Commit | Line | Data |
---|---|---|
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 | ||
b588c92b ML |
25 | #include <drm/drmP.h> |
26 | #include <drm/i915_drm.h> | |
5cc9ed4b CW |
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 | ||
ad46cb53 CW |
35 | struct i915_mm_struct { |
36 | struct mm_struct *mm; | |
37 | struct drm_device *dev; | |
38 | struct i915_mmu_notifier *mn; | |
39 | struct hlist_node node; | |
40 | struct kref kref; | |
41 | struct work_struct work; | |
42 | }; | |
43 | ||
5cc9ed4b CW |
44 | #if defined(CONFIG_MMU_NOTIFIER) |
45 | #include <linux/interval_tree.h> | |
46 | ||
47 | struct i915_mmu_notifier { | |
48 | spinlock_t lock; | |
49 | struct hlist_node node; | |
50 | struct mmu_notifier mn; | |
51 | struct rb_root objects; | |
5cc9ed4b CW |
52 | }; |
53 | ||
54 | struct i915_mmu_object { | |
ad46cb53 | 55 | struct i915_mmu_notifier *mn; |
768e159f | 56 | struct drm_i915_gem_object *obj; |
5cc9ed4b | 57 | struct interval_tree_node it; |
ec8b0dd5 | 58 | struct list_head link; |
380996aa | 59 | struct work_struct work; |
768e159f | 60 | bool attached; |
5cc9ed4b CW |
61 | }; |
62 | ||
768e159f | 63 | static void cancel_userptr(struct work_struct *work) |
ec8b0dd5 | 64 | { |
380996aa CW |
65 | struct i915_mmu_object *mo = container_of(work, typeof(*mo), work); |
66 | struct drm_i915_gem_object *obj = mo->obj; | |
ec8b0dd5 | 67 | struct drm_device *dev = obj->base.dev; |
ec8b0dd5 CW |
68 | |
69 | mutex_lock(&dev->struct_mutex); | |
70 | /* Cancel any active worker and force us to re-evaluate gup */ | |
71 | obj->userptr.work = NULL; | |
72 | ||
73 | if (obj->pages != NULL) { | |
74 | struct drm_i915_private *dev_priv = to_i915(dev); | |
75 | struct i915_vma *vma, *tmp; | |
76 | bool was_interruptible; | |
77 | ||
78 | was_interruptible = dev_priv->mm.interruptible; | |
79 | dev_priv->mm.interruptible = false; | |
80 | ||
1c7f4bca | 81 | list_for_each_entry_safe(vma, tmp, &obj->vma_list, obj_link) { |
ec8b0dd5 CW |
82 | int ret = i915_vma_unbind(vma); |
83 | WARN_ON(ret && ret != -EIO); | |
84 | } | |
85 | WARN_ON(i915_gem_object_put_pages(obj)); | |
86 | ||
87 | dev_priv->mm.interruptible = was_interruptible; | |
88 | } | |
89 | ||
ec8b0dd5 CW |
90 | drm_gem_object_unreference(&obj->base); |
91 | mutex_unlock(&dev->struct_mutex); | |
ec8b0dd5 CW |
92 | } |
93 | ||
768e159f | 94 | static void add_object(struct i915_mmu_object *mo) |
ec8b0dd5 | 95 | { |
768e159f CW |
96 | if (mo->attached) |
97 | return; | |
ec8b0dd5 | 98 | |
768e159f CW |
99 | interval_tree_insert(&mo->it, &mo->mn->objects); |
100 | mo->attached = true; | |
101 | } | |
102 | ||
103 | static void del_object(struct i915_mmu_object *mo) | |
104 | { | |
105 | if (!mo->attached) | |
106 | return; | |
107 | ||
108 | interval_tree_remove(&mo->it, &mo->mn->objects); | |
109 | mo->attached = false; | |
ec8b0dd5 CW |
110 | } |
111 | ||
5cc9ed4b CW |
112 | static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn, |
113 | struct mm_struct *mm, | |
114 | unsigned long start, | |
115 | unsigned long end) | |
116 | { | |
380996aa CW |
117 | struct i915_mmu_notifier *mn = |
118 | container_of(_mn, struct i915_mmu_notifier, mn); | |
119 | struct i915_mmu_object *mo; | |
768e159f CW |
120 | struct interval_tree_node *it; |
121 | LIST_HEAD(cancelled); | |
122 | ||
123 | if (RB_EMPTY_ROOT(&mn->objects)) | |
124 | return; | |
380996aa CW |
125 | |
126 | /* interval ranges are inclusive, but invalidate range is exclusive */ | |
127 | end--; | |
128 | ||
129 | spin_lock(&mn->lock); | |
768e159f CW |
130 | it = interval_tree_iter_first(&mn->objects, start, end); |
131 | while (it) { | |
