Re: [PATCH RFC] drm/msm/dpu: Force modeset if new CTLs have been reserved

[Dmitry Baryshkov]

On Wed, 5 Mar 2025 at 00:43, Abhinav Kumar <quic_abhinavk@xxxxxxxxxxx> wrote:
>
>
>
> On 3/4/2025 12:42 PM, Dmitry Baryshkov wrote:
> > On Tue, Mar 04, 2025 at 11:38:24AM -0800, Abhinav Kumar wrote:
> >>
> >>
> >> On 3/3/2025 9:32 PM, Dmitry Baryshkov wrote:
> >>> On Tue, 4 Mar 2025 at 03:44, Jessica Zhang <quic_jesszhan@xxxxxxxxxxx> wrote:
> >>>>
> >>>>
> >>>>
> >>>> On 3/3/2025 3:49 PM, Dmitry Baryshkov wrote:
> >>>>> On Mon, Mar 03, 2025 at 10:28:00AM -0800, Jessica Zhang wrote:
> >>>>>> If new CTLs are reserved by CRTC but atomic_enable() is skipped, the
> >>>>>> encoders will configure the stale CTL instead of the newly reserved one.
> >>>>>
> >>>>> The CTLs are propagates in .atomic_mode_set(), not in .atomic_enable().
> >>>>
> >>>> Hi Dmitry,
> >>>>
> >>>> Yes, sorry mixed up the two function ops here and in my reply in the CWB
> >>>> thread.
> >>>>
> >>>>>
> >>>>>>
> >>>>>> Avoid this by setting mode_changed to true if new CTLs have been
> >>>>>> reserved by CRTC.
> >>>>>
> >>>>> This looks very strange. First we reserve new CTLs when there is a
> >>>>> modeset requested. Then on one of the next commits we detect that
> >>>>> encoder has stale CTLs and try to upgrade the commit to full modeset
> >>>>> (while the user might not have .allow_modeset set to true for whatever
> >>>>> reason, e.g. because only ACTIVE is changed).
> >>>>
> >>>> Ah I see what you mean. I think this is an issue with how/when we're
> >>>> calling dpu_rm_reserve(). Since RM reservation is tied to
> >>>> atomic_check(), we aren't able to force a modeset based on HW block
> >>>> reservation. The only reason we were able to avoid this issue with
> >>>> needs_cdm is because needs_cdm didn't depend on the CDM HW block index.
> >>>>
> >>>> I think there's not really a good way to avoid this other than flipping
> >>>> the order of the msm_atomic_check to drm_helper_atomic_check ->
> >>>> dpu_kms.check_mode_changed -> drm_atomic_helper_check_modeset().
> >>>
> >>> No-no-no. This would require a full drm_atomic_helper_check() call
> >>> again, after the check_mode_changed() callback. But again, this should
> >>> not be required at all. The whole point of .check_mode_changed() is to
> >>> be called before performing full atomic_check() chains.
> >>>
> >>
> >> Right but the documentation also allows calling
> >> drm_atomic_helper_check_modeset() again. We are looking at all options even
> >> moving forward and not just this issue.
> >
> > It does. But I'd rather not do it. Especially not in this case: we
> > perfectly know in advance if hw resources were reallocated or not.
> >
> >>
> >>>>
> >>>> What do you think? It seems to be valid given the examples in the DRM
> >>>> docs [1]
> >>>>
> >>>> [1]
> >>>> https://elixir.bootlin.com/linux/v6.13.5/source/drivers/gpu/drm/drm_atomic_helper.c#L610
> >>>>
> >>>>>
> >>>>> Could you please check if the following change fixes the issue: in
> >>>>> crtc_set_mode() replace the raw !new_crtc_state->mode_changed check with
> >>>>> the drm_atomic_crtc_needs_modeset() call?
> >>>>
> >>>> This also fixes the DPMS failures. IIRC Abhinav had suggested a similar
> >>>> change to fix a different issue [2] and you gave some feedback on
> >>>> avoiding mode_set() for enable/disable calls which don't have mode_changed.
> >>>
> >>> After reading the documentation for
> >>> drm_encoder_helper_funcs.atomic_mode_set() and looking around, I think
> >>> the issue is in the handling of the DPMS functions. I might have a fix
> >>> for the issue.
> >>>
> >>>> Also, while this may fix the CWB CI failures, wouldn't the issue still
> >>>> remain regarding how to force modeset for changes in HW block reservation?
> >>>
> >>> I think it is the other way around: HW block reservation is only
> >>> changed if there is a modeset. I'm currently testing my theory :-) We
> >>> were performing HW reassignment if drm_atomic_crtc_needs_modeset() was
> >>> true. However this function returns true in one of the cases, where
> >>> there is no actual modeset happening (and it's even documented this
> >>> way) - when only DPMS call has happened (in other words, when
> >>> .active_changed = true, but two other bits are false). It is required
> >>> not to reassign HW resources in such a case. So, I think, a correct
> >>> fix is to change the condition in dpu_crtc_atomic_check().
> >>>
> >>
> >> Yes, Jessica had also suggested this option. This will work because now the
> >> resource re-assignment will not happen and hence will avoid the issue. The
> >> documentation of DPMS was not fully clear. So it said, the same thing you
> >> mentioned, that when active has changed there is no need to reassign
> >> hardware resources but I was not sure if that would impact normal
> >> suspend/resume because across suspend/resume hardware resources need to be
> >> cleared / re-assigned.
> >
> > Suspend / resume is handled via the helpers, which commit an
> > all-disabled state and then commit a previous state. Other than that,
> > there is no need to clear resource assignments during suspend resume.
> > They should be disabled, but there is no need to drop the assignment.
> >
>
> We will need to do some testing to make sure this does not introduce
> some other regressions. You can post your change, we can discuss it there.

