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
34 * For coherent userptr handling registers an MMU notifier to inform the driver
35 * about updates on the page tables of a process.
37 * When somebody tries to invalidate the page tables we block the update until
38 * all operations on the pages in question are completed, then those pages are
39 * marked as accessed and also dirty if it wasn't a read only access.
41 * New command submissions using the userptrs in question are delayed until all
42 * page table invalidation are completed and we once more see a coherent process
46 #include <linux/firmware.h>
47 #include <linux/module.h>
48 #include <linux/mmu_notifier.h>
49 #include <linux/interval_tree.h>
54 #include "amdgpu_amdkfd.h"
59 * @adev: amdgpu device pointer
60 * @mm: process address space
61 * @mn: MMU notifier structure
62 * @type: type of MMU notifier
63 * @work: destruction work item
64 * @node: hash table node to find structure by adev and mn
65 * @lock: rw semaphore protecting the notifier nodes
66 * @objects: interval tree containing amdgpu_mn_nodes
67 * @read_lock: mutex for recursive locking of @lock
68 * @recursion: depth of recursion
70 * Data for each amdgpu device and process address space.
73 /* constant after initialisation */
74 struct amdgpu_device *adev;
76 struct mmu_notifier mn;
77 enum amdgpu_mn_type type;
79 /* only used on destruction */
80 struct work_struct work;
82 /* protected by adev->mn_lock */
83 struct hlist_node node;
85 /* objects protected by lock */
86 struct rw_semaphore lock;
87 struct rb_root_cached objects;
88 struct mutex read_lock;
93 * struct amdgpu_mn_node
95 * @it: interval node defining start-last of the affected address range
96 * @bos: list of all BOs in the affected address range
98 * Manages all BOs which are affected of a certain range of address space.
100 struct amdgpu_mn_node {
101 struct interval_tree_node it;
102 struct list_head bos;
106 * amdgpu_mn_destroy - destroy the MMU notifier
108 * @work: previously sheduled work item
110 * Lazy destroys the notifier from a work item
112 static void amdgpu_mn_destroy(struct work_struct *work)
114 struct amdgpu_mn *amn = container_of(work, struct amdgpu_mn, work);
115 struct amdgpu_device *adev = amn->adev;
116 struct amdgpu_mn_node *node, *next_node;
117 struct amdgpu_bo *bo, *next_bo;
119 mutex_lock(&adev->mn_lock);
120 down_write(&amn->lock);
121 hash_del(&amn->node);
122 rbtree_postorder_for_each_entry_safe(node, next_node,
123 &amn->objects.rb_root, it.rb) {
124 list_for_each_entry_safe(bo, next_bo, &node->bos, mn_list) {
126 list_del_init(&bo->mn_list);
130 up_write(&amn->lock);
131 mutex_unlock(&adev->mn_lock);
132 mmu_notifier_unregister_no_release(&amn->mn, amn->mm);
137 * amdgpu_mn_release - callback to notify about mm destruction
140 * @mm: the mm this callback is about
142 * Shedule a work item to lazy destroy our notifier.
