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 | ||
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 | ||
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; | |
ec8b0dd5 | 52 | struct list_head linear; |
5cc9ed4b | 53 | unsigned long serial; |
ec8b0dd5 | 54 | bool has_linear; |
5cc9ed4b CW |
55 | }; |
56 | ||
57 | struct i915_mmu_object { | |
ad46cb53 | 58 | struct i915_mmu_notifier *mn; |
5cc9ed4b | 59 | struct interval_tree_node it; |
ec8b0dd5 | 60 | struct list_head link; |
5cc9ed4b | 61 | struct drm_i915_gem_object *obj; |
ec8b0dd5 | 62 | bool is_linear; |
5cc9ed4b CW |
63 | }; |
64 | ||
ec8b0dd5 CW |
65 | static unsigned long cancel_userptr(struct drm_i915_gem_object *obj) |
66 | { | |
67 | struct drm_device *dev = obj->base.dev; | |
68 | unsigned long end; | |
69 | ||
70 | mutex_lock(&dev->struct_mutex); | |
71 | /* Cancel any active worker and force us to re-evaluate gup */ | |
72 | obj->userptr.work = NULL; | |
73 | ||
74 | if (obj->pages != NULL) { | |
75 | struct drm_i915_private *dev_priv = to_i915(dev); | |
76 | struct i915_vma *vma, *tmp; | |
77 | bool was_interruptible; | |
78 | ||
79 | was_interruptible = dev_priv->mm.interruptible; | |
80 | dev_priv->mm.interruptible = false; | |
81 | ||
82 | list_for_each_entry_safe(vma, tmp, &obj->vma_list, vma_link) { | |
83 | int ret = i915_vma_unbind(vma); | |
84 | WARN_ON(ret && ret != -EIO); | |
85 | } | |
86 | WARN_ON(i915_gem_object_put_pages(obj)); | |
87 | ||
88 | dev_priv->mm.interruptible = was_interruptible; | |
89 | } | |
90 | ||
91 | end = obj->userptr.ptr + obj->base.size; | |
92 | ||
93 | drm_gem_object_unreference(&obj->base); | |
94 | mutex_unlock(&dev->struct_mutex); | |
95 | ||
96 | return end; | |
97 | } | |
98 | ||
48777767 CW |
99 | static void *invalidate_range__linear(struct i915_mmu_notifier *mn, |
100 | struct mm_struct *mm, | |
101 | unsigned long start, | |
102 | unsigned long end) | |
ec8b0dd5 | 103 | { |
ad46cb53 | 104 | struct i915_mmu_object *mo; |
ec8b0dd5 CW |
105 | unsigned long serial; |
106 | ||
107 | restart: | |
108 | serial = mn->serial; | |
ad46cb53 | 109 | list_for_each_entry(mo, &mn->linear, link) { |
ec8b0dd5 CW |
110 | struct drm_i915_gem_object *obj; |
111 | ||
ad46cb53 | 112 | if (mo->it.last < start || mo->it.start > end) |
ec8b0dd5 CW |
113 | continue; |
114 | ||
ad46cb53 | 115 | obj = mo->obj; |
ec8b0dd5 CW |
116 | drm_gem_object_reference(&obj->base); |
117 | spin_unlock(&mn->lock); | |
118 | ||
119 | cancel_userptr(obj); | |
120 | ||
121 | spin_lock(&mn->lock); | |
122 | if (serial != mn->serial) | |
123 | goto restart; | |
124 | } | |
125 | ||
48777767 | 126 | return NULL; |
ec8b0dd5 CW |
127 | } |
128 | ||
5cc9ed4b CW |
129 | static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn, |
130 | struct mm_struct *mm, | |
131 | unsigned long start, | |
132 | unsigned long end) | |
133 | { | |
134 | struct i915_mmu_notifier *mn = container_of(_mn, struct i915_mmu_notifier, mn); | |
135 | struct interval_tree_node *it = NULL; | |
ec8b0dd5 | 136 | unsigned long next = start; |
5cc9ed4b CW |
137 | unsigned long serial = 0; |
138 | ||
139 | end--; /* interval ranges are inclusive, but invalidate range is exclusive */ | |
ec8b0dd5 | 140 | while (next < end) { |
48777767 | 141 | struct drm_i915_gem_object *obj = NULL; |
5cc9ed4b | 142 | |
5cc9ed4b | 143 | spin_lock(&mn->lock); |
ec8b0dd5 | 144 | if (mn->has_linear) |
48777767 CW |
145 | it = invalidate_range__linear(mn, mm, start, end); |
