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                 list_add_tail(&vma->obj_link, &obj->vma_list);
 203         }
 204
 205         rb = NULL;
 206         p = &obj->vma_tree.rb_node;
 207         while (*p) {
 208                 struct i915_vma *pos;
 209
 210                 rb = *p;
 211                 pos = rb_entry(rb, struct i915_vma, obj_node);
 212                 if (i915_vma_compare(pos, vm, view) < 0)
 213                         p = &rb->rb_right;
 214                 else
 215                         p = &rb->rb_left;
 216         }
 217         rb_link_node(&vma->obj_node, rb, p);
 218         rb_insert_color(&vma->obj_node, &obj->vma_tree);
 219         list_add(&vma->vm_link, &vm->unbound_list);
 220
 221         return vma;
 222
 223 err_vma:
 224         kmem_cache_free(vm->i915->vmas, vma);
 225         return ERR_PTR(-E2BIG);
 226 }
 227
 228 static struct i915_vma *
 229 vma_lookup(struct drm_i915_gem_object *obj,
 230            struct i915_address_space *vm,
 231            const struct i915_ggtt_view *view)
 232 {
 233         struct rb_node *rb;
 234
 235         rb = obj->vma_tree.rb_node;
 236         while (rb) {
 237                 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
 238                 long cmp;
 239
 240                 cmp = i915_vma_compare(vma, vm, view);
 241                 if (cmp == 0)
 242                         return vma;
 243
 244                 if (cmp < 0)
 245                         rb = rb->rb_right;
 246                 else
 247                         rb = rb->rb_left;
 248         }
 249
 250         return NULL;
 251 }
 252
 253 /**
 254  * i915_vma_instance - return the singleton instance of the VMA
 255  * @obj: parent &struct drm_i915_gem_object to be mapped
 256  * @vm: address space in which the mapping is located
 257  * @view: additional mapping requirements
 258  *
 259  * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
 260  * the same @view characteristics. If a match is not found, one is created.
 261  * Once created, the VMA is kept until either the object is freed, or the
 262  * address space is closed.
 263  *
 264  * Must be called with struct_mutex held.
 265  *
 266  * Returns the vma, or an error pointer.
 267  */
 268 struct i915_vma *
 269 i915_vma_instance(struct drm_i915_gem_object *obj,
 270                   struct i915_address_space *vm,
 271                   const struct i915_ggtt_view *view)
 272 {
 273         struct i915_vma *vma;
 274
 275         lockdep_assert_held(&obj->base.dev->struct_mutex);
 276         GEM_BUG_ON(view && !i915_is_ggtt(vm));
 277         GEM_BUG_ON(vm->closed);
 278
 279         vma = vma_lookup(obj, vm, view);
 280         if (!vma)
 281                 vma = vma_create(obj, vm, view);
 282
 283         GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
 284         GEM_BUG_ON(!IS_ERR(vma) && vma_lookup(obj, vm, view) != vma);
 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         list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
 663
 664         if (vma->obj) {
 665                 struct drm_i915_gem_object *obj = vma->obj;
 666
 667                 spin_lock(&dev_priv->mm.obj_lock);
 668                 list_move_tail(&obj->mm.link, &dev_priv->mm.bound_list);
 669                 obj->bind_count++;
 670                 spin_unlock(&dev_priv->mm.obj_lock);
 671
 672                 assert_bind_count(obj);
 673         }
 674
 675         return 0;
 676
 677 err_clear:
 678         vma->ops->clear_pages(vma);
 679 err_unpin:
 680         if (vma->obj)
 681                 i915_gem_object_unpin_pages(vma->obj);
 682         return ret;
 683 }
 684
 685 static void
 686 i915_vma_remove(struct i915_vma *vma)
 687 {
 688         struct drm_i915_private *i915 = vma->vm->i915;
 689
 690         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 691         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 692
 693         vma->ops->clear_pages(vma);
 694
 695         drm_mm_remove_node(&vma->node);
 696         list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
 697
 698         /*
 699          * Since the unbound list is global, only move to that list if
 700          * no more VMAs exist.
