Hello,
This patch series are for enabling the asynchronous threads for the phases
resume_noirq, resume_early, suspend_noirq and suspend_late.
Just like commit 5af84b82701a and 97df8c12995, with async threads it will
reduce the system suspending and resuming time significantly.
With these patches, in my test platform, it saved 80% time in resume_noirq
phase.
Best Regards,
---
V2:
-- Based on Rafael's minor changes related to coding style, white space etc;
-- Rafael pointed out the v1 series break the device parent-children
suspending/resuming order when enabling asyncing, here using the
dev completion to sync parent-children order;
V3:
-- In patch v2 5/5, there is one missing "dpm_wait_for_children";
V4:
-- Rafael pointed out that dpm_wait_for_children()/dpm_wait()
should be put after some simple returning statement;
---
[PATCH v4 1/5] PM / sleep: Two flags for async suspend_noirq and
[PATCH v4 2/5] PM / sleep: Asynchronous threads for resume_noirq
[PATCH v4 3/5] PM / sleep: Asynchronous threads for resume_early
[PATCH v4 4/5] PM / sleep: Asynchronous threads for suspend_noirq
[PATCH v4 5/5] PM / sleep: Asynchronous threads for suspend_late
drivers/base/power/main.c | 275 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++----------------------------
include/linux/pm.h | 2 ++
2 files changed, 227 insertions(+), 50 deletions(-)
The patch is a helper adding two new flags for implementing
async threads for suspend_noirq and suspend_late.
Signed-off-by: Chuansheng Liu <[email protected]>
---
drivers/base/power/main.c | 24 ++++++++++++++++++++++--
include/linux/pm.h | 2 ++
2 files changed, 24 insertions(+), 2 deletions(-)
diff --git a/drivers/base/power/main.c b/drivers/base/power/main.c
index 1b41fca..00c53eb 100644
--- a/drivers/base/power/main.c
+++ b/drivers/base/power/main.c
@@ -91,6 +91,8 @@ void device_pm_sleep_init(struct device *dev)
{
dev->power.is_prepared = false;
dev->power.is_suspended = false;
+ dev->power.is_noirq_suspended = false;
+ dev->power.is_late_suspended = false;
init_completion(&dev->power.completion);
complete_all(&dev->power.completion);
dev->power.wakeup = NULL;
@@ -479,6 +481,9 @@ static int device_resume_noirq(struct device *dev, pm_message_t state)
if (dev->power.syscore)
goto Out;
+ if (!dev->power.is_noirq_suspended)
+ goto Out;
+
if (dev->pm_domain) {
info = "noirq power domain ";
callback = pm_noirq_op(&dev->pm_domain->ops, state);
@@ -499,6 +504,7 @@ static int device_resume_noirq(struct device *dev, pm_message_t state)
}
error = dpm_run_callback(callback, dev, state, info);
+ dev->power.is_noirq_suspended = false;
Out:
TRACE_RESUME(error);
@@ -561,6 +567,9 @@ static int device_resume_early(struct device *dev, pm_message_t state)
if (dev->power.syscore)
goto Out;
+ if (!dev->power.is_late_suspended)
+ goto Out;
+
if (dev->pm_domain) {
info = "early power domain ";
callback = pm_late_early_op(&dev->pm_domain->ops, state);
@@ -581,6 +590,7 @@ static int device_resume_early(struct device *dev, pm_message_t state)
}
error = dpm_run_callback(callback, dev, state, info);
+ dev->power.is_late_suspended = false;
Out:
TRACE_RESUME(error);
@@ -917,6 +927,7 @@ static int device_suspend_noirq(struct device *dev, pm_message_t state)
{
pm_callback_t callback = NULL;
char *info = NULL;
+ int error;
if (dev->power.syscore)
return 0;
@@ -940,7 +951,11 @@ static int device_suspend_noirq(struct device *dev, pm_message_t state)
callback = pm_noirq_op(dev->driver->pm, state);
}
- return dpm_run_callback(callback, dev, state, info);
+ error = dpm_run_callback(callback, dev, state, info);
+ if (!error)
+ dev->power.is_noirq_suspended = true;
+
+ return error;
}
/**
