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 "intel_ringbuffer.h"
  29 #include "intel_frontbuffer.h"
  30
  31 #include <drm/drm_gem.h>
  32
  33 #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM)
  34
  35 #include <linux/stackdepot.h>
  36
  37 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
  38 {
  39         unsigned long entries[12];
  40         struct stack_trace trace = {
  41                 .entries = entries,
  42                 .max_entries = ARRAY_SIZE(entries),
  43         };
  44         char buf[512];
  45
  46         if (!vma->node.stack) {
  47                 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",
  48                                  vma->node.start, vma->node.size, reason);
  49                 return;
  50         }
  51
  52         depot_fetch_stack(vma->node.stack, &trace);
  53         snprint_stack_trace(buf, sizeof(buf), &trace, 0);
  54         DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",
  55                          vma->node.start, vma->node.size, reason, buf);
  56 }
  57
  58 #else
  59
  60 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
  61 {
  62 }
  63
  64 #endif
  65
  66 static void obj_bump_mru(struct drm_i915_gem_object *obj)
  67 {
  68         struct drm_i915_private *i915 = to_i915(obj->base.dev);
  69
  70         spin_lock(&i915->mm.obj_lock);
  71         if (obj->bind_count)
  72                 list_move_tail(&obj->mm.link, &i915->mm.bound_list);
  73         spin_unlock(&i915->mm.obj_lock);
  74
  75         obj->mm.dirty = true; /* be paranoid  */
  76 }
  77
  78 static void __i915_vma_retire(struct i915_active *ref)
  79 {
  80         struct i915_vma *vma = container_of(ref, typeof(*vma), active);
  81         struct drm_i915_gem_object *obj = vma->obj;
  82
  83         GEM_BUG_ON(!i915_gem_object_is_active(obj));
  84         if (--obj->active_count)
  85                 return;
  86
  87         /* Prune the shared fence arrays iff completely idle (inc. external) */
  88         if (reservation_object_trylock(obj->resv)) {
  89                 if (reservation_object_test_signaled_rcu(obj->resv, true))
  90                         reservation_object_add_excl_fence(obj->resv, NULL);
  91                 reservation_object_unlock(obj->resv);
  92         }
  93
  94         /*
  95          * Bump our place on the bound list to keep it roughly in LRU order
  96          * so that we don't steal from recently used but inactive objects
  97          * (unless we are forced to ofc!)
  98          */
  99         obj_bump_mru(obj);
 100
 101         if (i915_gem_object_has_active_reference(obj)) {
 102                 i915_gem_object_clear_active_reference(obj);
 103                 i915_gem_object_put(obj);
 104         }
 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 *vma;
 113         struct rb_node *rb, **p;
 114
 115         /* The aliasing_ppgtt should never be used directly! */
 116         GEM_BUG_ON(vm == &vm->i915->mm.aliasing_ppgtt->vm);
 117
 118         vma = kmem_cache_zalloc(vm->i915->vmas, GFP_KERNEL);
 119         if (vma == NULL)
 120                 return ERR_PTR(-ENOMEM);
 121
 122         i915_active_init(vm->i915, &vma->active, __i915_vma_retire);
 123         INIT_ACTIVE_REQUEST(&vma->last_fence);
 124
 125         vma->vm = vm;
 126         vma->ops = &vm->vma_ops;
 127         vma->obj = obj;
 128         vma->resv = obj->resv;
 129         vma->size = obj->base.size;
 130         vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
 131
 132         if (view && view->type != I915_GGTT_VIEW_NORMAL) {
 133                 vma->ggtt_view = *view;
 134                 if (view->type == I915_GGTT_VIEW_PARTIAL) {
 135                         GEM_BUG_ON(range_overflows_t(u64,
 136                                                      view->partial.offset,
 137                                                      view->partial.size,
 138                                                      obj->base.size >> PAGE_SHIFT));
 139                         vma->size = view->partial.size;
 140                         vma->size <<= PAGE_SHIFT;
 141                         GEM_BUG_ON(vma->size > obj->base.size);
 142                 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
 143                         vma->size = intel_rotation_info_size(&view->rotated);
 144                         vma->size <<= PAGE_SHIFT;
 145                 }
 146         }
 147
 148         if (unlikely(vma->size > vm->total))
 149                 goto err_vma;
 150
 151         GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
 152
 153         if (i915_is_ggtt(vm)) {
 154                 if (unlikely(overflows_type(vma->size, u32)))
 155                         goto err_vma;
 156
 157                 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
 158                                                       i915_gem_object_get_tiling(obj),
 159                                                       i915_gem_object_get_stride(obj));
 160                 if (unlikely(vma->fence_size < vma->size || /* overflow */
 161                              vma->fence_size > vm->total))
 162                         goto err_vma;
 163
 164                 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
 165
 166                 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
 167                                                                 i915_gem_object_get_tiling(obj),
 168                                                                 i915_gem_object_get_stride(obj));
 169                 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
 170
 171                 vma->flags |= I915_VMA_GGTT;
 172         }
 173
 174         spin_lock(&obj->vma.lock);
 175
 176         rb = NULL;
 177         p = &obj->vma.tree.rb_node;
 178         while (*p) {
 179                 struct i915_vma *pos;
 180                 long cmp;
 181
 182                 rb = *p;
 183                 pos = rb_entry(rb, struct i915_vma, obj_node);
 184
 185                 /*
 186                  * If the view already exists in the tree, another thread
 187                  * already created a matching vma, so return the older instance
 188                  * and dispose of ours.
