drm/i915: Allow overlapping userptr objects
[deliverable/linux.git] / drivers / gpu / drm / i915 / i915_gem_userptr.c
CommitLineData
5cc9ed4b
CW
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 "drmP.h"
26#include "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
35#if defined(CONFIG_MMU_NOTIFIER)
36#include <linux/interval_tree.h>
37
38struct i915_mmu_notifier {
39 spinlock_t lock;
40 struct hlist_node node;
41 struct mmu_notifier mn;
42 struct rb_root objects;
ec8b0dd5 43 struct list_head linear;
5cc9ed4b
CW
44 struct drm_device *dev;
45 struct mm_struct *mm;
46 struct work_struct work;
47 unsigned long count;
48 unsigned long serial;
ec8b0dd5 49 bool has_linear;
5cc9ed4b
CW
50};
51
52struct i915_mmu_object {
53 struct i915_mmu_notifier *mmu;
54 struct interval_tree_node it;
ec8b0dd5 55 struct list_head link;
5cc9ed4b 56 struct drm_i915_gem_object *obj;
ec8b0dd5 57 bool is_linear;
5cc9ed4b
CW
58};
59
ec8b0dd5
CW
60static unsigned long cancel_userptr(struct drm_i915_gem_object *obj)
61{
62 struct drm_device *dev = obj->base.dev;
63 unsigned long end;
64
65 mutex_lock(&dev->struct_mutex);
66 /* Cancel any active worker and force us to re-evaluate gup */
67 obj->userptr.work = NULL;
68
69 if (obj->pages != NULL) {
70 struct drm_i915_private *dev_priv = to_i915(dev);
71 struct i915_vma *vma, *tmp;
72 bool was_interruptible;
73
74 was_interruptible = dev_priv->mm.interruptible;
75 dev_priv->mm.interruptible = false;
76
77 list_for_each_entry_safe(vma, tmp, &obj->vma_list, vma_link) {
78 int ret = i915_vma_unbind(vma);
79 WARN_ON(ret && ret != -EIO);
80 }
81 WARN_ON(i915_gem_object_put_pages(obj));
82
83 dev_priv->mm.interruptible = was_interruptible;
84 }
85
86 end = obj->userptr.ptr + obj->base.size;
87
88 drm_gem_object_unreference(&obj->base);
89 mutex_unlock(&dev->struct_mutex);
90
91 return end;
92}
93
94static void invalidate_range__linear(struct i915_mmu_notifier *mn,
95 struct mm_struct *mm,
96 unsigned long start,
97 unsigned long end)
98{
99 struct i915_mmu_object *mmu;
100 unsigned long serial;
101
102restart:
103 serial = mn->serial;
104 list_for_each_entry(mmu, &mn->linear, link) {
105 struct drm_i915_gem_object *obj;
106
107 if (mmu->it.last < start || mmu->it.start > end)
108 continue;
109
110 obj = mmu->obj;
111 drm_gem_object_reference(&obj->base);
112 spin_unlock(&mn->lock);
113
114 cancel_userptr(obj);
115
116 spin_lock(&mn->lock);
117 if (serial != mn->serial)
118 goto restart;
119 }
120
121 spin_unlock(&mn->lock);
122}
123
5cc9ed4b
CW
124static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn,
125 struct mm_struct *mm,
126 unsigned long start,
127 unsigned long end)
128{
129 struct i915_mmu_notifier *mn = container_of(_mn, struct i915_mmu_notifier, mn);
130 struct interval_tree_node *it = NULL;
ec8b0dd5 131 unsigned long next = start;
5cc9ed4b
CW
132 unsigned long serial = 0;
133
134 end--; /* interval ranges are inclusive, but invalidate range is exclusive */
ec8b0dd5 135 while (next < end) {
5cc9ed4b
CW
136 struct drm_i915_gem_object *obj;
