Hello, nVidia Support,
I am using a Xilinx FPGA as EP and DMA data to Xavier as RC. The Xilinx FPGA and DMA function was verified on an Intel x86_64 CPU, so it is confirmed the FPGA functions well. The issue comes when we port the driver from x86_64 to ARM64 on Xavier AGX. When set-up the DMA, the DMA looks like is running, can receive interrupts, but when read data from the buffer where the DMA writes data to, the data are pre-filled patterns 0x5a5a. It looks like the DMA is writing data to somewhere else.
Here are the address I use for setup the DMA, I use the kernbuf Address and program to FPGA as DMA destination address. The userBuf Address is the virtual address that I read from. These address were obtained by calling a DMABufAlloc function that is from a driver from a third-party company.
buf_0 bufferSize : 0x100000
buf_0 userBuf Addr: 0x7f9edee000
buf_0 kernBuf Addr: 0x4fff00000
buf_0 kernBuf64 Addr: 0x4fff00000
My question is, for the kernBuf address that is to be used as the DMA destination address, does it fall within 0x8000_0000 to 0xFFF0_0000?
as per this post:
In this nvidia forum, post 2 states that “Bus address for PCIe is reserved from 0x8000_0000 to 0xFFF0_0000. So, whether the memory is allocated directly or mapped later on, iova address always comes from this region.”
2). some forum posts state by disabling SMMU might solve this issue. Is this the only solution for this? Since the CPU designed the SMMU/IOMMU, there are good reasons to use this, and we would like to keep these functions. What are the solutions to resolve the above issue?
Thank you.
Mei Guodong