 189                  */
 190                 cmp = i915_vma_compare(pos, vm, view);
 191                 if (cmp == 0) {
 192                         spin_unlock(&obj->vma.lock);
 193                         kmem_cache_free(vm->i915->vmas, vma);
 194                         return pos;
 195                 }
 196
 197                 if (cmp < 0)
 198                         p = &rb->rb_right;
 199                 else
 200                         p = &rb->rb_left;
 201         }
 202         rb_link_node(&vma->obj_node, rb, p);
 203         rb_insert_color(&vma->obj_node, &obj->vma.tree);
 204
 205         if (i915_vma_is_ggtt(vma))
 206                 /*
 207                  * We put the GGTT vma at the start of the vma-list, followed
 208                  * by the ppGGTT vma. This allows us to break early when
 209                  * iterating over only the GGTT vma for an object, see
 210                  * for_each_ggtt_vma()
 211                  */
 212                 list_add(&vma->obj_link, &obj->vma.list);
 213         else
 214                 list_add_tail(&vma->obj_link, &obj->vma.list);
 215
 216         spin_unlock(&obj->vma.lock);
 217
 218         mutex_lock(&vm->mutex);
 219         list_add(&vma->vm_link, &vm->unbound_list);
 220         mutex_unlock(&vm->mutex);
 221
 222         return vma;
 223
 224 err_vma:
 225         kmem_cache_free(vm->i915->vmas, vma);
 226         return ERR_PTR(-E2BIG);
 227 }
 228
 229 static struct i915_vma *
 230 vma_lookup(struct drm_i915_gem_object *obj,
 231            struct i915_address_space *vm,
 232            const struct i915_ggtt_view *view)
 233 {
 234         struct rb_node *rb;
 235
 236         rb = obj->vma.tree.rb_node;
 237         while (rb) {
 238                 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
 239                 long cmp;
 240
 241                 cmp = i915_vma_compare(vma, vm, view);
 242                 if (cmp == 0)
 243                         return vma;
 244
 245                 if (cmp < 0)
 246                         rb = rb->rb_right;
 247                 else
 248                         rb = rb->rb_left;
 249         }
 250
 251         return NULL;
 252 }
 253
 254 /**
 255  * i915_vma_instance - return the singleton instance of the VMA
 256  * @obj: parent &struct drm_i915_gem_object to be mapped
 257  * @vm: address space in which the mapping is located
 258  * @view: additional mapping requirements
 259  *
 260  * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
 261  * the same @view characteristics. If a match is not found, one is created.
 262  * Once created, the VMA is kept until either the object is freed, or the
 263  * address space is closed.
 264  *
 265  * Must be called with struct_mutex held.
 266  *
 267  * Returns the vma, or an error pointer.
