Re: [PATCH 1/3] btrfs: tree-checker: Fix false alerts on log trees

[Nikolay Borisov <nborisov@suse.com>]

On 4.10.19 г. 17:13 ч., Filipe Manana wrote:
> On Fri, Oct 4, 2019 at 2:54 PM Nikolay Borisov <nborisov@suse.com> wrote:
>>
>>
>>
>> On 4.10.19 г. 12:31 ч., Qu Wenruo wrote:
>>> [BUG]
>>> When running btrfs/063 in a loop, we got the following random write time
>>> tree checker error:
>>>
>>>   BTRFS critical (device dm-4): corrupt leaf: root=18446744073709551610 block=33095680 slot=2 ino=307 file_offset=0, invalid previous key objectid, have 305 expect 307
>>>   BTRFS info (device dm-4): leaf 33095680 gen 7 total ptrs 47 free space 12146 owner 18446744073709551610
>>>   BTRFS info (device dm-4): refs 1 lock (w:0 r:0 bw:0 br:0 sw:0 sr:0) lock_owner 0 current 26176
>>>           item 0 key (305 1 0) itemoff 16123 itemsize 160
>>>                   inode generation 0 size 0 mode 40777
>>>           item 1 key (305 12 257) itemoff 16111 itemsize 12
>>>           item 2 key (307 108 0) itemoff 16058 itemsize 53 <<<
>>>                   extent data disk bytenr 0 nr 0
>>>                   extent data offset 0 nr 614400 ram 671744
>>>           item 3 key (307 108 614400) itemoff 16005 itemsize 53
>>>                   extent data disk bytenr 195342336 nr 57344
>>>                   extent data offset 0 nr 53248 ram 57344
>>>           item 4 key (307 108 667648) itemoff 15952 itemsize 53
>>>                   extent data disk bytenr 194048000 nr 4096
>>>                   extent data offset 0 nr 4096 ram 4096
>>>         [...]
>>>   BTRFS error (device dm-4): block=33095680 write time tree block corruption detected
>>>   BTRFS: error (device dm-4) in btrfs_commit_transaction:2332: errno=-5 IO failure (Error while writing out transaction)
>>>   BTRFS info (device dm-4): forced readonly
>>>   BTRFS warning (device dm-4): Skipping commit of aborted transaction.
>>>   BTRFS info (device dm-4): use zlib compression, level 3
>>>   BTRFS: error (device dm-4) in cleanup_transaction:1890: errno=-5 IO failure
>>>
>>> [CAUSE]
>>> Commit 59b0d030fb30 ("btrfs: tree-checker: Try to detect missing INODE_ITEM")
>>> assumes all XATTR_ITEM/DIR_INDEX/DIR_ITEM/INODE_REF/EXTENT_DATA items
>>> should have previous key with the same objectid as ino.
>>>
>>> But it's only true for fs trees. For log-tree, we can get above log tree
>>> block where an EXTENT_DATA item has no previous key with the same ino.
>>> As log tree only records modified items, it won't record unmodified
>>> items like INODE_ITEM.
>>>
>>> So this triggers write time tree check warning.
>>>
>>> [FIX]
>>> As a quick fix, check header owner to skip the previous key if it's not
>>> fs tree (log tree doesn't count as fs tree).
>>>
>>> This fix is only to be merged as a quick fix.
>>> There will be a more comprehensive fix to refactor the common check into
>>> one function.
>>>
>>> Reported-by: David Sterba <dsterba@suse.com>
>>> Fixes: 59b0d030fb30 ("btrfs: tree-checker: Try to detect missing INODE_ITEM")
>>> Signed-off-by: Qu Wenruo <wqu@suse.com>
>>
>>
>> It's not entirely clear why this bug manifests. My tests show that when
>> we write extents we always update the inode's c/m time so it's always
>> dirtied hence it's logged. OTOH when punching a hole the same thing is
>> valid.
>>
>> Filipe, under what conditions should it be possible to log an
>> EXTENT_DATA item without first logging the inode it belongs to? It seems
>> using the usual write paths (e.g. buffered write and punchole) that's
>> impossible?
> 
> The tests you did are pointless, none of those operations write to a
> log tree, only fsync does that.

You were quick to judge, I tried:
xfs_io -f -c "fpunch 1m 4k" -c "fsync" /media/foo (foo was a 4m, fully
sycned file)

Similar command with the just writing in the middle of the file i.e not
changing isize.

> 
> This change is perfectly fine. Logging (fsync) always logs the inode
> item since commit [1] (2015),
> however it might do so after logging extents and other items, and in
> between that, if writeback for
> the log tree leaf happens we get that error from the tree-checker.

Fair enough, however that clarification about the sequence of events
should be in the changelog.

> 
> [1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=e4545de5b035c7debb73d260c78377dbb69cbfb5
> 
>>
>>> ---
>>>  fs/btrfs/tree-checker.c | 6 ++++--
>>>  1 file changed, 4 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/fs/btrfs/tree-checker.c b/fs/btrfs/tree-checker.c
>>> index b8f82d9be9f0..5e34cd5e3e2e 100644
>>> --- a/fs/btrfs/tree-checker.c
>>> +++ b/fs/btrfs/tree-checker.c
>>> @@ -148,7 +148,8 @@ static int check_extent_data_item(struct extent_buffer *leaf,
>>>        * But if objectids mismatch, it means we have a missing
>>>        * INODE_ITEM.
>>>        */
>>> -     if (slot > 0 && prev_key->objectid != key->objectid) {
>>> +     if (slot > 0 && is_fstree(btrfs_header_owner(leaf)) &&
>>> +         prev_key->objectid != key->objectid) {
>>>               file_extent_err(leaf, slot,
>>>               "invalid previous key objectid, have %llu expect %llu",
>>>                               prev_key->objectid, key->objectid);
>>> @@ -322,7 +323,8 @@ static int check_dir_item(struct extent_buffer *leaf,
>>>       u32 cur = 0;
>>>
>>>       /* Same check as in check_extent_data_item() */
>>> -     if (slot > 0 && prev_key->objectid != key->objectid) {
>>> +     if (slot > 0 && is_fstree(btrfs_header_owner(leaf)) &&
>>> +         prev_key->objectid != key->objectid) {
>>>               dir_item_err(leaf, slot,
>>>               "invalid previous key objectid, have %llu expect %llu",
>>>                            prev_key->objectid, key->objectid);
>>>
> 
> 
>