[Freedreno] [PATCH 4/5] drm/msm/mdp5: dynamically assign hw pipes to planes

Archit Taneja architt at codeaurora.org
Mon Nov 7 10:29:56 UTC 2016


Hi,

Minor comments below. LGTM otherwise.

On 11/05/2016 09:56 PM, Rob Clark wrote:
> (re)assign the hw pipes to planes based on required caps, and to handle
> situations where we could not modify an in-use plane (ie. SMP block
> reallocation).
>
> This means all planes advertise the superset of formats and properties.
> Userspace must (as always) use atomic TEST_ONLY step for atomic updates,
> as not all planes may be available for use on every frame.
>
> The mapping of hwpipe to plane is stored in mdp5_state, so that state
> updates are atomically committed in the same way that plane/etc state
> updates are managed.  This is needed because the mdp5_plane_state keeps
> a pointer to the hwpipe, and we don't want global state to become out
> of sync with the plane state if an atomic update fails, we hit deadlock/
> backoff scenario, etc.  The use of state_lock keeps multiple parallel
> updates which both re-assign hwpipes properly serialized.
>
> Signed-off-by: Rob Clark <robdclark at gmail.com>
> ---
>  drivers/gpu/drm/msm/mdp/mdp5/mdp5_crtc.c  |   2 +-
>  drivers/gpu/drm/msm/mdp/mdp5/mdp5_kms.c   |   4 +-
>  drivers/gpu/drm/msm/mdp/mdp5/mdp5_kms.h   |   7 +-
>  drivers/gpu/drm/msm/mdp/mdp5/mdp5_pipe.c  |  70 +++++++++++++
>  drivers/gpu/drm/msm/mdp/mdp5/mdp5_pipe.h  |  10 ++
>  drivers/gpu/drm/msm/mdp/mdp5/mdp5_plane.c | 158 ++++++++++++++++--------------
>  6 files changed, 171 insertions(+), 80 deletions(-)
>
> diff --git a/drivers/gpu/drm/msm/mdp/mdp5/mdp5_crtc.c b/drivers/gpu/drm/msm/mdp/mdp5/mdp5_crtc.c
> index 6213f51..99958f7 100644
> --- a/drivers/gpu/drm/msm/mdp/mdp5/mdp5_crtc.c
> +++ b/drivers/gpu/drm/msm/mdp/mdp5/mdp5_crtc.c
> @@ -407,7 +407,7 @@ static int mdp5_crtc_atomic_check(struct drm_crtc *crtc,
>  	for (i = 0; i < cnt; i++) {
>  		pstates[i].state->stage = STAGE_BASE + i + base;
>  		DBG("%s: assign pipe %s on stage=%d", crtc->name,
> -				pipe2name(mdp5_plane_pipe(pstates[i].plane)),
> +				pstates[i].plane->name,
>  				pstates[i].state->stage);
>  	}
>
> diff --git a/drivers/gpu/drm/msm/mdp/mdp5/mdp5_kms.c b/drivers/gpu/drm/msm/mdp/mdp5/mdp5_kms.c
> index ca6dfeb..3542adf 100644
> --- a/drivers/gpu/drm/msm/mdp/mdp5/mdp5_kms.c
> +++ b/drivers/gpu/drm/msm/mdp/mdp5/mdp5_kms.c
> @@ -92,7 +92,7 @@ struct mdp5_state *mdp5_get_state(struct drm_atomic_state *s)
>  		return ERR_PTR(-ENOMEM);
>
>  	/* Copy state: */
> -	/* TODO */
> +	new_state->hwpipe = mdp5_kms->state->hwpipe;
>
>  	state->state = new_state;
>
> @@ -377,7 +377,7 @@ static int modeset_init(struct mdp5_kms *mdp5_kms)
>  		struct drm_plane *plane;
>  		struct drm_crtc *crtc;
>
> -		plane = mdp5_plane_init(dev, mdp5_kms->hwpipes[i], primary);
> +		plane = mdp5_plane_init(dev, primary);
>  		if (IS_ERR(plane)) {
>  			ret = PTR_ERR(plane);
>  			dev_err(dev->dev, "failed to construct plane %d (%d)\n", i, ret);
> diff --git a/drivers/gpu/drm/msm/mdp/mdp5/mdp5_kms.h b/drivers/gpu/drm/msm/mdp/mdp5/mdp5_kms.h
> index 52914ec..4475090 100644
> --- a/drivers/gpu/drm/msm/mdp/mdp5/mdp5_kms.h
> +++ b/drivers/gpu/drm/msm/mdp/mdp5/mdp5_kms.h
> @@ -82,7 +82,7 @@ struct mdp5_kms {
>   * For atomic updates which require modifying global state,
>   */
>  struct mdp5_state {
> -	uint32_t dummy;
> +	struct mdp5_hw_pipe_state hwpipe;

