1 /* 2 * Copyright © 2017 Intel Corporation 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 21 * IN THE SOFTWARE. 22 * 23 */ 24 25 #include <linux/highmem.h> 26 #include <linux/sched/mm.h> 27 28 #include <drm/drm_cache.h> 29 30 #include "display/intel_frontbuffer.h" 31 #include "pxp/intel_pxp.h" 32 33 #include "i915_drv.h" 34 #include "i915_file_private.h" 35 #include "i915_gem_clflush.h" 36 #include "i915_gem_context.h" 37 #include "i915_gem_dmabuf.h" 38 #include "i915_gem_mman.h" 39 #include "i915_gem_object.h" 40 #include "i915_gem_ttm.h" 41 #include "i915_memcpy.h" 42 #include "i915_trace.h" 43 44 static struct kmem_cache *slab_objects; 45 46 static const struct drm_gem_object_funcs i915_gem_object_funcs; 47 48 unsigned int i915_gem_get_pat_index(struct drm_i915_private *i915, 49 enum i915_cache_level level) 50 { 51 if (drm_WARN_ON(&i915->drm, level >= I915_MAX_CACHE_LEVEL)) 52 return 0; 53 54 return INTEL_INFO(i915)->cachelevel_to_pat[level]; 55 } 56 57 bool i915_gem_object_has_cache_level(const struct drm_i915_gem_object *obj, 58 enum i915_cache_level lvl) 59 { 60 /* 61 * In case the pat_index is set by user space, this kernel mode 62 * driver should leave the coherency to be managed by user space, 63 * simply return true here. 64 */ 65 if (obj->pat_set_by_user) 66 return true; 67 68 /* 69 * Otherwise the pat_index should have been converted from cache_level 70 * so that the following comparison is valid. 71 */ 72 return obj->pat_index == i915_gem_get_pat_index(obj_to_i915(obj), lvl); 73 } 74 75 struct drm_i915_gem_object *i915_gem_object_alloc(void) 76 { 77 struct drm_i915_gem_object *obj; 78 79 obj = kmem_cache_zalloc(slab_objects, GFP_KERNEL); 80 if (!obj) 81 return NULL; 82 obj->base.funcs = &i915_gem_object_funcs; 83 84 return obj; 85 } 86 87 void i915_gem_object_free(struct drm_i915_gem_object *obj) 88 { 89 return kmem_cache_free(slab_objects, obj); 90 } 91 92 void i915_gem_object_init(struct drm_i915_gem_object *obj, 93 const struct drm_i915_gem_object_ops *ops, 94 struct lock_class_key *key, unsigned flags) 95 { 96 /* 97 * A gem object is embedded both in a struct ttm_buffer_object :/ and 98 * in a drm_i915_gem_object. Make sure they are aliased. 99 */ 100 BUILD_BUG_ON(offsetof(typeof(*obj), base) != 101 offsetof(typeof(*obj), __do_not_access.base)); 102 103 spin_lock_init(&obj->vma.lock); 104 INIT_LIST_HEAD(&obj->vma.list); 105 106 INIT_LIST_HEAD(&obj->mm.link); 107 108 INIT_LIST_HEAD(&obj->lut_list); 109 spin_lock_init(&obj->lut_lock); 110 111 spin_lock_init(&obj->mmo.lock); 112 obj->mmo.offsets = RB_ROOT; 113 114 init_rcu_head(&obj->rcu); 115 116 obj->ops = ops; 117 GEM_BUG_ON(flags & ~I915_BO_ALLOC_FLAGS); 118 obj->flags = flags; 119 120 obj->mm.madv = I915_MADV_WILLNEED; 121 INIT_RADIX_TREE(&obj->mm.get_page.radix, GFP_KERNEL | __GFP_NOWARN); 122 mutex_init(&obj->mm.get_page.lock); 123 INIT_RADIX_TREE(&obj->mm.get_dma_page.radix, GFP_KERNEL | __GFP_NOWARN); 124 mutex_init(&obj->mm.get_dma_page.lock); 125 } 126 127 /** 128 * __i915_gem_object_fini - Clean up a GEM object initialization 129 * @obj: The gem object to cleanup 130 * 131 * This function cleans up gem object fields that are set up by 132 * drm_gem_private_object_init() and i915_gem_object_init(). 133 * It's primarily intended as a helper for backends that need to 134 * clean up the gem object in separate steps. 