 268  */
 269 struct i915_vma *
 270 i915_vma_instance(struct drm_i915_gem_object *obj,
 271                   struct i915_address_space *vm,
 272                   const struct i915_ggtt_view *view)
 273 {
 274         struct i915_vma *vma;
 275
 276         GEM_BUG_ON(view && !i915_is_ggtt(vm));
 277         GEM_BUG_ON(vm->closed);
 278
 279         spin_lock(&obj->vma.lock);
 280         vma = vma_lookup(obj, vm, view);
 281         spin_unlock(&obj->vma.lock);
 282
 283         /* vma_create() will resolve the race if another creates the vma */
 284         if (unlikely(!vma))
 285                 vma = vma_create(obj, vm, view);
 286
 287         GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
 288         return vma;
 289 }
 290
 291 /**
 292  * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
 293  * @vma: VMA to map
 294  * @cache_level: mapping cache level
 295  * @flags: flags like global or local mapping
 296  *
 297  * DMA addresses are taken from the scatter-gather table of this object (or of
 298  * this VMA in case of non-default GGTT views) and PTE entries set up.
 299  * Note that DMA addresses are also the only part of the SG table we care about.
 300  */
 301 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
 302                   u32 flags)
 303 {
 304         u32 bind_flags;
 305         u32 vma_flags;
 306         int ret;
 307
 308         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 309         GEM_BUG_ON(vma->size > vma->node.size);
 310
 311         if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
 312                                               vma->node.size,
 313                                               vma->vm->total)))
 314                 return -ENODEV;
 315
 316         if (GEM_DEBUG_WARN_ON(!flags))
 317                 return -EINVAL;
 318
 319         bind_flags = 0;
 320         if (flags & PIN_GLOBAL)
 321                 bind_flags |= I915_VMA_GLOBAL_BIND;
 322         if (flags & PIN_USER)
 323                 bind_flags |= I915_VMA_LOCAL_BIND;
 324
 325         vma_flags = vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
 326         if (flags & PIN_UPDATE)
 327                 bind_flags |= vma_flags;
 328         else
 329                 bind_flags &= ~vma_flags;
 330         if (bind_flags == 0)
 331                 return 0;
 332
 333         GEM_BUG_ON(!vma->pages);
 334
 335         trace_i915_vma_bind(vma, bind_flags);
 336         ret = vma->ops->bind_vma(vma, cache_level, bind_flags);
 337         if (ret)
 338                 return ret;
 339
 340         vma->flags |= bind_flags;
 341         return 0;
 342 }
 343
 344 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
 345 {
 346         void __iomem *ptr;
 347         int err;
 348
 349         /* Access through the GTT requires the device to be awake. */
 350         assert_rpm_wakelock_held(vma->vm->i915);
 351
 352         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 353         if (WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
 354                 err = -ENODEV;
 355                 goto err;
 356         }
 357
 358         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 359         GEM_BUG_ON((vma->flags & I915_VMA_GLOBAL_BIND) == 0);
 360
 361         ptr = vma->iomap;
 362         if (ptr == NULL) {
 363                 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
 364                                         vma->node.start,
 365                                         vma->node.size);
 366                 if (ptr == NULL) {
 367                         err = -ENOMEM;
 368                         goto err;
 369                 }
 370
 371                 vma->iomap = ptr;
 372         }
 373
 374         __i915_vma_pin(vma);
 375
 376         err = i915_vma_pin_fence(vma);
 377         if (err)
 378                 goto err_unpin;
 379
 380         i915_vma_set_ggtt_write(vma);
 381         return ptr;
 382
 383 err_unpin:
 384         __i915_vma_unpin(vma);
 385 err:
 386         return IO_ERR_PTR(err);
 387 }
 388
 389 void i915_vma_flush_writes(struct i915_vma *vma)
 390 {
 391         if (!i915_vma_has_ggtt_write(vma))
 392                 return;
 393
 394         i915_gem_flush_ggtt_writes(vma->vm->i915);
 395
 396         i915_vma_unset_ggtt_write(vma);
 397 }
 398
 399 void i915_vma_unpin_iomap(struct i915_vma *vma)
 400 {
 401         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 402
 403         GEM_BUG_ON(vma->iomap == NULL);
 404
 405         i915_vma_flush_writes(vma);
 406
 407         i915_vma_unpin_fence(vma);
 408         i915_vma_unpin(vma);
 409 }
 410
 411 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
 412 {
 413         struct i915_vma *vma;
 414         struct drm_i915_gem_object *obj;
 415
 416         vma = fetch_and_zero(p_vma);
 417         if (!vma)
 418                 return;
 419
 420         obj = vma->obj;
 421         GEM_BUG_ON(!obj);
 422
 423         i915_vma_unpin(vma);
 424         i915_vma_close(vma);
 425
 426         if (flags & I915_VMA_RELEASE_MAP)
 427                 i915_gem_object_unpin_map(obj);
 428
 429         __i915_gem_object_release_unless_active(obj);
 430 }
 431
 432 bool i915_vma_misplaced(const struct i915_vma *vma,
 433                         u64 size, u64 alignment, u64 flags)
 434 {
 435         if (!drm_mm_node_allocated(&vma->node))
 436                 return false;
 437
 438         if (vma->node.size < size)
 439                 return true;
 440
 441         GEM_BUG_ON(alignment && !is_power_of_2(alignment));
 442         if (alignment && !IS_ALIGNED(vma->node.start, alignment))
 443                 return true;
 444
 445         if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
 446                 return true;
 447
 448         if (flags & PIN_OFFSET_BIAS &&
 449             vma->node.start < (flags & PIN_OFFSET_MASK))
 450                 return true;
 451
 452         if (flags & PIN_OFFSET_FIXED &&
 453             vma->node.start != (flags & PIN_OFFSET_MASK))
 454                 return true;
 455
 456         return false;
 457 }
 458
 459 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
 460 {
 461         bool mappable, fenceable;
 462
 463         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 464         GEM_BUG_ON(!vma->fence_size);
 465
 466         /*
 467          * Explicitly disable for rotated VMA since the display does not
 468          * need the fence and the VMA is not accessible to other users.
