drivers/gpu/drm/radeon/radeon_object.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 #include <linux/list.h>
  33 #include <linux/slab.h>
  34 #include <drm/drmP.h>
  35 #include "radeon_drm.h"
  36 #include "radeon.h"
  37
  38
  39 int radeon_ttm_init(struct radeon_device *rdev);
  40 void radeon_ttm_fini(struct radeon_device *rdev);
  41 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
  42
  43 /*
  44  * To exclude mutual BO access we rely on bo_reserve exclusion, as all
  45  * function are calling it.
  46  */
  47
  48 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
  49 {
  50         struct radeon_bo *bo;
  51
  52         bo = container_of(tbo, struct radeon_bo, tbo);
  53         mutex_lock(&bo->rdev->gem.mutex);
  54         list_del_init(&bo->list);
  55         mutex_unlock(&bo->rdev->gem.mutex);
  56         radeon_bo_clear_surface_reg(bo);
  57         kfree(bo);
  58 }
  59
  60 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
  61 {
  62         if (bo->destroy == &radeon_ttm_bo_destroy)
  63                 return true;
  64         return false;
  65 }
  66
  67 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
  68 {
  69         u32 c = 0;
  70
  71         rbo->placement.fpfn = 0;
  72         rbo->placement.lpfn = 0;
  73         rbo->placement.placement = rbo->placements;
  74         rbo->placement.busy_placement = rbo->placements;
  75         if (domain & RADEON_GEM_DOMAIN_VRAM)
  76                 rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
  77                                         TTM_PL_FLAG_VRAM;
  78         if (domain & RADEON_GEM_DOMAIN_GTT)
  79                 rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
  80         if (domain & RADEON_GEM_DOMAIN_CPU)
  81                 rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
  82         if (!c)
  83                 rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
  84         rbo->placement.num_placement = c;
  85         rbo->placement.num_busy_placement = c;
  86 }
  87
  88 int radeon_bo_create(struct radeon_device *rdev, struct drm_gem_object *gobj,
  89                         unsigned long size, bool kernel, u32 domain,
  90                         struct radeon_bo **bo_ptr)
  91 {
  92         struct radeon_bo *bo;
  93         enum ttm_bo_type type;
  94         int r;
  95
  96         if (unlikely(rdev->mman.bdev.dev_mapping == NULL)) {
  97                 rdev->mman.bdev.dev_mapping = rdev->ddev->dev_mapping;
  98         }
  99         if (kernel) {
 100                 type = ttm_bo_type_kernel;
 101         } else {
 102                 type = ttm_bo_type_device;
 103         }
 104         *bo_ptr = NULL;
 105         bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
 106         if (bo == NULL)
 107                 return -ENOMEM;
 108         bo->rdev = rdev;
 109         bo->gobj = gobj;
 110         bo->surface_reg = -1;
 111         INIT_LIST_HEAD(&bo->list);
 112
 113         radeon_ttm_placement_from_domain(bo, domain);
 114         /* Kernel allocation are uninterruptible */
 115         mutex_lock(&rdev->vram_mutex);
 116         r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
 117                         &bo->placement, 0, 0, !kernel, NULL, size,
 118                         &radeon_ttm_bo_destroy);
 119         mutex_unlock(&rdev->vram_mutex);
 120         if (unlikely(r != 0)) {
 121                 if (r != -ERESTARTSYS)
 122                         dev_err(rdev->dev,
 123                                 "object_init failed for (%lu, 0x%08X)\n",
 124                                 size, domain);
 125                 return r;
 126         }
 127         *bo_ptr = bo;
 128         if (gobj) {
 129                 mutex_lock(&bo->rdev->gem.mutex);
 130                 list_add_tail(&bo->list, &rdev->gem.objects);
 131                 mutex_unlock(&bo->rdev->gem.mutex);
 132         }
 133         return 0;
 134 }
 135
 136 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
 137 {
 138         bool is_iomem;
 139         int r;
 140
 141         if (bo->kptr) {
 142                 if (ptr) {
 143                         *ptr = bo->kptr;
 144                 }
 145                 return 0;
 146         }
 147         r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
 148         if (r) {
 149                 return r;
 150         }
 151         bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
 152         if (ptr) {
 153                 *ptr = bo->kptr;
 154         }
 155         radeon_bo_check_tiling(bo, 0, 0);
 156         return 0;
 157 }
 158
 159 void radeon_bo_kunmap(struct radeon_bo *bo)
