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 if (!list_empty(&bo->swap)) { 202 list_del_init(&bo->swap); 203 kref_put(&bo->list_kref, ttm_bo_ref_bug); 204 } 205 if (!list_empty(&bo->lru)) { 206 list_del_init(&bo->lru); 207 kref_put(&bo->list_kref, ttm_bo_ref_bug); 208 } 209 210 /* 211 * TODO: Add a driver hook to delete from 212 * driver-specific LRU's here. 213 */ 214 } 215 216 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo) 217 { 218 struct ttm_bo_global *glob = bo->bdev->glob; 219 220 spin_lock(&glob->lru_lock); 221 ttm_bo_del_from_lru(bo); 222 spin_unlock(&glob->lru_lock); 223 } 224 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru); 225 226 static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos, 227 struct ttm_buffer_object *bo) 228 { 229 if (!pos->first) 230 pos->first = bo; 231 pos->last = bo; 232 } 233 234 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo, 235 struct ttm_lru_bulk_move *bulk) 236 { 237 reservation_object_assert_held(bo->resv); 238 239 ttm_bo_del_from_lru(bo); 240 ttm_bo_add_to_lru(bo); 241 242 if (bulk && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 243 switch (bo->mem.mem_type) { 244 case TTM_PL_TT: 245 ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo); 246 break; 247 248 case TTM_PL_VRAM: 249 ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo); 250 break; 251 } 252 if (bo->ttm && !(bo->ttm->page_flags & 253 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED))) 254 ttm_bo_bulk_move_set_pos(&bulk->swap[bo->priority], bo); 255 } 256 } 257 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail); 258 259 void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk) 260 { 261 unsigned i; 262 263 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 264 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i]; 265 struct ttm_mem_type_manager *man; 266 267 if (!pos->first) 268 continue; 269 270 reservation_object_assert_held(pos->first->resv); 271 reservation_object_assert_held(pos->last->resv); 272 273 man = &pos->first->bdev->man[TTM_PL_TT]; 274 list_bulk_move_tail(&man->lru[i], &pos->first->lru, 275 &pos->last->lru); 276 } 277 278 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 279 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i]; 280 struct ttm_mem_type_manager *man; 281 282 if (!pos->first) 283 continue; 284 285 reservation_object_assert_held(pos->first->resv); 286 reservation_object_assert_held(pos->last->resv); 287 288 man = &pos->first->bdev->man[TTM_PL_VRAM]; 289 list_bulk_move_tail(&man->lru[i], &pos->first->lru, 290 &pos->last->lru); 291 } 292 293 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 294 struct ttm_lru_bulk_move_pos *pos = &bulk->swap[i]; 295 struct list_head *lru; 296 297 if (!pos->first) 298 continue; 299 300 reservation_object_assert_held(pos->first->resv); 301 reservation_object_assert_held(pos->last->resv); 302 303 lru = &pos->first->bdev->glob->swap_lru[i]; 304 list_bulk_move_tail(lru, &pos->first->swap, &pos->last->swap); 305 } 306 } 307 EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail); 308 309 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo, 310 struct ttm_mem_reg *mem, bool evict, 311 struct ttm_operation_ctx *ctx) 312 { 313 struct ttm_bo_device *bdev = bo->bdev; 314 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem); 315 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem); 316 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type]; 317 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type]; 318 int ret = 0; 319 320 if (old_is_pci || new_is_pci || 321 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) { 322 ret = ttm_mem_io_lock(old_man, true); 323 if (unlikely(ret != 0)) 324 goto out_err; 325 ttm_bo_unmap_virtual_locked(bo); 326 ttm_mem_io_unlock(old_man); 327 } 328 329 /* 330 * Create and bind a ttm if required. 331 */ 332 333 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 334 if (bo->ttm == NULL) { 335 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED); 336 ret = ttm_tt_create(bo, zero); 337 if (ret) 338 goto out_err; 339 } 340 341 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement); 342 if (ret) 343 goto out_err; 344 345 if (mem->mem_type != TTM_PL_SYSTEM) { 346 ret = ttm_tt_bind(bo->ttm, mem, ctx); 347 if (ret) 348 goto out_err; 349 } 350 351 if (bo->mem.mem_type == TTM_PL_SYSTEM) { 352 if (bdev->driver->move_notify) 353 bdev->driver->move_notify(bo, evict, mem); 354 bo->mem = *mem; 355 mem->mm_node = NULL; 356 goto moved; 357 } 358 } 359 360 if (bdev->driver->move_notify) 361 bdev->driver->move_notify(bo, evict, mem); 362 363 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) && 364 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) 365 ret = ttm_bo_move_ttm(bo, ctx, mem); 366 else if (bdev->driver->move) 367 ret = bdev->driver->move(bo, evict, ctx, mem); 368 else 369 ret = ttm_bo_move_memcpy(bo, ctx, mem); 370 371 if (ret) { 372 if (bdev->driver->move_notify) { 373 swap(*mem, bo->mem); 374 bdev->driver->move_notify(bo, false, mem); 375 swap(*mem, bo->mem); 376 } 377 378 goto out_err; 379 } 380 381 moved: 382 if (bo->evicted) { 383 if (bdev->driver->invalidate_caches) { 384 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement); 385 if (ret) 386 pr_err("Can not flush read caches\n"); 387 } 388 bo->evicted = false; 389 } 390 391 if (bo->mem.mm_node) 392 bo->offset = (bo->mem.start << PAGE_SHIFT) + 393 bdev->man[bo->mem.mem_type].gpu_offset; 394 else 395 bo->offset = 0; 396 397 ctx->bytes_moved += bo->num_pages << PAGE_SHIFT; 398 return 0; 399 400 out_err: 401 new_man = &bdev->man[bo->mem.mem_type]; 402 if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) { 403 ttm_tt_destroy(bo->ttm); 404 bo->ttm = NULL; 405 } 406 407 return ret; 408 } 409 410 /** 411 * Call bo::reserved. 