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 <linux/sched/mm.h>
  26 #include <drm/drm_gem.h>
  27
  28 #include "display/intel_frontbuffer.h"
  29
  30 #include "gt/intel_engine.h"
  31 #include "gt/intel_engine_heartbeat.h"
  32 #include "gt/intel_gt.h"
  33 #include "gt/intel_gt_requests.h"
  34
  35 #include "i915_drv.h"
  36 #include "i915_globals.h"
  37 #include "i915_sw_fence_work.h"
  38 #include "i915_trace.h"
  39 #include "i915_vma.h"
  40
  41 static struct i915_global_vma {
  42         struct i915_global base;
  43         struct kmem_cache *slab_vmas;
  44 } global;
  45
  46 struct i915_vma *i915_vma_alloc(void)
  47 {
  48         return kmem_cache_zalloc(global.slab_vmas, GFP_KERNEL);
  49 }
  50
  51 void i915_vma_free(struct i915_vma *vma)
  52 {
  53         return kmem_cache_free(global.slab_vmas, vma);
  54 }
  55
  56 #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM)
  57
  58 #include <linux/stackdepot.h>
  59
  60 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
  61 {
  62         unsigned long *entries;
  63         unsigned int nr_entries;
  64         char buf[512];
  65
  66         if (!vma->node.stack) {
  67                 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",
  68                                  vma->node.start, vma->node.size, reason);
  69                 return;
  70         }
  71
  72         nr_entries = stack_depot_fetch(vma->node.stack, &entries);
  73         stack_trace_snprint(buf, sizeof(buf), entries, nr_entries, 0);
  74         DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",
  75                          vma->node.start, vma->node.size, reason, buf);
  76 }
  77
  78 #else
  79
  80 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
  81 {
  82 }
  83
  84 #endif
  85
  86 static inline struct i915_vma *active_to_vma(struct i915_active *ref)
  87 {
  88         return container_of(ref, typeof(struct i915_vma), active);
  89 }
  90
  91 static int __i915_vma_active(struct i915_active *ref)
  92 {
  93         return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT;
  94 }
  95
  96 __i915_active_call
  97 static void __i915_vma_retire(struct i915_active *ref)
  98 {
  99         i915_vma_put(active_to_vma(ref));
 100 }
 101
 102 static struct i915_vma *
 103 vma_create(struct drm_i915_gem_object *obj,
 104            struct i915_address_space *vm,
 105            const struct i915_ggtt_view *view)
 106 {
 107         struct i915_vma *vma;
 108         struct rb_node *rb, **p;
 109
 110         /* The aliasing_ppgtt should never be used directly! */
 111         GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm);
 112
 113         vma = i915_vma_alloc();
 114         if (vma == NULL)
 115                 return ERR_PTR(-ENOMEM);
 116
 117         kref_init(&vma->ref);
 118         mutex_init(&vma->pages_mutex);
 119         vma->vm = i915_vm_get(vm);
 120         vma->ops = &vm->vma_ops;
 121         vma->obj = obj;
 122         vma->resv = obj->base.resv;
 123         vma->size = obj->base.size;
 124         vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
 125
 126         i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire);
 127
 128         /* Declare ourselves safe for use inside shrinkers */
 129         if (IS_ENABLED(CONFIG_LOCKDEP)) {
 130                 fs_reclaim_acquire(GFP_KERNEL);
 131                 might_lock(&vma->active.mutex);
 132                 fs_reclaim_release(GFP_KERNEL);
 133         }
 134
 135         INIT_LIST_HEAD(&vma->closed_link);
 136
 137         if (view && view->type != I915_GGTT_VIEW_NORMAL) {
 138                 vma->ggtt_view = *view;
 139                 if (view->type == I915_GGTT_VIEW_PARTIAL) {
 140                         GEM_BUG_ON(range_overflows_t(u64,
 141                                                      view->partial.offset,
 142                                                      view->partial.size,
 143                                                      obj->base.size >> PAGE_SHIFT));
 144                         vma->size = view->partial.size;
 145                         vma->size <<= PAGE_SHIFT;
 146                         GEM_BUG_ON(vma->size > obj->base.size);
 147                 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
 148                         vma->size = intel_rotation_info_size(&view->rotated);
 149                         vma->size <<= PAGE_SHIFT;
 150                 } else if (view->type == I915_GGTT_VIEW_REMAPPED) {
 151                         vma->size = intel_remapped_info_size(&view->remapped);
 152                         vma->size <<= PAGE_SHIFT;
 153                 }
 154         }
 155
 156         if (unlikely(vma->size > vm->total))
 157                 goto err_vma;
 158
 159         GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
 160
 161         if (i915_is_ggtt(vm)) {
 162                 if (unlikely(overflows_type(vma->size, u32)))
 163                         goto err_vma;
 164
 165                 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
 166                                                       i915_gem_object_get_tiling(obj),
 167                                                       i915_gem_object_get_stride(obj));
 168                 if (unlikely(vma->fence_size < vma->size || /* overflow */
 169                              vma->fence_size > vm->total))
 170                         goto err_vma;
 171
 172                 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
 173
 174                 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
 175                                                                 i915_gem_object_get_tiling(obj),
 176                                                                 i915_gem_object_get_stride(obj));
 177                 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
 178
 179                 __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma));
 180         }
 181
 182         spin_lock(&obj->vma.lock);
 183
 184         rb = NULL;
 185         p = &obj->vma.tree.rb_node;
 186         while (*p) {
 187                 struct i915_vma *pos;
 188                 long cmp;
 189
 190                 rb = *p;
 191                 pos = rb_entry(rb, struct i915_vma, obj_node);
 192
 193                 /*
 194                  * If the view already exists in the tree, another thread
 195                  * already created a matching vma, so return the older instance
 196                  * and dispose of ours.
