drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c

   1 /*
   2  * Copyright 2016 Advanced Micro Devices, Inc.
   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 shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  * Authors: Christian König
  23  */
  24
  25 #include <linux/dma-mapping.h>
  26 #include <drm/ttm/ttm_range_manager.h>
  27
  28 #include "amdgpu.h"
  29 #include "amdgpu_vm.h"
  30 #include "amdgpu_res_cursor.h"
  31 #include "amdgpu_atomfirmware.h"
  32 #include "atom.h"
  33
  34 struct amdgpu_vram_reservation {
  35         u64 start;
  36         u64 size;
  37         struct list_head allocated;
  38         struct list_head blocks;
  39 };
  40
  41 static inline struct amdgpu_vram_mgr *
  42 to_vram_mgr(struct ttm_resource_manager *man)
  43 {
  44         return container_of(man, struct amdgpu_vram_mgr, manager);
  45 }
  46
  47 static inline struct amdgpu_device *
  48 to_amdgpu_device(struct amdgpu_vram_mgr *mgr)
  49 {
  50         return container_of(mgr, struct amdgpu_device, mman.vram_mgr);
  51 }
  52
  53 static inline struct drm_buddy_block *
  54 amdgpu_vram_mgr_first_block(struct list_head *list)
  55 {
  56         return list_first_entry_or_null(list, struct drm_buddy_block, link);
  57 }
  58
  59 static inline bool amdgpu_is_vram_mgr_blocks_contiguous(struct list_head *head)
  60 {
  61         struct drm_buddy_block *block;
  62         u64 start, size;
  63
  64         block = amdgpu_vram_mgr_first_block(head);
  65         if (!block)
  66                 return false;
  67
  68         while (head != block->link.next) {
  69                 start = amdgpu_vram_mgr_block_start(block);
  70                 size = amdgpu_vram_mgr_block_size(block);
  71
  72                 block = list_entry(block->link.next, struct drm_buddy_block, link);
  73                 if (start + size != amdgpu_vram_mgr_block_start(block))
  74                         return false;
  75         }
  76
  77         return true;
  78 }
  79
  80 static inline u64 amdgpu_vram_mgr_blocks_size(struct list_head *head)
  81 {
  82         struct drm_buddy_block *block;
  83         u64 size = 0;
  84
  85         list_for_each_entry(block, head, link)
  86                 size += amdgpu_vram_mgr_block_size(block);
  87
  88         return size;
  89 }
  90
  91 /**
  92  * DOC: mem_info_vram_total
  93  *
  94  * The amdgpu driver provides a sysfs API for reporting current total VRAM
  95  * available on the device
  96  * The file mem_info_vram_total is used for this and returns the total
  97  * amount of VRAM in bytes
  98  */
  99 static ssize_t amdgpu_mem_info_vram_total_show(struct device *dev,
 100                 struct device_attribute *attr, char *buf)
 101 {
 102         struct drm_device *ddev = dev_get_drvdata(dev);
 103         struct amdgpu_device *adev = drm_to_adev(ddev);
 104
 105         return sysfs_emit(buf, "%llu\n", adev->gmc.real_vram_size);
 106 }
 107
 108 /**
 109  * DOC: mem_info_vis_vram_total
 110  *
 111  * The amdgpu driver provides a sysfs API for reporting current total
 112  * visible VRAM available on the device
 113  * The file mem_info_vis_vram_total is used for this and returns the total
 114  * amount of visible VRAM in bytes
 115  */
 116 static ssize_t amdgpu_mem_info_vis_vram_total_show(struct device *dev,
 117                 struct device_attribute *attr, char *buf)
 118 {
 119         struct drm_device *ddev = dev_get_drvdata(dev);
 120         struct amdgpu_device *adev = drm_to_adev(ddev);
 121
 122         return sysfs_emit(buf, "%llu\n", adev->gmc.visible_vram_size);
 123 }
 124
 125 /**
 126  * DOC: mem_info_vram_used
 127  *
 128  * The amdgpu driver provides a sysfs API for reporting current total VRAM
 129  * available on the device
 130  * The file mem_info_vram_used is used for this and returns the total
 131  * amount of currently used VRAM in bytes
 132  */
 133 static ssize_t amdgpu_mem_info_vram_used_show(struct device *dev,
 134                                               struct device_attribute *attr,
 135                                               char *buf)
 136 {
 137         struct drm_device *ddev = dev_get_drvdata(dev);
 138         struct amdgpu_device *adev = drm_to_adev(ddev);
 139         struct ttm_resource_manager *man = &adev->mman.vram_mgr.manager;
 140
 141         return sysfs_emit(buf, "%llu\n", ttm_resource_manager_usage(man));
 142 }
 143
 144 /**
 145  * DOC: mem_info_vis_vram_used
 146  *
 147  * The amdgpu driver provides a sysfs API for reporting current total of
 148  * used visible VRAM
 149  * The file mem_info_vis_vram_used is used for this and returns the total
 150  * amount of currently used visible VRAM in bytes
 151  */
 152 static ssize_t amdgpu_mem_info_vis_vram_used_show(struct device *dev,
 153                                                   struct device_attribute *attr,
 154                                                   char *buf)
 155 {
 156         struct drm_device *ddev = dev_get_drvdata(dev);
 157         struct amdgpu_device *adev = drm_to_adev(ddev);
 158
 159         return sysfs_emit(buf, "%llu\n",
 160                           amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr));
 161 }
 162
 163 /**
 164  * DOC: mem_info_vram_vendor
 165  *
 166  * The amdgpu driver provides a sysfs API for reporting the vendor of the
 167  * installed VRAM
 168  * The file mem_info_vram_vendor is used for this and returns the name of the
 169  * vendor.
