drivers/gpu/drm/i915/i915_vma.c

   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 "i915_vma.h"
  26
  27 #include "i915_drv.h"
  28 #include "i915_globals.h"
  29 #include "intel_ringbuffer.h"
  30 #include "intel_frontbuffer.h"
  31
  32 #include <drm/drm_gem.h>
  33
  34 static struct i915_global_vma {
  35         struct i915_global base;
  36         struct kmem_cache *slab_vmas;
  37 } global;
  38
  39 struct i915_vma *i915_vma_alloc(void)
  40 {
  41         return kmem_cache_zalloc(global.slab_vmas, GFP_KERNEL);
  42 }
  43
  44 void i915_vma_free(struct i915_vma *vma)
  45 {
  46         return kmem_cache_free(global.slab_vmas, vma);
  47 }
  48
  49 #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM)
  50
  51 #include <linux/stackdepot.h>
  52
  53 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
  54 {
  55         unsigned long *entries;
  56         unsigned int nr_entries;
  57         char buf[512];
  58
  59         if (!vma->node.stack) {
  60                 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",
  61                                  vma->node.start, vma->node.size, reason);
  62                 return;
  63         }
  64
  65         nr_entries = stack_depot_fetch(vma->node.stack, &entries);
  66         stack_trace_snprint(buf, sizeof(buf), entries, nr_entries, 0);
  67         DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",
  68                          vma->node.start, vma->node.size, reason, buf);
  69 }
  70
  71 #else
  72
  73 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
  74 {
  75 }
  76
  77 #endif
  78
  79 static void obj_bump_mru(struct drm_i915_gem_object *obj)
  80 {
  81         struct drm_i915_private *i915 = to_i915(obj->base.dev);
  82
  83         spin_lock(&i915->mm.obj_lock);
  84         if (obj->bind_count)
  85                 list_move_tail(&obj->mm.link, &i915->mm.bound_list);
  86         spin_unlock(&i915->mm.obj_lock);
  87
  88         obj->mm.dirty = true; /* be paranoid  */
  89 }
  90
  91 static void __i915_vma_retire(struct i915_active *ref)
  92 {
  93         struct i915_vma *vma = container_of(ref, typeof(*vma), active);
  94         struct drm_i915_gem_object *obj = vma->obj;
  95
  96         GEM_BUG_ON(!i915_gem_object_is_active(obj));
  97         if (--obj->active_count)
  98                 return;
  99
 100         /* Prune the shared fence arrays iff completely idle (inc. external) */
 101         if (reservation_object_trylock(obj->resv)) {
 102                 if (reservation_object_test_signaled_rcu(obj->resv, true))
 103                         reservation_object_add_excl_fence(obj->resv, NULL);
 104                 reservation_object_unlock(obj->resv);
 105         }
 106
 107         /*
 108          * Bump our place on the bound list to keep it roughly in LRU order
 109          * so that we don't steal from recently used but inactive objects
 110          * (unless we are forced to ofc!)
 111          */
 112         obj_bump_mru(obj);
 113
 114         if (i915_gem_object_has_active_reference(obj)) {
 115                 i915_gem_object_clear_active_reference(obj);
 116                 i915_gem_object_put(obj);
 117         }
 118 }
 119
 120 static struct i915_vma *
 121 vma_create(struct drm_i915_gem_object *obj,
 122            struct i915_address_space *vm,
 123            const struct i915_ggtt_view *view)
 124 {
 125         struct i915_vma *vma;
 126         struct rb_node *rb, **p;
 127
 128         /* The aliasing_ppgtt should never be used directly! */
 129         GEM_BUG_ON(vm == &vm->i915->mm.aliasing_ppgtt->vm);
 130
 131         vma = i915_vma_alloc();
 132         if (vma == NULL)
 133                 return ERR_PTR(-ENOMEM);
 134
 135         i915_active_init(vm->i915, &vma->active, __i915_vma_retire);
 136         INIT_ACTIVE_REQUEST(&vma->last_fence);
 137
 138         vma->vm = vm;
 139         vma->ops = &vm->vma_ops;
 140         vma->obj = obj;
 141         vma->resv = obj->resv;
 142         vma->size = obj->base.size;
 143         vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
 144
 145         if (view && view->type != I915_GGTT_VIEW_NORMAL) {
 146                 vma->ggtt_view = *view;
 147                 if (view->type == I915_GGTT_VIEW_PARTIAL) {
 148                         GEM_BUG_ON(range_overflows_t(u64,
 149                                                      view->partial.offset,
 150                                                      view->partial.size,
 151                                                      obj->base.size >> PAGE_SHIFT));
 152                         vma->size = view->partial.size;
 153                         vma->size <<= PAGE_SHIFT;
 154                         GEM_BUG_ON(vma->size > obj->base.size);
 155                 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
