On 11/28/2011 04:33 AM, Benjamin Herrenschmidt wrote:

On Thu, 2011-11-17 at 16:58 +0530, Deepthi Dharwar wrote:

quoted

This patch implements a backhand cpuidle driver for pSeries
based on pseries_dedicated_idle_loop and pseries_shared_idle_loop
routines.  The driver is built only if CONFIG_CPU_IDLE is set. This
cpuidle driver uses global registration of idle states and
not per-cpu.

 .../...

quoted

+#define MAX_IDLE_STATE_COUNT	2
+
+static int max_cstate = MAX_IDLE_STATE_COUNT - 1;

Why "cstate" ? This isn't an x86 :-)

Sure, I shall change it right away :)

quoted

+static struct cpuidle_device __percpu *pseries_idle_cpuidle_devices;

Shorter name please. pseries_cpuidle_devs is fine.

I ll do so.

quoted

+static struct cpuidle_state *cpuidle_state_table;
+
+void update_smt_snooze_delay(int snooze)
+{
+	struct cpuidle_driver *drv = cpuidle_get_driver();
+	if (drv)
+		drv->states[0].target_residency = snooze;
+}
+
+static inline void idle_loop_prolog(unsigned long *in_purr, ktime_t *kt_before)
+{
+
+	*kt_before = ktime_get_real();
+	*in_purr = mfspr(SPRN_PURR);
+	/*
+	 * Indicate to the HV that we are idle. Now would be
+	 * a good time to find other work to dispatch.
+	 */
+	get_lppaca()->idle = 1;
+	get_lppaca()->donate_dedicated_cpu = 1;
+}

I notice that you call this on shared processors as well. The old ocde
used to not set donate_dedicated_cpu in that case. I assume that's not a
big deal and that the HV will just ignore it in the shared processor
case but please add a comment after you've verified it.

Yes, the old code does not set donate_dedicated_cpu. But yes I will
try testing it in a shared proc config but also remove this
initialization for shared idle loop.

 

quoted

+static inline  s64 idle_loop_epilog(unsigned long in_purr, ktime_t kt_before)
+{
+	get_lppaca()->wait_state_cycles += mfspr(SPRN_PURR) - in_purr;
+	get_lppaca()->donate_dedicated_cpu = 0;
+	get_lppaca()->idle = 0;
+
+	return ktime_to_us(ktime_sub(ktime_get_real(), kt_before));
+}
+
+
+static int snooze_loop(struct cpuidle_device *dev,
+			struct cpuidle_driver *drv,
+			int index)
+{
+	unsigned long in_purr;
+	ktime_t kt_before;
+
+	idle_loop_prolog(&in_purr, &kt_before);
+
+	local_irq_enable();
+	set_thread_flag(TIF_POLLING_NRFLAG);
+	while (!need_resched()) {
+		ppc64_runlatch_off();
+		HMT_low();
+		HMT_very_low();
+	}
+	HMT_medium();
+	clear_thread_flag(TIF_POLLING_NRFLAG);
+	smp_mb();
+	local_irq_disable();
+
+	dev->last_residency =
+		(int)idle_loop_epilog(in_purr, kt_before);
+	return index;
+}

So your snooze loop has no timeout, is that handled by the cpuidle
driver using some kind of timer ? That sounds a lot less efficient than
passing a max delay to the snooze loop to handle getting into a deeper
state after a bit of snoozing rather than interrupting etc...

My bad, snooze_loop is essential for a time out. Nope cpuidle
driver doesn't have any timer mechanism. I ll fix it.
Need to add loop for snooze time.

