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; | |
393afc2c | 52 | struct workqueue_struct *wq; |
5cc9ed4b CW |
53 | }; |
54 | ||
55 | struct i915_mmu_object { | |
ad46cb53 | 56 | struct i915_mmu_notifier *mn; |
768e159f | 57 | struct drm_i915_gem_object *obj; |
5cc9ed4b | 58 | struct interval_tree_node it; |
ec8b0dd5 | 59 | struct list_head link; |
380996aa | 60 | struct work_struct work; |
768e159f | 61 | bool attached; |
5cc9ed4b CW |
62 | }; |
63 | ||
393afc2c CW |
64 | static void wait_rendering(struct drm_i915_gem_object *obj) |
65 | { | |
66 | struct drm_device *dev = obj->base.dev; | |
67 | struct drm_i915_gem_request *requests[I915_NUM_ENGINES]; | |
68 | unsigned reset_counter; | |
69 | int i, n; | |
70 | ||
71 | if (!obj->active) | |
72 | return; | |
73 | ||
74 | n = 0; | |
75 | for (i = 0; i < I915_NUM_ENGINES; i++) { | |
76 | struct drm_i915_gem_request *req; | |
77 | ||
78 | req = obj->last_read_req[i]; | |
79 | if (req == NULL) | |
80 | continue; | |
81 | ||
82 | requests[n++] = i915_gem_request_reference(req); | |
83 | } | |
84 | ||
85 | reset_counter = atomic_read(&to_i915(dev)->gpu_error.reset_counter); | |
86 | mutex_unlock(&dev->struct_mutex); | |
87 | ||
88 | for (i = 0; i < n; i++) | |
89 | __i915_wait_request(requests[i], reset_counter, false, | |
90 | NULL, NULL); | |
91 | ||
92 | mutex_lock(&dev->struct_mutex); | |
93 | ||
94 | for (i = 0; i < n; i++) | |
95 | i915_gem_request_unreference(requests[i]); | |
96 | } | |
97 | ||
768e159f | 98 | static void cancel_userptr(struct work_struct *work) |
ec8b0dd5 | 99 | { |
380996aa CW |
100 | struct i915_mmu_object *mo = container_of(work, typeof(*mo), work); |
101 | struct drm_i915_gem_object *obj = mo->obj; | |
ec8b0dd5 | 102 | struct drm_device *dev = obj->base.dev; |
ec8b0dd5 CW |
103 | |
104 | mutex_lock(&dev->struct_mutex); | |
105 | /* Cancel any active worker and force us to re-evaluate gup */ | |
106 | obj->userptr.work = NULL; | |
107 | ||
108 | if (obj->pages != NULL) { | |
109 | struct drm_i915_private *dev_priv = to_i915(dev); | |
110 | struct i915_vma *vma, *tmp; | |
111 | bool was_interruptible; | |
112 | ||
393afc2c CW |
113 | wait_rendering(obj); |
114 | ||
ec8b0dd5 CW |
115 | was_interruptible = dev_priv->mm.interruptible; |
116 | dev_priv->mm.interruptible = false; | |
117 | ||
1c7f4bca | 118 | list_for_each_entry_safe(vma, tmp, &obj->vma_list, obj_link) { |
ec8b0dd5 CW |
119 | int ret = i915_vma_unbind(vma); |
120 | WARN_ON(ret && ret != -EIO); | |
121 | } | |
122 | WARN_ON(i915_gem_object_put_pages(obj)); | |
123 | ||
124 | dev_priv->mm.interruptible = was_interruptible; | |
125 | } | |
126 | ||
ec8b0dd5 CW |
127 | drm_gem_object_unreference(&obj->base); |
128 | mutex_unlock(&dev->struct_mutex); | |
ec8b0dd5 CW |
129 | } |
130 | ||
768e159f | 131 | static void add_object(struct i915_mmu_object *mo) |
ec8b0dd5 | 132 | { |
768e159f CW |
133 | if (mo->attached) |
134 | return; | |
ec8b0dd5 | 135 | |
768e159f CW |
136 | interval_tree_insert(&mo->it, &mo->mn->objects); |
137 | mo->attached = true; | |
138 | } | |
139 | ||
140 | static void del_object(struct i915_mmu_object *mo) | |
141 | { | |
142 | if (!mo->attached) | |
143 | return; | |
144 | ||
145 | interval_tree_remove(&mo->it, &mo->mn->objects); | |
146 | mo->attached = false; | |
ec8b0dd5 CW |
147 | } |
148 | ||
5cc9ed4b CW |
149 | static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn, |
150 | struct mm_struct *mm, | |
151 | unsigned long start, | |
152 | unsigned long end) | |
153 | { | |
380996aa CW |
154 | struct i915_mmu_notifier *mn = |
155 | container_of(_mn, struct i915_mmu_notifier, mn); | |
156 | struct i915_mmu_object *mo; | |
768e159f CW |
157 | struct interval_tree_node *it; |
158 | LIST_HEAD(cancelled); | |
