Existing VOP_VPTOCNP() interface has a fatal flow that is critical for

nullfs.  The problem is that resulting vnode is only required to be
held on return from the successfull call to vop, instead of being
referenced.

Nullfs VOP_INACTIVE() method reclaims the vnode, which in combination
with the VOP_VPTOCNP() interface means that the directory vnode
returned from VOP_VPTOCNP() is reclaimed in advance, causing
vn_fullpath() to error with EBADF or like.

Change the interface for VOP_VPTOCNP(), now the dvp must be
referenced. Convert all in-tree implementations of VOP_VPTOCNP(),
which is trivial, because vhold(9) and vref(9) are similar in the
locking prerequisites. Out-of-tree fs implementation of VOP_VPTOCNP(),
if any, should have no trouble with the fix.

Tested by:	pho
Reviewed by:	mckusick
MFC after:	3 weeks (subject of re approval)

1 files changed, 48 insertions, 17 deletions

diff --git a/sys/kern/vfs_cache.c b/sys/kern/vfs_cache.c
index 11efcab097a0..177fd5602e2f 100644
--- a/sys/kern/vfs_cache.c
+++ b/sys/kern/vfs_cache.c
@@ -1068,16 +1068,8 @@ vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen)
 
 	CACHE_RLOCK();
 	error = vn_vptocnp_locked(vp, cred, buf, buflen);
-	if (error == 0) {
-		/*
-		 * vn_vptocnp_locked() dropped hold acquired by
-		 * VOP_VPTOCNP immediately after locking the
-		 * cache. Since we are going to drop the cache rlock,
-		 * re-hold the result.
-		 */
-		vhold(*vp);
+	if (error == 0)
 		CACHE_RUNLOCK();
-	}
 	return (error);
 }
 
@@ -1096,6 +1088,9 @@ vn_vptocnp_locked(struct vnode **vp, struct ucred *cred, char *buf,
 	if (ncp != NULL) {
 		if (*buflen < ncp->nc_nlen) {
 			CACHE_RUNLOCK();
+			vfslocked = VFS_LOCK_GIANT((*vp)->v_mount);
+			vrele(*vp);
+			VFS_UNLOCK_GIANT(vfslocked);
 			numfullpathfail4++;
 			error = ENOMEM;
 			SDT_PROBE(vfs, namecache, fullpath, return, error,
@@ -1106,18 +1101,23 @@ vn_vptocnp_locked(struct vnode **vp, struct ucred *cred, char *buf,
 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
 		SDT_PROBE(vfs, namecache, fullpath, hit, ncp->nc_dvp,
 		    ncp->nc_name, vp, 0, 0);
+		dvp = *vp;
 		*vp = ncp->nc_dvp;
+		vref(*vp);
+		CACHE_RUNLOCK();
+		vfslocked = VFS_LOCK_GIANT(dvp->v_mount);
+		vrele(dvp);
+		VFS_UNLOCK_GIANT(vfslocked);
+		CACHE_RLOCK();
 		return (0);
 	}
 	SDT_PROBE(vfs, namecache, fullpath, miss, vp, 0, 0, 0, 0);
 
-	vhold(*vp);
 	CACHE_RUNLOCK();
 	vfslocked = VFS_LOCK_GIANT((*vp)->v_mount);
 	vn_lock(*vp, LK_SHARED | LK_RETRY);
 	error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen);
-	VOP_UNLOCK(*vp, 0);
-	vdrop(*vp);
+	vput(*vp);
 	VFS_UNLOCK_GIANT(vfslocked);
 	if (error) {
 		numfullpathfail2++;
@@ -1128,16 +1128,20 @@ vn_vptocnp_locked(struct vnode **vp, struct ucred *cred, char *buf,
 
