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_frontbuffer.h" 30 #include "gem/i915_gem_lmem.h" 31 #include "gem/i915_gem_tiling.h" 32 #include "gt/intel_engine.h" 33 #include "gt/intel_engine_heartbeat.h" 34 #include "gt/intel_gt.h" 35 #include "gt/intel_gt_requests.h" 36 37 #include "i915_drv.h" 38 #include "i915_gem_evict.h" 39 #include "i915_sw_fence_work.h" 40 #include "i915_trace.h" 41 #include "i915_vma.h" 42 #include "i915_vma_resource.h" 43 44 static inline void assert_vma_held_evict(const struct i915_vma *vma) 45 { 46 /* 47 * We may be forced to unbind when the vm is dead, to clean it up. 48 * This is the only exception to the requirement of the object lock 49 * being held. 50 */ 51 if (kref_read(&vma->vm->ref)) 52 assert_object_held_shared(vma->obj); 53 } 54 55 static struct kmem_cache *slab_vmas; 56 57 static struct i915_vma *i915_vma_alloc(void) 58 { 59 return kmem_cache_zalloc(slab_vmas, GFP_KERNEL); 60 } 61 62 static void i915_vma_free(struct i915_vma *vma) 63 { 64 return kmem_cache_free(slab_vmas, vma); 65 } 66 67 #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM) 68 69 #include <linux/stackdepot.h> 70 71 static void vma_print_allocator(struct i915_vma *vma, const char *reason) 72 { 73 char buf[512]; 74 75 if (!vma->node.stack) { 76 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n", 77 vma->node.start, vma->node.size, reason); 78 return; 79 } 80 81 stack_depot_snprint(vma->node.stack, buf, sizeof(buf), 0); 82 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n", 83 vma->node.start, vma->node.size, reason, buf); 84 } 85 86 #else 87 88 static void vma_print_allocator(struct i915_vma *vma, const char *reason) 89 { 90 } 91 92 #endif 93 94 static inline struct i915_vma *active_to_vma(struct i915_active *ref) 95 { 96 return container_of(ref, typeof(struct i915_vma), active); 97 } 98 99 static int __i915_vma_active(struct i915_active *ref) 100 { 101 return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT; 102 } 103 104 static void __i915_vma_retire(struct i915_active *ref) 105 { 106 i915_vma_put(active_to_vma(ref)); 107 } 108 109 static struct i915_vma * 110 vma_create(struct drm_i915_gem_object *obj, 111 struct i915_address_space *vm, 112 const struct i915_ggtt_view *view) 113 { 114 struct i915_vma *pos = ERR_PTR(-E2BIG); 115 struct i915_vma *vma; 116 struct rb_node *rb, **p; 117 int err; 118 119 /* The aliasing_ppgtt should never be used directly! */ 120 GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm); 121 122 vma = i915_vma_alloc(); 123 if (vma == NULL) 124 return ERR_PTR(-ENOMEM); 125 126 vma->ops = &vm->vma_ops; 127 vma->obj = obj; 128 vma->size = obj->base.size; 129 vma->display_alignment = I915_GTT_MIN_ALIGNMENT; 130 131 i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire, 0); 132 133 /* Declare ourselves safe for use inside shrinkers */ 134 if (IS_ENABLED(CONFIG_LOCKDEP)) { 135 fs_reclaim_acquire(GFP_KERNEL); 136 might_lock(&vma->active.mutex); 137 fs_reclaim_release(GFP_KERNEL); 138 } 139 140 INIT_LIST_HEAD(&vma->closed_link); 141 INIT_LIST_HEAD(&vma->obj_link); 142 RB_CLEAR_NODE(&vma->obj_node); 143 144 if (view && view->type != I915_GGTT_VIEW_NORMAL) { 145 vma->ggtt_view = *view; 146 if (view->type == I915_GGTT_VIEW_PARTIAL) { 147 GEM_BUG_ON(range_overflows_t(u64, 148 view->partial.offset, 149 view->partial.size, 150 obj->base.size >> PAGE_SHIFT)); 151 vma->size = view->partial.size; 152 vma->size <<= PAGE_SHIFT; 153 GEM_BUG_ON(vma->size > obj->base.size); 154 } else if (view->type == I915_GGTT_VIEW_ROTATED) { 155 vma->size = intel_rotation_info_size(&view->rotated); 156 vma->size <<= PAGE_SHIFT; 157 } else if (view->type == I915_GGTT_VIEW_REMAPPED) { 158 vma->size = intel_remapped_info_size(&view->remapped); 159 vma->size <<= PAGE_SHIFT; 160 } 161 } 162 163 if (unlikely(vma->size > vm->total)) 164 goto err_vma; 165 166 GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE)); 167 168 err = mutex_lock_interruptible(&vm->mutex); 169 if (err) { 170 pos = ERR_PTR(err); 171 goto err_vma; 172 } 173 174 vma->vm = vm; 175 list_add_tail(&vma->vm_link, &vm->unbound_list); 176 177 spin_lock(&obj->vma.lock); 178 if (i915_is_ggtt(vm)) { 179 if (unlikely(overflows_type(vma->size, u32))) 180 goto err_unlock; 181 182 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size, 183 i915_gem_object_get_tiling(obj), 184 i915_gem_object_get_stride(obj)); 185 if (unlikely(vma->fence_size < vma->size || /* overflow */ 186 vma->fence_size > vm->total)) 187 goto err_unlock; 188 189 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT)); 190 191 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size, 192 i915_gem_object_get_tiling(obj), 193 i915_gem_object_get_stride(obj)); 194 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment)); 195 196 __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma)); 197 } 198 199 rb = NULL; 200 p = &obj->vma.tree.rb_node; 201 while (*p) { 202 long cmp; 203 204 rb = *p; 205 pos = rb_entry(rb, struct i915_vma, obj_node); 206 207 /* 208 * If the view already exists in the tree, another thread 209 * already created a matching vma, so return the older instance 210 * and dispose of ours. 211 */ 212 cmp = i915_vma_compare(pos, vm, view); 213 if (cmp < 0) 214 p = &rb->rb_right; 215 else if (cmp > 0) 216 p = &rb->rb_left; 217 else 218 goto err_unlock; 219 } 220 rb_link_node(&vma->obj_node, rb, p); 221 rb_insert_color(&vma->obj_node, &obj->vma.tree); 222 223 if (i915_vma_is_ggtt(vma)) 224 /* 225 * We put the GGTT vma at the start of the vma-list, followed 226 * by the ppGGTT vma. This allows us to break early when 227 * iterating over only the GGTT vma for an object, see 228 * for_each_ggtt_vma() 229 */ 230 list_add(&vma->obj_link, &obj->vma.list); 231 else 232 list_add_tail(&vma->obj_link, &obj->vma.list); 233 234 spin_unlock(&obj->vma.lock); 235 mutex_unlock(&vm->mutex); 236 237 return vma; 238 239 err_unlock: 240 spin_unlock(&obj->vma.lock); 241 list_del_init(&vma->vm_link); 242 mutex_unlock(&vm->mutex); 243 err_vma: 244 i915_vma_free(vma); 245 return pos; 246 } 247 248 static struct i915_vma * 249 i915_vma_lookup(struct drm_i915_gem_object *obj, 250 struct i915_address_space *vm, 251 const struct i915_ggtt_view *view) 252 { 253 struct rb_node *rb; 254 255 rb = obj->vma.tree.rb_node; 256 while (rb) { 257 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node); 258 long cmp; 259 260 cmp = i915_vma_compare(vma, vm, view); 261 if (cmp == 0) 262 return vma; 263 264 if (cmp < 0) 265 rb = rb->rb_right; 266 else 267 rb = rb->rb_left; 268 } 269 270 return NULL; 271 } 272 273 /** 274 * i915_vma_instance - return the singleton instance of the VMA 275 * @obj: parent &struct drm_i915_gem_object to be mapped 276 * @vm: address space in which the mapping is located 277 * @view: additional mapping requirements 278 * 279 * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with 280 * the same @view characteristics. If a match is not found, one is created. 281 * Once created, the VMA is kept until either the object is freed, or the 282 * address space is closed. 283 * 284 * Returns the vma, or an error pointer. 