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
 320         i915_vm_free_pt_stash(vw->vm, &vw->stash);
 321         i915_vm_put(vw->vm);
 322 }
 323
 324 static const struct dma_fence_work_ops bind_ops = {
 325         .name = "bind",
 326         .work = __vma_bind,
 327         .release = __vma_release,
 328 };
 329
 330 struct i915_vma_work *i915_vma_work(void)
 331 {
 332         struct i915_vma_work *vw;
 333
 334         vw = kzalloc(sizeof(*vw), GFP_KERNEL);
 335         if (!vw)
 336                 return NULL;
 337
 338         dma_fence_work_init(&vw->base, &bind_ops);
 339         vw->base.dma.error = -EAGAIN; /* disable the worker by default */
 340
 341         return vw;
 342 }
 343
 344 int i915_vma_wait_for_bind(struct i915_vma *vma)
 345 {
 346         int err = 0;
 347
 348         if (rcu_access_pointer(vma->active.excl.fence)) {
 349                 struct dma_fence *fence;
 350
 351                 rcu_read_lock();
 352                 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence);
 353                 rcu_read_unlock();
 354                 if (fence) {
 355                         err = dma_fence_wait(fence, MAX_SCHEDULE_TIMEOUT);
 356                         dma_fence_put(fence);
 357                 }
 358         }
 359
 360         return err;
 361 }
 362
 363 /**
 364  * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
 365  * @vma: VMA to map
 366  * @cache_level: mapping cache level
 367  * @flags: flags like global or local mapping
 368  * @work: preallocated worker for allocating and binding the PTE
 369  *
 370  * DMA addresses are taken from the scatter-gather table of this object (or of
 371  * this VMA in case of non-default GGTT views) and PTE entries set up.
 372  * Note that DMA addresses are also the only part of the SG table we care about.
 373  */
 374 int i915_vma_bind(struct i915_vma *vma,
 375                   enum i915_cache_level cache_level,
 376                   u32 flags,
 377                   struct i915_vma_work *work)
 378 {
 379         u32 bind_flags;
 380         u32 vma_flags;
 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
 399         bind_flags &= ~vma_flags;
 400         if (bind_flags == 0)
 401                 return 0;
 402
 403         GEM_BUG_ON(!vma->pages);
 404
 405         trace_i915_vma_bind(vma, bind_flags);
 406         if (work && bind_flags & vma->vm->bind_async_flags) {
 407                 struct dma_fence *prev;
 408
 409                 work->vma = vma;
 410                 work->cache_level = cache_level;
 411                 work->flags = bind_flags;
 412
 413                 /*
 414                  * Note we only want to chain up to the migration fence on
 415                  * the pages (not the object itself). As we don't track that,
 416                  * yet, we have to use the exclusive fence instead.
 417                  *
 418                  * Also note that we do not want to track the async vma as
 419                  * part of the obj->resv->excl_fence as it only affects
 420                  * execution and not content or object's backing store lifetime.
 421                  */
 422                 prev = i915_active_set_exclusive(&vma->active, &work->base.dma);
 423                 if (prev) {
 424                         __i915_sw_fence_await_dma_fence(&work->base.chain,
 425                                                         prev,
 426                                                         &work->cb);
 427                         dma_fence_put(prev);
 428                 }
 429
 430                 work->base.dma.error = 0; /* enable the queue_work() */
 431
 432                 if (vma->obj) {
 433                         __i915_gem_object_pin_pages(vma->obj);
 434                         work->pinned = vma->obj;
 435                 }
 436         } else {
 437                 vma->ops->bind_vma(vma->vm, NULL, vma, cache_level, bind_flags);
 438         }
 439
 440         atomic_or(bind_flags, &vma->flags);
 441         return 0;
 442 }
 443
 444 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
 445 {
 446         void __iomem *ptr;
 447         int err;
 448
 449         if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
 450                 err = -ENODEV;
 451                 goto err;
 452         }
 453
 454         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 455         GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
 456
 457         ptr = READ_ONCE(vma->iomap);
 458         if (ptr == NULL) {
 459                 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
 460                                         vma->node.start,
 461                                         vma->node.size);
 462                 if (ptr == NULL) {
 463                         err = -ENOMEM;
 464                         goto err;
 465                 }
 466
 467                 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
