drm/i915/guc: Connect UAPI to GuC multi-lrc interface

Introduce 'set parallel submit' extension to connect UAPI to GuC
multi-lrc interface. Kernel doc in new uAPI should explain it all.

IGT: https://patchwork.freedesktop.org/patch/447008/?series=93071&rev=1
media UMD: https://github.com/intel/media-driver/pull/1252

v2:
(Daniel Vetter)
- Add IGT link and placeholder for media UMD link
v3:
(Kernel test robot)
- Fix warning in unpin engines call
(John Harrison)
- Reword a bunch of the kernel doc
v4:
(John Harrison)
- Add comment why perma-pin is done after setting gem context
- Update some comments / docs for proto contexts
v5:
(John Harrison)
- Rework perma-pin comment
- Add BUG_IN if context is pinned when setting gem context

Cc: Tvrtko Ursulin <[e-mail]>
Signed-off-by: Matthew Brost <[e-mail]>
Reviewed-by: John Harrison <[e-mail]>
Signed-off-by: John Harrison <[e-mail]>
Link: https://patchwork.freedesktop.org/patch/msgid/[e-mail]

drivers/gpu/drm/i915/gem/i915_gem_context.c

···
 	return 0;
 }
 
+static int
+set_proto_ctx_engines_parallel_submit(struct i915_user_extension __user *base,
+				      void *data)
+{
+	struct i915_context_engines_parallel_submit __user *ext =
+		container_of_user(base, typeof(*ext), base);
+	const struct set_proto_ctx_engines *set = data;
+	struct drm_i915_private *i915 = set->i915;
+	u64 flags;
+	int err = 0, n, i, j;
+	u16 slot, width, num_siblings;
+	struct intel_engine_cs **siblings = NULL;
+	intel_engine_mask_t prev_mask;
+
+	/* Disabling for now */
+	return -ENODEV;
+
+	/* FIXME: This is NIY for execlists */
+	if (!(intel_uc_uses_guc_submission(&i915->gt.uc)))
+		return -ENODEV;
+
+	if (get_user(slot, &ext->engine_index))
+		return -EFAULT;
+
+	if (get_user(width, &ext->width))
+		return -EFAULT;
+
+	if (get_user(num_siblings, &ext->num_siblings))
+		return -EFAULT;
+
+	if (slot >= set->num_engines) {
+		drm_dbg(&i915->drm, "Invalid placement value, %d >= %d\n",
+			slot, set->num_engines);
+		return -EINVAL;
+	}
+
+	if (set->engines[slot].type != I915_GEM_ENGINE_TYPE_INVALID) {
+		drm_dbg(&i915->drm,
+			"Invalid placement[%d], already occupied\n", slot);
+		return -EINVAL;
+	}
+
+	if (get_user(flags, &ext->flags))
+		return -EFAULT;
+
+	if (flags) {
+		drm_dbg(&i915->drm, "Unknown flags 0x%02llx", flags);
+		return -EINVAL;
+	}
+
+	for (n = 0; n < ARRAY_SIZE(ext->mbz64); n++) {
+		err = check_user_mbz(&ext->mbz64[n]);
+		if (err)
+			return err;
+	}
+
+	if (width < 2) {
+		drm_dbg(&i915->drm, "Width (%d) < 2\n", width);
+		return -EINVAL;
+	}
+
+	if (num_siblings < 1) {
+		drm_dbg(&i915->drm, "Number siblings (%d) < 1\n",
+			num_siblings);
+		return -EINVAL;
+	}
+
+	siblings = kmalloc_array(num_siblings * width,
+				 sizeof(*siblings),
+				 GFP_KERNEL);
+	if (!siblings)
+		return -ENOMEM;
+
+	/* Create contexts / engines */
+	for (i = 0; i < width; ++i) {
+		intel_engine_mask_t current_mask = 0;
+		struct i915_engine_class_instance prev_engine;
+
+		for (j = 0; j < num_siblings; ++j) {
+			struct i915_engine_class_instance ci;
+
+			n = i * num_siblings + j;
+			if (copy_from_user(&ci, &ext->engines[n], sizeof(ci))) {
+				err = -EFAULT;
+				goto out_err;
+			}
+
+			siblings[n] =
+				intel_engine_lookup_user(i915, ci.engine_class,
+							 ci.engine_instance);
+			if (!siblings[n]) {
+				drm_dbg(&i915->drm,
+					"Invalid sibling[%d]: { class:%d, inst:%d }\n",
+					n, ci.engine_class, ci.engine_instance);
+				err = -EINVAL;
+				goto out_err;
+			}
+
+			if (n) {
+				if (prev_engine.engine_class !=
+				    ci.engine_class) {
+					drm_dbg(&i915->drm,
+						"Mismatched class %d, %d\n",
+						prev_engine.engine_class,
+						ci.engine_class);
+					err = -EINVAL;
+					goto out_err;
+				}
+			}
+
+			prev_engine = ci;
+			current_mask |= siblings[n]->logical_mask;
+		}
+
+		if (i > 0) {
+			if (current_mask != prev_mask << 1) {
+				drm_dbg(&i915->drm,
+					"Non contiguous logical mask 0x%x, 0x%x\n",
+					prev_mask, current_mask);
+				err = -EINVAL;
+				goto out_err;
+			}
+		}
+		prev_mask = current_mask;
+	}
+
+	set->engines[slot].type = I915_GEM_ENGINE_TYPE_PARALLEL;
+	set->engines[slot].num_siblings = num_siblings;
+	set->engines[slot].width = width;
+	set->engines[slot].siblings = siblings;
+
+	return 0;
+
+out_err:
+	kfree(siblings);
+
+	return err;
+}
+
 static const i915_user_extension_fn set_proto_ctx_engines_extensions[] = {
 	[I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE] = set_proto_ctx_engines_balance,
 	[I915_CONTEXT_ENGINES_EXT_BOND] = set_proto_ctx_engines_bond,
+	[I915_CONTEXT_ENGINES_EXT_PARALLEL_SUBMIT] =
+		set_proto_ctx_engines_parallel_submit,
 };
 