Posted in another thread.

>
> >> I do still think that, even if this also works, we
> >> will still run into issues when we will need to force a mode_changed based
> >> on resource assignment of other encoder based blocks such as DSC or PP etc.
> >
> > This is what dpu_encoder_needs_modeset() is for. On the other hand,
> > note, if we reassigned resources, it means that either mode_changes or
> > connectors_changed is set. And the DRM framework will call encoder's
> > atomic_mode_set() in such a case.
> >
>
> Like we wrote, dpu_encoder_needs_modeset() is not useful without the
> sequence change in msm_atomic_check because dpu_encoder_needs_modeset()
> is called before the resource assignments are done today.

Resources are reassigned if the mode_changed or connector_changed is
true. There are no other ways to reassign resources.  Under these
circumstances the encoder's atomic_mode_set() callback is called,
propagating those changes to the encoder. What is the problem that you
are trying to solve here?

>
> >>
> >>>>
> >>>> [2] https://gitlab.freedesktop.org/drm/msm/-/issues/59
> >>>>
> >>>> Thanks,
> >>>>
> >>>> Jessica Zhang
> >>>>
> >>>>>
> >>>>>>
> >>>>>> Note: This patch only adds tracking for the CTL reservation, but eventually
> >>>>>> all HW blocks used by encoders (i.e. DSC, PINGPONG, CWB) should have a
> >>>>>> similar check to avoid the same issue.
> >>>>>>
> >>>>>> Suggested-by: Abhinav Kumar <quic_abhinavk@xxxxxxxxxxx>
> >>>>>> Closes: https://lists.freedesktop.org/archives/freedreno/2025-February/036719.html
> >>>>>> Signed-off-by: Jessica Zhang <quic_jesszhan@xxxxxxxxxxx>
> >>>>>> ---
> >>>>>>     drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c    | 13 +++++++++++++
> >>>>>>     drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c | 12 ++++++++++++
> >>>>>>     drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.h |  1 +
> >>>>>>     3 files changed, 26 insertions(+)
> >>>>>>
> >>>>>> diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c b/drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c
> >>>>>> index 4073d821158c0..a1a8be8f5ab9f 100644
> >>>>>> --- a/drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c
> >>>>>> +++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c
> >>>>>> @@ -1406,19 +1406,32 @@ int dpu_crtc_check_mode_changed(struct drm_crtc_state *old_crtc_state,
> >>>>>>        struct drm_crtc *crtc = new_crtc_state->crtc;
> >>>>>>        bool clone_mode_enabled = drm_crtc_in_clone_mode(old_crtc_state);
> >>>>>>        bool clone_mode_requested = drm_crtc_in_clone_mode(new_crtc_state);
> >>>>>> +    struct dpu_crtc_state *cstate = to_dpu_crtc_state(new_crtc_state);
> >>>>>> +    uint32_t enc_ctl_mask = 0;
> >>>>>> +    uint32_t crtc_ctl_mask = 0;
> >>>>>> +    struct dpu_crtc_mixer *m;
> >>>>>>
> >>>>>>        DRM_DEBUG_ATOMIC("%d\n", crtc->base.id);
> >>>>>>
> >>>>>> +    for (int i = 0; i < cstate->num_mixers; i++) {
> >>>>>> +            m = &cstate->mixers[i];
