aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorChris Mason <chris.mason@oracle.com>2011-11-06 03:05:08 -0500
committerChris Mason <chris.mason@oracle.com>2011-11-06 03:05:08 -0500
commit531f4b1ae5e0fc8c9b3f03838218e5ea178f80d3 (patch)
tree63efdfe9b192243fefb76be3921b9a2aaa26291e
parentc06a0e120a4e381a1c291c1fce3c6155c5791cae (diff)
parent7a26285eea8eb92e0088db011571d887d4551b0f (diff)
Merge branch 'for-chris' of git://github.com/sensille/linux into integration
Conflicts: fs/btrfs/ctree.h Signed-off-by: Chris Mason <chris.mason@oracle.com>
Diffstat
-rw-r--r--fs/btrfs/Makefile3
-rw-r--r--fs/btrfs/ctree.h21
-rw-r--r--fs/btrfs/disk-io.c84
-rw-r--r--fs/btrfs/disk-io.h2
-rw-r--r--fs/btrfs/extent_io.c9
-rw-r--r--fs/btrfs/extent_io.h4
-rw-r--r--fs/btrfs/reada.c949
-rw-r--r--fs/btrfs/scrub.c112
-rw-r--r--fs/btrfs/volumes.c8
-rw-r--r--fs/btrfs/volumes.h8
10 files changed, 1130 insertions, 70 deletions
diff --git a/fs/btrfs/Makefile b/fs/btrfs/Makefile
index 40e6ac08c21f..bdd6fb238ce1 100644
--- a/fs/btrfs/Makefile
+++ b/fs/btrfs/Makefile
@@ -7,6 +7,7 @@ btrfs-y += super.o ctree.o extent-tree.o print-tree.o root-tree.o dir-item.o \
7 extent_map.o sysfs.o struct-funcs.o xattr.o ordered-data.o \ 7 extent_map.o sysfs.o struct-funcs.o xattr.o ordered-data.o \
8 extent_io.o volumes.o async-thread.o ioctl.o locking.o orphan.o \ 8 extent_io.o volumes.o async-thread.o ioctl.o locking.o orphan.o \
9 export.o tree-log.o free-space-cache.o zlib.o lzo.o \ 9 export.o tree-log.o free-space-cache.o zlib.o lzo.o \
10 compression.o delayed-ref.o relocation.o delayed-inode.o scrub.o 10 compression.o delayed-ref.o relocation.o delayed-inode.o scrub.o \
11 reada.o
11 12
12btrfs-$(CONFIG_BTRFS_FS_POSIX_ACL) += acl.o 13btrfs-$(CONFIG_BTRFS_FS_POSIX_ACL) += acl.o
diff --git a/fs/btrfs/ctree.h b/fs/btrfs/ctree.h
index 6bb34fc1ff22..b9ba59ff9292 100644
--- a/fs/btrfs/ctree.h
+++ b/fs/btrfs/ctree.h
@@ -1074,6 +1074,7 @@ struct btrfs_fs_info {
1074 struct btrfs_workers endio_freespace_worker; 1074 struct btrfs_workers endio_freespace_worker;
1075 struct btrfs_workers submit_workers; 1075 struct btrfs_workers submit_workers;
1076 struct btrfs_workers caching_workers; 1076 struct btrfs_workers caching_workers;
1077 struct btrfs_workers readahead_workers;
1077 1078
1078 /* 1079 /*
1079 * fixup workers take dirty pages that didn't properly go through 1080 * fixup workers take dirty pages that didn't properly go through
@@ -1158,6 +1159,10 @@ struct btrfs_fs_info {
1158 1159
1159 struct btrfs_delayed_root *delayed_root; 1160 struct btrfs_delayed_root *delayed_root;
1160 1161
1162 /* readahead tree */
1163 spinlock_t reada_lock;
1164 struct radix_tree_root reada_tree;
1165
1161 /* next backup root to be overwritten */ 1166 /* next backup root to be overwritten */
1162 int backup_root_index; 1167 int backup_root_index;
1163}; 1168};
@@ -2812,4 +2817,20 @@ int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid);
2812int btrfs_scrub_progress(struct btrfs_root *root, u64 devid, 2817int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
2813 struct btrfs_scrub_progress *progress); 2818 struct btrfs_scrub_progress *progress);
2814 2819
2820/* reada.c */
2821struct reada_control {
2822 struct btrfs_root *root; /* tree to prefetch */
2823 struct btrfs_key key_start;
2824 struct btrfs_key key_end; /* exclusive */
2825 atomic_t elems;
2826 struct kref refcnt;
2827 wait_queue_head_t wait;
2828};
2829struct reada_control *btrfs_reada_add(struct btrfs_root *root,
2830 struct btrfs_key *start, struct btrfs_key *end);
2831int btrfs_reada_wait(void *handle);
2832void btrfs_reada_detach(void *handle);
2833int btree_readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
2834 u64 start, int err);
2835
2815#endif 2836#endif
diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
index 23b6776477b7..cedfbfb278eb 100644
--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c
@@ -366,7 +366,8 @@ static int btree_read_extent_buffer_pages(struct btrfs_root *root,
366 clear_bit(EXTENT_BUFFER_CORRUPT, &eb->bflags); 366 clear_bit(EXTENT_BUFFER_CORRUPT, &eb->bflags);
367 io_tree = &BTRFS_I(root->fs_info->btree_inode)->io_tree; 367 io_tree = &BTRFS_I(root->fs_info->btree_inode)->io_tree;
368 while (1) { 368 while (1) {
369 ret = read_extent_buffer_pages(io_tree, eb, start, 1, 369 ret = read_extent_buffer_pages(io_tree, eb, start,
370 WAIT_COMPLETE,
370 btree_get_extent, mirror_num); 371 btree_get_extent, mirror_num);
371 if (!ret && 372 if (!ret &&
372 !verify_parent_transid(io_tree, eb, parent_transid)) 373 !verify_parent_transid(io_tree, eb, parent_transid))
@@ -607,11 +608,47 @@ static int btree_readpage_end_io_hook(struct page *page, u64 start, u64 end,
607 end = min_t(u64, eb->len, PAGE_CACHE_SIZE); 608 end = min_t(u64, eb->len, PAGE_CACHE_SIZE);
608 end = eb->start + end - 1; 609 end = eb->start + end - 1;
609err: 610err:
611 if (test_bit(EXTENT_BUFFER_READAHEAD, &eb->bflags)) {
612 clear_bit(EXTENT_BUFFER_READAHEAD, &eb->bflags);
613 btree_readahead_hook(root, eb, eb->start, ret);
614 }
615
610 free_extent_buffer(eb); 616 free_extent_buffer(eb);
611out: 617out:
612 return ret; 618 return ret;
613} 619}
614 620
621static int btree_io_failed_hook(struct bio *failed_bio,
622 struct page *page, u64 start, u64 end,
623 struct extent_state *state)
624{
625 struct extent_io_tree *tree;
626 unsigned long len;
627 struct extent_buffer *eb;
628 struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
629
630 tree = &BTRFS_I(page->mapping->host)->io_tree;
