lock);114 e = __elevator_find(name);115 if (e && (!elevator_tryget(e)))116 e = NULL;117 spin_unlock(&elv_list_lock);118 return e;119}120 121static const struct kobj_type elv_ktype;122 123struct elevator_queue *elevator_alloc(struct request_queue *q,124 struct elevator_type *e)125{126 struct elevator_queue *eq;127 128 eq = kzalloc_node(sizeof(*eq), GFP_KERNEL, q->node);129 if (unlikely(!eq))130 return NULL;131 132 __elevator_get(e);133 eq->type = e;134 kobject_init(&eq->kobj, &elv_ktype);135 mutex_init(&eq->sysfs_lock);136 hash_init(eq->hash);137 138 return eq;139}140EXPORT_SYMBOL(elevator_alloc);141 142static void elevator_release(struct kobject *kobj)143{144 struct elevator_queue *e;145 146 e = container_of(kobj, struct elevator_queue, kobj);147 elevator_put(e->type);148 kfree(e);149}150 151void elevator_exit(struct request_queue *q)152{153 struct elevator_queue *e = q->elevator;154 155 ioc_clear_queue(q);156 blk_mq_sched_free_rqs(q);157 158 mutex_lock(&e->sysfs_lock);159 blk_mq_exit_sched(q, e);160 mutex_unlock(&e->sysfs_lock);161 162 kobject_put(&e->kobj);163}164 165static inline void __elv_rqhash_del(struct request *rq)166{167 hash_del(&rq->hash);168 rq->rq_flags &= ~RQF_HASHED;169}170 171void elv_rqhash_del(struct request_queue *q, struct request *rq)172{173 if (ELV_ON_HASH(rq))174 __elv_rqhash_del(rq);175}176EXPORT_SYMBOL_GPL(elv_rqhash_del);177 178void elv_rqhash_add(struct request_queue *q, struct request *rq)179{180 struct elevator_queue *e = q->elevator;181 182 BUG_ON(ELV_ON_HASH(rq));183 hash_add(e->hash, &rq->hash, rq_hash_key(rq));184 rq->rq_flags |= RQF_HASHED;185}186EXPORT_SYMBOL_GPL(elv_rqhash_add);187 188void elv_rqhash_reposition(struct request_queue *q, struct request *rq)189{190 __elv_rqhash_del(rq);191 elv_rqhash_add(q, rq);192}193 194struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)195{196 struct elevator_queue *e = q->elevator;197 struct hlist_node *next;198 struct request *rq;199 200 hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {201 BUG_ON(!ELV_ON_HASH(rq));202 203 if (unlikely(!rq_mergeable(rq))) {204 __elv_rqhash_del(rq);205 continue;206 }207 208 if (rq_hash_key(rq) == offset)209 return rq;210 }211 212 return NULL;213}214 215/*216 * RB-tree support functions for inserting/lookup/removal of requests217 * in a sorted RB tree.218 */219void elv_rb_add(struct rb_root *root, struct request *rq)220{221 struct rb_node **p = &root->rb_node;222 struct rb_node *parent = NULL;223 struct request *__rq;224 225 while (*p) {226 parent = *p;227 __rq = rb_entry(parent, struct request, rb_node);228 229 if (blk_rq_pos(rq) < blk_rq_pos(__rq))230 p = &(*p)->rb_left;231 else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))232 p = &(*p)->rb_right;233 }234 235 rb_link_node(&rq->rb_node, parent, p);236 rb_insert_color(&rq->rb_node, root);237}238EXPORT_SYMBOL(elv_rb_add);239 240void elv_rb_del(struct rb_root *root, struct request *rq)241{242 BUG_ON(RB_EMPTY_NODE(&rq->rb_node));243 rb_erase(&rq->rb_node, root);244 RB_CLEAR_NODE(&rq->rb_node);245}246EXPORT_SYMBOL(elv_rb_del);247 248struct request *elv_rb_find(struct rb_root *root, sector_t sector)249{250 struct rb_node *n = root->rb_node;251 struct request *rq;252 253 while (n) {254 rq = rb_entry(n, struct request, rb_node);255 256 if (sector < blk_rq_pos(rq))257 n = n->rb_left;258 else if (sector > blk_rq_pos(rq))259 n = n->rb_right;260 else261 return rq;262 }263 264 return NULL;265}266EXPORT_SYMBOL(elv_rb_find);267 268enum elv_merge elv_merge(struct request_queue *q, struct request **req,269 struct bio *bio)270{271 struct elevator_queue *e = q->elevator;272 struct request *__rq;273 274 /*275 * Levels of merges:276 * nomerges: No merges at all attempted277 * noxmerges: Only simple one-hit cache try278 * merges: All merge tries attempted279 */280 if (blk_queue_nomerges(q) || !bio_mergeable(bio))281 return ELEVATOR_NO_MERGE;282 283 /*284 * First try one-hit cache.285 */286 if (q->last_merge && elv_bio_merge_ok(q->last_merge, bio)) {287 enum elv_merge ret = blk_try_merge(q->last_merge, bio);288 289 if (ret != ELEVATOR_NO_MERGE) {290 *req = q->last_merge;291 return ret;292 }293 }294 295 if (blk_queue_noxmerges(q))296 return ELEVATOR_NO_MERGE;297 298 /*299 * See if our hash lookup can find a potential backmerge.300 */301 __rq = elv_rqhash_find(q, bio->bi_iter.bi_sector);302 if (__rq && elv_bio_merge_ok(__rq, bio)) {303 *req = __rq;304 305 if (blk_discard_mergable(__rq))306 return ELEVATOR_DISCARD_MERGE;307 return ELEVATOR_BACK_MERGE;308 }309 310 if (e->type->ops.request_merge)311 return e->type->ops.request_merge(q, req, bio);312 313 return ELEVATOR_NO_MERGE;314}315 316/*317 * Attempt to do an insertion back merge. Only check for the case where318 * we can append 'rq' to an existing request, so we can throw 'rq' away319 * afterwards.320 *321 * Returns true if we merged, false otherwise. 'free' will contain all322 * requests that need to be freed.323 */324bool elv_attempt_insert_merge(struct request_queue *q, struct request *rq,325 struct list_head *free)326{327 struct request *__rq;328 bool ret;329 330 if (blk_queue_nomerges(q))331 return false;332 333 /*334 * First try one-hit cache.335 */336 if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq)) {337 list_add(&rq->queuelist, free);338 return true;339 }340 341 if (blk_queue_noxmerges(q))342 return false;343 344 ret = false;345 /*346 * See if our hash lookup can find a potential backmerge.347 */348 while (1) {349 __rq = elv_rqhash_find(q, blk_rq_pos(rq));350 if (!__rq || !blk_attempt_req_merge(q, __rq, rq))351 break;352 353 list_add(&rq->queuelist, free);354 /* The merged request could be merged with others, try again */355 ret = true;356 rq = __rq;357 }358 359 return ret;360}361 362void elv_merged_request(struct request_queue *q, struct request *rq,363 enum elv_merge type)364{365 struct elevator_queue *e = q->elevator;366 367 if (e->type->ops.request_merged)368 e->type->ops.request_merged(q, rq, type);369 370 if (type == ELEVATOR_BACK_MERGE)371 elv_rqhash_reposition(q, rq);372 373 q->last_merge = rq;374}375 376void elv_merge_requests(struct request_queue *q, struct request *rq,377 struct request *next)378{379 struct elevator_queue *e = q->elevator;380 381 if (e->type->ops.requests_merged)382 e->type->ops.requests_merged(q, rq, next);383 384 elv_rqhash_reposition(q, rq);385 q->last_merge = rq;386}387 388struct request *elv_latter_request(struct request_queue *q, struct request *rq)389{390 struct elevator_queue *e = q->elevator;391 392 if (e->type->ops.next_request)393 return e->type->ops.next_request(q, rq);394 395 return NULL;396}397 398struct request *elv_former_request(struct request_queue *q, struct request *rq)399{400 