 469          */
 470         if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
 471                 return;
 472
 473         fenceable = (vma->node.size >= vma->fence_size &&
 474                      IS_ALIGNED(vma->node.start, vma->fence_alignment));
 475
 476         mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
 477
 478         if (mappable && fenceable)
 479                 vma->flags |= I915_VMA_CAN_FENCE;
 480         else
 481                 vma->flags &= ~I915_VMA_CAN_FENCE;
 482 }
 483
 484 static bool color_differs(struct drm_mm_node *node, unsigned long color)
 485 {
 486         return node->allocated && node->color != color;
 487 }
 488
 489 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long cache_level)
 490 {
 491         struct drm_mm_node *node = &vma->node;
 492         struct drm_mm_node *other;
 493
 494         /*
 495          * On some machines we have to be careful when putting differing types
 496          * of snoopable memory together to avoid the prefetcher crossing memory
 497          * domains and dying. During vm initialisation, we decide whether or not
 498          * these constraints apply and set the drm_mm.color_adjust
 499          * appropriately.
 500          */
 501         if (vma->vm->mm.color_adjust == NULL)
 502                 return true;
 503
 504         /* Only valid to be called on an already inserted vma */
 505         GEM_BUG_ON(!drm_mm_node_allocated(node));
 506         GEM_BUG_ON(list_empty(&node->node_list));
 507
 508         other = list_prev_entry(node, node_list);
 509         if (color_differs(other, cache_level) && !drm_mm_hole_follows(other))
 510                 return false;
 511
 512         other = list_next_entry(node, node_list);
 513         if (color_differs(other, cache_level) && !drm_mm_hole_follows(node))
 514                 return false;
 515
 516         return true;
 517 }
 518
 519 static void assert_bind_count(const struct drm_i915_gem_object *obj)
 520 {
 521         /*
 522          * Combine the assertion that the object is bound and that we have
 523          * pinned its pages. But we should never have bound the object
 524          * more than we have pinned its pages. (For complete accuracy, we
 525          * assume that no else is pinning the pages, but as a rough assertion
 526          * that we will not run into problems later, this will do!)
 527          */
 528         GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < obj->bind_count);
 529 }
 530
 531 /**
 532  * i915_vma_insert - finds a slot for the vma in its address space
 533  * @vma: the vma
 534  * @size: requested size in bytes (can be larger than the VMA)
 535  * @alignment: required alignment
 536  * @flags: mask of PIN_* flags to use
 537  *
 538  * First we try to allocate some free space that meets the requirements for
 539  * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
 540  * preferrably the oldest idle entry to make room for the new VMA.
 541  *
 542  * Returns:
 543  * 0 on success, negative error code otherwise.
