This change adds code to detect and fix the issue introduced in the kernel
release 3.17, where creation of read-only snapshots lead to a corrupted
filesystem if they were created at a moment when the source subvolume/snapshot
had orphan items. The issue was that the on-disk root items became incorrect,
referring to the pre orphan cleanup root node instead of the post orphan
cleanup root node.
A test filesystem can be generated with the test case recently submitted for
xfstests/fstests, which is essencially the following (bash script):
workout()
{
ops=$1
procs=$2
num_snapshots=$3
_scratch_mkfs >> $seqres.full 2>&1
_scratch_mount
snapshot_cmd="$BTRFS_UTIL_PROG subvolume snapshot -r $SCRATCH_MNT"
snapshot_cmd="$snapshot_cmd $SCRATCH_MNT/snap_\`date +'%H_%M_%S_%N'\`"
run_check $FSSTRESS_PROG -p $procs \
-x "$snapshot_cmd" -X $num_snapshots -d $SCRATCH_MNT -n $ops
}
ops=10000
procs=4
snapshots=500
workout $ops $procs $snapshots
Example of btrfsck's (btrfs check) behaviour against such filesystem:
$ btrfsck /dev/loop0
root item for root 311, current bytenr 44630016, current gen 60, current level 1, new bytenr 44957696, new gen 61, new level 1
root item for root 1480, current bytenr 1003569152, current gen 1271, current level 1, new bytenr 1004175360, new gen 1272, new level 1
root item for root 1509, current bytenr 1037434880, current gen 1300, current level 1, new bytenr 1038467072, new gen 1301, new level 1
root item for root 1562, current bytenr 33636352, current gen 1354, current level 1, new bytenr 34455552, new gen 1355, new level 1
root item for root 3094, current bytenr 1011712000, current gen 2935, current level 1, new bytenr 1008484352, new gen 2936, new level 1
root item for root 3716, current bytenr 80805888, current gen 3578, current level 1, new bytenr 73515008, new gen 3579, new level 1
root item for root 4085, current bytenr 714031104, current gen 3958, current level 1, new bytenr 716816384, new gen 3959, new level 1
Found 7 roots with an outdated root item.
Please run a filesystem check with the option --repair to fix them.
$ echo $?
1
$ btrfsck --repair /dev/loop0
enabling repair mode
fixing root item for root 311, current bytenr 44630016, current gen 60, current level 1, new bytenr 44957696, new gen 61, new level 1
fixing root item for root 1480, current bytenr 1003569152, current gen 1271, current level 1, new bytenr 1004175360, new gen 1272, new level 1
fixing root item for root 1509, current bytenr 1037434880, current gen 1300, current level 1, new bytenr 1038467072, new gen 1301, new level 1
fixing root item for root 1562, current bytenr 33636352, current gen 1354, current level 1, new bytenr 34455552, new gen 1355, new level 1
fixing root item for root 3094, current bytenr 1011712000, current gen 2935, current level 1, new bytenr 1008484352, new gen 2936, new level 1
fixing root item for root 3716, current bytenr 80805888, current gen 3578, current level 1, new bytenr 73515008, new gen 3579, new level 1
fixing root item for root 4085, current bytenr 714031104, current gen 3958, current level 1, new bytenr 716816384, new gen 3959, new level 1
Fixed 7 roots.
Checking filesystem on /dev/loop0
UUID: 2186e9b9-c977-4a35-9c7b-69c6609d4620
checking extents
checking free space cache
cache and super generation don't match, space cache will be invalidated
checking fs roots
checking csums
checking root refs
found 618537000 bytes used err is 0
total csum bytes: 130824
total tree bytes: 601620480
total fs tree bytes: 580288512
total extent tree bytes: 18464768
btree space waste bytes: 136939144
file data blocks allocated: 34150318080
referenced 27815415808
Btrfs v3.17-rc3-2-gbbe1dd8
$ echo $?
0
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
The following commit:
"btrfs-progs: fsck: remove unfriendly BUG_ON() for searching tree failure"
f495a2ac66
introduced a regression, detected through xfstests/btrfs/054, where
previously a negative return value (-1) was used to mean a particular
root didn't had any parent root, and now, after that change, a negative
value is also used to mean that an error happened. That change also made
the only caller of is_child_root() interpret any negative return value
as an error and therefore incorrectly made the caller leave with an
error, instead of continuing.
