CVE-2026-89810 PUBLISHED

drm/amdkfd: Fix error path at svm_migrate_copy_to_ram

Assigner: Linux
Reserved: 11.09.2026 Published: 16.09.2026 Updated: 16.09.2026

In the Linux kernel, the following vulnerability has been resolved:

drm/amdkfd: Fix error path at svm_migrate_copy_to_ram

If page migration from device to sys ram fails for some reasons driver needs release and unlock allocated system pages. To do that driver should use page physical address, or pfn, then get struct page*. Current driver uses dma address(for adev) that is not correct with IOMMU enabled, or even in general.

The patch releases and unlocks allocated system pages based on where migration failed by struct page* of sys ram pages. Also dma_unmap correspodent system ram pages at error path.

Metrics

CVSS Vector: CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H
CVSS Score: 7.8

AV:L - The bug is in amdkfd SVM VRAM-to-RAM migration (svm_migrate_copy_to_ram), reached from local /dev/kfd AMDKFD_IOC_SVM ioctls, CPU/GPU page faults, or TTM eviction in the caller’s process; it is not reachable from the network stack. AC:L - An attacker who can open /dev/kfd fully controls SVM prefetch and faults and can force the failing copy path (alloc_page_vma ENOMEM, dma_map_page failure, or GART IB allocation failure) via memory or DMA-mapping pressure; no uncontrollable race or rare config is required. PR:L - kfd_open() and AMDKFD_IOC_SVM perform no capability check; /dev/kfd and the amdgpu render node are accessible to unprivileged users in the render group, which is standard on AMD GPU, ROCm, HPC, and embedded systems. UI:N - The attacker opens /dev/kfd, establishes an SVM range, and triggers VRAM-to-RAM migration from their own process; no other user must mount, open, or interact with anything. S:U - Impact is kernel page-refcount and vmemmap corruption inside the host kernel that runs amdkfd. This is standard local kernel memory corruption, not a guest-to-host escape or IOMMU DMA-isolation bypass. C:H - On the error path the driver treats DMA addresses (IOMMU IOVAs, bounce-buffer addresses, or 0 for skipped pages) as PFNs and put_page()s the resulting struct page, a use-after-free/refcount underflow of unrelated pages that can disclose kernel memory. I:H - The same wrong-page unlock_page()+put_page() can free still-mapped pages or drop the lock on pages other code holds, and invalid PFNs write through pfn_to_page() into vmemmap; that is an exploitable write/UAF primitive for control-flow hijack. A:H - pfn_to_page() without pfn_valid() on a DMA IOVA, unlock_page() of an unlocked page, and later use of an underflowed page refcount each produce a kernel oops or panic, fully denying availability.

Product Status

Vendor Linux
Product Linux
Versions Default: unaffected
  • affected from 1da177e4c3f41524e886b7f1b8a0c1fc7321cac2 to da87bcad1f781d822e7ced6d1de9dbc6d381c72e (excl.)
  • affected from 1da177e4c3f41524e886b7f1b8a0c1fc7321cac2 to a2035918a7a1d8fd3bb79d22e1e149294a5a8418 (excl.)
  • affected from 1da177e4c3f41524e886b7f1b8a0c1fc7321cac2 to 960c4a8069bfd352c48cc88592618f1ebe24c69e (excl.)
  • affected from 0 to 6.18.51 (excl.)
  • affected from 0 to 7.2.5 (excl.)
Vendor Linux
Product Linux
Versions Default: affected
  • unaffected from 6.18.51 to 6.18.* (incl.)
  • unaffected from 7.2.5 to 7.2.* (incl.)
  • unaffected from 7.3-rc1 to * (incl.)

References