Re: [PATCH] perf: add /proc/perf_events file for dump perf events info

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

On Thu, 2012-07-05 at 21:38 +0800, Jovi Zhang wrote:
> On Thu, Jul 5, 2012 at 9:08 PM, Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:
> > On Thu, 2012-07-05 at 21:02 +0800, Jovi Zhang wrote:
> >> > Watch all perf events in system wide can be very useful for perf subsystem issue handling,
> >> > to know which perf event is active in system,
> >> > perf event is a resouce, it would like to be managed easily for user, with more visable, like /proc/timer, etc...
> >> >
> >> > .jovi
> >>
> >>
> >> Ping...
> >
> > Never saw your reply, due to HTML it landed in the spam folder.
> >
> > That's a pretty non-specific answer.. again, why are you interested,
> > what specific problem are you wanting to solve.
> >
> > I've never had this need myself.
> >
> sorry, I will try to explain more.
> 
> One problem I faced is about hw_breakpoint.
> As you known, hw_breakpoint use limit debug register in most architecture,
> In multi-user environment, sometime user cannot set hw_breakpoint because
> other user already occupy hw_breakpoint slots. currently, there don't
> have a way to know
> how many hw_breakpint perf event already is used in system, so that's
> why I thinking
> we might need a way to get perf event in systerm wide, with visable output.

Hrmm,. so this seems pretty specific to the horror of hw_breakpoint. And
yes those are unfortunate and weird.

But how would you use this proc file? Would you go read it
programmatically or just look at it as a user to figure out why stuff
doesn't work?

> Also this method is not only used for hw_breakpoint, others perf event
> might have similar problem,
> even other perf event don't have limit number, but it can make use of
> this /proc/perf_events

They have a limit alright, but we can round-robin them to hide this fact
(unless you tell it not to).

> Active perf events is cpu consumer at most time, at this point of
> view, system administrator also can use this
> /proc/perf_events to detect is there have any perf events is consuming cpu.

I doubt you can see which is consuming cycles, but you can see if
there's any in use.

> A method to detect perf event leak? of couse our perf subsystem is
> very stable right now, ingnore this :)

There's alway bugs ;-)

The problem I have with the patch is the global nature of it.. but if
something like this is require I guess I can live with it. But it might
be the current proposal is exposing too much information, I would
certainly not mark it readable for the entire world either.