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 static void
  34 i915_vma_retire(struct i915_gem_active *active, struct i915_request *rq)
  35 {
  36         const unsigned int idx = rq->engine->id;
  37         struct i915_vma *vma =
  38                 container_of(active, struct i915_vma, last_read[idx]);
  39         struct drm_i915_gem_object *obj = vma->obj;
  40
  41         GEM_BUG_ON(!i915_vma_has_active_engine(vma, idx));
  42
  43         i915_vma_clear_active(vma, idx);
  44         if (i915_vma_is_active(vma))
  45                 return;
  46
  47         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
  48         list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
  49
  50         GEM_BUG_ON(!i915_gem_object_is_active(obj));
  51         if (--obj->active_count)
  52                 return;
  53
  54         /* Prune the shared fence arrays iff completely idle (inc. external) */
  55         if (reservation_object_trylock(obj->resv)) {
  56                 if (reservation_object_test_signaled_rcu(obj->resv, true))
  57                         reservation_object_add_excl_fence(obj->resv, NULL);
  58                 reservation_object_unlock(obj->resv);
  59         }
  60
  61         /* Bump our place on the bound list to keep it roughly in LRU order
  62          * so that we don't steal from recently used but inactive objects
  63          * (unless we are forced to ofc!)
  64          */
  65         spin_lock(&rq->i915->mm.obj_lock);
  66         if (obj->bind_count)
  67                 list_move_tail(&obj->mm.link, &rq->i915->mm.bound_list);
  68         spin_unlock(&rq->i915->mm.obj_lock);
  69
  70         obj->mm.dirty = true; /* be paranoid  */
  71
  72         if (i915_gem_object_has_active_reference(obj)) {
  73                 i915_gem_object_clear_active_reference(obj);
  74                 i915_gem_object_put(obj);
  75         }
  76 }
  77
  78 static struct i915_vma *
  79 vma_create(struct drm_i915_gem_object *obj,
  80            struct i915_address_space *vm,
  81            const struct i915_ggtt_view *view)
  82 {
  83         struct i915_vma *vma;
  84         struct rb_node *rb, **p;
  85         int i;
  86
  87         /* The aliasing_ppgtt should never be used directly! */
  88         GEM_BUG_ON(vm == &vm->i915->mm.aliasing_ppgtt->base);
  89
  90         vma = kmem_cache_zalloc(vm->i915->vmas, GFP_KERNEL);
  91         if (vma == NULL)
  92                 return ERR_PTR(-ENOMEM);
  93
  94         for (i = 0; i < ARRAY_SIZE(vma->last_read); i++)
  95                 init_request_active(&vma->last_read[i], i915_vma_retire);
  96         init_request_active(&vma->last_fence, NULL);
  97         vma->vm = vm;
  98         vma->obj = obj;
  99         vma->resv = obj->resv;
 100         vma->size = obj->base.size;
 101         vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
 102
 103         if (view && view->type != I915_GGTT_VIEW_NORMAL) {
 104                 vma->ggtt_view = *view;
 105                 if (view->type == I915_GGTT_VIEW_PARTIAL) {
 106                         GEM_BUG_ON(range_overflows_t(u64,
 107                                                      view->partial.offset,
 108                                                      view->partial.size,
 109                                                      obj->base.size >> PAGE_SHIFT));
 110                         vma->size = view->partial.size;
 111                         vma->size <<= PAGE_SHIFT;
 112                         GEM_BUG_ON(vma->size > obj->base.size);
 113                 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
 114                         vma->size = intel_rotation_info_size(&view->rotated);
 115                         vma->size <<= PAGE_SHIFT;
 116                 }
 117         }
 118
 119         if (unlikely(vma->size > vm->total))
 120                 goto err_vma;
 121
 122         GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
 123
 124         if (i915_is_ggtt(vm)) {
 125                 if (unlikely(overflows_type(vma->size, u32)))
 126                         goto err_vma;
 127
 128                 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
 129                                                       i915_gem_object_get_tiling(obj),
 130                                                       i915_gem_object_get_stride(obj));
 131                 if (unlikely(vma->fence_size < vma->size || /* overflow */
