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

[John Fastabend <john.fastabend@gmail.com>]

Toke Høiland-Jørgensen wrote:
> Alexei Starovoitov <alexei.starovoitov@gmail.com> writes:
> 
> > On Thu, Oct 31, 2019 at 12:18 PM Jiri Olsa <jolsa@redhat.com> wrote:
> >> >
> >> > yes. older vmlinux and newer installed libbpf.so
> >> > or any version of libbpf.a that is statically linked into apps
> >> > is something that libbpf code has to support.
> >> > The server can be rebooted into older than libbpf kernel and
> >> > into newer than libbpf kernel. libbpf has to recognize all these
> >> > combinations and work appropriately.
> >> > That's what backward and forward compatibility is.
> >> > That's what makes libbpf so difficult to test, develop and code review.
> >> > What that particular server has in /usr/include is irrelevant.
> >>
> >> sure, anyway we can't compile following:
> >>
> >>         tredaell@aldebaran ~ $ 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;
> >>         ...
> >>
> >>         XDP_RING_NEED_WAKEUP is defined in kernel v5.4-rc1 (77cd0d7b3f257fd0e3096b4fdcff1a7d38e99e10).
> >>         XSK_UNALIGNED_BUF_ADDR_MASK and XSK_UNALIGNED_BUF_OFFSET_SHIFT are defined in kernel v5.4-rc1 (c05cd3645814724bdeb32a2b4d953b12bdea5f8c).
> >>
> >> with:
> >>   kernel-headers-5.3.6-300.fc31.x86_64
> >>   libbpf-0.0.5-1.fc31.x86_64
> >>
> >> if you're saying this is not supported, I guess we could be postponing
> >> libbpf rpm releases until we have the related fedora kernel released
> >
> > why? github/libbpf is the source of truth for building packages
> > and afaik it builds fine.
> >
> >> or how about inluding uapi headers in libbpf-devel.. but that might
> >> actualy cause more confusion
> >
> > Libraries (libbpf or any other) should not install headers that
> > typically go into /usr/include/
> > if_xdp.h case is not unique.
> > We'll surely add another #define, enum, etc to uapi/linux/bpf.h tomorrow.
> > And we will not copy paste these constants and types into tools/lib/bpf/.
> > In kernel tree libbpf development is using kernel tree headers.
> > No problem there for libbpf developers.
> > Packages are built out of github/libbpf that has a copy of uapi headers
> > necessary to create packages.
> > No problem there for package builders either.
> > But libbpf package is not going to install those uapi headers.
> > libbpf package installs only libbpf own headers (like libbpf.h)
> > The users that want to build against the latest libbpf package need
> > to install corresponding uapi headers package.
> > I don't think such dependency is specified in rpm scripts.
> > May be it is something to fix? Or may be not.
> > Some folks might not want to update all of /usr/include to bring libbpf-devel.
> > Then it would be their responsibility to get fresh /usr/include headers.
> 
> We can certainly tie libbpf to the kernel version. The obvious way to do
> that is to just ship the version of libbpf that's in the kernel tree of
> whatever kernel version the distro ships. But how will we handle
> bugfixes, then? You've explicitly stated that libbpf gets no bugfixes
> outside of bpf-next...
> 
> -Toke

We use libbpf and build for a wide variety of kernels so I don't think we
want to make libbpf kernel version specific. I always want the latest libbpf
features even when building on older kernels. I generally use the bpf-next
version though so maybe I'm not the target user.

.John