2 * Copyright 2014 Advanced Micro Devices, Inc.
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:
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.
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
31 #include <linux/firmware.h>
32 #include <linux/module.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/interval_tree.h>
41 /* constant after initialisation */
42 struct amdgpu_device *adev;
44 struct mmu_notifier mn;
46 /* only used on destruction */
47 struct work_struct work;
49 /* protected by adev->mn_lock */
50 struct hlist_node node;
52 /* objects protected by lock */
53 struct rw_semaphore lock;
54 struct rb_root_cached objects;
55 struct mutex read_lock;
59 struct amdgpu_mn_node {
60 struct interval_tree_node it;
65 * amdgpu_mn_destroy - destroy the rmn
67 * @work: previously sheduled work item
69 * Lazy destroys the notifier from a work item
71 static void amdgpu_mn_destroy(struct work_struct *work)
73 struct amdgpu_mn *rmn = container_of(work, struct amdgpu_mn, work);
74 struct amdgpu_device *adev = rmn->adev;
75 struct amdgpu_mn_node *node, *next_node;
76 struct amdgpu_bo *bo, *next_bo;
78 mutex_lock(&adev->mn_lock);
79 down_write(&rmn->lock);
81 rbtree_postorder_for_each_entry_safe(node, next_node,
82 &rmn->objects.rb_root, it.rb) {
83 list_for_each_entry_safe(bo, next_bo, &node->bos, mn_list) {
85 list_del_init(&bo->mn_list);
90 mutex_unlock(&adev->mn_lock);
91 mmu_notifier_unregister_no_release(&rmn->mn, rmn->mm);
96 * amdgpu_mn_release - callback to notify about mm destruction
99 * @mn: the mm this callback is about
101 * Shedule a work item to lazy destroy our notifier.
103 static void amdgpu_mn_release(struct mmu_notifier *mn,
104 struct mm_struct *mm)
106 struct amdgpu_mn *rmn = container_of(mn, struct amdgpu_mn, mn);
107 INIT_WORK(&rmn->work, amdgpu_mn_destroy);
108 schedule_work(&rmn->work);
113 * amdgpu_mn_lock - take the write side lock for this mn
115 void amdgpu_mn_lock(struct amdgpu_mn *mn)
118 down_write(&mn->lock);
122 * amdgpu_mn_unlock - drop the write side lock for this mn
124 void amdgpu_mn_unlock(struct amdgpu_mn *mn)
131 * amdgpu_mn_read_lock - take the rmn read lock
135 * Take the rmn read side lock.
137 static void amdgpu_mn_read_lock(struct amdgpu_mn *rmn)
139 mutex_lock(&rmn->read_lock);
140 if (atomic_inc_return(&rmn->recursion) == 1)
141 down_read_non_owner(&rmn->lock);
142 mutex_unlock(&rmn->read_lock);
146 * amdgpu_mn_read_unlock - drop the rmn read lock
150 * Drop the rmn read side lock.
152 static void amdgpu_mn_read_unlock(struct amdgpu_mn *rmn)
154 if (atomic_dec_return(&rmn->recursion) == 0)
155 up_read_non_owner(&rmn->lock);
159 * amdgpu_mn_invalidate_node - unmap all BOs of a node
161 * @node: the node with the BOs to unmap
163 * We block for all BOs and unmap them by move them
164 * into system domain again.
166 static void amdgpu_mn_invalidate_node(struct amdgpu_mn_node *node,
170 struct amdgpu_bo *bo;
173 list_for_each_entry(bo, &node->bos, mn_list) {
175 if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm, start, end))
178 r = reservation_object_wait_timeout_rcu(bo->tbo.resv,
179 true, false, MAX_SCHEDULE_TIMEOUT);
181 DRM_ERROR("(%ld) failed to wait for user bo\n", r);
183 amdgpu_ttm_tt_mark_user_pages(bo->tbo.ttm);
188 * amdgpu_mn_invalidate_range_start - callback to notify about mm change
191 * @mn: the mm this callback is about
192 * @start: start of updated range
193 * @end: end of updated range
195 * We block for all BOs between start and end to be idle and
196 * unmap them by move them into system domain again.
198 static void amdgpu_mn_invalidate_range_start(struct mmu_notifier *mn,
199 struct mm_struct *mm,
203 struct amdgpu_mn *rmn = container_of(mn, struct amdgpu_mn, mn);
204 struct interval_tree_node *it;
206 /* notification is exclusive, but interval is inclusive */
209 amdgpu_mn_read_lock(rmn);
211 it = interval_tree_iter_first(&rmn->objects, start, end);
213 struct amdgpu_mn_node *node;
215 node = container_of(it, struct amdgpu_mn_node, it);
216 it = interval_tree_iter_next(it, start, end);
218 amdgpu_mn_invalidate_node(node, start, end);
223 * amdgpu_mn_invalidate_range_end - callback to notify about mm change
226 * @mn: the mm this callback is about
227 * @start: start of updated range
228 * @end: end of updated range
230 * Release the lock again to allow new command submissions.
