2 * Copyright (C) 2011 STRATO. All rights reserved.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
19 #include <linux/sched.h>
20 #include <linux/pagemap.h>
21 #include <linux/writeback.h>
22 #include <linux/blkdev.h>
23 #include <linux/rbtree.h>
24 #include <linux/slab.h>
25 #include <linux/workqueue.h>
29 #include "transaction.h"
34 * This is the implementation for the generic read ahead framework.
36 * To trigger a readahead, btrfs_reada_add must be called. It will start
37 * a read ahead for the given range [start, end) on tree root. The returned
38 * handle can either be used to wait on the readahead to finish
39 * (btrfs_reada_wait), or to send it to the background (btrfs_reada_detach).
41 * The read ahead works as follows:
42 * On btrfs_reada_add, the root of the tree is inserted into a radix_tree.
43 * reada_start_machine will then search for extents to prefetch and trigger
44 * some reads. When a read finishes for a node, all contained node/leaf
45 * pointers that lie in the given range will also be enqueued. The reads will
46 * be triggered in sequential order, thus giving a big win over a naive
47 * enumeration. It will also make use of multi-device layouts. Each disk
48 * will have its on read pointer and all disks will by utilized in parallel.
49 * Also will no two disks read both sides of a mirror simultaneously, as this
50 * would waste seeking capacity. Instead both disks will read different parts
52 * Any number of readaheads can be started in parallel. The read order will be
53 * determined globally, i.e. 2 parallel readaheads will normally finish faster
54 * than the 2 started one after another.
58 #define MAX_IN_FLIGHT 6
61 struct list_head list;
62 struct reada_control *rc;
71 struct list_head extctl;
74 struct reada_zone *zones[MAX_MIRRORS];
76 struct btrfs_device *scheduled_for;
83 struct list_head list;
86 struct btrfs_device *device;
87 struct btrfs_device *devs[MAX_MIRRORS]; /* full list, incl self */
92 struct reada_machine_work {
93 struct btrfs_work work;
94 struct btrfs_fs_info *fs_info;
97 static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
98 static void reada_control_release(struct kref *kref);
99 static void reada_zone_release(struct kref *kref);
100 static void reada_start_machine(struct btrfs_fs_info *fs_info);
101 static void __reada_start_machine(struct btrfs_fs_info *fs_info);
103 static int reada_add_block(struct reada_control *rc, u64 logical,
104 struct btrfs_key *top, int level, u64 generation);
107 /* in case of err, eb might be NULL */
108 static int __readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
116 struct reada_extent *re;
117 struct btrfs_fs_info *fs_info = root->fs_info;
118 struct list_head list;
119 unsigned long index = start >> PAGE_CACHE_SHIFT;
120 struct btrfs_device *for_dev;
123 level = btrfs_header_level(eb);
126 spin_lock(&fs_info->reada_lock);
127 re = radix_tree_lookup(&fs_info->reada_tree, index);
129 kref_get(&re->refcnt);
130 spin_unlock(&fs_info->reada_lock);
135 spin_lock(&re->lock);
137 * just take the full list from the extent. afterwards we
138 * don't need the lock anymore
140 list_replace_init(&re->extctl, &list);
141 for_dev = re->scheduled_for;
142 re->scheduled_for = NULL;
143 spin_unlock(&re->lock);
146 nritems = level ? btrfs_header_nritems(eb) : 0;
147 generation = btrfs_header_generation(eb);
149 * FIXME: currently we just set nritems to 0 if this is a leaf,
150 * effectively ignoring the content. In a next step we could
151 * trigger more readahead depending from the content, e.g.
152 * fetch the checksums for the extents in the leaf.
156 * this is the error case, the extent buffer has not been
157 * read correctly. We won't access anything from it and
158 * just cleanup our data structures. Effectively this will
159 * cut the branch below this node from read ahead.
165 for (i = 0; i < nritems; i++) {
166 struct reada_extctl *rec;
168 struct btrfs_key key;
169 struct btrfs_key next_key;
171 btrfs_node_key_to_cpu(eb, &key, i);
173 btrfs_node_key_to_cpu(eb, &next_key, i + 1);
176 bytenr = btrfs_node_blockptr(eb, i);
177 n_gen = btrfs_node_ptr_generation(eb, i);
179 list_for_each_entry(rec, &list, list) {
180 struct reada_control *rc = rec->rc;
183 * if the generation doesn't match, just ignore this
184 * extctl. This will probably cut off a branch from
185 * prefetch. Alternatively one could start a new (sub-)
186 * prefetch for this branch, starting again from root.
