drivers/gpu/drm/i915/i915_vma.c

   1 /*
   2  * Copyright © 2016 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  */
  24
  25 #include "i915_vma.h"
  26
  27 #include "i915_drv.h"
  28 #include "intel_ringbuffer.h"
  29 #include "intel_frontbuffer.h"
  30
  31 #include <drm/drm_gem.h>
  32
  33 #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM)
  34
  35 #include <linux/stackdepot.h>
  36
  37 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
  38 {
  39         unsigned long entries[12];
  40         struct stack_trace trace = {
  41                 .entries = entries,
  42                 .max_entries = ARRAY_SIZE(entries),
  43         };
  44         char buf[512];
  45
  46         if (!vma->node.stack) {
  47                 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",
  48                                  vma->node.start, vma->node.size, reason);
  49                 return;
  50         }
  51
  52         depot_fetch_stack(vma->node.stack, &trace);
  53         snprint_stack_trace(buf, sizeof(buf), &trace, 0);
  54         DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",
  55                          vma->node.start, vma->node.size, reason, buf);
  56 }
  57
  58 #else
  59
  60 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
  61 {
  62 }
  63
  64 #endif
  65
  66 struct i915_vma_active {
  67         struct i915_gem_active base;
  68         struct i915_vma *vma;
  69         struct rb_node node;
  70         u64 timeline;
  71 };
  72
  73 static void
  74 __i915_vma_retire(struct i915_vma *vma, struct i915_request *rq)
  75 {
  76         struct drm_i915_gem_object *obj = vma->obj;
  77
  78         GEM_BUG_ON(!i915_vma_is_active(vma));
  79         if (--vma->active_count)
  80                 return;
  81
  82         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
  83         list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
  84
  85         GEM_BUG_ON(!i915_gem_object_is_active(obj));
  86         if (--obj->active_count)
  87                 return;
  88
  89         /* Prune the shared fence arrays iff completely idle (inc. external) */
  90         if (reservation_object_trylock(obj->resv)) {
  91                 if (reservation_object_test_signaled_rcu(obj->resv, true))
  92                         reservation_object_add_excl_fence(obj->resv, NULL);
  93                 reservation_object_unlock(obj->resv);
  94         }
  95
  96         /* Bump our place on the bound list to keep it roughly in LRU order
  97          * so that we don't steal from recently used but inactive objects
  98          * (unless we are forced to ofc!)
  99          */
 100         spin_lock(&rq->i915->mm.obj_lock);
 101         if (obj->bind_count)
 102                 list_move_tail(&obj->mm.link, &rq->i915->mm.bound_list);
 103         spin_unlock(&rq->i915->mm.obj_lock);
 104
 105         obj->mm.dirty = true; /* be paranoid  */
 106
 107         if (i915_gem_object_has_active_reference(obj)) {
 108                 i915_gem_object_clear_active_reference(obj);
 109                 i915_gem_object_put(obj);
 110         }
 111 }
 112
 113 static void
 114 i915_vma_retire(struct i915_gem_active *base, struct i915_request *rq)
 115 {
 116         struct i915_vma_active *active =
 117                 container_of(base, typeof(*active), base);
 118
 119         __i915_vma_retire(active->vma, rq);
 120 }
 121
 122 static void
 123 i915_vma_last_retire(struct i915_gem_active *base, struct i915_request *rq)
 124 {
 125         __i915_vma_retire(container_of(base, struct i915_vma, last_active), rq);
 126 }
 127
 128 static struct i915_vma *
 129 vma_create(struct drm_i915_gem_object *obj,
 130            struct i915_address_space *vm,
 131            const struct i915_ggtt_view *view)
 132 {
 133         struct i915_vma *vma;
 134         struct rb_node *rb, **p;
 135
 136         /* The aliasing_ppgtt should never be used directly! */
 137         GEM_BUG_ON(vm == &vm->i915->mm.aliasing_ppgtt->vm);
 138
 139         vma = kmem_cache_zalloc(vm->i915->vmas, GFP_KERNEL);
 140         if (vma == NULL)
 141                 return ERR_PTR(-ENOMEM);
 142
 143         vma->active = RB_ROOT;
 144
 145         init_request_active(&vma->last_active, i915_vma_last_retire);
 146         init_request_active(&vma->last_fence, NULL);
 147         vma->vm = vm;
 148         vma->ops = &vm->vma_ops;
 149         vma->obj = obj;
 150         vma->resv = obj->resv;
 151         vma->size = obj->base.size;
 152         vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
 153
 154         if (view && view->type != I915_GGTT_VIEW_NORMAL) {
 155                 vma->ggtt_view = *view;
 156                 if (view->type == I915_GGTT_VIEW_PARTIAL) {
 157                         GEM_BUG_ON(range_overflows_t(u64,
 158                                                      view->partial.offset,
 159                                                      view->partial.size,
 160                                                      obj->base.size >> PAGE_SHIFT));
 161                         vma->size = view->partial.size;
 162                         vma->size <<= PAGE_SHIFT;
 163                         GEM_BUG_ON(vma->size > obj->base.size);
