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 #ifdef CONFIG_X86 259 #define __ttm_kmap_atomic_prot(__page, __prot) kmap_atomic_prot(__page, __prot) 260 #define __ttm_kunmap_atomic(__addr) kunmap_atomic(__addr) 261 #else 262 #define __ttm_kmap_atomic_prot(__page, __prot) vmap(&__page, 1, 0, __prot) 263 #define __ttm_kunmap_atomic(__addr) vunmap(__addr) 264 #endif 265 266 267 /** 268 * ttm_kmap_atomic_prot - Efficient kernel map of a single page with 269 * specified page protection. 270 * 271 * @page: The page to map. 272 * @prot: The page protection. 273 * 274 * This function maps a TTM page using the kmap_atomic api if available, 275 * otherwise falls back to vmap. The user must make sure that the 276 * specified page does not have an aliased mapping with a different caching 277 * policy unless the architecture explicitly allows it. Also mapping and 278 * unmapping using this api must be correctly nested. Unmapping should 279 * occur in the reverse order of mapping. 280 */ 281 void *ttm_kmap_atomic_prot(struct page *page, pgprot_t prot) 282 { 283 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) 284 return kmap_atomic(page); 285 else 286 return __ttm_kmap_atomic_prot(page, prot); 287 } 288 EXPORT_SYMBOL(ttm_kmap_atomic_prot); 289 290 /** 291 * ttm_kunmap_atomic_prot - Unmap a page that was mapped using 292 * ttm_kmap_atomic_prot. 293 * 294 * @addr: The virtual address from the map. 295 * @prot: The page protection. 296 */ 297 void ttm_kunmap_atomic_prot(void *addr, pgprot_t prot) 298 { 299 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) 300 kunmap_atomic(addr); 301 else 302 __ttm_kunmap_atomic(addr); 303 } 304 EXPORT_SYMBOL(ttm_kunmap_atomic_prot); 305 306 static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src, 307 unsigned long page, 308 pgprot_t prot) 309 { 310 struct page *d = ttm->pages[page]; 311 void *dst; 312 313 if (!d) 314 return -ENOMEM; 315 316 src = (void *)((unsigned long)src + (page << PAGE_SHIFT)); 317 dst = ttm_kmap_atomic_prot(d, prot); 318 if (!dst) 319 return -ENOMEM; 320 321 memcpy_fromio(dst, src, PAGE_SIZE); 322 323 ttm_kunmap_atomic_prot(dst, prot); 324 325 return 0; 326 } 327 328 static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst, 329 unsigned long page, 330 pgprot_t prot) 331 { 332 struct page *s = ttm->pages[page]; 333 void *src; 334 335 if (!s) 336 return -ENOMEM; 337 338 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT)); 339 src = ttm_kmap_atomic_prot(s, prot); 340 if (!src) 341 return -ENOMEM; 342 343 memcpy_toio(dst, src, PAGE_SIZE); 344 345 ttm_kunmap_atomic_prot(src, prot); 346 347 return 0; 348 } 349 350 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo, 351 struct ttm_operation_ctx *ctx, 352 struct ttm_mem_reg *new_mem) 353 { 354 struct ttm_bo_device *bdev = bo->bdev; 355 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type]; 356 struct ttm_tt *ttm = bo->ttm; 357 struct ttm_mem_reg *old_mem = &bo->mem; 358 struct ttm_mem_reg old_copy = *old_mem; 359 void *old_iomap; 360 void *new_iomap; 361 int ret; 362 unsigned long i; 363 unsigned long page; 364 unsigned long add = 0; 365 int dir; 366 367 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu); 368 if (ret) 369 return ret; 370 371 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap); 372 if (ret) 373 return ret; 374 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap); 375 if (ret) 376 goto out; 377 378 /* 379 * Single TTM move. NOP. 380 */ 381 if (old_iomap == NULL && new_iomap == NULL) 382 goto out2; 383 384 /* 385 * Don't move nonexistent data. Clear destination instead. 386 */ 387 if (old_iomap == NULL && 388 (ttm == NULL || (ttm->state == tt_unpopulated && 389 !(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)))) { 390 memset_io(new_iomap, 0, new_mem->num_pages*PAGE_SIZE); 391 goto out2; 392 } 393 394 /* 395 * TTM might be null for moves within the same region. 