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"
21 * This is the implementation for the generic read ahead framework.
23 * To trigger a readahead, btrfs_reada_add must be called. It will start
24 * a read ahead for the given range [start, end) on tree root. The returned
25 * handle can either be used to wait on the readahead to finish
26 * (btrfs_reada_wait), or to send it to the background (btrfs_reada_detach).
28 * The read ahead works as follows:
29 * On btrfs_reada_add, the root of the tree is inserted into a radix_tree.
30 * reada_start_machine will then search for extents to prefetch and trigger
31 * some reads. When a read finishes for a node, all contained node/leaf
32 * pointers that lie in the given range will also be enqueued. The reads will
33 * be triggered in sequential order, thus giving a big win over a naive
34 * enumeration. It will also make use of multi-device layouts. Each disk
35 * will have its on read pointer and all disks will by utilized in parallel.
36 * Also will no two disks read both sides of a mirror simultaneously, as this
37 * would waste seeking capacity. Instead both disks will read different parts
39 * Any number of readaheads can be started in parallel. The read order will be
40 * determined globally, i.e. 2 parallel readaheads will normally finish faster
41 * than the 2 started one after another.
44 #define MAX_IN_FLIGHT 6
47 struct list_head list;
48 struct reada_control *rc;
55 struct list_head extctl;
58 struct reada_zone *zones[BTRFS_MAX_MIRRORS];
67 struct list_head list;
70 struct btrfs_device *device;
71 struct btrfs_device *devs[BTRFS_MAX_MIRRORS]; /* full list, incl
77 struct reada_machine_work {
78 struct btrfs_work work;
79 struct btrfs_fs_info *fs_info;
82 static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
83 static void reada_control_release(struct kref *kref);
84 static void reada_zone_release(struct kref *kref);
85 static void reada_start_machine(struct btrfs_fs_info *fs_info);
86 static void __reada_start_machine(struct btrfs_fs_info *fs_info);
88 static int reada_add_block(struct reada_control *rc, u64 logical,
89 struct btrfs_key *top, u64 generation);
92 /* in case of err, eb might be NULL */
93 static void __readahead_hook(struct btrfs_fs_info *fs_info,
94 struct reada_extent *re, struct extent_buffer *eb,
101 struct list_head list;
103 spin_lock(&re->lock);
105 * just take the full list from the extent. afterwards we
106 * don't need the lock anymore
108 list_replace_init(&re->extctl, &list);
110 spin_unlock(&re->lock);
113 * this is the error case, the extent buffer has not been
114 * read correctly. We won't access anything from it and
115 * just cleanup our data structures. Effectively this will
116 * cut the branch below this node from read ahead.
122 * FIXME: currently we just set nritems to 0 if this is a leaf,
123 * effectively ignoring the content. In a next step we could
124 * trigger more readahead depending from the content, e.g.
125 * fetch the checksums for the extents in the leaf.
127 if (!btrfs_header_level(eb))
130 nritems = btrfs_header_nritems(eb);
131 generation = btrfs_header_generation(eb);
132 for (i = 0; i < nritems; i++) {
133 struct reada_extctl *rec;
135 struct btrfs_key key;
136 struct btrfs_key next_key;
138 btrfs_node_key_to_cpu(eb, &key, i);
140 btrfs_node_key_to_cpu(eb, &next_key, i + 1);
143 bytenr = btrfs_node_blockptr(eb, i);
144 n_gen = btrfs_node_ptr_generation(eb, i);
146 list_for_each_entry(rec, &list, list) {
147 struct reada_control *rc = rec->rc;
150 * if the generation doesn't match, just ignore this
151 * extctl. This will probably cut off a branch from
152 * prefetch. Alternatively one could start a new (sub-)
153 * prefetch for this branch, starting again from root.
