xref: /dragonfly/sys/dev/drm/i915/i915_gem_userptr.c (revision 6a3cbbc2)
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 #include <linux/mmu_context.h>
31 #include <linux/mmu_notifier.h>
32 #include <linux/swap.h>
33 
34 struct i915_mm_struct {
35 	struct mm_struct *mm;
36 	struct drm_i915_private *i915;
37 	struct i915_mmu_notifier *mn;
38 	struct hlist_node node;
39 	struct kref kref;
40 	struct work_struct work;
41 };
42 
43 #if defined(CONFIG_MMU_NOTIFIER)
44 #include <linux/interval_tree.h>
45 
46 struct i915_mmu_notifier {
47 	spinlock_t lock;
48 	struct hlist_node node;
49 	struct mmu_notifier mn;
50 	struct rb_root objects;
51 	struct workqueue_struct *wq;
52 };
53 
54 struct i915_mmu_object {
55 	struct i915_mmu_notifier *mn;
56 	struct drm_i915_gem_object *obj;
57 	struct interval_tree_node it;
58 	struct list_head link;
59 	struct work_struct work;
60 	bool attached;
61 };
62 
63 static void wait_rendering(struct drm_i915_gem_object *obj)
64 {
65 	unsigned long active = __I915_BO_ACTIVE(obj);
66 	int idx;
67 
68 	for_each_active(active, idx)
69 		i915_gem_active_wait_unlocked(&obj->last_read[idx],
70 					      0, NULL, NULL);
71 }
72 
73 static void cancel_userptr(struct work_struct *work)
74 {
75 	struct i915_mmu_object *mo = container_of(work, typeof(*mo), work);
76 	struct drm_i915_gem_object *obj = mo->obj;
77 	struct drm_device *dev = obj->base.dev;
78 
79 	wait_rendering(obj);
80 
81 	mutex_lock(&dev->struct_mutex);
82 	/* Cancel any active worker and force us to re-evaluate gup */
83 	obj->userptr.work = NULL;
84 
85 	if (obj->pages != NULL) {
86 		/* We are inside a kthread context and can't be interrupted */
87 		WARN_ON(i915_gem_object_unbind(obj));
88 		WARN_ON(i915_gem_object_put_pages(obj));
89 	}
90 
91 	i915_gem_object_put(obj);
92 	mutex_unlock(&dev->struct_mutex);
93 }
94 
95 static void add_object(struct i915_mmu_object *mo)
96 {
97 	if (mo->attached)
98 		return;
99 
100 	interval_tree_insert(&mo->it, &mo->mn->objects);
101 	mo->attached = true;
102 }
103 
104 static void del_object(struct i915_mmu_object *mo)
105 {
106 	if (!mo->attached)
107 		return;
108 
109 	interval_tree_remove(&mo->it, &mo->mn->objects);
110 	mo->attached = false;
111 }
112 
113 static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn,
114 						       struct mm_struct *mm,
115 						       unsigned long start,
116 						       unsigned long end)
117 {
118 	struct i915_mmu_notifier *mn =
119 		container_of(_mn, struct i915_mmu_notifier, mn);
120 	struct i915_mmu_object *mo;
121 	struct interval_tree_node *it;
122 	LIST_HEAD(cancelled);
123 
124 	if (RB_EMPTY_ROOT(&mn->objects))
125 		return;
126 
127 	/* interval ranges are inclusive, but invalidate range is exclusive */
128 	end--;
129 
130 	lockmgr(&mn->lock, LK_EXCLUSIVE);
131 	it = interval_tree_iter_first(&mn->objects, start, end);
132 	while (it) {
133 		/* The mmu_object is released late when destroying the
134 		 * GEM object so it is entirely possible to gain a
135 		 * reference on an object in the process of being freed
136 		 * since our serialisation is via the spinlock and not
137 		 * the struct_mutex - and consequently use it after it
138 		 * is freed and then double free it. To prevent that
139 		 * use-after-free we only acquire a reference on the
140 		 * object if it is not in the process of being destroyed.
