1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */ 2 /************************************************************************** 3 * 4 * Copyright (c) 2007-2009 VMware, Inc., Palo Alto, CA., USA 5 * All Rights Reserved. 6 * 7 * Permission is hereby granted, free of charge, to any person obtaining a 8 * copy of this software and associated documentation files (the 9 * "Software"), to deal in the Software without restriction, including 10 * without limitation the rights to use, copy, modify, merge, publish, 11 * distribute, sub license, and/or sell copies of the Software, and to 12 * permit persons to whom the Software is furnished to do so, subject to 13 * the following conditions: 14 * 15 * The above copyright notice and this permission notice (including the 16 * next paragraph) shall be included in all copies or substantial portions 17 * of the Software. 18 * 19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 25 * USE OR OTHER DEALINGS IN THE SOFTWARE. 26 * 27 **************************************************************************/ 28 /* 29 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com> 30 */ 31 32 #include <drm/ttm/ttm_bo_driver.h> 33 #include <drm/ttm/ttm_placement.h> 34 #include <drm/drm_vma_manager.h> 35 #include <linux/io.h> 36 #include <linux/highmem.h> 37 #include <linux/wait.h> 38 #include <linux/slab.h> 39 #include <linux/vmalloc.h> 40 #include <linux/module.h> 41 #include <linux/dma-resv.h> 42 43 struct ttm_transfer_obj { 44 struct ttm_buffer_object base; 45 struct ttm_buffer_object *bo; 46 }; 47 48 void ttm_bo_free_old_node(struct ttm_buffer_object *bo) 49 { 50 ttm_bo_mem_put(bo, &bo->mem); 51 } 52 53 int ttm_bo_move_ttm(struct ttm_buffer_object *bo, 54 struct ttm_operation_ctx *ctx, 55 struct ttm_mem_reg *new_mem) 56 { 57 struct ttm_tt *ttm = bo->ttm; 58 struct ttm_mem_reg *old_mem = &bo->mem; 59 int ret; 60 61 if (old_mem->mem_type != TTM_PL_SYSTEM) { 62 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu); 63 64 if (unlikely(ret != 0)) { 65 if (ret != -ERESTARTSYS) 66 pr_err("Failed to expire sync object before unbinding TTM\n"); 67 return ret; 68 } 69 70 ttm_tt_unbind(ttm); 71 ttm_bo_free_old_node(bo); 72 ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM, 73 TTM_PL_MASK_MEM); 74 old_mem->mem_type = TTM_PL_SYSTEM; 75 } 76 77 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement); 78 if (unlikely(ret != 0)) 79 return ret; 80 81 if (new_mem->mem_type != TTM_PL_SYSTEM) { 82 ret = ttm_tt_bind(ttm, new_mem, ctx); 83 if (unlikely(ret != 0)) 84 return ret; 85 } 86 87 *old_mem = *new_mem; 88 new_mem->mm_node = NULL; 89 90 return 0; 91 } 92 EXPORT_SYMBOL(ttm_bo_move_ttm); 93 94 int ttm_mem_io_lock(struct ttm_mem_type_manager *man, bool interruptible) 95 { 96 if (likely(man->io_reserve_fastpath)) 97 return 0; 98 99 if (interruptible) 100 return mutex_lock_interruptible(&man->io_reserve_mutex); 101 102 mutex_lock(&man->io_reserve_mutex); 103 return 0; 104 } 105 106 void ttm_mem_io_unlock(struct ttm_mem_type_manager *man) 107 { 108 if (likely(man->io_reserve_fastpath)) 109 return; 110 111 mutex_unlock(&man->io_reserve_mutex); 112 } 113 114 static int ttm_mem_io_evict(struct ttm_mem_type_manager *man) 115 { 116 struct ttm_buffer_object *bo; 117 