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 "gem/i915_gem_lmem.h"
  31 #include "gt/intel_engine.h"
  32 #include "gt/intel_engine_heartbeat.h"
  33 #include "gt/intel_gt.h"
  34 #include "gt/intel_gt_requests.h"
  35
  36 #include "i915_drv.h"
  37 #include "i915_sw_fence_work.h"
  38 #include "i915_trace.h"
  39 #include "i915_vma.h"
  40
  41 static struct kmem_cache *slab_vmas;
  42
  43 struct i915_vma *i915_vma_alloc(void)
  44 {
  45         return kmem_cache_zalloc(slab_vmas, GFP_KERNEL);
  46 }
  47
  48 void i915_vma_free(struct i915_vma *vma)
  49 {
  50         return kmem_cache_free(slab_vmas, vma);
  51 }
  52
  53 #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM)
  54
  55 #include <linux/stackdepot.h>
  56
  57 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
  58 {
  59         char buf[512];
  60
  61         if (!vma->node.stack) {
  62                 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",
  63                                  vma->node.start, vma->node.size, reason);
  64                 return;
  65         }
  66
  67         stack_depot_snprint(vma->node.stack, buf, sizeof(buf), 0);
  68         DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",
  69                          vma->node.start, vma->node.size, reason, buf);
  70 }
  71
  72 #else
  73
  74 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
  75 {
  76 }
  77
  78 #endif
  79
  80 static inline struct i915_vma *active_to_vma(struct i915_active *ref)
  81 {
  82         return container_of(ref, typeof(struct i915_vma), active);
  83 }
  84
  85 static int __i915_vma_active(struct i915_active *ref)
  86 {
  87         return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT;
  88 }
  89
  90 static void __i915_vma_retire(struct i915_active *ref)
  91 {
  92         i915_vma_put(active_to_vma(ref));
  93 }
  94
  95 static struct i915_vma *
  96 vma_create(struct drm_i915_gem_object *obj,
  97            struct i915_address_space *vm,
  98            const struct i915_ggtt_view *view)
  99 {
 100         struct i915_vma *pos = ERR_PTR(-E2BIG);
 101         struct i915_vma *vma;
 102         struct rb_node *rb, **p;
 103
 104         /* The aliasing_ppgtt should never be used directly! */
 105         GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm);
 106
 107         vma = i915_vma_alloc();
 108         if (vma == NULL)
 109                 return ERR_PTR(-ENOMEM);
 110
 111         kref_init(&vma->ref);
 112         mutex_init(&vma->pages_mutex);
 113         vma->vm = i915_vm_get(vm);
 114         vma->ops = &vm->vma_ops;
 115         vma->obj = obj;
 116         vma->resv = obj->base.resv;
 117         vma->size = obj->base.size;
 118         vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
 119
 120         i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire, 0);
 121
 122         /* Declare ourselves safe for use inside shrinkers */
 123         if (IS_ENABLED(CONFIG_LOCKDEP)) {
 124                 fs_reclaim_acquire(GFP_KERNEL);
 125                 might_lock(&vma->active.mutex);
 126                 fs_reclaim_release(GFP_KERNEL);
 127         }
 128
 129         INIT_LIST_HEAD(&vma->closed_link);
 130
 131         if (view && view->type != I915_GGTT_VIEW_NORMAL) {
 132                 vma->ggtt_view = *view;
 133                 if (view->type == I915_GGTT_VIEW_PARTIAL) {
 134                         GEM_BUG_ON(range_overflows_t(u64,
 135                                                      view->partial.offset,
 136                                                      view->partial.size,
 137                                                      obj->base.size >> PAGE_SHIFT));
 138                         vma->size = view->partial.size;
 139                         vma->size <<= PAGE_SHIFT;
 140                         GEM_BUG_ON(vma->size > obj->base.size);
 141                 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
 142                         vma->size = intel_rotation_info_size(&view->rotated);
 143                         vma->size <<= PAGE_SHIFT;
 144                 } else if (view->type == I915_GGTT_VIEW_REMAPPED) {
 145                         vma->size = intel_remapped_info_size(&view->remapped);
 146                         vma->size <<= PAGE_SHIFT;
 147                 }
 148         }
 149
 150         if (unlikely(vma->size > vm->total))
 151                 goto err_vma;
 152
 153         GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
 154
 155         spin_lock(&obj->vma.lock);
 156
 157         if (i915_is_ggtt(vm)) {
 158                 if (unlikely(overflows_type(vma->size, u32)))
 159                         goto err_unlock;
 160
 161                 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
 162                                                       i915_gem_object_get_tiling(obj),
 163                                                       i915_gem_object_get_stride(obj));
 164                 if (unlikely(vma->fence_size < vma->size || /* overflow */
 165                              vma->fence_size > vm->total))
 166                         goto err_unlock;
 167
 168                 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
 169
 170                 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
 171                                                                 i915_gem_object_get_tiling(obj),
 172                                                                 i915_gem_object_get_stride(obj));
 173                 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
 174
 175                 __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma));
 176         }
 177
 178         rb = NULL;
 179         p = &obj->vma.tree.rb_node;
 180         while (*p) {
 181                 long cmp;
 182
 183                 rb = *p;
 184                 pos = rb_entry(rb, struct i915_vma, obj_node);
 185
 186                 /*
 187                  * If the view already exists in the tree, another thread
 188                  * already created a matching vma, so return the older instance
 189                  * and dispose of ours.
