drm/vc4: Add a BO cache.

We need to allocate new BOs in the kernel as part of each frame, but
the CMA allocator is way too slow for that.  As an optimization, keep
track of recently-freed BOs and reuse them, with a 1 second timeout to
fully free them back to the system.

This improves 3D performance by about 15%.

Signed-off-by: Eric Anholt <eric@anholt.net>

drivers/gpu/drm/vc4/vc4_bo.c

@@ -12,19 +12,229 @@
  * access to system memory with no MMU in between.  To support it, we
  * use the GEM CMA helper functions to allocate contiguous ranges of
  * physical memory for our BOs.
+ *
+ * Since the CMA allocator is very slow, we keep a cache of recently
+ * freed BOs around so that the kernel's allocation of objects for 3D
+ * rendering can return quickly.
  */
 
 #include "vc4_drv.h"
 
-struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t size)
+static void vc4_bo_stats_dump(struct vc4_dev *vc4)
 {
+	DRM_INFO("num bos allocated: %d\n",
+		 vc4->bo_stats.num_allocated);
+	DRM_INFO("size bos allocated: %dkb\n",
+		 vc4->bo_stats.size_allocated / 1024);
+	DRM_INFO("num bos used: %d\n",
+		 vc4->bo_stats.num_allocated - vc4->bo_stats.num_cached);
+	DRM_INFO("size bos used: %dkb\n",
+		 (vc4->bo_stats.size_allocated -
+		  vc4->bo_stats.size_cached) / 1024);
+	DRM_INFO("num bos cached: %d\n",
+		 vc4->bo_stats.num_cached);
+	DRM_INFO("size bos cached: %dkb\n",
+		 vc4->bo_stats.size_cached / 1024);
+}
+
+#ifdef CONFIG_DEBUG_FS
+int vc4_bo_stats_debugfs(struct seq_file *m, void *unused)
+{
+	struct drm_info_node *node = (struct drm_info_node *)m->private;
+	struct drm_device *dev = node->minor->dev;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct vc4_bo_stats stats;
+
+	/* Take a snapshot of the current stats with the lock held. */
+	mutex_lock(&vc4->bo_lock);
+	stats = vc4->bo_stats;
+	mutex_unlock(&vc4->bo_lock);
+
+	seq_printf(m, "num bos allocated: %d\n",
+		   stats.num_allocated);
+	seq_printf(m, "size bos allocated: %dkb\n",
+		   stats.size_allocated / 1024);
+	seq_printf(m, "num bos used: %d\n",
+		   stats.num_allocated - stats.num_cached);
+	seq_printf(m, "size bos used: %dkb\n",
+		   (stats.size_allocated - stats.size_cached) / 1024);
+	seq_printf(m, "num bos cached: %d\n",
+		   stats.num_cached);
+	seq_printf(m, "size bos cached: %dkb\n",
+		   stats.size_cached / 1024);
+
+	return 0;
+}
+#endif
+
+static uint32_t bo_page_index(size_t size)
+{
+	return (size / PAGE_SIZE) - 1;
+}
+
+/* Must be called with bo_lock held. */
+static void vc4_bo_destroy(struct vc4_bo *bo)
+{
+	struct drm_gem_object *obj = &bo->base.base;
+	struct vc4_dev *vc4 = to_vc4_dev(obj->dev);
+
+	vc4->bo_stats.num_allocated--;
+	vc4->bo_stats.size_allocated -= obj->size;
+	drm_gem_cma_free_object(obj);
+}
+
+/* Must be called with bo_lock held. */
+static void vc4_bo_remove_from_cache(struct vc4_bo *bo)
+{
+	struct drm_gem_object *obj = &bo->base.base;
+	struct vc4_dev *vc4 = to_vc4_dev(obj->dev);
+
+	vc4->bo_stats.num_cached--;
+	vc4->bo_stats.size_cached -= obj->size;
+
+	list_del(&bo->unref_head);
+	list_del(&bo->size_head);
+}
+
+static struct list_head *vc4_get_cache_list_for_size(struct drm_device *dev,
+						     size_t size)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	uint32_t page_index = bo_page_index(size);
+
+	if (vc4->bo_cache.size_list_size <= page_index) {
+		uint32_t new_size = max(vc4->bo_cache.size_list_size * 2,
+					page_index + 1);
+		struct list_head *new_list;
+		uint32_t i;
+
+		new_list = kmalloc_array(new_size, sizeof(struct list_head),
+					 GFP_KERNEL);
+		if (!new_list)
+			return NULL;
+
+		/* Rebase the old cached BO lists to their new list
+		 * head locations.
+		 */
+		for (i = 0; i < vc4->bo_cache.size_list_size; i++) {
+			struct list_head *old_list =
+				&vc4->bo_cache.size_list[i];
+
+			if (list_empty(old_list))
+				INIT_LIST_HEAD(&new_list[i]);
+			else
+				list_replace(old_list, &new_list[i]);
+		}
+		/* And initialize the brand new BO list heads. */
+		for (i = vc4->bo_cache.size_list_size; i < new_size; i++)
+			INIT_LIST_HEAD(&new_list[i]);
+
+		kfree(vc4->bo_cache.size_list);
+		vc4->bo_cache.size_list = new_list;
+		vc4->bo_cache.size_list_size = new_size;
+	}
+
+	return &vc4->bo_cache.size_list[page_index];
+}
+
+void vc4_bo_cache_purge(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	mutex_lock(&vc4->bo_lock);
+	while (!list_empty(&vc4->bo_cache.time_list)) {
+		struct vc4_bo *bo = list_last_entry(&vc4->bo_cache.time_list,
+						    struct vc4_bo, unref_head);
+		vc4_bo_remove_from_cache(bo);
+		vc4_bo_destroy(bo);
+	}
+	mutex_unlock(&vc4->bo_lock);
+}
+
+static struct vc4_bo *vc4_bo_get_from_cache(struct drm_device *dev,
+					    uint32_t size)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	uint32_t page_index = bo_page_index(size);
+	struct vc4_bo *bo = NULL;
+
+	size = roundup(size, PAGE_SIZE);
+
+	mutex_lock(&vc4->bo_lock);
+	if (page_index >= vc4->bo_cache.size_list_size)
+		goto out;
+
+	if (list_empty(&vc4->bo_cache.size_list[page_index]))
+		goto out;
+
+	bo = list_first_entry(&vc4->bo_cache.size_list[page_index],
+			      struct vc4_bo, size_head);
+	vc4_bo_remove_from_cache(bo);
+	kref_init(&bo->base.base.refcount);
+
+out:
+	mutex_unlock(&vc4->bo_lock);
+	return bo;
+}
+
+/**
+ * vc4_gem_create_object - Implementation of driver->gem_create_object.
+ *
+ * This lets the CMA helpers allocate object structs for us, and keep
+ * our BO stats correct.
+ */
+struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct vc4_bo *bo;
+
+	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
+	if (!bo)
+		return ERR_PTR(-ENOMEM);
+
+	mutex_lock(&vc4->bo_lock);
+	vc4->bo_stats.num_allocated++;
+	vc4->bo_stats.size_allocated += size;
+	mutex_unlock(&vc4->bo_lock);
+
+	return &bo->base.base;
+}
+
+struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t unaligned_size,
+			     bool from_cache)
+{
+	size_t size = roundup(unaligned_size, PAGE_SIZE);
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
 	struct drm_gem_cma_object *cma_obj;
 
