2 * Longest prefix match list implementation
4 * Copyright (c) 2016,2017 Daniel Mack
5 * Copyright (c) 2016 David Herrmann
7 * This file is subject to the terms and conditions of version 2 of the GNU
8 * General Public License. See the file COPYING in the main directory of the
9 * Linux distribution for more details.
12 #include <linux/bpf.h>
13 #include <linux/btf.h>
14 #include <linux/err.h>
15 #include <linux/slab.h>
16 #include <linux/spinlock.h>
17 #include <linux/vmalloc.h>
19 #include <uapi/linux/btf.h>
21 /* Intermediate node */
22 #define LPM_TREE_NODE_FLAG_IM BIT(0)
26 struct lpm_trie_node {
28 struct lpm_trie_node __rcu *child[2];
36 struct lpm_trie_node __rcu *root;
43 /* This trie implements a longest prefix match algorithm that can be used to
44 * match IP addresses to a stored set of ranges.
46 * Data stored in @data of struct bpf_lpm_key and struct lpm_trie_node is
47 * interpreted as big endian, so data[0] stores the most significant byte.
49 * Match ranges are internally stored in instances of struct lpm_trie_node
50 * which each contain their prefix length as well as two pointers that may
51 * lead to more nodes containing more specific matches. Each node also stores
52 * a value that is defined by and returned to userspace via the update_elem
53 * and lookup functions.
55 * For instance, let's start with a trie that was created with a prefix length
56 * of 32, so it can be used for IPv4 addresses, and one single element that
57 * matches 192.168.0.0/16. The data array would hence contain
58 * [0xc0, 0xa8, 0x00, 0x00] in big-endian notation. This documentation will
59 * stick to IP-address notation for readability though.
61 * As the trie is empty initially, the new node (1) will be places as root
62 * node, denoted as (R) in the example below. As there are no other node, both
63 * child pointers are %NULL.
72 * Next, let's add a new node (2) matching 192.168.0.0/24. As there is already
73 * a node with the same data and a smaller prefix (ie, a less specific one),
74 * node (2) will become a child of (1). In child index depends on the next bit
75 * that is outside of what (1) matches, and that bit is 0, so (2) will be
92 * The child[1] slot of (1) could be filled with another node which has bit #17
93 * (the next bit after the ones that (1) matches on) set to 1. For instance,
103 * +----------------+ +------------------+
105 * | 192.168.0.0/24 | | 192.168.128.0/24 |
106 * | value: 2 | | value: 3 |
107 * | [0] [1] | | [0] [1] |
108 * +----------------+ +------------------+
110 * Let's add another node (4) to the game for 192.168.1.0/24. In order to place
111 * it, node (1) is looked at first, and because (4) of the semantics laid out
112 * above (bit #17 is 0), it would normally be attached to (1) as child[0].
113 * However, that slot is already allocated, so a new node is needed in between.
114 * That node does not have a value attached to it and it will never be
115 * returned to users as result of a lookup. It is only there to differentiate
116 * the traversal further. It will get a prefix as wide as necessary to
117 * distinguish its two children:
126 * +----------------+ +------------------+
127 * | (4) (I) | | (3) |
128 * | 192.168.0.0/23 | | 192.168.128.0/24 |
129 * | value: --- | | value: 3 |
130 * | [0] [1] | | [0] [1] |
131 * +----------------+ +------------------+
133 * +----------------+ +----------------+
135 * | 192.168.0.0/24 | | 192.168.1.0/24 |
136 * | value: 2 | | value: 5 |
137 * | [0] [1] | | [0] [1] |
138 * +----------------+ +----------------+
140 * 192.168.1.1/32 would be a child of (5) etc.
142 * An intermediate node will be turned into a 'real' node on demand. In the
143 * example above, (4) would be re-used if 192.168.0.0/23 is added to the trie.
145 * A fully populated trie would have a height of 32 nodes, as the trie was
146 * created with a prefix length of 32.
