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 "gt/intel_engine.h"
  26
  27 #include "i915_vma.h"
  28
  29 #include "i915_drv.h"
  30 #include "i915_globals.h"
  31 #include "intel_frontbuffer.h"
  32
  33 #include <drm/drm_gem.h>
  34
  35 static struct i915_global_vma {
  36         struct i915_global base;
  37         struct kmem_cache *slab_vmas;
  38 } global;
  39
  40 struct i915_vma *i915_vma_alloc(void)
  41 {
  42         return kmem_cache_zalloc(global.slab_vmas, GFP_KERNEL);
  43 }
  44
  45 void i915_vma_free(struct i915_vma *vma)
  46 {
  47         return kmem_cache_free(global.slab_vmas, vma);
  48 }
  49
  50 #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM)
  51
  52 #include <linux/stackdepot.h>
  53
  54 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
  55 {
  56         unsigned long *entries;
  57         unsigned int nr_entries;
  58         char buf[512];
  59
  60         if (!vma->node.stack) {
  61                 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",
  62                                  vma->node.start, vma->node.size, reason);
  63                 return;
  64         }
  65
  66         nr_entries = stack_depot_fetch(vma->node.stack, &entries);
  67         stack_trace_snprint(buf, sizeof(buf), entries, nr_entries, 0);
  68         DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",
  69                          vma->node.start, vma->node.size, reason, buf);
  70 }
  71
  72 #else
  73
  74 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
  75 {
  76 }
  77
  78 #endif
  79
  80 static void obj_bump_mru(struct drm_i915_gem_object *obj)
  81 {
  82         struct drm_i915_private *i915 = to_i915(obj->base.dev);
  83
  84         spin_lock(&i915->mm.obj_lock);
  85         if (obj->bind_count)
  86                 list_move_tail(&obj->mm.link, &i915->mm.bound_list);
  87         spin_unlock(&i915->mm.obj_lock);
  88
  89         obj->mm.dirty = true; /* be paranoid  */
  90 }
  91
  92 static void __i915_vma_retire(struct i915_active *ref)
  93 {
  94         struct i915_vma *vma = container_of(ref, typeof(*vma), active);
  95         struct drm_i915_gem_object *obj = vma->obj;
  96
  97         GEM_BUG_ON(!i915_gem_object_is_active(obj));
  98         if (--obj->active_count)
  99                 return;
 100
 101         /* Prune the shared fence arrays iff completely idle (inc. external) */
 102         if (reservation_object_trylock(obj->resv)) {
 103                 if (reservation_object_test_signaled_rcu(obj->resv, true))
 104                         reservation_object_add_excl_fence(obj->resv, NULL);
 105                 reservation_object_unlock(obj->resv);
 106         }
 107
 108         /*
 109          * Bump our place on the bound list to keep it roughly in LRU order
 110          * so that we don't steal from recently used but inactive objects
 111          * (unless we are forced to ofc!)
 112          */
 113         obj_bump_mru(obj);
 114
 115         if (i915_gem_object_has_active_reference(obj)) {
 116                 i915_gem_object_clear_active_reference(obj);
 117                 i915_gem_object_put(obj);
 118         }
 119 }
 120
 121 static struct i915_vma *
 122 vma_create(struct drm_i915_gem_object *obj,
 123            struct i915_address_space *vm,
 124            const struct i915_ggtt_view *view)
 125 {
 126         struct i915_vma *vma;
 127         struct rb_node *rb, **p;
 128
 129         /* The aliasing_ppgtt should never be used directly! */
 130         GEM_BUG_ON(vm == &vm->i915->mm.aliasing_ppgtt->vm);
 131
 132         vma = i915_vma_alloc();
 133         if (vma == NULL)
 134                 return ERR_PTR(-ENOMEM);
 135
 136         i915_active_init(vm->i915, &vma->active, __i915_vma_retire);
 137         INIT_ACTIVE_REQUEST(&vma->last_fence);
 138
 139         vma->vm = vm;
 140         vma->ops = &vm->vma_ops;
 141         vma->obj = obj;
 142         vma->resv = obj->resv;
 143         vma->size = obj->base.size;
 144         vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
 145
 146         if (view && view->type != I915_GGTT_VIEW_NORMAL) {
 147                 vma->ggtt_view = *view;
 148                 if (view->type == I915_GGTT_VIEW_PARTIAL) {
 149                         GEM_BUG_ON(range_overflows_t(u64,
 150                                                      view->partial.offset,
 151                                                      view->partial.size,
 152                                                      obj->base.size >> PAGE_SHIFT));
 153                         vma->size = view->partial.size;
 154                         vma->size <<= PAGE_SHIFT;
 155                         GEM_BUG_ON(vma->size > obj->base.size);
 156                 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
 157                         vma->size = intel_rotation_info_size(&view->rotated);
