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/mempolicy.h>
33 #include <linux/swap.h>
34 #include <linux/sched/mm.h>
35
36 struct i915_mm_struct {
37 struct mm_struct *mm;
38 struct drm_i915_private *i915;
39 struct i915_mmu_notifier *mn;
40 struct hlist_node node;
41 struct kref kref;
42 struct work_struct work;
43 };
44
45 #if defined(CONFIG_MMU_NOTIFIER)
46 #include <linux/interval_tree.h>
47
48 struct i915_mmu_notifier {
49 spinlock_t lock;
50 struct hlist_node node;
51 struct mmu_notifier mn;
52 struct rb_root_cached objects;
53 struct workqueue_struct *wq;
54 };
55
56 struct i915_mmu_object {
57 struct i915_mmu_notifier *mn;
58 struct drm_i915_gem_object *obj;
59 struct interval_tree_node it;
60 struct list_head link;
61 struct work_struct work;
62 bool attached;
63 };
64
cancel_userptr(struct work_struct * work)65 static void cancel_userptr(struct work_struct *work)
66 {
67 struct i915_mmu_object *mo = container_of(work, typeof(*mo), work);
68 struct drm_i915_gem_object *obj = mo->obj;
69 struct work_struct *active;
70
71 /* Cancel any active worker and force us to re-evaluate gup */
72 mutex_lock(&obj->mm.lock);
73 active = fetch_and_zero(&obj->userptr.work);
74 mutex_unlock(&obj->mm.lock);
75 if (active)
76 goto out;
77
78 i915_gem_object_wait(obj, I915_WAIT_ALL, MAX_SCHEDULE_TIMEOUT, NULL);
79
80 mutex_lock(&obj->base.dev->struct_mutex);
81
82 /* We are inside a kthread context and can't be interrupted */
83 if (i915_gem_object_unbind(obj) == 0)
84 __i915_gem_object_put_pages(obj, I915_MM_NORMAL);
85 WARN_ONCE(i915_gem_object_has_pages(obj),
86 "Failed to release pages: bind_count=%d, pages_pin_count=%d, pin_global=%d\n",
87 obj->bind_count,
88 atomic_read(&obj->mm.pages_pin_count),
89 obj->pin_global);
90
91 mutex_unlock(&obj->base.dev->struct_mutex);
92
93 out:
94 i915_gem_object_put(obj);
95 }
96
add_object(struct i915_mmu_object * mo)97 static void add_object(struct i915_mmu_object *mo)
98 {
99 if (mo->attached)
100 return;
101
102 interval_tree_insert(&mo->it, &mo->mn->objects);
103 mo->attached = true;
104 }
105
del_object(struct i915_mmu_object * mo)106 static void del_object(struct i915_mmu_object *mo)
107 {
108 if (!mo->attached)
109 return;
110
111 interval_tree_remove(&mo->it, &mo->mn->objects);
112 mo->attached = false;
113 }
114
i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier * _mn,struct mm_struct * mm,unsigned long start,unsigned long end,bool blockable)115 static int i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn,
116 struct mm_struct *mm,
117 unsigned long start,
118 unsigned long end,
119 bool blockable)
120 {
121 struct i915_mmu_notifier *mn =
122 container_of(_mn, struct i915_mmu_notifier, mn);
123 struct i915_mmu_object *mo;
124 struct interval_tree_node *it;
125 LIST_HEAD(cancelled);
126
127 if (RB_EMPTY_ROOT(&mn->objects.rb_root))
128 return 0;
129
130 /* interval ranges are inclusive, but invalidate range is exclusive */
131 end--;
132
133 spin_lock(&mn->lock);
134 it = interval_tree_iter_first(&mn->objects, start, end);
135 while (it) {
136 if (!blockable) {
137 spin_unlock(&mn->lock);
138 return -EAGAIN;
139 }
140 /* The mmu_object is released late when destroying the
141 * GEM object so it is entirely possible to gain a
142 * reference on an object in the process of being freed
143 * since our serialisation is via the spinlock and not
144 * the struct_mutex - and consequently use it after it
145 * is freed and then double free it. To prevent that
146 * use-after-free we only acquire a reference on the
147 * object if it is not in the process of being destroyed.
