drivers/gpu/drm/radeon/radeon_ttm.c

   1 /*
   2  * Copyright 2009 Jerome Glisse.
   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
   7  * "Software"), to deal in the Software without restriction, including
   8  * without limitation the rights to use, copy, modify, merge, publish,
   9  * distribute, sub license, and/or sell copies of the Software, and to
  10  * permit persons to whom the Software is furnished to do so, subject to
  11  * the following conditions:
  12  *
  13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  20  *
  21  * The above copyright notice and this permission notice (including the
  22  * next paragraph) shall be included in all copies or substantial portions
  23  * of the Software.
  24  *
  25  */
  26 /*
  27  * Authors:
  28  *    Jerome Glisse <glisse@freedesktop.org>
  29  *    Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
  30  *    Dave Airlie
  31  */
  32
  33 #include <linux/dma-mapping.h>
  34 #include <linux/pagemap.h>
  35 #include <linux/pci.h>
  36 #include <linux/seq_file.h>
  37 #include <linux/slab.h>
  38 #include <linux/swap.h>
  39 #include <linux/swiotlb.h>
  40
  41 #include <drm/drm_agpsupport.h>
  42 #include <drm/drm_debugfs.h>
  43 #include <drm/drm_device.h>
  44 #include <drm/drm_file.h>
  45 #include <drm/drm_prime.h>
  46 #include <drm/radeon_drm.h>
  47 #include <drm/ttm/ttm_bo_api.h>
  48 #include <drm/ttm/ttm_bo_driver.h>
  49 #include <drm/ttm/ttm_module.h>
  50 #include <drm/ttm/ttm_page_alloc.h>
  51 #include <drm/ttm/ttm_placement.h>
  52
  53 #include "radeon_reg.h"
  54 #include "radeon.h"
  55
  56 static int radeon_ttm_debugfs_init(struct radeon_device *rdev);
  57 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev);
  58
  59 struct radeon_device *radeon_get_rdev(struct ttm_bo_device *bdev)
  60 {
  61         struct radeon_mman *mman;
  62         struct radeon_device *rdev;
  63
  64         mman = container_of(bdev, struct radeon_mman, bdev);
  65         rdev = container_of(mman, struct radeon_device, mman);
  66         return rdev;
  67 }
  68
  69 static int radeon_ttm_init_vram(struct radeon_device *rdev)
  70 {
  71         return ttm_range_man_init(&rdev->mman.bdev, TTM_PL_VRAM,
  72                                   TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC,
  73                                   TTM_PL_FLAG_WC, false,
  74                                   rdev->mc.real_vram_size >> PAGE_SHIFT);
  75 }
  76
  77 static int radeon_ttm_init_gtt(struct radeon_device *rdev)
  78 {
  79         return ttm_range_man_init(&rdev->mman.bdev, TTM_PL_TT,
  80                                   TTM_PL_MASK_CACHING,
  81                                   TTM_PL_FLAG_CACHED, true,
  82                                   rdev->mc.gtt_size >> PAGE_SHIFT);
  83 }
  84
  85 static void radeon_evict_flags(struct ttm_buffer_object *bo,
  86                                 struct ttm_placement *placement)
  87 {
  88         static const struct ttm_place placements = {
  89                 .fpfn = 0,
  90                 .lpfn = 0,
  91                 .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
  92         };
  93
  94         struct radeon_bo *rbo;
  95
  96         if (!radeon_ttm_bo_is_radeon_bo(bo)) {
  97                 placement->placement = &placements;
  98                 placement->busy_placement = &placements;
  99                 placement->num_placement = 1;
 100                 placement->num_busy_placement = 1;
 101                 return;
 102         }
 103         rbo = container_of(bo, struct radeon_bo, tbo);
 104         switch (bo->mem.mem_type) {
 105         case TTM_PL_VRAM:
 106                 if (rbo->rdev->ring[radeon_copy_ring_index(rbo->rdev)].ready == false)
 107                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
 108                 else if (rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size &&
 109                          bo->mem.start < (rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT)) {
 110                         unsigned fpfn = rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
 111                         int i;
 112
 113                         /* Try evicting to the CPU inaccessible part of VRAM
 114                          * first, but only set GTT as busy placement, so this
 115                          * BO will be evicted to GTT rather than causing other
 116                          * BOs to be evicted from VRAM
 117                          */
 118                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM |
 119                                                          RADEON_GEM_DOMAIN_GTT);
 120                         rbo->placement.num_busy_placement = 0;
 121                         for (i = 0; i < rbo->placement.num_placement; i++) {
 122                                 if (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) {
 123                                         if (rbo->placements[i].fpfn < fpfn)
 124                                                 rbo->placements[i].fpfn = fpfn;
 125                                 } else {
