Question about Arrow Lake PCIe P2P capability qualification in NVIDIA open kernel modules (#1452)

Hello,

I’m investigating an issue reported in the NVIDIA open GPU kernel modules repository:

The report concerns PCIe peer-to-peer DMA-BUF attachment restrictions on an Intel Arrow Lake platform with an NVIDIA RTX PRO 6000 Blackwell GPU. The reporter indicates that the proprietary driver of the same version does not exhibit the issue.

I’ve traced the relevant source paths and found that:

  • nvcst.h maps several Intel Arrow Lake device IDs to CS_INTEL_7D06.

  • The PCIe peer-access capability table in chipset_pcie.c has explicit entries for several Intel chipsets, but I could not find an explicit entry for CS_INTEL_7D06.

  • The table’s fallback entry specifies CAPABLE_NONE.

  • The resulting peer-read/write capability flags are used by _checkP2pChipsetSupport(), which can set NV_FLAG_PCI_P2P_UNSUPPORTED_CHIPSET.

  • The default DMA-BUF attachment path then applies a PCI-topology restriction when that flag is set.

This appears consistent with the reported behavior, but I haven’t reproduced the issue on the reporter’s hardware and cannot establish from source inspection alone whether Arrow Lake is qualified for peer reads and writes.

Could someone familiar with the NVIDIA Linux driver or PCIe P2P qualification clarify whether the absence of CS_INTEL_7D06 from the capability table is intentional? If support is appropriate, what hardware validation would be required before proposing a change?

I’m not proposing an unvalidated CAPABLE_READWRITE change, I want to understand the intended support policy first.

Thank you.

1 Like