Re: [PATCH RFT 6/6] serial: sh-sci: Increment the runtime usage counter for the earlycon device

[Claudiu Beznea <claudiu.beznea@tuxon.dev>]

Hi, Geert,

On 19.12.2024 16:30, Geert Uytterhoeven wrote:
> Hi Claudiu,
> 
> On Wed, Dec 4, 2024 at 4:58 PM Claudiu <claudiu.beznea@tuxon.dev> wrote:
>> From: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com>
>>
>> In the sh-sci driver, serial ports are mapped to the sci_ports[] array,
>> with earlycon mapped at index zero.
>>
>> The uart_add_one_port() function eventually calls __device_attach(),
>> which, in turn, calls pm_request_idle(). The identified code path is as
>> follows:
>>
>> uart_add_one_port() ->
>>   serial_ctrl_register_port() ->
>>     serial_core_register_port() ->
>>       serial_core_port_device_add() ->
>>         serial_base_port_add() ->
>>           device_add() ->
>>             bus_probe_device() ->
>>               device_initial_probe() ->
>>                 __device_attach() ->
>>                   // ...
>>                   if (dev->p->dead) {
>>                     // ...
>>                   } else if (dev->driver) {
>>                     // ...
>>                   } else {
>>                     // ...
>>                     pm_request_idle(dev);
>>                     // ...
>>                   }
>>
>> The earlycon device clocks are enabled by the bootloader. However, the
>> pm_request_idle() call in __device_attach() disables the SCI port clocks
>> while earlycon is still active.
>>
>> The earlycon write function, serial_console_write(), calls
>> sci_poll_put_char() via serial_console_putchar(). If the SCI port clocks
>> are disabled, writing to earlycon may sometimes cause the SR.TDFE bit to
>> remain unset indefinitely, causing the while loop in sci_poll_put_char()
>> to never exit. On single-core SoCs, this can result in the system being
>> blocked during boot when this issue occurs.
>>
>> To resolve this, increment the runtime PM usage counter for the earlycon
>> SCI device before registering the UART port.
>>
>> Fixes: 0b0cced19ab1 ("serial: sh-sci: Add CONFIG_SERIAL_EARLYCON support")
>> Cc: stable@vger.kernel.org
>> Signed-off-by: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com>
> 
> Thanks for your patch!
> 
>> --- a/drivers/tty/serial/sh-sci.c
>> +++ b/drivers/tty/serial/sh-sci.c
>> @@ -3435,7 +3435,24 @@ static int sci_probe_single(struct platform_device *dev,
>>                 sciport->port.flags |= UPF_HARD_FLOW;
>>         }
>>
>> +       /*
>> +        * In case:
>> +        * - this is the earlycon port (mapped on index 0 in sci_ports[]) and
>> +        * - it now maps to an alias other than zero and
>> +        * - the earlycon is still alive (e.g., "earlycon keep_bootcon" is
>> +        *   available in bootargs)
>> +        *
>> +        * we need to avoid disabling clocks and PM domains through the runtime
>> +        * PM APIs called in __device_attach(). For this, increment the runtime
>> +        * PM reference counter (the clocks and PM domains were already enabled
>> +        * by the bootloader). Otherwise the earlycon may access the HW when it
>> +        * has no clocks enabled leading to failures (infinite loop in
>> +        * sci_poll_put_char()).
>> +        */
>> +
>>         if (sci_ports[0].earlycon && sci_ports[0].port.mapbase == sci_res->start) {
> 
> Now there are two tests for mapbase: here and in sci_probe()...

I'm not sure how can we avoid it. We need to re-check it in this function
as the sci_probe_single() is the one that enables the runtime PM. Would you
prefer to move the devm_pm_runtime_enable() in sci_probe() and have the
pm_runtime_get_noresume() in sci_probe() as well?

Thank you,
Claudiu

> 
>> +               pm_runtime_get_noresume(&dev->dev);
>> +
>>                 /*
>>                  * Skip cleanup up the sci_port[0] in early_console_exit(), this
>>                  * port is the same as the earlycon one.
> 
> Gr{oetje,eeting}s,
> 
>                         Geert
>