 468                         io_mapping_unmap(ptr);
 469                         ptr = vma->iomap;
 470                 }
 471         }
 472
 473         __i915_vma_pin(vma);
 474
 475         err = i915_vma_pin_fence(vma);
 476         if (err)
 477                 goto err_unpin;
 478
 479         i915_vma_set_ggtt_write(vma);
 480
 481         /* NB Access through the GTT requires the device to be awake. */
 482         return ptr;
 483
 484 err_unpin:
 485         __i915_vma_unpin(vma);
 486 err:
 487         return IO_ERR_PTR(err);
 488 }
 489
 490 void i915_vma_flush_writes(struct i915_vma *vma)
 491 {
 492         if (i915_vma_unset_ggtt_write(vma))
 493                 intel_gt_flush_ggtt_writes(vma->vm->gt);
 494 }
 495
 496 void i915_vma_unpin_iomap(struct i915_vma *vma)
 497 {
 498         GEM_BUG_ON(vma->iomap == NULL);
 499
 500         i915_vma_flush_writes(vma);
 501
 502         i915_vma_unpin_fence(vma);
 503         i915_vma_unpin(vma);
 504 }
 505
 506 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
 507 {
 508         struct i915_vma *vma;
 509         struct drm_i915_gem_object *obj;
 510
 511         vma = fetch_and_zero(p_vma);
 512         if (!vma)
 513                 return;
 514
 515         obj = vma->obj;
 516         GEM_BUG_ON(!obj);
 517
 518         i915_vma_unpin(vma);
 519
 520         if (flags & I915_VMA_RELEASE_MAP)
 521                 i915_gem_object_unpin_map(obj);
 522
 523         i915_gem_object_put(obj);
 524 }
 525
 526 bool i915_vma_misplaced(const struct i915_vma *vma,
 527                         u64 size, u64 alignment, u64 flags)
 528 {
 529         if (!drm_mm_node_allocated(&vma->node))
 530                 return false;
 531
 532         if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
 533                 return true;
 534
 535         if (vma->node.size < size)
 536                 return true;
 537
 538         GEM_BUG_ON(alignment && !is_power_of_2(alignment));
 539         if (alignment && !IS_ALIGNED(vma->node.start, alignment))
 540                 return true;
 541
 542         if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
 543                 return true;
 544
 545         if (flags & PIN_OFFSET_BIAS &&
 546             vma->node.start < (flags & PIN_OFFSET_MASK))
 547                 return true;
 548
 549         if (flags & PIN_OFFSET_FIXED &&
 550             vma->node.start != (flags & PIN_OFFSET_MASK))
 551                 return true;
 552
 553         return false;
 554 }
 555
 556 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
 557 {
 558         bool mappable, fenceable;
 559
 560         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 561         GEM_BUG_ON(!vma->fence_size);
 562
 563         fenceable = (vma->node.size >= vma->fence_size &&
 564                      IS_ALIGNED(vma->node.start, vma->fence_alignment));
 565
 566         mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
 567
 568         if (mappable && fenceable)
 569                 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
 570         else
 571                 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
 572 }
 573
 574 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
 575 {
 576         struct drm_mm_node *node = &vma->node;
 577         struct drm_mm_node *other;
 578
 579         /*
 580          * On some machines we have to be careful when putting differing types
 581          * of snoopable memory together to avoid the prefetcher crossing memory
 582          * domains and dying. During vm initialisation, we decide whether or not
 583          * these constraints apply and set the drm_mm.color_adjust
 584          * appropriately.
 585          */
 586         if (!i915_vm_has_cache_coloring(vma->vm))
 587                 return true;
 588
 589         /* Only valid to be called on an already inserted vma */
 590         GEM_BUG_ON(!drm_mm_node_allocated(node));
 591         GEM_BUG_ON(list_empty(&node->node_list));
 592
 593         other = list_prev_entry(node, node_list);
 594         if (i915_node_color_differs(other, color) &&
 595             !drm_mm_hole_follows(other))
 596                 return false;
 597
 598         other = list_next_entry(node, node_list);
 599         if (i915_node_color_differs(other, color) &&
 600             !drm_mm_hole_follows(node))
 601                 return false;
 602
 603         return true;
 604 }
 605
 606 /**
 607  * i915_vma_insert - finds a slot for the vma in its address space
 608  * @vma: the vma
 609  * @size: requested size in bytes (can be larger than the VMA)
 610  * @alignment: required alignment
 611  * @flags: mask of PIN_* flags to use
 612  *
 613  * First we try to allocate some free space that meets the requirements for
 614  * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
 615  * preferrably the oldest idle entry to make room for the new VMA.
