1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2021 Intel Corporation 4 */ 5 6 #include <drm/ttm/ttm_bo_driver.h> 7 #include <drm/ttm/ttm_placement.h> 8 9 #include "i915_drv.h" 10 #include "intel_memory_region.h" 11 #include "intel_region_ttm.h" 12 13 #include "gem/i915_gem_object.h" 14 #include "gem/i915_gem_region.h" 15 #include "gem/i915_gem_ttm.h" 16 #include "gem/i915_gem_mman.h" 17 18 #include "gt/intel_migrate.h" 19 #include "gt/intel_engine_pm.h" 20 21 #define I915_PL_LMEM0 TTM_PL_PRIV 22 #define I915_PL_SYSTEM TTM_PL_SYSTEM 23 #define I915_PL_STOLEN TTM_PL_VRAM 24 #define I915_PL_GGTT TTM_PL_TT 25 26 #define I915_TTM_PRIO_PURGE 0 27 #define I915_TTM_PRIO_NO_PAGES 1 28 #define I915_TTM_PRIO_HAS_PAGES 2 29 30 /* 31 * Size of struct ttm_place vector in on-stack struct ttm_placement allocs 32 */ 33 #define I915_TTM_MAX_PLACEMENTS INTEL_REGION_UNKNOWN 34 35 /** 36 * struct i915_ttm_tt - TTM page vector with additional private information 37 * @ttm: The base TTM page vector. 38 * @dev: The struct device used for dma mapping and unmapping. 39 * @cached_st: The cached scatter-gather table. 40 * 41 * Note that DMA may be going on right up to the point where the page- 42 * vector is unpopulated in delayed destroy. Hence keep the 43 * scatter-gather table mapped and cached up to that point. This is 44 * different from the cached gem object io scatter-gather table which 45 * doesn't have an associated dma mapping. 46 */ 47 struct i915_ttm_tt { 48 struct ttm_tt ttm; 49 struct device *dev; 50 struct sg_table *cached_st; 51 }; 52 53 static const struct ttm_place sys_placement_flags = { 54 .fpfn = 0, 55 .lpfn = 0, 56 .mem_type = I915_PL_SYSTEM, 57 .flags = 0, 58 }; 59 60 static struct ttm_placement i915_sys_placement = { 61 .num_placement = 1, 62 .placement = &sys_placement_flags, 63 .num_busy_placement = 1, 64 .busy_placement = &sys_placement_flags, 65 }; 66 67 static int i915_ttm_err_to_gem(int err) 68 { 69 /* Fastpath */ 70 if (likely(!err)) 71 return 0; 72 73 switch (err) { 74 case -EBUSY: 75 /* 76 * TTM likes to convert -EDEADLK to -EBUSY, and wants us to 77 * restart the operation, since we don't record the contending 78 * lock. We use -EAGAIN to restart. 79 */ 80 return -EAGAIN; 81 case -ENOSPC: 82 /* 83 * Memory type / region is full, and we can't evict. 84 * Except possibly system, that returns -ENOMEM; 85 */ 86 return -ENXIO; 87 default: 88 break; 89 } 90 91 return err; 92 } 93 94 static bool gpu_binds_iomem(struct ttm_resource *mem) 95 { 96 return mem->mem_type != TTM_PL_SYSTEM; 97 } 98 99 static bool cpu_maps_iomem(struct ttm_resource *mem) 100 { 101 /* Once / if we support GGTT, this is also false for cached ttm_tts */ 102 return mem->mem_type != TTM_PL_SYSTEM; 103 } 104 105 static enum i915_cache_level 106 i915_ttm_cache_level(struct drm_i915_private *i915, struct ttm_resource *res, 107 struct ttm_tt *ttm) 108 { 109 return ((HAS_LLC(i915) || HAS_SNOOP(i915)) && !gpu_binds_iomem(res) && 110 ttm->caching == ttm_cached) ? I915_CACHE_LLC : 111 I915_CACHE_NONE; 112 } 113 114 static void i915_ttm_adjust_lru(struct drm_i915_gem_object *obj); 115 116 static enum ttm_caching 117 i915_ttm_select_tt_caching(const struct drm_i915_gem_object *obj) 118 { 119 /* 120 * Objects only allowed in system get cached cpu-mappings. 121 * Other objects get WC mapping for now. Even if in system. 