drivers/gpu/drm/nouveau/nouveau_bo.c

   1 /*
   2  * Copyright 2007 Dave Airlied
   3  * All Rights Reserved.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * VA LINUX SYSTEMS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22  * OTHER DEALINGS IN THE SOFTWARE.
  23  */
  24 /*
  25  * Authors: Dave Airlied <airlied@linux.ie>
  26  *          Ben Skeggs   <darktama@iinet.net.au>
  27  *          Jeremy Kolb  <jkolb@brandeis.edu>
  28  */
  29
  30 #include <linux/dma-mapping.h>
  31 #include <drm/ttm/ttm_tt.h>
  32
  33 #include "nouveau_drv.h"
  34 #include "nouveau_chan.h"
  35 #include "nouveau_fence.h"
  36
  37 #include "nouveau_bo.h"
  38 #include "nouveau_ttm.h"
  39 #include "nouveau_gem.h"
  40 #include "nouveau_mem.h"
  41 #include "nouveau_vmm.h"
  42
  43 #include <nvif/class.h>
  44 #include <nvif/if500b.h>
  45 #include <nvif/if900b.h>
  46
  47 static int nouveau_ttm_tt_bind(struct ttm_device *bdev, struct ttm_tt *ttm,
  48                                struct ttm_resource *reg);
  49 static void nouveau_ttm_tt_unbind(struct ttm_device *bdev, struct ttm_tt *ttm);
  50
  51 /*
  52  * NV10-NV40 tiling helpers
  53  */
  54
  55 static void
  56 nv10_bo_update_tile_region(struct drm_device *dev, struct nouveau_drm_tile *reg,
  57                            u32 addr, u32 size, u32 pitch, u32 flags)
  58 {
  59         struct nouveau_drm *drm = nouveau_drm(dev);
  60         int i = reg - drm->tile.reg;
  61         struct nvkm_fb *fb = nvxx_fb(&drm->client.device);
  62         struct nvkm_fb_tile *tile = &fb->tile.region[i];
  63
  64         nouveau_fence_unref(&reg->fence);
  65
  66         if (tile->pitch)
  67                 nvkm_fb_tile_fini(fb, i, tile);
  68
  69         if (pitch)
  70                 nvkm_fb_tile_init(fb, i, addr, size, pitch, flags, tile);
  71
  72         nvkm_fb_tile_prog(fb, i, tile);
  73 }
  74
  75 static struct nouveau_drm_tile *
  76 nv10_bo_get_tile_region(struct drm_device *dev, int i)
  77 {
  78         struct nouveau_drm *drm = nouveau_drm(dev);
  79         struct nouveau_drm_tile *tile = &drm->tile.reg[i];
  80
  81         spin_lock(&drm->tile.lock);
  82
  83         if (!tile->used &&
  84             (!tile->fence || nouveau_fence_done(tile->fence)))
  85                 tile->used = true;
  86         else
  87                 tile = NULL;
  88
  89         spin_unlock(&drm->tile.lock);
  90         return tile;
  91 }
  92
  93 static void
  94 nv10_bo_put_tile_region(struct drm_device *dev, struct nouveau_drm_tile *tile,
  95                         struct dma_fence *fence)
  96 {
  97         struct nouveau_drm *drm = nouveau_drm(dev);
  98
  99         if (tile) {
 100                 spin_lock(&drm->tile.lock);
 101                 tile->fence = (struct nouveau_fence *)dma_fence_get(fence);
 102                 tile->used = false;
 103                 spin_unlock(&drm->tile.lock);
 104         }
 105 }
 106
 107 static struct nouveau_drm_tile *
 108 nv10_bo_set_tiling(struct drm_device *dev, u32 addr,
 109                    u32 size, u32 pitch, u32 zeta)
 110 {
 111         struct nouveau_drm *drm = nouveau_drm(dev);
 112         struct nvkm_fb *fb = nvxx_fb(&drm->client.device);
 113         struct nouveau_drm_tile *tile, *found = NULL;
 114         int i;
 115
 116         for (i = 0; i < fb->tile.regions; i++) {
 117                 tile = nv10_bo_get_tile_region(dev, i);
 118
 119                 if (pitch && !found) {
 120                         found = tile;
 121                         continue;
 122
 123                 } else if (tile && fb->tile.region[i].pitch) {
 124                         /* Kill an unused tile region. */
 125                         nv10_bo_update_tile_region(dev, tile, 0, 0, 0, 0);
 126                 }
 127
 128                 nv10_bo_put_tile_region(dev, tile, NULL);
 129         }
 130
 131         if (found)
 132                 nv10_bo_update_tile_region(dev, found, addr, size, pitch, zeta);
 133         return found;
 134 }
 135
 136 static void
 137 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
 138 {
 139         struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
 140         struct drm_device *dev = drm->dev;
 141         struct nouveau_bo *nvbo = nouveau_bo(bo);
 142
 143         WARN_ON(nvbo->bo.pin_count > 0);
 144         nouveau_bo_del_io_reserve_lru(bo);
 145         nv10_bo_put_tile_region(dev, nvbo->tile, NULL);
 146
 147         /*
 148          * If nouveau_bo_new() allocated this buffer, the GEM object was never
 149          * initialized, so don't attempt to release it.
