openwrt/target/linux/brcm2708/patches-4.19/950-0570-drm-v3d-Add-support-for-submitting-jobs-to-the-TFU.patch

From 60c65dc612663be7136a19a117cee5d194530600 Mon Sep 17 00:00:00 2001
From: Eric Anholt <eric@anholt.net>
Date: Wed, 28 Nov 2018 15:09:25 -0800
Subject: [PATCH 570/703] drm/v3d: Add support for submitting jobs to the TFU.

The TFU can copy from raster, UIF, and SAND input images to UIF output
images, with optional mipmap generation.  This will certainly be
useful for media EGL image input, but is also useful immediately for
mipmap generation without bogging the V3D core down.

For now we only run the queue 1 job deep, and don't have any hang
recovery (though I don't think we should need it, with TFU).  Queuing
multiple jobs in the HW will require synchronizing the YUV coefficient
regs updates since they don't get FIFOed with the job.

v2: Change the ioctl to IOW instead of IOWR, always set COEF0, explain
    why TFU is AUTH, clarify the syncing docs, drop the unused TFU
    interrupt regs (you're expected to use the hub's), don't take
    &bo->base for NULL bos.
v3: Fix a little whitespace alignment (noticed by checkpatch), rebase
    on drm_sched_job_cleanup() changes.

Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Dave Emett <david.emett@broadcom.com> (v2)
Link: https://patchwork.freedesktop.org/patch/264607/
(cherry picked from commit 1584f16ca96ef124aad79efa3303cff5f3530e2c)
---
 drivers/gpu/drm/v3d/v3d_drv.c   |  15 ++-
 drivers/gpu/drm/v3d/v3d_drv.h   |  32 +++++-
 drivers/gpu/drm/v3d/v3d_gem.c   | 178 ++++++++++++++++++++++++++++----
 drivers/gpu/drm/v3d/v3d_irq.c   |  12 ++-
 drivers/gpu/drm/v3d/v3d_regs.h  |  49 +++++++++
 drivers/gpu/drm/v3d/v3d_sched.c | 148 ++++++++++++++++++++++----
 drivers/gpu/drm/v3d/v3d_trace.h |  20 ++++
 include/uapi/drm/v3d_drm.h      |  25 +++++
 8 files changed, 426 insertions(+), 53 deletions(-)

--- a/drivers/gpu/drm/v3d/v3d_drv.c
+++ b/drivers/gpu/drm/v3d/v3d_drv.c
@@ -112,10 +112,15 @@ static int v3d_get_param_ioctl(struct dr
 		return 0;
 	}
 
-	/* Any params that aren't just register reads would go here. */
 
-	DRM_DEBUG("Unknown parameter %d\n", args->param);
-	return -EINVAL;
+	switch (args->param) {
+	case DRM_V3D_PARAM_SUPPORTS_TFU:
+		args->value = 1;
+		return 0;
+	default:
+		DRM_DEBUG("Unknown parameter %d\n", args->param);
+		return -EINVAL;
+	}
 }
 
 static int
@@ -170,7 +175,8 @@ static const struct file_operations v3d_
 /* DRM_AUTH is required on SUBMIT_CL for now, while we don't have GMP
  * protection between clients.  Note that render nodes would be be
  * able to submit CLs that could access BOs from clients authenticated
- * with the master node.
+ * with the master node.  The TFU doesn't use the GMP, so it would
+ * need to stay DRM_AUTH until we do buffer size/offset validation.
  */
 static const struct drm_ioctl_desc v3d_drm_ioctls[] = {
 	DRM_IOCTL_DEF_DRV(V3D_SUBMIT_CL, v3d_submit_cl_ioctl, DRM_RENDER_ALLOW | DRM_AUTH),
@@ -179,6 +185,7 @@ static const struct drm_ioctl_desc v3d_d
 	DRM_IOCTL_DEF_DRV(V3D_MMAP_BO, v3d_mmap_bo_ioctl, DRM_RENDER_ALLOW),
 	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),
 };
 
 static const struct vm_operations_struct v3d_vm_ops = {
--- a/drivers/gpu/drm/v3d/v3d_drv.h
+++ b/drivers/gpu/drm/v3d/v3d_drv.h
@@ -7,19 +7,18 @@
 #include <drm/drm_encoder.h>
 #include <drm/drm_gem.h>
 #include <drm/gpu_scheduler.h>
+#include "uapi/drm/v3d_drm.h"
 
