xref: /linux/kernel/bpf/lpm_trie.c (revision 9d56c248e5030d17ea9cd132634e86fdf0622d0e)
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 __always_inline
166 size_t __longest_prefix_match(const struct lpm_trie *trie,
167 			      const struct lpm_trie_node *node,
168 			      const struct bpf_lpm_trie_key_u8 *key)
169 {
170 	u32 limit = min(node->prefixlen, key->prefixlen);
171 	u32 prefixlen = 0, i = 0;
172 
173 	BUILD_BUG_ON(offsetof(struct lpm_trie_node, data) % sizeof(u32));
174 	BUILD_BUG_ON(offsetof(struct bpf_lpm_trie_key_u8, data) % sizeof(u32));
175 
176 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && defined(CONFIG_64BIT)
177 
178 	/* data_size >= 16 has very small probability.
179 	 * We do not use a loop for optimal code generation.
180 	 */
181 	if (trie->data_size >= 8) {
182 		u64 diff = be64_to_cpu(*(__be64 *)node->data ^
183 				       *(__be64 *)key->data);
184 
185 		prefixlen = 64 - fls64(diff);
186 		if (prefixlen >= limit)
187 			return limit;
188 		if (diff)
189 			return prefixlen;
190 		i = 8;
191 	}
192 #endif
193 
194 	while (trie->data_size >= i + 4) {
195 		u32 diff = be32_to_cpu(*(__be32 *)&node->data[i] ^
196 				       *(__be32 *)&key->data[i]);
197 
198 		prefixlen += 32 - fls(diff);
199 		if (prefixlen >= limit)
200 			return limit;
201 		if (diff)
202 			return prefixlen;
203 		i += 4;
204 	}
205 
206 	if (trie->data_size >= i + 2) {
207 		u16 diff = be16_to_cpu(*(__be16 *)&node->data[i] ^
208 				       *(__be16 *)&key->data[i]);
209 
210 		prefixlen += 16 - fls(diff);
211 		if (prefixlen >= limit)
212 			return limit;
213 		if (diff)
214 			return prefixlen;
215 		i += 2;
216 	}
217 
218 	if (trie->data_size >= i + 1) {
219 		prefixlen += 8 - fls(node->data[i] ^ key->data[i]);
220 
221 		if (prefixlen >= limit)
222 			return limit;
223 	}
224 
225 	return prefixlen;
226 }
227 
228 static size_t longest_prefix_match(const struct lpm_trie *trie,
229 				   const struct lpm_trie_node *node,
230 				   const struct bpf_lpm_trie_key_u8 *key)
231 {
232 	return __longest_prefix_match(trie, node, key);
233 }
234 
235 /* Called from syscall or from eBPF program */
236 static void *trie_lookup_elem(struct bpf_map *map, void *_key)
237 {
238 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
239 	struct lpm_trie_node *node, *found = NULL;
240 	struct bpf_lpm_trie_key_u8 *key = _key;
241 
242 	if (key->prefixlen > trie->max_prefixlen)
243 		return NULL;
244 
245 	/* Start walking the trie from the root node ... */
246 
247 	for (node = rcu_dereference_check(trie->root, rcu_read_lock_bh_held());
248 	     node;) {
249 		unsigned int next_bit;
250 		size_t matchlen;
251 
252 		/* Determine the longest prefix of @node that matches @key.
253 		 * If it's the maximum possible prefix for this trie, we have
254 		 * an exact match and can return it directly.
255 		 */
256 		matchlen = __longest_prefix_match(trie, node, key);
257 		if (matchlen == trie->max_prefixlen) {
258 			found = node;
259 			break;
260 		}
261 
262 		/* If the number of bits that match is smaller than the prefix
263 		 * length of @node, bail out and return the node we have seen
264 		 * last in the traversal (ie, the parent).
265 		 */
266 		if (matchlen < node->prefixlen)
267 			break;
268 
269 		/* Consider this node as return candidate unless it is an
270 		 * artificially added intermediate one.
271 		 */
272 		if (!(node->flags & LPM_TREE_NODE_FLAG_IM))
273 			found = node;
274 
275 		/* If the node match is fully satisfied, let's see if we can
276 		 * become more specific. Determine the next bit in the key and
277 		 * traverse down.
