Date:   Thu, 07 Oct 2021 15:29:58 -0700
From:   abhinavk@...eaurora.org
To:     Bjorn Andersson <bjorn.andersson@...aro.org>
Cc:     Stephen Boyd <swboyd@...omium.org>,
        Daniel Vetter <daniel@...ll.ch>,
        David Airlie <airlied@...ux.ie>,
        Dmitry Baryshkov <dmitry.baryshkov@...aro.org>,
        Kalyan Thota <kalyan_t@...eaurora.org>,
        Kuogee Hsieh <khsieh@...eaurora.org>,
        Rob Clark <robdclark@...il.com>, Sean Paul <sean@...rly.run>,
        Rob Herring <robh+dt@...nel.org>,
        linux-arm-msm@...r.kernel.org, dri-devel@...ts.freedesktop.org,
        freedreno@...ts.freedesktop.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH v4 5/7] drm/msm/dp: Support up to 3 DP controllers

Hi Bjorn and Stephen

On 2021-10-06 13:39, Bjorn Andersson wrote:
> On Wed 06 Oct 11:59 PDT 2021, Stephen Boyd wrote:
> 
>> Quoting Bjorn Andersson (2021-10-06 11:05:09)
>> > On Wed 06 Oct 10:19 PDT 2021, Stephen Boyd wrote:
>> >
>> > > Quoting Bjorn Andersson (2021-10-06 10:07:17)
>> > > > On Tue 05 Oct 21:26 PDT 2021, Stephen Boyd wrote:
>> > > >
>> > > > > Quoting Bjorn Andersson (2021-10-05 19:37:52)
>> > > > > > On Tue 05 Oct 19:06 PDT 2021, Stephen Boyd wrote:
>> > > > > >
>> > > > > > > Quoting Bjorn Andersson (2021-10-05 18:43:16)
>> > > > > > > > On Tue 05 Oct 17:43 PDT 2021, Stephen Boyd wrote:
>> > > > > > > >
>> > > > > > > > > Quoting Bjorn Andersson (2021-10-05 16:13:21)
>> > > > > > > > > > diff --git a/drivers/gpu/drm/msm/dp/dp_display.c b/drivers/gpu/drm/msm/dp/dp_display.c
>> > > > > > > > > > index bdaf227f05dc..674cddfee5b0 100644
>> > > > > > > > > > --- a/drivers/gpu/drm/msm/dp/dp_display.c
>> > > > > > > > > > +++ b/drivers/gpu/drm/msm/dp/dp_display.c
>> > > > > > > > > > @@ -1233,7 +1239,7 @@ static int dp_display_probe(struct platform_device *pdev)
>> > > > > > > > > >         if (!dp)
>> > > > > > > > > >                 return -ENOMEM;
>> > > > > > > > > >
>> > > > > > > > > > -       desc = dp_display_get_desc(pdev);
>> > > > > > > > > > +       desc = dp_display_get_desc(pdev, &dp->id);
>> > > > > > > > >
>> > > > > > > > > I'm sad that dp->id has to match the number in the SoC specific
>> > > > > > > > > dpu_intf_cfg array in drivers/gpu/drm/msm/disp/dpu1/dpu_hw_catalog.c
>> > > > > > > > > still. Is there any way we can avoid that? Also, notice how those arrays
>> > > > > > > > > already have INTF_DP macros, which makes me think that it may be better
>> > > > > > > > > to connect this to those arrays instead of making an msm_dp_desc
>> > > > > > > > > structure and then make sure the 'type' member matches a connector
>> > > > > > > > > type number. Otherwise this code is super fragile.
>> > > > > > > > >
>> > > > > > > >
>> > > > > > > > I'm afraid I don't understand what you're proposing. Or which part you
>> > > > > > > > consider fragile, the indices of the INTF_DP instances aren't going to
>> > > > > > > > move around...
>> > > > > > > >
>> > > > > > > > I have N instances of the DP driver that I need to match to N entries
>> > > > > > > > from the platform specific intf array, I need some stable reference
>> > > > > > > > between them. When I started this journey I figured I could rely on the
>> > > > > > > > of_graph between the DPU and the interface controllers, but the values
>> > > > > > > > used there today are just bogus, so that was a no go.
