Messages in this thread | | | Subject | Re: [PATCH] perf_counter: allow and require one-page mmap on counting counters | From | Peter Zijlstra <> | Date | Wed, 01 Apr 2009 10:13:42 +0200 |
| |
On Wed, 2009-04-01 at 13:32 +1100, Paul Mackerras wrote: > Peter Zijlstra writes: > > > On Wed, 2009-03-25 at 20:42 +1100, Paul Mackerras wrote: > > > > > > And here's something else that is semi-related: the PAPI guys want a > > > kind of counter that counts until it overflows, and then sends a > > > signal to the process and disables itself (and the whole group it's > > > in). > > > > I tried doing this this evening, and its remarkably hard. Disabling a > > counter relies on reading the time, and taking ctx->lock and such. > > Things that are impossible to do in NMI context. > > So, if I have a group where the leader is a hardware counter set to > use NMIs, and there is a task_clock software counter in the group, > don't we hit exactly the same issue with reading the time?
Hmm, I think you're right there. Nasty.
> I'd be OK with saying that you can't use stop-and-signal with NMI > counters.
Right, so let them use regular IRQs, yes, that will work. Let me code that.
> There will still be some issues on powerpc because of our > lazy interrupt disabling scheme, so some work might have to get > deferred until we soft-enable interrupts, but we have a way to manage > that.
Hmm, you're saying ppc always uses NMIs, even when !hw_event.nmi?
> On another topic, I noticed that we have a race with perf_counter_read > where we do the IPI but don't check in __read() on the destination cpu > that the task we're after is still running on that cpu. It needs > checking and retry logic like we have in other places in > perf_counter.c.
Yep, you're right again. Should we perhaps generalize that whole code and provide a method vector for the various bits. That way we could collapse all that code replication.
This TODO list keeps growing :-)
BTW, how's progress with the lazy switching?
| |