1 // SPDX-License-Identifier: GPL-2.0-only
3 * Longest prefix match list implementation
5 * Copyright (c) 2016,2017 Daniel Mack
6 * Copyright (c) 2016 David Herrmann
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>
16 #include <uapi/linux/btf.h>
17 #include <linux/btf_ids.h>
19 /* Intermediate node */
20 #define LPM_TREE_NODE_FLAG_IM BIT(0)
24 struct lpm_trie_node {
26 struct lpm_trie_node __rcu *child[2];
34 struct lpm_trie_node __rcu *root;
41 /* This trie implements a longest prefix match algorithm that can be used to
42 * match IP addresses to a stored set of ranges.
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.
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.
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.
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.
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
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,
101 * +----------------+ +------------------+
103 * | 192.168.0.0/24 | | 192.168.128.0/24 |
104 * | value: 2 | | value: 3 |
105 * | [0] [1] | | [0] [1] |
106 * +----------------+ +------------------+
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:
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 * +----------------+ +------------------+
131 * +----------------+ +----------------+
133 * | 192.168.0.0/24 | | 192.168.1.0/24 |
134 * | value: 2 | | value: 5 |
135 * | [0] [1] | | [0] [1] |
136 * +----------------+ +----------------+
138 * 192.168.1.1/32 would be a child of (5) etc.
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.
143 * A fully populated trie would have a height of 32 nodes, as the trie was
144 * created with a prefix length of 32.
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
152 static inline int extract_bit(const u8 *data, size_t index)
154 return !!(data[index / 8] & (1 << (7 - (index % 8))));
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
163 * Determine the longest prefix of @node that matches the bits in @key.
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)
169 u32 limit = min(node->prefixlen, key->prefixlen);
170 u32 prefixlen = 0, i = 0;
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));
175 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && defined(CONFIG_64BIT)
177 /* data_size >= 16 has very small probability.
178 * We do not use a loop for optimal code generation.
180 if (trie->data_size >= 8) {
181 u64 diff = be64_to_cpu(*(__be64 *)node->data ^
182 *(__be64 *)key->data);
184 prefixlen = 64 - fls64(diff);
185 if (prefixlen >= limit)
193 while (trie->data_size >= i + 4) {
194 u32 diff = be32_to_cpu(*(__be32 *)&node->data[i] ^
195 *(__be32 *)&key->data[i]);
197 prefixlen += 32 - fls(diff);
198 if (prefixlen >= limit)
205 if (trie->data_size >= i + 2) {
206 u16 diff = be16_to_cpu(*(__be16 *)&node->data[i] ^
207 *(__be16 *)&key->data[i]);
209 prefixlen += 16 - fls(diff);
210 if (prefixlen >= limit)
217 if (trie->data_size >= i + 1) {
218 prefixlen += 8 - fls(node->data[i] ^ key->data[i]);
220 if (prefixlen >= limit)
227 /* Called from syscall or from eBPF program */
228 static void *trie_lookup_elem(struct bpf_map *map, void *_key)
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;
234 /* Start walking the trie from the root node ... */
236 for (node = rcu_dereference_check(trie->root, rcu_read_lock_bh_held());
238 unsigned int next_bit;
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.
245 matchlen = longest_prefix_match(trie, node, key);
246 if (matchlen == trie->max_prefixlen) {
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).
255 if (matchlen < node->prefixlen)
258 /* Consider this node as return candidate unless it is an
259 * artificially added intermediate one.
261 if (!(node->flags & LPM_TREE_NODE_FLAG_IM))
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
268 next_bit = extract_bit(key->data, node->prefixlen);
269 node = rcu_dereference_check(node->child[next_bit],
270 rcu_read_lock_bh_held());
276 return found->data + trie->data_size;
279 static struct lpm_trie_node *lpm_trie_node_alloc(const struct lpm_trie *trie,
282 struct lpm_trie_node *node;
283 size_t size = sizeof(struct lpm_trie_node) + trie->data_size;
286 size += trie->map.value_size;
288 node = bpf_map_kmalloc_node(&trie->map, size, GFP_NOWAIT | __GFP_NOWARN,
289 trie->map.numa_node);
296 memcpy(node->data + trie->data_size, value,
297 trie->map.value_size);
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)
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;
315 if (unlikely(flags > BPF_EXIST))
318 if (key->prefixlen > trie->max_prefixlen)
321 spin_lock_irqsave(&trie->lock, irq_flags);
323 /* Allocate and fill a new node */
325 if (trie->n_entries == trie->map.max_entries) {
330 new_node = lpm_trie_node_alloc(trie, value);
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);
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.