> >>>>>> +            crtc_ctl_mask |= BIT(m->lm_ctl->idx - CTL_0);
> >>>>>> +    }
> >>>>>> +
> >>>>>>        /* there might be cases where encoder needs a modeset too */
> >>>>>>        drm_for_each_encoder_mask(drm_enc, crtc->dev, new_crtc_state->encoder_mask) {
> >>>>>>                if (dpu_encoder_needs_modeset(drm_enc, new_crtc_state->state))
> >>>>>>                        new_crtc_state->mode_changed = true;
> >>>>>> +            enc_ctl_mask |= dpu_encoder_get_ctls(drm_enc);
> >>>>>>        }
> >>>>>>
> >>>>>>        if ((clone_mode_requested && !clone_mode_enabled) ||
> >>>>>>            (!clone_mode_requested && clone_mode_enabled))
> >>>>>>                new_crtc_state->mode_changed = true;
> >>>>>>
> >>>>>> +    if (crtc_ctl_mask != enc_ctl_mask)
> >>>>>> +            new_crtc_state->mode_changed = true;
> >>>>>> +
> >>>>>>        return 0;
> >>>>>>     }
> >>>>>>
> >>>>>> diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c
> >>>>>> index a61598710acda..2f3101caeba91 100644
> >>>>>> --- a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c
> >>>>>> +++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c
> >>>>>> @@ -188,6 +188,7 @@ struct dpu_encoder_virt {
> >>>>>>
> >>>>>>        unsigned int dsc_mask;
> >>>>>>        unsigned int cwb_mask;
> >>>>>> +    unsigned int ctl_mask;
> >>>>>>
> >>>>>>        bool intfs_swapped;
> >>>>>>
> >>>>>> @@ -707,6 +708,13 @@ void dpu_encoder_update_topology(struct drm_encoder *drm_enc,
> >>>>>>        }
> >>>>>>     }
> >>>>>>
> >>>>>> +uint32_t dpu_encoder_get_ctls(struct drm_encoder *drm_enc)
> >>>>>> +{
> >>>>>> +    struct dpu_encoder_virt *dpu_enc = to_dpu_encoder_virt(drm_enc);
> >>>>>> +
> >>>>>> +    return dpu_enc->ctl_mask;
> >>>>>> +}
> >>>>>> +
> >>>>>>     bool dpu_encoder_needs_modeset(struct drm_encoder *drm_enc, struct drm_atomic_state *state)
> >>>>>>     {
> >>>>>>        struct drm_connector *connector;
> >>>>>> @@ -1155,6 +1163,7 @@ static void dpu_encoder_virt_atomic_mode_set(struct drm_encoder *drm_enc,
> >>>>>>        bool is_cwb_encoder;
> >>>>>>        unsigned int dsc_mask = 0;
> >>>>>>        unsigned int cwb_mask = 0;
> >>>>>> +    unsigned int ctl_mask = 0;
> >>>>>>        int i;
> >>>>>>
> >>>>>>        if (!drm_enc) {
> >>>>>> @@ -1245,11 +1254,14 @@ static void dpu_encoder_virt_atomic_mode_set(struct drm_encoder *drm_enc,
> >>>>>>                                "no ctl block assigned at idx: %d\n", i);
> >>>>>>                        return;
> >>>>>>                }
> >>>>>> +            ctl_mask |= BIT(phys->hw_ctl->idx - CTL_0);
> >>>>>>
> >>>>>>                phys->cached_mode = crtc_state->adjusted_mode;
> >>>>>>                if (phys->ops.atomic_mode_set)
> >>>>>>                        phys->ops.atomic_mode_set(phys, crtc_state, conn_state);
> >>>>>>        }
> >>>>>> +
> >>>>>> +    dpu_enc->ctl_mask = ctl_mask;
> >>>>>>     }
> >>>>>>
> >>>>>>     static void _dpu_encoder_virt_enable_helper(struct drm_encoder *drm_enc)
