Re: [PATCH 2/2] PM / Domains: Avoid infinite loops in attach/detach code

[Rafael J. Wysocki]

On Tue, Jun 23, 2015 at 3:20 PM, Geert Uytterhoeven
<geert@xxxxxxxxxxxxxx> wrote:
> Hi Ulf,
>
> On Tue, Jun 23, 2015 at 2:50 PM, Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote:
>> On 22 June 2015 at 09:31, Geert Uytterhoeven <geert+renesas@xxxxxxxxx> wrote:
>>> If pm_genpd_{add,remove}_device() keeps on failing with -EAGAIN, we end
>>> up with an infinite loop in genpd_dev_pm_{at,de}tach().
>>>
>>> This may happen due to a genpd.prepared_count imbalance.  This is a bug
>>> elsewhere, but it will result in a system lock up, possibly during
>>> reboot of an otherwise functioning system.
>>>
>>> To avoid this, put a limit on the maximum number of loop iterations,
>>> including a simple back-off mechanism.  If the limit is reached, the
>>> operation will just fail.  An error message is already printed.
>>>
>>> Signed-off-by: Geert Uytterhoeven <geert+renesas@xxxxxxxxx>
>>> ---
>>>  drivers/base/power/domain.c | 16 ++++++++++++++--
>>>  1 file changed, 14 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/base/power/domain.c b/drivers/base/power/domain.c
>>> index cdd547bd67df8218..60e0309dd8dd0264 100644
>>> --- a/drivers/base/power/domain.c
>>> +++ b/drivers/base/power/domain.c
>>> @@ -6,6 +6,7 @@
>>>   * This file is released under the GPLv2.
>>>   */
>>>
>>> +#include <linux/delay.h>
>>>  #include <linux/kernel.h>
>>>  #include <linux/io.h>
>>>  #include <linux/platform_device.h>
>>> @@ -19,6 +20,9 @@
>>>  #include <linux/suspend.h>
>>>  #include <linux/export.h>
>>>
>>> +#define GENPD_RETRIES  20
>>> +#define GENPD_DELAY_US 10
>>> +
>>>  #define GENPD_DEV_CALLBACK(genpd, type, callback, dev)         \
>>>  ({                                                             \
>>>         type (*__routine)(struct device *__d);                  \
>>> @@ -2131,6 +2135,7 @@ EXPORT_SYMBOL_GPL(of_genpd_get_from_provider);
>>>  static void genpd_dev_pm_detach(struct device *dev, bool power_off)
>>>  {
>>>         struct generic_pm_domain *pd;
>>> +       unsigned int i;
>>>         int ret = 0;
>>>
>>>         pd = pm_genpd_lookup_dev(dev);
>>> @@ -2139,10 +2144,13 @@ static void genpd_dev_pm_detach(struct device *dev, bool power_off)
>>>
>>>         dev_dbg(dev, "removing from PM domain %s\n", pd->name);
>>>
>>> -       while (1) {
>>> +       for (i = 0; i < GENPD_RETRIES; i++) {
>>>                 ret = pm_genpd_remove_device(pd, dev);
>>>                 if (ret != -EAGAIN)
>>>                         break;
>>> +
>>> +               if (i > GENPD_RETRIES / 2)
>>> +                       udelay(GENPD_DELAY_US);
>>>                 cond_resched();
>>>         }
>>>
>>> @@ -2183,6 +2191,7 @@ int genpd_dev_pm_attach(struct device *dev)
>>>  {
>>>         struct of_phandle_args pd_args;
>>>         struct generic_pm_domain *pd;
>>> +       unsigned int i;
>>>         int ret;
>>>
>>>         if (!dev->of_node)
>>> @@ -2218,10 +2227,13 @@ int genpd_dev_pm_attach(struct device *dev)
>>>
>>>         dev_dbg(dev, "adding to PM domain %s\n", pd->name);
>>>
>>> -       while (1) {
>>> +       for (i = 0; i < GENPD_RETRIES; i++) {
>>>                 ret = pm_genpd_add_device(pd, dev);
>>>                 if (ret != -EAGAIN)
>>>                         break;
>>> +
>>> +               if (i > GENPD_RETRIES / 2)
>>> +                       udelay(GENPD_DELAY_US);
>>
>> In this execution path, we retry when getting -EAGAIN while believing
>> the reason to the error are only *temporary* as we are soon waiting
>> for all devices in the genpd to be system PM resumed. At least that's
>> my understanding to why we want to deal with -EAGAIN here, but I might
>> be wrong.
>>
>> In this regards, I wonder whether it could be better to re-try only a
>> few times but with a far longer interval time than a couple us. What
>> do you think?
>
> That's indeed viable. I have no idea for how long this temporary state can
> extend.

A usual approach to this kind of thing is to use exponential fallback
where you increase the delay twice with respect to the previous one
every time.

Rafael
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/