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 static struct i915_vma *i915_vma_alloc(void)
  44 {
  45         return kmem_cache_zalloc(slab_vmas, GFP_KERNEL);
  46 }
  47
  48 static 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         vma->vm = i915_vm_get(vm);
 113         vma->ops = &vm->vma_ops;
 114         vma->obj = obj;
 115         vma->size = obj->base.size;
 116         vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
 117
 118         i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire, 0);
 119
 120         /* Declare ourselves safe for use inside shrinkers */
 121         if (IS_ENABLED(CONFIG_LOCKDEP)) {
 122                 fs_reclaim_acquire(GFP_KERNEL);
 123                 might_lock(&vma->active.mutex);
 124                 fs_reclaim_release(GFP_KERNEL);
 125         }
 126
 127         INIT_LIST_HEAD(&vma->closed_link);
 128
 129         if (view && view->type != I915_GGTT_VIEW_NORMAL) {
 130                 vma->ggtt_view = *view;
 131                 if (view->type == I915_GGTT_VIEW_PARTIAL) {
 132                         GEM_BUG_ON(range_overflows_t(u64,
 133                                                      view->partial.offset,
 134                                                      view->partial.size,
 135                                                      obj->base.size >> PAGE_SHIFT));
 136                         vma->size = view->partial.size;
 137                         vma->size <<= PAGE_SHIFT;
 138                         GEM_BUG_ON(vma->size > obj->base.size);
 139                 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
 140                         vma->size = intel_rotation_info_size(&view->rotated);
 141                         vma->size <<= PAGE_SHIFT;
 142                 } else if (view->type == I915_GGTT_VIEW_REMAPPED) {
 143                         vma->size = intel_remapped_info_size(&view->remapped);
 144                         vma->size <<= PAGE_SHIFT;
 145                 }
 146         }
 147
 148         if (unlikely(vma->size > vm->total))
 149                 goto err_vma;
 150
 151         GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
 152
 153         spin_lock(&obj->vma.lock);
 154
 155         if (i915_is_ggtt(vm)) {
 156                 if (unlikely(overflows_type(vma->size, u32)))
 157                         goto err_unlock;
 158
 159                 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
 160                                                       i915_gem_object_get_tiling(obj),
 161                                                       i915_gem_object_get_stride(obj));
 162                 if (unlikely(vma->fence_size < vma->size || /* overflow */
 163                              vma->fence_size > vm->total))
 164                         goto err_unlock;
 165
 166                 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
 167
 168                 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
 169                                                                 i915_gem_object_get_tiling(obj),
 170                                                                 i915_gem_object_get_stride(obj));
 171                 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
 172
 173                 __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma));
 174         }
 175
 176         rb = NULL;
 177         p = &obj->vma.tree.rb_node;
 178         while (*p) {
 179                 long cmp;
 180
 181                 rb = *p;
 182                 pos = rb_entry(rb, struct i915_vma, obj_node);
 183
 184                 /*
 185                  * If the view already exists in the tree, another thread
 186                  * already created a matching vma, so return the older instance
 187                  * and dispose of ours.
 188                  */
 189                 cmp = i915_vma_compare(pos, vm, view);
 190                 if (cmp < 0)
 191                         p = &rb->rb_right;
 192                 else if (cmp > 0)
 193                         p = &rb->rb_left;
 194                 else
 195                         goto err_unlock;
 196         }
 197         rb_link_node(&vma->obj_node, rb, p);
 198         rb_insert_color(&vma->obj_node, &obj->vma.tree);
 199
 200         if (i915_vma_is_ggtt(vma))
 201                 /*
 202                  * We put the GGTT vma at the start of the vma-list, followed
 203                  * by the ppGGTT vma. This allows us to break early when
 204                  * iterating over only the GGTT vma for an object, see
 205                  * for_each_ggtt_vma()
 206                  */
 207                 list_add(&vma->obj_link, &obj->vma.list);
 208         else
 209                 list_add_tail(&vma->obj_link, &obj->vma.list);
 210
 211         spin_unlock(&obj->vma.lock);
 212
 213         return vma;
 214
 215 err_unlock:
 216         spin_unlock(&obj->vma.lock);
 217 err_vma:
 218         i915_vm_put(vm);
 219         i915_vma_free(vma);
 220         return pos;
 221 }
 222
 223 static struct i915_vma *
 224 i915_vma_lookup(struct drm_i915_gem_object *obj,
 225            struct i915_address_space *vm,
 226            const struct i915_ggtt_view *view)
 227 {
 228         struct rb_node *rb;
 229
 230         rb = obj->vma.tree.rb_node;
 231         while (rb) {
 232                 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
 233                 long cmp;
 234
 235                 cmp = i915_vma_compare(vma, vm, view);
 236                 if (cmp == 0)
 237                         return vma;
 238
 239                 if (cmp < 0)
 240                         rb = rb->rb_right;
 241                 else
 242                         rb = rb->rb_left;
 243         }
 244
 245         return NULL;
 246 }
 247
 248 /**
 249  * i915_vma_instance - return the singleton instance of the VMA
 250  * @obj: parent &struct drm_i915_gem_object to be mapped
 251  * @vm: address space in which the mapping is located
 252  * @view: additional mapping requirements
 253  *
 254  * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
 255  * the same @view characteristics. If a match is not found, one is created.
 256  * Once created, the VMA is kept until either the object is freed, or the
 257  * address space is closed.
 258  *
 259  * Returns the vma, or an error pointer.
