xen: cache cr0 value to avoid trap'n'emulate for read_cr0

stts() is implemented in terms of read_cr0/write_cr0 to update the
state of the TS bit.  This happens during context switch, and so
is fairly performance critical.  Rather than falling back to
a trap-and-emulate native read_cr0, implement our own by caching
the last-written value from write_cr0 (the TS bit is the only one
we really care about).

Impact: optimise Xen context switches
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>

arch/x86/xen/enlighten.c

@@ -658,10 +658,26 @@ static void xen_clts(void)
 	xen_mc_issue(PARAVIRT_LAZY_CPU);
 }
 
+static DEFINE_PER_CPU(unsigned long, xen_cr0_value);
+
+static unsigned long xen_read_cr0(void)
+{
+	unsigned long cr0 = percpu_read(xen_cr0_value);
+
+	if (unlikely(cr0 == 0)) {
+		cr0 = native_read_cr0();
+		percpu_write(xen_cr0_value, cr0);
+	}
+
+	return cr0;
+}
+
 static void xen_write_cr0(unsigned long cr0)
 {
 	struct multicall_space mcs;
 
+	percpu_write(xen_cr0_value, cr0);
+
 	/* Only pay attention to cr0.TS; everything else is
 	   ignored. */
 	mcs = xen_mc_entry(0);
@@ -847,7 +863,7 @@ static const struct pv_cpu_ops xen_cpu_ops __initdata = {
 
 	.clts = xen_clts,
 
-	.read_cr0 = native_read_cr0,
+	.read_cr0 = xen_read_cr0,
 	.write_cr0 = xen_write_cr0,
 
 	.read_cr4 = native_read_cr4,