 #define GMP_GRANULARITY (128 * 1024)
 
-/* Enum for each of the V3D queues.  We maintain various queue
- * tracking as an array because at some point we'll want to support
- * the TFU (texture formatting unit) as another queue.
- */
+/* Enum for each of the V3D queues. */
 enum v3d_queue {
 	V3D_BIN,
 	V3D_RENDER,
+	V3D_TFU,
 };
 
-#define V3D_MAX_QUEUES (V3D_RENDER + 1)
+#define V3D_MAX_QUEUES (V3D_TFU + 1)
 
 struct v3d_queue_state {
 	struct drm_gpu_scheduler sched;
@@ -68,6 +67,7 @@ struct v3d_dev {
 
 	struct v3d_exec_info *bin_job;
 	struct v3d_exec_info *render_job;
+	struct v3d_tfu_job *tfu_job;
 
 	struct v3d_queue_state queue[V3D_MAX_QUEUES];
 
@@ -218,6 +218,25 @@ struct v3d_exec_info {
 	u32 qma, qms, qts;
 };
 
+struct v3d_tfu_job {
+	struct drm_sched_job base;
+
+	struct drm_v3d_submit_tfu args;
+
+	/* An optional fence userspace can pass in for the job to depend on. */
+	struct dma_fence *in_fence;
+
+	/* v3d fence to be signaled by IRQ handler when the job is complete. */
+	struct dma_fence *done_fence;
+
+	struct v3d_dev *v3d;
+
+	struct kref refcount;
+
+	/* This is the array of BOs that were looked up at the start of exec. */
+	struct v3d_bo *bo[4];
+};
+
 /**
  * _wait_for - magic (register) wait macro
  *
@@ -281,9 +300,12 @@ int v3d_gem_init(struct drm_device *dev)
 void v3d_gem_destroy(struct drm_device *dev);
 int v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
 			struct drm_file *file_priv);
+int v3d_submit_tfu_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_exec_put(struct v3d_exec_info *exec);
+void v3d_tfu_job_put(struct v3d_tfu_job *exec);
 void v3d_reset(struct v3d_dev *v3d);
 void v3d_invalidate_caches(struct v3d_dev *v3d);
 void v3d_flush_caches(struct v3d_dev *v3d);
--- a/drivers/gpu/drm/v3d/v3d_gem.c
+++ b/drivers/gpu/drm/v3d/v3d_gem.c
@@ -207,26 +207,27 @@ v3d_flush_caches(struct v3d_dev *v3d)
 }
 
 static void
-v3d_attach_object_fences(struct v3d_exec_info *exec)
+v3d_attach_object_fences(struct v3d_bo **bos, int bo_count,
+			 struct dma_fence *fence)
 {
-	struct dma_fence *out_fence = exec->render_done_fence;
 	int i;
 
-	for (i = 0; i < exec->bo_count; i++) {
+	for (i = 0; i < bo_count; i++) {
 		/* XXX: Use shared fences for read-only objects. */
-		reservation_object_add_excl_fence(exec->bo[i]->resv, out_fence);
+		reservation_object_add_excl_fence(bos[i]->resv, fence);
 	}
 }
 
 static void
 v3d_unlock_bo_reservations(struct drm_device *dev,
-			   struct v3d_exec_info *exec,
+			   struct v3d_bo **bos,
+			   int bo_count,
 			   struct ww_acquire_ctx *acquire_ctx)
 {
 	int i;
 