631 if (page->private == EXTENT_PAGE_PRIVATE)
632 goto out;
633 if (!page->private)
634 goto out;
635
636 len = page->private >> 2;
637 WARN_ON(len == 0);
638
639 eb = alloc_extent_buffer(tree, start, len, page);
640 if (eb == NULL)
641 goto out;
642
643 if (test_bit(EXTENT_BUFFER_READAHEAD, &eb->bflags)) {
644 clear_bit(EXTENT_BUFFER_READAHEAD, &eb->bflags);
645 btree_readahead_hook(root, eb, eb->start, -EIO);
646 }
647
648out:
649 return -EIO; /* we fixed nothing */
650}
651
615static void end_workqueue_bio(struct bio *bio, int err) 652static void end_workqueue_bio(struct bio *bio, int err)
616{ 653{
617 struct end_io_wq *end_io_wq = bio->bi_private; 654 struct end_io_wq *end_io_wq = bio->bi_private;
@@ -973,11 +1010,43 @@ int readahead_tree_block(struct btrfs_root *root, u64 bytenr, u32 blocksize,
973 if (!buf) 1010 if (!buf)
974 return 0; 1011 return 0;
975 read_extent_buffer_pages(&BTRFS_I(btree_inode)->io_tree, 1012 read_extent_buffer_pages(&BTRFS_I(btree_inode)->io_tree,
976 buf, 0, 0, btree_get_extent, 0); 1013 buf, 0, WAIT_NONE, btree_get_extent, 0);
977 free_extent_buffer(buf); 1014 free_extent_buffer(buf);
978 return ret; 1015 return ret;
979} 1016}
980 1017
1018int reada_tree_block_flagged(struct btrfs_root *root, u64 bytenr, u32 blocksize,
1019 int mirror_num, struct extent_buffer **eb)
1020{
1021 struct extent_buffer *buf = NULL;
1022 struct inode *btree_inode = root->fs_info->btree_inode;
1023 struct extent_io_tree *io_tree = &BTRFS_I(btree_inode)->io_tree;
1024 int ret;
1025
1026 buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
1027 if (!buf)
1028 return 0;
1029
1030 set_bit(EXTENT_BUFFER_READAHEAD, &buf->bflags);
1031
1032 ret = read_extent_buffer_pages(io_tree, buf, 0, WAIT_PAGE_LOCK,
1033 btree_get_extent, mirror_num);
1034 if (ret) {
1035 free_extent_buffer(buf);
1036 return ret;
1037 }
1038
1039 if (test_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags)) {
1040 free_extent_buffer(buf);
1041 return -EIO;
1042 } else if (extent_buffer_uptodate(io_tree, buf, NULL)) {
1043 *eb = buf;
1044 } else {
1045 free_extent_buffer(buf);
1046 }
1047 return 0;
1048}
1049
981struct extent_buffer *btrfs_find_tree_block(struct btrfs_root *root, 1050struct extent_buffer *btrfs_find_tree_block(struct btrfs_root *root,
982 u64 bytenr, u32 blocksize) 1051 u64 bytenr, u32 blocksize)
983{ 1052{
@@ -1904,6 +1973,10 @@ struct btrfs_root *open_ctree(struct super_block *sb,
1904 fs_info->trans_no_join = 0; 1973 fs_info->trans_no_join = 0;
1905 fs_info->free_chunk_space = 0; 1974 fs_info->free_chunk_space = 0;
1906 1975
1976 /* readahead state */
1977 INIT_RADIX_TREE(&fs_info->reada_tree, GFP_NOFS & ~__GFP_WAIT);
1978 spin_lock_init(&fs_info->reada_lock);
1979
1907 fs_info->thread_pool_size = min_t(unsigned long, 1980 fs_info->thread_pool_size = min_t(unsigned long,
1908 num_online_cpus() + 2, 8); 1981 num_online_cpus() + 2, 8);
1909 1982
@@ -2103,6 +2176,9 @@ struct btrfs_root *open_ctree(struct super_block *sb,
2103 btrfs_init_workers(&fs_info->delayed_workers, "delayed-meta", 2176 btrfs_init_workers(&fs_info->delayed_workers, "delayed-meta",
2104 fs_info->thread_pool_size, 2177 fs_info->thread_pool_size,
2105 &fs_info->generic_worker); 2178 &fs_info->generic_worker);
2179 btrfs_init_workers(&fs_info->readahead_workers, "readahead",
2180 fs_info->thread_pool_size,
2181 &fs_info->generic_worker);
2106 2182
2107 /* 2183 /*
2108 * endios are largely parallel and should have a very 2184 * endios are largely parallel and should have a very
@@ -2113,6 +2189,7 @@ struct btrfs_root *open_ctree(struct super_block *sb,
2113 2189
2114 fs_info->endio_write_workers.idle_thresh = 2; 2190 fs_info->endio_write_workers.idle_thresh = 2;
2115 fs_info->endio_meta_write_workers.idle_thresh = 2; 2191 fs_info->endio_meta_write_workers.idle_thresh = 2;
2192 fs_info->readahead_workers.idle_thresh = 2;
2116 2193
2117 btrfs_start_workers(&fs_info->workers, 1); 2194 btrfs_start_workers(&fs_info->workers, 1);
2118 btrfs_start_workers(&fs_info->generic_worker, 1); 2195 btrfs_start_workers(&fs_info->generic_worker, 1);
@@ -2126,6 +2203,7 @@ struct btrfs_root *open_ctree(struct super_block *sb,
2126 btrfs_start_workers(&fs_info->endio_freespace_worker, 1); 2203 btrfs_start_workers(&fs_info->endio_freespace_worker, 1);
2127 btrfs_start_workers(&fs_info->delayed_workers, 1); 2204 btrfs_start_workers(&fs_info->delayed_workers, 1);
2128 btrfs_start_workers(&fs_info->caching_workers, 1); 2205 btrfs_start_workers(&fs_info->caching_workers, 1);
2206 btrfs_start_workers(&fs_info->readahead_workers, 1);
2129 2207
2130 fs_info->bdi.ra_pages *= btrfs_super_num_devices(disk_super); 2208 fs_info->bdi.ra_pages *= btrfs_super_num_devices(disk_super);
2131 fs_info->bdi.ra_pages = max(fs_info->bdi.ra_pages, 2209 fs_info->bdi.ra_pages = max(fs_info->bdi.ra_pages,
@@ -2855,6 +2933,7 @@ int close_ctree(struct btrfs_root *root)
2855 btrfs_stop_workers(&fs_info->submit_workers); 2933 btrfs_stop_workers(&fs_info->submit_workers);
2856 btrfs_stop_workers(&fs_info->delayed_workers); 2934 btrfs_stop_workers(&fs_info->delayed_workers);
2857 btrfs_stop_workers(&fs_info->caching_workers); 2935 btrfs_stop_workers(&fs_info->caching_workers);
2936 btrfs_stop_workers(&fs_info->readahead_workers);
2858 2937
2859 btrfs_close_devices(fs_info->fs_devices); 2938 btrfs_close_devices(fs_info->fs_devices);
2860 btrfs_mapping_tree_free(&fs_info->mapping_tree); 2939 btrfs_mapping_tree_free(&fs_info->mapping_tree);
@@ -3363,6 +3442,7 @@ static int btrfs_cleanup_transaction(struct btrfs_root *root)
