1 /************************************************************************** 2 * 3 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA 4 * All Rights Reserved. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the 8 * "Software"), to deal in the Software without restriction, including 9 * without limitation the rights to use, copy, modify, merge, publish, 10 * distribute, sub license, and/or sell copies of the Software, and to 11 * permit persons to whom the Software is furnished to do so, subject to 12 * the following conditions: 13 * 14 * The above copyright notice and this permission notice (including the 15 * next paragraph) shall be included in all copies or substantial portions 16 * of the Software. 17 * 18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 20 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 21 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 22 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 23 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 24 * USE OR OTHER DEALINGS IN THE SOFTWARE. 25 * 26 **************************************************************************/ 27 /* 28 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com> 29 */ 30 31 #define pr_fmt(fmt) "[TTM] " fmt 32 33 #include <drm/ttm/ttm_module.h> 34 #include <drm/ttm/ttm_bo_driver.h> 35 #include <drm/ttm/ttm_placement.h> 36 #include <linux/jiffies.h> 37 #include <linux/slab.h> 38 #include <linux/sched.h> 39 #include <linux/mm.h> 40 #include <linux/file.h> 41 #include <linux/module.h> 42 #include <linux/atomic.h> 43 #include <linux/reservation.h> 44 45 static void ttm_bo_global_kobj_release(struct kobject *kobj); 46 47 static struct attribute ttm_bo_count = { 48 .name = "bo_count", 49 .mode = S_IRUGO 50 }; 51 52 static inline int ttm_mem_type_from_place(const struct ttm_place *place, 53 uint32_t *mem_type) 54 { 55 int pos; 56 57 pos = ffs(place->flags & TTM_PL_MASK_MEM); 58 if (unlikely(!pos)) 59 return -EINVAL; 60 61 *mem_type = pos - 1; 62 return 0; 63 } 64 65 static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type) 66 { 67 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 68 struct drm_printer p = drm_debug_printer(TTM_PFX); 69 70 pr_err(" has_type: %d\n", man->has_type); 71 pr_err(" use_type: %d\n", man->use_type); 72 pr_err(" flags: 0x%08X\n", man->flags); 73 pr_err(" gpu_offset: 0x%08llX\n", man->gpu_offset); 74 pr_err(" size: %llu\n", man->size); 75 pr_err(" available_caching: 0x%08X\n", man->available_caching); 76 pr_err(" default_caching: 0x%08X\n", man->default_caching); 77 if (mem_type != TTM_PL_SYSTEM) 78 (*man->func->debug)(man, &p); 79 } 80 81 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo, 82 struct ttm_placement *placement) 83 { 84 int i, ret, mem_type; 85 86 pr_err("No space for %p (%lu pages, %luK, %luM)\n", 87 bo, bo->mem.num_pages, bo->mem.size >> 10, 88 bo->mem.size >> 20); 89 for (i = 0; i < placement->num_placement; i++) { 90 ret = ttm_mem_type_from_place(&placement->placement[i], 91 &mem_type); 92 if (ret) 93 return; 94 pr_err(" placement[%d]=0x%08X (%d)\n", 95 i, placement->placement[i].flags, mem_type); 96 ttm_mem_type_debug(bo->bdev, mem_type); 97 } 98 } 99 100 static ssize_t ttm_bo_global_show(struct kobject *kobj, 101 struct attribute *attr, 102 char *buffer) 103 { 104 struct ttm_bo_global *glob = 105 container_of(kobj, struct ttm_bo_global, kobj); 106 107 return snprintf(buffer, PAGE_SIZE, "%d\n", 108 atomic_read(&glob->bo_count)); 109 } 110 111 static struct attribute *ttm_bo_global_attrs[] = { 112 &ttm_bo_count, 113 NULL 114 }; 115 116 static const struct sysfs_ops ttm_bo_global_ops = { 117 .show = &ttm_bo_global_show 118 }; 119 120 static struct kobj_type ttm_bo_glob_kobj_type = { 121 .release = &ttm_bo_global_kobj_release, 122 .sysfs_ops = &ttm_bo_global_ops, 123 .default_attrs = ttm_bo_global_attrs 124 }; 125 126 127 static inline uint32_t ttm_bo_type_flags(unsigned type) 128 { 129 return 1 << (type); 130 } 131 132 static void ttm_bo_release_list(struct kref *list_kref) 133 { 134 struct ttm_buffer_object *bo = 135 container_of(list_kref, struct ttm_buffer_object, list_kref); 136 struct ttm_bo_device *bdev = bo->bdev; 137 size_t acc_size = bo->acc_size; 138 139 BUG_ON(kref_read(&bo->list_kref)); 140 BUG_ON(kref_read(&bo->kref)); 141 BUG_ON(atomic_read(&bo->cpu_writers)); 142 BUG_ON(bo->mem.mm_node != NULL); 143 BUG_ON(!list_empty(&bo->lru)); 144 BUG_ON(!list_empty(&bo->ddestroy)); 145 ttm_tt_destroy(bo->ttm); 146 atomic_dec(&bo->glob->bo_count); 147 dma_fence_put(bo->moving); 148 reservation_object_fini(&bo->ttm_resv); 149 mutex_destroy(&bo->wu_mutex); 150 if (bo->destroy) 151 bo->destroy(bo); 152 else { 153 kfree(bo); 154 } 155 ttm_mem_global_free(bdev->glob->mem_glob, acc_size); 156 } 157 158 void ttm_bo_add_to_lru(struct ttm_buffer_object *bo) 159 { 160 struct ttm_bo_device *bdev = bo->bdev; 161 struct ttm_mem_type_manager *man; 162 163 reservation_object_assert_held(bo->resv); 164 165 if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 166 167 BUG_ON(!list_empty(&bo->lru)); 168 169 man = &bdev->man[bo->mem.mem_type]; 170 list_add_tail(&bo->lru, &man->lru[bo->priority]); 171 kref_get(&bo->list_kref); 172 173 if (bo->ttm && !(bo->ttm->page_flags & 174 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED))) { 175 list_add_tail(&bo->swap, 176 &bo->glob->swap_lru[bo->priority]); 177 kref_get(&bo->list_kref); 178 } 179 } 180 } 181 EXPORT_SYMBOL(ttm_bo_add_to_lru); 182 183 static void ttm_bo_ref_bug(struct kref *list_kref) 184 { 185 BUG(); 186 } 187 188 void ttm_bo_del_from_lru(struct ttm_buffer_object *bo) 189 { 190 if (!list_empty(&bo->swap)) { 191 list_del_init(&bo->swap); 192 kref_put(&bo->list_kref, ttm_bo_ref_bug); 193 } 194 if (!list_empty(&bo->lru)) { 195 list_del_init(&bo->lru); 196 kref_put(&bo->list_kref, ttm_bo_ref_bug); 197 } 198 199 /* 200 * TODO: Add a driver hook to delete from 201 * driver-specific LRU's here. 202 */ 203 } 204 205 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo) 206 { 207 spin_lock(&bo->glob->lru_lock); 208 ttm_bo_del_from_lru(bo); 209 spin_unlock(&bo->glob->lru_lock); 210 } 211 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru); 212 213 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo) 214 { 215 reservation_object_assert_held(bo->resv); 216 217 ttm_bo_del_from_lru(bo); 218 ttm_bo_add_to_lru(bo); 219 } 220 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail); 221 222 /* 223 * Call bo->mutex locked. 