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->gem_base.driver_private = NULL; 146 bo->surface_reg = -1; 147 INIT_LIST_HEAD(&bo->list); 148 INIT_LIST_HEAD(&bo->va); 149 radeon_ttm_placement_from_domain(bo, domain); 150 /* Kernel allocation are uninterruptible */ 151 down_read(&rdev->pm.mclk_lock); 152 r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type, 153 &bo->placement, page_align, !kernel, NULL, 154 acc_size, sg, &radeon_ttm_bo_destroy); 155 up_read(&rdev->pm.mclk_lock); 156 if (unlikely(r != 0)) { 157 return r; 158 } 159 *bo_ptr = bo; 160 161 trace_radeon_bo_create(bo); 162 163 return 0; 164 } 165 166 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr) 167 { 168 bool is_iomem; 169 int r; 170 171 if (bo->kptr) { 172 if (ptr) { 173 *ptr = bo->kptr; 174 } 175 return 0; 176 } 177 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap); 178 if (r) { 179 return r; 180 } 181 bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem); 182 if (ptr) { 183 *ptr = bo->kptr; 184 } 185 radeon_bo_check_tiling(bo, 0, 0); 186 return 0; 187 } 188 189 void radeon_bo_kunmap(struct radeon_bo *bo) 190 { 191 if (bo->kptr == NULL) 192 return; 193 bo->kptr = NULL; 194 radeon_bo_check_tiling(bo, 0, 0); 195 ttm_bo_kunmap(&bo->kmap); 196 } 197 198 void radeon_bo_unref(struct radeon_bo **bo) 199 { 200 struct ttm_buffer_object *tbo; 201 struct radeon_device *rdev; 202 203 if ((*bo) == NULL) 204 return; 205 rdev = (*bo)->rdev; 206 tbo = &((*bo)->tbo); 207 down_read(&rdev->pm.mclk_lock); 208 ttm_bo_unref(&tbo); 209 up_read(&rdev->pm.mclk_lock); 210 if (tbo == NULL) 211 *bo = NULL; 212 } 213 214 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset, 215 u64 *gpu_addr) 216 { 217 int r, i; 218 219 if (bo->pin_count) { 220 bo->pin_count++; 221 if (gpu_addr) 222 *gpu_addr = radeon_bo_gpu_offset(bo); 223 224 if (max_offset != 0) { 225 u64 domain_start; 226 227 if (domain == RADEON_GEM_DOMAIN_VRAM) 228 domain_start = bo->rdev->mc.vram_start; 229 else 230 domain_start = bo->rdev->mc.gtt_start; 231 WARN_ON_ONCE(max_offset < 232 (radeon_bo_gpu_offset(bo) - domain_start)); 233 } 234 235 return 0; 236 } 237 radeon_ttm_placement_from_domain(bo, domain); 238 if (domain == RADEON_GEM_DOMAIN_VRAM) { 239 /* force to pin into visible video ram */ 240 bo->placement.lpfn = bo->rdev->mc.visible_vram_size >> PAGE_SHIFT; 241 } 242 if (max_offset) { 243 u64 lpfn = max_offset >> PAGE_SHIFT; 244 245 if (!bo->placement.lpfn) 246 bo->placement.lpfn = bo->rdev->mc.gtt_size >> PAGE_SHIFT; 247 248 if (lpfn < bo->placement.lpfn) 249 bo->placement.lpfn = lpfn; 250 } 251 for (i = 0; i < bo->placement.num_placement; i++) 252 bo->placements[i] |= TTM_PL_FLAG_NO_EVICT; 253 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false); 254 if (likely(r == 0)) { 255 bo->pin_count = 1; 256 if (gpu_addr != NULL) 257 *gpu_addr = radeon_bo_gpu_offset(bo); 258 } 259 if (unlikely(r != 0)) 260 dev_err(bo->rdev->dev, "%p pin failed\n", bo); 261 return r; 262 } 263 264 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr) 265 { 266 return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr); 267 } 268 269 int radeon_bo_unpin(struct radeon_bo *bo) 270 { 271 int r, i; 272 273 if (!bo->pin_count) { 274 dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo); 275 return 0; 276 } 277 bo->pin_count--; 278 if (bo->pin_count) 279 return 0; 280 for (i = 0; i < bo->placement.num_placement; i++) 281 bo->placements[i] &= ~TTM_PL_FLAG_NO_EVICT; 282 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false); 283 if (unlikely(r != 0)) 284 dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo); 285 return r; 286 } 287 288 int radeon_bo_evict_vram(struct radeon_device *rdev) 289 { 290 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */ 291 if (0 && (rdev->flags & RADEON_IS_IGP)) { 292 if (rdev->mc.igp_sideport_enabled == false) 293 /* Useless to evict on IGP chips */ 294 return 0; 295 } 296 return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM); 297 } 298 299 void radeon_bo_force_delete(struct radeon_device *rdev) 300 { 301 struct radeon_bo *bo, *n; 302 303 if (list_empty(&rdev->gem.objects)) { 304 return; 305 } 306 dev_err(rdev->dev, "Userspace still has active objects !