drivers/gpu/drm/vc4/vc4_bo.c

   1 /*
   2  *  Copyright © 2015 Broadcom
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License version 2 as
   6  * published by the Free Software Foundation.
   7  */
   8
   9 /**
  10  * DOC: VC4 GEM BO management support
  11  *
  12  * The VC4 GPU architecture (both scanout and rendering) has direct
  13  * access to system memory with no MMU in between.  To support it, we
  14  * use the GEM CMA helper functions to allocate contiguous ranges of
  15  * physical memory for our BOs.
  16  *
  17  * Since the CMA allocator is very slow, we keep a cache of recently
  18  * freed BOs around so that the kernel's allocation of objects for 3D
  19  * rendering can return quickly.
  20  */
  21
  22 #include "vc4_drv.h"
  23 #include "uapi/drm/vc4_drm.h"
  24
  25 static void vc4_bo_stats_dump(struct vc4_dev *vc4)
  26 {
  27         DRM_INFO("num bos allocated: %d\n",
  28                  vc4->bo_stats.num_allocated);
  29         DRM_INFO("size bos allocated: %dkb\n",
  30                  vc4->bo_stats.size_allocated / 1024);
  31         DRM_INFO("num bos used: %d\n",
  32                  vc4->bo_stats.num_allocated - vc4->bo_stats.num_cached);
  33         DRM_INFO("size bos used: %dkb\n",
  34                  (vc4->bo_stats.size_allocated -
  35                   vc4->bo_stats.size_cached) / 1024);
  36         DRM_INFO("num bos cached: %d\n",
  37                  vc4->bo_stats.num_cached);
  38         DRM_INFO("size bos cached: %dkb\n",
  39                  vc4->bo_stats.size_cached / 1024);
  40 }
  41
  42 #ifdef CONFIG_DEBUG_FS
  43 int vc4_bo_stats_debugfs(struct seq_file *m, void *unused)
  44 {
  45         struct drm_info_node *node = (struct drm_info_node *)m->private;
  46         struct drm_device *dev = node->minor->dev;
  47         struct vc4_dev *vc4 = to_vc4_dev(dev);
  48         struct vc4_bo_stats stats;
  49
  50         /* Take a snapshot of the current stats with the lock held. */
  51         mutex_lock(&vc4->bo_lock);
  52         stats = vc4->bo_stats;
  53         mutex_unlock(&vc4->bo_lock);
  54
  55         seq_printf(m, "num bos allocated: %d\n",
  56                    stats.num_allocated);
  57         seq_printf(m, "size bos allocated: %dkb\n",
  58                    stats.size_allocated / 1024);
  59         seq_printf(m, "num bos used: %d\n",
  60                    stats.num_allocated - stats.num_cached);
  61         seq_printf(m, "size bos used: %dkb\n",
  62                    (stats.size_allocated - stats.size_cached) / 1024);
  63         seq_printf(m, "num bos cached: %d\n",
  64                    stats.num_cached);
  65         seq_printf(m, "size bos cached: %dkb\n",
  66                    stats.size_cached / 1024);
  67
  68         return 0;
  69 }
  70 #endif
  71
  72 static uint32_t bo_page_index(size_t size)
  73 {
  74         return (size / PAGE_SIZE) - 1;
  75 }
  76
  77 /* Must be called with bo_lock held. */
  78 static void vc4_bo_destroy(struct vc4_bo *bo)
  79 {
  80         struct drm_gem_object *obj = &bo->base.base;
  81         struct vc4_dev *vc4 = to_vc4_dev(obj->dev);
  82
  83         if (bo->validated_shader) {
  84                 kfree(bo->validated_shader->texture_samples);
  85                 kfree(bo->validated_shader);
  86                 bo->validated_shader = NULL;
  87         }
  88
  89         vc4->bo_stats.num_allocated--;
  90         vc4->bo_stats.size_allocated -= obj->size;
  91         drm_gem_cma_free_object(obj);
  92 }
  93
  94 /* Must be called with bo_lock held. */
  95 static void vc4_bo_remove_from_cache(struct vc4_bo *bo)
  96 {
  97         struct drm_gem_object *obj = &bo->base.base;
  98         struct vc4_dev *vc4 = to_vc4_dev(obj->dev);
  99
 100         vc4->bo_stats.num_cached--;
 101         vc4->bo_stats.size_cached -= obj->size;
 102
 103         list_del(&bo->unref_head);
 104         list_del(&bo->size_head);
 105 }
 106
 107 static struct list_head *vc4_get_cache_list_for_size(struct drm_device *dev,
 108                                                      size_t size)
 109 {
 110         struct vc4_dev *vc4 = to_vc4_dev(dev);
 111         uint32_t page_index = bo_page_index(size);
 112
 113         if (vc4->bo_cache.size_list_size <= page_index) {
 114                 uint32_t new_size = max(vc4->bo_cache.size_list_size * 2,
 115                                         page_index + 1);
 116                 struct list_head *new_list;
 117                 uint32_t i;
 118
 119                 new_list = kmalloc_array(new_size, sizeof(struct list_head),
 120                                          GFP_KERNEL);
 121                 if (!new_list)
 122                         return NULL;
 123
 124                 /* Rebase the old cached BO lists to their new list
 125                  * head locations.
