Message-ID: <CAEf4BzbFEVjDkxeUu3XcU3QWYSBWWHhYehDyFWF0nnWrZmmtTg@mail.gmail.com>
Date:   Wed, 28 Aug 2019 13:23:51 -0700
From:   Andrii Nakryiko <andrii.nakryiko@...il.com>
To:     Jesper Dangaard Brouer <brouer@...hat.com>
Cc:     Toke Høiland-Jørgensen <toke@...hat.com>,
        Stephen Hemminger <stephen@...workplumber.org>,
        Daniel Borkmann <daniel@...earbox.net>,
        Alexei Starovoitov <ast@...nel.org>,
        Martin KaFai Lau <kafai@...com>,
        Song Liu <songliubraving@...com>, Yonghong Song <yhs@...com>,
        David Miller <davem@...emloft.net>,
        Networking <netdev@...r.kernel.org>, bpf <bpf@...r.kernel.org>,
        Anton Protopopov <aspsk2@...il.com>,
        Stanislav Fomichev <sdf@...ichev.me>,
        Yoel Caspersen <yoel@...knet.dk>
Subject: Re: [RFC bpf-next 0/5] Convert iproute2 to use libbpf (WIP)

On Fri, Aug 23, 2019 at 3:27 AM Jesper Dangaard Brouer
<brouer@...hat.com> wrote:
>
> On Wed, 21 Aug 2019 13:30:09 -0700
> Andrii Nakryiko <andrii.nakryiko@...il.com> wrote:
>
> > On Tue, Aug 20, 2019 at 4:47 AM Toke Høiland-Jørgensen <toke@...hat.com> wrote:
> > >
> > > iproute2 uses its own bpf loader to load eBPF programs, which has
> > > evolved separately from libbpf. Since we are now standardising on
> > > libbpf, this becomes a problem as iproute2 is slowly accumulating
> > > feature incompatibilities with libbpf-based loaders. In particular,
> > > iproute2 has its own (expanded) version of the map definition struct,
> > > which makes it difficult to write programs that can be loaded with both
> > > custom loaders and iproute2.
> > >
> > > This series seeks to address this by converting iproute2 to using libbpf
> > > for all its bpf needs. This version is an early proof-of-concept RFC, to
> > > get some feedback on whether people think this is the right direction.
> > >
> > > What this series does is the following:
> > >
> > > - Updates the libbpf map definition struct to match that of iproute2
> > >   (patch 1).
> >
> >
> > Hi Toke,
> >
> > Thanks for taking a stab at unifying libbpf and iproute2 loaders. I'm
> > totally in support of making iproute2 use libbpf to load/initialize
> > BPF programs. But I'm against adding iproute2-specific fields to
> > libbpf's bpf_map_def definitions to support this.
> >
> > I've proposed the plan of extending libbpf's supported features so
> > that it can be used to load iproute2-style BPF programs earlier,
> > please see discussions in [0] and [1]. I think instead of emulating
> > iproute2 way of matching everything based on user-specified internal
> > IDs, which doesn't provide good user experience and is quite easy to
> > get wrong, we should support same scenarios with better declarative
> > syntax and in a less error-prone way. I believe we can do that by
> > relying on BTF more heavily (again, please check some of my proposals
> > in [0], [1], and discussion with Daniel in those threads). It will
> > feel more natural and be more straightforward to follow. It would be
> > great if you can lend a hand in implementing pieces of that plan!
> >
> > I'm currently on vacation, so my availability is very sparse, but I'd
> > be happy to discuss this further, if need be.
> >
> >   [0] https://lore.kernel.org/bpf/CAEf4BzbfdG2ub7gCi0OYqBrUoChVHWsmOntWAkJt47=FE+km+A@mail.gmail.com/
> >   [1] https://www.spinics.net/lists/bpf/msg03976.html
> >
> > > - Adds functionality to libbpf to support automatic pinning of maps when
> > >   loading an eBPF program, while re-using pinned maps if they already
> > >   exist (patches 2-3).
>
> For production use-cases, libbpf really need an easier higher-level API
> for re-using pinned maps, for establishing shared maps between
> programs.  The existing libbpf API bpf_object__pin_maps() and
> bpf_object__unpin_maps(), which don't re-use pinned maps, are not
> really usable, because they pin/unpin ALL maps in the ELF file.
>
> What users really need is an easy way to specify, on a per map basis,
> what kind of pinning and reuse/sharing they want.  E.g. like iproute2
> have, "global", "object-scope", and "no-pinning". ("ifindex-scope" would
> be nice for XDP).

I totally agree and I think this is easy to add both for BTF-defined
and "classic" bpf_map_def maps. Daniel mentioned in one of the
previous threads that in practice object-scope doesn't seem to be
used, so I'd say we should start with no-pinning + global pinning as
two initial supported values for pinning attribute. ifindex-scope is
interesting, but I'd love to hear a bit more about the use cases.

>   Today users have to split/reimplement bpf_prog_load_xattr(), and
> use/add bpf_map__reuse_fd().  Which is that I ended doing for

Honestly, bpf_prog_load_xattr() existence seems redundant to me. It's
basically just bpf_object__open + bpf_object__load. There is a piece
in the middle with "guessing" program types, but it should just be
moved into bpf_object__open and happen by default. Using open + load
gives more control and isn't really harder than bpf_prog_load_xattr.
bpf_prog_load_xattr which might be slightly more convenient for simple
use case, but falls apart immediately if you need to tune anything
before load.

> xdp-cpumap-tc[2] (used in production at ISP) resulting in 142 lines of
> extra code[3] that should have been hidden inside libbpf.  And worse,
> in this solution[4] the maps for reuse-pinning is specified in the code
> by name.  Thus, they cannot use a generic loader.  That I why, I want
> to mark the maps via a pinning member, like iproute2.
>
> I really hope this moves in a practical direction, as I have the next
> production request lined up (also from an ISP), and I hate to have to
> advice them to choose the same route as [3].

It seems to me that map pinning doesn't need much discussion at this
point, let's start with no-pinning + global pinning. To accommodate
pinning at custom root, bpf_object__open_xattr should accept extra
argument with non-default pinning root path. That should solve your
case completely, shouldn't it? Ultimately, with BTF-defined maps it
should be possible to specify custom pinning path on per-map basis for
cases where user needs ultimate non-uniform manual control.

>
>
> [2] https://github.com/xdp-project/xdp-cpumap-tc/
> [3] https://github.com/xdp-project/xdp-cpumap-tc/blob/master/src/xdp_iphash_to_cpu_user.c#L262-L403
> [4] https://github.com/xdp-project/xdp-cpumap-tc/blob/master/src/xdp_iphash_to_cpu_user.c#L431-L441
> --
> Best regards,
>   Jesper Dangaard Brouer
>   MSc.CS, Principal Kernel Engineer at Red Hat
>   LinkedIn: http://www.linkedin.com/in/brouer

Powered by blists - more mailing lists