-	for (i = 0; i < exec->bo_count; i++)
-		ww_mutex_unlock(&exec->bo[i]->resv->lock);
+	for (i = 0; i < bo_count; i++)
+		ww_mutex_unlock(&bos[i]->resv->lock);
 
 	ww_acquire_fini(acquire_ctx);
 }
@@ -240,7 +241,8 @@ v3d_unlock_bo_reservations(struct drm_de
  */
 static int
 v3d_lock_bo_reservations(struct drm_device *dev,
-			 struct v3d_exec_info *exec,
+			 struct v3d_bo **bos,
+			 int bo_count,
 			 struct ww_acquire_ctx *acquire_ctx)
 {
 	int contended_lock = -1;
@@ -250,7 +252,7 @@ v3d_lock_bo_reservations(struct drm_devi
 
 retry:
 	if (contended_lock != -1) {
-		struct v3d_bo *bo = exec->bo[contended_lock];
+		struct v3d_bo *bo = bos[contended_lock];
 
 		ret = ww_mutex_lock_slow_interruptible(&bo->resv->lock,
 						       acquire_ctx);
@@ -260,20 +262,20 @@ retry:
 		}
 	}
 
-	for (i = 0; i < exec->bo_count; i++) {
+	for (i = 0; i < bo_count; i++) {
 		if (i == contended_lock)
 			continue;
 
-		ret = ww_mutex_lock_interruptible(&exec->bo[i]->resv->lock,
+		ret = ww_mutex_lock_interruptible(&bos[i]->resv->lock,
 						  acquire_ctx);
 		if (ret) {
 			int j;
 
 			for (j = 0; j < i; j++)
-				ww_mutex_unlock(&exec->bo[j]->resv->lock);
+				ww_mutex_unlock(&bos[j]->resv->lock);
 
 			if (contended_lock != -1 && contended_lock >= i) {
-				struct v3d_bo *bo = exec->bo[contended_lock];
+				struct v3d_bo *bo = bos[contended_lock];
 
 				ww_mutex_unlock(&bo->resv->lock);
 			}
@@ -293,10 +295,11 @@ retry:
 	/* Reserve space for our shared (read-only) fence references,
 	 * before we commit the CL to the hardware.
 	 */
-	for (i = 0; i < exec->bo_count; i++) {
-		ret = reservation_object_reserve_shared(exec->bo[i]->resv);
+	for (i = 0; i < bo_count; i++) {
+		ret = reservation_object_reserve_shared(bos[i]->resv);
 		if (ret) {
-			v3d_unlock_bo_reservations(dev, exec, acquire_ctx);
+			v3d_unlock_bo_reservations(dev, bos, bo_count,
+						   acquire_ctx);
 			return ret;
 		}
 	}
@@ -419,6 +422,33 @@ void v3d_exec_put(struct v3d_exec_info *
 	kref_put(&exec->refcount, v3d_exec_cleanup);
 }
 
+static void
+v3d_tfu_job_cleanup(struct kref *ref)
+{
+	struct v3d_tfu_job *job = container_of(ref, struct v3d_tfu_job,
+					       refcount);
+	struct v3d_dev *v3d = job->v3d;
+	unsigned int i;
+
+	dma_fence_put(job->in_fence);
+	dma_fence_put(job->done_fence);
+
+	for (i = 0; i < ARRAY_SIZE(job->bo); i++) {
+		if (job->bo[i])
+			drm_gem_object_put_unlocked(&job->bo[i]->base);
+	}
+
+	pm_runtime_mark_last_busy(v3d->dev);
+	pm_runtime_put_autosuspend(v3d->dev);
+
+	kfree(job);
+}
+
+void v3d_tfu_job_put(struct v3d_tfu_job *job)
+{
+	kref_put(&job->refcount, v3d_tfu_job_cleanup);
+}
+
 int
 v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
 		  struct drm_file *file_priv)
@@ -536,7 +566,8 @@ v3d_submit_cl_ioctl(struct drm_device *d
 	if (ret)
 		goto fail;
 
-	ret = v3d_lock_bo_reservations(dev, exec, &acquire_ctx);
+	ret = v3d_lock_bo_reservations(dev, exec->bo, exec->bo_count,
+				       &acquire_ctx);
 	if (ret)
 		goto fail;
 
