1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2021 Intel Corporation 4 */ 5 6 #include <linux/shmem_fs.h> 7 8 #include <drm/ttm/ttm_placement.h> 9 #include <drm/ttm/ttm_tt.h> 10 #include <drm/drm_buddy.h> 11 12 #include "i915_drv.h" 13 #include "i915_ttm_buddy_manager.h" 14 #include "intel_memory_region.h" 15 #include "intel_region_ttm.h" 16 17 #include "gem/i915_gem_mman.h" 18 #include "gem/i915_gem_object.h" 19 #include "gem/i915_gem_region.h" 20 #include "gem/i915_gem_ttm.h" 21 #include "gem/i915_gem_ttm_move.h" 22 #include "gem/i915_gem_ttm_pm.h" 23 #include "gt/intel_gpu_commands.h" 24 25 #define I915_TTM_PRIO_PURGE 0 26 #define I915_TTM_PRIO_NO_PAGES 1 27 #define I915_TTM_PRIO_HAS_PAGES 2 28 #define I915_TTM_PRIO_NEEDS_CPU_ACCESS 3 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_rsgt: The cached scatter-gather table. 40 * @is_shmem: Set if using shmem. 41 * @filp: The shmem file, if using shmem backend. 42 * 43 * Note that DMA may be going on right up to the point where the page- 44 * vector is unpopulated in delayed destroy. Hence keep the 45 * scatter-gather table mapped and cached up to that point. This is 46 * different from the cached gem object io scatter-gather table which 47 * doesn't have an associated dma mapping. 48 */ 49 struct i915_ttm_tt { 50 struct ttm_tt ttm; 51 struct device *dev; 52 struct i915_refct_sgt cached_rsgt; 53 54 bool is_shmem; 55 struct file *filp; 56 }; 57 58 static const struct ttm_place sys_placement_flags = { 59 .fpfn = 0, 60 .lpfn = 0, 61 .mem_type = I915_PL_SYSTEM, 62 .flags = 0, 63 }; 64 65 static struct ttm_placement i915_sys_placement = { 66 .num_placement = 1, 67 .placement = &sys_placement_flags, 68 .num_busy_placement = 1, 69 .busy_placement = &sys_placement_flags, 70 }; 71 72 /** 73 * i915_ttm_sys_placement - Return the struct ttm_placement to be 74 * used for an object in system memory. 75 * 76 * Rather than making the struct extern, use this 77 * function. 78 * 79 * Return: A pointer to a static variable for sys placement. 80 */ 81 struct ttm_placement *i915_ttm_sys_placement(void) 82 { 83 return &i915_sys_placement; 84 } 85 86 static int i915_ttm_err_to_gem(int err) 87 { 88 /* Fastpath */ 89 if (likely(!err)) 90 return 0; 91 92 switch (err) { 93 case -EBUSY: 94 /* 95 * TTM likes to convert -EDEADLK to -EBUSY, and wants us to 96 * restart the operation, since we don't record the contending 97 * lock. We use -EAGAIN to restart. 98 */ 99 return -EAGAIN; 100 case -ENOSPC: 101 /* 102 * Memory type / region is full, and we can't evict. 103 * Except possibly system, that returns -ENOMEM; 104 */ 105 return -ENXIO; 106 default: 107 break; 108 } 109 110 return err; 111 } 112 113 static enum ttm_caching 114 i915_ttm_select_tt_caching(const struct drm_i915_gem_object *obj) 115 { 116 /* 117 * Objects only allowed in system get cached cpu-mappings, or when 118 * evicting lmem-only buffers to system for swapping. Other objects get 119 * WC mapping for now. Even if in system. 120 */ 121 if (obj->mm.n_placements <= 1) 122 return ttm_cached; 123 124 return ttm_write_combined; 125 } 126 127 static void 128 i915_ttm_place_from_region(const struct intel_memory_region *mr, 129 struct ttm_place *place, 130 resource_size_t offset, 131 resource_size_t size, 132 unsigned int flags) 133 { 134 memset(place, 0, sizeof(*place)); 135 place->mem_type = intel_region_to_ttm_type(mr); 136 137 if (mr->type == INTEL_MEMORY_SYSTEM) 138 return; 139 140 if (flags & I915_BO_ALLOC_CONTIGUOUS) 141 place->flags |= TTM_PL_FLAG_CONTIGUOUS; 142 if (offset != I915_BO_INVALID_OFFSET) { 143 WARN_ON(overflows_type(offset >> PAGE_SHIFT, place->fpfn)); 144 place->fpfn = offset >> PAGE_SHIFT; 145 WARN_ON(overflows_type(place->fpfn + (size >> PAGE_SHIFT), place->lpfn)); 146 place->lpfn = place->fpfn + (size >> PAGE_SHIFT); 147 } else if (mr->io_size && mr->io_size < mr->total) { 148 if (flags & I915_BO_ALLOC_GPU_ONLY) { 149 place->flags |= TTM_PL_FLAG_TOPDOWN; 150 } else { 151 place->fpfn = 0; 152 WARN_ON(overflows_type(mr->io_size >> PAGE_SHIFT, place->lpfn)); 153 place->lpfn = mr->io_size >> PAGE_SHIFT; 154 } 155 } 156 } 157 158 static void 159 i915_ttm_placement_from_obj(const struct drm_i915_gem_object *obj, 160 struct ttm_place *requested, 161 struct ttm_place *busy, 162 struct ttm_placement *placement) 163 { 164 unsigned int num_allowed = obj->mm.n_placements; 165 unsigned int flags = obj->flags; 166 unsigned int i; 167 168 placement->num_placement = 1; 169 i915_ttm_place_from_region(num_allowed ? obj->mm.placements[0] : 170 obj->mm.region, requested, obj->bo_offset, 171 obj->base.size, flags); 172 173 /* Cache this on object? */ 174 placement->num_busy_placement = num_allowed; 175 for (i = 0; i < placement->num_busy_placement; ++i) 176 i915_ttm_place_from_region(obj->mm.placements[i], busy + i, 177 obj->bo_offset, obj->base.size, flags); 178 179 if (num_allowed == 0) { 180 *busy = *requested; 181 placement->num_busy_placement = 1; 182 } 183 184 placement->placement = requested; 185 placement->busy_placement = busy; 186 } 187 188 static int i915_ttm_tt_shmem_populate(struct ttm_device *bdev, 189 struct ttm_tt *ttm, 190 struct ttm_operation_ctx *ctx) 191 { 192 STUB(); 193 return -ENOSYS; 194 #ifdef notyet 195 struct drm_i915_private *i915 = container_of(bdev, typeof(*i915), bdev); 196 struct intel_memory_region *mr = i915->mm.regions[INTEL_MEMORY_SYSTEM]; 197 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 198 const unsigned int max_segment = i915_sg_segment_size(i915->drm.dev); 199 const size_t size = (size_t)ttm->num_pages << PAGE_SHIFT; 200 struct file *filp = i915_tt->filp; 201 struct sgt_iter sgt_iter; 202 struct sg_table *st; 203 struct vm_page *page; 204 unsigned long i; 205 int err; 206 207 if (!filp) { 208 struct address_space *mapping; 209 gfp_t mask; 210 211 filp = shmem_file_setup("i915-shmem-tt", size, VM_NORESERVE); 212 if (IS_ERR(filp)) 213 return PTR_ERR(filp); 214 215 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE; 216 217 mapping = filp->f_mapping; 218 mapping_set_gfp_mask(mapping, mask); 219 GEM_BUG_ON(!