Re: [RFC perf,bpf 5/5] perf util: generate bpf_prog_info_event for short living bpf programs

From: Alexei Starovoitov <ast@fb.com>
To: David Ahern <dsahern@gmail.com>, David Miller <davem@davemloft.net>
Cc: Song Liu <songliubraving@fb.com>,
	"netdev@vger.kernel.org" <netdev@vger.kernel.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	Kernel Team <Kernel-team@fb.com>,
	"ast@kernel.org" <ast@kernel.org>,
	"daniel@iogearbox.net" <daniel@iogearbox.net>,
	"peterz@infradead.org" <peterz@infradead.org>,
	"acme@kernel.org" <acme@kernel.org>
Subject: Re: [RFC perf,bpf 5/5] perf util: generate bpf_prog_info_event for short living bpf programs
Date: Wed, 7 Nov 2018 00:26:16 +0000	[thread overview]
Message-ID: <82745121-339e-2751-c9db-d1fca02d0b33@fb.com> (raw)
In-Reply-To: <0a05a14c-3a79-e894-ae48-cbe1df4feb91@gmail.com>

On 11/6/18 4:23 PM, David Ahern wrote:
> On 11/6/18 5:13 PM, Alexei Starovoitov wrote:
>> On 11/6/18 3:36 PM, David Miller wrote:
>>> From: Alexei Starovoitov <ast@fb.com>
>>> Date: Tue, 6 Nov 2018 23:29:07 +0000
>>>
>>>> I think concerns with perf overhead from collecting bpf events
>>>> are unfounded.
>>>> I would prefer for this flag to be on by default.
>>>
>>> I will sit in userspace looping over bpf load/unload and see how the
>>> person trying to monitor something else with perf feels about that.
>>>
>>> Really, it is inappropriate to turn this on by default, I completely
>>> agree with David Ahern.
>>>
>>> It's hard enough, _AS IS_, for me to fight back all of the bloat that
>>> is in perf right now and get it back to being able to handle simple
>>> full workloads without dropping events..
>>
>> It's a separate perf thread and separate event with its own epoll.
>> I don't see how it can affect main event collection.
>> Let's put it this way. If it does affect somehow, then yes,
>> it should not be on. If it is not, there is no downside to keep it on.
>> Typical user expects to type 'perf record' and see everything that
>> is happening on the system. Right now short lived bpf programs
>> will not be seen. How user suppose to even know when to use the flag?
> 
> The default is profiling where perf record collects task events and
> periodic samples. So for the default record/report, the bpf load /
> unload events are not relevant.

Exactly the opposite.
It's for default 'perf record' collection of periodic samples.
It can be off for -e collection. That's easy.