Do not translate the IN_ACCESS inode flag into the IN_MODIFIED while filesystem

is suspending/suspended. Doing so may result in deadlock. Instead, set the
(new) IN_LAZYACCESS flag, that becomes IN_MODIFIED when suspend is lifted.

Change the locking protocol in order to set the IN_ACCESS and timestamps
without upgrading shared vnode lock to exclusive (see comments in the
inode.h). Before that, inode was modified while holding only shared
lock.

Tested by:	Peter Holm
Reviewed by:	tegge, bde
Approved by:	pjd (mentor)
MFC after:	3 weeks

3 files changed, 33 insertions, 13 deletions

diff --git a/sys/ufs/ffs/ffs_inode.c b/sys/ufs/ffs/ffs_inode.c
index 33b65bf..ad6bff9 100644
--- a/sys/ufs/ffs/ffs_inode.c
+++ b/sys/ufs/ffs/ffs_inode.c
@@ -66,9 +66,11 @@ static int ffs_indirtrunc(struct inode *, ufs2_daddr_t, ufs2_daddr_t,
  * IN_ACCESS, IN_UPDATE, and IN_CHANGE flags respectively.  Write the inode
  * to disk if the IN_MODIFIED flag is set (it may be set initially, or by
  * the timestamp update).  The IN_LAZYMOD flag is set to force a write
- * later if not now.  If we write now, then clear both IN_MODIFIED and
- * IN_LAZYMOD to reflect the presumably successful write, and if waitfor is
- * set, then wait for the write to complete.
+ * later if not now.  The IN_LAZYACCESS is set instead of IN_MODIFIED if the fs
+ * is currently being suspended (or is suspended) and vnode has been accessed.
+ * If we write now, then clear IN_MODIFIED, IN_LAZYACCESS and IN_LAZYMOD to
+ * reflect the presumably successful write, and if waitfor is set, then wait
+ * for the write to complete.
  */
 int
 ffs_update(vp, waitfor)
@@ -80,12 +82,12 @@ ffs_update(vp, waitfor)
 	struct inode *ip;
 	int error;
 
-	ASSERT_VOP_LOCKED(vp, "ffs_update");
+	ASSERT_VOP_ELOCKED(vp, "ffs_update");
 	ufs_itimes(vp);
 	ip = VTOI(vp);
 	if ((ip->i_flag & IN_MODIFIED) == 0 && waitfor == 0)
 		return (0);
-	ip->i_flag &= ~(IN_LAZYMOD | IN_MODIFIED);
+	ip->i_flag &= ~(IN_LAZYACCESS | IN_LAZYMOD | IN_MODIFIED);
 	fs = ip->i_fs;
 	if (fs->fs_ronly)
 		return (0);
diff --git a/sys/ufs/ffs/ffs_snapshot.c b/sys/ufs/ffs/ffs_snapshot.c
index 473dbfd..6ab9af2 100644
--- a/sys/ufs/ffs/ffs_snapshot.c
+++ b/sys/ufs/ffs/ffs_snapshot.c
@@ -2309,15 +2309,16 @@ readblock(vp, bp, lbn)
 	return (bp->b_error);
 }
 
-
 /*
  * Process file deletes that were deferred by ufs_inactive() due to
- * the file system being suspended.
+ * the file system being suspended. Transfer IN_LAZYACCESS into
+ * IN_MODIFIED for vnodes that were accessed during suspension.
  */
 static void
 process_deferred_inactive(struct mount *mp)
 {
 	struct vnode *vp, *mvp;
+	struct inode *ip;
 	struct thread *td;
 	int error;
 
@@ -2327,9 +2328,15 @@ process_deferred_inactive(struct mount *mp)
  loop:
 	MNT_VNODE_FOREACH(vp, mp, mvp) {
 		VI_LOCK(vp);
-		if ((vp->v_iflag & (VI_DOOMED | VI_OWEINACT)) != VI_OWEINACT ||
-		    vp->v_usecount > 0 ||
-		    vp->v_type == VNON) {
+		/*
+		 * IN_LAZYACCESS is checked here without holding any
+		 * vnode lock, but this flag is set only while holding
+		 * vnode interlock.
+		 */
+		if (vp->v_type == VNON || (vp->v_iflag & VI_DOOMED) != 0 ||
+		    ((VTOI(vp)->i_flag & IN_LAZYACCESS) == 0 &&
+			((vp->v_iflag & VI_OWEINACT) == 0 ||
+			vp->v_usecount > 0))) {
 			VI_UNLOCK(vp);
 			continue;
 		}
@@ -2344,8 +2351,13 @@ process_deferred_inactive(struct mount *mp)
 			MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp);
 			goto loop;
 		}
+		ip = VTOI(vp);
+		if ((ip->i_flag & IN_LAZYACCESS) != 0) {
+			ip->i_flag &= ~IN_LAZYACCESS;
+			ip->i_flag |= IN_MODIFIED;
+		}
 		VI_LOCK(vp);
-		if ((vp->v_iflag & VI_OWEINACT) == 0) {
+		if ((vp->v_iflag & VI_OWEINACT) == 0 || vp->v_usecount > 0) {
 			VI_UNLOCK(vp);
 			VOP_UNLOCK(vp, 0, td);
 			vdrop(vp);
diff --git a/sys/ufs/ffs/ffs_vnops.c b/sys/ufs/ffs/ffs_vnops.c
index 91d2c01..4d8d9ef 100644
--- a/sys/ufs/ffs/ffs_vnops.c
+++ b/sys/ufs/ffs/ffs_vnops.c
@@ -593,8 +593,11 @@ ffs_read(ap)
 	}
 
 	if ((error == 0 || uio->uio_resid != orig_resid) &&
-	    (vp->v_mount->mnt_flag & MNT_NOATIME) == 0)
+	    (vp->v_mount->mnt_flag & MNT_NOATIME) == 0) {
+		VI_LOCK(vp);
 		ip->i_flag |= IN_ACCESS;
+		VI_UNLOCK(vp);
+	}
 	return (error);
 }
 
@@ -989,8 +992,11 @@ ffs_extread(struct vnode *vp, struct uio *uio, int ioflag)
 	}
 
 	if ((error == 0 || uio->uio_resid != orig_resid) &&
-	    (vp->v_mount->mnt_flag & MNT_NOATIME) == 0)
+	    (vp->v_mount->mnt_flag & MNT_NOATIME) == 0) {
+		VI_LOCK(vp);
 		ip->i_flag |= IN_ACCESS;
+		VI_UNLOCK(vp);
+	}
 	return (error);
 }