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/hmm.h>
49 #include <linux/interval_tree.h>
54 #include "amdgpu_amdkfd.h"
59 * @adev: amdgpu device pointer
60 * @mm: process address space
61 * @type: type of MMU notifier
62 * @work: destruction work item
63 * @node: hash table node to find structure by adev and mn
64 * @lock: rw semaphore protecting the notifier nodes
65 * @objects: interval tree containing amdgpu_mn_nodes
66 * @mirror: HMM mirror function support
68 * Data for each amdgpu device and process address space.
71 /* constant after initialisation */
72 struct amdgpu_device *adev;
74 enum amdgpu_mn_type type;
76 /* only used on destruction */
77 struct work_struct work;
79 /* protected by adev->mn_lock */
80 struct hlist_node node;
82 /* objects protected by lock */
83 struct rw_semaphore lock;
84 struct rb_root_cached objects;
87 struct hmm_mirror mirror;
91 * struct amdgpu_mn_node
93 * @it: interval node defining start-last of the affected address range
94 * @bos: list of all BOs in the affected address range
96 * Manages all BOs which are affected of a certain range of address space.
98 struct amdgpu_mn_node {
99 struct interval_tree_node it;
100 struct list_head bos;
104 * amdgpu_mn_destroy - destroy the HMM mirror
106 * @work: previously sheduled work item
108 * Lazy destroys the notifier from a work item
110 static void amdgpu_mn_destroy(struct work_struct *work)
112 struct amdgpu_mn *amn = container_of(work, struct amdgpu_mn, work);
113 struct amdgpu_device *adev = amn->adev;
114 struct amdgpu_mn_node *node, *next_node;
115 struct amdgpu_bo *bo, *next_bo;
117 mutex_lock(&adev->mn_lock);
118 down_write(&amn->lock);
119 hash_del(&amn->node);
120 rbtree_postorder_for_each_entry_safe(node, next_node,
121 &amn->objects.rb_root, it.rb) {
122 list_for_each_entry_safe(bo, next_bo, &node->bos, mn_list) {
124 list_del_init(&bo->mn_list);
128 up_write(&amn->lock);
129 mutex_unlock(&adev->mn_lock);
131 hmm_mirror_unregister(&amn->mirror);
136 * amdgpu_hmm_mirror_release - callback to notify about mm destruction
138 * @mirror: the HMM mirror (mm) this callback is about
140 * Shedule a work item to lazy destroy HMM mirror.
142 static void amdgpu_hmm_mirror_release(struct hmm_mirror *mirror)
144 struct amdgpu_mn *amn = container_of(mirror, struct amdgpu_mn, mirror);
146 INIT_WORK(&amn->work, amdgpu_mn_destroy);
147 schedule_work(&amn->work);
151 * amdgpu_mn_lock - take the write side lock for this notifier
155 void amdgpu_mn_lock(struct amdgpu_mn *mn)
158 down_write(&mn->lock);
162 * amdgpu_mn_unlock - drop the write side lock for this notifier
166 void amdgpu_mn_unlock(struct amdgpu_mn *mn)
173 * amdgpu_mn_read_lock - take the read side lock for this notifier
177 static int amdgpu_mn_read_lock(struct amdgpu_mn *amn, bool blockable)
180 down_read(&amn->lock);
181 else if (!down_read_trylock(&amn->lock))
188 * amdgpu_mn_read_unlock - drop the read side lock for this notifier
192 static void amdgpu_mn_read_unlock(struct amdgpu_mn *amn)
198 * amdgpu_mn_invalidate_node - unmap all BOs of a node
200 * @node: the node with the BOs to unmap
201 * @start: start of address range affected
202 * @end: end of address range affected
204 * Block for operations on BOs to finish and mark pages as accessed and
207 static void amdgpu_mn_invalidate_node(struct amdgpu_mn_node *node,
211 struct amdgpu_bo *bo;