 156                         vma->size = intel_rotation_info_size(&view->rotated);
 157                         vma->size <<= PAGE_SHIFT;
 158                 }
 159         }
 160
 161         if (unlikely(vma->size > vm->total))
 162                 goto err_vma;
 163
 164         GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
 165
 166         if (i915_is_ggtt(vm)) {
 167                 if (unlikely(overflows_type(vma->size, u32)))
 168                         goto err_vma;
 169
 170                 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
 171                                                       i915_gem_object_get_tiling(obj),
 172                                                       i915_gem_object_get_stride(obj));
 173                 if (unlikely(vma->fence_size < vma->size || /* overflow */
 174                              vma->fence_size > vm->total))
 175                         goto err_vma;
 176
 177                 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
 178
 179                 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
 180                                                                 i915_gem_object_get_tiling(obj),
 181                                                                 i915_gem_object_get_stride(obj));
 182                 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
 183
 184                 vma->flags |= I915_VMA_GGTT;
 185         }
 186
 187         spin_lock(&obj->vma.lock);
 188
 189         rb = NULL;
 190         p = &obj->vma.tree.rb_node;
 191         while (*p) {
 192                 struct i915_vma *pos;
 193                 long cmp;
 194
 195                 rb = *p;
 196                 pos = rb_entry(rb, struct i915_vma, obj_node);
 197
 198                 /*
 199                  * If the view already exists in the tree, another thread
 200                  * already created a matching vma, so return the older instance
 201                  * and dispose of ours.
 202                  */
 203                 cmp = i915_vma_compare(pos, vm, view);
 204                 if (cmp == 0) {
 205                         spin_unlock(&obj->vma.lock);
 206                         i915_vma_free(vma);
 207                         return pos;
 208                 }
 209
 210                 if (cmp < 0)
 211                         p = &rb->rb_right;
 212                 else
 213                         p = &rb->rb_left;
 214         }
 215         rb_link_node(&vma->obj_node, rb, p);
 216         rb_insert_color(&vma->obj_node, &obj->vma.tree);
 217
 218         if (i915_vma_is_ggtt(vma))
 219                 /*
 220                  * We put the GGTT vma at the start of the vma-list, followed
 221                  * by the ppGGTT vma. This allows us to break early when
 222                  * iterating over only the GGTT vma for an object, see
 223                  * for_each_ggtt_vma()
 224                  */
 225                 list_add(&vma->obj_link, &obj->vma.list);
 226         else
 227                 list_add_tail(&vma->obj_link, &obj->vma.list);
 228
 229         spin_unlock(&obj->vma.lock);
 230
 231         mutex_lock(&vm->mutex);
 232         list_add(&vma->vm_link, &vm->unbound_list);
 233         mutex_unlock(&vm->mutex);
 234
 235         return vma;
 236
 237 err_vma:
 238         i915_vma_free(vma);
 239         return ERR_PTR(-E2BIG);
 240 }
 241
 242 static struct i915_vma *
 243 vma_lookup(struct drm_i915_gem_object *obj,
 244            struct i915_address_space *vm,
 245            const struct i915_ggtt_view *view)
 246 {
 247         struct rb_node *rb;
 248
 249         rb = obj->vma.tree.rb_node;
 250         while (rb) {
 251                 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
 252                 long cmp;
 253
 254                 cmp = i915_vma_compare(vma, vm, view);
 255                 if (cmp == 0)
 256                         return vma;
 257
 258                 if (cmp < 0)
 259                         rb = rb->rb_right;
 260                 else
 261                         rb = rb->rb_left;
 262         }
 263
 264         return NULL;
 265 }
 266
 267 /**
 268  * i915_vma_instance - return the singleton instance of the VMA
 269  * @obj: parent &struct drm_i915_gem_object to be mapped
 270  * @vm: address space in which the mapping is located
 271  * @view: additional mapping requirements
 272  *
 273  * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
 274  * the same @view characteristics. If a match is not found, one is created.
 275  * Once created, the VMA is kept until either the object is freed, or the
 276  * address space is closed.
 277  *
 278  * Must be called with struct_mutex held.
 279  *
 280  * Returns the vma, or an error pointer.