137
138 obj = NULL;
139 spin_lock(&mn->lock);
ec8b0dd5
CW
140 if (mn->has_linear)
141 return invalidate_range__linear(mn, mm, start, end);
5cc9ed4b 142 if (serial == mn->serial)
ec8b0dd5 143 it = interval_tree_iter_next(it, next, end);
5cc9ed4b
CW
144 else
145 it = interval_tree_iter_first(&mn->objects, start, end);
146 if (it != NULL) {
147 obj = container_of(it, struct i915_mmu_object, it)->obj;
148 drm_gem_object_reference(&obj->base);
149 serial = mn->serial;
150 }
151 spin_unlock(&mn->lock);
152 if (obj == NULL)
153 return;
154
ec8b0dd5 155 next = cancel_userptr(obj);
5cc9ed4b
CW
156 }
157}
158
159static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
160 .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start,
161};
162
163static struct i915_mmu_notifier *
164__i915_mmu_notifier_lookup(struct drm_device *dev, struct mm_struct *mm)
165{
166 struct drm_i915_private *dev_priv = to_i915(dev);
167 struct i915_mmu_notifier *mmu;
168
169 /* Protected by dev->struct_mutex */
170 hash_for_each_possible(dev_priv->mmu_notifiers, mmu, node, (unsigned long)mm)
171 if (mmu->mm == mm)
172 return mmu;
173
174 return NULL;
175}
176
177static struct i915_mmu_notifier *
178i915_mmu_notifier_get(struct drm_device *dev, struct mm_struct *mm)
179{
180 struct drm_i915_private *dev_priv = to_i915(dev);
181 struct i915_mmu_notifier *mmu;
182 int ret;
183
184 lockdep_assert_held(&dev->struct_mutex);
185
186 mmu = __i915_mmu_notifier_lookup(dev, mm);
187 if (mmu)
188 return mmu;
189
190 mmu = kmalloc(sizeof(*mmu), GFP_KERNEL);
191 if (mmu == NULL)
192 return ERR_PTR(-ENOMEM);
193
194 spin_lock_init(&mmu->lock);
195 mmu->dev = dev;
196 mmu->mn.ops = &i915_gem_userptr_notifier;
197 mmu->mm = mm;
198 mmu->objects = RB_ROOT;
199 mmu->count = 0;
6c308fec 200 mmu->serial = 1;
ec8b0dd5
CW
201 INIT_LIST_HEAD(&mmu->linear);
202 mmu->has_linear = false;
5cc9ed4b
CW
203
204 /* Protected by mmap_sem (write-lock) */
205 ret = __mmu_notifier_register(&mmu->mn, mm);
206 if (ret) {
207 kfree(mmu);
208 return ERR_PTR(ret);
209 }
210
211 /* Protected by dev->struct_mutex */
212 hash_add(dev_priv->mmu_notifiers, &mmu->node, (unsigned long)mm);
213 return mmu;
214}
215
216static void
217__i915_mmu_notifier_destroy_worker(struct work_struct *work)
218{
219 struct i915_mmu_notifier *mmu = container_of(work, typeof(*mmu), work);
220 mmu_notifier_unregister(&mmu->mn, mmu->mm);
221 kfree(mmu);
222}
223
224static void
225__i915_mmu_notifier_destroy(struct i915_mmu_notifier *mmu)
226{
227 lockdep_assert_held(&mmu->dev->struct_mutex);
228
229 /* Protected by dev->struct_mutex */
230 hash_del(&mmu->node);
231
232 /* Our lock ordering is: mmap_sem, mmu_notifier_scru, struct_mutex.
233 * We enter the function holding struct_mutex, therefore we need
234 * to drop our mutex prior to calling mmu_notifier_unregister in
235 * order to prevent lock inversion (and system-wide deadlock)
236 * between the mmap_sem and struct-mutex. Hence we defer the
237 * unregistration to a workqueue where we hold no locks.