@@ -570,9 +601,10 @@ v3d_submit_cl_ioctl(struct drm_device *d
 				  &v3d_priv->sched_entity[V3D_RENDER]);
 	mutex_unlock(&v3d->sched_lock);
 
-	v3d_attach_object_fences(exec);
+	v3d_attach_object_fences(exec->bo, exec->bo_count,
+				 exec->render_done_fence);
 
-	v3d_unlock_bo_reservations(dev, exec, &acquire_ctx);
+	v3d_unlock_bo_reservations(dev, exec->bo, exec->bo_count, &acquire_ctx);
 
 	/* Update the return sync object for the */
 	sync_out = drm_syncobj_find(file_priv, args->out_sync);
@@ -588,12 +620,118 @@ v3d_submit_cl_ioctl(struct drm_device *d
 
 fail_unreserve:
 	mutex_unlock(&v3d->sched_lock);
-	v3d_unlock_bo_reservations(dev, exec, &acquire_ctx);
+	v3d_unlock_bo_reservations(dev, exec->bo, exec->bo_count, &acquire_ctx);
 fail:
 	v3d_exec_put(exec);
 
 	return ret;
 }
+
+/**
+ * v3d_submit_tfu_ioctl() - Submits a TFU (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 TFU, which we don't
+ * need to validate since the TFU is behind the MMU.
+ */
+int
+v3d_submit_tfu_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_tfu *args = data;
+	struct v3d_tfu_job *job;
+	struct ww_acquire_ctx acquire_ctx;
+	struct drm_syncobj *sync_out;
+	struct dma_fence *sched_done_fence;
+	int ret = 0;
+	int bo_count;
+
+	job = kcalloc(1, sizeof(*job), GFP_KERNEL);
+	if (!job)
+		return -ENOMEM;
+
+	ret = pm_runtime_get_sync(v3d->dev);
+	if (ret < 0) {
+		kfree(job);
+		return ret;
+	}
+
+	kref_init(&job->refcount);
+
+	ret = drm_syncobj_find_fence(file_priv, args->in_sync,
+				     0, &job->in_fence);
+	if (ret == -EINVAL)
+		goto fail;
+
+	job->args = *args;
+	job->v3d = v3d;
+
+	spin_lock(&file_priv->table_lock);
+	for (bo_count = 0; bo_count < ARRAY_SIZE(job->bo); bo_count++) {
+		struct drm_gem_object *bo;
+
+		if (!args->bo_handles[bo_count])
+			break;
+
+		bo = idr_find(&file_priv->object_idr,
+			      args->bo_handles[bo_count]);
+		if (!bo) {
+			DRM_DEBUG("Failed to look up GEM BO %d: %d\n",
+				  bo_count, args->bo_handles[bo_count]);
+			ret = -ENOENT;
+			spin_unlock(&file_priv->table_lock);
+			goto fail;
+		}
+		drm_gem_object_get(bo);
+		job->bo[bo_count] = to_v3d_bo(bo);
+	}
+	spin_unlock(&file_priv->table_lock);
+
+	ret = v3d_lock_bo_reservations(dev, job->bo, bo_count, &acquire_ctx);
+	if (ret)
+		goto fail;
+
+	mutex_lock(&v3d->sched_lock);
+	ret = drm_sched_job_init(&job->base,
+				 &v3d_priv->sched_entity[V3D_TFU],
+				 v3d_priv);
+	if (ret)
+		goto fail_unreserve;
+
+	sched_done_fence = dma_fence_get(&job->base.s_fence->finished);
+
+	kref_get(&job->refcount); /* put by scheduler job completion */
+	drm_sched_entity_push_job(&job->base, &v3d_priv->sched_entity[V3D_TFU]);
+	mutex_unlock(&v3d->sched_lock);
+
+	v3d_attach_object_fences(job->bo, bo_count, sched_done_fence);
+
+	v3d_unlock_bo_reservations(dev, job->bo, bo_count, &acquire_ctx);
+
+	/* Update the return sync object */
+	sync_out = drm_syncobj_find(file_priv, args->out_sync);
+	if (sync_out) {
+		drm_syncobj_replace_fence(sync_out, sched_done_fence);
+		drm_syncobj_put(sync_out);
+	}
+	dma_fence_put(sched_done_fence);
+
+	v3d_tfu_job_put(job);
+
+	return 0;
+
+fail_unreserve:
+	mutex_unlock(&v3d->sched_lock);
+	v3d_unlock_bo_reservations(dev, job->bo, bo_count, &acquire_ctx);
+fail:
+	v3d_tfu_job_put(job);
+
+	return ret;
+}
 
