Re: [PATCH v2 6/6] drm/i915: Add a cursor hack to allow converting legacy page flip to atomic, v3.

Op 13-12-16 om 14:01 schreef Archit Taneja:

Hi,

On 12/12/2016 4:04 PM, Maarten Lankhorst wrote:

...

Do something similar to vc4, only allow updating the cursor state in-place through a fastpath when the watermarks are unaffected. This will allow cursor movement to be smooth, but changing cursor size or showing/hiding cursor will still fall back so watermarks can be updated.

Only moving and changing fb is allowed.

I was implementing something similar for msm kms. I have a comment/query below.

...

Changes since v1:

• Set page flip to always_unused for trybot.
• Copy fence correctly, ignore plane_state->state, should be NULL.
• Check crtc_state for !active and modeset, go to slowpath if the case.

Changes since v2:

• Make error handling work correctly. (Matthew Auld)

Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com

drivers/gpu/drm/i915/intel_atomic_plane.c | 47 +++++++---- drivers/gpu/drm/i915/intel_display.c | 132 +++++++++++++++++++++++++++++- drivers/gpu/drm/i915/intel_drv.h | 2 + 3 files changed, 163 insertions(+), 18 deletions(-)

diff --git a/drivers/gpu/drm/i915/intel_atomic_plane.c b/drivers/gpu/drm/i915/intel_atomic_plane.c index dbe9fb41ae53..60d75ce8a989 100644 --- a/drivers/gpu/drm/i915/intel_atomic_plane.c +++ b/drivers/gpu/drm/i915/intel_atomic_plane.c @@ -103,36 +103,24 @@ intel_plane_destroy_state(struct drm_plane *plane, drm_atomic_helper_plane_destroy_state(plane, state); }

-static int intel_plane_atomic_check(struct drm_plane *plane,

•                struct drm_plane_state *state)
  

+int intel_plane_atomic_check_with_state(struct intel_crtc_state *crtc_state,

•                struct intel_plane_state *intel_state)
  

{

• struct drm_plane *plane = intel_state->base.plane; struct drm_i915_private *dev_priv = to_i915(plane->dev);

• struct drm_crtc *crtc = state->crtc;
• struct intel_crtc *intel_crtc;
• struct intel_crtc_state *crtc_state;

• struct drm_plane_state *state = &intel_state->base; struct intel_plane *intel_plane = to_intel_plane(plane);

• struct intel_plane_state *intel_state = to_intel_plane_state(state);

• struct drm_crtc_state *drm_crtc_state; int ret;

• crtc = crtc ? crtc : plane->state->crtc;

• intel_crtc = to_intel_crtc(crtc);

• /* * Both crtc and plane->crtc could be NULL if we're updating a * property while the plane is disabled. We don't actually have * anything driver-specific we need to test in that case, so * just return success. */

• if (!crtc)

• if (!intel_state->base.crtc && !plane->state->crtc) return 0;

• drm_crtc_state = drm_atomic_get_existing_crtc_state(state->state, crtc);
• if (WARN_ON(!drm_crtc_state))

•    return -EINVAL;
  

• crtc_state = to_intel_crtc_state(drm_crtc_state);
• /* Clip all planes to CRTC size, or 0x0 if CRTC is disabled */ intel_state->clip.x1 = 0; intel_state->clip.y1 = 0;

@@ -184,6 +172,31 @@ static int intel_plane_atomic_check(struct drm_plane *plane, return intel_plane_atomic_calc_changes(&crtc_state->base, state); }

+static int intel_plane_atomic_check(struct drm_plane *plane,

•                struct drm_plane_state *state)
  

+{

• struct drm_crtc *crtc = state->crtc;
• struct drm_crtc_state *drm_crtc_state;
• crtc = crtc ? crtc : plane->state->crtc;
• /*

• * Both crtc and plane->crtc could be NULL if we're updating a
  

• * property while the plane is disabled.  We don't actually have
  

• * anything driver-specific we need to test in that case, so
  

• * just return success.
  

• */
  

• if (!crtc)

•    return 0;
  

• drm_crtc_state = drm_atomic_get_existing_crtc_state(state->state, crtc);
• if (WARN_ON(!drm_crtc_state))

•    return -EINVAL;
  

• return intel_plane_atomic_check_with_state(to_intel_crtc_state(drm_crtc_state),

•                       to_intel_plane_state(state));
  

+}

static void intel_plane_atomic_update(struct drm_plane *plane, struct drm_plane_state *old_state) { diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c index 9eaf1e5bdae9..5568ecac2edc 100644 --- a/drivers/gpu/drm/i915/intel_display.c +++ b/drivers/gpu/drm/i915/intel_display.c @@ -15028,6 +15028,136 @@ const struct drm_plane_funcs intel_plane_funcs = { .atomic_destroy_state = intel_plane_destroy_state, };

+static int +intel_legacy_cursor_update(struct drm_plane *plane,

•           struct drm_crtc *crtc,
  

•           struct drm_framebuffer *fb,
  

•           int crtc_x, int crtc_y,
  

•           unsigned int crtc_w, unsigned int crtc_h,
  

•           uint32_t src_x, uint32_t src_y,
  

•           uint32_t src_w, uint32_t src_h)
  

+{

• struct drm_i915_private *dev_priv = to_i915(crtc->dev);
• int ret;
• struct drm_plane_state *old_plane_state, *new_plane_state;
• struct intel_plane *intel_plane = to_intel_plane(plane);
• struct drm_framebuffer *old_fb;
• struct drm_crtc_state *crtc_state = crtc->state;
• /*

• * When crtc is inactive or there is a modeset pending,
  

• * wait for it to complete in the slowpath
  

• */
  

• if (!crtc_state->active || needs_modeset(crtc_state) ||

•    to_intel_crtc_state(crtc_state)->update_pipe)
  

•    goto slow;
  

• old_plane_state = plane->state;
• /*

• * If any parameters change that may affect watermarks,
  

• * take the slowpath. Only changing fb or position should be
  

• * in the fastpath.
  

• */
  

• if (old_plane_state->crtc != crtc ||

•    old_plane_state->src_w != src_w ||
  

•    old_plane_state->src_h != src_h ||
  

•    old_plane_state->crtc_w != crtc_w ||
  

•    old_plane_state->crtc_h != crtc_h ||
  

•    !old_plane_state->visible ||
  

•    old_plane_state->fb->modifier != fb->modifier)
  

•    goto slow;
  

• new_plane_state = intel_plane_duplicate_state(plane);
• if (!new_plane_state)

•    return -ENOMEM;
  

About creating a new plane state here and swapping it later, I guess it's more for the sake of intel_plane_atomic_check_with_state() to go through cleanly, right? Rather than directly modifying the old plane state?

i915 keeps some derived state in intel_plane_state, we copy the full state to make sure the full derived state is updated as well.

I didn't see anything equivalent in vc4's implementation of update_plane, hence was wondering why exactly it is needed.

Also, about updating fb in the fast path, my assumption was that it was only crtc_x/y was the thing most likely to be changed really fast. Have you seen use cases where the fb changes at a faster than vsync rate?

It didn't cost much extra effort to implement fb changing too since that is handled in intel's update_plane callback too. This is why I did that as well. I'm not sure it's needed in general. Any changed fb still has to have the same size or we go to slowpath.

~Maarten