I was wondering if we could rename the above as hwpipes/hwpipe_state/hwpipe_map,
because it gets a confusing since variables of the type mdp5_hw_pipe are
also named hwpipe.

>  };
>
>  struct mdp5_state *__must_check
> @@ -94,6 +94,8 @@ mdp5_get_state(struct drm_atomic_state *s);
>  struct mdp5_plane_state {
>  	struct drm_plane_state base;
>
> +	struct mdp5_hw_pipe *hwpipe;
> +
>  	/* aligned with property */
>  	uint8_t premultiplied;
>  	uint8_t zpos;
> @@ -232,8 +234,7 @@ uint32_t mdp5_plane_get_flush(struct drm_plane *plane);
>  void mdp5_plane_complete_commit(struct drm_plane *plane,
>  	struct drm_plane_state *state);
>  enum mdp5_pipe mdp5_plane_pipe(struct drm_plane *plane);
> -struct drm_plane *mdp5_plane_init(struct drm_device *dev,
> -		struct mdp5_hw_pipe *hwpipe, bool primary);
> +struct drm_plane *mdp5_plane_init(struct drm_device *dev, bool primary);
>
>  uint32_t mdp5_crtc_vblank(struct drm_crtc *crtc);
>
> diff --git a/drivers/gpu/drm/msm/mdp/mdp5/mdp5_pipe.c b/drivers/gpu/drm/msm/mdp/mdp5/mdp5_pipe.c
> index 7f3e8e50..115de7d 100644
> --- a/drivers/gpu/drm/msm/mdp/mdp5/mdp5_pipe.c
> +++ b/drivers/gpu/drm/msm/mdp/mdp5/mdp5_pipe.c
> @@ -17,6 +17,76 @@
>
>  #include "mdp5_kms.h"
>
> +struct mdp5_hw_pipe *mdp5_pipe_assign(struct drm_atomic_state *s,
> +		struct drm_plane *plane, uint32_t caps)
> +{
> +	struct msm_drm_private *priv = s->dev->dev_private;
> +	struct mdp5_kms *mdp5_kms = to_mdp5_kms(to_mdp_kms(priv->kms));
> +	struct mdp5_state *state = mdp5_get_state(s);
> +	struct mdp5_hw_pipe_state *old_state, *new_state;
> +	struct mdp5_hw_pipe *hwpipe = NULL;
> +	int i;
> +

Can we assign "state = mdp5_get_state(s);" here instead for clarity? Since the func
does more than just getting the mdp5_state pointer?

> +	if (IS_ERR(state))
> +		return ERR_CAST(state);
> +
> +	/* grab old_state after mdp5_get_state(), since now we hold lock: */
> +	old_state = &mdp5_kms->state->hwpipe;
> +	new_state = &state->hwpipe;
> +
> +	for (i = 0; i < mdp5_kms->num_hwpipes; i++) {
> +		struct mdp5_hw_pipe *cur = mdp5_kms->hwpipes[i];
> +
> +		/* skip if already in-use.. check both new and old state,
> +		 * since we cannot immediately re-use a pipe that is
> +		 * released in the current update in some cases:
> +		 *  (1) mdp5 has SMP (non-double-buffered)
> +		 *  (2) hw pipe previously assigned to different CRTC
> +		 *      (vblanks might not be aligned)

For 1), we have non-SMP SoCs (a.k.a 8996), and for 2) we would have
the info whether the hwpipe is changing crtcs in old and new states.
I guess we could leave this as an optimization for the future.


