Message-ID: <e4390345-df3b-5ece-3464-83ff8c1992ce@fb.com>
Date:   Mon, 20 Jun 2022 09:06:13 -0700
From:   Yonghong Song <yhs@...com>
To:     Jörn-Thorben Hinz <jthinz@...lbox.tu-berlin.de>,
        Martin KaFai Lau <kafai@...com>
Cc:     bpf@...r.kernel.org, Alexei Starovoitov <ast@...nel.org>,
        Daniel Borkmann <daniel@...earbox.net>,
        Andrii Nakryiko <andrii@...nel.org>, netdev@...r.kernel.org
Subject: Re: [PATCH bpf-next v3 3/5] selftests/bpf: Test a BPF CC writing
 sk_pacing_*



On 6/18/22 9:43 AM, Jörn-Thorben Hinz wrote:
> On Fri, 2022-06-17 at 14:04 -0700, Martin KaFai Lau wrote:
>> On Tue, Jun 14, 2022 at 12:44:50PM +0200, Jörn-Thorben Hinz wrote:
>>> Test whether a TCP CC implemented in BPF is allowed to write
>>> sk_pacing_rate and sk_pacing_status in struct sock. This is needed
>>> when
>>> cong_control() is implemented and used.
>>>
>>> Signed-off-by: Jörn-Thorben Hinz <jthinz@...lbox.tu-berlin.de>
>>> ---
>>>   .../selftests/bpf/prog_tests/bpf_tcp_ca.c     | 21 +++++++
>>>   .../bpf/progs/tcp_ca_write_sk_pacing.c        | 60
>>> +++++++++++++++++++
>>>   2 files changed, 81 insertions(+)
>>>   create mode 100644
>>> tools/testing/selftests/bpf/progs/tcp_ca_write_sk_pacing.c
>>>
>>> diff --git a/tools/testing/selftests/bpf/prog_tests/bpf_tcp_ca.c
>>> b/tools/testing/selftests/bpf/prog_tests/bpf_tcp_ca.c
>>> index e9a9a31b2ffe..a797497e2864 100644
>>> --- a/tools/testing/selftests/bpf/prog_tests/bpf_tcp_ca.c
>>> +++ b/tools/testing/selftests/bpf/prog_tests/bpf_tcp_ca.c
>>> @@ -9,6 +9,7 @@
>>>   #include "bpf_cubic.skel.h"
>>>   #include "bpf_tcp_nogpl.skel.h"
>>>   #include "bpf_dctcp_release.skel.h"
>>> +#include "tcp_ca_write_sk_pacing.skel.h"
>>>   
>>>   #ifndef ENOTSUPP
>>>   #define ENOTSUPP 524
>>> @@ -322,6 +323,24 @@ static void test_rel_setsockopt(void)
>>>          bpf_dctcp_release__destroy(rel_skel);
>>>   }
>>>   
>>> +static void test_write_sk_pacing(void)
>>> +{
>>> +       struct tcp_ca_write_sk_pacing *skel;
>>> +       struct bpf_link *link;
>>> +
>>> +       skel = tcp_ca_write_sk_pacing__open_and_load();
>>> +       if (!ASSERT_OK_PTR(skel, "open_and_load")) {
>> nit. Remove this single line '{'.
>>
>> ./scripts/checkpatch.pl has reported that also:
>> WARNING: braces {} are not necessary for single statement blocks
>> #43: FILE: tools/testing/selftests/bpf/prog_tests/bpf_tcp_ca.c:332:
>> +       if (!ASSERT_OK_PTR(skel, "open_and_load")) {
>> +               return;
>> +       }
> Have to admit I knowingly disregarded that warning as more of a
> recommendation. Out of habit and since I personally don’t see any
> compelling reason to generally use single-line statements after ifs,
> only multiple disadvantages.
> 
> But wrong place to argue here, of course. Will bow to the warning.
> 
>>
>>
>>> +               return;
>>> +       }
>>> +
>>> +       link = bpf_map__attach_struct_ops(skel-
>>>> maps.write_sk_pacing);
>>> +       if (ASSERT_OK_PTR(link, "attach_struct_ops")) {
>> Same here.
>>
>> and no need to check the link before bpf_link__destroy.
>> bpf_link__destroy can handle error link.  Something like:
>>
>>          ASSERT_OK_PTR(link, "attach_struct_ops");
>>          bpf_link__destroy(link);
>>          tcp_ca_write_sk_pacing__destroy(skel);
>>
>> The earlier examples in test_cubic and test_dctcp were
>> written before bpf_link__destroy can handle error link.
> You are right, I followed the other two test_*() functions there. Good
> to know that it behaves similar to (k)free() and others. Will remove
> the ifs around bpf_link__destroy().
> 
>>
>>> +               bpf_link__destroy(link);
>>> +       }
>>> +
>>> +       tcp_ca_write_sk_pacing__destroy(skel);
>>> +}
>>> +
>>>   void test_bpf_tcp_ca(void)
>>>   {
>>>          if (test__start_subtest("dctcp"))
>>> @@ -334,4 +353,6 @@ void test_bpf_tcp_ca(void)
>>>                  test_dctcp_fallback();
>>>          if (test__start_subtest("rel_setsockopt"))
>>>                  test_rel_setsockopt();
>>> +       if (test__start_subtest("write_sk_pacing"))
>>> +               test_write_sk_pacing();
>>>   }
>>> diff --git
>>> a/tools/testing/selftests/bpf/progs/tcp_ca_write_sk_pacing.c
>>> b/tools/testing/selftests/bpf/progs/tcp_ca_write_sk_pacing.c
>>> new file mode 100644
>>> index 000000000000..43447704cf0e
>>> --- /dev/null
>>> +++ b/tools/testing/selftests/bpf/progs/tcp_ca_write_sk_pacing.c
>>> @@ -0,0 +1,60 @@
>>> +// SPDX-License-Identifier: GPL-2.0
>>> +
>>> +#include "vmlinux.h"
>>> +
>>> +#include <bpf/bpf_helpers.h>
>>> +#include <bpf/bpf_tracing.h>
>>> +
>>> +char _license[] SEC("license") = "GPL";
>>> +
>>> +#define USEC_PER_SEC 1000000UL
>>> +
>>> +#define min(a, b) ((a) < (b) ? (a) : (b))
>>> +
>>> +static inline struct tcp_sock *tcp_sk(const struct sock *sk)
>>> +{
>> This helper is already available in bpf_tcp_helpers.h.
>> Is there a reason not to use that one and redefine
>> it in both patch 3 and 4?  The mss_cache and srtt_us can be added
>> to bpf_tcp_helpers.h.  It will need another effort to move
>> all selftest's bpf-cc to vmlinux.h.
> I fully agree it’s not elegant to redefine tcp_sk() twice more.
> 
> It was between either using bpf_tcp_helpers.h and adding and
> maintaining additional struct members there. Or using the (as I
> understood it) more “modern” approach with vmlinux.h and redefining the
> trivial tcp_sk(). I chose the later. Didn’t see a reason not to slowly
> introduce vmlinux.h into the CA tests.
> 
> I had the same dilemma for the algorithm I’m implementing: Reuse
> bpf_tcp_helpers.h from the kernel tree and extend it. Or use vmlinux.h
> and copy only some of the (mostly trivial) helper functions. Also chose
> the later here.
> 
> While doing the above, I also considered extracting the type
> declarations from bpf_tcp_helpers.h into an, e.g.,
> bpf_tcp_types_helper.h, keeping only the functions in
> bpf_tcp_helpers.h. bpf_tcp_helpers.h could have been a base helper for
> any BPF CA implementation then and used with either vmlinux.h or the
> “old-school” includes. Similar to the way bpf_helpers.h is used. But at
> that point, a bpf_tcp_types_helper.h could have probably just been
> dropped for good and in favor of vmlinux.h. So I didn’t continue with
> that.
> 
> Do you insist to use bpf_tcp_helpers.h instead of vmlinux.h? Or could
> the described split into two headers make sense after all?