278 		 */
279 		next_bit = extract_bit(key->data, node->prefixlen);
280 		node = rcu_dereference_check(node->child[next_bit],
281 					     rcu_read_lock_bh_held());
282 	}
283 
284 	if (!found)
285 		return NULL;
286 
287 	return found->data + trie->data_size;
288 }
289 
290 static struct lpm_trie_node *lpm_trie_node_alloc(const struct lpm_trie *trie,
291 						 const void *value)
292 {
293 	struct lpm_trie_node *node;
294 	size_t size = sizeof(struct lpm_trie_node) + trie->data_size;
295 
296 	if (value)
297 		size += trie->map.value_size;
298 
299 	node = bpf_map_kmalloc_node(&trie->map, size, GFP_NOWAIT | __GFP_NOWARN,
300 				    trie->map.numa_node);
301 	if (!node)
302 		return NULL;
303 
304 	node->flags = 0;
305 
306 	if (value)
307 		memcpy(node->data + trie->data_size, value,
308 		       trie->map.value_size);
309 
310 	return node;
311 }
312 
313 /* Called from syscall or from eBPF program */
314 static long trie_update_elem(struct bpf_map *map,
315 			     void *_key, void *value, u64 flags)
316 {
317 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
318 	struct lpm_trie_node *node, *im_node = NULL, *new_node = NULL;
319 	struct lpm_trie_node __rcu **slot;
320 	struct bpf_lpm_trie_key_u8 *key = _key;
321 	unsigned long irq_flags;
322 	unsigned int next_bit;
323 	size_t matchlen = 0;
324 	int ret = 0;
325 
326 	if (unlikely(flags > BPF_EXIST))
327 		return -EINVAL;
328 
329 	if (key->prefixlen > trie->max_prefixlen)
330 		return -EINVAL;
331 
332 	spin_lock_irqsave(&trie->lock, irq_flags);
333 
334 	/* Allocate and fill a new node */
335 
336 	if (trie->n_entries == trie->map.max_entries) {
337 		ret = -ENOSPC;
338 		goto out;
339 	}
340 
341 	new_node = lpm_trie_node_alloc(trie, value);
342 	if (!new_node) {
343 		ret = -ENOMEM;
344 		goto out;
345 	}
346 
347 	trie->n_entries++;
348 
349 	new_node->prefixlen = key->prefixlen;
350 	RCU_INIT_POINTER(new_node->child[0], NULL);
351 	RCU_INIT_POINTER(new_node->child[1], NULL);
352 	memcpy(new_node->data, key->data, trie->data_size);
353 
354 	/* Now find a slot to attach the new node. To do that, walk the tree
355 	 * from the root and match as many bits as possible for each node until
356 	 * we either find an empty slot or a slot that needs to be replaced by
357 	 * an intermediate node.
358 	 */
359 	slot = &trie->root;
360 
361 	while ((node = rcu_dereference_protected(*slot,
362 					lockdep_is_held(&trie->lock)))) {
363 		matchlen = longest_prefix_match(trie, node, key);
364 
365 		if (node->prefixlen != matchlen ||
366 		    node->prefixlen == key->prefixlen ||
367 		    node->prefixlen == trie->max_prefixlen)
368 			break;
369 
370 		next_bit = extract_bit(key->data, node->prefixlen);
371 		slot = &node->child[next_bit];
372 	}
373 
374 	/* If the slot is empty (a free child pointer or an empty root),
375 	 * simply assign the @new_node to that slot and be done.
376 	 */
377 	if (!node) {
378 		rcu_assign_pointer(*slot, new_node);
379 		goto out;
380 	}
381 
382 	/* If the slot we picked already exists, replace it with @new_node
383 	 * which already has the correct data array set.
384 	 */
385 	if (node->prefixlen == matchlen) {
386 		new_node->child[0] = node->child[0];
387 		new_node->child[1] = node->child[1];
388 
389 		if (!(node->flags & LPM_TREE_NODE_FLAG_IM))
390 			trie->n_entries--;
391 
392 		rcu_assign_pointer(*slot, new_node);
393 		kfree_rcu(node, rcu);
394 
395 		goto out;
396 	}
397 
398 	/* If the new node matches the prefix completely, it must be inserted
399 	 * as an ancestor. Simply insert it between @node and *@slot.