@@ -1003,6 +1018,7 @@ static int device_suspend_late(struct device *dev, pm_message_t state)
{
pm_callback_t callback = NULL;
char *info = NULL;
+ int error;
__pm_runtime_disable(dev, false);
@@ -1028,7 +1044,11 @@ static int device_suspend_late(struct device *dev, pm_message_t state)
callback = pm_late_early_op(dev->driver->pm, state);
}
- return dpm_run_callback(callback, dev, state, info);
+ error = dpm_run_callback(callback, dev, state, info);
+ if (!error)
+ dev->power.is_late_suspended = true;
+
+ return error;
}
/**
diff --git a/include/linux/pm.h b/include/linux/pm.h
index 8c6583a..f23a4f1 100644
--- a/include/linux/pm.h
+++ b/include/linux/pm.h
@@ -542,6 +542,8 @@ struct dev_pm_info {
unsigned int async_suspend:1;
bool is_prepared:1; /* Owned by the PM core */
bool is_suspended:1; /* Ditto */
+ bool is_noirq_suspended:1;
+ bool is_late_suspended:1;
bool ignore_children:1;
bool early_init:1; /* Owned by the PM core */
spinlock_t lock;
--
1.9.rc0
In analogy with commits 5af84b82701a and 97df8c12995, using
asynchronous threads can improve the overall resume_noirq time
significantly.
One typical case is:
In resume_noirq phase and for the PCI devices, the function
pci_pm_resume_noirq() will be called, and there is one d3_delay
(10ms) at least.
With the way of asynchronous threads, we just need wait d3_delay
time once in parallel for each calling, which saves much time to
resume quickly.
Signed-off-by: Chuansheng Liu <[email protected]>
---
drivers/base/power/main.c | 66 +++++++++++++++++++++++++++++++++++------------
1 file changed, 50 insertions(+), 16 deletions(-)
diff --git a/drivers/base/power/main.c b/drivers/base/power/main.c
index 00c53eb..ea3f1d2 100644
--- a/drivers/base/power/main.c
+++ b/drivers/base/power/main.c
@@ -469,7 +469,7 @@ static void dpm_watchdog_clear(struct dpm_watchdog *wd)
* The driver of @dev will not receive interrupts while this function is being
* executed.
*/
-static int device_resume_noirq(struct device *dev, pm_message_t state)
+static int device_resume_noirq(struct device *dev, pm_message_t state, bool async)
{
pm_callback_t callback = NULL;
char *info = NULL;
@@ -484,6 +484,8 @@ static int device_resume_noirq(struct device *dev, pm_message_t state)
if (!dev->power.is_noirq_suspended)
goto Out;
+ dpm_wait(dev->parent, async);
+
if (dev->pm_domain) {
info = "noirq power domain ";
callback = pm_noirq_op(&dev->pm_domain->ops, state);
@@ -507,10 +509,29 @@ static int device_resume_noirq(struct device *dev, pm_message_t state)
dev->power.is_noirq_suspended = false;
Out:
+ complete_all(&dev->power.completion);
TRACE_RESUME(error);
return error;
}
+static bool is_async(struct device *dev)
+{
+ return dev->power.async_suspend && pm_async_enabled
+ && !pm_trace_is_enabled();
+}
+
+static void async_resume_noirq(void *data, async_cookie_t cookie)
+{
+ struct device *dev = (struct device *)data;
+ int error;
+
+ error = device_resume_noirq(dev, pm_transition, true);
+ if (error)
+ pm_dev_err(dev, pm_transition, " async", error);
+
+ put_device(dev);
+}
+
/**
* dpm_resume_noirq - Execute "noirq resume" callbacks for all devices.
* @state: PM transition of the system being carried out.
@@ -520,29 +541,48 @@ static int device_resume_noirq(struct device *dev, pm_message_t state)
*/
static void dpm_resume_noirq(pm_message_t state)
{
+ struct device *dev;
ktime_t starttime = ktime_get();
mutex_lock(&dpm_list_mtx);
- while (!list_empty(&dpm_noirq_list)) {
- struct device *dev = to_device(dpm_noirq_list.next);
- int error;
+ pm_transition = state;
+
+ /*
+ * Advanced the async threads upfront,
+ * in case the starting of async threads is
+ * delayed by non-async resuming devices.
