[Freedreno] [PATCH 01/25] drm/msm/dpu: fix hw ctl retrieval for mixer muxing
Sean Paul
sean at poorly.run
Thu Oct 11 14:51:47 UTC 2018
On Wed, Oct 10, 2018 at 11:35:56AM -0700, Jeykumar Sankaran wrote:
> On 2018-10-10 07:29, Sean Paul wrote:
> > On Tue, Oct 09, 2018 at 10:46:41PM -0700, Jeykumar Sankaran wrote:
> > > On 2018-10-09 11:07, Sean Paul wrote:
> > > > On Mon, Oct 08, 2018 at 09:27:18PM -0700, Jeykumar Sankaran wrote:
> > > > > Layer mixer/pingpong block counts and hw ctl block counts
> > > > > will not be same for all the topologies (e.g. layer
> > > > > mixer muxing to single interface)
> > > > >
> > > > > Use the encoder's split_role info to retrieve the
> > > > > respective control path for programming.
> > > > >
> > > > > Signed-off-by: Jeykumar Sankaran <jsanka at codeaurora.org>
> > > > > ---
> > > > > drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c | 9 ++++++---
> > > > > 1 file changed, 6 insertions(+), 3 deletions(-)
> > > > >
> > > > > diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c
> > > > b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c
> > > > > index 96cdf06..d12f896 100644
> > > > > --- a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c
> > > > > +++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c
> > > > > @@ -1060,6 +1060,7 @@ static void dpu_encoder_virt_mode_set(struct
> > > > drm_encoder *drm_enc,
> > > > >
> > > > > for (i = 0; i < dpu_enc->num_phys_encs; i++) {
> > > > > struct dpu_encoder_phys *phys = dpu_enc->phys_encs[i];
> > > > > + int ctl_index;
> > > > >
> > > > > if (phys) {
> > > > > if (!dpu_enc->hw_pp[i]) {
> > > > > @@ -1068,14 +1069,16 @@ static void dpu_encoder_virt_mode_set(struct
> > > > drm_encoder *drm_enc,
> > > > > return;
> > > > > }
> > > > >
> > > > > - if (!hw_ctl[i]) {
> > > > > + ctl_index = phys->split_role == ENC_ROLE_SLAVE ? 1
> > > > : 0;
> > > > > +
> > > >
> > > > What if MAX_CHANNELS_PER_ENC isn't 2? Similarly, what if num_phys_encs
> > >
> > > > MAX_CHANNELS_PER_ENC? It seems like there should be a more formal
> > > > relationship
> > > > between all of these verious values (num_of_h_tiles assumed to be <= 2
> > > > as
> > > > well).
> > > > If one of them changes beyond the assumed bound, the rest of the
> > driver
> > > > falls
> > > > over pretty hard.
> > > >
> > > MAX_CHANNELS_PER_ENC is set to 2 to represent HW limitation on the
> > chipset
> > > as
> > > we cannot gang up more than 2 LM chain to an interface. Supporting
> > > more
> > than
> > > 2
> > > might demand much larger changes than validating for boundaries.
> > >
> > > num_phys_enc is the max no of phys encoders we create as we are
> > > looping
> > > through
> > > num_of_h_tiles which cannot be more than priv->dsi array size.
> > >
> > > So its very unlikely we would expect these loops to go out of bound!
> >
> > For now, sure. However a new revision of hardware will be a pain to add
> > support
> > for if we add more assumptions, and secondly it makes it _really_ hard
> > to
> > understand the code if you don't have Qualcomm employee-level access to
> > the
> > hardware design :).
> >
> I am having a hard time understanding why you have to see these counts as
> "assumptions".
ctl_index = phys->split_role == ENC_ROLE_SLAVE ? 1 : 0
if (!hw_ctl[ctl_index]) {
Assumes that ARRAY_SIZE(hw_ctl) > 1 if phys is a slave
if (!dpu_enc->hw_pp[i]) {
Assumes that ARRAY_SIZE(hw_pp) >= ARRAY_SIZE(dpu_enc->phys_encs)
>
> Except for MAX_CHANNELS_PER_ENC, all the other counts are either calculated
> or derived from the other modules linked to the topology.
>
> h_tiles is the drm_connector terminology which represents the number of
> panels
> the display is driving. We use this information to determine the HW
> block chains in the MDP. HW blocks counts (pp or ctl) need not be same
> as the h_tile count to replace them with.
I must be missing something, because it seems that there's a 1:1 mapping of
num_h_tiles : num_phys_encs
And for each physical encoder, there's exactly one intf, pp, ctl assigned to it.
The issue I had was that when we loop through N phys_encs, we assume that there
are at least N hw blocks.
I think the takeaway is that I don't understand the hardware and its constraints
as well as you, so if this makes sense to you, let's go with it. Perhaps I'll
stare at it a while longer to try to refine my mental model.
Sean
>
> I believe maintaining the counts independently at each layer allows us to
> have more
> flexibility to support independent HW chaining for future revisions.
>
> Would it be more convincing if I get the MAX_CHANNELS_PER_ENC value from
> catalog.c?
>
> > So this is why I'm advocating for the reduction of the number of
> > "num_of_"
> > values we assume are all in the same range. It's a lot easier to
> > understand the
> > hardware when you can see that a phys encoder is needed per h tile, and
> > that a
> > ctl/pp is needed per phys encoder.
> This is exactly the idea I don't want to convey to the reader. For the LM
> merge path,
> each phys encoder will not be having its own control. Based on the topology
> we
> are supporting, HW block counts can vary. We can even drive:
> - 2 interfaces with 1 ctl and 1 ping pong
> - 1 interface with 1 ctl and 2 ping pongs
> - 1 interface with 1 ctl and 1 ping pong
>
> Thanks,
> Jeykumar S.
>
> >
> > Anyways, just my $0.02.
> >
> > Sean
> >
> > >
> > > Thanks,
> > > Jeykumar S.
> > > >
> > > > > + if (!hw_ctl[ctl_index]) {
> > > > > DPU_ERROR_ENC(dpu_enc, "no ctl block
> > > > assigned"
> > > > > - "at idx: %d\n", i);
> > > > > + "at idx: %d\n", ctl_index);
> > > > > return;
> > > >
> > > > When you return on error here, should you give back the resources that
> > > > you've
> > > > already provisioned?
> > > >
> > > > > }
> > > > >
> > > > > phys->hw_pp = dpu_enc->hw_pp[i];
> > > > > - phys->hw_ctl = hw_ctl[i];
> > > > > + phys->hw_ctl = hw_ctl[ctl_index];
> > > > >
> > > > > phys->connector = conn->state->connector;
> > > > > if (phys->ops.mode_set)
> > > > > --
> > > > > The Qualcomm Innovation Center, Inc. is a member of the Code Aurora
> > > > Forum,
> > > > > a Linux Foundation Collaborative Project
> > > > >
> > > > > _______________________________________________
> > > > > Freedreno mailing list
> > > > > Freedreno at lists.freedesktop.org
> > > > > https://lists.freedesktop.org/mailman/listinfo/freedreno
> > >
> > > --
> > > Jeykumar S
>
> --
> Jeykumar S
--
Sean Paul, Software Engineer, Google / Chromium OS
More information about the Freedreno
mailing list