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 inline bool 270 i915_gem_object_needs_ccs_pages(struct drm_i915_gem_object *obj) 271 { 272 bool lmem_placement = false; 273 int i; 274 275 for (i = 0; i < obj->mm.n_placements; i++) { 276 /* Compression is not allowed for the objects with smem placement */ 277 if (obj->mm.placements[i]->type == INTEL_MEMORY_SYSTEM) 278 return false; 279 if (!lmem_placement && 280 obj->mm.placements[i]->type == INTEL_MEMORY_LOCAL) 281 lmem_placement = true; 282 } 283 284 return lmem_placement; 285 } 286 287 static struct ttm_tt *i915_ttm_tt_create(struct ttm_buffer_object *bo, 288 uint32_t page_flags) 289 { 290 struct drm_i915_private *i915 = container_of(bo->bdev, typeof(*i915), 291 bdev); 292 struct ttm_resource_manager *man = 293 ttm_manager_type(bo->bdev, bo->resource->mem_type); 294 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 295 unsigned long ccs_pages = 0; 296 enum ttm_caching caching; 297 struct i915_ttm_tt *i915_tt; 298 int ret; 299 300 if (!obj) 301 return NULL; 302 303 i915_tt = kzalloc(sizeof(*i915_tt), GFP_KERNEL); 304 if (!i915_tt) 305 return NULL; 306 307 if (obj->flags & I915_BO_ALLOC_CPU_CLEAR && 308 man->use_tt) 309 page_flags |= TTM_TT_FLAG_ZERO_ALLOC; 310 311 caching = i915_ttm_select_tt_caching(obj); 312 if (i915_gem_object_is_shrinkable(obj) && caching == ttm_cached) { 313 page_flags |= TTM_TT_FLAG_EXTERNAL | 314 TTM_TT_FLAG_EXTERNAL_MAPPABLE; 315 i915_tt->is_shmem = true; 316 } 317 318 if (HAS_FLAT_CCS(i915) && i915_gem_object_needs_ccs_pages(obj)) 319 ccs_pages = DIV_ROUND_UP(DIV_ROUND_UP(bo->base.size, 320 NUM_BYTES_PER_CCS_BYTE), 321 PAGE_SIZE); 322 323 ret = ttm_tt_init(&i915_tt->ttm, bo, page_flags, caching, ccs_pages); 324 if (ret) 325 goto err_free; 326 327 __i915_refct_sgt_init(&i915_tt->cached_rsgt, bo->base.size, 328 &tt_rsgt_ops); 329 330 i915_tt->dev = obj->base.dev->dev; 331 332 return &i915_tt->ttm; 333 334 err_free: 335 kfree(i915_tt); 336 return NULL; 337 } 338 339 static int i915_ttm_tt_populate(struct ttm_device *bdev, 340 struct ttm_tt *ttm, 341 struct ttm_operation_ctx *ctx) 342 { 343 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 344 345 if (i915_tt->is_shmem) 346 return i915_ttm_tt_shmem_populate(bdev, ttm, ctx); 347 348 return ttm_pool_alloc(&bdev->pool, ttm, ctx); 349 } 350 351 static void i915_ttm_tt_unpopulate(struct ttm_device *bdev, struct ttm_tt *ttm) 352 { 353 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 354 struct sg_table *st = &i915_tt->cached_rsgt.table; 355 356 if (st->sgl) 357 dma_unmap_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 0); 358 359 if (i915_tt->is_shmem) { 360 i915_ttm_tt_shmem_unpopulate(ttm); 361 } else { 362 sg_free_table(st); 363 ttm_pool_free(&bdev->pool, ttm); 364 } 365 } 366 367 static void i915_ttm_tt_destroy(struct ttm_device *bdev, struct ttm_tt *ttm) 368 { 369 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 370 371 if (i915_tt->filp) 372 fput(i915_tt->filp); 373 374 ttm_tt_fini(ttm); 375 i915_refct_sgt_put(&i915_tt->cached_rsgt); 376 } 377 378 static bool i915_ttm_eviction_valuable(struct ttm_buffer_object *bo, 379 const struct ttm_place *place) 380 { 381 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 382 struct ttm_resource *res = bo->resource; 383 384 if (!obj) 385 return false; 386 387 /* 388 * EXTERNAL objects should never be swapped out by TTM, instead we need 389 * to handle that ourselves. TTM will already skip such objects for us, 390 * but we would like to avoid grabbing locks for no good reason. 