struct elevator_queue *e = q->elevator;401 402 if (e->type->ops.former_request)403 return e->type->ops.former_request(q, rq);404 405 return NULL;406}407 408#define to_elv(atr) container_of((atr), struct elv_fs_entry, attr)409 410static ssize_t411elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)412{413 struct elv_fs_entry *entry = to_elv(attr);414 struct elevator_queue *e;415 ssize_t error;416 417 if (!entry->show)418 return -EIO;419 420 e = container_of(kobj, struct elevator_queue, kobj);421 mutex_lock(&e->sysfs_lock);422 error = e->type ? entry->show(e, page) : -ENOENT;423 mutex_unlock(&e->sysfs_lock);424 return error;425}426 427static ssize_t428elv_attr_store(struct kobject *kobj, struct attribute *attr,429 const char *page, size_t length)430{431 struct elv_fs_entry *entry = to_elv(attr);432 struct elevator_queue *e;433 ssize_t error;434 435 if (!entry->store)436 return -EIO;437 438 e = container_of(kobj, struct elevator_queue, kobj);439 mutex_lock(&e->sysfs_lock);440 error = e->type ? entry->store(e, page, length) : -ENOENT;441 mutex_unlock(&e->sysfs_lock);442 return error;443}444 445static const struct sysfs_ops elv_sysfs_ops = {446 .show = elv_attr_show,447 .store = elv_attr_store,448};449 450static const struct kobj_type elv_ktype = {451 .sysfs_ops = &elv_sysfs_ops,452 .release = elevator_release,453};454 455int elv_register_queue(struct request_queue *q, bool uevent)456{457 struct elevator_queue *e = q->elevator;458 int error;459 460 lockdep_assert_held(&q->sysfs_lock);461 462 error = kobject_add(&e->kobj, &q->disk->queue_kobj, "iosched");463 if (!error) {464 struct elv_fs_entry *attr = e->type->elevator_attrs;465 if (attr) {466 while (attr->attr.name) {467 if (sysfs_create_file(&e->kobj, &attr->attr))468 break;469 attr++;470 }471 }472 if (uevent)473 kobject_uevent(&e->kobj, KOBJ_ADD);474 475 set_bit(ELEVATOR_FLAG_REGISTERED, &e->flags);476 }477 return error;478}479 480void elv_unregister_queue(struct request_queue *q)481{482 struct elevator_queue *e = q->elevator;483 484 lockdep_assert_held(&q->sysfs_lock);485 486 if (e && test_and_clear_bit(ELEVATOR_FLAG_REGISTERED, &e->flags)) {487 kobject_uevent(&e->kobj, KOBJ_REMOVE);488 kobject_del(&e->kobj);489 }490}491 492int elv_register(struct elevator_type *e)493{494 /* finish request is mandatory */495 if (WARN_ON_ONCE(!e->ops.finish_request))496 return -EINVAL;497 /* insert_requests and dispatch_request are mandatory */498 if (WARN_ON_ONCE(!e->ops.insert_requests || !e->ops.dispatch_request))499 return -EINVAL;500 501 /* create icq_cache if requested */502 if (e->icq_size) {503 if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||504 WARN_ON(e->icq_align < __alignof__(struct io_cq)))505 return -EINVAL;506 507 snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),508 "%s_io_cq", e->elevator_name);509 e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,510 e->icq_align, 0, NULL);511 if (!e->icq_cache)512 return -ENOMEM;513 }514 515 /* register, don't allow duplicate names */516 spin_lock(&elv_list_lock);517 if (__elevator_find(e->elevator_name)) {518 spin_unlock(&elv_list_lock);519 kmem_cache_destroy(e->icq_cache);520 return -EBUSY;521 }522 list_add_tail(&e->list, &elv_list);523 spin_unlock(&elv_list_lock);524 525 printk(KERN_INFO "io scheduler %s registered\n", e->elevator_name);526 527 return 