[RFC v2 2/2] dt-bindings: mipi-dsi: Add dual-channel DSI related info

Thu Jun 7 22:50:00 UTC 2018

Am Donnerstag, 7. Juni 2018, 23:10:20 CEST schrieb Brian Norris:
> Hi,
> 
> I only have a little bit to add, but Heiko did solicit my opinion.

yep ... and I realized that I am/was way to attached to my (working)
endpoint-based thingy to really appreciate the other arguments ;-)

And your DSI-related writings below, did provide some new thought-
directions for me, so thanks for that.

> On Thu, Jun 07, 2018 at 03:25:18PM +0200, Heiko Stuebner wrote:
> > Am Donnerstag, 7. Juni 2018, 12:39:03 CEST schrieb Andrzej Hajda:
> > > On 07.06.2018 01:08, Heiko Stuebner wrote:
> > > > Am Mittwoch, 6. Juni 2018, 18:07:46 CEST schrieb Archit Taneja:
> > > >> On Wednesday 06 June 2018 04:16 PM, Heiko Stübner wrote:
> > > >>> Am Mittwoch, 6. Juni 2018, 12:21:16 CEST schrieb Archit Taneja:
> > > >>>> On Wednesday 06 June 2018 02:00 PM, Heiko Stübner wrote:
> > > >>>>> Am Mittwoch, 6. Juni 2018, 07:59:29 CEST schrieb Archit Taneja:
> > > >>>>>> On Monday 04 June 2018 05:47 PM, Heiko Stuebner wrote:
> > > >>>>>>> Am Donnerstag, 18. Januar 2018, 05:53:55 CEST schrieb Archit Taneja:
> > > >>>>>>>> Add binding info for peripherals that support dual-channel DSI. Add
> > > >>>>>>>> corresponding optional bindings for DSI host controllers that may
> > > >>>>>>>> be configured in this mode. Add an example of an I2C controlled
> > > >>>>>>>> device operating in dual-channel DSI mode.
> > > >>>>>>>> 
> > > >>>>>>>> Signed-off-by: Archit Taneja <architt at codeaurora.org>
> > > >>>>>>> 
> > > >>>>>>> Looks like a great solution for that problem, so
> > > >>>>>>> Reviewed-by: Heiko Stuebner <heiko at sntech.de>
> > > >>>>>>> 
> > > >>>>>>> As I'm looking into that for my rk3399-scarlet device right now and
> > > >>>>>>> couldn't find this patchset in the kernel yet, is it planned to
> > > >>>>>>> merge or refresh these binding changes or were problems encountered.
> > > >>>>>>> 
> > > >>>>>>> At least an Ack/Review from Rob seems to be missing.
> > > >>>>>> 
> > > >>>>>> I forgot about these patches. Rob had reviewed the first one in
> > > >>>>>> the set the second one still needed an Ack. I'll post a v3
> > > >>>>>> that adds the Reviewed-bys and fixes a small typo.
> > > >>>>> 
> > > >>>>> very nice ... because it looks like yesterday I managed to make the
> > > >>>>> Rockchip dsi work in dual mode following this.
> > > >>>>> 
> > > >>>>> But one question came up, do you really want two input ports on the
> > > >>>>> panel
> > > >>>>> side? I.e. hardware-wise, I guess the panel will have one 8-lane or so
> > > >>>>> input thatonly gets split up on the soc side onto 2 dsi controllers?
> > > >>>> 
> > > >>>> I think all dual DSI panels actually have 2 DSI controllers/parsers
> > > >>>> within them, one on each port. The MIPI DSI spec doesn't support 8
> > > >>>> lanes. Also, the pixels coming out of the host are distributed among
> > > >>>> the lanes differently than what would have been the case with a
> > > >>>> 'theoretical' 8 lane receiver.
> > > >>>> 
> > > >>>> Other than that, some dual DSI panels only accept DSI commands on the
> > > >>>> 'master' port, where as others expect the same command to be sent
> > > >>>> across
> > > >>>> both the ports.
> > > >>>> 
> > > >>>> Therefore, I think it's better to represent dual DSI panels having 2
> > > >>>> DSI input ports.
> > > >>>> 
> > > >>>> Your DT looks good to me.
> > > >>> 
> > > >>> Hmm, that doesn't match up then ;-) ... as my dt uses 2 endpoints
