Re: [PATCH v2 1/2] PCI/ASPM: Override the ASPM and Clock PM states set by BIOS for devicetree platforms

[Niklas Cassel <cassel@kernel.org>]

Hello Shawn,

On Wed, Oct 15, 2025 at 05:11:39PM +0800, Shawn Lin wrote:
> > 
> > Thanks! Could you please try the below diff with f3ac2ff14834 applied?
> > 
> > diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
> > index 214ed060ca1b..0069d06c282d 100644
> > --- a/drivers/pci/quirks.c
> > +++ b/drivers/pci/quirks.c
> > @@ -2525,6 +2525,15 @@ static void quirk_disable_aspm_l0s_l1(struct pci_dev *dev)
> >    */
> >   DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_ASMEDIA, 0x1080, quirk_disable_aspm_l0s_l1);
> > 
> > +
> > +static void quirk_disable_aspm_all(struct pci_dev *dev)
> > +{
> > +       pci_info(dev, "Disabling ASPM\n");
> > +       pci_disable_link_state(dev, PCIE_LINK_STATE_ALL);
> > +}
> > +
> > +DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_ROCKCHIP, 0x3588, quirk_disable_aspm_all);
> 
> That's not true from my POV. Rockchip platform supports all ASPM policy
> after mass production verification. I also verified current upstream
> code this morning with RK3588-EVB and can check L0s/L1/L1ss work fine.
> 
> The log and lspci output could be found here:
> https://pastebin.com/qizeYED7
> 
> Moreover, I disscussed this issue with FUKAUMI today off-list and his
> board seems to work when only disable L1ss by patching:
> 
> --- a/drivers/pci/pcie/aspm.c
> +++ b/drivers/pci/pcie/aspm.c
> @@ -813,7 +813,7 @@ static void pcie_aspm_override_default_link_state(struct
> pcie_link_state *link)
> 
>         /* For devicetree platforms, enable all ASPM states by default */
>         if (of_have_populated_dt()) {
> -               link->aspm_default = PCIE_LINK_STATE_ASPM_ALL;
> +               link->aspm_default = PCIE_LINK_STATE_L0S |
> PCIE_LINK_STATE_L1;
>                 override = link->aspm_default & ~link->aspm_enabled;
>                 if (override)
>                         pci_info(pdev, "ASPM: DT platform,
> 
> 
> So, is there a proper way to just disable this feature for spec boards
> instead of this Soc?

This fix seems do the trick, without needing to patch common code (aspm.c):

diff --git a/drivers/pci/controller/dwc/pcie-dw-rockchip.c b/drivers/pci/controller/dwc/pcie-dw-rockchip.c
index 3e2752c7dd09..f5e1aaa97719 100644
--- a/drivers/pci/controller/dwc/pcie-dw-rockchip.c
+++ b/drivers/pci/controller/dwc/pcie-dw-rockchip.c
@@ -200,6 +200,19 @@ static bool rockchip_pcie_link_up(struct dw_pcie *pci)
 	return FIELD_GET(PCIE_LINKUP_MASK, val) == PCIE_LINKUP;
 }
 
+static void rockchip_pcie_disable_l1sub(struct dw_pcie *pci)
+{
+	u32 cap, l1subcap;
+
+	cap = dw_pcie_find_ext_capability(pci, PCI_EXT_CAP_ID_L1SS);
+	if (cap) {
+		l1subcap = dw_pcie_readl_dbi(pci, cap + PCI_L1SS_CAP);
+		l1subcap &= ~(PCI_L1SS_CAP_ASPM_L1_1 | PCI_L1SS_CAP_ASPM_L1_2 | PCI_L1SS_CAP_L1_PM_SS);
+		dw_pcie_writel_dbi(pci, cap + PCI_L1SS_CAP, l1subcap);
+		l1subcap = dw_pcie_readl_dbi(pci, cap + PCI_L1SS_CAP);
+	}
+}
+
 static void rockchip_pcie_enable_l0s(struct dw_pcie *pci)
 {
 	u32 cap, lnkcap;
@@ -264,6 +277,7 @@ static int rockchip_pcie_host_init(struct dw_pcie_rp *pp)
 	irq_set_chained_handler_and_data(irq, rockchip_pcie_intx_handler,
 					rockchip);
 
+	rockchip_pcie_disable_l1sub(pci);
 	rockchip_pcie_enable_l0s(pci);
 
 	return 0;
@@ -301,6 +315,7 @@ static void rockchip_pcie_ep_init(struct dw_pcie_ep *ep)
 	struct dw_pcie *pci = to_dw_pcie_from_ep(ep);
 	enum pci_barno bar;
 
+	rockchip_pcie_disable_l1sub(pci);
 	rockchip_pcie_enable_l0s(pci);
 	rockchip_pcie_ep_hide_broken_ats_cap_rk3588(ep);




In reality, I think that pcie-dw-rockchip.c should check 'supports-clkreq',
and only if it doesn't support clkreq, it should disable L1 substates, similar
to how the Tegra driver does things:
https://github.com/torvalds/linux/blob/v6.18-rc1/drivers/pci/controller/dwc/pcie-tegra194.c#L934-L938
https://github.com/torvalds/linux/blob/v6.18-rc1/drivers/pci/controller/dwc/pcie-tegra194.c#L1164-L1165

In fact, that is also how the downstream rockchip drives does things:
https://github.com/rockchip-linux/kernel/blob/develop-6.6/drivers/pci/controller/dwc/pcie-dw-rockchip.c#L200-L233
https://github.com/rockchip-linux/kernel/blob/develop-6.6/drivers/pci/controller/dwc/pcie-dw-rockchip.c#L725

So I guess we either:
1) Add code to pcie-dw-rockchip.c to unconditionally disable L1 substates, or
2) We add code to:
- If have 'supports-clkreq' property, set PCIE_CLIENT_POWER_CON.app_clk_req_n=1
- If don't have 'supports-clkreq' property, disable L1 substates.

I think we need to do either 1) or 2), because a user can build the kernel with
CONFIG_PCIEASPM_POWER_SUPERSAVE=y
and that would break things even on older kernels, that don't have Mani's recent
commit.



Mani, perhaps common code (aspm.c) should enable L1 substates only if
'supports-clkreq' DT property exists?



Kind regards,
Niklas