396 */ 397 if (ttm) { 398 ret = ttm_tt_populate(ttm, ctx); 399 if (ret) 400 goto out1; 401 } 402 403 add = 0; 404 dir = 1; 405 406 if ((old_mem->mem_type == new_mem->mem_type) && 407 (new_mem->start < old_mem->start + old_mem->size)) { 408 dir = -1; 409 add = new_mem->num_pages - 1; 410 } 411 412 for (i = 0; i < new_mem->num_pages; ++i) { 413 page = i * dir + add; 414 if (old_iomap == NULL) { 415 pgprot_t prot = ttm_io_prot(old_mem->placement, 416 PAGE_KERNEL); 417 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page, 418 prot); 419 } else if (new_iomap == NULL) { 420 pgprot_t prot = ttm_io_prot(new_mem->placement, 421 PAGE_KERNEL); 422 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page, 423 prot); 424 } else { 425 ret = ttm_copy_io_page(new_iomap, old_iomap, page); 426 } 427 if (ret) 428 goto out1; 429 } 430 mb(); 431 out2: 432 old_copy = *old_mem; 433 *old_mem = *new_mem; 434 new_mem->mm_node = NULL; 435 436 if (man->flags & TTM_MEMTYPE_FLAG_FIXED) { 437 ttm_tt_destroy(ttm); 438 bo->ttm = NULL; 439 } 440 441 out1: 442 ttm_mem_reg_iounmap(bdev, old_mem, new_iomap); 443 out: 444 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap); 445 446 /* 447 * On error, keep the mm node! 448 */ 449 if (!ret) 450 ttm_bo_mem_put(bo, &old_copy); 451 return ret; 452 } 453 EXPORT_SYMBOL(ttm_bo_move_memcpy); 454 455 static void ttm_transfered_destroy(struct ttm_buffer_object *bo) 456 { 457 kfree(bo); 458 } 459 460 /** 461 * ttm_buffer_object_transfer 462 * 463 * @bo: A pointer to a struct ttm_buffer_object. 464 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object, 465 * holding the data of @bo with the old placement. 466 * 467 * This is a utility function that may be called after an accelerated move 468 * has been scheduled. A new buffer object is created as a placeholder for 469 * the old data while it's being copied. When that buffer object is idle, 470 * it can be destroyed, releasing the space of the old placement. 471 * Returns: 472 * !0: Failure. 473 */ 474 475 static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo, 476 struct ttm_buffer_object **new_obj) 477 { 478 struct ttm_buffer_object *fbo; 479 int ret; 480 481 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL); 482 if (!fbo) 483 return -ENOMEM; 484 485 *fbo = *bo; 486 487 /** 488 * Fix up members that we shouldn't copy directly: 489 * TODO: Explicit member copy would probably be better here. 490 */ 491 492 atomic_inc(&bo->bdev->glob->bo_count); 493 INIT_LIST_HEAD(&fbo->ddestroy); 494 INIT_LIST_HEAD(&fbo->lru); 495 INIT_LIST_HEAD(&fbo->swap); 496 INIT_LIST_HEAD(&fbo->io_reserve_lru); 497 mutex_init(&fbo->wu_mutex); 498 fbo->moving = NULL; 499 drm_vma_node_reset(&fbo->vma_node); 500 atomic_set(&fbo->cpu_writers, 0); 501 502 kref_init(&fbo->list_kref); 503 kref_init(&fbo->kref); 504 fbo->destroy = &ttm_transfered_destroy; 505 fbo->acc_size = 0; 506 fbo->resv = &fbo->ttm_resv; 507 reservation_object_init(fbo->resv); 508 ret = reservation_object_trylock(fbo->resv); 509 WARN_ON(!ret); 510 511 *new_obj = fbo; 512 return 0; 513 } 514 515 pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp) 516 { 517 /* Cached mappings need no adjustment */ 518 if (caching_flags & TTM_PL_FLAG_CACHED) 519 return tmp; 520 521 #if defined(__i386__) || defined(__x86_64__) 522 if (caching_flags & TTM_PL_FLAG_WC) 523 tmp = pgprot_writecombine(tmp); 524 else if (boot_cpu_data.x86 > 3) 525 tmp = pgprot_noncached(tmp); 526 #endif 527 #if defined(__ia64__) || defined(__arm__) || defined(__aarch64__) || \ 528 defined(__powerpc__) 529 if (caching_flags & TTM_PL_FLAG_WC) 530 tmp = pgprot_writecombine(tmp); 531 else 532 tmp = pgprot_noncached(tmp); 533 #endif 534 #if defined(__sparc__) || defined(__mips__) 535 tmp = pgprot_noncached(tmp); 536 #endif 537 return tmp; 538 } 539 EXPORT_SYMBOL(ttm_io_prot); 540 541 static int ttm_bo_ioremap(struct ttm_buffer_object *bo, 542 unsigned long offset, 543 unsigned