148 */
149 mo = container_of(it, struct i915_mmu_object, it);
150 if (kref_get_unless_zero(&mo->obj->base.refcount))
151 queue_work(mn->wq, &mo->work);
152
153 list_add(&mo->link, &cancelled);
154 it = interval_tree_iter_next(it, start, end);
155 }
156 list_for_each_entry(mo, &cancelled, link)
157 del_object(mo);
158 spin_unlock(&mn->lock);
159
160 if (!list_empty(&cancelled))
161 flush_workqueue(mn->wq);
162
163 return 0;
164 }
165
166 static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
167 .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start,
168 };
169
170 static struct i915_mmu_notifier *
i915_mmu_notifier_create(struct mm_struct * mm)171 i915_mmu_notifier_create(struct mm_struct *mm)
172 {
173 struct i915_mmu_notifier *mn;
174
175 mn = kmalloc(sizeof(*mn), GFP_KERNEL);
176 if (mn == NULL)
177 return ERR_PTR(-ENOMEM);
178
179 spin_lock_init(&mn->lock);
180 mn->mn.ops = &i915_gem_userptr_notifier;
181 mn->objects = RB_ROOT_CACHED;
182 mn->wq = alloc_workqueue("i915-userptr-release",
183 WQ_UNBOUND | WQ_MEM_RECLAIM,
184 0);
185 if (mn->wq == NULL) {
186 kfree(mn);
187 return ERR_PTR(-ENOMEM);
188 }
189
190 return mn;
191 }
192
193 static void
i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object * obj)194 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
195 {
196 struct i915_mmu_object *mo;
197
198 mo = obj->userptr.mmu_object;
199 if (mo == NULL)
200 return;
201
202 spin_lock(&mo->mn->lock);
203 del_object(mo);
204 spin_unlock(&mo->mn->lock);
205 kfree(mo);
206
207 obj->userptr.mmu_object = NULL;
208 }
209
210 static struct i915_mmu_notifier *
i915_mmu_notifier_find(struct i915_mm_struct * mm)211 i915_mmu_notifier_find(struct i915_mm_struct *mm)
212 {
213 struct i915_mmu_notifier *mn;
214 int err = 0;
215
216 mn = mm->mn;
217 if (mn)
218 return mn;
219
220 mn = i915_mmu_notifier_create(mm->mm);
221 if (IS_ERR(mn))
222 err = PTR_ERR(mn);
223
224 down_write(&mm->mm->mmap_sem);
225 mutex_lock(&mm->i915->mm_lock);
226 if (mm->mn == NULL && !err) {
227 /* Protected by mmap_sem (write-lock) */
228 err = __mmu_notifier_register(&mn->mn, mm->mm);
229 if (!err) {
230 /* Protected by mm_lock */
231 mm->mn = fetch_and_zero(&mn);
232 }
233 } else if (mm->mn) {
234 /*
235 * Someone else raced and successfully installed the mmu
236 * notifier, we can cancel our own errors.