135 */ 136 void __i915_gem_object_fini(struct drm_i915_gem_object *obj) 137 { 138 mutex_destroy(&obj->mm.get_page.lock); 139 mutex_destroy(&obj->mm.get_dma_page.lock); 140 dma_resv_fini(&obj->base._resv); 141 } 142 143 /** 144 * i915_gem_object_set_cache_coherency - Mark up the object's coherency levels 145 * for a given cache_level 146 * @obj: #drm_i915_gem_object 147 * @cache_level: cache level 148 */ 149 void i915_gem_object_set_cache_coherency(struct drm_i915_gem_object *obj, 150 unsigned int cache_level) 151 { 152 struct drm_i915_private *i915 = to_i915(obj->base.dev); 153 154 obj->pat_index = i915_gem_get_pat_index(i915, cache_level); 155 156 if (cache_level != I915_CACHE_NONE) 157 obj->cache_coherent = (I915_BO_CACHE_COHERENT_FOR_READ | 158 I915_BO_CACHE_COHERENT_FOR_WRITE); 159 else if (HAS_LLC(i915)) 160 obj->cache_coherent = I915_BO_CACHE_COHERENT_FOR_READ; 161 else 162 obj->cache_coherent = 0; 163 164 obj->cache_dirty = 165 !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE) && 166 !IS_DGFX(i915); 167 } 168 169 /** 170 * i915_gem_object_set_pat_index - set PAT index to be used in PTE encode 171 * @obj: #drm_i915_gem_object 172 * @pat_index: PAT index 173 * 174 * This is a clone of i915_gem_object_set_cache_coherency taking pat index 175 * instead of cache_level as its second argument. 176 */ 177 void i915_gem_object_set_pat_index(struct drm_i915_gem_object *obj, 178 unsigned int pat_index) 179 { 180 struct drm_i915_private *i915 = to_i915(obj->base.dev); 181 182 if (obj->pat_index == pat_index) 183 return; 184 185 obj->pat_index = pat_index; 186 187 if (pat_index != i915_gem_get_pat_index(i915, I915_CACHE_NONE)) 188 obj->cache_coherent = (I915_BO_CACHE_COHERENT_FOR_READ | 189 I915_BO_CACHE_COHERENT_FOR_WRITE); 190 else if (HAS_LLC(i915)) 191 obj->cache_coherent = I915_BO_CACHE_COHERENT_FOR_READ; 192 else 193 obj->cache_coherent = 0; 194 195 obj->cache_dirty = 196 !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE) && 197 !IS_DGFX(i915); 198 } 199 200 bool i915_gem_object_can_bypass_llc(struct drm_i915_gem_object *obj) 201 { 202 struct drm_i915_private *i915 = to_i915(obj->base.dev); 203 204 /* 205 * This is purely from a security perspective, so we simply don't care 206 * about non-userspace objects being able to bypass the LLC. 207 */ 208 if (!(obj->flags & I915_BO_ALLOC_USER)) 209 return false; 210 211 /* 212 * EHL and JSL add the 'Bypass LLC' MOCS entry, which should make it 213 * possible for userspace to bypass the GTT caching bits set by the 214 * kernel, as per the given object cache_level. This is troublesome 215 * since the heavy flush we apply when first gathering the pages is 216 * skipped if the kernel thinks the object is coherent with the GPU. As 217 * a result it might be possible to bypass the cache and read the 218 * contents of the page directly, which could be stale data. If it's 219 * just a case of userspace shooting themselves in the foot then so be 220 * it, but since i915 takes the stance of always zeroing memory before 221 * handing it to userspace, we need to prevent this. 222 */ 223 return IS_JSL_EHL(i915); 224 } 225 226 static void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file) 227 { 228 struct drm_i915_gem_object *obj = to_intel_bo(gem); 229 struct drm_i915_file_private *fpriv = file->driver_priv; 230 struct i915_lut_handle bookmark = {}; 231 struct i915_mmap_offset *mmo, *mn; 232 struct i915_lut_handle *lut, *ln; 233 LIST_HEAD(close); 234 235 spin_lock(&obj->lut_lock); 236 list_for_each_entry_safe(lut, ln, &obj->lut_list, obj_link) { 237 struct i915_gem_context *ctx = lut->ctx; 238 239 if (ctx && ctx->file_priv == fpriv) { 240 i915_gem_context_get(ctx); 241 list_move(&lut->obj_link, &close); 242 } 243 244 /* Break long locks, and carefully continue on from this spot */ 245 if (&ln->obj_link != &obj->lut_list) { 246 list_add_tail(&bookmark.obj_link, &ln->obj_link); 247 if (cond_resched_lock(&obj->lut_lock)) 248 list_safe_reset_next(&bookmark, ln, obj_link); 249 __list_del_entry(&bookmark.obj_link); 250 } 251 } 252 spin_unlock(&obj->lut_lock); 253 254 