Re: [PATCH v2 net-next 1/2] bpf: allow extended BPF programs access skb fields

On 03/14/2015 05:59 AM, Alexei Starovoitov wrote:

On 3/13/15 7:27 PM, Alexei Starovoitov wrote:

quoted

On 3/13/15 7:16 PM, Daniel Borkmann wrote:

quoted

On 03/14/2015 03:08 AM, Alexei Starovoitov wrote:

quoted

On 3/13/15 7:06 PM, Daniel Borkmann wrote:

quoted

On 03/14/2015 02:46 AM, Daniel Borkmann wrote:
...

quoted

Previously, it was much more consistent, which I like better. And only
because of the simple BUILD_BUG_ON()? :/

Alternative is to move all of them into a central place, something like
in twsk_build_assert() or __mld2_query_bugs[].

nope. that defeats the purpose of bug_on.

Well, it doesn't. ;) It throws a build error thus the user is forced to
investigate that further.

according to this distorted logic all build_bug_on can be in one file
across the whole tree, since 'user is forced to investigate' ?!

That was not what I was suggesting, and I assume you know that ...

also note that this case and twsk_build_assert are different.
twsk_build_assert has no other choice then to have one function
that covers logic in the whole file, whereas in this patch:
+               BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
+               *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
+                                     offsetof(struct sk_buff, mark));

the build_bug_on protect the line directly below.
Separating them just doesn't make sense at all.

I also like the above approach better, I only suggested that as a
possible alternative since you were saying earlier in this thread:

   I thought about it, but didn't add it, since we already have them
   in the same file several lines above this spot. I think one build
   error per .c file should be enough to attract attention. Though
   I'll add a comment to convert_bpf_extensions() that build_bug_on
   errors should be addressed in two places.

Cheers,
Daniel