> +		 */
> +		if (new_state->hwpipe_to_plane[cur->idx] ||
> +				old_state->hwpipe_to_plane[cur->idx])
> +			continue;
> +
> +		/* skip if doesn't support some required caps: */
> +		if (caps & ~cur->caps)
> +			continue;
> +
> +		/* possible candidate, take the one with the
> +		 * fewest unneeded caps bits set:
> +		 */
> +		if (!hwpipe || (hweight_long(cur->caps & ~caps) <
> +				hweight_long(hwpipe->caps & ~caps)))
> +			hwpipe = cur;
> +	}
> +
> +	if (!hwpipe)
> +		return ERR_PTR(-ENOMEM);
> +
> +	DBG("%s: assign to plane %s for caps %x",
> +			hwpipe->name, plane->name, caps);
> +	new_state->hwpipe_to_plane[hwpipe->idx] = plane;
> +
> +	return hwpipe;
> +}
> +
> +void mdp5_pipe_release(struct drm_atomic_state *s, struct mdp5_hw_pipe *hwpipe)
> +{
> +	struct mdp5_state *state = mdp5_get_state(s);
> +	struct mdp5_hw_pipe_state *new_state = &state->hwpipe;
> +
> +	if (!hwpipe)
> +		return;
> +
> +	if (WARN_ON(!new_state->hwpipe_to_plane[hwpipe->idx]))
> +		return;
> +
> +	DBG("%s: release from plane %s", hwpipe->name,
> +		new_state->hwpipe_to_plane[hwpipe->idx]->name);
> +
> +	new_state->hwpipe_to_plane[hwpipe->idx] = NULL;
> +}
> +
>  void mdp5_pipe_destroy(struct mdp5_hw_pipe *hwpipe)
>  {
>  	kfree(hwpipe);
> diff --git a/drivers/gpu/drm/msm/mdp/mdp5/mdp5_pipe.h b/drivers/gpu/drm/msm/mdp/mdp5/mdp5_pipe.h
> index dcfa0e0..bac5900 100644
> --- a/drivers/gpu/drm/msm/mdp/mdp5/mdp5_pipe.h
> +++ b/drivers/gpu/drm/msm/mdp/mdp5/mdp5_pipe.h
> @@ -32,6 +32,16 @@ struct mdp5_hw_pipe {
>  	uint32_t flush_mask;      /* used to commit pipe registers */
>  };
>
> +/* global atomic state of assignment between pipes and planes: */
> +struct mdp5_hw_pipe_state {
> +	struct drm_plane *hwpipe_to_plane[16];

We could use SSPP_MAX here instead of 16. We would need to move its
definition from mdp5_crtc.c to mdp5_kms.h, though.