spin_lock(&obj->mmo.lock); 255 rbtree_postorder_for_each_entry_safe(mmo, mn, &obj->mmo.offsets, offset) 256 drm_vma_node_revoke(&mmo->vma_node, file); 257 spin_unlock(&obj->mmo.lock); 258 259 list_for_each_entry_safe(lut, ln, &close, obj_link) { 260 struct i915_gem_context *ctx = lut->ctx; 261 struct i915_vma *vma; 262 263 /* 264 * We allow the process to have multiple handles to the same 265 * vma, in the same fd namespace, by virtue of flink/open. 266 */ 267 268 mutex_lock(&ctx->lut_mutex); 269 vma = radix_tree_delete(&ctx->handles_vma, lut->handle); 270 if (vma) { 271 GEM_BUG_ON(vma->obj != obj); 272 GEM_BUG_ON(!atomic_read(&vma->open_count)); 273 i915_vma_close(vma); 274 } 275 mutex_unlock(&ctx->lut_mutex); 276 277 i915_gem_context_put(lut->ctx); 278 i915_lut_handle_free(lut); 279 i915_gem_object_put(obj); 280 } 281 } 282 283 void __i915_gem_free_object_rcu(struct rcu_head *head) 284 { 285 struct drm_i915_gem_object *obj = 286 container_of(head, typeof(*obj), rcu); 287 struct drm_i915_private *i915 = to_i915(obj->base.dev); 288 289 i915_gem_object_free(obj); 290 291 GEM_BUG_ON(!atomic_read(&i915->mm.free_count)); 292 atomic_dec(&i915->mm.free_count); 293 } 294 295 static void __i915_gem_object_free_mmaps(struct drm_i915_gem_object *obj) 296 { 297 /* Skip serialisation and waking the device if known to be not used. */ 298 299 if (obj->userfault_count && !IS_DGFX(to_i915(obj->base.dev))) 300 i915_gem_object_release_mmap_gtt(obj); 301 302 if (!RB_EMPTY_ROOT(&obj->mmo.offsets)) { 303 struct i915_mmap_offset *mmo, *mn; 304 305 i915_gem_object_release_mmap_offset(obj); 306 307 rbtree_postorder_for_each_entry_safe(mmo, mn, 308 &obj->mmo.offsets, 309 offset) { 310 drm_vma_offset_remove(obj->base.dev->vma_offset_manager, 311 &mmo->vma_node); 312 kfree(mmo); 313 } 314 obj->mmo.offsets = RB_ROOT; 315 } 316 } 317 318 /** 319 * __i915_gem_object_pages_fini - Clean up pages use of a gem object 320 * @obj: The gem object to clean up 321 * 322 * This function cleans up usage of the object mm.pages member. It 323 * is intended for backends that need to clean up a gem object in 324 * separate steps and needs to be called when the object is idle before 325 * the object's backing memory is freed. 326 */ 327 void __i915_gem_object_pages_fini(struct drm_i915_gem_object *obj) 328 { 329 assert_object_held_shared(obj); 330 331 if (!list_empty(&obj->vma.list)) { 332 struct i915_vma *vma; 333 334 spin_lock(&obj->vma.lock); 335 while ((vma = list_first_entry_or_null(&obj->vma.list, 336 struct i915_vma, 337 obj_link))) { 338 GEM_BUG_ON(vma->obj != obj); 339 spin_unlock(&obj->vma.lock); 340 341 i915_vma_destroy(vma); 342 343 spin_lock(&obj->vma.lock); 344 } 345 spin_unlock(&obj->vma.lock); 346 } 347 348 __i915_gem_object_free_mmaps(obj); 349 350 atomic_set(&obj->mm.pages_pin_count, 0); 351 352 /* 353 * dma_buf_unmap_attachment() requires reservation to be 354 * locked. The imported GEM shouldn't share reservation lock 355 * and ttm_bo_cleanup_memtype_use() shouldn't be invoked for 356 * dma-buf, so it's safe to take the lock. 357 */ 358 if (obj->base.import_attach) 359 i915_gem_object_lock(obj, NULL); 360 361 __i915_gem_object_put_pages(obj); 362 363 if (obj->base.import_attach) 364 i915_gem_object_unlock(obj); 365 366 GEM_BUG_ON(i915_gem_object_has_pages(obj)); 367 } 368 369 void __i915_gem_free_object(struct drm_i915_gem_object *obj) 370 { 371 trace_i915_gem_object_destroy(obj); 372 373 GEM_BUG_ON(!list_empty(&obj->lut_list)); 374 375 bitmap_free(obj->bit_17); 376 377 if (obj->base.import_attach) 378 drm_prime_gem_destroy(&obj->base, NULL); 379 380 drm_gem_free_mmap_offset(&obj->base); 381 382 if (obj->ops->release) 383 obj->ops->release(obj); 384 385 if (obj->mm.n_placements > 1) 386 