 170  */
 171 static ssize_t amdgpu_mem_info_vram_vendor(struct device *dev,
 172                                            struct device_attribute *attr,
 173                                            char *buf)
 174 {
 175         struct drm_device *ddev = dev_get_drvdata(dev);
 176         struct amdgpu_device *adev = drm_to_adev(ddev);
 177
 178         switch (adev->gmc.vram_vendor) {
 179         case SAMSUNG:
 180                 return sysfs_emit(buf, "samsung\n");
 181         case INFINEON:
 182                 return sysfs_emit(buf, "infineon\n");
 183         case ELPIDA:
 184                 return sysfs_emit(buf, "elpida\n");
 185         case ETRON:
 186                 return sysfs_emit(buf, "etron\n");
 187         case NANYA:
 188                 return sysfs_emit(buf, "nanya\n");
 189         case HYNIX:
 190                 return sysfs_emit(buf, "hynix\n");
 191         case MOSEL:
 192                 return sysfs_emit(buf, "mosel\n");
 193         case WINBOND:
 194                 return sysfs_emit(buf, "winbond\n");
 195         case ESMT:
 196                 return sysfs_emit(buf, "esmt\n");
 197         case MICRON:
 198                 return sysfs_emit(buf, "micron\n");
 199         default:
 200                 return sysfs_emit(buf, "unknown\n");
 201         }
 202 }
 203
 204 static DEVICE_ATTR(mem_info_vram_total, S_IRUGO,
 205                    amdgpu_mem_info_vram_total_show, NULL);
 206 static DEVICE_ATTR(mem_info_vis_vram_total, S_IRUGO,
 207                    amdgpu_mem_info_vis_vram_total_show,NULL);
 208 static DEVICE_ATTR(mem_info_vram_used, S_IRUGO,
 209                    amdgpu_mem_info_vram_used_show, NULL);
 210 static DEVICE_ATTR(mem_info_vis_vram_used, S_IRUGO,
 211                    amdgpu_mem_info_vis_vram_used_show, NULL);
 212 static DEVICE_ATTR(mem_info_vram_vendor, S_IRUGO,
 213                    amdgpu_mem_info_vram_vendor, NULL);
 214
 215 static struct attribute *amdgpu_vram_mgr_attributes[] = {
 216         &dev_attr_mem_info_vram_total.attr,
 217         &dev_attr_mem_info_vis_vram_total.attr,
 218         &dev_attr_mem_info_vram_used.attr,
 219         &dev_attr_mem_info_vis_vram_used.attr,
 220         &dev_attr_mem_info_vram_vendor.attr,
 221         NULL
 222 };
 223
 224 static umode_t amdgpu_vram_attrs_is_visible(struct kobject *kobj,
 225                                             struct attribute *attr, int i)
 226 {
 227         struct device *dev = kobj_to_dev(kobj);
 228         struct drm_device *ddev = dev_get_drvdata(dev);
 229         struct amdgpu_device *adev = drm_to_adev(ddev);
 230
 231         if (attr == &dev_attr_mem_info_vram_vendor.attr &&
 232             !adev->gmc.vram_vendor)
 233                 return 0;
 234
 235         return attr->mode;
 236 }
 237
 238 const struct attribute_group amdgpu_vram_mgr_attr_group = {
 239         .attrs = amdgpu_vram_mgr_attributes,
 240         .is_visible = amdgpu_vram_attrs_is_visible
 241 };
 242
 243 /**
 244  * amdgpu_vram_mgr_vis_size - Calculate visible block size
 245  *
 246  * @adev: amdgpu_device pointer
 247  * @block: DRM BUDDY block structure
 248  *
 249  * Calculate how many bytes of the DRM BUDDY block are inside visible VRAM
 250  */
 251 static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev,
 252                                     struct drm_buddy_block *block)
 253 {
 254         u64 start = amdgpu_vram_mgr_block_start(block);
 255         u64 end = start + amdgpu_vram_mgr_block_size(block);
 256
 257         if (start >= adev->gmc.visible_vram_size)
 258                 return 0;
 259
 260         return (end > adev->gmc.visible_vram_size ?