144 static void amdgpu_mn_release(struct mmu_notifier *mn,
145 struct mm_struct *mm)
147 struct amdgpu_mn *amn = container_of(mn, struct amdgpu_mn, mn);
149 INIT_WORK(&amn->work, amdgpu_mn_destroy);
150 schedule_work(&amn->work);
155 * amdgpu_mn_lock - take the write side lock for this notifier
159 void amdgpu_mn_lock(struct amdgpu_mn *mn)
162 down_write(&mn->lock);
166 * amdgpu_mn_unlock - drop the write side lock for this notifier
170 void amdgpu_mn_unlock(struct amdgpu_mn *mn)
177 * amdgpu_mn_read_lock - take the read side lock for this notifier
181 static void amdgpu_mn_read_lock(struct amdgpu_mn *amn)
183 mutex_lock(&amn->read_lock);
184 if (atomic_inc_return(&amn->recursion) == 1)
185 down_read_non_owner(&amn->lock);
186 mutex_unlock(&amn->read_lock);
190 * amdgpu_mn_read_unlock - drop the read side lock for this notifier
194 static void amdgpu_mn_read_unlock(struct amdgpu_mn *amn)
196 if (atomic_dec_return(&amn->recursion) == 0)
197 up_read_non_owner(&amn->lock);
201 * amdgpu_mn_invalidate_node - unmap all BOs of a node
203 * @node: the node with the BOs to unmap
204 * @start: start of address range affected
205 * @end: end of address range affected
207 * Block for operations on BOs to finish and mark pages as accessed and
210 static void amdgpu_mn_invalidate_node(struct amdgpu_mn_node *node,
214 struct amdgpu_bo *bo;
217 list_for_each_entry(bo, &node->bos, mn_list) {
219 if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm, start, end))
222 r = reservation_object_wait_timeout_rcu(bo->tbo.resv,
223 true, false, MAX_SCHEDULE_TIMEOUT);
225 DRM_ERROR("(%ld) failed to wait for user bo\n", r);
227 amdgpu_ttm_tt_mark_user_pages(bo->tbo.ttm);
232 * amdgpu_mn_invalidate_range_start_gfx - callback to notify about mm change
235 * @mm: the mm this callback is about
236 * @start: start of updated range
237 * @end: end of updated range
239 * Block for operations on BOs to finish and mark pages as accessed and
242 static void amdgpu_mn_invalidate_range_start_gfx(struct mmu_notifier *mn,
243 struct mm_struct *mm,
247 struct amdgpu_mn *amn = container_of(mn, struct amdgpu_mn, mn);
248 struct interval_tree_node *it;
250 /* notification is exclusive, but interval is inclusive */
253 amdgpu_mn_read_lock(amn);
255 it = interval_tree_iter_first(&amn->objects, start, end);
257 struct amdgpu_mn_node *node;
259 node = container_of(it, struct amdgpu_mn_node, it);
260 it = interval_tree_iter_next(it, start, end);
262 amdgpu_mn_invalidate_node(node, start, end);
267 * amdgpu_mn_invalidate_range_start_hsa - callback to notify about mm change
270 * @mm: the mm this callback is about
271 * @start: start of updated range
272 * @end: end of updated range
274 * We temporarily evict all BOs between start and end. This
275 * necessitates evicting all user-mode queues of the process. The BOs
276 * are restorted in amdgpu_mn_invalidate_range_end_hsa.
278 static void amdgpu_mn_invalidate_range_start_hsa(struct mmu_notifier *mn,
279 struct mm_struct *mm,
283 struct amdgpu_mn *amn = container_of(mn, struct amdgpu_mn, mn);
284 struct interval_tree_node *it;
286 /* notification is exclusive, but interval is inclusive */
289 amdgpu_mn_read_lock(amn);
291 it = interval_tree_iter_first(&amn->objects, start, end);
293 struct amdgpu_mn_node *node;
294 struct amdgpu_bo *bo;
296 node = container_of(it, struct amdgpu_mn_node, it);
297 it = interval_tree_iter_next(it, start, end);
299 list_for_each_entry(bo, &node->bos, mn_list) {
300 struct kgd_mem *mem = bo->kfd_bo;
302 if (amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm,
304 amdgpu_amdkfd_evict_userptr(mem, mm);
310 * amdgpu_mn_invalidate_range_end - callback to notify about mm change
313 * @mm: the mm this callback is about
314 * @start: start of updated range
315 * @end: end of updated range
317 * Release the lock again to allow new command submissions.
319 static void amdgpu_mn_invalidate_range_end(struct mmu_notifier *mn,
320 struct mm_struct *mm,
324 struct amdgpu_mn *amn = container_of(mn, struct amdgpu_mn, mn);
326 amdgpu_mn_read_unlock(amn);
329 static const struct mmu_notifier_ops amdgpu_mn_ops[] = {
330 [AMDGPU_MN_TYPE_GFX] = {
331 .release = amdgpu_mn_release,
332 .invalidate_range_start = amdgpu_mn_invalidate_range_start_gfx,
333 .invalidate_range_end = amdgpu_mn_invalidate_range_end,
335 [AMDGPU_MN_TYPE_HSA] = {
336 .release = amdgpu_mn_release,
337 .invalidate_range_start = amdgpu_mn_invalidate_range_start_hsa,
338 .invalidate_range_end = amdgpu_mn_invalidate_range_end,
342 /* Low bits of any reasonable mm pointer will be unused due to struct
343 * alignment. Use these bits to make a unique key from the mm pointer
346 #define AMDGPU_MN_KEY(mm, type) ((unsigned long)(mm) + (type))
349 * amdgpu_mn_get - create notifier context
351 * @adev: amdgpu device pointer
352 * @type: type of MMU notifier context
354 * Creates a notifier context for current->mm.
