1 /* 2 * Copyright © 2016 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/sched/mm.h> 26 #include <linux/dma-fence-array.h> 27 #include <drm/drm_gem.h> 28 29 #include "display/intel_display.h" 30 #include "display/intel_frontbuffer.h" 31 #include "gem/i915_gem_lmem.h" 32 #include "gem/i915_gem_tiling.h" 33 #include "gt/intel_engine.h" 34 #include "gt/intel_engine_heartbeat.h" 35 #include "gt/intel_gt.h" 36 #include "gt/intel_gt_requests.h" 37 38 #include "i915_drv.h" 39 #include "i915_gem_evict.h" 40 #include "i915_sw_fence_work.h" 41 #include "i915_trace.h" 42 #include "i915_vma.h" 43 #include "i915_vma_resource.h" 44 45 static inline void assert_vma_held_evict(const struct i915_vma *vma) 46 { 47 /* 48 * We may be forced to unbind when the vm is dead, to clean it up. 49 * This is the only exception to the requirement of the object lock 50 * being held. 51 */ 52 if (kref_read(&vma->vm->ref)) 53 assert_object_held_shared(vma->obj); 54 } 55 56 static struct kmem_cache *slab_vmas; 57 58 static struct i915_vma *i915_vma_alloc(void) 59 { 60 return kmem_cache_zalloc(slab_vmas, GFP_KERNEL); 61 } 62 63 static void i915_vma_free(struct i915_vma *vma) 64 { 65 return kmem_cache_free(slab_vmas, vma); 66 } 67 68 #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM) 69 70 #include <linux/stackdepot.h> 71 72 static void vma_print_allocator(struct i915_vma *vma, const char *reason) 73 { 74 char buf[512]; 75 76 if (!vma->node.stack) { 77 drm_dbg(&to_i915(vma->obj->base.dev)->drm, 78 "vma.node [%08llx + %08llx] %s: unknown owner\n", 79 vma->node.start, vma->node.size, reason); 80 return; 81 } 82 83 stack_depot_snprint(vma->node.stack, buf, sizeof(buf), 0); 84 drm_dbg(&to_i915(vma->obj->base.dev)->drm, 85 "vma.node [%08llx + %08llx] %s: inserted at %s\n", 86 vma->node.start, vma->node.size, reason, buf); 87 } 88 89 #else 90 91 static void vma_print_allocator(struct i915_vma *vma, const char *reason) 92 { 93 } 94 95 #endif 96 97 static inline struct i915_vma *active_to_vma(struct i915_active *ref) 98 { 99 return container_of(ref, typeof(struct i915_vma), active); 100 } 101 102 static int __i915_vma_active(struct i915_active *ref) 103 { 104 return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT; 105 } 106 107 static void __i915_vma_retire(struct i915_active *ref) 108 { 109 i915_vma_put(active_to_vma(ref)); 110 } 111 112 static struct i915_vma * 113 vma_create(struct drm_i915_gem_object *obj, 114 struct i915_address_space *vm, 115 const struct i915_gtt_view *view) 116 { 117 struct i915_vma *pos = ERR_PTR(-E2BIG); 118 struct i915_vma *vma; 119 struct rb_node *rb, **p; 120 int err; 121 122 /* The aliasing_ppgtt should never be used directly! */ 123 GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm); 124 125 vma = i915_vma_alloc(); 126 if (vma == NULL) 127 return ERR_PTR(-ENOMEM); 128 129 vma->ops = &vm->vma_ops; 130 vma->obj = obj; 131 vma->size = obj->base.size; 132 vma->display_alignment = I915_GTT_MIN_ALIGNMENT; 133 134 i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire, 0); 135 136 /* Declare ourselves safe for use inside shrinkers */ 137 if (IS_ENABLED(CONFIG_LOCKDEP)) { 138 fs_reclaim_acquire(GFP_KERNEL); 139 might_lock(&vma->active.mutex); 140 fs_reclaim_release(GFP_KERNEL); 141 } 142 143 INIT_LIST_HEAD(&vma->closed_link); 144 INIT_LIST_HEAD(&vma->obj_link); 145 RB_CLEAR_NODE(&vma->obj_node); 146 147 if (view && view->type != I915_GTT_VIEW_NORMAL) { 148 vma->gtt_view = *view; 149 if (view->type == I915_GTT_VIEW_PARTIAL) { 150 GEM_BUG_ON(range_overflows_t(u64, 151 view->partial.offset, 152 view->partial.size, 153 obj->base.size >> PAGE_SHIFT)); 154 vma->size = view->partial.size; 155 vma->size <<= PAGE_SHIFT; 156 GEM_BUG_ON(vma->size > obj->base.size); 157 } else if (view->type == I915_GTT_VIEW_ROTATED) { 158 vma->size = intel_rotation_info_size(&view->rotated); 159 vma->size <<= PAGE_SHIFT; 160 } else if (view->type == I915_GTT_VIEW_REMAPPED) { 161 vma->size = intel_remapped_info_size(&view->remapped); 162 vma->size <<= PAGE_SHIFT; 163 } 164 } 165 166 if (unlikely(vma->size > vm->total)) 167 goto err_vma; 168 169 GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE)); 170 171 err = mutex_lock_interruptible(&vm->mutex); 172 if (err) { 173 pos = ERR_PTR(err); 174 goto err_vma; 175 } 176 177 vma->vm = vm; 178 list_add_tail(&vma->vm_link, &vm->unbound_list); 179 180 spin_lock(&obj->vma.lock); 181 if (i915_is_ggtt(vm)) { 182 if (unlikely(overflows_type(vma->size, u32))) 183 goto err_unlock; 184 185 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size, 186 i915_gem_object_get_tiling(obj), 187 i915_gem_object_get_stride(obj)); 188 if (unlikely(vma->fence_size < vma->size || /* overflow */ 189 vma->fence_size > vm->total)) 190 goto err_unlock; 191 192 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT)); 193 194 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size, 195 i915_gem_object_get_tiling(obj), 196 i915_gem_object_get_stride(obj)); 197 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment)); 198 199 __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma)); 200 } 201 202 rb = NULL; 203 p = &obj->vma.tree.rb_node; 204 while (*p) { 205 long cmp; 206 207 rb = *p; 208 pos = rb_entry(rb, struct i915_vma, obj_node); 209 210 /* 211 * If the view already exists in the tree, another thread 212 * already created a matching vma, so return the older instance 213 * and dispose of ours. 214 */ 215 cmp = i915_vma_compare(pos, vm, view); 216 if (cmp < 0) 217 p = &rb->rb_right; 218 else if (cmp > 0) 219 p = &rb->rb_left; 220 else 221 goto err_unlock; 222 } 223 rb_link_node(&vma->obj_node, rb, p); 224 rb_insert_color(&vma->obj_node, &obj->vma.tree); 225 226 if (i915_vma_is_ggtt(vma)) 227 /* 228 * We put the GGTT vma at the start of the vma-list, followed 229 * by the ppGGTT vma. This allows us to break early when 230 * iterating over only the GGTT vma for an object, see 231 * for_each_ggtt_vma() 232 */ 233 list_add(&vma->obj_link, &obj->vma.list); 234 else 235 list_add_tail(&vma->obj_link, &obj->vma.list); 236 237 spin_unlock(&obj->vma.lock); 238 mutex_unlock(&vm->mutex); 239 240 return vma; 241 242 err_unlock: 243 spin_unlock(&obj->vma.lock); 244 list_del_init(&vma->vm_link); 245 mutex_unlock(&vm->mutex); 246 err_vma: 247 i915_vma_free(vma); 248 return pos; 249 } 250 251 static struct i915_vma * 252 i915_vma_lookup(struct drm_i915_gem_object *obj, 253 struct i915_address_space *vm, 254 const struct i915_gtt_view *view) 255 { 256 struct rb_node *rb; 257 258 rb = obj->vma.tree.rb_node; 259 while (rb) { 260 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node); 261 long cmp; 262 263 cmp = i915_vma_compare(vma, vm, view); 264 if (cmp == 0) 265 return vma; 266 267 if (cmp < 0) 268 rb = rb->rb_right; 269 else 270 rb = rb->rb_left; 271 } 272 273 return NULL; 274 } 275 276 /** 277 * i915_vma_instance - return the singleton instance of the VMA 278 * @obj: parent &struct drm_i915_gem_object to be mapped 279 * @vm: address space in which the mapping is located 280 * @view: additional mapping requirements 281 * 282 * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with 283 * the same @view characteristics. If a match is not found, one is created. 284 * Once created, the VMA is kept until either the object is freed, or the 285 * address space is closed. 286 * 287 * Returns the vma, or an error pointer. 