kfree(obj->mm.placements); 387 388 if (obj->shares_resv_from) 389 i915_vm_resv_put(obj->shares_resv_from); 390 391 __i915_gem_object_fini(obj); 392 } 393 394 static void __i915_gem_free_objects(struct drm_i915_private *i915, 395 struct llist_node *freed) 396 { 397 struct drm_i915_gem_object *obj, *on; 398 399 llist_for_each_entry_safe(obj, on, freed, freed) { 400 might_sleep(); 401 if (obj->ops->delayed_free) { 402 obj->ops->delayed_free(obj); 403 continue; 404 } 405 406 __i915_gem_object_pages_fini(obj); 407 __i915_gem_free_object(obj); 408 409 /* But keep the pointer alive for RCU-protected lookups */ 410 call_rcu(&obj->rcu, __i915_gem_free_object_rcu); 411 cond_resched(); 412 } 413 } 414 415 void i915_gem_flush_free_objects(struct drm_i915_private *i915) 416 { 417 struct llist_node *freed = llist_del_all(&i915->mm.free_list); 418 419 if (unlikely(freed)) 420 __i915_gem_free_objects(i915, freed); 421 } 422 423 static void __i915_gem_free_work(struct work_struct *work) 424 { 425 struct drm_i915_private *i915 = 426 container_of(work, struct drm_i915_private, mm.free_work); 427 428 i915_gem_flush_free_objects(i915); 429 } 430 431 static void i915_gem_free_object(struct drm_gem_object *gem_obj) 432 { 433 struct drm_i915_gem_object *obj = to_intel_bo(gem_obj); 434 struct drm_i915_private *i915 = to_i915(obj->base.dev); 435 436 GEM_BUG_ON(i915_gem_object_is_framebuffer(obj)); 437 438 /* 439 * Before we free the object, make sure any pure RCU-only 440 * read-side critical sections are complete, e.g. 441 * i915_gem_busy_ioctl(). For the corresponding synchronized 442 * lookup see i915_gem_object_lookup_rcu(). 443 */ 444 atomic_inc(&i915->mm.free_count); 445 446 /* 447 * Since we require blocking on struct_mutex to unbind the freed 448 * object from the GPU before releasing resources back to the 449 * system, we can not do that directly from the RCU callback (which may 450 * be a softirq context), but must instead then defer that work onto a 451 * kthread. We use the RCU callback rather than move the freed object 452 * directly onto the work queue so that we can mix between using the 453 * worker and performing frees directly from subsequent allocations for 454 * crude but effective memory throttling. 455 */ 456 457 if (llist_add(&obj->freed, &i915->mm.free_list)) 458 queue_work(i915->wq, &i915->mm.free_work); 459 } 460 461 void __i915_gem_object_flush_frontbuffer(struct drm_i915_gem_object *obj, 462 enum fb_op_origin origin) 463 { 464 struct intel_frontbuffer *front; 465 466 front = __intel_frontbuffer_get(obj); 467 if (front) { 468 intel_frontbuffer_flush(front, origin); 469 intel_frontbuffer_put(front); 470 } 471 } 472 473 void __i915_gem_object_invalidate_frontbuffer(struct drm_i915_gem_object *obj, 474 enum fb_op_origin origin) 475 { 476 struct intel_frontbuffer *front; 477 478 front = __intel_frontbuffer_get(obj); 479 if (front) { 480 intel_frontbuffer_invalidate(front, origin); 481 intel_frontbuffer_put(front); 482 } 483 } 484 485 static void 486 i915_gem_object_read_from_page_kmap(struct drm_i915_gem_object *obj, u64 offset, void *dst, int size) 487 { 488 pgoff_t idx = offset >> PAGE_SHIFT; 489 void *src_map; 490 void *src_ptr; 491 492 src_map = kmap_atomic(i915_gem_object_get_page(obj, idx)); 493 494 src_ptr = src_map + offset_in_page(offset); 495 if (!