long size, 544 struct ttm_bo_kmap_obj *map) 545 { 546 struct ttm_mem_reg *mem = &bo->mem; 547 548 if (bo->mem.bus.addr) { 549 map->bo_kmap_type = ttm_bo_map_premapped; 550 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset); 551 } else { 552 map->bo_kmap_type = ttm_bo_map_iomap; 553 if (mem->placement & TTM_PL_FLAG_WC) 554 map->virtual = ioremap_wc(bo->mem.bus.base + bo->mem.bus.offset + offset, 555 size); 556 else 557 map->virtual = ioremap_nocache(bo->mem.bus.base + bo->mem.bus.offset + offset, 558 size); 559 } 560 return (!map->virtual) ? -ENOMEM : 0; 561 } 562 563 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo, 564 unsigned long start_page, 565 unsigned long num_pages, 566 struct ttm_bo_kmap_obj *map) 567 { 568 struct ttm_mem_reg *mem = &bo->mem; 569 struct ttm_operation_ctx ctx = { 570 .interruptible = false, 571 .no_wait_gpu = false 572 }; 573 struct ttm_tt *ttm = bo->ttm; 574 pgprot_t prot; 575 int ret; 576 577 BUG_ON(!ttm); 578 579 ret = ttm_tt_populate(ttm, &ctx); 580 if (ret) 581 return ret; 582 583 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) { 584 /* 585 * We're mapping a single page, and the desired 586 * page protection is consistent with the bo. 587 */ 588 589 map->bo_kmap_type = ttm_bo_map_kmap; 590 map->page = ttm->pages[start_page]; 591 map->virtual = kmap(map->page); 592 } else { 593 /* 594 * We need to use vmap to get the desired page protection 595 * or to make the buffer object look contiguous. 596 */ 597 prot = ttm_io_prot(mem->placement, PAGE_KERNEL); 598 map->bo_kmap_type = ttm_bo_map_vmap; 599 map->virtual = vmap(ttm->pages + start_page, num_pages, 600 0, prot); 601 } 602 return (!map->virtual) ? -ENOMEM : 0; 603 } 604 605 int ttm_bo_kmap(struct ttm_buffer_object *bo, 606 unsigned long start_page, unsigned long num_pages, 607 struct ttm_bo_kmap_obj *map) 608 { 609 struct ttm_mem_type_manager *man = 610 &bo->bdev->man[bo->mem.mem_type]; 611 unsigned long offset, size; 612 int ret; 613 614 map->virtual = NULL; 615 map->bo = bo; 616 if (num_pages > bo->num_pages) 617 return -EINVAL; 618 if (start_page > bo->num_pages) 619 return -EINVAL; 620 #if 0 621 if (num_pages > 1 && !capable(CAP_SYS_ADMIN)) 622 return -EPERM; 623 #endif 624 (void) ttm_mem_io_lock(man, false); 625 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem); 626 ttm_mem_io_unlock(man); 627 if (ret) 628 return ret; 629 if (!bo->mem.bus.is_iomem) { 630 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map); 631 } else { 632 offset = start_page << PAGE_SHIFT; 633 size = num_pages << PAGE_SHIFT; 634 return ttm_bo_ioremap(bo, offset, size, map); 635 } 636 } 637 EXPORT_SYMBOL(ttm_bo_kmap); 638 639 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map) 640 { 641 struct ttm_buffer_object *bo = map->bo; 642 struct ttm_mem_type_manager *man = 643 &bo->bdev->man[bo->mem.mem_type]; 644 645 if (!map->virtual) 646 return; 647 switch (map->bo_kmap_type) { 648 case ttm_bo_map_iomap: 649 iounmap(map->virtual); 650 break; 651 case ttm_bo_map_vmap: 652 vunmap(map->virtual); 653 break; 654 case ttm_bo_map_kmap: 655 kunmap(map->page); 656 break; 657 case ttm_bo_map_premapped: 658 break; 659 default: 660 BUG(); 661 } 662 (void) ttm_mem_io_lock(man, false); 663 ttm_mem_io_free(map->bo->bdev, &map->bo->mem); 664 ttm_mem_io_unlock(man); 665 map->virtual = NULL; 666 map->page = NULL; 667 } 668 EXPORT_SYMBOL(ttm_bo_kunmap); 669 670 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo, 671 struct dma_fence *fence, 672 bool evict, 673 struct ttm_mem_reg *new_mem) 674 { 675 struct ttm_bo_device *bdev = bo->bdev; 676 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type]; 677 struct ttm_mem_reg *old_mem = &bo->mem; 678 int ret; 679 struct ttm_buffer_object *ghost_obj; 680 681 reservation_object_add_excl_fence(bo->resv, fence); 682 if (evict) { 683 ret = ttm_bo_wait(bo, false, false); 684 if (ret) 685 return ret; 686 687 if (man->flags & TTM_MEMTYPE_FLAG_FIXED) { 688 ttm_tt_destroy(bo->ttm); 689 bo->ttm = NULL; 690 } 691 ttm_bo_free_old_node(bo); 692 } else { 693 /** 694 * This should help pipeline ordinary buffer moves. 