 260  */
 261 struct i915_vma *
 262 i915_vma_instance(struct drm_i915_gem_object *obj,
 263                   struct i915_address_space *vm,
 264                   const struct i915_ggtt_view *view)
 265 {
 266         struct i915_vma *vma;
 267
 268         GEM_BUG_ON(view && !i915_is_ggtt_or_dpt(vm));
 269         GEM_BUG_ON(!atomic_read(&vm->open));
 270
 271         spin_lock(&obj->vma.lock);
 272         vma = i915_vma_lookup(obj, vm, view);
 273         spin_unlock(&obj->vma.lock);
 274
 275         /* vma_create() will resolve the race if another creates the vma */
 276         if (unlikely(!vma))
 277                 vma = vma_create(obj, vm, view);
 278
 279         GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
 280         return vma;
 281 }
 282
 283 struct i915_vma_work {
 284         struct dma_fence_work base;
 285         struct i915_address_space *vm;
 286         struct i915_vm_pt_stash stash;
 287         struct i915_vma *vma;
 288         struct drm_i915_gem_object *pinned;
 289         struct i915_sw_dma_fence_cb cb;
 290         enum i915_cache_level cache_level;
 291         unsigned int flags;
 292 };
 293
 294 static void __vma_bind(struct dma_fence_work *work)
 295 {
 296         struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
 297         struct i915_vma *vma = vw->vma;
 298
 299         vma->ops->bind_vma(vw->vm, &vw->stash,
 300                            vma, vw->cache_level, vw->flags);
 301 }
 302
 303 static void __vma_release(struct dma_fence_work *work)
 304 {
 305         struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
 306
 307         if (vw->pinned) {
 308                 __i915_gem_object_unpin_pages(vw->pinned);
 309                 i915_gem_object_put(vw->pinned);
 310         }
 311
 312         i915_vm_free_pt_stash(vw->vm, &vw->stash);
 313         i915_vm_put(vw->vm);
 314 }
 315
 316 static const struct dma_fence_work_ops bind_ops = {
 317         .name = "bind",
 318         .work = __vma_bind,
 319         .release = __vma_release,
 320 };
 321
 322 struct i915_vma_work *i915_vma_work(void)
 323 {
 324         struct i915_vma_work *vw;
 325
 326         vw = kzalloc(sizeof(*vw), GFP_KERNEL);
 327         if (!vw)
 328                 return NULL;
 329
 330         dma_fence_work_init(&vw->base, &bind_ops);
 331         vw->base.dma.error = -EAGAIN; /* disable the worker by default */
 332
 333         return vw;
 334 }
 335
 336 int i915_vma_wait_for_bind(struct i915_vma *vma)
 337 {
 338         int err = 0;
 339
 340         if (rcu_access_pointer(vma->active.excl.fence)) {
 341                 struct dma_fence *fence;
 342
 343                 rcu_read_lock();
 344                 fence = dma_fence_get_rcu_safe(&vma->active.excl.fence);
 345                 rcu_read_unlock();
 346                 if (fence) {
 347                         err = dma_fence_wait(fence, true);
 348                         dma_fence_put(fence);
 349                 }
 350         }
 351
 352         return err;
 353 }
 354
 355 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
 356 static int i915_vma_verify_bind_complete(struct i915_vma *vma)
 357 {
 358         struct dma_fence *fence = i915_active_fence_get(&vma->active.excl);
 359         int err;
 360
 361         if (!fence)
 362                 return 0;
 363
 364         if (dma_fence_is_signaled(fence))
 365                 err = fence->error;
 366         else
 367                 err = -EBUSY;
 368
 369         dma_fence_put(fence);
 370
 371         return err;
 372 }
 373 #else
 374 #define i915_vma_verify_bind_complete(_vma) 0
 375 #endif
 376
 377 /**
 378  * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
 379  * @vma: VMA to map
 380  * @cache_level: mapping cache level
 381  * @flags: flags like global or local mapping
 382  * @work: preallocated worker for allocating and binding the PTE
 383  *
 384  * DMA addresses are taken from the scatter-gather table of this object (or of
 385  * this VMA in case of non-default GGTT views) and PTE entries set up.
 386  * Note that DMA addresses are also the only part of the SG table we care about.
 387  */
 388 int i915_vma_bind(struct i915_vma *vma,
 389                   enum i915_cache_level cache_level,
 390                   u32 flags,
 391                   struct i915_vma_work *work)
 392 {
 393         u32 bind_flags;
 394         u32 vma_flags;
 395
 396         lockdep_assert_held(&vma->vm->mutex);
 397         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 398         GEM_BUG_ON(vma->size > vma->node.size);
 399
 400         if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
 401                                               vma->node.size,
 402                                               vma->vm->total)))
 403                 return -ENODEV;
 404
 405         if (GEM_DEBUG_WARN_ON(!flags))
 406                 return -EINVAL;
 407
 408         bind_flags = flags;
 409         bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
 410
 411         vma_flags = atomic_read(&vma->flags);
 412         vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
 413
 414         bind_flags &= ~vma_flags;
 415         if (bind_flags == 0)
 416                 return 0;
 417
 418         GEM_BUG_ON(!atomic_read(&vma->pages_count));
 419
 420         trace_i915_vma_bind(vma, bind_flags);
 421         if (work && bind_flags & vma->vm->bind_async_flags) {
 422                 struct dma_fence *prev;
 423
 424                 work->vma = vma;
 425                 work->cache_level = cache_level;
 426                 work->flags = bind_flags;
 427
 428                 /*
 429                  * Note we only want to chain up to the migration fence on
 430                  * the pages (not the object itself). As we don't track that,
 431                  * yet, we have to use the exclusive fence instead.
 432                  *
 433                  * Also note that we do not want to track the async vma as
 434                  * part of the obj->resv->excl_fence as it only affects
 435                  * execution and not content or object's backing store lifetime.
 436                  */
 437                 prev = i915_active_set_exclusive(&vma->active, &work->base.dma);
 438                 if (prev) {
 439                         __i915_sw_fence_await_dma_fence(&work->base.chain,
 440                                                         prev,
 441                                                         &work->cb);
 442                         dma_fence_put(prev);
 443                 }
 444
 445                 work->base.dma.error = 0; /* enable the queue_work() */
 446
 447                 __i915_gem_object_pin_pages(vma->obj);
 448                 work->pinned = i915_gem_object_get(vma->obj);
 449         } else {
 450                 if (vma->obj) {
 451                         int ret;
 452
 453                         ret = i915_gem_object_wait_moving_fence(vma->obj, true);
 454                         if (ret)
 455                                 return ret;
 456                 }
 457                 vma->ops->bind_vma(vma->vm, NULL, vma, cache_level, bind_flags);
 458         }
 459
 460         if (vma->obj)
 461                 set_bit(I915_BO_WAS_BOUND_BIT, &vma->obj->flags);
 462
 463         atomic_or(bind_flags, &vma->flags);
 464         return 0;
 465 }
 466
 467 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
 468 {
 469         void __iomem *ptr;
 470         int err;
 471
 472         if (!i915_gem_object_is_lmem(vma->obj)) {
 473                 if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
 474                         err = -ENODEV;
 475                         goto err;
 476                 }
 477         }
 478
 479         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 480         GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
 481         GEM_BUG_ON(i915_vma_verify_bind_complete(vma));
 482
 483         ptr = READ_ONCE(vma->iomap);
 484         if (ptr == NULL) {
 485                 /*
 486                  * TODO: consider just using i915_gem_object_pin_map() for lmem
 487                  * instead, which already supports mapping non-contiguous chunks
 488                  * of pages, that way we can also drop the
 489                  * I915_BO_ALLOC_CONTIGUOUS when allocating the object.
