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_fini(ttm);
 218         kfree(i915_tt);
 219 }
 220
 221 static bool i915_ttm_eviction_valuable(struct ttm_buffer_object *bo,
 222                                        const struct ttm_place *place)
 223 {
 224         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 225
 226         /* Will do for now. Our pinned objects are still on TTM's LRU lists */
 227         return i915_gem_object_evictable(obj);
 228 }
 229
 230 static void i915_ttm_evict_flags(struct ttm_buffer_object *bo,
 231                                  struct ttm_placement *placement)
 232 {
 233         *placement = i915_sys_placement;
 234 }
 235
 236 static int i915_ttm_move_notify(struct ttm_buffer_object *bo)
 237 {
 238         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 239         int ret;
 240
 241         ret = i915_gem_object_unbind(obj, I915_GEM_OBJECT_UNBIND_ACTIVE);
 242         if (ret)
 243                 return ret;
 244
 245         ret = __i915_gem_object_put_pages(obj);
 246         if (ret)
 247                 return ret;
 248
 249         return 0;
 250 }
 251
 252 static void i915_ttm_free_cached_io_st(struct drm_i915_gem_object *obj)
 253 {
 254         struct radix_tree_iter iter;
 255         void __rcu **slot;
 256
 257         if (!obj->ttm.cached_io_st)
 258                 return;
 259
 260         rcu_read_lock();
 261         radix_tree_for_each_slot(slot, &obj->ttm.get_io_page.radix, &iter, 0)
 262                 radix_tree_delete(&obj->ttm.get_io_page.radix, iter.index);
 263         rcu_read_unlock();
 264
 265         sg_free_table(obj->ttm.cached_io_st);
 266         kfree(obj->ttm.cached_io_st);
 267         obj->ttm.cached_io_st = NULL;
 268 }
 269
 270 static void
 271 i915_ttm_adjust_domains_after_move(struct drm_i915_gem_object *obj)
 272 {
 273         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
 274
 275         if (cpu_maps_iomem(bo->resource) || bo->ttm->caching != ttm_cached) {
 276                 obj->write_domain = I915_GEM_DOMAIN_WC;
 277                 obj->read_domains = I915_GEM_DOMAIN_WC;
 278         } else {
 279                 obj->write_domain = I915_GEM_DOMAIN_CPU;
 280                 obj->read_domains = I915_GEM_DOMAIN_CPU;
 281         }
 282 }
 283
 284 static void i915_ttm_adjust_gem_after_move(struct drm_i915_gem_object *obj)
 285 {
 286         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
 287         unsigned int cache_level;
 288         unsigned int i;
 289
 290         /*
 291          * If object was moved to an allowable region, update the object
 292          * region to consider it migrated. Note that if it's currently not
 293          * in an allowable region, it's evicted and we don't update the
 294          * object region.
 295          */
 296         if (intel_region_to_ttm_type(obj->mm.region) != bo->resource->mem_type) {
 297                 for (i = 0; i < obj->mm.n_placements; ++i) {
 298                         struct intel_memory_region *mr = obj->mm.placements[i];
 299
 300                         if (intel_region_to_ttm_type(mr) == bo->resource->mem_type &&
 301                             mr != obj->mm.region) {
 302                                 i915_gem_object_release_memory_region(obj);
 303                                 i915_gem_object_init_memory_region(obj, mr);
 304                                 break;
 305                         }
 306                 }
 307         }
 308
 309         obj->mem_flags &= ~(I915_BO_FLAG_STRUCT_PAGE | I915_BO_FLAG_IOMEM);
 310
 311         obj->mem_flags |= cpu_maps_iomem(bo->resource) ? I915_BO_FLAG_IOMEM :
 312                 I915_BO_FLAG_STRUCT_PAGE;
 313
 314         cache_level = i915_ttm_cache_level(to_i915(bo->base.dev), bo->resource,
 315                                            bo->ttm);
 316         i915_gem_object_set_cache_coherency(obj, cache_level);
 317 }
 318
