Re: [PATCH bpf-next v3] libbpf: fix compatibility for kernels without need_wakeup

["Toke Høiland-Jørgensen" <toke@redhat.com>]

Magnus Karlsson <magnus.karlsson@gmail.com> writes:

> On Thu, Oct 31, 2019 at 9:03 AM Björn Töpel <bjorn.topel@gmail.com> wrote:
>>
>> On Thu, 31 Oct 2019 at 08:17, Magnus Karlsson <magnus.karlsson@gmail.com> wrote:
>> >
>> > On Wed, Oct 30, 2019 at 2:36 PM Toke Høiland-Jørgensen <toke@redhat.com> wrote:
>> > >
>> > > Magnus Karlsson <magnus.karlsson@intel.com> writes:
>> > >
>> > > > When the need_wakeup flag was added to AF_XDP, the format of the
>> > > > XDP_MMAP_OFFSETS getsockopt was extended. Code was added to the
>> > > > kernel to take care of compatibility issues arrising from running
>> > > > applications using any of the two formats. However, libbpf was
>> > > > not extended to take care of the case when the application/libbpf
>> > > > uses the new format but the kernel only supports the old
>> > > > format. This patch adds support in libbpf for parsing the old
>> > > > format, before the need_wakeup flag was added, and emulating a
>> > > > set of static need_wakeup flags that will always work for the
>> > > > application.
>> > >
>> > > Hi Magnus
>> > >
>> > > While you're looking at backwards compatibility issues with xsk: libbpf
>> > > currently fails to compile on a system that has old kernel headers
>> > > installed (this is with kernel-headers 5.3):
>> > >
>> > > $ echo "#include <bpf/xsk.h>" | gcc -x c -
>> > > In file included from <stdin>:1:
>> > > /usr/include/bpf/xsk.h: In function ‘xsk_ring_prod__needs_wakeup’:
>> > > /usr/include/bpf/xsk.h:82:21: error: ‘XDP_RING_NEED_WAKEUP’ undeclared (first use in this function)
>> > >    82 |  return *r->flags & XDP_RING_NEED_WAKEUP;
>> > >       |                     ^~~~~~~~~~~~~~~~~~~~
>> > > /usr/include/bpf/xsk.h:82:21: note: each undeclared identifier is reported only once for each function it appears in
>> > > /usr/include/bpf/xsk.h: In function ‘xsk_umem__extract_addr’:
>> > > /usr/include/bpf/xsk.h:173:16: error: ‘XSK_UNALIGNED_BUF_ADDR_MASK’ undeclared (first use in this function)
>> > >   173 |  return addr & XSK_UNALIGNED_BUF_ADDR_MASK;
>> > >       |                ^~~~~~~~~~~~~~~~~~~~~~~~~~~
>> > > /usr/include/bpf/xsk.h: In function ‘xsk_umem__extract_offset’:
>> > > /usr/include/bpf/xsk.h:178:17: error: ‘XSK_UNALIGNED_BUF_OFFSET_SHIFT’ undeclared (first use in this function)
>> > >   178 |  return addr >> XSK_UNALIGNED_BUF_OFFSET_SHIFT;
>> > >       |                 ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>> > >
>> > >
>> > >
>> > > How would you prefer to handle this? A patch like the one below will fix
>> > > the compile errors, but I'm not sure it makes sense semantically?
>> >
>> > Thanks Toke for finding this. Of course it should be possible to
>> > compile this on an older kernel, but without getting any of the newer
>> > functionality that is not present in that older kernel.
>>
>> Is the plan to support source compatibility for the headers only, or
>> the whole the libbpf itself? Is the usecase here, that you've built
>> libbpf.so with system headers X, and then would like to use the
>> library on a system with older system headers X~10? XDP sockets? BTF?
>
> Good question. I let someone with more insight answer this. Providing
> the support Toke wants does make the header files less pleasant to
> look at for sure. But in any case, I think we should provide an error
> when you try to enable a new kernel feature using an old libbpf that
> has no support for it. Just in case someone mixes things up.

Yup, I agree. Removing the functions completely is fine with me. As for
the flags, I agree that having a check in libbpf would make sense; I can
see someone upgrading their kernel, but still using the distro-specified
libbpf and running into weird errors otherwise.

Maybe we should define XDP_FLAGS_ALL in if_xdp.h and use that for the
check in both libbpf and the kernel? We'd still need conditional defines
for backwards compatibility, but at least we wouldn't need to keep
updating that as new flags are added?

-Toke