 490                  */
 491                 if (i915_gem_object_is_lmem(vma->obj))
 492                         ptr = i915_gem_object_lmem_io_map(vma->obj, 0,
 493                                                           vma->obj->base.size);
 494                 else
 495                         ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
 496                                                 vma->node.start,
 497                                                 vma->node.size);
 498                 if (ptr == NULL) {
 499                         err = -ENOMEM;
 500                         goto err;
 501                 }
 502
 503                 if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
 504                         io_mapping_unmap(ptr);
 505                         ptr = vma->iomap;
 506                 }
 507         }
 508
 509         __i915_vma_pin(vma);
 510
 511         err = i915_vma_pin_fence(vma);
 512         if (err)
 513                 goto err_unpin;
 514
 515         i915_vma_set_ggtt_write(vma);
 516
 517         /* NB Access through the GTT requires the device to be awake. */
 518         return ptr;
 519
 520 err_unpin:
 521         __i915_vma_unpin(vma);
 522 err:
 523         return IO_ERR_PTR(err);
 524 }
 525
 526 void i915_vma_flush_writes(struct i915_vma *vma)
 527 {
 528         if (i915_vma_unset_ggtt_write(vma))
 529                 intel_gt_flush_ggtt_writes(vma->vm->gt);
 530 }
 531
 532 void i915_vma_unpin_iomap(struct i915_vma *vma)
 533 {
 534         GEM_BUG_ON(vma->iomap == NULL);
 535
 536         i915_vma_flush_writes(vma);
 537
 538         i915_vma_unpin_fence(vma);
 539         i915_vma_unpin(vma);
 540 }
 541
 542 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
 543 {
 544         struct i915_vma *vma;
 545         struct drm_i915_gem_object *obj;
 546
 547         vma = fetch_and_zero(p_vma);
 548         if (!vma)
 549                 return;
 550
 551         obj = vma->obj;
 552         GEM_BUG_ON(!obj);
 553
 554         i915_vma_unpin(vma);
 555
 556         if (flags & I915_VMA_RELEASE_MAP)
 557                 i915_gem_object_unpin_map(obj);
 558
 559         i915_gem_object_put(obj);
 560 }
 561
 562 bool i915_vma_misplaced(const struct i915_vma *vma,
 563                         u64 size, u64 alignment, u64 flags)
 564 {
 565         if (!drm_mm_node_allocated(&vma->node))
 566                 return false;
 567
 568         if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
 569                 return true;
 570
 571         if (vma->node.size < size)
 572                 return true;
 573
 574         GEM_BUG_ON(alignment && !is_power_of_2(alignment));
 575         if (alignment && !IS_ALIGNED(vma->node.start, alignment))
 576                 return true;
 577
 578         if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
 579                 return true;
 580
 581         if (flags & PIN_OFFSET_BIAS &&
 582             vma->node.start < (flags & PIN_OFFSET_MASK))
 583                 return true;
 584
 585         if (flags & PIN_OFFSET_FIXED &&
 586             vma->node.start != (flags & PIN_OFFSET_MASK))
 587                 return true;
 588
 589         return false;
 590 }
 591
 592 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
 593 {
 594         bool mappable, fenceable;
 595
 596         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 597         GEM_BUG_ON(!vma->fence_size);
 598
 599         fenceable = (vma->node.size >= vma->fence_size &&
 600                      IS_ALIGNED(vma->node.start, vma->fence_alignment));
 601
 602         mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
 603
 604         if (mappable && fenceable)
 605                 set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
 606         else
 607                 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
 608 }
 609
 610 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
 611 {
 612         struct drm_mm_node *node = &vma->node;
 613         struct drm_mm_node *other;
 614
 615         /*
 616          * On some machines we have to be careful when putting differing types
 617          * of snoopable memory together to avoid the prefetcher crossing memory
 618          * domains and dying. During vm initialisation, we decide whether or not
 619          * these constraints apply and set the drm_mm.color_adjust
 620          * appropriately.
 621          */
 622         if (!i915_vm_has_cache_coloring(vma->vm))
 623                 return true;
 624
 625         /* Only valid to be called on an already inserted vma */
 626         GEM_BUG_ON(!drm_mm_node_allocated(node));
 627         GEM_BUG_ON(list_empty(&node->node_list));
 628
 629         other = list_prev_entry(node, node_list);
 630         if (i915_node_color_differs(other, color) &&
 631             !drm_mm_hole_follows(other))
 632                 return false;
 633
 634         other = list_next_entry(node, node_list);
 635         if (i915_node_color_differs(other, color) &&
 636             !drm_mm_hole_follows(node))
 637                 return false;
 638
 639         return true;
 640 }
 641
 642 /**
 643  * i915_vma_insert - finds a slot for the vma in its address space
 644  * @vma: the vma
 645  * @size: requested size in bytes (can be larger than the VMA)
 646  * @alignment: required alignment
 647  * @flags: mask of PIN_* flags to use
 648  *
 649  * First we try to allocate some free space that meets the requirements for
 650  * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
 651  * preferrably the oldest idle entry to make room for the new VMA.
 652  *
 653  * Returns:
 654  * 0 on success, negative error code otherwise.
 655  */
 656 static int
 657 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 658 {
 659         unsigned long color;
 660         u64 start, end;
 661         int ret;
 662
 663         GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 664         GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
 665
 666         size = max(size, vma->size);
 667         alignment = max(alignment, vma->display_alignment);
 668         if (flags & PIN_MAPPABLE) {
 669                 size = max_t(typeof(size), size, vma->fence_size);
 670                 alignment = max_t(typeof(alignment),
 671                                   alignment, vma->fence_alignment);
 672         }
 673
 674         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
 675         GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
 676         GEM_BUG_ON(!is_power_of_2(alignment));
 677
 678         start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
 679         GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
 680
 681         end = vma->vm->total;
 682         if (flags & PIN_MAPPABLE)
 683                 end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end);
 684         if (flags & PIN_ZONE_4G)
 685                 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
 686         GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
 687
 688         /* If binding the object/GGTT view requires more space than the entire
 689          * aperture has, reject it early before evicting everything in a vain
 690          * attempt to find space.
 691          */
 692         if (size > end) {
 693                 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
 694                           size, flags & PIN_MAPPABLE ? "mappable" : "total",
 695                           end);
 696                 return -ENOSPC;
 697         }
 698
 699         color = 0;
 700         if (i915_vm_has_cache_coloring(vma->vm))
 701                 color = vma->obj->cache_level;
 702
 703         if (flags & PIN_OFFSET_FIXED) {
 704                 u64 offset = flags & PIN_OFFSET_MASK;
 705                 if (!IS_ALIGNED(offset, alignment) ||
 706                     range_overflows(offset, size, end))
 707                         return -EINVAL;
 708
 709                 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
 710                                            size, offset, color,
 711                                            flags);
 712                 if (ret)
 713                         return ret;
 714         } else {
 715                 /*
 716                  * We only support huge gtt pages through the 48b PPGTT,
 717                  * however we also don't want to force any alignment for
 718                  * objects which need to be tightly packed into the low 32bits.
 719                  *
 720                  * Note that we assume that GGTT are limited to 4GiB for the
 721                  * forseeable future. See also i915_ggtt_offset().
 722                  */
 723                 if (upper_32_bits(end - 1) &&
 724                     vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
 725                         /*
 726                          * We can't mix 64K and 4K PTEs in the same page-table
 727                          * (2M block), and so to avoid the ugliness and
 728                          * complexity of coloring we opt for just aligning 64K
 729                          * objects to 2M.
 730                          */
 731                         u64 page_alignment =
 732                                 rounddown_pow_of_two(vma->page_sizes.sg |
 733                                                      I915_GTT_PAGE_SIZE_2M);
 734
 735                         /*
 736                          * Check we don't expand for the limited Global GTT
 737                          * (mappable aperture is even more precious!). This
 738                          * also checks that we exclude the aliasing-ppgtt.