 static int set_proto_ctx_engines(struct drm_i915_file_private *fpriv,
···
 	GEM_BUG_ON(rcu_access_pointer(ce->gem_context));
 	RCU_INIT_POINTER(ce->gem_context, ctx);
 
+	GEM_BUG_ON(intel_context_is_pinned(ce));
 	ce->ring_size = SZ_16K;
 
 	i915_vm_put(ce->vm);
···
 		ret = intel_context_reconfigure_sseu(ce, sseu);
 
 	return ret;
+}
+
+static void __unpin_engines(struct i915_gem_engines *e, unsigned int count)
+{
+	while (count--) {
+		struct intel_context *ce = e->engines[count], *child;
+
+		if (!ce || !test_bit(CONTEXT_PERMA_PIN, &ce->flags))
+			continue;
+
+		for_each_child(ce, child)
+			intel_context_unpin(child);
+		intel_context_unpin(ce);
+	}
+}
+
+static void unpin_engines(struct i915_gem_engines *e)
+{
+	__unpin_engines(e, e->num_engines);
 }
 
 static void __free_engines(struct i915_gem_engines *e, unsigned int count)
···
 	return err;
 }
 
+static int perma_pin_contexts(struct intel_context *ce)
+{
+	struct intel_context *child;
+	int i = 0, j = 0, ret;
+
+	GEM_BUG_ON(!intel_context_is_parent(ce));
+
+	ret = intel_context_pin(ce);
+	if (unlikely(ret))
+		return ret;
+
+	for_each_child(ce, child) {
+		ret = intel_context_pin(child);
+		if (unlikely(ret))
+			goto unwind;
+		++i;
+	}
+
+	set_bit(CONTEXT_PERMA_PIN, &ce->flags);
+
+	return 0;
+
+unwind:
+	intel_context_unpin(ce);
+	for_each_child(ce, child) {
+		if (j++ < i)
+			intel_context_unpin(child);
+		else
+			break;
+	}
+
+	return ret;
+}
+
 static struct i915_gem_engines *user_engines(struct i915_gem_context *ctx,
 					     unsigned int num_engines,
 					     struct i915_gem_proto_engine *pe)
···
 	e->num_engines = num_engines;
 
 	for (n = 0; n < num_engines; n++) {
-		struct intel_context *ce;
+		struct intel_context *ce, *child;
 		int ret;
 