3363static struct extent_io_ops btree_extent_io_ops = { 3442static struct extent_io_ops btree_extent_io_ops = {
3364 .write_cache_pages_lock_hook = btree_lock_page_hook, 3443 .write_cache_pages_lock_hook = btree_lock_page_hook,
3365 .readpage_end_io_hook = btree_readpage_end_io_hook, 3444 .readpage_end_io_hook = btree_readpage_end_io_hook,
3445 .readpage_io_failed_hook = btree_io_failed_hook,
3366 .submit_bio_hook = btree_submit_bio_hook, 3446 .submit_bio_hook = btree_submit_bio_hook,
3367 /* note we're sharing with inode.c for the merge bio hook */ 3447 /* note we're sharing with inode.c for the merge bio hook */
3368 .merge_bio_hook = btrfs_merge_bio_hook, 3448 .merge_bio_hook = btrfs_merge_bio_hook,
diff --git a/fs/btrfs/disk-io.h b/fs/btrfs/disk-io.h
index e678539c8519..c99d0a8f13fa 100644
--- a/fs/btrfs/disk-io.h
+++ b/fs/btrfs/disk-io.h
@@ -40,6 +40,8 @@ struct extent_buffer *read_tree_block(struct btrfs_root *root, u64 bytenr,
40 u32 blocksize, u64 parent_transid); 40 u32 blocksize, u64 parent_transid);
41int readahead_tree_block(struct btrfs_root *root, u64 bytenr, u32 blocksize, 41int readahead_tree_block(struct btrfs_root *root, u64 bytenr, u32 blocksize,
42 u64 parent_transid); 42 u64 parent_transid);
43int reada_tree_block_flagged(struct btrfs_root *root, u64 bytenr, u32 blocksize,
44 int mirror_num, struct extent_buffer **eb);
43struct extent_buffer *btrfs_find_create_tree_block(struct btrfs_root *root, 45struct extent_buffer *btrfs_find_create_tree_block(struct btrfs_root *root,
44 u64 bytenr, u32 blocksize); 46 u64 bytenr, u32 blocksize);
45int clean_tree_block(struct btrfs_trans_handle *trans, 47int clean_tree_block(struct btrfs_trans_handle *trans,
diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c
index cc3c58970d4e..c12705682c65 100644
--- a/fs/btrfs/extent_io.c
+++ b/fs/btrfs/extent_io.c
@@ -1919,7 +1919,7 @@ static void end_bio_extent_readpage(struct bio *bio, int err)
1919 if (!uptodate && tree->ops && 1919 if (!uptodate && tree->ops &&
1920 tree->ops->readpage_io_failed_hook) { 1920 tree->ops->readpage_io_failed_hook) {
1921 ret = tree->ops->readpage_io_failed_hook(bio, page, 1921 ret = tree->ops->readpage_io_failed_hook(bio, page,
1922 start, end, NULL); 1922 start, end, state);
1923 if (ret == 0) { 1923 if (ret == 0) {
1924 uptodate = 1924 uptodate =
1925 test_bit(BIO_UPTODATE, &bio->bi_flags); 1925 test_bit(BIO_UPTODATE, &bio->bi_flags);
@@ -3551,8 +3551,7 @@ int extent_buffer_uptodate(struct extent_io_tree *tree,
3551} 3551}
3552 3552
3553int read_extent_buffer_pages(struct extent_io_tree *tree, 3553int read_extent_buffer_pages(struct extent_io_tree *tree,
3554 struct extent_buffer *eb, 3554 struct extent_buffer *eb, u64 start, int wait,
3555 u64 start, int wait,
3556 get_extent_t *get_extent, int mirror_num) 3555 get_extent_t *get_extent, int mirror_num)
3557{ 3556{
3558 unsigned long i; 3557 unsigned long i;
@@ -3588,7 +3587,7 @@ int read_extent_buffer_pages(struct extent_io_tree *tree,
3588 num_pages = num_extent_pages(eb->start, eb->len); 3587 num_pages = num_extent_pages(eb->start, eb->len);
3589 for (i = start_i; i < num_pages; i++) { 3588 for (i = start_i; i < num_pages; i++) {
3590 page = extent_buffer_page(eb, i); 3589 page = extent_buffer_page(eb, i);
3591 if (!wait) { 3590 if (wait == WAIT_NONE) {
3592 if (!trylock_page(page)) 3591 if (!trylock_page(page))
3593 goto unlock_exit; 3592 goto unlock_exit;
3594 } else { 3593 } else {
@@ -3632,7 +3631,7 @@ int read_extent_buffer_pages(struct extent_io_tree *tree,
3632 if (bio) 3631 if (bio)
3633 submit_one_bio(READ, bio, mirror_num, bio_flags); 3632 submit_one_bio(READ, bio, mirror_num, bio_flags);
3634 3633
3635 if (ret || !wait) 3634 if (ret || wait != WAIT_COMPLETE)
3636 return ret; 3635 return ret;
3637 3636
3638 for (i = start_i; i < num_pages; i++) { 3637 for (i = start_i; i < num_pages; i++) {
diff --git a/fs/btrfs/extent_io.h b/fs/btrfs/extent_io.h
index cbd4824a7c94..697570eed9e8 100644
--- a/fs/btrfs/extent_io.h
+++ b/fs/btrfs/extent_io.h
@@ -33,6 +33,7 @@
33#define EXTENT_BUFFER_BLOCKING 1 33#define EXTENT_BUFFER_BLOCKING 1
34#define EXTENT_BUFFER_DIRTY 2 34#define EXTENT_BUFFER_DIRTY 2
35#define EXTENT_BUFFER_CORRUPT 3 35#define EXTENT_BUFFER_CORRUPT 3
36#define EXTENT_BUFFER_READAHEAD 4 /* this got triggered by readahead */
36 37
37/* these are flags for extent_clear_unlock_delalloc */ 38/* these are flags for extent_clear_unlock_delalloc */
38#define EXTENT_CLEAR_UNLOCK_PAGE 0x1 39#define EXTENT_CLEAR_UNLOCK_PAGE 0x1
@@ -252,6 +253,9 @@ struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
252struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree, 253struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
253 u64 start, unsigned long len); 254 u64 start, unsigned long len);
254void free_extent_buffer(struct extent_buffer *eb); 255void free_extent_buffer(struct extent_buffer *eb);
256#define WAIT_NONE 0
257#define WAIT_COMPLETE 1
258#define WAIT_PAGE_LOCK 2
255int read_extent_buffer_pages(struct extent_io_tree *tree, 259int read_extent_buffer_pages(struct extent_io_tree *tree,
256 struct extent_buffer *eb, u64 start, int wait, 260 struct extent_buffer *eb, u64 start, int wait,
257 get_extent_t *get_extent, int mirror_num); 261 get_extent_t *get_extent, int mirror_num);
diff --git a/fs/btrfs/reada.c b/fs/btrfs/reada.c
new file mode 100644
index 000000000000..2b701d082227
--- /dev/null
+++ b/fs/btrfs/reada.c
@@ -0,0 +1,949 @@
1/*
2 * Copyright (C) 2011 STRATO. All rights reserved.