 197                  */
 198                 cmp = i915_vma_compare(pos, vm, view);
 199                 if (cmp == 0) {
 200                         spin_unlock(&obj->vma.lock);
 201                         i915_vma_free(vma);
 202                         return pos;
 203                 }
 204
 205                 if (cmp < 0)
 206                         p = &rb->rb_right;
 207                 else
 208                         p = &rb->rb_left;
 209         }
 210         rb_link_node(&vma->obj_node, rb, p);
 211         rb_insert_color(&vma->obj_node, &obj->vma.tree);
 212
 213         if (i915_vma_is_ggtt(vma))
 214                 /*
 215                  * We put the GGTT vma at the start of the vma-list, followed
 216                  * by the ppGGTT vma. This allows us to break early when
 217                  * iterating over only the GGTT vma for an object, see
 218                  * for_each_ggtt_vma()
 219                  */
 220                 list_add(&vma->obj_link, &obj->vma.list);
 221         else
 222                 list_add_tail(&vma->obj_link, &obj->vma.list);
 223
 224         spin_unlock(&obj->vma.lock);
 225
 226         return vma;
 227
 228 err_vma:
 229         i915_vma_free(vma);
 230         return ERR_PTR(-E2BIG);
 231 }
 232
 233 static struct i915_vma *
 234 vma_lookup(struct drm_i915_gem_object *obj,
 235            struct i915_address_space *vm,
 236            const struct i915_ggtt_view *view)
 237 {
 238         struct rb_node *rb;
 239
 240         rb = obj->vma.tree.rb_node;
 241         while (rb) {
 242                 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
 243                 long cmp;
 244
 245                 cmp = i915_vma_compare(vma, vm, view);
 246                 if (cmp == 0)
 247                         return vma;
 248
 249                 if (cmp < 0)
 250                         rb = rb->rb_right;
 251                 else
 252                         rb = rb->rb_left;
 253         }
 254
 255         return NULL;
 256 }
 257
 258 /**
 259  * i915_vma_instance - return the singleton instance of the VMA
 260  * @obj: parent &struct drm_i915_gem_object to be mapped
 261  * @vm: address space in which the mapping is located
 262  * @view: additional mapping requirements
 263  *
 264  * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
 265  * the same @view characteristics. If a match is not found, one is created.
 266  * Once created, the VMA is kept until either the object is freed, or the
 267  * address space is closed.
 268  *
 269  * Returns the vma, or an error pointer.
 270  */
 271 struct i915_vma *
 272 i915_vma_instance(struct drm_i915_gem_object *obj,
 273                   struct i915_address_space *vm,
 274                   const struct i915_ggtt_view *view)
 275 {
 276         struct i915_vma *vma;
 277
 278         GEM_BUG_ON(view && !i915_is_ggtt(vm));
 279         GEM_BUG_ON(!atomic_read(&vm->open));
 280
 281         spin_lock(&obj->vma.lock);
 282         vma = vma_lookup(obj, vm, view);
 283         spin_unlock(&obj->vma.lock);
 284
 285         /* vma_create() will resolve the race if another creates the vma */
 286         if (unlikely(!vma))
 287                 vma = vma_create(obj, vm, view);
 288
 289         GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
 290         return vma;
 291 }
 292
 293 struct i915_vma_work {
 294         struct dma_fence_work base;
 295         struct i915_vma *vma;
 296         struct drm_i915_gem_object *pinned;
 297         struct i915_sw_dma_fence_cb cb;
 298         enum i915_cache_level cache_level;
 299         unsigned int flags;
 300 };
 301
 302 static int __vma_bind(struct dma_fence_work *work)
 303 {
 304         struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
 305         struct i915_vma *vma = vw->vma;
 306         int err;
 307
 308         err = vma->ops->bind_vma(vma, vw->cache_level, vw->flags);
 309         if (err)
 310                 atomic_or(I915_VMA_ERROR, &vma->flags);
 311
 312         return err;
 313 }
 314
 315 static void __vma_release(struct dma_fence_work *work)
 316 {
 317         struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
 318
 319         if (vw->pinned)
 320                 __i915_gem_object_unpin_pages(vw->pinned);
 321 }
 322
 323 static const struct dma_fence_work_ops bind_ops = {
 324         .name = "bind",
 325         .work = __vma_bind,
 326         .release = __vma_release,
 327 };
 328
 329 struct i915_vma_work *i915_vma_work(void)
 330 {
 331         struct i915_vma_work *vw;
 332
 333         vw = kzalloc(sizeof(*vw), GFP_KERNEL);
 334         if (!vw)
 335                 return NULL;
 336
 337         dma_fence_work_init(&vw->base, &bind_ops);
 338         vw->base.dma.error = -EAGAIN; /* disable the worker by default */
 339
 340         return vw;
 341 }
 342
 343 int i915_vma_wait_for_bind(struct i915_vma *vma)
 344 {
 345         int err = 0;
 346
 347         if (rcu_access_pointer(vma->active.excl.fence)) {
 348                 struct dma_fence *fence;
 349
 350                 rcu_read_lock();
 351                 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence);
 352                 rcu_read_unlock();
 353                 if (fence) {
 354                         err = dma_fence_wait(fence, MAX_SCHEDULE_TIMEOUT);
 355                         dma_fence_put(fence);
 356                 }
 357         }
 358
 359         return err;
 360 }
 361
 362 /**
 363  * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
 364  * @vma: VMA to map
 365  * @cache_level: mapping cache level
 366  * @flags: flags like global or local mapping
 367  * @work: preallocated worker for allocating and binding the PTE
 368  *
 369  * DMA addresses are taken from the scatter-gather table of this object (or of
 370  * this VMA in case of non-default GGTT views) and PTE entries set up.
