1 /* 2 * Copyright © 2012-2014 Intel Corporation 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 21 * IN THE SOFTWARE. 22 * 23 */ 24 25 #include <drm/drmP.h> 26 #include <drm/i915_drm.h> 27 #include "i915_drv.h" 28 #include "i915_trace.h" 29 #include "intel_drv.h" 30 31 struct i915_mm_struct { 32 struct mm_struct *mm; 33 struct drm_i915_private *i915; 34 struct i915_mmu_notifier *mn; 35 struct hlist_node node; 36 struct kref kref; 37 struct work_struct work; 38 }; 39 40 #if defined(CONFIG_MMU_NOTIFIER) 41 #include <linux/interval_tree.h> 42 43 struct i915_mmu_notifier { 44 spinlock_t lock; 45 struct hlist_node node; 46 struct mmu_notifier mn; 47 struct rb_root objects; 48 struct workqueue_struct *wq; 49 }; 50 51 struct i915_mmu_object { 52 struct i915_mmu_notifier *mn; 53 struct drm_i915_gem_object *obj; 54 struct interval_tree_node it; 55 struct list_head link; 56 struct work_struct work; 57 bool attached; 58 }; 59 60 static void wait_rendering(struct drm_i915_gem_object *obj) 61 { 62 struct drm_device *dev = obj->base.dev; 63 struct drm_i915_gem_request *requests[I915_NUM_ENGINES]; 64 int i, n; 65 66 if (!obj->active) 67 return; 68 69 n = 0; 70 for (i = 0; i < I915_NUM_ENGINES; i++) { 71 struct drm_i915_gem_request *req; 72 73 req = obj->last_read_req[i]; 74 if (req == NULL) 75 continue; 76 77 requests[n++] = i915_gem_request_reference(req); 78 } 79 80 mutex_unlock(&dev->struct_mutex); 81 82 for (i = 0; i < n; i++) 83 __i915_wait_request(requests[i], false, NULL, NULL); 84 85 mutex_lock(&dev->struct_mutex); 86 87 for (i = 0; i < n; i++) 88 i915_gem_request_unreference(requests[i]); 89 } 90 91 static void cancel_userptr(struct work_struct *work) 92 { 93 struct i915_mmu_object *mo = container_of(work, typeof(*mo), work); 94 struct drm_i915_gem_object *obj = mo->obj; 95 struct drm_device *dev = obj->base.dev; 96 97 mutex_lock(&dev->struct_mutex); 98 /* Cancel any active worker and force us to re-evaluate gup */ 99 obj->userptr.work = NULL; 100 101 if (obj->pages != NULL) { 102 struct drm_i915_private *dev_priv = to_i915(dev); 103 struct i915_vma *vma, *tmp; 104 bool was_interruptible; 105 106 wait_rendering(obj); 107 108 was_interruptible = dev_priv->mm.interruptible; 109 dev_priv->mm.interruptible = false; 110 111 list_for_each_entry_safe(vma, tmp, &obj->vma_list, obj_link) 112 WARN_ON(i915_vma_unbind(vma)); 113 WARN_ON(i915_gem_object_put_pages(obj)); 114 115 dev_priv->mm.interruptible = was_interruptible; 116 } 117 118 drm_gem_object_unreference(&obj->base); 119 mutex_unlock(&dev->struct_mutex); 120 } 121 122 static void add_object(struct i915_mmu_object *mo) 123 { 124 if (mo->attached) 125 return; 126 127 interval_tree_insert(&mo->it, &mo->mn->objects); 128 mo->attached = true; 129 } 130 131 static void del_object(struct i915_mmu_object *mo) 132 { 133 if (!mo->attached) 134 return; 135 136 interval_tree_remove(&mo->it, &mo->mn->objects); 137 mo->attached = false; 138 } 139 140 static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn, 141 struct mm_struct *mm, 142 unsigned long start, 143 unsigned long end) 144 { 145 struct i915_mmu_notifier *mn = 146 container_of(_mn, struct i915_mmu_notifier, mn); 147 struct i915_mmu_object *mo; 148 struct interval_tree_node *it; 149 LIST_HEAD(cancelled); 150 151 if (RB_EMPTY_ROOT(&mn->objects)) 152 return; 153 154 /* interval ranges are inclusive, but invalidate range is exclusive */ 155 end--; 156 157 spin_lock(&mn->lock); 158 it = interval_tree_iter_first(&mn->objects, start, end); 159 while (it) { 160 /* The mmu_object is released late when destroying the 161 * GEM object so it is entirely possible to gain a 162 * reference on an object in the process of being freed 163 * since our serialisation is via the spinlock and not 164 * the struct_mutex - and consequently use it after it 165 * is freed and then double free it. To prevent that 166 * use-after-free we only acquire a reference on the 167 * object if it is not in the process of being destroyed. 168 */ 169 mo = container_of(it, struct i915_mmu_object, it); 170 if (kref_get_unless_zero(&mo->obj->base.refcount)) 171 queue_work(mn->wq, &mo->work); 172 173 list_add(&mo->link, &cancelled); 174 it = interval_tree_iter_next(it, start, end); 175 } 176 list_for_each_entry(mo, &cancelled, link) 177 del_object(mo); 178 spin_unlock(&mn->lock); 179 180 flush_workqueue(mn->wq); 181 } 182 183 static const struct mmu_notifier_ops i915_gem_userptr_notifier = { 184 .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start, 185 }; 186 187 static struct i915_mmu_notifier * 188 i915_mmu_notifier_create(struct mm_struct *mm) 189 { 190 struct i915_mmu_notifier *mn; 191 int ret; 192 193 mn = kmalloc(sizeof(*mn), GFP_KERNEL); 194 if (mn == NULL) 195 return ERR_PTR(-ENOMEM); 196 197 spin_lock_init(&mn->lock); 198 mn->mn.ops = &i915_gem_userptr_notifier; 199 mn->objects = RB_ROOT; 200 mn->wq = alloc_workqueue("i915-userptr-release", WQ_UNBOUND, 0); 201 if (mn->wq == NULL) { 202 kfree(mn); 203 return ERR_PTR(-ENOMEM); 204 } 205 206 /* Protected by mmap_sem (write-lock) */ 207 ret = __mmu_notifier_register(&mn->mn, mm); 208 if (ret) { 209 destroy_workqueue(mn->wq); 210 kfree(mn); 211 return ERR_PTR(ret); 212 } 213 214 return mn; 215 } 216 217 static void 218 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj) 219 { 220 struct i915_mmu_object *mo; 221 222 mo = obj->userptr.mmu_object; 223 if (mo == NULL) 224 return; 225 226 spin_lock(&mo->mn->lock); 227 del_object(mo); 228 spin_unlock(&mo->mn->lock); 229 kfree(mo); 230 231 obj->userptr.mmu_object = NULL; 232 } 233 234 static struct i915_mmu_notifier * 235 i915_mmu_notifier_find(struct i915_mm_struct *mm) 236 { 237 struct i915_mmu_notifier *mn = mm->mn; 238 239 mn = mm->mn; 240 if (mn) 241 return mn; 242 243 down_write(&mm->mm->mmap_sem); 244 mutex_lock(&mm->i915->mm_lock); 245 if ((mn = mm->mn) == NULL) { 246 mn = i915_mmu_notifier_create(mm->mm); 247 if (!IS_ERR(mn)) 248 mm->mn = mn; 249 } 250 mutex_unlock(&mm->i915->mm_lock); 251 up_write(&mm->mm->mmap_sem); 252 253 return mn; 254 } 255 256 static int 257 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj, 258 unsigned flags) 259 { 260 struct i915_mmu_notifier *mn; 261 struct i915_mmu_object *mo; 262 263 if (flags & I915_USERPTR_UNSYNCHRONIZED) 264 return capable(CAP_SYS_ADMIN) ? 0 : -EPERM; 265 266 if (WARN_ON(obj->userptr.mm == NULL)) 267 return -EINVAL; 268 269 mn = i915_mmu_notifier_find(obj->userptr.mm); 270 if (IS_ERR(mn)) 271 return PTR_ERR(mn); 272 273 mo = kzalloc(sizeof(*mo), GFP_KERNEL); 274 if (mo == NULL) 275 return -ENOMEM; 276 277 mo->mn = mn; 278 mo->obj = obj; 279 mo->it.start = obj->userptr.ptr; 280 mo->it.last = obj->userptr.ptr + obj->base.size - 1; 281 INIT_WORK(&mo->work, cancel_userptr); 282 283 obj->userptr.mmu_object = mo; 284 return 0; 285 } 286 287 static void 288 i915_mmu_notifier_free(struct i915_mmu_notifier *mn, 289 struct mm_struct *mm) 290 { 291 if (mn == NULL) 292 return; 293 294 mmu_notifier_unregister(&mn->mn, mm); 295 destroy_workqueue(mn->wq); 296 kfree(mn); 297 } 298 299 #else 300 301 static void 302 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj) 303 { 304 } 305 306 static int 307 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj, 308 unsigned flags) 309 { 310 if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0) 311 return -ENODEV; 312 313 if (!capable(CAP_SYS_ADMIN)) 314 return -EPERM; 315 316 return 0; 317 } 318 319 static void 320 i915_mmu_notifier_free(struct i915_mmu_notifier *mn, 321 struct mm_struct *mm) 322 { 323 } 324 325 #endif 326 327 #if 0 328 static struct i915_mm_struct * 329 __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real) 330 { 331 struct i915_mm_struct *mm; 332 333 /* Protected by dev_priv->mm_lock */ 334 hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real) 335 if (mm->mm == real) 336 return mm; 337 338 return NULL; 339 } 340 #endif 341 342 static int 343 i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj) 344 { 345 struct drm_i915_private *dev_priv = to_i915(obj->base.dev); 346 #if 0 347 struct i915_mm_struct *mm; 348 #endif 349 int ret = 0; 350 351 /* During release of the GEM object we hold the struct_mutex. This 352 * precludes us from calling mmput() at that time as that may be 353 * the last reference and so call exit_mmap(). exit_mmap() will 354 * attempt to reap the vma, and if we were holding a GTT mmap 355 * would then call drm_gem_vm_close() and attempt to reacquire 356 * the struct mutex. So in order to avoid that recursion, we have 357 * to defer releasing the mm reference until after we drop the 358 * struct_mutex, i.e. we need to schedule a worker to do the clean 359 * up. 360 */ 361 mutex_lock(&dev_priv->mm_lock); 362 #if 0 363 mm = __i915_mm_struct_find(dev_priv, current->mm); 364 if (mm == NULL) { 365 mm = kmalloc(sizeof(*mm), GFP_KERNEL); 366 if (mm == NULL) { 367 #endif 368 ret = -ENOMEM; 369 #if 0 370 goto out; 371 } 372 373 kref_init(&mm->kref); 374 mm->i915 = to_i915(obj->base.dev); 375 376 mm->mm = current->mm; 377 atomic_inc(¤t->mm->mm_count); 378 379 mm->mn = NULL; 380 381 /* Protected by dev_priv->mm_lock */ 382 hash_add(dev_priv->mm_structs, 383 &mm->node, (unsigned long)mm->mm); 384 } else 385 kref_get(&mm->kref); 386 387 obj->userptr.mm = mm; 388 out: 389 #endif 390 mutex_unlock(&dev_priv->mm_lock); 391 return ret; 392 } 393 394 static void 395 __i915_mm_struct_free__worker(struct work_struct *work) 396 { 397 struct i915_mm_struct *mm = container_of(work, typeof(*mm), work); 398 i915_mmu_notifier_free(mm->mn, mm->mm); 399 #if 0 400 mmdrop(mm->mm); 401 #endif 402 kfree(mm); 403 } 404 405 static void 406 __i915_mm_struct_free(struct kref *kref) 407 { 408 struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref); 409 410 /* Protected by dev_priv->mm_lock */ 411 #if 0 412 hash_del(&mm->node); 413 #endif 414 mutex_unlock(&mm->i915->mm_lock); 415 416 INIT_WORK(&mm->work, __i915_mm_struct_free__worker); 417 schedule_work(&mm->work); 418 } 419 420 static void 421 i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj) 422 { 423 if (obj->userptr.mm == NULL) 424 return; 425 426 kref_put_mutex(&obj->userptr.mm->kref, 427 __i915_mm_struct_free, 428 &to_i915(obj->base.dev)->mm_lock); 429 obj->userptr.mm = NULL; 430 } 431 432 struct get_pages_work { 433 struct work_struct work; 434 struct drm_i915_gem_object *obj; 435 struct task_struct *task; 436 }; 437 