Re: [PATCH] KVM: arm64: Check the untrusted offset in FF-A memory share

[Sebastian Ene <sebastianene@google.com>]

On Wed, Oct 22, 2025 at 04:21:41PM +0100, Vincent Donnefort wrote:
> On Fri, Oct 17, 2025 at 07:57:10AM +0000, Sebastian Ene wrote:
> > Verify the offset to prevent OOB access in the hypervisor
> 
> I believe that would be just a read, so probably it would be difficult to use
> this to compromise anything, except crashing the system?

The simplest way is to crash the system but a more advanced one might
lead to a confused deputy attack:

1. Use the original bug to trigger the overflow of the offset variable
which bypasses this check:
https://elixir.bootlin.com/linux/v6.18-rc2/source/arch/arm64/kvm/hyp/nvhe/ffa.c#L519

2. Use the host_share_hyp from the host to create a mapping in the hyp
address space so that : reg from reg = (void *)buf + offset; points to
memory mapped in the hyp address space & controlled from the host.

3. Make the __ffa_host_share_ranges fail (since we control the content of
the reg) to trigger the recovery mechanism for __ffa_host_unshare_ranges
(https://elixir.bootlin.com/linux/v6.18-rc2/source/arch/arm64/kvm/hyp/nvhe/ffa.c#L392)
and replace the content of the reg with pages that we want to remove the
host stage-2 FF-A annotation from.

With step(3) we can remove the host stage-2 FF-A annotation from pages
without having to invoke the FF-A reclaim mechanism. This allows a
confused deputy attack because the pages can be given to another entity
after the annotation is removed (eg. given to a protected VM).

> 
> > FF-A buffer in case an untrusted large enough value
> > [U32_MAX - sizeof(struct ffa_composite_mem_region) + 1, U32_MAX]
> > is set from the host kernel.
> > 
> > Signed-off-by: Sebastian Ene <sebastianene@google.com>
> > ---
> >  arch/arm64/kvm/hyp/nvhe/ffa.c | 9 +++++++--
> >  1 file changed, 7 insertions(+), 2 deletions(-)
> > 
> > diff --git a/arch/arm64/kvm/hyp/nvhe/ffa.c b/arch/arm64/kvm/hyp/nvhe/ffa.c
> > index 4e16f9b96f63..58b7d0c477d7 100644
> > --- a/arch/arm64/kvm/hyp/nvhe/ffa.c
> > +++ b/arch/arm64/kvm/hyp/nvhe/ffa.c
> > @@ -479,7 +479,7 @@ static void __do_ffa_mem_xfer(const u64 func_id,
> >  	struct ffa_mem_region_attributes *ep_mem_access;
> >  	struct ffa_composite_mem_region *reg;
> >  	struct ffa_mem_region *buf;
> > -	u32 offset, nr_ranges;
> > +	u32 offset, nr_ranges, checked_offset;
> >  	int ret = 0;
> >  
> >  	if (addr_mbz || npages_mbz || fraglen > len ||
> > @@ -516,7 +516,12 @@ static void __do_ffa_mem_xfer(const u64 func_id,
> >  		goto out_unlock;
> >  	}
> >  
> > -	if (fraglen < offset + sizeof(struct ffa_composite_mem_region)) {
> > +	if (check_add_overflow(offset, sizeof(struct ffa_composite_mem_region), &checked_offset)) {
> > +		ret = FFA_RET_INVALID_PARAMETERS;
> > +		goto out_unlock;
> > +	}
> > +
> > +	if (fraglen < checked_offset) {
> >  		ret = FFA_RET_INVALID_PARAMETERS;
> >  		goto out_unlock;
> >  
> 
> Perhaps this could be easier to reason about by moving this check with the nr_ranges?

I found it a bit more clear to use the helper on the offset variable, I
would like to keep it in this way if you are ok with this.


> 
>         reg = (void *)buf + offset;
>         if ((void *)reg->constituents > (void *)buf + fraglen) {
>                 ret = FFA_RET_INVALID_PARAMETERS;
>                 goto out_unlock;
>         }
>  
>         nr_ranges = ((void *)buf + fraglen) - (void *)reg->constituents;
>         if (nr_ranges % sizeof(reg->constituents[0])) {
>                 ret = FFA_RET_INVALID_PARAMETERS;
> 

Thanks,
Sebastian

> }
> > -- 
> > 2.51.0.858.gf9c4a03a3a-goog
> >