132 | /* The mmu_object is released late when destroying the | |
133 | * GEM object so it is entirely possible to gain a | |
134 | * reference on an object in the process of being freed | |
135 | * since our serialisation is via the spinlock and not | |
136 | * the struct_mutex - and consequently use it after it | |
137 | * is freed and then double free it. To prevent that | |
138 | * use-after-free we only acquire a reference on the | |
139 | * object if it is not in the process of being destroyed. | |
140 | */ | |
141 | mo = container_of(it, struct i915_mmu_object, it); | |
142 | if (kref_get_unless_zero(&mo->obj->base.refcount)) | |
143 | schedule_work(&mo->work); | |
5cc9ed4b | 144 | |
768e159f CW |
145 | list_add(&mo->link, &cancelled); |
146 | it = interval_tree_iter_next(it, start, end); | |
5cc9ed4b | 147 | } |
768e159f CW |
148 | list_for_each_entry(mo, &cancelled, link) |
149 | del_object(mo); | |
380996aa | 150 | spin_unlock(&mn->lock); |
5cc9ed4b CW |
151 | } |
152 | ||
153 | static const struct mmu_notifier_ops i915_gem_userptr_notifier = { | |
154 | .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start, | |
155 | }; | |
156 | ||
157 | static struct i915_mmu_notifier * | |
ad46cb53 | 158 | i915_mmu_notifier_create(struct mm_struct *mm) |
5cc9ed4b | 159 | { |
ad46cb53 | 160 | struct i915_mmu_notifier *mn; |
5cc9ed4b CW |
161 | int ret; |
162 | ||
ad46cb53 CW |
163 | mn = kmalloc(sizeof(*mn), GFP_KERNEL); |
164 | if (mn == NULL) | |
5cc9ed4b CW |
165 | return ERR_PTR(-ENOMEM); |
166 | ||
ad46cb53 CW |
167 | spin_lock_init(&mn->lock); |
168 | mn->mn.ops = &i915_gem_userptr_notifier; | |
169 | mn->objects = RB_ROOT; | |
ad46cb53 CW |
170 | |
171 | /* Protected by mmap_sem (write-lock) */ | |
172 | ret = __mmu_notifier_register(&mn->mn, mm); | |
5cc9ed4b | 173 | if (ret) { |
ad46cb53 | 174 | kfree(mn); |
5cc9ed4b CW |
175 | return ERR_PTR(ret); |
176 | } | |
177 | ||
ad46cb53 | 178 | return mn; |
5cc9ed4b CW |
179 | } |
180 | ||
5cc9ed4b CW |
181 | static void |
182 | i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj) | |
183 | { | |
ad46cb53 | 184 | struct i915_mmu_object *mo; |
5cc9ed4b | 185 | |
ad46cb53 CW |
186 | mo = obj->userptr.mmu_object; |
187 | if (mo == NULL) | |
5cc9ed4b CW |
188 | return; |
189 | ||
768e159f CW |
190 | spin_lock(&mo->mn->lock); |
191 | del_object(mo); | |
192 | spin_unlock(&mo->mn->lock); | |
ad46cb53 CW |
193 | kfree(mo); |
194 | ||
195 | obj->userptr.mmu_object = NULL; | |
196 | } | |
197 | ||
198 | static struct i915_mmu_notifier * | |
199 | i915_mmu_notifier_find(struct i915_mm_struct *mm) | |
200 | { | |
e9681366 CW |
201 | struct i915_mmu_notifier *mn = mm->mn; |
202 | ||
203 | mn = mm->mn; | |
204 | if (mn) | |
205 | return mn; | |
206 | ||
207 | down_write(&mm->mm->mmap_sem); | |
208 | mutex_lock(&to_i915(mm->dev)->mm_lock); | |
209 | if ((mn = mm->mn) == NULL) { | |
210 | mn = i915_mmu_notifier_create(mm->mm); | |
211 | if (!IS_ERR(mn)) | |
212 | mm->mn = mn; | |
ad46cb53 | 213 | } |
e9681366 CW |
214 | mutex_unlock(&to_i915(mm->dev)->mm_lock); |
215 | up_write(&mm->mm->mmap_sem); | |
216 | ||
217 | return mn; | |
5cc9ed4b CW |
218 | } |
219 | ||
220 | static int | |
221 | i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj, | |
222 | unsigned flags) | |
223 | { | |
ad46cb53 CW |
224 | struct i915_mmu_notifier *mn; |
225 | struct i915_mmu_object *mo; | |
5cc9ed4b CW |
226 | |
227 | if (flags & I915_USERPTR_UNSYNCHRONIZED) | |
228 | return capable(CAP_SYS_ADMIN) ? 0 : -EPERM; | |
229 | ||
ad46cb53 CW |
230 | if (WARN_ON(obj->userptr.mm == NULL)) |
231 | return -EINVAL; | |
5cc9ed4b | 232 | |
ad46cb53 CW |
233 | mn = i915_mmu_notifier_find(obj->userptr.mm); |