 739                          */
 740                         GEM_BUG_ON(i915_vma_is_ggtt(vma));
 741
 742                         alignment = max(alignment, page_alignment);
 743
 744                         if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
 745                                 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
 746                 }
 747
 748                 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
 749                                           size, alignment, color,
 750                                           start, end, flags);
 751                 if (ret)
 752                         return ret;
 753
 754                 GEM_BUG_ON(vma->node.start < start);
 755                 GEM_BUG_ON(vma->node.start + vma->node.size > end);
 756         }
 757         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 758         GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
 759
 760         list_add_tail(&vma->vm_link, &vma->vm->bound_list);
 761
 762         return 0;
 763 }
 764
 765 static void
 766 i915_vma_detach(struct i915_vma *vma)
 767 {
 768         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 769         GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 770
 771         /*
 772          * And finally now the object is completely decoupled from this
 773          * vma, we can drop its hold on the backing storage and allow
 774          * it to be reaped by the shrinker.
 775          */
 776         list_del(&vma->vm_link);
 777 }
 778
 779 static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
 780 {
 781         unsigned int bound;
 782         bool pinned = true;
 783
 784         bound = atomic_read(&vma->flags);
 785         do {
 786                 if (unlikely(flags & ~bound))
 787                         return false;
 788
 789                 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
 790                         return false;
 791
 792                 if (!(bound & I915_VMA_PIN_MASK))
 793                         goto unpinned;
 794
 795                 GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
 796         } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
 797
 798         return true;
 799
 800 unpinned:
 801         /*
 802          * If pin_count==0, but we are bound, check under the lock to avoid
 803          * racing with a concurrent i915_vma_unbind().
 804          */
 805         mutex_lock(&vma->vm->mutex);
 806         do {
 807                 if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) {
 808                         pinned = false;
 809                         break;
 810                 }
 811
 812                 if (unlikely(flags & ~bound)) {
 813                         pinned = false;
 814                         break;
 815                 }
 816         } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
 817         mutex_unlock(&vma->vm->mutex);
 818
 819         return pinned;
 820 }
 821
 822 static struct scatterlist *
 823 rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset,
 824              unsigned int width, unsigned int height,
 825              unsigned int src_stride, unsigned int dst_stride,
 826              struct sg_table *st, struct scatterlist *sg)
 827 {
 828         unsigned int column, row;
 829         unsigned int src_idx;
 830
 831         for (column = 0; column < width; column++) {
 832                 unsigned int left;
 833
 834                 src_idx = src_stride * (height - 1) + column + offset;
 835                 for (row = 0; row < height; row++) {
 836                         st->nents++;
 837                         /*
 838                          * We don't need the pages, but need to initialize
 839                          * the entries so the sg list can be happily traversed.
 840                          * The only thing we need are DMA addresses.
 841                          */
 842                         sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0);
 843                         sg_dma_address(sg) =
 844                                 i915_gem_object_get_dma_address(obj, src_idx);
 845                         sg_dma_len(sg) = I915_GTT_PAGE_SIZE;
 846                         sg = sg_next(sg);
 847                         src_idx -= src_stride;
 848                 }
 849
 850                 left = (dst_stride - height) * I915_GTT_PAGE_SIZE;
 851
 852                 if (!left)
 853                         continue;
 854
 855                 st->nents++;
 856
 857                 /*
 858                  * The DE ignores the PTEs for the padding tiles, the sg entry
 859                  * here is just a conenience to indicate how many padding PTEs
 860                  * to insert at this spot.
 861                  */
 862                 sg_set_page(sg, NULL, left, 0);
 863                 sg_dma_address(sg) = 0;
 864                 sg_dma_len(sg) = left;
 865                 sg = sg_next(sg);
 866         }
 867
 868         return sg;
 869 }
 870
 871 static noinline struct sg_table *
 872 intel_rotate_pages(struct intel_rotation_info *rot_info,
 873                    struct drm_i915_gem_object *obj)
 874 {
 875         unsigned int size = intel_rotation_info_size(rot_info);
 876         struct drm_i915_private *i915 = to_i915(obj->base.dev);
 877         struct sg_table *st;
 878         struct scatterlist *sg;
 879         int ret = -ENOMEM;
 880         int i;
 881
 882         /* Allocate target SG list. */
 883         st = kmalloc(sizeof(*st), GFP_KERNEL);
 884         if (!st)
 885                 goto err_st_alloc;
 886
 887         ret = sg_alloc_table(st, size, GFP_KERNEL);
 888         if (ret)
 889                 goto err_sg_alloc;
 890
 891         st->nents = 0;
 892         sg = st->sgl;
 893
 894         for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++)
 895                 sg = rotate_pages(obj, rot_info->plane[i].offset,
 896                                   rot_info->plane[i].width, rot_info->plane[i].height,
 897                                   rot_info->plane[i].src_stride,
 898                                   rot_info->plane[i].dst_stride,
 899                                   st, sg);
 900
 901         return st;
 902
 903 err_sg_alloc:
 904         kfree(st);
 905 err_st_alloc:
 906
 907         drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n",
 908                 obj->base.size, rot_info->plane[0].width,
 909                 rot_info->plane[0].height, size);
 910
 911         return ERR_PTR(ret);
 912 }
 913
 914 static struct scatterlist *
 915 remap_pages(struct drm_i915_gem_object *obj,
 916             unsigned int offset, unsigned int alignment_pad,
 917             unsigned int width, unsigned int height,
 918             unsigned int src_stride, unsigned int dst_stride,
 919             struct sg_table *st, struct scatterlist *sg)
 920 {
 921         unsigned int row;
 922
 923         if (!width || !height)
 924                 return sg;
 925
 926         if (alignment_pad) {
 927                 st->nents++;
 928
 929                 /*
 930                  * The DE ignores the PTEs for the padding tiles, the sg entry
 931                  * here is just a convenience to indicate how many padding PTEs
 932                  * to insert at this spot.
 933                  */
 934                 sg_set_page(sg, NULL, alignment_pad * 4096, 0);
 935                 sg_dma_address(sg) = 0;
 936                 sg_dma_len(sg) = alignment_pad * 4096;
 937                 sg = sg_next(sg);
 938         }
 939
 940         for (row = 0; row < height; row++) {
 941                 unsigned int left = width * I915_GTT_PAGE_SIZE;
 942
 943                 while (left) {
 944                         dma_addr_t addr;
 945                         unsigned int length;
 946
 947                         /*
 948                          * We don't need the pages, but need to initialize
 949                          * the entries so the sg list can be happily traversed.
 950                          * The only thing we need are DMA addresses.
