Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.1.
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 *  Copyright © 2015 Broadcom
   4 */
   5
   6/**
   7 * DOC: VC4 GEM BO management support
   8 *
   9 * The VC4 GPU architecture (both scanout and rendering) has direct
  10 * access to system memory with no MMU in between.  To support it, we
  11 * use the GEM CMA helper functions to allocate contiguous ranges of
  12 * physical memory for our BOs.
  13 *
  14 * Since the CMA allocator is very slow, we keep a cache of recently
  15 * freed BOs around so that the kernel's allocation of objects for 3D
  16 * rendering can return quickly.
  17 */
  18
  19#include <linux/dma-buf.h>
  20
  21#include "vc4_drv.h"
  22#include "uapi/drm/vc4_drm.h"
  23
  24static const struct drm_gem_object_funcs vc4_gem_object_funcs;
  25
  26static const char * const bo_type_names[] = {
  27	"kernel",
  28	"V3D",
  29	"V3D shader",
  30	"dumb",
  31	"binner",
  32	"RCL",
  33	"BCL",
  34	"kernel BO cache",
  35};
  36
  37static bool is_user_label(int label)
  38{
  39	return label >= VC4_BO_TYPE_COUNT;
  40}
  41
  42static void vc4_bo_stats_print(struct drm_printer *p, struct vc4_dev *vc4)
  43{
  44	int i;
  45
  46	for (i = 0; i < vc4->num_labels; i++) {
  47		if (!vc4->bo_labels[i].num_allocated)
  48			continue;
  49
  50		drm_printf(p, "%30s: %6dkb BOs (%d)\n",
  51			   vc4->bo_labels[i].name,
  52			   vc4->bo_labels[i].size_allocated / 1024,
  53			   vc4->bo_labels[i].num_allocated);
  54	}
  55
  56	mutex_lock(&vc4->purgeable.lock);
  57	if (vc4->purgeable.num)
  58		drm_printf(p, "%30s: %6zdkb BOs (%d)\n", "userspace BO cache",
  59			   vc4->purgeable.size / 1024, vc4->purgeable.num);
  60
  61	if (vc4->purgeable.purged_num)
  62		drm_printf(p, "%30s: %6zdkb BOs (%d)\n", "total purged BO",
  63			   vc4->purgeable.purged_size / 1024,
  64			   vc4->purgeable.purged_num);
  65	mutex_unlock(&vc4->purgeable.lock);
  66}
  67
  68static int vc4_bo_stats_debugfs(struct seq_file *m, void *unused)
  69{
  70	struct drm_info_node *node = (struct drm_info_node *)m->private;
  71	struct drm_device *dev = node->minor->dev;
  72	struct vc4_dev *vc4 = to_vc4_dev(dev);
  73	struct drm_printer p = drm_seq_file_printer(m);
  74
  75	vc4_bo_stats_print(&p, vc4);
  76
  77	return 0;
  78}
  79
  80/* Takes ownership of *name and returns the appropriate slot for it in
  81 * the bo_labels[] array, extending it as necessary.
  82 *
  83 * This is inefficient and could use a hash table instead of walking
  84 * an array and strcmp()ing.  However, the assumption is that user
  85 * labeling will be infrequent (scanout buffers and other long-lived
  86 * objects, or debug driver builds), so we can live with it for now.
  87 */
  88static int vc4_get_user_label(struct vc4_dev *vc4, const char *name)
  89{
  90	int i;
  91	int free_slot = -1;
  92
  93	for (i = 0; i < vc4->num_labels; i++) {
  94		if (!vc4->bo_labels[i].name) {
  95			free_slot = i;
  96		} else if (strcmp(vc4->bo_labels[i].name, name) == 0) {
  97			kfree(name);
  98			return i;
  99		}
 100	}
 101
 102	if (free_slot != -1) {
 103		WARN_ON(vc4->bo_labels[free_slot].num_allocated != 0);
 104		vc4->bo_labels[free_slot].name = name;
 105		return free_slot;
 106	} else {
 107		u32 new_label_count = vc4->num_labels + 1;
 108		struct vc4_label *new_labels =
 109			krealloc(vc4->bo_labels,
 110				 new_label_count * sizeof(*new_labels),
 111				 GFP_KERNEL);
 112
 113		if (!new_labels) {
 114			kfree(name);
 115			return -1;
 116		}
 117
 118		free_slot = vc4->num_labels;
 119		vc4->bo_labels = new_labels;
 120		vc4->num_labels = new_label_count;
 121
 122		vc4->bo_labels[free_slot].name = name;
 123		vc4->bo_labels[free_slot].num_allocated = 0;
 124		vc4->bo_labels[free_slot].size_allocated = 0;
 125
 126		return free_slot;
 127	}
 128}
 129
 130static void vc4_bo_set_label(struct drm_gem_object *gem_obj, int label)
 131{
 132	struct vc4_bo *bo = to_vc4_bo(gem_obj);
 133	struct vc4_dev *vc4 = to_vc4_dev(gem_obj->dev);
 134
 135	lockdep_assert_held(&vc4->bo_lock);
 136
 137	if (label != -1) {
 138		vc4->bo_labels[label].num_allocated++;
 139		vc4->bo_labels[label].size_allocated += gem_obj->size;
 140	}
 141
 142	vc4->bo_labels[bo->label].num_allocated--;
 143	vc4->bo_labels[bo->label].size_allocated -= gem_obj->size;
 144
 145	if (vc4->bo_labels[bo->label].num_allocated == 0 &&
 146	    is_user_label(bo->label)) {
 147		/* Free user BO label slots on last unreference.
