On 19/04/2021 13:37, Han-Wen Nienhuys via GitGitGadget wrote:

From: Han-Wen Nienhuys <redacted>

[...snip...]

void *src_rec,

+					  int hash_size)
+{
+	struct reftable_obj_record *obj = rec;
+	const struct reftable_obj_record *src =
+		(const struct reftable_obj_record *)src_rec;
+	int olen;
+
+	reftable_obj_record_release(obj);
+	*obj = *src;
+	obj->hash_prefix = reftable_malloc(obj->hash_prefix_len);
+	memcpy(obj->hash_prefix, src->hash_prefix, obj->hash_prefix_len);
+
+	olen = obj->offset_len * sizeof(uint64_t);
+	obj->offsets = reftable_malloc(olen);
+	memcpy(obj->offsets, src->offsets, olen);
+}

If src->offset_len is 0 then offsets will be NULL - and passing NULL 
into memcpy() results in undefined behaviour. I think we should either 
add an "if (src->offset_len)" check around the memcpy, or perhaps switch 
to COPY_ARRAY() which performs that check for us. (We can probably also 
skip the malloc and hence also olen calculation in this scenario though, 
because obj->offsets should already be NULL if src->offsets was NULL?)

Specifically, if I run t0032 as such:

   make CC=clang-11 UBSAN_OPTIONS=print_stacktrace=1 SANITIZE=undefined 
COPTS="-Og -g" GIT_TEST_OPTS=-v T=t0032-reftable-unittest.sh test

UBSAN reports the following error:

reftable/record.c:475:23: runtime error: null pointer passed as argument 
2, which is declared to never be null
/usr/include/string.h:43:28: note: nonnull attribute specified here
     #0 0x698e52 in reftable_obj_record_copy_from record.c:475:2
     #1 0x6ac806 in test_copy record_test.c:21:2
     #2 0x6abb73 in test_reftable_obj_record_roundtrip record_test.c:334:3
     #3 0x6abb73 in record_test_main record_test.c:403:2
     #4 0x437f08 in cmd__reftable test-reftable.c:10:2
     #5 0x427e74 in cmd_main test-tool.c:126:11
     #6 0x428043 in main common-main.c:52:11
     #7 0x7f3cd1044349 in __libc_start_main (/lib64/libc.so.6+0x24349)
     #8 0x407209 in _start start.S:120

SUMMARY: UndefinedBehaviorSanitizer: undefined-behavior 
reftable/record.c:475:23 in

[...snip...]

+static int reftable_obj_record_decode(void *rec, struct strbuf key,
+				      uint8_t val_type, struct string_view in,
+				      int hash_size)
+{
+	struct string_view start = in;
+	struct reftable_obj_record *r = rec;
+	uint64_t count = val_type;
+	int n = 0;
+	uint64_t last;
+	int j;
+	r->hash_prefix = reftable_malloc(key.len);
+	memcpy(r->hash_prefix, key.buf, key.len);
+	r->hash_prefix_len = key.len;
+
+	if (val_type == 0) {
+		n = get_var_int(&count, &in);
+		if (n < 0) {
+			return n;
+		}
+
+		string_view_consume(&in, n);
+	}
+
+	r->offsets = NULL;
+	r->offset_len = 0;
+	if (count == 0)
+		return start.len - in.len;

And it looks like creating a reftable_obj_record with offsets = NULL and 
offset_len = 0 is considered valid based on the code above.

[... snipped the rest ...]