 132                              vma->fence_size > vm->total))
 133                         goto err_vma;
 134
 135                 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
 136
 137                 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
 138                                                                 i915_gem_object_get_tiling(obj),
 139                                                                 i915_gem_object_get_stride(obj));
 140                 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
 141
 142                 /*
 143                  * We put the GGTT vma at the start of the vma-list, followed
 144                  * by the ppGGTT vma. This allows us to break early when
 145                  * iterating over only the GGTT vma for an object, see
 146                  * for_each_ggtt_vma()
 147                  */
 148                 vma->flags |= I915_VMA_GGTT;
 149                 list_add(&vma->obj_link, &obj->vma_list);
 150         } else {
 151                 i915_ppgtt_get(i915_vm_to_ppgtt(vm));
 152                 list_add_tail(&vma->obj_link, &obj->vma_list);
 153         }
 154
 155         rb = NULL;
 156         p = &obj->vma_tree.rb_node;
 157         while (*p) {
 158                 struct i915_vma *pos;
 159
 160                 rb = *p;
 161                 pos = rb_entry(rb, struct i915_vma, obj_node);
 162                 if (i915_vma_compare(pos, vm, view) < 0)
 163                         p = &rb->rb_right;
 164                 else
 165                         p = &rb->rb_left;
 166         }
 167         rb_link_node(&vma->obj_node, rb, p);
 168         rb_insert_color(&vma->obj_node, &obj->vma_tree);
 169         list_add(&vma->vm_link, &vm->unbound_list);
 170
 171         return vma;
 172
 173 err_vma:
 174         kmem_cache_free(vm->i915->vmas, vma);
 175         return ERR_PTR(-E2BIG);
 176 }
 177
 178 static struct i915_vma *
 179 vma_lookup(struct drm_i915_gem_object *obj,
 180            struct i915_address_space *vm,
 181            const struct i915_ggtt_view *view)
 182 {
 183         struct rb_node *rb;
 184
 185         rb = obj->vma_tree.rb_node;
 186         while (rb) {
 187                 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
 188                 long cmp;
 189
 190                 cmp = i915_vma_compare(vma, vm, view);
 191                 if (cmp == 0)
 192                         return vma;
 193
 194                 if (cmp < 0)
 195                         rb = rb->rb_right;
 196                 else
 197                         rb = rb->rb_left;
 198         }
 199
 200         return NULL;
 201 }
 202
 203 /**
 204  * i915_vma_instance - return the singleton instance of the VMA
 205  * @obj: parent &struct drm_i915_gem_object to be mapped
 206  * @vm: address space in which the mapping is located
 207  * @view: additional mapping requirements
 208  *
 209  * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
 210  * the same @view characteristics. If a match is not found, one is created.
 211  * Once created, the VMA is kept until either the object is freed, or the
 212  * address space is closed.
 213  *
 214  * Must be called with struct_mutex held.
 215  *
 216  * Returns the vma, or an error pointer.
 217  */
 218 struct i915_vma *
 219 i915_vma_instance(struct drm_i915_gem_object *obj,
 220                   struct i915_address_space *vm,
 221                   const struct i915_ggtt_view *view)
 222 {
 223         struct i915_vma *vma;
 224
 225         lockdep_assert_held(&obj->base.dev->struct_mutex);
 226         GEM_BUG_ON(view && !i915_is_ggtt(vm));
 227         GEM_BUG_ON(vm->closed);
 228
 229         vma = vma_lookup(obj, vm, view);
 230         if (!vma)
 231                 vma = vma_create(obj, vm, view);
 232
 233         GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
 234         GEM_BUG_ON(!IS_ERR(vma) && vma_lookup(obj, vm, view) != vma);
 235         return vma;
 236 }
 237
 238 /**
 239  * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
 240  * @vma: VMA to map
 241  * @cache_level: mapping cache level
 242  * @flags: flags like global or local mapping
 243  *
 244  * DMA addresses are taken from the scatter-gather table of this object (or of
 245  * this VMA in case of non-default GGTT views) and PTE entries set up.
 246  * Note that DMA addresses are also the only part of the SG table we care about.
