drm/v3d: Add support for compute shader dispatch.

The compute shader dispatch interface is pretty simple -- just pass in
the regs that userspace has passed us, with no CLs to run. However,
with no CL to run it means that we need to do manual cache flushing of
the L2 after the HW execution completes (for SSBO, atomic, and
image_load_store writes that are the output of compute shaders).

This doesn't yet expose the L2 cache's ability to have a region of the
address space not write back to memory (which could be used for
shared_var storage).

So far, the Mesa side has been tested on V3D v4.2 simpenrose (passing
the ES31 tests), and on the kernel side on 7278 (failing atomic
compswap tests in a way that doesn't reproduce on simpenrose).

v2: Fix excessive allocation for the clean_job (reported by Dan
Carpenter). Keep refs on jobs until clean_job is finished, to
avoid spurious MMU errors if the output BOs are freed by userspace
before L2 cleaning is finished.

Signed-off-by: Eric Anholt <[e-mail]>
Link: https://patchwork.freedesktop.org/patch/msgid/[e-mail]
Acked-by: Rob Clark <[e-mail]>

drivers/gpu/drm/v3d/v3d_debugfs.c

···
 	REGDEF(V3D_GMP_VIO_ADDR),
 };
 
+static const struct v3d_reg_def v3d_csd_reg_defs[] = {
+	REGDEF(V3D_CSD_STATUS),
+	REGDEF(V3D_CSD_CURRENT_CFG0),
+	REGDEF(V3D_CSD_CURRENT_CFG1),
+	REGDEF(V3D_CSD_CURRENT_CFG2),
+	REGDEF(V3D_CSD_CURRENT_CFG3),
+	REGDEF(V3D_CSD_CURRENT_CFG4),
+	REGDEF(V3D_CSD_CURRENT_CFG5),
+	REGDEF(V3D_CSD_CURRENT_CFG6),
+};
+
 static int v3d_v3d_debugfs_regs(struct seq_file *m, void *unused)
 {
 	struct drm_info_node *node = (struct drm_info_node *)m->private;
···
 				   v3d_core_reg_defs[i].reg,
 				   V3D_CORE_READ(core,
 						 v3d_core_reg_defs[i].reg));
+		}
+
+		if (v3d_has_csd(v3d)) {
+			for (i = 0; i < ARRAY_SIZE(v3d_csd_reg_defs); i++) {
+				seq_printf(m, "core %d %s (0x%04x): 0x%08x\n",
+					   core,
+					   v3d_csd_reg_defs[i].name,
+					   v3d_csd_reg_defs[i].reg,
+					   V3D_CORE_READ(core,
+							 v3d_csd_reg_defs[i].reg));
+			}
 		}
 	}
 

drivers/gpu/drm/v3d/v3d_drv.c

···
  * This driver supports the Broadcom V3D 3.3 and 4.1 OpenGL ES GPUs.
  * For V3D 2.x support, see the VC4 driver.
  *
- * Currently only single-core rendering using the binner and renderer,
- * along with TFU (texture formatting unit) rendering is supported.
- * V3D 4.x's CSD (compute shader dispatch) is not yet supported.
+ * The V3D GPU includes a tiled render (composed of a bin and render
+ * pipelines), the TFU (texture formatting unit), and the CSD (compute
+ * shader dispatch).
  */
 
 #include <linux/clk.h>
···
 	case DRM_V3D_PARAM_SUPPORTS_TFU:
 		args->value = 1;
 		return 0;
+	case DRM_V3D_PARAM_SUPPORTS_CSD:
+		args->value = v3d_has_csd(v3d);
+		return 0;
 	default:
 		DRM_DEBUG("Unknown parameter %d\n", args->param);
 		return -EINVAL;
···
 	DRM_IOCTL_DEF_DRV(V3D_GET_PARAM, v3d_get_param_ioctl, DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(V3D_GET_BO_OFFSET, v3d_get_bo_offset_ioctl, DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(V3D_SUBMIT_TFU, v3d_submit_tfu_ioctl, DRM_RENDER_ALLOW | DRM_AUTH),
+	DRM_IOCTL_DEF_DRV(V3D_SUBMIT_CSD, v3d_submit_csd_ioctl, DRM_RENDER_ALLOW | DRM_AUTH),
 };
 
 static struct drm_driver v3d_drm_driver = {

drivers/gpu/drm/v3d/v3d_drv.h

···
 	V3D_BIN,
 	V3D_RENDER,
 	V3D_TFU,
+	V3D_CSD,
+	V3D_CACHE_CLEAN,
 };
 
