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/rbtree.h>
11 #include <linux/slab.h>
12 #include <linux/workqueue.h>
16 #include "transaction.h"
17 #include "dev-replace.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;
56 struct list_head extctl;
59 struct reada_zone *zones[BTRFS_MAX_MIRRORS];
68 struct list_head list;
71 struct btrfs_device *device;
72 struct btrfs_device *devs[BTRFS_MAX_MIRRORS]; /* full list, incl
78 struct reada_machine_work {
79 struct btrfs_work work;
80 struct btrfs_fs_info *fs_info;
83 static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
84 static void reada_control_release(struct kref *kref);
85 static void reada_zone_release(struct kref *kref);
86 static void reada_start_machine(struct btrfs_fs_info *fs_info);
87 static void __reada_start_machine(struct btrfs_fs_info *fs_info);
89 static int reada_add_block(struct reada_control *rc, u64 logical,
90 struct btrfs_key *top, u64 generation);
93 /* in case of err, eb might be NULL */
94 static void __readahead_hook(struct btrfs_fs_info *fs_info,
95 struct reada_extent *re, struct extent_buffer *eb,
102 struct list_head list;
104 spin_lock(&re->lock);
106 * just take the full list from the extent. afterwards we
107 * don't need the lock anymore
109 list_replace_init(&re->extctl, &list);
111 spin_unlock(&re->lock);
114 * this is the error case, the extent buffer has not been
115 * read correctly. We won't access anything from it and
116 * just cleanup our data structures. Effectively this will
117 * cut the branch below this node from read ahead.
123 * FIXME: currently we just set nritems to 0 if this is a leaf,
124 * effectively ignoring the content. In a next step we could
125 * trigger more readahead depending from the content, e.g.
126 * fetch the checksums for the extents in the leaf.
128 if (!btrfs_header_level(eb))
131 nritems = btrfs_header_nritems(eb);
132 generation = btrfs_header_generation(eb);
133 for (i = 0; i < nritems; i++) {
134 struct reada_extctl *rec;
136 struct btrfs_key key;
137 struct btrfs_key next_key;
139 btrfs_node_key_to_cpu(eb, &key, i);
141 btrfs_node_key_to_cpu(eb, &next_key, i + 1);
144 bytenr = btrfs_node_blockptr(eb, i);
145 n_gen = btrfs_node_ptr_generation(eb, i);
147 list_for_each_entry(rec, &list, list) {
148 struct reada_control *rc = rec->rc;
151 * if the generation doesn't match, just ignore this
152 * extctl. This will probably cut off a branch from
153 * prefetch. Alternatively one could start a new (sub-)
154 * prefetch for this branch, starting again from root.
155 * FIXME: move the generation check out of this loop
158 if (rec->generation != generation) {
160 "generation mismatch for (%llu,%d,%llu) %llu != %llu",
161 key.objectid, key.type, key.offset,
162 rec->generation, generation);
165 if (rec->generation == generation &&
166 btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
167 btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
168 reada_add_block(rc, bytenr, &next_key, n_gen);
174 * free extctl records
176 while (!list_empty(&list)) {
177 struct reada_control *rc;
178 struct reada_extctl *rec;
180 rec = list_first_entry(&list, struct reada_extctl, list);
181 list_del(&rec->list);
185 kref_get(&rc->refcnt);
186 if (atomic_dec_and_test(&rc->elems)) {
187 kref_put(&rc->refcnt, reada_control_release);
190 kref_put(&rc->refcnt, reada_control_release);
192 reada_extent_put(fs_info, re); /* one ref for each entry */
198 int btree_readahead_hook(struct extent_buffer *eb, int err)
200 struct btrfs_fs_info *fs_info = eb->fs_info;
202 struct reada_extent *re;
205 spin_lock(&fs_info->reada_lock);
206 re = radix_tree_lookup(&fs_info->reada_tree,
207 eb->start >> PAGE_SHIFT);
210 spin_unlock(&fs_info->reada_lock);
216 __readahead_hook(fs_info, re, eb, err);
217 reada_extent_put(fs_info, re); /* our ref */
220 reada_start_machine(fs_info);
224 static struct reada_zone *reada_find_zone(struct btrfs_device *dev, u64 logical,
225 struct btrfs_bio *bbio)
227 struct btrfs_fs_info *fs_info = dev->fs_info;
229 struct reada_zone *zone;
