1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */ 2 /************************************************************************** 3 * 4 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA 5 * All Rights Reserved. 6 * 7 * Permission is hereby granted, free of charge, to any person obtaining a 8 * copy of this software and associated documentation files (the 9 * "Software"), to deal in the Software without restriction, including 10 * without limitation the rights to use, copy, modify, merge, publish, 11 * distribute, sub license, and/or sell copies of the Software, and to 12 * permit persons to whom the Software is furnished to do so, subject to 13 * the following conditions: 14 * 15 * The above copyright notice and this permission notice (including the 16 * next paragraph) shall be included in all copies or substantial portions 17 * of the Software. 18 * 19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 25 * USE OR OTHER DEALINGS IN THE SOFTWARE. 26 * 27 **************************************************************************/ 28 /* 29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com> 30 */ 31 32 #define pr_fmt(fmt) "[TTM] " fmt 33 34 #include <drm/ttm/ttm_module.h> 35 #include <drm/ttm/ttm_bo_driver.h> 36 #include <drm/ttm/ttm_placement.h> 37 #include <linux/jiffies.h> 38 #include <linux/slab.h> 39 #include <linux/sched.h> 40 #include <linux/mm.h> 41 #include <linux/file.h> 42 #include <linux/module.h> 43 #include <linux/atomic.h> 44 #include <linux/reservation.h> 45 46 static void ttm_bo_global_kobj_release(struct kobject *kobj); 47 48 /** 49 * ttm_global_mutex - protecting the global BO state 50 */ 51 DEFINE_MUTEX(ttm_global_mutex); 52 struct ttm_bo_global ttm_bo_glob = { 53 .use_count = 0 54 }; 55 56 static struct attribute ttm_bo_count = { 57 .name = "bo_count", 58 .mode = S_IRUGO 59 }; 60 61 /* default destructor */ 62 static void ttm_bo_default_destroy(struct ttm_buffer_object *bo) 63 { 64 kfree(bo); 65 } 66 67 static inline int ttm_mem_type_from_place(const struct ttm_place *place, 68 uint32_t *mem_type) 69 { 70 int pos; 71 72 pos = ffs(place->flags & TTM_PL_MASK_MEM); 73 if (unlikely(!pos)) 74 return -EINVAL; 75 76 *mem_type = pos - 1; 77 return 0; 78 } 79 80 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, struct drm_printer *p, 81 int mem_type) 82 { 83 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 84 85 drm_printf(p, " has_type: %d\n", man->has_type); 86 drm_printf(p, " use_type: %d\n", man->use_type); 87 drm_printf(p, " flags: 0x%08X\n", man->flags); 88 drm_printf(p, " gpu_offset: 0x%08llX\n", man->gpu_offset); 89 drm_printf(p, " size: %llu\n", man->size); 90 drm_printf(p, " available_caching: 0x%08X\n", man->available_caching); 91 drm_printf(p, " default_caching: 0x%08X\n", man->default_caching); 92 if (mem_type != TTM_PL_SYSTEM) 93 (*man->func->debug)(man, p); 94 } 95 96 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo, 97 struct ttm_placement *placement) 98 { 99 struct drm_printer p = drm_debug_printer(TTM_PFX); 100 int i, ret, mem_type; 101 102 drm_printf(&p, "No space for %p (%lu pages, %luK, %luM)\n", 103 bo, bo->mem.num_pages, bo->mem.size >> 10, 104 bo->mem.size >> 20); 105 for (i = 0; i < placement->num_placement; i++) { 106 ret = ttm_mem_type_from_place(&placement->placement[i], 107 &mem_type); 108 if (ret) 109 return; 110 drm_printf(&p, " placement[%d]=0x%08X (%d)\n", 111 i, placement->placement[i].flags, mem_type); 112 ttm_mem_type_debug(bo->bdev, &p, mem_type); 113 } 114 } 115 116 static ssize_t ttm_bo_global_show(struct kobject *kobj, 117 struct attribute *attr, 118 char *buffer) 119 { 120 struct ttm_bo_global *glob = 121 container_of(kobj, struct ttm_bo_global, kobj); 122 123 return snprintf(buffer, PAGE_SIZE, "%d\n", 124 atomic_read(&glob->bo_count)); 125 } 126 127 static struct attribute *ttm_bo_global_attrs[] = { 128 &ttm_bo_count, 129 NULL 130 }; 131 132 static const struct sysfs_ops ttm_bo_global_ops = { 133 .show = &ttm_bo_global_show 134 }; 135 136 static struct kobj_type ttm_bo_glob_kobj_type = { 137 .release = &ttm_bo_global_kobj_release, 138 .sysfs_ops = &ttm_bo_global_ops, 139 .default_attrs = ttm_bo_global_attrs 140 }; 141 142 143 static inline uint32_t ttm_bo_type_flags(unsigned type) 144 { 145 return 1 << (type); 146 } 147 148 static void ttm_bo_release_list(struct kref *list_kref) 149 { 150 struct ttm_buffer_object *bo = 151 container_of(list_kref, struct ttm_buffer_object, list_kref); 152 struct ttm_bo_device *bdev = bo->bdev; 153 size_t acc_size = bo->acc_size; 154 155 BUG_ON(kref_read(&bo->list_kref)); 156 BUG_ON(kref_read(&bo->kref)); 157 BUG_ON(atomic_read(&bo->cpu_writers)); 158 BUG_ON(bo->mem.mm_node != NULL); 159 BUG_ON(!list_empty(&bo->lru)); 160 BUG_ON(!list_empty(&bo->ddestroy)); 161 ttm_tt_destroy(bo->ttm); 162 atomic_dec(&bo->bdev->glob->bo_count); 163 dma_fence_put(bo->moving); 164 reservation_object_fini(&bo->ttm_resv); 165 mutex_destroy(&bo->wu_mutex); 166 bo->destroy(bo); 167 ttm_mem_global_free(bdev->glob->mem_glob, acc_size); 168 } 169 170 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo) 171 { 172 struct ttm_bo_device *bdev = bo->bdev; 173 struct ttm_mem_type_manager *man; 174 175 reservation_object_assert_held(bo->resv); 176 177 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 178 BUG_ON(!list_empty(&bo->lru)); 179 180 man = &bdev->man[bo->mem.mem_type]; 181 list_add_tail(&bo->lru, &man->lru[bo->priority]); 182 kref_get(&bo->list_kref); 183 184 if (bo->ttm && !(bo->ttm->page_flags & 185 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED))) { 186 list_add_tail(&bo->swap, 187 &bdev->glob->swap_lru[bo->priority]); 188 kref_get(&bo->list_kref); 189 } 190 } 191 } 192 EXPORT_SYMBOL(ttm_bo_add_to_lru); 193 194 static void ttm_bo_ref_bug(struct kref *list_kref) 195 { 196 BUG(); 197 } 198 199 void ttm_bo_del_from_lru(struct ttm_buffer_object *bo) 200 { 201 struct ttm_bo_device *bdev = bo->bdev; 202 bool notify = false; 203 204 if (!list_empty(&bo->swap)) { 205 list_del_init(&bo->swap); 206 kref_put(&bo->list_kref, ttm_bo_ref_bug); 207 notify = true; 208 } 209 if (!list_empty(&bo->lru)) { 210 list_del_init(&bo->lru); 211 kref_put(&bo->list_kref, ttm_bo_ref_bug); 212 notify = true; 213 } 214 215 if (notify && bdev->driver->del_from_lru_notify) 216 bdev->driver->del_from_lru_notify(bo); 217 } 218 219 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo) 220 { 221 struct ttm_bo_global *glob = bo->bdev->glob; 222 223 spin_lock(&glob->lru_lock); 224 ttm_bo_del_from_lru(bo); 225 spin_unlock(&glob->lru_lock); 226 } 227 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru); 228 229 static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos, 230 struct ttm_buffer_object *bo) 231 { 232 if (!pos->first) 233 pos->first = bo; 234 pos->last = bo; 235 } 236 237 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo, 238 struct ttm_lru_bulk_move *bulk) 239 { 240 reservation_object_assert_held(bo->resv); 241 242 ttm_bo_del_from_lru(bo); 243 ttm_bo_add_to_lru(bo); 244 245 if (bulk && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 246 switch (bo->mem.mem_type) { 247 case TTM_PL_TT: 248 ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo); 249 break; 250 251 case TTM_PL_VRAM: 252 ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo); 253 break; 254 } 255 if (bo->ttm && !