Public bug reported: [Impact]
A kdump crash kernel resets every SMMU while devices are still doing DMA set up by the crashed kernel. Those transactions then fail or land in the wrong memory: poisoned TLPs, completion timeouts, or a lost dump. linux-nvidia-7.3 has no fix. The 7.0 kernels carried an older version of the series (v6, LP: #2156531), which 7.3 dropped. [Fix] Backport v12, which adopts the crashed kernel's stream table and keeps the SMMU enabled until each device's driver attaches a new domain. Prerequisite from mainline: aa416fa911ad iommu: Finalize deferred attachments when mapping MSI pages Backported from https://lore.kernel.org/linux-iommu/[email protected]/ (12 patches). Plus two SAUCE patches: block MPAM updates on a device that is still using the crashed kernel's stream table entry, and move ARM_SMMU_FEAT_MPAM off the bit it shares with ARM_SMMU_FEAT_HAFT. A config change records the new ARM_SMMU_V3_KEXEC symbol in the annotations. Tested with kdump under GPU DMA load on Vera and on GH200: DMA keeps landing in host memory until the dump is captured, and no errors are logged. ** 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/2172875 Title: [linux-nvidia-7.3]: Backport the arm-smmu-v3 kdump stream table adoption series (v12) To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/linux-nvidia/+bug/2172875/+subscriptions -- ubuntu-bugs mailing list [email protected] https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
