1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
3 * Copyright(c) 2020 Cornelis Networks, Inc.
4 * Copyright(c) 2016 - 2017 Intel Corporation.
7 #include <linux/list.h>
8 #include <linux/rculist.h>
9 #include <linux/mmu_notifier.h>
10 #include <linux/interval_tree_generic.h>
11 #include <linux/sched/mm.h>
16 static unsigned long mmu_node_start(struct mmu_rb_node *);
17 static unsigned long mmu_node_last(struct mmu_rb_node *);
18 static int mmu_notifier_range_start(struct mmu_notifier *,
19 const struct mmu_notifier_range *);
20 static struct mmu_rb_node *__mmu_rb_search(struct mmu_rb_handler *,
21 unsigned long, unsigned long);
22 static void do_remove(struct mmu_rb_handler *handler,
23 struct list_head *del_list);
24 static void handle_remove(struct work_struct *work);
26 static const struct mmu_notifier_ops mn_opts = {
27 .invalidate_range_start = mmu_notifier_range_start,
30 INTERVAL_TREE_DEFINE(struct mmu_rb_node, node, unsigned long, __last,
31 mmu_node_start, mmu_node_last, static, __mmu_int_rb);
33 static unsigned long mmu_node_start(struct mmu_rb_node *node)
35 return node->addr & PAGE_MASK;
38 static unsigned long mmu_node_last(struct mmu_rb_node *node)
40 return PAGE_ALIGN(node->addr + node->len) - 1;
43 int hfi1_mmu_rb_register(void *ops_arg,
44 struct mmu_rb_ops *ops,
45 struct workqueue_struct *wq,
46 struct mmu_rb_handler **handler)
48 struct mmu_rb_handler *h;
51 h = kzalloc(sizeof(*h), GFP_KERNEL);
55 h->root = RB_ROOT_CACHED;
58 INIT_HLIST_NODE(&h->mn.hlist);
59 spin_lock_init(&h->lock);
61 INIT_WORK(&h->del_work, handle_remove);
62 INIT_LIST_HEAD(&h->del_list);
63 INIT_LIST_HEAD(&h->lru_list);
66 ret = mmu_notifier_register(&h->mn, current->mm);
76 void hfi1_mmu_rb_unregister(struct mmu_rb_handler *handler)
78 struct mmu_rb_node *rbnode;
81 struct list_head del_list;
83 /* Prevent freeing of mm until we are completely finished. */
84 mmgrab(handler->mn.mm);
86 /* Unregister first so we don't get any more notifications. */
87 mmu_notifier_unregister(&handler->mn, handler->mn.mm);
90 * Make sure the wq delete handler is finished running. It will not
91 * be triggered once the mmu notifiers are unregistered above.
93 flush_work(&handler->del_work);
95 INIT_LIST_HEAD(&del_list);
97 spin_lock_irqsave(&handler->lock, flags);
98 while ((node = rb_first_cached(&handler->root))) {
99 rbnode = rb_entry(node, struct mmu_rb_node, node);
100 rb_erase_cached(node, &handler->root);
101 /* move from LRU list to delete list */
102 list_move(&rbnode->list, &del_list);
104 spin_unlock_irqrestore(&handler->lock, flags);
106 do_remove(handler, &del_list);
108 /* Now the mm may be freed. */
109 mmdrop(handler->mn.mm);
114 int hfi1_mmu_rb_insert(struct mmu_rb_handler *handler,
115 struct mmu_rb_node *mnode)
117 struct mmu_rb_node *node;
121 trace_hfi1_mmu_rb_insert(mnode->addr, mnode->len);
123 if (current->mm != handler->mn.mm)
126 spin_lock_irqsave(&handler->lock, flags);
127 node = __mmu_rb_search(handler, mnode->addr, mnode->len);
132 __mmu_int_rb_insert(mnode, &handler->root);
133 list_add(&mnode->list, &handler->lru_list);
135 ret = handler->ops->insert(handler->ops_arg, mnode);
137 __mmu_int_rb_remove(mnode, &handler->root);
138 list_del(&mnode->list); /* remove from LRU list */
140 mnode->handler = handler;
142 spin_unlock_irqrestore(&handler->lock, flags);
146 /* Caller must hold handler lock */
147 static struct mmu_rb_node *__mmu_rb_search(struct mmu_rb_handler *handler,
151 struct mmu_rb_node *node = NULL;
153 trace_hfi1_mmu_rb_search(addr, len);
154 if (!handler->ops->filter) {
155 node = __mmu_int_rb_iter_first(&handler->root, addr,
158 for (node = __mmu_int_rb_iter_first(&handler->root, addr,
161 node = __mmu_int_rb_iter_next(node, addr,
163 if (handler->ops->filter(node, addr, len))
170 bool hfi1_mmu_rb_remove_unless_exact(struct mmu_rb_handler *handler,
