Re: [PATCH] tracing/perf: Move rcu_irq_enter/exit_irqson() to perf trace point hook

From: Peter Zijlstra
Date: Tue Feb 11 2020 - 07:00:38 EST


On Mon, Feb 10, 2020 at 07:30:32PM -0500, Mathieu Desnoyers wrote:

> > because perf only uses rcu to synchronize trace points.
>
> That last part seems inaccurate. The tracepoint synchronization is two-fold:
> one part is internal to tracepoint.c (see rcu_free_old_probes()), and the other
> is only needed if the probes are within modules which can be unloaded (see
> tracepoint_synchronize_unregister()). AFAIK, perf never implements probe callbacks
> within modules, so the latter is not needed by perf.
>
> The culprit of the problem here is that perf issues "rcu_read_lock()" and
> "rcu_read_unlock()" within the probe callbacks it registers to the tracepoints,
> including the rcuidle ones. Those require that RCU is "watching", which is
> triggering the regression when we remove the calls to rcu_irq_enter/exit_irqson()
> from the rcuidle tracepoint instrumentation sites.

It is not the fact that perf issues rcu_read_lock() that is the problem.
As we established yesterday, I can probably remove most rcu_read_lock()
calls from perf today (yay RCU flavour unification).

The problem is that the core perf code uses RCU managed data; and we
need an existence guarantee for it. It would be BAD (TM) if the
ring-buffer we're writing data to were to suddenly dissapear under our
feet etc..

> Which brings a question about handling of NMIs: in the proposed patch, if
> a NMI nests over rcuidle context, AFAIU it will be in a state
> !rcu_is_watching() && in_nmi(), which is handled by this patch with a simple
> "return", meaning important NMIs doing hardware event sampling can be
> completely lost.
>
> Considering that we cannot use rcu_irq_enter/exit_irqson() from NMI context,
> is it at all valid to use rcu_read_lock/unlock() as perf does from NMI handlers,

Again, rcu_read_lock() itself really isn't the problem. But we need
NMIs, just like regular interrupts, to imply rcu_read_lock(). That is,
any observable (RCU managed) pointer must stay valid during the NMI/IRQ
execution.

> considering that those can be nested on top of rcuidle context ?

As per nmi_enter() calling rcu_nmi_enter() I've always assumed that NMIs
are fully covered by RCU.

If this isn't so, RCU it terminally broken :-)