148 * The lookup starts at the root node. If the current node matches and if there
149 * is a child that can be used to become more specific, the trie is traversed
150 * downwards. The last node in the traversal that is a non-intermediate one is
154 static inline int extract_bit(const u8 *data, size_t index)
156 return !!(data[index / 8] & (1 << (7 - (index % 8))));
160 * longest_prefix_match() - determine the longest prefix
161 * @trie: The trie to get internal sizes from
162 * @node: The node to operate on
163 * @key: The key to compare to @node
165 * Determine the longest prefix of @node that matches the bits in @key.
167 static size_t longest_prefix_match(const struct lpm_trie *trie,
168 const struct lpm_trie_node *node,
169 const struct bpf_lpm_trie_key *key)
171 size_t prefixlen = 0;
174 for (i = 0; i < trie->data_size; i++) {
177 b = 8 - fls(node->data[i] ^ key->data[i]);
180 if (prefixlen >= node->prefixlen || prefixlen >= key->prefixlen)
181 return min(node->prefixlen, key->prefixlen);
190 /* Called from syscall or from eBPF program */
191 static void *trie_lookup_elem(struct bpf_map *map, void *_key)
193 struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
194 struct lpm_trie_node *node, *found = NULL;
195 struct bpf_lpm_trie_key *key = _key;
197 /* Start walking the trie from the root node ... */
199 for (node = rcu_dereference(trie->root); node;) {
200 unsigned int next_bit;
203 /* Determine the longest prefix of @node that matches @key.
204 * If it's the maximum possible prefix for this trie, we have
205 * an exact match and can return it directly.
207 matchlen = longest_prefix_match(trie, node, key);
208 if (matchlen == trie->max_prefixlen) {
213 /* If the number of bits that match is smaller than the prefix
214 * length of @node, bail out and return the node we have seen
215 * last in the traversal (ie, the parent).
217 if (matchlen < node->prefixlen)
220 /* Consider this node as return candidate unless it is an
221 * artificially added intermediate one.
223 if (!(node->flags & LPM_TREE_NODE_FLAG_IM))
226 /* If the node match is fully satisfied, let's see if we can
227 * become more specific. Determine the next bit in the key and
230 next_bit = extract_bit(key->data, node->prefixlen);
231 node = rcu_dereference(node->child[next_bit]);
237 return found->data + trie->data_size;
240 static struct lpm_trie_node *lpm_trie_node_alloc(const struct lpm_trie *trie,
243 struct lpm_trie_node *node;
244 size_t size = sizeof(struct lpm_trie_node) + trie->data_size;
247 size += trie->map.value_size;
249 node = kmalloc_node(size, GFP_ATOMIC | __GFP_NOWARN,
250 trie->map.numa_node);
257 memcpy(node->data + trie->data_size, value,
258 trie->map.value_size);
263 /* Called from syscall or from eBPF program */
264 static int trie_update_elem(struct bpf_map *map,
265 void *_key, void *value, u64 flags)
267 struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
268 struct lpm_trie_node *node, *im_node = NULL, *new_node = NULL;
269 struct lpm_trie_node __rcu **slot;
270 struct bpf_lpm_trie_key *key = _key;
271 unsigned long irq_flags;
272 unsigned int next_bit;
276 if (unlikely(flags > BPF_EXIST))
279 if (key->prefixlen > trie->max_prefixlen)
282 raw_spin_lock_irqsave(&trie->lock, irq_flags);
284 /* Allocate and fill a new node */
286 if (trie->n_entries == trie->map.max_entries) {
291 new_node = lpm_trie_node_alloc(trie, value);
299 new_node->prefixlen = key->prefixlen;
300 RCU_INIT_POINTER(new_node->child[0], NULL);
301 RCU_INIT_POINTER(new_node->child[1], NULL);
302 memcpy(new_node->data, key->data, trie->data_size);
304 /* Now find a slot to attach the new node. To do that, walk the tree
305 * from the root and match as many bits as possible for each node until
306 * we either find an empty slot or a slot that needs to be replaced by
307 * an intermediate node.