-	cma_obj = drm_gem_cma_create(dev, size);
+	if (size == 0)
-	if (IS_ERR(cma_obj))
 		return NULL;
-	else
+
-		return to_vc4_bo(&cma_obj->base);
+	/* First, try to get a vc4_bo from the kernel BO cache. */
+	if (from_cache) {
+		struct vc4_bo *bo = vc4_bo_get_from_cache(dev, size);
+
+		if (bo)
+			return bo;
+	}
+
+	cma_obj = drm_gem_cma_create(dev, size);
+	if (IS_ERR(cma_obj)) {
+		/*
+		 * If we've run out of CMA memory, kill the cache of
+		 * CMA allocations we've got laying around and try again.
+		 */
+		vc4_bo_cache_purge(dev);
+
+		cma_obj = drm_gem_cma_create(dev, size);
+		if (IS_ERR(cma_obj)) {
+			DRM_ERROR("Failed to allocate from CMA:\n");
+			vc4_bo_stats_dump(vc4);
+			return NULL;
+		}
+	}
+
+	return to_vc4_bo(&cma_obj->base);
 }
 
 int vc4_dumb_create(struct drm_file *file_priv,
@@ -41,7 +251,7 @@ int vc4_dumb_create(struct drm_file *file_priv,
 	if (args->size < args->pitch * args->height)
 		args->size = args->pitch * args->height;
 
