xref: /openbmc/linux/kernel/bpf/lpm_trie.c (revision 1c0bd035)
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 
18 /* Intermediate node */
19 #define LPM_TREE_NODE_FLAG_IM BIT(0)
20 
21 struct lpm_trie_node;
22 
23 struct lpm_trie_node {
24 	struct rcu_head rcu;
25 	struct lpm_trie_node __rcu	*child[2];
26 	u32				prefixlen;
27 	u32				flags;
28 	u8				data[];
29 };
30 
31 struct lpm_trie {
32 	struct bpf_map			map;
33 	struct lpm_trie_node __rcu	*root;
34 	size_t				n_entries;
35 	size_t				max_prefixlen;
36 	size_t				data_size;
37 	spinlock_t			lock;
38 };
39 
40 /* This trie implements a longest prefix match algorithm that can be used to
41  * match IP addresses to a stored set of ranges.
42  *
43  * Data stored in @data of struct bpf_lpm_key and struct lpm_trie_node is
44  * interpreted as big endian, so data[0] stores the most significant byte.
45  *
46  * Match ranges are internally stored in instances of struct lpm_trie_node
47  * which each contain their prefix length as well as two pointers that may
48  * lead to more nodes containing more specific matches. Each node also stores
49  * a value that is defined by and returned to userspace via the update_elem
50  * and lookup functions.
51  *
52  * For instance, let's start with a trie that was created with a prefix length
53  * of 32, so it can be used for IPv4 addresses, and one single element that
54  * matches 192.168.0.0/16. The data array would hence contain
55  * [0xc0, 0xa8, 0x00, 0x00] in big-endian notation. This documentation will
56  * stick to IP-address notation for readability though.
57  *
58  * As the trie is empty initially, the new node (1) will be places as root
59  * node, denoted as (R) in the example below. As there are no other node, both
60  * child pointers are %NULL.
61  *
62  *              +----------------+
63  *              |       (1)  (R) |
64  *              | 192.168.0.0/16 |
65  *              |    value: 1    |
66  *              |   [0]    [1]   |
67  *              +----------------+
68  *
69  * Next, let's add a new node (2) matching 192.168.0.0/24. As there is already
70  * a node with the same data and a smaller prefix (ie, a less specific one),
71  * node (2) will become a child of (1). In child index depends on the next bit
72  * that is outside of what (1) matches, and that bit is 0, so (2) will be
73  * child[0] of (1):
74  *
75  *              +----------------+
76  *              |       (1)  (R) |
77  *              | 192.168.0.0/16 |
78  *              |    value: 1    |
79  *              |   [0]    [1]   |
80  *              +----------------+
81  *                   |
82  *    +----------------+
83  *    |       (2)      |
84  *    | 192.168.0.0/24 |
85  *    |    value: 2    |
86  *    |   [0]    [1]   |
87  *    +----------------+
88  *
89  * The child[1] slot of (1) could be filled with another node which has bit #17
90  * (the next bit after the ones that (1) matches on) set to 1. For instance,
91  * 192.168.128.0/24:
92  *
93  *              +----------------+
94  *              |       (1)  (R) |
95  *              | 192.168.0.0/16 |
96  *              |    value: 1    |
97  *              |   [0]    [1]   |
98  *              +----------------+
99  *                   |      |
100  *    +----------------+  +------------------+
101  *    |       (2)      |  |        (3)       |
102  *    | 192.168.0.0/24 |  | 192.168.128.0/24 |
103  *    |    value: 2    |  |     value: 3     |
104  *    |   [0]    [1]   |  |    [0]    [1]    |
105  *    +----------------+  +------------------+
106  *
107  * Let's add another node (4) to the game for 192.168.1.0/24. In order to place
108  * it, node (1) is looked at first, and because (4) of the semantics laid out
109  * above (bit #17 is 0), it would normally be attached to (1) as child[0].
110  * However, that slot is already allocated, so a new node is needed in between.
