1 /* 2 * Copyright 2009 Jerome Glisse. 3 * All Rights Reserved. 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the 7 * "Software"), to deal in the Software without restriction, including 8 * without limitation the rights to use, copy, modify, merge, publish, 9 * distribute, sub license, and/or sell copies of the Software, and to 10 * permit persons to whom the Software is furnished to do so, subject to 11 * the following conditions: 12 * 13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 19 * USE OR OTHER DEALINGS IN THE SOFTWARE. 20 * 21 * The above copyright notice and this permission notice (including the 22 * next paragraph) shall be included in all copies or substantial portions 23 * of the Software. 24 * 25 */ 26 /* 27 * Authors: 28 * Jerome Glisse <glisse@freedesktop.org> 29 * Thomas Hellstrom <thomas-at-tungstengraphics-dot-com> 30 * Dave Airlie 31 * 32 * $FreeBSD: head/sys/dev/drm2/radeon/radeon_ttm.c 254885 2013-08-25 19:37:15Z dumbbell $ 33 */ 34 #include <drm/ttm/ttm_bo_api.h> 35 #include <drm/ttm/ttm_bo_driver.h> 36 #include <drm/ttm/ttm_placement.h> 37 #include <drm/ttm/ttm_module.h> 38 #include <drm/ttm/ttm_page_alloc.h> 39 #include <drm/drmP.h> 40 #include <drm/radeon_drm.h> 41 #include <linux/seq_file.h> 42 #include <linux/slab.h> 43 #include <linux/swap.h> 44 #include <linux/pagemap.h> 45 #include "radeon_reg.h" 46 #include "radeon.h" 47 48 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT) 49 50 static int radeon_ttm_debugfs_init(struct radeon_device *rdev); 51 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev); 52 53 static struct radeon_device *radeon_get_rdev(struct ttm_bo_device *bdev) 54 { 55 struct radeon_mman *mman; 56 struct radeon_device *rdev; 57 58 mman = container_of(bdev, struct radeon_mman, bdev); 59 rdev = container_of(mman, struct radeon_device, mman); 60 return rdev; 61 } 62 63 64 /* 65 * Global memory. 66 */ 67 static int radeon_ttm_mem_global_init(struct drm_global_reference *ref) 68 { 69 return ttm_mem_global_init(ref->object); 70 } 71 72 static void radeon_ttm_mem_global_release(struct drm_global_reference *ref) 73 { 74 ttm_mem_global_release(ref->object); 75 } 76 77 static int radeon_ttm_global_init(struct radeon_device *rdev) 78 { 79 struct drm_global_reference *global_ref; 80 int r; 81 82 rdev->mman.mem_global_referenced = false; 83 global_ref = &rdev->mman.mem_global_ref; 84 global_ref->global_type = DRM_GLOBAL_TTM_MEM; 85 global_ref->size = sizeof(struct ttm_mem_global); 86 global_ref->init = &radeon_ttm_mem_global_init; 87 global_ref->release = &radeon_ttm_mem_global_release; 88 r = drm_global_item_ref(global_ref); 89 if (r != 0) { 90 DRM_ERROR("Failed setting up TTM memory accounting " 91 "subsystem.\n"); 92 return r; 93 } 94 95 rdev->mman.bo_global_ref.mem_glob = 96 rdev->mman.mem_global_ref.object; 97 global_ref = &rdev->mman.bo_global_ref.ref; 98 global_ref->global_type = DRM_GLOBAL_TTM_BO; 99 global_ref->size = sizeof(struct ttm_bo_global); 100 global_ref->init = &ttm_bo_global_init; 101 global_ref->release = &ttm_bo_global_release; 102 r = drm_global_item_ref(global_ref); 103 if (r != 0) { 104 DRM_ERROR("Failed setting up TTM BO subsystem.\n"); 105 drm_global_item_unref(&rdev->mman.mem_global_ref); 106 return r; 107 } 108 109 rdev->mman.mem_global_referenced = true; 110 return 0; 111 } 112 113 static void radeon_ttm_global_fini(struct radeon_device *rdev) 114 { 115 if (rdev->mman.mem_global_referenced) { 116 drm_global_item_unref(&rdev->mman.bo_global_ref.ref); 117 drm_global_item_unref(&rdev->mman.mem_global_ref); 118 rdev->mman.mem_global_referenced = false; 119 } 120 } 121 122 static int radeon_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags) 123 { 124 return 0; 125 } 126 127 static int radeon_init_mem_type(struct ttm_bo_device *bdev, uint32_t type, 128 struct ttm_mem_type_manager *man) 129 { 130 struct radeon_device *rdev; 131 132 rdev = radeon_get_rdev(bdev); 133 134 switch (type) { 135 case TTM_PL_SYSTEM: 136 /* System memory */ 137 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE; 138 man->available_caching = TTM_PL_MASK_CACHING; 139 man->default_caching = TTM_PL_FLAG_CACHED; 140 break; 141 case TTM_PL_TT: 142 man->func = &ttm_bo_manager_func; 143 man->gpu_offset = rdev->mc.gtt_start; 144 man->available_caching = TTM_PL_MASK_CACHING; 