 951                          */
 952
 953                         addr = i915_gem_object_get_dma_address_len(obj, offset, &length);
 954
 955                         length = min(left, length);
 956
 957                         st->nents++;
 958
 959                         sg_set_page(sg, NULL, length, 0);
 960                         sg_dma_address(sg) = addr;
 961                         sg_dma_len(sg) = length;
 962                         sg = sg_next(sg);
 963
 964                         offset += length / I915_GTT_PAGE_SIZE;
 965                         left -= length;
 966                 }
 967
 968                 offset += src_stride - width;
 969
 970                 left = (dst_stride - width) * I915_GTT_PAGE_SIZE;
 971
 972                 if (!left)
 973                         continue;
 974
 975                 st->nents++;
 976
 977                 /*
 978                  * The DE ignores the PTEs for the padding tiles, the sg entry
 979                  * here is just a conenience to indicate how many padding PTEs
 980                  * to insert at this spot.
 981                  */
 982                 sg_set_page(sg, NULL, left, 0);
 983                 sg_dma_address(sg) = 0;
 984                 sg_dma_len(sg) = left;
 985                 sg = sg_next(sg);
 986         }
 987
 988         return sg;
 989 }
 990
 991 static noinline struct sg_table *
 992 intel_remap_pages(struct intel_remapped_info *rem_info,
 993                   struct drm_i915_gem_object *obj)
 994 {
 995         unsigned int size = intel_remapped_info_size(rem_info);
 996         struct drm_i915_private *i915 = to_i915(obj->base.dev);
 997         struct sg_table *st;
 998         struct scatterlist *sg;
 999         unsigned int gtt_offset = 0;
1000         int ret = -ENOMEM;
1001         int i;
1002
1003         /* Allocate target SG list. */
1004         st = kmalloc(sizeof(*st), GFP_KERNEL);
1005         if (!st)
1006                 goto err_st_alloc;
1007
1008         ret = sg_alloc_table(st, size, GFP_KERNEL);
1009         if (ret)
1010                 goto err_sg_alloc;
1011
1012         st->nents = 0;
1013         sg = st->sgl;
1014
1015         for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) {
1016                 unsigned int alignment_pad = 0;
1017
1018                 if (rem_info->plane_alignment)
1019                         alignment_pad = ALIGN(gtt_offset, rem_info->plane_alignment) - gtt_offset;
1020
1021                 sg = remap_pages(obj,
1022                                  rem_info->plane[i].offset, alignment_pad,
1023                                  rem_info->plane[i].width, rem_info->plane[i].height,
1024                                  rem_info->plane[i].src_stride, rem_info->plane[i].dst_stride,
1025                                  st, sg);
1026
1027                 gtt_offset += alignment_pad +
1028                               rem_info->plane[i].dst_stride * rem_info->plane[i].height;
1029         }
1030
1031         i915_sg_trim(st);
1032
1033         return st;
1034
1035 err_sg_alloc:
1036         kfree(st);
1037 err_st_alloc:
1038
1039         drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n",
1040                 obj->base.size, rem_info->plane[0].width,
1041                 rem_info->plane[0].height, size);
1042
1043         return ERR_PTR(ret);
1044 }
1045
1046 static noinline struct sg_table *
1047 intel_partial_pages(const struct i915_ggtt_view *view,
1048                     struct drm_i915_gem_object *obj)
1049 {
1050         struct sg_table *st;
1051         struct scatterlist *sg, *iter;
1052         unsigned int count = view->partial.size;
1053         unsigned int offset;
1054         int ret = -ENOMEM;
1055
1056         st = kmalloc(sizeof(*st), GFP_KERNEL);
1057         if (!st)
1058                 goto err_st_alloc;
1059
1060         ret = sg_alloc_table(st, count, GFP_KERNEL);
1061         if (ret)
1062                 goto err_sg_alloc;
1063
1064         iter = i915_gem_object_get_sg_dma(obj, view->partial.offset, &offset);
1065         GEM_BUG_ON(!iter);
1066
1067         sg = st->sgl;
1068         st->nents = 0;
1069         do {
1070                 unsigned int len;
1071
1072                 len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT),
1073                           count << PAGE_SHIFT);
1074                 sg_set_page(sg, NULL, len, 0);
1075                 sg_dma_address(sg) =
1076                         sg_dma_address(iter) + (offset << PAGE_SHIFT);
1077                 sg_dma_len(sg) = len;
1078
1079                 st->nents++;
1080                 count -= len >> PAGE_SHIFT;
1081                 if (count == 0) {
1082                         sg_mark_end(sg);
1083                         i915_sg_trim(st); /* Drop any unused tail entries. */
1084
1085                         return st;
1086                 }
1087
1088                 sg = __sg_next(sg);
1089                 iter = __sg_next(iter);
1090                 offset = 0;
1091         } while (1);
1092
1093 err_sg_alloc:
1094         kfree(st);
1095 err_st_alloc:
1096         return ERR_PTR(ret);
1097 }
1098
1099 static int
1100 __i915_vma_get_pages(struct i915_vma *vma)
1101 {
1102         struct sg_table *pages;
1103         int ret;
1104
1105         /*
1106          * The vma->pages are only valid within the lifespan of the borrowed
1107          * obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so
1108          * must be the vma->pages. A simple rule is that vma->pages must only
1109          * be accessed when the obj->mm.pages are pinned.
1110          */
1111         GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj));
1112
1113         switch (vma->ggtt_view.type) {
1114         default:
1115                 GEM_BUG_ON(vma->ggtt_view.type);
1116                 fallthrough;
1117         case I915_GGTT_VIEW_NORMAL:
1118                 pages = vma->obj->mm.pages;
1119                 break;
1120
1121         case I915_GGTT_VIEW_ROTATED:
1122                 pages =
1123                         intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj);
1124                 break;
1125
1126         case I915_GGTT_VIEW_REMAPPED:
1127                 pages =
1128                         intel_remap_pages(&vma->ggtt_view.remapped, vma->obj);
1129                 break;
1130
1131         case I915_GGTT_VIEW_PARTIAL:
1132                 pages = intel_partial_pages(&vma->ggtt_view, vma->obj);
1133                 break;
1134         }
1135
1136         ret = 0;
1137         if (IS_ERR(pages)) {
1138                 ret = PTR_ERR(pages);
1139                 pages = NULL;
1140                 drm_err(&vma->vm->i915->drm,
1141                         "Failed to get pages for VMA view type %u (%d)!\n",
1142                         vma->ggtt_view.type, ret);
1143         }
1144
1145         vma->pages = pages;
1146
1147         return ret;
1148 }
1149
1150 I915_SELFTEST_EXPORT int i915_vma_get_pages(struct i915_vma *vma)
1151 {
1152         int err;
1153
1154         if (atomic_add_unless(&vma->pages_count, 1, 0))
1155                 return 0;
1156
1157         err = i915_gem_object_pin_pages(vma->obj);
1158         if (err)
1159                 return err;
1160
1161         err = __i915_vma_get_pages(vma);
1162         if (err)
1163                 goto err_unpin;
1164
1165         vma->page_sizes = vma->obj->mm.page_sizes;
1166         atomic_inc(&vma->pages_count);
1167
1168         return 0;
1169
1170 err_unpin:
1171         __i915_gem_object_unpin_pages(vma->obj);
1172
1173         return err;
1174 }
1175
1176 static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
1177 {
1178         /* We allocate under vma_get_pages, so beware the shrinker */
1179         struct sg_table *pages = READ_ONCE(vma->pages);
1180
1181         GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
1182
1183         if (atomic_sub_return(count, &vma->pages_count) == 0) {
1184                 /*
1185                  * The atomic_sub_return is a read barrier for the READ_ONCE of
1186                  * vma->pages above.