 371  * Note that DMA addresses are also the only part of the SG table we care about.
 372  */
 373 int i915_vma_bind(struct i915_vma *vma,
 374                   enum i915_cache_level cache_level,
 375                   u32 flags,
 376                   struct i915_vma_work *work)
 377 {
 378         u32 bind_flags;
 379         u32 vma_flags;
 380         int ret;
 381
 382         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 383         GEM_BUG_ON(vma->size > vma->node.size);
 384
 385         if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
 386                                               vma->node.size,
 387                                               vma->vm->total)))
 388                 return -ENODEV;
 389
 390         if (GEM_DEBUG_WARN_ON(!flags))
 391                 return -EINVAL;
 392
 393         bind_flags = flags;
 394         bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
 395
 396         vma_flags = atomic_read(&vma->flags);
 397         vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
 398         if (flags & PIN_UPDATE)
 399                 bind_flags |= vma_flags;
 400         else
 401                 bind_flags &= ~vma_flags;
 402         if (bind_flags == 0)
 403                 return 0;
 404
 405         GEM_BUG_ON(!vma->pages);
 406
 407         trace_i915_vma_bind(vma, bind_flags);
 408         if (work && (bind_flags & ~vma_flags) & vma->vm->bind_async_flags) {
 409                 struct dma_fence *prev;
 410
 411                 work->vma = vma;
 412                 work->cache_level = cache_level;
 413                 work->flags = bind_flags | I915_VMA_ALLOC;
 414
 415                 /*
 416                  * Note we only want to chain up to the migration fence on
 417                  * the pages (not the object itself). As we don't track that,
 418                  * yet, we have to use the exclusive fence instead.
 419                  *
 420                  * Also note that we do not want to track the async vma as
 421                  * part of the obj->resv->excl_fence as it only affects
 422                  * execution and not content or object's backing store lifetime.
 423                  */
 424                 prev = i915_active_set_exclusive(&vma->active, &work->base.dma);
 425                 if (prev) {
 426                         __i915_sw_fence_await_dma_fence(&work->base.chain,
 427                                                         prev,
 428                                                         &work->cb);
 429                         dma_fence_put(prev);
 430                 }
 431
 432                 work->base.dma.error = 0; /* enable the queue_work() */
 433
 434                 if (vma->obj) {
 435                         __i915_gem_object_pin_pages(vma->obj);
 436                         work->pinned = vma->obj;
 437                 }
 438         } else {
 439                 ret = vma->ops->bind_vma(vma, cache_level, bind_flags);
 440                 if (ret)
 441                         return ret;
 442         }
 443
 444         atomic_or(bind_flags, &vma->flags);
 445         return 0;
 446 }
 447
 448 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
 449 {
 450         void __iomem *ptr;
 451         int err;
 452
 453         if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
 454                 err = -ENODEV;
 455                 goto err;
 456         }
 457
 458         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 459         GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
 460
 461         ptr = READ_ONCE(vma->iomap);
 462         if (ptr == NULL) {
 463                 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
 464                                         vma->node.start,
 465                                         vma->node.size);
 466                 if (ptr == NULL) {
 467                         err = -ENOMEM;
 468                         goto err;
 469                 }
 470
 471                 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
 472                         io_mapping_unmap(ptr);
 473                         ptr = vma->iomap;
 474                 }
 475         }
 476
 477         __i915_vma_pin(vma);
 478
 479         err = i915_vma_pin_fence(vma);
 480         if (err)
 481                 goto err_unpin;
 482
 483         i915_vma_set_ggtt_write(vma);
 484
 485         /* NB Access through the GTT requires the device to be awake. */
 486         return ptr;
 487
 488 err_unpin:
 489         __i915_vma_unpin(vma);
 490 err:
 491         return IO_ERR_PTR(err);
 492 }
 493
 494 void i915_vma_flush_writes(struct i915_vma *vma)
 495 {
 496         if (i915_vma_unset_ggtt_write(vma))
 497                 intel_gt_flush_ggtt_writes(vma->vm->gt);
 498 }
 499
 500 void i915_vma_unpin_iomap(struct i915_vma *vma)
 501 {
 502         GEM_BUG_ON(vma->iomap == NULL);
 503
 504         i915_vma_flush_writes(vma);
 505
 506         i915_vma_unpin_fence(vma);
 507         i915_vma_unpin(vma);
 508 }
 509
 510 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
 511 {
 512         struct i915_vma *vma;
 513         struct drm_i915_gem_object *obj;
 514
 515         vma = fetch_and_zero(p_vma);
 516         if (!vma)
 517                 return;
 518
 519         obj = vma->obj;
 520         GEM_BUG_ON(!obj);
 521
 522         i915_vma_unpin(vma);
 523         i915_vma_close(vma);
 524
 525         if (flags & I915_VMA_RELEASE_MAP)
 526                 i915_gem_object_unpin_map(obj);
 527
 528         i915_gem_object_put(obj);
 529 }
 530
 531 bool i915_vma_misplaced(const struct i915_vma *vma,
 532                         u64 size, u64 alignment, u64 flags)
 533 {
 534         if (!drm_mm_node_allocated(&vma->node))
 535                 return false;
 536
 537         if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
 538                 return true;
 539
 540         if (vma->node.size < size)
 541                 return true;
 542
 543         GEM_BUG_ON(alignment && !is_power_of_2(alignment));
 544         if (alignment && !IS_ALIGNED(vma->node.start, alignment))
 545                 return true;
 546
 547         if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
 548                 return true;
 549
 550         if (flags & PIN_OFFSET_BIAS &&
 551             vma->node.start < (flags & PIN_OFFSET_MASK))