 164                 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
 165                         vma->size = intel_rotation_info_size(&view->rotated);
 166                         vma->size <<= PAGE_SHIFT;
 167                 }
 168         }
 169
 170         if (unlikely(vma->size > vm->total))
 171                 goto err_vma;
 172
 173         GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
 174
 175         if (i915_is_ggtt(vm)) {
 176                 if (unlikely(overflows_type(vma->size, u32)))
 177                         goto err_vma;
 178
 179                 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
 180                                                       i915_gem_object_get_tiling(obj),
 181                                                       i915_gem_object_get_stride(obj));
 182                 if (unlikely(vma->fence_size < vma->size || /* overflow */
 183                              vma->fence_size > vm->total))
 184                         goto err_vma;
 185
 186                 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
 187
 188                 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
 189                                                                 i915_gem_object_get_tiling(obj),
 190                                                                 i915_gem_object_get_stride(obj));
 191                 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
 192
 193                 /*
 194                  * We put the GGTT vma at the start of the vma-list, followed
 195                  * by the ppGGTT vma. This allows us to break early when
 196                  * iterating over only the GGTT vma for an object, see
 197                  * for_each_ggtt_vma()
 198                  */
 199                 vma->flags |= I915_VMA_GGTT;
 200                 list_add(&vma->obj_link, &obj->vma_list);
 201         } else {
 202                 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
 203                 list_add_tail(&vma->obj_link, &obj->vma_list);
 204         }
 205
 206         rb = NULL;
 207         p = &obj->vma_tree.rb_node;
 208         while (*p) {
 209                 struct i915_vma *pos;
 210
 211                 rb = *p;
 212                 pos = rb_entry(rb, struct i915_vma, obj_node);
 213                 if (i915_vma_compare(pos, vm, view) < 0)
 214                         p = &rb->rb_right;
 215                 else
 216                         p = &rb->rb_left;
 217         }
 218         rb_link_node(&vma->obj_node, rb, p);
 219         rb_insert_color(&vma->obj_node, &obj->vma_tree);
 220         list_add(&vma->vm_link, &vm->unbound_list);
 221
 222         return vma;
 223
 224 err_vma:
 225         kmem_cache_free(vm->i915->vmas, vma);
 226         return ERR_PTR(-E2BIG);
 227 }
 228
 229 static struct i915_vma *
 230 vma_lookup(struct drm_i915_gem_object *obj,
 231            struct i915_address_space *vm,
 232            const struct i915_ggtt_view *view)
 233 {
 234         struct rb_node *rb;
 235
 236         rb = obj->vma_tree.rb_node;
 237         while (rb) {
 238                 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
 239                 long cmp;
 240
 241                 cmp = i915_vma_compare(vma, vm, view);
 242                 if (cmp == 0)
 243                         return vma;
 244
 245                 if (cmp < 0)
 246                         rb = rb->rb_right;
 247                 else
 248                         rb = rb->rb_left;
 249         }
 250
 251         return NULL;
 252 }
 253
 254 /**
 255  * i915_vma_instance - return the singleton instance of the VMA
 256  * @obj: parent &struct drm_i915_gem_object to be mapped
 257  * @vm: address space in which the mapping is located
 258  * @view: additional mapping requirements
 259  *
 260  * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
 261  * the same @view characteristics. If a match is not found, one is created.
 262  * Once created, the VMA is kept until either the object is freed, or the
 263  * address space is closed.
 264  *
 265  * Must be called with struct_mutex held.
 266  *
 267  * Returns the vma, or an error pointer.
 268  */
 269 struct i915_vma *
 270 i915_vma_instance(struct drm_i915_gem_object *obj,
 271                   struct i915_address_space *vm,
 272                   const struct i915_ggtt_view *view)
 273 {
 274         struct i915_vma *vma;
 275
 276         lockdep_assert_held(&obj->base.dev->struct_mutex);
 277         GEM_BUG_ON(view && !i915_is_ggtt(vm));
 278         GEM_BUG_ON(vm->closed);
 279
 280         vma = vma_lookup(obj, vm, view);
 281         if (!vma)
 282                 vma = vma_create(obj, vm, view);
 283
 284         GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
 285         GEM_BUG_ON(!IS_ERR(vma) && vma_lookup(obj, vm, view) != vma);
 286         return vma;
 287 }
 288
 289 /**
 290  * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
 291  * @vma: VMA to map
 292  * @cache_level: mapping cache level
 293  * @flags: flags like global or local mapping
 294  *
 295  * DMA addresses are taken from the scatter-gather table of this object (or of
 296  * this VMA in case of non-default GGTT views) and PTE entries set up.