159 | ||
160 | if (RB_EMPTY_ROOT(&mn->objects)) | |
161 | return; | |
380996aa CW |
162 | |
163 | /* interval ranges are inclusive, but invalidate range is exclusive */ | |
164 | end--; | |
165 | ||
166 | spin_lock(&mn->lock); | |
768e159f CW |
167 | it = interval_tree_iter_first(&mn->objects, start, end); |
168 | while (it) { | |
169 | /* The mmu_object is released late when destroying the | |
170 | * GEM object so it is entirely possible to gain a | |
171 | * reference on an object in the process of being freed | |
172 | * since our serialisation is via the spinlock and not | |
173 | * the struct_mutex - and consequently use it after it | |
174 | * is freed and then double free it. To prevent that | |
175 | * use-after-free we only acquire a reference on the | |
176 | * object if it is not in the process of being destroyed. | |
177 | */ | |
178 | mo = container_of(it, struct i915_mmu_object, it); | |
179 | if (kref_get_unless_zero(&mo->obj->base.refcount)) | |
393afc2c | 180 | queue_work(mn->wq, &mo->work); |
5cc9ed4b | 181 | |
768e159f CW |
182 | list_add(&mo->link, &cancelled); |
183 | it = interval_tree_iter_next(it, start, end); | |
5cc9ed4b | 184 | } |
768e159f CW |
185 | list_for_each_entry(mo, &cancelled, link) |
186 | del_object(mo); | |
380996aa | 187 | spin_unlock(&mn->lock); |
393afc2c CW |
188 | |
189 | flush_workqueue(mn->wq); | |
5cc9ed4b CW |
190 | } |
191 | ||
192 | static const struct mmu_notifier_ops i915_gem_userptr_notifier = { | |
193 | .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start, | |
194 | }; | |
195 | ||
196 | static struct i915_mmu_notifier * | |
ad46cb53 | 197 | i915_mmu_notifier_create(struct mm_struct *mm) |
5cc9ed4b | 198 | { |
ad46cb53 | 199 | struct i915_mmu_notifier *mn; |
5cc9ed4b CW |
200 | int ret; |
201 | ||
ad46cb53 CW |
202 | mn = kmalloc(sizeof(*mn), GFP_KERNEL); |
203 | if (mn == NULL) | |
5cc9ed4b CW |
204 | return ERR_PTR(-ENOMEM); |
205 | ||
ad46cb53 CW |
206 | spin_lock_init(&mn->lock); |
207 | mn->mn.ops = &i915_gem_userptr_notifier; | |
208 | mn->objects = RB_ROOT; | |
393afc2c CW |
209 | mn->wq = alloc_workqueue("i915-userptr-release", WQ_UNBOUND, 0); |
210 | if (mn->wq == NULL) { | |
211 | kfree(mn); | |
212 | return ERR_PTR(-ENOMEM); | |
213 | } | |
ad46cb53 CW |
214 | |
215 | /* Protected by mmap_sem (write-lock) */ | |
216 | ret = __mmu_notifier_register(&mn->mn, mm); | |
5cc9ed4b | 217 | if (ret) { |
393afc2c | 218 | destroy_workqueue(mn->wq); |
ad46cb53 | 219 | kfree(mn); |
5cc9ed4b CW |
220 | return ERR_PTR(ret); |
221 | } | |
222 | ||
ad46cb53 | 223 | return mn; |
5cc9ed4b CW |
224 | } |
225 | ||
5cc9ed4b CW |
226 | static void |
227 | i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj) | |
228 | { | |
ad46cb53 | 229 | struct i915_mmu_object *mo; |
5cc9ed4b | 230 | |
ad46cb53 CW |
231 | mo = obj->userptr.mmu_object; |
232 | if (mo == NULL) | |
5cc9ed4b CW |
233 | return; |
234 | ||
768e159f CW |
235 | spin_lock(&mo->mn->lock); |
236 | del_object(mo); | |
237 | spin_unlock(&mo->mn->lock); | |
ad46cb53 CW |
238 | kfree(mo); |
239 | ||
240 | obj->userptr.mmu_object = NULL; | |
241 | } | |
242 | ||
243 | static struct i915_mmu_notifier * | |
244 | i915_mmu_notifier_find(struct i915_mm_struct *mm) | |
245 | { | |
e9681366 CW |
246 | struct i915_mmu_notifier *mn = mm->mn; |
247 | ||
248 | mn = mm->mn; | |
249 | if (mn) | |
250 | return mn; | |
251 | ||
252 | down_write(&mm->mm->mmap_sem); | |
253 | mutex_lock(&to_i915(mm->dev)->mm_lock); | |
254 | if ((mn = mm->mn) == NULL) { | |
255 | mn = i915_mmu_notifier_create(mm->mm); | |
256 | if (!IS_ERR(mn)) | |
257 | mm->mn = mn; | |
ad46cb53 | 258 | } |
e9681366 CW |
259 | mutex_unlock(&to_i915(mm->dev)->mm_lock); |