145 man->default_caching = TTM_PL_FLAG_CACHED; 146 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA; 147 #if IS_ENABLED(CONFIG_AGP) 148 if (rdev->flags & RADEON_IS_AGP) { 149 if (!rdev->ddev->agp) { 150 DRM_ERROR("AGP is not enabled for memory type %u\n", 151 (unsigned)type); 152 return -EINVAL; 153 } 154 if (!rdev->ddev->agp->cant_use_aperture) 155 man->flags = TTM_MEMTYPE_FLAG_MAPPABLE; 156 man->available_caching = TTM_PL_FLAG_UNCACHED | 157 TTM_PL_FLAG_WC; 158 man->default_caching = TTM_PL_FLAG_WC; 159 } 160 #endif 161 break; 162 case TTM_PL_VRAM: 163 /* "On-card" video ram */ 164 man->func = &ttm_bo_manager_func; 165 man->gpu_offset = rdev->mc.vram_start; 166 man->flags = TTM_MEMTYPE_FLAG_FIXED | 167 TTM_MEMTYPE_FLAG_MAPPABLE; 168 man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC; 169 man->default_caching = TTM_PL_FLAG_WC; 170 break; 171 default: 172 DRM_ERROR("Unsupported memory type %u\n", (unsigned)type); 173 return -EINVAL; 174 } 175 return 0; 176 } 177 178 static void radeon_evict_flags(struct ttm_buffer_object *bo, 179 struct ttm_placement *placement) 180 { 181 static const struct ttm_place placements = { 182 .fpfn = 0, 183 .lpfn = 0, 184 .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM 185 }; 186 187 struct radeon_bo *rbo; 188 189 if (!radeon_ttm_bo_is_radeon_bo(bo)) { 190 placement->placement = &placements; 191 placement->busy_placement = &placements; 192 placement->num_placement = 1; 193 placement->num_busy_placement = 1; 194 return; 195 } 196 rbo = container_of(bo, struct radeon_bo, tbo); 197 switch (bo->mem.mem_type) { 198 case TTM_PL_VRAM: 199 if (rbo->rdev->ring[radeon_copy_ring_index(rbo->rdev)].ready == false) 200 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU); 201 else if (rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size && 202 bo->mem.start < (rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT)) { 203 unsigned fpfn = rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT; 204 int i; 205 206 /* Try evicting to the CPU inaccessible part of VRAM 207 * first, but only set GTT as busy placement, so this 208 * BO will be evicted to GTT rather than causing other 209 * BOs to be evicted from VRAM 210 */ 211 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM | 212 RADEON_GEM_DOMAIN_GTT); 213 rbo->placement.num_busy_placement = 0; 214 for (i = 0; i < rbo->placement.num_placement; i++) { 215 if (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) { 216 if (rbo->placements[i].fpfn < fpfn) 217 rbo->placements[i].fpfn = fpfn; 218 } else { 219 rbo->placement.busy_placement = 220 &rbo->placements[i]; 221 rbo->placement.num_busy_placement = 1; 222 } 223 } 224 } else 225 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT); 226 break; 227 case TTM_PL_TT: 228 default: 229 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU); 230 } 231 *placement = rbo->placement; 232 } 233 234 static int radeon_verify_access(struct ttm_buffer_object *bo, struct file *filp) 235 { 236 #if 0 237 struct radeon_bo *rbo = container_of(bo, struct radeon_bo, tbo); 238 #endif 239 240 if (radeon_ttm_tt_has_userptr(bo->ttm)) 241 return -EPERM; 242 243 return 0; 244 } 245 246 static void radeon_move_null(struct ttm_buffer_object *bo, 247 struct ttm_mem_reg *new_mem) 248 { 249 struct ttm_mem_reg *old_mem = &bo->mem; 250 251 BUG_ON(old_mem->mm_node != NULL); 252 *old_mem = *new_mem; 253 new_mem->mm_node = NULL; 254 } 255 256 static int radeon_move_blit(struct ttm_buffer_object *bo, 257 bool evict, bool no_wait_gpu, 258 struct ttm_mem_reg *new_mem, 259 struct ttm_mem_reg *old_mem) 260 { 261 struct radeon_device *rdev; 262 uint64_t old_start, new_start; 263 struct radeon_fence *fence; 264 unsigned num_pages; 265 int r, ridx; 266 267 rdev = radeon_get_rdev(bo->bdev); 268 ridx = radeon_copy_ring_index(rdev); 269 old_start = (u64)old_mem->start << PAGE_SHIFT; 270 new_start = (u64)new_mem->start << PAGE_SHIFT; 271 272 switch (old_mem->mem_type) { 273 case TTM_PL_VRAM: 274 old_start += rdev->mc.vram_start; 275 break; 276 case TTM_PL_TT: 277 old_start += rdev->mc.gtt_start; 278 break; 279 default: 280 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type); 281 return -EINVAL; 282 } 283 switch (new_mem->mem_type) { 284 case TTM_PL_VRAM: 285 new_start += rdev->mc.vram_start; 286 break; 287 case TTM_PL_TT: 288 new_start += rdev->mc.gtt_start; 289 break; 290 default: 291 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type); 292 return -EINVAL; 293 } 294 if (!rdev->ring[ridx].ready) { 295 DRM_ERROR("Trying to move memory with ring turned off.