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: c9b255e88761 ("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>

diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c
index bb63e2b93ff723b56e3ae5c01071d79566585f42..1b79dd5cf66160c4de72c5dcc50c456950a6a6fd 100644 (file)
--- a/fs/nfs/direct.c
+++ b/fs/nfs/direct.c
@@ -731,6 +731,8 @@ static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
                nfs_list_remove_request(req);
                if (request_commit) {
                        kref_get(&req->wb_kref);
+                       memcpy(&req->wb_verf, &hdr->verf.verifier,
+                              sizeof(req->wb_verf));
                        nfs_mark_request_commit(req, hdr->lseg, &cinfo,
                                hdr->ds_commit_idx);
                }