schedule(): hold runq lock for the entire duration of context switching

releasing the runq lock after loading page tables but before the actual
context switch can leave execution in an inconsistent if the current
process is descheduled from an IRQ between these two steps.
this patch holds the runq lock with IRQs disabled and makes the context
switch a single atomic operation.

arch/x86/kernel/cpu.c

@@ -1012,6 +1012,18 @@ void ihk_mc_init_context(ihk_mc_kernel_context_t *new_ctx,
 }
 
 extern char enter_user_mode[];
+
+/* 
+ * Release runq_lock before entering user space.
+ * This is needed because schedule() holds the runq lock throughout
+ * the context switch and when a new process is created it starts
+ * execution in enter_user_mode, which in turn calls this function.
+ */
+void release_runq_lock(void)
+{
+	ihk_mc_spinlock_unlock(&(cpu_local_var(runq_lock)), 
+			cpu_local_var(runq_irqstate));
+}
                                        
 /*@
   @ requires \valid(ctx);

arch/x86/kernel/interrupt.S

@@ -206,6 +206,7 @@ x86_syscall:
 		
 .globl enter_user_mode
 enter_user_mode:
+		callq	release_runq_lock
 		movq	$0,  %rdi
 		movq	%rsp, %rsi
 		call check_signal