\n"); 296 return -EINVAL; 297 } 298 299 BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0); 300 301 num_pages = new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE); 302 fence = radeon_copy(rdev, old_start, new_start, num_pages, bo->resv); 303 if (IS_ERR(fence)) 304 return PTR_ERR(fence); 305 306 r = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, new_mem); 307 radeon_fence_unref(&fence); 308 return r; 309 } 310 311 static int radeon_move_vram_ram(struct ttm_buffer_object *bo, 312 bool evict, bool interruptible, 313 bool no_wait_gpu, 314 struct ttm_mem_reg *new_mem) 315 { 316 struct radeon_device *rdev; 317 struct ttm_mem_reg *old_mem = &bo->mem; 318 struct ttm_mem_reg tmp_mem; 319 struct ttm_place placements; 320 struct ttm_placement placement; 321 int r; 322 323 rdev = radeon_get_rdev(bo->bdev); 324 tmp_mem = *new_mem; 325 tmp_mem.mm_node = NULL; 326 placement.num_placement = 1; 327 placement.placement = &placements; 328 placement.num_busy_placement = 1; 329 placement.busy_placement = &placements; 330 placements.fpfn = 0; 331 placements.lpfn = 0; 332 placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT; 333 r = ttm_bo_mem_space(bo, &placement, &tmp_mem, 334 interruptible, no_wait_gpu); 335 if (unlikely(r)) { 336 return r; 337 } 338 339 r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement); 340 if (unlikely(r)) { 341 goto out_cleanup; 342 } 343 344 r = ttm_tt_bind(bo->ttm, &tmp_mem); 345 if (unlikely(r)) { 346 goto out_cleanup; 347 } 348 r = radeon_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem); 349 if (unlikely(r)) { 350 goto out_cleanup; 351 } 352 r = ttm_bo_move_ttm(bo, interruptible, no_wait_gpu, new_mem); 353 out_cleanup: 354 ttm_bo_mem_put(bo, &tmp_mem); 355 return r; 356 } 357 358 static int radeon_move_ram_vram(struct ttm_buffer_object *bo, 359 bool evict, bool interruptible, 360 bool no_wait_gpu, 361 struct ttm_mem_reg *new_mem) 362 { 363 struct radeon_device *rdev; 364 struct ttm_mem_reg *old_mem = &bo->mem; 365 struct ttm_mem_reg tmp_mem; 366 struct ttm_placement placement; 367 struct ttm_place placements; 368 int r; 369 370 rdev = radeon_get_rdev(bo->bdev); 371 tmp_mem = *new_mem; 372 tmp_mem.mm_node = NULL; 373 placement.num_placement = 1; 374 placement.placement = &placements; 375 placement.num_busy_placement = 1; 376 placement.busy_placement = &placements; 377 placements.fpfn = 0; 378 placements.lpfn = 0; 379 placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT; 380 r = ttm_bo_mem_space(bo, &placement, &tmp_mem, 381 interruptible, no_wait_gpu); 382 if (unlikely(r)) { 383 return r; 384 } 385 r = ttm_bo_move_ttm(bo, interruptible, no_wait_gpu, &tmp_mem); 386 if (unlikely(r)) { 387 goto out_cleanup; 388 } 389 r = radeon_move_blit(bo, true, no_wait_gpu, new_mem, old_mem); 390 if (unlikely(r)) { 391 goto out_cleanup; 392 } 393 out_cleanup: 394 ttm_bo_mem_put(bo, &tmp_mem); 395 return r; 396 } 397 398 static int radeon_bo_move(struct ttm_buffer_object *bo, 399 bool evict, bool interruptible, 400 bool no_wait_gpu, 401 struct ttm_mem_reg *new_mem) 402 { 403 struct radeon_device *rdev; 404 struct radeon_bo *rbo; 405 struct ttm_mem_reg *old_mem = &bo->mem; 406 int r; 407 408 r = ttm_bo_wait(bo, interruptible, no_wait_gpu); 409 if (r) 410 return r; 411 412 /* Can't move a pinned BO */ 413 rbo = container_of(bo, struct radeon_bo, tbo); 414 if (WARN_ON_ONCE(rbo->pin_count > 0)) 415 return -EINVAL; 416 417 rdev = radeon_get_rdev(bo->bdev); 418 if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 419 radeon_move_null(bo, new_mem); 420 return 0; 421 } 422 if ((old_mem->mem_type == TTM_PL_TT && 423 new_mem->mem_type == TTM_PL_SYSTEM) || 424 (old_mem->mem_type == TTM_PL_SYSTEM && 425 new_mem->mem_type == TTM_PL_TT)) { 426 /* bind is enough */ 427 radeon_move_null(bo, new_mem); 428 return 0; 429 } 430 if (!rdev->ring[radeon_copy_ring_index(rdev)].ready || 431 rdev->asic->copy.copy == NULL) { 432 /* use memcpy */ 433 goto memcpy; 434 } 435 436 if (old_mem->mem_type == TTM_PL_VRAM && 437 new_mem->mem_type == TTM_PL_SYSTEM) { 438 r = radeon_move_vram_ram(bo, evict, interruptible, 439 no_wait_gpu, new_mem); 440 } else if (old_mem->mem_type == TTM_PL_SYSTEM && 441 new_mem->mem_type == TTM_PL_VRAM) { 442 r = radeon_move_ram_vram(bo, evict, interruptible, 443 no_wait_gpu, new_mem); 444 } else { 445 r = radeon_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem); 446 } 447 448 if (r) { 449 memcpy: 450 r = ttm_bo_move_memcpy(bo, interruptible, no_wait_gpu, new_mem); 451 if (r) { 452 return r; 453 } 454 } 455 456 /* update statistics */ 457 atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &rdev->num_bytes_moved); 458 return 0; 459 } 460 461 static int radeon_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 462 { 463 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 464 struct radeon_device *rdev = radeon_get_rdev(bdev); 465 466 mem->bus.addr = NULL; 467 mem->bus.offset = 0; 468 mem->bus.size = mem->num_pages << PAGE_SHIFT; 469 mem->bus.base = 0; 470 mem->bus.is_iomem = false; 471 if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE)) 472 return -EINVAL; 473 switch (mem->mem_type) { 474 case TTM_PL_SYSTEM: 475 /* system memory */ 476 return 0; 477 case TTM_PL_TT: 478 #if IS_ENABLED(CONFIG_AGP) 479 if (rdev->flags & RADEON_IS_AGP) { 480 /* RADEON_IS_AGP is set only if AGP is active */ 481 mem->bus.offset = mem->start << PAGE_SHIFT; 482 mem->bus.base = rdev->mc.agp_base; 483 mem->bus.is_iomem = !rdev->ddev->agp->cant_use_aperture; 484 } 485 #endif 486 break; 487 case TTM_PL_VRAM: 488 mem->bus.offset = mem->start << PAGE_SHIFT; 489 /* check if it's visible */ 490 if ((mem->bus.offset + mem->bus.size) > rdev->mc.visible_vram_size) 491 return -EINVAL; 492 mem->bus.base = rdev->mc.aper_base; 493 mem->bus.is_iomem = true; 494 #ifdef __alpha__ 495 /* 496 * Alpha: use bus.addr to hold the ioremap() return, 497 * so we can modify bus.base below. 498 */ 499 if (mem->placement & TTM_PL_FLAG_WC) 500 mem->bus.addr = 501 ioremap_wc(mem->bus.base + mem->bus.offset, 502 mem->bus.size); 503 else 504 mem->bus.addr = 505 ioremap_nocache(mem->bus.base + mem->bus.offset, 506 mem->bus.size); 507 if (!mem->bus.addr) 508 return -ENOMEM; 509 510 /* 511 * Alpha: Use just the bus offset plus 512 * the hose/domain memory base for bus.base. 513 * It then can be used to build PTEs for VRAM 514 * access, as done in ttm_bo_vm_fault(). 515 */ 516 mem->bus.base = (mem->bus.base & 0x0ffffffffUL) + 517 rdev->ddev->hose->dense_mem_base; 518 #endif 519 break; 520 default: 521 return -EINVAL; 522 } 523 return 0; 524 } 525 526 static void radeon_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 527 { 528 } 529 530 /* 531 * TTM backend functions. 532 */ 533 struct radeon_ttm_tt { 534 struct ttm_dma_tt ttm; 535 struct radeon_device *rdev; 536 u64 offset; 537 538 uint64_t userptr; 539 struct mm_struct *usermm; 540 uint32_t userflags; 541 }; 542 543 #if 0 544 /* prepare the sg table with the user pages */ 545 static int radeon_ttm_tt_pin_userptr(struct ttm_tt *ttm) 546 { 547 struct radeon_device *rdev = radeon_get_rdev(ttm->bdev); 548 struct radeon_ttm_tt *gtt = (void *)ttm; 549 unsigned pinned = 0, nents; 550 int r; 551 552 int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY); 553 enum dma_data_direction direction = write ? 554 DMA_BIDIRECTIONAL : DMA_TO_DEVICE; 555 556 if (current->mm != gtt->usermm) 557 return -EPERM; 558 559 if (gtt->userflags & RADEON_GEM_USERPTR_ANONONLY) { 560 /* check that we only pin down anonymous memory 561 to prevent problems with writeback */ 562 unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE; 563 struct vm_area_struct *vma; 564 vma = find_vma(gtt->usermm, gtt->userptr); 565 if (!vma || vma->vm_file || vma->vm_end < end) 566 return -EPERM; 567 } 568 569 do { 570 unsigned num_pages = ttm->num_pages - pinned; 571 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE; 572 struct page **pages = ttm->pages + pinned; 573 574 r = get_user_pages(userptr, num_pages, write ? FOLL_WRITE : 0, 575 pages, NULL); 576 if (r < 0) 577 goto release_pages; 578 579 pinned += r; 580 581 } while (pinned < ttm->num_pages); 582 583 r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0, 584 ttm->num_pages << PAGE_SHIFT, 585 GFP_KERNEL); 586 if (r) 587 goto release_sg; 588 589 r = -ENOMEM; 590 nents = dma_map_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction); 591 if (nents != ttm->sg->nents) 592 goto release_sg; 593 594 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages, 595 gtt->ttm.dma_address, ttm->num_pages); 596 597 return 0; 598 599 release_sg: 600 kfree(ttm->sg); 601 602 release_pages: 603 release_pages(ttm->pages, pinned); 604 return r; 605 } 606 607 static void radeon_ttm_tt_unpin_userptr(struct ttm_tt *ttm) 608 { 609 struct radeon_device *rdev = radeon_get_rdev(ttm->bdev); 610 struct radeon_ttm_tt *gtt = (void *)ttm; 611 struct sg_page_iter sg_iter; 612 613 int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY); 614 enum dma_data_direction direction = write ? 