Public bug reported: This series just made linux-next and is expected to land in v6.17. It adds support for device cacheable memory within a VM and is needed to support GPU PT on NVIDIA Grace systems.
Lore discussion: https://lore.kernel.org/all/[email protected]/ Upstream SHAs (linux-next): 8cc9dc1ae4fb KVM: arm64: Rename the device variable to s2_force_noncacheable 216887f79d98 KVM: arm64: Assume non-PFNMAP/MIXEDMAP VMAs can be mapped cacheable 2a8dfab26677 KVM: arm64: Block cacheable PFNMAP mapping 0c67288e0c8b KVM: arm64: Allow cacheable stage 2 mapping using VMA flags f55ce5a6cd33 KVM: arm64: Expose new KVM cap for cacheable PFNMAP We currently hold a NVIDIA SAUCE patch in the -nvidia 6.14 HWE (b244d1e5cdeb NVIDIA: SAUCE: KVM: arm64: determine memory type from VMA) that supports this function. Revert this patch and backport the series to 6.14 HWE. Note that only 1 patch (f55ce5a6cd33) did not pick cleanly due to some context collisions. These were addressed during the pick as indicated in the commit message. Tested by performing GPU PT on GB200 with 1 and 4 GPUs and running CUDA workloads within the VM. ** Affects: linux-nvidia (Ubuntu) Importance: Undecided Status: New -- You received this bug notification because you are a member of Ubuntu Bugs, which is subscribed to Ubuntu. https://bugs.launchpad.net/bugs/2116207 Title: Backport: KVM: arm64: Map GPU device memory as cacheable To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/linux-nvidia/+bug/2116207/+subscriptions -- ubuntu-bugs mailing list [email protected] https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