 297  * Note that DMA addresses are also the only part of the SG table we care about.
 298  */
 299 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
 300                   u32 flags)
 301 {
 302         u32 bind_flags;
 303         u32 vma_flags;
 304         int ret;
 305
 306         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 307         GEM_BUG_ON(vma->size > vma->node.size);
 308
 309         if (GEM_WARN_ON(range_overflows(vma->node.start,
 310                                         vma->node.size,
 311                                         vma->vm->total)))
 312                 return -ENODEV;
 313
 314         if (GEM_WARN_ON(!flags))
 315                 return -EINVAL;
 316
 317         bind_flags = 0;
 318         if (flags & PIN_GLOBAL)
 319                 bind_flags |= I915_VMA_GLOBAL_BIND;
 320         if (flags & PIN_USER)
 321                 bind_flags |= I915_VMA_LOCAL_BIND;
 322
 323         vma_flags = vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
 324         if (flags & PIN_UPDATE)
 325                 bind_flags |= vma_flags;
 326         else
 327                 bind_flags &= ~vma_flags;
 328         if (bind_flags == 0)
 329                 return 0;
 330
 331         GEM_BUG_ON(!vma->pages);
 332
 333         trace_i915_vma_bind(vma, bind_flags);
 334         ret = vma->ops->bind_vma(vma, cache_level, bind_flags);
 335         if (ret)
 336                 return ret;
 337
 338         vma->flags |= bind_flags;
 339         return 0;
 340 }
 341
 342 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
 343 {
 344         void __iomem *ptr;
 345         int err;
 346
 347         /* Access through the GTT requires the device to be awake. */
 348         assert_rpm_wakelock_held(vma->vm->i915);
 349
 350         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 351         if (WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
 352                 err = -ENODEV;
 353                 goto err;
 354         }
 355
 356         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 357         GEM_BUG_ON((vma->flags & I915_VMA_GLOBAL_BIND) == 0);
 358
 359         ptr = vma->iomap;
 360         if (ptr == NULL) {
 361                 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
 362                                         vma->node.start,
 363                                         vma->node.size);
 364                 if (ptr == NULL) {
 365                         err = -ENOMEM;
 366                         goto err;
 367                 }
 368
 369                 vma->iomap = ptr;
 370         }
 371
 372         __i915_vma_pin(vma);
 373
 374         err = i915_vma_pin_fence(vma);
 375         if (err)
 376                 goto err_unpin;
 377
 378         i915_vma_set_ggtt_write(vma);
 379         return ptr;
 380
 381 err_unpin:
 382         __i915_vma_unpin(vma);
 383 err:
 384         return IO_ERR_PTR(err);
 385 }
 386
 387 void i915_vma_flush_writes(struct i915_vma *vma)
 388 {
 389         if (!i915_vma_has_ggtt_write(vma))
 390                 return;
 391
 392         i915_gem_flush_ggtt_writes(vma->vm->i915);
 393
 394         i915_vma_unset_ggtt_write(vma);
 395 }
 396
 397 void i915_vma_unpin_iomap(struct i915_vma *vma)
 398 {
 399         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 400
 401         GEM_BUG_ON(vma->iomap == NULL);
 402
 403         i915_vma_flush_writes(vma);
 404
 405         i915_vma_unpin_fence(vma);
 406         i915_vma_unpin(vma);
 407 }
 408
 409 void i915_vma_unpin_and_release(struct i915_vma **p_vma)
 410 {
 411         struct i915_vma *vma;
 412         struct drm_i915_gem_object *obj;
 413
 414         vma = fetch_and_zero(p_vma);
 415         if (!vma)
 416                 return;
 417
 418         obj = vma->obj;
 419         GEM_BUG_ON(!obj);
 420
 421         i915_vma_unpin(vma);
 422         i915_vma_close(vma);
 423
 424         __i915_gem_object_release_unless_active(obj);
 425 }
 426
 427 bool i915_vma_misplaced(const struct i915_vma *vma,
 428                         u64 size, u64 alignment, u64 flags)
 429 {
 430         if (!drm_mm_node_allocated(&vma->node))
 431                 return false;
 432
 433         if (vma->node.size < size)
 434                 return true;
 435
 436         GEM_BUG_ON(alignment && !is_power_of_2(alignment));
 437         if (alignment && !IS_ALIGNED(vma->node.start, alignment))
 438                 return true;
 439
 440         if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
 441                 return true;
 442
 443         if (flags & PIN_OFFSET_BIAS &&
 444             vma->node.start < (flags & PIN_OFFSET_MASK))
 445                 return true;
 446
 447         if (flags & PIN_OFFSET_FIXED &&
 448             vma->node.start != (flags & PIN_OFFSET_MASK))
 449                 return true;
 450
 451         return false;
 452 }
 453
 454 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
 455 {
 456         bool mappable, fenceable;
 457
 458         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 459         GEM_BUG_ON(!vma->fence_size);
 460
 461         /*
 462          * Explicitly disable for rotated VMA since the display does not
 463          * need the fence and the VMA is not accessible to other users.