 150          */
 151         if (bo->base.dev)
 152                 drm_gem_object_release(&bo->base);
 153         else
 154                 dma_resv_fini(&bo->base._resv);
 155
 156         kfree(nvbo);
 157 }
 158
 159 static inline u64
 160 roundup_64(u64 x, u32 y)
 161 {
 162         x += y - 1;
 163         do_div(x, y);
 164         return x * y;
 165 }
 166
 167 static void
 168 nouveau_bo_fixup_align(struct nouveau_bo *nvbo, int *align, u64 *size)
 169 {
 170         struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 171         struct nvif_device *device = &drm->client.device;
 172
 173         if (device->info.family < NV_DEVICE_INFO_V0_TESLA) {
 174                 if (nvbo->mode) {
 175                         if (device->info.chipset >= 0x40) {
 176                                 *align = 65536;
 177                                 *size = roundup_64(*size, 64 * nvbo->mode);
 178
 179                         } else if (device->info.chipset >= 0x30) {
 180                                 *align = 32768;
 181                                 *size = roundup_64(*size, 64 * nvbo->mode);
 182
 183                         } else if (device->info.chipset >= 0x20) {
 184                                 *align = 16384;
 185                                 *size = roundup_64(*size, 64 * nvbo->mode);
 186
 187                         } else if (device->info.chipset >= 0x10) {
 188                                 *align = 16384;
 189                                 *size = roundup_64(*size, 32 * nvbo->mode);
 190                         }
 191                 }
 192         } else {
 193                 *size = roundup_64(*size, (1 << nvbo->page));
 194                 *align = max((1 <<  nvbo->page), *align);
 195         }
 196
 197         *size = roundup_64(*size, PAGE_SIZE);
 198 }
 199
 200 struct nouveau_bo *
 201 nouveau_bo_alloc(struct nouveau_cli *cli, u64 *size, int *align, u32 domain,
 202                  u32 tile_mode, u32 tile_flags)
 203 {
 204         struct nouveau_drm *drm = cli->drm;
 205         struct nouveau_bo *nvbo;
 206         struct nvif_mmu *mmu = &cli->mmu;
 207         struct nvif_vmm *vmm = cli->svm.cli ? &cli->svm.vmm : &cli->vmm.vmm;
 208         int i, pi = -1;
 209
 210         if (!*size) {
 211                 NV_WARN(drm, "skipped size %016llx\n", *size);
 212                 return ERR_PTR(-EINVAL);
 213         }
 214
 215         nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
 216         if (!nvbo)
 217                 return ERR_PTR(-ENOMEM);
 218         INIT_LIST_HEAD(&nvbo->head);
 219         INIT_LIST_HEAD(&nvbo->entry);
 220         INIT_LIST_HEAD(&nvbo->vma_list);
 221         nvbo->bo.bdev = &drm->ttm.bdev;
 222
 223         /* This is confusing, and doesn't actually mean we want an uncached
 224          * mapping, but is what NOUVEAU_GEM_DOMAIN_COHERENT gets translated
 225          * into in nouveau_gem_new().
 226          */
 227         if (domain & NOUVEAU_GEM_DOMAIN_COHERENT) {
 228                 /* Determine if we can get a cache-coherent map, forcing
 229                  * uncached mapping if we can't.
 230                  */
 231                 if (!nouveau_drm_use_coherent_gpu_mapping(drm))
 232                         nvbo->force_coherent = true;
 233         }
 234
 235         if (cli->device.info.family >= NV_DEVICE_INFO_V0_FERMI) {
 236                 nvbo->kind = (tile_flags & 0x0000ff00) >> 8;
 237                 if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
 238                         kfree(nvbo);
 239                         return ERR_PTR(-EINVAL);
 240                 }
 241
 242                 nvbo->comp = mmu->kind[nvbo->kind] != nvbo->kind;
 243         } else
 244         if (cli->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
 245                 nvbo->kind = (tile_flags & 0x00007f00) >> 8;
 246                 nvbo->comp = (tile_flags & 0x00030000) >> 16;
 247                 if (!nvif_mmu_kind_valid(mmu, nvbo->kind)) {
 248                         kfree(nvbo);
 249                         return ERR_PTR(-EINVAL);
 250                 }
 251         } else {
 252                 nvbo->zeta = (tile_flags & 0x00000007);
 253         }
 254         nvbo->mode = tile_mode;
 255         nvbo->contig = !(tile_flags & NOUVEAU_GEM_TILE_NONCONTIG);
 256
 257         /* Determine the desirable target GPU page size for the buffer. */
 258         for (i = 0; i < vmm->page_nr; i++) {
 259                 /* Because we cannot currently allow VMM maps to fail
 260                  * during buffer migration, we need to determine page
 261                  * size for the buffer up-front, and pre-allocate its
 262                  * page tables.
 263                  *
 264                  * Skip page sizes that can't support needed domains.
 265                  */
 266                 if (cli->device.info.family > NV_DEVICE_INFO_V0_CURIE &&
 267                     (domain & NOUVEAU_GEM_DOMAIN_VRAM) && !vmm->page[i].vram)
 268                         continue;
 269                 if ((domain & NOUVEAU_GEM_DOMAIN_GART) &&
 270                     (!vmm->page[i].host || vmm->page[i].shift > PAGE_SHIFT))
 271                         continue;
 272
 273                 /* Select this page size if it's the first that supports
 274                  * the potential memory domains, or when it's compatible
 275                  * with the requested compression settings.
 276                  */
 277                 if (pi < 0 || !nvbo->comp || vmm->page[i].comp)
 278                         pi = i;
 279
 280                 /* Stop once the buffer is larger than the current page size. */
 281                 if (*size >= 1ULL << vmm->page[i].shift)
 282                         break;
 283         }
 284
 285         if (WARN_ON(pi < 0)) {
 286                 kfree(nvbo);
 287                 return ERR_PTR(-EINVAL);
 288         }
 289
 290         /* Disable compression if suitable settings couldn't be found. */
 291         if (nvbo->comp && !vmm->page[pi].comp) {
 292                 if (mmu->object.oclass >= NVIF_CLASS_MMU_GF100)
 293                         nvbo->kind = mmu->kind[nvbo->kind];
 294                 nvbo->comp = 0;
 295         }