 319 static void i915_ttm_purge(struct drm_i915_gem_object *obj)
 320 {
 321         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
 322         struct ttm_operation_ctx ctx = {
 323                 .interruptible = true,
 324                 .no_wait_gpu = false,
 325         };
 326         struct ttm_placement place = {};
 327         int ret;
 328
 329         if (obj->mm.madv == __I915_MADV_PURGED)
 330                 return;
 331
 332         /* TTM's purge interface. Note that we might be reentering. */
 333         ret = ttm_bo_validate(bo, &place, &ctx);
 334         if (!ret) {
 335                 obj->write_domain = 0;
 336                 obj->read_domains = 0;
 337                 i915_ttm_adjust_gem_after_move(obj);
 338                 i915_ttm_free_cached_io_st(obj);
 339                 obj->mm.madv = __I915_MADV_PURGED;
 340         }
 341 }
 342
 343 static void i915_ttm_swap_notify(struct ttm_buffer_object *bo)
 344 {
 345         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 346         int ret = i915_ttm_move_notify(bo);
 347
 348         GEM_WARN_ON(ret);
 349         GEM_WARN_ON(obj->ttm.cached_io_st);
 350         if (!ret && obj->mm.madv != I915_MADV_WILLNEED)
 351                 i915_ttm_purge(obj);
 352 }
 353
 354 static void i915_ttm_delete_mem_notify(struct ttm_buffer_object *bo)
 355 {
 356         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 357
 358         if (likely(obj)) {
 359                 /* This releases all gem object bindings to the backend. */
 360                 i915_ttm_free_cached_io_st(obj);
 361                 __i915_gem_free_object(obj);
 362         }
 363 }
 364
 365 static struct intel_memory_region *
 366 i915_ttm_region(struct ttm_device *bdev, int ttm_mem_type)
 367 {
 368         struct drm_i915_private *i915 = container_of(bdev, typeof(*i915), bdev);
 369
 370         /* There's some room for optimization here... */
 371         GEM_BUG_ON(ttm_mem_type != I915_PL_SYSTEM &&
 372                    ttm_mem_type < I915_PL_LMEM0);
 373         if (ttm_mem_type == I915_PL_SYSTEM)
 374                 return intel_memory_region_lookup(i915, INTEL_MEMORY_SYSTEM,
 375                                                   0);
 376
 377         return intel_memory_region_lookup(i915, INTEL_MEMORY_LOCAL,
 378                                           ttm_mem_type - I915_PL_LMEM0);
 379 }
 380
 381 static struct sg_table *i915_ttm_tt_get_st(struct ttm_tt *ttm)
 382 {
 383         struct i915_ttm_tt *i915_tt = container_of(ttm, typeof(*i915_tt), ttm);
 384         struct sg_table *st;
 385         int ret;
 386
 387         if (i915_tt->cached_st)
 388                 return i915_tt->cached_st;
 389
 390         st = kzalloc(sizeof(*st), GFP_KERNEL);
 391         if (!st)
 392                 return ERR_PTR(-ENOMEM);
 393
 394         ret = sg_alloc_table_from_pages_segment(st,
 395                         ttm->pages, ttm->num_pages,
 396                         0, (unsigned long)ttm->num_pages << PAGE_SHIFT,
 397                         i915_sg_segment_size(), GFP_KERNEL);
 398         if (ret) {
 399                 kfree(st);
 400                 return ERR_PTR(ret);
 401         }
 402
 403         ret = dma_map_sgtable(i915_tt->dev, st, DMA_BIDIRECTIONAL, 0);
 404         if (ret) {
 405                 sg_free_table(st);
 406                 kfree(st);
 407                 return ERR_PTR(ret);
 408         }
 409
 410         i915_tt->cached_st = st;
 411         return st;
 412 }
 413
 414 static struct sg_table *
 415 i915_ttm_resource_get_st(struct drm_i915_gem_object *obj,
 416                          struct ttm_resource *res)
 417 {
 418         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
 419
 420         if (!gpu_binds_iomem(res))
 421                 return i915_ttm_tt_get_st(bo->ttm);
 422
 423         /*
 424          * If CPU mapping differs, we need to add the ttm_tt pages to
 425          * the resulting st. Might make sense for GGTT.
