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_bo_driver.h> 9 #include <drm/ttm/ttm_placement.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 place->fpfn = offset >> PAGE_SHIFT; 144 place->lpfn = place->fpfn + (size >> PAGE_SHIFT); 145 } else if (mr->io_size && mr->io_size < mr->total) { 146 if (flags & I915_BO_ALLOC_GPU_ONLY) { 147 place->flags |= TTM_PL_FLAG_TOPDOWN; 148 } else { 149 place->fpfn = 0; 150 place->lpfn = mr->io_size >> PAGE_SHIFT; 151 } 152 } 153 } 154 155 static void 156 i915_ttm_placement_from_obj(const struct drm_i915_gem_object *obj, 157 struct ttm_place *requested, 158 struct ttm_place *busy, 159 struct ttm_placement *placement) 160 { 161 unsigned int num_allowed = obj->mm.n_placements; 162 unsigned int flags = obj->flags; 163 unsigned int i; 164 165 placement->num_placement = 1; 166 i915_ttm_place_from_region(num_allowed ? obj->mm.placements[0] : 167 obj->mm.region, requested, obj->bo_offset, 168 obj->base.size, flags); 169 170 /* Cache this on object? */ 171 placement->num_busy_placement = num_allowed; 172 for (i = 0; i < placement->num_busy_placement; ++i) 173 i915_ttm_place_from_region(obj->mm.placements[i], busy + i, 174 obj->bo_offset, obj->base.size, flags); 175 176 if (num_allowed == 0) { 177 *busy = *requested; 178 placement->num_busy_placement = 1; 179 } 180 181 placement->placement = requested; 182 placement->busy_placement = busy; 183 } 184 185 static int i915_ttm_tt_shmem_populate(struct ttm_device *bdev, 186 struct ttm_tt *ttm, 187 struct ttm_operation_ctx *ctx) 188 { 189 struct drm_i915_private *i915 = container_of(bdev, typeof(*i915), bdev); 190 struct intel_memory_region *mr = i915->mm.regions[INTEL_MEMORY_SYSTEM]; 191 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 192 const unsigned int max_segment = i915_sg_segment_size(); 193 const size_t size = (size_t)ttm->num_pages << PAGE_SHIFT; 194 struct file *filp = i915_tt->filp; 195 struct sgt_iter sgt_iter; 196 struct sg_table *st; 197 struct page *page; 198 unsigned long i; 199 int err; 200 201 if (!filp) { 202 struct address_space *mapping; 203 gfp_t mask; 204 205 filp = shmem_file_setup("i915-shmem-tt", size, VM_NORESERVE); 206 if (IS_ERR(filp)) 207 return PTR_ERR(filp); 208 209 mask = GFP_HIGHUSER | __GFP_RECLAIMABLE; 210 211 mapping = filp->f_mapping; 212 mapping_set_gfp_mask(mapping, mask); 213 GEM_BUG_ON(!(mapping_gfp_mask(mapping) & __GFP_RECLAIM)); 214 215 i915_tt->filp = filp; 216 } 217 218 st = &i915_tt->cached_rsgt.table; 219 err = shmem_sg_alloc_table(i915, st, size, mr, filp->f_mapping, 220 max_segment); 221 if (err) 222 return err; 223 224 err = dma_map_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 225 DMA_ATTR_SKIP_CPU_SYNC); 226 if (err) 227 goto err_free_st; 228 229 i = 0; 230 for_each_sgt_page(page, sgt_iter, st) 231 ttm->pages[i++] = page; 232 233 if (ttm->page_flags & TTM_TT_FLAG_SWAPPED) 234 ttm->page_flags &= ~TTM_TT_FLAG_SWAPPED; 235 236 return 0; 237 238 err_free_st: 239 shmem_sg_free_table(st, filp->f_mapping, false, false); 240 241 return err; 242 } 243 244 static void i915_ttm_tt_shmem_unpopulate(struct ttm_tt *ttm) 245 { 246 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 247 bool backup = ttm->page_flags & TTM_TT_FLAG_SWAPPED; 248 struct sg_table *st = &i915_tt->cached_rsgt.table; 249 250 shmem_sg_free_table(st, file_inode(i915_tt->filp)->i_mapping, 251 backup, backup); 252 } 253 254 static void i915_ttm_tt_release(struct kref *ref) 255 { 256 struct i915_ttm_tt *i915_tt = 257 container_of(ref, typeof(*i915_tt), cached_rsgt.kref); 258 struct sg_table *st = &i915_tt->cached_rsgt.table; 259 260 GEM_WARN_ON(st->sgl); 261 262 kfree(i915_tt); 263 } 264 265 static const struct i915_refct_sgt_ops tt_rsgt_ops = { 266 .release = i915_ttm_tt_release 267 }; 268 269 static struct ttm_tt *i915_ttm_tt_create(struct ttm_buffer_object *bo, 270 uint32_t page_flags) 271 { 272 struct drm_i915_private *i915 = container_of(bo->bdev, typeof(*i915), 273 bdev); 274 struct ttm_resource_manager *man = 275 ttm_manager_type(bo->bdev, bo->resource->mem_type); 276 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 277 unsigned long ccs_pages = 0; 278 enum ttm_caching caching; 