237 */
238 err = 0;
239 }
240 mutex_unlock(&mm->i915->mm_lock);
241 up_write(&mm->mm->mmap_sem);
242
243 if (mn && !IS_ERR(mn)) {
244 destroy_workqueue(mn->wq);
245 kfree(mn);
246 }
247
248 return err ? ERR_PTR(err) : mm->mn;
249 }
250
251 static int
i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object * obj,unsigned flags)252 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
253 unsigned flags)
254 {
255 struct i915_mmu_notifier *mn;
256 struct i915_mmu_object *mo;
257
258 if (flags & I915_USERPTR_UNSYNCHRONIZED)
259 return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
260
261 if (WARN_ON(obj->userptr.mm == NULL))
262 return -EINVAL;
263
264 mn = i915_mmu_notifier_find(obj->userptr.mm);
265 if (IS_ERR(mn))
266 return PTR_ERR(mn);
267
268 mo = kzalloc(sizeof(*mo), GFP_KERNEL);
269 if (mo == NULL)
270 return -ENOMEM;
271
272 mo->mn = mn;
273 mo->obj = obj;
274 mo->it.start = obj->userptr.ptr;
275 mo->it.last = obj->userptr.ptr + obj->base.size - 1;
276 INIT_WORK(&mo->work, cancel_userptr);
277
278 obj->userptr.mmu_object = mo;
279 return 0;
280 }
281
282 static void
i915_mmu_notifier_free(struct i915_mmu_notifier * mn,struct mm_struct * mm)283 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
284 struct mm_struct *mm)
285 {
286 if (mn == NULL)
287 return;
288
289 mmu_notifier_unregister(&mn->mn, mm);
290 destroy_workqueue(mn->wq);
291 kfree(mn);
292 }
293
294 #else
295
296 static void
i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object * obj)297 i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
298 {
299 }
300
301 static int
i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object * obj,unsigned flags)302 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
303 unsigned flags)
304 {
305 if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
306 return -ENODEV;
307
308 if (!capable(CAP_SYS_ADMIN))
309 return -EPERM;
310
311 return 0;
312 }
313
314 static void
i915_mmu_notifier_free(struct i915_mmu_notifier * mn,struct mm_struct * mm)315 i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
316 struct mm_struct *mm)
317 {
318 }
319
320 #endif
321
322 static struct i915_mm_struct *
__i915_mm_struct_find(struct drm_i915_private * dev_priv,struct mm_struct * real)323 __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real)
324 {
325 struct i915_mm_struct *mm;
326
327 /* Protected by dev_priv->mm_lock */
328 hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real)
329 if (mm->mm == real)
330 return mm;
331
332 return NULL;
333 }
334
335 static int
i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object * obj)336 i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj)
337 {
338 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
339 struct i915_mm_struct *mm;
340 int ret = 0;
341
342 /* During release of the GEM object we hold the struct_mutex. This
343 * precludes us from calling mmput() at that time as that may be
344 * the last reference and so call exit_mmap(). exit_mmap() will
345 * attempt to reap the vma, and if we were holding a GTT mmap
346 * would then call drm_gem_vm_close() and attempt to reacquire
347 * the struct mutex. So in order to avoid that recursion, we have
348 * to defer releasing the mm reference until after we drop the
349 * struct_mutex, i.e. we need to schedule a worker to do the clean
350 * up.
351 */
352 mutex_lock(&dev_priv->mm_lock);
353 mm = __i915_mm_struct_find(dev_priv, current->mm);
354 if (mm == NULL) {
355 mm = kmalloc(sizeof(*mm), GFP_KERNEL);
356 if (mm == NULL) {
357 ret = -ENOMEM;
358 goto out;
359 }
360
361 kref_init(&mm->kref);
362 mm->i915 = to_i915(obj->base.dev);