 190                  */
 191                 cmp = i915_vma_compare(pos, vm, view);
 192                 if (cmp < 0)
 193                         p = &rb->rb_right;
 194                 else if (cmp > 0)
 195                         p = &rb->rb_left;
 196                 else
 197                         goto err_unlock;
 198         }
 199         rb_link_node(&vma->obj_node, rb, p);
 200         rb_insert_color(&vma->obj_node, &obj->vma.tree);
 201
 202         if (i915_vma_is_ggtt(vma))
 203                 /*
 204                  * We put the GGTT vma at the start of the vma-list, followed
 205                  * by the ppGGTT vma. This allows us to break early when
 206                  * iterating over only the GGTT vma for an object, see
 207                  * for_each_ggtt_vma()
 208                  */
 209                 list_add(&vma->obj_link, &obj->vma.list);
 210         else
 211                 list_add_tail(&vma->obj_link, &obj->vma.list);
 212
 213         spin_unlock(&obj->vma.lock);
 214
 215         return vma;
 216
 217 err_unlock:
 218         spin_unlock(&obj->vma.lock);
 219 err_vma:
 220         i915_vm_put(vm);
 221         i915_vma_free(vma);
 222         return pos;
 223 }
 224
 225 static struct i915_vma *
 226 i915_vma_lookup(struct drm_i915_gem_object *obj,
 227            struct i915_address_space *vm,
 228            const struct i915_ggtt_view *view)
 229 {
 230         struct rb_node *rb;
 231
 232         rb = obj->vma.tree.rb_node;
 233         while (rb) {
 234                 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
 235                 long cmp;
 236
 237                 cmp = i915_vma_compare(vma, vm, view);
 238                 if (cmp == 0)
 239                         return vma;
 240
 241                 if (cmp < 0)
 242                         rb = rb->rb_right;
 243                 else
 244                         rb = rb->rb_left;
 245         }
 246
 247         return NULL;
 248 }
 249
 250 /**
 251  * i915_vma_instance - return the singleton instance of the VMA
 252  * @obj: parent &struct drm_i915_gem_object to be mapped
 253  * @vm: address space in which the mapping is located
 254  * @view: additional mapping requirements
 255  *
 256  * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
 257  * the same @view characteristics. If a match is not found, one is created.
 258  * Once created, the VMA is kept until either the object is freed, or the
 259  * address space is closed.
 260  *
 261  * Returns the vma, or an error pointer.
 262  */
 263 struct i915_vma *
 264 i915_vma_instance(struct drm_i915_gem_object *obj,
 265                   struct i915_address_space *vm,
 266                   const struct i915_ggtt_view *view)
 267 {
 268         struct i915_vma *vma;
 269
 270         GEM_BUG_ON(view && !i915_is_ggtt_or_dpt(vm));
 271         GEM_BUG_ON(!atomic_read(&vm->open));
 272
 273         spin_lock(&obj->vma.lock);
 274         vma = i915_vma_lookup(obj, vm, view);
 275         spin_unlock(&obj->vma.lock);
 276
 277         /* vma_create() will resolve the race if another creates the vma */
 278         if (unlikely(!vma))
 279                 vma = vma_create(obj, vm, view);
 280
 281         GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
 282         return vma;
 283 }
 284
 285 struct i915_vma_work {
 286         struct dma_fence_work base;
 287         struct i915_address_space *vm;
 288         struct i915_vm_pt_stash stash;
 289         struct i915_vma *vma;
 290         struct drm_i915_gem_object *pinned;
 291         struct i915_sw_dma_fence_cb cb;
 292         enum i915_cache_level cache_level;
 293         unsigned int flags;
 294 };
 295
 296 static void __vma_bind(struct dma_fence_work *work)
 297 {
 298         struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
 299         struct i915_vma *vma = vw->vma;
 300
 301         vma->ops->bind_vma(vw->vm, &vw->stash,
 302                            vma, vw->cache_level, vw->flags);
 303 }
 304
 305 static void __vma_release(struct dma_fence_work *work)
 306 {
 307         struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
 308
 309         if (vw->pinned) {
 310                 __i915_gem_object_unpin_pages(vw->pinned);
 311                 i915_gem_object_put(vw->pinned);
 312         }
 313
 314         i915_vm_free_pt_stash(vw->vm, &vw->stash);
 315         i915_vm_put(vw->vm);
 316 }
 317
 318 static const struct dma_fence_work_ops bind_ops = {
 319         .name = "bind",
 320         .work = __vma_bind,
 321         .release = __vma_release,
 322 };
 323
 324 struct i915_vma_work *i915_vma_work(void)
 325 {
 326         struct i915_vma_work *vw;
 327
 328         vw = kzalloc(sizeof(*vw), GFP_KERNEL);
 329         if (!vw)
 330                 return NULL;
 331
 332         dma_fence_work_init(&vw->base, &bind_ops);
 333         vw->base.dma.error = -EAGAIN; /* disable the worker by default */
 334
 335         return vw;
 336 }
 337
 338 int i915_vma_wait_for_bind(struct i915_vma *vma)
 339 {
 340         int err = 0;
 341
 342         if (rcu_access_pointer(vma->active.excl.fence)) {
 343                 struct dma_fence *fence;
 344
 345                 rcu_read_lock();
 346                 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence);
 347                 rcu_read_unlock();
 348                 if (fence) {
 349                         err = dma_fence_wait(fence, MAX_SCHEDULE_TIMEOUT);
 350                         dma_fence_put(fence);
 351                 }
 352         }
 353
 354         return err;
 355 }
 356
 357 /**
 358  * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
 359  * @vma: VMA to map
 360  * @cache_level: mapping cache level
 361  * @flags: flags like global or local mapping
 362  * @work: preallocated worker for allocating and binding the PTE
 363  *
 364  * DMA addresses are taken from the scatter-gather table of this object (or of
 365  * this VMA in case of non-default GGTT views) and PTE entries set up.
 366  * Note that DMA addresses are also the only part of the SG table we care about.
