* ID field from 8 to 15 bits, allowing to target APIC IDs up 32768.
  */
 #define XEN_HVM_CPUID_EXT_DEST_ID      (1u << 5)
+/* Per-vCPU event channel upcalls */
+#define XEN_HVM_CPUID_UPCALL_VECTOR    (1u << 6)
 
 /*
  * Leaf 6 (0x40000x05)
 
 /* No need for a barrier -- XCHG is a barrier on x86. */
 #define xchg_xen_ulong(ptr, val) xchg((ptr), (val))
 
-extern int xen_have_vector_callback;
+extern bool xen_have_vector_callback;
 
 /*
  * Events delivered via platform PCI interrupts are always
        return (!xen_hvm_domain() || xen_have_vector_callback);
 }
 
+extern bool xen_percpu_upcall;
 #endif /* _ASM_X86_XEN_EVENTS_H */
 
 
 struct shared_info xen_dummy_shared_info;
 
-__read_mostly int xen_have_vector_callback;
+__read_mostly bool xen_have_vector_callback = true;
 EXPORT_SYMBOL_GPL(xen_have_vector_callback);
 
 /*
 
 
 #include <xen/features.h>
 #include <xen/events.h>
+#include <xen/hvm.h>
+#include <xen/interface/hvm/hvm_op.h>
 #include <xen/interface/memory.h>
 
 #include <asm/apic.h>
 
 static unsigned long shared_info_pfn;
 
+__ro_after_init bool xen_percpu_upcall;
+EXPORT_SYMBOL_GPL(xen_percpu_upcall);
+
 void xen_hvm_init_shared_info(void)
 {
        struct xen_add_to_physmap xatp;
 {
        struct pt_regs *old_regs = set_irq_regs(regs);
 
+       if (xen_percpu_upcall)
+               ack_APIC_irq();
+
        inc_irq_stat(irq_hv_callback_count);
 
        xen_hvm_evtchn_do_upcall();
        if (!xen_have_vector_callback)
                return 0;
 
+       if (xen_percpu_upcall) {
+               rc = xen_set_upcall_vector(cpu);
+               if (rc) {
+                       WARN(1, "HVMOP_set_evtchn_upcall_vector"
+                            " for CPU %d failed: %d\n", cpu, rc);
+                       return rc;
+               }
+       }
+
        if (xen_feature(XENFEAT_hvm_safe_pvclock))
                xen_setup_timer(cpu);
 
        return 0;
 }
 
-static bool no_vector_callback __initdata;
-
 static void __init xen_hvm_guest_init(void)
 {
        if (xen_pv_domain())
 
        xen_panic_handler_init();
 
-       if (!no_vector_callback && xen_feature(XENFEAT_hvm_callback_vector))
-               xen_have_vector_callback = 1;
-
        xen_hvm_smp_init();
        WARN_ON(xen_cpuhp_setup(xen_cpu_up_prepare_hvm, xen_cpu_dead_hvm));
        xen_unplug_emulated_devices();
 
 static __init int xen_parse_no_vector_callback(char *arg)
 {
-       no_vector_callback = true;
+       xen_have_vector_callback = false;
        return 0;
 }
 early_param("xen_no_vector_callback", xen_parse_no_vector_callback);
 
 #include <xen/hvm.h>
 #include <xen/features.h>
 #include <xen/interface/features.h>
+#include <xen/events.h>
 
 #include "xen-ops.h"
 
                xen_hvm_init_shared_info();
                xen_vcpu_restore();
        }
-       xen_setup_callback_vector();
+       if (xen_percpu_upcall) {
+               unsigned int cpu;
+
+               for_each_online_cpu(cpu)
+                       BUG_ON(xen_set_upcall_vector(cpu));
+       } else {
+               xen_setup_callback_vector();
+       }
        xen_unplug_emulated_devices();
 }
 
 #include <asm/irq.h>
 #include <asm/io_apic.h>
 #include <asm/i8259.h>
+#include <asm/xen/cpuid.h>
 #include <asm/xen/pci.h>
 #endif
 #include <asm/sync_bitops.h>
        .irq_ack                = ack_dynirq,
 };
 
