Re: [PATCH 2/4 v5] memstick: Prevent memstick host from getting runtime suspended during card detection

[Ulf Hansson]

On 30 October 2018 at 16:23, Kai Heng Feng <kai.heng.feng@xxxxxxxxxxxxx> wrote:
>
>
>> On Oct 30, 2018, at 21:03, Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote:
>>
>> On 29 October 2018 at 17:31, Kai Heng Feng <kai.heng.feng@xxxxxxxxxxxxx> wrote:
>>>
>>>
>>>> On Oct 29, 2018, at 20:25, Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote:
>>>>
>>>> On 24 October 2018 at 10:49, Kai-Heng Feng <kai.heng.feng@xxxxxxxxxxxxx> wrote:
>>>>> We can use MEMSTICK_POWER_{ON,OFF} along with pm_runtime_{get,put}
>>>>> helpers to let memstick host support runtime pm.
>>>>>
>>>>> There's a small window between memstick_detect_change() and its queued
>>>>> work, memstick_check(). In this window the rpm count may go down to zero
>>>>> before the memstick host powers on, so the host can be inadvertently
>>>>> suspended.
>>>>>
>>>>> Increment rpm count before calling memstick_check(), and decrement rpm
>>>>> count afterward, as now we are sure the memstick host should be
>>>>> suspended or not.
>>>>>
>>>>> Signed-off-by: Kai-Heng Feng <kai.heng.feng@xxxxxxxxxxxxx>
>>>>> ---
>>>>> drivers/memstick/core/memstick.c | 4 ++++
>>>>> 1 file changed, 4 insertions(+)
>>>>>
>>>>> diff --git a/drivers/memstick/core/memstick.c b/drivers/memstick/core/memstick.c
>>>>> index 76382c858c35..5f16a8826401 100644
>>>>> --- a/drivers/memstick/core/memstick.c
>>>>> +++ b/drivers/memstick/core/memstick.c
>>>>> @@ -18,6 +18,7 @@
>>>>> #include <linux/delay.h>
>>>>> #include <linux/slab.h>
>>>>> #include <linux/module.h>
>>>>> +#include <linux/pm_runtime.h>
>>>>>
>>>>> #define DRIVER_NAME "memstick"
>>>>>
>>>>> @@ -209,6 +210,7 @@ static int memstick_dummy_check(struct memstick_dev *card)
>>>>> */
>>>>> void memstick_detect_change(struct memstick_host *host)
>>>>> {
>>>>> +       pm_runtime_get_noresume(host->dev.parent);
>>>>>       queue_work(workqueue, &host->media_checker);
>>>>> }
>>>>> EXPORT_SYMBOL(memstick_detect_change);
>>>>> @@ -479,6 +481,8 @@ static void memstick_check(struct work_struct *work)
>>>>>               host->set_param(host, MEMSTICK_POWER, MEMSTICK_POWER_OFF);
>>>>>
>>>>>       mutex_unlock(&host->lock);
>>>>> +
>>>>> +       pm_runtime_put(host->dev.parent);
>>>>>       dev_dbg(&host->dev, "memstick_check finished\n");
>>>>> }
>>>>>
>>>>
>>>> I am not sure this works, sorry.
>>>>
>>>> More precisely, I don't think there is a guarantee that the calls to
>>>> pm_runtime_get|put*() becomes properly balanced. In principle
>>>> memstick_detect_change() could be called, without actually causing a
>>>> new work to be scheduled if there is already such a work in the queue
>>>> (depends on the workqueue configuration). Isn't it so?
>>>
>>> You are right.
>>>
>>> We can use test_and_set_bit() or alike to properly balance pm_runtime
>>> helpers, but the most straightforward solution in my mind is to merge
>>> memstick_detect_change() and memstick_check() as one function.
>>>
>>> memstick_detect_change() itâs the only user of memstick_check() anyway.
>>
>> I suspect memstick_detect_change() is supposed to be called by host
>> drivers, when they receive some kind of notification due to a card
>> being inserted or removed. I guess that happen (at least
>> hypothetically) also from atomic (IRQ) context.
>>
>> As memstick_check() is doing hole bunch of operations, I am not sure
>> bypassing the work-queue is a good idea, if that is what you are
>> proposing.
>
> Okay, itâs better to keep it that way.
>
>>
>>>
>>> Or is there a better way in your mind?
>>
>> I don't know.
>>
>> Well, I am not sure I understand why you need to call
>> pm_runtime_get_noresume() from memstick_detect_change() in the first
>> place. Could you explain that in more detail?
>
> I guess it didnât explain it well enough in the log, let me add some detail:
> There's a small window between memstick_detect_change() and its queued
> work, memstick_check(). In this window the rpm count may go down to zero
> before the memstick host powers on, where I use
> pm_runtime_get_noresume() to increment the rpm count.
>
> memstick_check() uses some functions in rtsx_usb_ms that have
> pm_runtime_put*() so the rpm count may go down to zero, before the
> memstick host powers on.

So then, why doesn't memstick_check() early on calls
pm_runtime_get_sync() and when it has finished with probing for a
card, balance that with a call pm_runtime_put()?

Kind regards
Uffe