 		switch (pe[n].type) {
···
 
 		case I915_GEM_ENGINE_TYPE_BALANCED:
 			ce = intel_engine_create_virtual(pe[n].siblings,
-							 pe[n].num_siblings);
+							 pe[n].num_siblings, 0);
+			break;
+
+		case I915_GEM_ENGINE_TYPE_PARALLEL:
+			ce = intel_engine_create_parallel(pe[n].siblings,
+							  pe[n].num_siblings,
+							  pe[n].width);
 			break;
 
 		case I915_GEM_ENGINE_TYPE_INVALID:
···
 		if (ret) {
 			err = ERR_PTR(ret);
 			goto free_engines;
+		}
+		for_each_child(ce, child) {
+			ret = intel_context_set_gem(child, ctx, pe->sseu);
+			if (ret) {
+				err = ERR_PTR(ret);
+				goto free_engines;
+			}
+		}
+
+		/*
+		 * XXX: Must be done after calling intel_context_set_gem as that
+		 * function changes the ring size. The ring is allocated when
+		 * the context is pinned. If the ring size is changed after
+		 * allocation we have a mismatch of the ring size and will cause
+		 * the context to hang. Presumably with a bit of reordering we
+		 * could move the perma-pin step to the backend function
+		 * intel_engine_create_parallel.
+		 */
+		if (pe[n].type == I915_GEM_ENGINE_TYPE_PARALLEL) {
+			ret = perma_pin_contexts(ce);
+			if (ret) {
+				err = ERR_PTR(ret);
+				goto free_engines;
+			}
 		}
 	}
 
···
 
 	/* Flush any concurrent set_engines() */
 	mutex_lock(&ctx->engines_mutex);
+	unpin_engines(__context_engines_static(ctx));
 	engines_idle_release(ctx, rcu_replace_pointer(ctx->engines, NULL, 1));
 	i915_gem_context_set_closed(ctx);
 	mutex_unlock(&ctx->engines_mutex);

drivers/gpu/drm/i915/gem/i915_gem_context_types.h

···
 
 	/** @I915_GEM_ENGINE_TYPE_BALANCED: A load-balanced engine set */
 	I915_GEM_ENGINE_TYPE_BALANCED,
+
+	/** @I915_GEM_ENGINE_TYPE_PARALLEL: A parallel engine set */
+	I915_GEM_ENGINE_TYPE_PARALLEL,
 };
 
 /**
  * struct i915_gem_proto_engine - prototype engine
  *
  * This struct describes an engine that a context may contain.  Engines
- * have three types:
+ * have four types:
  *
  *  - I915_GEM_ENGINE_TYPE_INVALID: Invalid engines can be created but they
  *    show up as a NULL in i915_gem_engines::engines[i] and any attempt to
···
  *
  *  - I915_GEM_ENGINE_TYPE_BALANCED: A load-balanced engine set, described
  *    i915_gem_proto_engine::num_siblings and i915_gem_proto_engine::siblings.
+ *
+ *  - I915_GEM_ENGINE_TYPE_PARALLEL: A parallel submission engine set, described
+ *    i915_gem_proto_engine::width, i915_gem_proto_engine::num_siblings, and
+ *    i915_gem_proto_engine::siblings.
  */
 struct i915_gem_proto_engine {
 	/** @type: Type of this engine */
···
 	/** @engine: Engine, for physical */
 	struct intel_engine_cs *engine;
 
-	/** @num_siblings: Number of balanced siblings */
+	/** @num_siblings: Number of balanced or parallel siblings */
 	unsigned int num_siblings;
 
-	/** @siblings: Balanced siblings */
+	/** @width: Width of each sibling */
+	unsigned int width;
+
+	/** @siblings: Balanced siblings or num_siblings * width for parallel */
 	struct intel_engine_cs **siblings;
 
 	/** @sseu: Client-set SSEU parameters */

drivers/gpu/drm/i915/gt/intel_context_types.h

···
 	void (*reset)(struct intel_context *ce);
 	void (*destroy)(struct kref *kref);
 