+ */
+ list_for_each_entry(dev, &dpm_noirq_list, power.entry) {
+ reinit_completion(&dev->power.completion);
+ if (is_async(dev)) {
+ get_device(dev);
+ async_schedule(async_resume_noirq, dev);
+ }
+ }
+ while (!list_empty(&dpm_noirq_list)) {
+ dev = to_device(dpm_noirq_list.next);
get_device(dev);
list_move_tail(&dev->power.entry, &dpm_late_early_list);
mutex_unlock(&dpm_list_mtx);
- error = device_resume_noirq(dev, state);
- if (error) {
- suspend_stats.failed_resume_noirq++;
- dpm_save_failed_step(SUSPEND_RESUME_NOIRQ);
- dpm_save_failed_dev(dev_name(dev));
- pm_dev_err(dev, state, " noirq", error);
+ if (!is_async(dev)) {
+ int error;
+
+ error = device_resume_noirq(dev, state, false);
+ if (error) {
+ suspend_stats.failed_resume_noirq++;
+ dpm_save_failed_step(SUSPEND_RESUME_NOIRQ);
+ dpm_save_failed_dev(dev_name(dev));
+ pm_dev_err(dev, state, " noirq", error);
+ }
}
mutex_lock(&dpm_list_mtx);
put_device(dev);
}
mutex_unlock(&dpm_list_mtx);
+ async_synchronize_full();
dpm_show_time(starttime, state, "noirq");
resume_device_irqs();
cpuidle_resume();
@@ -742,12 +782,6 @@ static void async_resume(void *data, async_cookie_t cookie)
put_device(dev);
}
-static bool is_async(struct device *dev)
-{
- return dev->power.async_suspend && pm_async_enabled
- && !pm_trace_is_enabled();
-}
-
/**
* dpm_resume - Execute "resume" callbacks for non-sysdev devices.
* @state: PM transition of the system being carried out.
--
1.9.rc0
In analogy with commits 5af84b82701a and 97df8c12995, using
asynchronous threads can improve the overall resume_early
time significantly.
This patch is for resume_early phase.
Signed-off-by: Chuansheng Liu <[email protected]>
---
drivers/base/power/main.c | 55 +++++++++++++++++++++++++++++++++++++----------
1 file changed, 44 insertions(+), 11 deletions(-)
diff --git a/drivers/base/power/main.c b/drivers/base/power/main.c
index ea3f1d2..6d41165 100644
--- a/drivers/base/power/main.c
+++ b/drivers/base/power/main.c
@@ -595,7 +595,7 @@ static void dpm_resume_noirq(pm_message_t state)
*
* Runtime PM is disabled for @dev while this function is being executed.
*/
-static int device_resume_early(struct device *dev, pm_message_t state)
+static int device_resume_early(struct device *dev, pm_message_t state, bool async)
{
pm_callback_t callback = NULL;
char *info = NULL;
@@ -610,6 +610,8 @@ static int device_resume_early(struct device *dev, pm_message_t state)
if (!dev->power.is_late_suspended)
goto Out;
+ dpm_wait(dev->parent, async);
+
if (dev->pm_domain) {
info = "early power domain ";
callback = pm_late_early_op(&dev->pm_domain->ops, state);
@@ -636,38 +638,69 @@ static int device_resume_early(struct device *dev, pm_message_t state)
TRACE_RESUME(error);
pm_runtime_enable(dev);
+ complete_all(&dev->power.completion);
return error;
}
+static void async_resume_early(void *data, async_cookie_t cookie)
+{
+ struct device *dev = (struct device *)data;
+ int error;
+
+ error = device_resume_early(dev, pm_transition, true);
+ if (error)
+ pm_dev_err(dev, pm_transition, " async", error);
+
+ put_device(dev);
+}
+
/**
* dpm_resume_early - Execute "early resume" callbacks for all devices.
* @state: PM transition of the system being carried out.
*/
static void dpm_resume_early(pm_message_t state)
{
+ struct device *dev;
ktime_t starttime = ktime_get();
mutex_lock(&dpm_list_mtx);
- while (!list_empty(&dpm_late_early_list)) {
- struct device *dev = to_device(dpm_late_early_list.next);
- int error;
+ pm_transition = state;
+
+ /*
+ * Advanced the async threads upfront,
+ * in case the starting of async threads is
+ * delayed by non-async resuming devices.
