1 // SPDX-License-Identifier: GPL-2.0-only
18 /* Intermediate node */
25 	struct lpm_trie_node __rcu	*child[2];  member
48  * lead to more nodes containing more specific matches. Each node also stores
55  * [0xc0, 0xa8, 0x00, 0x00] in big-endian notation. This documentation will
56  * stick to IP-address notation for readability though.
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.
62  *              +----------------+
67  *              +----------------+
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
73  * child[0] of (1):
75  *              +----------------+
80  *              +----------------+
82  *    +----------------+
87  *    +----------------+
89  * The child[1] slot of (1) could be filled with another node which has bit #17
93  *              +----------------+
98  *              +----------------+
100  *    +----------------+  +------------------+
105  *    +----------------+  +------------------+
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
116  *                      +----------------+
121  *                      +----------------+
123  *            +----------------+  +------------------+
126  *            |    value: ---  |  |     value: 3     |
128  *            +----------------+  +------------------+
130  *  +----------------+  +----------------+
135  *  +----------------+  +----------------+
137  * 192.168.1.1/32 would be a child of (5) etc.
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.
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
153 	return !!(data[index / 8] & (1 << (7 - (index % 8))));  in extract_bit()
157  * longest_prefix_match() - determine the longest prefix
159  * @node:	The node to operate on
160  * @key:	The key to compare to @node
162  * Determine the longest prefix of @node that matches the bits in @key.
165 				   const struct lpm_trie_node *node,  in longest_prefix_match()  argument
168 	u32 limit = min(node->prefixlen, key->prefixlen);  in longest_prefix_match()
179 	if (trie->data_size >= 8) {  in longest_prefix_match()
180 		u64 diff = be64_to_cpu(*(__be64 *)node->data ^  in longest_prefix_match()
181 				       *(__be64 *)key->data);  in longest_prefix_match()
183 		prefixlen = 64 - fls64(diff);  in longest_prefix_match()
192 	while (trie->data_size >= i + 4) {  in longest_prefix_match()
193 		u32 diff = be32_to_cpu(*(__be32 *)&node->data[i] ^  in longest_prefix_match()
194 				       *(__be32 *)&key->data[i]);  in longest_prefix_match()
196 		prefixlen += 32 - fls(diff);  in longest_prefix_match()
204 	if (trie->data_size >= i + 2) {  in longest_prefix_match()
205 		u16 diff = be16_to_cpu(*(__be16 *)&node->data[i] ^  in longest_prefix_match()
206 				       *(__be16 *)&key->data[i]);  in longest_prefix_match()
208 		prefixlen += 16 - fls(diff);  in longest_prefix_match()
216 	if (trie->data_size >= i + 1) {  in longest_prefix_match()
217 		prefixlen += 8 - fls(node->data[i] ^ key->data[i]);  in longest_prefix_match()
230 	struct lpm_trie_node *node, *found = NULL;  in trie_lookup_elem()  local
233 	/* Start walking the trie from the root node ... */  in trie_lookup_elem()
235 	for (node = rcu_dereference_check(trie->root, rcu_read_lock_bh_held());  in trie_lookup_elem()
236 	     node;) {  in trie_lookup_elem()
240 		/* Determine the longest prefix of @node that matches @key.  in trie_lookup_elem()
244 		matchlen = longest_prefix_match(trie, node, key);  in trie_lookup_elem()
245 		if (matchlen == trie->max_prefixlen) {  in trie_lookup_elem()
246 			found = node;  in trie_lookup_elem()
251 		 * length of @node, bail out and return the node we have seen  in trie_lookup_elem()
254 		if (matchlen < node->prefixlen)  in trie_lookup_elem()
257 		/* Consider this node as return candidate unless it is an  in trie_lookup_elem()
260 		if (!(node->flags & LPM_TREE_NODE_FLAG_IM))  in trie_lookup_elem()
261 			found = node;  in trie_lookup_elem()
263 		/* If the node match is fully satisfied, let's see if we can  in trie_lookup_elem()
267 		next_bit = extract_bit(key->data, node->prefixlen);  in trie_lookup_elem()
268 		node = rcu_dereference_check(node->child[next_bit],  in trie_lookup_elem()
275 	return found->data + trie->data_size;  in trie_lookup_elem()
281 	struct lpm_trie_node *node;  in lpm_trie_node_alloc()  local
282 	size_t size = sizeof(struct lpm_trie_node) + trie->data_size;  in lpm_trie_node_alloc()
285 		size += trie->map.value_size;  in lpm_trie_node_alloc()
287 	node = bpf_map_kmalloc_node(&trie->map, size, GFP_ATOMIC | __GFP_NOWARN,  in lpm_trie_node_alloc()
288 				    trie->map.numa_node);  in lpm_trie_node_alloc()
289 	if (!node)  in lpm_trie_node_alloc()
292 	node->flags = 0;  in lpm_trie_node_alloc()
295 		memcpy(node->data + trie->data_size, value,  in lpm_trie_node_alloc()
296 		       trie->map.value_size);  in lpm_trie_node_alloc()
298 	return node;  in lpm_trie_node_alloc()
306 	struct lpm_trie_node *node, *im_node = NULL, *new_node = NULL;  in trie_update_elem()  local
315 		return -EINVAL;  in trie_update_elem()
