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 *pos = ERR_PTR(-E2BIG);
 108         struct i915_vma *vma;
 109         struct rb_node *rb, **p;
 110
 111         /* The aliasing_ppgtt should never be used directly! */
 112         GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm);
 113
 114         vma = i915_vma_alloc();
 115         if (vma == NULL)
 116                 return ERR_PTR(-ENOMEM);
 117
 118         kref_init(&vma->ref);
 119         mutex_init(&vma->pages_mutex);
 120         vma->vm = i915_vm_get(vm);
 121         vma->ops = &vm->vma_ops;
 122         vma->obj = obj;
 123         vma->resv = obj->base.resv;
 124         vma->size = obj->base.size;
 125         vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
 126
 127         i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire);
 128
 129         /* Declare ourselves safe for use inside shrinkers */
 130         if (IS_ENABLED(CONFIG_LOCKDEP)) {
 131                 fs_reclaim_acquire(GFP_KERNEL);
 132                 might_lock(&vma->active.mutex);
 133                 fs_reclaim_release(GFP_KERNEL);
 134         }
 135
 136         INIT_LIST_HEAD(&vma->closed_link);
 137
 138         if (view && view->type != I915_GGTT_VIEW_NORMAL) {
 139                 vma->ggtt_view = *view;
 140                 if (view->type == I915_GGTT_VIEW_PARTIAL) {
 141                         GEM_BUG_ON(range_overflows_t(u64,
 142                                                      view->partial.offset,
 143                                                      view->partial.size,
 144                                                      obj->base.size >> PAGE_SHIFT));
 145                         vma->size = view->partial.size;
 146                         vma->size <<= PAGE_SHIFT;
 147                         GEM_BUG_ON(vma->size > obj->base.size);
 148                 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
 149                         vma->size = intel_rotation_info_size(&view->rotated);
 150                         vma->size <<= PAGE_SHIFT;
 151                 } else if (view->type == I915_GGTT_VIEW_REMAPPED) {
 152                         vma->size = intel_remapped_info_size(&view->remapped);
 153                         vma->size <<= PAGE_SHIFT;
 154                 }
 155         }
 156
 157         if (unlikely(vma->size > vm->total))
 158                 goto err_vma;
 159
 160         GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
 161
 162         spin_lock(&obj->vma.lock);
 163
 164         if (i915_is_ggtt(vm)) {
 165                 if (unlikely(overflows_type(vma->size, u32)))
 166                         goto err_unlock;
 167
 168                 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
 169                                                       i915_gem_object_get_tiling(obj),
 170                                                       i915_gem_object_get_stride(obj));
 171                 if (unlikely(vma->fence_size < vma->size || /* overflow */
 172                              vma->fence_size > vm->total))
 173                         goto err_unlock;
 174
 175                 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
 176
 177                 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
 178                                                                 i915_gem_object_get_tiling(obj),
 179                                                                 i915_gem_object_get_stride(obj));
 180                 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
 181
 182                 __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma));
 183         }
 184
 185         rb = NULL;
 186         p = &obj->vma.tree.rb_node;
 187         while (*p) {
 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                         p = &rb->rb_right;
 201                 else if (cmp > 0)
 202                         p = &rb->rb_left;
 203                 else
 204                         goto err_unlock;
 205         }
 206         rb_link_node(&vma->obj_node, rb, p);
 207         rb_insert_color(&vma->obj_node, &obj->vma.tree);
 208
 209         if (i915_vma_is_ggtt(vma))
 210                 /*
 211                  * We put the GGTT vma at the start of the vma-list, followed
 212                  * by the ppGGTT vma. This allows us to break early when
 213                  * iterating over only the GGTT vma for an object, see
 214                  * for_each_ggtt_vma()
 215                  */
 216                 list_add(&vma->obj_link, &obj->vma.list);
 217         else
 218                 list_add_tail(&vma->obj_link, &obj->vma.list);
 219
 220         spin_unlock(&obj->vma.lock);
 221
 222         return vma;
 223
 224 err_unlock:
 225         spin_unlock(&obj->vma.lock);
 226 err_vma:
 227         i915_vm_put(vm);
 228         i915_vma_free(vma);
 229         return pos;
 230 }
 231
 232 static struct i915_vma *
 233 vma_lookup(struct drm_i915_gem_object *obj,
 234            struct i915_address_space *vm,
 235            const struct i915_ggtt_view *view)
 236 {
 237         struct rb_node *rb;
 238
 239         rb = obj->vma.tree.rb_node;
 240         while (rb) {
 241                 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
 242                 long cmp;
 243
 244                 cmp = i915_vma_compare(vma, vm, view);
 245                 if (cmp == 0)
 246                         return vma;
 247
 248                 if (cmp < 0)
 249                         rb = rb->rb_right;
 250                 else
 251                         rb = rb->rb_left;
 252         }
 253
 254         return NULL;
 255 }
 256
 257 /**
 258  * i915_vma_instance - return the singleton instance of the VMA
 259  * @obj: parent &struct drm_i915_gem_object to be mapped
 260  * @vm: address space in which the mapping is located
 261  * @view: additional mapping requirements
 262  *
 263  * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
 264  * the same @view characteristics. If a match is not found, one is created.