 296         nvbo->page = vmm->page[pi].shift;
 297
 298         nouveau_bo_fixup_align(nvbo, align, size);
 299
 300         return nvbo;
 301 }
 302
 303 int
 304 nouveau_bo_init(struct nouveau_bo *nvbo, u64 size, int align, u32 domain,
 305                 struct sg_table *sg, struct dma_resv *robj)
 306 {
 307         int type = sg ? ttm_bo_type_sg : ttm_bo_type_device;
 308         int ret;
 309
 310         nouveau_bo_placement_set(nvbo, domain, 0);
 311         INIT_LIST_HEAD(&nvbo->io_reserve_lru);
 312
 313         ret = ttm_bo_init_validate(nvbo->bo.bdev, &nvbo->bo, type,
 314                                    &nvbo->placement, align >> PAGE_SHIFT, false,
 315                                    sg, robj, nouveau_bo_del_ttm);
 316         if (ret) {
 317                 /* ttm will call nouveau_bo_del_ttm if it fails.. */
 318                 return ret;
 319         }
 320
 321         return 0;
 322 }
 323
 324 int
 325 nouveau_bo_new(struct nouveau_cli *cli, u64 size, int align,
 326                uint32_t domain, uint32_t tile_mode, uint32_t tile_flags,
 327                struct sg_table *sg, struct dma_resv *robj,
 328                struct nouveau_bo **pnvbo)
 329 {
 330         struct nouveau_bo *nvbo;
 331         int ret;
 332
 333         nvbo = nouveau_bo_alloc(cli, &size, &align, domain, tile_mode,
 334                                 tile_flags);
 335         if (IS_ERR(nvbo))
 336                 return PTR_ERR(nvbo);
 337
 338         nvbo->bo.base.size = size;
 339         dma_resv_init(&nvbo->bo.base._resv);
 340         drm_vma_node_reset(&nvbo->bo.base.vma_node);
 341
 342         ret = nouveau_bo_init(nvbo, size, align, domain, sg, robj);
 343         if (ret)
 344                 return ret;
 345
 346         *pnvbo = nvbo;
 347         return 0;
 348 }
 349
 350 static void
 351 set_placement_list(struct ttm_place *pl, unsigned *n, uint32_t domain)
 352 {
 353         *n = 0;
 354
 355         if (domain & NOUVEAU_GEM_DOMAIN_VRAM) {
 356                 pl[*n].mem_type = TTM_PL_VRAM;
 357                 pl[*n].flags = 0;
 358                 (*n)++;
 359         }
 360         if (domain & NOUVEAU_GEM_DOMAIN_GART) {
 361                 pl[*n].mem_type = TTM_PL_TT;
 362                 pl[*n].flags = 0;
 363                 (*n)++;
 364         }
 365         if (domain & NOUVEAU_GEM_DOMAIN_CPU) {
 366                 pl[*n].mem_type = TTM_PL_SYSTEM;
 367                 pl[(*n)++].flags = 0;
 368         }
 369 }
 370
 371 static void
 372 set_placement_range(struct nouveau_bo *nvbo, uint32_t domain)
 373 {
 374         struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 375         u64 vram_size = drm->client.device.info.ram_size;
 376         unsigned i, fpfn, lpfn;
 377
 378         if (drm->client.device.info.family == NV_DEVICE_INFO_V0_CELSIUS &&
 379             nvbo->mode && (domain & NOUVEAU_GEM_DOMAIN_VRAM) &&
 380             nvbo->bo.base.size < vram_size / 4) {
 381                 /*
 382                  * Make sure that the color and depth buffers are handled
 383                  * by independent memory controller units. Up to a 9x
 384                  * speed up when alpha-blending and depth-test are enabled
 385                  * at the same time.
 386                  */
 387                 if (nvbo->zeta) {
 388                         fpfn = (vram_size / 2) >> PAGE_SHIFT;
 389                         lpfn = ~0;
 390                 } else {
 391                         fpfn = 0;
 392                         lpfn = (vram_size / 2) >> PAGE_SHIFT;
 393                 }
 394                 for (i = 0; i < nvbo->placement.num_placement; ++i) {
 395                         nvbo->placements[i].fpfn = fpfn;
 396                         nvbo->placements[i].lpfn = lpfn;
 397                 }
 398                 for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
 399                         nvbo->busy_placements[i].fpfn = fpfn;
 400                         nvbo->busy_placements[i].lpfn = lpfn;
 401                 }
 402         }
 403 }
 404
 405 void
 406 nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t domain,
 407                          uint32_t busy)
 408 {
 409         struct ttm_placement *pl = &nvbo->placement;
 410
 411         pl->placement = nvbo->placements;
 412         set_placement_list(nvbo->placements, &pl->num_placement, domain);
 413
 414         pl->busy_placement = nvbo->busy_placements;
 415         set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
 416                            domain | busy);
 417
 418         set_placement_range(nvbo, domain);
 419 }
 420
 421 int
 422 nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t domain, bool contig)
 423 {
 424         struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 425         struct ttm_buffer_object *bo = &nvbo->bo;
 426         bool force = false, evict = false;
 427         int ret;
 428
 429         ret = ttm_bo_reserve(bo, false, false, NULL);
 430         if (ret)
 431                 return ret;
 432
 433         if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
 434             domain == NOUVEAU_GEM_DOMAIN_VRAM && contig) {
 435                 if (!nvbo->contig) {
 436                         nvbo->contig = true;
 437                         force = true;
 438                         evict = true;
 439                 }
 440         }
 441
 442         if (nvbo->bo.pin_count) {
 443                 bool error = evict;
 444
 445                 switch (bo->resource->mem_type) {
 446                 case TTM_PL_VRAM:
 447                         error |= !(domain & NOUVEAU_GEM_DOMAIN_VRAM);
 448                         break;
 449                 case TTM_PL_TT:
 450                         error |= !(domain & NOUVEAU_GEM_DOMAIN_GART);
 451                         break;
 452                 default:
 453                         break;
 454                 }
 455
 456                 if (error) {
 457                         NV_ERROR(drm, "bo %p pinned elsewhere: "
 458                                       "0x%08x vs 0x%08x\n", bo,
 459                                  bo->resource->mem_type, domain);
 460                         ret = -EBUSY;
 461                 }
 462                 ttm_bo_pin(&nvbo->bo);
