xref: /openbmc/linux/kernel/bpf/lpm_trie.c (revision ae213c44)
1 /*
2  * Longest prefix match list implementation
3  *
4  * Copyright (c) 2016,2017 Daniel Mack
5  * Copyright (c) 2016 David Herrmann
6  *
7  * This file is subject to the terms and conditions of version 2 of the GNU
8  * General Public License.  See the file COPYING in the main directory of the
9  * Linux distribution for more details.
10  */
11 
12 #include <linux/bpf.h>
13 #include <linux/btf.h>
14 #include <linux/err.h>
15 #include <linux/slab.h>
16 #include <linux/spinlock.h>
17 #include <linux/vmalloc.h>
18 #include <net/ipv6.h>
19 #include <uapi/linux/btf.h>
20 
21 /* Intermediate node */
22 #define LPM_TREE_NODE_FLAG_IM BIT(0)
23 
24 struct lpm_trie_node;
25 
26 struct lpm_trie_node {
27 	struct rcu_head rcu;
28 	struct lpm_trie_node __rcu	*child[2];
29 	u32				prefixlen;
30 	u32				flags;
31 	u8				data[0];
32 };
33 
34 struct lpm_trie {
35 	struct bpf_map			map;
36 	struct lpm_trie_node __rcu	*root;
37 	size_t				n_entries;
38 	size_t				max_prefixlen;
39 	size_t				data_size;
40 	raw_spinlock_t			lock;
41 };
42 
43 /* This trie implements a longest prefix match algorithm that can be used to
44  * match IP addresses to a stored set of ranges.
45  *
46  * Data stored in @data of struct bpf_lpm_key and struct lpm_trie_node is
47  * interpreted as big endian, so data[0] stores the most significant byte.
48  *
49  * Match ranges are internally stored in instances of struct lpm_trie_node
50  * which each contain their prefix length as well as two pointers that may
51  * lead to more nodes containing more specific matches. Each node also stores
52  * a value that is defined by and returned to userspace via the update_elem
53  * and lookup functions.
54  *
55  * For instance, let's start with a trie that was created with a prefix length
56  * of 32, so it can be used for IPv4 addresses, and one single element that
57  * matches 192.168.0.0/16. The data array would hence contain
58  * [0xc0, 0xa8, 0x00, 0x00] in big-endian notation. This documentation will
59  * stick to IP-address notation for readability though.
60  *
61  * As the trie is empty initially, the new node (1) will be places as root
62  * node, denoted as (R) in the example below. As there are no other node, both
63  * child pointers are %NULL.
64  *
65  *              +----------------+
66  *              |       (1)  (R) |
67  *              | 192.168.0.0/16 |
68  *              |    value: 1    |
69  *              |   [0]    [1]   |
70  *              +----------------+
71  *
72  * Next, let's add a new node (2) matching 192.168.0.0/24. As there is already
73  * a node with the same data and a smaller prefix (ie, a less specific one),
74  * node (2) will become a child of (1). In child index depends on the next bit
75  * that is outside of what (1) matches, and that bit is 0, so (2) will be
76  * child[0] of (1):
77  *
78  *              +----------------+
79  *              |       (1)  (R) |
80  *              | 192.168.0.0/16 |
81  *              |    value: 1    |
82  *              |   [0]    [1]   |
83  *              +----------------+
84  *                   |
85  *    +----------------+
86  *    |       (2)      |
87  *    | 192.168.0.0/24 |
88  *    |    value: 2    |
89  *    |   [0]    [1]   |
90  *    +----------------+
91  *
92  * The child[1] slot of (1) could be filled with another node which has bit #17
93  * (the next bit after the ones that (1) matches on) set to 1. For instance,
94  * 192.168.128.0/24:
95  *
96  *              +----------------+
97  *              |       (1)  (R) |
98  *              | 192.168.0.0/16 |
99  *              |    value: 1    |
100  *              |   [0]    [1]   |
101  *              +----------------+
102  *                   |      |
103  *    +----------------+  +------------------+
104  *    |       (2)      |  |        (3)       |
105  *    | 192.168.0.0/24 |  | 192.168.128.0/24 |
106  *    |    value: 2    |  |     value: 3     |
107  *    |   [0]    [1]   |  |    [0]    [1]    |
108  *    +----------------+  +------------------+
109  *
110  * Let's add another node (4) to the game for 192.168.1.0/24. In order to place
111  * it, node (1) is looked at first, and because (4) of the semantics laid out
112  * above (bit #17 is 0), it would normally be attached to (1) as child[0].