-#define V3D_MAX_QUEUES (V3D_TFU + 1)
+#define V3D_MAX_QUEUES (V3D_CACHE_CLEAN + 1)
 
 struct v3d_queue_state {
 	struct drm_gpu_scheduler sched;
···
 	struct v3d_bin_job *bin_job;
 	struct v3d_render_job *render_job;
 	struct v3d_tfu_job *tfu_job;
+	struct v3d_csd_job *csd_job;
 
 	struct v3d_queue_state queue[V3D_MAX_QUEUES];
 
···
 	 */
 	struct mutex sched_lock;
 
+	/* Lock taken during a cache clean and when initiating an L2
+	 * flush, to keep L2 flushes from interfering with the
+	 * synchronous L2 cleans.
+	 */
+	struct mutex cache_clean_lock;
+
 	struct {
 		u32 num_allocated;
 		u32 pages_allocated;
···
 to_v3d_dev(struct drm_device *dev)
 {
 	return (struct v3d_dev *)dev->dev_private;
+}
+
+static inline bool
+v3d_has_csd(struct v3d_dev *v3d)
+{
+	return v3d->ver >= 41;
 }
 
 /* The per-fd struct, which tracks the MMU mappings. */
···
 	struct drm_v3d_submit_tfu args;
 };
 
+struct v3d_csd_job {
+	struct v3d_job base;
+
+	u32 timedout_batches;
+
+	struct drm_v3d_submit_csd args;
+};
+
 /**
  * _wait_for - magic (register) wait macro
  *
···
 			struct drm_file *file_priv);
 int v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
 			 struct drm_file *file_priv);
+int v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
+			 struct drm_file *file_priv);
 int v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
 		      struct drm_file *file_priv);
 void v3d_job_put(struct v3d_job *job);
 void v3d_reset(struct v3d_dev *v3d);
 void v3d_invalidate_caches(struct v3d_dev *v3d);
+void v3d_clean_caches(struct v3d_dev *v3d);
 
 /* v3d_irq.c */
 int v3d_irq_init(struct v3d_dev *v3d);

drivers/gpu/drm/v3d/v3d_fence.c

···
 		return "v3d-render";
 	case V3D_TFU:
 		return "v3d-tfu";
+	case V3D_CSD:
+		return "v3d-csd";
 	default:
 		return NULL;
 	}

drivers/gpu/drm/v3d/v3d_gem.c

···
 	/* While there is a busy bit (V3D_L2TCACTL_L2TFLS), we don't
 	 * need to wait for completion before dispatching the job --
 	 * L2T accesses will be stalled until the flush has completed.
+	 * However, we do need to make sure we don't try to trigger a
+	 * new flush while the L2_CLEAN queue is trying to
+	 * synchronously clean after a job.
 	 */
+	mutex_lock(&v3d->cache_clean_lock);
 	V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
 		       V3D_L2TCACTL_L2TFLS |
 		       V3D_SET_FIELD(V3D_L2TCACTL_FLM_FLUSH, V3D_L2TCACTL_FLM));
+	mutex_unlock(&v3d->cache_clean_lock);
+}
+
+/* Cleans texture L1 and L2 cachelines (writing back dirty data).
+ *
+ * For cleaning, which happens from the CACHE_CLEAN queue after CSD has
+ * executed, we need to make sure that the clean is done before
+ * signaling job completion.  So, we synchronously wait before
+ * returning, and we make sure that L2 invalidates don't happen in the
+ * meantime to confuse our are-we-done checks.
+ */
+void
+v3d_clean_caches(struct v3d_dev *v3d)
+{
+	struct drm_device *dev = &v3d->drm;
+	int core = 0;
+
+	trace_v3d_cache_clean_begin(dev);
+
+	V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL, V3D_L2TCACTL_TMUWCF);
+	if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
+		       V3D_L2TCACTL_L2TFLS), 100)) {
+		DRM_ERROR("Timeout waiting for L1T write combiner flush\n");
+	}
+
+	mutex_lock(&v3d->cache_clean_lock);
+	V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
+		       V3D_L2TCACTL_L2TFLS |
+		       V3D_SET_FIELD(V3D_L2TCACTL_FLM_CLEAN, V3D_L2TCACTL_FLM));
+
+	if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
+		       V3D_L2TCACTL_L2TFLS), 100)) {
+		DRM_ERROR("Timeout waiting for L2T clean\n");
+	}
+
+	mutex_unlock(&v3d->cache_clean_lock);
+
+	trace_v3d_cache_clean_end(dev);
 }
 