(mapping_gfp_mask(mapping) & __GFP_RECLAIM)); 220 221 i915_tt->filp = filp; 222 } 223 224 st = &i915_tt->cached_rsgt.table; 225 err = shmem_sg_alloc_table(i915, st, size, mr, filp->f_mapping, 226 max_segment); 227 if (err) 228 return err; 229 230 err = dma_map_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 231 DMA_ATTR_SKIP_CPU_SYNC); 232 if (err) 233 goto err_free_st; 234 235 i = 0; 236 for_each_sgt_page(page, sgt_iter, st) 237 ttm->pages[i++] = page; 238 239 if (ttm->page_flags & TTM_TT_FLAG_SWAPPED) 240 ttm->page_flags &= ~TTM_TT_FLAG_SWAPPED; 241 242 return 0; 243 244 err_free_st: 245 shmem_sg_free_table(st, filp->f_mapping, false, false); 246 247 return err; 248 #endif 249 } 250 251 static void i915_ttm_tt_shmem_unpopulate(struct ttm_tt *ttm) 252 { 253 STUB(); 254 #ifdef notyet 255 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 256 bool backup = ttm->page_flags & TTM_TT_FLAG_SWAPPED; 257 struct sg_table *st = &i915_tt->cached_rsgt.table; 258 259 shmem_sg_free_table(st, file_inode(i915_tt->filp)->i_mapping, 260 backup, backup); 261 #endif 262 } 263 264 static void i915_ttm_tt_release(struct kref *ref) 265 { 266 struct i915_ttm_tt *i915_tt = 267 container_of(ref, typeof(*i915_tt), cached_rsgt.kref); 268 struct sg_table *st = &i915_tt->cached_rsgt.table; 269 270 GEM_WARN_ON(st->sgl); 271 272 kfree(i915_tt); 273 } 274 275 static const struct i915_refct_sgt_ops tt_rsgt_ops = { 276 .release = i915_ttm_tt_release 277 }; 278 279 static struct ttm_tt *i915_ttm_tt_create(struct ttm_buffer_object *bo, 280 uint32_t page_flags) 281 { 282 struct drm_i915_private *i915 = container_of(bo->bdev, typeof(*i915), 283 bdev); 284 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 285 unsigned long ccs_pages = 0; 286 enum ttm_caching caching; 287 struct i915_ttm_tt *i915_tt; 288 int ret; 289 290 if (i915_ttm_is_ghost_object(bo)) 291 return NULL; 292 293 i915_tt = kzalloc(sizeof(*i915_tt), GFP_KERNEL); 294 if (!i915_tt) 295 return NULL; 296 297 if (obj->flags & I915_BO_ALLOC_CPU_CLEAR && (!bo->resource || 298 ttm_manager_type(bo->bdev, bo->resource->mem_type)->use_tt)) 299 page_flags |= TTM_TT_FLAG_ZERO_ALLOC; 300 301 caching = i915_ttm_select_tt_caching(obj); 302 if (i915_gem_object_is_shrinkable(obj) && caching == ttm_cached) { 303 page_flags |= TTM_TT_FLAG_EXTERNAL | 304 TTM_TT_FLAG_EXTERNAL_MAPPABLE; 305 i915_tt->is_shmem = true; 306 } 307 308 if (i915_gem_object_needs_ccs_pages(obj)) 309 ccs_pages = DIV_ROUND_UP(DIV_ROUND_UP(bo->base.size, 310 NUM_BYTES_PER_CCS_BYTE), 311 PAGE_SIZE); 312 313 ret = ttm_tt_init(&i915_tt->ttm, bo, page_flags, caching, ccs_pages); 314 if (ret) 315 goto err_free; 316 317 __i915_refct_sgt_init(&i915_tt->cached_rsgt, bo->base.size, 318 &tt_rsgt_ops); 319 320 i915_tt->dev = obj->base.dev->dev; 321 322 return &i915_tt->ttm; 323 324 err_free: 325 kfree(i915_tt); 326 return NULL; 327 } 328 329 static int i915_ttm_tt_populate(struct ttm_device *bdev, 330 struct ttm_tt *ttm, 331 struct ttm_operation_ctx *ctx) 332 { 333 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 334 335 if (i915_tt->is_shmem) 336 return i915_ttm_tt_shmem_populate(bdev, ttm, ctx); 337 338 return ttm_pool_alloc(&bdev->pool, ttm, ctx); 339 } 340 341 static void i915_ttm_tt_unpopulate(struct ttm_device *bdev, struct ttm_tt *ttm) 342 { 343 STUB(); 344 #ifdef notyet 345 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 346 struct sg_table *st = &i915_tt->cached_rsgt.table; 347 348 if (st->sgl) 349 dma_unmap_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 0); 350 351 if (i915_tt->is_shmem) { 352 i915_ttm_tt_shmem_unpopulate(ttm); 353 } else { 354 sg_free_table(st); 355 ttm_pool_free(&bdev->pool, ttm); 356 } 357 #endif 358 } 359 360 static void i915_ttm_tt_destroy(struct ttm_device *bdev, struct ttm_tt *ttm) 361 { 362 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 363 364 if (i915_tt->filp) 365 fput(i915_tt->filp); 366 367 ttm_tt_fini(ttm); 368 i915_refct_sgt_put(&i915_tt->cached_rsgt); 369 } 370 371 static bool i915_ttm_eviction_valuable(struct ttm_buffer_object *bo, 372 const struct ttm_place *place) 373 { 374 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 375 376 if (i915_ttm_is_ghost_object(bo)) 377 return false; 378 379 /* 380 * EXTERNAL objects should never be swapped out by TTM, instead we need 381 * to handle that ourselves. TTM will already skip such objects for us, 382 * but we would like to avoid grabbing locks for no good reason. 