 464          */
 465         if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
 466                 return;
 467
 468         fenceable = (vma->node.size >= vma->fence_size &&
 469                      IS_ALIGNED(vma->node.start, vma->fence_alignment));
 470
 471         mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
 472
 473         if (mappable && fenceable)
 474                 vma->flags |= I915_VMA_CAN_FENCE;
 475         else
 476                 vma->flags &= ~I915_VMA_CAN_FENCE;
 477 }
 478
 479 static bool color_differs(struct drm_mm_node *node, unsigned long color)
 480 {
 481         return node->allocated && node->color != color;
 482 }
 483
 484 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long cache_level)
 485 {
 486         struct drm_mm_node *node = &vma->node;
 487         struct drm_mm_node *other;
 488
 489         /*
 490          * On some machines we have to be careful when putting differing types
 491          * of snoopable memory together to avoid the prefetcher crossing memory
 492          * domains and dying. During vm initialisation, we decide whether or not
 493          * these constraints apply and set the drm_mm.color_adjust
 494          * appropriately.
 495          */
 496         if (vma->vm->mm.color_adjust == NULL)
 497                 return true;
 498
 499         /* Only valid to be called on an already inserted vma */
 500         GEM_BUG_ON(!drm_mm_node_allocated(node));
 501         GEM_BUG_ON(list_empty(&node->node_list));
 502
 503         other = list_prev_entry(node, node_list);
 504         if (color_differs(other, cache_level) && !drm_mm_hole_follows(other))
 505                 return false;
 506
 507         other = list_next_entry(node, node_list);
 508         if (color_differs(other, cache_level) && !drm_mm_hole_follows(node))
 509                 return false;
 510
 511         return true;
 512 }
 513
 514 static void assert_bind_count(const struct drm_i915_gem_object *obj)
 515 {
 516         /*
 517          * Combine the assertion that the object is bound and that we have
 518          * pinned its pages. But we should never have bound the object
 519          * more than we have pinned its pages. (For complete accuracy, we
 520          * assume that no else is pinning the pages, but as a rough assertion
 521          * that we will not run into problems later, this will do!)
 522          */
 523         GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < obj->bind_count);
 524 }
 525
 526 /**
 527  * i915_vma_insert - finds a slot for the vma in its address space
 528  * @vma: the vma
 529  * @size: requested size in bytes (can be larger than the VMA)
 530  * @alignment: required alignment
 531  * @flags: mask of PIN_* flags to use
 532  *
 533  * First we try to allocate some free space that meets the requirements for
 534  * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
 535  * preferrably the oldest idle entry to make room for the new VMA.
 536  *
 537  * Returns:
 538  * 0 on success, negative error code otherwise.
 539  */
 540 static int
 541 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 542 {
 543         struct drm_i915_private *dev_priv = vma->vm->i915;
 544         unsigned int cache_level;
 545         u64 start, end;
 546         int ret;
 547
 548         GEM_BUG_ON(i915_vma_is_closed(vma));
 549         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 550         GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
 551
 552         size = max(size, vma->size);
 553         alignment = max(alignment, vma->display_alignment);
 554         if (flags & PIN_MAPPABLE) {
 555                 size = max_t(typeof(size), size, vma->fence_size);
 556                 alignment = max_t(typeof(alignment),
 557                                   alignment, vma->fence_alignment);
 558         }
 559
 560         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
 561         GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
 562         GEM_BUG_ON(!is_power_of_2(alignment));
 563
 564         start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
 565         GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
 566
 567         end = vma->vm->total;
 568         if (flags & PIN_MAPPABLE)
 569                 end = min_t(u64, end, dev_priv->ggtt.mappable_end);
 570         if (flags & PIN_ZONE_4G)
 571                 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
 572         GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
 573
 574         /* If binding the object/GGTT view requires more space than the entire
 575          * aperture has, reject it early before evicting everything in a vain
 576          * attempt to find space.
 577          */
 578         if (size > end) {
 579                 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
 580                           size, flags & PIN_MAPPABLE ? "mappable" : "total",
 581                           end);
 582                 return -ENOSPC;
 583         }
 584
 585         if (vma->obj) {
 586                 ret = i915_gem_object_pin_pages(vma->obj);
 587                 if (ret)
 588                         return ret;
 589
 590                 cache_level = vma->obj->cache_level;
 591         } else {
 592                 cache_level = 0;
 593         }
 594
 595         GEM_BUG_ON(vma->pages);
 596
 597         ret = vma->ops->set_pages(vma);
 598         if (ret)
 599                 goto err_unpin;
 600
 601         if (flags & PIN_OFFSET_FIXED) {
 602                 u64 offset = flags & PIN_OFFSET_MASK;
 603                 if (!IS_ALIGNED(offset, alignment) ||
 604                     range_overflows(offset, size, end)) {
 605                         ret = -EINVAL;
 606                         goto err_clear;
 607                 }
 608
 609                 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
 610                                            size, offset, cache_level,
 611                                            flags);
 612                 if (ret)
 613                         goto err_clear;
 614         } else {
 615                 /*
 616                  * We only support huge gtt pages through the 48b PPGTT,
 617                  * however we also don't want to force any alignment for
 618                  * objects which need to be tightly packed into the low 32bits.
