try alternate sg creation scheme

this reduces CPU usage and make bring us back inside virtio limits

5 files changed, 104 insertions, 10 deletions

diff --git a/drivers/gpu/drm/virgl/virgl_3d.c b/drivers/gpu/drm/virgl/virgl_3d.c
index 0533639fa9e8..43b973b59d1f 100644
--- a/drivers/gpu/drm/virgl/virgl_3d.c
+++ b/drivers/gpu/drm/virgl/virgl_3d.c
@@ -260,9 +260,13 @@ struct virgl_vbuffer *allocate_vbuf(struct virgl_device *qdev,
 				     int size, bool inout, u32 *base_offset, u32 max_bo_len)
 {
 	struct virgl_vbuffer *vbuf;
-	int sgpages = bo ? bo->tbo.num_pages : 0;
+	int sgpages = 0;
 	int ret;
 
+	if (bo)
+		if (!bo->is_res_bound)
+			sgpages = bo->tbo.num_pages;
+
 	sgpages++;
 	
 	vbuf = kmalloc(sizeof(*vbuf) + sizeof(struct scatterlist) * sgpages + size, GFP_KERNEL);
@@ -274,7 +278,7 @@ struct virgl_vbuffer *allocate_vbuf(struct virgl_device *qdev,
 	vbuf->sgpages = sgpages;
 	vbuf->size = size;
 
-	if (bo) {
+	if (bo && !bo->is_res_bound) {
 		struct scatterlist *sg;
 		unsigned int i;
 		uint32_t offset = 0;
@@ -537,7 +541,7 @@ fail:
 	return ret;
 }
 
-static int virgl_bo_list_validate(struct list_head *head)
+int virgl_bo_list_validate(struct list_head *head)
 {
 	struct ttm_validate_buffer *buf;
 	struct ttm_buffer_object *bo;
@@ -561,7 +565,7 @@ static int virgl_bo_list_validate(struct list_head *head)
 	return 0;
 }
 
-static void virgl_unref_list(struct list_head *head)
+void virgl_unref_list(struct list_head *head)
 {
 	struct ttm_validate_buffer *buf;
 	struct ttm_buffer_object *bo;
diff --git a/drivers/gpu/drm/virgl/virgl_drv.c b/drivers/gpu/drm/virgl/virgl_drv.c
index b329f036c19f..fa7945a0e41f 100644
--- a/drivers/gpu/drm/virgl/virgl_drv.c
+++ b/drivers/gpu/drm/virgl/virgl_drv.c
@@ -49,6 +49,11 @@ int virgl_modeset = -1;
 MODULE_PARM_DESC(modeset, "Disable/Enable modesetting");
 module_param_named(modeset, virgl_modeset, int, 0400);
 
+int virgl_create_sg = 0;
+
+MODULE_PARM_DESC(sgc, "Use new SG creation feature");
+module_param_named(sgc, virgl_create_sg, int, 0400);
+
 static struct drm_driver virgl_driver;
 static struct pci_driver virgl_pci_driver;
 
diff --git a/drivers/gpu/drm/virgl/virgl_drv.h b/drivers/gpu/drm/virgl/virgl_drv.h
index a49a24ccf10b..6e7e5c55956f 100644
--- a/drivers/gpu/drm/virgl/virgl_drv.h
+++ b/drivers/gpu/drm/virgl/virgl_drv.h
@@ -77,6 +77,7 @@ struct virgl_bo {
 	uint32_t res_handle; /* used for backing dumb objects */
 	bool dumb;
 	uint32_t stride;
+	bool is_res_bound;
 };
 #define gem_to_virgl_bo(gobj) container_of((gobj), struct virgl_bo, gem_base)
 
@@ -231,6 +232,7 @@ struct virgl_fpriv {
 
 extern struct drm_ioctl_desc virgl_ioctls[];
 extern int virgl_max_ioctl;
+extern int virgl_create_sg;
 
 int virgl_driver_load(struct drm_device *dev, unsigned long flags);
 int virgl_driver_unload(struct drm_device *dev);
@@ -399,4 +401,7 @@ int virgl_enable_vblank(struct drm_device *dev, int crtc);
 void virgl_disable_vblank(struct drm_device *dev, int crtc);
 
 int virgl_get_caps(struct virgl_device *vdev);
+
+int virgl_bo_list_validate(struct list_head *head);
+void virgl_unref_list(struct list_head *head);
 #endif
diff --git a/drivers/gpu/drm/virgl/virgl_hw.h b/drivers/gpu/drm/virgl/virgl_hw.h
index 63a9c5cd2171..a6088323a1a4 100644
--- a/drivers/gpu/drm/virgl/virgl_hw.h
+++ b/drivers/gpu/drm/virgl/virgl_hw.h
@@ -74,7 +74,7 @@ struct virgl_resource_create {
 	uint32_t array_size;
 	uint32_t last_level;
 	uint32_t nr_samples;
-	uint32_t pad;
+        uint32_t nr_sg_entries;
 };
 
 struct virgl_resource_unref {
@@ -120,6 +120,12 @@ struct virgl_cmd_get_cap {
 
 #define VIRGL_COMMAND_EMIT_FENCE (1 << 0)
 
