The PCIe host bridge on RK3399 advertises a single address range
marked as 64-bit memory even though it lies entirely below 4GB. While
previously, the OF PCI range parser treated 64-bit ranges more
leniently (i.e., as 32-bit), since commit 9d57e61bf723 ("of/pci: Add
IORESOURCE_MEM_64 to resource flags for 64-bit memory addresses") the
code takes a stricter view and treats the ranges as advertised in the
device tree (i.e, as 64-bit).
The change in behaviour causes failure when allocating bus addresses
to devices connected behind a PCI-to-PCI bridge that require
non-prefetchable memory ranges. The allocation failure was observed
for certain Samsung NVMe drives connected to RockPro64 boards.
Update the host bridge window attributes to treat it as 32-bit address
memory. This fixes the allocation failure observed since commit
9d57e61bf723.
Reported-by: Alexandru Elisei <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Suggested-by: Robin Murphy <[email protected]>
Signed-off-by: Punit Agrawal <[email protected]>
Cc: Heiko Stuebner <[email protected]>
Cc: Rob Herring <[email protected]>
---
Hi,
The patch fixes the failure observed with detecting certain Samsung
NVMe drives on RK3399 based boards.
Hopefully, the folks on this thread can provide some input on the
reason the host bridge window was originally marked as 64-bit or if
there are any downsides to applying the patch.
Thanks,
Punit
arch/arm64/boot/dts/rockchip/rk3399.dtsi | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/arch/arm64/boot/dts/rockchip/rk3399.dtsi b/arch/arm64/boot/dts/rockchip/rk3399.dtsi
index 634a91af8e83..4b854eb21f72 100644
--- a/arch/arm64/boot/dts/rockchip/rk3399.dtsi
+++ b/arch/arm64/boot/dts/rockchip/rk3399.dtsi
@@ -227,7 +227,7 @@ pcie0: pcie@f8000000 {
<&pcie_phy 2>, <&pcie_phy 3>;
phy-names = "pcie-phy-0", "pcie-phy-1",
"pcie-phy-2", "pcie-phy-3";
- ranges = <0x83000000 0x0 0xfa000000 0x0 0xfa000000 0x0 0x1e00000>,
+ ranges = <0x82000000 0x0 0xfa000000 0x0 0xfa000000 0x0 0x1e00000>,
<0x81000000 0x0 0xfbe00000 0x0 0xfbe00000 0x0 0x100000>;
resets = <&cru SRST_PCIE_CORE>, <&cru SRST_PCIE_MGMT>,
<&cru SRST_PCIE_MGMT_STICKY>, <&cru SRST_PCIE_PIPE>,
--
2.30.2
On Wed, May 26, 2021 at 8:35 AM Punit Agrawal <[email protected]> wrote:
>
> The PCIe host bridge on RK3399 advertises a single address range
> marked as 64-bit memory even though it lies entirely below 4GB. While
> previously, the OF PCI range parser treated 64-bit ranges more
> leniently (i.e., as 32-bit), since commit 9d57e61bf723 ("of/pci: Add
> IORESOURCE_MEM_64 to resource flags for 64-bit memory addresses") the
> code takes a stricter view and treats the ranges as advertised in the
> device tree (i.e, as 64-bit).
>
> The change in behaviour causes failure when allocating bus addresses
> to devices connected behind a PCI-to-PCI bridge that require
> non-prefetchable memory ranges. The allocation failure was observed
> for certain Samsung NVMe drives connected to RockPro64 boards.
>
> Update the host bridge window attributes to treat it as 32-bit address
> memory. This fixes the allocation failure observed since commit
> 9d57e61bf723.
>
> Reported-by: Alexandru Elisei <[email protected]>
> Link: https://lore.kernel.org/r/[email protected]
> Suggested-by: Robin Murphy <[email protected]>
> Signed-off-by: Punit Agrawal <[email protected]>
> Cc: Heiko Stuebner <[email protected]>
> Cc: Rob Herring <[email protected]>
> ---
> Hi,
>
> The patch fixes the failure observed with detecting certain Samsung
> NVMe drives on RK3399 based boards.