230 struct btrfs_block_group_cache *cache = NULL;
236 spin_lock(&fs_info->reada_lock);
237 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
238 logical >> PAGE_SHIFT, 1);
239 if (ret == 1 && logical >= zone->start && logical <= zone->end) {
240 kref_get(&zone->refcnt);
241 spin_unlock(&fs_info->reada_lock);
245 spin_unlock(&fs_info->reada_lock);
247 cache = btrfs_lookup_block_group(fs_info, logical);
251 start = cache->key.objectid;
252 end = start + cache->key.offset - 1;
253 btrfs_put_block_group(cache);
255 zone = kzalloc(sizeof(*zone), GFP_KERNEL);
259 ret = radix_tree_preload(GFP_KERNEL);
267 INIT_LIST_HEAD(&zone->list);
268 spin_lock_init(&zone->lock);
270 kref_init(&zone->refcnt);
272 zone->device = dev; /* our device always sits at index 0 */
273 for (i = 0; i < bbio->num_stripes; ++i) {
274 /* bounds have already been checked */
275 zone->devs[i] = bbio->stripes[i].dev;
277 zone->ndevs = bbio->num_stripes;
279 spin_lock(&fs_info->reada_lock);
280 ret = radix_tree_insert(&dev->reada_zones,
281 (unsigned long)(zone->end >> PAGE_SHIFT),
284 if (ret == -EEXIST) {
286 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
287 logical >> PAGE_SHIFT, 1);
288 if (ret == 1 && logical >= zone->start && logical <= zone->end)
289 kref_get(&zone->refcnt);
293 spin_unlock(&fs_info->reada_lock);
294 radix_tree_preload_end();
299 static struct reada_extent *reada_find_extent(struct btrfs_fs_info *fs_info,
301 struct btrfs_key *top)
304 struct reada_extent *re = NULL;
305 struct reada_extent *re_exist = NULL;
306 struct btrfs_bio *bbio = NULL;
307 struct btrfs_device *dev;
308 struct btrfs_device *prev_dev;
312 unsigned long index = logical >> PAGE_SHIFT;
313 int dev_replace_is_ongoing;
316 spin_lock(&fs_info->reada_lock);
317 re = radix_tree_lookup(&fs_info->reada_tree, index);
320 spin_unlock(&fs_info->reada_lock);
325 re = kzalloc(sizeof(*re), GFP_KERNEL);
329 re->logical = logical;
331 INIT_LIST_HEAD(&re->extctl);
332 spin_lock_init(&re->lock);
338 length = fs_info->nodesize;
339 ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
341 if (ret || !bbio || length < fs_info->nodesize)
344 if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
346 "readahead: more than %d copies not supported",
351 real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
352 for (nzones = 0; nzones < real_stripes; ++nzones) {
353 struct reada_zone *zone;
355 dev = bbio->stripes[nzones].dev;
357 /* cannot read ahead on missing device. */
361 zone = reada_find_zone(dev, logical, bbio);
365 re->zones[re->nzones++] = zone;
366 spin_lock(&zone->lock);
368 kref_get(&zone->refcnt);
370 spin_unlock(&zone->lock);
371 spin_lock(&fs_info->reada_lock);
372 kref_put(&zone->refcnt, reada_zone_release);
373 spin_unlock(&fs_info->reada_lock);
375 if (re->nzones == 0) {
376 /* not a single zone found, error and out */
380 ret = radix_tree_preload(GFP_KERNEL);
384 /* insert extent in reada_tree + all per-device trees, all or nothing */
385 btrfs_dev_replace_read_lock(&fs_info->dev_replace);
386 spin_lock(&fs_info->reada_lock);
387 ret = radix_tree_insert(&fs_info->reada_tree, index, re);
388 if (ret == -EEXIST) {
389 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
391 spin_unlock(&fs_info->reada_lock);
392 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
393 radix_tree_preload_end();
397 spin_unlock(&fs_info->reada_lock);
398 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
399 radix_tree_preload_end();
402 radix_tree_preload_end();
404 dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
405 &fs_info->dev_replace);
406 for (nzones = 0; nzones < re->nzones; ++nzones) {
407 dev = re->zones[nzones]->device;
409 if (dev == prev_dev) {
411 * in case of DUP, just add the first zone. As both
412 * are on the same device, there's nothing to gain
414 * Also, it wouldn't work, as the tree is per device
415 * and adding would fail with EEXIST
422 if (dev_replace_is_ongoing &&
423 dev == fs_info->dev_replace.tgtdev) {
425 * as this device is selected for reading only as
426 * a last resort, skip it for read ahead.
