Skip to contentSkip to conversation

drm/v3d: clean caches at the end of render jobs on request from user space

PATCH Superseded Iago Toral Quiroga 2 participants 3 messages Sep 18, 2019

Subsystems (3)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 drm/v3d: clean caches at the end of render jobs on request from user space

Extends the user space ioctl for CL submissions so it can include a request to flush the cache once the CL execution has completed. Fixes memory write violation messages reported by the kernel in workloads involving shader memory writes (SSBOs, shader images, scratch, etc) which sometimes also lead to GPU resets during Piglit and CTS workloads.

v2: if v3d_job_init() fails we need to kfree() the job instead of

    v3d_job_put() it (Eric Anholt).
---
 drivers/gpu/drm/v3d/v3d_gem.c | 54 +++++++++++++++++++++++++++++------
 include/uapi/drm/v3d_drm.h    |  7 +++--
 2 files changed, 50 insertions(+), 11 deletions(-)
drivers/gpu/drm/v3d/v3d_gem.c +46 -8
index 5d80507b539b..d46d91346d09 100644
@@ -530,13 +530,16 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
 	struct drm_v3d_submit_cl *args = data;
 	struct v3d_bin_job *bin = NULL;
 	struct v3d_render_job *render;
+	struct v3d_job *clean_job = NULL;
+	struct v3d_job *last_job;
 	struct ww_acquire_ctx acquire_ctx;
 	int ret = 0;
 
 	trace_v3d_submit_cl_ioctl(&v3d->drm, args->rcl_start, args->rcl_end);
 
-	if (args->pad != 0) {
-		DRM_INFO("pad must be zero: %d\n", args->pad);
+	if (args->flags != 0 &&
+	    args->flags != DRM_V3D_SUBMIT_CL_FLUSH_CACHE_FLAG) {
+		DRM_INFO("invalid flags: %d\n", args->flags);
 		return -EINVAL;
 	}
 
@@ -575,12 +578,31 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
 		bin->render = render;
 	}
 
-	ret = v3d_lookup_bos(dev, file_priv, &render->base,
+	if (args->flags & DRM_V3D_SUBMIT_CL_FLUSH_CACHE_FLAG) {
+		clean_job = kcalloc(1, sizeof(*clean_job), GFP_KERNEL);
+		if (!clean_job) {
+			ret = -ENOMEM;
+			goto fail;
+		}
+
+		ret = v3d_job_init(v3d, file_priv, clean_job, v3d_job_free, 0);
+		if (ret) {
+			kfree(clean_job);
+			clean_job = NULL;
+			goto fail;
+		}
+
+		last_job = clean_job;
+	} else {
+		last_job = &render->base;
+	}
+
+	ret = v3d_lookup_bos(dev, file_priv, last_job,
 			     args->bo_handles, args->bo_handle_count);
 	if (ret)
 		goto fail;
 
-	ret = v3d_lock_bo_reservations(&render->base, &acquire_ctx);
+	ret = v3d_lock_bo_reservations(last_job, &acquire_ctx);
 	if (ret)
 		goto fail;
 
@@ -599,28 +621,44 @@ v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
 	ret = v3d_push_job(v3d_priv, &render->base, V3D_RENDER);
 	if (ret)
 		goto fail_unreserve;
+
+	if (clean_job) {
+		struct dma_fence *render_fence =
+			dma_fence_get(render->base.done_fence);
+		ret = drm_gem_fence_array_add(&clean_job->deps, render_fence);
+		if (ret)
+			goto fail_unreserve;
+		ret = v3d_push_job(v3d_priv, clean_job, V3D_CACHE_CLEAN);
+		if (ret)
+			goto fail_unreserve;
+	}
+
 	mutex_unlock(&v3d->sched_lock);
 
 	v3d_attach_fences_and_unlock_reservation(file_priv,
-						 &render->base,
+						 last_job,
 						 &acquire_ctx,
 						 args->out_sync,
-						 render->base.done_fence);
+						 last_job->done_fence);
 
 	if (bin)
 		v3d_job_put(&bin->base);
 	v3d_job_put(&render->base);
+	if (clean_job)
+		v3d_job_put(clean_job);
 
 	return 0;
 
 fail_unreserve:
 	mutex_unlock(&v3d->sched_lock);
-	drm_gem_unlock_reservations(render->base.bo,
-				    render->base.bo_count, &acquire_ctx);
+	drm_gem_unlock_reservations(last_job->bo,
+				    last_job->bo_count, &acquire_ctx);
 fail:
 	if (bin)
 		v3d_job_put(&bin->base);
 	v3d_job_put(&render->base);
+	if (clean_job)
+		v3d_job_put(clean_job);
 
 	return ret;
 }
include/uapi/drm/v3d_drm.h +4 -3
index 58fbe48c91e9..58d2040ea48c 100644
@@ -48,6 +48,8 @@ extern "C" {
 #define DRM_IOCTL_V3D_SUBMIT_TFU          DRM_IOW(DRM_COMMAND_BASE + DRM_V3D_SUBMIT_TFU, struct drm_v3d_submit_tfu)
 #define DRM_IOCTL_V3D_SUBMIT_CSD          DRM_IOW(DRM_COMMAND_BASE + DRM_V3D_SUBMIT_CSD, struct drm_v3d_submit_csd)
 