 247  */
 248 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
 249                   u32 flags)
 250 {
 251         u32 bind_flags;
 252         u32 vma_flags;
 253         int ret;
 254
 255         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 256         GEM_BUG_ON(vma->size > vma->node.size);
 257
 258         if (GEM_WARN_ON(range_overflows(vma->node.start,
 259                                         vma->node.size,
 260                                         vma->vm->total)))
 261                 return -ENODEV;
 262
 263         if (GEM_WARN_ON(!flags))
 264                 return -EINVAL;
 265
 266         bind_flags = 0;
 267         if (flags & PIN_GLOBAL)
 268                 bind_flags |= I915_VMA_GLOBAL_BIND;
 269         if (flags & PIN_USER)
 270                 bind_flags |= I915_VMA_LOCAL_BIND;
 271
 272         vma_flags = vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
 273         if (flags & PIN_UPDATE)
 274                 bind_flags |= vma_flags;
 275         else
 276                 bind_flags &= ~vma_flags;
 277         if (bind_flags == 0)
 278                 return 0;
 279
 280         GEM_BUG_ON(!vma->pages);
 281
 282         trace_i915_vma_bind(vma, bind_flags);
 283         ret = vma->vm->bind_vma(vma, cache_level, bind_flags);
 284         if (ret)
 285                 return ret;
 286
 287         vma->flags |= bind_flags;
 288         return 0;
 289 }
 290
 291 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
 292 {
 293         void __iomem *ptr;
 294         int err;
 295
 296         /* Access through the GTT requires the device to be awake. */
 297         assert_rpm_wakelock_held(vma->vm->i915);
 298
 299         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 300         if (WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
 301                 err = -ENODEV;
 302                 goto err;
 303         }
 304
 305         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 306         GEM_BUG_ON((vma->flags & I915_VMA_GLOBAL_BIND) == 0);
 307
 308         ptr = vma->iomap;
 309         if (ptr == NULL) {
 310                 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
 311                                         vma->node.start,
 312                                         vma->node.size);
 313                 if (ptr == NULL) {
 314                         err = -ENOMEM;
 315                         goto err;
 316                 }
 317
 318                 vma->iomap = ptr;
 319         }
 320
 321         __i915_vma_pin(vma);
 322
 323         err = i915_vma_pin_fence(vma);
 324         if (err)
 325                 goto err_unpin;
 326
 327         i915_vma_set_ggtt_write(vma);
 328         return ptr;
 329
 330 err_unpin:
 331         __i915_vma_unpin(vma);
 332 err:
 333         return IO_ERR_PTR(err);
 334 }
 335
 336 void i915_vma_flush_writes(struct i915_vma *vma)
 337 {
 338         if (!i915_vma_has_ggtt_write(vma))
 339                 return;
 340
 341         i915_gem_flush_ggtt_writes(vma->vm->i915);
 342
 343         i915_vma_unset_ggtt_write(vma);
 344 }
 345
 346 void i915_vma_unpin_iomap(struct i915_vma *vma)
 347 {
 348         lockdep_assert_held(&vma->obj->base.dev->struct_mutex);
 349
 350         GEM_BUG_ON(vma->iomap == NULL);
 351
 352         i915_vma_flush_writes(vma);
 353
 354         i915_vma_unpin_fence(vma);
 355         i915_vma_unpin(vma);
 356 }
 357
 358 void i915_vma_unpin_and_release(struct i915_vma **p_vma)
 359 {
 360         struct i915_vma *vma;
 361         struct drm_i915_gem_object *obj;
 362
 363         vma = fetch_and_zero(p_vma);
 364         if (!vma)
 365                 return;
 366
 367         obj = vma->obj;
 368
 369         i915_vma_unpin(vma);
 370         i915_vma_close(vma);
 371
 372         __i915_gem_object_release_unless_active(obj);
 373 }
 374
 375 bool i915_vma_misplaced(const struct i915_vma *vma,
 376                         u64 size, u64 alignment, u64 flags)
 377 {
 378         if (!drm_mm_node_allocated(&vma->node))
 379                 return false;
 380
 381         if (vma->node.size < size)
 382                 return true;
 383
 384         GEM_BUG_ON(alignment && !is_power_of_2(alignment));
 385         if (alignment && !IS_ALIGNED(vma->node.start, alignment))
 386                 return true;
 387
 388         if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
 389                 return true;
 390
 391         if (flags & PIN_OFFSET_BIAS &&
 392             vma->node.start < (flags & PIN_OFFSET_MASK))
 393                 return true;
 394
 395         if (flags & PIN_OFFSET_FIXED &&
 396             vma->node.start != (flags & PIN_OFFSET_MASK))
 397                 return true;
 398
 399         return false;
 400 }
 401
 402 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
 403 {
 404         bool mappable, fenceable;
 405
 406         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 407         GEM_BUG_ON(!vma->fence_size);
 408
 409         /*
 410          * Explicitly disable for rotated VMA since the display does not
 411          * need the fence and the VMA is not accessible to other users.