118 if (!man->use_io_reserve_lru || list_empty(&man->io_reserve_lru)) 119 return -EAGAIN; 120 121 bo = list_first_entry(&man->io_reserve_lru, 122 struct ttm_buffer_object, 123 io_reserve_lru); 124 list_del_init(&bo->io_reserve_lru); 125 ttm_bo_unmap_virtual_locked(bo); 126 127 return 0; 128 } 129 130 131 int ttm_mem_io_reserve(struct ttm_bo_device *bdev, 132 struct ttm_mem_reg *mem) 133 { 134 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 135 int ret = 0; 136 137 if (!bdev->driver->io_mem_reserve) 138 return 0; 139 if (likely(man->io_reserve_fastpath)) 140 return bdev->driver->io_mem_reserve(bdev, mem); 141 142 if (bdev->driver->io_mem_reserve && 143 mem->bus.io_reserved_count++ == 0) { 144 retry: 145 ret = bdev->driver->io_mem_reserve(bdev, mem); 146 if (ret == -EAGAIN) { 147 ret = ttm_mem_io_evict(man); 148 if (ret == 0) 149 goto retry; 150 } 151 } 152 return ret; 153 } 154 155 void ttm_mem_io_free(struct ttm_bo_device *bdev, 156 struct ttm_mem_reg *mem) 157 { 158 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 159 160 if (likely(man->io_reserve_fastpath)) 161 return; 162 163 if (bdev->driver->io_mem_reserve && 164 --mem->bus.io_reserved_count == 0 && 165 bdev->driver->io_mem_free) 166 bdev->driver->io_mem_free(bdev, mem); 167 168 } 169 170 int ttm_mem_io_reserve_vm(struct ttm_buffer_object *bo) 171 { 172 struct ttm_mem_reg *mem = &bo->mem; 173 int ret; 174 175 if (!mem->bus.io_reserved_vm) { 176 struct ttm_mem_type_manager *man = 177 &bo->bdev->man[mem->mem_type]; 178 179 ret = ttm_mem_io_reserve(bo->bdev, mem); 180 if (unlikely(ret != 0)) 181 return ret; 182 mem->bus.io_reserved_vm = true; 183 if (man->use_io_reserve_lru) 184 list_add_tail(&bo->io_reserve_lru, 185 &man->io_reserve_lru); 186 } 187 return 0; 188 } 189 190 void ttm_mem_io_free_vm(struct ttm_buffer_object *bo) 191 { 192 struct ttm_mem_reg *mem = &bo->mem; 193 194 if (mem->bus.io_reserved_vm) { 195 mem->bus.io_reserved_vm = false; 196 list_del_init(&bo->io_reserve_lru); 197 ttm_mem_io_free(bo->bdev, mem); 198 } 199 } 200 201 static int ttm_mem_reg_ioremap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem, 202 void **virtual) 203 { 204 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 205 int ret; 206 void *addr; 207 208 *virtual = NULL; 209 (void) ttm_mem_io_lock(man, false); 210 ret = ttm_mem_io_reserve(bdev, mem); 211 ttm_mem_io_unlock(man); 212 if (ret || !mem->bus.is_iomem) 213 return ret; 214 215 if (mem->bus.addr) { 216 addr = mem->bus.addr; 217 } else { 218 if (mem->placement & TTM_PL_FLAG_WC) 219 addr = ioremap_wc(mem->bus.base + mem->bus.offset, mem->bus.size); 220 else 221 addr = ioremap_nocache(mem->bus.base + mem->bus.offset, mem->bus.size); 222 if (!addr) { 223 (void) ttm_mem_io_lock(man, false); 224 ttm_mem_io_free(bdev, mem); 225 ttm_mem_io_unlock(man); 226 return -ENOMEM; 227 } 228 } 229 *virtual = addr; 230 return 0; 231 } 232 233 static void ttm_mem_reg_iounmap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem, 234 void *virtual) 235 { 236 struct ttm_mem_type_manager *man; 237 238 man = &bdev->man[mem->mem_type]; 239 240 if (virtual && mem->bus.addr == NULL) 241 iounmap(virtual); 242 (void) ttm_mem_io_lock(man, false); 243 ttm_mem_io_free(bdev, mem); 244 ttm_mem_io_unlock(man); 