xref: /linux/drivers/infiniband/hw/hfi1/mmu_rb.c (revision 79997eda0d31bc68203c95ecb978773ee6ce7a1f)
1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3  * Copyright(c) 2020 Cornelis Networks, Inc.
4  * Copyright(c) 2016 - 2017 Intel Corporation.
5  */
6 
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>
12 
13 #include "mmu_rb.h"
14 #include "trace.h"
15 
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 release_immediate(struct kref *refcount);
23 static void handle_remove(struct work_struct *work);
24 
25 static const struct mmu_notifier_ops mn_opts = {
26 	.invalidate_range_start = mmu_notifier_range_start,
27 };
28 
29 INTERVAL_TREE_DEFINE(struct mmu_rb_node, node, unsigned long, __last,
30 		     mmu_node_start, mmu_node_last, static, __mmu_int_rb);
31 
32 static unsigned long mmu_node_start(struct mmu_rb_node *node)
33 {
34 	return node->addr & PAGE_MASK;
35 }
36 
37 static unsigned long mmu_node_last(struct mmu_rb_node *node)
38 {
39 	return PAGE_ALIGN(node->addr + node->len) - 1;
40 }
41 
42 int hfi1_mmu_rb_register(void *ops_arg,
43 			 struct mmu_rb_ops *ops,
44 			 struct workqueue_struct *wq,
45 			 struct mmu_rb_handler **handler)
46 {
47 	struct mmu_rb_handler *h;
48 	void *free_ptr;
49 	int ret;
50 
51 	free_ptr = kzalloc(sizeof(*h) + cache_line_size() - 1, GFP_KERNEL);
52 	if (!free_ptr)
53 		return -ENOMEM;
54 
55 	h = PTR_ALIGN(free_ptr, cache_line_size());
56 	h->root = RB_ROOT_CACHED;
57 	h->ops = ops;
58 	h->ops_arg = ops_arg;
59 	INIT_HLIST_NODE(&h->mn.hlist);
60 	spin_lock_init(&h->lock);
61 	h->mn.ops = &mn_opts;
62 	INIT_WORK(&h->del_work, handle_remove);
63 	INIT_LIST_HEAD(&h->del_list);
64 	INIT_LIST_HEAD(&h->lru_list);
65 	h->wq = wq;
66 	h->free_ptr = free_ptr;
67 
68 	ret = mmu_notifier_register(&h->mn, current->mm);
69 	if (ret) {
70 		kfree(free_ptr);
71 		return ret;
72 	}
73 
74 	*handler = h;
75 	return 0;
76 }
77 
78 void hfi1_mmu_rb_unregister(struct mmu_rb_handler *handler)
79 {
80 	struct mmu_rb_node *rbnode;
81 	struct rb_node *node;
82 	unsigned long flags;
83 	struct list_head del_list;
84 
85 	/* Prevent freeing of mm until we are completely finished. */
86 	mmgrab(handler->mn.mm);
87 
88 	/* Unregister first so we don't get any more notifications. */
89 	mmu_notifier_unregister(&handler->mn, handler->mn.mm);
90 
91 	/*
92 	 * Make sure the wq delete handler is finished running.  It will not
93 	 * be triggered once the mmu notifiers are unregistered above.