260 | up_write(&mm->mm->mmap_sem); | |
261 | ||
262 | return mn; | |
5cc9ed4b CW |
263 | } |
264 | ||
265 | static int | |
266 | i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj, | |
267 | unsigned flags) | |
268 | { | |
ad46cb53 CW |
269 | struct i915_mmu_notifier *mn; |
270 | struct i915_mmu_object *mo; | |
5cc9ed4b CW |
271 | |
272 | if (flags & I915_USERPTR_UNSYNCHRONIZED) | |
273 | return capable(CAP_SYS_ADMIN) ? 0 : -EPERM; | |
274 | ||
ad46cb53 CW |
275 | if (WARN_ON(obj->userptr.mm == NULL)) |
276 | return -EINVAL; | |
5cc9ed4b | 277 | |
ad46cb53 CW |
278 | mn = i915_mmu_notifier_find(obj->userptr.mm); |
279 | if (IS_ERR(mn)) | |
280 | return PTR_ERR(mn); | |
5cc9ed4b | 281 | |
ad46cb53 CW |
282 | mo = kzalloc(sizeof(*mo), GFP_KERNEL); |
283 | if (mo == NULL) | |
284 | return -ENOMEM; | |
5cc9ed4b | 285 | |
ad46cb53 | 286 | mo->mn = mn; |
ad46cb53 | 287 | mo->obj = obj; |
768e159f CW |
288 | mo->it.start = obj->userptr.ptr; |
289 | mo->it.last = obj->userptr.ptr + obj->base.size - 1; | |
290 | INIT_WORK(&mo->work, cancel_userptr); | |
ad46cb53 CW |
291 | |
292 | obj->userptr.mmu_object = mo; | |
5cc9ed4b | 293 | return 0; |
ad46cb53 CW |
294 | } |
295 | ||
296 | static void | |
297 | i915_mmu_notifier_free(struct i915_mmu_notifier *mn, | |
298 | struct mm_struct *mm) | |
299 | { | |
300 | if (mn == NULL) | |
301 | return; | |
5cc9ed4b | 302 | |
ad46cb53 | 303 | mmu_notifier_unregister(&mn->mn, mm); |
393afc2c | 304 | destroy_workqueue(mn->wq); |
5cc9ed4b | 305 | kfree(mn); |
5cc9ed4b CW |
306 | } |
307 | ||
308 | #else | |
309 | ||
310 | static void | |
311 | i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj) | |
312 | { | |
313 | } | |
314 | ||
315 | static int | |
316 | i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj, | |
317 | unsigned flags) | |
318 | { | |
319 | if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0) | |
320 | return -ENODEV; | |
321 | ||
322 | if (!capable(CAP_SYS_ADMIN)) | |
323 | return -EPERM; | |
324 | ||
325 | return 0; | |
326 | } | |
ad46cb53 CW |
327 | |
328 | static void | |
329 | i915_mmu_notifier_free(struct i915_mmu_notifier *mn, | |
330 | struct mm_struct *mm) | |
331 | { | |
332 | } | |
333 | ||
5cc9ed4b CW |
334 | #endif |
335 | ||
ad46cb53 CW |
336 | static struct i915_mm_struct * |
337 | __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real) | |
338 | { | |
339 | struct i915_mm_struct *mm; | |
340 | ||
341 | /* Protected by dev_priv->mm_lock */ | |
342 | hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real) | |
343 | if (mm->mm == real) | |
344 | return mm; | |
345 | ||
346 | return NULL; | |
347 | } | |
348 | ||
349 | static int | |
350 | i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj) | |
351 | { | |
352 | struct drm_i915_private *dev_priv = to_i915(obj->base.dev); | |
353 | struct i915_mm_struct *mm; | |
354 | int ret = 0; | |
355 | ||
356 | /* During release of the GEM object we hold the struct_mutex. This | |
357 | * precludes us from calling mmput() at that time as that may be | |
358 | * the last reference and so call exit_mmap(). exit_mmap() will | |
359 | * attempt to reap the vma, and if we were holding a GTT mmap | |
360 | * would then call drm_gem_vm_close() and attempt to reacquire | |
361 | * the struct mutex. So in order to avoid that recursion, we have | |
362 | * to defer releasing the mm reference until after we drop the | |
363 | * struct_mutex, i.e. we need to schedule a worker to do the clean | |
364 | * up. | |
365 | */ | |
366 | mutex_lock(&dev_priv->mm_lock); | |
367 | mm = __i915_mm_struct_find(dev_priv, current->mm); | |
368 | if (mm == NULL) { | |
369 | mm = kmalloc(sizeof(*mm), GFP_KERNEL); | |
370 | if (mm == NULL) { | |
371 | ret = -ENOMEM; | |