 158                         vma->size <<= PAGE_SHIFT;
 159                 } else if (view->type == I915_GGTT_VIEW_REMAPPED) {
 160                         vma->size = intel_remapped_info_size(&view->remapped);
 161                         vma->size <<= PAGE_SHIFT;
 162                 }
 163         }
 164
 165         if (unlikely(vma->size > vm->total))
 166                 goto err_vma;
 167
 168         GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
 169
 170         if (i915_is_ggtt(vm)) {
 171                 if (unlikely(overflows_type(vma->size, u32)))
 172                         goto err_vma;
 173
 174                 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
 175                                                       i915_gem_object_get_tiling(obj),
 176                                                       i915_gem_object_get_stride(obj));
 177                 if (unlikely(vma->fence_size < vma->size || /* overflow */
 178                              vma->fence_size > vm->total))
 179                         goto err_vma;
 180
 181                 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
 182
 183                 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
 184                                                                 i915_gem_object_get_tiling(obj),
 185                                                                 i915_gem_object_get_stride(obj));
 186                 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
 187
 188                 vma->flags |= I915_VMA_GGTT;
 189         }
 190
 191         spin_lock(&obj->vma.lock);
 192
 193         rb = NULL;
 194         p = &obj->vma.tree.rb_node;
 195         while (*p) {
 196                 struct i915_vma *pos;
 197                 long cmp;
 198
 199                 rb = *p;
 200                 pos = rb_entry(rb, struct i915_vma, obj_node);
 201
 202                 /*
 203                  * If the view already exists in the tree, another thread
 204                  * already created a matching vma, so return the older instance
 205                  * and dispose of ours.
 206                  */
 207                 cmp = i915_vma_compare(pos, vm, view);
 208                 if (cmp == 0) {
 209                         spin_unlock(&obj->vma.lock);
 210                         i915_vma_free(vma);
 211                         return pos;
 212                 }
 213
 214                 if (cmp < 0)
 215                         p = &rb->rb_right;
 216                 else
 217                         p = &rb->rb_left;
 218         }
 219         rb_link_node(&vma->obj_node, rb, p);
 220         rb_insert_color(&vma->obj_node, &obj->vma.tree);
 221
 222         if (i915_vma_is_ggtt(vma))
 223                 /*
 224                  * We put the GGTT vma at the start of the vma-list, followed
 225                  * by the ppGGTT vma. This allows us to break early when
 226                  * iterating over only the GGTT vma for an object, see
 227                  * for_each_ggtt_vma()
 228                  */
 229                 list_add(&vma->obj_link, &obj->vma.list);
 230         else
 231                 list_add_tail(&vma->obj_link, &obj->vma.list);
 232
 233         spin_unlock(&obj->vma.lock);
 234
 235         mutex_lock(&vm->mutex);
 236         list_add(&vma->vm_link, &vm->unbound_list);
 237         mutex_unlock(&vm->mutex);
 238
 239         return vma;
 240
 241 err_vma:
 242         i915_vma_free(vma);
 243         return ERR_PTR(-E2BIG);
 244 }
 245
 246 static struct i915_vma *
 247 vma_lookup(struct drm_i915_gem_object *obj,
 248            struct i915_address_space *vm,
 249            const struct i915_ggtt_view *view)
 250 {
 251         struct rb_node *rb;
 252
 253         rb = obj->vma.tree.rb_node;
 254         while (rb) {
 255                 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
 256                 long cmp;
 257
 258                 cmp = i915_vma_compare(vma, vm, view);
 259                 if (cmp == 0)
 260                         return vma;
 261
 262                 if (cmp < 0)
 263                         rb = rb->rb_right;
 264                 else
 265                         rb = rb->rb_left;
 266         }
 267
 268         return NULL;
 269 }
 270
 271 /**
 272  * i915_vma_instance - return the singleton instance of the VMA
 273  * @obj: parent &struct drm_i915_gem_object to be mapped
 274  * @vm: address space in which the mapping is located
 275  * @view: additional mapping requirements
 276  *
 277  * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
 278  * the same @view characteristics. If a match is not found, one is created.
 279  * Once created, the VMA is kept until either the object is freed, or the
 280  * address space is closed.
 281  *
 282  * Must be called with struct_mutex held.