234 | if (IS_ERR(mn)) | |
235 | return PTR_ERR(mn); | |
5cc9ed4b | 236 | |
ad46cb53 CW |
237 | mo = kzalloc(sizeof(*mo), GFP_KERNEL); |
238 | if (mo == NULL) | |
239 | return -ENOMEM; | |
5cc9ed4b | 240 | |
ad46cb53 | 241 | mo->mn = mn; |
ad46cb53 | 242 | mo->obj = obj; |
768e159f CW |
243 | mo->it.start = obj->userptr.ptr; |
244 | mo->it.last = obj->userptr.ptr + obj->base.size - 1; | |
245 | INIT_WORK(&mo->work, cancel_userptr); | |
ad46cb53 CW |
246 | |
247 | obj->userptr.mmu_object = mo; | |
5cc9ed4b | 248 | return 0; |
ad46cb53 CW |
249 | } |
250 | ||
251 | static void | |
252 | i915_mmu_notifier_free(struct i915_mmu_notifier *mn, | |
253 | struct mm_struct *mm) | |
254 | { | |
255 | if (mn == NULL) | |
256 | return; | |
5cc9ed4b | 257 | |
ad46cb53 | 258 | mmu_notifier_unregister(&mn->mn, mm); |
5cc9ed4b | 259 | kfree(mn); |
5cc9ed4b CW |
260 | } |
261 | ||
262 | #else | |
263 | ||
264 | static void | |
265 | i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj) | |
266 | { | |
267 | } | |
268 | ||
269 | static int | |
270 | i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj, | |
271 | unsigned flags) | |
272 | { | |
273 | if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0) | |
274 | return -ENODEV; | |
275 | ||
276 | if (!capable(CAP_SYS_ADMIN)) | |
277 | return -EPERM; | |
278 | ||
279 | return 0; | |
280 | } | |
ad46cb53 CW |
281 | |
282 | static void | |
283 | i915_mmu_notifier_free(struct i915_mmu_notifier *mn, | |
284 | struct mm_struct *mm) | |
285 | { | |
286 | } | |
287 | ||
5cc9ed4b CW |
288 | #endif |
289 | ||
ad46cb53 CW |
290 | static struct i915_mm_struct * |
291 | __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real) | |
292 | { | |
293 | struct i915_mm_struct *mm; | |
294 | ||
295 | /* Protected by dev_priv->mm_lock */ | |
296 | hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real) | |
297 | if (mm->mm == real) | |
298 | return mm; | |
299 | ||
300 | return NULL; | |
301 | } | |
302 | ||
303 | static int | |
304 | i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj) | |
305 | { | |
306 | struct drm_i915_private *dev_priv = to_i915(obj->base.dev); | |
307 | struct i915_mm_struct *mm; | |
308 | int ret = 0; | |
309 | ||
310 | /* During release of the GEM object we hold the struct_mutex. This | |
311 | * precludes us from calling mmput() at that time as that may be | |
312 | * the last reference and so call exit_mmap(). exit_mmap() will | |
313 | * attempt to reap the vma, and if we were holding a GTT mmap | |
314 | * would then call drm_gem_vm_close() and attempt to reacquire | |
315 | * the struct mutex. So in order to avoid that recursion, we have | |
316 | * to defer releasing the mm reference until after we drop the | |
317 | * struct_mutex, i.e. we need to schedule a worker to do the clean | |
318 | * up. | |
319 | */ | |
320 | mutex_lock(&dev_priv->mm_lock); | |
321 | mm = __i915_mm_struct_find(dev_priv, current->mm); | |
322 | if (mm == NULL) { | |
323 | mm = kmalloc(sizeof(*mm), GFP_KERNEL); | |
324 | if (mm == NULL) { | |
325 | ret = -ENOMEM; | |
326 | goto out; | |
327 | } | |
328 | ||
329 | kref_init(&mm->kref); | |
330 | mm->dev = obj->base.dev; | |
331 | ||
332 | mm->mm = current->mm; | |
333 | atomic_inc(¤t->mm->mm_count); | |
334 | ||
335 | mm->mn = NULL; | |
336 | ||
337 | /* Protected by dev_priv->mm_lock */ | |
338 | hash_add(dev_priv->mm_structs, | |
339 | &mm->node, (unsigned long)mm->mm); | |
340 | } else | |
341 | kref_get(&mm->kref); | |
342 | ||
343 | obj->userptr.mm = mm; | |
344 | out: | |
345 | mutex_unlock(&dev_priv->mm_lock); | |
346 | return ret; | |
347 | } | |
348 | ||
349 | static void | |