412 * Will release GPU memory type usage on destruction. 413 * This is the place to put in driver specific hooks to release 414 * driver private resources. 415 * Will release the bo::reserved lock. 416 */ 417 418 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo) 419 { 420 if (bo->bdev->driver->move_notify) 421 bo->bdev->driver->move_notify(bo, false, NULL); 422 423 ttm_tt_destroy(bo->ttm); 424 bo->ttm = NULL; 425 ttm_bo_mem_put(bo, &bo->mem); 426 } 427 428 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo) 429 { 430 int r; 431 432 if (bo->resv == &bo->ttm_resv) 433 return 0; 434 435 BUG_ON(!reservation_object_trylock(&bo->ttm_resv)); 436 437 r = reservation_object_copy_fences(&bo->ttm_resv, bo->resv); 438 if (r) 439 reservation_object_unlock(&bo->ttm_resv); 440 441 return r; 442 } 443 444 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo) 445 { 446 struct reservation_object_list *fobj; 447 struct dma_fence *fence; 448 int i; 449 450 fobj = reservation_object_get_list(&bo->ttm_resv); 451 fence = reservation_object_get_excl(&bo->ttm_resv); 452 if (fence && !fence->ops->signaled) 453 dma_fence_enable_sw_signaling(fence); 454 455 for (i = 0; fobj && i < fobj->shared_count; ++i) { 456 fence = rcu_dereference_protected(fobj->shared[i], 457 reservation_object_held(bo->resv)); 458 459 if (!fence->ops->signaled) 460 dma_fence_enable_sw_signaling(fence); 461 } 462 } 463 464 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo) 465 { 466 struct ttm_bo_device *bdev = bo->bdev; 467 struct ttm_bo_global *glob = bdev->glob; 468 int ret; 469 470 ret = ttm_bo_individualize_resv(bo); 471 if (ret) { 472 /* Last resort, if we fail to allocate memory for the 473 * fences block for the BO to become idle 474 */ 475 reservation_object_wait_timeout_rcu(bo->resv, true, false, 476 30 * HZ); 477 spin_lock(&glob->lru_lock); 478 goto error; 479 } 480 481 spin_lock(&glob->lru_lock); 482 ret = reservation_object_trylock(bo->resv) ? 0 : -EBUSY; 483 if (!ret) { 484 if (reservation_object_test_signaled_rcu(&bo->ttm_resv, true)) { 485 ttm_bo_del_from_lru(bo); 486 spin_unlock(&glob->lru_lock); 487 if (bo->resv != &bo->ttm_resv) 488 reservation_object_unlock(&bo->ttm_resv); 489 490 ttm_bo_cleanup_memtype_use(bo); 491 reservation_object_unlock(bo->resv); 492 return; 493 } 494 495 ttm_bo_flush_all_fences(bo); 496 497 /* 498 * Make NO_EVICT bos immediately available to 499 * shrinkers, now that they are queued for 500 * destruction. 501 */ 502 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) { 503 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT; 504 ttm_bo_add_to_lru(bo); 505 } 506 507 reservation_object_unlock(bo->resv); 508 } 509 if (bo->resv != &bo->ttm_resv) 510 reservation_object_unlock(&bo->ttm_resv); 511 512 error: 513 kref_get(&bo->list_kref); 514 list_add_tail(&bo->ddestroy, &bdev->ddestroy); 515 spin_unlock(&glob->lru_lock); 516 517 schedule_delayed_work(&bdev->wq, 518 ((HZ / 100) < 1) ? 1 : HZ / 100); 519 } 520 521 /** 522 * function ttm_bo_cleanup_refs 523 * If bo idle, remove from delayed- and lru lists, and unref. 524 * If not idle, do nothing. 525 * 526 * Must be called with lru_lock and reservation held, this function 527 * will drop the lru lock and optionally the reservation lock before returning. 528 * 529 * @interruptible Any sleeps should occur interruptibly. 530 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead. 531 * @unlock_resv Unlock the reservation lock as well. 532 */ 533 534 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo, 535 bool interruptible, bool no_wait_gpu, 536 bool unlock_resv) 537 { 538 struct ttm_bo_global *glob = bo->bdev->glob; 539 struct reservation_object *resv; 540 int ret; 541 542 if (unlikely(list_empty(&bo->ddestroy))) 543 resv = bo->resv; 544 else 545 resv = &bo->ttm_resv; 546 547 if (reservation_object_test_signaled_rcu(resv, true)) 548 ret = 0; 549 else 550 ret = -EBUSY; 551 552 if (ret && !no_wait_gpu) { 553 long lret; 554 555 if (unlock_resv) 556 reservation_object_unlock(bo->resv); 557 spin_unlock(&glob->lru_lock); 558 559 lret = reservation_object_wait_timeout_rcu(resv, true, 560 interruptible, 561 30 * HZ); 562 563 if (lret < 0) 564 return lret; 565 else if (lret == 0) 566 return -EBUSY; 567 568 spin_lock(&glob->lru_lock); 569 if (unlock_resv && !reservation_object_trylock(bo->resv)) { 570 /* 571 * We raced, and lost, someone else holds the reservation now, 572 * and is probably busy in ttm_bo_cleanup_memtype_use. 573 * 574 * Even if it's not the case, because we finished waiting any 575 * delayed destruction would succeed, so just return success 576 * here. 577 */ 578 spin_unlock(&glob->lru_lock); 579 return 0; 580 } 581 ret = 0; 582 } 583 584 if (ret || unlikely(list_empty(&bo->ddestroy))) { 585 if (unlock_resv) 586 reservation_object_unlock(bo->resv); 587 spin_unlock(&glob->lru_lock); 588 return ret; 589 } 590 591 ttm_bo_del_from_lru(bo); 592 list_del_init(&bo->ddestroy); 593 kref_put(&bo->list_kref, ttm_bo_ref_bug); 594 595 spin_unlock(&glob->lru_lock); 596 ttm_bo_cleanup_memtype_use(bo); 597 598 if (unlock_resv) 599 reservation_object_unlock(bo->resv); 600 601 return 0; 602 } 603 604 /** 605 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all 606 * encountered buffers. 