 160 {
 161         if (bo->kptr == NULL)
 162                 return;
 163         bo->kptr = NULL;
 164         radeon_bo_check_tiling(bo, 0, 0);
 165         ttm_bo_kunmap(&bo->kmap);
 166 }
 167
 168 void radeon_bo_unref(struct radeon_bo **bo)
 169 {
 170         struct ttm_buffer_object *tbo;
 171         struct radeon_device *rdev;
 172
 173         if ((*bo) == NULL)
 174                 return;
 175         rdev = (*bo)->rdev;
 176         tbo = &((*bo)->tbo);
 177         mutex_lock(&rdev->vram_mutex);
 178         ttm_bo_unref(&tbo);
 179         mutex_unlock(&rdev->vram_mutex);
 180         if (tbo == NULL)
 181                 *bo = NULL;
 182 }
 183
 184 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
 185 {
 186         int r, i;
 187
 188         if (bo->pin_count) {
 189                 bo->pin_count++;
 190                 if (gpu_addr)
 191                         *gpu_addr = radeon_bo_gpu_offset(bo);
 192                 return 0;
 193         }
 194         radeon_ttm_placement_from_domain(bo, domain);
 195         if (domain == RADEON_GEM_DOMAIN_VRAM) {
 196                 /* force to pin into visible video ram */
 197                 bo->placement.lpfn = bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
 198         }
 199         for (i = 0; i < bo->placement.num_placement; i++)
 200                 bo->placements[i] |= TTM_PL_FLAG_NO_EVICT;
 201         r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false, false);
 202         if (likely(r == 0)) {
 203                 bo->pin_count = 1;
 204                 if (gpu_addr != NULL)
 205                         *gpu_addr = radeon_bo_gpu_offset(bo);
 206         }
 207         if (unlikely(r != 0))
 208                 dev_err(bo->rdev->dev, "%p pin failed\n", bo);
 209         return r;
 210 }
 211
 212 int radeon_bo_unpin(struct radeon_bo *bo)
 213 {
 214         int r, i;
 215
 216         if (!bo->pin_count) {
 217                 dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
 218                 return 0;
 219         }
 220         bo->pin_count--;
 221         if (bo->pin_count)
 222                 return 0;
 223         for (i = 0; i < bo->placement.num_placement; i++)
 224                 bo->placements[i] &= ~TTM_PL_FLAG_NO_EVICT;
 225         r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false, false);
 226         if (unlikely(r != 0))
 227                 dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
 228         return r;
 229 }
 230
 231 int radeon_bo_evict_vram(struct radeon_device *rdev)
 232 {
 233         /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
 234         if (0 && (rdev->flags & RADEON_IS_IGP)) {
 235                 if (rdev->mc.igp_sideport_enabled == false)
 236                         /* Useless to evict on IGP chips */
 237                         return 0;
 238         }
 239         return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
 240 }
 241
 242 void radeon_bo_force_delete(struct radeon_device *rdev)
 243 {
 244         struct radeon_bo *bo, *n;
 245         struct drm_gem_object *gobj;
 246
 247         if (list_empty(&rdev->gem.objects)) {
 248                 return;
 249         }
 250         dev_err(rdev->dev, "Userspace still has active objects !\n");
 251         list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
 252                 mutex_lock(&rdev->ddev->struct_mutex);
 253                 gobj = bo->gobj;
 254                 dev_err(rdev->dev, "%p %p %lu %lu force free\n",
 255                         gobj, bo, (unsigned long)gobj->size,
 256                         *((unsigned long *)&gobj->refcount));
 257                 mutex_lock(&bo->rdev->gem.mutex);
 258                 list_del_init(&bo->list);
 259                 mutex_unlock(&bo->rdev->gem.mutex);
 260                 radeon_bo_unref(&bo);
 261                 gobj->driver_private = NULL;
 262                 drm_gem_object_unreference(gobj);
 263                 mutex_unlock(&rdev->ddev->struct_mutex);
 264         }
 265 }
 266
 267 int radeon_bo_init(struct radeon_device *rdev)
 268 {
 269         /* Add an MTRR for the VRAM */
 270         rdev->mc.vram_mtrr = mtrr_add(rdev->mc.aper_base, rdev->mc.aper_size,
 271                         MTRR_TYPE_WRCOMB, 1);
 272         DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
 273                 rdev->mc.mc_vram_size >> 20,
 274                 (unsigned long long)rdev->mc.aper_size >> 20);
 275         DRM_INFO("RAM width %dbits %cDR\n",
 276                         rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
 277         return radeon_ttm_init(rdev);