 126                                         rbo->placement.busy_placement =
 127                                                 &rbo->placements[i];
 128                                         rbo->placement.num_busy_placement = 1;
 129                                 }
 130                         }
 131                 } else
 132                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
 133                 break;
 134         case TTM_PL_TT:
 135         default:
 136                 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
 137         }
 138         *placement = rbo->placement;
 139 }
 140
 141 static int radeon_verify_access(struct ttm_buffer_object *bo, struct file *filp)
 142 {
 143         struct radeon_bo *rbo = container_of(bo, struct radeon_bo, tbo);
 144
 145         if (radeon_ttm_tt_has_userptr(bo->ttm))
 146                 return -EPERM;
 147         return drm_vma_node_verify_access(&rbo->tbo.base.vma_node,
 148                                           filp->private_data);
 149 }
 150
 151 static int radeon_move_blit(struct ttm_buffer_object *bo,
 152                         bool evict, bool no_wait_gpu,
 153                         struct ttm_resource *new_mem,
 154                         struct ttm_resource *old_mem)
 155 {
 156         struct radeon_device *rdev;
 157         uint64_t old_start, new_start;
 158         struct radeon_fence *fence;
 159         unsigned num_pages;
 160         int r, ridx;
 161
 162         rdev = radeon_get_rdev(bo->bdev);
 163         ridx = radeon_copy_ring_index(rdev);
 164         old_start = (u64)old_mem->start << PAGE_SHIFT;
 165         new_start = (u64)new_mem->start << PAGE_SHIFT;
 166
 167         switch (old_mem->mem_type) {
 168         case TTM_PL_VRAM:
 169                 old_start += rdev->mc.vram_start;
 170                 break;
 171         case TTM_PL_TT:
 172                 old_start += rdev->mc.gtt_start;
 173                 break;
 174         default:
 175                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 176                 return -EINVAL;
 177         }
 178         switch (new_mem->mem_type) {
 179         case TTM_PL_VRAM:
 180                 new_start += rdev->mc.vram_start;
 181                 break;
 182         case TTM_PL_TT:
 183                 new_start += rdev->mc.gtt_start;
 184                 break;
 185         default:
 186                 DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
 187                 return -EINVAL;
 188         }
 189         if (!rdev->ring[ridx].ready) {
 190                 DRM_ERROR("Trying to move memory with ring turned off.\n");
 191                 return -EINVAL;
 192         }
 193
 194         BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
 195
 196         num_pages = new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
 197         fence = radeon_copy(rdev, old_start, new_start, num_pages, bo->base.resv);
 198         if (IS_ERR(fence))
 199                 return PTR_ERR(fence);
 200
 201         r = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, new_mem);
 202         radeon_fence_unref(&fence);
 203         return r;
 204 }
 205
 206 static int radeon_move_vram_ram(struct ttm_buffer_object *bo,
 207                                 bool evict, bool interruptible,
 208                                 bool no_wait_gpu,
 209                                 struct ttm_resource *new_mem)
 210 {
 211         struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
 212         struct ttm_resource *old_mem = &bo->mem;
 213         struct ttm_resource tmp_mem;
 214         struct ttm_place placements;
 215         struct ttm_placement placement;
 216         int r;
 217
 218         tmp_mem = *new_mem;
 219         tmp_mem.mm_node = NULL;
 220         placement.num_placement = 1;
 221         placement.placement = &placements;
 222         placement.num_busy_placement = 1;
 223         placement.busy_placement = &placements;
 224         placements.fpfn = 0;
 225         placements.lpfn = 0;
 226         placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 227         r = ttm_bo_mem_space(bo, &placement, &tmp_mem, &ctx);
 228         if (unlikely(r)) {
 229                 return r;
 230         }
 231
 232         r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
 233         if (unlikely(r)) {
 234                 goto out_cleanup;
 235         }
 236
 237         r = ttm_tt_bind(bo->bdev, bo->ttm, &tmp_mem, &ctx);
 238         if (unlikely(r)) {
 239                 goto out_cleanup;
 240         }
 241         r = radeon_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
 242         if (unlikely(r)) {
 243                 goto out_cleanup;
 244         }
 245         r = ttm_bo_move_ttm(bo, &ctx, new_mem);
 246 out_cleanup:
 247         ttm_resource_free(bo, &tmp_mem);
 248         return r;
 249 }
 250
 251 static int radeon_move_ram_vram(struct ttm_buffer_object *bo,
 252                                 bool evict, bool interruptible,
 253                                 bool no_wait_gpu,
 254                                 struct ttm_resource *new_mem)
 255 {