(bo->ttm->page_flags & 256 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED))) 257 ttm_bo_bulk_move_set_pos(&bulk->swap[bo->priority], bo); 258 } 259 } 260 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail); 261 262 void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk) 263 { 264 unsigned i; 265 266 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 267 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i]; 268 struct ttm_mem_type_manager *man; 269 270 if (!pos->first) 271 continue; 272 273 reservation_object_assert_held(pos->first->resv); 274 reservation_object_assert_held(pos->last->resv); 275 276 man = &pos->first->bdev->man[TTM_PL_TT]; 277 list_bulk_move_tail(&man->lru[i], &pos->first->lru, 278 &pos->last->lru); 279 } 280 281 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 282 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i]; 283 struct ttm_mem_type_manager *man; 284 285 if (!pos->first) 286 continue; 287 288 reservation_object_assert_held(pos->first->resv); 289 reservation_object_assert_held(pos->last->resv); 290 291 man = &pos->first->bdev->man[TTM_PL_VRAM]; 292 list_bulk_move_tail(&man->lru[i], &pos->first->lru, 293 &pos->last->lru); 294 } 295 296 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 297 struct ttm_lru_bulk_move_pos *pos = &bulk->swap[i]; 298 struct list_head *lru; 299 300 if (!pos->first) 301 continue; 302 303 reservation_object_assert_held(pos->first->resv); 304 reservation_object_assert_held(pos->last->resv); 305 306 lru = &pos->first->bdev->glob->swap_lru[i]; 307 list_bulk_move_tail(lru, &pos->first->swap, &pos->last->swap); 308 } 309 } 310 EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail); 311 312 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo, 313 struct ttm_mem_reg *mem, bool evict, 314 struct ttm_operation_ctx *ctx) 315 { 316 struct ttm_bo_device *bdev = bo->bdev; 317 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem); 318 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem); 319 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type]; 320 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type]; 321 int ret = 0; 322 323 if (old_is_pci || new_is_pci || 324 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) { 325 ret = ttm_mem_io_lock(old_man, true); 326 if (unlikely(ret != 0)) 327 goto out_err; 328 ttm_bo_unmap_virtual_locked(bo); 329 ttm_mem_io_unlock(old_man); 330 } 331 332 /* 333 * Create and bind a ttm if required. 334 */ 335 336 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 337 if (bo->ttm == NULL) { 338 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED); 339 ret = ttm_tt_create(bo, zero); 340 if (ret) 341 goto out_err; 342 } 343 344 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement); 345 if (ret) 346 goto out_err; 347 348 if (mem->mem_type != TTM_PL_SYSTEM) { 349 ret = ttm_tt_bind(bo->ttm, mem, ctx); 350 if (ret) 351 goto out_err; 352 } 353 354 if (bo->mem.mem_type == TTM_PL_SYSTEM) { 355 if (bdev->driver->move_notify) 356 bdev->driver->move_notify(bo, evict, mem); 357 bo->mem = *mem; 358 mem->mm_node = NULL; 359 goto moved; 360 } 361 } 362 363 if (bdev->driver->move_notify) 364 bdev->driver->move_notify(bo, evict, mem); 365 366 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) && 367 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) 368 ret = ttm_bo_move_ttm(bo, ctx, mem); 369 else if (bdev->driver->move) 370 ret = bdev->driver->move(bo, evict, ctx, mem); 371 else 372 ret = ttm_bo_move_memcpy(bo, ctx, mem); 373 374 if (ret) { 375 if (bdev->driver->move_notify) { 376 swap(*mem, bo->mem); 377 bdev->driver->move_notify(bo, false, mem); 378 swap(*mem, bo->mem); 379 } 380 381 goto out_err; 382 } 383 384 moved: 385 if (bo->evicted) { 386 if (bdev->driver->invalidate_caches) { 387 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement); 388 if (ret) 389 pr_err("Can not flush read caches\n"); 390 } 391 bo->evicted = false; 392 } 393 394 if (bo->mem.mm_node) 395 bo->offset = (bo->mem.start << PAGE_SHIFT) + 396 bdev->man[bo->mem.mem_type].gpu_offset; 397 else 398 bo->offset = 0; 399 400 ctx->bytes_moved += bo->num_pages << PAGE_SHIFT; 401 return 0; 402 403 out_err: 404 new_man = &bdev->man[bo->mem.mem_type]; 405 if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) { 406 ttm_tt_destroy(bo->ttm); 407 bo->ttm = NULL; 408 } 409 410 return ret; 411 } 412 413 /** 414 * Call bo::reserved. 415 * Will release GPU memory type usage on destruction. 416 * This is the place to put in driver specific hooks to release 417 * driver private resources. 418 * Will release the bo::reserved lock. 419 */ 420 421 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo) 422 { 423 if (bo->bdev->driver->move_notify) 424 bo->bdev->driver->move_notify(bo, false, NULL); 425 426 ttm_tt_destroy(bo->ttm); 427 bo->ttm = NULL; 428 ttm_bo_mem_put(bo, &bo->mem); 429 } 430 431 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo) 432 { 433 int r; 434 435 if (bo->resv == &bo->ttm_resv) 436 return 0; 437 438 BUG_ON(!reservation_object_trylock(&bo->ttm_resv)); 439 440 r = reservation_object_copy_fences(&bo->ttm_resv, bo->resv); 441 if (r) 442 reservation_object_unlock(&bo->ttm_resv); 443 444 return r; 445 } 446 447 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo) 448 { 449 struct reservation_object_list *fobj; 450 struct dma_fence *fence; 451 int i; 452 453 fobj = reservation_object_get_list(&bo->ttm_resv); 454 fence = reservation_object_get_excl(&bo->ttm_resv); 455 if (fence && !fence->ops->signaled) 456 dma_fence_enable_sw_signaling(fence); 457 458 for (i = 0; fobj && i < fobj->shared_count; ++i) { 459 fence = rcu_dereference_protected(fobj->shared[i], 460 reservation_object_held(bo->resv)); 461 462 if (!fence->ops->signaled) 463 dma_fence_enable_sw_signaling(fence); 464 } 465 } 466 467 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo) 468 { 469 struct ttm_bo_device *bdev = bo->bdev; 470 struct ttm_bo_global *glob = bdev->glob; 471 int ret; 472 473 ret = ttm_bo_individualize_resv(bo); 474 if (ret) { 475 /* Last resort, if we fail to allocate memory for the 476 * fences block for the BO to become idle 477 */ 478 reservation_object_wait_timeout_rcu(bo->resv, true, false, 479 30 * HZ); 480 spin_lock(&glob->lru_lock); 481 goto error; 482 } 483 484 spin_lock(&glob->lru_lock); 485 ret = reservation_object_trylock(bo->resv) ? 