Re: [PATCH v3 1/7] iversion: update comments with info about atime updates

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Tue, 2022-08-30 at 15:43 +0000, Trond Myklebust wrote:
> On Tue, 2022-08-30 at 11:17 -0400, J. Bruce Fields wrote:
> > On Tue, Aug 30, 2022 at 02:58:27PM +0000, Trond Myklebust wrote:
> > > On Tue, 2022-08-30 at 10:44 -0400, J. Bruce Fields wrote:
> > > > On Tue, Aug 30, 2022 at 09:50:02AM -0400, Jeff Layton wrote:
> > > > > On Tue, 2022-08-30 at 09:24 -0400, J. Bruce Fields wrote:
> > > > > > On Tue, Aug 30, 2022 at 07:40:02AM -0400, Jeff Layton wrote:
> > > > > > > Yes, saying only that it must be different is intentional.
> > > > > > > What
> > > > > > > we
> > > > > > > really want is for consumers to treat this as an opaque
> > > > > > > value
> > > > > > > for the
> > > > > > > most part [1]. Therefore an implementation based on hashing
> > > > > > > would
> > > > > > > conform to the spec, I'd think, as long as all of the
> > > > > > > relevant
> > > > > > > info is
> > > > > > > part of the hash.
> > > > > > 
> > > > > > It'd conform, but it might not be as useful as an increasing
> > > > > > value.
> > > > > > 
> > > > > > E.g. a client can use that to work out which of a series of
> > > > > > reordered
> > > > > > write replies is the most recent, and I seem to recall that
> > > > > > can
> > > > > > prevent
> > > > > > unnecessary invalidations in some cases.
> > > > > > 
> > > > > 
> > > > > That's a good point; the linux client does this. That said,
> > > > > NFSv4
> > > > > has a
> > > > > way for the server to advertise its change attribute behavior
> > > > > [1]
> > > > > (though nfsd hasn't implemented this yet).
> > > > 
> > > > It was implemented and reverted.  The issue was that I thought
> > > > nfsd
> > > > should mix in the ctime to prevent the change attribute going
> > > > backwards
> > > > on reboot (see fs/nfsd/nfsfh.h:nfsd4_change_attribute()), but
> > > > Trond
> > > > was
> > > > concerned about the possibility of time going backwards.  See
> > > > 1631087ba872 "Revert "nfsd4: support change_attr_type
> > > > attribute"".
> > > > There's some mailing list discussion to that I'm not turning up
> > > > right
> > > > now.
> > 
> > https://lore.kernel.org/linux-nfs/a6294c25cb5eb98193f609a52aa8f4b5d4e81279.camel@xxxxxxxxxxxxxxx/
> > is what I was thinking of but it isn't actually that interesting.
> > 
> > > My main concern was that some filesystems (e.g. ext3) were failing
> > > to
> > > provide sufficient timestamp resolution to actually label the
> > > resulting
> > > 'change attribute' as being updated monotonically. If the time
> > > stamp
> > > doesn't change when the file data or metadata are changed, then the
> > > client has to perform extra checks to try to figure out whether or
> > > not
> > > its caches are up to date.
> > 
> > That's a different issue from the one you were raising in that
> > discussion.
> > 
> > > > Did NFSv4 add change_attr_type because some implementations
> > > > needed
> > > > the
> > > > unordered case, or because they realized ordering was useful but
> > > > wanted
> > > > to keep backwards compatibility?  I don't know which it was.
> > > 
> > > We implemented it because, as implied above, knowledge of whether
> > > or
> > > not the change attribute behaves monotonically, or strictly
> > > monotonically, enables a number of optimisations.
> > 
> > Of course, but my question was about the value of the old behavior,
> > not
> > about the value of the monotonic behavior.
> > 
> > Put differently, if we could redesign the protocol from scratch would
> > we
> > actually have included the option of non-monotonic behavior?
> > 
> 
> If we could design the filesystems from scratch, we probably would not.
> The protocol ended up being as it is because people were trying to make
> it as easy to implement as possible.
> 
> So if we could design the filesystem from scratch, we would have
> probably designed it along the lines of what AFS does.
> i.e. each explicit change is accompanied by a single bump of the change
> attribute, so that the clients can not only decide the order of the
> resulting changes, but also if they have missed a change (that might
> have been made by a different client).
> 
> However that would be a requirement that is likely to be very specific
> to distributed caches (and hence distributed filesystems). I doubt
> there are many user space applications that would need that high
> precision. Maybe MPI, but that's the only candidate I can think of for
> now?
> 

The fact that NFS kept this more loosely-defined is what allowed us to
elide some of the i_version bumps and regain a fair bit of performance
for local filesystems [1]. If the change attribute had been more
strictly defined like you mention, then that particular optimization
would not have been possible.

This sort of thing is why I'm a fan of not defining this any more
strictly than we require. Later on, maybe we'll come up with a way for
filesystems to advertise that they can offer stronger guarantees.
-- 
Jeff Layton <jlayton@xxxxxxxxxx>

[1]:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=f02a9ad1f15d




[Index of Archives]     [Linux Filesystem Development]     [Linux USB Development]     [Linux Media Development]     [Video for Linux]     [Linux NILFS]     [Linux Audio Users]     [Yosemite Info]     [Linux SCSI]

  Powered by Linux