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 #include <linux/list.h> 33 #include <linux/slab.h> 34 #include <drm/drmP.h> 35 #include <drm/radeon_drm.h> 36 #include "radeon.h" 37 #include "radeon_trace.h" 38 39 40 int radeon_ttm_init(struct radeon_device *rdev); 41 void radeon_ttm_fini(struct radeon_device *rdev); 42 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo); 43 44 /* 45 * To exclude mutual BO access we rely on bo_reserve exclusion, as all 46 * function are calling it. 47 */ 48 49 void radeon_bo_clear_va(struct radeon_bo *bo) 50 { 51 struct radeon_bo_va *bo_va, *tmp; 52 53 list_for_each_entry_safe(bo_va, tmp, &bo->va, bo_list) { 54 /* remove from all vm address space */ 55 radeon_vm_bo_rmv(bo->rdev, bo_va); 56 } 57 } 58 59 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo) 60 { 61 struct radeon_bo *bo; 62 63 bo = container_of(tbo, struct radeon_bo, tbo); 64 mutex_lock(&bo->rdev->gem.mutex); 65 list_del_init(&bo->list); 66 mutex_unlock(&bo->rdev->gem.mutex); 67 radeon_bo_clear_surface_reg(bo); 68 radeon_bo_clear_va(bo); 69 drm_gem_object_release(&bo->gem_base); 70 kfree(bo); 71 } 72 73 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo) 74 { 75 if (bo->destroy == &radeon_ttm_bo_destroy) 76 return true; 77 return false; 78 } 79 80 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain) 81 { 82 u32 c = 0; 83 84 rbo->placement.fpfn = 0; 85 rbo->placement.lpfn = 0; 86 rbo->placement.placement = rbo->placements; 87 rbo->placement.busy_placement = rbo->placements; 88 if (domain & RADEON_GEM_DOMAIN_VRAM) 89 rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED | 90 TTM_PL_FLAG_VRAM; 91 if (domain & RADEON_GEM_DOMAIN_GTT) { 92 if (rbo->rdev->flags & RADEON_IS_AGP) { 93 rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_TT; 94 } else { 95 rbo->placements[c++] = TTM_PL_FLAG_CACHED | TTM_PL_FLAG_TT; 96 } 97 } 98 if (domain & RADEON_GEM_DOMAIN_CPU) { 99 if (rbo->rdev->flags & RADEON_IS_AGP) { 100 rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_SYSTEM; 101 } else { 102 rbo->placements[c++] = TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM; 103 } 104 } 105 if (!c) 106 rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM; 107 rbo->placement.num_placement = c; 108 rbo->placement.num_busy_placement = c; 109 } 110 111 int radeon_bo_create(struct radeon_device *rdev, 112 unsigned long size, int byte_align, bool kernel, u32 domain, 113 struct sg_table *sg, struct radeon_bo **bo_ptr) 114 { 115 struct radeon_bo *bo; 116 enum ttm_bo_type type; 117 unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT; 118 size_t acc_size; 119 int r; 120 121 size = ALIGN(size, PAGE_SIZE); 122 123 rdev->mman.bdev.dev_mapping = rdev->ddev->dev_mapping; 124 if (kernel) { 125 type = ttm_bo_type_kernel; 126 } else if (sg) { 127 type = ttm_bo_type_sg; 128 } else { 129 type = ttm_bo_type_device; 130 } 131 *bo_ptr = NULL; 132 133 acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size, 134 sizeof(struct radeon_bo)); 135 136 bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL); 137 if (bo == NULL) 138 return -ENOMEM; 139 r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size); 140 if (unlikely(r)) { 141 kfree(bo); 142 return r; 143 } 144 bo->rdev = rdev; 145 bo->surface_reg = -1; 146 INIT_LIST_HEAD(&bo->list); 147 INIT_LIST_HEAD(&bo->va); 148 radeon_ttm_placement_from_domain(bo, domain); 149 /* Kernel allocation are uninterruptible */ 150 down_read(&rdev->pm.mclk_lock); 151 r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type, 152 &bo->placement, page_align, !kernel, NULL, 153 acc_size, sg, &radeon_ttm_bo_destroy); 154 up_read(&rdev->pm.mclk_lock); 155 if (unlikely(r != 0)) { 156 return r; 157 } 158 *bo_ptr = bo; 159 160 trace_radeon_bo_create(bo); 161 162 return 0; 163 } 164 165 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr) 166 { 167 bool is_iomem; 168 