drivers/gpu/drm/i915/gem/i915_gem_ttm.c

   1 // SPDX-License-Identifier: MIT
   2 /*
   3  * Copyright © 2021 Intel Corporation
   4  */
   5
   6 #include <drm/ttm/ttm_bo_driver.h>
   7 #include <drm/ttm/ttm_placement.h>
   8
   9 #include "i915_drv.h"
  10 #include "intel_memory_region.h"
  11 #include "intel_region_ttm.h"
  12
  13 #include "gem/i915_gem_object.h"
  14 #include "gem/i915_gem_region.h"
  15 #include "gem/i915_gem_ttm.h"
  16 #include "gem/i915_gem_mman.h"
  17
  18 #include "gt/intel_migrate.h"
  19 #include "gt/intel_engine_pm.h"
  20
  21 #define I915_PL_LMEM0 TTM_PL_PRIV
  22 #define I915_PL_SYSTEM TTM_PL_SYSTEM
  23 #define I915_PL_STOLEN TTM_PL_VRAM
  24 #define I915_PL_GGTT TTM_PL_TT
  25
  26 #define I915_TTM_PRIO_PURGE     0
  27 #define I915_TTM_PRIO_NO_PAGES  1
  28 #define I915_TTM_PRIO_HAS_PAGES 2
  29
  30 /*
  31  * Size of struct ttm_place vector in on-stack struct ttm_placement allocs
  32  */
  33 #define I915_TTM_MAX_PLACEMENTS INTEL_REGION_UNKNOWN
  34
  35 /**
  36  * struct i915_ttm_tt - TTM page vector with additional private information
  37  * @ttm: The base TTM page vector.
  38  * @dev: The struct device used for dma mapping and unmapping.
  39  * @cached_st: The cached scatter-gather table.
  40  *
  41  * Note that DMA may be going on right up to the point where the page-
  42  * vector is unpopulated in delayed destroy. Hence keep the
  43  * scatter-gather table mapped and cached up to that point. This is
  44  * different from the cached gem object io scatter-gather table which
  45  * doesn't have an associated dma mapping.
  46  */
  47 struct i915_ttm_tt {
  48         struct ttm_tt ttm;
  49         struct device *dev;
  50         struct sg_table *cached_st;
  51 };
  52
  53 static const struct ttm_place sys_placement_flags = {
  54         .fpfn = 0,
  55         .lpfn = 0,
  56         .mem_type = I915_PL_SYSTEM,
  57         .flags = 0,
  58 };
  59
  60 static struct ttm_placement i915_sys_placement = {
  61         .num_placement = 1,
  62         .placement = &sys_placement_flags,
  63         .num_busy_placement = 1,
  64         .busy_placement = &sys_placement_flags,
  65 };
  66
  67 static int i915_ttm_err_to_gem(int err)
  68 {
  69         /* Fastpath */
  70         if (likely(!err))
  71                 return 0;
  72
  73         switch (err) {
  74         case -EBUSY:
  75                 /*
  76                  * TTM likes to convert -EDEADLK to -EBUSY, and wants us to
  77                  * restart the operation, since we don't record the contending
  78                  * lock. We use -EAGAIN to restart.
  79                  */
  80                 return -EAGAIN;
  81         case -ENOSPC:
  82                 /*
  83                  * Memory type / region is full, and we can't evict.
  84                  * Except possibly system, that returns -ENOMEM;
  85                  */
  86                 return -ENXIO;
  87         default:
  88                 break;
  89         }
  90
  91         return err;
  92 }
  93
  94 static bool gpu_binds_iomem(struct ttm_resource *mem)
  95 {
  96         return mem->mem_type != TTM_PL_SYSTEM;
  97 }
  98
  99 static bool cpu_maps_iomem(struct ttm_resource *mem)
 100 {
 101         /* Once / if we support GGTT, this is also false for cached ttm_tts */
 102         return mem->mem_type != TTM_PL_SYSTEM;
 103 }
 104
 105 static enum i915_cache_level
 106 i915_ttm_cache_level(struct drm_i915_private *i915, struct ttm_resource *res,
 107                      struct ttm_tt *ttm)
 108 {
 109         return ((HAS_LLC(i915) || HAS_SNOOP(i915)) && !gpu_binds_iomem(res) &&
 110                 ttm->caching == ttm_cached) ? I915_CACHE_LLC :
 111                 I915_CACHE_NONE;
 112 }
 113
 114 static void i915_ttm_adjust_lru(struct drm_i915_gem_object *obj);
 115
 116 static enum ttm_caching
 117 i915_ttm_select_tt_caching(const struct drm_i915_gem_object *obj)
 118 {
 119         /*
 120          * Objects only allowed in system get cached cpu-mappings.
 121          * Other objects get WC mapping for now. Even if in system.
