1 /************************************************************************** 2 * 3 * Copyright (c) 2007-2009 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 #include <drm/ttm/ttm_bo_driver.h> 32 #include <drm/ttm/ttm_placement.h> 33 #include <drm/drm_vma_manager.h> 34 #include <linux/io.h> 35 #include <linux/highmem.h> 36 #include <linux/wait.h> 37 #include <linux/slab.h> 38 #include <linux/vmalloc.h> 39 #include <linux/module.h> 40 #include <linux/reservation.h> 41 42 void ttm_bo_free_old_node(struct ttm_buffer_object *bo) 43 { 44 ttm_bo_mem_put(bo, &bo->mem); 45 } 46 47 int ttm_bo_move_ttm(struct ttm_buffer_object *bo, 48 struct ttm_operation_ctx *ctx, 49 struct ttm_mem_reg *new_mem) 50 { 51 struct ttm_tt *ttm = bo->ttm; 52 struct ttm_mem_reg *old_mem = &bo->mem; 53 int ret; 54 55 if (old_mem->mem_type != TTM_PL_SYSTEM) { 56 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu); 57 58 if (unlikely(ret != 0)) { 59 if (ret != -ERESTARTSYS) 60 pr_err("Failed to expire sync object before unbinding TTM\n"); 61 return ret; 62 } 63 64 ttm_tt_unbind(ttm); 65 ttm_bo_free_old_node(bo); 66 ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM, 67 TTM_PL_MASK_MEM); 68 old_mem->mem_type = TTM_PL_SYSTEM; 69 } 70 71 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement); 72 if (unlikely(ret != 0)) 73 return ret; 74 75 if (new_mem->mem_type != TTM_PL_SYSTEM) { 76 ret = ttm_tt_bind(ttm, new_mem, ctx); 77 if (unlikely(ret != 0)) 78 return ret; 79 } 80 81 *old_mem = *new_mem; 82 new_mem->mm_node = NULL; 83 84 return 0; 85 } 86 EXPORT_SYMBOL(ttm_bo_move_ttm); 87 88 int ttm_mem_io_lock(struct ttm_mem_type_manager *man, bool interruptible) 89 { 90 if (likely(man->io_reserve_fastpath)) 91 return 0; 92 93 if (interruptible) 94 return mutex_lock_interruptible(&man->io_reserve_mutex); 95 96 mutex_lock(&man->io_reserve_mutex); 97 return 0; 98 } 99 EXPORT_SYMBOL(ttm_mem_io_lock); 100 101 void ttm_mem_io_unlock(struct ttm_mem_type_manager *man) 102 { 103 if (likely(man->io_reserve_fastpath)) 104 return; 105 106 mutex_unlock(&man->io_reserve_mutex); 107 } 108 EXPORT_SYMBOL(ttm_mem_io_unlock); 109 110 static int ttm_mem_io_evict(struct ttm_mem_type_manager *man) 111 { 112 struct ttm_buffer_object *bo; 113 114 if (!man->use_io_reserve_lru || list_empty(&man->io_reserve_lru)) 115 return -EAGAIN; 116 117 bo = list_first_entry(&man->io_reserve_lru, 118 struct ttm_buffer_object, 119 io_reserve_lru); 120 list_del_init(&bo->io_reserve_lru); 121 ttm_bo_unmap_virtual_locked(bo); 122 123 return 0; 124 } 125 126 127 int ttm_mem_io_reserve(struct ttm_bo_device *bdev, 128 struct ttm_mem_reg *mem) 129 { 130 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 131 int ret = 0; 132 133 if (!bdev->driver->io_mem_reserve) 134 return 0; 135 if (likely(man->io_reserve_fastpath)) 136 return bdev->driver->io_mem_reserve(bdev, mem); 137 138 if (bdev->driver->io_mem_reserve && 139 mem->bus.io_reserved_count++ == 0) { 140 retry: 141 ret = bdev->driver->io_mem_reserve(bdev, mem); 142 if (ret == -EAGAIN) { 143 ret = ttm_mem_io_evict(man); 144 if (ret == 0) 145 goto retry; 146 } 147 } 148 return ret; 149 } 150 EXPORT_SYMBOL(ttm_mem_io_reserve); 151 152 void ttm_mem_io_free(struct ttm_bo_device *bdev, 153 struct ttm_mem_reg *mem) 154 { 155 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 156 157 if (likely(man->io_reserve_fastpath)) 158 return; 159 160 if (bdev->driver->io_mem_reserve && 161 --mem->bus.io_reserved_count == 0 && 162 bdev->driver->io_mem_free) 163 bdev->driver->io_mem_free(bdev, mem); 164 165 } 166 EXPORT_SYMBOL(ttm_mem_io_free); 167 168 int ttm_mem_io_reserve_vm(struct ttm_buffer_object *bo) 169 { 170 struct ttm_mem_reg *mem = &bo->mem; 171 int ret; 172 173 if (!mem->bus.io_reserved_vm) { 174 struct ttm_mem_type_manager *man = 175 &bo->bdev->man[mem->mem_type]; 176 177 ret = ttm_mem_io_reserve(bo->bdev, mem); 178 if (unlikely(ret != 0)) 179 return ret; 180 mem->bus.io_reserved_vm = true; 181 if (man->use_io_reserve_lru) 182 list_add_tail(&bo->io_reserve_lru, 183 &man->io_reserve_lru); 184 } 185 return 0; 186 } 187 188 void ttm_mem_io_free_vm(struct ttm_buffer_object *bo) 189 { 190 struct ttm_mem_reg *mem = &bo->mem; 191 192 if (mem->bus.io_reserved_vm) { 193 mem->bus.io_reserved_vm = false; 194 list_del_init(&bo->io_reserve_lru); 195 ttm_mem_io_free(bo->bdev, mem); 196 } 197 } 198 199 static int ttm_mem_reg_ioremap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem, 200 void **virtual) 201 { 202 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 203 int ret; 204 void *addr; 205 206 *virtual = NULL; 207 (void) ttm_mem_io_lock(man, false); 208 ret = ttm_mem_io_reserve(bdev, mem); 209 ttm_mem_io_unlock(man); 210 if (ret || !mem->bus.is_iomem) 211 return ret; 212 213 if (mem->bus.addr) { 214 addr = mem->bus.addr; 215 } else { 216 if (mem->placement & TTM_PL_FLAG_WC) 217 addr = ioremap_wc(mem->bus.base + mem->bus.offset, mem->bus.size); 218 else 219 addr = ioremap_nocache(mem->bus.base + mem->bus.offset, mem->bus.size); 220 if (!addr) { 221 (void) ttm_mem_io_lock(man, false); 222 ttm_mem_io_free(bdev, mem); 223 ttm_mem_io_unlock(man); 224 return -ENOMEM; 225 } 226 } 227 *virtual = addr; 228 return 0; 229 } 230 231 static void ttm_mem_reg_iounmap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem, 232 void *virtual) 233 { 234 struct ttm_mem_type_manager *man; 235 236 man = &bdev->man[mem->mem_type]; 237 238 if (virtual && mem->bus.addr == NULL) 239 iounmap(virtual); 240 (void) ttm_mem_io_lock(man, false); 241 ttm_mem_io_free(bdev, mem); 242 ttm_mem_io_unlock(man); 243 } 244 245 static int ttm_copy_io_page(void *dst, void *src, unsigned long page) 246 { 247 uint32_t *dstP = 248 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT)); 249 uint32_t *srcP = 250 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT)); 251 252 int i; 253 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i) 254 iowrite32(ioread32(srcP++), dstP++); 255 return 0; 256 } 257 258 static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src, 259 unsigned long page, 260 pgprot_t prot) 261 { 262 struct page *d = ttm->pages[page]; 263 void *dst; 264 265 if (!d) 266 return -ENOMEM; 267 268 src = (void *)((unsigned long)src + (page << PAGE_SHIFT)); 269 270 #ifdef CONFIG_X86 271 dst = kmap_atomic_prot(d, prot); 272 #else 273 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) 274 dst = vmap(&d, 1, 0, prot); 275 else 276 dst = kmap(d); 277 #endif 278 if (!dst) 279 return -ENOMEM; 280 281 memcpy_fromio(dst, src, PAGE_SIZE); 282 283 #ifdef CONFIG_X86 284 kunmap_atomic(dst); 285 #else 286 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) 287 vunmap(dst); 288 else 289 kunmap(d); 