94 	 */
95 	flush_work(&handler->del_work);
96 
97 	INIT_LIST_HEAD(&del_list);
98 
99 	spin_lock_irqsave(&handler->lock, flags);
100 	while ((node = rb_first_cached(&handler->root))) {
101 		rbnode = rb_entry(node, struct mmu_rb_node, node);
102 		rb_erase_cached(node, &handler->root);
103 		/* move from LRU list to delete list */
104 		list_move(&rbnode->list, &del_list);
105 	}
106 	spin_unlock_irqrestore(&handler->lock, flags);
107 
108 	while (!list_empty(&del_list)) {
109 		rbnode = list_first_entry(&del_list, struct mmu_rb_node, list);
110 		list_del(&rbnode->list);
111 		kref_put(&rbnode->refcount, release_immediate);
112 	}
113 
114 	/* Now the mm may be freed. */
115 	mmdrop(handler->mn.mm);
116 
117 	kfree(handler->free_ptr);
118 }
119 
120 int hfi1_mmu_rb_insert(struct mmu_rb_handler *handler,
121 		       struct mmu_rb_node *mnode)
122 {
123 	struct mmu_rb_node *node;
124 	unsigned long flags;
125 	int ret = 0;
126 
127 	trace_hfi1_mmu_rb_insert(mnode);
128 
129 	if (current->mm != handler->mn.mm)
130 		return -EPERM;
131 
132 	spin_lock_irqsave(&handler->lock, flags);
133 	node = __mmu_rb_search(handler, mnode->addr, mnode->len);
134 	if (node) {
135 		ret = -EEXIST;
136 		goto unlock;
137 	}
138 	__mmu_int_rb_insert(mnode, &handler->root);
139 	list_add_tail(&mnode->list, &handler->lru_list);
140 	mnode->handler = handler;
141 unlock:
142 	spin_unlock_irqrestore(&handler->lock, flags);
143 	return ret;
144 }
145 
146 /* Caller must hold handler lock */
147 struct mmu_rb_node *hfi1_mmu_rb_get_first(struct mmu_rb_handler *handler,
148 					  unsigned long addr, unsigned long len)
149 {
150 	struct mmu_rb_node *node;
151 
152 	trace_hfi1_mmu_rb_search(addr, len);
153 	node = __mmu_int_rb_iter_first(&handler->root, addr, (addr + len) - 1);
154 	if (node)
155 		list_move_tail(&node->list, &handler->lru_list);
156 	return node;
157 }
158 
159 /* Caller must hold handler lock */
160 static struct mmu_rb_node *__mmu_rb_search(struct mmu_rb_handler *handler,
161 					   unsigned long addr,
162 					   unsigned long len)
163 {
164 	struct mmu_rb_node *node = NULL;
165 
166 	trace_hfi1_mmu_rb_search(addr, len);
167 	if (!handler->ops->filter) {
168 		node = __mmu_int_rb_iter_first(&handler->root, addr,
169 					       (addr + len) - 1);
170 	} else {
171 		for (node = __mmu_int_rb_iter_first(&handler->root, addr,
172 						    (addr + len) - 1);
173 		     node;
174 		     node = __mmu_int_rb_iter_next(node, addr,
175 						   (addr + len) - 1)) {
176 			if (handler->ops->filter(node, addr, len))
177 				return node;
178 		}
179 	}
180 	return node;
181 }
182 
183 /*
184  * Must NOT call while holding mnode->handler->lock.
185  * mnode->handler->ops->remove() may sleep and mnode->handler->lock is a
186  * spinlock.
187  */
188 static void release_immediate(struct kref *refcount)
189 {
190 	struct mmu_rb_node *mnode =
191 		container_of(refcount, struct mmu_rb_node, refcount);
192 	trace_hfi1_mmu_release_node(mnode);
193 	mnode->handler->ops->remove(mnode->handler->ops_arg, mnode);
194 }
195 
196 /* Caller must hold mnode->handler->lock */
197 static void release_nolock(struct kref *refcount)
198 {
199 	struct mmu_rb_node *mnode =
200 		container_of(refcount, struct mmu_rb_node, refcount);
201 	list_move(&mnode->list, &mnode->handler->del_list);
202 	queue_work(mnode->handler->wq, &mnode->handler->del_work);
203 }
204 
205 /*
206  * struct mmu_rb_node->refcount kref_put() callback.
207  * Adds mmu_rb_node to mmu_rb_node->handler->del_list and queues
208  * handler->del_work on handler->wq.
209  * Does not remove mmu_rb_node from handler->lru_list or handler->rb_root.
210  * Acquires mmu_rb_node->handler->lock; do not call while already holding
211  * handler->lock.