113  * However, that slot is already allocated, so a new node is needed in between.
114  * That node does not have a value attached to it and it will never be
115  * returned to users as result of a lookup. It is only there to differentiate
116  * the traversal further. It will get a prefix as wide as necessary to
117  * distinguish its two children:
118  *
119  *                      +----------------+
120  *                      |       (1)  (R) |
121  *                      | 192.168.0.0/16 |
122  *                      |    value: 1    |
123  *                      |   [0]    [1]   |
124  *                      +----------------+
125  *                           |      |
126  *            +----------------+  +------------------+
127  *            |       (4)  (I) |  |        (3)       |
128  *            | 192.168.0.0/23 |  | 192.168.128.0/24 |
129  *            |    value: ---  |  |     value: 3     |
130  *            |   [0]    [1]   |  |    [0]    [1]    |
131  *            +----------------+  +------------------+
132  *                 |      |
133  *  +----------------+  +----------------+
134  *  |       (2)      |  |       (5)      |
135  *  | 192.168.0.0/24 |  | 192.168.1.0/24 |
136  *  |    value: 2    |  |     value: 5   |
137  *  |   [0]    [1]   |  |   [0]    [1]   |
138  *  +----------------+  +----------------+
139  *
140  * 192.168.1.1/32 would be a child of (5) etc.
141  *
142  * An intermediate node will be turned into a 'real' node on demand. In the
143  * example above, (4) would be re-used if 192.168.0.0/23 is added to the trie.
144  *
145  * A fully populated trie would have a height of 32 nodes, as the trie was
146  * created with a prefix length of 32.
147  *
148  * The lookup starts at the root node. If the current node matches and if there
149  * is a child that can be used to become more specific, the trie is traversed
150  * downwards. The last node in the traversal that is a non-intermediate one is
151  * returned.
152  */
153 
154 static inline int extract_bit(const u8 *data, size_t index)
155 {
156 	return !!(data[index / 8] & (1 << (7 - (index % 8))));
157 }
158 
159 /**
160  * longest_prefix_match() - determine the longest prefix
161  * @trie:	The trie to get internal sizes from
162  * @node:	The node to operate on
163  * @key:	The key to compare to @node
164  *
165  * Determine the longest prefix of @node that matches the bits in @key.
166  */
167 static size_t longest_prefix_match(const struct lpm_trie *trie,
168 				   const struct lpm_trie_node *node,
169 				   const struct bpf_lpm_trie_key *key)
170 {
171 	u32 limit = min(node->prefixlen, key->prefixlen);
172 	u32 prefixlen = 0, i = 0;
173 
174 	BUILD_BUG_ON(offsetof(struct lpm_trie_node, data) % sizeof(u32));
175 	BUILD_BUG_ON(offsetof(struct bpf_lpm_trie_key, data) % sizeof(u32));
176 
177 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && defined(CONFIG_64BIT)
178 
179 	/* data_size >= 16 has very small probability.
180 	 * We do not use a loop for optimal code generation.
