Re: [PATCH RFC] hist lookups

[Jiri Olsa]

On Tue, Nov 06, 2018 at 10:13:49PM -0800, David Miller wrote:
> From: Jiri Olsa <jolsa@xxxxxxxxxx>
> Date: Tue, 6 Nov 2018 21:42:55 +0100
> 
> > I pushed that fix in perf/fixes branch, but I'm still occasionaly
> > hitting the namespace crash.. working on it ;-)
> 
> Jiri, how can this new scheme work without setting copy_on_queue
> for the queued_events we use here?

aahh.. it won't, setting it up ;-)

> 
> I don't see copy_on_queue being set and that means the queued event
> structures reference the event memory directly in the mmaps, after the
> mmap thread has released them back to the queue.
> 
> That means new events can come in to the mmap ring and overwrite what
> was there previously, maybe even while deliver_event() is in the
> middle of parsing the event.
> 
> Setting copy_on_queue for data[0] and data[1] makes all of the crashes
> go away for me.
> 
> I get a lot of "[unknown]" shared objects shortly after perf top
> starts up during a full workload.  I've been wondering about one
> side effect of how the mmap queues are processed, consider the
> following:
> 
> 	cpu 0			cpu 1
> 
> 				exec
> 				create new mmap2 events
> 				scheduled to cpu 0 for whatever reason
> 	sample 1
> 	sample 2
> 
> And let's say that perf top is backlogged processing the mmap ring of
> events generated for cpu 0, and sees sample 1 and sample 2 before
> getting to any of cpu 1's events.
> 
> This means the thread and map and symbol objects won't exist and
> we'll get those '[Unknown]' histogram entries, and they won't go
> away.
> 
> When it finally stops looping over the mmap ring for cpu 0's events
> it gets to cpu 1's mmap ring and sees the exec and mmap2 events
> but at that point it's far too late.
> 
> I surmise from what I see with perf top right now that this happens
> a lot.

right, there's no reason why top should have different standards than
record/report.. above can definitely happen, I'll enable time sample
type and use ordered events for the queue

jirka