On Sat, 1 Sep 2018 21:16:39 -0700 "Paul E. McKenney" <[email protected]> wrote:
> On Sat, Sep 01, 2018 at 06:45:31PM -0400, Steven Rostedt wrote: > > On Sat, 1 Sep 2018 10:54:42 -0700 > > "Paul E. McKenney" <[email protected]> wrote: > > > > > On Sat, Sep 01, 2018 at 07:35:59PM +0200, Borislav Petkov wrote: > > > > This is a huge splat! It haz some perf* and sched* in it, I guess for > > > > peterz to stare at. And lemme add Paul for good measure too :) > > > > > > > > Kernel is -rc1 + 3 microcode loader patches ontop which should not be > > > > related. > > > > > > It really is tracing from the idle loop. But I thought that the event > > > tracing took care of that. Adding Steve and Joel for their thoughts. > > > > > > Thanx, Paul > > > > > > > Thx. > > > > > > > > --- > > > > [ 62.409125] ============================= > > > > [ 62.409129] WARNING: suspicious RCU usage > > > > [ 62.409133] 4.19.0-rc1+ #1 Not tainted > > > > [ 62.409136] ----------------------------- > > > > [ 62.409140] ./include/linux/rcupdate.h:631 rcu_read_lock() used > > > > illegally while idle! > > > > [ 62.409143] > > > > other info that might help us debug this: > > > > > > > > [ 62.409147] > > > > RCU used illegally from idle CPU! > > > > rcu_scheduler_active = 2, debug_locks = 1 > > > > [ 62.409151] RCU used illegally from extended quiescent state! > > > > [ 62.409155] 1 lock held by swapper/0/0: > > > > [ 62.409158] #0: 000000004557ee0e (rcu_read_lock){....}, at: > > > > perf_event_output_forward+0x0/0x130 > > > > [ 62.409175] > > > > stack backtrace: > > > > [ 62.409180] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.19.0-rc1+ #1 > > > > [ 62.409183] Hardware name: LENOVO 2320CTO/2320CTO, BIOS G2ET86WW > > > > (2.06 ) 11/13/2012 > > > > [ 62.409187] Call Trace: > > > > [ 62.409196] dump_stack+0x85/0xcb > > > > [ 62.409203] perf_event_output_forward+0xf6/0x130 > > > > I think this is because we switched the trace point code to be > > protected by srcu instead of rcu_lock_sched() and a song and dance to > > "make RCU watch again" if it is not, but perf is using normal > > rcu_read_lock() internally even though it is hooked into the > > tracepoint code. Should perf switch to SRCU, or perhaps it can do the > > song and dance to make RCU watch again? > > Well, this is a regression, so in theory we could push my three SRCU > patches into the current merge window, which would allow perf going > to SRCU, thus fixing the above splat. I am OK either way. What would > you prefer? > > I wonder if this partial revert will fix things? -- Steve diff --git a/include/linux/tracepoint.h b/include/linux/tracepoint.h index 7f2e16e76ac4..041f7e56a289 100644 --- a/include/linux/tracepoint.h +++ b/include/linux/tracepoint.h @@ -158,8 +158,10 @@ extern void syscall_unregfunc(void); * For rcuidle callers, use srcu since sched-rcu \ * doesn't work from the idle path. \ */ \ - if (rcuidle) \ + if (rcuidle) { \ idx = srcu_read_lock_notrace(&tracepoint_srcu); \ + rcu_irq_enter_irqson(); \ + } \ \ it_func_ptr = rcu_dereference_raw((tp)->funcs); \ \ @@ -171,8 +173,10 @@ extern void syscall_unregfunc(void); } while ((++it_func_ptr)->func); \ } \ \ - if (rcuidle) \ + if (rcuidle) { \ + rcu_irq_exit_irqson(); \ srcu_read_unlock_notrace(&tracepoint_srcu, idx);\ + } \ \ preempt_enable_notrace(); \ } while (0)