 265  * Once created, the VMA is kept until either the object is freed, or the
 266  * address space is closed.
 267  *
 268  * Returns the vma, or an error pointer.
 269  */
 270 struct i915_vma *
 271 i915_vma_instance(struct drm_i915_gem_object *obj,
 272                   struct i915_address_space *vm,
 273                   const struct i915_ggtt_view *view)
 274 {
 275         struct i915_vma *vma;
 276
 277         GEM_BUG_ON(view && !i915_is_ggtt(vm));
 278         GEM_BUG_ON(!atomic_read(&vm->open));
 279
 280         spin_lock(&obj->vma.lock);
 281         vma = vma_lookup(obj, vm, view);
 282         spin_unlock(&obj->vma.lock);
 283
 284         /* vma_create() will resolve the race if another creates the vma */
 285         if (unlikely(!vma))
 286                 vma = vma_create(obj, vm, view);
 287
 288         GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
 289         return vma;
 290 }
 291
 292 struct i915_vma_work {
 293         struct dma_fence_work base;
 294         struct i915_address_space *vm;
 295         struct i915_vm_pt_stash stash;
 296         struct i915_vma *vma;
 297         struct drm_i915_gem_object *pinned;
 298         struct i915_sw_dma_fence_cb cb;
 299         enum i915_cache_level cache_level;
 300         unsigned int flags;
 301 };
 302
 303 static int __vma_bind(struct dma_fence_work *work)
 304 {
 305         struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
 306         struct i915_vma *vma = vw->vma;
 307
 308         vma->ops->bind_vma(vw->vm, &vw->stash,
 309                            vma, vw->cache_level, vw->flags);
 310         return 0;
 311 }
 312
 313 static void __vma_release(struct dma_fence_work *work)
 314 {
 315         struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
 316
 317         if (vw->pinned) {
 318                 __i915_gem_object_unpin_pages(vw->pinned);
 319                 i915_gem_object_put(vw->pinned);
 320         }
 321
 322         i915_vm_free_pt_stash(vw->vm, &vw->stash);
 323         i915_vm_put(vw->vm);
 324 }
 325
 326 static const struct dma_fence_work_ops bind_ops = {
 327         .name = "bind",
 328         .work = __vma_bind,
 329         .release = __vma_release,
 330 };
 331
 332 struct i915_vma_work *i915_vma_work(void)
 333 {
 334         struct i915_vma_work *vw;
 335
 336         vw = kzalloc(sizeof(*vw), GFP_KERNEL);
 337         if (!vw)
 338                 return NULL;
 339
 340         dma_fence_work_init(&vw->base, &bind_ops);
 341         vw->base.dma.error = -EAGAIN; /* disable the worker by default */
 342
 343         return vw;
 344 }
 345
 346 int i915_vma_wait_for_bind(struct i915_vma *vma)
 347 {
 348         int err = 0;
 349
 350         if (rcu_access_pointer(vma->active.excl.fence)) {
 351                 struct dma_fence *fence;
 352
 353                 rcu_read_lock();
 354                 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence);
 355                 rcu_read_unlock();
 356                 if (fence) {
 357                         err = dma_fence_wait(fence, MAX_SCHEDULE_TIMEOUT);
 358                         dma_fence_put(fence);
 359                 }
 360         }
 361
 362         return err;
 363 }
 364
 365 /**
 366  * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
 367  * @vma: VMA to map
 368  * @cache_level: mapping cache level
 369  * @flags: flags like global or local mapping
 370  * @work: preallocated worker for allocating and binding the PTE
 371  *
 372  * DMA addresses are taken from the scatter-gather table of this object (or of
 373  * this VMA in case of non-default GGTT views) and PTE entries set up.
 374  * Note that DMA addresses are also the only part of the SG table we care about.