372 | goto out; | |
373 | } | |
374 | ||
375 | kref_init(&mm->kref); | |
376 | mm->dev = obj->base.dev; | |
377 | ||
378 | mm->mm = current->mm; | |
379 | atomic_inc(¤t->mm->mm_count); | |
380 | ||
381 | mm->mn = NULL; | |
382 | ||
383 | /* Protected by dev_priv->mm_lock */ | |
384 | hash_add(dev_priv->mm_structs, | |
385 | &mm->node, (unsigned long)mm->mm); | |
386 | } else | |
387 | kref_get(&mm->kref); | |
388 | ||
389 | obj->userptr.mm = mm; | |
390 | out: | |
391 | mutex_unlock(&dev_priv->mm_lock); | |
392 | return ret; | |
393 | } | |
394 | ||
395 | static void | |
396 | __i915_mm_struct_free__worker(struct work_struct *work) | |
397 | { | |
398 | struct i915_mm_struct *mm = container_of(work, typeof(*mm), work); | |
399 | i915_mmu_notifier_free(mm->mn, mm->mm); | |
400 | mmdrop(mm->mm); | |
401 | kfree(mm); | |
402 | } | |
403 | ||
404 | static void | |
405 | __i915_mm_struct_free(struct kref *kref) | |
406 | { | |
407 | struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref); | |
408 | ||
409 | /* Protected by dev_priv->mm_lock */ | |
410 | hash_del(&mm->node); | |
411 | mutex_unlock(&to_i915(mm->dev)->mm_lock); | |
412 | ||
413 | INIT_WORK(&mm->work, __i915_mm_struct_free__worker); | |
414 | schedule_work(&mm->work); | |
415 | } | |
416 | ||
417 | static void | |
418 | i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj) | |
419 | { | |
420 | if (obj->userptr.mm == NULL) | |
421 | return; | |
422 | ||
423 | kref_put_mutex(&obj->userptr.mm->kref, | |
424 | __i915_mm_struct_free, | |
425 | &to_i915(obj->base.dev)->mm_lock); | |
426 | obj->userptr.mm = NULL; | |
427 | } | |
428 | ||
5cc9ed4b CW |
429 | struct get_pages_work { |
430 | struct work_struct work; | |
431 | struct drm_i915_gem_object *obj; | |
432 | struct task_struct *task; | |
433 | }; | |
434 | ||
5cc9ed4b CW |
435 | #if IS_ENABLED(CONFIG_SWIOTLB) |
436 | #define swiotlb_active() swiotlb_nr_tbl() | |
437 | #else | |
438 | #define swiotlb_active() 0 | |
439 | #endif | |
440 | ||
441 | static int | |
442 | st_set_pages(struct sg_table **st, struct page **pvec, int num_pages) | |
443 | { | |
444 | struct scatterlist *sg; | |
445 | int ret, n; | |
446 | ||
447 | *st = kmalloc(sizeof(**st), GFP_KERNEL); | |
448 | if (*st == NULL) | |
449 | return -ENOMEM; | |
450 | ||
451 | if (swiotlb_active()) { | |
452 | ret = sg_alloc_table(*st, num_pages, GFP_KERNEL); | |
453 | if (ret) | |
454 | goto err; | |
455 | ||
456 | for_each_sg((*st)->sgl, sg, num_pages, n) | |
457 | sg_set_page(sg, pvec[n], PAGE_SIZE, 0); | |
458 | } else { | |
459 | ret = sg_alloc_table_from_pages(*st, pvec, num_pages, | |
460 | 0, num_pages << PAGE_SHIFT, | |
461 | GFP_KERNEL); | |
462 | if (ret) | |
463 | goto err; | |
464 | } | |
465 | ||
466 | return 0; | |
467 | ||
468 | err: | |
469 | kfree(*st); | |
470 | *st = NULL; | |
471 | return ret; | |
472 | } | |
473 | ||
e2273302 ID |
474 | static int |
475 | __i915_gem_userptr_set_pages(struct drm_i915_gem_object *obj, | |
476 | struct page **pvec, int num_pages) | |
477 | { | |
478 | int ret; | |
479 | ||
480 | ret = st_set_pages(&obj->pages, pvec, num_pages); | |
481 | if (ret) | |
482 | return ret; | |
483 | ||
484 | ret = i915_gem_gtt_prepare_object(obj); | |
485 | if (ret) { | |
486 | sg_free_table(obj->pages); | |
487 | kfree(obj->pages); | |
488 | obj->pages = NULL; | |
489 | } | |
490 | ||
491 | return ret; | |
492 | } | |
493 | ||
380996aa | 494 | static int |
e4b946bf CW |
495 | __i915_gem_userptr_set_active(struct drm_i915_gem_object *obj, |
496 | bool value) | |
497 | { | |
380996aa CW |
498 | int ret = 0; |
499 | ||
e4b946bf CW |
500 | /* During mm_invalidate_range we need to cancel any userptr that |
501 | * overlaps the range being invalidated. Doing so requires the | |