350 while ((node = rcu_dereference_protected(*slot,
351 lockdep_is_held(&trie->lock)))) {
352 matchlen = longest_prefix_match(trie, node, key);
354 if (node->prefixlen != matchlen ||
355 node->prefixlen == key->prefixlen ||
356 node->prefixlen == trie->max_prefixlen)
359 next_bit = extract_bit(key->data, node->prefixlen);
360 slot = &node->child[next_bit];
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.
367 rcu_assign_pointer(*slot, new_node);
371 /* If the slot we picked already exists, replace it with @new_node
372 * which already has the correct data array set.
374 if (node->prefixlen == matchlen) {
375 new_node->child[0] = node->child[0];
376 new_node->child[1] = node->child[1];
378 if (!(node->flags & LPM_TREE_NODE_FLAG_IM))
381 rcu_assign_pointer(*slot, new_node);
382 kfree_rcu(node, rcu);
387 /* If the new node matches the prefix completely, it must be inserted
388 * as an ancestor. Simply insert it between @node and *@slot.
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);
397 im_node = lpm_trie_node_alloc(trie, NULL);
403 im_node->prefixlen = matchlen;
404 im_node->flags |= LPM_TREE_NODE_FLAG_IM;
405 memcpy(im_node->data, node->data, trie->data_size);
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);
412 rcu_assign_pointer(im_node->child[0], new_node);
413 rcu_assign_pointer(im_node->child[1], node);
416 /* Finally, assign the intermediate node to the determined slot */
417 rcu_assign_pointer(*slot, im_node);
428 spin_unlock_irqrestore(&trie->lock, irq_flags);
433 /* Called from syscall or from eBPF program */
434 static long trie_delete_elem(struct bpf_map *map, void *_key)
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;
445 if (key->prefixlen > trie->max_prefixlen)
448 spin_lock_irqsave(&trie->lock, irq_flags);
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.
459 while ((node = rcu_dereference_protected(
460 *trim, lockdep_is_held(&trie->lock)))) {
461 matchlen = longest_prefix_match(trie, node, key);
463 if (node->prefixlen != matchlen ||
464 node->prefixlen == key->prefixlen)
469 next_bit = extract_bit(key->data, node->prefixlen);
470 trim = &node->child[next_bit];
473 if (!node || node->prefixlen != key->prefixlen ||
474 node->prefixlen != matchlen ||
475 (node->flags & LPM_TREE_NODE_FLAG_IM)) {
482 /* If the node we are removing has two children, simply mark it
483 * as intermediate and we are done.
485 if (rcu_access_pointer(node->child[0]) &&
486 rcu_access_pointer(node->child[1])) {
487 node->flags |= LPM_TREE_NODE_FLAG_IM;
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.
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]))
502 *trim2, rcu_access_pointer(parent->child[1]));
505 *trim2, rcu_access_pointer(parent->child[0]));
506 kfree_rcu(parent, rcu);
507 kfree_rcu(node, rcu);
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.
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]));
520 RCU_INIT_POINTER(*trim, NULL);
521 kfree_rcu(node, rcu);
524 spin_unlock_irqrestore(&trie->lock, irq_flags);
529 #define LPM_DATA_SIZE_MAX 256
530 #define LPM_DATA_SIZE_MIN 1
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
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)
540 #define LPM_CREATE_FLAG_MASK (BPF_F_NO_PREALLOC | BPF_F_NUMA_NODE | \
543 static struct bpf_map *trie_alloc(union bpf_attr *attr)
545 struct lpm_trie *trie;
547 /* check sanity of attributes */
548 if (attr->max_entries == 0 ||
549 !(attr->map_flags & BPF_F_NO_PREALLOC) ||
550 attr->map_flags & ~LPM_CREATE_FLAG_MASK ||
551 !bpf_map_flags_access_ok(attr->map_flags) ||
552 attr->key_size < LPM_KEY_SIZE_MIN ||
553 attr->key_size > LPM_KEY_SIZE_MAX ||
554 attr->value_size < LPM_VAL_SIZE_MIN ||
555 attr->value_size > LPM_VAL_SIZE_MAX)
556 return ERR_PTR(-EINVAL);
558 trie = bpf_map_area_alloc(sizeof(*trie), NUMA_NO_NODE);
560 return ERR_PTR(-ENOMEM);
562 /* copy mandatory map attributes */
563 bpf_map_init_from_attr(&trie->map, attr);
564 trie->data_size = attr->key_size -
565 offsetof(struct bpf_lpm_trie_key, data);
566 trie->max_prefixlen = trie->data_size * 8;
568 spin_lock_init(&trie->lock);
573 static void trie_free(struct bpf_map *map)
575 struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
576 struct lpm_trie_node __rcu **slot;
577 struct lpm_trie_node *node;
579 /* Always start at the root and walk down to a node that has no
580 * children. Then free that node, nullify its reference in the parent
588 node = rcu_dereference_protected(*slot, 1);
592 if (rcu_access_pointer(node->child[0])) {
593 slot = &node->child[0];
597 if (rcu_access_pointer(node->child[1])) {
598 slot = &node->child[1];
603 RCU_INIT_POINTER(*slot, NULL);
609 bpf_map_area_free(trie);
612 static int trie_get_next_key(struct bpf_map *map, void *_key, void *_next_key)
614 struct lpm_trie_node *node, *next_node = NULL, *parent, *search_root;
615 struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
616 struct bpf_lpm_trie_key *key = _key, *next_key = _next_key;
617 struct lpm_trie_node **node_stack = NULL;
618 int err = 0, stack_ptr = -1;
619 unsigned int next_bit;
622 /* The get_next_key follows postorder. For the 4 node example in
623 * the top of this file, the trie_get_next_key() returns the following
630 * The idea is to return more specific keys before less specific ones.