122 */ 123 if (obj->mm.region->type == INTEL_MEMORY_SYSTEM && 124 obj->mm.n_placements <= 1) 125 return ttm_cached; 126 127 return ttm_write_combined; 128 } 129 130 static void 131 i915_ttm_place_from_region(const struct intel_memory_region *mr, 132 struct ttm_place *place, 133 unsigned int flags) 134 { 135 memset(place, 0, sizeof(*place)); 136 place->mem_type = intel_region_to_ttm_type(mr); 137 138 if (flags & I915_BO_ALLOC_CONTIGUOUS) 139 place->flags = TTM_PL_FLAG_CONTIGUOUS; 140 } 141 142 static void 143 i915_ttm_placement_from_obj(const struct drm_i915_gem_object *obj, 144 struct ttm_place *requested, 145 struct ttm_place *busy, 146 struct ttm_placement *placement) 147 { 148 unsigned int num_allowed = obj->mm.n_placements; 149 unsigned int flags = obj->flags; 150 unsigned int i; 151 152 placement->num_placement = 1; 153 i915_ttm_place_from_region(num_allowed ? obj->mm.placements[0] : 154 obj->mm.region, requested, flags); 155 156 /* Cache this on object? */ 157 placement->num_busy_placement = num_allowed; 158 for (i = 0; i < placement->num_busy_placement; ++i) 159 i915_ttm_place_from_region(obj->mm.placements[i], busy + i, flags); 160 161 if (num_allowed == 0) { 162 *busy = *requested; 163 placement->num_busy_placement = 1; 164 } 165 166 placement->placement = requested; 167 placement->busy_placement = busy; 168 } 169 170 static struct ttm_tt *i915_ttm_tt_create(struct ttm_buffer_object *bo, 171 uint32_t page_flags) 172 { 173 struct ttm_resource_manager *man = 174 ttm_manager_type(bo->bdev, bo->resource->mem_type); 175 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 176 struct i915_ttm_tt *i915_tt; 177 int ret; 178 179 i915_tt = kzalloc(sizeof(*i915_tt), GFP_KERNEL); 180 if (!i915_tt) 181 return NULL; 182 183 if (obj->flags & I915_BO_ALLOC_CPU_CLEAR && 184 man->use_tt) 185 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC; 186 187 ret = ttm_tt_init(&i915_tt->ttm, bo, page_flags, 188 i915_ttm_select_tt_caching(obj)); 189 if (ret) { 190 kfree(i915_tt); 191 return NULL; 192 } 193 194 i915_tt->dev = obj->base.dev->dev; 195 196 return &i915_tt->ttm; 197 } 198 199 static void i915_ttm_tt_unpopulate(struct ttm_device *bdev, struct ttm_tt *ttm) 200 { 201 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 202 203 if (i915_tt->cached_st) { 204 dma_unmap_sgtable(i915_tt->dev, i915_tt->cached_st, 205 DMA_BIDIRECTIONAL, 0); 206 sg_free_table(i915_tt->cached_st); 207 kfree(i915_tt->cached_st); 208 i915_tt->cached_st = NULL; 209 } 210 ttm_pool_free(&bdev->pool, ttm); 211 } 212 213 static void i915_ttm_tt_destroy(struct ttm_device *bdev, struct ttm_tt *ttm) 214 { 215 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 216 217 ttm_tt_destroy_common(bdev, ttm); 218 ttm_tt_fini(ttm); 219 kfree(i915_tt); 220 } 221 222 static bool i915_ttm_eviction_valuable(struct ttm_buffer_object *bo, 223 const struct ttm_place *place) 224 { 225 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 226 227 /* Will do for now. Our pinned objects are still on TTM's LRU lists */ 228 return i915_gem_object_evictable(obj); 229 } 230 231 static void i915_ttm_evict_flags(struct ttm_buffer_object *bo, 232 struct ttm_placement *placement) 233 { 234 *placement = i915_sys_placement; 235 } 236 237 static int i915_ttm_move_notify(struct ttm_buffer_object *bo) 238 { 239 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 240 int ret; 241 242 ret = i915_gem_object_unbind(obj, I915_GEM_OBJECT_UNBIND_ACTIVE); 243 if (ret) 244 return ret; 245 246 ret = __i915_gem_object_put_pages(obj); 247 if (ret) 248 return ret; 249 250 return 0; 251 } 252 253 static void i915_ttm_free_cached_io_st(struct drm_i915_gem_object *obj) 254 { 255 struct radix_tree_iter iter; 256 void __rcu **slot; 257 258 if (!obj->ttm.cached_io_st) 259 return; 260 261 rcu_read_lock(); 262 radix_tree_for_each_slot(slot, &obj->ttm.get_io_page.radix, &iter, 0) 263 radix_tree_delete(&obj->ttm.get_io_page.radix, iter.index); 264 