X-Git-Url: https://git.subgeniuskitty.com/unix-history/.git/blobdiff_plain/79f80c71c54ed07d136009335bd60424662beefa..8a9c17f60a0e7f32fa5b6a8e64ec861e61a5246e:/usr/src/sys/miscfs/union/union_vnops.c

diff --git a/usr/src/sys/miscfs/union/union_vnops.c b/usr/src/sys/miscfs/union/union_vnops.c
index 1cb2e3d3e0..f165b98132 100644
--- a/usr/src/sys/miscfs/union/union_vnops.c
+++ b/usr/src/sys/miscfs/union/union_vnops.c
@@ -8,14 +8,13 @@
  *
  * %sccs.include.redist.c%
  *
- *	@(#)union_vnops.c	1.4 (Berkeley) %G%
+ *	@(#)union_vnops.c	8.10 (Berkeley) %G%
  */
 
 #include <sys/param.h>
 #include <sys/systm.h>
 #include <sys/proc.h>
 #include <sys/file.h>
-#include <sys/filedesc.h>
 #include <sys/time.h>
 #include <sys/types.h>
 #include <sys/vnode.h>
@@ -23,77 +22,81 @@
 #include <sys/namei.h>
 #include <sys/malloc.h>
 #include <sys/buf.h>
-#include "union.h"
+#include <sys/queue.h>
+#include <miscfs/union/union.h>
 
-/*
- * Create a shadow directory in the upper layer.
- * The new vnode is returned locked.
- */
-static int
-union_mkshadow(dvp, cnp, vpp)
-	struct vnode *dvp;
-	struct componentname *cnp;
-	struct vnode *vpp;
-{
-	int error;
-	struct vattr va;
-	struct proc *p = cnp->cn_proc;
+#define FIXUP(un) { \
+	if (((un)->un_flags & UN_ULOCK) == 0) { \
+		union_fixup(un); \
+	} \
+}
 
-	/*
-	 * policy: when creating the shadow directory in the
-	 * upper layer, create it owned by the current user,
-	 * group from parent directory, and mode 777 modified
-	 * by umask (ie mostly identical to the mkdir syscall).
-	 * (jsp, kb)
-	 * TODO: create the directory owned by the user who
-	 * did the mount (um->um_cred).
-	 */
+static void
+union_fixup(un)
+	struct union_node *un;
+{
 
-	VATTR_NULL(&va);
-	va.va_type = VDIR;
-	va.va_mode = UN_DIRMODE &~ p->p_fd->fd_cmask;
-	VOP_UNLOCK(dvp);
-	LEASE_CHECK(dvp, p, p->p_ucred, LEASE_WRITE);
-	VREF(dvp);
-	VOP_LOCK(dvp);
-	error = VOP_MKDIR(dvp, vpp, cnp, &va);
-	VOP_LOCK(dvp);
-	return (error);
+	VOP_LOCK(un->un_uppervp);
+	un->un_flags |= UN_ULOCK;
 }
 
 static int
-union_lookup1(udvp, dvp, vpp, cnp)
+union_lookup1(udvp, dvpp, vpp, cnp)
 	struct vnode *udvp;
-	struct vnode *dvp;
+	struct vnode **dvpp;
 	struct vnode **vpp;
 	struct componentname *cnp;
 {
 	int error;
 	struct vnode *tdvp;
+	struct vnode *dvp;
 	struct mount *mp;
 
-	if (cnp->cn_flags & ISDOTDOT) {
-		for (;;) {
-			if ((dvp->v_flag & VROOT) == 0 ||
-			    (cnp->cn_flags & NOCROSSMOUNT))
-				break;
+	dvp = *dvpp;
 
+	/*
+	 * If stepping up the directory tree, check for going
+	 * back across the mount point, in which case do what
+	 * lookup would do by stepping back down the mount
+	 * hierarchy.
+	 */
+	if (cnp->cn_flags & ISDOTDOT) {
+		while ((dvp != udvp) && (dvp->v_flag & VROOT)) {
+			/*
+			 * Don't do the NOCROSSMOUNT check
+			 * at this level.  By definition,
+			 * union fs deals with namespaces, not
+			 * filesystems.
+			 */
 			tdvp = dvp;
-			dvp = dvp->v_mount->mnt_vnodecovered;
+			*dvpp = dvp = dvp->v_mount->mnt_vnodecovered;
 			vput(tdvp);
 			VREF(dvp);
 			VOP_LOCK(dvp);
 		}
 	}
-	
+
         error = VOP_LOOKUP(dvp, &tdvp, cnp);
 	if (error)
 		return (error);
 
+	/*
+	 * The parent directory will have been unlocked, unless lookup
+	 * found the last component.  In which case, re-lock the node
+	 * here to allow it to be unlocked again (phew) in union_lookup.
+	 */
+	if (dvp != tdvp && !(cnp->cn_flags & ISLASTCN))
+		VOP_LOCK(dvp);
+
 	dvp = tdvp;
+
+	/*
+	 * Lastly check if the current node is a mount point in
+	 * which case walk up the mount hierarchy making sure not to
+	 * bump into the root of the mount tree (ie. dvp != udvp).
+	 */
 	while (dvp != udvp && (dvp->v_type == VDIR) &&
-	       (mp = dvp->v_mountedhere) &&
-	       (cnp->cn_flags & NOCROSSMOUNT) == 0) {
+	       (mp = dvp->v_mountedhere)) {
 
 		if (mp->mnt_flag & MNT_MLOCK) {
 			mp->mnt_flag |= MNT_MWAIT;
@@ -131,13 +134,16 @@ union_lookup(ap)
 	struct union_node *dun = VTOUNION(dvp);
 	struct componentname *cnp = ap->a_cnp;
 	int lockparent = cnp->cn_flags & LOCKPARENT;
+	int rdonly = cnp->cn_flags & RDONLY;
+	struct union_mount *um = MOUNTTOUNIONMOUNT(dvp->v_mount);
+	struct ucred *saved_cred;
 
 	cnp->cn_flags |= LOCKPARENT;
 
 	upperdvp = dun->un_uppervp;
 	lowerdvp = dun->un_lowervp;
-	uppervp = 0;
-	lowervp = 0;
+	uppervp = NULLVP;
+	lowervp = NULLVP;
 