> > > >>> in one port for the dsi-links.
> > > >> 
> > > >> Sorry, I didn't notice you'd created two endpoints within a single port.
> > > >> 
> > > >> I don't think I'm particular about 2 ports vs 1 port with 2 endpoints.
> > > >> They both seem okay to me as long as we follow it consistently. I'm
> > > >> myself not 100% sure of how to figure where one should prefer endpoints
> > > >> over ports. Maybe someone more familiar with the of graph bindings
> > > >> could comment here.
> > > 
> > > Port should describe physical port, endpoint are used to describe
> > > multiple links to the same port. If the panel has two physical DSI
> > > interfaces it should use two ports.
> > 
> > That again leaves a bit of space for interpretation ;-) .
> 
> Does it really? DSI maxes out at 4 lanes, so more than 4 lanes is
> clearly not a single standard port. Additionally, the code we're
> currently using in Chrome OS actually sends the same commands to each
> DSI separately (I'm not sure if it's actually required by the panels
> were using), so again, they are to some extent logically independent.
>
> > A dual-dsi display is probably pretty useless with only one controller
> > connected to it, as its 4 lanes cannot satisfy the required 8 lanes
> > of the panel.
> 
> Probably, although technically you can usually display on half the
> screen, if you have only have 1 working DSI.
>
> > See [0] for current WIP panel addition.
> > [Was already on the lists but needs cleanup]
> > 
> > 
> > 
> > [0] https://github.com/mmind/linux-rockchip/commit/459bc1a1599377c2c5408724e82619a4602a953d#diff-5d6d6ddab4fd282102d23d2c02e835f8R381
> > 
> > 
> > > >>> The issue I see with using ports and not endpoints for dual-dsi links
> > > >>> is with distinguishing between input and output ports.
> > > >>> 
> > > >>> For a panel that's easy, as you every port will be an input port and if
> > > >>> you have 2, it's supposed dual-dsi. But for example I guess there might
> > > >>> exist some dual-dsi-to-something bridges, where you would end up
> > > >>> with say 3 (or even more) ports ... two dual-dsi inputs and 1 or more
> > > >>> outputs.
> > > >> 
> > > >> Okay, I get your point here. Although, even if the remote device had
> > > >> exactly 2 ports. Is it safe to assume that port #0 is an input port and
> > > >> port #1 is an output port? Is that the norm?
> > > > 
> > > > I don't think that anything like that is specified anywhere, so you cannot
> > > > assume anything about what a port contains.
> > > 
> > > Ports are defined per device-node and port number assignment should be
> > > described in particular device bindings. So device driver knows well
> > > which functions should be assigned to which port. But it is private
> > > matter of the device/device driver. Different device bindings can assign
> > > different number for input and output ports, but since ports are private
> > > thing for devices it should not be a problem.
> 
> Agreed. And if Heiko needs a specific example close to home: all the
> Rockchip display bindings do this similarly; for example
> display/rockchip/analogix_dp-rockchip.txt describes exactly which port
> numbers and enpoints mean what:
> 
> - ports: there are 2 port nodes with endpoint definitions as defined in
>   Documentation/devicetree/bindings/media/video-interfaces.txt.
>     Port 0: contained 2 endpoints, connecting to the output of vop.
>     Port 1: contained 1 endpoint, connecting to the input of panel.