 616  *
 617  * Returns:
 618  * 0 on success, negative error code otherwise.
 619  */
 620 static int
 621 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 622 {
 623         unsigned long color;
 624         u64 start, end;
 625         int ret;
 626
 627         GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 628         GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
 629
 630         size = max(size, vma->size);
 631         alignment = max(alignment, vma->display_alignment);
 632         if (flags & PIN_MAPPABLE) {
 633                 size = max_t(typeof(size), size, vma->fence_size);
 634                 alignment = max_t(typeof(alignment),
 635                                   alignment, vma->fence_alignment);
 636         }
 637
 638         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
 639         GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
 640         GEM_BUG_ON(!is_power_of_2(alignment));
 641
 642         start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
 643         GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
 644
 645         end = vma->vm->total;
 646         if (flags & PIN_MAPPABLE)
 647                 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
 648         if (flags & PIN_ZONE_4G)
 649                 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
 650         GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
 651
 652         /* If binding the object/GGTT view requires more space than the entire
 653          * aperture has, reject it early before evicting everything in a vain
 654          * attempt to find space.
 655          */
 656         if (size > end) {
 657                 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
 658                           size, flags & PIN_MAPPABLE ? "mappable" : "total",
 659                           end);
 660                 return -ENOSPC;
 661         }
 662
 663         color = 0;
 664         if (vma->obj && i915_vm_has_cache_coloring(vma->vm))
 665                 color = vma->obj->cache_level;
 666
 667         if (flags & PIN_OFFSET_FIXED) {
 668                 u64 offset = flags & PIN_OFFSET_MASK;
 669                 if (!IS_ALIGNED(offset, alignment) ||
 670                     range_overflows(offset, size, end))
 671                         return -EINVAL;
 672
 673                 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
 674                                            size, offset, color,
 675                                            flags);
 676                 if (ret)
 677                         return ret;
 678         } else {
 679                 /*
 680                  * We only support huge gtt pages through the 48b PPGTT,
 681                  * however we also don't want to force any alignment for
 682                  * objects which need to be tightly packed into the low 32bits.
 683                  *
 684                  * Note that we assume that GGTT are limited to 4GiB for the
 685                  * forseeable future. See also i915_ggtt_offset().
 686                  */
 687                 if (upper_32_bits(end - 1) &&
 688                     vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
 689                         /*
 690                          * We can't mix 64K and 4K PTEs in the same page-table
 691                          * (2M block), and so to avoid the ugliness and
 692                          * complexity of coloring we opt for just aligning 64K
 693                          * objects to 2M.
 694                          */
 695                         u64 page_alignment =
 696                                 rounddown_pow_of_two(vma->page_sizes.sg |
 697                                                      I915_GTT_PAGE_SIZE_2M);
 698
 699                         /*
 700                          * Check we don't expand for the limited Global GTT
 701                          * (mappable aperture is even more precious!). This
 702                          * also checks that we exclude the aliasing-ppgtt.
 703                          */
 704                         GEM_BUG_ON(i915_vma_is_ggtt(vma));
 705
 706                         alignment = max(alignment, page_alignment);
 707
 708                         if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
 709                                 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
 710                 }
 711
 712                 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
 713                                           size, alignment, color,
 714                                           start, end, flags);
 715                 if (ret)
 716                         return ret;
 717
 718                 GEM_BUG_ON(vma->node.start < start);
 719                 GEM_BUG_ON(vma->node.start + vma->node.size > end);
 720         }
 721         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 722         GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
 723
 724         list_add_tail(&vma->vm_link, &vma->vm->bound_list);
 725
 726         return 0;
 727 }
 728
 729 static void
 730 i915_vma_detach(struct i915_vma *vma)
 731 {
 732         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 733         GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 734
 735         /*
 736          * And finally now the object is completely decoupled from this
 737          * vma, we can drop its hold on the backing storage and allow
 738          * it to be reaped by the shrinker.