1187                  *
1188                  * READ_ONCE is safe because this is either called from the same
1189                  * function (i915_vma_pin_ww), or guarded by vma->vm->mutex.
1190                  *
1191                  * TODO: We're leaving vma->pages dangling, until vma->obj->resv
1192                  * lock is required.
1193                  */
1194                 if (pages != vma->obj->mm.pages) {
1195                         sg_free_table(pages);
1196                         kfree(pages);
1197                 }
1198
1199                 i915_gem_object_unpin_pages(vma->obj);
1200         }
1201 }
1202
1203 I915_SELFTEST_EXPORT void i915_vma_put_pages(struct i915_vma *vma)
1204 {
1205         if (atomic_add_unless(&vma->pages_count, -1, 1))
1206                 return;
1207
1208         __vma_put_pages(vma, 1);
1209 }
1210
1211 static void vma_unbind_pages(struct i915_vma *vma)
1212 {
1213         unsigned int count;
1214
1215         lockdep_assert_held(&vma->vm->mutex);
1216
1217         /* The upper portion of pages_count is the number of bindings */
1218         count = atomic_read(&vma->pages_count);
1219         count >>= I915_VMA_PAGES_BIAS;
1220         GEM_BUG_ON(!count);
1221
1222         __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS);
1223 }
1224
1225 int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
1226                     u64 size, u64 alignment, u64 flags)
1227 {
1228         struct i915_vma_work *work = NULL;
1229         struct dma_fence *moving = NULL;
1230         intel_wakeref_t wakeref = 0;
1231         unsigned int bound;
1232         int err;
1233
1234         assert_vma_held(vma);
1235         GEM_BUG_ON(!ww);
1236
1237         BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND);
1238         BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND);
1239
1240         GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
1241
1242         /* First try and grab the pin without rebinding the vma */
1243         if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
1244                 return 0;
1245
1246         err = i915_vma_get_pages(vma);
1247         if (err)
1248                 return err;
1249
1250         if (flags & PIN_GLOBAL)
1251                 wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
1252
1253         moving = vma->obj ? i915_gem_object_get_moving_fence(vma->obj) : NULL;
1254         if (flags & vma->vm->bind_async_flags || moving) {
1255                 /* lock VM */
1256                 err = i915_vm_lock_objects(vma->vm, ww);
1257                 if (err)
1258                         goto err_rpm;
1259
1260                 work = i915_vma_work();
1261                 if (!work) {
1262                         err = -ENOMEM;
1263                         goto err_rpm;
1264                 }
1265
1266                 work->vm = i915_vm_get(vma->vm);
1267
1268                 dma_fence_work_chain(&work->base, moving);
1269
1270                 /* Allocate enough page directories to used PTE */
1271                 if (vma->vm->allocate_va_range) {
1272                         err = i915_vm_alloc_pt_stash(vma->vm,
1273                                                      &work->stash,
1274                                                      vma->size);
1275                         if (err)
1276                                 goto err_fence;
1277
1278                         err = i915_vm_map_pt_stash(vma->vm, &work->stash);
1279                         if (err)
1280                                 goto err_fence;
1281                 }
1282         }
1283
1284         /*
1285          * Differentiate between user/kernel vma inside the aliasing-ppgtt.
1286          *
1287          * We conflate the Global GTT with the user's vma when using the
1288          * aliasing-ppgtt, but it is still vitally important to try and
1289          * keep the use cases distinct. For example, userptr objects are
1290          * not allowed inside the Global GTT as that will cause lock
1291          * inversions when we have to evict them the mmu_notifier callbacks -
1292          * but they are allowed to be part of the user ppGTT which can never
1293          * be mapped. As such we try to give the distinct users of the same
1294          * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt
1295          * and i915_ppgtt separate].
1296          *
1297          * NB this may cause us to mask real lock inversions -- while the
1298          * code is safe today, lockdep may not be able to spot future
1299          * transgressions.
1300          */
1301         err = mutex_lock_interruptible_nested(&vma->vm->mutex,
1302                                               !(flags & PIN_GLOBAL));
1303         if (err)
1304                 goto err_fence;
1305
1306         /* No more allocations allowed now we hold vm->mutex */
1307
1308         if (unlikely(i915_vma_is_closed(vma))) {
1309                 err = -ENOENT;
1310                 goto err_unlock;
1311         }
1312
1313         bound = atomic_read(&vma->flags);
1314         if (unlikely(bound & I915_VMA_ERROR)) {
1315                 err = -ENOMEM;
1316                 goto err_unlock;
1317         }
1318
1319         if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))) {
1320                 err = -EAGAIN; /* pins are meant to be fairly temporary */
1321                 goto err_unlock;
1322         }
1323
1324         if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
1325                 __i915_vma_pin(vma);
1326                 goto err_unlock;
1327         }
1328
1329         err = i915_active_acquire(&vma->active);
1330         if (err)
1331                 goto err_unlock;
1332
1333         if (!(bound & I915_VMA_BIND_MASK)) {
1334                 err = i915_vma_insert(vma, size, alignment, flags);
1335                 if (err)
1336                         goto err_active;
1337
1338                 if (i915_is_ggtt(vma->vm))
1339                         __i915_vma_set_map_and_fenceable(vma);
1340         }
1341
1342         GEM_BUG_ON(!vma->pages);
1343         err = i915_vma_bind(vma,
1344                             vma->obj->cache_level,
1345                             flags, work);
1346         if (err)
1347                 goto err_remove;
1348
1349         /* There should only be at most 2 active bindings (user, global) */
1350         GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound);
1351         atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
1352         list_move_tail(&vma->vm_link, &vma->vm->bound_list);
1353
1354         __i915_vma_pin(vma);
1355         GEM_BUG_ON(!i915_vma_is_pinned(vma));
1356         GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
1357         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
1358
1359 err_remove:
1360         if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK)) {
1361                 i915_vma_detach(vma);
1362                 drm_mm_remove_node(&vma->node);
1363         }
1364 err_active:
1365         i915_active_release(&vma->active);
1366 err_unlock:
1367         mutex_unlock(&vma->vm->mutex);
1368 err_fence:
1369         if (work)
1370                 dma_fence_work_commit_imm(&work->base);
1371 err_rpm:
1372         if (wakeref)
1373                 intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
1374
1375         if (moving)
1376                 dma_fence_put(moving);
1377
1378         i915_vma_put_pages(vma);
1379         return err;
1380 }
1381
1382 static void flush_idle_contexts(struct intel_gt *gt)
