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 <sys/cdefs.h> 32 __FBSDID("$FreeBSD$"); 33 34 #include <dev/drm2/drmP.h> 35 #include <dev/drm2/ttm/ttm_bo_driver.h> 36 #include <dev/drm2/ttm/ttm_placement.h> 37 #include <sys/sf_buf.h> 38 39 void ttm_bo_free_old_node(struct ttm_buffer_object *bo) 40 { 41 ttm_bo_mem_put(bo, &bo->mem); 42 } 43 44 int ttm_bo_move_ttm(struct ttm_buffer_object *bo, 45 bool evict, 46 bool no_wait_gpu, struct ttm_mem_reg *new_mem) 47 { 48 struct ttm_tt *ttm = bo->ttm; 49 struct ttm_mem_reg *old_mem = &bo->mem; 50 int ret; 51 52 if (old_mem->mem_type != TTM_PL_SYSTEM) { 53 ttm_tt_unbind(ttm); 54 ttm_bo_free_old_node(bo); 55 ttm_flag_masked(&old_mem->placement, TTM_PL_FLAG_SYSTEM, 56 TTM_PL_MASK_MEM); 57 old_mem->mem_type = TTM_PL_SYSTEM; 58 } 59 60 ret = ttm_tt_set_placement_caching(ttm, new_mem->placement); 61 if (unlikely(ret != 0)) 62 return ret; 63 64 if (new_mem->mem_type != TTM_PL_SYSTEM) { 65 ret = ttm_tt_bind(ttm, new_mem); 66 if (unlikely(ret != 0)) 67 return ret; 68 } 69 70 *old_mem = *new_mem; 71 new_mem->mm_node = NULL; 72 73 return 0; 74 } 75 76 int ttm_mem_io_lock(struct ttm_mem_type_manager *man, bool interruptible) 77 { 78 if (likely(man->io_reserve_fastpath)) 79 return 0; 80 81 if (interruptible) { 82 if (sx_xlock_sig(&man->io_reserve_mutex)) 83 return (-EINTR); 84 else 85 return (0); 86 } 87 88 sx_xlock(&man->io_reserve_mutex); 89 return 0; 90 } 91 92 void ttm_mem_io_unlock(struct ttm_mem_type_manager *man) 93 { 94 if (likely(man->io_reserve_fastpath)) 95 return; 96 97 sx_xunlock(&man->io_reserve_mutex); 98 } 99 100 static int ttm_mem_io_evict(struct ttm_mem_type_manager *man) 101 { 102 struct ttm_buffer_object *bo; 103 104 if (!man->use_io_reserve_lru || list_empty(&man->io_reserve_lru)) 105 return -EAGAIN; 106 107 bo = list_first_entry(&man->io_reserve_lru, 108 struct ttm_buffer_object, 109 io_reserve_lru); 110 list_del_init(&bo->io_reserve_lru); 111 ttm_bo_unmap_virtual_locked(bo); 112 113 return 0; 114 } 115 116 static int ttm_mem_io_reserve(struct ttm_bo_device *bdev, 117 struct ttm_mem_reg *mem) 118 { 119 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 120 int ret = 0; 121 122 if (!bdev->driver->io_mem_reserve) 123 return 0; 124 if (likely(man->io_reserve_fastpath)) 125 return bdev->driver->io_mem_reserve(bdev, mem); 126 127 if (bdev->driver->io_mem_reserve && 128 mem->bus.io_reserved_count++ == 0) { 129 retry: 130 ret = bdev->driver->io_mem_reserve(bdev, mem); 131 if (ret == -EAGAIN) { 132 ret = ttm_mem_io_evict(man); 133 if (ret == 0) 134 goto retry; 135 } 136 } 137 return ret; 138 } 139 140 static void ttm_mem_io_free(struct ttm_bo_device *bdev, 141 struct ttm_mem_reg *mem) 142 { 143 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 144 145 if (likely(man->io_reserve_fastpath)) 146 return; 147 148 if (bdev->driver->io_mem_reserve && 149 --mem->bus.io_reserved_count == 0 && 150 bdev->driver->io_mem_free) 151 bdev->driver->io_mem_free(bdev, mem); 152 153 } 154 155 int ttm_mem_io_reserve_vm(struct ttm_buffer_object *bo) 156 { 157 struct ttm_mem_reg *mem = &bo->mem; 158 int ret; 159 160 if (!mem->bus.io_reserved_vm) { 161 struct ttm_mem_type_manager *man = 162 &bo->bdev->man[mem->mem_type]; 163 164 ret = ttm_mem_io_reserve(bo->bdev, mem); 165 if (unlikely(ret != 0)) 166 return ret; 167 mem->bus.io_reserved_vm = true; 168 if (man->use_io_reserve_lru) 169 list_add_tail(&bo->io_reserve_lru, 170 &man->io_reserve_lru); 171 } 172 return 0; 173 } 174 175 void ttm_mem_io_free_vm(struct ttm_buffer_object *bo) 176 { 177 struct ttm_mem_reg *mem = &bo->mem; 178 179 if (mem->bus.io_reserved_vm) { 180 mem->bus.io_reserved_vm = false; 181 list_del_init(&bo->io_reserve_lru); 182 ttm_mem_io_free(bo->bdev, mem); 183 } 184 } 185 186 static 187 int ttm_mem_reg_ioremap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem, 188 void **virtual) 189 { 190 struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type]; 191 int ret; 192 void *addr; 193 194 *virtual = NULL; 195 (void) ttm_mem_io_lock(man, false); 196 ret = ttm_mem_io_reserve(bdev, mem); 197 ttm_mem_io_unlock(man); 198 if (ret || !mem->bus.is_iomem) 199 return ret; 200 201 if (mem->bus.addr) { 202 addr = mem->bus.addr; 203 } else { 204 addr = pmap_mapdev_attr(mem->bus.base + mem->bus.offset, 205 mem->bus.size, (mem->placement & TTM_PL_FLAG_WC) ? 206 VM_MEMATTR_WRITE_COMBINING : VM_MEMATTR_UNCACHEABLE); 207 if (!addr) { 208 (void) ttm_mem_io_lock(man, false); 209 ttm_mem_io_free(bdev, mem); 210 ttm_mem_io_unlock(man); 211 return -ENOMEM; 212 } 213 } 214 *virtual = addr; 215 return 0; 216 } 217 218 static 219 void ttm_mem_reg_iounmap(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem, 220 void *virtual) 221 { 222 struct ttm_mem_type_manager *man; 223 224 man = &bdev->man[mem->mem_type]; 225 226 if (virtual && mem->bus.addr == NULL) 227 pmap_unmapdev(virtual, mem->bus.size); 228 (void) ttm_mem_io_lock(man, false); 229 ttm_mem_io_free(bdev, mem); 230 ttm_mem_io_unlock(man); 231 } 232 233 static int ttm_copy_io_page(void *dst, void *src, unsigned long page) 234 { 235 uint32_t *dstP = 236 (uint32_t *) ((unsigned long)dst + (page << PAGE_SHIFT)); 237 uint32_t *srcP = 238 (uint32_t *) ((unsigned long)src + (page << PAGE_SHIFT)); 239 240 int i; 241 for (i = 0; i < PAGE_SIZE / sizeof(uint32_t); ++i) 242 /* iowrite32(ioread32(srcP++), dstP++); */ 243 *dstP++ = *srcP++; 244 return 0; 245 } 246 247 static int ttm_copy_io_ttm_page(struct ttm_tt *ttm, void *src, 248 unsigned long page, 249 vm_memattr_t prot) 250 { 251 vm_page_t d = ttm->pages[page]; 252 void *dst; 253 254 if (!d) 255 return -ENOMEM; 256 257 src = (void *)((unsigned long)src + (page << PAGE_SHIFT)); 258 259 /* XXXKIB can't sleep ? */ 260 dst = pmap_mapdev_attr(VM_PAGE_TO_PHYS(d), PAGE_SIZE, prot); 261 if (!dst) 262 return -ENOMEM; 263 264 memcpy(dst, src, PAGE_SIZE); 265 266 pmap_unmapdev(dst, PAGE_SIZE); 267 268 return 0; 269 } 270 271 static int ttm_copy_ttm_io_page(struct ttm_tt *ttm, void *dst, 272 unsigned long page, 273 vm_memattr_t prot) 274 { 275 vm_page_t s = ttm->pages[page]; 276 void *src; 277 278 if (!s) 279 return -ENOMEM; 280 281 dst = (void *)((unsigned long)dst + (page << PAGE_SHIFT)); 282 src = pmap_mapdev_attr(VM_PAGE_TO_PHYS(s), PAGE_SIZE, prot); 283 if (!src) 284 return -ENOMEM; 285 286 memcpy(dst, src, PAGE_SIZE); 287 288 pmap_unmapdev(src, PAGE_SIZE); 289 290 return 0; 291 } 292 293 int ttm_bo_move_memcpy(struct ttm_buffer_object *bo, 294 bool evict, bool no_wait_gpu, 295 struct ttm_mem_reg *new_mem) 296 { 297 struct ttm_bo_device *bdev = bo->bdev; 298 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type]; 299 struct ttm_tt *ttm = bo->ttm; 300 struct ttm_mem_reg *old_mem = &bo->mem; 301 