This affects only the 3.17 release candidates (3.16 and older releases
don't have this issue).
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Wang Shilong <wangshilong1991@gmail.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
If we have a corrupt block that multiple snapshots point to we will only fix the
guy who originally pointed to the block, and then simply loop forever because we
keep finding the same bad block. So instead lookup all roots that point to this
block, and then search down to the block for each root and fix the block in all
snapshots. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
We can deal with corrupt items by deleting them in a few cases. Fsck can easily
recover from a missing extent item or a dir index item. So if we notice a item
is completely bogus and it is of a key that we know we can repair then just
delete it and carry on. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Usually if we find a bad block during the extent tree stuff we will error out,
but if the bad block is in an fs tree and doens't have extents in it then fsck
may still pass even though the block was complete garbage. So add the check
block logic to the fs root checking so we actually error out of fsck if there is
a bad block. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
A user had a corrupted fs where the items had been shifted improperly. This
patch adds the ability to fix this sort of problem within fsck. We will simply
shift the item over to the proper offset and update the offsets to make sure
they are correct. I tested this with a hand crafted fs that was broken in the
same way as the user, and I've included the file as a new test. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Sometimes we have a dir index and an inode ref that don't agree on the index.
In this case just assume that the inode ref is the ultimate authority on the
subject and delete the dir index. This means we have to not reset index if we
find a mismatched inode ref to make sure we delete the right dir index. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
If our location is bogus in our dir item we were just skipping the thing.
However in this case we want to just delete the dir index, so create a dummy
inode rec using BTRFS_MULTIPLE_OBJECTIDS and just add every backref we find to
the list so we know to straight up delete all of these items. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
We may run across dir indexes that are corrupt in such a way that it makes them
useless, such as having a bad location key or a bad name. In this case we can
just delete dir indexes that don't show up properly and then re-create what we
need. When we delete dir indexes however we need to restart scanning the fs
tree as we could have greated bogus inode recs if the location key was bad, so
set it up so that if we had to delete an dir index we go ahead and free up our
inode recs and return -EAGAIN to check_fs_roots so it knows to restart the loop.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
If we change something while scanning fs-roots we need to redo our search so
that we get valid root items and have valid root cache. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
If we hid a corrupt block that we fix and we restart the fsck loop you will get
lots of noise about duplicate block groups and such. This is because we don't
clear the block group and chunk cache when we do this restart. This patch fixes
that, which is a little tricky since the structs are linked together with
various linked lists, but this passed with a user who was hitting this problem.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
These were added to deal with duplicated functionality within btrfs-progs, but
we specifically copied rbtree.c from the kernel, so move these functions out
into their own file. This will make it easier to keep rbtree.c in sync. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
If we have an inode backref entry then we know enough to add back a missing dir
index. When messing with the inode backrefs we need to do all of that first
before we process the inode recs themselves as we may clear errors on the inode
recs as we fix the directory indexes. This adds the framework for fixing
backref errors and fixes missing dir index issues. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
In a filesystem corrupted by a faulty memory module, btrfsck would get
very confused attempting to access backrefs that weren't data backrefs
as if they were. Besides invoking undefined behavior for accessing
potentially-uninitialized data past the end of objects, or with
dynamic types unrelated with the static types held in the
corresponding memory, it used offsets and lengths from such fields
that did not correspond to anything in the filesystem proper.
Moving the test for full backrefs and checking that they're data
backrefs earlier avoided the crash I was running into, but that was
not enough to make the filesystem complete a successful repair.
Signed-off-by: Alexandre Oliva <oliva@gnu.org>
Signed-off-by: David Sterba <dsterba@suse.cz>
If we attempt to repair a filesystem with metadata blocks that need
recowing, we'll get into an infinite loop repeatedly recowing the
first entry in the list, without ever removing it from the list.
Oops. Fixed.
Signed-off-by: Alexandre Oliva <oliva@gnu.org>
Signed-off-by: David Sterba <dsterba@suse.cz>
If checksum root is corrupted, fsck will get segmentation. This
is because if we fail to load checksum root, root's node is NULL which
cause NULL pointer deferences later.
To fix this problem, we just did something like extent tree rebuilding.
Allocate a new one and clear uptodate flag. We will do sanity check
before fsck going on.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
The reason that we allow partial opening is that sometimes,
we may have some corrupted trees.(for example extent tree), for
fsck repair case, the broken tree may be rebuilt later.
