2018-07-03 16:45:05

by Rob Clark

[permalink] [raw]
Subject: [PATCH] drm/msm/mdp5: fix missing CTL flush

f9cb8d8d836e fixed various race conditions with CTL flush, in particular
flushing and sending the START signal before encoder state was updated.
But it did this a little too well in some cases that don't trigger
encoder->enable(), and CTL[n].FLUSH would never be set. When page flips
happen it would paper over the bug, since the first plag flip would
flush out the state to the hardware.

The issue could be reproduced with, for example, modetest (without the
'-v' argument).

Fixes: f9cb8d8d836e drm/msm/mdp5: rework CTL START signal handling
Signed-off-by: Rob Clark <[email protected]>
---
drivers/gpu/drm/msm/disp/mdp5/mdp5_encoder.c | 12 +++++++++++-
1 file changed, 11 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/msm/disp/mdp5/mdp5_encoder.c b/drivers/gpu/drm/msm/disp/mdp5/mdp5_encoder.c
index 9af94e35f678..fcd44d1d1068 100644
--- a/drivers/gpu/drm/msm/disp/mdp5/mdp5_encoder.c
+++ b/drivers/gpu/drm/msm/disp/mdp5/mdp5_encoder.c
@@ -319,7 +319,17 @@ static int mdp5_encoder_atomic_check(struct drm_encoder *encoder,

mdp5_cstate->ctl = ctl;
mdp5_cstate->pipeline.intf = intf;
- mdp5_cstate->defer_start = true;
+
+ /*
+ * This is a bit awkward, but we want to flush the CTL and hit the
+ * START bit at most once for an atomic update. In the non-full-
+ * modeset case, this is done from crtc->atomic_flush(), but that
+ * is too early in the case of full modeset, in which case we
+ * defer to encoder->enable(). But we need to *know* whether
+ * encoder->enable() will be called to do this:
+ */
+ if (drm_atomic_crtc_needs_modeset(crtc_state))
+ mdp5_cstate->defer_start = true;

return 0;
}
--
2.17.1



2018-07-03 17:02:03

by Sean Paul

[permalink] [raw]
Subject: Re: [PATCH] drm/msm/mdp5: fix missing CTL flush

On Tue, Jul 03, 2018 at 12:43:50PM -0400, Rob Clark wrote:
> f9cb8d8d836e fixed various race conditions with CTL flush, in particular
> flushing and sending the START signal before encoder state was updated.
> But it did this a little too well in some cases that don't trigger
> encoder->enable(), and CTL[n].FLUSH would never be set. When page flips
> happen it would paper over the bug, since the first plag flip would
> flush out the state to the hardware.
>
> The issue could be reproduced with, for example, modetest (without the
> '-v' argument).
>
> Fixes: f9cb8d8d836e drm/msm/mdp5: rework CTL START signal handling
> Signed-off-by: Rob Clark <[email protected]>

Reviewed-by: Sean Paul <[email protected]>

> ---
> drivers/gpu/drm/msm/disp/mdp5/mdp5_encoder.c | 12 +++++++++++-
> 1 file changed, 11 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/msm/disp/mdp5/mdp5_encoder.c b/drivers/gpu/drm/msm/disp/mdp5/mdp5_encoder.c
> index 9af94e35f678..fcd44d1d1068 100644
> --- a/drivers/gpu/drm/msm/disp/mdp5/mdp5_encoder.c
> +++ b/drivers/gpu/drm/msm/disp/mdp5/mdp5_encoder.c
> @@ -319,7 +319,17 @@ static int mdp5_encoder_atomic_check(struct drm_encoder *encoder,
>
> mdp5_cstate->ctl = ctl;
> mdp5_cstate->pipeline.intf = intf;
> - mdp5_cstate->defer_start = true;
> +
> + /*
> + * This is a bit awkward, but we want to flush the CTL and hit the
> + * START bit at most once for an atomic update. In the non-full-
> + * modeset case, this is done from crtc->atomic_flush(), but that
> + * is too early in the case of full modeset, in which case we
> + * defer to encoder->enable(). But we need to *know* whether
> + * encoder->enable() will be called to do this:
> + */
> + if (drm_atomic_crtc_needs_modeset(crtc_state))
> + mdp5_cstate->defer_start = true;
>
> return 0;
> }
> --
> 2.17.1
>

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