400 	 */
401 	if (matchlen == key->prefixlen) {
402 		next_bit = extract_bit(node->data, matchlen);
403 		rcu_assign_pointer(new_node->child[next_bit], node);
404 		rcu_assign_pointer(*slot, new_node);
405 		goto out;
406 	}
407 
408 	im_node = lpm_trie_node_alloc(trie, NULL);
409 	if (!im_node) {
410 		ret = -ENOMEM;
411 		goto out;
412 	}
413 
414 	im_node->prefixlen = matchlen;
415 	im_node->flags |= LPM_TREE_NODE_FLAG_IM;
416 	memcpy(im_node->data, node->data, trie->data_size);
417 
418 	/* Now determine which child to install in which slot */
419 	if (extract_bit(key->data, matchlen)) {
420 		rcu_assign_pointer(im_node->child[0], node);
421 		rcu_assign_pointer(im_node->child[1], new_node);
422 	} else {
423 		rcu_assign_pointer(im_node->child[0], new_node);
424 		rcu_assign_pointer(im_node->child[1], node);
425 	}
426 
427 	/* Finally, assign the intermediate node to the determined slot */
428 	rcu_assign_pointer(*slot, im_node);
429 
430 out:
431 	if (ret) {
432 		if (new_node)
433 			trie->n_entries--;
434 
435 		kfree(new_node);
436 		kfree(im_node);
437 	}
438 
439 	spin_unlock_irqrestore(&trie->lock, irq_flags);
440 
441 	return ret;
442 }
443 
444 /* Called from syscall or from eBPF program */
445 static long trie_delete_elem(struct bpf_map *map, void *_key)
446 {
447 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
448 	struct bpf_lpm_trie_key_u8 *key = _key;
449 	struct lpm_trie_node __rcu **trim, **trim2;
450 	struct lpm_trie_node *node, *parent;
451 	unsigned long irq_flags;
452 	unsigned int next_bit;
453 	size_t matchlen = 0;
454 	int ret = 0;
455 
456 	if (key->prefixlen > trie->max_prefixlen)
457 		return -EINVAL;
458 
459 	spin_lock_irqsave(&trie->lock, irq_flags);
460 
461 	/* Walk the tree looking for an exact key/length match and keeping
462 	 * track of the path we traverse.  We will need to know the node
463 	 * we wish to delete, and the slot that points to the node we want
464 	 * to delete.  We may also need to know the nodes parent and the
465 	 * slot that contains it.
466 	 */
467 	trim = &trie->root;
468 	trim2 = trim;
469 	parent = NULL;
470 	while ((node = rcu_dereference_protected(
471 		       *trim, lockdep_is_held(&trie->lock)))) {
472 		matchlen = longest_prefix_match(trie, node, key);
473 
474 		if (node->prefixlen != matchlen ||
475 		    node->prefixlen == key->prefixlen)
476 			break;
477 
478 		parent = node;
479 		trim2 = trim;
480 		next_bit = extract_bit(key->data, node->prefixlen);
481 		trim = &node->child[next_bit];
482 	}
483 
484 	if (!node || node->prefixlen != key->prefixlen ||
485 	    node->prefixlen != matchlen ||
486 	    (node->flags & LPM_TREE_NODE_FLAG_IM)) {
487 		ret = -ENOENT;
488 		goto out;
489 	}
490 
491 	trie->n_entries--;
492 
493 	/* If the node we are removing has two children, simply mark it
494 	 * as intermediate and we are done.
495 	 */
496 	if (rcu_access_pointer(node->child[0]) &&
497 	    rcu_access_pointer(node->child[1])) {
498 		node->flags |= LPM_TREE_NODE_FLAG_IM;
499 		goto out;
500 	}
501 
502 	/* If the parent of the node we are about to delete is an intermediate
503 	 * node, and the deleted node doesn't have any children, we can delete
504 	 * the intermediate parent as well and promote its other child
505 	 * up the tree.  Doing this maintains the invariant that all
506 	 * intermediate nodes have exactly 2 children and that there are no
507 	 * unnecessary intermediate nodes in the tree.
508 	 */
509 	if (parent && (parent->flags & LPM_TREE_NODE_FLAG_IM) &&
510 	    !node->child[0] && !node->child[1]) {
511 		if (node == rcu_access_pointer(parent->child[0]))
512 			rcu_assign_pointer(
513 				*trim2, rcu_access_pointer(parent->child[1]));
514 		else
515 			rcu_assign_pointer(
516 				*trim2, rcu_access_pointer(parent->child[0]));
517 		kfree_rcu(parent, rcu);
518 		kfree_rcu(node, rcu);
519 		goto out;
520 	}
521 
522 	/* The node we are removing has either zero or one child. If there
523 	 * is a child, move it into the removed node's slot then delete
524 	 * the node.  Otherwise just clear the slot and delete the node.
