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/dma-resv.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 unsigned ttm_bo_glob_use_count; 53 struct ttm_bo_global ttm_bo_glob; 54 55 static struct attribute ttm_bo_count = { 56 .name = "bo_count", 57 .mode = S_IRUGO 58 }; 59 60 /* default destructor */ 61 static void ttm_bo_default_destroy(struct ttm_buffer_object *bo) 62 { 63 kfree(bo); 64 } 65 66 static inline int ttm_mem_type_from_place(const struct ttm_place *place, 67 uint32_t *mem_type) 68 { 69 int pos; 70 71 pos = ffs(place->flags & TTM_PL_MASK_MEM); 72 if (unlikely(!pos)) 73 return -EINVAL; 74 75 *mem_type = pos - 1; 76 return 0; 77 } 78 79 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, struct drm_printer *p, 80 int mem_type) 81 { 82 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 83 84 drm_printf(p, " has_type: %d\n", man->has_type); 85 drm_printf(p, " use_type: %d\n", man->use_type); 86 drm_printf(p, " flags: 0x%08X\n", man->flags); 87 drm_printf(p, " gpu_offset: 0x%08llX\n", man->gpu_offset); 88 drm_printf(p, " size: %llu\n", man->size); 89 drm_printf(p, " available_caching: 0x%08X\n", man->available_caching); 90 drm_printf(p, " default_caching: 0x%08X\n", man->default_caching); 91 if (mem_type != TTM_PL_SYSTEM) 92 (*man->func->debug)(man, p); 93 } 94 95 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo, 96 struct ttm_placement *placement) 97 { 98 struct drm_printer p = drm_debug_printer(TTM_PFX); 99 int i, ret, mem_type; 100 101 drm_printf(&p, "No space for %p (%lu pages, %luK, %luM)\n", 102 bo, bo->mem.num_pages, bo->mem.size >> 10, 103 bo->mem.size >> 20); 104 for (i = 0; i < placement->num_placement; i++) { 105 ret = ttm_mem_type_from_place(&placement->placement[i], 106 &mem_type); 107 if (ret) 108 return; 109 drm_printf(&p, " placement[%d]=0x%08X (%d)\n", 110 i, placement->placement[i].flags, mem_type); 111 ttm_mem_type_debug(bo->bdev, &p, mem_type); 112 } 113 } 114 115 static ssize_t ttm_bo_global_show(struct kobject *kobj, 116 struct attribute *attr, 117 char *buffer) 118 { 119 struct ttm_bo_global *glob = 120 container_of(kobj, struct ttm_bo_global, kobj); 121 122 return snprintf(buffer, PAGE_SIZE, "%d\n", 123 atomic_read(&glob->bo_count)); 124 } 125 126 static struct attribute *ttm_bo_global_attrs[] = { 127 &ttm_bo_count, 128 NULL 129 }; 130 131 static const struct sysfs_ops ttm_bo_global_ops = { 132 .show = &ttm_bo_global_show 133 }; 134 135 static struct kobj_type ttm_bo_glob_kobj_type = { 136 .release = &ttm_bo_global_kobj_release, 137 .sysfs_ops = &ttm_bo_global_ops, 138 .default_attrs = ttm_bo_global_attrs 139 }; 140 141 142 static inline uint32_t ttm_bo_type_flags(unsigned type) 143 { 144 return 1 << (type); 145 } 146 147 static void ttm_bo_release_list(struct kref *list_kref) 148 { 149 struct ttm_buffer_object *bo = 150 container_of(list_kref, struct ttm_buffer_object, list_kref); 151 struct ttm_bo_device *bdev = bo->bdev; 152 size_t acc_size = bo->acc_size; 153 154 BUG_ON(kref_read(&bo->list_kref)); 155 BUG_ON(kref_read(&bo->kref)); 156 BUG_ON(atomic_read(&bo->cpu_writers)); 157 BUG_ON(bo->mem.mm_node != NULL); 158 BUG_ON(!list_empty(&bo->lru)); 159 BUG_ON(!list_empty(&bo->ddestroy)); 160 ttm_tt_destroy(bo->ttm); 161 atomic_dec(&bo->bdev->glob->bo_count); 162 dma_fence_put(bo->moving); 163 if (!ttm_bo_uses_embedded_gem_object(bo)) 164 dma_resv_fini(&bo->base._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 static void ttm_bo_add_mem_to_lru(struct ttm_buffer_object *bo, 171 struct ttm_mem_reg *mem) 172 { 173 struct ttm_bo_device *bdev = bo->bdev; 174 struct ttm_mem_type_manager *man; 175 176 dma_resv_assert_held(bo->base.resv); 177 178 if (!list_empty(&bo->lru)) 179 return; 180 181 if (mem->placement & TTM_PL_FLAG_NO_EVICT) 182 return; 183 184 man = &bdev->man[mem->mem_type]; 185 list_add_tail(&bo->lru, &man->lru[bo->priority]); 186 kref_get(&bo->list_kref); 187 188 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm && 189 !(bo->ttm->page_flags & (TTM_PAGE_FLAG_SG | 190 TTM_PAGE_FLAG_SWAPPED))) { 191 list_add_tail(&bo->swap, &bdev->glob->swap_lru[bo->priority]); 192 kref_get(&bo->list_kref); 193 } 194 } 195 196 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo) 197 { 198 ttm_bo_add_mem_to_lru(bo, &bo->mem); 199 } 200 EXPORT_SYMBOL(ttm_bo_add_to_lru); 201 202 static void ttm_bo_ref_bug(struct kref *list_kref) 203 { 204 BUG(); 205 } 206 207 void ttm_bo_del_from_lru(struct ttm_buffer_object *bo) 208 { 209 struct ttm_bo_device *bdev = bo->bdev; 210 bool notify = false; 211 212 if (!list_empty(&bo->swap)) { 213 list_del_init(&bo->swap); 214 kref_put(&bo->list_kref, ttm_bo_ref_bug); 215 notify = true; 216 } 217 if (!list_empty(&bo->lru)) { 218 list_del_init(&bo->lru); 219 kref_put(&bo->list_kref, ttm_bo_ref_bug); 220 notify = true; 221 } 222 223 if (notify && bdev->driver->del_from_lru_notify) 224 bdev->driver->del_from_lru_notify(bo); 225 } 226 227 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo) 228 { 229 struct ttm_bo_global *glob = bo->bdev->glob; 230 231 spin_lock(&glob->lru_lock); 232 ttm_bo_del_from_lru(bo); 233 spin_unlock(&glob->lru_lock); 234 } 235 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru); 236 237 static void ttm_bo_bulk_move_set_pos(struct ttm_lru_bulk_move_pos *pos, 238 struct ttm_buffer_object *bo) 239 { 240 if (!pos->first) 241 pos->first = bo; 242 pos->last = bo; 243 } 244 245 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo, 246 struct ttm_lru_bulk_move *bulk) 247 { 248 dma_resv_assert_held(bo->base.resv); 249 250 ttm_bo_del_from_lru(bo); 251 ttm_bo_add_to_lru(bo); 252 253 if (bulk && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 254 switch (bo->mem.mem_type) { 255 case TTM_PL_TT: 256 ttm_bo_bulk_move_set_pos(&bulk->tt[bo->priority], bo); 257 break; 258 259 case TTM_PL_VRAM: 260 ttm_bo_bulk_move_set_pos(&bulk->vram[bo->priority], bo); 261 break; 262 } 263 if (bo->ttm && !