431 ret = radix_tree_insert(&dev->reada_extents, index, re);
433 while (--nzones >= 0) {
434 dev = re->zones[nzones]->device;
436 /* ignore whether the entry was inserted */
437 radix_tree_delete(&dev->reada_extents, index);
439 radix_tree_delete(&fs_info->reada_tree, index);
440 spin_unlock(&fs_info->reada_lock);
441 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
446 spin_unlock(&fs_info->reada_lock);
447 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
452 btrfs_put_bbio(bbio);
456 for (nzones = 0; nzones < re->nzones; ++nzones) {
457 struct reada_zone *zone;
459 zone = re->zones[nzones];
460 kref_get(&zone->refcnt);
461 spin_lock(&zone->lock);
463 if (zone->elems == 0) {
465 * no fs_info->reada_lock needed, as this can't be
468 kref_put(&zone->refcnt, reada_zone_release);
470 spin_unlock(&zone->lock);
472 spin_lock(&fs_info->reada_lock);
473 kref_put(&zone->refcnt, reada_zone_release);
474 spin_unlock(&fs_info->reada_lock);
476 btrfs_put_bbio(bbio);
481 static void reada_extent_put(struct btrfs_fs_info *fs_info,
482 struct reada_extent *re)
485 unsigned long index = re->logical >> PAGE_SHIFT;
487 spin_lock(&fs_info->reada_lock);
489 spin_unlock(&fs_info->reada_lock);
493 radix_tree_delete(&fs_info->reada_tree, index);
494 for (i = 0; i < re->nzones; ++i) {
495 struct reada_zone *zone = re->zones[i];
497 radix_tree_delete(&zone->device->reada_extents, index);
500 spin_unlock(&fs_info->reada_lock);
502 for (i = 0; i < re->nzones; ++i) {
503 struct reada_zone *zone = re->zones[i];
505 kref_get(&zone->refcnt);
506 spin_lock(&zone->lock);
508 if (zone->elems == 0) {
509 /* no fs_info->reada_lock needed, as this can't be
511 kref_put(&zone->refcnt, reada_zone_release);
513 spin_unlock(&zone->lock);
515 spin_lock(&fs_info->reada_lock);
516 kref_put(&zone->refcnt, reada_zone_release);
517 spin_unlock(&fs_info->reada_lock);
523 static void reada_zone_release(struct kref *kref)
525 struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
527 radix_tree_delete(&zone->device->reada_zones,
528 zone->end >> PAGE_SHIFT);
533 static void reada_control_release(struct kref *kref)
535 struct reada_control *rc = container_of(kref, struct reada_control,
541 static int reada_add_block(struct reada_control *rc, u64 logical,
542 struct btrfs_key *top, u64 generation)
544 struct btrfs_fs_info *fs_info = rc->fs_info;
545 struct reada_extent *re;
546 struct reada_extctl *rec;
549 re = reada_find_extent(fs_info, logical, top);
553 rec = kzalloc(sizeof(*rec), GFP_KERNEL);
555 reada_extent_put(fs_info, re);
560 rec->generation = generation;
561 atomic_inc(&rc->elems);
563 spin_lock(&re->lock);
564 list_add_tail(&rec->list, &re->extctl);
565 spin_unlock(&re->lock);
567 /* leave the ref on the extent */
573 * called with fs_info->reada_lock held
575 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
578 unsigned long index = zone->end >> PAGE_SHIFT;
580 for (i = 0; i < zone->ndevs; ++i) {
581 struct reada_zone *peer;
582 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
583 if (peer && peer->device != zone->device)
589 * called with fs_info->reada_lock held
591 static int reada_pick_zone(struct btrfs_device *dev)
593 struct reada_zone *top_zone = NULL;
594 struct reada_zone *top_locked_zone = NULL;
596 u64 top_locked_elems = 0;
597 unsigned long index = 0;
600 if (dev->reada_curr_zone) {
601 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
602 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
603 dev->reada_curr_zone = NULL;
605 /* pick the zone with the most elements */
607 struct reada_zone *zone;
609 ret = radix_tree_gang_lookup(&dev->reada_zones,
610 (void **)&zone, index, 1);
613 index = (zone->end >> PAGE_SHIFT) + 1;
