path: root/drivers/gpu/drm/vc4/vc4_bo.c

// SPDX-License-Identifier: GPL-2.0-only
/*
 *  Copyright © 2015 Broadcom
 */

/**
 * DOC: VC4 GEM BO management support
 *
 * The VC4 GPU architecture (both scanout and rendering) has direct
 * 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 <linux/dma-buf.h>

#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_print(struct drm_printer *p, struct vc4_dev *vc4)
{
	int i;

	for (i = 0; i < vc4->num_labels; i++) {
		if (!vc4->bo_labels[i].num_allocated)
			continue;

		drm_printf(p, "%30s: %6dkb BOs (%d)\n",
			   vc4->bo_labels[i].name,
			   vc4->bo_labels[i].size_allocated / 1024,
			   vc4->bo_labels[i].num_allocated);
	}

	mutex_lock(&vc4->purgeable.lock);
	if (vc4->purgeable.num)
		drm_printf(p, "%30s: %6zdkb BOs (%d)\n", "userspace BO cache",
			   vc4->purgeable.size / 1024, vc4->purgeable.num);

	if (vc4->purgeable.purged_num)
		drm_printf(p, "%30s: %6zdkb BOs (%d)\n", "total purged BO",
			   vc4->purgeable.purged_size / 1024,
			   vc4->purgeable.purged_num);
	mutex_unlock(&vc4->purgeable.lock);
}

static 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 drm_printer p = drm_seq_file_printer(m);

	vc4_bo_stats_print(&p, vc4);

	return 0;
}

/* 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 =