\n"); 307 list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) { 308 mutex_lock(&rdev->ddev->struct_mutex); 309 dev_err(rdev->dev, "%p %p %lu %lu force free\n", 310 &bo->gem_base, bo, (unsigned long)bo->gem_base.size, 311 *((unsigned long *)&bo->gem_base.refcount)); 312 mutex_lock(&bo->rdev->gem.mutex); 313 list_del_init(&bo->list); 314 mutex_unlock(&bo->rdev->gem.mutex); 315 /* this should unref the ttm bo */ 316 drm_gem_object_unreference(&bo->gem_base); 317 mutex_unlock(&rdev->ddev->struct_mutex); 318 } 319 } 320 321 int radeon_bo_init(struct radeon_device *rdev) 322 { 323 /* Add an MTRR for the VRAM */ 324 if (!rdev->fastfb_working) { 325 rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base, 326 rdev->mc.aper_size); 327 } 328 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n", 329 rdev->mc.mc_vram_size >> 20, 330 (unsigned long long)rdev->mc.aper_size >> 20); 331 DRM_INFO("RAM width %dbits %cDR\n", 332 rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S'); 333 return radeon_ttm_init(rdev); 334 } 335 336 void radeon_bo_fini(struct radeon_device *rdev) 337 { 338 radeon_ttm_fini(rdev); 339 arch_phys_wc_del(rdev->mc.vram_mtrr); 340 } 341 342 void radeon_bo_list_add_object(struct radeon_bo_list *lobj, 343 struct list_head *head) 344 { 345 if (lobj->written) { 346 list_add(&lobj->tv.head, head); 347 } else { 348 list_add_tail(&lobj->tv.head, head); 349 } 350 } 351 352 int radeon_bo_list_validate(struct ww_acquire_ctx *ticket, 353 struct list_head *head, int ring) 354 { 355 struct radeon_bo_list *lobj; 356 struct radeon_bo *bo; 357 u32 domain; 358 int r; 359 360 r = ttm_eu_reserve_buffers(ticket, head); 361 if (unlikely(r != 0)) { 362 return r; 363 } 364 list_for_each_entry(lobj, head, tv.head) { 365 bo = lobj->bo; 366 if (!bo->pin_count) { 367 domain = lobj->domain; 368 369 retry: 370 radeon_ttm_placement_from_domain(bo, domain); 371 if (ring == R600_RING_TYPE_UVD_INDEX) 372 radeon_uvd_force_into_uvd_segment(bo); 373 r = ttm_bo_validate(&bo->tbo, &bo->placement, 374 true, false); 375 if (unlikely(r)) { 376 if (r != -ERESTARTSYS && domain != lobj->alt_domain) { 377 domain = lobj->alt_domain; 378 goto retry; 379 } 380 ttm_eu_backoff_reservation(ticket, head); 381 return r; 382 } 383 } 384 lobj->gpu_offset = radeon_bo_gpu_offset(bo); 385 lobj->tiling_flags = bo->tiling_flags; 386 } 387 return 0; 388 } 389 390 int radeon_bo_fbdev_mmap(struct radeon_bo *bo, 391 struct vm_area_struct *vma) 392 { 393 return ttm_fbdev_mmap(vma, &bo->tbo); 394 } 395 396 int radeon_bo_get_surface_reg(struct radeon_bo *bo) 397 { 398 struct radeon_device *rdev = bo->rdev; 399 struct radeon_surface_reg *reg; 400 struct radeon_bo *old_object; 401 int steal; 402 int i; 403 404 lockdep_assert_held(&bo->tbo.resv->lock.base); 405 406 if (!bo->tiling_flags) 407 return 0; 408 409 if (bo->surface_reg >= 0) { 410 reg = &rdev->surface_regs[bo->surface_reg]; 411 i = bo->surface_reg; 412 goto out; 413 } 414 415 steal = -1; 416 for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) { 417 418 reg = &rdev->surface_regs[i]; 419 if (!reg->bo) 420 break; 421 422 old_object = reg->bo; 423 if (old_object->pin_count == 0) 424 steal = i; 425 } 426 427 /* if we are all out */ 428 if (i == RADEON_GEM_MAX_SURFACES) { 429 if (steal == -1) 430 return -ENOMEM; 431 /* find someone with a surface reg and nuke their BO */ 432 reg = &rdev->surface_regs[steal]; 433 old_object = reg->bo; 434 /* blow away the mapping */ 435 DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object); 436 ttm_bo_unmap_virtual(&old_object->tbo); 437 old_object->surface_reg = -1; 438 i = steal; 439 } 440 441 bo->surface_reg = i; 442 reg->bo = bo; 443 444 out: 445 radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch, 446 bo->tbo.mem.start << PAGE_SHIFT, 447 bo->tbo.num_pages << PAGE_SHIFT); 448 return 0; 449 } 450 451 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo) 452 { 453 struct radeon_device *rdev = bo->rdev; 454 struct radeon_surface_reg *reg; 455 456 if (bo->surface_reg == -1) 