On Tue, 2023-07-18 at 11:03 -0700, Hugh Dickins wrote:
> On Tue, 18 Jul 2023, Jeff Layton wrote:
> > On Mon, 2023-07-17 at 20:43 -0700, Hugh Dickins wrote:
> > > Hi Jeff,
> > >
> > > I've been unable to run my kernel builds on ext4 on loop0 on tmpfs
> > > swapping load on linux-next recently, on one machine: various kinds
> > > of havoc, most common symptoms being ext4_find_dest_de:2107 errors,
> > > systemd-journald errors, segfaults. But no problem observed running
> > > on a more recent installation.
> > >
> > > Bisected yesterday to 979492850abd ("ext4: convert to ctime accessor
> > > functions").
> > >
> > > I've mostly averted my eyes from the EXT4_INODE macro changes there,
> > > but I think that's where the problem lies. Reading the comment in
> > > fs/ext4/ext4.h above EXT4_FITS_IN_INODE() led me to try "tune2fs -l"
> > > and look at /etc/mke2fs.conf. It's an old installation, its own
> > > inodes are 256, but that old mke2fs.conf does default to 128 for small
> > > FSes, and what I use for the load test is small. Passing -I 256 to the
> > > mkfs makes the problems go away.
> > >
> > > (What's most alarming about the corruption is that it appears to extend
> > > beyond just the throwaway test filesystem: segfaults on bash and libc.so
> > > from the root filesystem. But no permanent damage done there.)
> > >
> > > One oddity I noticed in scrutinizing that commit, didn't help with
> > > the issues above, but there's a hunk in ext4_rename() which changes
> > > - old.dir->i_ctime = old.dir->i_mtime = current_time(old.dir);
> > > + old.dir->i_mtime = inode_set_ctime_current(old.inode);
> > >
> > >
> >
> > I suspect the problem here is the i_crtime, which lives wholly in the
> > extended part of the inode. The old macros would just not store anything
> > if the i_crtime didn't fit, but the new ones would still store the
> > tv_sec field in that case, which could be a memory corruptor. This patch
> > should fix it, and I'm testing it now.
>
> That makes sense.
>
> >
> > Hugh, if you're able to give this a spin on your setup, then that would
> > be most helpful. This is also in the "ctime" branch in my kernel.org
> > tree if that helps. If this looks good, I'll ask Christian to fold this
> > into the ext4 conversion patch.
>
> Yes, it's now running fine on the problem machine, and on the no-problem.
>
> Tested-by: Hugh Dickins <hughd@xxxxxxxxxx>
>
> >
> > Thanks for the bug report!
>
> And thanks for the quick turnaround!
>
> But I'm puzzled by your dismissing that
> - old.dir->i_ctime = old.dir->i_mtime = current_time(old.dir);
> + old.dir->i_mtime = inode_set_ctime_current(old.inode);
> in ext4_rename() as "actually looks fine".
>
> Different issue, nothing to do with the corruption, sure. Much less
> important, sure. But updating ctime on the wrong inode is "fine"?
Ahh , sorry I wasn't looking at that properly. I think you're correct.
The right fix is probably to move ext4 to use generic_rename_timestamp.
I'll test and send another patch for that.
Thanks again!
--
Jeff Layton <jlayton@xxxxxxxxxx>