> >>>>>> diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.h b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.h
> >>>>>> index ca1ca2e51d7ea..70b03743dc346 100644
> >>>>>> --- a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.h
> >>>>>> +++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.h
> >>>>>> @@ -91,6 +91,7 @@ bool dpu_encoder_needs_modeset(struct drm_encoder *drm_enc, struct drm_atomic_st
> >>>>>>
> >>>>>>     void dpu_encoder_prepare_wb_job(struct drm_encoder *drm_enc,
> >>>>>>                struct drm_writeback_job *job);
> >>>>>> +uint32_t dpu_encoder_get_ctls(struct drm_encoder *drm_enc);
> >>>>>>
> >>>>>>     void dpu_encoder_cleanup_wb_job(struct drm_encoder *drm_enc,
> >>>>>>                struct drm_writeback_job *job);
> >>>>>>
> >>>>>> ---
> >>>>>> base-commit: 866e43b945bf98f8e807dfa45eca92f931f3a032
> >>>>>> change-id: 20250228-force-modeset-hw-ctl-d02b80a2bb4c
> >>>>>> prerequisite-change-id: 20241222-drm-dirty-modeset-88079bd27ae6:v2
> >>>>>> prerequisite-patch-id: 0c61aabfcd13651203f476985380cbf4d3c299e6
> >>>>>> prerequisite-patch-id: c6026f08011c288fd301676e9fa6f46d0cc1dab7
> >>>>>> prerequisite-patch-id: b0cb06d5c88791d6e4755d879ced0d5050aa3cbf
> >>>>>> prerequisite-patch-id: fd72ddde9dba0df053113bc505c213961a9760da
> >>>>>> prerequisite-change-id: 20250209-dpu-c3fac78fc617:v2
> >>>>>> prerequisite-patch-id: c84d2b4b06be06384968429085d1e8ebae23a583
> >>>>>> prerequisite-patch-id: fb8ea7b9e7c85fabd27589c6551108382a235002
> >>>>>> prerequisite-change-id: 20250211-dither-disable-b77b1e31977f:v1
> >>>>>> prerequisite-patch-id: 079e04296212b4b83d51394b5a9b5eea6870d98a
> >>>>>> prerequisite-change-id: 20240618-concurrent-wb-97d62387f952:v6
> >>>>>> prerequisite-patch-id: b52034179741dc182aea9411fd446e270fdc69d1
> >>>>>> prerequisite-patch-id: bc472765a7d5214691f3d92696cc8b0119f3252e
> >>>>>> prerequisite-patch-id: c959bc480e96b04297ebaf30fea3a68bbac69da6
> >>>>>> prerequisite-patch-id: f7db8449b241a41faac357d9257f8c7cb16503ec
> >>>>>> prerequisite-patch-id: 7beb73131d0ab100f266fcd3c1f67c818a3263f4
> >>>>>> prerequisite-patch-id: c08cbb5cf4e67e308afd61fdad6684b89429d3b6
> >>>>>> prerequisite-patch-id: a4e343143b8fbe98ae4aa068cc459c750105eb9d
> >>>>>> prerequisite-patch-id: 1d09edcf12ef7e7ab43547eefacae5b604b698e9
> >>>>>> prerequisite-patch-id: 0008f9802bfd3c5877267666cceb7608203e5830
> >>>>>> prerequisite-patch-id: 49402eb767c97915faf2378c5f5d05ced2dcfdac
> >>>>>> prerequisite-patch-id: 522be2a6b5fe4e3a2d609526bb1539f9bc6f828f
> >>>>>> prerequisite-patch-id: 031da00d0fffd522f74d682a551362f3ecda0c71
> >>>>>> prerequisite-patch-id: 9454cec22231a8f3f01c33d52a5df3e26dd88287
> >>>>>> prerequisite-patch-id: 7edbeaace3549332e581bee3183a76b0e4d18163
> >>>>>>
> >>>>>> Best regards,
> >>>>>> --
> >>>>>> Jessica Zhang <quic_jesszhan@xxxxxxxxxxx>
> >>>>>>
> >>>>>
> >>>>> --
> >>>>> With best wishes
> >>>>> Dmitry
> >>>>
> >>>
> >>>
> >>
> >
>


-- 
With best wishes
Dmitry