Date:   Thu, 26 Dec 2019 20:25:15 +0000
From:   Martin Lau <kafai@...com>
To:     Andrii Nakryiko <andrii.nakryiko@...il.com>
CC:     bpf <bpf@...r.kernel.org>, Alexei Starovoitov <ast@...nel.org>,
        "Daniel Borkmann" <daniel@...earbox.net>,
        David Miller <davem@...emloft.net>,
        "Kernel Team" <Kernel-team@...com>,
        Networking <netdev@...r.kernel.org>
Subject: Re: [PATCH bpf-next v2 11/11] bpf: Add bpf_dctcp example

On Thu, Dec 26, 2019 at 11:02:26AM -0800, Andrii Nakryiko wrote:
> On Tue, Dec 24, 2019 at 8:50 AM Martin Lau <kafai@...com> wrote:
> >
> > On Mon, Dec 23, 2019 at 11:01:55PM -0800, Andrii Nakryiko wrote:
> > > On Mon, Dec 23, 2019 at 5:31 PM Martin Lau <kafai@...com> wrote:
> > > >
> > > > On Mon, Dec 23, 2019 at 03:26:50PM -0800, Andrii Nakryiko wrote:
> > > > > On Fri, Dec 20, 2019 at 10:26 PM Martin KaFai Lau <kafai@...com> wrote:
> > > > > >
> > > > > > This patch adds a bpf_dctcp example.  It currently does not do
> > > > > > no-ECN fallback but the same could be done through the cgrp2-bpf.
> > > > > >
> > > > > > Signed-off-by: Martin KaFai Lau <kafai@...com>
> > > > > > ---
> > > > > >  tools/testing/selftests/bpf/bpf_tcp_helpers.h | 228 ++++++++++++++++++
> > > > > >  .../selftests/bpf/prog_tests/bpf_tcp_ca.c     | 218 +++++++++++++++++
> > > > > >  tools/testing/selftests/bpf/progs/bpf_dctcp.c | 210 ++++++++++++++++
> > > > > >  3 files changed, 656 insertions(+)
> > > > > >  create mode 100644 tools/testing/selftests/bpf/bpf_tcp_helpers.h
> > > > > >  create mode 100644 tools/testing/selftests/bpf/prog_tests/bpf_tcp_ca.c
> > > > > >  create mode 100644 tools/testing/selftests/bpf/progs/bpf_dctcp.c
> > > > > >
> > > > > > diff --git a/tools/testing/selftests/bpf/bpf_tcp_helpers.h b/tools/testing/selftests/bpf/bpf_tcp_helpers.h
> > > > > > new file mode 100644
> > > > > > index 000000000000..7ba8c1b4157a
> > > > > > --- /dev/null
> > > > > > +++ b/tools/testing/selftests/bpf/bpf_tcp_helpers.h
> > > > > > @@ -0,0 +1,228 @@
> > > > > > +/* SPDX-License-Identifier: GPL-2.0 */
> > > > > > +#ifndef __BPF_TCP_HELPERS_H
> > > > > > +#define __BPF_TCP_HELPERS_H
> > > > > > +
> > > > > > +#include <stdbool.h>
> > > > > > +#include <linux/types.h>
> > > > > > +#include <bpf_helpers.h>
> > > > > > +#include <bpf_core_read.h>
> > > > > > +#include "bpf_trace_helpers.h"
> > > > > > +
> > > > > > +#define BPF_TCP_OPS_0(fname, ret_type, ...) BPF_TRACE_x(0, #fname"_sec", fname, ret_type, __VA_ARGS__)
> > > > > > +#define BPF_TCP_OPS_1(fname, ret_type, ...) BPF_TRACE_x(1, #fname"_sec", fname, ret_type, __VA_ARGS__)
> > > > > > +#define BPF_TCP_OPS_2(fname, ret_type, ...) BPF_TRACE_x(2, #fname"_sec", fname, ret_type, __VA_ARGS__)
> > > > > > +#define BPF_TCP_OPS_3(fname, ret_type, ...) BPF_TRACE_x(3, #fname"_sec", fname, ret_type, __VA_ARGS__)
> > > > > > +#define BPF_TCP_OPS_4(fname, ret_type, ...) BPF_TRACE_x(4, #fname"_sec", fname, ret_type, __VA_ARGS__)
> > > > > > +#define BPF_TCP_OPS_5(fname, ret_type, ...) BPF_TRACE_x(5, #fname"_sec", fname, ret_type, __VA_ARGS__)
> > > > >
> > > > > Should we try to put those BPF programs into some section that would
> > > > > indicate they are used with struct opts? libbpf doesn't use or enforce
> > > > > that (even though it could to derive and enforce that they are
> > > > > STRUCT_OPS programs). So something like
> > > > > SEC("struct_ops/<ideally-operation-name-here>"). I think having this
> > > > > convention is very useful for consistency and to do a quick ELF dump
> > > > > and see what is where. WDYT?
> > > > I did not use it here because I don't want any misperception that it is
> > > > a required convention by libbpf.
> > > >
> > > > Sure, I can prefix it here and comment that it is just a
> > > > convention but not a libbpf's requirement.
> > >
> > > Well, we can actually make it a requirement of sorts. Currently your
> > > code expects that BPF program's type is UNSPEC and then it sets it to
> > > STRUCT_OPS. Alternatively we can say that any BPF program in
> > > SEC("struct_ops/<whatever>") will be automatically assigned
> > > STRUCT_OPTS BPF program type (which is done generically in
> > > bpf_object__open()), and then as .struct_ops section is parsed, all
> > > those programs will be "assembled" by the code you added into a
> > > struct_ops map.
> > Setting BPF_PROG_TYPE_STRUCT_OPS can be done automatically at open
> > phase (during collect_reloc time).  I will make this change.
> >
> 
> Can you please extend exiting logic in __bpf_object__open() to do
> this? See how libbpf_prog_type_by_name() is used for that.
Does it have to call libbpf_prog_type_by_name() if everything
has already been decided by the earlier
bpf_object__collect_struct_ops_map_reloc()?

