Re: [PATCH v5 tip 1/7] bpf: make internal bpf API independent of CONFIG_BPF_SYSCALL ifdefs

On Mon, Mar 2, 2015 at 4:26 AM, Daniel Borkmann [off-list ref] wrote:

On 03/02/2015 12:51 PM, Masami Hiramatsu wrote:

quoted

(2015/03/02 20:10), Daniel Borkmann wrote:

quoted

Well, currently all possible map types (hash table, array map) that
would actually call into bpf_register_map_type() are only built when
CONFIG_BPF_SYSCALL is enabled (see kernel/bpf/Makefile). I don't think
new map additions should be added that are not under kernel/bpf/ and/or
enabled outside the CONFIG_BPF_SYSCALL, as it should be considered
part of the eBPF core code.

agree. New map types will be only under kernel/bpf/
since this is really core infra that every component should be
able to share.

quoted

quoted

The difference here (this patch) is simply that we don't want users to
build ifdef spaghetti constructs in user code, so the API that is
actually used by eBPF _users_ is being properly ifdef'ed in the headers.

+1

quoted

Or, maybe we'd better move them into new include/linux/bpf_prog.h which
includes basic include/linux/bpf.h. Then, user can include the bpf_prog.h
instead of bpf.h. Also, we can check CONFIG_BPF_SYSCAL=y at the top of
bpf_prog.h. This makes things clearer :)

I'm preferring the 1st variant, though. We have currently two native eBPF
users, that is, socket filters and tc's cls_bpf (queued in net-next) and
looking at the code/API usage, it's really not that hard, where it would
justify to move this to an extra header file, really.

agree. new header seems overkill to fix something
that is not an issue today.