0 : -EBUSY; 486 if (!ret) { 487 if (reservation_object_test_signaled_rcu(&bo->ttm_resv, true)) { 488 ttm_bo_del_from_lru(bo); 489 spin_unlock(&glob->lru_lock); 490 if (bo->resv != &bo->ttm_resv) 491 reservation_object_unlock(&bo->ttm_resv); 492 493 ttm_bo_cleanup_memtype_use(bo); 494 reservation_object_unlock(bo->resv); 495 return; 496 } 497 498 ttm_bo_flush_all_fences(bo); 499 500 /* 501 * Make NO_EVICT bos immediately available to 502 * shrinkers, now that they are queued for 503 * destruction. 504 */ 505 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) { 506 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT; 507 ttm_bo_add_to_lru(bo); 508 } 509 510 reservation_object_unlock(bo->resv); 511 } 512 if (bo->resv != &bo->ttm_resv) 513 reservation_object_unlock(&bo->ttm_resv); 514 515 error: 516 kref_get(&bo->list_kref); 517 list_add_tail(&bo->ddestroy, &bdev->ddestroy); 518 spin_unlock(&glob->lru_lock); 519 520 schedule_delayed_work(&bdev->wq, 521 ((HZ / 100) < 1) ? 1 : HZ / 100); 522 } 523 524 /** 525 * function ttm_bo_cleanup_refs 526 * If bo idle, remove from delayed- and lru lists, and unref. 527 * If not idle, do nothing. 528 * 529 * Must be called with lru_lock and reservation held, this function 530 * will drop the lru lock and optionally the reservation lock before returning. 531 * 532 * @interruptible Any sleeps should occur interruptibly. 533 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead. 534 * @unlock_resv Unlock the reservation lock as well. 535 */ 536 537 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo, 538 bool interruptible, bool no_wait_gpu, 539 bool unlock_resv) 540 { 541 struct ttm_bo_global *glob = bo->bdev->glob; 542 struct reservation_object *resv; 543 int ret; 544 545 if (unlikely(list_empty(&bo->ddestroy))) 546 resv = bo->resv; 547 else 548 resv = &bo->ttm_resv; 549 550 if (reservation_object_test_signaled_rcu(resv, true)) 551 ret = 0; 552 else 553 ret = -EBUSY; 554 555 if (ret && !no_wait_gpu) { 556 long lret; 557 558 if (unlock_resv) 559 reservation_object_unlock(bo->resv); 560 spin_unlock(&glob->lru_lock); 561 562 lret = reservation_object_wait_timeout_rcu(resv, true, 563 interruptible, 564 30 * HZ); 565 566 if (lret < 0) 567 return lret; 568 else if (lret == 0) 569 return -EBUSY; 570 571 spin_lock(&glob->lru_lock); 572 if (unlock_resv && !reservation_object_trylock(bo->resv)) { 573 /* 574 * We raced, and lost, someone else holds the reservation now, 575 * and is probably busy in ttm_bo_cleanup_memtype_use. 576 * 577 * Even if it's not the case, because we finished waiting any 578 * delayed destruction would succeed, so just return success 579 * here. 580 */ 581 spin_unlock(&glob->lru_lock); 582 return 0; 583 } 584 ret = 0; 585 } 586 587 if (ret || unlikely(list_empty(&bo->ddestroy))) { 588 if (unlock_resv) 589 reservation_object_unlock(bo->resv); 590 spin_unlock(&glob->lru_lock); 591 return ret; 592 } 593 594 ttm_bo_del_from_lru(bo); 595 list_del_init(&bo->ddestroy); 596 kref_put(&bo->list_kref, ttm_bo_ref_bug); 597 598 spin_unlock(&glob->lru_lock); 599 ttm_bo_cleanup_memtype_use(bo); 600 601 if (unlock_resv) 602 reservation_object_unlock(bo->resv); 603 604 return 0; 605 } 606 607 /** 608 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all 609 * encountered buffers. 610 */ 611 static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all) 612 { 613 struct ttm_bo_global *glob = bdev->glob; 614 struct list_head removed; 615 bool empty; 616 617 INIT_LIST_HEAD(&removed); 618 619 spin_lock(&glob->lru_lock); 620 while (!list_empty(&bdev->ddestroy)) { 621 struct ttm_buffer_object *bo; 622 623 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object, 624 ddestroy); 625 kref_get(&bo->list_kref); 626 list_move_tail(&bo->ddestroy, &removed); 627 628 if (remove_all || bo->resv != &bo->ttm_resv) { 629 spin_unlock(&glob->lru_lock); 630 reservation_object_lock(bo->resv, NULL); 631 632 spin_lock(&glob->lru_lock); 633 ttm_bo_cleanup_refs(bo, false, !remove_all, true); 634 635 } else if (reservation_object_trylock(bo->resv)) { 636 ttm_bo_cleanup_refs(bo, false, !remove_all, true); 637 } else { 638 spin_unlock(&glob->lru_lock); 639 } 640 641 kref_put(&bo->list_kref, ttm_bo_release_list); 642 spin_lock(&glob->lru_lock); 643 } 644 list_splice_tail(&removed, &bdev->ddestroy); 645 empty = list_empty(&bdev->ddestroy); 646 spin_unlock(&glob->lru_lock); 647 648 return empty; 649 } 650 651 static void ttm_bo_delayed_workqueue(struct work_struct *work) 652 { 653 struct ttm_bo_device *bdev = 654 container_of(work, struct ttm_bo_device, wq.work); 655 656 if (!ttm_bo_delayed_delete(bdev, false)) 657 schedule_delayed_work(&bdev->wq, 658 ((HZ / 100) < 1) ? 1 : HZ / 100); 659 } 660 661 static void ttm_bo_release(struct kref *kref) 662 { 663 struct ttm_buffer_object *bo = 664 container_of(kref, struct ttm_buffer_object, kref); 665 struct ttm_bo_device *bdev = bo->bdev; 666 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 667 668 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node); 669 ttm_mem_io_lock(man, false); 670 ttm_mem_io_free_vm(bo); 671 ttm_mem_io_unlock(man); 672 ttm_bo_cleanup_refs_or_queue(bo); 673 kref_put(&bo->list_kref, ttm_bo_release_list); 674 } 675 676 void ttm_bo_put(struct ttm_buffer_object *bo) 677 { 678 kref_put(&bo->kref, ttm_bo_release); 679 } 680 EXPORT_SYMBOL(ttm_bo_put); 681 682 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) 683 { 684 return cancel_delayed_work_sync(&bdev->wq); 685 } 686 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue); 687 688 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched) 689 { 690 if (resched) 691 schedule_delayed_work(&bdev->wq, 692 ((HZ / 100) < 1) ? 