224 */ 225 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc) 226 { 227 struct ttm_bo_device *bdev = bo->bdev; 228 struct ttm_bo_global *glob = bo->glob; 229 int ret = 0; 230 uint32_t page_flags = 0; 231 232 reservation_object_assert_held(bo->resv); 233 bo->ttm = NULL; 234 235 if (bdev->need_dma32) 236 page_flags |= TTM_PAGE_FLAG_DMA32; 237 238 switch (bo->type) { 239 case ttm_bo_type_device: 240 if (zero_alloc) 241 page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC; 242 case ttm_bo_type_kernel: 243 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, 244 page_flags, glob->dummy_read_page); 245 if (unlikely(bo->ttm == NULL)) 246 ret = -ENOMEM; 247 break; 248 case ttm_bo_type_sg: 249 bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT, 250 page_flags | TTM_PAGE_FLAG_SG, 251 glob->dummy_read_page); 252 if (unlikely(bo->ttm == NULL)) { 253 ret = -ENOMEM; 254 break; 255 } 256 bo->ttm->sg = bo->sg; 257 break; 258 default: 259 pr_err("Illegal buffer object type\n"); 260 ret = -EINVAL; 261 break; 262 } 263 264 return ret; 265 } 266 267 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo, 268 struct ttm_mem_reg *mem, bool evict, 269 struct ttm_operation_ctx *ctx) 270 { 271 struct ttm_bo_device *bdev = bo->bdev; 272 bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem); 273 bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem); 274 struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type]; 275 struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type]; 276 int ret = 0; 277 278 if (old_is_pci || new_is_pci || 279 ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) { 280 ret = ttm_mem_io_lock(old_man, true); 281 if (unlikely(ret != 0)) 282 goto out_err; 283 ttm_bo_unmap_virtual_locked(bo); 284 ttm_mem_io_unlock(old_man); 285 } 286 287 /* 288 * Create and bind a ttm if required. 289 */ 290 291 if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 292 if (bo->ttm == NULL) { 293 bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED); 294 ret = ttm_bo_add_ttm(bo, zero); 295 if (ret) 296 goto out_err; 297 } 298 299 ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement); 300 if (ret) 301 goto out_err; 302 303 if (mem->mem_type != TTM_PL_SYSTEM) { 304 ret = ttm_tt_bind(bo->ttm, mem, ctx); 305 if (ret) 306 goto out_err; 307 } 308 309 if (bo->mem.mem_type == TTM_PL_SYSTEM) { 310 if (bdev->driver->move_notify) 311 bdev->driver->move_notify(bo, evict, mem); 312 bo->mem = *mem; 313 mem->mm_node = NULL; 314 goto moved; 315 } 316 } 317 318 if (bdev->driver->move_notify) 319 bdev->driver->move_notify(bo, evict, mem); 320 321 if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) && 322 !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) 323 ret = ttm_bo_move_ttm(bo, ctx, mem); 324 else if (bdev->driver->move) 325 ret = bdev->driver->move(bo, evict, ctx, mem); 326 else 327 ret = ttm_bo_move_memcpy(bo, ctx, mem); 328 329 if (ret) { 330 if (bdev->driver->move_notify) { 331 struct ttm_mem_reg tmp_mem = *mem; 332 *mem = bo->mem; 333 bo->mem = tmp_mem; 334 bdev->driver->move_notify(bo, false, mem); 335 bo->mem = *mem; 336 *mem = tmp_mem; 337 } 338 339 goto out_err; 340 } 341 342 moved: 343 if (bo->evicted) { 344 if (bdev->driver->invalidate_caches) { 345 ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement); 346 if (ret) 347 pr_err("Can not flush read caches\n"); 348 } 349 bo->evicted = false; 350 } 351 352 if (bo->mem.mm_node) 353 bo->offset = (bo->mem.start << PAGE_SHIFT) + 354 bdev->man[bo->mem.mem_type].gpu_offset; 355 else 356 bo->offset = 0; 357 358 ctx->bytes_moved += bo->num_pages << PAGE_SHIFT; 359 return 0; 360 361 out_err: 362 new_man = &bdev->man[bo->mem.mem_type]; 363 if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) { 364 ttm_tt_destroy(bo->ttm); 365 bo->ttm = NULL; 366 } 367 368 return ret; 369 } 370 371 /** 372 * Call bo::reserved. 373 * Will release GPU memory type usage on destruction. 374 * This is the place to put in driver specific hooks to release 375 * driver private resources. 376 * Will release the bo::reserved lock. 377 */ 378 379 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo) 380 { 381 if (bo->bdev->driver->move_notify) 382 bo->bdev->driver->move_notify(bo, false, NULL); 383 384 ttm_tt_destroy(bo->ttm); 385 bo->ttm = NULL; 386 ttm_bo_mem_put(bo, &bo->mem); 387 } 388 389 static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo) 390 { 391 int r; 392 393 if (bo->resv == &bo->ttm_resv) 394 return 0; 395 396 BUG_ON(!reservation_object_trylock(&bo->ttm_resv)); 397 398 r = reservation_object_copy_fences(&bo->ttm_resv, bo->resv); 399 if (r) 400 reservation_object_unlock(&bo->ttm_resv); 401 402 return r; 403 } 404 405 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo) 406 { 407 struct reservation_object_list *fobj; 408 struct dma_fence *fence; 409 int i; 410 411 fobj = reservation_object_get_list(&bo->ttm_resv); 412 fence = reservation_object_get_excl(&bo->ttm_resv); 413 if (fence && !fence->ops->signaled) 414 dma_fence_enable_sw_signaling(fence); 415 416 for (i = 0; fobj && i < fobj->shared_count; ++i) { 417 fence = rcu_dereference_protected(fobj->shared[i], 418 reservation_object_held(bo->resv)); 419 420 if (!fence->ops->signaled) 421 dma_fence_enable_sw_signaling(fence); 422 } 423 } 424 425 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo) 426 { 427 struct ttm_bo_device *bdev = bo->bdev; 428 struct ttm_bo_global *glob = bo->glob; 429 int ret; 430 431 ret = ttm_bo_individualize_resv(bo); 432 if (ret) { 433 /* Last resort, if we fail to allocate memory for the 434 * fences block for the BO to become idle 435 */ 436 reservation_object_wait_timeout_rcu(bo->resv, true, false, 437 30 * HZ); 438 spin_lock(&glob->lru_lock); 439 goto error; 440 } 441 442 spin_lock(&glob->lru_lock); 443 ret = reservation_object_trylock(bo->resv) ? 