 256         struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
 257         struct ttm_resource *old_mem = &bo->mem;
 258         struct ttm_resource tmp_mem;
 259         struct ttm_placement placement;
 260         struct ttm_place placements;
 261         int r;
 262
 263         tmp_mem = *new_mem;
 264         tmp_mem.mm_node = NULL;
 265         placement.num_placement = 1;
 266         placement.placement = &placements;
 267         placement.num_busy_placement = 1;
 268         placement.busy_placement = &placements;
 269         placements.fpfn = 0;
 270         placements.lpfn = 0;
 271         placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 272         r = ttm_bo_mem_space(bo, &placement, &tmp_mem, &ctx);
 273         if (unlikely(r)) {
 274                 return r;
 275         }
 276         r = ttm_bo_move_ttm(bo, &ctx, &tmp_mem);
 277         if (unlikely(r)) {
 278                 goto out_cleanup;
 279         }
 280         r = radeon_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
 281         if (unlikely(r)) {
 282                 goto out_cleanup;
 283         }
 284 out_cleanup:
 285         ttm_resource_free(bo, &tmp_mem);
 286         return r;
 287 }
 288
 289 static int radeon_bo_move(struct ttm_buffer_object *bo, bool evict,
 290                           struct ttm_operation_ctx *ctx,
 291                           struct ttm_resource *new_mem)
 292 {
 293         struct radeon_device *rdev;
 294         struct radeon_bo *rbo;
 295         struct ttm_resource *old_mem = &bo->mem;
 296         int r;
 297
 298         r = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
 299         if (r)
 300                 return r;
 301
 302         /* Can't move a pinned BO */
 303         rbo = container_of(bo, struct radeon_bo, tbo);
 304         if (WARN_ON_ONCE(rbo->pin_count > 0))
 305                 return -EINVAL;
 306
 307         rdev = radeon_get_rdev(bo->bdev);
 308         if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
 309                 ttm_bo_move_null(bo, new_mem);
 310                 return 0;
 311         }
 312         if ((old_mem->mem_type == TTM_PL_TT &&
 313              new_mem->mem_type == TTM_PL_SYSTEM) ||
 314             (old_mem->mem_type == TTM_PL_SYSTEM &&
 315              new_mem->mem_type == TTM_PL_TT)) {
 316                 /* bind is enough */
 317                 ttm_bo_move_null(bo, new_mem);
 318                 return 0;
 319         }
 320         if (!rdev->ring[radeon_copy_ring_index(rdev)].ready ||
 321             rdev->asic->copy.copy == NULL) {
 322                 /* use memcpy */
 323                 goto memcpy;
 324         }
 325
 326         if (old_mem->mem_type == TTM_PL_VRAM &&
 327             new_mem->mem_type == TTM_PL_SYSTEM) {
 328                 r = radeon_move_vram_ram(bo, evict, ctx->interruptible,
 329                                         ctx->no_wait_gpu, new_mem);
 330         } else if (old_mem->mem_type == TTM_PL_SYSTEM &&
 331                    new_mem->mem_type == TTM_PL_VRAM) {
 332                 r = radeon_move_ram_vram(bo, evict, ctx->interruptible,
 333                                             ctx->no_wait_gpu, new_mem);
 334         } else {
 335                 r = radeon_move_blit(bo, evict, ctx->no_wait_gpu,
 336                                      new_mem, old_mem);
 337         }
 338
 339         if (r) {
 340 memcpy:
 341                 r = ttm_bo_move_memcpy(bo, ctx, new_mem);
 342                 if (r) {
 343                         return r;
 344                 }
 345         }
 346
 347         /* update statistics */
 348         atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &rdev->num_bytes_moved);
 349         return 0;
 350 }
 351
 352 static int radeon_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_resource *mem)
 353 {
 354         struct radeon_device *rdev = radeon_get_rdev(bdev);
 355         size_t bus_size = (size_t)mem->num_pages << PAGE_SHIFT;
 356
 357         switch (mem->mem_type) {
 358         case TTM_PL_SYSTEM:
 359                 /* system memory */
 360                 return 0;
 361         case TTM_PL_TT:
 362 #if IS_ENABLED(CONFIG_AGP)
 363                 if (rdev->flags & RADEON_IS_AGP) {
 364                         /* RADEON_IS_AGP is set only if AGP is active */
 365                         mem->bus.offset = (mem->start << PAGE_SHIFT) +
 366                                 rdev->mc.agp_base;
 367                         mem->bus.is_iomem = !rdev->ddev->agp->cant_use_aperture;
 368                 }
 369 #endif
 370                 break;
 371         case TTM_PL_VRAM:
 372                 mem->bus.offset = mem->start << PAGE_SHIFT;
 373                 /* check if it's visible */
 374                 if ((mem->bus.offset + bus_size) > rdev->mc.visible_vram_size)
 375                         return -EINVAL;
 376                 mem->bus.offset += rdev->mc.aper_base;
 377                 mem->bus.is_iomem = true;
 378 #ifdef __alpha__
 379                 /*
 380                  * Alpha: use bus.addr to hold the ioremap() return,
 381                  * so we can modify bus.base below.
