Re: [PATCH 0/5] phy: qcom: qmp-pcie: Add PCIe Gen5 8-lane bifurcation support for Glymur

[Qiang Yu <qiang.yu@oss.qualcomm.com>]

On Thu, Mar 05, 2026 at 10:14:05AM +0100, Konrad Dybcio wrote:
> On 3/4/26 9:21 AM, Qiang Yu wrote:
> > This patch series adds support for PCIe Gen5 8-lane bifurcation mode on
> > the Glymur SoC's third PCIe controller. In this configuration, pcie3a PHY
> > acts as leader and pcie3b PHY as follower to form a single 8-lane PCIe
> > Gen5 interface.
> > 
> > To support 8-lanes mode, this patch series add multiple power domain and
> > multi nocsr reset infrastructure as the hardware programming guide
> > specifies a strict initialization sequence for bifurcation mode that
> > requires coordinated multi-PHY resource management:
> > 
> > 1. Turn on both pcie3a_phy_gdsc and pcie3b_phy_gdsc power domains
> > 2. Assert both pcie3a and pcie3b nocsr resets, then deassert them together
> > 3. Enable all pcie3a PHY clocks and pcie3b PHY aux clock (bifur_aux)
> > 4. Poll for PHY ready status
> 
> I think we never concluded the discussion where I suggested the
> bifurcated PHY may be better expressed as a single node with
> #phy-cells = <1>, removing the need for duplicated resource references
>
I understand your suggestion would look like below. I agree that the
unified PHY approach being more elegant from a device tree perspective,
provide better DT flexibility and eliminate the need for different
compatibles and dupicated resources between 1x8 and 2x4 modes.

However, this will include implementation complexity to phy driver.
The driver would need conditional logic to selectively enable different
clocks/resets based on the PHY parameter and maintain mode-specific
resource arrays. There's also the issue that assigned-clocks
GCC_PCIE_3A_PHY_RCHNG_CLK and GCC_PCIE_3B_PHY_RCHNG_CLK will be set before
probe no matter which mode is used, even though in 1x8 mode or only one of
them is actually needed. For pipe clock outputs, only pcie3a_pipe_clk would
be needed in 1x8 mode while pcie3b_pipe_clk would be unused. For
powerdomain, we also need to add additional logic to attach and turn
on/off them.

While these challenges could be resolved, I'm not sure the benefits
justify the added complexity.

pcie3_unified_phy {
    compatible = "qcom,glymur-qmp-gen5-pcie-phy";
    reg = <0 0x00f00000 0 0x10000>, <0 0x00f10000 0 0x10000>;  /* Both PHY ranges */

    clocks = <&gcc GCC_PCIE_PHY_3A_AUX_CLK>,
             <&gcc GCC_PCIE_3A_CFG_AHB_CLK>,
             <&tcsr TCSR_PCIE_3_CLKREF_EN>,
             <&gcc GCC_PCIE_3A_PHY_RCHNG_CLK>,
             <&gcc GCC_PCIE_3A_PIPE_CLK>,
             <&gcc GCC_PCIE_PHY_3B_AUX_CLK>,
             <&gcc GCC_PCIE_3B_CFG_AHB_CLK>,
             <&gcc GCC_PCIE_3B_PHY_RCHNG_CLK>,
             <&gcc GCC_PCIE_3B_PIPE_CLK>,
             <&gcc GCC_PCIE_3B_PIPE_DIV2_CLK>;

    power-domains = <&gcc GCC_PCIE_3A_PHY_GDSC>,
                    <&gcc GCC_PCIE_3B_PHY_GDSC>;

    resets = <&gcc GCC_PCIE_3A_PHY_BCR>,
             <&gcc GCC_PCIE_3A_NOCSR_COM_PHY_BCR>,
             <&gcc GCC_PCIE_3B_PHY_BCR>,
             <&gcc GCC_PCIE_3B_NOCSR_COM_PHY_BCR>;

	#clock-cells = <1>;
    clock-output-names = "pcie3a_pipe_clk", "pcie3b_pipe_clk";
    assigned-clocks = <&gcc GCC_PCIE_3A_PHY_RCHNG_CLK>,
                      <&gcc GCC_PCIE_3B_PHY_RCHNG_CLK>;
    assigned-clock-rates = <100000000>, <100000000>;

    #phy-cells = <1>;  /* Parameter: 0=PHY_A, 1=PHY_B, 2=UNIFIED_8LANE */
};

For 2x4 mode (independent 4-lane PHYs):
&pcie3a {
    phys = <&pcie3_unified_phy PHY_A>;  /* PHY A only */
    status = "okay";
};

&pcie3b {
    phys = <&pcie3_unified_phy PHY_B>;  /* PHY B only */
    status = "okay";
};

For 1x8 mode (unified 8-lane PHY):

&pcie3a {
    phys = <&pcie3_unified_phy PHY_AB>;
    num-lanes = <8>;
    status = "okay";
};

&pcie3b {
    status = "disabled";
};

- Qiang Yu