 619                  *
 620                  * Note that we assume that GGTT are limited to 4GiB for the
 621                  * forseeable future. See also i915_ggtt_offset().
 622                  */
 623                 if (upper_32_bits(end - 1) &&
 624                     vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
 625                         /*
 626                          * We can't mix 64K and 4K PTEs in the same page-table
 627                          * (2M block), and so to avoid the ugliness and
 628                          * complexity of coloring we opt for just aligning 64K
 629                          * objects to 2M.
 630                          */
 631                         u64 page_alignment =
 632                                 rounddown_pow_of_two(vma->page_sizes.sg |
 633                                                      I915_GTT_PAGE_SIZE_2M);
 634
 635                         /*
 636                          * Check we don't expand for the limited Global GTT
 637                          * (mappable aperture is even more precious!). This
 638                          * also checks that we exclude the aliasing-ppgtt.
 639                          */
 640                         GEM_BUG_ON(i915_vma_is_ggtt(vma));
 641
 642                         alignment = max(alignment, page_alignment);
 643
 644                         if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
 645                                 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
 646                 }
 647
 648                 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
 649                                           size, alignment, cache_level,
 650                                           start, end, flags);
 651                 if (ret)
 652                         goto err_clear;
 653
 654                 GEM_BUG_ON(vma->node.start < start);
 655                 GEM_BUG_ON(vma->node.start + vma->node.size > end);
 656         }
 657         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 658         GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, cache_level));
 659
 660         list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
 661
 662         if (vma->obj) {
 663                 struct drm_i915_gem_object *obj = vma->obj;
 664
 665                 spin_lock(&dev_priv->mm.obj_lock);
 666                 list_move_tail(&obj->mm.link, &dev_priv->mm.bound_list);
 667                 obj->bind_count++;
 668                 spin_unlock(&dev_priv->mm.obj_lock);
 669
 670                 assert_bind_count(obj);
 671         }
 672
 673         return 0;
 674
 675 err_clear:
 676         vma->ops->clear_pages(vma);
 677 err_unpin:
 678         if (vma->obj)
 679                 i915_gem_object_unpin_pages(vma->obj);
 680         return ret;
 681 }
 682
 683 static void
 684 i915_vma_remove(struct i915_vma *vma)
 685 {
 686         struct drm_i915_private *i915 = vma->vm->i915;
 687
 688         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 689         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 690
 691         vma->ops->clear_pages(vma);
 692
 693         drm_mm_remove_node(&vma->node);
 694         list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
 695
 696         /*
 697          * Since the unbound list is global, only move to that list if
 698          * no more VMAs exist.
 699          */
 700         if (vma->obj) {
 701                 struct drm_i915_gem_object *obj = vma->obj;
 702
 703                 spin_lock(&i915->mm.obj_lock);
 704                 if (--obj->bind_count == 0)
 705                         list_move_tail(&obj->mm.link, &i915->mm.unbound_list);
 706                 spin_unlock(&i915->mm.obj_lock);
 707
 708                 /*
 709                  * And finally now the object is completely decoupled from this
 710                  * vma, we can drop its hold on the backing storage and allow
 711                  * it to be reaped by the shrinker.