350 | __i915_mm_struct_free__worker(struct work_struct *work) | |
351 | { | |
352 | struct i915_mm_struct *mm = container_of(work, typeof(*mm), work); | |
353 | i915_mmu_notifier_free(mm->mn, mm->mm); | |
354 | mmdrop(mm->mm); | |
355 | kfree(mm); | |
356 | } | |
357 | ||
358 | static void | |
359 | __i915_mm_struct_free(struct kref *kref) | |
360 | { | |
361 | struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref); | |
362 | ||
363 | /* Protected by dev_priv->mm_lock */ | |
364 | hash_del(&mm->node); | |
365 | mutex_unlock(&to_i915(mm->dev)->mm_lock); | |
366 | ||
367 | INIT_WORK(&mm->work, __i915_mm_struct_free__worker); | |
368 | schedule_work(&mm->work); | |
369 | } | |
370 | ||
371 | static void | |
372 | i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj) | |
373 | { | |
374 | if (obj->userptr.mm == NULL) | |
375 | return; | |
376 | ||
377 | kref_put_mutex(&obj->userptr.mm->kref, | |
378 | __i915_mm_struct_free, | |
379 | &to_i915(obj->base.dev)->mm_lock); | |
380 | obj->userptr.mm = NULL; | |
381 | } | |
382 | ||
5cc9ed4b CW |
383 | struct get_pages_work { |
384 | struct work_struct work; | |
385 | struct drm_i915_gem_object *obj; | |
386 | struct task_struct *task; | |
387 | }; | |
388 | ||
5cc9ed4b CW |
389 | #if IS_ENABLED(CONFIG_SWIOTLB) |
390 | #define swiotlb_active() swiotlb_nr_tbl() | |
391 | #else | |
392 | #define swiotlb_active() 0 | |
393 | #endif | |
394 | ||
395 | static int | |
396 | st_set_pages(struct sg_table **st, struct page **pvec, int num_pages) | |
397 | { | |
398 | struct scatterlist *sg; | |
399 | int ret, n; | |
400 | ||
401 | *st = kmalloc(sizeof(**st), GFP_KERNEL); | |
402 | if (*st == NULL) | |
403 | return -ENOMEM; | |
404 | ||
405 | if (swiotlb_active()) { | |
406 | ret = sg_alloc_table(*st, num_pages, GFP_KERNEL); | |
407 | if (ret) | |
408 | goto err; | |
409 | ||
410 | for_each_sg((*st)->sgl, sg, num_pages, n) | |
411 | sg_set_page(sg, pvec[n], PAGE_SIZE, 0); | |
412 | } else { | |
413 | ret = sg_alloc_table_from_pages(*st, pvec, num_pages, | |
414 | 0, num_pages << PAGE_SHIFT, | |
415 | GFP_KERNEL); | |
416 | if (ret) | |
417 | goto err; | |
418 | } | |
419 | ||
420 | return 0; | |
421 | ||
422 | err: | |
423 | kfree(*st); | |
424 | *st = NULL; | |
425 | return ret; | |
426 | } | |
427 | ||
e2273302 ID |
428 | static int |
429 | __i915_gem_userptr_set_pages(struct drm_i915_gem_object *obj, | |
430 | struct page **pvec, int num_pages) | |
431 | { | |
432 | int ret; | |
433 | ||
434 | ret = st_set_pages(&obj->pages, pvec, num_pages); | |
435 | if (ret) | |
436 | return ret; | |
437 | ||
438 | ret = i915_gem_gtt_prepare_object(obj); | |
439 | if (ret) { | |
440 | sg_free_table(obj->pages); | |
441 | kfree(obj->pages); | |
442 | obj->pages = NULL; | |
443 | } | |
444 | ||
445 | return ret; | |
446 | } | |
447 | ||
380996aa | 448 | static int |
e4b946bf CW |
449 | __i915_gem_userptr_set_active(struct drm_i915_gem_object *obj, |
450 | bool value) | |
451 | { | |
380996aa CW |
452 | int ret = 0; |
453 | ||
e4b946bf CW |
454 | /* During mm_invalidate_range we need to cancel any userptr that |
455 | * overlaps the range being invalidated. Doing so requires the | |
456 | * struct_mutex, and that risks recursion. In order to cause | |
457 | * recursion, the user must alias the userptr address space with | |
458 | * a GTT mmapping (possible with a MAP_FIXED) - then when we have | |
459 | * to invalidate that mmaping, mm_invalidate_range is called with | |
460 | * the userptr address *and* the struct_mutex held. To prevent that | |
461 | * we set a flag under the i915_mmu_notifier spinlock to indicate | |
462 | * whether this object is valid. | |
