Re: [PATCH V6 XDMA 0/2] xilinx XDMA driver

From: Lizhi Hou
Date: Thu Oct 06 2022 - 13:43:08 EST



On 10/6/22 09:37, Martin Tůma wrote:
On 04. 10. 22 23:43, Lizhi Hou wrote:
Hello,

This V6 of patch series is to provide the platform driver to support the
Xilinx XDMA subsystem. The XDMA subsystem is used in conjunction with the
PCI Express IP block to provide high performance data transfer between host
memory and the card's DMA subsystem. It also provides up to 16 user
interrupt wires to user logic that generate interrupts to the host.

             +-------+       +-------+       +-----------+
    PCIe     |       |       |       |       |           |
    Tx/Rx    |       |       |       |  AXI  |           |
  <=======>  | PCIE  | <===> | XDMA  | <====>| User Logic|
             |       |       |       |       |           |
             +-------+       +-------+       +-----------+

The XDMA has been used for Xilinx Alveo PCIe devices.
And it is also integrated into Versal ACAP DMA and Bridge Subsystem.
     https://www.xilinx.com/products/boards-and-kits/alveo.html
https://docs.xilinx.com/r/en-US/pg344-pcie-dma-versal/Introduction-to-the-DMA-and-Bridge-Subsystems

The device driver for any FPGA based PCIe device which leverages XDMA can
call the standard dmaengine APIs to discover and use the XDMA subsystem
without duplicating the XDMA driver code in its own driver.

Changes since v5:
- Modified user logic interrupt APIs to handle user logic IP which does not
   have its own register to enable/disable interrupt.
- Clean up code based on review comments.

Changes since v4:
- Modified user logic interrupt APIs.

Changes since v3:
- Added one patch to support user logic interrupt.

Changes since v2:
- Removed tasklet.
- Fixed regression bug introduced to V2.
- Test Robot warning.

Changes since v1:
- Moved filling hardware descriptor to xdma_prep_device_sg().
- Changed hardware descriptor enum to "struct xdma_hw_desc".
- Minor changes from code review comments.

Lizhi Hou (2):
   dmaengine: xilinx: xdma: Add xilinx xdma driver
   dmaengine: xilinx: xdma: Add user logic interrupt support

  MAINTAINERS                            |   11 +
  drivers/dma/Kconfig                    |   13 +
  drivers/dma/xilinx/Makefile            |    1 +
  drivers/dma/xilinx/xdma-regs.h         |  171 ++++
  drivers/dma/xilinx/xdma.c              | 1034 ++++++++++++++++++++++++
  include/linux/dma/amd_xdma.h           |   16 +
  include/linux/platform_data/amd_xdma.h |   34 +
  7 files changed, 1280 insertions(+)
  create mode 100644 drivers/dma/xilinx/xdma-regs.h
  create mode 100644 drivers/dma/xilinx/xdma.c
  create mode 100644 include/linux/dma/amd_xdma.h
  create mode 100644 include/linux/platform_data/amd_xdma.h


Hi,
I have rewritten our V4L2 driver to use this new XDMA driver, but it does not work on our HW (where the previous Xilinx XDMA driver derived from the Xilinx sample code worked fine). The driver is sucessfully loaded and 4(2+2) DMA channels are successfully created. But when a DMA transfer is initiated, I get an error from my PC's DMA chip:

AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x000a address=0x36a00000 flags=0x0000]

and no error from XDMA.

Does the driver expect some special FPGA IP core configuration? Or is there something else I'm missing? My code is quiet similar to what you use in your XRT repo on GitHub (there is btw. a bug in the XRT code - you do not clear the dma_slave_config struct before using) but in my case the DMA transfer triggers the AMD-Vi error and timeouts.

The code of our driver is attached, the relevant parts are in mgb4_dma.c
and mgb4_core.c.

M.

Hi Martin,

Thanks for trying this and got a lot thing works. I have read your driver. Could you call pci_map_sg() before calling prep_sg()? (and pci_unmap_sg()) after transfer complete?

example: https://github.com/houlz0507/XRT-1/blob/xdma_v4_usage/src/runtime_src/core/pcie/driver/linux/xocl/subdev/xdma.c#L103


Thanks,

Lizhi