by Pin-yen Lin

[permalink] [raw]

Subject: Re: [PATCH v15 04/10] dt-bindings: display: bridge: anx7625: Add mode-switch support

On Fri, May 12, 2023 at 1:30 AM Stephen Boyd <[email protected]> wrote:
>
> Quoting Pin-yen Lin (2023-05-09 20:41:54)
> > On Sat, Apr 29, 2023 at 12:47 PM Stephen Boyd <[email protected]> wrote:
> > >
> > > Good point. I'm suggesting to make the drm bridge chain into a tree. We
> > > need to teach drm_bridge core about a mux, and have some logic to
> > > navigate atomically switching from one output to another. I was talking
> > > with dianders@ and he was suggesting to use bridge attach/detach for
> > > this. I'm not sure that will work though because that hook is only
> > > called when the encoder is attached to the bridge.
> > >
> > > It may also turn out that this helps with DP multi-stream transport
> > > (MST). As far as I can recall DP MST doesn't mesh well with drm_bridge
> > > designs because it wants to operate on a drm_connector and
> > > drm_bridge_connector_init() wants to make only one drm_connector for a
> > > chain of bridges. If you squint, the anx7625 could be an MST "branch"
> > > that only supports one drm_connector being enabled at a time. Maybe that
> > > is what we should do here, make drm_bridge support creating more than
> > > one drm_connector and when there is a mux in the tree it walks both
> > > sides and runs a callback similar to struct
> > > drm_dp_mst_topology_cbs::add_connector() to tell the encoder that
> > > there's another possible drm_connector here.
> >
> > I have been surveying the approaches to change the bridge chain in
> > runtime, and I found this thread[1]. Quoting from Daniel:
>
> I find the lore links easier to read.
>
> > "... exchanging the bridge chain isn't supported, there's no locking
> > for that since it's assumed to be invariant over the lifetime of the
> > drm_device instance. The simplest way to make that happen right now is to
> > have 2 drm_encoder instances, one with the lvds bridge chain, the other
> > with the hdmi bridge chain, and select the right encoder/bridge chain
> > depending upon which output userspace picks.
> > ...
> > I wouldn't try to make bridge chains exchangeable instead, that's
> > headaches - e.g. with dp mst we've also opted for a bunch of fake
> > drm_encoders to model that kind of switching."
> >
> > I'm not sure how we register two encoders properly, though. Do we make
> > the encoder driver check all the downstream bridges and register two
> > drm_encoder when a bridge is acting as a mux?
>
> I honestly don't know because I'm not a DRM expert. Maybe Jagan or DRM
> bridge maintainers can add to the thread here.
>
> I kept reading the thread[2] and I think they settled on 2 drm_encoders
> because they're different display formats (LVDS vs. HDMI) and 2
> drm_connector structures. But then I watched the youtube video from this
> thread[3] and it seems like 1 drm_encoder is actually what should be
> done because there is really only DSI at the root. There's at least
> three people working on this topic of muxing displays though. Good news?
>
> The analogy between their gpio controlled mux and this anx part with a
> crosspoint switch is that the gpio is like the crosspoint switch, but
> the gpio is like a virtual mux? If the gpio is asserted for them, one
> display bridge is enabled (HDMI) and the other one is not (LVDS).
>
> In this case, the type-c cables may be connected to both
> usb-c-connectors and HPD may be asserted on both, but only one
> drm_connector will be able to attach to the 1 drm_encoder at a time. If
> we had 2 drm_encoders it would be possible to attach both encoders to
> both drm_connectors at the same time, which is impossible because it's a
> mux. Indicating that each connector is connected is OK when HPD is high
> on both usb-c-connectors. Userspace will have to attach an encoder to
> the drm_connector it wants to use, and the drm_connector will indicate
> which drm_encoders are possible for it, so all the information will be
> provided to userspace in this design.
>
> I think it really comes down to implementing the tree walking logic in
> the drm bridge APIs. The problem is things like
> drm_bridge_get_next_bridge(), drm_bridge_get_prev_bridge(), and
> drm_for_each_bridge_in_chain() which will have to operate on a tree
> instead of a list. There's also drm_bridge_chain_get_first_bridge() and
> drm_bridge_attach(). The good news is most of these APIs are used
> sparingly.
>
> Maybe the simplest way to handle this is to maintain a tree of bridges
> and generate bridge chains when an encoder is attached to a connector in
> the tree. Presumably there is only one connector possible for a leaf of
> the bridge tree, and one encoder at the root of the bridge chain. From
> the drm_bridge structure, you'll only be able to iterate over the bridge
> chain that is currently configured. Same for the encoder, you'll only be
> able to walk the currently configured bridge chain from struct
> drm_encoder::bridge_chain.

