On Mon, Aug 29, 2011 at 01:40:11PM +0800, 黃一軒 (I-Hsuan Huang) wrote:
> Dear Darrick,
>
> I still met the following failure symptom with your patch.
>
> <2>[ 3.311674] EXT4-fs error (device mmcblk0p23): ext4_iget: inode
> #22944: (comm init) bad extended attribute block 173879342
>
> The validation code looks like not work.
> Is changing the flag enough to force validating extent header?
Extended attribute blocks are not extent tree blocks.
The patch in question only made ext4 complain consistently if the extent tree
got corrupt.
--D
>
> Thanks,
> IH Huang
>
> 2011/8/25 Darrick J. Wong <[email protected]>
>
> > On Wed, Aug 24, 2011 at 05:54:15PM +0800, 黃一軒 (I-Hsuan Huang) wrote:
> > > Dear Darrick,
> > >
> > > I have read your ext4 patch on Patchwork and have a question about your
> > ext4
> > > patch to validate the extent header.
> > > Is you patch related to the Bug 20992:
> > > https://bugzilla.kernel.org/show_bug.cgi?id=20992
> >
> > Yes, I think it would fix that bug. That said, I'm pretty close to posting
> > a
> > patch set to implement full metadata checksuming, which includes a better
> > solution than what's in patchwork. Basically it implements a new
> > bufferhead
> > flag that tracks whether or not a metadata buffer has been validated, and
> > skips
> > the checks if it has.
> >
> > Obviously that won't help you if you overwrite the filesystem while it's
> > mounted... but oh well, at least it'll fix your case. It was meant to fix
> > the
> > case where the first time a bad extent block gets read it fails the check
> > and
> > complains, but subsequent accesses of the same block don't even check.
> >
> > --D
> > >
> > > Ted mentioned that
> > >
> > > *Yep, looks like a bug alright.
> > >
> > > From what I can tell, you were in the middle of async I/O, at the time
> > when
> > > the
> > > disk was corrupted. The problem seemed to come after the I/O was
> > completed,
> > > and ext4_convert_unwritten_extents() was trying to set the initialized
> > bit
> > > on
> > > the extent tree. At that point the extent tree must have gotten
> > corrupted
> > > on
> > > disk, and this seriously confused the extent conversion code, which ended
> > up
> > > passing 0 to ext4_ext_put_in_cache() as the length of the extent, and
> > that
> > > tripped the BUG_ON in ext4_ext_put_in_cache().
> > >
> > > How did you corrupt the file system while it was mounted? Was it via
> > some
> > > dd
> > > to the disk device directly?
> > >
> > > We do have code that checks to make sure the extent tree is sane, but we
> > > skip
> > > it if the data was already in the buffer cache, to save CPU costs. But
> > if
> > > you
> > > wrote to the disk device directly, it would have gone through the buffer
> > > cache,
> > > since the extent tree was already cached, we would have skipped the
> > > validation
> > > step, and that could be the explanation for how the bug got triggered.
> > >
> > > If so, I'm loathe to turn on the validation check unconditionally, since
> > > that
> > > would kill performance. I can probably change the BUG_ON in
> > > ext4_put_in_cache() to rather set the cache state to "invalid", which
> > would
> > > at
> > > least prevent the BUG_ON. The filesystem was probably well and truly
> > > trashed,
> > > though, so sooner or later the ext4 fs code would have hit something to
> > > cause
> > > it to be very unhappy. Hoepfully it would be an ext4_error() call to
> > mark
> > > the
> > > file system as corrupted, as opposed to another BUG_ON.
> > > *
> > > Have a nice day.
> > >
> > > Thanks,
> > > Randall
> >