1383 {
1384         struct intel_engine_cs *engine;
1385         enum intel_engine_id id;
1386
1387         for_each_engine(engine, gt, id)
1388                 intel_engine_flush_barriers(engine);
1389
1390         intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
1391 }
1392
1393 static int __i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
1394                            u32 align, unsigned int flags)
1395 {
1396         struct i915_address_space *vm = vma->vm;
1397         int err;
1398
1399         do {
1400                 err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL);
1401
1402                 if (err != -ENOSPC) {
1403                         if (!err) {
1404                                 err = i915_vma_wait_for_bind(vma);
1405                                 if (err)
1406                                         i915_vma_unpin(vma);
1407                         }
1408                         return err;
1409                 }
1410
1411                 /* Unlike i915_vma_pin, we don't take no for an answer! */
1412                 flush_idle_contexts(vm->gt);
1413                 if (mutex_lock_interruptible(&vm->mutex) == 0) {
1414                         i915_gem_evict_vm(vm);
1415                         mutex_unlock(&vm->mutex);
1416                 }
1417         } while (1);
1418 }
1419
1420 int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
1421                   u32 align, unsigned int flags)
1422 {
1423         struct i915_gem_ww_ctx _ww;
1424         int err;
1425
1426         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
1427
1428         if (ww)
1429                 return __i915_ggtt_pin(vma, ww, align, flags);
1430
1431 #ifdef CONFIG_LOCKDEP
1432         WARN_ON(dma_resv_held(vma->obj->base.resv));
1433 #endif
1434
1435         for_i915_gem_ww(&_ww, err, true) {
1436                 err = i915_gem_object_lock(vma->obj, &_ww);
1437                 if (!err)
1438                         err = __i915_ggtt_pin(vma, &_ww, align, flags);
1439         }
1440
1441         return err;
1442 }
1443
1444 static void __vma_close(struct i915_vma *vma, struct intel_gt *gt)
1445 {
1446         /*
1447          * We defer actually closing, unbinding and destroying the VMA until
1448          * the next idle point, or if the object is freed in the meantime. By
1449          * postponing the unbind, we allow for it to be resurrected by the
1450          * client, avoiding the work required to rebind the VMA. This is
1451          * advantageous for DRI, where the client/server pass objects
1452          * between themselves, temporarily opening a local VMA to the
1453          * object, and then closing it again. The same object is then reused
1454          * on the next frame (or two, depending on the depth of the swap queue)
1455          * causing us to rebind the VMA once more. This ends up being a lot
1456          * of wasted work for the steady state.
1457          */
1458         GEM_BUG_ON(i915_vma_is_closed(vma));
1459         list_add(&vma->closed_link, &gt->closed_vma);
1460 }
1461
1462 void i915_vma_close(struct i915_vma *vma)
1463 {
1464         struct intel_gt *gt = vma->vm->gt;
1465         unsigned long flags;
1466
1467         if (i915_vma_is_ggtt(vma))
1468                 return;
1469
1470         GEM_BUG_ON(!atomic_read(&vma->open_count));
1471         if (atomic_dec_and_lock_irqsave(&vma->open_count,
1472                                         &gt->closed_lock,
1473                                         flags)) {
1474                 __vma_close(vma, gt);
1475                 spin_unlock_irqrestore(&gt->closed_lock, flags);
1476         }
1477 }
1478
1479 static void __i915_vma_remove_closed(struct i915_vma *vma)
1480 {
1481         struct intel_gt *gt = vma->vm->gt;
1482
1483         spin_lock_irq(&gt->closed_lock);
1484         list_del_init(&vma->closed_link);
1485         spin_unlock_irq(&gt->closed_lock);
1486 }
1487
1488 void i915_vma_reopen(struct i915_vma *vma)
1489 {
1490         if (i915_vma_is_closed(vma))
1491                 __i915_vma_remove_closed(vma);
1492 }
1493
1494 void i915_vma_release(struct kref *ref)
1495 {
1496         struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
1497         struct drm_i915_gem_object *obj = vma->obj;
1498
1499         if (drm_mm_node_allocated(&vma->node)) {
1500                 mutex_lock(&vma->vm->mutex);
1501                 atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
1502                 WARN_ON(__i915_vma_unbind(vma));
1503                 mutex_unlock(&vma->vm->mutex);
1504                 GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
1505         }
1506         GEM_BUG_ON(i915_vma_is_active(vma));
1507
1508         spin_lock(&obj->vma.lock);
1509         list_del(&vma->obj_link);
1510         if (!RB_EMPTY_NODE(&vma->obj_node))
1511                 rb_erase(&vma->obj_node, &obj->vma.tree);
1512         spin_unlock(&obj->vma.lock);
1513
1514         __i915_vma_remove_closed(vma);
1515         i915_vm_put(vma->vm);
1516
1517         i915_active_fini(&vma->active);
1518         i915_vma_free(vma);
1519 }
1520
1521 void i915_vma_parked(struct intel_gt *gt)
1522 {
1523         struct i915_vma *vma, *next;
1524         LIST_HEAD(closed);
1525
1526         spin_lock_irq(&gt->closed_lock);
1527         list_for_each_entry_safe(vma, next, &gt->closed_vma, closed_link) {
1528                 struct drm_i915_gem_object *obj = vma->obj;
1529                 struct i915_address_space *vm = vma->vm;
1530
1531                 /* XXX All to avoid keeping a reference on i915_vma itself */
1532
1533                 if (!kref_get_unless_zero(&obj->base.refcount))
1534                         continue;
1535
1536                 if (!i915_vm_tryopen(vm)) {
1537                         i915_gem_object_put(obj);
1538                         continue;
1539                 }
1540
1541                 list_move(&vma->closed_link, &closed);
1542         }
1543         spin_unlock_irq(&gt->closed_lock);
1544
1545         /* As the GT is held idle, no vma can be reopened as we destroy them */
1546         list_for_each_entry_safe(vma, next, &closed, closed_link) {
1547                 struct drm_i915_gem_object *obj = vma->obj;
1548                 struct i915_address_space *vm = vma->vm;
1549
1550                 INIT_LIST_HEAD(&vma->closed_link);
1551                 __i915_vma_put(vma);
1552
1553                 i915_gem_object_put(obj);
1554                 i915_vm_close(vm);
1555         }
1556 }
1557
1558 static void __i915_vma_iounmap(struct i915_vma *vma)
1559 {
1560         GEM_BUG_ON(i915_vma_is_pinned(vma));
1561
1562         if (vma->iomap == NULL)
1563                 return;
1564
1565         io_mapping_unmap(vma->iomap);
1566         vma->iomap = NULL;
1567 }
1568
1569 void i915_vma_revoke_mmap(struct i915_vma *vma)
1570 {
1571         struct drm_vma_offset_node *node;
1572         u64 vma_offset;
1573
1574         if (!i915_vma_has_userfault(vma))
1575                 return;
1576
1577         GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
1578         GEM_BUG_ON(!vma->obj->userfault_count);
1579
1580         node = &vma->mmo->vma_node;
1581         vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
1582         unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
1583                             drm_vma_node_offset_addr(node) + vma_offset,
1584                             vma->size,
1585                             1);
1586
1587         i915_vma_unset_userfault(vma);
1588         if (!--vma->obj->userfault_count)