279 struct i915_ttm_tt *i915_tt; 280 int ret; 281 282 if (!obj) 283 return NULL; 284 285 i915_tt = kzalloc(sizeof(*i915_tt), GFP_KERNEL); 286 if (!i915_tt) 287 return NULL; 288 289 if (obj->flags & I915_BO_ALLOC_CPU_CLEAR && 290 man->use_tt) 291 page_flags |= TTM_TT_FLAG_ZERO_ALLOC; 292 293 caching = i915_ttm_select_tt_caching(obj); 294 if (i915_gem_object_is_shrinkable(obj) && caching == ttm_cached) { 295 page_flags |= TTM_TT_FLAG_EXTERNAL | 296 TTM_TT_FLAG_EXTERNAL_MAPPABLE; 297 i915_tt->is_shmem = true; 298 } 299 300 if (HAS_FLAT_CCS(i915) && i915_gem_object_needs_ccs_pages(obj)) 301 ccs_pages = DIV_ROUND_UP(DIV_ROUND_UP(bo->base.size, 302 NUM_BYTES_PER_CCS_BYTE), 303 PAGE_SIZE); 304 305 ret = ttm_tt_init(&i915_tt->ttm, bo, page_flags, caching, ccs_pages); 306 if (ret) 307 goto err_free; 308 309 __i915_refct_sgt_init(&i915_tt->cached_rsgt, bo->base.size, 310 &tt_rsgt_ops); 311 312 i915_tt->dev = obj->base.dev->dev; 313 314 return &i915_tt->ttm; 315 316 err_free: 317 kfree(i915_tt); 318 return NULL; 319 } 320 321 static int i915_ttm_tt_populate(struct ttm_device *bdev, 322 struct ttm_tt *ttm, 323 struct ttm_operation_ctx *ctx) 324 { 325 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 326 327 if (i915_tt->is_shmem) 328 return i915_ttm_tt_shmem_populate(bdev, ttm, ctx); 329 330 return ttm_pool_alloc(&bdev->pool, ttm, ctx); 331 } 332 333 static void i915_ttm_tt_unpopulate(struct ttm_device *bdev, struct ttm_tt *ttm) 334 { 335 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 336 struct sg_table *st = &i915_tt->cached_rsgt.table; 337 338 if (st->sgl) 339 dma_unmap_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 0); 340 341 if (i915_tt->is_shmem) { 342 i915_ttm_tt_shmem_unpopulate(ttm); 343 } else { 344 sg_free_table(st); 345 ttm_pool_free(&bdev->pool, ttm); 346 } 347 } 348 349 static void i915_ttm_tt_destroy(struct ttm_device *bdev, struct ttm_tt *ttm) 350 { 351 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 352 353 if (i915_tt->filp) 354 fput(i915_tt->filp); 355 356 ttm_tt_fini(ttm); 357 i915_refct_sgt_put(&i915_tt->cached_rsgt); 358 } 359 360 static bool i915_ttm_eviction_valuable(struct ttm_buffer_object *bo, 361 const struct ttm_place *place) 362 { 363 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 364 struct ttm_resource *res = bo->resource; 365 366 if (!obj) 367 return false; 368 369 /* 370 * EXTERNAL objects should never be swapped out by TTM, instead we need 371 * to handle that ourselves. TTM will already skip such objects for us, 372 * but we would like to avoid grabbing locks for no good reason. 373 */ 374 if (bo->ttm && bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL) 375 return false; 376 377 /* Will do for now. Our pinned objects are still on TTM's LRU lists */ 378 if (!i915_gem_object_evictable(obj)) 379 return false; 380 381 switch (res->mem_type) { 382 case I915_PL_LMEM0: { 383 struct ttm_resource_manager *man = 384 ttm_manager_type(bo->bdev, res->mem_type); 385 struct i915_ttm_buddy_resource *bman_res = 386 to_ttm_buddy_resource(res); 387 struct drm_buddy *mm = bman_res->mm; 388 struct drm_buddy_block *block; 389 390 if (!place->fpfn && !place->lpfn) 391 return true; 392 393 GEM_BUG_ON(!place->lpfn); 394 395 /* 396 * If we just want something mappable then we can quickly check 397 * if the current victim resource is using any of the CPU 398 * visible portion. 399 */ 400 if (!place->fpfn && 401 place->lpfn == i915_ttm_buddy_man_visible_size(man)) 402 return bman_res->used_visible_size > 0; 403 404 /* Real range allocation */ 405 list_for_each_entry(block, &bman_res->blocks, link) { 406 unsigned long fpfn = 407 drm_buddy_block_offset(block) >> PAGE_SHIFT; 408 unsigned long lpfn = fpfn + 409 (drm_buddy_block_size(mm, block) >> PAGE_SHIFT); 410 411 if (place->fpfn < lpfn && place->lpfn > fpfn) 412 return true; 413 } 414 return false; 415 } default: 416 break; 417 } 418 419 return true; 420 } 421 422 static void i915_ttm_evict_flags(struct ttm_buffer_object *bo, 423 struct ttm_placement *placement) 424 { 425 *placement = i915_sys_placement; 426 } 427 428 /** 429 * i915_ttm_free_cached_io_rsgt - Free object cached LMEM information 430 * @obj: The GEM object 431 * This function frees any LMEM-related information that is cached on 432 * the object. For example the radix tree for fast page lookup and the 433 * cached refcounted sg-table 434 */ 435 void i915_ttm_free_cached_io_rsgt(struct drm_i915_gem_object *obj) 436 { 437 struct radix_tree_iter iter; 438 void __rcu **slot; 439 440 if (!obj->ttm.cached_io_rsgt) 441 return; 442 443 rcu_read_lock(); 444 radix_tree_for_each_slot(slot, &obj->ttm.get_io_page.radix, &iter, 0) 445 radix_tree_delete(&obj->ttm.get_io_page.radix, iter.index); 446 rcu_read_unlock(); 447 448 i915_refct_sgt_put(obj->ttm.cached_io_rsgt); 449 obj->ttm.cached_io_rsgt = NULL; 450 } 451 452 /** 453 * i915_ttm_purge - Clear an object of its memory 454 * @obj: The object 455 * 456 * This function is called to clear an object of it's memory when it is 457 * marked as not needed anymore. 458 * 459 * Return: 0 on success, negative error code on failure. 460 */ 461 int i915_ttm_purge(struct drm_i915_gem_object *obj) 462 { 463 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 464 struct i915_ttm_tt *i915_tt = 465 container_of(bo->ttm, typeof(*i915_tt), ttm); 466 struct ttm_operation_ctx ctx = { 467 .interruptible = true, 468 .no_wait_gpu = false, 469 }; 470 struct ttm_placement place = {}; 471 int ret; 472 473 if (obj->mm.madv == __I915_MADV_PURGED) 474 return 0; 475 476 ret = ttm_bo_validate(bo, &place, &ctx); 477 if (ret) 478 return ret; 479 480 if (bo->ttm && i915_tt->filp) { 481 /* 482 * The below fput(which eventually calls shmem_truncate) might 483 * be delayed by worker, so when directly called to purge the 484 * pages(like by the shrinker) we should try to be more 485 * aggressive and release the pages immediately. 486 */ 487 shmem_truncate_range(file_inode(i915_tt->filp), 488 0, (loff_t)-1); 489 fput(fetch_and_zero(&i915_tt->filp)); 490 } 491 492 obj->write_domain = 0; 493 obj->read_domains = 0; 494 i915_ttm_adjust_gem_after_move(obj); 495 i915_ttm_free_cached_io_rsgt(obj); 496 obj->mm.madv = __I915_MADV_PURGED; 497 498 return 0; 499 } 500 501 static int i915_ttm_shrink(struct drm_i915_gem_object *obj, unsigned int flags) 502 { 503 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 504 struct i915_ttm_tt *i915_tt = 505 container_of(bo->ttm, typeof(*i915_tt), ttm); 506 struct ttm_operation_ctx ctx = { 507 .interruptible = true, 508 .no_wait_gpu = flags & I915_GEM_OBJECT_SHRINK_NO_GPU_WAIT, 509 }; 510 struct ttm_placement place = {}; 511 int ret; 512 513 if (!bo->ttm || bo->resource->mem_type != TTM_PL_SYSTEM) 514 return 0; 515 516 GEM_BUG_ON(!i915_tt->is_shmem); 517 518 if (!i915_tt->filp) 519 return 0; 520 521 ret = ttm_bo_wait_ctx(bo, &ctx); 522 if (ret) 523 return ret; 524 525 switch (obj->mm.madv) { 526 case I915_MADV_DONTNEED: 527 return i915_ttm_purge(obj); 528 case __I915_MADV_PURGED: 529 return 0; 530 } 531 532 if (bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED) 533 return 0; 534 535 bo->ttm->page_flags |= TTM_TT_FLAG_SWAPPED; 536 ret = ttm_bo_validate(bo, &place, &ctx); 537 if (ret) { 538 bo->ttm->page_flags &= ~TTM_TT_FLAG_SWAPPED; 539 return ret; 540 } 541 542 if (flags & I915_GEM_OBJECT_SHRINK_WRITEBACK) 543 __shmem_writeback(obj->base.size, i915_tt->filp->f_mapping); 544 545 return 0; 546 } 547 548 static void i915_ttm_delete_mem_notify(struct ttm_buffer_object *bo) 549 { 550 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 551 552 if (likely(obj)) { 553 __i915_gem_object_pages_fini(obj); 554 i915_ttm_free_cached_io_rsgt(obj); 555 } 556 } 557 558 static struct i915_refct_sgt *i915_ttm_tt_get_st(struct ttm_tt *ttm) 559 { 560 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 561 struct sg_table *st; 562 int ret; 563 564 if (i915_tt->cached_rsgt.table.sgl) 565 return i915_refct_sgt_get(&i915_tt->cached_rsgt); 566 567 st = &i915_tt->cached_rsgt.table; 568 ret = sg_alloc_table_from_pages_segment(st, 569 ttm->pages, ttm->num_pages, 570 0, (unsigned long)ttm->num_pages << PAGE_SHIFT, 571 i915_sg_segment_size(), GFP_KERNEL); 572 if (ret) { 573 st->sgl = NULL; 574 return ERR_PTR(ret); 575 } 576 577 ret = dma_map_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 0); 578 if (ret) { 579 sg_free_table(st); 580 return ERR_PTR(ret); 581 } 582 583 return i915_refct_sgt_get(&i915_tt->cached_rsgt); 584 } 585 586 /** 587 * i915_ttm_resource_get_st - Get a refcounted sg-table pointing to the 588 * resource memory 589 * @obj: The GEM object used for sg-table caching 590 * @res: The struct ttm_resource for which an sg-table is requested. 