 463                 goto out;
 464         }
 465
 466         if (evict) {
 467                 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART, 0);
 468                 ret = nouveau_bo_validate(nvbo, false, false);
 469                 if (ret)
 470                         goto out;
 471         }
 472
 473         nouveau_bo_placement_set(nvbo, domain, 0);
 474         ret = nouveau_bo_validate(nvbo, false, false);
 475         if (ret)
 476                 goto out;
 477
 478         ttm_bo_pin(&nvbo->bo);
 479
 480         switch (bo->resource->mem_type) {
 481         case TTM_PL_VRAM:
 482                 drm->gem.vram_available -= bo->base.size;
 483                 break;
 484         case TTM_PL_TT:
 485                 drm->gem.gart_available -= bo->base.size;
 486                 break;
 487         default:
 488                 break;
 489         }
 490
 491 out:
 492         if (force && ret)
 493                 nvbo->contig = false;
 494         ttm_bo_unreserve(bo);
 495         return ret;
 496 }
 497
 498 int
 499 nouveau_bo_unpin(struct nouveau_bo *nvbo)
 500 {
 501         struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 502         struct ttm_buffer_object *bo = &nvbo->bo;
 503         int ret;
 504
 505         ret = ttm_bo_reserve(bo, false, false, NULL);
 506         if (ret)
 507                 return ret;
 508
 509         ttm_bo_unpin(&nvbo->bo);
 510         if (!nvbo->bo.pin_count) {
 511                 switch (bo->resource->mem_type) {
 512                 case TTM_PL_VRAM:
 513                         drm->gem.vram_available += bo->base.size;
 514                         break;
 515                 case TTM_PL_TT:
 516                         drm->gem.gart_available += bo->base.size;
 517                         break;
 518                 default:
 519                         break;
 520                 }
 521         }
 522
 523         ttm_bo_unreserve(bo);
 524         return 0;
 525 }
 526
 527 int
 528 nouveau_bo_map(struct nouveau_bo *nvbo)
 529 {
 530         int ret;
 531
 532         ret = ttm_bo_reserve(&nvbo->bo, false, false, NULL);
 533         if (ret)
 534                 return ret;
 535
 536         ret = ttm_bo_kmap(&nvbo->bo, 0, PFN_UP(nvbo->bo.base.size), &nvbo->kmap);
 537
 538         ttm_bo_unreserve(&nvbo->bo);
 539         return ret;
 540 }
 541
 542 void
 543 nouveau_bo_unmap(struct nouveau_bo *nvbo)
 544 {
 545         if (!nvbo)
 546                 return;
 547
 548         ttm_bo_kunmap(&nvbo->kmap);
 549 }
 550
 551 void
 552 nouveau_bo_sync_for_device(struct nouveau_bo *nvbo)
 553 {
 554         struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 555         struct ttm_tt *ttm_dma = (struct ttm_tt *)nvbo->bo.ttm;
 556         int i, j;
 557
 558         if (!ttm_dma || !ttm_dma->dma_address)
 559                 return;
 560         if (!ttm_dma->pages) {
 561                 NV_DEBUG(drm, "ttm_dma 0x%p: pages NULL\n", ttm_dma);
 562                 return;
 563         }
 564
 565         /* Don't waste time looping if the object is coherent */
 566         if (nvbo->force_coherent)
 567                 return;
 568
 569         i = 0;
 570         while (i < ttm_dma->num_pages) {
 571                 struct page *p = ttm_dma->pages[i];
 572                 size_t num_pages = 1;
 573
 574                 for (j = i + 1; j < ttm_dma->num_pages; ++j) {
 575                         if (++p != ttm_dma->pages[j])
 576                                 break;
 577
 578                         ++num_pages;
 579                 }
 580                 dma_sync_single_for_device(drm->dev->dev,
 581                                            ttm_dma->dma_address[i],
 582                                            num_pages * PAGE_SIZE, DMA_TO_DEVICE);
 583                 i += num_pages;
 584         }
 585 }
 586
 587 void
 588 nouveau_bo_sync_for_cpu(struct nouveau_bo *nvbo)
 589 {
 590         struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 591         struct ttm_tt *ttm_dma = (struct ttm_tt *)nvbo->bo.ttm;
 592         int i, j;
 593
 594         if (!ttm_dma || !ttm_dma->dma_address)
 595                 return;
 596         if (!ttm_dma->pages) {
 597                 NV_DEBUG(drm, "ttm_dma 0x%p: pages NULL\n", ttm_dma);
 598                 return;
 599         }
 600
 601         /* Don't waste time looping if the object is coherent */
 602         if (nvbo->force_coherent)
 603                 return;
 604
 605         i = 0;
 606         while (i < ttm_dma->num_pages) {
 607                 struct page *p = ttm_dma->pages[i];
 608                 size_t num_pages = 1;
 609
 610                 for (j = i + 1; j < ttm_dma->num_pages; ++j) {
 611                         if (++p != ttm_dma->pages[j])
 612                                 break;
 613
 614                         ++num_pages;
 615                 }
 616
 617                 dma_sync_single_for_cpu(drm->dev->dev, ttm_dma->dma_address[i],
 618                                         num_pages * PAGE_SIZE, DMA_FROM_DEVICE);
 619                 i += num_pages;
 620         }
 621 }
 622
 623 void nouveau_bo_add_io_reserve_lru(struct ttm_buffer_object *bo)
 624 {
 625         struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
 626         struct nouveau_bo *nvbo = nouveau_bo(bo);
 627
 628         mutex_lock(&drm->ttm.io_reserve_mutex);
 629         list_move_tail(&nvbo->io_reserve_lru, &drm->ttm.io_reserve_lru);
 630         mutex_unlock(&drm->ttm.io_reserve_mutex);
 631 }
 632
 633 void nouveau_bo_del_io_reserve_lru(struct ttm_buffer_object *bo)
 634 {
 635         struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
 636         struct nouveau_bo *nvbo = nouveau_bo(bo);
 637
 638         mutex_lock(&drm->ttm.io_reserve_mutex);
 639         list_del_init(&nvbo->io_reserve_lru);
 640         mutex_unlock(&drm->ttm.io_reserve_mutex);
 641 }
 642
 643 int
 644 nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
 645                     bool no_wait_gpu)
 646 {
 647         struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
 648         int ret;
 649
 650         ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement, &ctx);
 651         if (ret)
 652                 return ret;
 653