146 | else if (serial == mn->serial) | |
ec8b0dd5 | 147 | it = interval_tree_iter_next(it, next, end); |
5cc9ed4b CW |
148 | else |
149 | it = interval_tree_iter_first(&mn->objects, start, end); | |
150 | if (it != NULL) { | |
151 | obj = container_of(it, struct i915_mmu_object, it)->obj; | |
152 | drm_gem_object_reference(&obj->base); | |
153 | serial = mn->serial; | |
154 | } | |
155 | spin_unlock(&mn->lock); | |
156 | if (obj == NULL) | |
157 | return; | |
158 | ||
ec8b0dd5 | 159 | next = cancel_userptr(obj); |
5cc9ed4b CW |
160 | } |
161 | } | |
162 | ||
163 | static const struct mmu_notifier_ops i915_gem_userptr_notifier = { | |
164 | .invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start, | |
165 | }; | |
166 | ||
167 | static struct i915_mmu_notifier * | |
ad46cb53 | 168 | i915_mmu_notifier_create(struct mm_struct *mm) |
5cc9ed4b | 169 | { |
ad46cb53 | 170 | struct i915_mmu_notifier *mn; |
5cc9ed4b CW |
171 | int ret; |
172 | ||
ad46cb53 CW |
173 | mn = kmalloc(sizeof(*mn), GFP_KERNEL); |
174 | if (mn == NULL) | |
5cc9ed4b CW |
175 | return ERR_PTR(-ENOMEM); |
176 | ||
ad46cb53 CW |
177 | spin_lock_init(&mn->lock); |
178 | mn->mn.ops = &i915_gem_userptr_notifier; | |
179 | mn->objects = RB_ROOT; | |
180 | mn->serial = 1; | |
181 | INIT_LIST_HEAD(&mn->linear); | |
182 | mn->has_linear = false; | |
183 | ||
184 | /* Protected by mmap_sem (write-lock) */ | |
185 | ret = __mmu_notifier_register(&mn->mn, mm); | |
5cc9ed4b | 186 | if (ret) { |
ad46cb53 | 187 | kfree(mn); |
5cc9ed4b CW |
188 | return ERR_PTR(ret); |
189 | } | |
190 | ||
ad46cb53 | 191 | return mn; |
5cc9ed4b CW |
192 | } |
193 | ||
ad46cb53 | 194 | static void __i915_mmu_notifier_update_serial(struct i915_mmu_notifier *mn) |
5cc9ed4b | 195 | { |
ad46cb53 CW |
196 | if (++mn->serial == 0) |
197 | mn->serial = 1; | |
5cc9ed4b CW |
198 | } |
199 | ||
200 | static int | |
ad46cb53 CW |
201 | i915_mmu_notifier_add(struct drm_device *dev, |
202 | struct i915_mmu_notifier *mn, | |
203 | struct i915_mmu_object *mo) | |
5cc9ed4b CW |
204 | { |
205 | struct interval_tree_node *it; | |
206 | int ret; | |
207 | ||
ad46cb53 | 208 | ret = i915_mutex_lock_interruptible(dev); |
5cc9ed4b CW |
209 | if (ret) |
210 | return ret; | |
211 | ||
212 | /* Make sure we drop the final active reference (and thereby | |
213 | * remove the objects from the interval tree) before we do | |
214 | * the check for overlapping objects. | |
215 | */ | |
ad46cb53 | 216 | i915_gem_retire_requests(dev); |
5cc9ed4b | 217 | |
ad46cb53 CW |
218 | spin_lock(&mn->lock); |
219 | it = interval_tree_iter_first(&mn->objects, | |
220 | mo->it.start, mo->it.last); | |
5cc9ed4b CW |
221 | if (it) { |
222 | struct drm_i915_gem_object *obj; | |
223 | ||
224 | /* We only need to check the first object in the range as it | |
225 | * either has cancelled gup work queued and we need to | |
226 | * return back to the user to give time for the gup-workers | |
227 | * to flush their object references upon which the object will | |
228 | * be removed from the interval-tree, or the the range is | |
229 | * still in use by another client and the overlap is invalid. | |
ec8b0dd5 CW |
230 | * |
231 | * If we do have an overlap, we cannot use the interval tree | |
232 | * for fast range invalidation. | |
5cc9ed4b CW |
233 | */ |
234 | ||
235 | obj = container_of(it, struct i915_mmu_object, it)->obj; | |
ec8b0dd5 | 236 | if (!obj->userptr.workers) |
ad46cb53 | 237 | mn->has_linear = mo->is_linear = true; |
