AMD PTDMA Driver Landing For Linux 5.15 After Two Years In The Works

[phoronix]

Phoronix: AMD PTDMA Driver Landing For Linux 5.15 After Two Years In The Works

Going back to September 2019 was work on the AMD PTDMA driver for supporting this controller found on modern AMD processors for high bandwidth memory-to-memory and I/O copy operations. With the Linux 5.15 cycle the AMD PTDMA driver is finally being merged to the mainline kernel...

https://www.phoronix.com/scan.php?page=news_item&px=AMD-PTDMA-For-Linux-5.15

[Azrael5]

Which AMD CPus families support this feature?

[kieffer]

At least it is available since Epyc-Rome:

$ grep "model name" /proc/cpuinfo |uniq
model name : AMD EPYC 7262 8-Core Processor
$ lspci|grep PTDMA
03:00.2 Encryption controller: Advanced Micro Devices, Inc. [AMD] Starship/Matisse PTDMA
04:00.2 Encryption controller: Advanced Micro Devices, Inc. [AMD] Starship/Matisse PTDMA
44:00.2 Encryption controller: Advanced Micro Devices, Inc. [AMD] Starship/Matisse PTDMA
45:00.2 Encryption controller: Advanced Micro Devices, Inc. [AMD] Starship/Matisse PTDMA
82:00.2 Encryption controller: Advanced Micro Devices, Inc. [AMD] Starship/Matisse PTDMA
83:00.2 Encryption controller: Advanced Micro Devices, Inc. [AMD] Starship/Matisse PTDMA
c4:00.2 Encryption controller: Advanced Micro Devices, Inc. [AMD] Starship/Matisse PTDMA
c5:00.2 Encryption controller: Advanced Micro Devices, Inc. [AMD] Starship/Matisse PTDMA

[abott]

Sounds really important for general compute and huge data movement in general. It's also only 2,000 lines of code.

It's crazy how this stuff isn't done and finished faster. Get the docs out, get people on your payroll getting this in the kernel before the hardware is out, make a driver example on Linux for the Windows devs to reference for its hardware working if you have to. Get it in faster, and give it some non-zero priority.

You've had two Zen CPU's to get your stuff in line with the software side for the processors. But after so long, it seems only the GPU guys get how hardware support should happen and in a timely manner. The reverse engineering and bring-up from hackers with the PS4 was faster and better than the whole CPU/Chipset software support.

It's really mind boggling that this isn't an issue that is being worked through.

[kieffer]

Intel provides software driver for their processors far ahead of their launch. AMD is poor at providing drivers. It has been so for decades, and apparently it not improving that fast. On the other hand, AMD had patches for ages and has been trying to push it to the kernel since kernel 5.4. I suspect the AMD developers are not used to push code to the kernel.
I suspect this device is in charge of managing the unified memory of the AMD compute GPGPU (CDNA). This looks like the OpenCAPI that IBM has in their Power9.