 426          */
 427         GEM_WARN_ON(!cpu_maps_iomem(res));
 428         return intel_region_ttm_resource_to_st(obj->mm.region, res);
 429 }
 430
 431 static int i915_ttm_accel_move(struct ttm_buffer_object *bo,
 432                                struct ttm_resource *dst_mem,
 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 *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, ttm);
 450         if (!ttm || !ttm_tt_is_populated(ttm)) {
 451                 if (bo->type == ttm_bo_type_kernel)
 452                         return -EINVAL;
 453
 454                 if (ttm && !(ttm->page_flags & TTM_PAGE_FLAG_ZERO_ALLOC))
 455                         return 0;
 456
 457                 intel_engine_pm_get(i915->gt.migrate.context->engine);
 458                 ret = intel_context_migrate_clear(i915->gt.migrate.context, NULL,
 459                                                   dst_st->sgl, dst_level,
 460                                                   gpu_binds_iomem(dst_mem),
 461                                                   0, &rq);
 462
 463                 if (!ret && rq) {
 464                         i915_request_wait(rq, 0, MAX_SCHEDULE_TIMEOUT);
 465                         i915_request_put(rq);
 466                 }
 467                 intel_engine_pm_put(i915->gt.migrate.context->engine);
 468         } else {
 469                 src_st = src_man->use_tt ? i915_ttm_tt_get_st(ttm) :
 470                         obj->ttm.cached_io_st;
 471
 472                 src_level = i915_ttm_cache_level(i915, bo->resource, ttm);
 473                 intel_engine_pm_get(i915->gt.migrate.context->engine);
 474                 ret = intel_context_migrate_copy(i915->gt.migrate.context,
 475                                                  NULL, src_st->sgl, src_level,
 476                                                  gpu_binds_iomem(bo->resource),
 477                                                  dst_st->sgl, dst_level,
 478                                                  gpu_binds_iomem(dst_mem),
 479                                                  &rq);
 480                 if (!ret && rq) {
 481                         i915_request_wait(rq, 0, MAX_SCHEDULE_TIMEOUT);
 482                         i915_request_put(rq);
 483                 }
 484                 intel_engine_pm_put(i915->gt.migrate.context->engine);
 485         }
 486
 487         return ret;
 488 }
 489
 490 static int i915_ttm_move(struct ttm_buffer_object *bo, bool evict,
 491                          struct ttm_operation_ctx *ctx,
 492                          struct ttm_resource *dst_mem,
 493                          struct ttm_place *hop)
 494 {
 495         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 496         struct ttm_resource_manager *dst_man =
 497                 ttm_manager_type(bo->bdev, dst_mem->mem_type);
 498         struct intel_memory_region *dst_reg, *src_reg;
 499         union {
 500                 struct ttm_kmap_iter_tt tt;
 501                 struct ttm_kmap_iter_iomap io;
 502         } _dst_iter, _src_iter;
 503         struct ttm_kmap_iter *dst_iter, *src_iter;
 504         struct sg_table *dst_st;
 505         int ret;
 506
 507         dst_reg = i915_ttm_region(bo->bdev, dst_mem->mem_type);
 508         src_reg = i915_ttm_region(bo->bdev, bo->resource->mem_type);
 509         GEM_BUG_ON(!dst_reg || !src_reg);
 510
 511         /* Sync for now. We could do the actual copy async. */
 512         ret = ttm_bo_wait_ctx(bo, ctx);
 513         if (ret)
 514                 return ret;
 515
 516         ret = i915_ttm_move_notify(bo);
 517         if (ret)
 518                 return ret;
 519
 520         if (obj->mm.madv != I915_MADV_WILLNEED) {
 521                 i915_ttm_purge(obj);
 522                 ttm_resource_free(bo, &dst_mem);
 523                 return 0;
 524         }
 525
 526         /* Populate ttm with pages if needed. Typically system memory. */
 527         if (bo->ttm && (dst_man->use_tt ||
 528                         (bo->ttm->page_flags & TTM_PAGE_FLAG_SWAPPED))) {
 529                 ret = ttm_tt_populate(bo->bdev, bo->ttm, ctx);
 530                 if (ret)
 531                         return ret;
 532         }
 533
 534         dst_st = i915_ttm_resource_get_st(obj, dst_mem);
 535         if (IS_ERR(dst_st))
 536                 return PTR_ERR(dst_st);
 537
 538         ret = i915_ttm_accel_move(bo, dst_mem, dst_st);
 539         if (ret) {
 540                 /* If we start mapping GGTT, we can no longer use man::use_tt here. */
 541                 dst_iter = !cpu_maps_iomem(dst_mem) ?
 542                         ttm_kmap_iter_tt_init(&_dst_iter.tt, bo->ttm) :
 543                         ttm_kmap_iter_iomap_init(&_dst_iter.io, &dst_reg->iomap,
 544                                                  dst_st, dst_reg->region.start);
 545
 546                 src_iter = !cpu_maps_iomem(bo->resource) ?
