On Thu, Jul 18, 2019 at 12:14:22PM -0400, Joel Fernandes wrote: > Trimming the list a bit to keep my noise level low, > > On Sat, Jul 13, 2019 at 1:41 PM Paul E. McKenney <paulmck@xxxxxxxxxxxxx> wrote: > [snip] > > > It still feels like you guys are hyperfocusing on this one particular > > > > knob. I instead need you to look at the interrelating knobs as a group. > > > > > > Thanks for the hints, we'll do that. > > > > > > > On the debugging side, suppose someone gives you an RCU bug report. > > > > What information will you need? How can you best get that information > > > > without excessive numbers of over-and-back interactions with the guy > > > > reporting the bug? As part of this last question, what information is > > > > normally supplied with the bug? Alternatively, what information are > > > > bug reporters normally expected to provide when asked? > > > > > > I suppose I could dig out some of our Android bug reports of the past where > > > there were RCU issues but if there's any fires you are currently fighting do > > > send it our way as debugging homework ;-) > > > > Suppose that you were getting RCU CPU stall > > warnings featuring multi_cpu_stop() called from cpu_stopper_thread(). > > Of course, this really means that some other CPU/task is holding up > > multi_cpu_stop() without also blocking the current grace period. > > > > So I took a shot at this trying to learn how CPU stoppers work in > relation to this problem. > > I am assuming here say CPU X has entered MULTI_STOP_DISABLE_IRQ state > in multi_cpu_stop() but another CPU Y has not yet entered this state. > So CPU X is stalling RCU but it is really because of CPU Y. Now in the > problem statement, you mentioned CPU Y is not holding up the grace > period, which means Y doesn't have any of IRQ, BH or preemption > disabled ; but is still somehow stalling RCU indirectly by troubling > X. > > This can only happen if : > - CPU Y has a thread executing on it that is higher priority than CPU > X's stopper thread which prevents it from getting scheduled. - but the > CPU stopper thread (migration/..) is highest priority RT so this would > be some kind of an odd scheduler bug. I think this bug hardly can happen. > - There is a bug in the CPU stopper machinery itself preventing it > from scheduling the stopper on Y. Even though Y is not holding up the > grace period. Or any thread on Y is busy with preemption/irq disabled preventing the stopper from being scheduled on Y. Or something is stuck in ttwu() to wake up the stopper on Y due to any scheduler locks such as pi_lock or rq->lock or something. I think what you mentioned can happen easily. Basically we would need information about preemption/irq disabled sections on Y and scheduler's current activity on every cpu at that time. Am I missing something? > Did I get that right? Would be exciting to run the rcutorture test > once Paul has it available to reproduce this problem. Thanks, Byungchul