383 */ 384 if (bo->ttm && bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL) 385 return false; 386 387 /* Will do for now. Our pinned objects are still on TTM's LRU lists */ 388 if (!i915_gem_object_evictable(obj)) 389 return false; 390 391 return ttm_bo_eviction_valuable(bo, place); 392 } 393 394 static void i915_ttm_evict_flags(struct ttm_buffer_object *bo, 395 struct ttm_placement *placement) 396 { 397 *placement = i915_sys_placement; 398 } 399 400 /** 401 * i915_ttm_free_cached_io_rsgt - Free object cached LMEM information 402 * @obj: The GEM object 403 * This function frees any LMEM-related information that is cached on 404 * the object. For example the radix tree for fast page lookup and the 405 * cached refcounted sg-table 406 */ 407 void i915_ttm_free_cached_io_rsgt(struct drm_i915_gem_object *obj) 408 { 409 struct radix_tree_iter iter; 410 void __rcu **slot; 411 412 if (!obj->ttm.cached_io_rsgt) 413 return; 414 415 rcu_read_lock(); 416 radix_tree_for_each_slot(slot, &obj->ttm.get_io_page.radix, &iter, 0) 417 radix_tree_delete(&obj->ttm.get_io_page.radix, iter.index); 418 rcu_read_unlock(); 419 420 i915_refct_sgt_put(obj->ttm.cached_io_rsgt); 421 obj->ttm.cached_io_rsgt = NULL; 422 } 423 424 /** 425 * i915_ttm_purge - Clear an object of its memory 426 * @obj: The object 427 * 428 * This function is called to clear an object of it's memory when it is 429 * marked as not needed anymore. 430 * 431 * Return: 0 on success, negative error code on failure. 432 */ 433 int i915_ttm_purge(struct drm_i915_gem_object *obj) 434 { 435 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 436 struct i915_ttm_tt *i915_tt = 437 container_of(bo->ttm, typeof(*i915_tt), ttm); 438 struct ttm_operation_ctx ctx = { 439 .interruptible = true, 440 .no_wait_gpu = false, 441 }; 442 struct ttm_placement place = {}; 443 int ret; 444 445 if (obj->mm.madv == __I915_MADV_PURGED) 446 return 0; 447 448 ret = ttm_bo_validate(bo, &place, &ctx); 449 if (ret) 450 return ret; 451 452 if (bo->ttm && i915_tt->filp) { 453 /* 454 * The below fput(which eventually calls shmem_truncate) might 455 * be delayed by worker, so when directly called to purge the 456 * pages(like by the shrinker) we should try to be more 457 * aggressive and release the pages immediately. 458 */ 459 #ifdef __linux__ 460 shmem_truncate_range(file_inode(i915_tt->filp), 461 0, (loff_t)-1); 462 #else 463 rw_enter(obj->base.uao->vmobjlock, RW_WRITE); 464 obj->base.uao->pgops->pgo_flush(obj->base.uao, 0, obj->base.size, 465 PGO_ALLPAGES | PGO_FREE); 466 rw_exit(obj->base.uao->vmobjlock); 467 #endif 468 fput(fetch_and_zero(&i915_tt->filp)); 469 } 470 471 obj->write_domain = 0; 472 obj->read_domains = 0; 473 i915_ttm_adjust_gem_after_move(obj); 474 i915_ttm_free_cached_io_rsgt(obj); 475 obj->mm.madv = __I915_MADV_PURGED; 476 477 return 0; 478 } 479 480 static int i915_ttm_shrink(struct drm_i915_gem_object *obj, unsigned int flags) 481 { 482 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 483 struct i915_ttm_tt *i915_tt = 484 container_of(bo->ttm, typeof(*i915_tt), ttm); 485 struct ttm_operation_ctx ctx = { 486 .interruptible = true, 487 .no_wait_gpu = flags & I915_GEM_OBJECT_SHRINK_NO_GPU_WAIT, 488 }; 489 struct ttm_placement place = {}; 490 int ret; 491 492 if (!bo->ttm || i915_ttm_cpu_maps_iomem(bo->resource)) 493 return 0; 494 495 GEM_BUG_ON(!i915_tt->is_shmem); 496 497 if (!i915_tt->filp) 498 return 0; 499 500 ret = ttm_bo_wait_ctx(bo, &ctx); 501 if (ret) 502 return ret; 503 504 switch (obj->mm.madv) { 505 case I915_MADV_DONTNEED: 506 return i915_ttm_purge(obj); 507 case __I915_MADV_PURGED: 508 return 0; 509 } 510 511 if (bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED) 512 return 0; 513 514 bo->ttm->page_flags |= TTM_TT_FLAG_SWAPPED; 515 ret = ttm_bo_validate(bo, &place, &ctx); 516 if (ret) { 517 bo->ttm->page_flags &= ~TTM_TT_FLAG_SWAPPED; 518 return ret; 519 } 520 521 if (flags & I915_GEM_OBJECT_SHRINK_WRITEBACK) 522 #ifdef notyet 523 __shmem_writeback(obj->base.size, i915_tt->filp->f_mapping); 524 #else 525 STUB(); 526 #endif 527 528 return 0; 529 } 530 531 static void i915_ttm_delete_mem_notify(struct ttm_buffer_object *bo) 532 { 533 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 534 535 /* 536 * This gets called twice by ttm, so long as we have a ttm resource or 537 * ttm_tt then we can still safely call this. Due to pipeline-gutting, 538 * we maybe have NULL bo->resource, but in that case we should always 539 * have a ttm alive (like if the pages are swapped out). 540 */ 541 if ((bo->resource || bo->ttm) && !i915_ttm_is_ghost_object(bo)) { 542 __i915_gem_object_pages_fini(obj); 543 i915_ttm_free_cached_io_rsgt(obj); 544 } 545 } 546 547 static struct i915_refct_sgt *i915_ttm_tt_get_st(struct ttm_tt *ttm) 548 { 549 STUB(); 550 return ERR_PTR(-ENOSYS); 551 #ifdef notyet 552 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 553 struct sg_table *st; 554 int ret; 555 556 if (i915_tt->cached_rsgt.table.sgl) 557 return i915_refct_sgt_get(&i915_tt->cached_rsgt); 558 559 st = &i915_tt->cached_rsgt.table; 560 ret = sg_alloc_table_from_pages_segment(st, 561 ttm->pages, ttm->num_pages, 562 0, (unsigned long)ttm->num_pages << PAGE_SHIFT, 563 i915_sg_segment_size(i915_tt->dev), GFP_KERNEL); 564 if (ret) { 565 st->sgl = NULL; 566 return ERR_PTR(ret); 567 } 568 569 ret = dma_map_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 0); 570 if (ret) { 571 sg_free_table(st); 572 return ERR_PTR(ret); 573 } 574 575 return i915_refct_sgt_get(&i915_tt->cached_rsgt); 576 #endif 577 } 578 579 /** 580 * i915_ttm_resource_get_st - Get a refcounted sg-table pointing to the 581 * resource memory 582 * @obj: The GEM object used for sg-table caching 583 * @res: The struct ttm_resource for which an sg-table is requested. 