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