311 while ((node = rcu_dereference_protected(*slot,
312 lockdep_is_held(&trie->lock)))) {
313 matchlen = longest_prefix_match(trie, node, key);
315 if (node->prefixlen != matchlen ||
316 node->prefixlen == key->prefixlen ||
317 node->prefixlen == trie->max_prefixlen)
320 next_bit = extract_bit(key->data, node->prefixlen);
321 slot = &node->child[next_bit];
324 /* If the slot is empty (a free child pointer or an empty root),
325 * simply assign the @new_node to that slot and be done.
328 rcu_assign_pointer(*slot, new_node);
332 /* If the slot we picked already exists, replace it with @new_node
333 * which already has the correct data array set.
335 if (node->prefixlen == matchlen) {
336 new_node->child[0] = node->child[0];
337 new_node->child[1] = node->child[1];
339 if (!(node->flags & LPM_TREE_NODE_FLAG_IM))
342 rcu_assign_pointer(*slot, new_node);
343 kfree_rcu(node, rcu);
348 /* If the new node matches the prefix completely, it must be inserted
349 * as an ancestor. Simply insert it between @node and *@slot.
351 if (matchlen == key->prefixlen) {
352 next_bit = extract_bit(node->data, matchlen);
353 rcu_assign_pointer(new_node->child[next_bit], node);
354 rcu_assign_pointer(*slot, new_node);
358 im_node = lpm_trie_node_alloc(trie, NULL);
364 im_node->prefixlen = matchlen;
365 im_node->flags |= LPM_TREE_NODE_FLAG_IM;
366 memcpy(im_node->data, node->data, trie->data_size);
368 /* Now determine which child to install in which slot */
369 if (extract_bit(key->data, matchlen)) {
370 rcu_assign_pointer(im_node->child[0], node);
371 rcu_assign_pointer(im_node->child[1], new_node);
373 rcu_assign_pointer(im_node->child[0], new_node);
374 rcu_assign_pointer(im_node->child[1], node);
377 /* Finally, assign the intermediate node to the determined spot */
378 rcu_assign_pointer(*slot, im_node);
389 raw_spin_unlock_irqrestore(&trie->lock, irq_flags);
394 /* Called from syscall or from eBPF program */
395 static int trie_delete_elem(struct bpf_map *map, void *_key)
397 struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
398 struct bpf_lpm_trie_key *key = _key;
399 struct lpm_trie_node __rcu **trim, **trim2;
400 struct lpm_trie_node *node, *parent;
401 unsigned long irq_flags;
402 unsigned int next_bit;
406 if (key->prefixlen > trie->max_prefixlen)
409 raw_spin_lock_irqsave(&trie->lock, irq_flags);
411 /* Walk the tree looking for an exact key/length match and keeping
412 * track of the path we traverse. We will need to know the node
413 * we wish to delete, and the slot that points to the node we want
414 * to delete. We may also need to know the nodes parent and the
415 * slot that contains it.
420 while ((node = rcu_dereference_protected(
421 *trim, lockdep_is_held(&trie->lock)))) {
422 matchlen = longest_prefix_match(trie, node, key);
424 if (node->prefixlen != matchlen ||
425 node->prefixlen == key->prefixlen)
430 next_bit = extract_bit(key->data, node->prefixlen);
431 trim = &node->child[next_bit];
434 if (!node || node->prefixlen != key->prefixlen ||
435 node->prefixlen != matchlen ||
436 (node->flags & LPM_TREE_NODE_FLAG_IM)) {
443 /* If the node we are removing has two children, simply mark it
444 * as intermediate and we are done.