 552                 return true;
 553
 554         if (flags & PIN_OFFSET_FIXED &&
 555             vma->node.start != (flags & PIN_OFFSET_MASK))
 556                 return true;
 557
 558         return false;
 559 }
 560
 561 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
 562 {
 563         bool mappable, fenceable;
 564
 565         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 566         GEM_BUG_ON(!vma->fence_size);
 567
 568         fenceable = (vma->node.size >= vma->fence_size &&
 569                      IS_ALIGNED(vma->node.start, vma->fence_alignment));
 570
 571         mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
 572
 573         if (mappable && fenceable)
 574                 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
 575         else
 576                 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
 577 }
 578
 579 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
 580 {
 581         struct drm_mm_node *node = &vma->node;
 582         struct drm_mm_node *other;
 583
 584         /*
 585          * On some machines we have to be careful when putting differing types
 586          * of snoopable memory together to avoid the prefetcher crossing memory
 587          * domains and dying. During vm initialisation, we decide whether or not
 588          * these constraints apply and set the drm_mm.color_adjust
 589          * appropriately.
 590          */
 591         if (!i915_vm_has_cache_coloring(vma->vm))
 592                 return true;
 593
 594         /* Only valid to be called on an already inserted vma */
 595         GEM_BUG_ON(!drm_mm_node_allocated(node));
 596         GEM_BUG_ON(list_empty(&node->node_list));
 597
 598         other = list_prev_entry(node, node_list);
 599         if (i915_node_color_differs(other, color) &&
 600             !drm_mm_hole_follows(other))
 601                 return false;
 602
 603         other = list_next_entry(node, node_list);
 604         if (i915_node_color_differs(other, color) &&
 605             !drm_mm_hole_follows(node))
 606                 return false;
 607
 608         return true;
 609 }
 610
 611 static void assert_bind_count(const struct drm_i915_gem_object *obj)
 612 {
 613         /*
 614          * Combine the assertion that the object is bound and that we have
 615          * pinned its pages. But we should never have bound the object
 616          * more than we have pinned its pages. (For complete accuracy, we
 617          * assume that no else is pinning the pages, but as a rough assertion
 618          * that we will not run into problems later, this will do!)
 619          */
 620         GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < atomic_read(&obj->bind_count));
 621 }
 622
 623 /**
 624  * i915_vma_insert - finds a slot for the vma in its address space
 625  * @vma: the vma
 626  * @size: requested size in bytes (can be larger than the VMA)
 627  * @alignment: required alignment
 628  * @flags: mask of PIN_* flags to use
 629  *
 630  * First we try to allocate some free space that meets the requirements for
 631  * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
 632  * preferrably the oldest idle entry to make room for the new VMA.
 633  *
 634  * Returns:
 635  * 0 on success, negative error code otherwise.
 636  */
 637 static int
 638 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 639 {
 640         unsigned long color;
 641         u64 start, end;
 642         int ret;
 643
 644         GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 645         GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
 646
 647         size = max(size, vma->size);
 648         alignment = max(alignment, vma->display_alignment);
 649         if (flags & PIN_MAPPABLE) {
 650                 size = max_t(typeof(size), size, vma->fence_size);
 651                 alignment = max_t(typeof(alignment),
 652                                   alignment, vma->fence_alignment);
 653         }
 654
 655         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
 656         GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
 657         GEM_BUG_ON(!is_power_of_2(alignment));
 658
 659         start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
 660         GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
 661
 662         end = vma->vm->total;
 663         if (flags & PIN_MAPPABLE)
 664                 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
 665         if (flags & PIN_ZONE_4G)
 666                 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
 667         GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
 668
 669         /* If binding the object/GGTT view requires more space than the entire
 670          * aperture has, reject it early before evicting everything in a vain
 671          * attempt to find space.
 672          */
 673         if (size > end) {
 674                 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
 675                           size, flags & PIN_MAPPABLE ? "mappable" : "total",
 676                           end);
 677                 return -ENOSPC;
 678         }
 679
 680         color = 0;
 681         if (vma->obj && i915_vm_has_cache_coloring(vma->vm))
 682                 color = vma->obj->cache_level;
 683
 684         if (flags & PIN_OFFSET_FIXED) {
 685                 u64 offset = flags & PIN_OFFSET_MASK;
 686                 if (!IS_ALIGNED(offset, alignment) ||
 687                     range_overflows(offset, size, end))
 688                         return -EINVAL;
 689
 690                 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
 691                                            size, offset, color,
 692                                            flags);
 693                 if (ret)
 694                         return ret;
 695         } else {
 696                 /*
 697                  * We only support huge gtt pages through the 48b PPGTT,
 698                  * however we also don't want to force any alignment for
 699                  * objects which need to be tightly packed into the low 32bits.
 700                  *
 701                  * Note that we assume that GGTT are limited to 4GiB for the
 702                  * forseeable future. See also i915_ggtt_offset().
