Re: [PATCH v2 4/4] KVM: arm64: Release the ownership of the hyp rx buffer to Trustzone

[Will Deacon]

On Thu, Mar 06, 2025 at 09:40:43AM +0000, Sudeep Holla wrote:
> On Wed, Mar 05, 2025 at 07:34:26PM +0000, Will Deacon wrote:
> > On Wed, Mar 05, 2025 at 09:41:04AM +0000, Sudeep Holla wrote:
> > > On Wed, Mar 05, 2025 at 12:45:23AM +0000, Will Deacon wrote:
> > > > Hmm, the FFA spec is characteristically unclear as to whether or not we
> > > > need to release the rx buffer in the case that the flags indicate use of
> > > > the rx buffer but the returned partition count is 0.
> > > > 
> > > > Sudeep -- do you know what we should be doing in that case?
> > > > 
> > > 
> > > We need to call RX_RELEASE here. I went back to the spec to confirm the
> > > same again.
> > > 
> > > v1.2 EAC0 spec Section 7.2.2.4.2 Transfer of buffer ownership
> > > (Or just look for the section title in any version of the spec)
> > > "
> > > 2. Ownership transfer for the RX buffer takes place as follows.
> > >     2. For a framework message,
> > >        1. Completion of the FFA_PARTITION_INFO_GET ABI transfers the ownership
> > >        of the caller’s RX buffer from the Producer to the Consumer.
> > > 3. For both types of messages, an invocation of the following FF-A ABIs
> > >     transfers the ownership from the Consumer to the Producer.
> > >        1. FFA_MSG_WAIT ...
> > >        2. FFA_RX_RELEASE.
> > > "
> > > 
> > > Hope that helps, can dig deeper if there are any ambiguities around this.
> > 
> > Thanks Sudeep, but that also makes it sound like we need the RX_RELEASE
> > even if we're not using the RX buffer per the input flags. :/
> > 
> 
> Good spot, I had forgotten about the input flags that can avoid using the
> buffer. I will see if we can improve the spec in that regards.

Thanks. In the meantime, what do you think is the correct behaviour in that
case? I guess _not_ doing the release when the flags don't request the RX
buffer? In other words:


	if (flags & PARTITION_INFO_GET_RETURN_COUNT_ONLY)
		goto out_unlock;

	if (!count)
		goto release_rx;

	[...]

	if (copy_sz > KVM_FFA_MBOX_NR_PAGES * PAGE_SIZE) {
		ffa_to_smccc_res(res, FFA_RET_ABORTED);
		goto release_rx;
	}

	memcpy(host_buffers.rx, hyp_buffers.rx, copy_sz);
release_rx:
	ffa_rx_release(&_res);
out_unlock:
	hyp_spin_unlock(&host_buffers.lock);
}


What do you reckon?

Will