On Tue, Oct 26, 2021 at 6:18 AM Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx> wrote: > > On Mon, Oct 25, 2021 at 9:03 AM Toke Høiland-Jørgensen <toke@xxxxxxxxxx> wrote: > > > > Magnus Karlsson <magnus.karlsson@xxxxxxxxx> writes: > > > > > On Fri, Oct 22, 2021 at 7:49 PM Andrii Nakryiko > > > <andrii.nakryiko@xxxxxxxxx> wrote: > > >> > > >> Hey guys, > > >> > > >> It's been a while since we chatted about libxsk move. I believe last > > >> time we were already almost ready to recommend libxdp for this, but > > >> I'd like to double-check. Can one of you please own [0], validate that > > >> whatever APIs are provided by libxdp are equivalent to what libbpf > > >> provides, and start marking xdk.h APIs as deprecated? Thanks! > > > > > > Resending since Gmail had jumped out of plain text mode again. > > > > > > No problem, I will own this. I will verify the APIs are the same then > > > submit a patch marking the ones in libbpf's xsk.h as deprecated. > > > > > > One question is what to do with the samples and the selftests for xsk. > > > They currently rely on libbpf's xsk support. Two options that I see: > > > > > > 1: Require libxdp on the system. Do not try to compile the xsk samples > > > and selftests if libxdp is not available so the rest of the bpf > > > samples and selftests are not impacted. > > > 2: Provide a standalone mock-up file of xsk.c and xsk.h that samples > > > and selftests could use. > > > > > > I prefer #1 as it is better for the long-term. #2 means I would have > > > to maintain that mock-up file as libxdp features are added. Sounds > > > like double the amount of work to me. Thoughts? > > > > I agree #1 is preferable of those two. Another option is to move the > > samples to the xdp-tools repo instead? Doesn't work for selftests, of > > course; if it's acceptable to conditionally-compile the XSK tests > > depending on system library availability that would be fine by me... > > Seems like the only thing that uses xsk.h is xdpxceiver.c which is > tested through test_xsk.sh. It's not part of test_progs and so isn't > run regularly by BPF CI or maintainers. It makes sense to me to move > such test closer to the library it's supposed to be testing (i.e., > libxdp)? xdpxceiver.c tests kernel functionality, not libxdp functionality, though it does use libxdp (and libbpf) to make the implementation simpler. So it should remain here and use strategy #1. libxdp tests are on another level and should definitely go into the libxdp repo. The xsk samples in samples/bpf/, we could just stop developing/retire (or even remove) in the Linux repo and move them to the xdp-tools repo. They just show how to use the xsk.h api:s and it makes more sense to have them together with libxdp. > > > > I pinged the Debian maintainer of libbpf to see if I can get him to pick > > up libxdp as well, or sponsor me to maintain it. Should make the > > transition smoother; guess I also need to get hold of the OpenSuse > > people. > > > > -Toke > >