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 <drm/drm_cache.h> 37 #include "radeon.h" 38 #include "radeon_trace.h" 39 40 41 int radeon_ttm_init(struct radeon_device *rdev); 42 void radeon_ttm_fini(struct radeon_device *rdev); 43 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo); 44 45 /* 46 * To exclude mutual BO access we rely on bo_reserve exclusion, as all 47 * function are calling it. 48 */ 49 50 static void radeon_update_memory_usage(struct radeon_bo *bo, 51 unsigned mem_type, int sign) 52 { 53 struct radeon_device *rdev = bo->rdev; 54 u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT; 55 56 switch (mem_type) { 57 case TTM_PL_TT: 58 if (sign > 0) 59 atomic64_add(size, &rdev->gtt_usage); 60 else 61 atomic64_sub(size, &rdev->gtt_usage); 62 break; 63 case TTM_PL_VRAM: 64 if (sign > 0) 65 atomic64_add(size, &rdev->vram_usage); 66 else 67 atomic64_sub(size, &rdev->vram_usage); 68 break; 69 } 70 } 71 72 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo) 73 { 74 struct radeon_bo *bo; 75 76 bo = container_of(tbo, struct radeon_bo, tbo); 77 78 radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1); 79 80 mutex_lock(&bo->rdev->gem.mutex); 81 list_del_init(&bo->list); 82 mutex_unlock(&bo->rdev->gem.mutex); 83 radeon_bo_clear_surface_reg(bo); 84 WARN_ON_ONCE(!list_empty(&bo->va)); 85 drm_gem_object_release(&bo->gem_base); 86 kfree(bo); 87 } 88 89 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo) 90 { 91 if (bo->destroy == &radeon_ttm_bo_destroy) 92 return true; 93 return false; 94 } 95 96 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain) 97 { 98 u32 c = 0, i; 99 100 rbo->placement.placement = rbo->placements; 101 rbo->placement.busy_placement = rbo->placements; 102 if (domain & RADEON_GEM_DOMAIN_VRAM) { 103 /* Try placing BOs which don't need CPU access outside of the 104 * CPU accessible part of VRAM 105 */ 106 if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) && 107 rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) { 108 rbo->placements[c].fpfn = 109 rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT; 110 rbo->placements[c++].flags = TTM_PL_FLAG_WC | 111 TTM_PL_FLAG_UNCACHED | 112 TTM_PL_FLAG_VRAM; 113 } 114 115 rbo->placements[c].fpfn = 0; 116 rbo->placements[c++].flags = TTM_PL_FLAG_WC | 117 TTM_PL_FLAG_UNCACHED | 118 TTM_PL_FLAG_VRAM; 119 } 120 121 if (domain & RADEON_GEM_DOMAIN_GTT) { 122 if (rbo->flags & RADEON_GEM_GTT_UC) { 123 rbo->placements[c].fpfn = 0; 124 rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED | 125 TTM_PL_FLAG_TT; 126 127 } else if ((rbo->flags & RADEON_GEM_GTT_WC) || 128 (rbo->rdev->flags & RADEON_IS_AGP)) { 129 rbo->placements[c].fpfn = 0; 130 rbo->placements[c++].flags = TTM_PL_FLAG_WC | 131 TTM_PL_FLAG_UNCACHED | 132 TTM_PL_FLAG_TT; 133 } else { 134 rbo->placements[c].fpfn = 0; 135 rbo->placements[c++].flags = TTM_PL_FLAG_CACHED | 136 TTM_PL_FLAG_TT; 137 } 138 } 139 140 if (domain & RADEON_GEM_DOMAIN_CPU) { 141 if (rbo->flags & RADEON_GEM_GTT_UC) { 142 rbo->placements[c].fpfn = 0; 143 rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED | 144 TTM_PL_FLAG_SYSTEM; 145 146 } else if ((rbo->flags & RADEON_GEM_GTT_WC) || 147 rbo->rdev->flags & RADEON_IS_AGP) { 148 rbo->placements[c].fpfn = 0; 149 rbo->placements[c++].flags = TTM_PL_FLAG_WC | 150 TTM_PL_FLAG_UNCACHED | 151 TTM_PL_FLAG_SYSTEM; 152 } else { 153 rbo->placements[c].fpfn = 0; 154 rbo->placements[c++].flags = TTM_PL_FLAG_CACHED | 155 TTM_PL_FLAG_SYSTEM; 156 } 157 } 158 if (!c) { 159 rbo->placements[c].fpfn = 0; 160 rbo->placements[c++].flags = TTM_PL_MASK_CACHING | 161 TTM_PL_FLAG_SYSTEM; 162 } 163 164 rbo->placement.num_placement = c; 165 rbo->placement.num_busy_placement = c; 166 167 for (i = 0; i < c; ++i) { 168 if ((rbo->flags & RADEON_GEM_CPU_ACCESS) && 169 (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) && 170 !rbo->placements[i].fpfn) 171 rbo->placements[i].lpfn = 172 rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT; 173 else 174 rbo->placements[i].lpfn = 0; 175 } 176 } 177 178 int radeon_bo_create(struct radeon_device *rdev, 179 unsigned long size, int byte_align, bool kernel, 180 u32 domain, u32 flags, struct sg_table *sg, 181 struct reservation_object *resv, 182 struct radeon_bo **bo_ptr) 183 { 184 struct radeon_bo *bo; 185 enum ttm_bo_type type; 186 unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT; 187 size_t acc_size; 188 int r; 189 190 size = ALIGN(size, PAGE_SIZE); 191 192 if (kernel) { 193 type = ttm_bo_type_kernel; 194 } else if (sg) { 195 type = ttm_bo_type_sg; 196 } else { 197 type = ttm_bo_type_device; 198 } 199 *bo_ptr = NULL; 200 201 acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size, 202 sizeof(struct radeon_bo)); 203 204 bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL); 205 if (bo == NULL) 206 return -ENOMEM; 207 r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size); 208 if (unlikely(r)) { 209 kfree(bo); 210 return r; 211 } 212 bo->rdev = rdev; 213 bo->surface_reg = -1; 214 INIT_LIST_HEAD(&bo->list); 215 INIT_LIST_HEAD(&bo->va); 216 bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM | 217 RADEON_GEM_DOMAIN_GTT | 218 RADEON_GEM_DOMAIN_CPU); 219 220 bo->flags = flags; 221 /* PCI GART is always snooped */ 222 if (!(rdev->flags & RADEON_IS_PCIE)) 223 bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC); 224 225 /* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx 226 * See https://bugs.freedesktop.org/show_bug.cgi?id=91268 227 */ 228 if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635) 229 bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC); 230 231 #ifdef CONFIG_X86_32 232 /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit 233 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627 234 */ 235 bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC); 236 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT) 237 /* Don't try to enable write-combining when it can't work, or things 238 * may be slow 239 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758 240 */ 241 242 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \ 243 thanks to write-combining 244 245 if (bo->flags & RADEON_GEM_GTT_WC) 246 DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for " 247 "better performance thanks to write-combining\n"); 248 bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC); 249 #else 250 /* For architectures that don't support WC memory, 251 * mask out the WC flag from the BO 252 */ 253 if (!drm_arch_can_wc_memory()) 254 bo->flags &= ~RADEON_GEM_GTT_WC; 255 #endif 256 257 radeon_ttm_placement_from_domain(bo, domain); 258 /* Kernel allocation are uninterruptible */ 259 down_read(&rdev->pm.mclk_lock); 260 r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type, 261 &bo->placement, page_align, !kernel, NULL, 262 acc_size, sg, resv, &radeon_ttm_bo_destroy); 263 up_read(&rdev->pm.mclk_lock); 264 if (unlikely(r != 0)) { 265 return r; 266 } 267 *bo_ptr = bo; 268 269 trace_radeon_bo_create(bo); 270 271 return 0; 272 } 273 274 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr) 275 { 276 bool is_iomem; 277 int r; 278 279 if (bo->kptr) { 280 if (ptr) { 281 *ptr = bo->kptr; 282 } 283 return 0; 284 } 285 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap); 286 if (r) { 287 return r; 288 } 289 bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem); 290 if (ptr) { 291 *ptr = bo->kptr; 292 } 293 radeon_bo_check_tiling(bo, 0, 0); 294 return 0; 295 } 296 297 void radeon_bo_kunmap(struct radeon_bo *bo) 298 { 299 if (bo->kptr == NULL) 300 return; 301 bo->kptr = NULL; 302 radeon_bo_check_tiling(bo, 0, 0); 303 ttm_bo_kunmap(&bo->kmap); 304 } 305 306 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo) 307 { 308 if (bo == NULL) 309 return NULL; 310 311 ttm_bo_reference(&bo->tbo); 312 return