forked from ~ljy/RK356X_SDK_RELEASE
blame | history | patch | commit | commitdiff
hc
2024-12-19 9370bb92b2d16684ee45cf24e879c93c509162da 
 kernel/drivers/infiniband/hw/hfi1/mmu_rb.c
....@@ -1,4 +1,5 @@
11 /*
2
+ * Copyright(c) 2020 Cornelis Networks, Inc.
23  * Copyright(c) 2016 - 2017 Intel Corporation.
34  *
45  * This file is provided under a dual BSD/GPLv2 license.  When using or
....@@ -48,36 +49,21 @@
4849 #include <linux/rculist.h>
4950 #include <linux/mmu_notifier.h>
5051 #include <linux/interval_tree_generic.h>
52
+#include <linux/sched/mm.h>
5153 
5254 #include "mmu_rb.h"
5355 #include "trace.h"
5456 
55
-struct mmu_rb_handler {
56
-	struct mmu_notifier mn;
57
-	struct rb_root_cached root;
58
-	void *ops_arg;
59
-	spinlock_t lock;        /* protect the RB tree */
60
-	struct mmu_rb_ops *ops;
61
-	struct mm_struct *mm;
62
-	struct list_head lru_list;
63
-	struct work_struct del_work;
64
-	struct list_head del_list;
65
-	struct workqueue_struct *wq;
66
-};
67
-
6857 static unsigned long mmu_node_start(struct mmu_rb_node *);
6958 static unsigned long mmu_node_last(struct mmu_rb_node *);
7059 static int mmu_notifier_range_start(struct mmu_notifier *,
71
-				     struct mm_struct *,
72
-				     unsigned long, unsigned long, bool);
60
+		const struct mmu_notifier_range *);
7361 static struct mmu_rb_node *__mmu_rb_search(struct mmu_rb_handler *,
7462 					   unsigned long, unsigned long);
75
-static void do_remove(struct mmu_rb_handler *handler,
76
-		      struct list_head *del_list);
63
+static void release_immediate(struct kref *refcount);
7764 static void handle_remove(struct work_struct *work);
7865 
7966 static const struct mmu_notifier_ops mn_opts = {
80
-	.flags = MMU_INVALIDATE_DOES_NOT_BLOCK,
8167 	.invalidate_range_start = mmu_notifier_range_start,
8268 };
8369 
....@@ -94,37 +80,36 @@
9480 	return PAGE_ALIGN(node->addr + node->len) - 1;
9581 }
9682 
97
-int hfi1_mmu_rb_register(void *ops_arg, struct mm_struct *mm,
83
+int hfi1_mmu_rb_register(void *ops_arg,
9884 			 struct mmu_rb_ops *ops,
9985 			 struct workqueue_struct *wq,
10086 			 struct mmu_rb_handler **handler)
10187 {
102
-	struct mmu_rb_handler *handlr;
88
+	struct mmu_rb_handler *h;
10389 	int ret;
10490 
105
-	handlr = kmalloc(sizeof(*handlr), GFP_KERNEL);
106
-	if (!handlr)
91
+	h = kzalloc(sizeof(*h), GFP_KERNEL);
92
+	if (!h)
10793 		return -ENOMEM;
10894 
109
-	handlr->root = RB_ROOT_CACHED;
110
-	handlr->ops = ops;
111
-	handlr->ops_arg = ops_arg;
112
-	INIT_HLIST_NODE(&handlr->mn.hlist);
113
-	spin_lock_init(&handlr->lock);
114
-	handlr->mn.ops = &mn_opts;
115
-	handlr->mm = mm;
116
-	INIT_WORK(&handlr->del_work, handle_remove);
117
-	INIT_LIST_HEAD(&handlr->del_list);
118
-	INIT_LIST_HEAD(&handlr->lru_list);
119
-	handlr->wq = wq;
95
+	h->root = RB_ROOT_CACHED;
96
+	h->ops = ops;
97
+	h->ops_arg = ops_arg;
98
+	INIT_HLIST_NODE(&h->mn.hlist);
99
+	spin_lock_init(&h->lock);
100
+	h->mn.ops = &mn_opts;
101
+	INIT_WORK(&h->del_work, handle_remove);
102
+	INIT_LIST_HEAD(&h->del_list);
103
+	INIT_LIST_HEAD(&h->lru_list);