288 */ 289 struct i915_vma * 290 i915_vma_instance(struct drm_i915_gem_object *obj, 291 struct i915_address_space *vm, 292 const struct i915_gtt_view *view) 293 { 294 struct i915_vma *vma; 295 296 GEM_BUG_ON(view && !i915_is_ggtt_or_dpt(vm)); 297 GEM_BUG_ON(!kref_read(&vm->ref)); 298 299 spin_lock(&obj->vma.lock); 300 vma = i915_vma_lookup(obj, vm, view); 301 spin_unlock(&obj->vma.lock); 302 303 /* vma_create() will resolve the race if another creates the vma */ 304 if (unlikely(!vma)) 305 vma = vma_create(obj, vm, view); 306 307 GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view)); 308 return vma; 309 } 310 311 struct i915_vma_work { 312 struct dma_fence_work base; 313 struct i915_address_space *vm; 314 struct i915_vm_pt_stash stash; 315 struct i915_vma_resource *vma_res; 316 struct drm_i915_gem_object *obj; 317 struct i915_sw_dma_fence_cb cb; 318 enum i915_cache_level cache_level; 319 unsigned int flags; 320 }; 321 322 static void __vma_bind(struct dma_fence_work *work) 323 { 324 struct i915_vma_work *vw = container_of(work, typeof(*vw), base); 325 struct i915_vma_resource *vma_res = vw->vma_res; 326 327 /* 328 * We are about the bind the object, which must mean we have already 329 * signaled the work to potentially clear/move the pages underneath. If 330 * something went wrong at that stage then the object should have 331 * unknown_state set, in which case we need to skip the bind. 332 */ 333 if (i915_gem_object_has_unknown_state(vw->obj)) 334 return; 335 336 vma_res->ops->bind_vma(vma_res->vm, &vw->stash, 337 vma_res, vw->cache_level, vw->flags); 338 } 339 340 static void __vma_release(struct dma_fence_work *work) 341 { 342 struct i915_vma_work *vw = container_of(work, typeof(*vw), base); 343 344 if (vw->obj) 345 i915_gem_object_put(vw->obj); 346 347 i915_vm_free_pt_stash(vw->vm, &vw->stash); 348 if (vw->vma_res) 349 i915_vma_resource_put(vw->vma_res); 350 } 351 352 static const struct dma_fence_work_ops bind_ops = { 353 .name = "bind", 354 .work = __vma_bind, 355 .release = __vma_release, 356 }; 357 358 struct i915_vma_work *i915_vma_work(void) 359 { 360 struct i915_vma_work *vw; 361 362 vw = kzalloc(sizeof(*vw), GFP_KERNEL); 363 if (!vw) 364 return NULL; 365 366 dma_fence_work_init(&vw->base, &bind_ops); 367 vw->base.dma.error = -EAGAIN; /* disable the worker by default */ 368 369 return vw; 370 } 371 372 int i915_vma_wait_for_bind(struct i915_vma *vma) 373 { 374 int err = 0; 375 376 if (rcu_access_pointer(vma->active.excl.fence)) { 377 struct dma_fence *fence; 378 379 rcu_read_lock(); 380 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence); 381 rcu_read_unlock(); 382 if (fence) { 383 err = dma_fence_wait(fence, true); 384 dma_fence_put(fence); 385 } 386 } 387 388 return err; 389 } 390 391 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) 392 static int i915_vma_verify_bind_complete(struct i915_vma *vma) 393 { 394 struct dma_fence *fence = i915_active_fence_get(&vma->active.excl); 395 int err; 396 397 if (!fence) 398 return 0; 399 400 if (dma_fence_is_signaled(fence)) 401 err = fence->error; 402 else 403 err = -EBUSY; 404 405 dma_fence_put(fence); 406 407 return err; 408 } 409 #else 410 #define i915_vma_verify_bind_complete(_vma) 0 411 #endif 412 413 I915_SELFTEST_EXPORT void 414 i915_vma_resource_init_from_vma(struct i915_vma_resource *vma_res, 415 struct i915_vma *vma) 416 { 417 struct drm_i915_gem_object *obj = vma->obj; 418 419 i915_vma_resource_init(vma_res, vma->vm, vma->pages, &vma->page_sizes, 420 obj->mm.rsgt, i915_gem_object_is_readonly(obj), 421 i915_gem_object_is_lmem(obj), obj->mm.region, 422 vma->ops, vma->private, __i915_vma_offset(vma), 423 __i915_vma_size(vma), vma->size, vma->guard); 424 } 425 426 /** 427 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space. 428 * @vma: VMA to map 429 * @cache_level: mapping cache level 430 * @flags: flags like global or local mapping 431 * @work: preallocated worker for allocating and binding the PTE 432 * @vma_res: pointer to a preallocated vma resource. The resource is either 433 * consumed or freed. 434 * 435 * DMA addresses are taken from the scatter-gather table of this object (or of 436 * this VMA in case of non-default GGTT views) and PTE entries set up. 437 * Note that DMA addresses are also the only part of the SG table we care about. 438 */ 439 int i915_vma_bind(struct i915_vma *vma, 440 enum i915_cache_level cache_level, 441 u32 flags, 442 struct i915_vma_work *work, 443 struct i915_vma_resource *vma_res) 444 { 445 u32 bind_flags; 446 u32 vma_flags; 447 int ret; 448 449 lockdep_assert_held(&vma->vm->mutex); 450 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 451 GEM_BUG_ON(vma->size > i915_vma_size(vma)); 452 453 if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start, 454 vma->node.size, 455 vma->vm->total))) { 456 i915_vma_resource_free(vma_res); 457 return -ENODEV; 458 } 459 460 if (GEM_DEBUG_WARN_ON(!flags)) { 461 i915_vma_resource_free(vma_res); 462 return -EINVAL; 463 } 464 465 bind_flags = flags; 466 bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND; 467 468 vma_flags = atomic_read(&vma->flags); 469 vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND; 470 471 bind_flags &= ~vma_flags; 472 if (bind_flags == 0) { 473 i915_vma_resource_free(vma_res); 474 return 0; 475 } 476 477 GEM_BUG_ON(!atomic_read(&vma->pages_count)); 478 479 /* Wait for or await async unbinds touching our range */ 480 if (work && bind_flags & vma->vm->bind_async_flags) 481 ret = i915_vma_resource_bind_dep_await(vma->vm, 482 &work->base.chain, 483 vma->node.start, 484 vma->node.size, 485 true, 486 GFP_NOWAIT | 487 __GFP_RETRY_MAYFAIL | 488 __GFP_NOWARN); 489 else 490 ret = i915_vma_resource_bind_dep_sync(vma->vm, vma->node.start, 491 vma->node.size, true); 492 if (ret) { 493 i915_vma_resource_free(vma_res); 494 return ret; 495 } 496 497 if (vma->resource || !vma_res) { 498 /* Rebinding with an additional I915_VMA_*_BIND */ 499 GEM_WARN_ON(!vma_flags); 500 i915_vma_resource_free(vma_res); 501 } else { 502 i915_vma_resource_init_from_vma(vma_res, vma); 503 vma->resource = vma_res; 504 } 505 trace_i915_vma_bind(vma, bind_flags); 506 if (work && bind_flags & vma->vm->bind_async_flags) { 507 struct dma_fence *prev; 508 509 work->vma_res = i915_vma_resource_get(vma->resource); 510 work->cache_level = cache_level; 511 work->flags = bind_flags; 512 513 /* 514 * Note we only want to chain up to the migration fence on 515 * the pages (not the object itself). As we don't track that, 516 * yet, we have to use the exclusive fence instead. 517 * 518 * Also note that we do not want to track the async vma as 519 * part of the obj->resv->excl_fence as it only affects 520 * execution and not content or object's backing store lifetime. 521 */ 522 prev = i915_active_set_exclusive(&vma->active, &work->base.dma); 523 if (prev) { 524 __i915_sw_fence_await_dma_fence(&work->base.chain, 525 prev, 526 &work->cb); 527 dma_fence_put(prev); 528 } 529 530 work->base.dma.error = 0; /* enable the queue_work() */ 531 work->obj = i915_gem_object_get(vma->obj); 532 } else { 533 ret = i915_gem_object_wait_moving_fence(vma->obj, true); 534 if (ret) { 535 i915_vma_resource_free(vma->resource); 536 vma->resource = NULL; 537 538 return ret; 539 } 540 vma->ops->bind_vma(vma->vm, NULL, vma->resource, cache_level, 541 bind_flags); 542 } 543 544 atomic_or(bind_flags, &vma->flags); 545 return 0; 546 } 547 548 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma) 549 { 550 void __iomem *ptr; 551 int err; 552 553 if (WARN_ON_ONCE(vma->obj->flags & I915_BO_ALLOC_GPU_ONLY)) 554 return IOMEM_ERR_PTR(-EINVAL); 555 556 GEM_BUG_ON(!i915_vma_is_ggtt(vma)); 557 GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)); 558 GEM_BUG_ON(i915_vma_verify_bind_complete(vma)); 559 560 ptr = READ_ONCE(vma->iomap); 561 if (ptr == NULL) { 562 /* 563 * TODO: consider just using i915_gem_object_pin_map() for lmem 564 * instead, which already supports mapping non-contiguous chunks 565 * of pages, that way we can also drop the 566 * I915_BO_ALLOC_CONTIGUOUS when allocating the object. 