MFC: r259084

For software builds, the NFS client does many small
synchronous (with FILE_SYNC) writes because non-contiguous
byte ranges in the same buffer cache block are being
written. This patch adds a new mount option "noncontigwr"
which allows the non-contiguous byte ranges to be combined,
with the dirty byte range becoming the superset of the bytes
that are dirty, if the file has not been file locked.
This reduces the number of writes significantly for software
builds. The only case where this change might break existing
applications is where an application is writing
non-overlapping byte ranges within the same buffer cache block
of a file from multiple clients concurrently.
Since such an application would normally do file locking on
the file, avoiding the byte range merge for files that have
been file locked should be sufficient for most (maybe all?) cases.

1 files changed, 25 insertions, 10 deletions

diff --git a/sys/fs/nfsclient/nfs_clbio.c b/sys/fs/nfsclient/nfs_clbio.c
index 47c2f35..576fd59 100644
--- a/sys/fs/nfsclient/nfs_clbio.c
+++ b/sys/fs/nfsclient/nfs_clbio.c
@@ -874,7 +874,7 @@ ncl_write(struct vop_write_args *ap)
 	struct vattr vattr;
 	struct nfsmount *nmp = VFSTONFS(vp->v_mount);
 	daddr_t lbn;
-	int bcount;
+	int bcount, noncontig_write, obcount;
 	int bp_cached, n, on, error = 0, error1;
 	size_t orig_resid, local_resid;
 	off_t orig_size, tmp_off;
@@ -1037,7 +1037,15 @@ again:
 		 * unaligned buffer size.
 		 */
 		mtx_lock(&np->n_mtx);
-		if (uio->uio_offset == np->n_size && n) {
+		if ((np->n_flag & NHASBEENLOCKED) == 0 &&
+		    (nmp->nm_flag & NFSMNT_NONCONTIGWR) != 0)
+			noncontig_write = 1;
+		else
+			noncontig_write = 0;
+		if ((uio->uio_offset == np->n_size ||
+		    (noncontig_write != 0 &&
+		    lbn == (np->n_size / biosize) &&
+		    uio->uio_offset + n > np->n_size)) && n) {
 			mtx_unlock(&np->n_mtx);
 			/*
 			 * Get the buffer (in its pre-append state to maintain
@@ -1045,8 +1053,8 @@ again:
 			 * nfsnode after we have locked the buffer to prevent
 			 * readers from reading garbage.
 			 */
-			bcount = on;
-			bp = nfs_getcacheblk(vp, lbn, bcount, td);
+			obcount = np->n_size - (lbn * biosize);
+			bp = nfs_getcacheblk(vp, lbn, obcount, td);
 
 			if (bp != NULL) {
 				long save;
@@ -1058,9 +1066,12 @@ again:
 				mtx_unlock(&np->n_mtx);
 
 				save = bp->b_flags & B_CACHE;
-				bcount += n;
+				bcount = on + n;
 				allocbuf(bp, bcount);
 				bp->b_flags |= save;
+				if (noncontig_write != 0 && on > obcount)
+					vfs_bio_bzero_buf(bp, obcount, on -
+					    obcount);
 			}
 		} else {
 			/*
@@ -1159,19 +1170,23 @@ again:
 		 * area, just update the b_dirtyoff and b_dirtyend,
 		 * otherwise force a write rpc of the old dirty area.
 		 *
+		 * If there has been a file lock applied to this file
+		 * or vfs.nfs.old_noncontig_writing is set, do the following:
 		 * While it is possible to merge discontiguous writes due to
 		 * our having a B_CACHE buffer ( and thus valid read data
 		 * for the hole), we don't because it could lead to
 		 * significant cache coherency problems with multiple clients,
 		 * especially if locking is implemented later on.
 		 *
-		 * As an optimization we could theoretically maintain
-		 * a linked list of discontinuous areas, but we would still
-		 * have to commit them separately so there isn't much
-		 * advantage to it except perhaps a bit of asynchronization.
+		 * If vfs.nfs.old_noncontig_writing is not set and there has
+		 * not been file locking done on this file:
+		 * Relax coherency a bit for the sake of performance and
+		 * expand the current dirty region to contain the new
+		 * write even if it means we mark some non-dirty data as
+		 * dirty.
 		 */
 
-		if (bp->b_dirtyend > 0 &&
+		if (noncontig_write == 0 && bp->b_dirtyend > 0 &&
 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
 			if (bwrite(bp) == EINTR) {
 				error = EINTR;