181 	 */
182 	if (trie->data_size >= 8) {
183 		u64 diff = be64_to_cpu(*(__be64 *)node->data ^
184 				       *(__be64 *)key->data);
185 
186 		prefixlen = 64 - fls64(diff);
187 		if (prefixlen >= limit)
188 			return limit;
189 		if (diff)
190 			return prefixlen;
191 		i = 8;
192 	}
193 #endif
194 
195 	while (trie->data_size >= i + 4) {
196 		u32 diff = be32_to_cpu(*(__be32 *)&node->data[i] ^
197 				       *(__be32 *)&key->data[i]);
198 
199 		prefixlen += 32 - fls(diff);
200 		if (prefixlen >= limit)
201 			return limit;
202 		if (diff)
203 			return prefixlen;
204 		i += 4;
205 	}
206 
207 	if (trie->data_size >= i + 2) {
208 		u16 diff = be16_to_cpu(*(__be16 *)&node->data[i] ^
209 				       *(__be16 *)&key->data[i]);
210 
211 		prefixlen += 16 - fls(diff);
212 		if (prefixlen >= limit)
213 			return limit;
214 		if (diff)
215 			return prefixlen;
216 		i += 2;
217 	}
218 
219 	if (trie->data_size >= i + 1) {
220 		prefixlen += 8 - fls(node->data[i] ^ key->data[i]);
221 
222 		if (prefixlen >= limit)
223 			return limit;
224 	}
225 
226 	return prefixlen;
227 }
228 
229 /* Called from syscall or from eBPF program */
230 static void *trie_lookup_elem(struct bpf_map *map, void *_key)
231 {
232 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
233 	struct lpm_trie_node *node, *found = NULL;
234 	struct bpf_lpm_trie_key *key = _key;
235 
236 	/* Start walking the trie from the root node ... */
237 
238 	for (node = rcu_dereference(trie->root); node;) {
239 		unsigned int next_bit;
240 		size_t matchlen;
241 
242 		/* Determine the longest prefix of @node that matches @key.
243 		 * If it's the maximum possible prefix for this trie, we have
244 		 * an exact match and can return it directly.
245 		 */
246 		matchlen = longest_prefix_match(trie, node, key);
247 		if (matchlen == trie->max_prefixlen) {
248 			found = node;
249 			break;
250 		}
251 
252 		/* If the number of bits that match is smaller than the prefix
253 		 * length of @node, bail out and return the node we have seen
254 		 * last in the traversal (ie, the parent).
255 		 */
256 		if (matchlen < node->prefixlen)
257 			break;
258 
259 		/* Consider this node as return candidate unless it is an
260 		 * artificially added intermediate one.
261 		 */
262 		if (!(node->flags & LPM_TREE_NODE_FLAG_IM))
263 			found = node;
264 
265 		/* If the node match is fully satisfied, let's see if we can
266 		 * become more specific. Determine the next bit in the key and
267 		 * traverse down.
268 		 */
269 		next_bit = extract_bit(key->data, node->prefixlen);
270 		node = rcu_dereference(node->child[next_bit]);
271 	}
272 
273 	if (!found)
274 		return NULL;
275 
276 	return found->data + trie->data_size;
277 }
278 
279 static struct lpm_trie_node *lpm_trie_node_alloc(const struct lpm_trie *trie,
280 						 const void *value)
281 {
282 	struct lpm_trie_node *node;
283 	size_t size = sizeof(struct lpm_trie_node) + trie->data_size;
284 
285 	if (value)
286 		size += trie->map.value_size;
287 
288 	node = kmalloc_node(size, GFP_ATOMIC | __GFP_NOWARN,
289 			    trie->map.numa_node);
290 	if (!node)
291 		return NULL;
292 
293 	node->flags = 0;
294 
295 	if (value)
296 		memcpy(node->data + trie->data_size, value,
297 		       trie->map.value_size);
298 
299 	return node;
300 }
301 
302 /* Called from syscall or from eBPF program */
303 static int trie_update_elem(struct bpf_map *map,
304 			    void *_key, void *value, u64 flags)
305 {
306 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
307 	struct lpm_trie_node *node, *im_node = NULL, *new_node = NULL;
308 	struct lpm_trie_node __rcu **slot;
309 	struct bpf_lpm_trie_key *key = _key;
310 	unsigned long irq_flags;
311 	unsigned int next_bit;
312 	size_t matchlen = 0;
313 	int ret = 0;
314 
315 	if (unlikely(flags > BPF_EXIST))
316 		return -EINVAL;
317 
318 	if (key->prefixlen > trie->max_prefixlen)
319 		return -EINVAL;
320 
321 	raw_spin_lock_irqsave(&trie->lock, irq_flags);
322 
323 	/* Allocate and fill a new node */
324 
325 	if (trie->n_entries == trie->map.max_entries) {
326 		ret = -ENOSPC;
327 		goto out;
328 	}
329 
330 	new_node = lpm_trie_node_alloc(trie, value);
331 	if (!new_node) {
332 		ret = -ENOMEM;
333 		goto out;
334 	}
335 
336 	trie->n_entries++;
337 
338 	new_node->prefixlen = key->prefixlen;
339 	RCU_INIT_POINTER(new_node->child[0], NULL);
340 	RCU_INIT_POINTER(new_node->child[1], NULL);
341 	memcpy(new_node->data, key->data, trie->data_size);
342 
343 	/* Now find a slot to attach the new node. To do that, walk the tree
344 	 * from the root and match as many bits as possible for each node until
345 	 * we either find an empty slot or a slot that needs to be replaced by
346 	 * an intermediate node.