187 * FIXME: move the generation check out of this loop
190 if (rec->generation != generation) {
191 printk(KERN_DEBUG "generation mismatch for "
192 "(%llu,%d,%llu) %llu != %llu\n",
193 key.objectid, key.type, key.offset,
194 rec->generation, generation);
197 if (rec->generation == generation &&
198 btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
199 btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
200 reada_add_block(rc, bytenr, &next_key,
205 * free extctl records
207 while (!list_empty(&list)) {
208 struct reada_control *rc;
209 struct reada_extctl *rec;
211 rec = list_first_entry(&list, struct reada_extctl, list);
212 list_del(&rec->list);
216 kref_get(&rc->refcnt);
217 if (atomic_dec_and_test(&rc->elems)) {
218 kref_put(&rc->refcnt, reada_control_release);
221 kref_put(&rc->refcnt, reada_control_release);
223 reada_extent_put(fs_info, re); /* one ref for each entry */
225 reada_extent_put(fs_info, re); /* our ref */
227 atomic_dec(&for_dev->reada_in_flight);
233 * start is passed separately in case eb in NULL, which may be the case with
236 int btree_readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
241 ret = __readahead_hook(root, eb, start, err);
243 reada_start_machine(root->fs_info);
248 static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info,
249 struct btrfs_device *dev, u64 logical,
250 struct btrfs_bio *multi)
254 struct reada_zone *zone;
255 struct btrfs_block_group_cache *cache = NULL;
262 spin_lock(&fs_info->reada_lock);
263 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
264 logical >> PAGE_CACHE_SHIFT, 1);
266 kref_get(&zone->refcnt);
267 spin_unlock(&fs_info->reada_lock);
270 if (logical >= zone->start && logical < zone->end)
272 spin_lock(&fs_info->reada_lock);
273 kref_put(&zone->refcnt, reada_zone_release);
274 spin_unlock(&fs_info->reada_lock);
280 cache = btrfs_lookup_block_group(fs_info, logical);
284 start = cache->key.objectid;
285 end = start + cache->key.offset - 1;
286 btrfs_put_block_group(cache);
288 zone = kzalloc(sizeof(*zone), GFP_NOFS);
294 INIT_LIST_HEAD(&zone->list);
295 spin_lock_init(&zone->lock);
297 kref_init(&zone->refcnt);
299 zone->device = dev; /* our device always sits at index 0 */
300 for (i = 0; i < multi->num_stripes; ++i) {
301 /* bounds have already been checked */
302 zone->devs[i] = multi->stripes[i].dev;
304 zone->ndevs = multi->num_stripes;
306 spin_lock(&fs_info->reada_lock);
307 ret = radix_tree_insert(&dev->reada_zones,
308 (unsigned long)zone->end >> PAGE_CACHE_SHIFT,
310 spin_unlock(&fs_info->reada_lock);
321 static struct reada_extent *reada_find_extent(struct btrfs_root *root,
323 struct btrfs_key *top, int level)
327 struct reada_extent *re = NULL;
328 struct btrfs_fs_info *fs_info = root->fs_info;
329 struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
330 struct btrfs_bio *multi = NULL;
331 struct btrfs_device *dev;
336 unsigned long index = logical >> PAGE_CACHE_SHIFT;
339 spin_lock(&fs_info->reada_lock);
340 re = radix_tree_lookup(&fs_info->reada_tree, index);
342 kref_get(&re->refcnt);
343 spin_unlock(&fs_info->reada_lock);
348 re = kzalloc(sizeof(*re), GFP_NOFS);
352 blocksize = btrfs_level_size(root, level);
353 re->logical = logical;
354 re->blocksize = blocksize;
356 INIT_LIST_HEAD(&re->extctl);
357 spin_lock_init(&re->lock);
358 kref_init(&re->refcnt);
364 ret = btrfs_map_block(map_tree, REQ_WRITE, logical, &length, &multi, 0);
365 if (ret || !multi || length < blocksize)
368 if (multi->num_stripes > MAX_MIRRORS) {
369 printk(KERN_ERR "btrfs readahead: more than %d copies not "
370 "supported", MAX_MIRRORS);
374 for (nzones = 0; nzones < multi->num_stripes; ++nzones) {
375 struct reada_zone *zone;
377 dev = multi->stripes[nzones].dev;
378 zone = reada_find_zone(fs_info, dev, logical, multi);