(bo->ttm->page_flags & 264 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED))) 265 ttm_bo_bulk_move_set_pos(&bulk->swap[bo->priority], bo); 266 } 267 } 268 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail); 269 270 void ttm_bo_bulk_move_lru_tail(struct ttm_lru_bulk_move *bulk) 271 { 272 unsigned i; 273 274 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 275 struct ttm_lru_bulk_move_pos *pos = &bulk->tt[i]; 276 struct ttm_mem_type_manager *man; 277 278 if (!pos->first) 279 continue; 280 281 dma_resv_assert_held(pos->first->base.resv); 282 dma_resv_assert_held(pos->last->base.resv); 283 284 man = &pos->first->bdev->man[TTM_PL_TT]; 285 list_bulk_move_tail(&man->lru[i], &pos->first->lru, 286 &pos->last->lru); 287 } 288 289 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 290 struct ttm_lru_bulk_move_pos *pos = &bulk->vram[i]; 291 struct ttm_mem_type_manager *man; 292 293 if (!pos->first) 294 continue; 295 296 dma_resv_assert_held(pos->first->base.resv); 297 dma_resv_assert_held(pos->last->base.resv); 298 299 man = &pos->first->bdev->man[TTM_PL_VRAM]; 300 list_bulk_move_tail(&man->lru[i], &pos->first->lru, 301 &pos->last->lru); 302 } 303 304 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 305 struct ttm_lru_bulk_move_pos *pos = &bulk->swap[i]; 306 struct list_head *lru; 307 308 if (!pos->first) 309 continue; 310 311 dma_resv_assert_held(pos->first->base.resv); 312 dma_resv_assert_held(pos->last->base.resv); 313 314 lru = &pos->first->bdev->glob->swap_lru[i]; 315 list_bulk_move_tail(lru, &pos->first->swap, &pos->last->swap); 316 } 317 } 318 EXPORT_SYMBOL(ttm_bo_bulk_move_lru_tail); 319 320 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo, 321 struct ttm_mem_reg *mem, bool evict, 322 struct ttm_operation_ctx *ctx) 323 { 324 struct ttm_bo_device *bdev = bo->bdev; 325 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem); 326 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem); 327 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type]; 328 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type]; 329 int ret = 0; 330 331 if (old_is_pci || new_is_pci || 332 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) { 333 ret = ttm_mem_io_lock(old_man, true); 334 if (unlikely(ret != 0)) 335 goto out_err; 336 ttm_bo_unmap_virtual_locked(bo); 337 ttm_mem_io_unlock(old_man); 338 } 339 340 /* 341 * Create and bind a ttm if required. 342 */ 343 344 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 345 if (bo->ttm == NULL) { 346 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED); 347 ret = ttm_tt_create(bo, zero); 348 if (ret) 349 goto out_err; 350 } 351 352 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement); 353 if (ret) 354 goto out_err; 355 356 if (mem->mem_type != TTM_PL_SYSTEM) { 357 ret = ttm_tt_bind(bo->ttm, mem, ctx); 358 if (ret) 359 goto out_err; 360 } 361 362 if (bo->mem.mem_type == TTM_PL_SYSTEM) { 363 if (bdev->driver->move_notify) 364 bdev->driver->move_notify(bo, evict, mem); 365 bo->mem = *mem; 366 mem->mm_node = NULL; 367 goto moved; 368 } 369 } 370 371 if (bdev->driver->move_notify) 372 bdev->driver->move_notify(bo, evict, mem); 373 374 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) && 375 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) 376 ret = ttm_bo_move_ttm(bo, ctx, mem); 377 else if (bdev->driver->move) 378 ret = bdev->driver->move(bo, evict, ctx, mem); 379 else 380 ret = ttm_bo_move_memcpy(bo, ctx, mem); 381 382 if (ret) { 383 if (bdev->driver->move_notify) { 384 swap(*mem, bo->mem); 385 bdev->driver->move_notify(bo, false, mem); 386 swap(*mem, bo->mem); 387 } 388 389 goto out_err; 390 } 391 392 moved: 393 if (bo->evicted) { 394 if (bdev->driver->invalidate_caches) { 395 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement); 396 if (ret) 397 pr_err("Can not flush read caches\n"); 398 } 399 bo->evicted = false; 400 } 401 402 if (bo->mem.mm_node) 403 bo->offset = (bo->mem.start << PAGE_SHIFT) + 404 bdev->man[bo->mem.mem_type].gpu_offset; 405 else 406 bo->offset = 0; 407 408 ctx->bytes_moved += bo->num_pages << PAGE_SHIFT; 409 return 0; 410 411 out_err: 412 new_man = &bdev->man[bo->mem.mem_type]; 413 if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) { 414 ttm_tt_destroy(bo->ttm); 415 bo->ttm = NULL; 416 } 417 418 return ret; 419 } 420 421 /** 422 * Call bo::reserved. 423 * Will release GPU memory type usage on destruction. 424 * This is the place to put in driver specific hooks to release 425 * driver private resources. 426 * Will release the bo::reserved lock. 427 */ 428 429 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo) 430 { 431 if (bo->bdev->driver->move_notify) 432 bo->bdev->driver->move_notify(bo, false, NULL); 433 434 ttm_tt_destroy(bo->ttm); 435 bo->ttm = NULL; 436 ttm_bo_mem_put(bo, &bo->mem); 437 } 438 439 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo) 440 { 441 int r; 442 443 if (bo->base.resv == &bo->base._resv) 444 return 0; 445 446 BUG_ON(!dma_resv_trylock(&bo->base._resv)); 447 448 r = dma_resv_copy_fences(&bo->base._resv, bo->base.resv); 449 if (r) 450 dma_resv_unlock(&bo->base._resv); 451 452 return r; 453 } 454 455 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo) 456 { 457 struct dma_resv_list *fobj; 458 struct dma_fence *fence; 459 int i; 460 461 fobj = dma_resv_get_list(&bo->base._resv); 462 fence = dma_resv_get_excl(&bo->base._resv); 463 if (fence && !fence->ops->signaled) 464 dma_fence_enable_sw_signaling(fence); 465 466 for (i = 0; fobj && i < fobj->shared_count; ++i) { 467 fence = rcu_dereference_protected(fobj->shared[i], 468 dma_resv_held(bo->base.resv)); 469 470 if (!fence->ops->signaled) 471 dma_fence_enable_sw_signaling(fence); 472 } 473 } 474 475 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo) 476 { 477 struct ttm_bo_device *bdev = bo->bdev; 478 struct ttm_bo_global *glob = bdev->glob; 479 int ret; 480 481 ret = ttm_bo_individualize_resv(bo); 482 if (ret) { 483 /* Last resort, if we fail to allocate memory for the 484 * fences block for the BO to become idle 485 */ 486 dma_resv_wait_timeout_rcu(bo->base.resv, true, false, 487 30 * HZ); 488 spin_lock(&glob->lru_lock); 489 goto error; 490 } 491 492 spin_lock(&glob->lru_lock); 493 ret = dma_resv_trylock(bo->base.resv) ? 0 : -EBUSY; 494 if (!ret) { 495 if (dma_resv_test_signaled_rcu(&bo->base._resv, true)) { 496 ttm_bo_del_from_lru(bo); 497 spin_unlock(&glob->lru_lock); 498 if (bo->base.resv != &bo->base._resv) 499 dma_resv_unlock(&bo->base._resv); 500 501 ttm_bo_cleanup_memtype_use(bo); 502 dma_resv_unlock(bo->base.resv); 503 return; 504 } 505 506 ttm_bo_flush_all_fences(bo); 507 508 /* 509 * Make NO_EVICT bos immediately available to 510 * shrinkers, now that they are queued for 511 * destruction. 512 */ 513 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) { 514 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT; 515 ttm_bo_add_to_lru(bo); 516 } 517 518 dma_resv_unlock(bo->base.resv); 519 } 520 if (bo->base.resv != &bo->base._resv) 521 dma_resv_unlock(&bo->base._resv); 522 523 error: 524 kref_get(&bo->list_kref); 525 list_add_tail(&bo->ddestroy, &bdev->ddestroy); 526 spin_unlock(&glob->lru_lock); 527 528 schedule_delayed_work(&bdev->wq, 529 ((HZ / 100) < 1) ? 