In general, a (non-ideal) workaround is to disable KASLR. For the CUDA use-case, there is a workaround of setting uvm_disable_hmm=1 for the nvidia_uvm kernel module.
-- You received this bug notification because you are a member of Kernel Packages, which is subscribed to linux in Ubuntu. https://bugs.launchpad.net/bugs/2120209 Title: x86 systems with PCIe BAR addresses located outside a certain range see P2PDMA allocation failures and CUDA initialization errors Status in linux package in Ubuntu: Invalid Status in linux source package in Jammy: In Progress Status in linux source package in Noble: In Progress Status in linux source package in Plucky: In Progress Bug description: SRU Justification [Impact] On some x86 systems, it is possible for PCIe device BAR addresses to exceed the range reserved by KASLR for direct mappings. This causes attempts to map the impacted BAR region using devm_memremap_pages() to fail. These memmap-backed mappings are required for multiple use-cases, including P2PDMA, and CUDA with Heterogeneous Memory Management (HMM) enabled. [Fix] This is resolved upstream by commit 7ffb791423c7 ("x86/kaslr: Reduce KASLR entropy on most x86 systems"). It changes the behavior of KASLR to not shrink direct mapping space when CONFIG_PCI_P2PDMA is enabled. The consequence of this is that there is less room for KASLR to maneuver, and thus the amount of entropy in the randomized layout is reduced. In discussion on the upstream patch submission [1], it is noted that on the submitter's system this reduces entropy from 16 bits down to 15 bits. Cherry-picking the mentioned commit allows CUDA with HMM enabled and P2PDMA to function on the systems described above, as with it the direct mapping space is not shrunk, so all BAR regions fall within its bounds, and thus the devm_memremap_pages() operation succeeds. Jammy 5.15 has CONFIG_PCI_P2PDMA set to n, so a cherry-pick alone will not resolve the issue. In addition to the cherry-pick, set CONFIG_PCI_P2PDMA=y. Jammy: Cherry-pick and CONFIG_PCI_P2PDMA=y needed Noble: Cherry-pick needed Plucky: Cherry-pick needed Questing: Not affected, fix commit already in tree. [Test Case] The issue only occurs on systems with PCIe BAR addresses located outside of the current minimum address range of [0, ceil(max_pfn / 1TiB) + CONFIG_RANDOMIZE_MEMORY_PHYSICAL_PADDING (10 TiB)]. With the NVIDIA Container Toolkit installed and enabled for Docker, the following reproduces the issue on affected systems where one or more NVIDIA GPUs have BAR addresses outside of the current minimum range: $ sudo docker run --runtime nvidia --rm -it nvcr.io/nvidia/pytorch:25.03-py3 ERROR: The NVIDIA Driver is present, but CUDA failed to initialize. GPU functionality will not be available. [[ Initialization error (error 3) ]] [Where things could go wrong] This reduces the entropy of the memory layouts KASLR generates on most x86 systems. A bug would likely show up as misbehavior of KASLR. On Jammy, this changeset also enables CONFIG_PCI_P2PDMA, which could have additional side-effects. There is an LP bug [2] noting the change of CONFIG_PCI_P2PDMA in newer kernels. [Other Notes] [1] https://lore.kernel.org/lkml/202502061145.8AFAF053E4@keescook/ [2] https://bugs.launchpad.net/bugs/1987394 To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2120209/+subscriptions -- Mailing list: https://launchpad.net/~kernel-packages Post to : [email protected] Unsubscribe : https://launchpad.net/~kernel-packages More help : https://help.launchpad.net/ListHelp