bo; 313 } 314 315 void radeon_bo_unref(struct radeon_bo **bo) 316 { 317 struct ttm_buffer_object *tbo; 318 struct radeon_device *rdev; 319 320 if ((*bo) == NULL) 321 return; 322 rdev = (*bo)->rdev; 323 tbo = &((*bo)->tbo); 324 ttm_bo_unref(&tbo); 325 if (tbo == NULL) 326 *bo = NULL; 327 } 328 329 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset, 330 u64 *gpu_addr) 331 { 332 int r, i; 333 334 if (radeon_ttm_tt_has_userptr(bo->tbo.ttm)) 335 return -EPERM; 336 337 if (bo->pin_count) { 338 bo->pin_count++; 339 if (gpu_addr) 340 *gpu_addr = radeon_bo_gpu_offset(bo); 341 342 if (max_offset != 0) { 343 u64 domain_start; 344 345 if (domain == RADEON_GEM_DOMAIN_VRAM) 346 domain_start = bo->rdev->mc.vram_start; 347 else 348 domain_start = bo->rdev->mc.gtt_start; 349 WARN_ON_ONCE(max_offset < 350 (radeon_bo_gpu_offset(bo) - domain_start)); 351 } 352 353 return 0; 354 } 355 if (bo->prime_shared_count && domain == RADEON_GEM_DOMAIN_VRAM) { 356 /* A BO shared as a dma-buf cannot be sensibly migrated to VRAM */ 357 return -EINVAL; 358 } 359 360 radeon_ttm_placement_from_domain(bo, domain); 361 for (i = 0; i < bo->placement.num_placement; i++) { 362 /* force to pin into visible video ram */ 363 if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) && 364 !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) && 365 (!max_offset || max_offset > bo->rdev->mc.visible_vram_size)) 366 bo->placements[i].lpfn = 367 bo->rdev->mc.visible_vram_size >> PAGE_SHIFT; 368 else 369 bo->placements[i].lpfn = max_offset >> PAGE_SHIFT; 370 371 bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT; 372 } 373 374 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false); 375 if (likely(r == 0)) { 376 bo->pin_count = 1; 377 if (gpu_addr != NULL) 378 *gpu_addr = radeon_bo_gpu_offset(bo); 379 if (domain == RADEON_GEM_DOMAIN_VRAM) 380 bo->rdev->vram_pin_size += radeon_bo_size(bo); 381 else 382 bo->rdev->gart_pin_size += radeon_bo_size(bo); 383 } else { 384 dev_err(bo->rdev->dev, "%p pin failed\n", bo); 385 } 386 return r; 387 } 388 389 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr) 390 { 391 return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr); 392 } 393 394 int radeon_bo_unpin(struct radeon_bo *bo) 395 { 396 int r, i; 397 398 if (!bo->pin_count) { 399 dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo); 400 return 0; 401 } 402 bo->pin_count--; 403 if (bo->pin_count) 404 return 0; 405 for (i = 0; i < bo->placement.num_placement; i++) { 406 bo->placements[i].lpfn = 0; 407 bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT; 408 } 409 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false); 410 if (likely(r == 0)) { 411 if (bo->tbo.mem.mem_type == TTM_PL_VRAM) 412 bo->rdev->vram_pin_size -= radeon_bo_size(bo); 413 else 414 bo->rdev->gart_pin_size -= radeon_bo_size(bo); 415 } else { 416 dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo); 417 } 418 return r; 419 } 420 421 int radeon_bo_evict_vram(struct radeon_device *rdev) 422 { 423 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */ 424 if (0 && (rdev->flags & RADEON_IS_IGP)) { 425 if (rdev->mc.igp_sideport_enabled == false) 426 /* Useless to evict on IGP chips */ 427 return 0; 428 } 429 return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM); 430 } 431 432 void radeon_bo_force_delete(struct radeon_device *rdev) 433 { 434 struct radeon_bo *bo, *n; 435 436 if (list_empty(&rdev->gem.objects)) { 437 return; 438 } 439 dev_err(rdev->dev, "Userspace still has active objects !