1 : HZ / 100); 693 } 694 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue); 695 696 static int ttm_bo_evict(struct ttm_buffer_object *bo, 697 struct ttm_operation_ctx *ctx) 698 { 699 struct ttm_bo_device *bdev = bo->bdev; 700 struct ttm_mem_reg evict_mem; 701 struct ttm_placement placement; 702 int ret = 0; 703 704 reservation_object_assert_held(bo->resv); 705 706 placement.num_placement = 0; 707 placement.num_busy_placement = 0; 708 bdev->driver->evict_flags(bo, &placement); 709 710 if (!placement.num_placement && !placement.num_busy_placement) { 711 ret = ttm_bo_pipeline_gutting(bo); 712 if (ret) 713 return ret; 714 715 return ttm_tt_create(bo, false); 716 } 717 718 evict_mem = bo->mem; 719 evict_mem.mm_node = NULL; 720 evict_mem.bus.io_reserved_vm = false; 721 evict_mem.bus.io_reserved_count = 0; 722 723 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx); 724 if (ret) { 725 if (ret != -ERESTARTSYS) { 726 pr_err("Failed to find memory space for buffer 0x%p eviction\n", 727 bo); 728 ttm_bo_mem_space_debug(bo, &placement); 729 } 730 goto out; 731 } 732 733 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx); 734 if (unlikely(ret)) { 735 if (ret != -ERESTARTSYS) 736 pr_err("Buffer eviction failed\n"); 737 ttm_bo_mem_put(bo, &evict_mem); 738 goto out; 739 } 740 bo->evicted = true; 741 out: 742 return ret; 743 } 744 745 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo, 746 const struct ttm_place *place) 747 { 748 /* Don't evict this BO if it's outside of the 749 * requested placement range 750 */ 751 if (place->fpfn >= (bo->mem.start + bo->mem.size) || 752 (place->lpfn && place->lpfn <= bo->mem.start)) 753 return false; 754 755 return true; 756 } 757 EXPORT_SYMBOL(ttm_bo_eviction_valuable); 758 759 /** 760 * Check the target bo is allowable to be evicted or swapout, including cases: 761 * 762 * a. if share same reservation object with ctx->resv, have assumption 763 * reservation objects should already be locked, so not lock again and 764 * return true directly when either the opreation allow_reserved_eviction 765 * or the target bo already is in delayed free list; 766 * 767 * b. Otherwise, trylock it. 768 */ 769 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo, 770 struct ttm_operation_ctx *ctx, bool *locked) 771 { 772 bool ret = false; 773 774 *locked = false; 775 if (bo->resv == ctx->resv) { 776 reservation_object_assert_held(bo->resv); 777 if (ctx->flags & TTM_OPT_FLAG_ALLOW_RES_EVICT 778 || !list_empty(&bo->ddestroy)) 779 ret = true; 780 } else { 781 *locked = reservation_object_trylock(bo->resv); 782 ret = *locked; 783 } 784 785 return ret; 786 } 787 788 static int ttm_mem_evict_first(struct ttm_bo_device *bdev, 789 uint32_t mem_type, 790 const struct ttm_place *place, 791 struct ttm_operation_ctx *ctx) 792 { 793 struct ttm_bo_global *glob = bdev->glob; 794 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 795 struct ttm_buffer_object *bo = NULL; 796 bool locked = false; 797 unsigned i; 798 int ret; 799 800 spin_lock(&glob->lru_lock); 801 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 802 list_for_each_entry(bo, &man->lru[i], lru) { 803 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked)) 804 continue; 805 806 if (place && !bdev->driver->eviction_valuable(bo, 807 place)) { 808 if (locked) 809 reservation_object_unlock(bo->resv); 810 continue; 811 } 812 break; 813 } 814 815 /* If the inner loop terminated early, we have our candidate */ 816 if (&bo->lru != &man->lru[i]) 817 break; 818 819 bo = NULL; 820 } 821 822 if (!bo) { 823 spin_unlock(&glob->lru_lock); 824 return -EBUSY; 825 } 826 827 kref_get(&bo->list_kref); 828 829 if (!list_empty(&bo->ddestroy)) { 830 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible, 831 ctx->no_wait_gpu, locked); 832 kref_put(&bo->list_kref, ttm_bo_release_list); 833 return ret; 834 } 835 836 ttm_bo_del_from_lru(bo); 837 spin_unlock(&glob->lru_lock); 838 839 ret = ttm_bo_evict(bo, ctx); 840 if (locked) { 841 ttm_bo_unreserve(bo); 842 } else { 843 spin_lock(&glob->lru_lock); 844 ttm_bo_add_to_lru(bo); 845 spin_unlock(&glob->lru_lock); 846 } 847 848 kref_put(&bo->list_kref, ttm_bo_release_list); 849 return ret; 850 } 851 852 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem) 853 { 854 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type]; 855 856 if (mem->mm_node) 857 (*man->func->put_node)(man, mem); 858 } 859 EXPORT_SYMBOL(ttm_bo_mem_put); 860 861 /** 862 * Add the last move fence to the BO and reserve a new shared slot. 863 */ 864 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo, 865 struct ttm_mem_type_manager *man, 866 struct ttm_mem_reg *mem) 867 { 868 struct dma_fence *fence; 869 int ret; 870 871 spin_lock(&man->move_lock); 872 fence = dma_fence_get(man->move); 873 spin_unlock(&man->move_lock); 874 875 if (fence) { 876 reservation_object_add_shared_fence(bo->resv, fence); 877 878 ret = reservation_object_reserve_shared(bo->resv, 1); 879 if (unlikely(ret)) { 880 dma_fence_put(fence); 881 return ret; 882 } 883 884 dma_fence_put(bo->moving); 885 bo->moving = fence; 886 } 887 888 return 0; 889 } 890 891 /** 892 * Repeatedly evict memory from the LRU for @mem_type until we create enough 893 * space, or we've evicted everything and there isn't enough space. 894 */ 895 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo, 896 uint32_t mem_type, 897 const struct ttm_place *place, 898 struct ttm_mem_reg *mem, 899 struct ttm_operation_ctx *ctx) 900 { 901 struct ttm_bo_device *bdev = bo->bdev; 902 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 903 int ret; 904 905 do { 906 ret = (*man->func->get_node)(man, bo, place, mem); 907 if (unlikely(ret != 0)) 908 return ret; 909 if (mem->mm_node) 910 break; 911 ret = ttm_mem_evict_first(bdev, mem_type, place, ctx); 912 if (unlikely(ret != 0)) 913 return ret; 914 } while (1); 915 mem->mem_type = mem_type; 916 return ttm_bo_add_move_fence(bo, man, mem); 917 } 918 919 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man, 920 uint32_t cur_placement, 921 uint32_t proposed_placement) 922 { 923 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING; 924 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING; 925 926 /** 927 * Keep current caching if possible. 