 261                 adev->gmc.visible_vram_size : end) - start;
 262 }
 263
 264 /**
 265  * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size
 266  *
 267  * @bo: &amdgpu_bo buffer object (must be in VRAM)
 268  *
 269  * Returns:
 270  * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM.
 271  */
 272 u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo)
 273 {
 274         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
 275         struct ttm_resource *res = bo->tbo.resource;
 276         struct amdgpu_vram_mgr_resource *vres = to_amdgpu_vram_mgr_resource(res);
 277         struct drm_buddy_block *block;
 278         u64 usage = 0;
 279
 280         if (amdgpu_gmc_vram_full_visible(&adev->gmc))
 281                 return amdgpu_bo_size(bo);
 282
 283         if (res->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
 284                 return 0;
 285
 286         list_for_each_entry(block, &vres->blocks, link)
 287                 usage += amdgpu_vram_mgr_vis_size(adev, block);
 288
 289         return usage;
 290 }
 291
 292 /* Commit the reservation of VRAM pages */
 293 static void amdgpu_vram_mgr_do_reserve(struct ttm_resource_manager *man)
 294 {
 295         struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
 296         struct amdgpu_device *adev = to_amdgpu_device(mgr);
 297         struct drm_buddy *mm = &mgr->mm;
 298         struct amdgpu_vram_reservation *rsv, *temp;
 299         struct drm_buddy_block *block;
 300         uint64_t vis_usage;
 301
 302         list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, blocks) {
 303                 if (drm_buddy_alloc_blocks(mm, rsv->start, rsv->start + rsv->size,
 304                                            rsv->size, mm->chunk_size, &rsv->allocated,
 305                                            DRM_BUDDY_RANGE_ALLOCATION))
 306                         continue;
 307
 308                 block = amdgpu_vram_mgr_first_block(&rsv->allocated);
 309                 if (!block)
 310                         continue;
 311
 312                 dev_dbg(adev->dev, "Reservation 0x%llx - %lld, Succeeded\n",
 313                         rsv->start, rsv->size);
 314
 315                 vis_usage = amdgpu_vram_mgr_vis_size(adev, block);
 316                 atomic64_add(vis_usage, &mgr->vis_usage);
 317                 spin_lock(&man->bdev->lru_lock);
 318                 man->usage += rsv->size;
 319                 spin_unlock(&man->bdev->lru_lock);
 320                 list_move(&rsv->blocks, &mgr->reserved_pages);
 321         }
 322 }
 323
 324 /**
 325  * amdgpu_vram_mgr_reserve_range - Reserve a range from VRAM
 326  *
 327  * @mgr: amdgpu_vram_mgr pointer
 328  * @start: start address of the range in VRAM
 329  * @size: size of the range
 330  *
 331  * Reserve memory from start address with the specified size in VRAM
 332  */
 333 int amdgpu_vram_mgr_reserve_range(struct amdgpu_vram_mgr *mgr,
 334                                   uint64_t start, uint64_t size)
 335 {
 336         struct amdgpu_vram_reservation *rsv;
 337
 338         rsv = kzalloc(sizeof(*rsv), GFP_KERNEL);
 339         if (!rsv)
 340                 return -ENOMEM;
 341
 342         INIT_LIST_HEAD(&rsv->allocated);
 343         INIT_LIST_HEAD(&rsv->blocks);
 344
 345         rsv->start = start;
 346         rsv->size = size;
 347
 348         mutex_lock(&mgr->lock);
 349         list_add_tail(&rsv->blocks, &mgr->reservations_pending);
 350         amdgpu_vram_mgr_do_reserve(&mgr->manager);
 351         mutex_unlock(&mgr->lock);
 352
 353         return 0;
 354 }
 355
 356 /**
 357  * amdgpu_vram_mgr_query_page_status - query the reservation status
 358  *
 359  * @mgr: amdgpu_vram_mgr pointer
 360  * @start: start address of a page in VRAM
 361  *
 362  * Returns:
 363  *      -EBUSY: the page is still hold and in pending list
 364  *      0: the page has been reserved
 365  *      -ENOENT: the input page is not a reservation
 366  */
 367 int amdgpu_vram_mgr_query_page_status(struct amdgpu_vram_mgr *mgr,
 368                                       uint64_t start)
 369 {
 370         struct amdgpu_vram_reservation *rsv;
 371         int ret;
 372
 373         mutex_lock(&mgr->lock);
 374
 375         list_for_each_entry(rsv, &mgr->reservations_pending, blocks) {
 376                 if (rsv->start <= start &&
 377                     (start < (rsv->start + rsv->size))) {