If I understand correctly, encoders are attached to connectors when
the driver initializes itself. Do you mean that the chain should be
generated when a connector is connected?

If so, "attach" becomes adding a bridge to the bridge "tree", and the
chain is only determined after a connector is plugged. When HPD fires,
we can use the .detect callback to find the right connector and form
the bridge chain. The following changes would need to be made to the
existing APIs:
(1) Some of the chain-traversing helpers (e.g.,
drm_for_each_bridge_in_chain()) need to be changed to traversing the
bridge tree.
(2) Drivers that hold a pointer to the connector need to either hold a
list of possible connectors or use a helper to get the active
connector at runtime.

I don't really know if this would work out. e.g., Do we really not
need the chain when the connector is disconnected? I will appreciate
any suggestions on this.

Regards,
Pin-yen

> This hinges on the idea that the bridge_chain is a list, not a tree, and
> that it only needs to exist when an encoder is attached to a connector.
> When an encoder isn't attached to a connector the bridges will be in the
> tree, and probably that tree structure will be maintained in the bridge
> driver itself knowing that there is one input side bridge and two output
> side bridges. When the input bridge is attached, it should be able to
> query the output side bridges for the connector that the encoder is
> attaching to and configure the mux and hook the input bridge to the
> correct output bridge.
>
> [2] https://lore.kernel.org/all/CAPj87rO7a=NbarOyZp1pE=19Lf2aGjWu7sru-eHwGjX0fpN+-A@mail.gmail.com/
> [3] https://lore.kernel.org/all/CAMty3ZAQyADGLVcB13qJdEy_BiZEKpyDfCr9QrM-ucFLFPZLcw@mail.gmail.com/

2023-05-30 06:36:53

by Jagan Teki

[permalink] [raw]

Subject: Re: [PATCH v15 04/10] dt-bindings: display: bridge: anx7625: Add mode-switch support