 278 }
 279
 280 void radeon_bo_fini(struct radeon_device *rdev)
 281 {
 282         radeon_ttm_fini(rdev);
 283 }
 284
 285 void radeon_bo_list_add_object(struct radeon_bo_list *lobj,
 286                                 struct list_head *head)
 287 {
 288         if (lobj->wdomain) {
 289                 list_add(&lobj->list, head);
 290         } else {
 291                 list_add_tail(&lobj->list, head);
 292         }
 293 }
 294
 295 int radeon_bo_list_reserve(struct list_head *head)
 296 {
 297         struct radeon_bo_list *lobj;
 298         int r;
 299
 300         list_for_each_entry(lobj, head, list){
 301                 r = radeon_bo_reserve(lobj->bo, false);
 302                 if (unlikely(r != 0))
 303                         return r;
 304         }
 305         return 0;
 306 }
 307
 308 void radeon_bo_list_unreserve(struct list_head *head)
 309 {
 310         struct radeon_bo_list *lobj;
 311
 312         list_for_each_entry(lobj, head, list) {
 313                 /* only unreserve object we successfully reserved */
 314                 if (radeon_bo_is_reserved(lobj->bo))
 315                         radeon_bo_unreserve(lobj->bo);
 316         }
 317 }
 318
 319 int radeon_bo_list_validate(struct list_head *head)
 320 {
 321         struct radeon_bo_list *lobj;
 322         struct radeon_bo *bo;
 323         int r;
 324
 325         r = radeon_bo_list_reserve(head);
 326         if (unlikely(r != 0)) {
 327                 return r;
 328         }
 329         list_for_each_entry(lobj, head, list) {
 330                 bo = lobj->bo;
 331                 if (!bo->pin_count) {
 332                         if (lobj->wdomain) {
 333                                 radeon_ttm_placement_from_domain(bo,
 334                                                                 lobj->wdomain);
 335                         } else {
 336                                 radeon_ttm_placement_from_domain(bo,
 337                                                                 lobj->rdomain);
 338                         }
 339                         r = ttm_bo_validate(&bo->tbo, &bo->placement,
 340                                                 true, false, false);
 341                         if (unlikely(r))
 342                                 return r;
 343                 }
 344                 lobj->gpu_offset = radeon_bo_gpu_offset(bo);
 345                 lobj->tiling_flags = bo->tiling_flags;
 346         }
 347         return 0;
 348 }
 349
 350 void radeon_bo_list_fence(struct list_head *head, void *fence)
 351 {
 352         struct radeon_bo_list *lobj;
 353         struct radeon_bo *bo;
 354         struct radeon_fence *old_fence = NULL;
 355
 356         list_for_each_entry(lobj, head, list) {
 357                 bo = lobj->bo;
 358                 spin_lock(&bo->tbo.lock);
 359                 old_fence = (struct radeon_fence *)bo->tbo.sync_obj;
 360                 bo->tbo.sync_obj = radeon_fence_ref(fence);
 361                 bo->tbo.sync_obj_arg = NULL;
 362                 spin_unlock(&bo->tbo.lock);
 363                 if (old_fence) {
 364                         radeon_fence_unref(&old_fence);
 365                 }
 366         }
 367 }
 368
 369 int radeon_bo_fbdev_mmap(struct radeon_bo *bo,
 370                              struct vm_area_struct *vma)
 371 {
 372         return ttm_fbdev_mmap(vma, &bo->tbo);
 373 }
 374
 375 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
 376 {
 377         struct radeon_device *rdev = bo->rdev;
 378         struct radeon_surface_reg *reg;
 379         struct radeon_bo *old_object;
 380         int steal;
 381         int i;
 382
 383         BUG_ON(!atomic_read(&bo->tbo.reserved));
 384
 385         if (!bo->tiling_flags)
 386                 return 0;
 387
 388         if (bo->surface_reg >= 0) {
 389                 reg = &rdev->surface_regs[bo->surface_reg];
 390                 i = bo->surface_reg;
 391                 goto out;
 392         }
 393
 394         steal = -1;
 395         for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
 396
 397                 reg = &rdev->surface_regs[i];
 398                 if (!reg->bo)
 399                         break;
 400
 401                 old_object = reg->bo;
 402                 if (old_object->pin_count == 0)
 403                         steal = i;
 404         }
 405
 406         /* if we are all out */
 407         if (i == RADEON_GEM_MAX_SURFACES) {
 408                 if (steal == -1)
 409                         return -ENOMEM;
 410                 /* find someone with a surface reg and nuke their BO */