928 */ 929 930 if ((cur_placement & caching) != 0) 931 result |= (cur_placement & caching); 932 else if ((man->default_caching & caching) != 0) 933 result |= man->default_caching; 934 else if ((TTM_PL_FLAG_CACHED & caching) != 0) 935 result |= TTM_PL_FLAG_CACHED; 936 else if ((TTM_PL_FLAG_WC & caching) != 0) 937 result |= TTM_PL_FLAG_WC; 938 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0) 939 result |= TTM_PL_FLAG_UNCACHED; 940 941 return result; 942 } 943 944 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man, 945 uint32_t mem_type, 946 const struct ttm_place *place, 947 uint32_t *masked_placement) 948 { 949 uint32_t cur_flags = ttm_bo_type_flags(mem_type); 950 951 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0) 952 return false; 953 954 if ((place->flags & man->available_caching) == 0) 955 return false; 956 957 cur_flags |= (place->flags & man->available_caching); 958 959 *masked_placement = cur_flags; 960 return true; 961 } 962 963 /** 964 * Creates space for memory region @mem according to its type. 965 * 966 * This function first searches for free space in compatible memory types in 967 * the priority order defined by the driver. If free space isn't found, then 968 * ttm_bo_mem_force_space is attempted in priority order to evict and find 969 * space. 970 */ 971 int ttm_bo_mem_space(struct ttm_buffer_object *bo, 972 struct ttm_placement *placement, 973 struct ttm_mem_reg *mem, 974 struct ttm_operation_ctx *ctx) 975 { 976 struct ttm_bo_device *bdev = bo->bdev; 977 struct ttm_mem_type_manager *man; 978 uint32_t mem_type = TTM_PL_SYSTEM; 979 uint32_t cur_flags = 0; 980 bool type_found = false; 981 bool type_ok = false; 982 bool has_erestartsys = false; 983 int i, ret; 984 985 ret = reservation_object_reserve_shared(bo->resv, 1); 986 if (unlikely(ret)) 987 return ret; 988 989 mem->mm_node = NULL; 990 for (i = 0; i < placement->num_placement; ++i) { 991 const struct ttm_place *place = &placement->placement[i]; 992 993 ret = ttm_mem_type_from_place(place, &mem_type); 994 if (ret) 995 return ret; 996 man = &bdev->man[mem_type]; 997 if (!man->has_type || !man->use_type) 998 continue; 999 1000 type_ok = ttm_bo_mt_compatible(man, mem_type, place, 1001 &cur_flags); 1002 1003 if (!type_ok) 1004 continue; 1005 1006 type_found = true; 1007 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1008 cur_flags); 1009 /* 1010 * Use the access and other non-mapping-related flag bits from 1011 * the memory placement flags to the current flags 1012 */ 1013 ttm_flag_masked(&cur_flags, place->flags, 1014 ~TTM_PL_MASK_MEMTYPE); 1015 1016 if (mem_type == TTM_PL_SYSTEM) 1017 break; 1018 1019 ret = (*man->func->get_node)(man, bo, place, mem); 1020 if (unlikely(ret)) 1021 return ret; 1022 1023 if (mem->mm_node) { 1024 ret = ttm_bo_add_move_fence(bo, man, mem); 1025 if (unlikely(ret)) { 1026 (*man->func->put_node)(man, mem); 1027 return ret; 1028 } 1029 break; 1030 } 1031 } 1032 1033 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) { 1034 mem->mem_type = mem_type; 1035 mem->placement = cur_flags; 1036 return 0; 1037 } 1038 1039 for (i = 0; i < placement->num_busy_placement; ++i) { 1040 const struct ttm_place *place = &placement->busy_placement[i]; 1041 1042 ret = ttm_mem_type_from_place(place, &mem_type); 1043 if (ret) 1044 return ret; 1045 man = &bdev->man[mem_type]; 1046 if (!man->has_type || !man->use_type) 1047 continue; 1048 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags)) 1049 continue; 1050 1051 type_found = true; 1052 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1053 cur_flags); 1054 /* 1055 * Use the access and other non-mapping-related flag bits from 1056 * the memory placement flags to the current flags 1057 */ 1058 ttm_flag_masked(&cur_flags, place->flags, 1059 ~TTM_PL_MASK_MEMTYPE); 1060 1061 if (mem_type == TTM_PL_SYSTEM) { 1062 mem->mem_type = mem_type; 1063 mem->placement = cur_flags; 1064 mem->mm_node = NULL; 1065 return 0; 1066 } 1067 1068 ret = ttm_bo_mem_force_space(bo, mem_type, place, mem, ctx); 1069 if (ret == 0 && mem->mm_node) { 1070 mem->placement = cur_flags; 1071 return 0; 1072 } 1073 if (ret == -ERESTARTSYS) 1074 has_erestartsys = true; 1075 } 1076 1077 if (!type_found) { 1078 pr_err(TTM_PFX "No compatible memory type found\n"); 1079 return -EINVAL; 1080 } 1081 1082 return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM; 1083 } 1084 EXPORT_SYMBOL(ttm_bo_mem_space); 1085 1086 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo, 1087 struct ttm_placement *placement, 1088 struct ttm_operation_ctx *ctx) 1089 { 1090 int ret = 0; 1091 struct ttm_mem_reg mem; 1092 1093 reservation_object_assert_held(bo->resv); 1094 1095 mem.num_pages = bo->num_pages; 1096 mem.size = mem.num_pages << PAGE_SHIFT; 1097 mem.page_alignment = bo->mem.page_alignment; 1098 mem.bus.io_reserved_vm = false; 1099 mem.bus.io_reserved_count = 0; 1100 /* 1101 * Determine where to move the buffer. 1102 */ 1103 ret = ttm_bo_mem_space(bo, placement, &mem, ctx); 1104 if (ret) 1105 goto out_unlock; 1106 ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx); 1107 out_unlock: 1108 if (ret && mem.mm_node) 1109 ttm_bo_mem_put(bo, &mem); 1110 return ret; 1111 } 1112 1113 static bool ttm_bo_places_compat(const struct ttm_place *places, 1114 unsigned num_placement, 1115 struct ttm_mem_reg *mem, 1116 uint32_t *new_flags) 1117 { 1118 unsigned i; 1119 1120 for (i = 0; i < num_placement; i++) { 1121 const struct ttm_place *heap = &places[i]; 1122 1123 if (mem->mm_node && (mem->start < heap->fpfn || 1124 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn))) 1125 continue; 1126 1127 *new_flags = heap->flags; 1128 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) && 1129 (*new_flags & mem->placement & TTM_PL_MASK_MEM) && 1130 (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) || 1131 (mem->placement & TTM_PL_FLAG_CONTIGUOUS))) 1132 return true; 1133 } 1134 return false; 1135 } 1136 1137 bool ttm_bo_mem_compat(struct ttm_placement *placement, 1138 struct ttm_mem_reg *mem, 1139 uint32_t *new_flags) 1140 { 1141 if (ttm_bo_places_compat(placement->placement, placement->num_placement, 1142 mem, new_flags)) 1143 return true; 1144 1145 if ((placement->busy_placement != placement->placement || 1146 placement->num_busy_placement > placement->num_placement) && 1147 ttm_bo_places_compat(placement->busy_placement, 1148 placement->num_busy_placement, 1149 mem, new_flags)) 1150 return true; 1151 1152 return false; 1153 } 1154 EXPORT_SYMBOL(ttm_bo_mem_compat); 1155 1156 int ttm_bo_validate(struct ttm_buffer_object *bo, 1157 struct ttm_placement *placement, 1158 struct ttm_operation_ctx *ctx) 1159 { 1160 int ret; 1161 uint32_t new_flags; 1162 1163 reservation_object_assert_held(bo->resv); 1164 /* 1165 * Check whether we need to move buffer. 