+ */
+ list_for_each_entry(dev, &dpm_late_early_list, power.entry) {
+ reinit_completion(&dev->power.completion);
+ if (is_async(dev)) {
+ get_device(dev);
+ async_schedule(async_resume_early, dev);
+ }
+ }
+ while (!list_empty(&dpm_late_early_list)) {
+ dev = to_device(dpm_late_early_list.next);
get_device(dev);
list_move_tail(&dev->power.entry, &dpm_suspended_list);
mutex_unlock(&dpm_list_mtx);
- error = device_resume_early(dev, state);
- if (error) {
- suspend_stats.failed_resume_early++;
- dpm_save_failed_step(SUSPEND_RESUME_EARLY);
- dpm_save_failed_dev(dev_name(dev));
- pm_dev_err(dev, state, " early", error);
- }
+ if (!is_async(dev)) {
+ int error;
+ error = device_resume_early(dev, state, false);
+ if (error) {
+ suspend_stats.failed_resume_early++;
+ dpm_save_failed_step(SUSPEND_RESUME_EARLY);
+ dpm_save_failed_dev(dev_name(dev));
+ pm_dev_err(dev, state, " early", error);
+ }
+ }
mutex_lock(&dpm_list_mtx);
put_device(dev);
}
mutex_unlock(&dpm_list_mtx);
+ async_synchronize_full();
dpm_show_time(starttime, state, "early");
}
--
1.9.rc0
In analogy with commits 5af84b82701a and 97df8c12995, using
asynchronous threads can improve the overall suspend_noirq
time significantly.
This patch is for suspend_noirq phase.
Signed-off-by: Chuansheng Liu <[email protected]>
---
drivers/base/power/main.c | 68 +++++++++++++++++++++++++++++++++++++++--------
1 file changed, 57 insertions(+), 11 deletions(-)
diff --git a/drivers/base/power/main.c b/drivers/base/power/main.c
index 6d41165..9335b32 100644
--- a/drivers/base/power/main.c
+++ b/drivers/base/power/main.c
@@ -990,14 +990,24 @@ static pm_message_t resume_event(pm_message_t sleep_state)
* The driver of @dev will not receive interrupts while this function is being
* executed.
*/
-static int device_suspend_noirq(struct device *dev, pm_message_t state)
+static int __device_suspend_noirq(struct device *dev, pm_message_t state, bool async)
{
pm_callback_t callback = NULL;
char *info = NULL;
- int error;
+ int error = 0;
+
+ if (async_error)
+ goto Complete;
+
+ if (pm_wakeup_pending()) {
+ async_error = -EBUSY;
+ goto Complete;
+ }
if (dev->power.syscore)
- return 0;
+ goto Complete;
+
+ dpm_wait_for_children(dev, async);
if (dev->pm_domain) {
info = "noirq power domain ";
@@ -1021,10 +1031,40 @@ static int device_suspend_noirq(struct device *dev, pm_message_t state)
error = dpm_run_callback(callback, dev, state, info);
if (!error)
dev->power.is_noirq_suspended = true;
+ else
+ async_error = error;
+Complete:
+ complete_all(&dev->power.completion);
return error;
}
+static void async_suspend_noirq(void *data, async_cookie_t cookie)
+{
+ struct device *dev = (struct device *)data;
+ int error;
+
+ error = __device_suspend_noirq(dev, pm_transition, true);
+ if (error) {
+ dpm_save_failed_dev(dev_name(dev));
+ pm_dev_err(dev, pm_transition, " async", error);
+ }
+
+ put_device(dev);
+}
+
+static int device_suspend_noirq(struct device *dev)
+{
+ reinit_completion(&dev->power.completion);
+
+ if (pm_async_enabled && dev->power.async_suspend) {
+ get_device(dev);
+ async_schedule(async_suspend_noirq, dev);
+ return 0;
+ }
+ return __device_suspend_noirq(dev, pm_transition, false);
+}
+
/**
* dpm_suspend_noirq - Execute "noirq suspend" callbacks for all devices.