-	/* virtual engine/context interface */
+	/* virtual/parallel engine/context interface */
 	struct intel_context *(*create_virtual)(struct intel_engine_cs **engine,
-						unsigned int count);
+						unsigned int count,
+						unsigned long flags);
+	struct intel_context *(*create_parallel)(struct intel_engine_cs **engines,
+						 unsigned int num_siblings,
+						 unsigned int width);
 	struct intel_engine_cs *(*get_sibling)(struct intel_engine_cs *engine,
 					       unsigned int sibling);
 };
···
 #define CONTEXT_NOPREEMPT		8
 #define CONTEXT_LRCA_DIRTY		9
 #define CONTEXT_GUC_INIT		10
+#define CONTEXT_PERMA_PIN		11
 
 	struct {
 		u64 timeout_us;

drivers/gpu/drm/i915/gt/intel_engine.h

···
 	return intel_engine_has_preemption(engine);
 }
 
+#define FORCE_VIRTUAL	BIT(0)
 struct intel_context *
 intel_engine_create_virtual(struct intel_engine_cs **siblings,
-			    unsigned int count);
+			    unsigned int count, unsigned long flags);
+
+static inline struct intel_context *
+intel_engine_create_parallel(struct intel_engine_cs **engines,
+			     unsigned int num_engines,
+			     unsigned int width)
+{
+	GEM_BUG_ON(!engines[0]->cops->create_parallel);
+	return engines[0]->cops->create_parallel(engines, num_engines, width);
+}
 
 static inline bool
 intel_virtual_engine_has_heartbeat(const struct intel_engine_cs *engine)

drivers/gpu/drm/i915/gt/intel_engine_cs.c

···
 
 struct intel_context *
 intel_engine_create_virtual(struct intel_engine_cs **siblings,
-			    unsigned int count)
+			    unsigned int count, unsigned long flags)
 {
 	if (count == 0)
 		return ERR_PTR(-EINVAL);
 
-	if (count == 1)
+	if (count == 1 && !(flags & FORCE_VIRTUAL))
 		return intel_context_create(siblings[0]);
 
 	GEM_BUG_ON(!siblings[0]->cops->create_virtual);
-	return siblings[0]->cops->create_virtual(siblings, count);
+	return siblings[0]->cops->create_virtual(siblings, count, flags);
 }
 
 struct i915_request *

drivers/gpu/drm/i915/gt/intel_execlists_submission.c

···
 }
 
 static struct intel_context *
-execlists_create_virtual(struct intel_engine_cs **siblings, unsigned int count);
+execlists_create_virtual(struct intel_engine_cs **siblings, unsigned int count,
+			 unsigned long flags);
 
 static struct i915_request *
 __active_request(const struct intel_timeline * const tl,
···
 }
 
 static struct intel_context *
-execlists_create_virtual(struct intel_engine_cs **siblings, unsigned int count)
+execlists_create_virtual(struct intel_engine_cs **siblings, unsigned int count,
+			 unsigned long flags)
 {
 	struct virtual_engine *ve;
 	unsigned int n;

drivers/gpu/drm/i915/gt/selftest_execlists.c

···
 	GEM_BUG_ON(!nctx || nctx > ARRAY_SIZE(ve));
 
 	for (n = 0; n < nctx; n++) {
-		ve[n] = intel_engine_create_virtual(siblings, nsibling);
+		ve[n] = intel_engine_create_virtual(siblings, nsibling, 0);
 		if (IS_ERR(ve[n])) {
 			err = PTR_ERR(ve[n]);
 			nctx = n;
···
 	 * restrict it to our desired engine within the virtual engine.
 	 */
 
-	ve = intel_engine_create_virtual(siblings, nsibling);
+	ve = intel_engine_create_virtual(siblings, nsibling, 0);
 	if (IS_ERR(ve)) {
 		err = PTR_ERR(ve);
 		goto out_close;
···
 		i915_request_add(rq);
 	}
 
-	ce = intel_engine_create_virtual(siblings, nsibling);
+	ce = intel_engine_create_virtual(siblings, nsibling, 0);
 	if (IS_ERR(ce)) {
 		err = PTR_ERR(ce);
 		goto out;
···
 
