drivers/gpu/drm/vmwgfx/vmwgfx_ttm_buffer.c

   1 // SPDX-License-Identifier: GPL-2.0 OR MIT
   2 /**************************************************************************
   3  *
   4  * Copyright 2009-2023 VMware, Inc., Palo Alto, CA., USA
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  20  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  21  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  22  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  23  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  24  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 #include "vmwgfx_bo.h"
  29 #include "vmwgfx_drv.h"
  30 #include <drm/ttm/ttm_placement.h>
  31
  32 static const struct ttm_place vram_placement_flags = {
  33         .fpfn = 0,
  34         .lpfn = 0,
  35         .mem_type = TTM_PL_VRAM,
  36         .flags = 0
  37 };
  38
  39 static const struct ttm_place sys_placement_flags = {
  40         .fpfn = 0,
  41         .lpfn = 0,
  42         .mem_type = TTM_PL_SYSTEM,
  43         .flags = 0
  44 };
  45
  46 struct ttm_placement vmw_vram_placement = {
  47         .num_placement = 1,
  48         .placement = &vram_placement_flags,
  49 };
  50
  51 struct ttm_placement vmw_sys_placement = {
  52         .num_placement = 1,
  53         .placement = &sys_placement_flags,
  54 };
  55
  56 const size_t vmw_tt_size = sizeof(struct vmw_ttm_tt);
  57
  58 /**
  59  * __vmw_piter_non_sg_next: Helper functions to advance
  60  * a struct vmw_piter iterator.
  61  *
  62  * @viter: Pointer to the iterator.
  63  *
  64  * These functions return false if past the end of the list,
  65  * true otherwise. Functions are selected depending on the current
  66  * DMA mapping mode.
  67  */
  68 static bool __vmw_piter_non_sg_next(struct vmw_piter *viter)
  69 {
  70         return ++(viter->i) < viter->num_pages;
  71 }
  72
  73 static bool __vmw_piter_sg_next(struct vmw_piter *viter)
  74 {
  75         bool ret = __vmw_piter_non_sg_next(viter);
  76
  77         return __sg_page_iter_dma_next(&viter->iter) && ret;
  78 }
  79
  80
  81 static dma_addr_t __vmw_piter_dma_addr(struct vmw_piter *viter)
  82 {
  83         return viter->addrs[viter->i];
  84 }
  85
  86 static dma_addr_t __vmw_piter_sg_addr(struct vmw_piter *viter)
  87 {
  88         return sg_page_iter_dma_address(&viter->iter);
  89 }
  90
  91
  92 /**
  93  * vmw_piter_start - Initialize a struct vmw_piter.
  94  *
  95  * @viter: Pointer to the iterator to initialize
  96  * @vsgt: Pointer to a struct vmw_sg_table to initialize from
  97  * @p_offset: Pointer offset used to update current array position
  98  *
  99  * Note that we're following the convention of __sg_page_iter_start, so that
 100  * the iterator doesn't point to a valid page after initialization; it has
 101  * to be advanced one step first.
 102  */
 103 void vmw_piter_start(struct vmw_piter *viter, const struct vmw_sg_table *vsgt,
 104                      unsigned long p_offset)
 105 {
 106         viter->i = p_offset - 1;
 107         viter->num_pages = vsgt->num_pages;
 108         viter->pages = vsgt->pages;
 109         switch (vsgt->mode) {
 110         case vmw_dma_alloc_coherent:
 111                 viter->next = &__vmw_piter_non_sg_next;
 112                 viter->dma_address = &__vmw_piter_dma_addr;
 113                 viter->addrs = vsgt->addrs;
 114                 break;
 115         case vmw_dma_map_populate:
 116         case vmw_dma_map_bind:
 117                 viter->next = &__vmw_piter_sg_next;
 118                 viter->dma_address = &__vmw_piter_sg_addr;
 119                 __sg_page_iter_start(&viter->iter.base, vsgt->sgt->sgl,
 120                                      vsgt->sgt->orig_nents, p_offset);
 121                 break;
 122         default:
 123                 BUG();
 124         }
 125 }
 126
 127 /**
 128  * vmw_ttm_unmap_from_dma - unmap  device addresses previsouly mapped for
 129  * TTM pages
 130  *
 131  * @vmw_tt: Pointer to a struct vmw_ttm_backend
 132  *
 133  * Used to free dma mappings previously mapped by vmw_ttm_map_for_dma.