238 */
239 INIT_WORK(&mmu->work, __i915_mmu_notifier_destroy_worker);
240 schedule_work(&mmu->work);
241}
242
243static void __i915_mmu_notifier_update_serial(struct i915_mmu_notifier *mmu)
244{
245 if (++mmu->serial == 0)
246 mmu->serial = 1;
247}
248
ec8b0dd5
CW
249static bool i915_mmu_notifier_has_linear(struct i915_mmu_notifier *mmu)
250{
251 struct i915_mmu_object *mn;
252
253 list_for_each_entry(mn, &mmu->linear, link)
254 if (mn->is_linear)
255 return true;
256
257 return false;
258}
259
5cc9ed4b
CW
260static void
261i915_mmu_notifier_del(struct i915_mmu_notifier *mmu,
262 struct i915_mmu_object *mn)
263{
264 lockdep_assert_held(&mmu->dev->struct_mutex);
265
266 spin_lock(&mmu->lock);
ec8b0dd5
CW
267 list_del(&mn->link);
268 if (mn->is_linear)
269 mmu->has_linear = i915_mmu_notifier_has_linear(mmu);
270 else
271 interval_tree_remove(&mn->it, &mmu->objects);
5cc9ed4b
CW
272 __i915_mmu_notifier_update_serial(mmu);
273 spin_unlock(&mmu->lock);
274
275 /* Protected against _add() by dev->struct_mutex */
276 if (--mmu->count == 0)
277 __i915_mmu_notifier_destroy(mmu);
278}
279
280static int
281i915_mmu_notifier_add(struct i915_mmu_notifier *mmu,
282 struct i915_mmu_object *mn)
283{
284 struct interval_tree_node *it;
285 int ret;
286
287 ret = i915_mutex_lock_interruptible(mmu->dev);
288 if (ret)
289 return ret;
290
291 /* Make sure we drop the final active reference (and thereby
292 * remove the objects from the interval tree) before we do
293 * the check for overlapping objects.
294 */
295 i915_gem_retire_requests(mmu->dev);
296
5cc9ed4b
CW
297 spin_lock(&mmu->lock);
298 it = interval_tree_iter_first(&mmu->objects,
299 mn->it.start, mn->it.last);
300 if (it) {
301 struct drm_i915_gem_object *obj;
302
303 /* We only need to check the first object in the range as it
304 * either has cancelled gup work queued and we need to
305 * return back to the user to give time for the gup-workers
306 * to flush their object references upon which the object will
307 * be removed from the interval-tree, or the the range is
308 * still in use by another client and the overlap is invalid.
ec8b0dd5
CW
309 *
310 * If we do have an overlap, we cannot use the interval tree
311 * for fast range invalidation.
5cc9ed4b
CW
312 */
313
314 obj = container_of(it, struct i915_mmu_object, it)->obj;
ec8b0dd5
CW
315 if (!obj->userptr.workers)
316 mmu->has_linear = mn->is_linear = true;
317 else
318 ret = -EAGAIN;
319 } else
5cc9ed4b 320 interval_tree_insert(&mn->it, &mmu->objects);
ec8b0dd5
CW
321
322 if (ret == 0) {
323 list_add(&mn->link, &mmu->linear);
5cc9ed4b 324 __i915_mmu_notifier_update_serial(mmu);
5cc9ed4b
CW
325 }
326 spin_unlock(&mmu->lock);
327 mutex_unlock(&mmu->dev->struct_mutex);
328
329 return ret;
330}
331
332static void
333i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
334{
335 struct i915_mmu_object *mn;
336
337 mn = obj->userptr.mn;
338 if (mn == NULL)
339 return;
340
341 i915_mmu_notifier_del(mn->mmu, mn);
342 obj->userptr.mn = NULL;
343}
344
345static int
346i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
347 unsigned flags)
348{
349 struct i915_mmu_notifier *mmu;
350 struct i915_mmu_object *mn;
351 int ret;
352
353 if (flags & I915_USERPTR_UNSYNCHRONIZED)
354 return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
355
356 down_write(&obj->userptr.mm->mmap_sem);
357 ret = i915_mutex_lock_interruptible(obj->base.dev);
358 if (ret == 0) {
359 mmu = i915_mmu_notifier_get(obj->base.dev, obj->userptr.mm);
360 if (!IS_ERR(mmu))
361 mmu->count++; /* preemptive add to act as a refcount */
362 else
363 ret = PTR_ERR(mmu);
364 mutex_unlock(&obj->base.dev->struct_mutex);
365 }
366 up_write(&obj->userptr.mm->mmap_sem);
367 if (ret)
368 return ret;
369
370 mn = kzalloc(sizeof(*mn), GFP_KERNEL);
371 if (mn == NULL) {
372 ret = -ENOMEM;
373 goto destroy_mmu;
