The following series consists of a bunch of fixes for
cpuidle back-end driver for ppc64 (pSeries) platform.
The series applies on 3.6-rc7 and has been tested on ppc64 pSeries
POWER7 system with snooze and nap states.
With these fixes, avg idle residency for nap on 50 percent idle system
increased from 96 to 99.4 percent and idle residency for snooze for
the same decreased from 4 to 0.6 percent, determined running spec power
benchmark runs.
Deepthi Dharwar (3):
cpuidle: (POWER) Fix snooze state problem persistant in the current cpuidle design on pseries.
cpuidle: (POWER) Fix smt_snooze_delay functionality.
cpuidle: (POWER) Fix target residency initialisation in pseries cpuidle
arch/powerpc/include/asm/processor.h | 4 +
arch/powerpc/kernel/sysfs.c | 2 -
arch/powerpc/platforms/pseries/processor_idle.c | 62 +++++++++++------------
3 files changed, 32 insertions(+), 36 deletions(-)
-- Deepthi
Remove the redundant target residency initialisation in pseries_cpuidle_driver_init().
This is currently over-writing the residency time updated as part of the static
table, resulting in all the idle states having the same target
residency of 100us which is incorrect. This may result in the menu governor making
wrong state decisions.
Signed-off-by: Deepthi Dharwar <[email protected]>
---
arch/powerpc/platforms/pseries/processor_idle.c | 4 ----
1 file changed, 4 deletions(-)
diff --git a/arch/powerpc/platforms/pseries/processor_idle.c b/arch/powerpc/platforms/pseries/processor_idle.c
index 455760b..02e425a 100644
--- a/arch/powerpc/platforms/pseries/processor_idle.c
+++ b/arch/powerpc/platforms/pseries/processor_idle.c
@@ -246,10 +246,6 @@ static int pseries_cpuidle_driver_init(void)
drv->states[drv->state_count] = /* structure copy */
cpuidle_state_table[idle_state];
- if (cpuidle_state_table == dedicated_states)
- drv->states[drv->state_count].target_residency =
- __get_cpu_var(smt_snooze_delay);
-
drv->state_count += 1;
}
Earlier without cpuidle framework on pseries, the native arch
idle routine comprised of both snooze and nap
states. smt_snooze_delay variable was used to delay
the idle process entry to deeper idle state like nap.
With the coming of cpuidle, this arch specific idle was replaced
by two different idle routines, one for supporting snooze and other
for nap. This enabled addition of more
low level idle states on pseries in the future.
On adopting the generic cpuidle framework for POWER systems,
the decision of which idle state to choose from, given a predicted
idle time is taken by the menu governor based on
target_residency and exit_latency of the idle states.
target_residency is the minimum time to be resident in that idle state.
Exit_latency is time taken to exit out of idle state.
Deeper the idle state, both the target residency and exit latency
would be higher.
In the current design, smt_snooze_delay is used as target_residency
for the snooze state which is incorrect, as it is not the
minimum but the maximum duration to be in snooze state.
This would result in the governor in taking bad decision,
as presently target_residency of nap < target_residency of snooze
inspite of nap being deeper idle state.
This patch aims to fix this problem by replacing the smt_snooze_delay loop
in snooze state, with the need_resched() as the governor is aware of
entry and exit of various idle transitions based on which
next idle time prediction.
The governor is intelligent enough to determine the idle state the needs to
be transitioned to and maintains a whole of heuristics including
io load, previous idle states predictions etc for the same, based on
which idle state entry decision is taken.
With this fix, of setting target_residency of snooze to 0
nap to smt_snooze_delay
if the predicted idle time is less
than smt_snooze_delay (target_residency of nap)
value governor would pick snooze state, else nap. This adhers to the
previous native idle design.
