Re: [Intel-gfx] [PATCH v2 02/17] drm/display/dp_mst: Handle old/new payload states in drm_dp_remove_payload()
On Tue, 2023-01-31 at 17:05 +0200, Imre Deak wrote: > Atm, drm_dp_remove_payload() uses the same payload state to both get the > vc_start_slot required for the payload removal DPCD message and to > deduct time_slots from vc_start_slot of all payloads after the one being > removed. > > The above isn't always correct, as vc_start_slot must be the up-to-date > version contained in the new payload state, but time_slots must be the > one used when the payload was previously added, contained in the old > payload state. The new payload's time_slots can change vs. the old one > if the current atomic commit changes the corresponding mode. > > This patch let's drivers pass the old and new payload states to > drm_dp_remove_payload(), but keeps these the same for now in all drivers > not to change the behavior. A follow-up i915 patch will pass in that > driver the correct old and new states to the function. Oh wow, this was definitely a mistake on my part, thanks for catching this! TBH, I think this behavior is correct so (now that I actually have a setup capable of testing amdgpu's MST fully thanks to gitlab issue 2171…) if you'd like to change it on other drivers as well I can test it fully. Or feel free to leave it to me, shouldn't be too difficult I think :). For 0-2: Reviewed-by: Lyude Paul > > Cc: Lyude Paul > Cc: Ville Syrjälä > Cc: Ben Skeggs > Cc: Karol Herbst > Cc: Harry Wentland > Cc: Alex Deucher > Cc: Wayne Lin > Cc: sta...@vger.kernel.org # 6.1 > Cc: dri-de...@lists.freedesktop.org > Reviewed-by: Ville Syrjälä > Signed-off-by: Imre Deak > --- > .../amd/display/amdgpu_dm/amdgpu_dm_helpers.c | 2 +- > drivers/gpu/drm/display/drm_dp_mst_topology.c | 26 ++- > drivers/gpu/drm/i915/display/intel_dp_mst.c | 4 ++- > drivers/gpu/drm/nouveau/dispnv50/disp.c | 2 +- > include/drm/display/drm_dp_mst_helper.h | 3 ++- > 5 files changed, 21 insertions(+), 16 deletions(-) > > diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > index a50319fc42b11..180d3893b68da 100644 > --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > @@ -208,7 +208,7 @@ bool dm_helpers_dp_mst_write_payload_allocation_table( > if (enable) > drm_dp_add_payload_part1(mst_mgr, mst_state, payload); > else > - drm_dp_remove_payload(mst_mgr, mst_state, payload); > + drm_dp_remove_payload(mst_mgr, mst_state, payload, payload); > > /* mst_mgr->->payloads are VC payload notify MST branch using DPCD or >* AUX message. The sequence is slot 1-63 allocated sequence for each > diff --git a/drivers/gpu/drm/display/drm_dp_mst_topology.c > b/drivers/gpu/drm/display/drm_dp_mst_topology.c > index 847c10aa2098c..1990ff5dc7ddd 100644 > --- a/drivers/gpu/drm/display/drm_dp_mst_topology.c > +++ b/drivers/gpu/drm/display/drm_dp_mst_topology.c > @@ -3342,7 +3342,8 @@ EXPORT_SYMBOL(drm_dp_add_payload_part1); > * drm_dp_remove_payload() - Remove an MST payload > * @mgr: Manager to use. > * @mst_state: The MST atomic state > - * @payload: The payload to write > + * @old_payload: The payload with its old state > + * @new_payload: The payload to write > * > * Removes a payload from an MST topology if it was successfully assigned a > start slot. Also updates > * the starting time slots of all other payloads which would have been > shifted towards the start of > @@ -3350,36 +3351,37 @@ EXPORT_SYMBOL(drm_dp_add_payload_part1); > */ > void drm_dp_remove_payload(struct drm_dp_mst_topology_mgr *mgr, > struct drm_dp_mst_topology_state *mst_state, > -struct drm_dp_mst_atomic_payload *payload) > +const struct drm_dp_mst_atomic_payload *old_payload, > +struct drm_dp_mst_atomic_payload *new_payload) > { > struct drm_dp_mst_atomic_payload *pos; > bool send_remove = false; > > /* We failed to make the payload, so nothing to do */ > - if (payload->vc_start_slot == -1) > + if (new_payload->vc_start_slot == -1) > return; > > mutex_lock(&mgr->lock); > - send_remove = drm_dp_mst_port_downstream_of_branch(payload->port, > mgr->mst_primary); > + send_remove = drm_dp_mst_port_downstream_of_branch(new_payload->port, > mgr->mst_primary); > mutex_unlock(&mgr->lock); > > if (send_remove) > - drm_dp_destroy_payload_step1(mgr, mst_state, payload); > + drm_dp_destroy_payload_step1(mgr, mst_state, new_payload); > else > drm_dbg_kms(mgr->dev, "Payload for VCPI %d not in topology, not > sending remove\n", > - payload->vcpi); > + new_payload->vcpi); > > list_for_each_entry(pos, &mst_state->payloads, next) { > - if (pos != payload &&
