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 <drm/drm_gem.h> 27 28 #include "display/intel_frontbuffer.h" 29 30 #include "gem/i915_gem_lmem.h" 31 #include "gt/intel_engine.h" 32 #include "gt/intel_engine_heartbeat.h" 33 #include "gt/intel_gt.h" 34 #include "gt/intel_gt_requests.h" 35 36 #include "i915_drv.h" 37 #include "i915_sw_fence_work.h" 38 #include "i915_trace.h" 39 #include "i915_vma.h" 40 #include "i915_vma_resource.h" 41 42 static inline void assert_vma_held_evict(const struct i915_vma *vma) 43 { 44 /* 45 * We may be forced to unbind when the vm is dead, to clean it up. 46 * This is the only exception to the requirement of the object lock 47 * being held. 48 */ 49 if (atomic_read(&vma->vm->open)) 50 assert_object_held_shared(vma->obj); 51 } 52 53 static struct kmem_cache *slab_vmas; 54 55 static struct i915_vma *i915_vma_alloc(void) 56 { 57 return kmem_cache_zalloc(slab_vmas, GFP_KERNEL); 58 } 59 60 static void i915_vma_free(struct i915_vma *vma) 61 { 62 return kmem_cache_free(slab_vmas, vma); 63 } 64 65 #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM) 66 67 #include <linux/stackdepot.h> 68 69 static void vma_print_allocator(struct i915_vma *vma, const char *reason) 70 { 71 char buf[512]; 72 73 if (!vma->node.stack) { 74 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n", 75 vma->node.start, vma->node.size, reason); 76 return; 77 } 78 79 stack_depot_snprint(vma->node.stack, buf, sizeof(buf), 0); 80 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n", 81 vma->node.start, vma->node.size, reason, buf); 82 } 83 84 #else 85 86 static void vma_print_allocator(struct i915_vma *vma, const char *reason) 87 { 88 } 89 90 #endif 91 92 static inline struct i915_vma *active_to_vma(struct i915_active *ref) 93 { 94 return container_of(ref, typeof(struct i915_vma), active); 95 } 96 97 static int __i915_vma_active(struct i915_active *ref) 98 { 99 return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT; 100 } 101 102 static void __i915_vma_retire(struct i915_active *ref) 103 { 104 i915_vma_put(active_to_vma(ref)); 105 } 106 107 static struct i915_vma * 108 vma_create(struct drm_i915_gem_object *obj, 109 struct i915_address_space *vm, 110 const struct i915_ggtt_view *view) 111 { 112 struct i915_vma *pos = ERR_PTR(-E2BIG); 113 struct i915_vma *vma; 114 struct rb_node *rb, **p; 115 116 /* The aliasing_ppgtt should never be used directly! */ 117 GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm); 118 119 vma = i915_vma_alloc(); 120 if (vma == NULL) 121 return ERR_PTR(-ENOMEM); 122 123 kref_init(&vma->ref); 124 vma->vm = i915_vm_get(vm); 125 vma->ops = &vm->vma_ops; 126 vma->obj = obj; 127 vma->size = obj->base.size; 128 vma->display_alignment = I915_GTT_MIN_ALIGNMENT; 129 130 i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire, 0); 131 132 /* Declare ourselves safe for use inside shrinkers */ 133 if (IS_ENABLED(CONFIG_LOCKDEP)) { 134 fs_reclaim_acquire(GFP_KERNEL); 135 might_lock(&vma->active.mutex); 136 fs_reclaim_release(GFP_KERNEL); 137 } 138 139 INIT_LIST_HEAD(&vma->closed_link); 140 141 if (view && view->type != I915_GGTT_VIEW_NORMAL) { 142 vma->ggtt_view = *view; 143 if (view->type == I915_GGTT_VIEW_PARTIAL) { 144 GEM_BUG_ON(range_overflows_t(u64, 145 view->partial.offset, 146 view->partial.size, 147 obj->base.size >> PAGE_SHIFT)); 148 vma->size = view->partial.size; 149 vma->size <<= PAGE_SHIFT; 150 GEM_BUG_ON(vma->size > obj->base.size); 151 } else if (view->type == I915_GGTT_VIEW_ROTATED) { 152 vma->size = intel_rotation_info_size(&view->rotated); 153 vma->size <<= PAGE_SHIFT; 154 } else if (view->type == I915_GGTT_VIEW_REMAPPED) { 155 vma->size = intel_remapped_info_size(&view->remapped); 156 vma->size <<= PAGE_SHIFT; 157 } 158 } 159 160 if (unlikely(vma->size > vm->total)) 161 goto err_vma; 162 163 GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE)); 164 165 spin_lock(&obj->vma.lock); 166 167 if (i915_is_ggtt(vm)) { 168 if (unlikely(overflows_type(vma->size, u32))) 169 goto err_unlock; 170 171 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size, 172 i915_gem_object_get_tiling(obj), 173 i915_gem_object_get_stride(obj)); 174 if (unlikely(vma->fence_size < vma->size || /* overflow */ 175 vma->fence_size > vm->total)) 176 goto err_unlock; 177 178 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT)); 179 180 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size, 181 i915_gem_object_get_tiling(obj), 182 i915_gem_object_get_stride(obj)); 183 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment)); 184 185 __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma)); 186 } 187 188 rb = NULL; 189 p = &obj->vma.tree.rb_node; 190 while (*p) { 191 long cmp; 192 193 rb = *p; 194 pos = rb_entry(rb, struct i915_vma, obj_node); 195 196 /* 197 * If the view already exists in the tree, another thread 198 * already created a matching vma, so return the older instance 199 * and dispose of ours. 200 */ 201 cmp = i915_vma_compare(pos, vm, view); 202 if (cmp < 0) 203 p = &rb->rb_right; 204 else if (cmp > 0) 205 p = &rb->rb_left; 206 else 207 goto err_unlock; 208 } 209 rb_link_node(&vma->obj_node, rb, p); 210 rb_insert_color(&vma->obj_node, &obj->vma.tree); 211 212 if (i915_vma_is_ggtt(vma)) 213 /* 214 * We put the GGTT vma at the start of the vma-list, followed 215 * by the ppGGTT vma. This allows us to break early when 216 * iterating over only the GGTT vma for an object, see 217 * for_each_ggtt_vma() 218 */ 219 list_add(&vma->obj_link, &obj->vma.list); 220 else 221 list_add_tail(&vma->obj_link, &obj->vma.list); 222 223 spin_unlock(&obj->vma.lock); 224 225 return vma; 226 227 err_unlock: 228 spin_unlock(&obj->vma.lock); 229 err_vma: 230 i915_vm_put(vm); 231 i915_vma_free(vma); 232 return pos; 233 } 234 235 static struct i915_vma * 236 i915_vma_lookup(struct drm_i915_gem_object *obj, 237 struct i915_address_space *vm, 238 const struct i915_ggtt_view *view) 239 { 240 struct rb_node *rb; 241 242 rb = obj->vma.tree.rb_node; 243 while (rb) { 244 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node); 245 long cmp; 246 247 cmp = i915_vma_compare(vma, vm, view); 248 if (cmp == 0) 249 return vma; 250 251 if (cmp < 0) 252 rb = rb->rb_right; 253 else 254 rb = rb->rb_left; 255 } 256 257 return NULL; 258 } 259 260 /** 261 * i915_vma_instance - return the singleton instance of the VMA 262 * @obj: parent &struct drm_i915_gem_object to be mapped 263 * @vm: address space in which the mapping is located 264 * @view: additional mapping requirements 265 * 266 * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with 267 * the same @view characteristics. If a match is not found, one is created. 268 * Once created, the VMA is kept until either the object is freed, or the 269 * address space is closed. 270 * 271 * Returns the vma, or an error pointer. 