0 : -EBUSY; 444 if (!ret) { 445 if (reservation_object_test_signaled_rcu(&bo->ttm_resv, true)) { 446 ttm_bo_del_from_lru(bo); 447 spin_unlock(&glob->lru_lock); 448 if (bo->resv != &bo->ttm_resv) 449 reservation_object_unlock(&bo->ttm_resv); 450 451 ttm_bo_cleanup_memtype_use(bo); 452 reservation_object_unlock(bo->resv); 453 return; 454 } 455 456 ttm_bo_flush_all_fences(bo); 457 458 /* 459 * Make NO_EVICT bos immediately available to 460 * shrinkers, now that they are queued for 461 * destruction. 462 */ 463 if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) { 464 bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT; 465 ttm_bo_add_to_lru(bo); 466 } 467 468 reservation_object_unlock(bo->resv); 469 } 470 if (bo->resv != &bo->ttm_resv) 471 reservation_object_unlock(&bo->ttm_resv); 472 473 error: 474 kref_get(&bo->list_kref); 475 list_add_tail(&bo->ddestroy, &bdev->ddestroy); 476 spin_unlock(&glob->lru_lock); 477 478 schedule_delayed_work(&bdev->wq, 479 ((HZ / 100) < 1) ? 1 : HZ / 100); 480 } 481 482 /** 483 * function ttm_bo_cleanup_refs 484 * If bo idle, remove from delayed- and lru lists, and unref. 485 * If not idle, do nothing. 486 * 487 * Must be called with lru_lock and reservation held, this function 488 * will drop the lru lock and optionally the reservation lock before returning. 489 * 490 * @interruptible Any sleeps should occur interruptibly. 491 * @no_wait_gpu Never wait for gpu. Return -EBUSY instead. 492 * @unlock_resv Unlock the reservation lock as well. 493 */ 494 495 static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo, 496 bool interruptible, bool no_wait_gpu, 497 bool unlock_resv) 498 { 499 struct ttm_bo_global *glob = bo->glob; 500 struct reservation_object *resv; 501 int ret; 502 503 if (unlikely(list_empty(&bo->ddestroy))) 504 resv = bo->resv; 505 else 506 resv = &bo->ttm_resv; 507 508 if (reservation_object_test_signaled_rcu(resv, true)) 509 ret = 0; 510 else 511 ret = -EBUSY; 512 513 if (ret && !no_wait_gpu) { 514 long lret; 515 516 if (unlock_resv) 517 reservation_object_unlock(bo->resv); 518 spin_unlock(&glob->lru_lock); 519 520 lret = reservation_object_wait_timeout_rcu(resv, true, 521 interruptible, 522 30 * HZ); 523 524 if (lret < 0) 525 return lret; 526 else if (lret == 0) 527 return -EBUSY; 528 529 spin_lock(&glob->lru_lock); 530 if (unlock_resv && !reservation_object_trylock(bo->resv)) { 531 /* 532 * We raced, and lost, someone else holds the reservation now, 533 * and is probably busy in ttm_bo_cleanup_memtype_use. 534 * 535 * Even if it's not the case, because we finished waiting any 536 * delayed destruction would succeed, so just return success 537 * here. 538 */ 539 spin_unlock(&glob->lru_lock); 540 return 0; 541 } 542 ret = 0; 543 } 544 545 if (ret || unlikely(list_empty(&bo->ddestroy))) { 546 if (unlock_resv) 547 reservation_object_unlock(bo->resv); 548 spin_unlock(&glob->lru_lock); 549 return ret; 550 } 551 552 ttm_bo_del_from_lru(bo); 553 list_del_init(&bo->ddestroy); 554 kref_put(&bo->list_kref, ttm_bo_ref_bug); 555 556 spin_unlock(&glob->lru_lock); 557 ttm_bo_cleanup_memtype_use(bo); 558 559 if (unlock_resv) 560 reservation_object_unlock(bo->resv); 561 562 return 0; 563 } 564 565 /** 566 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all 567 * encountered buffers. 568 */ 569 static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all) 570 { 571 struct ttm_bo_global *glob = bdev->glob; 572 struct list_head removed; 573 bool empty; 574 575 INIT_LIST_HEAD(&removed); 576 577 spin_lock(&glob->lru_lock); 578 while (!list_empty(&bdev->ddestroy)) { 579 struct ttm_buffer_object *bo; 580 581 bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object, 582 ddestroy); 583 kref_get(&bo->list_kref); 584 list_move_tail(&bo->ddestroy, &removed); 585 586 if (remove_all || bo->resv != &bo->ttm_resv) { 587 spin_unlock(&glob->lru_lock); 588 reservation_object_lock(bo->resv, NULL); 589 590 spin_lock(&glob->lru_lock); 591 ttm_bo_cleanup_refs(bo, false, !remove_all, true); 592 593 } else if (reservation_object_trylock(bo->resv)) { 594 ttm_bo_cleanup_refs(bo, false, !remove_all, true); 595 } else { 596 spin_unlock(&glob->lru_lock); 597 } 598 599 kref_put(&bo->list_kref, ttm_bo_release_list); 600 spin_lock(&glob->lru_lock); 601 } 602 list_splice_tail(&removed, &bdev->ddestroy); 603 empty = list_empty(&bdev->ddestroy); 604 spin_unlock(&glob->lru_lock); 605 606 return empty; 607 } 608 609 static void ttm_bo_delayed_workqueue(struct work_struct *work) 610 { 611 struct ttm_bo_device *bdev = 612 container_of(work, struct ttm_bo_device, wq.work); 613 614 if (!ttm_bo_delayed_delete(bdev, false)) { 615 schedule_delayed_work(&bdev->wq, 616 ((HZ / 100) < 1) ? 1 : HZ / 100); 617 } 618 } 619 620 static void ttm_bo_release(struct kref *kref) 621 { 622 struct ttm_buffer_object *bo = 623 container_of(kref, struct ttm_buffer_object, kref); 624 struct ttm_bo_device *bdev = bo->bdev; 625 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 626 627 drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node); 628 ttm_mem_io_lock(man, false); 629 ttm_mem_io_free_vm(bo); 630 ttm_mem_io_unlock(man); 631 ttm_bo_cleanup_refs_or_queue(bo); 632 kref_put(&bo->list_kref, ttm_bo_release_list); 633 } 634 635 void ttm_bo_unref(struct ttm_buffer_object **p_bo) 636 { 637 struct ttm_buffer_object *bo = *p_bo; 638 639 *p_bo = NULL; 640 kref_put(&bo->kref, ttm_bo_release); 641 } 642 EXPORT_SYMBOL(ttm_bo_unref); 643 644 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev) 645 { 646 return cancel_delayed_work_sync(&bdev->wq); 647 } 648 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue); 649 650 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched) 651 { 652 if (resched) 653 schedule_delayed_work(&bdev->wq, 654 ((HZ / 100) < 1) ? 