inotify: Work around the vput() bug directly

For 15.0, apply a minimal fix which at least ensures that inotify can't
trigger the latent race described in commit 99cb3dca4773 ("vnode: Rework
vput() to avoid holding the vnode lock after decrementing").

Approved by:	re (cperciva)
Reviewed by:	olce, kib
MFC after:	3 days
Differential Revision:	https://reviews.freebsd.org/D53774

(cherry picked from commit ebc17879f0885ca87644980f6275b9759b311eb3)
(cherry picked from commit 1f6e3abf41718e8e4a309be122f0a6048e9c5772)

1 files changed, 8 insertions, 1 deletions

diff --git a/sys/kern/vfs_inotify.c b/sys/kern/vfs_inotify.c
index b265a5ff3a62..de8f99ea8d2f 100644
--- a/sys/kern/vfs_inotify.c
+++ b/sys/kern/vfs_inotify.c
@@ -380,7 +380,14 @@ inotify_unlink_watch_locked(struct inotify_softc *sc, struct inotify_watch *watc
 static void
 inotify_free_watch(struct inotify_watch *watch)
 {
-	vrele(watch->vp);
+	/*
+	 * Formally, we don't need to lock the vnode here.  However, if we
+	 * don't, and vrele() releases the last reference, it's possible the
+	 * vnode will be recycled while a different thread holds the vnode lock.
+	 * Work around this bug by acquiring the lock here.
+	 */
+	(void)vn_lock(watch->vp, LK_EXCLUSIVE | LK_RETRY);
+	vput(watch->vp);
 	free(watch, M_INOTIFY);
 }