The recent NFS forced unmount improvements introduced a side-effect

where some client operations might be unexpectedly cancelled during
an unsuccessful non-forced unmount attempt. This causes problems
for amd(8), because it periodically attempts a non-forced unmount
to check if the filesystem is still in use.

Fix this by adding a new mountpoint flag MNTK_UNMOUNTF that is set
only during the operation of a forced unmount. Use this instead of
MNTK_UNMOUNT to trigger the cancellation of hung NFS operations.

Also correct a problem where dounmount() might inadvertently clear
the MNTK_UNMOUNT flag.

Reported by:	simokawa
MFC after:	1 week

sys/kern/vfs_extattr.c

@@ -1103,11 +1103,18 @@ dounmount(mp, flags, td)
 	int async_flag;
 
 	mtx_lock(&mountlist_mtx);
+	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
+		mtx_unlock(&mountlist_mtx);
+		return (EBUSY);
+	}
 	mp->mnt_kern_flag |= MNTK_UNMOUNT;
+	/* Allow filesystems to detect that a forced unmount is in progress. */
+	if (flags & MNT_FORCE)
+		mp->mnt_kern_flag |= MNTK_UNMOUNTF;
 	error = lockmgr(&mp->mnt_lock, LK_DRAIN | LK_INTERLOCK |
 	    ((flags & MNT_FORCE) ? 0 : LK_NOWAIT), &mountlist_mtx, td);
 	if (error) {
-		mp->mnt_kern_flag &= ~MNTK_UNMOUNT;
+		mp->mnt_kern_flag &= ~(MNTK_UNMOUNT | MNTK_UNMOUNTF);
 		if (mp->mnt_kern_flag & MNTK_MWAIT)
 			wakeup((caddr_t)mp);
 		return (error);
@@ -1153,7 +1160,7 @@ dounmount(mp, flags, td)
 		if ((mp->mnt_flag & MNT_RDONLY) == 0 && mp->mnt_syncer == NULL)
 			(void) vfs_allocate_syncvnode(mp);
 		mtx_lock(&mountlist_mtx);
-		mp->mnt_kern_flag &= ~MNTK_UNMOUNT;
+		mp->mnt_kern_flag &= ~(MNTK_UNMOUNT | MNTK_UNMOUNTF);
 		mp->mnt_flag |= async_flag;
 		lockmgr(&mp->mnt_lock, LK_RELEASE | LK_INTERLOCK,
 		    &mountlist_mtx, td);

sys/kern/vfs_syscalls.c

@@ -1103,11 +1103,18 @@ dounmount(mp, flags, td)
 	int async_flag;
 
 	mtx_lock(&mountlist_mtx);
+	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
+		mtx_unlock(&mountlist_mtx);
+		return (EBUSY);
+	}
 	mp->mnt_kern_flag |= MNTK_UNMOUNT;
+	/* Allow filesystems to detect that a forced unmount is in progress. */
+	if (flags & MNT_FORCE)
+		mp->mnt_kern_flag |= MNTK_UNMOUNTF;
 	error = lockmgr(&mp->mnt_lock, LK_DRAIN | LK_INTERLOCK |
 	    ((flags & MNT_FORCE) ? 0 : LK_NOWAIT), &mountlist_mtx, td);
 	if (error) {
-		mp->mnt_kern_flag &= ~MNTK_UNMOUNT;
+		mp->mnt_kern_flag &= ~(MNTK_UNMOUNT | MNTK_UNMOUNTF);
 		if (mp->mnt_kern_flag & MNTK_MWAIT)
 			wakeup((caddr_t)mp);
 		return (error);
@@ -1153,7 +1160,7 @@ dounmount(mp, flags, td)
 		if ((mp->mnt_flag & MNT_RDONLY) == 0 && mp->mnt_syncer == NULL)
 			(void) vfs_allocate_syncvnode(mp);
 		mtx_lock(&mountlist_mtx);
-		mp->mnt_kern_flag &= ~MNTK_UNMOUNT;
+		mp->mnt_kern_flag &= ~(MNTK_UNMOUNT | MNTK_UNMOUNTF);
 		mp->mnt_flag |= async_flag;
 		lockmgr(&mp->mnt_lock, LK_RELEASE | LK_INTERLOCK,
 		    &mountlist_mtx, td);

sys/nfsclient/nfs_socket.c

@@ -850,8 +850,8 @@ nfs_request(struct vnode *vp, struct mbuf *mrest, int procnum,
 	int trylater_delay = NQ_TRYLATERDEL, trylater_cnt = 0;
 	u_int32_t xid;
 
-	/* Reject requests while attempting to unmount. */
+	/* Reject requests while attempting a forced unmount. */
-	if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNT) {
+	if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
 		m_freem(mrest);
 		return (ESTALE);
 	}
@@ -1230,8 +1230,8 @@ nfs_sigintr(struct nfsmount *nmp, struct nfsreq *rep, struct proc *p)
 
 	if (rep && (rep->r_flags & R_SOFTTERM))
 		return (EINTR);
-	/* Terminate all requests while attempting to unmount. */
+	/* Terminate all requests while attempting a forced unmount. */
-	if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNT)
+	if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)
 		return (EINTR);
 	if (!(nmp->nm_flag & NFSMNT_INT))
 		return (0);

sys/sys/mount.h

@@ -240,7 +240,13 @@ struct mount {
  * MNTK_UNMOUNT locks the mount entry so that name lookup cannot proceed
  * past the mount point.  This keeps the subtree stable during mounts
  * and unmounts.
+ *
+ * MNTK_UNMOUNTF permits filesystems to detect a forced unmount while
+ * dounmount() is still waiting to lock the mountpoint. This allows
+ * the filesystem to cancel operations that might otherwise deadlock
+ * with the unmount attempt (used by NFS).
  */
+#define MNTK_UNMOUNTF	0x00000001	/* forced unmount in progress */
 #define MNTK_UNMOUNT	0x01000000	/* unmount in progress */
 #define	MNTK_MWAIT	0x02000000	/* waiting for unmount to finish */
 #define MNTK_WANTRDWR	0x04000000	/* upgrade to read/write requested */