1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2011 STRATO. All rights reserved.
6 #include <linux/sched.h>
7 #include <linux/pagemap.h>
8 #include <linux/writeback.h>
9 #include <linux/blkdev.h>
10 #include <linux/slab.h>
11 #include <linux/workqueue.h>
15 #include "transaction.h"
16 #include "dev-replace.h"
17 #include "block-group.h"
22 * This is the implementation for the generic read ahead framework.
24 * To trigger a readahead, btrfs_reada_add must be called. It will start
25 * a read ahead for the given range [start, end) on tree root. The returned
26 * handle can either be used to wait on the readahead to finish
27 * (btrfs_reada_wait), or to send it to the background (btrfs_reada_detach).
29 * The read ahead works as follows:
30 * On btrfs_reada_add, the root of the tree is inserted into a radix_tree.
31 * reada_start_machine will then search for extents to prefetch and trigger
32 * some reads. When a read finishes for a node, all contained node/leaf
33 * pointers that lie in the given range will also be enqueued. The reads will
34 * be triggered in sequential order, thus giving a big win over a naive
35 * enumeration. It will also make use of multi-device layouts. Each disk
36 * will have its on read pointer and all disks will by utilized in parallel.
37 * Also will no two disks read both sides of a mirror simultaneously, as this
38 * would waste seeking capacity. Instead both disks will read different parts
40 * Any number of readaheads can be started in parallel. The read order will be
41 * determined globally, i.e. 2 parallel readaheads will normally finish faster
42 * than the 2 started one after another.
45 #define MAX_IN_FLIGHT 6
48 struct list_head list;
49 struct reada_control *rc;
57 struct list_head extctl;
60 struct reada_zone *zones[BTRFS_MAX_MIRRORS];
70 struct list_head list;
73 struct btrfs_device *device;
74 struct btrfs_device *devs[BTRFS_MAX_MIRRORS]; /* full list, incl
80 struct reada_machine_work {
81 struct btrfs_work work;
82 struct btrfs_fs_info *fs_info;
85 static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
86 static void reada_control_release(struct kref *kref);
87 static void reada_zone_release(struct kref *kref);
88 static void reada_start_machine(struct btrfs_fs_info *fs_info);
89 static void __reada_start_machine(struct btrfs_fs_info *fs_info);
91 static int reada_add_block(struct reada_control *rc, u64 logical,
92 struct btrfs_key *top, u64 owner_root,
93 u64 generation, int level);
96 /* in case of err, eb might be NULL */
97 static void __readahead_hook(struct btrfs_fs_info *fs_info,
98 struct reada_extent *re, struct extent_buffer *eb,
105 struct list_head list;
107 spin_lock(&re->lock);
109 * just take the full list from the extent. afterwards we
110 * don't need the lock anymore
112 list_replace_init(&re->extctl, &list);
114 spin_unlock(&re->lock);
117 * this is the error case, the extent buffer has not been
118 * read correctly. We won't access anything from it and
119 * just cleanup our data structures. Effectively this will
120 * cut the branch below this node from read ahead.
126 * FIXME: currently we just set nritems to 0 if this is a leaf,
127 * effectively ignoring the content. In a next step we could
128 * trigger more readahead depending from the content, e.g.
129 * fetch the checksums for the extents in the leaf.
131 if (!btrfs_header_level(eb))
134 nritems = btrfs_header_nritems(eb);
135 generation = btrfs_header_generation(eb);
136 for (i = 0; i < nritems; i++) {
137 struct reada_extctl *rec;
139 struct btrfs_key key;
140 struct btrfs_key next_key;
142 btrfs_node_key_to_cpu(eb, &key, i);
144 btrfs_node_key_to_cpu(eb, &next_key, i + 1);
147 bytenr = btrfs_node_blockptr(eb, i);
148 n_gen = btrfs_node_ptr_generation(eb, i);
150 list_for_each_entry(rec, &list, list) {
151 struct reada_control *rc = rec->rc;
154 * if the generation doesn't match, just ignore this
155 * extctl. This will probably cut off a branch from
156 * prefetch. Alternatively one could start a new (sub-)
157 * prefetch for this branch, starting again from root.