 /* Invalidates the slice caches.  These are read-only caches. */
···
 v3d_attach_fences_and_unlock_reservation(struct drm_file *file_priv,
 					 struct v3d_job *job,
 					 struct ww_acquire_ctx *acquire_ctx,
-					 u32 out_sync)
+					 u32 out_sync,
+					 struct dma_fence *done_fence)
 {
 	struct drm_syncobj *sync_out;
 	int i;
···
 	/* Update the return sync object for the job */
 	sync_out = drm_syncobj_find(file_priv, out_sync);
 	if (sync_out) {
-		drm_syncobj_replace_fence(sync_out, job->done_fence);
+		drm_syncobj_replace_fence(sync_out, done_fence);
 		drm_syncobj_put(sync_out);
 	}
 }
···
 	mutex_unlock(&v3d->sched_lock);
 
 	v3d_attach_fences_and_unlock_reservation(file_priv,
-						 &render->base, &acquire_ctx,
-						 args->out_sync);
+						 &render->base,
+						 &acquire_ctx,
+						 args->out_sync,
+						 render->base.done_fence);
 
 	if (bin)
 		v3d_job_put(&bin->base);
···
 
 	v3d_attach_fences_and_unlock_reservation(file_priv,
 						 &job->base, &acquire_ctx,
-						 args->out_sync);
+						 args->out_sync,
+						 job->base.done_fence);
 
 	v3d_job_put(&job->base);
 
···
 				    &acquire_ctx);
 fail:
 	v3d_job_put(&job->base);
+
+	return ret;
+}
+
+/**
+ * v3d_submit_csd_ioctl() - Submits a CSD (texture formatting) job to the V3D.
+ * @dev: DRM device
+ * @data: ioctl argument
+ * @file_priv: DRM file for this fd
+ *
+ * Userspace provides the register setup for the CSD, which we don't
+ * need to validate since the CSD is behind the MMU.
+ */
+int
+v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
+		     struct drm_file *file_priv)
+{
+	struct v3d_dev *v3d = to_v3d_dev(dev);
+	struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
+	struct drm_v3d_submit_csd *args = data;
+	struct v3d_csd_job *job;
+	struct v3d_job *clean_job;
+	struct ww_acquire_ctx acquire_ctx;
+	int ret;
+
+	trace_v3d_submit_csd_ioctl(&v3d->drm, args->cfg[5], args->cfg[6]);
+
+	if (!v3d_has_csd(v3d)) {
+		DRM_DEBUG("Attempting CSD submit on non-CSD hardware\n");
+		return -EINVAL;
+	}
+
+	job = kcalloc(1, sizeof(*job), GFP_KERNEL);
+	if (!job)
+		return -ENOMEM;
+
+	ret = v3d_job_init(v3d, file_priv, &job->base,
+			   v3d_job_free, args->in_sync);
+	if (ret) {
+		kfree(job);
+		return ret;
+	}
+
+	clean_job = kcalloc(1, sizeof(*clean_job), GFP_KERNEL);
+	if (!clean_job) {
+		v3d_job_put(&job->base);
+		kfree(job);
+		return -ENOMEM;
+	}
+
+	ret = v3d_job_init(v3d, file_priv, clean_job, v3d_job_free, 0);
+	if (ret) {
+		v3d_job_put(&job->base);
+		kfree(clean_job);
+		return ret;
+	}
+
+	job->args = *args;
+
+	ret = v3d_lookup_bos(dev, file_priv, clean_job,
+			     args->bo_handles, args->bo_handle_count);
+	if (ret)
+		goto fail;
+
+	ret = v3d_lock_bo_reservations(clean_job->base.bo,
+				       clean_job->base.bo_count,
+				       &acquire_ctx);
+	if (ret)
+		goto fail;
+
+	mutex_lock(&v3d->sched_lock);
+	ret = v3d_push_job(v3d_priv, &job->base, V3D_CSD);
+	if (ret)
+		goto fail_unreserve;
+
+	clean_job->in_fence = dma_fence_get(job->base.done_fence);
+	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,
+						 clean_job,
+						 &acquire_ctx,
+						 args->out_sync,
+						 clean_job->done_fence);
+
+	v3d_job_put(&job->base);
+	v3d_job_put(clean_job);
+
+	return 0;
+
+fail_unreserve:
+	mutex_unlock(&v3d->sched_lock);
+	drm_gem_unlock_reservations(clean_job->bo, clean_job->bo_count,
+				    &acquire_ctx);
+fail:
+	v3d_job_put(&job->base);
+	v3d_job_put(clean_job);
 