3 *
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.
7 *
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.
12 *
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.
17 */
18
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>
26#include "ctree.h"
27#include "volumes.h"
28#include "disk-io.h"
29#include "transaction.h"
30
31#undef DEBUG
32
33/*
34 * This is the implementation for the generic read ahead framework.
35 *
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).
40 *
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
51 * of the filesystem.
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.
55 */
56
57#define MAX_MIRRORS 2
58#define MAX_IN_FLIGHT 6
59
60struct reada_extctl {
61 struct list_head list;
62 struct reada_control *rc;
63 u64 generation;
64};
65
66struct reada_extent {
67 u64 logical;
68 struct btrfs_key top;
69 u32 blocksize;
70 int err;
71 struct list_head extctl;
72 struct kref refcnt;
73 spinlock_t lock;
74 struct reada_zone *zones[MAX_MIRRORS];
75 int nzones;
76 struct btrfs_device *scheduled_for;
77};
78
79struct reada_zone {
80 u64 start;
81 u64 end;
82 u64 elems;
83 struct list_head list;
84 spinlock_t lock;
85 int locked;
86 struct btrfs_device *device;
87 struct btrfs_device *devs[MAX_MIRRORS]; /* full list, incl self */
88 int ndevs;
89 struct kref refcnt;
90};
91
92struct reada_machine_work {
93 struct btrfs_work work;
94 struct btrfs_fs_info *fs_info;
95};
96
97static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
98static void reada_control_release(struct kref *kref);
99static void reada_zone_release(struct kref *kref);
100static void reada_start_machine(struct btrfs_fs_info *fs_info);
101static void __reada_start_machine(struct btrfs_fs_info *fs_info);
102
103static int reada_add_block(struct reada_control *rc, u64 logical,
104 struct btrfs_key *top, int level, u64 generation);
105
106/* recurses */
107/* in case of err, eb might be NULL */
108static int __readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
109 u64 start, int err)
110{
111 int level = 0;
112 int nritems;
113 int i;
114 u64 bytenr;
115 u64 generation;
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;
121
122 if (eb)
123 level = btrfs_header_level(eb);
124
125 /* find extent */
126 spin_lock(&fs_info->reada_lock);
127 re = radix_tree_lookup(&fs_info->reada_tree, index);
128 if (re)
129 kref_get(&re->refcnt);
130 spin_unlock(&fs_info->reada_lock);
131
132 if (!re)
133 return -1;
134
135 spin_lock(&re->lock);
136 /*
137 * just take the full list from the extent. afterwards we
138 * don't need the lock anymore
139 */
140 list_replace_init(&re->extctl, &list);
141 for_dev = re->scheduled_for;
142 re->scheduled_for = NULL;
143 spin_unlock(&re->lock);
144
145 if (err == 0) {
146 nritems = level ? btrfs_header_nritems(eb) : 0;
147 generation = btrfs_header_generation(eb);
148 /*
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.
153 */
154 } else {
155 /*
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.
160 */
161 nritems = 0;
162 generation = 0;
163 }
164
165 for (i = 0; i < nritems; i++) {
166 struct reada_extctl *rec;
167 u64 n_gen;
168 struct btrfs_key key;
169 struct btrfs_key next_key;
170
171 btrfs_node_key_to_cpu(eb, &key, i);
172 if (i + 1 < nritems)
173 btrfs_node_key_to_cpu(eb, &next_key, i + 1);
174 else
175 next_key = re->top;
176 bytenr = btrfs_node_blockptr(eb, i);
177 n_gen = btrfs_node_ptr_generation(eb, i);
178
179 list_for_each_entry(rec, &list, list) {
180 struct reada_control *rc = rec->rc;
181
182 /*
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
188 */
189#ifdef DEBUG
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);
195 }
196#endif
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,
201 level - 1, n_gen);
202 }
203 }
204 /*
205 * free extctl records
206 */
207 while (!list_empty(&list)) {
208 struct reada_control *rc;
209 struct reada_extctl *rec;
210
211 rec = list_first_entry(&list, struct reada_extctl, list);
212 list_del(&rec->list);
213 rc = rec->rc;
214 kfree(rec);
215
216 kref_get(&rc->refcnt);
217 if (atomic_dec_and_test(&rc->elems)) {
218 kref_put(&rc->refcnt, reada_control_release);
219 wake_up(&rc->wait);
220 }
221 kref_put(&rc->refcnt, reada_control_release);
222
223 reada_extent_put(fs_info, re); /* one ref for each entry */
224 }
225 reada_extent_put(fs_info, re); /* our ref */
226 if (for_dev)
227 atomic_dec(&for_dev->reada_in_flight);
228
229 return 0;
230}
231
232/*
233 * start is passed separately in case eb in NULL, which may be the case with
234 * failed I/O
235 */
236int btree_readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
237 u64 start, int err)
238{
239 int ret;
240
241 ret = __readahead_hook(root, eb, start, err);
242
243 reada_start_machine(root->fs_info);
244
245 return ret;
246}
247
248static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info,
249 struct btrfs_device *dev, u64 logical,
250 struct btrfs_multi_bio *multi)
251{
252 int ret;
253 int looped = 0;
254 struct reada_zone *zone;
255 struct btrfs_block_group_cache *cache = NULL;
256 u64 start;
257 u64 end;
258 int i;
259
260again:
261 zone = 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);
265 if (ret == 1)
266 kref_get(&zone->refcnt);
267 spin_unlock(&fs_info->reada_lock);
268
269 if (ret == 1) {
270 if (logical >= zone->start && logical < zone->end)
