[Linux-cachefs] Adventures in NFS re-exporting

[Trond Myklebust]

On Thu, 2020-10-01 at 06:36 -0400, Jeff Layton wrote:
> On Thu, 2020-10-01 at 01:09 +0100, Daire Byrne wrote:
> > ----- On 30 Sep, 2020, at 20:30, Jeff Layton jlayton at kernel.org
> > wrote:
> > 
> > > On Tue, 2020-09-22 at 13:31 +0100, Daire Byrne wrote:
> > > > Hi,
> > > > 
> > > > I just thought I'd flesh out the other two issues I have found
> > > > with re-exporting
> > > > that are ultimately responsible for the biggest performance
> > > > bottlenecks. And
> > > > both of them revolve around the caching of metadata file
> > > > lookups in the NFS
> > > > client.
> > > > 
> > > > Especially for the case where we are re-exporting a server many
> > > > milliseconds
> > > > away (i.e. on-premise -> cloud), we want to be able to control
> > > > how much the
> > > > client caches metadata and file data so that it's many LAN
> > > > clients all benefit
> > > > from the re-export server only having to do the WAN lookups
> > > > once (within a
> > > > specified coherency time).
> > > > 
> > > > Keeping the file data in the vfs page cache or on disk using
> > > > fscache/cachefiles
> > > > is fairly straightforward, but keeping the metadata cached is
> > > > particularly
> > > > difficult. And without the cached metadata we introduce long
> > > > delays before we
> > > > can serve the already present and locally cached file data to
> > > > many waiting
> > > > clients.
> > > > 
> > > > ----- On 7 Sep, 2020, at 18:31, Daire Byrne daire at dneg.com
> > > > wrote:
> > > > > 2) If we cache metadata on the re-export server using
> > > > > actimeo=3600,nocto we can
> > > > > cut the network packets back to the origin server to zero for
> > > > > repeated lookups.
> > > > > However, if a client of the re-export server walks paths and
> > > > > memory maps those
> > > > > files (i.e. loading an application), the re-export server
> > > > > starts issuing
> > > > > unexpected calls back to the origin server again,
> > > > > ignoring/invalidating the
> > > > > re-export server's NFS client cache. We worked around this
> > > > > this by patching an
> > > > > inode/iversion validity check in inode.c so that the NFS
> > > > > client cache on the
> > > > > re-export server is used. I'm not sure about the correctness
> > > > > of this patch but
> > > > > it works for our corner case.
> > > > 
> > > > If we use actimeo=3600,nocto (say) to mount a remote software
> > > > volume on the
> > > > re-export server, we can successfully cache the loading of
> > > > applications and
> > > > walking of paths directly on the re-export server such that
> > > > after a couple of
> > > > runs, there are practically zero packets back to the
> > > > originating NFS server
> > > > (great!). But, if we then do the same thing on a client which
> > > > is mounting that
> > > > re-export server, the re-export server now starts issuing lots
> > > > of calls back to
> > > > the originating server and invalidating it's client cache
> > > > (bad!).
> > > > 
> > > > I'm not exactly sure why, but the iversion of the inode gets
> > > > changed locally
> > > > (due to atime modification?) most likely via invocation of
> > > > method
> > > > inode_inc_iversion_raw. Each time it gets incremented the
> > > > following call to
> > > > validate attributes detects changes causing it to be reloaded
> > > > from the
> > > > originating server.
> > > > 
> > > 
> > > I'd expect the change attribute to track what's in actual inode
> > > on the
> > > "home" server. The NFS client is supposed to (mostly) keep the
> > > raw
> > > change attribute in its i_version field.
> > > 
> > > The only place we call inode_inc_iversion_raw is in
> > > nfs_inode_add_request, which I don't think you'd be hitting
> > > unless you
> > > were writing to the file while holding a write delegation.
> > > 
> > > What sort of server is hosting the actual data in your setup?
> > 
> > We mostly use RHEL7.6 NFS servers with XFS backed filesystems and a
> > couple of (older) Netapps too. The re-export server is running the
> > latest mainline kernel(s).
> > 
> > As far as I can make out, both these originating (home) server
> > types exhibit a similar (but not exactly the same) effect on the
> > Linux NFS client cache when it is being re-exported and accessed by
> > other clients. I can replicate it when only using a read-only mount
> > at every hop so I don't think that writes are related.
> > 
> > Our RHEL7 NFS servers actually mount XFS with noatime too so any
> > atime updates that might be causing this client invalidation (which
> > is what I initially thought) are ultimately a wasted effort.
> > 
> 
> Ok. I suspect there is a bug here somewhere, but with such a
> complicated
> setup though it's not clear to me where that bug would be though. You
> might need to do some packet sniffing and look at what the servers
> are
> sending for change attributes.
> 
> nfsd4_change_attribute does mix in the ctime, so your hunch about the
> atime may be correct. atime updates imply a ctime update and that
> could
> cause nfsd to continually send a new one, even on files that aren't
> being changed.

No. Ordinary atime updates due to read() do not trigger a ctime or
change attribute update. Only an explicit atime update through, e.g. a
call to utimensat() will do that.

> 
> It might be interesting to doctor nfsd4_change_attribute() to not mix
> in
> the ctime and see whether that improves things. If it does, then we
> may
> want to teach nfsd how to avoid doing that for certain types of
> filesystems.

NACK. That would cause very incorrect behaviour for the change
attribute. It is supposed to change in all circumstances where you
ordinarily see a ctime change.

-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust at hammerspace.com