So if users only want to do check but not repair anything, this
patch will make fsck return failure as soon as possible and
tell users that some critial roots have been corrupted.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
When btrfs-progs walk down the tree, it does not check whether the child
node/leaf is valid.
In fact, there is some corrupted image whose csum is all valid but
parent node points to a invalid leaf.
In my case, the parent node in fs tree point to a invalid leaf(gen 11),
whose generation(15) and first key(EXTENT_TREE ROOT_ITEM 0) is
completely invalid, and will cause BUG_ON in process_inode_item().
Unfortunately, we are unable to fix when it happens.
So we can only output meaningful error message and avoid the insane
node/leaf, which is still much better than the original BUG_ON().
Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
We have --init-csum-tree, which just empties the csum tree. I'm not sure why we
would ever need this, but we definitely need to be able to rebuild the csum tree
in some cases. This patch adds the ability to completely rebuild the crc tree
by reading all of the data and adding csum entries for them. This patch doesn't
pay attention to NODATASUM inodes, it'll happily add csums for everything.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
When we have non-inlined extent references, we were failing to find the
corresponding extent item for an existing csum item in the csum tree.
Reproducer:
mkfs.btrfs -f /dev/sdd
mount /dev/sdd /mnt
xfs_io -f -c "falloc 780366 135302" /mnt/foo
xfs_io -c "falloc 327680 151552" /mnt/foo
xfs_io -c "pwrite -S 0xff -b 131072 0 131072" /mnt/foo
sync
for i in `seq 1 40`; do btrfs subvolume snapshot /mnt /mnt/snap$i ; done
umount /mnt
btrfs check /dev/sdd
The check command exited with status 1 and the following output:
Checking filesystem on /dev/sdd
UUID: 2416ab5f-9d71-457e-bb13-a27d4f6b399a
checking extents
checking free space cache
checking fs roots
checking csums
There are no extents for csum range 12980224-12984320
Csum exists for 12980224-12984320 but there is no extent record
found 1388544 bytes used err is 1
total csum bytes: 132
total tree bytes: 704512
total fs tree bytes: 573440
total extent tree bytes: 16384
btree space waste bytes: 564479
file data blocks allocated: 19341312
referenced 14606336
Btrfs v3.14.1-94-g80597e7
After this change it no longer erroneously reports a missing extent for the
csum item and exits with a status of 0.
Also added missing btrfs_prev_leaf() return value checks, as we were ignoring
errors and non-existence of left siblings completely.
Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
When encountering system crash or balance enospc errors,
there maybe still some reloc roots left.
The way we store reloc root is different from fs root:
reloc root's root key(BTRFS_RELOC_TREE_OBJECTID, ROOT_ITEM, objectid)
fs root's root key(objectid, ROOT_ITEM, -1)
reloc data's root key(BTRFS_DATA_RELOC_TREE_OBJECTID, ROOT_ITEM, 0)
So this patch use right key to search corresponding root node, and
avoid using normal fs root cache for reloc roots.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
If btrfsck fail to repair, we hit something like following:
Check tree block failed, want=29442048, have=0
Check tree block failed, want=29442048, have=0
Check tree block failed, want=29442048, have=0
Check tree block failed, want=29442048, have=0
Check tree block failed, want=29442048, have=0
read block failed check_tree_block
found 98304 bytes used err is 1
total csum bytes: 0
total tree bytes: 0
total fs tree bytes: 0
total extent tree bytes: 0
btree space waste bytes: 0
file data blocks allocated: 0
referenced 0
Btrfs v3.14.2-rc2-63-g3944f15
btrfs: transaction.h:38: btrfs_start_transaction: Assertion `!(root->commit_root)' failed.
Aborted (core dumped)
This is because under repair mode, we will start a transaction, and if we error out,
we don't finish this transaction. So in close_ctree(), it will try
to start and commit transaction which causes the above segmentation.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Now btrfsck would hit assertation failure for some searching tree failure.
It is true that filesystem may get some metadata block corrupted,
and btrfsck could not deal with these corruptings. But, Users really
don't want a BUG_ON() here, Instead, just return errors to caller.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Repair mode will commit transaction which will make us
fail to load log tree anymore.
Give a warning to common users, if they really want to
coninue, we will clear out log tree.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
This can not only give some speedups but also avoid forever loop
with a really broken filesystem.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
[BUG]
Some fsfuzzed btrfs image will cause btrfsck segfault.
