[GEM] Make pread/pwrite manage memory domains. No luck with movnti though.

pread and pwrite must update the memory domains to ensure consistency with
the GPU. At some point, it should be possible to avoid clflush through this
path, but that isn't working for me.

5 files changed, 100 insertions, 27 deletions

diff --git a/linux-core/drmP.h b/linux-core/drmP.h
index 419b4be3..fc7043d7 100644
--- a/linux-core/drmP.h
+++ b/linux-core/drmP.h
@@ -781,6 +781,13 @@ struct drm_driver {
 			       uint32_t read_domains,
 			       uint32_t write_domain);
 
+	/**
+	 * Driver-specific callback to flush pwrite through chipset
+	 */
+	int (*gem_flush_pwrite) (struct drm_gem_object *obj,
+				 uint64_t offset,
+				 uint64_t size);
+
 	struct drm_fence_driver *fence_driver;
 	struct drm_bo_driver *bo_driver;
 
diff --git a/linux-core/drm_gem.c b/linux-core/drm_gem.c
index 6c462921..14cf7e47 100644
--- a/linux-core/drm_gem.c
+++ b/linux-core/drm_gem.c
@@ -266,6 +266,7 @@ drm_gem_pread_ioctl(struct drm_device *dev, void *data,
 	struct drm_gem_object *obj;
 	ssize_t read;
 	loff_t offset;
+	int ret;
 
 	if (!(dev->driver->driver_features & DRIVER_GEM))
 		return -ENODEV;
@@ -274,6 +275,15 @@ drm_gem_pread_ioctl(struct drm_device *dev, void *data,
 	if (obj == NULL)
 		return -EINVAL;
 
+	if (dev->driver->gem_set_domain) {
+		ret = dev->driver->gem_set_domain (obj,
+						   DRM_GEM_DOMAIN_CPU,
+						   0);
+		if (ret) {
+			drm_gem_object_unreference(obj);
+			return ret;
+		}
+	}
 	offset = args->offset;
 
 	read = vfs_read(obj->filp, (char __user *)(uintptr_t)args->data_ptr,
@@ -343,6 +353,7 @@ drm_gem_pwrite_ioctl(struct drm_device *dev, void *data,
 	struct drm_gem_object *obj;
 	ssize_t written;
 	loff_t offset;
+	int ret;
 
 	if (!(dev->driver->driver_features & DRIVER_GEM))
 		return -ENODEV;
@@ -351,11 +362,21 @@ drm_gem_pwrite_ioctl(struct drm_device *dev, void *data,
 	if (obj == NULL)
 		return -EINVAL;
 
+	if (dev->driver->gem_set_domain) {
+		ret = dev->driver->gem_set_domain (obj,
+						   DRM_GEM_DOMAIN_CPU,
+						   0);
+		if (ret) {
+			drm_gem_object_unreference(obj);
+			return ret;
+		}
+	}
 	offset = args->offset;
 
 	written = vfs_write(obj->filp,
 			    (char __user *)(uintptr_t) args->data_ptr,
 			    args->size, &offset);
+	
 	if (written != args->size) {
 		drm_gem_object_unreference(obj);
 		if (written < 0)
@@ -364,6 +385,11 @@ drm_gem_pwrite_ioctl(struct drm_device *dev, void *data,
 			return -EINVAL;
 	}
 
+	if (dev->driver->gem_flush_pwrite)
+		dev->driver->gem_flush_pwrite(obj,
+					      args->offset,
+					      args->size);
+
 	drm_gem_object_unreference(obj);
 
 	return 0;
diff --git a/linux-core/i915_drv.c b/linux-core/i915_drv.c
index ae8cf3e0..cc47ed64 100644
--- a/linux-core/i915_drv.c
+++ b/linux-core/i915_drv.c
@@ -589,6 +589,7 @@ static struct drm_driver driver = {
 	.gem_init_object = i915_gem_init_object,
 	.gem_free_object = i915_gem_free_object,
 	.gem_set_domain = i915_gem_set_domain_ioctl,
+	.gem_flush_pwrite = i915_gem_flush_pwrite,
 	.fops = {
 		.owner = THIS_MODULE,
 		.open = drm_open,
diff --git a/linux-core/i915_gem.c b/linux-core/i915_gem.c
index 90d1d52b..9d701ba1 100644
--- a/linux-core/i915_gem.c
+++ b/linux-core/i915_gem.c
@@ -377,6 +377,37 @@ i915_gem_evict_something(struct drm_device *dev)
 	return 0;
 }
 