158 * FIXME: move the generation check out of this loop
161 if (rec->generation != generation) {
163 "generation mismatch for (%llu,%d,%llu) %llu != %llu",
164 key.objectid, key.type, key.offset,
165 rec->generation, generation);
168 if (rec->generation == generation &&
169 btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
170 btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
171 reada_add_block(rc, bytenr, &next_key,
172 btrfs_header_owner(eb), n_gen,
173 btrfs_header_level(eb) - 1);
179 * free extctl records
181 while (!list_empty(&list)) {
182 struct reada_control *rc;
183 struct reada_extctl *rec;
185 rec = list_first_entry(&list, struct reada_extctl, list);
186 list_del(&rec->list);
190 kref_get(&rc->refcnt);
191 if (atomic_dec_and_test(&rc->elems)) {
192 kref_put(&rc->refcnt, reada_control_release);
195 kref_put(&rc->refcnt, reada_control_release);
197 reada_extent_put(fs_info, re); /* one ref for each entry */
203 int btree_readahead_hook(struct extent_buffer *eb, int err)
205 struct btrfs_fs_info *fs_info = eb->fs_info;
207 struct reada_extent *re;
210 spin_lock(&fs_info->reada_lock);
211 re = radix_tree_lookup(&fs_info->reada_tree,
212 eb->start >> fs_info->sectorsize_bits);
215 spin_unlock(&fs_info->reada_lock);
221 __readahead_hook(fs_info, re, eb, err);
222 reada_extent_put(fs_info, re); /* our ref */
225 reada_start_machine(fs_info);
229 static struct reada_zone *reada_find_zone(struct btrfs_device *dev, u64 logical,
230 struct btrfs_io_context *bioc)
232 struct btrfs_fs_info *fs_info = dev->fs_info;
234 struct reada_zone *zone;
235 struct btrfs_block_group *cache = NULL;
241 spin_lock(&fs_info->reada_lock);
242 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
243 logical >> fs_info->sectorsize_bits, 1);
244 if (ret == 1 && logical >= zone->start && logical <= zone->end) {
245 kref_get(&zone->refcnt);
246 spin_unlock(&fs_info->reada_lock);
250 spin_unlock(&fs_info->reada_lock);
252 cache = btrfs_lookup_block_group(fs_info, logical);
256 start = cache->start;
257 end = start + cache->length - 1;
258 btrfs_put_block_group(cache);
260 zone = kzalloc(sizeof(*zone), GFP_KERNEL);
264 ret = radix_tree_preload(GFP_KERNEL);
272 INIT_LIST_HEAD(&zone->list);
273 spin_lock_init(&zone->lock);
275 kref_init(&zone->refcnt);
277 zone->device = dev; /* our device always sits at index 0 */
278 for (i = 0; i < bioc->num_stripes; ++i) {
279 /* bounds have already been checked */
280 zone->devs[i] = bioc->stripes[i].dev;
282 zone->ndevs = bioc->num_stripes;
284 spin_lock(&fs_info->reada_lock);
285 ret = radix_tree_insert(&dev->reada_zones,
286 (unsigned long)(zone->end >> fs_info->sectorsize_bits),
289 if (ret == -EEXIST) {
291 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
292 logical >> fs_info->sectorsize_bits, 1);
293 if (ret == 1 && logical >= zone->start && logical <= zone->end)
294 kref_get(&zone->refcnt);
298 spin_unlock(&fs_info->reada_lock);
299 radix_tree_preload_end();
304 static struct reada_extent *reada_find_extent(struct btrfs_fs_info *fs_info,
306 struct btrfs_key *top,
307 u64 owner_root, int level)
310 struct reada_extent *re = NULL;
311 struct reada_extent *re_exist = NULL;
312 struct btrfs_io_context *bioc = NULL;
313 struct btrfs_device *dev;
314 struct btrfs_device *prev_dev;
318 unsigned long index = logical >> fs_info->sectorsize_bits;
319 int dev_replace_is_ongoing;
322 spin_lock(&fs_info->reada_lock);
323 re = radix_tree_lookup(&fs_info->reada_tree, index);
326 spin_unlock(&fs_info->reada_lock);
331 re = kzalloc(sizeof(*re), GFP_KERNEL);
335 re->logical = logical;
337 INIT_LIST_HEAD(&re->extctl);
338 spin_lock_init(&re->lock);
340 re->owner_root = owner_root;
346 length = fs_info->nodesize;
347 ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
349 if (ret || !bioc || length < fs_info->nodesize)
352 if (bioc->num_stripes > BTRFS_MAX_MIRRORS) {
354 "readahead: more than %d copies not supported",
359 real_stripes = bioc->num_stripes - bioc->num_tgtdevs;
