From:   Jeff Layton <jlayton@kernel.org>
To:     trond.myklebust@hammerspace.com, chuck.lever@oracle.com
Cc:     willy@infradead.org, linux-nfs@vger.kernel.org
Subject: [PATCH 0/3] nfsd: write verifier fixes and optimization
Date:   Mon, 13 Feb 2023 16:13:42 -0500
Message-Id: <20230213211345.385005-1-jlayton@kernel.org>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

While looking at the recent problems with the fsync during nfsd_file
cleanup, it occured to me that we could greatly simplify and improve
the server's write verifier handling. I also noticed an existing bug
which is fixed in patch #1.

Instead of trying to check for errors via fsync and resetting the write
verifier when we get one, we can just fold the current value of the
inode's errseq_t into the hashed verifier that is generated at startup
time.

Testing this new scheme has been a real challenge. Once a writeback
error is recorded on all local filesystems, further attempts to write to
the inode return -EIO (and some filesystems even flip to r/o) and you
never see the new verifier.

Trond, you originally added the code to make it reset the verifier on a
writeback error. Do you have a good way to test that? Did you guys use
NFSv3 reexport for testing this somehow?

Jeff Layton (3):
  nfsd: copy the whole verifier in nfsd_copy_write_verifier
  errseq: add a new errseq_fetch helper
  nfsd: simplify write verifier handling

 fs/nfsd/filecache.c    | 22 +------------------
 fs/nfsd/netns.h        |  4 ----
 fs/nfsd/nfs4proc.c     | 17 ++++++---------
 fs/nfsd/nfsctl.c       |  1 -
 fs/nfsd/nfssvc.c       | 48 +++++++++++++-----------------------------
 fs/nfsd/trace.h        | 28 ------------------------
 fs/nfsd/vfs.c          | 28 +++++-------------------
 fs/nfsd/vfs.h          |  1 +
 include/linux/errseq.h |  1 +
 lib/errseq.c           | 33 ++++++++++++++++++++++++++++-
 10 files changed, 62 insertions(+), 121 deletions(-)

-- 
2.39.1