Re: [Intel-gfx] [PATCH v2 02/17] drm/display/dp_mst: Handle old/new payload states in drm_dp_remove_payload()
On Tue, Jan 31, 2023 at 06:13:10PM -0500, Lyude Paul wrote: > On Tue, 2023-01-31 at 17:05 +0200, Imre Deak wrote: > > Atm, drm_dp_remove_payload() uses the same payload state to both get the > > vc_start_slot required for the payload removal DPCD message and to > > deduct time_slots from vc_start_slot of all payloads after the one being > > removed. > > > > The above isn't always correct, as vc_start_slot must be the up-to-date > > version contained in the new payload state, but time_slots must be the > > one used when the payload was previously added, contained in the old > > payload state. The new payload's time_slots can change vs. the old one > > if the current atomic commit changes the corresponding mode. > > > > This patch let's drivers pass the old and new payload states to > > drm_dp_remove_payload(), but keeps these the same for now in all drivers > > not to change the behavior. A follow-up i915 patch will pass in that > > driver the correct old and new states to the function. > > Oh wow, this was definitely a mistake on my part, thanks for catching this! > TBH, I think this behavior is correct so (now that I actually have a setup > capable of testing amdgpu's MST fully thanks to gitlab issue 2171…) if you'd > like to change it on other drivers as well I can test it fully. I only checked that the other drivers pass the new payload state to drm_dp_remove_payload(), so not sure how that works atm if the same commit has to both remove the payload (with the old time_slots value) and add it back (with a new time_slots value). Maybe that can't happen in those drivers, or time_slots get updated between remove and readd. > Or feel free to leave it to me, shouldn't be too difficult I think :). Yes, this patch should have no functional change, so please check what would apply to other drivers as well. Could you also check Ville's comment about storing start_slot elsewhere than the atomic state (leaving only time_slots there). I wonder if that would work, at least it would simplify things I think. > For 0-2: > > Reviewed-by: Lyude Paul Thanks. > > > > > Cc: Lyude Paul > > Cc: Ville Syrjälä > > Cc: Ben Skeggs > > Cc: Karol Herbst > > Cc: Harry Wentland > > Cc: Alex Deucher > > Cc: Wayne Lin > > Cc: sta...@vger.kernel.org # 6.1 > > Cc: dri-de...@lists.freedesktop.org > > Reviewed-by: Ville Syrjälä > > Signed-off-by: Imre Deak > > --- > > .../amd/display/amdgpu_dm/amdgpu_dm_helpers.c | 2 +- > > drivers/gpu/drm/display/drm_dp_mst_topology.c | 26 ++- > > drivers/gpu/drm/i915/display/intel_dp_mst.c | 4 ++- > > drivers/gpu/drm/nouveau/dispnv50/disp.c | 2 +- > > include/drm/display/drm_dp_mst_helper.h | 3 ++- > > 5 files changed, 21 insertions(+), 16 deletions(-) > > > > diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > index a50319fc42b11..180d3893b68da 100644 > > --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > @@ -208,7 +208,7 @@ bool dm_helpers_dp_mst_write_payload_allocation_table( > > if (enable) > > drm_dp_add_payload_part1(mst_mgr, mst_state, payload); > > else > > - drm_dp_remove_payload(mst_mgr, mst_state, payload); > > + drm_dp_remove_payload(mst_mgr, mst_state, payload, payload); > > > > /* mst_mgr->->payloads are VC payload notify MST branch using DPCD or > > * AUX message. The sequence is slot 1-63 allocated sequence for each > > diff --git a/drivers/gpu/drm/display/drm_dp_mst_topology.c > > b/drivers/gpu/drm/display/drm_dp_mst_topology.c > > index 847c10aa2098c..1990ff5dc7ddd 100644 > > --- a/drivers/gpu/drm/display/drm_dp_mst_topology.c > > +++ b/drivers/gpu/drm/display/drm_dp_mst_topology.c > > @@ -3342,7 +3342,8 @@ EXPORT_SYMBOL(drm_dp_add_payload_part1); > > * drm_dp_remove_payload() - Remove an MST payload > > * @mgr: Manager to use. > > * @mst_state: The MST atomic state > > - * @payload: The payload to write > > + * @old_payload: The payload with its old state > > + * @new_payload: The payload to write > > * > > * Removes a payload from an MST topology if it was successfully assigned > > a start slot. Also updates > > * the starting time slots of all other payloads which would have been > > shifted towards the start of > > @@ -3350,36 +3351,37 @@ EXPORT_SYMBOL(drm_dp_add_payload_part1); > > */ > > void drm_dp_remove_payload(struct drm_dp_mst_topology_mgr *mgr, > >struct drm_dp_mst_topology_state *mst_state, > > - struct drm_dp_mst_atomic_payload *payload) > > + const struct drm_dp_mst_atomic_payload *old_payload, > > + struct drm_dp_mst_atomic_payload *new_payload) > > { > > struct drm_dp_mst_atomic_payload *pos; > > bool send_remove = false; > > > > /* We failed to make the payload, so not
Re: [Intel-gfx] [PATCH v2 02/17] drm/display/dp_mst: Handle old/new payload states in drm_dp_remove_payload()
On Wed, 2023-02-01 at 17:04 +0200, Imre Deak wrote: > > Yes, this patch should have no functional change, so please check what > would apply to other drivers as well. > > Could you also check Ville's comment about storing start_slot elsewhere > than the atomic state (leaving only time_slots there). I wonder if that > would work, at least it would simplify things I think. Ideally it'd be nice if we could have all this info in the atomic commit, but yeah - you're not the first person to find this a bit confusing. FWIW though, the way we store start_slot right now is intended to follow the same kind of behavior we have with atomic CRTC commit dependencies, I think I also made it that way so all of the state would be in a single place but there's no hard requirement for it. So if you guys think it'd be better design-wise to store this something else, I've got no strong feelings either way > > > For 0-2: > > > > Reviewed-by: Lyude Paul > > Thanks. > > > > > > > > > Cc: Lyude Paul > > > Cc: Ville Syrjälä > > > Cc: Ben Skeggs > > > Cc: Karol Herbst > > > Cc: Harry Wentland > > > Cc: Alex Deucher > > > Cc: Wayne Lin > > > Cc: sta...@vger.kernel.org # 6.1 > > > Cc: dri-de...@lists.freedesktop.org > > > Reviewed-by: Ville Syrjälä > > > Signed-off-by: Imre Deak > > > --- > > > .../amd/display/amdgpu_dm/amdgpu_dm_helpers.c | 2 +- > > > drivers/gpu/drm/display/drm_dp_mst_topology.c | 26 ++- > > > drivers/gpu/drm/i915/display/intel_dp_mst.c | 4 ++- > > > drivers/gpu/drm/nouveau/dispnv50/disp.c | 2 +- > > > include/drm/display/drm_dp_mst_helper.h | 3 ++- > > > 5 files changed, 21 insertions(+), 16 deletions(-) > > > > > > diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > > b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > > index a50319fc42b11..180d3893b68da 100644 > > > --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > > +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > > @@ -208,7 +208,7 @@ bool dm_helpers_dp_mst_write_payload_allocation_table( > > > if (enable) > > > drm_dp_add_payload_part1(mst_mgr, mst_state, payload); > > > else > > > - drm_dp_remove_payload(mst_mgr, mst_state, payload); > > > + drm_dp_remove_payload(mst_mgr, mst_state, payload, payload); > > > > > > /* mst_mgr->->payloads are VC payload notify MST branch using DPCD or > > >* AUX message. The sequence is slot 1-63 allocated sequence for each > > > diff --git a/drivers/gpu/drm/display/drm_dp_mst_topology.c > > > b/drivers/gpu/drm/display/drm_dp_mst_topology.c > > > index 847c10aa2098c..1990ff5dc7ddd 100644 > > > --- a/drivers/gpu/drm/display/drm_dp_mst_topology.c > > > +++ b/drivers/gpu/drm/display/drm_dp_mst_topology.c > > > @@ -3342,7 +3342,8 @@ EXPORT_SYMBOL(drm_dp_add_payload_part1); > > > * drm_dp_remove_payload() - Remove an MST payload > > > * @mgr: Manager to use. > > > * @mst_state: The MST atomic state > > > - * @payload: The payload to write > > > + * @old_payload: The payload with its old state > > > + * @new_payload: The payload to write > > > * > > > * Removes a payload from an MST topology if it was successfully > > > assigned a start slot. Also updates > > > * the starting time slots of all other payloads which would have been > > > shifted towards the start of > > > @@ -3350,36 +3351,37 @@ EXPORT_SYMBOL(drm_dp_add_payload_part1); > > > */ > > > void drm_dp_remove_payload(struct drm_dp_mst_topology_mgr *mgr, > > > struct drm_dp_mst_topology_state *mst_state, > > > -struct drm_dp_mst_atomic_payload *payload) > > > +const struct drm_dp_mst_atomic_payload *old_payload, > > > +struct drm_dp_mst_atomic_payload *new_payload) > > > { > > > struct drm_dp_mst_atomic_payload *pos; > > > bool send_remove = false; > > > > > > /* We failed to make the payload, so nothing to do */ > > > - if (payload->vc_start_slot == -1) > > > + if (new_payload->vc_start_slot == -1) > > > return; > > > > > > mutex_lock(&mgr->lock); > > > - send_remove = drm_dp_mst_port_downstream_of_branch(payload->port, > > > mgr->mst_primary); > > > + send_remove = drm_dp_mst_port_downstream_of_branch(new_payload->port, > > > mgr->mst_primary); > > > mutex_unlock(&mgr->lock); > > > > > > if (send_remove) > > > - drm_dp_destroy_payload_step1(mgr, mst_state, payload); > > > + drm_dp_destroy_payload_step1(mgr, mst_state, new_payload); > > > else > > > drm_dbg_kms(mgr->dev, "Payload for VCPI %d not in topology, not > > > sending remove\n", > > > - payload->vcpi); > > > + new_payload->vcpi); > > > > > > list_for_each_entry(pos, &mst_state->payloads, next) { > > > - if (pos != payload && pos->vc_start_slot > > > > payload->vc_start_slot) > > > - pos->vc_start_slot -= payload->time_slots;
Re: [Intel-gfx] [PATCH v2 02/17] drm/display/dp_mst: Handle old/new payload states in drm_dp_remove_payload()
On Mon, Feb 06, 2023 at 07:42:32PM -0500, Lyude Paul wrote: > On Wed, 2023-02-01 at 17:04 +0200, Imre Deak wrote: > > > > Yes, this patch should have no functional change, so please check what > > would apply to other drivers as well. > > > > Could you also check Ville's comment about storing start_slot elsewhere > > than the atomic state (leaving only time_slots there). I wonder if that > > would work, at least it would simplify things I think. > > Ideally it'd be nice if we could have all this info in the atomic commit, but > yeah - you're not the first person to find this a bit confusing. FWIW though, > the way we store start_slot right now is intended to follow the same kind of > behavior we have with atomic CRTC commit dependencies, I think I also made it > that way so all of the state would be in a single place but there's no hard > requirement for it. As I understood the atomic state should be precomputed during atomic check and not changed later during the commit phase. In case of start_slot this would mean knowing in advance the actual (driver specific) disabling / enabling sequence of the payloads, not sure how feasible it is to figure that out. start_slot also changes dynamically as each payload is disabled, so these intermediate versions of the field would need to be tracked somehow separately from the final (precomputed) versions. > So if you guys think it'd be better design-wise to store this something else, > I've got no strong feelings either way > > > > > For 0-2: > > > > > > Reviewed-by: Lyude Paul > > > > Thanks. > > > > > > > > > > > > > Cc: Lyude Paul > > > > Cc: Ville Syrjälä > > > > Cc: Ben Skeggs > > > > Cc: Karol Herbst > > > > Cc: Harry Wentland > > > > Cc: Alex Deucher > > > > Cc: Wayne Lin > > > > Cc: sta...@vger.kernel.org # 6.1 > > > > Cc: dri-de...@lists.freedesktop.org > > > > Reviewed-by: Ville Syrjälä > > > > Signed-off-by: Imre Deak > > > > --- > > > > .../amd/display/amdgpu_dm/amdgpu_dm_helpers.c | 2 +- > > > > drivers/gpu/drm/display/drm_dp_mst_topology.c | 26 ++- > > > > drivers/gpu/drm/i915/display/intel_dp_mst.c | 4 ++- > > > > drivers/gpu/drm/nouveau/dispnv50/disp.c | 2 +- > > > > include/drm/display/drm_dp_mst_helper.h | 3 ++- > > > > 5 files changed, 21 insertions(+), 16 deletions(-) > > > > > > > > diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > > > b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > > > index a50319fc42b11..180d3893b68da 100644 > > > > --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > > > +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > > > @@ -208,7 +208,7 @@ bool > > > > dm_helpers_dp_mst_write_payload_allocation_table( > > > > if (enable) > > > > drm_dp_add_payload_part1(mst_mgr, mst_state, payload); > > > > else > > > > - drm_dp_remove_payload(mst_mgr, mst_state, payload); > > > > + drm_dp_remove_payload(mst_mgr, mst_state, payload, > > > > payload); > > > > > > > > /* mst_mgr->->payloads are VC payload notify MST branch using > > > > DPCD or > > > > * AUX message. The sequence is slot 1-63 allocated sequence > > > > for each > > > > diff --git a/drivers/gpu/drm/display/drm_dp_mst_topology.c > > > > b/drivers/gpu/drm/display/drm_dp_mst_topology.c > > > > index 847c10aa2098c..1990ff5dc7ddd 100644 > > > > --- a/drivers/gpu/drm/display/drm_dp_mst_topology.c > > > > +++ b/drivers/gpu/drm/display/drm_dp_mst_topology.c > > > > @@ -3342,7 +3342,8 @@ EXPORT_SYMBOL(drm_dp_add_payload_part1); > > > > * drm_dp_remove_payload() - Remove an MST payload > > > > * @mgr: Manager to use. > > > > * @mst_state: The MST atomic state > > > > - * @payload: The payload to write > > > > + * @old_payload: The payload with its old state > > > > + * @new_payload: The payload to write > > > > * > > > > * Removes a payload from an MST topology if it was successfully > > > > assigned a start slot. Also updates > > > > * the starting time slots of all other payloads which would have been > > > > shifted towards the start of > > > > @@ -3350,36 +3351,37 @@ EXPORT_SYMBOL(drm_dp_add_payload_part1); > > > > */ > > > > void drm_dp_remove_payload(struct drm_dp_mst_topology_mgr *mgr, > > > >struct drm_dp_mst_topology_state *mst_state, > > > > - struct drm_dp_mst_atomic_payload *payload) > > > > + const struct drm_dp_mst_atomic_payload > > > > *old_payload, > > > > + struct drm_dp_mst_atomic_payload > > > > *new_payload) > > > > { > > > > struct drm_dp_mst_atomic_payload *pos; > > > > bool send_remove = false; > > > > > > > > /* We failed to make the payload, so nothing to do */ > > > > - if (payload->vc_start_slot == -1) > > > > + if (new_payload->vc_start_slot == -1) > > > > return; >
Re: [Intel-gfx] [PATCH v2 02/17] drm/display/dp_mst: Handle old/new payload states in drm_dp_remove_payload()
On Tue, 2023-02-07 at 14:11 +0200, Imre Deak wrote: > On Mon, Feb 06, 2023 at 07:42:32PM -0500, Lyude Paul wrote: > > On Wed, 2023-02-01 at 17:04 +0200, Imre Deak wrote: > > > > > > Yes, this patch should have no functional change, so please check what > > > would apply to other drivers as well. > > > > > > Could you also check Ville's comment about storing start_slot elsewhere > > > than the atomic state (leaving only time_slots there). I wonder if that > > > would work, at least it would simplify things I think. > > > > Ideally it'd be nice if we could have all this info in the atomic commit, > > but > > yeah - you're not the first person to find this a bit confusing. FWIW > > though, > > the way we store start_slot right now is intended to follow the same kind of > > behavior we have with atomic CRTC commit dependencies, I think I also made > > it > > that way so all of the state would be in a single place but there's no hard > > requirement for it. > > As I understood the atomic state should be precomputed during atomic > check and not changed later during the commit phase. In case of > start_slot this would mean knowing in advance the actual (driver > specific) disabling / enabling sequence of the payloads, not sure how > feasible it is to figure that out. start_slot also changes dynamically It isn't feasible afaict, which was the main motivation for having the strange update procedure - this is pretty much something that needs to be determined at commit time. > as each payload is disabled, so these intermediate versions of the field > would need to be tracked somehow separately from the final (precomputed) > versions. FWIW, the current way this works is that we call drm_dp_mst_atomic_wait_for_dependencies() in order to ensure that no one's currently writing to the start_slot field (remember, at this point we may not have atomic locks held anymore). Then at that point, start_slot in the new atomic state should hold the current in-hw start_slot value. start_slot isn't actually set to the new starting slot until drm_dp_add_payload_part1(). That of course as you pointed out, doesn't help us unless we read all of the start_slot values before we start changing payloads - since disabling payloads