Re: [PATCH] arm64: dts: rockchip: rk356x: Fix PCIe register map and ranges

From: Ondřej Jirman
Date: Wed Oct 05 2022 - 04:59:06 EST


Hello Peter,

On Tue, Oct 04, 2022 at 03:52:39PM -0400, Peter Geis wrote:
> On Tue, Oct 4, 2022 at 10:43 AM Ondrej Jirman <megi@xxxxxx> wrote:
>
> Good Afternoon,
>
>
> >
> > I have two Realtek PCIe wifi cards connected over the 4 port
> > PCIe bridge to Quartz64-A. The cards fail to work, when nvme
> > SSD is connected at the same time to the bridge. Without nvme
> > connected, cards work fine. The issue seems to be related
> > to mixed use of devices which make use of I/O ranges and memory
> > ranges.
> >
> > This mapping is designed to be more straightforward, inspired by
> > dt-bindings docs for sample pcie3x2 node:
> >
> > reg = <0x3 0xc0800000 0x0 0x390000>,
> > <0x0 0xfe280000 0x0 0x10000>,
> > <0x3 0x80000000 0x0 0x100000>;
> > ranges = <0x81000000 0x0 0x80800000 0x3 0x80800000 0x0 0x100000>,
> > <0x83000000 0x0 0x80900000 0x3 0x80900000 0x0 0x3f700000>;
> >
> > I noticed that this is crafted so that there doesn't need to be
> > any translation other than dropping the high dword bits, and I
> > modified the ranges for pcie2x1 to follow the same principle.
> >
> > This change to the regs/ranges makes the issue go away and both
> > nvme and wifi cards work when connected at the same time to the
> > bridge.
> >
> > Signed-off-by: Ondrej Jirman <megi@xxxxxx>
> > ---
> > arch/arm64/boot/dts/rockchip/rk356x.dtsi | 7 ++++---
> > 1 file changed, 4 insertions(+), 3 deletions(-)
> >
> > diff --git a/arch/arm64/boot/dts/rockchip/rk356x.dtsi b/arch/arm64/boot/dts/rockchip/rk356x.dtsi
> > index 319981c3e9f7..e88e8c4fe25b 100644
> > --- a/arch/arm64/boot/dts/rockchip/rk356x.dtsi
> > +++ b/arch/arm64/boot/dts/rockchip/rk356x.dtsi
> > @@ -855,7 +855,8 @@ pcie2x1: pcie@fe260000 {
> > compatible = "rockchip,rk3568-pcie";
> > reg = <0x3 0xc0000000 0x0 0x00400000>,
> > <0x0 0xfe260000 0x0 0x00010000>,
> > - <0x3 0x3f000000 0x0 0x01000000>;
> > + <0x3 0x00000000 0x0 0x01000000>;
> > +
> > reg-names = "dbi", "apb", "config";
> > interrupts = <GIC_SPI 75 IRQ_TYPE_LEVEL_HIGH>,
> > <GIC_SPI 74 IRQ_TYPE_LEVEL_HIGH>,
> > @@ -884,8 +885,8 @@ pcie2x1: pcie@fe260000 {
> > phys = <&combphy2 PHY_TYPE_PCIE>;
> > phy-names = "pcie-phy";
> > power-domains = <&power RK3568_PD_PIPE>;
> > - ranges = <0x01000000 0x0 0x3ef00000 0x3 0x3ef00000 0x0 0x00100000
> > - 0x02000000 0x0 0x00000000 0x3 0x00000000 0x0 0x3ef00000>;
> > + ranges = <0x01000000 0x0 0x01000000 0x3 0x01000000 0x0 0x00100000
> > + 0x02000000 0x0 0x02000000 0x3 0x02000000 0x0 0x3e000000>;
>
> Have you verified these ranges do not regress the NVMe drive when it
> is connected directly to the controller? The reason we went with the
> configuration space we did was because the original space from
> downstream caused errors on NVMe drives when reading large amounts
> (>1GB) of data at a time.

I did. Anyway... looking at the ranges more carefully, I came up with a scheme
that uses whole of 0x300000000 range for MEM and the smaller 32M range at
0xf4000000 for I/O and config, and tested it again to work for both:

- nvme without a switch (directly connected)
- nvme+wifi cards with a switch

See v2 for more details.

That ranges setup is closer to what BSP does.

Kind regards,
o.

> Very Respectfully,
> Peter Geis
>
> > resets = <&cru SRST_PCIE20_POWERUP>;
> > reset-names = "pipe";
> > #address-cells = <3>;
> > --
> > 2.37.3
> >