On Fri, 7 Apr 2017 08:11:46 -0700 "Paul E. McKenney" <[email protected]> wrote:
> On Fri, Apr 07, 2017 at 10:58:26AM -0400, Steven Rostedt wrote: > > On Fri, 7 Apr 2017 07:43:35 -0700 > > "Paul E. McKenney" <[email protected]> wrote: > > > > > On Fri, Apr 07, 2017 at 10:01:06AM -0400, Steven Rostedt wrote: > > > > Paul, > > > > > > > > Here's my latest. You OK with it? > > > > > > Given your update to 3/5, I suspect that we could live with it. I am > > > expecting some complaints about increases in idle-entry latency, but might > > > be best to wait for complaints rather than complexifying too proactively. > > > > > > > We only added a this_cpu_inc() and this_cpu_dec() which are very fast > > operations. I highly doubt it will be measurable. Although, I'm talking > > about x86, IIRC, the this_cpu_inc/dec were be poorly written for other > > archs in the past. I'm not sure if that was fixed though. > > That is an issue for CPUs that don't have a to-memory increment > instruction. How about __this_cpu_inc() and __this_cpu_dec(), given > that preemption is disabled? Ah, so the issue still exists. I remember complaining to Christoph Lamater about this. Yeah, switching to __this_cpu_inc() works as well. > > > > That said, there isn't supposed to be any tracing during the now very > > > small interval where RCU's idle-entry is incomplete. Mightn't it be > > > better to (under CONFIG_PROVE_RCU or some such) give splats if tracing > > > showed up in that interval? > > > > Again, tracing is not the issue. I do function tracing in that location > > without any problems. The issue here was the stack tracer. > > > > Maybe we can create a new variable that is more cache local to the RCU > > code. > > > > What about calling it "rcu_disabled"? Then tracing that depends on RCU > > can simply check that. > > > > s/stack_trace_disable/disable_rcu/ > > s/stack_trace_enable/enable_rcu/ > > > > export a per cpu variable rcu_disabled > > > > Then I can have the stack tracer check that variable as well. And we > > could even put in a WARN_ON(this_cpu_read(rcu_disabled) in the > > TRACE_EVENT() macros. > > > > Thoughts? > > At this point, if you can use the "__" versions, the change should be > small. With that change, if no one else complains, I am OK. Yeah, I was thinking this_cpu_inc() was the same as __this_cpu_inc(), (where I simply forgot about the existence of __this_cpu_inc() otherwise I would have used it from the beginning. But that said, I wonder if a rcu_disabled would make more sense. I kinda feel we are currently doing this backwards. The real state change is that rcu is currently doing something that will definitely break any rcu_irq_enter() call. Anything doing that to wake up RCU should be informed that that doesn't currently work. Having a way for RCU to tell others that "rcu_irq_enter() wont do anything right now" seems more appropriate to me, as we can then do checks in the trace_event_rcuidle() code to do the same thing. Having RCU just pick stack tracing seems to be backward to me. -- Steve