1 : HZ / 100); 530 } 531 532 /** 533 * function ttm_bo_cleanup_refs 534 * If bo idle, remove from delayed- and lru lists, and unref. 535 * If not idle, do nothing. 536 * 537 * Must be called with lru_lock and reservation held, this function 538 * will drop the lru lock and optionally the reservation lock before returning. 539 * 540 * @interruptible Any sleeps should occur interruptibly. 541 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead. 542 * @unlock_resv Unlock the reservation lock as well. 543 */ 544 545 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo, 546 bool interruptible, bool no_wait_gpu, 547 bool unlock_resv) 548 { 549 struct ttm_bo_global *glob = bo->bdev->glob; 550 struct dma_resv *resv; 551 int ret; 552 553 if (unlikely(list_empty(&bo->ddestroy))) 554 resv = bo->base.resv; 555 else 556 resv = &bo->base._resv; 557 558 if (dma_resv_test_signaled_rcu(resv, true)) 559 ret = 0; 560 else 561 ret = -EBUSY; 562 563 if (ret && !no_wait_gpu) { 564 long lret; 565 566 if (unlock_resv) 567 dma_resv_unlock(bo->base.resv); 568 spin_unlock(&glob->lru_lock); 569 570 lret = dma_resv_wait_timeout_rcu(resv, true, 571 interruptible, 572 30 * HZ); 573 574 if (lret < 0) 575 return lret; 576 else if (lret == 0) 577 return -EBUSY; 578 579 spin_lock(&glob->lru_lock); 580 if (unlock_resv && !dma_resv_trylock(bo->base.resv)) { 581 /* 582 * We raced, and lost, someone else holds the reservation now, 583 * and is probably busy in ttm_bo_cleanup_memtype_use. 584 * 585 * Even if it's not the case, because we finished waiting any 586 * delayed destruction would succeed, so just return success 587 * here. 588 */ 589 spin_unlock(&glob->lru_lock); 590 return 0; 591 } 592 ret = 0; 593 } 594 595 if (ret || unlikely(list_empty(&bo->ddestroy))) { 596 if (unlock_resv) 597 dma_resv_unlock(bo->base.resv); 598 spin_unlock(&glob->lru_lock); 599 return ret; 600 } 601 602 ttm_bo_del_from_lru(bo); 603 list_del_init(&bo->ddestroy); 604 kref_put(&bo->list_kref, ttm_bo_ref_bug); 605 606 spin_unlock(&glob->lru_lock); 607 ttm_bo_cleanup_memtype_use(bo); 608 609 if (unlock_resv) 610 dma_resv_unlock(bo->base.resv); 611 612 return 0; 613 } 614 615 /** 616 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all 617 * encountered buffers. 618 */ 619 static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all) 620 { 621 struct ttm_bo_global *glob = bdev->glob; 622 struct list_head removed; 623 bool empty; 624 625 INIT_LIST_HEAD(&removed); 626 627 spin_lock(&glob->lru_lock); 628 while (!list_empty(&bdev->ddestroy)) { 629 struct ttm_buffer_object *bo; 630 631 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object, 632 ddestroy); 633 kref_get(&bo->list_kref); 634 list_move_tail(&bo->ddestroy, &removed); 635 636 if (remove_all || bo->base.resv != &bo->base._resv) { 637 spin_unlock(&glob->lru_lock); 638 dma_resv_lock(bo->base.resv, NULL); 639 640 spin_lock(&glob->lru_lock); 641 ttm_bo_cleanup_refs(bo, false, !remove_all, true); 642 643 } else if (dma_resv_trylock(bo->base.resv)) { 644 ttm_bo_cleanup_refs(bo, false, !remove_all, true); 645 } else { 646 spin_unlock(&glob->lru_lock); 647 } 648 649 kref_put(&bo->list_kref, ttm_bo_release_list); 650 spin_lock(&glob->lru_lock); 651 } 652 list_splice_tail(&removed, &bdev->ddestroy); 653 empty = list_empty(&bdev->ddestroy); 654 spin_unlock(&glob->lru_lock); 655 656 return empty; 657 } 658 659 static void ttm_bo_delayed_workqueue(struct work_struct *work) 660 { 661 struct ttm_bo_device *bdev = 662 container_of(work, struct ttm_bo_device, wq.work); 663 664 if (!ttm_bo_delayed_delete(bdev, false)) 665 schedule_delayed_work(&bdev->wq, 666 ((HZ / 100) < 1) ? 1 : HZ / 100); 667 } 668 669 static void ttm_bo_release(struct kref *kref) 670 { 671 struct ttm_buffer_object *bo = 672 container_of(kref, struct ttm_buffer_object, kref); 673 struct ttm_bo_device *bdev = bo->bdev; 674 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 675 676 if (bo->bdev->driver->release_notify) 677 bo->bdev->driver->release_notify(bo); 678 679 drm_vma_offset_remove(&bdev->vma_manager, &bo->base.vma_node); 680 ttm_mem_io_lock(man, false); 681 ttm_mem_io_free_vm(bo); 682 ttm_mem_io_unlock(man); 683 ttm_bo_cleanup_refs_or_queue(bo); 684 kref_put(&bo->list_kref, ttm_bo_release_list); 685 } 686 687 void ttm_bo_put(struct ttm_buffer_object *bo) 688 { 689 kref_put(&bo->kref, ttm_bo_release); 690 } 691 EXPORT_SYMBOL(ttm_bo_put); 692 693 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) 694 { 695 return cancel_delayed_work_sync(&bdev->wq); 696 } 697 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue); 698 699 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched) 700 { 701 if (resched) 702 schedule_delayed_work(&bdev->wq, 703 ((HZ / 100) < 1) ? 1 : HZ / 100); 704 } 705 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue); 706 707 static int ttm_bo_evict(struct ttm_buffer_object *bo, 708 struct ttm_operation_ctx *ctx) 709 { 710 struct ttm_bo_device *bdev = bo->bdev; 711 struct ttm_mem_reg evict_mem; 712 struct ttm_placement placement; 713 int ret = 0; 714 715 dma_resv_assert_held(bo->base.resv); 716 717 placement.num_placement = 0; 718 placement.num_busy_placement = 0; 719 bdev->driver->evict_flags(bo, &placement); 720 721 if (!placement.num_placement && !placement.num_busy_placement) { 722 ret = ttm_bo_pipeline_gutting(bo); 723 if (ret) 724 return ret; 725 726 return ttm_tt_create(bo, false); 727 } 728 729 evict_mem = bo->mem; 730 evict_mem.mm_node = NULL; 731 evict_mem.bus.io_reserved_vm = false; 732 evict_mem.bus.io_reserved_count = 0; 733 734 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx); 735 if (ret) { 736 if (ret != -ERESTARTSYS) { 737 pr_err("Failed to find memory space for buffer 0x%p eviction\n", 738 bo); 739 ttm_bo_mem_space_debug(bo, &placement); 740 } 741 goto out; 742 } 743 744 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx); 745 if (unlikely(ret)) { 746 if (ret != -ERESTARTSYS) 747 pr_err("Buffer eviction failed\n"); 748 ttm_bo_mem_put(bo, &evict_mem); 749 goto out; 750 } 751 bo->evicted = true; 752 out: 753 return ret; 754 } 755 756 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo, 757 const struct ttm_place *place) 758 { 759 /* Don't evict this BO if it's outside of the 760 * requested placement range 761 */ 762 if (place->fpfn >= (bo->mem.start + bo->mem.size) || 763 (place->lpfn && place->lpfn <= bo->mem.start)) 764 return false; 765 766 return true; 767 } 768 EXPORT_SYMBOL(ttm_bo_eviction_valuable); 769 770 /** 771 * Check the target bo