525 	 */
526 	if (node->child[0])
527 		rcu_assign_pointer(*trim, rcu_access_pointer(node->child[0]));
528 	else if (node->child[1])
529 		rcu_assign_pointer(*trim, rcu_access_pointer(node->child[1]));
530 	else
531 		RCU_INIT_POINTER(*trim, NULL);
532 	kfree_rcu(node, rcu);
533 
534 out:
535 	spin_unlock_irqrestore(&trie->lock, irq_flags);
536 
537 	return ret;
538 }
539 
540 #define LPM_DATA_SIZE_MAX	256
541 #define LPM_DATA_SIZE_MIN	1
542 
543 #define LPM_VAL_SIZE_MAX	(KMALLOC_MAX_SIZE - LPM_DATA_SIZE_MAX - \
544 				 sizeof(struct lpm_trie_node))
545 #define LPM_VAL_SIZE_MIN	1
546 
547 #define LPM_KEY_SIZE(X)		(sizeof(struct bpf_lpm_trie_key_u8) + (X))
548 #define LPM_KEY_SIZE_MAX	LPM_KEY_SIZE(LPM_DATA_SIZE_MAX)
549 #define LPM_KEY_SIZE_MIN	LPM_KEY_SIZE(LPM_DATA_SIZE_MIN)
550 
551 #define LPM_CREATE_FLAG_MASK	(BPF_F_NO_PREALLOC | BPF_F_NUMA_NODE |	\
552 				 BPF_F_ACCESS_MASK)
553 
554 static struct bpf_map *trie_alloc(union bpf_attr *attr)
555 {
556 	struct lpm_trie *trie;
557 
558 	/* check sanity of attributes */
559 	if (attr->max_entries == 0 ||
560 	    !(attr->map_flags & BPF_F_NO_PREALLOC) ||
561 	    attr->map_flags & ~LPM_CREATE_FLAG_MASK ||
562 	    !bpf_map_flags_access_ok(attr->map_flags) ||
563 	    attr->key_size < LPM_KEY_SIZE_MIN ||
564 	    attr->key_size > LPM_KEY_SIZE_MAX ||
565 	    attr->value_size < LPM_VAL_SIZE_MIN ||
566 	    attr->value_size > LPM_VAL_SIZE_MAX)
567 		return ERR_PTR(-EINVAL);
568 
569 	trie = bpf_map_area_alloc(sizeof(*trie), NUMA_NO_NODE);
570 	if (!trie)
571 		return ERR_PTR(-ENOMEM);
572 
573 	/* copy mandatory map attributes */
574 	bpf_map_init_from_attr(&trie->map, attr);
575 	trie->data_size = attr->key_size -
576 			  offsetof(struct bpf_lpm_trie_key_u8, data);
577 	trie->max_prefixlen = trie->data_size * 8;
578 
579 	spin_lock_init(&trie->lock);
580 
581 	return &trie->map;
582 }
583 
584 static void trie_free(struct bpf_map *map)
585 {
586 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
587 	struct lpm_trie_node __rcu **slot;
588 	struct lpm_trie_node *node;
589 
590 	/* Always start at the root and walk down to a node that has no
591 	 * children. Then free that node, nullify its reference in the parent
592 	 * and start over.
593 	 */
594 
595 	for (;;) {
596 		slot = &trie->root;
597 
598 		for (;;) {
599 			node = rcu_dereference_protected(*slot, 1);
600 			if (!node)
601 				goto out;
602 
603 			if (rcu_access_pointer(node->child[0])) {
604 				slot = &node->child[0];
605 				continue;
606 			}
607 
608 			if (rcu_access_pointer(node->child[1])) {
609 				slot = &node->child[1];
610 				continue;
611 			}
612 
613 			kfree(node);
614 			RCU_INIT_POINTER(*slot, NULL);
615 			break;
616 		}
617 	}
618 
619 out:
620 	bpf_map_area_free(trie);
621 }
622 
623 static int trie_get_next_key(struct bpf_map *map, void *_key, void *_next_key)
624 {
625 	struct lpm_trie_node *node, *next_node = NULL, *parent, *search_root;
626 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
627 	struct bpf_lpm_trie_key_u8 *key = _key, *next_key = _next_key;
628 	struct lpm_trie_node **node_stack = NULL;
629 	int err = 0, stack_ptr = -1;
630 	unsigned int next_bit;
631 	size_t matchlen;
632 
633 	/* The get_next_key follows postorder. For the 4 node example in
634 	 * the top of this file, the trie_get_next_key() returns the following
635 	 * one after another:
636 	 *   192.168.0.0/24
637 	 *   192.168.1.0/24
638 	 *   192.168.128.0/24
639 	 *   192.168.0.0/16
640 	 *
641 	 * The idea is to return more specific keys before less specific ones.