 703                  */
 704                 if (upper_32_bits(end - 1) &&
 705                     vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
 706                         /*
 707                          * We can't mix 64K and 4K PTEs in the same page-table
 708                          * (2M block), and so to avoid the ugliness and
 709                          * complexity of coloring we opt for just aligning 64K
 710                          * objects to 2M.
 711                          */
 712                         u64 page_alignment =
 713                                 rounddown_pow_of_two(vma->page_sizes.sg |
 714                                                      I915_GTT_PAGE_SIZE_2M);
 715
 716                         /*
 717                          * Check we don't expand for the limited Global GTT
 718                          * (mappable aperture is even more precious!). This
 719                          * also checks that we exclude the aliasing-ppgtt.
 720                          */
 721                         GEM_BUG_ON(i915_vma_is_ggtt(vma));
 722
 723                         alignment = max(alignment, page_alignment);
 724
 725                         if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
 726                                 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
 727                 }
 728
 729                 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
 730                                           size, alignment, color,
 731                                           start, end, flags);
 732                 if (ret)
 733                         return ret;
 734
 735                 GEM_BUG_ON(vma->node.start < start);
 736                 GEM_BUG_ON(vma->node.start + vma->node.size > end);
 737         }
 738         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 739         GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
 740
 741         if (vma->obj) {
 742                 struct drm_i915_gem_object *obj = vma->obj;
 743
 744                 atomic_inc(&obj->bind_count);
 745                 assert_bind_count(obj);
 746         }
 747         list_add_tail(&vma->vm_link, &vma->vm->bound_list);
 748
 749         return 0;
 750 }
 751
 752 static void
 753 i915_vma_detach(struct i915_vma *vma)
 754 {
 755         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 756         GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 757
 758         /*
 759          * And finally now the object is completely decoupled from this
 760          * vma, we can drop its hold on the backing storage and allow
 761          * it to be reaped by the shrinker.
 762          */
 763         list_del(&vma->vm_link);
 764         if (vma->obj) {
 765                 struct drm_i915_gem_object *obj = vma->obj;
 766
 767                 assert_bind_count(obj);
 768                 atomic_dec(&obj->bind_count);
 769         }
 770 }
 771
 772 static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
 773 {
 774         unsigned int bound;
 775         bool pinned = true;
 776
 777         bound = atomic_read(&vma->flags);
 778         do {
 779                 if (unlikely(flags & ~bound))
 780                         return false;
 781
 782                 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
 783                         return false;
 784
 785                 if (!(bound & I915_VMA_PIN_MASK))
 786                         goto unpinned;
 787
 788                 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
 789         } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
 790
 791         return true;
 792
 793 unpinned:
 794         /*
 795          * If pin_count==0, but we are bound, check under the lock to avoid
 796          * racing with a concurrent i915_vma_unbind().
 797          */
 798         mutex_lock(&vma->vm->mutex);
 799         do {
 800                 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) {
 801                         pinned = false;
 802                         break;
 803                 }
 804
 805                 if (unlikely(flags & ~bound)) {
 806                         pinned = false;
 807                         break;
 808                 }
 809         } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
 810         mutex_unlock(&vma->vm->mutex);
 811
 812         return pinned;
 813 }
 814
 815 static int vma_get_pages(struct i915_vma *vma)
 816 {
 817         int err = 0;
 818
 819         if (atomic_add_unless(&vma->pages_count, 1, 0))
 820                 return 0;
 821
 822         /* Allocations ahoy! */
 823         if (mutex_lock_interruptible(&vma->pages_mutex))
 824                 return -EINTR;
 825
 826         if (!atomic_read(&vma->pages_count)) {
 827                 if (vma->obj) {
 828                         err = i915_gem_object_pin_pages(vma->obj);
 829                         if (err)
 830                                 goto unlock;
 831                 }
 832
 833                 err = vma->ops->set_pages(vma);
 834                 if (err) {
 835                         if (vma->obj)
 836                                 i915_gem_object_unpin_pages(vma->obj);
 837                         goto unlock;
 838                 }
 839         }
 840         atomic_inc(&vma->pages_count);
 841
 842 unlock:
 843         mutex_unlock(&vma->pages_mutex);
 844
 845         return err;
 846 }
 847
 848 static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
 849 {
 850         /* We allocate under vma_get_pages, so beware the shrinker */
 851         mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING);
 852         GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
 853         if (atomic_sub_return(count, &vma->pages_count) == 0) {
 854                 vma->ops->clear_pages(vma);
 855                 GEM_BUG_ON(vma->pages);
 856                 if (vma->obj)
 857                         i915_gem_object_unpin_pages(vma->obj);
 858         }
 859         mutex_unlock(&vma->pages_mutex);
 860 }
 861
 862 static void vma_put_pages(struct i915_vma *vma)
 863 {
 864         if (atomic_add_unless(&vma->pages_count, -1, 1))
 865                 return;
 866
 867         __vma_put_pages(vma, 1);
 868 }
 869
 870 static void vma_unbind_pages(struct i915_vma *vma)
 871 {
 872         unsigned int count;
 873
 874         lockdep_assert_held(&vma->vm->mutex);
 875
 876         /* The upper portion of pages_count is the number of bindings */