567 */ 568 if (i915_gem_object_is_lmem(vma->obj)) { 569 ptr = i915_gem_object_lmem_io_map(vma->obj, 0, 570 vma->obj->base.size); 571 } else if (i915_vma_is_map_and_fenceable(vma)) { 572 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap, 573 i915_vma_offset(vma), 574 i915_vma_size(vma)); 575 } else { 576 ptr = (void __iomem *) 577 i915_gem_object_pin_map(vma->obj, I915_MAP_WC); 578 if (IS_ERR(ptr)) { 579 err = PTR_ERR(ptr); 580 goto err; 581 } 582 ptr = page_pack_bits(ptr, 1); 583 } 584 585 if (ptr == NULL) { 586 err = -ENOMEM; 587 goto err; 588 } 589 590 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) { 591 if (page_unmask_bits(ptr)) 592 __i915_gem_object_release_map(vma->obj); 593 else 594 io_mapping_unmap(ptr); 595 ptr = vma->iomap; 596 } 597 } 598 599 __i915_vma_pin(vma); 600 601 err = i915_vma_pin_fence(vma); 602 if (err) 603 goto err_unpin; 604 605 i915_vma_set_ggtt_write(vma); 606 607 /* NB Access through the GTT requires the device to be awake. */ 608 return page_mask_bits(ptr); 609 610 err_unpin: 611 __i915_vma_unpin(vma); 612 err: 613 return IOMEM_ERR_PTR(err); 614 } 615 616 void i915_vma_flush_writes(struct i915_vma *vma) 617 { 618 if (i915_vma_unset_ggtt_write(vma)) 619 intel_gt_flush_ggtt_writes(vma->vm->gt); 620 } 621 622 void i915_vma_unpin_iomap(struct i915_vma *vma) 623 { 624 GEM_BUG_ON(vma->iomap == NULL); 625 626 /* XXX We keep the mapping until __i915_vma_unbind()/evict() */ 627 628 i915_vma_flush_writes(vma); 629 630 i915_vma_unpin_fence(vma); 631 i915_vma_unpin(vma); 632 } 633 634 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags) 635 { 636 struct i915_vma *vma; 637 struct drm_i915_gem_object *obj; 638 639 vma = fetch_and_zero(p_vma); 640 if (!vma) 641 return; 642 643 obj = vma->obj; 644 GEM_BUG_ON(!obj); 645 646 i915_vma_unpin(vma); 647 648 if (flags & I915_VMA_RELEASE_MAP) 649 i915_gem_object_unpin_map(obj); 650 651 i915_gem_object_put(obj); 652 } 653 654 bool i915_vma_misplaced(const struct i915_vma *vma, 655 u64 size, u64 alignment, u64 flags) 656 { 657 if (!drm_mm_node_allocated(&vma->node)) 658 return false; 659 660 if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma))) 661 return true; 662 663 if (i915_vma_size(vma) < size) 664 return true; 665 666 GEM_BUG_ON(alignment && !is_power_of_2(alignment)); 667 if (alignment && !IS_ALIGNED(i915_vma_offset(vma), alignment)) 668 return true; 669 670 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma)) 671 return true; 672 673 if (flags & PIN_OFFSET_BIAS && 674 i915_vma_offset(vma) < (flags & PIN_OFFSET_MASK)) 675 return true; 676 677 if (flags & PIN_OFFSET_FIXED && 678 i915_vma_offset(vma) != (flags & PIN_OFFSET_MASK)) 679 return true; 680 681 if (flags & PIN_OFFSET_GUARD && 682 vma->guard < (flags & PIN_OFFSET_MASK)) 683 return true; 684 685 return false; 686 } 687 688 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma) 689 { 690 bool mappable, fenceable; 691 692 GEM_BUG_ON(!i915_vma_is_ggtt(vma)); 693 GEM_BUG_ON(!vma->fence_size); 694 695 fenceable = (i915_vma_size(vma) >= vma->fence_size && 696 IS_ALIGNED(i915_vma_offset(vma), vma->fence_alignment)); 697 698 mappable = i915_ggtt_offset(vma) + vma->fence_size <= 699 i915_vm_to_ggtt(vma->vm)->mappable_end; 700 701 if (mappable && fenceable) 702 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma)); 703 else 704 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma)); 705 } 706 707 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color) 708 { 709 struct drm_mm_node *node = &vma->node; 710 struct drm_mm_node *other; 711 712 /* 713 * On some machines we have to be careful when putting differing types 714 * of snoopable memory together to avoid the prefetcher crossing memory 715 * domains and dying. During vm initialisation, we decide whether or not 716 * these constraints apply and set the drm_mm.color_adjust 717 * appropriately. 718 */ 719 if (!i915_vm_has_cache_coloring(vma->vm)) 720 return true; 721 722 /* Only valid to be called on an already inserted vma */ 723 GEM_BUG_ON(!drm_mm_node_allocated(node)); 724 GEM_BUG_ON(list_empty(&node->node_list)); 725 726 other = list_prev_entry(node, node_list); 727 if (i915_node_color_differs(other, color) && 728 !drm_mm_hole_follows(other)) 729 return false; 730 731 other = list_next_entry(node, node_list); 732 if (i915_node_color_differs(other, color) && 733 !drm_mm_hole_follows(node)) 734 return false; 735 736 return true; 737 } 738 739 /** 740 * i915_vma_insert - finds a slot for the vma in its address space 741 * @vma: the vma 742 * @ww: An optional struct i915_gem_ww_ctx 743 * @size: requested size in bytes (can be larger than the VMA) 744 * @alignment: required alignment 745 * @flags: mask of PIN_* flags to use 746 * 747 * First we try to allocate some free space that meets the requirements for 748 * the VMA. Failiing that, if the flags permit, it will evict an old VMA, 749 * preferrably the oldest idle entry to make room for the new VMA. 750 * 751 * Returns: 752 * 0 on success, negative error code otherwise. 753 */ 754 static int 755 i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, 756 u64 size, u64 alignment, u64 flags) 757 { 758 unsigned long color, guard; 759 u64 start, end; 760 int ret; 761 762 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND)); 763 GEM_BUG_ON(drm_mm_node_allocated(&vma->node)); 764 GEM_BUG_ON(hweight64(flags & (PIN_OFFSET_GUARD | PIN_OFFSET_FIXED | PIN_OFFSET_BIAS)) > 1); 765 766 size = max(size, vma->size); 767 alignment = max_t(typeof(alignment), alignment, vma->display_alignment); 768 if (flags & PIN_MAPPABLE) { 769 size = max_t(typeof(size), size, vma->fence_size); 770 alignment = max_t(typeof(alignment), 771 alignment, vma->fence_alignment); 772 } 773 774 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE)); 775 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT)); 776 GEM_BUG_ON(!is_power_of_2(alignment)); 777 778 guard = vma->guard; /* retain guard across rebinds */ 779 if (flags & PIN_OFFSET_GUARD) { 780 GEM_BUG_ON(overflows_type(flags & PIN_OFFSET_MASK, u32)); 781 guard = max_t(u32, guard, flags & PIN_OFFSET_MASK); 782 } 783 /* 784 * As we align the node upon insertion, but the hardware gets 785 * node.start + guard, the easiest way to make that work is 786 * to make the guard a multiple of the alignment size. 787 */ 788 guard = ALIGN(guard, alignment); 789 790 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0; 791 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE)); 792 793 end = vma->vm->total; 794 if (flags & PIN_MAPPABLE) 795 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end); 796 if (flags & PIN_ZONE_4G) 797 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE); 798 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE)); 799 800 alignment = max(alignment, i915_vm_obj_min_alignment(vma->vm, vma->obj)); 801 802 /* 803 * If binding the object/GGTT view requires more space than the entire 804 * aperture has, reject it early before evicting everything in a vain 805 * attempt to find space. 806 */ 807 if (size > end - 2 * guard) { 808 drm_dbg(&to_i915(vma->obj->base.dev)->drm, 809 "Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n", 810 size, flags & PIN_MAPPABLE ? "mappable" : "total", end); 811 return -ENOSPC; 812 } 813 814 color = 0; 815 816 if (i915_vm_has_cache_coloring(vma->vm)) 817 color = vma->obj->cache_level; 818 819 if (flags & PIN_OFFSET_FIXED) { 820 u64 offset = flags & PIN_OFFSET_MASK; 821 if (!IS_ALIGNED(offset, alignment) || 822 range_overflows(offset, size, end)) 823 return -EINVAL; 824 /* 825 * The caller knows not of the guard added by others and 826 * requests for the offset of the start of its buffer 827 * to be fixed, which may not be the same as the position 828 * of the vma->node due to the guard pages. 829 */ 830 if (offset < guard || offset + size > end - guard) 831 return -ENOSPC; 832 833 ret = i915_gem_gtt_reserve(vma->vm, ww, &vma->node, 834 size + 2 * guard, 835 offset - guard, 836 color, flags); 837 if (ret) 838 return ret; 839 } else { 840 size += 2 * guard; 841 /* 842 * We only support huge gtt pages through the 48b PPGTT, 843 * however we also don't want to force any alignment for 844 * objects which need to be tightly packed into the low 32bits. 845 * 846 * Note that we assume that GGTT are limited to 4GiB for the 847 * forseeable future. See also i915_ggtt_offset(). 