 int
 v3d_gem_init(struct drm_device *dev)
--- a/drivers/gpu/drm/v3d/v3d_irq.c
+++ b/drivers/gpu/drm/v3d/v3d_irq.c
@@ -4,8 +4,8 @@
 /**
  * DOC: Interrupt management for the V3D engine
  *
- * When we take a binning or rendering flush done interrupt, we need
- * to signal the fence for that job so that the scheduler can queue up
+ * 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 the binner out of memory interrupt, we need to
@@ -23,7 +23,8 @@
 
 #define V3D_HUB_IRQS ((u32)(V3D_HUB_INT_MMU_WRV |	\
 			    V3D_HUB_INT_MMU_PTI |	\
-			    V3D_HUB_INT_MMU_CAP))
+			    V3D_HUB_INT_MMU_CAP |	\
+			    V3D_HUB_INT_TFUC))
 
 static void
 v3d_overflow_mem_work(struct work_struct *work)
@@ -117,6 +118,11 @@ v3d_hub_irq(int irq, void *arg)
 	/* Acknowledge the interrupts we're handling here. */
 	V3D_WRITE(V3D_HUB_INT_CLR, intsts);
 
+	if (intsts & V3D_HUB_INT_TFUC) {
+		dma_fence_signal(v3d->tfu_job->done_fence);
+		status = IRQ_HANDLED;
+	}
+
 	if (intsts & (V3D_HUB_INT_MMU_WRV |
 		      V3D_HUB_INT_MMU_PTI |
 		      V3D_HUB_INT_MMU_CAP)) {
--- a/drivers/gpu/drm/v3d/v3d_regs.h
+++ b/drivers/gpu/drm/v3d/v3d_regs.h
@@ -86,6 +86,55 @@
 # define V3D_TOP_GR_BRIDGE_SW_INIT_1                   0x0000c
 # define V3D_TOP_GR_BRIDGE_SW_INIT_1_V3D_CLK_108_SW_INIT BIT(0)
 
+#define V3D_TFU_CS                                     0x00400
+/* Stops current job, empties input fifo. */
+# define V3D_TFU_CS_TFURST                             BIT(31)
+# define V3D_TFU_CS_CVTCT_MASK                         V3D_MASK(23, 16)
+# define V3D_TFU_CS_CVTCT_SHIFT                        16
+# define V3D_TFU_CS_NFREE_MASK                         V3D_MASK(13, 8)
+# define V3D_TFU_CS_NFREE_SHIFT                        8
+# define V3D_TFU_CS_BUSY                               BIT(0)
+
+#define V3D_TFU_SU                                     0x00404
+/* Interrupt when FINTTHR input slots are free (0 = disabled) */
+# define V3D_TFU_SU_FINTTHR_MASK                       V3D_MASK(13, 8)
+# define V3D_TFU_SU_FINTTHR_SHIFT                      8
+/* Skips resetting the CRC at the start of CRC generation. */
+# define V3D_TFU_SU_CRCCHAIN                           BIT(4)
+/* skips writes, computes CRC of the image.  miplevels must be 0. */
+# define V3D_TFU_SU_CRC                                BIT(3)
+# define V3D_TFU_SU_THROTTLE_MASK                      V3D_MASK(1, 0)
+# define V3D_TFU_SU_THROTTLE_SHIFT                     0
+
+#define V3D_TFU_ICFG                                   0x00408
+/* Interrupt when the conversion is complete. */
+# define V3D_TFU_ICFG_IOC                              BIT(0)
+
+/* Input Image Address */
+#define V3D_TFU_IIA                                    0x0040c
+/* Input Chroma Address */
+#define V3D_TFU_ICA                                    0x00410
+/* Input Image Stride */
+#define V3D_TFU_IIS                                    0x00414
+/* Input Image U-Plane Address */
+#define V3D_TFU_IUA                                    0x00418
+/* Output Image Address */
+#define V3D_TFU_IOA                                    0x0041c
+/* Image Output Size */
+#define V3D_TFU_IOS                                    0x00420
+/* TFU YUV Coefficient 0 */
+#define V3D_TFU_COEF0                                  0x00424
+/* Use these regs instead of the defaults. */
+# define V3D_TFU_COEF0_USECOEF                         BIT(31)
+/* TFU YUV Coefficient 1 */
+#define V3D_TFU_COEF1                                  0x00428
+/* TFU YUV Coefficient 2 */
+#define V3D_TFU_COEF2                                  0x0042c
+/* TFU YUV Coefficient 3 */
+#define V3D_TFU_COEF3                                  0x00430
+
+#define V3D_TFU_CRC                                    0x00434
+
 /* Per-MMU registers. */
 