rcu_read_unlock(); 265 266 sg_free_table(obj->ttm.cached_io_st); 267 kfree(obj->ttm.cached_io_st); 268 obj->ttm.cached_io_st = NULL; 269 } 270 271 static void 272 i915_ttm_adjust_domains_after_move(struct drm_i915_gem_object *obj) 273 { 274 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 275 276 if (cpu_maps_iomem(bo->resource) || bo->ttm->caching != ttm_cached) { 277 obj->write_domain = I915_GEM_DOMAIN_WC; 278 obj->read_domains = I915_GEM_DOMAIN_WC; 279 } else { 280 obj->write_domain = I915_GEM_DOMAIN_CPU; 281 obj->read_domains = I915_GEM_DOMAIN_CPU; 282 } 283 } 284 285 static void i915_ttm_adjust_gem_after_move(struct drm_i915_gem_object *obj) 286 { 287 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 288 unsigned int cache_level; 289 unsigned int i; 290 291 /* 292 * If object was moved to an allowable region, update the object 293 * region to consider it migrated. Note that if it's currently not 294 * in an allowable region, it's evicted and we don't update the 295 * object region. 296 */ 297 if (intel_region_to_ttm_type(obj->mm.region) != bo->resource->mem_type) { 298 for (i = 0; i < obj->mm.n_placements; ++i) { 299 struct intel_memory_region *mr = obj->mm.placements[i]; 300 301 if (intel_region_to_ttm_type(mr) == bo->resource->mem_type && 302 mr != obj->mm.region) { 303 i915_gem_object_release_memory_region(obj); 304 i915_gem_object_init_memory_region(obj, mr); 305 break; 306 } 307 } 308 } 309 310 obj->mem_flags &= ~(I915_BO_FLAG_STRUCT_PAGE | I915_BO_FLAG_IOMEM); 311 312 obj->mem_flags |= cpu_maps_iomem(bo->resource) ? I915_BO_FLAG_IOMEM : 313 I915_BO_FLAG_STRUCT_PAGE; 314 315 cache_level = i915_ttm_cache_level(to_i915(bo->base.dev), bo->resource, 316 bo->ttm); 317 i915_gem_object_set_cache_coherency(obj, cache_level); 318 } 319 320 static void i915_ttm_purge(struct drm_i915_gem_object *obj) 321 { 322 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 323 struct ttm_operation_ctx ctx = { 324 .interruptible = true, 325 .no_wait_gpu = false, 326 }; 327 struct ttm_placement place = {}; 328 int ret; 329 330 if (obj->mm.madv == __I915_MADV_PURGED) 331 return; 332 333 /* TTM's purge interface. Note that we might be reentering. */ 334 ret = ttm_bo_validate(bo, &place, &ctx); 335 if (!ret) { 336 obj->write_domain = 0; 337 obj->read_domains = 0; 338 i915_ttm_adjust_gem_after_move(obj); 339 i915_ttm_free_cached_io_st(obj); 340 obj->mm.madv = __I915_MADV_PURGED; 341 } 342 } 343 344 static void i915_ttm_swap_notify(struct ttm_buffer_object *bo) 345 { 346 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 347 int ret = i915_ttm_move_notify(bo); 348 349 GEM_WARN_ON(ret); 350 GEM_WARN_ON(obj->ttm.cached_io_st); 351 if (!ret && obj->mm.madv != I915_MADV_WILLNEED) 352 i915_ttm_purge(obj); 353 } 354 355 static void i915_ttm_delete_mem_notify(struct ttm_buffer_object *bo) 356 { 357 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 358 359 if (likely(obj)) 360 i915_ttm_free_cached_io_st(obj); 361 } 362 363 static struct intel_memory_region * 364 i915_ttm_region(struct ttm_device *bdev, int ttm_mem_type) 365 { 366 struct drm_i915_private *i915 = container_of(bdev, typeof(*i915), bdev); 367 368 /* There's some room for optimization here... */ 369 GEM_BUG_ON(ttm_mem_type != I915_PL_SYSTEM && 370 ttm_mem_type < I915_PL_LMEM0); 371 if (ttm_mem_type == I915_PL_SYSTEM) 372 return intel_memory_region_lookup(i915, INTEL_MEMORY_SYSTEM, 373 0); 374 375 return intel_memory_region_lookup(i915, INTEL_MEMORY_LOCAL, 376 ttm_mem_type - I915_PL_LMEM0); 377 } 378 379 static struct sg_table *i915_ttm_tt_get_st(struct ttm_tt *ttm) 380 { 381 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 