>
> Hopefully, the folks on this thread can provide some input on the
> reason the host bridge window was originally marked as 64-bit or if
> there are any downsides to applying the patch.
We can't require *only* a DT update to fix this. Ideally, the Rockchip
PCI driver should clear the 64-bit flag in the resources though I'm
not sure if the bridge driver would have access early enough.
This is not the first time we've had to work-around RK3399 PCI DT (see
commit d1ac0002dd297069 "of: address: Work around missing device_type
property in pcie nodes").
Rob
Hi Rob,
Thanks for taking a look.
Rob Herring <[email protected]> writes:
> On Wed, May 26, 2021 at 8:35 AM Punit Agrawal <[email protected]> wrote:
>>
>> The PCIe host bridge on RK3399 advertises a single address range
>> marked as 64-bit memory even though it lies entirely below 4GB. While
>> previously, the OF PCI range parser treated 64-bit ranges more
>> leniently (i.e., as 32-bit), since commit 9d57e61bf723 ("of/pci: Add
>> IORESOURCE_MEM_64 to resource flags for 64-bit memory addresses") the
>> code takes a stricter view and treats the ranges as advertised in the
>> device tree (i.e, as 64-bit).
>>
>> The change in behaviour causes failure when allocating bus addresses
>> to devices connected behind a PCI-to-PCI bridge that require
>> non-prefetchable memory ranges. The allocation failure was observed
>> for certain Samsung NVMe drives connected to RockPro64 boards.
>>
>> Update the host bridge window attributes to treat it as 32-bit address
>> memory. This fixes the allocation failure observed since commit
>> 9d57e61bf723.
>>
>> Reported-by: Alexandru Elisei <[email protected]>
>> Link: https://lore.kernel.org/r/[email protected]
>> Suggested-by: Robin Murphy <[email protected]>
>> Signed-off-by: Punit Agrawal <[email protected]>
>> Cc: Heiko Stuebner <[email protected]>
>> Cc: Rob Herring <[email protected]>
>> ---
>> Hi,
>>
>> The patch fixes the failure observed with detecting certain Samsung
>> NVMe drives on RK3399 based boards.
>>
>> Hopefully, the folks on this thread can provide some input on the
>> reason the host bridge window was originally marked as 64-bit or if
>> there are any downsides to applying the patch.
>
> We can't require *only* a DT update to fix this. Ideally, the Rockchip
> PCI driver should clear the 64-bit flag in the resources though I'm
> not sure if the bridge driver would have access early enough.
Following the discussion in the other thread, I tested the following
changes to fixup 64-bit flag for non-prefetchable memory resources that
fit below 4GB.
If the changes look good, I'll send it out as a proper patch later
today.
---->8----
diff --git a/drivers/pci/of.c b/drivers/pci/of.c
index da5b414d585a..b9d0bee5a088 100644
--- a/drivers/pci/of.c
+++ b/drivers/pci/of.c
@@ -565,10 +565,14 @@ static int pci_parse_request_of_pci_ranges(struct device *dev,
case IORESOURCE_MEM:
res_valid |= !(res->flags & IORESOURCE_PREFETCH);
- if (!(res->flags & IORESOURCE_PREFETCH))
+ if (!(res->flags & IORESOURCE_PREFETCH)) {
if (upper_32_bits(resource_size(res)))
dev_warn(dev, "Memory resource size exceeds max for 32 bits\n");
-
+ if ((res->flags & IORESOURCE_MEM_64) && !upper_32_bits(res->end)) {
+ dev_warn(dev, "Overriding 64-bit flag for non-prefetchable memory below 4GB\n");
+ res->flags &= ~IORESOURCE_MEM_64;
+ }
+ }
break;
}
}