347 	 */
348 	slot = &trie->root;
349 
350 	while ((node = rcu_dereference_protected(*slot,
351 					lockdep_is_held(&trie->lock)))) {
352 		matchlen = longest_prefix_match(trie, node, key);
353 
354 		if (node->prefixlen != matchlen ||
355 		    node->prefixlen == key->prefixlen ||
356 		    node->prefixlen == trie->max_prefixlen)
357 			break;
358 
359 		next_bit = extract_bit(key->data, node->prefixlen);
360 		slot = &node->child[next_bit];
361 	}
362 
363 	/* If the slot is empty (a free child pointer or an empty root),
364 	 * simply assign the @new_node to that slot and be done.
365 	 */
366 	if (!node) {
367 		rcu_assign_pointer(*slot, new_node);
368 		goto out;
369 	}
370 
371 	/* If the slot we picked already exists, replace it with @new_node
372 	 * which already has the correct data array set.
373 	 */
374 	if (node->prefixlen == matchlen) {
375 		new_node->child[0] = node->child[0];
376 		new_node->child[1] = node->child[1];
377 
378 		if (!(node->flags & LPM_TREE_NODE_FLAG_IM))
379 			trie->n_entries--;
380 
381 		rcu_assign_pointer(*slot, new_node);
382 		kfree_rcu(node, rcu);
383 
384 		goto out;
385 	}
386 
387 	/* If the new node matches the prefix completely, it must be inserted
388 	 * as an ancestor. Simply insert it between @node and *@slot.
389 	 */
390 	if (matchlen == key->prefixlen) {
391 		next_bit = extract_bit(node->data, matchlen);
392 		rcu_assign_pointer(new_node->child[next_bit], node);
393 		rcu_assign_pointer(*slot, new_node);
394 		goto out;
395 	}
396 
397 	im_node = lpm_trie_node_alloc(trie, NULL);
398 	if (!im_node) {
399 		ret = -ENOMEM;
400 		goto out;
401 	}
402 
403 	im_node->prefixlen = matchlen;
404 	im_node->flags |= LPM_TREE_NODE_FLAG_IM;
405 	memcpy(im_node->data, node->data, trie->data_size);
406 
407 	/* Now determine which child to install in which slot */
408 	if (extract_bit(key->data, matchlen)) {
409 		rcu_assign_pointer(im_node->child[0], node);
410 		rcu_assign_pointer(im_node->child[1], new_node);
411 	} else {
412 		rcu_assign_pointer(im_node->child[0], new_node);
413 		rcu_assign_pointer(im_node->child[1], node);
414 	}
415 
416 	/* Finally, assign the intermediate node to the determined spot */
417 	rcu_assign_pointer(*slot, im_node);
418 
419 out:
420 	if (ret) {
421 		if (new_node)
422 			trie->n_entries--;
423 
424 		kfree(new_node);
425 		kfree(im_node);
426 	}
427 
428 	raw_spin_unlock_irqrestore(&trie->lock, irq_flags);
429 
430 	return ret;
431 }
432 
433 /* Called from syscall or from eBPF program */
434 static int trie_delete_elem(struct bpf_map *map, void *_key)
435 {
436 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
437 	struct bpf_lpm_trie_key *key = _key;
438 	struct lpm_trie_node __rcu **trim, **trim2;
439 	struct lpm_trie_node *node, *parent;
440 	unsigned long irq_flags;
441 	unsigned int next_bit;
442 	size_t matchlen = 0;
443 	int ret = 0;
444 
445 	if (key->prefixlen > trie->max_prefixlen)
446 		return -EINVAL;
447 
448 	raw_spin_lock_irqsave(&trie->lock, irq_flags);
449 
450 	/* Walk the tree looking for an exact key/length match and keeping
451 	 * track of the path we traverse.  We will need to know the node
452 	 * we wish to delete, and the slot that points to the node we want
453 	 * to delete.  We may also need to know the nodes parent and the
454 	 * slot that contains it.