111  * That node does not have a value attached to it and it will never be
112  * returned to users as result of a lookup. It is only there to differentiate
113  * the traversal further. It will get a prefix as wide as necessary to
114  * distinguish its two children:
115  *
116  *                      +----------------+
117  *                      |       (1)  (R) |
118  *                      | 192.168.0.0/16 |
119  *                      |    value: 1    |
120  *                      |   [0]    [1]   |
121  *                      +----------------+
122  *                           |      |
123  *            +----------------+  +------------------+
124  *            |       (4)  (I) |  |        (3)       |
125  *            | 192.168.0.0/23 |  | 192.168.128.0/24 |
126  *            |    value: ---  |  |     value: 3     |
127  *            |   [0]    [1]   |  |    [0]    [1]    |
128  *            +----------------+  +------------------+
129  *                 |      |
130  *  +----------------+  +----------------+
131  *  |       (2)      |  |       (5)      |
132  *  | 192.168.0.0/24 |  | 192.168.1.0/24 |
133  *  |    value: 2    |  |     value: 5   |
134  *  |   [0]    [1]   |  |   [0]    [1]   |
135  *  +----------------+  +----------------+
136  *
137  * 192.168.1.1/32 would be a child of (5) etc.
138  *
139  * An intermediate node will be turned into a 'real' node on demand. In the
140  * example above, (4) would be re-used if 192.168.0.0/23 is added to the trie.
141  *
142  * A fully populated trie would have a height of 32 nodes, as the trie was
143  * created with a prefix length of 32.
144  *
145  * The lookup starts at the root node. If the current node matches and if there
146  * is a child that can be used to become more specific, the trie is traversed
147  * downwards. The last node in the traversal that is a non-intermediate one is
148  * returned.
149  */
150 
151 static inline int extract_bit(const u8 *data, size_t index)
152 {
153 	return !!(data[index / 8] & (1 << (7 - (index % 8))));
154 }
155 
156 /**
157  * longest_prefix_match() - determine the longest prefix
158  * @trie:	The trie to get internal sizes from
159  * @node:	The node to operate on
160  * @key:	The key to compare to @node
161  *
162  * Determine the longest prefix of @node that matches the bits in @key.
163  */
164 static size_t longest_prefix_match(const struct lpm_trie *trie,
165 				   const struct lpm_trie_node *node,
166 				   const struct bpf_lpm_trie_key *key)
167 {
168 	u32 limit = min(node->prefixlen, key->prefixlen);
169 	u32 prefixlen = 0, i = 0;
170 
171 	BUILD_BUG_ON(offsetof(struct lpm_trie_node, data) % sizeof(u32));
172 	BUILD_BUG_ON(offsetof(struct bpf_lpm_trie_key, data) % sizeof(u32));
173 
174 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && defined(CONFIG_64BIT)
175 
176 	/* data_size >= 16 has very small probability.
177 	 * We do not use a loop for optimal code generation.
178 	 */
179 	if (trie->data_size >= 8) {
180 		u64 diff = be64_to_cpu(*(__be64 *)node->data ^
181 				       *(__be64 *)key->data);
182 
183 		prefixlen = 64 - fls64(diff);
184 		if (prefixlen >= limit)
185 			return limit;
186 		if (diff)
187 			return prefixlen;
188 		i = 8;
189 	}
190 #endif
191 
192 	while (trie->data_size >= i + 4) {
193 		u32 diff = be32_to_cpu(*(__be32 *)&node->data[i] ^
194 				       *(__be32 *)&key->data[i]);
195 
196 		prefixlen += 32 - fls(diff);
197 		if (prefixlen >= limit)
198 			return limit;
199 		if (diff)
200 			return prefixlen;
201 		i += 4;
202 	}
203 
204 	if (trie->data_size >= i + 2) {
205 		u16 diff = be16_to_cpu(*(__be16 *)&node->data[i] ^
206 				       *(__be16 *)&key->data[i]);
207 
208 		prefixlen += 16 - fls(diff);
209 		if (prefixlen >= limit)
210 			return limit;
211 		if (diff)
212 			return prefixlen;
213 		i += 2;
214 	}
215 
216 	if (trie->data_size >= i + 1) {
217 		prefixlen += 8 - fls(node->data[i] ^ key->data[i]);
218 
219 		if (prefixlen >= limit)
220 			return limit;
221 	}
222 
223 	return prefixlen;
224 }
225 
226 /* Called from syscall or from eBPF program */
227 static void *trie_lookup_elem(struct bpf_map *map, void *_key)
228 {
229 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
230 	struct lpm_trie_node *node, *found = NULL;
231 	struct bpf_lpm_trie_key *key = _key;
232 
233 	/* Start walking the trie from the root node ... */
234 
235 	for (node = rcu_dereference_check(trie->root, rcu_read_lock_bh_held());
236 	     node;) {
237 		unsigned int next_bit;
238 		size_t matchlen;
239 
240 		/* Determine the longest prefix of @node that matches @key.