271 return zone;
272 spin_lock(&fs_info->reada_lock);
273 kref_put(&zone->refcnt, reada_zone_release);
274 spin_unlock(&fs_info->reada_lock);
275 }
276
277 if (looped)
278 return NULL;
279
280 cache = btrfs_lookup_block_group(fs_info, logical);
281 if (!cache)
282 return NULL;
283
284 start = cache->key.objectid;
285 end = start + cache->key.offset - 1;
286 btrfs_put_block_group(cache);
287
288 zone = kzalloc(sizeof(*zone), GFP_NOFS);
289 if (!zone)
290 return NULL;
291
292 zone->start = start;
293 zone->end = end;
294 INIT_LIST_HEAD(&zone->list);
295 spin_lock_init(&zone->lock);
296 zone->locked = 0;
297 kref_init(&zone->refcnt);
298 zone->elems = 0;
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;
303 }
304 zone->ndevs = multi->num_stripes;
305
306 spin_lock(&fs_info->reada_lock);
307 ret = radix_tree_insert(&dev->reada_zones,
308 (unsigned long)zone->end >> PAGE_CACHE_SHIFT,
309 zone);
310 spin_unlock(&fs_info->reada_lock);
311
312 if (ret) {
313 kfree(zone);
314 looped = 1;
315 goto again;
316 }
317
318 return zone;
319}
320
321static struct reada_extent *reada_find_extent(struct btrfs_root *root,
322 u64 logical,
323 struct btrfs_key *top, int level)
324{
325 int ret;
326 int looped = 0;
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_multi_bio *multi = NULL;
331 struct btrfs_device *dev;
332 u32 blocksize;
333 u64 length;
334 int nzones = 0;
335 int i;
336 unsigned long index = logical >> PAGE_CACHE_SHIFT;
337
338again:
339 spin_lock(&fs_info->reada_lock);
340 re = radix_tree_lookup(&fs_info->reada_tree, index);
341 if (re)
342 kref_get(&re->refcnt);
343 spin_unlock(&fs_info->reada_lock);
344
345 if (re || looped)
346 return re;
347
348 re = kzalloc(sizeof(*re), GFP_NOFS);
349 if (!re)
350 return NULL;
351
352 blocksize = btrfs_level_size(root, level);
353 re->logical = logical;
354 re->blocksize = blocksize;
355 re->top = *top;
356 INIT_LIST_HEAD(&re->extctl);
357 spin_lock_init(&re->lock);
358 kref_init(&re->refcnt);
359
360 /*
361 * map block
362 */
363 length = blocksize;
364 ret = btrfs_map_block(map_tree, REQ_WRITE, logical, &length, &multi, 0);
365 if (ret || !multi || length < blocksize)
366 goto error;
367
368 if (multi->num_stripes > MAX_MIRRORS) {
369 printk(KERN_ERR "btrfs readahead: more than %d copies not "
370 "supported", MAX_MIRRORS);
371 goto error;
372 }
373
374 for (nzones = 0; nzones < multi->num_stripes; ++nzones) {
375 struct reada_zone *zone;
376
377 dev = multi->stripes[nzones].dev;
378 zone = reada_find_zone(fs_info, dev, logical, multi);
379 if (!zone)
380 break;
381
382 re->zones[nzones] = zone;
383 spin_lock(&zone->lock);
384 if (!zone->elems)
385 kref_get(&zone->refcnt);
386 ++zone->elems;
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);
391 }
392 re->nzones = nzones;
393 if (nzones == 0) {
394 /* not a single zone found, error and out */
395 goto error;
396 }
397
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);
401 if (ret) {
402 spin_unlock(&fs_info->reada_lock);
403 if (ret != -ENOMEM) {
404 /* someone inserted the extent in the meantime */
405 looped = 1;
406 }
407 goto error;
408 }
409 for (i = 0; i < nzones; ++i) {
410 dev = multi->stripes[i].dev;
411 ret = radix_tree_insert(&dev->reada_extents, index, re);
412 if (ret) {
413 while (--i >= 0) {
414 dev = multi->stripes[i].dev;
415 BUG_ON(dev == NULL);
416 radix_tree_delete(&dev->reada_extents, index);
417 }
418 BUG_ON(fs_info == NULL);
419 radix_tree_delete(&fs_info->reada_tree, index);
420 spin_unlock(&fs_info->reada_lock);
421 goto error;
422 }
423 }
424 spin_unlock(&fs_info->reada_lock);
425
426 return re;
427
428error:
429 while (nzones) {
430 struct reada_zone *zone;
431
432 --nzones;
433 zone = re->zones[nzones];
434 kref_get(&zone->refcnt);
435 spin_lock(&zone->lock);
436 --zone->elems;
437 if (zone->elems == 0) {
438 /*
439 * no fs_info->reada_lock needed, as this can't be
440 * the last ref
441 */
442 kref_put(&zone->refcnt, reada_zone_release);
443 }
444 spin_unlock(&zone->lock);
445
446 spin_lock(&fs_info->reada_lock);
447 kref_put(&zone->refcnt, reada_zone_release);
448 spin_unlock(&fs_info->reada_lock);
449 }
450 kfree(re);
451 if (looped)
452 goto again;
453 return NULL;
454}
455
456static void reada_kref_dummy(struct kref *kr)
457{
458}
459
460static void reada_extent_put(struct btrfs_fs_info *fs_info,
461 struct reada_extent *re)
462{
463 int i;
464 unsigned long index = re->logical >> PAGE_CACHE_SHIFT;
465
466 spin_lock(&fs_info->reada_lock);
467 if (!kref_put(&re->refcnt, reada_kref_dummy)) {
468 spin_unlock(&fs_info->reada_lock);
469 return;
470 }
471
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];
475
476 radix_tree_delete(&zone->device->reada_extents, index);
477 }
478
479 spin_unlock(&fs_info->reada_lock);
480
481 for (i = 0; i < re->nzones; ++i) {
482 struct reada_zone *zone = re->zones[i];
483
484 kref_get(&zone->refcnt);
485 spin_lock(&zone->lock);
486 --zone->elems;
487 if (zone->elems == 0) {
488 /* no fs_info->reada_lock needed, as this can't be
489 * the last ref */
490 kref_put(&zone->refcnt, reada_zone_release);
491 }
492 spin_unlock(&zone->lock);
493
494 spin_lock(&fs_info->reada_lock);
495 kref_put(&zone->refcnt, reada_zone_release);
496 spin_unlock(&fs_info->reada_lock);
497 }
498 if (re->scheduled_for)
499 atomic_dec(&re->scheduled_for->reada_in_flight);
500