 382                  */
 383                 if (mem->placement & TTM_PL_FLAG_WC)
 384                         mem->bus.addr =
 385                                 ioremap_wc(mem->bus.offset, bus_size);
 386                 else
 387                         mem->bus.addr =
 388                                 ioremap(mem->bus.offset, bus_size);
 389                 if (!mem->bus.addr)
 390                         return -ENOMEM;
 391
 392                 /*
 393                  * Alpha: Use just the bus offset plus
 394                  * the hose/domain memory base for bus.base.
 395                  * It then can be used to build PTEs for VRAM
 396                  * access, as done in ttm_bo_vm_fault().
 397                  */
 398                 mem->bus.offset = (mem->bus.offset & 0x0ffffffffUL) +
 399                         rdev->ddev->hose->dense_mem_base;
 400 #endif
 401                 break;
 402         default:
 403                 return -EINVAL;
 404         }
 405         return 0;
 406 }
 407
 408 /*
 409  * TTM backend functions.
 410  */
 411 struct radeon_ttm_tt {
 412         struct ttm_dma_tt               ttm;
 413         u64                             offset;
 414
 415         uint64_t                        userptr;
 416         struct mm_struct                *usermm;
 417         uint32_t                        userflags;
 418 };
 419
 420 /* prepare the sg table with the user pages */
 421 static int radeon_ttm_tt_pin_userptr(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
 422 {
 423         struct radeon_device *rdev = radeon_get_rdev(bdev);
 424         struct radeon_ttm_tt *gtt = (void *)ttm;
 425         unsigned pinned = 0;
 426         int r;
 427
 428         int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
 429         enum dma_data_direction direction = write ?
 430                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 431
 432         if (current->mm != gtt->usermm)
 433                 return -EPERM;
 434
 435         if (gtt->userflags & RADEON_GEM_USERPTR_ANONONLY) {
 436                 /* check that we only pin down anonymous memory
 437                    to prevent problems with writeback */
 438                 unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
 439                 struct vm_area_struct *vma;
 440                 vma = find_vma(gtt->usermm, gtt->userptr);
 441                 if (!vma || vma->vm_file || vma->vm_end < end)
 442                         return -EPERM;
 443         }
 444
 445         do {
 446                 unsigned num_pages = ttm->num_pages - pinned;
 447                 uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
 448                 struct page **pages = ttm->pages + pinned;
 449
 450                 r = get_user_pages(userptr, num_pages, write ? FOLL_WRITE : 0,
 451                                    pages, NULL);
 452                 if (r < 0)
 453                         goto release_pages;
 454
 455                 pinned += r;
 456
 457         } while (pinned < ttm->num_pages);
 458
 459         r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
 460                                       ttm->num_pages << PAGE_SHIFT,
 461                                       GFP_KERNEL);
 462         if (r)
 463                 goto release_sg;
 464
 465         r = dma_map_sgtable(rdev->dev, ttm->sg, direction, 0);
 466         if (r)
 467                 goto release_sg;
 468
 469         drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
 470                                          gtt->ttm.dma_address, ttm->num_pages);
 471
 472         return 0;
 473
 474 release_sg:
 475         kfree(ttm->sg);
 476
 477 release_pages:
 478         release_pages(ttm->pages, pinned);
 479         return r;
 480 }
 481
 482 static void radeon_ttm_tt_unpin_userptr(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
 483 {
 484         struct radeon_device *rdev = radeon_get_rdev(bdev);
 485         struct radeon_ttm_tt *gtt = (void *)ttm;
 486         struct sg_page_iter sg_iter;
 487
 488         int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
 489         enum dma_data_direction direction = write ?