 	/* XXX We do not handle oversubscription and fairness with normal rq */
 	for (n = 0; n < nsibling; n++) {
-		ce = intel_engine_create_virtual(siblings, nsibling);
+		ce = intel_engine_create_virtual(siblings, nsibling, 0);
 		if (IS_ERR(ce)) {
 			err = PTR_ERR(ce);
 			goto out;
···
 	if (err)
 		goto out_scratch;
 
-	ve = intel_engine_create_virtual(siblings, nsibling);
+	ve = intel_engine_create_virtual(siblings, nsibling, 0);
 	if (IS_ERR(ve)) {
 		err = PTR_ERR(ve);
 		goto out_scratch;
···
 	if (igt_spinner_init(&spin, gt))
 		return -ENOMEM;
 
-	ve = intel_engine_create_virtual(siblings, nsibling);
+	ve = intel_engine_create_virtual(siblings, nsibling, 0);
 	if (IS_ERR(ve)) {
 		err = PTR_ERR(ve);
 		goto out_spin;

drivers/gpu/drm/i915/gt/uc/intel_guc_submission.c

···
 };
 
 static struct intel_context *
-guc_create_virtual(struct intel_engine_cs **siblings, unsigned int count);
+guc_create_virtual(struct intel_engine_cs **siblings, unsigned int count,
+		   unsigned long flags);
+
+static struct intel_context *
+guc_create_parallel(struct intel_engine_cs **engines,
+		    unsigned int num_siblings,
+		    unsigned int width);
 
 #define GUC_REQUEST_SIZE 64 /* bytes */
 
···
 	.destroy = guc_context_destroy,
 
 	.create_virtual = guc_create_virtual,
+	.create_parallel = guc_create_parallel,
 };
 
 static void submit_work_cb(struct irq_work *wrk)
···
 	.get_sibling = guc_virtual_get_sibling,
 };
 
-/* Future patches will use this function */
-__maybe_unused
 static int guc_parent_context_pin(struct intel_context *ce, void *vaddr)
 {
 	struct intel_engine_cs *engine = guc_virtual_get_sibling(ce->engine, 0);
···
 	return __guc_context_pin(ce, engine, vaddr);
 }
 
-/* Future patches will use this function */
-__maybe_unused
 static int guc_child_context_pin(struct intel_context *ce, void *vaddr)
 {
 	struct intel_engine_cs *engine = guc_virtual_get_sibling(ce->engine, 0);
···
 	return __guc_context_pin(ce, engine, vaddr);
 }
 
-/* Future patches will use this function */
-__maybe_unused
 static void guc_parent_context_unpin(struct intel_context *ce)
 {
 	struct intel_guc *guc = ce_to_guc(ce);
···
 	lrc_unpin(ce);
 }
 
-/* Future patches will use this function */
-__maybe_unused
 static void guc_child_context_unpin(struct intel_context *ce)
 {
 	GEM_BUG_ON(context_enabled(ce));
···
 	lrc_unpin(ce);
 }
 
-/* Future patches will use this function */
-__maybe_unused
 static void guc_child_context_post_unpin(struct intel_context *ce)
 {
 	GEM_BUG_ON(!intel_context_is_child(ce));
···
 