[REPRODUCER]
Run btrfsck on a csum tree block corrupted image.
[REASON]
check_csums() function call btrfs_search_slot() on csum_tree but doesn't
check whether the csum_tree contains a valid extent_buffer, which causes
the segfault.
[FIX]
Check the csum_root->node before any search.
Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
A user reported a WARN_ON() when trying to run btrfsck --repair on his fs with
bad key ordering. This was because the root that was broken wasn't part of the
transaction yet. We do this open coded thing in a few other places in fsck, so
just make it a helper function and make sure all the places that need to call it
do call it. With this patch he was able to run repair without it dying.
Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
David sent a quick patch that removed a BUG_ON(). I took a peek and
found that the function was already leaking an eb ref and only returned
0. So this fixes the leak and makes the function void and fixes up the
callers.
Accidentally-motivated-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Zach Brown <zab@zabbo.net>
Signed-off-by: David Sterba <dsterba@suse.cz>
Since this patch:
btrfs-progs: move the check_argc_* functions into utils.c
All tools including the independent tools(e.g. btrfs-image, btrfs-convert)
can share the convenience of the check_argc_* functions, so this patch
adopt the argc check functions globally.
Signed-off-by: Gui Hecheng <guihc.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
The qgroup verification code can trivially be extended to provide
extended information on the extents which a subvolume root
references. Along with qgroup-verify, I have found this tool to be
invaluable when tracking down extent references.
The patch adds a switch to the check subcommand '--subvol-extents'
which takes as args a single subvolume id. When run with the switch,
we'll print out each extent that the subvolume references. The extent
printout gives standard extent info you would expect along with
information on which other roots reference it.
Sample output follows - this is a few lines from a run on a subvolume
I've been testing qgroup changes on:
Print extent state for subvolume 281 on /dev/vdb2
UUID: 8203ca66-9858-4e3f-b447-5bbaacf79c02
Offset Len Root Refs Roots
12582912 20480 12 257 279 280 281 282 283 284 285 286 287 288 289
12603392 8192 12 257 279 280 281 282 283 284 285 286 287 288 289
12611584 12288 12 257 279 280 281 282 283 284 285 286 287 288 289
<snip a bunch of extents to show some variety>
124583936 16384 4 281 282 283 280
125075456 16384 4 280 281 282 283
126255104 16384 11 257 280 281 282 283 284 285 286 287 288 289
4763508736 4096 3 279 280 281
In case it wasn't clear, this applies on top of my qgroup verify patch:
"btrfs-progs: add quota group verify code"
A branch with all this can be found on github:
https://github.com/markfasheh/btrfs-progs-patches/tree/qgroup-verify
Please apply,
Signed-off-by: Mark Fasheh <mfasheh@suse.de>
Signed-off-by: David Sterba <dsterba@suse.cz>
Steps to reproduce:
# mkfs.btrfs -f /dev/sda9 -b 2g
# mount /dev/sda9 /mnt
# dd if=/dev/zero of=/mnt/data bs=4k oflag=direct
# btrfs file df /mnt
Data, single: total=1.66GiB, used=1.66GiB
System, single: total=4.00MiB, used=16.00KiB
Metadata, single: total=200.00MiB, used=67.88MiB
For a filesystem without snapshots, 70M metadata, extent
checking eats max memory about 110M, this is a nightmare
for some system with low memory.
It is very likely that extent record can be freed quickly
for a filesystem without snapshots, improve this by trying
if it can free memory after adding data/tree backrefs.
This patch reduces max memory cost from 110M to 40M for
extents checking for the above case.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
There's no reason to assume that the bad key order is in a leaf block,
so accessing level 0 of the path is going to be an error if it's actually
a node block that's bad.
Reported-by: Chris Mason <clm@fb.com>
Signed-off-by: Hugo Mills <hugo@carfax.org.uk>
Signed-off-by: David Sterba <dsterba@suse.cz>
This patch adds an option '--check-data-csum' to verify data checksums.
fsck won't check data csums unless users specify this option explictly.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
This patch adds functionality (in qgroup-verify.c) to compute bytecounts in
subvolume quota groups. The original groups are read in and stored in memory
so that after we compute our own bytecounts, we can compare them with those
on disk. A print function is provided to do this comparison and show the
results on the console.