391 */ 392 if (bo->ttm && bo->ttm->page_flags & TTM_TT_FLAG_EXTERNAL) 393 return false; 394 395 /* Will do for now. Our pinned objects are still on TTM's LRU lists */ 396 if (!i915_gem_object_evictable(obj)) 397 return false; 398 399 switch (res->mem_type) { 400 case I915_PL_LMEM0: { 401 struct ttm_resource_manager *man = 402 ttm_manager_type(bo->bdev, res->mem_type); 403 struct i915_ttm_buddy_resource *bman_res = 404 to_ttm_buddy_resource(res); 405 struct drm_buddy *mm = bman_res->mm; 406 struct drm_buddy_block *block; 407 408 if (!place->fpfn && !place->lpfn) 409 return true; 410 411 GEM_BUG_ON(!place->lpfn); 412 413 /* 414 * If we just want something mappable then we can quickly check 415 * if the current victim resource is using any of the CPU 416 * visible portion. 417 */ 418 if (!place->fpfn && 419 place->lpfn == i915_ttm_buddy_man_visible_size(man)) 420 return bman_res->used_visible_size > 0; 421 422 /* Real range allocation */ 423 list_for_each_entry(block, &bman_res->blocks, link) { 424 unsigned long fpfn = 425 drm_buddy_block_offset(block) >> PAGE_SHIFT; 426 unsigned long lpfn = fpfn + 427 (drm_buddy_block_size(mm, block) >> PAGE_SHIFT); 428 429 if (place->fpfn < lpfn && place->lpfn > fpfn) 430 return true; 431 } 432 return false; 433 } default: 434 break; 435 } 436 437 return true; 438 } 439 440 static void i915_ttm_evict_flags(struct ttm_buffer_object *bo, 441 struct ttm_placement *placement) 442 { 443 *placement = i915_sys_placement; 444 } 445 446 /** 447 * i915_ttm_free_cached_io_rsgt - Free object cached LMEM information 448 * @obj: The GEM object 449 * This function frees any LMEM-related information that is cached on 450 * the object. For example the radix tree for fast page lookup and the 451 * cached refcounted sg-table 452 */ 453 void i915_ttm_free_cached_io_rsgt(struct drm_i915_gem_object *obj) 454 { 455 struct radix_tree_iter iter; 456 void __rcu **slot; 457 458 if (!obj->ttm.cached_io_rsgt) 459 return; 460 461 rcu_read_lock(); 462 radix_tree_for_each_slot(slot, &obj->ttm.get_io_page.radix, &iter, 0) 463 radix_tree_delete(&obj->ttm.get_io_page.radix, iter.index); 464 rcu_read_unlock(); 465 466 i915_refct_sgt_put(obj->ttm.cached_io_rsgt); 467 obj->ttm.cached_io_rsgt = NULL; 468 } 469 470 /** 471 * i915_ttm_purge - Clear an object of its memory 472 * @obj: The object 473 * 474 * This function is called to clear an object of it's memory when it is 475 * marked as not needed anymore. 476 * 477 * Return: 0 on success, negative error code on failure. 478 */ 479 int i915_ttm_purge(struct drm_i915_gem_object *obj) 480 { 481 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 482 struct i915_ttm_tt *i915_tt = 483 container_of(bo->ttm, typeof(*i915_tt), ttm); 484 struct ttm_operation_ctx ctx = { 485 .interruptible = true, 486 .no_wait_gpu = false, 487 }; 488 struct ttm_placement place = {}; 489 int ret; 490 491 if (obj->mm.madv == __I915_MADV_PURGED) 492 return 0; 493 494 ret = ttm_bo_validate(bo, &place, &ctx); 495 if (ret) 496 return ret; 497 498 if (bo->ttm && i915_tt->filp) { 499 /* 500 * The below fput(which eventually calls shmem_truncate) might 501 * be delayed by worker, so when directly called to purge the 502 * pages(like by the shrinker) we should try to be more 503 * aggressive and release the pages immediately. 504 */ 505 shmem_truncate_range(file_inode(i915_tt->filp), 506 0, (loff_t)-1); 507 fput(fetch_and_zero(&i915_tt->filp)); 508 } 509 510 obj->write_domain = 0; 511 obj->read_domains = 0; 512 i915_ttm_adjust_gem_after_move(obj); 513 i915_ttm_free_cached_io_rsgt(obj); 514 obj->mm.madv = __I915_MADV_PURGED; 515 516 return 0; 517 } 518 519 static int i915_ttm_shrink(struct drm_i915_gem_object *obj, unsigned int flags) 520 { 521 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 522 struct i915_ttm_tt *i915_tt = 523 container_of(bo->ttm, typeof(*i915_tt), ttm); 524 struct ttm_operation_ctx ctx = { 525 .interruptible = true, 526 .no_wait_gpu = flags & I915_GEM_OBJECT_SHRINK_NO_GPU_WAIT, 527 }; 528 struct ttm_placement place = {}; 529 int ret; 530 531 if (!bo->ttm || bo->resource->mem_type != TTM_PL_SYSTEM) 532 return 0; 533 534 GEM_BUG_ON(!i915_tt->is_shmem); 535 536 if (!i915_tt->filp) 537 return 0; 538 539 ret = ttm_bo_wait_ctx(bo, &ctx); 540 if (ret) 541 return ret; 542 543 switch (obj->mm.madv) { 544 case I915_MADV_DONTNEED: 545 return i915_ttm_purge(obj); 546 case __I915_MADV_PURGED: 547 return 0; 548 } 549 550 if (bo->ttm->page_flags & TTM_TT_FLAG_SWAPPED) 551 return 0; 552 553 bo->ttm->page_flags |= TTM_TT_FLAG_SWAPPED; 554 ret = ttm_bo_validate(bo, &place, &ctx); 555 if (ret) { 556 bo->ttm->page_flags &= ~TTM_TT_FLAG_SWAPPED; 557 return ret; 558 } 559 560 if (flags & I915_GEM_OBJECT_SHRINK_WRITEBACK) 561 __shmem_writeback(obj->base.size, i915_tt->filp->f_mapping); 562 563 return 0; 564 } 565 566 static void i915_ttm_delete_mem_notify(struct ttm_buffer_object *bo) 567 { 568 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 569 570 if (likely(obj)) { 571 __i915_gem_object_pages_fini(obj); 572 i915_ttm_free_cached_io_rsgt(obj); 573 } 574 } 575 576 static struct i915_refct_sgt *i915_ttm_tt_get_st(struct ttm_tt *ttm) 577 { 578 struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm); 579 struct sg_table *st; 580 int ret; 581 582 if (i915_tt->cached_rsgt.table.sgl) 583 return i915_refct_sgt_get(&i915_tt->cached_rsgt); 584 585 st = &i915_tt->cached_rsgt.table; 586 ret = sg_alloc_table_from_pages_segment(st, 587 ttm->pages, ttm->num_pages, 588 0, (unsigned long)ttm->num_pages << PAGE_SHIFT, 589 i915_sg_segment_size(), GFP_KERNEL); 590 if (ret) { 591 st->sgl = NULL; 592 return ERR_PTR(ret); 593 } 594 595 ret = dma_map_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 0); 596 if (ret) { 597 sg_free_table(st); 598 return ERR_PTR(ret); 599 } 600 601 return i915_refct_sgt_get(&i915_tt->cached_rsgt); 602 } 603 604 /** 605 * i915_ttm_resource_get_st - Get a refcounted sg-table pointing to the 606 * resource memory 607 * @obj: The GEM object used for sg-table caching 608 * @res: The struct ttm_resource for which an sg-table is requested. 609 * 610 * This function returns a refcounted sg-table representing the memory 611 * pointed to by @res. If @res is the object's current resource it may also 612 * cache the sg_table on the object or attempt to access an already cached 613 * sg-table. The refcounted sg-table needs to be put when no-longer in use. 614 * 615 * Return: A valid pointer to a struct i915_refct_sgt or error pointer on 616 * failure. 617 */ 618 struct i915_refct_sgt * 619 i915_ttm_resource_get_st(struct drm_i915_gem_object *obj, 620 struct ttm_resource *res) 621 { 622 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 623 624 if (!i915_ttm_gtt_binds_lmem(res)) 625 return i915_ttm_tt_get_st(bo->ttm); 626 627 /* 628 * If CPU mapping differs, we need to add the ttm_tt pages to 629 * the resulting st. Might make sense for GGTT. 630 */ 631 GEM_WARN_ON(!i915_ttm_cpu_maps_iomem(res)); 632 if (bo->resource == res) { 633 if (!obj->ttm.cached_io_rsgt) { 634 struct i915_refct_sgt *rsgt; 635 636 rsgt = intel_region_ttm_resource_to_rsgt(obj->mm.region, 637 res); 638 if (IS_ERR(rsgt)) 639 return rsgt; 640 641 obj->ttm.cached_io_rsgt = rsgt; 642 } 643 return i915_refct_sgt_get(obj->ttm.cached_io_rsgt); 644 } 645 646 return intel_region_ttm_resource_to_rsgt(obj->mm.region, res); 647 } 648 649 static int i915_ttm_truncate(struct drm_i915_gem_object *obj) 650 { 651 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 652 int err; 653 654 WARN_ON_ONCE(obj->mm.madv == I915_MADV_WILLNEED); 655 656 err = i915_ttm_move_notify(bo); 657 if (err) 658 return err; 659 660 return i915_ttm_purge(obj); 661 } 662 663 static void i915_ttm_swap_notify(struct ttm_buffer_object *bo) 664 { 665 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 666 int ret; 667 668 if (!obj) 669 return; 670 671 ret = i915_ttm_move_notify(bo); 672 GEM_WARN_ON(ret); 673 GEM_WARN_ON(obj->ttm.cached_io_rsgt); 674 if (!ret && obj->mm.madv != I915_MADV_WILLNEED) 675 i915_ttm_purge(obj); 676 } 