382 struct sg_table *st; 383 int ret; 384 385 if (i915_tt->cached_st) 386 return i915_tt->cached_st; 387 388 st = kzalloc(sizeof(*st), GFP_KERNEL); 389 if (!st) 390 return ERR_PTR(-ENOMEM); 391 392 ret = sg_alloc_table_from_pages_segment(st, 393 ttm->pages, ttm->num_pages, 394 0, (unsigned long)ttm->num_pages << PAGE_SHIFT, 395 i915_sg_segment_size(), GFP_KERNEL); 396 if (ret) { 397 kfree(st); 398 return ERR_PTR(ret); 399 } 400 401 ret = dma_map_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 0); 402 if (ret) { 403 sg_free_table(st); 404 kfree(st); 405 return ERR_PTR(ret); 406 } 407 408 i915_tt->cached_st = st; 409 return st; 410 } 411 412 static struct sg_table * 413 i915_ttm_resource_get_st(struct drm_i915_gem_object *obj, 414 struct ttm_resource *res) 415 { 416 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 417 418 if (!gpu_binds_iomem(res)) 419 return i915_ttm_tt_get_st(bo->ttm); 420 421 /* 422 * If CPU mapping differs, we need to add the ttm_tt pages to 423 * the resulting st. Might make sense for GGTT. 424 */ 425 GEM_WARN_ON(!cpu_maps_iomem(res)); 426 return intel_region_ttm_resource_to_st(obj->mm.region, res); 427 } 428 429 static int i915_ttm_accel_move(struct ttm_buffer_object *bo, 430 struct ttm_resource *dst_mem, 431 struct sg_table *dst_st) 432 { 433 struct drm_i915_private *i915 = container_of(bo->bdev, typeof(*i915), 434 bdev); 435 struct ttm_resource_manager *src_man = 436 ttm_manager_type(bo->bdev, bo->resource->mem_type); 437 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 438 struct sg_table *src_st; 439 struct i915_request *rq; 440 struct ttm_tt *ttm = bo->ttm; 441 enum i915_cache_level src_level, dst_level; 442 int ret; 443 444 if (!i915->gt.migrate.context) 445 return -EINVAL; 446 447 dst_level = i915_ttm_cache_level(i915, dst_mem, ttm); 448 if (!ttm || !ttm_tt_is_populated(ttm)) { 449 if (bo->type == ttm_bo_type_kernel) 450 return -EINVAL; 451 452 if (ttm && !(ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)) 453 return 0; 454 455 intel_engine_pm_get(i915->gt.migrate.context->engine); 456 ret = intel_context_migrate_clear(i915->gt.migrate.context, NULL, 457 dst_st->sgl, dst_level, 458 gpu_binds_iomem(dst_mem), 459 0, &rq); 460 461 if (!ret && rq) { 462 i915_request_wait(rq, 0, MAX_SCHEDULE_TIMEOUT); 463 i915_request_put(rq); 464 } 465 intel_engine_pm_put(i915->gt.migrate.context->engine); 466 } else { 467 src_st = src_man->use_tt ? i915_ttm_tt_get_st(ttm) : 468 obj->ttm.cached_io_st; 469 470 src_level = i915_ttm_cache_level(i915, bo->resource, ttm); 471 intel_engine_pm_get(i915->gt.migrate.context->engine); 472 ret = intel_context_migrate_copy(i915->gt.migrate.context, 473 NULL, src_st->sgl, src_level, 474 gpu_binds_iomem(bo->resource), 475 dst_st->sgl, dst_level, 476 gpu_binds_iomem(dst_mem), 477 &rq); 478 if (!ret && rq) { 479 i915_request_wait(rq, 0, MAX_SCHEDULE_TIMEOUT); 480 i915_request_put(rq); 481 } 482 intel_engine_pm_put(i915->gt.migrate.context->engine); 483 } 484 485 return ret; 486 } 487 488 static int i915_ttm_move(struct ttm_buffer_object *bo, bool evict, 489 struct ttm_operation_ctx *ctx, 490 struct ttm_resource *dst_mem, 491 struct ttm_place *hop) 492 { 493 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 494 struct ttm_resource_manager *dst_man = 495 ttm_manager_type(bo->bdev, dst_mem->mem_type); 496 struct intel_memory_region *dst_reg, *src_reg; 497 union { 498 struct ttm_kmap_iter_tt tt; 499 struct ttm_kmap_iter_iomap io; 500 } _dst_iter, _src_iter; 501 struct ttm_kmap_iter *dst_iter, *src_iter; 502 struct sg_table *dst_st; 503 int ret; 504 505 dst_reg = i915_ttm_region(bo->bdev, dst_mem->mem_type); 