 134  */
 135 static void vmw_ttm_unmap_from_dma(struct vmw_ttm_tt *vmw_tt)
 136 {
 137         struct device *dev = vmw_tt->dev_priv->drm.dev;
 138
 139         dma_unmap_sgtable(dev, &vmw_tt->sgt, DMA_BIDIRECTIONAL, 0);
 140         vmw_tt->sgt.nents = vmw_tt->sgt.orig_nents;
 141 }
 142
 143 /**
 144  * vmw_ttm_map_for_dma - map TTM pages to get device addresses
 145  *
 146  * @vmw_tt: Pointer to a struct vmw_ttm_backend
 147  *
 148  * This function is used to get device addresses from the kernel DMA layer.
 149  * However, it's violating the DMA API in that when this operation has been
 150  * performed, it's illegal for the CPU to write to the pages without first
 151  * unmapping the DMA mappings, or calling dma_sync_sg_for_cpu(). It is
 152  * therefore only legal to call this function if we know that the function
 153  * dma_sync_sg_for_cpu() is a NOP, and dma_sync_sg_for_device() is at most
 154  * a CPU write buffer flush.
 155  */
 156 static int vmw_ttm_map_for_dma(struct vmw_ttm_tt *vmw_tt)
 157 {
 158         struct device *dev = vmw_tt->dev_priv->drm.dev;
 159
 160         return dma_map_sgtable(dev, &vmw_tt->sgt, DMA_BIDIRECTIONAL, 0);
 161 }
 162
 163 /**
 164  * vmw_ttm_map_dma - Make sure TTM pages are visible to the device
 165  *
 166  * @vmw_tt: Pointer to a struct vmw_ttm_tt
 167  *
 168  * Select the correct function for and make sure the TTM pages are
 169  * visible to the device. Allocate storage for the device mappings.
 170  * If a mapping has already been performed, indicated by the storage
 171  * pointer being non NULL, the function returns success.
 172  */
 173 static int vmw_ttm_map_dma(struct vmw_ttm_tt *vmw_tt)
 174 {
 175         struct vmw_private *dev_priv = vmw_tt->dev_priv;
 176         struct vmw_sg_table *vsgt = &vmw_tt->vsgt;
 177         int ret = 0;
 178
 179         if (vmw_tt->mapped)
 180                 return 0;
 181
 182         vsgt->mode = dev_priv->map_mode;
 183         vsgt->pages = vmw_tt->dma_ttm.pages;
 184         vsgt->num_pages = vmw_tt->dma_ttm.num_pages;
 185         vsgt->addrs = vmw_tt->dma_ttm.dma_address;
 186         vsgt->sgt = NULL;
 187
 188         switch (dev_priv->map_mode) {
 189         case vmw_dma_map_bind:
 190         case vmw_dma_map_populate:
 191                 if (vmw_tt->dma_ttm.page_flags  & TTM_TT_FLAG_EXTERNAL) {
 192                         vsgt->sgt = vmw_tt->dma_ttm.sg;
 193                 } else {
 194                         vsgt->sgt = &vmw_tt->sgt;
 195                         ret = sg_alloc_table_from_pages_segment(&vmw_tt->sgt,
 196                                 vsgt->pages, vsgt->num_pages, 0,
 197                                 (unsigned long)vsgt->num_pages << PAGE_SHIFT,
 198                                 dma_get_max_seg_size(dev_priv->drm.dev),
 199                                 GFP_KERNEL);
 200                         if (ret)
 201                                 goto out_sg_alloc_fail;
 202                 }
 203
 204                 ret = vmw_ttm_map_for_dma(vmw_tt);
 205                 if (unlikely(ret != 0))
 206                         goto out_map_fail;
 207
 208                 break;
 209         default:
 210                 break;
 211         }
 212
 213         vmw_tt->mapped = true;
 214         return 0;
 215
 216 out_map_fail:
 217         drm_warn(&dev_priv->drm, "VSG table map failed!");
 218         sg_free_table(vsgt->sgt);
 219         vsgt->sgt = NULL;
 220 out_sg_alloc_fail:
 221         return ret;
 222 }
 223
 224 /**
 225  * vmw_ttm_unmap_dma - Tear down any TTM page device mappings
 226  *
 227  * @vmw_tt: Pointer to a struct vmw_ttm_tt
 228  *
 229  * Tear down any previously set up device DMA mappings and free
 230  * any storage space allocated for them. If there are no mappings set up,
 231  * this function is a NOP.
