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         unsigned long flags;
  84
  85         spin_lock_irqsave(&i915->mm.obj_lock, flags);
  86
  87         if (obj->bind_count)
  88                 list_move_tail(&obj->mm.link, &i915->mm.bound_list);
  89
  90         spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
  91
  92         obj->mm.dirty = true; /* be paranoid  */
  93 }
  94
  95 static void __i915_vma_retire(struct i915_active *ref)
  96 {
  97         struct i915_vma *vma = container_of(ref, typeof(*vma), active);
  98         struct drm_i915_gem_object *obj = vma->obj;
  99
 100         GEM_BUG_ON(!i915_gem_object_is_active(obj));
 101         if (--obj->active_count)
 102                 return;
 103
 104         /* Prune the shared fence arrays iff completely idle (inc. external) */
 105         if (reservation_object_trylock(obj->resv)) {
 106                 if (reservation_object_test_signaled_rcu(obj->resv, true))
 107                         reservation_object_add_excl_fence(obj->resv, NULL);
 108                 reservation_object_unlock(obj->resv);
 109         }
 110
 111         /*
 112          * Bump our place on the bound list to keep it roughly in LRU order
 113          * so that we don't steal from recently used but inactive objects
 114          * (unless we are forced to ofc!)
 115          */
 116         if (i915_gem_object_is_shrinkable(obj))
 117                 obj_bump_mru(obj);
 118
 119         i915_gem_object_put(obj); /* and drop the active reference */
 120 }
 121
 122 static struct i915_vma *
 123 vma_create(struct drm_i915_gem_object *obj,
 124            struct i915_address_space *vm,
 125            const struct i915_ggtt_view *view)
 126 {
 127         struct i915_vma *vma;
 128         struct rb_node *rb, **p;
 129
 130         /* The aliasing_ppgtt should never be used directly! */
 131         GEM_BUG_ON(vm == &vm->i915->mm.aliasing_ppgtt->vm);
 132
 133         vma = i915_vma_alloc();
 134         if (vma == NULL)
 135                 return ERR_PTR(-ENOMEM);
 136
 137         vma->vm = vm;
 138         vma->ops = &vm->vma_ops;
 139         vma->obj = obj;
 140         vma->resv = obj->resv;
 141         vma->size = obj->base.size;
 142         vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
 143
 144         i915_active_init(vm->i915, &vma->active, __i915_vma_retire);
 145         INIT_ACTIVE_REQUEST(&vma->last_fence);
 146
 147         INIT_LIST_HEAD(&vma->closed_link);
 148
 149         if (view && view->type != I915_GGTT_VIEW_NORMAL) {
 150                 vma->ggtt_view = *view;
 151                 if (view->type == I915_GGTT_VIEW_PARTIAL) {
 152                         GEM_BUG_ON(range_overflows_t(u64,
 153                                                      view->partial.offset,
 154                                                      view->partial.size,
 155                                                      obj->base.size >> PAGE_SHIFT));
 156                         vma->size = view->partial.size;
 157                         vma->size <<= PAGE_SHIFT;
 158                         GEM_BUG_ON(vma->size > obj->base.size);
 159                 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
 160                         vma->size = intel_rotation_info_size(&view->rotated);
 161                         vma->size <<= PAGE_SHIFT;
 162                 } else if (view->type == I915_GGTT_VIEW_REMAPPED) {
 163                         vma->size = intel_remapped_info_size(&view->remapped);
 164                         vma->size <<= PAGE_SHIFT;
 165                 }
 166         }
 167
 168         if (unlikely(vma->size > vm->total))
 169                 goto err_vma;
 170
 171         GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
 172
 173         if (i915_is_ggtt(vm)) {
 174                 if (unlikely(overflows_type(vma->size, u32)))
 175                         goto err_vma;
 176
 177                 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
 178                                                       i915_gem_object_get_tiling(obj),
 179                                                       i915_gem_object_get_stride(obj));
 180                 if (unlikely(vma->fence_size < vma->size || /* overflow */
 181                              vma->fence_size > vm->total))
 182                         goto err_vma;
 183
 184                 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
 185
 186                 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
 187                                                                 i915_gem_object_get_tiling(obj),
 188                                                                 i915_gem_object_get_stride(obj));
 189                 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
 190
 191                 vma->flags |= I915_VMA_GGTT;
 192         }
 193
 194         spin_lock(&obj->vma.lock);
 195
 196         rb = NULL;
 197         p = &obj->vma.tree.rb_node;
 198         while (*p) {
 199                 struct i915_vma *pos;
 200                 long cmp;
 201
 202                 rb = *p;
 203                 pos = rb_entry(rb, struct i915_vma, obj_node);
 204
 205                 /*
 206                  * If the view already exists in the tree, another thread
 207                  * already created a matching vma, so return the older instance
 208                  * and dispose of ours.