382 re->zones[nzones] = zone;
383 spin_lock(&zone->lock);
385 kref_get(&zone->refcnt);
387 spin_unlock(&zone->lock);
388 spin_lock(&fs_info->reada_lock);
389 kref_put(&zone->refcnt, reada_zone_release);
390 spin_unlock(&fs_info->reada_lock);
394 /* not a single zone found, error and out */
398 /* insert extent in reada_tree + all per-device trees, all or nothing */
399 spin_lock(&fs_info->reada_lock);
400 ret = radix_tree_insert(&fs_info->reada_tree, index, re);
402 spin_unlock(&fs_info->reada_lock);
403 if (ret != -ENOMEM) {
404 /* someone inserted the extent in the meantime */
409 for (i = 0; i < nzones; ++i) {
410 dev = multi->stripes[i].dev;
411 ret = radix_tree_insert(&dev->reada_extents, index, re);
414 dev = multi->stripes[i].dev;
416 radix_tree_delete(&dev->reada_extents, index);
418 BUG_ON(fs_info == NULL);
419 radix_tree_delete(&fs_info->reada_tree, index);
420 spin_unlock(&fs_info->reada_lock);
424 spin_unlock(&fs_info->reada_lock);
430 struct reada_zone *zone;
433 zone = re->zones[nzones];
434 kref_get(&zone->refcnt);
435 spin_lock(&zone->lock);
437 if (zone->elems == 0) {
439 * no fs_info->reada_lock needed, as this can't be
442 kref_put(&zone->refcnt, reada_zone_release);
444 spin_unlock(&zone->lock);
446 spin_lock(&fs_info->reada_lock);
447 kref_put(&zone->refcnt, reada_zone_release);
448 spin_unlock(&fs_info->reada_lock);
456 static void reada_kref_dummy(struct kref *kr)
460 static void reada_extent_put(struct btrfs_fs_info *fs_info,
461 struct reada_extent *re)
464 unsigned long index = re->logical >> PAGE_CACHE_SHIFT;
466 spin_lock(&fs_info->reada_lock);
467 if (!kref_put(&re->refcnt, reada_kref_dummy)) {
468 spin_unlock(&fs_info->reada_lock);
472 radix_tree_delete(&fs_info->reada_tree, index);
473 for (i = 0; i < re->nzones; ++i) {
474 struct reada_zone *zone = re->zones[i];
476 radix_tree_delete(&zone->device->reada_extents, index);
479 spin_unlock(&fs_info->reada_lock);
481 for (i = 0; i < re->nzones; ++i) {
482 struct reada_zone *zone = re->zones[i];
484 kref_get(&zone->refcnt);
485 spin_lock(&zone->lock);
487 if (zone->elems == 0) {
488 /* no fs_info->reada_lock needed, as this can't be
490 kref_put(&zone->refcnt, reada_zone_release);
492 spin_unlock(&zone->lock);
494 spin_lock(&fs_info->reada_lock);
495 kref_put(&zone->refcnt, reada_zone_release);
496 spin_unlock(&fs_info->reada_lock);
498 if (re->scheduled_for)
499 atomic_dec(&re->scheduled_for->reada_in_flight);
504 static void reada_zone_release(struct kref *kref)
506 struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
508 radix_tree_delete(&zone->device->reada_zones,
509 zone->end >> PAGE_CACHE_SHIFT);
514 static void reada_control_release(struct kref *kref)
516 struct reada_control *rc = container_of(kref, struct reada_control,
522 static int reada_add_block(struct reada_control *rc, u64 logical,
523 struct btrfs_key *top, int level, u64 generation)
525 struct btrfs_root *root = rc->root;
526 struct reada_extent *re;
527 struct reada_extctl *rec;
529 re = reada_find_extent(root, logical, top, level); /* takes one ref */
533 rec = kzalloc(sizeof(*rec), GFP_NOFS);
535 reada_extent_put(root->fs_info, re);
540 rec->generation = generation;
541 atomic_inc(&rc->elems);
543 spin_lock(&re->lock);
544 list_add_tail(&rec->list, &re->extctl);
545 spin_unlock(&re->lock);
547 /* leave the ref on the extent */
553 * called with fs_info->reada_lock held
555 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
558 unsigned long index = zone->end >> PAGE_CACHE_SHIFT;
560 for (i = 0; i < zone->ndevs; ++i) {
561 struct reada_zone *peer;
562 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
563 if (peer && peer->device != zone->device)
569 * called with fs_info->reada_lock held
571 static int reada_pick_zone(struct btrfs_device *dev)