int r; 169 170 if (bo->kptr) { 171 if (ptr) { 172 *ptr = bo->kptr; 173 } 174 return 0; 175 } 176 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap); 177 if (r) { 178 return r; 179 } 180 bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem); 181 if (ptr) { 182 *ptr = bo->kptr; 183 } 184 radeon_bo_check_tiling(bo, 0, 0); 185 return 0; 186 } 187 188 void radeon_bo_kunmap(struct radeon_bo *bo) 189 { 190 if (bo->kptr == NULL) 191 return; 192 bo->kptr = NULL; 193 radeon_bo_check_tiling(bo, 0, 0); 194 ttm_bo_kunmap(&bo->kmap); 195 } 196 197 void radeon_bo_unref(struct radeon_bo **bo) 198 { 199 struct ttm_buffer_object *tbo; 200 struct radeon_device *rdev; 201 202 if ((*bo) == NULL) 203 return; 204 rdev = (*bo)->rdev; 205 tbo = &((*bo)->tbo); 206 down_read(&rdev->pm.mclk_lock); 207 ttm_bo_unref(&tbo); 208 up_read(&rdev->pm.mclk_lock); 209 if (tbo == NULL) 210 *bo = NULL; 211 } 212 213 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset, 214 u64 *gpu_addr) 215 { 216 int r, i; 217 218 if (bo->pin_count) { 219 bo->pin_count++; 220 if (gpu_addr) 221 *gpu_addr = radeon_bo_gpu_offset(bo); 222 223 if (max_offset != 0) { 224 u64 domain_start; 225 226 if (domain == RADEON_GEM_DOMAIN_VRAM) 227 domain_start = bo->rdev->mc.vram_start; 228 else 229 domain_start = bo->rdev->mc.gtt_start; 230 WARN_ON_ONCE(max_offset < 231 (radeon_bo_gpu_offset(bo) - domain_start)); 232 } 233 234 return 0; 235 } 236 radeon_ttm_placement_from_domain(bo, domain); 237 if (domain == RADEON_GEM_DOMAIN_VRAM) { 238 /* force to pin into visible video ram */ 239 bo->placement.lpfn = bo->rdev->mc.visible_vram_size >> PAGE_SHIFT; 240 } 241 if (max_offset) { 242 u64 lpfn = max_offset >> PAGE_SHIFT; 243 244 if (!bo->placement.lpfn) 245 bo->placement.lpfn = bo->rdev->mc.gtt_size >> PAGE_SHIFT; 246 247 if (lpfn < bo->placement.lpfn) 248 bo->placement.lpfn = lpfn; 249 } 250 for (i = 0; i < bo->placement.num_placement; i++) 251 bo->placements[i] |= TTM_PL_FLAG_NO_EVICT; 252 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false); 253 if (likely(r == 0)) { 254 bo->pin_count = 1; 255 if (gpu_addr != NULL) 256 *gpu_addr = radeon_bo_gpu_offset(bo); 257 } 258 if (unlikely(r != 0)) 259 dev_err(bo->rdev->dev, "%p pin failed\n", bo); 260 return r; 261 } 262 263 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr) 264 { 265 return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr); 266 } 267 268 int radeon_bo_unpin(struct radeon_bo *bo) 269 { 270 int r, i; 271 272 if (!bo->pin_count) { 273 dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo); 274 return 0; 275 } 276 bo->pin_count--; 277 if (bo->pin_count) 278 return 0; 279 for (i = 0; i < bo->placement.num_placement; i++) 280 bo->placements[i] &= ~TTM_PL_FLAG_NO_EVICT; 281 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false); 282 if (unlikely(r != 0)) 283 dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo); 284 return r; 285 } 286 287 int radeon_bo_evict_vram(struct radeon_device *rdev) 288 { 289 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */ 290 if (0 && (rdev->flags & RADEON_IS_IGP)) { 291 if (rdev->mc.igp_sideport_enabled == false) 292 /* Useless to evict on IGP chips */ 293 return 0; 294 } 295 return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM); 296 } 297 298 void radeon_bo_force_delete(struct radeon_device *rdev) 299 { 300 struct radeon_bo *bo, *n; 301 302 if (list_empty(&rdev->gem.objects)) { 303 return; 304 } 305 dev_err(rdev->dev, "Userspace still has active objects !