141 		 */
142 		mo = container_of(it, struct i915_mmu_object, it);
143 		if (kref_get_unless_zero(&mo->obj->base.refcount))
144 			queue_work(mn->wq, &mo->work);
145 
146 		list_add(&mo->link, &cancelled);
147 		it = interval_tree_iter_next(it, start, end);
148 	}
149 	list_for_each_entry(mo, &cancelled, link)
150 		del_object(mo);
151 	lockmgr(&mn->lock, LK_RELEASE);
152 
153 	flush_workqueue(mn->wq);
154 }
155 
156 static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
157 	.invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start,
158 };
159 
160 static struct i915_mmu_notifier *
161 i915_mmu_notifier_create(struct mm_struct *mm)
162 {
163 	struct i915_mmu_notifier *mn;
164 	int ret;
165 
166 	mn = kmalloc(sizeof(*mn), M_DRM, GFP_KERNEL);
167 	if (mn == NULL)
168 		return ERR_PTR(-ENOMEM);
169 
170 	spin_lock_init(&mn->lock);
171 	mn->mn.ops = &i915_gem_userptr_notifier;
172 	mn->objects = RB_ROOT;
173 	mn->wq = alloc_workqueue("i915-userptr-release", WQ_UNBOUND, 0);
174 	if (mn->wq == NULL) {
175 		kfree(mn);
176 		return ERR_PTR(-ENOMEM);
177 	}
178 
179 	 /* Protected by mmap_sem (write-lock) */
180 	ret = __mmu_notifier_register(&mn->mn, mm);
181 	if (ret) {
182 		destroy_workqueue(mn->wq);
183 		kfree(mn);
184 		return ERR_PTR(ret);
185 	}
186 
187 	return mn;
188 }
189 
190 static void
191 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
192 {
193 	struct i915_mmu_object *mo;
194 
195 	mo = obj->userptr.mmu_object;
196 	if (mo == NULL)
197 		return;
198 
199 	lockmgr(&mo->mn->lock, LK_EXCLUSIVE);
200 	del_object(mo);
201 	lockmgr(&mo->mn->lock, LK_RELEASE);
202 	kfree(mo);
203 
204 	obj->userptr.mmu_object = NULL;
205 }
206 
207 static struct i915_mmu_notifier *
208 i915_mmu_notifier_find(struct i915_mm_struct *mm)
209 {
210 	struct i915_mmu_notifier *mn = mm->mn;
211 
212 	mn = mm->mn;
213 	if (mn)
214 		return mn;
215 
216 	down_write(&mm->mm->mmap_sem);
217 	mutex_lock(&mm->i915->mm_lock);
218 	if ((mn = mm->mn) == NULL) {
219 		mn = i915_mmu_notifier_create(mm->mm);
220 		if (!IS_ERR(mn))
221 			mm->mn = mn;
222 	}
223 	mutex_unlock(&mm->i915->mm_lock);
224 	up_write(&mm->mm->mmap_sem);
225 
226 	return mn;
227 }
228 
229 static int
230 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
231 				    unsigned flags)
232 {
233 	struct i915_mmu_notifier *mn;
234 	struct i915_mmu_object *mo;
235 
236 	if (flags & I915_USERPTR_UNSYNCHRONIZED)
237 		return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
238 
239 	if (WARN_ON(obj->userptr.mm == NULL))
240 		return -EINVAL;
241 
242 	mn = i915_mmu_notifier_find(obj->userptr.mm);
243 	if (IS_ERR(mn))
244 		return PTR_ERR(mn);
245 
246 	mo = kzalloc(sizeof(*mo), GFP_KERNEL);
247 	if (mo == NULL)
248 		return -ENOMEM;
249 
250 	mo->mn = mn;
251 	mo->obj = obj;
252 	mo->it.start = obj->userptr.ptr;
253 	mo->it.last = obj->userptr.ptr + obj->base.size - 1;
254 	INIT_WORK(&mo->work, cancel_userptr);
255 
256 	obj->userptr.mmu_object = mo;
257 	return 0;
258 }
259 
260 static void
