From: Matthew Dillon <dillon@dragonflybsd.org>
Date: Mon, 14 Jul 2008 17:45:49 +0000 (+0000)
Subject: NFS performance fixes.
X-Git-Tag: v2.1.1~904
X-Git-Url: https://gitweb.dragonflybsd.org/dragonfly.git/commitdiff_plain/a482a28a416459195cf2d1169045ac7170a48d20

NFS performance fixes.

* sync on an NFS mount was a big NOP due to a silly bug.

* utimes (setattr w/ mtime-changed) was unconditionally flushing the file,
  causing programs such as cpdup, rsync, rdist, and tar xp to sync on each
  file.

  change it so it does not unconditionally flush the file.
---

diff --git a/sys/vfs/nfs/nfs.h b/sys/vfs/nfs/nfs.h
index c81156478e..9aeb797dc5 100644
--- a/sys/vfs/nfs/nfs.h
+++ b/sys/vfs/nfs/nfs.h
@@ -35,7 +35,7 @@
  *
  *	@(#)nfs.h	8.4 (Berkeley) 5/1/95
  * $FreeBSD: src/sys/nfs/nfs.h,v 1.53.2.5 2002/02/20 01:35:34 iedowse Exp $
- * $DragonFly: src/sys/vfs/nfs/nfs.h,v 1.19 2006/12/23 00:41:29 swildner Exp $
+ * $DragonFly: src/sys/vfs/nfs/nfs.h,v 1.20 2008/07/14 17:45:49 dillon Exp $
  */
 
 #ifndef _NFS_NFS_H_
@@ -305,6 +305,7 @@ extern vm_zone_t nfsmount_zone;
 #endif
 
 extern struct callout nfs_timer_handle;
+extern int nfs_async;
 
 struct uio;
 struct buf;
diff --git a/sys/vfs/nfs/nfs_bio.c b/sys/vfs/nfs/nfs_bio.c
index 818e796a80..9305345dfb 100644
--- a/sys/vfs/nfs/nfs_bio.c
+++ b/sys/vfs/nfs/nfs_bio.c
@@ -35,7 +35,7 @@
  *
  *	@(#)nfs_bio.c	8.9 (Berkeley) 3/30/95
  * $FreeBSD: /repoman/r/ncvs/src/sys/nfsclient/nfs_bio.c,v 1.130 2004/04/14 23:23:55 peadar Exp $
- * $DragonFly: src/sys/vfs/nfs/nfs_bio.c,v 1.43 2008/04/22 18:46:53 dillon Exp $
+ * $DragonFly: src/sys/vfs/nfs/nfs_bio.c,v 1.44 2008/07/14 17:45:49 dillon Exp $
  */
 
 
@@ -1045,6 +1045,11 @@ again:
 		 *
 		 * IO_INVAL appears to be unused.  The idea appears to be
 		 * to turn off caching in this case.  Very odd.  XXX
+		 *
+		 * If nfs_async is set bawrite() will use an unstable write
+		 * (build dirty bufs on the server), so we might as well
+		 * push it out with bawrite().  If nfs_async is not set we
+		 * use bdwrite() to cache dirty bufs on the client.
 		 */
 		if ((np->n_flag & NDONTCACHE) || (ioflag & IO_SYNC)) {
 			if (ioflag & IO_INVAL)
@@ -1057,9 +1062,8 @@ again:
 				if (error)
 					break;
 			}
-		} else if ((n + on) == biosize) {
-			bp->b_flags |= B_ASYNC;
-			bwrite(bp);
+		} else if ((n + on) == biosize && nfs_async) {
+			bawrite(bp);
 		} else {
 			bdwrite(bp);
 		}
diff --git a/sys/vfs/nfs/nfs_serv.c b/sys/vfs/nfs/nfs_serv.c
index ebddb36a1a..d8a3c884d0 100644
--- a/sys/vfs/nfs/nfs_serv.c
+++ b/sys/vfs/nfs/nfs_serv.c
@@ -35,7 +35,7 @@
  *
  *	@(#)nfs_serv.c  8.8 (Berkeley) 7/31/95
  * $FreeBSD: src/sys/nfs/nfs_serv.c,v 1.93.2.6 2002/12/29 18:19:53 dillon Exp $
- * $DragonFly: src/sys/vfs/nfs/nfs_serv.c,v 1.46 2007/11/20 21:36:08 dillon Exp $
+ * $DragonFly: src/sys/vfs/nfs/nfs_serv.c,v 1.47 2008/07/14 17:45:49 dillon Exp $
  */
 
 /*
@@ -134,7 +134,7 @@ static struct timespec	nfsver;
 
 SYSCTL_DECL(_vfs_nfs);
 
-static int nfs_async;
+int nfs_async;
 SYSCTL_INT(_vfs_nfs, OID_AUTO, async, CTLFLAG_RW, &nfs_async, 0, "");
 static int nfs_commit_blks;
 static int nfs_commit_miss;
diff --git a/sys/vfs/nfs/nfs_vfsops.c b/sys/vfs/nfs/nfs_vfsops.c
index f116515a9c..9c35b4a5df 100644
--- a/sys/vfs/nfs/nfs_vfsops.c
+++ b/sys/vfs/nfs/nfs_vfsops.c
@@ -35,7 +35,7 @@
  *
  *	@(#)nfs_vfsops.c	8.12 (Berkeley) 5/20/95
  * $FreeBSD: src/sys/nfs/nfs_vfsops.c,v 1.91.2.7 2003/01/27 20:04:08 dillon Exp $
- * $DragonFly: src/sys/vfs/nfs/nfs_vfsops.c,v 1.51 2007/09/04 00:48:18 dillon Exp $
+ * $DragonFly: src/sys/vfs/nfs/nfs_vfsops.c,v 1.52 2008/07/14 17:45:49 dillon Exp $
  */
 
