[Intel-gfx] [PATCH] drm/atomic: Try to preserve the crtc enabled state in drm_atomic_remove_fb, v2.

[Maarten Lankhorst]

Op 01-11-17 om 18:00 schreef Ville Syrjälä:
> On Wed, Nov 01, 2017 at 04:55:06PM +0100, Maarten Lankhorst wrote:
>> Op 01-11-17 om 16:29 schreef Ville Syrjälä:
>>> On Wed, Nov 01, 2017 at 04:04:33PM +0100, Maarten Lankhorst wrote:
>>>> This introduces a slight behavioral change to rmfb. Instead of
>>>> disabling a crtc when the primary plane is disabled, we try to
>>>> preserve it.
>>>>
>>>> Apart from old versions of the vmwgfx xorg driver, there is
>>>> nothing depending on rmfb disabling a crtc.
>>>>
>>>> Vmwgfx' and simple kms helper atomic implementation rejects CRTC
>>>> enabled without plane, so we can do this safely.
> The code for those seems a bit inconsistent. The crtc check requires
> that the crtc state and plane state match. But the plane check allows
> the plane to be enabled w/o the crtc being enabled. I guess it doesn't
> matter really since you can't enable the plane without a crtc, and the
> crtc check would then catch the case where the crtc would be disabled.

Exactly. :-) Only when a plane is unbound and stays unbound, the crtc check won't
be invoked. Hence it's the most accurate way of making sure that
crtc enabled <=> primary plane bound.

If the check was done in the primary plane, an atomic modeset could enable
the crtc without enabling the primary plane, which shouldn't be allowed but
a plane check won't catch it.

This has been a bug in simple-kms-helper, fixed in the below commit:

commit 765831dc27ab141b3a0be1ab55b922b012427902
Author: Maarten Lankhorst <maarten.lankhorst at linux.intel.com>
Date:   Wed Jul 12 10:13:29 2017 +0200

    drm/simple-kms-helper: Fix the check for the mismatch between plane and CRTC enabled.

> Oh and looks like drm_plane_helper_check_state() is a bit buggy. It
> still uses crtc->enabled instead of crtc_state->enable to check the
> state of the crtc. I guess to keep drm_plane_helper_check_update()
> working we may have to pass in the crtc state manually.
>
> The vmwgfx plane check looks a bit bogus in other ways too. I guess
> I'll have to fire off a couple of patches.
>
>>>> If the atomic commit is rejected by the driver then we will still
>>>> fall back to the old behavior and turn off the crtc.
>>>>
>>>> Changes since v1:
>>>> - Restart completely when rmfb with crtc on fails (Sean Paul).
>>>>
>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst at linux.intel.com>
>>>> Cc: Sean Paul <seanpaul at chromium.org>
>>>> Cc: Daniel Vetter <daniel.vetter at ffwll.ch>
>>>> ---
>>>>  drivers/gpu/drm/drm_framebuffer.c | 23 +++++++++++++++++------
>>>>  1 file changed, 17 insertions(+), 6 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/drm_framebuffer.c b/drivers/gpu/drm/drm_framebuffer.c
>>>> index 2affe53f3fda..f0679468f421 100644
>>>> --- a/drivers/gpu/drm/drm_framebuffer.c
>>>> +++ b/drivers/gpu/drm/drm_framebuffer.c
>>>> @@ -765,14 +765,18 @@ static int atomic_remove_fb(struct drm_framebuffer *fb)
>>>>  	struct drm_plane *plane;
>>>>  	struct drm_connector *conn;
>>>>  	struct drm_connector_state *conn_state;
>>>> -	int i, ret = 0;
>>>> +	int i, ret;
>>>>  	unsigned plane_mask;
>>>> +	bool disable_crtcs = false;
>>>>  
>>>> -	state = drm_atomic_state_alloc(dev);
>>>> -	if (!state)
>>>> -		return -ENOMEM;
>>>> -
>>>> +retry_disable:
>>>>  	drm_modeset_acquire_init(&ctx, 0);
>>>> +
>>>> +	state = drm_atomic_state_alloc(dev);
>>>> +	if (!state) {
>>>> +		ret = -ENOMEM;
>>>> +		goto out;
>>>> +	}
>>>>  	state->acquire_ctx = &ctx;
>>>>  
>>>>  retry:
>>>> @@ -793,7 +797,7 @@ static int atomic_remove_fb(struct drm_framebuffer *fb)
>>>>  			goto unlock;
>>>>  		}
>>>>  
>>>> -		if (plane_state->crtc->primary == plane) {
>>>> +		if (disable_crtcs && plane_state->crtc->primary == plane) {
>>>>  			struct drm_crtc_state *crtc_state;
>>>>  
>>>>  			crtc_state = drm_atomic_get_existing_crtc_state(state, plane_state->crtc);
>>>> @@ -818,6 +822,7 @@ static int atomic_remove_fb(struct drm_framebuffer *fb)
>>>>  		plane->old_fb = plane->fb;
>>>>  	}
>>>>  
>>>> +	/* This list is only filled when disable_crtcs is set. */
>>>>  	for_each_new_connector_in_state(state, conn, conn_state, i) {
>>> WARN_ON(!disable_crtcs) maybe?
>> Would be overkill, nothing before it adds connector state, and if atomic check does then that's fine, but it won't be run here. :)
> It would serve as a way to document that fact, even without the comment.
> But I won't insist on it.
>
>>>>  		ret = drm_atomic_set_crtc_for_connector(conn_state, NULL);
>>>>  
>>>> @@ -840,9 +845,15 @@ static int atomic_remove_fb(struct drm_framebuffer *fb)
>>>>  
>>>>  	drm_atomic_state_put(state);
>>>>  
>>>> +out:
>>>>  	drm_modeset_drop_locks(&ctx);
>>>>  	drm_modeset_acquire_fini(&ctx);
>>>>  
>>>> +	if (ret == -EINVAL && !disable_crtcs) {
>>> Hmm. -EINVAL seems rather specific. Not sure if we could just check for
>>> any error?
>>>
>>> Or... I'm not sure if we have any central place where we do the
>>> "can I disable the primary plane w/o disabling the crtc?" check. If we
>>> do then we could also add a comment there informing people that the
>>> -EINVAL is important.
>> We don't have a central place, I check for EINVAL since that is the generic atomic_check() failed error. If it fails for any other reason then we don't have to retry, but pass it along. :)
> Oh well. I guess people just have to be careful with their error
> values. I suppoe anyone depending on the retry will notice this
> issue rather quickly.
>