502 | * struct_mutex, and that risks recursion. In order to cause | |
503 | * recursion, the user must alias the userptr address space with | |
504 | * a GTT mmapping (possible with a MAP_FIXED) - then when we have | |
505 | * to invalidate that mmaping, mm_invalidate_range is called with | |
506 | * the userptr address *and* the struct_mutex held. To prevent that | |
507 | * we set a flag under the i915_mmu_notifier spinlock to indicate | |
508 | * whether this object is valid. | |
509 | */ | |
510 | #if defined(CONFIG_MMU_NOTIFIER) | |
511 | if (obj->userptr.mmu_object == NULL) | |
380996aa | 512 | return 0; |
e4b946bf CW |
513 | |
514 | spin_lock(&obj->userptr.mmu_object->mn->lock); | |
380996aa CW |
515 | /* In order to serialise get_pages with an outstanding |
516 | * cancel_userptr, we must drop the struct_mutex and try again. | |
517 | */ | |
768e159f CW |
518 | if (!value) |
519 | del_object(obj->userptr.mmu_object); | |
520 | else if (!work_pending(&obj->userptr.mmu_object->work)) | |
521 | add_object(obj->userptr.mmu_object); | |
380996aa CW |
522 | else |
523 | ret = -EAGAIN; | |
e4b946bf CW |
524 | spin_unlock(&obj->userptr.mmu_object->mn->lock); |
525 | #endif | |
380996aa CW |
526 | |
527 | return ret; | |
e4b946bf CW |
528 | } |
529 | ||
5cc9ed4b CW |
530 | static void |
531 | __i915_gem_userptr_get_pages_worker(struct work_struct *_work) | |
532 | { | |
533 | struct get_pages_work *work = container_of(_work, typeof(*work), work); | |
534 | struct drm_i915_gem_object *obj = work->obj; | |
535 | struct drm_device *dev = obj->base.dev; | |
68d6c840 | 536 | const int npages = obj->base.size >> PAGE_SHIFT; |
5cc9ed4b CW |
537 | struct page **pvec; |
538 | int pinned, ret; | |
539 | ||
540 | ret = -ENOMEM; | |
541 | pinned = 0; | |
542 | ||
f2a85e19 | 543 | pvec = drm_malloc_gfp(npages, sizeof(struct page *), GFP_TEMPORARY); |
5cc9ed4b | 544 | if (pvec != NULL) { |
ad46cb53 | 545 | struct mm_struct *mm = obj->userptr.mm->mm; |
5cc9ed4b CW |
546 | |
547 | down_read(&mm->mmap_sem); | |
68d6c840 | 548 | while (pinned < npages) { |
1e987790 DH |
549 | ret = get_user_pages_remote(work->task, mm, |
550 | obj->userptr.ptr + pinned * PAGE_SIZE, | |
551 | npages - pinned, | |
552 | !obj->userptr.read_only, 0, | |
553 | pvec + pinned, NULL); | |
5cc9ed4b CW |
554 | if (ret < 0) |
555 | break; | |
556 | ||
557 | pinned += ret; | |
558 | } | |
559 | up_read(&mm->mmap_sem); | |
560 | } | |
561 | ||
562 | mutex_lock(&dev->struct_mutex); | |
68d6c840 CW |
563 | if (obj->userptr.work == &work->work) { |
564 | if (pinned == npages) { | |
565 | ret = __i915_gem_userptr_set_pages(obj, pvec, npages); | |
566 | if (ret == 0) { | |
567 | list_add_tail(&obj->global_list, | |
568 | &to_i915(dev)->mm.unbound_list); | |
569 | obj->get_page.sg = obj->pages->sgl; | |
570 | obj->get_page.last = 0; | |
571 | pinned = 0; | |
572 | } | |
5cc9ed4b | 573 | } |
68d6c840 | 574 | obj->userptr.work = ERR_PTR(ret); |
e4b946bf CW |
575 | if (ret) |
576 | __i915_gem_userptr_set_active(obj, false); | |
5cc9ed4b CW |
577 | } |
578 | ||
5cc9ed4b CW |
579 | obj->userptr.workers--; |
580 | drm_gem_object_unreference(&obj->base); | |
581 | mutex_unlock(&dev->struct_mutex); | |
582 | ||
583 | release_pages(pvec, pinned, 0); | |
584 | drm_free_large(pvec); | |
585 | ||
586 | put_task_struct(work->task); | |
587 | kfree(work); | |
588 | } | |
589 | ||
e4b946bf CW |
590 | static int |
591 | __i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object *obj, | |
592 | bool *active) | |
593 | { | |
594 | struct get_pages_work *work; | |
595 | ||
596 | /* Spawn a worker so that we can acquire the | |
597 | * user pages without holding our mutex. Access | |
598 | * to the user pages requires mmap_sem, and we have | |