290 #endif 291 292 return 0; 293 } 294 295 static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst, 296 unsigned long page, 297 pgprot_t prot) 298 { 299 struct page *s = ttm->pages[page]; 300 void *src; 301 302 if (!s) 303 return -ENOMEM; 304 305 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT)); 306 #ifdef CONFIG_X86 307 src = kmap_atomic_prot(s, prot); 308 #else 309 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) 310 src = vmap(&s, 1, 0, prot); 311 else 312 src = kmap(s); 313 #endif 314 if (!src) 315 return -ENOMEM; 316 317 memcpy_toio(dst, src, PAGE_SIZE); 318 319 #ifdef CONFIG_X86 320 kunmap_atomic(src); 321 #else 322 if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) 323 vunmap(src); 324 else 325 kunmap(s); 326 #endif 327 328 return 0; 329 } 330 331 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo, 332 struct ttm_operation_ctx *ctx, 333 struct ttm_mem_reg *new_mem) 334 { 335 struct ttm_bo_device *bdev = bo->bdev; 336 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type]; 337 struct ttm_tt *ttm = bo->ttm; 338 struct ttm_mem_reg *old_mem = &bo->mem; 339 struct ttm_mem_reg old_copy = *old_mem; 340 void *old_iomap; 341 void *new_iomap; 342 int ret; 343 unsigned long i; 344 unsigned long page; 345 unsigned long add = 0; 346 int dir; 347 348 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu); 349 if (ret) 350 return ret; 351 352 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap); 353 if (ret) 354 return ret; 355 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap); 356 if (ret) 357 goto out; 358 359 /* 360 * Single TTM move. NOP. 361 */ 362 if (old_iomap == NULL && new_iomap == NULL) 363 goto out2; 364 365 /* 366 * Don't move nonexistent data. Clear destination instead. 367 */ 368 if (old_iomap == NULL && 369 (ttm == NULL || (ttm->state == tt_unpopulated && 370 !(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)))) { 371 memset_io(new_iomap, 0, new_mem->num_pages*PAGE_SIZE); 372 goto out2; 373 } 374 375 /* 376 * TTM might be null for moves within the same region. 377 */ 378 if (ttm && ttm->state == tt_unpopulated) { 379 ret = ttm->bdev->driver->ttm_tt_populate(ttm, ctx); 380 if (ret) 381 goto out1; 382 } 383 384 add = 0; 385 dir = 1; 386 387 if ((old_mem->mem_type == new_mem->mem_type) && 388 (new_mem->start < old_mem->start + old_mem->size)) { 389 dir = -1; 390 add = new_mem->num_pages - 1; 391 } 392 393 for (i = 0; i < new_mem->num_pages; ++i) { 394 page = i * dir + add; 395 if (old_iomap == NULL) { 396 pgprot_t prot = ttm_io_prot(old_mem->placement, 397 PAGE_KERNEL); 398 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page, 399 prot); 400 } else if (new_iomap == NULL) { 401 pgprot_t prot = ttm_io_prot(new_mem->placement, 402 PAGE_KERNEL); 403 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page, 404 prot); 405 } else 406 ret = ttm_copy_io_page(new_iomap, old_iomap, page); 407 if (ret) 408 goto out1; 409 } 410 mb(); 411 out2: 412 old_copy = *old_mem; 413 *old_mem = *new_mem; 414 new_mem->mm_node = NULL; 415 416 if (man->flags & TTM_MEMTYPE_FLAG_FIXED) { 417 ttm_tt_destroy(ttm); 418 bo->ttm = NULL; 419 } 420 421 out1: 422 ttm_mem_reg_iounmap(bdev, old_mem, new_iomap); 423 out: 424 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap); 425 426 /* 427 * On error, keep the mm node! 428 */ 429 if (!ret) 430 ttm_bo_mem_put(bo, &old_copy); 431 return ret; 432 } 433 EXPORT_SYMBOL(ttm_bo_move_memcpy); 434 435 static void ttm_transfered_destroy(struct ttm_buffer_object *bo) 436 { 437 kfree(bo); 438 } 439 440 /** 441 * ttm_buffer_object_transfer 442 * 443 * @bo: A pointer to a struct ttm_buffer_object. 