272 */ 273 struct i915_vma * 274 i915_vma_instance(struct drm_i915_gem_object *obj, 275 struct i915_address_space *vm, 276 const struct i915_ggtt_view *view) 277 { 278 struct i915_vma *vma; 279 280 GEM_BUG_ON(view && !i915_is_ggtt_or_dpt(vm)); 281 GEM_BUG_ON(!atomic_read(&vm->open)); 282 283 spin_lock(&obj->vma.lock); 284 vma = i915_vma_lookup(obj, vm, view); 285 spin_unlock(&obj->vma.lock); 286 287 /* vma_create() will resolve the race if another creates the vma */ 288 if (unlikely(!vma)) 289 vma = vma_create(obj, vm, view); 290 291 GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view)); 292 return vma; 293 } 294 295 struct i915_vma_work { 296 struct dma_fence_work base; 297 struct i915_address_space *vm; 298 struct i915_vm_pt_stash stash; 299 struct i915_vma_resource *vma_res; 300 struct drm_i915_gem_object *pinned; 301 struct i915_sw_dma_fence_cb cb; 302 enum i915_cache_level cache_level; 303 unsigned int flags; 304 }; 305 306 static void __vma_bind(struct dma_fence_work *work) 307 { 308 struct i915_vma_work *vw = container_of(work, typeof(*vw), base); 309 struct i915_vma_resource *vma_res = vw->vma_res; 310 311 vma_res->ops->bind_vma(vma_res->vm, &vw->stash, 312 vma_res, vw->cache_level, vw->flags); 313 314 } 315 316 static void __vma_release(struct dma_fence_work *work) 317 { 318 struct i915_vma_work *vw = container_of(work, typeof(*vw), base); 319 320 if (vw->pinned) 321 i915_gem_object_put(vw->pinned); 322 323 i915_vm_free_pt_stash(vw->vm, &vw->stash); 324 i915_vm_put(vw->vm); 325 if (vw->vma_res) 326 i915_vma_resource_put(vw->vma_res); 327 } 328 329 static const struct dma_fence_work_ops bind_ops = { 330 .name = "bind", 331 .work = __vma_bind, 332 .release = __vma_release, 333 }; 334 335 struct i915_vma_work *i915_vma_work(void) 336 { 337 struct i915_vma_work *vw; 338 339 vw = kzalloc(sizeof(*vw), GFP_KERNEL); 340 if (!vw) 341 return NULL; 342 343 dma_fence_work_init(&vw->base, &bind_ops); 344 vw->base.dma.error = -EAGAIN; /* disable the worker by default */ 345 346 return vw; 347 } 348 349 int i915_vma_wait_for_bind(struct i915_vma *vma) 350 { 351 int err = 0; 352 353 if (rcu_access_pointer(vma->active.excl.fence)) { 354 struct dma_fence *fence; 355 356 rcu_read_lock(); 357 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence); 358 rcu_read_unlock(); 359 if (fence) { 360 err = dma_fence_wait(fence, true); 361 dma_fence_put(fence); 362 } 363 } 364 365 return err; 366 } 367 368 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) 369 static int i915_vma_verify_bind_complete(struct i915_vma *vma) 370 { 371 struct dma_fence *fence = i915_active_fence_get(&vma->active.excl); 372 int err; 373 374 if (!fence) 375 return 0; 376 377 if (dma_fence_is_signaled(fence)) 378 err = fence->error; 379 else 380 err = -EBUSY; 381 382 dma_fence_put(fence); 383 384 return err; 385 } 386 #else 387 #define i915_vma_verify_bind_complete(_vma) 0 388 #endif 389 390 I915_SELFTEST_EXPORT void 391 i915_vma_resource_init_from_vma(struct i915_vma_resource *vma_res, 392 struct i915_vma *vma) 393 { 394 struct drm_i915_gem_object *obj = vma->obj; 395 396 i915_vma_resource_init(vma_res, vma->vm, vma->pages, &vma->page_sizes, 397 obj->mm.rsgt, i915_gem_object_is_readonly(obj), 398 i915_gem_object_is_lmem(obj), obj->mm.region, 399 vma->ops, vma->private, vma->node.start, 400 vma->node.size, vma->size); 401 } 402 403 /** 404 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space. 405 * @vma: VMA to map 406 * @cache_level: mapping cache level 407 * @flags: flags like global or local mapping 408 * @work: preallocated worker for allocating and binding the PTE 409 * @vma_res: pointer to a preallocated vma resource. The resource is either 410 * consumed or freed. 411 * 412 * DMA addresses are taken from the scatter-gather table of this object (or of 413 * this VMA in case of non-default GGTT views) and PTE entries set up. 414 * Note that DMA addresses are also the only part of the SG table we care about. 415 */ 416 int i915_vma_bind(struct i915_vma *vma, 417 enum i915_cache_level cache_level, 418 u32 flags, 419 struct i915_vma_work *work, 420 struct i915_vma_resource *vma_res) 421 { 422 u32 bind_flags; 423 u32 vma_flags; 424 int ret; 425 426 lockdep_assert_held(&vma->vm->mutex); 427 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 428 GEM_BUG_ON(vma->size > vma->node.size); 429 430 if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start, 431 vma->node.size, 432 vma->vm->total))) { 433 i915_vma_resource_free(vma_res); 434 return -ENODEV; 435 } 436 437 if (GEM_DEBUG_WARN_ON(!flags)) { 438 i915_vma_resource_free(vma_res); 439 return -EINVAL; 440 } 441 442 bind_flags = flags; 443 bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND; 444 445 vma_flags = atomic_read(&vma->flags); 446 vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND; 447 448 bind_flags &= ~vma_flags; 449 if (bind_flags == 0) { 450 i915_vma_resource_free(vma_res); 451 return 0; 452 } 453 454 GEM_BUG_ON(!atomic_read(&vma->pages_count)); 455 456 /* Wait for or await async unbinds touching our range */ 457 if (work && bind_flags & vma->vm->bind_async_flags) 458 ret = i915_vma_resource_bind_dep_await(vma->vm, 459 &work->base.chain, 460 vma->node.start, 461 vma->node.size, 462 true, 463 GFP_NOWAIT | 464 __GFP_RETRY_MAYFAIL | 465 __GFP_NOWARN); 466 else 467 ret = i915_vma_resource_bind_dep_sync(vma->vm, vma->node.start, 468 vma->node.size, true); 469 if (ret) { 470 i915_vma_resource_free(vma_res); 471 return ret; 472 } 473 474 if (vma->resource || !vma_res) { 475 /* Rebinding with an additional I915_VMA_*_BIND */ 476 GEM_WARN_ON(!vma_flags); 477 i915_vma_resource_free(vma_res); 478 } else { 479 i915_vma_resource_init_from_vma(vma_res, vma); 480 vma->resource = vma_res; 481 } 482 trace_i915_vma_bind(vma, bind_flags); 483 if (work && bind_flags & vma->vm->bind_async_flags) { 484 struct dma_fence *prev; 485 486 work->vma_res = i915_vma_resource_get(vma->resource); 487 work->cache_level = cache_level; 488 work->flags = bind_flags; 489 490 /* 491 * Note we only want to chain up to the migration fence on 492 * the pages (not the object itself). As we don't track that, 493 * yet, we have to use the exclusive fence instead. 494 * 495 * Also note that we do not want to track the async vma as 496 * part of the obj->resv->excl_fence as it only affects 497 * execution and not content or object's backing store lifetime. 