Re: [QUESTION] IMA: kexec appraisal and the unauthenticated target command line

[Mimi Zohar <zohar@linux.ibm.com>]

Hi Danny,

On Fri, 2026-10-02 at 11:18 -0700, Danny Hu wrote:
> On Thu, Sep 24, 2026 at 7:27 PM Mimi Zohar <zohar@linux.ibm.com> wrote:
> > 
> > The lack of KEXEC_CMDLINE appraisal is an architectural constraint of IMA's
> > inode-based appraisal model. The kexec command line is a plain buffer passed
> > directly to kexec_file_load() via syscall. Unlike the kernel image or initramfs,
> > it is not a file. There is no natural way to associate a signature with it,
> > whether via xattr or as an appended signature. This is not an intentional
> > security decision but a consequence of how IMA appraisal works.
> 
> Would appended-signature appraisal for KEXEC_CMDLINE be a reasonable
> extension, or is it deemed out of scope for IMA? I see code comments
> stating that buffers can only be measured and not appraised and
> recognize that IMA was architectured to appraise file backed objects,
> however, it appears technically feasible to use ima_read_modsig() to
> parse an appended signature from the cmdline buffer. The signature can
> then be verified by existing modsig verification helpers, and excluded
> before validating the terminating NUL and loading the cmdline.

The original IMA buffer was limited to measuring the buffer, but that changed
with appended signature support for the kexec kernel image.  So signing the boot
command line with an appended signature should be possible.

> 
> I understand this would require extending the current appraisal path,
> policy support, and error handling. Are there additional architectural
> reasons that would make this approach unsuitable?

AI mentioned a couple of issues, but none necessarily blockers:
- Size: boot command line size limitation (defined on per arch basis), might not
be large enough for the appended signature.

- Ordering: the NUL-termination check and length handling are in generic kexec
code, so the signature would have to be stripped before that check, which means
changes outside IMA.

- Final cmdline: several arch loaders append to the buffer after it is passed
in, so the signature would cover the user-supplied portion, not necessarily what
the new kernel finally sees.

- Binding: the signed cmdline should be bound to a specific kernel; otherwise a
validly signed cmdline could be used with any kernel.

In lieu of generic cmdline signature support, a forced built-in cmdline (e.g.
CONFIG_CMDLINE_OVERRIDE on x86) lets the kernel signature cover the string, if
that fits your use case. The tradeoff is that any change to the cmdline requires
rebuilding and re-signing the kernel.
 
> 
> > 
> > The threat model concern you raise is valid — a caller with access to
> > kexec_file_load() can supply an unauthenticated command line that weakens the
> > target kernel's enforcement state, even when the kernel and initramfs are
> > appraised. This gap is real and currently unaddressed.
> > 
> > IMA continuity across kexec has not been discussed. Appraising the kernel image
> > and initramfs does not guarantee the target kernel will enforce an equivalent
> > IMA policy after boot.
> 
> I agree that appraising a kernel image alone does not guarantee
> equivalent IMA enforcement after boot. However, the deployment would
> be responsible for signing and trusting only kernels that implement
> and require IMA continuity. It would also need to exclude older
> trusted images that lack this support.
> 
> > 
> > We would welcome proposals for addressing these gaps.
> 
> An idea that does not involve signing the cmdline is to preserve IMA
> enforcement across kexec. The source kernel could hand over its active
> policy, which the target would restore during IMA initialization. To
> ensure the IMA subsystem has initialized in the target kernel, we
> could imagine a CONFIG_IMA_REQUIRE_ENABLED that ntroduces a readiness
> flag that init_ima() would set only after initialization and policy
> restoration. An unconditional check in kernel_init_freeable() would
> panic if the flag remained false, indicating that IMA has been
> tampered with.
> 
> If the inherited policy appraises /init, initramfs signatures and
> xattrs would also need to be available before execution. I saw the
> earlier initramfs xattr proposal (https://lwn.net/Articles/788922/)
> but it seemed never to have merged, so it will likely need revisiting.

Refer to https://github.com/linux-integrity/linux/issues/11 for a more recent
version.

> This idea would prevent weakening IMA enforcement across the
> transition, although the problem of broader cmdline authentication
> still remains open. Would this be a reasonable direction to
> investigate?

Any help in reviewing and testing Roberto's patch set would be appreciated.

thanks,

Mimi