(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ)) 496 drm_clflush_virt_range(src_ptr, size); 497 memcpy(dst, src_ptr, size); 498 499 kunmap_atomic(src_map); 500 } 501 502 static void 503 i915_gem_object_read_from_page_iomap(struct drm_i915_gem_object *obj, u64 offset, void *dst, int size) 504 { 505 pgoff_t idx = offset >> PAGE_SHIFT; 506 dma_addr_t dma = i915_gem_object_get_dma_address(obj, idx); 507 void __iomem *src_map; 508 void __iomem *src_ptr; 509 510 src_map = io_mapping_map_wc(&obj->mm.region->iomap, 511 dma - obj->mm.region->region.start, 512 PAGE_SIZE); 513 514 src_ptr = src_map + offset_in_page(offset); 515 if (!i915_memcpy_from_wc(dst, (void __force *)src_ptr, size)) 516 memcpy_fromio(dst, src_ptr, size); 517 518 io_mapping_unmap(src_map); 519 } 520 521 static bool object_has_mappable_iomem(struct drm_i915_gem_object *obj) 522 { 523 GEM_BUG_ON(!i915_gem_object_has_iomem(obj)); 524 525 if (IS_DGFX(to_i915(obj->base.dev))) 526 return i915_ttm_resource_mappable(i915_gem_to_ttm(obj)->resource); 527 528 return true; 529 } 530 531 /** 532 * i915_gem_object_read_from_page - read data from the page of a GEM object 533 * @obj: GEM object to read from 534 * @offset: offset within the object 535 * @dst: buffer to store the read data 536 * @size: size to read 537 * 538 * Reads data from @obj at the specified offset. The requested region to read 539 * from can't cross a page boundary. The caller must ensure that @obj pages 540 * are pinned and that @obj is synced wrt. any related writes. 541 * 542 * Return: %0 on success or -ENODEV if the type of @obj's backing store is 543 * unsupported. 544 */ 545 int i915_gem_object_read_from_page(struct drm_i915_gem_object *obj, u64 offset, void *dst, int size) 546 { 547 GEM_BUG_ON(overflows_type(offset >> PAGE_SHIFT, pgoff_t)); 548 GEM_BUG_ON(offset >= obj->base.size); 549 GEM_BUG_ON(offset_in_page(offset) > PAGE_SIZE - size); 550 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj)); 551 552 if (i915_gem_object_has_struct_page(obj)) 553 i915_gem_object_read_from_page_kmap(obj, offset, dst, size); 554 else if (i915_gem_object_has_iomem(obj) && object_has_mappable_iomem(obj)) 555 i915_gem_object_read_from_page_iomap(obj, offset, dst, size); 556 else 557 return -ENODEV; 558 559 return 0; 560 } 561 562 /** 563 * i915_gem_object_evictable - Whether object is likely evictable after unbind. 564 * @obj: The object to check 565 * 566 * This function checks whether the object is likely unvictable after unbind. 567 * If the object is not locked when checking, the result is only advisory. 568 * If the object is locked when checking, and the function returns true, 569 * then an eviction should indeed be possible. But since unlocked vma 570 * unpinning and unbinding is currently possible, the object can actually 571 * become evictable even if this function returns false. 572 * 573 * Return: true if the object may be evictable. False otherwise. 574 */ 575 bool i915_gem_object_evictable(struct drm_i915_gem_object *obj) 576 { 577 struct i915_vma *vma; 578 int pin_count = atomic_read(&obj->mm.pages_pin_count); 579 580 if (!pin_count) 581 return true; 582 583 spin_lock(&obj->vma.lock); 584 list_for_each_entry(vma, &obj->vma.list, obj_link) { 585 if (i915_vma_is_pinned(vma)) { 586 spin_unlock(&obj->vma.lock); 587 return false; 588 } 589 if (atomic_read(&vma->pages_count)) 590 pin_count--; 591 } 592 spin_unlock(&obj->vma.lock); 593 GEM_WARN_ON(pin_count < 0); 594 595 return pin_count == 0; 596 } 597 598 /** 599 * i915_gem_object_migratable - Whether the object is migratable out of the 600 * current region. 601 * @obj: Pointer to the object. 602 * 603 * Return: Whether the object is allowed to be resident in other 604 * regions than the current while pages are present. 605 */ 606 bool i915_gem_object_migratable(struct drm_i915_gem_object *obj) 607 { 608 struct intel_memory_region *mr = READ_ONCE(obj->mm.region); 609 610 if (!mr) 611 return false; 612 613 return obj->mm.n_placements > 1; 614 } 615 616 /** 617 * i915_gem_object_has_struct_page - Whether the object is page-backed 618 * @obj: The object to query. 