374 }
375
376 mn->mmu = mmu;
377 mn->it.start = obj->userptr.ptr;
378 mn->it.last = mn->it.start + obj->base.size - 1;
379 mn->obj = obj;
380
381 ret = i915_mmu_notifier_add(mmu, mn);
382 if (ret)
383 goto free_mn;
384
385 obj->userptr.mn = mn;
386 return 0;
387
388free_mn:
389 kfree(mn);
390destroy_mmu:
391 mutex_lock(&obj->base.dev->struct_mutex);
392 if (--mmu->count == 0)
393 __i915_mmu_notifier_destroy(mmu);
394 mutex_unlock(&obj->base.dev->struct_mutex);
395 return ret;
396}
397
398#else
399
400static void
401i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
402{
403}
404
405static int
406i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
407 unsigned flags)
408{
409 if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
410 return -ENODEV;
411
412 if (!capable(CAP_SYS_ADMIN))
413 return -EPERM;
414
415 return 0;
416}
417#endif
418
419struct get_pages_work {
420 struct work_struct work;
421 struct drm_i915_gem_object *obj;
422 struct task_struct *task;
423};
424
425
426#if IS_ENABLED(CONFIG_SWIOTLB)
427#define swiotlb_active() swiotlb_nr_tbl()
428#else
429#define swiotlb_active() 0
430#endif
431
432static int
433st_set_pages(struct sg_table **st, struct page **pvec, int num_pages)
434{
435 struct scatterlist *sg;
436 int ret, n;
437
438 *st = kmalloc(sizeof(**st), GFP_KERNEL);
439 if (*st == NULL)
440 return -ENOMEM;
441
442 if (swiotlb_active()) {
443 ret = sg_alloc_table(*st, num_pages, GFP_KERNEL);
444 if (ret)
445 goto err;
446
447 for_each_sg((*st)->sgl, sg, num_pages, n)
448 sg_set_page(sg, pvec[n], PAGE_SIZE, 0);
449 } else {
450 ret = sg_alloc_table_from_pages(*st, pvec, num_pages,
451 0, num_pages << PAGE_SHIFT,
452 GFP_KERNEL);
453 if (ret)
454 goto err;
455 }
456
457 return 0;
458
459err:
460 kfree(*st);
461 *st = NULL;
462 return ret;
463}
464
465static void
466__i915_gem_userptr_get_pages_worker(struct work_struct *_work)
467{
468 struct get_pages_work *work = container_of(_work, typeof(*work), work);
469 struct drm_i915_gem_object *obj = work->obj;
470 struct drm_device *dev = obj->base.dev;
471 const int num_pages = obj->base.size >> PAGE_SHIFT;
472 struct page **pvec;
473 int pinned, ret;
474
475 ret = -ENOMEM;
476 pinned = 0;
477
478 pvec = kmalloc(num_pages*sizeof(struct page *),
479 GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
480 if (pvec == NULL)
481 pvec = drm_malloc_ab(num_pages, sizeof(struct page *));
482 if (pvec != NULL) {
483 struct mm_struct *mm = obj->userptr.mm;
484
485 down_read(&mm->mmap_sem);
486 while (pinned < num_pages) {
487 ret = get_user_pages(work->task, mm,
488 obj->userptr.ptr + pinned * PAGE_SIZE,
489 num_pages - pinned,
490 !obj->userptr.read_only, 0,
491 pvec + pinned, NULL);
492 if (ret < 0)
493 break;
494
495 pinned += ret;
496 }
497 up_read(&mm->mmap_sem);
498 }
499
500 mutex_lock(&dev->struct_mutex);
501 if (obj->userptr.work != &work->work) {
502 ret = 0;
503 } else if (pinned == num_pages) {
504 ret = st_set_pages(&obj->pages, pvec, num_pages);
505 if (ret == 0) {
506 list_add_tail(&obj->global_list, &to_i915(dev)->mm.unbound_list);
507 pinned = 0;
508 }
509 }
510
511 obj->userptr.work = ERR_PTR(ret);
512 obj->userptr.workers--;
513 drm_gem_object_unreference(&obj->base);
514 mutex_unlock(&dev->struct_mutex);
515
516 release_pages(pvec, pinned, 0);
517 drm_free_large(pvec);
518
519 put_task_struct(work->task);
520 kfree(work);
521}
522
523static int
524i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
525{
526 const int num_pages = obj->base.size >> PAGE_SHIFT;
527 struct page **pvec;
528 int pinned, ret;
529
530 /* If userspace should engineer that these pages are replaced in
531 * the vma between us binding this page into the GTT and completion
532 * of rendering... Their loss. If they change the mapping of their
533 * pages they need to create a new bo to point to the new vma.