285 */ 286 struct i915_vma * 287 i915_vma_instance(struct drm_i915_gem_object *obj, 288 struct i915_address_space *vm, 289 const struct i915_ggtt_view *view) 290 { 291 struct i915_vma *vma; 292 293 GEM_BUG_ON(view && !i915_is_ggtt_or_dpt(vm)); 294 GEM_BUG_ON(!kref_read(&vm->ref)); 295 296 spin_lock(&obj->vma.lock); 297 vma = i915_vma_lookup(obj, vm, view); 298 spin_unlock(&obj->vma.lock); 299 300 /* vma_create() will resolve the race if another creates the vma */ 301 if (unlikely(!vma)) 302 vma = vma_create(obj, vm, view); 303 304 GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view)); 305 return vma; 306 } 307 308 struct i915_vma_work { 309 struct dma_fence_work base; 310 struct i915_address_space *vm; 311 struct i915_vm_pt_stash stash; 312 struct i915_vma_resource *vma_res; 313 struct drm_i915_gem_object *obj; 314 struct i915_sw_dma_fence_cb cb; 315 enum i915_cache_level cache_level; 316 unsigned int flags; 317 }; 318 319 static void __vma_bind(struct dma_fence_work *work) 320 { 321 struct i915_vma_work *vw = container_of(work, typeof(*vw), base); 322 struct i915_vma_resource *vma_res = vw->vma_res; 323 324 /* 325 * We are about the bind the object, which must mean we have already 326 * signaled the work to potentially clear/move the pages underneath. If 327 * something went wrong at that stage then the object should have 328 * unknown_state set, in which case we need to skip the bind. 329 */ 330 if (i915_gem_object_has_unknown_state(vw->obj)) 331 return; 332 333 vma_res->ops->bind_vma(vma_res->vm, &vw->stash, 334 vma_res, vw->cache_level, vw->flags); 335 } 336 337 static void __vma_release(struct dma_fence_work *work) 338 { 339 struct i915_vma_work *vw = container_of(work, typeof(*vw), base); 340 341 if (vw->obj) 342 i915_gem_object_put(vw->obj); 343 344 i915_vm_free_pt_stash(vw->vm, &vw->stash); 345 if (vw->vma_res) 346 i915_vma_resource_put(vw->vma_res); 347 } 348 349 static const struct dma_fence_work_ops bind_ops = { 350 .name = "bind", 351 .work = __vma_bind, 352 .release = __vma_release, 353 }; 354 355 struct i915_vma_work *i915_vma_work(void) 356 { 357 struct i915_vma_work *vw; 358 359 vw = kzalloc(sizeof(*vw), GFP_KERNEL); 360 if (!vw) 361 return NULL; 362 363 dma_fence_work_init(&vw->base, &bind_ops); 364 vw->base.dma.error = -EAGAIN; /* disable the worker by default */ 365 366 return vw; 367 } 368 369 int i915_vma_wait_for_bind(struct i915_vma *vma) 370 { 371 int err = 0; 372 373 if (rcu_access_pointer(vma->active.excl.fence)) { 374 struct dma_fence *fence; 375 376 rcu_read_lock(); 377 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence); 378 rcu_read_unlock(); 379 if (fence) { 380 err = dma_fence_wait(fence, true); 381 dma_fence_put(fence); 382 } 383 } 384 385 return err; 386 } 387 388 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) 389 static int i915_vma_verify_bind_complete(struct i915_vma *vma) 390 { 391 struct dma_fence *fence = i915_active_fence_get(&vma->active.excl); 392 int err; 393 394 if (!fence) 395 return 0; 396 397 if (dma_fence_is_signaled(fence)) 398 err = fence->error; 399 else 400 err = -EBUSY; 401 402 dma_fence_put(fence); 403 404 return err; 405 } 406 #else 407 #define i915_vma_verify_bind_complete(_vma) 0 408 #endif 409 410 I915_SELFTEST_EXPORT void 411 i915_vma_resource_init_from_vma(struct i915_vma_resource *vma_res, 412 struct i915_vma *vma) 413 { 414 struct drm_i915_gem_object *obj = vma->obj; 415 416 i915_vma_resource_init(vma_res, vma->vm, vma->pages, &vma->page_sizes, 417 obj->mm.rsgt, i915_gem_object_is_readonly(obj), 418 i915_gem_object_is_lmem(obj), obj->mm.region, 419 vma->ops, vma->private, vma->node.start, 420 vma->node.size, vma->size); 421 } 422 423 /** 424 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space. 425 * @vma: VMA to map 426 * @cache_level: mapping cache level 427 * @flags: flags like global or local mapping 428 * @work: preallocated worker for allocating and binding the PTE 429 * @vma_res: pointer to a preallocated vma resource. The resource is either 430 * consumed or freed. 431 * 432 * DMA addresses are taken from the scatter-gather table of this object (or of 433 * this VMA in case of non-default GGTT views) and PTE entries set up. 434 * Note that DMA addresses are also the only part of the SG table we care about. 435 */ 436 int i915_vma_bind(struct i915_vma *vma, 437 enum i915_cache_level cache_level, 438 u32 flags, 439 struct i915_vma_work *work, 440 struct i915_vma_resource *vma_res) 441 { 442 u32 bind_flags; 443 u32 vma_flags; 444 int ret; 445 446 lockdep_assert_held(&vma->vm->mutex); 447 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 448 GEM_BUG_ON(vma->size > vma->node.size); 449 450 if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start, 451 vma->node.size, 452 vma->vm->total))) { 453 i915_vma_resource_free(vma_res); 454 return -ENODEV; 455 } 456 457 if (GEM_DEBUG_WARN_ON(!flags)) { 458 i915_vma_resource_free(vma_res); 459 return -EINVAL; 460 } 461 462 bind_flags = flags; 463 bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND; 464 465 vma_flags = atomic_read(&vma->flags); 466 vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND; 467 468 bind_flags &= ~vma_flags; 469 if (bind_flags == 0) { 470 i915_vma_resource_free(vma_res); 471 return 0; 472 } 473 474 GEM_BUG_ON(!atomic_read(&vma->pages_count)); 475 476 /* Wait for or await async unbinds touching our range */ 477 if (work && bind_flags & vma->vm->bind_async_flags) 478 ret = i915_vma_resource_bind_dep_await(vma->vm, 479 &work->base.chain, 480 vma->node.start, 481 vma->node.size, 482 true, 483 GFP_NOWAIT | 484 __GFP_RETRY_MAYFAIL | 485 __GFP_NOWARN); 486 else 487 ret = i915_vma_resource_bind_dep_sync(vma->vm, vma->node.start, 488 vma->node.size, true); 489 if (ret) { 490 i915_vma_resource_free(vma_res); 491 return ret; 492 } 493 494 if (vma->resource || !vma_res) { 495 /* Rebinding with an additional I915_VMA_*_BIND */ 496 GEM_WARN_ON(!vma_flags); 497 i915_vma_resource_free(vma_res); 498 } else { 499 i915_vma_resource_init_from_vma(vma_res, vma); 500 vma->resource = vma_res; 501 } 502 trace_i915_vma_bind(vma, bind_flags); 503 if (work && bind_flags & vma->vm->bind_async_flags) { 504 struct dma_fence *prev; 505 506 work->vma_res = i915_vma_resource_get(vma->resource); 507 work->cache_level = cache_level; 508 work->flags = bind_flags; 509 510 /* 511 * Note we only want to chain up to the migration fence on 512 * the pages (not the object itself). As we don't track that, 513 * yet, we have to use the exclusive fence instead. 514 * 515 * Also note that we do not want to track the async vma as 516 * part of the obj->resv->excl_fence as it only affects 517 * execution and not content or object's backing store lifetime. 518 */ 519 prev = i915_active_set_exclusive(&vma->active, &work->base.dma); 520 if (prev) { 521 __i915_sw_fence_await_dma_fence(&work->base.chain, 522 prev, 523 &work->cb); 524 dma_fence_put(prev); 525 } 526 527 work->base.dma.error = 0; /* enable the queue_work() */ 528 work->obj = i915_gem_object_get(vma->obj); 529 } else { 530 ret = i915_gem_object_wait_moving_fence(vma->obj, true); 531 if (ret) { 532 i915_vma_resource_free(vma->resource); 533 vma->resource = NULL; 534 535 return ret; 536 } 537 vma->ops->bind_vma(vma->vm, NULL, vma->resource, cache_level, 538 bind_flags); 539 } 540 541 atomic_or(bind_flags, &vma->flags); 542 return 0; 543 } 544 545 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma) 546 { 547 void __iomem *ptr; 548 int err; 549 550 if (WARN_ON_ONCE(vma->obj->flags & I915_BO_ALLOC_GPU_ONLY)) 551 return IOMEM_ERR_PTR(-EINVAL); 552 553 GEM_BUG_ON(!i915_vma_is_ggtt(vma)); 554 GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)); 555 GEM_BUG_ON(i915_vma_verify_bind_complete(vma)); 556 557 ptr = READ_ONCE(vma->iomap); 558 if (ptr == NULL) { 559 /* 560 * TODO: consider just using i915_gem_object_pin_map() for lmem 561 * instead, which already supports mapping non-contiguous chunks 562 * of pages, that way we can also drop the 563 * I915_BO_ALLOC_CONTIGUOUS when allocating the object. 