463 | */ | |
464 | #if defined(CONFIG_MMU_NOTIFIER) | |
465 | if (obj->userptr.mmu_object == NULL) | |
380996aa | 466 | return 0; |
e4b946bf CW |
467 | |
468 | spin_lock(&obj->userptr.mmu_object->mn->lock); | |
380996aa CW |
469 | /* In order to serialise get_pages with an outstanding |
470 | * cancel_userptr, we must drop the struct_mutex and try again. | |
471 | */ | |
768e159f CW |
472 | if (!value) |
473 | del_object(obj->userptr.mmu_object); | |
474 | else if (!work_pending(&obj->userptr.mmu_object->work)) | |
475 | add_object(obj->userptr.mmu_object); | |
380996aa CW |
476 | else |
477 | ret = -EAGAIN; | |
e4b946bf CW |
478 | spin_unlock(&obj->userptr.mmu_object->mn->lock); |
479 | #endif | |
380996aa CW |
480 | |
481 | return ret; | |
e4b946bf CW |
482 | } |
483 | ||
5cc9ed4b CW |
484 | static void |
485 | __i915_gem_userptr_get_pages_worker(struct work_struct *_work) | |
486 | { | |
487 | struct get_pages_work *work = container_of(_work, typeof(*work), work); | |
488 | struct drm_i915_gem_object *obj = work->obj; | |
489 | struct drm_device *dev = obj->base.dev; | |
68d6c840 | 490 | const int npages = obj->base.size >> PAGE_SHIFT; |
5cc9ed4b CW |
491 | struct page **pvec; |
492 | int pinned, ret; | |
493 | ||
494 | ret = -ENOMEM; | |
495 | pinned = 0; | |
496 | ||
68d6c840 | 497 | pvec = kmalloc(npages*sizeof(struct page *), |
5cc9ed4b CW |
498 | GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY); |
499 | if (pvec == NULL) | |
68d6c840 | 500 | pvec = drm_malloc_ab(npages, sizeof(struct page *)); |
5cc9ed4b | 501 | if (pvec != NULL) { |
ad46cb53 | 502 | struct mm_struct *mm = obj->userptr.mm->mm; |
5cc9ed4b | 503 | |
db9f9203 CW |
504 | ret = -EFAULT; |
505 | if (atomic_inc_not_zero(&mm->mm_users)) { | |
506 | down_read(&mm->mmap_sem); | |
507 | while (pinned < npages) { | |
508 | ret = get_user_pages_remote | |
509 | (work->task, mm, | |
510 | obj->userptr.ptr + pinned * PAGE_SIZE, | |
511 | npages - pinned, | |
512 | !obj->userptr.read_only, 0, | |
513 | pvec + pinned, NULL); | |
514 | if (ret < 0) | |
515 | break; | |
516 | ||
517 | pinned += ret; | |
518 | } | |
519 | up_read(&mm->mmap_sem); | |
520 | mmput(mm); | |
5cc9ed4b | 521 | } |
5cc9ed4b CW |
522 | } |
523 | ||
524 | mutex_lock(&dev->struct_mutex); | |
68d6c840 CW |
525 | if (obj->userptr.work == &work->work) { |
526 | if (pinned == npages) { | |
527 | ret = __i915_gem_userptr_set_pages(obj, pvec, npages); | |
528 | if (ret == 0) { | |
529 | list_add_tail(&obj->global_list, | |
530 | &to_i915(dev)->mm.unbound_list); | |
531 | obj->get_page.sg = obj->pages->sgl; | |
532 | obj->get_page.last = 0; | |
533 | pinned = 0; | |
534 | } | |
5cc9ed4b | 535 | } |
68d6c840 | 536 | obj->userptr.work = ERR_PTR(ret); |
e4b946bf CW |
537 | if (ret) |
538 | __i915_gem_userptr_set_active(obj, false); | |
5cc9ed4b CW |
539 | } |
540 | ||
5cc9ed4b CW |
541 | obj->userptr.workers--; |
542 | drm_gem_object_unreference(&obj->base); | |
543 | mutex_unlock(&dev->struct_mutex); | |
544 | ||
545 | release_pages(pvec, pinned, 0); | |
546 | drm_free_large(pvec); | |
547 | ||
548 | put_task_struct(work->task); | |
549 | kfree(work); | |
550 | } | |
551 | ||
e4b946bf CW |
552 | static int |
553 | __i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object *obj, | |
554 | bool *active) | |
555 | { | |
556 | struct get_pages_work *work; | |
557 | ||
558 | /* Spawn a worker so that we can acquire the | |
559 | * user pages without holding our mutex. Access | |
560 | * to the user pages requires mmap_sem, and we have | |
561 | * a strict lock ordering of mmap_sem, struct_mutex - | |
562 | * we already hold struct_mutex here and so cannot | |
