Hi Geert,
Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> writes:
> Hi Ulf, Rafael,
>
> On Tue, May 17, 2016 at 1:41 PM, Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote:
>> When the pm_runtime_force_suspend|resume() helpers were invented, we still
>> had CONFIG_PM_RUNTIME and CONFIG_PM_SLEEP as separate Kconfig options.
>>
>> To make sure these helpers worked for all combinations and without
>> introducing too much of complexity, the device was always resumed in
>> pm_runtime_force_resume().
>>
>> More precisely, when CONFIG_PM_SLEEP was set and CONFIG_PM_RUNTIME was
>> unset, we needed to resume the device as the subsystem/driver couldn't
>> rely on using runtime PM to do it.
>>
>> As the CONFIG_PM_RUNTIME option was merged into CONFIG_PM a while ago, it
>> removed this combination, of using CONFIG_PM_SLEEP without the earlier
>> CONFIG_PM_RUNTIME.
>>
>> For this reason we can now rely on the subsystem/driver to use runtime PM
>> to resume the device, instead of forcing that to be done in all cases. In
>> other words, let's defer this to a later point when it's actually needed.
>>
>> Signed-off-by: Ulf Hansson <ulf.hansson@xxxxxxxxxx>
>> ---
>> drivers/base/power/runtime.c | 11 +++++++++++
>> 1 file changed, 11 insertions(+)
>>
>> diff --git a/drivers/base/power/runtime.c b/drivers/base/power/runtime.c
>> index 09e4eb1..1db7b46 100644
>> --- a/drivers/base/power/runtime.c
>> +++ b/drivers/base/power/runtime.c
>> @@ -1509,6 +1509,17 @@ int pm_runtime_force_resume(struct device *dev)
>> if (!pm_runtime_status_suspended(dev))
>> goto out;
>>
>> + /*
>> + * The PM core increases the runtime PM usage count in the system PM
>> + * prepare phase. If the count is greather than 1 at this point, someone
>> + * else has also increased it. In that case, invoke the runtime resume
>> + * callback for the device as that is likely what is expected. In other
>> + * case we trust the subsystem/driver to runtime resume the device when
>> + * it's actually needed.
>> + */
>> + if (atomic_read(&dev->power.usage_count) < 2)
>> + goto out;
>> +
>> ret = pm_runtime_set_active(dev);
>> if (ret)
>> goto out;
>
> This patch (commit eb13a0a1b6d5d5c2 in pm/linux-next) breaks resume on
> sh73a0/kzm9g and r8a73a4/ape6evm. On these boards, the Ethernet controller is a
> child of a local bus (bsc), whose clock (zb) is controlled through pm_clk and
> simple-pm-bus, cfr.
>
> arch/arm/boot/dts/r8a73a4-ape6evm.dts
> arch/arm/boot/dts/r8a73a4.dtsi
> arch/arm/boot/dts/sh73a0-kzm9g.dts
> arch/arm/boot/dts/sh73a0.dtsi
>
> During resume, the bus clock is not enabled, causing an imprecise abort
> when accessing the Ethernet controller's registers.
I have a hunch (without too much digging) that this may be an odd
interaction with the direct_complete stuff since the simple-pm-bus has
no callbacks.
For kicks, could you add something like the hack below (untested) which
will avoid the direct_complete path, and at least help indicate if
that path is worth investigating further.
Thanks,
Kevin
diff --git a/drivers/bus/simple-pm-bus.c b/drivers/bus/simple-pm-bus.c
index c5eb46cbf388..63b95fb21510 100644
--- a/drivers/bus/simple-pm-bus.c
+++ b/drivers/bus/simple-pm-bus.c
@@ -36,6 +36,15 @@ static int simple_pm_bus_remove(struct platform_device *pdev)
return 0;
}
+static int simple_pm_bus_prepare(struct device *dev)
+{
+ return pm_generic_prepare(dev);
+}
+
+static const struct dev_pm_ops simple_pm_bus_ops = {
+ .prepare = simple_pm_bus_prepare,
+};
+
static const struct of_device_id simple_pm_bus_of_match[] = {
{ .compatible = "simple-pm-bus", },
{ /* sentinel */ }
@@ -48,6 +57,7 @@ static struct platform_driver simple_pm_bus_driver = {
.driver = {
.name = "simple-pm-bus",
.of_match_table = simple_pm_bus_of_match,
+ .pm = &simple_pm_bus_ops,
},
};