501 kfree(re);
502}
503
504static void reada_zone_release(struct kref *kref)
505{
506 struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
507
508 radix_tree_delete(&zone->device->reada_zones,
509 zone->end >> PAGE_CACHE_SHIFT);
510
511 kfree(zone);
512}
513
514static void reada_control_release(struct kref *kref)
515{
516 struct reada_control *rc = container_of(kref, struct reada_control,
517 refcnt);
518
519 kfree(rc);
520}
521
522static int reada_add_block(struct reada_control *rc, u64 logical,
523 struct btrfs_key *top, int level, u64 generation)
524{
525 struct btrfs_root *root = rc->root;
526 struct reada_extent *re;
527 struct reada_extctl *rec;
528
529 re = reada_find_extent(root, logical, top, level); /* takes one ref */
530 if (!re)
531 return -1;
532
533 rec = kzalloc(sizeof(*rec), GFP_NOFS);
534 if (!rec) {
535 reada_extent_put(root->fs_info, re);
536 return -1;
537 }
538
539 rec->rc = rc;
540 rec->generation = generation;
541 atomic_inc(&rc->elems);
542
543 spin_lock(&re->lock);
544 list_add_tail(&rec->list, &re->extctl);
545 spin_unlock(&re->lock);
546
547 /* leave the ref on the extent */
548
549 return 0;
550}
551
552/*
553 * called with fs_info->reada_lock held
554 */
555static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
556{
557 int i;
558 unsigned long index = zone->end >> PAGE_CACHE_SHIFT;
559
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)
564 peer->locked = lock;
565 }
566}
567
568/*
569 * called with fs_info->reada_lock held
570 */
571static int reada_pick_zone(struct btrfs_device *dev)
572{
573 struct reada_zone *top_zone = NULL;
574 struct reada_zone *top_locked_zone = NULL;
575 u64 top_elems = 0;
576 u64 top_locked_elems = 0;
577 unsigned long index = 0;
578 int ret;
579
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;
584 }
585 /* pick the zone with the most elements */
586 while (1) {
587 struct reada_zone *zone;
588
589 ret = radix_tree_gang_lookup(&dev->reada_zones,
590 (void **)&zone, index, 1);
591 if (ret == 0)
592 break;
593 index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
594 if (zone->locked) {
595 if (zone->elems > top_locked_elems) {
596 top_locked_elems = zone->elems;
597 top_locked_zone = zone;
598 }
599 } else {
600 if (zone->elems > top_elems) {
601 top_elems = zone->elems;
602 top_zone = zone;
603 }
604 }
605 }
606 if (top_zone)
607 dev->reada_curr_zone = top_zone;
608 else if (top_locked_zone)
609 dev->reada_curr_zone = top_locked_zone;
610 else
611 return 0;
612
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);
616
617 return 1;
618}
619
620static int reada_start_machine_dev(struct btrfs_fs_info *fs_info,
621 struct btrfs_device *dev)
622{
623 struct reada_extent *re = NULL;
624 int mirror_num = 0;
625 struct extent_buffer *eb = NULL;
626 u64 logical;
627 u32 blocksize;
628 int ret;
629 int i;
630 int need_kick = 0;
631
632 spin_lock(&fs_info->reada_lock);
633 if (dev->reada_curr_zone == NULL) {
634 ret = reada_pick_zone(dev);
635 if (!ret) {
636 spin_unlock(&fs_info->reada_lock);
637 return 0;
638 }
639 }
640 /*
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
644 */
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);
649 if (!ret) {
650 spin_unlock(&fs_info->reada_lock);
651 return 0;
652 }
653 re = NULL;
654 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
655 dev->reada_next >> PAGE_CACHE_SHIFT, 1);
656 }
657 if (ret == 0) {
658 spin_unlock(&fs_info->reada_lock);
659 return 0;
660 }
661 dev->reada_next = re->logical + re->blocksize;
662 kref_get(&re->refcnt);
663
664 spin_unlock(&fs_info->reada_lock);
665
666 /*
667 * find mirror num
668 */
669 for (i = 0; i < re->nzones; ++i) {
670 if (re->zones[i]->device == dev) {
671 mirror_num = i + 1;
672 break;
673 }
674 }
675 logical = re->logical;
676 blocksize = re->blocksize;
677
678 spin_lock(&re->lock);
679 if (re->scheduled_for == NULL) {
680 re->scheduled_for = dev;
681 need_kick = 1;
682 }
683 spin_unlock(&re->lock);
684
685 reada_extent_put(fs_info, re);
686
687 if (!need_kick)
688 return 0;
689
690 atomic_inc(&dev->reada_in_flight);
691 ret = reada_tree_block_flagged(fs_info->extent_root, logical, blocksize,
692 mirror_num, &eb);
693 if (ret)
694 __readahead_hook(fs_info->extent_root, NULL, logical, ret);
695 else if (eb)
696 __readahead_hook(fs_info->extent_root, eb, eb->start, ret);
697
698 if (eb)
699 free_extent_buffer(eb);
700
701 return 1;
702
703}
704
705static void reada_start_machine_worker(struct btrfs_work *work)
706{
707 struct reada_machine_work *rmw;
708 struct btrfs_fs_info *fs_info;
709
710 rmw = container_of(work, struct reada_machine_work, work);
711 fs_info = rmw->fs_info;
712
713 kfree(rmw);
714
715 __reada_start_machine(fs_info);
716}
717
718static void __reada_start_machine(struct btrfs_fs_info *fs_info)
719{
720 struct btrfs_device *device;
721 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
722 u64 enqueued;
723 u64 total = 0;
724 int i;
725
726 do {
727 enqueued = 0;
728 list_for_each_entry(device, &fs_devices->devices, dev_list) {
729 if (atomic_read(&device->reada_in_flight) <
730 MAX_IN_FLIGHT)
731 enqueued += reada_start_machine_dev(fs_info,
732 device);
733 }
734 total += enqueued;
735 } while (enqueued && total < 10000);
736
737 if (enqueued == 0)
738 return;
739
740 /*
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
745 * the cores.
