Re: UBIFS issues within kernel 4.14.69?

[Richard Weinberger <richard@nod.at>]

Koen,

Am Samstag, 15. September 2018, 09:13:09 CEST schrieb Richard Weinberger:
> Koen,
> 
> Am Dienstag, 11. September 2018, 16:26:34 CEST schrieb Koen Vandeputte:
> > 
> > On 2018-09-11 15:46, Koen Vandeputte wrote:
> > > Hi Richard,
> > ...
> > 
> > > I'm only seeing these issues on UBIFS enabled volumes.
> > >
> > > It seems it's related to one of your 5 commits, but I'm still in the 
> > > process of bisecting to find the actual culprit.
> > > As soon as I've found it, I'll let you know, but maybe you already 
> > > have an idea here?
> > >
> > Reverting ("ubifs: xattr: Don't operate on deleted inodes") fixes the 
> > weird issues.
> 
> Thanks for finding that bad commit!
> I fear by fixing one bug I've uncovered another one.
> 
> So, I guess you are using overlayfs?
> Which overlayfs features are you using?

I guess I've figured myself.
overlayfs is using temp files (O_TMPFILE), and a recent overlayfs
feature uses xattrs to indicate directory redirects.
So it can happen that a temp file, which has link count 0, gains
xattrs.

UBIFS models xattrs like regular files in directories. Since you cannot
add new files to a unlinked directory, UBIFS kind of enforced that for
xattrs too.
I say "kind of" because technically it works but can trigger an assertion
in UBIFS's journal code.
Recently I saw this assertion but failed to conclude that xattr operations on
unlinked files are perfectly fine and "fixed" the assert.

The right solution is reverting "ubifs: xattr: Don't operate on deleted inodes"
and removing the false positive asserts from UBIFS' journal code.

Sadly xfstests does not test for that, I'll prepare a new test case.
Maybe other file systems got this wrong too.

Thanks,
//richard