 148		 * Slots are just where we track the stats for a given
 149		 * name, and once a name is unused we can reuse that
 150		 * slot.
 151		 */
 152		kfree(vc4->bo_labels[bo->label].name);
 153		vc4->bo_labels[bo->label].name = NULL;
 154	}
 155
 156	bo->label = label;
 157}
 158
 159static uint32_t bo_page_index(size_t size)
 160{
 161	return (size / PAGE_SIZE) - 1;
 162}
 163
 164static void vc4_bo_destroy(struct vc4_bo *bo)
 165{
 166	struct drm_gem_object *obj = &bo->base.base;
 167	struct vc4_dev *vc4 = to_vc4_dev(obj->dev);
 168
 169	lockdep_assert_held(&vc4->bo_lock);
 170
 171	vc4_bo_set_label(obj, -1);
 172
 173	if (bo->validated_shader) {
 174		kfree(bo->validated_shader->uniform_addr_offsets);
 175		kfree(bo->validated_shader->texture_samples);
 176		kfree(bo->validated_shader);
 177		bo->validated_shader = NULL;
 178	}
 179
 180	drm_gem_cma_free_object(obj);
 181}
 182
 183static void vc4_bo_remove_from_cache(struct vc4_bo *bo)
 184{
 185	struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev);
 186
 187	lockdep_assert_held(&vc4->bo_lock);
 188	list_del(&bo->unref_head);
 189	list_del(&bo->size_head);
 190}
 191
 192static struct list_head *vc4_get_cache_list_for_size(struct drm_device *dev,
 193						     size_t size)
 194{
 195	struct vc4_dev *vc4 = to_vc4_dev(dev);
 196	uint32_t page_index = bo_page_index(size);
 197
 198	if (vc4->bo_cache.size_list_size <= page_index) {
 199		uint32_t new_size = max(vc4->bo_cache.size_list_size * 2,
 200					page_index + 1);
 201		struct list_head *new_list;
 202		uint32_t i;
 203
 204		new_list = kmalloc_array(new_size, sizeof(struct list_head),
 205					 GFP_KERNEL);
 206		if (!new_list)
 207			return NULL;
 208
 209		/* Rebase the old cached BO lists to their new list
 210		 * head locations.
 211		 */
 212		for (i = 0; i < vc4->bo_cache.size_list_size; i++) {
 213			struct list_head *old_list =
 214				&vc4->bo_cache.size_list[i];
 215
 216			if (list_empty(old_list))
 217				INIT_LIST_HEAD(&new_list[i]);
 218			else
 219				list_replace(old_list, &new_list[i]);
 220		}
 221		/* And initialize the brand new BO list heads. */
 222		for (i = vc4->bo_cache.size_list_size; i < new_size; i++)
 223			INIT_LIST_HEAD(&new_list[i]);
 224
 225		kfree(vc4->bo_cache.size_list);
 226		vc4->bo_cache.size_list = new_list;
 227		vc4->bo_cache.size_list_size = new_size;
 228	}
 229
 230	return &vc4->bo_cache.size_list[page_index];
 231}
 232
 233static void vc4_bo_cache_purge(struct drm_device *dev)
 234{
 235	struct vc4_dev *vc4 = to_vc4_dev(dev);
 236
 237	mutex_lock(&vc4->bo_lock);
 238	while (!list_empty(&vc4->bo_cache.time_list)) {
 239		struct vc4_bo *bo = list_last_entry(&vc4->bo_cache.time_list,
 240						    struct vc4_bo, unref_head);
 241		vc4_bo_remove_from_cache(bo);
 242		vc4_bo_destroy(bo);
 243	}
 244	mutex_unlock(&vc4->bo_lock);
 245}
 246
 247void vc4_bo_add_to_purgeable_pool(struct vc4_bo *bo)
 248{
 249	struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev);
 250
 251	mutex_lock(&vc4->purgeable.lock);
 252	list_add_tail(&bo->size_head, &vc4->purgeable.list);
 253	vc4->purgeable.num++;
 254	vc4->purgeable.size += bo->base.base.size;
 255	mutex_unlock(&vc4->purgeable.lock);
 256}
 257
 258static void vc4_bo_remove_from_purgeable_pool_locked(struct vc4_bo *bo)
 259{
 260	struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev);
 261
 262	/* list_del_init() is used here because the caller might release
 263	 * the purgeable lock in order to acquire the madv one and update the
 264	 * madv status.