Agreed here. My confusion stems from the issue that with the way
the code is right now, the dsi controller would need to know how
the ports of the panel are structured, to find its way to the 2nd
controller.

But then again, maybe the implementation just needs to change ;-) .

> > > I do not see why do you have such issue, could you describe it more.
> > > 
> > > >> I've at least seen one device (toshiba,tc358767 bridge) that can
> > > >> actually take either DPI as input or DSI based on how you configure it.
> > > >> There are 2 input ports here, port #0 for DSI and port #1 for DPI. Would
> > > >> it have made sense here to have a single port and 2 endpoints here too?
> > > > 
> > > > Nope, I guess having a port for DPI input, one port for DSI input makes
> > > > quite a lot of sense. And then you can have the input-specifics living in
> > > > the endpoints like dual links or so.
> > > > 
> > > >>> While the following argument might not be 100% valid from a dt-purity
> > > >>> standpoint implementing this might get hairy for _any_ operating system,
> > > >>> as you will need each panel/bridge to tell what the ports are used for.
> > > >> 
> > > >> Yeah.
> > > 
> > > Each panel/bridge should know what his ports are used for, but to who
> > > and why you want to communicate it?
> > 
> > I'll do that below :-)
> > 
> > 
> > > >>> I.e. in my endpoint based mapping, right now I have this nice and
> > > >>> generic
> > > >>> WIP function to parse the of_graph and get the master+slave nodes:
> > > >>> 
> > > >>> https://github.com/mmind/linux-rockchip/blob/tmp/rk3399-gru-bob-scarlet/
> > > >>> drivers/gpu/drm/rockchip/dw-mipi-dsi-rockchip.c#L697 [0]
> > > >> 
> > > >> I'd tried out something locally before posting this patch, I don't have
> > > >> the code for it, but I can describe the steps I took. This code expects
> > > >> the panel/bridge to have 2 input ports.
> > > > 
> > > > Which again would be very much panel-specific as you cannot assume
> > > > to much about the ports.
> > > > 
> > > >> 1. DSI0 host driver looks up its output port, gets the remote endpoint,
> > > >> and get this endpoint's parent (using
> > > >> of_graph_get_remote_port_parent()). Keeps the parent device node in a
> > > >> temp variable.
> > > >> 
> > > >> 2. DSI1 host driver does the same thing.
> > > >> 
> > > >> 3. DSI0 and DSI1 check whether their outputs are connected to the
> > > >> same device. If so, they're in dual DSI mode. If not, they are
> > > >> operating independently.
> > > >> 
> > > >> The positive of this approach is that we don't need to make any
> > > >> assumptions about the panel/bridge's port numbers, which is great.
> > > >> The negative is that our DSI controller instances now need to query
> > > >> each other, which can be messy, but not too hard to implement.
>
> > > I think detection of dual dsi and whole configuration thing of dual-dsi
> > > should be done using in-kernel frameworks (maybe new ones :) ).
> > > Could you describe necessary configuration steps to be performed, what
> > > info should be passed from who to who, etc. It would be easier to discuss.
> > 
> > The base setup is the rk3399 soc, which has 2 completely separate
> > dw-mipi-dsi controllers.
> > 
> > Gru-Scarlet (a ChromeOS tablet device) uses a high-res dual-DSI display
> > (2500x1500 pixels or so) which is driven by both controllers together to
> > achieve the required number of lanes to it.
> > 
> > So you end up with one (specific) controller being the master and the
> > other being the slave, deferring to the master for most everything.
> > As the controllers are somehow build to work together in that case
> > via special soc-specific settings.
> > 
> > See [1] for current WIP code, I lifted out of the chromeos kernel and
> > modified to not have that special rockchip,dual-channel =<&...>; property.
> > That patch was on the lists already last year I think.
> > 
> > So each dw-mipi-dsi instance needs to at least know that it is in a
> > master-slave-relationship and right now also needs access to the other
> > driver instance. Doing the master->slave access should be fine in that
> > special case, as the IP is the same type.
> > 
> > Also the crtc<->dsi interaction is quite a bit of handling-different between
> > one crtc talking to 2 DSIs or 2 crtc talking to the 2 DSIs separately.
> 
> That's probably the bigger key: to treat them as completely separate
> ports means that you get separate CRTCs, IIUC (I'll admit, I'm still a
> bit rusty on navigating some DRM concepts).

One thing I realized with your mention of DSI maxing out at 4 lanes is,
that this makes it easy to detect the existence of a dual-dsi situation ...
simply when the device reports 8 lanes.
(via of_find_mipi_dsi_device_by_node presumably)

As a dual-dsi situation requires a clock-master property in one,
both master and slave also would be able to determine their
master or slave status, thus the global dual-dsi config could be
done by the master (GRF stuff in the Rockchip case)

The only thing that makes my head explode now, is how to
make the slave actually react to settings sent to the master
in a sane way.

But that is a drm-specific implementation-detail, I guess ;-) .
And hopefully someone might have a great idea how this
could be done better than in my current implementation.

Heiko