\n"); 306 list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) { 307 mutex_lock(&rdev->ddev->struct_mutex); 308 dev_err(rdev->dev, "%p %p %lu %lu force free\n", 309 &bo->gem_base, bo, (unsigned long)bo->gem_base.size, 310 *((unsigned long *)&bo->gem_base.refcount)); 311 mutex_lock(&bo->rdev->gem.mutex); 312 list_del_init(&bo->list); 313 mutex_unlock(&bo->rdev->gem.mutex); 314 /* this should unref the ttm bo */ 315 drm_gem_object_unreference(&bo->gem_base); 316 mutex_unlock(&rdev->ddev->struct_mutex); 317 } 318 } 319 320 int radeon_bo_init(struct radeon_device *rdev) 321 { 322 /* Add an MTRR for the VRAM */ 323 if (!rdev->fastfb_working) { 324 rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base, 325 rdev->mc.aper_size); 326 } 327 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n", 328 rdev->mc.mc_vram_size >> 20, 329 (unsigned long long)rdev->mc.aper_size >> 20); 330 DRM_INFO("RAM width %dbits %cDR\n", 331 rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S'); 332 return radeon_ttm_init(rdev); 333 } 334 335 void radeon_bo_fini(struct radeon_device *rdev) 336 { 337 radeon_ttm_fini(rdev); 338 arch_phys_wc_del(rdev->mc.vram_mtrr); 339 } 340 341 void radeon_bo_list_add_object(struct radeon_bo_list *lobj, 342 struct list_head *head) 343 { 344 if (lobj->written) { 345 list_add(&lobj->tv.head, head); 346 } else { 347 list_add_tail(&lobj->tv.head, head); 348 } 349 } 350 351 int radeon_bo_list_validate(struct ww_acquire_ctx *ticket, 352 struct list_head *head, int ring) 353 { 354 struct radeon_bo_list *lobj; 355 struct radeon_bo *bo; 356 u32 domain; 357 int r; 358 359 r = ttm_eu_reserve_buffers(ticket, head); 360 if (unlikely(r != 0)) { 361 return r; 362 } 363 list_for_each_entry(lobj, head, tv.head) { 364 bo = lobj->bo; 365 if (!bo->pin_count) { 366 domain = lobj->domain; 367 368 retry: 369 radeon_ttm_placement_from_domain(bo, domain); 370 if (ring == R600_RING_TYPE_UVD_INDEX) 371 radeon_uvd_force_into_uvd_segment(bo); 372 r = ttm_bo_validate(&bo->tbo, &bo->placement, 373 true, false); 374 if (unlikely(r)) { 375 if (r != -ERESTARTSYS && domain != lobj->alt_domain) { 376 domain = lobj->alt_domain; 377 goto retry; 378 } 379 ttm_eu_backoff_reservation(ticket, head); 380 return r; 381 } 382 } 383 lobj->gpu_offset = radeon_bo_gpu_offset(bo); 384 lobj->tiling_flags = bo->tiling_flags; 385 } 386 return 0; 387 } 388 389 int radeon_bo_fbdev_mmap(struct radeon_bo *bo, 390 struct vm_area_struct *vma) 391 { 392 return ttm_fbdev_mmap(vma, &bo->tbo); 393 } 394 395 int radeon_bo_get_surface_reg(struct radeon_bo *bo) 396 { 397 struct radeon_device *rdev = bo->rdev; 398 struct radeon_surface_reg *reg; 399 struct radeon_bo *old_object; 400 int steal; 401 int i; 402 403 lockdep_assert_held(&bo->tbo.resv->lock.base); 404 405 if (!bo->tiling_flags) 406 return 0; 407 408 if (bo->surface_reg >= 0) { 409 reg = &rdev->surface_regs[bo->surface_reg]; 410 i = bo->surface_reg; 411 goto out; 412 } 413 414 steal = -1; 415 for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) { 416 417 reg = &rdev->surface_regs[i]; 418 if (!reg->bo) 419 break; 420 421 old_object = reg->bo; 422 if (old_object->pin_count == 0) 423 steal = i; 424 } 425 426 /* if we are all out */ 427 if (i == RADEON_GEM_MAX_SURFACES) { 428 if (steal == -1) 429 return -ENOMEM; 430 /* find someone with a surface reg and nuke their BO */ 431 reg = &rdev->surface_regs[steal]; 432 old_object = reg->bo; 433 /* blow away the mapping */ 434 DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object); 435 ttm_bo_unmap_virtual(&old_object->tbo); 436 old_object->surface_reg = -1; 437 i = steal; 438 } 439 440 bo->surface_reg = i; 441 reg->bo = bo; 442 443 out: 444 radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch, 445 bo->tbo.mem.start << PAGE_SHIFT, 446 bo->tbo.num_pages << PAGE_SHIFT); 447 return 0; 448 } 449 450 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo) 451 { 452 struct radeon_device *rdev = bo->rdev; 453 struct radeon_surface_reg *reg; 454 455 if (bo->surface_reg == -1) 456 return; 457 458 reg = &rdev->surface_regs[bo->surface_reg]; 459 radeon_clear_surface_reg(rdev, bo->surface_reg); 460 461 reg->bo = NULL; 462 bo->surface_reg = -1; 463 } 464 465 int radeon_bo_set_tiling_flags(struct radeon_bo *bo, 