241 		 * If it's the maximum possible prefix for this trie, we have
242 		 * an exact match and can return it directly.
243 		 */
244 		matchlen = longest_prefix_match(trie, node, key);
245 		if (matchlen == trie->max_prefixlen) {
246 			found = node;
247 			break;
248 		}
249 
250 		/* If the number of bits that match is smaller than the prefix
251 		 * length of @node, bail out and return the node we have seen
252 		 * last in the traversal (ie, the parent).
253 		 */
254 		if (matchlen < node->prefixlen)
255 			break;
256 
257 		/* Consider this node as return candidate unless it is an
258 		 * artificially added intermediate one.
259 		 */
260 		if (!(node->flags & LPM_TREE_NODE_FLAG_IM))
261 			found = node;
262 
263 		/* If the node match is fully satisfied, let's see if we can
264 		 * become more specific. Determine the next bit in the key and
265 		 * traverse down.
266 		 */
267 		next_bit = extract_bit(key->data, node->prefixlen);
268 		node = rcu_dereference_check(node->child[next_bit],
269 					     rcu_read_lock_bh_held());
270 	}
271 
272 	if (!found)
273 		return NULL;
274 
275 	return found->data + trie->data_size;
276 }
277 
278 static struct lpm_trie_node *lpm_trie_node_alloc(const struct lpm_trie *trie,
279 						 const void *value)
280 {
281 	struct lpm_trie_node *node;
282 	size_t size = sizeof(struct lpm_trie_node) + trie->data_size;
283 
284 	if (value)
285 		size += trie->map.value_size;
286 
287 	node = bpf_map_kmalloc_node(&trie->map, size, GFP_ATOMIC | __GFP_NOWARN,
288 				    trie->map.numa_node);
289 	if (!node)
290 		return NULL;
291 
292 	node->flags = 0;
293 
294 	if (value)
295 		memcpy(node->data + trie->data_size, value,
296 		       trie->map.value_size);
297 
298 	return node;
299 }
300 
301 /* Called from syscall or from eBPF program */
302 static int trie_update_elem(struct bpf_map *map,
303 			    void *_key, void *value, u64 flags)
304 {
305 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
306 	struct lpm_trie_node *node, *im_node = NULL, *new_node = NULL;
307 	struct lpm_trie_node __rcu **slot;
308 	struct bpf_lpm_trie_key *key = _key;
309 	unsigned long irq_flags;
310 	unsigned int next_bit;
311 	size_t matchlen = 0;
312 	int ret = 0;
313 
314 	if (unlikely(flags > BPF_EXIST))
315 		return -EINVAL;
316 
317 	if (key->prefixlen > trie->max_prefixlen)
318 		return -EINVAL;
319 
320 	spin_lock_irqsave(&trie->lock, irq_flags);
321 
322 	/* Allocate and fill a new node */
323 
324 	if (trie->n_entries == trie->map.max_entries) {
325 		ret = -ENOSPC;
326 		goto out;
327 	}
328 
329 	new_node = lpm_trie_node_alloc(trie, value);
330 	if (!new_node) {
331 		ret = -ENOMEM;
332 		goto out;
333 	}
334 
335 	trie->n_entries++;
336 
337 	new_node->prefixlen = key->prefixlen;
338 	RCU_INIT_POINTER(new_node->child[0], NULL);
339 	RCU_INIT_POINTER(new_node->child[1], NULL);
340 	memcpy(new_node->data, key->data, trie->data_size);
341 
342 	/* Now find a slot to attach the new node. To do that, walk the tree
343 	 * from the root and match as many bits as possible for each node until
344 	 * we either find an empty slot or a slot that needs to be replaced by
345 	 * an intermediate node.