591 * 592 * This function returns a refcounted sg-table representing the memory 593 * pointed to by @res. If @res is the object's current resource it may also 594 * cache the sg_table on the object or attempt to access an already cached 595 * sg-table. The refcounted sg-table needs to be put when no-longer in use. 596 * 597 * Return: A valid pointer to a struct i915_refct_sgt or error pointer on 598 * failure. 599 */ 600 struct i915_refct_sgt * 601 i915_ttm_resource_get_st(struct drm_i915_gem_object *obj, 602 struct ttm_resource *res) 603 { 604 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 605 u32 page_alignment; 606 607 if (!i915_ttm_gtt_binds_lmem(res)) 608 return i915_ttm_tt_get_st(bo->ttm); 609 610 page_alignment = bo->page_alignment << PAGE_SHIFT; 611 if (!page_alignment) 612 page_alignment = obj->mm.region->min_page_size; 613 614 /* 615 * If CPU mapping differs, we need to add the ttm_tt pages to 616 * the resulting st. Might make sense for GGTT. 617 */ 618 GEM_WARN_ON(!i915_ttm_cpu_maps_iomem(res)); 619 if (bo->resource == res) { 620 if (!obj->ttm.cached_io_rsgt) { 621 struct i915_refct_sgt *rsgt; 622 623 rsgt = intel_region_ttm_resource_to_rsgt(obj->mm.region, 624 res, 625 page_alignment); 626 if (IS_ERR(rsgt)) 627 return rsgt; 628 629 obj->ttm.cached_io_rsgt = rsgt; 630 } 631 return i915_refct_sgt_get(obj->ttm.cached_io_rsgt); 632 } 633 634 return intel_region_ttm_resource_to_rsgt(obj->mm.region, res, 635 page_alignment); 636 } 637 638 static int i915_ttm_truncate(struct drm_i915_gem_object *obj) 639 { 640 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 641 int err; 642 643 WARN_ON_ONCE(obj->mm.madv == I915_MADV_WILLNEED); 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 (!obj) 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 == bman_res->base.num_pages; 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 (!obj) 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 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 719 struct scatterlist *sg; 720 unsigned long base; 721 unsigned int ofs; 722 723 GEM_BUG_ON(!obj); 724 GEM_WARN_ON(bo->ttm); 725 726 base = obj->mm.region->iomap.base - obj->mm.region->region.start; 727 sg = __i915_gem_object_get_sg(obj, &obj->ttm.get_io_page, page_offset, &ofs, true); 728 729 return ((base + sg_dma_address(sg)) >> PAGE_SHIFT) + ofs; 730 } 731 732 /* 733 * All callbacks need to take care not to downcast a struct ttm_buffer_object 734 * without checking its subclass, since it might be a TTM ghost object. 735 */ 736 static struct ttm_device_funcs i915_ttm_bo_driver = { 737 .ttm_tt_create = i915_ttm_tt_create, 738 .ttm_tt_populate = i915_ttm_tt_populate, 739 .ttm_tt_unpopulate = i915_ttm_tt_unpopulate, 740 .ttm_tt_destroy = i915_ttm_tt_destroy, 741 .eviction_valuable = i915_ttm_eviction_valuable, 742 .evict_flags = i915_ttm_evict_flags, 743 .move = i915_ttm_move, 744 .swap_notify = i915_ttm_swap_notify, 745 .delete_mem_notify = i915_ttm_delete_mem_notify, 746 .io_mem_reserve = i915_ttm_io_mem_reserve, 747 .io_mem_pfn = i915_ttm_io_mem_pfn, 748 }; 749 750 /** 751 * i915_ttm_driver - Return a pointer to the TTM device funcs 752 * 753 * Return: Pointer to statically allocated TTM device funcs. 754 */ 755 struct ttm_device_funcs *i915_ttm_driver(void) 756 { 757 return &i915_ttm_bo_driver; 758 } 759 760 static int __i915_ttm_get_pages(struct drm_i915_gem_object *obj, 761 struct ttm_placement *placement) 762 { 763 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 764 struct ttm_operation_ctx ctx = { 765 .interruptible = true, 766 .no_wait_gpu = false, 767 }; 768 int real_num_busy; 769 int ret; 770 771 /* First try only the requested placement. No eviction. */ 772 real_num_busy = fetch_and_zero(&placement->num_busy_placement); 773 ret = ttm_bo_validate(bo, placement, &ctx); 774 if (ret) { 775 ret = i915_ttm_err_to_gem(ret); 776 /* 777 * Anything that wants to restart the operation gets to 778 * do that. 779 */ 780 if (ret == -EDEADLK || ret == -EINTR || ret == -ERESTARTSYS || 781 ret == -EAGAIN) 782 return ret; 783 784 /* 785 * If the initial attempt fails, allow all accepted placements, 786 * evicting if necessary. 