214 list_for_each_entry(bo, &node->bos, mn_list) {
216 if (!amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm, start, end))
219 r = reservation_object_wait_timeout_rcu(bo->tbo.resv,
220 true, false, MAX_SCHEDULE_TIMEOUT);
222 DRM_ERROR("(%ld) failed to wait for user bo\n", r);
227 * amdgpu_mn_sync_pagetables_gfx - callback to notify about mm change
229 * @mirror: the hmm_mirror (mm) is about to update
230 * @update: the update start, end address
232 * Block for operations on BOs to finish and mark pages as accessed and
235 static int amdgpu_mn_sync_pagetables_gfx(struct hmm_mirror *mirror,
236 const struct hmm_update *update)
238 struct amdgpu_mn *amn = container_of(mirror, struct amdgpu_mn, mirror);
239 unsigned long start = update->start;
240 unsigned long end = update->end;
241 bool blockable = update->blockable;
242 struct interval_tree_node *it;
244 /* notification is exclusive, but interval is inclusive */
247 /* TODO we should be able to split locking for interval tree and
248 * amdgpu_mn_invalidate_node
250 if (amdgpu_mn_read_lock(amn, blockable))
253 it = interval_tree_iter_first(&amn->objects, start, end);
255 struct amdgpu_mn_node *node;
258 amdgpu_mn_read_unlock(amn);
262 node = container_of(it, struct amdgpu_mn_node, it);
263 it = interval_tree_iter_next(it, start, end);
265 amdgpu_mn_invalidate_node(node, start, end);
268 amdgpu_mn_read_unlock(amn);
274 * amdgpu_mn_sync_pagetables_hsa - callback to notify about mm change
276 * @mirror: the hmm_mirror (mm) is about to update
277 * @update: the update start, end address
279 * We temporarily evict all BOs between start and end. This
280 * necessitates evicting all user-mode queues of the process. The BOs
281 * are restorted in amdgpu_mn_invalidate_range_end_hsa.
283 static int amdgpu_mn_sync_pagetables_hsa(struct hmm_mirror *mirror,
284 const struct hmm_update *update)
286 struct amdgpu_mn *amn = container_of(mirror, struct amdgpu_mn, mirror);
287 unsigned long start = update->start;
288 unsigned long end = update->end;
289 bool blockable = update->blockable;
290 struct interval_tree_node *it;
292 /* notification is exclusive, but interval is inclusive */
295 if (amdgpu_mn_read_lock(amn, blockable))
298 it = interval_tree_iter_first(&amn->objects, start, end);
300 struct amdgpu_mn_node *node;
301 struct amdgpu_bo *bo;
304 amdgpu_mn_read_unlock(amn);
308 node = container_of(it, struct amdgpu_mn_node, it);
309 it = interval_tree_iter_next(it, start, end);
311 list_for_each_entry(bo, &node->bos, mn_list) {
312 struct kgd_mem *mem = bo->kfd_bo;
314 if (amdgpu_ttm_tt_affect_userptr(bo->tbo.ttm,
316 amdgpu_amdkfd_evict_userptr(mem, amn->mm);
320 amdgpu_mn_read_unlock(amn);
325 /* Low bits of any reasonable mm pointer will be unused due to struct
326 * alignment. Use these bits to make a unique key from the mm pointer
329 #define AMDGPU_MN_KEY(mm, type) ((unsigned long)(mm) + (type))
331 static struct hmm_mirror_ops amdgpu_hmm_mirror_ops[] = {
332 [AMDGPU_MN_TYPE_GFX] = {
333 .sync_cpu_device_pagetables = amdgpu_mn_sync_pagetables_gfx,
334 .release = amdgpu_hmm_mirror_release
336 [AMDGPU_MN_TYPE_HSA] = {
337 .sync_cpu_device_pagetables = amdgpu_mn_sync_pagetables_hsa,
338 .release = amdgpu_hmm_mirror_release
343 * amdgpu_mn_get - create HMM mirror context
345 * @adev: amdgpu device pointer
346 * @type: type of MMU notifier context
348 * Creates a HMM mirror context for current->mm.
