Look up another CVE · open CVE-2026-97496 in the interactive tracker
CVE-2026-97496
drm/amdkfd: Fix OOB memory exposure in get_wave_state()
Published 2026-09-24 (24 days ago) by the Linux kernel CNA. Subsystem drivers/gpu. Debian urgency: not yet assigned.
What to do, per Debian release
Debian's own status for CVE-2026-97496 in every release this site tracks, and what each one means in practice. Every status below is the one the Debian Security Team publishes for this CVE. source
| Release | Status | Version | What this means |
|---|---|---|---|
| Debian 12bookworm | vulnerable | 6.1.190-1 | No fix published anywhere yet, neither in Debian nor upstream for the series bookworm tracks. Watch the Debian tracker page for this CVE. Advisories for bookworm have historically landed a median of 23 days after publication. |
| Debian 13trixie | fixedDSA-6528-1 · 2026-09-29 · text | 6.12.111-1 | Update the kernel and reboot. DSA-6528-1 shipped the fix on 2026-09-29, in 6.12.111-1. A running kernel keeps the old code until the machine restarts. |
| Debian 14forky | fixed | 7.2.6-1 | Update the kernel and reboot. Fixed in 7.2.6-1. A running kernel keeps the old code until the machine restarts. |
| Debian unstablesid | fixed | 7.2.6-1 | Update the kernel and reboot. Fixed in 7.2.6-1. A running kernel keeps the old code until the machine restarts. |
| Debian 12bookworm · linux-6.12 | fixedDLA-4817-1 · 2026-10-04 · text | 6.12.111-1~deb12u1 | Update the kernel and reboot. DLA-4817-1 shipped the fix on 2026-10-04, in 6.12.111-1~deb12u1. A running kernel keeps the old code until the machine restarts. |
The vulnerability
What the bug is
The get_wave_state() function for v9 trusts cp_hqd_cntl_stack_size and cp_hqd_cntl_stack_offset values read directly from the MQD, which are written by GPU microcode and fully attacker-controlled on the CRIU-restore path (via AMDKFD_IOC_RESTORE_PROCESS with H3). this leads to an unbounded copy_to_user() that can leak adjacent GTT/kernel memory. If offset > size, integer underflow produces a ~4 GiB read length, if size is set to 1 MiB against a 4 KiB allocation, we leak 1 MiB of adjacent kernel memory (other queues' MQDs, ring buffers, KASLR pointers). Fix by clamping both cp_hqd_cntl_stack_size to the actual allocated buffer size (q->ctl_stack_size) and cp_hqd_cntl_stack_offset to the clamped size before performing arithmetic and copy_to_user(). This ensures we never read beyond the allocated kernel BO regardless of attacker-supplied MQD field values.
Where it lives
The fix touches these files. If the subsystem is one you do not use, the practical exposure is lower, though the package is still the vulnerable one.
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c
- What this is
- AMD KFDMAINTAINERS section matched on drivers/gpu/drm/amd/amdkfd/ · amd-gfx@lists.freedesktop.org · Supported
Show whether Debian builds this, per release
Not established. This site could not find the build rule for the affected file in the kernel trees Debian ships, so it has no answer about whether Debian compiles this code. Read that as unknown, never as absent.
- bookworm, trixie, forky, sid, bookworm · linux-6.12?
- ?
- not established - no Makefile entry for the affected file
Where a release above reads "not established", no build rule for the affected file was found in that release's kernel tree. That is a gap in this lookup, not a finding that the code is absent.