 	/*
 	 * do the lookup in the upper level.
@@ -145,13 +151,12 @@ union_lookup(ap)
 	 * then assume that something special is going
 	 * on and just return that vnode.
 	 */
-	if (upperdvp) {
-		VOP_LOCK(upperdvp);
-		uerror = union_lookup1(
-			MOUNTTOUNIONMOUNT(dvp->v_mount)->um_uppervp,
-			upperdvp, &uppervp, cnp);
-		if (uppervp != upperdvp)
-			VOP_UNLOCK(upperdvp);
+	if (upperdvp != NULLVP) {
+		FIXUP(dun);
+		uerror = union_lookup1(um->um_uppervp, &upperdvp,
+					&uppervp, cnp);
+		/*if (uppervp == upperdvp)
+			dun->un_flags |= UN_KLOCK;*/
 
 		if (cnp->cn_consume != 0) {
 			*ap->a_vpp = uppervp;
@@ -170,18 +175,37 @@ union_lookup(ap)
 	 * back from the upper layer and return the lower vnode
 	 * instead.
 	 */
-	if (lowerdvp) {
+	if (lowerdvp != NULLVP) {
+		int nameiop;
+
 		VOP_LOCK(lowerdvp);
-		lerror = union_lookup1(
-			MOUNTTOUNIONMOUNT(dvp->v_mount)->um_lowervp,
-			lowerdvp, &lowervp, cnp);
+
+		/*
+		 * Only do a LOOKUP on the bottom node, since
+		 * we won't be making changes to it anyway.
+		 */
+		nameiop = cnp->cn_nameiop;
+		cnp->cn_nameiop = LOOKUP;
+		if (um->um_op == UNMNT_BELOW) {
+			saved_cred = cnp->cn_cred;
+			cnp->cn_cred = um->um_cred;
+		}
+		lerror = union_lookup1(um->um_lowervp, &lowerdvp,
+				&lowervp, cnp);
+		if (um->um_op == UNMNT_BELOW)
+			cnp->cn_cred = saved_cred;
+		cnp->cn_nameiop = nameiop;
+
 		if (lowervp != lowerdvp)
 			VOP_UNLOCK(lowerdvp);
 
 		if (cnp->cn_consume != 0) {
-			if (uppervp) {
-				vput(uppervp);
-				uppervp = 0;
+			if (uppervp != NULLVP) {
+				if (uppervp == upperdvp)
+					vrele(uppervp);
+				else
+					vput(uppervp);
+				uppervp = NULLVP;
 			}
 			*ap->a_vpp = lowervp;
 			if (!lockparent)
@@ -219,46 +243,47 @@ union_lookup(ap)
 	 *    whatever the bottom layer returned.
 	 */
 
+	*ap->a_vpp = NULLVP;
+
 	/* case 1. */
 	if ((uerror != 0) && (lerror != 0)) {
-		*ap->a_vpp = 0;
 		return (uerror);
 	}
 
 	/* case 2. */
 	if (uerror != 0 /* && (lerror == 0) */ ) {
 		if (lowervp->v_type == VDIR) { /* case 2b. */
-			if (uppervp != upperdvp)
-				VOP_LOCK(upperdvp);
-			uerror = union_mkshadow(upperdvp, cnp, &uppervp);
-			if (uppervp != upperdvp)
-				VOP_UNLOCK(upperdvp);
+			dun->un_flags &= ~UN_ULOCK;
+			VOP_UNLOCK(upperdvp);
+			uerror = union_mkshadow(um, upperdvp, cnp, &uppervp);
+			VOP_LOCK(upperdvp);
+			dun->un_flags |= UN_ULOCK;
+
 			if (uerror) {
-				if (lowervp) {
+				if (lowervp != NULLVP) {
 					vput(lowervp);
-					lowervp = 0;
+					lowervp = NULLVP;
 				}
 				return (uerror);
 			}
 		}
 	}
 
+	if (lowervp != NULLVP)
+		VOP_UNLOCK(lowervp);
+
 	error = union_allocvp(ap->a_vpp, dvp->v_mount, dvp, upperdvp, cnp,
 			      uppervp, lowervp);
 
-	if (uppervp)
-		VOP_UNLOCK(uppervp);
-	if (lowervp)
-		VOP_UNLOCK(lowervp);
-
 	if (error) {
-		if (uppervp)
-			vrele(uppervp);
-		if (lowervp)
+		if (uppervp != NULLVP)
+			vput(uppervp);
+		if (lowervp != NULLVP)
 			vrele(lowervp);
 	} else {
-		if (!lockparent && (*ap->a_vpp != dvp))
-			VOP_UNLOCK(dvp);
+		if (*ap->a_vpp != dvp)
+			if (!lockparent || !(cnp->cn_flags & ISLASTCN))
+				VOP_UNLOCK(dvp);
 	}
 
 	return (error);
@@ -276,12 +301,14 @@ union_create(ap)
 	struct union_node *un = VTOUNION(ap->a_dvp);
 	struct vnode *dvp = un->un_uppervp;
 