261 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
262 		       struct mm_struct *mm)
263 {
264 	if (mn == NULL)
265 		return;
266 
267 	mmu_notifier_unregister(&mn->mn, mm);
268 	destroy_workqueue(mn->wq);
269 	kfree(mn);
270 }
271 
272 #else
273 
274 static void
275 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
276 {
277 }
278 
279 static int
280 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
281 				    unsigned flags)
282 {
283 	if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
284 		return -ENODEV;
285 
286 	if (!capable(CAP_SYS_ADMIN))
287 		return -EPERM;
288 
289 	return 0;
290 }
291 
292 static void
293 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
294 		       struct mm_struct *mm)
295 {
296 }
297 
298 #endif
299 
300 static struct i915_mm_struct *
301 __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real)
302 {
303 	struct i915_mm_struct *mm;
304 
305 	/* Protected by dev_priv->mm_lock */
306 	hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real)
307 		if (mm->mm == real)
308 			return mm;
309 
310 	return NULL;
311 }
312 
313 static int
314 i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj)
315 {
316 	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
317 	struct i915_mm_struct *mm;
318 	int ret = 0;
319 
320 	/* During release of the GEM object we hold the struct_mutex. This
321 	 * precludes us from calling mmput() at that time as that may be
322 	 * the last reference and so call exit_mmap(). exit_mmap() will
323 	 * attempt to reap the vma, and if we were holding a GTT mmap
324 	 * would then call drm_gem_vm_close() and attempt to reacquire
325 	 * the struct mutex. So in order to avoid that recursion, we have
326 	 * to defer releasing the mm reference until after we drop the
327 	 * struct_mutex, i.e. we need to schedule a worker to do the clean
328 	 * up.
329 	 */
330 	mutex_lock(&dev_priv->mm_lock);
331 	mm = __i915_mm_struct_find(dev_priv, current->mm);
332 #if 0
333 	if (mm == NULL) {
334 		mm = kmalloc(sizeof(*mm), M_DRM, GFP_KERNEL);
335 		if (mm == NULL) {
336 #endif
337 			ret = -ENOMEM;
338 #if 0
339 			goto out;
340 		}
341 
342 		kref_init(&mm->kref);
343 		mm->i915 = to_i915(obj->base.dev);
344 
345 		mm->mm = current->mm;
346 		atomic_inc(&current->mm->mm_count);
347 
348 		mm->mn = NULL;
349 
350 		/* Protected by dev_priv->mm_lock */
351 		hash_add(dev_priv->mm_structs,
352 			 &mm->node, (unsigned long)mm->mm);
353 	} else
354 		kref_get(&mm->kref);
355 
356 	obj->userptr.mm = mm;
357 out:
358 	mutex_unlock(&dev_priv->mm_lock);
359 #endif
360 	return ret;
361 }
362 
363 static void
364 __i915_mm_struct_free__worker(struct work_struct *work)
365 {
366 	struct i915_mm_struct *mm = container_of(work, typeof(*mm), work);
367 	i915_mmu_notifier_free(mm->mn, mm->mm);
368 #if 0
369 	mmdrop(mm->mm);
370 #endif
371 	kfree(mm);
372 }
373 
374 static void
375 __i915_mm_struct_free(struct kref *kref)
376 {
377 	struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref);
378 