 281  */
 282 struct i915_vma *
 283 i915_vma_instance(struct drm_i915_gem_object *obj,
 284                   struct i915_address_space *vm,
 285                   const struct i915_ggtt_view *view)
 286 {
 287         struct i915_vma *vma;
 288
 289         GEM_BUG_ON(view && !i915_is_ggtt(vm));
 290         GEM_BUG_ON(vm->closed);
 291
 292         spin_lock(&obj->vma.lock);
 293         vma = vma_lookup(obj, vm, view);
 294         spin_unlock(&obj->vma.lock);
 295
 296         /* vma_create() will resolve the race if another creates the vma */
 297         if (unlikely(!vma))
 298                 vma = vma_create(obj, vm, view);
 299
 300         GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
 301         return vma;
 302 }
 303
 304 /**
 305  * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
 306  * @vma: VMA to map
 307  * @cache_level: mapping cache level
 308  * @flags: flags like global or local mapping
 309  *
 310  * DMA addresses are taken from the scatter-gather table of this object (or of
 311  * this VMA in case of non-default GGTT views) and PTE entries set up.
 312  * Note that DMA addresses are also the only part of the SG table we care about.
 313  */
 314 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
 315                   u32 flags)
 316 {
 317         u32 bind_flags;
 318         u32 vma_flags;
 319         int ret;
 320
 321         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 322         GEM_BUG_ON(vma->size > vma->node.size);
 323
 324         if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
 325                                               vma->node.size,
 326                                               vma->vm->total)))
 327                 return -ENODEV;
 328
 329         if (GEM_DEBUG_WARN_ON(!flags))
 330                 return -EINVAL;
 331
 332         bind_flags = 0;
 333         if (flags & PIN_GLOBAL)
 334                 bind_flags |= I915_VMA_GLOBAL_BIND;
 335         if (flags & PIN_USER)
 336                 bind_flags |= I915_VMA_LOCAL_BIND;
 337
 338         vma_flags = vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
 339         if (flags & PIN_UPDATE)
 340                 bind_flags |= vma_flags;
 341         else
 342                 bind_flags &= ~vma_flags;
 343         if (bind_flags == 0)
 344                 return 0;
 345
 346         GEM_BUG_ON(!vma->pages);
 347
 348         trace_i915_vma_bind(vma, bind_flags);
 349         ret = vma->ops->bind_vma(vma, cache_level, bind_flags);
 350         if (ret)
 351                 return ret;
 352
 353         vma->flags |= bind_flags;
 354         return 0;
 355 }
 356
 357 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
 358 {
 359         void __iomem *ptr;
 360         int err;
 361
 362         /* Access through the GTT requires the device to be awake. */
 363         assert_rpm_wakelock_held(vma->vm->i915);
 364
 365         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 366         if (WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
 367                 err = -ENODEV;
 368                 goto err;
 369         }
 370
 371         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 372         GEM_BUG_ON((vma->flags & I915_VMA_GLOBAL_BIND) == 0);
 373
 374         ptr = vma->iomap;
 375         if (ptr == NULL) {
 376                 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
 377                                         vma->node.start,
 378                                         vma->node.size);
 379                 if (ptr == NULL) {
 380                         err = -ENOMEM;
 381                         goto err;
 382                 }
 383
 384                 vma->iomap = ptr;
 385         }
 386
 387         __i915_vma_pin(vma);
 388
 389         err = i915_vma_pin_fence(vma);
 390         if (err)
 391                 goto err_unpin;
 392
 393         i915_vma_set_ggtt_write(vma);
 394         return ptr;
 395
 396 err_unpin:
 397         __i915_vma_unpin(vma);
 398 err:
 399         return IO_ERR_PTR(err);
 400 }
 401
 402 void i915_vma_flush_writes(struct i915_vma *vma)
 403 {
 404         if (!i915_vma_has_ggtt_write(vma))
 405                 return;
 406
 407         i915_gem_flush_ggtt_writes(vma->vm->i915);
 408
 409         i915_vma_unset_ggtt_write(vma);
 410 }
 411
 412 void i915_vma_unpin_iomap(struct i915_vma *vma)
 413 {
 414         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 415
 416         GEM_BUG_ON(vma->iomap == NULL);
 417
 418         i915_vma_flush_writes(vma);
 419
 420         i915_vma_unpin_fence(vma);
 421         i915_vma_unpin(vma);
 422 }
 423
 424 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
 425 {
 426         struct i915_vma *vma;
 427         struct drm_i915_gem_object *obj;
