From 8fe6bddff703fa63fca6ddb08f5fb8dc6fa31496 Mon Sep 17 00:00:00 2001
From: Rick Macklem <rmacklem@FreeBSD.org>
Date: Sun, 1 Sep 2013 23:02:59 +0000
Subject: [PATCH] Forced dismounts of NFS mounts can fail when thread(s) are
 stuck waiting for an RPC reply from the server while holding the mount point
 busy (mnt_lockref incremented). This happens because dounmount() msleep()s
 waiting for mnt_lockref to become 0, before calling VFS_UNMOUNT(). This patch
 adds a new VFS operation called VFS_PURGE(), which the NFS client implements
 as purging RPCs in progress. Making this call before checking mnt_lockref
 fixes the problem, by ensuring that the VOP_xxx() calls will fail and unbusy
 the mount point.

Reported by:	sbruno
Reviewed by:	kib
MFC after:	2 weeks
---
 sys/fs/nfsclient/nfs_clvfsops.c | 15 +++++++++++++++
 sys/kern/vfs_mount.c            | 10 +++++++++-
 sys/sys/mount.h                 | 10 ++++++++++
 3 files changed, 34 insertions(+), 1 deletion(-)

diff --git a/sys/fs/nfsclient/nfs_clvfsops.c b/sys/fs/nfsclient/nfs_clvfsops.c
index 863c418d1b04..4a180c52ed6e 100644
--- a/sys/fs/nfsclient/nfs_clvfsops.c
+++ b/sys/fs/nfsclient/nfs_clvfsops.c
@@ -120,6 +120,7 @@ static vfs_root_t nfs_root;
 static vfs_statfs_t nfs_statfs;
 static vfs_sync_t nfs_sync;
 static vfs_sysctl_t nfs_sysctl;
+static vfs_purge_t nfs_purge;
 
 /*
  * nfs vfs operations.
@@ -134,6 +135,7 @@ static struct vfsops nfs_vfsops = {
 	.vfs_uninit =		ncl_uninit,
 	.vfs_unmount =		nfs_unmount,
 	.vfs_sysctl =		nfs_sysctl,
+	.vfs_purge =		nfs_purge,
 };
 VFS_SET(nfs_vfsops, nfs, VFCF_NETWORK | VFCF_SBDRY);
 
@@ -1675,6 +1677,19 @@ nfs_sysctl(struct mount *mp, fsctlop_t op, struct sysctl_req *req)
 	return (0);
 }
 
+/*
+ * Purge any RPCs in progress, so that they will all return errors.
+ * This allows dounmount() to continue as far as VFS_UNMOUNT() for a
+ * forced dismount.
+ */
+static void
+nfs_purge(struct mount *mp)
+{
+	struct nfsmount *nmp = VFSTONFS(mp);
+
+	newnfs_nmcancelreqs(nmp);
+}
+
 /*
  * Extract the information needed by the nlm from the nfs vnode.
  */
diff --git a/sys/kern/vfs_mount.c b/sys/kern/vfs_mount.c
index 493bb9888836..8f92e10b69b9 100644
--- a/sys/kern/vfs_mount.c
+++ b/sys/kern/vfs_mount.c
@@ -1269,8 +1269,16 @@ dounmount(mp, flags, td)
 	}
 	mp->mnt_kern_flag |= MNTK_UNMOUNT | MNTK_NOINSMNTQ;
 	/* Allow filesystems to detect that a forced unmount is in progress. */
-	if (flags & MNT_FORCE)
+	if (flags & MNT_FORCE) {
 		mp->mnt_kern_flag |= MNTK_UNMOUNTF;
+		MNT_IUNLOCK(mp);
+		/*
+		 * Must be done after setting MNTK_UNMOUNTF and before
+		 * waiting for mnt_lockref to become 0.
+		 */
+		VFS_PURGE(mp);
+		MNT_ILOCK(mp);
+	}
 	error = 0;
 	if (mp->mnt_lockref) {
 		mp->mnt_kern_flag |= MNTK_DRAINING;
diff --git a/sys/sys/mount.h b/sys/sys/mount.h
index 1a835b7e5280..8f9445154b06 100644
--- a/sys/sys/mount.h
+++ b/sys/sys/mount.h
@@ -609,6 +609,7 @@ typedef int vfs_sysctl_t(struct mount *mp, fsctlop_t op,
 		    struct sysctl_req *req);
 typedef void vfs_susp_clean_t(struct mount *mp);
 typedef void vfs_notify_lowervp_t(struct mount *mp, struct vnode *lowervp);
+typedef void vfs_purge_t(struct mount *mp);
 
 struct vfsops {
 	vfs_mount_t		*vfs_mount;
@@ -628,6 +629,7 @@ struct vfsops {
 	vfs_susp_clean_t	*vfs_susp_clean;
 	vfs_notify_lowervp_t	*vfs_reclaim_lowervp;
 	vfs_notify_lowervp_t	*vfs_unlink_lowervp;
+	vfs_purge_t		*vfs_purge;
 	vfs_mount_t		*vfs_spare[6];	/* spares for ABI compat */
 };
 
@@ -757,6 +759,14 @@ vfs_statfs_t	__vfs_statfs;
 	}								\
 } while (0)
 
+#define	VFS_PURGE(MP) do {						\
+	if (*(MP)->mnt_op->vfs_purge != NULL) {				\
+		VFS_PROLOGUE(MP);					\
+		(*(MP)->mnt_op->vfs_purge)(MP);				\
+		VFS_EPILOGUE(MP);					\
+	}								\
+} while (0)
+
 #define VFS_KNOTE_LOCKED(vp, hint) do					\
 {									\
 	if (((vp)->v_vflag & VV_NOKNOTE) == 0)				\