ec8b0dd5 CW |
238 | else |
239 | ret = -EAGAIN; | |
240 | } else | |
ad46cb53 | 241 | interval_tree_insert(&mo->it, &mn->objects); |
ec8b0dd5 CW |
242 | |
243 | if (ret == 0) { | |
ad46cb53 CW |
244 | list_add(&mo->link, &mn->linear); |
245 | __i915_mmu_notifier_update_serial(mn); | |
5cc9ed4b | 246 | } |
ad46cb53 CW |
247 | spin_unlock(&mn->lock); |
248 | mutex_unlock(&dev->struct_mutex); | |
5cc9ed4b CW |
249 | |
250 | return ret; | |
251 | } | |
252 | ||
ad46cb53 CW |
253 | static bool i915_mmu_notifier_has_linear(struct i915_mmu_notifier *mn) |
254 | { | |
255 | struct i915_mmu_object *mo; | |
256 | ||
257 | list_for_each_entry(mo, &mn->linear, link) | |
258 | if (mo->is_linear) | |
259 | return true; | |
260 | ||
261 | return false; | |
262 | } | |
263 | ||
264 | static void | |
265 | i915_mmu_notifier_del(struct i915_mmu_notifier *mn, | |
266 | struct i915_mmu_object *mo) | |
267 | { | |
268 | spin_lock(&mn->lock); | |
269 | list_del(&mo->link); | |
270 | if (mo->is_linear) | |
271 | mn->has_linear = i915_mmu_notifier_has_linear(mn); | |
272 | else | |
273 | interval_tree_remove(&mo->it, &mn->objects); | |
274 | __i915_mmu_notifier_update_serial(mn); | |
275 | spin_unlock(&mn->lock); | |
276 | } | |
277 | ||
5cc9ed4b CW |
278 | static void |
279 | i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj) | |
280 | { | |
ad46cb53 | 281 | struct i915_mmu_object *mo; |
5cc9ed4b | 282 | |
ad46cb53 CW |
283 | mo = obj->userptr.mmu_object; |
284 | if (mo == NULL) | |
5cc9ed4b CW |
285 | return; |
286 | ||
ad46cb53 CW |
287 | i915_mmu_notifier_del(mo->mn, mo); |
288 | kfree(mo); | |
289 | ||
290 | obj->userptr.mmu_object = NULL; | |
291 | } | |
292 | ||
293 | static struct i915_mmu_notifier * | |
294 | i915_mmu_notifier_find(struct i915_mm_struct *mm) | |
295 | { | |
296 | if (mm->mn == NULL) { | |
297 | down_write(&mm->mm->mmap_sem); | |
298 | mutex_lock(&to_i915(mm->dev)->mm_lock); | |
299 | if (mm->mn == NULL) | |
300 | mm->mn = i915_mmu_notifier_create(mm->mm); | |
301 | mutex_unlock(&to_i915(mm->dev)->mm_lock); | |
302 | up_write(&mm->mm->mmap_sem); | |
303 | } | |
304 | return mm->mn; | |
5cc9ed4b CW |
305 | } |
306 | ||
307 | static int | |
308 | i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj, | |
309 | unsigned flags) | |
310 | { | |
ad46cb53 CW |
311 | struct i915_mmu_notifier *mn; |
312 | struct i915_mmu_object *mo; | |
5cc9ed4b CW |
313 | int ret; |
314 | ||
315 | if (flags & I915_USERPTR_UNSYNCHRONIZED) | |
316 | return capable(CAP_SYS_ADMIN) ? 0 : -EPERM; | |
317 | ||
ad46cb53 CW |
318 | if (WARN_ON(obj->userptr.mm == NULL)) |
319 | return -EINVAL; | |
5cc9ed4b | 320 | |
ad46cb53 CW |
321 | mn = i915_mmu_notifier_find(obj->userptr.mm); |
322 | if (IS_ERR(mn)) | |
323 | return PTR_ERR(mn); | |
5cc9ed4b | 324 | |
ad46cb53 CW |
325 | mo = kzalloc(sizeof(*mo), GFP_KERNEL); |
326 | if (mo == NULL) | |
327 | return -ENOMEM; | |
5cc9ed4b | 328 | |
ad46cb53 CW |
329 | mo->mn = mn; |
330 | mo->it.start = obj->userptr.ptr; | |
331 | mo->it.last = mo->it.start + obj->base.size - 1; | |
332 | mo->obj = obj; | |
5cc9ed4b | 333 | |
ad46cb53 CW |
334 | ret = i915_mmu_notifier_add(obj->base.dev, mn, mo); |
335 | if (ret) { | |
336 | kfree(mo); | |
337 | return ret; | |
338 | } | |
339 | ||
340 | obj->userptr.mmu_object = mo; | |
5cc9ed4b | 341 | return 0; |
ad46cb53 CW |
342 | } |
343 | ||
344 | static void | |
345 | i915_mmu_notifier_free(struct i915_mmu_notifier *mn, | |
346 | struct mm_struct *mm) | |
