[Bug 2160704] [NEW] AMD-Vi "Completion-Wait loop timed out" IOMMU hang → full system freeze on 6.8.0-134; regression vs 6.8.0-124 (Ryzen 7 PRO 5850U)

Public bug reported: [Impact] On linux-image-6.8.0-134-generic the AMD IOMMU (AMD-Vi) begins logging "AMD-Vi: Completion-Wait loop timed out" within hours of boot, and the machine then hard-freezes every 3-6 days requiring a physical power cycle: no SSH, dead console, no clean shutdown recorded. The final kernel log line before each freeze is that AMD-Vi timeout. This is a REGRESSION. The immediately prior kernel, 6.8.0-124-generic, is completely clean on the same hardware and identical workload. [Regression evidence: count of "Completion-Wait loop timed out" per boot] 6.8.0-124-generic (Jun 17 -> Jul 4, 17 days) : 0 occurrences, stable 6.8.0-134-generic (Jul 5 -> Jul 11) : 83 occurrences, ended in hard freeze 6.8.0-134-generic (Jul 11 -> Jul 14) : 43 occurrences, ended in hard freeze 6.8.0-134-generic (current boot) : 1 occurrence ~18 min after boot The errors begin a few hours after booting -134 and accumulate until lockup. [Hardware] - Beelink SER5 mini PC - AMD Ryzen 7 PRO 5850U (Cezanne, integrated Radeon Vega / RENOIR amdgpu) - 32 GB RAM, BIOS dated 2024-09-06 - No VFIO / PCI passthrough. No custom kernel command line. - IOMMU default config: "Default domain type: Translated", "DMA domain TLB invalidation policy: lazy mode". [Ruled out] Not OOM (memory healthy, systemd-oomd inactive, zero OOM kills), not disk, not thermal (the thermal_zone0 read failure appears on the good -124 boots too). [Steps to reproduce] 1. Boot 6.8.0-134-generic on the above hardware. 2. Run a normal container / desktop workload (rootless podman, amdgpu active). 3. Within hours "AMD-Vi: Completion-Wait loop timed out" appears in dmesg; within 3-6 days the machine freezes hard. [Expected] No IOMMU completion-wait timeouts and no freeze, as on 6.8.0-124-generic. [Workaround] Reverting to 6.8.0-124-generic eliminates the errors entirely. "amd_iommu=off" is expected to also work (no passthrough in use). [Request] Please bisect the AMD IOMMU changes between 6.8.0-124 and 6.8.0-134, or advise if this is a known regression with a fix in flight. ** Affects: linux (Ubuntu) Importance: Undecided Status: New ** Tags: regression-update ** Attachment added: "Kernel log, 6.8.0-134 boot of Jul 11-14 that ended in a hard freeze — AMD-Vi Completion-Wait timeouts throughout; final kernel line is the timeout." https://bugs.launchpad.net/bugs/2160704/+attachment/5981991/+files/dooku-amdvi-134-prev.txt -- You received this bug notification because you are subscribed to linux in Ubuntu. Matching subscriptions: Bgg, Bmail, Nb https://bugs.launchpad.net/bugs/2160704 Title: AMD-Vi "Completion-Wait loop timed out" IOMMU hang → full system freeze on 6.8.0-134; regression vs 6.8.0-124 (Ryzen 7 PRO 5850U) Status in linux package in Ubuntu: New Bug description: [Impact] On linux-image-6.8.0-134-generic the AMD IOMMU (AMD-Vi) begins logging "AMD-Vi: Completion-Wait loop timed out" within hours of boot, and the machine then hard-freezes every 3-6 days requiring a physical power cycle: no SSH, dead console, no clean shutdown recorded. The final kernel log line before each freeze is that AMD-Vi timeout. This is a REGRESSION. The immediately prior kernel, 6.8.0-124-generic, is completely clean on the same hardware and identical workload. [Regression evidence: count of "Completion-Wait loop timed out" per boot] 6.8.0-124-generic (Jun 17 -> Jul 4, 17 days) : 0 occurrences, stable 6.8.0-134-generic (Jul 5 -> Jul 11) : 83 occurrences, ended in hard freeze 6.8.0-134-generic (Jul 11 -> Jul 14) : 43 occurrences, ended in hard freeze 6.8.0-134-generic (current boot) : 1 occurrence ~18 min after boot The errors begin a few hours after booting -134 and accumulate until lockup. [Hardware] - Beelink SER5 mini PC - AMD Ryzen 7 PRO 5850U (Cezanne, integrated Radeon Vega / RENOIR amdgpu) - 32 GB RAM, BIOS dated 2024-09-06 - No VFIO / PCI passthrough. No custom kernel command line. - IOMMU default config: "Default domain type: Translated", "DMA domain TLB invalidation policy: lazy mode". [Ruled out] Not OOM (memory healthy, systemd-oomd inactive, zero OOM kills), not disk, not thermal (the thermal_zone0 read failure appears on the good -124 boots too). [Steps to reproduce] 1. Boot 6.8.0-134-generic on the above hardware. 2. Run a normal container / desktop workload (rootless podman, amdgpu active). 3. Within hours "AMD-Vi: Completion-Wait loop timed out" appears in dmesg; within 3-6 days the machine freezes hard. [Expected] No IOMMU completion-wait timeouts and no freeze, as on 6.8.0-124-generic. [Workaround] Reverting to 6.8.0-124-generic eliminates the errors entirely. "amd_iommu=off" is expected to also work (no passthrough in use). [Request] Please bisect the AMD IOMMU changes between 6.8.0-124 and 6.8.0-134, or advise if this is a known regression with a fix in flight. To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2160704/+subscriptions