drm/tegra: Implement buffer object cache

This cache is used to avoid mapping and unmapping buffer objects
unnecessarily. Mappings are cached per client and stay hot until
the buffer object is destroyed.

Signed-off-by: Thierry Reding <[e-mail]>

drivers/gpu/drm/tegra/gem.c

···
 	if (!map)
 		return ERR_PTR(-ENOMEM);
 
+	kref_init(&map->ref);
 	map->bo = host1x_bo_get(bo);
 	map->direction = direction;
 	map->dev = dev;
···
 
 static void tegra_bo_unpin(struct host1x_bo_mapping *map)
 {
-	if (!map)
-		return;
-
 	if (map->attach) {
 		dma_buf_unmap_attachment(map->attach, map->sgt, map->direction);
 		dma_buf_detach(map->attach->dmabuf, map->attach);
···
 void tegra_bo_free_object(struct drm_gem_object *gem)
 {
 	struct tegra_drm *tegra = gem->dev->dev_private;
+	struct host1x_bo_mapping *mapping, *tmp;
 	struct tegra_bo *bo = to_tegra_bo(gem);
+
+	/* remove all mappings of this buffer object from any caches */
+	list_for_each_entry_safe(mapping, tmp, &bo->base.mappings, list) {
+		if (mapping->cache)
+			host1x_bo_unpin(mapping);
+		else
+			dev_err(gem->dev->dev, "mapping %p stale for device %s\n", mapping,
+				dev_name(mapping->dev));
+	}
 
 	if (tegra->domain)
 		tegra_bo_iommu_unmap(tegra, bo);

drivers/gpu/drm/tegra/plane.c

···
 		struct tegra_bo *bo = tegra_fb_get_plane(state->base.fb, i);
 		struct host1x_bo_mapping *map;
 
-		map = host1x_bo_pin(dc->dev, &bo->base, DMA_TO_DEVICE);
+		map = host1x_bo_pin(dc->dev, &bo->base, DMA_TO_DEVICE, &dc->client.cache);
 		if (IS_ERR(map)) {
 			err = PTR_ERR(map);
 			goto unpin;

drivers/gpu/drm/tegra/submit.c

···
 	if (!map)
 		return ERR_PTR(-ENOMEM);
 
+	kref_init(&map->ref);
 	map->bo = host1x_bo_get(bo);
 	map->direction = direction;
 	map->dev = dev;

drivers/gpu/drm/tegra/uapi.c

···
 		goto put_gem;
 	}
 
-	mapping->map = host1x_bo_pin(context->client->base.dev, mapping->bo, direction);
+	mapping->map = host1x_bo_pin(context->client->base.dev, mapping->bo, direction, NULL);
 	if (IS_ERR(mapping->map)) {
 		err = PTR_ERR(mapping->map);
 		goto put_gem;

drivers/gpu/host1x/bus.c

···
  */
 void __host1x_client_init(struct host1x_client *client, struct lock_class_key *key)
 {
+	host1x_bo_cache_init(&client->cache);
 	INIT_LIST_HEAD(&client->list);
 	__mutex_init(&client->lock, "host1x client lock", key);
 	client->usecount = 0;
···
 
 	mutex_unlock(&clients_lock);
 
+	host1x_bo_cache_destroy(&client->cache);
+
 	return 0;
 }
 EXPORT_SYMBOL(host1x_client_unregister);
···
 	return err;
 }
 EXPORT_SYMBOL(host1x_client_resume);
+
+struct host1x_bo_mapping *host1x_bo_pin(struct device *dev, struct host1x_bo *bo,
+					enum dma_data_direction dir,
+					struct host1x_bo_cache *cache)
+{
+	struct host1x_bo_mapping *mapping;
+
+	if (cache) {
+		mutex_lock(&cache->lock);
+
+		list_for_each_entry(mapping, &cache->mappings, entry) {
+			if (mapping->bo == bo && mapping->direction == dir) {
+				kref_get(&mapping->ref);
+				goto unlock;
+			}
+		}
+	}
+
+	mapping = bo->ops->pin(dev, bo, dir);
+	if (IS_ERR(mapping))
+		goto unlock;
+
+	spin_lock(&mapping->bo->lock);
+	list_add_tail(&mapping->list, &bo->mappings);
+	spin_unlock(&mapping->bo->lock);
+
+	if (cache) {
+		INIT_LIST_HEAD(&mapping->entry);
+		mapping->cache = cache;
+
+		list_add_tail(&mapping->entry, &cache->mappings);
+
+		/* bump reference count to track the copy in the cache */
+		kref_get(&mapping->ref);
+	}
+
+unlock:
+	if (cache)
+		mutex_unlock(&cache->lock);
+
+	return mapping;
+}
+EXPORT_SYMBOL(host1x_bo_pin);
+
+static void __host1x_bo_unpin(struct kref *ref)
+{
+	struct host1x_bo_mapping *mapping = to_host1x_bo_mapping(ref);
+
+	/*
+	 * When the last reference of the mapping goes away, make sure to remove the mapping from
+	 * the cache.
+	 */
+	if (mapping->cache)
+		list_del(&mapping->entry);
+
+	spin_lock(&mapping->bo->lock);
+	list_del(&mapping->list);
+	spin_unlock(&mapping->bo->lock);
+
+	mapping->bo->ops->unpin(mapping);
+}
+
+void host1x_bo_unpin(struct host1x_bo_mapping *mapping)
+{
+	struct host1x_bo_cache *cache = mapping->cache;
+
+	if (cache)
+		mutex_lock(&cache->lock);
+
+	kref_put(&mapping->ref, __host1x_bo_unpin);
+
+	if (cache)
+		mutex_unlock(&cache->lock);
+}
+EXPORT_SYMBOL(host1x_bo_unpin);

drivers/gpu/host1x/dev.c

···
 	if (syncpt_irq < 0)
 		return syncpt_irq;
 
