AMD KFD CVE-2026-89810

Requires a system running the AMD KFD (amdkfd) driver with GPU Shared Virtual Memory (SVM) migration in use - typically ROCm/HSA compute workloads - AND an enabled IOMMU, which is the precise condition under which the DMA (bus) address diverges from the physical address so the wrong struct page is resolved. The vulnerable cleanup executes ONLY on the error path of svm_migrate_copy_to_ram, i.e. after a device-to-system-RAM page migration has already failed, so an attacker must be able to induce such a failure (memory pressure / allocation or copy failure / race). Exploitation is local (AV:L) and needs a local account able to reach the KFD compute interface (PR:L); there is no remote or unauthenticated path and no user interaction. Absence of IOMMU or of the SVM migration workload removes the wrong-page mismatch and the practical risk.

The vendor CVSS of 7.8 (AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H) is a local-privilege full-triad memory-corruption score, but the description shows the bug is reachable only on the error path of svm_migrate_copy_to_ram - the device-to-system-RAM migration must actually FAIL before the faulty cleanup runs, which realistically warrants AC:H rather than AC:L. There is no EPSS score, no CISA KEV listing, and no public exploit code, so exploitation status is 'no public exploit identified at time of analysis' - this is a code-level correctness fix, not evidence of active attack. The dominant real-world limiter is not privilege (an ordinary local user with AMD KFD/ROCm compute access qualifies as PR:L) but the need to induce a migration failure (ENOMEM/IOMMU pressure/race) on a system that both uses an AMD GPU with SVM and has IOMMU enabled - exactly the condition under which the DMA-address-vs-physical-address mismatch produces the wrong struct page. Net: a genuine but conditional local memory-corruption bug, patch-worthy on affected AMD compute hosts, but not a broadly automatable or opportunistic threat. The input 'Information Disclosure' tag is inconsistent with the C:H/I:H/A:H vector - this is memory corruption, not disclosure - so treat the tag, not the vector, as the error.

A local user with access to the AMD KFD compute interface (e.g. a ROCm/HSA workload) on an IOMMU-enabled host repeatedly drives shared-virtual-memory migrations from GPU device memory back to system RAM under memory pressure so that a copy fails. On the failure path the driver, using the DMA (bus) address instead of the physical pfn, releases and unlocks the wrong struct pages and skips the correct dma_unmap, corrupting kernel page state and leaking IOMMU mappings. No public exploit code exists; achieving a controllable outcome would require winning the error-path timing and grooming page state.