 209                  */
 210                 cmp = i915_vma_compare(pos, vm, view);
 211                 if (cmp == 0) {
 212                         spin_unlock(&obj->vma.lock);
 213                         i915_vma_free(vma);
 214                         return pos;
 215                 }
 216
 217                 if (cmp < 0)
 218                         p = &rb->rb_right;
 219                 else
 220                         p = &rb->rb_left;
 221         }
 222         rb_link_node(&vma->obj_node, rb, p);
 223         rb_insert_color(&vma->obj_node, &obj->vma.tree);
 224
 225         if (i915_vma_is_ggtt(vma))
 226                 /*
 227                  * We put the GGTT vma at the start of the vma-list, followed
 228                  * by the ppGGTT vma. This allows us to break early when
 229                  * iterating over only the GGTT vma for an object, see
 230                  * for_each_ggtt_vma()
 231                  */
 232                 list_add(&vma->obj_link, &obj->vma.list);
 233         else
 234                 list_add_tail(&vma->obj_link, &obj->vma.list);
 235
 236         spin_unlock(&obj->vma.lock);
 237
 238         mutex_lock(&vm->mutex);
 239         list_add(&vma->vm_link, &vm->unbound_list);
 240         mutex_unlock(&vm->mutex);
 241
 242         return vma;
 243
 244 err_vma:
 245         i915_vma_free(vma);
 246         return ERR_PTR(-E2BIG);
 247 }
 248
 249 static struct i915_vma *
 250 vma_lookup(struct drm_i915_gem_object *obj,
 251            struct i915_address_space *vm,
 252            const struct i915_ggtt_view *view)
 253 {
 254         struct rb_node *rb;
 255
 256         rb = obj->vma.tree.rb_node;
 257         while (rb) {
 258                 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
 259                 long cmp;
 260
 261                 cmp = i915_vma_compare(vma, vm, view);
 262                 if (cmp == 0)
 263                         return vma;
 264
 265                 if (cmp < 0)
 266                         rb = rb->rb_right;
 267                 else
 268                         rb = rb->rb_left;
 269         }
 270
 271         return NULL;
 272 }
 273
 274 /**
 275  * i915_vma_instance - return the singleton instance of the VMA
 276  * @obj: parent &struct drm_i915_gem_object to be mapped
 277  * @vm: address space in which the mapping is located
 278  * @view: additional mapping requirements
 279  *
 280  * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
 281  * the same @view characteristics. If a match is not found, one is created.
 282  * Once created, the VMA is kept until either the object is freed, or the
 283  * address space is closed.
 284  *
 285  * Must be called with struct_mutex held.
 286  *
 287  * Returns the vma, or an error pointer.
 288  */
 289 struct i915_vma *
 290 i915_vma_instance(struct drm_i915_gem_object *obj,
 291                   struct i915_address_space *vm,
 292                   const struct i915_ggtt_view *view)
 293 {
 294         struct i915_vma *vma;
 295
 296         GEM_BUG_ON(view && !i915_is_ggtt(vm));
 297         GEM_BUG_ON(vm->closed);
 298
 299         spin_lock(&obj->vma.lock);
 300         vma = vma_lookup(obj, vm, view);
 301         spin_unlock(&obj->vma.lock);
 302
 303         /* vma_create() will resolve the race if another creates the vma */
 304         if (unlikely(!vma))
 305                 vma = vma_create(obj, vm, view);
 306
 307         GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
 308         return vma;
 309 }
 310
 311 /**
 312  * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
 313  * @vma: VMA to map
 314  * @cache_level: mapping cache level
 315  * @flags: flags like global or local mapping
 316  *
 317  * DMA addresses are taken from the scatter-gather table of this object (or of
 318  * this VMA in case of non-default GGTT views) and PTE entries set up.
 319  * Note that DMA addresses are also the only part of the SG table we care about.