498 */ 499 prev = i915_active_set_exclusive(&vma->active, &work->base.dma); 500 if (prev) { 501 __i915_sw_fence_await_dma_fence(&work->base.chain, 502 prev, 503 &work->cb); 504 dma_fence_put(prev); 505 } 506 507 work->base.dma.error = 0; /* enable the queue_work() */ 508 509 /* 510 * If we don't have the refcounted pages list, keep a reference 511 * on the object to avoid waiting for the async bind to 512 * complete in the object destruction path. 513 */ 514 if (!work->vma_res->bi.pages_rsgt) 515 work->pinned = i915_gem_object_get(vma->obj); 516 } else { 517 if (vma->obj) { 518 ret = i915_gem_object_wait_moving_fence(vma->obj, true); 519 if (ret) { 520 i915_vma_resource_free(vma->resource); 521 vma->resource = NULL; 522 523 return ret; 524 } 525 } 526 vma->ops->bind_vma(vma->vm, NULL, vma->resource, cache_level, 527 bind_flags); 528 } 529 530 atomic_or(bind_flags, &vma->flags); 531 return 0; 532 } 533 534 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma) 535 { 536 void __iomem *ptr; 537 int err; 538 539 if (!i915_gem_object_is_lmem(vma->obj)) { 540 if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) { 541 err = -ENODEV; 542 goto err; 543 } 544 } 545 546 GEM_BUG_ON(!i915_vma_is_ggtt(vma)); 547 GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)); 548 GEM_BUG_ON(i915_vma_verify_bind_complete(vma)); 549 550 ptr = READ_ONCE(vma->iomap); 551 if (ptr == NULL) { 552 /* 553 * TODO: consider just using i915_gem_object_pin_map() for lmem 554 * instead, which already supports mapping non-contiguous chunks 555 * of pages, that way we can also drop the 556 * I915_BO_ALLOC_CONTIGUOUS when allocating the object. 557 */ 558 if (i915_gem_object_is_lmem(vma->obj)) 559 ptr = i915_gem_object_lmem_io_map(vma->obj, 0, 560 vma->obj->base.size); 561 else 562 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap, 563 vma->node.start, 564 vma->node.size); 565 if (ptr == NULL) { 566 err = -ENOMEM; 567 goto err; 568 } 569 570 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) { 571 io_mapping_unmap(ptr); 572 ptr = vma->iomap; 573 } 574 } 575 576 __i915_vma_pin(vma); 577 578 err = i915_vma_pin_fence(vma); 579 if (err) 580 goto err_unpin; 581 582 i915_vma_set_ggtt_write(vma); 583 584 /* NB Access through the GTT requires the device to be awake. */ 585 return ptr; 586 587 err_unpin: 588 __i915_vma_unpin(vma); 589 err: 590 return IO_ERR_PTR(err); 591 } 592 593 void i915_vma_flush_writes(struct i915_vma *vma) 594 { 595 if (i915_vma_unset_ggtt_write(vma)) 596 intel_gt_flush_ggtt_writes(vma->vm->gt); 597 } 598 599 void i915_vma_unpin_iomap(struct i915_vma *vma) 600 { 601 GEM_BUG_ON(vma->iomap == NULL); 602 603 i915_vma_flush_writes(vma); 604 605 i915_vma_unpin_fence(vma); 606 i915_vma_unpin(vma); 607 } 608 609 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags) 610 { 611 struct i915_vma *vma; 612 struct drm_i915_gem_object *obj; 613 614 vma = fetch_and_zero(p_vma); 615 if (!vma) 616 return; 617 618 obj = vma->obj; 619 GEM_BUG_ON(!obj); 620 621 i915_vma_unpin(vma); 622 623 if (flags & I915_VMA_RELEASE_MAP) 624 i915_gem_object_unpin_map(obj); 625 626 i915_gem_object_put(obj); 627 } 628 629 bool i915_vma_misplaced(const struct i915_vma *vma, 630 u64 size, u64 alignment, u64 flags) 631 { 632 if (!drm_mm_node_allocated(&vma->node)) 633 return false; 634 635 if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma))) 636 return true; 637 638 if (vma->node.size < size) 639 return true; 640 641 GEM_BUG_ON(alignment && !is_power_of_2(alignment)); 642 if (alignment && !IS_ALIGNED(vma->node.start, alignment)) 643 return true; 644 645 if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma)) 646 return true; 647 648 if (flags & PIN_OFFSET_BIAS && 649 vma->node.start < (flags & PIN_OFFSET_MASK)) 650 return true; 651 652 if (flags & PIN_OFFSET_FIXED && 653 vma->node.start != (flags & PIN_OFFSET_MASK)) 654 return true; 655 656 return false; 657 } 658 659 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma) 660 { 661 bool mappable, fenceable; 662 663 GEM_BUG_ON(!i915_vma_is_ggtt(vma)); 664 GEM_BUG_ON(!vma->fence_size); 665 666 fenceable = (vma->node.size >= vma->fence_size && 667 IS_ALIGNED(vma->node.start, vma->fence_alignment)); 668 669 mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end; 670 671 if (mappable && fenceable) 672 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma)); 673 else 674 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma)); 675 } 676 677 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color) 678 { 679 struct drm_mm_node *node = &vma->node; 680 struct drm_mm_node *other; 681 682 /* 683 * On some machines we have to be careful when putting differing types 684 * of snoopable memory together to avoid the prefetcher crossing memory 685 * domains and dying. During vm initialisation, we decide whether or not 686 * these constraints apply and set the drm_mm.color_adjust 687 * appropriately. 688 */ 689 if (!i915_vm_has_cache_coloring(vma->vm)) 690 return true; 691 692 /* Only valid to be called on an already inserted vma */ 693 GEM_BUG_ON(!drm_mm_node_allocated(node)); 694 GEM_BUG_ON(list_empty(&node->node_list)); 695 696 other = list_prev_entry(node, node_list); 697 if (i915_node_color_differs(other, color) && 698 !drm_mm_hole_follows(other)) 699 return false; 700 701 other = list_next_entry(node, node_list); 702 if (i915_node_color_differs(other, color) && 703 !drm_mm_hole_follows(node)) 704 return false; 705 706 return true; 707 } 708 709 /** 710 * i915_vma_insert - finds a slot for the vma in its address space 711 * @vma: the vma 712 * @size: requested size in bytes (can be larger than the VMA) 713 * @alignment: required alignment 714 * @flags: mask of PIN_* flags to use 715 * 716 * First we try to allocate some free space that meets the requirements for 717 * the VMA. Failiing that, if the flags permit, it will evict an old VMA, 718 * preferrably the oldest idle entry to make room for the new VMA. 719 * 720 * Returns: 721 * 0 on success, negative error code otherwise. 722 */ 723 static int 724 i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, 725 u64 size, u64 alignment, u64 flags) 726 { 727 unsigned long color; 728 u64 start, end; 729 int ret; 730 731 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND)); 732 GEM_BUG_ON(drm_mm_node_allocated(&vma->node)); 733 734 size = max(size, vma->size); 735 alignment = max(alignment, vma->display_alignment); 736 if (flags & PIN_MAPPABLE) { 737 size = max_t(typeof(size), size, vma->fence_size); 738 alignment = max_t(typeof(alignment), 739 alignment, vma->fence_alignment); 740 } 741 742 GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE)); 743 GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT)); 744 GEM_BUG_ON(!is_power_of_2(alignment)); 745 746 start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0; 747 GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE)); 748 749 end = vma->vm->total; 750 if (flags & PIN_MAPPABLE) 751 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end); 752 if (flags & PIN_ZONE_4G) 753 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE); 754 GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE)); 755 756 /* If binding the object/GGTT view requires more space than the entire 757 * aperture has, reject it early before evicting everything in a vain 758 * attempt to find space. 