564 */ 565 if (i915_gem_object_is_lmem(vma->obj)) { 566 ptr = i915_gem_object_lmem_io_map(vma->obj, 0, 567 vma->obj->base.size); 568 } else if (i915_vma_is_map_and_fenceable(vma)) { 569 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap, 570 vma->node.start, 571 vma->node.size); 572 } else { 573 ptr = (void __iomem *) 574 i915_gem_object_pin_map(vma->obj, I915_MAP_WC); 575 if (IS_ERR(ptr)) { 576 err = PTR_ERR(ptr); 577 goto err; 578 } 579 ptr = page_pack_bits(ptr, 1); 580 } 581 582 if (ptr == NULL) { 583 err = -ENOMEM; 584 goto err; 585 } 586 587 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) { 588 if (page_unmask_bits(ptr)) 589 __i915_gem_object_release_map(vma->obj); 590 else 591 io_mapping_unmap(ptr); 592 ptr = vma->iomap; 593 } 594 } 595 596 __i915_vma_pin(vma); 597 598 err = i915_vma_pin_fence(vma); 599 if (err) 600 goto err_unpin; 601 602 i915_vma_set_ggtt_write(vma); 603 604 /* NB Access through the GTT requires the device to be awake. */ 605 return page_mask_bits(ptr); 606 607 err_unpin: 608 __i915_vma_unpin(vma); 609 err: 610 return IOMEM_ERR_PTR(err); 611 } 612 613 void i915_vma_flush_writes(struct i915_vma *vma) 614 { 615 if (i915_vma_unset_ggtt_write(vma)) 616 intel_gt_flush_ggtt_writes(vma->vm->gt); 617 } 618 619 void i915_vma_unpin_iomap(struct i915_vma *vma) 620 { 621 GEM_BUG_ON(vma->iomap == NULL); 622 623 /* XXX We keep the mapping until __i915_vma_unbind()/evict() */ 624 625 i915_vma_flush_writes(vma); 626 627 i915_vma_unpin_fence(vma); 628 i915_vma_unpin(vma); 629 } 630 631 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags) 632 { 633 struct i915_vma *vma; 634 struct drm_i915_gem_object *obj; 635 636 vma = fetch_and_zero(p_vma); 637 if (!vma) 638 return; 639 640 obj = vma->obj; 641 GEM_BUG_ON(!obj); 642 643 i915_vma_unpin(vma); 644 645 if (flags & I915_VMA_RELEASE_MAP) 646 i915_gem_object_unpin_map(obj); 647 648 i915_gem_object_put(obj); 649 } 650 651 bool i915_vma_misplaced(const struct i915_vma *vma, 652 u64 size, u64 alignment, u64 flags) 653 { 654 if (!drm_mm_node_allocated(&vma->node)) 655 return false; 656 657 if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma))) 658 return true; 659 660 if (vma->node.size < size) 661 return true; 662 663 GEM_BUG_ON(alignment && !is_power_of_2(alignment)); 664 if (alignment && !IS_ALIGNED(vma->node.start, alignment)) 665 return true; 666 667 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma)) 668 return true; 669 670 if (flags & PIN_OFFSET_BIAS && 671 vma->node.start < (flags & PIN_OFFSET_MASK)) 672 return true; 673 674 if (flags & PIN_OFFSET_FIXED && 675 vma->node.start != (flags & PIN_OFFSET_MASK)) 676 return true; 677 678 return false; 679 } 680 681 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma) 682 { 683 bool mappable, fenceable; 684 685 GEM_BUG_ON(!i915_vma_is_ggtt(vma)); 686 GEM_BUG_ON(!vma->fence_size); 687 688 fenceable = (vma->node.size >= vma->fence_size && 689 IS_ALIGNED(vma->node.start, vma->fence_alignment)); 690 691 mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end; 692 693 if (mappable && fenceable) 694 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma)); 695 else 696 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma)); 697 } 698 699 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color) 700 { 701 struct drm_mm_node *node = &vma->node; 702 struct drm_mm_node *other; 703 704 /* 705 * On some machines we have to be careful when putting differing types 706 * of snoopable memory together to avoid the prefetcher crossing memory 707 * domains and dying. During vm initialisation, we decide whether or not 708 * these constraints apply and set the drm_mm.color_adjust 709 * appropriately. 710 */ 711 if (!i915_vm_has_cache_coloring(vma->vm)) 712 return true; 713 714 /* Only valid to be called on an already inserted vma */ 715 GEM_BUG_ON(!drm_mm_node_allocated(node)); 716 GEM_BUG_ON(list_empty(&node->node_list)); 717 718 other = list_prev_entry(node, node_list); 719 if (i915_node_color_differs(other, color) && 720 !drm_mm_hole_follows(other)) 721 return false; 722 723 other = list_next_entry(node, node_list); 724 if (i915_node_color_differs(other, color) && 725 !drm_mm_hole_follows(node)) 726 return false; 727 728 return true; 729 } 730 731 /** 732 * i915_vma_insert - finds a slot for the vma in its address space 733 * @vma: the vma 734 * @size: requested size in bytes (can be larger than the VMA) 735 * @alignment: required alignment 736 * @flags: mask of PIN_* flags to use 737 * 738 * First we try to allocate some free space that meets the requirements for 739 * the VMA. Failiing that, if the flags permit, it will evict an old VMA, 740 * preferrably the oldest idle entry to make room for the new VMA. 741 * 742 * Returns: 743 * 0 on success, negative error code otherwise. 744 */ 745 static int 746 i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, 747 u64 size, u64 alignment, u64 flags) 748 { 749 unsigned long color; 750 u64 start, end; 751 int ret; 752 753 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND)); 754 GEM_BUG_ON(drm_mm_node_allocated(&vma->node)); 755 756 size = max(size, vma->size); 757 alignment = max(alignment, vma->display_alignment); 758 if (flags & PIN_MAPPABLE) { 759 size = max_t(typeof(size), size, vma->fence_size); 760 alignment = max_t(typeof(alignment), 761 alignment, vma->fence_alignment); 762 } 763 764 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE)); 765 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT)); 766 GEM_BUG_ON(!is_power_of_2(alignment)); 767 768 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0; 769 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE)); 770 771 end = vma->vm->total; 772 if (flags & PIN_MAPPABLE) 773 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end); 774 if (flags & PIN_ZONE_4G) 775 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE); 776 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE)); 777 778 alignment = max(alignment, i915_vm_obj_min_alignment(vma->vm, vma->obj)); 779 /* 780 * for compact-pt we round up the reservation to prevent 781 * any smaller pages being used within the same PDE 782 */ 783 if (NEEDS_COMPACT_PT(vma->vm->i915)) 784 size = round_up(size, alignment); 785 786 /* If binding the object/GGTT view requires more space than the entire 787 * aperture has, reject it early before evicting everything in a vain 788 * attempt to find space. 789 */ 790 if (size > end) { 791 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n", 792 size, flags & PIN_MAPPABLE ? "mappable" : "total", 793 end); 794 return -ENOSPC; 795 } 796 797 color = 0; 798 799 if (i915_vm_has_cache_coloring(vma->vm)) 800 color = vma->obj->cache_level; 801 802 if (flags & PIN_OFFSET_FIXED) { 803 u64 offset = flags & PIN_OFFSET_MASK; 804 if (!IS_ALIGNED(offset, alignment) || 805 range_overflows(offset, size, end)) 806 return -EINVAL; 807 808 ret = i915_gem_gtt_reserve(vma->vm, ww, &vma->node, 809 size, offset, color, 810 flags); 811 if (ret) 812 return ret; 813 } else { 814 /* 815 * We only support huge gtt pages through the 48b PPGTT, 816 * however we also don't want to force any alignment for 817 * objects which need to be tightly packed into the low 32bits. 818 * 819 * Note that we assume that GGTT are limited to 4GiB for the 820 * forseeable future. See also i915_ggtt_offset(). 821 */ 822 if (upper_32_bits(end - 1) && 823 vma->page_sizes.sg > I915_GTT_PAGE_SIZE) { 824 /* 825 * We can't mix 64K and 4K PTEs in the same page-table 826 * (2M block), and so to avoid the ugliness and 827 * complexity of coloring we opt for just aligning 64K 828 * objects to 2M. 829 */ 830 u64 page_alignment = 831 rounddown_pow_of_two(vma->page_sizes.sg | 832 I915_GTT_PAGE_SIZE_2M); 833 834 /* 835 * Check we don't expand for the limited Global GTT 836 * (mappable aperture is even more precious!). This 837 * also checks that we exclude the aliasing-ppgtt. 