 122          */
 123         if (obj->mm.region->type == INTEL_MEMORY_SYSTEM &&
 124             obj->mm.n_placements <= 1)
 125                 return ttm_cached;
 126
 127         return ttm_write_combined;
 128 }
 129
 130 static void
 131 i915_ttm_place_from_region(const struct intel_memory_region *mr,
 132                            struct ttm_place *place,
 133                            unsigned int flags)
 134 {
 135         memset(place, 0, sizeof(*place));
 136         place->mem_type = intel_region_to_ttm_type(mr);
 137
 138         if (flags & I915_BO_ALLOC_CONTIGUOUS)
 139                 place->flags = TTM_PL_FLAG_CONTIGUOUS;
 140 }
 141
 142 static void
 143 i915_ttm_placement_from_obj(const struct drm_i915_gem_object *obj,
 144                             struct ttm_place *requested,
 145                             struct ttm_place *busy,
 146                             struct ttm_placement *placement)
 147 {
 148         unsigned int num_allowed = obj->mm.n_placements;
 149         unsigned int flags = obj->flags;
 150         unsigned int i;
 151
 152         placement->num_placement = 1;
 153         i915_ttm_place_from_region(num_allowed ? obj->mm.placements[0] :
 154                                    obj->mm.region, requested, flags);
 155
 156         /* Cache this on object? */
 157         placement->num_busy_placement = num_allowed;
 158         for (i = 0; i < placement->num_busy_placement; ++i)
 159                 i915_ttm_place_from_region(obj->mm.placements[i], busy + i, flags);
 160
 161         if (num_allowed == 0) {
 162                 *busy = *requested;
 163                 placement->num_busy_placement = 1;
 164         }
 165
 166         placement->placement = requested;
 167         placement->busy_placement = busy;
 168 }
 169
 170 static struct ttm_tt *i915_ttm_tt_create(struct ttm_buffer_object *bo,
 171                                          uint32_t page_flags)
 172 {
 173         struct ttm_resource_manager *man =
 174                 ttm_manager_type(bo->bdev, bo->resource->mem_type);
 175         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 176         struct i915_ttm_tt *i915_tt;
 177         int ret;
 178
 179         i915_tt = kzalloc(sizeof(*i915_tt), GFP_KERNEL);
 180         if (!i915_tt)
 181                 return NULL;
 182
 183         if (obj->flags & I915_BO_ALLOC_CPU_CLEAR &&
 184             man->use_tt)
 185                 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
 186
 187         ret = ttm_tt_init(&i915_tt->ttm, bo, page_flags,
 188                           i915_ttm_select_tt_caching(obj));
 189         if (ret) {
 190                 kfree(i915_tt);
 191                 return NULL;
 192         }
 193
 194         i915_tt->dev = obj->base.dev->dev;
 195
 196         return &i915_tt->ttm;
 197 }
 198
 199 static void i915_ttm_tt_unpopulate(struct ttm_device *bdev, struct ttm_tt *ttm)
 200 {
 201         struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
 202
 203         if (i915_tt->cached_st) {
 204                 dma_unmap_sgtable(i915_tt->dev, i915_tt->cached_st,
 205                                   DMA_BIDIRECTIONAL, 0);
 206                 sg_free_table(i915_tt->cached_st);
 207                 kfree(i915_tt->cached_st);
 208                 i915_tt->cached_st = NULL;
 209         }
 210         ttm_pool_free(&bdev->pool, ttm);
 211 }
 212
 213 static void i915_ttm_tt_destroy(struct ttm_device *bdev, struct ttm_tt *ttm)
 214 {
 215         struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
 216
 217         ttm_tt_destroy_common(bdev, ttm);
 218         ttm_tt_fini(ttm);
 219         kfree(i915_tt);
 220 }
 221
 222 static bool i915_ttm_eviction_valuable(struct ttm_buffer_object *bo,
 223                                        const struct ttm_place *place)
 224 {
 225         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 226
 227         /* Will do for now. Our pinned objects are still on TTM's LRU lists */
 228         return i915_gem_object_evictable(obj);
 229 }
 230
 231 static void i915_ttm_evict_flags(struct ttm_buffer_object *bo,
 232                                  struct ttm_placement *placement)
 233 {
 234         *placement = i915_sys_placement;
 235 }
 236
 237 static int i915_ttm_move_notify(struct ttm_buffer_object *bo)
 238 {
 239         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 240         int ret;
 241
 242         ret = i915_gem_object_unbind(obj, I915_GEM_OBJECT_UNBIND_ACTIVE);
 243         if (ret)
 244                 return ret;
 245
 246         ret = __i915_gem_object_put_pages(obj);
 247         if (ret)
 248                 return ret;
 249
 250         return 0;
 251 }
 252
 253 static void i915_ttm_free_cached_io_st(struct drm_i915_gem_object *obj)
 254 {
 255         struct radix_tree_iter iter;
 256         void __rcu **slot;
 257
 258         if (!obj->ttm.cached_io_st)
 259                 return;
 260
 261         rcu_read_lock();
 262         radix_tree_for_each_slot(slot, &obj->ttm.get_io_page.radix, &iter, 0)
 263                 radix_tree_delete(&obj->ttm.get_io_page.radix, iter.index);
 264         rcu_read_unlock();
 265
 266         sg_free_table(obj->ttm.cached_io_st);
 267         kfree(obj->ttm.cached_io_st);
 268         obj->ttm.cached_io_st = NULL;
 269 }
 270
 271 static void
 272 i915_ttm_adjust_domains_after_move(struct drm_i915_gem_object *obj)
 273 {
 274         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
 275
 276         if (cpu_maps_iomem(bo->resource) || bo->ttm->caching != ttm_cached) {
 277                 obj->write_domain = I915_GEM_DOMAIN_WC;
 278                 obj->read_domains = I915_GEM_DOMAIN_WC;
 279         } else {
 280                 obj->write_domain = I915_GEM_DOMAIN_CPU;
 281                 obj->read_domains = I915_GEM_DOMAIN_CPU;
 282         }
 283 }
 284
 285 static void i915_ttm_adjust_gem_after_move(struct drm_i915_gem_object *obj)
 286 {
 287         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
 288         unsigned int cache_level;
 289         unsigned int i;
 290
 291         /*
 292          * If object was moved to an allowable region, update the object
 293          * region to consider it migrated. Note that if it's currently not
 294          * in an allowable region, it's evicted and we don't update the
 295          * object region.