245 } 246 247 static int ttm_copy_io_page(void *dst, void *src, unsigned long page) 248 { 249 uint32_t *dstP = 250 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT)); 251 uint32_t *srcP = 252 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT)); 253 254 int i; 255 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i) 256 iowrite32(ioread32(srcP++), dstP++); 257 return 0; 258 } 259 260 #ifdef CONFIG_X86 261 #define __ttm_kmap_atomic_prot(__page, __prot) kmap_atomic_prot(__page, __prot) 262 #define __ttm_kunmap_atomic(__addr) kunmap_atomic(__addr) 263 #else 264 #define __ttm_kmap_atomic_prot(__page, __prot) vmap(&__page, 1, 0, __prot) 265 #define __ttm_kunmap_atomic(__addr) vunmap(__addr) 266 #endif 267 268 269 /** 270 * ttm_kmap_atomic_prot - Efficient kernel map of a single page with 271 * specified page protection. 272 * 273 * @page: The page to map. 274 * @prot: The page protection. 275 * 276 * This function maps a TTM page using the kmap_atomic api if available, 277 * otherwise falls back to vmap. The user must make sure that the 278 * specified page does not have an aliased mapping with a different caching 279 * policy unless the architecture explicitly allows it. Also mapping and 280 * unmapping using this api must be correctly nested. Unmapping should 281 * occur in the reverse order of mapping. 282 */ 283 void *ttm_kmap_atomic_prot(struct page *page, pgprot_t prot) 284 { 285 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) 286 return kmap_atomic(page); 287 else 288 return __ttm_kmap_atomic_prot(page, prot); 289 } 290 EXPORT_SYMBOL(ttm_kmap_atomic_prot); 291 292 /** 293 * ttm_kunmap_atomic_prot - Unmap a page that was mapped using 294 * ttm_kmap_atomic_prot. 295 * 296 * @addr: The virtual address from the map. 297 * @prot: The page protection. 298 */ 299 void ttm_kunmap_atomic_prot(void *addr, pgprot_t prot) 300 { 301 if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) 302 kunmap_atomic(addr); 303 else 304 __ttm_kunmap_atomic(addr); 305 } 306 EXPORT_SYMBOL(ttm_kunmap_atomic_prot); 307 308 static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src, 309 unsigned long page, 310 pgprot_t prot) 311 { 312 struct page *d = ttm->pages[page]; 313 void *dst; 314 315 if (!d) 316 return -ENOMEM; 317 318 src = (void *)((unsigned long)src + (page << PAGE_SHIFT)); 319 dst = ttm_kmap_atomic_prot(d, prot); 320 if (!dst) 321 return -ENOMEM; 322 323 memcpy_fromio(dst, src, PAGE_SIZE); 324 325 ttm_kunmap_atomic_prot(dst, prot); 326 327 return 0; 328 } 329 330 static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst, 331 unsigned long page, 332 pgprot_t prot) 333 { 334 struct page *s = ttm->pages[page]; 335 void *src; 336 337 if (!s) 338 return -ENOMEM; 339 340 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT)); 341 src = ttm_kmap_atomic_prot(s, prot); 342 if (!src) 343 return -ENOMEM; 344 345 memcpy_toio(dst, src, PAGE_SIZE); 346 347 ttm_kunmap_atomic_prot(src, prot); 348 349 return 0; 350 } 351 352 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo, 353 struct ttm_operation_ctx *ctx, 354 struct ttm_mem_reg *new_mem) 355 { 356 struct ttm_bo_device *bdev = bo->bdev; 357 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type]; 358 struct ttm_tt *ttm = bo->ttm; 359 struct ttm_mem_reg *old_mem = &bo->mem; 360 struct ttm_mem_reg old_copy = *old_mem; 361 void *old_iomap; 362 void *new_iomap; 363 int ret; 364 unsigned long i; 365 unsigned long page; 366 unsigned long add = 0; 367 int dir; 368 369 ret = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu); 370 if (ret) 371 return ret; 372 373 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap); 374 if (ret) 375 return ret; 376 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap); 377 if (ret) 378 goto out; 379 380 /* 381 * Single TTM move. NOP. 382 */ 383 if (old_iomap == NULL && new_iomap == NULL) 384 goto out2; 385 386 /* 387 * Don't move nonexistent data. Clear destination instead. 388 */ 389 if (old_iomap == NULL && 390 (ttm == NULL || (ttm->state == tt_unpopulated && 391 !(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)))) { 392 memset_io(new_iomap, 0, new_mem->num_pages*PAGE_SIZE); 393 goto out2; 394 } 395 396 /* 397 * TTM might be null for moves within the same region. 398 */ 399 if (ttm) { 400 ret = ttm_tt_populate(ttm, ctx); 401 if (ret) 402 goto out1; 403 } 404 405 add = 0; 406 dir = 1; 407 408 if ((old_mem->mem_type == new_mem->mem_type) && 409 (new_mem->start < old_mem->start + old_mem->size)) { 410 dir = -1; 411 add = new_mem->num_pages - 1; 412 } 413 414 for (i = 0; i < new_mem->num_pages; ++i) { 415 page = i * dir + add; 416 if (old_iomap == NULL) { 417 pgprot_t prot = ttm_io_prot(old_mem->placement, 418 PAGE_KERNEL); 419 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page, 420 prot); 421 } else if (new_iomap == NULL) { 422 pgprot_t prot = ttm_io_prot(new_mem->placement, 423 PAGE_KERNEL); 424 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page, 425 prot); 426 } else { 427 ret = ttm_copy_io_page(new_iomap, old_iomap, page); 428 } 429 if (ret) 430 goto out1; 431 } 432 mb(); 433 out2: 434 old_copy = *old_mem; 435 *old_mem = *new_mem; 436 new_mem->mm_node = NULL; 437 438 if (man->flags & TTM_MEMTYPE_FLAG_FIXED) { 439 ttm_tt_destroy(ttm); 440 bo->ttm = NULL; 441 } 442 443 out1: 444 ttm_mem_reg_iounmap(bdev, old_mem, new_iomap); 445 out: 446 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap); 447 448 /* 449 * On error, keep the mm node! 450 */ 451 if (!ret) 452 ttm_bo_mem_put(bo, &old_copy); 453 return ret; 454 } 455 EXPORT_SYMBOL(ttm_bo_move_memcpy); 456 457 static void ttm_transfered_destroy(struct ttm_buffer_object *bo) 458 { 459 struct ttm_transfer_obj *fbo; 460 461 fbo = container_of(bo, struct ttm_transfer_obj, base); 462 ttm_bo_put(fbo->bo); 463 kfree(fbo); 464 } 465 466 /** 467 * ttm_buffer_object_transfer 468 * 469 * @bo: A pointer to a struct ttm_buffer_object. 470 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object, 471 * holding the data of @bo with the old placement. 472 * 473 * This is a utility function that may be called after an accelerated move 474 * has been scheduled. A new buffer object is created as a placeholder for 475 * the old data while it's being copied. When that buffer object is idle, 476 * it can be destroyed, releasing the space of the old placement. 477 * Returns: 478 * !0: Failure. 