drm/vc4: Add an ioctl for labeling GEM BOs for summary stats

This has proven immensely useful for debugging memory leaks and
overallocation (which is a rather serious concern on the platform,
given that we typically run at about 256MB of CMA out of up to 1GB
total memory, with framebuffers that are about 8MB ecah).

The state of the art without this is to dump debug logs from every GL
application, guess as to kernel allocations based on bo_stats, and try
to merge that all together into a global picture of memory allocation
state.  With this, you can add a couple of calls to the debug build of
the 3D driver and get a pretty detailed view of GPU memory usage from
/debug/dri/0/bo_stats (or when we debug print to dmesg on allocation
failure).

The Mesa side currently labels at the gallium resource level (so you
see that a 1920x20 pixmap has been created, presumably for the window
system panel), but we could extend that to be even more useful with
glObjectLabel() names being sent all the way down to the kernel.

(partial) example of sorted debugfs output with Mesa labeling all
resources:

               kernel BO cache:  16392kb BOs (3)
       tiling shadow 1920x1080:   8160kb BOs (1)
       resource 1920x1080@32/0:   8160kb BOs (1)
scanout resource 1920x1080@32/0:   8100kb BOs (1)
                        kernel:   8100kb BOs (1)

v2: Use strndup_user(), use lockdep assertion instead of just a
    comment, fix an array[-1] reference, extend comment about name
    freeing.

Signed-off-by: Eric Anholt <eric@anholt.net>
Link: https://patchwork.freedesktop.org/patch/msgid/20170725182718.31468-2-eric@anholt.net
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>

1 files changed, 204 insertions, 54 deletions

diff --git a/drivers/gpu/drm/vc4/vc4_bo.c b/drivers/gpu/drm/vc4/vc4_bo.c
index 487f96412d35..f4387e1e178d 100644
--- a/drivers/gpu/drm/vc4/vc4_bo.c
+++ b/drivers/gpu/drm/vc4/vc4_bo.c
@@ -24,21 +24,35 @@
 #include "vc4_drv.h"
 #include "uapi/drm/vc4_drm.h"
 
+static const char * const bo_type_names[] = {
+	"kernel",
+	"V3D",
+	"V3D shader",
+	"dumb",
+	"binner",
+	"RCL",
+	"BCL",
+	"kernel BO cache",
+};
+
+static bool is_user_label(int label)
+{
+	return label >= VC4_BO_TYPE_COUNT;
+}
+
 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);
+	int i;
+
+	for (i = 0; i < vc4->num_labels; i++) {
+		if (!vc4->bo_labels[i].num_allocated)
+			continue;
+
+		DRM_INFO("%30s: %6dkb BOs (%d)\n",
+			 vc4->bo_labels[i].name,
+			 vc4->bo_labels[i].size_allocated / 1024,
+			 vc4->bo_labels[i].num_allocated);
+	}
 }
 
 #ifdef CONFIG_DEBUG_FS
@@ -47,30 +61,103 @@ 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;
+	int i;
 
-	/* Take a snapshot of the current stats with the lock held. */
 	mutex_lock(&vc4->bo_lock);
-	stats = vc4->bo_stats;
+	for (i = 0; i < vc4->num_labels; i++) {
+		if (!vc4->bo_labels[i].num_allocated)
+			continue;
+
+		seq_printf(m, "%30s: %6dkb BOs (%d)\n",
+			   vc4->bo_labels[i].name,
+			   vc4->bo_labels[i].size_allocated / 1024,
+			   vc4->bo_labels[i].num_allocated);
+	}
 	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
 