506 src_reg = i915_ttm_region(bo->bdev, bo->resource->mem_type); 507 GEM_BUG_ON(!dst_reg || !src_reg); 508 509 /* Sync for now. We could do the actual copy async. */ 510 ret = ttm_bo_wait_ctx(bo, ctx); 511 if (ret) 512 return ret; 513 514 ret = i915_ttm_move_notify(bo); 515 if (ret) 516 return ret; 517 518 if (obj->mm.madv != I915_MADV_WILLNEED) { 519 i915_ttm_purge(obj); 520 ttm_resource_free(bo, &dst_mem); 521 return 0; 522 } 523 524 /* Populate ttm with pages if needed. Typically system memory. */ 525 if (bo->ttm && (dst_man->use_tt || 526 (bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED))) { 527 ret = ttm_tt_populate(bo->bdev, bo->ttm, ctx); 528 if (ret) 529 return ret; 530 } 531 532 dst_st = i915_ttm_resource_get_st(obj, dst_mem); 533 if (IS_ERR(dst_st)) 534 return PTR_ERR(dst_st); 535 536 ret = i915_ttm_accel_move(bo, dst_mem, dst_st); 537 if (ret) { 538 /* If we start mapping GGTT, we can no longer use man::use_tt here. */ 539 dst_iter = !cpu_maps_iomem(dst_mem) ? 540 ttm_kmap_iter_tt_init(&_dst_iter.tt, bo->ttm) : 541 ttm_kmap_iter_iomap_init(&_dst_iter.io, &dst_reg->iomap, 542 dst_st, dst_reg->region.start); 543 544 src_iter = !cpu_maps_iomem(bo->resource) ? 545 ttm_kmap_iter_tt_init(&_src_iter.tt, bo->ttm) : 546 ttm_kmap_iter_iomap_init(&_src_iter.io, &src_reg->iomap, 547 obj->ttm.cached_io_st, 548 src_reg->region.start); 549 550 ttm_move_memcpy(bo, dst_mem->num_pages, dst_iter, src_iter); 551 } 552 /* Below dst_mem becomes bo->resource. */ 553 ttm_bo_move_sync_cleanup(bo, dst_mem); 554 i915_ttm_adjust_domains_after_move(obj); 555 i915_ttm_free_cached_io_st(obj); 556 557 if (gpu_binds_iomem(dst_mem) || cpu_maps_iomem(dst_mem)) { 558 obj->ttm.cached_io_st = dst_st; 559 obj->ttm.get_io_page.sg_pos = dst_st->sgl; 560 obj->ttm.get_io_page.sg_idx = 0; 561 } 562 563 i915_ttm_adjust_gem_after_move(obj); 564 return 0; 565 } 566 567 static int i915_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *mem) 568 { 569 if (!cpu_maps_iomem(mem)) 570 return 0; 571 572 mem->bus.caching = ttm_write_combined; 573 mem->bus.is_iomem = true; 574 575 return 0; 576 } 577 578 static unsigned long i915_ttm_io_mem_pfn(struct ttm_buffer_object *bo, 579 unsigned long page_offset) 580 { 581 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 582 unsigned long base = obj->mm.region->iomap.base - obj->mm.region->region.start; 583 struct scatterlist *sg; 584 unsigned int ofs; 585 586 GEM_WARN_ON(bo->ttm); 587 588 sg = __i915_gem_object_get_sg(obj, &obj->ttm.get_io_page, page_offset, &ofs, true); 589 590 return ((base + sg_dma_address(sg)) >> PAGE_SHIFT) + ofs; 591 } 592 593 static struct ttm_device_funcs i915_ttm_bo_driver = { 594 .ttm_tt_create = i915_ttm_tt_create, 595 .ttm_tt_unpopulate = i915_ttm_tt_unpopulate, 596 .ttm_tt_destroy = i915_ttm_tt_destroy, 597 .eviction_valuable = i915_ttm_eviction_valuable, 598 .evict_flags = i915_ttm_evict_flags, 599 .move = i915_ttm_move, 600 .swap_notify = i915_ttm_swap_notify, 601 .delete_mem_notify = i915_ttm_delete_mem_notify, 602 .io_mem_reserve = i915_ttm_io_mem_reserve, 603 .io_mem_pfn = i915_ttm_io_mem_pfn, 604 }; 605 606 /** 607 * i915_ttm_driver - Return a pointer to the TTM device funcs 608 * 609 * Return: Pointer to statically allocated TTM device funcs. 