Re: [PATCH v2 02/12] btrfs-progs: do not infinite loop on corrupt keys with lowmem mode

[Qu Wenruo <quwenruo.btrfs@gmx.com>]

On 2021/8/24 上午2:44, Josef Bacik wrote:
> On 8/23/21 11:04 AM, David Sterba wrote:
>> On Thu, Aug 19, 2021 at 01:42:39PM +0800, Qu Wenruo wrote:
>>>
>>>
>>> On 2021/8/19 上午5:33, Josef Bacik wrote:
>>>> By enabling the lowmem checks properly I uncovered the case where test
>>>> 007 will infinite loop at the detection stage.  This is because when
>>>> checking the inode item we will just btrfs_next_item(), and because we
>>>> ignore check tree block failures at read time we don't get an -EIO from
>>>> btrfs_next_leaf.  Generally what check usually does is validate the
>>>> leaves/nodes as we hit them, but in this case we're not doing that.
>>>> Fix
>>>> this by checking the leaf if we move to the next one and if it fails
>>>> bail.  This allows us to pass the 007 test with lowmem.
>>>
>>> Doesn't this mean btrfs_next_item() is not doing what it should do?
>>>
>>> Normally we would expect btrfs_next_item() to return -EIO other than
>>> manually checking the returned leaf.
>>
>> That's an interesting point, I think we rely on that behaviour
>> elsewhere too.
>>
>
> This is the result of how we deal with corrupt blocks.  We will happily
> read corrupt blocks with check, because we expect check to do it's own
> btrfs_check_node/btrfs_check_leaf().  The problem here is that if the
> block is corrupt it's still in cache, and so btrfs_next_leaf() will
> return it because the buffer is marked uptodate.

Shouldn't we prevent the corrupted block from entering the cache?

>
> It looks like search does the extra check_block() specifically to catch
> this case, so I'll fix next_leaf to do the same thing as well.  Thanks,

OK for now I think it's fine to have the extra check.

It won't cause any harm even if we solved the cache problem in the future.

Thanks,
Qu

>
> Josef