Re: [PATCH v2] gpio: realtek-otto: switch to 32-bit I/O

[Sander Vanheule]

On Sun, 2022-07-24 at 13:31 +0200, Sander Vanheule wrote:
> By using 16-bit I/O on the GPIO peripheral, which is apparently not safe
> on MIPS, the IMR can end up containing garbage. This then results in
> interrupt triggers for lines that don't have an interrupt source
> associated. The irq_desc lookup fails, and the ISR will not be cleared,
> keeping the CPU busy until reboot, or until another IMR operation
> restores the correct value. This situation appears to happen very
> rarely, for < 0.5% of IMR writes.
> 
> Instead of using 8-bit or 16-bit I/O operations on the 32-bit memory
> mapped peripheral registers, switch to using 32-bit I/O only, operating
> on the entire bank for all single bit line settings. For 2-bit line
> settings, with 16-bit port values, stick to manual (un)packing.
> 
> This issue has been seen on RTL8382M (HPE 1920-16G), RTL8391M (Netgear
> GS728TP v2), and RTL8393M (D-Link DGS-1210-52 F3, Zyxel GS1900-48).
> 
> Reported-by: Luiz Angelo Daros de Luca <luizluca@xxxxxxxxx> # DGS-1210-52
> Reported-by: Birger Koblitz <mail@xxxxxxxxxxxxxxxxx> # GS728TP
> Reported-by: Jan Hoffmann <jan@xxxxxx> # 1920-16G
> Cc: Paul Cercueil <paul@xxxxxxxxxxxxxxx>
> Signed-off-by: Sander Vanheule <sander@xxxxxxxxxxxxx>
> ---

...

> @@ -307,16 +308,17 @@ static int realtek_gpio_irq_set_affinity(struct irq_data
> *data,
>  static int realtek_gpio_irq_init(struct gpio_chip *gc)
>  {
>         struct realtek_gpio_ctrl *ctrl = gpiochip_get_data(gc);
> -       unsigned int port;
> +       u32 mask_all = GENMASK(gc->ngpio, 0);

This should be GENMASK(gc->ngpio - 1, 0).

I'll wait a bit more for other comments before sending a v3.

Best,
Sander