[Intel-gfx] [PATCH v1 2/2] drm/i915: Reject the atomic modeset if an associated Type-C port is disconnected

[Kasireddy, Vivek]

Hi Imre,

> 
> On Fri, May 20, 2022 at 10:28:31AM +0300, Kasireddy, Vivek wrote:
> > Hi Imre,
> > [...]
> > > > > > @@ -131,6 +137,20 @@ int intel_digital_connector_atomic_check(struct
> drm_connector *conn,
> > > > > >
> > > > > >  	crtc_state = drm_atomic_get_new_crtc_state(state, new_state->crtc);
> > > > > >
> > > > > > +	/*
> > > > > > +	 * The spec says that it is not safe to use a disconnected Type-C port.
> > > > > > +	 * Therefore, check to see if this connector is connected and reject
> > > > > > +	 * the modeset if there is no sink detected.
> > > > > > +	 */
> > > > > > +	if (dig_port && !dig_port->connected(encoder) &&
> > > > >
> > > > > This check is racy, as right after dig_port->connected() returns true,
> > > > > the port can become disconnected.
> > > >
> > > > [Kasireddy, Vivek] Given that, do you think the only way to reliably determine
> > > > if the Type-C port has a sink is to check the live status and ignore dig_port-
> >tc_mode?
> > > >
> > > > If that is the case, should I just add a function pointer to dig_port to call
> > > > tc_port_live_status_mask()? Or, should I just change intel_tc_port_connected()
> > > > to ignore dig_port->tc_mode like below:
> > > > @@ -764,8 +764,7 @@ bool intel_tc_port_connected(struct intel_encoder
> *encoder)
> > > >
> > > >         intel_tc_port_lock(dig_port);
> > > >
> > > > -       is_connected = tc_port_live_status_mask(dig_port) &
> > > > -                      BIT(dig_port->tc_mode);
> > > > +       is_connected = tc_port_live_status_mask(dig_port);
> > > >
> > > > Or, are there any other elegant ways that you can think of to determine whether
> > > > a tc port has a sink or not?
> > >
> > > I meant that I don't think there is a way to prevent a modeset on a
> > > disconnected port.
> >
> > But we need to find a way right given that the spec clearly states that the driver
> > must not use or access (PHY/FIA registers of) a disconnected tc port.
> 
> The driver does not access the PHY/FIA regs on a disconnected port/PHY.
[Kasireddy, Vivek] I think it does after the first patch in this series is applied if
the userspace (Weston) forces a modeset on a disconnected tc legacy port (HDMI).
For instance, some of the FIA/PHY regs accesses I noticed include programming
the lane count (intel_tc_port_set_fia_lane_count() called by intel_ddi_pre_pll_enable()),
reading the pin assignment mask (intel_tc_port_get_pin_assignment_mask() called
by icl_program_mg_dp_mode() which is called by intel_ddi_pre_enable_hdmi()), etc.

Of-course, these accesses would probably not occur if the disconnected tc port
defaults to TBT mode but this brings other problems like I described in the
commit description of the first patch and the cover letter.
 
> 
> > > Live status is what provides the connected state, but
> > > it can change right after it is read out.
> >
> > Does this change happen after giving up the ownership (in
> > icl_tc_phy_disconnect)?
> 
> The HPD live status changes whenever a user plugs/unplugs a sink.
> 
> > But shouldn't we distinguish between the cases where we are
> > deliberately disconnecting the phy for power-savings reason vs when
> > the port actually becomes disconnected? The port can still be
> > considered connected in the former case right?
> 
> The driver - based on the spec - needs to avoid accessing the PHY/FIA
> regs whenever the PHY is disconnected either by FW/HW (because the user
> unplugged the sink) or the driver (during the suspend, modeset disable
> sequence).
[Kasireddy, Vivek] Regardless of whether the PHY/FIA regs are accessed or
not, I don't think the driver should let the userspace's modeset to succeed on
a disconnected tc port. Do you not agree?

> 
> > Under what other situations would the live status change or become
> > unreliable after the port has a connected sink?
> 
> It's not unreliable, it reflects the state of a sink being plugged to
> the connector or not.
[Kasireddy, Vivek] Ok, assuming that the state of the sink is "connected"
during intel_atomic_check() phase (which is where this patch checks for
connected status), are you concerned about the case where the user may
unplug the sink before we get to the intel_atomic_commit() phase? Is
this what you meant when you said this earlier: "This check is racy, as
right after dig_port->connected() returns true, the port can become
disconnected"? I am just trying to figure out the scenarios when this
might happen.

> 
> > And, since we rely on SDEISR to detect the live status for tc legacy
> > ports, could this not be considered reliable?
> 
> Changes in the HPD live status is used as a hint to user space to
> follow up with connector detection and modeset enable/disable requests
> as necessary.
[Kasireddy, Vivek] Right, that is the ideal case but user/userspace can commit
mistakes where for example they can assume that HDMI-A-1 is connected 
(while it is not) instead of HDMI-A-3 which is the one actually connected.
During such cases, I think the driver should not let the userspace hang the
system or lead to unexpected states and instead should return an error.

Thanks,
Vivek

> 
> --Imre