455 	 */
456 	trim = &trie->root;
457 	trim2 = trim;
458 	parent = NULL;
459 	while ((node = rcu_dereference_protected(
460 		       *trim, lockdep_is_held(&trie->lock)))) {
461 		matchlen = longest_prefix_match(trie, node, key);
462 
463 		if (node->prefixlen != matchlen ||
464 		    node->prefixlen == key->prefixlen)
465 			break;
466 
467 		parent = node;
468 		trim2 = trim;
469 		next_bit = extract_bit(key->data, node->prefixlen);
470 		trim = &node->child[next_bit];
471 	}
472 
473 	if (!node || node->prefixlen != key->prefixlen ||
474 	    node->prefixlen != matchlen ||
475 	    (node->flags & LPM_TREE_NODE_FLAG_IM)) {
476 		ret = -ENOENT;
477 		goto out;
478 	}
479 
480 	trie->n_entries--;
481 
482 	/* If the node we are removing has two children, simply mark it
483 	 * as intermediate and we are done.
484 	 */
485 	if (rcu_access_pointer(node->child[0]) &&
486 	    rcu_access_pointer(node->child[1])) {
487 		node->flags |= LPM_TREE_NODE_FLAG_IM;
488 		goto out;
489 	}
490 
491 	/* If the parent of the node we are about to delete is an intermediate
492 	 * node, and the deleted node doesn't have any children, we can delete
493 	 * the intermediate parent as well and promote its other child
494 	 * up the tree.  Doing this maintains the invariant that all
495 	 * intermediate nodes have exactly 2 children and that there are no
496 	 * unnecessary intermediate nodes in the tree.
497 	 */
498 	if (parent && (parent->flags & LPM_TREE_NODE_FLAG_IM) &&
499 	    !node->child[0] && !node->child[1]) {
500 		if (node == rcu_access_pointer(parent->child[0]))
501 			rcu_assign_pointer(
502 				*trim2, rcu_access_pointer(parent->child[1]));
503 		else
504 			rcu_assign_pointer(
505 				*trim2, rcu_access_pointer(parent->child[0]));
506 		kfree_rcu(parent, rcu);
507 		kfree_rcu(node, rcu);
508 		goto out;
509 	}
510 
511 	/* The node we are removing has either zero or one child. If there
512 	 * is a child, move it into the removed node's slot then delete
513 	 * the node.  Otherwise just clear the slot and delete the node.
