kern/132068: [zfs] page fault when using ZFS over NFS on 7.1-RELEASE/amd64

[Jaakko Heinonen]

The following reply was made to PR kern/132068; it has been noted by GNATS.

From: Jaakko Heinonen <jh at saunalahti.fi>
To: Edward Fisk <7ogcg7g02 at sneakemail.com>
Cc: bug-followup at FreeBSD.org, Weldon Godfrey <wgodfrey at ena.com>
Subject: Re: kern/132068: [zfs] page fault when using ZFS over NFS on
	7.1-RELEASE/amd64
Date: Fri, 10 Apr 2009 14:46:02 +0300

 On 2009-03-26, Jaakko Heinonen wrote:
 > I now know what is going on. The vnode may be reclaimed during
 > zfs_zget() because it doesn't hold the vnode lock (except when a new
 > znode is created).
 
 OK, I have now put together a patch which should avoid the original
 panic you reported. The same panic was also reported by Weldon Godfrey
 (Cc'd) on -fs:
 http://lists.freebsd.org/pipermail/freebsd-fs/2008-August/004998.html
 
 --- patch begins here ---
 Index: sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zfs_znode.c
 ===================================================================
 --- sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zfs_znode.c	(revision 190593)
 +++ sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zfs_znode.c	(working copy)
 @@ -890,8 +890,9 @@ again:
  		if (zp->z_unlinked) {
  			err = ENOENT;
  		} else {
 -			if (ZTOV(zp) != NULL)
 -				VN_HOLD(ZTOV(zp));
 +			vp = ZTOV(zp);
 +			if (vp != NULL)
 +				VN_HOLD(vp);
  			else {
  				if (first) {
  					ZFS_LOG(1, "dying znode detected (zp=%p)", zp);
 @@ -907,12 +908,25 @@ again:
  				tsleep(zp, 0, "zcollide", 1);
  				goto again;
  			}
 -			*zpp = zp;
  			err = 0;
  		}
  		dmu_buf_rele(db, NULL);
  		mutex_exit(&zp->z_lock);
  		ZFS_OBJ_HOLD_EXIT(zfsvfs, obj_num);
 +		if (err == 0) {
 +			/*
 +			 * Check if we lost lost race against reclaim.
 +			 */
 +			VI_LOCK(vp);
 +			if (vp->v_iflag & VI_DOOMED) {
 +				VI_UNLOCK(vp);
 +				VN_RELE(vp);
 +				ZFS_LOG(1, "doomed vnode detected (zp=%p)", zp);
 +				goto again;
 +			}
 +			VI_UNLOCK(vp);
 +			*zpp = zp;
 +		}
  		return (err);
  	}
 --- patch ends here ---
 
 The fix isn't perfect. Vnodes may be still reclaimed during zfs_zget()
 in forced unmount case. However zfs doesn't support forced unmounts at
 all right now. The patch is againt 8.0-CURRENT.
 
 -- 
 Jaakko