-	if (dvp) {
+	if (dvp != NULLVP) {
 		int error;
 		struct vnode *vp;
 
+		FIXUP(un);
+
 		VREF(dvp);
-		VOP_LOCK(dvp);
+		un->un_flags |= UN_KLOCK;
 		vput(ap->a_dvp);
 		error = VOP_CREATE(dvp, &vp, ap->a_cnp, ap->a_vap);
 		if (error)
@@ -295,9 +322,8 @@ union_create(ap)
 				ap->a_cnp,
 				vp,
 				NULLVP);
-		VOP_UNLOCK(vp);
 		if (error)
-			vrele(vp);
+			vput(vp);
 		return (error);
 	}
 
@@ -317,18 +343,20 @@ union_mknod(ap)
 	struct union_node *un = VTOUNION(ap->a_dvp);
 	struct vnode *dvp = un->un_uppervp;
 
-	if (dvp) {
+	if (dvp != NULLVP) {
 		int error;
 		struct vnode *vp;
 
+		FIXUP(un);
+
 		VREF(dvp);
-		VOP_LOCK(dvp);
+		un->un_flags |= UN_KLOCK;
 		vput(ap->a_dvp);
 		error = VOP_MKNOD(dvp, &vp, ap->a_cnp, ap->a_vap);
 		if (error)
 			return (error);
 
-		if (vp) {
+		if (vp != NULLVP) {
 			error = union_allocvp(
 					ap->a_vpp,
 					ap->a_dvp->v_mount,
@@ -337,9 +365,8 @@ union_mknod(ap)
 					ap->a_cnp,
 					vp,
 					NULLVP);
-			VOP_UNLOCK(vp);
 			if (error)
-				vrele(vp);
+				vput(vp);
 		}
 		return (error);
 	}
@@ -348,75 +375,6 @@ union_mknod(ap)
 	return (EROFS);
 }
 
-/*
- * copyfile.  copy the vnode (fvp) to the vnode (tvp)
- * using a sequence of reads and writes.  both (fvp)
- * and (tvp) are locked on entry and exit.
- */
-static int
-union_copyfile(p, cred, fvp, tvp)
-	struct proc *p;
-	struct ucred *cred;
-	struct vnode *fvp;
-	struct vnode *tvp;
-{
-	char *buf;
-	struct uio uio;
-	struct iovec iov;
-	int error = 0;
-	off_t offset;
-
-	/*
-	 * strategy:
-	 * allocate a buffer of size MAXBSIZE.
-	 * loop doing reads and writes, keeping track
-	 * of the current uio offset.
-	 * give up at the first sign of trouble.
-	 */
-
-	uio.uio_procp = p;
-	uio.uio_segflg = UIO_SYSSPACE;
-	offset = 0;
-
-	VOP_UNLOCK(fvp);				/* XXX */
-	LEASE_CHECK(fvp, p, cred, LEASE_READ);
-	VOP_LOCK(fvp);					/* XXX */
-	VOP_UNLOCK(tvp);				/* XXX */
-	LEASE_CHECK(tvp, p, cred, LEASE_WRITE);
-	VOP_LOCK(tvp);					/* XXX */
-
-	buf = malloc(MAXBSIZE, M_TEMP, M_WAITOK);
-	do {
-		uio.uio_iov = &iov;
-		uio.uio_iovcnt = 1;
-		iov.iov_base = buf;
-		iov.iov_len = MAXBSIZE;
-		uio.uio_resid = iov.iov_len;
-		uio.uio_offset = offset;
-		uio.uio_rw = UIO_READ;
-		error = VOP_READ(fvp, &uio, 0, cred);
-
-		if (error == 0) {
-			uio.uio_iov = &iov;
-			uio.uio_iovcnt = 1;
-			iov.iov_base = buf;
-			iov.iov_len = MAXBSIZE - uio.uio_resid;
-			uio.uio_rw = UIO_WRITE;
-			uio.uio_resid = iov.iov_len;
-			uio.uio_offset = offset;
-
-			do {
-				error = VOP_WRITE(tvp, &uio, 0, cred);
-			} while (error == 0 && uio.uio_resid > 0);
-			if (error == 0)
-				offset = uio.uio_offset;
-		}
-	} while ((uio.uio_resid == 0) && (error == 0));
-
-	free(buf, M_TEMP);
-	return (error);
-}
-
 int
 union_open(ap)
 	struct vop_open_args /* {
@@ -446,10 +404,8 @@ union_open(ap)
 		 */
 		tvp = un->un_lowervp;
 		if ((ap->a_mode & FWRITE) && (tvp->v_type == VREG)) {
-			struct nameidata nd;
-			struct filedesc *fdp = p->p_fd;
-			int fmode;
-			int cmode;
+			struct vnode *vp;
+			int i;
 
 			/*
 			 * Open the named file in the upper layer.  Note that
@@ -458,17 +414,16 @@ union_open(ap)
 			 * be a loopback mount of some other filesystem...
 			 * so open the file with exclusive create and barf if
 			 * it already exists.
-			 * XXX - perhaps shoudl re-lookup the node (once more
+			 * XXX - perhaps should re-lookup the node (once more
 			 * with feeling) and simply open that.  Who knows.
 			 */
-			NDINIT(&nd, CREATE, 0, UIO_SYSSPACE, un->un_path, p);
-			fmode = (O_CREAT|O_TRUNC|O_EXCL);
-			cmode = UN_FILEMODE & ~fdp->fd_cmask;
-			error = vn_open(&nd, fmode, cmode);
+			error = union_vn_create(&vp, un, p);
 			if (error)
 				return (error);
-			un->un_uppervp = nd.ni_vp;	/* XXX */
+
 			/* at this point, uppervp is locked */
+			union_newupper(un, vp);
+			un->un_flags |= UN_ULOCK;
 