563 | * call gup without encountering a lock inversion. | |
564 | * | |
565 | * Userspace will keep on repeating the operation | |
566 | * (thanks to EAGAIN) until either we hit the fast | |
567 | * path or the worker completes. If the worker is | |
568 | * cancelled or superseded, the task is still run | |
569 | * but the results ignored. (This leads to | |
570 | * complications that we may have a stray object | |
571 | * refcount that we need to be wary of when | |
572 | * checking for existing objects during creation.) | |
573 | * If the worker encounters an error, it reports | |
574 | * that error back to this function through | |
575 | * obj->userptr.work = ERR_PTR. | |
576 | */ | |
577 | if (obj->userptr.workers >= I915_GEM_USERPTR_MAX_WORKERS) | |
578 | return -EAGAIN; | |
579 | ||
580 | work = kmalloc(sizeof(*work), GFP_KERNEL); | |
581 | if (work == NULL) | |
582 | return -ENOMEM; | |
583 | ||
584 | obj->userptr.work = &work->work; | |
585 | obj->userptr.workers++; | |
586 | ||
587 | work->obj = obj; | |
588 | drm_gem_object_reference(&obj->base); | |
589 | ||
590 | work->task = current; | |
591 | get_task_struct(work->task); | |
592 | ||
593 | INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker); | |
594 | schedule_work(&work->work); | |
595 | ||
596 | *active = true; | |
597 | return -EAGAIN; | |
598 | } | |
599 | ||
5cc9ed4b CW |
600 | static int |
601 | i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj) | |
602 | { | |
603 | const int num_pages = obj->base.size >> PAGE_SHIFT; | |
604 | struct page **pvec; | |
605 | int pinned, ret; | |
e4b946bf | 606 | bool active; |
5cc9ed4b CW |
607 | |
608 | /* If userspace should engineer that these pages are replaced in | |
609 | * the vma between us binding this page into the GTT and completion | |
610 | * of rendering... Their loss. If they change the mapping of their | |
611 | * pages they need to create a new bo to point to the new vma. | |
612 | * | |
613 | * However, that still leaves open the possibility of the vma | |
614 | * being copied upon fork. Which falls under the same userspace | |
615 | * synchronisation issue as a regular bo, except that this time | |
616 | * the process may not be expecting that a particular piece of | |
617 | * memory is tied to the GPU. | |
618 | * | |
619 | * Fortunately, we can hook into the mmu_notifier in order to | |
620 | * discard the page references prior to anything nasty happening | |
621 | * to the vma (discard or cloning) which should prevent the more | |
622 | * egregious cases from causing harm. | |
623 | */ | |
e4b946bf CW |
624 | if (IS_ERR(obj->userptr.work)) { |
625 | /* active flag will have been dropped already by the worker */ | |
626 | ret = PTR_ERR(obj->userptr.work); | |
627 | obj->userptr.work = NULL; | |
628 | return ret; | |
629 | } | |
630 | if (obj->userptr.work) | |
631 | /* active flag should still be held for the pending work */ | |
632 | return -EAGAIN; | |
633 | ||
634 | /* Let the mmu-notifier know that we have begun and need cancellation */ | |
380996aa CW |
635 | ret = __i915_gem_userptr_set_active(obj, true); |
636 | if (ret) | |
637 | return ret; | |
5cc9ed4b CW |
638 | |
639 | pvec = NULL; | |
640 | pinned = 0; | |
ad46cb53 | 641 | if (obj->userptr.mm->mm == current->mm) { |
5cc9ed4b CW |
642 | pvec = kmalloc(num_pages*sizeof(struct page *), |
643 | GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY); | |
644 | if (pvec == NULL) { | |
645 | pvec = drm_malloc_ab(num_pages, sizeof(struct page *)); | |
e4b946bf CW |
646 | if (pvec == NULL) { |
647 | __i915_gem_userptr_set_active(obj, false); | |
5cc9ed4b | 648 | return -ENOMEM; |
