Hi,

On Thu, Nov 05, 2026 at 08:51:14AM +0800, Shawn Lin wrote:
> [...]
> > +static void rockchip_pcie_pme_turn_off(struct dw_pcie_rp *pp)
> > +{
> > +	struct dw_pcie *pci = to_dw_pcie_from_pp(pp);
> > +	struct rockchip_pcie *rockchip = to_rockchip_pcie(pci);
> > +	struct device *dev = rockchip->pci.dev;
> > +	u32 status;
> > +	int ret;
> > +
> > +	/* 1. Broadcast PME_Turn_Off Message, bit 4 self-clear once done */
> > +	rockchip_pcie_writel_apb(rockchip, PME_TURN_OFF, PCIE_CLIENT_MSG_GEN);
> > +	ret = readl_poll_timeout(rockchip->apb_base + PCIE_CLIENT_MSG_GEN,
> > +				 status, !(status & BIT(4)), PCIE_PME_TO_L2_TIMEOUT_US / 10,
> > +				 PCIE_PME_TO_L2_TIMEOUT_US);
> > +	if (ret) {
> > +		dev_warn(dev, "Failed to send PME_Turn_Off\n");
> > +		return;
> > +	}
> > +
> > +	/* 2. Wait for PME_TO_Ack, bit 9 will be set once received */
> > +	ret = readl_poll_timeout(rockchip->apb_base + PCIE_CLIENT_INTR_STATUS_MSG_RX,
> > +				 status, status & BIT(9), PCIE_PME_TO_L2_TIMEOUT_US / 10,
> > +				 PCIE_PME_TO_L2_TIMEOUT_US);
> > +	if (ret) {
> > +		dev_warn(dev, "Failed to receive PME_TO_Ack\n");
> > +		return;
> > +	}
> > +
> > +	/* 3. Clear PME_TO_Ack and Wait for ready to enter L23 message */
> > +	rockchip_pcie_writel_apb(rockchip, PME_TO_ACK, PCIE_CLIENT_INTR_STATUS_MSG_RX);
> > +	ret = readl_poll_timeout(rockchip->apb_base + PCIE_CLIENT_POWER_CON,
> > +				 status, status & PME_READY_ENTER_L23,
> > +				 PCIE_PME_TO_L2_TIMEOUT_US / 10,
> > +				 PCIE_PME_TO_L2_TIMEOUT_US);
> > +	if (ret)
> > +		dev_err(dev, "Failed to get ready to enter L23 message\n");
> > +}
> 
> This is a question for the DWC core: should we change
> dw_pcie_host_ops::pme_turn_off() to return a value instead of silently
> ignoring a failure?
> 
> Looking at pcie-designware-host.c, we have:
> 
> if (pci->pp.ops->pme_turn_off) {
>     pci->pp.ops->pme_turn_off(&pci->pp);
> } else {
>     ret = dw_pcie_pme_turn_off(pci);
>     if (ret)
>         return ret;
> }
> 
> The handling feels asymmetric. If ->pme_turn_off fails, shouldn't we abort
> dw_pcie_suspend_noirq as well?

That sounds sensible to me. I had a look and apart from Rockchip,
Layescape can run into an error. Most platforms cannot produce any
errors in that function, which explains why this was void to begin
with.

> Other than that, this patch looks good to me in current shape,
> Reviewed-by: Shawn Lin <shawn.lin@rock-chips.com>

Thanks, taken over for v9.

> We can further discuss the callback's return value issue on top.

Agreed. This will touch all the glue drivers to add a return 0. I
think its better to do that in a separate series.

Greetings,

-- Sebastian