619 * 620 * This function should only be called while the object is locked or pinned, 621 * otherwise the page backing may change under the caller. 622 * 623 * Return: True if page-backed, false otherwise. 624 */ 625 bool i915_gem_object_has_struct_page(const struct drm_i915_gem_object *obj) 626 { 627 #ifdef CONFIG_LOCKDEP 628 if (IS_DGFX(to_i915(obj->base.dev)) && 629 i915_gem_object_evictable((void __force *)obj)) 630 assert_object_held_shared(obj); 631 #endif 632 return obj->mem_flags & I915_BO_FLAG_STRUCT_PAGE; 633 } 634 635 /** 636 * i915_gem_object_has_iomem - Whether the object is iomem-backed 637 * @obj: The object to query. 638 * 639 * This function should only be called while the object is locked or pinned, 640 * otherwise the iomem backing may change under the caller. 641 * 642 * Return: True if iomem-backed, false otherwise. 643 */ 644 bool i915_gem_object_has_iomem(const struct drm_i915_gem_object *obj) 645 { 646 #ifdef CONFIG_LOCKDEP 647 if (IS_DGFX(to_i915(obj->base.dev)) && 648 i915_gem_object_evictable((void __force *)obj)) 649 assert_object_held_shared(obj); 650 #endif 651 return obj->mem_flags & I915_BO_FLAG_IOMEM; 652 } 653 654 /** 655 * i915_gem_object_can_migrate - Whether an object likely can be migrated 656 * 657 * @obj: The object to migrate 658 * @id: The region intended to migrate to 659 * 660 * Check whether the object backend supports migration to the 661 * given region. Note that pinning may affect the ability to migrate as 662 * returned by this function. 663 * 664 * This function is primarily intended as a helper for checking the 665 * possibility to migrate objects and might be slightly less permissive 666 * than i915_gem_object_migrate() when it comes to objects with the 667 * I915_BO_ALLOC_USER flag set. 668 * 669 * Return: true if migration is possible, false otherwise. 670 */ 671 bool i915_gem_object_can_migrate(struct drm_i915_gem_object *obj, 672 enum intel_region_id id) 673 { 674 struct drm_i915_private *i915 = to_i915(obj->base.dev); 675 unsigned int num_allowed = obj->mm.n_placements; 676 struct intel_memory_region *mr; 677 unsigned int i; 678 679 GEM_BUG_ON(id >= INTEL_REGION_UNKNOWN); 680 GEM_BUG_ON(obj->mm.madv != I915_MADV_WILLNEED); 681 682 mr = i915->mm.regions[id]; 683 if (!mr) 684 return false; 685 686 if (!IS_ALIGNED(obj->base.size, mr->min_page_size)) 687 return false; 688 689 if (obj->mm.region == mr) 690 return true; 691 692 if (!i915_gem_object_evictable(obj)) 693 return false; 694 695 if (!obj->ops->migrate) 696 return false; 697 698 if (!(obj->flags & I915_BO_ALLOC_USER)) 699 return true; 700 701 if (num_allowed == 0) 702 return false; 703 704 for (i = 0; i < num_allowed; ++i) { 705 if (mr == obj->mm.placements[i]) 706 return true; 707 } 708 709 return false; 710 } 711 712 /** 713 * i915_gem_object_migrate - Migrate an object to the desired region id 714 * @obj: The object to migrate. 715 * @ww: An optional struct i915_gem_ww_ctx. If NULL, the backend may 716 * not be successful in evicting other objects to make room for this object. 717 * @id: The region id to migrate to. 718 * 719 * Attempt to migrate the object to the desired memory region. The 720 * object backend must support migration and the object may not be 721 * pinned, (explicitly pinned pages or pinned vmas). The object must 722 * be locked. 723 * On successful completion, the object will have pages pointing to 724 * memory in the new region, but an async migration task may not have 725 * completed yet, and to accomplish that, i915_gem_object_wait_migration() 726 * must be called. 727 * 728 * Note: the @ww parameter is not used yet, but included to make sure 729 * callers put some effort into obtaining a valid ww ctx if one is 730 * available. 