 490                 DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
 491
 492         /* double check that we don't free the table twice */
 493         if (!ttm->sg->sgl)
 494                 return;
 495
 496         /* free the sg table and pages again */
 497         dma_unmap_sgtable(rdev->dev, ttm->sg, direction, 0);
 498
 499         for_each_sgtable_page(ttm->sg, &sg_iter, 0) {
 500                 struct page *page = sg_page_iter_page(&sg_iter);
 501                 if (!(gtt->userflags & RADEON_GEM_USERPTR_READONLY))
 502                         set_page_dirty(page);
 503
 504                 mark_page_accessed(page);
 505                 put_page(page);
 506         }
 507
 508         sg_free_table(ttm->sg);
 509 }
 510
 511 static int radeon_ttm_backend_bind(struct ttm_bo_device *bdev,
 512                                    struct ttm_tt *ttm,
 513                                    struct ttm_resource *bo_mem)
 514 {
 515         struct radeon_ttm_tt *gtt = (void*)ttm;
 516         struct radeon_device *rdev = radeon_get_rdev(bdev);
 517         uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ |
 518                 RADEON_GART_PAGE_WRITE;
 519         int r;
 520
 521         if (gtt->userptr) {
 522                 radeon_ttm_tt_pin_userptr(bdev, ttm);
 523                 flags &= ~RADEON_GART_PAGE_WRITE;
 524         }
 525
 526         gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
 527         if (!ttm->num_pages) {
 528                 WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
 529                      ttm->num_pages, bo_mem, ttm);
 530         }
 531         if (ttm->caching_state == tt_cached)
 532                 flags |= RADEON_GART_PAGE_SNOOP;
 533         r = radeon_gart_bind(rdev, gtt->offset, ttm->num_pages,
 534                              ttm->pages, gtt->ttm.dma_address, flags);
 535         if (r) {
 536                 DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
 537                           ttm->num_pages, (unsigned)gtt->offset);
 538                 return r;
 539         }
 540         return 0;
 541 }
 542
 543 static void radeon_ttm_backend_unbind(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
 544 {
 545         struct radeon_ttm_tt *gtt = (void *)ttm;
 546         struct radeon_device *rdev = radeon_get_rdev(bdev);
 547
 548         radeon_gart_unbind(rdev, gtt->offset, ttm->num_pages);
 549
 550         if (gtt->userptr)
 551                 radeon_ttm_tt_unpin_userptr(bdev, ttm);
 552 }
 553
 554 static void radeon_ttm_backend_destroy(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
 555 {
 556         struct radeon_ttm_tt *gtt = (void *)ttm;
 557
 558         ttm_dma_tt_fini(&gtt->ttm);
 559         kfree(gtt);
 560 }
 561
 562 static struct ttm_backend_func radeon_backend_func = {
 563         .bind = &radeon_ttm_backend_bind,
 564         .unbind = &radeon_ttm_backend_unbind,
 565         .destroy = &radeon_ttm_backend_destroy,
 566 };
 567
 568 static struct ttm_tt *radeon_ttm_tt_create(struct ttm_buffer_object *bo,
 569                                            uint32_t page_flags)
 570 {
 571         struct radeon_device *rdev;
 572         struct radeon_ttm_tt *gtt;
 573
 574         rdev = radeon_get_rdev(bo->bdev);
 575 #if IS_ENABLED(CONFIG_AGP)
 576         if (rdev->flags & RADEON_IS_AGP) {
 577                 return ttm_agp_tt_create(bo, rdev->ddev->agp->bridge,
 578                                          page_flags);
 579         }
 580 #endif
 581
 582         gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
 583         if (gtt == NULL) {
 584                 return NULL;
 585         }
 586         gtt->ttm.ttm.func = &radeon_backend_func;
 587         if (ttm_dma_tt_init(&gtt->ttm, bo, page_flags)) {
 588                 kfree(gtt);
 589                 return NULL;
 590         }
 591         return &gtt->ttm.ttm;
 592 }
 593
 594 static struct radeon_ttm_tt *radeon_ttm_tt_to_gtt(struct ttm_tt *ttm)
 595 {
 596         if (!ttm || ttm->func != &radeon_backend_func)
 597                 return NULL;
 598         return (struct radeon_ttm_tt *)ttm;
 599 }
 600
 601 static int radeon_ttm_tt_populate(struct ttm_bo_device *bdev,
 602                                   struct ttm_tt *ttm,
 603                                   struct ttm_operation_ctx *ctx)
 604 {
 605         struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
 606         struct radeon_device *rdev;
 607         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
 608
 609         if (gtt && gtt->userptr) {
 610                 ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
 611                 if (!ttm->sg)
 612                         return -ENOMEM;
 613
 614                 ttm->page_flags |= TTM_PAGE_FLAG_SG;
 615                 ttm->state = tt_unbound;
 616                 return 0;