+struct virgl_iov_entry {
+	VIRGLPHYSICAL addr;
+	uint32_t length;
+	uint32_t pad;
+};
+	
 struct virgl_command {
 	uint32_t type;
 	uint32_t flags;
diff --git a/drivers/gpu/drm/virgl/virgl_ioctl.c b/drivers/gpu/drm/virgl/virgl_ioctl.c
index c72ff98e1c9c..8932883bbca6 100644
--- a/drivers/gpu/drm/virgl/virgl_ioctl.c
+++ b/drivers/gpu/drm/virgl/virgl_ioctl.c
@@ -25,6 +25,7 @@
 
 #include "virgl_drv.h"
 #include "virgl_object.h"
+#include "ttm/ttm_execbuf_util.h"
 
 static void convert_to_hw_box(struct virgl_box *dst,
 			      const struct drm_virgl_3d_box *src)
@@ -101,8 +102,19 @@ static int virgl_resource_create_ioctl(struct drm_device *dev, void *data,
 	int ret;
 	uint32_t res_id;
 	struct virgl_bo *qobj;
-	uint32_t handle;
-	uint32_t size;
+	uint32_t handle = 0;
+	uint32_t size, pg_size;
+	struct virgl_bo *pg_bo = NULL;
+	void *optr;
+	int si;
+	struct scatterlist *sg;
+	struct list_head validate_list;
+	struct ttm_validate_buffer mainbuf, page_info_buf;
+	struct virgl_fence *fence;
+
+	INIT_LIST_HEAD(&validate_list);
+	memset(&mainbuf, 0, sizeof(struct ttm_validate_buffer));
+	memset(&page_info_buf, 0, sizeof(struct ttm_validate_buffer));
 
 	ret = virgl_resource_id_get(qdev, &res_id);
 	if (ret) 
@@ -110,16 +122,58 @@ static int virgl_resource_create_ioctl(struct drm_device *dev, void *data,
 
 	size = rc->size;
 
+	/* allocate a single page size object */
+	if (size == 0)
+		size = PAGE_SIZE;
+
 	ret = virgl_gem_object_create_with_handle(qdev, file_priv,
 						  0,
 						  size,
 						  &qobj, &handle);
+	if (ret)
+		goto fail_id;
+
+	/* use a gem reference since unref list undoes them */
+	drm_gem_object_reference(&qobj->gem_base);
+	mainbuf.bo = &qobj->tbo;
+	list_add(&mainbuf.head, &validate_list);
+
+	if (virgl_create_sg == 1) {
+		ret = virgl_bo_get_sg_table(qdev, qobj);
+		if (ret)
+			goto fail_obj;
+
+		pg_size = sizeof(struct virgl_iov_entry) * qobj->sgt->nents;
+
+		ret = virgl_bo_create(qdev, pg_size, true, 0, &pg_bo);
+		if (ret)
+			goto fail_unref;
+
+		drm_gem_object_reference(&pg_bo->gem_base);
+		page_info_buf.bo = &pg_bo->tbo;
+		list_add(&page_info_buf.head, &validate_list);
+	}
+
+	ret = virgl_bo_list_validate(&validate_list);
 	if (ret) {
-		virgl_resource_id_put(qdev, res_id);
-		return ret;
+		printk("failed to validate\n");
+		goto fail_unref;
 	}
 
-	cmd_p = virgl_alloc_cmd(qdev, NULL, false, NULL, 0, &vbuf);
+	if (virgl_create_sg == 1) {
+		ret = virgl_bo_kmap(pg_bo, &optr);
+		for_each_sg(qobj->sgt->sgl, sg, qobj->sgt->nents, si) {
+			struct virgl_iov_entry *iov = ((struct virgl_iov_entry *)optr) + si;
+			iov->addr = sg_phys(sg);
+			iov->length = sg->length;
+			iov->pad = 0;
+		}
+		virgl_bo_kunmap(pg_bo);
+
+		qobj->is_res_bound = true;
+	}
+
+	cmd_p = virgl_alloc_cmd(qdev, pg_bo, false, NULL, 0, &vbuf);
 	memset(cmd_p, 0, sizeof(*cmd_p));
 	cmd_p->type = VIRGL_CMD_CREATE_RESOURCE;
 	cmd_p->u.res_create.handle = res_id;
@@ -132,15 +186,35 @@ static int virgl_resource_create_ioctl(struct drm_device *dev, void *data,
 	cmd_p->u.res_create.array_size = rc->array_size;
 	cmd_p->u.res_create.last_level = rc->last_level;
 	cmd_p->u.res_create.nr_samples = rc->nr_samples;
+	cmd_p->u.res_create.nr_sg_entries = qobj->sgt ? qobj->sgt->nents : 0;
+
+	ret = virgl_fence_emit(qdev, cmd_p, &fence);
 
 	virgl_queue_cmd_buf(qdev, vbuf);
 
+	ttm_eu_fence_buffer_objects(&validate_list, fence);
+
 	qobj->res_handle = res_id;
 	qobj->stride = rc->stride;
 	rc->res_handle = res_id; /* similiar to a VM address */
 	rc->bo_handle = handle;
 
+	virgl_unref_list(&validate_list);
+	if (virgl_create_sg == 1)
+		virgl_bo_unref(&pg_bo);
+
 	return 0;
+fail_unref:
+	virgl_unref_list(&validate_list);
+fail_pg_obj:
+	if (virgl_create_sg == 1)
+		if (pg_bo)
+			virgl_bo_unref(&pg_bo);
+fail_obj:
+	drm_gem_object_handle_unreference_unlocked(&qobj->gem_base);
+fail_id:
+	virgl_resource_id_put(qdev, res_id);
+	return ret;
 }
 
 static int virgl_resource_info_ioctl(struct drm_device *dev, void *data,