Re: [PATCH] perf session: Add option to copy events when queueing

[David Ahern <dsahern@gmail.com>]

On 9/16/13 10:40 AM, Frederic Weisbecker wrote:
>> Yes. I could make it the default behavior; just overhead in doing
>> that (malloc/copy for each event).
>
> Are there any tool that don't suffer from this bug somehow? If not then it must
> be applied unconditionally.

I believe only 'live' commands would be affected: top, kvm stat live, my 
local scheduling daemon. perf-trace bypasses session and time ordering 
(though I question that decision).

---8<---

>> size of event is determined by mmap_event (mmap2_event in latest
>> code) which is > 4096 because of the filename argument. Including
>> the event directly in sample_queue would balloon memory usage
>> (learned this the hard way!).
>
> Ah then perhaps we can allocate with the dynamic size of the event?

Yes, that's how I have it my patch -- allocate memory based on header 
size and copy in the queue_event function. I think we agree on this part 
now.

---8<---

>>> Although the mirrored os->sample_buffer condition check is a bit ugly and should move to
>>> a function. But the idea is there.
>>
>> Ok. That should be a separate patch. Are you going to submit that one?
>
> Yeah, unless you beat me at it :)

That's not going to happen on my end due to a recent time constraint 
(jury duty - minimum 1 week trial). At best I can do a patch in early 
October.

David