848 */ 849 if (upper_32_bits(end - 1) && 850 vma->page_sizes.sg > I915_GTT_PAGE_SIZE && 851 !HAS_64K_PAGES(vma->vm->i915)) { 852 /* 853 * We can't mix 64K and 4K PTEs in the same page-table 854 * (2M block), and so to avoid the ugliness and 855 * complexity of coloring we opt for just aligning 64K 856 * objects to 2M. 857 */ 858 u64 page_alignment = 859 rounddown_pow_of_two(vma->page_sizes.sg | 860 I915_GTT_PAGE_SIZE_2M); 861 862 /* 863 * Check we don't expand for the limited Global GTT 864 * (mappable aperture is even more precious!). This 865 * also checks that we exclude the aliasing-ppgtt. 866 */ 867 GEM_BUG_ON(i915_vma_is_ggtt(vma)); 868 869 alignment = max(alignment, page_alignment); 870 871 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K) 872 size = round_up(size, I915_GTT_PAGE_SIZE_2M); 873 } 874 875 ret = i915_gem_gtt_insert(vma->vm, ww, &vma->node, 876 size, alignment, color, 877 start, end, flags); 878 if (ret) 879 return ret; 880 881 GEM_BUG_ON(vma->node.start < start); 882 GEM_BUG_ON(vma->node.start + vma->node.size > end); 883 } 884 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 885 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color)); 886 887 list_move_tail(&vma->vm_link, &vma->vm->bound_list); 888 vma->guard = guard; 889 890 return 0; 891 } 892 893 static void 894 i915_vma_detach(struct i915_vma *vma) 895 { 896 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 897 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND)); 898 899 /* 900 * And finally now the object is completely decoupled from this 901 * vma, we can drop its hold on the backing storage and allow 902 * it to be reaped by the shrinker. 903 */ 904 list_move_tail(&vma->vm_link, &vma->vm->unbound_list); 905 } 906 907 static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) 908 { 909 unsigned int bound; 910 911 bound = atomic_read(&vma->flags); 912 913 if (flags & PIN_VALIDATE) { 914 flags &= I915_VMA_BIND_MASK; 915 916 return (flags & bound) == flags; 917 } 918 919 /* with the lock mandatory for unbind, we don't race here */ 920 flags &= I915_VMA_BIND_MASK; 921 do { 922 if (unlikely(flags & ~bound)) 923 return false; 924 925 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) 926 return false; 927 928 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0); 929 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1)); 930 931 return true; 932 } 933 934 static struct scatterlist * 935 rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset, 936 unsigned int width, unsigned int height, 937 unsigned int src_stride, unsigned int dst_stride, 938 struct sg_table *st, struct scatterlist *sg) 939 { 940 unsigned int column, row; 941 pgoff_t src_idx; 942 943 for (column = 0; column < width; column++) { 944 unsigned int left; 945 946 src_idx = src_stride * (height - 1) + column + offset; 947 for (row = 0; row < height; row++) { 948 st->nents++; 949 /* 950 * We don't need the pages, but need to initialize 951 * the entries so the sg list can be happily traversed. 952 * The only thing we need are DMA addresses. 953 */ 954 sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0); 955 sg_dma_address(sg) = 956 i915_gem_object_get_dma_address(obj, src_idx); 957 sg_dma_len(sg) = I915_GTT_PAGE_SIZE; 958 sg = sg_next(sg); 959 src_idx -= src_stride; 960 } 961 962 left = (dst_stride - height) * I915_GTT_PAGE_SIZE; 963 964 if (!left) 965 continue; 966 967 st->nents++; 968 969 /* 970 * The DE ignores the PTEs for the padding tiles, the sg entry 971 * here is just a conenience to indicate how many padding PTEs 972 * to insert at this spot. 973 */ 974 sg_set_page(sg, NULL, left, 0); 975 sg_dma_address(sg) = 0; 976 sg_dma_len(sg) = left; 977 sg = sg_next(sg); 978 } 979 980 return sg; 981 } 982 983 static noinline struct sg_table * 984 intel_rotate_pages(struct intel_rotation_info *rot_info, 985 struct drm_i915_gem_object *obj) 986 { 987 unsigned int size = intel_rotation_info_size(rot_info); 988 struct drm_i915_private *i915 = to_i915(obj->base.dev); 989 struct sg_table *st; 990 struct scatterlist *sg; 991 int ret = -ENOMEM; 992 int i; 993 994 /* Allocate target SG list. */ 995 st = kmalloc(sizeof(*st), GFP_KERNEL); 996 if (!st) 997 goto err_st_alloc; 998 999 ret = sg_alloc_table(st, size, GFP_KERNEL); 1000 if (ret) 1001 goto err_sg_alloc; 1002 1003 st->nents = 0; 1004 sg = st->sgl; 1005 1006 for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++) 1007 sg = rotate_pages(obj, rot_info->plane[i].offset, 1008 rot_info->plane[i].width, rot_info->plane[i].height, 1009 rot_info->plane[i].src_stride, 1010 rot_info->plane[i].dst_stride, 1011 st, sg); 1012 1013 return st; 1014 1015 err_sg_alloc: 1016 kfree(st); 1017 err_st_alloc: 1018 1019 drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n", 1020 obj->base.size, rot_info->plane[0].width, 1021 rot_info->plane[0].height, size); 1022 1023 return ERR_PTR(ret); 1024 } 1025 1026 static struct scatterlist * 1027 add_padding_pages(unsigned int count, 1028 struct sg_table *st, struct scatterlist *sg) 1029 { 1030 st->nents++; 1031 1032 /* 1033 * The DE ignores the PTEs for the padding tiles, the sg entry 1034 * here is just a convenience to indicate how many padding PTEs 1035 * to insert at this spot. 1036 */ 1037 sg_set_page(sg, NULL, count * I915_GTT_PAGE_SIZE, 0); 1038 sg_dma_address(sg) = 0; 1039 sg_dma_len(sg) = count * I915_GTT_PAGE_SIZE; 1040 sg = sg_next(sg); 1041 1042 return sg; 1043 } 1044 1045 static struct scatterlist * 1046 remap_tiled_color_plane_pages(struct drm_i915_gem_object *obj, 1047 unsigned long offset, unsigned int alignment_pad, 1048 unsigned int width, unsigned int height, 1049 unsigned int src_stride, unsigned int dst_stride, 1050 struct sg_table *st, struct scatterlist *sg, 1051 unsigned int *gtt_offset) 1052 { 1053 unsigned int row; 1054 1055 if (!width || !height) 1056 return sg; 1057 1058 if (alignment_pad) 1059 sg = add_padding_pages(alignment_pad, st, sg); 1060 1061 for (row = 0; row < height; row++) { 1062 unsigned int left = width * I915_GTT_PAGE_SIZE; 1063 1064 while (left) { 1065 dma_addr_t addr; 1066 unsigned int length; 1067 1068 /* 1069 * We don't need the pages, but need to initialize 1070 * the entries so the sg list can be happily traversed. 1071 * The only thing we need are DMA addresses. 1072 */ 1073 1074 addr = i915_gem_object_get_dma_address_len(obj, offset, &length); 1075 1076 length = min(left, length); 1077 1078 st->nents++; 1079 1080 sg_set_page(sg, NULL, length, 0); 1081 sg_dma_address(sg) = addr; 1082 sg_dma_len(sg) = length; 1083 sg = sg_next(sg); 1084 1085 offset += length / I915_GTT_PAGE_SIZE; 1086 left -= length; 1087 } 1088 1089 offset += src_stride - width; 1090 1091 left = (dst_stride - width) * I915_GTT_PAGE_SIZE; 1092 1093 if (!left) 1094 continue; 1095 1096 sg = add_padding_pages(left >> PAGE_SHIFT, st, sg); 1097 } 1098 1099 *gtt_offset += alignment_pad + dst_stride * height; 1100 1101 return sg; 1102 } 1103 1104 static struct scatterlist * 1105 remap_contiguous_pages(struct drm_i915_gem_object *obj, 1106 pgoff_t obj_offset, 1107 unsigned int count, 1108 struct sg_table *st, struct scatterlist *sg) 1109 { 1110 struct scatterlist *iter; 1111 unsigned int offset; 1112 1113 iter = i915_gem_object_get_sg_dma(obj, obj_offset, &offset); 1114 GEM_BUG_ON(!iter); 1115 1116 do { 1117 unsigned int len; 1118 1119 len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT), 1120 count << PAGE_SHIFT); 1121 sg_set_page(sg, NULL, len, 0); 1122 sg_dma_address(sg) = 1123 sg_dma_address(iter) + (offset << PAGE_SHIFT); 1124 sg_dma_len(sg) = len; 1125 1126 