[PATCH] firmware: arm_scmi: fix voltage description in failure cases

[Cristian Marussi]

On Fri, Jul 26, 2024 at 04:16:03PM +0100, Sudeep Holla wrote:
> On Thu, Jul 25, 2024 at 08:53:17AM +0200, Etienne Carriere wrote:
> > Reset the reception buffer max size when a voltage domain description
> > request fails, for example when the voltage domain returns an access
> > permission error (SCMI_ERR_ACCESS) unless what only a single 32bit
> > word is read back for the remaining voltage description requests
> > responses leading to invalid information. The side effect of this
> > issue is that the voltage regulators registered from those remaining
> > SCMI voltage domain were assigned a wrong regulator name.
> > 
> > Signed-off-by: Etienne Carriere <etienne.carriere at foss.st.com>
> > ---
> >  drivers/firmware/arm_scmi/voltage.c | 4 +++-
> >  1 file changed, 3 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/firmware/arm_scmi/voltage.c b/drivers/firmware/arm_scmi/voltage.c
> > index 2175ffd6cef5..f1a7c04ae820 100644
> > --- a/drivers/firmware/arm_scmi/voltage.c
> > +++ b/drivers/firmware/arm_scmi/voltage.c
> > @@ -229,8 +229,10 @@ static int scmi_voltage_descriptors_get(const struct scmi_protocol_handle *ph,
> >  		/* Retrieve domain attributes at first ... */
> >  		put_unaligned_le32(dom, td->tx.buf);
> >  		/* Skip domain on comms error */
> > -		if (ph->xops->do_xfer(ph, td))
> > +		if (ph->xops->do_xfer(ph, td)) {
> > +			ph->xops->reset_rx_to_maxsz(ph, td);
> 
> I am fine with this to keep it simple, but thought I will check my thoughts.
> We usually use reset_rx_to_maxsz in iterators as we don't know the expected
> size of the response, whereas here it must be max sizeof(*resp_dom).
> 
> That said, we don't have any helpers and changing xfer->rx.len directly
> doesn't looks good ? Or may be it is OK ? Thoughts ?

We do not access those xfer internal field explicitly from the protocol layer
(beside once in Base)...and surely not on write....in the past I was even
tempted to try to make these internal stuff untouchable by the protocol layer...
...that's the reason of course for the existence of reset_rx_to_maxsz() helpers
....not sure if it is worth adding another helper for this, given that the
using the maxsz should have any adverse effect (unless I am missing
something of course :P).

This kind of 'issues' are really common whenever in the SCMI stack we
reuse the same allocated xfer across multiple do_xfers in a loop
(reusing seq_nums is another thing...) since we wanted to avoid the
penalty of resetting some of these automatically on each do_xfer()...

....we could think of some mechanism to transparently reset/fill such xfer
fields automatically if the core detects a 'reuse'....got to check first,
though, if this does not break some of the current usage patterns...and
I would not say it is a high prio thing to explore as of now...

Thanks,
Cristian