347 | { | |
348 | if (mn == NULL) | |
349 | return; | |
5cc9ed4b | 350 | |
ad46cb53 | 351 | mmu_notifier_unregister(&mn->mn, mm); |
5cc9ed4b | 352 | kfree(mn); |
5cc9ed4b CW |
353 | } |
354 | ||
355 | #else | |
356 | ||
357 | static void | |
358 | i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj) | |
359 | { | |
360 | } | |
361 | ||
362 | static int | |
363 | i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj, | |
364 | unsigned flags) | |
365 | { | |
366 | if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0) | |
367 | return -ENODEV; | |
368 | ||
369 | if (!capable(CAP_SYS_ADMIN)) | |
370 | return -EPERM; | |
371 | ||
372 | return 0; | |
373 | } | |
ad46cb53 CW |
374 | |
375 | static void | |
376 | i915_mmu_notifier_free(struct i915_mmu_notifier *mn, | |
377 | struct mm_struct *mm) | |
378 | { | |
379 | } | |
380 | ||
5cc9ed4b CW |
381 | #endif |
382 | ||
ad46cb53 CW |
383 | static struct i915_mm_struct * |
384 | __i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real) | |
385 | { | |
386 | struct i915_mm_struct *mm; | |
387 | ||
388 | /* Protected by dev_priv->mm_lock */ | |
389 | hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real) | |
390 | if (mm->mm == real) | |
391 | return mm; | |
392 | ||
393 | return NULL; | |
394 | } | |
395 | ||
396 | static int | |
397 | i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj) | |
398 | { | |
399 | struct drm_i915_private *dev_priv = to_i915(obj->base.dev); | |
400 | struct i915_mm_struct *mm; | |
401 | int ret = 0; | |
402 | ||
403 | /* During release of the GEM object we hold the struct_mutex. This | |
404 | * precludes us from calling mmput() at that time as that may be | |
405 | * the last reference and so call exit_mmap(). exit_mmap() will | |
406 | * attempt to reap the vma, and if we were holding a GTT mmap | |
407 | * would then call drm_gem_vm_close() and attempt to reacquire | |
408 | * the struct mutex. So in order to avoid that recursion, we have | |
409 | * to defer releasing the mm reference until after we drop the | |
410 | * struct_mutex, i.e. we need to schedule a worker to do the clean | |
411 | * up. | |
412 | */ | |
413 | mutex_lock(&dev_priv->mm_lock); | |
414 | mm = __i915_mm_struct_find(dev_priv, current->mm); | |
415 | if (mm == NULL) { | |
416 | mm = kmalloc(sizeof(*mm), GFP_KERNEL); | |
417 | if (mm == NULL) { | |
418 | ret = -ENOMEM; | |
419 | goto out; | |
420 | } | |
421 | ||
422 | kref_init(&mm->kref); | |
423 | mm->dev = obj->base.dev; | |
424 | ||
425 | mm->mm = current->mm; | |
426 | atomic_inc(¤t->mm->mm_count); | |
427 | ||
428 | mm->mn = NULL; | |
429 | ||
430 | /* Protected by dev_priv->mm_lock */ | |
431 | hash_add(dev_priv->mm_structs, | |
432 | &mm->node, (unsigned long)mm->mm); | |
433 | } else | |
434 | kref_get(&mm->kref); | |
435 | ||
436 | obj->userptr.mm = mm; | |
437 | out: | |
438 | mutex_unlock(&dev_priv->mm_lock); | |
439 | return ret; | |
440 | } | |
441 | ||
442 | static void | |
443 | __i915_mm_struct_free__worker(struct work_struct *work) | |
444 | { | |
445 | struct i915_mm_struct *mm = container_of(work, typeof(*mm), work); | |
446 | i915_mmu_notifier_free(mm->mn, mm->mm); | |
447 | mmdrop(mm->mm); | |
448 | kfree(mm); | |
449 | } | |
450 | ||
451 | static void | |
452 | __i915_mm_struct_free(struct kref *kref) | |
453 | { | |
454 | struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref); | |
455 | ||
456 | /* Protected by dev_priv->mm_lock */ | |
457 | hash_del(&mm->node); | |