quoted

+static int dedicated_cede_loop(struct cpuidle_device *dev,
+				struct cpuidle_driver *drv,
+				int index)
+{
+	unsigned long in_purr;
+	ktime_t kt_before;
+
+	idle_loop_prolog(&in_purr, &kt_before);
+
+	ppc64_runlatch_off();
+	HMT_medium();
+	cede_processor();
+
+	dev->last_residency =
+		(int)idle_loop_epilog(in_purr, kt_before);
+	return index;
+}
+
+static int shared_cede_loop(struct cpuidle_device *dev,
+			struct cpuidle_driver *drv,
+			int index)
+{
+	unsigned long in_purr;
+	ktime_t kt_before;
+
+	idle_loop_prolog(&in_purr, &kt_before);
+
+	/*
+	 * Yield the processor to the hypervisor.  We return if
+	 * an external interrupt occurs (which are driven prior
+	 * to returning here) or if a prod occurs from another
+	 * processor. When returning here, external interrupts
+	 * are enabled.
+	 */
+	cede_processor();
+
+	dev->last_residency =
+		(int)idle_loop_epilog(in_purr, kt_before);
+	return index;
+}
+
+/*
+ * States for dedicated partition case.
+ */
+static struct cpuidle_state dedicated_states[MAX_IDLE_STATE_COUNT] = {
+	{ /* Snooze */
+		.name = "snooze",
+		.desc = "snooze",
+		.flags = CPUIDLE_FLAG_TIME_VALID,
+		.exit_latency = 0,
+		.target_residency = 0,
+		.enter = &snooze_loop },
+	{ /* CEDE */
+		.name = "CEDE",
+		.desc = "CEDE",
+		.flags = CPUIDLE_FLAG_TIME_VALID,
+		.exit_latency = 1,
+		.target_residency = 10,
+		.enter = &dedicated_cede_loop },
+};
+
+/*
+ * States for shared partition case.
+ */
+static struct cpuidle_state shared_states[MAX_IDLE_STATE_COUNT] = {
+	{ /* Shared Cede */
+		.name = "Shared Cede",
+		.desc = "Shared Cede",
+		.flags = CPUIDLE_FLAG_TIME_VALID,
+		.exit_latency = 0,
+		.target_residency = 0,
+		.enter = &shared_cede_loop },
+};
+
+int pseries_notify_cpuidle_add_cpu(int cpu)
+{
+	struct cpuidle_device *dev =
+			per_cpu_ptr(pseries_idle_cpuidle_devices, cpu);
+	if (dev && cpuidle_get_driver()) {
+		cpuidle_disable_device(dev);
+		cpuidle_enable_device(dev);
+	}
+	return 0;
+}
+
+/*
+ * pseries_idle_cpuidle_driver_init()
+ */
+static int pseries_idle_cpuidle_driver_init(void)
+{

Drop the "idle", those names are too long.

 Sure.

quoted

+	int cstate;

And use something else than 'cstate', we are among civilized people
here ;-)

Yes :)

quoted

+	struct cpuidle_driver *drv = &pseries_idle_driver;
+
+	drv->state_count = 0;
+
+	for (cstate = 0; cstate < MAX_IDLE_STATE_COUNT; ++cstate) {
+
+		if (cstate > max_cstate)
+			break;
+
+		/* is the state not enabled? */
+		if (cpuidle_state_table[cstate].enter == NULL)
+			continue;
+
+		drv->states[drv->state_count] =	/* structure copy */
+			cpuidle_state_table[cstate];
+
+		if (cpuidle_state_table == dedicated_states)
+			drv->states[drv->state_count].target_residency =
+				__get_cpu_var(smt_snooze_delay);
+
+		drv->state_count += 1;
+	}
+
+	return 0;
+}
+
+/* pseries_idle_devices_uninit(void)
+ * unregister cpuidle devices and de-allocate memory
+ */
+static void pseries_idle_devices_uninit(void)
+{
+	int i;
+	struct cpuidle_device *dev;
+
+	for_each_possible_cpu(i) {
+		dev = per_cpu_ptr(pseries_idle_cpuidle_devices, i);
+		cpuidle_unregister_device(dev);
+	}
+
+	free_percpu(pseries_idle_cpuidle_devices);
+	return;
+}
+
+/* pseries_idle_devices_init()
+ * allocate, initialize and register cpuidle device
+ */
+static int pseries_idle_devices_init(void)
+{
+	int i;
+	struct cpuidle_driver *drv = &pseries_idle_driver;
+	struct cpuidle_device *dev;
+
+	pseries_idle_cpuidle_devices = alloc_percpu(struct cpuidle_device);
+	if (pseries_idle_cpuidle_devices == NULL)
+		return -ENOMEM;
+
+	for_each_possible_cpu(i) {
+		dev = per_cpu_ptr(pseries_idle_cpuidle_devices, i);
+		dev->state_count = drv->state_count;
+		dev->cpu = i;
+		if (cpuidle_register_device(dev)) {
+			printk(KERN_DEBUG \
+				"cpuidle_register_device %d failed!\n", i);
+			return -EIO;
+		}
+	}
+
+	return 0;
+}
+
+/*
+ * pseries_idle_probe()
+ * Choose state table for shared versus dedicated partition
+ */
+static int pseries_idle_probe(void)
+{
+	if (max_cstate == 0) {
+		printk(KERN_DEBUG "pseries processor idle disabled.\n");
+		return -EPERM;
+	}
+
+	if (!firmware_has_feature(FW_FEATURE_SPLPAR)) {
+		printk(KERN_DEBUG "Using default idle\n");
+		return -ENODEV;
+	}

Don't even printk here, this will happen on all !pseries machines and
the debug output isn't useful. Also move the check of max-cstate to
after the splpar test.

I ll move the check above.

 

Overall, I quite like these patches, my comments are all pretty minor,
hopefully the next round should be the one.

Thank you.

Cheers,
Ben.

Regards,
Deepthi