Re: [PATCH] arm64: dts: rockchip: align bindings to PCIe spec

[Geraldo Nascimento <geraldogabriel@gmail.com>]

On Wed, Nov 12, 2025 at 05:03:32AM -0300, Geraldo Nascimento wrote:
> On Tue, Nov 11, 2025 at 03:47:04PM +0800, 张烨 wrote:
> > Hi Geraldo,
> > 
> > In standard GPIO operations, the typical practice is to set the output level first before configuring the direction as output. This approach helps avoid outputting an uncertain voltage level at the instant when the direction switches from input to output.
> 
> Thanks for the explanation Ye Zhang, it makes sense to me. It avoids the
> pin to not be floating so to speak. I kept hammering at this problem, by
> the way is PCIe PERST# side-band signal refusing to co-operate and
> failing PCIe initial link-training.
> 
> You're not going to like this:
> 
> diff --git a/drivers/gpio/gpio-rockchip.c b/drivers/gpio/gpio-rockchip.c
> index 47174eb3ba76..fea2c55992e8 100644
> --- a/drivers/gpio/gpio-rockchip.c
> +++ b/drivers/gpio/gpio-rockchip.c
> @@ -183,11 +183,13 @@ static int rockchip_gpio_set(struct gpio_chip *gc, unsigned int offset,
>  	struct rockchip_pin_bank *bank = gpiochip_get_data(gc);
>  	unsigned long flags;
>  
> +	rockchip_gpio_set_direction(gc, offset, true);
> +
>  	raw_spin_lock_irqsave(&bank->slock, flags);
>  	rockchip_gpio_writel_bit(bank, offset, value, bank->gpio_regs->port_dr);
>  	raw_spin_unlock_irqrestore(&bank->slock, flags);
>  
> -	return 0;
> +	return rockchip_gpio_set_direction(gc, offset, false);
>  }
>  
>  static int rockchip_gpio_get(struct gpio_chip *gc, unsigned int offset)
> 
> By setting direction INPUT, then writing out, then setting OUTPUT again
> miraculously it doesn't fail initial link training, with no other
> changes that already have been rejected by PCI folks and Shawn Lin.

Hi Ye, Shawn,

Here's more contained workaround without resorting to clearing DDR to
INPUT for every GPIO:

diff --git a/drivers/pci/controller/pcie-rockchip-host.c b/drivers/pci/controller/pcie-rockchip-host.c
index ee1822ca01db..1d89131ec6ac 100644
--- a/drivers/pci/controller/pcie-rockchip-host.c
+++ b/drivers/pci/controller/pcie-rockchip-host.c
@@ -315,7 +315,8 @@ static int rockchip_pcie_host_init_port(struct rockchip_pcie *rockchip)
 			    PCIE_CLIENT_CONFIG);
 
 	msleep(PCIE_T_PVPERL_MS);
-	gpiod_set_value_cansleep(rockchip->perst_gpio, 1);
+	gpiod_direction_input(rockchip->perst_gpio);
+	gpiod_direction_output(rockchip->perst_gpio, 1);
 
 	msleep(PCIE_RESET_CONFIG_WAIT_MS);
 
This results in working PCIe for me, pass initial link training.

I think I need to provide more details:

GPIO in question is GPIO0-12 / GPIO0-PB4. On RK3399PRO VMARC schematic
it's called PCIE_PERST#_3.3V and it's in the PMUIO1 domain.

Without workaround I have about 6 milliseconds, from driver probe and
parsing of DT that sets initial value 0 for GPIO, to deassert PERST#
by setting it high. Which is why just removing msleep(PCIE_T_PVPERL_MS)
produced working PCIe for me.

After the ~6 ms it becomes necessary to hack direction to input then
setting direction to output to properly deassert PERST# and proceed
with initial link training - which is why, again, hacking PERST# as
either open-drain with pull-up or open-source with pull-down worked:
gpiolib hacks that by setting direction INPUT.

Thanks,
Geraldo Nascimento