104
+	h->wq = wq;
120105 
121
-	ret = mmu_notifier_register(&handlr->mn, handlr->mm);
106
+	ret = mmu_notifier_register(&h->mn, current->mm);
122107 	if (ret) {
123
-		kfree(handlr);
108
+		kfree(h);
124109 		return ret;
125110 	}
126111 
127
-	*handler = handlr;
112
+	*handler = h;
128113 	return 0;
129114 }
130115 
....@@ -135,8 +120,11 @@
135120 	unsigned long flags;
136121 	struct list_head del_list;
137122 
123
+	/* Prevent freeing of mm until we are completely finished. */
124
+	mmgrab(handler->mn.mm);
125
+
138126 	/* Unregister first so we don't get any more notifications. */
139
-	mmu_notifier_unregister(&handler->mn, handler->mm);
127
+	mmu_notifier_unregister(&handler->mn, handler->mn.mm);
140128 
141129 	/*
142130 	 * Make sure the wq delete handler is finished running.  It will not
....@@ -155,7 +143,14 @@
155143 	}
156144 	spin_unlock_irqrestore(&handler->lock, flags);
157145 
158
-	do_remove(handler, &del_list);
146
+	while (!list_empty(&del_list)) {
147
+		rbnode = list_first_entry(&del_list, struct mmu_rb_node, list);
148
+		list_del(&rbnode->list);
149
+		kref_put(&rbnode->refcount, release_immediate);
150
+	}
151
+
152
+	/* Now the mm may be freed. */
153
+	mmdrop(handler->mn.mm);
159154 
160155 	kfree(handler);
161156 }
....@@ -168,23 +163,35 @@
168163 	int ret = 0;
169164 
170165 	trace_hfi1_mmu_rb_insert(mnode->addr, mnode->len);
166
+
167
+	if (current->mm != handler->mn.mm)
168
+		return -EPERM;
169
+
171170 	spin_lock_irqsave(&handler->lock, flags);
172171 	node = __mmu_rb_search(handler, mnode->addr, mnode->len);
173172 	if (node) {
174
-		ret = -EINVAL;
173
+		ret = -EEXIST;
175174 		goto unlock;
176175 	}
177176 	__mmu_int_rb_insert(mnode, &handler->root);
178
-	list_add(&mnode->list, &handler->lru_list);
179
-
180
-	ret = handler->ops->insert(handler->ops_arg, mnode);
181
-	if (ret) {
182
-		__mmu_int_rb_remove(mnode, &handler->root);
183
-		list_del(&mnode->list); /* remove from LRU list */
184
-	}
177
+	list_add_tail(&mnode->list, &handler->lru_list);
178
+	mnode->handler = handler;
185179 unlock:
186180 	spin_unlock_irqrestore(&handler->lock, flags);
187181 	return ret;
182
+}
183
+
184
+/* Caller must hold handler lock */
185
+struct mmu_rb_node *hfi1_mmu_rb_get_first(struct mmu_rb_handler *handler,
186
+					  unsigned long addr, unsigned long len)
187
+{
188
+	struct mmu_rb_node *node;
189
+
190
+	trace_hfi1_mmu_rb_search(addr, len);
191
+	node = __mmu_int_rb_iter_first(&handler->root, addr, (addr + len) - 1);
192
+	if (node)
193
+		list_move_tail(&node->list, &handler->lru_list);
194
+	return node;
188195 }
189196 
190197 /* Caller must hold handler lock */
....@@ -211,27 +218,46 @@
211218 	return node;
212219 }
213220 
214
-bool hfi1_mmu_rb_remove_unless_exact(struct mmu_rb_handler *handler,
215
-				     unsigned long addr, unsigned long len,
216
-				     struct mmu_rb_node **rb_node)
221
+/*
222
+ * Must NOT call while holding mnode->handler->lock.
223
+ * mnode->handler->ops->remove() may sleep and mnode->handler->lock is a
224
+ * spinlock.