 	return ret;
 }
···
 	mutex_init(&v3d->bo_lock);
 	mutex_init(&v3d->reset_lock);
 	mutex_init(&v3d->sched_lock);
+	mutex_init(&v3d->cache_clean_lock);
 
 	/* Note: We don't allocate address 0.  Various bits of HW
 	 * treat 0 as special, such as the occlusion query counters

drivers/gpu/drm/v3d/v3d_irq.c

···
 /**
  * DOC: Interrupt management for the V3D engine
  *
- * When we take a bin, render, or TFU done interrupt, we need to
- * signal the fence for that job so that the scheduler can queue up
- * the next one and unblock any waiters.
+ * When we take a bin, render, TFU done, or CSD done interrupt, we
+ * need to signal the fence for that job so that the scheduler can
+ * queue up the next one and unblock any waiters.
  *
  * When we take the binner out of memory interrupt, we need to
  * allocate some new memory and pass it to the binner so that the
···
 #define V3D_CORE_IRQS ((u32)(V3D_INT_OUTOMEM |	\
 			     V3D_INT_FLDONE |	\
 			     V3D_INT_FRDONE |	\
+			     V3D_INT_CSDDONE |	\
 			     V3D_INT_GMPV))
 
 #define V3D_HUB_IRQS ((u32)(V3D_HUB_INT_MMU_WRV |	\
···
 			to_v3d_fence(v3d->render_job->base.irq_fence);
 
 		trace_v3d_rcl_irq(&v3d->drm, fence->seqno);
+		dma_fence_signal(&fence->base);
+		status = IRQ_HANDLED;
+	}
+
+	if (intsts & V3D_INT_CSDDONE) {
+		struct v3d_fence *fence =
+			to_v3d_fence(v3d->csd_job->base.irq_fence);
+
+		trace_v3d_csd_irq(&v3d->drm, fence->seqno);
 		dma_fence_signal(&fence->base);
 		status = IRQ_HANDLED;
 	}