1166 */ 1167 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) { 1168 ret = ttm_bo_move_buffer(bo, placement, ctx); 1169 if (ret) 1170 return ret; 1171 } else { 1172 /* 1173 * Use the access and other non-mapping-related flag bits from 1174 * the compatible memory placement flags to the active flags 1175 */ 1176 ttm_flag_masked(&bo->mem.placement, new_flags, 1177 ~TTM_PL_MASK_MEMTYPE); 1178 } 1179 /* 1180 * We might need to add a TTM. 1181 */ 1182 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 1183 ret = ttm_tt_create(bo, true); 1184 if (ret) 1185 return ret; 1186 } 1187 return 0; 1188 } 1189 EXPORT_SYMBOL(ttm_bo_validate); 1190 1191 int ttm_bo_init_reserved(struct ttm_bo_device *bdev, 1192 struct ttm_buffer_object *bo, 1193 unsigned long size, 1194 enum ttm_bo_type type, 1195 struct ttm_placement *placement, 1196 uint32_t page_alignment, 1197 struct ttm_operation_ctx *ctx, 1198 size_t acc_size, 1199 struct sg_table *sg, 1200 struct reservation_object *resv, 1201 void (*destroy) (struct ttm_buffer_object *)) 1202 { 1203 int ret = 0; 1204 unsigned long num_pages; 1205 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob; 1206 bool locked; 1207 1208 ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx); 1209 if (ret) { 1210 pr_err("Out of kernel memory\n"); 1211 if (destroy) 1212 (*destroy)(bo); 1213 else 1214 kfree(bo); 1215 return -ENOMEM; 1216 } 1217 1218 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 1219 if (num_pages == 0) { 1220 pr_err("Illegal buffer object size\n"); 1221 if (destroy) 1222 (*destroy)(bo); 1223 else 1224 kfree(bo); 1225 ttm_mem_global_free(mem_glob, acc_size); 1226 return -EINVAL; 1227 } 1228 bo->destroy = destroy ? destroy : ttm_bo_default_destroy; 1229 1230 kref_init(&bo->kref); 1231 kref_init(&bo->list_kref); 1232 atomic_set(&bo->cpu_writers, 0); 1233 INIT_LIST_HEAD(&bo->lru); 1234 INIT_LIST_HEAD(&bo->ddestroy); 1235 INIT_LIST_HEAD(&bo->swap); 1236 INIT_LIST_HEAD(&bo->io_reserve_lru); 1237 mutex_init(&bo->wu_mutex); 1238 bo->bdev = bdev; 1239 bo->type = type; 1240 bo->num_pages = num_pages; 1241 bo->mem.size = num_pages << PAGE_SHIFT; 1242 bo->mem.mem_type = TTM_PL_SYSTEM; 1243 bo->mem.num_pages = bo->num_pages; 1244 bo->mem.mm_node = NULL; 1245 bo->mem.page_alignment = page_alignment; 1246 bo->mem.bus.io_reserved_vm = false; 1247 bo->mem.bus.io_reserved_count = 0; 1248 bo->moving = NULL; 1249 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED); 1250 bo->acc_size = acc_size; 1251 bo->sg = sg; 1252 if (resv) { 1253 bo->resv = resv; 1254 reservation_object_assert_held(bo->resv); 1255 } else { 1256 bo->resv = &bo->ttm_resv; 1257 } 1258 reservation_object_init(&bo->ttm_resv); 1259 atomic_inc(&bo->bdev->glob->bo_count); 1260 drm_vma_node_reset(&bo->vma_node); 1261 1262 /* 1263 * For ttm_bo_type_device buffers, allocate 1264 * address space from the device. 1265 */ 1266 if (bo->type == ttm_bo_type_device || 1267 bo->type == ttm_bo_type_sg) 1268 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node, 1269 bo->mem.num_pages); 1270 1271 /* passed reservation objects should already be locked, 1272 * since otherwise lockdep will be angered in radeon. 1273 */ 1274 if (!resv) { 1275 locked = reservation_object_trylock(bo->resv); 1276 WARN_ON(!locked); 1277 } 1278 1279 if (likely(!ret)) 1280 ret = ttm_bo_validate(bo, placement, ctx); 1281 1282 if (unlikely(ret)) { 1283 if (!resv) 1284 ttm_bo_unreserve(bo); 1285 1286 ttm_bo_put(bo); 1287 return ret; 1288 } 1289 1290 if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 1291 spin_lock(&bdev->glob->lru_lock); 1292 ttm_bo_add_to_lru(bo); 1293 spin_unlock(&bdev->glob->lru_lock); 1294 } 1295 1296 return ret; 1297 } 1298 EXPORT_SYMBOL(ttm_bo_init_reserved); 1299 1300 int ttm_bo_init(struct ttm_bo_device *bdev, 1301 struct ttm_buffer_object *bo, 1302 unsigned long size, 1303 enum ttm_bo_type type, 1304 struct ttm_placement *placement, 1305 uint32_t page_alignment, 1306 bool interruptible, 1307 size_t acc_size, 1308 struct sg_table *sg, 1309 struct reservation_object *resv, 1310 void (*destroy) (struct ttm_buffer_object *)) 1311 { 1312 struct ttm_operation_ctx ctx = { interruptible, false }; 1313 int ret; 1314 1315 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement, 1316 page_alignment, &ctx, acc_size, 1317 sg, resv, destroy); 1318 if (ret) 1319 return ret; 1320 1321 if (!resv) 1322 ttm_bo_unreserve(bo); 1323 1324 return 0; 1325 } 1326 EXPORT_SYMBOL(ttm_bo_init); 1327 1328 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev, 1329 unsigned long bo_size, 1330 unsigned struct_size) 1331 { 1332 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1333 size_t size = 0; 1334 1335 size += ttm_round_pot(struct_size); 1336 size += ttm_round_pot(npages * sizeof(void *)); 1337 size += ttm_round_pot(sizeof(struct ttm_tt)); 1338 return size; 1339 } 1340 EXPORT_SYMBOL(ttm_bo_acc_size); 1341 1342 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev, 1343 unsigned long bo_size, 1344 unsigned struct_size) 1345 { 1346 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1347 size_t size = 0; 1348 1349 size += ttm_round_pot(struct_size); 1350 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t))); 1351 size += ttm_round_pot(sizeof(struct ttm_dma_tt)); 1352 return size; 1353 } 1354 EXPORT_SYMBOL(ttm_bo_dma_acc_size); 1355 1356 int ttm_bo_create(struct ttm_bo_device *bdev, 1357 unsigned long size, 1358 enum ttm_bo_type type, 1359 struct ttm_placement *placement, 1360 uint32_t page_alignment, 1361 bool interruptible, 1362 struct ttm_buffer_object **p_bo) 1363 { 1364 struct ttm_buffer_object *bo; 1365 size_t acc_size; 1366 int ret; 1367 1368 bo = kzalloc(sizeof(*bo), GFP_KERNEL); 1369 if (unlikely(bo == NULL)) 1370 return -ENOMEM; 1371 1372 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object)); 1373 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment, 1374 interruptible, acc_size, 1375 NULL, NULL, NULL); 1376 if (likely(ret == 0)) 1377 *p_bo = bo; 1378 1379 return ret; 1380 } 1381 EXPORT_SYMBOL(ttm_bo_create); 1382 1383 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev, 1384 unsigned mem_type) 1385 { 1386 struct ttm_operation_ctx ctx = { 1387 .interruptible = false, 1388 .no_wait_gpu = false, 1389 .flags = TTM_OPT_FLAG_FORCE_ALLOC 1390 }; 1391 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1392 struct ttm_bo_global *glob = bdev->glob; 1393 struct dma_fence *fence; 1394 int ret; 1395 unsigned i; 1396 1397 /* 1398 * Can't use standard list traversal since we're unlocking. 