5 files changed, 63 insertions, 6 deletions

diff --git a/drivers/gpu/drm/vc4/vc4_drv.c b/drivers/gpu/drm/vc4/vc4_drv.c
index 40ddeaafd65f..d9b8b701d2ce 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.c
+++ b/drivers/gpu/drm/vc4/vc4_drv.c
@@ -175,7 +175,8 @@ static struct drm_driver vc4_drm_driver = {
 			    DRIVER_GEM |
 			    DRIVER_HAVE_IRQ |
 			    DRIVER_RENDER |
-			    DRIVER_PRIME),
+			    DRIVER_PRIME |
+			    DRIVER_SYNCOBJ),
 	.lastclose = drm_fb_helper_lastclose,
 	.open = vc4_open,
 	.postclose = vc4_close,
diff --git a/drivers/gpu/drm/vc4/vc4_drv.h b/drivers/gpu/drm/vc4/vc4_drv.h
index 22589d39083c..554a4e810d5b 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.h
+++ b/drivers/gpu/drm/vc4/vc4_drv.h
@@ -11,6 +11,7 @@
 #include <drm/drm_encoder.h>
 #include <drm/drm_gem_cma_helper.h>
 #include <drm/drm_atomic.h>
+#include <drm/drm_syncobj.h>
 
 #include "uapi/drm/vc4_drm.h"
 
diff --git a/drivers/gpu/drm/vc4/vc4_dsi.c b/drivers/gpu/drm/vc4/vc4_dsi.c
index 94085f8bcd68..8aa897835118 100644
--- a/drivers/gpu/drm/vc4/vc4_dsi.c
+++ b/drivers/gpu/drm/vc4/vc4_dsi.c
@@ -753,6 +753,11 @@ static void vc4_dsi_ulps(struct vc4_dsi *dsi, bool ulps)
 			 (dsi->lanes > 2 ? DSI1_STAT_PHY_D2_STOP : 0) |
 			 (dsi->lanes > 3 ? DSI1_STAT_PHY_D3_STOP : 0));
 	int ret;
+	bool ulps_currently_enabled = (DSI_PORT_READ(PHY_AFEC0) &
+				       DSI_PORT_BIT(PHY_AFEC0_LATCH_ULPS));
+
+	if (ulps == ulps_currently_enabled)
+		return;
 
 	DSI_PORT_WRITE(STAT, stat_ulps);
 	DSI_PORT_WRITE(PHYC, DSI_PORT_READ(PHYC) | phyc_ulps);
diff --git a/drivers/gpu/drm/vc4/vc4_gem.c b/drivers/gpu/drm/vc4/vc4_gem.c
index 2107b0daf8ef..7910b9acedd6 100644
--- a/drivers/gpu/drm/vc4/vc4_gem.c
+++ b/drivers/gpu/drm/vc4/vc4_gem.c
@@ -27,6 +27,7 @@
 #include <linux/device.h>
 #include <linux/io.h>
 #include <linux/sched/signal.h>
+#include <linux/dma-fence-array.h>
 
 #include "uapi/drm/vc4_drm.h"
 #include "vc4_drv.h"
@@ -655,7 +656,8 @@ retry:
  */
 static int
 vc4_queue_submit(struct drm_device *dev, struct vc4_exec_info *exec,
-		 struct ww_acquire_ctx *acquire_ctx)
+		 struct ww_acquire_ctx *acquire_ctx,
+		 struct drm_syncobj *out_sync)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
 	struct vc4_exec_info *renderjob;
@@ -678,6 +680,9 @@ vc4_queue_submit(struct drm_device *dev, struct vc4_exec_info *exec,
 	fence->seqno = exec->seqno;
 	exec->fence = &fence->base;
 
+	if (out_sync)
+		drm_syncobj_replace_fence(out_sync, exec->fence);
+
 	vc4_update_bo_seqnos(exec, seqno);
 
 	vc4_unlock_bo_reservations(dev, exec, acquire_ctx);
@@ -1113,8 +1118,10 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
 	struct vc4_file *vc4file = file_priv->driver_priv;
 	struct drm_vc4_submit_cl *args = data;
+	struct drm_syncobj *out_sync = NULL;
 	struct vc4_exec_info *exec;
 	struct ww_acquire_ctx acquire_ctx;
+	struct dma_fence *in_fence;
 	int ret = 0;
 
 	if ((args->flags & ~(VC4_SUBMIT_CL_USE_CLEAR_COLOR |
@@ -1126,7 +1133,7 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 	}
 
 	if (args->pad2 != 0) {
-		DRM_DEBUG("->pad2 must be set to zero\n");
+		DRM_DEBUG("Invalid pad: 0x%08x\n", args->pad2);
 		return -EINVAL;
 	}
 
@@ -1164,6 +1171,29 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 		}
 	}
 
+	if (args->in_sync) {
+		ret = drm_syncobj_find_fence(file_priv, args->in_sync,
+					     &in_fence);
+		if (ret)
+			goto fail;
+
+		/* When the fence (or fence array) is exclusively from our
+		 * context we can skip the wait since jobs are executed in
+		 * order of their submission through this ioctl and this can
+		 * only have fences from a prior job.
+		 */
+		if (!dma_fence_match_context(in_fence,
+					     vc4->dma_fence_context)) {
+			ret = dma_fence_wait(in_fence, true);
+			if (ret) {
+				dma_fence_put(in_fence);
+				goto fail;
+			}
+		}
+
+		dma_fence_put(in_fence);
+	}
+
 	if (exec->args->bin_cl_size != 0) {
 		ret = vc4_get_bcl(dev, exec);
 		if (ret)
@@ -1181,12 +1211,33 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 	if (ret)
 		goto fail;
 
+	if (args->out_sync) {
+		out_sync = drm_syncobj_find(file_priv, args->out_sync);
+		if (!out_sync) {
+			ret = -EINVAL;
+			goto fail;
+		}
+
+		/* We replace the fence in out_sync in vc4_queue_submit since
+		 * the render job could execute immediately after that call.
+		 * If it finishes before our ioctl processing resumes the
+		 * render job fence could already have been freed.
+		 */
+	}
+
 	/* Clear this out of the struct we'll be putting in the queue,
 	 * since it's part of our stack.
 	 */
 	exec->args = NULL;
 
-	ret = vc4_queue_submit(dev, exec, &acquire_ctx);
+	ret = vc4_queue_submit(dev, exec, &acquire_ctx, out_sync);
+
+	/* The syncobj isn't part of the exec data and we need to free our
+	 * reference even if job submission failed.
+	 */
+	if (out_sync)
+		drm_syncobj_put(out_sync);
+
 	if (ret)
 		goto fail;
 
diff --git a/drivers/gpu/drm/vc4/vc4_v3d.c b/drivers/gpu/drm/vc4/vc4_v3d.c
index bfc2fa73d2ae..e47e29426078 100644
--- a/drivers/gpu/drm/vc4/vc4_v3d.c
+++ b/drivers/gpu/drm/vc4/vc4_v3d.c
@@ -218,8 +218,7 @@ try_again:
  * overall CMA pool before they make scenes complicated enough to run
  * out of bin space.
  */
-int
-vc4_allocate_bin_bo(struct drm_device *drm)
+static int vc4_allocate_bin_bo(struct drm_device *drm)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(drm);
 	struct vc4_v3d *v3d = vc4->v3d;