 378                         ret = -EBUSY;
 379                         goto out;
 380                 }
 381         }
 382
 383         list_for_each_entry(rsv, &mgr->reserved_pages, blocks) {
 384                 if (rsv->start <= start &&
 385                     (start < (rsv->start + rsv->size))) {
 386                         ret = 0;
 387                         goto out;
 388                 }
 389         }
 390
 391         ret = -ENOENT;
 392 out:
 393         mutex_unlock(&mgr->lock);
 394         return ret;
 395 }
 396
 397 static void amdgpu_dummy_vram_mgr_debug(struct ttm_resource_manager *man,
 398                                   struct drm_printer *printer)
 399 {
 400         DRM_DEBUG_DRIVER("Dummy vram mgr debug\n");
 401 }
 402
 403 static bool amdgpu_dummy_vram_mgr_compatible(struct ttm_resource_manager *man,
 404                                        struct ttm_resource *res,
 405                                        const struct ttm_place *place,
 406                                        size_t size)
 407 {
 408         DRM_DEBUG_DRIVER("Dummy vram mgr compatible\n");
 409         return false;
 410 }
 411
 412 static bool amdgpu_dummy_vram_mgr_intersects(struct ttm_resource_manager *man,
 413                                        struct ttm_resource *res,
 414                                        const struct ttm_place *place,
 415                                        size_t size)
 416 {
 417         DRM_DEBUG_DRIVER("Dummy vram mgr intersects\n");
 418         return true;
 419 }
 420
 421 static void amdgpu_dummy_vram_mgr_del(struct ttm_resource_manager *man,
 422                                 struct ttm_resource *res)
 423 {
 424         DRM_DEBUG_DRIVER("Dummy vram mgr deleted\n");
 425 }
 426
 427 static int amdgpu_dummy_vram_mgr_new(struct ttm_resource_manager *man,
 428                                struct ttm_buffer_object *tbo,
 429                                const struct ttm_place *place,
 430                                struct ttm_resource **res)
 431 {
 432         DRM_DEBUG_DRIVER("Dummy vram mgr new\n");
 433         return -ENOSPC;
 434 }
 435
 436 /**
 437  * amdgpu_vram_mgr_new - allocate new ranges
 438  *
 439  * @man: TTM memory type manager
 440  * @tbo: TTM BO we need this range for
 441  * @place: placement flags and restrictions
 442  * @res: the resulting mem object
 443  *
 444  * Allocate VRAM for the given BO.
 445  */
 446 static int amdgpu_vram_mgr_new(struct ttm_resource_manager *man,
 447                                struct ttm_buffer_object *tbo,
 448                                const struct ttm_place *place,
 449                                struct ttm_resource **res)
 450 {
 451         struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
 452         struct amdgpu_device *adev = to_amdgpu_device(mgr);
 453         struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
 454         u64 vis_usage = 0, max_bytes, min_block_size;
 455         struct amdgpu_vram_mgr_resource *vres;
 456         u64 size, remaining_size, lpfn, fpfn;
 457         struct drm_buddy *mm = &mgr->mm;
 458         struct drm_buddy_block *block;
 459         unsigned long pages_per_block;
 460         int r;
 461
 462         lpfn = (u64)place->lpfn << PAGE_SHIFT;
 463         if (!lpfn)
 464                 lpfn = man->size;
 465
 466         fpfn = (u64)place->fpfn << PAGE_SHIFT;
 467
 468         max_bytes = adev->gmc.mc_vram_size;
 469         if (tbo->type != ttm_bo_type_kernel)
 470                 max_bytes -= AMDGPU_VM_RESERVED_VRAM;
 471
 472         if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) {
 473                 pages_per_block = ~0ul;
 474         } else {
 475 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 476                 pages_per_block = HPAGE_PMD_NR;
 477 #else
 478                 /* default to 2MB */
 479                 pages_per_block = 2UL << (20UL - PAGE_SHIFT);
 480 #endif
 481                 pages_per_block = max_t(u32, pages_per_block,
 482                                         tbo->page_alignment);
 483         }
 484
 485         vres = kzalloc(sizeof(*vres), GFP_KERNEL);
 486         if (!vres)
 487                 return -ENOMEM;
 488
 489         ttm_resource_init(tbo, place, &vres->base);
 490
 491         /* bail out quickly if there's likely not enough VRAM for this BO */