457 return; 458 459 reg = &rdev->surface_regs[bo->surface_reg]; 460 radeon_clear_surface_reg(rdev, bo->surface_reg); 461 462 reg->bo = NULL; 463 bo->surface_reg = -1; 464 } 465 466 int radeon_bo_set_tiling_flags(struct radeon_bo *bo, 467 uint32_t tiling_flags, uint32_t pitch) 468 { 469 struct radeon_device *rdev = bo->rdev; 470 int r; 471 472 if (rdev->family >= CHIP_CEDAR) { 473 unsigned bankw, bankh, mtaspect, tilesplit, stilesplit; 474 475 bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK; 476 bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK; 477 mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK; 478 tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK; 479 stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK; 480 switch (bankw) { 481 case 0: 482 case 1: 483 case 2: 484 case 4: 485 case 8: 486 break; 487 default: 488 return -EINVAL; 489 } 490 switch (bankh) { 491 case 0: 492 case 1: 493 case 2: 494 case 4: 495 case 8: 496 break; 497 default: 498 return -EINVAL; 499 } 500 switch (mtaspect) { 501 case 0: 502 case 1: 503 case 2: 504 case 4: 505 case 8: 506 break; 507 default: 508 return -EINVAL; 509 } 510 if (tilesplit > 6) { 511 return -EINVAL; 512 } 513 if (stilesplit > 6) { 514 return -EINVAL; 515 } 516 } 517 r = radeon_bo_reserve(bo, false); 518 if (unlikely(r != 0)) 519 return r; 520 bo->tiling_flags = tiling_flags; 521 bo->pitch = pitch; 522 radeon_bo_unreserve(bo); 523 return 0; 524 } 525 526 void radeon_bo_get_tiling_flags(struct radeon_bo *bo, 527 uint32_t *tiling_flags, 528 uint32_t *pitch) 529 { 530 lockdep_assert_held(&bo->tbo.resv->lock.base); 531 532 if (tiling_flags) 533 *tiling_flags = bo->tiling_flags; 534 if (pitch) 535 *pitch = bo->pitch; 536 } 537 538 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved, 539 bool force_drop) 540 { 541 if (!force_drop) 542 lockdep_assert_held(&bo->tbo.resv->lock.base); 543 544 if (!(bo->tiling_flags & RADEON_TILING_SURFACE)) 545 return 0; 546 547 if (force_drop) { 548 radeon_bo_clear_surface_reg(bo); 549 return 0; 550 } 551 552 if (bo->tbo.mem.mem_type != TTM_PL_VRAM) { 553 if (!has_moved) 554 return 0; 555 556 if (bo->surface_reg >= 0) 557 radeon_bo_clear_surface_reg(bo); 558 return 0; 559 } 560 561 if ((bo->surface_reg >= 0) && !has_moved) 562 return 0; 563 564 return radeon_bo_get_surface_reg(bo); 565 } 566 567 void radeon_bo_move_notify(struct ttm_buffer_object *bo, 568 struct ttm_mem_reg *mem) 569 { 570 struct radeon_bo *rbo; 571 if (!radeon_ttm_bo_is_radeon_bo(bo)) 572 return; 573 rbo = container_of(bo, struct radeon_bo, tbo); 574 radeon_bo_check_tiling(rbo, 0, 1); 575 radeon_vm_bo_invalidate(rbo->rdev, rbo); 576 } 577 578 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo) 579 { 580 struct radeon_device *rdev; 581 struct radeon_bo *rbo; 582 unsigned long offset, size; 583 int r; 584 585 if (!radeon_ttm_bo_is_radeon_bo(bo)) 586 return 0; 587 rbo = container_of(bo, struct radeon_bo, tbo); 588 radeon_bo_check_tiling(rbo, 0, 0); 589 rdev = rbo->rdev; 590 if (bo->mem.mem_type == TTM_PL_VRAM) { 591 size = bo->mem.num_pages << PAGE_SHIFT; 592 offset = bo->mem.start << PAGE_SHIFT; 593 if ((offset + size) > rdev->mc.visible_vram_size) { 594 /* hurrah the memory is not visible ! */ 595 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM); 596 rbo->placement.lpfn = rdev->mc.visible_vram_size >> PAGE_SHIFT; 597 r = ttm_bo_validate(bo, &rbo->placement, false, false); 598 if (unlikely(r != 0)) 599 return r; 600 offset = bo->mem.start << PAGE_SHIFT; 601 /* this should not happen */ 602 if ((offset + size) > rdev->mc.visible_vram_size) 603 return -EINVAL; 604 } 605 } 606 return 0; 607 } 608 609 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait) 610 { 611 int r; 612 613 r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, 0); 614 if (unlikely(r != 0)) 615 return r; 616 spin_lock(&bo->tbo.bdev->fence_lock); 617 if (mem_type) 618 *mem_type = bo->tbo.mem.mem_type; 619 if (bo->tbo.sync_obj) 620 r = ttm_bo_wait(&bo->tbo, true, true, no_wait); 621 spin_unlock(&bo->tbo.bdev->fence_lock); 622 ttm_bo_unreserve(&bo->tbo); 623 return r; 624 } 625