363
364 mm->mm = current->mm;
365 mmgrab(current->mm);
366
367 mm->mn = NULL;
368
369 /* Protected by dev_priv->mm_lock */
370 hash_add(dev_priv->mm_structs,
371 &mm->node, (unsigned long)mm->mm);
372 } else
373 kref_get(&mm->kref);
374
375 obj->userptr.mm = mm;
376 out:
377 mutex_unlock(&dev_priv->mm_lock);
378 return ret;
379 }
380
381 static void
__i915_mm_struct_free__worker(struct work_struct * work)382 __i915_mm_struct_free__worker(struct work_struct *work)
383 {
384 struct i915_mm_struct *mm = container_of(work, typeof(*mm), work);
385 i915_mmu_notifier_free(mm->mn, mm->mm);
386 mmdrop(mm->mm);
387 kfree(mm);
388 }
389
390 static void
__i915_mm_struct_free(struct kref * kref)391 __i915_mm_struct_free(struct kref *kref)
392 {
393 struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref);
394
395 /* Protected by dev_priv->mm_lock */
396 hash_del(&mm->node);
397 mutex_unlock(&mm->i915->mm_lock);
398
399 INIT_WORK(&mm->work, __i915_mm_struct_free__worker);
400 queue_work(mm->i915->mm.userptr_wq, &mm->work);
401 }
402
403 static void
i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object * obj)404 i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj)
405 {
406 if (obj->userptr.mm == NULL)
407 return;
408
409 kref_put_mutex(&obj->userptr.mm->kref,
410 __i915_mm_struct_free,
411 &to_i915(obj->base.dev)->mm_lock);
412 obj->userptr.mm = NULL;
413 }
414
415 struct get_pages_work {
416 struct work_struct work;
417 struct drm_i915_gem_object *obj;
418 struct task_struct *task;
419 };
420
421 static struct sg_table *
__i915_gem_userptr_alloc_pages(struct drm_i915_gem_object * obj,struct page ** pvec,int num_pages)422 __i915_gem_userptr_alloc_pages(struct drm_i915_gem_object *obj,
423 struct page **pvec, int num_pages)
424 {
425 unsigned int max_segment = i915_sg_segment_size();
426 struct sg_table *st;
427 unsigned int sg_page_sizes;
428 int ret;
429
430 st = kmalloc(sizeof(*st), GFP_KERNEL);
431 if (!st)
432 return ERR_PTR(-ENOMEM);
433
434 alloc_table:
435 ret = __sg_alloc_table_from_pages(st, pvec, num_pages,
436 0, num_pages << PAGE_SHIFT,
437 max_segment,
438 GFP_KERNEL);
439 if (ret) {
440 kfree(st);
441 return ERR_PTR(ret);
442 }
443
444 ret = i915_gem_gtt_prepare_pages(obj, st);
445 if (ret) {
446 sg_free_table(st);
447
448 if (max_segment > PAGE_SIZE) {
449 max_segment = PAGE_SIZE;
450 goto alloc_table;
451 }
452
453 kfree(st);
454 return ERR_PTR(ret);
455 }
456
457 sg_page_sizes = i915_sg_page_sizes(st->sgl);
458
459 __i915_gem_object_set_pages(obj, st, sg_page_sizes);
460
461 return st;
462 }
463
464 static int
__i915_gem_userptr_set_active(struct drm_i915_gem_object * obj,bool value)465 __i915_gem_userptr_set_active(struct drm_i915_gem_object *obj,
466 bool value)
467 {
468 int ret = 0;
469
470 /* During mm_invalidate_range we need to cancel any userptr that
471 * overlaps the range being invalidated. Doing so requires the
472 * struct_mutex, and that risks recursion. In order to cause
473 * recursion, the user must alias the userptr address space with
474 * a GTT mmapping (possible with a MAP_FIXED) - then when we have
475 * to invalidate that mmaping, mm_invalidate_range is called with
476 * the userptr address *and* the struct_mutex held. To prevent that
477 * we set a flag under the i915_mmu_notifier spinlock to indicate
478 * whether this object is valid.
479 */
480 #if defined(CONFIG_MMU_NOTIFIER)
481 if (obj->userptr.mmu_object == NULL)
482 return 0;
483
484 spin_lock(&obj->userptr.mmu_object->mn->lock);
485 /* In order to serialise get_pages with an outstanding
486 * cancel_userptr, we must drop the struct_mutex and try again.