 265	 * During this short period of time a user might decide to mark
 266	 * the BO as unpurgeable, and if bo->madv is set to
 267	 * VC4_MADV_DONTNEED it will try to remove the BO from the
 268	 * purgeable list which will fail if the ->next/prev fields
 269	 * are set to LIST_POISON1/LIST_POISON2 (which is what
 270	 * list_del() does).
 271	 * Re-initializing the list element guarantees that list_del()
 272	 * will work correctly even if it's a NOP.
 273	 */
 274	list_del_init(&bo->size_head);
 275	vc4->purgeable.num--;
 276	vc4->purgeable.size -= bo->base.base.size;
 277}
 278
 279void vc4_bo_remove_from_purgeable_pool(struct vc4_bo *bo)
 280{
 281	struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev);
 282
 283	mutex_lock(&vc4->purgeable.lock);
 284	vc4_bo_remove_from_purgeable_pool_locked(bo);
 285	mutex_unlock(&vc4->purgeable.lock);
 286}
 287
 288static void vc4_bo_purge(struct drm_gem_object *obj)
 289{
 290	struct vc4_bo *bo = to_vc4_bo(obj);
 291	struct drm_device *dev = obj->dev;
 292
 293	WARN_ON(!mutex_is_locked(&bo->madv_lock));
 294	WARN_ON(bo->madv != VC4_MADV_DONTNEED);
 295
 296	drm_vma_node_unmap(&obj->vma_node, dev->anon_inode->i_mapping);
 297
 298	dma_free_wc(dev->dev, obj->size, bo->base.vaddr, bo->base.paddr);
 299	bo->base.vaddr = NULL;
 300	bo->madv = __VC4_MADV_PURGED;
 301}
 302
 303static void vc4_bo_userspace_cache_purge(struct drm_device *dev)
 304{
 305	struct vc4_dev *vc4 = to_vc4_dev(dev);
 306
 307	mutex_lock(&vc4->purgeable.lock);
 308	while (!list_empty(&vc4->purgeable.list)) {
 309		struct vc4_bo *bo = list_first_entry(&vc4->purgeable.list,
 310						     struct vc4_bo, size_head);
 311		struct drm_gem_object *obj = &bo->base.base;
 312		size_t purged_size = 0;
 313
 314		vc4_bo_remove_from_purgeable_pool_locked(bo);
 315
 316		/* Release the purgeable lock while we're purging the BO so
 317		 * that other people can continue inserting things in the
 318		 * purgeable pool without having to wait for all BOs to be
 319		 * purged.
 320		 */
 321		mutex_unlock(&vc4->purgeable.lock);
 322		mutex_lock(&bo->madv_lock);
 323
 324		/* Since we released the purgeable pool lock before acquiring
 325		 * the BO madv one, the user may have marked the BO as WILLNEED
 326		 * and re-used it in the meantime.
 327		 * Before purging the BO we need to make sure
 328		 * - it is still marked as DONTNEED
 329		 * - it has not been re-inserted in the purgeable list
 330		 * - it is not used by HW blocks
 331		 * If one of these conditions is not met, just skip the entry.
 332		 */
 333		if (bo->madv == VC4_MADV_DONTNEED &&
 334		    list_empty(&bo->size_head) &&
 335		    !refcount_read(&bo->usecnt)) {
 336			purged_size = bo->base.base.size;
 337			vc4_bo_purge(obj);
 338		}
 339		mutex_unlock(&bo->madv_lock);
 340		mutex_lock(&vc4->purgeable.lock);
 341
 342		if (purged_size) {
 343			vc4->purgeable.purged_size += purged_size;
 344			vc4->purgeable.purged_num++;
 345		}
 346	}
 347	mutex_unlock(&vc4->purgeable.lock);
 348}
 349
 350static struct vc4_bo *vc4_bo_get_from_cache(struct drm_device *dev,
 351					    uint32_t size,
 352					    enum vc4_kernel_bo_type type)
 353{
 354	struct vc4_dev *vc4 = to_vc4_dev(dev);
 355	uint32_t page_index = bo_page_index(size);
 356	struct vc4_bo *bo = NULL;
 357
 358	size = roundup(size, PAGE_SIZE);
 359
 360	mutex_lock(&vc4->bo_lock);
 361	if (page_index >= vc4->bo_cache.size_list_size)
 362		goto out;
 363
 364	if (list_empty(&vc4->bo_cache.size_list[page_index]))
 365		goto out;
 366
 367	bo = list_first_entry(&vc4->bo_cache.size_list[page_index],
 368			      struct vc4_bo, size_head);
 369	vc4_bo_remove_from_cache(bo);
 370	kref_init(&bo->base.base.refcount);
 371
 372out:
 373	if (bo)
 374		vc4_bo_set_label(&bo->base.base, type);
 375	mutex_unlock(&vc4->bo_lock);
 376	return bo;
 377}
 378
 379/**
 380 * vc4_create_object - Implementation of driver->gem_create_object.
