Re: [man-pages RFC PATCH v4] statx, inode: document the new STATX_INO_VERSION field

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

 



On Sat, 10 Sep 2022, Jeff Layton wrote:
> On Fri, 2022-09-09 at 16:41 +1000, NeilBrown wrote:
> > > On Fri, 09 Sep 2022, Trond Myklebust wrote:
> > > > > On Fri, 2022-09-09 at 01:10 +0000, Trond Myklebust wrote:
> > > > > > > On Fri, 2022-09-09 at 11:07 +1000, NeilBrown wrote:
> > > > > > > > > On Fri, 09 Sep 2022, NeilBrown wrote:
> > > > > > > > > > > On Fri, 09 Sep 2022, Trond Myklebust wrote:
> > > > > > > > > > > 
> > > > > > > > > > > > > 
> > > > > > > > > > > > > IOW: the minimal condition needs to be that for all cases
> > > > > > > > > > > > > below,
> > > > > > > > > > > > > the
> > > > > > > > > > > > > application reads 'state B' as having occurred if any data was
> > > > > > > > > > > > > committed to disk before the crash.
> > > > > > > > > > > > > 
> > > > > > > > > > > > > Application                             Filesystem
> > > > > > > > > > > > > ===========                             =========
> > > > > > > > > > > > > read change attr <- 'state A'
> > > > > > > > > > > > > read data <- 'state A'
> > > > > > > > > > > > >                                         write data -> 'state B'
> > > > > > > > > > > > >                                         <crash>+<reboot>
> > > > > > > > > > > > > read change attr <- 'state B'
> > > > > > > > > > > 
> > > > > > > > > > > The important thing here is to not see 'state A'.  Seeing 'state
> > > > > > > > > > > C'
> > > > > > > > > > > should be acceptable.  Worst case we could merge in wall-clock
> > > > > > > > > > > time
> > > > > > > > > > > of
> > > > > > > > > > > system boot, but the filesystem should be able to be more helpful
> > > > > > > > > > > than
> > > > > > > > > > > that.
> > > > > > > > > > > 
> > > > > > > > > 
> > > > > > > > > Actually, without the crash+reboot it would still be acceptable to
> > > > > > > > > see
> > > > > > > > > "state A" at the end there - but preferably not for long.
> > > > > > > > > From the NFS perspective, the changeid needs to update by the time
> > > > > > > > > of
> > > > > > > > > a
> > > > > > > > > close or unlock (so it is visible to open or lock), but before that
> > > > > > > > > it
> > > > > > > > > is just best-effort.
> > > > > > > 
> > > > > > > Nope. That will inevitably lead to data corruption, since the
> > > > > > > application might decide to use the data from state A instead of
> > > > > > > revalidating it.
> > > > > > > 
> > > > > 
> > > > > The point is, NFS is not the only potential use case for change
> > > > > attributes. We wouldn't be bothering to discuss statx() if it was.
> > > 
> > > My understanding is that it was primarily a desire to add fstests to
> > > exercise the i_version which motivated the statx extension.
> > > Obviously we should prepare for other uses though.
> > > 
> 
> Mainly. Also, userland nfs servers might also like this for obvious
> reasons. For now though, in the v5 set, I've backed off on trying to
> expose this to userland in favor of trying to just clean up the internal
> implementation.
> 
> I'd still like to expose this via statx if possible, but I don't want to
> get too bogged down in interface design just now as we have Real Bugs to
> fix. That patchset should make it simple to expose it later though.
> 
> > > > > 
> > > > > I could be using O_DIRECT, and all the tricks in order to ensure
> > > > > that
> > > > > my stock broker application (to choose one example) has access
> > > > > to the
> > > > > absolute very latest prices when I'm trying to execute a trade.
> > > > > When the filesystem then says 'the prices haven't changed since
> > > > > your
> > > > > last read because the change attribute on the database file is
> > > > > the
> > > > > same' in response to a statx() request with the
> > > > > AT_STATX_FORCE_SYNC
> > > > > flag set, then why shouldn't my application be able to assume it
> > > > > can
> > > > > serve those prices right out of memory instead of having to go
> > > > > to disk?
> > > 
> > > I would think that such an application would be using inotify rather
> > > than having to poll.  But certainly we should have a clear statement
> > > of
> > > quality-of-service parameters in the documentation.
> > > If we agree that perfect atomicity is what we want to promise, and
> > > that
> > > the cost to the filesystem and the statx call is acceptable, then so
> > > be it.
> > > 
> > > My point wasn't to say that atomicity is bad.  It was that:
> > >  - if the i_version change is visible before the change itself is
> > >    visible, then that is a correctness problem.
> > >  - if the i_version change is only visible some time after the
> > > change
> > >    itself is visible, then that is a quality-of-service issue.
> > > I cannot see any room for debating the first.  I do see some room to
> > > debate the second.
> > > 
> > > Cached writes, directory ops, and attribute changes are, I think,
> > > easy
> > > enough to provide truly atomic i_version updates with the change
> > > being
> > > visible.
> > > 
> > > Changes to a shared memory-mapped files is probably the hardest to
> > > provide timely i_version updates for.  We might want to document an
> > > explicit exception for those.  Alternately each request for
> > > i_version
> > > would need to find all pages that are writable, remap them read-only
> > > to
> > > catch future writes, then update i_version if any were writable
> > > (i.e.
> > > ->mkwrite had been called).  That is the only way I can think of to
> > > provide atomicity.
> > > 
> 
> I don't think we really want to make i_version bumps that expensive.
> Documenting that you can't expect perfect consistency vs. mmap with NFS
> seems like the best thing to do. We do our best, but that sort of
> synchronization requires real locking.
> 
> > > O_DIRECT writes are a little easier than mmapped files.  I suspect we
> > > should update the i_version once the device reports that the write is
> > > complete, but a parallel reader could have seem some of the write before
> > > that moment.  True atomicity could only be provided by taking some
> > > exclusive lock that blocked all O_DIRECT writes.  Jeff seems to be
> > > suggesting this, but I doubt the stock broker application would be
> > > willing to make the call in that case.  I don't think I would either.
> 
> Well, only blocked for long enough to run the getattr. Granted, with a
> slow underlying filesystem that can take a while.