458 | mutex_unlock(&to_i915(mm->dev)->mm_lock); | |
459 | ||
460 | INIT_WORK(&mm->work, __i915_mm_struct_free__worker); | |
461 | schedule_work(&mm->work); | |
462 | } | |
463 | ||
464 | static void | |
465 | i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj) | |
466 | { | |
467 | if (obj->userptr.mm == NULL) | |
468 | return; | |
469 | ||
470 | kref_put_mutex(&obj->userptr.mm->kref, | |
471 | __i915_mm_struct_free, | |
472 | &to_i915(obj->base.dev)->mm_lock); | |
473 | obj->userptr.mm = NULL; | |
474 | } | |
475 | ||
5cc9ed4b CW |
476 | struct get_pages_work { |
477 | struct work_struct work; | |
478 | struct drm_i915_gem_object *obj; | |
479 | struct task_struct *task; | |
480 | }; | |
481 | ||
5cc9ed4b CW |
482 | #if IS_ENABLED(CONFIG_SWIOTLB) |
483 | #define swiotlb_active() swiotlb_nr_tbl() | |
484 | #else | |
485 | #define swiotlb_active() 0 | |
486 | #endif | |
487 | ||
488 | static int | |
489 | st_set_pages(struct sg_table **st, struct page **pvec, int num_pages) | |
490 | { | |
491 | struct scatterlist *sg; | |
492 | int ret, n; | |
493 | ||
494 | *st = kmalloc(sizeof(**st), GFP_KERNEL); | |
495 | if (*st == NULL) | |
496 | return -ENOMEM; | |
497 | ||
498 | if (swiotlb_active()) { | |
499 | ret = sg_alloc_table(*st, num_pages, GFP_KERNEL); | |
500 | if (ret) | |
501 | goto err; | |
502 | ||
503 | for_each_sg((*st)->sgl, sg, num_pages, n) | |
504 | sg_set_page(sg, pvec[n], PAGE_SIZE, 0); | |
505 | } else { | |
506 | ret = sg_alloc_table_from_pages(*st, pvec, num_pages, | |
507 | 0, num_pages << PAGE_SHIFT, | |
508 | GFP_KERNEL); | |
509 | if (ret) | |
510 | goto err; | |
511 | } | |
512 | ||
513 | return 0; | |
514 | ||
515 | err: | |
516 | kfree(*st); | |
517 | *st = NULL; | |
518 | return ret; | |
519 | } | |
520 | ||
521 | static void | |
522 | __i915_gem_userptr_get_pages_worker(struct work_struct *_work) | |
523 | { | |
524 | struct get_pages_work *work = container_of(_work, typeof(*work), work); | |
525 | struct drm_i915_gem_object *obj = work->obj; | |
526 | struct drm_device *dev = obj->base.dev; | |
527 | const int num_pages = obj->base.size >> PAGE_SHIFT; | |
528 | struct page **pvec; | |
529 | int pinned, ret; | |
530 | ||
531 | ret = -ENOMEM; | |
532 | pinned = 0; | |
533 | ||
534 | pvec = kmalloc(num_pages*sizeof(struct page *), | |
535 | GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY); | |
536 | if (pvec == NULL) | |
537 | pvec = drm_malloc_ab(num_pages, sizeof(struct page *)); | |
538 | if (pvec != NULL) { | |
ad46cb53 | 539 | struct mm_struct *mm = obj->userptr.mm->mm; |
5cc9ed4b CW |
540 | |
541 | down_read(&mm->mmap_sem); | |
542 | while (pinned < num_pages) { | |
543 | ret = get_user_pages(work->task, mm, | |
544 | obj->userptr.ptr + pinned * PAGE_SIZE, | |
545 | num_pages - pinned, | |
546 | !obj->userptr.read_only, 0, | |
547 | pvec + pinned, NULL); | |
548 | if (ret < 0) | |
549 | break; | |
550 | ||
551 | pinned += ret; | |
552 | } | |
553 | up_read(&mm->mmap_sem); | |
554 | } | |
555 | ||
556 | mutex_lock(&dev->struct_mutex); | |
557 | if (obj->userptr.work != &work->work) { | |
558 | ret = 0; | |
559 | } else if (pinned == num_pages) { | |
560 | ret = st_set_pages(&obj->pages, pvec, num_pages); | |
561 | if (ret == 0) { | |
562 | list_add_tail(&obj->global_list, &to_i915(dev)->mm.unbound_list); | |
563 | pinned = 0; | |
564 | } | |
565 | } | |
566 | ||
567 | obj->userptr.work = ERR_PTR(ret); | |
568 | obj->userptr.workers--; | |