I prefer to use vmlinux.h. Eventually we would like to use vmlinux.h
for progs which include bpf_tcp_healpers.h. Basically remove the struct
definitions in bpf_tcp_helpers.h and replacing "bpf.h, stddef.h, tcp.h 
..." with vmlinux.h. We may not be there yet, but that is the goal.

> 
> (Will wait for your reply here before sending a v4.)
> 
>>
>>> +       return (struct tcp_sock *)sk;
>>> +}
>>> +
>>> +SEC("struct_ops/write_sk_pacing_init")
>>> +void BPF_PROG(write_sk_pacing_init, struct sock *sk)
>>> +{
>>> +#ifdef ENABLE_ATOMICS_TESTS
>>> +       __sync_bool_compare_and_swap(&sk->sk_pacing_status,
>>> SK_PACING_NONE,
>>> +                                    SK_PACING_NEEDED);
>>> +#else
>>> +       sk->sk_pacing_status = SK_PACING_NEEDED;
>>> +#endif
>>> +}
>>> +
>>> +SEC("struct_ops/write_sk_pacing_cong_control")
>>> +void BPF_PROG(write_sk_pacing_cong_control, struct sock *sk,
>>> +             const struct rate_sample *rs)
>>> +{
>>> +       const struct tcp_sock *tp = tcp_sk(sk);
>>> +       unsigned long rate =
>>> +               ((tp->snd_cwnd * tp->mss_cache * USEC_PER_SEC) <<
>>> 3) /
>>> +               (tp->srtt_us ?: 1U << 3);
>>> +       sk->sk_pacing_rate = min(rate, sk->sk_max_pacing_rate);
>>> +}
>>> +
>>> +SEC("struct_ops/write_sk_pacing_ssthresh")
>>> +__u32 BPF_PROG(write_sk_pacing_ssthresh, struct sock *sk)
>>> +{
>>> +       return tcp_sk(sk)->snd_ssthresh;
>>> +}
>>> +
>>> +SEC("struct_ops/write_sk_pacing_undo_cwnd")
>>> +__u32 BPF_PROG(write_sk_pacing_undo_cwnd, struct sock *sk)
>>> +{
>>> +       return tcp_sk(sk)->snd_cwnd;
>>> +}
>>> +
>>> +SEC(".struct_ops")
>>> +struct tcp_congestion_ops write_sk_pacing = {
>>> +       .init = (void *)write_sk_pacing_init,
>>> +       .cong_control = (void *)write_sk_pacing_cong_control,
>>> +       .ssthresh = (void *)write_sk_pacing_ssthresh,
>>> +       .undo_cwnd = (void *)write_sk_pacing_undo_cwnd,
>>> +       .name = "bpf_w_sk_pacing",
>>> +};
>>> -- 
>>> 2.30.2
>>>
> 
> 

Powered by blists - more mailing lists