1 /*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
12 *
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
20 *
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
23 * of the Software.
24 *
25 */
26 /*
27 * Authors:
28 * Christian König <christian.koenig@amd.com>
29 */
30
31 #include <linux/firmware.h>
32 #include <linux/module.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/interval_tree.h>
35 #include <drm/drmP.h>
36 #include <drm/drm.h>
37
38 #include "amdgpu.h"
39
40 struct amdgpu_mn {
41 /* constant after initialisation */
42 struct amdgpu_device *adev;
43 struct mm_struct *mm;
44 struct mmu_notifier mn;
45
46 /* only used on destruction */
47 struct work_struct work;
48
49 /* protected by adev->mn_lock */
50 struct hlist_node node;
51
52 /* objects protected by lock */
53 struct mutex lock;
54 struct rb_root_cached objects;
55 };
56
57 struct amdgpu_mn_node {
58 struct interval_tree_node it;
59 struct list_head bos;
60 };
61
62 /**
63 * amdgpu_mn_destroy - destroy the rmn
64 *
65 * @work: previously sheduled work item
66 *
67 * Lazy destroys the notifier from a work item
68 */
amdgpu_mn_destroy(struct work_struct * work)69 static void amdgpu_mn_destroy(struct work_struct *work)
70 {
71 struct amdgpu_mn *rmn = container_of(work, struct amdgpu_mn, work);
72 struct amdgpu_device *adev = rmn->adev;
73 struct amdgpu_mn_node *node, *next_node;
74 struct amdgpu_bo *bo, *next_bo;
75
76 mutex_lock(&adev->mn_lock);
77 mutex_lock(&rmn->lock);
78 hash_del(&rmn->node);
79 rbtree_postorder_for_each_entry_safe(node, next_node,
80 &rmn->objects.rb_root, it.rb) {
81 list_for_each_entry_safe(bo, next_bo, &node->bos, mn_list) {
82 bo->mn = NULL;
83 list_del_init(&bo->mn_list);
84 }
85 kfree(node);
86 }
87 mutex_unlock(&rmn->lock);
88 mutex_unlock(&adev->mn_lock);
89 mmu_notifier_unregister_no_release(&rmn->mn, rmn->mm);
90 kfree(rmn);
91 }
92
93 /**
94 * amdgpu_mn_release - callback to notify about mm destruction
95 *
96 * @mn: our notifier
97 * @mn: the mm this callback is about
98 *
99 * Shedule a work item to lazy destroy our notifier.
100 */
amdgpu_mn_release(struct mmu_notifier * mn,struct mm_struct * mm)101 static void amdgpu_mn_release(struct mmu_notifier *mn,
102 struct mm_struct *mm)
103 {
104 struct amdgpu_mn *rmn = container_of(mn, struct amdgpu_mn, mn);
105 INIT_WORK(&rmn->work, amdgpu_mn_destroy);
106 schedule_work(&rmn->work);
107 }
108
109 /**
110 * amdgpu_mn_invalidate_node - unmap all BOs of a node
111 *
112 * @node: the node with the BOs to unmap
113 *
114 * We block for all BOs and unmap them by move them
115 * into system domain again.
116 */
amdgpu_mn_invalidate_node(struct amdgpu_mn_node * node,unsigned long start,unsigned long end)117 static void amdgpu_mn_invalidate_node(struct amdgpu_mn_node *node,
118 unsigned long start,
119 unsigned long end)
120 {
121 struct amdgpu_bo *bo;
122 long r;
123
124 list_for_each_entry(bo, &node->bos, mn_list) {
125
126 if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm, start, end))
127 continue;
128
129 r = amdgpu_bo_reserve(bo, true);
130 if (r) {
131 DRM_ERROR("(%ld) failed to reserve user bo\n", r);
132 continue;
133 }
134
135 r = reservation_object_wait_timeout_rcu(bo->tbo.resv,
136 true, false, MAX_SCHEDULE_TIMEOUT);
137 if (r <= 0)
138 DRM_ERROR("(%ld) failed to wait for user bo\n", r);
139
140 amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
141 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
142 if (r)
143 DRM_ERROR("(%ld) failed to validate user bo\n", r);
144
145 amdgpu_bo_unreserve(bo);
146 }
147 }
148
149 /**
150 * amdgpu_mn_invalidate_range_start - callback to notify about mm change
151 *
152 * @mn: our notifier
153 * @mn: the mm this callback is about
154 * @start: start of updated range
155 * @end: end of updated range
156 *
157 * We block for all BOs between start and end to be idle and
158 * unmap them by move them into system domain again.
159 */
amdgpu_mn_invalidate_range_start(struct mmu_notifier * mn,struct mm_struct * mm,unsigned long start,unsigned long end)160 static void amdgpu_mn_invalidate_range_start(struct mmu_notifier *mn,
161 struct mm_struct *mm,
162 unsigned long start,
163 unsigned long end)
164 {
165 struct amdgpu_mn *rmn = container_of(mn, struct amdgpu_mn, mn);
166 struct interval_tree_node *it;
167
168 /* notification is exclusive, but interval is inclusive */
169 end -= 1;
170
171 mutex_lock(&rmn->lock);
172
173 it = interval_tree_iter_first(&rmn->objects, start, end);
174 while (it) {
175 struct amdgpu_mn_node *node;
176
177 node = container_of(it, struct amdgpu_mn_node, it);
178 it = interval_tree_iter_next(it, start, end);
179
180 amdgpu_mn_invalidate_node(node, start, end);
181 }
182
183 mutex_unlock(&rmn->lock);
184 }
185
186 static const struct mmu_notifier_ops amdgpu_mn_ops = {
187 .release = amdgpu_mn_release,
188 .invalidate_range_start = amdgpu_mn_invalidate_range_start,
189 };
190
191 /**
192 * amdgpu_mn_get - create notifier context
193 *
194 * @adev: amdgpu device pointer
195 *
196 * Creates a notifier context for current->mm.