446 if (rcu_access_pointer(node->child[0]) &&
447 rcu_access_pointer(node->child[1])) {
448 node->flags |= LPM_TREE_NODE_FLAG_IM;
452 /* If the parent of the node we are about to delete is an intermediate
453 * node, and the deleted node doesn't have any children, we can delete
454 * the intermediate parent as well and promote its other child
455 * up the tree. Doing this maintains the invariant that all
456 * intermediate nodes have exactly 2 children and that there are no
457 * unnecessary intermediate nodes in the tree.
459 if (parent && (parent->flags & LPM_TREE_NODE_FLAG_IM) &&
460 !node->child[0] && !node->child[1]) {
461 if (node == rcu_access_pointer(parent->child[0]))
463 *trim2, rcu_access_pointer(parent->child[1]));
466 *trim2, rcu_access_pointer(parent->child[0]));
467 kfree_rcu(parent, rcu);
468 kfree_rcu(node, rcu);
472 /* The node we are removing has either zero or one child. If there
473 * is a child, move it into the removed node's slot then delete
474 * the node. Otherwise just clear the slot and delete the node.
477 rcu_assign_pointer(*trim, rcu_access_pointer(node->child[0]));
478 else if (node->child[1])
479 rcu_assign_pointer(*trim, rcu_access_pointer(node->child[1]));
481 RCU_INIT_POINTER(*trim, NULL);
482 kfree_rcu(node, rcu);
485 raw_spin_unlock_irqrestore(&trie->lock, irq_flags);
490 #define LPM_DATA_SIZE_MAX 256
491 #define LPM_DATA_SIZE_MIN 1
493 #define LPM_VAL_SIZE_MAX (KMALLOC_MAX_SIZE - LPM_DATA_SIZE_MAX - \
494 sizeof(struct lpm_trie_node))
495 #define LPM_VAL_SIZE_MIN 1
497 #define LPM_KEY_SIZE(X) (sizeof(struct bpf_lpm_trie_key) + (X))
498 #define LPM_KEY_SIZE_MAX LPM_KEY_SIZE(LPM_DATA_SIZE_MAX)
499 #define LPM_KEY_SIZE_MIN LPM_KEY_SIZE(LPM_DATA_SIZE_MIN)
501 #define LPM_CREATE_FLAG_MASK (BPF_F_NO_PREALLOC | BPF_F_NUMA_NODE | \
502 BPF_F_RDONLY | BPF_F_WRONLY)
504 static struct bpf_map *trie_alloc(union bpf_attr *attr)
506 struct lpm_trie *trie;
507 u64 cost = sizeof(*trie), cost_per_node;
510 if (!capable(CAP_SYS_ADMIN))
511 return ERR_PTR(-EPERM);
513 /* check sanity of attributes */
514 if (attr->max_entries == 0 ||
515 !(attr->map_flags & BPF_F_NO_PREALLOC) ||
516 attr->map_flags & ~LPM_CREATE_FLAG_MASK ||
517 attr->key_size < LPM_KEY_SIZE_MIN ||
518 attr->key_size > LPM_KEY_SIZE_MAX ||
519 attr->value_size < LPM_VAL_SIZE_MIN ||
520 attr->value_size > LPM_VAL_SIZE_MAX)
521 return ERR_PTR(-EINVAL);
523 trie = kzalloc(sizeof(*trie), GFP_USER | __GFP_NOWARN);
525 return ERR_PTR(-ENOMEM);
527 /* copy mandatory map attributes */
528 bpf_map_init_from_attr(&trie->map, attr);
529 trie->data_size = attr->key_size -
530 offsetof(struct bpf_lpm_trie_key, data);
531 trie->max_prefixlen = trie->data_size * 8;
533 cost_per_node = sizeof(struct lpm_trie_node) +
534 attr->value_size + trie->data_size;
535 cost += (u64) attr->max_entries * cost_per_node;
536 if (cost >= U32_MAX - PAGE_SIZE) {
541 trie->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
543 ret = bpf_map_precharge_memlock(trie->map.pages);
547 raw_spin_lock_init(&trie->lock);
555 static void trie_free(struct bpf_map *map)
557 struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
558 struct lpm_trie_node __rcu **slot;
559 struct lpm_trie_node *node;
561 /* Wait for outstanding programs to complete
562 * update/lookup/delete/get_next_key and free the trie.