-	bo = vc4_bo_create(dev, roundup(args->size, PAGE_SIZE));
+	bo = vc4_bo_create(dev, args->size, false);
 	if (!bo)
 		return -ENOMEM;
 
@@ -50,3 +260,117 @@ int vc4_dumb_create(struct drm_file *file_priv,
 
 	return ret;
 }
+
+/* Must be called with bo_lock held. */
+static void vc4_bo_cache_free_old(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	unsigned long expire_time = jiffies - msecs_to_jiffies(1000);
+
+	while (!list_empty(&vc4->bo_cache.time_list)) {
+		struct vc4_bo *bo = list_last_entry(&vc4->bo_cache.time_list,
+						    struct vc4_bo, unref_head);
+		if (time_before(expire_time, bo->free_time)) {
+			mod_timer(&vc4->bo_cache.time_timer,
+				  round_jiffies_up(jiffies +
+						   msecs_to_jiffies(1000)));
+			return;
+		}
+
+		vc4_bo_remove_from_cache(bo);
+		vc4_bo_destroy(bo);
+	}
+}
+
+/* Called on the last userspace/kernel unreference of the BO.  Returns
+ * it to the BO cache if possible, otherwise frees it.
+ *
+ * Note that this is called with the struct_mutex held.
+ */
+void vc4_free_object(struct drm_gem_object *gem_bo)
+{
+	struct drm_device *dev = gem_bo->dev;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct vc4_bo *bo = to_vc4_bo(gem_bo);
+	struct list_head *cache_list;
+
+	mutex_lock(&vc4->bo_lock);
+	/* If the object references someone else's memory, we can't cache it.
+	 */
+	if (gem_bo->import_attach) {
+		vc4_bo_destroy(bo);
+		goto out;
+	}
+
+	/* Don't cache if it was publicly named. */
+	if (gem_bo->name) {
+		vc4_bo_destroy(bo);
+		goto out;
+	}
+
+	cache_list = vc4_get_cache_list_for_size(dev, gem_bo->size);
+	if (!cache_list) {
+		vc4_bo_destroy(bo);
+		goto out;
+	}
+
+	bo->free_time = jiffies;
+	list_add(&bo->size_head, cache_list);
+	list_add(&bo->unref_head, &vc4->bo_cache.time_list);
+
+	vc4->bo_stats.num_cached++;
+	vc4->bo_stats.size_cached += gem_bo->size;
+
+	vc4_bo_cache_free_old(dev);
+
+out:
+	mutex_unlock(&vc4->bo_lock);
+}
+
+static void vc4_bo_cache_time_work(struct work_struct *work)
+{
+	struct vc4_dev *vc4 =
+		container_of(work, struct vc4_dev, bo_cache.time_work);
+	struct drm_device *dev = vc4->dev;
+
+	mutex_lock(&vc4->bo_lock);
+	vc4_bo_cache_free_old(dev);
+	mutex_unlock(&vc4->bo_lock);
+}
+
+static void vc4_bo_cache_time_timer(unsigned long data)
+{
+	struct drm_device *dev = (struct drm_device *)data;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	schedule_work(&vc4->bo_cache.time_work);
+}
+
+void vc4_bo_cache_init(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	mutex_init(&vc4->bo_lock);
+
+	INIT_LIST_HEAD(&vc4->bo_cache.time_list);
+
+	INIT_WORK(&vc4->bo_cache.time_work, vc4_bo_cache_time_work);
+	setup_timer(&vc4->bo_cache.time_timer,
+		    vc4_bo_cache_time_timer,
+		    (unsigned long)dev);
+}
+
+void vc4_bo_cache_destroy(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	del_timer(&vc4->bo_cache.time_timer);
+	cancel_work_sync(&vc4->bo_cache.time_work);
+
+	vc4_bo_cache_purge(dev);
+
+	if (vc4->bo_stats.num_allocated) {
+		DRM_ERROR("Destroying BO cache while BOs still allocated:\n");
+		vc4_bo_stats_dump(vc4);
+	}
+}

drivers/gpu/drm/vc4/vc4_debugfs.c

@@ -16,6 +16,7 @@
 #include "vc4_regs.h"
 
 static const struct drm_info_list vc4_debugfs_list[] = {
+	{"bo_stats", vc4_bo_stats_debugfs, 0},
 	{"hdmi_regs", vc4_hdmi_debugfs_regs, 0},
 	{"hvs_regs", vc4_hvs_debugfs_regs, 0},
 	{"crtc0_regs", vc4_crtc_debugfs_regs, 0, (void *)(uintptr_t)0},

drivers/gpu/drm/vc4/vc4_drv.c

@@ -92,7 +92,8 @@ static struct drm_driver vc4_drm_driver = {
 	.debugfs_cleanup = vc4_debugfs_cleanup,
 #endif
 