 320  */
 321 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
 322                   u32 flags)
 323 {
 324         u32 bind_flags;
 325         u32 vma_flags;
 326         int ret;
 327
 328         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 329         GEM_BUG_ON(vma->size > vma->node.size);
 330
 331         if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
 332                                               vma->node.size,
 333                                               vma->vm->total)))
 334                 return -ENODEV;
 335
 336         if (GEM_DEBUG_WARN_ON(!flags))
 337                 return -EINVAL;
 338
 339         bind_flags = 0;
 340         if (flags & PIN_GLOBAL)
 341                 bind_flags |= I915_VMA_GLOBAL_BIND;
 342         if (flags & PIN_USER)
 343                 bind_flags |= I915_VMA_LOCAL_BIND;
 344
 345         vma_flags = vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
 346         if (flags & PIN_UPDATE)
 347                 bind_flags |= vma_flags;
 348         else
 349                 bind_flags &= ~vma_flags;
 350         if (bind_flags == 0)
 351                 return 0;
 352
 353         GEM_BUG_ON(!vma->pages);
 354
 355         trace_i915_vma_bind(vma, bind_flags);
 356         ret = vma->ops->bind_vma(vma, cache_level, bind_flags);
 357         if (ret)
 358                 return ret;
 359
 360         vma->flags |= bind_flags;
 361         return 0;
 362 }
 363
 364 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
 365 {
 366         void __iomem *ptr;
 367         int err;
 368
 369         /* Access through the GTT requires the device to be awake. */
 370         assert_rpm_wakelock_held(vma->vm->i915);
 371
 372         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 373         if (WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
 374                 err = -ENODEV;
 375                 goto err;
 376         }
 377
 378         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 379         GEM_BUG_ON((vma->flags & I915_VMA_GLOBAL_BIND) == 0);
 380
 381         ptr = vma->iomap;
 382         if (ptr == NULL) {
 383                 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
 384                                         vma->node.start,
 385                                         vma->node.size);
 386                 if (ptr == NULL) {
 387                         err = -ENOMEM;
 388                         goto err;
 389                 }
 390
 391                 vma->iomap = ptr;
 392         }
 393
 394         __i915_vma_pin(vma);
 395
 396         err = i915_vma_pin_fence(vma);
 397         if (err)
 398                 goto err_unpin;
 399
 400         i915_vma_set_ggtt_write(vma);
 401         return ptr;
 402
 403 err_unpin:
 404         __i915_vma_unpin(vma);
 405 err:
 406         return IO_ERR_PTR(err);
 407 }
 408
 409 void i915_vma_flush_writes(struct i915_vma *vma)
 410 {
 411         if (!i915_vma_has_ggtt_write(vma))
 412                 return;
 413
 414         i915_gem_flush_ggtt_writes(vma->vm->i915);
 415
 416         i915_vma_unset_ggtt_write(vma);
 417 }
 418
 419 void i915_vma_unpin_iomap(struct i915_vma *vma)
 420 {
 421         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 422
 423         GEM_BUG_ON(vma->iomap == NULL);
 424
 425         i915_vma_flush_writes(vma);
 426
 427         i915_vma_unpin_fence(vma);
 428         i915_vma_unpin(vma);
 429 }
 430
 431 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
 432 {
 433         struct i915_vma *vma;
 434         struct drm_i915_gem_object *obj;
 435
 436         vma = fetch_and_zero(p_vma);
 437         if (!vma)
 438                 return;
 439
 440         obj = vma->obj;
 441         GEM_BUG_ON(!obj);
 442
 443         i915_vma_unpin(vma);
 444         i915_vma_close(vma);
 445
 446         if (flags & I915_VMA_RELEASE_MAP)
 447                 i915_gem_object_unpin_map(obj);
 448
 449         i915_gem_object_put(obj);
 450 }
 451
 452 bool i915_vma_misplaced(const struct i915_vma *vma,
 453                         u64 size, u64 alignment, u64 flags)
 454 {
 455         if (!drm_mm_node_allocated(&vma->node))
 456                 return false;
 457
 458         if (vma->node.size < size)
 459                 return true;
 460
 461         GEM_BUG_ON(alignment && !is_power_of_2(alignment));
 462         if (alignment && !IS_ALIGNED(vma->node.start, alignment))
 463                 return true;
 464
 465         if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
 466                 return true;
 467
 468         if (flags & PIN_OFFSET_BIAS &&
 469             vma->node.start < (flags & PIN_OFFSET_MASK))
 470                 return true;
 471
 472         if (flags & PIN_OFFSET_FIXED &&
 473             vma->node.start != (flags & PIN_OFFSET_MASK))
 474                 return true;
 475
 476         return false;
 477 }
 478
 479 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
 480 {
 481         bool mappable, fenceable;
 482
 483         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 484         GEM_BUG_ON(!vma->fence_size);
 485
 486         fenceable = (vma->node.size >= vma->fence_size &&
 487                      IS_ALIGNED(vma->node.start, vma->fence_alignment));
 488
 489         mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
 490
 491         if (mappable && fenceable)
 492                 vma->flags |= I915_VMA_CAN_FENCE;
 493         else
 494                 vma->flags &= ~I915_VMA_CAN_FENCE;
 495 }
 496
 497 static bool color_differs(struct drm_mm_node *node, unsigned long color)
 498 {
 499         return node->allocated && node->color != color;
 500 }
 501
 502 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long cache_level)
 503 {
 504         struct drm_mm_node *node = &vma->node;
 505         struct drm_mm_node *other;
 506
 507         /*
 508          * On some machines we have to be careful when putting differing types
 509          * of snoopable memory together to avoid the prefetcher crossing memory
 510          * domains and dying. During vm initialisation, we decide whether or not
 511          * these constraints apply and set the drm_mm.color_adjust
 512          * appropriately.