+static int
+i915_gem_object_get_page_list(struct drm_gem_object *obj)
+{
+	struct drm_i915_gem_object *obj_priv = obj->driver_private;
+	int page_count, i;
+	if (obj_priv->page_list)
+		return 0;
+
+	/* Get the list of pages out of our struct file.  They'll be pinned
+	 * at this point until we release them.
+	 */
+	page_count = obj->size / PAGE_SIZE;
+	BUG_ON(obj_priv->page_list != NULL);
+	obj_priv->page_list = drm_calloc(page_count, sizeof(struct page *),
+					 DRM_MEM_DRIVER);
+	if (obj_priv->page_list == NULL)
+		return -ENOMEM;
+
+	for (i = 0; i < page_count; i++) {
+		obj_priv->page_list[i] =
+		    find_or_create_page(obj->filp->f_mapping, i, GFP_HIGHUSER);
+
+		if (obj_priv->page_list[i] == NULL) {
+			i915_gem_object_free_page_list(obj);
+			return -ENOMEM;
+		}
+		unlock_page(obj_priv->page_list[i]);
+	}
+	return 0;
+}
+
 /**
  * Finds free space in the GTT aperture and binds the object there.
  */
@@ -387,7 +418,7 @@ i915_gem_object_bind_to_gtt(struct drm_gem_object *obj, unsigned alignment)
 	drm_i915_private_t *dev_priv = dev->dev_private;
 	struct drm_i915_gem_object *obj_priv = obj->driver_private;
 	struct drm_memrange_node *free_space;
-	int page_count, i, ret;
+	int page_count, ret;
 
 	if (alignment == 0)
 		alignment = PAGE_SIZE;
@@ -432,33 +463,14 @@ i915_gem_object_bind_to_gtt(struct drm_gem_object *obj, unsigned alignment)
 	DRM_INFO("Binding object of size %d at 0x%08x\n",
 		 obj->size, obj_priv->gtt_offset);
 #endif
-
-	/* Get the list of pages out of our struct file.  They'll be pinned
-	 * at this point until we release them.
-	 */
-	page_count = obj->size / PAGE_SIZE;
-	BUG_ON(obj_priv->page_list != NULL);
-	obj_priv->page_list = drm_calloc(page_count, sizeof(struct page *),
-					 DRM_MEM_DRIVER);
-	if (obj_priv->page_list == NULL) {
-		drm_memrange_put_block(obj_priv->gtt_space);
+	ret = i915_gem_object_get_page_list (obj);
+	if (ret) {
+		drm_memrange_put_block (obj_priv->gtt_space);
 		obj_priv->gtt_space = NULL;
-		return -ENOMEM;
-	}
-
-	for (i = 0; i < page_count; i++) {
-		obj_priv->page_list[i] =
-		    find_or_create_page(obj->filp->f_mapping, i, GFP_HIGHUSER);
-
-		if (obj_priv->page_list[i] == NULL) {
-			i915_gem_object_free_page_list(obj);
-			drm_memrange_put_block(obj_priv->gtt_space);
-			obj_priv->gtt_space = NULL;
-			return -ENOMEM;
-		}
-		unlock_page(obj_priv->page_list[i]);
+		return ret;
 	}
 
+	page_count = obj->size / PAGE_SIZE;
 	/* Create an AGP memory structure pointing at our pages, and bind it
 	 * into the GTT.
 	 */
@@ -1052,14 +1064,39 @@ void i915_gem_free_object(struct drm_gem_object *obj)
 
 int
 i915_gem_set_domain_ioctl(struct drm_gem_object *obj,
-			   uint32_t read_domains,
-			   uint32_t write_domain)
+			  uint32_t read_domains,
+			  uint32_t write_domain)
 {
 	i915_gem_object_set_domain(obj, read_domains, write_domain);
 	i915_gem_dev_set_domain(obj->dev);
 	return 0;
 }
 
+int
+i915_gem_flush_pwrite(struct drm_gem_object *obj,
+		      uint64_t offset, uint64_t size)
+{
+	struct drm_device *dev = obj->dev;
+	struct drm_i915_gem_object *obj_priv = obj->driver_private;
+	/*
+	 * As far as I can tell, writes of > 64 bytes will use non-temporal
+	 * stores which should obviate the need for this clflush.
+	 * It doesn't work for me though...
+	 */
+/*	if (size <= 64)  */{
+		if (obj_priv->gtt_space == NULL) {
+			int ret = i915_gem_object_get_page_list (obj);
+			if (ret)
+				 return ret;
+		}
+		i915_gem_clflush_object(obj);
+		if (obj_priv->gtt_space == NULL)
+			i915_gem_object_free_page_list (obj);
+	}
+	drm_agp_chipset_flush(dev);
+	return 0;
+}
+
 void
 i915_gem_lastclose(struct drm_device *dev)
 {
diff --git a/shared-core/i915_drv.h b/shared-core/i915_drv.h
index 413fca89..78c5d3df 100644
--- a/shared-core/i915_drv.h
+++ b/shared-core/i915_drv.h
@@ -391,6 +391,8 @@ void i915_gem_free_object(struct drm_gem_object *obj);
 int i915_gem_set_domain_ioctl (struct drm_gem_object *obj,
 			       uint32_t read_domains,
 			       uint32_t write_domain);
+int i915_gem_flush_pwrite(struct drm_gem_object *obj,
+			  uint64_t offset, uint64_t size);
 void i915_gem_lastclose(struct drm_device *dev);
 #endif