444 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object, 445 * holding the data of @bo with the old placement. 446 * 447 * This is a utility function that may be called after an accelerated move 448 * has been scheduled. A new buffer object is created as a placeholder for 449 * the old data while it's being copied. When that buffer object is idle, 450 * it can be destroyed, releasing the space of the old placement. 451 * Returns: 452 * !0: Failure. 453 */ 454 455 static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo, 456 struct ttm_buffer_object **new_obj) 457 { 458 struct ttm_buffer_object *fbo; 459 int ret; 460 461 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL); 462 if (!fbo) 463 return -ENOMEM; 464 465 *fbo = *bo; 466 467 /** 468 * Fix up members that we shouldn't copy directly: 469 * TODO: Explicit member copy would probably be better here. 470 */ 471 472 atomic_inc(&bo->glob->bo_count); 473 INIT_LIST_HEAD(&fbo->ddestroy); 474 INIT_LIST_HEAD(&fbo->lru); 475 INIT_LIST_HEAD(&fbo->swap); 476 INIT_LIST_HEAD(&fbo->io_reserve_lru); 477 mutex_init(&fbo->wu_mutex); 478 fbo->moving = NULL; 479 drm_vma_node_reset(&fbo->vma_node); 480 atomic_set(&fbo->cpu_writers, 0); 481 482 kref_init(&fbo->list_kref); 483 kref_init(&fbo->kref); 484 fbo->destroy = &ttm_transfered_destroy; 485 fbo->acc_size = 0; 486 fbo->resv = &fbo->ttm_resv; 487 reservation_object_init(fbo->resv); 488 ret = reservation_object_trylock(fbo->resv); 489 WARN_ON(!ret); 490 491 *new_obj = fbo; 492 return 0; 493 } 494 495 pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp) 496 { 497 /* Cached mappings need no adjustment */ 498 if (caching_flags & TTM_PL_FLAG_CACHED) 499 return tmp; 500 501 #if defined(__i386__) || defined(__x86_64__) 502 if (caching_flags & TTM_PL_FLAG_WC) 503 tmp = pgprot_writecombine(tmp); 504 else if (boot_cpu_data.x86 > 3) 505 tmp = pgprot_noncached(tmp); 506 #endif 507 #if defined(__ia64__) || defined(__arm__) || defined(__aarch64__) || \ 508 defined(__powerpc__) 509 if (caching_flags & TTM_PL_FLAG_WC) 510 tmp = pgprot_writecombine(tmp); 511 else 512 tmp = pgprot_noncached(tmp); 513 #endif 514 #if defined(__sparc__) || defined(__mips__) 515 tmp = pgprot_noncached(tmp); 516 #endif 517 return tmp; 518 } 519 EXPORT_SYMBOL(ttm_io_prot); 520 521 static int ttm_bo_ioremap(struct ttm_buffer_object *bo, 522 unsigned long offset, 523 unsigned long size, 524 struct ttm_bo_kmap_obj *map) 525 { 526 struct ttm_mem_reg *mem = &bo->mem; 527 528 if (bo->mem.bus.addr) { 529 map->bo_kmap_type = ttm_bo_map_premapped; 530 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset); 531 } else { 532 map->bo_kmap_type = ttm_bo_map_iomap; 533 if (mem->placement & TTM_PL_FLAG_WC) 534 map->virtual = ioremap_wc(bo->mem.bus.base + bo->mem.bus.offset + offset, 535 size); 536 else 537 map->virtual = ioremap_nocache(bo->mem.bus.base + bo->mem.bus.offset + offset, 538 size); 539 } 540 return (!map->virtual) ? -ENOMEM : 0; 541 } 542 543 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo, 544 unsigned