479 */ 480 481 static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo, 482 struct ttm_buffer_object **new_obj) 483 { 484 struct ttm_transfer_obj *fbo; 485 int ret; 486 487 fbo = kmalloc(sizeof(*fbo), GFP_KERNEL); 488 if (!fbo) 489 return -ENOMEM; 490 491 fbo->base = *bo; 492 fbo->base.mem.placement |= TTM_PL_FLAG_NO_EVICT; 493 494 ttm_bo_get(bo); 495 fbo->bo = bo; 496 497 /** 498 * Fix up members that we shouldn't copy directly: 499 * TODO: Explicit member copy would probably be better here. 500 */ 501 502 atomic_inc(&ttm_bo_glob.bo_count); 503 INIT_LIST_HEAD(&fbo->base.ddestroy); 504 INIT_LIST_HEAD(&fbo->base.lru); 505 INIT_LIST_HEAD(&fbo->base.swap); 506 INIT_LIST_HEAD(&fbo->base.io_reserve_lru); 507 mutex_init(&fbo->base.wu_mutex); 508 fbo->base.moving = NULL; 509 drm_vma_node_reset(&fbo->base.base.vma_node); 510 511 kref_init(&fbo->base.list_kref); 512 kref_init(&fbo->base.kref); 513 fbo->base.destroy = &ttm_transfered_destroy; 514 fbo->base.acc_size = 0; 515 if (bo->base.resv == &bo->base._resv) 516 fbo->base.base.resv = &fbo->base.base._resv; 517 518 dma_resv_init(&fbo->base.base._resv); 519 ret = dma_resv_trylock(&fbo->base.base._resv); 520 WARN_ON(!ret); 521 522 *new_obj = &fbo->base; 523 return 0; 524 } 525 526 pgprot_t ttm_io_prot(uint32_t caching_flags, pgprot_t tmp) 527 { 528 /* Cached mappings need no adjustment */ 529 if (caching_flags & TTM_PL_FLAG_CACHED) 530 return tmp; 531 532 #if defined(__i386__) || defined(__x86_64__) 533 if (caching_flags & TTM_PL_FLAG_WC) 534 tmp = pgprot_writecombine(tmp); 535 else if (boot_cpu_data.x86 > 3) 536 tmp = pgprot_noncached(tmp); 537 #endif 538 #if defined(__ia64__) || defined(__arm__) || defined(__aarch64__) || \ 539 defined(__powerpc__) || defined(__mips__) 540 if (caching_flags & TTM_PL_FLAG_WC) 541 tmp = pgprot_writecombine(tmp); 542 else 543 tmp = pgprot_noncached(tmp); 544 #endif 545 #if defined(__sparc__) 546 tmp = pgprot_noncached(tmp); 547 #endif 548 return tmp; 549 } 550 EXPORT_SYMBOL(ttm_io_prot); 551 552 static int ttm_bo_ioremap(struct ttm_buffer_object *bo, 553 unsigned long offset, 554 unsigned long size, 555 struct ttm_bo_kmap_obj *map) 556 { 557 struct ttm_mem_reg *mem = &bo->mem; 558 559 if (bo->mem.bus.addr) { 560 map->bo_kmap_type = ttm_bo_map_premapped; 561 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset); 562 } else { 563 map->bo_kmap_type = ttm_bo_map_iomap; 564 if (mem->placement & TTM_PL_FLAG_WC) 565 map->virtual = ioremap_wc(bo->mem.bus.base + bo->mem.bus.offset + offset, 566 size); 567 else 568 map->virtual = ioremap_nocache(bo->mem.bus.base + bo->mem.bus.offset + offset, 569 size); 570 } 571 return (!map->virtual) ? -ENOMEM : 0; 572 } 573 574 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo, 575 unsigned long start_page, 576 unsigned long num_pages, 577 struct ttm_bo_kmap_obj *map) 578 { 579 struct ttm_mem_reg *mem = &bo->mem; 580 struct ttm_operation_ctx ctx = { 581 .interruptible = false, 582 .no_wait_gpu = false 583 }; 584 struct ttm_tt *ttm = bo->ttm; 585 pgprot_t prot; 586 int ret; 587 588 BUG_ON(!ttm); 589 590 ret = ttm_tt_populate(ttm, &ctx); 591 if (ret) 592 return ret; 593 594 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) { 595 /* 596 * We're mapping a single page, and the desired 597 * page protection is consistent with the bo. 598 */ 599 600 map->bo_kmap_type = ttm_bo_map_kmap; 601 map->page = ttm->pages[start_page]; 602 map->virtual = kmap(map->page); 603 } else { 604 /* 605 * We need to use vmap to get the desired page protection 606 * or to make the buffer object look contiguous. 