584 * 585 * This function returns a refcounted sg-table representing the memory 586 * pointed to by @res. If @res is the object's current resource it may also 587 * cache the sg_table on the object or attempt to access an already cached 588 * sg-table. The refcounted sg-table needs to be put when no-longer in use. 589 * 590 * Return: A valid pointer to a struct i915_refct_sgt or error pointer on 591 * failure. 592 */ 593 struct i915_refct_sgt * 594 i915_ttm_resource_get_st(struct drm_i915_gem_object *obj, 595 struct ttm_resource *res) 596 { 597 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 598 u32 page_alignment; 599 600 if (!i915_ttm_gtt_binds_lmem(res)) 601 return i915_ttm_tt_get_st(bo->ttm); 602 603 page_alignment = bo->page_alignment << PAGE_SHIFT; 604 if (!page_alignment) 605 page_alignment = obj->mm.region->min_page_size; 606 607 /* 608 * If CPU mapping differs, we need to add the ttm_tt pages to 609 * the resulting st. Might make sense for GGTT. 610 */ 611 GEM_WARN_ON(!i915_ttm_cpu_maps_iomem(res)); 612 if (bo->resource == res) { 613 if (!obj->ttm.cached_io_rsgt) { 614 struct i915_refct_sgt *rsgt; 615 616 rsgt = intel_region_ttm_resource_to_rsgt(obj->mm.region, 617 res, 618 page_alignment); 619 if (IS_ERR(rsgt)) 620 return rsgt; 621 622 obj->ttm.cached_io_rsgt = rsgt; 623 } 624 return i915_refct_sgt_get(obj->ttm.cached_io_rsgt); 625 } 626 627 return intel_region_ttm_resource_to_rsgt(obj->mm.region, res, 628 page_alignment); 629 } 630 631 static int i915_ttm_truncate(struct drm_i915_gem_object *obj) 632 { 633 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 634 long err; 635 636 WARN_ON_ONCE(obj->mm.madv == I915_MADV_WILLNEED); 637 638 err = dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP, 639 true, 15 * HZ); 640 if (err < 0) 641 return err; 642 if (err == 0) 643 return -EBUSY; 644 645 err = i915_ttm_move_notify(bo); 646 if (err) 647 return err; 648 649 return i915_ttm_purge(obj); 650 } 651 652 static void i915_ttm_swap_notify(struct ttm_buffer_object *bo) 653 { 654 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 655 int ret; 656 657 if (i915_ttm_is_ghost_object(bo)) 658 return; 659 660 ret = i915_ttm_move_notify(bo); 661 GEM_WARN_ON(ret); 662 GEM_WARN_ON(obj->ttm.cached_io_rsgt); 663 if (!ret && obj->mm.madv != I915_MADV_WILLNEED) 664 i915_ttm_purge(obj); 665 } 666 667 /** 668 * i915_ttm_resource_mappable - Return true if the ttm resource is CPU 669 * accessible. 670 * @res: The TTM resource to check. 671 * 672 * This is interesting on small-BAR systems where we may encounter lmem objects 673 * that can't be accessed via the CPU. 674 */ 675 bool i915_ttm_resource_mappable(struct ttm_resource *res) 676 { 677 struct i915_ttm_buddy_resource *bman_res = to_ttm_buddy_resource(res); 678 679 if (!i915_ttm_cpu_maps_iomem(res)) 680 return true; 681 682 return bman_res->used_visible_size == PFN_UP(bman_res->base.size); 683 } 684 685 static int i915_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *mem) 686 { 687 struct drm_i915_gem_object *obj = i915_ttm_to_gem(mem->bo); 688 bool unknown_state; 689 690 if (i915_ttm_is_ghost_object(mem->bo)) 691 return -EINVAL; 692 693 if (!kref_get_unless_zero(&obj->base.refcount)) 694 return -EINVAL; 695 696 assert_object_held(obj); 697 698 unknown_state = i915_gem_object_has_unknown_state(obj); 699 i915_gem_object_put(obj); 700 if (unknown_state) 701 return -EINVAL; 702 703 if (!i915_ttm_cpu_maps_iomem(mem)) 704 return 0; 705 706 if (!i915_ttm_resource_mappable(mem)) 707 return -EINVAL; 708 709 mem->bus.caching = ttm_write_combined; 710 mem->bus.is_iomem = true; 711 712 return 0; 713 } 714 715 static unsigned long i915_ttm_io_mem_pfn(struct ttm_buffer_object *bo, 716 unsigned long page_offset) 717 { 718 STUB(); 719 return 0; 720 #ifdef notyet 721 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 722 struct scatterlist *sg; 723 unsigned long base; 724 unsigned int ofs; 725 726 GEM_BUG_ON(i915_ttm_is_ghost_object(bo)); 727 GEM_WARN_ON(bo->ttm); 728 729 base = obj->mm.region->iomap.base - obj->mm.region->region.start; 730 sg = i915_gem_object_page_iter_get_sg(obj, &obj->ttm.get_io_page, page_offset, &ofs); 731 732 return ((base + sg_dma_address(sg)) >> PAGE_SHIFT) + ofs; 733 #endif 734 } 735 736 static int i915_ttm_access_memory(struct ttm_buffer_object *bo, 737 unsigned long offset, void *buf, 738 int len, int write) 739 { 740 STUB(); 741 return -ENOSYS; 742 #ifdef notyet 743 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 744 resource_size_t iomap = obj->mm.region->iomap.base - 745 obj->mm.region->region.start; 746 unsigned long page = offset >> PAGE_SHIFT; 747 unsigned long bytes_left = len; 748 749 /* 750 * TODO: For now just let it fail if the resource is non-mappable, 751 * otherwise we need to perform the memcpy from the gpu here, without 752 * interfering with the object (like moving the entire thing). 753 */ 754 if (!i915_ttm_resource_mappable(bo->resource)) 755 return -EIO; 756 757 offset -= page << PAGE_SHIFT; 758 do { 759 unsigned long bytes = min(bytes_left, PAGE_SIZE - offset); 760 void __iomem *ptr; 761 dma_addr_t daddr; 762 763 daddr = i915_gem_object_get_dma_address(obj, page); 764 ptr = ioremap_wc(iomap + daddr + offset, bytes); 765 if (!ptr) 766 return -EIO; 767 768 if (write) 769 memcpy_toio(ptr, buf, bytes); 770 else 771 memcpy_fromio(buf, ptr, bytes); 772 iounmap(ptr); 773 774 page++; 775 buf += bytes; 776 bytes_left -= bytes; 777 offset = 0; 778 } while (bytes_left); 779 780 return len; 781 #endif 782 } 783 784 /* 785 * All callbacks need to take care not to downcast a struct ttm_buffer_object 786 * without checking its subclass, since it might be a TTM ghost object. 