346 	 */
347 	slot = &trie->root;
348 
349 	while ((node = rcu_dereference_protected(*slot,
350 					lockdep_is_held(&trie->lock)))) {
351 		matchlen = longest_prefix_match(trie, node, key);
352 
353 		if (node->prefixlen != matchlen ||
354 		    node->prefixlen == key->prefixlen ||
355 		    node->prefixlen == trie->max_prefixlen)
356 			break;
357 
358 		next_bit = extract_bit(key->data, node->prefixlen);
359 		slot = &node->child[next_bit];
360 	}
361 
362 	/* If the slot is empty (a free child pointer or an empty root),
363 	 * simply assign the @new_node to that slot and be done.
364 	 */
365 	if (!node) {
366 		rcu_assign_pointer(*slot, new_node);
367 		goto out;
368 	}
369 
370 	/* If the slot we picked already exists, replace it with @new_node
371 	 * which already has the correct data array set.
372 	 */
373 	if (node->prefixlen == matchlen) {
374 		new_node->child[0] = node->child[0];
375 		new_node->child[1] = node->child[1];
376 
377 		if (!(node->flags & LPM_TREE_NODE_FLAG_IM))
378 			trie->n_entries--;
379 
380 		rcu_assign_pointer(*slot, new_node);
381 		kfree_rcu(node, rcu);
382 
383 		goto out;
384 	}
385 
386 	/* If the new node matches the prefix completely, it must be inserted
387 	 * as an ancestor. Simply insert it between @node and *@slot.
388 	 */
389 	if (matchlen == key->prefixlen) {
390 		next_bit = extract_bit(node->data, matchlen);
391 		rcu_assign_pointer(new_node->child[next_bit], node);
392 		rcu_assign_pointer(*slot, new_node);
393 		goto out;
394 	}
395 
396 	im_node = lpm_trie_node_alloc(trie, NULL);
397 	if (!im_node) {
398 		ret = -ENOMEM;
399 		goto out;
400 	}
401 
402 	im_node->prefixlen = matchlen;
403 	im_node->flags |= LPM_TREE_NODE_FLAG_IM;
404 	memcpy(im_node->data, node->data, trie->data_size);
405 
406 	/* Now determine which child to install in which slot */
407 	if (extract_bit(key->data, matchlen)) {
408 		rcu_assign_pointer(im_node->child[0], node);
409 		rcu_assign_pointer(im_node->child[1], new_node);
410 	} else {
411 		rcu_assign_pointer(im_node->child[0], new_node);
412 		rcu_assign_pointer(im_node->child[1], node);
413 	}
414 
415 	/* Finally, assign the intermediate node to the determined slot */
416 	rcu_assign_pointer(*slot, im_node);
417 
418 out:
419 	if (ret) {
420 		if (new_node)
421 			trie->n_entries--;
422 
423 		kfree(new_node);
424 		kfree(im_node);
425 	}
426 
427 	spin_unlock_irqrestore(&trie->lock, irq_flags);
428 
429 	return ret;
430 }
431 
432 /* Called from syscall or from eBPF program */
433 static int trie_delete_elem(struct bpf_map *map, void *_key)
434 {
435 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
436 	struct bpf_lpm_trie_key *key = _key;
437 	struct lpm_trie_node __rcu **trim, **trim2;
438 	struct lpm_trie_node *node, *parent;
439 	unsigned long irq_flags;
440 	unsigned int next_bit;
441 	size_t matchlen = 0;
442 	int ret = 0;
443 
444 	if (key->prefixlen > trie->max_prefixlen)
445 		return -EINVAL;
446 
447 	spin_lock_irqsave(&trie->lock, irq_flags);
448 
449 	/* Walk the tree looking for an exact key/length match and keeping
450 	 * track of the path we traverse.  We will need to know the node
451 	 * we wish to delete, and the slot that points to the node we want
452 	 * to delete.  We may also need to know the nodes parent and the
453 	 * slot that contains it.