438 #if IS_ENABLED(CONFIG_SWIOTLB) 439 #define swiotlb_active() swiotlb_nr_tbl() 440 #else 441 #define swiotlb_active() 0 442 #endif 443 444 #if 0 445 static int 446 st_set_pages(struct sg_table **st, struct page **pvec, int num_pages) 447 { 448 struct scatterlist *sg; 449 int ret, n; 450 451 *st = kmalloc(sizeof(**st), M_DRM, M_WAITOK); 452 if (*st == NULL) 453 return -ENOMEM; 454 455 if (swiotlb_active()) { 456 ret = sg_alloc_table(*st, num_pages, GFP_KERNEL); 457 if (ret) 458 goto err; 459 460 for_each_sg((*st)->sgl, sg, num_pages, n) 461 sg_set_page(sg, pvec[n], PAGE_SIZE, 0); 462 } else { 463 ret = sg_alloc_table_from_pages(*st, pvec, num_pages, 464 0, num_pages << PAGE_SHIFT, 465 GFP_KERNEL); 466 if (ret) 467 goto err; 468 } 469 470 return 0; 471 472 err: 473 kfree(*st); 474 *st = NULL; 475 return ret; 476 } 477 478 static int 479 __i915_gem_userptr_set_pages(struct drm_i915_gem_object *obj, 480 struct page **pvec, int num_pages) 481 { 482 int ret; 483 484 ret = st_set_pages(&obj->pages, pvec, num_pages); 485 if (ret) 486 return ret; 487 488 ret = i915_gem_gtt_prepare_object(obj); 489 if (ret) { 490 sg_free_table(obj->pages); 491 kfree(obj->pages); 492 obj->pages = NULL; 493 } 494 495 return ret; 496 } 497 #endif 498 499 static int 500 __i915_gem_userptr_set_active(struct drm_i915_gem_object *obj, 501 bool value) 502 { 503 int ret = 0; 504 505 /* During mm_invalidate_range we need to cancel any userptr that 506 * overlaps the range being invalidated. Doing so requires the 507 * struct_mutex, and that risks recursion. In order to cause 508 * recursion, the user must alias the userptr address space with 509 * a GTT mmapping (possible with a MAP_FIXED) - then when we have 510 * to invalidate that mmaping, mm_invalidate_range is called with 511 * the userptr address *and* the struct_mutex held. To prevent that 512 * we set a flag under the i915_mmu_notifier spinlock to indicate 513 * whether this object is valid. 514 */ 515 #if defined(CONFIG_MMU_NOTIFIER) 516 if (obj->userptr.mmu_object == NULL) 517 return 0; 518 519 spin_lock(&obj->userptr.mmu_object->mn->lock); 520 /* In order to serialise get_pages with an outstanding 521 * cancel_userptr, we must drop the struct_mutex and try again. 522 */ 523 if (!value) 524 del_object(obj->userptr.mmu_object); 525 else if (!work_pending(&obj->userptr.mmu_object->work)) 526 add_object(obj->userptr.mmu_object); 527 else 528 ret = -EAGAIN; 529 spin_unlock(&obj->userptr.mmu_object->mn->lock); 530 #endif 531 532 return ret; 533 } 534 535 #if 0 536 static void 537 __i915_gem_userptr_get_pages_worker(struct work_struct *_work) 538 { 539 struct get_pages_work *work = container_of(_work, typeof(*work), work); 540 struct drm_i915_gem_object *obj = work->obj; 541 struct drm_device *dev = obj->base.dev; 542 const int npages = obj->base.size >> PAGE_SHIFT; 543 struct page **pvec; 544 int pinned, ret; 545 546 ret = -ENOMEM; 547 pinned = 0; 548 549 pvec = drm_malloc_gfp(npages, sizeof(struct page *), GFP_TEMPORARY); 550 if (pvec != NULL) { 551 struct mm_struct *mm = obj->userptr.mm->mm; 552 553 ret = -EFAULT; 554 if (atomic_inc_not_zero(&mm->mm_users)) { 555 down_read(&mm->mmap_sem); 556 while (pinned < npages) { 557 ret = get_user_pages_remote 558 (work->task, mm, 559 obj->userptr.ptr + pinned * PAGE_SIZE, 560 npages - pinned, 561 !obj->userptr.read_only, 0, 562 pvec + pinned, NULL); 563 if (ret < 0) 564 break; 565 566 pinned += ret; 567 } 568 up_read(&mm->mmap_sem); 