 367  */
 368 int i915_vma_bind(struct i915_vma *vma,
 369                   enum i915_cache_level cache_level,
 370                   u32 flags,
 371                   struct i915_vma_work *work)
 372 {
 373         u32 bind_flags;
 374         u32 vma_flags;
 375
 376         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 377         GEM_BUG_ON(vma->size > vma->node.size);
 378
 379         if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
 380                                               vma->node.size,
 381                                               vma->vm->total)))
 382                 return -ENODEV;
 383
 384         if (GEM_DEBUG_WARN_ON(!flags))
 385                 return -EINVAL;
 386
 387         bind_flags = flags;
 388         bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
 389
 390         vma_flags = atomic_read(&vma->flags);
 391         vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
 392
 393         bind_flags &= ~vma_flags;
 394         if (bind_flags == 0)
 395                 return 0;
 396
 397         GEM_BUG_ON(!vma->pages);
 398
 399         trace_i915_vma_bind(vma, bind_flags);
 400         if (work && bind_flags & vma->vm->bind_async_flags) {
 401                 struct dma_fence *prev;
 402
 403                 work->vma = vma;
 404                 work->cache_level = cache_level;
 405                 work->flags = bind_flags;
 406
 407                 /*
 408                  * Note we only want to chain up to the migration fence on
 409                  * the pages (not the object itself). As we don't track that,
 410                  * yet, we have to use the exclusive fence instead.
 411                  *
 412                  * Also note that we do not want to track the async vma as
 413                  * part of the obj->resv->excl_fence as it only affects
 414                  * execution and not content or object's backing store lifetime.
 415                  */
 416                 prev = i915_active_set_exclusive(&vma->active, &work->base.dma);
 417                 if (prev) {
 418                         __i915_sw_fence_await_dma_fence(&work->base.chain,
 419                                                         prev,
 420                                                         &work->cb);
 421                         dma_fence_put(prev);
 422                 }
 423
 424                 work->base.dma.error = 0; /* enable the queue_work() */
 425
 426                 if (vma->obj) {
 427                         __i915_gem_object_pin_pages(vma->obj);
 428                         work->pinned = i915_gem_object_get(vma->obj);
 429                 }
 430         } else {
 431                 vma->ops->bind_vma(vma->vm, NULL, vma, cache_level, bind_flags);
 432         }
 433
 434         atomic_or(bind_flags, &vma->flags);
 435         return 0;
 436 }
 437
 438 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
 439 {
 440         void __iomem *ptr;
 441         int err;
 442
 443         if (!i915_gem_object_is_lmem(vma->obj)) {
 444                 if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
 445                         err = -ENODEV;
 446                         goto err;
 447                 }
 448         }
 449
 450         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 451         GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
 452
 453         ptr = READ_ONCE(vma->iomap);
 454         if (ptr == NULL) {
 455                 /*
 456                  * TODO: consider just using i915_gem_object_pin_map() for lmem
 457                  * instead, which already supports mapping non-contiguous chunks
 458                  * of pages, that way we can also drop the
 459                  * I915_BO_ALLOC_CONTIGUOUS when allocating the object.
 460                  */
 461                 if (i915_gem_object_is_lmem(vma->obj))
 462                         ptr = i915_gem_object_lmem_io_map(vma->obj, 0,
 463                                                           vma->obj->base.size);
 464                 else
 465                         ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
 466                                                 vma->node.start,
 467                                                 vma->node.size);
 468                 if (ptr == NULL) {
 469                         err = -ENOMEM;
 470                         goto err;
 471                 }
 472
 473                 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
 474                         io_mapping_unmap(ptr);
 475                         ptr = vma->iomap;
 476                 }
 477         }
 478
 479         __i915_vma_pin(vma);
 480
 481         err = i915_vma_pin_fence(vma);
 482         if (err)
 483                 goto err_unpin;
 484
 485         i915_vma_set_ggtt_write(vma);
 486
 487         /* NB Access through the GTT requires the device to be awake. */
 488         return ptr;
 489
 490 err_unpin:
 491         __i915_vma_unpin(vma);
 492 err:
 493         return IO_ERR_PTR(err);
 494 }
 495
 496 void i915_vma_flush_writes(struct i915_vma *vma)
 497 {
 498         if (i915_vma_unset_ggtt_write(vma))
 499                 intel_gt_flush_ggtt_writes(vma->vm->gt);
 500 }
 501
 502 void i915_vma_unpin_iomap(struct i915_vma *vma)
 503 {
 504         GEM_BUG_ON(vma->iomap == NULL);
 505
 506         i915_vma_flush_writes(vma);
 507
 508         i915_vma_unpin_fence(vma);
 509         i915_vma_unpin(vma);
 510 }
 511
 512 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
 513 {
 514         struct i915_vma *vma;
 515         struct drm_i915_gem_object *obj;
 516
 517         vma = fetch_and_zero(p_vma);
 518         if (!vma)
 519                 return;
 520
 521         obj = vma->obj;
 522         GEM_BUG_ON(!obj);
 523
 524         i915_vma_unpin(vma);
 525
 526         if (flags & I915_VMA_RELEASE_MAP)
 527                 i915_gem_object_unpin_map(obj);
 528
 529         i915_gem_object_put(obj);
 530 }
 531
 532 bool i915_vma_misplaced(const struct i915_vma *vma,
 533                         u64 size, u64 alignment, u64 flags)
 534 {
 535         if (!drm_mm_node_allocated(&vma->node))
 536                 return false;
 537
 538         if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
 539                 return true;
 540
 541         if (vma->node.size < size)
 542                 return true;
 543
 544         GEM_BUG_ON(alignment && !is_power_of_2(alignment));
 545         if (alignment && !IS_ALIGNED(vma->node.start, alignment))
 546                 return true;
 547
 548         if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
 549                 return true;
 550
 551         if (flags & PIN_OFFSET_BIAS &&
 552             vma->node.start < (flags & PIN_OFFSET_MASK))
 553                 return true;
 554
 555         if (flags & PIN_OFFSET_FIXED &&
 556             vma->node.start != (flags & PIN_OFFSET_MASK))
 557                 return true;
 558
 559         return false;
 560 }
 561
 562 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
 563 {
 564         bool mappable, fenceable;
 565
 566         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 567         GEM_BUG_ON(!vma->fence_size);
 568
 569         fenceable = (vma->node.size >= vma->fence_size &&
 570                      IS_ALIGNED(vma->node.start, vma->fence_alignment));
 571
 572         mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
 573
 574         if (mappable && fenceable)
 575                 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
 576         else
 577                 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
 578 }
 579
 580 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
 581 {
 582         struct drm_mm_node *node = &vma->node;
 583         struct drm_mm_node *other;
 584
 585         /*
 586          * On some machines we have to be careful when putting differing types
 587          * of snoopable memory together to avoid the prefetcher crossing memory
 588          * domains and dying. During vm initialisation, we decide whether or not
 589          * these constraints apply and set the drm_mm.color_adjust
 590          * appropriately.
