Re: I think he's dead, Jim

[Justin Engwer <justin@mautobu.com>]

So, in my case at least, I'd guess that dropping the kernel from 4.16
to 4.4 combined with a failed disk is what the root cause was.

I've done what little recovery I can of the current state of files
using btrfs restore. Is there a means of rebuilding the metadata using
the existing data on the drives? Can I put that metadata elsewhere in
a different location so not to overwrite anything? I'm thinking of
moving onto destructive recovery at this point anyway.

Cheers,
Justin

On Wed, May 20, 2020 at 11:20 PM Zygo Blaxell
<ce3g8jdj@umail.furryterror.org> wrote:
>
> On Wed, May 20, 2020 at 10:53:19PM +0200, Johannes Hirte wrote:
> > On 2020 Mai 19, Zygo Blaxell wrote:
> > >
> > > Corollary:  Never use space_cache=v1 with raid5 or raid6 data.
> > > space_cache=v1 puts some metadata (free space cache) in data block
> > > groups, so it violates the "never use raid5 or raid6 for metadata" rule.
> > > space_cache=v2 eliminates this problem by storing the free space tree
> > > in metadata block groups.
> > >
> >
> > This should not be a real problem, as the space-cache can be discarded
> > and rebuild anytime. Or do I miss something?
>
> Keep in mind that there are multiple reasons to not use space_cache=v1;
> space_cache=v1 is quite slow, especially on filesystems big enough that
> raid5 is in play, even when it's not recovering from integrity failures.
>
> The free space cache (v1) is stored in nodatacow inodes, so it has all
> the btrfs RAID data integrity problems of nodatasum, plus the parity
> corruption and write hole issues of raid5.  Free space tree (v2) is
> stored in metadata, so it has csums to detect data corruption and transid
> checks for dropped writes, and if you are using raid1 metadata you also
> avoid the parity corruption bug in btrfs's raid5/6 implementation and
> the write hole.  v2 is faster too, especially at commit time.
>
> The probability of undetected space_cache=v1 failure is low, but not zero.
> In the event of failure, the filesystem should detect the error when it
> tries to create new entries in the extent tree--they'll overlap existing
> allocated blocks, and the filesystem will force itself read-only, so
> there should be no permanent damage other than killing any application
> that was writing to the disk at the time.
>
> Come to think of it, though, the space_cache=v1 problems are not specific
> to raid5.  You shouldn't use space_cache=v1 with raid1 or raid10 data
> either, for the same reasons.
>
> In the raid5/6 case it's a bit simpler:   kernels that can't do
> space_cache=v2 (4.4 and earlier) don't have working raid5 recovery either.
>
> > --
> > Regards,
> >   Johannes Hirte
> >



-- 

Justin Engwer
Mautobu Business Services
250-415-3709