759 */ 760 if (size > end) { 761 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n", 762 size, flags & PIN_MAPPABLE ? "mappable" : "total", 763 end); 764 return -ENOSPC; 765 } 766 767 color = 0; 768 if (i915_vm_has_cache_coloring(vma->vm)) 769 color = vma->obj->cache_level; 770 771 if (flags & PIN_OFFSET_FIXED) { 772 u64 offset = flags & PIN_OFFSET_MASK; 773 if (!IS_ALIGNED(offset, alignment) || 774 range_overflows(offset, size, end)) 775 return -EINVAL; 776 777 ret = i915_gem_gtt_reserve(vma->vm, ww, &vma->node, 778 size, offset, color, 779 flags); 780 if (ret) 781 return ret; 782 } else { 783 /* 784 * We only support huge gtt pages through the 48b PPGTT, 785 * however we also don't want to force any alignment for 786 * objects which need to be tightly packed into the low 32bits. 787 * 788 * Note that we assume that GGTT are limited to 4GiB for the 789 * forseeable future. See also i915_ggtt_offset(). 790 */ 791 if (upper_32_bits(end - 1) && 792 vma->page_sizes.sg > I915_GTT_PAGE_SIZE) { 793 /* 794 * We can't mix 64K and 4K PTEs in the same page-table 795 * (2M block), and so to avoid the ugliness and 796 * complexity of coloring we opt for just aligning 64K 797 * objects to 2M. 798 */ 799 u64 page_alignment = 800 rounddown_pow_of_two(vma->page_sizes.sg | 801 I915_GTT_PAGE_SIZE_2M); 802 803 /* 804 * Check we don't expand for the limited Global GTT 805 * (mappable aperture is even more precious!). This 806 * also checks that we exclude the aliasing-ppgtt. 807 */ 808 GEM_BUG_ON(i915_vma_is_ggtt(vma)); 809 810 alignment = max(alignment, page_alignment); 811 812 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K) 813 size = round_up(size, I915_GTT_PAGE_SIZE_2M); 814 } 815 816 ret = i915_gem_gtt_insert(vma->vm, ww, &vma->node, 817 size, alignment, color, 818 start, end, flags); 819 if (ret) 820 return ret; 821 822 GEM_BUG_ON(vma->node.start < start); 823 GEM_BUG_ON(vma->node.start + vma->node.size > end); 824 } 825 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 826 GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color)); 827 828 list_add_tail(&vma->vm_link, &vma->vm->bound_list); 829 830 return 0; 831 } 832 833 static void 834 i915_vma_detach(struct i915_vma *vma) 835 { 836 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); 837 GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND)); 838 839 /* 840 * And finally now the object is completely decoupled from this 841 * vma, we can drop its hold on the backing storage and allow 842 * it to be reaped by the shrinker. 843 */ 844 list_del(&vma->vm_link); 845 } 846 847 static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) 848 { 849 unsigned int bound; 850 851 bound = atomic_read(&vma->flags); 852 853 if (flags & PIN_VALIDATE) { 854 flags &= I915_VMA_BIND_MASK; 855 856 return (flags & bound) == flags; 857 } 858 859 /* with the lock mandatory for unbind, we don't race here */ 860 flags &= I915_VMA_BIND_MASK; 861 do { 862 if (unlikely(flags & ~bound)) 863 return false; 864 865 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) 866 return false; 867 868 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0); 869 } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1)); 870 871 return true; 872 } 873 874 static struct scatterlist * 875 rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset, 876 unsigned int width, unsigned int height, 877 unsigned int src_stride, unsigned int dst_stride, 878 struct sg_table *st, struct scatterlist *sg) 879 { 880 unsigned int column, row; 881 unsigned int src_idx; 882 883 for (column = 0; column < width; column++) { 884 unsigned int left; 885 886 src_idx = src_stride * (height - 1) + column + offset; 887 for (row = 0; row < height; row++) { 888 st->nents++; 889 /* 890 * We don't need the pages, but need to initialize 891 * the entries so the sg list can be happily traversed. 892 * The only thing we need are DMA addresses. 893 */ 894 sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0); 895 sg_dma_address(sg) = 896 i915_gem_object_get_dma_address(obj, src_idx); 897 sg_dma_len(sg) = I915_GTT_PAGE_SIZE; 898 sg = sg_next(sg); 899 src_idx -= src_stride; 900 } 901 902 left = (dst_stride - height) * I915_GTT_PAGE_SIZE; 903 904 if (!left) 905 continue; 906 907 st->nents++; 908 909 /* 910 * The DE ignores the PTEs for the padding tiles, the sg entry 911 * here is just a conenience to indicate how many padding PTEs 912 * to insert at this spot. 913 */ 914 sg_set_page(sg, NULL, left, 0); 915 sg_dma_address(sg) = 0; 916 sg_dma_len(sg) = left; 917 sg = sg_next(sg); 918 } 919 920 return sg; 921 } 922 923 static noinline struct sg_table * 924 intel_rotate_pages(struct intel_rotation_info *rot_info, 925 struct drm_i915_gem_object *obj) 926 { 927 unsigned int size = intel_rotation_info_size(rot_info); 928 struct drm_i915_private *i915 = to_i915(obj->base.dev); 929 struct sg_table *st; 930 struct scatterlist *sg; 931 int ret = -ENOMEM; 932 int i; 933 934 /* Allocate target SG list. */ 935 st = kmalloc(sizeof(*st), GFP_KERNEL); 936 if (!st) 937 goto err_st_alloc; 938 939 ret = sg_alloc_table(st, size, GFP_KERNEL); 940 if (ret) 941 goto err_sg_alloc; 942 943 st->nents = 0; 944 sg = st->sgl; 945 946 for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++) 947 sg = rotate_pages(obj, rot_info->plane[i].offset, 948 rot_info->plane[i].width, rot_info->plane[i].height, 949 rot_info->plane[i].src_stride, 950 rot_info->plane[i].dst_stride, 951 st, sg); 952 953 return st; 954 955 err_sg_alloc: 956 kfree(st); 957 err_st_alloc: 958 959 drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n", 960 obj->base.size, rot_info->plane[0].width, 961 rot_info->plane[0].height, size); 962 963 return ERR_PTR(ret); 964 } 965 966 static struct scatterlist * 967 remap_pages(struct drm_i915_gem_object *obj, 968 unsigned int offset, unsigned int alignment_pad, 969 unsigned int width, unsigned int height, 970 unsigned int src_stride, unsigned int dst_stride, 971 struct sg_table *st, struct scatterlist *sg) 972 { 973 unsigned int row; 974 975 if (!width || !height) 976 return sg; 977 978 if (alignment_pad) { 979 st->nents++; 980 981 /* 982 * The DE ignores the PTEs for the padding tiles, the sg entry 983 * here is just a convenience to indicate how many padding PTEs 984 * to insert at this spot. 985 */ 986 sg_set_page(sg, NULL, alignment_pad * 4096, 0); 987 sg_dma_address(sg) = 0; 988 sg_dma_len(sg) = alignment_pad * 4096; 989 sg = sg_next(sg); 990 } 991 992 for (row = 0; row < height; row++) { 993 unsigned int left = width * I915_GTT_PAGE_SIZE; 994 995 while (left) { 996 dma_addr_t addr; 997 unsigned int length; 998 999 /* 1000 * We don't need the pages, but need to initialize 1001 * the entries so the sg list can be happily traversed. 1002 * The only thing we need are DMA addresses. 