607 */ 608 static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all) 609 { 610 struct ttm_bo_global *glob = bdev->glob; 611 struct list_head removed; 612 bool empty; 613 614 INIT_LIST_HEAD(&removed); 615 616 spin_lock(&glob->lru_lock); 617 while (!list_empty(&bdev->ddestroy)) { 618 struct ttm_buffer_object *bo; 619 620 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object, 621 ddestroy); 622 kref_get(&bo->list_kref); 623 list_move_tail(&bo->ddestroy, &removed); 624 625 if (remove_all || bo->resv != &bo->ttm_resv) { 626 spin_unlock(&glob->lru_lock); 627 reservation_object_lock(bo->resv, NULL); 628 629 spin_lock(&glob->lru_lock); 630 ttm_bo_cleanup_refs(bo, false, !remove_all, true); 631 632 } else if (reservation_object_trylock(bo->resv)) { 633 ttm_bo_cleanup_refs(bo, false, !remove_all, true); 634 } else { 635 spin_unlock(&glob->lru_lock); 636 } 637 638 kref_put(&bo->list_kref, ttm_bo_release_list); 639 spin_lock(&glob->lru_lock); 640 } 641 list_splice_tail(&removed, &bdev->ddestroy); 642 empty = list_empty(&bdev->ddestroy); 643 spin_unlock(&glob->lru_lock); 644 645 return empty; 646 } 647 648 static void ttm_bo_delayed_workqueue(struct work_struct *work) 649 { 650 struct ttm_bo_device *bdev = 651 container_of(work, struct ttm_bo_device, wq.work); 652 653 if (!ttm_bo_delayed_delete(bdev, false)) 654 schedule_delayed_work(&bdev->wq, 655 ((HZ / 100) < 1) ? 1 : HZ / 100); 656 } 657 658 static void ttm_bo_release(struct kref *kref) 659 { 660 struct ttm_buffer_object *bo = 661 container_of(kref, struct ttm_buffer_object, kref); 662 struct ttm_bo_device *bdev = bo->bdev; 663 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 664 665 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node); 666 ttm_mem_io_lock(man, false); 667 ttm_mem_io_free_vm(bo); 668 ttm_mem_io_unlock(man); 669 ttm_bo_cleanup_refs_or_queue(bo); 670 kref_put(&bo->list_kref, ttm_bo_release_list); 671 } 672 673 void ttm_bo_put(struct ttm_buffer_object *bo) 674 { 675 kref_put(&bo->kref, ttm_bo_release); 676 } 677 EXPORT_SYMBOL(ttm_bo_put); 678 679 void ttm_bo_unref(struct ttm_buffer_object **p_bo) 680 { 681 struct ttm_buffer_object *bo = *p_bo; 682 683 *p_bo = NULL; 684 ttm_bo_put(bo); 685 } 686 EXPORT_SYMBOL(ttm_bo_unref); 687 688 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) 689 { 690 return cancel_delayed_work_sync(&bdev->wq); 691 } 692 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue); 693 694 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched) 695 { 696 if (resched) 697 schedule_delayed_work(&bdev->wq, 698 ((HZ / 100) < 1) ? 1 : HZ / 100); 699 } 700 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue); 701 702 static int ttm_bo_evict(struct ttm_buffer_object *bo, 703 struct ttm_operation_ctx *ctx) 704 { 705 struct ttm_bo_device *bdev = bo->bdev; 706 struct ttm_mem_reg evict_mem; 707 struct ttm_placement placement; 708 int ret = 0; 709 710 reservation_object_assert_held(bo->resv); 711 712 placement.num_placement = 0; 713 placement.num_busy_placement = 0; 714 bdev->driver->evict_flags(bo, &placement); 715 716 if (!placement.num_placement && !placement.num_busy_placement) { 717 ret = ttm_bo_pipeline_gutting(bo); 718 if (ret) 719 return ret; 720 721 return ttm_tt_create(bo, false); 722 } 723 724 evict_mem = bo->mem; 725 evict_mem.mm_node = NULL; 726 evict_mem.bus.io_reserved_vm = false; 727 evict_mem.bus.io_reserved_count = 0; 728 729 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx); 730 if (ret) { 731 if (ret != -ERESTARTSYS) { 732 pr_err("Failed to find memory space for buffer 0x%p eviction\n", 733 bo); 734 ttm_bo_mem_space_debug(bo, &placement); 735 } 736 goto out; 737 } 738 739 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx); 740 if (unlikely(ret)) { 741 if (ret != -ERESTARTSYS) 742 pr_err("Buffer eviction failed\n"); 743 ttm_bo_mem_put(bo, &evict_mem); 744 goto out; 745 } 746 bo->evicted = true; 747 out: 748 return ret; 749 } 750 751 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo, 752 const struct ttm_place *place) 753 { 754 /* Don't evict this BO if it's outside of the 755 * requested placement range 756 */ 757 if (place->fpfn >= (bo->mem.start + bo->mem.size) || 758 (place->lpfn && place->lpfn <= bo->mem.start)) 759 return false; 760 761 return true; 762 } 763 EXPORT_SYMBOL(ttm_bo_eviction_valuable); 764 765 /** 766 * Check the target bo is allowable to be evicted or swapout, including cases: 767 * 768 * a. if share same reservation object with ctx->resv, have assumption 769 * reservation objects should already be locked, so not lock again and 770 * return true directly when either the opreation allow_reserved_eviction 771 * or the target bo already is in delayed free list; 772 * 773 * b. Otherwise, trylock it. 774 */ 775 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo, 776 struct ttm_operation_ctx *ctx, bool *locked) 777 { 778 bool ret = false; 779 780 *locked = false; 781 if (bo->resv == ctx->resv) { 782 reservation_object_assert_held(bo->resv); 783 if (ctx->flags & TTM_OPT_FLAG_ALLOW_RES_EVICT 784 || !list_empty(&bo->ddestroy)) 785 ret = true; 786 } else { 787 *locked = reservation_object_trylock(bo->resv); 788 ret = *locked; 789 } 790 791 return ret; 792 } 793 794 static int ttm_mem_evict_first(struct ttm_bo_device *bdev, 795 uint32_t mem_type, 796 const struct ttm_place *place, 797 struct ttm_operation_ctx *ctx) 798 { 799 struct ttm_bo_global *glob = bdev->glob; 800 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 801 struct ttm_buffer_object *bo = NULL; 802 bool locked = false; 803 unsigned i; 804 int ret; 805 806 spin_lock(&glob->lru_lock); 807 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 808 list_for_each_entry(bo, &man->lru[i], lru) { 809 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked)) 810 continue; 811 812 if (place && !bdev->driver->eviction_valuable(bo, 813 place)) { 814 if (locked) 815 reservation_object_unlock(bo->resv); 816 continue; 817 } 818 break; 819 } 820 821 /* If the inner loop terminated early, we have our candidate */ 822 if (&bo->lru != &man->lru[i]) 823 break; 824 825 bo = NULL; 826 } 827 828 if (!bo) { 829 spin_unlock(&glob->lru_lock); 