 126                  */
 127                 for (i = 0; i < vc4->bo_cache.size_list_size; i++) {
 128                         struct list_head *old_list =
 129                                 &vc4->bo_cache.size_list[i];
 130
 131                         if (list_empty(old_list))
 132                                 INIT_LIST_HEAD(&new_list[i]);
 133                         else
 134                                 list_replace(old_list, &new_list[i]);
 135                 }
 136                 /* And initialize the brand new BO list heads. */
 137                 for (i = vc4->bo_cache.size_list_size; i < new_size; i++)
 138                         INIT_LIST_HEAD(&new_list[i]);
 139
 140                 kfree(vc4->bo_cache.size_list);
 141                 vc4->bo_cache.size_list = new_list;
 142                 vc4->bo_cache.size_list_size = new_size;
 143         }
 144
 145         return &vc4->bo_cache.size_list[page_index];
 146 }
 147
 148 static void vc4_bo_cache_purge(struct drm_device *dev)
 149 {
 150         struct vc4_dev *vc4 = to_vc4_dev(dev);
 151
 152         mutex_lock(&vc4->bo_lock);
 153         while (!list_empty(&vc4->bo_cache.time_list)) {
 154                 struct vc4_bo *bo = list_last_entry(&vc4->bo_cache.time_list,
 155                                                     struct vc4_bo, unref_head);
 156                 vc4_bo_remove_from_cache(bo);
 157                 vc4_bo_destroy(bo);
 158         }
 159         mutex_unlock(&vc4->bo_lock);
 160 }
 161
 162 static struct vc4_bo *vc4_bo_get_from_cache(struct drm_device *dev,
 163                                             uint32_t size)
 164 {
 165         struct vc4_dev *vc4 = to_vc4_dev(dev);
 166         uint32_t page_index = bo_page_index(size);
 167         struct vc4_bo *bo = NULL;
 168
 169         size = roundup(size, PAGE_SIZE);
 170
 171         mutex_lock(&vc4->bo_lock);
 172         if (page_index >= vc4->bo_cache.size_list_size)
 173                 goto out;
 174
 175         if (list_empty(&vc4->bo_cache.size_list[page_index]))
 176                 goto out;
 177
 178         bo = list_first_entry(&vc4->bo_cache.size_list[page_index],
 179                               struct vc4_bo, size_head);
 180         vc4_bo_remove_from_cache(bo);
 181         kref_init(&bo->base.base.refcount);
 182
 183 out:
 184         mutex_unlock(&vc4->bo_lock);
 185         return bo;
 186 }
 187
 188 /**
 189  * vc4_gem_create_object - Implementation of driver->gem_create_object.
 190  * @dev: DRM device
 191  * @size: Size in bytes of the memory the object will reference
 192  *
 193  * This lets the CMA helpers allocate object structs for us, and keep
 194  * our BO stats correct.