154 * FIXME: move the generation check out of this loop
157 if (rec->generation != generation) {
159 "generation mismatch for (%llu,%d,%llu) %llu != %llu",
160 key.objectid, key.type, key.offset,
161 rec->generation, generation);
164 if (rec->generation == generation &&
165 btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
166 btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
167 reada_add_block(rc, bytenr, &next_key, n_gen);
173 * free extctl records
175 while (!list_empty(&list)) {
176 struct reada_control *rc;
177 struct reada_extctl *rec;
179 rec = list_first_entry(&list, struct reada_extctl, list);
180 list_del(&rec->list);
184 kref_get(&rc->refcnt);
185 if (atomic_dec_and_test(&rc->elems)) {
186 kref_put(&rc->refcnt, reada_control_release);
189 kref_put(&rc->refcnt, reada_control_release);
191 reada_extent_put(fs_info, re); /* one ref for each entry */
197 int btree_readahead_hook(struct extent_buffer *eb, int err)
199 struct btrfs_fs_info *fs_info = eb->fs_info;
201 struct reada_extent *re;
204 spin_lock(&fs_info->reada_lock);
205 re = radix_tree_lookup(&fs_info->reada_tree,
206 eb->start >> PAGE_SHIFT);
209 spin_unlock(&fs_info->reada_lock);
215 __readahead_hook(fs_info, re, eb, err);
216 reada_extent_put(fs_info, re); /* our ref */
219 reada_start_machine(fs_info);
223 static struct reada_zone *reada_find_zone(struct btrfs_device *dev, u64 logical,
224 struct btrfs_bio *bbio)
226 struct btrfs_fs_info *fs_info = dev->fs_info;
228 struct reada_zone *zone;
229 struct btrfs_block_group_cache *cache = NULL;
235 spin_lock(&fs_info->reada_lock);
236 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
237 logical >> PAGE_SHIFT, 1);
238 if (ret == 1 && logical >= zone->start && logical <= zone->end) {
239 kref_get(&zone->refcnt);
240 spin_unlock(&fs_info->reada_lock);
244 spin_unlock(&fs_info->reada_lock);
246 cache = btrfs_lookup_block_group(fs_info, logical);
250 start = cache->key.objectid;
251 end = start + cache->key.offset - 1;
252 btrfs_put_block_group(cache);
254 zone = kzalloc(sizeof(*zone), GFP_KERNEL);
258 ret = radix_tree_preload(GFP_KERNEL);
266 INIT_LIST_HEAD(&zone->list);
267 spin_lock_init(&zone->lock);
269 kref_init(&zone->refcnt);
271 zone->device = dev; /* our device always sits at index 0 */
272 for (i = 0; i < bbio->num_stripes; ++i) {
273 /* bounds have already been checked */
274 zone->devs[i] = bbio->stripes[i].dev;
276 zone->ndevs = bbio->num_stripes;
278 spin_lock(&fs_info->reada_lock);
279 ret = radix_tree_insert(&dev->reada_zones,
280 (unsigned long)(zone->end >> PAGE_SHIFT),
283 if (ret == -EEXIST) {
285 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
286 logical >> PAGE_SHIFT, 1);
287 if (ret == 1 && logical >= zone->start && logical <= zone->end)
288 kref_get(&zone->refcnt);
292 spin_unlock(&fs_info->reada_lock);
293 radix_tree_preload_end();
298 static struct reada_extent *reada_find_extent(struct btrfs_fs_info *fs_info,
300 struct btrfs_key *top)
303 struct reada_extent *re = NULL;
304 struct reada_extent *re_exist = NULL;
305 struct btrfs_bio *bbio = NULL;
306 struct btrfs_device *dev;
307 struct btrfs_device *prev_dev;
311 unsigned long index = logical >> PAGE_SHIFT;
312 int dev_replace_is_ongoing;
315 spin_lock(&fs_info->reada_lock);
316 re = radix_tree_lookup(&fs_info->reada_tree, index);
319 spin_unlock(&fs_info->reada_lock);
324 re = kzalloc(sizeof(*re), GFP_KERNEL);
328 re->logical = logical;
330 INIT_LIST_HEAD(&re->extctl);
331 spin_lock_init(&re->lock);
337 length = fs_info->nodesize;
338 ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
340 if (ret || !bbio || length < fs_info->nodesize)
343 if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
345 "readahead: more than %d copies not supported",
350 real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
351 for (nzones = 0; nzones < real_stripes; ++nzones) {
352 struct reada_zone *zone;
354 dev = bbio->stripes[nzones].dev;
356 /* cannot read ahead on missing device. */
360 zone = reada_find_zone(dev, logical, bbio);
364 re->zones[re->nzones++] = zone;
365 spin_lock(&zone->lock);
367 kref_get(&zone->refcnt);
369 spin_unlock(&zone->lock);
370 spin_lock(&fs_info->reada_lock);
371 kref_put(&zone->refcnt, reada_zone_release);
372 spin_unlock(&fs_info->reada_lock);
374 if (re->nzones == 0) {
375 /* not a single zone found, error and out */
379 /* Insert extent in reada tree + all per-device trees, all or nothing */
380 down_read(&fs_info->dev_replace.rwsem);
381 ret = radix_tree_preload(GFP_KERNEL);
383 up_read(&fs_info->dev_replace.rwsem);
387 spin_lock(&fs_info->reada_lock);
388 ret = radix_tree_insert(&fs_info->reada_tree, index, re);
389 if (ret == -EEXIST) {
390 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
392 spin_unlock(&fs_info->reada_lock);
393 radix_tree_preload_end();
394 up_read(&fs_info->dev_replace.rwsem);
398 spin_unlock(&fs_info->reada_lock);
399 radix_tree_preload_end();
400 up_read(&fs_info->dev_replace.rwsem);
403 radix_tree_preload_end();
405 dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
406 &fs_info->dev_replace);
407 for (nzones = 0; nzones < re->nzones; ++nzones) {
408 dev = re->zones[nzones]->device;
410 if (dev == prev_dev) {
412 * in case of DUP, just add the first zone. As both
413 * are on the same device, there's nothing to gain
415 * Also, it wouldn't work, as the tree is per device
416 * and adding would fail with EEXIST
423 if (dev_replace_is_ongoing &&
424 dev == fs_info->dev_replace.tgtdev) {
426 * as this device is selected for reading only as
427 * a last resort, skip it for read ahead.