 739          */
 740         list_del(&vma->vm_link);
 741 }
 742
 743 static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
 744 {
 745         unsigned int bound;
 746         bool pinned = true;
 747
 748         bound = atomic_read(&vma->flags);
 749         do {
 750                 if (unlikely(flags & ~bound))
 751                         return false;
 752
 753                 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
 754                         return false;
 755
 756                 if (!(bound & I915_VMA_PIN_MASK))
 757                         goto unpinned;
 758
 759                 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
 760         } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
 761
 762         return true;
 763
 764 unpinned:
 765         /*
 766          * If pin_count==0, but we are bound, check under the lock to avoid
 767          * racing with a concurrent i915_vma_unbind().
 768          */
 769         mutex_lock(&vma->vm->mutex);
 770         do {
 771                 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) {
 772                         pinned = false;
 773                         break;
 774                 }
 775
 776                 if (unlikely(flags & ~bound)) {
 777                         pinned = false;
 778                         break;
 779                 }
 780         } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
 781         mutex_unlock(&vma->vm->mutex);
 782
 783         return pinned;
 784 }
 785
 786 static int vma_get_pages(struct i915_vma *vma)
 787 {
 788         int err = 0;
 789
 790         if (atomic_add_unless(&vma->pages_count, 1, 0))
 791                 return 0;
 792
 793         /* Allocations ahoy! */
 794         if (mutex_lock_interruptible(&vma->pages_mutex))
 795                 return -EINTR;
 796
 797         if (!atomic_read(&vma->pages_count)) {
 798                 if (vma->obj) {
 799                         err = i915_gem_object_pin_pages(vma->obj);
 800                         if (err)
 801                                 goto unlock;
 802                 }
 803
 804                 err = vma->ops->set_pages(vma);
 805                 if (err) {
 806                         if (vma->obj)
 807                                 i915_gem_object_unpin_pages(vma->obj);
 808                         goto unlock;
 809                 }
 810         }
 811         atomic_inc(&vma->pages_count);
 812
 813 unlock:
 814         mutex_unlock(&vma->pages_mutex);
 815
 816         return err;
 817 }
 818
 819 static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
 820 {
 821         /* We allocate under vma_get_pages, so beware the shrinker */
 822         mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING);
 823         GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
 824         if (atomic_sub_return(count, &vma->pages_count) == 0) {
 825                 vma->ops->clear_pages(vma);
 826                 GEM_BUG_ON(vma->pages);
 827                 if (vma->obj)
 828                         i915_gem_object_unpin_pages(vma->obj);
 829         }
 830         mutex_unlock(&vma->pages_mutex);
 831 }
 832
 833 static void vma_put_pages(struct i915_vma *vma)
 834 {
 835         if (atomic_add_unless(&vma->pages_count, -1, 1))
 836                 return;
 837
 838         __vma_put_pages(vma, 1);
 839 }
 840
 841 static void vma_unbind_pages(struct i915_vma *vma)
 842 {
 843         unsigned int count;
 844
 845         lockdep_assert_held(&vma->vm->mutex);
 846
 847         /* The upper portion of pages_count is the number of bindings */
 848         count = atomic_read(&vma->pages_count);
 849         count >>= I915_VMA_PAGES_BIAS;
 850         GEM_BUG_ON(!count);
 851
 852         __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
 853 }
 854
 855 int i915_vma_pin(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 856 {
 857         struct i915_vma_work *work = NULL;
 858         intel_wakeref_t wakeref = 0;
 859         unsigned int bound;
 860         int err;
 861
 862         BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
 863         BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
 864
 865         GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
 866
 867         /* First try and grab the pin without rebinding the vma */
 868         if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
 869                 return 0;
 870
 871         err = vma_get_pages(vma);
 872         if (err)
 873                 return err;
 874
 875         if (flags & vma->vm->bind_async_flags) {
 876                 work = i915_vma_work();
 877                 if (!work) {
 878                         err = -ENOMEM;
 879                         goto err_pages;
 880                 }
 881
 882                 work->vm = i915_vm_get(vma->vm);
 883
 884                 /* Allocate enough page directories to used PTE */
 885                 if (vma->vm->allocate_va_range)
 886                         i915_vm_alloc_pt_stash(vma->vm,
 887                                                &work->stash,
 888                                                vma->size);
 889         }
 890
 891         if (flags & PIN_GLOBAL)
 892                 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
 893
 894         /*
 895          * Differentiate between user/kernel vma inside the aliasing-ppgtt.
