struct start_info *xen_start_info;
 EXPORT_SYMBOL_GPL(xen_start_info);
 
-void xen_vcpu_setup(int cpu)
+static /* __initdata */ struct shared_info dummy_shared_info;
+
+/*
+ * Point at some empty memory to start with. We map the real shared_info
+ * page as soon as fixmap is up and running.
+ */
+struct shared_info *HYPERVISOR_shared_info = (void *)&dummy_shared_info;
+
+/*
+ * Flag to determine whether vcpu info placement is available on all
+ * VCPUs.  We assume it is to start with, and then set it to zero on
+ * the first failure.  This is because it can succeed on some VCPUs
+ * and not others, since it can involve hypervisor memory allocation,
+ * or because the guest failed to guarantee all the appropriate
+ * constraints on all VCPUs (ie buffer can't cross a page boundary).
+ *
+ * Note that any particular CPU may be using a placed vcpu structure,
+ * but we can only optimise if the all are.
+ *
+ * 0: not available, 1: available
+ */
+static int have_vcpu_info_placement = 1;
+
+static void __init xen_vcpu_setup(int cpu)
 {
+       struct vcpu_register_vcpu_info info;
+       int err;
+       struct vcpu_info *vcpup;
+
        per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
+
+       if (!have_vcpu_info_placement)
+               return;         /* already tested, not available */
+
+       vcpup = &per_cpu(xen_vcpu_info, cpu);
+
+       info.mfn = virt_to_mfn(vcpup);
+       info.offset = offset_in_page(vcpup);
+
+       printk(KERN_DEBUG "trying to map vcpu_info %d at %p, mfn %x, offset %d\n",
+              cpu, vcpup, info.mfn, info.offset);
+
+       /* Check to see if the hypervisor will put the vcpu_info
+          structure where we want it, which allows direct access via
+          a percpu-variable. */
+       err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
+
+       if (err) {
+               printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
+               have_vcpu_info_placement = 0;
+       } else {
+               /* This cpu is using the registered vcpu info, even if
+                  later ones fail to. */
+               per_cpu(xen_vcpu, cpu) = vcpup;
+               printk(KERN_DEBUG "cpu %d using vcpu_info at %p\n",
+                      cpu, vcpup);
+       }
 }
 
 static void __init xen_banner(void)
        return (-flags) & X86_EFLAGS_IF;
 }
 
+static unsigned long xen_save_fl_direct(void)
+{
+       unsigned long flags;
+
+       /* flag has opposite sense of mask */
+       flags = !x86_read_percpu(xen_vcpu_info.evtchn_upcall_mask);
+
+       /* convert to IF type flag
+          -0 -> 0x00000000
+          -1 -> 0xffffffff
+       */
+       return (-flags) & X86_EFLAGS_IF;
+}
+
 static void xen_restore_fl(unsigned long flags)
 {
        struct vcpu_info *vcpu;
        }
 }
 
+static void xen_restore_fl_direct(unsigned long flags)
+{
+       /* convert from IF type flag */
+       flags = !(flags & X86_EFLAGS_IF);
+
+       /* This is an atomic update, so no need to worry about
+          preemption. */
+       x86_write_percpu(xen_vcpu_info.evtchn_upcall_mask, flags);
+
+       /* If we get preempted here, then any pending event will be
+          handled anyway. */
+
+       if (flags == 0) {
+               barrier(); /* unmask then check (avoid races) */
+               if (unlikely(x86_read_percpu(xen_vcpu_info.evtchn_upcall_pending)))
+                       force_evtchn_callback();
+       }
+}
+
 static void xen_irq_disable(void)
 {
        /* There's a one instruction preempt window here.  We need to
        preempt_enable_no_resched();
 }
 
+static void xen_irq_disable_direct(void)
+{
+       /* Atomic update, so preemption not a concern. */
+       x86_write_percpu(xen_vcpu_info.evtchn_upcall_mask, 1);
+}
+
 static void xen_irq_enable(void)
 {
        struct vcpu_info *vcpu;
                force_evtchn_callback();
 }
 
+static void xen_irq_enable_direct(void)
+{
+       /* Atomic update, so preemption not a concern. */
+       x86_write_percpu(xen_vcpu_info.evtchn_upcall_mask, 0);
+
+       /* Doesn't matter if we get preempted here, because any
+          pending event will get dealt with anyway. */
+
+       barrier(); /* unmask then check (avoid races) */
+       if (unlikely(x86_read_percpu(xen_vcpu_info.evtchn_upcall_pending)))
+               force_evtchn_callback();
+}
+
 static void xen_safe_halt(void)
 {
        /* Blocking includes an implicit local_irq_enable(). */
        xen_mc_issue(PARAVIRT_LAZY_MMU);
 }
 
+static void xen_write_cr2(unsigned long cr2)
+{
+       x86_read_percpu(xen_vcpu)->arch.cr2 = cr2;
+}
+
 static unsigned long xen_read_cr2(void)
 {
        return x86_read_percpu(xen_vcpu)->arch.cr2;
 }
 
+static unsigned long xen_read_cr2_direct(void)
+{
+       return x86_read_percpu(xen_vcpu_info.arch.cr2);
+}
+
 static void xen_write_cr4(unsigned long cr4)
 {
        /* never allow TSC to be disabled */
                if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
                        BUG();
        }
+}
 
-       xen_vcpu_setup(smp_processor_id());
+/* This is called once we have the cpu_possible_map */
+void __init xen_setup_vcpu_info_placement(void)
+{
+       int cpu;
+
+       for_each_possible_cpu(cpu)
+               xen_vcpu_setup(cpu);
+
+       /* xen_vcpu_setup managed to place the vcpu_info within the
+          percpu area for all cpus, so make use of it */
+       if (have_vcpu_info_placement) {
+               printk(KERN_INFO "Xen: using vcpu_info placement\n");
+
+               paravirt_ops.save_fl = xen_save_fl_direct;
+               paravirt_ops.restore_fl = xen_restore_fl_direct;
+               paravirt_ops.irq_disable = xen_irq_disable_direct;
+               paravirt_ops.irq_enable = xen_irq_enable_direct;
+               paravirt_ops.read_cr2 = xen_read_cr2_direct;
+       }
 }
 
 static const struct paravirt_ops xen_paravirt_ops __initdata = {
        .write_cr0 = native_write_cr0,
 
        .read_cr2 = xen_read_cr2,
-       .write_cr2 = native_write_cr2,
+       .write_cr2 = xen_write_cr2,
 
        .read_cr3 = xen_read_cr3,
        .write_cr3 = xen_write_cr3,
        /* keep using Xen gdt for now; no urgent need to change it */
 
        x86_write_percpu(xen_cr3, __pa(pgd));
-       xen_vcpu_setup(0);
+
+#ifdef CONFIG_SMP
+       /* Don't do the full vcpu_info placement stuff until we have a
+          possible map. */
+       per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
+#else
+       /* May as well do it now, since there's no good time to call
+          it later on UP. */
+       xen_setup_vcpu_info_placement();
+#endif
 
        paravirt_ops.kernel_rpl = 1;
        if (xen_feature(XENFEAT_supervisor_mode_kernel))