 381 * @dev: DRM device
 382 * @size: Size in bytes of the memory the object will reference
 383 *
 384 * This lets the CMA helpers allocate object structs for us, and keep
 385 * our BO stats correct.
 386 */
 387struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size)
 388{
 389	struct vc4_dev *vc4 = to_vc4_dev(dev);
 390	struct vc4_bo *bo;
 391
 392	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
 393	if (!bo)
 394		return ERR_PTR(-ENOMEM);
 395
 396	bo->madv = VC4_MADV_WILLNEED;
 397	refcount_set(&bo->usecnt, 0);
 398	mutex_init(&bo->madv_lock);
 399	mutex_lock(&vc4->bo_lock);
 400	bo->label = VC4_BO_TYPE_KERNEL;
 401	vc4->bo_labels[VC4_BO_TYPE_KERNEL].num_allocated++;
 402	vc4->bo_labels[VC4_BO_TYPE_KERNEL].size_allocated += size;
 403	mutex_unlock(&vc4->bo_lock);
 404
 405	bo->base.base.funcs = &vc4_gem_object_funcs;
 406
 407	return &bo->base.base;
 408}
 409
 410struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t unaligned_size,
 411			     bool allow_unzeroed, enum vc4_kernel_bo_type type)
 412{
 413	size_t size = roundup(unaligned_size, PAGE_SIZE);
 414	struct vc4_dev *vc4 = to_vc4_dev(dev);
 415	struct drm_gem_cma_object *cma_obj;
 416	struct vc4_bo *bo;
 417
 418	if (size == 0)
 419		return ERR_PTR(-EINVAL);
 420
 421	/* First, try to get a vc4_bo from the kernel BO cache. */
 422	bo = vc4_bo_get_from_cache(dev, size, type);
 423	if (bo) {
 424		if (!allow_unzeroed)
 425			memset(bo->base.vaddr, 0, bo->base.base.size);
 426		return bo;
 427	}
 428
 429	cma_obj = drm_gem_cma_create(dev, size);
 430	if (IS_ERR(cma_obj)) {
 431		/*
 432		 * If we've run out of CMA memory, kill the cache of
 433		 * CMA allocations we've got laying around and try again.
 434		 */
 435		vc4_bo_cache_purge(dev);
 436		cma_obj = drm_gem_cma_create(dev, size);
 437	}
 438
 439	if (IS_ERR(cma_obj)) {
 440		/*
 441		 * Still not enough CMA memory, purge the userspace BO
 442		 * cache and retry.
 443		 * This is sub-optimal since we purge the whole userspace
 444		 * BO cache which forces user that want to re-use the BO to
 445		 * restore its initial content.
 446		 * Ideally, we should purge entries one by one and retry
 447		 * after each to see if CMA allocation succeeds. Or even
 448		 * better, try to find an entry with at least the same
 449		 * size.
 450		 */
 451		vc4_bo_userspace_cache_purge(dev);
 452		cma_obj = drm_gem_cma_create(dev, size);
 453	}
 454
 455	if (IS_ERR(cma_obj)) {
 456		struct drm_printer p = drm_info_printer(vc4->base.dev);
 457		DRM_ERROR("Failed to allocate from CMA:\n");
 458		vc4_bo_stats_print(&p, vc4);
 459		return ERR_PTR(-ENOMEM);
 460	}
 461	bo = to_vc4_bo(&cma_obj->base);
 462
 463	/* By default, BOs do not support the MADV ioctl. This will be enabled
 464	 * only on BOs that are exposed to userspace (V3D, V3D_SHADER and DUMB
 465	 * BOs).
 466	 */
 467	bo->madv = __VC4_MADV_NOTSUPP;
 468
 469	mutex_lock(&vc4->bo_lock);
 470	vc4_bo_set_label(&cma_obj->base, type);
 471	mutex_unlock(&vc4->bo_lock);
 472
 473	return bo;
 474}
 475
 476int vc4_dumb_create(struct drm_file *file_priv,
 477		    struct drm_device *dev,
 478		    struct drm_mode_create_dumb *args)
 479{
 480	int min_pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
 481	struct vc4_bo *bo = NULL;
 482	int ret;
 483
 484	if (args->pitch < min_pitch)
 485		args->pitch = min_pitch;
 486
 487	if (args->size < args->pitch * args->height)
 488		args->size = args->pitch * args->height;
 489
 490	bo = vc4_bo_create(dev, args->size, false, VC4_BO_TYPE_DUMB);
 491	if (IS_ERR(bo))
 492		return PTR_ERR(bo);
 493
 494	bo->madv = VC4_MADV_WILLNEED;
 495
 496	ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
 497	drm_gem_object_put(&bo->base.base);
 498
 499	return ret;
 500}
 501
 502static void vc4_bo_cache_free_old(struct drm_device *dev)
 503{
 504	struct vc4_dev *vc4 = to_vc4_dev(dev);
 505	unsigned long expire_time = jiffies - msecs_to_jiffies(1000);
 506
 507	lockdep_assert_held(&vc4->bo_lock);
 508
 509	while (!list_empty(&vc4->bo_cache.time_list)) {
 510		struct vc4_bo *bo = list_last_entry(&vc4->bo_cache.time_list,
 511						    struct vc4_bo, unref_head);
 512		if (time_before(expire_time, bo->free_time)) {
 513			mod_timer(&vc4->bo_cache.time_timer,
 514				  round_jiffies_up(jiffies +
 515						   msecs_to_jiffies(1000)));
 516			return;
 517		}
 518
 519		vc4_bo_remove_from_cache(bo);
 520		vc4_bo_destroy(bo);
 521	}
 522}
 523
 524/* Called on the last userspace/kernel unreference of the BO.  Returns
 525 * it to the BO cache if possible, otherwise frees it.