 896          *
 897          * We conflate the Global GTT with the user's vma when using the
 898          * aliasing-ppgtt, but it is still vitally important to try and
 899          * keep the use cases distinct. For example, userptr objects are
 900          * not allowed inside the Global GTT as that will cause lock
 901          * inversions when we have to evict them the mmu_notifier callbacks -
 902          * but they are allowed to be part of the user ppGTT which can never
 903          * be mapped. As such we try to give the distinct users of the same
 904          * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt
 905          * and i915_ppgtt separate].
 906          *
 907          * NB this may cause us to mask real lock inversions -- while the
 908          * code is safe today, lockdep may not be able to spot future
 909          * transgressions.
 910          */
 911         err = mutex_lock_interruptible_nested(&vma->vm->mutex,
 912                                               !(flags & PIN_GLOBAL));
 913         if (err)
 914                 goto err_fence;
 915
 916         /* No more allocations allowed now we hold vm->mutex */
 917
 918         if (unlikely(i915_vma_is_closed(vma))) {
 919                 err = -ENOENT;
 920                 goto err_unlock;
 921         }
 922
 923         bound = atomic_read(&vma->flags);
 924         if (unlikely(bound & I915_VMA_ERROR)) {
 925                 err = -ENOMEM;
 926                 goto err_unlock;
 927         }
 928
 929         if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
 930                 err = -EAGAIN; /* pins are meant to be fairly temporary */
 931                 goto err_unlock;
 932         }
 933
 934         if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
 935                 __i915_vma_pin(vma);
 936                 goto err_unlock;
 937         }
 938
 939         err = i915_active_acquire(&vma->active);
 940         if (err)
 941                 goto err_unlock;
 942
 943         if (!(bound & I915_VMA_BIND_MASK)) {
 944                 err = i915_vma_insert(vma, size, alignment, flags);
 945                 if (err)
 946                         goto err_active;
 947
 948                 if (i915_is_ggtt(vma->vm))
 949                         __i915_vma_set_map_and_fenceable(vma);
 950         }
 951
 952         GEM_BUG_ON(!vma->pages);
 953         err = i915_vma_bind(vma,
 954                             vma->obj ? vma->obj->cache_level : 0,
 955                             flags, work);
 956         if (err)
 957                 goto err_remove;
 958
 959         /* There should only be at most 2 active bindings (user, global) */
 960         GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
 961         atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
 962         list_move_tail(&vma->vm_link, &vma->vm->bound_list);
 963
 964         __i915_vma_pin(vma);
 965         GEM_BUG_ON(!i915_vma_is_pinned(vma));
 966         GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
 967         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
 968
 969 err_remove:
 970         if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
 971                 i915_vma_detach(vma);
 972                 drm_mm_remove_node(&vma->node);
 973         }
 974 err_active:
 975         i915_active_release(&vma->active);
 976 err_unlock:
 977         mutex_unlock(&vma->vm->mutex);
 978 err_fence:
 979         if (work)
 980                 dma_fence_work_commit_imm(&work->base);
 981         if (wakeref)
 982                 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
 983 err_pages:
 984         vma_put_pages(vma);
 985         return err;
 986 }
 987
 988 static void flush_idle_contexts(struct intel_gt *gt)
 989 {
 990         struct intel_engine_cs *engine;
 991         enum intel_engine_id id;
 992
 993         for_each_engine(engine, gt, id)
 994                 intel_engine_flush_barriers(engine);
 995
 996         intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
 997 }
 998
 999 int i915_ggtt_pin(struct i915_vma *vma, u32 align, unsigned int flags)
1000 {
1001         struct i915_address_space *vm = vma->vm;
1002         int err;
1003
1004         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
1005
1006         do {
1007                 err = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL);
1008                 if (err != -ENOSPC) {
1009                         if (!err) {
1010                                 err = i915_vma_wait_for_bind(vma);
1011                                 if (err)
1012                                         i915_vma_unpin(vma);
1013                         }
1014                         return err;
1015                 }
1016
1017                 /* Unlike i915_vma_pin, we don't take no for an answer! */
1018                 flush_idle_contexts(vm->gt);
1019                 if (mutex_lock_interruptible(&vm->mutex) == 0) {
1020                         i915_gem_evict_vm(vm);
1021                         mutex_unlock(&vm->mutex);
1022                 }
1023         } while (1);
1024 }
1025
1026 static void __vma_close(struct i915_vma *vma, struct intel_gt *gt)
1027 {
1028         /*
1029          * We defer actually closing, unbinding and destroying the VMA until
1030          * the next idle point, or if the object is freed in the meantime. By
1031          * postponing the unbind, we allow for it to be resurrected by the
1032          * client, avoiding the work required to rebind the VMA. This is
1033          * advantageous for DRI, where the client/server pass objects
1034          * between themselves, temporarily opening a local VMA to the
1035          * object, and then closing it again. The same object is then reused
1036          * on the next frame (or two, depending on the depth of the swap queue)
1037          * causing us to rebind the VMA once more. This ends up being a lot
1038          * of wasted work for the steady state.