356 struct amdgpu_mn *amdgpu_mn_get(struct amdgpu_device *adev,
357 enum amdgpu_mn_type type)
359 struct mm_struct *mm = current->mm;
360 struct amdgpu_mn *amn;
361 unsigned long key = AMDGPU_MN_KEY(mm, type);
364 mutex_lock(&adev->mn_lock);
365 if (down_write_killable(&mm->mmap_sem)) {
366 mutex_unlock(&adev->mn_lock);
367 return ERR_PTR(-EINTR);
370 hash_for_each_possible(adev->mn_hash, amn, node, key)
371 if (AMDGPU_MN_KEY(amn->mm, amn->type) == key)
374 amn = kzalloc(sizeof(*amn), GFP_KERNEL);
376 amn = ERR_PTR(-ENOMEM);
382 init_rwsem(&amn->lock);
384 amn->mn.ops = &amdgpu_mn_ops[type];
385 amn->objects = RB_ROOT_CACHED;
386 mutex_init(&amn->read_lock);
387 atomic_set(&amn->recursion, 0);
389 r = __mmu_notifier_register(&amn->mn, mm);
393 hash_add(adev->mn_hash, &amn->node, AMDGPU_MN_KEY(mm, type));
396 up_write(&mm->mmap_sem);
397 mutex_unlock(&adev->mn_lock);
402 up_write(&mm->mmap_sem);
403 mutex_unlock(&adev->mn_lock);
410 * amdgpu_mn_register - register a BO for notifier updates
412 * @bo: amdgpu buffer object
413 * @addr: userptr addr we should monitor
415 * Registers an MMU notifier for the given BO at the specified address.
416 * Returns 0 on success, -ERRNO if anything goes wrong.
418 int amdgpu_mn_register(struct amdgpu_bo *bo, unsigned long addr)
420 unsigned long end = addr + amdgpu_bo_size(bo) - 1;
421 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
422 enum amdgpu_mn_type type =
423 bo->kfd_bo ? AMDGPU_MN_TYPE_HSA : AMDGPU_MN_TYPE_GFX;
424 struct amdgpu_mn *amn;
425 struct amdgpu_mn_node *node = NULL, *new_node;
426 struct list_head bos;
427 struct interval_tree_node *it;
429 amn = amdgpu_mn_get(adev, type);
433 new_node = kmalloc(sizeof(*new_node), GFP_KERNEL);
437 INIT_LIST_HEAD(&bos);
439 down_write(&amn->lock);
441 while ((it = interval_tree_iter_first(&amn->objects, addr, end))) {
443 node = container_of(it, struct amdgpu_mn_node, it);
444 interval_tree_remove(&node->it, &amn->objects);
445 addr = min(it->start, addr);
446 end = max(it->last, end);
447 list_splice(&node->bos, &bos);
457 node->it.start = addr;
459 INIT_LIST_HEAD(&node->bos);
460 list_splice(&bos, &node->bos);
461 list_add(&bo->mn_list, &node->bos);
463 interval_tree_insert(&node->it, &amn->objects);
465 up_write(&amn->lock);
471 * amdgpu_mn_unregister - unregister a BO for notifier updates
473 * @bo: amdgpu buffer object
475 * Remove any registration of MMU notifier updates from the buffer object.
477 void amdgpu_mn_unregister(struct amdgpu_bo *bo)
479 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
480 struct amdgpu_mn *amn;
481 struct list_head *head;
483 mutex_lock(&adev->mn_lock);
487 mutex_unlock(&adev->mn_lock);
491 down_write(&amn->lock);
493 /* save the next list entry for later */
494 head = bo->mn_list.next;
497 list_del_init(&bo->mn_list);
499 if (list_empty(head)) {
500 struct amdgpu_mn_node *node;
502 node = container_of(head, struct amdgpu_mn_node, bos);
503 interval_tree_remove(&node->it, &amn->objects);
507 up_write(&amn->lock);
508 mutex_unlock(&adev->mn_lock);