On Thu, May 04, 2023 at 08:16:33AM -0700, Bjorn Andersson wrote:
> On Thu, May 04, 2023 at 03:44:53PM +0200, Johan Hovold wrote:
> > On Wed, May 03, 2023 at 08:29:07PM -0700, Bjorn Andersson wrote:
> > > On Tue, May 02, 2023 at 01:48:16PM +0200, Johan Hovold wrote:
> > > > On Mon, Apr 24, 2023 at 08:40:06PM -0700, Bjorn Andersson wrote:
> >
> > > > > static void qmp_v3_dp_aux_init(struct qmp_combo *qmp);
> > > > > @@ -1955,29 +1962,23 @@ static void qmp_v3_configure_dp_tx(struct qmp_combo *qmp)
> > > > > static bool qmp_combo_configure_dp_mode(struct qmp_combo *qmp)
> > > > > {
> > > > > u32 val;
> > > > > - bool reverse = false;
> > > > > + bool reverse = qmp->orientation == TYPEC_ORIENTATION_REVERSE;
> >
> > > > It also looks like these callbacks end up being called without holding
> > > > the qmp->phy_mutex via phy->power_on(). Perhaps there is no risk for a
> > > > concurrent switch notification and dp phy power-on but it's not that
> > > > obvious.
> >
> > > It seems we're arriving here from hpd_event_thread(), while
> > > phy_power_on() and phy_power_off() will be called in some other context.
> > > I've not been able to convince myself if DP driver ensures ordering, or
> > > if we have an existing race here...
> >
> > > Unless you insist, I would prefer to follow up with an additional patch
> > > once we've landed this series. The fix will depend on the phy_mutex
> > > shuffling patch anyways...
> >
> > Sure.
> >
> > But perhaps you can just move the orientation == qmp->orientation check
> > under the mutex in qmp_combo_typec_switch_set() for now (in case I
> > forgot to point that out earlier).
> >
>
> qmp_combo_probe() and qmp_combo_typec_switch_set() are the only writers
> to qmp->orientation, so that check can't race with any updates and hence
> doesn't need to be protected.
Only if you happen to know that the callers of
qmp_combo_typec_switch_set() are serialised, right? That happens to be
the case for pmic_glink, but it may not be the case generally.
> Reading the code again, qmp_combo_configure_dp_mode() is invoked from
> phy_power_on(), not the hpd_event_thread(), as I claimed yesterday.
Yeah, but phy_power_on() is typically called from that thread. But
perhaps not only from there.
> But we shouldn't do qmp_combo_dp_power_on() in parallel with the
> reinitialization following a switch in orientation, qmp->orientation
> might change, but we definitely would have two contexts reconfiguring
> the hardware simultaneously - perhaps this was the cause for the 10%
> crashes I hit when trying to extend this to handle typec_mux as well...
>
> I will grab the phy_mux in qmp_combo_configure_dp_mode() as well, thanks
> for "insisting" :)
:)
Johan