struct ttm_mem_reg old_copy = *old_mem; 302 void *old_iomap; 303 void *new_iomap; 304 int ret; 305 unsigned long i; 306 unsigned long page; 307 unsigned long add = 0; 308 int dir; 309 310 ret = ttm_mem_reg_ioremap(bdev, old_mem, &old_iomap); 311 if (ret) 312 return ret; 313 ret = ttm_mem_reg_ioremap(bdev, new_mem, &new_iomap); 314 if (ret) 315 goto out; 316 317 if (old_iomap == NULL && new_iomap == NULL) 318 goto out2; 319 if (old_iomap == NULL && ttm == NULL) 320 goto out2; 321 322 if (ttm->state == tt_unpopulated) { 323 ret = ttm->bdev->driver->ttm_tt_populate(ttm); 324 if (ret) { 325 /* if we fail here don't nuke the mm node 326 * as the bo still owns it */ 327 old_copy.mm_node = NULL; 328 goto out1; 329 } 330 } 331 332 add = 0; 333 dir = 1; 334 335 if ((old_mem->mem_type == new_mem->mem_type) && 336 (new_mem->start < old_mem->start + old_mem->size)) { 337 dir = -1; 338 add = new_mem->num_pages - 1; 339 } 340 341 for (i = 0; i < new_mem->num_pages; ++i) { 342 page = i * dir + add; 343 if (old_iomap == NULL) { 344 vm_memattr_t prot = ttm_io_prot(old_mem->placement); 345 ret = ttm_copy_ttm_io_page(ttm, new_iomap, page, 346 prot); 347 } else if (new_iomap == NULL) { 348 vm_memattr_t prot = ttm_io_prot(new_mem->placement); 349 ret = ttm_copy_io_ttm_page(ttm, old_iomap, page, 350 prot); 351 } else 352 ret = ttm_copy_io_page(new_iomap, old_iomap, page); 353 if (ret) { 354 /* failing here, means keep old copy as-is */ 355 old_copy.mm_node = NULL; 356 goto out1; 357 } 358 } 359 mb(); 360 out2: 361 old_copy = *old_mem; 362 *old_mem = *new_mem; 363 new_mem->mm_node = NULL; 364 365 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && (ttm != NULL)) { 366 ttm_tt_unbind(ttm); 367 ttm_tt_destroy(ttm); 368 bo->ttm = NULL; 369 } 370 371 out1: 372 ttm_mem_reg_iounmap(bdev, old_mem, new_iomap); 373 out: 374 ttm_mem_reg_iounmap(bdev, &old_copy, old_iomap); 375 ttm_bo_mem_put(bo, &old_copy); 376 return ret; 377 } 378 379 MALLOC_DEFINE(M_TTM_TRANSF_OBJ, "ttm_transf_obj", "TTM Transfer Objects"); 380 381 static void ttm_transfered_destroy(struct ttm_buffer_object *bo) 382 { 383 free(bo, M_TTM_TRANSF_OBJ); 384 } 385 386 /** 387 * ttm_buffer_object_transfer 388 * 389 * @bo: A pointer to a struct ttm_buffer_object. 390 * @new_obj: A pointer to a pointer to a newly created ttm_buffer_object, 391 * holding the data of @bo with the old placement. 392 * 393 * This is a utility function that may be called after an accelerated move 394 * has been scheduled. A new buffer object is created as a placeholder for 395 * the old data while it's being copied. When that buffer object is idle, 396 * it can be destroyed, releasing the space of the old placement. 397 * Returns: 398 * !0: Failure. 399 */ 400 401 static int 402 ttm_buffer_object_transfer(struct ttm_buffer_object *bo, 403 struct ttm_buffer_object **new_obj) 404 { 405 struct ttm_buffer_object *fbo; 406 struct ttm_bo_device *bdev = bo->bdev; 407 struct ttm_bo_driver *driver = bdev->driver; 408 409 fbo = malloc(sizeof(*fbo), M_TTM_TRANSF_OBJ, M_WAITOK); 410 *fbo = *bo; 411 412 /** 413 * Fix up members that we shouldn't copy directly: 414 * TODO: Explicit member copy would probably be better here. 415 */ 416 417 INIT_LIST_HEAD(&fbo->ddestroy); 418 INIT_LIST_HEAD(&fbo->lru); 419 INIT_LIST_HEAD(&fbo->swap); 420 INIT_LIST_HEAD(&fbo->io_reserve_lru); 421 fbo->vm_node = NULL; 