\n"); 440 list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) { 441 dev_err(rdev->dev, "%p %p %lu %lu force free\n", 442 &bo->gem_base, bo, (unsigned long)bo->gem_base.size, 443 *((unsigned long *)&bo->gem_base.refcount)); 444 mutex_lock(&bo->rdev->gem.mutex); 445 list_del_init(&bo->list); 446 mutex_unlock(&bo->rdev->gem.mutex); 447 /* this should unref the ttm bo */ 448 drm_gem_object_put_unlocked(&bo->gem_base); 449 } 450 } 451 452 int radeon_bo_init(struct radeon_device *rdev) 453 { 454 /* reserve PAT memory space to WC for VRAM */ 455 arch_io_reserve_memtype_wc(rdev->mc.aper_base, 456 rdev->mc.aper_size); 457 458 /* Add an MTRR for the VRAM */ 459 if (!rdev->fastfb_working) { 460 rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base, 461 rdev->mc.aper_size); 462 } 463 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n", 464 rdev->mc.mc_vram_size >> 20, 465 (unsigned long long)rdev->mc.aper_size >> 20); 466 DRM_INFO("RAM width %dbits %cDR\n", 467 rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S'); 468 return radeon_ttm_init(rdev); 469 } 470 471 void radeon_bo_fini(struct radeon_device *rdev) 472 { 473 radeon_ttm_fini(rdev); 474 arch_phys_wc_del(rdev->mc.vram_mtrr); 475 arch_io_free_memtype_wc(rdev->mc.aper_base, rdev->mc.aper_size); 476 } 477 478 /* Returns how many bytes TTM can move per IB. 479 */ 480 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev) 481 { 482 u64 real_vram_size = rdev->mc.real_vram_size; 483 u64 vram_usage = atomic64_read(&rdev->vram_usage); 484 485 /* This function is based on the current VRAM usage. 486 * 487 * - If all of VRAM is free, allow relocating the number of bytes that 488 * is equal to 1/4 of the size of VRAM for this IB. 489 490 * - If more than one half of VRAM is occupied, only allow relocating 491 * 1 MB of data for this IB. 492 * 493 * - From 0 to one half of used VRAM, the threshold decreases 494 * linearly. 495 * __________________ 496 * 1/4 of -|\ | 497 * VRAM | \ | 498 * | \ | 499 * | \ | 500 * | \ | 501 * | \ | 502 * | \ | 503 * | \________|1 MB 504 * |----------------| 505 * VRAM 0 % 100 % 506 * used used 507 * 508 * Note: It's a threshold, not a limit. The threshold must be crossed 509 * for buffer relocations to stop, so any buffer of an arbitrary size 510 * can be moved as long as the threshold isn't crossed before 511 * the relocation takes place. We don't want to disable buffer 512 * relocations completely. 513 * 514 * The idea is that buffers should be placed in VRAM at creation time 515 * and TTM should only do a minimum number of relocations during 516 * command submission. In practice, you need to submit at least 517 * a dozen IBs to move all buffers to VRAM if they are in GTT. 518 * 519 * Also, things can get pretty crazy under memory pressure and actual 520 * VRAM usage can change a lot, so playing safe even at 50% does 521 * consistently increase performance. 522 */ 523 524 u64 half_vram = real_vram_size >> 1; 525 u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage; 526 u64 bytes_moved_threshold = half_free_vram >> 1; 527 return max(bytes_moved_threshold, 1024*1024ull); 528 } 529 530 int radeon_bo_list_validate(struct radeon_device *rdev, 531 struct ww_acquire_ctx *ticket, 532 struct list_head *head, int ring) 533 { 534 struct radeon_bo_list *lobj; 535 struct list_head duplicates; 536 int r; 537 u64 bytes_moved = 0, initial_bytes_moved; 538 u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev); 539 540 INIT_LIST_HEAD(&duplicates); 541 r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates); 542 if (unlikely(r != 0)) { 543 return r; 544 } 545 546 list_for_each_entry(lobj, head, tv.head) { 547 struct radeon_bo *bo = lobj->robj; 548 if (!bo->pin_count) { 549 u32 domain = lobj->preferred_domains; 550 u32 allowed = lobj->allowed_domains; 551 u32 current_domain = 552 radeon_mem_type_to_domain(bo->tbo.mem.mem_type); 553 554 /* Check if this buffer will be moved and don't move it 555 * if we have moved too many buffers for this IB already. 556 * 557 * Note that this allows moving at least one buffer of 558 * any size, because it doesn't take the current "bo" 559 * into account. We don't want to disallow buffer moves 560 * completely. 