 #define V3D_MMUC_CONTROL                               0x01000
--- a/drivers/gpu/drm/v3d/v3d_sched.c
+++ b/drivers/gpu/drm/v3d/v3d_sched.c
@@ -30,6 +30,12 @@ to_v3d_job(struct drm_sched_job *sched_j
 	return container_of(sched_job, struct v3d_job, base);
 }
 
+static struct v3d_tfu_job *
+to_tfu_job(struct drm_sched_job *sched_job)
+{
+	return container_of(sched_job, struct v3d_tfu_job, base);
+}
+
 static void
 v3d_job_free(struct drm_sched_job *sched_job)
 {
@@ -38,6 +44,14 @@ v3d_job_free(struct drm_sched_job *sched
 	v3d_exec_put(job->exec);
 }
 
+static void
+v3d_tfu_job_free(struct drm_sched_job *sched_job)
+{
+	struct v3d_tfu_job *job = to_tfu_job(sched_job);
+
+	v3d_tfu_job_put(job);
+}
+
 /**
  * Returns the fences that the bin or render job depends on, one by one.
  * v3d_job_run() won't be called until all of them have been signaled.
@@ -76,6 +90,27 @@ v3d_job_dependency(struct drm_sched_job
 	return fence;
 }
 
+/**
+ * Returns the fences that the TFU job depends on, one by one.
+ * v3d_tfu_job_run() won't be called until all of them have been
+ * signaled.
+ */
+static struct dma_fence *
+v3d_tfu_job_dependency(struct drm_sched_job *sched_job,
+		       struct drm_sched_entity *s_entity)
+{
+	struct v3d_tfu_job *job = to_tfu_job(sched_job);
+	struct dma_fence *fence;
+
+	fence = job->in_fence;
+	if (fence) {
+		job->in_fence = NULL;
+		return fence;
+	}
+
+	return NULL;
+}
+
 static struct dma_fence *v3d_job_run(struct drm_sched_job *sched_job)
 {
 	struct v3d_job *job = to_v3d_job(sched_job);
@@ -147,31 +182,47 @@ static struct dma_fence *v3d_job_run(str
 	return fence;
 }
 
-static void
-v3d_job_timedout(struct drm_sched_job *sched_job)
+static struct dma_fence *
+v3d_tfu_job_run(struct drm_sched_job *sched_job)
 {
-	struct v3d_job *job = to_v3d_job(sched_job);
-	struct v3d_exec_info *exec = job->exec;
-	struct v3d_dev *v3d = exec->v3d;
-	enum v3d_queue job_q = job == &exec->bin ? V3D_BIN : V3D_RENDER;
-	enum v3d_queue q;
-	u32 ctca = V3D_CORE_READ(0, V3D_CLE_CTNCA(job_q));
-	u32 ctra = V3D_CORE_READ(0, V3D_CLE_CTNRA(job_q));
+	struct v3d_tfu_job *job = to_tfu_job(sched_job);
+	struct v3d_dev *v3d = job->v3d;
+	struct drm_device *dev = &v3d->drm;
+	struct dma_fence *fence;
 