is allowable to be evicted or swapout, including cases: 772 * 773 * a. if share same reservation object with ctx->resv, have assumption 774 * reservation objects should already be locked, so not lock again and 775 * return true directly when either the opreation allow_reserved_eviction 776 * or the target bo already is in delayed free list; 777 * 778 * b. Otherwise, trylock it. 779 */ 780 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo, 781 struct ttm_operation_ctx *ctx, bool *locked, bool *busy) 782 { 783 bool ret = false; 784 785 if (bo->base.resv == ctx->resv) { 786 dma_resv_assert_held(bo->base.resv); 787 if (ctx->flags & TTM_OPT_FLAG_ALLOW_RES_EVICT 788 || !list_empty(&bo->ddestroy)) 789 ret = true; 790 *locked = false; 791 if (busy) 792 *busy = false; 793 } else { 794 ret = dma_resv_trylock(bo->base.resv); 795 *locked = ret; 796 if (busy) 797 *busy = !ret; 798 } 799 800 return ret; 801 } 802 803 /** 804 * ttm_mem_evict_wait_busy - wait for a busy BO to become available 805 * 806 * @busy_bo: BO which couldn't be locked with trylock 807 * @ctx: operation context 808 * @ticket: acquire ticket 809 * 810 * Try to lock a busy buffer object to avoid failing eviction. 811 */ 812 static int ttm_mem_evict_wait_busy(struct ttm_buffer_object *busy_bo, 813 struct ttm_operation_ctx *ctx, 814 struct ww_acquire_ctx *ticket) 815 { 816 int r; 817 818 if (!busy_bo || !ticket) 819 return -EBUSY; 820 821 if (ctx->interruptible) 822 r = dma_resv_lock_interruptible(busy_bo->base.resv, 823 ticket); 824 else 825 r = dma_resv_lock(busy_bo->base.resv, ticket); 826 827 /* 828 * TODO: It would be better to keep the BO locked until allocation is at 829 * least tried one more time, but that would mean a much larger rework 830 * of TTM. 831 */ 832 if (!r) 833 dma_resv_unlock(busy_bo->base.resv); 834 835 return r == -EDEADLK ? -EBUSY : r; 836 } 837 838 static int ttm_mem_evict_first(struct ttm_bo_device *bdev, 839 uint32_t mem_type, 840 const struct ttm_place *place, 841 struct ttm_operation_ctx *ctx, 842 struct ww_acquire_ctx *ticket) 843 { 844 struct ttm_buffer_object *bo = NULL, *busy_bo = NULL; 845 struct ttm_bo_global *glob = bdev->glob; 846 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 847 bool locked = false; 848 unsigned i; 849 int ret; 850 851 spin_lock(&glob->lru_lock); 852 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 853 list_for_each_entry(bo, &man->lru[i], lru) { 854 bool busy; 855 856 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked, 857 &busy)) { 858 if (busy && !busy_bo && ticket != 859 dma_resv_locking_ctx(bo->base.resv)) 860 busy_bo = bo; 861 continue; 862 } 863 864 if (place && !bdev->driver->eviction_valuable(bo, 865 place)) { 866 if (locked) 867 dma_resv_unlock(bo->base.resv); 868 continue; 869 } 870 break; 871 } 872 873 /* If the inner loop terminated early, we have our candidate */ 874 if (&bo->lru != &man->lru[i]) 875 break; 876 877 bo = NULL; 878 } 879 880 if (!bo) { 881 if (busy_bo) 882 kref_get(&busy_bo->list_kref); 883 spin_unlock(&glob->lru_lock); 884 ret = ttm_mem_evict_wait_busy(busy_bo, ctx, ticket); 885 if (busy_bo) 886 kref_put(&busy_bo->list_kref, ttm_bo_release_list); 887 return ret; 888 } 889 890 kref_get(&bo->list_kref); 891 892 if (!list_empty(&bo->ddestroy)) { 893 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible, 894 ctx->no_wait_gpu, locked); 895 kref_put(&bo->list_kref, ttm_bo_release_list); 896 return ret; 897 } 898 899 ttm_bo_del_from_lru(bo); 900 spin_unlock(&glob->lru_lock); 901 902 ret = ttm_bo_evict(bo, ctx); 903 if (locked) { 904 ttm_bo_unreserve(bo); 905 } else { 906 spin_lock(&glob->lru_lock); 907 ttm_bo_add_to_lru(bo); 908 spin_unlock(&glob->lru_lock); 909 } 910 911 kref_put(&bo->list_kref, ttm_bo_release_list); 912 return ret; 913 } 914 915 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem) 916 { 917 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type]; 918 919 if (mem->mm_node) 920 (*man->func->put_node)(man, mem); 921 } 922 EXPORT_SYMBOL(ttm_bo_mem_put); 923 924 /** 925 * Add the last move fence to the BO and reserve a new shared slot. 926 */ 927 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo, 928 struct ttm_mem_type_manager *man, 929 struct ttm_mem_reg *mem) 930 { 931 struct dma_fence *fence; 932 int ret; 933 934 spin_lock(&man->move_lock); 935 fence = dma_fence_get(man->move); 936 spin_unlock(&man->move_lock); 937 938 if (fence) { 939 dma_resv_add_shared_fence(bo->base.resv, fence); 940 941 ret = dma_resv_reserve_shared(bo->base.resv, 1); 942 if (unlikely(ret)) { 943 dma_fence_put(fence); 944 return ret; 945 } 946 947 dma_fence_put(bo->moving); 948 bo->moving = fence; 949 } 950 951 return 0; 952 } 953 954 /** 955 * Repeatedly evict memory from the LRU for @mem_type until we create enough 956 * space, or we've evicted everything and there isn't enough space. 957 */ 958 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo, 959 const struct ttm_place *place, 960 struct ttm_mem_reg *mem, 961 struct ttm_operation_ctx *ctx) 962 { 963 struct ttm_bo_device *bdev = bo->bdev; 964 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 965 struct ww_acquire_ctx *ticket; 966 int ret; 967 968 ticket = dma_resv_locking_ctx(bo->base.resv); 969 do { 970 ret = (*man->func->get_node)(man, bo, place, mem); 971 if (unlikely(ret != 0)) 972 return ret; 973 if (mem->mm_node) 974 break; 975 ret = ttm_mem_evict_first(bdev, mem->mem_type, place, ctx, 976 ticket); 977 if (unlikely(ret != 0)) 978 return ret; 979 } while (1); 980 981 return ttm_bo_add_move_fence(bo, man, mem); 982 } 983 984 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man, 985 uint32_t cur_placement, 986 uint32_t proposed_placement) 987 { 988 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING; 989 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING; 990 991 /** 992 * Keep current caching if possible. 993 */ 994 995 if ((cur_placement & caching) != 0) 996 result |= (cur_placement & caching); 997 else if ((man->default_caching & caching) != 0) 998 result |= man->default_caching; 999 else if ((TTM_PL_FLAG_CACHED & caching) != 0) 1000 result |= TTM_PL_FLAG_CACHED; 1001 else if ((TTM_PL_FLAG_WC & caching) != 0) 1002 result |= TTM_PL_FLAG_WC; 1003 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0) 1004 result |= TTM_PL_FLAG_UNCACHED; 1005 1006 return result; 1007 } 1008 1009 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man, 1010 uint32_t mem_type, 1011 const struct ttm_place *place, 1012 uint32_t *masked_placement) 1013 { 1014 uint32_t cur_flags = ttm_bo_type_flags(mem_type); 1015 1016 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0) 1017 return false; 1018 1019 if ((place->flags & man->available_caching) == 0) 1020 return false; 1021 1022 cur_flags |= (place->flags & man->available_caching); 1023 1024 *masked_placement = cur_flags; 1025 return true; 1026 } 1027 1028 /** 1029 * ttm_bo_mem_placement - check if placement is compatible 1030 * @bo: BO to find memory for 1031 * @place: where to search 1032 * @mem: the memory object to fill in 1033 * @ctx: operation context 1034 * 1035 * Check if placement is compatible and fill in mem structure. 