On Thu, Jan 6, 2022 at 8:32 AM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Thu, Jan 6, 2022 at 12:41 AM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
> >
> > On Wed, Jan 05, 2022 at 08:30:56PM -0800, Andrii Nakryiko wrote:
> > > On Tue, Jan 4, 2022 at 12:10 AM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
> > > >
> > > > Adding new link type BPF_LINK_TYPE_KPROBE to attach kprobes
> > > > directly through register_kprobe/kretprobe API.
> > > >
> > > > Adding new attach type BPF_TRACE_RAW_KPROBE that enables
> > > > such link for kprobe program.
> > > >
> > > > The new link allows to create multiple kprobes link by using
> > > > new link_create interface:
> > > >
> > > > struct {
> > > > __aligned_u64 addrs;
> > > > __u32 cnt;
> > > > __u64 bpf_cookie;
> > >
> > > I'm afraid bpf_cookie has to be different for each addr, otherwise
> > > it's severely limiting. So it would be an array of cookies alongside
> > > an array of addresses
> >
> > ok
> >
> > >
> > > > } kprobe;
> > > >
> > > > Plus new flag BPF_F_KPROBE_RETURN for link_create.flags to
> > > > create return probe.
> > > >
> > > > Signed-off-by: Jiri Olsa <jolsa@xxxxxxxxxx>
> > > > ---
> > > > include/linux/bpf_types.h | 1 +
> > > > include/uapi/linux/bpf.h | 12 +++
> > > > kernel/bpf/syscall.c | 191 ++++++++++++++++++++++++++++++++-
> > > > tools/include/uapi/linux/bpf.h | 12 +++
> > > > 4 files changed, 211 insertions(+), 5 deletions(-)
> > > >
> > >
> > > [...]
> > >
> > > > @@ -1111,6 +1113,11 @@ enum bpf_link_type {
> > > > */
> > > > #define BPF_F_SLEEPABLE (1U << 4)
> > > >
> > > > +/* link_create flags used in LINK_CREATE command for BPF_TRACE_RAW_KPROBE
> > > > + * attach type.
> > > > + */
> > > > +#define BPF_F_KPROBE_RETURN (1U << 0)
> > > > +
> > >
> > > we have plenty of flexibility to have per-link type fields, so why not
> > > add `bool is_retprobe` next to addrs and cnt?
> >
> > well I thought if I do that, people would suggest to use the empty
> > flags field instead ;-)
> >
> > we can move it there as you suggest, but I wonder it's good idea to
> > use bool in uapi headers, because the bool size definition is vague
>
> Good point. No 'bool' please.
> grep bool include/uapi/linux/
> Only gives openvswitch.h and it's guarded by ifdef KERNEL
> So not a single uapi header has bool in it.
Yeah, I don't insist on bool specifically.
But I was trying to avoid link_create.flags to become map_flags where
we have various flags, each taking a separate bit, but most flags
don't apply to most map types. Ideally link_create.flags would have
few flags that apply to all or most link types (i.e., something
orthogonal to a specific link type), and for this case we could have
kprobe-specific flags (link_create.kprobe.flags) to adjust kprobe link
creation behavior.
But I don't know, maybe I'm overthinking this.