>> > > > > > > >
>> > > > > > > > We can use whatever, as long as _dpu_kms_initialize_displayport() can
>> > > > > > > > come up with an identifier to put in h_tile_instance[0] so that
>> > > > > > > > dpu_encoder_setup_display() can find the relevant INTF.
>> > > > > > > >
>> > > > > > >
>> > > > > > > To make it more concrete we can look at sc7180
>> > > > > > >
>> > > > > > > static const struct dpu_intf_cfg sc7180_intf[] = {
>> > > > > > >         INTF_BLK("intf_0", INTF_0, 0x6A000, INTF_DP, 0, 24,
>> > > > > > > INTF_SC7180_MASK, MDP_SSPP_TOP0_INTR, 24, 25),
>> > > > > > >                                                      ^
>> > > > > > >                                                      |
>> > > > > > >
>> > > > > > > intf0 is irrelevant. Also the address is irrelevant. But here we have a
>> > > > > > > zero, the number after INTF_DP, and that is very relevant. That number
>> > > > > > > needs to match the dp->id. Somewhere we have a match between
>> > > > > > > controller_id and dp->id in the code.
>> > > > > >
>> > > > > > That number (the 0, not INTF_0) is what the code matches against dp->id
>> > > > > > in _dpu_kms_initialize_displayport(), in order to figure out that this
>> > > > > > is INTF_0 in dpu_encoder_setup_display().
>> > > > > >
>> > > > > > I.e. look at the sc8180x patch:
>> > > > > >
>> > > > > > INTF_BLK("intf_0", INTF_0, 0x6A000, INTF_DP, 0, 24, INTF_SC8180X_MASK, MDP_SSPP_TOP0_INTR, 24, 25),
>> > > > > > INTF_BLK("intf_1", INTF_1, 0x6A800, INTF_DSI, 0, 24, INTF_SC8180X_MASK, MDP_SSPP_TOP0_INTR, 26, 27),
>> > > > > > INTF_BLK("intf_2", INTF_2, 0x6B000, INTF_DSI, 1, 24, INTF_SC8180X_MASK, MDP_SSPP_TOP0_INTR, 28, 29),
>> > > > > > /* INTF_3 is for MST, wired to INTF_DP 0 and 1, use dummy index until this is supported */
>> > > > > > INTF_BLK("intf_3", INTF_3, 0x6B800, INTF_DP, 999, 24, INTF_SC8180X_MASK, MDP_SSPP_TOP0_INTR, 30, 31),
>> > > > > > INTF_BLK("intf_4", INTF_4, 0x6C000, INTF_DP, 1, 24, INTF_SC8180X_MASK, MDP_SSPP_TOP0_INTR, 20, 21),
>> > > > > > INTF_BLK("intf_5", INTF_5, 0x6C800, INTF_DP, 2, 24, INTF_SC8180X_MASK, MDP_SSPP_TOP0_INTR, 22, 23),
>> > > > > >
>> > > > > > Where the DP driver defines the 3 controllers with dp->id of 0, 1 and 2,
>> > > > > > which the DPU code will match against to INTF_0, INTF_4 and INTF_5.
>> > > > > >
>> > > > >
>> > > > > Yep. I'm saying that having to make that number in this intf array match
>> > > > > the order of the register mapping descriptor array is fragile. Why can't
>> > > > > we indicate the interface is DP or eDP with INTF_DP or INTF_EDP and then
>> > > > > map from the descriptor array to this intf array somehow so that the
>> > > > > order of the descriptor array doesn't matter? Then we don't have to put
>> > > > > the connector type in the descriptor array, and we don't have to keep
>> > > > > the order of the array a certain way to match this intf descriptor.
>> > > > >
>> > > > > Maybe
>> > > > >
>> > > > >       struct msm_dp_desc {
>> > > > >               phys_addr_t io_start;
>> > > > >               unsigned int id;
>> > > >
>> > > > The INTF_<N> constants are a property of the DPU driver and not
>> > > > available in the DP driver and the msm_dp struct is a property of the DP
>> > > > driver and can't be dereferenced in the DPU driver.
>> > > >
>> > > > The proposed way around this is that the descs array defines the order
>> > > > in priv->dp[N] and this N is used as controller_id.
>> > >
>> > > I'm pretty sure I'm following along.
>> > >
>> > > >
>> > > > So the only thing that I don't find straight forward here is that the
>> > > > eDP controller is considered just a DP controller, so you have to use
>> > > > INTF_DP, <N> for that, and not just INTF_EDP, 0.