 296          */
 297         if (intel_region_to_ttm_type(obj->mm.region) != bo->resource->mem_type) {
 298                 for (i = 0; i < obj->mm.n_placements; ++i) {
 299                         struct intel_memory_region *mr = obj->mm.placements[i];
 300
 301                         if (intel_region_to_ttm_type(mr) == bo->resource->mem_type &&
 302                             mr != obj->mm.region) {
 303                                 i915_gem_object_release_memory_region(obj);
 304                                 i915_gem_object_init_memory_region(obj, mr);
 305                                 break;
 306                         }
 307                 }
 308         }
 309
 310         obj->mem_flags &= ~(I915_BO_FLAG_STRUCT_PAGE | I915_BO_FLAG_IOMEM);
 311
 312         obj->mem_flags |= cpu_maps_iomem(bo->resource) ? I915_BO_FLAG_IOMEM :
 313                 I915_BO_FLAG_STRUCT_PAGE;
 314
 315         cache_level = i915_ttm_cache_level(to_i915(bo->base.dev), bo->resource,
 316                                            bo->ttm);
 317         i915_gem_object_set_cache_coherency(obj, cache_level);
 318 }
 319
 320 static void i915_ttm_purge(struct drm_i915_gem_object *obj)
 321 {
 322         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
 323         struct ttm_operation_ctx ctx = {
 324                 .interruptible = true,
 325                 .no_wait_gpu = false,
 326         };
 327         struct ttm_placement place = {};
 328         int ret;
 329
 330         if (obj->mm.madv == __I915_MADV_PURGED)
 331                 return;
 332
 333         /* TTM's purge interface. Note that we might be reentering. */
 334         ret = ttm_bo_validate(bo, &place, &ctx);
 335         if (!ret) {
 336                 obj->write_domain = 0;
 337                 obj->read_domains = 0;
 338                 i915_ttm_adjust_gem_after_move(obj);
 339                 i915_ttm_free_cached_io_st(obj);
 340                 obj->mm.madv = __I915_MADV_PURGED;
 341         }
 342 }
 343
 344 static void i915_ttm_swap_notify(struct ttm_buffer_object *bo)
 345 {
 346         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 347         int ret = i915_ttm_move_notify(bo);
 348
 349         GEM_WARN_ON(ret);
 350         GEM_WARN_ON(obj->ttm.cached_io_st);
 351         if (!ret && obj->mm.madv != I915_MADV_WILLNEED)
 352                 i915_ttm_purge(obj);
 353 }
 354
 355 static void i915_ttm_delete_mem_notify(struct ttm_buffer_object *bo)
 356 {
 357         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 358
 359         if (likely(obj))
 360                 i915_ttm_free_cached_io_st(obj);
 361 }
 362
 363 static struct intel_memory_region *
 364 i915_ttm_region(struct ttm_device *bdev, int ttm_mem_type)
 365 {
 366         struct drm_i915_private *i915 = container_of(bdev, typeof(*i915), bdev);
 367
 368         /* There's some room for optimization here... */
 369         GEM_BUG_ON(ttm_mem_type != I915_PL_SYSTEM &&
 370                    ttm_mem_type < I915_PL_LMEM0);
 371         if (ttm_mem_type == I915_PL_SYSTEM)
 372                 return intel_memory_region_lookup(i915, INTEL_MEMORY_SYSTEM,
 373                                                   0);
 374
 375         return intel_memory_region_lookup(i915, INTEL_MEMORY_LOCAL,
 376                                           ttm_mem_type - I915_PL_LMEM0);
 377 }
 378
 379 static struct sg_table *i915_ttm_tt_get_st(struct ttm_tt *ttm)
 380 {
 381         struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
 382         struct sg_table *st;
 383         int ret;
 384
 385         if (i915_tt->cached_st)
 386                 return i915_tt->cached_st;
 387
 388         st = kzalloc(sizeof(*st), GFP_KERNEL);
 389         if (!st)
 390                 return ERR_PTR(-ENOMEM);
 391
 392         ret = sg_alloc_table_from_pages_segment(st,
 393                         ttm->pages, ttm->num_pages,
 394                         0, (unsigned long)ttm->num_pages << PAGE_SHIFT,
 395                         i915_sg_segment_size(), GFP_KERNEL);
 396         if (ret) {
 397                 kfree(st);
 398                 return ERR_PTR(ret);
 399         }
 400
 401         ret = dma_map_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 0);
 402         if (ret) {
 403                 sg_free_table(st);
 404                 kfree(st);
 405                 return ERR_PTR(ret);
 406         }
 407
 408         i915_tt->cached_st = st;
 409         return st;
 410 }
 411
 412 static struct sg_table *
 413 i915_ttm_resource_get_st(struct drm_i915_gem_object *obj,
 414                          struct ttm_resource *res)
 415 {
 416         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
 417
 418         if (!gpu_binds_iomem(res))
 419                 return i915_ttm_tt_get_st(bo->ttm);
 420
 421         /*
 422          * If CPU mapping differs, we need to add the ttm_tt pages to
 423          * the resulting st. Might make sense for GGTT.