607 */ 608 prot = ttm_io_prot(mem->placement, PAGE_KERNEL); 609 map->bo_kmap_type = ttm_bo_map_vmap; 610 map->virtual = vmap(ttm->pages + start_page, num_pages, 611 0, prot); 612 } 613 return (!map->virtual) ? -ENOMEM : 0; 614 } 615 616 int ttm_bo_kmap(struct ttm_buffer_object *bo, 617 unsigned long start_page, unsigned long num_pages, 618 struct ttm_bo_kmap_obj *map) 619 { 620 struct ttm_mem_type_manager *man = 621 &bo->bdev->man[bo->mem.mem_type]; 622 unsigned long offset, size; 623 int ret; 624 625 map->virtual = NULL; 626 map->bo = bo; 627 if (num_pages > bo->num_pages) 628 return -EINVAL; 629 if (start_page > bo->num_pages) 630 return -EINVAL; 631 632 (void) ttm_mem_io_lock(man, false); 633 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem); 634 ttm_mem_io_unlock(man); 635 if (ret) 636 return ret; 637 if (!bo->mem.bus.is_iomem) { 638 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map); 639 } else { 640 offset = start_page << PAGE_SHIFT; 641 size = num_pages << PAGE_SHIFT; 642 return ttm_bo_ioremap(bo, offset, size, map); 643 } 644 } 645 EXPORT_SYMBOL(ttm_bo_kmap); 646 647 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map) 648 { 649 struct ttm_buffer_object *bo = map->bo; 650 struct ttm_mem_type_manager *man = 651 &bo->bdev->man[bo->mem.mem_type]; 652 653 if (!map->virtual) 654 return; 655 switch (map->bo_kmap_type) { 656 case ttm_bo_map_iomap: 657 iounmap(map->virtual); 658 break; 659 case ttm_bo_map_vmap: 660 vunmap(map->virtual); 661 break; 662 case ttm_bo_map_kmap: 663 kunmap(map->page); 664 break; 665 case ttm_bo_map_premapped: 666 break; 667 default: 668 BUG(); 669 } 670 (void) ttm_mem_io_lock(man, false); 671 ttm_mem_io_free(map->bo->bdev, &map->bo->mem); 672 ttm_mem_io_unlock(man); 673 map->virtual = NULL; 674 map->page = NULL; 675 } 676 EXPORT_SYMBOL(ttm_bo_kunmap); 677 678 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo, 679 struct dma_fence *fence, 680 bool evict, 681 struct ttm_mem_reg *new_mem) 682 { 683 struct ttm_bo_device *bdev = bo->bdev; 684 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type]; 685 struct ttm_mem_reg *old_mem = &bo->mem; 686 int ret; 687 struct ttm_buffer_object *ghost_obj; 688 689 dma_resv_add_excl_fence(bo->base.resv, fence); 690 if (evict) { 691 ret = ttm_bo_wait(bo, false, false); 692 if (ret) 693 return ret; 694 695 if (man->flags & TTM_MEMTYPE_FLAG_FIXED) { 696 ttm_tt_destroy(bo->ttm); 697 bo->ttm = NULL; 698 } 699 ttm_bo_free_old_node(bo); 700 } else { 701 /** 702 * This should help pipeline ordinary buffer moves. 703 * 704 * Hang old buffer memory on a new buffer object, 705 * and leave it to be released when the GPU 706 * operation has completed. 707 */ 708 709 dma_fence_put(bo->moving); 710 bo->moving = dma_fence_get(fence); 711 712 ret = ttm_buffer_object_transfer(bo, &ghost_obj); 713 if (ret) 714 return ret; 715 716 dma_resv_add_excl_fence(&ghost_obj->base._resv, fence); 717 718 /** 719 * If we're not moving to fixed memory, the TTM object 720 * needs to stay alive. Otherwhise hang it on the ghost 721 * bo to be unbound and destroyed. 