On Fri, Mar 11, 2022 at 05:05:24AM IST, Alexei Starovoitov wrote:
> On Thu, Mar 10, 2022 at 3:30 PM Toke Høiland-Jørgensen <toke@xxxxxxxxxx> wrote:
> >
> > Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx> writes:
> >
> > > On Tue, Mar 8, 2022 at 5:40 AM Toke Høiland-Jørgensen <toke@xxxxxxxxxx> wrote:
> > >>
> > >> Kumar Kartikeya Dwivedi <memxor@xxxxxxxxx> writes:
> > >>
> > >> > On Tue, Mar 08, 2022 at 11:18:52AM IST, Andrii Nakryiko wrote:
> > >> >> On Sun, Mar 6, 2022 at 3:43 PM Kumar Kartikeya Dwivedi <memxor@xxxxxxxxx> wrote:
> > >> >> >
> > >> >> > Expose existing 'bpf_xdp_pointer' as a BPF helper named 'bpf_packet_pointer'
> > >> >> > returning a packet pointer with a fixed immutable range. This can be useful to
> > >> >> > enable DPA without having to use memcpy (currently the case in
> > >> >> > bpf_xdp_load_bytes and bpf_xdp_store_bytes).
> > >> >> >
> > >> >> > The intended usage to read and write data for multi-buff XDP is:
> > >> >> >
> > >> >> > int err = 0;
> > >> >> > char buf[N];
> > >> >> >
> > >> >> > off &= 0xffff;
> > >> >> > ptr = bpf_packet_pointer(ctx, off, sizeof(buf), &err);
> > >> >> > if (unlikely(!ptr)) {
> > >> >> > if (err < 0)
> > >> >> > return XDP_ABORTED;
> > >> >> > err = bpf_xdp_load_bytes(ctx, off, buf, sizeof(buf));
> > >> >> > if (err < 0)
> > >> >> > return XDP_ABORTED;
> > >> >> > ptr = buf;
> > >> >> > }
> > >> >> > ...
> > >> >> > // Do some stores and loads in [ptr, ptr + N) region
> > >> >> > ...
> > >> >> > if (unlikely(ptr == buf)) {
> > >> >> > err = bpf_xdp_store_bytes(ctx, off, buf, sizeof(buf));
> > >> >> > if (err < 0)
> > >> >> > return XDP_ABORTED;
> > >> >> > }
> > >> >> >
> > >> >> > Note that bpf_packet_pointer returns a PTR_TO_PACKET, not PTR_TO_MEM, because
> > >> >> > these pointers need to be invalidated on clear_all_pkt_pointers invocation, and
> > >> >> > it is also more meaningful to the user to see return value as R0=pkt.
> > >> >> >
> > >> >> > This series is meant to collect feedback on the approach, next version can
> > >> >> > include a bpf_skb_pointer and exposing it as bpf_packet_pointer helper for TC
> > >> >> > hooks, and explore not resetting range to zero on r0 += rX, instead check access
> > >> >> > like check_mem_region_access (var_off + off < range), since there would be no
> > >> >> > data_end to compare against and obtain a new range.
> > >> >> >
> > >> >> > The common name and func_id is supposed to allow writing generic code using
> > >> >> > bpf_packet_pointer that works for both XDP and TC programs.
> > >> >> >
> > >> >> > Please see the individual patches for implementation details.
> > >> >> >
> > >> >>
> > >> >> Joanne is working on a "bpf_dynptr" framework that will support
> > >> >> exactly this feature, in addition to working with dynamically
> > >> >> allocated memory, working with memory of statically unknown size (but
> > >> >> safe and checked at runtime), etc. And all that within a generic
> > >> >> common feature implemented uniformly within the verifier. E.g., it
> > >> >> won't need any of the custom bits of logic added in patch #2 and #3.
> > >> >> So I'm thinking that instead of custom-implementing a partial case of
> > >> >> bpf_dynptr just for skb and xdp packets, let's maybe wait for dynptr
> > >> >> and do it only once there?
> > >> >>
> > >> >
> > >> > Interesting stuff, looking forward to it.
> > >> >
> > >> >> See also my ARG_CONSTANT comment. It seems like a pretty common thing
> > >> >> where input constant is used to characterize some pointer returned
> > >> >> from the helper (e.g., bpf_ringbuf_reserve() case), and we'll need
> > >> >> that for bpf_dynptr for exactly this "give me direct access of N
> > >> >> bytes, if possible" case. So improving/generalizing it now before
> > >> >> dynptr lands makes a lot of sense, outside of bpf_packet_pointer()
> > >> >> feature itself.
> > >> >
> > >> > No worries, we can continue the discussion in patch 1, I'll split out the arg
> > >> > changes into a separate patch, and wait for dynptr to be posted before reworking
> > >> > this.
> > >>
> > >> This does raise the question of what we do in the meantime, though? Your
> > >> patch includes a change to bpf_xdp_{load,store}_bytes() which, if we're
> > >> making it, really has to go in before those hit a release and become
> > >> UAPI.
> > >>
> > >> One option would be to still make the change to those other helpers;
> > >> they'd become a bit slower, but if we have a solution for that coming,
> > >> that may be OK for a single release? WDYT?
> > >
> > > I must have missed important changes to bpf_xdp_{load,store}_bytes().
> > > Does anything change about its behavior? If there are some fixes
> > > specific to those helpers, we should fix them as well as a separate
> > > patch. My main objection is adding a bpf_packet_pointer() special case
> > > when we have a generic mechanism in the works that will come this use
> > > case (among other use cases).
> >
> > Well it's not a functional change per se, but Kartikeya's patch is
> > removing an optimisation from bpf_xdp_{load_store}_bytes() (i.e., the
> > use of the bpf_xdp_pointer()) in favour of making it available directly
> > to BPF. So if we don't do that change before those helpers are
> > finalised, we will end up either introducing a performance regression
> > for code using those helpers, or being stuck with the bpf_xdp_pointer()
> > use inside them even though it makes more sense to move it out to BPF.
> >
> > So the "safe" thing to do would do the change to the store/load helpers
> > now, and get rid of the bpf_xdp_pointer() entirely until it can be
> > introduced as a BPF helper in a generic way. Of course this depends on
> > whether you consider performance regressions to be something to avoid,
> > but this being XDP IMO we should :)
>
> I don't follow this logic.
> Would you mean by "get rid of the bpf_xdp_pointer" ?
> It's just an internal static function.
>
> Also I don't believe that this patch set and exposing
> bpf_xdp_pointer as a helper actually gives measurable performance wins.
> It looks quirky to me and hard to use.
This is actually inspired from your idea to avoid memcpy when reading and
writing to multi-buff XDP [0]. But instead of passing in the stack or mem
pointer (as discussed in that thread), I let the user set it and detect it
themselves, which makes the implementation simpler.
I am sure accessing a few bytes directly is going to be faster than first
memcpy'ing it to a local buffer, reading, and then possibly writing things
back again using a memcpy, but I will be happy to come with some numbers when
I respin this later, when Joanne posts the dynptr series.
Ofcourse, we could just make return value PTR_TO_MEM even for the 'pass buf
pointer' idea, but then we have to conservatively invalidate the pointer even if
it points to stack buffer on clear_all_pkt_pointers. The current approach looked
better to me.
[0]: https://lore.kernel.org/bpf/CAADnVQKbrkOxfNoixUx-RLJEWULJLyhqjZ=M_X2cFG_APwNyCg@xxxxxxxxxxxxxx
--
Kartikeya