 591          */
 592         if (!i915_vm_has_cache_coloring(vma->vm))
 593                 return true;
 594
 595         /* Only valid to be called on an already inserted vma */
 596         GEM_BUG_ON(!drm_mm_node_allocated(node));
 597         GEM_BUG_ON(list_empty(&node->node_list));
 598
 599         other = list_prev_entry(node, node_list);
 600         if (i915_node_color_differs(other, color) &&
 601             !drm_mm_hole_follows(other))
 602                 return false;
 603
 604         other = list_next_entry(node, node_list);
 605         if (i915_node_color_differs(other, color) &&
 606             !drm_mm_hole_follows(node))
 607                 return false;
 608
 609         return true;
 610 }
 611
 612 /**
 613  * i915_vma_insert - finds a slot for the vma in its address space
 614  * @vma: the vma
 615  * @size: requested size in bytes (can be larger than the VMA)
 616  * @alignment: required alignment
 617  * @flags: mask of PIN_* flags to use
 618  *
 619  * First we try to allocate some free space that meets the requirements for
 620  * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
 621  * preferrably the oldest idle entry to make room for the new VMA.
 622  *
 623  * Returns:
 624  * 0 on success, negative error code otherwise.
 625  */
 626 static int
 627 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 628 {
 629         unsigned long color;
 630         u64 start, end;
 631         int ret;
 632
 633         GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 634         GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
 635
 636         size = max(size, vma->size);
 637         alignment = max(alignment, vma->display_alignment);
 638         if (flags & PIN_MAPPABLE) {
 639                 size = max_t(typeof(size), size, vma->fence_size);
 640                 alignment = max_t(typeof(alignment),
 641                                   alignment, vma->fence_alignment);
 642         }
 643
 644         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
 645         GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
 646         GEM_BUG_ON(!is_power_of_2(alignment));
 647
 648         start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
 649         GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
 650
 651         end = vma->vm->total;
 652         if (flags & PIN_MAPPABLE)
 653                 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
 654         if (flags & PIN_ZONE_4G)
 655                 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
 656         GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
 657
 658         /* If binding the object/GGTT view requires more space than the entire
 659          * aperture has, reject it early before evicting everything in a vain
 660          * attempt to find space.
 661          */
 662         if (size > end) {
 663                 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
 664                           size, flags & PIN_MAPPABLE ? "mappable" : "total",
 665                           end);
 666                 return -ENOSPC;
 667         }
 668
 669         color = 0;
 670         if (vma->obj && i915_vm_has_cache_coloring(vma->vm))
 671                 color = vma->obj->cache_level;
 672
 673         if (flags & PIN_OFFSET_FIXED) {
 674                 u64 offset = flags & PIN_OFFSET_MASK;
 675                 if (!IS_ALIGNED(offset, alignment) ||
 676                     range_overflows(offset, size, end))
 677                         return -EINVAL;
 678
 679                 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
 680                                            size, offset, color,
 681                                            flags);
 682                 if (ret)
 683                         return ret;
 684         } else {
 685                 /*
 686                  * We only support huge gtt pages through the 48b PPGTT,
 687                  * however we also don't want to force any alignment for
 688                  * objects which need to be tightly packed into the low 32bits.
 689                  *
 690                  * Note that we assume that GGTT are limited to 4GiB for the
 691                  * forseeable future. See also i915_ggtt_offset().
 692                  */
 693                 if (upper_32_bits(end - 1) &&
 694                     vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
 695                         /*
 696                          * We can't mix 64K and 4K PTEs in the same page-table
 697                          * (2M block), and so to avoid the ugliness and
 698                          * complexity of coloring we opt for just aligning 64K
 699                          * objects to 2M.
 700                          */
 701                         u64 page_alignment =
 702                                 rounddown_pow_of_two(vma->page_sizes.sg |
 703                                                      I915_GTT_PAGE_SIZE_2M);
 704
 705                         /*
 706                          * Check we don't expand for the limited Global GTT
 707                          * (mappable aperture is even more precious!). This
 708                          * also checks that we exclude the aliasing-ppgtt.