615 DMA_BIDIRECTIONAL : DMA_TO_DEVICE; 616 617 /* double check that we don't free the table twice */ 618 if (!ttm->sg->sgl) 619 return; 620 621 /* free the sg table and pages again */ 622 dma_unmap_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction); 623 624 for_each_sg_page(ttm->sg->sgl, &sg_iter, ttm->sg->nents, 0) { 625 struct page *page = sg_page_iter_page(&sg_iter); 626 if (!(gtt->userflags & RADEON_GEM_USERPTR_READONLY)) 627 set_page_dirty(page); 628 629 mark_page_accessed(page); 630 put_page(page); 631 } 632 633 sg_free_table(ttm->sg); 634 } 635 #endif 636 637 static int radeon_ttm_backend_bind(struct ttm_tt *ttm, 638 struct ttm_mem_reg *bo_mem) 639 { 640 struct radeon_ttm_tt *gtt = (void*)ttm; 641 uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ | 642 RADEON_GART_PAGE_WRITE; 643 int r; 644 645 #if 0 646 if (gtt->userptr) { 647 radeon_ttm_tt_pin_userptr(ttm); 648 flags &= ~RADEON_GART_PAGE_WRITE; 649 } 650 #endif 651 652 gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT); 653 if (!ttm->num_pages) { 654 WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n", 655 ttm->num_pages, bo_mem, ttm); 656 } 657 if (ttm->caching_state == tt_cached) 658 flags |= RADEON_GART_PAGE_SNOOP; 659 r = radeon_gart_bind(gtt->rdev, gtt->offset, ttm->num_pages, 660 ttm->pages, gtt->ttm.dma_address, flags); 661 if (r) { 662 DRM_ERROR("failed to bind %lu pages at 0x%08X\n", 663 ttm->num_pages, (unsigned)gtt->offset); 664 return r; 665 } 666 return 0; 667 } 668 669 static int radeon_ttm_backend_unbind(struct ttm_tt *ttm) 670 { 671 struct radeon_ttm_tt *gtt = (void *)ttm; 672 673 radeon_gart_unbind(gtt->rdev, gtt->offset, ttm->num_pages); 674 #if 0 675 if (gtt->userptr) 676 radeon_ttm_tt_unpin_userptr(ttm); 677 #endif 678 679 return 0; 680 } 681 682 static void radeon_ttm_backend_destroy(struct ttm_tt *ttm) 683 { 684 struct radeon_ttm_tt *gtt = (void *)ttm; 685 686 ttm_dma_tt_fini(>t->ttm); 687 kfree(gtt); 688 } 689 690 static struct ttm_backend_func radeon_backend_func = { 691 .bind = &radeon_ttm_backend_bind, 692 .unbind = &radeon_ttm_backend_unbind, 693 .destroy = &radeon_ttm_backend_destroy, 694 }; 695 696 static struct ttm_tt *radeon_ttm_tt_create(struct ttm_bo_device *bdev, 697 unsigned long size, uint32_t page_flags, 698 struct page *dummy_read_page) 699 { 700 struct radeon_device *rdev; 701 struct radeon_ttm_tt *gtt; 702 703 rdev = radeon_get_rdev(bdev); 704 #if IS_ENABLED(CONFIG_AGP) 705 if (rdev->flags & RADEON_IS_AGP) { 706 return ttm_agp_tt_create(bdev, rdev->ddev->agp->agpdev, 707 size, page_flags, dummy_read_page); 708 } 709 #endif 710 711 gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL); 712 if (gtt == NULL) { 713 return NULL; 714 } 715 gtt->ttm.ttm.func = &radeon_backend_func; 716 gtt->rdev = rdev; 717 if (ttm_dma_tt_init(>t->ttm, bdev, size, page_flags, dummy_read_page)) { 718 kfree(gtt); 719 return NULL; 720 } 721 return >t->ttm.ttm; 722 } 723 724 static struct radeon_ttm_tt *radeon_ttm_tt_to_gtt(struct ttm_tt *ttm) 725 { 726 if (!ttm || ttm->func != &radeon_backend_func) 727 return NULL; 728 return (struct radeon_ttm_tt *)ttm; 729 } 730 731 static int radeon_ttm_tt_populate(struct ttm_tt *ttm) 732 { 733 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm); 734 struct radeon_device *rdev; 735 unsigned i; 736 int r; 737 #ifdef DUMBBELL_WIP 738 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG); 739 #endif /* DUMBBELL_WIP */ 740 741 if (ttm->state != tt_unpopulated) 742 return 0; 743 744 #if 0 745 if (gtt && gtt->userptr) { 746 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL); 747 if (!ttm->sg) 748 return -ENOMEM; 749 750 ttm->page_flags |= TTM_PAGE_FLAG_SG; 751 ttm->state = tt_unbound; 752 return 0; 753 } 754 #endif 755 756 #ifdef DUMBBELL_WIP 757 /* 758 * Maybe unneeded on FreeBSD. 