* @state: PM transition of the system being carried out.
@@ -1040,19 +1080,20 @@ static int dpm_suspend_noirq(pm_message_t state)
cpuidle_pause();
suspend_device_irqs();
mutex_lock(&dpm_list_mtx);
+ pm_transition = state;
+ async_error = 0;
+
while (!list_empty(&dpm_late_early_list)) {
struct device *dev = to_device(dpm_late_early_list.prev);
get_device(dev);
mutex_unlock(&dpm_list_mtx);
- error = device_suspend_noirq(dev, state);
+ error = device_suspend_noirq(dev);
mutex_lock(&dpm_list_mtx);
if (error) {
pm_dev_err(dev, state, " noirq", error);
- suspend_stats.failed_suspend_noirq++;
- dpm_save_failed_step(SUSPEND_SUSPEND_NOIRQ);
dpm_save_failed_dev(dev_name(dev));
put_device(dev);
break;
@@ -1061,16 +1102,21 @@ static int dpm_suspend_noirq(pm_message_t state)
list_move(&dev->power.entry, &dpm_noirq_list);
put_device(dev);
- if (pm_wakeup_pending()) {
- error = -EBUSY;
+ if (async_error)
break;
- }
}
mutex_unlock(&dpm_list_mtx);
- if (error)
+ async_synchronize_full();
+ if (!error)
+ error = async_error;
+
+ if (error) {
+ suspend_stats.failed_suspend_noirq++;
+ dpm_save_failed_step(SUSPEND_SUSPEND_NOIRQ);
dpm_resume_noirq(resume_event(state));
- else
+ } else {
dpm_show_time(starttime, state, "noirq");
+ }
return error;
}
--
1.9.rc0
In analogy with commits 5af84b82701a and 97df8c12995, using
asynchronous threads can improve the overall suspend_late
time significantly.
This patch is for suspend_late phase.
Signed-off-by: Chuansheng Liu <[email protected]>
---
drivers/base/power/main.c | 66 ++++++++++++++++++++++++++++++++++++++---------
1 file changed, 54 insertions(+), 12 deletions(-)
diff --git a/drivers/base/power/main.c b/drivers/base/power/main.c
index 9335b32..42355e4 100644
--- a/drivers/base/power/main.c
+++ b/drivers/base/power/main.c
@@ -1127,16 +1127,26 @@ static int dpm_suspend_noirq(pm_message_t state)
*
* Runtime PM is disabled for @dev while this function is being executed.
*/
-static int device_suspend_late(struct device *dev, pm_message_t state)
+static int __device_suspend_late(struct device *dev, pm_message_t state, bool async)
{
pm_callback_t callback = NULL;
char *info = NULL;
- int error;
+ int error = 0;
__pm_runtime_disable(dev, false);
+ if (async_error)
+ goto Complete;
+
+ if (pm_wakeup_pending()) {
+ async_error = -EBUSY;
+ goto Complete;
+ }
+
if (dev->power.syscore)
- return 0;
+ goto Complete;
+
+ dpm_wait_for_children(dev, async);
if (dev->pm_domain) {
info = "late power domain ";
@@ -1160,10 +1170,40 @@ static int device_suspend_late(struct device *dev, pm_message_t state)
error = dpm_run_callback(callback, dev, state, info);
if (!error)
dev->power.is_late_suspended = true;
+ else
+ async_error = error;
+Complete:
+ complete_all(&dev->power.completion);
return error;
}
+static void async_suspend_late(void *data, async_cookie_t cookie)
+{
+ struct device *dev = (struct device *)data;
+ int error;
+
+ error = __device_suspend_late(dev, pm_transition, true);
+ if (error) {
+ dpm_save_failed_dev(dev_name(dev));
+ pm_dev_err(dev, pm_transition, " async", error);
+ }
+ put_device(dev);
+}
+
+static int device_suspend_late(struct device *dev)
+{
+ reinit_completion(&dev->power.completion);
+
+ if (pm_async_enabled && dev->power.async_suspend) {
+ get_device(dev);
+ async_schedule(async_suspend_late, dev);
+ return 0;
+ }
+
+ return __device_suspend_late(dev, pm_transition, false);
+}
+
/**
* dpm_suspend_late - Execute "late suspend" callbacks for all devices.