 654         nouveau_bo_sync_for_device(nvbo);
 655
 656         return 0;
 657 }
 658
 659 void
 660 nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
 661 {
 662         bool is_iomem;
 663         u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
 664
 665         mem += index;
 666
 667         if (is_iomem)
 668                 iowrite16_native(val, (void __force __iomem *)mem);
 669         else
 670                 *mem = val;
 671 }
 672
 673 u32
 674 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
 675 {
 676         bool is_iomem;
 677         u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
 678
 679         mem += index;
 680
 681         if (is_iomem)
 682                 return ioread32_native((void __force __iomem *)mem);
 683         else
 684                 return *mem;
 685 }
 686
 687 void
 688 nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
 689 {
 690         bool is_iomem;
 691         u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
 692
 693         mem += index;
 694
 695         if (is_iomem)
 696                 iowrite32_native(val, (void __force __iomem *)mem);
 697         else
 698                 *mem = val;
 699 }
 700
 701 static struct ttm_tt *
 702 nouveau_ttm_tt_create(struct ttm_buffer_object *bo, uint32_t page_flags)
 703 {
 704 #if IS_ENABLED(CONFIG_AGP)
 705         struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
 706
 707         if (drm->agp.bridge) {
 708                 return ttm_agp_tt_create(bo, drm->agp.bridge, page_flags);
 709         }
 710 #endif
 711
 712         return nouveau_sgdma_create_ttm(bo, page_flags);
 713 }
 714
 715 static int
 716 nouveau_ttm_tt_bind(struct ttm_device *bdev, struct ttm_tt *ttm,
 717                     struct ttm_resource *reg)
 718 {
 719 #if IS_ENABLED(CONFIG_AGP)
 720         struct nouveau_drm *drm = nouveau_bdev(bdev);
 721 #endif
 722         if (!reg)
 723                 return -EINVAL;
 724 #if IS_ENABLED(CONFIG_AGP)
 725         if (drm->agp.bridge)
 726                 return ttm_agp_bind(ttm, reg);
 727 #endif
 728         return nouveau_sgdma_bind(bdev, ttm, reg);
 729 }
 730
 731 static void
 732 nouveau_ttm_tt_unbind(struct ttm_device *bdev, struct ttm_tt *ttm)
 733 {
 734 #if IS_ENABLED(CONFIG_AGP)
 735         struct nouveau_drm *drm = nouveau_bdev(bdev);
 736
 737         if (drm->agp.bridge) {
 738                 ttm_agp_unbind(ttm);
 739                 return;
 740         }
 741 #endif
 742         nouveau_sgdma_unbind(bdev, ttm);
 743 }
 744
 745 static void
 746 nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
 747 {
 748         struct nouveau_bo *nvbo = nouveau_bo(bo);
 749
 750         switch (bo->resource->mem_type) {
 751         case TTM_PL_VRAM:
 752                 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART,
 753                                          NOUVEAU_GEM_DOMAIN_CPU);
 754                 break;
 755         default:
 756                 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_CPU, 0);
 757                 break;
 758         }
 759
 760         *pl = nvbo->placement;
 761 }
 762
 763 static int
 764 nouveau_bo_move_prep(struct nouveau_drm *drm, struct ttm_buffer_object *bo,
 765                      struct ttm_resource *reg)
 766 {
 767         struct nouveau_mem *old_mem = nouveau_mem(bo->resource);
 768         struct nouveau_mem *new_mem = nouveau_mem(reg);
 769         struct nvif_vmm *vmm = &drm->client.vmm.vmm;
 770         int ret;
 771
 772         ret = nvif_vmm_get(vmm, LAZY, false, old_mem->mem.page, 0,
 773                            old_mem->mem.size, &old_mem->vma[0]);
 774         if (ret)
 775                 return ret;
 776
 777         ret = nvif_vmm_get(vmm, LAZY, false, new_mem->mem.page, 0,
 778                            new_mem->mem.size, &old_mem->vma[1]);
 779         if (ret)
 780                 goto done;
 781
 782         ret = nouveau_mem_map(old_mem, vmm, &old_mem->vma[0]);
 783         if (ret)
 784                 goto done;
 785
 786         ret = nouveau_mem_map(new_mem, vmm, &old_mem->vma[1]);
 787 done:
 788         if (ret) {
 789                 nvif_vmm_put(vmm, &old_mem->vma[1]);
 790                 nvif_vmm_put(vmm, &old_mem->vma[0]);
 791         }
 792         return 0;
 793 }
 794
 795 static int
 796 nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict,
 797                      struct ttm_operation_ctx *ctx,
 798                      struct ttm_resource *new_reg)
 799 {
 800         struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
 801         struct nouveau_channel *chan = drm->ttm.chan;
 802         struct nouveau_cli *cli = (void *)chan->user.client;
 803         struct nouveau_fence *fence;
 804         int ret;
 805
 806         /* create temporary vmas for the transfer and attach them to the
 807          * old nvkm_mem node, these will get cleaned up after ttm has
 808          * destroyed the ttm_resource
 809          */
 810         if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
 811                 ret = nouveau_bo_move_prep(drm, bo, new_reg);
 812                 if (ret)
 813                         return ret;
 814         }
 815
 816         if (drm_drv_uses_atomic_modeset(drm->dev))
 817                 mutex_lock(&cli->mutex);
 818         else
 819                 mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
 820         ret = nouveau_fence_sync(nouveau_bo(bo), chan, true, ctx->interruptible);
 821         if (ret == 0) {
 822                 ret = drm->ttm.move(chan, bo, bo->resource, new_reg);
 823                 if (ret == 0) {
 824                         ret = nouveau_fence_new(chan, false, &fence);
 825                         if (ret == 0) {
 826                                 /* TODO: figure out a better solution here
 827                                  *
 828                                  * wait on the fence here explicitly as going through
 829                                  * ttm_bo_move_accel_cleanup somehow doesn't seem to do it.
 830                                  *
 831                                  * Without this the operation can timeout and we'll fallback to a
 832                                  * software copy, which might take several minutes to finish.