759 * -- dumbbell@ 760 */ 761 if (slave && ttm->sg) { 762 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages, 763 gtt->ttm.dma_address, ttm->num_pages); 764 ttm->state = tt_unbound; 765 return 0; 766 } 767 #endif /* DUMBBELL_WIP */ 768 769 rdev = radeon_get_rdev(ttm->bdev); 770 #if IS_ENABLED(CONFIG_AGP) 771 if (rdev->flags & RADEON_IS_AGP) { 772 return ttm_agp_tt_populate(ttm); 773 } 774 #endif 775 776 #ifdef CONFIG_SWIOTLB 777 if (swiotlb_nr_tbl()) { 778 return ttm_dma_populate(>t->ttm, rdev->dev); 779 } 780 #endif 781 782 r = ttm_pool_populate(ttm); 783 if (r) { 784 return r; 785 } 786 787 for (i = 0; i < ttm->num_pages; i++) { 788 gtt->ttm.dma_address[i] = pci_map_page(rdev->pdev, ttm->pages[i], 789 0, PAGE_SIZE, 790 PCI_DMA_BIDIRECTIONAL); 791 #ifdef DUMBBELL_WIP 792 if (pci_dma_mapping_error(rdev->pdev, gtt->ttm.dma_address[i])) { 793 while (i--) { 794 pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i], 795 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL); 796 gtt->ttm.dma_address[i] = 0; 797 } 798 ttm_pool_unpopulate(ttm); 799 return -EFAULT; 800 } 801 #endif /* DUMBBELL_WIP */ 802 } 803 return 0; 804 } 805 806 static void radeon_ttm_tt_unpopulate(struct ttm_tt *ttm) 807 { 808 struct radeon_device *rdev; 809 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm); 810 unsigned i; 811 bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG); 812 813 #if 0 814 if (gtt && gtt->userptr) { 815 kfree(ttm->sg); 816 ttm->page_flags &= ~TTM_PAGE_FLAG_SG; 817 return; 818 } 819 #endif 820 821 if (slave) 822 return; 823 824 rdev = radeon_get_rdev(ttm->bdev); 825 #if IS_ENABLED(CONFIG_AGP) 826 if (rdev->flags & RADEON_IS_AGP) { 827 ttm_agp_tt_unpopulate(ttm); 828 return; 829 } 830 #endif 831 832 #ifdef CONFIG_SWIOTLB 833 if (swiotlb_nr_tbl()) { 834 ttm_dma_unpopulate(>t->ttm, rdev->dev); 835 return; 836 } 837 #endif 838 839 for (i = 0; i < ttm->num_pages; i++) { 840 if (gtt->ttm.dma_address[i]) { 841 gtt->ttm.dma_address[i] = 0; 842 #ifdef DUMBBELL_WIP 843 pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i], 844 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL); 845 #endif /* DUMBBELL_WIP */ 846 } 847 } 848 849 ttm_pool_unpopulate(ttm); 850 } 851 852 bool radeon_ttm_tt_has_userptr(struct ttm_tt *ttm) 853 { 854 #if 0 855 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm); 856 857 if (gtt == NULL) 858 #endif 859 return false; 860 861 #if 0 862 return !!gtt->userptr; 863 #endif 864 } 865 866 bool radeon_ttm_tt_is_readonly(struct ttm_tt *ttm) 867 { 868 struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm); 869 870 if (gtt == NULL) 871 return false; 872 873 return !!(gtt->userflags & RADEON_GEM_USERPTR_READONLY); 874 } 875 876 static struct ttm_bo_driver radeon_bo_driver = { 877 .ttm_tt_create = &radeon_ttm_tt_create, 878 .ttm_tt_populate = &radeon_ttm_tt_populate, 879 .ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate, 880 .invalidate_caches = &radeon_invalidate_caches, 881 .init_mem_type = &radeon_init_mem_type, 882 .eviction_valuable = ttm_bo_eviction_valuable, 883 .evict_flags = &radeon_evict_flags, 884 .move = &radeon_bo_move, 885 .verify_access = &radeon_verify_access, 886 .move_notify = &radeon_bo_move_notify, 887 .fault_reserve_notify = &radeon_bo_fault_reserve_notify, 888 .io_mem_reserve = &radeon_ttm_io_mem_reserve, 889 .io_mem_free = &radeon_ttm_io_mem_free, 890 .io_mem_pfn = ttm_bo_default_io_mem_pfn, 891 }; 892 893 int radeon_ttm_init(struct radeon_device *rdev) 894 { 895 int r; 896 897 r = radeon_ttm_global_init(rdev); 898 if (r) { 899 return r; 900 } 901 /* No others user of address space so set it to 0 */ 902 r = ttm_bo_device_init(&rdev->mman.bdev, 903 rdev->mman.bo_global_ref.ref.object, 904 &radeon_bo_driver, 905 #ifdef __DragonFly__ 906 NULL, 907 #else 908 rdev->ddev->anon_inode->i_mapping, 909 #endif 910 DRM_FILE_PAGE_OFFSET, 911 rdev->need_dma32); 912 if (r) { 913 DRM_ERROR("failed initializing buffer object driver(%d).\n", r); 914 return r; 915 } 916 rdev->mman.initialized = true; 917 rdev->ddev->drm_ttm_bdev = &rdev->mman.bdev; 918 r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_VRAM, 919 rdev->mc.real_vram_size >> PAGE_SHIFT); 920 if (r) { 921 DRM_ERROR("Failed initializing VRAM heap.