+/* Takes ownership of *name and returns the appropriate slot for it in
+ * the bo_labels[] array, extending it as necessary.
+ *
+ * This is inefficient and could use a hash table instead of walking
+ * an array and strcmp()ing.  However, the assumption is that user
+ * labeling will be infrequent (scanout buffers and other long-lived
+ * objects, or debug driver builds), so we can live with it for now.
+ */
+static int vc4_get_user_label(struct vc4_dev *vc4, const char *name)
+{
+	int i;
+	int free_slot = -1;
+
+	for (i = 0; i < vc4->num_labels; i++) {
+		if (!vc4->bo_labels[i].name) {
+			free_slot = i;
+		} else if (strcmp(vc4->bo_labels[i].name, name) == 0) {
+			kfree(name);
+			return i;
+		}
+	}
+
+	if (free_slot != -1) {
+		WARN_ON(vc4->bo_labels[free_slot].num_allocated != 0);
+		vc4->bo_labels[free_slot].name = name;
+		return free_slot;
+	} else {
+		u32 new_label_count = vc4->num_labels + 1;
+		struct vc4_label *new_labels =
+			krealloc(vc4->bo_labels,
+				 new_label_count * sizeof(*new_labels),
+				 GFP_KERNEL);
+
+		if (!new_labels) {
+			kfree(name);
+			return -1;
+		}
+
+		free_slot = vc4->num_labels;
+		vc4->bo_labels = new_labels;
+		vc4->num_labels = new_label_count;
+
+		vc4->bo_labels[free_slot].name = name;
+		vc4->bo_labels[free_slot].num_allocated = 0;
+		vc4->bo_labels[free_slot].size_allocated = 0;
+
+		return free_slot;
+	}
+}
+
+static void vc4_bo_set_label(struct drm_gem_object *gem_obj, int label)
+{
+	struct vc4_bo *bo = to_vc4_bo(gem_obj);
+	struct vc4_dev *vc4 = to_vc4_dev(gem_obj->dev);
+
+	lockdep_assert_held(&vc4->bo_lock);
+
+	if (label != -1) {
+		vc4->bo_labels[label].num_allocated++;
+		vc4->bo_labels[label].size_allocated += gem_obj->size;
+	}
+
+	vc4->bo_labels[bo->label].num_allocated--;
+	vc4->bo_labels[bo->label].size_allocated -= gem_obj->size;
+
+	if (vc4->bo_labels[bo->label].num_allocated == 0 &&
+	    is_user_label(bo->label)) {
+		/* Free user BO label slots on last unreference.
+		 * Slots are just where we track the stats for a given
+		 * name, and once a name is unused we can reuse that
+		 * slot.
+		 */
+		kfree(vc4->bo_labels[bo->label].name);
+		vc4->bo_labels[bo->label].name = NULL;
+	}
+
+	bo->label = label;
+}
+
 static uint32_t bo_page_index(size_t size)
 {
 	return (size / PAGE_SIZE) - 1;
@@ -80,7 +167,8 @@ static uint32_t bo_page_index(size_t size)
 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_set_label(obj, -1);
 
 	if (bo->validated_shader) {
 		kfree(bo->validated_shader->texture_samples);
@@ -88,9 +176,6 @@ static void vc4_bo_destroy(struct vc4_bo *bo)
 		bo->validated_shader = NULL;
 	}
 
-	vc4->bo_stats.num_allocated--;
-	vc4->bo_stats.size_allocated -= obj->size;
-
 	reservation_object_fini(&bo->_resv);
 
 	drm_gem_cma_free_object(obj);
@@ -99,12 +184,6 @@ static void vc4_bo_destroy(struct vc4_bo *bo)
 /* 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);
 }
@@ -165,7 +244,8 @@ static void vc4_bo_cache_purge(struct drm_device *dev)
 }
 
 static struct vc4_bo *vc4_bo_get_from_cache(struct drm_device *dev,
-					    uint32_t size)
+					    uint32_t size,
+					    enum vc4_kernel_bo_type type)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
 	uint32_t page_index = bo_page_index(size);
@@ -186,6 +266,8 @@ static struct vc4_bo *vc4_bo_get_from_cache(struct drm_device *dev,
 	kref_init(&bo->base.base.refcount);
 
 out:
+	if (bo)
+		vc4_bo_set_label(&bo->base.base, type);
 	mutex_unlock(&vc4->bo_lock);
 	return bo;
 }
@@ -208,8 +290,9 @@ struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size)
 		return ERR_PTR(-ENOMEM);
 
 	mutex_lock(&vc4->bo_lock);
-	vc4->bo_stats.num_allocated++;
-	vc4->bo_stats.size_allocated += size;
+	bo->label = VC4_BO_TYPE_KERNEL;
+	vc4->bo_labels[VC4_BO_TYPE_KERNEL].num_allocated++;
+	vc4->bo_labels[VC4_BO_TYPE_KERNEL].size_allocated += size;
 	mutex_unlock(&vc4->bo_lock);
 	bo->resv = &bo->_resv;
 	reservation_object_init(bo->resv);
@@ -218,7 +301,7 @@ struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size)
 }
 
 struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t unaligned_size,
-			     bool allow_unzeroed)
+			     bool allow_unzeroed, enum vc4_kernel_bo_type type)
 {
 	size_t size = roundup(unaligned_size, PAGE_SIZE);
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
@@ -229,7 +312,7 @@ struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t unaligned_size,
 		return ERR_PTR(-EINVAL);
 
 	/* First, try to get a vc4_bo from the kernel BO cache. */
-	bo = vc4_bo_get_from_cache(dev, size);
+	bo = vc4_bo_get_from_cache(dev, size, type);
 	if (bo) {
 		if (!allow_unzeroed)
 			memset(bo->base.vaddr, 0, bo->base.base.size);
@@ -251,7 +334,13 @@ struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t unaligned_size,
 			return ERR_PTR(-ENOMEM);
 		}
 	}
-	return to_vc4_bo(&cma_obj->base);
+	bo = to_vc4_bo(&cma_obj->base);
+
+	mutex_lock(&vc4->bo_lock);
+	vc4_bo_set_label(&cma_obj->base, type);
+	mutex_unlock(&vc4->bo_lock);
+
+	return bo;
 }
 
 int vc4_dumb_create(struct drm_file *file_priv,
@@ -268,7 +357,7 @@ 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, args->size, false);
+	bo = vc4_bo_create(dev, args->size, false, VC4_BO_TYPE_DUMB);
 	if (IS_ERR(bo))
 		return PTR_ERR(bo);
 