 544  */
 545 static int
 546 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 547 {
 548         struct drm_i915_private *dev_priv = vma->vm->i915;
 549         unsigned int cache_level;
 550         u64 start, end;
 551         int ret;
 552
 553         GEM_BUG_ON(i915_vma_is_closed(vma));
 554         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 555         GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
 556
 557         size = max(size, vma->size);
 558         alignment = max(alignment, vma->display_alignment);
 559         if (flags & PIN_MAPPABLE) {
 560                 size = max_t(typeof(size), size, vma->fence_size);
 561                 alignment = max_t(typeof(alignment),
 562                                   alignment, vma->fence_alignment);
 563         }
 564
 565         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
 566         GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
 567         GEM_BUG_ON(!is_power_of_2(alignment));
 568
 569         start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
 570         GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
 571
 572         end = vma->vm->total;
 573         if (flags & PIN_MAPPABLE)
 574                 end = min_t(u64, end, dev_priv->ggtt.mappable_end);
 575         if (flags & PIN_ZONE_4G)
 576                 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
 577         GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
 578
 579         /* If binding the object/GGTT view requires more space than the entire
 580          * aperture has, reject it early before evicting everything in a vain
 581          * attempt to find space.
 582          */
 583         if (size > end) {
 584                 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
 585                           size, flags & PIN_MAPPABLE ? "mappable" : "total",
 586                           end);
 587                 return -ENOSPC;
 588         }
 589
 590         if (vma->obj) {
 591                 ret = i915_gem_object_pin_pages(vma->obj);
 592                 if (ret)
 593                         return ret;
 594
 595                 cache_level = vma->obj->cache_level;
 596         } else {
 597                 cache_level = 0;
 598         }
 599
 600         GEM_BUG_ON(vma->pages);
 601
 602         ret = vma->ops->set_pages(vma);
 603         if (ret)
 604                 goto err_unpin;
 605
 606         if (flags & PIN_OFFSET_FIXED) {
 607                 u64 offset = flags & PIN_OFFSET_MASK;
 608                 if (!IS_ALIGNED(offset, alignment) ||
 609                     range_overflows(offset, size, end)) {
 610                         ret = -EINVAL;
 611                         goto err_clear;
 612                 }
 613
 614                 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
 615                                            size, offset, cache_level,
 616                                            flags);
 617                 if (ret)
 618                         goto err_clear;
 619         } else {
 620                 /*
 621                  * We only support huge gtt pages through the 48b PPGTT,
 622                  * however we also don't want to force any alignment for
 623                  * objects which need to be tightly packed into the low 32bits.
 624                  *
 625                  * Note that we assume that GGTT are limited to 4GiB for the
 626                  * forseeable future. See also i915_ggtt_offset().
 627                  */
 628                 if (upper_32_bits(end - 1) &&
 629                     vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
 630                         /*
 631                          * We can't mix 64K and 4K PTEs in the same page-table
 632                          * (2M block), and so to avoid the ugliness and
 633                          * complexity of coloring we opt for just aligning 64K
 634                          * objects to 2M.
 635                          */
 636                         u64 page_alignment =
 637                                 rounddown_pow_of_two(vma->page_sizes.sg |
 638                                                      I915_GTT_PAGE_SIZE_2M);
 639
 640                         /*
 641                          * Check we don't expand for the limited Global GTT
 642                          * (mappable aperture is even more precious!). This
 643                          * also checks that we exclude the aliasing-ppgtt.
 644                          */
 645                         GEM_BUG_ON(i915_vma_is_ggtt(vma));
 646
 647                         alignment = max(alignment, page_alignment);
 648
 649                         if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
 650                                 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
 651                 }
 652
 653                 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
 654                                           size, alignment, cache_level,
 655                                           start, end, flags);
 656                 if (ret)
 657                         goto err_clear;
 658
 659                 GEM_BUG_ON(vma->node.start < start);
 660                 GEM_BUG_ON(vma->node.start + vma->node.size > end);
 661         }
 662         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 663         GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, cache_level));
 664
 665         mutex_lock(&vma->vm->mutex);
 666         list_move_tail(&vma->vm_link, &vma->vm->bound_list);
 667         mutex_unlock(&vma->vm->mutex);
 668
 669         if (vma->obj) {
 670                 struct drm_i915_gem_object *obj = vma->obj;
 671
 672                 spin_lock(&dev_priv->mm.obj_lock);
 673                 list_move_tail(&obj->mm.link, &dev_priv->mm.bound_list);
 674                 obj->bind_count++;
 675                 spin_unlock(&dev_priv->mm.obj_lock);
 676
 677                 assert_bind_count(obj);
 678         }
 679
 680         return 0;
 681
 682 err_clear:
 683         vma->ops->clear_pages(vma);
 684 err_unpin:
 685         if (vma->obj)
 686                 i915_gem_object_unpin_pages(vma->obj);
 687         return ret;
 688 }
 689
 690 static void
 691 i915_vma_remove(struct i915_vma *vma)
 692 {
 693         struct drm_i915_private *i915 = vma->vm->i915;
 694
 695         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 696         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 697
 698         vma->ops->clear_pages(vma);
 699
 700         mutex_lock(&vma->vm->mutex);
 701         drm_mm_remove_node(&vma->node);
 702         list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
 703         mutex_unlock(&vma->vm->mutex);
 704
 705         /*
 706          * Since the unbound list is global, only move to that list if
 707          * no more VMAs exist.