225
+ */
226
+static void release_immediate(struct kref *refcount)
217227 {
218
-	struct mmu_rb_node *node;
228
+	struct mmu_rb_node *mnode =
229
+		container_of(refcount, struct mmu_rb_node, refcount);
230
+	mnode->handler->ops->remove(mnode->handler->ops_arg, mnode);
231
+}
232
+
233
+/* Caller must hold mnode->handler->lock */
234
+static void release_nolock(struct kref *refcount)
235
+{
236
+	struct mmu_rb_node *mnode =
237
+		container_of(refcount, struct mmu_rb_node, refcount);
238
+	list_move(&mnode->list, &mnode->handler->del_list);
239
+	queue_work(mnode->handler->wq, &mnode->handler->del_work);
240
+}
241
+
242
+/*
243
+ * struct mmu_rb_node->refcount kref_put() callback.
244
+ * Adds mmu_rb_node to mmu_rb_node->handler->del_list and queues
245
+ * handler->del_work on handler->wq.
246
+ * Does not remove mmu_rb_node from handler->lru_list or handler->rb_root.
247
+ * Acquires mmu_rb_node->handler->lock; do not call while already holding
248
+ * handler->lock.
249
+ */
250
+void hfi1_mmu_rb_release(struct kref *refcount)
251
+{
252
+	struct mmu_rb_node *mnode =
253
+		container_of(refcount, struct mmu_rb_node, refcount);
254
+	struct mmu_rb_handler *handler = mnode->handler;
219255 	unsigned long flags;
220
-	bool ret = false;
221256 
222257 	spin_lock_irqsave(&handler->lock, flags);
223
-	node = __mmu_rb_search(handler, addr, len);
224
-	if (node) {
225
-		if (node->addr == addr && node->len == len)
226
-			goto unlock;
227
-		__mmu_int_rb_remove(node, &handler->root);
228
-		list_del(&node->list); /* remove from LRU list */
229
-		ret = true;
230
-	}
231
-unlock:
258
+	list_move(&mnode->list, &mnode->handler->del_list);
232259 	spin_unlock_irqrestore(&handler->lock, flags);
233
-	*rb_node = node;
234
-	return ret;
260
+	queue_work(handler->wq, &handler->del_work);
235261 }
236262 
237263 void hfi1_mmu_rb_evict(struct mmu_rb_handler *handler, void *evict_arg)
....@@ -241,11 +267,17 @@
241267 	unsigned long flags;
242268 	bool stop = false;
243269 
270
+	if (current->mm != handler->mn.mm)
271
+		return;
272
+
244273 	INIT_LIST_HEAD(&del_list);
245274 
246275 	spin_lock_irqsave(&handler->lock, flags);
247
-	list_for_each_entry_safe_reverse(rbnode, ptr, &handler->lru_list,
248
-					 list) {
276
+	list_for_each_entry_safe(rbnode, ptr, &handler->lru_list, list) {
277
+		/* refcount == 1 implies mmu_rb_handler has only rbnode ref */
278
+		if (kref_read(&rbnode->refcount) > 1)
279
+			continue;
280
+
249281 		if (handler->ops->evict(handler->ops_arg, rbnode, evict_arg,
250282 					&stop)) {
251283 			__mmu_int_rb_remove(rbnode, &handler->root);
....@@ -257,87 +289,40 @@
257289 	}
258290 	spin_unlock_irqrestore(&handler->lock, flags);
259291 
260
-	while (!list_empty(&del_list)) {
261
-		rbnode = list_first_entry(&del_list, struct mmu_rb_node, list);
262
-		list_del(&rbnode->list);
263
-		handler->ops->remove(handler->ops_arg, rbnode);
292
+	list_for_each_entry_safe(rbnode, ptr, &del_list, list) {
293
+		kref_put(&rbnode->refcount, release_immediate);
264294 	}
265295 }
266296 
267
-/*
268
- * It is up to the caller to ensure that this function does not race with the
269
- * mmu invalidate notifier which may be calling the users remove callback on
270
- * 'node'.