 547                         ttm_kmap_iter_tt_init(&_src_iter.tt, bo->ttm) :
 548                         ttm_kmap_iter_iomap_init(&_src_iter.io, &src_reg->iomap,
 549                                                  obj->ttm.cached_io_st,
 550                                                  src_reg->region.start);
 551
 552                 ttm_move_memcpy(bo, dst_mem->num_pages, dst_iter, src_iter);
 553         }
 554         /* Below dst_mem becomes bo->resource. */
 555         ttm_bo_move_sync_cleanup(bo, dst_mem);
 556         i915_ttm_adjust_domains_after_move(obj);
 557         i915_ttm_free_cached_io_st(obj);
 558
 559         if (gpu_binds_iomem(dst_mem) || cpu_maps_iomem(dst_mem)) {
 560                 obj->ttm.cached_io_st = dst_st;
 561                 obj->ttm.get_io_page.sg_pos = dst_st->sgl;
 562                 obj->ttm.get_io_page.sg_idx = 0;
 563         }
 564
 565         i915_ttm_adjust_gem_after_move(obj);
 566         return 0;
 567 }
 568
 569 static int i915_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *mem)
 570 {
 571         if (!cpu_maps_iomem(mem))
 572                 return 0;
 573
 574         mem->bus.caching = ttm_write_combined;
 575         mem->bus.is_iomem = true;
 576
 577         return 0;
 578 }
 579
 580 static unsigned long i915_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
 581                                          unsigned long page_offset)
 582 {
 583         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 584         unsigned long base = obj->mm.region->iomap.base - obj->mm.region->region.start;
 585         struct scatterlist *sg;
 586         unsigned int ofs;
 587
 588         GEM_WARN_ON(bo->ttm);
 589
 590         sg = __i915_gem_object_get_sg(obj, &obj->ttm.get_io_page, page_offset, &ofs, true);
 591
 592         return ((base + sg_dma_address(sg)) >> PAGE_SHIFT) + ofs;
 593 }
 594
 595 static struct ttm_device_funcs i915_ttm_bo_driver = {
 596         .ttm_tt_create = i915_ttm_tt_create,
 597         .ttm_tt_unpopulate = i915_ttm_tt_unpopulate,
 598         .ttm_tt_destroy = i915_ttm_tt_destroy,
 599         .eviction_valuable = i915_ttm_eviction_valuable,
 600         .evict_flags = i915_ttm_evict_flags,
 601         .move = i915_ttm_move,
 602         .swap_notify = i915_ttm_swap_notify,
 603         .delete_mem_notify = i915_ttm_delete_mem_notify,
 604         .io_mem_reserve = i915_ttm_io_mem_reserve,
 605         .io_mem_pfn = i915_ttm_io_mem_pfn,
 606 };
 607
 608 /**
 609  * i915_ttm_driver - Return a pointer to the TTM device funcs
 610  *
 611  * Return: Pointer to statically allocated TTM device funcs.
 612  */
 613 struct ttm_device_funcs *i915_ttm_driver(void)
 614 {
 615         return &i915_ttm_bo_driver;
 616 }
 617
 618 static int __i915_ttm_get_pages(struct drm_i915_gem_object *obj,
 619                                 struct ttm_placement *placement)
 620 {
 621         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
 622         struct ttm_operation_ctx ctx = {
 623                 .interruptible = true,
 624                 .no_wait_gpu = false,
 625         };
 626         struct sg_table *st;
 627         int real_num_busy;
 628         int ret;
 629
 630         /* First try only the requested placement. No eviction. */
 631         real_num_busy = fetch_and_zero(&placement->num_busy_placement);
 632         ret = ttm_bo_validate(bo, placement, &ctx);
 633         if (ret) {
 634                 ret = i915_ttm_err_to_gem(ret);
 635                 /*
 636                  * Anything that wants to restart the operation gets to
 637                  * do that.
 638                  */
 639                 if (ret == -EDEADLK || ret == -EINTR || ret == -ERESTARTSYS ||
 640                     ret == -EAGAIN)
 641                         return ret;
 642
 643                 /*
 644                  * If the initial attempt fails, allow all accepted placements,
 645                  * evicting if necessary.
