On 11/20/16 10:50 AM, Liu Bo wrote:
> On Fri, Nov 18, 2016 at 09:52:40PM -0500, Jeff Mahoney wrote:
>> From: Jeff Mahoney <jeffm@xxxxxxxx>
>> Subject: btrfs: Ensure proper sector alignment for
>> btrfs_free_reserved_data_space
>> References: bsc#1005666
>> Patch-mainline: Submitted 18 Nov 2016, linux-btrfs
>>
>> This fixes the WARN_ON on BTRFS_I(inode)->reserved_extents in
>> btrfs_destroy_inode and the WARN_ON on nonzero delalloc bytes on umount
>> with qgroups enabled.
>>
>> I was able to reproduce this by setting up a small (~500kb) quota limit
>> and writing a file one byte at a time until I hit the limit. The warnings
>> would all hit on umount.
>>
>> The root cause is that we would reserve a block-sized range in both
>> the reservation and the quota in btrfs_check_data_free_space, but if we
>> encountered a problem (like e.g. EDQUOT), we would only release the single
>> byte in the qgroup reservation. That caused an iotree state split, which
>> increased the number of outstanding extents, in turn disallowing releasing
>> the metadata reservation.
>>
>> Signed-off-by: Jeff Mahoney <jeffm@xxxxxxxx>
>> ---
>> fs/btrfs/extent-tree.c | 7 +++++++
>> 1 file changed, 7 insertions(+)
>>
>> --- a/fs/btrfs/extent-tree.c
>> +++ b/fs/btrfs/extent-tree.c
>> @@ -3822,6 +3822,13 @@ void btrfs_free_reserved_data_space_noqu
>> */
>> void btrfs_free_reserved_data_space(struct inode *inode, u64 start, u64 len)
>> {
>> + struct btrfs_root *root = BTRFS_I(inode)->root;
>> +
>> + /* Make sure the range is aligned to sectorsize */
>> + len = round_up(start + len, root->sectorsize) -
>> + round_down(start, root->sectorsize);
>> + start = round_down(start, root->sectorsize);
>> +
>> btrfs_free_reserved_data_space_noquota(inode, start, len);
>> btrfs_qgroup_free_data(inode, start, len);
>
> The patch looks reasonable, but I'm afraid btrfs_fallocate can be
> affected since in btrfs_fallocate(), btrfs_qgroup_reserve_data() takes
> 'cur_offset' and 'last_byte - cur_offset' which are possible unaligned
> to root->sectorsize, but if any errors occur during allocation,
> btrfs_qgroup_free_data() in btrfs_free_reserved_data_space() is gonna
> free aligned range and it ends up a negative qgroup value.
Ok, yeah. I was thinking about this later that evening but hadn't
gotten a chance to dig back into it. I think the biggest thing is that
the handling of space reservation and qgroups is way too complicated.
It seems nearly impossible for new bugs *not* to sneak in whenever we
touch them.
For this particular bit, though, both cur_offset and last_byte are
sector aligned in btrfs_fallocate, at least in the current mainline
HEAD. I think fixing up the alignment in the reservation and qgroups
routines is probably the wrong way to do it. Instead, we should expect
the callers to handle the alignment properly and complain very loudly if
they fail to do that. I started in on patches to do that after I
submitted the one above, but wanted to get the fix sent first.
-Jeff
--
Jeff Mahoney
SUSE Labs
Attachment:
signature.asc
Description: OpenPGP digital signature
