[PATCH v7 12/13] ext4: switch to multigrain timestamps
Jeff Layton
jlayton at kernel.org
Wed Sep 20 07:12:03 PDT 2023
On Wed, 2023-09-20 at 14:48 +0200, Jan Kara wrote:
> On Wed 20-09-23 06:35:18, Jeff Layton wrote:
> > On Wed, 2023-09-20 at 12:17 +0200, Jan Kara wrote:
> > > If I were a sysadmin, I'd rather opt for something like
> > > finegrained timestamps + lazytime (if I needed the finegrained timestamps
> > > functionality). That should avoid the IO overhead of finegrained timestamps
> > > as well and I'd know I can have problems with timestamps only after a
> > > system crash.
> >
> > > I've just got another idea how we could solve the problem: Couldn't we
> > > always just report coarsegrained timestamp to userspace and provide access
> > > to finegrained value only to NFS which should know what it's doing?
> > >
> >
> > I think that'd be hard. First of all, where would we store the second
> > timestamp? We can't just truncate the fine-grained ones to come up with
> > a coarse-grained one. It might also be confusing having nfsd and local
> > filesystems present different attributes.
>
> So what I had in mind (and I definitely miss all the NFS intricacies so the
> idea may be bogus) was that inode->i_ctime would be maintained exactly as
> is now. There will be new (kernel internal at least for now) STATX flag
> STATX_MULTIGRAIN_TS. fill_mg_cmtime() will return timestamp truncated to
> sb->s_time_gran unless STATX_MULTIGRAIN_TS is set. Hence unless you set
> STATX_MULTIGRAIN_TS, there is no difference in the returned timestamps
> compared to the state before multigrain timestamps were introduced. With
> STATX_MULTIGRAIN_TS we return full precision timestamp as stored in the
> inode. Then NFS in fh_fill_pre_attrs() and fh_fill_post_attrs() needs to
> make sure STATX_MULTIGRAIN_TS is set when calling vfs_getattr() to get
> multigrain time.
> I agree nfsd may now be presenting slightly different timestamps than user
> is able to see with stat(2) directly on the filesystem. But is that a
> problem? Essentially it is a similar solution as the mgtime mount option
> but now sysadmin doesn't have to decide on filesystem mount how to report
> timestamps but the stat caller knowingly opts into possibly inconsistent
> (among files) but high precision timestamps. And in the particular NFS
> usecase where stat is called all the time anyway, timestamps will likely
> even be consistent among files.
>
I like this idea...
Would we also need to raise sb->s_time_gran to something corresponding
to HZ on these filesystems? If we truncate the timestamps at a
granularity corresponding to HZ before presenting them via statx and the
like then that should work around the problem with programs that compare
timestamps between inodes.
With NFSv4, when a filesystem doesn't report a STATX_CHANGE_COOKIE, nfsd
will fake one up using the ctime. It's fine for that to use a full fine-
grained timestamp since we don't expect to be able to compare that value
with one of a different inode.
I think we'd want nfsd to present the mtime/ctime values as truncated,
just like we would with a local fs. We could hit the same problem of an
earlier-looking timestamp with NFS if we try to present the actual fine-
grained values to the clients. IOW, I'm convinced that we need to avoid
this behavior in most situations.
If we do this, then we technically don't need the mount option either.
We could still add it though, and have it govern whether fill_mg_cmtime
truncates the timestamps before storing them in the kstat.
--
Jeff Layton <jlayton at kernel.org>
More information about the linux-afs
mailing list