 412          */
 413         if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED)
 414                 return;
 415
 416         fenceable = (vma->node.size >= vma->fence_size &&
 417                      IS_ALIGNED(vma->node.start, vma->fence_alignment));
 418
 419         mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
 420
 421         if (mappable && fenceable)
 422                 vma->flags |= I915_VMA_CAN_FENCE;
 423         else
 424                 vma->flags &= ~I915_VMA_CAN_FENCE;
 425 }
 426
 427 static bool color_differs(struct drm_mm_node *node, unsigned long color)
 428 {
 429         return node->allocated && node->color != color;
 430 }
 431
 432 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long cache_level)
 433 {
 434         struct drm_mm_node *node = &vma->node;
 435         struct drm_mm_node *other;
 436
 437         /*
 438          * On some machines we have to be careful when putting differing types
 439          * of snoopable memory together to avoid the prefetcher crossing memory
 440          * domains and dying. During vm initialisation, we decide whether or not
 441          * these constraints apply and set the drm_mm.color_adjust
 442          * appropriately.
 443          */
 444         if (vma->vm->mm.color_adjust == NULL)
 445                 return true;
 446
 447         /* Only valid to be called on an already inserted vma */
 448         GEM_BUG_ON(!drm_mm_node_allocated(node));
 449         GEM_BUG_ON(list_empty(&node->node_list));
 450
 451         other = list_prev_entry(node, node_list);
 452         if (color_differs(other, cache_level) && !drm_mm_hole_follows(other))
 453                 return false;
 454
 455         other = list_next_entry(node, node_list);
 456         if (color_differs(other, cache_level) && !drm_mm_hole_follows(node))
 457                 return false;
 458
 459         return true;
 460 }
 461
 462 /**
 463  * i915_vma_insert - finds a slot for the vma in its address space
 464  * @vma: the vma
 465  * @size: requested size in bytes (can be larger than the VMA)
 466  * @alignment: required alignment
 467  * @flags: mask of PIN_* flags to use
 468  *
 469  * First we try to allocate some free space that meets the requirements for
 470  * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
 471  * preferrably the oldest idle entry to make room for the new VMA.
 472  *
 473  * Returns:
 474  * 0 on success, negative error code otherwise.
 475  */
 476 static int
 477 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 478 {
 479         struct drm_i915_private *dev_priv = vma->vm->i915;
 480         struct drm_i915_gem_object *obj = vma->obj;
 481         u64 start, end;
 482         int ret;
 483
 484         GEM_BUG_ON(i915_vma_is_closed(vma));
 485         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 486         GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
 487
 488         size = max(size, vma->size);
 489         alignment = max(alignment, vma->display_alignment);
 490         if (flags & PIN_MAPPABLE) {
 491                 size = max_t(typeof(size), size, vma->fence_size);
 492                 alignment = max_t(typeof(alignment),
 493                                   alignment, vma->fence_alignment);
 494         }
 495
 496         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
 497         GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
 498         GEM_BUG_ON(!is_power_of_2(alignment));
 499
 500         start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
 501         GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
 502
 503         end = vma->vm->total;
 504         if (flags & PIN_MAPPABLE)
 505                 end = min_t(u64, end, dev_priv->ggtt.mappable_end);
 506         if (flags & PIN_ZONE_4G)
 507                 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
 508         GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
 509
 510         /* If binding the object/GGTT view requires more space than the entire
 511          * aperture has, reject it early before evicting everything in a vain
 512          * attempt to find space.
 513          */
 514         if (size > end) {
 515                 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu [object=%zd] > %s aperture=%llu\n",
 516                           size, obj->base.size,
 517                           flags & PIN_MAPPABLE ? "mappable" : "total",
 518                           end);
 519                 return -ENOSPC;
 520         }
 521
 522         ret = i915_gem_object_pin_pages(obj);
 523         if (ret)
 524                 return ret;
 525
 526         GEM_BUG_ON(vma->pages);
 527
 528         ret = vma->vm->set_pages(vma);
 529         if (ret)
 530                 goto err_unpin;
 531
 532         if (flags & PIN_OFFSET_FIXED) {
 533                 u64 offset = flags & PIN_OFFSET_MASK;
 534                 if (!IS_ALIGNED(offset, alignment) ||
 535                     range_overflows(offset, size, end)) {
 536                         ret = -EINVAL;
 537                         goto err_clear;
 538                 }
 539
 540                 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
 541                                            size, offset, obj->cache_level,
 542                                            flags);
 543                 if (ret)
 544                         goto err_clear;
 545         } else {
 546                 /*
 547                  * We only support huge gtt pages through the 48b PPGTT,
 548                  * however we also don't want to force any alignment for
 549                  * objects which need to be tightly packed into the low 32bits.
