Re: [PATCH v8 01/15] drm/msm/disp/dpu: clear dpu_assign_crtc and get crtc from connector state instead of dpu_enc

From: Dmitry Baryshkov
Date: Thu Oct 27 2022 - 13:40:51 EST


On 27/10/2022 16:34, Vinod Polimera wrote:
-----Original Message-----
From: Dmitry Baryshkov <dmitry.baryshkov@xxxxxxxxxx>
Sent: Monday, October 24, 2022 8:52 PM
To: Vinod Polimera (QUIC) <quic_vpolimer@xxxxxxxxxxx>; dri-
devel@xxxxxxxxxxxxxxxxxxxxx; linux-arm-msm@xxxxxxxxxxxxxxx;
freedreno@xxxxxxxxxxxxxxxxxxxxx; devicetree@xxxxxxxxxxxxxxx
Cc: linux-kernel@xxxxxxxxxxxxxxx; robdclark@xxxxxxxxx;
dianders@xxxxxxxxxxxx; swboyd@xxxxxxxxxxxx; Kalyan Thota (QUIC)
<quic_kalyant@xxxxxxxxxxx>; Kuogee Hsieh (QUIC)
<quic_khsieh@xxxxxxxxxxx>; Vishnuvardhan Prodduturi (QUIC)
<quic_vproddut@xxxxxxxxxxx>; Bjorn Andersson (QUIC)
<quic_bjorande@xxxxxxxxxxx>; Aravind Venkateswaran (QUIC)
<quic_aravindh@xxxxxxxxxxx>; Abhinav Kumar (QUIC)
<quic_abhinavk@xxxxxxxxxxx>; Sankeerth Billakanti (QUIC)
<quic_sbillaka@xxxxxxxxxxx>
Subject: Re: [PATCH v8 01/15] drm/msm/disp/dpu: clear dpu_assign_crtc and
get crtc from connector state instead of dpu_enc

WARNING: This email originated from outside of Qualcomm. Please be wary
of any links or attachments, and do not enable macros.

On 12/10/2022 15:02, Vinod Polimera wrote:
Update crtc retrieval from dpu_enc to dpu_enc connector state,
since new links get set as part of the dpu enc virt mode set.
The dpu_enc->crtc cache is no more needed, hence cleaning it as
part of this change.

Signed-off-by: Vinod Polimera <quic_vpolimer@xxxxxxxxxxx>
---
drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c | 4 ---
drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c | 42 +++++++++---------
-----------
drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.h | 8 ------
3 files changed, 13 insertions(+), 41 deletions(-)

diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c
b/drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c
index 13ce321..8ec9a13 100644
--- a/drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c
+++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_crtc.c
@@ -1029,7 +1029,6 @@ static void dpu_crtc_disable(struct drm_crtc *crtc,
*/
if (dpu_encoder_get_intf_mode(encoder) == INTF_MODE_VIDEO)
release_bandwidth = true;
- dpu_encoder_assign_crtc(encoder, NULL);
}

/* wait for frame_event_done completion */
@@ -1099,9 +1098,6 @@ static void dpu_crtc_enable(struct drm_crtc *crtc,
trace_dpu_crtc_enable(DRMID(crtc), true, dpu_crtc);
dpu_crtc->enabled = true;

- drm_for_each_encoder_mask(encoder, crtc->dev, crtc->state-
encoder_mask)
- dpu_encoder_assign_crtc(encoder, crtc);
-
/* Enable/restore vblank irq handling */
drm_crtc_vblank_on(crtc);
}
diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c
b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c
index 9c6817b..d05b353 100644
--- a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c
+++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder.c
@@ -132,11 +132,6 @@ enum dpu_enc_rc_states {
* @intfs_swapped: Whether or not the phys_enc interfaces have been
swapped
* for partial update right-only cases, such as pingpong
* split where virtual pingpong does not generate IRQs
- * @crtc: Pointer to the currently assigned crtc. Normally you
- * would use crtc->state->encoder_mask to determine the
- * link between encoder/crtc. However in this case we need
- * to track crtc in the disable() hook which is called
- * _after_ encoder_mask is cleared.
* @connector: If a mode is set, cached pointer to the active
connector
* @crtc_kickoff_cb: Callback into CRTC that will flush & start
* all CTL paths
@@ -181,7 +176,6 @@ struct dpu_encoder_virt {

bool intfs_swapped;

- struct drm_crtc *crtc;
struct drm_connector *connector;

struct dentry *debugfs_root;
@@ -1288,7 +1282,7 @@ static void dpu_encoder_vblank_callback(struct
drm_encoder *drm_enc,
struct dpu_encoder_phys *phy_enc)
{
struct dpu_encoder_virt *dpu_enc = NULL;
- unsigned long lock_flags;
+ struct drm_crtc *crtc;

if (!drm_enc || !phy_enc)
return;
@@ -1296,12 +1290,13 @@ static void dpu_encoder_vblank_callback(struct
drm_encoder *drm_enc,
DPU_ATRACE_BEGIN("encoder_vblank_callback");
dpu_enc = to_dpu_encoder_virt(drm_enc);

- atomic_inc(&phy_enc->vsync_cnt);
+ if (!dpu_enc->connector || !dpu_enc->connector->state ||
+ !dpu_enc->connector->state->crtc)
+ return;

- spin_lock_irqsave(&dpu_enc->enc_spinlock, lock_flags);
- if (dpu_enc->crtc)
- dpu_crtc_vblank_callback(dpu_enc->crtc);
- spin_unlock_irqrestore(&dpu_enc->enc_spinlock, lock_flags);
+ atomic_inc(&phy_enc->vsync_cnt);
+ crtc = dpu_enc->connector->state->crtc;
+ dpu_crtc_vblank_callback(crtc);

So, what if the user commits the mode setting change on another CPU,
while we are handling the vblank callback here? Can this happen?

If user issues a commit on another CPU, it will wait in the drm_atomic_helper_swap_state
as drm_atomic_helper_commit_hw_done which does the complete_all(&commit->hw_done)
for the current commit didn't finish yet.

Yes. But there is no interlock between commit->hw_done and vblank IRQ processing, isn't it? This call happens when DPU processes the vblank IRQ, so nobody stops other core from swapping the encode state on another core.


DPU_ATRACE_END("encoder_vblank_callback");
}
--
With best wishes
Dmitry

Thanks,
Vinod P.

--
With best wishes
Dmitry