 232  */
 233 static void vmw_ttm_unmap_dma(struct vmw_ttm_tt *vmw_tt)
 234 {
 235         struct vmw_private *dev_priv = vmw_tt->dev_priv;
 236
 237         if (!vmw_tt->vsgt.sgt)
 238                 return;
 239
 240         switch (dev_priv->map_mode) {
 241         case vmw_dma_map_bind:
 242         case vmw_dma_map_populate:
 243                 vmw_ttm_unmap_from_dma(vmw_tt);
 244                 sg_free_table(vmw_tt->vsgt.sgt);
 245                 vmw_tt->vsgt.sgt = NULL;
 246                 break;
 247         default:
 248                 break;
 249         }
 250         vmw_tt->mapped = false;
 251 }
 252
 253 /**
 254  * vmw_bo_sg_table - Return a struct vmw_sg_table object for a
 255  * TTM buffer object
 256  *
 257  * @bo: Pointer to a struct ttm_buffer_object
 258  *
 259  * Returns a pointer to a struct vmw_sg_table object. The object should
 260  * not be freed after use.
 261  * Note that for the device addresses to be valid, the buffer object must
 262  * either be reserved or pinned.
 263  */
 264 const struct vmw_sg_table *vmw_bo_sg_table(struct ttm_buffer_object *bo)
 265 {
 266         struct vmw_ttm_tt *vmw_tt =
 267                 container_of(bo->ttm, struct vmw_ttm_tt, dma_ttm);
 268
 269         return &vmw_tt->vsgt;
 270 }
 271
 272
 273 static int vmw_ttm_bind(struct ttm_device *bdev,
 274                         struct ttm_tt *ttm, struct ttm_resource *bo_mem)
 275 {
 276         struct vmw_ttm_tt *vmw_be =
 277                 container_of(ttm, struct vmw_ttm_tt, dma_ttm);
 278         int ret = 0;
 279
 280         if (!bo_mem)
 281                 return -EINVAL;
 282
 283         if (vmw_be->bound)
 284                 return 0;
 285
 286         ret = vmw_ttm_map_dma(vmw_be);
 287         if (unlikely(ret != 0))
 288                 return ret;
 289
 290         vmw_be->gmr_id = bo_mem->start;
 291         vmw_be->mem_type = bo_mem->mem_type;
 292
 293         switch (bo_mem->mem_type) {
 294         case VMW_PL_GMR:
 295                 ret = vmw_gmr_bind(vmw_be->dev_priv, &vmw_be->vsgt,
 296                                     ttm->num_pages, vmw_be->gmr_id);
 297                 break;
 298         case VMW_PL_MOB:
 299                 if (unlikely(vmw_be->mob == NULL)) {
 300                         vmw_be->mob =
 301                                 vmw_mob_create(ttm->num_pages);
 302                         if (unlikely(vmw_be->mob == NULL))
 303                                 return -ENOMEM;
 304                 }
 305
 306                 ret = vmw_mob_bind(vmw_be->dev_priv, vmw_be->mob,
 307                                     &vmw_be->vsgt, ttm->num_pages,
 308                                     vmw_be->gmr_id);
 309                 break;
 310         case VMW_PL_SYSTEM:
 311                 /* Nothing to be done for a system bind */
 312                 break;