 424          */
 425         GEM_WARN_ON(!cpu_maps_iomem(res));
 426         return intel_region_ttm_resource_to_st(obj->mm.region, res);
 427 }
 428
 429 static int i915_ttm_accel_move(struct ttm_buffer_object *bo,
 430                                bool clear,
 431                                struct ttm_resource *dst_mem,
 432                                struct ttm_tt *dst_ttm,
 433                                struct sg_table *dst_st)
 434 {
 435         struct drm_i915_private *i915 = container_of(bo->bdev, typeof(*i915),
 436                                                      bdev);
 437         struct ttm_resource_manager *src_man =
 438                 ttm_manager_type(bo->bdev, bo->resource->mem_type);
 439         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 440         struct sg_table *src_st;
 441         struct i915_request *rq;
 442         struct ttm_tt *src_ttm = bo->ttm;
 443         enum i915_cache_level src_level, dst_level;
 444         int ret;
 445
 446         if (!i915->gt.migrate.context)
 447                 return -EINVAL;
 448
 449         dst_level = i915_ttm_cache_level(i915, dst_mem, dst_ttm);
 450         if (clear) {
 451                 if (bo->type == ttm_bo_type_kernel)
 452                         return -EINVAL;
 453
 454                 intel_engine_pm_get(i915->gt.migrate.context->engine);
 455                 ret = intel_context_migrate_clear(i915->gt.migrate.context, NULL,
 456                                                   dst_st->sgl, dst_level,
 457                                                   gpu_binds_iomem(dst_mem),
 458                                                   0, &rq);
 459
 460                 if (!ret && rq) {
 461                         i915_request_wait(rq, 0, MAX_SCHEDULE_TIMEOUT);
 462                         i915_request_put(rq);
 463                 }
 464                 intel_engine_pm_put(i915->gt.migrate.context->engine);
 465         } else {
 466                 src_st = src_man->use_tt ? i915_ttm_tt_get_st(src_ttm) :
 467                         obj->ttm.cached_io_st;
 468
 469                 src_level = i915_ttm_cache_level(i915, bo->resource, src_ttm);
 470                 intel_engine_pm_get(i915->gt.migrate.context->engine);
 471                 ret = intel_context_migrate_copy(i915->gt.migrate.context,
 472                                                  NULL, src_st->sgl, src_level,
 473                                                  gpu_binds_iomem(bo->resource),
 474                                                  dst_st->sgl, dst_level,
 475                                                  gpu_binds_iomem(dst_mem),
 476                                                  &rq);
 477                 if (!ret && rq) {
 478                         i915_request_wait(rq, 0, MAX_SCHEDULE_TIMEOUT);
 479                         i915_request_put(rq);
 480                 }
 481                 intel_engine_pm_put(i915->gt.migrate.context->engine);
 482         }
 483
 484         return ret;
 485 }
 486
 487 static void __i915_ttm_move(struct ttm_buffer_object *bo, bool clear,
 488                             struct ttm_resource *dst_mem,
 489                             struct ttm_tt *dst_ttm,
 490                             struct sg_table *dst_st,
 491                             bool allow_accel)
 492 {
 493         int ret = -EINVAL;
 494
 495         if (allow_accel)
 496                 ret = i915_ttm_accel_move(bo, clear, dst_mem, dst_ttm, dst_st);
 497         if (ret) {
 498                 struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 499                 struct intel_memory_region *dst_reg, *src_reg;
 500                 union {
 501                         struct ttm_kmap_iter_tt tt;
 502                         struct ttm_kmap_iter_iomap io;
 503                 } _dst_iter, _src_iter;
 504                 struct ttm_kmap_iter *dst_iter, *src_iter;
 505
 506                 dst_reg = i915_ttm_region(bo->bdev, dst_mem->mem_type);
 507                 src_reg = i915_ttm_region(bo->bdev, bo->resource->mem_type);
 508                 GEM_BUG_ON(!dst_reg || !src_reg);
 509
 510                 dst_iter = !cpu_maps_iomem(dst_mem) ?
 511                         ttm_kmap_iter_tt_init(&_dst_iter.tt, dst_ttm) :
 512                         ttm_kmap_iter_iomap_init(&_dst_iter.io, &dst_reg->iomap,
 513                                                  dst_st, dst_reg->region.start);
 514
 515                 src_iter = !cpu_maps_iomem(bo->resource) ?
