On 07/15/2016 11:10 AM, Andrei Borzenkov wrote:

15.07.2016 16:20, Chris Mason пишет:

quoted

quoted

quoted

Interesting, thanks for taking the time to write this up.  Is the
failure specific to scrub?  Or is parity rebuild in general also failing
in this case?

How do you rebuild parity without scrub as long as all devices appear to
be present?

If one block is corrupted, the crcs will fail and the kernel will
rebuild parity when you read the file.  You can also use balance instead
of scrub.

As we have seen recently, btrfs does not compute, stores or verifies
checksum of RAID56 parity. So if parity is corrupted, the only way to
detect and correct it is to use scrub. Balance may work by side effect,
because it simply recomputes parity on new data, but it will not fix
wrong parity on existing data.

Ah, I misread your question  Yes, this is definitely where scrub is the 
best tool.  But even if we have to add debugging to force parity 
recomputation, we should see if the problem is only in scrub or deeper.

-chris