0;528}529EXPORT_SYMBOL_GPL(elv_register);530 531void elv_unregister(struct elevator_type *e)532{533 /* unregister */534 spin_lock(&elv_list_lock);535 list_del_init(&e->list);536 spin_unlock(&elv_list_lock);537 538 /*539 * Destroy icq_cache if it exists. icq's are RCU managed. Make540 * sure all RCU operations are complete before proceeding.541 */542 if (e->icq_cache) {543 rcu_barrier();544 kmem_cache_destroy(e->icq_cache);545 e->icq_cache = NULL;546 }547}548EXPORT_SYMBOL_GPL(elv_unregister);549 550static inline bool elv_support_iosched(struct request_queue *q)551{552 if (!queue_is_mq(q) ||553 (q->tag_set->flags & BLK_MQ_F_NO_SCHED))554 return false;555 return true;556}557 558/*559 * For single queue devices, default to using mq-deadline. If we have multiple560 * queues or mq-deadline is not available, default to "none".561 */562static struct elevator_type *elevator_get_default(struct request_queue *q)563{564 if (q->tag_set->flags & BLK_MQ_F_NO_SCHED_BY_DEFAULT)565 return NULL;566 567 if (q->nr_hw_queues != 1 &&568 !blk_mq_is_shared_tags(q->tag_set->flags))569 return NULL;570 571 return elevator_find_get("mq-deadline");572}573 574/*575 * Use the default elevator settings. If the chosen elevator initialization576 * fails, fall back to the "none" elevator (no elevator).577 */578void elevator_init_mq(struct request_queue *q)579{580 struct elevator_type *e;581 int err;582 583 if (!elv_support_iosched(q))584 return;585 586 WARN_ON_ONCE(blk_queue_registered(q));587 588 if (unlikely(q->elevator))589 return;590 591 e = elevator_get_default(q);592 if (!e)593 return;594 595 /*596 * We are called before adding disk, when there isn't any FS I/O,597 * so freezing queue plus canceling dispatch work is enough to598 * drain any dispatch activities originated from passthrough599 * requests, then no need to quiesce queue which may add long boot600 * latency, especially when lots of disks are involved.601 */602 blk_mq_freeze_queue(q);603 blk_mq_cancel_work_sync(q);604 605 err = blk_mq_init_sched(q, e);606 607 blk_mq_unfreeze_queue(q);608 609 if (err) {610 pr_warn("\"%s\" elevator initialization failed, "611 "falling back to \"none\"\n", e->elevator_name);612 }613 614 elevator_put(e);615}616 617/*618 * Switch to new_e io scheduler.619 *620 * If switching fails, we are most likely running out of memory and not able621 * to restore the old io scheduler, so leaving the io scheduler being none.622 */623int elevator_switch(struct request_queue *q, struct elevator_type *new_e)624{625 int ret;626 627 lockdep_assert_held(&q->sysfs_lock);628 629 blk_mq_freeze_queue(q);630 blk_mq_quiesce_queue(q);631 632 if (q->elevator) {633 elv_unregister_queue(q);634 elevator_exit(q);635 }636 637 ret = blk_mq_init_sched(q, new_e);638 if (ret)639 goto out_unfreeze;640 641 ret = elv_register_queue(q, true);642 if (ret) {643 elevator_exit(q);644 goto out_unfreeze;645 }646 blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);647 648out_unfreeze:649 blk_mq_unquiesce_queue(q);650 blk_mq_unfreeze_queue(q);651 652 if (ret) {653 pr_warn("elv: switch to \"%s\" failed, falling back to \"none\"\n",654 new_e->elevator_name);655 }656 657 return ret;658}659 660void elevator_disable(struct request_queue *q)661{662 lockdep_assert_held(&q->sysfs_lock);663 664 blk_mq_freeze_queue(q);665 blk_mq_quiesce_queue(q);666 667 elv_unregister_queue(q);668 