st->nents++; 1127 count -= len >> PAGE_SHIFT; 1128 if (count == 0) 1129 return sg; 1130 1131 sg = __sg_next(sg); 1132 iter = __sg_next(iter); 1133 offset = 0; 1134 } while (1); 1135 } 1136 1137 static struct scatterlist * 1138 remap_linear_color_plane_pages(struct drm_i915_gem_object *obj, 1139 pgoff_t obj_offset, unsigned int alignment_pad, 1140 unsigned int size, 1141 struct sg_table *st, struct scatterlist *sg, 1142 unsigned int *gtt_offset) 1143 { 1144 if (!size) 1145 return sg; 1146 1147 if (alignment_pad) 1148 sg = add_padding_pages(alignment_pad, st, sg); 1149 1150 sg = remap_contiguous_pages(obj, obj_offset, size, st, sg); 1151 sg = sg_next(sg); 1152 1153 *gtt_offset += alignment_pad + size; 1154 1155 return sg; 1156 } 1157 1158 static struct scatterlist * 1159 remap_color_plane_pages(const struct intel_remapped_info *rem_info, 1160 struct drm_i915_gem_object *obj, 1161 int color_plane, 1162 struct sg_table *st, struct scatterlist *sg, 1163 unsigned int *gtt_offset) 1164 { 1165 unsigned int alignment_pad = 0; 1166 1167 if (rem_info->plane_alignment) 1168 alignment_pad = ALIGN(*gtt_offset, rem_info->plane_alignment) - *gtt_offset; 1169 1170 if (rem_info->plane[color_plane].linear) 1171 sg = remap_linear_color_plane_pages(obj, 1172 rem_info->plane[color_plane].offset, 1173 alignment_pad, 1174 rem_info->plane[color_plane].size, 1175 st, sg, 1176 gtt_offset); 1177 1178 else 1179 sg = remap_tiled_color_plane_pages(obj, 1180 rem_info->plane[color_plane].offset, 1181 alignment_pad, 1182 rem_info->plane[color_plane].width, 1183 rem_info->plane[color_plane].height, 1184 rem_info->plane[color_plane].src_stride, 1185 rem_info->plane[color_plane].dst_stride, 1186 st, sg, 1187 gtt_offset); 1188 1189 return sg; 1190 } 1191 1192 static noinline struct sg_table * 1193 intel_remap_pages(struct intel_remapped_info *rem_info, 1194 struct drm_i915_gem_object *obj) 1195 { 1196 unsigned int size = intel_remapped_info_size(rem_info); 1197 struct drm_i915_private *i915 = to_i915(obj->base.dev); 1198 struct sg_table *st; 1199 struct scatterlist *sg; 1200 unsigned int gtt_offset = 0; 1201 int ret = -ENOMEM; 1202 int i; 1203 1204 /* Allocate target SG list. */ 1205 st = kmalloc(sizeof(*st), GFP_KERNEL); 1206 if (!st) 1207 goto err_st_alloc; 1208 1209 ret = sg_alloc_table(st, size, GFP_KERNEL); 1210 if (ret) 1211 goto err_sg_alloc; 1212 1213 st->nents = 0; 1214 sg = st->sgl; 1215 1216 for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) 1217 sg = remap_color_plane_pages(rem_info, obj, i, st, sg, >t_offset); 1218 1219 i915_sg_trim(st); 1220 1221 return st; 1222 1223 err_sg_alloc: 1224 kfree(st); 1225 err_st_alloc: 1226 1227 drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n", 1228 obj->base.size, rem_info->plane[0].width, 1229 rem_info->plane[0].height, size); 1230 1231 return ERR_PTR(ret); 1232 } 1233 1234 static noinline struct sg_table * 1235 intel_partial_pages(const struct i915_gtt_view *view, 1236 struct drm_i915_gem_object *obj) 1237 { 1238 struct sg_table *st; 1239 struct scatterlist *sg; 1240 unsigned int count = view->partial.size; 1241 int ret = -ENOMEM; 1242 1243 st = kmalloc(sizeof(*st), GFP_KERNEL); 1244 if (!st) 1245 goto err_st_alloc; 1246 1247 ret = sg_alloc_table(st, count, GFP_KERNEL); 1248 if (ret) 1249 goto err_sg_alloc; 1250 1251 st->nents = 0; 1252 1253 sg = remap_contiguous_pages(obj, view->partial.offset, count, st, st->sgl); 1254 1255 sg_mark_end(sg); 1256 i915_sg_trim(st); /* Drop any unused tail entries. */ 1257 1258 return st; 1259 1260 err_sg_alloc: 1261 kfree(st); 1262 err_st_alloc: 1263 return ERR_PTR(ret); 1264 } 1265 1266 static int 1267 __i915_vma_get_pages(struct i915_vma *vma) 1268 { 1269 struct sg_table *pages; 1270 1271 /* 1272 * The vma->pages are only valid within the lifespan of the borrowed 1273 * obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so 1274 * must be the vma->pages. A simple rule is that vma->pages must only 1275 * be accessed when the obj->mm.pages are pinned. 1276 */ 1277 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj)); 1278 1279 switch (vma->gtt_view.type) { 1280 default: 1281 GEM_BUG_ON(vma->gtt_view.type); 1282 fallthrough; 1283 case I915_GTT_VIEW_NORMAL: 1284 pages = vma->obj->mm.pages; 1285 break; 1286 1287 case I915_GTT_VIEW_ROTATED: 1288 pages = 1289 intel_rotate_pages(&vma->gtt_view.rotated, vma->obj); 1290 break; 1291 1292 case I915_GTT_VIEW_REMAPPED: 1293 pages = 1294 intel_remap_pages(&vma->gtt_view.remapped, vma->obj); 1295 break; 1296 1297 case I915_GTT_VIEW_PARTIAL: 1298 pages = intel_partial_pages(&vma->gtt_view, vma->obj); 1299 break; 1300 } 1301 1302 if (IS_ERR(pages)) { 1303 drm_err(&vma->vm->i915->drm, 1304 "Failed to get pages for VMA view type %u (%ld)!\n", 1305 vma->gtt_view.type, PTR_ERR(pages)); 1306 return PTR_ERR(pages); 1307 } 1308 1309 vma->pages = pages; 1310 1311 return 0; 1312 } 1313 1314 I915_SELFTEST_EXPORT int i915_vma_get_pages(struct i915_vma *vma) 1315 { 1316 int err; 1317 1318 if (atomic_add_unless(&vma->pages_count, 1, 0)) 1319 return 0; 1320 1321 err = i915_gem_object_pin_pages(vma->obj); 1322 if (err) 1323 return err; 1324 1325 err = __i915_vma_get_pages(vma); 1326 if (err) 1327 goto err_unpin; 1328 1329 vma->page_sizes = vma->obj->mm.page_sizes; 1330 atomic_inc(&vma->pages_count); 1331 1332 return 0; 1333 1334 err_unpin: 1335 __i915_gem_object_unpin_pages(vma->obj); 1336 1337 return err; 1338 } 1339 1340 void vma_invalidate_tlb(struct i915_address_space *vm, u32 *tlb) 1341 { 1342 /* 1343 * Before we release the pages that were bound by this vma, we 1344 * must invalidate all the TLBs that may still have a reference 1345 * back to our physical address. It only needs to be done once, 1346 * so after updating the PTE to point away from the pages, record 1347 * the most recent TLB invalidation seqno, and if we have not yet 1348 * flushed the TLBs upon release, perform a full invalidation. 1349 */ 1350 WRITE_ONCE(*tlb, intel_gt_next_invalidate_tlb_full(vm->gt)); 1351 } 1352 1353 static void __vma_put_pages(struct i915_vma *vma, unsigned int count) 1354 { 1355 /* We allocate under vma_get_pages, so beware the shrinker */ 1356 GEM_BUG_ON(atomic_read(&vma->pages_count) < count); 1357 1358 if (atomic_sub_return(count, &vma->pages_count) == 0) { 1359 if (vma->pages != vma->obj->mm.pages) { 1360 sg_free_table(vma->pages); 1361 kfree(vma->pages); 1362 } 1363 vma->pages = NULL; 1364 1365 i915_gem_object_unpin_pages(vma->obj); 1366 } 1367 } 1368 1369 I915_SELFTEST_EXPORT void i915_vma_put_pages(struct i915_vma *vma) 1370 { 1371 if (atomic_add_unless(&vma->pages_count, -1, 1)) 1372 return; 1373 1374 __vma_put_pages(vma, 1); 1375 } 1376 1377 static void vma_unbind_pages(struct i915_vma *vma) 1378 { 1379 unsigned int count; 1380 1381 lockdep_assert_held(&vma->vm->mutex); 1382 1383 /* The upper portion of pages_count is the number of bindings */ 1384 count = atomic_read(&vma->pages_count); 1385 count >>= I915_VMA_PAGES_BIAS; 1386 GEM_BUG_ON(!count); 1387 1388 __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS); 1389 } 1390 1391 int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, 1392 u64 size, u64 alignment, u64 flags) 1393 { 1394 struct i915_vma_work *work = NULL; 1395 struct dma_fence *moving = NULL; 1396 struct i915_vma_resource *vma_res = NULL; 1397 intel_wakeref_t wakeref = 0; 1398 unsigned int bound; 1399 int err; 1400 1401 assert_vma_held(vma); 1402 GEM_BUG_ON(!ww); 1403 1404 BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND); 1405 BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND); 1406 1407 GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL))); 1408 1409 /* First try and grab the pin without rebinding the vma */ 1410 if (try_qad_pin(vma, flags)) 1411 return 0; 1412 1413 err = i915_vma_get_pages(vma); 1414 if (err) 1415 return err; 1416 1417 if (flags & PIN_GLOBAL) 1418 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm); 1419 1420 if (flags & vma->vm->bind_async_flags) { 1421 /* lock VM */ 1422 err = i915_vm_lock_objects(vma->vm, ww); 1423 if (err) 1424 goto err_rpm; 1425 1426 work = i915_vma_work(); 1427 if (!work) { 1428 err = -ENOMEM; 1429 goto err_rpm; 1430 } 1431 1432 work->vm = vma->vm; 1433 1434 err = i915_gem_object_get_moving_fence(vma->obj, &moving); 1435 if (err) 1436 goto err_rpm; 1437 1438 dma_fence_work_chain(&work->base, moving); 1439 1440 /* Allocate enough page directories to used PTE */ 1441 if (vma->vm->allocate_va_range) { 1442 err = i915_vm_alloc_pt_stash(vma->vm, 1443 &work->stash, 1444 vma->size); 1445 if (err) 1446 goto err_fence; 1447 1448 err = i915_vm_map_pt_stash(vma->vm, &work->stash); 1449 if (err) 1450 goto err_fence; 1451 } 1452 } 1453 1454 vma_res = i915_vma_resource_alloc(); 1455 if (IS_ERR(vma_res)) { 1456 err = PTR_ERR(vma_res); 1457 goto err_fence; 1458 } 1459 1460 /* 1461 * Differentiate between user/kernel vma inside the aliasing-ppgtt. 