838 */ 839 GEM_BUG_ON(i915_vma_is_ggtt(vma)); 840 841 alignment = max(alignment, page_alignment); 842 843 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K) 844 size = round_up(size, I915_GTT_PAGE_SIZE_2M); 845 } 846 847 ret = i915_gem_gtt_insert(vma->vm, ww, &vma->node, 848 size, alignment, color, 849 start, end, flags); 850 if (ret) 851 return ret; 852 853 GEM_BUG_ON(vma->node.start < start); 854 GEM_BUG_ON(vma->node.start + vma->node.size > end); 855 } 856 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 857 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color)); 858 859 list_move_tail(&vma->vm_link, &vma->vm->bound_list); 860 861 return 0; 862 } 863 864 static void 865 i915_vma_detach(struct i915_vma *vma) 866 { 867 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 868 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND)); 869 870 /* 871 * And finally now the object is completely decoupled from this 872 * vma, we can drop its hold on the backing storage and allow 873 * it to be reaped by the shrinker. 874 */ 875 list_move_tail(&vma->vm_link, &vma->vm->unbound_list); 876 } 877 878 static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) 879 { 880 unsigned int bound; 881 882 bound = atomic_read(&vma->flags); 883 884 if (flags & PIN_VALIDATE) { 885 flags &= I915_VMA_BIND_MASK; 886 887 return (flags & bound) == flags; 888 } 889 890 /* with the lock mandatory for unbind, we don't race here */ 891 flags &= I915_VMA_BIND_MASK; 892 do { 893 if (unlikely(flags & ~bound)) 894 return false; 895 896 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) 897 return false; 898 899 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0); 900 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1)); 901 902 return true; 903 } 904 905 static struct scatterlist * 906 rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset, 907 unsigned int width, unsigned int height, 908 unsigned int src_stride, unsigned int dst_stride, 909 struct sg_table *st, struct scatterlist *sg) 910 { 911 unsigned int column, row; 912 unsigned int src_idx; 913 914 for (column = 0; column < width; column++) { 915 unsigned int left; 916 917 src_idx = src_stride * (height - 1) + column + offset; 918 for (row = 0; row < height; row++) { 919 st->nents++; 920 /* 921 * We don't need the pages, but need to initialize 922 * the entries so the sg list can be happily traversed. 923 * The only thing we need are DMA addresses. 924 */ 925 sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0); 926 sg_dma_address(sg) = 927 i915_gem_object_get_dma_address(obj, src_idx); 928 sg_dma_len(sg) = I915_GTT_PAGE_SIZE; 929 sg = sg_next(sg); 930 src_idx -= src_stride; 931 } 932 933 left = (dst_stride - height) * I915_GTT_PAGE_SIZE; 934 935 if (!left) 936 continue; 937 938 st->nents++; 939 940 /* 941 * The DE ignores the PTEs for the padding tiles, the sg entry 942 * here is just a conenience to indicate how many padding PTEs 943 * to insert at this spot. 944 */ 945 sg_set_page(sg, NULL, left, 0); 946 sg_dma_address(sg) = 0; 947 sg_dma_len(sg) = left; 948 sg = sg_next(sg); 949 } 950 951 return sg; 952 } 953 954 static noinline struct sg_table * 955 intel_rotate_pages(struct intel_rotation_info *rot_info, 956 struct drm_i915_gem_object *obj) 957 { 958 unsigned int size = intel_rotation_info_size(rot_info); 959 struct drm_i915_private *i915 = to_i915(obj->base.dev); 960 struct sg_table *st; 961 struct scatterlist *sg; 962 int ret = -ENOMEM; 963 int i; 964 965 /* Allocate target SG list. */ 966 st = kmalloc(sizeof(*st), GFP_KERNEL); 967 if (!st) 968 goto err_st_alloc; 969 970 ret = sg_alloc_table(st, size, GFP_KERNEL); 971 if (ret) 972 goto err_sg_alloc; 973 974 st->nents = 0; 975 sg = st->sgl; 976 977 for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++) 978 sg = rotate_pages(obj, rot_info->plane[i].offset, 979 rot_info->plane[i].width, rot_info->plane[i].height, 980 rot_info->plane[i].src_stride, 981 rot_info->plane[i].dst_stride, 982 st, sg); 983 984 return st; 985 986 err_sg_alloc: 987 kfree(st); 988 err_st_alloc: 989 990 drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n", 991 obj->base.size, rot_info->plane[0].width, 992 rot_info->plane[0].height, size); 993 994 return ERR_PTR(ret); 995 } 996 997 static struct scatterlist * 998 add_padding_pages(unsigned int count, 999 struct sg_table *st, struct scatterlist *sg) 1000 { 1001 st->nents++; 1002 1003 /* 1004 * The DE ignores the PTEs for the padding tiles, the sg entry 1005 * here is just a convenience to indicate how many padding PTEs 1006 * to insert at this spot. 1007 */ 1008 sg_set_page(sg, NULL, count * I915_GTT_PAGE_SIZE, 0); 1009 sg_dma_address(sg) = 0; 1010 sg_dma_len(sg) = count * I915_GTT_PAGE_SIZE; 1011 sg = sg_next(sg); 1012 1013 return sg; 1014 } 1015 1016 static struct scatterlist * 1017 remap_tiled_color_plane_pages(struct drm_i915_gem_object *obj, 1018 unsigned int offset, unsigned int alignment_pad, 1019 unsigned int width, unsigned int height, 1020 unsigned int src_stride, unsigned int dst_stride, 1021 struct sg_table *st, struct scatterlist *sg, 1022 unsigned int *gtt_offset) 1023 { 1024 unsigned int row; 1025 1026 if (!width || !height) 1027 return sg; 1028 1029 if (alignment_pad) 1030 sg = add_padding_pages(alignment_pad, st, sg); 1031 1032 for (row = 0; row < height; row++) { 1033 unsigned int left = width * I915_GTT_PAGE_SIZE; 1034 1035 while (left) { 1036 dma_addr_t addr; 1037 unsigned int length; 1038 1039 /* 1040 * We don't need the pages, but need to initialize 1041 * the entries so the sg list can be happily traversed. 1042 * The only thing we need are DMA addresses. 1043 */ 1044 1045 addr = i915_gem_object_get_dma_address_len(obj, offset, &length); 1046 1047 length = min(left, length); 1048 1049 st->nents++; 1050 1051 sg_set_page(sg, NULL, length, 0); 1052 sg_dma_address(sg) = addr; 1053 sg_dma_len(sg) = length; 1054 sg = sg_next(sg); 1055 1056 offset += length / I915_GTT_PAGE_SIZE; 1057 left -= length; 1058 } 1059 1060 offset += src_stride - width; 1061 1062 left = (dst_stride - width) * I915_GTT_PAGE_SIZE; 1063 1064 if (!left) 1065 continue; 1066 1067 sg = add_padding_pages(left >> PAGE_SHIFT, st, sg); 1068 } 1069 1070 *gtt_offset += alignment_pad + dst_stride * height; 1071 1072 return sg; 1073 } 1074 1075 static struct scatterlist * 1076 remap_contiguous_pages(struct drm_i915_gem_object *obj, 1077 unsigned int obj_offset, 1078 unsigned int count, 1079 struct sg_table *st, struct scatterlist *sg) 1080 { 1081 struct scatterlist *iter; 1082 unsigned int offset; 1083 1084 iter = i915_gem_object_get_sg_dma(obj, obj_offset, &offset); 1085 GEM_BUG_ON(!iter); 1086 1087 do { 1088 unsigned int len; 1089 1090 len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT), 1091 count << PAGE_SHIFT); 1092 sg_set_page(sg, NULL, len, 0); 1093 sg_dma_address(sg) = 1094 sg_dma_address(iter) + (offset << PAGE_SHIFT); 1095 sg_dma_len(sg) = len; 1096 1097 st->nents++; 1098 count -= len >> PAGE_SHIFT; 1099 if (count == 0) 1100 return sg; 1101 1102 sg = __sg_next(sg); 1103 iter = __sg_next(iter); 1104 offset = 0; 1105 } while (1); 1106 } 1107 1108 static struct scatterlist * 1109 remap_linear_color_plane_pages(struct drm_i915_gem_object *obj, 1110 unsigned int obj_offset, unsigned int alignment_pad, 1111 unsigned int size, 1112 struct sg_table *st, struct scatterlist *sg, 1113 unsigned int *gtt_offset) 1114 { 1115 if (!size) 1116 return sg; 1117 1118 if (alignment_pad) 1119 sg = add_padding_pages(alignment_pad, st, sg); 1120 1121 sg = remap_contiguous_pages(obj, obj_offset, size, st, sg); 1122 sg = sg_next(sg); 1123 1124 *gtt_offset += alignment_pad + size; 1125 1126 return sg; 1127 } 1128 1129 static struct scatterlist * 1130 remap_color_plane_pages(const struct intel_remapped_info *rem_info, 1131 struct drm_i915_gem_object *obj, 1132 int color_plane, 1133 struct sg_table *st, struct scatterlist *sg, 1134 unsigned int *gtt_offset) 1135 { 1136 unsigned int alignment_pad = 0; 1137 1138 if (rem_info->plane_alignment) 1139 alignment_pad = ALIGN(*gtt_offset, rem_info->plane_alignment) - *gtt_offset; 1140 1141 if (rem_info->plane[color_plane].linear) 1142 sg = remap_linear_color_plane_pages(obj, 1143 rem_info->plane[color_plane].offset, 1144 alignment_pad, 1145 rem_info->plane[color_plane].size, 1146 st, sg, 1147 gtt_offset); 1148 1149 else 1150 sg = remap_tiled_color_plane_pages(obj, 1151 rem_info->plane[color_plane].offset, 1152 alignment_pad, 1153 rem_info->plane[color_plane].width, 1154 rem_info->plane[color_plane].height, 1155 rem_info->plane[color_plane].src_stride, 1156 rem_info->plane[color_plane].dst_stride, 1157 st, sg, 1158 gtt_offset); 1159 1160 return sg; 1161 } 1162 1163 static noinline struct sg_table * 1164 intel_remap_pages(struct intel_remapped_info *rem_info, 1165 struct drm_i915_gem_object *obj) 1166 { 1167 unsigned int size = intel_remapped_info_size(rem_info); 1168 struct drm_i915_private *i915 = to_i915(obj->base.dev); 1169 struct sg_table *st; 1170 struct scatterlist *sg; 1171 unsigned int gtt_offset = 0; 1172 int ret = -ENOMEM; 1173 int i; 1174 1175 /* Allocate target SG list. */ 1176 st = kmalloc(sizeof(*st), GFP_KERNEL); 1177 if (!st) 1178 goto err_st_alloc; 1179 1180 ret = sg_alloc_table(st, size, GFP_KERNEL); 1181 if (ret) 1182 goto err_sg_alloc; 1183 1184 st->nents = 0; 1185 sg = st->sgl; 1186 1187 for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) 1188 sg = remap_color_plane_pages(rem_info, obj, i, st, sg, >t_offset); 1189 1190 i915_sg_trim(st); 1191 1192 return st; 1193 1194 err_sg_alloc: 1195 kfree(st); 1196 err_st_alloc: 1197 1198 drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n", 1199 obj->base.size, rem_info->plane[0].width, 1200 rem_info->plane[0].height, size); 1201 1202 return ERR_PTR(ret); 1203 } 1204 1205 static noinline struct sg_table * 1206 intel_partial_pages(const struct i915_ggtt_view *view, 1207 struct drm_i915_gem_object *obj) 1208 { 1209 struct sg_table *st; 1210 struct scatterlist *sg; 1211 unsigned int count = view->partial.size; 1212 int ret = -ENOMEM; 1213 1214 st = kmalloc(sizeof(*st), GFP_KERNEL); 1215 if (!st) 1216 goto err_st_alloc; 1217 1218 ret = sg_alloc_table(st, count, GFP_KERNEL); 1219 if (ret) 1220 goto err_sg_alloc; 1221 1222 st->nents = 0; 1223 1224 sg = remap_contiguous_pages(obj, view->partial.offset, count, st, st->sgl); 1225 1226 sg_mark_end(sg); 1227 i915_sg_trim(st); /* Drop any unused tail entries. */ 1228 1229 return st; 1230 1231 err_sg_alloc: 1232 kfree(st); 1233 err_st_alloc: 1234 return ERR_PTR(ret); 1235 } 1236 1237 static int 1238 __i915_vma_get_pages(struct i915_vma *vma) 1239 { 1240 struct sg_table *pages; 1241 1242 /* 1243 * The vma->pages are only valid within the lifespan of the borrowed 1244 * obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so 1245 * must be the vma->pages. A simple rule is that vma->pages must only 1246 * be accessed when the obj->mm.pages are pinned. 1247 */ 1248 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj)); 1249 1250 switch (vma->ggtt_view.type) { 1251 default: 1252 GEM_BUG_ON(vma->ggtt_view.type); 1253 fallthrough; 1254 case I915_GGTT_VIEW_NORMAL: 1255 pages = vma->obj->mm.pages; 1256 break; 1257 1258 case I915_GGTT_VIEW_ROTATED: 1259 pages = 1260 intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj); 1261 break; 1262 1263 case I915_GGTT_VIEW_REMAPPED: 1264 pages = 1265 intel_remap_pages(&vma->ggtt_view.remapped, vma->obj); 1266 break; 1267 1268 case I915_GGTT_VIEW_PARTIAL: 1269 pages = intel_partial_pages(&vma->ggtt_view, vma->obj); 1270 break; 1271 } 1272 1273 if (IS_ERR(pages)) { 1274 drm_err(&vma->vm->i915->drm, 1275 "Failed to get pages for VMA view type %u (%ld)!\n", 1276 vma->ggtt_view.type, PTR_ERR(pages)); 1277 return PTR_ERR(pages); 1278 } 1279 1280 vma->pages = pages; 1281 1282 return 0; 1283 } 1284 1285 I915_SELFTEST_EXPORT int i915_vma_get_pages(struct i915_vma *vma) 1286 { 1287 int err; 1288 1289 if (atomic_add_unless(&vma->pages_count, 1, 0)) 1290 return 0; 1291 1292 err = i915_gem_object_pin_pages(vma->obj); 1293 if (err) 1294 return err; 1295 1296 err = __i915_vma_get_pages(vma); 1297 if (err) 1298 goto err_unpin; 1299 1300 vma->page_sizes = vma->obj->mm.page_sizes; 1301 atomic_inc(&vma->pages_count); 1302 1303 return 0; 1304 1305 err_unpin: 1306 __i915_gem_object_unpin_pages(vma->obj); 1307 1308 return err; 1309 } 1310 1311 void vma_invalidate_tlb(struct i915_address_space *vm, u32 *tlb) 1312 { 1313 /* 1314 * Before we release the pages that were bound by this vma, we 1315 * must invalidate all the TLBs that may still have a reference 1316 * back to our physical address. It only needs to be done once, 1317 * so after updating the PTE to point away from the pages, record 1318 * the most recent TLB invalidation seqno, and if we have not yet 1319 * flushed the TLBs upon release, perform a full invalidation. 1320 */ 1321 WRITE_ONCE(*tlb, intel_gt_next_invalidate_tlb_full(vm->gt)); 1322 } 1323 1324 static void __vma_put_pages(struct i915_vma *vma, unsigned int count) 1325 { 1326 /* We allocate under vma_get_pages, so beware the shrinker */ 1327 GEM_BUG_ON(atomic_read(&vma->pages_count) < count); 1328 1329 if (atomic_sub_return(count, &vma->pages_count) == 0) { 1330 if (vma->pages != vma->obj->mm.pages) { 1331 sg_free_table(vma->pages); 1332 kfree(vma->pages); 1333 } 1334 vma->pages = NULL; 1335 1336 i915_gem_object_unpin_pages(vma->obj); 1337 } 1338 } 1339 1340 I915_SELFTEST_EXPORT void i915_vma_put_pages(struct i915_vma *vma) 1341 { 1342 if (atomic_add_unless(&vma->pages_count, -1, 1)) 1343 return; 1344 1345 __vma_put_pages(vma, 1); 1346 } 1347 1348 static void vma_unbind_pages(struct i915_vma *vma) 1349 { 1350 unsigned int count; 1351 1352 lockdep_assert_held(&vma->vm->mutex); 1353 1354 /* The upper portion of pages_count is the number of bindings */ 1355 count = atomic_read(&vma->pages_count); 1356 count >>= I915_VMA_PAGES_BIAS; 1357 GEM_BUG_ON(!count); 1358 1359 __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS); 1360 } 1361 1362 int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, 1363 u64 size, u64 alignment, u64 flags) 1364 { 1365 struct i915_vma_work *work = NULL; 1366 struct dma_fence *moving = NULL; 1367 struct i915_vma_resource *vma_res = NULL; 1368 intel_wakeref_t wakeref = 0; 1369 unsigned int bound; 1370 int err; 1371 1372 assert_vma_held(vma); 1373 GEM_BUG_ON(!ww); 1374 1375 BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND); 1376 BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND); 1377 1378 GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL))); 1379 1380 /* First try and grab the pin without rebinding the vma */ 1381 if (try_qad_pin(vma, flags)) 1382 return 0; 1383 1384 err = i915_vma_get_pages(vma); 1385 if (err) 1386 return err; 1387 1388 if (flags & PIN_GLOBAL) 1389 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm); 1390 1391 if (flags & vma->vm->bind_async_flags) { 1392 /* lock VM */ 1393 err = i915_vm_lock_objects(vma->vm, ww); 1394 if (err) 1395 goto err_rpm; 1396 1397 work = i915_vma_work(); 1398 if (!work) { 1399 err = -ENOMEM; 1400 goto err_rpm; 1401 } 1402 1403 work->vm = vma->vm; 1404 1405 err = i915_gem_object_get_moving_fence(vma->obj, &moving); 1406 if (err) 1407 goto err_rpm; 1408 1409 dma_fence_work_chain(&work->base, moving); 1410 1411 /* Allocate enough page directories to used PTE */ 1412 if (vma->vm->allocate_va_range) { 1413 err = i915_vm_alloc_pt_stash(vma->vm, 1414 &work->stash, 1415 vma->size); 1416 if (err) 1417 goto err_fence; 1418 1419 err = i915_vm_map_pt_stash(vma->vm, &work->stash); 1420 if (err) 1421 goto err_fence; 1422 } 1423 } 1424 1425 vma_res = i915_vma_resource_alloc(); 1426 if (IS_ERR(vma_res)) { 1427 err = PTR_ERR(vma_res); 1428 goto err_fence; 1429 } 1430 1431 /* 1432 * Differentiate between user/kernel vma inside the aliasing-ppgtt. 