Re: [PATCH] x86: add workaround monitor bug

From: Peter Zijlstra <peterz@infradead.org>
To: Ingo Molnar <mingo@kernel.org>
Cc: Jacob Pan <jacob.jun.pan@linux.intel.com>,
	LKML <linux-kernel@vger.kernel.org>,
	Ingo Molnar <mingo@redhat.com>, "H. Peter Anvin" <hpa@zytor.com>,
	X86 Kernel <x86@kernel.org>,
	Arjan van de Ven <arjan@linux.intel.com>,
	Len Brown <lenb@kernel.org>
Subject: Re: [PATCH] x86: add workaround monitor bug
Date: Fri, 8 Jul 2016 13:45:58 +0200	[thread overview]
Message-ID: <20160708114558.GI30909@twins.programming.kicks-ass.net> (raw)
In-Reply-To: <20160708085515.GA4682@gmail.com>

On Fri, Jul 08, 2016 at 10:55:15AM +0200, Ingo Molnar wrote:

> >  static inline void mwait_idle_with_hints(unsigned long eax, unsigned long ecx)
> >  {
> > -	if (!current_set_polling_and_test()) {
> > +	if (static_cpu_has_bug(X86_BUG_MONITOR) || !current_set_polling_and_test()) {
> 
> Hm, this might be suboptimal: if MONITOR/MWAIT is implemented by setting the 
> exclusive flag for the monitored memory address and then snooping for cache 
> invalidation requests for that cache line, then not modifying the ->flags value 
> with TIF_POLLING_NRFLAG makes MWAIT not wake up - only the IPI would wake it up.

Confused.. POLLING_NRFLAGS is not used to wake up ever. It is only used
to determine if we want to send IPIs or not.

And since we _must_ send an IPI in this case, because the monitor is
busted, we cannot set this.

> I think a better approach would be to still optimistically modify the ->flags 
> value _AND_ to also send an IPI, to make sure the wakeup is not lost. This means 
> that the woken CPU will wake up much faster (no IPI latency).

This is exactly what is done. See resched_curr()'s use of
set_nr_and_not_polling(). That does:

	if (!(fetch_or(&flags, NEED_RESCHED) & POLLING_NRFLAG))
		smp_send_reschedule(cpu);

So we unconditionally set NEED_RESCHED, if, when we set that, POLLING
was set, we skip the IPI.

So again, since monitor is busted, simply setting NEED_RESCHED will not
wake us, we must send the IPI, this is achieved by not setting
POLLING_NRFLAG.