 513          */
 514         if (vma->vm->mm.color_adjust == NULL)
 515                 return true;
 516
 517         /* Only valid to be called on an already inserted vma */
 518         GEM_BUG_ON(!drm_mm_node_allocated(node));
 519         GEM_BUG_ON(list_empty(&node->node_list));
 520
 521         other = list_prev_entry(node, node_list);
 522         if (color_differs(other, cache_level) && !drm_mm_hole_follows(other))
 523                 return false;
 524
 525         other = list_next_entry(node, node_list);
 526         if (color_differs(other, cache_level) && !drm_mm_hole_follows(node))
 527                 return false;
 528
 529         return true;
 530 }
 531
 532 static void assert_bind_count(const struct drm_i915_gem_object *obj)
 533 {
 534         /*
 535          * Combine the assertion that the object is bound and that we have
 536          * pinned its pages. But we should never have bound the object
 537          * more than we have pinned its pages. (For complete accuracy, we
 538          * assume that no else is pinning the pages, but as a rough assertion
 539          * that we will not run into problems later, this will do!)
 540          */
 541         GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < obj->bind_count);
 542 }
 543
 544 /**
 545  * i915_vma_insert - finds a slot for the vma in its address space
 546  * @vma: the vma
 547  * @size: requested size in bytes (can be larger than the VMA)
 548  * @alignment: required alignment
 549  * @flags: mask of PIN_* flags to use
 550  *
 551  * First we try to allocate some free space that meets the requirements for
 552  * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
 553  * preferrably the oldest idle entry to make room for the new VMA.
 554  *
 555  * Returns:
 556  * 0 on success, negative error code otherwise.
 557  */
 558 static int
 559 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 560 {
 561         struct drm_i915_private *dev_priv = vma->vm->i915;
 562         unsigned int cache_level;
 563         u64 start, end;
 564         int ret;
 565
 566         GEM_BUG_ON(i915_vma_is_closed(vma));
 567         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 568         GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
 569
 570         size = max(size, vma->size);
 571         alignment = max(alignment, vma->display_alignment);
 572         if (flags & PIN_MAPPABLE) {
 573                 size = max_t(typeof(size), size, vma->fence_size);
 574                 alignment = max_t(typeof(alignment),
 575                                   alignment, vma->fence_alignment);
 576         }
 577
 578         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
 579         GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
 580         GEM_BUG_ON(!is_power_of_2(alignment));
 581
 582         start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
 583         GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
 584
 585         end = vma->vm->total;
 586         if (flags & PIN_MAPPABLE)
 587                 end = min_t(u64, end, dev_priv->ggtt.mappable_end);
 588         if (flags & PIN_ZONE_4G)
 589                 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
 590         GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
 591
 592         /* If binding the object/GGTT view requires more space than the entire
 593          * aperture has, reject it early before evicting everything in a vain
 594          * attempt to find space.