534 *
535 * However, that still leaves open the possibility of the vma
536 * being copied upon fork. Which falls under the same userspace
537 * synchronisation issue as a regular bo, except that this time
538 * the process may not be expecting that a particular piece of
539 * memory is tied to the GPU.
540 *
541 * Fortunately, we can hook into the mmu_notifier in order to
542 * discard the page references prior to anything nasty happening
543 * to the vma (discard or cloning) which should prevent the more
544 * egregious cases from causing harm.
545 */
546
547 pvec = NULL;
548 pinned = 0;
549 if (obj->userptr.mm == current->mm) {
550 pvec = kmalloc(num_pages*sizeof(struct page *),
551 GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
552 if (pvec == NULL) {
553 pvec = drm_malloc_ab(num_pages, sizeof(struct page *));
554 if (pvec == NULL)
555 return -ENOMEM;
556 }
557
558 pinned = __get_user_pages_fast(obj->userptr.ptr, num_pages,
559 !obj->userptr.read_only, pvec);
560 }
561 if (pinned < num_pages) {
562 if (pinned < 0) {
563 ret = pinned;
564 pinned = 0;
565 } else {
566 /* Spawn a worker so that we can acquire the
567 * user pages without holding our mutex. Access
568 * to the user pages requires mmap_sem, and we have
569 * a strict lock ordering of mmap_sem, struct_mutex -
570 * we already hold struct_mutex here and so cannot
571 * call gup without encountering a lock inversion.
572 *
573 * Userspace will keep on repeating the operation
574 * (thanks to EAGAIN) until either we hit the fast
575 * path or the worker completes. If the worker is
576 * cancelled or superseded, the task is still run
577 * but the results ignored. (This leads to
578 * complications that we may have a stray object
579 * refcount that we need to be wary of when
580 * checking for existing objects during creation.)
581 * If the worker encounters an error, it reports
582 * that error back to this function through
583 * obj->userptr.work = ERR_PTR.
584 */
585 ret = -EAGAIN;
586 if (obj->userptr.work == NULL &&
587 obj->userptr.workers < I915_GEM_USERPTR_MAX_WORKERS) {
588 struct get_pages_work *work;
589
590 work = kmalloc(sizeof(*work), GFP_KERNEL);
591 if (work != NULL) {
592 obj->userptr.work = &work->work;
593 obj->userptr.workers++;
594
595 work->obj = obj;
596 drm_gem_object_reference(&obj->base);
597
598 work->task = current;
599 get_task_struct(work->task);
600
601 INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
602 schedule_work(&work->work);
603 } else
604 ret = -ENOMEM;
605 } else {
606 if (IS_ERR(obj->userptr.work)) {
607 ret = PTR_ERR(obj->userptr.work);
608 obj->userptr.work = NULL;
609 }
610 }
611 }
612 } else {
613 ret = st_set_pages(&obj->pages, pvec, num_pages);
614 if (ret == 0) {
615 obj->userptr.work = NULL;
616 pinned = 0;
617 }
618 }
619
620 release_pages(pvec, pinned, 0);
621 drm_free_large(pvec);
622 return ret;
623}
624
625static void
626i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj)
627{
628 struct scatterlist *sg;
629 int i;
630
631 BUG_ON(obj->userptr.work != NULL);
632
633 if (obj->madv != I915_MADV_WILLNEED)
634 obj->dirty = 0;
635
636 for_each_sg(obj->pages->sgl, sg, obj->pages->nents, i) {
637 struct page *page = sg_page(sg);
638
639 if (obj->dirty)
640 set_page_dirty(page);
641
642 mark_page_accessed(page);
643 page_cache_release(page);
644 }
645 obj->dirty = 0;
646
647 sg_free_table(obj->pages);
648 kfree(obj->pages);
649}
650
651static void
652i915_gem_userptr_release(struct drm_i915_gem_object *obj)
653{
654 i915_gem_userptr_release__mmu_notifier(obj);
655
656 if (obj->userptr.mm) {
657 mmput(obj->userptr.mm);
658 obj->userptr.mm = NULL;
659 }
660}
661
662static int
663i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
664{
665 if (obj->userptr.mn)
666 return 0;
667
668 return i915_gem_userptr_init__mmu_notifier(obj, 0);
669}
670
671static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
672 .dmabuf_export = i915_gem_userptr_dmabuf_export,
673 .get_pages = i915_gem_userptr_get_pages,
674 .put_pages = i915_gem_userptr_put_pages,
675 .release = i915_gem_userptr_release,
676};
677
678/**
679 * Creates a new mm object that wraps some normal memory from the process
680 * context - user memory.
