Stop ignoring ERELOOKUP from VOP_INACTIVE()

When possible, relock the vnode and retry inactivation.  Only vunref() is
required not to drop the vnode lock, so handle it specially by not retrying.

This is a part of the efforts to ensure that unlinked not referenced vnode
does not prevent inode from reusing.

Reviewed by:	chs, mckusick
Tested by:	pho
MFC after:	2 weeks
Sponsored by:	The FreeBSD Foundation

3 files changed, 42 insertions, 16 deletions

diff --git a/sys/kern/vfs_subr.c b/sys/kern/vfs_subr.c
index 047e4c54f0c5..04cd0e0175f9 100644
--- a/sys/kern/vfs_subr.c
+++ b/sys/kern/vfs_subr.c
@@ -3159,9 +3159,21 @@ vput_final(struct vnode *vp, enum vput_op func)
 		break;
 	}
 	if (error == 0) {
-		vinactive(vp);
-		if (want_unlock)
-			VOP_UNLOCK(vp);
+		if (func == VUNREF) {
+			VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
+			    ("recursive vunref"));
+			vp->v_vflag |= VV_UNREF;
+		}
+		for (;;) {
+			error = vinactive(vp);
+			if (want_unlock)
+				VOP_UNLOCK(vp);
+			if (error != ERELOOKUP || !want_unlock)
+				break;
+			VOP_LOCK(vp, LK_EXCLUSIVE);
+		}
+		if (func == VUNREF)
+			vp->v_vflag &= ~VV_UNREF;
 		vdropl(vp);
 	} else {
 		vdefer_inactive(vp);
@@ -3546,10 +3558,11 @@ vdropl(struct vnode *vp)
  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
  */
-static void
+static int
 vinactivef(struct vnode *vp)
 {
 	struct vm_object *obj;
+	int error;
 
 	ASSERT_VOP_ELOCKED(vp, "vinactive");
 	ASSERT_VI_LOCKED(vp, "vinactive");
@@ -3575,14 +3588,15 @@ vinactivef(struct vnode *vp)
 		vm_object_page_clean(obj, 0, 0, 0);
 		VM_OBJECT_WUNLOCK(obj);
 	}
-	VOP_INACTIVE(vp);
+	error = VOP_INACTIVE(vp);
 	VI_LOCK(vp);
 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
 	    ("vinactive: lost VI_DOINGINACT"));
 	vp->v_iflag &= ~VI_DOINGINACT;
+	return (error);
 }
 
-void
+int
 vinactive(struct vnode *vp)
 {
 
@@ -3591,14 +3605,14 @@ vinactive(struct vnode *vp)
 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
 
 	if ((vp->v_iflag & VI_OWEINACT) == 0)
-		return;
+		return (0);
 	if (vp->v_iflag & VI_DOINGINACT)
-		return;
+		return (0);
 	if (vp->v_usecount > 0) {
 		vp->v_iflag &= ~VI_OWEINACT;
-		return;
+		return (0);
 	}
-	vinactivef(vp);
+	return (vinactivef(vp));
 }
 
 /*
@@ -3911,10 +3925,15 @@ vgonel(struct vnode *vp)
 	 */
 	if (active)
 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
-	if ((oweinact || active) && !doinginact) {
-		VI_LOCK(vp);
-		vinactivef(vp);
-		VI_UNLOCK(vp);
+	if (!doinginact) {
+		do {
+			if (oweinact || active) {
+				VI_LOCK(vp);
+				vinactivef(vp);
+				oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
+				VI_UNLOCK(vp);
+			}
+		} while (oweinact);
 	}
 	if (vp->v_type == VSOCK)
 		vfs_unp_reclaim(vp);
diff --git a/sys/sys/vnode.h b/sys/sys/vnode.h
index 78fbec1bd0ba..639a16881e09 100644
--- a/sys/sys/vnode.h
+++ b/sys/sys/vnode.h
@@ -269,6 +269,7 @@ struct xvnode {
 #define	VV_MD		0x0800	/* vnode backs the md device */
 #define	VV_FORCEINSMQ	0x1000	/* force the insmntque to succeed */
 #define	VV_READLINK	0x2000	/* fdescfs linux vnode */
+#define	VV_UNREF	0x4000	/* vunref, do not drop lock in inactive() */
 
 #define	VMP_LAZYLIST	0x0001	/* Vnode is on mnt's lazy list */
 
@@ -710,7 +711,7 @@ void	vgone(struct vnode *vp);
 void	vhold(struct vnode *);
 void	vholdnz(struct vnode *);
 bool	vhold_smr(struct vnode *);
-void	vinactive(struct vnode *vp);
+int	vinactive(struct vnode *vp);
 int	vinvalbuf(struct vnode *vp, int save, int slpflag, int slptimeo);
 int	vtruncbuf(struct vnode *vp, off_t length, int blksize);
 void	v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
diff --git a/sys/ufs/ffs/ffs_snapshot.c b/sys/ufs/ffs/ffs_snapshot.c
index b5daec14decf..72c8061917d8 100644
--- a/sys/ufs/ffs/ffs_snapshot.c
+++ b/sys/ufs/ffs/ffs_snapshot.c
@@ -2595,6 +2595,7 @@ process_deferred_inactive(struct mount *mp)
 			continue;
 		}
 		vholdl(vp);
+retry_vnode:
 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
 		if (error != 0) {
 			vdrop(vp);
@@ -2609,7 +2610,12 @@ process_deferred_inactive(struct mount *mp)
 			UFS_INODE_SET_FLAG(ip, IN_MODIFIED);
 		}
 		VI_LOCK(vp);
-		vinactive(vp);
+		error = vinactive(vp);
+		if (error == ERELOOKUP && vp->v_usecount == 0) {
+			VI_UNLOCK(vp);
+			VOP_UNLOCK(vp);
+			goto retry_vnode;
+		}
 		VI_UNLOCK(vp);
 		VOP_UNLOCK(vp);
 		vdrop(vp);