569 | drm_gem_object_unreference(&obj->base); | |
570 | mutex_unlock(&dev->struct_mutex); | |
571 | ||
572 | release_pages(pvec, pinned, 0); | |
573 | drm_free_large(pvec); | |
574 | ||
575 | put_task_struct(work->task); | |
576 | kfree(work); | |
577 | } | |
578 | ||
579 | static int | |
580 | i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj) | |
581 | { | |
582 | const int num_pages = obj->base.size >> PAGE_SHIFT; | |
583 | struct page **pvec; | |
584 | int pinned, ret; | |
585 | ||
586 | /* If userspace should engineer that these pages are replaced in | |
587 | * the vma between us binding this page into the GTT and completion | |
588 | * of rendering... Their loss. If they change the mapping of their | |
589 | * pages they need to create a new bo to point to the new vma. | |
590 | * | |
591 | * However, that still leaves open the possibility of the vma | |
592 | * being copied upon fork. Which falls under the same userspace | |
593 | * synchronisation issue as a regular bo, except that this time | |
594 | * the process may not be expecting that a particular piece of | |
595 | * memory is tied to the GPU. | |
596 | * | |
597 | * Fortunately, we can hook into the mmu_notifier in order to | |
598 | * discard the page references prior to anything nasty happening | |
599 | * to the vma (discard or cloning) which should prevent the more | |
600 | * egregious cases from causing harm. | |
601 | */ | |
602 | ||
603 | pvec = NULL; | |
604 | pinned = 0; | |
ad46cb53 | 605 | if (obj->userptr.mm->mm == current->mm) { |
5cc9ed4b CW |
606 | pvec = kmalloc(num_pages*sizeof(struct page *), |
607 | GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY); | |
608 | if (pvec == NULL) { | |
609 | pvec = drm_malloc_ab(num_pages, sizeof(struct page *)); | |
610 | if (pvec == NULL) | |
611 | return -ENOMEM; | |
612 | } | |
613 | ||
614 | pinned = __get_user_pages_fast(obj->userptr.ptr, num_pages, | |
615 | !obj->userptr.read_only, pvec); | |
616 | } | |
617 | if (pinned < num_pages) { | |
618 | if (pinned < 0) { | |
619 | ret = pinned; | |
620 | pinned = 0; | |
621 | } else { | |
622 | /* Spawn a worker so that we can acquire the | |
623 | * user pages without holding our mutex. Access | |
624 | * to the user pages requires mmap_sem, and we have | |
625 | * a strict lock ordering of mmap_sem, struct_mutex - | |
626 | * we already hold struct_mutex here and so cannot | |
627 | * call gup without encountering a lock inversion. | |
628 | * | |
629 | * Userspace will keep on repeating the operation | |
630 | * (thanks to EAGAIN) until either we hit the fast | |
631 | * path or the worker completes. If the worker is | |
632 | * cancelled or superseded, the task is still run | |
633 | * but the results ignored. (This leads to | |
634 | * complications that we may have a stray object | |
635 | * refcount that we need to be wary of when | |
636 | * checking for existing objects during creation.) | |
637 | * If the worker encounters an error, it reports | |
638 | * that error back to this function through | |
639 | * obj->userptr.work = ERR_PTR. | |
640 | */ | |
641 | ret = -EAGAIN; | |
642 | if (obj->userptr.work == NULL && | |
643 | obj->userptr.workers < I915_GEM_USERPTR_MAX_WORKERS) { | |
644 | struct get_pages_work *work; | |
645 | ||
646 | work = kmalloc(sizeof(*work), GFP_KERNEL); | |
647 | if (work != NULL) { | |
648 | obj->userptr.work = &work->work; | |
649 | obj->userptr.workers++; | |
650 | ||
651 | work->obj = obj; | |