\n"); 922 return r; 923 } 924 /* Change the size here instead of the init above so only lpfn is affected */ 925 radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size); 926 927 r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true, 928 RADEON_GEM_DOMAIN_VRAM, 0, NULL, 929 NULL, &rdev->stolen_vga_memory); 930 if (r) { 931 return r; 932 } 933 r = radeon_bo_reserve(rdev->stolen_vga_memory, false); 934 if (r) 935 return r; 936 r = radeon_bo_pin(rdev->stolen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL); 937 radeon_bo_unreserve(rdev->stolen_vga_memory); 938 if (r) { 939 radeon_bo_unref(&rdev->stolen_vga_memory); 940 return r; 941 } 942 DRM_INFO("radeon: %uM of VRAM memory ready\n", 943 (unsigned) (rdev->mc.real_vram_size / (1024 * 1024))); 944 r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_TT, 945 rdev->mc.gtt_size >> PAGE_SHIFT); 946 if (r) { 947 DRM_ERROR("Failed initializing GTT heap.\n"); 948 return r; 949 } 950 DRM_INFO("radeon: %uM of GTT memory ready.\n", 951 (unsigned)(rdev->mc.gtt_size / (1024 * 1024))); 952 953 r = radeon_ttm_debugfs_init(rdev); 954 if (r) { 955 DRM_ERROR("Failed to init debugfs\n"); 956 return r; 957 } 958 return 0; 959 } 960 961 void radeon_ttm_fini(struct radeon_device *rdev) 962 { 963 int r; 964 965 if (!rdev->mman.initialized) 966 return; 967 radeon_ttm_debugfs_fini(rdev); 968 if (rdev->stolen_vga_memory) { 969 r = radeon_bo_reserve(rdev->stolen_vga_memory, false); 970 if (r == 0) { 971 radeon_bo_unpin(rdev->stolen_vga_memory); 972 radeon_bo_unreserve(rdev->stolen_vga_memory); 973 } 974 radeon_bo_unref(&rdev->stolen_vga_memory); 975 } 976 ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_VRAM); 977 ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_TT); 978 ttm_bo_device_release(&rdev->mman.bdev); 979 radeon_gart_fini(rdev); 980 radeon_ttm_global_fini(rdev); 981 rdev->mman.initialized = false; 982 DRM_INFO("radeon: ttm finalized\n"); 983 } 984 985 /* this should only be called at bootup or when userspace 986 * isn't running */ 987 void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size) 988 { 989 struct ttm_mem_type_manager *man; 990 991 if (!rdev->mman.initialized) 992 return; 993 994 man = &rdev->mman.bdev.man[TTM_PL_VRAM]; 995 /* this just adjusts TTM size idea, which sets lpfn to the correct value */ 996 man->size = size >> PAGE_SHIFT; 997 } 998 999 #ifdef DUMBBELL_WIP 1000 static struct vm_operations_struct radeon_ttm_vm_ops; 1001 static const struct vm_operations_struct *ttm_vm_ops = NULL; 1002 1003 static int radeon_ttm_fault(struct vm_fault *vmf) 1004 { 1005 struct ttm_buffer_object *bo; 1006 struct radeon_device *rdev; 1007 int r; 1008 1009 bo = (struct ttm_buffer_object *)vmf->vma->vm_private_data; 1010 if (bo == NULL) { 1011 return VM_FAULT_NOPAGE; 1012 } 1013 rdev = radeon_get_rdev(bo->bdev); 1014 down_read(&rdev->pm.mclk_lock); 1015 r = ttm_vm_ops->fault(vmf); 1016 up_read(&rdev->pm.mclk_lock); 1017 return r; 1018 } 1019 1020 int radeon_mmap(struct file *filp, struct vm_area_struct *vma) 1021 { 1022 struct drm_file *file_priv; 1023 struct radeon_device *rdev; 1024 int r; 1025 1026 if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET)) { 1027 return -EINVAL; 1028 } 1029 1030 file_priv = filp->private_data; 1031 rdev = file_priv->minor->dev->dev_private; 1032 if (rdev == NULL) { 1033 return -EINVAL; 1034 } 1035 r = ttm_bo_mmap(filp, vma, &rdev->mman.bdev); 1036 if (unlikely(r != 0)) { 1037 return r; 1038 } 1039 if (unlikely(ttm_vm_ops == NULL)) { 1040 ttm_vm_ops = vma->vm_ops; 1041 radeon_ttm_vm_ops = *ttm_vm_ops; 1042 radeon_ttm_vm_ops.fault = &radeon_ttm_fault; 1043 } 1044 vma->vm_ops = &radeon_ttm_vm_ops; 1045 return 0; 1046 } 1047 #endif /* DUMBBELL_WIP */ 1048 1049 #if defined(CONFIG_DEBUG_FS) 1050 1051 static int radeon_mm_dump_table(struct seq_file *m, void *data) 1052 { 1053 struct drm_info_node *node = (struct drm_info_node *)m->private; 1054 unsigned ttm_pl = *(int*)node->info_ent->data; 1055 struct drm_device *dev = node->minor->dev; 1056 struct radeon_device *rdev = dev->dev_private; 1057 struct ttm_mem_type_manager *man = &rdev->mman.bdev.man[ttm_pl]; 1058 struct drm_printer p = drm_seq_file_printer(m); 1059 1060 man->func->debug(man, &p); 1061 return 0; 1062 } 1063 1064 