 411                 reg = &rdev->surface_regs[steal];
 412                 old_object = reg->bo;
 413                 /* blow away the mapping */
 414                 DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
 415                 ttm_bo_unmap_virtual(&old_object->tbo);
 416                 old_object->surface_reg = -1;
 417                 i = steal;
 418         }
 419
 420         bo->surface_reg = i;
 421         reg->bo = bo;
 422
 423 out:
 424         radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
 425                                bo->tbo.mem.mm_node->start << PAGE_SHIFT,
 426                                bo->tbo.num_pages << PAGE_SHIFT);
 427         return 0;
 428 }
 429
 430 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
 431 {
 432         struct radeon_device *rdev = bo->rdev;
 433         struct radeon_surface_reg *reg;
 434
 435         if (bo->surface_reg == -1)
 436                 return;
 437
 438         reg = &rdev->surface_regs[bo->surface_reg];
 439         radeon_clear_surface_reg(rdev, bo->surface_reg);
 440
 441         reg->bo = NULL;
 442         bo->surface_reg = -1;
 443 }
 444
 445 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
 446                                 uint32_t tiling_flags, uint32_t pitch)
 447 {
 448         int r;
 449
 450         r = radeon_bo_reserve(bo, false);
 451         if (unlikely(r != 0))
 452                 return r;
 453         bo->tiling_flags = tiling_flags;
 454         bo->pitch = pitch;
 455         radeon_bo_unreserve(bo);
 456         return 0;
 457 }
 458
 459 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
 460                                 uint32_t *tiling_flags,
 461                                 uint32_t *pitch)
 462 {
 463         BUG_ON(!atomic_read(&bo->tbo.reserved));
 464         if (tiling_flags)
 465                 *tiling_flags = bo->tiling_flags;
 466         if (pitch)
 467                 *pitch = bo->pitch;
 468 }
 469
 470 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
 471                                 bool force_drop)
 472 {
 473         BUG_ON(!atomic_read(&bo->tbo.reserved));
 474
 475         if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
 476                 return 0;
 477
 478         if (force_drop) {
 479                 radeon_bo_clear_surface_reg(bo);
 480                 return 0;
 481         }
 482
 483         if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
 484                 if (!has_moved)
 485                         return 0;
 486
 487                 if (bo->surface_reg >= 0)
 488                         radeon_bo_clear_surface_reg(bo);
 489                 return 0;
 490         }
 491
 492         if ((bo->surface_reg >= 0) && !has_moved)
 493                 return 0;
 494
 495         return radeon_bo_get_surface_reg(bo);
 496 }
 497
 498 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
 499                            struct ttm_mem_reg *mem)
 500 {
 501         struct radeon_bo *rbo;
 502         if (!radeon_ttm_bo_is_radeon_bo(bo))
 503                 return;
 504         rbo = container_of(bo, struct radeon_bo, tbo);
 505         radeon_bo_check_tiling(rbo, 0, 1);
 506 }
 507
 508 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
 509 {
 510         struct radeon_device *rdev;
 511         struct radeon_bo *rbo;
 512         unsigned long offset, size;
 513         int r;
 514
 515         if (!radeon_ttm_bo_is_radeon_bo(bo))
 516                 return 0;
 517         rbo = container_of(bo, struct radeon_bo, tbo);
 518         radeon_bo_check_tiling(rbo, 0, 0);
 519         rdev = rbo->rdev;
 520         if (bo->mem.mem_type == TTM_PL_VRAM) {
 521                 size = bo->mem.num_pages << PAGE_SHIFT;
 522                 offset = bo->mem.mm_node->start << PAGE_SHIFT;
 523                 if ((offset + size) > rdev->mc.visible_vram_size) {
 524                         /* hurrah the memory is not visible ! */
 525                         radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
 526                         rbo->placement.lpfn = rdev->mc.visible_vram_size >> PAGE_SHIFT;
 527                         r = ttm_bo_validate(bo, &rbo->placement, false, true, false);
 528                         if (unlikely(r != 0))
 529                                 return r;
 530                         offset = bo->mem.mm_node->start << PAGE_SHIFT;
 531                         /* this should not happen */
 532                         if ((offset + size) > rdev->mc.visible_vram_size)
 533                                 return -EINVAL;
 534                 }
 535         }
 536         return 0;
 537 }