 877         count = atomic_read(&vma->pages_count);
 878         count >>= I915_VMA_PAGES_BIAS;
 879         GEM_BUG_ON(!count);
 880
 881         __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
 882 }
 883
 884 int i915_vma_pin(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 885 {
 886         struct i915_vma_work *work = NULL;
 887         intel_wakeref_t wakeref = 0;
 888         unsigned int bound;
 889         int err;
 890
 891         BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
 892         BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
 893
 894         GEM_BUG_ON(flags & PIN_UPDATE);
 895         GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
 896
 897         /* First try and grab the pin without rebinding the vma */
 898         if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
 899                 return 0;
 900
 901         err = vma_get_pages(vma);
 902         if (err)
 903                 return err;
 904
 905         if (flags & vma->vm->bind_async_flags) {
 906                 work = i915_vma_work();
 907                 if (!work) {
 908                         err = -ENOMEM;
 909                         goto err_pages;
 910                 }
 911         }
 912
 913         if (flags & PIN_GLOBAL)
 914                 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
 915
 916         /* No more allocations allowed once we hold vm->mutex */
 917         err = mutex_lock_interruptible(&vma->vm->mutex);
 918         if (err)
 919                 goto err_fence;
 920
 921         if (unlikely(i915_vma_is_closed(vma))) {
 922                 err = -ENOENT;
 923                 goto err_unlock;
 924         }
 925
 926         bound = atomic_read(&vma->flags);
 927         if (unlikely(bound & I915_VMA_ERROR)) {
 928                 err = -ENOMEM;
 929                 goto err_unlock;
 930         }
 931
 932         if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
 933                 err = -EAGAIN; /* pins are meant to be fairly temporary */
 934                 goto err_unlock;
 935         }
 936
 937         if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
 938                 __i915_vma_pin(vma);
 939                 goto err_unlock;
 940         }
 941
 942         err = i915_active_acquire(&vma->active);
 943         if (err)
 944                 goto err_unlock;
 945
 946         if (!(bound & I915_VMA_BIND_MASK)) {
 947                 err = i915_vma_insert(vma, size, alignment, flags);
 948                 if (err)
 949                         goto err_active;
 950
 951                 if (i915_is_ggtt(vma->vm))
 952                         __i915_vma_set_map_and_fenceable(vma);
 953         }
 954
 955         GEM_BUG_ON(!vma->pages);
 956         err = i915_vma_bind(vma,
 957                             vma->obj ? vma->obj->cache_level : 0,
 958                             flags, work);
 959         if (err)
 960                 goto err_remove;
 961
 962         /* There should only be at most 2 active bindings (user, global) */
 963         GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
 964         atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
 965         list_move_tail(&vma->vm_link, &vma->vm->bound_list);
 966
 967         __i915_vma_pin(vma);
 968         GEM_BUG_ON(!i915_vma_is_pinned(vma));
 969         GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
 970         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
 971
 972 err_remove:
 973         if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
 974                 i915_vma_detach(vma);
 975                 drm_mm_remove_node(&vma->node);
 976         }
 977 err_active:
 978         i915_active_release(&vma->active);
 979 err_unlock:
 980         mutex_unlock(&vma->vm->mutex);
 981 err_fence:
 982         if (work)
 983                 dma_fence_work_commit(&work->base);
 984         if (wakeref)
 985                 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
 986 err_pages:
 987         vma_put_pages(vma);
 988         return err;
 989 }
 990
 991 static void flush_idle_contexts(struct intel_gt *gt)
 992 {
 993         struct intel_engine_cs *engine;
 994         enum intel_engine_id id;
 995
 996         for_each_engine(engine, gt, id)
 997                 intel_engine_flush_barriers(engine);
 998
 999         intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
1000 }
1001
1002 int i915_ggtt_pin(struct i915_vma *vma, u32 align, unsigned int flags)
1003 {
1004         struct i915_address_space *vm = vma->vm;
1005         int err;
1006
1007         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
1008
1009         do {
1010                 err = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL);
1011                 if (err != -ENOSPC) {
1012                         if (!err) {
1013                                 err = i915_vma_wait_for_bind(vma);
1014                                 if (err)
1015                                         i915_vma_unpin(vma);
1016                         }
1017                         return err;
1018                 }
1019
1020                 /* Unlike i915_vma_pin, we don't take no for an answer! */
1021                 flush_idle_contexts(vm->gt);
1022                 if (mutex_lock_interruptible(&vm->mutex) == 0) {
1023                         i915_gem_evict_vm(vm);
1024                         mutex_unlock(&vm->mutex);
1025                 }
1026         } while (1);
1027 }
1028
1029 void i915_vma_close(struct i915_vma *vma)
1030 {
1031         struct intel_gt *gt = vma->vm->gt;
1032         unsigned long flags;
1033
1034         GEM_BUG_ON(i915_vma_is_closed(vma));
1035
1036         /*
1037          * We defer actually closing, unbinding and destroying the VMA until
1038          * the next idle point, or if the object is freed in the meantime. By
1039          * postponing the unbind, we allow for it to be resurrected by the
1040          * client, avoiding the work required to rebind the VMA. This is
1041          * advantageous for DRI, where the client/server pass objects
1042          * between themselves, temporarily opening a local VMA to the
1043          * object, and then closing it again. The same object is then reused
1044          * on the next frame (or two, depending on the depth of the swap queue)
1045          * causing us to rebind the VMA once more. This ends up being a lot
1046          * of wasted work for the steady state.