566 /* Always start at the root and walk down to a node that has no
567 * children. Then free that node, nullify its reference in the parent
575 node = rcu_dereference_protected(*slot, 1);
579 if (rcu_access_pointer(node->child[0])) {
580 slot = &node->child[0];
584 if (rcu_access_pointer(node->child[1])) {
585 slot = &node->child[1];
590 RCU_INIT_POINTER(*slot, NULL);
599 static int trie_get_next_key(struct bpf_map *map, void *_key, void *_next_key)
601 struct lpm_trie_node *node, *next_node = NULL, *parent, *search_root;
602 struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
603 struct bpf_lpm_trie_key *key = _key, *next_key = _next_key;
604 struct lpm_trie_node **node_stack = NULL;
605 int err = 0, stack_ptr = -1;
606 unsigned int next_bit;
609 /* The get_next_key follows postorder. For the 4 node example in
610 * the top of this file, the trie_get_next_key() returns the following
617 * The idea is to return more specific keys before less specific ones.
621 search_root = rcu_dereference(trie->root);
625 /* For invalid key, find the leftmost node in the trie */
626 if (!key || key->prefixlen > trie->max_prefixlen)
629 node_stack = kmalloc_array(trie->max_prefixlen,
630 sizeof(struct lpm_trie_node *),
631 GFP_ATOMIC | __GFP_NOWARN);
635 /* Try to find the exact node for the given key */
636 for (node = search_root; node;) {
637 node_stack[++stack_ptr] = node;
638 matchlen = longest_prefix_match(trie, node, key);
639 if (node->prefixlen != matchlen ||
640 node->prefixlen == key->prefixlen)
643 next_bit = extract_bit(key->data, node->prefixlen);
644 node = rcu_dereference(node->child[next_bit]);
646 if (!node || node->prefixlen != key->prefixlen ||
647 (node->flags & LPM_TREE_NODE_FLAG_IM))
650 /* The node with the exactly-matching key has been found,
651 * find the first node in postorder after the matched node.
653 node = node_stack[stack_ptr];
654 while (stack_ptr > 0) {
655 parent = node_stack[stack_ptr - 1];
656 if (rcu_dereference(parent->child[0]) == node) {
657 search_root = rcu_dereference(parent->child[1]);
661 if (!(parent->flags & LPM_TREE_NODE_FLAG_IM)) {
670 /* did not find anything */
675 /* Find the leftmost non-intermediate node, all intermediate nodes
676 * have exact two children, so this function will never return NULL.
678 for (node = search_root; node;) {
679 if (node->flags & LPM_TREE_NODE_FLAG_IM) {
680 node = rcu_dereference(node->child[0]);
683 node = rcu_dereference(node->child[0]);
685 node = rcu_dereference(next_node->child[1]);
689 next_key->prefixlen = next_node->prefixlen;
690 memcpy((void *)next_key + offsetof(struct bpf_lpm_trie_key, data),
691 next_node->data, trie->data_size);
697 static int trie_check_btf(const struct bpf_map *map,
698 const struct btf_type *key_type,
699 const struct btf_type *value_type)
701 /* Keys must have struct bpf_lpm_trie_key embedded. */
702 return BTF_INFO_KIND(key_type->info) != BTF_KIND_STRUCT ?
706 const struct bpf_map_ops trie_map_ops = {
707 .map_alloc = trie_alloc,
708 .map_free = trie_free,
709 .map_get_next_key = trie_get_next_key,
710 .map_lookup_elem = trie_lookup_elem,
711 .map_update_elem = trie_update_elem,
712 .map_delete_elem = trie_delete_elem,
713 .map_check_btf = trie_check_btf,