 283  *
 284  * Returns the vma, or an error pointer.
 285  */
 286 struct i915_vma *
 287 i915_vma_instance(struct drm_i915_gem_object *obj,
 288                   struct i915_address_space *vm,
 289                   const struct i915_ggtt_view *view)
 290 {
 291         struct i915_vma *vma;
 292
 293         GEM_BUG_ON(view && !i915_is_ggtt(vm));
 294         GEM_BUG_ON(vm->closed);
 295
 296         spin_lock(&obj->vma.lock);
 297         vma = vma_lookup(obj, vm, view);
 298         spin_unlock(&obj->vma.lock);
 299
 300         /* vma_create() will resolve the race if another creates the vma */
 301         if (unlikely(!vma))
 302                 vma = vma_create(obj, vm, view);
 303
 304         GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
 305         return vma;
 306 }
 307
 308 /**
 309  * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
 310  * @vma: VMA to map
 311  * @cache_level: mapping cache level
 312  * @flags: flags like global or local mapping
 313  *
 314  * DMA addresses are taken from the scatter-gather table of this object (or of
 315  * this VMA in case of non-default GGTT views) and PTE entries set up.
 316  * Note that DMA addresses are also the only part of the SG table we care about.
 317  */
 318 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
 319                   u32 flags)
 320 {
 321         u32 bind_flags;
 322         u32 vma_flags;
 323         int ret;
 324
 325         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 326         GEM_BUG_ON(vma->size > vma->node.size);
 327
 328         if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
 329                                               vma->node.size,
 330                                               vma->vm->total)))
 331                 return -ENODEV;
 332
 333         if (GEM_DEBUG_WARN_ON(!flags))
 334                 return -EINVAL;
 335
 336         bind_flags = 0;
 337         if (flags & PIN_GLOBAL)
 338                 bind_flags |= I915_VMA_GLOBAL_BIND;
 339         if (flags & PIN_USER)
 340                 bind_flags |= I915_VMA_LOCAL_BIND;
 341
 342         vma_flags = vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
 343         if (flags & PIN_UPDATE)
 344                 bind_flags |= vma_flags;
 345         else
 346                 bind_flags &= ~vma_flags;
 347         if (bind_flags == 0)
 348                 return 0;
 349
 350         GEM_BUG_ON(!vma->pages);
 351
 352         trace_i915_vma_bind(vma, bind_flags);
 353         ret = vma->ops->bind_vma(vma, cache_level, bind_flags);
 354         if (ret)
 355                 return ret;
 356
 357         vma->flags |= bind_flags;
 358         return 0;
 359 }
 360
 361 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
 362 {
 363         void __iomem *ptr;
 364         int err;
 365
 366         /* Access through the GTT requires the device to be awake. */
 367         assert_rpm_wakelock_held(vma->vm->i915);
 368
 369         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 370         if (WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
 371                 err = -ENODEV;
 372                 goto err;
 373         }
 374
 375         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 376         GEM_BUG_ON((vma->flags & I915_VMA_GLOBAL_BIND) == 0);
 377
 378         ptr = vma->iomap;
 379         if (ptr == NULL) {
 380                 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
 381                                         vma->node.start,
 382                                         vma->node.size);
 383                 if (ptr == NULL) {
 384                         err = -ENOMEM;
 385                         goto err;
 386                 }
 387
 388                 vma->iomap = ptr;
 389         }
 390
 391         __i915_vma_pin(vma);
 392
 393         err = i915_vma_pin_fence(vma);
 394         if (err)
 395                 goto err_unpin;
 396
 397         i915_vma_set_ggtt_write(vma);
 398         return ptr;
 399
 400 err_unpin:
 401         __i915_vma_unpin(vma);
 402 err:
 403         return IO_ERR_PTR(err);
 404 }
 405
 406 void i915_vma_flush_writes(struct i915_vma *vma)
 407 {
 408         if (!i915_vma_has_ggtt_write(vma))
 409                 return;
 410
 411         i915_gem_flush_ggtt_writes(vma->vm->i915);
 412
 413         i915_vma_unset_ggtt_write(vma);
 414 }
 415
 416 void i915_vma_unpin_iomap(struct i915_vma *vma)
 417 {
 418         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 419
 420         GEM_BUG_ON(vma->iomap == NULL);
 421
 422         i915_vma_flush_writes(vma);
 423
 424         i915_vma_unpin_fence(vma);
 425         i915_vma_unpin(vma);
