News › linux
By The Linux Camp News, October 9, 2026
The October 8 proposal would authenticate restored system state rather than disable lockdown's protections. Its TPM trust design remains open for review, and the supplied sources establish no shipping kernel version.
On October 8, 2026, NVIDIA's Matthew Garrett submitted a Linux request for comments proposing hibernation under kernel lockdown using TPM-backed image authentication. The series contains 17 patches, ending with conditional permission to hibernate using signed images. This is a proposed capability, not a release announcement: the sources retrieved on October 9 establish neither a merge nor a shipping kernel version. 1 3 2
The restriction exists because resuming is more than reading saved application data. As Garrett explains, hibernation writes system state to disk, then restores it into RAM and transfers execution to that state. Without authentication, root could alter the swap contents and make resume load attacker-chosen kernel-space data. Lockdown therefore blocks this route across its trust boundary. The proposal aims to prove both that the kernel produced the image and that nobody changed it afterward. 2 1
The first building block is a kernel-only TPM nonvolatile index, a persistent storage location in the Trusted Platform Module. Reserving it gives the design TPM state that userspace is not supposed to control. That reservation is unfinished: Garrett says the proposed index is a placeholder and needs an explicitly registered allocation to avoid conflicts with userspace software. 2
The next problem is proving that the booted kernel actually implements that protection. Garrett proposes updating PCR 5, a TPM measurement register, from the boot stub before ExitBootServices. Firmware then extends the register as boot services end. In his design, that ordering prevents userspace from reproducing the earlier proof step after boot. It is intended to distinguish a kernel supporting the reservation from one that does not, rather than simply trusting a claim made by userspace. 2
Audited TPM sessions then connect that protected environment to the image-signing key. These sessions produce signed digests, or cryptographic summaries, of the commands executed and their results. The series creates a signing key within such a session and uses the audit evidence to establish that the session took place in the kernel. That key signs the hibernation image. The important distinction is provenance: an arbitrary signature would not establish that the saved state came from the kernel. 2
The intended operational benefit is hibernation without abandoning lockdown, including Secure Boot deployments that enable lockdown and systems where administrators enable it manually. This is a signed-image exception, not removal of lockdown's other restrictions. Virtual-machine operators also face an explicit boundary: patch 16 proposes refusing image verification with a virtual TPM. The supplied archive includes that patch's subject, but not its implementation or rationale. 1 2
The security design still has an acknowledged weak point to resolve. Garrett warns that someone able to send TPM commands directly could violate its assumptions, making userspace command filtering a critical dependency. He asks whether a secret inaccessible to userspace, but consistent across platform reboots, could instead authorize the protected operations. This is an open design question from the author, not evidence that reviewers have validated the proposed trust chain. 2
For deployment planning, treat this as work to evaluate rather than a feature to enable from the announcement. The available material provides no supported configuration checklist or release target. Also keep image authentication separate from confidentiality: the described operation is signing and verification, so it does not establish that saved memory is encrypted. Operators assessing protection for sensitive data on disk still need that question answered independently. 1 2