 709                          */
 710                         GEM_BUG_ON(i915_vma_is_ggtt(vma));
 711
 712                         alignment = max(alignment, page_alignment);
 713
 714                         if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
 715                                 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
 716                 }
 717
 718                 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
 719                                           size, alignment, color,
 720                                           start, end, flags);
 721                 if (ret)
 722                         return ret;
 723
 724                 GEM_BUG_ON(vma->node.start < start);
 725                 GEM_BUG_ON(vma->node.start + vma->node.size > end);
 726         }
 727         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 728         GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
 729
 730         list_add_tail(&vma->vm_link, &vma->vm->bound_list);
 731
 732         return 0;
 733 }
 734
 735 static void
 736 i915_vma_detach(struct i915_vma *vma)
 737 {
 738         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 739         GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 740
 741         /*
 742          * And finally now the object is completely decoupled from this
 743          * vma, we can drop its hold on the backing storage and allow
 744          * it to be reaped by the shrinker.
 745          */
 746         list_del(&vma->vm_link);
 747 }
 748
 749 static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
 750 {
 751         unsigned int bound;
 752         bool pinned = true;
 753
 754         bound = atomic_read(&vma->flags);
 755         do {
 756                 if (unlikely(flags & ~bound))
 757                         return false;
 758
 759                 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
 760                         return false;
 761
 762                 if (!(bound & I915_VMA_PIN_MASK))
 763                         goto unpinned;
 764
 765                 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
 766         } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
 767
 768         return true;
 769
 770 unpinned:
 771         /*
 772          * If pin_count==0, but we are bound, check under the lock to avoid
 773          * racing with a concurrent i915_vma_unbind().
 774          */
 775         mutex_lock(&vma->vm->mutex);
 776         do {
 777                 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) {
 778                         pinned = false;
 779                         break;
 780                 }
 781
 782                 if (unlikely(flags & ~bound)) {
 783                         pinned = false;
 784                         break;
 785                 }
 786         } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
 787         mutex_unlock(&vma->vm->mutex);
 788
 789         return pinned;
 790 }
 791
 792 static int vma_get_pages(struct i915_vma *vma)
 793 {
 794         int err = 0;
 795         bool pinned_pages = false;
 796
 797         if (atomic_add_unless(&vma->pages_count, 1, 0))
 798                 return 0;
 799
 800         if (vma->obj) {
 801                 err = i915_gem_object_pin_pages(vma->obj);
 802                 if (err)
 803                         return err;
 804                 pinned_pages = true;
 805         }
 806
 807         /* Allocations ahoy! */
 808         if (mutex_lock_interruptible(&vma->pages_mutex)) {
 809                 err = -EINTR;
 810                 goto unpin;
 811         }
 812
 813         if (!atomic_read(&vma->pages_count)) {
 814                 err = vma->ops->set_pages(vma);
 815                 if (err)
 816                         goto unlock;
 817                 pinned_pages = false;
 818         }
 819         atomic_inc(&vma->pages_count);
 820
 821 unlock:
 822         mutex_unlock(&vma->pages_mutex);
 823 unpin:
 824         if (pinned_pages)
 825                 __i915_gem_object_unpin_pages(vma->obj);
 826
 827         return err;
 828 }
 829
 830 static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
 831 {
 832         /* We allocate under vma_get_pages, so beware the shrinker */
 833         mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING);
 834         GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
 835         if (atomic_sub_return(count, &vma->pages_count) == 0) {
 836                 vma->ops->clear_pages(vma);
 837                 GEM_BUG_ON(vma->pages);
 838                 if (vma->obj)
 839                         i915_gem_object_unpin_pages(vma->obj);
 840         }
 841         mutex_unlock(&vma->pages_mutex);
 842 }
 843
 844 static void vma_put_pages(struct i915_vma *vma)
 845 {
 846         if (atomic_add_unless(&vma->pages_count, -1, 1))
 847                 return;
 848
 849         __vma_put_pages(vma, 1);
 850 }
 851
 852 static void vma_unbind_pages(struct i915_vma *vma)
 853 {
 854         unsigned int count;
 855
 856         lockdep_assert_held(&vma->vm->mutex);
 857
 858         /* The upper portion of pages_count is the number of bindings */
 859         count = atomic_read(&vma->pages_count);
 860         count >>= I915_VMA_PAGES_BIAS;
 861         GEM_BUG_ON(!count);
 862
 863         __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
 864 }
 865
 866 int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
 867                     u64 size, u64 alignment, u64 flags)
 868 {
 869         struct i915_vma_work *work = NULL;
 870         intel_wakeref_t wakeref = 0;
 871         unsigned int bound;
 872         int err;
 873
 874 #ifdef CONFIG_PROVE_LOCKING
 875         if (debug_locks && !WARN_ON(!ww) && vma->resv)
 876                 assert_vma_held(vma);
 877 #endif
 878
 879         BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
 880         BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
 881
 882         GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
 883
 884         /* First try and grab the pin without rebinding the vma */
 885         if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
 886                 return 0;
 887
 888         err = vma_get_pages(vma);
 889         if (err)
 890                 return err;
 891
 892         if (flags & PIN_GLOBAL)
 893                 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
 894
 895         if (flags & vma->vm->bind_async_flags) {
 896                 /* lock VM */
 897                 err = i915_vm_lock_objects(vma->vm, ww);
 898                 if (err)
 899                         goto err_rpm;
 900
 901                 work = i915_vma_work();
 902                 if (!work) {
 903                         err = -ENOMEM;
 904                         goto err_rpm;
 905                 }
 906
 907                 work->vm = i915_vm_get(vma->vm);
 908
 909                 /* Allocate enough page directories to used PTE */
 910                 if (vma->vm->allocate_va_range) {
 911                         err = i915_vm_alloc_pt_stash(vma->vm,
 912                                                      &work->stash,
 913                                                      vma->size);
 914                         if (err)
 915                                 goto err_fence;
 916
 917                         err = i915_vm_map_pt_stash(vma->vm, &work->stash);
 918                         if (err)
 919                                 goto err_fence;
 920                 }
 921         }
 922
 923         /*
 924          * Differentiate between user/kernel vma inside the aliasing-ppgtt.
