From 214297b8eb7dc08618ae717483f5deb3ad1574cd Mon Sep 17 00:00:00 2001
From: Alexander Bluhm <bluhm@cvs.openbsd.org>
Date: Sat, 15 Apr 2017 13:56:44 +0000
Subject: After forced unmount of a file system that has other mount points in
 it, dangling mounts could remain.  When unmounting check the hierarcy and
 unmount recursively.  Also prevent that a new mount appears during the scan.
 Joint work with natano@; testing and OK krw@

---
 sys/kern/vfs_syscalls.c | 85 ++++++++++++++++++++++++++++++++++++++++++++++++-
 sys/sys/mount.h         |  3 +-
 2 files changed, 86 insertions(+), 2 deletions(-)

(limited to 'sys')

diff --git a/sys/kern/vfs_syscalls.c b/sys/kern/vfs_syscalls.c
index 25d53b46afe..b41c4a24b96 100644
--- a/sys/kern/vfs_syscalls.c
+++ b/sys/kern/vfs_syscalls.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: vfs_syscalls.c,v 1.271 2017/02/15 03:36:58 guenther Exp $	*/
+/*	$OpenBSD: vfs_syscalls.c,v 1.272 2017/04/15 13:56:43 bluhm Exp $	*/
 /*	$NetBSD: vfs_syscalls.c,v 1.71 1996/04/23 10:29:02 mycroft Exp $	*/
 
 /*
@@ -88,6 +88,7 @@ int domkdirat(struct proc *, int, const char *, mode_t);
 int doutimensat(struct proc *, int, const char *, struct timespec [2], int);
 int dovutimens(struct proc *, struct vnode *, struct timespec [2]);
 int dofutimens(struct proc *, int, struct timespec [2]);
+int dounmount_leaf(struct mount *, int, struct proc *);
 
 /*
  * Virtual File System System Calls
@@ -204,7 +205,22 @@ sys_mount(struct proc *p, void *v, register_t *retval)
 	strncpy(mp->mnt_stat.f_fstypename, vfsp->vfc_name, MFSNAMELEN);
 	mp->mnt_vnodecovered = vp;
 	mp->mnt_stat.f_owner = p->p_ucred->cr_uid;
+
 update:
+	/* Ensure that the parent mountpoint does not get unmounted. */
+	error = vfs_busy(vp->v_mount, VB_READ|VB_NOWAIT);
+	if (error) {
+		if (mp->mnt_flag & MNT_UPDATE) {
+			mp->mnt_flag = mntflag;
+			vfs_unbusy(mp);
+		} else {
+			vfs_unbusy(mp);
+			free(mp, M_MOUNT, sizeof(*mp));
+		}
+		vput(vp);
+		return (error);
+	}
+
 	/*
 	 * Set the mount level flags.
 	 */
@@ -226,6 +242,7 @@ update:
 		mp->mnt_stat.f_ctime = time_second;
 	}
 	if (mp->mnt_flag & MNT_UPDATE) {
+		vfs_unbusy(vp->v_mount);
 		vput(vp);
 		if (mp->mnt_flag & MNT_WANTRDWR)
 			mp->mnt_flag &= ~MNT_RDONLY;
@@ -257,6 +274,7 @@ update:
 		vfsp->vfc_refcount++;
 		TAILQ_INSERT_TAIL(&mountlist, mp, mnt_list);
 		checkdirs(vp);
+		vfs_unbusy(vp->v_mount);
 		VOP_UNLOCK(vp, p);
 		if ((mp->mnt_flag & MNT_RDONLY) == 0)
 			error = vfs_allocate_syncvnode(mp);
@@ -268,6 +286,7 @@ update:
 		mp->mnt_vnodecovered->v_mountedhere = NULL;
 		vfs_unbusy(mp);
 		free(mp, M_MOUNT, sizeof(*mp));
+		vfs_unbusy(vp->v_mount);
 		vput(vp);
 	}
 	return (error);
@@ -373,6 +392,70 @@ sys_unmount(struct proc *p, void *v, register_t *retval)
  */
 int
 dounmount(struct mount *mp, int flags, struct proc *p)
+{
+	SLIST_HEAD(, mount) mplist;
+	struct mount *nmp;
+	int error;
+
+	SLIST_INIT(&mplist);
+	SLIST_INSERT_HEAD(&mplist, mp, mnt_dounmount);
+
+	/*
+	 * Collect nested mount points. This takes advantage of the mount list
+	 * being ordered - nested mount points come after their parent.
+	 */
+	while ((mp = TAILQ_NEXT(mp, mnt_list)) != NULL) {
+		SLIST_FOREACH(nmp, &mplist, mnt_dounmount) {
+			if (mp->mnt_vnodecovered == NULLVP ||
+			    mp->mnt_vnodecovered->v_mount != nmp)
+				continue;
+
+			if ((flags & MNT_FORCE) == 0) {
+				error = EBUSY;
+				goto err;
+			}
+			error = vfs_busy(mp, VB_WRITE|VB_WAIT);
+			if (error) {
+				if ((flags & MNT_DOOMED)) {
+					/*
+					 * If the mount point was busy due to 
+					 * being unmounted, it has been removed
+					 * from the mount list already.
+					 * Restart the iteration from the last 
+					 * collected busy entry.
+					 */
+					mp = SLIST_FIRST(&mplist);
+					break;
+				}
+				goto err;
+			}
+			SLIST_INSERT_HEAD(&mplist, mp, mnt_dounmount);
+			break;
+		}
+	}
+
+	/* 
+	 * Nested mount points cannot appear during this loop as mounting
+	 * requires a read lock for the parent mount point.
+	 */
+	while ((mp = SLIST_FIRST(&mplist)) != NULL) {
+		SLIST_REMOVE(&mplist, mp, mount, mnt_dounmount);
+		error = dounmount_leaf(mp, flags, p);
+		if (error)
+			goto err;
+	}
+	return (0);
+
+err:
+	while ((mp = SLIST_FIRST(&mplist)) != NULL) {
+		SLIST_REMOVE(&mplist, mp, mount, mnt_dounmount);
+		vfs_unbusy(mp);
+	}
+	return (error);
+}
+
+int
+dounmount_leaf(struct mount *mp, int flags, struct proc *p)
 {
 	struct vnode *coveredvp;
 	int error;
diff --git a/sys/sys/mount.h b/sys/sys/mount.h
index 4c5d62658cf..90ce21e2117 100644
--- a/sys/sys/mount.h
+++ b/sys/sys/mount.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: mount.h,v 1.128 2017/01/10 19:48:32 bluhm Exp $	*/
+/*	$OpenBSD: mount.h,v 1.129 2017/04/15 13:56:43 bluhm Exp $	*/
 /*	$NetBSD: mount.h,v 1.48 1996/02/18 11:55:47 fvdl Exp $	*/
 
 /*
@@ -347,6 +347,7 @@ LIST_HEAD(vnodelst, vnode);
 
 struct mount {
 	TAILQ_ENTRY(mount) mnt_list;		/* mount list */
+	SLIST_ENTRY(mount) mnt_dounmount;	/* unmount work queue */
 	const struct vfsops *mnt_op;		/* operations on fs */
 	struct vfsconf  *mnt_vfc;               /* configuration info */
 	struct vnode	*mnt_vnodecovered;	/* vnode we mounted on */
-- 
cgit v1.2.3