1047          */
1048         spin_lock_irqsave(&gt->closed_lock, flags);
1049         list_add(&vma->closed_link, &gt->closed_vma);
1050         spin_unlock_irqrestore(&gt->closed_lock, flags);
1051 }
1052
1053 static void __i915_vma_remove_closed(struct i915_vma *vma)
1054 {
1055         struct intel_gt *gt = vma->vm->gt;
1056
1057         spin_lock_irq(&gt->closed_lock);
1058         list_del_init(&vma->closed_link);
1059         spin_unlock_irq(&gt->closed_lock);
1060 }
1061
1062 void i915_vma_reopen(struct i915_vma *vma)
1063 {
1064         if (i915_vma_is_closed(vma))
1065                 __i915_vma_remove_closed(vma);
1066 }
1067
1068 void i915_vma_release(struct kref *ref)
1069 {
1070         struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
1071
1072         if (drm_mm_node_allocated(&vma->node)) {
1073                 mutex_lock(&vma->vm->mutex);
1074                 atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
1075                 WARN_ON(__i915_vma_unbind(vma));
1076                 mutex_unlock(&vma->vm->mutex);
1077                 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
1078         }
1079         GEM_BUG_ON(i915_vma_is_active(vma));
1080
1081         if (vma->obj) {
1082                 struct drm_i915_gem_object *obj = vma->obj;
1083
1084                 spin_lock(&obj->vma.lock);
1085                 list_del(&vma->obj_link);
1086                 rb_erase(&vma->obj_node, &obj->vma.tree);
1087                 spin_unlock(&obj->vma.lock);
1088         }
1089
1090         __i915_vma_remove_closed(vma);
1091         i915_vm_put(vma->vm);
1092
1093         i915_active_fini(&vma->active);
1094         i915_vma_free(vma);
1095 }
1096
1097 void i915_vma_parked(struct intel_gt *gt)
1098 {
1099         struct i915_vma *vma, *next;
1100         LIST_HEAD(closed);
1101
1102         spin_lock_irq(&gt->closed_lock);
1103         list_for_each_entry_safe(vma, next, &gt->closed_vma, closed_link) {
1104                 struct drm_i915_gem_object *obj = vma->obj;
1105                 struct i915_address_space *vm = vma->vm;
1106
1107                 /* XXX All to avoid keeping a reference on i915_vma itself */
1108
1109                 if (!kref_get_unless_zero(&obj->base.refcount))
1110                         continue;
1111
1112                 if (!i915_vm_tryopen(vm)) {
1113                         i915_gem_object_put(obj);
1114                         continue;
1115                 }
1116
1117                 list_move(&vma->closed_link, &closed);
1118         }
1119         spin_unlock_irq(&gt->closed_lock);
1120
1121         /* As the GT is held idle, no vma can be reopened as we destroy them */
1122         list_for_each_entry_safe(vma, next, &closed, closed_link) {
1123                 struct drm_i915_gem_object *obj = vma->obj;
1124                 struct i915_address_space *vm = vma->vm;
1125
1126                 INIT_LIST_HEAD(&vma->closed_link);
1127                 __i915_vma_put(vma);
1128
1129                 i915_gem_object_put(obj);
1130                 i915_vm_close(vm);
1131         }
1132 }
1133
1134 static void __i915_vma_iounmap(struct i915_vma *vma)
1135 {
1136         GEM_BUG_ON(i915_vma_is_pinned(vma));
1137
1138         if (vma->iomap == NULL)
1139                 return;
1140
1141         io_mapping_unmap(vma->iomap);
1142         vma->iomap = NULL;
1143 }
1144
1145 void i915_vma_revoke_mmap(struct i915_vma *vma)
1146 {
1147         struct drm_vma_offset_node *node;
1148         u64 vma_offset;
1149
1150         if (!i915_vma_has_userfault(vma))
1151                 return;
1152
1153         GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
1154         GEM_BUG_ON(!vma->obj->userfault_count);
1155
1156         node = &vma->mmo->vma_node;
1157         vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
1158         unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
1159                             drm_vma_node_offset_addr(node) + vma_offset,
1160                             vma->size,
1161                             1);
1162
1163         i915_vma_unset_userfault(vma);
1164         if (!--vma->obj->userfault_count)
1165                 list_del(&vma->obj->userfault_link);
1166 }
1167
1168 int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
1169 {
1170         int err;
1171
1172         GEM_BUG_ON(!i915_vma_is_pinned(vma));
1173
1174         /* Wait for the vma to be bound before we start! */
1175         err = i915_request_await_active(rq, &vma->active, 0);
1176         if (err)
1177                 return err;
1178
1179         return i915_active_add_request(&vma->active, rq);
1180 }
1181
1182 int i915_vma_move_to_active(struct i915_vma *vma,
1183                             struct i915_request *rq,
1184                             unsigned int flags)
1185 {
1186         struct drm_i915_gem_object *obj = vma->obj;
1187         int err;
1188
1189         assert_object_held(obj);
1190
1191         err = __i915_vma_move_to_active(vma, rq);
1192         if (unlikely(err))
1193                 return err;
1194
1195         if (flags & EXEC_OBJECT_WRITE) {
1196                 struct intel_frontbuffer *front;
1197
1198                 front = __intel_frontbuffer_get(obj);
1199                 if (unlikely(front)) {
1200                         if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
1201                                 i915_active_add_request(&front->write, rq);
1202                         intel_frontbuffer_put(front);
1203                 }
1204
1205                 dma_resv_add_excl_fence(vma->resv, &rq->fence);
1206                 obj->write_domain = I915_GEM_DOMAIN_RENDER;
1207                 obj->read_domains = 0;
1208         } else {
1209                 err = dma_resv_reserve_shared(vma->resv, 1);
1210                 if (unlikely(err))
1211                         return err;
1212
1213                 dma_resv_add_shared_fence(vma->resv, &rq->fence);
1214                 obj->write_domain = 0;
1215         }
1216         obj->read_domains |= I915_GEM_GPU_DOMAINS;
1217         obj->mm.dirty = true;
1218
1219         GEM_BUG_ON(!i915_vma_is_active(vma));
1220         return 0;
1221 }
1222
1223 int __i915_vma_unbind(struct i915_vma *vma)
1224 {
1225         int ret;
1226
1227         lockdep_assert_held(&vma->vm->mutex);
1228
1229         /*
1230          * First wait upon any activity as retiring the request may
1231          * have side-effects such as unpinning or even unbinding this vma.