 426 }
 427
 428 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
 429 {
 430         struct i915_vma *vma;
 431         struct drm_i915_gem_object *obj;
 432
 433         vma = fetch_and_zero(p_vma);
 434         if (!vma)
 435                 return;
 436
 437         obj = vma->obj;
 438         GEM_BUG_ON(!obj);
 439
 440         i915_vma_unpin(vma);
 441         i915_vma_close(vma);
 442
 443         if (flags & I915_VMA_RELEASE_MAP)
 444                 i915_gem_object_unpin_map(obj);
 445
 446         __i915_gem_object_release_unless_active(obj);
 447 }
 448
 449 bool i915_vma_misplaced(const struct i915_vma *vma,
 450                         u64 size, u64 alignment, u64 flags)
 451 {
 452         if (!drm_mm_node_allocated(&vma->node))
 453                 return false;
 454
 455         if (vma->node.size < size)
 456                 return true;
 457
 458         GEM_BUG_ON(alignment && !is_power_of_2(alignment));
 459         if (alignment && !IS_ALIGNED(vma->node.start, alignment))
 460                 return true;
 461
 462         if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
 463                 return true;
 464
 465         if (flags & PIN_OFFSET_BIAS &&
 466             vma->node.start < (flags & PIN_OFFSET_MASK))
 467                 return true;
 468
 469         if (flags & PIN_OFFSET_FIXED &&
 470             vma->node.start != (flags & PIN_OFFSET_MASK))
 471                 return true;
 472
 473         return false;
 474 }
 475
 476 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
 477 {
 478         bool mappable, fenceable;
 479
 480         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 481         GEM_BUG_ON(!vma->fence_size);
 482
 483         fenceable = (vma->node.size >= vma->fence_size &&
 484                      IS_ALIGNED(vma->node.start, vma->fence_alignment));
 485
 486         mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
 487
 488         if (mappable && fenceable)
 489                 vma->flags |= I915_VMA_CAN_FENCE;
 490         else
 491                 vma->flags &= ~I915_VMA_CAN_FENCE;
 492 }
 493
 494 static bool color_differs(struct drm_mm_node *node, unsigned long color)
 495 {
 496         return node->allocated && node->color != color;
 497 }
 498
 499 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long cache_level)
 500 {
 501         struct drm_mm_node *node = &vma->node;
 502         struct drm_mm_node *other;
 503
 504         /*
 505          * On some machines we have to be careful when putting differing types
 506          * of snoopable memory together to avoid the prefetcher crossing memory
 507          * domains and dying. During vm initialisation, we decide whether or not
 508          * these constraints apply and set the drm_mm.color_adjust
 509          * appropriately.
 510          */
 511         if (vma->vm->mm.color_adjust == NULL)
 512                 return true;
 513
 514         /* Only valid to be called on an already inserted vma */
 515         GEM_BUG_ON(!drm_mm_node_allocated(node));
 516         GEM_BUG_ON(list_empty(&node->node_list));
 517
 518         other = list_prev_entry(node, node_list);
 519         if (color_differs(other, cache_level) && !drm_mm_hole_follows(other))
 520                 return false;
 521
 522         other = list_next_entry(node, node_list);
 523         if (color_differs(other, cache_level) && !drm_mm_hole_follows(node))
 524                 return false;
 525
 526         return true;
 527 }
 528
 529 static void assert_bind_count(const struct drm_i915_gem_object *obj)
 530 {
 531         /*
 532          * Combine the assertion that the object is bound and that we have
 533          * pinned its pages. But we should never have bound the object
 534          * more than we have pinned its pages. (For complete accuracy, we
 535          * assume that no else is pinning the pages, but as a rough assertion
 536          * that we will not run into problems later, this will do!)
 537          */
 538         GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < obj->bind_count);
 539 }
 540
 541 /**
 542  * i915_vma_insert - finds a slot for the vma in its address space
 543  * @vma: the vma
 544  * @size: requested size in bytes (can be larger than the VMA)
 545  * @alignment: required alignment
 546  * @flags: mask of PIN_* flags to use
 547  *
 548  * First we try to allocate some free space that meets the requirements for
 549  * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
 550  * preferrably the oldest idle entry to make room for the new VMA.
 551  *
 552  * Returns:
 553  * 0 on success, negative error code otherwise.