-	.gem_free_object = drm_gem_cma_free_object,
+	.gem_create_object = vc4_create_object,
+	.gem_free_object = vc4_free_object,
 	.gem_vm_ops = &drm_gem_cma_vm_ops,
 
 	.prime_handle_to_fd = drm_gem_prime_handle_to_fd,
@@ -170,6 +171,8 @@ static int vc4_drm_bind(struct device *dev)
 
 	drm_dev_set_unique(drm, dev_name(dev));
 
+	vc4_bo_cache_init(drm);
+
 	drm_mode_config_init(drm);
 	if (ret)
 		goto unref;
@@ -202,6 +205,7 @@ static int vc4_drm_bind(struct device *dev)
 	component_unbind_all(dev, drm);
 unref:
 	drm_dev_unref(drm);
+	vc4_bo_cache_destroy(drm);
 	return ret;
 }
 

drivers/gpu/drm/vc4/vc4_drv.h

@@ -17,6 +17,37 @@ struct vc4_dev {
 	struct vc4_crtc *crtc[3];
 
 	struct drm_fbdev_cma *fbdev;
+
+	/* The kernel-space BO cache.  Tracks buffers that have been
+	 * unreferenced by all other users (refcounts of 0!) but not
+	 * yet freed, so we can do cheap allocations.
+	 */
+	struct vc4_bo_cache {
+		/* Array of list heads for entries in the BO cache,
+		 * based on number of pages, so we can do O(1) lookups
+		 * in the cache when allocating.
+		 */
+		struct list_head *size_list;
+		uint32_t size_list_size;
+
+		/* List of all BOs in the cache, ordered by age, so we
+		 * can do O(1) lookups when trying to free old
+		 * buffers.
+		 */
+		struct list_head time_list;
+		struct work_struct time_work;
+		struct timer_list time_timer;
+	} bo_cache;
+
+	struct vc4_bo_stats {
+		u32 num_allocated;
+		u32 size_allocated;
+		u32 num_cached;
+		u32 size_cached;
+	} bo_stats;
+
+	/* Protects bo_cache and the BO stats. */
+	struct mutex bo_lock;
 };
 
 static inline struct vc4_dev *
@@ -27,6 +58,17 @@ to_vc4_dev(struct drm_device *dev)
 
 struct vc4_bo {
 	struct drm_gem_cma_object base;
+
+	/* List entry for the BO's position in either
+	 * vc4_exec_info->unref_list or vc4_dev->bo_cache.time_list
+	 */
+	struct list_head unref_head;
+
+	/* Time in jiffies when the BO was put in vc4->bo_cache. */
+	unsigned long free_time;
+
+	/* List entry for the BO's position in vc4_dev->bo_cache.size_list */
+	struct list_head size_head;
 };
 
 static inline struct vc4_bo *
@@ -104,13 +146,18 @@ to_vc4_encoder(struct drm_encoder *encoder)
 #define wait_for(COND, MS) _wait_for(COND, MS, 1)
 
 /* vc4_bo.c */
+struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size);
 void vc4_free_object(struct drm_gem_object *gem_obj);
-struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t size);
+struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t size,
+			     bool from_cache);
 int vc4_dumb_create(struct drm_file *file_priv,
 		    struct drm_device *dev,
 		    struct drm_mode_create_dumb *args);
 struct dma_buf *vc4_prime_export(struct drm_device *dev,
 				 struct drm_gem_object *obj, int flags);
+void vc4_bo_cache_init(struct drm_device *dev);
+void vc4_bo_cache_destroy(struct drm_device *dev);
+int vc4_bo_stats_debugfs(struct seq_file *m, void *arg);
 
 /* vc4_crtc.c */
 extern struct platform_driver vc4_crtc_driver;