 712                  */
 713                 i915_gem_object_unpin_pages(obj);
 714                 assert_bind_count(obj);
 715         }
 716 }
 717
 718 int __i915_vma_do_pin(struct i915_vma *vma,
 719                       u64 size, u64 alignment, u64 flags)
 720 {
 721         const unsigned int bound = vma->flags;
 722         int ret;
 723
 724         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 725         GEM_BUG_ON((flags & (PIN_GLOBAL | PIN_USER)) == 0);
 726         GEM_BUG_ON((flags & PIN_GLOBAL) && !i915_vma_is_ggtt(vma));
 727
 728         if (WARN_ON(bound & I915_VMA_PIN_OVERFLOW)) {
 729                 ret = -EBUSY;
 730                 goto err_unpin;
 731         }
 732
 733         if ((bound & I915_VMA_BIND_MASK) == 0) {
 734                 ret = i915_vma_insert(vma, size, alignment, flags);
 735                 if (ret)
 736                         goto err_unpin;
 737         }
 738         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 739
 740         ret = i915_vma_bind(vma, vma->obj ? vma->obj->cache_level : 0, flags);
 741         if (ret)
 742                 goto err_remove;
 743
 744         GEM_BUG_ON((vma->flags & I915_VMA_BIND_MASK) == 0);
 745
 746         if ((bound ^ vma->flags) & I915_VMA_GLOBAL_BIND)
 747                 __i915_vma_set_map_and_fenceable(vma);
 748
 749         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
 750         return 0;
 751
 752 err_remove:
 753         if ((bound & I915_VMA_BIND_MASK) == 0) {
 754                 i915_vma_remove(vma);
 755                 GEM_BUG_ON(vma->pages);
 756                 GEM_BUG_ON(vma->flags & I915_VMA_BIND_MASK);
 757         }
 758 err_unpin:
 759         __i915_vma_unpin(vma);
 760         return ret;
 761 }
 762
 763 void i915_vma_close(struct i915_vma *vma)
 764 {
 765         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 766
 767         GEM_BUG_ON(i915_vma_is_closed(vma));
 768         vma->flags |= I915_VMA_CLOSED;
 769
 770         /*
 771          * We defer actually closing, unbinding and destroying the VMA until
 772          * the next idle point, or if the object is freed in the meantime. By
 773          * postponing the unbind, we allow for it to be resurrected by the
 774          * client, avoiding the work required to rebind the VMA. This is
 775          * advantageous for DRI, where the client/server pass objects
 776          * between themselves, temporarily opening a local VMA to the
 777          * object, and then closing it again. The same object is then reused
 778          * on the next frame (or two, depending on the depth of the swap queue)
 779          * causing us to rebind the VMA once more. This ends up being a lot
 780          * of wasted work for the steady state.
 781          */
 782         list_add_tail(&vma->closed_link, &vma->vm->i915->gt.closed_vma);
 783 }
 784
 785 void i915_vma_reopen(struct i915_vma *vma)
 786 {
 787         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 788
 789         if (vma->flags & I915_VMA_CLOSED) {
 790                 vma->flags &= ~I915_VMA_CLOSED;
 791                 list_del(&vma->closed_link);
 792         }
 793 }
 794
 795 static void __i915_vma_destroy(struct i915_vma *vma)
 796 {
 797         struct drm_i915_private *i915 = vma->vm->i915;
 798         struct i915_vma_active *iter, *n;
 799
 800         GEM_BUG_ON(vma->node.allocated);
 801         GEM_BUG_ON(vma->fence);
 802
 803         GEM_BUG_ON(i915_gem_active_isset(&vma->last_fence));
 804
 805         list_del(&vma->obj_link);
 806         list_del(&vma->vm_link);
 807         if (vma->obj)
 808                 rb_erase(&vma->obj_node, &vma->obj->vma_tree);
 809
 810         if (!i915_vma_is_ggtt(vma))
 811                 i915_ppgtt_put(i915_vm_to_ppgtt(vma->vm));
 812
 813         rbtree_postorder_for_each_entry_safe(iter, n, &vma->active, node) {
 814                 GEM_BUG_ON(i915_gem_active_isset(&iter->base));
 815                 kfree(iter);
 816         }
 817
 818         kmem_cache_free(i915->vmas, vma);
 819 }
 820
 821 void i915_vma_destroy(struct i915_vma *vma)
 822 {
 823         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 824
 825         GEM_BUG_ON(i915_vma_is_active(vma));
 826         GEM_BUG_ON(i915_vma_is_pinned(vma));
 827
 828         if (i915_vma_is_closed(vma))
 829                 list_del(&vma->closed_link);
 830
 831         WARN_ON(i915_vma_unbind(vma));
 832         __i915_vma_destroy(vma);
 833 }
 834
 835 void i915_vma_parked(struct drm_i915_private *i915)
 836 {
 837         struct i915_vma *vma, *next;
 838
 839         list_for_each_entry_safe(vma, next, &i915->gt.closed_vma, closed_link) {
 840                 GEM_BUG_ON(!i915_vma_is_closed(vma));
 841                 i915_vma_destroy(vma);
 842         }
 843
 844         GEM_BUG_ON(!list_empty(&i915->gt.closed_vma));
 845 }
 846
 847 static void __i915_vma_iounmap(struct i915_vma *vma)
 848 {
 849         GEM_BUG_ON(i915_vma_is_pinned(vma));
 850
 851         if (vma->iomap == NULL)
 852                 return;
 853
 854         io_mapping_unmap(vma->iomap);
 855         vma->iomap = NULL;
 856 }
 857
 858 void i915_vma_revoke_mmap(struct i915_vma *vma)
 859 {
 860         struct drm_vma_offset_node *node = &vma->obj->base.vma_node;
 861         u64 vma_offset;
 862
 863         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 864
 865         if (!i915_vma_has_userfault(vma))
 866                 return;
 867
 868         GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
 869         GEM_BUG_ON(!vma->obj->userfault_count);
 870
 871         vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
 872         unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
 873                             drm_vma_node_offset_addr(node) + vma_offset,
 874                             vma->size,
 875                             1);
 876
 877         i915_vma_unset_userfault(vma);
 878         if (!--vma->obj->userfault_count)
 879                 list_del(&vma->obj->userfault_link);
 880 }
 881
 882 static void export_fence(struct i915_vma *vma,
 883                          struct i915_request *rq,
 884                          unsigned int flags)
 885 {
 886         struct reservation_object *resv = vma->resv;
 887
 888         /*
 889          * Ignore errors from failing to allocate the new fence, we can't
 890          * handle an error right now. Worst case should be missed
 891          * synchronisation leading to rendering corruption.