 617         }
 618
 619         if (slave && ttm->sg) {
 620                 drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
 621                                                  gtt->ttm.dma_address, ttm->num_pages);
 622                 ttm->state = tt_unbound;
 623                 return 0;
 624         }
 625
 626         rdev = radeon_get_rdev(bdev);
 627 #if IS_ENABLED(CONFIG_AGP)
 628         if (rdev->flags & RADEON_IS_AGP) {
 629                 return ttm_agp_tt_populate(bdev, ttm, ctx);
 630         }
 631 #endif
 632
 633 #ifdef CONFIG_SWIOTLB
 634         if (rdev->need_swiotlb && swiotlb_nr_tbl()) {
 635                 return ttm_dma_populate(&gtt->ttm, rdev->dev, ctx);
 636         }
 637 #endif
 638
 639         return ttm_populate_and_map_pages(rdev->dev, &gtt->ttm, ctx);
 640 }
 641
 642 static void radeon_ttm_tt_unpopulate(struct ttm_bo_device *bdev, struct ttm_tt *ttm)
 643 {
 644         struct radeon_device *rdev;
 645         struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
 646         bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
 647
 648         if (gtt && gtt->userptr) {
 649                 kfree(ttm->sg);
 650                 ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
 651                 return;
 652         }
 653
 654         if (slave)
 655                 return;
 656
 657         rdev = radeon_get_rdev(bdev);
 658 #if IS_ENABLED(CONFIG_AGP)
 659         if (rdev->flags & RADEON_IS_AGP) {
 660                 ttm_agp_tt_unpopulate(bdev, ttm);
 661                 return;
 662         }
 663 #endif
 664
 665 #ifdef CONFIG_SWIOTLB
 666         if (rdev->need_swiotlb && swiotlb_nr_tbl()) {
 667                 ttm_dma_unpopulate(&gtt->ttm, rdev->dev);
 668                 return;
 669         }
 670 #endif
 671
 672         ttm_unmap_and_unpopulate_pages(rdev->dev, &gtt->ttm);
 673 }
 674
 675 int radeon_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
 676                               uint32_t flags)
 677 {
 678         struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
 679
 680         if (gtt == NULL)
 681                 return -EINVAL;
 682
 683         gtt->userptr = addr;
 684         gtt->usermm = current->mm;
 685         gtt->userflags = flags;
 686         return 0;
 687 }
 688
 689 bool radeon_ttm_tt_has_userptr(struct ttm_tt *ttm)
 690 {
 691         struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
 692
 693         if (gtt == NULL)
 694                 return false;
 695
 696         return !!gtt->userptr;
 697 }
 698
 699 bool radeon_ttm_tt_is_readonly(struct ttm_tt *ttm)
 700 {
 701         struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
 702
 703         if (gtt == NULL)
 704                 return false;
 705
 706         return !!(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
 707 }
 708
 709 static struct ttm_bo_driver radeon_bo_driver = {
 710         .ttm_tt_create = &radeon_ttm_tt_create,
 711         .ttm_tt_populate = &radeon_ttm_tt_populate,
 712         .ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
 713         .eviction_valuable = ttm_bo_eviction_valuable,
 714         .evict_flags = &radeon_evict_flags,
 715         .move = &radeon_bo_move,
 716         .verify_access = &radeon_verify_access,
 717         .move_notify = &radeon_bo_move_notify,
 718         .fault_reserve_notify = &radeon_bo_fault_reserve_notify,
 719         .io_mem_reserve = &radeon_ttm_io_mem_reserve,
 720 };
 721
 722 int radeon_ttm_init(struct radeon_device *rdev)
 723 {
 724         int r;
 725
 726         /* No others user of address space so set it to 0 */
 727         r = ttm_bo_device_init(&rdev->mman.bdev,
 728                                &radeon_bo_driver,
 729                                rdev->ddev->anon_inode->i_mapping,
 730                                rdev->ddev->vma_offset_manager,
 731                                dma_addressing_limited(&rdev->pdev->dev));
 732         if (r) {
 733                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
 734                 return r;
 735         }
 736         rdev->mman.initialized = true;
 737
 738         r = radeon_ttm_init_vram(rdev);
 739         if (r) {
 740                 DRM_ERROR("Failed initializing VRAM heap.\n");
 741                 return r;
 742         }
 743         /* Change the size here instead of the init above so only lpfn is affected */
 744         radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
 745
 746         r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
 747                              RADEON_GEM_DOMAIN_VRAM, 0, NULL,
 748                              NULL, &rdev->stolen_vga_memory);
 749         if (r) {
 750                 return r;
 751         }
 752         r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
 753         if (r)
 754                 return r;