599 | * a strict lock ordering of mmap_sem, struct_mutex - | |
600 | * we already hold struct_mutex here and so cannot | |
601 | * call gup without encountering a lock inversion. | |
602 | * | |
603 | * Userspace will keep on repeating the operation | |
604 | * (thanks to EAGAIN) until either we hit the fast | |
605 | * path or the worker completes. If the worker is | |
606 | * cancelled or superseded, the task is still run | |
607 | * but the results ignored. (This leads to | |
608 | * complications that we may have a stray object | |
609 | * refcount that we need to be wary of when | |
610 | * checking for existing objects during creation.) | |
611 | * If the worker encounters an error, it reports | |
612 | * that error back to this function through | |
613 | * obj->userptr.work = ERR_PTR. | |
614 | */ | |
615 | if (obj->userptr.workers >= I915_GEM_USERPTR_MAX_WORKERS) | |
616 | return -EAGAIN; | |
617 | ||
618 | work = kmalloc(sizeof(*work), GFP_KERNEL); | |
619 | if (work == NULL) | |
620 | return -ENOMEM; | |
621 | ||
622 | obj->userptr.work = &work->work; | |
623 | obj->userptr.workers++; | |
624 | ||
625 | work->obj = obj; | |
626 | drm_gem_object_reference(&obj->base); | |
627 | ||
628 | work->task = current; | |
629 | get_task_struct(work->task); | |
630 | ||
631 | INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker); | |
632 | schedule_work(&work->work); | |
633 | ||
634 | *active = true; | |
635 | return -EAGAIN; | |
636 | } | |
637 | ||
5cc9ed4b CW |
638 | static int |
639 | i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj) | |
640 | { | |
641 | const int num_pages = obj->base.size >> PAGE_SHIFT; | |
642 | struct page **pvec; | |
643 | int pinned, ret; | |
e4b946bf | 644 | bool active; |
5cc9ed4b CW |
645 | |
646 | /* If userspace should engineer that these pages are replaced in | |
647 | * the vma between us binding this page into the GTT and completion | |
648 | * of rendering... Their loss. If they change the mapping of their | |
649 | * pages they need to create a new bo to point to the new vma. | |
650 | * | |
651 | * However, that still leaves open the possibility of the vma | |
652 | * being copied upon fork. Which falls under the same userspace | |
653 | * synchronisation issue as a regular bo, except that this time | |
654 | * the process may not be expecting that a particular piece of | |
655 | * memory is tied to the GPU. | |
656 | * | |
657 | * Fortunately, we can hook into the mmu_notifier in order to | |
658 | * discard the page references prior to anything nasty happening | |
659 | * to the vma (discard or cloning) which should prevent the more | |
660 | * egregious cases from causing harm. | |
661 | */ | |
e4b946bf CW |
662 | if (IS_ERR(obj->userptr.work)) { |
663 | /* active flag will have been dropped already by the worker */ | |
664 | ret = PTR_ERR(obj->userptr.work); | |
665 | obj->userptr.work = NULL; | |
666 | return ret; | |
667 | } | |
668 | if (obj->userptr.work) | |
669 | /* active flag should still be held for the pending work */ | |
670 | return -EAGAIN; | |
671 | ||
672 | /* Let the mmu-notifier know that we have begun and need cancellation */ | |
380996aa CW |
673 | ret = __i915_gem_userptr_set_active(obj, true); |
674 | if (ret) | |
675 | return ret; | |
5cc9ed4b CW |
676 | |
677 | pvec = NULL; | |
678 | pinned = 0; | |
ad46cb53 | 679 | if (obj->userptr.mm->mm == current->mm) { |
f2a85e19 CW |
680 | pvec = drm_malloc_gfp(num_pages, sizeof(struct page *), |
681 | GFP_TEMPORARY); | |
5cc9ed4b | 682 | if (pvec == NULL) { |
f2a85e19 CW |
683 | __i915_gem_userptr_set_active(obj, false); |
684 | return -ENOMEM; | |
5cc9ed4b CW |
685 | } |
686 | ||
687 | pinned = __get_user_pages_fast(obj->userptr.ptr, num_pages, | |