360 for (nzones = 0; nzones < real_stripes; ++nzones) {
361 struct reada_zone *zone;
363 dev = bioc->stripes[nzones].dev;
365 /* cannot read ahead on missing device. */
369 zone = reada_find_zone(dev, logical, bioc);
373 re->zones[re->nzones++] = zone;
374 spin_lock(&zone->lock);
376 kref_get(&zone->refcnt);
378 spin_unlock(&zone->lock);
379 spin_lock(&fs_info->reada_lock);
380 kref_put(&zone->refcnt, reada_zone_release);
381 spin_unlock(&fs_info->reada_lock);
383 if (re->nzones == 0) {
384 /* not a single zone found, error and out */
388 /* Insert extent in reada tree + all per-device trees, all or nothing */
389 down_read(&fs_info->dev_replace.rwsem);
390 ret = radix_tree_preload(GFP_KERNEL);
392 up_read(&fs_info->dev_replace.rwsem);
396 spin_lock(&fs_info->reada_lock);
397 ret = radix_tree_insert(&fs_info->reada_tree, index, re);
398 if (ret == -EEXIST) {
399 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
401 spin_unlock(&fs_info->reada_lock);
402 radix_tree_preload_end();
403 up_read(&fs_info->dev_replace.rwsem);
407 spin_unlock(&fs_info->reada_lock);
408 radix_tree_preload_end();
409 up_read(&fs_info->dev_replace.rwsem);
412 radix_tree_preload_end();
414 dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
415 &fs_info->dev_replace);
416 for (nzones = 0; nzones < re->nzones; ++nzones) {
417 dev = re->zones[nzones]->device;
419 if (dev == prev_dev) {
421 * in case of DUP, just add the first zone. As both
422 * are on the same device, there's nothing to gain
424 * Also, it wouldn't work, as the tree is per device
425 * and adding would fail with EEXIST
432 if (test_bit(BTRFS_DEV_STATE_NO_READA, &dev->dev_state))
435 if (dev_replace_is_ongoing &&
436 dev == fs_info->dev_replace.tgtdev) {
438 * as this device is selected for reading only as
439 * a last resort, skip it for read ahead.
444 ret = radix_tree_insert(&dev->reada_extents, index, re);
446 while (--nzones >= 0) {
447 dev = re->zones[nzones]->device;
449 /* ignore whether the entry was inserted */
450 radix_tree_delete(&dev->reada_extents, index);
452 radix_tree_delete(&fs_info->reada_tree, index);
453 spin_unlock(&fs_info->reada_lock);
454 up_read(&fs_info->dev_replace.rwsem);
460 radix_tree_delete(&fs_info->reada_tree, index);
461 spin_unlock(&fs_info->reada_lock);
462 up_read(&fs_info->dev_replace.rwsem);
467 btrfs_put_bioc(bioc);
471 for (nzones = 0; nzones < re->nzones; ++nzones) {
472 struct reada_zone *zone;
474 zone = re->zones[nzones];
475 kref_get(&zone->refcnt);
476 spin_lock(&zone->lock);
478 if (zone->elems == 0) {
480 * no fs_info->reada_lock needed, as this can't be
483 kref_put(&zone->refcnt, reada_zone_release);
485 spin_unlock(&zone->lock);
487 spin_lock(&fs_info->reada_lock);
488 kref_put(&zone->refcnt, reada_zone_release);
489 spin_unlock(&fs_info->reada_lock);
491 btrfs_put_bioc(bioc);
496 static void reada_extent_put(struct btrfs_fs_info *fs_info,
497 struct reada_extent *re)
500 unsigned long index = re->logical >> fs_info->sectorsize_bits;
502 spin_lock(&fs_info->reada_lock);
504 spin_unlock(&fs_info->reada_lock);
508 radix_tree_delete(&fs_info->reada_tree, index);
509 for (i = 0; i < re->nzones; ++i) {
510 struct reada_zone *zone = re->zones[i];
512 radix_tree_delete(&zone->device->reada_extents, index);
515 spin_unlock(&fs_info->reada_lock);
517 for (i = 0; i < re->nzones; ++i) {
518 struct reada_zone *zone = re->zones[i];
520 kref_get(&zone->refcnt);
521 spin_lock(&zone->lock);
523 if (zone->elems == 0) {
524 /* no fs_info->reada_lock needed, as this can't be
526 kref_put(&zone->refcnt, reada_zone_release);
528 spin_unlock(&zone->lock);
530 spin_lock(&fs_info->reada_lock);
531 kref_put(&zone->refcnt, reada_zone_release);
532 spin_unlock(&fs_info->reada_lock);
538 static void reada_zone_release(struct kref *kref)
540 struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
541 struct btrfs_fs_info *fs_info = zone->device->fs_info;
543 lockdep_assert_held(&fs_info->reada_lock);