 755         r = radeon_bo_pin(rdev->stolen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
 756         radeon_bo_unreserve(rdev->stolen_vga_memory);
 757         if (r) {
 758                 radeon_bo_unref(&rdev->stolen_vga_memory);
 759                 return r;
 760         }
 761         DRM_INFO("radeon: %uM of VRAM memory ready\n",
 762                  (unsigned) (rdev->mc.real_vram_size / (1024 * 1024)));
 763
 764         r = radeon_ttm_init_gtt(rdev);
 765         if (r) {
 766                 DRM_ERROR("Failed initializing GTT heap.\n");
 767                 return r;
 768         }
 769         DRM_INFO("radeon: %uM of GTT memory ready.\n",
 770                  (unsigned)(rdev->mc.gtt_size / (1024 * 1024)));
 771
 772         r = radeon_ttm_debugfs_init(rdev);
 773         if (r) {
 774                 DRM_ERROR("Failed to init debugfs\n");
 775                 return r;
 776         }
 777         return 0;
 778 }
 779
 780 void radeon_ttm_fini(struct radeon_device *rdev)
 781 {
 782         int r;
 783
 784         if (!rdev->mman.initialized)
 785                 return;
 786         radeon_ttm_debugfs_fini(rdev);
 787         if (rdev->stolen_vga_memory) {
 788                 r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
 789                 if (r == 0) {
 790                         radeon_bo_unpin(rdev->stolen_vga_memory);
 791                         radeon_bo_unreserve(rdev->stolen_vga_memory);
 792                 }
 793                 radeon_bo_unref(&rdev->stolen_vga_memory);
 794         }
 795         ttm_range_man_fini(&rdev->mman.bdev, TTM_PL_VRAM);
 796         ttm_range_man_fini(&rdev->mman.bdev, TTM_PL_TT);
 797         ttm_bo_device_release(&rdev->mman.bdev);
 798         radeon_gart_fini(rdev);
 799         rdev->mman.initialized = false;
 800         DRM_INFO("radeon: ttm finalized\n");
 801 }
 802
 803 /* this should only be called at bootup or when userspace
 804  * isn't running */
 805 void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
 806 {
 807         struct ttm_resource_manager *man;
 808
 809         if (!rdev->mman.initialized)
 810                 return;
 811
 812         man = ttm_manager_type(&rdev->mman.bdev, TTM_PL_VRAM);
 813         /* this just adjusts TTM size idea, which sets lpfn to the correct value */
 814         man->size = size >> PAGE_SHIFT;
 815 }
 816
 817 static vm_fault_t radeon_ttm_fault(struct vm_fault *vmf)
 818 {
 819         struct ttm_buffer_object *bo;
 820         struct radeon_device *rdev;
 821         vm_fault_t ret;
 822
 823         bo = (struct ttm_buffer_object *)vmf->vma->vm_private_data;
 824         if (bo == NULL)
 825                 return VM_FAULT_NOPAGE;
 826
 827         rdev = radeon_get_rdev(bo->bdev);
 828         down_read(&rdev->pm.mclk_lock);
 829         ret = ttm_bo_vm_fault(vmf);
 830         up_read(&rdev->pm.mclk_lock);
 831         return ret;
 832 }
 833
 834 static struct vm_operations_struct radeon_ttm_vm_ops = {
 835         .fault = radeon_ttm_fault,
 836         .open = ttm_bo_vm_open,
 837         .close = ttm_bo_vm_close,
 838         .access = ttm_bo_vm_access
 839 };
 840
 841 int radeon_mmap(struct file *filp, struct vm_area_struct *vma)
 842 {
 843         int r;
 844         struct drm_file *file_priv = filp->private_data;
 845         struct radeon_device *rdev = file_priv->minor->dev->dev_private;
 846
 847         if (rdev == NULL)
 848                 return -EINVAL;
 849
 850         r = ttm_bo_mmap(filp, vma, &rdev->mman.bdev);
 851         if (unlikely(r != 0))
 852                 return r;
 853
 854         vma->vm_ops = &radeon_ttm_vm_ops;
 855         return 0;
 856 }
 857
 858 #if defined(CONFIG_DEBUG_FS)
 859
 860 static int radeon_mm_dump_table(struct seq_file *m, void *data)
 861 {
 862         struct drm_info_node *node = (struct drm_info_node *)m->private;
 863         unsigned ttm_pl = *(int*)node->info_ent->data;
 864         struct drm_device *dev = node->minor->dev;
 865         struct radeon_device *rdev = dev->dev_private;
 866         struct ttm_resource_manager *man = ttm_manager_type(&rdev->mman.bdev, ttm_pl);
 867         struct drm_printer p = drm_seq_file_printer(m);
 868
 869         man->func->debug(man, &p);
 870         return 0;
 871 }
 872
 873
 874 static int ttm_pl_vram = TTM_PL_VRAM;
 875 static int ttm_pl_tt = TTM_PL_TT;
 876
 877 static struct drm_info_list radeon_ttm_debugfs_list[] = {
 878         {"radeon_vram_mm", radeon_mm_dump_table, 0, &ttm_pl_vram},
 879         {"radeon_gtt_mm", radeon_mm_dump_table, 0, &ttm_pl_tt},
 880         {"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
 881 #ifdef CONFIG_SWIOTLB
 882         {"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
 883 #endif
 884 };
 885
 886 static int radeon_ttm_vram_open(struct inode *inode, struct file *filep)
 887 {
