Message-ID: <B98810F3-EA8F-432B-A66D-7111B779AC1C@fb.com>
Date:   Wed, 18 Mar 2020 23:45:32 +0000
From:   Song Liu <songliubraving@...com>
To:     Stanislav Fomichev <sdf@...ichev.me>
CC:     Daniel Borkmann <daniel@...earbox.net>,
        "linux-fsdevel@...r.kernel.org" <linux-fsdevel@...r.kernel.org>,
        Networking <netdev@...r.kernel.org>,
        "bpf@...r.kernel.org" <bpf@...r.kernel.org>,
        Kernel Team <Kernel-team@...com>,
        "ast@...nel.org" <ast@...nel.org>,
        "mcgrof@...nel.org" <mcgrof@...nel.org>,
        "keescook@...omium.org" <keescook@...omium.org>,
        "yzaikin@...gle.com" <yzaikin@...gle.com>
Subject: Re: [PATCH v2 bpf-next] bpf: sharing bpf runtime stats with
 /dev/bpf_stats



> On Mar 18, 2020, at 3:29 PM, Stanislav Fomichev <sdf@...ichev.me> wrote:
> 
> On 03/18, Song Liu wrote:
>> 
>> 
>>> On Mar 18, 2020, at 1:58 PM, Daniel Borkmann <daniel@...earbox.net> wrote:
>>> 
>>> On 3/18/20 7:33 AM, Song Liu wrote:
>>>>> On Mar 17, 2020, at 4:08 PM, Song Liu <songliubraving@...com> wrote:
>>>>>> On Mar 17, 2020, at 2:47 PM, Daniel Borkmann <daniel@...earbox.net> wrote:
>>>>>>>> 
>>>>>>>> Hm, true as well. Wouldn't long-term extending "bpftool prog profile" fentry/fexit
>>>>>>>> programs supersede this old bpf_stats infrastructure? Iow, can't we implement the
>>>>>>>> same (or even more elaborate stats aggregation) in BPF via fentry/fexit and then
>>>>>>>> potentially deprecate bpf_stats counters?
>>>>>>> I think run_time_ns has its own value as a simple monitoring framework. We can
>>>>>>> use it in tools like top (and variations). It will be easier for these tools to
>>>>>>> adopt run_time_ns than using fentry/fexit.
>>>>>> 
>>>>>> Agree that this is easier; I presume there is no such official integration today
>>>>>> in tools like top, right, or is there anything planned?
>>>>> 
>>>>> Yes, we do want more supports in different tools to increase the visibility.
>>>>> Here is the effort for atop: https://github.com/Atoptool/atop/pull/88 .
>>>>> 
>>>>> I wasn't pushing push hard on this one mostly because the sysctl interface requires
>>>>> a user space "owner".
>>>>> 
>>>>>>> On the other hand, in long term, we may include a few fentry/fexit based programs
>>>>>>> in the kernel binary (or the rpm), so that these tools can use them easily. At
>>>>>>> that time, we can fully deprecate run_time_ns. Maybe this is not too far away?
>>>>>> 
>>>>>> Did you check how feasible it is to have something like `bpftool prog profile top`
>>>>>> which then enables fentry/fexit for /all/ existing BPF programs in the system? It
>>>>>> could then sort the sample interval by run_cnt, cycles, cache misses, aggregated
>>>>>> runtime, etc in a top-like output. Wdyt?
>>>>> 
>>>>> I wonder whether we can achieve this with one bpf prog (or a trampoline) that covers
>>>>> all BPF programs, like a trampoline inside __BPF_PROG_RUN()?
>>>>> 
>>>>> For long term direction, I think we could compare two different approaches: add new
>>>>> tools (like bpftool prog profile top) vs. add BPF support to existing tools. The
>>>>> first approach is easier. The latter approach would show BPF information to users
>>>>> who are not expecting BPF programs in the systems. For many sysadmins, seeing BPF
>>>>> programs in top/ps, and controlling them via kill is more natural than learning
>>>>> bpftool. What's your thought on this?
>>>> More thoughts on this.
>>>> If we have a special trampoline that attach to all BPF programs at once, we really
>>>> don't need the run_time_ns stats anymore. Eventually, tools that monitor BPF
>>>> programs will depend on libbpf, so using fentry/fexit to monitor BPF programs doesn't
>>>> introduce extra dependency. I guess we also need a way to include BPF program in
>>>> libbpf.
>>>> To summarize this plan, we need:
>>>> 1) A global trampoline that attaches to all BPF programs at once;
>>> 
>>> Overall sounds good, I think the `at once` part might be tricky, at least it would
>>> need to patch one prog after another, each prog also needs to store its own metrics
>>> somewhere for later collection. The start-to-sample could be a shared global var (aka
>>> shared map between all the programs) which would flip the switch though.
>> 
>> I was thinking about adding bpf_global_trampoline and use it in __BPF_PROG_RUN. 
>> Something like:
>> 
>> diff --git i/include/linux/filter.h w/include/linux/filter.h
>> index 9b5aa5c483cc..ac9497d1fa7b 100644
>> --- i/include/linux/filter.h
>> +++ w/include/linux/filter.h
>> @@ -559,9 +559,14 @@ struct sk_filter {
>> 
>> DECLARE_STATIC_KEY_FALSE(bpf_stats_enabled_key);
>> 
>> +extern struct bpf_trampoline *bpf_global_trampoline;
>> +DECLARE_STATIC_KEY_FALSE(bpf_global_tr_active);
>> +
>> #define __BPF_PROG_RUN(prog, ctx, dfunc)       ({                      \
>>        u32 ret;                                                        \
>>        cant_migrate();                                                 \
>> +       if (static_branch_unlikely(&bpf_global_tr_active))              \
>> +               run_the_trampoline();                                   \
>>        if (static_branch_unlikely(&bpf_stats_enabled_key)) {           \
>>                struct bpf_prog_stats *stats;                           \
>>                u64 start = sched_clock();                              \
>> 
>> 
>> I am not 100% sure this is OK. 
>> 
>> I am also not sure whether this is an overkill. Do we really want more complex
>> metric for all BPF programs? Or run_time_ns is enough? 
> I was thinking about exporting a real distribution of the prog runtimes
> instead of doing an average. It would be interesting to see
> 50%/95%/99%/max stats.

Good point. Distribution logic fits well in fentry/fexit programs. 

Let me think more about. this. 

Thanks,
Song

Powered by blists - more mailing lists