677 678 static bool i915_ttm_resource_mappable(struct ttm_resource *res) 679 { 680 struct i915_ttm_buddy_resource *bman_res = to_ttm_buddy_resource(res); 681 682 if (!i915_ttm_cpu_maps_iomem(res)) 683 return true; 684 685 return bman_res->used_visible_size == bman_res->base.num_pages; 686 } 687 688 static int i915_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *mem) 689 { 690 if (!i915_ttm_cpu_maps_iomem(mem)) 691 return 0; 692 693 if (!i915_ttm_resource_mappable(mem)) 694 return -EINVAL; 695 696 mem->bus.caching = ttm_write_combined; 697 mem->bus.is_iomem = true; 698 699 return 0; 700 } 701 702 static unsigned long i915_ttm_io_mem_pfn(struct ttm_buffer_object *bo, 703 unsigned long page_offset) 704 { 705 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 706 struct scatterlist *sg; 707 unsigned long base; 708 unsigned int ofs; 709 710 GEM_BUG_ON(!obj); 711 GEM_WARN_ON(bo->ttm); 712 713 base = obj->mm.region->iomap.base - obj->mm.region->region.start; 714 sg = __i915_gem_object_get_sg(obj, &obj->ttm.get_io_page, page_offset, &ofs, true); 715 716 return ((base + sg_dma_address(sg)) >> PAGE_SHIFT) + ofs; 717 } 718 719 /* 720 * All callbacks need to take care not to downcast a struct ttm_buffer_object 721 * without checking its subclass, since it might be a TTM ghost object. 722 */ 723 static struct ttm_device_funcs i915_ttm_bo_driver = { 724 .ttm_tt_create = i915_ttm_tt_create, 725 .ttm_tt_populate = i915_ttm_tt_populate, 726 .ttm_tt_unpopulate = i915_ttm_tt_unpopulate, 727 .ttm_tt_destroy = i915_ttm_tt_destroy, 728 .eviction_valuable = i915_ttm_eviction_valuable, 729 .evict_flags = i915_ttm_evict_flags, 730 .move = i915_ttm_move, 731 .swap_notify = i915_ttm_swap_notify, 732 .delete_mem_notify = i915_ttm_delete_mem_notify, 733 .io_mem_reserve = i915_ttm_io_mem_reserve, 734 .io_mem_pfn = i915_ttm_io_mem_pfn, 735 }; 736 737 /** 738 * i915_ttm_driver - Return a pointer to the TTM device funcs 739 * 740 * Return: Pointer to statically allocated TTM device funcs. 741 */ 742 struct ttm_device_funcs *i915_ttm_driver(void) 743 { 744 return &i915_ttm_bo_driver; 745 } 746 747 static int __i915_ttm_get_pages(struct drm_i915_gem_object *obj, 748 struct ttm_placement *placement) 749 { 750 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 751 struct ttm_operation_ctx ctx = { 752 .interruptible = true, 753 .no_wait_gpu = false, 754 }; 755 int real_num_busy; 756 int ret; 757 758 /* First try only the requested placement. No eviction. */ 759 real_num_busy = fetch_and_zero(&placement->num_busy_placement); 760 ret = ttm_bo_validate(bo, placement, &ctx); 761 if (ret) { 762 ret = i915_ttm_err_to_gem(ret); 763 /* 764 * Anything that wants to restart the operation gets to 765 * do that. 766 */ 767 if (ret == -EDEADLK || ret == -EINTR || ret == -ERESTARTSYS || 768 ret == -EAGAIN) 769 return ret; 770 771 /* 772 * If the initial attempt fails, allow all accepted placements, 773 * evicting if necessary. 774 */ 775 placement->num_busy_placement = real_num_busy; 776 ret = ttm_bo_validate(bo, placement, &ctx); 777 if (ret) 778 return i915_ttm_err_to_gem(ret); 779 } 780 781 if (bo->ttm && !ttm_tt_is_populated(bo->ttm)) { 782 ret = ttm_tt_populate(bo->bdev, bo->ttm, &ctx); 783 if (ret) 784 return ret; 785 786 i915_ttm_adjust_domains_after_move(obj); 787 i915_ttm_adjust_gem_after_move(obj); 788 } 789 790 if (!i915_gem_object_has_pages(obj)) { 791 struct i915_refct_sgt *rsgt = 792 i915_ttm_resource_get_st(obj, bo->resource); 793 794 if (IS_ERR(rsgt)) 795 return PTR_ERR(rsgt); 796 797 GEM_BUG_ON(obj->mm.rsgt); 798 obj->mm.rsgt = rsgt; 799 __i915_gem_object_set_pages(obj, &rsgt->table, 800 i915_sg_dma_sizes(rsgt->table.sgl)); 801 } 802 803 GEM_BUG_ON(bo->ttm && ((obj->base.size >> PAGE_SHIFT) < bo->ttm->num_pages)); 804 i915_ttm_adjust_lru(obj); 805 return ret; 806 } 807 808 static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) 