static int ttm_pl_vram = TTM_PL_VRAM; 1065 static int ttm_pl_tt = TTM_PL_TT; 1066 1067 static struct drm_info_list radeon_ttm_debugfs_list[] = { 1068 {"radeon_vram_mm", radeon_mm_dump_table, 0, &ttm_pl_vram}, 1069 {"radeon_gtt_mm", radeon_mm_dump_table, 0, &ttm_pl_tt}, 1070 {"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL}, 1071 #ifdef CONFIG_SWIOTLB 1072 {"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL} 1073 #endif 1074 }; 1075 1076 static int radeon_ttm_vram_open(struct inode *inode, struct file *filep) 1077 { 1078 struct radeon_device *rdev = inode->i_private; 1079 i_size_write(inode, rdev->mc.mc_vram_size); 1080 filep->private_data = inode->i_private; 1081 return 0; 1082 } 1083 1084 static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf, 1085 size_t size, loff_t *pos) 1086 { 1087 struct radeon_device *rdev = f->private_data; 1088 ssize_t result = 0; 1089 int r; 1090 1091 if (size & 0x3 || *pos & 0x3) 1092 return -EINVAL; 1093 1094 while (size) { 1095 unsigned long flags; 1096 uint32_t value; 1097 1098 if (*pos >= rdev->mc.mc_vram_size) 1099 return result; 1100 1101 spin_lock_irqsave(&rdev->mmio_idx_lock, flags); 1102 WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000); 1103 if (rdev->family >= CHIP_CEDAR) 1104 WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31); 1105 value = RREG32(RADEON_MM_DATA); 1106 spin_unlock_irqrestore(&rdev->mmio_idx_lock, flags); 1107 1108 r = put_user(value, (uint32_t *)buf); 1109 if (r) 1110 return r; 1111 1112 result += 4; 1113 buf += 4; 1114 *pos += 4; 1115 size -= 4; 1116 } 1117 1118 return result; 1119 } 1120 1121 static const struct file_operations radeon_ttm_vram_fops = { 1122 .owner = THIS_MODULE, 1123 .open = radeon_ttm_vram_open, 1124 .read = radeon_ttm_vram_read, 1125 .llseek = default_llseek 1126 }; 1127 1128 static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep) 1129 { 1130 struct radeon_device *rdev = inode->i_private; 1131 i_size_write(inode, rdev->mc.gtt_size); 1132 filep->private_data = inode->i_private; 1133 return 0; 1134 } 1135 1136 static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf, 1137 size_t size, loff_t *pos) 1138 { 1139 struct radeon_device *rdev = f->private_data; 1140 ssize_t result = 0; 1141 int r; 1142 1143 while (size) { 1144 loff_t p = *pos / PAGE_SIZE; 1145 unsigned off = *pos & ~LINUX_PAGE_MASK; 1146 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off); 1147 struct page *page; 1148 void *ptr; 1149 1150 if (p >= rdev->gart.num_cpu_pages) 1151 return result; 1152 1153 page = rdev->gart.pages[p]; 1154 if (page) { 1155 ptr = kmap(page); 1156 ptr += off; 1157 1158 r = copy_to_user(buf, ptr, cur_size); 1159 kunmap(rdev->gart.pages[p]); 1160 } else 1161 r = clear_user(buf, cur_size); 1162 1163 if (r) 1164 return -EFAULT; 1165 1166 result += cur_size; 1167 buf += cur_size; 1168 *pos += cur_size; 1169 size -= cur_size; 1170 } 1171 1172 return result; 1173 } 1174 1175 static const struct file_operations radeon_ttm_gtt_fops = { 1176 .owner = THIS_MODULE, 1177 .open = radeon_ttm_gtt_open, 1178 .read = radeon_ttm_gtt_read, 1179 .llseek = default_llseek 1180 }; 1181 1182 #endif 1183 1184 static int radeon_ttm_debugfs_init(struct radeon_device *rdev) 1185 { 1186 #if defined(CONFIG_DEBUG_FS) 1187 unsigned count; 1188 1189 struct drm_minor *minor = rdev->ddev->primary; 1190 struct dentry *ent, *root = minor->debugfs_root; 1191 1192 ent = debugfs_create_file("radeon_vram", S_IFREG | S_IRUGO, root, 1193 rdev, &radeon_ttm_vram_fops); 1194 if (IS_ERR(ent)) 1195 return PTR_ERR(ent); 1196 rdev->mman.vram = ent; 1197 1198 ent = debugfs_create_file("radeon_gtt", S_IFREG | S_IRUGO, root, 1199 rdev, &radeon_ttm_gtt_fops); 1200 if (IS_ERR(ent)) 1201 return PTR_ERR(ent); 1202 rdev->mman.gtt = ent; 1203 1204 count = ARRAY_SIZE(radeon_ttm_debugfs_list); 1205 1206 #ifdef CONFIG_SWIOTLB 1207 if (!swiotlb_nr_tbl()) 1208 --count; 1209 #endif 1210 1211 return radeon_debugfs_add_files(rdev, radeon_ttm_debugfs_list, count); 1212 #else 1213 1214 return 0; 1215 #endif 1216 } 1217 1218 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev) 1219 { 1220 #if defined(CONFIG_DEBUG_FS) 1221 1222 debugfs_remove(rdev->mman.vram); 1223 rdev->mman.vram = NULL; 1224 1225 debugfs_remove(rdev->mman.gtt); 1226 rdev->mman.gtt = NULL; 1227 #endif 1228 } 1229