 428
 429         vma = fetch_and_zero(p_vma);
 430         if (!vma)
 431                 return;
 432
 433         obj = vma->obj;
 434         GEM_BUG_ON(!obj);
 435
 436         i915_vma_unpin(vma);
 437         i915_vma_close(vma);
 438
 439         if (flags & I915_VMA_RELEASE_MAP)
 440                 i915_gem_object_unpin_map(obj);
 441
 442         __i915_gem_object_release_unless_active(obj);
 443 }
 444
 445 bool i915_vma_misplaced(const struct i915_vma *vma,
 446                         u64 size, u64 alignment, u64 flags)
 447 {
 448         if (!drm_mm_node_allocated(&vma->node))
 449                 return false;
 450
 451         if (vma->node.size < size)
 452                 return true;
 453
 454         GEM_BUG_ON(alignment && !is_power_of_2(alignment));
 455         if (alignment && !IS_ALIGNED(vma->node.start, alignment))
 456                 return true;
 457
 458         if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
 459                 return true;
 460
 461         if (flags & PIN_OFFSET_BIAS &&
 462             vma->node.start < (flags & PIN_OFFSET_MASK))
 463                 return true;
 464
 465         if (flags & PIN_OFFSET_FIXED &&
 466             vma->node.start != (flags & PIN_OFFSET_MASK))
 467                 return true;
 468
 469         return false;
 470 }
 471
 472 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
 473 {
 474         bool mappable, fenceable;
 475
 476         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 477         GEM_BUG_ON(!vma->fence_size);
 478
 479         /*
 480          * Explicitly disable for rotated VMA since the display does not
 481          * need the fence and the VMA is not accessible to other users.
 482          */
 483         if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
 484                 return;
 485
 486         fenceable = (vma->node.size >= vma->fence_size &&
 487                      IS_ALIGNED(vma->node.start, vma->fence_alignment));
 488
 489         mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
 490
 491         if (mappable && fenceable)
 492                 vma->flags |= I915_VMA_CAN_FENCE;
 493         else
 494                 vma->flags &= ~I915_VMA_CAN_FENCE;
 495 }
 496
 497 static bool color_differs(struct drm_mm_node *node, unsigned long color)
 498 {
 499         return node->allocated && node->color != color;
 500 }
 501
 502 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long cache_level)
 503 {
 504         struct drm_mm_node *node = &vma->node;
 505         struct drm_mm_node *other;
 506
 507         /*
 508          * On some machines we have to be careful when putting differing types
 509          * of snoopable memory together to avoid the prefetcher crossing memory
 510          * domains and dying. During vm initialisation, we decide whether or not
 511          * these constraints apply and set the drm_mm.color_adjust
 512          * appropriately.
 513          */
 514         if (vma->vm->mm.color_adjust == NULL)
 515                 return true;
 516
 517         /* Only valid to be called on an already inserted vma */
 518         GEM_BUG_ON(!drm_mm_node_allocated(node));
 519         GEM_BUG_ON(list_empty(&node->node_list));
 520
 521         other = list_prev_entry(node, node_list);
 522         if (color_differs(other, cache_level) && !drm_mm_hole_follows(other))
 523                 return false;
 524
 525         other = list_next_entry(node, node_list);
 526         if (color_differs(other, cache_level) && !drm_mm_hole_follows(node))
 527                 return false;
 528
 529         return true;
 530 }
 531
 532 static void assert_bind_count(const struct drm_i915_gem_object *obj)
 533 {
 534         /*
 535          * Combine the assertion that the object is bound and that we have
 536          * pinned its pages. But we should never have bound the object
 537          * more than we have pinned its pages. (For complete accuracy, we
 538          * assume that no else is pinning the pages, but as a rough assertion
 539          * that we will not run into problems later, this will do!)
 540          */
 541         GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < obj->bind_count);
 542 }
 543
 544 /**
 545  * i915_vma_insert - finds a slot for the vma in its address space
 546  * @vma: the vma
 547  * @size: requested size in bytes (can be larger than the VMA)
 548  * @alignment: required alignment
 549  * @flags: mask of PIN_* flags to use
 550  *
 551  * First we try to allocate some free space that meets the requirements for
 552  * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
 553  * preferrably the oldest idle entry to make room for the new VMA.
 554  *
 555  * Returns:
 556  * 0 on success, negative error code otherwise.
