Re: [PATCH net] bpf: don't zero out the info struct in bpf_obj_get_info_by_fd()

On Tue, 25 Jul 2017 16:15:47 -0700, Jakub Kicinski wrote:

On Wed, 26 Jul 2017 00:59:49 +0200, Daniel Borkmann wrote:

quoted

quoted

@@ -418,6 +420,8 @@ static void test_bpf_obj_id(void)
  		nr_id_found++;

  		err = bpf_obj_get_info_by_fd(prog_fd, &prog_info, &info_len);
+		prog_infos[i].jited_prog_insns = 0;
+		prog_infos[i].xlated_prog_insns = 0;    

Can you elaborate why this one above is needed?  

Ah, I removed the comment about it at the last minute.  The check below
compares the info we get here with info we got reading the programs in
the earlier loop - using memcmp().  This call, however, doesn't fill in
the pointers for jited and xlated images, so the memcmp() would fail.

It used to work when bpf_obj_get_info_by_fd() was zeroing info, since
the pointers would be cleared by it, and no dump ever returned, it
didn't matter that the call sites differ.

FWIW the comment was this:
+               /* Clear the insns pointers, we're not requesting dumps here.   
+                * Otherwise the byte-by-byte comparison below would fail.      
+                */ 

quoted

quoted

  		CHECK(err || info_len != sizeof(struct bpf_prog_info) ||
  		      memcmp(&prog_info, &prog_infos[i], info_len),
  		      "get-prog-info(next_id->fd)",