 492         if (ttm_resource_manager_usage(man) > max_bytes) {
 493                 r = -ENOSPC;
 494                 goto error_fini;
 495         }
 496
 497         INIT_LIST_HEAD(&vres->blocks);
 498
 499         if (place->flags & TTM_PL_FLAG_TOPDOWN)
 500                 vres->flags |= DRM_BUDDY_TOPDOWN_ALLOCATION;
 501
 502         if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
 503                 vres->flags |= DRM_BUDDY_CONTIGUOUS_ALLOCATION;
 504
 505         if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED)
 506                 vres->flags |= DRM_BUDDY_CLEAR_ALLOCATION;
 507
 508         if (fpfn || lpfn != mgr->mm.size)
 509                 /* Allocate blocks in desired range */
 510                 vres->flags |= DRM_BUDDY_RANGE_ALLOCATION;
 511
 512         remaining_size = (u64)vres->base.size;
 513
 514         mutex_lock(&mgr->lock);
 515         while (remaining_size) {
 516                 if (tbo->page_alignment)
 517                         min_block_size = (u64)tbo->page_alignment << PAGE_SHIFT;
 518                 else
 519                         min_block_size = mgr->default_page_size;
 520
 521                 /* Limit maximum size to 2GiB due to SG table limitations */
 522                 size = min(remaining_size, 2ULL << 30);
 523
 524                 if ((size >= (u64)pages_per_block << PAGE_SHIFT) &&
 525                     !(size & (((u64)pages_per_block << PAGE_SHIFT) - 1)))
 526                         min_block_size = (u64)pages_per_block << PAGE_SHIFT;
 527
 528                 BUG_ON(min_block_size < mm->chunk_size);
 529
 530                 r = drm_buddy_alloc_blocks(mm, fpfn,
 531                                            lpfn,
 532                                            size,
 533                                            min_block_size,
 534                                            &vres->blocks,
 535                                            vres->flags);
 536
 537                 if (unlikely(r == -ENOSPC) && pages_per_block == ~0ul &&
 538                     !(place->flags & TTM_PL_FLAG_CONTIGUOUS)) {
 539                         vres->flags &= ~DRM_BUDDY_CONTIGUOUS_ALLOCATION;
 540                         pages_per_block = max_t(u32, 2UL << (20UL - PAGE_SHIFT),
 541                                                 tbo->page_alignment);
 542
 543                         continue;
 544                 }
 545
 546                 if (unlikely(r))
 547                         goto error_free_blocks;
 548
 549                 if (size > remaining_size)
 550                         remaining_size = 0;
 551                 else
 552                         remaining_size -= size;
 553         }
 554         mutex_unlock(&mgr->lock);
 555
 556         vres->base.start = 0;
 557         size = max_t(u64, amdgpu_vram_mgr_blocks_size(&vres->blocks),
 558                      vres->base.size);
 559         list_for_each_entry(block, &vres->blocks, link) {
 560                 unsigned long start;
 561
 562                 start = amdgpu_vram_mgr_block_start(block) +
 563                         amdgpu_vram_mgr_block_size(block);
 564                 start >>= PAGE_SHIFT;
 565
 566                 if (start > PFN_UP(size))
 567                         start -= PFN_UP(size);
 568                 else
 569                         start = 0;
 570                 vres->base.start = max(vres->base.start, start);
 571
 572                 vis_usage += amdgpu_vram_mgr_vis_size(adev, block);
 573         }
 574
 575         if (amdgpu_is_vram_mgr_blocks_contiguous(&vres->blocks))
 576                 vres->base.placement |= TTM_PL_FLAG_CONTIGUOUS;
 577
 578         if (adev->gmc.xgmi.connected_to_cpu)
 579                 vres->base.bus.caching = ttm_cached;
 580         else
 581                 vres->base.bus.caching = ttm_write_combined;
 582
 583         atomic64_add(vis_usage, &mgr->vis_usage);
 584         *res = &vres->base;
 585         return 0;
 586
 587 error_free_blocks:
 588         drm_buddy_free_list(mm, &vres->blocks, 0);
 589         mutex_unlock(&mgr->lock);
 590 error_fini:
 591         ttm_resource_fini(man, &vres->base);
 592         kfree(vres);
 593
 594         return r;
 595 }
 596
 597 /**
 598  * amdgpu_vram_mgr_del - free ranges
 599  *
 600  * @man: TTM memory type manager
 601  * @res: TTM memory object
 602  *
 603  * Free the allocated VRAM again.