 554  */
 555 static int
 556 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 557 {
 558         struct drm_i915_private *dev_priv = vma->vm->i915;
 559         unsigned int cache_level;
 560         u64 start, end;
 561         int ret;
 562
 563         GEM_BUG_ON(i915_vma_is_closed(vma));
 564         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 565         GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
 566
 567         size = max(size, vma->size);
 568         alignment = max(alignment, vma->display_alignment);
 569         if (flags & PIN_MAPPABLE) {
 570                 size = max_t(typeof(size), size, vma->fence_size);
 571                 alignment = max_t(typeof(alignment),
 572                                   alignment, vma->fence_alignment);
 573         }
 574
 575         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
 576         GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
 577         GEM_BUG_ON(!is_power_of_2(alignment));
 578
 579         start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
 580         GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
 581
 582         end = vma->vm->total;
 583         if (flags & PIN_MAPPABLE)
 584                 end = min_t(u64, end, dev_priv->ggtt.mappable_end);
 585         if (flags & PIN_ZONE_4G)
 586                 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
 587         GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
 588
 589         /* If binding the object/GGTT view requires more space than the entire
 590          * aperture has, reject it early before evicting everything in a vain
 591          * attempt to find space.
 592          */
 593         if (size > end) {
 594                 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
 595                           size, flags & PIN_MAPPABLE ? "mappable" : "total",
 596                           end);
 597                 return -ENOSPC;
 598         }
 599
 600         if (vma->obj) {
 601                 ret = i915_gem_object_pin_pages(vma->obj);
 602                 if (ret)
 603                         return ret;
 604
 605                 cache_level = vma->obj->cache_level;
 606         } else {
 607                 cache_level = 0;
 608         }
 609
 610         GEM_BUG_ON(vma->pages);
 611
 612         ret = vma->ops->set_pages(vma);
 613         if (ret)
 614                 goto err_unpin;
 615
 616         if (flags & PIN_OFFSET_FIXED) {
 617                 u64 offset = flags & PIN_OFFSET_MASK;
 618                 if (!IS_ALIGNED(offset, alignment) ||
 619                     range_overflows(offset, size, end)) {
 620                         ret = -EINVAL;
 621                         goto err_clear;
 622                 }
 623
 624                 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
 625                                            size, offset, cache_level,
 626                                            flags);
 627                 if (ret)
 628                         goto err_clear;
 629         } else {
 630                 /*
 631                  * We only support huge gtt pages through the 48b PPGTT,
 632                  * however we also don't want to force any alignment for
 633                  * objects which need to be tightly packed into the low 32bits.
 634                  *
 635                  * Note that we assume that GGTT are limited to 4GiB for the
 636                  * forseeable future. See also i915_ggtt_offset().
 637                  */
 638                 if (upper_32_bits(end - 1) &&
 639                     vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
 640                         /*
 641                          * We can't mix 64K and 4K PTEs in the same page-table
 642                          * (2M block), and so to avoid the ugliness and
 643                          * complexity of coloring we opt for just aligning 64K
 644                          * objects to 2M.
 645                          */
 646                         u64 page_alignment =
 647                                 rounddown_pow_of_two(vma->page_sizes.sg |
 648                                                      I915_GTT_PAGE_SIZE_2M);
 649
 650                         /*
 651                          * Check we don't expand for the limited Global GTT
 652                          * (mappable aperture is even more precious!). This
 653                          * also checks that we exclude the aliasing-ppgtt.
 654                          */
 655                         GEM_BUG_ON(i915_vma_is_ggtt(vma));
 656
 657                         alignment = max(alignment, page_alignment);
 658
 659                         if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
 660                                 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
 661                 }
 662
 663                 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
 664                                           size, alignment, cache_level,
 665                                           start, end, flags);
 666                 if (ret)
 667                         goto err_clear;
 668
 669                 GEM_BUG_ON(vma->node.start < start);
 670                 GEM_BUG_ON(vma->node.start + vma->node.size > end);
 671         }
 672         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 673         GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, cache_level));
 674
 675         mutex_lock(&vma->vm->mutex);
 676         list_move_tail(&vma->vm_link, &vma->vm->bound_list);
 677         mutex_unlock(&vma->vm->mutex);
 678
 679         if (vma->obj) {
 680                 struct drm_i915_gem_object *obj = vma->obj;
 681
 682                 spin_lock(&dev_priv->mm.obj_lock);
 683                 list_move_tail(&obj->mm.link, &dev_priv->mm.bound_list);
 684                 obj->bind_count++;
 685                 spin_unlock(&dev_priv->mm.obj_lock);
 686
 687                 assert_bind_count(obj);
 688         }
 689
 690         return 0;
 691
 692 err_clear:
 693         vma->ops->clear_pages(vma);
 694 err_unpin:
 695         if (vma->obj)
 696                 i915_gem_object_unpin_pages(vma->obj);
 697         return ret;
 698 }
 699
 700 static void
 701 i915_vma_remove(struct i915_vma *vma)
 702 {
 703         struct drm_i915_private *i915 = vma->vm->i915;
 704
 705         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 706         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 707
 708         vma->ops->clear_pages(vma);
 709
 710         mutex_lock(&vma->vm->mutex);
 711         drm_mm_remove_node(&vma->node);
 712         list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
 713         mutex_unlock(&vma->vm->mutex);
 714
 715         /*
 716          * Since the unbound list is global, only move to that list if
 717          * no more VMAs exist.