787 */ 788 placement->num_busy_placement = real_num_busy; 789 ret = ttm_bo_validate(bo, placement, &ctx); 790 if (ret) 791 return i915_ttm_err_to_gem(ret); 792 } 793 794 if (bo->ttm && !ttm_tt_is_populated(bo->ttm)) { 795 ret = ttm_tt_populate(bo->bdev, bo->ttm, &ctx); 796 if (ret) 797 return ret; 798 799 i915_ttm_adjust_domains_after_move(obj); 800 i915_ttm_adjust_gem_after_move(obj); 801 } 802 803 if (!i915_gem_object_has_pages(obj)) { 804 struct i915_refct_sgt *rsgt = 805 i915_ttm_resource_get_st(obj, bo->resource); 806 807 if (IS_ERR(rsgt)) 808 return PTR_ERR(rsgt); 809 810 GEM_BUG_ON(obj->mm.rsgt); 811 obj->mm.rsgt = rsgt; 812 __i915_gem_object_set_pages(obj, &rsgt->table, 813 i915_sg_dma_sizes(rsgt->table.sgl)); 814 } 815 816 GEM_BUG_ON(bo->ttm && ((obj->base.size >> PAGE_SHIFT) < bo->ttm->num_pages)); 817 i915_ttm_adjust_lru(obj); 818 return ret; 819 } 820 821 static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) 822 { 823 struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS]; 824 struct ttm_placement placement; 825 826 GEM_BUG_ON(obj->mm.n_placements > I915_TTM_MAX_PLACEMENTS); 827 828 /* Move to the requested placement. */ 829 i915_ttm_placement_from_obj(obj, &requested, busy, &placement); 830 831 return __i915_ttm_get_pages(obj, &placement); 832 } 833 834 /** 835 * DOC: Migration vs eviction 836 * 837 * GEM migration may not be the same as TTM migration / eviction. If 838 * the TTM core decides to evict an object it may be evicted to a 839 * TTM memory type that is not in the object's allowable GEM regions, or 840 * in fact theoretically to a TTM memory type that doesn't correspond to 841 * a GEM memory region. In that case the object's GEM region is not 842 * updated, and the data is migrated back to the GEM region at 843 * get_pages time. TTM may however set up CPU ptes to the object even 844 * when it is evicted. 845 * Gem forced migration using the i915_ttm_migrate() op, is allowed even 846 * to regions that are not in the object's list of allowable placements. 847 */ 848 static int __i915_ttm_migrate(struct drm_i915_gem_object *obj, 849 struct intel_memory_region *mr, 850 unsigned int flags) 851 { 852 struct ttm_place requested; 853 struct ttm_placement placement; 854 int ret; 855 856 i915_ttm_place_from_region(mr, &requested, obj->bo_offset, 857 obj->base.size, flags); 858 placement.num_placement = 1; 859 placement.num_busy_placement = 1; 860 placement.placement = &requested; 861 placement.busy_placement = &requested; 862 863 ret = __i915_ttm_get_pages(obj, &placement); 864 if (ret) 865 return ret; 866 867 /* 868 * Reinitialize the region bindings. This is primarily 869 * required for objects where the new region is not in 870 * its allowable placements. 871 */ 872 if (obj->mm.region != mr) { 873 i915_gem_object_release_memory_region(obj); 874 i915_gem_object_init_memory_region(obj, mr); 875 } 876 877 return 0; 878 } 879 880 static int i915_ttm_migrate(struct drm_i915_gem_object *obj, 881 struct intel_memory_region *mr) 882 { 883 return __i915_ttm_migrate(obj, mr, obj->flags); 884 } 885 886 static void i915_ttm_put_pages(struct drm_i915_gem_object *obj, 887 struct sg_table *st) 888 { 889 /* 890 * We're currently not called from a shrinker, so put_pages() 891 * typically means the object is about to destroyed, or called 892 * from move_notify(). So just avoid doing much for now. 893 * If the object is not destroyed next, The TTM eviction logic 894 * and shrinkers will move it out if needed. 895 */ 896 897 if (obj->mm.rsgt) 898 i915_refct_sgt_put(fetch_and_zero(&obj->mm.rsgt)); 899 } 900 901 /** 902 * i915_ttm_adjust_lru - Adjust an object's position on relevant LRU lists. 