Signed-off-by: Deepthi Dharwar <[email protected]>
---
arch/powerpc/platforms/pseries/processor_idle.c | 36 ++++++++---------------
1 file changed, 13 insertions(+), 23 deletions(-)
diff --git a/arch/powerpc/platforms/pseries/processor_idle.c b/arch/powerpc/platforms/pseries/processor_idle.c
index a32d56d..45d00e5 100644
--- a/arch/powerpc/platforms/pseries/processor_idle.c
+++ b/arch/powerpc/platforms/pseries/processor_idle.c
@@ -59,32 +59,22 @@ static int snooze_loop(struct cpuidle_device *dev,
{
unsigned long in_purr;
ktime_t kt_before;
- unsigned long start_snooze;
- long snooze = drv->states[0].target_residency;
+ int cpu = dev->cpu;
idle_loop_prolog(&in_purr, &kt_before);
+ local_irq_enable();
+ set_thread_flag(TIF_POLLING_NRFLAG);
- if (snooze) {
- start_snooze = get_tb() + snooze * tb_ticks_per_usec;
- local_irq_enable();
- set_thread_flag(TIF_POLLING_NRFLAG);
-
- while ((snooze < 0) || (get_tb() < start_snooze)) {
- if (need_resched() || cpu_is_offline(dev->cpu))
- goto out;
- ppc64_runlatch_off();
- HMT_low();
- HMT_very_low();
- }
-
- HMT_medium();
- clear_thread_flag(TIF_POLLING_NRFLAG);
- smp_mb();
- local_irq_disable();
+ while ((!need_resched()) && cpu_online(cpu)) {
+ ppc64_runlatch_off();
+ HMT_low();
+ HMT_very_low();
}
-out:
HMT_medium();
+ clear_thread_flag(TIF_POLLING_NRFLAG);
+ smp_mb();
+
dev->last_residency =
(int)idle_loop_epilog(in_purr, kt_before);
return index;
@@ -165,8 +155,8 @@ static struct cpuidle_state dedicated_states[MAX_IDLE_STATE_COUNT] = {
.name = "CEDE",
.desc = "CEDE",
.flags = CPUIDLE_FLAG_TIME_VALID,
- .exit_latency = 1,
- .target_residency = 10,
+ .exit_latency = 10,
+ .target_residency = 100,
.enter = &dedicated_cede_loop },
};
@@ -197,7 +187,7 @@ void update_smt_snooze_delay(int cpu, int residency)
dev->states_usage[1].disable = 1;
} else
if (drv)
- drv->states[0].target_residency = residency;
+ drv->states[1].target_residency = residency;
}
static int pseries_cpuidle_add_cpu_notifier(struct notifier_block *n,
smt_snooze_delay was designed to delay idle loop's nap entry
in the native idle code before it got ported over to use as part of
the cpuidle framework.
A -ve value assigned to smt_snooze_delay should result in
busy looping, in other words disabling the entry to nap state.
- https://lists.ozlabs.org/pipermail/linuxppc-dev/2010-May/082450.html
This particular functionality can be achieved currently by
echo 1 > /sys/devices/system/cpu/cpu*/state1/disable
but it is broken when one assigns -ve value to the smt_snooze_delay
variable either via sysfs entry or ppc64_cpu util.
This patch aims to fix this, by disabling nap state when smt_snooze_delay
variable is set to -ve value.
Signed-off-by: Deepthi Dharwar <[email protected]>
---
arch/powerpc/include/asm/processor.h | 4 ++--
arch/powerpc/kernel/sysfs.c | 2 +-
arch/powerpc/platforms/pseries/processor_idle.c | 24 ++++++++++++++++-------
3 files changed, 20 insertions(+), 10 deletions(-)
diff --git a/arch/powerpc/include/asm/processor.h b/arch/powerpc/include/asm/processor.h
index 9dc5cd1..7e44855 100644
--- a/arch/powerpc/include/asm/processor.h
+++ b/arch/powerpc/include/asm/processor.h
@@ -391,9 +391,9 @@ extern int powersave_nap; /* set if nap mode can be used in idle loop */
extern void power7_nap(void);
#ifdef CONFIG_PSERIES_IDLE
-extern void update_smt_snooze_delay(int snooze);
+extern void update_smt_snooze_delay(int cpu, int residency);
#else
-static inline void update_smt_snooze_delay(int snooze) {}
+static inline void update_smt_snooze_delay(int cpu, int residency) {}
#endif
extern void flush_instruction_cache(void);
diff --git a/arch/powerpc/kernel/sysfs.c b/arch/powerpc/kernel/sysfs.c
index 8302af6..cf357a0 100644
--- a/arch/powerpc/kernel/sysfs.c
+++ b/arch/powerpc/kernel/sysfs.c
@@ -50,7 +50,7 @@ static ssize_t store_smt_snooze_delay(struct device *dev,
return -EINVAL;
per_cpu(smt_snooze_delay, cpu->dev.id) = snooze;
- update_smt_snooze_delay(snooze);
+ update_smt_snooze_delay(cpu->dev.id, snooze);
return count;
}
diff --git a/arch/powerpc/platforms/pseries/processor_idle.c b/arch/powerpc/platforms/pseries/processor_idle.c
index 02e425a..a32d56d 100644
--- a/arch/powerpc/platforms/pseries/processor_idle.c
+++ b/arch/powerpc/platforms/pseries/processor_idle.c
@@ -33,13 +33,6 @@ static int max_idle_state = MAX_IDLE_STATE_COUNT - 1;
static struct cpuidle_device __percpu *pseries_cpuidle_devices;
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)
{
@@ -190,6 +183,23 @@ static struct cpuidle_state shared_states[MAX_IDLE_STATE_COUNT] = {
.enter = &shared_cede_loop },
};
+void update_smt_snooze_delay(int cpu, int residency)
+{
+ struct cpuidle_driver *drv = cpuidle_get_driver();
+ struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
+
+ if (cpuidle_state_table != dedicated_states)
+ return;
+
+ if (residency < 0) {
+ /* Disable the Nap state on that cpu */
+ if (dev)
+ dev->states_usage[1].disable = 1;
+ } else
+ if (drv)
+ drv->states[0].target_residency = residency;
+}
+
static int pseries_cpuidle_add_cpu_notifier(struct notifier_block *n,
unsigned long action, void *hcpu)
{