830 return -EBUSY; 831 } 832 833 kref_get(&bo->list_kref); 834 835 if (!list_empty(&bo->ddestroy)) { 836 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible, 837 ctx->no_wait_gpu, locked); 838 kref_put(&bo->list_kref, ttm_bo_release_list); 839 return ret; 840 } 841 842 ttm_bo_del_from_lru(bo); 843 spin_unlock(&glob->lru_lock); 844 845 ret = ttm_bo_evict(bo, ctx); 846 if (locked) { 847 ttm_bo_unreserve(bo); 848 } else { 849 spin_lock(&glob->lru_lock); 850 ttm_bo_add_to_lru(bo); 851 spin_unlock(&glob->lru_lock); 852 } 853 854 kref_put(&bo->list_kref, ttm_bo_release_list); 855 return ret; 856 } 857 858 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem) 859 { 860 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type]; 861 862 if (mem->mm_node) 863 (*man->func->put_node)(man, mem); 864 } 865 EXPORT_SYMBOL(ttm_bo_mem_put); 866 867 /** 868 * Add the last move fence to the BO and reserve a new shared slot. 869 */ 870 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo, 871 struct ttm_mem_type_manager *man, 872 struct ttm_mem_reg *mem) 873 { 874 struct dma_fence *fence; 875 int ret; 876 877 spin_lock(&man->move_lock); 878 fence = dma_fence_get(man->move); 879 spin_unlock(&man->move_lock); 880 881 if (fence) { 882 reservation_object_add_shared_fence(bo->resv, fence); 883 884 ret = reservation_object_reserve_shared(bo->resv, 1); 885 if (unlikely(ret)) 886 return ret; 887 888 dma_fence_put(bo->moving); 889 bo->moving = fence; 890 } 891 892 return 0; 893 } 894 895 /** 896 * Repeatedly evict memory from the LRU for @mem_type until we create enough 897 * space, or we've evicted everything and there isn't enough space. 898 */ 899 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo, 900 uint32_t mem_type, 901 const struct ttm_place *place, 902 struct ttm_mem_reg *mem, 903 struct ttm_operation_ctx *ctx) 904 { 905 struct ttm_bo_device *bdev = bo->bdev; 906 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 907 int ret; 908 909 do { 910 ret = (*man->func->get_node)(man, bo, place, mem); 911 if (unlikely(ret != 0)) 912 return ret; 913 if (mem->mm_node) 914 break; 915 ret = ttm_mem_evict_first(bdev, mem_type, place, ctx); 916 if (unlikely(ret != 0)) 917 return ret; 918 } while (1); 919 mem->mem_type = mem_type; 920 return ttm_bo_add_move_fence(bo, man, mem); 921 } 922 923 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man, 924 uint32_t cur_placement, 925 uint32_t proposed_placement) 926 { 927 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING; 928 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING; 929 930 /** 931 * Keep current caching if possible. 932 */ 933 934 if ((cur_placement & caching) != 0) 935 result |= (cur_placement & caching); 936 else if ((man->default_caching & caching) != 0) 937 result |= man->default_caching; 938 else if ((TTM_PL_FLAG_CACHED & caching) != 0) 939 result |= TTM_PL_FLAG_CACHED; 940 else if ((TTM_PL_FLAG_WC & caching) != 0) 941 result |= TTM_PL_FLAG_WC; 942 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0) 943 result |= TTM_PL_FLAG_UNCACHED; 944 945 return result; 946 } 947 948 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man, 949 uint32_t mem_type, 950 const struct ttm_place *place, 951 uint32_t *masked_placement) 952 { 953 uint32_t cur_flags = ttm_bo_type_flags(mem_type); 954 955 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0) 956 return false; 957 958 if ((place->flags & man->available_caching) == 0) 959 return false; 960 961 cur_flags |= (place->flags & man->available_caching); 962 963 *masked_placement = cur_flags; 964 return true; 965 } 966 967 /** 968 * Creates space for memory region @mem according to its type. 969 * 970 * This function first searches for free space in compatible memory types in 971 * the priority order defined by the driver. If free space isn't found, then 972 * ttm_bo_mem_force_space is attempted in priority order to evict and find 973 * space. 974 */ 975 int ttm_bo_mem_space(struct ttm_buffer_object *bo, 976 struct ttm_placement *placement, 977 struct ttm_mem_reg *mem, 978 struct ttm_operation_ctx *ctx) 979 { 980 struct ttm_bo_device *bdev = bo->bdev; 981 struct ttm_mem_type_manager *man; 982 uint32_t mem_type = TTM_PL_SYSTEM; 983 uint32_t cur_flags = 0; 984 bool type_found = false; 985 bool type_ok = false; 986 bool has_erestartsys = false; 987 int i, ret; 988 989 ret = reservation_object_reserve_shared(bo->resv, 1); 990 if (unlikely(ret)) 991 return ret; 992 993 mem->mm_node = NULL; 994 for (i = 0; i < placement->num_placement; ++i) { 995 const struct ttm_place *place = &placement->placement[i]; 996 997 ret = ttm_mem_type_from_place(place, &mem_type); 998 if (ret) 999 return ret; 1000 man = &bdev->man[mem_type]; 1001 if (!man->has_type || !man->use_type) 1002 continue; 1003 1004 type_ok = ttm_bo_mt_compatible(man, mem_type, place, 1005 &cur_flags); 1006 1007 if (!type_ok) 1008 continue; 1009 1010 type_found = true; 1011 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1012 cur_flags); 1013 /* 1014 * Use the access and other non-mapping-related flag bits from 1015 * the memory placement flags to the current flags 1016 */ 1017 ttm_flag_masked(&cur_flags, place->flags, 1018 ~TTM_PL_MASK_MEMTYPE); 1019 1020 if (mem_type == TTM_PL_SYSTEM) 1021 break; 1022 1023 ret = (*man->func->get_node)(man, bo, place, mem); 1024 if (unlikely(ret)) 1025 return ret; 1026 1027 if (mem->mm_node) { 1028 ret = ttm_bo_add_move_fence(bo, man, mem); 1029 if (unlikely(ret)) { 1030 (*man->func->put_node)(man, mem); 1031 return ret; 1032 } 1033 break; 1034 } 1035 } 1036 1037 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) { 1038 mem->mem_type = mem_type; 1039 mem->placement = cur_flags; 1040 return 0; 1041 } 1042 1043 