1036 * Returns -EBUSY if placement won't work or negative error code. 1037 * 0 when placement can be used. 1038 */ 1039 static int ttm_bo_mem_placement(struct ttm_buffer_object *bo, 1040 const struct ttm_place *place, 1041 struct ttm_mem_reg *mem, 1042 struct ttm_operation_ctx *ctx) 1043 { 1044 struct ttm_bo_device *bdev = bo->bdev; 1045 uint32_t mem_type = TTM_PL_SYSTEM; 1046 struct ttm_mem_type_manager *man; 1047 uint32_t cur_flags = 0; 1048 int ret; 1049 1050 ret = ttm_mem_type_from_place(place, &mem_type); 1051 if (ret) 1052 return ret; 1053 1054 man = &bdev->man[mem_type]; 1055 if (!man->has_type || !man->use_type) 1056 return -EBUSY; 1057 1058 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags)) 1059 return -EBUSY; 1060 1061 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, cur_flags); 1062 /* 1063 * Use the access and other non-mapping-related flag bits from 1064 * the memory placement flags to the current flags 1065 */ 1066 ttm_flag_masked(&cur_flags, place->flags, ~TTM_PL_MASK_MEMTYPE); 1067 1068 mem->mem_type = mem_type; 1069 mem->placement = cur_flags; 1070 1071 if (bo->mem.mem_type < mem_type && !list_empty(&bo->lru)) { 1072 spin_lock(&bo->bdev->glob->lru_lock); 1073 ttm_bo_del_from_lru(bo); 1074 ttm_bo_add_mem_to_lru(bo, mem); 1075 spin_unlock(&bo->bdev->glob->lru_lock); 1076 } 1077 1078 return 0; 1079 } 1080 1081 /** 1082 * Creates space for memory region @mem according to its type. 1083 * 1084 * This function first searches for free space in compatible memory types in 1085 * the priority order defined by the driver. If free space isn't found, then 1086 * ttm_bo_mem_force_space is attempted in priority order to evict and find 1087 * space. 1088 */ 1089 int ttm_bo_mem_space(struct ttm_buffer_object *bo, 1090 struct ttm_placement *placement, 1091 struct ttm_mem_reg *mem, 1092 struct ttm_operation_ctx *ctx) 1093 { 1094 struct ttm_bo_device *bdev = bo->bdev; 1095 bool type_found = false; 1096 int i, ret; 1097 1098 ret = dma_resv_reserve_shared(bo->base.resv, 1); 1099 if (unlikely(ret)) 1100 return ret; 1101 1102 mem->mm_node = NULL; 1103 for (i = 0; i < placement->num_placement; ++i) { 1104 const struct ttm_place *place = &placement->placement[i]; 1105 struct ttm_mem_type_manager *man; 1106 1107 ret = ttm_bo_mem_placement(bo, place, mem, ctx); 1108 if (ret == -EBUSY) 1109 continue; 1110 if (ret) 1111 goto error; 1112 1113 type_found = true; 1114 mem->mm_node = NULL; 1115 if (mem->mem_type == TTM_PL_SYSTEM) 1116 return 0; 1117 1118 man = &bdev->man[mem->mem_type]; 1119 ret = (*man->func->get_node)(man, bo, place, mem); 1120 if (unlikely(ret)) 1121 goto error; 1122 1123 if (mem->mm_node) { 1124 ret = ttm_bo_add_move_fence(bo, man, mem); 1125 if (unlikely(ret)) { 1126 (*man->func->put_node)(man, mem); 1127 goto error; 1128 } 1129 return 0; 1130 } 1131 } 1132 1133 for (i = 0; i < placement->num_busy_placement; ++i) { 1134 const struct ttm_place *place = &placement->busy_placement[i]; 1135 1136 ret = ttm_bo_mem_placement(bo, place, mem, ctx); 1137 if (ret == -EBUSY) 1138 continue; 1139 if (ret) 1140 goto error; 1141 1142 type_found = true; 1143 mem->mm_node = NULL; 1144 if (mem->mem_type == TTM_PL_SYSTEM) 1145 return 0; 1146 1147 ret = ttm_bo_mem_force_space(bo, place, mem, ctx); 1148 if (ret == 0 && mem->mm_node) 1149 return 0; 1150 1151 if (ret && ret != -EBUSY) 1152 goto error; 1153 } 1154 1155 ret = -ENOMEM; 1156 if (!type_found) { 1157 pr_err(TTM_PFX "No compatible memory type found\n"); 1158 ret = -EINVAL; 1159 } 1160 1161 error: 1162 if (bo->mem.mem_type == TTM_PL_SYSTEM && !list_empty(&bo->lru)) { 1163 spin_lock(&bo->bdev->glob->lru_lock); 1164 ttm_bo_move_to_lru_tail(bo, NULL); 1165 spin_unlock(&bo->bdev->glob->lru_lock); 1166 } 1167 1168 return ret; 1169 } 1170 EXPORT_SYMBOL(ttm_bo_mem_space); 1171 1172 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo, 1173 struct ttm_placement *placement, 1174 struct ttm_operation_ctx *ctx) 1175 { 1176 int ret = 0; 1177 struct ttm_mem_reg mem; 1178 1179 dma_resv_assert_held(bo->base.resv); 1180 1181 mem.num_pages = bo->num_pages; 1182 mem.size = mem.num_pages << PAGE_SHIFT; 1183 mem.page_alignment = bo->mem.page_alignment; 1184 mem.bus.io_reserved_vm = false; 1185 mem.bus.io_reserved_count = 0; 1186 /* 1187 * Determine where to move the buffer. 1188 */ 1189 ret = ttm_bo_mem_space(bo, placement, &mem, ctx); 1190 if (ret) 1191 goto out_unlock; 1192 ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx); 1193 out_unlock: 1194 if (ret && mem.mm_node) 1195 ttm_bo_mem_put(bo, &mem); 1196 return ret; 1197 } 1198 1199 static bool ttm_bo_places_compat(const struct ttm_place *places, 1200 unsigned num_placement, 1201 struct ttm_mem_reg *mem, 1202 uint32_t *new_flags) 1203 { 1204 unsigned i; 1205 1206 for (i = 0; i < num_placement; i++) { 1207 const struct ttm_place *heap = &places[i]; 1208 1209 if (mem->mm_node && (mem->start < heap->fpfn || 1210 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn))) 1211 continue; 1212 1213 *new_flags = heap->flags; 1214 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) && 1215 (*new_flags & mem->placement & TTM_PL_MASK_MEM) && 1216 (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) || 1217 (mem->placement & TTM_PL_FLAG_CONTIGUOUS))) 1218 return true; 1219 } 1220 return false; 1221 } 1222 1223 bool ttm_bo_mem_compat(struct ttm_placement *placement, 1224 struct ttm_mem_reg *mem, 1225 uint32_t *new_flags) 1226 { 1227 if (ttm_bo_places_compat(placement->placement, placement->num_placement, 1228 mem, new_flags)) 1229 return true; 1230 1231 if ((placement->busy_placement != placement->placement || 1232 placement->num_busy_placement > placement->num_placement) && 1233 ttm_bo_places_compat(placement->busy_placement, 1234 placement->num_busy_placement, 1235 mem, new_flags)) 1236 return true; 1237 1238 return false; 1239 } 1240 EXPORT_SYMBOL(ttm_bo_mem_compat); 1241 1242 int ttm_bo_validate(struct ttm_buffer_object *bo, 1243 struct ttm_placement *placement, 1244 struct ttm_operation_ctx *ctx) 1245 { 1246 int ret; 1247 uint32_t new_flags; 1248 1249 dma_resv_assert_held(bo->base.resv); 1250 /* 1251 * Check whether we need to move buffer. 