 375  */
 376 int i915_vma_bind(struct i915_vma *vma,
 377                   enum i915_cache_level cache_level,
 378                   u32 flags,
 379                   struct i915_vma_work *work)
 380 {
 381         u32 bind_flags;
 382         u32 vma_flags;
 383
 384         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 385         GEM_BUG_ON(vma->size > vma->node.size);
 386
 387         if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
 388                                               vma->node.size,
 389                                               vma->vm->total)))
 390                 return -ENODEV;
 391
 392         if (GEM_DEBUG_WARN_ON(!flags))
 393                 return -EINVAL;
 394
 395         bind_flags = flags;
 396         bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
 397
 398         vma_flags = atomic_read(&vma->flags);
 399         vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
 400
 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->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;
 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 = i915_gem_object_get(vma->obj);
 437                 }
 438         } else {
 439                 vma->ops->bind_vma(vma->vm, NULL, vma, cache_level, bind_flags);
 440         }
 441
 442         atomic_or(bind_flags, &vma->flags);
 443         return 0;
 444 }
 445
 446 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
 447 {
 448         void __iomem *ptr;
 449         int err;
 450
 451         if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
 452                 err = -ENODEV;
 453                 goto err;
 454         }
 455
 456         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 457         GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
 458
 459         ptr = READ_ONCE(vma->iomap);
 460         if (ptr == NULL) {
 461                 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
 462                                         vma->node.start,
 463                                         vma->node.size);
 464                 if (ptr == NULL) {
 465                         err = -ENOMEM;
 466                         goto err;
 467                 }
 468
 469                 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
 470                         io_mapping_unmap(ptr);
 471                         ptr = vma->iomap;
 472                 }
 473         }
 474
 475         __i915_vma_pin(vma);
 476
 477         err = i915_vma_pin_fence(vma);
 478         if (err)
 479                 goto err_unpin;
 480
 481         i915_vma_set_ggtt_write(vma);
 482
 483         /* NB Access through the GTT requires the device to be awake. */
 484         return ptr;
 485
 486 err_unpin:
 487         __i915_vma_unpin(vma);
 488 err:
 489         return IO_ERR_PTR(err);
 490 }
 491
 492 void i915_vma_flush_writes(struct i915_vma *vma)
 493 {
 494         if (i915_vma_unset_ggtt_write(vma))
 495                 intel_gt_flush_ggtt_writes(vma->vm->gt);
 496 }
 497
 498 void i915_vma_unpin_iomap(struct i915_vma *vma)
 499 {
 500         GEM_BUG_ON(vma->iomap == NULL);
 501
 502         i915_vma_flush_writes(vma);
 503
 504         i915_vma_unpin_fence(vma);
 505         i915_vma_unpin(vma);
 506 }
 507
 508 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
 509 {
 510         struct i915_vma *vma;
 511         struct drm_i915_gem_object *obj;
 512
 513         vma = fetch_and_zero(p_vma);
 514         if (!vma)
 515                 return;
 516
 517         obj = vma->obj;
 518         GEM_BUG_ON(!obj);
 519
 520         i915_vma_unpin(vma);
 521
 522         if (flags & I915_VMA_RELEASE_MAP)
 523                 i915_gem_object_unpin_map(obj);
 524
 525         i915_gem_object_put(obj);
 526 }
 527
 528 bool i915_vma_misplaced(const struct i915_vma *vma,
 529                         u64 size, u64 alignment, u64 flags)
 530 {
 531         if (!drm_mm_node_allocated(&vma->node))
 532                 return false;
 533
 534         if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
 535                 return true;
 536
 537         if (vma->node.size < size)
 538                 return true;
 539
 540         GEM_BUG_ON(alignment && !is_power_of_2(alignment));
 541         if (alignment && !IS_ALIGNED(vma->node.start, alignment))
 542                 return true;
 543
 544         if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
 545                 return true;
 546
 547         if (flags & PIN_OFFSET_BIAS &&
 548             vma->node.start < (flags & PIN_OFFSET_MASK))
 549                 return true;
 550
 551         if (flags & PIN_OFFSET_FIXED &&
 552             vma->node.start != (flags & PIN_OFFSET_MASK))
 553                 return true;
 554
 555         return false;
 556 }
 557
 558 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
 559 {
 560         bool mappable, fenceable;
 561
 562         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 563         GEM_BUG_ON(!vma->fence_size);
 564
 565         fenceable = (vma->node.size >= vma->fence_size &&
 566                      IS_ALIGNED(vma->node.start, vma->fence_alignment));
 567
 568         mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
 569
 570         if (mappable && fenceable)
 571                 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
 572         else
 573                 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
 574 }
 575
 576 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
 577 {
 578         struct drm_mm_node *node = &vma->node;
 579         struct drm_mm_node *other;
 580
 581         /*
 582          * On some machines we have to be careful when putting differing types
 583          * of snoopable memory together to avoid the prefetcher crossing memory
 584          * domains and dying. During vm initialisation, we decide whether or not
 585          * these constraints apply and set the drm_mm.color_adjust
 586          * appropriately.
 587          */
 588         if (!i915_vm_has_cache_coloring(vma->vm))
 589                 return true;
 590
 591         /* Only valid to be called on an already inserted vma */
 592         GEM_BUG_ON(!drm_mm_node_allocated(node));
 593         GEM_BUG_ON(list_empty(&node->node_list));
 594
 595         other = list_prev_entry(node, node_list);
 596         if (i915_node_color_differs(other, color) &&
 597             !drm_mm_hole_follows(other))
 598                 return false;
 599
 600         other = list_next_entry(node, node_list);
 601         if (i915_node_color_differs(other, color) &&
 602             !drm_mm_hole_follows(node))
 603                 return false;
 604
 605         return true;
 606 }
 607
 608 /**
 609  * i915_vma_insert - finds a slot for the vma in its address space
 610  * @vma: the vma
 611  * @size: requested size in bytes (can be larger than the VMA)
 612  * @alignment: required alignment
 613  * @flags: mask of PIN_* flags to use
 614  *
 615  * First we try to allocate some free space that meets the requirements for
 616  * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
 617  * preferrably the oldest idle entry to make room for the new VMA.