On Wed, May 10, 2023 at 9:12 AM Pin-yen Lin <[email protected]> wrote:
>
> +Jagan who worked on a similar design and initiated the thread.
>
> Hi Stephen,
>
> On Sat, Apr 29, 2023 at 12:47 PM Stephen Boyd <[email protected]> wrote:
> >
> > Quoting Pin-yen Lin (2023-04-20 02:10:46)
> > > On Thu, Apr 20, 2023 at 2:10 PM Stephen Boyd <[email protected]> wrote:
> > > >
> > > > Quoting Stephen Boyd (2023-04-13 17:22:46)
> > > > > Quoting Pin-yen Lin (2023-04-13 02:50:44)
> > > > > >
> > > > > > Actually the `mode-switch` property here is mainly because
> > > > > > `fwnode_typec_mux_get`[1] and `typec_mux_match`[2] only return matches
> > > > > > when the property is present. I am not sure what side effects would be
> > > > > > if I remove the ID-matching condition in `typec_mux_match`, so I added
> > > > > > the property here.
> > > > > >
> > > > > > Is it feasible to remove the `mode-switch` property here given the
> > > > > > existing implementation of the Type-C framework?
> > > > >
> > > > > Omitting the mode-switch property would require changes to the type-c
> > > > > framework.
> > > > >
> > > > > I'm wondering if we can have this anx driver register mode switches for
> > > > > however many endpoints exist in the output port all the time when the
> > > > > aux-bus node doesn't exist. Then the type-c framework can walk from the
> > > > > usb-c-connector to each connected node looking for a device that is both
> > > > > a drm_bridge and a mode-switch. When it finds that combination, it knows
> > > > > that the mode-switch has been found. This hinges on the idea that a
> > > > > device that would have the mode-switch property is a drm_bridge and
> > > > > would register a mode-switch with the type-c framework.
> > > > >
> > > > > It may be a little complicated though, because we would only register
> > > > > one drm_bridge for the input to this anx device. The type-c walking code
> > > > > would need to look at the graph endpoint, and find the parent device to
> > > > > see if it is a drm_bridge.
> > > >
> > > > I've been thinking more about this. I think we should only have the
> > > > 'mode-switch' property possible when the USB input pins (port@2) are
> > > > connected and the DPI input pins are connected (port@0). Probably you
> > > > don't have that case though?
> > >
> > > No we don't have the use case that uses the USB input pins on anx7625.
> > > >
> > > > In your case, this device should register either one or two drm_bridges
> > > > that connect to whatever downstream is actually muxing the 2 DP lanes
> > > > with the USB SS lanes onto the usb-c-connector.
> > >
> > > What do you mean by "muxing the 2 DP lanes with the USB SS lanes''? In
> > > our use case, the USB data lanes from both ports are connected to a
> > > USB hub, but the DP lanes are muxed by the crosspoint switch on
> > > anx7625. HPD and AUX for the external display are muxed by the EC. You
> > > can find the diagram at
> > > https://lore.kernel.org/linux-usb/[email protected]/
> >
> > I mean that you must have some sort of orientation switch hardware that
> > takes the 2 DP lanes and the 2 USB SuperSpeed "lanes" or "pairs" and
> > puts them all onto a usb-c-connector. The usb-c-connector node can't be
> > connected directly to the anx7625 in your diagram because there must be
> > some sort of "flipper" that does the orientation control. Otherwise the
> > usb-c-connector wouldn't work if the user flipped the cable. Probably
> > this is some TCPC or redriver controlled by the EC.
> >
> > >
> > > > If that is the EC for
> > > > ChromeOS, then the EC should have a binding that accepts some number of
> > > > input ports for DP. The EC would act as a drm_bridge, or in this case
> > > > probably two bridges, and also as two type-c switches for each
> > > > drm_bridge corresponding to the usb-c-connector nodes. When DP is on the
> > > > cable, the type-c switch/mux would signal to the drm_bridge that the
> > > > display is 'connected' via DRM_BRIDGE_OP_DETECT and struct
> > > > drm_bridge_funcs::detect(). Then the drm_bridge in this anx part would
> > > > implement struct drm_bridge_funcs::atomic_enable() and configure the
> > > > crosspoint switch the right way depending on the reg property of the
> > > > output node in port@1.
> > >
> > > So there will be two drm bridges that act as the downstreams for
> > > anx7625, and we find the downstream with connector_status_connected to
> > > configure the crosspoint switch? How do we support that kind of
> > > topology given that the drm bridge chain is currently a list? Are you
> > > suggesting making the bridge topology to a tree, or maintaining the
> > > two downstreams inside the anx7625 driver and not attaching them to
> > > the bridge chain?
> >
> > Good point. I'm suggesting to make the drm bridge chain into a tree. We