722 */ 723 724 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) 725 ghost_obj->ttm = NULL; 726 else 727 bo->ttm = NULL; 728 729 dma_resv_unlock(&ghost_obj->base._resv); 730 ttm_bo_put(ghost_obj); 731 } 732 733 *old_mem = *new_mem; 734 new_mem->mm_node = NULL; 735 736 return 0; 737 } 738 EXPORT_SYMBOL(ttm_bo_move_accel_cleanup); 739 740 int ttm_bo_pipeline_move(struct ttm_buffer_object *bo, 741 struct dma_fence *fence, bool evict, 742 struct ttm_mem_reg *new_mem) 743 { 744 struct ttm_bo_device *bdev = bo->bdev; 745 struct ttm_mem_reg *old_mem = &bo->mem; 746 747 struct ttm_mem_type_manager *from = &bdev->man[old_mem->mem_type]; 748 struct ttm_mem_type_manager *to = &bdev->man[new_mem->mem_type]; 749 750 int ret; 751 752 dma_resv_add_excl_fence(bo->base.resv, fence); 753 754 if (!evict) { 755 struct ttm_buffer_object *ghost_obj; 756 757 /** 758 * This should help pipeline ordinary buffer moves. 759 * 760 * Hang old buffer memory on a new buffer object, 761 * and leave it to be released when the GPU 762 * operation has completed. 763 */ 764 765 dma_fence_put(bo->moving); 766 bo->moving = dma_fence_get(fence); 767 768 ret = ttm_buffer_object_transfer(bo, &ghost_obj); 769 if (ret) 770 return ret; 771 772 dma_resv_add_excl_fence(&ghost_obj->base._resv, fence); 773 774 /** 775 * If we're not moving to fixed memory, the TTM object 776 * needs to stay alive. Otherwhise hang it on the ghost 777 * bo to be unbound and destroyed. 778 */ 779 780 if (!(to->flags & TTM_MEMTYPE_FLAG_FIXED)) 781 ghost_obj->ttm = NULL; 782 else 783 bo->ttm = NULL; 784 785 dma_resv_unlock(&ghost_obj->base._resv); 786 ttm_bo_put(ghost_obj); 787 788 } else if (from->flags & TTM_MEMTYPE_FLAG_FIXED) { 789 790 /** 791 * BO doesn't have a TTM we need to bind/unbind. Just remember 792 * this eviction and free up the allocation 793 */ 794 795 spin_lock(&from->move_lock); 796 if (!from->move || dma_fence_is_later(fence, from->move)) { 797 dma_fence_put(from->move); 798 from->move = dma_fence_get(fence); 799 } 800 spin_unlock(&from->move_lock); 801 802 ttm_bo_free_old_node(bo); 803 804 dma_fence_put(bo->moving); 805 bo->moving = dma_fence_get(fence); 806 807 } else { 808 /** 809 * Last resort, wait for the move to be completed. 810 * 811 * Should never happen in pratice. 812 */ 813 814 ret = ttm_bo_wait(bo, false, false); 815 if (ret) 816 return ret; 817 818 if (to->flags & TTM_MEMTYPE_FLAG_FIXED) { 819 ttm_tt_destroy(bo->ttm); 820 bo->ttm = NULL; 821 } 822 ttm_bo_free_old_node(bo); 823 } 824 825 *old_mem = *new_mem; 826 new_mem->mm_node = NULL; 827 828 return 0; 829 } 830 EXPORT_SYMBOL(ttm_bo_pipeline_move); 831 832 int ttm_bo_pipeline_gutting(struct ttm_buffer_object *bo) 833 { 834 struct ttm_buffer_object *ghost; 835 int ret; 836 837 ret = ttm_buffer_object_transfer(bo, &ghost); 838 if (ret) 839 return ret; 840 841 ret = dma_resv_copy_fences(&ghost->base._resv, bo->base.resv); 842 /* Last resort, wait for the BO to be idle when we are OOM */ 843 if (ret) 844 ttm_bo_wait(bo, false, false); 845 846 memset(&bo->mem, 0, sizeof(bo->mem)); 847 bo->mem.mem_type = TTM_PL_SYSTEM; 848 bo->ttm = NULL; 849 850 dma_resv_unlock(&ghost->base._resv); 851 ttm_bo_put(ghost); 852 853 return 0; 854 } 855