 	*vp = dvp;
 	CACHE_RLOCK();
-	if ((*vp)->v_iflag & VI_DOOMED) {
+	if (dvp->v_iflag & VI_DOOMED) {
 		/* forced unmount */
 		CACHE_RUNLOCK();
-		vdrop(*vp);
+		vfslocked = VFS_LOCK_GIANT(dvp->v_mount);
+		vrele(dvp);
+		VFS_UNLOCK_GIANT(vfslocked);
 		error = ENOENT;
 		SDT_PROBE(vfs, namecache, fullpath, return, error, vp,
 		    NULL, 0, 0);
 		return (error);
 	}
-	vdrop(*vp);
+	/*
+	 * *vp has its use count incremented still.
+	 */
 
 	return (0);
 }
@@ -1149,10 +1153,11 @@ static int
 vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
     char *buf, char **retbuf, u_int buflen)
 {
-	int error, slash_prefixed;
+	int error, slash_prefixed, vfslocked;
 #ifdef KDTRACE_HOOKS
 	struct vnode *startvp = vp;
 #endif
+	struct vnode *vp1;
 
 	buflen--;
 	buf[buflen] = '\0';
@@ -1161,6 +1166,7 @@ vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
 
 	SDT_PROBE(vfs, namecache, fullpath, entry, vp, 0, 0, 0, 0);
 	numfullpathcalls++;
+	vref(vp);
 	CACHE_RLOCK();
 	if (vp->v_type != VDIR) {
 		error = vn_vptocnp_locked(&vp, td->td_ucred, buf, &buflen);
@@ -1168,6 +1174,9 @@ vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
 			return (error);
 		if (buflen == 0) {
 			CACHE_RUNLOCK();
+			vfslocked = VFS_LOCK_GIANT(vp->v_mount);
+			vrele(vp);
+			VFS_UNLOCK_GIANT(vfslocked);
 			return (ENOMEM);
 		}
 		buf[--buflen] = '/';
@@ -1177,16 +1186,29 @@ vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
 		if (vp->v_vflag & VV_ROOT) {
 			if (vp->v_iflag & VI_DOOMED) {	/* forced unmount */
 				CACHE_RUNLOCK();
+				vfslocked = VFS_LOCK_GIANT(vp->v_mount);
+				vrele(vp);
+				VFS_UNLOCK_GIANT(vfslocked);
 				error = ENOENT;
 				SDT_PROBE(vfs, namecache, fullpath, return,
 				    error, vp, NULL, 0, 0);
 				break;
 			}
-			vp = vp->v_mount->mnt_vnodecovered;
+			vp1 = vp->v_mount->mnt_vnodecovered;
+			vref(vp1);
+			CACHE_RUNLOCK();
+			vfslocked = VFS_LOCK_GIANT(vp->v_mount);
+			vrele(vp);
+			VFS_UNLOCK_GIANT(vfslocked);
+			vp = vp1;
+			CACHE_RLOCK();
 			continue;
 		}
 		if (vp->v_type != VDIR) {
 			CACHE_RUNLOCK();
+			vfslocked = VFS_LOCK_GIANT(vp->v_mount);
+			vrele(vp);
+			VFS_UNLOCK_GIANT(vfslocked);
 			numfullpathfail1++;
 			error = ENOTDIR;
 			SDT_PROBE(vfs, namecache, fullpath, return,
@@ -1198,6 +1220,9 @@ vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
 			break;
 		if (buflen == 0) {
 			CACHE_RUNLOCK();
+			vfslocked = VFS_LOCK_GIANT(vp->v_mount);
+			vrele(vp);
+			VFS_UNLOCK_GIANT(vfslocked);
 			error = ENOMEM;
 			SDT_PROBE(vfs, namecache, fullpath, return, error,
 			    startvp, NULL, 0, 0);
@@ -1211,6 +1236,9 @@ vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
 	if (!slash_prefixed) {
 		if (buflen == 0) {
 			CACHE_RUNLOCK();
+			vfslocked = VFS_LOCK_GIANT(vp->v_mount);
+			vrele(vp);
+			VFS_UNLOCK_GIANT(vfslocked);
 			numfullpathfail4++;
 			SDT_PROBE(vfs, namecache, fullpath, return, ENOMEM,
 			    startvp, NULL, 0, 0);
@@ -1220,6 +1248,9 @@ vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir,
 	}
 	numfullpathfound++;
 	CACHE_RUNLOCK();
+	vfslocked = VFS_LOCK_GIANT(vp->v_mount);
+	vrele(vp);
+	VFS_UNLOCK_GIANT(vfslocked);
 
 	SDT_PROBE(vfs, namecache, fullpath, return, 0, startvp, buf + buflen,
 	    0, 0);