561 */ 562 if ((allowed & current_domain) != 0 && 563 (domain & current_domain) == 0 && /* will be moved */ 564 bytes_moved > bytes_moved_threshold) { 565 /* don't move it */ 566 domain = current_domain; 567 } 568 569 retry: 570 radeon_ttm_placement_from_domain(bo, domain); 571 if (ring == R600_RING_TYPE_UVD_INDEX) 572 radeon_uvd_force_into_uvd_segment(bo, allowed); 573 574 initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved); 575 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false); 576 bytes_moved += atomic64_read(&rdev->num_bytes_moved) - 577 initial_bytes_moved; 578 579 if (unlikely(r)) { 580 if (r != -ERESTARTSYS && 581 domain != lobj->allowed_domains) { 582 domain = lobj->allowed_domains; 583 goto retry; 584 } 585 ttm_eu_backoff_reservation(ticket, head); 586 return r; 587 } 588 } 589 lobj->gpu_offset = radeon_bo_gpu_offset(bo); 590 lobj->tiling_flags = bo->tiling_flags; 591 } 592 593 list_for_each_entry(lobj, &duplicates, tv.head) { 594 lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj); 595 lobj->tiling_flags = lobj->robj->tiling_flags; 596 } 597 598 return 0; 599 } 600 601 int radeon_bo_get_surface_reg(struct radeon_bo *bo) 602 { 603 struct radeon_device *rdev = bo->rdev; 604 struct radeon_surface_reg *reg; 605 struct radeon_bo *old_object; 606 int steal; 607 int i; 608 609 lockdep_assert_held(&bo->tbo.resv->lock.base); 610 611 if (!bo->tiling_flags) 612 return 0; 613 614 if (bo->surface_reg >= 0) { 615 reg = &rdev->surface_regs[bo->surface_reg]; 616 i = bo->surface_reg; 617 goto out; 618 } 619 620 steal = -1; 621 for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) { 622 623 reg = &rdev->surface_regs[i]; 624 if (!reg->bo) 625 break; 626 627 old_object = reg->bo; 628 if (old_object->pin_count == 0) 629 steal = i; 630 } 631 632 /* if we are all out */ 633 if (i == RADEON_GEM_MAX_SURFACES) { 634 if (steal == -1) 635 return -ENOMEM; 636 /* find someone with a surface reg and nuke their BO */ 637 reg = &rdev->surface_regs[steal]; 638 old_object = reg->bo; 639 /* blow away the mapping */ 640 DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object); 641 ttm_bo_unmap_virtual(&old_object->tbo); 642 old_object->surface_reg = -1; 643 i = steal; 644 } 645 646 bo->surface_reg = i; 647 reg->bo = bo; 648 649 out: 650 radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch, 651 bo->tbo.mem.start << PAGE_SHIFT, 652 bo->tbo.num_pages << PAGE_SHIFT); 653 return 0; 654 } 655 656 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo) 657 { 658 struct radeon_device *rdev = bo->rdev; 659 struct radeon_surface_reg *reg; 660 661 if (bo->surface_reg == -1) 662 return; 663 664 reg = &rdev->surface_regs[bo->surface_reg]; 665 radeon_clear_surface_reg(rdev, bo->surface_reg); 666 667 reg->bo = NULL; 668 bo->surface_reg = -1; 669 } 670 671 int radeon_bo_set_tiling_flags(struct radeon_bo *bo, 672 uint32_t tiling_flags, uint32_t pitch) 673 { 674 struct radeon_device *rdev = bo->rdev; 675 int r; 676 677 if (rdev->family >= CHIP_CEDAR) { 678 unsigned bankw, bankh, mtaspect, tilesplit, stilesplit; 679 680 bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK; 681 bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK; 682 mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK; 683 tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK; 684 stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK; 685 switch (bankw) { 686 case 0: 687 case 1: 688 case 2: 689 case 4: 690 case 8: 691 break; 692 default: 693 return -EINVAL; 694 } 695 switch (bankh) { 696 case 0: 697 case 1: 698 case 2: 699 case 4: 700 case 8: 701 break; 702 default: 703 return -EINVAL; 704 } 705 switch (mtaspect) { 706 case 0: 707 case 1: 708 case 2: 709 case 4: 710 case 8: 711 break; 712 default: 713 return -EINVAL; 714 } 715 if (tilesplit > 6) { 716 return -EINVAL; 717 } 718 if (stilesplit > 6) { 719 return -EINVAL; 720 } 721 } 722 r = radeon_bo_reserve(bo, false); 723 if (unlikely(r != 0)) 724 return r; 725 bo->tiling_flags = tiling_flags; 726 bo->pitch = pitch; 727 radeon_bo_unreserve(bo); 728 return 0; 729 } 730 731 void radeon_bo_get_tiling_flags(struct radeon_bo *bo, 732 uint32_t *tiling_flags, 733 uint32_t *pitch) 734 { 735 lockdep_assert_held(&bo->tbo.resv->lock.base); 736 737 if (tiling_flags) 738 *tiling_flags = bo->tiling_flags; 739 if (pitch) 740 *pitch = bo->pitch; 741 } 742 743 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved, 744 bool force_drop) 745 { 746 if (!force_drop) 747 lockdep_assert_held(&bo->tbo.resv->lock.base); 748 749 if (!