for (i = 0; i < placement->num_busy_placement; ++i) { 1044 const struct ttm_place *place = &placement->busy_placement[i]; 1045 1046 ret = ttm_mem_type_from_place(place, &mem_type); 1047 if (ret) 1048 return ret; 1049 man = &bdev->man[mem_type]; 1050 if (!man->has_type || !man->use_type) 1051 continue; 1052 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags)) 1053 continue; 1054 1055 type_found = true; 1056 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1057 cur_flags); 1058 /* 1059 * Use the access and other non-mapping-related flag bits from 1060 * the memory placement flags to the current flags 1061 */ 1062 ttm_flag_masked(&cur_flags, place->flags, 1063 ~TTM_PL_MASK_MEMTYPE); 1064 1065 if (mem_type == TTM_PL_SYSTEM) { 1066 mem->mem_type = mem_type; 1067 mem->placement = cur_flags; 1068 mem->mm_node = NULL; 1069 return 0; 1070 } 1071 1072 ret = ttm_bo_mem_force_space(bo, mem_type, place, mem, ctx); 1073 if (ret == 0 && mem->mm_node) { 1074 mem->placement = cur_flags; 1075 return 0; 1076 } 1077 if (ret == -ERESTARTSYS) 1078 has_erestartsys = true; 1079 } 1080 1081 if (!type_found) { 1082 pr_err(TTM_PFX "No compatible memory type found\n"); 1083 return -EINVAL; 1084 } 1085 1086 return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM; 1087 } 1088 EXPORT_SYMBOL(ttm_bo_mem_space); 1089 1090 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo, 1091 struct ttm_placement *placement, 1092 struct ttm_operation_ctx *ctx) 1093 { 1094 int ret = 0; 1095 struct ttm_mem_reg mem; 1096 1097 reservation_object_assert_held(bo->resv); 1098 1099 mem.num_pages = bo->num_pages; 1100 mem.size = mem.num_pages << PAGE_SHIFT; 1101 mem.page_alignment = bo->mem.page_alignment; 1102 mem.bus.io_reserved_vm = false; 1103 mem.bus.io_reserved_count = 0; 1104 /* 1105 * Determine where to move the buffer. 1106 */ 1107 ret = ttm_bo_mem_space(bo, placement, &mem, ctx); 1108 if (ret) 1109 goto out_unlock; 1110 ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx); 1111 out_unlock: 1112 if (ret && mem.mm_node) 1113 ttm_bo_mem_put(bo, &mem); 1114 return ret; 1115 } 1116 1117 static bool ttm_bo_places_compat(const struct ttm_place *places, 1118 unsigned num_placement, 1119 struct ttm_mem_reg *mem, 1120 uint32_t *new_flags) 1121 { 1122 unsigned i; 1123 1124 for (i = 0; i < num_placement; i++) { 1125 const struct ttm_place *heap = &places[i]; 1126 1127 if (mem->mm_node && (mem->start < heap->fpfn || 1128 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn))) 1129 continue; 1130 1131 *new_flags = heap->flags; 1132 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) && 1133 (*new_flags & mem->placement & TTM_PL_MASK_MEM) && 1134 (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) || 1135 (mem->placement & TTM_PL_FLAG_CONTIGUOUS))) 1136 return true; 1137 } 1138 return false; 1139 } 1140 1141 bool ttm_bo_mem_compat(struct ttm_placement *placement, 1142 struct ttm_mem_reg *mem, 1143 uint32_t *new_flags) 1144 { 1145 if (ttm_bo_places_compat(placement->placement, placement->num_placement, 1146 mem, new_flags)) 1147 return true; 1148 1149 if ((placement->busy_placement != placement->placement || 1150 placement->num_busy_placement > placement->num_placement) && 1151 ttm_bo_places_compat(placement->busy_placement, 1152 placement->num_busy_placement, 1153 mem, new_flags)) 1154 return true; 1155 1156 return false; 1157 } 1158 EXPORT_SYMBOL(ttm_bo_mem_compat); 1159 1160 int ttm_bo_validate(struct ttm_buffer_object *bo, 1161 struct ttm_placement *placement, 1162 struct ttm_operation_ctx *ctx) 1163 { 1164 int ret; 1165 uint32_t new_flags; 1166 1167 reservation_object_assert_held(bo->resv); 1168 /* 1169 * Check whether we need to move buffer. 1170 */ 1171 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) { 1172 ret = ttm_bo_move_buffer(bo, placement, ctx); 1173 if (ret) 1174 return ret; 1175 } else { 1176 /* 1177 * Use the access and other non-mapping-related flag bits from 1178 * the compatible memory placement flags to the active flags 1179 */ 1180 ttm_flag_masked(&bo->mem.placement, new_flags, 1181 ~TTM_PL_MASK_MEMTYPE); 1182 } 1183 /* 1184 * We might need to add a TTM. 1185 */ 1186 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 1187 ret = ttm_tt_create(bo, true); 1188 if (ret) 1189 return ret; 1190 } 1191 return 0; 1192 } 1193 EXPORT_SYMBOL(ttm_bo_validate); 1194 1195 int ttm_bo_init_reserved(struct ttm_bo_device *bdev, 1196 struct ttm_buffer_object *bo, 1197 unsigned long size, 1198 enum ttm_bo_type type, 1199 struct ttm_placement *placement, 1200 uint32_t page_alignment, 1201 struct ttm_operation_ctx *ctx, 1202 size_t acc_size, 1203 struct sg_table *sg, 1204 struct reservation_object *resv, 1205 void (*destroy) (struct ttm_buffer_object *)) 1206 { 1207 int ret = 0; 1208 unsigned long num_pages; 1209 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob; 1210 bool locked; 1211 1212 ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx); 1213 if (ret) { 1214 pr_err("Out of kernel memory\n"); 1215 if (destroy) 1216 (*destroy)(bo); 1217 else 1218 kfree(bo); 1219 return -ENOMEM; 1220 } 1221 1222 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 1223 if (num_pages == 0) { 1224 pr_err("Illegal buffer object size\n"); 1225 if (destroy) 1226 (*destroy)(bo); 1227 else 1228 kfree(bo); 1229 ttm_mem_global_free(mem_glob, acc_size); 1230 return -EINVAL; 1231 } 1232 bo->destroy = destroy ? destroy : ttm_bo_default_destroy; 1233 1234 kref_init(&bo->kref); 1235 kref_init(&bo->list_kref); 1236 atomic_set(&bo->cpu_writers, 0); 1237 INIT_LIST_HEAD(&bo->lru); 1238 INIT_LIST_HEAD(&bo->ddestroy); 1239 INIT_LIST_HEAD(&bo->swap); 1240 INIT_LIST_HEAD(&bo->io_reserve_lru); 1241 mutex_init(&bo->wu_mutex); 1242 bo->bdev = bdev; 1243 bo->type = type; 1244 bo->num_pages = num_pages; 1245 bo->mem.size = num_pages << PAGE_SHIFT; 1246 bo->mem.mem_type = TTM_PL_SYSTEM; 1247 bo->mem.num_pages = bo->num_pages; 1248 bo->mem.mm_node = NULL; 1249 bo->mem.page_alignment = page_alignment; 1250 bo->mem.bus.io_reserved_vm = false; 1251 bo->mem.bus.io_reserved_count = 0; 1252 bo->moving = NULL; 1253 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED); 1254 bo->acc_size = acc_size; 1255 bo->sg = sg; 1256 if (resv) { 1257 bo->resv = resv; 1258 reservation_object_assert_held(bo->resv); 1259 } else { 1260 bo->resv = &bo->ttm_resv; 1261 } 1262 reservation_object_init(&bo->ttm_resv); 1263 atomic_inc(&bo->bdev->glob->bo_count); 1264 drm_vma_node_reset(&bo->vma_node); 1265 1266 /* 1267 * For ttm_bo_type_device buffers, allocate 1268 * address space from the device. 