> +};
> +
> +struct mdp5_hw_pipe *__must_check
> +mdp5_pipe_assign(struct drm_atomic_state *s, struct drm_plane *plane,
> +		uint32_t caps);
> +void mdp5_pipe_release(struct drm_atomic_state *s, struct mdp5_hw_pipe *hwpipe);
> +
>  struct mdp5_hw_pipe *mdp5_pipe_init(enum mdp5_pipe pipe,
>  		uint32_t reg_offset, uint32_t caps);
>  void mdp5_pipe_destroy(struct mdp5_hw_pipe *hwpipe);
> diff --git a/drivers/gpu/drm/msm/mdp/mdp5/mdp5_plane.c b/drivers/gpu/drm/msm/mdp/mdp5/mdp5_plane.c
> index e4ecfeb..64e97ef 100644
> --- a/drivers/gpu/drm/msm/mdp/mdp5/mdp5_plane.c
> +++ b/drivers/gpu/drm/msm/mdp/mdp5/mdp5_plane.c
> @@ -22,8 +22,6 @@
>  struct mdp5_plane {
>  	struct drm_plane base;
>
> -	struct mdp5_hw_pipe *hwpipe;
> -
>  	uint32_t nformats;
>  	uint32_t formats[32];
>  };
> @@ -63,12 +61,6 @@ static void mdp5_plane_destroy(struct drm_plane *plane)
>  static void mdp5_plane_install_rotation_property(struct drm_device *dev,
>  		struct drm_plane *plane)
>  {
> -	struct mdp5_plane *mdp5_plane = to_mdp5_plane(plane);
> -
> -	if (!(mdp5_plane->hwpipe->caps & MDP_PIPE_CAP_HFLIP) &&
> -		!(mdp5_plane->hwpipe->caps & MDP_PIPE_CAP_VFLIP))
> -		return;
> -
>  	drm_plane_create_rotation_property(plane,
>  					   DRM_ROTATE_0,
>  					   DRM_ROTATE_0 |
> @@ -184,6 +176,8 @@ mdp5_plane_atomic_print_state(struct drm_printer *p,
>  {
>  	struct mdp5_plane_state *pstate = to_mdp5_plane_state(state);
>
> +	drm_printf(p, "\thwpipe=%s\n", pstate->hwpipe ?
> +			pstate->hwpipe->name : "(null)");
>  	drm_printf(p, "\tpremultiplied=%u\n", pstate->premultiplied);
>  	drm_printf(p, "\tzpos=%u\n", pstate->zpos);
>  	drm_printf(p, "\talpha=%u\n", pstate->alpha);
> @@ -237,10 +231,12 @@ mdp5_plane_duplicate_state(struct drm_plane *plane)
>  static void mdp5_plane_destroy_state(struct drm_plane *plane,
>  		struct drm_plane_state *state)
>  {
> +	struct mdp5_plane_state *pstate = to_mdp5_plane_state(state);
> +
>  	if (state->fb)
>  		drm_framebuffer_unreference(state->fb);
>
> -	kfree(to_mdp5_plane_state(state));
> +	kfree(pstate);
>  }
>
>  static const struct drm_plane_funcs mdp5_plane_funcs = {
> @@ -285,70 +281,81 @@ static void mdp5_plane_cleanup_fb(struct drm_plane *plane,
>  static int mdp5_plane_atomic_check(struct drm_plane *plane,
>  		struct drm_plane_state *state)
>  {
> -	struct mdp5_plane *mdp5_plane = to_mdp5_plane(plane);
> +	struct mdp5_plane_state *mdp5_state = to_mdp5_plane_state(state);
>  	struct drm_plane_state *old_state = plane->state;
> -	const struct mdp_format *format;
> -	bool vflip, hflip;
> +	bool new_hwpipe = false;
> +	uint32_t caps = 0;
>
>  	DBG("%s: check (%d -> %d)", plane->name,
>  			plane_enabled(old_state), plane_enabled(state));
>
> +	/* We don't allow faster-than-vblank updates.. if we did add this
> +	 * some day, we would need to disallow in cases where hwpipe
> +	 * changes
> +	 */
> +	if (WARN_ON(to_mdp5_plane_state(old_state)->pending))
> +		return -EBUSY;
> +
>  	if (plane_enabled(state)) {
>  		unsigned int rotation;
> +		const struct mdp_format *format;
>
>  		format = to_mdp_format(msm_framebuffer_format(state->fb));
> -		if (MDP_FORMAT_IS_YUV(format) &&
> -			!pipe_supports_yuv(mdp5_plane->hwpipe->caps)) {
> -			DBG("Pipe doesn't support YUV\n");
> -
> -			return -EINVAL;
> -		}
> -
> -		if (!(mdp5_plane->hwpipe->caps & MDP_PIPE_CAP_SCALE) &&
