PM: sleep: call devfreq_suspend/resume and cpufreq_suspend/resume in pairs.
Subsystems (4)
PATCH 1/1 PM: sleep: call devfreq_suspend/resume and cpufreq_suspend/resume in pairs.
From: Vincent Wang <[email protected]>
If dpm_prepare() fails in dpm_suspend_start(), dpm_suspend() can't be called. And then, devfreq_suspend() and cpufreq_suspend() will not be called in the suspend flow.
But in the resiume flow, devfreq_resume() and cpufreq_resume() will be called.
This patch will ensure that devfreq_suspend/devfreq_resume and cpufreq_suspend/cpufreq_resume are called in pairs.
--- drivers/base/power/main.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/base/power/main.c +3 -3
@@ -1866,9 +1866,6 @@ int dpm_suspend(pm_message_t state)trace_suspend_resume (TPS ("dpm_suspend" ), state.event, true);might_sleep (); -devfreq_suspend (); -cpufreq_suspend (); -mutex_lock (&dpm_list_mtx); pm_transition = state; async_error =0 ;
@@ -1988,6 +1985,9 @@ int dpm_prepare(pm_message_t state)trace_suspend_resume (TPS ("dpm_prepare" ), state.event, true);might_sleep (); +devfreq_suspend (); +cpufreq_suspend (); +/* * Give a chance for the known devices to complete their probes, before * disable probing of devices. This sync point is important at least
Signature · 1 line
2.20.1
Rafael J. Wysocki
maintainerQuoted 5 lines · Chunyan Zhang
From: Vincent Wang <[email protected]>
If
dpm_prepare()fails indpm_suspend_start(),dpm_suspend()can't be called.
That's correct.
Quoted 2 lines
And then,
devfreq_suspend()andcpufreq_suspend()will not be called in the suspend flow.
Right.
Quoted 2 lines
But in the resiume flow,
devfreq_resume()andcpufreq_resume()will be called.
Right, and they are expected to cope with the situation.
Quoted 2 lines
This patch will ensure that devfreq_suspend/devfreq_resume and cpufreq_suspend/cpufreq_resume are called in pairs.
So why is it better to do this than to make devfreq_resume() meet the expectations?
Quoted 34 lines
Signed-off-by Vincent Wang <[email protected]>Signed-off-by Samer Xie <[email protected]>Signed-off-by Chunyan Zhang <[email protected]>--- drivers/base/power/main.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)drivers/base/power/main.c +3
-4@@ -1866,9 +1866,6 @@ int dpm_suspend(pm_message_t state)trace_suspend_resume (TPS ("dpm_suspend" ), state.event, true);might_sleep (); -devfreq_suspend (); -cpufreq_suspend (); -mutex_lock (&dpm_list_mtx); pm_transition = state; async_error =0 ;@@ -1988,6 +1985,9 @@ int dpm_prepare(pm_message_t state)trace_suspend_resume (TPS ("dpm_prepare" ), state.event, true);might_sleep (); +devfreq_suspend (); +cpufreq_suspend (); +/* * Give a chance for the known devices to complete their probes, before * disable probing of devices. This sync point is important at least -- 2.20.1
Hi Rafael,
(Behalf Of Vincent Wang)
Thanks for your comments, please see my answers below.
Quoted 18 lines · Rafael J. Wysocki
Quoted 5 lines · Chunyan Zhang
From: Vincent Wang <[email protected]>
If
dpm_prepare()fails indpm_suspend_start(),dpm_suspend()can't be called.That's correct.
Quoted 2 lines
And then,
devfreq_suspend()andcpufreq_suspend()will not be called in the suspend flow.Right.
Quoted 2 lines
But in the resiume flow,
devfreq_resume()andcpufreq_resume()will be called.
Quoted 7 lines
Right, and they are expected to cope with the situation.
Quoted 2 lines
This patch will ensure that devfreq_suspend/devfreq_resume and cpufreq_suspend/cpufreq_resume are called in pairs.
So why is it better to do this than to make
devfreq_resume()meet the expectations?
Yes,we found an issue with cpufreq schedutil governor on kernel4.14, and I think the issue should haven't been changed on the latest version of kernel.
In the function dpm_suspend_start(), dpm_suspend() would not be exceuted if return error from device_prepare() [1]. So cpufreq_cpufreq() will not be called, then cpufreq_remove_update_util_hook() will not be called either, and so cpufreq_update_util_data will not be NULL.
In the dpm resume flow, sugov_start() will be called, in which sg_cpu.update_util will be set to 0.
And since cpufreq_update_util_data is not NULL, data->func will not be set and is still NULL which actually is sg_cpu.update_util.
void cpufreq_add_update_util_hook(int cpu, struct update_util_data *data,
void (*func)(struct update_util_data *data, u64 time,
unsigned int flags))
{
[...]
if (WARN_ON(per_cpu(cpufreq_update_util_data, cpu)))
return;
data->func = func;[...]