 604  */
 605 static void amdgpu_vram_mgr_del(struct ttm_resource_manager *man,
 606                                 struct ttm_resource *res)
 607 {
 608         struct amdgpu_vram_mgr_resource *vres = to_amdgpu_vram_mgr_resource(res);
 609         struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
 610         struct amdgpu_device *adev = to_amdgpu_device(mgr);
 611         struct drm_buddy *mm = &mgr->mm;
 612         struct drm_buddy_block *block;
 613         uint64_t vis_usage = 0;
 614
 615         mutex_lock(&mgr->lock);
 616         list_for_each_entry(block, &vres->blocks, link)
 617                 vis_usage += amdgpu_vram_mgr_vis_size(adev, block);
 618
 619         amdgpu_vram_mgr_do_reserve(man);
 620
 621         drm_buddy_free_list(mm, &vres->blocks, vres->flags);
 622         mutex_unlock(&mgr->lock);
 623
 624         atomic64_sub(vis_usage, &mgr->vis_usage);
 625
 626         ttm_resource_fini(man, res);
 627         kfree(vres);
 628 }
 629
 630 /**
 631  * amdgpu_vram_mgr_alloc_sgt - allocate and fill a sg table
 632  *
 633  * @adev: amdgpu device pointer
 634  * @res: TTM memory object
 635  * @offset: byte offset from the base of VRAM BO
 636  * @length: number of bytes to export in sg_table
 637  * @dev: the other device
 638  * @dir: dma direction
 639  * @sgt: resulting sg table
 640  *
 641  * Allocate and fill a sg table from a VRAM allocation.
 642  */
 643 int amdgpu_vram_mgr_alloc_sgt(struct amdgpu_device *adev,
 644                               struct ttm_resource *res,
 645                               u64 offset, u64 length,
 646                               struct device *dev,
 647                               enum dma_data_direction dir,
 648                               struct sg_table **sgt)
 649 {
 650         struct amdgpu_res_cursor cursor;
 651         struct scatterlist *sg;
 652         int num_entries = 0;
 653         int i, r;
 654
 655         *sgt = kmalloc(sizeof(**sgt), GFP_KERNEL);
 656         if (!*sgt)
 657                 return -ENOMEM;
 658
 659         /* Determine the number of DRM_BUDDY blocks to export */
 660         amdgpu_res_first(res, offset, length, &cursor);
 661         while (cursor.remaining) {
 662                 num_entries++;
 663                 amdgpu_res_next(&cursor, cursor.size);
 664         }
 665
 666         r = sg_alloc_table(*sgt, num_entries, GFP_KERNEL);
 667         if (r)
 668                 goto error_free;
 669
 670         /* Initialize scatterlist nodes of sg_table */
 671         for_each_sgtable_sg((*sgt), sg, i)
 672                 sg->length = 0;
 673
 674         /*
 675          * Walk down DRM_BUDDY blocks to populate scatterlist nodes
 676          * @note: Use iterator api to get first the DRM_BUDDY block
 677          * and the number of bytes from it. Access the following
 678          * DRM_BUDDY block(s) if more buffer needs to exported
 679          */
 680         amdgpu_res_first(res, offset, length, &cursor);
 681         for_each_sgtable_sg((*sgt), sg, i) {
 682                 phys_addr_t phys = cursor.start + adev->gmc.aper_base;
 683                 size_t size = cursor.size;
 684                 dma_addr_t addr;
 685
 686                 addr = dma_map_resource(dev, phys, size, dir,
 687                                         DMA_ATTR_SKIP_CPU_SYNC);
 688                 r = dma_mapping_error(dev, addr);
 689                 if (r)
 690                         goto error_unmap;
 691
 692                 sg_set_page(sg, NULL, size, 0);
 693                 sg_dma_address(sg) = addr;
 694                 sg_dma_len(sg) = size;
 695
 696                 amdgpu_res_next(&cursor, cursor.size);
 697         }
 698
 699         return 0;
 700
 701 error_unmap:
 702         for_each_sgtable_sg((*sgt), sg, i) {
 703                 if (!sg->length)
 704                         continue;
 705
 706                 dma_unmap_resource(dev, sg->dma_address,
 707                                    sg->length, dir,
 708                                    DMA_ATTR_SKIP_CPU_SYNC);
 709         }
 710         sg_free_table(*sgt);
 711
 712 error_free:
 713         kfree(*sgt);
 714         return r;
 715 }
 716
 717 /**
 718  * amdgpu_vram_mgr_free_sgt - allocate and fill a sg table
 719  *
 720  * @dev: device pointer
 721  * @dir: data direction of resource to unmap
 722  * @sgt: sg table to free
 723  *
 724  * Free a previously allocate sg table.