 618  *
 619  * Returns:
 620  * 0 on success, negative error code otherwise.
 621  */
 622 static int
 623 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 624 {
 625         unsigned long color;
 626         u64 start, end;
 627         int ret;
 628
 629         GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 630         GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
 631
 632         size = max(size, vma->size);
 633         alignment = max(alignment, vma->display_alignment);
 634         if (flags & PIN_MAPPABLE) {
 635                 size = max_t(typeof(size), size, vma->fence_size);
 636                 alignment = max_t(typeof(alignment),
 637                                   alignment, vma->fence_alignment);
 638         }
 639
 640         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
 641         GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
 642         GEM_BUG_ON(!is_power_of_2(alignment));
 643
 644         start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
 645         GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
 646
 647         end = vma->vm->total;
 648         if (flags & PIN_MAPPABLE)
 649                 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
 650         if (flags & PIN_ZONE_4G)
 651                 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
 652         GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
 653
 654         /* If binding the object/GGTT view requires more space than the entire
 655          * aperture has, reject it early before evicting everything in a vain
 656          * attempt to find space.
 657          */
 658         if (size > end) {
 659                 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
 660                           size, flags & PIN_MAPPABLE ? "mappable" : "total",
 661                           end);
 662                 return -ENOSPC;
 663         }
 664
 665         color = 0;
 666         if (vma->obj && i915_vm_has_cache_coloring(vma->vm))
 667                 color = vma->obj->cache_level;
 668
 669         if (flags & PIN_OFFSET_FIXED) {
 670                 u64 offset = flags & PIN_OFFSET_MASK;
 671                 if (!IS_ALIGNED(offset, alignment) ||
 672                     range_overflows(offset, size, end))
 673                         return -EINVAL;
 674
 675                 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
 676                                            size, offset, color,
 677                                            flags);
 678                 if (ret)
 679                         return ret;
 680         } else {
 681                 /*
 682                  * We only support huge gtt pages through the 48b PPGTT,
 683                  * however we also don't want to force any alignment for
 684                  * objects which need to be tightly packed into the low 32bits.
 685                  *
 686                  * Note that we assume that GGTT are limited to 4GiB for the
 687                  * forseeable future. See also i915_ggtt_offset().
 688                  */
 689                 if (upper_32_bits(end - 1) &&
 690                     vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
 691                         /*
 692                          * We can't mix 64K and 4K PTEs in the same page-table
 693                          * (2M block), and so to avoid the ugliness and
 694                          * complexity of coloring we opt for just aligning 64K
 695                          * objects to 2M.
 696                          */
 697                         u64 page_alignment =
 698                                 rounddown_pow_of_two(vma->page_sizes.sg |
 699                                                      I915_GTT_PAGE_SIZE_2M);
 700
 701                         /*
 702                          * Check we don't expand for the limited Global GTT
 703                          * (mappable aperture is even more precious!). This
 704                          * also checks that we exclude the aliasing-ppgtt.
 705                          */
 706                         GEM_BUG_ON(i915_vma_is_ggtt(vma));
 707
 708                         alignment = max(alignment, page_alignment);
 709
 710                         if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
 711                                 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
 712                 }
 713
 714                 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
 715                                           size, alignment, color,
 716                                           start, end, flags);
 717                 if (ret)
 718                         return ret;
 719
 720                 GEM_BUG_ON(vma->node.start < start);
 721                 GEM_BUG_ON(vma->node.start + vma->node.size > end);
 722         }
 723         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 724         GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
 725
 726         list_add_tail(&vma->vm_link, &vma->vm->bound_list);
 727
 728         return 0;
 729 }
 730
 731 static void
 732 i915_vma_detach(struct i915_vma *vma)
 733 {
 734         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 735         GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 736
 737         /*
 738          * And finally now the object is completely decoupled from this
 739          * vma, we can drop its hold on the backing storage and allow
 740          * it to be reaped by the shrinker.
 741          */
 742         list_del(&vma->vm_link);
 743 }
 744
 745 static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
 746 {
 747         unsigned int bound;
 748         bool pinned = true;
 749
 750         bound = atomic_read(&vma->flags);
 751         do {
 752                 if (unlikely(flags & ~bound))
 753                         return false;
 754
 755                 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
 756                         return false;
 757
 758                 if (!(bound & I915_VMA_PIN_MASK))
 759                         goto unpinned;
 760
 761                 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
 762         } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
 763
 764         return true;
 765
 766 unpinned:
 767         /*
 768          * If pin_count==0, but we are bound, check under the lock to avoid
 769          * racing with a concurrent i915_vma_unbind().