elevator_exit(q);669 blk_queue_flag_clear(QUEUE_FLAG_SQ_SCHED, q);670 q->elevator = NULL;671 q->nr_requests = q->tag_set->queue_depth;672 blk_add_trace_msg(q, "elv switch: none");673 674 blk_mq_unquiesce_queue(q);675 blk_mq_unfreeze_queue(q);676}677 678/*679 * Switch this queue to the given IO scheduler.680 */681static int elevator_change(struct request_queue *q, const char *elevator_name)682{683 struct elevator_type *e;684 int ret;685 686 /* Make sure queue is not in the middle of being removed */687 if (!blk_queue_registered(q))688 return -ENOENT;689 690 if (!strncmp(elevator_name, "none", 4)) {691 if (q->elevator)692 elevator_disable(q);693 return 0;694 }695 696 if (q->elevator && elevator_match(q->elevator->type, elevator_name))697 return 0;698 699 e = elevator_find_get(elevator_name);700 if (!e)701 return -EINVAL;702 ret = elevator_switch(q, e);703 elevator_put(e);704 return ret;705}706 707int elv_iosched_load_module(struct gendisk *disk, const char *buf,708 size_t count)709{710 char elevator_name[ELV_NAME_MAX];711 struct elevator_type *found;712 const char *name;713 714 if (!elv_support_iosched(disk->queue))715 return -EOPNOTSUPP;716 717 strscpy(elevator_name, buf, sizeof(elevator_name));718 name = strstrip(elevator_name);719 720 spin_lock(&elv_list_lock);721 found = __elevator_find(name);722 spin_unlock(&elv_list_lock);723 724 if (!found)725 request_module("%s-iosched", name);726 727 return 0;728}729 730ssize_t elv_iosched_store(struct gendisk *disk, const char *buf,731 size_t count)732{733 char elevator_name[ELV_NAME_MAX];734 int ret;735 736 if (!elv_support_iosched(disk->queue))737 return count;738 739 strscpy(elevator_name, buf, sizeof(elevator_name));740 ret = elevator_change(disk->queue, strstrip(elevator_name));741 if (!ret)742 return count;743 return ret;744}745 746ssize_t elv_iosched_show(struct gendisk *disk, char *name)747{748 struct request_queue *q = disk->queue;749 struct elevator_queue *eq = q->elevator;750 struct elevator_type *cur = NULL, *e;751 int len = 0;752 753 if (!elv_support_iosched(q))754 return sprintf(name, "none\n");755 756 if (!q->elevator) {757 len += sprintf(name+len, "[none] ");758 } else {759 len += sprintf(name+len, "none ");760 cur = eq->type;761 }762 763 spin_lock(&elv_list_lock);764 list_for_each_entry(e, &elv_list, list) {765 if (e == cur)766 len += sprintf(name+len, "[%s] ", e->elevator_name);767 else768 len += sprintf(name+len, "%s ", e->elevator_name);769 }770 spin_unlock(&elv_list_lock);771 772 len += sprintf(name+len, "\n");773 return len;774}775 776struct request *elv_rb_former_request(struct request_queue *q,777 struct request *rq)778{779 struct rb_node *rbprev = rb_prev(&rq->rb_node);780 781 if (rbprev)782 return rb_entry_rq(rbprev);783 784 return NULL;785}786EXPORT_SYMBOL(elv_rb_former_request);787 788struct request *elv_rb_latter_request(struct request_queue *q,789 struct request *rq)790{791 struct rb_node *rbnext = rb_next(&rq->rb_node);792 793 if (rbnext)794 return rb_entry_rq(rbnext);795 796 return NULL;797}798EXPORT_SYMBOL(elv_rb_latter_request);799 800static int __init elevator_setup(char *str)801{802 pr_warn("Kernel parameter elevator= does not have any effect anymore.\n"803 "Please use sysfs to set IO scheduler for individual devices.\n");804 return 1;805}806 807__setup("elevator=", elevator_setup);808