 725  */
 726 void amdgpu_vram_mgr_free_sgt(struct device *dev,
 727                               enum dma_data_direction dir,
 728                               struct sg_table *sgt)
 729 {
 730         struct scatterlist *sg;
 731         int i;
 732
 733         for_each_sgtable_sg(sgt, sg, i)
 734                 dma_unmap_resource(dev, sg->dma_address,
 735                                    sg->length, dir,
 736                                    DMA_ATTR_SKIP_CPU_SYNC);
 737         sg_free_table(sgt);
 738         kfree(sgt);
 739 }
 740
 741 /**
 742  * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part
 743  *
 744  * @mgr: amdgpu_vram_mgr pointer
 745  *
 746  * Returns how many bytes are used in the visible part of VRAM
 747  */
 748 uint64_t amdgpu_vram_mgr_vis_usage(struct amdgpu_vram_mgr *mgr)
 749 {
 750         return atomic64_read(&mgr->vis_usage);
 751 }
 752
 753 /**
 754  * amdgpu_vram_mgr_intersects - test each drm buddy block for intersection
 755  *
 756  * @man: TTM memory type manager
 757  * @res: The resource to test
 758  * @place: The place to test against
 759  * @size: Size of the new allocation
 760  *
 761  * Test each drm buddy block for intersection for eviction decision.
 762  */
 763 static bool amdgpu_vram_mgr_intersects(struct ttm_resource_manager *man,
 764                                        struct ttm_resource *res,
 765                                        const struct ttm_place *place,
 766                                        size_t size)
 767 {
 768         struct amdgpu_vram_mgr_resource *mgr = to_amdgpu_vram_mgr_resource(res);
 769         struct drm_buddy_block *block;
 770
 771         /* Check each drm buddy block individually */
 772         list_for_each_entry(block, &mgr->blocks, link) {
 773                 unsigned long fpfn =
 774                         amdgpu_vram_mgr_block_start(block) >> PAGE_SHIFT;
 775                 unsigned long lpfn = fpfn +
 776                         (amdgpu_vram_mgr_block_size(block) >> PAGE_SHIFT);
 777
 778                 if (place->fpfn < lpfn &&
 779                     (!place->lpfn || place->lpfn > fpfn))
 780                         return true;
 781         }
 782
 783         return false;
 784 }
 785
 786 /**
 787  * amdgpu_vram_mgr_compatible - test each drm buddy block for compatibility
 788  *
 789  * @man: TTM memory type manager
 790  * @res: The resource to test
 791  * @place: The place to test against
 792  * @size: Size of the new allocation
 793  *
 794  * Test each drm buddy block for placement compatibility.
 795  */
 796 static bool amdgpu_vram_mgr_compatible(struct ttm_resource_manager *man,
 797                                        struct ttm_resource *res,
 798                                        const struct ttm_place *place,
 799                                        size_t size)
 800 {
 801         struct amdgpu_vram_mgr_resource *mgr = to_amdgpu_vram_mgr_resource(res);
 802         struct drm_buddy_block *block;
 803
 804         /* Check each drm buddy block individually */
 805         list_for_each_entry(block, &mgr->blocks, link) {
 806                 unsigned long fpfn =
 807                         amdgpu_vram_mgr_block_start(block) >> PAGE_SHIFT;
 808                 unsigned long lpfn = fpfn +
 809                         (amdgpu_vram_mgr_block_size(block) >> PAGE_SHIFT);
 810
 811                 if (fpfn < place->fpfn ||
 812                     (place->lpfn && lpfn > place->lpfn))
 813                         return false;
 814         }
 815
 816         return true;
 817 }
 818
 819 /**
 820  * amdgpu_vram_mgr_debug - dump VRAM table
 821  *
 822  * @man: TTM memory type manager
 823  * @printer: DRM printer to use
 824  *
 825  * Dump the table content using printk.