466 uint32_t tiling_flags, uint32_t pitch) 467 { 468 struct radeon_device *rdev = bo->rdev; 469 int r; 470 471 if (rdev->family >= CHIP_CEDAR) { 472 unsigned bankw, bankh, mtaspect, tilesplit, stilesplit; 473 474 bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK; 475 bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK; 476 mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK; 477 tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK; 478 stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK; 479 switch (bankw) { 480 case 0: 481 case 1: 482 case 2: 483 case 4: 484 case 8: 485 break; 486 default: 487 return -EINVAL; 488 } 489 switch (bankh) { 490 case 0: 491 case 1: 492 case 2: 493 case 4: 494 case 8: 495 break; 496 default: 497 return -EINVAL; 498 } 499 switch (mtaspect) { 500 case 0: 501 case 1: 502 case 2: 503 case 4: 504 case 8: 505 break; 506 default: 507 return -EINVAL; 508 } 509 if (tilesplit > 6) { 510 return -EINVAL; 511 } 512 if (stilesplit > 6) { 513 return -EINVAL; 514 } 515 } 516 r = radeon_bo_reserve(bo, false); 517 if (unlikely(r != 0)) 518 return r; 519 bo->tiling_flags = tiling_flags; 520 bo->pitch = pitch; 521 radeon_bo_unreserve(bo); 522 return 0; 523 } 524 525 void radeon_bo_get_tiling_flags(struct radeon_bo *bo, 526 uint32_t *tiling_flags, 527 uint32_t *pitch) 528 { 529 lockdep_assert_held(&bo->tbo.resv->lock.base); 530 531 if (tiling_flags) 532 *tiling_flags = bo->tiling_flags; 533 if (pitch) 534 *pitch = bo->pitch; 535 } 536 537 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved, 538 bool force_drop) 539 { 540 if (!force_drop) 541 lockdep_assert_held(&bo->tbo.resv->lock.base); 542 543 if (!(bo->tiling_flags & RADEON_TILING_SURFACE)) 544 return 0; 545 546 if (force_drop) { 547 radeon_bo_clear_surface_reg(bo); 548 return 0; 549 } 550 551 if (bo->tbo.mem.mem_type != TTM_PL_VRAM) { 552 if (!has_moved) 553 return 0; 554 555 if (bo->surface_reg >= 0) 556 radeon_bo_clear_surface_reg(bo); 557 return 0; 558 } 559 560 if ((bo->surface_reg >= 0) && !has_moved) 561 return 0; 562 563 return radeon_bo_get_surface_reg(bo); 564 } 565 566 void radeon_bo_move_notify(struct ttm_buffer_object *bo, 567 struct ttm_mem_reg *mem) 568 { 569 struct radeon_bo *rbo; 570 if (!radeon_ttm_bo_is_radeon_bo(bo)) 571 return; 572 rbo = container_of(bo, struct radeon_bo, tbo); 573 radeon_bo_check_tiling(rbo, 0, 1); 574 radeon_vm_bo_invalidate(rbo->rdev, rbo); 575 } 576 577 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo) 578 { 579 struct radeon_device *rdev; 580 struct radeon_bo *rbo; 581 unsigned long offset, size; 582 int r; 583 584 if (!radeon_ttm_bo_is_radeon_bo(bo)) 585 return 0; 586 rbo = container_of(bo, struct radeon_bo, tbo); 587 radeon_bo_check_tiling(rbo, 0, 0); 588 rdev = rbo->rdev; 589 if (bo->mem.mem_type == TTM_PL_VRAM) { 590 size = bo->mem.num_pages << PAGE_SHIFT; 591 offset = bo->mem.start << PAGE_SHIFT; 592 if ((offset + size) > rdev->mc.visible_vram_size) { 593 /* hurrah the memory is not visible ! */ 594 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM); 595 rbo->placement.lpfn = rdev->mc.visible_vram_size >> PAGE_SHIFT; 596 r = ttm_bo_validate(bo, &rbo->placement, false, false); 597 if (unlikely(r != 0)) 598 return r; 599 offset = bo->mem.start << PAGE_SHIFT; 600 /* this should not happen */ 601 if ((offset + size) > rdev->mc.visible_vram_size) 602 return -EINVAL; 603 } 604 } 605 return 0; 606 } 607 608 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait) 609 { 610 int r; 611 612 r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, 0); 613 if (unlikely(r != 0)) 614 return r; 615 spin_lock(&bo->tbo.bdev->fence_lock); 616 if (mem_type) 617 *mem_type = bo->tbo.mem.mem_type; 618 if (bo->tbo.sync_obj) 619 r = ttm_bo_wait(&bo->tbo, true, true, no_wait); 620 spin_unlock(&bo->tbo.bdev->fence_lock); 621 ttm_bo_unreserve(&bo->tbo); 622 return r; 623 } 624