903 * @obj: The object 904 */ 905 void i915_ttm_adjust_lru(struct drm_i915_gem_object *obj) 906 { 907 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 908 struct i915_ttm_tt *i915_tt = 909 container_of(bo->ttm, typeof(*i915_tt), ttm); 910 bool shrinkable = 911 bo->ttm && i915_tt->filp && ttm_tt_is_populated(bo->ttm); 912 913 /* 914 * Don't manipulate the TTM LRUs while in TTM bo destruction. 915 * We're called through i915_ttm_delete_mem_notify(). 916 */ 917 if (!kref_read(&bo->kref)) 918 return; 919 920 /* 921 * We skip managing the shrinker LRU in set_pages() and just manage 922 * everything here. This does at least solve the issue with having 923 * temporary shmem mappings(like with evicted lmem) not being visible to 924 * the shrinker. Only our shmem objects are shrinkable, everything else 925 * we keep as unshrinkable. 926 * 927 * To make sure everything plays nice we keep an extra shrink pin in TTM 928 * if the underlying pages are not currently shrinkable. Once we release 929 * our pin, like when the pages are moved to shmem, the pages will then 930 * be added to the shrinker LRU, assuming the caller isn't also holding 931 * a pin. 932 * 933 * TODO: consider maybe also bumping the shrinker list here when we have 934 * already unpinned it, which should give us something more like an LRU. 935 * 936 * TODO: There is a small window of opportunity for this function to 937 * get called from eviction after we've dropped the last GEM refcount, 938 * but before the TTM deleted flag is set on the object. Avoid 939 * adjusting the shrinker list in such cases, since the object is 940 * not available to the shrinker anyway due to its zero refcount. 941 * To fix this properly we should move to a TTM shrinker LRU list for 942 * these objects. 943 */ 944 if (kref_get_unless_zero(&obj->base.refcount)) { 945 if (shrinkable != obj->mm.ttm_shrinkable) { 946 if (shrinkable) { 947 if (obj->mm.madv == I915_MADV_WILLNEED) 948 __i915_gem_object_make_shrinkable(obj); 949 else 950 __i915_gem_object_make_purgeable(obj); 951 } else { 952 i915_gem_object_make_unshrinkable(obj); 953 } 954 955 obj->mm.ttm_shrinkable = shrinkable; 956 } 957 i915_gem_object_put(obj); 958 } 959 960 /* 961 * Put on the correct LRU list depending on the MADV status 962 */ 963 spin_lock(&bo->bdev->lru_lock); 964 if (shrinkable) { 965 /* Try to keep shmem_tt from being considered for shrinking. */ 966 bo->priority = TTM_MAX_BO_PRIORITY - 1; 967 } else if (obj->mm.madv != I915_MADV_WILLNEED) { 968 bo->priority = I915_TTM_PRIO_PURGE; 969 } else if (!i915_gem_object_has_pages(obj)) { 970 bo->priority = I915_TTM_PRIO_NO_PAGES; 971 } else { 972 struct ttm_resource_manager *man = 973 ttm_manager_type(bo->bdev, bo->resource->mem_type); 974 975 /* 976 * If we need to place an LMEM resource which doesn't need CPU 977 * access then we should try not to victimize mappable objects 978 * first, since we likely end up stealing more of the mappable 979 * portion. And likewise when we try to find space for a mappble 980 * object, we know not to ever victimize objects that don't 981 * occupy any mappable pages. 982 */ 983 if (i915_ttm_cpu_maps_iomem(bo->resource) && 984 i915_ttm_buddy_man_visible_size(man) < man->size && 985 !(obj->flags & I915_BO_ALLOC_GPU_ONLY)) 986 bo->priority = I915_TTM_PRIO_NEEDS_CPU_ACCESS; 987 else 988 bo->priority = I915_TTM_PRIO_HAS_PAGES; 989 } 990 991 ttm_bo_move_to_lru_tail(bo); 992 spin_unlock(&bo->bdev->lru_lock); 993 } 994 995 /* 996 * TTM-backed gem object destruction requires some clarification. 997 * Basically we have two possibilities here. We can either rely on the 998 * i915 delayed destruction and put the TTM object when the object 999 * is idle. This would be detected by TTM which would bypass the 1000 * TTM delayed destroy handling. The other approach is to put the TTM 1001 * object early and rely on the TTM destroyed handling, and then free 1002 * the leftover parts of the GEM object once TTM's destroyed list handling is 1003 * complete. For now, we rely on the latter for two reasons: 1004 * a) TTM can evict an object even when it's on the delayed destroy list, 1005 * which in theory allows for complete eviction. 