379 	/* Protected by dev_priv->mm_lock */
380 	hash_del(&mm->node);
381 	mutex_unlock(&mm->i915->mm_lock);
382 
383 	INIT_WORK(&mm->work, __i915_mm_struct_free__worker);
384 	schedule_work(&mm->work);
385 }
386 
387 static void
388 i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj)
389 {
390 	if (obj->userptr.mm == NULL)
391 		return;
392 
393 	kref_put_mutex(&obj->userptr.mm->kref,
394 		       __i915_mm_struct_free,
395 		       &to_i915(obj->base.dev)->mm_lock);
396 	obj->userptr.mm = NULL;
397 }
398 
399 struct get_pages_work {
400 	struct work_struct work;
401 	struct drm_i915_gem_object *obj;
402 	struct task_struct *task;
403 };
404 
405 #if IS_ENABLED(CONFIG_SWIOTLB)
406 #define swiotlb_active() swiotlb_nr_tbl()
407 #else
408 #define swiotlb_active() 0
409 #endif
410 
411 #if 0
412 static int
413 st_set_pages(struct sg_table **st, struct page **pvec, int num_pages)
414 {
415 	struct scatterlist *sg;
416 	int ret, n;
417 
418 	*st = kmalloc(sizeof(**st), M_DRM, GFP_KERNEL);
419 	if (*st == NULL)
420 		return -ENOMEM;
421 
422 	if (swiotlb_active()) {
423 		ret = sg_alloc_table(*st, num_pages, GFP_KERNEL);
424 		if (ret)
425 			goto err;
426 
427 		for_each_sg((*st)->sgl, sg, num_pages, n)
428 			sg_set_page(sg, pvec[n], PAGE_SIZE, 0);
429 	} else {
430 		ret = sg_alloc_table_from_pages(*st, pvec, num_pages,
431 						0, num_pages << PAGE_SHIFT,
432 						GFP_KERNEL);
433 		if (ret)
434 			goto err;
435 	}
436 
437 	return 0;
438 
439 err:
440 	kfree(*st);
441 	*st = NULL;
442 	return ret;
443 }
444 
445 static int
446 __i915_gem_userptr_set_pages(struct drm_i915_gem_object *obj,
447 			     struct page **pvec, int num_pages)
448 {
449 	int ret;
450 
451 	ret = st_set_pages(&obj->pages, pvec, num_pages);
452 	if (ret)
453 		return ret;
454 
455 	ret = i915_gem_gtt_prepare_object(obj);
456 	if (ret) {
457 		sg_free_table(obj->pages);
458 		kfree(obj->pages);
459 		obj->pages = NULL;
460 	}
461 
462 	return ret;
463 }
464 #endif
465 
466 static int
467 __i915_gem_userptr_set_active(struct drm_i915_gem_object *obj,
468 			      bool value)
469 {
470 	int ret = 0;
471 
472 	/* During mm_invalidate_range we need to cancel any userptr that
473 	 * overlaps the range being invalidated. Doing so requires the
474 	 * struct_mutex, and that risks recursion. In order to cause
475 	 * recursion, the user must alias the userptr address space with
476 	 * a GTT mmapping (possible with a MAP_FIXED) - then when we have
477 	 * to invalidate that mmaping, mm_invalidate_range is called with
478 	 * the userptr address *and* the struct_mutex held.  To prevent that
479 	 * we set a flag under the i915_mmu_notifier spinlock to indicate
480 	 * whether this object is valid.
481 	 */
482 #if defined(CONFIG_MMU_NOTIFIER)
483 	if (obj->userptr.mmu_object == NULL)
484 		return 0;
485 
486 	lockmgr(&obj->userptr.mmu_object->mn->lock, LK_EXCLUSIVE);
487 	/* In order to serialise get_pages with an outstanding
488 	 * cancel_userptr, we must drop the struct_mutex and try again.