 557  */
 558 static int
 559 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 560 {
 561         struct drm_i915_private *dev_priv = vma->vm->i915;
 562         unsigned int cache_level;
 563         u64 start, end;
 564         int ret;
 565
 566         GEM_BUG_ON(i915_vma_is_closed(vma));
 567         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 568         GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
 569
 570         size = max(size, vma->size);
 571         alignment = max(alignment, vma->display_alignment);
 572         if (flags & PIN_MAPPABLE) {
 573                 size = max_t(typeof(size), size, vma->fence_size);
 574                 alignment = max_t(typeof(alignment),
 575                                   alignment, vma->fence_alignment);
 576         }
 577
 578         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
 579         GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
 580         GEM_BUG_ON(!is_power_of_2(alignment));
 581
 582         start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
 583         GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
 584
 585         end = vma->vm->total;
 586         if (flags & PIN_MAPPABLE)
 587                 end = min_t(u64, end, dev_priv->ggtt.mappable_end);
 588         if (flags & PIN_ZONE_4G)
 589                 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
 590         GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
 591
 592         /* If binding the object/GGTT view requires more space than the entire
 593          * aperture has, reject it early before evicting everything in a vain
 594          * attempt to find space.
 595          */
 596         if (size > end) {
 597                 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
 598                           size, flags & PIN_MAPPABLE ? "mappable" : "total",
 599                           end);
 600                 return -ENOSPC;
 601         }
 602
 603         if (vma->obj) {
 604                 ret = i915_gem_object_pin_pages(vma->obj);
 605                 if (ret)
 606                         return ret;
 607
 608                 cache_level = vma->obj->cache_level;
 609         } else {
 610                 cache_level = 0;
 611         }
 612
 613         GEM_BUG_ON(vma->pages);
 614
 615         ret = vma->ops->set_pages(vma);
 616         if (ret)
 617                 goto err_unpin;
 618
 619         if (flags & PIN_OFFSET_FIXED) {
 620                 u64 offset = flags & PIN_OFFSET_MASK;
 621                 if (!IS_ALIGNED(offset, alignment) ||
 622                     range_overflows(offset, size, end)) {
 623                         ret = -EINVAL;
 624                         goto err_clear;
 625                 }
 626
 627                 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
 628                                            size, offset, cache_level,
 629                                            flags);
 630                 if (ret)
 631                         goto err_clear;
 632         } else {
 633                 /*
 634                  * We only support huge gtt pages through the 48b PPGTT,
 635                  * however we also don't want to force any alignment for
 636                  * objects which need to be tightly packed into the low 32bits.
 637                  *
 638                  * Note that we assume that GGTT are limited to 4GiB for the
 639                  * forseeable future. See also i915_ggtt_offset().
 640                  */
 641                 if (upper_32_bits(end - 1) &&
 642                     vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
 643                         /*
 644                          * We can't mix 64K and 4K PTEs in the same page-table
 645                          * (2M block), and so to avoid the ugliness and
 646                          * complexity of coloring we opt for just aligning 64K
 647                          * objects to 2M.
 648                          */
 649                         u64 page_alignment =
 650                                 rounddown_pow_of_two(vma->page_sizes.sg |
 651                                                      I915_GTT_PAGE_SIZE_2M);
 652
 653                         /*
 654                          * Check we don't expand for the limited Global GTT
 655                          * (mappable aperture is even more precious!). This
 656                          * also checks that we exclude the aliasing-ppgtt.
 657                          */
 658                         GEM_BUG_ON(i915_vma_is_ggtt(vma));
 659
 660                         alignment = max(alignment, page_alignment);
 661
 662                         if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
 663                                 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
 664                 }
 665
 666                 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
 667                                           size, alignment, cache_level,
 668                                           start, end, flags);
 669                 if (ret)
 670                         goto err_clear;
 671
 672                 GEM_BUG_ON(vma->node.start < start);
 673                 GEM_BUG_ON(vma->node.start + vma->node.size > end);
 674         }
 675         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 676         GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, cache_level));
 677
 678         mutex_lock(&vma->vm->mutex);
 679         list_move_tail(&vma->vm_link, &vma->vm->bound_list);
 680         mutex_unlock(&vma->vm->mutex);
 681
 682         if (vma->obj) {
 683                 struct drm_i915_gem_object *obj = vma->obj;
 684
 685                 spin_lock(&dev_priv->mm.obj_lock);
 686                 list_move_tail(&obj->mm.link, &dev_priv->mm.bound_list);
 687                 obj->bind_count++;
 688                 spin_unlock(&dev_priv->mm.obj_lock);
 689
 690                 assert_bind_count(obj);
 691         }
 692
 693         return 0;
 694
 695 err_clear:
 696         vma->ops->clear_pages(vma);
 697 err_unpin:
 698         if (vma->obj)
 699                 i915_gem_object_unpin_pages(vma->obj);
 700         return ret;
 701 }
 702
 703 static void
 704 i915_vma_remove(struct i915_vma *vma)
 705 {
 706         struct drm_i915_private *i915 = vma->vm->i915;
 707
 708         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 709         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 710
 711         vma->ops->clear_pages(vma);
 712
 713         mutex_lock(&vma->vm->mutex);
 714         drm_mm_remove_node(&vma->node);
 715         list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
 716         mutex_unlock(&vma->vm->mutex);
 717
 718         /*
 719          * Since the unbound list is global, only move to that list if
 720          * no more VMAs exist.