514 	 */
515 	if (node->child[0])
516 		rcu_assign_pointer(*trim, rcu_access_pointer(node->child[0]));
517 	else if (node->child[1])
518 		rcu_assign_pointer(*trim, rcu_access_pointer(node->child[1]));
519 	else
520 		RCU_INIT_POINTER(*trim, NULL);
521 	kfree_rcu(node, rcu);
522 
523 out:
524 	raw_spin_unlock_irqrestore(&trie->lock, irq_flags);
525 
526 	return ret;
527 }
528 
529 #define LPM_DATA_SIZE_MAX	256
530 #define LPM_DATA_SIZE_MIN	1
531 
532 #define LPM_VAL_SIZE_MAX	(KMALLOC_MAX_SIZE - LPM_DATA_SIZE_MAX - \
533 				 sizeof(struct lpm_trie_node))
534 #define LPM_VAL_SIZE_MIN	1
535 
536 #define LPM_KEY_SIZE(X)		(sizeof(struct bpf_lpm_trie_key) + (X))
537 #define LPM_KEY_SIZE_MAX	LPM_KEY_SIZE(LPM_DATA_SIZE_MAX)
538 #define LPM_KEY_SIZE_MIN	LPM_KEY_SIZE(LPM_DATA_SIZE_MIN)
539 
540 #define LPM_CREATE_FLAG_MASK	(BPF_F_NO_PREALLOC | BPF_F_NUMA_NODE |	\
541 				 BPF_F_ACCESS_MASK)
542 
543 static struct bpf_map *trie_alloc(union bpf_attr *attr)
544 {
545 	struct lpm_trie *trie;
546 	u64 cost = sizeof(*trie), cost_per_node;
547 	int ret;
548 
549 	if (!capable(CAP_SYS_ADMIN))
550 		return ERR_PTR(-EPERM);
551 
552 	/* check sanity of attributes */
553 	if (attr->max_entries == 0 ||
554 	    !(attr->map_flags & BPF_F_NO_PREALLOC) ||
555 	    attr->map_flags & ~LPM_CREATE_FLAG_MASK ||
556 	    !bpf_map_flags_access_ok(attr->map_flags) ||
557 	    attr->key_size < LPM_KEY_SIZE_MIN ||
558 	    attr->key_size > LPM_KEY_SIZE_MAX ||
559 	    attr->value_size < LPM_VAL_SIZE_MIN ||
560 	    attr->value_size > LPM_VAL_SIZE_MAX)
561 		return ERR_PTR(-EINVAL);
562 
563 	trie = kzalloc(sizeof(*trie), GFP_USER | __GFP_NOWARN);
564 	if (!trie)
565 		return ERR_PTR(-ENOMEM);
566 
567 	/* copy mandatory map attributes */
568 	bpf_map_init_from_attr(&trie->map, attr);
569 	trie->data_size = attr->key_size -
570 			  offsetof(struct bpf_lpm_trie_key, data);
571 	trie->max_prefixlen = trie->data_size * 8;
572 
573 	cost_per_node = sizeof(struct lpm_trie_node) +
574 			attr->value_size + trie->data_size;
575 	cost += (u64) attr->max_entries * cost_per_node;
576 	if (cost >= U32_MAX - PAGE_SIZE) {
577 		ret = -E2BIG;
578 		goto out_err;
579 	}
580 
581 	trie->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
582 
583 	ret = bpf_map_precharge_memlock(trie->map.pages);
584 	if (ret)
585 		goto out_err;
586 
587 	raw_spin_lock_init(&trie->lock);
588 
589 	return &trie->map;
590 out_err:
591 	kfree(trie);
592 	return ERR_PTR(ret);
593 }
594 
595 static void trie_free(struct bpf_map *map)
596 {
597 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
598 	struct lpm_trie_node __rcu **slot;
599 	struct lpm_trie_node *node;
600 
601 	/* Wait for outstanding programs to complete
602 	 * update/lookup/delete/get_next_key and free the trie.
603 	 */
604 	synchronize_rcu();
605 
606 	/* Always start at the root and walk down to a node that has no
607 	 * children. Then free that node, nullify its reference in the parent
608 	 * and start over.
609 	 */
610 
611 	for (;;) {
612 		slot = &trie->root;
613 
614 		for (;;) {
615 			node = rcu_dereference_protected(*slot, 1);
616 			if (!node)
617 				goto out;
618 
619 			if (rcu_access_pointer(node->child[0])) {
620 				slot = &node->child[0];
621 				continue;
622 			}
623 
624 			if (rcu_access_pointer(node->child[1])) {
625 				slot = &node->child[1];
626 				continue;
627 			}
628 
629 			kfree(node);
630 			RCU_INIT_POINTER(*slot, NULL);
631 			break;
632 		}
633 	}
634 
635 out:
636 	kfree(trie);
637 }
638 
639 static int trie_get_next_key(struct bpf_map *map, void *_key, void *_next_key)
640 {
641 	struct lpm_trie_node *node, *next_node = NULL, *parent, *search_root;
642 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
643 	struct bpf_lpm_trie_key *key = _key, *next_key = _next_key;
644 	struct lpm_trie_node **node_stack = NULL;
645 	int err = 0, stack_ptr = -1;
646 	unsigned int next_bit;
647 	size_t matchlen;
648 
649 	/* The get_next_key follows postorder. For the 4 node example in
650 	 * the top of this file, the trie_get_next_key() returns the following
651 	 * one after another:
652 	 *   192.168.0.0/24
653 	 *   192.168.1.0/24
654 	 *   192.168.128.0/24
655 	 *   192.168.0.0/16
656 	 *
657 	 * The idea is to return more specific keys before less specific ones.
