Re: [PATCH 11/14] xfs: widen ondisk timestamps to deal with y2038 problem

[Amir Goldstein <amir73il@gmail.com>]

On Tue, Jun 2, 2020 at 2:17 AM Darrick J. Wong <darrick.wong@oracle.com> wrote:
>
> On Sun, May 31, 2020 at 03:30:42PM +0300, Amir Goldstein wrote:
> > > @@ -265,17 +278,35 @@ xfs_inode_from_disk(
> > >         if (to->di_version == 3) {
> > >                 inode_set_iversion_queried(inode,
> > >                                            be64_to_cpu(from->di_changecount));
> > > -               xfs_inode_from_disk_timestamp(&to->di_crtime, &from->di_crtime);
> > > +               xfs_inode_from_disk_timestamp(from, &to->di_crtime,
> > > +                               &from->di_crtime);
> > >                 to->di_flags2 = be64_to_cpu(from->di_flags2);
> > >                 to->di_cowextsize = be32_to_cpu(from->di_cowextsize);
> > > +               /*
> > > +                * Convert this inode's timestamps to bigtime format the next
> > > +                * time we write it out to disk.
> > > +                */
> > > +               if (xfs_sb_version_hasbigtime(&mp->m_sb))
> > > +                       to->di_flags2 |= XFS_DIFLAG2_BIGTIME;
> > >         }
> >
> > This feels wrong.
> > incore inode has a union for timestamp.
> > This flag should indicate how the union should be interpreted
> > otherwise it is going to be very easy to stumble on that in future code.
>
> Hm?  I think you've gotten it backwards.
>
> The incore inode (struct xfs_icdinode) has a struct timespec64.
>
> The ondisk inode (struct xfs_dinode) has a union xfs_timestamp.
>
> xfs_inode_from_disk_timestamp uses the ondisk inode (from) to convert
> the ondisk timestamp (&from->di_crtime) to the incore inode
> (&to->di_crtime).
>
> In other words, we use the ondisk flag and union to load the ondisk
> timestamp into a format-agnostic incore structure.  Then we set BIGTIME
> in the incore inode (to->di_flags2).
>
> If and when we write the inode back out to disk, we'll see that BIGTIME
> is set in the incore inode, and convert the incore structure into the
> bigtime encoding on disk.
>
> > So either convert incore timestamp now or check hasbigtime when
> > we write to disk.
>
> That's more or less what this is doing.  We read the timestamp in from
> disk in whatever format it's in, and then set ourselves up to write it
> back out in bigtime format.
>

I stand corrected. xfs_ictimestamp got me confused.
Everything does look consistent.

Thanks,
Amir.