+#define DRM_V3D_SUBMIT_CL_FLUSH_CACHE_FLAG        0x01
+
 /**
  * struct drm_v3d_submit_cl - ioctl argument for submitting commands to the 3D
  * engine.
@@ -61,7 +63,7 @@ extern "C" {
  * flushed by the time the render done IRQ happens, which is the
  * trigger for out_sync.  Any dirtying of cachelines by the job (only
  * possible using TMU writes) must be flushed by the caller using the
- * CL's cache flush commands.
+ * DRM_V3D_SUBMIT_CL_FLUSH_CACHE_FLAG flag.
  */
 struct drm_v3d_submit_cl {
 	/* Pointer to the binner command list.
@@ -124,8 +126,7 @@ struct drm_v3d_submit_cl {
 	/* Number of BO handles passed in (size is that times 4). */
 	__u32 bo_handle_count;
 
-	/* Pad, must be zero-filled. */
-	__u32 pad;
+	__u32 flags;
 };
 
 /**
Signature · 6 lines
2.17.1

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#1 lore raw

Eric, could you push this for me? I don't have push rights to DRM yet.

Iago

Quoted 166 lines · Iago Toral Quiroga

Extends the user space ioctl for CL submissions so it can include a request
to flush the cache once the CL execution has completed. Fixes memory write violation messages reported by the kernel in workloads involving
shader memory writes (SSBOs, shader images, scratch, etc) which sometimes
also lead to GPU resets during Piglit and CTS workloads.

v2: if v3d_job_init() fails we need to kfree() the job instead of

    v3d_job_put() it (Eric Anholt).

Link:
https://patchwork.freedesktop.org/patch/msgid/[email protected]

---
 drivers/gpu/drm/v3d/v3d_gem.c | 54 +++++++++++++++++++++++++++++--
----
 include/uapi/drm/v3d_drm.h    |  7 +++--
 2 files changed, 50 insertions(+), 11 deletions(-)
drivers/gpu/drm/v3d/v3d_gem.c +5 -2
index 5d80507b539b..d46d91346d09 100644
@@ -530,13 +530,16 @@ v3d_submit_cl_ioctl(struct drm_device *dev,
 void *data,
 	struct drm_v3d_submit_cl *args = data;
 	struct v3d_bin_job *bin = NULL;
 	struct v3d_render_job *render;
+	struct v3d_job *clean_job = NULL;
+	struct v3d_job *last_job;
 	struct ww_acquire_ctx acquire_ctx;
 	int ret = 0;
 
 	trace_v3d_submit_cl_ioctl(&v3d->drm, args->rcl_start, args-
 >rcl_end);
 
-	if (args->pad != 0) {
-		DRM_INFO("pad must be zero: %d\n", args->pad);
+	if (args->flags != 0 &&
+	    args->flags != DRM_V3D_SUBMIT_CL_FLUSH_CACHE_FLAG) {
+		DRM_INFO("invalid flags: %d\n", args->flags);
 		return -EINVAL;
drivers/gpu/drm/v3d/v3d_gem.c +21 -2
 	}
@@ -575,12 +578,31 @@ v3d_submit_cl_ioctl(struct drm_device *dev,
 void *data,
 		bin->render = render;
 	}
 
-	ret = v3d_lookup_bos(dev, file_priv, &render->base,
+	if (args->flags & DRM_V3D_SUBMIT_CL_FLUSH_CACHE_FLAG) {
+		clean_job = kcalloc(1, sizeof(*clean_job), GFP_KERNEL);
+		if (!clean_job) {
+			ret = -ENOMEM;
+			goto fail;
+		}
+
+		ret = v3d_job_init(v3d, file_priv, clean_job,
 v3d_job_free, 0);
+		if (ret) {
+			kfree(clean_job);
+			clean_job = NULL;
+			goto fail;
+		}
+
+		last_job = clean_job;
+	} else {
+		last_job = &render->base;
+	}
+
+	ret = v3d_lookup_bos(dev, file_priv, last_job,
 			     args->bo_handles, args->bo_handle_count);
 	if (ret)
 		goto fail;
 