1039          */
1040         GEM_BUG_ON(i915_vma_is_closed(vma));
1041         list_add(&vma->closed_link, &gt->closed_vma);
1042 }
1043
1044 void i915_vma_close(struct i915_vma *vma)
1045 {
1046         struct intel_gt *gt = vma->vm->gt;
1047         unsigned long flags;
1048
1049         if (i915_vma_is_ggtt(vma))
1050                 return;
1051
1052         GEM_BUG_ON(!atomic_read(&vma->open_count));
1053         if (atomic_dec_and_lock_irqsave(&vma->open_count,
1054                                         &gt->closed_lock,
1055                                         flags)) {
1056                 __vma_close(vma, gt);
1057                 spin_unlock_irqrestore(&gt->closed_lock, flags);
1058         }
1059 }
1060
1061 static void __i915_vma_remove_closed(struct i915_vma *vma)
1062 {
1063         struct intel_gt *gt = vma->vm->gt;
1064
1065         spin_lock_irq(&gt->closed_lock);
1066         list_del_init(&vma->closed_link);
1067         spin_unlock_irq(&gt->closed_lock);
1068 }
1069
1070 void i915_vma_reopen(struct i915_vma *vma)
1071 {
1072         if (i915_vma_is_closed(vma))
1073                 __i915_vma_remove_closed(vma);
1074 }
1075
1076 void i915_vma_release(struct kref *ref)
1077 {
1078         struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
1079
1080         if (drm_mm_node_allocated(&vma->node)) {
1081                 mutex_lock(&vma->vm->mutex);
1082                 atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
1083                 WARN_ON(__i915_vma_unbind(vma));
1084                 mutex_unlock(&vma->vm->mutex);
1085                 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
1086         }
1087         GEM_BUG_ON(i915_vma_is_active(vma));
1088
1089         if (vma->obj) {
1090                 struct drm_i915_gem_object *obj = vma->obj;
1091
1092                 spin_lock(&obj->vma.lock);
1093                 list_del(&vma->obj_link);
1094                 if (!RB_EMPTY_NODE(&vma->obj_node))
1095                         rb_erase(&vma->obj_node, &obj->vma.tree);
1096                 spin_unlock(&obj->vma.lock);
1097         }
1098
1099         __i915_vma_remove_closed(vma);
1100         i915_vm_put(vma->vm);
1101
1102         i915_active_fini(&vma->active);
1103         i915_vma_free(vma);
1104 }
1105
1106 void i915_vma_parked(struct intel_gt *gt)
1107 {
1108         struct i915_vma *vma, *next;
1109         LIST_HEAD(closed);
1110
1111         spin_lock_irq(&gt->closed_lock);
1112         list_for_each_entry_safe(vma, next, &gt->closed_vma, closed_link) {
1113                 struct drm_i915_gem_object *obj = vma->obj;
1114                 struct i915_address_space *vm = vma->vm;
1115
1116                 /* XXX All to avoid keeping a reference on i915_vma itself */
1117
1118                 if (!kref_get_unless_zero(&obj->base.refcount))
1119                         continue;
1120
1121                 if (!i915_vm_tryopen(vm)) {
1122                         i915_gem_object_put(obj);
1123                         continue;
1124                 }
1125
1126                 list_move(&vma->closed_link, &closed);
1127         }
1128         spin_unlock_irq(&gt->closed_lock);
1129
1130         /* As the GT is held idle, no vma can be reopened as we destroy them */
1131         list_for_each_entry_safe(vma, next, &closed, closed_link) {
1132                 struct drm_i915_gem_object *obj = vma->obj;
1133                 struct i915_address_space *vm = vma->vm;
1134
1135                 INIT_LIST_HEAD(&vma->closed_link);
1136                 __i915_vma_put(vma);
1137
1138                 i915_gem_object_put(obj);
1139                 i915_vm_close(vm);
1140         }
1141 }
1142
1143 static void __i915_vma_iounmap(struct i915_vma *vma)
1144 {
1145         GEM_BUG_ON(i915_vma_is_pinned(vma));
1146
1147         if (vma->iomap == NULL)
