[PATCH] drm/msm: dpu: Don't enable vblank interrupts in irq

Discussion:

[PATCH] drm/msm: dpu: Don't enable vblank interrupts in irq_control

Sean Paul

2018-12-07 17:16:57 UTC

From: Sean Paul <***@chromium.org>

The irq_control function is called upon encoder enable/disable and turns
on/off the vblank interrupts. Unfortunately, it enables them when the
drm code is not expecting them to be on. As a result, we can get into
nasty locking situations.

vblank interrupts should be solely managed by the drm core via
control_vblank_irq(). This patch removes the vblank irq handling from
irq_control.

Here's some tracing before this patch:
-- CRTC is enabled, irq_control is called, drm_vblank reference is 0
25.078455: dpu_crtc_enable: id:46 enable:true state{enabled:false}
-- vblank irqs incorrectly firing, drm_crtc_handle_vblank is called from them
25.290067: dpu_crtc_vblank_cb: id=46
25.306691: dpu_crtc_vblank_cb: id=46
-- drm core enables vblank interrupts, drm_vblank_reference > 0
25.311527: dpu_crtc_vblank: id:46 enable:true state{enabled:true}
-- vblank irqs continue (correctly)
25.323351: dpu_crtc_vblank_cb: id=46
25.340033: dpu_crtc_vblank_cb: id=46
25.356701: dpu_crtc_vblank_cb: id=46

And after this patch:
-- CRTC is enabled, irq_control does not enable vblank interrupts
123.861353: dpu_crtc_enable: id:46 enable:true state{enabled:false}
-- drm core enables vblank, we enable the irq
124.192916: dpu_crtc_vblank: id:46 enable:true state{enabled:true}
-- vblank irqs start at the correct time
124.200548: dpu_crtc_vblank_cb: id=46
124.217195: dpu_crtc_vblank_cb: id=46
124.233848: dpu_crtc_vblank_cb: id=46

Signed-off-by: Sean Paul <***@chromium.org>
---
drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_cmd.c | 2 --
drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_vid.c | 6 ------
2 files changed, 8 deletions(-)

diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_cmd.c b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_cmd.c
index 99ab5ca9bed3..c9eaa3516e9a 100644
--- a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_cmd.c
+++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_cmd.c
@@ -339,7 +339,6 @@ static void dpu_encoder_phys_cmd_irq_control(struct dpu_encoder_phys *phys_enc,
if (enable) {
dpu_encoder_helper_register_irq(phys_enc, INTR_IDX_PINGPONG);
dpu_encoder_helper_register_irq(phys_enc, INTR_IDX_UNDERRUN);
- dpu_encoder_phys_cmd_control_vblank_irq(phys_enc, true);

if (dpu_encoder_phys_cmd_is_master(phys_enc))
dpu_encoder_helper_register_irq(phys_enc,
@@ -350,7 +349,6 @@ static void dpu_encoder_phys_cmd_irq_control(struct dpu_encoder_phys *phys_enc,
INTR_IDX_CTL_START);

dpu_encoder_helper_unregister_irq(phys_enc, INTR_IDX_UNDERRUN);
- dpu_encoder_phys_cmd_control_vblank_irq(phys_enc, false);
dpu_encoder_helper_unregister_irq(phys_enc, INTR_IDX_PINGPONG);
}
}
diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_vid.c b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_vid.c
index acdab5b0db18..e3125a1ab3dc 100644
--- a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_vid.c
+++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_vid.c
@@ -706,7 +706,6 @@ static void dpu_encoder_phys_vid_irq_control(struct dpu_encoder_phys *phys_enc,
bool enable)
{
struct dpu_encoder_phys_vid *vid_enc;
- int ret;

if (!phys_enc)
return;
@@ -719,13 +718,8 @@ static void dpu_encoder_phys_vid_irq_control(struct dpu_encoder_phys *phys_enc,
atomic_read(&phys_enc->vblank_refcount));

if (enable) {
- ret = dpu_encoder_phys_vid_control_vblank_irq(phys_enc, true);
- if (ret)
- return;
-
dpu_encoder_helper_register_irq(phys_enc, INTR_IDX_UNDERRUN);
} else {
- dpu_encoder_phys_vid_control_vblank_irq(phys_enc, false);
dpu_encoder_helper_unregister_irq(phys_enc, INTR_IDX_UNDERRUN);
}
}