 925          *
 926          * We conflate the Global GTT with the user's vma when using the
 927          * aliasing-ppgtt, but it is still vitally important to try and
 928          * keep the use cases distinct. For example, userptr objects are
 929          * not allowed inside the Global GTT as that will cause lock
 930          * inversions when we have to evict them the mmu_notifier callbacks -
 931          * but they are allowed to be part of the user ppGTT which can never
 932          * be mapped. As such we try to give the distinct users of the same
 933          * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt
 934          * and i915_ppgtt separate].
 935          *
 936          * NB this may cause us to mask real lock inversions -- while the
 937          * code is safe today, lockdep may not be able to spot future
 938          * transgressions.
 939          */
 940         err = mutex_lock_interruptible_nested(&vma->vm->mutex,
 941                                               !(flags & PIN_GLOBAL));
 942         if (err)
 943                 goto err_fence;
 944
 945         /* No more allocations allowed now we hold vm->mutex */
 946
 947         if (unlikely(i915_vma_is_closed(vma))) {
 948                 err = -ENOENT;
 949                 goto err_unlock;
 950         }
 951
 952         bound = atomic_read(&vma->flags);
 953         if (unlikely(bound & I915_VMA_ERROR)) {
 954                 err = -ENOMEM;
 955                 goto err_unlock;
 956         }
 957
 958         if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
 959                 err = -EAGAIN; /* pins are meant to be fairly temporary */
 960                 goto err_unlock;
 961         }
 962
 963         if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
 964                 __i915_vma_pin(vma);
 965                 goto err_unlock;
 966         }
 967
 968         err = i915_active_acquire(&vma->active);
 969         if (err)
 970                 goto err_unlock;
 971
 972         if (!(bound & I915_VMA_BIND_MASK)) {
 973                 err = i915_vma_insert(vma, size, alignment, flags);
 974                 if (err)
 975                         goto err_active;
 976
 977                 if (i915_is_ggtt(vma->vm))
 978                         __i915_vma_set_map_and_fenceable(vma);
 979         }
 980
 981         GEM_BUG_ON(!vma->pages);
 982         err = i915_vma_bind(vma,
 983                             vma->obj ? vma->obj->cache_level : 0,
 984                             flags, work);
 985         if (err)
 986                 goto err_remove;
 987
 988         /* There should only be at most 2 active bindings (user, global) */
 989         GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
 990         atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
 991         list_move_tail(&vma->vm_link, &vma->vm->bound_list);
 992
 993         __i915_vma_pin(vma);
 994         GEM_BUG_ON(!i915_vma_is_pinned(vma));
 995         GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
 996         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
 997
 998 err_remove:
 999         if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
1000                 i915_vma_detach(vma);
1001                 drm_mm_remove_node(&vma->node);
1002         }
1003 err_active:
1004         i915_active_release(&vma->active);
1005 err_unlock:
1006         mutex_unlock(&vma->vm->mutex);
1007 err_fence:
1008         if (work)
1009                 dma_fence_work_commit_imm(&work->base);
1010 err_rpm:
1011         if (wakeref)
1012                 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
1013         vma_put_pages(vma);
1014         return err;
1015 }
1016
1017 static void flush_idle_contexts(struct intel_gt *gt)
1018 {
1019         struct intel_engine_cs *engine;
1020         enum intel_engine_id id;
1021
1022         for_each_engine(engine, gt, id)
1023                 intel_engine_flush_barriers(engine);
1024
1025         intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
1026 }
1027
1028 int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
1029                   u32 align, unsigned int flags)
1030 {
1031         struct i915_address_space *vm = vma->vm;
1032         int err;
1033
1034         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
1035
1036 #ifdef CONFIG_LOCKDEP
1037         WARN_ON(!ww && vma->resv && dma_resv_held(vma->resv));
1038 #endif
1039
1040         do {
1041                 if (ww)
1042                         err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL);
1043                 else
1044                         err = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL);
1045                 if (err != -ENOSPC) {
1046                         if (!err) {
1047                                 err = i915_vma_wait_for_bind(vma);
1048                                 if (err)
1049                                         i915_vma_unpin(vma);
1050                         }
1051                         return err;
1052                 }
1053
1054                 /* Unlike i915_vma_pin, we don't take no for an answer! */
1055                 flush_idle_contexts(vm->gt);
1056                 if (mutex_lock_interruptible(&vm->mutex) == 0) {
1057                         i915_gem_evict_vm(vm);
1058                         mutex_unlock(&vm->mutex);
1059                 }
1060         } while (1);
1061 }
1062
1063 static void __vma_close(struct i915_vma *vma, struct intel_gt *gt)
1064 {
1065         /*
1066          * We defer actually closing, unbinding and destroying the VMA until
1067          * the next idle point, or if the object is freed in the meantime. By
1068          * postponing the unbind, we allow for it to be resurrected by the
1069          * client, avoiding the work required to rebind the VMA. This is
1070          * advantageous for DRI, where the client/server pass objects
1071          * between themselves, temporarily opening a local VMA to the
1072          * object, and then closing it again. The same object is then reused
1073          * on the next frame (or two, depending on the depth of the swap queue)
1074          * causing us to rebind the VMA once more. This ends up being a lot
1075          * of wasted work for the steady state.