569 mmput(mm); 570 } 571 } 572 573 mutex_lock(&dev->struct_mutex); 574 if (obj->userptr.work == &work->work) { 575 if (pinned == npages) { 576 ret = __i915_gem_userptr_set_pages(obj, pvec, npages); 577 if (ret == 0) { 578 list_add_tail(&obj->global_list, 579 &to_i915(dev)->mm.unbound_list); 580 obj->get_page.sg = obj->pages->sgl; 581 obj->get_page.last = 0; 582 pinned = 0; 583 } 584 } 585 obj->userptr.work = ERR_PTR(ret); 586 if (ret) 587 __i915_gem_userptr_set_active(obj, false); 588 } 589 590 obj->userptr.workers--; 591 drm_gem_object_unreference(&obj->base); 592 mutex_unlock(&dev->struct_mutex); 593 594 release_pages(pvec, pinned, 0); 595 drm_free_large(pvec); 596 597 put_task_struct(work->task); 598 kfree(work); 599 } 600 601 static int 602 __i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object *obj, 603 bool *active) 604 { 605 struct get_pages_work *work; 606 607 /* Spawn a worker so that we can acquire the 608 * user pages without holding our mutex. Access 609 * to the user pages requires mmap_sem, and we have 610 * a strict lock ordering of mmap_sem, struct_mutex - 611 * we already hold struct_mutex here and so cannot 612 * call gup without encountering a lock inversion. 613 * 614 * Userspace will keep on repeating the operation 615 * (thanks to EAGAIN) until either we hit the fast 616 * path or the worker completes. If the worker is 617 * cancelled or superseded, the task is still run 618 * but the results ignored. (This leads to 619 * complications that we may have a stray object 620 * refcount that we need to be wary of when 621 * checking for existing objects during creation.) 622 * If the worker encounters an error, it reports 623 * that error back to this function through 624 * obj->userptr.work = ERR_PTR. 625 */ 626 if (obj->userptr.workers >= I915_GEM_USERPTR_MAX_WORKERS) 627 return -EAGAIN; 628 629 work = kmalloc(sizeof(*work), GFP_KERNEL); 630 if (work == NULL) 631 return -ENOMEM; 632 633 obj->userptr.work = &work->work; 634 obj->userptr.workers++; 635 636 work->obj = obj; 637 drm_gem_object_reference(&obj->base); 638 639 work->task = current; 640 get_task_struct(work->task); 641 642 INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker); 643 schedule_work(&work->work); 644 645 *active = true; 646 return -EAGAIN; 647 } 648 #endif 649 650 static int 651 i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj) 652 { 653 #if 0 654 const int num_pages = obj->base.size >> PAGE_SHIFT; 655 struct page **pvec; 656 int pinned, ret; 657 bool active; 658 659 /* If userspace should engineer that these pages are replaced in 660 * the vma between us binding this page into the GTT and completion 661 * of rendering... Their loss. If they change the mapping of their 662 * pages they need to create a new bo to point to the new vma. 663 * 664 * However, that still leaves open the possibility of the vma 665 * being copied upon fork. Which falls under the same userspace 666 * synchronisation issue as a regular bo, except that this time 667 * the process may not be expecting that a particular piece of 668 * memory is tied to the GPU. 669 * 670 * Fortunately, we can hook into the mmu_notifier in order to 671 * discard the page references prior to anything nasty happening 672 * to the vma (discard or cloning) which should prevent the more 673 * egregious cases from causing harm. 674 */ 675 if (IS_ERR(obj->userptr.work)) { 676 /* active flag will have been dropped already by the