 721          */
 722         if (vma->obj) {
 723                 struct drm_i915_gem_object *obj = vma->obj;
 724
 725                 spin_lock(&i915->mm.obj_lock);
 726                 if (--obj->bind_count == 0)
 727                         list_move_tail(&obj->mm.link, &i915->mm.unbound_list);
 728                 spin_unlock(&i915->mm.obj_lock);
 729
 730                 /*
 731                  * And finally now the object is completely decoupled from this
 732                  * vma, we can drop its hold on the backing storage and allow
 733                  * it to be reaped by the shrinker.
 734                  */
 735                 i915_gem_object_unpin_pages(obj);
 736                 assert_bind_count(obj);
 737         }
 738 }
 739
 740 int __i915_vma_do_pin(struct i915_vma *vma,
 741                       u64 size, u64 alignment, u64 flags)
 742 {
 743         const unsigned int bound = vma->flags;
 744         int ret;
 745
 746         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 747         GEM_BUG_ON((flags & (PIN_GLOBAL | PIN_USER)) == 0);
 748         GEM_BUG_ON((flags & PIN_GLOBAL) && !i915_vma_is_ggtt(vma));
 749
 750         if (WARN_ON(bound & I915_VMA_PIN_OVERFLOW)) {
 751                 ret = -EBUSY;
 752                 goto err_unpin;
 753         }
 754
 755         if ((bound & I915_VMA_BIND_MASK) == 0) {
 756                 ret = i915_vma_insert(vma, size, alignment, flags);
 757                 if (ret)
 758                         goto err_unpin;
 759         }
 760         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 761
 762         ret = i915_vma_bind(vma, vma->obj ? vma->obj->cache_level : 0, flags);
 763         if (ret)
 764                 goto err_remove;
 765
 766         GEM_BUG_ON((vma->flags & I915_VMA_BIND_MASK) == 0);
 767
 768         if ((bound ^ vma->flags) & I915_VMA_GLOBAL_BIND)
 769                 __i915_vma_set_map_and_fenceable(vma);
 770
 771         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
 772         return 0;
 773
 774 err_remove:
 775         if ((bound & I915_VMA_BIND_MASK) == 0) {
 776                 i915_vma_remove(vma);
 777                 GEM_BUG_ON(vma->pages);
 778                 GEM_BUG_ON(vma->flags & I915_VMA_BIND_MASK);
 779         }
 780 err_unpin:
 781         __i915_vma_unpin(vma);
 782         return ret;
 783 }
 784
 785 void i915_vma_close(struct i915_vma *vma)
 786 {
 787         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 788
 789         GEM_BUG_ON(i915_vma_is_closed(vma));
 790         vma->flags |= I915_VMA_CLOSED;
 791
 792         /*
 793          * We defer actually closing, unbinding and destroying the VMA until
 794          * the next idle point, or if the object is freed in the meantime. By
 795          * postponing the unbind, we allow for it to be resurrected by the
 796          * client, avoiding the work required to rebind the VMA. This is
 797          * advantageous for DRI, where the client/server pass objects
 798          * between themselves, temporarily opening a local VMA to the
 799          * object, and then closing it again. The same object is then reused
 800          * on the next frame (or two, depending on the depth of the swap queue)
 801          * causing us to rebind the VMA once more. This ends up being a lot
 802          * of wasted work for the steady state.