432 ret = radix_tree_insert(&dev->reada_extents, index, re);
434 while (--nzones >= 0) {
435 dev = re->zones[nzones]->device;
437 /* ignore whether the entry was inserted */
438 radix_tree_delete(&dev->reada_extents, index);
440 radix_tree_delete(&fs_info->reada_tree, index);
441 spin_unlock(&fs_info->reada_lock);
442 up_read(&fs_info->dev_replace.rwsem);
447 spin_unlock(&fs_info->reada_lock);
448 up_read(&fs_info->dev_replace.rwsem);
453 btrfs_put_bbio(bbio);
457 for (nzones = 0; nzones < re->nzones; ++nzones) {
458 struct reada_zone *zone;
460 zone = re->zones[nzones];
461 kref_get(&zone->refcnt);
462 spin_lock(&zone->lock);
464 if (zone->elems == 0) {
466 * no fs_info->reada_lock needed, as this can't be
469 kref_put(&zone->refcnt, reada_zone_release);
471 spin_unlock(&zone->lock);
473 spin_lock(&fs_info->reada_lock);
474 kref_put(&zone->refcnt, reada_zone_release);
475 spin_unlock(&fs_info->reada_lock);
477 btrfs_put_bbio(bbio);
482 static void reada_extent_put(struct btrfs_fs_info *fs_info,
483 struct reada_extent *re)
486 unsigned long index = re->logical >> PAGE_SHIFT;
488 spin_lock(&fs_info->reada_lock);
490 spin_unlock(&fs_info->reada_lock);
494 radix_tree_delete(&fs_info->reada_tree, index);
495 for (i = 0; i < re->nzones; ++i) {
496 struct reada_zone *zone = re->zones[i];
498 radix_tree_delete(&zone->device->reada_extents, index);
501 spin_unlock(&fs_info->reada_lock);
503 for (i = 0; i < re->nzones; ++i) {
504 struct reada_zone *zone = re->zones[i];
506 kref_get(&zone->refcnt);
507 spin_lock(&zone->lock);
509 if (zone->elems == 0) {
510 /* no fs_info->reada_lock needed, as this can't be
512 kref_put(&zone->refcnt, reada_zone_release);
514 spin_unlock(&zone->lock);
516 spin_lock(&fs_info->reada_lock);
517 kref_put(&zone->refcnt, reada_zone_release);
518 spin_unlock(&fs_info->reada_lock);
524 static void reada_zone_release(struct kref *kref)
526 struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
528 radix_tree_delete(&zone->device->reada_zones,
529 zone->end >> PAGE_SHIFT);
534 static void reada_control_release(struct kref *kref)
536 struct reada_control *rc = container_of(kref, struct reada_control,
542 static int reada_add_block(struct reada_control *rc, u64 logical,
543 struct btrfs_key *top, u64 generation)
545 struct btrfs_fs_info *fs_info = rc->fs_info;
546 struct reada_extent *re;
547 struct reada_extctl *rec;
550 re = reada_find_extent(fs_info, logical, top);
554 rec = kzalloc(sizeof(*rec), GFP_KERNEL);
556 reada_extent_put(fs_info, re);
561 rec->generation = generation;
562 atomic_inc(&rc->elems);
564 spin_lock(&re->lock);
565 list_add_tail(&rec->list, &re->extctl);
566 spin_unlock(&re->lock);
568 /* leave the ref on the extent */
574 * called with fs_info->reada_lock held
576 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
579 unsigned long index = zone->end >> PAGE_SHIFT;
581 for (i = 0; i < zone->ndevs; ++i) {
582 struct reada_zone *peer;
583 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
584 if (peer && peer->device != zone->device)
590 * called with fs_info->reada_lock held
592 static int reada_pick_zone(struct btrfs_device *dev)
594 struct reada_zone *top_zone = NULL;
595 struct reada_zone *top_locked_zone = NULL;
597 u64 top_locked_elems = 0;
598 unsigned long index = 0;
601 if (dev->reada_curr_zone) {
602 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
603 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
604 dev->reada_curr_zone = NULL;
606 /* pick the zone with the most elements */
608 struct reada_zone *zone;
610 ret = radix_tree_gang_lookup(&dev->reada_zones,
611 (void **)&zone, index, 1);
614 index = (zone->end >> PAGE_SHIFT) + 1;