1269 */ 1270 if (bo->type == ttm_bo_type_device || 1271 bo->type == ttm_bo_type_sg) 1272 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node, 1273 bo->mem.num_pages); 1274 1275 /* passed reservation objects should already be locked, 1276 * since otherwise lockdep will be angered in radeon. 1277 */ 1278 if (!resv) { 1279 locked = reservation_object_trylock(bo->resv); 1280 WARN_ON(!locked); 1281 } 1282 1283 if (likely(!ret)) 1284 ret = ttm_bo_validate(bo, placement, ctx); 1285 1286 if (unlikely(ret)) { 1287 if (!resv) 1288 ttm_bo_unreserve(bo); 1289 1290 ttm_bo_put(bo); 1291 return ret; 1292 } 1293 1294 if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 1295 spin_lock(&bdev->glob->lru_lock); 1296 ttm_bo_add_to_lru(bo); 1297 spin_unlock(&bdev->glob->lru_lock); 1298 } 1299 1300 return ret; 1301 } 1302 EXPORT_SYMBOL(ttm_bo_init_reserved); 1303 1304 int ttm_bo_init(struct ttm_bo_device *bdev, 1305 struct ttm_buffer_object *bo, 1306 unsigned long size, 1307 enum ttm_bo_type type, 1308 struct ttm_placement *placement, 1309 uint32_t page_alignment, 1310 bool interruptible, 1311 size_t acc_size, 1312 struct sg_table *sg, 1313 struct reservation_object *resv, 1314 void (*destroy) (struct ttm_buffer_object *)) 1315 { 1316 struct ttm_operation_ctx ctx = { interruptible, false }; 1317 int ret; 1318 1319 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement, 1320 page_alignment, &ctx, acc_size, 1321 sg, resv, destroy); 1322 if (ret) 1323 return ret; 1324 1325 if (!resv) 1326 ttm_bo_unreserve(bo); 1327 1328 return 0; 1329 } 1330 EXPORT_SYMBOL(ttm_bo_init); 1331 1332 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev, 1333 unsigned long bo_size, 1334 unsigned struct_size) 1335 { 1336 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1337 size_t size = 0; 1338 1339 size += ttm_round_pot(struct_size); 1340 size += ttm_round_pot(npages * sizeof(void *)); 1341 size += ttm_round_pot(sizeof(struct ttm_tt)); 1342 return size; 1343 } 1344 EXPORT_SYMBOL(ttm_bo_acc_size); 1345 1346 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev, 1347 unsigned long bo_size, 1348 unsigned struct_size) 1349 { 1350 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1351 size_t size = 0; 1352 1353 size += ttm_round_pot(struct_size); 1354 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t))); 1355 size += ttm_round_pot(sizeof(struct ttm_dma_tt)); 1356 return size; 1357 } 1358 EXPORT_SYMBOL(ttm_bo_dma_acc_size); 1359 1360 int ttm_bo_create(struct ttm_bo_device *bdev, 1361 unsigned long size, 1362 enum ttm_bo_type type, 1363 struct ttm_placement *placement, 1364 uint32_t page_alignment, 1365 bool interruptible, 1366 struct ttm_buffer_object **p_bo) 1367 { 1368 struct ttm_buffer_object *bo; 1369 size_t acc_size; 1370 int ret; 1371 1372 bo = kzalloc(sizeof(*bo), GFP_KERNEL); 1373 if (unlikely(bo == NULL)) 1374 return -ENOMEM; 1375 1376 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object)); 1377 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment, 1378 interruptible, acc_size, 1379 NULL, NULL, NULL); 1380 if (likely(ret == 0)) 1381 *p_bo = bo; 1382 1383 return ret; 1384 } 1385 EXPORT_SYMBOL(ttm_bo_create); 1386 1387 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev, 1388 unsigned mem_type) 1389 { 1390 struct ttm_operation_ctx ctx = { 1391 .interruptible = false, 1392 .no_wait_gpu = false, 1393 .flags = TTM_OPT_FLAG_FORCE_ALLOC 1394 }; 1395 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1396 struct ttm_bo_global *glob = bdev->glob; 1397 struct dma_fence *fence; 1398 int ret; 1399 unsigned i; 1400 1401 /* 1402 * Can't use standard list traversal since we're unlocking. 1403 */ 1404 1405 spin_lock(&glob->lru_lock); 1406 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 1407 while (!list_empty(&man->lru[i])) { 1408 spin_unlock(&glob->lru_lock); 1409 ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx); 1410 if (ret) 1411 return ret; 1412 spin_lock(&glob->lru_lock); 1413 } 1414 } 1415 spin_unlock(&glob->lru_lock); 1416 1417 spin_lock(&man->move_lock); 1418 fence = dma_fence_get(man->move); 1419 spin_unlock(&man->move_lock); 1420 1421 if (fence) { 1422 ret = dma_fence_wait(fence, false); 1423 dma_fence_put(fence); 1424 if (ret) 1425 return ret; 1426 } 1427 1428 return 0; 1429 } 1430 1431 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1432 { 1433 struct ttm_mem_type_manager *man; 1434 int ret = -EINVAL; 1435 1436 if (mem_type >= TTM_NUM_MEM_TYPES) { 1437 pr_err("Illegal memory type %d\n", mem_type); 1438 return ret; 1439 } 1440 man = &bdev->man[mem_type]; 1441 1442 if (!man->has_type) { 1443 pr_err("Trying to take down uninitialized memory manager type %u\n", 1444 mem_type); 1445 return ret; 1446 } 1447 1448 man->use_type = false; 1449 man->has_type = false; 1450 1451 ret = 0; 1452 if (mem_type > 0) { 1453 ret = ttm_bo_force_list_clean(bdev, mem_type); 1454 if (ret) { 1455 pr_err("Cleanup eviction failed\n"); 1456 return ret; 1457 } 1458 1459 ret = (*man->func->takedown)(man); 1460 } 1461 1462 dma_fence_put(man->move); 1463 man->move = NULL; 1464 1465 return ret; 1466 } 1467 EXPORT_SYMBOL(ttm_bo_clean_mm); 1468 1469 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1470 { 1471 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1472 1473 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) { 1474 pr_err("Illegal memory manager memory type %u\n", mem_type); 1475 return -EINVAL; 1476 } 1477 1478 if (!man->has_type) { 1479 pr_err("Memory type %u has not been initialized\n", mem_type); 1480 return 0; 1481 } 1482 1483 return ttm_bo_force_list_clean(bdev, mem_type); 1484 } 1485 EXPORT_SYMBOL(ttm_bo_evict_mm); 1486 1487 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type, 1488 unsigned long p_size) 1489 { 1490 int ret; 1491 struct ttm_mem_type_manager *man; 1492 unsigned i; 1493 1494 BUG_ON(type >= TTM_NUM_MEM_TYPES); 1495 man = &bdev->man[type]; 1496 BUG_ON(man->has_type); 1497 man->io_reserve_fastpath = true; 1498 man->use_io_reserve_lru = false; 1499 mutex_init(&man->io_reserve_mutex); 1500 spin_lock_init(&man->move_lock); 1501 INIT_LIST_HEAD(&man->io_reserve_lru); 1502 1503 ret = bdev->driver->init_mem_type(bdev, type, man); 1504 if (ret) 1505 return ret; 1506 man->bdev = bdev; 1507 1508 if (type != TTM_PL_SYSTEM) { 1509 ret = (*man->func->init)(man, p_size); 1510 if (ret) 1511 return ret; 1512 } 1513 man->has_type = true; 1514 man->use_type = true; 1515 man->size = p_size; 1516 1517 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1518 INIT_LIST_HEAD(&man->lru[i]); 1519 man->move = NULL; 1520 1521 return 0; 1522 } 1523 EXPORT_SYMBOL(ttm_bo_init_mm); 1524 1525 static void ttm_bo_global_kobj_release(struct kobject *kobj) 1526 { 1527 struct ttm_bo_global *glob = 1528 container_of(kobj, struct ttm_bo_global, kobj); 1529 1530 __free_page(glob->dummy_read_page); 1531 } 1532 1533 static void ttm_bo_global_release(void) 1534 { 1535 struct ttm_bo_global *glob = &ttm_bo_glob; 1536 1537 mutex_lock(&ttm_global_mutex); 1538 if (--glob->use_count > 0) 1539 goto out; 1540 1541 kobject_del(&glob->kobj); 1542 kobject_put(&glob->kobj); 1543 ttm_mem_global_release(&ttm_mem_glob); 1544 out: 1545 mutex_unlock(&ttm_global_mutex); 1546 } 1547 1548 static int ttm_bo_global_init(void) 1549 { 1550 struct ttm_bo_global *glob = &ttm_bo_glob; 1551 int ret = 0; 1552 unsigned i; 1553 1554 mutex_lock(&ttm_global_mutex); 1555 if (++glob->use_count > 1) 1556 goto out; 1557 1558 ret = ttm_mem_global_init(&ttm_mem_glob); 1559 if (ret) 1560 goto out; 1561 1562 spin_lock_init(&glob->lru_lock); 1563 glob->mem_glob = &ttm_mem_glob; 1564 glob->mem_glob->bo_glob = glob; 1565 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32); 1566 1567 if (unlikely(glob->dummy_read_page == NULL)) { 1568 ret = -ENOMEM; 1569 goto out; 1570 } 1571 1572 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1573 INIT_LIST_HEAD(&glob->swap_lru[i]); 1574 INIT_LIST_HEAD(&glob->device_list); 1575 atomic_set(&glob->bo_count, 0); 1576 1577 ret = kobject_init_and_add( 1578 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects"); 1579 if (unlikely(ret != 0)) 1580 kobject_put(&glob->kobj); 1581 out: 1582 mutex_unlock(&ttm_global_mutex); 1583 return ret; 1584 } 1585 1586 int ttm_bo_device_release(struct ttm_bo_device *bdev) 1587 { 1588 int ret = 0; 1589 unsigned i = TTM_NUM_MEM_TYPES; 1590 struct ttm_mem_type_manager *man; 1591 struct ttm_bo_global *glob = bdev->glob; 1592 1593 while (i--) { 1594 man = &bdev->man[i]; 1595 if (man->has_type) { 1596 man->use_type = false; 1597 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { 1598 ret = -EBUSY; 1599 pr_err("DRM memory manager type %d is not clean\n", 1600 i); 1601 } 1602 man->has_type = false; 1603 } 1604 } 1605 1606 mutex_lock(&ttm_global_mutex); 1607 list_del(&bdev->device_list); 1608 mutex_unlock(&ttm_global_mutex); 1609 1610 cancel_delayed_work_sync(&bdev->wq); 1611 1612 if (ttm_bo_delayed_delete(bdev, true)) 1613 pr_debug("Delayed destroy list was clean\n"); 1614 1615 spin_lock(&glob->lru_lock); 1616 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1617 if (list_empty(&bdev->man[0].lru[0])) 1618 pr_debug("Swap list %d was clean\n", i); 1619 spin_unlock(&glob->lru_lock); 1620 1621 drm_vma_offset_manager_destroy(&bdev->vma_manager); 1622 1623 if (!ret) 1624 ttm_bo_global_release(); 1625 1626 return ret; 1627 } 1628 EXPORT_SYMBOL(ttm_bo_device_release); 1629 1630 int ttm_bo_device_init(struct ttm_bo_device *bdev, 1631 struct ttm_bo_driver *driver, 1632 struct address_space *mapping, 1633 uint64_t file_page_offset, 1634 bool need_dma32) 1635 { 1636 struct ttm_bo_global *glob = &ttm_bo_glob; 1637 int ret; 1638 1639 ret = ttm_bo_global_init(); 1640 if (ret) 1641 return ret; 1642 1643 bdev->driver = driver; 1644 1645 memset(bdev->man, 0, sizeof(bdev->man)); 1646 1647 /* 1648 * Initialize the system memory buffer type. 1649 * Other types need to be driver / IOCTL initialized. 1650 */ 1651 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); 1652 if (unlikely(ret != 0)) 1653 goto out_no_sys; 1654 1655 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset, 1656 0x10000000); 1657 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue); 1658 INIT_LIST_HEAD(&bdev->ddestroy); 1659 bdev->dev_mapping = mapping; 1660 bdev->glob = glob; 1661 bdev->need_dma32 = need_dma32; 1662 mutex_lock(&ttm_global_mutex); 1663 list_add_tail(&bdev->device_list, &glob->device_list); 1664 mutex_unlock(&ttm_global_mutex); 1665 1666 return 0; 1667 out_no_sys: 1668 ttm_bo_global_release(); 1669 return ret; 1670 } 1671 EXPORT_SYMBOL(ttm_bo_device_init); 1672 1673 /* 1674 * buffer object vm functions. 