 701          */
 702         if (vma->obj) {
 703                 struct drm_i915_gem_object *obj = vma->obj;
 704
 705                 spin_lock(&i915->mm.obj_lock);
 706                 if (--obj->bind_count == 0)
 707                         list_move_tail(&obj->mm.link, &i915->mm.unbound_list);
 708                 spin_unlock(&i915->mm.obj_lock);
 709
 710                 /*
 711                  * And finally now the object is completely decoupled from this
 712                  * vma, we can drop its hold on the backing storage and allow
 713                  * it to be reaped by the shrinker.
 714                  */
 715                 i915_gem_object_unpin_pages(obj);
 716                 assert_bind_count(obj);
 717         }
 718 }
 719
 720 int __i915_vma_do_pin(struct i915_vma *vma,
 721                       u64 size, u64 alignment, u64 flags)
 722 {
 723         const unsigned int bound = vma->flags;
 724         int ret;
 725
 726         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 727         GEM_BUG_ON((flags & (PIN_GLOBAL | PIN_USER)) == 0);
 728         GEM_BUG_ON((flags & PIN_GLOBAL) && !i915_vma_is_ggtt(vma));
 729
 730         if (WARN_ON(bound & I915_VMA_PIN_OVERFLOW)) {
 731                 ret = -EBUSY;
 732                 goto err_unpin;
 733         }
 734
 735         if ((bound & I915_VMA_BIND_MASK) == 0) {
 736                 ret = i915_vma_insert(vma, size, alignment, flags);
 737                 if (ret)
 738                         goto err_unpin;
 739         }
 740         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 741
 742         ret = i915_vma_bind(vma, vma->obj ? vma->obj->cache_level : 0, flags);
 743         if (ret)
 744                 goto err_remove;
 745
 746         GEM_BUG_ON((vma->flags & I915_VMA_BIND_MASK) == 0);
 747
 748         if ((bound ^ vma->flags) & I915_VMA_GLOBAL_BIND)
 749                 __i915_vma_set_map_and_fenceable(vma);
 750
 751         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
 752         return 0;
 753
 754 err_remove:
 755         if ((bound & I915_VMA_BIND_MASK) == 0) {
 756                 i915_vma_remove(vma);
 757                 GEM_BUG_ON(vma->pages);
 758                 GEM_BUG_ON(vma->flags & I915_VMA_BIND_MASK);
 759         }
 760 err_unpin:
 761         __i915_vma_unpin(vma);
 762         return ret;
 763 }
 764
 765 void i915_vma_close(struct i915_vma *vma)
 766 {
 767         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 768
 769         GEM_BUG_ON(i915_vma_is_closed(vma));
 770         vma->flags |= I915_VMA_CLOSED;
 771
 772         /*
 773          * We defer actually closing, unbinding and destroying the VMA until
 774          * the next idle point, or if the object is freed in the meantime. By
 775          * postponing the unbind, we allow for it to be resurrected by the
 776          * client, avoiding the work required to rebind the VMA. This is
 777          * advantageous for DRI, where the client/server pass objects
 778          * between themselves, temporarily opening a local VMA to the
 779          * object, and then closing it again. The same object is then reused
 780          * on the next frame (or two, depending on the depth of the swap queue)
 781          * causing us to rebind the VMA once more. This ends up being a lot
 782          * of wasted work for the steady state.