worker */ 677 ret = PTR_ERR(obj->userptr.work); 678 obj->userptr.work = NULL; 679 return ret; 680 } 681 if (obj->userptr.work) 682 /* active flag should still be held for the pending work */ 683 return -EAGAIN; 684 685 /* Let the mmu-notifier know that we have begun and need cancellation */ 686 ret = __i915_gem_userptr_set_active(obj, true); 687 if (ret) 688 return ret; 689 690 pvec = NULL; 691 pinned = 0; 692 if (obj->userptr.mm->mm == current->mm) { 693 pvec = kmalloc(num_pages*sizeof(struct page *), 694 GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY); 695 if (pvec == NULL) { 696 pvec = drm_malloc_ab(num_pages, sizeof(struct page *)); 697 if (pvec == NULL) { 698 __i915_gem_userptr_set_active(obj, false); 699 return -ENOMEM; 700 } 701 } 702 703 pinned = __get_user_pages_fast(obj->userptr.ptr, num_pages, 704 !obj->userptr.read_only, pvec); 705 } 706 707 active = false; 708 if (pinned < 0) 709 ret = pinned, pinned = 0; 710 else if (pinned < num_pages) 711 ret = __i915_gem_userptr_get_pages_schedule(obj, &active); 712 else 713 ret = __i915_gem_userptr_set_pages(obj, pvec, num_pages); 714 if (ret) { 715 __i915_gem_userptr_set_active(obj, active); 716 release_pages(pvec, pinned, 0); 717 } 718 drm_free_large(pvec); 719 return ret; 720 #else 721 return 0; 722 #endif /* 0 */ 723 } 724 725 static void 726 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj) 727 { 728 struct sg_page_iter sg_iter; 729 730 BUG_ON(obj->userptr.work != NULL); 731 __i915_gem_userptr_set_active(obj, false); 732 733 if (obj->madv != I915_MADV_WILLNEED) 734 obj->dirty = 0; 735 736 i915_gem_gtt_finish_object(obj); 737 738 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) { 739 struct page *page = sg_page_iter_page(&sg_iter); 740 741 if (obj->dirty) 742 set_page_dirty(page); 743 744 mark_page_accessed(page); 745 #if 0 746 page_cache_release(page); 747 #endif 748 } 749 obj->dirty = 0; 750 751 sg_free_table(obj->pages); 752 kfree(obj->pages); 753 } 754 755 static void 756 i915_gem_userptr_release(struct drm_i915_gem_object *obj) 757 { 758 i915_gem_userptr_release__mmu_notifier(obj); 759 i915_gem_userptr_release__mm_struct(obj); 760 } 761 762 static int 763 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj) 764 { 765 if (obj->userptr.mmu_object) 766 return 0; 767 768 return i915_gem_userptr_init__mmu_notifier(obj, 0); 769 } 770 771 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = { 772 .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE, 773 .get_pages = i915_gem_userptr_get_pages, 774 .put_pages = i915_gem_userptr_put_pages, 775 .dmabuf_export = i915_gem_userptr_dmabuf_export, 776 .release = i915_gem_userptr_release, 777 }; 778 779 /** 780 * Creates a new mm object that wraps some normal memory from the process 781 * context - user memory. 782 * 783 * We impose several restrictions upon the memory being mapped 784 * into the GPU. 785 * 1. It must be page aligned (both start/end addresses, i.e ptr and size). 786 * 2. It must be normal system memory, not a pointer into another map of IO 787 * space (e.g. it must not be a GTT mmapping of another object). 788 * 3. We only allow a bo as large as we could in theory map into the GTT, 789 * that is we limit the size to the total size of the GTT. 790 * 4. The bo is marked as being snoopable. The backing pages are left 791 * accessible directly by the CPU, but reads and writes by the GPU may 792 * incur the cost of a snoop (unless you have an LLC architecture). 