489 	 */
490 	if (!value)
491 		del_object(obj->userptr.mmu_object);
492 	else if (!work_pending(&obj->userptr.mmu_object->work))
493 		add_object(obj->userptr.mmu_object);
494 	else
495 		ret = -EAGAIN;
496 	lockmgr(&obj->userptr.mmu_object->mn->lock, LK_RELEASE);
497 #endif
498 
499 	return ret;
500 }
501 
502 #if 0
503 static void
504 __i915_gem_userptr_get_pages_worker(struct work_struct *_work)
505 {
506 	struct get_pages_work *work = container_of(_work, typeof(*work), work);
507 	struct drm_i915_gem_object *obj = work->obj;
508 	struct drm_device *dev = obj->base.dev;
509 	const int npages = obj->base.size >> PAGE_SHIFT;
510 	struct page **pvec;
511 	int pinned, ret;
512 
513 	ret = -ENOMEM;
514 	pinned = 0;
515 
516 	pvec = drm_malloc_gfp(npages, sizeof(struct page *), GFP_TEMPORARY);
517 	if (pvec != NULL) {
518 		struct mm_struct *mm = obj->userptr.mm->mm;
519 
520 		ret = -EFAULT;
521 		if (atomic_inc_not_zero(&mm->mm_users)) {
522 			down_read(&mm->mmap_sem);
523 			while (pinned < npages) {
524 				ret = get_user_pages_remote
525 					(work->task, mm,
526 					 obj->userptr.ptr + pinned * PAGE_SIZE,
527 					 npages - pinned,
528 					 !obj->userptr.read_only, 0,
529 					 pvec + pinned, NULL);
530 				if (ret < 0)
531 					break;
532 
533 				pinned += ret;
534 			}
535 			up_read(&mm->mmap_sem);
536 			mmput(mm);
537 		}
538 	}
539 
540 	mutex_lock(&dev->struct_mutex);
541 	if (obj->userptr.work == &work->work) {
542 		if (pinned == npages) {
543 			ret = __i915_gem_userptr_set_pages(obj, pvec, npages);
544 			if (ret == 0) {
545 				list_add_tail(&obj->global_list,
546 					      &to_i915(dev)->mm.unbound_list);
547 				obj->get_page.sg = obj->pages->sgl;
548 				obj->get_page.last = 0;
549 				pinned = 0;
550 			}
551 		}
552 		obj->userptr.work = ERR_PTR(ret);
553 	}
554 
555 	obj->userptr.workers--;
556 	i915_gem_object_put(obj);
557 	mutex_unlock(&dev->struct_mutex);
558 
559 	release_pages(pvec, pinned, 0);
560 	drm_free_large(pvec);
561 
562 	put_task_struct(work->task);
563 	kfree(work);
564 }
565 
566 static int
567 __i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object *obj,
568 				      bool *active)
569 {
570 	struct get_pages_work *work;
571 
572 	/* Spawn a worker so that we can acquire the
573 	 * user pages without holding our mutex. Access
574 	 * to the user pages requires mmap_sem, and we have
575 	 * a strict lock ordering of mmap_sem, struct_mutex -
576 	 * we already hold struct_mutex here and so cannot
577 	 * call gup without encountering a lock inversion.
578 	 *
579 	 * Userspace will keep on repeating the operation
580 	 * (thanks to EAGAIN) until either we hit the fast
581 	 * path or the worker completes. If the worker is
582 	 * cancelled or superseded, the task is still run
583 	 * but the results ignored. (This leads to
584 	 * complications that we may have a stray object
585 	 * refcount that we need to be wary of when
586 	 * checking for existing objects during creation.)
587 	 * If the worker encounters an error, it reports
588 	 * that error back to this function through
589 	 * obj->userptr.work = ERR_PTR.
590 	 */
591 	if (obj->userptr.workers >= I915_GEM_USERPTR_MAX_WORKERS)
592 		return -EAGAIN;
593 
594 	work = kmalloc(sizeof(*work), GFP_KERNEL);
595 	if (work == NULL)
596 		return -ENOMEM;
597 
598 	obj->userptr.work = &work->work;
599 	obj->userptr.workers++;
600 
601 	work->obj = i915_gem_object_get(obj);
602 
603 	work->task = current;
604 	get_task_struct(work->task);
605 
606 	INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
607 	schedule_work(&work->work);
608 
609 	*active = true;
610 	return -EAGAIN;
611 }
612 #endif
613 
614 static int
615 i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
616 {
617 #if 0
618 	const int num_pages = obj->base.size >> PAGE_SHIFT;
619 	struct page **pvec;
620 	int pinned, ret;
621 	bool active;
622 
623 	/* If userspace should engineer that these pages are replaced in
624 	 * the vma between us binding this page into the GTT and completion
625 	 * of rendering... Their loss. If they change the mapping of their
626 	 * pages they need to create a new bo to point to the new vma.