1076          */
1077         GEM_BUG_ON(i915_vma_is_closed(vma));
1078         list_add(&vma->closed_link, &gt->closed_vma);
1079 }
1080
1081 void i915_vma_close(struct i915_vma *vma)
1082 {
1083         struct intel_gt *gt = vma->vm->gt;
1084         unsigned long flags;
1085
1086         if (i915_vma_is_ggtt(vma))
1087                 return;
1088
1089         GEM_BUG_ON(!atomic_read(&vma->open_count));
1090         if (atomic_dec_and_lock_irqsave(&vma->open_count,
1091                                         &gt->closed_lock,
1092                                         flags)) {
1093                 __vma_close(vma, gt);
1094                 spin_unlock_irqrestore(&gt->closed_lock, flags);
1095         }
1096 }
1097
1098 static void __i915_vma_remove_closed(struct i915_vma *vma)
1099 {
1100         struct intel_gt *gt = vma->vm->gt;
1101
1102         spin_lock_irq(&gt->closed_lock);
1103         list_del_init(&vma->closed_link);
1104         spin_unlock_irq(&gt->closed_lock);
1105 }
1106
1107 void i915_vma_reopen(struct i915_vma *vma)
1108 {
1109         if (i915_vma_is_closed(vma))
1110                 __i915_vma_remove_closed(vma);
1111 }
1112
1113 void i915_vma_release(struct kref *ref)
1114 {
1115         struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
1116
1117         if (drm_mm_node_allocated(&vma->node)) {
1118                 mutex_lock(&vma->vm->mutex);
1119                 atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
1120                 WARN_ON(__i915_vma_unbind(vma));
1121                 mutex_unlock(&vma->vm->mutex);
1122                 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
1123         }
1124         GEM_BUG_ON(i915_vma_is_active(vma));
1125
1126         if (vma->obj) {
1127                 struct drm_i915_gem_object *obj = vma->obj;
1128
1129                 spin_lock(&obj->vma.lock);
1130                 list_del(&vma->obj_link);
1131                 if (!RB_EMPTY_NODE(&vma->obj_node))
1132                         rb_erase(&vma->obj_node, &obj->vma.tree);
1133                 spin_unlock(&obj->vma.lock);
1134         }
1135
1136         __i915_vma_remove_closed(vma);
1137         i915_vm_put(vma->vm);
1138
1139         i915_active_fini(&vma->active);
1140         i915_vma_free(vma);
1141 }
1142
1143 void i915_vma_parked(struct intel_gt *gt)
1144 {
1145         struct i915_vma *vma, *next;
1146         LIST_HEAD(closed);
1147
1148         spin_lock_irq(&gt->closed_lock);
1149         list_for_each_entry_safe(vma, next, &gt->closed_vma, closed_link) {
1150                 struct drm_i915_gem_object *obj = vma->obj;
1151                 struct i915_address_space *vm = vma->vm;
1152
1153                 /* XXX All to avoid keeping a reference on i915_vma itself */
1154
1155                 if (!kref_get_unless_zero(&obj->base.refcount))
1156                         continue;
1157
1158                 if (!i915_vm_tryopen(vm)) {
1159                         i915_gem_object_put(obj);
1160                         continue;
1161                 }
1162
1163                 list_move(&vma->closed_link, &closed);
1164         }
1165         spin_unlock_irq(&gt->closed_lock);
1166
1167         /* As the GT is held idle, no vma can be reopened as we destroy them */
1168         list_for_each_entry_safe(vma, next, &closed, closed_link) {
1169                 struct drm_i915_gem_object *obj = vma->obj;
1170                 struct i915_address_space *vm = vma->vm;
1171
1172                 INIT_LIST_HEAD(&vma->closed_link);
1173                 __i915_vma_put(vma);
1174
1175                 i915_gem_object_put(obj);
1176                 i915_vm_close(vm);
1177         }
1178 }
1179
1180 static void __i915_vma_iounmap(struct i915_vma *vma)
1181 {
1182         GEM_BUG_ON(i915_vma_is_pinned(vma));
1183
1184         if (vma->iomap == NULL)
1185                 return;
1186
1187         io_mapping_unmap(vma->iomap);
1188         vma->iomap = NULL;
1189 }
1190
1191 void i915_vma_revoke_mmap(struct i915_vma *vma)
1192 {
1193         struct drm_vma_offset_node *node;
1194         u64 vma_offset;
1195
1196         if (!i915_vma_has_userfault(vma))
1197                 return;
1198
1199         GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
1200         GEM_BUG_ON(!vma->obj->userfault_count);
1201
1202         node = &vma->mmo->vma_node;
1203         vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
1204         unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
1205                             drm_vma_node_offset_addr(node) + vma_offset,
1206                             vma->size,
1207                             1);
1208
1209         i915_vma_unset_userfault(vma);
1210         if (!--vma->obj->userfault_count)
1211                 list_del(&vma->obj->userfault_link);
1212 }
1213
1214 static int
1215 __i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma)
1216 {
1217         return __i915_request_await_exclusive(rq, &vma->active);
1218 }
1219
1220 int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
1221 {
1222         int err;
1223
1224         GEM_BUG_ON(!i915_vma_is_pinned(vma));
1225
1226         /* Wait for the vma to be bound before we start! */
1227         err = __i915_request_await_bind(rq, vma);
1228         if (err)
1229                 return err;
1230
1231         return i915_active_add_request(&vma->active, rq);
1232 }
1233
1234 int _i915_vma_move_to_active(struct i915_vma *vma,
1235                              struct i915_request *rq,
1236                              struct dma_fence *fence,
1237                              unsigned int flags)
1238 {
1239         struct drm_i915_gem_object *obj = vma->obj;
1240         int err;
1241
1242         assert_object_held(obj);
1243
1244         err = __i915_vma_move_to_active(vma, rq);
1245         if (unlikely(err))
1246                 return err;
1247
1248         if (flags & EXEC_OBJECT_WRITE) {
1249                 struct intel_frontbuffer *front;
1250
1251                 front = __intel_frontbuffer_get(obj);
1252                 if (unlikely(front)) {
1253                         if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
