On Wed, Nov 19, 2025 at 09:38:10AM +0200, Suraj Kandpal wrote:
> > Subject: Re: [PATCH] drm/display/dp_mst: Add protection against 0 vcpi
> > 
> > On Mon, Nov 17, 2025 at 07:09:38AM +0200, Suraj Kandpal wrote:
> > > > -----Original Message-----
> > > > From: Jani Nikula <[email protected]>
> > > > Sent: Thursday, November 13, 2025 9:55 PM
> > > > To: Deak, Imre <[email protected]>; Kandpal, Suraj
> > > > <[email protected]>
> > > > Cc: [email protected]; [email protected];
> > > > intel- [email protected]; Nautiyal, Ankit K
> > > > <[email protected]>; Murthy, Arun R
> > > > <[email protected]>
> > > > Subject: Re: [PATCH] drm/display/dp_mst: Add protection against 0
> > > > vcpi
> > > >
> > > > On Thu, 13 Nov 2025, Imre Deak <[email protected]> wrote:
> > > > > On Thu, Nov 13, 2025 at 10:09:19AM +0530, Suraj Kandpal wrote:
> > > > >> When releasing a timeslot there is a slight chance we may end up
> > > > >> with the wrong payload mask due to overflow if the
> > > > >> delayed_destroy_work ends up coming into play after a DP 2.1
> > > > >> monitor gets disconnected which causes vcpi to become 0 then we
> > > > >> try to make the payload = ~BIT(vcpi - 1) which is a negative shift.
> > > > >>
> > > > >> Signed-off-by: Suraj Kandpal <[email protected]>
> > > > >> ---
> > > > >>  drivers/gpu/drm/display/drm_dp_mst_topology.c | 4 +++-
> > > > >>  1 file changed, 3 insertions(+), 1 deletion(-)
> > > > >>
> > > > >> diff --git a/drivers/gpu/drm/display/drm_dp_mst_topology.c
> > > > >> b/drivers/gpu/drm/display/drm_dp_mst_topology.c
> > > > >> index 64e5c176d5cc..3cf1eafcfcb5 100644
> > > > >> --- a/drivers/gpu/drm/display/drm_dp_mst_topology.c
> > > > >> +++ b/drivers/gpu/drm/display/drm_dp_mst_topology.c
> > > > >> @@ -4531,6 +4531,7 @@ int drm_dp_atomic_release_time_slots(struct
> > > > drm_atomic_state *state,
> > > > >>      struct drm_dp_mst_atomic_payload *payload;
> > > > >>      struct drm_connector_state *old_conn_state, *new_conn_state;
> > > > >>      bool update_payload = true;
> > > > >> +    int bit;
> > > > >>
> > > > >>      old_conn_state = drm_atomic_get_old_connector_state(state,
> > > > >> port-
> > > > >connector);
> > > > >>      if (!old_conn_state->crtc)
> > > > >> @@ -4572,7 +4573,8 @@ int drm_dp_atomic_release_time_slots(struct
> > > > drm_atomic_state *state,
> > > > >>      if (!payload->delete) {
> > > > >>              payload->pbn = 0;
> > > > >>              payload->delete = true;
> > > > >> -            topology_state->payload_mask &= ~BIT(payload->vcpi - 1);
> > > > >> +            bit = payload->vcpi ? payload->vcpi - 1 : 0;
> > > > >> +            topology_state->payload_mask &= ~BIT(bit);
> > > > >
> > > > > This looks wrong, clearing the bit for an unrelated payload.
> > > >
> > > > Agreed.
> > > >
> > > > The logs have, among other things,
> > > >
> > > > <7> [515.138211] xe 0000:03:00.0:
> > > > [drm:intel_dp_sink_set_dsc_decompression
> > > > [xe]] Failed to enable sink decompression state
> > > >
> > > > <7> [515.193484] xe 0000:03:00.0: [drm:drm_dp_add_payload_part1
> > > > [drm_display_helper]] VCPI 0 for port ffff888126ce9000 not in
> > > > topology, not creating a payload to remote
> > > >
> > > > <7> [515.194671] xe 0000:03:00.0: [drm:drm_dp_add_payload_part2
> > > > [drm_display_helper]] Part 1 of payload creation for DP-5 failed,
> > > > skipping part 2
> > > >
> > > > <7> [515.347331] xe 0000:03:00.0: [drm:drm_dp_remove_payload_part1
> > > > [drm_display_helper]] Payload for VCPI 0 not in topology, not
> > > > sending remove
> > > >
> > > > So it's no wonder the port's not in topology and everything fails.
> > > > We obviously need to skip payload_mask updates when the VCPI is 0,
> > > > but that's just a symptom of other stuff going wrong first. Perhaps
> > > > we could do with some earlier error handling too?
> > >
> > > Yes I agree the question is how high will the error handling needs to be
> > added.
> > > A lot of weird things going on here.
> > >
> > > 1st one is how is it finding a payload which we do not create while we
> > > call destroy function
> > >
> > > 2nd how is VCPI with id 0 possible from what I see VCPI are 1 at least
> > > that's what I gather from
> > > drm_dp_mst_atomic_check_payload_alloc_limits.So what are we missing
> > when we create a payload?
> > >
> > > Imre, Jani any idea still new to how payload creation work so am I
> > > missing something.
> > 
> > A VCPI ID will be assigned to a payload during an atomic commit only if the
> > corresponding MST connector is still connected. If the MST connector gets
> > disconnected by the time of the atomic commit - as in the above case - no 
> > VCPI
> > ID will assigned and the allocation table in the branch device cannot be
> > updated either for the payload, as indicated by the above payload
> > creation/removal failed messages.
> > 
> > I think the fix should be not to clear the VCPI ID if it's 0. Valid VCPI 
> > IDs start
> > from 1.
> 
> Hmm then in that case should we just return 0 early if vcpi turns out to be 0 
> here.

The payload should be still deleted, so only the clearing of VCPI ID
from payload_mask needs to be avoided if the ID is 0.

> Regards,
> Suraj Kandpal
> 
> > > Regards
> > > Suraj Kandpal
> > >
> > > > BR,
> > > > Jani.
> > > >
> > > >
> > > > >
> > > > >>      }
> > > > >>
> > > > >>      return 0;
> > > > >> --
> > > > >> 2.34.1
> > > > >>
> > > >
> > > > --
> > > > Jani Nikula, Intel

Reply via email to