809 { 810 struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS]; 811 struct ttm_placement placement; 812 813 GEM_BUG_ON(obj->mm.n_placements > I915_TTM_MAX_PLACEMENTS); 814 815 /* Move to the requested placement. */ 816 i915_ttm_placement_from_obj(obj, &requested, busy, &placement); 817 818 return __i915_ttm_get_pages(obj, &placement); 819 } 820 821 /** 822 * DOC: Migration vs eviction 823 * 824 * GEM migration may not be the same as TTM migration / eviction. If 825 * the TTM core decides to evict an object it may be evicted to a 826 * TTM memory type that is not in the object's allowable GEM regions, or 827 * in fact theoretically to a TTM memory type that doesn't correspond to 828 * a GEM memory region. In that case the object's GEM region is not 829 * updated, and the data is migrated back to the GEM region at 830 * get_pages time. TTM may however set up CPU ptes to the object even 831 * when it is evicted. 832 * Gem forced migration using the i915_ttm_migrate() op, is allowed even 833 * to regions that are not in the object's list of allowable placements. 834 */ 835 static int __i915_ttm_migrate(struct drm_i915_gem_object *obj, 836 struct intel_memory_region *mr, 837 unsigned int flags) 838 { 839 struct ttm_place requested; 840 struct ttm_placement placement; 841 int ret; 842 843 i915_ttm_place_from_region(mr, &requested, obj->bo_offset, 844 obj->base.size, flags); 845 placement.num_placement = 1; 846 placement.num_busy_placement = 1; 847 placement.placement = &requested; 848 placement.busy_placement = &requested; 849 850 ret = __i915_ttm_get_pages(obj, &placement); 851 if (ret) 852 return ret; 853 854 /* 855 * Reinitialize the region bindings. This is primarily 856 * required for objects where the new region is not in 857 * its allowable placements. 858 */ 859 if (obj->mm.region != mr) { 860 i915_gem_object_release_memory_region(obj); 861 i915_gem_object_init_memory_region(obj, mr); 862 } 863 864 return 0; 865 } 866 867 static int i915_ttm_migrate(struct drm_i915_gem_object *obj, 868 struct intel_memory_region *mr) 869 { 870 return __i915_ttm_migrate(obj, mr, obj->flags); 871 } 872 873 static void i915_ttm_put_pages(struct drm_i915_gem_object *obj, 874 struct sg_table *st) 875 { 876 /* 877 * We're currently not called from a shrinker, so put_pages() 878 * typically means the object is about to destroyed, or called 879 * from move_notify(). So just avoid doing much for now. 880 * If the object is not destroyed next, The TTM eviction logic 881 * and shrinkers will move it out if needed. 882 */ 883 884 if (obj->mm.rsgt) 885 i915_refct_sgt_put(fetch_and_zero(&obj->mm.rsgt)); 886 } 887 888 /** 889 * i915_ttm_adjust_lru - Adjust an object's position on relevant LRU lists. 890 * @obj: The object 891 */ 892 void i915_ttm_adjust_lru(struct drm_i915_gem_object *obj) 893 { 894 struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); 895 struct i915_ttm_tt *i915_tt = 896 container_of(bo->ttm, typeof(*i915_tt), ttm); 897 bool shrinkable = 898 bo->ttm && i915_tt->filp && ttm_tt_is_populated(bo->ttm); 899 900 /* 901 * Don't manipulate the TTM LRUs while in TTM bo destruction. 902 * We're called through i915_ttm_delete_mem_notify(). 903 */ 904 if (!kref_read(&bo->kref)) 905 return; 906 907 /* 908 * We skip managing the shrinker LRU in set_pages() and just manage 909 * everything here. This does at least solve the issue with having 910 * temporary shmem mappings(like with evicted lmem) not being visible to 911 * the shrinker. Only our shmem objects are shrinkable, everything else 912 * we keep as unshrinkable. 