From mboxrd@z Thu Jan  1 00:00:00 1970
From: linux@arm.linux.org.uk (Russell King - ARM Linux)
Date: Wed, 6 Oct 2010 23:33:58 +0100
Subject: [PATCH] ARM: Check for is_smp for tlb_ops and cache_ops boardcast
In-Reply-To: <20101006144414.GA2913@atomide.com>
References: <20100817104414.19061.38999.stgit@baageli.muru.com>
	<20100906104413.GD20903@n2100.arm.linux.org.uk>
	<20100906180346.GC20849@atomide.com>
	<20100908030951.GH22507@atomide.com>
	<20100908031214.GI22507@atomide.com>
	<20101005221952.GV3117@atomide.com>
	<20101005223310.GA14065@n2100.arm.linux.org.uk>
	<20101006144414.GA2913@atomide.com>
Message-ID: <20101006223358.GA7581@n2100.arm.linux.org.uk>
To: linux-arm-kernel@lists.infradead.org
List-Id: linux-arm-kernel.lists.infradead.org

On Wed, Oct 06, 2010 at 07:44:14AM -0700, Tony Lindgren wrote:
> * Russell King - ARM Linux <linux@arm.linux.org.uk> [101005 15:24]:
> > On Tue, Oct 05, 2010 at 03:19:52PM -0700, Tony Lindgren wrote:
> > > * Tony Lindgren <tony@atomide.com> [100907 20:04]:
> > > > This should not be needed when running on UP systems.
> > > > 
> > > > Additionally we will also get an undefined instruction on ARM cores
> > > > without the extended CPUID registers with CONFIG_SMP_ON_UP.
> > > > 
> > > > Also, we can now remove the is_smp() test from mmu.c.
> > > 
> > > Just FYI, I've updated this one more time with to use cpus_empty
> > > instead of !smp_on_up() here as well.
> > 
> > What's the rationale?
> 
> With CPU hotplug if the other SMP cores are unplugged for PM or
> other reasons, no need to do the broadcast.

Yes, but why this expensive test when the smp_on_up() is much cheaper?

smp_call_function_many() already takes care of the "no other CPUs"
case, which is used by on_each_cpu_mask() and on_each_cpu(), which
means these functions won't broadcast the operations to other CPUs
when they're offline.

In any case, if you think that we broadcast every operation to all
CPUs, you're mistaken - TLB and cache ops are broadcast to only
those CPUs which the thread is running on, or in the case of non-MM
specific, to all online CPUs.

So the only thing we have to worry about is "is there an ID register
available" which is covered by the is_smp() test.  Checking the CPU
mask is far more expensive and imho ends up needlessly adding to the
complexity.