drm/i915: Reuse the active golden render state batch

The golden render state is constant, but we recreate the batch setting
it up for every new context. If we keep that batch in a volatile cache
we can safely reuse it whenever we need to initialise a new context. We
mark the pages as purgeable and use the shrinker to recover pages from
the batch whenever we face memory pressues, recreating that batch afresh
on the next new context.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtien@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20161028125858.23563-8-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/i915_gem_render_state.c

@@ -28,17 +28,19 @@
 #include "i915_drv.h"
 #include "intel_renderstate.h"
 
-struct render_state {
+struct intel_render_state {
 	const struct intel_renderstate_rodata *rodata;
 	struct i915_vma *vma;
-	u32 aux_batch_size;
-	u32 aux_batch_offset;
+	u32 batch_offset;
+	u32 batch_size;
+	u32 aux_offset;
+	u32 aux_size;
 };
 
 static const struct intel_renderstate_rodata *
-render_state_get_rodata(const struct drm_i915_gem_request *req)
+render_state_get_rodata(const struct intel_engine_cs *engine)
 {
-	switch (INTEL_GEN(req->i915)) {
+	switch (INTEL_GEN(engine->i915)) {
 	case 6:
 		return &gen6_null_state;
 	case 7:
@@ -63,29 +65,27 @@ render_state_get_rodata(const struct drm_i915_gem_request *req)
  */
 #define OUT_BATCH(batch, i, val)				\
 	do {							\
-		if (WARN_ON((i) >= PAGE_SIZE / sizeof(u32))) {	\
-			ret = -ENOSPC;				\
-			goto err_out;				\
-		}						\
+		if ((i) >= PAGE_SIZE / sizeof(u32))		\
+			goto err;				\
 		(batch)[(i)++] = (val);				\
 	} while(0)
 
-static int render_state_setup(struct render_state *so)
+static int render_state_setup(struct intel_render_state *so,
+			      struct drm_i915_private *i915)
 {
-	struct drm_i915_private *dev_priv = to_i915(so->vma->vm->dev);
 	const struct intel_renderstate_rodata *rodata = so->rodata;
-	const bool has_64bit_reloc = INTEL_GEN(dev_priv) >= 8;
+	const bool has_64bit_reloc = INTEL_GEN(i915) >= 8;
+	struct drm_i915_gem_object *obj = so->vma->obj;
 	unsigned int i = 0, reloc_index = 0;
-	struct page *page;
+	unsigned int needs_clflush;
 	u32 *d;
 	int ret;
 
-	ret = i915_gem_object_set_to_cpu_domain(so->vma->obj, true);
+	ret = i915_gem_obj_prepare_shmem_write(obj, &needs_clflush);
 	if (ret)
 		return ret;
 
-	page = i915_gem_object_get_dirty_page(so->vma->obj, 0);
-	d = kmap(page);
+	d = kmap_atomic(i915_gem_object_get_dirty_page(obj, 0));
 
 	while (i < rodata->batch_items) {
 		u32 s = rodata->batch[i];
@@ -95,10 +95,8 @@ static int render_state_setup(struct render_state *so)
 			s = lower_32_bits(r);
 			if (has_64bit_reloc) {
 				if (i + 1 >= rodata->batch_items ||
-				    rodata->batch[i + 1] != 0) {
-					ret = -EINVAL;
-					goto err_out;
-				}
+				    rodata->batch[i + 1] != 0)
+					goto err;
 
 				d[i++] = s;
 				s = upper_32_bits(r);
@@ -110,12 +108,20 @@ static int render_state_setup(struct render_state *so)
 		d[i++] = s;
 	}
 
+	if (rodata->reloc[reloc_index] != -1) {
+		DRM_ERROR("only %d relocs resolved\n", reloc_index);
+		goto err;
+	}
+
+	so->batch_offset = so->vma->node.start;
+	so->batch_size = rodata->batch_items * sizeof(u32);
+
 	while (i % CACHELINE_DWORDS)
 		OUT_BATCH(d, i, MI_NOOP);
 
-	so->aux_batch_offset = i * sizeof(u32);
+	so->aux_offset = i * sizeof(u32);
 
-	if (HAS_POOLED_EU(dev_priv)) {
+	if (HAS_POOLED_EU(i915)) {
 		/*
 		 * We always program 3x6 pool config but depending upon which
 		 * subslice is disabled HW drops down to appropriate config
@@ -143,89 +149,131 @@ static int render_state_setup(struct render_state *so)
 	}
 