652 | drm_gem_object_reference(&obj->base); | |
653 | ||
654 | work->task = current; | |
655 | get_task_struct(work->task); | |
656 | ||
657 | INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker); | |
658 | schedule_work(&work->work); | |
659 | } else | |
660 | ret = -ENOMEM; | |
661 | } else { | |
662 | if (IS_ERR(obj->userptr.work)) { | |
663 | ret = PTR_ERR(obj->userptr.work); | |
664 | obj->userptr.work = NULL; | |
665 | } | |
666 | } | |
667 | } | |
668 | } else { | |
669 | ret = st_set_pages(&obj->pages, pvec, num_pages); | |
670 | if (ret == 0) { | |
671 | obj->userptr.work = NULL; | |
672 | pinned = 0; | |
673 | } | |
674 | } | |
675 | ||
676 | release_pages(pvec, pinned, 0); | |
677 | drm_free_large(pvec); | |
678 | return ret; | |
679 | } | |
680 | ||
681 | static void | |
682 | i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj) | |
683 | { | |
684 | struct scatterlist *sg; | |
685 | int i; | |
686 | ||
687 | BUG_ON(obj->userptr.work != NULL); | |
688 | ||
689 | if (obj->madv != I915_MADV_WILLNEED) | |
690 | obj->dirty = 0; | |
691 | ||
692 | for_each_sg(obj->pages->sgl, sg, obj->pages->nents, i) { | |
693 | struct page *page = sg_page(sg); | |
694 | ||
695 | if (obj->dirty) | |
696 | set_page_dirty(page); | |
697 | ||
698 | mark_page_accessed(page); | |
699 | page_cache_release(page); | |
700 | } | |
701 | obj->dirty = 0; | |
702 | ||
703 | sg_free_table(obj->pages); | |
704 | kfree(obj->pages); | |
705 | } | |
706 | ||
707 | static void | |
708 | i915_gem_userptr_release(struct drm_i915_gem_object *obj) | |
709 | { | |
710 | i915_gem_userptr_release__mmu_notifier(obj); | |
ad46cb53 | 711 | i915_gem_userptr_release__mm_struct(obj); |
5cc9ed4b CW |
712 | } |
713 | ||
714 | static int | |
715 | i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj) | |
716 | { | |
ad46cb53 | 717 | if (obj->userptr.mmu_object) |
5cc9ed4b CW |
718 | return 0; |
719 | ||
720 | return i915_gem_userptr_init__mmu_notifier(obj, 0); | |
721 | } | |
722 | ||
723 | static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = { | |
724 | .dmabuf_export = i915_gem_userptr_dmabuf_export, | |
725 | .get_pages = i915_gem_userptr_get_pages, | |
726 | .put_pages = i915_gem_userptr_put_pages, | |
727 | .release = i915_gem_userptr_release, | |
728 | }; | |
729 | ||
730 | /** | |
731 | * Creates a new mm object that wraps some normal memory from the process | |
732 | * context - user memory. | |
733 | * | |
734 | * We impose several restrictions upon the memory being mapped | |
735 | * into the GPU. | |
736 | * 1. It must be page aligned (both start/end addresses, i.e ptr and size). | |
ec8b0dd5 | 737 | * 2. It must be normal system memory, not a pointer into another map of IO |
5cc9ed4b | 738 | * space (e.g. it must not be a GTT mmapping of another object). |
ec8b0dd5 | 739 | * 3. We only allow a bo as large as we could in theory map into the GTT, |
5cc9ed4b | 740 | * that is we limit the size to the total size of the GTT. |
ec8b0dd5 | 741 | * 4. The bo is marked as being snoopable. The backing pages are left |
5cc9ed4b CW |
742 | * accessible directly by the CPU, but reads and writes by the GPU may |
743 | * incur the cost of a snoop (unless you have an LLC architecture). | |
744 | * | |
745 | * Synchronisation between multiple users and the GPU is left to userspace | |
746 | * through the normal set-domain-ioctl. The kernel will enforce that the | |
747 | * GPU relinquishes the VMA before it is returned back to the system | |