1433 * 1434 * We conflate the Global GTT with the user's vma when using the 1435 * aliasing-ppgtt, but it is still vitally important to try and 1436 * keep the use cases distinct. For example, userptr objects are 1437 * not allowed inside the Global GTT as that will cause lock 1438 * inversions when we have to evict them the mmu_notifier callbacks - 1439 * but they are allowed to be part of the user ppGTT which can never 1440 * be mapped. As such we try to give the distinct users of the same 1441 * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt 1442 * and i915_ppgtt separate]. 1443 * 1444 * NB this may cause us to mask real lock inversions -- while the 1445 * code is safe today, lockdep may not be able to spot future 1446 * transgressions. 1447 */ 1448 err = mutex_lock_interruptible_nested(&vma->vm->mutex, 1449 !(flags & PIN_GLOBAL)); 1450 if (err) 1451 goto err_vma_res; 1452 1453 /* No more allocations allowed now we hold vm->mutex */ 1454 1455 if (unlikely(i915_vma_is_closed(vma))) { 1456 err = -ENOENT; 1457 goto err_unlock; 1458 } 1459 1460 bound = atomic_read(&vma->flags); 1461 if (unlikely(bound & I915_VMA_ERROR)) { 1462 err = -ENOMEM; 1463 goto err_unlock; 1464 } 1465 1466 if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) { 1467 err = -EAGAIN; /* pins are meant to be fairly temporary */ 1468 goto err_unlock; 1469 } 1470 1471 if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) { 1472 if (!(flags & PIN_VALIDATE)) 1473 __i915_vma_pin(vma); 1474 goto err_unlock; 1475 } 1476 1477 err = i915_active_acquire(&vma->active); 1478 if (err) 1479 goto err_unlock; 1480 1481 if (!(bound & I915_VMA_BIND_MASK)) { 1482 err = i915_vma_insert(vma, ww, size, alignment, flags); 1483 if (err) 1484 goto err_active; 1485 1486 if (i915_is_ggtt(vma->vm)) 1487 __i915_vma_set_map_and_fenceable(vma); 1488 } 1489 1490 GEM_BUG_ON(!vma->pages); 1491 err = i915_vma_bind(vma, 1492 vma->obj->cache_level, 1493 flags, work, vma_res); 1494 vma_res = NULL; 1495 if (err) 1496 goto err_remove; 1497 1498 /* There should only be at most 2 active bindings (user, global) */ 1499 GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound); 1500 atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count); 1501 list_move_tail(&vma->vm_link, &vma->vm->bound_list); 1502 1503 if (!(flags & PIN_VALIDATE)) { 1504 __i915_vma_pin(vma); 1505 GEM_BUG_ON(!i915_vma_is_pinned(vma)); 1506 } 1507 GEM_BUG_ON(!i915_vma_is_bound(vma, flags)); 1508 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags)); 1509 1510 err_remove: 1511 if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) { 1512 i915_vma_detach(vma); 1513 drm_mm_remove_node(&vma->node); 1514 } 1515 err_active: 1516 i915_active_release(&vma->active); 1517 err_unlock: 1518 mutex_unlock(&vma->vm->mutex); 1519 err_vma_res: 1520 i915_vma_resource_free(vma_res); 1521 err_fence: 1522 if (work) 1523 dma_fence_work_commit_imm(&work->base); 1524 err_rpm: 1525 if (wakeref) 1526 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref); 1527 1528 if (moving) 1529 dma_fence_put(moving); 1530 1531 i915_vma_put_pages(vma); 1532 return err; 1533 } 1534 1535 static void flush_idle_contexts(struct intel_gt *gt) 1536 { 1537 struct intel_engine_cs *engine; 1538 enum intel_engine_id id; 1539 1540 for_each_engine(engine, gt, id) 1541 intel_engine_flush_barriers(engine); 1542 1543 intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT); 1544 } 1545 1546 static int __i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, 1547 u32 align, unsigned int flags) 1548 { 1549 struct i915_address_space *vm = vma->vm; 1550 int err; 1551 1552 do { 1553 err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL); 1554 1555 if (err != -ENOSPC) { 1556 if (!err) { 1557 err = i915_vma_wait_for_bind(vma); 1558 if (err) 1559 i915_vma_unpin(vma); 1560 } 1561 return err; 1562 } 1563 1564 /* Unlike i915_vma_pin, we don't take no for an answer! */ 1565 flush_idle_contexts(vm->gt); 1566 if (mutex_lock_interruptible(&vm->mutex) == 0) { 1567 /* 1568 * We pass NULL ww here, as we don't want to unbind 1569 * locked objects when called from execbuf when pinning 1570 * is removed. This would probably regress badly. 1571 */ 1572 i915_gem_evict_vm(vm, NULL); 1573 mutex_unlock(&vm->mutex); 1574 } 1575 } while (1); 1576 } 1577 1578 int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, 1579 u32 align, unsigned int flags) 1580 { 1581 struct i915_gem_ww_ctx _ww; 1582 int err; 1583 1584 GEM_BUG_ON(!i915_vma_is_ggtt(vma)); 1585 1586 if (ww) 1587 return __i915_ggtt_pin(vma, ww, align, flags); 1588 1589 lockdep_assert_not_held(&vma->obj->base.resv->lock.base); 1590 1591 for_i915_gem_ww(&_ww, err, true) { 1592 err = i915_gem_object_lock(vma->obj, &_ww); 1593 if (!err) 1594 err = __i915_ggtt_pin(vma, &_ww, align, flags); 1595 } 1596 1597 return err; 1598 } 1599 1600 static void __vma_close(struct i915_vma *vma, struct intel_gt *gt) 1601 { 1602 /* 1603 * We defer actually closing, unbinding and destroying the VMA until 1604 * the next idle point, or if the object is freed in the meantime. By 1605 * postponing the unbind, we allow for it to be resurrected by the 1606 * client, avoiding the work required to rebind the VMA. This is 1607 * advantageous for DRI, where the client/server pass objects 1608 * between themselves, temporarily opening a local VMA to the 1609 * object, and then closing it again. The same object is then reused 1610 * on the next frame (or two, depending on the depth of the swap queue) 1611 * causing us to rebind the VMA once more. This ends up being a lot 1612 * of wasted work for the steady state. 1613 */ 1614 GEM_BUG_ON(i915_vma_is_closed(vma)); 1615 list_add(&vma->closed_link, >->closed_vma); 1616 } 1617 1618 void i915_vma_close(struct i915_vma *vma) 1619 { 1620 struct intel_gt *gt = vma->vm->gt; 1621 unsigned long flags; 1622 1623 if (i915_vma_is_ggtt(vma)) 1624 return; 1625 1626 GEM_BUG_ON(!atomic_read(&vma->open_count)); 1627 if (atomic_dec_and_lock_irqsave(&vma->open_count, 1628 >->closed_lock, 1629 flags)) { 1630 __vma_close(vma, gt); 1631 spin_unlock_irqrestore(>->closed_lock, flags); 1632 } 1633 } 1634 1635 static void __i915_vma_remove_closed(struct i915_vma *vma) 1636 { 1637 list_del_init(&vma->closed_link); 1638 } 1639 1640 void i915_vma_reopen(struct i915_vma *vma) 1641 { 1642 struct intel_gt *gt = vma->vm->gt; 1643 1644 spin_lock_irq(>->closed_lock); 1645 if (i915_vma_is_closed(vma)) 1646 __i915_vma_remove_closed(vma); 1647 spin_unlock_irq(>->closed_lock); 1648 } 1649 1650 static void force_unbind(struct i915_vma *vma) 1651 { 1652 if (!drm_mm_node_allocated(&vma->node)) 1653 return; 1654 1655 atomic_and(~I915_VMA_PIN_MASK, &vma->flags); 1656 WARN_ON(__i915_vma_unbind(vma)); 1657 GEM_BUG_ON(drm_mm_node_allocated(&vma->node)); 1658 } 1659 1660 static void release_references(struct i915_vma *vma, struct intel_gt *gt, 1661 bool vm_ddestroy) 1662 { 1663 struct drm_i915_gem_object *obj = vma->obj; 1664 1665 GEM_BUG_ON(i915_vma_is_active(vma)); 1666 1667 spin_lock(&obj->vma.lock); 1668 list_del(&vma->obj_link); 1669 if (!RB_EMPTY_NODE(&vma->obj_node)) 1670 rb_erase(&vma->obj_node, &obj->vma.tree); 1671 1672 spin_unlock(&obj->vma.lock); 1673 1674 spin_lock_irq(>->closed_lock); 1675 __i915_vma_remove_closed(vma); 1676 spin_unlock_irq(>->closed_lock); 1677 1678 if (vm_ddestroy) 1679 i915_vm_resv_put(vma->vm); 1680 1681 i915_active_fini(&vma->active); 1682 GEM_WARN_ON(vma->resource); 1683 i915_vma_free(vma); 1684 } 1685 1686 /** 1687 * i915_vma_destroy_locked - Remove all weak reference to the vma and put 1688 * the initial reference. 