487 */
488 if (!value)
489 del_object(obj->userptr.mmu_object);
490 else if (!work_pending(&obj->userptr.mmu_object->work))
491 add_object(obj->userptr.mmu_object);
492 else
493 ret = -EAGAIN;
494 spin_unlock(&obj->userptr.mmu_object->mn->lock);
495 #endif
496
497 return ret;
498 }
499
500 static void
__i915_gem_userptr_get_pages_worker(struct work_struct * _work)501 __i915_gem_userptr_get_pages_worker(struct work_struct *_work)
502 {
503 struct get_pages_work *work = container_of(_work, typeof(*work), work);
504 struct drm_i915_gem_object *obj = work->obj;
505 const int npages = obj->base.size >> PAGE_SHIFT;
506 struct page **pvec;
507 int pinned, ret;
508
509 ret = -ENOMEM;
510 pinned = 0;
511
512 pvec = kvmalloc_array(npages, sizeof(struct page *), GFP_KERNEL);
513 if (pvec != NULL) {
514 struct mm_struct *mm = obj->userptr.mm->mm;
515 unsigned int flags = 0;
516
517 if (!i915_gem_object_is_readonly(obj))
518 flags |= FOLL_WRITE;
519
520 ret = -EFAULT;
521 if (mmget_not_zero(mm)) {
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 flags,
529 pvec + pinned, NULL, 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(&obj->mm.lock);
541 if (obj->userptr.work == &work->work) {
542 struct sg_table *pages = ERR_PTR(ret);
543
544 if (pinned == npages) {
545 pages = __i915_gem_userptr_alloc_pages(obj, pvec,
546 npages);
547 if (!IS_ERR(pages)) {
548 pinned = 0;
549 pages = NULL;
550 }
551 }
552
553 obj->userptr.work = ERR_CAST(pages);
554 if (IS_ERR(pages))
555 __i915_gem_userptr_set_active(obj, false);
556 }
557 mutex_unlock(&obj->mm.lock);
558
559 release_pages(pvec, pinned);
560 kvfree(pvec);
561
562 i915_gem_object_put(obj);
563 put_task_struct(work->task);
564 kfree(work);
565 }
566
567 static struct sg_table *
__i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object * obj)568 __i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object *obj)
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 work = kmalloc(sizeof(*work), GFP_KERNEL);
592 if (work == NULL)
593 return ERR_PTR(-ENOMEM);
594
595 obj->userptr.work = &work->work;
596
597 work->obj = i915_gem_object_get(obj);
598
599 work->task = current;
600 get_task_struct(work->task);
601
602 INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
603 queue_work(to_i915(obj->base.dev)->mm.userptr_wq, &work->work);
604
605 return ERR_PTR(-EAGAIN);
606 }
607
i915_gem_userptr_get_pages(struct drm_i915_gem_object * obj)608 static int i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
609 {
610 const int num_pages = obj->base.size >> PAGE_SHIFT;
611 struct mm_struct *mm = obj->userptr.mm->mm;
612 struct page **pvec;
613 struct sg_table *pages;
614 bool active;
615 int pinned;
616
617 /* If userspace should engineer that these pages are replaced in
618 * the vma between us binding this page into the GTT and completion
619 * of rendering... Their loss. If they change the mapping of their
620 * pages they need to create a new bo to point to the new vma.
621 *
622 * However, that still leaves open the possibility of the vma
623 * being copied upon fork. Which falls under the same userspace
624 * synchronisation issue as a regular bo, except that this time
625 * the process may not be expecting that a particular piece of
626 * memory is tied to the GPU.
627 *
628 * Fortunately, we can hook into the mmu_notifier in order to
629 * discard the page references prior to anything nasty happening
630 * to the vma (discard or cloning) which should prevent the more
631 * egregious cases from causing harm.