 #include "opt_bootp.h"
@@ -1130,7 +1130,7 @@ nfs_sync(struct mount *mp, int waitfor)
 	struct scaninfo scaninfo;
 	int error;
 
-	scaninfo.rescan = 0;
+	scaninfo.rescan = 1;
 	scaninfo.waitfor = waitfor;
 	scaninfo.allerror = 0;
 
diff --git a/sys/vfs/nfs/nfs_vnops.c b/sys/vfs/nfs/nfs_vnops.c
index 2b5658ed4f..a05e57d03a 100644
--- a/sys/vfs/nfs/nfs_vnops.c
+++ b/sys/vfs/nfs/nfs_vnops.c
@@ -35,7 +35,7 @@
  *
  *	@(#)nfs_vnops.c	8.16 (Berkeley) 5/27/95
  * $FreeBSD: src/sys/nfs/nfs_vnops.c,v 1.150.2.5 2001/12/20 19:56:28 dillon Exp $
- * $DragonFly: src/sys/vfs/nfs/nfs_vnops.c,v 1.77 2008/06/08 08:38:06 sephe Exp $
+ * $DragonFly: src/sys/vfs/nfs/nfs_vnops.c,v 1.78 2008/07/14 17:45:49 dillon Exp $
  */
 
 
@@ -226,6 +226,10 @@ int nfs_numasync = 0;
 
 SYSCTL_DECL(_vfs_nfs);
 
+static int nfs_flush_on_rename = 1;
+SYSCTL_INT(_vfs_nfs, OID_AUTO, flush_on_rename, CTLFLAG_RW, 
+	   &nfs_flush_on_rename, 0, "flush fvp prior to rename");
+
 static int	nfsaccess_cache_timeout = NFS_DEFATTRTIMO;
 SYSCTL_INT(_vfs_nfs, OID_AUTO, access_cache_timeout, CTLFLAG_RW, 
 	   &nfsaccess_cache_timeout, 0, "NFS ACCESS cache timeout");
@@ -670,7 +674,11 @@ nfs_setattr(struct vop_setattr_args *ap)
 	    vap->va_mtime.tv_sec != VNOVAL || vap->va_mode != (mode_t)VNOVAL) &&
 	    (vp->v_mount->mnt_flag & MNT_RDONLY))
 		return (EROFS);
+
 	if (vap->va_size != VNOVAL) {
+		/*
+		 * truncation requested
+		 */
  		switch (vp->v_type) {
  		case VDIR:
  			return (EISDIR);
@@ -729,12 +737,26 @@ again:
 			np->n_vattr.va_size = vap->va_size;
 			break;
 		}
-  	} else if ((vap->va_mtime.tv_sec != VNOVAL ||
-		vap->va_atime.tv_sec != VNOVAL) && (np->n_flag & NLMODIFIED) &&
-		vp->v_type == VREG &&
-  		(error = nfs_vinvalbuf(vp, V_SAVE, 1)) == EINTR
-	) {
-		return (error);
+	} else if ((np->n_flag & NLMODIFIED) && vp->v_type == VREG) {
+		/*
+		 * What to do.  If we are modifying the mtime we lose
+		 * mtime detection of changes made by the server or other
+		 * clients.  But programs like rsync/rdist/cpdup are going
+		 * to call utimes a lot.  We don't want to piecemeal sync.
+		 *
+		 * For now sync if any prior remote changes were detected,
+		 * but allow us to lose track of remote changes made during
+		 * the utimes operation.
+		 */
+		if (np->n_flag & NRMODIFIED)
+			error = nfs_vinvalbuf(vp, V_SAVE, 1);
+		if (error == EINTR)
+			return (error);
+		if (error == 0) {
+			if (vap->va_mtime.tv_sec != VNOVAL) {
+				np->n_mtime = vap->va_mtime.tv_sec;
+			}
+		}
 	}
 	error = nfs_setattrrpc(vp, vap, ap->a_cred, td);
 
@@ -1700,15 +1722,14 @@ nfs_rename(struct vop_old_rename_args *ap)
 	}
 
 	/*
-	 * We have to flush B_DELWRI data prior to renaming
-	 * the file.  If we don't, the delayed-write buffers
-	 * can be flushed out later after the file has gone stale
-	 * under NFSV3.  NFSV2 does not have this problem because
-	 * ( as far as I can tell ) it flushes dirty buffers more
-	 * often.
+	 * We shouldn't have to flush fvp on rename as the file handle should
+	 * not change, but the default is to do so.
+	 *
+	 * We must flush tvp on rename because it might become stale on the
+	 * server after the rename.
 	 */
-
-	VOP_FSYNC(fvp, MNT_WAIT);
+	if (nfs_flush_on_rename)
+	    VOP_FSYNC(fvp, MNT_WAIT);
 	if (tvp)
 	    VOP_FSYNC(tvp, MNT_WAIT);
 
@@ -3053,9 +3074,8 @@ nfs_flush_bp(struct buf *bp, void *data)
 			}
 			bremfree(bp);
 
-			bp->b_flags |= B_ASYNC;
 			crit_exit();
-			bwrite(bp);
+			bawrite(bp);
 		} else {
 			crit_exit();
 			error = 0;