545 radix_tree_delete(&zone->device->reada_zones,
546 zone->end >> fs_info->sectorsize_bits);
551 static void reada_control_release(struct kref *kref)
553 struct reada_control *rc = container_of(kref, struct reada_control,
559 static int reada_add_block(struct reada_control *rc, u64 logical,
560 struct btrfs_key *top, u64 owner_root,
561 u64 generation, int level)
563 struct btrfs_fs_info *fs_info = rc->fs_info;
564 struct reada_extent *re;
565 struct reada_extctl *rec;
568 re = reada_find_extent(fs_info, logical, top, owner_root, level);
572 rec = kzalloc(sizeof(*rec), GFP_KERNEL);
574 reada_extent_put(fs_info, re);
579 rec->generation = generation;
580 atomic_inc(&rc->elems);
582 spin_lock(&re->lock);
583 list_add_tail(&rec->list, &re->extctl);
584 spin_unlock(&re->lock);
586 /* leave the ref on the extent */
592 * called with fs_info->reada_lock held
594 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
597 unsigned long index = zone->end >> zone->device->fs_info->sectorsize_bits;
599 for (i = 0; i < zone->ndevs; ++i) {
600 struct reada_zone *peer;
601 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
602 if (peer && peer->device != zone->device)
608 * called with fs_info->reada_lock held
610 static int reada_pick_zone(struct btrfs_device *dev)
612 struct reada_zone *top_zone = NULL;
613 struct reada_zone *top_locked_zone = NULL;
615 u64 top_locked_elems = 0;
616 unsigned long index = 0;
619 if (dev->reada_curr_zone) {
620 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
621 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
622 dev->reada_curr_zone = NULL;
624 /* pick the zone with the most elements */
626 struct reada_zone *zone;
628 ret = radix_tree_gang_lookup(&dev->reada_zones,
629 (void **)&zone, index, 1);
632 index = (zone->end >> dev->fs_info->sectorsize_bits) + 1;
634 if (zone->elems > top_locked_elems) {
635 top_locked_elems = zone->elems;
636 top_locked_zone = zone;
639 if (zone->elems > top_elems) {
640 top_elems = zone->elems;
646 dev->reada_curr_zone = top_zone;
647 else if (top_locked_zone)
648 dev->reada_curr_zone = top_locked_zone;
652 dev->reada_next = dev->reada_curr_zone->start;
653 kref_get(&dev->reada_curr_zone->refcnt);
654 reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
659 static int reada_tree_block_flagged(struct btrfs_fs_info *fs_info, u64 bytenr,
660 u64 owner_root, int level, int mirror_num,
661 struct extent_buffer **eb)
663 struct extent_buffer *buf = NULL;
666 buf = btrfs_find_create_tree_block(fs_info, bytenr, owner_root, level);
670 set_bit(EXTENT_BUFFER_READAHEAD, &buf->bflags);
672 ret = read_extent_buffer_pages(buf, WAIT_PAGE_LOCK, mirror_num);
674 free_extent_buffer_stale(buf);
678 if (test_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags)) {
679 free_extent_buffer_stale(buf);
681 } else if (extent_buffer_uptodate(buf)) {
684 free_extent_buffer(buf);
689 static int reada_start_machine_dev(struct btrfs_device *dev)
691 struct btrfs_fs_info *fs_info = dev->fs_info;
692 struct reada_extent *re = NULL;
694 struct extent_buffer *eb = NULL;
699 spin_lock(&fs_info->reada_lock);
700 if (dev->reada_curr_zone == NULL) {
701 ret = reada_pick_zone(dev);
703 spin_unlock(&fs_info->reada_lock);
708 * FIXME currently we issue the reads one extent at a time. If we have
709 * a contiguous block of extents, we could also coagulate them or use
710 * plugging to speed things up
712 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
713 dev->reada_next >> fs_info->sectorsize_bits, 1);
714 if (ret == 0 || re->logical > dev->reada_curr_zone->end) {
715 ret = reada_pick_zone(dev);
717 spin_unlock(&fs_info->reada_lock);
721 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
722 dev->reada_next >> fs_info->sectorsize_bits, 1);
725 spin_unlock(&fs_info->reada_lock);
728 dev->reada_next = re->logical + fs_info->nodesize;
731 spin_unlock(&fs_info->reada_lock);
733 spin_lock(&re->lock);
