Re: [PATCH tip 0/3] Improvements of scheduler related Tracepoints

[Alexei Starovoitov <ast@fb.com>]

On 12/14/17 11:39 PM, Peter Zijlstra wrote:
> On Thu, Dec 14, 2017 at 07:16:00PM -0800, Alexei Starovoitov wrote:
>> On 12/14/17 12:49 PM, Peter Zijlstra wrote:
>>> On Thu, Dec 14, 2017 at 12:20:41PM -0800, Teng Qin wrote:
>>>> This set of commits attempts to improve three scheduler related
>>>> Tracepoints: sched_switch, sched_process_fork, sched_process_exit.
>>>>
>>>> Firstly, these commit add additional flag values, namely preempt,
>>>> clone_flags and group_dead to these Tracepoints, to make information
>>>> exposed via the Tracepoints more useful and complete.
>>>>
>>>> Secondly, these commits exposes task_struct pointers in these
>>>> Tracepoints. The task_struct pointers are arguments of the Tracepoints
>>>> and currently only used to compute struct field values. But for BPF
>>>> programs attached to these Tracepoints, we may want to read additional
>>>> task information via the task_struct pointers. This is currently either
>>>> impossible, or we have to make assumption of whether the Tracepoint is
>>>> running from previous / parent or next / child, and use current pointer
>>>> instead. Exposing the task_struct pointers explicitly makes such use
>>>> case easier and more reliable.
>>>>
>>>
>>> NAK
>>
>> not sure what is the concern here.
>> Is it first or second part of the above ?
>
> Definitely the second, but also the first. You know I would have ripped
> out all scheduler tracepoints if I could have. They're a pain in the
> arse.
>
> A lot of people want to add to the tracepoints, with the end result that
> they'll end up a big bloated pile of useless crap. The first part is
> just the pieces you want added.
>
> As to the second, that's complete crap; that just makes everything
> slower for bodies benefit. If you register a traceprobe you already get
> access to these things.
>
> I think your problem is that you use perf to get access to the
> tracepoints, which them means you have to do disgusting things like
> this.

yeah. Currently bpf progs are called at the end of
perf_trace_##call()
{
   .. regular tracepoint copy craft
   perf_trace_run_bpf_submit( &copied args )
}

from bpf pov we'd rather get access to raw args passed into
perf_trace_##call.
Sounds like you're suggesting to let bpf side register its
progs directly via tracepoint_probe_register() ?
That would solve the whole thing really nicely indeed.

How such api would look like ?
Something like extending kprobe/uprobe fd-based perf_event_open?
https://www.spinics.net/lists/netdev/msg470567.html
btw could you please apply that set to tip tree
or you want us to route it via bpf-next -> net-next ?

Thanks