On Tue, 2023-02-14 at 14:39 -0500, Benjamin Coddington wrote:

We've observed that if the NFS client experiences a network partition
and
the server revokes the client's state, the client may not revalidate
cached
data for an open file during recovery.  If the file is extended by a
second
client during this network partition, the first client will correctly
update the file's size and attributes during recovery, but another
extending write will discard the second client's data.

I'm having trouble fully understanding your problem description. Is the
issue that both clients are opening the same file with something like
O_WRONLY|O_DSYNC|O_APPEND?

If so, what if the network partition happens during the write() system
call of client 1, so that the page cache is updated but the flush of
the write data ends up being delayed by the partition?
In that case, client 2 doesn't know that client 1 has writes
outstanding so it may write its data to where the server thinks the eof
offset is. However once client 1 is able to recover its open state, it
will still have dirty page cache data that is going to overwrite that
same offset.

In the case where another client opened the file during the network
partition and the server revoked the first client's state, the
recovery can
forego optimizations and instead attempt to avoid corruption.

It's a little tricky to solve this in a per-file way during recovery
without plumbing arguments or introducing new flags.  This patch
side-steps
the per-file complexity by simply checking if the client is within a
NOGRACE recovery window, and if so, invalidates data during the open
recovery.

I don't see how this scenario can ever be made fully safe. If people
care, then we should probably have the open recovery of client 1 fail
altogether in this case (subject to some switch similar to the existing
'recover_lost_locks' kernel option).



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