A 'qgroup check' pass is added to btrfsck. If any subvolume quota groups
differ from what we compute, the differences for them are printed. We also
provide an option '--qgroup-report' which will run only the quota check code
and print a report on all quota groups. Other than making it possible to
verify that our qgroup changes work correctly, this mode can also be used in
xfstests for automated checking after qgroup tests.
This patch does not address the following:
- compressed counts are identical to non compressed, because kernel doesn't
make the distinction yet. Adding the code to verify compressed counts
shouldn't be hard at all though once kernel can do this.
- It is only concerned with subvolume quota groups (like most of
btrfs-progs).
Signed-off-by: Mark Fasheh <mfasheh@suse.de>
Signed-off-by: David Sterba <dsterba@suse.cz>
It complains errno never gets assigned to zero in find-root and since
errno anyway is zero at program started up, lets remove it.
Check "copy is less then zero" isn't possible because strtoull used by
arg_strtou64 wouldn't return -ve number.
Trivial space fixes.
Signed-off-by: Rakesh Pandit <rakesh@tuxera.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Lets use "errors" instead of "error" because more then one ref errors
are possible. Also print error messages for unresolved refs in
check_root_refs.
Signed-off-by: Rakesh Pandit <rakesh@tuxera.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
The following kernel commit changed the definition of the inline function
btrfs_file_extent_inline_len():
commit 514ac8ad8793a097c0c9d89202c642479d6dfa34
Author: Chris Mason <clm@fb.com>
Date: Fri Jan 3 21:07:00 2014 -0800
Btrfs: don't use ram_bytes for uncompressed inline items
If we truncate an uncompressed inline item, ram_bytes isn't updated to reflect
the new size. The fixe uses the size directly from the item header when
reading uncompressed inlines, and also fixes truncate to update the
size as it goes.
Not having this new definition implies that the restore tool might misbehave when
restoring files with an inline extent that got truncated on a kernel older than
release 3.14.
Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Though all tree blocks have same size, we'd better use right
index here.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Two changes:
1.use bit filed for @found_rec
2.u32 is enough to calculate duplicate extent number.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
We still need free allocated cache memory in case error happens.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
@seen cache is used to avoid iterating same block more than once, and
we can not free them until we have finished searching.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Free already allocated memory to item1_data if malloc fails for
item2_data in swap_values. Seems to be a typo from commit 70749a77.
Signed-off-by: Rakesh Pandit <rakesh@tuxera.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
Previously, --init-extent-tree works just because btrfs_lookup_extent_info()
blindly return 0, and this make it work if there are not any *FULL BACKREF*
mode in broken filesystem.
It is just a coincidence that --init-extent-tree option works, let's
do it in the right way firstly.
For now, we have not supported to rebuild extent tree if there are
any *FULL BACKREF* mode which means if there are snapshots with broken
filesystem, avoid using --init-extent-tree option now.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
commit roots won't update root item in tree root if it finds
updated root's bytenr is same as before.
However, this is not right for fsck, we need update tree root in
the following case:
1.overwrite previous root node.
2.reinit reloc data tree, this is because we skip pin relo data
tree before which means we can allocate same block as before.
Fix this by updating tree root ourselves for the above cases.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
There are two bugs when resetting balance:
1.we will skip reinitting reloc data tree if no reloc root found, however
this is not right because we don't pin reloc data tree before.
2.we should insert root dir into reloc data tree,otherwise we will fail
to fsck.
Fix problems by forcely reiniting reloc data root and inserting root dir.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
reset balance need cow block which will insert extent item into
extent tree. If we do this before reinitting extent root, we may
encounter EEIXST.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
To reinit extent root, we need find a free extent, however,
we may have a really corrupted extent tree, so we can't rely
on existed extent tree to cache block group any more.
During test, we fail to reinit extent tree which is because we
can not find a free extent so let's make block group cache ourselves
firstly.
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>
When working with a user who had a broken file system I noticed that we were
reading a bad copy of a block when the other copy was perfectly fine. This is
because we don't keep track of the parent generation for tree blocks, so we just
read whichever copy we damned well please with no regards for which is best.
This fixes this problem by recording the parent generation of the tree block so
we can be sure to read the most correct copy before we check it, which will give
us a better chance of fixing really broken filesystems. Thanks,
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David Sterba <dsterba@suse.cz>
Signed-off-by: Chris Mason <clm@fb.com>