748 | * i.e. upon free(), munmap() or process termination. However, the userspace | |
749 | * malloc() library may not immediately relinquish the VMA after free() and | |
750 | * instead reuse it whilst the GPU is still reading and writing to the VMA. | |
751 | * Caveat emptor. | |
752 | * | |
753 | * Also note, that the object created here is not currently a "first class" | |
754 | * object, in that several ioctls are banned. These are the CPU access | |
755 | * ioctls: mmap(), pwrite and pread. In practice, you are expected to use | |
756 | * direct access via your pointer rather than use those ioctls. | |
757 | * | |
758 | * If you think this is a good interface to use to pass GPU memory between | |
759 | * drivers, please use dma-buf instead. In fact, wherever possible use | |
760 | * dma-buf instead. | |
761 | */ | |
762 | int | |
763 | i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file) | |
764 | { | |
765 | struct drm_i915_private *dev_priv = dev->dev_private; | |
766 | struct drm_i915_gem_userptr *args = data; | |
767 | struct drm_i915_gem_object *obj; | |
768 | int ret; | |
769 | u32 handle; | |
770 | ||
771 | if (args->flags & ~(I915_USERPTR_READ_ONLY | | |
772 | I915_USERPTR_UNSYNCHRONIZED)) | |
773 | return -EINVAL; | |
774 | ||
775 | if (offset_in_page(args->user_ptr | args->user_size)) | |
776 | return -EINVAL; | |
777 | ||
778 | if (args->user_size > dev_priv->gtt.base.total) | |
779 | return -E2BIG; | |
780 | ||
781 | if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE, | |
782 | (char __user *)(unsigned long)args->user_ptr, args->user_size)) | |
783 | return -EFAULT; | |
784 | ||
785 | if (args->flags & I915_USERPTR_READ_ONLY) { | |
786 | /* On almost all of the current hw, we cannot tell the GPU that a | |
787 | * page is readonly, so this is just a placeholder in the uAPI. | |
788 | */ | |
789 | return -ENODEV; | |
790 | } | |
791 | ||
5cc9ed4b CW |
792 | obj = i915_gem_object_alloc(dev); |
793 | if (obj == NULL) | |
794 | return -ENOMEM; | |
795 | ||
796 | drm_gem_private_object_init(dev, &obj->base, args->user_size); | |
797 | i915_gem_object_init(obj, &i915_gem_userptr_ops); | |
798 | obj->cache_level = I915_CACHE_LLC; | |
799 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
800 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
801 | ||
802 | obj->userptr.ptr = args->user_ptr; | |
803 | obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY); | |
804 | ||
805 | /* And keep a pointer to the current->mm for resolving the user pages | |
806 | * at binding. This means that we need to hook into the mmu_notifier | |
807 | * in order to detect if the mmu is destroyed. | |
808 | */ | |
ad46cb53 CW |
809 | ret = i915_gem_userptr_init__mm_struct(obj); |
810 | if (ret == 0) | |
5cc9ed4b CW |
811 | ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags); |
812 | if (ret == 0) | |
813 | ret = drm_gem_handle_create(file, &obj->base, &handle); | |
814 | ||
815 | /* drop reference from allocate - handle holds it now */ | |
816 | drm_gem_object_unreference_unlocked(&obj->base); | |
817 | if (ret) | |
818 | return ret; | |
819 | ||
820 | args->handle = handle; | |
821 | return 0; | |
822 | } | |
823 | ||
824 | int | |
825 | i915_gem_init_userptr(struct drm_device *dev) | |
826 | { | |
5cc9ed4b | 827 | struct drm_i915_private *dev_priv = to_i915(dev); |
ad46cb53 CW |
828 | mutex_init(&dev_priv->mm_lock); |
829 | hash_init(dev_priv->mm_structs); | |
5cc9ed4b CW |
830 | return 0; |
831 | } |