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>
40 /* constant after initialisation */
41 struct amdgpu_device *adev;
43 struct mmu_notifier mn;
45 /* only used on destruction */
46 struct work_struct work;
48 /* protected by adev->mn_lock */
49 struct hlist_node node;
51 /* objects protected by lock */
53 struct rb_root objects;
56 struct amdgpu_mn_node {
57 struct interval_tree_node it;
62 * amdgpu_mn_destroy - destroy the rmn
64 * @work: previously sheduled work item
66 * Lazy destroys the notifier from a work item
68 static void amdgpu_mn_destroy(struct work_struct *work)
70 struct amdgpu_mn *rmn = container_of(work, struct amdgpu_mn, work);
71 struct amdgpu_device *adev = rmn->adev;
72 struct amdgpu_mn_node *node, *next_node;
73 struct amdgpu_bo *bo, *next_bo;
75 mutex_lock(&adev->mn_lock);
76 mutex_lock(&rmn->lock);
78 rbtree_postorder_for_each_entry_safe(node, next_node, &rmn->objects,
80 list_for_each_entry_safe(bo, next_bo, &node->bos, mn_list) {
82 list_del_init(&bo->mn_list);
86 mutex_unlock(&rmn->lock);
87 mutex_unlock(&adev->mn_lock);
88 mmu_notifier_unregister_no_release(&rmn->mn, rmn->mm);
93 * amdgpu_mn_release - callback to notify about mm destruction
96 * @mn: the mm this callback is about
98 * Shedule a work item to lazy destroy our notifier.
100 static void amdgpu_mn_release(struct mmu_notifier *mn,
101 struct mm_struct *mm)
103 struct amdgpu_mn *rmn = container_of(mn, struct amdgpu_mn, mn);
104 INIT_WORK(&rmn->work, amdgpu_mn_destroy);
105 schedule_work(&rmn->work);
109 * amdgpu_mn_invalidate_node - unmap all BOs of a node
111 * @node: the node with the BOs to unmap
113 * We block for all BOs and unmap them by move them
114 * into system domain again.
116 static void amdgpu_mn_invalidate_node(struct amdgpu_mn_node *node,
120 struct amdgpu_bo *bo;
123 list_for_each_entry(bo, &node->bos, mn_list) {
125 if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm, start, end))
128 r = amdgpu_bo_reserve(bo, true);
130 DRM_ERROR("(%ld) failed to reserve user bo\n", r);
134 r = reservation_object_wait_timeout_rcu(bo->tbo.resv,
135 true, false, MAX_SCHEDULE_TIMEOUT);
137 DRM_ERROR("(%ld) failed to wait for user bo\n", r);
139 amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
140 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
142 DRM_ERROR("(%ld) failed to validate user bo\n", r);
144 amdgpu_bo_unreserve(bo);
149 * amdgpu_mn_invalidate_page - callback to notify about mm change
152 * @mn: the mm this callback is about
153 * @address: address of invalidate page
155 * Invalidation of a single page. Blocks for all BOs mapping it
156 * and unmap them by move them into system domain again.
158 static void amdgpu_mn_invalidate_page(struct mmu_notifier *mn,
159 struct mm_struct *mm,
160 unsigned long address)
162 struct amdgpu_mn *rmn = container_of(mn, struct amdgpu_mn, mn);
163 struct interval_tree_node *it;
165 mutex_lock(&rmn->lock);
167 it = interval_tree_iter_first(&rmn->objects, address, address);
169 struct amdgpu_mn_node *node;
171 node = container_of(it, struct amdgpu_mn_node, it);
172 amdgpu_mn_invalidate_node(node, address, address);
175 mutex_unlock(&rmn->lock);
179 * amdgpu_mn_invalidate_range_start - callback to notify about mm change
182 * @mn: the mm this callback is about
183 * @start: start of updated range
184 * @end: end of updated range
186 * We block for all BOs between start and end to be idle and
187 * unmap them by move them into system domain again.