681 *
682 * We impose several restrictions upon the memory being mapped
683 * into the GPU.
684 * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
ec8b0dd5 685 * 2. It must be normal system memory, not a pointer into another map of IO
5cc9ed4b 686 * space (e.g. it must not be a GTT mmapping of another object).
ec8b0dd5 687 * 3. We only allow a bo as large as we could in theory map into the GTT,
5cc9ed4b 688 * that is we limit the size to the total size of the GTT.
ec8b0dd5 689 * 4. The bo is marked as being snoopable. The backing pages are left
5cc9ed4b
CW
690 * accessible directly by the CPU, but reads and writes by the GPU may
691 * incur the cost of a snoop (unless you have an LLC architecture).
692 *
693 * Synchronisation between multiple users and the GPU is left to userspace
694 * through the normal set-domain-ioctl. The kernel will enforce that the
695 * GPU relinquishes the VMA before it is returned back to the system
696 * i.e. upon free(), munmap() or process termination. However, the userspace
697 * malloc() library may not immediately relinquish the VMA after free() and
698 * instead reuse it whilst the GPU is still reading and writing to the VMA.
699 * Caveat emptor.
700 *
701 * Also note, that the object created here is not currently a "first class"
702 * object, in that several ioctls are banned. These are the CPU access
703 * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
704 * direct access via your pointer rather than use those ioctls.
705 *
706 * If you think this is a good interface to use to pass GPU memory between
707 * drivers, please use dma-buf instead. In fact, wherever possible use
708 * dma-buf instead.
709 */
710int
711i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
712{
713 struct drm_i915_private *dev_priv = dev->dev_private;
714 struct drm_i915_gem_userptr *args = data;
715 struct drm_i915_gem_object *obj;
716 int ret;
717 u32 handle;
718
719 if (args->flags & ~(I915_USERPTR_READ_ONLY |
720 I915_USERPTR_UNSYNCHRONIZED))
721 return -EINVAL;
722
723 if (offset_in_page(args->user_ptr | args->user_size))
724 return -EINVAL;
725
726 if (args->user_size > dev_priv->gtt.base.total)
727 return -E2BIG;
728
729 if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
730 (char __user *)(unsigned long)args->user_ptr, args->user_size))
731 return -EFAULT;
732
733 if (args->flags & I915_USERPTR_READ_ONLY) {
734 /* On almost all of the current hw, we cannot tell the GPU that a
735 * page is readonly, so this is just a placeholder in the uAPI.
736 */
737 return -ENODEV;
738 }
739
740 /* Allocate the new object */
741 obj = i915_gem_object_alloc(dev);
742 if (obj == NULL)
743 return -ENOMEM;
744
745 drm_gem_private_object_init(dev, &obj->base, args->user_size);
746 i915_gem_object_init(obj, &i915_gem_userptr_ops);
747 obj->cache_level = I915_CACHE_LLC;
748 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
749 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
750
751 obj->userptr.ptr = args->user_ptr;
752 obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY);
753
754 /* And keep a pointer to the current->mm for resolving the user pages
755 * at binding. This means that we need to hook into the mmu_notifier
756 * in order to detect if the mmu is destroyed.
757 */
758 ret = -ENOMEM;
759 if ((obj->userptr.mm = get_task_mm(current)))
760 ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
761 if (ret == 0)
762 ret = drm_gem_handle_create(file, &obj->base, &handle);
763
764 /* drop reference from allocate - handle holds it now */
765 drm_gem_object_unreference_unlocked(&obj->base);
766 if (ret)
767 return ret;
768
769 args->handle = handle;
770 return 0;
771}
772
773int
774i915_gem_init_userptr(struct drm_device *dev)
775{
776#if defined(CONFIG_MMU_NOTIFIER)
777 struct drm_i915_private *dev_priv = to_i915(dev);
778 hash_init(dev_priv->mmu_notifiers);
779#endif
780 return 0;
781}
This page took 0.114073 seconds and 5 git commands to generate.