 718          */
 719         if (vma->obj) {
 720                 struct drm_i915_gem_object *obj = vma->obj;
 721
 722                 spin_lock(&i915->mm.obj_lock);
 723                 if (--obj->bind_count == 0)
 724                         list_move_tail(&obj->mm.link, &i915->mm.unbound_list);
 725                 spin_unlock(&i915->mm.obj_lock);
 726
 727                 /*
 728                  * And finally now the object is completely decoupled from this
 729                  * vma, we can drop its hold on the backing storage and allow
 730                  * it to be reaped by the shrinker.
 731                  */
 732                 i915_gem_object_unpin_pages(obj);
 733                 assert_bind_count(obj);
 734         }
 735 }
 736
 737 int __i915_vma_do_pin(struct i915_vma *vma,
 738                       u64 size, u64 alignment, u64 flags)
 739 {
 740         const unsigned int bound = vma->flags;
 741         int ret;
 742
 743         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 744         GEM_BUG_ON((flags & (PIN_GLOBAL | PIN_USER)) == 0);
 745         GEM_BUG_ON((flags & PIN_GLOBAL) && !i915_vma_is_ggtt(vma));
 746
 747         if (WARN_ON(bound & I915_VMA_PIN_OVERFLOW)) {
 748                 ret = -EBUSY;
 749                 goto err_unpin;
 750         }
 751
 752         if ((bound & I915_VMA_BIND_MASK) == 0) {
 753                 ret = i915_vma_insert(vma, size, alignment, flags);
 754                 if (ret)
 755                         goto err_unpin;
 756         }
 757         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 758
 759         ret = i915_vma_bind(vma, vma->obj ? vma->obj->cache_level : 0, flags);
 760         if (ret)
 761                 goto err_remove;
 762
 763         GEM_BUG_ON((vma->flags & I915_VMA_BIND_MASK) == 0);
 764
 765         if ((bound ^ vma->flags) & I915_VMA_GLOBAL_BIND)
 766                 __i915_vma_set_map_and_fenceable(vma);
 767
 768         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
 769         return 0;
 770
 771 err_remove:
 772         if ((bound & I915_VMA_BIND_MASK) == 0) {
 773                 i915_vma_remove(vma);
 774                 GEM_BUG_ON(vma->pages);
 775                 GEM_BUG_ON(vma->flags & I915_VMA_BIND_MASK);
 776         }
 777 err_unpin:
 778         __i915_vma_unpin(vma);
 779         return ret;
 780 }
 781
 782 void i915_vma_close(struct i915_vma *vma)
 783 {
 784         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 785
 786         GEM_BUG_ON(i915_vma_is_closed(vma));
 787         vma->flags |= I915_VMA_CLOSED;
 788
 789         /*
 790          * We defer actually closing, unbinding and destroying the VMA until
 791          * the next idle point, or if the object is freed in the meantime. By
 792          * postponing the unbind, we allow for it to be resurrected by the
 793          * client, avoiding the work required to rebind the VMA. This is
 794          * advantageous for DRI, where the client/server pass objects
 795          * between themselves, temporarily opening a local VMA to the
 796          * object, and then closing it again. The same object is then reused
 797          * on the next frame (or two, depending on the depth of the swap queue)
 798          * causing us to rebind the VMA once more. This ends up being a lot
 799          * of wasted work for the steady state.