 			/*
 			 * Now, if the file is being opened with truncation,
@@ -482,7 +437,7 @@ union_open(ap)
 				 * XXX - should not ignore errors
 				 * from VOP_CLOSE
 				 */
-				VOP_LOCK(un->un_lowervp);
+				VOP_LOCK(tvp);
 				error = VOP_OPEN(tvp, FREAD, cred, p);
 				if (error == 0) {
 					error = union_copyfile(p, cred,
@@ -492,20 +447,53 @@ union_open(ap)
 				} else {
 					VOP_UNLOCK(tvp);
 				}
-				VOP_UNLOCK(un->un_uppervp);
-				(void) VOP_CLOSE(un->un_uppervp, FWRITE);
-				VOP_LOCK(un->un_uppervp);
+
+#ifdef UNION_DIAGNOSTIC
+				if (!error)
+					uprintf("union: copied up %s\n",
+								un->un_path);
+#endif
+			}
+
+			un->un_flags &= ~UN_ULOCK;
+			VOP_UNLOCK(un->un_uppervp);
+			union_vn_close(un->un_uppervp, FWRITE, cred, p);
+			VOP_LOCK(un->un_uppervp);
+			un->un_flags |= UN_ULOCK;
+
+			/*
+			 * Subsequent IOs will go to the top layer, so
+			 * call close on the lower vnode and open on the
+			 * upper vnode to ensure that the filesystem keeps
+			 * its references counts right.  This doesn't do
+			 * the right thing with (cred) and (FREAD) though.
+			 * Ignoring error returns is not righ, either.
+			 */
+			for (i = 0; i < un->un_openl; i++) {
+				(void) VOP_CLOSE(tvp, FREAD);
+				(void) VOP_OPEN(un->un_uppervp, FREAD, cred, p);
 			}
+			un->un_openl = 0;
+
 			if (error == 0)
 				error = VOP_OPEN(un->un_uppervp, mode, cred, p);
-			VOP_UNLOCK(un->un_uppervp);
 			return (error);
 		}
+
+		/*
+		 * Just open the lower vnode
+		 */
+		un->un_openl++;
+		VOP_LOCK(tvp);
+		error = VOP_OPEN(tvp, mode, cred, p);
+		VOP_UNLOCK(tvp);
+
+		return (error);
 	}
 
-	VOP_LOCK(tvp);
+	FIXUP(un);
+
 	error = VOP_OPEN(tvp, mode, cred, p);
-	VOP_UNLOCK(tvp);
 
 	return (error);
 }
@@ -519,8 +507,21 @@ union_close(ap)
 		struct proc *a_p;
 	} */ *ap;
 {
+	struct union_node *un = VTOUNION(ap->a_vp);
+	struct vnode *vp;
 
-	return (VOP_CLOSE(OTHERVP(ap->a_vp), ap->a_fflag, ap->a_cred, ap->a_p));
+	if (un->un_uppervp != NULLVP) {
+		vp = un->un_uppervp;
+	} else {
+#ifdef UNION_DIAGNOSTIC
+		if (un->un_openl <= 0)
+			panic("union: un_openl cnt");
+#endif
+		--un->un_openl;
+		vp = un->un_lowervp;
+	}
+
+	return (VOP_CLOSE(vp, ap->a_fflag, ap->a_cred, ap->a_p));
 }
 
 /*
@@ -542,28 +543,35 @@ union_access(ap)
 	} */ *ap;
 {
 	struct union_node *un = VTOUNION(ap->a_vp);
-	int error = 0;
+	int error = EACCES;
 	struct vnode *vp;
 
-	if (vp = un->un_lowervp) {
-		VOP_LOCK(vp);
-		error = VOP_ACCESS(vp, ap->a_mode, ap->a_cred, ap->a_p);
-		VOP_UNLOCK(vp);
-		if (error)
-			return (error);
+	if ((vp = un->un_uppervp) != NULLVP) {
+		FIXUP(un);
+		return (VOP_ACCESS(vp, ap->a_mode, ap->a_cred, ap->a_p));
 	}
 
-	if (vp = un->un_uppervp) {
+	if ((vp = un->un_lowervp) != NULLVP) {
 		VOP_LOCK(vp);
 		error = VOP_ACCESS(vp, ap->a_mode, ap->a_cred, ap->a_p);
+		if (error == 0) {
+			struct union_mount *um = MOUNTTOUNIONMOUNT(vp->v_mount);
+
+			if (um->um_op == UNMNT_BELOW)
+				error = VOP_ACCESS(vp, ap->a_mode,
+						um->um_cred, ap->a_p);
+		}
 		VOP_UNLOCK(vp);
+		if (error)
+			return (error);
 	}
 
 	return (error);
 }
 
 /*
- *  We handle getattr only to change the fsid.
+ * We handle getattr only to change the fsid and
+ * track object sizes
  */
 int
 union_getattr(ap)
@@ -575,14 +583,64 @@ union_getattr(ap)
 	} */ *ap;
 {
 	int error;
-	struct vnode *vp = OTHERVP(ap->a_vp);
+	struct union_node *un = VTOUNION(ap->a_vp);
+	struct vnode *vp = un->un_uppervp;
+	struct vattr *vap;
+	struct vattr va;
+
+
+	/*
+	 * Some programs walk the filesystem hierarchy by counting
+	 * links to directories to avoid stat'ing all the time.
+	 * This means the link count on directories needs to be "correct".
+	 * The only way to do that is to call getattr on both layers
+	 * and fix up the link count.  The link count will not necessarily
+	 * be accurate but will be large enough to defeat the tree walkers.
+	 */
+
+	vap = ap->a_vap;
+
+	vp = un->un_uppervp;
+	if (vp != NULLVP) {
+		/*
+		 * It's not clear whether VOP_GETATTR is to be
+		 * called with the vnode locked or not.  stat() calls
+		 * it with (vp) locked, and fstat calls it with
+		 * (vp) unlocked.
+		 * In the mean time, compensate here by checking
+		 * the union_node's lock flag.
+		 */
+		if (un->un_flags & UN_LOCKED)
+			FIXUP(un);
+
+		error = VOP_GETATTR(vp, vap, ap->a_cred, ap->a_p);
+		if (error)
+			return (error);
+		union_newsize(ap->a_vp, vap->va_size, VNOVAL);
+	}
 
