[PATCH v2 1/3] drm/vc4: Syncobj import support

[Stefan Schake]

Allow userland to specify a syncobj that is waited on before a render job
starts processing.

v2: Use 0 as invalid syncobj to drop flag (Eric)
Drop extra newline (Eric)

Signed-off-by: Stefan Schake <***@gmail.com>
---
drivers/gpu/drm/vc4/vc4_drv.h | 1 +
drivers/gpu/drm/vc4/vc4_gem.c | 30 +++++++++++++++++++++++++-----
include/uapi/drm/vc4_drm.h | 7 +++----
3 files changed, 29 insertions(+), 9 deletions(-)

diff --git a/drivers/gpu/drm/vc4/vc4_drv.h b/drivers/gpu/drm/vc4/vc4_drv.h
index 22589d3..554a4e8 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_gem.c b/drivers/gpu/drm/vc4/vc4_gem.c
index 2107b0d..e305ccde 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"
@@ -1115,6 +1116,7 @@ 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 |
@@ -1125,11 +1127,6 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
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 +1161,29 @@ 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)
diff --git a/include/uapi/drm/vc4_drm.h b/include/uapi/drm/vc4_drm.h
index b95a0e1..d97065b 100644
--- a/include/uapi/drm/vc4_drm.h
+++ b/include/uapi/drm/vc4_drm.h
@@ -183,11 +183,10 @@ 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. If set, processing of this render job
+ * will not start until the syncobj is signaled. 0 means ignore.
*/
- __u32 pad2;
+ __u32 in_sync;
};

/**

[Stefan Schake]

This doesn't require any additional functionality from the driver but
is a prerequisite to userland calling the syncobj ioctls.

Signed-off-by: Stefan Schake <***@gmail.com>
---
drivers/gpu/drm/vc4/vc4_drv.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/vc4/vc4_drv.c b/drivers/gpu/drm/vc4/vc4_drv.c
index 40ddeaa..d9b8b70 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,

[Stefan Schake]

Allow specifying a syncobj on render job submission where we store the
fence for the job. This gives userland flexible access to the fence.

v2: Use 0 as invalid syncobj to drop flag (Eric)
Don't reintroduce the padding (Eric)

Signed-off-by: Stefan Schake <***@gmail.com>
---
drivers/gpu/drm/vc4/vc4_gem.c | 30 ++++++++++++++++++++++++++++--
include/uapi/drm/vc4_drm.h | 6 ++++++
2 files changed, 34 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/vc4/vc4_gem.c b/drivers/gpu/drm/vc4/vc4_gem.c
index e305ccde..a4c4be3 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;
@@ -1201,12 +1206,33 @@ 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/include/uapi/drm/vc4_drm.h b/include/uapi/drm/vc4_drm.h
index d97065b..2be4fe3 100644
--- a/include/uapi/drm/vc4_drm.h
+++ b/include/uapi/drm/vc4_drm.h
@@ -187,6 +187,12 @@ struct drm_vc4_submit_cl {
* will not start until the syncobj is signaled. 0 means ignore.
*/
__u32 in_sync;
+
+ /* Syncobj handle to export fence to. If set, the fence in the syncobj
+ * will be replaced with a fence that signals upon completion of this
+ * render job. 0 means ignore.
+ */
+ __u32 out_sync;
};

/**

[Eric Anholt]

v2 drops the extra syncobj parameter and gets rid of the submit flags
since 0 is never a valid syncobj handle.
This series allows userspace to submit syncobj handles for importing
a fence to wait on or exporting the job fence as part of submission.
The primary use of this is to enable native fence fd support in Mesa.

This series was a joy to read. I've reviewed and pushed it to
drm-misc-next, and we can land the Mesa side once it's made its way to
drm-next.