[Bug 2173894] Re: Intel IOMMU duplicate DMA PTE warnings overlap RMRR in multifunction NVIDIA GPU group

** Tags added: kernel-daily-bug -- You received this bug notification because you are subscribed to linux in Ubuntu. Matching subscriptions: Bgg, Bmail, Nb https://bugs.launchpad.net/bugs/2173894 Title: Intel IOMMU duplicate DMA PTE warnings overlap RMRR in multifunction NVIDIA GPU group Status in linux package in Ubuntu: New Bug description: Ubuntu 24.04.5 LTS amd64, kernel 6.8.0-149-generic (6.8.0-149.149), ASRock X99 Extreme3, Intel i7-5930K, BIOS P1.40, NVIDIA GTX 1660 Ti, driver 580.178.04. During one supervised Ollama cold model load, 15 duplicate-PTE warnings occurred over approximately 0.180 seconds. The collected log identifies llama-server and NVIDIA DMA allocation/scatterlist calls leading to Intel IOMMU mapping. Inference completed. The active DMAR table is 204 bytes with a valid checksum. Its single RMRR spans 0x7ae64000–0x7ae72fff in segment 0000. Explicit endpoint scopes are 01:00.2, 00:14.0, 00:1a.0 and 00:1d.0. The 15 conflicting vPFNs, 0x7ae64–0x7ae72, exactly cover this region. GPU graphics, audio, USB and USB-C functions 01:00.0–01:00.3 share IOMMU group 28, type DMA-FQ. The region is listed as direct-relaxable. Only GPU USB function 01:00.2 is explicitly scoped by the RMRR. Boot logs report a broken RMRR entry for that function and application of a workaround. Confirmed observations are the firmware scopes, group membership, address overlap and NVIDIA mapping stack. The unproven hypothesis is that the USB-specific reservation is mapped in the shared domain but omitted from its IOVA allocator exclusions. Initialization order and allocator state have not been instrumented. The USB driver is not established as the mapping caller. No supported correction has been verified. Please assess applicability of the shared-domain reservation issue and identify a reviewed correction or supported backport. No additional inference or stress reproduction was performed for collection. Attachments are sanitized extracts; topology was recorded later than the original Apport snapshot. Related upstream discussion: https://lkml.iu.edu/2410.3/01077.html To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2173894/+subscriptions