1 : HZ / 100); 655 } 656 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue); 657 658 static int ttm_bo_evict(struct ttm_buffer_object *bo, 659 struct ttm_operation_ctx *ctx) 660 { 661 struct ttm_bo_device *bdev = bo->bdev; 662 struct ttm_mem_reg evict_mem; 663 struct ttm_placement placement; 664 int ret = 0; 665 666 reservation_object_assert_held(bo->resv); 667 668 evict_mem = bo->mem; 669 evict_mem.mm_node = NULL; 670 evict_mem.bus.io_reserved_vm = false; 671 evict_mem.bus.io_reserved_count = 0; 672 673 placement.num_placement = 0; 674 placement.num_busy_placement = 0; 675 bdev->driver->evict_flags(bo, &placement); 676 ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx); 677 if (ret) { 678 if (ret != -ERESTARTSYS) { 679 pr_err("Failed to find memory space for buffer 0x%p eviction\n", 680 bo); 681 ttm_bo_mem_space_debug(bo, &placement); 682 } 683 goto out; 684 } 685 686 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx); 687 if (unlikely(ret)) { 688 if (ret != -ERESTARTSYS) 689 pr_err("Buffer eviction failed\n"); 690 ttm_bo_mem_put(bo, &evict_mem); 691 goto out; 692 } 693 bo->evicted = true; 694 out: 695 return ret; 696 } 697 698 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo, 699 const struct ttm_place *place) 700 { 701 /* Don't evict this BO if it's outside of the 702 * requested placement range 703 */ 704 if (place->fpfn >= (bo->mem.start + bo->mem.size) || 705 (place->lpfn && place->lpfn <= bo->mem.start)) 706 return false; 707 708 return true; 709 } 710 EXPORT_SYMBOL(ttm_bo_eviction_valuable); 711 712 /** 713 * Check the target bo is allowable to be evicted or swapout, including cases: 714 * 715 * a. if share same reservation object with ctx->resv, have assumption 716 * reservation objects should already be locked, so not lock again and 717 * return true directly when either the opreation allow_reserved_eviction 718 * or the target bo already is in delayed free list; 719 * 720 * b. Otherwise, trylock it. 721 */ 722 static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo, 723 struct ttm_operation_ctx *ctx, bool *locked) 724 { 725 bool ret = false; 726 727 *locked = false; 728 if (bo->resv == ctx->resv) { 729 reservation_object_assert_held(bo->resv); 730 if (ctx->allow_reserved_eviction || !list_empty(&bo->ddestroy)) 731 ret = true; 732 } else { 733 *locked = reservation_object_trylock(bo->resv); 734 ret = *locked; 735 } 736 737 return ret; 738 } 739 740 static int ttm_mem_evict_first(struct ttm_bo_device *bdev, 741 uint32_t mem_type, 742 const struct ttm_place *place, 743 struct ttm_operation_ctx *ctx) 744 { 745 struct ttm_bo_global *glob = bdev->glob; 746 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 747 struct ttm_buffer_object *bo = NULL; 748 bool locked = false; 749 unsigned i; 750 int ret; 751 752 spin_lock(&glob->lru_lock); 753 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 754 list_for_each_entry(bo, &man->lru[i], lru) { 755 if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked)) 756 continue; 757 758 if (place && !bdev->driver->eviction_valuable(bo, 759 place)) { 760 if (locked) 761 reservation_object_unlock(bo->resv); 762 continue; 763 } 764 break; 765 } 766 767 /* If the inner loop terminated early, we have our candidate */ 768 if (&bo->lru != &man->lru[i]) 769 break; 770 771 bo = NULL; 772 } 773 774 if (!bo) { 775 spin_unlock(&glob->lru_lock); 776 return -EBUSY; 777 } 778 779 kref_get(&bo->list_kref); 780 781 if (!list_empty(&bo->ddestroy)) { 782 ret = ttm_bo_cleanup_refs(bo, ctx->interruptible, 783 ctx->no_wait_gpu, locked); 784 kref_put(&bo->list_kref, ttm_bo_release_list); 785 return ret; 786 } 787 788 ttm_bo_del_from_lru(bo); 789 spin_unlock(&glob->lru_lock); 790 791 ret = ttm_bo_evict(bo, ctx); 792 if (locked) { 793 ttm_bo_unreserve(bo); 794 } else { 795 spin_lock(&glob->lru_lock); 796 ttm_bo_add_to_lru(bo); 797 spin_unlock(&glob->lru_lock); 798 } 799 800 kref_put(&bo->list_kref, ttm_bo_release_list); 801 return ret; 802 } 803 804 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem) 805 { 806 struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type]; 807 808 if (mem->mm_node) 809 (*man->func->put_node)(man, mem); 810 } 811 EXPORT_SYMBOL(ttm_bo_mem_put); 812 813 /** 814 * Add the last move fence to the BO and reserve a new shared slot. 815 */ 816 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo, 817 struct ttm_mem_type_manager *man, 818 struct ttm_mem_reg *mem) 819 { 820 struct dma_fence *fence; 821 int ret; 822 823 spin_lock(&man->move_lock); 824 fence = dma_fence_get(man->move); 825 spin_unlock(&man->move_lock); 826 827 if (fence) { 828 reservation_object_add_shared_fence(bo->resv, fence); 829 830 ret = reservation_object_reserve_shared(bo->resv); 831 if (unlikely(ret)) 832 return ret; 833 834 dma_fence_put(bo->moving); 835 bo->moving = fence; 836 } 837 838 return 0; 839 } 840 841 /** 842 * Repeatedly evict memory from the LRU for @mem_type until we create enough 843 * space, or we've evicted everything and there isn't enough space. 844 */ 845 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo, 846 uint32_t mem_type, 847 const struct ttm_place *place, 848 struct ttm_mem_reg *mem, 849 struct ttm_operation_ctx *ctx) 850 { 851 struct ttm_bo_device *bdev = bo->bdev; 852 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 853 int ret; 854 855 do { 856 ret = (*man->func->get_node)(man, bo, place, mem); 857 if (unlikely(ret != 0)) 858 return ret; 859 if (mem->mm_node) 860 break; 861 ret = ttm_mem_evict_first(bdev, mem_type, place, ctx); 862 if (unlikely(ret != 0)) 863 return ret; 864 } while (1); 865 mem->mem_type = mem_type; 866 return ttm_bo_add_move_fence(bo, man, mem); 867 } 868 869 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man, 870 uint32_t cur_placement, 871 uint32_t proposed_placement) 872 { 873 uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING; 874 uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING; 875 876 /** 877 * Keep current caching if possible. 878 */ 879 880 if ((cur_placement & caching) != 0) 881 result |= (cur_placement & caching); 882 else if ((man->default_caching & caching) != 0) 883 result |= man->default_caching; 884 else if ((TTM_PL_FLAG_CACHED & caching) != 0) 885 result |= TTM_PL_FLAG_CACHED; 886 else if ((TTM_PL_FLAG_WC & caching) != 0) 887 result |= TTM_PL_FLAG_WC; 888 else if ((TTM_PL_FLAG_UNCACHED & caching) != 0) 889 result |= TTM_PL_FLAG_UNCACHED; 890 891 return result; 892 } 893 894 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man, 895 uint32_t mem_type, 896 const struct ttm_place *place, 897 uint32_t *masked_placement) 898 { 899 uint32_t cur_flags = ttm_bo_type_flags(mem_type); 900 901 if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0) 902 return false; 903 904 if ((place->flags & man->available_caching) == 0) 905 return false; 906 907 cur_flags |= (place->flags & man->available_caching); 908 909 *masked_placement = cur_flags; 910 return true; 911 } 912 913 /** 914 * Creates space for memory region @mem according to its type. 915 * 916 * This function first searches for free space in compatible memory types in 917 * the priority order defined by the driver. If free space isn't found, then 918 * ttm_bo_mem_force_space is attempted in priority order to evict and find 919 * space. 