616 if (zone->elems > top_locked_elems) {
617 top_locked_elems = zone->elems;
618 top_locked_zone = zone;
621 if (zone->elems > top_elems) {
622 top_elems = zone->elems;
628 dev->reada_curr_zone = top_zone;
629 else if (top_locked_zone)
630 dev->reada_curr_zone = top_locked_zone;
634 dev->reada_next = dev->reada_curr_zone->start;
635 kref_get(&dev->reada_curr_zone->refcnt);
636 reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
641 static int reada_start_machine_dev(struct btrfs_device *dev)
643 struct btrfs_fs_info *fs_info = dev->fs_info;
644 struct reada_extent *re = NULL;
646 struct extent_buffer *eb = NULL;
651 spin_lock(&fs_info->reada_lock);
652 if (dev->reada_curr_zone == NULL) {
653 ret = reada_pick_zone(dev);
655 spin_unlock(&fs_info->reada_lock);
660 * FIXME currently we issue the reads one extent at a time. If we have
661 * a contiguous block of extents, we could also coagulate them or use
662 * plugging to speed things up
664 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
665 dev->reada_next >> PAGE_SHIFT, 1);
666 if (ret == 0 || re->logical > dev->reada_curr_zone->end) {
667 ret = reada_pick_zone(dev);
669 spin_unlock(&fs_info->reada_lock);
673 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
674 dev->reada_next >> PAGE_SHIFT, 1);
677 spin_unlock(&fs_info->reada_lock);
680 dev->reada_next = re->logical + fs_info->nodesize;
683 spin_unlock(&fs_info->reada_lock);
685 spin_lock(&re->lock);
686 if (re->scheduled || list_empty(&re->extctl)) {
687 spin_unlock(&re->lock);
688 reada_extent_put(fs_info, re);
692 spin_unlock(&re->lock);
697 for (i = 0; i < re->nzones; ++i) {
698 if (re->zones[i]->device == dev) {
703 logical = re->logical;
705 atomic_inc(&dev->reada_in_flight);
706 ret = reada_tree_block_flagged(fs_info, logical, mirror_num, &eb);
708 __readahead_hook(fs_info, re, NULL, ret);
710 __readahead_hook(fs_info, re, eb, ret);
713 free_extent_buffer(eb);
715 atomic_dec(&dev->reada_in_flight);
716 reada_extent_put(fs_info, re);
722 static void reada_start_machine_worker(struct btrfs_work *work)
724 struct reada_machine_work *rmw;
725 struct btrfs_fs_info *fs_info;
728 rmw = container_of(work, struct reada_machine_work, work);
729 fs_info = rmw->fs_info;
733 old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
734 task_nice_ioprio(current));
735 set_task_ioprio(current, BTRFS_IOPRIO_READA);
736 __reada_start_machine(fs_info);
737 set_task_ioprio(current, old_ioprio);
739 atomic_dec(&fs_info->reada_works_cnt);
742 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
744 struct btrfs_device *device;
745 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
752 mutex_lock(&fs_devices->device_list_mutex);
753 list_for_each_entry(device, &fs_devices->devices, dev_list) {
754 if (atomic_read(&device->reada_in_flight) <
756 enqueued += reada_start_machine_dev(device);
758 mutex_unlock(&fs_devices->device_list_mutex);
760 } while (enqueued && total < 10000);
766 * If everything is already in the cache, this is effectively single
767 * threaded. To a) not hold the caller for too long and b) to utilize
768 * more cores, we broke the loop above after 10000 iterations and now
769 * enqueue to workers to finish it. This will distribute the load to
772 for (i = 0; i < 2; ++i) {
773 reada_start_machine(fs_info);
774 if (atomic_read(&fs_info->reada_works_cnt) >
775 BTRFS_MAX_MIRRORS * 2)
780 static void reada_start_machine(struct btrfs_fs_info *fs_info)
782 struct reada_machine_work *rmw;
784 rmw = kzalloc(sizeof(*rmw), GFP_KERNEL);
786 /* FIXME we cannot handle this properly right now */
789 btrfs_init_work(&rmw->work, btrfs_readahead_helper,
790 reada_start_machine_worker, NULL, NULL);