189 static void amdgpu_mn_invalidate_range_start(struct mmu_notifier *mn,
190 struct mm_struct *mm,
194 struct amdgpu_mn *rmn = container_of(mn, struct amdgpu_mn, mn);
195 struct interval_tree_node *it;
197 /* notification is exclusive, but interval is inclusive */
200 mutex_lock(&rmn->lock);
202 it = interval_tree_iter_first(&rmn->objects, start, end);
204 struct amdgpu_mn_node *node;
206 node = container_of(it, struct amdgpu_mn_node, it);
207 it = interval_tree_iter_next(it, start, end);
209 amdgpu_mn_invalidate_node(node, start, end);
212 mutex_unlock(&rmn->lock);
215 static const struct mmu_notifier_ops amdgpu_mn_ops = {
216 .release = amdgpu_mn_release,
217 .invalidate_page = amdgpu_mn_invalidate_page,
218 .invalidate_range_start = amdgpu_mn_invalidate_range_start,
222 * amdgpu_mn_get - create notifier context
224 * @adev: amdgpu device pointer
226 * Creates a notifier context for current->mm.
228 static struct amdgpu_mn *amdgpu_mn_get(struct amdgpu_device *adev)
230 struct mm_struct *mm = current->mm;
231 struct amdgpu_mn *rmn;
234 mutex_lock(&adev->mn_lock);
235 if (down_write_killable(&mm->mmap_sem)) {
236 mutex_unlock(&adev->mn_lock);
237 return ERR_PTR(-EINTR);
240 hash_for_each_possible(adev->mn_hash, rmn, node, (unsigned long)mm)
244 rmn = kzalloc(sizeof(*rmn), GFP_KERNEL);
246 rmn = ERR_PTR(-ENOMEM);
252 rmn->mn.ops = &amdgpu_mn_ops;
253 mutex_init(&rmn->lock);
254 rmn->objects = RB_ROOT;
256 r = __mmu_notifier_register(&rmn->mn, mm);
260 hash_add(adev->mn_hash, &rmn->node, (unsigned long)mm);
263 up_write(&mm->mmap_sem);
264 mutex_unlock(&adev->mn_lock);
269 up_write(&mm->mmap_sem);
270 mutex_unlock(&adev->mn_lock);
277 * amdgpu_mn_register - register a BO for notifier updates
279 * @bo: amdgpu buffer object
280 * @addr: userptr addr we should monitor
282 * Registers an MMU notifier for the given BO at the specified address.
283 * Returns 0 on success, -ERRNO if anything goes wrong.
285 int amdgpu_mn_register(struct amdgpu_bo *bo, unsigned long addr)
287 unsigned long end = addr + amdgpu_bo_size(bo) - 1;
288 struct amdgpu_device *adev = bo->adev;
289 struct amdgpu_mn *rmn;
290 struct amdgpu_mn_node *node = NULL;
291 struct list_head bos;
292 struct interval_tree_node *it;
294 rmn = amdgpu_mn_get(adev);
298 INIT_LIST_HEAD(&bos);
300 mutex_lock(&rmn->lock);
302 while ((it = interval_tree_iter_first(&rmn->objects, addr, end))) {
304 node = container_of(it, struct amdgpu_mn_node, it);
305 interval_tree_remove(&node->it, &rmn->objects);
306 addr = min(it->start, addr);
307 end = max(it->last, end);
308 list_splice(&node->bos, &bos);
312 node = kmalloc(sizeof(struct amdgpu_mn_node), GFP_KERNEL);
314 mutex_unlock(&rmn->lock);
321 node->it.start = addr;
323 INIT_LIST_HEAD(&node->bos);
324 list_splice(&bos, &node->bos);
325 list_add(&bo->mn_list, &node->bos);
327 interval_tree_insert(&node->it, &rmn->objects);
329 mutex_unlock(&rmn->lock);
335 * amdgpu_mn_unregister - unregister a BO for notifier updates
337 * @bo: amdgpu buffer object
339 * Remove any registration of MMU notifier updates from the buffer object.
341 void amdgpu_mn_unregister(struct amdgpu_bo *bo)
343 struct amdgpu_device *adev = bo->adev;
344 struct amdgpu_mn *rmn;
345 struct list_head *head;
347 mutex_lock(&adev->mn_lock);
351 mutex_unlock(&adev->mn_lock);
355 mutex_lock(&rmn->lock);
357 /* save the next list entry for later */
358 head = bo->mn_list.next;
361 list_del(&bo->mn_list);
363 if (list_empty(head)) {
364 struct amdgpu_mn_node *node;
365 node = container_of(head, struct amdgpu_mn_node, bos);
366 interval_tree_remove(&node->it, &rmn->objects);
370 mutex_unlock(&rmn->lock);
371 mutex_unlock(&adev->mn_lock);