 892          */
 893         reservation_object_lock(resv, NULL);
 894         if (flags & EXEC_OBJECT_WRITE)
 895                 reservation_object_add_excl_fence(resv, &rq->fence);
 896         else if (reservation_object_reserve_shared(resv) == 0)
 897                 reservation_object_add_shared_fence(resv, &rq->fence);
 898         reservation_object_unlock(resv);
 899 }
 900
 901 static struct i915_gem_active *active_instance(struct i915_vma *vma, u64 idx)
 902 {
 903         struct i915_vma_active *active;
 904         struct rb_node **p, *parent;
 905         struct i915_request *old;
 906
 907         /*
 908          * We track the most recently used timeline to skip a rbtree search
 909          * for the common case, under typical loads we never need the rbtree
 910          * at all. We can reuse the last_active slot if it is empty, that is
 911          * after the previous activity has been retired, or if the active
 912          * matches the current timeline.
 913          *
 914          * Note that we allow the timeline to be active simultaneously in
 915          * the rbtree and the last_active cache. We do this to avoid having
 916          * to search and replace the rbtree element for a new timeline, with
 917          * the cost being that we must be aware that the vma may be retired
 918          * twice for the same timeline (as the older rbtree element will be
 919          * retired before the new request added to last_active).
 920          */
 921         old = i915_gem_active_raw(&vma->last_active,
 922                                   &vma->vm->i915->drm.struct_mutex);
 923         if (!old || old->fence.context == idx)
 924                 goto out;
 925
 926         /* Move the currently active fence into the rbtree */
 927         idx = old->fence.context;
 928
 929         parent = NULL;
 930         p = &vma->active.rb_node;
 931         while (*p) {
 932                 parent = *p;
 933
 934                 active = rb_entry(parent, struct i915_vma_active, node);
 935                 if (active->timeline == idx)
 936                         goto replace;
 937
 938                 if (active->timeline < idx)
 939                         p = &parent->rb_right;
 940                 else
 941                         p = &parent->rb_left;
 942         }
 943
 944         active = kmalloc(sizeof(*active), GFP_KERNEL);
 945
 946         /* kmalloc may retire the vma->last_active request (thanks shrinker)! */
 947         if (unlikely(!i915_gem_active_raw(&vma->last_active,
 948                                           &vma->vm->i915->drm.struct_mutex))) {
 949                 kfree(active);
 950                 goto out;
 951         }
 952
 953         if (unlikely(!active))
 954                 return ERR_PTR(-ENOMEM);
 955
 956         init_request_active(&active->base, i915_vma_retire);
 957         active->vma = vma;
 958         active->timeline = idx;
 959
 960         rb_link_node(&active->node, parent, p);
 961         rb_insert_color(&active->node, &vma->active);
 962
 963 replace:
 964         /*
 965          * Overwrite the previous active slot in the rbtree with last_active,
 966          * leaving last_active zeroed. If the previous slot is still active,
 967          * we must be careful as we now only expect to receive one retire
 968          * callback not two, and so much undo the active counting for the
 969          * overwritten slot.
 970          */
 971         if (i915_gem_active_isset(&active->base)) {
 972                 /* Retire ourselves from the old rq->active_list */
 973                 __list_del_entry(&active->base.link);
 974                 vma->active_count--;
 975                 GEM_BUG_ON(!vma->active_count);
 976         }
 977         GEM_BUG_ON(list_empty(&vma->last_active.link));
 978         list_replace_init(&vma->last_active.link, &active->base.link);
 979         active->base.request = fetch_and_zero(&vma->last_active.request);
 980
 981 out:
 982         return &vma->last_active;
 983 }
 984
 985 int i915_vma_move_to_active(struct i915_vma *vma,
 986                             struct i915_request *rq,
 987                             unsigned int flags)
 988 {
 989         struct drm_i915_gem_object *obj = vma->obj;
 990         struct i915_gem_active *active;
 991
 992         lockdep_assert_held(&rq->i915->drm.struct_mutex);
 993         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 994
 995         active = active_instance(vma, rq->fence.context);
 996         if (IS_ERR(active))
 997                 return PTR_ERR(active);
 998
 999         /*
1000          * Add a reference if we're newly entering the active list.