454 	 */
455 	trim = &trie->root;
456 	trim2 = trim;
457 	parent = NULL;
458 	while ((node = rcu_dereference_protected(
459 		       *trim, lockdep_is_held(&trie->lock)))) {
460 		matchlen = longest_prefix_match(trie, node, key);
461 
462 		if (node->prefixlen != matchlen ||
463 		    node->prefixlen == key->prefixlen)
464 			break;
465 
466 		parent = node;
467 		trim2 = trim;
468 		next_bit = extract_bit(key->data, node->prefixlen);
469 		trim = &node->child[next_bit];
470 	}
471 
472 	if (!node || node->prefixlen != key->prefixlen ||
473 	    node->prefixlen != matchlen ||
474 	    (node->flags & LPM_TREE_NODE_FLAG_IM)) {
475 		ret = -ENOENT;
476 		goto out;
477 	}
478 
479 	trie->n_entries--;
480 
481 	/* If the node we are removing has two children, simply mark it
482 	 * as intermediate and we are done.
483 	 */
484 	if (rcu_access_pointer(node->child[0]) &&
485 	    rcu_access_pointer(node->child[1])) {
486 		node->flags |= LPM_TREE_NODE_FLAG_IM;
487 		goto out;
488 	}
489 
490 	/* If the parent of the node we are about to delete is an intermediate
491 	 * node, and the deleted node doesn't have any children, we can delete
492 	 * the intermediate parent as well and promote its other child
493 	 * up the tree.  Doing this maintains the invariant that all
494 	 * intermediate nodes have exactly 2 children and that there are no
495 	 * unnecessary intermediate nodes in the tree.
496 	 */
497 	if (parent && (parent->flags & LPM_TREE_NODE_FLAG_IM) &&
498 	    !node->child[0] && !node->child[1]) {
499 		if (node == rcu_access_pointer(parent->child[0]))
500 			rcu_assign_pointer(
501 				*trim2, rcu_access_pointer(parent->child[1]));
502 		else
503 			rcu_assign_pointer(
504 				*trim2, rcu_access_pointer(parent->child[0]));
505 		kfree_rcu(parent, rcu);
506 		kfree_rcu(node, rcu);
507 		goto out;
508 	}
509 
510 	/* The node we are removing has either zero or one child. If there
511 	 * is a child, move it into the removed node's slot then delete
512 	 * the node.  Otherwise just clear the slot and delete the node.