610 */ 611 struct ttm_device_funcs *i915_ttm_driver(void) 612 { 613 return &i915_ttm_bo_driver; 614 } 615 616 static int __i915_ttm_get_pages(struct drm_i915_gem_object *obj, 617 struct ttm_placement *placement) 618 { 619 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 620 struct ttm_operation_ctx ctx = { 621 .interruptible = true, 622 .no_wait_gpu = false, 623 }; 624 struct sg_table *st; 625 int real_num_busy; 626 int ret; 627 628 /* First try only the requested placement. No eviction. */ 629 real_num_busy = fetch_and_zero(&placement->num_busy_placement); 630 ret = ttm_bo_validate(bo, placement, &ctx); 631 if (ret) { 632 ret = i915_ttm_err_to_gem(ret); 633 /* 634 * Anything that wants to restart the operation gets to 635 * do that. 636 */ 637 if (ret == -EDEADLK || ret == -EINTR || ret == -ERESTARTSYS || 638 ret == -EAGAIN) 639 return ret; 640 641 /* 642 * If the initial attempt fails, allow all accepted placements, 643 * evicting if necessary. 644 */ 645 placement->num_busy_placement = real_num_busy; 646 ret = ttm_bo_validate(bo, placement, &ctx); 647 if (ret) 648 return i915_ttm_err_to_gem(ret); 649 } 650 651 i915_ttm_adjust_lru(obj); 652 if (bo->ttm && !ttm_tt_is_populated(bo->ttm)) { 653 ret = ttm_tt_populate(bo->bdev, bo->ttm, &ctx); 654 if (ret) 655 return ret; 656 657 i915_ttm_adjust_domains_after_move(obj); 658 i915_ttm_adjust_gem_after_move(obj); 659 } 660 661 if (!i915_gem_object_has_pages(obj)) { 662 /* Object either has a page vector or is an iomem object */ 663 st = bo->ttm ? i915_ttm_tt_get_st(bo->ttm) : obj->ttm.cached_io_st; 664 if (IS_ERR(st)) 665 return PTR_ERR(st); 666 667 __i915_gem_object_set_pages(obj, st, i915_sg_dma_sizes(st->sgl)); 668 } 669 670 return ret; 671 } 672 673 static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) 674 { 675 struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS]; 676 struct ttm_placement placement; 677 678 GEM_BUG_ON(obj->mm.n_placements > I915_TTM_MAX_PLACEMENTS); 679 680 /* Move to the requested placement. */ 681 i915_ttm_placement_from_obj(obj, &requested, busy, &placement); 682 683 return __i915_ttm_get_pages(obj, &placement); 684 } 685 686 /** 687 * DOC: Migration vs eviction 688 * 689 * GEM migration may not be the same as TTM migration / eviction. If 690 * the TTM core decides to evict an object it may be evicted to a 691 * TTM memory type that is not in the object's allowable GEM regions, or 692 * in fact theoretically to a TTM memory type that doesn't correspond to 693 * a GEM memory region. In that case the object's GEM region is not 694 * updated, and the data is migrated back to the GEM region at 695 * get_pages time. TTM may however set up CPU ptes to the object even 696 * when it is evicted. 697 * Gem forced migration using the i915_ttm_migrate() op, is allowed even 698 * to regions that are not in the object's list of allowable placements. 699 */ 700 static int i915_ttm_migrate(struct drm_i915_gem_object *obj, 701 struct intel_memory_region *mr) 702 { 703 struct ttm_place requested; 704 struct ttm_placement placement; 705 int ret; 706 707 i915_ttm_place_from_region(mr, &requested, obj->flags); 708 placement.num_placement = 1; 709 placement.num_busy_placement = 1; 710 placement.placement = &requested; 711 placement.busy_placement = &requested; 712 713 ret = __i915_ttm_get_pages(obj, &placement); 714 if (ret) 715 return ret; 716 717 /* 718 * Reinitialize the region bindings. This is primarily 719 * required for objects where the new region is not in 720 * its allowable placements. 721 */ 722 if (obj->mm.region != mr) { 723 i915_gem_object_release_memory_region(obj); 724 i915_gem_object_init_memory_region(obj, mr); 725 } 726 727 return 0; 728 } 729 730 static void i915_ttm_put_pages(struct drm_i915_gem_object *obj, 731 struct sg_table *st) 732 { 733 /* 734 * We're currently not called from a shrinker, so put_pages() 735 * typically means the object is about to destroyed, or called 736 * from move_notify(). So just avoid doing much for now. 737 * If the object is not destroyed next, The TTM eviction logic 738 * and shrinkers will move it out if needed. 