CVE-2026-89811: drm/amdkfd: Add TLB flush after MES queue eviction/suspension
In the Linux kernel, the following vulnerability has been resolved:
drm/amdkfd: Add TLB flush after MES queue eviction/suspension
MES (Micro Engine Scheduler) does not perform heavy-weight TLB invalidation after unmapping queues, unlike HWS which does this automatically. This causes a race condition where in-flight DMA descriptors can access memory that has been unmapped, leading to page faults and GPU queue hangs during SVM page migration.
The issue manifests as KFDSVMRangeTest.MultiThreadMigrationTest failures on gfx1151 (Strix Point) with XNACK mode 1 enabled - the GPU compute queue hangs with packets submitted but never consumed.
Add kfdflushtlb() calls after MES queue removal in two locations: - evictprocessqueuescpsch(): after all queues removed during eviction - suspendqueues(): after debug/criu queue suspension (with memfence barrier)
This ensures all in-flight memory accesses from unmapped queues are flushed before memory is freed or migrated.
(cherry picked from commit f5c4f88e0f9c45a8fb9dfac0c1df726c95e41b77)
Affected Software
Remediation
Recommended actions to resolve this vulnerability, in priority order.
- Configuration
Update the Linux kernel drm/amdkfd code to add kfd_flush_tlb() after MES queue removal in both specified locations: (1) after MES queue eviction removes all queues (in evict_process_queues_cpsch(): after all queues removed during eviction), and (2) after debug/criu queue suspension with mem_fence barrier (in suspend_queues(): after debug/criu queue suspension with mem_fence barrier) to ensure in-flight memory accesses from unmapped queues are flushed before memory is freed or migrated.
drm/amdkfd (Linux kernel) TLB flush after MES queue eviction/suspension = Add kfd_flush_tlb() calls after MES queue removal in two locations (after MES queue eviction and after queue suspension/debug-criu queue suspension with mem_fence barrier)
Event History
Frequently Asked Questions
Which hardware and configuration are known to exhibit this issue?
The issue is reported on gfx1151 (Strix Point) systems with XNACK mode 1 enabled. It occurs in the AMD KFD path when MES-managed GPU queues are involved in SVM page migration.
How can an affected system present operationally?
Affected systems can show KFDSVMRangeTest.MultiThreadMigrationTest failures, GPU compute queues with submitted packets that are never consumed, page faults, and GPU queue hangs.
Which queue-management operations are implicated?
The race is associated with MES queue eviction and suspension after queues are unmapped. In-flight DMA descriptors may still access the unmapped memory until a TLB flush is performed.