This patch adds the "LOCKSHARED" option to namei which causes it to only acquire shared locks on leafs.

The stat() and open() calls have been changed to make use of this new functionality.  Using shared locks in
these cases is sufficient and can significantly reduce their latency if IO is pending to these vnodes.  Also,
this reduces the number of exclusive locks that are floating around in the system, which helps reduce the
number of deadlocks that occur.

A new kernel option "LOOKUP_SHARED" has been added.  It defaults to off so this patch can be turned on for
testing, and should eventually go away once it is proven to be stable.  I have personally been running this
patch for over a year now, so it is believed to be fully stable.

Reviewed by:	jake, obrien
Approved by:	jake

1 files changed, 34 insertions, 0 deletions

diff --git a/sys/kern/vfs_vnops.c b/sys/kern/vfs_vnops.c
index 461cb1e..336086c 100644
--- a/sys/kern/vfs_vnops.c
+++ b/sys/kern/vfs_vnops.c
@@ -105,6 +105,11 @@ vn_open_cred(ndp, flagp, cmode, cred)
 	struct vattr vat;
 	struct vattr *vap = &vat;
 	int mode, fmode, error;
+#ifdef LOOKUP_SHARED
+	int exclusive;	/* The current intended lock state */
+
+	exclusive = 0;
+#endif
 
 restart:
 	fmode = *flagp;
@@ -143,6 +148,9 @@ restart:
 			ASSERT_VOP_LOCKED(ndp->ni_vp, "create");
 			fmode &= ~O_TRUNC;
 			vp = ndp->ni_vp;
+#ifdef LOOKUP_SHARED
+			exclusive = 1;
+#endif
 		} else {
 			if (ndp->ni_dvp == ndp->ni_vp)
 				vrele(ndp->ni_dvp);
@@ -158,8 +166,14 @@ restart:
 		}
 	} else {
 		ndp->ni_cnd.cn_nameiop = LOOKUP;
+#ifdef LOOKUP_SHARED
+		ndp->ni_cnd.cn_flags =
+		    ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) |
+		    LOCKSHARED | LOCKLEAF;
+#else
 		ndp->ni_cnd.cn_flags =
 		    ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
+#endif
 		if ((error = namei(ndp)) != 0)
 			return (error);
 		vp = ndp->ni_vp;
@@ -198,6 +212,21 @@ restart:
 	 * Make sure that a VM object is created for VMIO support.
 	 */
 	if (vn_canvmio(vp) == TRUE) {
+#ifdef LOOKUP_SHARED
+		int flock;
+
+		if (!exclusive && vp->v_object == NULL)
+			VOP_LOCK(vp, LK_UPGRADE, td);
+		/*
+		 * In cases where the object is marked as dead object_create
+		 * will unlock and relock exclusive.  It is safe to call in
+		 * here with a shared lock because we only examine fields that
+		 * the shared lock guarantees will be stable.  In the UPGRADE
+		 * case it is not likely that anyone has used this vnode yet
+		 * so there will be no contention.  The logic after this call
+		 * restores the requested locking state.
+		 */
+#endif
 		if ((error = vfs_object_create(vp, td, cred)) != 0) {
 			VOP_UNLOCK(vp, 0, td);
 			VOP_CLOSE(vp, fmode, cred, td);
@@ -206,6 +235,11 @@ restart:
 			*flagp = fmode;
 			return (error);
 		}
+#ifdef LOOKUP_SHARED
+		flock = VOP_ISLOCKED(vp, td);
+		if (!exclusive && flock == LK_EXCLUSIVE)
+			VOP_LOCK(vp, LK_DOWNGRADE, td);
+#endif
 	}
 
 	if (fmode & FWRITE)