1675 */ 1676 1677 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 1678 { 1679 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 1680 1681 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 1682 if (mem->mem_type == TTM_PL_SYSTEM) 1683 return false; 1684 1685 if (man->flags & TTM_MEMTYPE_FLAG_CMA) 1686 return false; 1687 1688 if (mem->placement & TTM_PL_FLAG_CACHED) 1689 return false; 1690 } 1691 return true; 1692 } 1693 1694 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) 1695 { 1696 struct ttm_bo_device *bdev = bo->bdev; 1697 1698 drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping); 1699 ttm_mem_io_free_vm(bo); 1700 } 1701 1702 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) 1703 { 1704 struct ttm_bo_device *bdev = bo->bdev; 1705 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 1706 1707 ttm_mem_io_lock(man, false); 1708 ttm_bo_unmap_virtual_locked(bo); 1709 ttm_mem_io_unlock(man); 1710 } 1711 1712 1713 EXPORT_SYMBOL(ttm_bo_unmap_virtual); 1714 1715 int ttm_bo_wait(struct ttm_buffer_object *bo, 1716 bool interruptible, bool no_wait) 1717 { 1718 long timeout = 15 * HZ; 1719 1720 if (no_wait) { 1721 if (reservation_object_test_signaled_rcu(bo->resv, true)) 1722 return 0; 1723 else 1724 return -EBUSY; 1725 } 1726 1727 timeout = reservation_object_wait_timeout_rcu(bo->resv, true, 1728 interruptible, timeout); 1729 if (timeout < 0) 1730 return timeout; 1731 1732 if (timeout == 0) 1733 return -EBUSY; 1734 1735 reservation_object_add_excl_fence(bo->resv, NULL); 1736 return 0; 1737 } 1738 EXPORT_SYMBOL(ttm_bo_wait); 1739 1740 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) 1741 { 1742 int ret = 0; 1743 1744 /* 1745 * Using ttm_bo_reserve makes sure the lru lists are updated. 1746 */ 1747 1748 ret = ttm_bo_reserve(bo, true, no_wait, NULL); 1749 if (unlikely(ret != 0)) 1750 return ret; 1751 ret = ttm_bo_wait(bo, true, no_wait); 1752 if (likely(ret == 0)) 1753 atomic_inc(&bo->cpu_writers); 1754 ttm_bo_unreserve(bo); 1755 return ret; 1756 } 1757 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab); 1758 1759 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) 1760 { 1761 atomic_dec(&bo->cpu_writers); 1762 } 1763 EXPORT_SYMBOL(ttm_bo_synccpu_write_release); 1764 1765 /** 1766 * A buffer object shrink method that tries to swap out the first 1767 * buffer object on the bo_global::swap_lru list. 1768 */ 1769 int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx) 1770 { 1771 struct ttm_buffer_object *bo; 1772 int ret = -EBUSY; 1773 bool locked; 1774 unsigned i; 1775 1776 spin_lock(&glob->lru_lock); 1777 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 1778 list_for_each_entry(bo, &glob->swap_lru[i], swap) { 1779 if (ttm_bo_evict_swapout_allowable(bo, ctx, &locked)) { 1780 ret = 0; 1781 break; 1782 } 1783 } 1784 if (!ret) 1785 break; 1786 } 1787 1788 if (ret) { 1789 spin_unlock(&glob->lru_lock); 1790 return ret; 1791 } 1792 1793 kref_get(&bo->list_kref); 1794 1795 if (!list_empty(&bo->ddestroy)) { 1796 ret = ttm_bo_cleanup_refs(bo, false, false, locked); 1797 kref_put(&bo->list_kref, ttm_bo_release_list); 1798 return ret; 1799 } 1800 1801 ttm_bo_del_from_lru(bo); 1802 spin_unlock(&glob->lru_lock); 1803 1804 /** 1805 * Move to system cached 1806 */ 1807 1808 if (bo->mem.mem_type != TTM_PL_SYSTEM || 1809 bo->ttm->caching_state != tt_cached) { 1810 struct ttm_operation_ctx ctx = { false, false }; 1811 struct ttm_mem_reg evict_mem; 1812 1813 evict_mem = bo->mem; 1814 evict_mem.mm_node = NULL; 1815 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; 1816 evict_mem.mem_type = TTM_PL_SYSTEM; 1817 1818 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx); 1819 if (unlikely(ret != 0)) 1820 goto out; 1821 } 1822 1823 /** 1824 * Make sure BO is idle. 1825 */ 1826 1827 ret = ttm_bo_wait(bo, false, false); 1828 if (unlikely(ret != 0)) 1829 goto out; 1830 1831 ttm_bo_unmap_virtual(bo); 1832 1833 /** 1834 * Swap out. Buffer will be swapped in again as soon as 1835 * anyone tries to access a ttm page. 1836 */ 1837 1838 if (bo->bdev->driver->swap_notify) 1839 bo->bdev->driver->swap_notify(bo); 1840 1841 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); 1842 out: 1843 1844 /** 1845 * 1846 * Unreserve without putting on LRU to avoid swapping out an 1847 * already swapped buffer. 1848 */ 1849 if (locked) 1850 reservation_object_unlock(bo->resv); 1851 kref_put(&bo->list_kref, ttm_bo_release_list); 1852 return ret; 1853 } 1854 EXPORT_SYMBOL(ttm_bo_swapout); 1855 1856 void ttm_bo_swapout_all(struct ttm_bo_device *bdev) 1857 { 1858 struct ttm_operation_ctx ctx = { 1859 .interruptible = false, 1860 .no_wait_gpu = false 1861 }; 1862 1863 while (ttm_bo_swapout(bdev->glob, &ctx) == 0) 1864 ; 1865 } 1866 EXPORT_SYMBOL(ttm_bo_swapout_all); 1867 1868 /** 1869 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become 1870 * unreserved 1871 * 1872 * @bo: Pointer to buffer 1873 */ 1874 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo) 1875 { 1876 int ret; 1877 1878 /* 1879 * In the absense of a wait_unlocked API, 1880 * Use the bo::wu_mutex to avoid triggering livelocks due to 1881 * concurrent use of this function. Note that this use of 1882 * bo::wu_mutex can go away if we change locking order to 1883 * mmap_sem -> bo::reserve. 1884 */ 1885 ret = mutex_lock_interruptible(&bo->wu_mutex); 1886 if (unlikely(ret != 0)) 1887 return -ERESTARTSYS; 1888 if (!ww_mutex_is_locked(&bo->resv->lock)) 1889 goto out_unlock; 1890 ret = reservation_object_lock_interruptible(bo->resv, NULL); 1891 if (ret == -EINTR) 1892 ret = -ERESTARTSYS; 1893 if (unlikely(ret != 0)) 1894 goto out_unlock; 1895 reservation_object_unlock(bo->resv); 1896 1897 out_unlock: 1898 mutex_unlock(&bo->wu_mutex); 1899 return ret; 1900 } 1901