 313         default:
 314                 BUG();
 315         }
 316         vmw_be->bound = true;
 317         return ret;
 318 }
 319
 320 static void vmw_ttm_unbind(struct ttm_device *bdev,
 321                            struct ttm_tt *ttm)
 322 {
 323         struct vmw_ttm_tt *vmw_be =
 324                 container_of(ttm, struct vmw_ttm_tt, dma_ttm);
 325
 326         if (!vmw_be->bound)
 327                 return;
 328
 329         switch (vmw_be->mem_type) {
 330         case VMW_PL_GMR:
 331                 vmw_gmr_unbind(vmw_be->dev_priv, vmw_be->gmr_id);
 332                 break;
 333         case VMW_PL_MOB:
 334                 vmw_mob_unbind(vmw_be->dev_priv, vmw_be->mob);
 335                 break;
 336         case VMW_PL_SYSTEM:
 337                 break;
 338         default:
 339                 BUG();
 340         }
 341
 342         if (vmw_be->dev_priv->map_mode == vmw_dma_map_bind)
 343                 vmw_ttm_unmap_dma(vmw_be);
 344         vmw_be->bound = false;
 345 }
 346
 347
 348 static void vmw_ttm_destroy(struct ttm_device *bdev, struct ttm_tt *ttm)
 349 {
 350         struct vmw_ttm_tt *vmw_be =
 351                 container_of(ttm, struct vmw_ttm_tt, dma_ttm);
 352
 353         vmw_ttm_unmap_dma(vmw_be);
 354         ttm_tt_fini(ttm);
 355         if (vmw_be->mob)
 356                 vmw_mob_destroy(vmw_be->mob);
 357
 358         kfree(vmw_be);
 359 }
 360
 361
 362 static int vmw_ttm_populate(struct ttm_device *bdev,
 363                             struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
 364 {
 365         bool external = (ttm->page_flags & TTM_TT_FLAG_EXTERNAL) != 0;
 366
 367         if (ttm_tt_is_populated(ttm))
 368                 return 0;
 369
 370         if (external && ttm->sg)
 371                 return  drm_prime_sg_to_dma_addr_array(ttm->sg,
 372                                                        ttm->dma_address,
 373                                                        ttm->num_pages);
 374
 375         return ttm_pool_alloc(&bdev->pool, ttm, ctx);
 376 }
 377
 378 static void vmw_ttm_unpopulate(struct ttm_device *bdev,
 379                                struct ttm_tt *ttm)
 380 {
 381         struct vmw_ttm_tt *vmw_tt = container_of(ttm, struct vmw_ttm_tt,
 382                                                  dma_ttm);
 383         bool external = (ttm->page_flags & TTM_TT_FLAG_EXTERNAL) != 0;
 384
 385         if (external)
 386                 return;
 387
 388         vmw_ttm_unbind(bdev, ttm);
 389
 390         if (vmw_tt->mob) {
 391                 vmw_mob_destroy(vmw_tt->mob);
 392                 vmw_tt->mob = NULL;
 393         }
 394
 395         vmw_ttm_unmap_dma(vmw_tt);
 396
 397         ttm_pool_free(&bdev->pool, ttm);
 398 }
 399
 400 static struct ttm_tt *vmw_ttm_tt_create(struct ttm_buffer_object *bo,