drivers/gpu/drm/v3d/v3d_regs.h

···
 #define V3D_CTL_L2TCACTL                               0x00030
 # define V3D_L2TCACTL_TMUWCF                           BIT(8)
 # define V3D_L2TCACTL_L2T_NO_WM                        BIT(4)
+/* Invalidates cache lines. */
 # define V3D_L2TCACTL_FLM_FLUSH                        0
+/* Removes cachelines without writing dirty lines back. */
 # define V3D_L2TCACTL_FLM_CLEAR                        1
+/* Writes out dirty cachelines and marks them clean, but doesn't invalidate. */
 # define V3D_L2TCACTL_FLM_CLEAN                        2
 # define V3D_L2TCACTL_FLM_MASK                         V3D_MASK(2, 1)
 # define V3D_L2TCACTL_FLM_SHIFT                        1
···
 #define V3D_CTL_INT_MSK_CLR                            0x00064
 # define V3D_INT_QPU_MASK                              V3D_MASK(27, 16)
 # define V3D_INT_QPU_SHIFT                             16
+# define V3D_INT_CSDDONE                               BIT(7)
+# define V3D_INT_PCTR                                  BIT(6)
 # define V3D_INT_GMPV                                  BIT(5)
 # define V3D_INT_TRFB                                  BIT(4)
 # define V3D_INT_SPILLUSE                              BIT(3)
···
 #define V3D_GMP_CLEAR_LOAD                             0x00814
 #define V3D_GMP_PRESERVE_LOAD                          0x00818
 #define V3D_GMP_VALID_LINES                            0x00820
+
+#define V3D_CSD_STATUS                                 0x00900
+# define V3D_CSD_STATUS_NUM_COMPLETED_MASK             V3D_MASK(11, 4)
+# define V3D_CSD_STATUS_NUM_COMPLETED_SHIFT            4
+# define V3D_CSD_STATUS_NUM_ACTIVE_MASK                V3D_MASK(3, 2)
+# define V3D_CSD_STATUS_NUM_ACTIVE_SHIFT               2
+# define V3D_CSD_STATUS_HAVE_CURRENT_DISPATCH          BIT(1)
+# define V3D_CSD_STATUS_HAVE_QUEUED_DISPATCH           BIT(0)
+
+#define V3D_CSD_QUEUED_CFG0                            0x00904
+# define V3D_CSD_QUEUED_CFG0_NUM_WGS_X_MASK            V3D_MASK(31, 16)
+# define V3D_CSD_QUEUED_CFG0_NUM_WGS_X_SHIFT           16
+# define V3D_CSD_QUEUED_CFG0_WG_X_OFFSET_MASK          V3D_MASK(15, 0)
+# define V3D_CSD_QUEUED_CFG0_WG_X_OFFSET_SHIFT         0
+
+#define V3D_CSD_QUEUED_CFG1                            0x00908
+# define V3D_CSD_QUEUED_CFG1_NUM_WGS_Y_MASK            V3D_MASK(31, 16)
+# define V3D_CSD_QUEUED_CFG1_NUM_WGS_Y_SHIFT           16
+# define V3D_CSD_QUEUED_CFG1_WG_Y_OFFSET_MASK          V3D_MASK(15, 0)
+# define V3D_CSD_QUEUED_CFG1_WG_Y_OFFSET_SHIFT         0
+
+#define V3D_CSD_QUEUED_CFG2                            0x0090c
+# define V3D_CSD_QUEUED_CFG2_NUM_WGS_Z_MASK            V3D_MASK(31, 16)
+# define V3D_CSD_QUEUED_CFG2_NUM_WGS_Z_SHIFT           16
+# define V3D_CSD_QUEUED_CFG2_WG_Z_OFFSET_MASK          V3D_MASK(15, 0)
+# define V3D_CSD_QUEUED_CFG2_WG_Z_OFFSET_SHIFT         0
+
+#define V3D_CSD_QUEUED_CFG3                            0x00910
+# define V3D_CSD_QUEUED_CFG3_OVERLAP_WITH_PREV         BIT(26)
+# define V3D_CSD_QUEUED_CFG3_MAX_SG_ID_MASK            V3D_MASK(25, 20)
+# define V3D_CSD_QUEUED_CFG3_MAX_SG_ID_SHIFT           20
+# define V3D_CSD_QUEUED_CFG3_BATCHES_PER_SG_M1_MASK    V3D_MASK(19, 12)
+# define V3D_CSD_QUEUED_CFG3_BATCHES_PER_SG_M1_SHIFT   12
+# define V3D_CSD_QUEUED_CFG3_WGS_PER_SG_MASK           V3D_MASK(11, 8)
+# define V3D_CSD_QUEUED_CFG3_WGS_PER_SG_SHIFT          8
+# define V3D_CSD_QUEUED_CFG3_WG_SIZE_MASK              V3D_MASK(7, 0)
+# define V3D_CSD_QUEUED_CFG3_WG_SIZE_SHIFT             0
+
+/* Number of batches, minus 1 */
+#define V3D_CSD_QUEUED_CFG4                            0x00914
+
+/* Shader address, pnan, singleseg, threading, like a shader record. */
+#define V3D_CSD_QUEUED_CFG5                            0x00918
+
+/* Uniforms address (4 byte aligned) */
+#define V3D_CSD_QUEUED_CFG6                            0x0091c
+
+#define V3D_CSD_CURRENT_CFG0                          0x00920
+#define V3D_CSD_CURRENT_CFG1                          0x00924
+#define V3D_CSD_CURRENT_CFG2                          0x00928
+#define V3D_CSD_CURRENT_CFG3                          0x0092c
+#define V3D_CSD_CURRENT_CFG4                          0x00930
+#define V3D_CSD_CURRENT_CFG5                          0x00934
+#define V3D_CSD_CURRENT_CFG6                          0x00938
+
+#define V3D_CSD_CURRENT_ID0                            0x0093c
+# define V3D_CSD_CURRENT_ID0_WG_X_MASK                 V3D_MASK(31, 16)
+# define V3D_CSD_CURRENT_ID0_WG_X_SHIFT                16
+# define V3D_CSD_CURRENT_ID0_WG_IN_SG_MASK             V3D_MASK(11, 8)
+# define V3D_CSD_CURRENT_ID0_WG_IN_SG_SHIFT            8
+# define V3D_CSD_CURRENT_ID0_L_IDX_MASK                V3D_MASK(7, 0)
+# define V3D_CSD_CURRENT_ID0_L_IDX_SHIFT               0
+
+#define V3D_CSD_CURRENT_ID1                            0x00940
+# define V3D_CSD_CURRENT_ID0_WG_Z_MASK                 V3D_MASK(31, 16)
+# define V3D_CSD_CURRENT_ID0_WG_Z_SHIFT                16
+# define V3D_CSD_CURRENT_ID0_WG_Y_MASK                 V3D_MASK(15, 0)
+# define V3D_CSD_CURRENT_ID0_WG_Y_SHIFT                0
 