--
Sean Paul, Software Engineer, Google / Chromium OS

Abhinav Kumar

2018-12-07 21:16:54 UTC

Permalink

Post by Sean Paul
The irq_control function is called upon encoder enable/disable and turns
on/off the vblank interrupts. Unfortunately, it enables them when the
drm code is not expecting them to be on. As a result, we can get into
nasty locking situations.
vblank interrupts should be solely managed by the drm core via
control_vblank_irq(). This patch removes the vblank irq handling from
irq_control.
-- CRTC is enabled, irq_control is called, drm_vblank reference is 0
25.078455: dpu_crtc_enable: id:46 enable:true state{enabled:false}
-- vblank irqs incorrectly firing, drm_crtc_handle_vblank is called from them
25.290067: dpu_crtc_vblank_cb: id=46
25.306691: dpu_crtc_vblank_cb: id=46
-- drm core enables vblank interrupts, drm_vblank_reference > 0
25.311527: dpu_crtc_vblank: id:46 enable:true state{enabled:true}
-- vblank irqs continue (correctly)
25.323351: dpu_crtc_vblank_cb: id=46
25.340033: dpu_crtc_vblank_cb: id=46
25.356701: dpu_crtc_vblank_cb: id=46
-- CRTC is enabled, irq_control does not enable vblank interrupts
123.861353: dpu_crtc_enable: id:46 enable:true state{enabled:false}
-- drm core enables vblank, we enable the irq
124.192916: dpu_crtc_vblank: id:46 enable:true state{enabled:true}
-- vblank irqs start at the correct time
124.200548: dpu_crtc_vblank_cb: id=46
124.217195: dpu_crtc_vblank_cb: id=46
124.233848: dpu_crtc_vblank_cb: id=46

With this change "FROMLIST: drm/msm: Grab a vblank reference when
waiting for commit_done"
in place, this patch looks fine to me as we are protected to disable
vblank
right after commit hence cases such as idle power collapse for smart
panels ideally
shouldnt get impacted.

However, since we are not validating smart panels do we want to remove
changes to dpu_encoder_phys_cmd.c and just keep it for video mode?

Post by Sean Paul
---
drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_cmd.c | 2 --
drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_vid.c | 6 ------
2 files changed, 8 deletions(-)
diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_cmd.c
b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_cmd.c
index 99ab5ca9bed3..c9eaa3516e9a 100644
--- a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_cmd.c
+++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_cmd.c
@@ -339,7 +339,6 @@ static void
dpu_encoder_phys_cmd_irq_control(struct dpu_encoder_phys *phys_enc,
if (enable) {
dpu_encoder_helper_register_irq(phys_enc, INTR_IDX_PINGPONG);
dpu_encoder_helper_register_irq(phys_enc, INTR_IDX_UNDERRUN);
- dpu_encoder_phys_cmd_control_vblank_irq(phys_enc, true);
if (dpu_encoder_phys_cmd_is_master(phys_enc))
dpu_encoder_helper_register_irq(phys_enc,
@@ -350,7 +349,6 @@ static void
dpu_encoder_phys_cmd_irq_control(struct dpu_encoder_phys *phys_enc,
INTR_IDX_CTL_START);
dpu_encoder_helper_unregister_irq(phys_enc, INTR_IDX_UNDERRUN);
- dpu_encoder_phys_cmd_control_vblank_irq(phys_enc, false);
dpu_encoder_helper_unregister_irq(phys_enc, INTR_IDX_PINGPONG);
}
}
diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_vid.c
b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_vid.c
index acdab5b0db18..e3125a1ab3dc 100644
--- a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_vid.c
+++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_vid.c
@@ -706,7 +706,6 @@ static void
dpu_encoder_phys_vid_irq_control(struct dpu_encoder_phys *phys_enc,
bool enable)
{
struct dpu_encoder_phys_vid *vid_enc;
- int ret;
if (!phys_enc)
return;
@@ -719,13 +718,8 @@ static void
dpu_encoder_phys_vid_irq_control(struct dpu_encoder_phys *phys_enc,
atomic_read(&phys_enc->vblank_refcount));
if (enable) {
- ret = dpu_encoder_phys_vid_control_vblank_irq(phys_enc, true);
- if (ret)
- return;
-
dpu_encoder_helper_register_irq(phys_enc, INTR_IDX_UNDERRUN);
} else {
- dpu_encoder_phys_vid_control_vblank_irq(phys_enc, false);
dpu_encoder_helper_unregister_irq(phys_enc, INTR_IDX_UNDERRUN);
}
}

Sean Paul

2018-12-07 21:50:08 UTC

Permalink

With this change "FROMLIST: drm/msm: Grab a vblank reference when waiting
for commit_done"
in place, this patch looks fine to me as we are protected to disable vblank
right after commit hence cases such as idle power collapse for smart panels
ideally
shouldnt get impacted.
However, since we are not validating smart panels do we want to remove
changes to dpu_encoder_phys_cmd.c and just keep it for video mode?

We really don't want the vblank handler running when it's not supposed
to. So having the irq always enabled may be covering up another bug in
cmd mode, but I'd much rather fix this now and deal with regressions
if they surface than leave a known issue in the code.

Sean

--
Sean Paul, Software Engineer, Google / Chromium OS

Abhinav Kumar

2018-12-07 21:58:33 UTC

Permalink

Post by Sean Paul

With this change "FROMLIST: drm/msm: Grab a vblank reference when waiting
for commit_done"
in place, this patch looks fine to me as we are protected to disable vblank
right after commit hence cases such as idle power collapse for smart panels
ideally
shouldnt get impacted.
However, since we are not validating smart panels do we want to remove
changes to dpu_encoder_phys_cmd.c and just keep it for video mode?

Alright, in that case no further concerns from my side

Post by Sean Paul