1252 */ 1253 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) { 1254 ret = ttm_bo_move_buffer(bo, placement, ctx); 1255 if (ret) 1256 return ret; 1257 } else { 1258 /* 1259 * Use the access and other non-mapping-related flag bits from 1260 * the compatible memory placement flags to the active flags 1261 */ 1262 ttm_flag_masked(&bo->mem.placement, new_flags, 1263 ~TTM_PL_MASK_MEMTYPE); 1264 } 1265 /* 1266 * We might need to add a TTM. 1267 */ 1268 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 1269 ret = ttm_tt_create(bo, true); 1270 if (ret) 1271 return ret; 1272 } 1273 return 0; 1274 } 1275 EXPORT_SYMBOL(ttm_bo_validate); 1276 1277 int ttm_bo_init_reserved(struct ttm_bo_device *bdev, 1278 struct ttm_buffer_object *bo, 1279 unsigned long size, 1280 enum ttm_bo_type type, 1281 struct ttm_placement *placement, 1282 uint32_t page_alignment, 1283 struct ttm_operation_ctx *ctx, 1284 size_t acc_size, 1285 struct sg_table *sg, 1286 struct dma_resv *resv, 1287 void (*destroy) (struct ttm_buffer_object *)) 1288 { 1289 int ret = 0; 1290 unsigned long num_pages; 1291 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob; 1292 bool locked; 1293 1294 ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx); 1295 if (ret) { 1296 pr_err("Out of kernel memory\n"); 1297 if (destroy) 1298 (*destroy)(bo); 1299 else 1300 kfree(bo); 1301 return -ENOMEM; 1302 } 1303 1304 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 1305 if (num_pages == 0) { 1306 pr_err("Illegal buffer object size\n"); 1307 if (destroy) 1308 (*destroy)(bo); 1309 else 1310 kfree(bo); 1311 ttm_mem_global_free(mem_glob, acc_size); 1312 return -EINVAL; 1313 } 1314 bo->destroy = destroy ? destroy : ttm_bo_default_destroy; 1315 1316 kref_init(&bo->kref); 1317 kref_init(&bo->list_kref); 1318 atomic_set(&bo->cpu_writers, 0); 1319 INIT_LIST_HEAD(&bo->lru); 1320 INIT_LIST_HEAD(&bo->ddestroy); 1321 INIT_LIST_HEAD(&bo->swap); 1322 INIT_LIST_HEAD(&bo->io_reserve_lru); 1323 mutex_init(&bo->wu_mutex); 1324 bo->bdev = bdev; 1325 bo->type = type; 1326 bo->num_pages = num_pages; 1327 bo->mem.size = num_pages << PAGE_SHIFT; 1328 bo->mem.mem_type = TTM_PL_SYSTEM; 1329 bo->mem.num_pages = bo->num_pages; 1330 bo->mem.mm_node = NULL; 1331 bo->mem.page_alignment = page_alignment; 1332 bo->mem.bus.io_reserved_vm = false; 1333 bo->mem.bus.io_reserved_count = 0; 1334 bo->moving = NULL; 1335 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED); 1336 bo->acc_size = acc_size; 1337 bo->sg = sg; 1338 if (resv) { 1339 bo->base.resv = resv; 1340 dma_resv_assert_held(bo->base.resv); 1341 } else { 1342 bo->base.resv = &bo->base._resv; 1343 } 1344 if (!ttm_bo_uses_embedded_gem_object(bo)) { 1345 /* 1346 * bo.gem is not initialized, so we have to setup the 1347 * struct elements we want use regardless. 1348 */ 1349 dma_resv_init(&bo->base._resv); 1350 drm_vma_node_reset(&bo->base.vma_node); 1351 } 1352 atomic_inc(&bo->bdev->glob->bo_count); 1353 1354 /* 1355 * For ttm_bo_type_device buffers, allocate 1356 * address space from the device. 1357 */ 1358 if (bo->type == ttm_bo_type_device || 1359 bo->type == ttm_bo_type_sg) 1360 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->base.vma_node, 1361 bo->mem.num_pages); 1362 1363 /* passed reservation objects should already be locked, 1364 * since otherwise lockdep will be angered in radeon. 1365 */ 1366 if (!resv) { 1367 locked = dma_resv_trylock(bo->base.resv); 1368 WARN_ON(!locked); 1369 } 1370 1371 if (likely(!ret)) 1372 ret = ttm_bo_validate(bo, placement, ctx); 1373 1374 if (unlikely(ret)) { 1375 if (!resv) 1376 ttm_bo_unreserve(bo); 1377 1378 ttm_bo_put(bo); 1379 return ret; 1380 } 1381 1382 if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 1383 spin_lock(&bdev->glob->lru_lock); 1384 ttm_bo_add_to_lru(bo); 1385 spin_unlock(&bdev->glob->lru_lock); 1386 } 1387 1388 return ret; 1389 } 1390 EXPORT_SYMBOL(ttm_bo_init_reserved); 1391 1392 int ttm_bo_init(struct ttm_bo_device *bdev, 1393 struct ttm_buffer_object *bo, 1394 unsigned long size, 1395 enum ttm_bo_type type, 1396 struct ttm_placement *placement, 1397 uint32_t page_alignment, 1398 bool interruptible, 1399 size_t acc_size, 1400 struct sg_table *sg, 1401 struct dma_resv *resv, 1402 void (*destroy) (struct ttm_buffer_object *)) 1403 { 1404 struct ttm_operation_ctx ctx = { interruptible, false }; 1405 int ret; 1406 1407 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement, 1408 page_alignment, &ctx, acc_size, 1409 sg, resv, destroy); 1410 if (ret) 1411 return ret; 1412 1413 if (!resv) 1414 ttm_bo_unreserve(bo); 1415 1416 return 0; 1417 } 1418 EXPORT_SYMBOL(ttm_bo_init); 1419 1420 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev, 1421 unsigned long bo_size, 1422 unsigned struct_size) 1423 { 1424 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1425 size_t size = 0; 1426 1427 size += ttm_round_pot(struct_size); 1428 size += ttm_round_pot(npages * sizeof(void *)); 1429 size += ttm_round_pot(sizeof(struct ttm_tt)); 1430 return size; 1431 } 1432 EXPORT_SYMBOL(ttm_bo_acc_size); 1433 1434 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev, 1435 unsigned long bo_size, 1436 unsigned struct_size) 1437 { 1438 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1439 size_t size = 0; 1440 1441 size += ttm_round_pot(struct_size); 1442 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t))); 1443 size += ttm_round_pot(sizeof(struct ttm_dma_tt)); 1444 return size; 1445 } 1446 EXPORT_SYMBOL(ttm_bo_dma_acc_size); 1447 1448 int ttm_bo_create(struct ttm_bo_device *bdev, 1449 unsigned long size, 1450 enum ttm_bo_type type, 1451 struct ttm_placement *placement, 1452 uint32_t page_alignment, 1453 bool interruptible, 1454 struct ttm_buffer_object **p_bo) 1455 { 1456 struct ttm_buffer_object *bo; 1457 size_t acc_size; 1458 int ret; 1459 1460 bo = kzalloc(sizeof(*bo), GFP_KERNEL); 1461 if (unlikely(bo == NULL)) 1462 return -ENOMEM; 1463 1464 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object)); 1465 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment, 1466 interruptible, acc_size, 1467 NULL, NULL, NULL); 1468 if (likely(ret == 0)) 1469 *p_bo = bo; 1470 1471 return ret; 1472 } 1473 EXPORT_SYMBOL(ttm_bo_create); 1474 1475 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev, 1476 unsigned mem_type) 1477 { 1478 struct ttm_operation_ctx ctx = { 1479 .interruptible = false, 1480 .no_wait_gpu = false, 1481 .flags = TTM_OPT_FLAG_FORCE_ALLOC 1482 }; 1483 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1484 struct ttm_bo_global *glob = bdev->glob; 1485 struct dma_fence *fence; 1486 int ret; 1487 unsigned i; 1488 1489 /* 1490 * Can't use standard list traversal since we're unlocking. 