 770          */
 771         mutex_lock(&vma->vm->mutex);
 772         do {
 773                 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) {
 774                         pinned = false;
 775                         break;
 776                 }
 777
 778                 if (unlikely(flags & ~bound)) {
 779                         pinned = false;
 780                         break;
 781                 }
 782         } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
 783         mutex_unlock(&vma->vm->mutex);
 784
 785         return pinned;
 786 }
 787
 788 static int vma_get_pages(struct i915_vma *vma)
 789 {
 790         int err = 0;
 791
 792         if (atomic_add_unless(&vma->pages_count, 1, 0))
 793                 return 0;
 794
 795         /* Allocations ahoy! */
 796         if (mutex_lock_interruptible(&vma->pages_mutex))
 797                 return -EINTR;
 798
 799         if (!atomic_read(&vma->pages_count)) {
 800                 if (vma->obj) {
 801                         err = i915_gem_object_pin_pages(vma->obj);
 802                         if (err)
 803                                 goto unlock;
 804                 }
 805
 806                 err = vma->ops->set_pages(vma);
 807                 if (err) {
 808                         if (vma->obj)
 809                                 i915_gem_object_unpin_pages(vma->obj);
 810                         goto unlock;
 811                 }
 812         }
 813         atomic_inc(&vma->pages_count);
 814
 815 unlock:
 816         mutex_unlock(&vma->pages_mutex);
 817
 818         return err;
 819 }
 820
 821 static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
 822 {
 823         /* We allocate under vma_get_pages, so beware the shrinker */
 824         mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING);
 825         GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
 826         if (atomic_sub_return(count, &vma->pages_count) == 0) {
 827                 vma->ops->clear_pages(vma);
 828                 GEM_BUG_ON(vma->pages);
 829                 if (vma->obj)
 830                         i915_gem_object_unpin_pages(vma->obj);
 831         }
 832         mutex_unlock(&vma->pages_mutex);
 833 }
 834
 835 static void vma_put_pages(struct i915_vma *vma)
 836 {
 837         if (atomic_add_unless(&vma->pages_count, -1, 1))
 838                 return;
 839
 840         __vma_put_pages(vma, 1);
 841 }
 842
 843 static void vma_unbind_pages(struct i915_vma *vma)
 844 {
 845         unsigned int count;
 846
 847         lockdep_assert_held(&vma->vm->mutex);
 848
 849         /* The upper portion of pages_count is the number of bindings */
 850         count = atomic_read(&vma->pages_count);
 851         count >>= I915_VMA_PAGES_BIAS;
 852         GEM_BUG_ON(!count);
 853
 854         __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
 855 }
 856
 857 int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
 858                     u64 size, u64 alignment, u64 flags)
 859 {
 860         struct i915_vma_work *work = NULL;
 861         intel_wakeref_t wakeref = 0;
 862         unsigned int bound;
 863         int err;
 864
 865 #ifdef CONFIG_PROVE_LOCKING
 866         if (debug_locks && lockdep_is_held(&vma->vm->i915->drm.struct_mutex))
 867                 WARN_ON(!ww);
 868 #endif
 869
 870         BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
 871         BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
 872
 873         GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
 874
 875         /* First try and grab the pin without rebinding the vma */
 876         if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
 877                 return 0;
 878
 879         err = vma_get_pages(vma);
 880         if (err)
 881                 return err;
 882
 883         if (flags & PIN_GLOBAL)
 884                 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
 885
 886         if (flags & vma->vm->bind_async_flags) {
 887                 work = i915_vma_work();
 888                 if (!work) {
 889                         err = -ENOMEM;
 890                         goto err_rpm;
 891                 }
 892
 893                 work->vm = i915_vm_get(vma->vm);
 894
 895                 /* Allocate enough page directories to used PTE */
 896                 if (vma->vm->allocate_va_range) {
 897                         err = i915_vm_alloc_pt_stash(vma->vm,
 898                                                      &work->stash,
 899                                                      vma->size);
 900                         if (err)
 901                                 goto err_fence;
 902
 903                         err = i915_vm_pin_pt_stash(vma->vm,
 904                                                    &work->stash);
 905                         if (err)
 906                                 goto err_fence;
 907                 }
 908         }
 909
 910         /*
 911          * Differentiate between user/kernel vma inside the aliasing-ppgtt.
 912          *
 913          * We conflate the Global GTT with the user's vma when using the
 914          * aliasing-ppgtt, but it is still vitally important to try and
 915          * keep the use cases distinct. For example, userptr objects are
 916          * not allowed inside the Global GTT as that will cause lock
 917          * inversions when we have to evict them the mmu_notifier callbacks -
 918          * but they are allowed to be part of the user ppGTT which can never
 919          * be mapped. As such we try to give the distinct users of the same
 920          * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt
 921          * and i915_ppgtt separate].
 922          *
 923          * NB this may cause us to mask real lock inversions -- while the
 924          * code is safe today, lockdep may not be able to spot future
 925          * transgressions.
