Fix unallocated object detection for large_dnode datasets

Fix dmu_object_next() to correctly handle unallocated objects on
large_dnode datasets.

We implement this by scanning the dnode block until we find the correct
offset to be used in dnode_next_offset(). This is necessary because we
can't assume *objectp is a hole even if dmu_object_info() returns
ENOENT.

This fixes a couple of issues with zfs receive on large_dnode datasets.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: loli10K <ezomori.nozomu@gmail.com>
Closes #5027 
Closes #5532

module/zfs/dmu_object.c

@@ -237,28 +237,39 @@ int
 dmu_object_next(objset_t *os, uint64_t *objectp, boolean_t hole, uint64_t txg)
 {
 	uint64_t offset;
-	dmu_object_info_t doi;
+	uint64_t start_obj;
 	struct dsl_dataset *ds = os->os_dsl_dataset;
-	int dnodesize;
 	int error;
 
-	/*
+	if (*objectp == 0) {
-	 * Avoid expensive dnode hold if this dataset doesn't use large dnodes.
+		start_obj = 1;
-	 */
+	} else if (ds && ds->ds_feature_inuse[SPA_FEATURE_LARGE_DNODE]) {
-	if (ds && ds->ds_feature_inuse[SPA_FEATURE_LARGE_DNODE]) {
+		/*
-		error = dmu_object_info(os, *objectp, &doi);
+		 * For large_dnode datasets, scan from the beginning of the
-		if (error && !(error == EINVAL && *objectp == 0))
+		 * dnode block to find the starting offset. This is needed
-			return (SET_ERROR(error));
+		 * because objectp could be part of a large dnode so we can't
-		else
+		 * assume it's a hole even if dmu_object_info() returns ENOENT.
-			dnodesize = doi.doi_dnodesize;
+		 */
+		int epb = DNODE_BLOCK_SIZE >> DNODE_SHIFT;
+		int skip;
+		uint64_t i;
+
+		for (i = *objectp & ~(epb - 1); i <= *objectp; i += skip) {
+			dmu_object_info_t doi;
+
+			error = dmu_object_info(os, i, &doi);
+			if (error)
+				skip = 1;
+			else
+				skip = doi.doi_dnodesize >> DNODE_SHIFT;
+		}
+
+		start_obj = i;
 	} else {
-		dnodesize = DNODE_MIN_SIZE;
+		start_obj = *objectp + 1;
 	}
 
-	if (*objectp == 0)
+	offset = start_obj << DNODE_SHIFT;
-		offset = 1 << DNODE_SHIFT;
-	else
-		offset = (*objectp << DNODE_SHIFT) + dnodesize;
 
 	error = dnode_next_offset(DMU_META_DNODE(os),
 	    (hole ? DNODE_FIND_HOLE : 0), &offset, 0, DNODES_PER_BLOCK, txg);

module/zfs/dnode.c

@@ -1184,6 +1184,7 @@ dnode_is_free(dmu_buf_impl_t *db, int idx, int slots)
  * errors:
  * EINVAL - invalid object number.
  * ENOSPC - hole too small to fulfill "slots" request
+ * ENOENT - the requested dnode is not allocated
  * EIO - i/o error.
  * succeeds even for free dnodes.
  */

tests/zfs-tests/tests/functional/features/large_dnode/large_dnode_005_pos.ksh

@@ -27,9 +27,11 @@ verify_runnable "both"
 TEST_SEND_FS=$TESTPOOL/send_large_dnode
 TEST_RECV_FS=$TESTPOOL/recv_large_dnode
 TEST_SNAP=$TEST_SEND_FS@ldnsnap
+TEST_SNAPINCR=$TEST_SEND_FS@ldnsnap_incr
 TEST_STREAM=$TESTDIR/ldnsnap
+TEST_STREAMINCR=$TESTDIR/ldnsnap_incr
 TEST_FILE=foo
-
+TEST_FILEINCR=bar
 
 function cleanup
 {
@@ -42,6 +44,7 @@ function cleanup
 	fi
 
 	rm -f $TEST_STREAM
+	rm -f $TEST_STREAMINCR
 }
 
 log_onexit cleanup
@@ -49,10 +52,13 @@ log_onexit cleanup
 log_assert "zfs send stream with large dnodes accepted by new pool"
 
 log_must $ZFS create -o dnodesize=1k $TEST_SEND_FS
-log_must touch /$TEST_SEND_FS/$TEST_FILE
+log_must $TOUCH /$TEST_SEND_FS/$TEST_FILE
-log_must $ZFS umount $TEST_SEND_FS
 log_must $ZFS snap $TEST_SNAP
 log_must $ZFS send $TEST_SNAP > $TEST_STREAM
+log_must $RM -f /$TEST_SEND_FS/$TEST_FILE
+log_must $TOUCH /$TEST_SEND_FS/$TEST_FILEINCR
+log_must $ZFS snap $TEST_SNAPINCR
+log_must $ZFS send -i $TEST_SNAP $TEST_SNAPINCR > $TEST_STREAMINCR
 
 log_must eval "$ZFS recv $TEST_RECV_FS < $TEST_STREAM"
 inode=$(ls -li /$TEST_RECV_FS/$TEST_FILE | awk '{print $1}')
@@ -61,4 +67,9 @@ if [[ "$dnsize" != "1K" ]]; then
 	log_fail "dnode size is $dnsize (expected 1K)"
 fi
 
+log_must eval "$ZFS recv -F $TEST_RECV_FS < $TEST_STREAMINCR"
+log_must $DIFF -r /$TEST_SEND_FS /$TEST_RECV_FS
+log_must $ZFS umount $TEST_SEND_FS
+log_must $ZFS umount $TEST_RECV_FS
+
 log_pass