573 struct reada_zone *top_zone = NULL;
574 struct reada_zone *top_locked_zone = NULL;
576 u64 top_locked_elems = 0;
577 unsigned long index = 0;
580 if (dev->reada_curr_zone) {
581 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
582 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
583 dev->reada_curr_zone = NULL;
585 /* pick the zone with the most elements */
587 struct reada_zone *zone;
589 ret = radix_tree_gang_lookup(&dev->reada_zones,
590 (void **)&zone, index, 1);
593 index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
595 if (zone->elems > top_locked_elems) {
596 top_locked_elems = zone->elems;
597 top_locked_zone = zone;
600 if (zone->elems > top_elems) {
601 top_elems = zone->elems;
607 dev->reada_curr_zone = top_zone;
608 else if (top_locked_zone)
609 dev->reada_curr_zone = top_locked_zone;
613 dev->reada_next = dev->reada_curr_zone->start;
614 kref_get(&dev->reada_curr_zone->refcnt);
615 reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
620 static int reada_start_machine_dev(struct btrfs_fs_info *fs_info,
621 struct btrfs_device *dev)
623 struct reada_extent *re = NULL;
625 struct extent_buffer *eb = NULL;
632 spin_lock(&fs_info->reada_lock);
633 if (dev->reada_curr_zone == NULL) {
634 ret = reada_pick_zone(dev);
636 spin_unlock(&fs_info->reada_lock);
641 * FIXME currently we issue the reads one extent at a time. If we have
642 * a contiguous block of extents, we could also coagulate them or use
643 * plugging to speed things up
645 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
646 dev->reada_next >> PAGE_CACHE_SHIFT, 1);
647 if (ret == 0 || re->logical >= dev->reada_curr_zone->end) {
648 ret = reada_pick_zone(dev);
650 spin_unlock(&fs_info->reada_lock);
654 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
655 dev->reada_next >> PAGE_CACHE_SHIFT, 1);
658 spin_unlock(&fs_info->reada_lock);
661 dev->reada_next = re->logical + re->blocksize;
662 kref_get(&re->refcnt);
664 spin_unlock(&fs_info->reada_lock);
669 for (i = 0; i < re->nzones; ++i) {
670 if (re->zones[i]->device == dev) {
675 logical = re->logical;
676 blocksize = re->blocksize;
678 spin_lock(&re->lock);
679 if (re->scheduled_for == NULL) {
680 re->scheduled_for = dev;
683 spin_unlock(&re->lock);
685 reada_extent_put(fs_info, re);
690 atomic_inc(&dev->reada_in_flight);
691 ret = reada_tree_block_flagged(fs_info->extent_root, logical, blocksize,
694 __readahead_hook(fs_info->extent_root, NULL, logical, ret);
696 __readahead_hook(fs_info->extent_root, eb, eb->start, ret);
699 free_extent_buffer(eb);
705 static void reada_start_machine_worker(struct btrfs_work *work)
707 struct reada_machine_work *rmw;
708 struct btrfs_fs_info *fs_info;
710 rmw = container_of(work, struct reada_machine_work, work);
711 fs_info = rmw->fs_info;
715 __reada_start_machine(fs_info);
718 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
720 struct btrfs_device *device;
721 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
728 list_for_each_entry(device, &fs_devices->devices, dev_list) {
729 if (atomic_read(&device->reada_in_flight) <
731 enqueued += reada_start_machine_dev(fs_info,
735 } while (enqueued && total < 10000);
741 * If everything is already in the cache, this is effectively single
742 * threaded. To a) not hold the caller for too long and b) to utilize
743 * more cores, we broke the loop above after 10000 iterations and now
744 * enqueue to workers to finish it. This will distribute the load to
747 for (i = 0; i < 2; ++i)
748 reada_start_machine(fs_info);
751 static void reada_start_machine(struct btrfs_fs_info *fs_info)
753 struct reada_machine_work *rmw;
755 rmw = kzalloc(sizeof(*rmw), GFP_NOFS);
757 /* FIXME we cannot handle this properly right now */
760 rmw->work.func = reada_start_machine_worker;