 516                         ttm_kmap_iter_tt_init(&_src_iter.tt, bo->ttm) :
 517                         ttm_kmap_iter_iomap_init(&_src_iter.io, &src_reg->iomap,
 518                                                  obj->ttm.cached_io_st,
 519                                                  src_reg->region.start);
 520
 521                 ttm_move_memcpy(clear, dst_mem->num_pages, dst_iter, src_iter);
 522         }
 523 }
 524
 525 static int i915_ttm_move(struct ttm_buffer_object *bo, bool evict,
 526                          struct ttm_operation_ctx *ctx,
 527                          struct ttm_resource *dst_mem,
 528                          struct ttm_place *hop)
 529 {
 530         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 531         struct ttm_resource_manager *dst_man =
 532                 ttm_manager_type(bo->bdev, dst_mem->mem_type);
 533         struct ttm_tt *ttm = bo->ttm;
 534         struct sg_table *dst_st;
 535         bool clear;
 536         int ret;
 537
 538         /* Sync for now. We could do the actual copy async. */
 539         ret = ttm_bo_wait_ctx(bo, ctx);
 540         if (ret)
 541                 return ret;
 542
 543         ret = i915_ttm_move_notify(bo);
 544         if (ret)
 545                 return ret;
 546
 547         if (obj->mm.madv != I915_MADV_WILLNEED) {
 548                 i915_ttm_purge(obj);
 549                 ttm_resource_free(bo, &dst_mem);
 550                 return 0;
 551         }
 552
 553         /* Populate ttm with pages if needed. Typically system memory. */
 554         if (ttm && (dst_man->use_tt || (ttm->page_flags & TTM_PAGE_FLAG_SWAPPED))) {
 555                 ret = ttm_tt_populate(bo->bdev, ttm, ctx);
 556                 if (ret)
 557                         return ret;
 558         }
 559
 560         dst_st = i915_ttm_resource_get_st(obj, dst_mem);
 561         if (IS_ERR(dst_st))
 562                 return PTR_ERR(dst_st);
 563
 564         clear = !cpu_maps_iomem(bo->resource) && (!ttm || !ttm_tt_is_populated(ttm));
 565         if (!(clear && ttm && !(ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)))
 566                 __i915_ttm_move(bo, clear, dst_mem, bo->ttm, dst_st, true);
 567
 568         ttm_bo_move_sync_cleanup(bo, dst_mem);
 569         i915_ttm_adjust_domains_after_move(obj);
 570         i915_ttm_free_cached_io_st(obj);
 571
 572         if (gpu_binds_iomem(dst_mem) || cpu_maps_iomem(dst_mem)) {
 573                 obj->ttm.cached_io_st = dst_st;
 574                 obj->ttm.get_io_page.sg_pos = dst_st->sgl;
 575                 obj->ttm.get_io_page.sg_idx = 0;
 576         }
 577
 578         i915_ttm_adjust_gem_after_move(obj);
 579         return 0;
 580 }
 581
 582 static int i915_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *mem)
 583 {
 584         if (!cpu_maps_iomem(mem))
 585                 return 0;
 586
 587         mem->bus.caching = ttm_write_combined;
 588         mem->bus.is_iomem = true;
 589
 590         return 0;
 591 }
 592
 593 static unsigned long i915_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
 594                                          unsigned long page_offset)
 595 {
 596         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 597         unsigned long base = obj->mm.region->iomap.base - obj->mm.region->region.start;
 598         struct scatterlist *sg;
 599         unsigned int ofs;
 600
 601         GEM_WARN_ON(bo->ttm);
 602
 603         sg = __i915_gem_object_get_sg(obj, &obj->ttm.get_io_page, page_offset, &ofs, true);
 604
 605         return ((base + sg_dma_address(sg)) >> PAGE_SHIFT) + ofs;
 606 }
 607
 608 static struct ttm_device_funcs i915_ttm_bo_driver = {
 609         .ttm_tt_create = i915_ttm_tt_create,
 610         .ttm_tt_unpopulate = i915_ttm_tt_unpopulate,
 611         .ttm_tt_destroy = i915_ttm_tt_destroy,
 612         .eviction_valuable = i915_ttm_eviction_valuable,
 613         .evict_flags = i915_ttm_evict_flags,
 614         .move = i915_ttm_move,
 615         .swap_notify = i915_ttm_swap_notify,
 616         .delete_mem_notify = i915_ttm_delete_mem_notify,
 617         .io_mem_reserve = i915_ttm_io_mem_reserve,
 618         .io_mem_pfn = i915_ttm_io_mem_pfn,
 619 };
 620
 621 /**
 622  * i915_ttm_driver - Return a pointer to the TTM device funcs
 623  *
 624  * Return: Pointer to statically allocated TTM device funcs.
 625  */
 626 struct ttm_device_funcs *i915_ttm_driver(void)
 627 {
 628         return &i915_ttm_bo_driver;
 629 }
 630
 631 static int __i915_ttm_get_pages(struct drm_i915_gem_object *obj,
 632                                 struct ttm_placement *placement)
 633 {
 634         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
 635         struct ttm_operation_ctx ctx = {
 636                 .interruptible = true,
 637                 .no_wait_gpu = false,
 638         };
 639         struct sg_table *st;
 640         int real_num_busy;
 641         int ret;
 642
 643         /* First try only the requested placement. No eviction. */
 644         real_num_busy = fetch_and_zero(&placement->num_busy_placement);
 645         ret = ttm_bo_validate(bo, placement, &ctx);
 646         if (ret) {
 647                 ret = i915_ttm_err_to_gem(ret);
 648                 /*
 649                  * Anything that wants to restart the operation gets to
 650                  * do that.
