On Mon, Mar 2, 2020 at 8:10 PM Andrii Nakryiko
<andrii.nakryiko@xxxxxxxxx> wrote:
>
> On Mon, Mar 2, 2020 at 4:59 PM Alexei Starovoitov
> <alexei.starovoitov@xxxxxxxxx> wrote:
> >
> > On Sun, Mar 01, 2020 at 12:10:42AM -0800, Andrii Nakryiko wrote:
> > > Fix issue with not preserving btf_trace_##call structs when compiled under
> > > Clang. Additionally, capture raw tracepoint arguments in raw_tp_##call
> > > structs, directly usable from BPF programs. Convert runqslower to use those
> > > for proof of concept and to simplify code further.
> >
> > Not only folks compile kernel with clang they use the latest BPF/BTF features
> > with it. This is very nice to see!
> > I've applied 1st patch to make clang compiled kernel emit proper BTF.
> >
> > As far as patch 2 I'm not sure about 'raw_tp_' prefix. tp_btf type of progs can
> > use the same structs. So I think there could be a better name. Also bpftool can
> > generate them as well while emitting vmlinux.h. I think that will avoid adding
> > few kilobytes to vmlinux BTF that kernel isn't going to use atm.
>
> Fair enough, I'll follow up with bpftool changes to generate such
> structs. I'm thinking to use tp_args_xxx name pattern, unless someone
> has a better idea :)
Bad news. BTF_KIND_FUNC_PROTOs don't capture argument names and having
something like:
struct tp_args_sched_switch {
bool arg1;
struct task_struct *arg2;
struct task_struct *arg3;
};
doesn't seem like a good solution...