920 */ 921 int ttm_bo_mem_space(struct ttm_buffer_object *bo, 922 struct ttm_placement *placement, 923 struct ttm_mem_reg *mem, 924 struct ttm_operation_ctx *ctx) 925 { 926 struct ttm_bo_device *bdev = bo->bdev; 927 struct ttm_mem_type_manager *man; 928 uint32_t mem_type = TTM_PL_SYSTEM; 929 uint32_t cur_flags = 0; 930 bool type_found = false; 931 bool type_ok = false; 932 bool has_erestartsys = false; 933 int i, ret; 934 935 ret = reservation_object_reserve_shared(bo->resv); 936 if (unlikely(ret)) 937 return ret; 938 939 mem->mm_node = NULL; 940 for (i = 0; i < placement->num_placement; ++i) { 941 const struct ttm_place *place = &placement->placement[i]; 942 943 ret = ttm_mem_type_from_place(place, &mem_type); 944 if (ret) 945 return ret; 946 man = &bdev->man[mem_type]; 947 if (!man->has_type || !man->use_type) 948 continue; 949 950 type_ok = ttm_bo_mt_compatible(man, mem_type, place, 951 &cur_flags); 952 953 if (!type_ok) 954 continue; 955 956 type_found = true; 957 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 958 cur_flags); 959 /* 960 * Use the access and other non-mapping-related flag bits from 961 * the memory placement flags to the current flags 962 */ 963 ttm_flag_masked(&cur_flags, place->flags, 964 ~TTM_PL_MASK_MEMTYPE); 965 966 if (mem_type == TTM_PL_SYSTEM) 967 break; 968 969 ret = (*man->func->get_node)(man, bo, place, mem); 970 if (unlikely(ret)) 971 return ret; 972 973 if (mem->mm_node) { 974 ret = ttm_bo_add_move_fence(bo, man, mem); 975 if (unlikely(ret)) { 976 (*man->func->put_node)(man, mem); 977 return ret; 978 } 979 break; 980 } 981 } 982 983 if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) { 984 mem->mem_type = mem_type; 985 mem->placement = cur_flags; 986 return 0; 987 } 988 989 for (i = 0; i < placement->num_busy_placement; ++i) { 990 const struct ttm_place *place = &placement->busy_placement[i]; 991 992 ret = ttm_mem_type_from_place(place, &mem_type); 993 if (ret) 994 return ret; 995 man = &bdev->man[mem_type]; 996 if (!man->has_type || !man->use_type) 997 continue; 998 if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags)) 999 continue; 1000 1001 type_found = true; 1002 cur_flags = ttm_bo_select_caching(man, bo->mem.placement, 1003 cur_flags); 1004 /* 1005 * Use the access and other non-mapping-related flag bits from 1006 * the memory placement flags to the current flags 1007 */ 1008 ttm_flag_masked(&cur_flags, place->flags, 1009 ~TTM_PL_MASK_MEMTYPE); 1010 1011 if (mem_type == TTM_PL_SYSTEM) { 1012 mem->mem_type = mem_type; 1013 mem->placement = cur_flags; 1014 mem->mm_node = NULL; 1015 return 0; 1016 } 1017 1018 ret = ttm_bo_mem_force_space(bo, mem_type, place, mem, ctx); 1019 if (ret == 0 && mem->mm_node) { 1020 mem->placement = cur_flags; 1021 return 0; 1022 } 1023 if (ret == -ERESTARTSYS) 1024 has_erestartsys = true; 1025 } 1026 1027 if (!type_found) { 1028 pr_err(TTM_PFX "No compatible memory type found\n"); 1029 return -EINVAL; 1030 } 1031 1032 return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM; 1033 } 1034 EXPORT_SYMBOL(ttm_bo_mem_space); 1035 1036 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo, 1037 struct ttm_placement *placement, 1038 struct ttm_operation_ctx *ctx) 1039 { 1040 int ret = 0; 1041 struct ttm_mem_reg mem; 1042 1043 reservation_object_assert_held(bo->resv); 1044 1045 mem.num_pages = bo->num_pages; 1046 mem.size = mem.num_pages << PAGE_SHIFT; 1047 mem.page_alignment = bo->mem.page_alignment; 1048 mem.bus.io_reserved_vm = false; 1049 mem.bus.io_reserved_count = 0; 1050 /* 1051 * Determine where to move the buffer. 1052 */ 1053 ret = ttm_bo_mem_space(bo, placement, &mem, ctx); 1054 if (ret) 1055 goto out_unlock; 1056 ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx); 1057 out_unlock: 1058 if (ret && mem.mm_node) 1059 ttm_bo_mem_put(bo, &mem); 1060 return ret; 1061 } 1062 1063 static bool ttm_bo_places_compat(const struct ttm_place *places, 1064 unsigned num_placement, 1065 struct ttm_mem_reg *mem, 1066 uint32_t *new_flags) 1067 { 1068 unsigned i; 1069 1070 for (i = 0; i < num_placement; i++) { 1071 const struct ttm_place *heap = &places[i]; 1072 1073 if (mem->mm_node && (mem->start < heap->fpfn || 1074 (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn))) 1075 continue; 1076 1077 *new_flags = heap->flags; 1078 if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) && 1079 (*new_flags & mem->placement & TTM_PL_MASK_MEM) && 1080 (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) || 1081 (mem->placement & TTM_PL_FLAG_CONTIGUOUS))) 1082 return true; 1083 } 1084 return false; 1085 } 1086 1087 bool ttm_bo_mem_compat(struct ttm_placement *placement, 1088 struct ttm_mem_reg *mem, 1089 uint32_t *new_flags) 1090 { 1091 if (ttm_bo_places_compat(placement->placement, placement->num_placement, 1092 mem, new_flags)) 1093 return true; 1094 1095 if ((placement->busy_placement != placement->placement || 1096 placement->num_busy_placement > placement->num_placement) && 1097 ttm_bo_places_compat(placement->busy_placement, 1098 placement->num_busy_placement, 1099 mem, new_flags)) 1100 return true; 1101 1102 return false; 1103 } 1104 EXPORT_SYMBOL(ttm_bo_mem_compat); 1105 1106 int ttm_bo_validate(struct ttm_buffer_object *bo, 1107 struct ttm_placement *placement, 1108 struct ttm_operation_ctx *ctx) 1109 { 1110 int ret; 1111 uint32_t new_flags; 1112 1113 reservation_object_assert_held(bo->resv); 1114 /* 1115 * Check whether we need to move buffer. 