658 	 */
659 
660 	/* Empty trie */
661 	search_root = rcu_dereference(trie->root);
662 	if (!search_root)
663 		return -ENOENT;
664 
665 	/* For invalid key, find the leftmost node in the trie */
666 	if (!key || key->prefixlen > trie->max_prefixlen)
667 		goto find_leftmost;
668 
669 	node_stack = kmalloc_array(trie->max_prefixlen,
670 				   sizeof(struct lpm_trie_node *),
671 				   GFP_ATOMIC | __GFP_NOWARN);
672 	if (!node_stack)
673 		return -ENOMEM;
674 
675 	/* Try to find the exact node for the given key */
676 	for (node = search_root; node;) {
677 		node_stack[++stack_ptr] = node;
678 		matchlen = longest_prefix_match(trie, node, key);
679 		if (node->prefixlen != matchlen ||
680 		    node->prefixlen == key->prefixlen)
681 			break;
682 
683 		next_bit = extract_bit(key->data, node->prefixlen);
684 		node = rcu_dereference(node->child[next_bit]);
685 	}
686 	if (!node || node->prefixlen != key->prefixlen ||
687 	    (node->flags & LPM_TREE_NODE_FLAG_IM))
688 		goto find_leftmost;
689 
690 	/* The node with the exactly-matching key has been found,
691 	 * find the first node in postorder after the matched node.
692 	 */
693 	node = node_stack[stack_ptr];
694 	while (stack_ptr > 0) {
695 		parent = node_stack[stack_ptr - 1];
696 		if (rcu_dereference(parent->child[0]) == node) {
697 			search_root = rcu_dereference(parent->child[1]);
698 			if (search_root)
699 				goto find_leftmost;
700 		}
701 		if (!(parent->flags & LPM_TREE_NODE_FLAG_IM)) {
702 			next_node = parent;
703 			goto do_copy;
704 		}
705 
706 		node = parent;
707 		stack_ptr--;
708 	}
709 
710 	/* did not find anything */
711 	err = -ENOENT;
712 	goto free_stack;
713 
714 find_leftmost:
715 	/* Find the leftmost non-intermediate node, all intermediate nodes
716 	 * have exact two children, so this function will never return NULL.
717 	 */
718 	for (node = search_root; node;) {
719 		if (!(node->flags & LPM_TREE_NODE_FLAG_IM))
720 			next_node = node;
721 		node = rcu_dereference(node->child[0]);
722 	}
723 do_copy:
724 	next_key->prefixlen = next_node->prefixlen;
725 	memcpy((void *)next_key + offsetof(struct bpf_lpm_trie_key, data),
726 	       next_node->data, trie->data_size);
727 free_stack:
728 	kfree(node_stack);
729 	return err;
730 }
731 
732 static int trie_check_btf(const struct bpf_map *map,
733 			  const struct btf *btf,
734 			  const struct btf_type *key_type,
735 			  const struct btf_type *value_type)
736 {
737 	/* Keys must have struct bpf_lpm_trie_key embedded. */
738 	return BTF_INFO_KIND(key_type->info) != BTF_KIND_STRUCT ?
739 	       -EINVAL : 0;
740 }
741 
742 const struct bpf_map_ops trie_map_ops = {
743 	.map_alloc = trie_alloc,
744 	.map_free = trie_free,
745 	.map_get_next_key = trie_get_next_key,
746 	.map_lookup_elem = trie_lookup_elem,
747 	.map_update_elem = trie_update_elem,
748 	.map_delete_elem = trie_delete_elem,
749 	.map_check_btf = trie_check_btf,
750 };
751