350 struct amdgpu_mn *amdgpu_mn_get(struct amdgpu_device *adev,
351 enum amdgpu_mn_type type)
353 struct mm_struct *mm = current->mm;
354 struct amdgpu_mn *amn;
355 unsigned long key = AMDGPU_MN_KEY(mm, type);
358 mutex_lock(&adev->mn_lock);
359 if (down_write_killable(&mm->mmap_sem)) {
360 mutex_unlock(&adev->mn_lock);
361 return ERR_PTR(-EINTR);
364 hash_for_each_possible(adev->mn_hash, amn, node, key)
365 if (AMDGPU_MN_KEY(amn->mm, amn->type) == key)
368 amn = kzalloc(sizeof(*amn), GFP_KERNEL);
370 amn = ERR_PTR(-ENOMEM);
376 init_rwsem(&amn->lock);
378 amn->objects = RB_ROOT_CACHED;
380 amn->mirror.ops = &amdgpu_hmm_mirror_ops[type];
381 r = hmm_mirror_register(&amn->mirror, mm);
385 hash_add(adev->mn_hash, &amn->node, AMDGPU_MN_KEY(mm, type));
388 up_write(&mm->mmap_sem);
389 mutex_unlock(&adev->mn_lock);
394 up_write(&mm->mmap_sem);
395 mutex_unlock(&adev->mn_lock);
402 * amdgpu_mn_register - register a BO for notifier updates
404 * @bo: amdgpu buffer object
405 * @addr: userptr addr we should monitor
407 * Registers an HMM mirror for the given BO at the specified address.
408 * Returns 0 on success, -ERRNO if anything goes wrong.
410 int amdgpu_mn_register(struct amdgpu_bo *bo, unsigned long addr)
412 unsigned long end = addr + amdgpu_bo_size(bo) - 1;
413 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
414 enum amdgpu_mn_type type =
415 bo->kfd_bo ? AMDGPU_MN_TYPE_HSA : AMDGPU_MN_TYPE_GFX;
416 struct amdgpu_mn *amn;
417 struct amdgpu_mn_node *node = NULL, *new_node;
418 struct list_head bos;
419 struct interval_tree_node *it;
421 amn = amdgpu_mn_get(adev, type);
425 new_node = kmalloc(sizeof(*new_node), GFP_KERNEL);
429 INIT_LIST_HEAD(&bos);
431 down_write(&amn->lock);
433 while ((it = interval_tree_iter_first(&amn->objects, addr, end))) {
435 node = container_of(it, struct amdgpu_mn_node, it);
436 interval_tree_remove(&node->it, &amn->objects);
437 addr = min(it->start, addr);
438 end = max(it->last, end);
439 list_splice(&node->bos, &bos);
449 node->it.start = addr;
451 INIT_LIST_HEAD(&node->bos);
452 list_splice(&bos, &node->bos);
453 list_add(&bo->mn_list, &node->bos);
455 interval_tree_insert(&node->it, &amn->objects);
457 up_write(&amn->lock);
463 * amdgpu_mn_unregister - unregister a BO for HMM mirror updates
465 * @bo: amdgpu buffer object
467 * Remove any registration of HMM mirror updates from the buffer object.
469 void amdgpu_mn_unregister(struct amdgpu_bo *bo)
471 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
472 struct amdgpu_mn *amn;
473 struct list_head *head;
475 mutex_lock(&adev->mn_lock);
479 mutex_unlock(&adev->mn_lock);
483 down_write(&amn->lock);
485 /* save the next list entry for later */
486 head = bo->mn_list.next;
489 list_del_init(&bo->mn_list);
491 if (list_empty(head)) {
492 struct amdgpu_mn_node *node;
494 node = container_of(head, struct amdgpu_mn_node, bos);
495 interval_tree_remove(&node->it, &amn->objects);
499 up_write(&amn->lock);
500 mutex_unlock(&adev->mn_lock);
503 /* flags used by HMM internal, not related to CPU/GPU PTE flags */
504 static const uint64_t hmm_range_flags[HMM_PFN_FLAG_MAX] = {
505 (1 << 0), /* HMM_PFN_VALID */
506 (1 << 1), /* HMM_PFN_WRITE */
507 0 /* HMM_PFN_DEVICE_PRIVATE */
510 static const uint64_t hmm_range_values[HMM_PFN_VALUE_MAX] = {
511 0xfffffffffffffffeUL, /* HMM_PFN_ERROR */
512 0, /* HMM_PFN_NONE */
513 0xfffffffffffffffcUL /* HMM_PFN_SPECIAL */
516 void amdgpu_hmm_init_range(struct hmm_range *range)
519 range->flags = hmm_range_flags;
520 range->values = hmm_range_values;
521 range->pfn_shift = PAGE_SHIFT;
522 INIT_LIST_HEAD(&range->list);