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