1491 */ 1492 1493 spin_lock(&glob->lru_lock); 1494 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 1495 while (!list_empty(&man->lru[i])) { 1496 spin_unlock(&glob->lru_lock); 1497 ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx, 1498 NULL); 1499 if (ret) 1500 return ret; 1501 spin_lock(&glob->lru_lock); 1502 } 1503 } 1504 spin_unlock(&glob->lru_lock); 1505 1506 spin_lock(&man->move_lock); 1507 fence = dma_fence_get(man->move); 1508 spin_unlock(&man->move_lock); 1509 1510 if (fence) { 1511 ret = dma_fence_wait(fence, false); 1512 dma_fence_put(fence); 1513 if (ret) 1514 return ret; 1515 } 1516 1517 return 0; 1518 } 1519 1520 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1521 { 1522 struct ttm_mem_type_manager *man; 1523 int ret = -EINVAL; 1524 1525 if (mem_type >= TTM_NUM_MEM_TYPES) { 1526 pr_err("Illegal memory type %d\n", mem_type); 1527 return ret; 1528 } 1529 man = &bdev->man[mem_type]; 1530 1531 if (!man->has_type) { 1532 pr_err("Trying to take down uninitialized memory manager type %u\n", 1533 mem_type); 1534 return ret; 1535 } 1536 1537 man->use_type = false; 1538 man->has_type = false; 1539 1540 ret = 0; 1541 if (mem_type > 0) { 1542 ret = ttm_bo_force_list_clean(bdev, mem_type); 1543 if (ret) { 1544 pr_err("Cleanup eviction failed\n"); 1545 return ret; 1546 } 1547 1548 ret = (*man->func->takedown)(man); 1549 } 1550 1551 dma_fence_put(man->move); 1552 man->move = NULL; 1553 1554 return ret; 1555 } 1556 EXPORT_SYMBOL(ttm_bo_clean_mm); 1557 1558 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1559 { 1560 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1561 1562 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) { 1563 pr_err("Illegal memory manager memory type %u\n", mem_type); 1564 return -EINVAL; 1565 } 1566 1567 if (!man->has_type) { 1568 pr_err("Memory type %u has not been initialized\n", mem_type); 1569 return 0; 1570 } 1571 1572 return ttm_bo_force_list_clean(bdev, mem_type); 1573 } 1574 EXPORT_SYMBOL(ttm_bo_evict_mm); 1575 1576 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type, 1577 unsigned long p_size) 1578 { 1579 int ret; 1580 struct ttm_mem_type_manager *man; 1581 unsigned i; 1582 1583 BUG_ON(type >= TTM_NUM_MEM_TYPES); 1584 man = &bdev->man[type]; 1585 BUG_ON(man->has_type); 1586 man->io_reserve_fastpath = true; 1587 man->use_io_reserve_lru = false; 1588 mutex_init(&man->io_reserve_mutex); 1589 spin_lock_init(&man->move_lock); 1590 INIT_LIST_HEAD(&man->io_reserve_lru); 1591 1592 ret = bdev->driver->init_mem_type(bdev, type, man); 1593 if (ret) 1594 return ret; 1595 man->bdev = bdev; 1596 1597 if (type != TTM_PL_SYSTEM) { 1598 ret = (*man->func->init)(man, p_size); 1599 if (ret) 1600 return ret; 1601 } 1602 man->has_type = true; 1603 man->use_type = true; 1604 man->size = p_size; 1605 1606 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1607 INIT_LIST_HEAD(&man->lru[i]); 1608 man->move = NULL; 1609 1610 return 0; 1611 } 1612 EXPORT_SYMBOL(ttm_bo_init_mm); 1613 1614 static void ttm_bo_global_kobj_release(struct kobject *kobj) 1615 { 1616 struct ttm_bo_global *glob = 1617 container_of(kobj, struct ttm_bo_global, kobj); 1618 1619 __free_page(glob->dummy_read_page); 1620 } 1621 1622 static void ttm_bo_global_release(void) 1623 { 1624 struct ttm_bo_global *glob = &ttm_bo_glob; 1625 1626 mutex_lock(&ttm_global_mutex); 1627 if (--ttm_bo_glob_use_count > 0) 1628 goto out; 1629 1630 kobject_del(&glob->kobj); 1631 kobject_put(&glob->kobj); 1632 ttm_mem_global_release(&ttm_mem_glob); 1633 memset(glob, 0, sizeof(*glob)); 1634 out: 1635 mutex_unlock(&ttm_global_mutex); 1636 } 1637 1638 static int ttm_bo_global_init(void) 1639 { 1640 struct ttm_bo_global *glob = &ttm_bo_glob; 1641 int ret = 0; 1642 unsigned i; 1643 1644 mutex_lock(&ttm_global_mutex); 1645 if (++ttm_bo_glob_use_count > 1) 1646 goto out; 1647 1648 ret = ttm_mem_global_init(&ttm_mem_glob); 1649 if (ret) 1650 goto out; 1651 1652 spin_lock_init(&glob->lru_lock); 1653 glob->mem_glob = &ttm_mem_glob; 1654 glob->mem_glob->bo_glob = glob; 1655 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32); 1656 1657 if (unlikely(glob->dummy_read_page == NULL)) { 1658 ret = -ENOMEM; 1659 goto out; 1660 } 1661 1662 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1663 INIT_LIST_HEAD(&glob->swap_lru[i]); 1664 INIT_LIST_HEAD(&glob->device_list); 1665 atomic_set(&glob->bo_count, 0); 1666 1667 ret = kobject_init_and_add( 1668 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects"); 1669 if (unlikely(ret != 0)) 1670 kobject_put(&glob->kobj); 1671 out: 1672 mutex_unlock(&ttm_global_mutex); 1673 return ret; 1674 } 1675 1676 int ttm_bo_device_release(struct ttm_bo_device *bdev) 1677 { 1678 int ret = 0; 1679 unsigned i = TTM_NUM_MEM_TYPES; 1680 struct ttm_mem_type_manager *man; 1681 struct ttm_bo_global *glob = bdev->glob; 1682 1683 while (i--) { 1684 man = &bdev->man[i]; 1685 if (man->has_type) { 1686 man->use_type = false; 1687 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { 1688 ret = -EBUSY; 1689 pr_err("DRM memory manager type %d is not clean\n", 1690 i); 1691 } 1692 man->has_type = false; 1693 } 1694 } 1695 1696 mutex_lock(&ttm_global_mutex); 1697 list_del(&bdev->device_list); 1698 mutex_unlock(&ttm_global_mutex); 1699 1700 cancel_delayed_work_sync(&bdev->wq); 1701 1702 if (ttm_bo_delayed_delete(bdev, true)) 1703 pr_debug("Delayed destroy list was clean\n"); 1704 1705 spin_lock(&glob->lru_lock); 1706 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1707 if (list_empty(&bdev->man[0].lru[0])) 1708 pr_debug("Swap list %d was clean\n", i); 1709 spin_unlock(&glob->lru_lock); 1710 1711 drm_vma_offset_manager_destroy(&bdev->vma_manager); 1712 1713 if (!ret) 1714 ttm_bo_global_release(); 1715 1716 return ret; 1717 } 1718 EXPORT_SYMBOL(ttm_bo_device_release); 1719 1720 int ttm_bo_device_init(struct ttm_bo_device *bdev, 1721 struct ttm_bo_driver *driver, 1722 struct address_space *mapping, 1723 bool need_dma32) 1724 { 1725 struct ttm_bo_global *glob = &ttm_bo_glob; 1726 int ret; 1727 1728 ret = ttm_bo_global_init(); 1729 if (ret) 1730 return ret; 1731 1732 bdev->driver = driver; 1733 1734 memset(bdev->man, 0, sizeof(bdev->man)); 1735 1736 /* 1737 * Initialize the system memory buffer type. 1738 * Other types need to be driver / IOCTL initialized. 