 526 */
 527static void vc4_free_object(struct drm_gem_object *gem_bo)
 528{
 529	struct drm_device *dev = gem_bo->dev;
 530	struct vc4_dev *vc4 = to_vc4_dev(dev);
 531	struct vc4_bo *bo = to_vc4_bo(gem_bo);
 532	struct list_head *cache_list;
 533
 534	/* Remove the BO from the purgeable list. */
 535	mutex_lock(&bo->madv_lock);
 536	if (bo->madv == VC4_MADV_DONTNEED && !refcount_read(&bo->usecnt))
 537		vc4_bo_remove_from_purgeable_pool(bo);
 538	mutex_unlock(&bo->madv_lock);
 539
 540	mutex_lock(&vc4->bo_lock);
 541	/* If the object references someone else's memory, we can't cache it.
 542	 */
 543	if (gem_bo->import_attach) {
 544		vc4_bo_destroy(bo);
 545		goto out;
 546	}
 547
 548	/* Don't cache if it was publicly named. */
 549	if (gem_bo->name) {
 550		vc4_bo_destroy(bo);
 551		goto out;
 552	}
 553
 554	/* If this object was partially constructed but CMA allocation
 555	 * had failed, just free it. Can also happen when the BO has been
 556	 * purged.
 557	 */
 558	if (!bo->base.vaddr) {
 559		vc4_bo_destroy(bo);
 560		goto out;
 561	}
 562
 563	cache_list = vc4_get_cache_list_for_size(dev, gem_bo->size);
 564	if (!cache_list) {
 565		vc4_bo_destroy(bo);
 566		goto out;
 567	}
 568
 569	if (bo->validated_shader) {
 570		kfree(bo->validated_shader->uniform_addr_offsets);
 571		kfree(bo->validated_shader->texture_samples);
 572		kfree(bo->validated_shader);
 573		bo->validated_shader = NULL;
 574	}
 575
 576	/* Reset madv and usecnt before adding the BO to the cache. */
 577	bo->madv = __VC4_MADV_NOTSUPP;
 578	refcount_set(&bo->usecnt, 0);
 579
 580	bo->t_format = false;
 581	bo->free_time = jiffies;
 582	list_add(&bo->size_head, cache_list);
 583	list_add(&bo->unref_head, &vc4->bo_cache.time_list);
 584
 585	vc4_bo_set_label(&bo->base.base, VC4_BO_TYPE_KERNEL_CACHE);
 586
 587	vc4_bo_cache_free_old(dev);
 588
 589out:
 590	mutex_unlock(&vc4->bo_lock);
 591}
 592
 593static void vc4_bo_cache_time_work(struct work_struct *work)
 594{
 595	struct vc4_dev *vc4 =
 596		container_of(work, struct vc4_dev, bo_cache.time_work);
 597	struct drm_device *dev = &vc4->base;
 598
 599	mutex_lock(&vc4->bo_lock);
 600	vc4_bo_cache_free_old(dev);
 601	mutex_unlock(&vc4->bo_lock);
 602}
 603
 604int vc4_bo_inc_usecnt(struct vc4_bo *bo)
 605{
 606	int ret;
 607
 608	/* Fast path: if the BO is already retained by someone, no need to
 609	 * check the madv status.
 610	 */
 611	if (refcount_inc_not_zero(&bo->usecnt))
 612		return 0;
 613
 614	mutex_lock(&bo->madv_lock);
 615	switch (bo->madv) {
 616	case VC4_MADV_WILLNEED:
 617		if (!refcount_inc_not_zero(&bo->usecnt))
 618			refcount_set(&bo->usecnt, 1);
 619		ret = 0;
 620		break;
 621	case VC4_MADV_DONTNEED:
 622		/* We shouldn't use a BO marked as purgeable if at least
 623		 * someone else retained its content by incrementing usecnt.
 624		 * Luckily the BO hasn't been purged yet, but something wrong
 625		 * is happening here. Just throw an error instead of
 626		 * authorizing this use case.
 627		 */
 628	case __VC4_MADV_PURGED:
 629		/* We can't use a purged BO. */
 630	default:
 631		/* Invalid madv value. */
 632		ret = -EINVAL;
 633		break;
 634	}
 635	mutex_unlock(&bo->madv_lock);
 636
 637	return ret;
 638}
 639
 640void vc4_bo_dec_usecnt(struct vc4_bo *bo)
 641{
 642	/* Fast path: if the BO is still retained by someone, no need to test
 643	 * the madv value.