761 rmw->fs_info = fs_info;
763 btrfs_queue_worker(&fs_info->readahead_workers, &rmw->work);
767 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
769 struct btrfs_device *device;
770 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
777 spin_lock(&fs_info->reada_lock);
778 list_for_each_entry(device, &fs_devices->devices, dev_list) {
779 printk(KERN_DEBUG "dev %lld has %d in flight\n", device->devid,
780 atomic_read(&device->reada_in_flight));
783 struct reada_zone *zone;
784 ret = radix_tree_gang_lookup(&device->reada_zones,
785 (void **)&zone, index, 1);
788 printk(KERN_DEBUG " zone %llu-%llu elems %llu locked "
789 "%d devs", zone->start, zone->end, zone->elems,
791 for (j = 0; j < zone->ndevs; ++j) {
792 printk(KERN_CONT " %lld",
793 zone->devs[j]->devid);
795 if (device->reada_curr_zone == zone)
796 printk(KERN_CONT " curr off %llu",
797 device->reada_next - zone->start);
798 printk(KERN_CONT "\n");
799 index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
804 struct reada_extent *re = NULL;
806 ret = radix_tree_gang_lookup(&device->reada_extents,
807 (void **)&re, index, 1);
811 " re: logical %llu size %u empty %d for %lld",
812 re->logical, re->blocksize,
813 list_empty(&re->extctl), re->scheduled_for ?
814 re->scheduled_for->devid : -1);
816 for (i = 0; i < re->nzones; ++i) {
817 printk(KERN_CONT " zone %llu-%llu devs",
820 for (j = 0; j < re->zones[i]->ndevs; ++j) {
821 printk(KERN_CONT " %lld",
822 re->zones[i]->devs[j]->devid);
825 printk(KERN_CONT "\n");
826 index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
835 struct reada_extent *re = NULL;
837 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
841 if (!re->scheduled_for) {
842 index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
846 "re: logical %llu size %u list empty %d for %lld",
847 re->logical, re->blocksize, list_empty(&re->extctl),
848 re->scheduled_for ? re->scheduled_for->devid : -1);
849 for (i = 0; i < re->nzones; ++i) {
850 printk(KERN_CONT " zone %llu-%llu devs",
853 for (i = 0; i < re->nzones; ++i) {
854 printk(KERN_CONT " zone %llu-%llu devs",
857 for (j = 0; j < re->zones[i]->ndevs; ++j) {
858 printk(KERN_CONT " %lld",
859 re->zones[i]->devs[j]->devid);
863 printk(KERN_CONT "\n");
864 index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
866 spin_unlock(&fs_info->reada_lock);
873 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
874 struct btrfs_key *key_start, struct btrfs_key *key_end)
876 struct reada_control *rc;
880 struct extent_buffer *node;
881 static struct btrfs_key max_key = {
887 rc = kzalloc(sizeof(*rc), GFP_NOFS);
889 return ERR_PTR(-ENOMEM);
892 rc->key_start = *key_start;
893 rc->key_end = *key_end;
894 atomic_set(&rc->elems, 0);
895 init_waitqueue_head(&rc->wait);
896 kref_init(&rc->refcnt);
897 kref_get(&rc->refcnt); /* one ref for having elements */
899 node = btrfs_root_node(root);
901 level = btrfs_header_level(node);
902 generation = btrfs_header_generation(node);
903 free_extent_buffer(node);
905 reada_add_block(rc, start, &max_key, level, generation);
907 reada_start_machine(root->fs_info);
913 int btrfs_reada_wait(void *handle)
915 struct reada_control *rc = handle;
917 while (atomic_read(&rc->elems)) {
918 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
920 dump_devs(rc->root->fs_info, rc->elems < 10 ? 1 : 0);
923 dump_devs(rc->root->fs_info, rc->elems < 10 ? 1 : 0);
925 kref_put(&rc->refcnt, reada_control_release);
930 int btrfs_reada_wait(void *handle)
932 struct reada_control *rc = handle;
934 while (atomic_read(&rc->elems)) {
935 wait_event(rc->wait, atomic_read(&rc->elems) == 0);
938 kref_put(&rc->refcnt, reada_control_release);
944 void btrfs_reada_detach(void *handle)
946 struct reada_control *rc = handle;
948 kref_put(&rc->refcnt, reada_control_release);