1148                 return;
1149
1150         io_mapping_unmap(vma->iomap);
1151         vma->iomap = NULL;
1152 }
1153
1154 void i915_vma_revoke_mmap(struct i915_vma *vma)
1155 {
1156         struct drm_vma_offset_node *node;
1157         u64 vma_offset;
1158
1159         if (!i915_vma_has_userfault(vma))
1160                 return;
1161
1162         GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
1163         GEM_BUG_ON(!vma->obj->userfault_count);
1164
1165         node = &vma->mmo->vma_node;
1166         vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
1167         unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
1168                             drm_vma_node_offset_addr(node) + vma_offset,
1169                             vma->size,
1170                             1);
1171
1172         i915_vma_unset_userfault(vma);
1173         if (!--vma->obj->userfault_count)
1174                 list_del(&vma->obj->userfault_link);
1175 }
1176
1177 static int
1178 __i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma)
1179 {
1180         return __i915_request_await_exclusive(rq, &vma->active);
1181 }
1182
1183 int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
1184 {
1185         int err;
1186
1187         GEM_BUG_ON(!i915_vma_is_pinned(vma));
1188
1189         /* Wait for the vma to be bound before we start! */
1190         err = __i915_request_await_bind(rq, vma);
1191         if (err)
1192                 return err;
1193
1194         return i915_active_add_request(&vma->active, rq);
1195 }
1196
1197 int i915_vma_move_to_active(struct i915_vma *vma,
1198                             struct i915_request *rq,
1199                             unsigned int flags)
1200 {
1201         struct drm_i915_gem_object *obj = vma->obj;
1202         int err;
1203
1204         assert_object_held(obj);
1205
1206         err = __i915_vma_move_to_active(vma, rq);
1207         if (unlikely(err))
1208                 return err;
1209
1210         if (flags & EXEC_OBJECT_WRITE) {
1211                 struct intel_frontbuffer *front;
1212
1213                 front = __intel_frontbuffer_get(obj);
1214                 if (unlikely(front)) {
1215                         if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
1216                                 i915_active_add_request(&front->write, rq);
1217                         intel_frontbuffer_put(front);
1218                 }
1219
1220                 dma_resv_add_excl_fence(vma->resv, &rq->fence);
1221                 obj->write_domain = I915_GEM_DOMAIN_RENDER;
1222                 obj->read_domains = 0;
1223         } else {
1224                 err = dma_resv_reserve_shared(vma->resv, 1);
1225                 if (unlikely(err))
1226                         return err;
1227
1228                 dma_resv_add_shared_fence(vma->resv, &rq->fence);
1229                 obj->write_domain = 0;
1230         }
1231
1232         if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence)
1233                 i915_active_add_request(&vma->fence->active, rq);
1234
1235         obj->read_domains |= I915_GEM_GPU_DOMAINS;
1236         obj->mm.dirty = true;
1237
1238         GEM_BUG_ON(!i915_vma_is_active(vma));
1239         return 0;
1240 }
1241
1242 void __i915_vma_evict(struct i915_vma *vma)
1243 {
1244         GEM_BUG_ON(i915_vma_is_pinned(vma));
1245
1246         if (i915_vma_is_map_and_fenceable(vma)) {
1247                 /* Force a pagefault for domain tracking on next user access */
1248                 i915_vma_revoke_mmap(vma);
1249
1250                 /*
1251                  * Check that we have flushed all writes through the GGTT
1252                  * before the unbind, other due to non-strict nature of those
1253                  * indirect writes they may end up referencing the GGTT PTE
1254                  * after the unbind.