 195  */
 196 struct drm_gem_object *vc4_create_object(struct drm_device *dev, size_t size)
 197 {
 198         struct vc4_dev *vc4 = to_vc4_dev(dev);
 199         struct vc4_bo *bo;
 200
 201         bo = kzalloc(sizeof(*bo), GFP_KERNEL);
 202         if (!bo)
 203                 return ERR_PTR(-ENOMEM);
 204
 205         mutex_lock(&vc4->bo_lock);
 206         vc4->bo_stats.num_allocated++;
 207         vc4->bo_stats.size_allocated += size;
 208         mutex_unlock(&vc4->bo_lock);
 209
 210         return &bo->base.base;
 211 }
 212
 213 struct vc4_bo *vc4_bo_create(struct drm_device *dev, size_t unaligned_size,
 214                              bool allow_unzeroed)
 215 {
 216         size_t size = roundup(unaligned_size, PAGE_SIZE);
 217         struct vc4_dev *vc4 = to_vc4_dev(dev);
 218         struct drm_gem_cma_object *cma_obj;
 219         struct vc4_bo *bo;
 220
 221         if (size == 0)
 222                 return ERR_PTR(-EINVAL);
 223
 224         /* First, try to get a vc4_bo from the kernel BO cache. */
 225         bo = vc4_bo_get_from_cache(dev, size);
 226         if (bo) {
 227                 if (!allow_unzeroed)
 228                         memset(bo->base.vaddr, 0, bo->base.base.size);
 229                 return bo;
 230         }
 231
 232         cma_obj = drm_gem_cma_create(dev, size);
 233         if (IS_ERR(cma_obj)) {
 234                 /*
 235                  * If we've run out of CMA memory, kill the cache of
 236                  * CMA allocations we've got laying around and try again.
 237                  */
 238                 vc4_bo_cache_purge(dev);
 239
 240                 cma_obj = drm_gem_cma_create(dev, size);
 241                 if (IS_ERR(cma_obj)) {
 242                         DRM_ERROR("Failed to allocate from CMA:\n");
 243                         vc4_bo_stats_dump(vc4);
 244                         return ERR_PTR(-ENOMEM);
 245                 }
 246         }
 247
 248         return to_vc4_bo(&cma_obj->base);
 249 }
 250
 251 int vc4_dumb_create(struct drm_file *file_priv,
 252                     struct drm_device *dev,
 253                     struct drm_mode_create_dumb *args)
 254 {
 255         int min_pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
 256         struct vc4_bo *bo = NULL;
 257         int ret;
 258
 259         if (args->pitch < min_pitch)
 260                 args->pitch = min_pitch;
 261
 262         if (args->size < args->pitch * args->height)
 263                 args->size = args->pitch * args->height;
 264
 265         bo = vc4_bo_create(dev, args->size, false);
 266         if (IS_ERR(bo))
 267                 return PTR_ERR(bo);
 268
 269         ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
 270         drm_gem_object_unreference_unlocked(&bo->base.base);
 271
 272         return ret;
 273 }
 274
 275 /* Must be called with bo_lock held. */
 276 static void vc4_bo_cache_free_old(struct drm_device *dev)
 277 {
 278         struct vc4_dev *vc4 = to_vc4_dev(dev);
 279         unsigned long expire_time = jiffies - msecs_to_jiffies(1000);
 280
 281         while (!list_empty(&vc4->bo_cache.time_list)) {
 282                 struct vc4_bo *bo = list_last_entry(&vc4->bo_cache.time_list,
 283                                                     struct vc4_bo, unref_head);
 284                 if (time_before(expire_time, bo->free_time)) {
 285                         mod_timer(&vc4->bo_cache.time_timer,
 286                                   round_jiffies_up(jiffies +
 287                                                    msecs_to_jiffies(1000)));
 288                         return;
 289                 }
 290
 291                 vc4_bo_remove_from_cache(bo);
 292                 vc4_bo_destroy(bo);
 293         }
 294 }
 295
 296 /* Called on the last userspace/kernel unreference of the BO.  Returns
 297  * it to the BO cache if possible, otherwise frees it.
 298  */
 299 void vc4_free_object(struct drm_gem_object *gem_bo)
 300 {
 301         struct drm_device *dev = gem_bo->dev;
 302         struct vc4_dev *vc4 = to_vc4_dev(dev);
 303         struct vc4_bo *bo = to_vc4_bo(gem_bo);
 304         struct list_head *cache_list;
 305
 306         mutex_lock(&vc4->bo_lock);
 307         /* If the object references someone else's memory, we can't cache it.
 308          */
 309         if (gem_bo->import_attach) {
 310                 vc4_bo_destroy(bo);
 311                 goto out;
 312         }
 313
 314         /* Don't cache if it was publicly named. */
 315         if (gem_bo->name) {
 316                 vc4_bo_destroy(bo);
 317                 goto out;
 318         }
 319
 320         /* If this object was partially constructed but CMA allocation
 321          * had failed, just free it.