e4b946bf | 649 | } |
5cc9ed4b CW |
650 | } |
651 | ||
652 | pinned = __get_user_pages_fast(obj->userptr.ptr, num_pages, | |
653 | !obj->userptr.read_only, pvec); | |
654 | } | |
e4b946bf CW |
655 | |
656 | active = false; | |
657 | if (pinned < 0) | |
658 | ret = pinned, pinned = 0; | |
659 | else if (pinned < num_pages) | |
660 | ret = __i915_gem_userptr_get_pages_schedule(obj, &active); | |
661 | else | |
e2273302 | 662 | ret = __i915_gem_userptr_set_pages(obj, pvec, num_pages); |
e4b946bf CW |
663 | if (ret) { |
664 | __i915_gem_userptr_set_active(obj, active); | |
665 | release_pages(pvec, pinned, 0); | |
5cc9ed4b | 666 | } |
5cc9ed4b CW |
667 | drm_free_large(pvec); |
668 | return ret; | |
669 | } | |
670 | ||
671 | static void | |
672 | i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj) | |
673 | { | |
c479f438 | 674 | struct sg_page_iter sg_iter; |
5cc9ed4b CW |
675 | |
676 | BUG_ON(obj->userptr.work != NULL); | |
e4b946bf | 677 | __i915_gem_userptr_set_active(obj, false); |
5cc9ed4b CW |
678 | |
679 | if (obj->madv != I915_MADV_WILLNEED) | |
680 | obj->dirty = 0; | |
681 | ||
e2273302 ID |
682 | i915_gem_gtt_finish_object(obj); |
683 | ||
c479f438 TU |
684 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) { |
685 | struct page *page = sg_page_iter_page(&sg_iter); | |
5cc9ed4b CW |
686 | |
687 | if (obj->dirty) | |
688 | set_page_dirty(page); | |
689 | ||
690 | mark_page_accessed(page); | |
09cbfeaf | 691 | put_page(page); |
5cc9ed4b CW |
692 | } |
693 | obj->dirty = 0; | |
694 | ||
695 | sg_free_table(obj->pages); | |
696 | kfree(obj->pages); | |
697 | } | |
698 | ||
699 | static void | |
700 | i915_gem_userptr_release(struct drm_i915_gem_object *obj) | |
701 | { | |
702 | i915_gem_userptr_release__mmu_notifier(obj); | |
ad46cb53 | 703 | i915_gem_userptr_release__mm_struct(obj); |
5cc9ed4b CW |
704 | } |
705 | ||
706 | static int | |
707 | i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj) | |
708 | { | |
ad46cb53 | 709 | if (obj->userptr.mmu_object) |
5cc9ed4b CW |
710 | return 0; |
711 | ||
712 | return i915_gem_userptr_init__mmu_notifier(obj, 0); | |
713 | } | |
714 | ||
715 | static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = { | |
93232aeb | 716 | .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE, |
5cc9ed4b CW |
717 | .get_pages = i915_gem_userptr_get_pages, |
718 | .put_pages = i915_gem_userptr_put_pages, | |
93232aeb | 719 | .dmabuf_export = i915_gem_userptr_dmabuf_export, |
5cc9ed4b CW |
720 | .release = i915_gem_userptr_release, |
721 | }; | |
722 | ||
723 | /** | |
724 | * Creates a new mm object that wraps some normal memory from the process | |
725 | * context - user memory. | |
726 | * | |
727 | * We impose several restrictions upon the memory being mapped | |
728 | * into the GPU. | |
729 | * 1. It must be page aligned (both start/end addresses, i.e ptr and size). | |
ec8b0dd5 | 730 | * 2. It must be normal system memory, not a pointer into another map of IO |
5cc9ed4b | 731 | * space (e.g. it must not be a GTT mmapping of another object). |
ec8b0dd5 | 732 | * 3. We only allow a bo as large as we could in theory map into the GTT, |
5cc9ed4b | 733 | * that is we limit the size to the total size of the GTT. |
ec8b0dd5 | 734 | * 4. The bo is marked as being snoopable. The backing pages are left |
5cc9ed4b CW |
735 | * accessible directly by the CPU, but reads and writes by the GPU may |
736 | * incur the cost of a snoop (unless you have an LLC architecture). | |
737 | * | |
738 | * Synchronisation between multiple users and the GPU is left to userspace | |
739 | * through the normal set-domain-ioctl. The kernel will enforce that the | |
