Re: [PATCH x86/entry: Force rcu_irq_enter() when in idle task

[Paul E. McKenney]

On Fri, Jun 12, 2020 at 09:34:56PM +0200, Thomas Gleixner wrote:
> Thomas Gleixner <tglx@xxxxxxxxxxxxx> writes:
> 
> > "Paul E. McKenney" <paulmck@xxxxxxxxxx> writes:
> >> On Fri, Jun 12, 2020 at 10:49:53AM -0700, Paul E. McKenney wrote:
> >>> I will therefore address this issue in a follow-on patch.
> >>
> >> I should add that -your- patch from yesterday did -not- cause this
> >> problem, in case that is of interest.
> >
> > So I still can add it back and amend the changelog and comment:
> >
> > Change the condition from !rcu_is_watching() to is_idle_task(current) ||
> > rcu_is_watching() which enforces that interrupts in the idle task
> > unconditionally invoke rcu_irq_enter() independent of the RCU state. For
> > most scenarios is_idle_task() would be sufficient but Task RCU needs it
> > according to Paul.
> 
> After talking to Paul some more we came to the conclusion that the
> failure scenario of task rcu is not completely clear and the trigger
> scenario is obscure enough. This needs more investigation and the
> important part which we were chasing is fixed and agreed on. So I go
> with the simple version now and Paul will follow up once it can be
> properly explained.

So the whole TASKS03 failure issue turned out to be me applying the
wrong patches onto the wrong commits.  Retesting with -tip x86/entry
passes TASKS03, as in more than 100 instances of it.

I am rerunning the whole stack, but I don't see the need to wait for
that.  (I will be running increasingly long tests over Friday night,
Pacific Time.)

							Thanx, Paul