 646                  */
 647                 placement->num_busy_placement = real_num_busy;
 648                 ret = ttm_bo_validate(bo, placement, &ctx);
 649                 if (ret)
 650                         return i915_ttm_err_to_gem(ret);
 651         }
 652
 653         i915_ttm_adjust_lru(obj);
 654         if (bo->ttm && !ttm_tt_is_populated(bo->ttm)) {
 655                 ret = ttm_tt_populate(bo->bdev, bo->ttm, &ctx);
 656                 if (ret)
 657                         return ret;
 658
 659                 i915_ttm_adjust_domains_after_move(obj);
 660                 i915_ttm_adjust_gem_after_move(obj);
 661         }
 662
 663         if (!i915_gem_object_has_pages(obj)) {
 664                 /* Object either has a page vector or is an iomem object */
 665                 st = bo->ttm ? i915_ttm_tt_get_st(bo->ttm) : obj->ttm.cached_io_st;
 666                 if (IS_ERR(st))
 667                         return PTR_ERR(st);
 668
 669                 __i915_gem_object_set_pages(obj, st, i915_sg_dma_sizes(st->sgl));
 670         }
 671
 672         return ret;
 673 }
 674
 675 static int i915_ttm_get_pages(struct drm_i915_gem_object *obj)
 676 {
 677         struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS];
 678         struct ttm_placement placement;
 679
 680         GEM_BUG_ON(obj->mm.n_placements > I915_TTM_MAX_PLACEMENTS);
 681
 682         /* Move to the requested placement. */
 683         i915_ttm_placement_from_obj(obj, &requested, busy, &placement);
 684
 685         return __i915_ttm_get_pages(obj, &placement);
 686 }
 687
 688 /**
 689  * DOC: Migration vs eviction
 690  *
 691  * GEM migration may not be the same as TTM migration / eviction. If
 692  * the TTM core decides to evict an object it may be evicted to a
 693  * TTM memory type that is not in the object's allowable GEM regions, or
 694  * in fact theoretically to a TTM memory type that doesn't correspond to
 695  * a GEM memory region. In that case the object's GEM region is not
 696  * updated, and the data is migrated back to the GEM region at
 697  * get_pages time. TTM may however set up CPU ptes to the object even
 698  * when it is evicted.
 699  * Gem forced migration using the i915_ttm_migrate() op, is allowed even
 700  * to regions that are not in the object's list of allowable placements.
 701  */
 702 static int i915_ttm_migrate(struct drm_i915_gem_object *obj,
 703                             struct intel_memory_region *mr)
 704 {
 705         struct ttm_place requested;
 706         struct ttm_placement placement;
 707         int ret;
 708
 709         i915_ttm_place_from_region(mr, &requested, obj->flags);
 710         placement.num_placement = 1;
 711         placement.num_busy_placement = 1;
 712         placement.placement = &requested;
 713         placement.busy_placement = &requested;
 714
 715         ret = __i915_ttm_get_pages(obj, &placement);
 716         if (ret)
 717                 return ret;
 718
 719         /*
 720          * Reinitialize the region bindings. This is primarily
 721          * required for objects where the new region is not in
 722          * its allowable placements.
 723          */
 724         if (obj->mm.region != mr) {
 725                 i915_gem_object_release_memory_region(obj);
 726                 i915_gem_object_init_memory_region(obj, mr);
 727         }
 728
 729         return 0;
 730 }
 731
 732 static void i915_ttm_put_pages(struct drm_i915_gem_object *obj,
 733                                struct sg_table *st)
 734 {
 735         /*
 736          * We're currently not called from a shrinker, so put_pages()
 737          * typically means the object is about to destroyed, or called
 738          * from move_notify(). So just avoid doing much for now.
 739          * If the object is not destroyed next, The TTM eviction logic
 740          * and shrinkers will move it out if needed.
 741          */
 742
 743         i915_ttm_adjust_lru(obj);
 744 }
 745
 746 static void i915_ttm_adjust_lru(struct drm_i915_gem_object *obj)
 747 {
 748         struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
 749
 750         /*
 751          * Don't manipulate the TTM LRUs while in TTM bo destruction.
 752          * We're called through i915_ttm_delete_mem_notify().