1689 * 1690 * This function should be called when it's decided the vma isn't needed 1691 * anymore. The caller must assure that it doesn't race with another lookup 1692 * plus destroy, typically by taking an appropriate reference. 1693 * 1694 * Current callsites are 1695 * - __i915_gem_object_pages_fini() 1696 * - __i915_vm_close() - Blocks the above function by taking a reference on 1697 * the object. 1698 * - __i915_vma_parked() - Blocks the above functions by taking a reference 1699 * on the vm and a reference on the object. Also takes the object lock so 1700 * destruction from __i915_vma_parked() can be blocked by holding the 1701 * object lock. Since the object lock is only allowed from within i915 with 1702 * an object refcount, holding the object lock also implicitly blocks the 1703 * vma freeing from __i915_gem_object_pages_fini(). 1704 * 1705 * Because of locks taken during destruction, a vma is also guaranteed to 1706 * stay alive while the following locks are held if it was looked up while 1707 * holding one of the locks: 1708 * - vm->mutex 1709 * - obj->vma.lock 1710 * - gt->closed_lock 1711 */ 1712 void i915_vma_destroy_locked(struct i915_vma *vma) 1713 { 1714 lockdep_assert_held(&vma->vm->mutex); 1715 1716 force_unbind(vma); 1717 list_del_init(&vma->vm_link); 1718 release_references(vma, vma->vm->gt, false); 1719 } 1720 1721 void i915_vma_destroy(struct i915_vma *vma) 1722 { 1723 struct intel_gt *gt; 1724 bool vm_ddestroy; 1725 1726 mutex_lock(&vma->vm->mutex); 1727 force_unbind(vma); 1728 list_del_init(&vma->vm_link); 1729 vm_ddestroy = vma->vm_ddestroy; 1730 vma->vm_ddestroy = false; 1731 1732 /* vma->vm may be freed when releasing vma->vm->mutex. */ 1733 gt = vma->vm->gt; 1734 mutex_unlock(&vma->vm->mutex); 1735 release_references(vma, gt, vm_ddestroy); 1736 } 1737 1738 void i915_vma_parked(struct intel_gt *gt) 1739 { 1740 struct i915_vma *vma, *next; 1741 LIST_HEAD(closed); 1742 1743 spin_lock_irq(>->closed_lock); 1744 list_for_each_entry_safe(vma, next, >->closed_vma, closed_link) { 1745 struct drm_i915_gem_object *obj = vma->obj; 1746 struct i915_address_space *vm = vma->vm; 1747 1748 /* XXX All to avoid keeping a reference on i915_vma itself */ 1749 1750 if (!kref_get_unless_zero(&obj->base.refcount)) 1751 continue; 1752 1753 if (!i915_vm_tryget(vm)) { 1754 i915_gem_object_put(obj); 1755 continue; 1756 } 1757 1758 list_move(&vma->closed_link, &closed); 1759 } 1760 spin_unlock_irq(>->closed_lock); 1761 1762 /* As the GT is held idle, no vma can be reopened as we destroy them */ 1763 list_for_each_entry_safe(vma, next, &closed, closed_link) { 1764 struct drm_i915_gem_object *obj = vma->obj; 1765 struct i915_address_space *vm = vma->vm; 1766 1767 if (i915_gem_object_trylock(obj, NULL)) { 1768 INIT_LIST_HEAD(&vma->closed_link); 1769 i915_vma_destroy(vma); 1770 i915_gem_object_unlock(obj); 1771 } else { 1772 /* back you go.. */ 1773 spin_lock_irq(>->closed_lock); 1774 list_add(&vma->closed_link, >->closed_vma); 1775 spin_unlock_irq(>->closed_lock); 1776 } 1777 1778 i915_gem_object_put(obj); 1779 i915_vm_put(vm); 1780 } 1781 } 1782 1783 static void __i915_vma_iounmap(struct i915_vma *vma) 1784 { 1785 GEM_BUG_ON(i915_vma_is_pinned(vma)); 1786 1787 if (vma->iomap == NULL) 1788 return; 1789 1790 if (page_unmask_bits(vma->iomap)) 1791 __i915_gem_object_release_map(vma->obj); 1792 else 1793 io_mapping_unmap(vma->iomap); 1794 vma->iomap = NULL; 1795 } 1796 1797 void i915_vma_revoke_mmap(struct i915_vma *vma) 1798 { 1799 struct drm_vma_offset_node *node; 1800 u64 vma_offset; 1801 1802 if (!i915_vma_has_userfault(vma)) 1803 return; 1804 1805 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma)); 1806 GEM_BUG_ON(!vma->obj->userfault_count); 1807 1808 node = &vma->mmo->vma_node; 1809 vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT; 1810 unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping, 1811 drm_vma_node_offset_addr(node) + vma_offset, 1812 vma->size, 1813 1); 1814 1815 i915_vma_unset_userfault(vma); 1816 if (!--vma->obj->userfault_count) 1817 list_del(&vma->obj->userfault_link); 1818 } 1819 1820 static int 1821 __i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma) 1822 { 1823 return __i915_request_await_exclusive(rq, &vma->active); 1824 } 1825 1826 static int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq) 1827 { 1828 int err; 1829 1830 /* Wait for the vma to be bound before we start! */ 1831 err = __i915_request_await_bind(rq, vma); 1832 if (err) 1833 return err; 1834 1835 return i915_active_add_request(&vma->active, rq); 1836 } 1837 1838 int _i915_vma_move_to_active(struct i915_vma *vma, 1839 struct i915_request *rq, 1840 struct dma_fence *fence, 1841 unsigned int flags) 1842 { 1843 struct drm_i915_gem_object *obj = vma->obj; 1844 int err; 1845 1846 assert_object_held(obj); 1847 1848 GEM_BUG_ON(!vma->pages); 1849 1850 err = __i915_vma_move_to_active(vma, rq); 1851 if (unlikely(err)) 1852 return err; 1853 1854 /* 1855 * Reserve fences slot early to prevent an allocation after preparing 1856 * the workload and associating fences with dma_resv. 1857 */ 1858 if (fence && !(flags & __EXEC_OBJECT_NO_RESERVE)) { 1859 struct dma_fence *curr; 1860 int idx; 1861 1862 dma_fence_array_for_each(curr, idx, fence) 1863 ; 1864 err = dma_resv_reserve_fences(vma->obj->base.resv, idx); 1865 if (unlikely(err)) 1866 return err; 1867 } 1868 1869 if (flags & EXEC_OBJECT_WRITE) { 1870 struct intel_frontbuffer *front; 1871 1872 front = __intel_frontbuffer_get(obj); 1873 if (unlikely(front)) { 1874 if (intel_frontbuffer_invalidate(front, ORIGIN_CS)) 1875 i915_active_add_request(&front->write, rq); 1876 intel_frontbuffer_put(front); 1877 } 1878 } 1879 1880 if (fence) { 1881 struct dma_fence *curr; 1882 enum dma_resv_usage usage; 1883 int idx; 1884 1885 if (flags & EXEC_OBJECT_WRITE) { 1886 usage = DMA_RESV_USAGE_WRITE; 1887 obj->write_domain = I915_GEM_DOMAIN_RENDER; 1888 obj->read_domains = 0; 1889 } else { 1890 usage = DMA_RESV_USAGE_READ; 1891 obj->write_domain = 0; 1892 } 1893 1894 dma_fence_array_for_each(curr, idx, fence) 1895 dma_resv_add_fence(vma->obj->base.resv, curr, usage); 1896 } 1897 1898 if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence) 1899 i915_active_add_request(&vma->fence->active, rq); 1900 1901 obj->read_domains |= I915_GEM_GPU_DOMAINS; 1902 obj->mm.dirty = true; 1903 1904 GEM_BUG_ON(!i915_vma_is_active(vma)); 1905 return 0; 1906 } 1907 1908 struct dma_fence *__i915_vma_evict(struct i915_vma *vma, bool async) 1909 { 1910 struct i915_vma_resource *vma_res = vma->resource; 1911 struct dma_fence *unbind_fence; 1912 1913 GEM_BUG_ON(i915_vma_is_pinned(vma)); 1914 assert_vma_held_evict(vma); 1915 1916 if (i915_vma_is_map_and_fenceable(vma)) { 1917 /* Force a pagefault for domain tracking on next user access */ 1918 i915_vma_revoke_mmap(vma); 1919 1920 /* 1921 * Check that we have flushed all writes through the GGTT 1922 * before the unbind, other due to non-strict nature of those 1923 * indirect writes they may end up referencing the GGTT PTE 1924 * after the unbind. 