 651                  */
 652                 if (ret == -EDEADLK || ret == -EINTR || ret == -ERESTARTSYS ||
 653                     ret == -EAGAIN)
 654                         return ret;
 655
 656                 /*
 657                  * If the initial attempt fails, allow all accepted placements,
 658                  * evicting if necessary.
 659                  */
 660                 placement->num_busy_placement = real_num_busy;
 661                 ret = ttm_bo_validate(bo, placement, &ctx);
 662                 if (ret)
 663                         return i915_ttm_err_to_gem(ret);
 664         }
 665
 666         i915_ttm_adjust_lru(obj);
 667         if (bo->ttm && !ttm_tt_is_populated(bo->ttm)) {
 668                 ret = ttm_tt_populate(bo->bdev, bo->ttm, &ctx);
 669                 if (ret)
 670                         return ret;
 671
 672                 i915_ttm_adjust_domains_after_move(obj);
 673                 i915_ttm_adjust_gem_after_move(obj);
 674         }
 675
 676         if (!i915_gem_object_has_pages(obj)) {
 677                 /* Object either has a page vector or is an iomem object */
 678                 st = bo->ttm ? i915_ttm_tt_get_st(bo->ttm) : obj->ttm.cached_io_st;
 679                 if (IS_ERR(st))
 680                         return PTR_ERR(st);
 681
 682                 __i915_gem_object_set_pages(obj, st, i915_sg_dma_sizes(st->sgl));
 683         }
 684
 685         return ret;
 686 }
 687
 688 static int i915_ttm_get_pages(struct drm_i915_gem_object *obj)
 689 {
 690         struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS];
 691         struct ttm_placement placement;
 692
 693         GEM_BUG_ON(obj->mm.n_placements > I915_TTM_MAX_PLACEMENTS);
 694
 695         /* Move to the requested placement. */
 696         i915_ttm_placement_from_obj(obj, &requested, busy, &placement);
 697
 698         return __i915_ttm_get_pages(obj, &placement);
 699 }
 700
 701 /**
 702  * DOC: Migration vs eviction
 703  *
 704  * GEM migration may not be the same as TTM migration / eviction. If
 705  * the TTM core decides to evict an object it may be evicted to a
 706  * TTM memory type that is not in the object's allowable GEM regions, or
 707  * in fact theoretically to a TTM memory type that doesn't correspond to
 708  * a GEM memory region. In that case the object's GEM region is not
 709  * updated, and the data is migrated back to the GEM region at
 710  * get_pages time. TTM may however set up CPU ptes to the object even
 711  * when it is evicted.
 712  * Gem forced migration using the i915_ttm_migrate() op, is allowed even
 713  * to regions that are not in the object's list of allowable placements.
 714  */
 715 static int i915_ttm_migrate(struct drm_i915_gem_object *obj,
 716                             struct intel_memory_region *mr)
 717 {
 718         struct ttm_place requested;
 719         struct ttm_placement placement;
 720         int ret;
 721
 722         i915_ttm_place_from_region(mr, &requested, obj->flags);
 723         placement.num_placement = 1;
 724         placement.num_busy_placement = 1;
 725         placement.placement = &requested;
 726         placement.busy_placement = &requested;
 727
 728         ret = __i915_ttm_get_pages(obj, &placement);
 729         if (ret)
 730                 return ret;
 731
 732         /*
 733          * Reinitialize the region bindings. This is primarily
 734          * required for objects where the new region is not in
 735          * its allowable placements.
 736          */
 737         if (obj->mm.region != mr) {
 738                 i915_gem_object_release_memory_region(obj);
 739                 i915_gem_object_init_memory_region(obj, mr);
 740         }
 741
 742         return 0;
 743 }
 744
 745 static void i915_ttm_put_pages(struct drm_i915_gem_object *obj,
 746                                struct sg_table *st)
 747 {
 748         /*
 749          * We're currently not called from a shrinker, so put_pages()
 750          * typically means the object is about to destroyed, or called
 751          * from move_notify(). So just avoid doing much for now.
 752          * If the object is not destroyed next, The TTM eviction logic
 753          * and shrinkers will move it out if needed.
 754          */
 755
 756         i915_ttm_adjust_lru(obj);
 757 }
 758
 759 static void i915_ttm_adjust_lru(struct drm_i915_gem_object *obj)
 760 {
 761         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
 762
 763         /*
 764          * Don't manipulate the TTM LRUs while in TTM bo destruction.
 765          * We're called through i915_ttm_delete_mem_notify().