197 */
amdgpu_mn_get(struct amdgpu_device * adev)198 static struct amdgpu_mn *amdgpu_mn_get(struct amdgpu_device *adev)
199 {
200 struct mm_struct *mm = current->mm;
201 struct amdgpu_mn *rmn;
202 int r;
203
204 mutex_lock(&adev->mn_lock);
205 if (down_write_killable(&mm->mmap_sem)) {
206 mutex_unlock(&adev->mn_lock);
207 return ERR_PTR(-EINTR);
208 }
209
210 hash_for_each_possible(adev->mn_hash, rmn, node, (unsigned long)mm)
211 if (rmn->mm == mm)
212 goto release_locks;
213
214 rmn = kzalloc(sizeof(*rmn), GFP_KERNEL);
215 if (!rmn) {
216 rmn = ERR_PTR(-ENOMEM);
217 goto release_locks;
218 }
219
220 rmn->adev = adev;
221 rmn->mm = mm;
222 rmn->mn.ops = &amdgpu_mn_ops;
223 mutex_init(&rmn->lock);
224 rmn->objects = RB_ROOT_CACHED;
225
226 r = __mmu_notifier_register(&rmn->mn, mm);
227 if (r)
228 goto free_rmn;
229
230 hash_add(adev->mn_hash, &rmn->node, (unsigned long)mm);
231
232 release_locks:
233 up_write(&mm->mmap_sem);
234 mutex_unlock(&adev->mn_lock);
235
236 return rmn;
237
238 free_rmn:
239 up_write(&mm->mmap_sem);
240 mutex_unlock(&adev->mn_lock);
241 kfree(rmn);
242
243 return ERR_PTR(r);
244 }
245
246 /**
247 * amdgpu_mn_register - register a BO for notifier updates
248 *
249 * @bo: amdgpu buffer object
250 * @addr: userptr addr we should monitor
251 *
252 * Registers an MMU notifier for the given BO at the specified address.
253 * Returns 0 on success, -ERRNO if anything goes wrong.
254 */
amdgpu_mn_register(struct amdgpu_bo * bo,unsigned long addr)255 int amdgpu_mn_register(struct amdgpu_bo *bo, unsigned long addr)
256 {
257 unsigned long end = addr + amdgpu_bo_size(bo) - 1;
258 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
259 struct amdgpu_mn *rmn;
260 struct amdgpu_mn_node *node = NULL;
261 struct list_head bos;
262 struct interval_tree_node *it;
263
264 rmn = amdgpu_mn_get(adev);
265 if (IS_ERR(rmn))
266 return PTR_ERR(rmn);
267
268 INIT_LIST_HEAD(&bos);
269
270 mutex_lock(&rmn->lock);
271
272 while ((it = interval_tree_iter_first(&rmn->objects, addr, end))) {
273 kfree(node);
274 node = container_of(it, struct amdgpu_mn_node, it);
275 interval_tree_remove(&node->it, &rmn->objects);
276 addr = min(it->start, addr);
277 end = max(it->last, end);
278 list_splice(&node->bos, &bos);
279 }
280
281 if (!node) {
282 node = kmalloc(sizeof(struct amdgpu_mn_node), GFP_KERNEL);
283 if (!node) {
284 mutex_unlock(&rmn->lock);
285 return -ENOMEM;
286 }
287 }
288
289 bo->mn = rmn;
290
291 node->it.start = addr;
292 node->it.last = end;
293 INIT_LIST_HEAD(&node->bos);
294 list_splice(&bos, &node->bos);
295 list_add(&bo->mn_list, &node->bos);
296
297 interval_tree_insert(&node->it, &rmn->objects);
298
299 mutex_unlock(&rmn->lock);
300
301 return 0;
302 }
303
304 /**
305 * amdgpu_mn_unregister - unregister a BO for notifier updates
306 *
307 * @bo: amdgpu buffer object
308 *
309 * Remove any registration of MMU notifier updates from the buffer object.
310 */
amdgpu_mn_unregister(struct amdgpu_bo * bo)311 void amdgpu_mn_unregister(struct amdgpu_bo *bo)
312 {
313 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
314 struct amdgpu_mn *rmn;
315 struct list_head *head;
316
317 mutex_lock(&adev->mn_lock);
318
319 rmn = bo->mn;
320 if (rmn == NULL) {
321 mutex_unlock(&adev->mn_lock);
322 return;
323 }
324
325 mutex_lock(&rmn->lock);
326
327 /* save the next list entry for later */
328 head = bo->mn_list.next;
329
330 bo->mn = NULL;
331 list_del_init(&bo->mn_list);
332
333 if (list_empty(head)) {
334 struct amdgpu_mn_node *node;
335 node = container_of(head, struct amdgpu_mn_node, bos);
336 interval_tree_remove(&node->it, &rmn->objects);
337 kfree(node);
338 }
339
340 mutex_unlock(&rmn->lock);
341 mutex_unlock(&adev->mn_lock);
342 }
343