}
When cpufreq_update_util() is called by scheduler, there will be a NULL pointer issue.
[1] https://elixir.bootlin.com/linux/v5.7-rc1/source/drivers/base/power/main.c#L2052
Quoted 35 lines
Quoted 35 lines
Signed-off-by Vincent Wang <[email protected]>Signed-off-by Samer Xie <[email protected]>Signed-off-by Chunyan Zhang <[email protected]>--- drivers/base/power/main.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)drivers/base/power/main.c +3
-4@@ -1866,9 +1866,6 @@ int dpm_suspend(pm_message_t state)trace_suspend_resume (TPS ("dpm_suspend" ), state.event, true);might_sleep (); -devfreq_suspend (); -cpufreq_suspend (); -mutex_lock (&dpm_list_mtx); pm_transition = state; async_error =0 ;@@ -1988,6 +1985,9 @@ int dpm_prepare(pm_message_t state)trace_suspend_resume (TPS ("dpm_prepare" ), state.event, true);might_sleep (); +devfreq_suspend (); +cpufreq_suspend (); +/* * Give a chance for the known devices to complete their probes, before * disable probing of devices. This sync point is important at least -- 2.20.1
Rafael J. Wysocki
maintainerQuoted 34 lines · Chunyan Zhang
Hi Rafael,
(Behalf Of Vincent Wang)
Thanks for your comments, please see my answers below.
Quoted 18 lines · Rafael J. Wysocki
Quoted 5 lines · Chunyan Zhang
From: Vincent Wang <[email protected]>
If
dpm_prepare()fails indpm_suspend_start(),dpm_suspend()can't be called.That's correct.
Quoted 2 lines
And then,
devfreq_suspend()andcpufreq_suspend()will not be called in the suspend flow.Right.
Quoted 2 lines
But in the resiume flow,
devfreq_resume()andcpufreq_resume()will be called.Quoted 7 lines
Right, and they are expected to cope with the situation.
Quoted 2 lines
This patch will ensure that devfreq_suspend/devfreq_resume and cpufreq_suspend/cpufreq_resume are called in pairs.
So why is it better to do this than to make
devfreq_resume()meet the expectations?
Quoted 7 lines
Yes,we found an issue with cpufreq schedutil governor on kernel4.14, and I think the issue should haven't been changed on the latest version of kernel.
In the function
dpm_suspend_start(),dpm_suspend()would not be exceuted if return error fromdevice_prepare()[1]. Socpufreq_cpufreq()will not be called,
I guess you mean cpufreq_suspend().
That should be OK .
Quoted 5 lines
then
cpufreq_remove_update_util_hook()will not be called either, and socpufreq_update_util_datawill not be NULL.In the dpm resume flow,
sugov_start()will be called, in which sg_cpu.update_util will be set to 0.
Which code patch does this?
Surely not cpufreq_resume(), because that checks cpufreq_suspended which cannot be set if cpufreq_suspend() has not been called (because it is the only place setting cpufreq_suspended).
Quoted 51 lines · Rafael J. Wysocki
Quoted 34 lines · Chunyan Zhang
Hi Rafael,
(Behalf Of Vincent Wang)
Thanks for your comments, please see my answers below.
Quoted 18 lines · Rafael J. Wysocki
Quoted 5 lines · Chunyan Zhang
From: Vincent Wang <[email protected]>
If
dpm_prepare()fails indpm_suspend_start(),dpm_suspend()can't be called.That's correct.
Quoted 2 lines
And then,
devfreq_suspend()andcpufreq_suspend()will not be called in the suspend flow.Right.
Quoted 2 lines
But in the resiume flow,
devfreq_resume()andcpufreq_resume()will be called.Quoted 7 lines
Right, and they are expected to cope with the situation.
Quoted 2 lines
This patch will ensure that devfreq_suspend/devfreq_resume and cpufreq_suspend/cpufreq_resume are called in pairs.
So why is it better to do this than to make
devfreq_resume()meet the expectations?Quoted 7 lines
Yes,we found an issue with cpufreq schedutil governor on kernel4.14, and I think the issue should haven't been changed on the latest version of kernel.
In the function
dpm_suspend_start(),dpm_suspend()would not be exceuted if return error fromdevice_prepare()[1]. Socpufreq_cpufreq()will not be called,I guess you mean
cpufreq_suspend().That should be OK .
Quoted 3 lines
then
cpufreq_remove_update_util_hook()will not be called either, and socpufreq_update_util_datawill not be NULL.
Quoted 8 lines
Quoted 2 lines
In the dpm resume flow,
sugov_start()will be called, in which sg_cpu.update_util will be set to 0.Which code patch does this?
Surely not
cpufreq_resume(), because that checkscpufreq_suspendedwhich cannot be set ifcpufreq_suspend()has not been called (because it is the only place settingcpufreq_suspended).
Right, I saw that, then there's no issue indeed. I just need to backport the patch which added checking of cpufreq_suspended to cpufreq_resume.
Thanks for your review!