 595          */
 596         if (size > end) {
 597                 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
 598                           size, flags & PIN_MAPPABLE ? "mappable" : "total",
 599                           end);
 600                 return -ENOSPC;
 601         }
 602
 603         if (vma->obj) {
 604                 ret = i915_gem_object_pin_pages(vma->obj);
 605                 if (ret)
 606                         return ret;
 607
 608                 cache_level = vma->obj->cache_level;
 609         } else {
 610                 cache_level = 0;
 611         }
 612
 613         GEM_BUG_ON(vma->pages);
 614
 615         ret = vma->ops->set_pages(vma);
 616         if (ret)
 617                 goto err_unpin;
 618
 619         if (flags & PIN_OFFSET_FIXED) {
 620                 u64 offset = flags & PIN_OFFSET_MASK;
 621                 if (!IS_ALIGNED(offset, alignment) ||
 622                     range_overflows(offset, size, end)) {
 623                         ret = -EINVAL;
 624                         goto err_clear;
 625                 }
 626
 627                 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
 628                                            size, offset, cache_level,
 629                                            flags);
 630                 if (ret)
 631                         goto err_clear;
 632         } else {
 633                 /*
 634                  * We only support huge gtt pages through the 48b PPGTT,
 635                  * however we also don't want to force any alignment for
 636                  * objects which need to be tightly packed into the low 32bits.
 637                  *
 638                  * Note that we assume that GGTT are limited to 4GiB for the
 639                  * forseeable future. See also i915_ggtt_offset().
 640                  */
 641                 if (upper_32_bits(end - 1) &&
 642                     vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
 643                         /*
 644                          * We can't mix 64K and 4K PTEs in the same page-table
 645                          * (2M block), and so to avoid the ugliness and
 646                          * complexity of coloring we opt for just aligning 64K
 647                          * objects to 2M.
 648                          */
 649                         u64 page_alignment =
 650                                 rounddown_pow_of_two(vma->page_sizes.sg |
 651                                                      I915_GTT_PAGE_SIZE_2M);
 652
 653                         /*
 654                          * Check we don't expand for the limited Global GTT
 655                          * (mappable aperture is even more precious!). This
 656                          * also checks that we exclude the aliasing-ppgtt.
 657                          */
 658                         GEM_BUG_ON(i915_vma_is_ggtt(vma));
 659
 660                         alignment = max(alignment, page_alignment);
 661
 662                         if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
 663                                 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
 664                 }
 665
 666                 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
 667                                           size, alignment, cache_level,
 668                                           start, end, flags);
 669                 if (ret)
 670                         goto err_clear;
 671
 672                 GEM_BUG_ON(vma->node.start < start);
 673                 GEM_BUG_ON(vma->node.start + vma->node.size > end);
 674         }
 675         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 676         GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, cache_level));
 677
 678         mutex_lock(&vma->vm->mutex);
 679         list_move_tail(&vma->vm_link, &vma->vm->bound_list);
 680         mutex_unlock(&vma->vm->mutex);
 681
 682         if (vma->obj) {
 683                 struct drm_i915_gem_object *obj = vma->obj;
 684                 unsigned long flags;
 685
 686                 spin_lock_irqsave(&dev_priv->mm.obj_lock, flags);
 687
 688                 if (i915_gem_object_is_shrinkable(obj))
 689                         list_move_tail(&obj->mm.link, &dev_priv->mm.bound_list);
 690
 691                 obj->bind_count++;
 692                 assert_bind_count(obj);
 693
 694                 spin_unlock_irqrestore(&dev_priv->mm.obj_lock, flags);
 695         }
 696
 697         return 0;
 698
 699 err_clear:
 700         vma->ops->clear_pages(vma);
 701 err_unpin:
 702         if (vma->obj)
 703                 i915_gem_object_unpin_pages(vma->obj);
 704         return ret;
 705 }
 706
 707 static void
 708 i915_vma_remove(struct i915_vma *vma)
 709 {
 710         struct drm_i915_private *i915 = vma->vm->i915;
 711
 712         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 713         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 714
 715         vma->ops->clear_pages(vma);
 716
 717         mutex_lock(&vma->vm->mutex);
 718         drm_mm_remove_node(&vma->node);
 719         list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
 720         mutex_unlock(&vma->vm->mutex);
 721
 722         /*
 723          * Since the unbound list is global, only move to that list if
 724          * no more VMAs exist.
 725          */
 726         if (vma->obj) {
 727                 struct drm_i915_gem_object *obj = vma->obj;
 728                 unsigned long flags;
 729
 730                 spin_lock_irqsave(&i915->mm.obj_lock, flags);
 731
 732                 GEM_BUG_ON(obj->bind_count == 0);
 733                 if (--obj->bind_count == 0 &&
 734                     i915_gem_object_is_shrinkable(obj) &&
 735                     obj->mm.madv == I915_MADV_WILLNEED)
 736                         list_move_tail(&obj->mm.link, &i915->mm.unbound_list);
 737
 738                 spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
 739
 740                 /*
 741                  * And finally now the object is completely decoupled from this
 742                  * vma, we can drop its hold on the backing storage and allow
 743                  * it to be reaped by the shrinker.