> -			(((state->src_w >> 16) != state->crtc_w) ||
> -			((state->src_h >> 16) != state->crtc_h))) {
> -			DBG("Pipe doesn't support scaling (%dx%d -> %dx%d)\n",
> -				state->src_w >> 16, state->src_h >> 16,
> -				state->crtc_w, state->crtc_h);
> +		if (MDP_FORMAT_IS_YUV(format))
> +			caps |= MDP_PIPE_CAP_SCALE | MDP_PIPE_CAP_CSC;
>
> -			return -EINVAL;
> -		}
> +		if (((state->src_w >> 16) != state->crtc_w) ||
> +				((state->src_h >> 16) != state->crtc_h))
> +			caps |= MDP_PIPE_CAP_SCALE;
>
>  		rotation = drm_rotation_simplify(state->rotation,
>  						 DRM_ROTATE_0 |
>  						 DRM_REFLECT_X |
>  						 DRM_REFLECT_Y);
>
> -		hflip = !!(rotation & DRM_REFLECT_X);
> -		vflip = !!(rotation & DRM_REFLECT_Y);
> -
> -		if ((vflip && !(mdp5_plane->hwpipe->caps & MDP_PIPE_CAP_VFLIP)) ||
> -			(hflip && !(mdp5_plane->hwpipe->caps & MDP_PIPE_CAP_HFLIP))) {
> -			DBG("Pipe doesn't support flip\n");
> -
> -			return -EINVAL;
> +		if (rotation & DRM_REFLECT_X)
> +			caps |= MDP_PIPE_CAP_HFLIP;
> +
> +		if (rotation & DRM_REFLECT_Y)
> +			caps |= MDP_PIPE_CAP_VFLIP;
> +
> +		/* (re)allocate hw pipe if we don't have one or caps-mismatch: */
> +		if (!mdp5_state->hwpipe || (caps & ~mdp5_state->hwpipe->caps))
> +			new_hwpipe = true;
> +
> +		if (plane_enabled(old_state)) {
> +			bool full_modeset = false;
> +			if (state->fb->pixel_format != old_state->fb->pixel_format) {
> +				DBG("%s: pixel_format change!", plane->name);
> +				full_modeset = true;
> +			}
> +			if (state->src_w != old_state->src_w) {
> +				DBG("%s: src_w change!", plane->name);
> +				full_modeset = true;
> +			}
> +			if (full_modeset) {
> +				/* cannot change SMP block allocation during
> +				 * scanout:
> +				 */
> +				if (get_kms(plane)->smp)
> +					new_hwpipe = true;
> +			}
>  		}
> -	}
>
> -	if (plane_enabled(state) && plane_enabled(old_state)) {
> -		/* we cannot change SMP block configuration during scanout: */
> -		bool full_modeset = false;
> -		if (state->fb->pixel_format != old_state->fb->pixel_format) {
> -			DBG("%s: pixel_format change!", plane->name);
> -			full_modeset = true;
> -		}
> -		if (state->src_w != old_state->src_w) {
> -			DBG("%s: src_w change!", plane->name);
> -			full_modeset = true;
> -		}
> -		if (to_mdp5_plane_state(old_state)->pending) {
> -			DBG("%s: still pending!", plane->name);
> -			full_modeset = true;
> -		}
> -		if (full_modeset) {
> -			struct drm_crtc_state *crtc_state =
> -					drm_atomic_get_crtc_state(state->state, state->crtc);
> -			crtc_state->mode_changed = true;
> +		/* (re)assign hwpipe if needed, otherwise keep old one: */
> +		if (new_hwpipe) {
> +			/* TODO maybe we want to re-assign hwpipe sometimes
> +			 * in cases when we no-longer need some caps to make
> +			 * it available for other planes?
> +			 */
> +			struct mdp5_hw_pipe *hwpipe = mdp5_state->hwpipe;

Maybe hwpipe above could be curr_hwpipe or old_hwpipe?

> +			mdp5_state->hwpipe =
> +				mdp5_pipe_assign(state->state, plane, caps);
> +			if (IS_ERR(mdp5_state->hwpipe)) {
> +				DBG("%s: failed to assign hwpipe!", plane->name);
> +				return PTR_ERR(mdp5_state->hwpipe);
> +			}
> +			mdp5_pipe_release(state->state, hwpipe);
>  		}
>  	}

Thanks,
Archit

-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project


More information about the Freedreno mailing list