734 if (re->scheduled || list_empty(&re->extctl)) {
735 spin_unlock(&re->lock);
736 reada_extent_put(fs_info, re);
740 spin_unlock(&re->lock);
745 for (i = 0; i < re->nzones; ++i) {
746 if (re->zones[i]->device == dev) {
751 logical = re->logical;
753 atomic_inc(&dev->reada_in_flight);
754 ret = reada_tree_block_flagged(fs_info, logical, re->owner_root,
755 re->level, mirror_num, &eb);
757 __readahead_hook(fs_info, re, NULL, ret);
759 __readahead_hook(fs_info, re, eb, ret);
762 free_extent_buffer(eb);
764 atomic_dec(&dev->reada_in_flight);
765 reada_extent_put(fs_info, re);
771 static void reada_start_machine_worker(struct btrfs_work *work)
773 struct reada_machine_work *rmw;
776 rmw = container_of(work, struct reada_machine_work, work);
778 old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
779 task_nice_ioprio(current));
780 set_task_ioprio(current, BTRFS_IOPRIO_READA);
781 __reada_start_machine(rmw->fs_info);
782 set_task_ioprio(current, old_ioprio);
784 atomic_dec(&rmw->fs_info->reada_works_cnt);
789 /* Try to start up to 10k READA requests for a group of devices */
790 static int reada_start_for_fsdevs(struct btrfs_fs_devices *fs_devices)
794 struct btrfs_device *device;
798 list_for_each_entry(device, &fs_devices->devices, dev_list) {
799 if (atomic_read(&device->reada_in_flight) <
801 enqueued += reada_start_machine_dev(device);
804 } while (enqueued && total < 10000);
809 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
811 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices, *seed_devs;
815 mutex_lock(&fs_devices->device_list_mutex);
817 enqueued += reada_start_for_fsdevs(fs_devices);
818 list_for_each_entry(seed_devs, &fs_devices->seed_list, seed_list)
819 enqueued += reada_start_for_fsdevs(seed_devs);
821 mutex_unlock(&fs_devices->device_list_mutex);
826 * If everything is already in the cache, this is effectively single
827 * threaded. To a) not hold the caller for too long and b) to utilize
828 * more cores, we broke the loop above after 10000 iterations and now
829 * enqueue to workers to finish it. This will distribute the load to
832 for (i = 0; i < 2; ++i) {
833 reada_start_machine(fs_info);
834 if (atomic_read(&fs_info->reada_works_cnt) >
835 BTRFS_MAX_MIRRORS * 2)
840 static void reada_start_machine(struct btrfs_fs_info *fs_info)
842 struct reada_machine_work *rmw;
844 rmw = kzalloc(sizeof(*rmw), GFP_KERNEL);
846 /* FIXME we cannot handle this properly right now */
849 btrfs_init_work(&rmw->work, reada_start_machine_worker, NULL, NULL);
850 rmw->fs_info = fs_info;
852 btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
853 atomic_inc(&fs_info->reada_works_cnt);
857 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
859 struct btrfs_device *device;
860 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
867 spin_lock(&fs_info->reada_lock);
868 list_for_each_entry(device, &fs_devices->devices, dev_list) {
869 btrfs_debug(fs_info, "dev %lld has %d in flight", device->devid,
870 atomic_read(&device->reada_in_flight));
873 struct reada_zone *zone;
874 ret = radix_tree_gang_lookup(&device->reada_zones,
875 (void **)&zone, index, 1);
878 pr_debug(" zone %llu-%llu elems %llu locked %d devs",
879 zone->start, zone->end, zone->elems,
881 for (j = 0; j < zone->ndevs; ++j) {
883 zone->devs[j]->devid);
885 if (device->reada_curr_zone == zone)
886 pr_cont(" curr off %llu",
887 device->reada_next - zone->start);
889 index = (zone->end >> fs_info->sectorsize_bits) + 1;
894 struct reada_extent *re = NULL;
896 ret = radix_tree_gang_lookup(&device->reada_extents,
897 (void **)&re, index, 1);
900 pr_debug(" re: logical %llu size %u empty %d scheduled %d",
901 re->logical, fs_info->nodesize,
902 list_empty(&re->extctl), re->scheduled);
904 for (i = 0; i < re->nzones; ++i) {
905 pr_cont(" zone %llu-%llu devs",
908 for (j = 0; j < re->zones[i]->ndevs; ++j) {
910 re->zones[i]->devs[j]->devid);
914 index = (re->logical >> fs_info->sectorsize_bits) + 1;
923 struct reada_extent *re = NULL;