688 | !obj->userptr.read_only, pvec); | |
689 | } | |
e4b946bf CW |
690 | |
691 | active = false; | |
692 | if (pinned < 0) | |
693 | ret = pinned, pinned = 0; | |
694 | else if (pinned < num_pages) | |
695 | ret = __i915_gem_userptr_get_pages_schedule(obj, &active); | |
696 | else | |
e2273302 | 697 | ret = __i915_gem_userptr_set_pages(obj, pvec, num_pages); |
e4b946bf CW |
698 | if (ret) { |
699 | __i915_gem_userptr_set_active(obj, active); | |
700 | release_pages(pvec, pinned, 0); | |
5cc9ed4b | 701 | } |
5cc9ed4b CW |
702 | drm_free_large(pvec); |
703 | return ret; | |
704 | } | |
705 | ||
706 | static void | |
707 | i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj) | |
708 | { | |
c479f438 | 709 | struct sg_page_iter sg_iter; |
5cc9ed4b CW |
710 | |
711 | BUG_ON(obj->userptr.work != NULL); | |
e4b946bf | 712 | __i915_gem_userptr_set_active(obj, false); |
5cc9ed4b CW |
713 | |
714 | if (obj->madv != I915_MADV_WILLNEED) | |
715 | obj->dirty = 0; | |
716 | ||
e2273302 ID |
717 | i915_gem_gtt_finish_object(obj); |
718 | ||
c479f438 TU |
719 | for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) { |
720 | struct page *page = sg_page_iter_page(&sg_iter); | |
5cc9ed4b CW |
721 | |
722 | if (obj->dirty) | |
723 | set_page_dirty(page); | |
724 | ||
725 | mark_page_accessed(page); | |
09cbfeaf | 726 | put_page(page); |
5cc9ed4b CW |
727 | } |
728 | obj->dirty = 0; | |
729 | ||
730 | sg_free_table(obj->pages); | |
731 | kfree(obj->pages); | |
732 | } | |
733 | ||
734 | static void | |
735 | i915_gem_userptr_release(struct drm_i915_gem_object *obj) | |
736 | { | |
737 | i915_gem_userptr_release__mmu_notifier(obj); | |
ad46cb53 | 738 | i915_gem_userptr_release__mm_struct(obj); |
5cc9ed4b CW |
739 | } |
740 | ||
741 | static int | |
742 | i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj) | |
743 | { | |
ad46cb53 | 744 | if (obj->userptr.mmu_object) |
5cc9ed4b CW |
745 | return 0; |
746 | ||
747 | return i915_gem_userptr_init__mmu_notifier(obj, 0); | |
748 | } | |
749 | ||
750 | static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = { | |
de472664 | 751 | .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE, |
5cc9ed4b CW |
752 | .get_pages = i915_gem_userptr_get_pages, |
753 | .put_pages = i915_gem_userptr_put_pages, | |
de472664 | 754 | .dmabuf_export = i915_gem_userptr_dmabuf_export, |
5cc9ed4b CW |
755 | .release = i915_gem_userptr_release, |
756 | }; | |
757 | ||
758 | /** | |
759 | * Creates a new mm object that wraps some normal memory from the process | |
760 | * context - user memory. | |
761 | * | |
762 | * We impose several restrictions upon the memory being mapped | |
763 | * into the GPU. | |
764 | * 1. It must be page aligned (both start/end addresses, i.e ptr and size). | |
ec8b0dd5 | 765 | * 2. It must be normal system memory, not a pointer into another map of IO |
5cc9ed4b | 766 | * space (e.g. it must not be a GTT mmapping of another object). |
ec8b0dd5 | 767 | * 3. We only allow a bo as large as we could in theory map into the GTT, |
5cc9ed4b | 768 | * that is we limit the size to the total size of the GTT. |
ec8b0dd5 | 769 | * 4. The bo is marked as being snoopable. The backing pages are left |
5cc9ed4b CW |
770 | * accessible directly by the CPU, but reads and writes by the GPU may |
771 | * incur the cost of a snoop (unless you have an LLC architecture). | |
772 | * | |
773 | * Synchronisation between multiple users and the GPU is left to userspace | |
774 | * through the normal set-domain-ioctl. The kernel will enforce that the | |
775 | * GPU relinquishes the VMA before it is returned back to the system | |
776 | * i.e. upon free(), munmap() or process termination. However, the userspace | |
