intel_idle: Voluntary leave_mm before entering deeper

Avoid TLB flush IPIs for the cores in deeper c-states by voluntary leave_mm()
before entering into that state. CPUs tend to flush TLB in those c-states
anyways.

acpi_idle does this with C3-type states, but it was not caried over
when intel_idle was introduced.  intel_idle can apply it
to C-states in addition to those that ACPI might export as C3...
Signed-off-by: NSuresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: NLen Brown <len.brown@intel.com>

drivers/idle/intel_idle.c

@@ -108,7 +108,7 @@ static struct cpuidle_state nehalem_cstates[MWAIT_MAX_NUM_CSTATES] = {
 		.name = "NHM-C3",
 		.desc = "MWAIT 0x10",
 		.driver_data = (void *) 0x10,
-		.flags = CPUIDLE_FLAG_TIME_VALID,
+		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
 		.exit_latency = 20,
 		.power_usage = 500,
 		.target_residency = 80,
@@ -117,7 +117,7 @@ static struct cpuidle_state nehalem_cstates[MWAIT_MAX_NUM_CSTATES] = {
 		.name = "NHM-C6",
 		.desc = "MWAIT 0x20",
 		.driver_data = (void *) 0x20,
-		.flags = CPUIDLE_FLAG_TIME_VALID,
+		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
 		.exit_latency = 200,
 		.power_usage = 350,
 		.target_residency = 800,
@@ -149,7 +149,7 @@ static struct cpuidle_state atom_cstates[MWAIT_MAX_NUM_CSTATES] = {
 		.name = "ATM-C4",
 		.desc = "MWAIT 0x30",
 		.driver_data = (void *) 0x30,
-		.flags = CPUIDLE_FLAG_TIME_VALID,
+		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
 		.exit_latency = 100,
 		.power_usage = 250,
 		.target_residency = 400,
@@ -159,7 +159,7 @@ static struct cpuidle_state atom_cstates[MWAIT_MAX_NUM_CSTATES] = {
 		.name = "ATM-C6",
 		.desc = "MWAIT 0x40",
 		.driver_data = (void *) 0x40,
-		.flags = CPUIDLE_FLAG_TIME_VALID,
+		.flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
 		.exit_latency = 200,
 		.power_usage = 150,
 		.target_residency = 800,
@@ -185,6 +185,16 @@ static int intel_idle(struct cpuidle_device *dev, struct cpuidle_state *state)
 
 	local_irq_disable();
 
+	/*
+	 * If the state flag indicates that the TLB will be flushed or if this
+	 * is the deepest c-state supported, do a voluntary leave mm to avoid
+	 * costly and mostly unnecessary wakeups for flushing the user TLB's
+	 * associated with the active mm.
+	 */
+	if (state->flags & CPUIDLE_FLAG_TLB_FLUSHED ||
+	    (&dev->states[dev->state_count - 1] == state))
+		leave_mm(cpu);
+
 	if (!(lapic_timer_reliable_states & (1 << (cstate))))
 		clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER, &cpu);
 

include/linux/cpuidle.h

@@ -53,6 +53,7 @@ struct cpuidle_state {
 #define CPUIDLE_FLAG_BALANCED	(0x40) /* medium latency, moderate savings */
 #define CPUIDLE_FLAG_DEEP	(0x80) /* high latency, large savings */
 #define CPUIDLE_FLAG_IGNORE	(0x100) /* ignore during this idle period */
+#define CPUIDLE_FLAG_TLB_FLUSHED (0x200) /* tlb will be flushed */
 
 #define CPUIDLE_DRIVER_FLAGS_MASK (0xFFFF0000)