skip FREEOBJECTS for objects which can't exist

when sending an incremental stream based on a snapshot, the receiving side must have the same base snapshot. thus we do not need to send FREEOBJECTS records for any objects past the maximum one which exists locally. this allows us to send incremental streams (again) to older ZFS implementations (e.g. ZoL < 0.7) which actually try to free all objects in a FREEOBJECTS record, instead of bailing out early. Signed-off-by: Fabian Grünbichler <[email protected]>
openzfs · Oct 3, 2017 · 4adab0c · 4adab0c
1 parent 39f5662
commit 4adab0c
Showing 1 changed file with 16 additions and 0 deletions.
diff --git a/module/zfs/dmu_send.c b/module/zfs/dmu_send.c
@@ -454,6 +454,22 @@ static int
 dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
 {
 	struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
+	uint64_t maxobj = DNODES_PER_BLOCK *
+	    (DMU_META_DNODE(dsp->dsa_os)->dn_maxblkid + 1);
+
+	/*
+	 * ZoL < 0.7 does not handle large FREEOBJECTS records correctly,
+	 * leading to zfs recv never completing. to avoid this issue, don't
+	 * send FREEOBJECTS records for object IDs which cannot exist on the
+	 * receiving side.
+	 */
+	if (maxobj > 0) {
+		if (maxobj < firstobj)
+			return (0);
+
+		if (maxobj < firstobj + numobjs)
+			numobjs = maxobj - firstobj;
+	}
 
 	/*
 	 * If there is a pending op, but it's not PENDING_FREEOBJECTS,