632 */
633
634 if (obj->userptr.work) {
635 /* active flag should still be held for the pending work */
636 if (IS_ERR(obj->userptr.work))
637 return PTR_ERR(obj->userptr.work);
638 else
639 return -EAGAIN;
640 }
641
642 pvec = NULL;
643 pinned = 0;
644
645 if (mm == current->mm) {
646 pvec = kvmalloc_array(num_pages, sizeof(struct page *),
647 GFP_KERNEL |
648 __GFP_NORETRY |
649 __GFP_NOWARN);
650 if (pvec) /* defer to worker if malloc fails */
651 pinned = __get_user_pages_fast(obj->userptr.ptr,
652 num_pages,
653 !i915_gem_object_is_readonly(obj),
654 pvec);
655 }
656
657 active = false;
658 if (pinned < 0) {
659 pages = ERR_PTR(pinned);
660 pinned = 0;
661 } else if (pinned < num_pages) {
662 pages = __i915_gem_userptr_get_pages_schedule(obj);
663 active = pages == ERR_PTR(-EAGAIN);
664 } else {
665 pages = __i915_gem_userptr_alloc_pages(obj, pvec, num_pages);
666 active = !IS_ERR(pages);
667 }
668 if (active)
669 __i915_gem_userptr_set_active(obj, true);
670
671 if (IS_ERR(pages))
672 release_pages(pvec, pinned);
673 kvfree(pvec);
674
675 return PTR_ERR_OR_ZERO(pages);
676 }
677
678 static void
i915_gem_userptr_put_pages(struct drm_i915_gem_object * obj,struct sg_table * pages)679 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj,
680 struct sg_table *pages)
681 {
682 struct sgt_iter sgt_iter;
683 struct page *page;
684
685 BUG_ON(obj->userptr.work != NULL);
686 __i915_gem_userptr_set_active(obj, false);
687
688 if (obj->mm.madv != I915_MADV_WILLNEED)
689 obj->mm.dirty = false;
690
691 i915_gem_gtt_finish_pages(obj, pages);
692
693 for_each_sgt_page(page, sgt_iter, pages) {
694 if (obj->mm.dirty)
695 set_page_dirty(page);
696
697 mark_page_accessed(page);
698 put_page(page);
699 }
700 obj->mm.dirty = false;
701
702 sg_free_table(pages);
703 kfree(pages);
704 }
705
706 static void
i915_gem_userptr_release(struct drm_i915_gem_object * obj)707 i915_gem_userptr_release(struct drm_i915_gem_object *obj)
708 {
709 i915_gem_userptr_release__mmu_notifier(obj);
710 i915_gem_userptr_release__mm_struct(obj);
711 }
712
713 static int
i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object * obj)714 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
715 {
716 if (obj->userptr.mmu_object)
717 return 0;
718
719 return i915_gem_userptr_init__mmu_notifier(obj, 0);
720 }
721
722 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
723 .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE |
724 I915_GEM_OBJECT_IS_SHRINKABLE,
725 .get_pages = i915_gem_userptr_get_pages,
726 .put_pages = i915_gem_userptr_put_pages,
727 .dmabuf_export = i915_gem_userptr_dmabuf_export,
728 .release = i915_gem_userptr_release,
729 };
730
731 /*
732 * Creates a new mm object that wraps some normal memory from the process
733 * context - user memory.
734 *
735 * We impose several restrictions upon the memory being mapped
736 * into the GPU.
737 * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
738 * 2. It must be normal system memory, not a pointer into another map of IO
739 * space (e.g. it must not be a GTT mmapping of another object).
740 * 3. We only allow a bo as large as we could in theory map into the GTT,
741 * that is we limit the size to the total size of the GTT.
742 * 4. The bo is marked as being snoopable. The backing pages are left
743 * accessible directly by the CPU, but reads and writes by the GPU may
744 * incur the cost of a snoop (unless you have an LLC architecture).
745 *
746 * Synchronisation between multiple users and the GPU is left to userspace
747 * through the normal set-domain-ioctl. The kernel will enforce that the
748 * GPU relinquishes the VMA before it is returned back to the system
749 * i.e. upon free(), munmap() or process termination. However, the userspace
750 * malloc() library may not immediately relinquish the VMA after free() and
751 * instead reuse it whilst the GPU is still reading and writing to the VMA.
752 * Caveat emptor.
753 *
754 * Also note, that the object created here is not currently a "first class"
755 * object, in that several ioctls are banned. These are the CPU access
756 * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
757 * direct access via your pointer rather than use those ioctls. Another
758 * restriction is that we do not allow userptr surfaces to be pinned to the
759 * hardware and so we reject any attempt to create a framebuffer out of a
760 * userptr.
