From: Jan Kara jack@suse.cz
[ Upstream commit a527c3ba41c4c61e2069bfce4091e5515f06a8dd ]
When we are doing WB_SYNC_ALL writeback, nfs submits write requests with NFS_FILE_SYNC flag to the server (which then generally treats it as an O_SYNC write). This helps to reduce latency for single requests but when submitting more requests, additional fsyncs on the server side hurt latency. NFS generally avoids this additional overhead by not setting NFS_FILE_SYNC if desc->pg_moreio is set.
However this logic doesn't always work. When we do random 4k writes to a huge file and then call fsync(2), each page writeback is going to be sent with NFS_FILE_SYNC because after preparing one page for writeback, we start writing back next, nfs_do_writepage() will call nfs_pageio_cond_complete() which finds the page is not contiguous with previously prepared IO and submits is *without* setting desc->pg_moreio. Hence NFS_FILE_SYNC is used resulting in poor performance.
Fix the problem by setting desc->pg_moreio in nfs_pageio_cond_complete() before submitting outstanding IO. This improves throughput of fsync-after-random-writes on my test SSD from ~70MB/s to ~250MB/s.
Signed-off-by: Jan Kara jack@suse.cz Signed-off-by: Trond Myklebust trond.myklebust@hammerspace.com Signed-off-by: Sasha Levin sashal@kernel.org --- fs/nfs/pagelist.c | 5 +++++ 1 file changed, 5 insertions(+)
diff --git a/fs/nfs/pagelist.c b/fs/nfs/pagelist.c index 6efb5068c116e..040b6b79c75e5 100644 --- a/fs/nfs/pagelist.c +++ b/fs/nfs/pagelist.c @@ -1545,6 +1545,11 @@ void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index) continue; } else if (index == prev->wb_index + 1) continue; + /* + * We will submit more requests after these. Indicate + * this to the underlying layers. + */ + desc->pg_moreio = 1; nfs_pageio_complete(desc); break; }