 	lrc_post_unpin(ce);
 	intel_context_unpin(ce->parallel.parent);
+}
+
+static void guc_child_context_destroy(struct kref *kref)
+{
+	struct intel_context *ce = container_of(kref, typeof(*ce), ref);
+
+	__guc_context_destroy(ce);
+}
+
+static const struct intel_context_ops virtual_parent_context_ops = {
+	.alloc = guc_virtual_context_alloc,
+
+	.pre_pin = guc_context_pre_pin,
+	.pin = guc_parent_context_pin,
+	.unpin = guc_parent_context_unpin,
+	.post_unpin = guc_context_post_unpin,
+
+	.ban = guc_context_ban,
+
+	.cancel_request = guc_context_cancel_request,
+
+	.enter = guc_virtual_context_enter,
+	.exit = guc_virtual_context_exit,
+
+	.sched_disable = guc_context_sched_disable,
+
+	.destroy = guc_context_destroy,
+
+	.get_sibling = guc_virtual_get_sibling,
+};
+
+static const struct intel_context_ops virtual_child_context_ops = {
+	.alloc = guc_virtual_context_alloc,
+
+	.pre_pin = guc_context_pre_pin,
+	.pin = guc_child_context_pin,
+	.unpin = guc_child_context_unpin,
+	.post_unpin = guc_child_context_post_unpin,
+
+	.cancel_request = guc_context_cancel_request,
+
+	.enter = guc_virtual_context_enter,
+	.exit = guc_virtual_context_exit,
+
+	.destroy = guc_child_context_destroy,
+
+	.get_sibling = guc_virtual_get_sibling,
+};
+
+static struct intel_context *
+guc_create_parallel(struct intel_engine_cs **engines,
+		    unsigned int num_siblings,
+		    unsigned int width)
+{
+	struct intel_engine_cs **siblings = NULL;
+	struct intel_context *parent = NULL, *ce, *err;
+	int i, j;
+
+	siblings = kmalloc_array(num_siblings,
+				 sizeof(*siblings),
+				 GFP_KERNEL);
+	if (!siblings)
+		return ERR_PTR(-ENOMEM);
+
+	for (i = 0; i < width; ++i) {
+		for (j = 0; j < num_siblings; ++j)
+			siblings[j] = engines[i * num_siblings + j];
+
+		ce = intel_engine_create_virtual(siblings, num_siblings,
+						 FORCE_VIRTUAL);
+		if (!ce) {
+			err = ERR_PTR(-ENOMEM);
+			goto unwind;
+		}
+
+		if (i == 0) {
+			parent = ce;
+			parent->ops = &virtual_parent_context_ops;
+		} else {
+			ce->ops = &virtual_child_context_ops;
+			intel_context_bind_parent_child(parent, ce);
+		}
+	}
+
+	kfree(siblings);
+	return parent;
+
+unwind:
+	if (parent)
+		intel_context_put(parent);
+	kfree(siblings);
+	return err;
 }
 
 static bool
···
 }
 
 static struct intel_context *
-guc_create_virtual(struct intel_engine_cs **siblings, unsigned int count)
+guc_create_virtual(struct intel_engine_cs **siblings, unsigned int count,
+		   unsigned long flags)
 {
 	struct guc_virtual_engine *ve;
 	struct intel_guc *guc;

include/uapi/drm/i915_drm.h

···
  * Extensions:
  *   i915_context_engines_load_balance (I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE)
  *   i915_context_engines_bond (I915_CONTEXT_ENGINES_EXT_BOND)
+ *   i915_context_engines_parallel_submit (I915_CONTEXT_ENGINES_EXT_PARALLEL_SUBMIT)
  */
 #define I915_CONTEXT_PARAM_ENGINES	0xa
 