 644	 */
 645	if (refcount_dec_not_one(&bo->usecnt))
 646		return;
 647
 648	mutex_lock(&bo->madv_lock);
 649	if (refcount_dec_and_test(&bo->usecnt) &&
 650	    bo->madv == VC4_MADV_DONTNEED)
 651		vc4_bo_add_to_purgeable_pool(bo);
 652	mutex_unlock(&bo->madv_lock);
 653}
 654
 655static void vc4_bo_cache_time_timer(struct timer_list *t)
 656{
 657	struct vc4_dev *vc4 = from_timer(vc4, t, bo_cache.time_timer);
 658
 659	schedule_work(&vc4->bo_cache.time_work);
 660}
 661
 662static struct dma_buf *vc4_prime_export(struct drm_gem_object *obj, int flags)
 663{
 664	struct vc4_bo *bo = to_vc4_bo(obj);
 665	struct dma_buf *dmabuf;
 666	int ret;
 667
 668	if (bo->validated_shader) {
 669		DRM_DEBUG("Attempting to export shader BO\n");
 670		return ERR_PTR(-EINVAL);
 671	}
 672
 673	/* Note: as soon as the BO is exported it becomes unpurgeable, because
 674	 * noone ever decrements the usecnt even if the reference held by the
 675	 * exported BO is released. This shouldn't be a problem since we don't
 676	 * expect exported BOs to be marked as purgeable.
 677	 */
 678	ret = vc4_bo_inc_usecnt(bo);
 679	if (ret) {
 680		DRM_ERROR("Failed to increment BO usecnt\n");
 681		return ERR_PTR(ret);
 682	}
 683
 684	dmabuf = drm_gem_prime_export(obj, flags);
 685	if (IS_ERR(dmabuf))
 686		vc4_bo_dec_usecnt(bo);
 687
 688	return dmabuf;
 689}
 690
 691static vm_fault_t vc4_fault(struct vm_fault *vmf)
 692{
 693	struct vm_area_struct *vma = vmf->vma;
 694	struct drm_gem_object *obj = vma->vm_private_data;
 695	struct vc4_bo *bo = to_vc4_bo(obj);
 696
 697	/* The only reason we would end up here is when user-space accesses
 698	 * BO's memory after it's been purged.
 699	 */
 700	mutex_lock(&bo->madv_lock);
 701	WARN_ON(bo->madv != __VC4_MADV_PURGED);
 702	mutex_unlock(&bo->madv_lock);
 703
 704	return VM_FAULT_SIGBUS;
 705}
 706
 707static int vc4_gem_object_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
 708{
 709	struct vc4_bo *bo = to_vc4_bo(obj);
 710
 711	if (bo->validated_shader && (vma->vm_flags & VM_WRITE)) {
 712		DRM_DEBUG("mmaping of shader BOs for writing not allowed.\n");
 713		return -EINVAL;
 714	}
 715
 716	if (bo->madv != VC4_MADV_WILLNEED) {
 717		DRM_DEBUG("mmaping of %s BO not allowed\n",
 718			  bo->madv == VC4_MADV_DONTNEED ?
 719			  "purgeable" : "purged");
 720		return -EINVAL;
 721	}
 722
 723	return drm_gem_cma_mmap(obj, vma);
 724}
 725
 726static const struct vm_operations_struct vc4_vm_ops = {
 727	.fault = vc4_fault,
 728	.open = drm_gem_vm_open,
 729	.close = drm_gem_vm_close,
 730};
 731
 732static const struct drm_gem_object_funcs vc4_gem_object_funcs = {
 733	.free = vc4_free_object,
 734	.export = vc4_prime_export,
 735	.get_sg_table = drm_gem_cma_get_sg_table,
 736	.vmap = drm_gem_cma_vmap,
 737	.mmap = vc4_gem_object_mmap,
 738	.vm_ops = &vc4_vm_ops,
 739};
 740
 741static int vc4_grab_bin_bo(struct vc4_dev *vc4, struct vc4_file *vc4file)
 742{
 743	int ret;
 744
 745	if (!vc4->v3d)
 746		return -ENODEV;
 747
 748	if (vc4file->bin_bo_used)
 749		return 0;
 750
 751	ret = vc4_v3d_bin_bo_get(vc4, &vc4file->bin_bo_used);
 752	if (ret)
 753		return ret;
 754
 755	return 0;
 756}
 757
 758int vc4_create_bo_ioctl(struct drm_device *dev, void *data,
 759			struct drm_file *file_priv)
 760{
 761	struct drm_vc4_create_bo *args = data;
 762	struct vc4_file *vc4file = file_priv->driver_priv;
 763	struct vc4_dev *vc4 = to_vc4_dev(dev);
 764	struct vc4_bo *bo = NULL;
 765	int ret;
 766
 767	ret = vc4_grab_bin_bo(vc4, vc4file);
 768	if (ret)
 769		return ret;
 770
 771	/*
 772	 * We can't allocate from the BO cache, because the BOs don't
 773	 * get zeroed, and that might leak data between users.