-	VOP_LOCK(vp);
-	error = VOP_GETATTR(vp, ap->a_vap, ap->a_cred, ap->a_p);
-	VOP_UNLOCK(vp);
+	if (vp == NULLVP) {
+		vp = un->un_lowervp;
+	} else if (vp->v_type == VDIR) {
+		vp = un->un_lowervp;
+		vap = &va;
+	} else {
+		vp = NULLVP;
+	}
 
-	/* Requires that arguments be restored. */
-	ap->a_vap->va_fsid = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0];
+	if (vp != NULLVP) {
+		VOP_LOCK(vp);
+		error = VOP_GETATTR(vp, vap, ap->a_cred, ap->a_p);
+		VOP_UNLOCK(vp);
+		if (error)
+			return (error);
+		union_newsize(ap->a_vp, VNOVAL, vap->va_size);
+	}
+
+	if ((vap != ap->a_vap) && (vap->va_type == VDIR))
+		ap->a_vap->va_nlink += vap->va_nlink;
+
+	vap->va_fsid = ap->a_vp->v_mount->mnt_stat.f_fsid.val[0];
 	return (0);
 }
 
@@ -598,17 +656,41 @@ union_setattr(ap)
 	struct union_node *un = VTOUNION(ap->a_vp);
 	int error;
 
-	if (un->un_uppervp) {
-		VOP_LOCK(un->un_uppervp);
+	/*
+	 * Handle case of truncating lower object to zero size,
+	 * by creating a zero length upper object.  This is to
+	 * handle the case of open with O_TRUNC and O_CREAT.
+	 */
+	if ((un->un_uppervp == NULLVP) &&
+	    /* assert(un->un_lowervp != NULLVP) */
+	    (un->un_lowervp->v_type == VREG) &&
+	    (ap->a_vap->va_size == 0)) {
+		struct vnode *vp;
+
+		error = union_vn_create(&vp, un, ap->a_p);
+		if (error)
+			return (error);
+
+		/* at this point, uppervp is locked */
+		union_newupper(un, vp);
+
+		VOP_UNLOCK(vp);
+		union_vn_close(un->un_uppervp, FWRITE, ap->a_cred, ap->a_p);
+		VOP_LOCK(vp);
+		un->un_flags |= UN_ULOCK;
+	}
+
+	/*
+	 * Try to set attributes in upper layer,
+	 * otherwise return read-only filesystem error.
+	 */
+	if (un->un_uppervp != NULLVP) {
+		FIXUP(un);
 		error = VOP_SETATTR(un->un_uppervp, ap->a_vap,
 					ap->a_cred, ap->a_p);
-		VOP_UNLOCK(un->un_uppervp);
+		if ((error == 0) && (ap->a_vap->va_size != VNOVAL))
+			union_newsize(ap->a_vp, ap->a_vap->va_size, VNOVAL);
 	} else {
-		/*
-		 * XXX should do a copyfile (perhaps only if
-		 * the file permission change, which would not
-		 * track va_ctime correctly).
-		 */
 		error = EROFS;
 	}
 
@@ -626,10 +708,34 @@ union_read(ap)
 {
 	int error;
 	struct vnode *vp = OTHERVP(ap->a_vp);
+	int dolock = (vp == LOWERVP(ap->a_vp));
 
-	VOP_LOCK(vp);
+	if (dolock)
+		VOP_LOCK(vp);
+	else
+		FIXUP(VTOUNION(ap->a_vp));
 	error = VOP_READ(vp, ap->a_uio, ap->a_ioflag, ap->a_cred);
-	VOP_UNLOCK(vp);
+	if (dolock)
+		VOP_UNLOCK(vp);
+
+	/*
+	 * XXX
+	 * perhaps the size of the underlying object has changed under
+	 * our feet.  take advantage of the offset information present
+	 * in the uio structure.
+	 */
+	if (error == 0) {
+		struct union_node *un = VTOUNION(ap->a_vp);
+		off_t cur = ap->a_uio->uio_offset;
+
+		if (vp == un->un_uppervp) {
+			if (cur > un->un_uppersz)
+				union_newsize(ap->a_vp, cur, VNOVAL);
+		} else {
+			if (cur > un->un_lowersz)
+				union_newsize(ap->a_vp, VNOVAL, cur);
+		}
+	}
 
 	return (error);
 }
@@ -645,10 +751,32 @@ union_write(ap)
 {
 	int error;
 	struct vnode *vp = OTHERVP(ap->a_vp);
+	int dolock = (vp == LOWERVP(ap->a_vp));
 