+#ifdef CONFIG_X86
 #ifdef CONFIG_XEN_PVHVM
 /* Vector callbacks are better than PCI interrupts to receive event
  * channel notifications because we can receive vector callbacks on any
                callback_via = HVM_CALLBACK_VECTOR(HYPERVISOR_CALLBACK_VECTOR);
                if (xen_set_callback_via(callback_via)) {
                        pr_err("Request for Xen HVM callback vector failed\n");
-                       xen_have_vector_callback = 0;
+                       xen_have_vector_callback = false;
                }
        }
 }
 
+/*
+ * Setup per-vCPU vector-type callbacks. If this setup is unavailable,
+ * fallback to the global vector-type callback.
+ */
+static __init void xen_init_setup_upcall_vector(void)
+{
+       if (!xen_have_vector_callback)
+               return;
+
+       if ((cpuid_eax(xen_cpuid_base() + 4) & XEN_HVM_CPUID_UPCALL_VECTOR) &&
+           !xen_set_upcall_vector(0))
+               xen_percpu_upcall = true;
+       else if (xen_feature(XENFEAT_hvm_callback_vector))
+               xen_setup_callback_vector();
+       else
+               xen_have_vector_callback = false;
+}
+
+int xen_set_upcall_vector(unsigned int cpu)
+{
+       int rc;
+       xen_hvm_evtchn_upcall_vector_t op = {
+               .vector = HYPERVISOR_CALLBACK_VECTOR,
+               .vcpu = per_cpu(xen_vcpu_id, cpu),
+       };
+
+       rc = HYPERVISOR_hvm_op(HVMOP_set_evtchn_upcall_vector, &op);
+       if (rc)
+               return rc;
+
+       /* Trick toolstack to think we are enlightened. */
+       if (!cpu)
+               rc = xen_set_callback_via(1);
+
+       return rc;
+}
+
 static __init void xen_alloc_callback_vector(void)
 {
        if (!xen_have_vector_callback)
 }
 #else
 void xen_setup_callback_vector(void) {}
+static inline void xen_init_setup_upcall_vector(void) {}
+int xen_set_upcall_vector(unsigned int cpu) {}
 static inline void xen_alloc_callback_vector(void) {}
-#endif
+#endif /* CONFIG_XEN_PVHVM */
+#endif /* CONFIG_X86 */
 
 bool xen_fifo_events = true;
 module_param_named(fifo_events, xen_fifo_events, bool, 0);
                if (xen_initial_domain())
                        pci_xen_initial_domain();
        }
-       if (xen_feature(XENFEAT_hvm_callback_vector)) {
-               xen_setup_callback_vector();
-               xen_alloc_callback_vector();
-       }
+       xen_init_setup_upcall_vector();
+       xen_alloc_callback_vector();
+
 
        if (xen_hvm_domain()) {
                native_init_IRQ();
 
 
 void xen_setup_callback_vector(void);
 
+int xen_set_upcall_vector(unsigned int cpu);
+
 #endif /* XEN_HVM_H__ */
 
 };
 DEFINE_GUEST_HANDLE_STRUCT(xen_hvm_get_mem_type);
 
+#if defined(__i386__) || defined(__x86_64__)
+
+/*
+ * HVMOP_set_evtchn_upcall_vector: Set a <vector> that should be used for event
+ *                                 channel upcalls on the specified <vcpu>. If set,
+ *                                 this vector will be used in preference to the
+ *                                 domain global callback via (see
+ *                                 HVM_PARAM_CALLBACK_IRQ).
+ */
+#define HVMOP_set_evtchn_upcall_vector 23
+struct xen_hvm_evtchn_upcall_vector {
+    uint32_t vcpu;
+    uint8_t vector;
+};
+typedef struct xen_hvm_evtchn_upcall_vector xen_hvm_evtchn_upcall_vector_t;
+DEFINE_GUEST_HANDLE_STRUCT(xen_hvm_evtchn_upcall_vector_t);
+
+#endif /* defined(__i386__) || defined(__x86_64__) */
+
 #endif /* __XEN_PUBLIC_HVM_HVM_OP_H__ */