 833                                  */
 834                                 nouveau_fence_wait(fence, false, false);
 835                                 ret = ttm_bo_move_accel_cleanup(bo,
 836                                                                 &fence->base,
 837                                                                 evict, false,
 838                                                                 new_reg);
 839                                 nouveau_fence_unref(&fence);
 840                         }
 841                 }
 842         }
 843         mutex_unlock(&cli->mutex);
 844         return ret;
 845 }
 846
 847 void
 848 nouveau_bo_move_init(struct nouveau_drm *drm)
 849 {
 850         static const struct _method_table {
 851                 const char *name;
 852                 int engine;
 853                 s32 oclass;
 854                 int (*exec)(struct nouveau_channel *,
 855                             struct ttm_buffer_object *,
 856                             struct ttm_resource *, struct ttm_resource *);
 857                 int (*init)(struct nouveau_channel *, u32 handle);
 858         } _methods[] = {
 859                 {  "COPY", 4, 0xc7b5, nve0_bo_move_copy, nve0_bo_move_init },
 860                 {  "GRCE", 0, 0xc7b5, nve0_bo_move_copy, nvc0_bo_move_init },
 861                 {  "COPY", 4, 0xc6b5, nve0_bo_move_copy, nve0_bo_move_init },
 862                 {  "GRCE", 0, 0xc6b5, nve0_bo_move_copy, nvc0_bo_move_init },
 863                 {  "COPY", 4, 0xc5b5, nve0_bo_move_copy, nve0_bo_move_init },
 864                 {  "GRCE", 0, 0xc5b5, nve0_bo_move_copy, nvc0_bo_move_init },
 865                 {  "COPY", 4, 0xc3b5, nve0_bo_move_copy, nve0_bo_move_init },
 866                 {  "GRCE", 0, 0xc3b5, nve0_bo_move_copy, nvc0_bo_move_init },
 867                 {  "COPY", 4, 0xc1b5, nve0_bo_move_copy, nve0_bo_move_init },
 868                 {  "GRCE", 0, 0xc1b5, nve0_bo_move_copy, nvc0_bo_move_init },
 869                 {  "COPY", 4, 0xc0b5, nve0_bo_move_copy, nve0_bo_move_init },
 870                 {  "GRCE", 0, 0xc0b5, nve0_bo_move_copy, nvc0_bo_move_init },
 871                 {  "COPY", 4, 0xb0b5, nve0_bo_move_copy, nve0_bo_move_init },
 872                 {  "GRCE", 0, 0xb0b5, nve0_bo_move_copy, nvc0_bo_move_init },
 873                 {  "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
 874                 {  "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
 875                 { "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
 876                 { "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
 877                 {  "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
 878                 { "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
 879                 {  "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
 880                 {  "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
 881                 {  "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
 882                 {},
 883         };
 884         const struct _method_table *mthd = _methods;
 885         const char *name = "CPU";
 886         int ret;
 887
 888         do {
 889                 struct nouveau_channel *chan;
 890
 891                 if (mthd->engine)
 892                         chan = drm->cechan;
 893                 else
 894                         chan = drm->channel;
 895                 if (chan == NULL)
 896                         continue;
 897
 898                 ret = nvif_object_ctor(&chan->user, "ttmBoMove",
 899                                        mthd->oclass | (mthd->engine << 16),
 900                                        mthd->oclass, NULL, 0,
 901                                        &drm->ttm.copy);
 902                 if (ret == 0) {
 903                         ret = mthd->init(chan, drm->ttm.copy.handle);
 904                         if (ret) {
 905                                 nvif_object_dtor(&drm->ttm.copy);
 906                                 continue;
 907                         }
 908
 909                         drm->ttm.move = mthd->exec;
 910                         drm->ttm.chan = chan;
 911                         name = mthd->name;
 912                         break;
 913                 }
 914         } while ((++mthd)->exec);
 915
 916         NV_INFO(drm, "MM: using %s for buffer copies\n", name);
 917 }
 918
 919 static void nouveau_bo_move_ntfy(struct ttm_buffer_object *bo,
 920                                  struct ttm_resource *new_reg)
 921 {
 922         struct nouveau_mem *mem = new_reg ? nouveau_mem(new_reg) : NULL;
 923         struct nouveau_bo *nvbo = nouveau_bo(bo);
 924         struct nouveau_vma *vma;
 925         long ret;
 926
 927         /* ttm can now (stupidly) pass the driver bos it didn't create... */
 928         if (bo->destroy != nouveau_bo_del_ttm)
 929                 return;
 930
 931         nouveau_bo_del_io_reserve_lru(bo);
 932
 933         if (mem && new_reg->mem_type != TTM_PL_SYSTEM &&
 934             mem->mem.page == nvbo->page) {
 935                 list_for_each_entry(vma, &nvbo->vma_list, head) {
 936                         nouveau_vma_map(vma, mem);
 937                 }
 938         } else {
 939                 list_for_each_entry(vma, &nvbo->vma_list, head) {
 940                         ret = dma_resv_wait_timeout(bo->base.resv,
 941                                                     DMA_RESV_USAGE_BOOKKEEP,
 942                                                     false, 15 * HZ);
 943                         WARN_ON(ret <= 0);
 944                         nouveau_vma_unmap(vma);
 945                 }
 946         }
 947
 948         if (new_reg)
 949                 nvbo->offset = (new_reg->start << PAGE_SHIFT);
 950
 951 }
 952
 953 static int
 954 nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_resource *new_reg,
 955                    struct nouveau_drm_tile **new_tile)
 956 {
 957         struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
 958         struct drm_device *dev = drm->dev;
 959         struct nouveau_bo *nvbo = nouveau_bo(bo);
 960         u64 offset = new_reg->start << PAGE_SHIFT;
 961
 962         *new_tile = NULL;
 963         if (new_reg->mem_type != TTM_PL_VRAM)
 964                 return 0;
 965
 966         if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_CELSIUS) {
 967                 *new_tile = nv10_bo_set_tiling(dev, offset, bo->base.size,
 968                                                nvbo->mode, nvbo->zeta);
 969         }
 970
 971         return 0;
 972 }
 973
 974 static void
 975 nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
 976                       struct nouveau_drm_tile *new_tile,
 977                       struct nouveau_drm_tile **old_tile)
 978 {
 979         struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
 980         struct drm_device *dev = drm->dev;
 981         struct dma_fence *fence;
 982         int ret;
 983
 984         ret = dma_resv_get_singleton(bo->base.resv, DMA_RESV_USAGE_WRITE,
 985                                      &fence);
 986         if (ret)
 987                 dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_WRITE,
 988                                       false, MAX_SCHEDULE_TIMEOUT);
 989
 990         nv10_bo_put_tile_region(dev, *old_tile, fence);
 991         *old_tile = new_tile;
 992 }
 993
 994 static int
 995 nouveau_bo_move(struct ttm_buffer_object *bo, bool evict,
 996                 struct ttm_operation_ctx *ctx,
 997                 struct ttm_resource *new_reg,
 998                 struct ttm_place *hop)
 999 {
1000         struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1001         struct nouveau_bo *nvbo = nouveau_bo(bo);
1002         struct ttm_resource *old_reg = bo->resource;
1003         struct nouveau_drm_tile *new_tile = NULL;
1004         int ret = 0;
1005
1006
1007         if (new_reg->mem_type == TTM_PL_TT) {
1008                 ret = nouveau_ttm_tt_bind(bo->bdev, bo->ttm, new_reg);
1009                 if (ret)
1010                         return ret;
1011         }
1012
1013         nouveau_bo_move_ntfy(bo, new_reg);
1014         ret = ttm_bo_wait_ctx(bo, ctx);
1015         if (ret)
1016                 goto out_ntfy;
1017
1018         if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1019                 ret = nouveau_bo_vm_bind(bo, new_reg, &new_tile);
1020                 if (ret)
1021                         goto out_ntfy;
1022         }
1023