···
 } __attribute__((packed)) name__
 
 /**
+ * struct i915_context_engines_parallel_submit - Configure engine for
+ * parallel submission.
+ *
+ * Setup a slot in the context engine map to allow multiple BBs to be submitted
+ * in a single execbuf IOCTL. Those BBs will then be scheduled to run on the GPU
+ * in parallel. Multiple hardware contexts are created internally in the i915 to
+ * run these BBs. Once a slot is configured for N BBs only N BBs can be
+ * submitted in each execbuf IOCTL and this is implicit behavior e.g. The user
+ * doesn't tell the execbuf IOCTL there are N BBs, the execbuf IOCTL knows how
+ * many BBs there are based on the slot's configuration. The N BBs are the last
+ * N buffer objects or first N if I915_EXEC_BATCH_FIRST is set.
+ *
+ * The default placement behavior is to create implicit bonds between each
+ * context if each context maps to more than 1 physical engine (e.g. context is
+ * a virtual engine). Also we only allow contexts of same engine class and these
+ * contexts must be in logically contiguous order. Examples of the placement
+ * behavior are described below. Lastly, the default is to not allow BBs to be
+ * preempted mid-batch. Rather insert coordinated preemption points on all
+ * hardware contexts between each set of BBs. Flags could be added in the future
+ * to change both of these default behaviors.
+ *
+ * Returns -EINVAL if hardware context placement configuration is invalid or if
+ * the placement configuration isn't supported on the platform / submission
+ * interface.
+ * Returns -ENODEV if extension isn't supported on the platform / submission
+ * interface.
+ *
+ * .. code-block:: none
+ *
+ *	Examples syntax:
+ *	CS[X] = generic engine of same class, logical instance X
+ *	INVALID = I915_ENGINE_CLASS_INVALID, I915_ENGINE_CLASS_INVALID_NONE
+ *
+ *	Example 1 pseudo code:
+ *	set_engines(INVALID)
+ *	set_parallel(engine_index=0, width=2, num_siblings=1,
+ *		     engines=CS[0],CS[1])
+ *
+ *	Results in the following valid placement:
+ *	CS[0], CS[1]
+ *
+ *	Example 2 pseudo code:
+ *	set_engines(INVALID)
+ *	set_parallel(engine_index=0, width=2, num_siblings=2,
+ *		     engines=CS[0],CS[2],CS[1],CS[3])
+ *
+ *	Results in the following valid placements:
+ *	CS[0], CS[1]
+ *	CS[2], CS[3]
+ *
+ *	This can be thought of as two virtual engines, each containing two
+ *	engines thereby making a 2D array. However, there are bonds tying the
+ *	entries together and placing restrictions on how they can be scheduled.
+ *	Specifically, the scheduler can choose only vertical columns from the 2D
+ *	array. That is, CS[0] is bonded to CS[1] and CS[2] to CS[3]. So if the
+ *	scheduler wants to submit to CS[0], it must also choose CS[1] and vice
+ *	versa. Same for CS[2] requires also using CS[3].
+ *	VE[0] = CS[0], CS[2]
+ *	VE[1] = CS[1], CS[3]
+ *
+ *	Example 3 pseudo code:
+ *	set_engines(INVALID)
+ *	set_parallel(engine_index=0, width=2, num_siblings=2,
+ *		     engines=CS[0],CS[1],CS[1],CS[3])
+ *
+ *	Results in the following valid and invalid placements:
+ *	CS[0], CS[1]
+ *	CS[1], CS[3] - Not logically contiguous, return -EINVAL
+ */
+struct i915_context_engines_parallel_submit {
+	/**
+	 * @base: base user extension.
+	 */
+	struct i915_user_extension base;
+
+	/**
+	 * @engine_index: slot for parallel engine
+	 */
+	__u16 engine_index;
+
+	/**
+	 * @width: number of contexts per parallel engine or in other words the
+	 * number of batches in each submission
+	 */
+	__u16 width;
+
+	/**
+	 * @num_siblings: number of siblings per context or in other words the
+	 * number of possible placements for each submission
+	 */
+	__u16 num_siblings;
+
+	/**
+	 * @mbz16: reserved for future use; must be zero
+	 */
+	__u16 mbz16;
+
+	/**
+	 * @flags: all undefined flags must be zero, currently not defined flags
+	 */
+	__u64 flags;
+
+	/**
+	 * @mbz64: reserved for future use; must be zero
+	 */
+	__u64 mbz64[3];
+
+	/**
+	 * @engines: 2-d array of engine instances to configure parallel engine
+	 *
+	 * length = width (i) * num_siblings (j)
+	 * index = j + i * num_siblings
+	 */
+	struct i915_engine_class_instance engines[0];
+
+} __packed;
+
+#define I915_DEFINE_CONTEXT_ENGINES_PARALLEL_SUBMIT(name__, N__) struct { \
+	struct i915_user_extension base; \
+	__u16 engine_index; \
+	__u16 width; \
+	__u16 num_siblings; \
+	__u16 mbz16; \
+	__u64 flags; \
+	__u64 mbz64[3]; \
+	struct i915_engine_class_instance engines[N__]; \
+} __attribute__((packed)) name__
+
+/**
  * DOC: Context Engine Map uAPI
  *
  * Context engine map is a new way of addressing engines when submitting batch-
···
 	__u64 extensions; /* linked chain of extension blocks, 0 terminates */
 #define I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE 0 /* see i915_context_engines_load_balance */
 #define I915_CONTEXT_ENGINES_EXT_BOND 1 /* see i915_context_engines_bond */
+#define I915_CONTEXT_ENGINES_EXT_PARALLEL_SUBMIT 2 /* see i915_context_engines_parallel_submit */
 	struct i915_engine_class_instance engines[0];
 } __attribute__((packed));