 744                  */
 745                 i915_gem_object_unpin_pages(obj);
 746                 assert_bind_count(obj);
 747         }
 748 }
 749
 750 int __i915_vma_do_pin(struct i915_vma *vma,
 751                       u64 size, u64 alignment, u64 flags)
 752 {
 753         const unsigned int bound = vma->flags;
 754         int ret;
 755
 756         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 757         GEM_BUG_ON((flags & (PIN_GLOBAL | PIN_USER)) == 0);
 758         GEM_BUG_ON((flags & PIN_GLOBAL) && !i915_vma_is_ggtt(vma));
 759
 760         if (WARN_ON(bound & I915_VMA_PIN_OVERFLOW)) {
 761                 ret = -EBUSY;
 762                 goto err_unpin;
 763         }
 764
 765         if ((bound & I915_VMA_BIND_MASK) == 0) {
 766                 ret = i915_vma_insert(vma, size, alignment, flags);
 767                 if (ret)
 768                         goto err_unpin;
 769         }
 770         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 771
 772         ret = i915_vma_bind(vma, vma->obj ? vma->obj->cache_level : 0, flags);
 773         if (ret)
 774                 goto err_remove;
 775
 776         GEM_BUG_ON((vma->flags & I915_VMA_BIND_MASK) == 0);
 777
 778         if ((bound ^ vma->flags) & I915_VMA_GLOBAL_BIND)
 779                 __i915_vma_set_map_and_fenceable(vma);
 780
 781         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
 782         return 0;
 783
 784 err_remove:
 785         if ((bound & I915_VMA_BIND_MASK) == 0) {
 786                 i915_vma_remove(vma);
 787                 GEM_BUG_ON(vma->pages);
 788                 GEM_BUG_ON(vma->flags & I915_VMA_BIND_MASK);
 789         }
 790 err_unpin:
 791         __i915_vma_unpin(vma);
 792         return ret;
 793 }
 794
 795 void i915_vma_close(struct i915_vma *vma)
 796 {
 797         struct drm_i915_private *i915 = vma->vm->i915;
 798         unsigned long flags;
 799
 800         GEM_BUG_ON(i915_vma_is_closed(vma));
 801
 802         /*
 803          * We defer actually closing, unbinding and destroying the VMA until
 804          * the next idle point, or if the object is freed in the meantime. By
 805          * postponing the unbind, we allow for it to be resurrected by the
 806          * client, avoiding the work required to rebind the VMA. This is
 807          * advantageous for DRI, where the client/server pass objects
 808          * between themselves, temporarily opening a local VMA to the
 809          * object, and then closing it again. The same object is then reused
 810          * on the next frame (or two, depending on the depth of the swap queue)
 811          * causing us to rebind the VMA once more. This ends up being a lot
 812          * of wasted work for the steady state.
 813          */
 814         spin_lock_irqsave(&i915->gt.closed_lock, flags);
 815         list_add(&vma->closed_link, &i915->gt.closed_vma);
 816         spin_unlock_irqrestore(&i915->gt.closed_lock, flags);
 817 }
 818
 819 static void __i915_vma_remove_closed(struct i915_vma *vma)
 820 {
 821         struct drm_i915_private *i915 = vma->vm->i915;
 822
 823         if (!i915_vma_is_closed(vma))
 824                 return;
 825
 826         spin_lock_irq(&i915->gt.closed_lock);
 827         list_del_init(&vma->closed_link);
 828         spin_unlock_irq(&i915->gt.closed_lock);
 829 }
 830
 831 void i915_vma_reopen(struct i915_vma *vma)
 832 {
 833         __i915_vma_remove_closed(vma);
 834 }
 835
 836 static void __i915_vma_destroy(struct i915_vma *vma)
 837 {
 838         GEM_BUG_ON(vma->node.allocated);
 839         GEM_BUG_ON(vma->fence);
 840
 841         GEM_BUG_ON(i915_active_request_isset(&vma->last_fence));
 842
 843         mutex_lock(&vma->vm->mutex);
 844         list_del(&vma->vm_link);
 845         mutex_unlock(&vma->vm->mutex);
 846
 847         if (vma->obj) {
 848                 struct drm_i915_gem_object *obj = vma->obj;
 849
 850                 spin_lock(&obj->vma.lock);