 783          */
 784         list_add_tail(&vma->closed_link, &vma->vm->i915->gt.closed_vma);
 785 }
 786
 787 void i915_vma_reopen(struct i915_vma *vma)
 788 {
 789         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 790
 791         if (vma->flags & I915_VMA_CLOSED) {
 792                 vma->flags &= ~I915_VMA_CLOSED;
 793                 list_del(&vma->closed_link);
 794         }
 795 }
 796
 797 static void __i915_vma_destroy(struct i915_vma *vma)
 798 {
 799         struct drm_i915_private *i915 = vma->vm->i915;
 800         struct i915_vma_active *iter, *n;
 801
 802         GEM_BUG_ON(vma->node.allocated);
 803         GEM_BUG_ON(vma->fence);
 804
 805         GEM_BUG_ON(i915_gem_active_isset(&vma->last_fence));
 806
 807         list_del(&vma->obj_link);
 808         list_del(&vma->vm_link);
 809         if (vma->obj)
 810                 rb_erase(&vma->obj_node, &vma->obj->vma_tree);
 811
 812         rbtree_postorder_for_each_entry_safe(iter, n, &vma->active, node) {
 813                 GEM_BUG_ON(i915_gem_active_isset(&iter->base));
 814                 kfree(iter);
 815         }
 816
 817         kmem_cache_free(i915->vmas, vma);
 818 }
 819
 820 void i915_vma_destroy(struct i915_vma *vma)
 821 {
 822         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 823
 824         GEM_BUG_ON(i915_vma_is_active(vma));
 825         GEM_BUG_ON(i915_vma_is_pinned(vma));
 826
 827         if (i915_vma_is_closed(vma))
 828                 list_del(&vma->closed_link);
 829
 830         WARN_ON(i915_vma_unbind(vma));
 831         __i915_vma_destroy(vma);
 832 }
 833
 834 void i915_vma_parked(struct drm_i915_private *i915)
 835 {
 836         struct i915_vma *vma, *next;
 837
 838         list_for_each_entry_safe(vma, next, &i915->gt.closed_vma, closed_link) {
 839                 GEM_BUG_ON(!i915_vma_is_closed(vma));
 840                 i915_vma_destroy(vma);
 841         }
 842
 843         GEM_BUG_ON(!list_empty(&i915->gt.closed_vma));
 844 }
 845
 846 static void __i915_vma_iounmap(struct i915_vma *vma)
 847 {
 848         GEM_BUG_ON(i915_vma_is_pinned(vma));
 849
 850         if (vma->iomap == NULL)
 851                 return;
 852
 853         io_mapping_unmap(vma->iomap);
 854         vma->iomap = NULL;
 855 }
 856
 857 void i915_vma_revoke_mmap(struct i915_vma *vma)
 858 {
 859         struct drm_vma_offset_node *node = &vma->obj->base.vma_node;
 860         u64 vma_offset;
 861
 862         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 863
 864         if (!i915_vma_has_userfault(vma))
 865                 return;
 866
 867         GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
 868         GEM_BUG_ON(!vma->obj->userfault_count);
 869
 870         vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
 871         unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
 872                             drm_vma_node_offset_addr(node) + vma_offset,
 873                             vma->size,
 874                             1);
 875
 876         i915_vma_unset_userfault(vma);
 877         if (!--vma->obj->userfault_count)
 878                 list_del(&vma->obj->userfault_link);
 879 }
 880
 881 static void export_fence(struct i915_vma *vma,
 882                          struct i915_request *rq,
 883                          unsigned int flags)
 884 {
 885         struct reservation_object *resv = vma->resv;
 886
 887         /*
 888          * Ignore errors from failing to allocate the new fence, we can't
 889          * handle an error right now. Worst case should be missed
 890          * synchronisation leading to rendering corruption.
 891          */
 892         reservation_object_lock(resv, NULL);
 893         if (flags & EXEC_OBJECT_WRITE)
 894                 reservation_object_add_excl_fence(resv, &rq->fence);
 895         else if (reservation_object_reserve_shared(resv, 1) == 0)
 896                 reservation_object_add_shared_fence(resv, &rq->fence);
 897         reservation_object_unlock(resv);
 898 }
 899
 900 static struct i915_gem_active *active_instance(struct i915_vma *vma, u64 idx)
 901 {
 902         struct i915_vma_active *active;
 903         struct rb_node **p, *parent;
 904         struct i915_request *old;
 905
 906         /*
 907          * We track the most recently used timeline to skip a rbtree search
 908          * for the common case, under typical loads we never need the rbtree
 909          * at all. We can reuse the last_active slot if it is empty, that is
 910          * after the previous activity has been retired, or if the active
 911          * matches the current timeline.