 800          */
 801         list_add_tail(&vma->closed_link, &vma->vm->i915->gt.closed_vma);
 802 }
 803
 804 void i915_vma_reopen(struct i915_vma *vma)
 805 {
 806         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 807
 808         if (vma->flags & I915_VMA_CLOSED) {
 809                 vma->flags &= ~I915_VMA_CLOSED;
 810                 list_del(&vma->closed_link);
 811         }
 812 }
 813
 814 static void __i915_vma_destroy(struct i915_vma *vma)
 815 {
 816         GEM_BUG_ON(vma->node.allocated);
 817         GEM_BUG_ON(vma->fence);
 818
 819         GEM_BUG_ON(i915_active_request_isset(&vma->last_fence));
 820
 821         mutex_lock(&vma->vm->mutex);
 822         list_del(&vma->vm_link);
 823         mutex_unlock(&vma->vm->mutex);
 824
 825         if (vma->obj) {
 826                 struct drm_i915_gem_object *obj = vma->obj;
 827
 828                 spin_lock(&obj->vma.lock);
 829                 list_del(&vma->obj_link);
 830                 rb_erase(&vma->obj_node, &vma->obj->vma.tree);
 831                 spin_unlock(&obj->vma.lock);
 832         }
 833
 834         i915_active_fini(&vma->active);
 835
 836         i915_vma_free(vma);
 837 }
 838
 839 void i915_vma_destroy(struct i915_vma *vma)
 840 {
 841         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 842
 843         GEM_BUG_ON(i915_vma_is_active(vma));
 844         GEM_BUG_ON(i915_vma_is_pinned(vma));
 845
 846         if (i915_vma_is_closed(vma))
 847                 list_del(&vma->closed_link);
 848
 849         WARN_ON(i915_vma_unbind(vma));
 850         __i915_vma_destroy(vma);
 851 }
 852
 853 void i915_vma_parked(struct drm_i915_private *i915)
 854 {
 855         struct i915_vma *vma, *next;
 856
 857         list_for_each_entry_safe(vma, next, &i915->gt.closed_vma, closed_link) {
 858                 GEM_BUG_ON(!i915_vma_is_closed(vma));
 859                 i915_vma_destroy(vma);
 860         }
 861
 862         GEM_BUG_ON(!list_empty(&i915->gt.closed_vma));
 863 }
 864
 865 static void __i915_vma_iounmap(struct i915_vma *vma)
 866 {
 867         GEM_BUG_ON(i915_vma_is_pinned(vma));
 868
 869         if (vma->iomap == NULL)
 870                 return;
 871
 872         io_mapping_unmap(vma->iomap);
 873         vma->iomap = NULL;
 874 }
 875
 876 void i915_vma_revoke_mmap(struct i915_vma *vma)
 877 {
 878         struct drm_vma_offset_node *node = &vma->obj->base.vma_node;
 879         u64 vma_offset;
 880
 881         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 882
 883         if (!i915_vma_has_userfault(vma))
 884                 return;
 885
 886         GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
 887         GEM_BUG_ON(!vma->obj->userfault_count);
 888
 889         vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
 890         unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
 891                             drm_vma_node_offset_addr(node) + vma_offset,
 892                             vma->size,
 893                             1);
 894
 895         i915_vma_unset_userfault(vma);
 896         if (!--vma->obj->userfault_count)
 897                 list_del(&vma->obj->userfault_link);
 898 }
 899
 900 static void export_fence(struct i915_vma *vma,
 901                          struct i915_request *rq,
 902                          unsigned int flags)
 903 {
 904         struct reservation_object *resv = vma->resv;
 905
 906         /*
 907          * Ignore errors from failing to allocate the new fence, we can't
 908          * handle an error right now. Worst case should be missed
 909          * synchronisation leading to rendering corruption.
 910          */
 911         reservation_object_lock(resv, NULL);
 912         if (flags & EXEC_OBJECT_WRITE)
 913                 reservation_object_add_excl_fence(resv, &rq->fence);
 914         else if (reservation_object_reserve_shared(resv, 1) == 0)
 915                 reservation_object_add_shared_fence(resv, &rq->fence);
 916         reservation_object_unlock(resv);
 917 }
 918
 919 int i915_vma_move_to_active(struct i915_vma *vma,
 920                             struct i915_request *rq,
 921                             unsigned int flags)
 922 {
 923         struct drm_i915_gem_object *obj = vma->obj;
 924
 925         lockdep_assert_held(&rq->i915->drm.struct_mutex);
 926         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 927
 928         /*
 929          * Add a reference if we're newly entering the active list.
 930          * The order in which we add operations to the retirement queue is
 931          * vital here: mark_active adds to the start of the callback list,
 932          * such that subsequent callbacks are called first. Therefore we
 933          * add the active reference first and queue for it to be dropped
 934          * *last*.