787 */ 788 static struct ttm_device_funcs i915_ttm_bo_driver = { 789 .ttm_tt_create = i915_ttm_tt_create, 790 .ttm_tt_populate = i915_ttm_tt_populate, 791 .ttm_tt_unpopulate = i915_ttm_tt_unpopulate, 792 .ttm_tt_destroy = i915_ttm_tt_destroy, 793 .eviction_valuable = i915_ttm_eviction_valuable, 794 .evict_flags = i915_ttm_evict_flags, 795 .move = i915_ttm_move, 796 .swap_notify = i915_ttm_swap_notify, 797 .delete_mem_notify = i915_ttm_delete_mem_notify, 798 .io_mem_reserve = i915_ttm_io_mem_reserve, 799 .io_mem_pfn = i915_ttm_io_mem_pfn, 800 .access_memory = i915_ttm_access_memory, 801 }; 802 803 /** 804 * i915_ttm_driver - Return a pointer to the TTM device funcs 805 * 806 * Return: Pointer to statically allocated TTM device funcs. 807 */ 808 struct ttm_device_funcs *i915_ttm_driver(void) 809 { 810 return &i915_ttm_bo_driver; 811 } 812 813 static int __i915_ttm_get_pages(struct drm_i915_gem_object *obj, 814 struct ttm_placement *placement) 815 { 816 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 817 struct ttm_operation_ctx ctx = { 818 .interruptible = true, 819 .no_wait_gpu = false, 820 }; 821 int real_num_busy; 822 int ret; 823 824 /* First try only the requested placement. No eviction. */ 825 real_num_busy = fetch_and_zero(&placement->num_busy_placement); 826 ret = ttm_bo_validate(bo, placement, &ctx); 827 if (ret) { 828 ret = i915_ttm_err_to_gem(ret); 829 /* 830 * Anything that wants to restart the operation gets to 831 * do that. 832 */ 833 if (ret == -EDEADLK || ret == -EINTR || ret == -ERESTARTSYS || 834 ret == -EAGAIN) 835 return ret; 836 837 /* 838 * If the initial attempt fails, allow all accepted placements, 839 * evicting if necessary. 840 */ 841 placement->num_busy_placement = real_num_busy; 842 ret = ttm_bo_validate(bo, placement, &ctx); 843 if (ret) 844 return i915_ttm_err_to_gem(ret); 845 } 846 847 if (bo->ttm && !ttm_tt_is_populated(bo->ttm)) { 848 ret = ttm_tt_populate(bo->bdev, bo->ttm, &ctx); 849 if (ret) 850 return ret; 851 852 i915_ttm_adjust_domains_after_move(obj); 853 i915_ttm_adjust_gem_after_move(obj); 854 } 855 856 if (!i915_gem_object_has_pages(obj)) { 857 struct i915_refct_sgt *rsgt = 858 i915_ttm_resource_get_st(obj, bo->resource); 859 860 if (IS_ERR(rsgt)) 861 return PTR_ERR(rsgt); 862 863 GEM_BUG_ON(obj->mm.rsgt); 864 obj->mm.rsgt = rsgt; 865 __i915_gem_object_set_pages(obj, &rsgt->table); 866 } 867 868 GEM_BUG_ON(bo->ttm && ((obj->base.size >> PAGE_SHIFT) < bo->ttm->num_pages)); 869 i915_ttm_adjust_lru(obj); 870 return ret; 871 } 872 873 static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) 874 { 875 struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS]; 876 struct ttm_placement placement; 877 878 /* restricted by sg_alloc_table */ 879 if (overflows_type(obj->base.size >> PAGE_SHIFT, unsigned int)) 880 return -E2BIG; 881 882 GEM_BUG_ON(obj->mm.n_placements > I915_TTM_MAX_PLACEMENTS); 883 884 /* Move to the requested placement. */ 885 i915_ttm_placement_from_obj(obj, &requested, busy, &placement); 886 887 return __i915_ttm_get_pages(obj, &placement); 888 } 889 890 /** 891 * DOC: Migration vs eviction 892 * 893 * GEM migration may not be the same as TTM migration / eviction. If 894 * the TTM core decides to evict an object it may be evicted to a 895 * TTM memory type that is not in the object's allowable GEM regions, or 896 * in fact theoretically to a TTM memory type that doesn't correspond to 897 * a GEM memory region. In that case the object's GEM region is not 898 * updated, and the data is migrated back to the GEM region at 899 * get_pages time. TTM may however set up CPU ptes to the object even 900 * when it is evicted. 901 * Gem forced migration using the i915_ttm_migrate() op, is allowed even 902 * to regions that are not in the object's list of allowable placements. 903 */ 904 static int __i915_ttm_migrate(struct drm_i915_gem_object *obj, 905 struct intel_memory_region *mr, 906 unsigned int flags) 907 { 908 struct ttm_place requested; 909 struct ttm_placement placement; 910 int ret; 911 912 i915_ttm_place_from_region(mr, &requested, obj->bo_offset, 913 obj->base.size, flags); 914 placement.num_placement = 1; 915 placement.num_busy_placement = 1; 916 placement.placement = &requested; 917 placement.busy_placement = &requested; 918 919 ret = __i915_ttm_get_pages(obj, &placement); 920 if (ret) 921 return ret; 922 923 /* 924 * Reinitialize the region bindings. This is primarily 925 * required for objects where the new region is not in 926 * its allowable placements. 927 */ 928 if (obj->mm.region != mr) { 929 i915_gem_object_release_memory_region(obj); 930 i915_gem_object_init_memory_region(obj, mr); 931 } 932 933 return 0; 934 } 935 936 static int i915_ttm_migrate(struct drm_i915_gem_object *obj, 937 struct intel_memory_region *mr, 938 unsigned int flags) 939 { 940 return __i915_ttm_migrate(obj, mr, flags); 941 } 942 943 static void i915_ttm_put_pages(struct drm_i915_gem_object *obj, 944 struct sg_table *st) 945 { 946 /* 947 * We're currently not called from a shrinker, so put_pages() 948 * typically means the object is about to destroyed, or called 949 * from move_notify(). So just avoid doing much for now. 950 * If the object is not destroyed next, The TTM eviction logic 951 * and shrinkers will move it out if needed. 