1462 * 1463 * We conflate the Global GTT with the user's vma when using the 1464 * aliasing-ppgtt, but it is still vitally important to try and 1465 * keep the use cases distinct. For example, userptr objects are 1466 * not allowed inside the Global GTT as that will cause lock 1467 * inversions when we have to evict them the mmu_notifier callbacks - 1468 * but they are allowed to be part of the user ppGTT which can never 1469 * be mapped. As such we try to give the distinct users of the same 1470 * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt 1471 * and i915_ppgtt separate]. 1472 * 1473 * NB this may cause us to mask real lock inversions -- while the 1474 * code is safe today, lockdep may not be able to spot future 1475 * transgressions. 1476 */ 1477 err = mutex_lock_interruptible_nested(&vma->vm->mutex, 1478 !(flags & PIN_GLOBAL)); 1479 if (err) 1480 goto err_vma_res; 1481 1482 /* No more allocations allowed now we hold vm->mutex */ 1483 1484 if (unlikely(i915_vma_is_closed(vma))) { 1485 err = -ENOENT; 1486 goto err_unlock; 1487 } 1488 1489 bound = atomic_read(&vma->flags); 1490 if (unlikely(bound & I915_VMA_ERROR)) { 1491 err = -ENOMEM; 1492 goto err_unlock; 1493 } 1494 1495 if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) { 1496 err = -EAGAIN; /* pins are meant to be fairly temporary */ 1497 goto err_unlock; 1498 } 1499 1500 if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) { 1501 if (!(flags & PIN_VALIDATE)) 1502 __i915_vma_pin(vma); 1503 goto err_unlock; 1504 } 1505 1506 err = i915_active_acquire(&vma->active); 1507 if (err) 1508 goto err_unlock; 1509 1510 if (!(bound & I915_VMA_BIND_MASK)) { 1511 err = i915_vma_insert(vma, ww, size, alignment, flags); 1512 if (err) 1513 goto err_active; 1514 1515 if (i915_is_ggtt(vma->vm)) 1516 __i915_vma_set_map_and_fenceable(vma); 1517 } 1518 1519 GEM_BUG_ON(!vma->pages); 1520 err = i915_vma_bind(vma, 1521 vma->obj->cache_level, 1522 flags, work, vma_res); 1523 vma_res = NULL; 1524 if (err) 1525 goto err_remove; 1526 1527 /* There should only be at most 2 active bindings (user, global) */ 1528 GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound); 1529 atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count); 1530 list_move_tail(&vma->vm_link, &vma->vm->bound_list); 1531 1532 if (!(flags & PIN_VALIDATE)) { 1533 __i915_vma_pin(vma); 1534 GEM_BUG_ON(!i915_vma_is_pinned(vma)); 1535 } 1536 GEM_BUG_ON(!i915_vma_is_bound(vma, flags)); 1537 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags)); 1538 1539 err_remove: 1540 if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) { 1541 i915_vma_detach(vma); 1542 drm_mm_remove_node(&vma->node); 1543 } 1544 err_active: 1545 i915_active_release(&vma->active); 1546 err_unlock: 1547 mutex_unlock(&vma->vm->mutex); 1548 err_vma_res: 1549 i915_vma_resource_free(vma_res); 1550 err_fence: 1551 if (work) 1552 dma_fence_work_commit_imm(&work->base); 1553 err_rpm: 1554 if (wakeref) 1555 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref); 1556 1557 if (moving) 1558 dma_fence_put(moving); 1559 1560 i915_vma_put_pages(vma); 1561 return err; 1562 } 1563 1564 static void flush_idle_contexts(struct intel_gt *gt) 1565 { 1566 struct intel_engine_cs *engine; 1567 enum intel_engine_id id; 1568 1569 for_each_engine(engine, gt, id) 1570 intel_engine_flush_barriers(engine); 1571 1572 intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT); 1573 } 1574 1575 static int __i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, 1576 u32 align, unsigned int flags) 1577 { 1578 struct i915_address_space *vm = vma->vm; 1579 struct intel_gt *gt; 1580 struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm); 1581 int err; 1582 1583 do { 1584 err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL); 1585 1586 if (err != -ENOSPC) { 1587 if (!err) { 1588 err = i915_vma_wait_for_bind(vma); 1589 if (err) 1590 i915_vma_unpin(vma); 1591 } 1592 return err; 1593 } 1594 1595 /* Unlike i915_vma_pin, we don't take no for an answer! */ 1596 list_for_each_entry(gt, &ggtt->gt_list, ggtt_link) 1597 flush_idle_contexts(gt); 1598 if (mutex_lock_interruptible(&vm->mutex) == 0) { 1599 /* 1600 * We pass NULL ww here, as we don't want to unbind 1601 * locked objects when called from execbuf when pinning 1602 * is removed. This would probably regress badly. 1603 */ 1604 i915_gem_evict_vm(vm, NULL, NULL); 1605 mutex_unlock(&vm->mutex); 1606 } 1607 } while (1); 1608 } 1609 1610 int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, 1611 u32 align, unsigned int flags) 1612 { 1613 struct i915_gem_ww_ctx _ww; 1614 int err; 1615 1616 GEM_BUG_ON(!i915_vma_is_ggtt(vma)); 1617 1618 if (ww) 1619 return __i915_ggtt_pin(vma, ww, align, flags); 1620 1621 lockdep_assert_not_held(&vma->obj->base.resv->lock.base); 1622 1623 for_i915_gem_ww(&_ww, err, true) { 1624 err = i915_gem_object_lock(vma->obj, &_ww); 1625 if (!err) 1626 err = __i915_ggtt_pin(vma, &_ww, align, flags); 1627 } 1628 1629 return err; 1630 } 1631 1632 static void __vma_close(struct i915_vma *vma, struct intel_gt *gt) 1633 { 1634 /* 1635 * We defer actually closing, unbinding and destroying the VMA until 1636 * the next idle point, or if the object is freed in the meantime. By 1637 * postponing the unbind, we allow for it to be resurrected by the 1638 * client, avoiding the work required to rebind the VMA. This is 1639 * advantageous for DRI, where the client/server pass objects 1640 * between themselves, temporarily opening a local VMA to the 1641 * object, and then closing it again. The same object is then reused 1642 * on the next frame (or two, depending on the depth of the swap queue) 1643 * causing us to rebind the VMA once more. This ends up being a lot 1644 * of wasted work for the steady state. 1645 */ 1646 GEM_BUG_ON(i915_vma_is_closed(vma)); 1647 list_add(&vma->closed_link, >->closed_vma); 1648 } 1649 1650 void i915_vma_close(struct i915_vma *vma) 1651 { 1652 struct intel_gt *gt = vma->vm->gt; 1653 unsigned long flags; 1654 1655 if (i915_vma_is_ggtt(vma)) 1656 return; 1657 1658 GEM_BUG_ON(!atomic_read(&vma->open_count)); 1659 if (atomic_dec_and_lock_irqsave(&vma->open_count, 1660 >->closed_lock, 1661 flags)) { 1662 __vma_close(vma, gt); 1663 spin_unlock_irqrestore(>->closed_lock, flags); 1664 } 1665 } 1666 1667 static void __i915_vma_remove_closed(struct i915_vma *vma) 1668 { 1669 list_del_init(&vma->closed_link); 1670 } 1671 1672 void i915_vma_reopen(struct i915_vma *vma) 1673 { 1674 struct intel_gt *gt = vma->vm->gt; 1675 1676 spin_lock_irq(>->closed_lock); 1677 if (i915_vma_is_closed(vma)) 1678 __i915_vma_remove_closed(vma); 1679 spin_unlock_irq(>->closed_lock); 1680 } 1681 1682 static void force_unbind(struct i915_vma *vma) 1683 { 1684 if (!drm_mm_node_allocated(&vma->node)) 1685 return; 1686 1687 atomic_and(~I915_VMA_PIN_MASK, &vma->flags); 1688 WARN_ON(__i915_vma_unbind(vma)); 1689 GEM_BUG_ON(drm_mm_node_allocated(&vma->node)); 1690 } 1691 1692 static void release_references(struct i915_vma *vma, struct intel_gt *gt, 1693 bool vm_ddestroy) 1694 { 1695 struct drm_i915_gem_object *obj = vma->obj; 1696 1697 GEM_BUG_ON(i915_vma_is_active(vma)); 1698 1699 spin_lock(&obj->vma.lock); 1700 list_del(&vma->obj_link); 1701 if (!RB_EMPTY_NODE(&vma->obj_node)) 1702 rb_erase(&vma->obj_node, &obj->vma.tree); 1703 1704 spin_unlock(&obj->vma.lock); 1705 1706 spin_lock_irq(>->closed_lock); 1707 __i915_vma_remove_closed(vma); 1708 spin_unlock_irq(>->closed_lock); 1709 1710 if (vm_ddestroy) 1711 i915_vm_resv_put(vma->vm); 1712 1713 i915_active_fini(&vma->active); 1714 GEM_WARN_ON(vma->resource); 1715 i915_vma_free(vma); 1716 } 1717 1718 /* 1719 * i915_vma_destroy_locked - Remove all weak reference to the vma and put 1720 * the initial reference. 