627 	 *
628 	 * However, that still leaves open the possibility of the vma
629 	 * being copied upon fork. Which falls under the same userspace
630 	 * synchronisation issue as a regular bo, except that this time
631 	 * the process may not be expecting that a particular piece of
632 	 * memory is tied to the GPU.
633 	 *
634 	 * Fortunately, we can hook into the mmu_notifier in order to
635 	 * discard the page references prior to anything nasty happening
636 	 * to the vma (discard or cloning) which should prevent the more
637 	 * egregious cases from causing harm.
638 	 */
639 
640 	if (obj->userptr.work) {
641 		/* active flag should still be held for the pending work */
642 		if (IS_ERR(obj->userptr.work))
643 			return PTR_ERR(obj->userptr.work);
644 		else
645 			return -EAGAIN;
646 	}
647 
648 	/* Let the mmu-notifier know that we have begun and need cancellation */
649 	ret = __i915_gem_userptr_set_active(obj, true);
650 	if (ret)
651 		return ret;
652 
653 	pvec = NULL;
654 	pinned = 0;
655 	if (obj->userptr.mm->mm == current->mm) {
656 		pvec = drm_malloc_gfp(num_pages, sizeof(struct page *),
657 				      GFP_TEMPORARY);
658 		if (pvec == NULL) {
659 			__i915_gem_userptr_set_active(obj, false);
660 			return -ENOMEM;
661 		}
662 
663 		pinned = __get_user_pages_fast(obj->userptr.ptr, num_pages,
664 					       !obj->userptr.read_only, pvec);
665 	}
666 
667 	active = false;
668 	if (pinned < 0)
669 		ret = pinned, pinned = 0;
670 	else if (pinned < num_pages)
671 		ret = __i915_gem_userptr_get_pages_schedule(obj, &active);
672 	else
673 		ret = __i915_gem_userptr_set_pages(obj, pvec, num_pages);
674 	if (ret) {
675 		__i915_gem_userptr_set_active(obj, active);
676 		release_pages(pvec, pinned, 0);
677 	}
678 	drm_free_large(pvec);
679 	return ret;
680 #else
681 	return 0;
682 #endif	/* 0 */
683 }
684 
685 static void
686 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj)
687 {
688 	struct sgt_iter sgt_iter;
689 	struct page *page;
690 
691 	BUG_ON(obj->userptr.work != NULL);
692 	__i915_gem_userptr_set_active(obj, false);
693 
694 	if (obj->madv != I915_MADV_WILLNEED)
695 		obj->dirty = 0;
696 
697 	i915_gem_gtt_finish_object(obj);
698 
699 	for_each_sgt_page(page, sgt_iter, obj->pages) {
700 		if (obj->dirty)
701 			set_page_dirty(page);
702 
703 		mark_page_accessed(page);
704 #if 0
705 		put_page(page);
706 #endif
707 	}
708 	obj->dirty = 0;
709 
710 	sg_free_table(obj->pages);
711 	kfree(obj->pages);
712 }
713 
714 static void
715 i915_gem_userptr_release(struct drm_i915_gem_object *obj)
716 {
717 	i915_gem_userptr_release__mmu_notifier(obj);
718 	i915_gem_userptr_release__mm_struct(obj);
719 }
720 
721 static int
722 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
723 {
724 	if (obj->userptr.mmu_object)
725 		return 0;
726 
727 	return i915_gem_userptr_init__mmu_notifier(obj, 0);
728 }
729 
730 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
731 	.flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE,
732 	.get_pages = i915_gem_userptr_get_pages,
733 	.put_pages = i915_gem_userptr_put_pages,
734 	.dmabuf_export = i915_gem_userptr_dmabuf_export,
735 	.release = i915_gem_userptr_release,
736 };
737 
738 /**
739  * Creates a new mm object that wraps some normal memory from the process
740  * context - user memory.
741  *
742  * We impose several restrictions upon the memory being mapped
743  * into the GPU.
744  * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
745  * 2. It must be normal system memory, not a pointer into another map of IO
746  *    space (e.g. it must not be a GTT mmapping of another object).