1024         /* Fake bo copy. */
1025         if (!old_reg || (old_reg->mem_type == TTM_PL_SYSTEM &&
1026                          !bo->ttm)) {
1027                 ttm_bo_move_null(bo, new_reg);
1028                 goto out;
1029         }
1030
1031         if (old_reg->mem_type == TTM_PL_SYSTEM &&
1032             new_reg->mem_type == TTM_PL_TT) {
1033                 ttm_bo_move_null(bo, new_reg);
1034                 goto out;
1035         }
1036
1037         if (old_reg->mem_type == TTM_PL_TT &&
1038             new_reg->mem_type == TTM_PL_SYSTEM) {
1039                 nouveau_ttm_tt_unbind(bo->bdev, bo->ttm);
1040                 ttm_resource_free(bo, &bo->resource);
1041                 ttm_bo_assign_mem(bo, new_reg);
1042                 goto out;
1043         }
1044
1045         /* Hardware assisted copy. */
1046         if (drm->ttm.move) {
1047                 if ((old_reg->mem_type == TTM_PL_SYSTEM &&
1048                      new_reg->mem_type == TTM_PL_VRAM) ||
1049                     (old_reg->mem_type == TTM_PL_VRAM &&
1050                      new_reg->mem_type == TTM_PL_SYSTEM)) {
1051                         hop->fpfn = 0;
1052                         hop->lpfn = 0;
1053                         hop->mem_type = TTM_PL_TT;
1054                         hop->flags = 0;
1055                         return -EMULTIHOP;
1056                 }
1057                 ret = nouveau_bo_move_m2mf(bo, evict, ctx,
1058                                            new_reg);
1059         } else
1060                 ret = -ENODEV;
1061
1062         if (ret) {
1063                 /* Fallback to software copy. */
1064                 ret = ttm_bo_move_memcpy(bo, ctx, new_reg);
1065         }
1066
1067 out:
1068         if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1069                 if (ret)
1070                         nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1071                 else
1072                         nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1073         }
1074 out_ntfy:
1075         if (ret) {
1076                 nouveau_bo_move_ntfy(bo, bo->resource);
1077         }
1078         return ret;
1079 }
1080
1081 static void
1082 nouveau_ttm_io_mem_free_locked(struct nouveau_drm *drm,
1083                                struct ttm_resource *reg)
1084 {
1085         struct nouveau_mem *mem = nouveau_mem(reg);
1086
1087         if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1088                 switch (reg->mem_type) {
1089                 case TTM_PL_TT:
1090                         if (mem->kind)
1091                                 nvif_object_unmap_handle(&mem->mem.object);
1092                         break;
1093                 case TTM_PL_VRAM:
1094                         nvif_object_unmap_handle(&mem->mem.object);
1095                         break;
1096                 default:
1097                         break;
1098                 }
1099         }
1100 }
1101
1102 static int
1103 nouveau_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *reg)
1104 {
1105         struct nouveau_drm *drm = nouveau_bdev(bdev);
1106         struct nvkm_device *device = nvxx_device(&drm->client.device);
1107         struct nouveau_mem *mem = nouveau_mem(reg);
1108         struct nvif_mmu *mmu = &drm->client.mmu;
1109         int ret;
1110
1111         mutex_lock(&drm->ttm.io_reserve_mutex);
1112 retry:
1113         switch (reg->mem_type) {
1114         case TTM_PL_SYSTEM:
1115                 /* System memory */
1116                 ret = 0;
1117                 goto out;
1118         case TTM_PL_TT:
1119 #if IS_ENABLED(CONFIG_AGP)
1120                 if (drm->agp.bridge) {
1121                         reg->bus.offset = (reg->start << PAGE_SHIFT) +
1122                                 drm->agp.base;
1123                         reg->bus.is_iomem = !drm->agp.cma;
1124                         reg->bus.caching = ttm_write_combined;
1125                 }
1126 #endif
1127                 if (drm->client.mem->oclass < NVIF_CLASS_MEM_NV50 ||
1128                     !mem->kind) {
1129                         /* untiled */
1130                         ret = 0;
1131                         break;
1132                 }
1133                 fallthrough;    /* tiled memory */
1134         case TTM_PL_VRAM:
1135                 reg->bus.offset = (reg->start << PAGE_SHIFT) +
1136                         device->func->resource_addr(device, 1);
1137                 reg->bus.is_iomem = true;
1138
1139                 /* Some BARs do not support being ioremapped WC */
1140                 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
1141                     mmu->type[drm->ttm.type_vram].type & NVIF_MEM_UNCACHED)
1142                         reg->bus.caching = ttm_uncached;
1143                 else
1144                         reg->bus.caching = ttm_write_combined;
1145
1146                 if (drm->client.mem->oclass >= NVIF_CLASS_MEM_NV50) {
1147                         union {
1148                                 struct nv50_mem_map_v0 nv50;
1149                                 struct gf100_mem_map_v0 gf100;
1150                         } args;
1151                         u64 handle, length;
1152                         u32 argc = 0;
1153
1154                         switch (mem->mem.object.oclass) {
1155                         case NVIF_CLASS_MEM_NV50:
1156                                 args.nv50.version = 0;
1157                                 args.nv50.ro = 0;
1158                                 args.nv50.kind = mem->kind;
1159                                 args.nv50.comp = mem->comp;
1160                                 argc = sizeof(args.nv50);
1161                                 break;
1162                         case NVIF_CLASS_MEM_GF100:
1163                                 args.gf100.version = 0;
1164                                 args.gf100.ro = 0;
1165                                 args.gf100.kind = mem->kind;
1166                                 argc = sizeof(args.gf100);
1167                                 break;
1168                         default:
1169                                 WARN_ON(1);
1170                                 break;
1171                         }
1172
1173                         ret = nvif_object_map_handle(&mem->mem.object,
1174                                                      &args, argc,
1175                                                      &handle, &length);
1176                         if (ret != 1) {
1177                                 if (WARN_ON(ret == 0))
1178                                         ret = -EINVAL;
1179                                 goto out;
1180                         }
1181
1182                         reg->bus.offset = handle;
1183                 }
1184                 ret = 0;
1185                 break;
1186         default:
1187                 ret = -EINVAL;
1188         }
1189
1190 out:
1191         if (ret == -ENOSPC) {
1192                 struct nouveau_bo *nvbo;
1193
1194                 nvbo = list_first_entry_or_null(&drm->ttm.io_reserve_lru,
1195                                                 typeof(*nvbo),
1196                                                 io_reserve_lru);
1197                 if (nvbo) {
1198                         list_del_init(&nvbo->io_reserve_lru);
1199                         drm_vma_node_unmap(&nvbo->bo.base.vma_node,
1200                                            bdev->dev_mapping);
1201                         nouveau_ttm_io_mem_free_locked(drm, nvbo->bo.resource);
1202                         goto retry;
1203                 }
1204
1205         }
1206         mutex_unlock(&drm->ttm.io_reserve_mutex);
1207         return ret;
1208 }
1209
1210 static void
1211 nouveau_ttm_io_mem_free(struct ttm_device *bdev, struct ttm_resource *reg)
1212 {
1213         struct nouveau_drm *drm = nouveau_bdev(bdev);
1214
1215         mutex_lock(&drm->ttm.io_reserve_mutex);
1216         nouveau_ttm_io_mem_free_locked(drm, reg);
1217         mutex_unlock(&drm->ttm.io_reserve_mutex);
1218 }
1219
1220 vm_fault_t nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1221 {
1222         struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1223         struct nouveau_bo *nvbo = nouveau_bo(bo);
1224         struct nvkm_device *device = nvxx_device(&drm->client.device);
1225         u32 mappable = device->func->resource_size(device, 1) >> PAGE_SHIFT;
1226         int i, ret;
1227
1228         /* as long as the bo isn't in vram, and isn't tiled, we've got
1229          * nothing to do here.