 708          */
 709         if (vma->obj) {
 710                 struct drm_i915_gem_object *obj = vma->obj;
 711
 712                 spin_lock(&i915->mm.obj_lock);
 713                 if (--obj->bind_count == 0)
 714                         list_move_tail(&obj->mm.link, &i915->mm.unbound_list);
 715                 spin_unlock(&i915->mm.obj_lock);
 716
 717                 /*
 718                  * And finally now the object is completely decoupled from this
 719                  * vma, we can drop its hold on the backing storage and allow
 720                  * it to be reaped by the shrinker.
 721                  */
 722                 i915_gem_object_unpin_pages(obj);
 723                 assert_bind_count(obj);
 724         }
 725 }
 726
 727 int __i915_vma_do_pin(struct i915_vma *vma,
 728                       u64 size, u64 alignment, u64 flags)
 729 {
 730         const unsigned int bound = vma->flags;
 731         int ret;
 732
 733         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 734         GEM_BUG_ON((flags & (PIN_GLOBAL | PIN_USER)) == 0);
 735         GEM_BUG_ON((flags & PIN_GLOBAL) && !i915_vma_is_ggtt(vma));
 736
 737         if (WARN_ON(bound & I915_VMA_PIN_OVERFLOW)) {
 738                 ret = -EBUSY;
 739                 goto err_unpin;
 740         }
 741
 742         if ((bound & I915_VMA_BIND_MASK) == 0) {
 743                 ret = i915_vma_insert(vma, size, alignment, flags);
 744                 if (ret)
 745                         goto err_unpin;
 746         }
 747         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 748
 749         ret = i915_vma_bind(vma, vma->obj ? vma->obj->cache_level : 0, flags);
 750         if (ret)
 751                 goto err_remove;
 752
 753         GEM_BUG_ON((vma->flags & I915_VMA_BIND_MASK) == 0);
 754
 755         if ((bound ^ vma->flags) & I915_VMA_GLOBAL_BIND)
 756                 __i915_vma_set_map_and_fenceable(vma);
 757
 758         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
 759         return 0;
 760
 761 err_remove:
 762         if ((bound & I915_VMA_BIND_MASK) == 0) {
 763                 i915_vma_remove(vma);
 764                 GEM_BUG_ON(vma->pages);
 765                 GEM_BUG_ON(vma->flags & I915_VMA_BIND_MASK);
 766         }
 767 err_unpin:
 768         __i915_vma_unpin(vma);
 769         return ret;
 770 }
 771
 772 void i915_vma_close(struct i915_vma *vma)
 773 {
 774         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 775
 776         GEM_BUG_ON(i915_vma_is_closed(vma));
 777         vma->flags |= I915_VMA_CLOSED;
 778
 779         /*
 780          * We defer actually closing, unbinding and destroying the VMA until
 781          * the next idle point, or if the object is freed in the meantime. By
 782          * postponing the unbind, we allow for it to be resurrected by the
 783          * client, avoiding the work required to rebind the VMA. This is
 784          * advantageous for DRI, where the client/server pass objects
 785          * between themselves, temporarily opening a local VMA to the
 786          * object, and then closing it again. The same object is then reused
 787          * on the next frame (or two, depending on the depth of the swap queue)
 788          * causing us to rebind the VMA once more. This ends up being a lot
 789          * of wasted work for the steady state.