513 	 */
514 	if (node->child[0])
515 		rcu_assign_pointer(*trim, rcu_access_pointer(node->child[0]));
516 	else if (node->child[1])
517 		rcu_assign_pointer(*trim, rcu_access_pointer(node->child[1]));
518 	else
519 		RCU_INIT_POINTER(*trim, NULL);
520 	kfree_rcu(node, rcu);
521 
522 out:
523 	spin_unlock_irqrestore(&trie->lock, irq_flags);
524 
525 	return ret;
526 }
527 
528 #define LPM_DATA_SIZE_MAX	256
529 #define LPM_DATA_SIZE_MIN	1
530 
531 #define LPM_VAL_SIZE_MAX	(KMALLOC_MAX_SIZE - LPM_DATA_SIZE_MAX - \
532 				 sizeof(struct lpm_trie_node))
533 #define LPM_VAL_SIZE_MIN	1
534 
535 #define LPM_KEY_SIZE(X)		(sizeof(struct bpf_lpm_trie_key) + (X))
536 #define LPM_KEY_SIZE_MAX	LPM_KEY_SIZE(LPM_DATA_SIZE_MAX)
537 #define LPM_KEY_SIZE_MIN	LPM_KEY_SIZE(LPM_DATA_SIZE_MIN)
538 
539 #define LPM_CREATE_FLAG_MASK	(BPF_F_NO_PREALLOC | BPF_F_NUMA_NODE |	\
540 				 BPF_F_ACCESS_MASK)
541 
542 static struct bpf_map *trie_alloc(union bpf_attr *attr)
543 {
544 	struct lpm_trie *trie;
545 
546 	if (!bpf_capable())
547 		return ERR_PTR(-EPERM);
548 
549 	/* check sanity of attributes */
550 	if (attr->max_entries == 0 ||
551 	    !(attr->map_flags & BPF_F_NO_PREALLOC) ||
552 	    attr->map_flags & ~LPM_CREATE_FLAG_MASK ||
553 	    !bpf_map_flags_access_ok(attr->map_flags) ||
554 	    attr->key_size < LPM_KEY_SIZE_MIN ||
555 	    attr->key_size > LPM_KEY_SIZE_MAX ||
556 	    attr->value_size < LPM_VAL_SIZE_MIN ||
557 	    attr->value_size > LPM_VAL_SIZE_MAX)
558 		return ERR_PTR(-EINVAL);
559 
560 	trie = kzalloc(sizeof(*trie), GFP_USER | __GFP_NOWARN | __GFP_ACCOUNT);
561 	if (!trie)
562 		return ERR_PTR(-ENOMEM);
563 
564 	/* copy mandatory map attributes */
565 	bpf_map_init_from_attr(&trie->map, attr);
566 	trie->data_size = attr->key_size -
567 			  offsetof(struct bpf_lpm_trie_key, data);
568 	trie->max_prefixlen = trie->data_size * 8;
569 
570 	spin_lock_init(&trie->lock);
571 
572 	return &trie->map;
573 }
574 
575 static void trie_free(struct bpf_map *map)
576 {
577 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
578 	struct lpm_trie_node __rcu **slot;
579 	struct lpm_trie_node *node;
580 
581 	/* Always start at the root and walk down to a node that has no
582 	 * children. Then free that node, nullify its reference in the parent
583 	 * and start over.
584 	 */
585 
586 	for (;;) {
587 		slot = &trie->root;
588 
589 		for (;;) {
590 			node = rcu_dereference_protected(*slot, 1);
591 			if (!node)
592 				goto out;
593 
594 			if (rcu_access_pointer(node->child[0])) {
595 				slot = &node->child[0];
596 				continue;
597 			}
598 
599 			if (rcu_access_pointer(node->child[1])) {
600 				slot = &node->child[1];
601 				continue;
602 			}
603 
604 			kfree(node);
605 			RCU_INIT_POINTER(*slot, NULL);
606 			break;
607 		}
608 	}
609 
610 out:
611 	kfree(trie);
612 }
613 
614 static int trie_get_next_key(struct bpf_map *map, void *_key, void *_next_key)
615 {
616 	struct lpm_trie_node *node, *next_node = NULL, *parent, *search_root;
617 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
618 	struct bpf_lpm_trie_key *key = _key, *next_key = _next_key;
619 	struct lpm_trie_node **node_stack = NULL;
620 	int err = 0, stack_ptr = -1;
621 	unsigned int next_bit;
622 	size_t matchlen;
623 
624 	/* The get_next_key follows postorder. For the 4 node example in
625 	 * the top of this file, the trie_get_next_key() returns the following
626 	 * one after another:
627 	 *   192.168.0.0/24
628 	 *   192.168.1.0/24
629 	 *   192.168.128.0/24
630 	 *   192.168.0.0/16
631 	 *
632 	 * The idea is to return more specific keys before less specific ones.