1589                 list_del(&vma->obj->userfault_link);
1590 }
1591
1592 static int
1593 __i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma)
1594 {
1595         return __i915_request_await_exclusive(rq, &vma->active);
1596 }
1597
1598 static int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq)
1599 {
1600         int err;
1601
1602         GEM_BUG_ON(!i915_vma_is_pinned(vma));
1603
1604         /* Wait for the vma to be bound before we start! */
1605         err = __i915_request_await_bind(rq, vma);
1606         if (err)
1607                 return err;
1608
1609         return i915_active_add_request(&vma->active, rq);
1610 }
1611
1612 int _i915_vma_move_to_active(struct i915_vma *vma,
1613                              struct i915_request *rq,
1614                              struct dma_fence *fence,
1615                              unsigned int flags)
1616 {
1617         struct drm_i915_gem_object *obj = vma->obj;
1618         int err;
1619
1620         assert_object_held(obj);
1621
1622         err = __i915_vma_move_to_active(vma, rq);
1623         if (unlikely(err))
1624                 return err;
1625
1626         if (flags & EXEC_OBJECT_WRITE) {
1627                 struct intel_frontbuffer *front;
1628
1629                 front = __intel_frontbuffer_get(obj);
1630                 if (unlikely(front)) {
1631                         if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
1632                                 i915_active_add_request(&front->write, rq);
1633                         intel_frontbuffer_put(front);
1634                 }
1635
1636                 if (fence) {
1637                         dma_resv_add_excl_fence(vma->obj->base.resv, fence);
1638                         obj->write_domain = I915_GEM_DOMAIN_RENDER;
1639                         obj->read_domains = 0;
1640                 }
1641         } else {
1642                 if (!(flags & __EXEC_OBJECT_NO_RESERVE)) {
1643                         err = dma_resv_reserve_shared(vma->obj->base.resv, 1);
1644                         if (unlikely(err))
1645                                 return err;
1646                 }
1647
1648                 if (fence) {
1649                         dma_resv_add_shared_fence(vma->obj->base.resv, fence);
1650                         obj->write_domain = 0;
1651                 }
1652         }
1653
1654         if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence)
1655                 i915_active_add_request(&vma->fence->active, rq);
1656
1657         obj->read_domains |= I915_GEM_GPU_DOMAINS;
1658         obj->mm.dirty = true;
1659
1660         GEM_BUG_ON(!i915_vma_is_active(vma));
1661         return 0;
1662 }
1663
1664 void __i915_vma_evict(struct i915_vma *vma)
1665 {
1666         GEM_BUG_ON(i915_vma_is_pinned(vma));
1667
1668         if (i915_vma_is_map_and_fenceable(vma)) {
1669                 /* Force a pagefault for domain tracking on next user access */
1670                 i915_vma_revoke_mmap(vma);
1671
1672                 /*
1673                  * Check that we have flushed all writes through the GGTT
1674                  * before the unbind, other due to non-strict nature of those
1675                  * indirect writes they may end up referencing the GGTT PTE
1676                  * after the unbind.
1677                  *
1678                  * Note that we may be concurrently poking at the GGTT_WRITE
1679                  * bit from set-domain, as we mark all GGTT vma associated
1680                  * with an object. We know this is for another vma, as we
1681                  * are currently unbinding this one -- so if this vma will be
1682                  * reused, it will be refaulted and have its dirty bit set
1683                  * before the next write.
1684                  */
1685                 i915_vma_flush_writes(vma);
1686
1687                 /* release the fence reg _after_ flushing */
1688                 i915_vma_revoke_fence(vma);
1689
1690                 __i915_vma_iounmap(vma);
1691                 clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
1692         }
1693         GEM_BUG_ON(vma->fence);
1694         GEM_BUG_ON(i915_vma_has_userfault(vma));
1695
1696         if (likely(atomic_read(&vma->vm->open))) {
1697                 trace_i915_vma_unbind(vma);
1698                 vma->ops->unbind_vma(vma->vm, vma);
1699         }
1700         atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE),
1701                    &vma->flags);
1702
1703         i915_vma_detach(vma);
1704         vma_unbind_pages(vma);
1705 }
1706
1707 int __i915_vma_unbind(struct i915_vma *vma)
1708 {
1709         int ret;
1710
1711         lockdep_assert_held(&vma->vm->mutex);
1712
1713         if (!drm_mm_node_allocated(&vma->node))
1714                 return 0;
1715
1716         if (i915_vma_is_pinned(vma)) {
1717                 vma_print_allocator(vma, "is pinned");
1718                 return -EAGAIN;
1719         }
1720
1721         /*
1722          * After confirming that no one else is pinning this vma, wait for
1723          * any laggards who may have crept in during the wait (through
1724          * a residual pin skipping the vm->mutex) to complete.
1725          */
1726         ret = i915_vma_sync(vma);
1727         if (ret)
1728                 return ret;
1729
1730         GEM_BUG_ON(i915_vma_is_active(vma));
1731         __i915_vma_evict(vma);
1732
1733         drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
1734         return 0;
1735 }
1736
1737 int i915_vma_unbind(struct i915_vma *vma)
1738 {
1739         struct i915_address_space *vm = vma->vm;
1740         intel_wakeref_t wakeref = 0;
1741         int err;
1742
1743         /* Optimistic wait before taking the mutex */
1744         err = i915_vma_sync(vma);
1745         if (err)
1746                 return err;
1747
1748         if (!drm_mm_node_allocated(&vma->node))
1749                 return 0;
1750
1751         if (i915_vma_is_pinned(vma)) {
1752                 vma_print_allocator(vma, "is pinned");
1753                 return -EAGAIN;
1754         }
1755
1756         if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
1757                 /* XXX not always required: nop_clear_range */
1758                 wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
1759
1760         err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref);
1761         if (err)
1762                 goto out_rpm;
1763
1764         err = __i915_vma_unbind(vma);
1765         mutex_unlock(&vm->mutex);
1766
1767 out_rpm:
1768         if (wakeref)
1769                 intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
1770         return err;
1771 }
1772
1773 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1774 {
1775         i915_gem_object_make_unshrinkable(vma->obj);
1776         return vma;
1777 }
1778
1779 void i915_vma_make_shrinkable(struct i915_vma *vma)
1780 {
1781         i915_gem_object_make_shrinkable(vma->obj);
1782 }
1783
1784 void i915_vma_make_purgeable(struct i915_vma *vma)
1785 {
1786         i915_gem_object_make_purgeable(vma->obj);
1787 }
1788
1789 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1790 #include "selftests/i915_vma.c"
1791 #endif
1792
1793 void i915_vma_module_exit(void)
1794 {
1795         kmem_cache_destroy(slab_vmas);
1796 }
1797
1798 int __init i915_vma_module_init(void)
1799 {
1800         slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1801         if (!slab_vmas)
1802                 return -ENOMEM;
1803
1804         return 0;
1805 }