747  * 3. We only allow a bo as large as we could in theory map into the GTT,
748  *    that is we limit the size to the total size of the GTT.
749  * 4. The bo is marked as being snoopable. The backing pages are left
750  *    accessible directly by the CPU, but reads and writes by the GPU may
751  *    incur the cost of a snoop (unless you have an LLC architecture).
752  *
753  * Synchronisation between multiple users and the GPU is left to userspace
754  * through the normal set-domain-ioctl. The kernel will enforce that the
755  * GPU relinquishes the VMA before it is returned back to the system
756  * i.e. upon free(), munmap() or process termination. However, the userspace
757  * malloc() library may not immediately relinquish the VMA after free() and
758  * instead reuse it whilst the GPU is still reading and writing to the VMA.
759  * Caveat emptor.
760  *
761  * Also note, that the object created here is not currently a "first class"
762  * object, in that several ioctls are banned. These are the CPU access
763  * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
764  * direct access via your pointer rather than use those ioctls. Another
765  * restriction is that we do not allow userptr surfaces to be pinned to the
766  * hardware and so we reject any attempt to create a framebuffer out of a
767  * userptr.
768  *
769  * If you think this is a good interface to use to pass GPU memory between
770  * drivers, please use dma-buf instead. In fact, wherever possible use
771  * dma-buf instead.
772  */
773 int
774 i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
775 {
776 	struct drm_i915_gem_userptr *args = data;
777 	struct drm_i915_gem_object *obj;
778 	int ret;
779 	u32 handle;
780 
781 	if (!HAS_LLC(dev) && !HAS_SNOOP(dev)) {
782 		/* We cannot support coherent userptr objects on hw without
783 		 * LLC and broken snooping.
784 		 */
785 		return -ENODEV;
786 	}
787 
788 	if (args->flags & ~(I915_USERPTR_READ_ONLY |
789 			    I915_USERPTR_UNSYNCHRONIZED))
790 		return -EINVAL;
791 
792 	if (offset_in_page(args->user_ptr | args->user_size))
793 		return -EINVAL;
794 
795 #if 0
796 	if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
797 		       (char __user *)(unsigned long)args->user_ptr, args->user_size))
798 		return -EFAULT;
799 #endif
800 
801 	if (args->flags & I915_USERPTR_READ_ONLY) {
802 		/* On almost all of the current hw, we cannot tell the GPU that a
803 		 * page is readonly, so this is just a placeholder in the uAPI.
804 		 */
805 		return -ENODEV;
806 	}
807 
808 	obj = i915_gem_object_alloc(dev);
809 	if (obj == NULL)
810 		return -ENOMEM;
811 
812 	drm_gem_private_object_init(dev, &obj->base, args->user_size);
813 	i915_gem_object_init(obj, &i915_gem_userptr_ops);
814 	obj->cache_level = I915_CACHE_LLC;
815 	obj->base.write_domain = I915_GEM_DOMAIN_CPU;
816 	obj->base.read_domains = I915_GEM_DOMAIN_CPU;
817 
818 	obj->userptr.ptr = args->user_ptr;
819 	obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY);
820 
821 	/* And keep a pointer to the current->mm for resolving the user pages
822 	 * at binding. This means that we need to hook into the mmu_notifier
823 	 * in order to detect if the mmu is destroyed.
824 	 */
825 	ret = i915_gem_userptr_init__mm_struct(obj);
826 	if (ret == 0)
827 		ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
828 	if (ret == 0)
829 		ret = drm_gem_handle_create(file, &obj->base, &handle);
830 
831 	/* drop reference from allocate - handle holds it now */
832 	i915_gem_object_put_unlocked(obj);
833 	if (ret)
834 		return ret;
835 
836 	args->handle = handle;
837 	return 0;
838 }
839 
840 void i915_gem_init_userptr(struct drm_i915_private *dev_priv)
841 {
842 	lockinit(&dev_priv->mm_lock, "i915dmm", 0, LK_CANRECURSE);
843 	hash_init(dev_priv->mm_structs);
844 }
845