 322          */
 323         if (!bo->base.vaddr) {
 324                 vc4_bo_destroy(bo);
 325                 goto out;
 326         }
 327
 328         cache_list = vc4_get_cache_list_for_size(dev, gem_bo->size);
 329         if (!cache_list) {
 330                 vc4_bo_destroy(bo);
 331                 goto out;
 332         }
 333
 334         if (bo->validated_shader) {
 335                 kfree(bo->validated_shader->texture_samples);
 336                 kfree(bo->validated_shader);
 337                 bo->validated_shader = NULL;
 338         }
 339
 340         bo->free_time = jiffies;
 341         list_add(&bo->size_head, cache_list);
 342         list_add(&bo->unref_head, &vc4->bo_cache.time_list);
 343
 344         vc4->bo_stats.num_cached++;
 345         vc4->bo_stats.size_cached += gem_bo->size;
 346
 347         vc4_bo_cache_free_old(dev);
 348
 349 out:
 350         mutex_unlock(&vc4->bo_lock);
 351 }
 352
 353 static void vc4_bo_cache_time_work(struct work_struct *work)
 354 {
 355         struct vc4_dev *vc4 =
 356                 container_of(work, struct vc4_dev, bo_cache.time_work);
 357         struct drm_device *dev = vc4->dev;
 358
 359         mutex_lock(&vc4->bo_lock);
 360         vc4_bo_cache_free_old(dev);
 361         mutex_unlock(&vc4->bo_lock);
 362 }
 363
 364 static void vc4_bo_cache_time_timer(unsigned long data)
 365 {
 366         struct drm_device *dev = (struct drm_device *)data;
 367         struct vc4_dev *vc4 = to_vc4_dev(dev);
 368
 369         schedule_work(&vc4->bo_cache.time_work);
 370 }
 371
 372 struct dma_buf *
 373 vc4_prime_export(struct drm_device *dev, struct drm_gem_object *obj, int flags)
 374 {
 375         struct vc4_bo *bo = to_vc4_bo(obj);
 376
 377         if (bo->validated_shader) {
 378                 DRM_ERROR("Attempting to export shader BO\n");
 379                 return ERR_PTR(-EINVAL);
 380         }
 381
 382         return drm_gem_prime_export(dev, obj, flags);
 383 }
 384
 385 int vc4_mmap(struct file *filp, struct vm_area_struct *vma)
 386 {
 387         struct drm_gem_object *gem_obj;
 388         struct vc4_bo *bo;
 389         int ret;
 390
 391         ret = drm_gem_mmap(filp, vma);
 392         if (ret)
 393                 return ret;
 394
 395         gem_obj = vma->vm_private_data;
 396         bo = to_vc4_bo(gem_obj);
 397
 398         if (bo->validated_shader && (vma->vm_flags & VM_WRITE)) {
 399                 DRM_ERROR("mmaping of shader BOs for writing not allowed.\n");
 400                 return -EINVAL;
 401         }
 402
 403         /*
 404          * Clear the VM_PFNMAP flag that was set by drm_gem_mmap(), and set the
 405          * vm_pgoff (used as a fake buffer offset by DRM) to 0 as we want to map
 406          * the whole buffer.
 407          */
 408         vma->vm_flags &= ~VM_PFNMAP;
 409         vma->vm_pgoff = 0;
 410
 411         ret = dma_mmap_wc(bo->base.base.dev->dev, vma, bo->base.vaddr,
 412                           bo->base.paddr, vma->vm_end - vma->vm_start);
 413         if (ret)
 414                 drm_gem_vm_close(vma);
 415
 416         return ret;
 417 }
 418
 419 int vc4_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
 420 {
 421         struct vc4_bo *bo = to_vc4_bo(obj);
 422
 423         if (bo->validated_shader && (vma->vm_flags & VM_WRITE)) {
 424                 DRM_ERROR("mmaping of shader BOs for writing not allowed.\n");
 425                 return -EINVAL;
 426         }
 427
 428         return drm_gem_cma_prime_mmap(obj, vma);
 429 }
 430
 431 void *vc4_prime_vmap(struct drm_gem_object *obj)
 432 {
 433         struct vc4_bo *bo = to_vc4_bo(obj);
 434
 435         if (bo->validated_shader) {
 436                 DRM_ERROR("mmaping of shader BOs not allowed.\n");
 437                 return ERR_PTR(-EINVAL);
 438         }
 439
 440         return drm_gem_cma_prime_vmap(obj);
 441 }
 442
 443 int vc4_create_bo_ioctl(struct drm_device *dev, void *data,
 444                         struct drm_file *file_priv)
 445 {
 446         struct drm_vc4_create_bo *args = data;
 447         struct vc4_bo *bo = NULL;
 448         int ret;
 449
 450         /*
 451          * We can't allocate from the BO cache, because the BOs don't
 452          * get zeroed, and that might leak data between users.