913 * 914 * To make sure everything plays nice we keep an extra shrink pin in TTM 915 * if the underlying pages are not currently shrinkable. Once we release 916 * our pin, like when the pages are moved to shmem, the pages will then 917 * be added to the shrinker LRU, assuming the caller isn't also holding 918 * a pin. 919 * 920 * TODO: consider maybe also bumping the shrinker list here when we have 921 * already unpinned it, which should give us something more like an LRU. 922 * 923 * TODO: There is a small window of opportunity for this function to 924 * get called from eviction after we've dropped the last GEM refcount, 925 * but before the TTM deleted flag is set on the object. Avoid 926 * adjusting the shrinker list in such cases, since the object is 927 * not available to the shrinker anyway due to its zero refcount. 928 * To fix this properly we should move to a TTM shrinker LRU list for 929 * these objects. 930 */ 931 if (kref_get_unless_zero(&obj->base.refcount)) { 932 if (shrinkable != obj->mm.ttm_shrinkable) { 933 if (shrinkable) { 934 if (obj->mm.madv == I915_MADV_WILLNEED) 935 __i915_gem_object_make_shrinkable(obj); 936 else 937 __i915_gem_object_make_purgeable(obj); 938 } else { 939 i915_gem_object_make_unshrinkable(obj); 940 } 941 942 obj->mm.ttm_shrinkable = shrinkable; 943 } 944 i915_gem_object_put(obj); 945 } 946 947 /* 948 * Put on the correct LRU list depending on the MADV status 949 */ 950 spin_lock(&bo->bdev->lru_lock); 951 if (shrinkable) { 952 /* Try to keep shmem_tt from being considered for shrinking. */ 953 bo->priority = TTM_MAX_BO_PRIORITY - 1; 954 } else if (obj->mm.madv != I915_MADV_WILLNEED) { 955 bo->priority = I915_TTM_PRIO_PURGE; 956 } else if (!i915_gem_object_has_pages(obj)) { 957 bo->priority = I915_TTM_PRIO_NO_PAGES; 958 } else { 959 struct ttm_resource_manager *man = 960 ttm_manager_type(bo->bdev, bo->resource->mem_type); 961 962 /* 963 * If we need to place an LMEM resource which doesn't need CPU 964 * access then we should try not to victimize mappable objects 965 * first, since we likely end up stealing more of the mappable 966 * portion. And likewise when we try to find space for a mappble 967 * object, we know not to ever victimize objects that don't 968 * occupy any mappable pages. 969 */ 970 if (i915_ttm_cpu_maps_iomem(bo->resource) && 971 i915_ttm_buddy_man_visible_size(man) < man->size && 972 !(obj->flags & I915_BO_ALLOC_GPU_ONLY)) 973 bo->priority = I915_TTM_PRIO_NEEDS_CPU_ACCESS; 974 else 975 bo->priority = I915_TTM_PRIO_HAS_PAGES; 976 } 977 978 ttm_bo_move_to_lru_tail(bo); 979 spin_unlock(&bo->bdev->lru_lock); 980 } 981 982 /* 983 * TTM-backed gem object destruction requires some clarification. 984 * Basically we have two possibilities here. We can either rely on the 985 * i915 delayed destruction and put the TTM object when the object 986 * is idle. This would be detected by TTM which would bypass the 987 * TTM delayed destroy handling. The other approach is to put the TTM 988 * object early and rely on the TTM destroyed handling, and then free 989 * the leftover parts of the GEM object once TTM's destroyed list handling is 990 * complete. For now, we rely on the latter for two reasons: 991 * a) TTM can evict an object even when it's on the delayed destroy list, 992 * which in theory allows for complete eviction. 993 * b) There is work going on in TTM to allow freeing an object even when 994 * it's not idle, and using the TTM destroyed list handling could help us 995 * benefit from that. 996 */ 997 static void i915_ttm_delayed_free(struct drm_i915_gem_object *obj) 998 { 999 GEM_BUG_ON(!obj->ttm.created); 1000 1001 ttm_bo_put(i915_gem_to_ttm(obj)); 1002 } 1003 1004 static vm_fault_t vm_fault_ttm(struct vm_fault *vmf) 1005 { 1006 struct vm_area_struct *area = vmf->vma; 1007 struct ttm_buffer_object *bo = area->vm_private_data; 1008 struct drm_device *dev = bo->base.dev; 1009 struct drm_i915_gem_object *obj; 1010 vm_fault_t ret; 1011 int idx; 1012 1013 obj = i915_ttm_to_gem(bo); 1014 if (!obj) 1015 return VM_FAULT_SIGBUS; 1016 1017 /* Sanity check that we allow writing into this object */ 1018 if (unlikely(i915_gem_object_is_readonly(obj) && 1019 area->vm_flags & VM_WRITE)) 1020 return VM_FAULT_SIGBUS; 1021 1022 ret = ttm_bo_vm_reserve(bo, vmf); 1023 if (ret) 1024 return ret; 1025 1026 if (obj->mm.madv != I915_MADV_WILLNEED) { 1027 dma_resv_unlock(bo->base.resv); 1028 return VM_FAULT_SIGBUS; 1029 } 1030 1031 if (!i915_ttm_resource_mappable(bo->resource)) { 1032 int err = -ENODEV; 1033 int i; 1034 1035 for (i = 0; i < obj->mm.n_placements; i++) { 1036 struct intel_memory_region *mr = obj->mm.placements[i]; 1037 unsigned int flags; 1038 1039 if (!mr->io_size && mr->type != INTEL_MEMORY_SYSTEM) 1040 continue; 1041 1042 flags = obj->flags; 1043 flags &= ~I915_BO_ALLOC_GPU_ONLY; 1044 err = __i915_ttm_migrate(obj, mr, flags); 1045 if (!err) 1046 break; 1047 } 1048 1049 if (err) { 1050 drm_dbg(dev, "Unable to make resource CPU accessible\n"); 1051 dma_resv_unlock(bo->base.resv); 1052 return VM_FAULT_SIGBUS; 1053 } 1054 } 1055 1056 if (drm_dev_enter(dev, &idx)) { 1057 ret = ttm_bo_vm_fault_reserved(vmf, vmf->vma->vm_page_prot, 1058 TTM_BO_VM_NUM_PREFAULT); 1059 drm_dev_exit(idx); 1060 } else { 1061 ret = ttm_bo_vm_dummy_page(vmf, vmf->vma->vm_page_prot); 1062 } 1063 if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) 1064 return ret; 1065 1066 i915_ttm_adjust_lru(obj); 1067 1068 dma_resv_unlock(bo->base.resv); 1069 return ret; 1070 } 1071 1072 static int 1073 vm_access_ttm(struct vm_area_struct *area, unsigned long addr, 1074 void *buf, int len, int write) 1075 { 1076 struct drm_i915_gem_object *obj = 1077 i915_ttm_to_gem(area->vm_private_data); 1078 1079 if (i915_gem_object_is_readonly(obj) && write) 1080 return -EACCES; 1081 1082 return ttm_bo_vm_access(area, addr, buf, len, write); 1083 } 1084 1085 static void ttm_vm_open(struct vm_area_struct *vma) 1086 { 1087 struct drm_i915_gem_object *obj = 1088 i915_ttm_to_gem(vma->vm_private_data); 1089 1090 GEM_BUG_ON(!obj); 1091 i915_gem_object_get(obj); 1092 } 1093 1094 static void ttm_vm_close(struct vm_area_struct *vma) 1095 { 1096 struct drm_i915_gem_object *obj = 1097 i915_ttm_to_gem(vma->vm_private_data); 1098 1099 GEM_BUG_ON(!obj); 1100 i915_gem_object_put(obj); 1101 } 1102 1103 static const struct vm_operations_struct vm_ops_ttm = { 1104 .fault = vm_fault_ttm, 1105 .access = vm_access_ttm, 1106 .open = ttm_vm_open, 1107 .close = ttm_vm_close, 1108 }; 1109 1110 static u64 i915_ttm_mmap_offset(struct drm_i915_gem_object *obj) 1111 { 1112 /* The ttm_bo must be allocated with I915_BO_ALLOC_USER */ 1113 GEM_BUG_ON(!drm_mm_node_allocated(&obj->base.vma_node.vm_node)); 1114 1115 return drm_vma_node_offset_addr(&obj->base.vma_node); 1116 } 1117 1118 static void i915_ttm_unmap_virtual(struct drm_i915_gem_object *obj) 1119 { 1120 ttm_bo_unmap_virtual(i915_gem_to_ttm(obj)); 1121 } 1122 1123 static const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = { 1124 .name = "i915_gem_object_ttm", 1125 .flags = I915_GEM_OBJECT_IS_SHRINKABLE | 1126 I915_GEM_OBJECT_SELF_MANAGED_SHRINK_LIST, 1127 1128 .get_pages = i915_ttm_get_pages, 1129 .put_pages = i915_ttm_put_pages, 1130 .truncate = i915_ttm_truncate, 1131 .shrink = i915_ttm_shrink, 1132 1133 .adjust_lru = i915_ttm_adjust_lru, 1134 .delayed_free = i915_ttm_delayed_free, 1135 .migrate = i915_ttm_migrate, 1136 1137 .mmap_offset = i915_ttm_mmap_offset, 1138 .unmap_virtual = i915_ttm_unmap_virtual, 1139 .mmap_ops = &vm_ops_ttm, 1140 }; 1141 1142 void i915_ttm_bo_destroy(struct ttm_buffer_object *bo) 1143 { 1144 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo); 1145 1146 i915_gem_object_release_memory_region(obj); 1147 mutex_destroy(&obj->ttm.get_io_page.lock); 1148 1149 if (obj->ttm.created) { 1150 /* 1151 * We freely manage the shrinker LRU outide of the mm.pages life 1152 * cycle. As a result when destroying the object we should be 1153 * extra paranoid and ensure we remove it from the LRU, before 1154 * we free the object. 