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