 790          */
 791         list_add_tail(&vma->closed_link, &vma->vm->i915->gt.closed_vma);
 792 }
 793
 794 void i915_vma_reopen(struct i915_vma *vma)
 795 {
 796         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 797
 798         if (vma->flags & I915_VMA_CLOSED) {
 799                 vma->flags &= ~I915_VMA_CLOSED;
 800                 list_del(&vma->closed_link);
 801         }
 802 }
 803
 804 static void __i915_vma_destroy(struct i915_vma *vma)
 805 {
 806         struct drm_i915_private *i915 = vma->vm->i915;
 807
 808         GEM_BUG_ON(vma->node.allocated);
 809         GEM_BUG_ON(vma->fence);
 810
 811         GEM_BUG_ON(i915_active_request_isset(&vma->last_fence));
 812
 813         mutex_lock(&vma->vm->mutex);
 814         list_del(&vma->vm_link);
 815         mutex_unlock(&vma->vm->mutex);
 816
 817         if (vma->obj) {
 818                 struct drm_i915_gem_object *obj = vma->obj;
 819
 820                 spin_lock(&obj->vma.lock);
 821                 list_del(&vma->obj_link);
 822                 rb_erase(&vma->obj_node, &vma->obj->vma.tree);
 823                 spin_unlock(&obj->vma.lock);
 824         }
 825
 826         i915_active_fini(&vma->active);
 827
 828         kmem_cache_free(i915->vmas, vma);
 829 }
 830
 831 void i915_vma_destroy(struct i915_vma *vma)
 832 {
 833         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 834
 835         GEM_BUG_ON(i915_vma_is_active(vma));
 836         GEM_BUG_ON(i915_vma_is_pinned(vma));
 837
 838         if (i915_vma_is_closed(vma))
 839                 list_del(&vma->closed_link);
 840
 841         WARN_ON(i915_vma_unbind(vma));
 842         __i915_vma_destroy(vma);
 843 }
 844
 845 void i915_vma_parked(struct drm_i915_private *i915)
 846 {
 847         struct i915_vma *vma, *next;
 848
 849         list_for_each_entry_safe(vma, next, &i915->gt.closed_vma, closed_link) {
 850                 GEM_BUG_ON(!i915_vma_is_closed(vma));
 851                 i915_vma_destroy(vma);
 852         }
 853
 854         GEM_BUG_ON(!list_empty(&i915->gt.closed_vma));
 855 }
 856
 857 static void __i915_vma_iounmap(struct i915_vma *vma)
 858 {
 859         GEM_BUG_ON(i915_vma_is_pinned(vma));
 860
 861         if (vma->iomap == NULL)
 862                 return;
 863
 864         io_mapping_unmap(vma->iomap);
 865         vma->iomap = NULL;
 866 }
 867
 868 void i915_vma_revoke_mmap(struct i915_vma *vma)
 869 {
 870         struct drm_vma_offset_node *node = &vma->obj->base.vma_node;
 871         u64 vma_offset;
 872
 873         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 874
 875         if (!i915_vma_has_userfault(vma))
 876                 return;
 877
 878         GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
 879         GEM_BUG_ON(!vma->obj->userfault_count);
 880
 881         vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
 882         unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
 883                             drm_vma_node_offset_addr(node) + vma_offset,
 884                             vma->size,
 885                             1);
 886
 887         i915_vma_unset_userfault(vma);
 888         if (!--vma->obj->userfault_count)
 889                 list_del(&vma->obj->userfault_link);
 890 }
 891
 892 static void export_fence(struct i915_vma *vma,
 893                          struct i915_request *rq,
 894                          unsigned int flags)
 895 {
 896         struct reservation_object *resv = vma->resv;
 897
 898         /*
 899          * Ignore errors from failing to allocate the new fence, we can't
 900          * handle an error right now. Worst case should be missed
 901          * synchronisation leading to rendering corruption.
 902          */
 903         reservation_object_lock(resv, NULL);
 904         if (flags & EXEC_OBJECT_WRITE)
 905                 reservation_object_add_excl_fence(resv, &rq->fence);
 906         else if (reservation_object_reserve_shared(resv, 1) == 0)
 907                 reservation_object_add_shared_fence(resv, &rq->fence);
 908         reservation_object_unlock(resv);
 909 }
 910
 911 int i915_vma_move_to_active(struct i915_vma *vma,
 912                             struct i915_request *rq,
 913                             unsigned int flags)
 914 {
 915         struct drm_i915_gem_object *obj = vma->obj;
 916
 917         lockdep_assert_held(&rq->i915->drm.struct_mutex);
 918         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 919
 920         /*
 921          * Add a reference if we're newly entering the active list.