1232          *
1233          * XXX Actually waiting under the vm->mutex is a hinderance and
1234          * should be pipelined wherever possible. In cases where that is
1235          * unavoidable, we should lift the wait to before the mutex.
1236          */
1237         ret = i915_vma_sync(vma);
1238         if (ret)
1239                 return ret;
1240
1241         if (i915_vma_is_pinned(vma)) {
1242                 vma_print_allocator(vma, "is pinned");
1243                 return -EAGAIN;
1244         }
1245
1246         /*
1247          * After confirming that no one else is pinning this vma, wait for
1248          * any laggards who may have crept in during the wait (through
1249          * a residual pin skipping the vm->mutex) to complete.
1250          */
1251         ret = i915_vma_sync(vma);
1252         if (ret)
1253                 return ret;
1254
1255         if (!drm_mm_node_allocated(&vma->node))
1256                 return 0;
1257
1258         GEM_BUG_ON(i915_vma_is_pinned(vma));
1259         GEM_BUG_ON(i915_vma_is_active(vma));
1260
1261         if (i915_vma_is_map_and_fenceable(vma)) {
1262                 /*
1263                  * Check that we have flushed all writes through the GGTT
1264                  * before the unbind, other due to non-strict nature of those
1265                  * indirect writes they may end up referencing the GGTT PTE
1266                  * after the unbind.
1267                  *
1268                  * Note that we may be concurrently poking at the GGTT_WRITE
1269                  * bit from set-domain, as we mark all GGTT vma associated
1270                  * with an object. We know this is for another vma, as we
1271                  * are currently unbinding this one -- so if this vma will be
1272                  * reused, it will be refaulted and have its dirty bit set
1273                  * before the next write.
1274                  */
1275                 i915_vma_flush_writes(vma);
1276
1277                 /* release the fence reg _after_ flushing */
1278                 ret = i915_vma_revoke_fence(vma);
1279                 if (ret)
1280                         return ret;
1281
1282                 /* Force a pagefault for domain tracking on next user access */
1283                 i915_vma_revoke_mmap(vma);
1284
1285                 __i915_vma_iounmap(vma);
1286                 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
1287         }
1288         GEM_BUG_ON(vma->fence);
1289         GEM_BUG_ON(i915_vma_has_userfault(vma));
1290
1291         if (likely(atomic_read(&vma->vm->open))) {
1292                 trace_i915_vma_unbind(vma);
1293                 vma->ops->unbind_vma(vma);
1294         }
1295         atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE),
1296                    &vma->flags);
1297
1298         i915_vma_detach(vma);
1299         vma_unbind_pages(vma);
1300
1301         drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
1302         return 0;
1303 }
1304
1305 int i915_vma_unbind(struct i915_vma *vma)
1306 {
1307         struct i915_address_space *vm = vma->vm;
1308         intel_wakeref_t wakeref = 0;
1309         int err;
1310
1311         if (!drm_mm_node_allocated(&vma->node))
1312                 return 0;
1313
1314         if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
1315                 /* XXX not always required: nop_clear_range */
1316                 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
1317
1318         /* Optimistic wait before taking the mutex */
1319         err = i915_vma_sync(vma);
1320         if (err)
1321                 goto out_rpm;
1322
1323         err = mutex_lock_interruptible(&vm->mutex);
1324         if (err)
1325                 goto out_rpm;
1326
1327         err = __i915_vma_unbind(vma);
1328         mutex_unlock(&vm->mutex);
1329
1330 out_rpm:
1331         if (wakeref)
1332                 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
1333         return err;
1334 }
1335
1336 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1337 {
1338         i915_gem_object_make_unshrinkable(vma->obj);
1339         return vma;
1340 }
1341
1342 void i915_vma_make_shrinkable(struct i915_vma *vma)
1343 {
1344         i915_gem_object_make_shrinkable(vma->obj);
1345 }
1346
1347 void i915_vma_make_purgeable(struct i915_vma *vma)
1348 {
1349         i915_gem_object_make_purgeable(vma->obj);
1350 }
1351
1352 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1353 #include "selftests/i915_vma.c"
1354 #endif
1355
1356 static void i915_global_vma_shrink(void)
1357 {
1358         kmem_cache_shrink(global.slab_vmas);
1359 }
1360
1361 static void i915_global_vma_exit(void)
1362 {
1363         kmem_cache_destroy(global.slab_vmas);
1364 }
1365
1366 static struct i915_global_vma global = { {
1367         .shrink = i915_global_vma_shrink,
1368         .exit = i915_global_vma_exit,
1369 } };
1370
1371 int __init i915_global_vma_init(void)
1372 {
1373         global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1374         if (!global.slab_vmas)
1375                 return -ENOMEM;
1376
1377         i915_global_register(&global.base);
1378         return 0;
1379 }