On 2020-08-28 18:53, Manish Narani wrote:
Hi Robin,
Thanks for the review. Please find my comment below inline.
-----Original Message-----
From: Robin Murphy <robin.murphy@xxxxxxx>
Sent: Friday, August 28, 2020 12:17 AM
To: Manish Narani <MNARANI@xxxxxxxxxx>; gregkh@xxxxxxxxxxxxxxxxxxx;
robh+dt@xxxxxxxxxx; Michal Simek <michals@xxxxxxxxxx>; balbi@xxxxxxxxxx;
p.zabel@xxxxxxxxxxxxxx
Cc: devicetree@xxxxxxxxxxxxxxx; linux-usb@xxxxxxxxxxxxxxx; linux-
kernel@xxxxxxxxxxxxxxx; git <git@xxxxxxxxxx>; linux-arm-
kernel@xxxxxxxxxxxxxxxxxxx
Subject: Re: [PATCH 2/2] usb: dwc3: Add driver for Xilinx platforms
On 2020-08-26 19:44, Manish Narani wrote:
[...]
+ /*
+ * This routes the usb dma traffic to go through CCI path instead
+ * of reaching DDR directly. This traffic routing is needed to
+ * make SMMU and CCI work with USB dma.
+ */
+ if (of_dma_is_coherent(dev->of_node) || dev->iommu_group) {
+ reg = readl(priv_data->regs + XLNX_USB_COHERENCY);
+ reg |= XLNX_USB_COHERENCY_ENABLE;
+ writel(reg, priv_data->regs + XLNX_USB_COHERENCY);
+ }
This looks rather suspect - coherency should be based on coherency, not
on whether an IOMMU group is present. If the device isn't described as
coherent in the DT, then any SMMU mappings will end up using attributes
that will downgrade traffic to be non-snooping anyway. And if the SMMU
is enabled but not translating (e.g. "iommu.passthrough=1") then
enabling hardware coherency when the DMA layer hasn't been told about it
can potentially lead to nasty subtle problems and data loss.
May be the description needs to be updated in this. This is not the actual coherency enabling bit, but this is needed when coherency is enabled.
This is a register inside Xilinx USB controller which handles USB (which is in LPD) traffic route switching from LPD (Low Power Domain) to FPD (Full Power Domain) path in the Xilinx SoC in either of the below scenarios:
1. Device is described coherent in DT.
2. SMMU is enabled.
I will update the same in v2.
Ah, OK, so it's just that the control bit itself has a terrible name :)
From the available information I had assumed that this controlled the
output attributes, and that the interconnect might then steer traffic
based on those. Explaining a bit more clearly in the comment probably
would be a good idea. In that case, I'd concur that the current logic is
in fact appropriate, but please use the device_iommu_mapped() helper for
cleanliness.
Cheers,
Robin.