1399 */ 1400 1401 spin_lock(&glob->lru_lock); 1402 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 1403 while (!list_empty(&man->lru[i])) { 1404 spin_unlock(&glob->lru_lock); 1405 ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx); 1406 if (ret) 1407 return ret; 1408 spin_lock(&glob->lru_lock); 1409 } 1410 } 1411 spin_unlock(&glob->lru_lock); 1412 1413 spin_lock(&man->move_lock); 1414 fence = dma_fence_get(man->move); 1415 spin_unlock(&man->move_lock); 1416 1417 if (fence) { 1418 ret = dma_fence_wait(fence, false); 1419 dma_fence_put(fence); 1420 if (ret) 1421 return ret; 1422 } 1423 1424 return 0; 1425 } 1426 1427 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1428 { 1429 struct ttm_mem_type_manager *man; 1430 int ret = -EINVAL; 1431 1432 if (mem_type >= TTM_NUM_MEM_TYPES) { 1433 pr_err("Illegal memory type %d\n", mem_type); 1434 return ret; 1435 } 1436 man = &bdev->man[mem_type]; 1437 1438 if (!man->has_type) { 1439 pr_err("Trying to take down uninitialized memory manager type %u\n", 1440 mem_type); 1441 return ret; 1442 } 1443 1444 man->use_type = false; 1445 man->has_type = false; 1446 1447 ret = 0; 1448 if (mem_type > 0) { 1449 ret = ttm_bo_force_list_clean(bdev, mem_type); 1450 if (ret) { 1451 pr_err("Cleanup eviction failed\n"); 1452 return ret; 1453 } 1454 1455 ret = (*man->func->takedown)(man); 1456 } 1457 1458 dma_fence_put(man->move); 1459 man->move = NULL; 1460 1461 return ret; 1462 } 1463 EXPORT_SYMBOL(ttm_bo_clean_mm); 1464 1465 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1466 { 1467 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1468 1469 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) { 1470 pr_err("Illegal memory manager memory type %u\n", mem_type); 1471 return -EINVAL; 1472 } 1473 1474 if (!man->has_type) { 1475 pr_err("Memory type %u has not been initialized\n", mem_type); 1476 return 0; 1477 } 1478 1479 return ttm_bo_force_list_clean(bdev, mem_type); 1480 } 1481 EXPORT_SYMBOL(ttm_bo_evict_mm); 1482 1483 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type, 1484 unsigned long p_size) 1485 { 1486 int ret; 1487 struct ttm_mem_type_manager *man; 1488 unsigned i; 1489 1490 BUG_ON(type >= TTM_NUM_MEM_TYPES); 1491 man = &bdev->man[type]; 1492 BUG_ON(man->has_type); 1493 man->io_reserve_fastpath = true; 1494 man->use_io_reserve_lru = false; 1495 mutex_init(&man->io_reserve_mutex); 1496 spin_lock_init(&man->move_lock); 1497 INIT_LIST_HEAD(&man->io_reserve_lru); 1498 1499 ret = bdev->driver->init_mem_type(bdev, type, man); 1500 if (ret) 1501 return ret; 1502 man->bdev = bdev; 1503 1504 if (type != TTM_PL_SYSTEM) { 1505 ret = (*man->func->init)(man, p_size); 1506 if (ret) 1507 return ret; 1508 } 1509 man->has_type = true; 1510 man->use_type = true; 1511 man->size = p_size; 1512 1513 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1514 INIT_LIST_HEAD(&man->lru[i]); 1515 man->move = NULL; 1516 1517 return 0; 1518 } 1519 EXPORT_SYMBOL(ttm_bo_init_mm); 1520 1521 static void ttm_bo_global_kobj_release(struct kobject *kobj) 1522 { 1523 struct ttm_bo_global *glob = 1524 container_of(kobj, struct ttm_bo_global, kobj); 1525 1526 __free_page(glob->dummy_read_page); 1527 } 1528 1529 static void ttm_bo_global_release(void) 1530 { 1531 struct ttm_bo_global *glob = &ttm_bo_glob; 1532 1533 mutex_lock(&ttm_global_mutex); 1534 if (--glob->use_count > 0) 1535 goto out; 1536 1537 kobject_del(&glob->kobj); 1538 kobject_put(&glob->kobj); 1539 ttm_mem_global_release(&ttm_mem_glob); 1540 out: 1541 mutex_unlock(&ttm_global_mutex); 1542 } 1543 1544 static int ttm_bo_global_init(void) 1545 { 1546 struct ttm_bo_global *glob = &ttm_bo_glob; 1547 int ret = 0; 1548 unsigned i; 1549 1550 mutex_lock(&ttm_global_mutex); 1551 if (++glob->use_count > 1) 1552 goto out; 1553 1554 ret = ttm_mem_global_init(&ttm_mem_glob); 1555 if (ret) 1556 goto out; 1557 1558 spin_lock_init(&glob->lru_lock); 1559 glob->mem_glob = &ttm_mem_glob; 1560 glob->mem_glob->bo_glob = glob; 1561 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32); 1562 1563 if (unlikely(glob->dummy_read_page == NULL)) { 1564 ret = -ENOMEM; 1565 goto out; 1566 } 1567 1568 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1569 INIT_LIST_HEAD(&glob->swap_lru[i]); 1570 INIT_LIST_HEAD(&glob->device_list); 1571 atomic_set(&glob->bo_count, 0); 1572 1573 ret = kobject_init_and_add( 1574 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects"); 1575 if (unlikely(ret != 0)) 1576 kobject_put(&glob->kobj); 1577 out: 1578 mutex_unlock(&ttm_global_mutex); 1579 return ret; 1580 } 1581 1582 int ttm_bo_device_release(struct ttm_bo_device *bdev) 1583 { 1584 int ret = 0; 1585 unsigned i = TTM_NUM_MEM_TYPES; 1586 struct ttm_mem_type_manager *man; 1587 struct ttm_bo_global *glob = bdev->glob; 1588 1589 while (i--) { 1590 man = &bdev->man[i]; 1591 if (man->has_type) { 1592 man->use_type = false; 1593 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { 1594 ret = -EBUSY; 1595 pr_err("DRM memory manager type %d is not clean\n", 1596 i); 1597 } 1598 man->has_type = false; 1599 } 1600 } 1601 1602 mutex_lock(&ttm_global_mutex); 1603 list_del(&bdev->device_list); 1604 mutex_unlock(&ttm_global_mutex); 1605 1606 cancel_delayed_work_sync(&bdev->wq); 1607 1608 if (ttm_bo_delayed_delete(bdev, true)) 1609 pr_debug("Delayed destroy list was clean\n"); 1610 1611 spin_lock(&glob->lru_lock); 1612 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1613 if (list_empty(&bdev->man[0].lru[0])) 1614 pr_debug("Swap list %d was clean\n", i); 1615 spin_unlock(&glob->lru_lock); 1616 1617 drm_vma_offset_manager_destroy(&bdev->vma_manager); 1618 1619 if (!ret) 1620 ttm_bo_global_release(); 1621 1622 return ret; 1623 } 1624 EXPORT_SYMBOL(ttm_bo_device_release); 1625 1626 int ttm_bo_device_init(struct ttm_bo_device *bdev, 1627 struct ttm_bo_driver *driver, 1628 struct address_space *mapping, 1629 uint64_t file_page_offset, 1630 bool need_dma32) 1631 { 1632 struct ttm_bo_global *glob = &ttm_bo_glob; 1633 int ret; 1634 1635 ret = ttm_bo_global_init(); 1636 if (ret) 1637 return ret; 1638 1639 bdev->driver = driver; 1640 1641 memset(bdev->man, 0, sizeof(bdev->man)); 1642 1643 /* 1644 * Initialize the system memory buffer type. 1645 * Other types need to be driver / IOCTL initialized. 