[Bug 2164934] [NEW] [SRU] Add TLB flush after MES queue eviction/suspension in amdkfd (fixes GPU hang on SVM migration, gfx1151/Strix Point)

Public bug reported: [Impact] On AMD GPUs using MES (Micro Engine Scheduler), queue suspension does not perform a heavy-weight TLB invalidation. This allows DMA descriptors to access memory that has already been unmapped, causing page faults and GPU queue hangs during SVM (Shared Virtual Memory) page migration. Observed as KFDSVMRangeTest.MultiTHreadMigrationTest failures on gfx1151 (Strix Point) with XNACK mode 1 enabled: the runtime queue hangs with packets submitted but never consumed. [Fix] As suggested by the AMD team: upstream commit (https://lore.kernel.org/amd-gfx/20260816172608.14470-2-Priya.Hosur@amd.com/) (amd-gfx mailing list, Priya Hosur): "drm/amdkfd: Add TLB flush after MES queue eviction/suspension" Message-Id: <20260816172608.14470-2-Priya.Hosur@amd.com> Adds kfd_flush_tlb() calls after MES queue removal in: - evict_process_queues_cpsch() - suspend_queues() (with mem_fence barrier) This ensures all in-flight memory accesses from unmapped queues are completed/flushed before the underlying memory is freed or fully migrated. [Where problems could occur] This change adds and extra kfd_flush_tlb() fall, gated behind dqp->dev->kfd->shared_resources.enable_mes, which means it only affects MES enabled devices from newer AMD GPU generations. The risk is limited to a possible minor latency increase during queue suspension on MES enabled hardware. [Other Info] Patch is currently in review upstream, not yet merged into the Linus' tree. Requesting it to be tracked once it lands, or considered for inclusion in Ubuntu kernel. Change-Id: Iaa884d4a7ad1199446bc45f4ad8a9a179ab386e6 ** Affects: linux (Ubuntu) Importance: Undecided Status: New -- You received this bug notification because you are subscribed to linux in Ubuntu. Matching subscriptions: Bgg, Bmail, Nb https://bugs.launchpad.net/bugs/2164934 Title: [SRU] Add TLB flush after MES queue eviction/suspension in amdkfd (fixes GPU hang on SVM migration, gfx1151/Strix Point) Status in linux package in Ubuntu: New Bug description: [Impact] On AMD GPUs using MES (Micro Engine Scheduler), queue suspension does not perform a heavy-weight TLB invalidation. This allows DMA descriptors to access memory that has already been unmapped, causing page faults and GPU queue hangs during SVM (Shared Virtual Memory) page migration. Observed as KFDSVMRangeTest.MultiTHreadMigrationTest failures on gfx1151 (Strix Point) with XNACK mode 1 enabled: the runtime queue hangs with packets submitted but never consumed. [Fix] As suggested by the AMD team: upstream commit (https://lore.kernel.org/amd-gfx/20260816172608.14470-2-Priya.Hosur@amd.com/) (amd-gfx mailing list, Priya Hosur): "drm/amdkfd: Add TLB flush after MES queue eviction/suspension" Message-Id: <20260816172608.14470-2-Priya.Hosur@amd.com> Adds kfd_flush_tlb() calls after MES queue removal in: - evict_process_queues_cpsch() - suspend_queues() (with mem_fence barrier) This ensures all in-flight memory accesses from unmapped queues are completed/flushed before the underlying memory is freed or fully migrated. [Where problems could occur] This change adds and extra kfd_flush_tlb() fall, gated behind dqp->dev->kfd->shared_resources.enable_mes, which means it only affects MES enabled devices from newer AMD GPU generations. The risk is limited to a possible minor latency increase during queue suspension on MES enabled hardware. [Other Info] Patch is currently in review upstream, not yet merged into the Linus' tree. Requesting it to be tracked once it lands, or considered for inclusion in Ubuntu kernel. Change-Id: Iaa884d4a7ad1199446bc45f4ad8a9a179ab386e6 To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2164934/+subscriptions