Allow specifying a syncobj on render job submission where we store the
fence for the job. This gives userland flexible access to the fence.
Signed-off-by: Stefan Schake <***@gmail.com>
---
drivers/gpu/drm/vc4/vc4_gem.c | 38 +++++++++++++++++++++++++++++++++++---
include/uapi/drm/vc4_drm.h | 13 +++++++++++++
2 files changed, 48 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/vc4/vc4_gem.c b/drivers/gpu/drm/vc4/vc4_gem.c
index 232363488125..b39515a4ddcb 100644
--- a/drivers/gpu/drm/vc4/vc4_gem.c
+++ b/drivers/gpu/drm/vc4/vc4_gem.c
@@ -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;
diff --git a/include/uapi/drm/vc4_drm.h b/include/uapi/drm/vc4_drm.h
index 389f21931c25..3a2ef9b5b60b 100644
--- a/include/uapi/drm/vc4_drm.h
+++ b/include/uapi/drm/vc4_drm.h
@@ -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;
};
/**
--
2.14.1