drm/vc4: Syncobj import export

index 94b99c90425a..806c8004b793 100644
@@ -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,

index 4288615b66a2..3105df99cb12 100644
@@ -10,6 +10,8 @@
 #include <drm/drmP.h>
 #include <drm/drm_encoder.h>
 #include <drm/drm_gem_cma_helper.h>
+#include <drm/drm_syncobj.h>
+
 
 #include "uapi/drm/vc4_drm.h"
 

index 2107b0daf8ef..232363488125 100644
@@ -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"

@@ -1115,21 +1116,18 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 	struct drm_vc4_submit_cl *args = data;
 	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 |
 			     VC4_SUBMIT_CL_FIXED_RCL_ORDER |
 			     VC4_SUBMIT_CL_RCL_ORDER_INCREASING_X |
-			     VC4_SUBMIT_CL_RCL_ORDER_INCREASING_Y)) != 0) {
+			     VC4_SUBMIT_CL_RCL_ORDER_INCREASING_Y |
+			     VC4_SUBMIT_CL_IMPORT_SYNCOBJ)) != 0) {
 		DRM_DEBUG("Unknown flags: 0x%02x\n", args->flags);
 		return -EINVAL;
 	}
 
-	if (args->pad2 != 0) {
-		DRM_DEBUG("->pad2 must be set to zero\n");
-		return -EINVAL;
-	}
-
 	exec = kcalloc(1, sizeof(*exec), GFP_KERNEL);
 	if (!exec) {
 		DRM_ERROR("malloc failure on exec struct\n");

@@ -1164,6 +1162,29 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 		}
 	}
 
+	if (args->flags & VC4_SUBMIT_CL_IMPORT_SYNCOBJ) {
+		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)

index b95a0e11cb07..389f21931c25 100644
@@ -173,6 +173,7 @@ struct drm_vc4_submit_cl {
 #define VC4_SUBMIT_CL_FIXED_RCL_ORDER			(1 << 1)
 #define VC4_SUBMIT_CL_RCL_ORDER_INCREASING_X		(1 << 2)
 #define VC4_SUBMIT_CL_RCL_ORDER_INCREASING_Y		(1 << 3)
+#define VC4_SUBMIT_CL_IMPORT_SYNCOBJ			(1 << 4)
 	__u32 flags;
 
 	/* Returned value of the seqno of this render job (for the

@@ -183,11 +184,11 @@ struct drm_vc4_submit_cl {
 	/* ID of the perfmon to attach to this job. 0 means no perfmon. */
 	__u32 perfmonid;
 
-	/* Unused field to align this struct on 64 bits. Must be set to 0.
-	 * If one ever needs to add an u32 field to this struct, this field
-	 * can be used.
+	/* Syncobj handle to wait on. Set together with IMPORT_SYNCOBJ flag.
+	 * If set, processing of this render job will not start until the
+	 * syncobj is signalled.
 	 */
-	__u32 pad2;
+	__u32 in_sync;
 };
 
 /**

index 232363488125..b39515a4ddcb 100644
@@ -656,7 +656,8 @@ vc4_lock_bo_reservations(struct drm_device *dev,
  */
 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;

@@ -679,6 +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);

@@ -1114,6 +1118,7 @@ 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;

@@ -1123,11 +1128,17 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 			     VC4_SUBMIT_CL_FIXED_RCL_ORDER |
 			     VC4_SUBMIT_CL_RCL_ORDER_INCREASING_X |
 			     VC4_SUBMIT_CL_RCL_ORDER_INCREASING_Y |
-			     VC4_SUBMIT_CL_IMPORT_SYNCOBJ)) != 0) {
+			     VC4_SUBMIT_CL_IMPORT_SYNCOBJ |
+			     VC4_SUBMIT_CL_EXPORT_SYNCOBJ)) != 0) {
 		DRM_DEBUG("Unknown flags: 0x%02x\n", args->flags);
 		return -EINVAL;
 	}
 
+	if (args->pad2 != 0) {
+		DRM_DEBUG("->pad2 must be set to zero\n");
+		return -EINVAL;
+	}
+
 	exec = kcalloc(1, sizeof(*exec), GFP_KERNEL);
 	if (!exec) {
 		DRM_ERROR("malloc failure on exec struct\n");

@@ -1202,12 +1213,33 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 	if (ret)
 		goto fail;
 
+	if (args->flags & VC4_SUBMIT_CL_EXPORT_SYNCOBJ) {
+		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;
 

index 389f21931c25..3a2ef9b5b60b 100644
@@ -174,6 +174,7 @@ struct drm_vc4_submit_cl {
 #define VC4_SUBMIT_CL_RCL_ORDER_INCREASING_X		(1 << 2)
 #define VC4_SUBMIT_CL_RCL_ORDER_INCREASING_Y		(1 << 3)
 #define VC4_SUBMIT_CL_IMPORT_SYNCOBJ			(1 << 4)
+#define VC4_SUBMIT_CL_EXPORT_SYNCOBJ			(1 << 5)
 	__u32 flags;
 
 	/* Returned value of the seqno of this render job (for the

@@ -189,6 +190,18 @@ struct drm_vc4_submit_cl {
 	 * syncobj is signalled.
 	 */
 	__u32 in_sync;
+
+	/* Syncobj handle to export fence to. Set together with EXPORT_SYNCOBJ
+	 * flag. If set, the fence in the syncobj will be replaced with a fence
+	 * that signals upon completion of this render job.
+	 */
+	__u32 out_sync;
+
+	/* Unused field to align this struct on 64 bits. Must be set to 0.
+	 * If one ever needs to add an u32 field to this struct, this field
+	 * can be used.
+	 */
+	__u32 pad2;
 };
 
 /**

index 806c8004b793..4e2ae2a9a164 100644
@@ -102,6 +102,7 @@ static int vc4_get_param_ioctl(struct drm_device *dev, void *data,
 	case DRM_VC4_PARAM_SUPPORTS_FIXED_RCL_ORDER:
 	case DRM_VC4_PARAM_SUPPORTS_MADVISE:
 	case DRM_VC4_PARAM_SUPPORTS_PERFMON:
+	case DRM_VC4_PARAM_SUPPORTS_SYNCOBJ:
 		args->value = true;
 		break;
 	default:

index 3a2ef9b5b60b..6f81cf05f7a3 100644
@@ -338,6 +338,7 @@ struct drm_vc4_get_hang_state {
 #define DRM_VC4_PARAM_SUPPORTS_FIXED_RCL_ORDER	6
 #define DRM_VC4_PARAM_SUPPORTS_MADVISE		7
 #define DRM_VC4_PARAM_SUPPORTS_PERFMON		8
+#define DRM_VC4_PARAM_SUPPORTS_SYNCOBJ		9
 
 struct drm_vc4_get_param {
 	__u32 param;