 926          */
 927         err = mutex_lock_interruptible_nested(&vma->vm->mutex,
 928                                               !(flags & PIN_GLOBAL));
 929         if (err)
 930                 goto err_fence;
 931
 932         /* No more allocations allowed now we hold vm->mutex */
 933
 934         if (unlikely(i915_vma_is_closed(vma))) {
 935                 err = -ENOENT;
 936                 goto err_unlock;
 937         }
 938
 939         bound = atomic_read(&vma->flags);
 940         if (unlikely(bound & I915_VMA_ERROR)) {
 941                 err = -ENOMEM;
 942                 goto err_unlock;
 943         }
 944
 945         if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
 946                 err = -EAGAIN; /* pins are meant to be fairly temporary */
 947                 goto err_unlock;
 948         }
 949
 950         if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
 951                 __i915_vma_pin(vma);
 952                 goto err_unlock;
 953         }
 954
 955         err = i915_active_acquire(&vma->active);
 956         if (err)
 957                 goto err_unlock;
 958
 959         if (!(bound & I915_VMA_BIND_MASK)) {
 960                 err = i915_vma_insert(vma, size, alignment, flags);
 961                 if (err)
 962                         goto err_active;
 963
 964                 if (i915_is_ggtt(vma->vm))
 965                         __i915_vma_set_map_and_fenceable(vma);
 966         }
 967
 968         GEM_BUG_ON(!vma->pages);
 969         err = i915_vma_bind(vma,
 970                             vma->obj ? vma->obj->cache_level : 0,
 971                             flags, work);
 972         if (err)
 973                 goto err_remove;
 974
 975         /* There should only be at most 2 active bindings (user, global) */
 976         GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
 977         atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
 978         list_move_tail(&vma->vm_link, &vma->vm->bound_list);
 979
 980         __i915_vma_pin(vma);
 981         GEM_BUG_ON(!i915_vma_is_pinned(vma));
 982         GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
 983         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
 984
 985 err_remove:
 986         if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
 987                 i915_vma_detach(vma);
 988                 drm_mm_remove_node(&vma->node);
 989         }
 990 err_active:
 991         i915_active_release(&vma->active);
 992 err_unlock:
 993         mutex_unlock(&vma->vm->mutex);
 994 err_fence:
 995         if (work)
 996                 dma_fence_work_commit_imm(&work->base);
 997 err_rpm:
 998         if (wakeref)
 999                 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
1000         vma_put_pages(vma);
1001         return err;
1002 }
1003
1004 static void flush_idle_contexts(struct intel_gt *gt)
1005 {
1006         struct intel_engine_cs *engine;
1007         enum intel_engine_id id;
1008
1009         for_each_engine(engine, gt, id)
1010                 intel_engine_flush_barriers(engine);
1011
1012         intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
1013 }
1014
1015 int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
1016                   u32 align, unsigned int flags)
1017 {
1018         struct i915_address_space *vm = vma->vm;
1019         int err;
1020
1021         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
1022
1023         do {
1024                 err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL);
1025                 if (err != -ENOSPC) {
1026                         if (!err) {
1027                                 err = i915_vma_wait_for_bind(vma);
1028                                 if (err)
1029                                         i915_vma_unpin(vma);
1030                         }
1031                         return err;
1032                 }
1033
1034                 /* Unlike i915_vma_pin, we don't take no for an answer! */
1035                 flush_idle_contexts(vm->gt);
1036                 if (mutex_lock_interruptible(&vm->mutex) == 0) {
1037                         i915_gem_evict_vm(vm);
1038                         mutex_unlock(&vm->mutex);
1039                 }
1040         } while (1);
1041 }
1042
1043 static void __vma_close(struct i915_vma *vma, struct intel_gt *gt)
1044 {
1045         /*
1046          * We defer actually closing, unbinding and destroying the VMA until
1047          * the next idle point, or if the object is freed in the meantime. By
1048          * postponing the unbind, we allow for it to be resurrected by the
1049          * client, avoiding the work required to rebind the VMA. This is
1050          * advantageous for DRI, where the client/server pass objects
1051          * between themselves, temporarily opening a local VMA to the
1052          * object, and then closing it again. The same object is then reused
1053          * on the next frame (or two, depending on the depth of the swap queue)
1054          * causing us to rebind the VMA once more. This ends up being a lot
1055          * of wasted work for the steady state.