long start_page, 545 unsigned long num_pages, 546 struct ttm_bo_kmap_obj *map) 547 { 548 struct ttm_mem_reg *mem = &bo->mem; 549 struct ttm_operation_ctx ctx = { 550 .interruptible = false, 551 .no_wait_gpu = false 552 }; 553 struct ttm_tt *ttm = bo->ttm; 554 pgprot_t prot; 555 int ret; 556 557 BUG_ON(!ttm); 558 559 if (ttm->state == tt_unpopulated) { 560 ret = ttm->bdev->driver->ttm_tt_populate(ttm, &ctx); 561 if (ret) 562 return ret; 563 } 564 565 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) { 566 /* 567 * We're mapping a single page, and the desired 568 * page protection is consistent with the bo. 569 */ 570 571 map->bo_kmap_type = ttm_bo_map_kmap; 572 map->page = ttm->pages[start_page]; 573 map->virtual = kmap(map->page); 574 } else { 575 /* 576 * We need to use vmap to get the desired page protection 577 * or to make the buffer object look contiguous. 578 */ 579 prot = ttm_io_prot(mem->placement, PAGE_KERNEL); 580 map->bo_kmap_type = ttm_bo_map_vmap; 581 map->virtual = vmap(ttm->pages + start_page, num_pages, 582 0, prot); 583 } 584 return (!map->virtual) ? -ENOMEM : 0; 585 } 586 587 int ttm_bo_kmap(struct ttm_buffer_object *bo, 588 unsigned long start_page, unsigned long num_pages, 589 struct ttm_bo_kmap_obj *map) 590 { 591 struct ttm_mem_type_manager *man = 592 &bo->bdev->man[bo->mem.mem_type]; 593 unsigned long offset, size; 594 int ret; 595 596 map->virtual = NULL; 597 map->bo = bo; 598 if (num_pages > bo->num_pages) 599 return -EINVAL; 600 if (start_page > bo->num_pages) 601 return -EINVAL; 602 #if 0 603 if (num_pages > 1 && !capable(CAP_SYS_ADMIN)) 604 return -EPERM; 605 #endif 606 (void) ttm_mem_io_lock(man, false); 607 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem); 608 ttm_mem_io_unlock(man); 609 if (ret) 610 return ret; 611 if (!bo->mem.bus.is_iomem) { 612 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map); 613 } else { 614 offset = start_page << PAGE_SHIFT; 615 size = num_pages << PAGE_SHIFT; 616 return ttm_bo_ioremap(bo, offset, size, map); 617 } 618 } 619 EXPORT_SYMBOL(ttm_bo_kmap); 620 621 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map) 622 { 623 struct ttm_buffer_object *bo = map->bo; 624 struct ttm_mem_type_manager *man = 625 &bo->bdev->man[bo->mem.mem_type]; 626 627 if (!map->virtual) 628 return; 629 switch (map->bo_kmap_type) { 630 case ttm_bo_map_iomap: 631 iounmap(map->virtual); 632 break; 633 case ttm_bo_map_vmap: 634 vunmap(map->virtual); 635 break; 636 case ttm_bo_map_kmap: 637 kunmap(map->page); 638 break; 639 case ttm_bo_map_premapped: 640 break; 641 default: 642 BUG(); 643 } 644 (void) ttm_mem_io_lock(man, false); 645 ttm_mem_io_free(map->bo->bdev, &map->bo->mem); 646 ttm_mem_io_unlock(man); 647 map->virtual = NULL; 648 map->page = NULL; 649 } 650 EXPORT_SYMBOL(ttm_bo_kunmap); 651 652 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo, 653 struct dma_fence *fence, 654 bool evict, 655 struct ttm_mem_reg *new_mem) 656 { 657 struct ttm_bo_device *bdev = bo->bdev; 658 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type]; 659 struct ttm_mem_reg *old_mem = &bo->mem; 660 int ret; 661 struct ttm_buffer_object *ghost_obj; 662 663 reservation_object_add_excl_fence(bo->resv, fence); 664 if (evict) { 665 ret = ttm_bo_wait(bo, false, false); 666 if (ret) 667 return ret; 668 669 if (man->flags & TTM_MEMTYPE_FLAG_FIXED) { 670 ttm_tt_destroy(bo->ttm); 671 bo->ttm = NULL; 672 } 673 ttm_bo_free_old_node(bo); 674 } else { 675 /** 676 * This should help pipeline ordinary buffer moves. 