 550                  *
 551                  * Note that we assume that GGTT are limited to 4GiB for the
 552                  * forseeable future. See also i915_ggtt_offset().
 553                  */
 554                 if (upper_32_bits(end - 1) &&
 555                     vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
 556                         /*
 557                          * We can't mix 64K and 4K PTEs in the same page-table
 558                          * (2M block), and so to avoid the ugliness and
 559                          * complexity of coloring we opt for just aligning 64K
 560                          * objects to 2M.
 561                          */
 562                         u64 page_alignment =
 563                                 rounddown_pow_of_two(vma->page_sizes.sg |
 564                                                      I915_GTT_PAGE_SIZE_2M);
 565
 566                         /*
 567                          * Check we don't expand for the limited Global GTT
 568                          * (mappable aperture is even more precious!). This
 569                          * also checks that we exclude the aliasing-ppgtt.
 570                          */
 571                         GEM_BUG_ON(i915_vma_is_ggtt(vma));
 572
 573                         alignment = max(alignment, page_alignment);
 574
 575                         if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
 576                                 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
 577                 }
 578
 579                 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
 580                                           size, alignment, obj->cache_level,
 581                                           start, end, flags);
 582                 if (ret)
 583                         goto err_clear;
 584
 585                 GEM_BUG_ON(vma->node.start < start);
 586                 GEM_BUG_ON(vma->node.start + vma->node.size > end);
 587         }
 588         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 589         GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, obj->cache_level));
 590
 591         list_move_tail(&vma->vm_link, &vma->vm->inactive_list);
 592
 593         spin_lock(&dev_priv->mm.obj_lock);
 594         list_move_tail(&obj->mm.link, &dev_priv->mm.bound_list);
 595         obj->bind_count++;
 596         spin_unlock(&dev_priv->mm.obj_lock);
 597
 598         GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < obj->bind_count);
 599
 600         return 0;
 601
 602 err_clear:
 603         vma->vm->clear_pages(vma);
 604 err_unpin:
 605         i915_gem_object_unpin_pages(obj);
 606         return ret;
 607 }
 608
 609 static void
 610 i915_vma_remove(struct i915_vma *vma)
 611 {
 612         struct drm_i915_private *i915 = vma->vm->i915;
 613         struct drm_i915_gem_object *obj = vma->obj;
 614
 615         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 616         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 617
 618         vma->vm->clear_pages(vma);
 619
 620         drm_mm_remove_node(&vma->node);
 621         list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
 622
 623         /* Since the unbound list is global, only move to that list if
 624          * no more VMAs exist.
 625          */
 626         spin_lock(&i915->mm.obj_lock);
 627         if (--obj->bind_count == 0)
 628                 list_move_tail(&obj->mm.link, &i915->mm.unbound_list);
 629         spin_unlock(&i915->mm.obj_lock);
 630
 631         /* And finally now the object is completely decoupled from this vma,
 632          * we can drop its hold on the backing storage and allow it to be
 633          * reaped by the shrinker.