952 */ 953 954 if (obj->mm.rsgt) 955 i915_refct_sgt_put(fetch_and_zero(&obj->mm.rsgt)); 956 } 957 958 /** 959 * i915_ttm_adjust_lru - Adjust an object's position on relevant LRU lists. 960 * @obj: The object 961 */ 962 void i915_ttm_adjust_lru(struct drm_i915_gem_object *obj) 963 { 964 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 965 struct i915_ttm_tt *i915_tt = 966 container_of(bo->ttm, typeof(*i915_tt), ttm); 967 bool shrinkable = 968 bo->ttm && i915_tt->filp && ttm_tt_is_populated(bo->ttm); 969 970 /* 971 * Don't manipulate the TTM LRUs while in TTM bo destruction. 972 * We're called through i915_ttm_delete_mem_notify(). 973 */ 974 if (!kref_read(&bo->kref)) 975 return; 976 977 /* 978 * We skip managing the shrinker LRU in set_pages() and just manage 979 * everything here. This does at least solve the issue with having 980 * temporary shmem mappings(like with evicted lmem) not being visible to 981 * the shrinker. Only our shmem objects are shrinkable, everything else 982 * we keep as unshrinkable. 983 * 984 * To make sure everything plays nice we keep an extra shrink pin in TTM 985 * if the underlying pages are not currently shrinkable. Once we release 986 * our pin, like when the pages are moved to shmem, the pages will then 987 * be added to the shrinker LRU, assuming the caller isn't also holding 988 * a pin. 989 * 990 * TODO: consider maybe also bumping the shrinker list here when we have 991 * already unpinned it, which should give us something more like an LRU. 992 * 993 * TODO: There is a small window of opportunity for this function to 994 * get called from eviction after we've dropped the last GEM refcount, 995 * but before the TTM deleted flag is set on the object. Avoid 996 * adjusting the shrinker list in such cases, since the object is 997 * not available to the shrinker anyway due to its zero refcount. 998 * To fix this properly we should move to a TTM shrinker LRU list for 999 * these objects. 1000 */ 1001 if (kref_get_unless_zero(&obj->base.refcount)) { 1002 if (shrinkable != obj->mm.ttm_shrinkable) { 1003 if (shrinkable) { 1004 if (obj->mm.madv == I915_MADV_WILLNEED) 1005 __i915_gem_object_make_shrinkable(obj); 1006 else 1007 __i915_gem_object_make_purgeable(obj); 1008 } else { 1009 i915_gem_object_make_unshrinkable(obj); 1010 } 1011 1012 obj->mm.ttm_shrinkable = shrinkable; 1013 } 1014 i915_gem_object_put(obj); 1015 } 1016 1017 /* 1018 * Put on the correct LRU list depending on the MADV status 1019 */ 1020 spin_lock(&bo->bdev->lru_lock); 1021 if (shrinkable) { 1022 /* Try to keep shmem_tt from being considered for shrinking. */ 1023 bo->priority = TTM_MAX_BO_PRIORITY - 1; 1024 } else if (obj->mm.madv != I915_MADV_WILLNEED) { 1025 bo->priority = I915_TTM_PRIO_PURGE; 1026 } else if (!i915_gem_object_has_pages(obj)) { 1027 bo->priority = I915_TTM_PRIO_NO_PAGES; 1028 } else { 1029 struct ttm_resource_manager *man = 1030 ttm_manager_type(bo->bdev, bo->resource->mem_type); 1031 1032 /* 1033 * If we need to place an LMEM resource which doesn't need CPU 1034 * access then we should try not to victimize mappable objects 1035 * first, since we likely end up stealing more of the mappable 1036 * portion. And likewise when we try to find space for a mappble 1037 * object, we know not to ever victimize objects that don't 1038 * occupy any mappable pages. 1039 */ 1040 if (i915_ttm_cpu_maps_iomem(bo->resource) && 1041 i915_ttm_buddy_man_visible_size(man) < man->size && 1042 !(obj->flags & I915_BO_ALLOC_GPU_ONLY)) 1043 bo->priority = I915_TTM_PRIO_NEEDS_CPU_ACCESS; 1044 else 1045 bo->priority = I915_TTM_PRIO_HAS_PAGES; 1046 } 1047 1048 ttm_bo_move_to_lru_tail(bo); 1049 spin_unlock(&bo->bdev->lru_lock); 1050 } 1051 1052 /* 1053 * TTM-backed gem object destruction requires some clarification. 1054 * Basically we have two possibilities here. We can either rely on the 1055 * i915 delayed destruction and put the TTM object when the object 1056 * is idle. This would be detected by TTM which would bypass the 1057 * TTM delayed destroy handling. The other approach is to put the TTM 1058 * object early and rely on the TTM destroyed handling, and then free 1059 * the leftover parts of the GEM object once TTM's destroyed list handling is 1060 * complete. For now, we rely on the latter for two reasons: 1061 * a) TTM can evict an object even when it's on the delayed destroy list, 1062 * which in theory allows for complete eviction. 1063 * b) There is work going on in TTM to allow freeing an object even when 1064 * it's not idle, and using the TTM destroyed list handling could help us 1065 * benefit from that. 1066 */ 1067 static void i915_ttm_delayed_free(struct drm_i915_gem_object *obj) 1068 { 1069 GEM_BUG_ON(!obj->ttm.created); 1070 1071 ttm_bo_put(i915_gem_to_ttm(obj)); 1072 } 1073 1074 #ifdef notyet 1075 static vm_fault_t vm_fault_ttm(struct vm_fault *vmf) 1076 { 1077 struct vm_area_struct *area = vmf->vma; 1078 struct ttm_buffer_object *bo = area->vm_private_data; 1079 struct drm_device *dev = bo->base.dev; 1080 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 1081 intel_wakeref_t wakeref = 0; 1082 vm_fault_t ret; 1083 int idx; 1084 1085 /* Sanity check that we allow writing into this object */ 1086 if (unlikely(i915_gem_object_is_readonly(obj) && 1087 area->vm_flags & VM_WRITE)) 1088 return VM_FAULT_SIGBUS; 1089 1090 ret = ttm_bo_vm_reserve(bo, vmf); 1091 if (ret) 1092 return ret; 1093 1094 if (obj->mm.madv != I915_MADV_WILLNEED) { 1095 dma_resv_unlock(bo->base.resv); 1096 return VM_FAULT_SIGBUS; 1097 } 1098 1099 /* 1100 * This must be swapped out with shmem ttm_tt (pipeline-gutting). 