From the configuration Debian publishes for every kernel flavour it builds. It covers only the architectures Debian ships a linux-config package for, and nothing about a kernel you built yourself. Debian's own status above stays authoritative about the source package.MAINTAINERS
Signals
- Published
- 2026-09-24 (24 days ago)source
- CVSS
- 7.1 high
CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:N/A:Hsource recompute it - Attack vector
- local (AV:L)source
- EPSS
- 0.12%, higher than 2.0% of all CVEssource
- CISA KEV
- not listedthe KEV catalogue
- Debian urgency
- not yet assignedsource
- Subsystem
- drivers/gpu
Upstream fixes
The versions the CNA records as introducing and fixing this, and the commit for each. source
| Introduced | Fixed in | Commit |
|---|---|---|
| 5.18 | 6.6.158 | e2bef0f168be |
| 5.18 | 6.12.111 | d9183d974ddd |
| 5.18 | 6.18.53 | ec646686613d |
| 5.18 | 7.2 | 7ef144458f48 |
Debian advisories
| Advisory | Date | Shipped to | Debian |
|---|---|---|---|
| DLA-4817-1 | 2026-10-04 | bookworm 6.12.111-1~deb12u1 | tracker |
| DSA-6528-1 | 2026-09-29 | trixie 6.12.111-1 | tracker |
Why the CNA scored it this way
Quoted from the kernel CNA's own scoring file, word for word. A highlighted metric carries the most severe value CVSS v3.1 defines for it, which is what pushes the score up. source
- AV:LAttack vector, local
- Attacker-controlled cp_hqd_cntl_stack_size/offset enter via AMDKFD_IOC_CRIU_OP KFD_CRIU_OP_RESTORE (kfd_ioctl_criu → criu_restore → kfd_criu_restore_queue → restore_mqd memcpy of the user MQD), then v9 get_wave_state() runs from kfd_ioctl_get_queue_wave_state (AMDKFD_IOC_GET_QUEUE_WAVE_STATE) on /dev/kfd; no remote protocol carries those fields.
- AC:LAttack complexity, low
- restore_mqd() installs the crafted MQD then sets qp->is_active=0, so AMDKFD_IOC_GET_QUEUE_WAVE_STATE deterministically passes dqm get_wave_state()'s !is_active and cwsr_enabled checks (cwsr_enable defaults to 1) and v9 get_wave_state() copy_to_user()s using unclamped size-offset; no race or firmware quirk is required.
- PR:LPrivileges required, low
- AMDKFD_IOC_GET_QUEUE_WAVE_STATE has ioctl flags 0, but the MQD size/offset that make get_wave_state() over-read are attacker-controlled only after KFD_CRIU_OP_RESTORE; kfd_ioctl() admits that op only with capable(CAP_CHECKPOINT_RESTORE) or capable(CAP_SYS_ADMIN), a delegated non-root cap routinely granted to CRIU/container checkpoint services.
- UI:NUser interaction, none
- The attacker opens /dev/kfd and issues KFD_CRIU_OP_RESTORE then AMDKFD_IOC_GET_QUEUE_WAVE_STATE against its own KFD process; no other user must mount a filesystem, open a file, or otherwise interact.
- S:UScope, unchanged
- The unbounded copy_to_user() from the kernel MQD control-stack BO in v9 get_wave_state() discloses host kernel/GTT memory inside the same kernel security authority; no VM, IOMMU, or sandbox boundary is crossed.
- C:HConfidentiality impact, high
- Pre-fix v9 get_wave_state() copy_to_user()s (cp_hqd_cntl_stack_size - cp_hqd_cntl_stack_offset) bytes from mqd+AMDGPU_GPU_PAGE_SIZE without clamping to q->ctl_stack_size, so a restored 1MiB size against a 4KiB allocate_mqd() BO (or offset>size underflow to ~4GiB) dumps adjacent GTT/kernel memory including other MQDs and KASLR pointers.
- I:NIntegrity impact, none
- The defective operation is copy_to_user() from the kernel control-stack BO to the caller's ctl_stack_address; restore_mqd() writes only sizeof(struct v9_mqd) plus ctl_stack_size into the buffer allocate_mqd() sized from q->ctl_stack_size, so this bug does not modify kernel memory out of bounds.
- A:HAvailability impact, high
- A restored cp_hqd_cntl_stack_size of ~1MiB or an offset>size underflow to ~4GiB makes get_wave_state() copy_to_user() walk off the small kmap/vmap GTT/VRAM BO into unmapped kernel VA (and can trip CONFIG_HARDENED_USERCOPY), producing a kernel oops.