422 atomic_set(&fbo->cpu_writers, 0); 423 424 mtx_lock(&bdev->fence_lock); 425 if (bo->sync_obj) 426 fbo->sync_obj = driver->sync_obj_ref(bo->sync_obj); 427 else 428 fbo->sync_obj = NULL; 429 mtx_unlock(&bdev->fence_lock); 430 refcount_init(&fbo->list_kref, 1); 431 refcount_init(&fbo->kref, 1); 432 fbo->destroy = &ttm_transfered_destroy; 433 fbo->acc_size = 0; 434 435 *new_obj = fbo; 436 return 0; 437 } 438 439 vm_memattr_t 440 ttm_io_prot(uint32_t caching_flags) 441 { 442 #if defined(__i386__) || defined(__amd64__) || defined(__powerpc__) || \ 443 defined(__arm__) 444 if (caching_flags & TTM_PL_FLAG_WC) 445 return (VM_MEMATTR_WRITE_COMBINING); 446 else 447 /* 448 * We do not support i386, look at the linux source 449 * for the reason of the comment. 450 */ 451 return (VM_MEMATTR_UNCACHEABLE); 452 #else 453 #error Port me 454 #endif 455 } 456 457 static int ttm_bo_ioremap(struct ttm_buffer_object *bo, 458 unsigned long offset, 459 unsigned long size, 460 struct ttm_bo_kmap_obj *map) 461 { 462 struct ttm_mem_reg *mem = &bo->mem; 463 464 if (bo->mem.bus.addr) { 465 map->bo_kmap_type = ttm_bo_map_premapped; 466 map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset); 467 } else { 468 map->bo_kmap_type = ttm_bo_map_iomap; 469 map->virtual = pmap_mapdev_attr(bo->mem.bus.base + 470 bo->mem.bus.offset + offset, size, 471 (mem->placement & TTM_PL_FLAG_WC) ? 472 VM_MEMATTR_WRITE_COMBINING : VM_MEMATTR_UNCACHEABLE); 473 map->size = size; 474 } 475 return (!map->virtual) ? -ENOMEM : 0; 476 } 477 478 static int ttm_bo_kmap_ttm(struct ttm_buffer_object *bo, 479 unsigned long start_page, 480 unsigned long num_pages, 481 struct ttm_bo_kmap_obj *map) 482 { 483 struct ttm_mem_reg *mem = &bo->mem; 484 vm_memattr_t prot; 485 struct ttm_tt *ttm = bo->ttm; 486 int i, ret; 487 488 MPASS(ttm != NULL); 489 490 if (ttm->state == tt_unpopulated) { 491 ret = ttm->bdev->driver->ttm_tt_populate(ttm); 492 if (ret) 493 return ret; 494 } 495 496 if (num_pages == 1 && (mem->placement & TTM_PL_FLAG_CACHED)) { 497 /* 498 * We're mapping a single page, and the desired 499 * page protection is consistent with the bo. 500 */ 501 502 map->bo_kmap_type = ttm_bo_map_kmap; 503 map->page = ttm->pages[start_page]; 504 map->sf = sf_buf_alloc(map->page, 0); 505 map->virtual = (void *)sf_buf_kva(map->sf); 506 } else { 507 /* 508 * We need to use vmap to get the desired page protection 509 * or to make the buffer object look contiguous. 510 */ 511 prot = (mem->placement & TTM_PL_FLAG_CACHED) ? 512 VM_MEMATTR_DEFAULT : ttm_io_prot(mem->placement); 513 map->bo_kmap_type = ttm_bo_map_vmap; 514 map->num_pages = num_pages; 515 map->virtual = (void *)kva_alloc(num_pages * PAGE_SIZE); 516 if (map->virtual != NULL) { 517 for (i = 0; i < num_pages; i++) { 518 /* XXXKIB hack */ 519 pmap_page_set_memattr(ttm->pages[start_page + 520 i], prot); 521 } 522 pmap_qenter((vm_offset_t)map->virtual, 523 &ttm->pages[start_page], num_pages); 524 } 525 } 526 return (!map->virtual) ? -ENOMEM : 0; 527 } 528 529 int ttm_bo_kmap(struct ttm_buffer_object *bo, 530 unsigned long start_page, unsigned long num_pages, 531 struct ttm_bo_kmap_obj *map) 532 { 533 struct ttm_mem_type_manager *man = 534 &bo->bdev->man[bo->mem.mem_type]; 535 unsigned long offset, size; 536 int ret; 537 538 MPASS(list_empty(&bo->swap)); 539 map->virtual = NULL; 540 map->bo = bo; 541 if (num_pages > bo->num_pages) 542 return -EINVAL; 543 if (start_page > bo->num_pages) 544 return -EINVAL; 545 #if 0 546 if (num_pages > 1 && !DRM_SUSER(DRM_CURPROC)) 547 return -EPERM; 548 #endif 549 (void) ttm_mem_io_lock(man, false); 550 ret = ttm_mem_io_reserve(bo->bdev, &bo->mem); 551 ttm_mem_io_unlock(man); 552 if (ret) 553 return ret; 554 if (!bo->mem.bus.is_iomem) { 555 return ttm_bo_kmap_ttm(bo, start_page, num_pages, map); 556 } else { 557 offset = start_page << PAGE_SHIFT; 558 size = num_pages << PAGE_SHIFT; 559 return ttm_bo_ioremap(bo, offset, size, map); 560 } 561 } 562 563 void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map) 564 { 565 struct ttm_buffer_object *bo = map->bo; 566 struct ttm_mem_type_manager *man = 567 &bo->bdev->man[bo->mem.mem_type]; 568 569 if (!map->virtual) 570 return; 571 switch (map->bo_kmap_type) { 572 case ttm_bo_map_iomap: 573 pmap_unmapdev(map->virtual, map->size); 574 break; 575 case ttm_bo_map_vmap: 576 pmap_qremove((vm_offset_t)(map->virtual), map->num_pages); 577 kva_free((vm_offset_t)map->virtual, 578 map->num_pages * PAGE_SIZE); 579 break; 580 case ttm_bo_map_kmap: 581 sf_buf_free(map->sf); 582 break; 583 case ttm_bo_map_premapped: 584 break; 585 default: 586 MPASS(0); 587 } 588 (void) ttm_mem_io_lock(man, false); 589 ttm_mem_io_free(map->bo->bdev, &map->bo->mem); 590 ttm_mem_io_unlock(man); 591 map->virtual = NULL; 592 map->page = NULL; 593 map->sf = NULL; 594 } 595 596 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo, 597 void *sync_obj, 598 bool evict, 599 bool no_wait_gpu, 600 struct ttm_mem_reg *new_mem) 601 { 602 struct ttm_bo_device *bdev = bo->bdev; 603 struct ttm_bo_driver *driver = bdev->driver; 604 struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type]; 605 struct ttm_mem_reg *old_mem = &bo->mem; 606 int ret; 607 struct ttm_buffer_object *ghost_obj; 608 void *tmp_obj = NULL; 609 610 mtx_lock(&bdev->fence_lock); 611 if (bo->sync_obj) { 612 tmp_obj = bo->sync_obj; 613 bo->sync_obj = NULL; 614 } 615 bo->sync_obj = driver->sync_obj_ref(sync_obj); 616 if (evict) { 617 ret = ttm_bo_wait(bo, false, false, false); 618 mtx_unlock(&bdev->fence_lock); 619 if (tmp_obj) 620 driver->sync_obj_unref(&tmp_obj); 621 if (ret) 622 return ret; 623 624 if ((man->flags & TTM_MEMTYPE_FLAG_FIXED) && 625 (bo->ttm != NULL)) { 626 ttm_tt_unbind(bo->ttm); 627 ttm_tt_destroy(bo->ttm); 628 bo->ttm = NULL; 629 } 630 ttm_bo_free_old_node(bo); 631 } else { 632 /** 633 * This should help pipeline ordinary buffer moves. 634 * 635 * Hang old buffer memory on a new buffer object, 636 * and leave it to be released when the GPU 637 * operation has completed. 638 */ 639 640 set_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags); 641 mtx_unlock(&bdev->fence_lock); 642 if (tmp_obj) 643 driver->sync_obj_unref(&tmp_obj); 644 645 ret = ttm_buffer_object_transfer(bo, &ghost_obj); 646 if (ret) 647 return ret; 648 649 /** 650 * If we're not moving to fixed memory, the TTM object 651 * needs to stay alive. Otherwhise hang it on the ghost 652 * bo to be unbound and destroyed. 653 */ 654 655 if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) 656 ghost_obj->ttm = NULL; 657 else 658 bo->ttm = NULL; 659 660 ttm_bo_unreserve(ghost_obj); 661 ttm_bo_unref(&ghost_obj); 662 } 663 664 *old_mem = *new_mem; 665 new_mem->mm_node = NULL; 666 667 return 0; 668 } 669