925 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
929 if (!re->scheduled) {
930 index = (re->logical >> fs_info->sectorsize_bits) + 1;
933 pr_debug("re: logical %llu size %u list empty %d scheduled %d",
934 re->logical, fs_info->nodesize,
935 list_empty(&re->extctl), re->scheduled);
936 for (i = 0; i < re->nzones; ++i) {
937 pr_cont(" zone %llu-%llu devs",
940 for (j = 0; j < re->zones[i]->ndevs; ++j) {
942 re->zones[i]->devs[j]->devid);
946 index = (re->logical >> fs_info->sectorsize_bits) + 1;
948 spin_unlock(&fs_info->reada_lock);
955 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
956 struct btrfs_key *key_start, struct btrfs_key *key_end)
958 struct reada_control *rc;
963 struct extent_buffer *node;
964 static struct btrfs_key max_key = {
970 rc = kzalloc(sizeof(*rc), GFP_KERNEL);
972 return ERR_PTR(-ENOMEM);
974 rc->fs_info = root->fs_info;
975 rc->key_start = *key_start;
976 rc->key_end = *key_end;
977 atomic_set(&rc->elems, 0);
978 init_waitqueue_head(&rc->wait);
979 kref_init(&rc->refcnt);
980 kref_get(&rc->refcnt); /* one ref for having elements */
982 node = btrfs_root_node(root);
984 generation = btrfs_header_generation(node);
985 level = btrfs_header_level(node);
986 free_extent_buffer(node);
988 ret = reada_add_block(rc, start, &max_key, root->root_key.objectid,
995 reada_start_machine(root->fs_info);
1001 int btrfs_reada_wait(void *handle)
1003 struct reada_control *rc = handle;
1004 struct btrfs_fs_info *fs_info = rc->fs_info;
1006 while (atomic_read(&rc->elems)) {
1007 if (!atomic_read(&fs_info->reada_works_cnt))
1008 reada_start_machine(fs_info);
1009 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
1011 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
1014 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
1016 kref_put(&rc->refcnt, reada_control_release);
1021 int btrfs_reada_wait(void *handle)
1023 struct reada_control *rc = handle;
1024 struct btrfs_fs_info *fs_info = rc->fs_info;
1026 while (atomic_read(&rc->elems)) {
1027 if (!atomic_read(&fs_info->reada_works_cnt))
1028 reada_start_machine(fs_info);
1029 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
1033 kref_put(&rc->refcnt, reada_control_release);
1039 void btrfs_reada_detach(void *handle)
1041 struct reada_control *rc = handle;
1043 kref_put(&rc->refcnt, reada_control_release);
1047 * Before removing a device (device replace or device remove ioctls), call this
1048 * function to wait for all existing readahead requests on the device and to
1049 * make sure no one queues more readahead requests for the device.
1051 * Must be called without holding neither the device list mutex nor the device
1052 * replace semaphore, otherwise it will deadlock.
1054 void btrfs_reada_remove_dev(struct btrfs_device *dev)
1056 struct btrfs_fs_info *fs_info = dev->fs_info;
1058 /* Serialize with readahead extent creation at reada_find_extent(). */
1059 spin_lock(&fs_info->reada_lock);
1060 set_bit(BTRFS_DEV_STATE_NO_READA, &dev->dev_state);
1061 spin_unlock(&fs_info->reada_lock);
1064 * There might be readahead requests added to the radix trees which
1065 * were not yet added to the readahead work queue. We need to start
1066 * them and wait for their completion, otherwise we can end up with
1067 * use-after-free problems when dropping the last reference on the
1068 * readahead extents and their zones, as they need to access the
1071 reada_start_machine(fs_info);
1072 btrfs_flush_workqueue(fs_info->readahead_workers);
1076 * If when removing a device (device replace or device remove ioctls) an error
1077 * happens after calling btrfs_reada_remove_dev(), call this to undo what that
1078 * function did. This is safe to call even if btrfs_reada_remove_dev() was not
1081 void btrfs_reada_undo_remove_dev(struct btrfs_device *dev)
1083 spin_lock(&dev->fs_info->reada_lock);
1084 clear_bit(BTRFS_DEV_STATE_NO_READA, &dev->dev_state);
1085 spin_unlock(&dev->fs_info->reada_lock);