* @state: PM transition of the system being carried out.
@@ -1174,19 +1214,20 @@ static int dpm_suspend_late(pm_message_t state)
int error = 0;
mutex_lock(&dpm_list_mtx);
+ pm_transition = state;
+ async_error = 0;
+
while (!list_empty(&dpm_suspended_list)) {
struct device *dev = to_device(dpm_suspended_list.prev);
get_device(dev);
mutex_unlock(&dpm_list_mtx);
- error = device_suspend_late(dev, state);
+ error = device_suspend_late(dev);
mutex_lock(&dpm_list_mtx);
if (error) {
pm_dev_err(dev, state, " late", error);
- suspend_stats.failed_suspend_late++;
- dpm_save_failed_step(SUSPEND_SUSPEND_LATE);
dpm_save_failed_dev(dev_name(dev));
put_device(dev);
break;
@@ -1195,17 +1236,18 @@ static int dpm_suspend_late(pm_message_t state)
list_move(&dev->power.entry, &dpm_late_early_list);
put_device(dev);
- if (pm_wakeup_pending()) {
- error = -EBUSY;
+ if (async_error)
break;
- }
}
mutex_unlock(&dpm_list_mtx);
- if (error)
+ async_synchronize_full();
+ if (error) {
+ suspend_stats.failed_suspend_late++;
+ dpm_save_failed_step(SUSPEND_SUSPEND_LATE);
dpm_resume_early(resume_event(state));
- else
+ } else {
dpm_show_time(starttime, state, "late");
-
+ }
return error;
}
--
1.9.rc0
Hello Rafael,
> -----Original Message-----
> From: Liu, Chuansheng
> Sent: Tuesday, February 18, 2014 10:29 AM
> To: [email protected]; [email protected]; Brown, Len;
> [email protected]
> Cc: [email protected]; [email protected]; Li, Zhuangzhi; Liu,
> Chuansheng
> Subject: [PATCH v4 0/5] Enabling the asynchronous threads for other phases
>
> Hello,
>
> This patch series are for enabling the asynchronous threads for the phases
> resume_noirq, resume_early, suspend_noirq and suspend_late.
>
After dig more about async threads, I found dpm_prepare/complete phases need
them also.
dpm_prepare()
-- > call pci_pm_prepare() one by one
-- > call pm_runtime_resume() one by one
-- > call pci_set_power_state to D3HOT -- > D0
It will cause much time delaying due to d3_delay.
And I made one draft patch to implement the asynchronous threads for dpm_prepare(),
it is really save much time.
If you like it, I can prepare another series patches for dpm_prepare() and dpm_complete(),
thanks.
Best Regards,
On Friday, February 21, 2014 01:38:45 AM Liu, Chuansheng wrote:
> Hello Rafael,
>
> > -----Original Message-----
> > From: Liu, Chuansheng
> > Sent: Tuesday, February 18, 2014 10:29 AM
> > To: [email protected]; [email protected]; Brown, Len;
> > [email protected]
> > Cc: [email protected]; [email protected]; Li, Zhuangzhi; Liu,
> > Chuansheng
> > Subject: [PATCH v4 0/5] Enabling the asynchronous threads for other phases
> >
> > Hello,
> >
> > This patch series are for enabling the asynchronous threads for the phases
> > resume_noirq, resume_early, suspend_noirq and suspend_late.
> >
>
> After dig more about async threads, I found dpm_prepare/complete phases need
> them also.
>
> dpm_prepare()
> -- > call pci_pm_prepare() one by one
> -- > call pm_runtime_resume() one by one
> -- > call pci_set_power_state to D3HOT -- > D0
>
> It will cause much time delaying due to d3_delay.
> And I made one draft patch to implement the asynchronous threads for dpm_prepare(),
> it is really save much time.
>
> If you like it, I can prepare another series patches for dpm_prepare() and dpm_complete(),
> thanks.
I think a better solution here would be to move the pm_runtime_resume()
out of the .prepare() callback for PCI. We may be doing that in the
process of suspend optimization anyway, so I'd wait with making those
phases async.
Thanks!
--
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.