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

6 files changed, 112 insertions, 3 deletions

diff --git a/sys/conf/options b/sys/conf/options
index c3101df..fede7fd 100644
--- a/sys/conf/options
+++ b/sys/conf/options
@@ -383,6 +383,7 @@ BLKDEV_IOSIZE		opt_global.h
 DEBUG			opt_global.h
 DEBUG_LOCKS		opt_global.h
 DEBUG_VFS_LOCKS		opt_global.h
+LOOKUP_SHARED		opt_global.h
 DIAGNOSTIC		opt_global.h
 ENABLE_VFS_IOOPT	opt_global.h
 INVARIANT_SUPPORT	opt_global.h
diff --git a/sys/kern/vfs_cache.c b/sys/kern/vfs_cache.c
index 08e0397..529efd1 100644
--- a/sys/kern/vfs_cache.c
+++ b/sys/kern/vfs_cache.c
@@ -570,8 +570,35 @@ vfs_cache_lookup(ap)
 
 	error = cache_lookup(dvp, vpp, cnp);
 
+#ifdef LOOKUP_SHARED
+	if (!error) {
+		/* We do this because the rest of the system now expects to get
+		 * a shared lock, which is later upgraded if LOCKSHARED is not
+		 * set.  We have so many cases here because of bugs that yield
+		 * inconsistant lock states.  This all badly needs to be fixed
+		 */
+		error = VOP_CACHEDLOOKUP(dvp, vpp, cnp);
+		if (!error) {
+			int flock;
+
+			flock = VOP_ISLOCKED(*vpp, td);
+			if (flock != LK_EXCLUSIVE) {
+				if (flock == 0) {
+					if ((flags & ISLASTCN) &&
+					    (flags & LOCKSHARED))
+						VOP_LOCK(*vpp, LK_SHARED, td);
+					else
+						VOP_LOCK(*vpp, LK_EXCLUSIVE, td);
+				}
+			} else if ((flags & ISLASTCN) && (flags & LOCKSHARED))
+				VOP_LOCK(*vpp, LK_DOWNGRADE, td);
+		}
+		return (error);
+	}
+#else
 	if (!error) 
 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
+#endif
 
 	if (error == ENOENT)
 		return (error);
@@ -585,13 +612,28 @@ vfs_cache_lookup(ap)
 	} else if (flags & ISDOTDOT) {
 		VOP_UNLOCK(dvp, 0, td);
 		cnp->cn_flags |= PDIRUNLOCK;
+#ifdef LOOKUP_SHARED
+		if ((flags & ISLASTCN) && (flags & LOCKSHARED))
+			error = vget(vp, LK_SHARED, td);
+		else
+			error = vget(vp, LK_EXCLUSIVE, td);
+#else
 		error = vget(vp, LK_EXCLUSIVE, td);
+#endif
+
 		if (!error && lockparent && (flags & ISLASTCN)) {
 			if ((error = vn_lock(dvp, LK_EXCLUSIVE, td)) == 0)
 				cnp->cn_flags &= ~PDIRUNLOCK;
 		}
 	} else {
+#ifdef LOOKUP_SHARED
+		if ((flags & ISLASTCN) && (flags & LOCKSHARED))
+			error = vget(vp, LK_SHARED, td);
+		else
+			error = vget(vp, LK_EXCLUSIVE, td);
+#else
 		error = vget(vp, LK_EXCLUSIVE, td);
+#endif
 		if (!lockparent || error || !(flags & ISLASTCN)) {
 			VOP_UNLOCK(dvp, 0, td);
 			cnp->cn_flags |= PDIRUNLOCK;
@@ -616,7 +658,28 @@ vfs_cache_lookup(ap)
 			return (error);
 		cnp->cn_flags &= ~PDIRUNLOCK;
 	}
+#ifdef LOOKUP_SHARED
+	error = VOP_CACHEDLOOKUP(dvp, vpp, cnp);
+
+	if (!error) {
+		int flock = 0;
+
+		flock = VOP_ISLOCKED(*vpp, td);
+		if (flock != LK_EXCLUSIVE) {
+			if (flock == 0) {
+				if ((flags & ISLASTCN) && (flags & LOCKSHARED))
+					VOP_LOCK(*vpp, LK_SHARED, td);
+				else
+					VOP_LOCK(*vpp, LK_EXCLUSIVE, td);
+			}
+		} else if ((flags & ISLASTCN) && (flags & LOCKSHARED))
+			VOP_LOCK(*vpp, LK_DOWNGRADE, td);
+	}
+
+	return (error);
+#else
 	return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
+#endif
 }
 
