* Ingo Molnar (mingo@xxxxxxx) wrote: > > * Ingo Molnar <mingo@xxxxxxx> wrote: > > > Note that you do not want the context switch event, but the CPU > > migration event: that will notify user-space when it gets migrated > > to another CPU. This is the case that RCU really needs. > > Also note that the main current use-case of perf events is > instrumentation, thus if you make use of this facility for user-space > RCU you need to check whether the events are all precise and arrive > in time to the target process, etc. > > Statistical behavior isnt a big problem for instrumentation but it's > a showstopper for RCU, obviously! :-) > > If you find such bugs then we want to fix them, so there's no > fundamental *desire* from us for these events to be statistical and > inaccurate anywhere. The accuracy vs speed tradeoff is actually quite different from the instrumentation vs low-level-synchronization point of views. It might be acceptable in some sampling situations to get some inaccuracy due to lack of locking if it makes the data collection tool reasonably fast for real-life use (note that I am talking about "sampling", not event-based tracing here). So there are situations where adding locking will make the overhead prohibitive for sampling, but would be required for the perfect accuracy needed by RCU. So although the desire might not be to get inaccurate data, the actual desire to get it in a low-overhead fashion can lead to different decisions regarding the accuracy vs speed tradeoff. Thanks, Mathieu -- Mathieu Desnoyers Operating System Efficiency R&D Consultant EfficiOS Inc. http://www.efficios.com -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