 753          */
 754         if (!kref_read(&bo->kref))
 755                 return;
 756
 757         /*
 758          * Put on the correct LRU list depending on the MADV status
 759          */
 760         spin_lock(&bo->bdev->lru_lock);
 761         if (obj->mm.madv != I915_MADV_WILLNEED) {
 762                 bo->priority = I915_TTM_PRIO_PURGE;
 763         } else if (!i915_gem_object_has_pages(obj)) {
 764                 if (bo->priority < I915_TTM_PRIO_HAS_PAGES)
 765                         bo->priority = I915_TTM_PRIO_HAS_PAGES;
 766         } else {
 767                 if (bo->priority > I915_TTM_PRIO_NO_PAGES)
 768                         bo->priority = I915_TTM_PRIO_NO_PAGES;
 769         }
 770
 771         ttm_bo_move_to_lru_tail(bo, bo->resource, NULL);
 772         spin_unlock(&bo->bdev->lru_lock);
 773 }
 774
 775 /*
 776  * TTM-backed gem object destruction requires some clarification.
 777  * Basically we have two possibilities here. We can either rely on the
 778  * i915 delayed destruction and put the TTM object when the object
 779  * is idle. This would be detected by TTM which would bypass the
 780  * TTM delayed destroy handling. The other approach is to put the TTM
 781  * object early and rely on the TTM destroyed handling, and then free
 782  * the leftover parts of the GEM object once TTM's destroyed list handling is
 783  * complete. For now, we rely on the latter for two reasons:
 784  * a) TTM can evict an object even when it's on the delayed destroy list,
 785  * which in theory allows for complete eviction.
 786  * b) There is work going on in TTM to allow freeing an object even when
 787  * it's not idle, and using the TTM destroyed list handling could help us
 788  * benefit from that.
 789  */
 790 static void i915_ttm_delayed_free(struct drm_i915_gem_object *obj)
 791 {
 792         if (obj->ttm.created) {
 793                 ttm_bo_put(i915_gem_to_ttm(obj));
 794         } else {
 795                 __i915_gem_free_object(obj);
 796                 call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
 797         }
 798 }
 799
 800 static vm_fault_t vm_fault_ttm(struct vm_fault *vmf)
 801 {
 802         struct vm_area_struct *area = vmf->vma;
 803         struct drm_i915_gem_object *obj =
 804                 i915_ttm_to_gem(area->vm_private_data);
 805
 806         /* Sanity check that we allow writing into this object */
 807         if (unlikely(i915_gem_object_is_readonly(obj) &&
 808                      area->vm_flags & VM_WRITE))
 809                 return VM_FAULT_SIGBUS;
 810
 811         return ttm_bo_vm_fault(vmf);
 812 }
 813
 814 static int
 815 vm_access_ttm(struct vm_area_struct *area, unsigned long addr,
 816               void *buf, int len, int write)
 817 {
 818         struct drm_i915_gem_object *obj =
 819                 i915_ttm_to_gem(area->vm_private_data);
 820
 821         if (i915_gem_object_is_readonly(obj) && write)
 822                 return -EACCES;
 823
 824         return ttm_bo_vm_access(area, addr, buf, len, write);
 825 }
 826
 827 static void ttm_vm_open(struct vm_area_struct *vma)
 828 {
 829         struct drm_i915_gem_object *obj =
 830                 i915_ttm_to_gem(vma->vm_private_data);
 831
 832         GEM_BUG_ON(!obj);
 833         i915_gem_object_get(obj);
 834 }
 835
 836 static void ttm_vm_close(struct vm_area_struct *vma)
 837 {
 838         struct drm_i915_gem_object *obj =
 839                 i915_ttm_to_gem(vma->vm_private_data);
 840
 841         GEM_BUG_ON(!obj);
 842         i915_gem_object_put(obj);
 843 }
 844
 845 static const struct vm_operations_struct vm_ops_ttm = {
 846         .fault = vm_fault_ttm,
 847         .access = vm_access_ttm,
 848         .open = ttm_vm_open,
 849         .close = ttm_vm_close,
 850 };
 851
 852 static u64 i915_ttm_mmap_offset(struct drm_i915_gem_object *obj)
 853 {
 854         /* The ttm_bo must be allocated with I915_BO_ALLOC_USER */
 855         GEM_BUG_ON(!drm_mm_node_allocated(&obj->base.vma_node.vm_node));
 856
 857         return drm_vma_node_offset_addr(&obj->base.vma_node);
 858 }
 859
 860 static const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = {
 861         .name = "i915_gem_object_ttm",
 862
 863         .get_pages = i915_ttm_get_pages,
 864         .put_pages = i915_ttm_put_pages,
 865         .truncate = i915_ttm_purge,
 866         .adjust_lru = i915_ttm_adjust_lru,
 867         .delayed_free = i915_ttm_delayed_free,
 868         .migrate = i915_ttm_migrate,
 869         .mmap_offset = i915_ttm_mmap_offset,
 870         .mmap_ops = &vm_ops_ttm,
 871 };
 872
 873 void i915_ttm_bo_destroy(struct ttm_buffer_object *bo)
 874 {
 875         struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
 876
 877         i915_gem_object_release_memory_region(obj);
 878         mutex_destroy(&obj->ttm.get_io_page.lock);
 879         if (obj->ttm.created)
 880                 call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
 881 }
 882
 883 /**
 884  * __i915_gem_ttm_object_init - Initialize a ttm-backed i915 gem object
 885  * @mem: The initial memory region for the object.