634 search_root = rcu_dereference(trie->root);
638 /* For invalid key, find the leftmost node in the trie */
639 if (!key || key->prefixlen > trie->max_prefixlen)
642 node_stack = kmalloc_array(trie->max_prefixlen,
643 sizeof(struct lpm_trie_node *),
644 GFP_ATOMIC | __GFP_NOWARN);
648 /* Try to find the exact node for the given key */
649 for (node = search_root; node;) {
650 node_stack[++stack_ptr] = node;
651 matchlen = longest_prefix_match(trie, node, key);
652 if (node->prefixlen != matchlen ||
653 node->prefixlen == key->prefixlen)
656 next_bit = extract_bit(key->data, node->prefixlen);
657 node = rcu_dereference(node->child[next_bit]);
659 if (!node || node->prefixlen != key->prefixlen ||
660 (node->flags & LPM_TREE_NODE_FLAG_IM))
663 /* The node with the exactly-matching key has been found,
664 * find the first node in postorder after the matched node.
666 node = node_stack[stack_ptr];
667 while (stack_ptr > 0) {
668 parent = node_stack[stack_ptr - 1];
669 if (rcu_dereference(parent->child[0]) == node) {
670 search_root = rcu_dereference(parent->child[1]);
674 if (!(parent->flags & LPM_TREE_NODE_FLAG_IM)) {
683 /* did not find anything */
688 /* Find the leftmost non-intermediate node, all intermediate nodes
689 * have exact two children, so this function will never return NULL.
691 for (node = search_root; node;) {
692 if (node->flags & LPM_TREE_NODE_FLAG_IM) {
693 node = rcu_dereference(node->child[0]);
696 node = rcu_dereference(node->child[0]);
698 node = rcu_dereference(next_node->child[1]);
702 next_key->prefixlen = next_node->prefixlen;
703 memcpy((void *)next_key + offsetof(struct bpf_lpm_trie_key, data),
704 next_node->data, trie->data_size);
710 static int trie_check_btf(const struct bpf_map *map,
711 const struct btf *btf,
712 const struct btf_type *key_type,
713 const struct btf_type *value_type)
715 /* Keys must have struct bpf_lpm_trie_key embedded. */
716 return BTF_INFO_KIND(key_type->info) != BTF_KIND_STRUCT ?
720 static u64 trie_mem_usage(const struct bpf_map *map)
722 struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
725 elem_size = sizeof(struct lpm_trie_node) + trie->data_size +
726 trie->map.value_size;
727 return elem_size * READ_ONCE(trie->n_entries);
730 BTF_ID_LIST_SINGLE(trie_map_btf_ids, struct, lpm_trie)
731 const struct bpf_map_ops trie_map_ops = {
732 .map_meta_equal = bpf_map_meta_equal,
733 .map_alloc = trie_alloc,
734 .map_free = trie_free,
735 .map_get_next_key = trie_get_next_key,
736 .map_lookup_elem = trie_lookup_elem,
737 .map_update_elem = trie_update_elem,
738 .map_delete_elem = trie_delete_elem,
739 .map_lookup_batch = generic_map_lookup_batch,
740 .map_update_batch = generic_map_update_batch,
741 .map_delete_batch = generic_map_delete_batch,
742 .map_check_btf = trie_check_btf,
743 .map_mem_usage = trie_mem_usage,
744 .map_btf_id = &trie_map_btf_ids[0],