 851                 list_del(&vma->obj_link);
 852                 rb_erase(&vma->obj_node, &vma->obj->vma.tree);
 853                 spin_unlock(&obj->vma.lock);
 854         }
 855
 856         i915_active_fini(&vma->active);
 857
 858         i915_vma_free(vma);
 859 }
 860
 861 void i915_vma_destroy(struct i915_vma *vma)
 862 {
 863         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 864
 865         GEM_BUG_ON(i915_vma_is_pinned(vma));
 866
 867         __i915_vma_remove_closed(vma);
 868
 869         WARN_ON(i915_vma_unbind(vma));
 870         GEM_BUG_ON(i915_vma_is_active(vma));
 871
 872         __i915_vma_destroy(vma);
 873 }
 874
 875 void i915_vma_parked(struct drm_i915_private *i915)
 876 {
 877         struct i915_vma *vma, *next;
 878
 879         spin_lock_irq(&i915->gt.closed_lock);
 880         list_for_each_entry_safe(vma, next, &i915->gt.closed_vma, closed_link) {
 881                 list_del_init(&vma->closed_link);
 882                 spin_unlock_irq(&i915->gt.closed_lock);
 883
 884                 i915_vma_destroy(vma);
 885
 886                 spin_lock_irq(&i915->gt.closed_lock);
 887         }
 888         spin_unlock_irq(&i915->gt.closed_lock);
 889 }
 890
 891 static void __i915_vma_iounmap(struct i915_vma *vma)
 892 {
 893         GEM_BUG_ON(i915_vma_is_pinned(vma));
 894
 895         if (vma->iomap == NULL)
 896                 return;
 897
 898         io_mapping_unmap(vma->iomap);
 899         vma->iomap = NULL;
 900 }
 901
 902 void i915_vma_revoke_mmap(struct i915_vma *vma)
 903 {
 904         struct drm_vma_offset_node *node = &vma->obj->base.vma_node;
 905         u64 vma_offset;
 906
 907         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 908
 909         if (!i915_vma_has_userfault(vma))
 910                 return;
 911
 912         GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
 913         GEM_BUG_ON(!vma->obj->userfault_count);
 914
 915         vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
 916         unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
 917                             drm_vma_node_offset_addr(node) + vma_offset,
 918                             vma->size,
 919                             1);
 920
 921         i915_vma_unset_userfault(vma);
 922         if (!--vma->obj->userfault_count)
 923                 list_del(&vma->obj->userfault_link);
 924 }
 925
 926 static void export_fence(struct i915_vma *vma,
 927                          struct i915_request *rq,
 928                          unsigned int flags)
 929 {
 930         struct reservation_object *resv = vma->resv;
 931
 932         /*
 933          * Ignore errors from failing to allocate the new fence, we can't
 934          * handle an error right now. Worst case should be missed
 935          * synchronisation leading to rendering corruption.
 936          */
 937         if (flags & EXEC_OBJECT_WRITE)
 938                 reservation_object_add_excl_fence(resv, &rq->fence);
 939         else if (reservation_object_reserve_shared(resv, 1) == 0)
 940                 reservation_object_add_shared_fence(resv, &rq->fence);
 941 }
 942
 943 int i915_vma_move_to_active(struct i915_vma *vma,
 944                             struct i915_request *rq,
 945                             unsigned int flags)
 946 {
 947         struct drm_i915_gem_object *obj = vma->obj;
 948
 949         assert_vma_held(vma);
 950         assert_object_held(obj);
 951         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 952
 953         /*
 954          * Add a reference if we're newly entering the active list.
 955          * The order in which we add operations to the retirement queue is
 956          * vital here: mark_active adds to the start of the callback list,
 957          * such that subsequent callbacks are called first. Therefore we
 958          * add the active reference first and queue for it to be dropped
 959          * *last*.