 803          */
 804         list_add_tail(&vma->closed_link, &vma->vm->i915->gt.closed_vma);
 805 }
 806
 807 void i915_vma_reopen(struct i915_vma *vma)
 808 {
 809         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 810
 811         if (vma->flags & I915_VMA_CLOSED) {
 812                 vma->flags &= ~I915_VMA_CLOSED;
 813                 list_del(&vma->closed_link);
 814         }
 815 }
 816
 817 static void __i915_vma_destroy(struct i915_vma *vma)
 818 {
 819         GEM_BUG_ON(vma->node.allocated);
 820         GEM_BUG_ON(vma->fence);
 821
 822         GEM_BUG_ON(i915_active_request_isset(&vma->last_fence));
 823
 824         mutex_lock(&vma->vm->mutex);
 825         list_del(&vma->vm_link);
 826         mutex_unlock(&vma->vm->mutex);
 827
 828         if (vma->obj) {
 829                 struct drm_i915_gem_object *obj = vma->obj;
 830
 831                 spin_lock(&obj->vma.lock);
 832                 list_del(&vma->obj_link);
 833                 rb_erase(&vma->obj_node, &vma->obj->vma.tree);
 834                 spin_unlock(&obj->vma.lock);
 835         }
 836
 837         i915_active_fini(&vma->active);
 838
 839         i915_vma_free(vma);
 840 }
 841
 842 void i915_vma_destroy(struct i915_vma *vma)
 843 {
 844         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 845
 846         GEM_BUG_ON(i915_vma_is_active(vma));
 847         GEM_BUG_ON(i915_vma_is_pinned(vma));
 848
 849         if (i915_vma_is_closed(vma))
 850                 list_del(&vma->closed_link);
 851
 852         WARN_ON(i915_vma_unbind(vma));
 853         __i915_vma_destroy(vma);
 854 }
 855
 856 void i915_vma_parked(struct drm_i915_private *i915)
 857 {
 858         struct i915_vma *vma, *next;
 859
 860         list_for_each_entry_safe(vma, next, &i915->gt.closed_vma, closed_link) {
 861                 GEM_BUG_ON(!i915_vma_is_closed(vma));
 862                 i915_vma_destroy(vma);
 863         }
 864
 865         GEM_BUG_ON(!list_empty(&i915->gt.closed_vma));
 866 }
 867
 868 static void __i915_vma_iounmap(struct i915_vma *vma)
 869 {
 870         GEM_BUG_ON(i915_vma_is_pinned(vma));
 871
 872         if (vma->iomap == NULL)
 873                 return;
 874
 875         io_mapping_unmap(vma->iomap);
 876         vma->iomap = NULL;
 877 }
 878
 879 void i915_vma_revoke_mmap(struct i915_vma *vma)
 880 {
 881         struct drm_vma_offset_node *node = &vma->obj->base.vma_node;
 882         u64 vma_offset;
 883
 884         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 885
 886         if (!i915_vma_has_userfault(vma))
 887                 return;
 888
 889         GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
 890         GEM_BUG_ON(!vma->obj->userfault_count);
 891
 892         vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
 893         unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
 894                             drm_vma_node_offset_addr(node) + vma_offset,
 895                             vma->size,
 896                             1);
 897
 898         i915_vma_unset_userfault(vma);
 899         if (!--vma->obj->userfault_count)
 900                 list_del(&vma->obj->userfault_link);
 901 }
 902
 903 static void export_fence(struct i915_vma *vma,
 904                          struct i915_request *rq,
 905                          unsigned int flags)
 906 {
 907         struct reservation_object *resv = vma->resv;
 908
 909         /*
 910          * Ignore errors from failing to allocate the new fence, we can't
 911          * handle an error right now. Worst case should be missed
 912          * synchronisation leading to rendering corruption.
 913          */
 914         reservation_object_lock(resv, NULL);
 915         if (flags & EXEC_OBJECT_WRITE)
 916                 reservation_object_add_excl_fence(resv, &rq->fence);
 917         else if (reservation_object_reserve_shared(resv, 1) == 0)
 918                 reservation_object_add_shared_fence(resv, &rq->fence);
 919         reservation_object_unlock(resv);
 920 }
 921
 922 int i915_vma_move_to_active(struct i915_vma *vma,
 923                             struct i915_request *rq,
 924                             unsigned int flags)
 925 {
 926         struct drm_i915_gem_object *obj = vma->obj;
 927
 928         lockdep_assert_held(&rq->i915->drm.struct_mutex);
 929         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 930
 931         /*
 932          * Add a reference if we're newly entering the active list.
 933          * The order in which we add operations to the retirement queue is
 934          * vital here: mark_active adds to the start of the callback list,
 935          * such that subsequent callbacks are called first. Therefore we
 936          * add the active reference first and queue for it to be dropped
 937          * *last*.