 886  * @obj: The gem object.
 887  * @size: Object size in bytes.
 888  * @flags: gem object flags.
 889  *
 890  * Return: 0 on success, negative error code on failure.
 891  */
 892 int __i915_gem_ttm_object_init(struct intel_memory_region *mem,
 893                                struct drm_i915_gem_object *obj,
 894                                resource_size_t size,
 895                                resource_size_t page_size,
 896                                unsigned int flags)
 897 {
 898         static struct lock_class_key lock_class;
 899         struct drm_i915_private *i915 = mem->i915;
 900         struct ttm_operation_ctx ctx = {
 901                 .interruptible = true,
 902                 .no_wait_gpu = false,
 903         };
 904         enum ttm_bo_type bo_type;
 905         int ret;
 906
 907         drm_gem_private_object_init(&i915->drm, &obj->base, size);
 908         i915_gem_object_init(obj, &i915_gem_ttm_obj_ops, &lock_class, flags);
 909         i915_gem_object_init_memory_region(obj, mem);
 910         i915_gem_object_make_unshrinkable(obj);
 911         INIT_RADIX_TREE(&obj->ttm.get_io_page.radix, GFP_KERNEL | __GFP_NOWARN);
 912         mutex_init(&obj->ttm.get_io_page.lock);
 913         bo_type = (obj->flags & I915_BO_ALLOC_USER) ? ttm_bo_type_device :
 914                 ttm_bo_type_kernel;
 915
 916         obj->base.vma_node.driver_private = i915_gem_to_ttm(obj);
 917
 918         /* Forcing the page size is kernel internal only */
 919         GEM_BUG_ON(page_size && obj->mm.n_placements);
 920
 921         /*
 922          * If this function fails, it will call the destructor, but
 923          * our caller still owns the object. So no freeing in the
 924          * destructor until obj->ttm.created is true.
 925          * Similarly, in delayed_destroy, we can't call ttm_bo_put()
 926          * until successful initialization.
 927          */
 928         ret = ttm_bo_init_reserved(&i915->bdev, i915_gem_to_ttm(obj), size,
 929                                    bo_type, &i915_sys_placement,
 930                                    page_size >> PAGE_SHIFT,
 931                                    &ctx, NULL, NULL, i915_ttm_bo_destroy);
 932         if (ret)
 933                 return i915_ttm_err_to_gem(ret);
 934
 935         obj->ttm.created = true;
 936         i915_ttm_adjust_domains_after_move(obj);
 937         i915_ttm_adjust_gem_after_move(obj);
 938         i915_gem_object_unlock(obj);
 939
 940         return 0;
 941 }
 942
 943 static const struct intel_memory_region_ops ttm_system_region_ops = {
 944         .init_object = __i915_gem_ttm_object_init,
 945 };
 946
 947 struct intel_memory_region *
 948 i915_gem_ttm_system_setup(struct drm_i915_private *i915,
 949                           u16 type, u16 instance)
 950 {
 951         struct intel_memory_region *mr;
 952
 953         mr = intel_memory_region_create(i915, 0,
 954                                         totalram_pages() << PAGE_SHIFT,
 955                                         PAGE_SIZE, 0,
 956                                         type, instance,
 957                                         &ttm_system_region_ops);
 958         if (IS_ERR(mr))
 959                 return mr;
 960
 961         intel_memory_region_set_name(mr, "system-ttm");
 962         return mr;
 963 }