1230          */
1231         if (bo->resource->mem_type != TTM_PL_VRAM) {
1232                 if (drm->client.device.info.family < NV_DEVICE_INFO_V0_TESLA ||
1233                     !nvbo->kind)
1234                         return 0;
1235
1236                 if (bo->resource->mem_type != TTM_PL_SYSTEM)
1237                         return 0;
1238
1239                 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_GART, 0);
1240
1241         } else {
1242                 /* make sure bo is in mappable vram */
1243                 if (drm->client.device.info.family >= NV_DEVICE_INFO_V0_TESLA ||
1244                     bo->resource->start + PFN_UP(bo->resource->size) < mappable)
1245                         return 0;
1246
1247                 for (i = 0; i < nvbo->placement.num_placement; ++i) {
1248                         nvbo->placements[i].fpfn = 0;
1249                         nvbo->placements[i].lpfn = mappable;
1250                 }
1251
1252                 for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
1253                         nvbo->busy_placements[i].fpfn = 0;
1254                         nvbo->busy_placements[i].lpfn = mappable;
1255                 }
1256
1257                 nouveau_bo_placement_set(nvbo, NOUVEAU_GEM_DOMAIN_VRAM, 0);
1258         }
1259
1260         ret = nouveau_bo_validate(nvbo, false, false);
1261         if (unlikely(ret == -EBUSY || ret == -ERESTARTSYS))
1262                 return VM_FAULT_NOPAGE;
1263         else if (unlikely(ret))
1264                 return VM_FAULT_SIGBUS;
1265
1266         ttm_bo_move_to_lru_tail_unlocked(bo);
1267         return 0;
1268 }
1269
1270 static int
1271 nouveau_ttm_tt_populate(struct ttm_device *bdev,
1272                         struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
1273 {
1274         struct ttm_tt *ttm_dma = (void *)ttm;
1275         struct nouveau_drm *drm;
1276         bool slave = !!(ttm->page_flags & TTM_TT_FLAG_EXTERNAL);
1277
1278         if (ttm_tt_is_populated(ttm))
1279                 return 0;
1280
1281         if (slave && ttm->sg) {
1282                 drm_prime_sg_to_dma_addr_array(ttm->sg, ttm_dma->dma_address,
1283                                                ttm->num_pages);
1284                 return 0;
1285         }
1286
1287         drm = nouveau_bdev(bdev);
1288
1289         return ttm_pool_alloc(&drm->ttm.bdev.pool, ttm, ctx);
1290 }
1291
1292 static void
1293 nouveau_ttm_tt_unpopulate(struct ttm_device *bdev,
1294                           struct ttm_tt *ttm)
1295 {
1296         struct nouveau_drm *drm;
1297         bool slave = !!(ttm->page_flags & TTM_TT_FLAG_EXTERNAL);
1298
1299         if (slave)
1300                 return;
1301
1302         nouveau_ttm_tt_unbind(bdev, ttm);
1303
1304         drm = nouveau_bdev(bdev);
1305
1306         return ttm_pool_free(&drm->ttm.bdev.pool, ttm);
1307 }
1308
1309 static void
1310 nouveau_ttm_tt_destroy(struct ttm_device *bdev,
1311                        struct ttm_tt *ttm)
1312 {
1313 #if IS_ENABLED(CONFIG_AGP)
1314         struct nouveau_drm *drm = nouveau_bdev(bdev);
1315         if (drm->agp.bridge) {
1316                 ttm_agp_destroy(ttm);
1317                 return;
1318         }
1319 #endif
1320         nouveau_sgdma_destroy(bdev, ttm);
1321 }
1322
1323 void
1324 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence, bool exclusive)
1325 {
1326         struct dma_resv *resv = nvbo->bo.base.resv;
1327
1328         if (!fence)
1329                 return;
1330
1331         dma_resv_add_fence(resv, &fence->base, exclusive ?
1332                            DMA_RESV_USAGE_WRITE : DMA_RESV_USAGE_READ);
1333 }
1334
1335 static void
1336 nouveau_bo_delete_mem_notify(struct ttm_buffer_object *bo)
1337 {
1338         nouveau_bo_move_ntfy(bo, NULL);
1339 }
1340
1341 struct ttm_device_funcs nouveau_bo_driver = {
1342         .ttm_tt_create = &nouveau_ttm_tt_create,
1343         .ttm_tt_populate = &nouveau_ttm_tt_populate,
1344         .ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1345         .ttm_tt_destroy = &nouveau_ttm_tt_destroy,
1346         .eviction_valuable = ttm_bo_eviction_valuable,
1347         .evict_flags = nouveau_bo_evict_flags,
1348         .delete_mem_notify = nouveau_bo_delete_mem_notify,
1349         .move = nouveau_bo_move,
1350         .io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1351         .io_mem_free = &nouveau_ttm_io_mem_free,
1352 };