695 * 696 * Hang old buffer memory on a new buffer object, 697 * and leave it to be released when the GPU 698 * operation has completed. 699 */ 700 701 dma_fence_put(bo->moving); 702 bo->moving = dma_fence_get(fence); 703 704 ret = ttm_buffer_object_transfer(bo, &ghost_obj); 705 if (ret) 706 return ret; 707 708 reservation_object_add_excl_fence(ghost_obj->resv, fence); 709 710 /** 711 * If we're not moving to fixed memory, the TTM object 712 * needs to stay alive. Otherwhise hang it on the ghost 713 * bo to be unbound and destroyed. 714 */ 715 716 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) 717 ghost_obj->ttm = NULL; 718 else 719 bo->ttm = NULL; 720 721 ttm_bo_unreserve(ghost_obj); 722 ttm_bo_unref(&ghost_obj); 723 } 724 725 *old_mem = *new_mem; 726 new_mem->mm_node = NULL; 727 728 return 0; 729 } 730 EXPORT_SYMBOL(ttm_bo_move_accel_cleanup); 731 732 int ttm_bo_pipeline_move(struct ttm_buffer_object *bo, 733 struct dma_fence *fence, bool evict, 734 struct ttm_mem_reg *new_mem) 735 { 736 struct ttm_bo_device *bdev = bo->bdev; 737 struct ttm_mem_reg *old_mem = &bo->mem; 738 739 struct ttm_mem_type_manager *from = &bdev->man[old_mem->mem_type]; 740 struct ttm_mem_type_manager *to = &bdev->man[new_mem->mem_type]; 741 742 int ret; 743 744 reservation_object_add_excl_fence(bo->resv, fence); 745 746 if (!evict) { 747 struct ttm_buffer_object *ghost_obj; 748 749 /** 750 * This should help pipeline ordinary buffer moves. 751 * 752 * Hang old buffer memory on a new buffer object, 753 * and leave it to be released when the GPU 754 * operation has completed. 755 */ 756 757 dma_fence_put(bo->moving); 758 bo->moving = dma_fence_get(fence); 759 760 ret = ttm_buffer_object_transfer(bo, &ghost_obj); 761 if (ret) 762 return ret; 763 764 reservation_object_add_excl_fence(ghost_obj->resv, fence); 765 766 /** 767 * If we're not moving to fixed memory, the TTM object 768 * needs to stay alive. Otherwhise hang it on the ghost 769 * bo to be unbound and destroyed. 770 */ 771 772 if (!(to->flags & TTM_MEMTYPE_FLAG_FIXED)) 773 ghost_obj->ttm = NULL; 774 else 775 bo->ttm = NULL; 776 777 ttm_bo_unreserve(ghost_obj); 778 ttm_bo_unref(&ghost_obj); 779 780 } else if (from->flags & TTM_MEMTYPE_FLAG_FIXED) { 781 782 /** 783 * BO doesn't have a TTM we need to bind/unbind. Just remember 784 * this eviction and free up the allocation 785 */ 786 787 spin_lock(&from->move_lock); 788 if (!from->move || dma_fence_is_later(fence, from->move)) { 789 dma_fence_put(from->move); 790 from->move = dma_fence_get(fence); 791 } 792 spin_unlock(&from->move_lock); 793 794 ttm_bo_free_old_node(bo); 795 796 dma_fence_put(bo->moving); 797 bo->moving = dma_fence_get(fence); 798 799 } else { 800 /** 801 * Last resort, wait for the move to be completed. 802 * 803 * Should never happen in pratice. 804 */ 805 806 ret = ttm_bo_wait(bo, false, false); 807 if (ret) 808 return ret; 809 810 if (to->flags & TTM_MEMTYPE_FLAG_FIXED) { 811 ttm_tt_destroy(bo->ttm); 812 bo->ttm = NULL; 813 } 814 ttm_bo_free_old_node(bo); 815 } 816 817 *old_mem = *new_mem; 818 new_mem->mm_node = NULL; 819 820 return 0; 821 } 822 EXPORT_SYMBOL(ttm_bo_pipeline_move); 823 824 int ttm_bo_pipeline_gutting(struct ttm_buffer_object *bo) 825 { 826 struct ttm_buffer_object *ghost; 827 int ret; 828 829 ret = ttm_buffer_object_transfer(bo, &ghost); 830 if (ret) 831 return ret; 832 833 ret = reservation_object_copy_fences(ghost->resv, bo->resv); 834 /* Last resort, wait for the BO to be idle when we are OOM */ 835 if (ret) 836 ttm_bo_wait(bo, false, false); 837 838 memset(&bo->mem, 0, sizeof(bo->mem)); 839 bo->mem.mem_type = TTM_PL_SYSTEM; 840 bo->ttm = NULL; 841 842 ttm_bo_unreserve(ghost); 843 ttm_bo_unref(&ghost); 844 845 return 0; 846 } 847