1003 */ 1004 1005 addr = i915_gem_object_get_dma_address_len(obj, offset, &length); 1006 1007 length = min(left, length); 1008 1009 st->nents++; 1010 1011 sg_set_page(sg, NULL, length, 0); 1012 sg_dma_address(sg) = addr; 1013 sg_dma_len(sg) = length; 1014 sg = sg_next(sg); 1015 1016 offset += length / I915_GTT_PAGE_SIZE; 1017 left -= length; 1018 } 1019 1020 offset += src_stride - width; 1021 1022 left = (dst_stride - width) * I915_GTT_PAGE_SIZE; 1023 1024 if (!left) 1025 continue; 1026 1027 st->nents++; 1028 1029 /* 1030 * The DE ignores the PTEs for the padding tiles, the sg entry 1031 * here is just a conenience to indicate how many padding PTEs 1032 * to insert at this spot. 1033 */ 1034 sg_set_page(sg, NULL, left, 0); 1035 sg_dma_address(sg) = 0; 1036 sg_dma_len(sg) = left; 1037 sg = sg_next(sg); 1038 } 1039 1040 return sg; 1041 } 1042 1043 static noinline struct sg_table * 1044 intel_remap_pages(struct intel_remapped_info *rem_info, 1045 struct drm_i915_gem_object *obj) 1046 { 1047 unsigned int size = intel_remapped_info_size(rem_info); 1048 struct drm_i915_private *i915 = to_i915(obj->base.dev); 1049 struct sg_table *st; 1050 struct scatterlist *sg; 1051 unsigned int gtt_offset = 0; 1052 int ret = -ENOMEM; 1053 int i; 1054 1055 /* Allocate target SG list. */ 1056 st = kmalloc(sizeof(*st), GFP_KERNEL); 1057 if (!st) 1058 goto err_st_alloc; 1059 1060 ret = sg_alloc_table(st, size, GFP_KERNEL); 1061 if (ret) 1062 goto err_sg_alloc; 1063 1064 st->nents = 0; 1065 sg = st->sgl; 1066 1067 for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) { 1068 unsigned int alignment_pad = 0; 1069 1070 if (rem_info->plane_alignment) 1071 alignment_pad = ALIGN(gtt_offset, rem_info->plane_alignment) - gtt_offset; 1072 1073 sg = remap_pages(obj, 1074 rem_info->plane[i].offset, alignment_pad, 1075 rem_info->plane[i].width, rem_info->plane[i].height, 1076 rem_info->plane[i].src_stride, rem_info->plane[i].dst_stride, 1077 st, sg); 1078 1079 gtt_offset += alignment_pad + 1080 rem_info->plane[i].dst_stride * rem_info->plane[i].height; 1081 } 1082 1083 i915_sg_trim(st); 1084 1085 return st; 1086 1087 err_sg_alloc: 1088 kfree(st); 1089 err_st_alloc: 1090 1091 drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n", 1092 obj->base.size, rem_info->plane[0].width, 1093 rem_info->plane[0].height, size); 1094 1095 return ERR_PTR(ret); 1096 } 1097 1098 static noinline struct sg_table * 1099 intel_partial_pages(const struct i915_ggtt_view *view, 1100 struct drm_i915_gem_object *obj) 1101 { 1102 struct sg_table *st; 1103 struct scatterlist *sg, *iter; 1104 unsigned int count = view->partial.size; 1105 unsigned int offset; 1106 int ret = -ENOMEM; 1107 1108 st = kmalloc(sizeof(*st), GFP_KERNEL); 1109 if (!st) 1110 goto err_st_alloc; 1111 1112 ret = sg_alloc_table(st, count, GFP_KERNEL); 1113 if (ret) 1114 goto err_sg_alloc; 1115 1116 iter = i915_gem_object_get_sg_dma(obj, view->partial.offset, &offset); 1117 GEM_BUG_ON(!iter); 1118 1119 sg = st->sgl; 1120 st->nents = 0; 1121 do { 1122 unsigned int len; 1123 1124 len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT), 1125 count << PAGE_SHIFT); 1126 sg_set_page(sg, NULL, len, 0); 1127 sg_dma_address(sg) = 1128 sg_dma_address(iter) + (offset << PAGE_SHIFT); 1129 sg_dma_len(sg) = len; 1130 1131 st->nents++; 1132 count -= len >> PAGE_SHIFT; 1133 if (count == 0) { 1134 sg_mark_end(sg); 1135 i915_sg_trim(st); /* Drop any unused tail entries. */ 1136 1137 return st; 1138 } 1139 1140 sg = __sg_next(sg); 1141 iter = __sg_next(iter); 1142 offset = 0; 1143 } while (1); 1144 1145 err_sg_alloc: 1146 kfree(st); 1147 err_st_alloc: 1148 return ERR_PTR(ret); 1149 } 1150 1151 static int 1152 __i915_vma_get_pages(struct i915_vma *vma) 1153 { 1154 struct sg_table *pages; 1155 1156 /* 1157 * The vma->pages are only valid within the lifespan of the borrowed 1158 * obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so 1159 * must be the vma->pages. A simple rule is that vma->pages must only 1160 * be accessed when the obj->mm.pages are pinned. 1161 */ 1162 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj)); 1163 1164 switch (vma->ggtt_view.type) { 1165 default: 1166 GEM_BUG_ON(vma->ggtt_view.type); 1167 fallthrough; 1168 case I915_GGTT_VIEW_NORMAL: 1169 pages = vma->obj->mm.pages; 1170 break; 1171 1172 case I915_GGTT_VIEW_ROTATED: 1173 pages = 1174 intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj); 1175 break; 1176 1177 case I915_GGTT_VIEW_REMAPPED: 1178 pages = 1179 intel_remap_pages(&vma->ggtt_view.remapped, vma->obj); 1180 break; 1181 1182 case I915_GGTT_VIEW_PARTIAL: 1183 pages = intel_partial_pages(&vma->ggtt_view, vma->obj); 1184 break; 1185 } 1186 1187 if (IS_ERR(pages)) { 1188 drm_err(&vma->vm->i915->drm, 1189 "Failed to get pages for VMA view type %u (%ld)!\n", 1190 vma->ggtt_view.type, PTR_ERR(pages)); 1191 return PTR_ERR(pages); 1192 } 1193 1194 vma->pages = pages; 1195 1196 return 0; 1197 } 1198 1199 I915_SELFTEST_EXPORT int i915_vma_get_pages(struct i915_vma *vma) 1200 { 1201 int err; 1202 1203 if (atomic_add_unless(&vma->pages_count, 1, 0)) 1204 return 0; 1205 1206 err = i915_gem_object_pin_pages(vma->obj); 1207 if (err) 1208 return err; 1209 1210 err = __i915_vma_get_pages(vma); 1211 if (err) 1212 goto err_unpin; 1213 1214 vma->page_sizes = vma->obj->mm.page_sizes; 1215 atomic_inc(&vma->pages_count); 1216 1217 return 0; 1218 1219 err_unpin: 1220 __i915_gem_object_unpin_pages(vma->obj); 1221 1222 return err; 1223 } 1224 1225 static void __vma_put_pages(struct i915_vma *vma, unsigned int count) 1226 { 1227 /* We allocate under vma_get_pages, so beware the shrinker */ 1228 GEM_BUG_ON(atomic_read(&vma->pages_count) < count); 1229 1230 if (atomic_sub_return(count, &vma->pages_count) == 0) { 1231 if (vma->pages != vma->obj->mm.pages) { 1232 sg_free_table(vma->pages); 1233 kfree(vma->pages); 1234 } 1235 vma->pages = NULL; 1236 1237 i915_gem_object_unpin_pages(vma->obj); 1238 } 1239 } 1240 1241 I915_SELFTEST_EXPORT void i915_vma_put_pages(struct i915_vma *vma) 1242 { 1243 if (atomic_add_unless(&vma->pages_count, -1, 1)) 1244 return; 1245 1246 __vma_put_pages(vma, 1); 1247 } 1248 1249 static void vma_unbind_pages(struct i915_vma *vma) 1250 { 1251 unsigned int count; 1252 1253 lockdep_assert_held(&vma->vm->mutex); 1254 1255 /* The upper portion of pages_count is the number of bindings */ 1256 count = atomic_read(&vma->pages_count); 1257 count >>= I915_VMA_PAGES_BIAS; 1258 GEM_BUG_ON(!count); 1259 1260 __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS); 1261 } 1262 1263 int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, 1264 u64 size, u64 alignment, u64 flags) 1265 { 1266 struct i915_vma_work *work = NULL; 1267 struct dma_fence *moving = NULL; 1268 struct i915_vma_resource *vma_res = NULL; 1269 intel_wakeref_t wakeref = 0; 1270 unsigned int bound; 1271 int err; 1272 1273 assert_vma_held(vma); 1274 GEM_BUG_ON(!ww); 1275 1276 BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND); 1277 BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND); 1278 1279 GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL))); 1280 1281 /* First try and grab the pin without rebinding the vma */ 1282 if (try_qad_pin(vma, flags)) 1283 return 0; 1284 1285 err = i915_vma_get_pages(vma); 1286 if (err) 1287 return err; 1288 1289 if (flags & PIN_GLOBAL) 1290 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm); 1291 1292 moving = vma->obj ? i915_gem_object_get_moving_fence(vma->obj) : NULL; 1293 if (flags & vma->vm->bind_async_flags || moving) { 1294 /* lock VM */ 1295 err = i915_vm_lock_objects(vma->vm, ww); 1296 if (err) 1297 goto err_rpm; 1298 1299 work = i915_vma_work(); 1300 if (!work) { 1301 err = -ENOMEM; 1302 goto err_rpm; 1303 } 1304 1305 work->vm = i915_vm_get(vma->vm); 1306 1307 dma_fence_work_chain(&work->base, moving); 1308 1309 /* Allocate enough page directories to used PTE */ 1310 if (vma->vm->allocate_va_range) { 1311 err = i915_vm_alloc_pt_stash(vma->vm, 1312 &work->stash, 1313 vma->size); 1314 if (err) 1315 goto err_fence; 1316 1317 err = i915_vm_map_pt_stash(vma->vm, &work->stash); 1318 if (err) 1319 goto err_fence; 1320 } 1321 } 1322 1323 vma_res = i915_vma_resource_alloc(); 1324 if (IS_ERR(vma_res)) { 1325 err = PTR_ERR(vma_res); 1326 goto err_fence; 1327 } 1328 1329 /* 1330 * Differentiate between user/kernel vma inside the aliasing-ppgtt. 1331 * 1332 * We conflate the Global GTT with the user's vma when using the 1333 * aliasing-ppgtt, but it is still vitally important to try and 1334 * keep the use cases distinct. For example, userptr objects are 1335 * not allowed inside the Global GTT as that will cause lock 1336 * inversions when we have to evict them the mmu_notifier callbacks - 1337 * but they are allowed to be part of the user ppGTT which can never 1338 * be mapped. As such we try to give the distinct users of the same 1339 * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt 1340 * and i915_ppgtt separate]. 1341 * 1342 * NB this may cause us to mask real lock inversions -- while the 1343 * code is safe today, lockdep may not be able to spot future 1344 * transgressions. 1345 */ 1346 err = mutex_lock_interruptible_nested(&vma->vm->mutex, 1347 !(flags & PIN_GLOBAL)); 1348 if (err) 1349 goto err_vma_res; 1350 1351 /* No more allocations allowed now we hold vm->mutex */ 1352 1353 if (unlikely(i915_vma_is_closed(vma))) { 1354 err = -ENOENT; 1355 goto err_unlock; 1356 } 1357 1358 bound = atomic_read(&vma->flags); 1359 if (unlikely(bound & I915_VMA_ERROR)) { 1360 err = -ENOMEM; 1361 goto err_unlock; 1362 } 1363 1364 if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) { 1365 err = -EAGAIN; /* pins are meant to be fairly temporary */ 1366 goto err_unlock; 1367 } 1368 1369 if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) { 1370 if (!(flags & PIN_VALIDATE)) 1371 __i915_vma_pin(vma); 1372 goto err_unlock; 1373 } 1374 1375 err = i915_active_acquire(&vma->active); 1376 if (err) 1377 goto err_unlock; 1378 1379 if (!(bound & I915_VMA_BIND_MASK)) { 1380 err = i915_vma_insert(vma, ww, size, alignment, flags); 1381 if (err) 1382 goto err_active; 1383 1384 if (i915_is_ggtt(vma->vm)) 1385 __i915_vma_set_map_and_fenceable(vma); 1386 } 1387 1388 GEM_BUG_ON(!vma->pages); 1389 err = i915_vma_bind(vma, 1390 vma->obj->cache_level, 1391 flags, work, vma_res); 1392 vma_res = NULL; 1393 if (err) 1394 goto err_remove; 1395 1396 /* There should only be at most 2 active bindings (user, global) */ 1397 GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound); 1398 atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count); 1399 list_move_tail(&vma->vm_link, &vma->vm->bound_list); 1400 1401 if (!(flags & PIN_VALIDATE)) { 1402 __i915_vma_pin(vma); 1403 GEM_BUG_ON(!i915_vma_is_pinned(vma)); 1404 } 1405 GEM_BUG_ON(!i915_vma_is_bound(vma, flags)); 1406 GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags)); 1407 1408 err_remove: 1409 if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) { 1410 i915_vma_detach(vma); 1411 drm_mm_remove_node(&vma->node); 1412 } 1413 err_active: 1414 i915_active_release(&vma->active); 1415 err_unlock: 1416 mutex_unlock(&vma->vm->mutex); 1417 err_vma_res: 1418 i915_vma_resource_free(vma_res); 1419 err_fence: 1420 if (work) 1421 dma_fence_work_commit_imm(&work->base); 1422 err_rpm: 1423 if (wakeref) 1424 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref); 1425 1426 if (moving) 1427 dma_fence_put(moving); 1428 1429 i915_vma_put_pages(vma); 1430 return err; 1431 } 1432 1433 static void flush_idle_contexts(struct intel_gt *gt) 1434 { 1435 struct intel_engine_cs *engine; 1436 enum intel_engine_id id; 1437 1438 for_each_engine(engine, gt, id) 1439 intel_engine_flush_barriers(engine); 1440 1441 intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT); 1442 } 1443 1444 static int __i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, 1445 u32 align, unsigned int flags) 1446 { 1447 struct i915_address_space *vm = vma->vm; 1448 int err; 1449 1450 do { 1451 err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL); 1452 1453 if (err != -ENOSPC) { 1454 if (!err) { 1455 err = i915_vma_wait_for_bind(vma); 1456 if (err) 1457 i915_vma_unpin(vma); 1458 } 1459 return err; 1460 } 1461 1462 /* Unlike i915_vma_pin, we don't take no for an answer! */ 1463 flush_idle_contexts(vm->gt); 1464 if (mutex_lock_interruptible(&vm->mutex) == 0) { 1465 /* 1466 * We pass NULL ww here, as we don't want to unbind 1467 * locked objects when called from execbuf when pinning 1468 * is removed. This would probably regress badly. 1469 */ 1470 i915_gem_evict_vm(vm, NULL); 1471 mutex_unlock(&vm->mutex); 1472 } 1473 } while (1); 1474 } 1475 1476 int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, 1477 u32 align, unsigned int flags) 1478 { 1479 struct i915_gem_ww_ctx _ww; 1480 int err; 1481 1482 GEM_BUG_ON(!i915_vma_is_ggtt(vma)); 1483 1484 if (ww) 1485 return __i915_ggtt_pin(vma, ww, align, flags); 1486 1487 #ifdef CONFIG_LOCKDEP 1488 WARN_ON(dma_resv_held(vma->obj->base.resv)); 1489 #endif 1490 1491 for_i915_gem_ww(&_ww, err, true) { 1492 err = i915_gem_object_lock(vma->obj, &_ww); 1493 if (!err) 1494 err = __i915_ggtt_pin(vma, &_ww, align, flags); 1495 } 1496 1497 return err; 1498 } 1499 1500 static void __vma_close(struct i915_vma *vma, struct intel_gt *gt) 1501 { 1502 /* 1503 * We defer actually closing, unbinding and destroying the VMA until 1504 * the next idle point, or if the object is freed in the meantime. By 1505 * postponing the unbind, we allow for it to be resurrected by the 1506 * client, avoiding the work required to rebind the VMA. This is 1507 * advantageous for DRI, where the client/server pass objects 1508 * between themselves, temporarily opening a local VMA to the 1509 * object, and then closing it again. The same object is then reused 1510 * on the next frame (or two, depending on the depth of the swap queue) 1511 * causing us to rebind the VMA once more. This ends up being a lot 1512 * of wasted work for the steady state. 