On 09/24/2014 04:21 PM, Christoph Hellwig wrote:
On Tue, Sep 23, 2014 at 07:07:35PM -0400, Martin K. Petersen wrote:
"Heinz" == heinzm <heinzm@xxxxxxxxxx> writes:
Heinz> In case of RAID levels 4,5 and 6 we have to check for the
Heinz> capability to zero data on discards to avoid stripe data
Heinz> corruption.
I'm quite concerned about relying on discard_zeroes_data since it's a
hint and not a hard guarantee.
If we want to provide useful discards through raid we need to make
it a hard guarantee. That is:
- only set it when using a WRITE SAME variant on scsi
- start adding a whitelist for ATA
Relying on zeroout which also works when it doesn't unmap below is not
the answer, especially if it may end up doing a physical write of zeroes
on disk devices. Alternatively we could get rid of discard_zeroes_data
and stop the pretence of any sane behavior from discard, and instead
expose an unmap bit through the blkdev_issue_zeroout stack for those
callers who want to do a discard, but rely on zeroes. The end effect
will be the same, just a different presentation.
Neil already came up with an unmap bitmap to add to the RAID4/5/6
personalities in order to harden MD in this regard.
In order to avoid any related data corruption with any future disk types
it seems to be the approprite solution.
Waiting for him to comment more on this.
Heinz
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel
