xref: /linux/kernel/bpf/lpm_trie.c (revision 5e3992fe72748ed3892be876f09d4d990548b7af)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Longest prefix match list implementation
4  *
5  * Copyright (c) 2016,2017 Daniel Mack
6  * Copyright (c) 2016 David Herrmann
7  */
8 
9 #include <linux/bpf.h>
10 #include <linux/btf.h>
11 #include <linux/err.h>
12 #include <linux/slab.h>
13 #include <linux/spinlock.h>
14 #include <linux/vmalloc.h>
15 #include <net/ipv6.h>
16 #include <uapi/linux/btf.h>
17 #include <linux/btf_ids.h>
18 
19 /* Intermediate node */
20 #define LPM_TREE_NODE_FLAG_IM BIT(0)
21 
22 struct lpm_trie_node;
23 
24 struct lpm_trie_node {
25 	struct rcu_head rcu;
26 	struct lpm_trie_node __rcu	*child[2];
27 	u32				prefixlen;
28 	u32				flags;
29 	u8				data[];
30 };
31 
32 struct lpm_trie {
33 	struct bpf_map			map;
34 	struct lpm_trie_node __rcu	*root;
35 	size_t				n_entries;
36 	size_t				max_prefixlen;
37 	size_t				data_size;
38 	spinlock_t			lock;
39 };
40 
41 /* This trie implements a longest prefix match algorithm that can be used to
42  * match IP addresses to a stored set of ranges.
43  *
44  * Data stored in @data of struct bpf_lpm_key and struct lpm_trie_node is
45  * interpreted as big endian, so data[0] stores the most significant byte.
46  *
47  * Match ranges are internally stored in instances of struct lpm_trie_node
48  * which each contain their prefix length as well as two pointers that may
49  * lead to more nodes containing more specific matches. Each node also stores
50  * a value that is defined by and returned to userspace via the update_elem
51  * and lookup functions.
52  *
53  * For instance, let's start with a trie that was created with a prefix length
54  * of 32, so it can be used for IPv4 addresses, and one single element that
55  * matches 192.168.0.0/16. The data array would hence contain
56  * [0xc0, 0xa8, 0x00, 0x00] in big-endian notation. This documentation will
57  * stick to IP-address notation for readability though.
58  *
59  * As the trie is empty initially, the new node (1) will be places as root
60  * node, denoted as (R) in the example below. As there are no other node, both
61  * child pointers are %NULL.
62  *
63  *              +----------------+
64  *              |       (1)  (R) |
65  *              | 192.168.0.0/16 |
66  *              |    value: 1    |
67  *              |   [0]    [1]   |
68  *              +----------------+
69  *
70  * Next, let's add a new node (2) matching 192.168.0.0/24. As there is already
71  * a node with the same data and a smaller prefix (ie, a less specific one),
72  * node (2) will become a child of (1). In child index depends on the next bit
73  * that is outside of what (1) matches, and that bit is 0, so (2) will be
74  * child[0] of (1):
75  *
76  *              +----------------+
77  *              |       (1)  (R) |
78  *              | 192.168.0.0/16 |
79  *              |    value: 1    |
80  *              |   [0]    [1]   |
81  *              +----------------+
82  *                   |
83  *    +----------------+
84  *    |       (2)      |
85  *    | 192.168.0.0/24 |
86  *    |    value: 2    |
87  *    |   [0]    [1]   |
88  *    +----------------+
89  *
90  * The child[1] slot of (1) could be filled with another node which has bit #17
91  * (the next bit after the ones that (1) matches on) set to 1. For instance,
92  * 192.168.128.0/24:
93  *
94  *              +----------------+
95  *              |       (1)  (R) |
96  *              | 192.168.0.0/16 |
97  *              |    value: 1    |
98  *              |   [0]    [1]   |
99  *              +----------------+
100  *                   |      |
101  *    +----------------+  +------------------+
102  *    |       (2)      |  |        (3)       |
103  *    | 192.168.0.0/24 |  | 192.168.128.0/24 |
104  *    |    value: 2    |  |     value: 3     |
105  *    |   [0]    [1]   |  |    [0]    [1]    |
106  *    +----------------+  +------------------+
107  *
108  * Let's add another node (4) to the game for 192.168.1.0/24. In order to place
109  * it, node (1) is looked at first, and because (4) of the semantics laid out
110  * above (bit #17 is 0), it would normally be attached to (1) as child[0].
111  * However, that slot is already allocated, so a new node is needed in between.
112  * That node does not have a value attached to it and it will never be
113  * returned to users as result of a lookup. It is only there to differentiate
114  * the traversal further. It will get a prefix as wide as necessary to
115  * distinguish its two children:
116  *
117  *                      +----------------+
118  *                      |       (1)  (R) |
119  *                      | 192.168.0.0/16 |
120  *                      |    value: 1    |
121  *                      |   [0]    [1]   |
122  *                      +----------------+
123  *                           |      |
124  *            +----------------+  +------------------+
125  *            |       (4)  (I) |  |        (3)       |
126  *            | 192.168.0.0/23 |  | 192.168.128.0/24 |
127  *            |    value: ---  |  |     value: 3     |
128  *            |   [0]    [1]   |  |    [0]    [1]    |
129  *            +----------------+  +------------------+
130  *                 |      |
131  *  +----------------+  +----------------+
132  *  |       (2)      |  |       (5)      |
133  *  | 192.168.0.0/24 |  | 192.168.1.0/24 |
134  *  |    value: 2    |  |     value: 5   |
135  *  |   [0]    [1]   |  |   [0]    [1]   |
136  *  +----------------+  +----------------+
137  *
138  * 192.168.1.1/32 would be a child of (5) etc.
139  *
140  * An intermediate node will be turned into a 'real' node on demand. In the
141  * example above, (4) would be re-used if 192.168.0.0/23 is added to the trie.
142  *
143  * A fully populated trie would have a height of 32 nodes, as the trie was
144  * created with a prefix length of 32.
145  *
146  * The lookup starts at the root node. If the current node matches and if there
147  * is a child that can be used to become more specific, the trie is traversed
148  * downwards. The last node in the traversal that is a non-intermediate one is
149  * returned.
150  */
151 
152 static inline int extract_bit(const u8 *data, size_t index)
153 {
154 	return !!(data[index / 8] & (1 << (7 - (index % 8))));
155 }
156 
157 /**
158  * longest_prefix_match() - determine the longest prefix
159  * @trie:	The trie to get internal sizes from
160  * @node:	The node to operate on
161  * @key:	The key to compare to @node
162  *
163  * Determine the longest prefix of @node that matches the bits in @key.
164  */
165 static size_t longest_prefix_match(const struct lpm_trie *trie,
166 				   const struct lpm_trie_node *node,
167 				   const struct bpf_lpm_trie_key *key)
168 {
169 	u32 limit = min(node->prefixlen, key->prefixlen);
170 	u32 prefixlen = 0, i = 0;
171 
172 	BUILD_BUG_ON(offsetof(struct lpm_trie_node, data) % sizeof(u32));
173 	BUILD_BUG_ON(offsetof(struct bpf_lpm_trie_key, data) % sizeof(u32));
174 
175 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && defined(CONFIG_64BIT)
176 
177 	/* data_size >= 16 has very small probability.
178 	 * We do not use a loop for optimal code generation.
179 	 */
180 	if (trie->data_size >= 8) {
181 		u64 diff = be64_to_cpu(*(__be64 *)node->data ^
182 				       *(__be64 *)key->data);
183 
184 		prefixlen = 64 - fls64(diff);
185 		if (prefixlen >= limit)
186 			return limit;
187 		if (diff)
188 			return prefixlen;
189 		i = 8;
190 	}
191 #endif
192 
193 	while (trie->data_size >= i + 4) {
194 		u32 diff = be32_to_cpu(*(__be32 *)&node->data[i] ^
195 				       *(__be32 *)&key->data[i]);
196 
197 		prefixlen += 32 - fls(diff);
198 		if (prefixlen >= limit)
199 			return limit;
200 		if (diff)
201 			return prefixlen;
202 		i += 4;
203 	}
204 
205 	if (trie->data_size >= i + 2) {
206 		u16 diff = be16_to_cpu(*(__be16 *)&node->data[i] ^
207 				       *(__be16 *)&key->data[i]);
208 
209 		prefixlen += 16 - fls(diff);
210 		if (prefixlen >= limit)
211 			return limit;
212 		if (diff)
213 			return prefixlen;
214 		i += 2;
215 	}
216 
217 	if (trie->data_size >= i + 1) {
218 		prefixlen += 8 - fls(node->data[i] ^ key->data[i]);
219 
220 		if (prefixlen >= limit)
221 			return limit;
222 	}
223 
224 	return prefixlen;
225 }
226 
227 /* Called from syscall or from eBPF program */
228 static void *trie_lookup_elem(struct bpf_map *map, void *_key)
229 {
230 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
231 	struct lpm_trie_node *node, *found = NULL;
232 	struct bpf_lpm_trie_key *key = _key;
233 
234 	/* Start walking the trie from the root node ... */
235 
236 	for (node = rcu_dereference_check(trie->root, rcu_read_lock_bh_held());
237 	     node;) {
238 		unsigned int next_bit;
239 		size_t matchlen;
240 
241 		/* Determine the longest prefix of @node that matches @key.
242 		 * If it's the maximum possible prefix for this trie, we have
243 		 * an exact match and can return it directly.
244 		 */
245 		matchlen = longest_prefix_match(trie, node, key);
246 		if (matchlen == trie->max_prefixlen) {
247 			found = node;
248 			break;
249 		}
250 
251 		/* If the number of bits that match is smaller than the prefix
252 		 * length of @node, bail out and return the node we have seen
253 		 * last in the traversal (ie, the parent).
254 		 */
255 		if (matchlen < node->prefixlen)
256 			break;
257 
258 		/* Consider this node as return candidate unless it is an
259 		 * artificially added intermediate one.
260 		 */
261 		if (!(node->flags & LPM_TREE_NODE_FLAG_IM))
262 			found = node;
263 
264 		/* If the node match is fully satisfied, let's see if we can
265 		 * become more specific. Determine the next bit in the key and
266 		 * traverse down.
267 		 */
268 		next_bit = extract_bit(key->data, node->prefixlen);
269 		node = rcu_dereference_check(node->child[next_bit],
270 					     rcu_read_lock_bh_held());
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 = bpf_map_kmalloc_node(&trie->map, size, GFP_NOWAIT | __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 long 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 	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 slot */
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 	spin_unlock_irqrestore(&trie->lock, irq_flags);
429 
430 	return ret;
431 }
432 
433 /* Called from syscall or from eBPF program */
434 static long 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 	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 	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 
547 	if (!bpf_capable())
548 		return ERR_PTR(-EPERM);
549 
550 	/* check sanity of attributes */
551 	if (attr->max_entries == 0 ||
552 	    !(attr->map_flags & BPF_F_NO_PREALLOC) ||
553 	    attr->map_flags & ~LPM_CREATE_FLAG_MASK ||
554 	    !bpf_map_flags_access_ok(attr->map_flags) ||
555 	    attr->key_size < LPM_KEY_SIZE_MIN ||
556 	    attr->key_size > LPM_KEY_SIZE_MAX ||
557 	    attr->value_size < LPM_VAL_SIZE_MIN ||
558 	    attr->value_size > LPM_VAL_SIZE_MAX)
559 		return ERR_PTR(-EINVAL);
560 
561 	trie = bpf_map_area_alloc(sizeof(*trie), NUMA_NO_NODE);
562 	if (!trie)
563 		return ERR_PTR(-ENOMEM);
564 
565 	/* copy mandatory map attributes */
566 	bpf_map_init_from_attr(&trie->map, attr);
567 	trie->data_size = attr->key_size -
568 			  offsetof(struct bpf_lpm_trie_key, data);
569 	trie->max_prefixlen = trie->data_size * 8;
570 
571 	spin_lock_init(&trie->lock);
572 
573 	return &trie->map;
574 }
575 
576 static void trie_free(struct bpf_map *map)
577 {
578 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
579 	struct lpm_trie_node __rcu **slot;
580 	struct lpm_trie_node *node;
581 
582 	/* Always start at the root and walk down to a node that has no
583 	 * children. Then free that node, nullify its reference in the parent
584 	 * and start over.
585 	 */
586 
587 	for (;;) {
588 		slot = &trie->root;
589 
590 		for (;;) {
591 			node = rcu_dereference_protected(*slot, 1);
592 			if (!node)
593 				goto out;
594 
595 			if (rcu_access_pointer(node->child[0])) {
596 				slot = &node->child[0];
597 				continue;
598 			}
599 
600 			if (rcu_access_pointer(node->child[1])) {
601 				slot = &node->child[1];
602 				continue;
603 			}
604 
605 			kfree(node);
606 			RCU_INIT_POINTER(*slot, NULL);
607 			break;
608 		}
609 	}
610 
611 out:
612 	bpf_map_area_free(trie);
613 }
614 
615 static int trie_get_next_key(struct bpf_map *map, void *_key, void *_next_key)
616 {
617 	struct lpm_trie_node *node, *next_node = NULL, *parent, *search_root;
618 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
619 	struct bpf_lpm_trie_key *key = _key, *next_key = _next_key;
620 	struct lpm_trie_node **node_stack = NULL;
621 	int err = 0, stack_ptr = -1;
622 	unsigned int next_bit;
623 	size_t matchlen;
624 
625 	/* The get_next_key follows postorder. For the 4 node example in
626 	 * the top of this file, the trie_get_next_key() returns the following
627 	 * one after another:
628 	 *   192.168.0.0/24
629 	 *   192.168.1.0/24
630 	 *   192.168.128.0/24
631 	 *   192.168.0.0/16
632 	 *
633 	 * The idea is to return more specific keys before less specific ones.
634 	 */
635 
636 	/* Empty trie */
637 	search_root = rcu_dereference(trie->root);
638 	if (!search_root)
639 		return -ENOENT;
640 
641 	/* For invalid key, find the leftmost node in the trie */
642 	if (!key || key->prefixlen > trie->max_prefixlen)
643 		goto find_leftmost;
644 
645 	node_stack = kmalloc_array(trie->max_prefixlen,
646 				   sizeof(struct lpm_trie_node *),
647 				   GFP_ATOMIC | __GFP_NOWARN);
648 	if (!node_stack)
649 		return -ENOMEM;
650 
651 	/* Try to find the exact node for the given key */
652 	for (node = search_root; node;) {
653 		node_stack[++stack_ptr] = node;
654 		matchlen = longest_prefix_match(trie, node, key);
655 		if (node->prefixlen != matchlen ||
656 		    node->prefixlen == key->prefixlen)
657 			break;
658 
659 		next_bit = extract_bit(key->data, node->prefixlen);
660 		node = rcu_dereference(node->child[next_bit]);
661 	}
662 	if (!node || node->prefixlen != key->prefixlen ||
663 	    (node->flags & LPM_TREE_NODE_FLAG_IM))
664 		goto find_leftmost;
665 
666 	/* The node with the exactly-matching key has been found,
667 	 * find the first node in postorder after the matched node.
668 	 */
669 	node = node_stack[stack_ptr];
670 	while (stack_ptr > 0) {
671 		parent = node_stack[stack_ptr - 1];
672 		if (rcu_dereference(parent->child[0]) == node) {
673 			search_root = rcu_dereference(parent->child[1]);
674 			if (search_root)
675 				goto find_leftmost;
676 		}
677 		if (!(parent->flags & LPM_TREE_NODE_FLAG_IM)) {
678 			next_node = parent;
679 			goto do_copy;
680 		}
681 
682 		node = parent;
683 		stack_ptr--;
684 	}
685 
686 	/* did not find anything */
687 	err = -ENOENT;
688 	goto free_stack;
689 
690 find_leftmost:
691 	/* Find the leftmost non-intermediate node, all intermediate nodes
692 	 * have exact two children, so this function will never return NULL.
693 	 */
694 	for (node = search_root; node;) {
695 		if (node->flags & LPM_TREE_NODE_FLAG_IM) {
696 			node = rcu_dereference(node->child[0]);
697 		} else {
698 			next_node = node;
699 			node = rcu_dereference(node->child[0]);
700 			if (!node)
701 				node = rcu_dereference(next_node->child[1]);
702 		}
703 	}
704 do_copy:
705 	next_key->prefixlen = next_node->prefixlen;
706 	memcpy((void *)next_key + offsetof(struct bpf_lpm_trie_key, data),
707 	       next_node->data, trie->data_size);
708 free_stack:
709 	kfree(node_stack);
710 	return err;
711 }
712 
713 static int trie_check_btf(const struct bpf_map *map,
714 			  const struct btf *btf,
715 			  const struct btf_type *key_type,
716 			  const struct btf_type *value_type)
717 {
718 	/* Keys must have struct bpf_lpm_trie_key embedded. */
719 	return BTF_INFO_KIND(key_type->info) != BTF_KIND_STRUCT ?
720 	       -EINVAL : 0;
721 }
722 
723 static u64 trie_mem_usage(const struct bpf_map *map)
724 {
725 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
726 	u64 elem_size;
727 
728 	elem_size = sizeof(struct lpm_trie_node) + trie->data_size +
729 			    trie->map.value_size;
730 	return elem_size * READ_ONCE(trie->n_entries);
731 }
732 
733 BTF_ID_LIST_SINGLE(trie_map_btf_ids, struct, lpm_trie)
734 const struct bpf_map_ops trie_map_ops = {
735 	.map_meta_equal = bpf_map_meta_equal,
736 	.map_alloc = trie_alloc,
737 	.map_free = trie_free,
738 	.map_get_next_key = trie_get_next_key,
739 	.map_lookup_elem = trie_lookup_elem,
740 	.map_update_elem = trie_update_elem,
741 	.map_delete_elem = trie_delete_elem,
742 	.map_lookup_batch = generic_map_lookup_batch,
743 	.map_update_batch = generic_map_update_batch,
744 	.map_delete_batch = generic_map_delete_batch,
745 	.map_check_btf = trie_check_btf,
746 	.map_mem_usage = trie_mem_usage,
747 	.map_btf_id = &trie_map_btf_ids[0],
748 };
749