Re: [PATCH/RFC 1/2] perfcounters: provide a way to read the current value of interrupting counters

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Tue, 2009-03-17 at 18:47 +1100, Paul Mackerras wrote:
> Peter Zijlstra writes:
> 
> > I'm not sure I understand why. It seems to me you're either interested
> > in sample data, that is {tid,ip,counter} like things, or you want raw
> > count values.
> 
> It was specifically requested by people porting PAPI to PCL, and it
> seems like a reasonable request.

OK, then why didn't the changelog say so :-)

Could you ask them why though, if they need it I won't object too much,
but I'd like to know the use case.

As to the method proposed, I think Ingo and I talked about 'abusing'
non-blocking reads for this purpose, would that work? Then if you need
two fds you could dup() and flip one to non-blocking.

The non-blocking read would either output whatever is already pending,
but in case there is no data, it would generate some on the spot.

> I think we should make record_type a bitset rather than a single
> value, and define bits for recording the callchain as well as various
> other interesting things.

Sounds sensible at first though, I'll see if I can poke a hole in that.

The only two things I've got on my todo list is that challchain and mmap
info, and those can indeed be done as a bitmask, tid information too.

> > Currently the group record type writes things like {hw_event->type,
> > counter} which is ambiguous since we really have a 65bit id space. So I
> > was thinking of making that {fd, counter} to at least have a unique
> > identifier in there.
> 
> The fd is already not a unique identifier - think about dup().

You're quite right indeed - risks of mailing before waking up I
suppose :-)

> The hw_event->type values are not really necessary, in fact, since the
> group members will always be read out in the order that they were
> added to the group.  The only time there might be any possibility of
> confusion is if a new member is added after some samples have already
> been taken (or a member is removed) - then we'll get new records being
> added to the event queue being a different size from those already in
> the queue.  But that's a somewhat different problem which isn't really
> solved by having the type values in there.

Ah, ok, so we can leave that out entirely. I was already planning to add
a sibling_count field, with that we could do something like:

{ record_type, sibling_count, n*{counter} }