NFS: Fix direct WRITE throughput regression
authorChuck Lever <chuck.lever@oracle.com>
Fri, 29 May 2020 18:14:40 +0000 (14:14 -0400)
committerAnna Schumaker <Anna.Schumaker@Netapp.com>
Thu, 11 Jun 2020 17:33:48 +0000 (13:33 -0400)
commitba838a75e73f55a780f1ee896b8e3ecb032dba0f
tree77afff1e2a6b81553cfa3d640411553a473d39d7
parent94afd9c489fe6720ea3f49084125344e22610880
NFS: Fix direct WRITE throughput regression

I measured a 50% throughput regression for large direct writes.

The observed on-the-wire behavior is that the client sends every
NFS WRITE twice: once as an UNSTABLE WRITE plus a COMMIT, and once
as a FILE_SYNC WRITE.

This is because the nfs_write_match_verf() check in
nfs_direct_commit_complete() fails for every WRITE.

Buffered writes use nfs_write_completion(), which sets req->wb_verf
correctly. Direct writes use nfs_direct_write_completion(), which
does not set req->wb_verf at all. This leaves req->wb_verf set to
all zeroes for every direct WRITE, and thus
nfs_direct_commit_completion() always sets NFS_ODIRECT_RESCHED_WRITES.

This fix appears to restore nearly all of the lost performance.

Fixes: 1f28476dcb98 ("NFS: Fix O_DIRECT commit verifier handling")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
fs/nfs/direct.c