inevitably changes the start slot of payloads that come afterwards. I'll admit I'm a bit surprised this logic was wrong - if I recall the whole reason I assumed this was OK when writing this code was that since we're updating one payload at a time, ACT, repeat, that the start slots of each payload in hardware _would_ actually change in the same way we modify them in helpers. e.g., my expectation was if we had a bunch of payloads like this: Payload #1: 15 slots, start_slot=0 Payload #2: 15 slots, start_slot=15 Payload #3: 15 slots, start_slot=30 And then disabled say, payload #1, that immediately after we get the ACT that the payload table in hardware would look like this: Payload #2: 15 slots, start_slot=0 Payload #3: 15 slots, start_slot=15 But it sounds like it actually would look like this in the real world? Payload #2: 15 slots, start_slot=15 Payload #3: 15 slots, start_slot=30 So I'm curious, is there something I missed here? At what point does the MST hub at the other end decide that it's time to move the start slots back? (keep in mind, the MST specification does explicitly mention that there should never be holes in the payload table - so something has to be shifting the payloads back). > > > So if you guys think it'd be better design-wise to store this something > > else, > > I've got no strong feelings either way > > > > > > > For 0-2: > > > > > > > > Reviewed-by: Lyude Paul > > > > > > Thanks. > > > > > > > > > > > > > > > > > Cc: Lyude Paul > > > > > Cc: Ville Syrjälä > > > > > Cc: Ben Skeggs > > > > > Cc: Karol Herbst > > > > > Cc: Harry Wentland > > > > > Cc: Alex Deucher > > > > > Cc: Wayne Lin > > > > > Cc: sta...@vger.kernel.org # 6.1 > > > > > Cc: dri-de...@lists.freedesktop.org > > > > > Reviewed-by: Ville Syrjälä > > > > > Signed-off-by: Imre Deak > > > > > --- > > > > > .../amd/display/amdgpu_dm/amdgpu_dm_helpers.c | 2 +- > > > > > drivers/gpu/drm/display/drm_dp_mst_topology.c | 26 > > > > > ++- > > > > > drivers/gpu/drm/i915/display/intel_dp_mst.c | 4 ++- > > > > > drivers/gpu/drm/nouveau/dispnv50/disp.c | 2 +- > > > > > include/drm/display/drm_dp_mst_helper.h | 3 ++- > > > > > 5 files changed, 21 insertions(+), 16 deletions(-) > > > > > > > > > > diff --git > > > > > a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > > > > b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > > > > index a50319fc42b11..180d3893b68da 100644 > > > > > --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > > > > +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c > > > > > @@ -208,7 +208,7 @@ bool > > > > > dm_helpers_dp_mst_write_payload_allocation_table( > > > > > if (enable) >
Re: [Intel-gfx] [PATCH v2 02/17] drm/display/dp_mst: Handle old/new payload states in drm_dp_remove_payload()
On Tue, Feb 07, 2023 at 07:21:48PM -0500, Lyude Paul wrote: > On Tue, 2023-02-07 at 14:11 +0200, Imre Deak wrote: > > On Mon, Feb 06, 2023 at 07:42:32PM -0500, Lyude Paul wrote: > > > On Wed, 2023-02-01 at 17:04 +0200, Imre Deak wrote: > > > > > > > > Yes, this patch should have no functional change, so please check what > > > > would apply to other drivers as well. > > > > > > > > Could you also check Ville's comment about storing start_slot elsewhere > > > > than the atomic state (leaving only time_slots there). I wonder if that > > > > would work, at least it would simplify things I think. > > > > > > Ideally it'd be nice if we could have all this info in the atomic commit, > > > but > > > yeah - you're not the first person to find this a bit confusing. FWIW > > > though, > > > the way we store start_slot right now is intended to follow the same kind > > > of > > > behavior we have with atomic CRTC commit dependencies, I think I also > > > made it > > > that way so all of the state would be in a single place but there's no > > > hard > > > requirement for it. > > > > As I understood the atomic state should be precomputed during atomic > > check and not changed later during the commit phase. In case of > > start_slot this would mean knowing in advance the actual (driver > > specific) disabling / enabling sequence of the payloads, not sure how > > feasible it is to figure that out. start_slot also changes dynamically > > It isn't