> 
> > >
> > > It's a requirement "of sorts", because even if user doesn't do that,
> > > stuff will still work, if user manually will call
> > > bpf_program__set_struct_ops(prog). Which actually reminds me that it
> > > would be good to add bpf_program__set_struct_ops() and
> > Although there is BPF_PROG_TYPE_FNS macro,
> > I don't see moving bpf_prog__set_struct_ops(prog) to LIBBPF_API is useful
> > while actually may cause confusion and error.  How could __set_struct_ops()
> > a prog to struct_ops prog_type help a program, which is not used in
> > SEC(".struct_ops"), to be loaded successfully as a struct_ops prog?
> >
> > Assigning a bpf_prog to a function ptr under the SEC(".struct_ops")
> > is the only way for a program to be successfully loaded as
> > struct_ops prog type.  Extra way to allow a prog to be changed to
> > struct_ops prog_type is either useless or redundant.
> 
> Well, first of all, just for consistency with everything else. We have
> such methods for all prog_types, so I'd like to avoid a special
> snowflake one that doesn't.
Yes, for consistency is fine as I mentioned in the earlier reply,
as long as it is understood the usefulness of it.

> Second, while high-level libbpf API provides all the magic to
> construct STRUCT_OPS map based on .struct_ops section types,
> technically, user might decide to do that using low-level map creation
> API, right?
How?

Correct that the map api is reused as is in SEC(".struct_ops").

For prog, AFAICT, it is not possible to create struct_ops
prog from raw and use it in struct_ops map unless more LIBBPF_API
is added.  Lets put aside the need to find the btf_vmlinux
and its btf-types...etc.  At least, there is no LIBBPF_API to
set prog->attach_btf_id.  Considering the amount of preparation
is needed to create a struct_ops map from raw,  I would like
to see a real use case first before even considering what else
is needed and add another LIBBPF_API that may not be used.

Powered by blists - more mailing lists