740 | * GPU relinquishes the VMA before it is returned back to the system | |
741 | * i.e. upon free(), munmap() or process termination. However, the userspace | |
742 | * malloc() library may not immediately relinquish the VMA after free() and | |
743 | * instead reuse it whilst the GPU is still reading and writing to the VMA. | |
744 | * Caveat emptor. | |
745 | * | |
746 | * Also note, that the object created here is not currently a "first class" | |
747 | * object, in that several ioctls are banned. These are the CPU access | |
748 | * ioctls: mmap(), pwrite and pread. In practice, you are expected to use | |
cc917ab4 CW |
749 | * direct access via your pointer rather than use those ioctls. Another |
750 | * restriction is that we do not allow userptr surfaces to be pinned to the | |
751 | * hardware and so we reject any attempt to create a framebuffer out of a | |
752 | * userptr. | |
5cc9ed4b CW |
753 | * |
754 | * If you think this is a good interface to use to pass GPU memory between | |
755 | * drivers, please use dma-buf instead. In fact, wherever possible use | |
756 | * dma-buf instead. | |
757 | */ | |
758 | int | |
759 | i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file) | |
760 | { | |
5cc9ed4b CW |
761 | struct drm_i915_gem_userptr *args = data; |
762 | struct drm_i915_gem_object *obj; | |
763 | int ret; | |
764 | u32 handle; | |
765 | ||
766 | if (args->flags & ~(I915_USERPTR_READ_ONLY | | |
767 | I915_USERPTR_UNSYNCHRONIZED)) | |
768 | return -EINVAL; | |
769 | ||
770 | if (offset_in_page(args->user_ptr | args->user_size)) | |
771 | return -EINVAL; | |
772 | ||
5cc9ed4b CW |
773 | if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE, |
774 | (char __user *)(unsigned long)args->user_ptr, args->user_size)) | |
775 | return -EFAULT; | |
776 | ||
777 | if (args->flags & I915_USERPTR_READ_ONLY) { | |
778 | /* On almost all of the current hw, we cannot tell the GPU that a | |
779 | * page is readonly, so this is just a placeholder in the uAPI. | |
780 | */ | |
781 | return -ENODEV; | |
782 | } | |
783 | ||
5cc9ed4b CW |
784 | obj = i915_gem_object_alloc(dev); |
785 | if (obj == NULL) | |
786 | return -ENOMEM; | |
787 | ||
788 | drm_gem_private_object_init(dev, &obj->base, args->user_size); | |
789 | i915_gem_object_init(obj, &i915_gem_userptr_ops); | |
790 | obj->cache_level = I915_CACHE_LLC; | |
791 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
792 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
793 | ||
794 | obj->userptr.ptr = args->user_ptr; | |
795 | obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY); | |
796 | ||
797 | /* And keep a pointer to the current->mm for resolving the user pages | |
798 | * at binding. This means that we need to hook into the mmu_notifier | |
799 | * in order to detect if the mmu is destroyed. | |
800 | */ | |
ad46cb53 CW |
801 | ret = i915_gem_userptr_init__mm_struct(obj); |
802 | if (ret == 0) | |
5cc9ed4b CW |
803 | ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags); |
804 | if (ret == 0) | |
805 | ret = drm_gem_handle_create(file, &obj->base, &handle); | |
806 | ||
807 | /* drop reference from allocate - handle holds it now */ | |
808 | drm_gem_object_unreference_unlocked(&obj->base); | |
809 | if (ret) | |
810 | return ret; | |
811 | ||
812 | args->handle = handle; | |
813 | return 0; | |
814 | } | |
815 | ||
816 | int | |
817 | i915_gem_init_userptr(struct drm_device *dev) | |
818 | { | |
5cc9ed4b | 819 | struct drm_i915_private *dev_priv = to_i915(dev); |
ad46cb53 CW |
820 | mutex_init(&dev_priv->mm_lock); |
821 | hash_init(dev_priv->mm_structs); | |
5cc9ed4b CW |
822 | return 0; |
823 | } |