1155 * 1156 * Touching the ttm_shrinkable outside of the object lock here 1157 * should be safe now that the last GEM object ref was dropped. 1158 */ 1159 if (obj->mm.ttm_shrinkable) 1160 i915_gem_object_make_unshrinkable(obj); 1161 1162 i915_ttm_backup_free(obj); 1163 1164 /* This releases all gem object bindings to the backend. */ 1165 __i915_gem_free_object(obj); 1166 1167 call_rcu(&obj->rcu, __i915_gem_free_object_rcu); 1168 } else { 1169 __i915_gem_object_fini(obj); 1170 } 1171 } 1172 1173 /** 1174 * __i915_gem_ttm_object_init - Initialize a ttm-backed i915 gem object 1175 * @mem: The initial memory region for the object. 1176 * @obj: The gem object. 1177 * @size: Object size in bytes. 1178 * @flags: gem object flags. 1179 * 1180 * Return: 0 on success, negative error code on failure. 1181 */ 1182 int __i915_gem_ttm_object_init(struct intel_memory_region *mem, 1183 struct drm_i915_gem_object *obj, 1184 resource_size_t offset, 1185 resource_size_t size, 1186 resource_size_t page_size, 1187 unsigned int flags) 1188 { 1189 static struct lock_class_key lock_class; 1190 struct drm_i915_private *i915 = mem->i915; 1191 struct ttm_operation_ctx ctx = { 1192 .interruptible = true, 1193 .no_wait_gpu = false, 1194 }; 1195 enum ttm_bo_type bo_type; 1196 int ret; 1197 1198 drm_gem_private_object_init(&i915->drm, &obj->base, size); 1199 i915_gem_object_init(obj, &i915_gem_ttm_obj_ops, &lock_class, flags); 1200 1201 obj->bo_offset = offset; 1202 1203 /* Don't put on a region list until we're either locked or fully initialized. */ 1204 obj->mm.region = mem; 1205 INIT_LIST_HEAD(&obj->mm.region_link); 1206 1207 INIT_RADIX_TREE(&obj->ttm.get_io_page.radix, GFP_KERNEL | __GFP_NOWARN); 1208 mutex_init(&obj->ttm.get_io_page.lock); 1209 bo_type = (obj->flags & I915_BO_ALLOC_USER) ? ttm_bo_type_device : 1210 ttm_bo_type_kernel; 1211 1212 obj->base.vma_node.driver_private = i915_gem_to_ttm(obj); 1213 1214 /* Forcing the page size is kernel internal only */ 1215 GEM_BUG_ON(page_size && obj->mm.n_placements); 1216 1217 /* 1218 * Keep an extra shrink pin to prevent the object from being made 1219 * shrinkable too early. If the ttm_tt is ever allocated in shmem, we 1220 * drop the pin. The TTM backend manages the shrinker LRU itself, 1221 * outside of the normal mm.pages life cycle. 1222 */ 1223 i915_gem_object_make_unshrinkable(obj); 1224 1225 /* 1226 * If this function fails, it will call the destructor, but 1227 * our caller still owns the object. So no freeing in the 1228 * destructor until obj->ttm.created is true. 1229 * Similarly, in delayed_destroy, we can't call ttm_bo_put() 1230 * until successful initialization. 1231 */ 1232 ret = ttm_bo_init_reserved(&i915->bdev, i915_gem_to_ttm(obj), size, 1233 bo_type, &i915_sys_placement, 1234 page_size >> PAGE_SHIFT, 1235 &ctx, NULL, NULL, i915_ttm_bo_destroy); 1236 if (ret) 1237 return i915_ttm_err_to_gem(ret); 1238 1239 obj->ttm.created = true; 1240 i915_gem_object_release_memory_region(obj); 1241 i915_gem_object_init_memory_region(obj, mem); 1242 i915_ttm_adjust_domains_after_move(obj); 1243 i915_ttm_adjust_gem_after_move(obj); 1244 i915_gem_object_unlock(obj); 1245 1246 return 0; 1247 } 1248 1249 static const struct intel_memory_region_ops ttm_system_region_ops = { 1250 .init_object = __i915_gem_ttm_object_init, 1251 .release = intel_region_ttm_fini, 1252 }; 1253 1254 struct intel_memory_region * 1255 i915_gem_ttm_system_setup(struct drm_i915_private *i915, 1256 u16 type, u16 instance) 1257 { 1258 struct intel_memory_region *mr; 1259 1260 mr = intel_memory_region_create(i915, 0, 1261 totalram_pages() << PAGE_SHIFT, 1262 PAGE_SIZE, 0, 0, 1263 type, instance, 1264 &ttm_system_region_ops); 1265 if (IS_ERR(mr)) 1266 return mr; 1267 1268 intel_memory_region_set_name(mr, "system-ttm"); 1269 return mr; 1270 } 1271