feasible afaict, which was the main motivation for having the strange > update procedure - this is pretty much something that needs to be determined > at commit time. > > > as each payload is disabled, so these intermediate versions of the field > > would need to be tracked somehow separately from the final (precomputed) > > versions. > > FWIW, the current way this works is that we call > drm_dp_mst_atomic_wait_for_dependencies() in order to ensure that no one's > currently writing to the start_slot field (remember, at this point we may not > have atomic locks held anymore). Then at that point, start_slot in the new > atomic state should hold the current in-hw start_slot value. start_slot isn't > actually set to the new starting slot until drm_dp_add_payload_part1(). That > of course as you pointed out, doesn't help us unless we read all of the > start_slot values before we start changing payloads - since disabling payloads > inevitably changes the start slot of payloads that come afterwards. > > I'll admit I'm a bit surprised this logic was wrong - if I recall the whole > reason I assumed this was OK when writing this code was that since we're > updating one payload at a time, ACT, repeat, that the start slots of each > payload in hardware _would_ actually change in the same way we modify them in > helpers. e.g., my expectation was if we had a bunch of payloads like this: > > Payload #1: 15 slots, start_slot=0 > Payload #2: 15 slots, start_slot=15 > Payload #3: 15 slots, start_slot=30 > > And then disabled say, payload #1, that immediately after we get the ACT that > the payload table in hardware would look like this: > > Payload #2: 15 slots, start_slot=0 > Payload #3: 15 slots, start_slot=15 The above is the actual and expected HW state state yes. > But it sounds like it actually would look like this in the real world? > > Payload #2: 15 slots, start_slot=15 > Payload #3: 15 slots, start_slot=30 No, the problem currently is not that start_slot of the subsequent payloads are not shifted towards the beginning. Rather the atomic state doesn't get updated properly, becoming out of sync with the HW. For instance in a commit resizing payload #1, in the commit phase (intel_atomic_commit_tail()) will begin by removing payload #1. The initial state is old payload state new payload state Payload #1: 15 slots, start_slot=020 slots, start_slot=0 Payload #2: 15 slots, start_slot=15 15 slots, start_slot=15 Payload #3: 15 slots, start_slot=30 15 slots, start_slot=30 mgr->next_start_slot = 45 intel_mst_disable_dp() will pass the old MST and payload state to drm_dp_remove_payload(): The MST state was added during atomic check, since payload #1 changed, then intel_atomic_commit() -> drm_atomic_helper_swap_state() sets the MST current state (returned by drm_atomic_get_mst_topology_state()) to point to the old state. So at the point drm_dp_remove_payload() returns we have: old payload state new payload state Payload #1: 15 slots, start_slot=-1 20 slots, start_slot=0 Payload #2: 15 slots, start_slot=015 slots, start_slot=15 Payload #3: 15 slots, start_slot=15 15 slots, start_slot=30 mgr->next_start_slot = 30 then after re-enabling payload #1, after drm_dp_add_payload_part1() returns (passing to it the new MST and payload state) we have: old payload state new payload state Payload #1: 15 slots, start_slot=-1 20 slots, start_slot=30 Payload #2: 15 slots, start_slot=015 slots, start_slot=15 Pay
Re: [Intel-gfx] [PATCH v2 02/17] drm/display/dp_mst: Handle old/new payload states in drm_dp_remove_payload()
On Wed, 2023-02-08 at 09:41 +0200, Imre Deak wrote: > On Tue, Feb 07, 2023 at 07:21:48PM -0500, Lyude Paul wrote: > > On Tue, 2023-02-07 at 14:11 +0200, Imre Deak wrote: > > > > And then disabled say, payload #1, that immediately after we get the ACT > > that > > the payload table in hardware would look like this: > > > > Payload #2: 15 slots, start_slot=0 > > Payload #3: 15 slots, start_slot=15 > > The above is the actual and expected HW state state yes. > > > But it sounds like it actually would look like this in the real world? > > > > Payload #2: 15 slots, start_slot=15 > > Payload #3: 15 slots, start_slot=30 > > No, the problem currently is not that