615 if (zone->elems > top_locked_elems) {
616 top_locked_elems = zone->elems;
617 top_locked_zone = zone;
620 if (zone->elems > top_elems) {
621 top_elems = zone->elems;
627 dev->reada_curr_zone = top_zone;
628 else if (top_locked_zone)
629 dev->reada_curr_zone = top_locked_zone;
633 dev->reada_next = dev->reada_curr_zone->start;
634 kref_get(&dev->reada_curr_zone->refcnt);
635 reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
640 static int reada_start_machine_dev(struct btrfs_device *dev)
642 struct btrfs_fs_info *fs_info = dev->fs_info;
643 struct reada_extent *re = NULL;
645 struct extent_buffer *eb = NULL;
650 spin_lock(&fs_info->reada_lock);
651 if (dev->reada_curr_zone == NULL) {
652 ret = reada_pick_zone(dev);
654 spin_unlock(&fs_info->reada_lock);
659 * FIXME currently we issue the reads one extent at a time. If we have
660 * a contiguous block of extents, we could also coagulate them or use
661 * plugging to speed things up
663 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
664 dev->reada_next >> PAGE_SHIFT, 1);
665 if (ret == 0 || re->logical > dev->reada_curr_zone->end) {
666 ret = reada_pick_zone(dev);
668 spin_unlock(&fs_info->reada_lock);
672 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
673 dev->reada_next >> PAGE_SHIFT, 1);
676 spin_unlock(&fs_info->reada_lock);
679 dev->reada_next = re->logical + fs_info->nodesize;
682 spin_unlock(&fs_info->reada_lock);
684 spin_lock(&re->lock);
685 if (re->scheduled || list_empty(&re->extctl)) {
686 spin_unlock(&re->lock);
687 reada_extent_put(fs_info, re);
691 spin_unlock(&re->lock);
696 for (i = 0; i < re->nzones; ++i) {
697 if (re->zones[i]->device == dev) {
702 logical = re->logical;
704 atomic_inc(&dev->reada_in_flight);
705 ret = reada_tree_block_flagged(fs_info, logical, mirror_num, &eb);
707 __readahead_hook(fs_info, re, NULL, ret);
709 __readahead_hook(fs_info, re, eb, ret);
712 free_extent_buffer(eb);
714 atomic_dec(&dev->reada_in_flight);
715 reada_extent_put(fs_info, re);
721 static void reada_start_machine_worker(struct btrfs_work *work)
723 struct reada_machine_work *rmw;
724 struct btrfs_fs_info *fs_info;
727 rmw = container_of(work, struct reada_machine_work, work);
728 fs_info = rmw->fs_info;
732 old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
733 task_nice_ioprio(current));
734 set_task_ioprio(current, BTRFS_IOPRIO_READA);
735 __reada_start_machine(fs_info);
736 set_task_ioprio(current, old_ioprio);
738 atomic_dec(&fs_info->reada_works_cnt);
741 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
743 struct btrfs_device *device;
744 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
751 mutex_lock(&fs_devices->device_list_mutex);
752 list_for_each_entry(device, &fs_devices->devices, dev_list) {
753 if (atomic_read(&device->reada_in_flight) <
755 enqueued += reada_start_machine_dev(device);
757 mutex_unlock(&fs_devices->device_list_mutex);
759 } while (enqueued && total < 10000);
765 * If everything is already in the cache, this is effectively single
766 * threaded. To a) not hold the caller for too long and b) to utilize
767 * more cores, we broke the loop above after 10000 iterations and now
768 * enqueue to workers to finish it. This will distribute the load to
771 for (i = 0; i < 2; ++i) {
772 reada_start_machine(fs_info);
773 if (atomic_read(&fs_info->reada_works_cnt) >
774 BTRFS_MAX_MIRRORS * 2)
779 static void reada_start_machine(struct btrfs_fs_info *fs_info)
781 struct reada_machine_work *rmw;
783 rmw = kzalloc(sizeof(*rmw), GFP_KERNEL);
785 /* FIXME we cannot handle this properly right now */
788 btrfs_init_work(&rmw->work, btrfs_readahead_helper,
789 reada_start_machine_worker, NULL, NULL);