746 */
747 for (i = 0; i < 2; ++i)
748 reada_start_machine(fs_info);
749}
750
751static void reada_start_machine(struct btrfs_fs_info *fs_info)
752{
753 struct reada_machine_work *rmw;
754
755 rmw = kzalloc(sizeof(*rmw), GFP_NOFS);
756 if (!rmw) {
757 /* FIXME we cannot handle this properly right now */
758 BUG();
759 }
760 rmw->work.func = reada_start_machine_worker;
761 rmw->fs_info = fs_info;
762
763 btrfs_queue_worker(&fs_info->readahead_workers, &rmw->work);
764}
765
766#ifdef DEBUG
767static void dump_devs(struct btrfs_fs_info *fs_info, int all)
768{
769 struct btrfs_device *device;
770 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
771 unsigned long index;
772 int ret;
773 int i;
774 int j;
775 int cnt;
776
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));
781 index = 0;
782 while (1) {
783 struct reada_zone *zone;
784 ret = radix_tree_gang_lookup(&device->reada_zones,
785 (void **)&zone, index, 1);
786 if (ret == 0)
787 break;
788 printk(KERN_DEBUG " zone %llu-%llu elems %llu locked "
789 "%d devs", zone->start, zone->end, zone->elems,
790 zone->locked);
791 for (j = 0; j < zone->ndevs; ++j) {
792 printk(KERN_CONT " %lld",
793 zone->devs[j]->devid);
794 }
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;
800 }
801 cnt = 0;
802 index = 0;
803 while (all) {
804 struct reada_extent *re = NULL;
805
806 ret = radix_tree_gang_lookup(&device->reada_extents,
807 (void **)&re, index, 1);
808 if (ret == 0)
809 break;
810 printk(KERN_DEBUG
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);
815
816 for (i = 0; i < re->nzones; ++i) {
817 printk(KERN_CONT " zone %llu-%llu devs",
818 re->zones[i]->start,
819 re->zones[i]->end);
820 for (j = 0; j < re->zones[i]->ndevs; ++j) {
821 printk(KERN_CONT " %lld",
822 re->zones[i]->devs[j]->devid);
823 }
824 }
825 printk(KERN_CONT "\n");
826 index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
827 if (++cnt > 15)
828 break;
829 }
830 }
831
832 index = 0;
833 cnt = 0;
834 while (all) {
835 struct reada_extent *re = NULL;
836
837 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
838 index, 1);
839 if (ret == 0)
840 break;
841 if (!re->scheduled_for) {
842 index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
843 continue;
844 }
845 printk(KERN_DEBUG
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",
851 re->zones[i]->start,
852 re->zones[i]->end);
853 for (i = 0; i < re->nzones; ++i) {
854 printk(KERN_CONT " zone %llu-%llu devs",
855 re->zones[i]->start,
856 re->zones[i]->end);
857 for (j = 0; j < re->zones[i]->ndevs; ++j) {
858 printk(KERN_CONT " %lld",
859 re->zones[i]->devs[j]->devid);
860 }
861 }
862 }
863 printk(KERN_CONT "\n");
864 index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
865 }
866 spin_unlock(&fs_info->reada_lock);
867}
868#endif
869
870/*
871 * interface
872 */
873struct reada_control *btrfs_reada_add(struct btrfs_root *root,
874 struct btrfs_key *key_start, struct btrfs_key *key_end)
875{
876 struct reada_control *rc;
877 u64 start;
878 u64 generation;
879 int level;
880 struct extent_buffer *node;
881 static struct btrfs_key max_key = {
882 .objectid = (u64)-1,
883 .type = (u8)-1,
884 .offset = (u64)-1
885 };
886
887 rc = kzalloc(sizeof(*rc), GFP_NOFS);
888 if (!rc)
889 return ERR_PTR(-ENOMEM);
890
891 rc->root = root;
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 */
898
899 node = btrfs_root_node(root);
900 start = node->start;
901 level = btrfs_header_level(node);
902 generation = btrfs_header_generation(node);
903 free_extent_buffer(node);
904
905 reada_add_block(rc, start, &max_key, level, generation);
906
907 reada_start_machine(root->fs_info);
908
909 return rc;
910}
911
912#ifdef DEBUG
913int btrfs_reada_wait(void *handle)
914{
915 struct reada_control *rc = handle;
916
917 while (atomic_read(&rc->elems)) {
918 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
919 5 * HZ);
920 dump_devs(rc->root->fs_info, rc->elems < 10 ? 1 : 0);
921 }
922
923 dump_devs(rc->root->fs_info, rc->elems < 10 ? 1 : 0);
924
925 kref_put(&rc->refcnt, reada_control_release);
926
927 return 0;
928}
929#else
930int btrfs_reada_wait(void *handle)
931{
932 struct reada_control *rc = handle;
933
934 while (atomic_read(&rc->elems)) {
935 wait_event(rc->wait, atomic_read(&rc->elems) == 0);
936 }
937
938 kref_put(&rc->refcnt, reada_control_release);
939
940 return 0;
941}
942#endif
943
944void btrfs_reada_detach(void *handle)
945{
946 struct reada_control *rc = handle;
947
948 kref_put(&rc->refcnt, reada_control_release);
949}
diff --git a/fs/btrfs/scrub.c b/fs/btrfs/scrub.c
index 69a600f07763..5bc4ec827b3d 100644
--- a/fs/btrfs/scrub.c
+++ b/fs/btrfs/scrub.c
@@ -29,15 +29,12 @@
29 * any can be found. 29 * any can be found.
30 * 30 *
31 * Future enhancements: 31 * Future enhancements:
32 * - To enhance the performance, better read-ahead strategies for the
33 * extent-tree can be employed.
34 * - In case an unrepairable extent is encountered, track which files are 32 * - In case an unrepairable extent is encountered, track which files are
35 * affected and report them 33 * affected and report them
36 * - In case of a read error on files with nodatasum, map the file and read 34 * - In case of a read error on files with nodatasum, map the file and read
37 * the extent to trigger a writeback of the good copy 35 * the extent to trigger a writeback of the good copy
38 * - track and record media errors, throw out bad devices 36 * - track and record media errors, throw out bad devices
39 * - add a mode to also read unallocated space 37 * - add a mode to also read unallocated space
40 * - make the prefetch cancellable
41 */ 38 */
42 39
43struct scrub_bio; 40struct scrub_bio;
@@ -741,13 +738,16 @@ static noinline_for_stack int scrub_stripe(struct scrub_dev *sdev,
741 int slot; 738 int slot;
742 int i; 739 int i;
743 u64 nstripes; 740 u64 nstripes;
744 int start_stripe;
745 struct extent_buffer *l; 741 struct extent_buffer *l;
746 struct btrfs_key key; 742 struct btrfs_key key;
747 u64 physical; 743 u64 physical;
748 u64 logical; 744 u64 logical;
749 u64 generation; 745 u64 generation;
750 u64 mirror_num; 746 u64 mirror_num;
747 struct reada_control *reada1;
748 struct reada_control *reada2;
749 struct btrfs_key key_start;
750 struct btrfs_key key_end;
751 751
752 u64 increment = map->stripe_len; 752 u64 increment = map->stripe_len;
753 u64 offset; 753 u64 offset;
@@ -779,81 +779,67 @@ static noinline_for_stack int scrub_stripe(struct scrub_dev *sdev,
779 if (!path) 779 if (!path)
780 return -ENOMEM; 780 return -ENOMEM;
781 781
782 path->reada = 2;
783 path->search_commit_root = 1; 782 path->search_commit_root = 1;