@@ -348,8 +437,7 @@ void vc4_free_object(struct drm_gem_object *gem_bo)
 	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_set_label(&bo->base.base, VC4_BO_TYPE_KERNEL_CACHE);
 
 	vc4_bo_cache_free_old(dev);
 
@@ -483,7 +571,7 @@ int vc4_create_bo_ioctl(struct drm_device *dev, void *data,
 	 * We can't allocate from the BO cache, because the BOs don't
 	 * get zeroed, and that might leak data between users.
 	 */
-	bo = vc4_bo_create(dev, args->size, false);
+	bo = vc4_bo_create(dev, args->size, false, VC4_BO_TYPE_V3D);
 	if (IS_ERR(bo))
 		return PTR_ERR(bo);
 
@@ -536,7 +624,7 @@ vc4_create_shader_bo_ioctl(struct drm_device *dev, void *data,
 		return -EINVAL;
 	}
 
-	bo = vc4_bo_create(dev, args->size, true);
+	bo = vc4_bo_create(dev, args->size, true, VC4_BO_TYPE_V3D_SHADER);
 	if (IS_ERR(bo))
 		return PTR_ERR(bo);
 
@@ -651,9 +739,24 @@ int vc4_get_tiling_ioctl(struct drm_device *dev, void *data,
 	return 0;
 }
 
-void vc4_bo_cache_init(struct drm_device *dev)
+int vc4_bo_cache_init(struct drm_device *dev)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	int i;
+
+	/* Create the initial set of BO labels that the kernel will
+	 * use.  This lets us avoid a bunch of string reallocation in
+	 * the kernel's draw and BO allocation paths.
+	 */
+	vc4->bo_labels = kcalloc(VC4_BO_TYPE_COUNT, sizeof(*vc4->bo_labels),
+				 GFP_KERNEL);
+	if (!vc4->bo_labels)
+		return -ENOMEM;
+	vc4->num_labels = VC4_BO_TYPE_COUNT;
+
+	BUILD_BUG_ON(ARRAY_SIZE(bo_type_names) != VC4_BO_TYPE_COUNT);
+	for (i = 0; i < VC4_BO_TYPE_COUNT; i++)
+		vc4->bo_labels[i].name = bo_type_names[i];
 
 	mutex_init(&vc4->bo_lock);
 
@@ -663,19 +766,66 @@ void vc4_bo_cache_init(struct drm_device *dev)
 	setup_timer(&vc4->bo_cache.time_timer,
 		    vc4_bo_cache_time_timer,
 		    (unsigned long)dev);
+
+	return 0;
 }
 
 void vc4_bo_cache_destroy(struct drm_device *dev)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	int i;
 
 	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);
+	for (i = 0; i < vc4->num_labels; i++) {
+		if (vc4->bo_labels[i].num_allocated) {
+			DRM_ERROR("Destroying BO cache with %d %s "
+				  "BOs still allocated\n",
+				  vc4->bo_labels[i].num_allocated,
+				  vc4->bo_labels[i].name);
+		}
+
+		if (is_user_label(i))
+			kfree(vc4->bo_labels[i].name);
 	}
+	kfree(vc4->bo_labels);
+}
+
+int vc4_label_bo_ioctl(struct drm_device *dev, void *data,
+		       struct drm_file *file_priv)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct drm_vc4_label_bo *args = data;
+	char *name;
+	struct drm_gem_object *gem_obj;
+	int ret = 0, label;
+
+	if (!args->len)
+		return -EINVAL;
+
+	name = strndup_user(u64_to_user_ptr(args->name), args->len + 1);
+	if (IS_ERR(name))
+		return PTR_ERR(name);
+
+	gem_obj = drm_gem_object_lookup(file_priv, args->handle);
+	if (!gem_obj) {
+		DRM_ERROR("Failed to look up GEM BO %d\n", args->handle);
+		kfree(name);
+		return -ENOENT;
+	}
+
+	mutex_lock(&vc4->bo_lock);
+	label = vc4_get_user_label(vc4, name);
+	if (label != -1)
+		vc4_bo_set_label(gem_obj, label);
+	else
+		ret = -ENOMEM;
+	mutex_unlock(&vc4->bo_lock);
+
+	drm_gem_object_unreference_unlocked(gem_obj);
+
+	return ret;
 }