1116 */ 1117 if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) { 1118 ret = ttm_bo_move_buffer(bo, placement, ctx); 1119 if (ret) 1120 return ret; 1121 } else { 1122 /* 1123 * Use the access and other non-mapping-related flag bits from 1124 * the compatible memory placement flags to the active flags 1125 */ 1126 ttm_flag_masked(&bo->mem.placement, new_flags, 1127 ~TTM_PL_MASK_MEMTYPE); 1128 } 1129 /* 1130 * We might need to add a TTM. 1131 */ 1132 if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) { 1133 ret = ttm_bo_add_ttm(bo, true); 1134 if (ret) 1135 return ret; 1136 } 1137 return 0; 1138 } 1139 EXPORT_SYMBOL(ttm_bo_validate); 1140 1141 int ttm_bo_init_reserved(struct ttm_bo_device *bdev, 1142 struct ttm_buffer_object *bo, 1143 unsigned long size, 1144 enum ttm_bo_type type, 1145 struct ttm_placement *placement, 1146 uint32_t page_alignment, 1147 struct ttm_operation_ctx *ctx, 1148 struct file *persistent_swap_storage, 1149 size_t acc_size, 1150 struct sg_table *sg, 1151 struct reservation_object *resv, 1152 void (*destroy) (struct ttm_buffer_object *)) 1153 { 1154 int ret = 0; 1155 unsigned long num_pages; 1156 struct ttm_mem_global *mem_glob = bdev->glob->mem_glob; 1157 bool locked; 1158 1159 ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx); 1160 if (ret) { 1161 pr_err("Out of kernel memory\n"); 1162 if (destroy) 1163 (*destroy)(bo); 1164 else 1165 kfree(bo); 1166 return -ENOMEM; 1167 } 1168 1169 num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 1170 if (num_pages == 0) { 1171 pr_err("Illegal buffer object size\n"); 1172 if (destroy) 1173 (*destroy)(bo); 1174 else 1175 kfree(bo); 1176 ttm_mem_global_free(mem_glob, acc_size); 1177 return -EINVAL; 1178 } 1179 bo->destroy = destroy; 1180 1181 kref_init(&bo->kref); 1182 kref_init(&bo->list_kref); 1183 atomic_set(&bo->cpu_writers, 0); 1184 INIT_LIST_HEAD(&bo->lru); 1185 INIT_LIST_HEAD(&bo->ddestroy); 1186 INIT_LIST_HEAD(&bo->swap); 1187 INIT_LIST_HEAD(&bo->io_reserve_lru); 1188 mutex_init(&bo->wu_mutex); 1189 bo->bdev = bdev; 1190 bo->glob = bdev->glob; 1191 bo->type = type; 1192 bo->num_pages = num_pages; 1193 bo->mem.size = num_pages << PAGE_SHIFT; 1194 bo->mem.mem_type = TTM_PL_SYSTEM; 1195 bo->mem.num_pages = bo->num_pages; 1196 bo->mem.mm_node = NULL; 1197 bo->mem.page_alignment = page_alignment; 1198 bo->mem.bus.io_reserved_vm = false; 1199 bo->mem.bus.io_reserved_count = 0; 1200 bo->moving = NULL; 1201 bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED); 1202 bo->persistent_swap_storage = persistent_swap_storage; 1203 bo->acc_size = acc_size; 1204 bo->sg = sg; 1205 if (resv) { 1206 bo->resv = resv; 1207 reservation_object_assert_held(bo->resv); 1208 } else { 1209 bo->resv = &bo->ttm_resv; 1210 } 1211 reservation_object_init(&bo->ttm_resv); 1212 atomic_inc(&bo->glob->bo_count); 1213 drm_vma_node_reset(&bo->vma_node); 1214 bo->priority = 0; 1215 1216 /* 1217 * For ttm_bo_type_device buffers, allocate 1218 * address space from the device. 1219 */ 1220 if (bo->type == ttm_bo_type_device || 1221 bo->type == ttm_bo_type_sg) 1222 ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node, 1223 bo->mem.num_pages); 1224 1225 /* passed reservation objects should already be locked, 1226 * since otherwise lockdep will be angered in radeon. 1227 */ 1228 if (!resv) { 1229 locked = reservation_object_trylock(bo->resv); 1230 WARN_ON(!locked); 1231 } 1232 1233 if (likely(!ret)) 1234 ret = ttm_bo_validate(bo, placement, ctx); 1235 1236 if (unlikely(ret)) { 1237 if (!resv) 1238 ttm_bo_unreserve(bo); 1239 1240 ttm_bo_unref(&bo); 1241 return ret; 1242 } 1243 1244 if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) { 1245 spin_lock(&bo->glob->lru_lock); 1246 ttm_bo_add_to_lru(bo); 1247 spin_unlock(&bo->glob->lru_lock); 1248 } 1249 1250 return ret; 1251 } 1252 EXPORT_SYMBOL(ttm_bo_init_reserved); 1253 1254 int ttm_bo_init(struct ttm_bo_device *bdev, 1255 struct ttm_buffer_object *bo, 1256 unsigned long size, 1257 enum ttm_bo_type type, 1258 struct ttm_placement *placement, 1259 uint32_t page_alignment, 1260 bool interruptible, 1261 struct file *persistent_swap_storage, 1262 size_t acc_size, 1263 struct sg_table *sg, 1264 struct reservation_object *resv, 1265 void (*destroy) (struct ttm_buffer_object *)) 1266 { 1267 struct ttm_operation_ctx ctx = { interruptible, false }; 1268 int ret; 1269 1270 ret = ttm_bo_init_reserved(bdev, bo, size, type, placement, 1271 page_alignment, &ctx, 1272 persistent_swap_storage, acc_size, 1273 sg, resv, destroy); 1274 if (ret) 1275 return ret; 1276 1277 if (!resv) 1278 ttm_bo_unreserve(bo); 1279 1280 return 0; 1281 } 1282 EXPORT_SYMBOL(ttm_bo_init); 1283 1284 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev, 1285 unsigned long bo_size, 1286 unsigned struct_size) 1287 { 1288 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1289 size_t size = 0; 1290 1291 size += ttm_round_pot(struct_size); 1292 size += ttm_round_pot(npages * sizeof(void *)); 1293 size += ttm_round_pot(sizeof(struct ttm_tt)); 1294 return size; 1295 } 1296 EXPORT_SYMBOL(ttm_bo_acc_size); 1297 1298 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev, 1299 unsigned long bo_size, 1300 unsigned struct_size) 1301 { 1302 unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT; 1303 size_t size = 0; 1304 1305 size += ttm_round_pot(struct_size); 1306 size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t))); 1307 size += ttm_round_pot(sizeof(struct ttm_dma_tt)); 1308 return size; 1309 } 1310 EXPORT_SYMBOL(ttm_bo_dma_acc_size); 1311 1312 int ttm_bo_create(struct ttm_bo_device *bdev, 1313 unsigned long size, 1314 enum ttm_bo_type type, 1315 struct ttm_placement *placement, 1316 uint32_t page_alignment, 1317 bool interruptible, 1318 struct file *persistent_swap_storage, 1319 struct ttm_buffer_object **p_bo) 1320 { 1321 struct ttm_buffer_object *bo; 1322 size_t acc_size; 1323 int ret; 1324 1325 bo = kzalloc(sizeof(*bo), GFP_KERNEL); 1326 if (unlikely(bo == NULL)) 1327 return -ENOMEM; 1328 1329 acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object)); 1330 ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment, 1331 interruptible, persistent_swap_storage, acc_size, 1332 NULL, NULL, NULL); 1333 if (likely(ret == 0)) 1334 *p_bo = bo; 1335 1336 return ret; 1337 } 1338 EXPORT_SYMBOL(ttm_bo_create); 1339 1340 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev, 1341 unsigned mem_type) 1342 { 1343 struct ttm_operation_ctx ctx = { false, false }; 1344 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1345 struct ttm_bo_global *glob = bdev->glob; 1346 struct dma_fence *fence; 1347 int ret; 1348 unsigned i; 1349 1350 /* 1351 * Can't use standard list traversal since we're unlocking. 