1001          * The order in which we add operations to the retirement queue is
1002          * vital here: mark_active adds to the start of the callback list,
1003          * such that subsequent callbacks are called first. Therefore we
1004          * add the active reference first and queue for it to be dropped
1005          * *last*.
1006          */
1007         if (!i915_gem_active_isset(active) && !vma->active_count++) {
1008                 list_move_tail(&vma->vm_link, &vma->vm->active_list);
1009                 obj->active_count++;
1010         }
1011         i915_gem_active_set(active, rq);
1012         GEM_BUG_ON(!i915_vma_is_active(vma));
1013         GEM_BUG_ON(!obj->active_count);
1014
1015         obj->write_domain = 0;
1016         if (flags & EXEC_OBJECT_WRITE) {
1017                 obj->write_domain = I915_GEM_DOMAIN_RENDER;
1018
1019                 if (intel_fb_obj_invalidate(obj, ORIGIN_CS))
1020                         i915_gem_active_set(&obj->frontbuffer_write, rq);
1021
1022                 obj->read_domains = 0;
1023         }
1024         obj->read_domains |= I915_GEM_GPU_DOMAINS;
1025
1026         if (flags & EXEC_OBJECT_NEEDS_FENCE)
1027                 i915_gem_active_set(&vma->last_fence, rq);
1028
1029         export_fence(vma, rq, flags);
1030         return 0;
1031 }
1032
1033 int i915_vma_unbind(struct i915_vma *vma)
1034 {
1035         int ret;
1036
1037         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
1038
1039         /*
1040          * First wait upon any activity as retiring the request may
1041          * have side-effects such as unpinning or even unbinding this vma.
1042          */
1043         might_sleep();
1044         if (i915_vma_is_active(vma)) {
1045                 struct i915_vma_active *active, *n;
1046
1047                 /*
1048                  * When a closed VMA is retired, it is unbound - eek.
1049                  * In order to prevent it from being recursively closed,
1050                  * take a pin on the vma so that the second unbind is
1051                  * aborted.
1052                  *
1053                  * Even more scary is that the retire callback may free
1054                  * the object (last active vma). To prevent the explosion
1055                  * we defer the actual object free to a worker that can
1056                  * only proceed once it acquires the struct_mutex (which
1057                  * we currently hold, therefore it cannot free this object
1058                  * before we are finished).
1059                  */
1060                 __i915_vma_pin(vma);
1061
1062                 ret = i915_gem_active_retire(&vma->last_active,
1063                                              &vma->vm->i915->drm.struct_mutex);
1064                 if (ret)
1065                         goto unpin;
1066
1067                 rbtree_postorder_for_each_entry_safe(active, n,
1068                                                      &vma->active, node) {
1069                         ret = i915_gem_active_retire(&active->base,
1070                                                      &vma->vm->i915->drm.struct_mutex);
1071                         if (ret)
1072                                 goto unpin;
1073                 }
1074
1075                 ret = i915_gem_active_retire(&vma->last_fence,
1076                                              &vma->vm->i915->drm.struct_mutex);
1077 unpin:
1078                 __i915_vma_unpin(vma);
1079                 if (ret)
1080                         return ret;
1081         }
1082         GEM_BUG_ON(i915_vma_is_active(vma));
1083
1084         if (i915_vma_is_pinned(vma)) {
1085                 vma_print_allocator(vma, "is pinned");
1086                 return -EBUSY;
1087         }
1088
1089         if (!drm_mm_node_allocated(&vma->node))
1090                 return 0;
1091
1092         if (i915_vma_is_map_and_fenceable(vma)) {
1093                 /*
1094                  * Check that we have flushed all writes through the GGTT
1095                  * before the unbind, other due to non-strict nature of those
1096                  * indirect writes they may end up referencing the GGTT PTE
1097                  * after the unbind.
1098                  */
1099                 i915_vma_flush_writes(vma);
1100                 GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
1101
1102                 /* release the fence reg _after_ flushing */
1103                 ret = i915_vma_put_fence(vma);
1104                 if (ret)
1105                         return ret;
1106
1107                 /* Force a pagefault for domain tracking on next user access */
1108                 i915_vma_revoke_mmap(vma);
1109
1110                 __i915_vma_iounmap(vma);
1111                 vma->flags &= ~I915_VMA_CAN_FENCE;
1112         }
1113         GEM_BUG_ON(vma->fence);
1114         GEM_BUG_ON(i915_vma_has_userfault(vma));
1115
1116         if (likely(!vma->vm->closed)) {
1117                 trace_i915_vma_unbind(vma);
1118                 vma->ops->unbind_vma(vma);
1119         }
1120         vma->flags &= ~(I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
1121
1122         i915_vma_remove(vma);
1123
1124         return 0;
1125 }
1126
1127 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1128 #include "selftests/i915_vma.c"
1129 #endif