171 unsigned long addr, unsigned long len,
172 struct mmu_rb_node **rb_node)
174 struct mmu_rb_node *node;
178 if (current->mm != handler->mn.mm)
181 spin_lock_irqsave(&handler->lock, flags);
182 node = __mmu_rb_search(handler, addr, len);
184 if (node->addr == addr && node->len == len)
186 __mmu_int_rb_remove(node, &handler->root);
187 list_del(&node->list); /* remove from LRU list */
191 spin_unlock_irqrestore(&handler->lock, flags);
196 void hfi1_mmu_rb_evict(struct mmu_rb_handler *handler, void *evict_arg)
198 struct mmu_rb_node *rbnode, *ptr;
199 struct list_head del_list;
203 if (current->mm != handler->mn.mm)
206 INIT_LIST_HEAD(&del_list);
208 spin_lock_irqsave(&handler->lock, flags);
209 list_for_each_entry_safe_reverse(rbnode, ptr, &handler->lru_list,
211 if (handler->ops->evict(handler->ops_arg, rbnode, evict_arg,
213 __mmu_int_rb_remove(rbnode, &handler->root);
214 /* move from LRU list to delete list */
215 list_move(&rbnode->list, &del_list);
220 spin_unlock_irqrestore(&handler->lock, flags);
222 while (!list_empty(&del_list)) {
223 rbnode = list_first_entry(&del_list, struct mmu_rb_node, list);
224 list_del(&rbnode->list);
225 handler->ops->remove(handler->ops_arg, rbnode);
230 * It is up to the caller to ensure that this function does not race with the
231 * mmu invalidate notifier which may be calling the users remove callback on
234 void hfi1_mmu_rb_remove(struct mmu_rb_handler *handler,
235 struct mmu_rb_node *node)
239 if (current->mm != handler->mn.mm)
242 /* Validity of handler and node pointers has been checked by caller. */
243 trace_hfi1_mmu_rb_remove(node->addr, node->len);
244 spin_lock_irqsave(&handler->lock, flags);
245 __mmu_int_rb_remove(node, &handler->root);
246 list_del(&node->list); /* remove from LRU list */
247 spin_unlock_irqrestore(&handler->lock, flags);
249 handler->ops->remove(handler->ops_arg, node);
252 static int mmu_notifier_range_start(struct mmu_notifier *mn,
253 const struct mmu_notifier_range *range)
255 struct mmu_rb_handler *handler =
256 container_of(mn, struct mmu_rb_handler, mn);
257 struct rb_root_cached *root = &handler->root;
258 struct mmu_rb_node *node, *ptr = NULL;
262 spin_lock_irqsave(&handler->lock, flags);
263 for (node = __mmu_int_rb_iter_first(root, range->start, range->end-1);
265 /* Guard against node removal. */
266 ptr = __mmu_int_rb_iter_next(node, range->start,
268 trace_hfi1_mmu_mem_invalidate(node->addr, node->len);
269 if (handler->ops->invalidate(handler->ops_arg, node)) {
270 __mmu_int_rb_remove(node, root);
271 /* move from LRU list to delete list */
272 list_move(&node->list, &handler->del_list);
276 spin_unlock_irqrestore(&handler->lock, flags);
279 queue_work(handler->wq, &handler->del_work);
285 * Call the remove function for the given handler and the list. This
286 * is expected to be called with a delete list extracted from handler.
287 * The caller should not be holding the handler lock.
289 static void do_remove(struct mmu_rb_handler *handler,
290 struct list_head *del_list)
292 struct mmu_rb_node *node;
294 while (!list_empty(del_list)) {
295 node = list_first_entry(del_list, struct mmu_rb_node, list);
296 list_del(&node->list);
297 handler->ops->remove(handler->ops_arg, node);
302 * Work queue function to remove all nodes that have been queued up to
303 * be removed. The key feature is that mm->mmap_lock is not being held
304 * and the remove callback can sleep while taking it, if needed.
306 static void handle_remove(struct work_struct *work)
308 struct mmu_rb_handler *handler = container_of(work,
309 struct mmu_rb_handler,
311 struct list_head del_list;
314 /* remove anything that is queued to get removed */
315 spin_lock_irqsave(&handler->lock, flags);
316 list_replace_init(&handler->del_list, &del_list);
317 spin_unlock_irqrestore(&handler->lock, flags);
319 do_remove(handler, &del_list);