 #endif /* V3D_REGS_H */

drivers/gpu/drm/v3d/v3d_sched.c

···
 	return container_of(sched_job, struct v3d_tfu_job, base.base);
 }
 
+static struct v3d_csd_job *
+to_csd_job(struct drm_sched_job *sched_job)
+{
+	return container_of(sched_job, struct v3d_csd_job, base.base);
+}
+
 static void
 v3d_job_free(struct drm_sched_job *sched_job)
 {
···
 	return fence;
 }
 
+static struct dma_fence *
+v3d_csd_job_run(struct drm_sched_job *sched_job)
+{
+	struct v3d_csd_job *job = to_csd_job(sched_job);
+	struct v3d_dev *v3d = job->base.v3d;
+	struct drm_device *dev = &v3d->drm;
+	struct dma_fence *fence;
+	int i;
+
+	v3d->csd_job = job;
+
+	v3d_invalidate_caches(v3d);
+
+	fence = v3d_fence_create(v3d, V3D_CSD);
+	if (IS_ERR(fence))
+		return NULL;
+
+	if (job->base.irq_fence)
+		dma_fence_put(job->base.irq_fence);
+	job->base.irq_fence = dma_fence_get(fence);
+
+	trace_v3d_submit_csd(dev, to_v3d_fence(fence)->seqno);
+
+	for (i = 1; i <= 6; i++)
+		V3D_CORE_WRITE(0, V3D_CSD_QUEUED_CFG0 + 4 * i, job->args.cfg[i]);
+	/* CFG0 write kicks off the job. */
+	V3D_CORE_WRITE(0, V3D_CSD_QUEUED_CFG0, job->args.cfg[0]);
+
+	return fence;
+}
+
+static struct dma_fence *
+v3d_cache_clean_job_run(struct drm_sched_job *sched_job)
+{
+	struct v3d_job *job = to_v3d_job(sched_job);
+	struct v3d_dev *v3d = job->v3d;
+
+	v3d_clean_caches(v3d);
+
+	return NULL;
+}
+
 static void
 v3d_gpu_reset_for_timeout(struct v3d_dev *v3d, struct drm_sched_job *sched_job)
 {
···
 }
 
 static void
-v3d_tfu_job_timedout(struct drm_sched_job *sched_job)
+v3d_generic_job_timedout(struct drm_sched_job *sched_job)
 {
 	struct v3d_job *job = to_v3d_job(sched_job);
 
 	v3d_gpu_reset_for_timeout(job->v3d, sched_job);
+}
+
+static void
+v3d_csd_job_timedout(struct drm_sched_job *sched_job)
+{
+	struct v3d_csd_job *job = to_csd_job(sched_job);
+	struct v3d_dev *v3d = job->base.v3d;
+	u32 batches = V3D_CORE_READ(0, V3D_CSD_CURRENT_CFG4);
+
+	/* If we've made progress, skip reset and let the timer get
+	 * rearmed.
+	 */
+	if (job->timedout_batches != batches) {
+		job->timedout_batches = batches;
+		return;
+	}
+
+	v3d_gpu_reset_for_timeout(v3d, sched_job);
 }
 