1646 */ 1647 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); 1648 if (unlikely(ret != 0)) 1649 goto out_no_sys; 1650 1651 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset, 1652 0x10000000); 1653 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue); 1654 INIT_LIST_HEAD(&bdev->ddestroy); 1655 bdev->dev_mapping = mapping; 1656 bdev->glob = glob; 1657 bdev->need_dma32 = need_dma32; 1658 mutex_lock(&ttm_global_mutex); 1659 list_add_tail(&bdev->device_list, &glob->device_list); 1660 mutex_unlock(&ttm_global_mutex); 1661 1662 return 0; 1663 out_no_sys: 1664 ttm_bo_global_release(); 1665 return ret; 1666 } 1667 EXPORT_SYMBOL(ttm_bo_device_init); 1668 1669 /* 1670 * buffer object vm functions. 1671 */ 1672 1673 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 1674 { 1675 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 1676 1677 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 1678 if (mem->mem_type == TTM_PL_SYSTEM) 1679 return false; 1680 1681 if (man->flags & TTM_MEMTYPE_FLAG_CMA) 1682 return false; 1683 1684 if (mem->placement & TTM_PL_FLAG_CACHED) 1685 return false; 1686 } 1687 return true; 1688 } 1689 1690 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) 1691 { 1692 struct ttm_bo_device *bdev = bo->bdev; 1693 1694 drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping); 1695 ttm_mem_io_free_vm(bo); 1696 } 1697 1698 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) 1699 { 1700 struct ttm_bo_device *bdev = bo->bdev; 1701 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 1702 1703 ttm_mem_io_lock(man, false); 1704 ttm_bo_unmap_virtual_locked(bo); 1705 ttm_mem_io_unlock(man); 1706 } 1707 1708 1709 EXPORT_SYMBOL(ttm_bo_unmap_virtual); 1710 1711 int ttm_bo_wait(struct ttm_buffer_object *bo, 1712 bool interruptible, bool no_wait) 1713 { 1714 long timeout = 15 * HZ; 1715 1716 if (no_wait) { 1717 if (reservation_object_test_signaled_rcu(bo->resv, true)) 1718 return 0; 1719 else 1720 return -EBUSY; 1721 } 1722 1723 timeout = reservation_object_wait_timeout_rcu(bo->resv, true, 1724 interruptible, timeout); 1725 if (timeout < 0) 1726 return timeout; 1727 1728 if (timeout == 0) 1729 return -EBUSY; 1730 1731 reservation_object_add_excl_fence(bo->resv, NULL); 1732 return 0; 1733 } 1734 EXPORT_SYMBOL(ttm_bo_wait); 1735 1736 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) 1737 { 1738 int ret = 0; 1739 1740 /* 1741 * Using ttm_bo_reserve makes sure the lru lists are updated. 1742 */ 1743 1744 ret = ttm_bo_reserve(bo, true, no_wait, NULL); 1745 if (unlikely(ret != 0)) 1746 return ret; 1747 ret = ttm_bo_wait(bo, true, no_wait); 1748 if (likely(ret == 0)) 1749 atomic_inc(&bo->cpu_writers); 1750 ttm_bo_unreserve(bo); 1751 return ret; 1752 } 1753 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab); 1754 1755 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) 1756 { 1757 atomic_dec(&bo->cpu_writers); 1758 } 1759 EXPORT_SYMBOL(ttm_bo_synccpu_write_release); 1760 1761 /** 1762 * A buffer object shrink method that tries to swap out the first 1763 * buffer object on the bo_global::swap_lru list. 1764 */ 1765 int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx) 1766 { 1767 struct ttm_buffer_object *bo; 1768 int ret = -EBUSY; 1769 bool locked; 1770 unsigned i; 1771 1772 spin_lock(&glob->lru_lock); 1773 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 1774 list_for_each_entry(bo, &glob->swap_lru[i], swap) { 1775 if (ttm_bo_evict_swapout_allowable(bo, ctx, &locked)) { 1776 ret = 0; 1777 break; 1778 } 1779 } 1780 if (!ret) 1781 break; 1782 } 1783 1784 if (ret) { 1785 spin_unlock(&glob->lru_lock); 1786 return ret; 1787 } 1788 1789 kref_get(&bo->list_kref); 1790 1791 if (!list_empty(&bo->ddestroy)) { 1792 ret = ttm_bo_cleanup_refs(bo, false, false, locked); 1793 kref_put(&bo->list_kref, ttm_bo_release_list); 1794 return ret; 1795 } 1796 1797 ttm_bo_del_from_lru(bo); 1798 spin_unlock(&glob->lru_lock); 1799 1800 /** 1801 * Move to system cached 1802 */ 1803 1804 if (bo->mem.mem_type != TTM_PL_SYSTEM || 1805 bo->ttm->caching_state != tt_cached) { 1806 struct ttm_operation_ctx ctx = { false, false }; 1807 struct ttm_mem_reg evict_mem; 1808 1809 evict_mem = bo->mem; 1810 evict_mem.mm_node = NULL; 1811 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; 1812 evict_mem.mem_type = TTM_PL_SYSTEM; 1813 1814 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx); 1815 if (unlikely(ret != 0)) 1816 goto out; 1817 } 1818 1819 /** 1820 * Make sure BO is idle. 1821 */ 1822 1823 ret = ttm_bo_wait(bo, false, false); 1824 if (unlikely(ret != 0)) 1825 goto out; 1826 1827 ttm_bo_unmap_virtual(bo); 1828 1829 /** 1830 * Swap out. Buffer will be swapped in again as soon as 1831 * anyone tries to access a ttm page. 1832 */ 1833 1834 if (bo->bdev->driver->swap_notify) 1835 bo->bdev->driver->swap_notify(bo); 1836 1837 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); 1838 out: 1839 1840 /** 1841 * 1842 * Unreserve without putting on LRU to avoid swapping out an 1843 * already swapped buffer. 1844 */ 1845 if (locked) 1846 reservation_object_unlock(bo->resv); 1847 kref_put(&bo->list_kref, ttm_bo_release_list); 1848 return ret; 1849 } 1850 EXPORT_SYMBOL(ttm_bo_swapout); 1851 1852 void ttm_bo_swapout_all(struct ttm_bo_device *bdev) 1853 { 1854 struct ttm_operation_ctx ctx = { 1855 .interruptible = false, 1856 .no_wait_gpu = false 1857 }; 1858 1859 while (ttm_bo_swapout(bdev->glob, &ctx) == 0) 1860 ; 1861 } 1862 EXPORT_SYMBOL(ttm_bo_swapout_all); 1863 1864 /** 1865 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become 1866 * unreserved 1867 * 1868 * @bo: Pointer to buffer 1869 */ 1870 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo) 1871 { 1872 int ret; 1873 1874 /* 1875 * In the absense of a wait_unlocked API, 1876 * Use the bo::wu_mutex to avoid triggering livelocks due to 1877 * concurrent use of this function. Note that this use of 1878 * bo::wu_mutex can go away if we change locking order to 1879 * mmap_sem -> bo::reserve. 1880 */ 1881 ret = mutex_lock_interruptible(&bo->wu_mutex); 1882 if (unlikely(ret != 0)) 1883 return -ERESTARTSYS; 1884 if (!ww_mutex_is_locked(&bo->resv->lock)) 1885 goto out_unlock; 1886 ret = reservation_object_lock_interruptible(bo->resv, NULL); 1887 if (ret == -EINTR) 1888 ret = -ERESTARTSYS; 1889 if (unlikely(ret != 0)) 1890 goto out_unlock; 1891 reservation_object_unlock(bo->resv); 1892 1893 out_unlock: 1894 mutex_unlock(&bo->wu_mutex); 1895 return ret; 1896 } 1897