 774	 */
 775	bo = vc4_bo_create(dev, args->size, false, VC4_BO_TYPE_V3D);
 776	if (IS_ERR(bo))
 777		return PTR_ERR(bo);
 778
 779	bo->madv = VC4_MADV_WILLNEED;
 780
 781	ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
 782	drm_gem_object_put(&bo->base.base);
 783
 784	return ret;
 785}
 786
 787int vc4_mmap_bo_ioctl(struct drm_device *dev, void *data,
 788		      struct drm_file *file_priv)
 789{
 790	struct drm_vc4_mmap_bo *args = data;
 791	struct drm_gem_object *gem_obj;
 792
 793	gem_obj = drm_gem_object_lookup(file_priv, args->handle);
 794	if (!gem_obj) {
 795		DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
 796		return -EINVAL;
 797	}
 798
 799	/* The mmap offset was set up at BO allocation time. */
 800	args->offset = drm_vma_node_offset_addr(&gem_obj->vma_node);
 801
 802	drm_gem_object_put(gem_obj);
 803	return 0;
 804}
 805
 806int
 807vc4_create_shader_bo_ioctl(struct drm_device *dev, void *data,
 808			   struct drm_file *file_priv)
 809{
 810	struct drm_vc4_create_shader_bo *args = data;
 811	struct vc4_file *vc4file = file_priv->driver_priv;
 812	struct vc4_dev *vc4 = to_vc4_dev(dev);
 813	struct vc4_bo *bo = NULL;
 814	int ret;
 815
 816	if (args->size == 0)
 817		return -EINVAL;
 818
 819	if (args->size % sizeof(u64) != 0)
 820		return -EINVAL;
 821
 822	if (args->flags != 0) {
 823		DRM_INFO("Unknown flags set: 0x%08x\n", args->flags);
 824		return -EINVAL;
 825	}
 826
 827	if (args->pad != 0) {
 828		DRM_INFO("Pad set: 0x%08x\n", args->pad);
 829		return -EINVAL;
 830	}
 831
 832	ret = vc4_grab_bin_bo(vc4, vc4file);
 833	if (ret)
 834		return ret;
 835
 836	bo = vc4_bo_create(dev, args->size, true, VC4_BO_TYPE_V3D_SHADER);
 837	if (IS_ERR(bo))
 838		return PTR_ERR(bo);
 839
 840	bo->madv = VC4_MADV_WILLNEED;
 841
 842	if (copy_from_user(bo->base.vaddr,
 843			     (void __user *)(uintptr_t)args->data,
 844			     args->size)) {
 845		ret = -EFAULT;
 846		goto fail;
 847	}
 848	/* Clear the rest of the memory from allocating from the BO
 849	 * cache.
 850	 */
 851	memset(bo->base.vaddr + args->size, 0,
 852	       bo->base.base.size - args->size);
 853
 854	bo->validated_shader = vc4_validate_shader(&bo->base);
 855	if (!bo->validated_shader) {
 856		ret = -EINVAL;
 857		goto fail;
 858	}
 859
 860	/* We have to create the handle after validation, to avoid
 861	 * races for users to do doing things like mmap the shader BO.
 862	 */
 863	ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
 864
 865fail:
 866	drm_gem_object_put(&bo->base.base);
 867
 868	return ret;
 869}
 870
 871/**
 872 * vc4_set_tiling_ioctl() - Sets the tiling modifier for a BO.
 873 * @dev: DRM device
 874 * @data: ioctl argument
 875 * @file_priv: DRM file for this fd
 876 *
 877 * The tiling state of the BO decides the default modifier of an fb if
 878 * no specific modifier was set by userspace, and the return value of
 879 * vc4_get_tiling_ioctl() (so that userspace can treat a BO it
 880 * received from dmabuf as the same tiling format as the producer
 881 * used).
 882 */
 883int vc4_set_tiling_ioctl(struct drm_device *dev, void *data,
 884			 struct drm_file *file_priv)
 885{
 886	struct drm_vc4_set_tiling *args = data;
 887	struct drm_gem_object *gem_obj;
 888	struct vc4_bo *bo;
 889	bool t_format;
 890
 891	if (args->flags != 0)
 892		return -EINVAL;
 893
 894	switch (args->modifier) {
 895	case DRM_FORMAT_MOD_NONE:
 896		t_format = false;
 897		break;
 898	case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED:
 899		t_format = true;
 900		break;
 901	default:
 902		return -EINVAL;
 903	}
 904
 905	gem_obj = drm_gem_object_lookup(file_priv, args->handle);
 906	if (!gem_obj) {
 907		DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
 908		return -ENOENT;
 909	}
 910	bo = to_vc4_bo(gem_obj);
 911	bo->t_format = t_format;
 912
 913	drm_gem_object_put(gem_obj);
 914
 915	return 0;
 916}
 917
 918/**
 919 * vc4_get_tiling_ioctl() - Gets the tiling modifier for a BO.
