On Thu, Aug 04, 2016 at 04:28:53PM +0200, Peter Zijlstra wrote: > On Wed, Aug 03, 2016 at 11:57:05AM -0700, Brendan Gregg wrote: > > > As for pmu tracepoints: if I were to instrument it (although I wasn't > > planning to), I'd put a tracepoint in perf_event_overflow() called > > "perf:perf_overflow", with the same arguments. That could then be used > > for all PMU overflow events, without needing to add specific > > tracepoints. > > Could we not teach BPF to replace event->overflow_handler and inject > itself there? > > We don't currently have nice interfaces for doing that, but it should be > possible to do I think. We already have the indirect function call, so > injecting ourself there has 0 overhead.
you're right. All makes sense. I guess I was too lazy to look into how to do it properly. Adding a tracepoint looked like quick and easy way to achieve the same. As far as api goes probably existing IOC_SET_BPF ioctl will do too. Currently overflow_handler is set at event alloc time. If we start changing it on the fly with atomic xchg(), afaik things shouldn't break, since each overflow_handler is run to completion and doesn't change global state, right?