784 path->skip_locking = 1; 783 path->skip_locking = 1;
785 784
786 /* 785 /*
787 * find all extents for each stripe and just read them to get 786 * trigger the readahead for extent tree csum tree and wait for
788 * them into the page cache 787 * completion. During readahead, the scrub is officially paused
789 * FIXME: we can do better. build a more intelligent prefetching 788 * to not hold off transaction commits
790 */ 789 */
791 logical = base + offset; 790 logical = base + offset;
792 physical = map->stripes[num].physical;
793 ret = 0;
794 for (i = 0; i < nstripes; ++i) {
795 key.objectid = logical;
796 key.type = BTRFS_EXTENT_ITEM_KEY;
797 key.offset = (u64)0;
798
799 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
800 if (ret < 0)
801 goto out_noplug;
802
803 /*
804 * we might miss half an extent here, but that doesn't matter,
805 * as it's only the prefetch
806 */
807 while (1) {
808 l = path->nodes[0];
809 slot = path->slots[0];
810 if (slot >= btrfs_header_nritems(l)) {
811 ret = btrfs_next_leaf(root, path);
812 if (ret == 0)
813 continue;
814 if (ret < 0)
815 goto out_noplug;
816 791
817 break; 792 wait_event(sdev->list_wait,
818 } 793 atomic_read(&sdev->in_flight) == 0);
819 btrfs_item_key_to_cpu(l, &key, slot); 794 atomic_inc(&fs_info->scrubs_paused);
795 wake_up(&fs_info->scrub_pause_wait);
820 796
821 if (key.objectid >= logical + map->stripe_len) 797 /* FIXME it might be better to start readahead at commit root */
822 break; 798 key_start.objectid = logical;
799 key_start.type = BTRFS_EXTENT_ITEM_KEY;
800 key_start.offset = (u64)0;
801 key_end.objectid = base + offset + nstripes * increment;
802 key_end.type = BTRFS_EXTENT_ITEM_KEY;
803 key_end.offset = (u64)0;
804 reada1 = btrfs_reada_add(root, &key_start, &key_end);
805
806 key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
807 key_start.type = BTRFS_EXTENT_CSUM_KEY;
808 key_start.offset = logical;
809 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
810 key_end.type = BTRFS_EXTENT_CSUM_KEY;
811 key_end.offset = base + offset + nstripes * increment;
812 reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
813
814 if (!IS_ERR(reada1))
815 btrfs_reada_wait(reada1);
816 if (!IS_ERR(reada2))
817 btrfs_reada_wait(reada2);
823 818
824 path->slots[0]++; 819 mutex_lock(&fs_info->scrub_lock);
825 } 820 while (atomic_read(&fs_info->scrub_pause_req)) {
826 btrfs_release_path(path); 821 mutex_unlock(&fs_info->scrub_lock);
827 logical += increment; 822 wait_event(fs_info->scrub_pause_wait,
828 physical += map->stripe_len; 823 atomic_read(&fs_info->scrub_pause_req) == 0);
829 cond_resched(); 824 mutex_lock(&fs_info->scrub_lock);
830 } 825 }
826 atomic_dec(&fs_info->scrubs_paused);
827 mutex_unlock(&fs_info->scrub_lock);
828 wake_up(&fs_info->scrub_pause_wait);
831 829
832 /* 830 /*
833 * collect all data csums for the stripe to avoid seeking during 831 * collect all data csums for the stripe to avoid seeking during
834 * the scrub. This might currently (crc32) end up to be about 1MB 832 * the scrub. This might currently (crc32) end up to be about 1MB
835 */ 833 */
836 start_stripe = 0;
837 blk_start_plug(&plug); 834 blk_start_plug(&plug);
838again:
839 logical = base + offset + start_stripe * increment;
840 for (i = start_stripe; i < nstripes; ++i) {
841 ret = btrfs_lookup_csums_range(csum_root, logical,
842 logical + map->stripe_len - 1,
843 &sdev->csum_list, 1);
844 if (ret)
845 goto out;
846 835
847 logical += increment;
848 cond_resched();
849 }
850 /* 836 /*
851 * now find all extents for each stripe and scrub them 837 * now find all extents for each stripe and scrub them
852 */ 838 */
853 logical = base + offset + start_stripe * increment; 839 logical = base + offset;
854 physical = map->stripes[num].physical + start_stripe * map->stripe_len; 840 physical = map->stripes[num].physical;
855 ret = 0; 841 ret = 0;
856 for (i = start_stripe; i < nstripes; ++i) { 842 for (i = 0; i < nstripes; ++i) {
857 /* 843 /*
858 * canceled? 844 * canceled?
859 */ 845 */
@@ -882,11 +868,14 @@ again:
882 atomic_dec(&fs_info->scrubs_paused); 868 atomic_dec(&fs_info->scrubs_paused);
883 mutex_unlock(&fs_info->scrub_lock); 869 mutex_unlock(&fs_info->scrub_lock);
884 wake_up(&fs_info->scrub_pause_wait); 870 wake_up(&fs_info->scrub_pause_wait);
885 scrub_free_csums(sdev);
886 start_stripe = i;
887 goto again;
888 } 871 }
889 872
873 ret = btrfs_lookup_csums_range(csum_root, logical,
874 logical + map->stripe_len - 1,
875 &sdev->csum_list, 1);
876 if (ret)
877 goto out;
878
890 key.objectid = logical; 879 key.objectid = logical;
891 key.type = BTRFS_EXTENT_ITEM_KEY; 880 key.type = BTRFS_EXTENT_ITEM_KEY;
892 key.offset = (u64)0; 881 key.offset = (u64)0;
@@ -982,7 +971,6 @@ next:
982 971
983out: 972out:
984 blk_finish_plug(&plug); 973 blk_finish_plug(&plug);
985out_noplug:
986 btrfs_free_path(path); 974 btrfs_free_path(path);
987 return ret < 0 ? ret : 0; 975 return ret < 0 ? ret : 0;
988} 976}
diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
index c3b45564048e..f1685a2b45c8 100644
--- a/fs/btrfs/volumes.c
+++ b/fs/btrfs/volumes.c
@@ -366,6 +366,14 @@ static noinline int device_list_add(const char *path,
366 } 366 }
367 INIT_LIST_HEAD(&device->dev_alloc_list); 367 INIT_LIST_HEAD(&device->dev_alloc_list);
368 368
369 /* init readahead state */
370 spin_lock_init(&device->reada_lock);
371 device->reada_curr_zone = NULL;
372 atomic_set(&device->reada_in_flight, 0);
373 device->reada_next = 0;
374 INIT_RADIX_TREE(&device->reada_zones, GFP_NOFS & ~__GFP_WAIT);
375 INIT_RADIX_TREE(&device->reada_extents, GFP_NOFS & ~__GFP_WAIT);
376
369 mutex_lock(&fs_devices->device_list_mutex); 377 mutex_lock(&fs_devices->device_list_mutex);
370 list_add_rcu(&device->dev_list, &fs_devices->devices); 378 list_add_rcu(&device->dev_list, &fs_devices->devices);
371 mutex_unlock(&fs_devices->device_list_mutex); 379 mutex_unlock(&fs_devices->device_list_mutex);
diff --git a/fs/btrfs/volumes.h b/fs/btrfs/volumes.h
index 6d866db4e177..2a751246188a 100644
--- a/fs/btrfs/volumes.h
+++ b/fs/btrfs/volumes.h
@@ -92,6 +92,14 @@ struct btrfs_device {
92 struct btrfs_work work; 92 struct btrfs_work work;
93 struct rcu_head rcu; 93 struct rcu_head rcu;
94 struct work_struct rcu_work; 94 struct work_struct rcu_work;
95
96 /* readahead state */
97 spinlock_t reada_lock;
98 atomic_t reada_in_flight;
99 u64 reada_next;
100 struct reada_zone *reada_curr_zone;
101 struct radix_tree_root reada_zones;
102 struct radix_tree_root reada_extents;
95}; 103};
96 104
97struct btrfs_fs_devices { 105struct btrfs_fs_devices {
generated by cgit v1.2.2 (git 2.25.0) at 2025-08-23 09:45:53 -0400