 920 * @dev: DRM device
 921 * @data: ioctl argument
 922 * @file_priv: DRM file for this fd
 923 *
 924 * Returns the tiling modifier for a BO as set by vc4_set_tiling_ioctl().
 925 */
 926int vc4_get_tiling_ioctl(struct drm_device *dev, void *data,
 927			 struct drm_file *file_priv)
 928{
 929	struct drm_vc4_get_tiling *args = data;
 930	struct drm_gem_object *gem_obj;
 931	struct vc4_bo *bo;
 932
 933	if (args->flags != 0 || args->modifier != 0)
 934		return -EINVAL;
 935
 936	gem_obj = drm_gem_object_lookup(file_priv, args->handle);
 937	if (!gem_obj) {
 938		DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
 939		return -ENOENT;
 940	}
 941	bo = to_vc4_bo(gem_obj);
 942
 943	if (bo->t_format)
 944		args->modifier = DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED;
 945	else
 946		args->modifier = DRM_FORMAT_MOD_NONE;
 947
 948	drm_gem_object_put(gem_obj);
 949
 950	return 0;
 951}
 952
 953static void vc4_bo_cache_destroy(struct drm_device *dev, void *unused);
 954int vc4_bo_cache_init(struct drm_device *dev)
 955{
 956	struct vc4_dev *vc4 = to_vc4_dev(dev);
 957	int i;
 958
 959	/* Create the initial set of BO labels that the kernel will
 960	 * use.  This lets us avoid a bunch of string reallocation in
 961	 * the kernel's draw and BO allocation paths.
 962	 */
 963	vc4->bo_labels = kcalloc(VC4_BO_TYPE_COUNT, sizeof(*vc4->bo_labels),
 964				 GFP_KERNEL);
 965	if (!vc4->bo_labels)
 966		return -ENOMEM;
 967	vc4->num_labels = VC4_BO_TYPE_COUNT;
 968
 969	BUILD_BUG_ON(ARRAY_SIZE(bo_type_names) != VC4_BO_TYPE_COUNT);
 970	for (i = 0; i < VC4_BO_TYPE_COUNT; i++)
 971		vc4->bo_labels[i].name = bo_type_names[i];
 972
 973	mutex_init(&vc4->bo_lock);
 974
 975	vc4_debugfs_add_file(dev, "bo_stats", vc4_bo_stats_debugfs, NULL);
 976
 977	INIT_LIST_HEAD(&vc4->bo_cache.time_list);
 978
 979	INIT_WORK(&vc4->bo_cache.time_work, vc4_bo_cache_time_work);
 980	timer_setup(&vc4->bo_cache.time_timer, vc4_bo_cache_time_timer, 0);
 981
 982	return drmm_add_action_or_reset(dev, vc4_bo_cache_destroy, NULL);
 983}
 984
 985static void vc4_bo_cache_destroy(struct drm_device *dev, void *unused)
 986{
 987	struct vc4_dev *vc4 = to_vc4_dev(dev);
 988	int i;
 989
 990	del_timer(&vc4->bo_cache.time_timer);
 991	cancel_work_sync(&vc4->bo_cache.time_work);
 992
 993	vc4_bo_cache_purge(dev);
 994
 995	for (i = 0; i < vc4->num_labels; i++) {
 996		if (vc4->bo_labels[i].num_allocated) {
 997			DRM_ERROR("Destroying BO cache with %d %s "
 998				  "BOs still allocated\n",
 999				  vc4->bo_labels[i].num_allocated,
1000				  vc4->bo_labels[i].name);
1001		}
1002
1003		if (is_user_label(i))
1004			kfree(vc4->bo_labels[i].name);
1005	}
1006	kfree(vc4->bo_labels);
1007}
1008
1009int vc4_label_bo_ioctl(struct drm_device *dev, void *data,
1010		       struct drm_file *file_priv)
1011{
1012	struct vc4_dev *vc4 = to_vc4_dev(dev);
1013	struct drm_vc4_label_bo *args = data;
1014	char *name;
1015	struct drm_gem_object *gem_obj;
1016	int ret = 0, label;
1017
1018	if (!args->len)
1019		return -EINVAL;
1020
1021	name = strndup_user(u64_to_user_ptr(args->name), args->len + 1);
1022	if (IS_ERR(name))
1023		return PTR_ERR(name);
1024
1025	gem_obj = drm_gem_object_lookup(file_priv, args->handle);
1026	if (!gem_obj) {
1027		DRM_ERROR("Failed to look up GEM BO %d\n", args->handle);
1028		kfree(name);
1029		return -ENOENT;
1030	}
1031
1032	mutex_lock(&vc4->bo_lock);
1033	label = vc4_get_user_label(vc4, name);
1034	if (label != -1)
1035		vc4_bo_set_label(gem_obj, label);
1036	else
1037		ret = -ENOMEM;
1038	mutex_unlock(&vc4->bo_lock);
1039
1040	drm_gem_object_put(gem_obj);
1041
1042	return ret;
1043}