731 * 732 * Return: 0 on success. Negative error code on failure. In particular may 733 * return -ENXIO on lack of region space, -EDEADLK for deadlock avoidance 734 * if @ww is set, -EINTR or -ERESTARTSYS if signal pending, and 735 * -EBUSY if the object is pinned. 736 */ 737 int i915_gem_object_migrate(struct drm_i915_gem_object *obj, 738 struct i915_gem_ww_ctx *ww, 739 enum intel_region_id id) 740 { 741 return __i915_gem_object_migrate(obj, ww, id, obj->flags); 742 } 743 744 /** 745 * __i915_gem_object_migrate - Migrate an object to the desired region id, with 746 * control of the extra flags 747 * @obj: The object to migrate. 748 * @ww: An optional struct i915_gem_ww_ctx. If NULL, the backend may 749 * not be successful in evicting other objects to make room for this object. 750 * @id: The region id to migrate to. 751 * @flags: The object flags. Normally just obj->flags. 752 * 753 * Attempt to migrate the object to the desired memory region. The 754 * object backend must support migration and the object may not be 755 * pinned, (explicitly pinned pages or pinned vmas). The object must 756 * be locked. 757 * On successful completion, the object will have pages pointing to 758 * memory in the new region, but an async migration task may not have 759 * completed yet, and to accomplish that, i915_gem_object_wait_migration() 760 * must be called. 761 * 762 * Note: the @ww parameter is not used yet, but included to make sure 763 * callers put some effort into obtaining a valid ww ctx if one is 764 * available. 765 * 766 * Return: 0 on success. Negative error code on failure. In particular may 767 * return -ENXIO on lack of region space, -EDEADLK for deadlock avoidance 768 * if @ww is set, -EINTR or -ERESTARTSYS if signal pending, and 769 * -EBUSY if the object is pinned. 770 */ 771 int __i915_gem_object_migrate(struct drm_i915_gem_object *obj, 772 struct i915_gem_ww_ctx *ww, 773 enum intel_region_id id, 774 unsigned int flags) 775 { 776 struct drm_i915_private *i915 = to_i915(obj->base.dev); 777 struct intel_memory_region *mr; 778 779 GEM_BUG_ON(id >= INTEL_REGION_UNKNOWN); 780 GEM_BUG_ON(obj->mm.madv != I915_MADV_WILLNEED); 781 assert_object_held(obj); 782 783 mr = i915->mm.regions[id]; 784 GEM_BUG_ON(!mr); 785 786 if (!i915_gem_object_can_migrate(obj, id)) 787 return -EINVAL; 788 789 if (!obj->ops->migrate) { 790 if (GEM_WARN_ON(obj->mm.region != mr)) 791 return -EINVAL; 792 return 0; 793 } 794 795 return obj->ops->migrate(obj, mr, flags); 796 } 797 798 /** 799 * i915_gem_object_placement_possible - Check whether the object can be 800 * placed at certain memory type 801 * @obj: Pointer to the object 802 * @type: The memory type to check 803 * 804 * Return: True if the object can be placed in @type. False otherwise. 805 */ 806 bool i915_gem_object_placement_possible(struct drm_i915_gem_object *obj, 807 enum intel_memory_type type) 808 { 809 unsigned int i; 810 811 if (!obj->mm.n_placements) { 812 switch (type) { 813 case INTEL_MEMORY_LOCAL: 814 return i915_gem_object_has_iomem(obj); 815 case INTEL_MEMORY_SYSTEM: 816 return i915_gem_object_has_pages(obj); 817 default: 818 /* Ignore stolen for now */ 819 GEM_BUG_ON(1); 820 return false; 821 } 822 } 823 824 for (i = 0; i < obj->mm.n_placements; i++) { 825 if (obj->mm.placements[i]->type == type) 826 return true; 827 } 828 829 return false; 830 } 831 832 /** 833 * i915_gem_object_needs_ccs_pages - Check whether the object requires extra 834 * pages when placed in system-memory, in order to save and later restore the 835 * flat-CCS aux state when the object is moved between local-memory and 836 * system-memory 837 * @obj: Pointer to the object 838 * 839 * Return: True if the object needs extra ccs pages. False otherwise. 