(bo->tiling_flags & RADEON_TILING_SURFACE)) 750 return 0; 751 752 if (force_drop) { 753 radeon_bo_clear_surface_reg(bo); 754 return 0; 755 } 756 757 if (bo->tbo.mem.mem_type != TTM_PL_VRAM) { 758 if (!has_moved) 759 return 0; 760 761 if (bo->surface_reg >= 0) 762 radeon_bo_clear_surface_reg(bo); 763 return 0; 764 } 765 766 if ((bo->surface_reg >= 0) && !has_moved) 767 return 0; 768 769 return radeon_bo_get_surface_reg(bo); 770 } 771 772 void radeon_bo_move_notify(struct ttm_buffer_object *bo, 773 bool evict, 774 struct ttm_mem_reg *new_mem) 775 { 776 struct radeon_bo *rbo; 777 778 if (!radeon_ttm_bo_is_radeon_bo(bo)) 779 return; 780 781 rbo = container_of(bo, struct radeon_bo, tbo); 782 radeon_bo_check_tiling(rbo, 0, 1); 783 radeon_vm_bo_invalidate(rbo->rdev, rbo); 784 785 /* update statistics */ 786 if (!new_mem) 787 return; 788 789 radeon_update_memory_usage(rbo, bo->mem.mem_type, -1); 790 radeon_update_memory_usage(rbo, new_mem->mem_type, 1); 791 } 792 793 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo) 794 { 795 struct radeon_device *rdev; 796 struct radeon_bo *rbo; 797 unsigned long offset, size, lpfn; 798 int i, r; 799 800 if (!radeon_ttm_bo_is_radeon_bo(bo)) 801 return 0; 802 rbo = container_of(bo, struct radeon_bo, tbo); 803 radeon_bo_check_tiling(rbo, 0, 0); 804 rdev = rbo->rdev; 805 if (bo->mem.mem_type != TTM_PL_VRAM) 806 return 0; 807 808 size = bo->mem.num_pages << PAGE_SHIFT; 809 offset = bo->mem.start << PAGE_SHIFT; 810 if ((offset + size) <= rdev->mc.visible_vram_size) 811 return 0; 812 813 /* Can't move a pinned BO to visible VRAM */ 814 if (rbo->pin_count > 0) 815 return -EINVAL; 816 817 /* hurrah the memory is not visible ! */ 818 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM); 819 lpfn = rdev->mc.visible_vram_size >> PAGE_SHIFT; 820 for (i = 0; i < rbo->placement.num_placement; i++) { 821 /* Force into visible VRAM */ 822 if ((rbo->placements[i].flags & TTM_PL_FLAG_VRAM) && 823 (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn)) 824 rbo->placements[i].lpfn = lpfn; 825 } 826 r = ttm_bo_validate(bo, &rbo->placement, false, false); 827 if (unlikely(r == -ENOMEM)) { 828 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT); 829 return ttm_bo_validate(bo, &rbo->placement, false, false); 830 } else if (unlikely(r != 0)) { 831 return r; 832 } 833 834 offset = bo->mem.start << PAGE_SHIFT; 835 /* this should never happen */ 836 if ((offset + size) > rdev->mc.visible_vram_size) 837 return -EINVAL; 838 839 return 0; 840 } 841 842 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait) 843 { 844 int r; 845 846 r = ttm_bo_reserve(&bo->tbo, true, no_wait, NULL); 847 if (unlikely(r != 0)) 848 return r; 849 if (mem_type) 850 *mem_type = bo->tbo.mem.mem_type; 851 852 r = ttm_bo_wait(&bo->tbo, true, no_wait); 853 ttm_bo_unreserve(&bo->tbo); 854 return r; 855 } 856 857 /** 858 * radeon_bo_fence - add fence to buffer object 859 * 860 * @bo: buffer object in question 861 * @fence: fence to add 862 * @shared: true if fence should be added shared 863 * 864 */ 865 void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence, 866 bool shared) 867 { 868 struct reservation_object *resv = bo->tbo.resv; 869 870 if (shared) 871 reservation_object_add_shared_fence(resv, &fence->base); 872 else 873 reservation_object_add_excl_fence(resv, &fence->base); 874 } 875