1513 */ 1514 GEM_BUG_ON(i915_vma_is_closed(vma)); 1515 list_add(&vma->closed_link, >->closed_vma); 1516 } 1517 1518 void i915_vma_close(struct i915_vma *vma) 1519 { 1520 struct intel_gt *gt = vma->vm->gt; 1521 unsigned long flags; 1522 1523 if (i915_vma_is_ggtt(vma)) 1524 return; 1525 1526 GEM_BUG_ON(!atomic_read(&vma->open_count)); 1527 if (atomic_dec_and_lock_irqsave(&vma->open_count, 1528 >->closed_lock, 1529 flags)) { 1530 __vma_close(vma, gt); 1531 spin_unlock_irqrestore(>->closed_lock, flags); 1532 } 1533 } 1534 1535 static void __i915_vma_remove_closed(struct i915_vma *vma) 1536 { 1537 struct intel_gt *gt = vma->vm->gt; 1538 1539 spin_lock_irq(>->closed_lock); 1540 list_del_init(&vma->closed_link); 1541 spin_unlock_irq(>->closed_lock); 1542 } 1543 1544 void i915_vma_reopen(struct i915_vma *vma) 1545 { 1546 if (i915_vma_is_closed(vma)) 1547 __i915_vma_remove_closed(vma); 1548 } 1549 1550 void i915_vma_release(struct kref *ref) 1551 { 1552 struct i915_vma *vma = container_of(ref, typeof(*vma), ref); 1553 struct drm_i915_gem_object *obj = vma->obj; 1554 1555 if (drm_mm_node_allocated(&vma->node)) { 1556 mutex_lock(&vma->vm->mutex); 1557 atomic_and(~I915_VMA_PIN_MASK, &vma->flags); 1558 WARN_ON(__i915_vma_unbind(vma)); 1559 mutex_unlock(&vma->vm->mutex); 1560 GEM_BUG_ON(drm_mm_node_allocated(&vma->node)); 1561 } 1562 GEM_BUG_ON(i915_vma_is_active(vma)); 1563 1564 spin_lock(&obj->vma.lock); 1565 list_del(&vma->obj_link); 1566 if (!RB_EMPTY_NODE(&vma->obj_node)) 1567 rb_erase(&vma->obj_node, &obj->vma.tree); 1568 spin_unlock(&obj->vma.lock); 1569 1570 __i915_vma_remove_closed(vma); 1571 i915_vm_put(vma->vm); 1572 1573 i915_active_fini(&vma->active); 1574 GEM_WARN_ON(vma->resource); 1575 i915_vma_free(vma); 1576 } 1577 1578 void i915_vma_parked(struct intel_gt *gt) 1579 { 1580 struct i915_vma *vma, *next; 1581 LIST_HEAD(closed); 1582 1583 spin_lock_irq(>->closed_lock); 1584 list_for_each_entry_safe(vma, next, >->closed_vma, closed_link) { 1585 struct drm_i915_gem_object *obj = vma->obj; 1586 struct i915_address_space *vm = vma->vm; 1587 1588 /* XXX All to avoid keeping a reference on i915_vma itself */ 1589 1590 if (!kref_get_unless_zero(&obj->base.refcount)) 1591 continue; 1592 1593 if (!i915_vm_tryopen(vm)) { 1594 i915_gem_object_put(obj); 1595 continue; 1596 } 1597 1598 list_move(&vma->closed_link, &closed); 1599 } 1600 spin_unlock_irq(>->closed_lock); 1601 1602 /* As the GT is held idle, no vma can be reopened as we destroy them */ 1603 list_for_each_entry_safe(vma, next, &closed, closed_link) { 1604 struct drm_i915_gem_object *obj = vma->obj; 1605 struct i915_address_space *vm = vma->vm; 1606 1607 if (i915_gem_object_trylock(obj, NULL)) { 1608 INIT_LIST_HEAD(&vma->closed_link); 1609 __i915_vma_put(vma); 1610 i915_gem_object_unlock(obj); 1611 } else { 1612 /* back you go.. */ 1613 spin_lock_irq(>->closed_lock); 1614 list_add(&vma->closed_link, >->closed_vma); 1615 spin_unlock_irq(>->closed_lock); 1616 } 1617 1618 i915_gem_object_put(obj); 1619 i915_vm_close(vm); 1620 } 1621 } 1622 1623 static void __i915_vma_iounmap(struct i915_vma *vma) 1624 { 1625 GEM_BUG_ON(i915_vma_is_pinned(vma)); 1626 1627 if (vma->iomap == NULL) 1628 return; 1629 1630 io_mapping_unmap(vma->iomap); 1631 vma->iomap = NULL; 1632 } 1633 1634 void i915_vma_revoke_mmap(struct i915_vma *vma) 1635 { 1636 struct drm_vma_offset_node *node; 1637 u64 vma_offset; 1638 1639 if (!i915_vma_has_userfault(vma)) 1640 return; 1641 1642 GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma)); 1643 GEM_BUG_ON(!vma->obj->userfault_count); 1644 1645 node = &vma->mmo->vma_node; 1646 vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT; 1647 unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping, 1648 drm_vma_node_offset_addr(node) + vma_offset, 1649 vma->size, 1650 1); 1651 1652 i915_vma_unset_userfault(vma); 1653 if (!--vma->obj->userfault_count) 1654 list_del(&vma->obj->userfault_link); 1655 } 1656 1657 static int 1658 __i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma) 1659 { 1660 return __i915_request_await_exclusive(rq, &vma->active); 1661 } 1662 1663 static int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq) 1664 { 1665 int err; 1666 1667 /* Wait for the vma to be bound before we start! */ 1668 err = __i915_request_await_bind(rq, vma); 1669 if (err) 1670 return err; 1671 1672 return i915_active_add_request(&vma->active, rq); 1673 } 1674 1675 int _i915_vma_move_to_active(struct i915_vma *vma, 1676 struct i915_request *rq, 1677 struct dma_fence *fence, 1678 unsigned int flags) 1679 { 1680 struct drm_i915_gem_object *obj = vma->obj; 1681 int err; 1682 1683 assert_object_held(obj); 1684 1685 GEM_BUG_ON(!vma->pages); 1686 1687 err = __i915_vma_move_to_active(vma, rq); 1688 if (unlikely(err)) 1689 return err; 1690 1691 if (flags & EXEC_OBJECT_WRITE) { 1692 struct intel_frontbuffer *front; 1693 1694 front = __intel_frontbuffer_get(obj); 1695 if (unlikely(front)) { 1696 if (intel_frontbuffer_invalidate(front, ORIGIN_CS)) 1697 i915_active_add_request(&front->write, rq); 1698 intel_frontbuffer_put(front); 1699 } 1700 1701 if (fence) { 1702 dma_resv_add_excl_fence(vma->obj->base.resv, fence); 1703 obj->write_domain = I915_GEM_DOMAIN_RENDER; 1704 obj->read_domains = 0; 1705 } 1706 } else { 1707 if (!(flags & __EXEC_OBJECT_NO_RESERVE)) { 1708 err = dma_resv_reserve_shared(vma->obj->base.resv, 1); 1709 if (unlikely(err)) 1710 return err; 1711 } 1712 1713 if (fence) { 1714 dma_resv_add_shared_fence(vma->obj->base.resv, fence); 1715 obj->write_domain = 0; 1716 } 1717 } 1718 1719 if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence) 1720 i915_active_add_request(&vma->fence->active, rq); 1721 1722 obj->read_domains |= I915_GEM_GPU_DOMAINS; 1723 obj->mm.dirty = true; 1724 1725 GEM_BUG_ON(!i915_vma_is_active(vma)); 1726 return 0; 1727 } 1728 1729 struct dma_fence *__i915_vma_evict(struct i915_vma *vma, bool async) 1730 { 1731 struct i915_vma_resource *vma_res = vma->resource; 1732 struct dma_fence *unbind_fence; 1733 1734 GEM_BUG_ON(i915_vma_is_pinned(vma)); 1735 assert_vma_held_evict(vma); 1736 1737 if (i915_vma_is_map_and_fenceable(vma)) { 1738 /* Force a pagefault for domain tracking on next user access */ 1739 i915_vma_revoke_mmap(vma); 1740 1741 /* 1742 * Check that we have flushed all writes through the GGTT 1743 * before the unbind, other due to non-strict nature of those 1744 * indirect writes they may end up referencing the GGTT PTE 1745 * after the unbind. 1746 * 1747 * Note that we may be concurrently poking at the GGTT_WRITE 1748 * bit from set-domain, as we mark all GGTT vma associated 1749 * with an object. We know this is for another vma, as we 1750 * are currently unbinding this one -- so if this vma will be 1751 * reused, it will be refaulted and have its dirty bit set 1752 * before the next write. 