212  */
213 void hfi1_mmu_rb_release(struct kref *refcount)
214 {
215 	struct mmu_rb_node *mnode =
216 		container_of(refcount, struct mmu_rb_node, refcount);
217 	struct mmu_rb_handler *handler = mnode->handler;
218 	unsigned long flags;
219 
220 	spin_lock_irqsave(&handler->lock, flags);
221 	list_move(&mnode->list, &mnode->handler->del_list);
222 	spin_unlock_irqrestore(&handler->lock, flags);
223 	queue_work(handler->wq, &handler->del_work);
224 }
225 
226 void hfi1_mmu_rb_evict(struct mmu_rb_handler *handler, void *evict_arg)
227 {
228 	struct mmu_rb_node *rbnode, *ptr;
229 	struct list_head del_list;
230 	unsigned long flags;
231 	bool stop = false;
232 
233 	if (current->mm != handler->mn.mm)
234 		return;
235 
236 	INIT_LIST_HEAD(&del_list);
237 
238 	spin_lock_irqsave(&handler->lock, flags);
239 	list_for_each_entry_safe(rbnode, ptr, &handler->lru_list, list) {
240 		/* refcount == 1 implies mmu_rb_handler has only rbnode ref */
241 		if (kref_read(&rbnode->refcount) > 1)
242 			continue;
243 
244 		if (handler->ops->evict(handler->ops_arg, rbnode, evict_arg,
245 					&stop)) {
246 			__mmu_int_rb_remove(rbnode, &handler->root);
247 			/* move from LRU list to delete list */
248 			list_move(&rbnode->list, &del_list);
249 		}
250 		if (stop)
251 			break;
252 	}
253 	spin_unlock_irqrestore(&handler->lock, flags);
254 
255 	list_for_each_entry_safe(rbnode, ptr, &del_list, list) {
256 		trace_hfi1_mmu_rb_evict(rbnode);
257 		kref_put(&rbnode->refcount, release_immediate);
258 	}
259 }
260 
261 static int mmu_notifier_range_start(struct mmu_notifier *mn,
262 		const struct mmu_notifier_range *range)
263 {
264 	struct mmu_rb_handler *handler =
265 		container_of(mn, struct mmu_rb_handler, mn);
266 	struct rb_root_cached *root = &handler->root;
267 	struct mmu_rb_node *node, *ptr = NULL;
268 	unsigned long flags;
269 
270 	spin_lock_irqsave(&handler->lock, flags);
271 	for (node = __mmu_int_rb_iter_first(root, range->start, range->end-1);
272 	     node; node = ptr) {
273 		/* Guard against node removal. */
274 		ptr = __mmu_int_rb_iter_next(node, range->start,
275 					     range->end - 1);
276 		trace_hfi1_mmu_mem_invalidate(node);
277 		/* Remove from rb tree and lru_list. */
278 		__mmu_int_rb_remove(node, root);
279 		list_del_init(&node->list);
280 		kref_put(&node->refcount, release_nolock);
281 	}
282 	spin_unlock_irqrestore(&handler->lock, flags);
283 
284 	return 0;
285 }
286 
287 /*
288  * Work queue function to remove all nodes that have been queued up to
289  * be removed.  The key feature is that mm->mmap_lock is not being held
290  * and the remove callback can sleep while taking it, if needed.
291  */
292 static void handle_remove(struct work_struct *work)
293 {
294 	struct mmu_rb_handler *handler = container_of(work,
295 						struct mmu_rb_handler,
296 						del_work);
297 	struct list_head del_list;
298 	unsigned long flags;
299 	struct mmu_rb_node *node;
300 
301 	/* remove anything that is queued to get removed */
302 	spin_lock_irqsave(&handler->lock, flags);
303 	list_replace_init(&handler->del_list, &del_list);
304 	spin_unlock_irqrestore(&handler->lock, flags);
305 
306 	while (!list_empty(&del_list)) {
307 		node = list_first_entry(&del_list, struct mmu_rb_node, list);
308 		list_del(&node->list);
309 		trace_hfi1_mmu_release_node(node);
310 		handler->ops->remove(handler->ops_arg, node);
311 	}
312 }
313