 938          */
 939         if (!vma->active.count)
 940                 obj->active_count++;
 941
 942         if (unlikely(i915_active_ref(&vma->active, rq->fence.context, rq))) {
 943                 if (!vma->active.count)
 944                         obj->active_count--;
 945                 return -ENOMEM;
 946         }
 947
 948         GEM_BUG_ON(!i915_vma_is_active(vma));
 949         GEM_BUG_ON(!obj->active_count);
 950
 951         obj->write_domain = 0;
 952         if (flags & EXEC_OBJECT_WRITE) {
 953                 obj->write_domain = I915_GEM_DOMAIN_RENDER;
 954
 955                 if (intel_fb_obj_invalidate(obj, ORIGIN_CS))
 956                         __i915_active_request_set(&obj->frontbuffer_write, rq);
 957
 958                 obj->read_domains = 0;
 959         }
 960         obj->read_domains |= I915_GEM_GPU_DOMAINS;
 961
 962         if (flags & EXEC_OBJECT_NEEDS_FENCE)
 963                 __i915_active_request_set(&vma->last_fence, rq);
 964
 965         export_fence(vma, rq, flags);
 966         return 0;
 967 }
 968
 969 int i915_vma_unbind(struct i915_vma *vma)
 970 {
 971         int ret;
 972
 973         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 974
 975         /*
 976          * First wait upon any activity as retiring the request may
 977          * have side-effects such as unpinning or even unbinding this vma.
 978          */
 979         might_sleep();
 980         if (i915_vma_is_active(vma)) {
 981                 /*
 982                  * When a closed VMA is retired, it is unbound - eek.
 983                  * In order to prevent it from being recursively closed,
 984                  * take a pin on the vma so that the second unbind is
 985                  * aborted.
 986                  *
 987                  * Even more scary is that the retire callback may free
 988                  * the object (last active vma). To prevent the explosion
 989                  * we defer the actual object free to a worker that can
 990                  * only proceed once it acquires the struct_mutex (which
 991                  * we currently hold, therefore it cannot free this object
 992                  * before we are finished).
 993                  */
 994                 __i915_vma_pin(vma);
 995
 996                 ret = i915_active_wait(&vma->active);
 997                 if (ret)
 998                         goto unpin;
 999
1000                 ret = i915_active_request_retire(&vma->last_fence,
1001                                               &vma->vm->i915->drm.struct_mutex);
1002 unpin:
1003                 __i915_vma_unpin(vma);
1004                 if (ret)
1005                         return ret;
1006         }
1007         GEM_BUG_ON(i915_vma_is_active(vma));
1008
1009         if (i915_vma_is_pinned(vma)) {
1010                 vma_print_allocator(vma, "is pinned");
1011                 return -EBUSY;
1012         }
1013
1014         if (!drm_mm_node_allocated(&vma->node))
1015                 return 0;
1016
1017         if (i915_vma_is_map_and_fenceable(vma)) {
1018                 /*
1019                  * Check that we have flushed all writes through the GGTT
1020                  * before the unbind, other due to non-strict nature of those
1021                  * indirect writes they may end up referencing the GGTT PTE
1022                  * after the unbind.
1023                  */
1024                 i915_vma_flush_writes(vma);
1025                 GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
1026
1027                 /* release the fence reg _after_ flushing */
1028                 ret = i915_vma_put_fence(vma);
1029                 if (ret)
1030                         return ret;
1031
1032                 /* Force a pagefault for domain tracking on next user access */
1033                 i915_vma_revoke_mmap(vma);
1034
1035                 __i915_vma_iounmap(vma);
1036                 vma->flags &= ~I915_VMA_CAN_FENCE;
1037         }
1038         GEM_BUG_ON(vma->fence);
1039         GEM_BUG_ON(i915_vma_has_userfault(vma));
1040
1041         if (likely(!vma->vm->closed)) {
1042                 trace_i915_vma_unbind(vma);
1043                 vma->ops->unbind_vma(vma);
1044         }
1045         vma->flags &= ~(I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
1046
1047         i915_vma_remove(vma);
1048
1049         return 0;
1050 }
1051
1052 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1053 #include "selftests/i915_vma.c"
1054 #endif
1055
1056 static void i915_global_vma_shrink(void)
1057 {
1058         kmem_cache_shrink(global.slab_vmas);
1059 }
1060
1061 static void i915_global_vma_exit(void)
1062 {
1063         kmem_cache_destroy(global.slab_vmas);
1064 }
1065
1066 static struct i915_global_vma global = { {
1067         .shrink = i915_global_vma_shrink,
1068         .exit = i915_global_vma_exit,
1069 } };
1070
1071 int __init i915_global_vma_init(void)
1072 {
1073         global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1074         if (!global.slab_vmas)
1075                 return -ENOMEM;
1076
1077         i915_global_register(&global.base);
1078         return 0;
1079 }