232 static void amdgpu_mn_invalidate_range_end(struct mmu_notifier *mn,
233 struct mm_struct *mm,
237 struct amdgpu_mn *rmn = container_of(mn, struct amdgpu_mn, mn);
239 amdgpu_mn_read_unlock(rmn);
242 static const struct mmu_notifier_ops amdgpu_mn_ops = {
243 .release = amdgpu_mn_release,
244 .invalidate_range_start = amdgpu_mn_invalidate_range_start,
245 .invalidate_range_end = amdgpu_mn_invalidate_range_end,
249 * amdgpu_mn_get - create notifier context
251 * @adev: amdgpu device pointer
253 * Creates a notifier context for current->mm.
255 struct amdgpu_mn *amdgpu_mn_get(struct amdgpu_device *adev)
257 struct mm_struct *mm = current->mm;
258 struct amdgpu_mn *rmn;
261 mutex_lock(&adev->mn_lock);
262 if (down_write_killable(&mm->mmap_sem)) {
263 mutex_unlock(&adev->mn_lock);
264 return ERR_PTR(-EINTR);
267 hash_for_each_possible(adev->mn_hash, rmn, node, (unsigned long)mm)
271 rmn = kzalloc(sizeof(*rmn), GFP_KERNEL);
273 rmn = ERR_PTR(-ENOMEM);
279 rmn->mn.ops = &amdgpu_mn_ops;
280 init_rwsem(&rmn->lock);
281 rmn->objects = RB_ROOT_CACHED;
282 mutex_init(&rmn->read_lock);
283 atomic_set(&rmn->recursion, 0);
285 r = __mmu_notifier_register(&rmn->mn, mm);
289 hash_add(adev->mn_hash, &rmn->node, (unsigned long)mm);
292 up_write(&mm->mmap_sem);
293 mutex_unlock(&adev->mn_lock);
298 up_write(&mm->mmap_sem);
299 mutex_unlock(&adev->mn_lock);
306 * amdgpu_mn_register - register a BO for notifier updates
308 * @bo: amdgpu buffer object
309 * @addr: userptr addr we should monitor
311 * Registers an MMU notifier for the given BO at the specified address.
312 * Returns 0 on success, -ERRNO if anything goes wrong.
314 int amdgpu_mn_register(struct amdgpu_bo *bo, unsigned long addr)
316 unsigned long end = addr + amdgpu_bo_size(bo) - 1;
317 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
318 struct amdgpu_mn *rmn;
319 struct amdgpu_mn_node *node = NULL;
320 struct list_head bos;
321 struct interval_tree_node *it;
323 rmn = amdgpu_mn_get(adev);
327 INIT_LIST_HEAD(&bos);
329 down_write(&rmn->lock);
331 while ((it = interval_tree_iter_first(&rmn->objects, addr, end))) {
333 node = container_of(it, struct amdgpu_mn_node, it);
334 interval_tree_remove(&node->it, &rmn->objects);
335 addr = min(it->start, addr);
336 end = max(it->last, end);
337 list_splice(&node->bos, &bos);
341 node = kmalloc(sizeof(struct amdgpu_mn_node), GFP_KERNEL);
343 up_write(&rmn->lock);
350 node->it.start = addr;
352 INIT_LIST_HEAD(&node->bos);
353 list_splice(&bos, &node->bos);
354 list_add(&bo->mn_list, &node->bos);
356 interval_tree_insert(&node->it, &rmn->objects);
358 up_write(&rmn->lock);
364 * amdgpu_mn_unregister - unregister a BO for notifier updates
366 * @bo: amdgpu buffer object
368 * Remove any registration of MMU notifier updates from the buffer object.
370 void amdgpu_mn_unregister(struct amdgpu_bo *bo)
372 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
373 struct amdgpu_mn *rmn;
374 struct list_head *head;
376 mutex_lock(&adev->mn_lock);
380 mutex_unlock(&adev->mn_lock);
384 down_write(&rmn->lock);
386 /* save the next list entry for later */
387 head = bo->mn_list.next;
390 list_del_init(&bo->mn_list);
392 if (list_empty(head)) {
393 struct amdgpu_mn_node *node;
394 node = container_of(head, struct amdgpu_mn_node, bos);
395 interval_tree_remove(&node->it, &rmn->objects);
399 up_write(&rmn->lock);
400 mutex_unlock(&adev->mn_lock);