739 */ 740 741 i915_ttm_adjust_lru(obj); 742 } 743 744 static void i915_ttm_adjust_lru(struct drm_i915_gem_object *obj) 745 { 746 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 747 748 /* 749 * Don't manipulate the TTM LRUs while in TTM bo destruction. 750 * We're called through i915_ttm_delete_mem_notify(). 751 */ 752 if (!kref_read(&bo->kref)) 753 return; 754 755 /* 756 * Put on the correct LRU list depending on the MADV status 757 */ 758 spin_lock(&bo->bdev->lru_lock); 759 if (obj->mm.madv != I915_MADV_WILLNEED) { 760 bo->priority = I915_TTM_PRIO_PURGE; 761 } else if (!i915_gem_object_has_pages(obj)) { 762 if (bo->priority < I915_TTM_PRIO_HAS_PAGES) 763 bo->priority = I915_TTM_PRIO_HAS_PAGES; 764 } else { 765 if (bo->priority > I915_TTM_PRIO_NO_PAGES) 766 bo->priority = I915_TTM_PRIO_NO_PAGES; 767 } 768 769 ttm_bo_move_to_lru_tail(bo, bo->resource, NULL); 770 spin_unlock(&bo->bdev->lru_lock); 771 } 772 773 /* 774 * TTM-backed gem object destruction requires some clarification. 775 * Basically we have two possibilities here. We can either rely on the 776 * i915 delayed destruction and put the TTM object when the object 777 * is idle. This would be detected by TTM which would bypass the 778 * TTM delayed destroy handling. The other approach is to put the TTM 779 * object early and rely on the TTM destroyed handling, and then free 780 * the leftover parts of the GEM object once TTM's destroyed list handling is 781 * complete. For now, we rely on the latter for two reasons: 782 * a) TTM can evict an object even when it's on the delayed destroy list, 783 * which in theory allows for complete eviction. 784 * b) There is work going on in TTM to allow freeing an object even when 785 * it's not idle, and using the TTM destroyed list handling could help us 786 * benefit from that. 787 */ 788 static void i915_ttm_delayed_free(struct drm_i915_gem_object *obj) 789 { 790 if (obj->ttm.created) { 791 ttm_bo_put(i915_gem_to_ttm(obj)); 792 } else { 793 __i915_gem_free_object(obj); 794 call_rcu(&obj->rcu, __i915_gem_free_object_rcu); 795 } 796 } 797 798 static vm_fault_t vm_fault_ttm(struct vm_fault *vmf) 799 { 800 struct vm_area_struct *area = vmf->vma; 801 struct drm_i915_gem_object *obj = 802 i915_ttm_to_gem(area->vm_private_data); 803 804 /* Sanity check that we allow writing into this object */ 805 if (unlikely(i915_gem_object_is_readonly(obj) && 806 area->vm_flags & VM_WRITE)) 807 return VM_FAULT_SIGBUS; 808 809 return ttm_bo_vm_fault(vmf); 810 } 811 812 static int 813 vm_access_ttm(struct vm_area_struct *area, unsigned long addr, 814 void *buf, int len, int write) 815 { 816 struct drm_i915_gem_object *obj = 817 i915_ttm_to_gem(area->vm_private_data); 818 819 if (i915_gem_object_is_readonly(obj) && write) 820 return -EACCES; 821 822 return ttm_bo_vm_access(area, addr, buf, len, write); 823 } 824 825 static void ttm_vm_open(struct vm_area_struct *vma) 826 { 827 struct drm_i915_gem_object *obj = 828 i915_ttm_to_gem(vma->vm_private_data); 829 830 GEM_BUG_ON(!obj); 831 i915_gem_object_get(obj); 832 } 833 834 static void ttm_vm_close(struct vm_area_struct *vma) 835 { 836 struct drm_i915_gem_object *obj = 837 i915_ttm_to_gem(vma->vm_private_data); 838 839 GEM_BUG_ON(!obj); 840 i915_gem_object_put(obj); 841 } 842 843 static const struct vm_operations_struct vm_ops_ttm = { 844 .fault = vm_fault_ttm, 845 .access = vm_access_ttm, 846 .open = ttm_vm_open, 847 .close = ttm_vm_close, 848 }; 849 850 static u64 i915_ttm_mmap_offset(struct drm_i915_gem_object *obj) 851 { 852 /* The ttm_bo