1006 * b) There is work going on in TTM to allow freeing an object even when 1007 * it's not idle, and using the TTM destroyed list handling could help us 1008 * benefit from that. 1009 */ 1010 static void i915_ttm_delayed_free(struct drm_i915_gem_object *obj) 1011 { 1012 GEM_BUG_ON(!obj->ttm.created); 1013 1014 ttm_bo_put(i915_gem_to_ttm(obj)); 1015 } 1016 1017 static vm_fault_t vm_fault_ttm(struct vm_fault *vmf) 1018 { 1019 struct vm_area_struct *area = vmf->vma; 1020 struct ttm_buffer_object *bo = area->vm_private_data; 1021 struct drm_device *dev = bo->base.dev; 1022 struct drm_i915_gem_object *obj; 1023 vm_fault_t ret; 1024 int idx; 1025 1026 obj = i915_ttm_to_gem(bo); 1027 if (!obj) 1028 return VM_FAULT_SIGBUS; 1029 1030 /* Sanity check that we allow writing into this object */ 1031 if (unlikely(i915_gem_object_is_readonly(obj) && 1032 area->vm_flags & VM_WRITE)) 1033 return VM_FAULT_SIGBUS; 1034 1035 ret = ttm_bo_vm_reserve(bo, vmf); 1036 if (ret) 1037 return ret; 1038 1039 if (obj->mm.madv != I915_MADV_WILLNEED) { 1040 dma_resv_unlock(bo->base.resv); 1041 return VM_FAULT_SIGBUS; 1042 } 1043 1044 if (!i915_ttm_resource_mappable(bo->resource)) { 1045 int err = -ENODEV; 1046 int i; 1047 1048 for (i = 0; i < obj->mm.n_placements; i++) { 1049 struct intel_memory_region *mr = obj->mm.placements[i]; 1050 unsigned int flags; 1051 1052 if (!mr->io_size && mr->type != INTEL_MEMORY_SYSTEM) 1053 continue; 1054 1055 flags = obj->flags; 1056 flags &= ~I915_BO_ALLOC_GPU_ONLY; 1057 err = __i915_ttm_migrate(obj, mr, flags); 1058 if (!err) 1059 break; 1060 } 1061 1062 if (err) { 1063 drm_dbg(dev, "Unable to make resource CPU accessible\n"); 1064 dma_resv_unlock(bo->base.resv); 1065 return VM_FAULT_SIGBUS; 1066 } 1067 } 1068 1069 if (drm_dev_enter(dev, &idx)) { 1070 ret = ttm_bo_vm_fault_reserved(vmf, vmf->vma->vm_page_prot, 1071 TTM_BO_VM_NUM_PREFAULT); 1072 drm_dev_exit(idx); 1073 } else { 1074 ret = ttm_bo_vm_dummy_page(vmf, vmf->vma->vm_page_prot); 1075 } 1076 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) 1077 return ret; 1078 1079 i915_ttm_adjust_lru(obj); 1080 1081 dma_resv_unlock(bo->base.resv); 1082 return ret; 1083 } 1084 1085 static int 1086 vm_access_ttm(struct vm_area_struct *area, unsigned long addr, 1087 void *buf, int len, int write) 1088 { 1089 struct drm_i915_gem_object *obj = 1090 i915_ttm_to_gem(area->vm_private_data); 1091 1092 if (i915_gem_object_is_readonly(obj) && write) 1093 return -EACCES; 1094 1095 return ttm_bo_vm_access(area, addr, buf, len, write); 1096 } 1097 1098 static void ttm_vm_open(struct vm_area_struct *vma) 1099 { 1100 struct drm_i915_gem_object *obj = 1101 i915_ttm_to_gem(vma->vm_private_data); 1102 1103 GEM_BUG_ON(!obj); 1104 i915_gem_object_get(obj); 1105 } 1106 1107 static void ttm_vm_close(struct vm_area_struct *vma) 1108 { 1109 struct drm_i915_gem_object *obj = 1110 i915_ttm_to_gem(vma->vm_private_data); 1111 1112 GEM_BUG_ON(!obj); 1113 i915_gem_object_put(obj); 1114 } 1115 1116 static const struct vm_operations_struct vm_ops_ttm = { 1117 .fault = vm_fault_ttm, 1118 .access = vm_access_ttm, 1119 .open = ttm_vm_open, 1120 .close = ttm_vm_close, 1121 }; 1122 1123 static u64 i915_ttm_mmap_offset(struct drm_i915_gem_object *obj) 1124 { 1125 /* The ttm_bo must be allocated with I915_BO_ALLOC_USER */ 1126 GEM_BUG_ON(!drm_mm_node_allocated(&obj->base.vma_node.vm_node)); 1127 1128 return drm_vma_node_offset_addr(&obj->base.vma_node); 1129 } 1130 1131 static void i915_ttm_unmap_virtual(struct drm_i915_gem_object *obj) 1132 { 1133 ttm_bo_unmap_virtual(i915_gem_to_ttm(obj)); 1134 } 1135 1136 static const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = { 1137 .name = "i915_gem_object_ttm", 1138 .flags = I915_GEM_OBJECT_IS_SHRINKABLE | 1139 I915_GEM_OBJECT_SELF_MANAGED_SHRINK_LIST, 1140 1141 .get_pages = i915_ttm_get_pages, 1142 .put_pages = i915_ttm_put_pages, 1143 .truncate = i915_ttm_truncate, 1144 .shrink = i915_ttm_shrink, 1145 1146 .adjust_lru = i915_ttm_adjust_lru, 1147 .delayed_free = i915_ttm_delayed_free, 1148 .migrate = i915_ttm_migrate, 1149 1150 .mmap_offset = i915_ttm_mmap_offset, 1151 .unmap_virtual = i915_ttm_unmap_virtual, 1152 .mmap_ops = &vm_ops_ttm, 1153 }; 1154 1155 void i915_ttm_bo_destroy(struct ttm_buffer_object *bo) 1156 { 1157 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 1158 1159 i915_gem_object_release_memory_region(obj); 1160 mutex_destroy(&obj->ttm.get_io_page.lock); 1161 1162 if (obj->ttm.created) { 1163 /* 1164 * We freely manage the shrinker LRU outide of the mm.pages life 1165 * cycle. As a result when destroying the object we should be 1166 * extra paranoid and ensure we remove it from the LRU, before 1167 * we free the object. 