 912          *
 913          * Note that we allow the timeline to be active simultaneously in
 914          * the rbtree and the last_active cache. We do this to avoid having
 915          * to search and replace the rbtree element for a new timeline, with
 916          * the cost being that we must be aware that the vma may be retired
 917          * twice for the same timeline (as the older rbtree element will be
 918          * retired before the new request added to last_active).
 919          */
 920         old = i915_gem_active_raw(&vma->last_active,
 921                                   &vma->vm->i915->drm.struct_mutex);
 922         if (!old || old->fence.context == idx)
 923                 goto out;
 924
 925         /* Move the currently active fence into the rbtree */
 926         idx = old->fence.context;
 927
 928         parent = NULL;
 929         p = &vma->active.rb_node;
 930         while (*p) {
 931                 parent = *p;
 932
 933                 active = rb_entry(parent, struct i915_vma_active, node);
 934                 if (active->timeline == idx)
 935                         goto replace;
 936
 937                 if (active->timeline < idx)
 938                         p = &parent->rb_right;
 939                 else
 940                         p = &parent->rb_left;
 941         }
 942
 943         active = kmalloc(sizeof(*active), GFP_KERNEL);
 944
 945         /* kmalloc may retire the vma->last_active request (thanks shrinker)! */
 946         if (unlikely(!i915_gem_active_raw(&vma->last_active,
 947                                           &vma->vm->i915->drm.struct_mutex))) {
 948                 kfree(active);
 949                 goto out;
 950         }
 951
 952         if (unlikely(!active))
 953                 return ERR_PTR(-ENOMEM);
 954
 955         init_request_active(&active->base, i915_vma_retire);
 956         active->vma = vma;
 957         active->timeline = idx;
 958
 959         rb_link_node(&active->node, parent, p);
 960         rb_insert_color(&active->node, &vma->active);
 961
 962 replace:
 963         /*
 964          * Overwrite the previous active slot in the rbtree with last_active,
 965          * leaving last_active zeroed. If the previous slot is still active,
 966          * we must be careful as we now only expect to receive one retire
 967          * callback not two, and so much undo the active counting for the
 968          * overwritten slot.
 969          */
 970         if (i915_gem_active_isset(&active->base)) {
 971                 /* Retire ourselves from the old rq->active_list */
 972                 __list_del_entry(&active->base.link);
 973                 vma->active_count--;
 974                 GEM_BUG_ON(!vma->active_count);
 975         }
 976         GEM_BUG_ON(list_empty(&vma->last_active.link));
 977         list_replace_init(&vma->last_active.link, &active->base.link);
 978         active->base.request = fetch_and_zero(&vma->last_active.request);
 979
 980 out:
 981         return &vma->last_active;
 982 }
 983
 984 int i915_vma_move_to_active(struct i915_vma *vma,
 985                             struct i915_request *rq,
 986                             unsigned int flags)
 987 {
 988         struct drm_i915_gem_object *obj = vma->obj;
 989         struct i915_gem_active *active;
 990
 991         lockdep_assert_held(&rq->i915->drm.struct_mutex);
 992         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 993
 994         active = active_instance(vma, rq->fence.context);
 995         if (IS_ERR(active))
 996                 return PTR_ERR(active);
 997
 998         /*
 999          * Add a reference if we're newly entering the active list.
1000          * The order in which we add operations to the retirement queue is
1001          * vital here: mark_active adds to the start of the callback list,
1002          * such that subsequent callbacks are called first. Therefore we
1003          * add the active reference first and queue for it to be dropped
1004          * *last*.