> > need to teach drm_bridge core about a mux, and have some logic to
> > navigate atomically switching from one output to another. I was talking
> > with dianders@ and he was suggesting to use bridge attach/detach for
> > this. I'm not sure that will work though because that hook is only
> > called when the encoder is attached to the bridge.
> >
> > It may also turn out that this helps with DP multi-stream transport
> > (MST). As far as I can recall DP MST doesn't mesh well with drm_bridge
> > designs because it wants to operate on a drm_connector and
> > drm_bridge_connector_init() wants to make only one drm_connector for a
> > chain of bridges. If you squint, the anx7625 could be an MST "branch"
> > that only supports one drm_connector being enabled at a time. Maybe that
> > is what we should do here, make drm_bridge support creating more than
> > one drm_connector and when there is a mux in the tree it walks both
> > sides and runs a callback similar to struct
> > drm_dp_mst_topology_cbs::add_connector() to tell the encoder that
> > there's another possible drm_connector here.
>
> I have been surveying the approaches to change the bridge chain in
> runtime, and I found this thread[1]. Quoting from Daniel:
> "... exchanging the bridge chain isn't supported, there's no locking
> for that since it's assumed to be invariant over the lifetime of the
> drm_device instance. The simplest way to make that happen right now is to
> have 2 drm_encoder instances, one with the lvds bridge chain, the other
> with the hdmi bridge chain, and select the right encoder/bridge chain
> depending upon which output userspace picks.
> ...
> I wouldn't try to make bridge chains exchangeable instead, that's
> headaches - e.g. with dp mst we've also opted for a bunch of fake
> drm_encoders to model that kind of switching."
>
> I'm not sure how we register two encoders properly, though. Do we make
> the encoder driver check all the downstream bridges and register two
> drm_encoder when a bridge is acting as a mux?
>
> [1]: https://www.spinics.net/lists/dri-devel/msg340511.html
>
> >
> > >
> > > Also, if we still register mode switches on the two downstream
> > > bridges, why do you prefer that over the original approach that
> > > register switches in the anx7625 driver?
> >
> > I prefer to not have a mode-switch property here for a couple reasons:
> >
> > 1. The binding is usb type-c specific, and in the case of the IT6505
> > part there is nothing that indicates this is a usb type-c piece of
> > hardware. The IT6505 is simply a display bridge. The anx7625 part
> > actually does accept usb signals though, but that isn't being used or
> > described here. That's where my disclaimer about mode-switch making
> > sense applies when the usb input is used.
> >
> > 2. Putting mode-switch into the graph endpoint nodes is awkward. It is
> > a device property, and graph nodes are not devices. Some patches in
> > this series have to work around this fact and special case the graph
> > walking logic to treat the graph itself as a place to look for the
> > property.
> >
> > 3. The mode-switch property probably isn't necessary at all. The DT
> > reviewers have been asking why it is needed. The EC driver that
> > registers the usb-c-connectors can be the mode-switch and the
> > orientation-switch. And in reality, it _is_ both. The DP signals and
> > the USB signals go to the TCPC/redriver that is controlled by the EC
> > and the EC is the device that's doing the mode switching to push DP and
> > USB through the TCPC/redriver out on the right pins of the
> > usb-c-connector.

I'm hoping that I'm replying to the right in-line comments.

The design I have tried was able to switch the runtime display, where
I initially booted the system with LVDS and once Linux comes up the
console then I can connect HDMI Out cable so-that switch bridge will
detect the HPD and Turn-Off LVDS and Turn-ON HDMI Display. If I unplug
HDMI Out cable then the switch bridge detects the change of HPD and
Turn-Off HDMI and Turn ON LVDS.

The design uses a switching bridge that acts as another gpio-interrupt
bridge that sits between the DSI controller and SN65DSI84/ADV7535
bridges. From the switch bridge, the final bridge pipeline will attach
during bridge ops attach call which is indeed uncommon with respect to
some of the mainline drivers. If the switch bridge detects the changes
in HPD then the handler will update kms and which indeed updates the
bridge chain at run-time. So may be unconventional but I don't find
any other solutions other than altering the bridge chain at run time.
Incidentally, I see more switch bridges coming on this design in the
market in order to produce more than one output.

May be it is a worthy topic to discuss further.

Thanks,
Jagan.