Print a warning when we miss vinactive() call, because of race in vget().

The race is very real, but conditions needed for triggering it are rather
hard to meet now.
When gjournal will be committed (where it is quite easy to trigger) we need
to fix it.

For now, verify if it is really hard to trigger.

Discussed with:	kan

1 files changed, 20 insertions, 0 deletions

diff --git a/sys/kern/vfs_subr.c b/sys/kern/vfs_subr.c
index 6f5f375..a57b2ba 100644
--- a/sys/kern/vfs_subr.c
+++ b/sys/kern/vfs_subr.c
@@ -1928,6 +1928,7 @@ vget(vp, flags, td)
 {
 	int oweinact;
 	int oldflags;
+	int usecount;
 	int error;
 
 	error = 0;
@@ -1949,6 +1950,7 @@ vget(vp, flags, td)
 		flags |= LK_EXCLUSIVE;
 		oweinact = 1;
 	}
+	usecount = vp->v_usecount;
 	v_incr_usecount(vp);
 	if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
 		VI_LOCK(vp);
@@ -1959,6 +1961,24 @@ vget(vp, flags, td)
 		 * active.
 		 */
 		v_decr_usecount(vp);
+		/*
+		 * Print warning when race below occur:
+		 *
+		 * thread1	thread2
+		 * -------	-------
+		 *					v_usecount=0
+		 * vref(vp)				v_usecount=1
+		 *		vget(vp)
+		 *		v_incr_usecount(vp)	v_usecount=2
+		 *		vn_lock(vp)
+		 * vrele(vp)				v_usecount=1
+		 *		v_decr_usecount(vp)	v_usecount=0
+		 *
+		 * In such situation VOP_INACTIVE() will not be called for
+		 * the vnode vp.
+		 */
+		if (usecount > 0 && vp->v_usecount == 0)
+			printf("vinactive() won't be called for vp=%p\n", vp);
 		return (error);
 	}
 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)