-	ret = v3d_lock_bo_reservations(&render->base, &acquire_ctx);
+	ret = v3d_lock_bo_reservations(last_job, &acquire_ctx);
 	if (ret)
drivers/gpu/drm/v3d/v3d_gem.c +20 -4
 		goto fail;
@@ -599,28 +621,44 @@ v3d_submit_cl_ioctl(struct drm_device *dev,
 void *data,
 	ret = v3d_push_job(v3d_priv, &render->base, V3D_RENDER);
 	if (ret)
 		goto fail_unreserve;
+
+	if (clean_job) {
+		struct dma_fence *render_fence =
+			dma_fence_get(render->base.done_fence);
+		ret = drm_gem_fence_array_add(&clean_job->deps,
 render_fence);
+		if (ret)
+			goto fail_unreserve;
+		ret = v3d_push_job(v3d_priv, clean_job,
 V3D_CACHE_CLEAN);
+		if (ret)
+			goto fail_unreserve;
+	}
+
 	mutex_unlock(&v3d->sched_lock);
 
 	v3d_attach_fences_and_unlock_reservation(file_priv,
-						 &render->base,
+						 last_job,
 						 &acquire_ctx,
 						 args->out_sync,
-						 render-
 >base.done_fence);
+						 last_job->done_fence);
 
 	if (bin)
 		v3d_job_put(&bin->base);
 	v3d_job_put(&render->base);
+	if (clean_job)
+		v3d_job_put(clean_job);
 
 	return 0;
 
 fail_unreserve:
 	mutex_unlock(&v3d->sched_lock);
-	drm_gem_unlock_reservations(render->base.bo,
-				    render->base.bo_count,
 &acquire_ctx);
+	drm_gem_unlock_reservations(last_job->bo,
+				    last_job->bo_count, &acquire_ctx);
 fail:
 	if (bin)
 		v3d_job_put(&bin->base);
 	v3d_job_put(&render->base);
+	if (clean_job)
+		v3d_job_put(clean_job);
 
 	return ret;
 }
include/uapi/drm/v3d_drm.h +2 -0
index 58fbe48c91e9..58d2040ea48c 100644
@@ -48,6 +48,8 @@ extern "C" {
 #define DRM_IOCTL_V3D_SUBMIT_TFU          DRM_IOW(DRM_COMMAND_BASE +
 DRM_V3D_SUBMIT_TFU, struct drm_v3d_submit_tfu)
 #define DRM_IOCTL_V3D_SUBMIT_CSD          DRM_IOW(DRM_COMMAND_BASE +
 DRM_V3D_SUBMIT_CSD, struct drm_v3d_submit_csd)
 
+#define DRM_V3D_SUBMIT_CL_FLUSH_CACHE_FLAG        0x01
+
 /**

* struct drm_v3d_submit_cl - ioctl argument for submitting commands
to the 3D
* engine.

include/uapi/drm/v3d_drm.h +0 -0
@@ -61,7 +63,7 @@ extern "C" {
  * flushed by the time the render done IRQ happens, which is the
  * trigger for out_sync.  Any dirtying of cachelines by the job

(only
* possible using TMU writes) must be flushed by the caller using
the
- * CL's cache flush commands.
+ * DRM_V3D_SUBMIT_CL_FLUSH_CACHE_FLAG flag.
*/
struct drm_v3d_submit_cl {

include/uapi/drm/v3d_drm.h +0 -0
 	/* Pointer to the binner command list.

include/uapi/drm/v3d_drm.h struct drm_v3d_submit_cl { · line 132

Quoted 9 lines
include/uapi/drm/v3d_drm.h +1 -2
@@ -124,8 +126,7 @@ struct drm_v3d_submit_cl {
 	/* Number of BO handles passed in (size is that times 4). */
 	__u32 bo_handle_count;
 
-	/* Pad, must be zero-filled. */
-	__u32 pad;
+	__u32 flags;
 };
 
 /**

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#2 lore raw
In reply to Iago Toral Quiroga
Quoted 18 lines · Iago Toral Quiroga

Extends the user space ioctl for CL submissions so it can include a request to flush the cache once the CL execution has completed. Fixes memory write violation messages reported by the kernel in workloads involving shader memory writes (SSBOs, shader images, scratch, etc) which sometimes also lead to GPU resets during Piglit and CTS workloads.

v2: if v3d_job_init() fails we need to kfree() the job instead of

    v3d_job_put() it (Eric Anholt).
include/uapi/drm/v3d_drm.h +0 -0
index 58fbe48c91e9..58d2040ea48c 100644

include/uapi/drm/v3d_drm.h extern "C" { · line 52

Quoted 5 lines
include/uapi/drm/v3d_drm.h +2 -0
@@ -48,6 +48,8 @@ extern "C" {
 #define DRM_IOCTL_V3D_SUBMIT_TFU          DRM_IOW(DRM_COMMAND_BASE + DRM_V3D_SUBMIT_TFU, struct drm_v3d_submit_tfu)
 #define DRM_IOCTL_V3D_SUBMIT_CSD          DRM_IOW(DRM_COMMAND_BASE + DRM_V3D_SUBMIT_CSD, struct drm_v3d_submit_csd)
 
+#define DRM_V3D_SUBMIT_CL_FLUSH_CACHE_FLAG        0x01
+

Hmm. vc4, msm, panfrost, etnaviv, i915 all name their submit flags without "_FLAG" in the name, can we drop that?

Also, I just noticed: You don't have a new param to indicate the availability of the new flag. You're going to need that (unless you've got some other clever plan?) so that new Mesa can detect an old kernel and not expose the GLES 3.1 features that require it.

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel