Linux kernel: powerpc: KVM guest to host memory corruption

Linux kernel: powerpc: KVM guest to host memory corruption

[Michael Ellerman]

The Linux kernel for powerpc since v3.10 has a bug which allows a malicious KVM guest to
corrupt host memory.

In the handling of the H_RTAS hypercall, args.rets is made to point into the args.args
buffer which is located on the stack:

	args.rets = &args.args[be32_to_cpu(args.nargs)];

However args.nargs has not been range checked. That allows the guest to point args.rets
anywhere up to +16GB from args.args.

The guest does not have control of what is written to args.rets, it is always (u32)-3,
because subsequent code does check nargs. Additionally the guest will be killed as a
result of the nargs being out of range, so a given guest only has a single shot at
corrupting memory.

Only machines using Linux as the hypervisor, aka. KVM or bare metal, are affected by the
bug.

The bug was introduced in:

    8e591cb72047 ("KVM: PPC: Book3S: Add infrastructure to implement kernel-side RTAS calls")

Which was first released in v3.10.

The upstream fix is:

  f62f3c20647e ("KVM: PPC: Book3S: Fix H_RTAS rets buffer overflow")

  https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=f62f3c20647ebd5fb6ecb8f0b477b9281c44c10a

Which will be included in the v5.14 release.

cheers

Re: Linux kernel: powerpc: KVM guest to host memory corruption

[Michael Ellerman]

Michael Ellerman <mpe@ellerman.id.au> writes:
> The Linux kernel for powerpc since v3.10 has a bug which allows a malicious KVM guest to
> corrupt host memory.
>
> In the handling of the H_RTAS hypercall, args.rets is made to point into the args.args
> buffer which is located on the stack:
>
> 	args.rets = &args.args[be32_to_cpu(args.nargs)];
>
> However args.nargs has not been range checked. That allows the guest to point args.rets
> anywhere up to +16GB from args.args.
>
> The guest does not have control of what is written to args.rets, it is always (u32)-3,
> because subsequent code does check nargs. Additionally the guest will be killed as a
> result of the nargs being out of range, so a given guest only has a single shot at
> corrupting memory.
>
> Only machines using Linux as the hypervisor, aka. KVM or bare metal, are affected by the
> bug.
>
> The bug was introduced in:
>
>     8e591cb72047 ("KVM: PPC: Book3S: Add infrastructure to implement kernel-side RTAS calls")
>
> Which was first released in v3.10.
>
> The upstream fix is:
>
>   f62f3c20647e ("KVM: PPC: Book3S: Fix H_RTAS rets buffer overflow")
>
>   https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=f62f3c20647ebd5fb6ecb8f0b477b9281c44c10a
>
> Which will be included in the v5.14 release.

This has been assigned CVE-2021-37576.

cheers