 888         struct radeon_device *rdev = inode->i_private;
 889         i_size_write(inode, rdev->mc.mc_vram_size);
 890         filep->private_data = inode->i_private;
 891         return 0;
 892 }
 893
 894 static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf,
 895                                     size_t size, loff_t *pos)
 896 {
 897         struct radeon_device *rdev = f->private_data;
 898         ssize_t result = 0;
 899         int r;
 900
 901         if (size & 0x3 || *pos & 0x3)
 902                 return -EINVAL;
 903
 904         while (size) {
 905                 unsigned long flags;
 906                 uint32_t value;
 907
 908                 if (*pos >= rdev->mc.mc_vram_size)
 909                         return result;
 910
 911                 spin_lock_irqsave(&rdev->mmio_idx_lock, flags);
 912                 WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000);
 913                 if (rdev->family >= CHIP_CEDAR)
 914                         WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31);
 915                 value = RREG32(RADEON_MM_DATA);
 916                 spin_unlock_irqrestore(&rdev->mmio_idx_lock, flags);
 917
 918                 r = put_user(value, (uint32_t *)buf);
 919                 if (r)
 920                         return r;
 921
 922                 result += 4;
 923                 buf += 4;
 924                 *pos += 4;
 925                 size -= 4;
 926         }
 927
 928         return result;
 929 }
 930
 931 static const struct file_operations radeon_ttm_vram_fops = {
 932         .owner = THIS_MODULE,
 933         .open = radeon_ttm_vram_open,
 934         .read = radeon_ttm_vram_read,
 935         .llseek = default_llseek
 936 };
 937
 938 static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep)
 939 {
 940         struct radeon_device *rdev = inode->i_private;
 941         i_size_write(inode, rdev->mc.gtt_size);
 942         filep->private_data = inode->i_private;
 943         return 0;
 944 }
 945
 946 static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf,
 947                                    size_t size, loff_t *pos)
 948 {
 949         struct radeon_device *rdev = f->private_data;
 950         ssize_t result = 0;
 951         int r;
 952
 953         while (size) {
 954                 loff_t p = *pos / PAGE_SIZE;
 955                 unsigned off = *pos & ~PAGE_MASK;
 956                 size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
 957                 struct page *page;
 958                 void *ptr;
 959
 960                 if (p >= rdev->gart.num_cpu_pages)
 961                         return result;
 962
 963                 page = rdev->gart.pages[p];
 964                 if (page) {
 965                         ptr = kmap(page);
 966                         ptr += off;
 967
 968                         r = copy_to_user(buf, ptr, cur_size);
 969                         kunmap(rdev->gart.pages[p]);
 970                 } else
 971                         r = clear_user(buf, cur_size);
 972
 973                 if (r)
 974                         return -EFAULT;
 975
 976                 result += cur_size;
 977                 buf += cur_size;
 978                 *pos += cur_size;
 979                 size -= cur_size;
 980         }
 981
 982         return result;
 983 }
 984
 985 static const struct file_operations radeon_ttm_gtt_fops = {
 986         .owner = THIS_MODULE,
 987         .open = radeon_ttm_gtt_open,
 988         .read = radeon_ttm_gtt_read,
 989         .llseek = default_llseek
 990 };
 991
 992 #endif
 993
 994 static int radeon_ttm_debugfs_init(struct radeon_device *rdev)
 995 {
 996 #if defined(CONFIG_DEBUG_FS)
 997         unsigned count;
 998
 999         struct drm_minor *minor = rdev->ddev->primary;
1000         struct dentry *root = minor->debugfs_root;
1001
1002         rdev->mman.vram = debugfs_create_file("radeon_vram", S_IFREG | S_IRUGO,
1003                                               root, rdev,
1004                                               &radeon_ttm_vram_fops);
1005
1006         rdev->mman.gtt = debugfs_create_file("radeon_gtt", S_IFREG | S_IRUGO,
1007                                              root, rdev, &radeon_ttm_gtt_fops);
1008
1009         count = ARRAY_SIZE(radeon_ttm_debugfs_list);
1010
1011 #ifdef CONFIG_SWIOTLB
1012         if (!(rdev->need_swiotlb && swiotlb_nr_tbl()))
1013                 --count;
1014 #endif
1015
1016         return radeon_debugfs_add_files(rdev, radeon_ttm_debugfs_list, count);
1017 #else
1018
1019         return 0;
1020 #endif
1021 }
1022
1023 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev)
1024 {
1025 #if defined(CONFIG_DEBUG_FS)
1026
1027         debugfs_remove(rdev->mman.vram);
1028         rdev->mman.vram = NULL;
1029
1030         debugfs_remove(rdev->mman.gtt);
1031         rdev->mman.gtt = NULL;
1032 #endif
1033 }