271
- */
272
-void hfi1_mmu_rb_remove(struct mmu_rb_handler *handler,
273
-			struct mmu_rb_node *node)
274
-{
275
-	unsigned long flags;
276
-
277
-	/* Validity of handler and node pointers has been checked by caller. */
278
-	trace_hfi1_mmu_rb_remove(node->addr, node->len);
279
-	spin_lock_irqsave(&handler->lock, flags);
280
-	__mmu_int_rb_remove(node, &handler->root);
281
-	list_del(&node->list); /* remove from LRU list */
282
-	spin_unlock_irqrestore(&handler->lock, flags);
283
-
284
-	handler->ops->remove(handler->ops_arg, node);
285
-}
286
-
287297 static int mmu_notifier_range_start(struct mmu_notifier *mn,
288
-				     struct mm_struct *mm,
289
-				     unsigned long start,
290
-				     unsigned long end,
291
-				     bool blockable)
298
+		const struct mmu_notifier_range *range)
292299 {
293300 	struct mmu_rb_handler *handler =
294301 		container_of(mn, struct mmu_rb_handler, mn);
295302 	struct rb_root_cached *root = &handler->root;
296303 	struct mmu_rb_node *node, *ptr = NULL;
297304 	unsigned long flags;
298
-	bool added = false;
299305 
300306 	spin_lock_irqsave(&handler->lock, flags);
301
-	for (node = __mmu_int_rb_iter_first(root, start, end - 1);
307
+	for (node = __mmu_int_rb_iter_first(root, range->start, range->end-1);
302308 	     node; node = ptr) {
303309 		/* Guard against node removal. */
304
-		ptr = __mmu_int_rb_iter_next(node, start, end - 1);
310
+		ptr = __mmu_int_rb_iter_next(node, range->start,
311
+					     range->end - 1);
305312 		trace_hfi1_mmu_mem_invalidate(node->addr, node->len);
306
-		if (handler->ops->invalidate(handler->ops_arg, node)) {
307
-			__mmu_int_rb_remove(node, root);
308
-			/* move from LRU list to delete list */
309
-			list_move(&node->list, &handler->del_list);
310
-			added = true;
311
-		}
313
+		/* Remove from rb tree and lru_list. */
314
+		__mmu_int_rb_remove(node, root);
315
+		list_del_init(&node->list);
316
+		kref_put(&node->refcount, release_nolock);
312317 	}
313318 	spin_unlock_irqrestore(&handler->lock, flags);
314
-
315
-	if (added)
316
-		queue_work(handler->wq, &handler->del_work);
317319 
318320 	return 0;
319321 }
320322 
321323 /*
322
- * Call the remove function for the given handler and the list.  This
323
- * is expected to be called with a delete list extracted from handler.
324
- * The caller should not be holding the handler lock.
325
- */
326
-static void do_remove(struct mmu_rb_handler *handler,
327
-		      struct list_head *del_list)
328
-{
329
-	struct mmu_rb_node *node;
330
-
331
-	while (!list_empty(del_list)) {
332
-		node = list_first_entry(del_list, struct mmu_rb_node, list);
333
-		list_del(&node->list);
334
-		handler->ops->remove(handler->ops_arg, node);
335
-	}
336
-}
337
-
338
-/*
339324  * Work queue function to remove all nodes that have been queued up to
340
- * be removed.  The key feature is that mm->mmap_sem is not being held
325
+ * be removed.  The key feature is that mm->mmap_lock is not being held
341326  * and the remove callback can sleep while taking it, if needed.
342327  */
343328 static void handle_remove(struct work_struct *work)
....@@ -347,11 +332,16 @@
347332 						del_work);
348333 	struct list_head del_list;
349334 	unsigned long flags;
335
+	struct mmu_rb_node *node;
350336 
351337 	/* remove anything that is queued to get removed */
352338 	spin_lock_irqsave(&handler->lock, flags);
353339 	list_replace_init(&handler->del_list, &del_list);
354340 	spin_unlock_irqrestore(&handler->lock, flags);
355341 
356
-	do_remove(handler, &del_list);
342
+	while (!list_empty(&del_list)) {
343
+		node = list_first_entry(&del_list, struct mmu_rb_node, list);
344
+		list_del(&node->list);
345
+		handler->ops->remove(handler->ops_arg, node);
346
+	}
357347 }