 960          */
 961         if (!vma->active.count && !obj->active_count++)
 962                 i915_gem_object_get(obj); /* once more for the active ref */
 963
 964         if (unlikely(i915_active_ref(&vma->active, rq->fence.context, rq))) {
 965                 if (!vma->active.count && !--obj->active_count)
 966                         i915_gem_object_put(obj);
 967                 return -ENOMEM;
 968         }
 969
 970         GEM_BUG_ON(!i915_vma_is_active(vma));
 971         GEM_BUG_ON(!obj->active_count);
 972
 973         obj->write_domain = 0;
 974         if (flags & EXEC_OBJECT_WRITE) {
 975                 obj->write_domain = I915_GEM_DOMAIN_RENDER;
 976
 977                 if (intel_fb_obj_invalidate(obj, ORIGIN_CS))
 978                         __i915_active_request_set(&obj->frontbuffer_write, rq);
 979
 980                 obj->read_domains = 0;
 981         }
 982         obj->read_domains |= I915_GEM_GPU_DOMAINS;
 983
 984         if (flags & EXEC_OBJECT_NEEDS_FENCE)
 985                 __i915_active_request_set(&vma->last_fence, rq);
 986
 987         export_fence(vma, rq, flags);
 988         return 0;
 989 }
 990
 991 int i915_vma_unbind(struct i915_vma *vma)
 992 {
 993         int ret;
 994
 995         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 996
 997         /*
 998          * First wait upon any activity as retiring the request may
 999          * have side-effects such as unpinning or even unbinding this vma.
1000          */
1001         might_sleep();
1002         if (i915_vma_is_active(vma)) {
1003                 /*
1004                  * When a closed VMA is retired, it is unbound - eek.
1005                  * In order to prevent it from being recursively closed,
1006                  * take a pin on the vma so that the second unbind is
1007                  * aborted.
1008                  *
1009                  * Even more scary is that the retire callback may free
1010                  * the object (last active vma). To prevent the explosion
1011                  * we defer the actual object free to a worker that can
1012                  * only proceed once it acquires the struct_mutex (which
1013                  * we currently hold, therefore it cannot free this object
1014                  * before we are finished).
1015                  */
1016                 __i915_vma_pin(vma);
1017
1018                 ret = i915_active_wait(&vma->active);
1019                 if (ret)
1020                         goto unpin;
1021
1022                 ret = i915_active_request_retire(&vma->last_fence,
1023                                               &vma->vm->i915->drm.struct_mutex);
1024 unpin:
1025                 __i915_vma_unpin(vma);
1026                 if (ret)
1027                         return ret;
1028         }
1029         GEM_BUG_ON(i915_vma_is_active(vma));
1030
1031         if (i915_vma_is_pinned(vma)) {
1032                 vma_print_allocator(vma, "is pinned");
1033                 return -EBUSY;
1034         }
1035
1036         if (!drm_mm_node_allocated(&vma->node))
1037                 return 0;
1038
1039         if (i915_vma_is_map_and_fenceable(vma)) {
1040                 /*
1041                  * Check that we have flushed all writes through the GGTT
1042                  * before the unbind, other due to non-strict nature of those
1043                  * indirect writes they may end up referencing the GGTT PTE
1044                  * after the unbind.
1045                  */
1046                 i915_vma_flush_writes(vma);
1047                 GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
1048
1049                 /* release the fence reg _after_ flushing */
1050                 ret = i915_vma_put_fence(vma);
1051                 if (ret)
1052                         return ret;
1053
1054                 /* Force a pagefault for domain tracking on next user access */
1055                 i915_vma_revoke_mmap(vma);
1056
1057                 __i915_vma_iounmap(vma);
1058                 vma->flags &= ~I915_VMA_CAN_FENCE;
1059         }
1060         GEM_BUG_ON(vma->fence);
1061         GEM_BUG_ON(i915_vma_has_userfault(vma));
1062
1063         if (likely(!vma->vm->closed)) {
1064                 trace_i915_vma_unbind(vma);
1065                 vma->ops->unbind_vma(vma);
1066         }
1067         vma->flags &= ~(I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
1068
1069         i915_vma_remove(vma);
1070
1071         return 0;
1072 }
1073
1074 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1075 #include "selftests/i915_vma.c"
1076 #endif
1077
1078 static void i915_global_vma_shrink(void)
1079 {
1080         kmem_cache_shrink(global.slab_vmas);
1081 }
1082
1083 static void i915_global_vma_exit(void)
1084 {
1085         kmem_cache_destroy(global.slab_vmas);
1086 }
1087
1088 static struct i915_global_vma global = { {
1089         .shrink = i915_global_vma_shrink,
1090         .exit = i915_global_vma_exit,
1091 } };
1092
1093 int __init i915_global_vma_init(void)
1094 {
1095         global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1096         if (!global.slab_vmas)
1097                 return -ENOMEM;
1098
1099         i915_global_register(&global.base);
1100         return 0;
1101 }