642 	 */
643 
644 	/* Empty trie */
645 	search_root = rcu_dereference(trie->root);
646 	if (!search_root)
647 		return -ENOENT;
648 
649 	/* For invalid key, find the leftmost node in the trie */
650 	if (!key || key->prefixlen > trie->max_prefixlen)
651 		goto find_leftmost;
652 
653 	node_stack = kmalloc_array(trie->max_prefixlen,
654 				   sizeof(struct lpm_trie_node *),
655 				   GFP_ATOMIC | __GFP_NOWARN);
656 	if (!node_stack)
657 		return -ENOMEM;
658 
659 	/* Try to find the exact node for the given key */
660 	for (node = search_root; node;) {
661 		node_stack[++stack_ptr] = node;
662 		matchlen = longest_prefix_match(trie, node, key);
663 		if (node->prefixlen != matchlen ||
664 		    node->prefixlen == key->prefixlen)
665 			break;
666 
667 		next_bit = extract_bit(key->data, node->prefixlen);
668 		node = rcu_dereference(node->child[next_bit]);
669 	}
670 	if (!node || node->prefixlen != key->prefixlen ||
671 	    (node->flags & LPM_TREE_NODE_FLAG_IM))
672 		goto find_leftmost;
673 
674 	/* The node with the exactly-matching key has been found,
675 	 * find the first node in postorder after the matched node.
676 	 */
677 	node = node_stack[stack_ptr];
678 	while (stack_ptr > 0) {
679 		parent = node_stack[stack_ptr - 1];
680 		if (rcu_dereference(parent->child[0]) == node) {
681 			search_root = rcu_dereference(parent->child[1]);
682 			if (search_root)
683 				goto find_leftmost;
684 		}
685 		if (!(parent->flags & LPM_TREE_NODE_FLAG_IM)) {
686 			next_node = parent;
687 			goto do_copy;
688 		}
689 
690 		node = parent;
691 		stack_ptr--;
692 	}
693 
694 	/* did not find anything */
695 	err = -ENOENT;
696 	goto free_stack;
697 
698 find_leftmost:
699 	/* Find the leftmost non-intermediate node, all intermediate nodes
700 	 * have exact two children, so this function will never return NULL.
701 	 */
702 	for (node = search_root; node;) {
703 		if (node->flags & LPM_TREE_NODE_FLAG_IM) {
704 			node = rcu_dereference(node->child[0]);
705 		} else {
706 			next_node = node;
707 			node = rcu_dereference(node->child[0]);
708 			if (!node)
709 				node = rcu_dereference(next_node->child[1]);
710 		}
711 	}
712 do_copy:
713 	next_key->prefixlen = next_node->prefixlen;
714 	memcpy((void *)next_key + offsetof(struct bpf_lpm_trie_key_u8, data),
715 	       next_node->data, trie->data_size);
716 free_stack:
717 	kfree(node_stack);
718 	return err;
719 }
720 
721 static int trie_check_btf(const struct bpf_map *map,
722 			  const struct btf *btf,
723 			  const struct btf_type *key_type,
724 			  const struct btf_type *value_type)
725 {
726 	/* Keys must have struct bpf_lpm_trie_key_u8 embedded. */
727 	return BTF_INFO_KIND(key_type->info) != BTF_KIND_STRUCT ?
728 	       -EINVAL : 0;
729 }
730 
731 static u64 trie_mem_usage(const struct bpf_map *map)
732 {
733 	struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
734 	u64 elem_size;
735 
736 	elem_size = sizeof(struct lpm_trie_node) + trie->data_size +
737 			    trie->map.value_size;
738 	return elem_size * READ_ONCE(trie->n_entries);
739 }
740 
741 BTF_ID_LIST_SINGLE(trie_map_btf_ids, struct, lpm_trie)
742 const struct bpf_map_ops trie_map_ops = {
743 	.map_meta_equal = bpf_map_meta_equal,
744 	.map_alloc = trie_alloc,
745 	.map_free = trie_free,
746 	.map_get_next_key = trie_get_next_key,
747 	.map_lookup_elem = trie_lookup_elem,
748 	.map_update_elem = trie_update_elem,
749 	.map_delete_elem = trie_delete_elem,
750 	.map_lookup_batch = generic_map_lookup_batch,
751 	.map_update_batch = generic_map_update_batch,
752 	.map_delete_batch = generic_map_delete_batch,
753 	.map_check_btf = trie_check_btf,
754 	.map_mem_usage = trie_mem_usage,
755 	.map_btf_id = &trie_map_btf_ids[0],
756 };
757