 static const struct drm_sched_backend_ops v3d_bin_sched_ops = {
···
 static const struct drm_sched_backend_ops v3d_tfu_sched_ops = {
 	.dependency = v3d_job_dependency,
 	.run_job = v3d_tfu_job_run,
-	.timedout_job = v3d_tfu_job_timedout,
+	.timedout_job = v3d_generic_job_timedout,
 	.free_job = v3d_job_free,
+};
+
+static const struct drm_sched_backend_ops v3d_csd_sched_ops = {
+	.dependency = v3d_job_dependency,
+	.run_job = v3d_csd_job_run,
+	.timedout_job = v3d_csd_job_timedout,
+	.free_job = v3d_job_free
+};
+
+static const struct drm_sched_backend_ops v3d_cache_clean_sched_ops = {
+	.dependency = v3d_job_dependency,
+	.run_job = v3d_cache_clean_job_run,
+	.timedout_job = v3d_generic_job_timedout,
+	.free_job = v3d_job_free
 };
 
 int
···
 	if (ret) {
 		dev_err(v3d->dev, "Failed to create render scheduler: %d.",
 			ret);
-		drm_sched_fini(&v3d->queue[V3D_BIN].sched);
+		v3d_sched_fini(v3d);
 		return ret;
 	}
 
···
 	if (ret) {
 		dev_err(v3d->dev, "Failed to create TFU scheduler: %d.",
 			ret);
-		drm_sched_fini(&v3d->queue[V3D_RENDER].sched);
-		drm_sched_fini(&v3d->queue[V3D_BIN].sched);
+		v3d_sched_fini(v3d);
 		return ret;
+	}
+
+	if (v3d_has_csd(v3d)) {
+		ret = drm_sched_init(&v3d->queue[V3D_CSD].sched,
+				     &v3d_csd_sched_ops,
+				     hw_jobs_limit, job_hang_limit,
+				     msecs_to_jiffies(hang_limit_ms),
+				     "v3d_csd");
+		if (ret) {
+			dev_err(v3d->dev, "Failed to create CSD scheduler: %d.",
+				ret);
+			v3d_sched_fini(v3d);
+			return ret;
+		}
+
+		ret = drm_sched_init(&v3d->queue[V3D_CACHE_CLEAN].sched,
+				     &v3d_cache_clean_sched_ops,
+				     hw_jobs_limit, job_hang_limit,
+				     msecs_to_jiffies(hang_limit_ms),
+				     "v3d_cache_clean");
+		if (ret) {
+			dev_err(v3d->dev, "Failed to create CACHE_CLEAN scheduler: %d.",
+				ret);
+			v3d_sched_fini(v3d);
+			return ret;
+		}
 	}
 
 	return 0;
···
 {
 	enum v3d_queue q;
 
-	for (q = 0; q < V3D_MAX_QUEUES; q++)
-		drm_sched_fini(&v3d->queue[q].sched);
+	for (q = 0; q < V3D_MAX_QUEUES; q++) {
+		if (v3d->queue[q].sched.ready)
+			drm_sched_fini(&v3d->queue[q].sched);
+	}
 }

drivers/gpu/drm/v3d/v3d_trace.h

···
 		      __entry->seqno)
 );
 