 401                                         uint32_t page_flags)
 402 {
 403         struct vmw_ttm_tt *vmw_be;
 404         int ret;
 405         bool external = bo->type == ttm_bo_type_sg;
 406
 407         vmw_be = kzalloc(sizeof(*vmw_be), GFP_KERNEL);
 408         if (!vmw_be)
 409                 return NULL;
 410
 411         vmw_be->dev_priv = vmw_priv_from_ttm(bo->bdev);
 412         vmw_be->mob = NULL;
 413
 414         if (external)
 415                 page_flags |= TTM_TT_FLAG_EXTERNAL | TTM_TT_FLAG_EXTERNAL_MAPPABLE;
 416
 417         if (vmw_be->dev_priv->map_mode == vmw_dma_alloc_coherent || external)
 418                 ret = ttm_sg_tt_init(&vmw_be->dma_ttm, bo, page_flags,
 419                                      ttm_cached);
 420         else
 421                 ret = ttm_tt_init(&vmw_be->dma_ttm, bo, page_flags,
 422                                   ttm_cached, 0);
 423         if (unlikely(ret != 0))
 424                 goto out_no_init;
 425
 426         return &vmw_be->dma_ttm;
 427 out_no_init:
 428         kfree(vmw_be);
 429         return NULL;
 430 }
 431
 432 static void vmw_evict_flags(struct ttm_buffer_object *bo,
 433                      struct ttm_placement *placement)
 434 {
 435         *placement = vmw_sys_placement;
 436 }
 437
 438 static int vmw_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *mem)
 439 {
 440         struct vmw_private *dev_priv = vmw_priv_from_ttm(bdev);
 441
 442         switch (mem->mem_type) {
 443         case TTM_PL_SYSTEM:
 444         case VMW_PL_SYSTEM:
 445         case VMW_PL_GMR:
 446         case VMW_PL_MOB:
 447                 return 0;
 448         case TTM_PL_VRAM:
 449                 mem->bus.offset = (mem->start << PAGE_SHIFT) +
 450                         dev_priv->vram_start;
 451                 mem->bus.is_iomem = true;
 452                 mem->bus.caching = ttm_cached;
 453                 break;
 454         default:
 455                 return -EINVAL;
 456         }
 457         return 0;
 458 }
 459
 460 /**
 461  * vmw_move_notify - TTM move_notify_callback
 462  *
 463  * @bo: The TTM buffer object about to move.
 464  * @old_mem: The old memory where we move from
 465  * @new_mem: The struct ttm_resource indicating to what memory
 466  *       region the move is taking place.
 467  *
 468  * Calls move_notify for all subsystems needing it.
 469  * (currently only resources).
 470  */
 471 static void vmw_move_notify(struct ttm_buffer_object *bo,
 472                             struct ttm_resource *old_mem,
 473                             struct ttm_resource *new_mem)
 474 {
 475         vmw_bo_move_notify(bo, new_mem);
 476         vmw_query_move_notify(bo, old_mem, new_mem);
 477 }
 478
 479
 480 /**
 481  * vmw_swap_notify - TTM move_notify_callback
 482  *
 483  * @bo: The TTM buffer object about to be swapped out.
