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

[Alexei Starovoitov]

On 11/6/18 3:17 PM, Song Liu wrote:
> 
> 
>> On Nov 6, 2018, at 1:54 PM, David Ahern <dsahern@xxxxxxxxx> wrote:
>>
>> On 11/6/18 1:52 PM, Song Liu wrote:
>>> +
>>> static int record__mmap_read_all(struct record *rec)
>>> {
>>> 	int err;
>>>
>>> +	err = record__mmap_process_vip_events(rec);
>>> +	if (err)
>>> +		return err;
>>> +
>>> 	err = record__mmap_read_evlist(rec, rec->evlist, false);
>>> 	if (err)
>>> 		return err;
>>
>> Seems to me that is going to increase the overhead of perf on any system
>> doing BPF updates. The BPF events cause a wakeup every load and unload,
>> and perf processes not only the VIP events but then walks all of the
>> other maps.
> 
> BPF prog load/unload events should be rare events in real world use cases.
> So I think the overhead is OK. Also, I don't see an easy way to improve
> this.
> 
>>
>>> @@ -1686,6 +1734,8 @@ static struct option __record_options[] = {
>>> 			  "signal"),
>>> 	OPT_BOOLEAN(0, "dry-run", &dry_run,
>>> 		    "Parse options then exit"),
>>> +	OPT_BOOLEAN(0, "no-bpf-event", &record.no_bpf_event,
>>> +		    "do not record event on bpf program load/unload"),
>>
>> Why should this default on? If am recording FIB events, I don't care
>> about BPF events.
>>
> 
> I am OK with default off if that's the preferred way.

I think concerns with perf overhead from collecting bpf events
are unfounded.
I would prefer for this flag to be on by default.