On Wed, 27 Mar 2013, Rob Herring wrote:

On 03/27/2013 11:23 AM, Stefano Stabellini wrote:

quoted

Would you agree on a patch that moves virt_smp_ops out of mach-virt and
renames them to psci_smp_ops (maybe to arch/arm/kernel/psci_smp_ops.c)?

Would you agree on initializing psci from setup_arch, right after the
call to arm_dt_init_cpu_maps()?

Finally the most controversial point: would you agree on using
psci_smp_ops by default if they are available?
If not, would you at least agree on letting Xen overwrite the default
machine smp_ops?
We need one or the other for dom0 support.

It should not be *always* use PSCI smp ops if available, but use them
only if the platform does not define its own smp ops.

Well, that is the one additional problem that we have on Xen.

On x86 Xen replaces a lot of core native function calls with its own
implementations (see paravirt_ops).
On ARM we only need *one* set of calls: the smp_ops calls.

So if we don't want to give priority to PSCI over the platform smp_ops,
then we need a simple workaround just for Xen in common code like the
one appended below.
Not pretty, but at least small:

diff --git a/arch/arm/kernel/setup.c b/arch/arm/kernel/setup.c
index 3f6cbb2..08cf7e0 100644
--- a/arch/arm/kernel/setup.c
+++ b/arch/arm/kernel/setup.c

@@ -43,6 +43,8 @@
 #include <asm/cacheflush.h>
 #include <asm/cachetype.h>
 #include <asm/tlbflush.h>
+#include <xen/xen.h>
+#include <asm/xen/hypervisor.h>
 
 #include <asm/prom.h>
 #include <asm/mach/arch.h>

@@ -766,9 +768,13 @@ void __init setup_arch(char **cmdline_p)
 	unflatten_device_tree();
 
 	arm_dt_init_cpu_maps();
+	xen_early_init();
 #ifdef CONFIG_SMP
 	if (is_smp()) {
-		smp_set_ops(mdesc->smp);
+		if (xen_domain())
+			smp_set_ops(&xen_smp_ops);
+		else
+			smp_set_ops(mdesc->smp);
 		smp_init_cpus();
 	}
 #endif