 484  */
 485 static void vmw_swap_notify(struct ttm_buffer_object *bo)
 486 {
 487         vmw_bo_swap_notify(bo);
 488         (void) ttm_bo_wait(bo, false, false);
 489 }
 490
 491 static bool vmw_memtype_is_system(uint32_t mem_type)
 492 {
 493         return mem_type == TTM_PL_SYSTEM || mem_type == VMW_PL_SYSTEM;
 494 }
 495
 496 static int vmw_move(struct ttm_buffer_object *bo,
 497                     bool evict,
 498                     struct ttm_operation_ctx *ctx,
 499                     struct ttm_resource *new_mem,
 500                     struct ttm_place *hop)
 501 {
 502         struct ttm_resource_manager *new_man;
 503         struct ttm_resource_manager *old_man = NULL;
 504         int ret = 0;
 505
 506         new_man = ttm_manager_type(bo->bdev, new_mem->mem_type);
 507         if (bo->resource)
 508                 old_man = ttm_manager_type(bo->bdev, bo->resource->mem_type);
 509
 510         if (new_man->use_tt && !vmw_memtype_is_system(new_mem->mem_type)) {
 511                 ret = vmw_ttm_bind(bo->bdev, bo->ttm, new_mem);
 512                 if (ret)
 513                         return ret;
 514         }
 515
 516         if (!bo->resource || (bo->resource->mem_type == TTM_PL_SYSTEM &&
 517                               bo->ttm == NULL)) {
 518                 ttm_bo_move_null(bo, new_mem);
 519                 return 0;
 520         }
 521
 522         vmw_move_notify(bo, bo->resource, new_mem);
 523
 524         if (old_man && old_man->use_tt && new_man->use_tt) {
 525                 if (vmw_memtype_is_system(bo->resource->mem_type)) {
 526                         ttm_bo_move_null(bo, new_mem);
 527                         return 0;
 528                 }
 529                 ret = ttm_bo_wait_ctx(bo, ctx);
 530                 if (ret)
 531                         goto fail;
 532
 533                 vmw_ttm_unbind(bo->bdev, bo->ttm);
 534                 ttm_resource_free(bo, &bo->resource);
 535                 ttm_bo_assign_mem(bo, new_mem);
 536                 return 0;
 537         } else {
 538                 ret = ttm_bo_move_memcpy(bo, ctx, new_mem);
 539                 if (ret)
 540                         goto fail;
 541         }
 542         return 0;
 543 fail:
 544         vmw_move_notify(bo, new_mem, bo->resource);
 545         return ret;
 546 }
 547
 548 struct ttm_device_funcs vmw_bo_driver = {
 549         .ttm_tt_create = &vmw_ttm_tt_create,
 550         .ttm_tt_populate = &vmw_ttm_populate,
 551         .ttm_tt_unpopulate = &vmw_ttm_unpopulate,
 552         .ttm_tt_destroy = &vmw_ttm_destroy,
 553         .eviction_valuable = ttm_bo_eviction_valuable,
 554         .evict_flags = vmw_evict_flags,
 555         .move = vmw_move,
 556         .swap_notify = vmw_swap_notify,
 557         .io_mem_reserve = &vmw_ttm_io_mem_reserve,
 558 };
 559
 560 int vmw_bo_create_and_populate(struct vmw_private *dev_priv,
 561                                size_t bo_size, u32 domain,
 562                                struct vmw_bo **bo_p)
 563 {
 564         struct ttm_operation_ctx ctx = {
 565                 .interruptible = false,
 566                 .no_wait_gpu = false
 567         };
 568         struct vmw_bo *vbo;
 569         int ret;
 570         struct vmw_bo_params bo_params = {
 571                 .domain = domain,
 572                 .busy_domain = domain,
 573                 .bo_type = ttm_bo_type_kernel,
 574                 .size = bo_size,
 575                 .pin = true,
 576                 .keep_resv = true,
 577         };
 578
 579         ret = vmw_bo_create(dev_priv, &bo_params, &vbo);
 580         if (unlikely(ret != 0))
 581                 return ret;
 582
 583         ret = vmw_ttm_populate(vbo->tbo.bdev, vbo->tbo.ttm, &ctx);
 584         if (likely(ret == 0)) {
 585                 struct vmw_ttm_tt *vmw_tt =
 586                         container_of(vbo->tbo.ttm, struct vmw_ttm_tt, dma_ttm);
 587                 ret = vmw_ttm_map_dma(vmw_tt);
 588         }
 589
 590         ttm_bo_unreserve(&vbo->tbo);
 591
 592         if (likely(ret == 0))
 593                 *bo_p = vbo;
 594         return ret;
 595 }