-	VOP_LOCK(vp);
+	if (dolock)
+		VOP_LOCK(vp);
+	else
+		FIXUP(VTOUNION(ap->a_vp));
 	error = VOP_WRITE(vp, ap->a_uio, ap->a_ioflag, ap->a_cred);
-	VOP_UNLOCK(vp);
+	if (dolock)
+		VOP_UNLOCK(vp);
+
+	/*
+	 * the size of the underlying object may be changed by the
+	 * write.
+	 */
+	if (error == 0) {
+		struct union_node *un = VTOUNION(ap->a_vp);
+		off_t cur = ap->a_uio->uio_offset;
+
+		if (vp == un->un_uppervp) {
+			if (cur > un->un_uppersz)
+				union_newsize(ap->a_vp, cur, VNOVAL);
+		} else {
+			if (cur > un->un_lowersz)
+				union_newsize(ap->a_vp, VNOVAL, cur);
+		}
+	}
 
 	return (error);
 }
@@ -710,11 +838,17 @@ union_fsync(ap)
 	int error = 0;
 	struct vnode *targetvp = OTHERVP(ap->a_vp);
 
-	if (targetvp) {
-		VOP_LOCK(targetvp);
+	if (targetvp != NULLVP) {
+		int dolock = (targetvp == LOWERVP(ap->a_vp));
+
+		if (dolock)
+			VOP_LOCK(targetvp);
+		else
+			FIXUP(VTOUNION(ap->a_vp));
 		error = VOP_FSYNC(targetvp, ap->a_cred,
 					ap->a_waitfor, ap->a_p);
-		VOP_UNLOCK(targetvp);
+		if (dolock)
+			VOP_UNLOCK(targetvp);
 	}
 
 	return (error);
@@ -745,18 +879,26 @@ union_remove(ap)
 	struct union_node *dun = VTOUNION(ap->a_dvp);
 	struct union_node *un = VTOUNION(ap->a_vp);
 
-	if (dun->un_uppervp && un->un_uppervp) {
+	if (dun->un_uppervp != NULLVP && un->un_uppervp != NULLVP) {
 		struct vnode *dvp = dun->un_uppervp;
 		struct vnode *vp = un->un_uppervp;
 
+		FIXUP(dun);
 		VREF(dvp);
-		VOP_LOCK(dvp);
+		dun->un_flags |= UN_KLOCK;
 		vput(ap->a_dvp);
+		FIXUP(un);
 		VREF(vp);
-		VOP_LOCK(vp);
+		un->un_flags |= UN_KLOCK;
 		vput(ap->a_vp);
 
 		error = VOP_REMOVE(dvp, vp, ap->a_cnp);
+		if (!error)
+			union_removed_upper(un);
+
+		/*
+		 * XXX: should create a whiteout here
+		 */
 	} else {
 		/*
 		 * XXX: should create a whiteout here
@@ -781,20 +923,22 @@ union_link(ap)
 	struct union_node *dun = VTOUNION(ap->a_vp);
 	struct union_node *un = VTOUNION(ap->a_tdvp);
 
-	if (dun->un_uppervp && un->un_uppervp) {
+	if (dun->un_uppervp != NULLVP && un->un_uppervp != NULLVP) {
 		struct vnode *dvp = dun->un_uppervp;
 		struct vnode *vp = un->un_uppervp;
 
+		FIXUP(dun);
 		VREF(dvp);
-		VOP_LOCK(dvp);
+		dun->un_flags |= UN_KLOCK;
 		vput(ap->a_vp);
+		FIXUP(un);
 		VREF(vp);
 		vrele(ap->a_tdvp);
 
 		error = VOP_LINK(dvp, vp, ap->a_cnp);
 	} else {
 		/*
-		 * XXX: need to copy to upper layer
+		 * XXX: perhaps could copy to upper layer
 		 * and do the link there.
 		 */
 		vput(ap->a_vp);
@@ -825,7 +969,7 @@ union_rename(ap)
 
 	if (fdvp->v_op == union_vnodeop_p) {	/* always true */
 		struct union_node *un = VTOUNION(fdvp);
-		if (un->un_uppervp == 0) {
+		if (un->un_uppervp == NULLVP) {
 			error = EROFS;
 			goto bad;
 		}
@@ -837,7 +981,7 @@ union_rename(ap)
 
 	if (fvp->v_op == union_vnodeop_p) {	/* always true */
 		struct union_node *un = VTOUNION(fvp);
-		if (un->un_uppervp == 0) {
+		if (un->un_uppervp == NULLVP) {
 			error = EROFS;
 			goto bad;
 		}
@@ -849,27 +993,31 @@ union_rename(ap)
 
 	if (tdvp->v_op == union_vnodeop_p) {
 		struct union_node *un = VTOUNION(tdvp);
-		if (un->un_uppervp == 0) {
+		if (un->un_uppervp == NULLVP) {
+			/*
+			 * this should never happen in normal
+			 * operation but might if there was
+			 * a problem creating the top-level shadow
+			 * directory.
+			 */
 			error = EROFS;
 			goto bad;
 		}
 
 		tdvp = un->un_uppervp;
 		VREF(tdvp);
-		VOP_LOCK(tdvp);
-		vput(ap->a_fdvp);
+		un->un_flags |= UN_KLOCK;
+		vput(ap->a_tdvp);
 	}
 
-	if (tvp && tvp->v_op == union_vnodeop_p) {
+	if (tvp != NULLVP && tvp->v_op == union_vnodeop_p) {
 		struct union_node *un = VTOUNION(tvp);
-		if (un->un_uppervp == 0) {
-			error = EROFS;
-			goto bad;
-		}
 
 		tvp = un->un_uppervp;
-		VREF(tvp);
-		VOP_LOCK(tvp);
+		if (tvp != NULLVP) {
+			VREF(tvp);
+			un->un_flags |= UN_KLOCK;
+		}
 		vput(ap->a_tvp);
 	}
 