1101 * Calling ttm_bo_validate() here with TTM_PL_SYSTEM should only go as 1102 * far as far doing a ttm_bo_move_null(), which should skip all the 1103 * other junk. 1104 */ 1105 if (!bo->resource) { 1106 struct ttm_operation_ctx ctx = { 1107 .interruptible = true, 1108 .no_wait_gpu = true, /* should be idle already */ 1109 }; 1110 int err; 1111 1112 GEM_BUG_ON(!bo->ttm || !(bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED)); 1113 1114 err = ttm_bo_validate(bo, i915_ttm_sys_placement(), &ctx); 1115 if (err) { 1116 dma_resv_unlock(bo->base.resv); 1117 return VM_FAULT_SIGBUS; 1118 } 1119 } else if (!i915_ttm_resource_mappable(bo->resource)) { 1120 int err = -ENODEV; 1121 int i; 1122 1123 for (i = 0; i < obj->mm.n_placements; i++) { 1124 struct intel_memory_region *mr = obj->mm.placements[i]; 1125 unsigned int flags; 1126 1127 if (!mr->io_size && mr->type != INTEL_MEMORY_SYSTEM) 1128 continue; 1129 1130 flags = obj->flags; 1131 flags &= ~I915_BO_ALLOC_GPU_ONLY; 1132 err = __i915_ttm_migrate(obj, mr, flags); 1133 if (!err) 1134 break; 1135 } 1136 1137 if (err) { 1138 drm_dbg(dev, "Unable to make resource CPU accessible(err = %pe)\n", 1139 ERR_PTR(err)); 1140 dma_resv_unlock(bo->base.resv); 1141 ret = VM_FAULT_SIGBUS; 1142 goto out_rpm; 1143 } 1144 } 1145 1146 if (i915_ttm_cpu_maps_iomem(bo->resource)) 1147 wakeref = intel_runtime_pm_get(&to_i915(obj->base.dev)->runtime_pm); 1148 1149 if (drm_dev_enter(dev, &idx)) { 1150 ret = ttm_bo_vm_fault_reserved(vmf, vmf->vma->vm_page_prot, 1151 TTM_BO_VM_NUM_PREFAULT); 1152 drm_dev_exit(idx); 1153 } else { 1154 ret = ttm_bo_vm_dummy_page(vmf, vmf->vma->vm_page_prot); 1155 } 1156 1157 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) 1158 goto out_rpm; 1159 1160 /* 1161 * ttm_bo_vm_reserve() already has dma_resv_lock. 1162 * userfault_count is protected by dma_resv lock and rpm wakeref. 1163 */ 1164 if (ret == VM_FAULT_NOPAGE && wakeref && !obj->userfault_count) { 1165 obj->userfault_count = 1; 1166 spin_lock(&to_i915(obj->base.dev)->runtime_pm.lmem_userfault_lock); 1167 list_add(&obj->userfault_link, &to_i915(obj->base.dev)->runtime_pm.lmem_userfault_list); 1168 spin_unlock(&to_i915(obj->base.dev)->runtime_pm.lmem_userfault_lock); 1169 1170 GEM_WARN_ON(!i915_ttm_cpu_maps_iomem(bo->resource)); 1171 } 1172 1173 if (wakeref & CONFIG_DRM_I915_USERFAULT_AUTOSUSPEND) 1174 intel_wakeref_auto(&to_i915(obj->base.dev)->runtime_pm.userfault_wakeref, 1175 msecs_to_jiffies_timeout(CONFIG_DRM_I915_USERFAULT_AUTOSUSPEND)); 1176 1177 i915_ttm_adjust_lru(obj); 1178 1179 dma_resv_unlock(bo->base.resv); 1180 1181 out_rpm: 1182 if (wakeref) 1183 intel_runtime_pm_put(&to_i915(obj->base.dev)->runtime_pm, wakeref); 1184 1185 return ret; 1186 } 1187 1188 static int 1189 vm_access_ttm(struct vm_area_struct *area, unsigned long addr, 1190 void *buf, int len, int write) 1191 { 1192 struct drm_i915_gem_object *obj = 1193 i915_ttm_to_gem(area->vm_private_data); 1194 1195 if (i915_gem_object_is_readonly(obj) && write) 1196 return -EACCES; 1197 1198 return ttm_bo_vm_access(area, addr, buf, len, write); 1199 } 1200 1201 static void ttm_vm_open(struct vm_area_struct *vma) 1202 { 1203 struct drm_i915_gem_object *obj = 1204 i915_ttm_to_gem(vma->vm_private_data); 1205 1206 GEM_BUG_ON(i915_ttm_is_ghost_object(vma->vm_private_data)); 1207 i915_gem_object_get(obj); 1208 } 1209 1210 static void ttm_vm_close(struct vm_area_struct *vma) 1211 { 1212 struct drm_i915_gem_object *obj = 1213 i915_ttm_to_gem(vma->vm_private_data); 1214 1215 GEM_BUG_ON(i915_ttm_is_ghost_object(vma->vm_private_data)); 1216 i915_gem_object_put(obj); 1217 } 1218 1219 static const struct vm_operations_struct vm_ops_ttm = { 1220 .fault = vm_fault_ttm, 1221 .access = vm_access_ttm, 1222 .open = ttm_vm_open, 1223 .close = ttm_vm_close, 1224 }; 1225 1226 #endif 1227 1228 static u64 i915_ttm_mmap_offset(struct drm_i915_gem_object *obj) 1229 { 1230 /* The ttm_bo must be allocated with I915_BO_ALLOC_USER */ 1231 GEM_BUG_ON(!drm_mm_node_allocated(&obj->base.vma_node.vm_node)); 1232 1233 return drm_vma_node_offset_addr(&obj->base.vma_node); 1234 } 1235 1236 static void i915_ttm_unmap_virtual(struct drm_i915_gem_object *obj) 1237 { 1238 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 1239 intel_wakeref_t wakeref = 0; 1240 1241 assert_object_held_shared(obj); 1242 1243 if (i915_ttm_cpu_maps_iomem(bo->resource)) { 1244 wakeref = intel_runtime_pm_get(&to_i915(obj->base.dev)->runtime_pm); 1245 1246 /* userfault_count is protected by obj lock and rpm wakeref. */ 1247 if (obj->userfault_count) { 1248 spin_lock(&to_i915(obj->base.dev)->runtime_pm.lmem_userfault_lock); 1249 list_del(&obj->userfault_link); 1250 spin_unlock(&to_i915(obj->base.dev)->runtime_pm.lmem_userfault_lock); 1251 obj->userfault_count = 0; 1252 } 1253 } 1254 1255 GEM_WARN_ON(obj->userfault_count); 1256 1257 ttm_bo_unmap_virtual(i915_gem_to_ttm(obj)); 1258 1259 if (wakeref) 1260 intel_runtime_pm_put(&to_i915(obj->base.dev)->runtime_pm, wakeref); 