 634          */
 635         i915_gem_object_unpin_pages(obj);
 636         GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < obj->bind_count);
 637 }
 638
 639 int __i915_vma_do_pin(struct i915_vma *vma,
 640                       u64 size, u64 alignment, u64 flags)
 641 {
 642         const unsigned int bound = vma->flags;
 643         int ret;
 644
 645         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 646         GEM_BUG_ON((flags & (PIN_GLOBAL | PIN_USER)) == 0);
 647         GEM_BUG_ON((flags & PIN_GLOBAL) && !i915_vma_is_ggtt(vma));
 648
 649         if (WARN_ON(bound & I915_VMA_PIN_OVERFLOW)) {
 650                 ret = -EBUSY;
 651                 goto err_unpin;
 652         }
 653
 654         if ((bound & I915_VMA_BIND_MASK) == 0) {
 655                 ret = i915_vma_insert(vma, size, alignment, flags);
 656                 if (ret)
 657                         goto err_unpin;
 658         }
 659         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 660
 661         ret = i915_vma_bind(vma, vma->obj->cache_level, flags);
 662         if (ret)
 663                 goto err_remove;
 664
 665         GEM_BUG_ON((vma->flags & I915_VMA_BIND_MASK) == 0);
 666
 667         if ((bound ^ vma->flags) & I915_VMA_GLOBAL_BIND)
 668                 __i915_vma_set_map_and_fenceable(vma);
 669
 670         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
 671         return 0;
 672
 673 err_remove:
 674         if ((bound & I915_VMA_BIND_MASK) == 0) {
 675                 i915_vma_remove(vma);
 676                 GEM_BUG_ON(vma->pages);
 677                 GEM_BUG_ON(vma->flags & I915_VMA_BIND_MASK);
 678         }
 679 err_unpin:
 680         __i915_vma_unpin(vma);
 681         return ret;
 682 }
 683
 684 void i915_vma_close(struct i915_vma *vma)
 685 {
 686         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 687
 688         GEM_BUG_ON(i915_vma_is_closed(vma));
 689         vma->flags |= I915_VMA_CLOSED;
 690
 691         /*
 692          * We defer actually closing, unbinding and destroying the VMA until
 693          * the next idle point, or if the object is freed in the meantime. By
 694          * postponing the unbind, we allow for it to be resurrected by the
 695          * client, avoiding the work required to rebind the VMA. This is
 696          * advantageous for DRI, where the client/server pass objects
 697          * between themselves, temporarily opening a local VMA to the
 698          * object, and then closing it again. The same object is then reused
 699          * on the next frame (or two, depending on the depth of the swap queue)
 700          * causing us to rebind the VMA once more. This ends up being a lot
 701          * of wasted work for the steady state.
 702          */
 703         list_add_tail(&vma->closed_link, &vma->vm->i915->gt.closed_vma);
 704 }
 705
 706 void i915_vma_reopen(struct i915_vma *vma)
 707 {
 708         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 709
 710         if (vma->flags & I915_VMA_CLOSED) {
 711                 vma->flags &= ~I915_VMA_CLOSED;
 712                 list_del(&vma->closed_link);
 713         }
 714 }
 715
 716 static void __i915_vma_destroy(struct i915_vma *vma)
 717 {
 718         int i;
 719
 720         GEM_BUG_ON(vma->node.allocated);
 721         GEM_BUG_ON(vma->fence);
 722
 723         for (i = 0; i < ARRAY_SIZE(vma->last_read); i++)
 724                 GEM_BUG_ON(i915_gem_active_isset(&vma->last_read[i]));
 725         GEM_BUG_ON(i915_gem_active_isset(&vma->last_fence));
 726
 727         list_del(&vma->obj_link);
 728         list_del(&vma->vm_link);
 729         rb_erase(&vma->obj_node, &vma->obj->vma_tree);
 730
 731         if (!i915_vma_is_ggtt(vma))
 732                 i915_ppgtt_put(i915_vm_to_ppgtt(vma->vm));
 733
 734         kmem_cache_free(to_i915(vma->obj->base.dev)->vmas, vma);
 735 }
 736
 737 void i915_vma_destroy(struct i915_vma *vma)
 738 {
 739         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 740
 741         GEM_BUG_ON(i915_vma_is_active(vma));
 742         GEM_BUG_ON(i915_vma_is_pinned(vma));
 743
 744         if (i915_vma_is_closed(vma))
 745                 list_del(&vma->closed_link);
 746
 747         WARN_ON(i915_vma_unbind(vma));
 748         __i915_vma_destroy(vma);
 749 }
 750
 751 void i915_vma_parked(struct drm_i915_private *i915)
 752 {
 753         struct i915_vma *vma, *next;
 754
 755         list_for_each_entry_safe(vma, next, &i915->gt.closed_vma, closed_link) {
 756                 GEM_BUG_ON(!i915_vma_is_closed(vma));
 757                 i915_vma_destroy(vma);
 758         }
 759
 760         GEM_BUG_ON(!list_empty(&i915->gt.closed_vma));
 761 }
 762
 763 static void __i915_vma_iounmap(struct i915_vma *vma)
 764 {
 765         GEM_BUG_ON(i915_vma_is_pinned(vma));
 766
 767         if (vma->iomap == NULL)
