[PATCH 3/3] serial: Check result of console_unregister()

[Andrey Smirnov]

On Wed, Apr 11, 2018 at 1:34 AM, Sascha Hauer <s.hauer at pengutronix.de> wrote:
> Hi Andrey,
>
> On Mon, Apr 09, 2018 at 09:00:52AM -0700, Andrey Smirnov wrote:
>> On Tue, Apr 3, 2018 at 12:04 AM, Sascha Hauer <s.hauer at pengutronix.de> wrote:
>> > On Mon, Mar 26, 2018 at 06:09:15AM -0700, Andrey Smirnov wrote:
>> >> In order to allow 'serdev' devices to prevent parent console device
>> >> removal and correspondign memory deallocation add code to all serial
>> >> driver to check result of console_unregister() and bail out early if
>> >> it is unsuccessful.
>> >>
>> >> One example of a use-case for this would be a reset handler relying on
>> >> a serdev device for transport. Without this patch underlying console
>> >> device would be removed and de-allocated before reset handler is even
>> >> run thus leading to unpredictable behaviour and crashes.
>> >
>> > Can't we make this sure at driver core level?
>>
>> I need to be able to prevent serial driver's "remove" function from
>> ever executing to prevent any de-initialization/memory freeing from
>> happening. The simplest way to solve this in driver core that comes to
>> my mind is implementing simple reference counting API that children
>> could use to force driver core to bail out on removing parents if they
>> are still in use. Does that sound like a reasonable idea?
>>
>> > So if a device decides not
>> > to return -EBUSY in the remove callback then the parent devices won't be
>> > removed?
>>
>> Remove callback currently returns void, we could change it to return
>> int and use it to implement a sort of implicit refcounting, but doing
>> so would result in quite a bit of code churn since all of the current
>> drivers would have to be converted to return int in their .remove
>> callbacks. Would you rather I do this or explicit refcounting?
>
> Normally it helps looking at the Linux kernel to see how a problem is
> solved there. Not so this time it seems. Linux distinguishes between
> "remove" and "shutdown". "shutdown" is what we want during barebox
> shutdown. I found a Linux driver that is similar to your situation: It
> registers a restart_handler while being a i2c device itself. There seems
> to be no way to prevent a device from being shutdown, it's only that the
> i2c bus drivers simply do not implement it.
>
> Where do we go from here? I think reference counting is a bit over the
> top.
>
> At the moment I would opt for a *very* simple solution: Let's drop the
> call to console_unregister() and the freeing of resources entirely as
> it gives us nothing. The only console driver I can see where removing
> is valid is drivers/usb/gadget/u_serial.c and you won't use this for
> restarting a SoC ;)
>

OK, makes sense and works for me. Will do in v2.

Thanks,
Andrey Smirnov