Re: RAID 6 corrupted

[Adam Borowski <kilobyte@angband.pl>]

On Wed, May 17, 2017 at 09:52:28AM +0000, Duncan wrote:
> Łukasz Wróblewski posted on Wed, 17 May 2017 10:27:53 +0200 as excerpted:
> > About two years ago I created RAID 6 consisting of 5 disks with BTRFS.
> > One of the disks has crashed.
> > I started to exchange it for another, but I did something wrong.
> > Or at the time, RAID56 support was experimental in BTRFS.

> > Linux localhost 4.10.12-coreos
> 
> For anything raid56 related, you'll need at /least/ 4.11, as it has some 
> major raid56 stability patch updates.

Uhm, am I missing something?
    git shortlog v4.10..v4.11 -- fs/btrfs/
shows nothing related to RAID5/6 (save for a single commit that removes an
unused variable).

> There's still some issues with raid56, the biggest being that unlike most 
> of btrfs, the parity isn't COW, meaning the traditional parity-raid write 
> hole remains as an issue, a BIG issue for btrfs due to its design, but 
> that's not really fixable without a full raid56 rewrite, so it'll remain 
> the case with the existing raid56 mode likely forever, and a new 
> implementation that COWs parity as well may eventually happen.

Ideas like plug extents (real or virtual) would fix this without a format
change.  In fact, I don't see how a format change would help: there's no way
around having every stripe belong to exactly 0 or 1 transactions, as
otherwise it's RMW rather than COW.  That'd require some auto reclaim, but
reclaim in general sounds like a thing we want, with recent research knorrie
did about SSD.

Well, there's RAIDz, but that's no RAID5 but a hybrid between that and
RAID1, and it doesn't play well with btrfs' concept of block groups.

> But the patches that went into 4.11 fix the known existing issues other 
> than the usual write hole, and they're effectively mandatory for any 
> attempt at repair or recovery of existing raid56 once there are issues of 
> any sort.  They're not a guarantee of a fix, but before that, any attempt 
> at a fix has a rather good chance of making the problem worse instead of 
> better, so you really do want 4.11 if you're doing btrfs raid56 at all.

If you mean 4.12-rc1, then yeah, these patches are a big step forward.


Meow!
-- 
Don't be racist.  White, amber or black, all beers should be judged based
solely on their merits.  Heck, even if occasionally a cider applies for a
beer's job, why not?
On the other hand, corpo lager is not a race.