1352 */ 1353 1354 spin_lock(&glob->lru_lock); 1355 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 1356 while (!list_empty(&man->lru[i])) { 1357 spin_unlock(&glob->lru_lock); 1358 ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx); 1359 if (ret) 1360 return ret; 1361 spin_lock(&glob->lru_lock); 1362 } 1363 } 1364 spin_unlock(&glob->lru_lock); 1365 1366 spin_lock(&man->move_lock); 1367 fence = dma_fence_get(man->move); 1368 spin_unlock(&man->move_lock); 1369 1370 if (fence) { 1371 ret = dma_fence_wait(fence, false); 1372 dma_fence_put(fence); 1373 if (ret) 1374 return ret; 1375 } 1376 1377 return 0; 1378 } 1379 1380 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1381 { 1382 struct ttm_mem_type_manager *man; 1383 int ret = -EINVAL; 1384 1385 if (mem_type >= TTM_NUM_MEM_TYPES) { 1386 pr_err("Illegal memory type %d\n", mem_type); 1387 return ret; 1388 } 1389 man = &bdev->man[mem_type]; 1390 1391 if (!man->has_type) { 1392 pr_err("Trying to take down uninitialized memory manager type %u\n", 1393 mem_type); 1394 return ret; 1395 } 1396 1397 man->use_type = false; 1398 man->has_type = false; 1399 1400 ret = 0; 1401 if (mem_type > 0) { 1402 ret = ttm_bo_force_list_clean(bdev, mem_type); 1403 if (ret) { 1404 pr_err("Cleanup eviction failed\n"); 1405 return ret; 1406 } 1407 1408 ret = (*man->func->takedown)(man); 1409 } 1410 1411 dma_fence_put(man->move); 1412 man->move = NULL; 1413 1414 return ret; 1415 } 1416 EXPORT_SYMBOL(ttm_bo_clean_mm); 1417 1418 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type) 1419 { 1420 struct ttm_mem_type_manager *man = &bdev->man[mem_type]; 1421 1422 if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) { 1423 pr_err("Illegal memory manager memory type %u\n", mem_type); 1424 return -EINVAL; 1425 } 1426 1427 if (!man->has_type) { 1428 pr_err("Memory type %u has not been initialized\n", mem_type); 1429 return 0; 1430 } 1431 1432 return ttm_bo_force_list_clean(bdev, mem_type); 1433 } 1434 EXPORT_SYMBOL(ttm_bo_evict_mm); 1435 1436 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type, 1437 unsigned long p_size) 1438 { 1439 int ret; 1440 struct ttm_mem_type_manager *man; 1441 unsigned i; 1442 1443 BUG_ON(type >= TTM_NUM_MEM_TYPES); 1444 man = &bdev->man[type]; 1445 BUG_ON(man->has_type); 1446 man->io_reserve_fastpath = true; 1447 man->use_io_reserve_lru = false; 1448 mutex_init(&man->io_reserve_mutex); 1449 spin_lock_init(&man->move_lock); 1450 INIT_LIST_HEAD(&man->io_reserve_lru); 1451 1452 ret = bdev->driver->init_mem_type(bdev, type, man); 1453 if (ret) 1454 return ret; 1455 man->bdev = bdev; 1456 1457 if (type != TTM_PL_SYSTEM) { 1458 ret = (*man->func->init)(man, p_size); 1459 if (ret) 1460 return ret; 1461 } 1462 man->has_type = true; 1463 man->use_type = true; 1464 man->size = p_size; 1465 1466 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1467 INIT_LIST_HEAD(&man->lru[i]); 1468 man->move = NULL; 1469 1470 return 0; 1471 } 1472 EXPORT_SYMBOL(ttm_bo_init_mm); 1473 1474 static void ttm_bo_global_kobj_release(struct kobject *kobj) 1475 { 1476 struct ttm_bo_global *glob = 1477 container_of(kobj, struct ttm_bo_global, kobj); 1478 1479 __free_page(glob->dummy_read_page); 1480 kfree(glob); 1481 } 1482 1483 void ttm_bo_global_release(struct drm_global_reference *ref) 1484 { 1485 struct ttm_bo_global *glob = ref->object; 1486 1487 kobject_del(&glob->kobj); 1488 kobject_put(&glob->kobj); 1489 } 1490 EXPORT_SYMBOL(ttm_bo_global_release); 1491 1492 int ttm_bo_global_init(struct drm_global_reference *ref) 1493 { 1494 struct ttm_bo_global_ref *bo_ref = 1495 container_of(ref, struct ttm_bo_global_ref, ref); 1496 struct ttm_bo_global *glob = ref->object; 1497 int ret; 1498 unsigned i; 1499 1500 mutex_init(&glob->device_list_mutex); 1501 spin_lock_init(&glob->lru_lock); 1502 glob->mem_glob = bo_ref->mem_glob; 1503 glob->mem_glob->bo_glob = glob; 1504 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32); 1505 1506 if (unlikely(glob->dummy_read_page == NULL)) { 1507 ret = -ENOMEM; 1508 goto out_no_drp; 1509 } 1510 1511 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1512 INIT_LIST_HEAD(&glob->swap_lru[i]); 1513 INIT_LIST_HEAD(&glob->device_list); 1514 atomic_set(&glob->bo_count, 0); 1515 1516 ret = kobject_init_and_add( 1517 &glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects"); 1518 if (unlikely(ret != 0)) 1519 kobject_put(&glob->kobj); 1520 return ret; 1521 out_no_drp: 1522 kfree(glob); 1523 return ret; 1524 } 1525 EXPORT_SYMBOL(ttm_bo_global_init); 1526 1527 1528 int ttm_bo_device_release(struct ttm_bo_device *bdev) 1529 { 1530 int ret = 0; 1531 unsigned i = TTM_NUM_MEM_TYPES; 1532 struct ttm_mem_type_manager *man; 1533 struct ttm_bo_global *glob = bdev->glob; 1534 1535 while (i--) { 1536 man = &bdev->man[i]; 1537 if (man->has_type) { 1538 man->use_type = false; 1539 if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) { 1540 ret = -EBUSY; 1541 pr_err("DRM memory manager type %d is not clean\n", 1542 i); 1543 } 1544 man->has_type = false; 1545 } 1546 } 1547 1548 mutex_lock(&glob->device_list_mutex); 1549 list_del(&bdev->device_list); 1550 mutex_unlock(&glob->device_list_mutex); 1551 1552 cancel_delayed_work_sync(&bdev->wq); 1553 1554 if (ttm_bo_delayed_delete(bdev, true)) 1555 pr_debug("Delayed destroy list was clean\n"); 1556 1557 spin_lock(&glob->lru_lock); 1558 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 1559 if (list_empty(&bdev->man[0].lru[0])) 1560 pr_debug("Swap list %d was clean\n", i); 1561 spin_unlock(&glob->lru_lock); 1562 1563 drm_vma_offset_manager_destroy(&bdev->vma_manager); 1564 1565 return ret; 1566 } 1567 EXPORT_SYMBOL(ttm_bo_device_release); 1568 1569 int ttm_bo_device_init(struct ttm_bo_device *bdev, 1570 struct ttm_bo_global *glob, 1571 struct ttm_bo_driver *driver, 1572 struct address_space *mapping, 1573 uint64_t file_page_offset, 1574 bool need_dma32) 1575 { 1576 int ret = -EINVAL; 1577 1578 bdev->driver = driver; 1579 1580 memset(bdev->man, 0, sizeof(bdev->man)); 1581 1582 /* 1583 * Initialize the system memory buffer type. 1584 * Other types need to be driver / IOCTL initialized. 