777 | * malloc() library may not immediately relinquish the VMA after free() and | |
778 | * instead reuse it whilst the GPU is still reading and writing to the VMA. | |
779 | * Caveat emptor. | |
780 | * | |
781 | * Also note, that the object created here is not currently a "first class" | |
782 | * object, in that several ioctls are banned. These are the CPU access | |
783 | * ioctls: mmap(), pwrite and pread. In practice, you are expected to use | |
cc917ab4 CW |
784 | * direct access via your pointer rather than use those ioctls. Another |
785 | * restriction is that we do not allow userptr surfaces to be pinned to the | |
786 | * hardware and so we reject any attempt to create a framebuffer out of a | |
787 | * userptr. | |
5cc9ed4b CW |
788 | * |
789 | * If you think this is a good interface to use to pass GPU memory between | |
790 | * drivers, please use dma-buf instead. In fact, wherever possible use | |
791 | * dma-buf instead. | |
792 | */ | |
793 | int | |
794 | i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file) | |
795 | { | |
5cc9ed4b CW |
796 | struct drm_i915_gem_userptr *args = data; |
797 | struct drm_i915_gem_object *obj; | |
798 | int ret; | |
799 | u32 handle; | |
800 | ||
ca377809 TU |
801 | if (!HAS_LLC(dev) && !HAS_SNOOP(dev)) { |
802 | /* We cannot support coherent userptr objects on hw without | |
803 | * LLC and broken snooping. | |
804 | */ | |
805 | return -ENODEV; | |
806 | } | |
807 | ||
5cc9ed4b CW |
808 | if (args->flags & ~(I915_USERPTR_READ_ONLY | |
809 | I915_USERPTR_UNSYNCHRONIZED)) | |
810 | return -EINVAL; | |
811 | ||
812 | if (offset_in_page(args->user_ptr | args->user_size)) | |
813 | return -EINVAL; | |
814 | ||
5cc9ed4b CW |
815 | if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE, |
816 | (char __user *)(unsigned long)args->user_ptr, args->user_size)) | |
817 | return -EFAULT; | |
818 | ||
819 | if (args->flags & I915_USERPTR_READ_ONLY) { | |
820 | /* On almost all of the current hw, we cannot tell the GPU that a | |
821 | * page is readonly, so this is just a placeholder in the uAPI. | |
822 | */ | |
823 | return -ENODEV; | |
824 | } | |
825 | ||
5cc9ed4b CW |
826 | obj = i915_gem_object_alloc(dev); |
827 | if (obj == NULL) | |
828 | return -ENOMEM; | |
829 | ||
830 | drm_gem_private_object_init(dev, &obj->base, args->user_size); | |
831 | i915_gem_object_init(obj, &i915_gem_userptr_ops); | |
832 | obj->cache_level = I915_CACHE_LLC; | |
833 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
834 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
835 | ||
836 | obj->userptr.ptr = args->user_ptr; | |
837 | obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY); | |
838 | ||
839 | /* And keep a pointer to the current->mm for resolving the user pages | |
840 | * at binding. This means that we need to hook into the mmu_notifier | |
841 | * in order to detect if the mmu is destroyed. | |
842 | */ | |
ad46cb53 CW |
843 | ret = i915_gem_userptr_init__mm_struct(obj); |
844 | if (ret == 0) | |
5cc9ed4b CW |
845 | ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags); |
846 | if (ret == 0) | |
847 | ret = drm_gem_handle_create(file, &obj->base, &handle); | |
848 | ||
849 | /* drop reference from allocate - handle holds it now */ | |
850 | drm_gem_object_unreference_unlocked(&obj->base); | |
851 | if (ret) | |
852 | return ret; | |
853 | ||
854 | args->handle = handle; | |
855 | return 0; | |
856 | } | |
857 | ||
858 | int | |
859 | i915_gem_init_userptr(struct drm_device *dev) | |
860 | { | |
5cc9ed4b | 861 | struct drm_i915_private *dev_priv = to_i915(dev); |
ad46cb53 CW |
862 | mutex_init(&dev_priv->mm_lock); |
863 | hash_init(dev_priv->mm_structs); | |
5cc9ed4b CW |
864 | return 0; |
865 | } |