diff --git a/arch/ia64/kernel/crash.c b/arch/ia64/kernel/crash.c
index b2a8b3da8af3..9c851b73f276 100644
--- a/arch/ia64/kernel/crash.c
+++ b/arch/ia64/kernel/crash.c
@@ -23,6 +23,7 @@
 int kdump_status[NR_CPUS];
 static atomic_t kdump_cpu_frozen;
 atomic_t kdump_in_progress;
+static int kdump_freeze_monarch;
 static int kdump_on_init = 1;
 static int kdump_on_fatal_mca = 1;
 
@@ -108,6 +109,33 @@ machine_crash_shutdown(struct pt_regs *pt)
 	 */
 	kexec_disable_iosapic();
 #ifdef CONFIG_SMP
+	/*
+	 * If kdump_on_init is set and an INIT is asserted here, kdump will
+	 * be started again via INIT monarch.
+	 */
+	local_irq_disable();
+	ia64_set_psr_mc();	/* mask MCA/INIT */
+	if (atomic_inc_return(&kdump_in_progress) != 1)
+		unw_init_running(kdump_cpu_freeze, NULL);
+
+	/*
+	 * Now this cpu is ready for kdump.
+	 * Stop all others by IPI or INIT.  They could receive INIT from
+	 * outside and might be INIT monarch, but only thing they have to
+	 * do is falling into kdump_cpu_freeze().
+	 *
+	 * If an INIT is asserted here:
+	 * - All receivers might be slaves, since some of cpus could already
+	 *   be frozen and INIT might be masked on monarch.  In this case,
+	 *   all slaves will park in while (monarch_cpu == -1) loop before
+	 *   DIE_INIT_SLAVE_ENTER that for waiting monarch enters.
+	 *	=> TBD: freeze all slaves
+	 * - One might be a monarch, but INIT rendezvous will fail since
+	 *   at least this cpu already have INIT masked so it never join
+	 *   to the rendezvous.  In this case, all slaves and monarch will
+	 *   be frozen after timeout of the INIT rendezvous.
+	 *	=> TBD: freeze them without waiting timeout
+	 */
 	kdump_smp_send_stop();
 	/* not all cpu response to IPI, send INIT to freeze them */
 	if (kdump_wait_cpu_freeze() && kdump_on_init) 	{
@@ -177,13 +205,18 @@ kdump_init_notifier(struct notifier_block *self, unsigned long val, void *data)
 	switch (val) {
 	case DIE_INIT_MONARCH_PROCESS:
 		if (kdump_on_init) {
-			atomic_set(&kdump_in_progress, 1);
+			if (atomic_inc_return(&kdump_in_progress) != 1)
+				kdump_freeze_monarch = 1;
 			*(nd->monarch_cpu) = -1;
 		}
 		break;
 	case DIE_INIT_MONARCH_LEAVE:
-		if (kdump_on_init)
-			machine_kdump_on_init();
+		if (kdump_on_init) {
+			if (kdump_freeze_monarch)
+				unw_init_running(kdump_cpu_freeze, NULL);
+			else
+				machine_kdump_on_init();
+		}
 		break;
 	case DIE_INIT_SLAVE_LEAVE:
 		if (atomic_read(&kdump_in_progress))
@@ -196,9 +229,11 @@ kdump_init_notifier(struct notifier_block *self, unsigned long val, void *data)
 	case DIE_MCA_MONARCH_LEAVE:
 		/* *(nd->data) indicate if MCA is recoverable */
 		if (kdump_on_fatal_mca && !(*(nd->data))) {
-			atomic_set(&kdump_in_progress, 1);
-			*(nd->monarch_cpu) = -1;
-			machine_kdump_on_init();
+			if (atomic_inc_return(&kdump_in_progress) == 1) {
+				*(nd->monarch_cpu) = -1;
+				machine_kdump_on_init();
+			}
+			/* We got fatal MCA while kdump!? No way!! */
 		}
 		break;
 	}
diff --git a/arch/ia64/kernel/relocate_kernel.S b/arch/ia64/kernel/relocate_kernel.S
index 903babd22d62..32f6fc131fbe 100644
--- a/arch/ia64/kernel/relocate_kernel.S
+++ b/arch/ia64/kernel/relocate_kernel.S
@@ -52,7 +52,7 @@ GLOBAL_ENTRY(relocate_new_kernel)
 	srlz.i
 	;;
 	mov ar.rnat=r18
-	rfi
+	rfi				// note: this unmask MCA/INIT (psr.mc)
 	;;
 1:
 	//physical mode code begin