677 * 678 * Hang old buffer memory on a new buffer object, 679 * and leave it to be released when the GPU 680 * operation has completed. 681 */ 682 683 dma_fence_put(bo->moving); 684 bo->moving = dma_fence_get(fence); 685 686 ret = ttm_buffer_object_transfer(bo, &ghost_obj); 687 if (ret) 688 return ret; 689 690 reservation_object_add_excl_fence(ghost_obj->resv, fence); 691 692 /** 693 * If we're not moving to fixed memory, the TTM object 694 * needs to stay alive. Otherwhise hang it on the ghost 695 * bo to be unbound and destroyed. 696 */ 697 698 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) 699 ghost_obj->ttm = NULL; 700 else 701 bo->ttm = NULL; 702 703 ttm_bo_unreserve(ghost_obj); 704 ttm_bo_unref(&ghost_obj); 705 } 706 707 *old_mem = *new_mem; 708 new_mem->mm_node = NULL; 709 710 return 0; 711 } 712 EXPORT_SYMBOL(ttm_bo_move_accel_cleanup); 713 714 int ttm_bo_pipeline_move(struct ttm_buffer_object *bo, 715 struct dma_fence *fence, bool evict, 716 struct ttm_mem_reg *new_mem) 717 { 718 struct ttm_bo_device *bdev = bo->bdev; 719 struct ttm_mem_reg *old_mem = &bo->mem; 720 721 struct ttm_mem_type_manager *from = &bdev->man[old_mem->mem_type]; 722 struct ttm_mem_type_manager *to = &bdev->man[new_mem->mem_type]; 723 724 int ret; 725 726 reservation_object_add_excl_fence(bo->resv, fence); 727 728 if (!evict) { 729 struct ttm_buffer_object *ghost_obj; 730 731 /** 732 * This should help pipeline ordinary buffer moves. 733 * 734 * Hang old buffer memory on a new buffer object, 735 * and leave it to be released when the GPU 736 * operation has completed. 737 */ 738 739 dma_fence_put(bo->moving); 740 bo->moving = dma_fence_get(fence); 741 742 ret = ttm_buffer_object_transfer(bo, &ghost_obj); 743 if (ret) 744 return ret; 745 746 reservation_object_add_excl_fence(ghost_obj->resv, fence); 747 748 /** 749 * If we're not moving to fixed memory, the TTM object 750 * needs to stay alive. Otherwhise hang it on the ghost 751 * bo to be unbound and destroyed. 752 */ 753 754 if (!(to->flags & TTM_MEMTYPE_FLAG_FIXED)) 755 ghost_obj->ttm = NULL; 756 else 757 bo->ttm = NULL; 758 759 ttm_bo_unreserve(ghost_obj); 760 ttm_bo_unref(&ghost_obj); 761 762 } else if (from->flags & TTM_MEMTYPE_FLAG_FIXED) { 763 764 /** 765 * BO doesn't have a TTM we need to bind/unbind. Just remember 766 * this eviction and free up the allocation 767 */ 768 769 spin_lock(&from->move_lock); 770 if (!from->move || dma_fence_is_later(fence, from->move)) { 771 dma_fence_put(from->move); 772 from->move = dma_fence_get(fence); 773 } 774 spin_unlock(&from->move_lock); 775 776 ttm_bo_free_old_node(bo); 777 778 dma_fence_put(bo->moving); 779 bo->moving = dma_fence_get(fence); 780 781 } else { 782 /** 783 * Last resort, wait for the move to be completed. 784 * 785 * Should never happen in pratice. 786 */ 787 788 ret = ttm_bo_wait(bo, false, false); 789 if (ret) 790 return ret; 791 792 if (to->flags & TTM_MEMTYPE_FLAG_FIXED) { 793 ttm_tt_destroy(bo->ttm); 794 bo->ttm = NULL; 795 } 796 ttm_bo_free_old_node(bo); 797 } 798 799 *old_mem = *new_mem; 800 new_mem->mm_node = NULL; 801 802 return 0; 803 } 804 EXPORT_SYMBOL(ttm_bo_pipeline_move); 805