1168 * 1169 * Touching the ttm_shrinkable outside of the object lock here 1170 * should be safe now that the last GEM object ref was dropped. 1171 */ 1172 if (obj->mm.ttm_shrinkable) 1173 i915_gem_object_make_unshrinkable(obj); 1174 1175 i915_ttm_backup_free(obj); 1176 1177 /* This releases all gem object bindings to the backend. */ 1178 __i915_gem_free_object(obj); 1179 1180 call_rcu(&obj->rcu, __i915_gem_free_object_rcu); 1181 } else { 1182 __i915_gem_object_fini(obj); 1183 } 1184 } 1185 1186 /** 1187 * __i915_gem_ttm_object_init - Initialize a ttm-backed i915 gem object 1188 * @mem: The initial memory region for the object. 1189 * @obj: The gem object. 1190 * @size: Object size in bytes. 1191 * @flags: gem object flags. 1192 * 1193 * Return: 0 on success, negative error code on failure. 1194 */ 1195 int __i915_gem_ttm_object_init(struct intel_memory_region *mem, 1196 struct drm_i915_gem_object *obj, 1197 resource_size_t offset, 1198 resource_size_t size, 1199 resource_size_t page_size, 1200 unsigned int flags) 1201 { 1202 static struct lock_class_key lock_class; 1203 struct drm_i915_private *i915 = mem->i915; 1204 struct ttm_operation_ctx ctx = { 1205 .interruptible = true, 1206 .no_wait_gpu = false, 1207 }; 1208 enum ttm_bo_type bo_type; 1209 int ret; 1210 1211 drm_gem_private_object_init(&i915->drm, &obj->base, size); 1212 i915_gem_object_init(obj, &i915_gem_ttm_obj_ops, &lock_class, flags); 1213 1214 obj->bo_offset = offset; 1215 1216 /* Don't put on a region list until we're either locked or fully initialized. */ 1217 obj->mm.region = mem; 1218 INIT_LIST_HEAD(&obj->mm.region_link); 1219 1220 INIT_RADIX_TREE(&obj->ttm.get_io_page.radix, GFP_KERNEL | __GFP_NOWARN); 1221 mutex_init(&obj->ttm.get_io_page.lock); 1222 bo_type = (obj->flags & I915_BO_ALLOC_USER) ? ttm_bo_type_device : 1223 ttm_bo_type_kernel; 1224 1225 obj->base.vma_node.driver_private = i915_gem_to_ttm(obj); 1226 1227 /* Forcing the page size is kernel internal only */ 1228 GEM_BUG_ON(page_size && obj->mm.n_placements); 1229 1230 /* 1231 * Keep an extra shrink pin to prevent the object from being made 1232 * shrinkable too early. If the ttm_tt is ever allocated in shmem, we 1233 * drop the pin. The TTM backend manages the shrinker LRU itself, 1234 * outside of the normal mm.pages life cycle. 1235 */ 1236 i915_gem_object_make_unshrinkable(obj); 1237 1238 /* 1239 * If this function fails, it will call the destructor, but 1240 * our caller still owns the object. So no freeing in the 1241 * destructor until obj->ttm.created is true. 1242 * Similarly, in delayed_destroy, we can't call ttm_bo_put() 1243 * until successful initialization. 1244 */ 1245 ret = ttm_bo_init_reserved(&i915->bdev, i915_gem_to_ttm(obj), size, 1246 bo_type, &i915_sys_placement, 1247 page_size >> PAGE_SHIFT, 1248 &ctx, NULL, NULL, i915_ttm_bo_destroy); 1249 if (ret) 1250 return i915_ttm_err_to_gem(ret); 1251 1252 obj->ttm.created = true; 1253 i915_gem_object_release_memory_region(obj); 1254 i915_gem_object_init_memory_region(obj, mem); 1255 i915_ttm_adjust_domains_after_move(obj); 1256 i915_ttm_adjust_gem_after_move(obj); 1257 i915_gem_object_unlock(obj); 1258 1259 return 0; 1260 } 1261 1262 static const struct intel_memory_region_ops ttm_system_region_ops = { 1263 .init_object = __i915_gem_ttm_object_init, 1264 .release = intel_region_ttm_fini, 1265 }; 1266 1267 struct intel_memory_region * 1268 i915_gem_ttm_system_setup(struct drm_i915_private *i915, 1269 u16 type, u16 instance) 1270 { 1271 struct intel_memory_region *mr; 1272 1273 mr = intel_memory_region_create(i915, 0, 1274 totalram_pages() << PAGE_SHIFT, 1275 PAGE_SIZE, 0, 0, 1276 type, instance, 1277 &ttm_system_region_ops); 1278 if (IS_ERR(mr)) 1279 return mr; 1280 1281 intel_memory_region_set_name(mr, "system-ttm"); 1282 return mr; 1283 } 1284