must be allocated with I915_BO_ALLOC_USER */ 853 GEM_BUG_ON(!drm_mm_node_allocated(&obj->base.vma_node.vm_node)); 854 855 return drm_vma_node_offset_addr(&obj->base.vma_node); 856 } 857 858 static const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = { 859 .name = "i915_gem_object_ttm", 860 861 .get_pages = i915_ttm_get_pages, 862 .put_pages = i915_ttm_put_pages, 863 .truncate = i915_ttm_purge, 864 .adjust_lru = i915_ttm_adjust_lru, 865 .delayed_free = i915_ttm_delayed_free, 866 .migrate = i915_ttm_migrate, 867 .mmap_offset = i915_ttm_mmap_offset, 868 .mmap_ops = &vm_ops_ttm, 869 }; 870 871 void i915_ttm_bo_destroy(struct ttm_buffer_object *bo) 872 { 873 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 874 875 /* This releases all gem object bindings to the backend. */ 876 __i915_gem_free_object(obj); 877 878 i915_gem_object_release_memory_region(obj); 879 mutex_destroy(&obj->ttm.get_io_page.lock); 880 881 if (obj->ttm.created) 882 call_rcu(&obj->rcu, __i915_gem_free_object_rcu); 883 } 884 885 /** 886 * __i915_gem_ttm_object_init - Initialize a ttm-backed i915 gem object 887 * @mem: The initial memory region for the object. 888 * @obj: The gem object. 889 * @size: Object size in bytes. 890 * @flags: gem object flags. 891 * 892 * Return: 0 on success, negative error code on failure. 893 */ 894 int __i915_gem_ttm_object_init(struct intel_memory_region *mem, 895 struct drm_i915_gem_object *obj, 896 resource_size_t size, 897 resource_size_t page_size, 898 unsigned int flags) 899 { 900 static struct lock_class_key lock_class; 901 struct drm_i915_private *i915 = mem->i915; 902 struct ttm_operation_ctx ctx = { 903 .interruptible = true, 904 .no_wait_gpu = false, 905 }; 906 enum ttm_bo_type bo_type; 907 int ret; 908 909 drm_gem_private_object_init(&i915->drm, &obj->base, size); 910 i915_gem_object_init(obj, &i915_gem_ttm_obj_ops, &lock_class, flags); 911 i915_gem_object_init_memory_region(obj, mem); 912 i915_gem_object_make_unshrinkable(obj); 913 INIT_RADIX_TREE(&obj->ttm.get_io_page.radix, GFP_KERNEL | __GFP_NOWARN); 914 mutex_init(&obj->ttm.get_io_page.lock); 915 bo_type = (obj->flags & I915_BO_ALLOC_USER) ? ttm_bo_type_device : 916 ttm_bo_type_kernel; 917 918 obj->base.vma_node.driver_private = i915_gem_to_ttm(obj); 919 920 /* Forcing the page size is kernel internal only */ 921 GEM_BUG_ON(page_size && obj->mm.n_placements); 922 923 /* 924 * If this function fails, it will call the destructor, but 925 * our caller still owns the object. So no freeing in the 926 * destructor until obj->ttm.created is true. 927 * Similarly, in delayed_destroy, we can't call ttm_bo_put() 928 * until successful initialization. 929 */ 930 ret = ttm_bo_init_reserved(&i915->bdev, i915_gem_to_ttm(obj), size, 931 bo_type, &i915_sys_placement, 932 page_size >> PAGE_SHIFT, 933 &ctx, NULL, NULL, i915_ttm_bo_destroy); 934 if (ret) 935 return i915_ttm_err_to_gem(ret); 936 937 obj->ttm.created = true; 938 i915_ttm_adjust_domains_after_move(obj); 939 i915_ttm_adjust_gem_after_move(obj); 940 i915_gem_object_unlock(obj); 941 942 return 0; 943 } 944 945 static const struct intel_memory_region_ops ttm_system_region_ops = { 946 .init_object = __i915_gem_ttm_object_init, 947 }; 948 949 struct intel_memory_region * 950 i915_gem_ttm_system_setup(struct drm_i915_private *i915, 951 u16 type, u16 instance) 952 { 953 struct intel_memory_region *mr; 954 955 mr = intel_memory_region_create(i915, 0, 956 totalram_pages() << PAGE_SHIFT, 957 PAGE_SIZE, 0, 958 type, instance, 959 &ttm_system_region_ops); 960 if (IS_ERR(mr)) 961 return mr; 962 963 intel_memory_region_set_name(mr, "system-ttm"); 964 return mr; 965 } 966