CVE-2026-63873 Overview
CVE-2026-63873 is a memory reference leak vulnerability in the Linux kernel's AMD XDNA accelerator driver (accel/amdxdna). The flaw resides in the aie2_populate_range() function, which jumps back to the again label without calling mmput(mm). This omission leaks a reference to the mm_struct each time the retry path executes. Over time, leaked references prevent the memory descriptor from being freed and can degrade kernel memory management. The fix adds the missing mmput() call before the jump to again.
Critical Impact
Repeated invocation of the affected code path leaks mm_struct references, holding memory management structures beyond their expected lifetime and potentially leading to resource exhaustion.
Affected Products
- Linux kernel versions containing the accel/amdxdna driver prior to the fixing commits
- Systems using AMD XDNA (AI Engine) accelerator hardware
- Distributions shipping stable kernel branches referenced in the fix commits
Discovery Timeline
- 2026-07-19 - CVE-2026-63873 published to the National Vulnerability Database
- 2026-07-19 - Last updated in NVD database
Technical Details for CVE-2026-63873
Vulnerability Analysis
The vulnerability is a memory leak [CWE-401] in the AMD XDNA accelerator driver used to expose the AI Engine hardware present on AMD Ryzen AI platforms. The aie2_populate_range() function acquires a reference to a process mm_struct via get_task_mm() or an equivalent path before performing memory range population. When the function needs to retry the operation, it uses a goto again control-flow construct. The original code path failed to release the previously acquired mm_struct reference before jumping back, causing the reference count to increment without a matching decrement.
The EPSS score for this issue is 0.155%, reflecting a low probability of active exploitation. The vulnerability requires local code paths that repeatedly exercise the accelerator driver's populate routine.
Root Cause
The root cause is missing cleanup on a retry path. The function calls into memory management routines that pin an mm_struct reference, then re-enters the operation without calling mmput(mm). Each retry increments the reference counter, but the corresponding release only occurs on the terminal exit path. The upstream fix inserts an mmput(mm) call immediately before the goto again statement so the reference is balanced on every iteration.
Attack Vector
Triggering the leak requires local access to the AMD XDNA device node, typically /dev/accel/accel*, and permission to submit workloads to the AI Engine. A local user or process with driver access can repeatedly invoke code paths that force aie2_populate_range() to retry, gradually accumulating leaked mm_struct references. Sustained exploitation can degrade kernel memory accounting and prevent process teardown from releasing memory.
No public proof-of-concept exploit is available. See the kernel git commit reference for the exact patch content.
Detection Methods for CVE-2026-63873
Indicators of Compromise
- Growth in the reference count of process mm_struct objects that outlives the originating process lifetime.
- Kernel memory accounting anomalies visible in /proc/meminfo under KernelStack or slab caches for mm_struct.
- Processes stuck in exit state with unreleased memory descriptors after interacting with /dev/accel/* device nodes.
Detection Strategies
- Audit the running kernel version against the patched commits 21dfec59, 2f41af63, and e83fc4c2 referenced in the kernel git repository.
- Track workloads that hold open file descriptors to AMD XDNA accelerator device nodes and correlate with unusual slab growth.
- Enable CONFIG_DEBUG_KMEMLEAK on test systems to observe reference imbalances during driver stress testing.
Monitoring Recommendations
- Monitor kernel logs for repeated retry patterns in amdxdna driver messages.
- Alert on sustained increases in slab allocator usage on hosts with AMD Ryzen AI hardware.
- Track long-running processes that never fully release memory after accelerator workload submission.
How to Mitigate CVE-2026-63873
Immediate Actions Required
- Update the Linux kernel to a stable release that includes the fix commits 21dfec59939120b20d2c7794caaa421f9450be0a, 2f41af638c92bac6f1f9275ea2d1901baef578f3, or e83fc4c28226be75fbc0c41f2846935ba2b5f949.
- Restrict access to /dev/accel/* device nodes to trusted users and workloads through file system permissions.
- Reboot affected hosts after patching to clear any accumulated leaked references.
Patch Information
The upstream fix adds the missing mmput(mm) call before the goto again in aie2_populate_range(). Patch references are available at the primary fix commit, the stable backport commit, and the additional stable branch commit. Apply the vendor kernel package that includes these commits for your distribution.
Workarounds
- Unload the amdxdna kernel module on systems that do not require AI Engine acceleration using modprobe -r amdxdna.
- Blacklist the module via /etc/modprobe.d/ to prevent automatic loading until the kernel is patched.
- Limit local user access to accelerator device nodes with strict Unix permissions or udev rules.
# Prevent the amdxdna module from loading until the kernel is patched
echo "blacklist amdxdna" | sudo tee /etc/modprobe.d/blacklist-amdxdna.conf
sudo modprobe -r amdxdna 2>/dev/null || true
sudo update-initramfs -u
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