-	/* If the current address or return address have changed, then
-	 * the GPU has probably made progress and we should delay the
-	 * reset.  This could fail if the GPU got in an infinite loop
-	 * in the CL, but that is pretty unlikely outside of an i-g-t
-	 * testcase.
-	 */
-	if (job->timedout_ctca != ctca || job->timedout_ctra != ctra) {
-		job->timedout_ctca = ctca;
-		job->timedout_ctra = ctra;
+	fence = v3d_fence_create(v3d, V3D_TFU);
+	if (IS_ERR(fence))
+		return NULL;
 
-		schedule_delayed_work(&job->base.work_tdr,
-				      job->base.sched->timeout);
-		return;
+	v3d->tfu_job = job;
+	if (job->done_fence)
+		dma_fence_put(job->done_fence);
+	job->done_fence = dma_fence_get(fence);
+
+	trace_v3d_submit_tfu(dev, to_v3d_fence(fence)->seqno);
+
+	V3D_WRITE(V3D_TFU_IIA, job->args.iia);
+	V3D_WRITE(V3D_TFU_IIS, job->args.iis);
+	V3D_WRITE(V3D_TFU_ICA, job->args.ica);
+	V3D_WRITE(V3D_TFU_IUA, job->args.iua);
+	V3D_WRITE(V3D_TFU_IOA, job->args.ioa);
+	V3D_WRITE(V3D_TFU_IOS, job->args.ios);
+	V3D_WRITE(V3D_TFU_COEF0, job->args.coef[0]);
+	if (job->args.coef[0] & V3D_TFU_COEF0_USECOEF) {
+		V3D_WRITE(V3D_TFU_COEF1, job->args.coef[1]);
+		V3D_WRITE(V3D_TFU_COEF2, job->args.coef[2]);
+		V3D_WRITE(V3D_TFU_COEF3, job->args.coef[3]);
 	}
+	/* ICFG kicks off the job. */
+	V3D_WRITE(V3D_TFU_ICFG, job->args.icfg | V3D_TFU_ICFG_IOC);
+
+	return fence;
+}
+
+static void
+v3d_gpu_reset_for_timeout(struct v3d_dev *v3d, struct drm_sched_job *sched_job)
+{
+	enum v3d_queue q;
 
 	mutex_lock(&v3d->reset_lock);
 
@@ -196,6 +247,41 @@ v3d_job_timedout(struct drm_sched_job *s
 	mutex_unlock(&v3d->reset_lock);
 }
 
+static void
+v3d_job_timedout(struct drm_sched_job *sched_job)
+{
+	struct v3d_job *job = to_v3d_job(sched_job);
+	struct v3d_exec_info *exec = job->exec;
+	struct v3d_dev *v3d = exec->v3d;
+	enum v3d_queue job_q = job == &exec->bin ? V3D_BIN : V3D_RENDER;
+	u32 ctca = V3D_CORE_READ(0, V3D_CLE_CTNCA(job_q));
+	u32 ctra = V3D_CORE_READ(0, V3D_CLE_CTNRA(job_q));
+
+	/* If the current address or return address have changed, then
+	 * the GPU has probably made progress and we should delay the
+	 * reset.  This could fail if the GPU got in an infinite loop
+	 * in the CL, but that is pretty unlikely outside of an i-g-t
+	 * testcase.
+	 */
+	if (job->timedout_ctca != ctca || job->timedout_ctra != ctra) {
+		job->timedout_ctca = ctca;
+		job->timedout_ctra = ctra;
+		schedule_delayed_work(&job->base.work_tdr,
+				      job->base.sched->timeout);
+		return;
+	}
+
+	v3d_gpu_reset_for_timeout(v3d, sched_job);
+}
+
+static void
+v3d_tfu_job_timedout(struct drm_sched_job *sched_job)
+{
+	struct v3d_tfu_job *job = to_tfu_job(sched_job);
+
+	v3d_gpu_reset_for_timeout(job->v3d, sched_job);
+}
+
 static const struct drm_sched_backend_ops v3d_sched_ops = {
 	.dependency = v3d_job_dependency,
 	.run_job = v3d_job_run,
@@ -203,6 +289,13 @@ static const struct drm_sched_backend_op
 	.free_job = v3d_job_free
 };
 