1753 */ 1754 i915_vma_flush_writes(vma); 1755 1756 /* release the fence reg _after_ flushing */ 1757 i915_vma_revoke_fence(vma); 1758 1759 __i915_vma_iounmap(vma); 1760 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma)); 1761 } 1762 GEM_BUG_ON(vma->fence); 1763 GEM_BUG_ON(i915_vma_has_userfault(vma)); 1764 1765 /* Object backend must be async capable. */ 1766 GEM_WARN_ON(async && !vma->resource->bi.pages_rsgt); 1767 1768 /* If vm is not open, unbind is a nop. */ 1769 vma_res->needs_wakeref = i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND) && 1770 atomic_read(&vma->vm->open); 1771 trace_i915_vma_unbind(vma); 1772 1773 unbind_fence = i915_vma_resource_unbind(vma_res); 1774 vma->resource = NULL; 1775 1776 atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE), 1777 &vma->flags); 1778 1779 i915_vma_detach(vma); 1780 1781 if (!async && unbind_fence) { 1782 dma_fence_wait(unbind_fence, false); 1783 dma_fence_put(unbind_fence); 1784 unbind_fence = NULL; 1785 } 1786 1787 /* 1788 * Binding itself may not have completed until the unbind fence signals, 1789 * so don't drop the pages until that happens, unless the resource is 1790 * async_capable. 1791 */ 1792 1793 vma_unbind_pages(vma); 1794 return unbind_fence; 1795 } 1796 1797 int __i915_vma_unbind(struct i915_vma *vma) 1798 { 1799 int ret; 1800 1801 lockdep_assert_held(&vma->vm->mutex); 1802 assert_vma_held_evict(vma); 1803 1804 if (!drm_mm_node_allocated(&vma->node)) 1805 return 0; 1806 1807 if (i915_vma_is_pinned(vma)) { 1808 vma_print_allocator(vma, "is pinned"); 1809 return -EAGAIN; 1810 } 1811 1812 /* 1813 * After confirming that no one else is pinning this vma, wait for 1814 * any laggards who may have crept in during the wait (through 1815 * a residual pin skipping the vm->mutex) to complete. 1816 */ 1817 ret = i915_vma_sync(vma); 1818 if (ret) 1819 return ret; 1820 1821 GEM_BUG_ON(i915_vma_is_active(vma)); 1822 __i915_vma_evict(vma, false); 1823 1824 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */ 1825 return 0; 1826 } 1827 1828 static struct dma_fence *__i915_vma_unbind_async(struct i915_vma *vma) 1829 { 1830 struct dma_fence *fence; 1831 1832 lockdep_assert_held(&vma->vm->mutex); 1833 1834 if (!drm_mm_node_allocated(&vma->node)) 1835 return NULL; 1836 1837 if (i915_vma_is_pinned(vma) || 1838 &vma->obj->mm.rsgt->table != vma->resource->bi.pages) 1839 return ERR_PTR(-EAGAIN); 1840 1841 /* 1842 * We probably need to replace this with awaiting the fences of the 1843 * object's dma_resv when the vma active goes away. When doing that 1844 * we need to be careful to not add the vma_resource unbind fence 1845 * immediately to the object's dma_resv, because then unbinding 1846 * the next vma from the object, in case there are many, will 1847 * actually await the unbinding of the previous vmas, which is 1848 * undesirable. 1849 */ 1850 if (i915_sw_fence_await_active(&vma->resource->chain, &vma->active, 1851 I915_ACTIVE_AWAIT_EXCL | 1852 I915_ACTIVE_AWAIT_ACTIVE) < 0) { 1853 return ERR_PTR(-EBUSY); 1854 } 1855 1856 fence = __i915_vma_evict(vma, true); 1857 1858 drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */ 1859 1860 return fence; 1861 } 1862 1863 int i915_vma_unbind(struct i915_vma *vma) 1864 { 1865 struct i915_address_space *vm = vma->vm; 1866 intel_wakeref_t wakeref = 0; 1867 int err; 1868 1869 assert_object_held_shared(vma->obj); 1870 1871 /* Optimistic wait before taking the mutex */ 1872 err = i915_vma_sync(vma); 1873 if (err) 1874 return err; 1875 1876 if (!drm_mm_node_allocated(&vma->node)) 1877 return 0; 1878 1879 if (i915_vma_is_pinned(vma)) { 1880 vma_print_allocator(vma, "is pinned"); 1881 return -EAGAIN; 1882 } 1883 1884 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)) 1885 /* XXX not always required: nop_clear_range */ 1886 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm); 1887 1888 err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref); 1889 if (err) 1890 goto out_rpm; 1891 1892 err = __i915_vma_unbind(vma); 1893 mutex_unlock(&vm->mutex); 1894 1895 out_rpm: 1896 if (wakeref) 1897 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref); 1898 return err; 1899 } 1900 1901 int i915_vma_unbind_async(struct i915_vma *vma, bool trylock_vm) 1902 { 1903 struct drm_i915_gem_object *obj = vma->obj; 1904 struct i915_address_space *vm = vma->vm; 1905 intel_wakeref_t wakeref = 0; 1906 struct dma_fence *fence; 1907 int err; 1908 1909 /* 1910 * We need the dma-resv lock since we add the 1911 * unbind fence to the dma-resv object. 1912 */ 1913 assert_object_held(obj); 1914 1915 if (!drm_mm_node_allocated(&vma->node)) 1916 return 0; 1917 1918 if (i915_vma_is_pinned(vma)) { 1919 vma_print_allocator(vma, "is pinned"); 1920 return -EAGAIN; 1921 } 1922 1923 if (!obj->mm.rsgt) 1924 return -EBUSY; 1925 1926 err = dma_resv_reserve_shared(obj->base.resv, 1); 1927 if (err) 1928 return -EBUSY; 1929 1930 /* 1931 * It would be great if we could grab this wakeref from the 1932 * async unbind work if needed, but we can't because it uses 1933 * kmalloc and it's in the dma-fence signalling critical path. 1934 */ 1935 if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)) 1936 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm); 1937 1938 if (trylock_vm && !mutex_trylock(&vm->mutex)) { 1939 err = -EBUSY; 1940 goto out_rpm; 1941 } else if (!trylock_vm) { 1942 err = mutex_lock_interruptible_nested(&vm->mutex, !wakeref); 1943 if (err) 1944 goto out_rpm; 1945 } 1946 1947 fence = __i915_vma_unbind_async(vma); 1948 mutex_unlock(&vm->mutex); 1949 if (IS_ERR_OR_NULL(fence)) { 1950 err = PTR_ERR_OR_ZERO(fence); 1951 goto out_rpm; 1952 } 1953 1954 dma_resv_add_shared_fence(obj->base.resv, fence); 1955 dma_fence_put(fence); 1956 1957 out_rpm: 1958 if (wakeref) 1959 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref); 1960 return err; 1961 } 1962 1963 int i915_vma_unbind_unlocked(struct i915_vma *vma) 1964 { 1965 int err; 1966 1967 i915_gem_object_lock(vma->obj, NULL); 1968 err = i915_vma_unbind(vma); 1969 i915_gem_object_unlock(vma->obj); 1970 1971 return err; 1972 } 1973 1974 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma) 1975 { 1976 i915_gem_object_make_unshrinkable(vma->obj); 1977 return vma; 1978 } 1979 1980 void i915_vma_make_shrinkable(struct i915_vma *vma) 1981 { 1982 i915_gem_object_make_shrinkable(vma->obj); 1983 } 1984 1985 void i915_vma_make_purgeable(struct i915_vma *vma) 1986 { 1987 i915_gem_object_make_purgeable(vma->obj); 1988 } 1989 1990 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 1991 #include "selftests/i915_vma.c" 1992 #endif 1993 1994 void i915_vma_module_exit(void) 1995 { 1996 kmem_cache_destroy(slab_vmas); 1997 } 1998 1999 int __init i915_vma_module_init(void) 2000 { 2001 slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN); 2002 if (!slab_vmas) 2003 return -ENOMEM; 2004 2005 return 0; 2006 } 2007