start_slot of the subsequent > payloads are not shifted towards the beginning. Rather the atomic state > doesn't get updated properly, becoming out of sync with the HW. For > instance in a commit resizing payload #1, in the commit phase > (intel_atomic_commit_tail()) will begin by removing payload #1. The > initial state is > > old payload state new payload state > Payload #1: 15 slots, start_slot=020 slots, start_slot=0 > Payload #2: 15 slots, start_slot=15 15 slots, start_slot=15 > Payload #3: 15 slots, start_slot=30 15 slots, start_slot=30 > > mgr->next_start_slot = 45 > > intel_mst_disable_dp() will pass the old MST and payload state to > drm_dp_remove_payload(): The MST state was added during atomic check, > since payload #1 changed, then intel_atomic_commit() -> > drm_atomic_helper_swap_state() sets the MST current state (returned by > drm_atomic_get_mst_topology_state()) to point to the old state. So at OK - this took me a while to wrap my head around but you're completely right. It appears I totally misunderstood where the state swapping actually happens during the check -> commit sequence. I think if that's how things work too then yeah, it definitely might not be a bad idea to move the start slot out of the atomic state :P. I guess we could just keep this in the mst manager struct instead of the commit state and make the rules for access be the same: protect them through commit ordering, and document that the proper way of accessing start values outside of the context of an atomic commit (if this was needed for some reason) is: * grab mst lock * call drm_dp_mst_atomic_wait_for_dependencies() * read values under lock Thank y'all again so much for helping out with this! It is super appreciated, and once you guys push these patches upstream I will look into adopting this for nouveau. I already poked some folks from AMD as well to make sure they're keeping an eye on this (although looking at the Cc I realize they were already added a while ago, whoops lol). > the point drm_dp_remove_payload() returns we have: > > old payload state new payload state > Payload #1: 15 slots, start_slot=-1 20 slots, start_slot=0 > Payload #2: 15 slots, start_slot=015 slots, start_slot=15 > Payload #3: 15 slots, start_slot=15 15 slots, start_slot=30 > > mgr->next_start_slot = 30 > > then after re-enabling payload #1, after drm_dp_add_payload_part1() > returns (passing to it the new MST and payload state) we have: > > old payload state new payload state > Payload #1: 15 slots, start_slot=-1 20 slots, start_slot=30 > Payload #2: 15 slots, start_slot=015 slots, start_slot=15 > Payload #3: 15 slots, start_slot=15 15 slots, start_slot=30 > > mgr->next_start_slot = 50 > > So in the new SW state payload #1 and #3 incorrectly overlap, with the > actual HW state being: > > Payload #1: 20 slots, start_slot=30 > Payload #2: 15 slots, start_slot=0 > Payload #3: 15 slots, start_slot=15 > > A subsequent commit will see the wrong start_slot in the SW state for > payload #2 (15) and #3 (30). > > > So I'm curious, is there something I missed here? At what point does the MST > > hub at the other end decide that it's time to move the start slots back? > > The hub shifts back payloads after the DPCD write command to 0x1c0 - > 0x1c2 to remove a payload. (The HW OTOH does the corresponding shift at > the point of disabling the stream, in intel_mst_post_disable_dp() -> > intel_disable_transcoder() for i915). > > > (keep in mind, the MST specification does explicitly mention that > > there should never be holes in the payload table - so something has to > > be shifting the payloads back). > > Right, the hubs I checked conform to this. > > > > > So if you guys think it'd be better design-wise to store this something > > > > else, > > > > I've got no strong feelings either way > > > > > > > > > > > For 0-2: > > > > > > > > > > > > Reviewed-by: Lyude Paul > > > > > > > > > > Thanks. > > > > > > > > > > > > > > > > > > > > > > > > > Cc: Lyude Paul > > > > > > > Cc: Ville Syrjälä > > > > > > > Cc: Ben Skeggs > > > > > > > Cc: Karol Herbst > > > > > > > Cc: Harry Wentland > > > > > > > Cc: Alex Deucher > > > > > > > Cc: Wayne Lin > > > > > > > Cc: sta...@vger.kernel.org # 6.1