633 	 */
634 
635 	/* Empty trie */
636 	search_root = rcu_dereference(trie->root);
637 	if (!search_root)
638 		return -ENOENT;
639 
640 	/* For invalid key, find the leftmost node in the trie */
641 	if (!key || key->prefixlen > trie->max_prefixlen)
642 		goto find_leftmost;
643 
644 	node_stack = kmalloc_array(trie->max_prefixlen,
645 				   sizeof(struct lpm_trie_node *),
646 				   GFP_ATOMIC | __GFP_NOWARN);
647 	if (!node_stack)
648 		return -ENOMEM;
649 
650 	/* Try to find the exact node for the given key */
651 	for (node = search_root; node;) {
652 		node_stack[++stack_ptr] = node;
653 		matchlen = longest_prefix_match(trie, node, key);
654 		if (node->prefixlen != matchlen ||
655 		    node->prefixlen == key->prefixlen)
656 			break;
657 
658 		next_bit = extract_bit(key->data, node->prefixlen);
659 		node = rcu_dereference(node->child[next_bit]);
660 	}
661 	if (!node || node->prefixlen != key->prefixlen ||
662 	    (node->flags & LPM_TREE_NODE_FLAG_IM))
663 		goto find_leftmost;
664 
665 	/* The node with the exactly-matching key has been found,
666 	 * find the first node in postorder after the matched node.
667 	 */
668 	node = node_stack[stack_ptr];
669 	while (stack_ptr > 0) {
670 		parent = node_stack[stack_ptr - 1];
671 		if (rcu_dereference(parent->child[0]) == node) {
672 			search_root = rcu_dereference(parent->child[1]);
673 			if (search_root)
674 				goto find_leftmost;
675 		}
676 		if (!(parent->flags & LPM_TREE_NODE_FLAG_IM)) {
677 			next_node = parent;
678 			goto do_copy;
679 		}
680 
681 		node = parent;
682 		stack_ptr--;
683 	}
684 
685 	/* did not find anything */
686 	err = -ENOENT;
687 	goto free_stack;
688 
689 find_leftmost:
690 	/* Find the leftmost non-intermediate node, all intermediate nodes
691 	 * have exact two children, so this function will never return NULL.
692 	 */
693 	for (node = search_root; node;) {
694 		if (node->flags & LPM_TREE_NODE_FLAG_IM) {
695 			node = rcu_dereference(node->child[0]);
696 		} else {
697 			next_node = node;
698 			node = rcu_dereference(node->child[0]);
699 			if (!node)
700 				node = rcu_dereference(next_node->child[1]);
701 		}
702 	}
703 do_copy:
704 	next_key->prefixlen = next_node->prefixlen;
705 	memcpy((void *)next_key + offsetof(struct bpf_lpm_trie_key, data),
706 	       next_node->data, trie->data_size);
707 free_stack:
708 	kfree(node_stack);
709 	return err;
710 }
711 
712 static int trie_check_btf(const struct bpf_map *map,
713 			  const struct btf *btf,
714 			  const struct btf_type *key_type,
715 			  const struct btf_type *value_type)
716 {
717 	/* Keys must have struct bpf_lpm_trie_key embedded. */
718 	return BTF_INFO_KIND(key_type->info) != BTF_KIND_STRUCT ?
719 	       -EINVAL : 0;
720 }
721 
722 static int trie_map_btf_id;
723 const struct bpf_map_ops trie_map_ops = {
724 	.map_meta_equal = bpf_map_meta_equal,
725 	.map_alloc = trie_alloc,
726 	.map_free = trie_free,
727 	.map_get_next_key = trie_get_next_key,
728 	.map_lookup_elem = trie_lookup_elem,
729 	.map_update_elem = trie_update_elem,
730 	.map_delete_elem = trie_delete_elem,
731 	.map_lookup_batch = generic_map_lookup_batch,
732 	.map_update_batch = generic_map_update_batch,
733 	.map_delete_batch = generic_map_delete_batch,
734 	.map_check_btf = trie_check_btf,
735 	.map_btf_name = "lpm_trie",
736 	.map_btf_id = &trie_map_btf_id,
737 };
738