1925 * 1926 * Note that we may be concurrently poking at the GGTT_WRITE 1927 * bit from set-domain, as we mark all GGTT vma associated 1928 * with an object. We know this is for another vma, as we 1929 * are currently unbinding this one -- so if this vma will be 1930 * reused, it will be refaulted and have its dirty bit set 1931 * before the next write. 1932 */ 1933 i915_vma_flush_writes(vma); 1934 1935 /* release the fence reg _after_ flushing */ 1936 i915_vma_revoke_fence(vma); 1937 1938 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma)); 1939 } 1940 1941 __i915_vma_iounmap(vma); 1942 1943 GEM_BUG_ON(vma->fence); 1944 GEM_BUG_ON(i915_vma_has_userfault(vma)); 1945 1946 /* Object backend must be async capable. */ 1947 GEM_WARN_ON(async && !vma->resource->bi.pages_rsgt); 1948 1949 /* If vm is not open, unbind is a nop. */ 1950 vma_res->needs_wakeref = i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND) && 1951 kref_read(&vma->vm->ref); 1952 vma_res->skip_pte_rewrite = !kref_read(&vma->vm->ref) || 1953 vma->vm->skip_pte_rewrite; 1954 trace_i915_vma_unbind(vma); 1955 1956 if (async) 1957 unbind_fence = i915_vma_resource_unbind(vma_res, 1958 &vma->obj->mm.tlb); 1959 else 1960 unbind_fence = i915_vma_resource_unbind(vma_res, NULL); 1961 1962 vma->resource = NULL; 1963 1964 atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE), 1965 &vma->flags); 1966 1967 i915_vma_detach(vma); 1968 1969 if (!async) { 1970 if (unbind_fence) { 1971 dma_fence_wait(unbind_fence, false); 1972 dma_fence_put(unbind_fence); 1973 unbind_fence = NULL; 1974 } 1975 vma_invalidate_tlb(vma->vm, &vma->obj->mm.tlb); 1976 } 1977 1978 /* 1979 * Binding itself may not have completed until the unbind fence signals, 1980 * so don't drop the pages until that happens, unless the resource is 1981 * async_capable. 1982 */ 1983 1984 vma_unbind_pages(vma); 1985 return unbind_fence; 1986 } 1987 1988 int __i915_vma_unbind(struct i915_vma *vma) 1989 { 1990 int ret; 1991 1992 lockdep_assert_held(&vma->vm->mutex); 1993 assert_vma_held_evict(vma); 1994 1995 if (!drm_mm_node_allocated(&vma->node)) 1996 return 0; 1997 1998 if (i915_vma_is_pinned(vma)) { 1999 vma_print_allocator(vma, "is pinned"); 2000 return -EAGAIN; 2001 } 2002 2003 /* 2004 * After confirming that no one else is pinning this vma, wait for 2005 * any laggards who may have crept in during the wait (through 2006 * a residual pin skipping the vm->mutex) to complete. 2007 */ 2008 ret = i915_vma_sync(vma); 2009 if (ret) 2010 return ret; 2011 2012 GEM_BUG_ON(i915_vma_is_active(vma)); 2013 __i915_vma_evict(vma, false); 2014 2015 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */ 2016 return 0; 2017 } 2018 2019 static struct dma_fence *__i915_vma_unbind_async(struct i915_vma *vma) 2020 { 2021 struct dma_fence *fence; 2022 2023 lockdep_assert_held(&vma->vm->mutex); 2024 2025 if (!drm_mm_node_allocated(&vma->node)) 2026 return NULL; 2027 2028 if (i915_vma_is_pinned(vma) || 2029 &vma->obj->mm.rsgt->table != vma->resource->bi.pages) 2030 return ERR_PTR(-EAGAIN); 2031 2032 /* 2033 * We probably need to replace this with awaiting the fences of the 2034 * object's dma_resv when the vma active goes away. When doing that 2035 * we need to be careful to not add the vma_resource unbind fence 2036 * immediately to the object's dma_resv, because then unbinding 2037 * the next vma from the object, in case there are many, will 2038 * actually await the unbinding of the previous vmas, which is 2039 * undesirable. 2040 */ 2041 if (i915_sw_fence_await_active(&vma->resource->chain, &vma->active, 2042 I915_ACTIVE_AWAIT_EXCL | 2043 I915_ACTIVE_AWAIT_ACTIVE) < 0) { 2044 return ERR_PTR(-EBUSY); 2045 } 2046 2047 fence = __i915_vma_evict(vma, true); 2048 2049 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */ 2050 2051 return fence; 2052 } 2053 2054 int i915_vma_unbind(struct i915_vma *vma) 2055 { 2056 struct i915_address_space *vm = vma->vm; 2057 intel_wakeref_t wakeref = 0; 2058 int err; 2059 2060 assert_object_held_shared(vma->obj); 2061 2062 /* Optimistic wait before taking the mutex */ 2063 err = i915_vma_sync(vma); 2064 if (err) 2065 return err; 2066 2067 if (!drm_mm_node_allocated(&vma->node)) 2068 return 0; 2069 2070 if (i915_vma_is_pinned(vma)) { 2071 vma_print_allocator(vma, "is pinned"); 2072 return -EAGAIN; 2073 } 2074 2075 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)) 2076 /* XXX not always required: nop_clear_range */ 2077 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm); 2078 2079 err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref); 2080 if (err) 2081 goto out_rpm; 2082 2083 err = __i915_vma_unbind(vma); 2084 mutex_unlock(&vm->mutex); 2085 2086 out_rpm: 2087 if (wakeref) 2088 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref); 2089 return err; 2090 } 2091 2092 int i915_vma_unbind_async(struct i915_vma *vma, bool trylock_vm) 2093 { 2094 struct drm_i915_gem_object *obj = vma->obj; 2095 struct i915_address_space *vm = vma->vm; 2096 intel_wakeref_t wakeref = 0; 2097 struct dma_fence *fence; 2098 int err; 2099 2100 /* 2101 * We need the dma-resv lock since we add the 2102 * unbind fence to the dma-resv object. 2103 */ 2104 assert_object_held(obj); 2105 2106 if (!drm_mm_node_allocated(&vma->node)) 2107 return 0; 2108 2109 if (i915_vma_is_pinned(vma)) { 2110 vma_print_allocator(vma, "is pinned"); 2111 return -EAGAIN; 2112 } 2113 2114 if (!obj->mm.rsgt) 2115 return -EBUSY; 2116 2117 err = dma_resv_reserve_fences(obj->base.resv, 1); 2118 if (err) 2119 return -EBUSY; 2120 2121 /* 2122 * It would be great if we could grab this wakeref from the 2123 * async unbind work if needed, but we can't because it uses 2124 * kmalloc and it's in the dma-fence signalling critical path. 2125 */ 2126 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)) 2127 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm); 2128 2129 if (trylock_vm && !mutex_trylock(&vm->mutex)) { 2130 err = -EBUSY; 2131 goto out_rpm; 2132 } else if (!trylock_vm) { 2133 err = mutex_lock_interruptible_nested(&vm->mutex, !wakeref); 2134 if (err) 2135 goto out_rpm; 2136 } 2137 2138 fence = __i915_vma_unbind_async(vma); 2139 mutex_unlock(&vm->mutex); 2140 if (IS_ERR_OR_NULL(fence)) { 2141 err = PTR_ERR_OR_ZERO(fence); 2142 goto out_rpm; 2143 } 2144 2145 dma_resv_add_fence(obj->base.resv, fence, DMA_RESV_USAGE_READ); 2146 dma_fence_put(fence); 2147 2148 out_rpm: 2149 if (wakeref) 2150 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref); 2151 return err; 2152 } 2153 2154 int i915_vma_unbind_unlocked(struct i915_vma *vma) 2155 { 2156 int err; 2157 2158 i915_gem_object_lock(vma->obj, NULL); 2159 err = i915_vma_unbind(vma); 2160 i915_gem_object_unlock(vma->obj); 2161 2162 return err; 2163 } 2164 2165 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma) 2166 { 2167 i915_gem_object_make_unshrinkable(vma->obj); 2168 return vma; 2169 } 2170 2171 void i915_vma_make_shrinkable(struct i915_vma *vma) 2172 { 2173 i915_gem_object_make_shrinkable(vma->obj); 2174 } 2175 2176 void i915_vma_make_purgeable(struct i915_vma *vma) 2177 { 2178 i915_gem_object_make_purgeable(vma->obj); 2179 } 2180 2181 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 2182 #include "selftests/i915_vma.c" 2183 #endif 2184 2185 void i915_vma_module_exit(void) 2186 { 2187 kmem_cache_destroy(slab_vmas); 2188 } 2189 2190 int __init i915_vma_module_init(void) 2191 { 2192 slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN); 2193 if (!slab_vmas) 2194 return -ENOMEM; 2195 2196 return 0; 2197 } 2198