1005          */
1006         if (!i915_gem_active_isset(active) && !vma->active_count++) {
1007                 list_move_tail(&vma->vm_link, &vma->vm->active_list);
1008                 obj->active_count++;
1009         }
1010         i915_gem_active_set(active, rq);
1011         GEM_BUG_ON(!i915_vma_is_active(vma));
1012         GEM_BUG_ON(!obj->active_count);
1013
1014         obj->write_domain = 0;
1015         if (flags & EXEC_OBJECT_WRITE) {
1016                 obj->write_domain = I915_GEM_DOMAIN_RENDER;
1017
1018                 if (intel_fb_obj_invalidate(obj, ORIGIN_CS))
1019                         i915_gem_active_set(&obj->frontbuffer_write, rq);
1020
1021                 obj->read_domains = 0;
1022         }
1023         obj->read_domains |= I915_GEM_GPU_DOMAINS;
1024
1025         if (flags & EXEC_OBJECT_NEEDS_FENCE)
1026                 i915_gem_active_set(&vma->last_fence, rq);
1027
1028         export_fence(vma, rq, flags);
1029         return 0;
1030 }
1031
1032 int i915_vma_unbind(struct i915_vma *vma)
1033 {
1034         int ret;
1035
1036         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
1037
1038         /*
1039          * First wait upon any activity as retiring the request may
1040          * have side-effects such as unpinning or even unbinding this vma.
1041          */
1042         might_sleep();
1043         if (i915_vma_is_active(vma)) {
1044                 struct i915_vma_active *active, *n;
1045
1046                 /*
1047                  * When a closed VMA is retired, it is unbound - eek.
1048                  * In order to prevent it from being recursively closed,
1049                  * take a pin on the vma so that the second unbind is
1050                  * aborted.
1051                  *
1052                  * Even more scary is that the retire callback may free
1053                  * the object (last active vma). To prevent the explosion
1054                  * we defer the actual object free to a worker that can
1055                  * only proceed once it acquires the struct_mutex (which
1056                  * we currently hold, therefore it cannot free this object
1057                  * before we are finished).
1058                  */
1059                 __i915_vma_pin(vma);
1060
1061                 ret = i915_gem_active_retire(&vma->last_active,
1062                                              &vma->vm->i915->drm.struct_mutex);
1063                 if (ret)
1064                         goto unpin;
1065
1066                 rbtree_postorder_for_each_entry_safe(active, n,
1067                                                      &vma->active, node) {
1068                         ret = i915_gem_active_retire(&active->base,
1069                                                      &vma->vm->i915->drm.struct_mutex);
1070                         if (ret)
1071                                 goto unpin;
1072                 }
1073
1074                 ret = i915_gem_active_retire(&vma->last_fence,
1075                                              &vma->vm->i915->drm.struct_mutex);
1076 unpin:
1077                 __i915_vma_unpin(vma);
1078                 if (ret)
1079                         return ret;
1080         }
1081         GEM_BUG_ON(i915_vma_is_active(vma));
1082
1083         if (i915_vma_is_pinned(vma)) {
1084                 vma_print_allocator(vma, "is pinned");
1085                 return -EBUSY;
1086         }
1087
1088         if (!drm_mm_node_allocated(&vma->node))
1089                 return 0;
1090
1091         if (i915_vma_is_map_and_fenceable(vma)) {
1092                 /*
1093                  * Check that we have flushed all writes through the GGTT
1094                  * before the unbind, other due to non-strict nature of those
1095                  * indirect writes they may end up referencing the GGTT PTE
1096                  * after the unbind.
1097                  */
1098                 i915_vma_flush_writes(vma);
1099                 GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
1100
1101                 /* release the fence reg _after_ flushing */
1102                 ret = i915_vma_put_fence(vma);
1103                 if (ret)
1104                         return ret;
1105
1106                 /* Force a pagefault for domain tracking on next user access */
1107                 i915_vma_revoke_mmap(vma);
1108
1109                 __i915_vma_iounmap(vma);
1110                 vma->flags &= ~I915_VMA_CAN_FENCE;
1111         }
1112         GEM_BUG_ON(vma->fence);
1113         GEM_BUG_ON(i915_vma_has_userfault(vma));
1114
1115         if (likely(!vma->vm->closed)) {
1116                 trace_i915_vma_unbind(vma);
1117                 vma->ops->unbind_vma(vma);
1118         }
1119         vma->flags &= ~(I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
1120
1121         i915_vma_remove(vma);
1122
1123         return 0;
1124 }
1125
1126 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1127 #include "selftests/i915_vma.c"
1128 #endif