+	host1x_bo_cache_init(&host->cache);
 	mutex_init(&host->devices_lock);
 	INIT_LIST_HEAD(&host->devices);
 	INIT_LIST_HEAD(&host->list);
···
 	reset_control_assert(host->rst);
 	clk_disable_unprepare(host->clk);
 	host1x_iommu_exit(host);
+	host1x_bo_cache_destroy(&host->cache);
 
 	return 0;
 }

drivers/gpu/host1x/dev.h

···
 	struct list_head list;
 
 	struct device_dma_parameters dma_parms;
+
+	struct host1x_bo_cache cache;
 };
 
 void host1x_hypervisor_writel(struct host1x *host1x, u32 r, u32 v);

drivers/gpu/host1x/job.c

···
 			goto unpin;
 		}
 
-		map = host1x_bo_pin(dev, bo, direction);
+		map = host1x_bo_pin(dev, bo, direction, &client->cache);
 		if (IS_ERR(map)) {
 			err = PTR_ERR(map);
 			goto unpin;
···
 			goto unpin;
 		}
 
-		map = host1x_bo_pin(host->dev, g->bo, DMA_TO_DEVICE);
+		map = host1x_bo_pin(host->dev, g->bo, DMA_TO_DEVICE, &host->cache);
 		if (IS_ERR(map)) {
 			err = PTR_ERR(map);
 			goto unpin;

include/linux/host1x.h

···
 
 #include <linux/device.h>
 #include <linux/dma-direction.h>
+#include <linux/spinlock.h>
 #include <linux/types.h>
 
 enum host1x_class {
···
 struct iommu_group;
 
 u64 host1x_get_dma_mask(struct host1x *host1x);
+
+/**
+ * struct host1x_bo_cache - host1x buffer object cache
+ * @mappings: list of mappings
+ * @lock: synchronizes accesses to the list of mappings
+ */
+struct host1x_bo_cache {
+	struct list_head mappings;
+	struct mutex lock;
+};
+
+static inline void host1x_bo_cache_init(struct host1x_bo_cache *cache)
+{
+	INIT_LIST_HEAD(&cache->mappings);
+	mutex_init(&cache->lock);
+}
+
+static inline void host1x_bo_cache_destroy(struct host1x_bo_cache *cache)
+{
+	/* XXX warn if not empty? */
+	mutex_destroy(&cache->lock);
+}
 
 /**
  * struct host1x_client_ops - host1x client operations
···
 	struct host1x_client *parent;
 	unsigned int usecount;
 	struct mutex lock;
+
+	struct host1x_bo_cache cache;
 };
 
 /*
···
 struct sg_table;
 
 struct host1x_bo_mapping {
+	struct kref ref;
 	struct dma_buf_attachment *attach;
 	enum dma_data_direction direction;
+	struct list_head list;
 	struct host1x_bo *bo;
 	struct sg_table *sgt;
 	unsigned int chunks;
 	struct device *dev;
 	dma_addr_t phys;
 	size_t size;
+
+	struct host1x_bo_cache *cache;
+	struct list_head entry;
 };
+
+static inline struct host1x_bo_mapping *to_host1x_bo_mapping(struct kref *ref)
+{
+	return container_of(ref, struct host1x_bo_mapping, ref);
+}
 
 struct host1x_bo_ops {
 	struct host1x_bo *(*get)(struct host1x_bo *bo);
···
 
 struct host1x_bo {
 	const struct host1x_bo_ops *ops;
+	struct list_head mappings;
+	spinlock_t lock;
 };
 
 static inline void host1x_bo_init(struct host1x_bo *bo,
 				  const struct host1x_bo_ops *ops)
 {
+	INIT_LIST_HEAD(&bo->mappings);
+	spin_lock_init(&bo->lock);
 	bo->ops = ops;
 }
 
···
 	bo->ops->put(bo);
 }
 
-static inline struct host1x_bo_mapping *host1x_bo_pin(struct device *dev, struct host1x_bo *bo,
-						      enum dma_data_direction dir)
-{
-	return bo->ops->pin(dev, bo, dir);
-}
-
-static inline void host1x_bo_unpin(struct host1x_bo_mapping *map)
-{
-	map->bo->ops->unpin(map);
-}
+struct host1x_bo_mapping *host1x_bo_pin(struct device *dev, struct host1x_bo *bo,
+					enum dma_data_direction dir,
+					struct host1x_bo_cache *cache);
+void host1x_bo_unpin(struct host1x_bo_mapping *map);
 
 static inline void *host1x_bo_mmap(struct host1x_bo *bo)
 {