791 rmw->fs_info = fs_info;
793 btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
794 atomic_inc(&fs_info->reada_works_cnt);
798 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
800 struct btrfs_device *device;
801 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
808 spin_lock(&fs_info->reada_lock);
809 list_for_each_entry(device, &fs_devices->devices, dev_list) {
810 btrfs_debug(fs_info, "dev %lld has %d in flight", device->devid,
811 atomic_read(&device->reada_in_flight));
814 struct reada_zone *zone;
815 ret = radix_tree_gang_lookup(&device->reada_zones,
816 (void **)&zone, index, 1);
819 pr_debug(" zone %llu-%llu elems %llu locked %d devs",
820 zone->start, zone->end, zone->elems,
822 for (j = 0; j < zone->ndevs; ++j) {
824 zone->devs[j]->devid);
826 if (device->reada_curr_zone == zone)
827 pr_cont(" curr off %llu",
828 device->reada_next - zone->start);
830 index = (zone->end >> PAGE_SHIFT) + 1;
835 struct reada_extent *re = NULL;
837 ret = radix_tree_gang_lookup(&device->reada_extents,
838 (void **)&re, index, 1);
841 pr_debug(" re: logical %llu size %u empty %d scheduled %d",
842 re->logical, fs_info->nodesize,
843 list_empty(&re->extctl), re->scheduled);
845 for (i = 0; i < re->nzones; ++i) {
846 pr_cont(" zone %llu-%llu devs",
849 for (j = 0; j < re->zones[i]->ndevs; ++j) {
851 re->zones[i]->devs[j]->devid);
855 index = (re->logical >> PAGE_SHIFT) + 1;
864 struct reada_extent *re = NULL;
866 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
870 if (!re->scheduled) {
871 index = (re->logical >> PAGE_SHIFT) + 1;
874 pr_debug("re: logical %llu size %u list empty %d scheduled %d",
875 re->logical, fs_info->nodesize,
876 list_empty(&re->extctl), re->scheduled);
877 for (i = 0; i < re->nzones; ++i) {
878 pr_cont(" zone %llu-%llu devs",
881 for (j = 0; j < re->zones[i]->ndevs; ++j) {
883 re->zones[i]->devs[j]->devid);
887 index = (re->logical >> PAGE_SHIFT) + 1;
889 spin_unlock(&fs_info->reada_lock);
896 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
897 struct btrfs_key *key_start, struct btrfs_key *key_end)
899 struct reada_control *rc;
903 struct extent_buffer *node;
904 static struct btrfs_key max_key = {
910 rc = kzalloc(sizeof(*rc), GFP_KERNEL);
912 return ERR_PTR(-ENOMEM);
914 rc->fs_info = root->fs_info;
915 rc->key_start = *key_start;
916 rc->key_end = *key_end;
917 atomic_set(&rc->elems, 0);
918 init_waitqueue_head(&rc->wait);
919 kref_init(&rc->refcnt);
920 kref_get(&rc->refcnt); /* one ref for having elements */
922 node = btrfs_root_node(root);
924 generation = btrfs_header_generation(node);
925 free_extent_buffer(node);
927 ret = reada_add_block(rc, start, &max_key, generation);
933 reada_start_machine(root->fs_info);
939 int btrfs_reada_wait(void *handle)
941 struct reada_control *rc = handle;
942 struct btrfs_fs_info *fs_info = rc->fs_info;
944 while (atomic_read(&rc->elems)) {
945 if (!atomic_read(&fs_info->reada_works_cnt))
946 reada_start_machine(fs_info);
947 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
949 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
952 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
954 kref_put(&rc->refcnt, reada_control_release);
959 int btrfs_reada_wait(void *handle)
961 struct reada_control *rc = handle;
962 struct btrfs_fs_info *fs_info = rc->fs_info;
964 while (atomic_read(&rc->elems)) {
965 if (!atomic_read(&fs_info->reada_works_cnt))
966 reada_start_machine(fs_info);
967 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
971 kref_put(&rc->refcnt, reada_control_release);
977 void btrfs_reada_detach(void *handle)
979 struct reada_control *rc = handle;
981 kref_put(&rc->refcnt, reada_control_release);