1585 */ 1586 ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0); 1587 if (unlikely(ret != 0)) 1588 goto out_no_sys; 1589 1590 drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset, 1591 0x10000000); 1592 INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue); 1593 INIT_LIST_HEAD(&bdev->ddestroy); 1594 bdev->dev_mapping = mapping; 1595 bdev->glob = glob; 1596 bdev->need_dma32 = need_dma32; 1597 mutex_lock(&glob->device_list_mutex); 1598 list_add_tail(&bdev->device_list, &glob->device_list); 1599 mutex_unlock(&glob->device_list_mutex); 1600 1601 return 0; 1602 out_no_sys: 1603 return ret; 1604 } 1605 EXPORT_SYMBOL(ttm_bo_device_init); 1606 1607 /* 1608 * buffer object vm functions. 1609 */ 1610 1611 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem) 1612 { 1613 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 1614 1615 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) { 1616 if (mem->mem_type == TTM_PL_SYSTEM) 1617 return false; 1618 1619 if (man->flags & TTM_MEMTYPE_FLAG_CMA) 1620 return false; 1621 1622 if (mem->placement & TTM_PL_FLAG_CACHED) 1623 return false; 1624 } 1625 return true; 1626 } 1627 1628 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo) 1629 { 1630 struct ttm_bo_device *bdev = bo->bdev; 1631 1632 drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping); 1633 ttm_mem_io_free_vm(bo); 1634 } 1635 1636 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo) 1637 { 1638 struct ttm_bo_device *bdev = bo->bdev; 1639 struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type]; 1640 1641 ttm_mem_io_lock(man, false); 1642 ttm_bo_unmap_virtual_locked(bo); 1643 ttm_mem_io_unlock(man); 1644 } 1645 1646 1647 EXPORT_SYMBOL(ttm_bo_unmap_virtual); 1648 1649 int ttm_bo_wait(struct ttm_buffer_object *bo, 1650 bool interruptible, bool no_wait) 1651 { 1652 long timeout = 15 * HZ; 1653 1654 if (no_wait) { 1655 if (reservation_object_test_signaled_rcu(bo->resv, true)) 1656 return 0; 1657 else 1658 return -EBUSY; 1659 } 1660 1661 timeout = reservation_object_wait_timeout_rcu(bo->resv, true, 1662 interruptible, timeout); 1663 if (timeout < 0) 1664 return timeout; 1665 1666 if (timeout == 0) 1667 return -EBUSY; 1668 1669 reservation_object_add_excl_fence(bo->resv, NULL); 1670 return 0; 1671 } 1672 EXPORT_SYMBOL(ttm_bo_wait); 1673 1674 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait) 1675 { 1676 int ret = 0; 1677 1678 /* 1679 * Using ttm_bo_reserve makes sure the lru lists are updated. 1680 */ 1681 1682 ret = ttm_bo_reserve(bo, true, no_wait, NULL); 1683 if (unlikely(ret != 0)) 1684 return ret; 1685 ret = ttm_bo_wait(bo, true, no_wait); 1686 if (likely(ret == 0)) 1687 atomic_inc(&bo->cpu_writers); 1688 ttm_bo_unreserve(bo); 1689 return ret; 1690 } 1691 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab); 1692 1693 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo) 1694 { 1695 atomic_dec(&bo->cpu_writers); 1696 } 1697 EXPORT_SYMBOL(ttm_bo_synccpu_write_release); 1698 1699 /** 1700 * A buffer object shrink method that tries to swap out the first 1701 * buffer object on the bo_global::swap_lru list. 1702 */ 1703 int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx) 1704 { 1705 struct ttm_buffer_object *bo; 1706 int ret = -EBUSY; 1707 bool locked; 1708 unsigned i; 1709 1710 spin_lock(&glob->lru_lock); 1711 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) { 1712 list_for_each_entry(bo, &glob->swap_lru[i], swap) { 1713 if (ttm_bo_evict_swapout_allowable(bo, ctx, &locked)) { 1714 ret = 0; 1715 break; 1716 } 1717 } 1718 if (!ret) 1719 break; 1720 } 1721 1722 if (ret) { 1723 spin_unlock(&glob->lru_lock); 1724 return ret; 1725 } 1726 1727 kref_get(&bo->list_kref); 1728 1729 if (!list_empty(&bo->ddestroy)) { 1730 ret = ttm_bo_cleanup_refs(bo, false, false, locked); 1731 kref_put(&bo->list_kref, ttm_bo_release_list); 1732 return ret; 1733 } 1734 1735 ttm_bo_del_from_lru(bo); 1736 spin_unlock(&glob->lru_lock); 1737 1738 /** 1739 * Move to system cached 1740 */ 1741 1742 if (bo->mem.mem_type != TTM_PL_SYSTEM || 1743 bo->ttm->caching_state != tt_cached) { 1744 struct ttm_operation_ctx ctx = { false, false }; 1745 struct ttm_mem_reg evict_mem; 1746 1747 evict_mem = bo->mem; 1748 evict_mem.mm_node = NULL; 1749 evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED; 1750 evict_mem.mem_type = TTM_PL_SYSTEM; 1751 1752 ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx); 1753 if (unlikely(ret != 0)) 1754 goto out; 1755 } 1756 1757 /** 1758 * Make sure BO is idle. 1759 */ 1760 1761 ret = ttm_bo_wait(bo, false, false); 1762 if (unlikely(ret != 0)) 1763 goto out; 1764 1765 ttm_bo_unmap_virtual(bo); 1766 1767 /** 1768 * Swap out. Buffer will be swapped in again as soon as 1769 * anyone tries to access a ttm page. 1770 */ 1771 1772 if (bo->bdev->driver->swap_notify) 1773 bo->bdev->driver->swap_notify(bo); 1774 1775 ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage); 1776 out: 1777 1778 /** 1779 * 1780 * Unreserve without putting on LRU to avoid swapping out an 1781 * already swapped buffer. 1782 */ 1783 if (locked) 1784 reservation_object_unlock(bo->resv); 1785 kref_put(&bo->list_kref, ttm_bo_release_list); 1786 return ret; 1787 } 1788 EXPORT_SYMBOL(ttm_bo_swapout); 1789 1790 void ttm_bo_swapout_all(struct ttm_bo_device *bdev) 1791 { 1792 struct ttm_operation_ctx ctx = { 1793 .interruptible = false, 1794 .no_wait_gpu = false 1795 }; 1796 1797 while (ttm_bo_swapout(bdev->glob, &ctx) == 0) 1798 ; 1799 } 1800 EXPORT_SYMBOL(ttm_bo_swapout_all); 1801 1802 /** 1803 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become 1804 * unreserved 1805 * 1806 * @bo: Pointer to buffer 1807 */ 1808 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo) 1809 { 1810 int ret; 1811 1812 /* 1813 * In the absense of a wait_unlocked API, 1814 * Use the bo::wu_mutex to avoid triggering livelocks due to 1815 * concurrent use of this function. Note that this use of 1816 * bo::wu_mutex can go away if we change locking order to 1817 * mmap_sem -> bo::reserve. 1818 */ 1819 ret = mutex_lock_interruptible(&bo->wu_mutex); 1820 if (unlikely(ret != 0)) 1821 return -ERESTARTSYS; 1822 if (!ww_mutex_is_locked(&bo->resv->lock)) 1823 goto out_unlock; 1824 ret = reservation_object_lock_interruptible(bo->resv, NULL); 1825 if (ret == -EINTR) 1826 ret = -ERESTARTSYS; 1827 if (unlikely(ret != 0)) 1828 goto out_unlock; 1829 reservation_object_unlock(bo->resv); 1830 1831 out_unlock: 1832 mutex_unlock(&bo->wu_mutex); 1833 return ret; 1834 } 1835