@@ -879,7 +1027,7 @@ bad:
 	vrele(fdvp);
 	vrele(fvp);
 	vput(tdvp);
-	if (tvp)
+	if (tvp != NULLVP)
 		vput(tvp);
 
 	return (error);
@@ -897,12 +1045,13 @@ union_mkdir(ap)
 	struct union_node *un = VTOUNION(ap->a_dvp);
 	struct vnode *dvp = un->un_uppervp;
 
-	if (dvp) {
+	if (dvp != NULLVP) {
 		int error;
 		struct vnode *vp;
 
+		FIXUP(un);
 		VREF(dvp);
-		VOP_LOCK(dvp);
+		un->un_flags |= UN_KLOCK;
 		vput(ap->a_dvp);
 		error = VOP_MKDIR(dvp, &vp, ap->a_cnp, ap->a_vap);
 		if (error)
@@ -916,9 +1065,8 @@ union_mkdir(ap)
 				ap->a_cnp,
 				vp,
 				NULLVP);
-		VOP_UNLOCK(vp);
 		if (error)
-			vrele(vp);
+			vput(vp);
 		return (error);
 	}
 
@@ -938,18 +1086,26 @@ union_rmdir(ap)
 	struct union_node *dun = VTOUNION(ap->a_dvp);
 	struct union_node *un = VTOUNION(ap->a_vp);
 
-	if (dun->un_uppervp && un->un_uppervp) {
+	if (dun->un_uppervp != NULLVP && un->un_uppervp != NULLVP) {
 		struct vnode *dvp = dun->un_uppervp;
 		struct vnode *vp = un->un_uppervp;
 
+		FIXUP(dun);
 		VREF(dvp);
-		VOP_LOCK(dvp);
+		dun->un_flags |= UN_KLOCK;
 		vput(ap->a_dvp);
+		FIXUP(un);
 		VREF(vp);
-		VOP_LOCK(vp);
+		un->un_flags |= UN_KLOCK;
 		vput(ap->a_vp);
 
-		error = VOP_REMOVE(dvp, vp, ap->a_cnp);
+		error = VOP_RMDIR(dvp, vp, ap->a_cnp);
+		if (!error)
+			union_removed_upper(un);
+
+		/*
+		 * XXX: should create a whiteout here
+		 */
 	} else {
 		/*
 		 * XXX: should create a whiteout here
@@ -975,17 +1131,18 @@ union_symlink(ap)
 	struct union_node *un = VTOUNION(ap->a_dvp);
 	struct vnode *dvp = un->un_uppervp;
 
-	if (dvp) {
+	if (dvp != NULLVP) {
 		int error;
 		struct vnode *vp;
 		struct mount *mp = ap->a_dvp->v_mount;
 
+		FIXUP(un);
 		VREF(dvp);
-		VOP_LOCK(dvp);
+		un->un_flags |= UN_KLOCK;
 		vput(ap->a_dvp);
 		error = VOP_SYMLINK(dvp, &vp, ap->a_cnp,
 					ap->a_vap, ap->a_target);
-		*ap->a_vpp = 0;
+		*ap->a_vpp = NULLVP;
 		return (error);
 	}
 
@@ -1012,12 +1169,9 @@ union_readdir(ap)
 	int error = 0;
 	struct union_node *un = VTOUNION(ap->a_vp);
 
-	if (un->un_uppervp) {
-		struct vnode *vp = OTHERVP(ap->a_vp);
-
-		VOP_LOCK(vp);
-		error = VOP_READLINK(vp, ap->a_uio, ap->a_cred);
-		VOP_UNLOCK(vp);
+	if (un->un_uppervp != NULLVP) {
+		FIXUP(un);
+		error = VOP_READDIR(un->un_uppervp, ap->a_uio, ap->a_cred);
 	}
 
 	return (error);
@@ -1033,10 +1187,15 @@ union_readlink(ap)
 {
 	int error;
 	struct vnode *vp = OTHERVP(ap->a_vp);
+	int dolock = (vp == LOWERVP(ap->a_vp));
 
-	VOP_LOCK(vp);
+	if (dolock)
+		VOP_LOCK(vp);
+	else
+		FIXUP(VTOUNION(ap->a_vp));
 	error = VOP_READLINK(vp, ap->a_uio, ap->a_cred);
-	VOP_UNLOCK(vp);
+	if (dolock)
+		VOP_UNLOCK(vp);
 
 	return (error);
 }
@@ -1052,11 +1211,16 @@ union_abortop(ap)
 	struct vnode *vp = OTHERVP(ap->a_dvp);
 	struct union_node *un = VTOUNION(ap->a_dvp);
 	int islocked = un->un_flags & UN_LOCKED;
+	int dolock = (vp == LOWERVP(ap->a_dvp));
 
-	if (islocked)
-		VOP_LOCK(vp);
+	if (islocked) {
+		if (dolock)
+			VOP_LOCK(vp);
+		else
+			FIXUP(VTOUNION(ap->a_dvp));
+	}
 	error = VOP_ABORTOP(vp, ap->a_cnp);
-	if (islocked)
+	if (islocked && dolock)
 		VOP_UNLOCK(vp);
 
 	return (error);
@@ -1068,6 +1232,7 @@ union_inactive(ap)
 		struct vnode *a_vp;
 	} */ *ap;
 {
+	struct union_node *un = VTOUNION(ap->a_vp);
 
 	/*
 	 * Do nothing (and _don't_ bypass).
@@ -1082,13 +1247,16 @@ union_inactive(ap)
 	 * That's too much work for now.
 	 */
 
-#ifdef DIAGNOSTIC
-	struct union_node *un = VTOUNION(ap->a_vp);
-
+#ifdef UNION_DIAGNOSTIC
 	if (un->un_flags & UN_LOCKED)
 		panic("union: inactivating locked node");
+	if (un->un_flags & UN_ULOCK)
+		panic("union: inactivating w/locked upper node");
 #endif
 