+static const struct drm_sched_backend_ops v3d_tfu_sched_ops = {
+	.dependency = v3d_tfu_job_dependency,
+	.run_job = v3d_tfu_job_run,
+	.timedout_job = v3d_tfu_job_timedout,
+	.free_job = v3d_tfu_job_free
+};
+
 int
 v3d_sched_init(struct v3d_dev *v3d)
 {
@@ -232,6 +325,19 @@ v3d_sched_init(struct v3d_dev *v3d)
 		drm_sched_fini(&v3d->queue[V3D_BIN].sched);
 		return ret;
 	}
+
+	ret = drm_sched_init(&v3d->queue[V3D_TFU].sched,
+			     &v3d_tfu_sched_ops,
+			     hw_jobs_limit, job_hang_limit,
+			     msecs_to_jiffies(hang_limit_ms),
+			     "v3d_tfu");
+	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);
+		return ret;
+	}
 
 	return 0;
 }
--- a/drivers/gpu/drm/v3d/v3d_trace.h
+++ b/drivers/gpu/drm/v3d/v3d_trace.h
@@ -42,6 +42,26 @@ TRACE_EVENT(v3d_submit_cl,
 		      __entry->ctnqea)
 );
 
+TRACE_EVENT(v3d_submit_tfu,
+	    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_reset_begin,
 	    TP_PROTO(struct drm_device *dev),
 	    TP_ARGS(dev),
--- a/include/uapi/drm/v3d_drm.h
+++ b/include/uapi/drm/v3d_drm.h
@@ -36,6 +36,7 @@ extern "C" {
 #define DRM_V3D_MMAP_BO                           0x03
 #define DRM_V3D_GET_PARAM                         0x04
 #define DRM_V3D_GET_BO_OFFSET                     0x05
+#define DRM_V3D_SUBMIT_TFU                        0x06
 
 #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)
@@ -43,6 +44,7 @@ extern "C" {
 #define DRM_IOCTL_V3D_MMAP_BO             DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_MMAP_BO, struct drm_v3d_mmap_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)
 
 /**
  * struct drm_v3d_submit_cl - ioctl argument for submitting commands to the 3D
@@ -169,6 +171,7 @@ enum drm_v3d_param {
 	DRM_V3D_PARAM_V3D_CORE0_IDENT0,
 	DRM_V3D_PARAM_V3D_CORE0_IDENT1,
 	DRM_V3D_PARAM_V3D_CORE0_IDENT2,
+	DRM_V3D_PARAM_SUPPORTS_TFU,
 };
 
 struct drm_v3d_get_param {
@@ -187,6 +190,28 @@ struct drm_v3d_get_bo_offset {
 	__u32 offset;
 };
 
+struct drm_v3d_submit_tfu {
+	__u32 icfg;
+	__u32 iia;
+	__u32 iis;
+	__u32 ica;
+	__u32 iua;
+	__u32 ioa;
+	__u32 ios;
+	__u32 coef[4];
+	/* First handle is the output BO, following are other inputs.
+	 * 0 for unused.
+	 */
+	__u32 bo_handles[4];
+	/* sync object to block on before running the TFU job.  Each TFU
+	 * job will execute in the order submitted to its FD.  Synchronization
+	 * against rendering jobs requires using sync objects.
+	 */
+	__u32 in_sync;
+	/* Sync object to signal when the TFU job is done. */
+	__u32 out_sync;
+};
+
 #if defined(__cplusplus)
 }
 #endif