 
diff --git a/sys/kern/vfs_extattr.c b/sys/kern/vfs_extattr.c
index fd22602..31b1244 100644
--- a/sys/kern/vfs_extattr.c
+++ b/sys/kern/vfs_extattr.c
@@ -2029,8 +2029,13 @@ stat(td, uap)
 	int error;
 	struct nameidata nd;
 
+#ifdef LOOKUP_SHARED
+	NDINIT(&nd, LOOKUP, FOLLOW | LOCKSHARED | LOCKLEAF | NOOBJ,
+	    UIO_USERSPACE, SCARG(uap, path), td);
+#else
 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | NOOBJ, UIO_USERSPACE,
 	    SCARG(uap, path), td);
+#endif
 	if ((error = namei(&nd)) != 0)
 		return (error);
 	error = vn_stat(nd.ni_vp, &sb, td);
diff --git a/sys/kern/vfs_syscalls.c b/sys/kern/vfs_syscalls.c
index fd22602..31b1244 100644
--- a/sys/kern/vfs_syscalls.c
+++ b/sys/kern/vfs_syscalls.c
@@ -2029,8 +2029,13 @@ stat(td, uap)
 	int error;
 	struct nameidata nd;
 
+#ifdef LOOKUP_SHARED
+	NDINIT(&nd, LOOKUP, FOLLOW | LOCKSHARED | LOCKLEAF | NOOBJ,
+	    UIO_USERSPACE, SCARG(uap, path), td);
+#else
 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | NOOBJ, UIO_USERSPACE,
 	    SCARG(uap, path), td);
+#endif
 	if ((error = namei(&nd)) != 0)
 		return (error);
 	error = vn_stat(nd.ni_vp, &sb, td);
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)
diff --git a/sys/sys/namei.h b/sys/sys/namei.h
index 5b133cc..1ae6120 100644
--- a/sys/sys/namei.h
+++ b/sys/sys/namei.h
@@ -113,8 +113,9 @@ struct nameidata {
 #define	NOCACHE		0x0020	/* name must not be left in cache */
 #define	FOLLOW		0x0040	/* follow symbolic links */
 #define	NOOBJ		0x0080	/* don't create object */
+#define	LOCKSHARED	0x0100	/* Shared lock leaf */
 #define	NOFOLLOW	0x0000	/* do not follow symbolic links (pseudo) */
-#define	MODMASK		0x00fc	/* mask of operational modifiers */
+#define	MODMASK		0x01fc	/* mask of operational modifiers */
 /*
  * Namei parameter descriptors.
  *
@@ -129,7 +130,6 @@ struct nameidata {
  * name being sought. The caller is responsible for releasing the
  * buffer and for vrele'ing ni_startdir.
  */
-#define	NOCROSSMOUNT	0x000100 /* do not cross mount points */
 #define	RDONLY		0x000200 /* lookup with read-only semantics */
 #define	HASBUF		0x000400 /* has allocated pathname buffer */
 #define	SAVENAME	0x000800 /* save pathname buffer */
@@ -143,7 +143,8 @@ struct nameidata {
 #define	WILLBEDIR	0x080000 /* new files will be dirs; allow trailing / */
 #define	ISUNICODE	0x100000 /* current component name is unicode*/
 #define	PDIRUNLOCK	0x200000 /* file system lookup() unlocked parent dir */
-#define PARAMASK	0x1fff00 /* mask of parameter descriptors */
+#define	NOCROSSMOUNT	0x400000 /* do not cross mount points */
+#define	PARAMASK	0x3ffe00 /* mask of parameter descriptors */
 
 /*
  * Initialization of an nameidata structure.