>> > > >
>> > > > >       };
>> > > > >
>> > > > > and then have msm_dp_desc::id equal INTF_<N> and then look through the
>> > > > > intf from DPU here in the DP driver to find the id and type of connector
>> > > > > that should be used by default? Still sort of fragile because the only
>> > > > > connection is an unsigned int which isn't great, but at least it's
>> > > > > explicit instead of implicit based on the array order.
>> > > >
>> > > > No matter how I look at this, you need to put some number somewhere here
>> > > > that will be used to match up the INTF with the right DSI/DP encoder.
>> > >
>> > > Correct.
>> > >
>> > > >
>> > > > Using the proposed number scheme follows the numbering of all the DP
>> > > > controllers from the documentation.
>> > > >
>> > >
>> > > Maybe I can make a better example. I have this for sc7280 in dpu_hw_catalog.c:
>> > >
>> > >       static const struct dpu_intf_cfg sc7280_intf[] = {
>> > >               INTF_BLK("intf_0", INTF_0, 0x34000, INTF_DP, CONTROLLER_ID_A, 24,
>> > > INTF_SC7280_MASK, MDP_SSPP_TOP0_INTR, 24, 25),
>> > >               INTF_BLK("intf_1", INTF_1, 0x35000, INTF_DSI, 0, 24,
>> > > INTF_SC7280_MASK, MDP_SSPP_TOP0_INTR, 26, 27),
>> > >               INTF_BLK("intf_5", INTF_5, 0x39000, INTF_DP, CONTROLLER_ID_B, 24,
>> > > INTF_SC7280_MASK, MDP_SSPP_TOP0_INTR, 22, 23),
>> > >       };
>> > >
>> > > And then this array for sc7280 in dp_display.c:
>> > >
>> > >       static const struct msm_dp_desc sc7280_dp_cfg = {
>> > >               .desc = {
>> > >                       [CONTROLLER_ID_A] = { 0xaea0000, DRM_MODE_CONNECTOR_eDP },
>> > >                       [CONTROLLER_ID_B] = { 0xae90000, DRM_MODE_CONNECTOR_DisplayPort },
>> > >               },
>> > >               .num_dp = 2,
>> > >       };
>> > >
>> > > So these two arrays must match based on CONTROLLER_ID_{A,B}. I don't
>> > > like having to make these two numbers match so if it was explicit, even
>> > > possibly by having a bunch of macros put in both places then I would be
>> > > happy. I spent a few hours when I messed up the order of the
>> > > sc7280_dp_cfg.desc array trying to figure out why things weren't
>> > > working.
>> >
>> > So essentially, you didn't know that the controller_id has to match the
>> > index in priv->dsi[] and priv->dp[] and providing a define for them
>> > would make this more obvious?
>> 
>> Now you got it!
>> 
>> >
>> > I think per your argument the 0 following INTF_DSI should also be using
>> > this scheme, so we'd have multiple CONTROLLER_ID_A, which probably is
>> > confusing as well.
>> 
>> Agreed.
>> 
>> >
>> > I tried it out with below patch; it documents the relationship, provides
>> > constants for the magic 2 and 3 for number of DSI and DP controllers in
>> > struct msm_drm_private.
>> >
>> > I like it.
>> 
>> Thanks. I prefer this approach as well.
> 
> Sweet, I'll update my patch set accordingly.
Yes, I also agree with this approach to better document the relationship 
between the hw_catalog array
and the dp_display msm_dp_desc

> 
>> I can see now why qcom always wants to change the output ports on the
>> DPU node in DT to match the INTF number. If they would have described
>> this problem it may have made sense to have the graph endpoints with
>> reg properties matching the interface number in the intf array. Sigh.
> 
> Yes, I think the supposed design is that you should use the of_graph 
> and
> then call of_find_possible_crtcs() to figure out your links.
> 
> Unfortunately that doesn't work with the design of the DPU driver,
> because the crtcs doesn't represent the INTFs - and as you say, the
> existing of_graphs are full of incorrect data.
> 
> Which also means that I don't know why we keep filling out the 
> of_graph,
> because afaict it's not used and it contains invalid information.
> 
> Thanks,
> Bjorn

Powered by blists - more mailing lists