Maybe I misunderstand, but this doesn't seem to make much sense.

If you want i_version updates to appear to be atomic w.r.t O_DIRECT
writes, then you need to prevent accessing the i_version while any write
is on-going. At that time there is no meaningful value for i_version.
So you need a lock (At least shared) around the actual write, and you
need an exclusive lock around the get_i_version().
So accessing the i_version would have to wait for all pending O_DIRECT
writes to complete, and would block any new O_DIRECT writes from
starting.

This could be expensive.

There is not currently any locking around O_DIRECT writes.  You cannot
synchronise with them.

The best you can do is update the i_version immediately after all the
O_DIRECT writes in a single request complete.

> 
> To summarize, there are two main uses for the change attr in NFSv4:
> 
> 1/ to provide change_info4 for directory morphing operations (CREATE,
> LINK, OPEN, REMOVE, and RENAME). It turns out that this is already
> atomic in the current nfsd code (AFAICT) by virtue of the fact that we
> hold the i_rwsem exclusively over these operations. The change attr is
> also queried pre and post while the lock is held, so that should ensure
> that we get true atomicity for this.

Yes, directory ops are relatively easy.

> 
> 2/ as an adjunct for the ctime when fetching attributes to validate
> caches. We don't expect perfect consistency between read (and readlike)
> operations and GETATTR, even when they're in the same compound.
> 
> IOW, a READ+GETATTR compound can legally give you a short (or zero-
> length) read, and then the getattr indicates a size that is larger than
> where the READ data stops, due to a write or truncate racing in after
> the read.

I agree that atomicity is neither necessary nor practical.  Ordering is
important though.  I don't think a truncate(0) racing with a READ can
credibly result in a non-zero size AFTER a zero-length read.  A truncate
that extends the size could have that effect though.

> 
> Ideally, the attributes in the GETATTR reply should be consistent
> between themselves though. IOW, all of the attrs should accurately
> represent the state of the file at a single point in time.
> change+size+times+etc. should all be consistent with one another.
> 
> I think we get all of this by taking the inode_lock around the
> vfs_getattr call in nfsd4_encode_fattr. It may not be the most elegant
> solution, but it should give us the atomicity we need, and it doesn't
> require adding extra operations or locking to the write codepaths.

Explicit attribute changes (chown/chmod/utimes/truncate etc) are always
done under the inode lock.  Implicit changes via inode_update_time() are
not (though xfs does take the lock, ext4 doesn't, haven't checked
others).  So taking the inode lock won't ensure those are internally
consistent.

I think using inode_lock_shared() is acceptable.  It doesn't promise
perfect atomicity, but it is probably good enough.

We'd need a good reason to want perfect atomicity to go further, and I
cannot think of one.

NeilBrown


> 
> We could also consider less invasive ways to achieve this (maybe some
> sort of seqretry loop around the vfs_getattr call?), but I'd rather not
> do extra work in the write codepaths if we can get away with it.
> -- 
> Jeff Layton <jlayton@xxxxxxxxxx>
> 
> 




[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