1056          */
1057         GEM_BUG_ON(i915_vma_is_closed(vma));
1058         list_add(&vma->closed_link, &gt->closed_vma);
1059 }
1060
1061 void i915_vma_close(struct i915_vma *vma)
1062 {
1063         struct intel_gt *gt = vma->vm->gt;
1064         unsigned long flags;
1065
1066         if (i915_vma_is_ggtt(vma))
1067                 return;
1068
1069         GEM_BUG_ON(!atomic_read(&vma->open_count));
1070         if (atomic_dec_and_lock_irqsave(&vma->open_count,
1071                                         &gt->closed_lock,
1072                                         flags)) {
1073                 __vma_close(vma, gt);
1074                 spin_unlock_irqrestore(&gt->closed_lock, flags);
1075         }
1076 }
1077
1078 static void __i915_vma_remove_closed(struct i915_vma *vma)
1079 {
1080         struct intel_gt *gt = vma->vm->gt;
1081
1082         spin_lock_irq(&gt->closed_lock);
1083         list_del_init(&vma->closed_link);
1084         spin_unlock_irq(&gt->closed_lock);
1085 }
1086
1087 void i915_vma_reopen(struct i915_vma *vma)
1088 {
1089         if (i915_vma_is_closed(vma))
1090                 __i915_vma_remove_closed(vma);
1091 }
1092
1093 void i915_vma_release(struct kref *ref)
1094 {
1095         struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
1096
1097         if (drm_mm_node_allocated(&vma->node)) {
1098                 mutex_lock(&vma->vm->mutex);
1099                 atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
1100                 WARN_ON(__i915_vma_unbind(vma));
1101                 mutex_unlock(&vma->vm->mutex);
1102                 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
1103         }
1104         GEM_BUG_ON(i915_vma_is_active(vma));
1105
1106         if (vma->obj) {
1107                 struct drm_i915_gem_object *obj = vma->obj;
1108
1109                 spin_lock(&obj->vma.lock);
1110                 list_del(&vma->obj_link);
1111                 if (!RB_EMPTY_NODE(&vma->obj_node))
1112                         rb_erase(&vma->obj_node, &obj->vma.tree);
1113                 spin_unlock(&obj->vma.lock);
1114         }
1115
1116         __i915_vma_remove_closed(vma);
1117         i915_vm_put(vma->vm);
1118
1119         i915_active_fini(&vma->active);
1120         i915_vma_free(vma);
1121 }
1122
1123 void i915_vma_parked(struct intel_gt *gt)
1124 {
1125         struct i915_vma *vma, *next;
1126         LIST_HEAD(closed);
1127
1128         spin_lock_irq(&gt->closed_lock);
1129         list_for_each_entry_safe(vma, next, &gt->closed_vma, closed_link) {
1130                 struct drm_i915_gem_object *obj = vma->obj;
1131                 struct i915_address_space *vm = vma->vm;
1132
1133                 /* XXX All to avoid keeping a reference on i915_vma itself */
1134
1135                 if (!kref_get_unless_zero(&obj->base.refcount))
1136                         continue;
1137
1138                 if (!i915_vm_tryopen(vm)) {
1139                         i915_gem_object_put(obj);
1140                         continue;
1141                 }
1142
1143                 list_move(&vma->closed_link, &closed);
1144         }
1145         spin_unlock_irq(&gt->closed_lock);
1146
1147         /* As the GT is held idle, no vma can be reopened as we destroy them */
1148         list_for_each_entry_safe(vma, next, &closed, closed_link) {
1149                 struct drm_i915_gem_object *obj = vma->obj;
1150                 struct i915_address_space *vm = vma->vm;
1151
1152                 INIT_LIST_HEAD(&vma->closed_link);
1153                 __i915_vma_put(vma);
1154
1155                 i915_gem_object_put(obj);
1156                 i915_vm_close(vm);
1157         }
1158 }
1159
1160 static void __i915_vma_iounmap(struct i915_vma *vma)
1161 {
1162         GEM_BUG_ON(i915_vma_is_pinned(vma));
1163
1164         if (vma->iomap == NULL)
1165                 return;
1166
1167         io_mapping_unmap(vma->iomap);
1168         vma->iomap = NULL;
1169 }
1170
1171 void i915_vma_revoke_mmap(struct i915_vma *vma)
1172 {
1173         struct drm_vma_offset_node *node;
1174         u64 vma_offset;
1175
1176         if (!i915_vma_has_userfault(vma))
1177                 return;
1178
1179         GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
1180         GEM_BUG_ON(!vma->obj->userfault_count);
1181
1182         node = &vma->mmo->vma_node;
1183         vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
1184         unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
1185                             drm_vma_node_offset_addr(node) + vma_offset,
1186                             vma->size,
1187                             1);
1188
1189         i915_vma_unset_userfault(vma);
1190         if (!--vma->obj->userfault_count)
1191                 list_del(&vma->obj->userfault_link);
1192 }
1193
1194 static int
1195 __i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma)
1196 {
1197         return __i915_request_await_exclusive(rq, &vma->active);
1198 }
1199
1200 int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
1201 {
1202         int err;
1203
1204         GEM_BUG_ON(!i915_vma_is_pinned(vma));
1205
1206         /* Wait for the vma to be bound before we start! */
1207         err = __i915_request_await_bind(rq, vma);
1208         if (err)
1209                 return err;
1210
1211         return i915_active_add_request(&vma->active, rq);
1212 }
1213
1214 int i915_vma_move_to_active(struct i915_vma *vma,
1215                             struct i915_request *rq,
1216                             unsigned int flags)
1217 {
1218         struct drm_i915_gem_object *obj = vma->obj;
1219         int err;
1220
1221         assert_object_held(obj);
1222
1223         err = __i915_vma_move_to_active(vma, rq);
1224         if (unlikely(err))
1225                 return err;
1226
1227         if (flags & EXEC_OBJECT_WRITE) {
1228                 struct intel_frontbuffer *front;
1229
1230                 front = __intel_frontbuffer_get(obj);
1231                 if (unlikely(front)) {
1232                         if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
1233                                 i915_active_add_request(&front->write, rq);
1234                         intel_frontbuffer_put(front);
1235                 }
1236
1237                 dma_resv_add_excl_fence(vma->resv, &rq->fence);
1238                 obj->write_domain = I915_GEM_DOMAIN_RENDER;
1239                 obj->read_domains = 0;
1240         } else {
1241                 err = dma_resv_reserve_shared(vma->resv, 1);
1242                 if (unlikely(err))
1243                         return err;
1244
1245                 dma_resv_add_shared_fence(vma->resv, &rq->fence);
1246                 obj->write_domain = 0;
1247         }
1248
1249         if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence)
1250                 i915_active_add_request(&vma->fence->active, rq);
1251
1252         obj->read_domains |= I915_GEM_GPU_DOMAINS;
1253         obj->mm.dirty = true;
1254
1255         GEM_BUG_ON(!i915_vma_is_active(vma));
1256         return 0;
1257 }
1258
1259 void __i915_vma_evict(struct i915_vma *vma)
1260 {
1261         GEM_BUG_ON(i915_vma_is_pinned(vma));
1262
1263         if (i915_vma_is_map_and_fenceable(vma)) {
1264                 /* Force a pagefault for domain tracking on next user access */
1265                 i915_vma_revoke_mmap(vma);
1266
1267                 /*
1268                  * Check that we have flushed all writes through the GGTT
1269                  * before the unbind, other due to non-strict nature of those
1270                  * indirect writes they may end up referencing the GGTT PTE
1271                  * after the unbind.