 768                 return;
 769
 770         io_mapping_unmap(vma->iomap);
 771         vma->iomap = NULL;
 772 }
 773
 774 void i915_vma_revoke_mmap(struct i915_vma *vma)
 775 {
 776         struct drm_vma_offset_node *node = &vma->obj->base.vma_node;
 777         u64 vma_offset;
 778
 779         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 780
 781         if (!i915_vma_has_userfault(vma))
 782                 return;
 783
 784         GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
 785         GEM_BUG_ON(!vma->obj->userfault_count);
 786
 787         vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
 788         unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
 789                             drm_vma_node_offset_addr(node) + vma_offset,
 790                             vma->size,
 791                             1);
 792
 793         i915_vma_unset_userfault(vma);
 794         if (!--vma->obj->userfault_count)
 795                 list_del(&vma->obj->userfault_link);
 796 }
 797
 798 int i915_vma_unbind(struct i915_vma *vma)
 799 {
 800         struct drm_i915_gem_object *obj = vma->obj;
 801         unsigned long active;
 802         int ret;
 803
 804         lockdep_assert_held(&obj->base.dev->struct_mutex);
 805
 806         /* First wait upon any activity as retiring the request may
 807          * have side-effects such as unpinning or even unbinding this vma.
 808          */
 809         might_sleep();
 810         active = i915_vma_get_active(vma);
 811         if (active) {
 812                 int idx;
 813
 814                 /* When a closed VMA is retired, it is unbound - eek.
 815                  * In order to prevent it from being recursively closed,
 816                  * take a pin on the vma so that the second unbind is
 817                  * aborted.
 818                  *
 819                  * Even more scary is that the retire callback may free
 820                  * the object (last active vma). To prevent the explosion
 821                  * we defer the actual object free to a worker that can
 822                  * only proceed once it acquires the struct_mutex (which
 823                  * we currently hold, therefore it cannot free this object
 824                  * before we are finished).
 825                  */
 826                 __i915_vma_pin(vma);
 827
 828                 for_each_active(active, idx) {
 829                         ret = i915_gem_active_retire(&vma->last_read[idx],
 830                                                      &vma->vm->i915->drm.struct_mutex);
 831                         if (ret)
 832                                 break;
 833                 }
 834
 835                 if (!ret) {
 836                         ret = i915_gem_active_retire(&vma->last_fence,
 837                                                      &vma->vm->i915->drm.struct_mutex);
 838                 }
 839
 840                 __i915_vma_unpin(vma);
 841                 if (ret)
 842                         return ret;
 843         }
 844         GEM_BUG_ON(i915_vma_is_active(vma));
 845
 846         if (i915_vma_is_pinned(vma))
 847                 return -EBUSY;
 848
 849         if (!drm_mm_node_allocated(&vma->node))
 850                 return 0;
 851
 852         GEM_BUG_ON(obj->bind_count == 0);
 853         GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
 854
 855         if (i915_vma_is_map_and_fenceable(vma)) {
 856                 /*
 857                  * Check that we have flushed all writes through the GGTT
 858                  * before the unbind, other due to non-strict nature of those
 859                  * indirect writes they may end up referencing the GGTT PTE
 860                  * after the unbind.
 861                  */
 862                 i915_vma_flush_writes(vma);
 863                 GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
 864
 865                 /* release the fence reg _after_ flushing */
 866                 ret = i915_vma_put_fence(vma);
 867                 if (ret)
 868                         return ret;
 869
 870                 /* Force a pagefault for domain tracking on next user access */
 871                 i915_vma_revoke_mmap(vma);
 872
 873                 __i915_vma_iounmap(vma);
 874                 vma->flags &= ~I915_VMA_CAN_FENCE;
 875         }
 876         GEM_BUG_ON(vma->fence);
 877         GEM_BUG_ON(i915_vma_has_userfault(vma));
 878
 879         if (likely(!vma->vm->closed)) {
 880                 trace_i915_vma_unbind(vma);
 881                 vma->vm->unbind_vma(vma);
 882         }
 883         vma->flags &= ~(I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
 884
 885         i915_vma_remove(vma);
 886
 887         return 0;
 888 }
 889
 890 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
 891 #include "selftests/i915_vma.c"
 892 #endif