 	OUT_BATCH(d, i, MI_BATCH_BUFFER_END);
-	so->aux_batch_size = (i * sizeof(u32)) - so->aux_batch_offset;
-
+	so->aux_size = i * sizeof(u32) - so->aux_offset;
+	so->aux_offset += so->batch_offset;
 	/*
 	 * Since we are sending length, we need to strictly conform to
 	 * all requirements. For Gen2 this must be a multiple of 8.
 	 */
-	so->aux_batch_size = ALIGN(so->aux_batch_size, 8);
+	so->aux_size = ALIGN(so->aux_size, 8);
 
-	kunmap(page);
+	if (needs_clflush)
+		drm_clflush_virt_range(d, i * sizeof(u32));
+	kunmap_atomic(d);
 
-	ret = i915_gem_object_set_to_gtt_domain(so->vma->obj, false);
-	if (ret)
-		return ret;
-
-	if (rodata->reloc[reloc_index] != -1) {
-		DRM_ERROR("only %d relocs resolved\n", reloc_index);
-		return -EINVAL;
-	}
-
-	return 0;
-
-err_out:
-	kunmap(page);
+	ret = i915_gem_object_set_to_gtt_domain(obj, false);
+out:
+	i915_gem_obj_finish_shmem_access(obj);
 	return ret;
+
+err:
+	kunmap_atomic(d);
+	ret = -EINVAL;
+	goto out;
 }
 
 #undef OUT_BATCH
 
-int i915_gem_render_state_init(struct drm_i915_gem_request *req)
+int i915_gem_render_state_init(struct intel_engine_cs *engine)
 {
-	struct render_state so;
+	struct intel_render_state *so;
+	const struct intel_renderstate_rodata *rodata;
 	struct drm_i915_gem_object *obj;
 	int ret;
 
-	if (WARN_ON(req->engine->id != RCS))
-		return -ENOENT;
-
-	so.rodata = render_state_get_rodata(req);
-	if (!so.rodata)
+	if (engine->id != RCS)
 		return 0;
 
-	if (so.rodata->batch_items * 4 > 4096)
+	rodata = render_state_get_rodata(engine);
+	if (!rodata)
+		return 0;
+
+	if (rodata->batch_items * 4 > 4096)
 		return -EINVAL;
 
-	obj = i915_gem_object_create_internal(req->i915, 4096);
-	if (IS_ERR(obj))
-		return PTR_ERR(obj);
+	so = kmalloc(sizeof(*so), GFP_KERNEL);
+	if (!so)
+		return -ENOMEM;
 
-	so.vma = i915_vma_create(obj, &req->i915->ggtt.base, NULL);
-	if (IS_ERR(so.vma)) {
-		ret = PTR_ERR(so.vma);
+	obj = i915_gem_object_create_internal(engine->i915, 4096);
+	if (IS_ERR(obj)) {
+		ret = PTR_ERR(obj);
+		goto err_free;
+	}
+
+	so->vma = i915_vma_create(obj, &engine->i915->ggtt.base, NULL);
+	if (IS_ERR(so->vma)) {
+		ret = PTR_ERR(so->vma);
 		goto err_obj;
 	}
 
-	ret = i915_vma_pin(so.vma, 0, 0, PIN_GLOBAL);
-	if (ret)
-		goto err_obj;
+	so->rodata = rodata;
+	engine->render_state = so;
+	return 0;
 
-	ret = render_state_setup(&so);
-	if (ret)
-		goto err_unpin;
+err_obj:
+	i915_gem_object_put(obj);
+err_free:
+	kfree(so);
+	return ret;
+}
 
-	ret = req->engine->emit_bb_start(req, so.vma->node.start,
-					 so.rodata->batch_items * 4,
+int i915_gem_render_state_emit(struct drm_i915_gem_request *req)
+{
+	struct intel_render_state *so;
+	int ret;
+
+	so = req->engine->render_state;
+	if (!so)
+		return 0;
+
+	/* Recreate the page after shrinking */
+	if (!so->vma->obj->pages)
+		so->batch_offset = -1;
+
+	ret = i915_vma_pin(so->vma, 0, 0, PIN_GLOBAL | PIN_HIGH);
+	if (ret)
+		return ret;
+
+	if (so->vma->node.start != so->batch_offset) {
+		ret = render_state_setup(so, req->i915);
+		if (ret)
+			goto err_unpin;
+	}
+
+	ret = req->engine->emit_bb_start(req,
+					 so->batch_offset, so->batch_size,
 					 I915_DISPATCH_SECURE);
 	if (ret)
 		goto err_unpin;
 