 766          */
 767         if (!kref_read(&bo->kref))
 768                 return;
 769
 770         /*
 771          * Put on the correct LRU list depending on the MADV status
 772          */
 773         spin_lock(&bo->bdev->lru_lock);
 774         if (obj->mm.madv != I915_MADV_WILLNEED) {
 775                 bo->priority = I915_TTM_PRIO_PURGE;
 776         } else if (!i915_gem_object_has_pages(obj)) {
 777                 if (bo->priority < I915_TTM_PRIO_HAS_PAGES)
 778                         bo->priority = I915_TTM_PRIO_HAS_PAGES;
 779         } else {
 780                 if (bo->priority > I915_TTM_PRIO_NO_PAGES)
 781                         bo->priority = I915_TTM_PRIO_NO_PAGES;
 782         }
 783
 784         ttm_bo_move_to_lru_tail(bo, bo->resource, NULL);
 785         spin_unlock(&bo->bdev->lru_lock);
 786 }
 787
 788 /*
 789  * TTM-backed gem object destruction requires some clarification.
 790  * Basically we have two possibilities here. We can either rely on the
 791  * i915 delayed destruction and put the TTM object when the object
 792  * is idle. This would be detected by TTM which would bypass the
 793  * TTM delayed destroy handling. The other approach is to put the TTM
 794  * object early and rely on the TTM destroyed handling, and then free
 795  * the leftover parts of the GEM object once TTM's destroyed list handling is
 796  * complete. For now, we rely on the latter for two reasons:
 797  * a) TTM can evict an object even when it's on the delayed destroy list,
 798  * which in theory allows for complete eviction.
 799  * b) There is work going on in TTM to allow freeing an object even when
 800  * it's not idle, and using the TTM destroyed list handling could help us
 801  * benefit from that.
 802  */
 803 static void i915_ttm_delayed_free(struct drm_i915_gem_object *obj)
 804 {
 805         if (obj->ttm.created) {
 806                 ttm_bo_put(i915_gem_to_ttm(obj));
 807         } else {
 808                 __i915_gem_free_object(obj);
 809                 call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
 810         }
 811 }
 812
 813 static vm_fault_t vm_fault_ttm(struct vm_fault *vmf)
 814 {
 815         struct vm_area_struct *area = vmf->vma;
 816         struct drm_i915_gem_object *obj =
 817                 i915_ttm_to_gem(area->vm_private_data);
 818
 819         /* Sanity check that we allow writing into this object */
 820         if (unlikely(i915_gem_object_is_readonly(obj) &&
 821                      area->vm_flags & VM_WRITE))
 822                 return VM_FAULT_SIGBUS;
 823
 824         return ttm_bo_vm_fault(vmf);
 825 }
 826
 827 static int
 828 vm_access_ttm(struct vm_area_struct *area, unsigned long addr,
 829               void *buf, int len, int write)
 830 {
 831         struct drm_i915_gem_object *obj =
 832                 i915_ttm_to_gem(area->vm_private_data);
 833
 834         if (i915_gem_object_is_readonly(obj) && write)
 835                 return -EACCES;
 836
 837         return ttm_bo_vm_access(area, addr, buf, len, write);
 838 }
 839
 840 static void ttm_vm_open(struct vm_area_struct *vma)
 841 {
 842         struct drm_i915_gem_object *obj =
 843                 i915_ttm_to_gem(vma->vm_private_data);
 844
 845         GEM_BUG_ON(!obj);
 846         i915_gem_object_get(obj);
 847 }
 848
 849 static void ttm_vm_close(struct vm_area_struct *vma)
 850 {
 851         struct drm_i915_gem_object *obj =
 852                 i915_ttm_to_gem(vma->vm_private_data);
 853
 854         GEM_BUG_ON(!obj);
 855         i915_gem_object_put(obj);
 856 }
 857
 858 static const struct vm_operations_struct vm_ops_ttm = {
 859         .fault = vm_fault_ttm,
 860         .access = vm_access_ttm,
 861         .open = ttm_vm_open,
 862         .close = ttm_vm_close,
 863 };
 864
 865 static u64 i915_ttm_mmap_offset(struct drm_i915_gem_object *obj)
 866 {
 867         /* The ttm_bo must be allocated with I915_BO_ALLOC_USER */
 868         GEM_BUG_ON(!drm_mm_node_allocated(&obj->base.vma_node.vm_node));
 869
 870         return drm_vma_node_offset_addr(&obj->base.vma_node);
 871 }
 872
 873 static const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = {
 874         .name = "i915_gem_object_ttm",
 875
 876         .get_pages = i915_ttm_get_pages,
 877         .put_pages = i915_ttm_put_pages,
 878         .truncate = i915_ttm_purge,
 879         .adjust_lru = i915_ttm_adjust_lru,
 880         .delayed_free = i915_ttm_delayed_free,
 881         .migrate = i915_ttm_migrate,
 882         .mmap_offset = i915_ttm_mmap_offset,
 883         .mmap_ops = &vm_ops_ttm,
 884 };
 885
 886 void i915_ttm_bo_destroy(struct ttm_buffer_object *bo)
 887 {
 888         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 889
 890         /* This releases all gem object bindings to the backend. */
 891         __i915_gem_free_object(obj);
 892
 893         i915_gem_object_release_memory_region(obj);
 894         mutex_destroy(&obj->ttm.get_io_page.lock);
 895
 896         if (obj->ttm.created)
 897                 call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
 898 }
 899
 900 /**
 901  * __i915_gem_ttm_object_init - Initialize a ttm-backed i915 gem object
 902  * @mem: The initial memory region for the object.
 903  * @obj: The gem object.
 904  * @size: Object size in bytes.
 905  * @flags: gem object flags.