 826  */
 827 static void amdgpu_vram_mgr_debug(struct ttm_resource_manager *man,
 828                                   struct drm_printer *printer)
 829 {
 830         struct amdgpu_vram_mgr *mgr = to_vram_mgr(man);
 831         struct drm_buddy *mm = &mgr->mm;
 832         struct amdgpu_vram_reservation *rsv;
 833
 834         drm_printf(printer, "  vis usage:%llu\n",
 835                    amdgpu_vram_mgr_vis_usage(mgr));
 836
 837         mutex_lock(&mgr->lock);
 838         drm_printf(printer, "default_page_size: %lluKiB\n",
 839                    mgr->default_page_size >> 10);
 840
 841         drm_buddy_print(mm, printer);
 842
 843         drm_printf(printer, "reserved:\n");
 844         list_for_each_entry(rsv, &mgr->reserved_pages, blocks)
 845                 drm_printf(printer, "%#018llx-%#018llx: %llu\n",
 846                         rsv->start, rsv->start + rsv->size, rsv->size);
 847         mutex_unlock(&mgr->lock);
 848 }
 849
 850 static const struct ttm_resource_manager_func amdgpu_dummy_vram_mgr_func = {
 851         .alloc  = amdgpu_dummy_vram_mgr_new,
 852         .free   = amdgpu_dummy_vram_mgr_del,
 853         .intersects = amdgpu_dummy_vram_mgr_intersects,
 854         .compatible = amdgpu_dummy_vram_mgr_compatible,
 855         .debug  = amdgpu_dummy_vram_mgr_debug
 856 };
 857
 858 static const struct ttm_resource_manager_func amdgpu_vram_mgr_func = {
 859         .alloc  = amdgpu_vram_mgr_new,
 860         .free   = amdgpu_vram_mgr_del,
 861         .intersects = amdgpu_vram_mgr_intersects,
 862         .compatible = amdgpu_vram_mgr_compatible,
 863         .debug  = amdgpu_vram_mgr_debug
 864 };
 865
 866 /**
 867  * amdgpu_vram_mgr_init - init VRAM manager and DRM MM
 868  *
 869  * @adev: amdgpu_device pointer
 870  *
 871  * Allocate and initialize the VRAM manager.
 872  */
 873 int amdgpu_vram_mgr_init(struct amdgpu_device *adev)
 874 {
 875         struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
 876         struct ttm_resource_manager *man = &mgr->manager;
 877         int err;
 878
 879         ttm_resource_manager_init(man, &adev->mman.bdev,
 880                                   adev->gmc.real_vram_size);
 881
 882         mutex_init(&mgr->lock);
 883         INIT_LIST_HEAD(&mgr->reservations_pending);
 884         INIT_LIST_HEAD(&mgr->reserved_pages);
 885         mgr->default_page_size = PAGE_SIZE;
 886
 887         if (!adev->gmc.is_app_apu) {
 888                 man->func = &amdgpu_vram_mgr_func;
 889
 890                 err = drm_buddy_init(&mgr->mm, man->size, PAGE_SIZE);
 891                 if (err)
 892                         return err;
 893         } else {
 894                 man->func = &amdgpu_dummy_vram_mgr_func;
 895                 DRM_INFO("Setup dummy vram mgr\n");
 896         }
 897
 898         ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, &mgr->manager);
 899         ttm_resource_manager_set_used(man, true);
 900         return 0;
 901 }
 902
 903 /**
 904  * amdgpu_vram_mgr_fini - free and destroy VRAM manager
 905  *
 906  * @adev: amdgpu_device pointer
 907  *
 908  * Destroy and free the VRAM manager, returns -EBUSY if ranges are still
 909  * allocated inside it.
 910  */
 911 void amdgpu_vram_mgr_fini(struct amdgpu_device *adev)
 912 {
 913         struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
 914         struct ttm_resource_manager *man = &mgr->manager;
 915         int ret;
 916         struct amdgpu_vram_reservation *rsv, *temp;
 917
 918         ttm_resource_manager_set_used(man, false);
 919
 920         ret = ttm_resource_manager_evict_all(&adev->mman.bdev, man);
 921         if (ret)
 922                 return;
 923
 924         mutex_lock(&mgr->lock);
 925         list_for_each_entry_safe(rsv, temp, &mgr->reservations_pending, blocks)
 926                 kfree(rsv);
 927
 928         list_for_each_entry_safe(rsv, temp, &mgr->reserved_pages, blocks) {
 929                 drm_buddy_free_list(&mgr->mm, &rsv->allocated, 0);
 930                 kfree(rsv);
 931         }
 932         if (!adev->gmc.is_app_apu)
 933                 drm_buddy_fini(&mgr->mm);
 934         mutex_unlock(&mgr->lock);
 935
 936         ttm_resource_manager_cleanup(man);
 937         ttm_set_driver_manager(&adev->mman.bdev, TTM_PL_VRAM, NULL);
 938 }