1739 */ 1740 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); 1741 if (unlikely(ret != 0)) 1742 goto out_no_sys; 1743 1744 drm_vma_offset_manager_init(&bdev->vma_manager, 1745 DRM_FILE_PAGE_OFFSET_START, 1746 DRM_FILE_PAGE_OFFSET_SIZE); 1747 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue); 1748 INIT_LIST_HEAD(&bdev->ddestroy); 1749 bdev->dev_mapping = mapping; 1750 bdev->glob = glob; 1751 bdev->need_dma32 = need_dma32; 1752 mutex_lock(&ttm_global_mutex); 1753 list_add_tail(&bdev->device_list, &glob->device_list); 1754 mutex_unlock(&ttm_global_mutex); 1755 1756 return 0; 1757 out_no_sys: 1758 ttm_bo_global_release(); 1759 return ret; 1760 } 1761 EXPORT_SYMBOL(ttm_bo_device_init); 1762 1763 /* 1764 * buffer object vm functions. 1765 */ 1766 1767 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 1768 { 1769 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 1770 1771 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 1772 if (mem->mem_type == TTM_PL_SYSTEM) 1773 return false; 1774 1775 if (man->flags & TTM_MEMTYPE_FLAG_CMA) 1776 return false; 1777 1778 if (mem->placement & TTM_PL_FLAG_CACHED) 1779 return false; 1780 } 1781 return true; 1782 } 1783 1784 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) 1785 { 1786 struct ttm_bo_device *bdev = bo->bdev; 1787 1788 drm_vma_node_unmap(&bo->base.vma_node, bdev->dev_mapping); 1789 ttm_mem_io_free_vm(bo); 1790 } 1791 1792 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) 1793 { 1794 struct ttm_bo_device *bdev = bo->bdev; 1795 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 1796 1797 ttm_mem_io_lock(man, false); 1798 ttm_bo_unmap_virtual_locked(bo); 1799 ttm_mem_io_unlock(man); 1800 } 1801 1802 1803 EXPORT_SYMBOL(ttm_bo_unmap_virtual); 1804 1805 int ttm_bo_wait(struct ttm_buffer_object *bo, 1806 bool interruptible, bool no_wait) 1807 { 1808 long timeout = 15 * HZ; 1809 1810 if (no_wait) { 1811 if (dma_resv_test_signaled_rcu(bo->base.resv, true)) 1812 return 0; 1813 else 1814 return -EBUSY; 1815 } 1816 1817 timeout = dma_resv_wait_timeout_rcu(bo->base.resv, true, 1818 interruptible, timeout); 1819 if (timeout < 0) 1820 return timeout; 1821 1822 if (timeout == 0) 1823 return -EBUSY; 1824 1825 dma_resv_add_excl_fence(bo->base.resv, NULL); 1826 return 0; 1827 } 1828 EXPORT_SYMBOL(ttm_bo_wait); 1829 1830 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) 1831 { 1832 int ret = 0; 1833 1834 /* 1835 * Using ttm_bo_reserve makes sure the lru lists are updated. 1836 */ 1837 1838 ret = ttm_bo_reserve(bo, true, no_wait, NULL); 1839 if (unlikely(ret != 0)) 1840 return ret; 1841 ret = ttm_bo_wait(bo, true, no_wait); 1842 if (likely(ret == 0)) 1843 atomic_inc(&bo->cpu_writers); 1844 ttm_bo_unreserve(bo); 1845 return ret; 1846 } 1847 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab); 1848 1849 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) 1850 { 1851 atomic_dec(&bo->cpu_writers); 1852 } 1853 EXPORT_SYMBOL(ttm_bo_synccpu_write_release); 1854 1855 /** 1856 * A buffer object shrink method that tries to swap out the first 1857 * buffer object on the bo_global::swap_lru list. 1858 */ 1859 int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx) 1860 { 1861 struct ttm_buffer_object *bo; 1862 int ret = -EBUSY; 1863 bool locked; 1864 unsigned i; 1865 1866 spin_lock(&glob->lru_lock); 1867 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 1868 list_for_each_entry(bo, &glob->swap_lru[i], swap) { 1869 if (ttm_bo_evict_swapout_allowable(bo, ctx, &locked, 1870 NULL)) { 1871 ret = 0; 1872 break; 1873 } 1874 } 1875 if (!ret) 1876 break; 1877 } 1878 1879 if (ret) { 1880 spin_unlock(&glob->lru_lock); 1881 return ret; 1882 } 1883 1884 kref_get(&bo->list_kref); 1885 1886 if (!list_empty(&bo->ddestroy)) { 1887 ret = ttm_bo_cleanup_refs(bo, false, false, locked); 1888 kref_put(&bo->list_kref, ttm_bo_release_list); 1889 return ret; 1890 } 1891 1892 ttm_bo_del_from_lru(bo); 1893 spin_unlock(&glob->lru_lock); 1894 1895 /** 1896 * Move to system cached 1897 */ 1898 1899 if (bo->mem.mem_type != TTM_PL_SYSTEM || 1900 bo->ttm->caching_state != tt_cached) { 1901 struct ttm_operation_ctx ctx = { false, false }; 1902 struct ttm_mem_reg evict_mem; 1903 1904 evict_mem = bo->mem; 1905 evict_mem.mm_node = NULL; 1906 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; 1907 evict_mem.mem_type = TTM_PL_SYSTEM; 1908 1909 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx); 1910 if (unlikely(ret != 0)) 1911 goto out; 1912 } 1913 1914 /** 1915 * Make sure BO is idle. 1916 */ 1917 1918 ret = ttm_bo_wait(bo, false, false); 1919 if (unlikely(ret != 0)) 1920 goto out; 1921 1922 ttm_bo_unmap_virtual(bo); 1923 1924 /** 1925 * Swap out. Buffer will be swapped in again as soon as 1926 * anyone tries to access a ttm page. 1927 */ 1928 1929 if (bo->bdev->driver->swap_notify) 1930 bo->bdev->driver->swap_notify(bo); 1931 1932 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); 1933 out: 1934 1935 /** 1936 * 1937 * Unreserve without putting on LRU to avoid swapping out an 1938 * already swapped buffer. 1939 */ 1940 if (locked) 1941 dma_resv_unlock(bo->base.resv); 1942 kref_put(&bo->list_kref, ttm_bo_release_list); 1943 return ret; 1944 } 1945 EXPORT_SYMBOL(ttm_bo_swapout); 1946 1947 void ttm_bo_swapout_all(struct ttm_bo_device *bdev) 1948 { 1949 struct ttm_operation_ctx ctx = { 1950 .interruptible = false, 1951 .no_wait_gpu = false 1952 }; 1953 1954 while (ttm_bo_swapout(bdev->glob, &ctx) == 0) 1955 ; 1956 } 1957 EXPORT_SYMBOL(ttm_bo_swapout_all); 1958 1959 /** 1960 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become 1961 * unreserved 1962 * 1963 * @bo: Pointer to buffer 1964 */ 1965 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo) 1966 { 1967 int ret; 1968 1969 /* 1970 * In the absense of a wait_unlocked API, 1971 * Use the bo::wu_mutex to avoid triggering livelocks due to 1972 * concurrent use of this function. Note that this use of 1973 * bo::wu_mutex can go away if we change locking order to 1974 * mmap_sem -> bo::reserve. 1975 */ 1976 ret = mutex_lock_interruptible(&bo->wu_mutex); 1977 if (unlikely(ret != 0)) 1978 return -ERESTARTSYS; 1979 if (!dma_resv_is_locked(bo->base.resv)) 1980 goto out_unlock; 1981 ret = dma_resv_lock_interruptible(bo->base.resv, NULL); 1982 if (ret == -EINTR) 1983 ret = -ERESTARTSYS; 1984 if (unlikely(ret != 0)) 1985 goto out_unlock; 1986 dma_resv_unlock(bo->base.resv); 1987 1988 out_unlock: 1989 mutex_unlock(&bo->wu_mutex); 1990 return ret; 1991 } 1992