 935          */
 936         if (!vma->active.count)
 937                 obj->active_count++;
 938
 939         if (unlikely(i915_active_ref(&vma->active, rq->fence.context, rq))) {
 940                 if (!vma->active.count)
 941                         obj->active_count--;
 942                 return -ENOMEM;
 943         }
 944
 945         GEM_BUG_ON(!i915_vma_is_active(vma));
 946         GEM_BUG_ON(!obj->active_count);
 947
 948         obj->write_domain = 0;
 949         if (flags & EXEC_OBJECT_WRITE) {
 950                 obj->write_domain = I915_GEM_DOMAIN_RENDER;
 951
 952                 if (intel_fb_obj_invalidate(obj, ORIGIN_CS))
 953                         __i915_active_request_set(&obj->frontbuffer_write, rq);
 954
 955                 obj->read_domains = 0;
 956         }
 957         obj->read_domains |= I915_GEM_GPU_DOMAINS;
 958
 959         if (flags & EXEC_OBJECT_NEEDS_FENCE)
 960                 __i915_active_request_set(&vma->last_fence, rq);
 961
 962         export_fence(vma, rq, flags);
 963         return 0;
 964 }
 965
 966 int i915_vma_unbind(struct i915_vma *vma)
 967 {
 968         int ret;
 969
 970         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 971
 972         /*
 973          * First wait upon any activity as retiring the request may
 974          * have side-effects such as unpinning or even unbinding this vma.
 975          */
 976         might_sleep();
 977         if (i915_vma_is_active(vma)) {
 978                 /*
 979                  * When a closed VMA is retired, it is unbound - eek.
 980                  * In order to prevent it from being recursively closed,
 981                  * take a pin on the vma so that the second unbind is
 982                  * aborted.
 983                  *
 984                  * Even more scary is that the retire callback may free
 985                  * the object (last active vma). To prevent the explosion
 986                  * we defer the actual object free to a worker that can
 987                  * only proceed once it acquires the struct_mutex (which
 988                  * we currently hold, therefore it cannot free this object
 989                  * before we are finished).
 990                  */
 991                 __i915_vma_pin(vma);
 992
 993                 ret = i915_active_wait(&vma->active);
 994                 if (ret)
 995                         goto unpin;
 996
 997                 ret = i915_active_request_retire(&vma->last_fence,
 998                                               &vma->vm->i915->drm.struct_mutex);
 999 unpin:
1000                 __i915_vma_unpin(vma);
1001                 if (ret)
1002                         return ret;
1003         }
1004         GEM_BUG_ON(i915_vma_is_active(vma));
1005
1006         if (i915_vma_is_pinned(vma)) {
1007                 vma_print_allocator(vma, "is pinned");
1008                 return -EBUSY;
1009         }
1010
1011         if (!drm_mm_node_allocated(&vma->node))
1012                 return 0;
1013
1014         if (i915_vma_is_map_and_fenceable(vma)) {
1015                 /*
1016                  * Check that we have flushed all writes through the GGTT
1017                  * before the unbind, other due to non-strict nature of those
1018                  * indirect writes they may end up referencing the GGTT PTE
1019                  * after the unbind.
1020                  */
1021                 i915_vma_flush_writes(vma);
1022                 GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
1023
1024                 /* release the fence reg _after_ flushing */
1025                 ret = i915_vma_put_fence(vma);
1026                 if (ret)
1027                         return ret;
1028
1029                 /* Force a pagefault for domain tracking on next user access */
1030                 i915_vma_revoke_mmap(vma);
1031
1032                 __i915_vma_iounmap(vma);
1033                 vma->flags &= ~I915_VMA_CAN_FENCE;
1034         }
1035         GEM_BUG_ON(vma->fence);
1036         GEM_BUG_ON(i915_vma_has_userfault(vma));
1037
1038         if (likely(!vma->vm->closed)) {
1039                 trace_i915_vma_unbind(vma);
1040                 vma->ops->unbind_vma(vma);
1041         }
1042         vma->flags &= ~(I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
1043
1044         i915_vma_remove(vma);
1045
1046         return 0;
1047 }
1048
1049 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1050 #include "selftests/i915_vma.c"
1051 #endif
1052
1053 static void i915_global_vma_shrink(void)
1054 {
1055         kmem_cache_shrink(global.slab_vmas);
1056 }
1057
1058 static void i915_global_vma_exit(void)
1059 {
1060         kmem_cache_destroy(global.slab_vmas);
1061 }
1062
1063 static struct i915_global_vma global = { {
1064         .shrink = i915_global_vma_shrink,
1065         .exit = i915_global_vma_exit,
1066 } };
1067
1068 int __init i915_global_vma_init(void)
1069 {
1070         global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1071         if (!global.slab_vmas)
1072                 return -ENOMEM;
1073
1074         i915_global_register(&global.base);
1075         return 0;
1076 }