 922          * The order in which we add operations to the retirement queue is
 923          * vital here: mark_active adds to the start of the callback list,
 924          * such that subsequent callbacks are called first. Therefore we
 925          * add the active reference first and queue for it to be dropped
 926          * *last*.
 927          */
 928         if (!vma->active.count)
 929                 obj->active_count++;
 930
 931         if (unlikely(i915_active_ref(&vma->active, rq->fence.context, rq))) {
 932                 if (!vma->active.count)
 933                         obj->active_count--;
 934                 return -ENOMEM;
 935         }
 936
 937         GEM_BUG_ON(!i915_vma_is_active(vma));
 938         GEM_BUG_ON(!obj->active_count);
 939
 940         obj->write_domain = 0;
 941         if (flags & EXEC_OBJECT_WRITE) {
 942                 obj->write_domain = I915_GEM_DOMAIN_RENDER;
 943
 944                 if (intel_fb_obj_invalidate(obj, ORIGIN_CS))
 945                         __i915_active_request_set(&obj->frontbuffer_write, rq);
 946
 947                 obj->read_domains = 0;
 948         }
 949         obj->read_domains |= I915_GEM_GPU_DOMAINS;
 950
 951         if (flags & EXEC_OBJECT_NEEDS_FENCE)
 952                 __i915_active_request_set(&vma->last_fence, rq);
 953
 954         export_fence(vma, rq, flags);
 955         return 0;
 956 }
 957
 958 int i915_vma_unbind(struct i915_vma *vma)
 959 {
 960         int ret;
 961
 962         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 963
 964         /*
 965          * First wait upon any activity as retiring the request may
 966          * have side-effects such as unpinning or even unbinding this vma.
 967          */
 968         might_sleep();
 969         if (i915_vma_is_active(vma)) {
 970                 /*
 971                  * When a closed VMA is retired, it is unbound - eek.
 972                  * In order to prevent it from being recursively closed,
 973                  * take a pin on the vma so that the second unbind is
 974                  * aborted.
 975                  *
 976                  * Even more scary is that the retire callback may free
 977                  * the object (last active vma). To prevent the explosion
 978                  * we defer the actual object free to a worker that can
 979                  * only proceed once it acquires the struct_mutex (which
 980                  * we currently hold, therefore it cannot free this object
 981                  * before we are finished).
 982                  */
 983                 __i915_vma_pin(vma);
 984
 985                 ret = i915_active_wait(&vma->active);
 986                 if (ret)
 987                         goto unpin;
 988
 989                 ret = i915_active_request_retire(&vma->last_fence,
 990                                               &vma->vm->i915->drm.struct_mutex);
 991 unpin:
 992                 __i915_vma_unpin(vma);
 993                 if (ret)
 994                         return ret;
 995         }
 996         GEM_BUG_ON(i915_vma_is_active(vma));
 997
 998         if (i915_vma_is_pinned(vma)) {
 999                 vma_print_allocator(vma, "is pinned");
1000                 return -EBUSY;
1001         }
1002
1003         if (!drm_mm_node_allocated(&vma->node))
1004                 return 0;
1005
1006         if (i915_vma_is_map_and_fenceable(vma)) {
1007                 /*
1008                  * Check that we have flushed all writes through the GGTT
1009                  * before the unbind, other due to non-strict nature of those
1010                  * indirect writes they may end up referencing the GGTT PTE
1011                  * after the unbind.
1012                  */
1013                 i915_vma_flush_writes(vma);
1014                 GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
1015
1016                 /* release the fence reg _after_ flushing */
1017                 ret = i915_vma_put_fence(vma);
1018                 if (ret)
1019                         return ret;
1020
1021                 /* Force a pagefault for domain tracking on next user access */
1022                 i915_vma_revoke_mmap(vma);
1023
1024                 __i915_vma_iounmap(vma);
1025                 vma->flags &= ~I915_VMA_CAN_FENCE;
1026         }
1027         GEM_BUG_ON(vma->fence);
1028         GEM_BUG_ON(i915_vma_has_userfault(vma));
1029
1030         if (likely(!vma->vm->closed)) {
1031                 trace_i915_vma_unbind(vma);
1032                 vma->ops->unbind_vma(vma);
1033         }
1034         vma->flags &= ~(I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
1035
1036         i915_vma_remove(vma);
1037
1038         return 0;
1039 }
1040
1041 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1042 #include "selftests/i915_vma.c"
1043 #endif