Hi

On 2016-07-16 17:51, Jarkko Lavinen wrote:

On 07/12/2016 05:50 PM, Goffredo Baroncelli wrote:

quoted

Using "btrfs insp phy" I developed a script to trigger the bug.

Thank you for the script and all for sharing the raid5 and scrubbing
issues. I have been using two raid5 arrays and ran scrub occasionally
without any problems lately and been in false confidence. I converted
successfully raid5 arrays into raid10 without any glitch.

I tried to modify the shell script so that instead of corrupting data
with dd, a simulated bad block is created with device mapper. Modern
disks are likely to either return the correct data or an error if
they cannot.

You are right; but doing so we are complicating further the test case:
- my tests show what happen when there is a corruption, but the drive behaves well
- your tests show what happen when there is a corruption AND the drive has a failure

I agree that your simulation is more realistic, but I fear that doing so we are complicating the bug finding.

The modified script behaves very much like the original dd version.
With dd version I see wrong data instead of expected data. 

When toy say "I see wrong data", you means with 
1) "cat mnt/out.txt" 
or 2) with "dd if=/dev/loop....." ?

In the first case I see always good data; in the second case I see wrong data but of course no reading error

With simulated bad block I see no data at all instead of expected data
since dd quits on read error.

Jarkko Lavinen

-- 
gpg @keyserver.linux.it: Goffredo Baroncelli <kreijackATinwind.it>
Key fingerprint BBF5 1610 0B64 DAC6 5F7D  17B2 0EDA 9B37 8B82 E0B5