-	if (so.aux_batch_size > 8) {
+	if (so->aux_size > 8) {
 		ret = req->engine->emit_bb_start(req,
-						 (so.vma->node.start +
-						  so.aux_batch_offset),
-						 so.aux_batch_size,
+						 so->aux_offset, so->aux_size,
 						 I915_DISPATCH_SECURE);
 		if (ret)
 			goto err_unpin;
 	}
 
-	i915_vma_move_to_active(so.vma, req, 0);
+	i915_vma_move_to_active(so->vma, req, 0);
 err_unpin:
-	i915_vma_unpin(so.vma);
-	i915_vma_close(so.vma);
-err_obj:
-	__i915_gem_object_release_unless_active(obj);
+	i915_vma_unpin(so->vma);
 	return ret;
 }
+
+void i915_gem_render_state_fini(struct intel_engine_cs *engine)
+{
+	struct intel_render_state *so;
+	struct drm_i915_gem_object *obj;
+
+	so = fetch_and_zero(&engine->render_state);
+	if (!so)
+		return;
+
+	obj = so->vma->obj;
+
+	i915_vma_close(so->vma);
+	__i915_gem_object_release_unless_active(obj);
+
+	kfree(so);
+}

drivers/gpu/drm/i915/i915_gem_render_state.h

@@ -26,6 +26,8 @@
 
 struct drm_i915_gem_request;
 
-int i915_gem_render_state_init(struct drm_i915_gem_request *req);
+int i915_gem_render_state_init(struct intel_engine_cs *engine);
+int i915_gem_render_state_emit(struct drm_i915_gem_request *req);
+void i915_gem_render_state_fini(struct intel_engine_cs *engine);
 
 #endif /* _I915_GEM_RENDER_STATE_H_ */

drivers/gpu/drm/i915/intel_engine_cs.c

@@ -314,6 +314,10 @@ int intel_engine_init_common(struct intel_engine_cs *engine)
 	if (ret)
 		return ret;
 
+	ret = i915_gem_render_state_init(engine);
+	if (ret)
+		return ret;
+
 	return 0;
 }
 
@@ -328,6 +332,7 @@ void intel_engine_cleanup_common(struct intel_engine_cs *engine)
 {
 	intel_engine_cleanup_scratch(engine);
 
+	i915_gem_render_state_fini(engine);
 	intel_engine_fini_breadcrumbs(engine);
 	intel_engine_cleanup_cmd_parser(engine);
 	i915_gem_batch_pool_fini(&engine->batch_pool);

drivers/gpu/drm/i915/intel_lrc.c

@@ -1637,7 +1637,7 @@ static int gen8_init_rcs_context(struct drm_i915_gem_request *req)
 	if (ret)
 		DRM_ERROR("MOCS failed to program: expect performance issues.\n");
 
-	return i915_gem_render_state_init(req);
+	return i915_gem_render_state_emit(req);
 }
 
 /**

drivers/gpu/drm/i915/intel_ringbuffer.c

@@ -648,7 +648,7 @@ static int intel_rcs_ctx_init(struct drm_i915_gem_request *req)
 	if (ret != 0)
 		return ret;
 
-	ret = i915_gem_render_state_init(req);
+	ret = i915_gem_render_state_emit(req);
 	if (ret)
 		return ret;
 

drivers/gpu/drm/i915/intel_ringbuffer.h

@@ -157,6 +157,7 @@ struct i915_ctx_workarounds {
 };
 
 struct drm_i915_gem_request;
+struct intel_render_state;
 
 struct intel_engine_cs {
 	struct drm_i915_private *i915;
@@ -184,6 +185,8 @@ struct intel_engine_cs {
 	unsigned int irq_shift;
 	struct intel_ring *buffer;
 
+	struct intel_render_state *render_state;
+
 	/* Rather than have every client wait upon all user interrupts,
 	 * with the herd waking after every interrupt and each doing the
 	 * heavyweight seqno dance, we delegate the task (of being the