+	if ((un->un_flags & UN_CACHED) == 0)
+		vgone(ap->a_vp);
+
 	return (0);
 }
 
@@ -1098,31 +1266,9 @@ union_reclaim(ap)
 		struct vnode *a_vp;
 	} */ *ap;
 {
-	struct vnode *vp = ap->a_vp;
-	struct union_node *un = VTOUNION(vp);
-	struct vnode *uppervp = un->un_uppervp;
-	struct vnode *lowervp = un->un_lowervp;
-	struct vnode *dirvp = un->un_dirvp;
-	char *path = un->un_path;
 
-	/*
-	 * Note: in vop_reclaim, vp->v_op == dead_vnodeop_p,
-	 * so we can't call VOPs on ourself.
-	 */
-	/* After this assignment, this node will not be re-used. */
-	un->un_uppervp = 0;
-	un->un_lowervp = 0;
-	un->un_dirvp = 0;
-	un->un_path = NULL;
-	union_freevp(vp);
-	if (uppervp)
-		vrele(uppervp);
-	if (lowervp)
-		vrele(lowervp);
-	if (dirvp)
-		vrele(dirvp);
-	if (path)
-		free(path, M_TEMP);
+	union_freevp(ap->a_vp);
+
 	return (0);
 }
 
@@ -1130,9 +1276,29 @@ int
 union_lock(ap)
 	struct vop_lock_args *ap;
 {
-	struct union_node *un = VTOUNION(ap->a_vp);
+	struct vnode *vp = ap->a_vp;
+	struct union_node *un;
+
+start:
+	while (vp->v_flag & VXLOCK) {
+		vp->v_flag |= VXWANT;
+		sleep((caddr_t)vp, PINOD);
+	}
+
+	un = VTOUNION(vp);
 
-	while (un->un_flags & UN_LOCKED) {
+	if (un->un_uppervp != NULLVP) {
+		if ((un->un_flags & UN_ULOCK) == 0) {
+			un->un_flags |= UN_ULOCK;
+			VOP_LOCK(un->un_uppervp);
+		}
+#ifdef DIAGNOSTIC
+		if (un->un_flags & UN_KLOCK)
+			panic("union: dangling upper lock");
+#endif
+	}
+
+	if (un->un_flags & UN_LOCKED) {
 #ifdef DIAGNOSTIC
 		if (curproc && un->un_pid == curproc->p_pid &&
 			    un->un_pid > -1 && curproc->p_pid > -1)
@@ -1140,8 +1306,8 @@ union_lock(ap)
 #endif
 		un->un_flags |= UN_WANT;
 		sleep((caddr_t) &un->un_flags, PINOD);
+		goto start;
 	}
-	un->un_flags |= UN_LOCKED;
 
 #ifdef DIAGNOSTIC
 	if (curproc)
@@ -1149,6 +1315,9 @@ union_lock(ap)
 	else
 		un->un_pid = -1;
 #endif
+
+	un->un_flags |= UN_LOCKED;
+	return (0);
 }
 
 int
@@ -1166,6 +1335,12 @@ union_unlock(ap)
 #endif
 
 	un->un_flags &= ~UN_LOCKED;
+
+	if ((un->un_flags & (UN_ULOCK|UN_KLOCK)) == UN_ULOCK)
+		VOP_UNLOCK(un->un_uppervp);
+
+	un->un_flags &= ~(UN_ULOCK|UN_KLOCK);
+
 	if (un->un_flags & UN_WANT) {
 		un->un_flags &= ~UN_WANT;
 		wakeup((caddr_t) &un->un_flags);
@@ -1174,6 +1349,8 @@ union_unlock(ap)
 #ifdef DIAGNOSTIC
 	un->un_pid = 0;
 #endif
+
+	return (0);
 }
 
 int
@@ -1188,10 +1365,15 @@ union_bmap(ap)
 {
 	int error;
 	struct vnode *vp = OTHERVP(ap->a_vp);
+	int dolock = (vp == LOWERVP(ap->a_vp));
 
-	VOP_LOCK(vp);
+	if (dolock)
+		VOP_LOCK(vp);
+	else
+		FIXUP(VTOUNION(ap->a_vp));
 	error = VOP_BMAP(vp, ap->a_bn, ap->a_vpp, ap->a_bnp, ap->a_runp);
-	VOP_UNLOCK(vp);
+	if (dolock)
+		VOP_UNLOCK(vp);
 
 	return (error);
 }
@@ -1229,10 +1411,15 @@ union_pathconf(ap)
 {
 	int error;
 	struct vnode *vp = OTHERVP(ap->a_vp);
+	int dolock = (vp == LOWERVP(ap->a_vp));
 
-	VOP_LOCK(vp);
+	if (dolock)
+		VOP_LOCK(vp);
+	else
+		FIXUP(VTOUNION(ap->a_vp));
 	error = VOP_PATHCONF(vp, ap->a_name, ap->a_retval);
-	VOP_UNLOCK(vp);
+	if (dolock)
+		VOP_UNLOCK(vp);
 
 	return (error);
 }
@@ -1272,7 +1459,7 @@ union_strategy(ap)
 	bp->b_vp = OTHERVP(bp->b_vp);
 
 #ifdef DIAGNOSTIC
-	if (bp->b_vp == 0)
+	if (bp->b_vp == NULLVP)
 		panic("union_strategy: nil vp");
 	if (((bp->b_flags & B_READ) == 0) &&
 	    (bp->b_vp == LOWERVP(savedvp)))