1255                  *
1256                  * Note that we may be concurrently poking at the GGTT_WRITE
1257                  * bit from set-domain, as we mark all GGTT vma associated
1258                  * with an object. We know this is for another vma, as we
1259                  * are currently unbinding this one -- so if this vma will be
1260                  * reused, it will be refaulted and have its dirty bit set
1261                  * before the next write.
1262                  */
1263                 i915_vma_flush_writes(vma);
1264
1265                 /* release the fence reg _after_ flushing */
1266                 i915_vma_revoke_fence(vma);
1267
1268                 __i915_vma_iounmap(vma);
1269                 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
1270         }
1271         GEM_BUG_ON(vma->fence);
1272         GEM_BUG_ON(i915_vma_has_userfault(vma));
1273
1274         if (likely(atomic_read(&vma->vm->open))) {
1275                 trace_i915_vma_unbind(vma);
1276                 vma->ops->unbind_vma(vma->vm, vma);
1277         }
1278         atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE),
1279                    &vma->flags);
1280
1281         i915_vma_detach(vma);
1282         vma_unbind_pages(vma);
1283 }
1284
1285 int __i915_vma_unbind(struct i915_vma *vma)
1286 {
1287         int ret;
1288
1289         lockdep_assert_held(&vma->vm->mutex);
1290
1291         if (!drm_mm_node_allocated(&vma->node))
1292                 return 0;
1293
1294         if (i915_vma_is_pinned(vma)) {
1295                 vma_print_allocator(vma, "is pinned");
1296                 return -EAGAIN;
1297         }
1298
1299         /*
1300          * After confirming that no one else is pinning this vma, wait for
1301          * any laggards who may have crept in during the wait (through
1302          * a residual pin skipping the vm->mutex) to complete.
1303          */
1304         ret = i915_vma_sync(vma);
1305         if (ret)
1306                 return ret;
1307
1308         GEM_BUG_ON(i915_vma_is_active(vma));
1309         __i915_vma_evict(vma);
1310
1311         drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
1312         return 0;
1313 }
1314
1315 int i915_vma_unbind(struct i915_vma *vma)
1316 {
1317         struct i915_address_space *vm = vma->vm;
1318         intel_wakeref_t wakeref = 0;
1319         int err;
1320
1321         /* Optimistic wait before taking the mutex */
1322         err = i915_vma_sync(vma);
1323         if (err)
1324                 return err;
1325
1326         if (!drm_mm_node_allocated(&vma->node))
1327                 return 0;
1328
1329         if (i915_vma_is_pinned(vma)) {
1330                 vma_print_allocator(vma, "is pinned");
1331                 return -EAGAIN;
1332         }
1333
1334         if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
1335                 /* XXX not always required: nop_clear_range */
1336                 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
1337
1338         err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref);
1339         if (err)
1340                 goto out_rpm;
1341
1342         err = __i915_vma_unbind(vma);
1343         mutex_unlock(&vm->mutex);
1344
1345 out_rpm:
1346         if (wakeref)
1347                 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
1348         return err;
1349 }
1350
1351 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1352 {
1353         i915_gem_object_make_unshrinkable(vma->obj);
1354         return vma;
1355 }
1356
1357 void i915_vma_make_shrinkable(struct i915_vma *vma)
1358 {
1359         i915_gem_object_make_shrinkable(vma->obj);
1360 }
1361
1362 void i915_vma_make_purgeable(struct i915_vma *vma)
1363 {
1364         i915_gem_object_make_purgeable(vma->obj);
1365 }
1366
1367 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1368 #include "selftests/i915_vma.c"
1369 #endif
1370
1371 static void i915_global_vma_shrink(void)
1372 {
1373         kmem_cache_shrink(global.slab_vmas);
1374 }
1375
1376 static void i915_global_vma_exit(void)
1377 {
1378         kmem_cache_destroy(global.slab_vmas);
1379 }
1380
1381 static struct i915_global_vma global = { {
1382         .shrink = i915_global_vma_shrink,
1383         .exit = i915_global_vma_exit,
1384 } };
1385
1386 int __init i915_global_vma_init(void)
1387 {
1388         global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1389         if (!global.slab_vmas)
1390                 return -ENOMEM;
1391
1392         i915_global_register(&global.base);
1393         return 0;
1394 }