790 rmw->fs_info = fs_info;
792 btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
793 atomic_inc(&fs_info->reada_works_cnt);
797 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
799 struct btrfs_device *device;
800 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
807 spin_lock(&fs_info->reada_lock);
808 list_for_each_entry(device, &fs_devices->devices, dev_list) {
809 btrfs_debug(fs_info, "dev %lld has %d in flight", device->devid,
810 atomic_read(&device->reada_in_flight));
813 struct reada_zone *zone;
814 ret = radix_tree_gang_lookup(&device->reada_zones,
815 (void **)&zone, index, 1);
818 pr_debug(" zone %llu-%llu elems %llu locked %d devs",
819 zone->start, zone->end, zone->elems,
821 for (j = 0; j < zone->ndevs; ++j) {
823 zone->devs[j]->devid);
825 if (device->reada_curr_zone == zone)
826 pr_cont(" curr off %llu",
827 device->reada_next - zone->start);
829 index = (zone->end >> PAGE_SHIFT) + 1;
834 struct reada_extent *re = NULL;
836 ret = radix_tree_gang_lookup(&device->reada_extents,
837 (void **)&re, index, 1);
840 pr_debug(" re: logical %llu size %u empty %d scheduled %d",
841 re->logical, fs_info->nodesize,
842 list_empty(&re->extctl), re->scheduled);
844 for (i = 0; i < re->nzones; ++i) {
845 pr_cont(" zone %llu-%llu devs",
848 for (j = 0; j < re->zones[i]->ndevs; ++j) {
850 re->zones[i]->devs[j]->devid);
854 index = (re->logical >> PAGE_SHIFT) + 1;
863 struct reada_extent *re = NULL;
865 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
869 if (!re->scheduled) {
870 index = (re->logical >> PAGE_SHIFT) + 1;
873 pr_debug("re: logical %llu size %u list empty %d scheduled %d",
874 re->logical, fs_info->nodesize,
875 list_empty(&re->extctl), re->scheduled);
876 for (i = 0; i < re->nzones; ++i) {
877 pr_cont(" zone %llu-%llu devs",
880 for (j = 0; j < re->zones[i]->ndevs; ++j) {
882 re->zones[i]->devs[j]->devid);
886 index = (re->logical >> PAGE_SHIFT) + 1;
888 spin_unlock(&fs_info->reada_lock);
895 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
896 struct btrfs_key *key_start, struct btrfs_key *key_end)
898 struct reada_control *rc;
902 struct extent_buffer *node;
903 static struct btrfs_key max_key = {
909 rc = kzalloc(sizeof(*rc), GFP_KERNEL);
911 return ERR_PTR(-ENOMEM);
913 rc->fs_info = root->fs_info;
914 rc->key_start = *key_start;
915 rc->key_end = *key_end;
916 atomic_set(&rc->elems, 0);
917 init_waitqueue_head(&rc->wait);
918 kref_init(&rc->refcnt);
919 kref_get(&rc->refcnt); /* one ref for having elements */
921 node = btrfs_root_node(root);
923 generation = btrfs_header_generation(node);
924 free_extent_buffer(node);
926 ret = reada_add_block(rc, start, &max_key, generation);
932 reada_start_machine(root->fs_info);
938 int btrfs_reada_wait(void *handle)
940 struct reada_control *rc = handle;
941 struct btrfs_fs_info *fs_info = rc->fs_info;
943 while (atomic_read(&rc->elems)) {
944 if (!atomic_read(&fs_info->reada_works_cnt))
945 reada_start_machine(fs_info);
946 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
948 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
951 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
953 kref_put(&rc->refcnt, reada_control_release);
958 int btrfs_reada_wait(void *handle)
960 struct reada_control *rc = handle;
961 struct btrfs_fs_info *fs_info = rc->fs_info;
963 while (atomic_read(&rc->elems)) {
964 if (!atomic_read(&fs_info->reada_works_cnt))
965 reada_start_machine(fs_info);
966 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
970 kref_put(&rc->refcnt, reada_control_release);
976 void btrfs_reada_detach(void *handle)
978 struct reada_control *rc = handle;
980 kref_put(&rc->refcnt, reada_control_release);