HIGH Introduced in 5.19
amdkfd MESQueue Race
CVE-2026-89811
CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H
KernelScan AI4.7MEDIUM
01Description
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 kfd_flush_tlb() calls after MES queue removal in two locations: - evict_process_queues_cpsch(): after all queues removed during eviction - suspend_queues(): after debug/criu queue suspension (with mem_fence barrier) This ensures all in-flight memory accesses from unmapped queues are flushed before memory is freed or migrated. (cherry picked from commit f5c4f88e0f9c45a8fb9dfac0c1df726c95e41b77)
02KernelScan AI Analysis
Risk summary
A local user with access to an affected AMD GPU can trigger a race condition during SVM page migration that causes in-flight DMA accesses to unmapped memory, resulting in GPU queue hangs and system instability. The issue requires specific AMD hardware (gfx1151/Strix Point) with MES enabled and XNACK mode 1.
Vulnerability analysis
When the MES scheduler removes or suspends GPU compute queues, it does not perform a heavy-weight TLB invalidation, unlike the older HWS scheduler. This allows in-flight DMA descriptors from those queues to keep accessing memory that has already been unmapped or migrated, causing page faults and permanent GPU queue hangs. The fix adds explicit TLB flush calls after MES queue removal and suspension, ensuring all outstanding memory accesses complete before the underlying pages are freed or migrated. The vulnerable path is reachable locally by any user with access to the AMD GPU compute device who can trigger SVM page migration operations.
Lifecycle
03Fix Versions
| Branch | Introduced | Fixed in | Patch commit |
|---|---|---|---|
| 6.12 | 5.19 | 6.12.114 | bbf5ad5268f3 |
| 6.18 | 5.19 | 6.18.51 | e230c546ed93 |
| 6.6 | 5.19 | 6.6.160 | 5bbd7f283312 |
| 7.2 | 5.19 | 7.2.5 | 6aae545c8708 |
| mainline | 5.19 | 7.3 | 94e25cb6ab7f |
04CIP Kernels
State of the newest release of each Civil Infrastructure Platform SLTS tree, from CIP's own fix records.
| CIP tree | State |
|---|---|
| cip/6.12 | Affected — no CIP fix yet (latest 6.12.112-cip33) |
| cip/6.1 | Affected — no CIP fix yet (latest 6.1.189-cip61) |
| cip/6.1-rt | Affected — no CIP fix yet (latest 6.1.189-cip61-rt32) |