+TRACE_EVENT(v3d_csd_irq,
+	    TP_PROTO(struct drm_device *dev,
+		     uint64_t seqno),
+	    TP_ARGS(dev, seqno),
+
+	    TP_STRUCT__entry(
+			     __field(u32, dev)
+			     __field(u64, seqno)
+			     ),
+
+	    TP_fast_assign(
+			   __entry->dev = dev->primary->index;
+			   __entry->seqno = seqno;
+			   ),
+
+	    TP_printk("dev=%u, seqno=%llu",
+		      __entry->dev,
+		      __entry->seqno)
+);
+
 TRACE_EVENT(v3d_submit_tfu_ioctl,
 	    TP_PROTO(struct drm_device *dev, u32 iia),
 	    TP_ARGS(dev, iia),
···
 	    TP_printk("dev=%u, seqno=%llu",
 		      __entry->dev,
 		      __entry->seqno)
+);
+
+TRACE_EVENT(v3d_submit_csd_ioctl,
+	    TP_PROTO(struct drm_device *dev, u32 cfg5, u32 cfg6),
+	    TP_ARGS(dev, cfg5, cfg6),
+
+	    TP_STRUCT__entry(
+			     __field(u32, dev)
+			     __field(u32, cfg5)
+			     __field(u32, cfg6)
+			     ),
+
+	    TP_fast_assign(
+			   __entry->dev = dev->primary->index;
+			   __entry->cfg5 = cfg5;
+			   __entry->cfg6 = cfg6;
+			   ),
+
+	    TP_printk("dev=%u, CFG5 0x%08x, CFG6 0x%08x",
+		      __entry->dev,
+		      __entry->cfg5,
+		      __entry->cfg6)
+);
+
+TRACE_EVENT(v3d_submit_csd,
+	    TP_PROTO(struct drm_device *dev,
+		     uint64_t seqno),
+	    TP_ARGS(dev, seqno),
+
+	    TP_STRUCT__entry(
+			     __field(u32, dev)
+			     __field(u64, seqno)
+			     ),
+
+	    TP_fast_assign(
+			   __entry->dev = dev->primary->index;
+			   __entry->seqno = seqno;
+			   ),
+
+	    TP_printk("dev=%u, seqno=%llu",
+		      __entry->dev,
+		      __entry->seqno)
+);
+
+TRACE_EVENT(v3d_cache_clean_begin,
+	    TP_PROTO(struct drm_device *dev),
+	    TP_ARGS(dev),
+
+	    TP_STRUCT__entry(
+			     __field(u32, dev)
+			     ),
+
+	    TP_fast_assign(
+			   __entry->dev = dev->primary->index;
+			   ),
+
+	    TP_printk("dev=%u",
+		      __entry->dev)
+);
+
+TRACE_EVENT(v3d_cache_clean_end,
+	    TP_PROTO(struct drm_device *dev),
+	    TP_ARGS(dev),
+
+	    TP_STRUCT__entry(
+			     __field(u32, dev)
+			     ),
+
+	    TP_fast_assign(
+			   __entry->dev = dev->primary->index;
+			   ),
+
+	    TP_printk("dev=%u",
+		      __entry->dev)
 );
 
 TRACE_EVENT(v3d_reset_begin,

include/uapi/drm/v3d_drm.h

···
 #define DRM_V3D_GET_PARAM                         0x04
 #define DRM_V3D_GET_BO_OFFSET                     0x05
 #define DRM_V3D_SUBMIT_TFU                        0x06
+#define DRM_V3D_SUBMIT_CSD                        0x07
 
 #define DRM_IOCTL_V3D_SUBMIT_CL           DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_SUBMIT_CL, struct drm_v3d_submit_cl)
 #define DRM_IOCTL_V3D_WAIT_BO             DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_WAIT_BO, struct drm_v3d_wait_bo)
···
 #define DRM_IOCTL_V3D_GET_PARAM           DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_GET_PARAM, struct drm_v3d_get_param)
 #define DRM_IOCTL_V3D_GET_BO_OFFSET       DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_GET_BO_OFFSET, struct drm_v3d_get_bo_offset)
 #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)
 
 /**
  * struct drm_v3d_submit_cl - ioctl argument for submitting commands to the 3D
···
 	DRM_V3D_PARAM_V3D_CORE0_IDENT1,
 	DRM_V3D_PARAM_V3D_CORE0_IDENT2,
 	DRM_V3D_PARAM_SUPPORTS_TFU,
+	DRM_V3D_PARAM_SUPPORTS_CSD,
 };
 
 struct drm_v3d_get_param {
···
 	 */
 	__u32 in_sync;
 	/* Sync object to signal when the TFU job is done. */
+	__u32 out_sync;
+};
+
+/* Submits a compute shader for dispatch.  This job will block on any
+ * previous compute shaders submitted on this fd, and any other
+ * synchronization must be performed with in_sync/out_sync.
+ */
+struct drm_v3d_submit_csd {
+	__u32 cfg[7];
+	__u32 coef[4];
+
+	/* Pointer to a u32 array of the BOs that are referenced by the job.
+	 */
+	__u64 bo_handles;
+
+	/* Number of BO handles passed in (size is that times 4). */
+	__u32 bo_handle_count;
+
+	/* sync object to block on before running the CSD job.  Each
+	 * CSD job will execute in the order submitted to its FD.
+	 * Synchronization against rendering/TFU jobs or CSD from
+	 * other fds requires using sync objects.
+	 */
+	__u32 in_sync;
+	/* Sync object to signal when the CSD job is done. */
 	__u32 out_sync;
 };