1721 * 1722 * This function should be called when it's decided the vma isn't needed 1723 * anymore. The caller must assure that it doesn't race with another lookup 1724 * plus destroy, typically by taking an appropriate reference. 1725 * 1726 * Current callsites are 1727 * - __i915_gem_object_pages_fini() 1728 * - __i915_vm_close() - Blocks the above function by taking a reference on 1729 * the object. 1730 * - __i915_vma_parked() - Blocks the above functions by taking a reference 1731 * on the vm and a reference on the object. Also takes the object lock so 1732 * destruction from __i915_vma_parked() can be blocked by holding the 1733 * object lock. Since the object lock is only allowed from within i915 with 1734 * an object refcount, holding the object lock also implicitly blocks the 1735 * vma freeing from __i915_gem_object_pages_fini(). 1736 * 1737 * Because of locks taken during destruction, a vma is also guaranteed to 1738 * stay alive while the following locks are held if it was looked up while 1739 * holding one of the locks: 1740 * - vm->mutex 1741 * - obj->vma.lock 1742 * - gt->closed_lock 1743 */ 1744 void i915_vma_destroy_locked(struct i915_vma *vma) 1745 { 1746 lockdep_assert_held(&vma->vm->mutex); 1747 1748 force_unbind(vma); 1749 list_del_init(&vma->vm_link); 1750 release_references(vma, vma->vm->gt, false); 1751 } 1752 1753 void i915_vma_destroy(struct i915_vma *vma) 1754 { 1755 struct intel_gt *gt; 1756 bool vm_ddestroy; 1757 1758 mutex_lock(&vma->vm->mutex); 1759 force_unbind(vma); 1760 list_del_init(&vma->vm_link); 1761 vm_ddestroy = vma->vm_ddestroy; 1762 vma->vm_ddestroy = false; 1763 1764 /* vma->vm may be freed when releasing vma->vm->mutex. */ 1765 gt = vma->vm->gt; 1766 mutex_unlock(&vma->vm->mutex); 1767 release_references(vma, gt, vm_ddestroy); 1768 } 1769 1770 void i915_vma_parked(struct intel_gt *gt) 1771 { 1772 struct i915_vma *vma, *next; 1773 LIST_HEAD(closed); 1774 1775 spin_lock_irq(>->closed_lock); 1776 list_for_each_entry_safe(vma, next, >->closed_vma, closed_link) { 1777 struct drm_i915_gem_object *obj = vma->obj; 1778 struct i915_address_space *vm = vma->vm; 1779 1780 /* XXX All to avoid keeping a reference on i915_vma itself */ 1781 1782 if (!kref_get_unless_zero(&obj->base.refcount)) 1783 continue; 1784 1785 if (!i915_vm_tryget(vm)) { 1786 i915_gem_object_put(obj); 1787 continue; 1788 } 1789 1790 list_move(&vma->closed_link, &closed); 1791 } 1792 spin_unlock_irq(>->closed_lock); 1793 1794 /* As the GT is held idle, no vma can be reopened as we destroy them */ 1795 list_for_each_entry_safe(vma, next, &closed, closed_link) { 1796 struct drm_i915_gem_object *obj = vma->obj; 1797 struct i915_address_space *vm = vma->vm; 1798 1799 if (i915_gem_object_trylock(obj, NULL)) { 1800 INIT_LIST_HEAD(&vma->closed_link); 1801 i915_vma_destroy(vma); 1802 i915_gem_object_unlock(obj); 1803 } else { 1804 /* back you go.. */ 1805 spin_lock_irq(>->closed_lock); 1806 list_add(&vma->closed_link, >->closed_vma); 1807 spin_unlock_irq(>->closed_lock); 1808 } 1809 1810 i915_gem_object_put(obj); 1811 i915_vm_put(vm); 1812 } 1813 } 1814 1815 static void __i915_vma_iounmap(struct i915_vma *vma) 1816 { 1817 GEM_BUG_ON(i915_vma_is_pinned(vma)); 1818 1819 if (vma->iomap == NULL) 1820 return; 1821 1822 if (page_unmask_bits(vma->iomap)) 1823 __i915_gem_object_release_map(vma->obj); 1824 else 1825 io_mapping_unmap(vma->iomap); 1826 vma->iomap = NULL; 1827 } 1828 1829 void i915_vma_revoke_mmap(struct i915_vma *vma) 1830 { 1831 struct drm_vma_offset_node *node; 1832 u64 vma_offset; 1833 1834 if (!i915_vma_has_userfault(vma)) 1835 return; 1836 1837 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma)); 1838 GEM_BUG_ON(!vma->obj->userfault_count); 1839 1840 node = &vma->mmo->vma_node; 1841 vma_offset = vma->gtt_view.partial.offset << PAGE_SHIFT; 1842 unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping, 1843 drm_vma_node_offset_addr(node) + vma_offset, 1844 vma->size, 1845 1); 1846 1847 i915_vma_unset_userfault(vma); 1848 if (!--vma->obj->userfault_count) 1849 list_del(&vma->obj->userfault_link); 1850 } 1851 1852 static int 1853 __i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma) 1854 { 1855 return __i915_request_await_exclusive(rq, &vma->active); 1856 } 1857 1858 static int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq) 1859 { 1860 int err; 1861 1862 /* Wait for the vma to be bound before we start! */ 1863 err = __i915_request_await_bind(rq, vma); 1864 if (err) 1865 return err; 1866 1867 return i915_active_add_request(&vma->active, rq); 1868 } 1869 1870 int _i915_vma_move_to_active(struct i915_vma *vma, 1871 struct i915_request *rq, 1872 struct dma_fence *fence, 1873 unsigned int flags) 1874 { 1875 struct drm_i915_gem_object *obj = vma->obj; 1876 int err; 1877 1878 assert_object_held(obj); 1879 1880 GEM_BUG_ON(!vma->pages); 1881 1882 if (!(flags & __EXEC_OBJECT_NO_REQUEST_AWAIT)) { 1883 err = i915_request_await_object(rq, vma->obj, flags & EXEC_OBJECT_WRITE); 1884 if (unlikely(err)) 1885 return err; 1886 } 1887 err = __i915_vma_move_to_active(vma, rq); 1888 if (unlikely(err)) 1889 return err; 1890 1891 /* 1892 * Reserve fences slot early to prevent an allocation after preparing 1893 * the workload and associating fences with dma_resv. 1894 */ 1895 if (fence && !(flags & __EXEC_OBJECT_NO_RESERVE)) { 1896 struct dma_fence *curr; 1897 int idx; 1898 1899 dma_fence_array_for_each(curr, idx, fence) 1900 ; 1901 err = dma_resv_reserve_fences(vma->obj->base.resv, idx); 1902 if (unlikely(err)) 1903 return err; 1904 } 1905 1906 if (flags & EXEC_OBJECT_WRITE) { 1907 struct intel_frontbuffer *front; 1908 1909 front = __intel_frontbuffer_get(obj); 1910 if (unlikely(front)) { 1911 if (intel_frontbuffer_invalidate(front, ORIGIN_CS)) 1912 i915_active_add_request(&front->write, rq); 1913 intel_frontbuffer_put(front); 1914 } 1915 } 1916 1917 if (fence) { 1918 struct dma_fence *curr; 1919 enum dma_resv_usage usage; 1920 int idx; 1921 1922 if (flags & EXEC_OBJECT_WRITE) { 1923 usage = DMA_RESV_USAGE_WRITE; 1924 obj->write_domain = I915_GEM_DOMAIN_RENDER; 1925 obj->read_domains = 0; 1926 } else { 1927 usage = DMA_RESV_USAGE_READ; 1928 obj->write_domain = 0; 1929 } 1930 1931 dma_fence_array_for_each(curr, idx, fence) 1932 dma_resv_add_fence(vma->obj->base.resv, curr, usage); 1933 } 1934 1935 if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence) 1936 i915_active_add_request(&vma->fence->active, rq); 1937 1938 obj->read_domains |= I915_GEM_GPU_DOMAINS; 1939 obj->mm.dirty = true; 1940 1941 GEM_BUG_ON(!i915_vma_is_active(vma)); 1942 return 0; 1943 } 1944 1945 struct dma_fence *__i915_vma_evict(struct i915_vma *vma, bool async) 1946 { 1947 struct i915_vma_resource *vma_res = vma->resource; 1948 struct dma_fence *unbind_fence; 1949 1950 GEM_BUG_ON(i915_vma_is_pinned(vma)); 1951 assert_vma_held_evict(vma); 1952 1953 if (i915_vma_is_map_and_fenceable(vma)) { 1954 /* Force a pagefault for domain tracking on next user access */ 1955 i915_vma_revoke_mmap(vma); 1956 1957 /* 1958 * Check that we have flushed all writes through the GGTT 1959 * before the unbind, other due to non-strict nature of those 1960 * indirect writes they may end up referencing the GGTT PTE 1961 * after the unbind. 1962 * 1963 * Note that we may be concurrently poking at the GGTT_WRITE 1964 * bit from set-domain, as we mark all GGTT vma associated 1965 * with an object. We know this is for another vma, as we 1966 * are currently unbinding this one -- so if this vma will be 1967 * reused, it will be refaulted and have its dirty bit set 1968 * before the next write. 1969 */ 1970 i915_vma_flush_writes(vma); 1971 1972 /* release the fence reg _after_ flushing */ 1973 i915_vma_revoke_fence(vma); 1974 1975 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma)); 1976 } 1977 1978 __i915_vma_iounmap(vma); 1979 1980 GEM_BUG_ON(vma->fence); 1981 GEM_BUG_ON(i915_vma_has_userfault(vma)); 1982 1983 /* Object backend must be async capable. */ 1984 GEM_WARN_ON(async && !vma->resource->bi.pages_rsgt); 1985 1986 /* If vm is not open, unbind is a nop. */ 1987 vma_res->needs_wakeref = i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND) && 1988 kref_read(&vma->vm->ref); 1989 vma_res->skip_pte_rewrite = !kref_read(&vma->vm->ref) || 1990 vma->vm->skip_pte_rewrite; 1991 trace_i915_vma_unbind(vma); 1992 1993 if (async) 1994 unbind_fence = i915_vma_resource_unbind(vma_res, 1995 &vma->obj->mm.tlb); 1996 else 1997 unbind_fence = i915_vma_resource_unbind(vma_res, NULL); 1998 1999 vma->resource = NULL; 2000 2001 atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE), 2002 &vma->flags); 2003 2004 i915_vma_detach(vma); 2005 2006 if (!async) { 2007 if (unbind_fence) { 2008 dma_fence_wait(unbind_fence, false); 2009 dma_fence_put(unbind_fence); 2010 unbind_fence = NULL; 2011 } 2012 vma_invalidate_tlb(vma->vm, &vma->obj->mm.tlb); 2013 } 2014 2015 /* 2016 * Binding itself may not have completed until the unbind fence signals, 2017 * so don't drop the pages until that happens, unless the resource is 2018 * async_capable. 2019 */ 2020 2021 vma_unbind_pages(vma); 2022 return unbind_fence; 2023 } 2024 2025 int __i915_vma_unbind(struct i915_vma *vma) 2026 { 2027 int ret; 2028 2029 lockdep_assert_held(&vma->vm->mutex); 2030 assert_vma_held_evict(vma); 2031 2032 if (!drm_mm_node_allocated(&vma->node)) 2033 return 0; 2034 2035 if (i915_vma_is_pinned(vma)) { 2036 vma_print_allocator(vma, "is pinned"); 2037 return -EAGAIN; 2038 } 2039 2040 /* 2041 * After confirming that no one else is pinning this vma, wait for 2042 * any laggards who may have crept in during the wait (through 2043 * a residual pin skipping the vm->mutex) to complete. 2044 */ 2045 ret = i915_vma_sync(vma); 2046 if (ret) 2047 return ret; 2048 2049 GEM_BUG_ON(i915_vma_is_active(vma)); 2050 __i915_vma_evict(vma, false); 2051 2052 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */ 2053 return 0; 2054 } 2055 2056 static struct dma_fence *__i915_vma_unbind_async(struct i915_vma *vma) 2057 { 2058 struct dma_fence *fence; 2059 2060 lockdep_assert_held(&vma->vm->mutex); 2061 2062 if (!drm_mm_node_allocated(&vma->node)) 2063 return NULL; 2064 2065 if (i915_vma_is_pinned(vma) || 2066 &vma->obj->mm.rsgt->table != vma->resource->bi.pages) 2067 return ERR_PTR(-EAGAIN); 2068 2069 /* 2070 * We probably need to replace this with awaiting the fences of the 2071 * object's dma_resv when the vma active goes away. When doing that 2072 * we need to be careful to not add the vma_resource unbind fence 2073 * immediately to the object's dma_resv, because then unbinding 2074 * the next vma from the object, in case there are many, will 2075 * actually await the unbinding of the previous vmas, which is 2076 * undesirable. 2077 */ 2078 if (i915_sw_fence_await_active(&vma->resource->chain, &vma->active, 2079 I915_ACTIVE_AWAIT_EXCL | 2080 I915_ACTIVE_AWAIT_ACTIVE) < 0) { 2081 return ERR_PTR(-EBUSY); 2082 } 2083 2084 fence = __i915_vma_evict(vma, true); 2085 2086 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */ 2087 2088 return fence; 2089 } 2090 2091 int i915_vma_unbind(struct i915_vma *vma) 2092 { 2093 struct i915_address_space *vm = vma->vm; 2094 intel_wakeref_t wakeref = 0; 2095 int err; 2096 2097 assert_object_held_shared(vma->obj); 2098 2099 /* Optimistic wait before taking the mutex */ 2100 err = i915_vma_sync(vma); 2101 if (err) 2102 return err; 2103 2104 if (!drm_mm_node_allocated(&vma->node)) 2105 return 0; 2106 2107 if (i915_vma_is_pinned(vma)) { 2108 vma_print_allocator(vma, "is pinned"); 2109 return -EAGAIN; 2110 } 2111 2112 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)) 2113 /* XXX not always required: nop_clear_range */ 2114 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm); 2115 2116 err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref); 2117 if (err) 2118 goto out_rpm; 2119 2120 err = __i915_vma_unbind(vma); 2121 mutex_unlock(&vm->mutex); 2122 2123 out_rpm: 2124 if (wakeref) 2125 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref); 2126 return err; 2127 } 2128 2129 int i915_vma_unbind_async(struct i915_vma *vma, bool trylock_vm) 2130 { 2131 struct drm_i915_gem_object *obj = vma->obj; 2132 struct i915_address_space *vm = vma->vm; 2133 intel_wakeref_t wakeref = 0; 2134 struct dma_fence *fence; 2135 int err; 2136 2137 /* 2138 * We need the dma-resv lock since we add the 2139 * unbind fence to the dma-resv object. 2140 */ 2141 assert_object_held(obj); 2142 2143 if (!drm_mm_node_allocated(&vma->node)) 2144 return 0; 2145 2146 if (i915_vma_is_pinned(vma)) { 2147 vma_print_allocator(vma, "is pinned"); 2148 return -EAGAIN; 2149 } 2150 2151 if (!obj->mm.rsgt) 2152 return -EBUSY; 2153 2154 err = dma_resv_reserve_fences(obj->base.resv, 2); 2155 if (err) 2156 return -EBUSY; 2157 2158 /* 2159 * It would be great if we could grab this wakeref from the 2160 * async unbind work if needed, but we can't because it uses 2161 * kmalloc and it's in the dma-fence signalling critical path. 2162 */ 2163 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)) 2164 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm); 2165 2166 if (trylock_vm && !mutex_trylock(&vm->mutex)) { 2167 err = -EBUSY; 2168 goto out_rpm; 2169 } else if (!trylock_vm) { 2170 err = mutex_lock_interruptible_nested(&vm->mutex, !wakeref); 2171 if (err) 2172 goto out_rpm; 2173 } 2174 2175 fence = __i915_vma_unbind_async(vma); 2176 mutex_unlock(&vm->mutex); 2177 if (IS_ERR_OR_NULL(fence)) { 2178 err = PTR_ERR_OR_ZERO(fence); 2179 goto out_rpm; 2180 } 2181 2182 dma_resv_add_fence(obj->base.resv, fence, DMA_RESV_USAGE_READ); 2183 dma_fence_put(fence); 2184 2185 out_rpm: 2186 if (wakeref) 2187 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref); 2188 return err; 2189 } 2190 2191 int i915_vma_unbind_unlocked(struct i915_vma *vma) 2192 { 2193 int err; 2194 2195 i915_gem_object_lock(vma->obj, NULL); 2196 err = i915_vma_unbind(vma); 2197 i915_gem_object_unlock(vma->obj); 2198 2199 return err; 2200 } 2201 2202 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma) 2203 { 2204 i915_gem_object_make_unshrinkable(vma->obj); 2205 return vma; 2206 } 2207 2208 void i915_vma_make_shrinkable(struct i915_vma *vma) 2209 { 2210 i915_gem_object_make_shrinkable(vma->obj); 2211 } 2212 2213 void i915_vma_make_purgeable(struct i915_vma *vma) 2214 { 2215 i915_gem_object_make_purgeable(vma->obj); 2216 } 2217 2218 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 2219 #include "selftests/i915_vma.c" 2220 #endif 2221 2222 void i915_vma_module_exit(void) 2223 { 2224 kmem_cache_destroy(slab_vmas); 2225 } 2226 2227 int __init i915_vma_module_init(void) 2228 { 2229 slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN); 2230 if (!slab_vmas) 2231 return -ENOMEM; 2232 2233 return 0; 2234 } 2235