1261 } 1262 1263 static const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = { 1264 .name = "i915_gem_object_ttm", 1265 .flags = I915_GEM_OBJECT_IS_SHRINKABLE | 1266 I915_GEM_OBJECT_SELF_MANAGED_SHRINK_LIST, 1267 1268 .get_pages = i915_ttm_get_pages, 1269 .put_pages = i915_ttm_put_pages, 1270 .truncate = i915_ttm_truncate, 1271 .shrink = i915_ttm_shrink, 1272 1273 .adjust_lru = i915_ttm_adjust_lru, 1274 .delayed_free = i915_ttm_delayed_free, 1275 .migrate = i915_ttm_migrate, 1276 1277 .mmap_offset = i915_ttm_mmap_offset, 1278 .unmap_virtual = i915_ttm_unmap_virtual, 1279 #ifdef notyet 1280 .mmap_ops = &vm_ops_ttm, 1281 #endif 1282 }; 1283 1284 void i915_ttm_bo_destroy(struct ttm_buffer_object *bo) 1285 { 1286 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 1287 1288 i915_gem_object_release_memory_region(obj); 1289 mutex_destroy(&obj->ttm.get_io_page.lock); 1290 1291 if (obj->ttm.created) { 1292 /* 1293 * We freely manage the shrinker LRU outide of the mm.pages life 1294 * cycle. As a result when destroying the object we should be 1295 * extra paranoid and ensure we remove it from the LRU, before 1296 * we free the object. 1297 * 1298 * Touching the ttm_shrinkable outside of the object lock here 1299 * should be safe now that the last GEM object ref was dropped. 1300 */ 1301 if (obj->mm.ttm_shrinkable) 1302 i915_gem_object_make_unshrinkable(obj); 1303 1304 i915_ttm_backup_free(obj); 1305 1306 /* This releases all gem object bindings to the backend. */ 1307 __i915_gem_free_object(obj); 1308 1309 call_rcu(&obj->rcu, __i915_gem_free_object_rcu); 1310 } else { 1311 __i915_gem_object_fini(obj); 1312 } 1313 } 1314 1315 /* 1316 * __i915_gem_ttm_object_init - Initialize a ttm-backed i915 gem object 1317 * @mem: The initial memory region for the object. 1318 * @obj: The gem object. 1319 * @size: Object size in bytes. 1320 * @flags: gem object flags. 1321 * 1322 * Return: 0 on success, negative error code on failure. 1323 */ 1324 int __i915_gem_ttm_object_init(struct intel_memory_region *mem, 1325 struct drm_i915_gem_object *obj, 1326 resource_size_t offset, 1327 resource_size_t size, 1328 resource_size_t page_size, 1329 unsigned int flags) 1330 { 1331 static struct lock_class_key lock_class; 1332 struct drm_i915_private *i915 = mem->i915; 1333 struct ttm_operation_ctx ctx = { 1334 .interruptible = true, 1335 .no_wait_gpu = false, 1336 }; 1337 enum ttm_bo_type bo_type; 1338 int ret; 1339 1340 drm_gem_private_object_init(&i915->drm, &obj->base, size); 1341 i915_gem_object_init(obj, &i915_gem_ttm_obj_ops, &lock_class, flags); 1342 1343 obj->bo_offset = offset; 1344 1345 /* Don't put on a region list until we're either locked or fully initialized. */ 1346 obj->mm.region = mem; 1347 INIT_LIST_HEAD(&obj->mm.region_link); 1348 1349 INIT_RADIX_TREE(&obj->ttm.get_io_page.radix, GFP_KERNEL | __GFP_NOWARN); 1350 rw_init(&obj->ttm.get_io_page.lock, "i915ttm"); 1351 bo_type = (obj->flags & I915_BO_ALLOC_USER) ? ttm_bo_type_device : 1352 ttm_bo_type_kernel; 1353 1354 obj->base.vma_node.driver_private = i915_gem_to_ttm(obj); 1355 1356 /* Forcing the page size is kernel internal only */ 1357 GEM_BUG_ON(page_size && obj->mm.n_placements); 1358 1359 /* 1360 * Keep an extra shrink pin to prevent the object from being made 1361 * shrinkable too early. If the ttm_tt is ever allocated in shmem, we 1362 * drop the pin. The TTM backend manages the shrinker LRU itself, 1363 * outside of the normal mm.pages life cycle. 1364 */ 1365 i915_gem_object_make_unshrinkable(obj); 1366 1367 /* 1368 * If this function fails, it will call the destructor, but 1369 * our caller still owns the object. So no freeing in the 1370 * destructor until obj->ttm.created is true. 1371 * Similarly, in delayed_destroy, we can't call ttm_bo_put() 1372 * until successful initialization. 1373 */ 1374 ret = ttm_bo_init_reserved(&i915->bdev, i915_gem_to_ttm(obj), bo_type, 1375 &i915_sys_placement, page_size >> PAGE_SHIFT, 1376 &ctx, NULL, NULL, i915_ttm_bo_destroy); 1377 1378 /* 1379 * XXX: The ttm_bo_init_reserved() functions returns -ENOSPC if the size 1380 * is too big to add vma. The direct function that returns -ENOSPC is 1381 * drm_mm_insert_node_in_range(). To handle the same error as other code 1382 * that returns -E2BIG when the size is too large, it converts -ENOSPC to 1383 * -E2BIG. 1384 */ 1385 if (size >> PAGE_SHIFT > INT_MAX && ret == -ENOSPC) 1386 ret = -E2BIG; 1387 1388 if (ret) 1389 return i915_ttm_err_to_gem(ret); 1390 1391 obj->ttm.created = true; 1392 i915_gem_object_release_memory_region(obj); 1393 i915_gem_object_init_memory_region(obj, mem); 1394 i915_ttm_adjust_domains_after_move(obj); 1395 i915_ttm_adjust_gem_after_move(obj); 1396 i915_gem_object_unlock(obj); 1397 1398 return 0; 1399 } 1400 1401 static const struct intel_memory_region_ops ttm_system_region_ops = { 1402 .init_object = __i915_gem_ttm_object_init, 1403 .release = intel_region_ttm_fini, 1404 }; 1405 1406 struct intel_memory_region * 1407 i915_gem_ttm_system_setup(struct drm_i915_private *i915, 1408 u16 type, u16 instance) 1409 { 1410 struct intel_memory_region *mr; 1411 1412 mr = intel_memory_region_create(i915, 0, 1413 totalram_pages() << PAGE_SHIFT, 1414 PAGE_SIZE, 0, 0, 1415 type, instance, 1416 &ttm_system_region_ops); 1417 if (IS_ERR(mr)) 1418 return mr; 1419 1420 intel_memory_region_set_name(mr, "system-ttm"); 1421 return mr; 1422 } 1423