761 *
762 * If you think this is a good interface to use to pass GPU memory between
763 * drivers, please use dma-buf instead. In fact, wherever possible use
764 * dma-buf instead.
765 */
766 int
i915_gem_userptr_ioctl(struct drm_device * dev,void * data,struct drm_file * file)767 i915_gem_userptr_ioctl(struct drm_device *dev,
768 void *data,
769 struct drm_file *file)
770 {
771 struct drm_i915_private *dev_priv = to_i915(dev);
772 struct drm_i915_gem_userptr *args = data;
773 struct drm_i915_gem_object *obj;
774 int ret;
775 u32 handle;
776
777 if (!HAS_LLC(dev_priv) && !HAS_SNOOP(dev_priv)) {
778 /* We cannot support coherent userptr objects on hw without
779 * LLC and broken snooping.
780 */
781 return -ENODEV;
782 }
783
784 if (args->flags & ~(I915_USERPTR_READ_ONLY |
785 I915_USERPTR_UNSYNCHRONIZED))
786 return -EINVAL;
787
788 if (!args->user_size)
789 return -EINVAL;
790
791 if (offset_in_page(args->user_ptr | args->user_size))
792 return -EINVAL;
793
794 if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
795 (char __user *)(unsigned long)args->user_ptr, args->user_size))
796 return -EFAULT;
797
798 if (args->flags & I915_USERPTR_READ_ONLY) {
799 struct i915_hw_ppgtt *ppgtt;
800
801 /*
802 * On almost all of the older hw, we cannot tell the GPU that
803 * a page is readonly.
804 */
805 ppgtt = dev_priv->kernel_context->ppgtt;
806 if (!ppgtt || !ppgtt->vm.has_read_only)
807 return -ENODEV;
808 }
809
810 obj = i915_gem_object_alloc(dev_priv);
811 if (obj == NULL)
812 return -ENOMEM;
813
814 drm_gem_private_object_init(dev, &obj->base, args->user_size);
815 i915_gem_object_init(obj, &i915_gem_userptr_ops);
816 obj->read_domains = I915_GEM_DOMAIN_CPU;
817 obj->write_domain = I915_GEM_DOMAIN_CPU;
818 i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
819
820 obj->userptr.ptr = args->user_ptr;
821 if (args->flags & I915_USERPTR_READ_ONLY)
822 i915_gem_object_set_readonly(obj);
823
824 /* And keep a pointer to the current->mm for resolving the user pages
825 * at binding. This means that we need to hook into the mmu_notifier
826 * in order to detect if the mmu is destroyed.
827 */
828 ret = i915_gem_userptr_init__mm_struct(obj);
829 if (ret == 0)
830 ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
831 if (ret == 0)
832 ret = drm_gem_handle_create(file, &obj->base, &handle);
833
834 /* drop reference from allocate - handle holds it now */
835 i915_gem_object_put(obj);
836 if (ret)
837 return ret;
838
839 args->handle = handle;
840 return 0;
841 }
842
i915_gem_init_userptr(struct drm_i915_private * dev_priv)843 int i915_gem_init_userptr(struct drm_i915_private *dev_priv)
844 {
845 mutex_init(&dev_priv->mm_lock);
846 hash_init(dev_priv->mm_structs);
847
848 dev_priv->mm.userptr_wq =
849 alloc_workqueue("i915-userptr-acquire",
850 WQ_HIGHPRI | WQ_UNBOUND,
851 0);
852 if (!dev_priv->mm.userptr_wq)
853 return -ENOMEM;
854
855 return 0;
856 }
857
i915_gem_cleanup_userptr(struct drm_i915_private * dev_priv)858 void i915_gem_cleanup_userptr(struct drm_i915_private *dev_priv)
859 {
860 destroy_workqueue(dev_priv->mm.userptr_wq);
861 }
862