793 * 794 * Synchronisation between multiple users and the GPU is left to userspace 795 * through the normal set-domain-ioctl. The kernel will enforce that the 796 * GPU relinquishes the VMA before it is returned back to the system 797 * i.e. upon free(), munmap() or process termination. However, the userspace 798 * malloc() library may not immediately relinquish the VMA after free() and 799 * instead reuse it whilst the GPU is still reading and writing to the VMA. 800 * Caveat emptor. 801 * 802 * Also note, that the object created here is not currently a "first class" 803 * object, in that several ioctls are banned. These are the CPU access 804 * ioctls: mmap(), pwrite and pread. In practice, you are expected to use 805 * direct access via your pointer rather than use those ioctls. Another 806 * restriction is that we do not allow userptr surfaces to be pinned to the 807 * hardware and so we reject any attempt to create a framebuffer out of a 808 * userptr. 809 * 810 * If you think this is a good interface to use to pass GPU memory between 811 * drivers, please use dma-buf instead. In fact, wherever possible use 812 * dma-buf instead. 813 */ 814 int 815 i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file) 816 { 817 struct drm_i915_gem_userptr *args = data; 818 struct drm_i915_gem_object *obj; 819 int ret; 820 u32 handle; 821 822 if (!HAS_LLC(dev) && !HAS_SNOOP(dev)) { 823 /* We cannot support coherent userptr objects on hw without 824 * LLC and broken snooping. 825 */ 826 return -ENODEV; 827 } 828 829 if (args->flags & ~(I915_USERPTR_READ_ONLY | 830 I915_USERPTR_UNSYNCHRONIZED)) 831 return -EINVAL; 832 833 if (offset_in_page(args->user_ptr | args->user_size)) 834 return -EINVAL; 835 836 #if 0 837 if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE, 838 (char __user *)(unsigned long)args->user_ptr, args->user_size)) 839 return -EFAULT; 840 #endif 841 842 if (args->flags & I915_USERPTR_READ_ONLY) { 843 /* On almost all of the current hw, we cannot tell the GPU that a 844 * page is readonly, so this is just a placeholder in the uAPI. 845 */ 846 return -ENODEV; 847 } 848 849 obj = i915_gem_object_alloc(dev); 850 if (obj == NULL) 851 return -ENOMEM; 852 853 drm_gem_private_object_init(dev, &obj->base, args->user_size); 854 i915_gem_object_init(obj, &i915_gem_userptr_ops); 855 obj->cache_level = I915_CACHE_LLC; 856 obj->base.write_domain = I915_GEM_DOMAIN_CPU; 857 obj->base.read_domains = I915_GEM_DOMAIN_CPU; 858 859 obj->userptr.ptr = args->user_ptr; 860 obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY); 861 862 /* And keep a pointer to the current->mm for resolving the user pages 863 * at binding. This means that we need to hook into the mmu_notifier 864 * in order to detect if the mmu is destroyed. 865 */ 866 ret = i915_gem_userptr_init__mm_struct(obj); 867 if (ret == 0) 868 ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags); 869 if (ret == 0) 870 ret = drm_gem_handle_create(file, &obj->base, &handle); 871 872 /* drop reference from allocate - handle holds it now */ 873 drm_gem_object_unreference_unlocked(&obj->base); 874 if (ret) 875 return ret; 876 877 args->handle = handle; 878 return 0; 879 } 880 881 int 882 i915_gem_init_userptr(struct drm_device *dev) 883 { 884 struct drm_i915_private *dev_priv = to_i915(dev); 885 lockinit(&dev_priv->mm_lock, "i915dmm", 0, LK_CANRECURSE); 886 #if 0 887 hash_init(dev_priv->mm_structs); 888 #endif 889 return 0; 890 } 891