1254                                 i915_active_add_request(&front->write, rq);
1255                         intel_frontbuffer_put(front);
1256                 }
1257
1258                 if (fence) {
1259                         dma_resv_add_excl_fence(vma->resv, fence);
1260                         obj->write_domain = I915_GEM_DOMAIN_RENDER;
1261                         obj->read_domains = 0;
1262                 }
1263         } else {
1264                 if (!(flags & __EXEC_OBJECT_NO_RESERVE)) {
1265                         err = dma_resv_reserve_shared(vma->resv, 1);
1266                         if (unlikely(err))
1267                                 return err;
1268                 }
1269
1270                 if (fence) {
1271                         dma_resv_add_shared_fence(vma->resv, fence);
1272                         obj->write_domain = 0;
1273                 }
1274         }
1275
1276         if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence)
1277                 i915_active_add_request(&vma->fence->active, rq);
1278
1279         obj->read_domains |= I915_GEM_GPU_DOMAINS;
1280         obj->mm.dirty = true;
1281
1282         GEM_BUG_ON(!i915_vma_is_active(vma));
1283         return 0;
1284 }
1285
1286 void __i915_vma_evict(struct i915_vma *vma)
1287 {
1288         GEM_BUG_ON(i915_vma_is_pinned(vma));
1289
1290         if (i915_vma_is_map_and_fenceable(vma)) {
1291                 /* Force a pagefault for domain tracking on next user access */
1292                 i915_vma_revoke_mmap(vma);
1293
1294                 /*
1295                  * Check that we have flushed all writes through the GGTT
1296                  * before the unbind, other due to non-strict nature of those
1297                  * indirect writes they may end up referencing the GGTT PTE
1298                  * after the unbind.
1299                  *
1300                  * Note that we may be concurrently poking at the GGTT_WRITE
1301                  * bit from set-domain, as we mark all GGTT vma associated
1302                  * with an object. We know this is for another vma, as we
1303                  * are currently unbinding this one -- so if this vma will be
1304                  * reused, it will be refaulted and have its dirty bit set
1305                  * before the next write.
1306                  */
1307                 i915_vma_flush_writes(vma);
1308
1309                 /* release the fence reg _after_ flushing */
1310                 i915_vma_revoke_fence(vma);
1311
1312                 __i915_vma_iounmap(vma);
1313                 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
1314         }
1315         GEM_BUG_ON(vma->fence);
1316         GEM_BUG_ON(i915_vma_has_userfault(vma));
1317
1318         if (likely(atomic_read(&vma->vm->open))) {
1319                 trace_i915_vma_unbind(vma);
1320                 vma->ops->unbind_vma(vma->vm, vma);
1321         }
1322         atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE),
1323                    &vma->flags);
1324
1325         i915_vma_detach(vma);
1326         vma_unbind_pages(vma);
1327 }
1328
1329 int __i915_vma_unbind(struct i915_vma *vma)
1330 {
1331         int ret;
1332
1333         lockdep_assert_held(&vma->vm->mutex);
1334
1335         if (!drm_mm_node_allocated(&vma->node))
1336                 return 0;
1337
1338         if (i915_vma_is_pinned(vma)) {
1339                 vma_print_allocator(vma, "is pinned");
1340                 return -EAGAIN;
1341         }
1342
1343         /*
1344          * After confirming that no one else is pinning this vma, wait for
1345          * any laggards who may have crept in during the wait (through
1346          * a residual pin skipping the vm->mutex) to complete.
1347          */
1348         ret = i915_vma_sync(vma);
1349         if (ret)
1350                 return ret;
1351
1352         GEM_BUG_ON(i915_vma_is_active(vma));
1353         __i915_vma_evict(vma);
1354
1355         drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
1356         return 0;
1357 }
1358
1359 int i915_vma_unbind(struct i915_vma *vma)
1360 {
1361         struct i915_address_space *vm = vma->vm;
1362         intel_wakeref_t wakeref = 0;
1363         int err;
1364
1365         /* Optimistic wait before taking the mutex */
1366         err = i915_vma_sync(vma);
1367         if (err)
1368                 return err;
1369
1370         if (!drm_mm_node_allocated(&vma->node))
1371                 return 0;
1372
1373         if (i915_vma_is_pinned(vma)) {
1374                 vma_print_allocator(vma, "is pinned");
1375                 return -EAGAIN;
1376         }
1377
1378         if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
1379                 /* XXX not always required: nop_clear_range */
1380                 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
1381
1382         err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref);
1383         if (err)
1384                 goto out_rpm;
1385
1386         err = __i915_vma_unbind(vma);
1387         mutex_unlock(&vm->mutex);
1388
1389 out_rpm:
1390         if (wakeref)
1391                 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
1392         return err;
1393 }
1394
1395 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1396 {
1397         i915_gem_object_make_unshrinkable(vma->obj);
1398         return vma;
1399 }
1400
1401 void i915_vma_make_shrinkable(struct i915_vma *vma)
1402 {
1403         i915_gem_object_make_shrinkable(vma->obj);
1404 }
1405
1406 void i915_vma_make_purgeable(struct i915_vma *vma)
1407 {
1408         i915_gem_object_make_purgeable(vma->obj);
1409 }
1410
1411 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1412 #include "selftests/i915_vma.c"
1413 #endif
1414
1415 void i915_vma_module_exit(void)
1416 {
1417         kmem_cache_destroy(slab_vmas);
1418 }
1419
1420 int __init i915_vma_module_init(void)
1421 {
1422         slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1423         if (!slab_vmas)
1424                 return -ENOMEM;
1425
1426         return 0;
1427 }