On Fri, Jun 07, 2013 at 04:03:05PM -0700, David Daney wrote:

Date:   Fri,  7 Jun 2013 16:03:05 -0700
From: David Daney <redacted>
To: linux-mips@linux-mips.org, ralf@linux-mips.org, kvm@vger.kernel.org,
 Sanjay Lal [off-list ref]
Cc: linux-kernel@vger.kernel.org, David Daney <redacted>
Subject: [PATCH 01/31] MIPS: Move allocate_kscratch to cpu-probe.c and make
 it public.

From: David Daney <redacted>

I'd just like to add a note about compatibility.  Code optimized for
LL/SC-less CPUs has made use of the fact that exception handlers will
clobber k0/k1 to a non-zero value.  On a MIPS II or better CPU a branch
likely instruction could be used to atomically test k0/k1 and depending
on the test, execute a store instruction like:

	.set	noreorder
	beqzl	$k0, ok
	sw	$reg, offset($reg)
	/* if we get here, our SC emulation has failed  */
ok:	...

In particular Sony had elected to do this for the R5900 (after I explained
the concept to somebody and told it'd be a _bad_ idea for compatibility
reasons).  Bad ideas are infectious so I'm sure others have used it, too.

I don't think this should stop your patch nor should we unless this turns
out to be an actual problem add any kludges to support such cowboy style
hacks.  But I wanted to mention and document the issue; maybe this should
be mentioned in the log message of the next version of this patch.

Acked-by: Ralf Baechle <redacted>

  Ralf