On Fri, Aug 07, 2020 at 09:23:38PM +0200, Peter Zijlstra wrote:
> Much of the complexity in irqenter_{enter,exit}() is due to #PF being
> the sole exception that can schedule from kernel context.
>
> One additional wrinkle with #PF is that it is non-maskable, it can
> happen _anywhere_. Due to this, and the wonders of tracing, we can get
> the 'normal' NMI nesting vs TRACE_IRQFLAGS:
>
> local_irq_disable()
> raw_local_irq_disable();
> trace_hardirqs_off();
>
> local_irq_enable();
> trace_hardirqs_on();
> <#PF>
> trace_hardirqs_off()
> ...
> if (!regs_irqs_disabled(regs)
> trace_hardirqs_on();
> </#PF>
> // WHOOPS -- lockdep thinks IRQs are disabled again!
> raw_local_irqs_enable();
>
> Rework irqenter_{enter,exit}() to save/restore the software state.
So with #3 v1.1, we can maybe do away with this patch.
So the actual case that triggered the above was:
raw_local_irq_disable();
trace_lock_acquire()
... tracing ...
<#PF/>
Now, as Marco spotted, DEBUG_LOCKDEP would trigger in this case, because
'... tracing ...' includes rcu_dereference(), and that in turn calls
lock_is_held()/check_flags() and goes *boom*, because we did
raw_local_irq_disable().
Now, the new patch, moves the tracepoint out from under the
raw_local_irq_disable() too, and given RCU-lockdep complains in this
situation, I'm thinking we're actually free of such cases... fingers
crossed.