840 */ 841 bool i915_gem_object_needs_ccs_pages(struct drm_i915_gem_object *obj) 842 { 843 bool lmem_placement = false; 844 int i; 845 846 if (!HAS_FLAT_CCS(to_i915(obj->base.dev))) 847 return false; 848 849 if (obj->flags & I915_BO_ALLOC_CCS_AUX) 850 return true; 851 852 for (i = 0; i < obj->mm.n_placements; i++) { 853 /* Compression is not allowed for the objects with smem placement */ 854 if (obj->mm.placements[i]->type == INTEL_MEMORY_SYSTEM) 855 return false; 856 if (!lmem_placement && 857 obj->mm.placements[i]->type == INTEL_MEMORY_LOCAL) 858 lmem_placement = true; 859 } 860 861 return lmem_placement; 862 } 863 864 void i915_gem_init__objects(struct drm_i915_private *i915) 865 { 866 INIT_WORK(&i915->mm.free_work, __i915_gem_free_work); 867 } 868 869 void i915_objects_module_exit(void) 870 { 871 kmem_cache_destroy(slab_objects); 872 } 873 874 int __init i915_objects_module_init(void) 875 { 876 slab_objects = KMEM_CACHE(drm_i915_gem_object, SLAB_HWCACHE_ALIGN); 877 if (!slab_objects) 878 return -ENOMEM; 879 880 return 0; 881 } 882 883 static const struct drm_gem_object_funcs i915_gem_object_funcs = { 884 .free = i915_gem_free_object, 885 .close = i915_gem_close_object, 886 .export = i915_gem_prime_export, 887 }; 888 889 /** 890 * i915_gem_object_get_moving_fence - Get the object's moving fence if any 891 * @obj: The object whose moving fence to get. 892 * @fence: The resulting fence 893 * 894 * A non-signaled moving fence means that there is an async operation 895 * pending on the object that needs to be waited on before setting up 896 * any GPU- or CPU PTEs to the object's pages. 897 * 898 * Return: Negative error code or 0 for success. 899 */ 900 int i915_gem_object_get_moving_fence(struct drm_i915_gem_object *obj, 901 struct dma_fence **fence) 902 { 903 return dma_resv_get_singleton(obj->base.resv, DMA_RESV_USAGE_KERNEL, 904 fence); 905 } 906 907 /** 908 * i915_gem_object_wait_moving_fence - Wait for the object's moving fence if any 909 * @obj: The object whose moving fence to wait for. 910 * @intr: Whether to wait interruptible. 911 * 912 * If the moving fence signaled without an error, it is detached from the 913 * object and put. 914 * 915 * Return: 0 if successful, -ERESTARTSYS if the wait was interrupted, 916 * negative error code if the async operation represented by the 917 * moving fence failed. 918 */ 919 int i915_gem_object_wait_moving_fence(struct drm_i915_gem_object *obj, 920 bool intr) 921 { 922 long ret; 923 924 assert_object_held(obj); 925 926 ret = dma_resv_wait_timeout(obj->base. resv, DMA_RESV_USAGE_KERNEL, 927 intr, MAX_SCHEDULE_TIMEOUT); 928 if (!ret) 929 ret = -ETIME; 930 else if (ret > 0 && i915_gem_object_has_unknown_state(obj)) 931 ret = -EIO; 932 933 return ret < 0 ? ret : 0; 934 } 935 936 /* 937 * i915_gem_object_has_unknown_state - Return true if the object backing pages are 938 * in an unknown_state. This means that userspace must NEVER be allowed to touch 939 * the pages, with either the GPU or CPU. 940 * 941 * ONLY valid to be called after ensuring that all kernel fences have signalled 942 * (in particular the fence for moving/clearing the object). 943 */ 944 bool i915_gem_object_has_unknown_state(struct drm_i915_gem_object *obj) 945 { 946 /* 947 * The below barrier pairs with the dma_fence_signal() in 948 * __memcpy_work(). We should only sample the unknown_state after all 949 * the kernel fences have signalled. 950 */ 951 smp_rmb(); 952 return obj->mm.unknown_state; 953 } 954 955 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 956 #include "selftests/huge_gem_object.c" 957 #include "selftests/huge_pages.c" 958 #include "selftests/i915_gem_migrate.c" 959 #include "selftests/i915_gem_object.c" 960 #include "selftests/i915_gem_coherency.c" 961 #endif 962