1272                  *
1273                  * Note that we may be concurrently poking at the GGTT_WRITE
1274                  * bit from set-domain, as we mark all GGTT vma associated
1275                  * with an object. We know this is for another vma, as we
1276                  * are currently unbinding this one -- so if this vma will be
1277                  * reused, it will be refaulted and have its dirty bit set
1278                  * before the next write.
1279                  */
1280                 i915_vma_flush_writes(vma);
1281
1282                 /* release the fence reg _after_ flushing */
1283                 i915_vma_revoke_fence(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->vm, 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
1302 int __i915_vma_unbind(struct i915_vma *vma)
1303 {
1304         int ret;
1305
1306         lockdep_assert_held(&vma->vm->mutex);
1307
1308         if (!drm_mm_node_allocated(&vma->node))
1309                 return 0;
1310
1311         if (i915_vma_is_pinned(vma)) {
1312                 vma_print_allocator(vma, "is pinned");
1313                 return -EAGAIN;
1314         }
1315
1316         /*
1317          * After confirming that no one else is pinning this vma, wait for
1318          * any laggards who may have crept in during the wait (through
1319          * a residual pin skipping the vm->mutex) to complete.
1320          */
1321         ret = i915_vma_sync(vma);
1322         if (ret)
1323                 return ret;
1324
1325         GEM_BUG_ON(i915_vma_is_active(vma));
1326         __i915_vma_evict(vma);
1327
1328         drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
1329         return 0;
1330 }
1331
1332 int i915_vma_unbind(struct i915_vma *vma)
1333 {
1334         struct i915_address_space *vm = vma->vm;
1335         intel_wakeref_t wakeref = 0;
1336         int err;
1337
1338         /* Optimistic wait before taking the mutex */
1339         err = i915_vma_sync(vma);
1340         if (err)
1341                 return err;
1342
1343         if (!drm_mm_node_allocated(&vma->node))
1344                 return 0;
1345
1346         if (i915_vma_is_pinned(vma)) {
1347                 vma_print_allocator(vma, "is pinned");
1348                 return -EAGAIN;
1349         }
1350
1351         if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
1352                 /* XXX not always required: nop_clear_range */
1353                 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
1354
1355         err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref);
1356         if (err)
1357                 goto out_rpm;
1358
1359         err = __i915_vma_unbind(vma);
1360         mutex_unlock(&vm->mutex);
1361
1362 out_rpm:
1363         if (wakeref)
1364                 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
1365         return err;
1366 }
1367
1368 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1369 {
1370         i915_gem_object_make_unshrinkable(vma->obj);
1371         return vma;
1372 }
1373
1374 void i915_vma_make_shrinkable(struct i915_vma *vma)
1375 {
1376         i915_gem_object_make_shrinkable(vma->obj);
1377 }
1378
1379 void i915_vma_make_purgeable(struct i915_vma *vma)
1380 {
1381         i915_gem_object_make_purgeable(vma->obj);
1382 }
1383
1384 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1385 #include "selftests/i915_vma.c"
1386 #endif
1387
1388 static void i915_global_vma_shrink(void)
1389 {
1390         kmem_cache_shrink(global.slab_vmas);
1391 }
1392
1393 static void i915_global_vma_exit(void)
1394 {
1395         kmem_cache_destroy(global.slab_vmas);
1396 }
1397
1398 static struct i915_global_vma global = { {
1399         .shrink = i915_global_vma_shrink,
1400         .exit = i915_global_vma_exit,
1401 } };
1402
1403 int __init i915_global_vma_init(void)
1404 {
1405         global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1406         if (!global.slab_vmas)
1407                 return -ENOMEM;
1408
1409         i915_global_register(&global.base);
1410         return 0;
1411 }