 453          */
 454         bo = vc4_bo_create(dev, args->size, false);
 455         if (IS_ERR(bo))
 456                 return PTR_ERR(bo);
 457
 458         ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
 459         drm_gem_object_unreference_unlocked(&bo->base.base);
 460
 461         return ret;
 462 }
 463
 464 int vc4_mmap_bo_ioctl(struct drm_device *dev, void *data,
 465                       struct drm_file *file_priv)
 466 {
 467         struct drm_vc4_mmap_bo *args = data;
 468         struct drm_gem_object *gem_obj;
 469
 470         gem_obj = drm_gem_object_lookup(file_priv, args->handle);
 471         if (!gem_obj) {
 472                 DRM_ERROR("Failed to look up GEM BO %d\n", args->handle);
 473                 return -EINVAL;
 474         }
 475
 476         /* The mmap offset was set up at BO allocation time. */
 477         args->offset = drm_vma_node_offset_addr(&gem_obj->vma_node);
 478
 479         drm_gem_object_unreference_unlocked(gem_obj);
 480         return 0;
 481 }
 482
 483 int
 484 vc4_create_shader_bo_ioctl(struct drm_device *dev, void *data,
 485                            struct drm_file *file_priv)
 486 {
 487         struct drm_vc4_create_shader_bo *args = data;
 488         struct vc4_bo *bo = NULL;
 489         int ret;
 490
 491         if (args->size == 0)
 492                 return -EINVAL;
 493
 494         if (args->size % sizeof(u64) != 0)
 495                 return -EINVAL;
 496
 497         if (args->flags != 0) {
 498                 DRM_INFO("Unknown flags set: 0x%08x\n", args->flags);
 499                 return -EINVAL;
 500         }
 501
 502         if (args->pad != 0) {
 503                 DRM_INFO("Pad set: 0x%08x\n", args->pad);
 504                 return -EINVAL;
 505         }
 506
 507         bo = vc4_bo_create(dev, args->size, true);
 508         if (IS_ERR(bo))
 509                 return PTR_ERR(bo);
 510
 511         if (copy_from_user(bo->base.vaddr,
 512                              (void __user *)(uintptr_t)args->data,
 513                              args->size)) {
 514                 ret = -EFAULT;
 515                 goto fail;
 516         }
 517         /* Clear the rest of the memory from allocating from the BO
 518          * cache.
 519          */
 520         memset(bo->base.vaddr + args->size, 0,
 521                bo->base.base.size - args->size);
 522
 523         bo->validated_shader = vc4_validate_shader(&bo->base);
 524         if (!bo->validated_shader) {
 525                 ret = -EINVAL;
 526                 goto fail;
 527         }
 528
 529         /* We have to create the handle after validation, to avoid
 530          * races for users to do doing things like mmap the shader BO.
 531          */
 532         ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
 533
 534  fail:
 535         drm_gem_object_unreference_unlocked(&bo->base.base);
 536
 537         return ret;
 538 }
 539
 540 void vc4_bo_cache_init(struct drm_device *dev)
 541 {
 542         struct vc4_dev *vc4 = to_vc4_dev(dev);
 543
 544         mutex_init(&vc4->bo_lock);
 545
 546         INIT_LIST_HEAD(&vc4->bo_cache.time_list);
 547
 548         INIT_WORK(&vc4->bo_cache.time_work, vc4_bo_cache_time_work);
 549         setup_timer(&vc4->bo_cache.time_timer,
 550                     vc4_bo_cache_time_timer,
 551                     (unsigned long)dev);
 552 }
 553
 554 void vc4_bo_cache_destroy(struct drm_device *dev)
 555 {
 556         struct vc4_dev *vc4 = to_vc4_dev(dev);
 557
 558         del_timer(&vc4->bo_cache.time_timer);
 559         cancel_work_sync(&vc4->bo_cache.time_work);
 560
 561         vc4_bo_cache_purge(dev);
 562
 563         if (vc4->bo_stats.num_allocated) {
 564                 DRM_ERROR("Destroying BO cache while BOs still allocated:\n");
 565                 vc4_bo_stats_dump(vc4);
 566         }
 567 }