 906  *
 907  * Return: 0 on success, negative error code on failure.
 908  */
 909 int __i915_gem_ttm_object_init(struct intel_memory_region *mem,
 910                                struct drm_i915_gem_object *obj,
 911                                resource_size_t size,
 912                                resource_size_t page_size,
 913                                unsigned int flags)
 914 {
 915         static struct lock_class_key lock_class;
 916         struct drm_i915_private *i915 = mem->i915;
 917         struct ttm_operation_ctx ctx = {
 918                 .interruptible = true,
 919                 .no_wait_gpu = false,
 920         };
 921         enum ttm_bo_type bo_type;
 922         int ret;
 923
 924         drm_gem_private_object_init(&i915->drm, &obj->base, size);
 925         i915_gem_object_init(obj, &i915_gem_ttm_obj_ops, &lock_class, flags);
 926         i915_gem_object_init_memory_region(obj, mem);
 927         i915_gem_object_make_unshrinkable(obj);
 928         INIT_RADIX_TREE(&obj->ttm.get_io_page.radix, GFP_KERNEL | __GFP_NOWARN);
 929         mutex_init(&obj->ttm.get_io_page.lock);
 930         bo_type = (obj->flags & I915_BO_ALLOC_USER) ? ttm_bo_type_device :
 931                 ttm_bo_type_kernel;
 932
 933         obj->base.vma_node.driver_private = i915_gem_to_ttm(obj);
 934
 935         /* Forcing the page size is kernel internal only */
 936         GEM_BUG_ON(page_size && obj->mm.n_placements);
 937
 938         /*
 939          * If this function fails, it will call the destructor, but
 940          * our caller still owns the object. So no freeing in the
 941          * destructor until obj->ttm.created is true.
 942          * Similarly, in delayed_destroy, we can't call ttm_bo_put()
 943          * until successful initialization.
 944          */
 945         ret = ttm_bo_init_reserved(&i915->bdev, i915_gem_to_ttm(obj), size,
 946                                    bo_type, &i915_sys_placement,
 947                                    page_size >> PAGE_SHIFT,
 948                                    &ctx, NULL, NULL, i915_ttm_bo_destroy);
 949         if (ret)
 950                 return i915_ttm_err_to_gem(ret);
 951
 952         obj->ttm.created = true;
 953         i915_ttm_adjust_domains_after_move(obj);
 954         i915_ttm_adjust_gem_after_move(obj);
 955         i915_gem_object_unlock(obj);
 956
 957         return 0;
 958 }
 959
 960 static const struct intel_memory_region_ops ttm_system_region_ops = {
 961         .init_object = __i915_gem_ttm_object_init,
 962 };
 963
 964 struct intel_memory_region *
 965 i915_gem_ttm_system_setup(struct drm_i915_private *i915,
 966                           u16 type, u16 instance)
 967 {
 968         struct intel_memory_region *mr;
 969
 970         mr = intel_memory_region_create(i915, 0,
 971                                         totalram_pages() << PAGE_SHIFT,
 972                                         PAGE_SIZE, 0,
 973                                         type, instance,
 974                                         &ttm_system_region_ops);
 975         if (IS_ERR(mr))
 976                 return mr;
 977
 978         intel_memory_region_set_name(mr, "system-ttm");
 979         return mr;
 980 }
 981
 982 /**
 983  * i915_gem_obj_copy_ttm - Copy the contents of one ttm-based gem object to
 984  * another
 985  * @dst: The destination object
 986  * @src: The source object
 987  * @allow_accel: Allow using the blitter. Otherwise TTM memcpy is used.
 988  * @intr: Whether to perform waits interruptible:
 989  *
 990  * Note: The caller is responsible for assuring that the underlying
 991  * TTM objects are populated if needed and locked.
 992  *
 993  * Return: Zero on success. Negative error code on error. If @intr == true,
 994  * then it may return -ERESTARTSYS or -EINTR.
 995  */
 996 int i915_gem_obj_copy_ttm(struct drm_i915_gem_object *dst,
 997                           struct drm_i915_gem_object *src,
 998                           bool allow_accel, bool intr)
 999 {
1000         struct ttm_buffer_object *dst_bo = i915_gem_to_ttm(dst);
1001         struct ttm_buffer_object *src_bo = i915_gem_to_ttm(src);
1002         struct ttm_operation_ctx ctx = {
1003                 .interruptible = intr,
1004         };
1005         struct sg_table *dst_st;
1006         int ret;
1007
1008         assert_object_held(dst);
1009         assert_object_held(src);
1010
1011         /*
1012          * Sync for now. This will change with async moves.
1013          */
1014         ret = ttm_bo_wait_ctx(dst_bo, &ctx);
1015         if (!ret)
1016                 ret = ttm_bo_wait_ctx(src_bo, &ctx);
1017         if (ret)
1018                 return ret;
1019
1020         dst_st = gpu_binds_iomem(dst_bo->resource) ?
1021                 dst->ttm.cached_io_st : i915_ttm_tt_get_st(dst_bo->ttm);
1022
1023         __i915_ttm_move(src_bo, false, dst_bo->resource, dst_bo->ttm,
1024                         dst_st, allow_accel);
1025
1026         return 0;
1027 }