317 	if (key->prefixlen > trie->max_prefixlen)  in trie_update_elem()
318 		return -EINVAL;  in trie_update_elem()
320 	spin_lock_irqsave(&trie->lock, irq_flags);  in trie_update_elem()
322 	/* Allocate and fill a new node */  in trie_update_elem()
324 	if (trie->n_entries == trie->map.max_entries) {  in trie_update_elem()
325 		ret = -ENOSPC;  in trie_update_elem()
331 		ret = -ENOMEM;  in trie_update_elem()
335 	trie->n_entries++;  in trie_update_elem()
337 	new_node->prefixlen = key->prefixlen;  in trie_update_elem()
338 	RCU_INIT_POINTER(new_node->child[0], NULL);  in trie_update_elem()
339 	RCU_INIT_POINTER(new_node->child[1], NULL);  in trie_update_elem()
340 	memcpy(new_node->data, key->data, trie->data_size);  in trie_update_elem()
342 	/* Now find a slot to attach the new node. To do that, walk the tree  in trie_update_elem()
343 	 * from the root and match as many bits as possible for each node until  in trie_update_elem()
345 	 * an intermediate node.  in trie_update_elem()
347 	slot = &trie->root;  in trie_update_elem()
349 	while ((node = rcu_dereference_protected(*slot,  in trie_update_elem()
350 					lockdep_is_held(&trie->lock)))) {  in trie_update_elem()
351 		matchlen = longest_prefix_match(trie, node, key);  in trie_update_elem()
353 		if (node->prefixlen != matchlen ||  in trie_update_elem()
354 		    node->prefixlen == key->prefixlen ||  in trie_update_elem()
355 		    node->prefixlen == trie->max_prefixlen)  in trie_update_elem()
358 		next_bit = extract_bit(key->data, node->prefixlen);  in trie_update_elem()
359 		slot = &node->child[next_bit];  in trie_update_elem()
362 	/* If the slot is empty (a free child pointer or an empty root),  in trie_update_elem()
365 	if (!node) {  in trie_update_elem()
373 	if (node->prefixlen == matchlen) {  in trie_update_elem()
374 		new_node->child[0] = node->child[0];  in trie_update_elem()
375 		new_node->child[1] = node->child[1];  in trie_update_elem()
377 		if (!(node->flags & LPM_TREE_NODE_FLAG_IM))  in trie_update_elem()
378 			trie->n_entries--;  in trie_update_elem()
381 		kfree_rcu(node, rcu);  in trie_update_elem()
386 	/* If the new node matches the prefix completely, it must be inserted  in trie_update_elem()
387 	 * as an ancestor. Simply insert it between @node and *@slot.  in trie_update_elem()
389 	if (matchlen == key->prefixlen) {  in trie_update_elem()
390 		next_bit = extract_bit(node->data, matchlen);  in trie_update_elem()
391 		rcu_assign_pointer(new_node->child[next_bit], node);  in trie_update_elem()
398 		ret = -ENOMEM;  in trie_update_elem()
402 	im_node->prefixlen = matchlen;  in trie_update_elem()
403 	im_node->flags |= LPM_TREE_NODE_FLAG_IM;  in trie_update_elem()
404 	memcpy(im_node->data, node->data, trie->data_size);  in trie_update_elem()
406 	/* Now determine which child to install in which slot */  in trie_update_elem()
407 	if (extract_bit(key->data, matchlen)) {  in trie_update_elem()
408 		rcu_assign_pointer(im_node->child[0], node);  in trie_update_elem()
409 		rcu_assign_pointer(im_node->child[1], new_node);  in trie_update_elem()
411 		rcu_assign_pointer(im_node->child[0], new_node);  in trie_update_elem()
412 		rcu_assign_pointer(im_node->child[1], node);  in trie_update_elem()
415 	/* Finally, assign the intermediate node to the determined spot */  in trie_update_elem()
421 			trie->n_entries--;  in trie_update_elem()
427 	spin_unlock_irqrestore(&trie->lock, irq_flags);  in trie_update_elem()
438 	struct lpm_trie_node *node, *parent;  in trie_delete_elem()  local
444 	if (key->prefixlen > trie->max_prefixlen)  in trie_delete_elem()
445 		return -EINVAL;  in trie_delete_elem()
447 	spin_lock_irqsave(&trie->lock, irq_flags);  in trie_delete_elem()
450 	 * track of the path we traverse.  We will need to know the node  in trie_delete_elem()
451 	 * we wish to delete, and the slot that points to the node we want  in trie_delete_elem()
455 	trim = &trie->root;  in trie_delete_elem()
458 	while ((node = rcu_dereference_protected(  in trie_delete_elem()
459 		       *trim, lockdep_is_held(&trie->lock)))) {  in trie_delete_elem()
460 		matchlen = longest_prefix_match(trie, node, key);  in trie_delete_elem()
462 		if (node->prefixlen != matchlen ||  in trie_delete_elem()
463 		    node->prefixlen == key->prefixlen)  in trie_delete_elem()
466 		parent = node;  in trie_delete_elem()
468 		next_bit = extract_bit(key->data, node->prefixlen);  in trie_delete_elem()
469 		trim = &node->child[next_bit];  in trie_delete_elem()
472 	if (!node || node->prefixlen != key->prefixlen ||  in trie_delete_elem()
473 	    node->prefixlen != matchlen ||  in trie_delete_elem()
474 	    (node->flags & LPM_TREE_NODE_FLAG_IM)) {  in trie_delete_elem()
475 		ret = -ENOENT;  in trie_delete_elem()
479 	trie->n_entries--;  in trie_delete_elem()
481 	/* If the node we are removing has two children, simply mark it  in trie_delete_elem()
484 	if (rcu_access_pointer(node->child[0]) &&  in trie_delete_elem()
485 	    rcu_access_pointer(node->child[1])) {  in trie_delete_elem()
486 		node->flags |= LPM_TREE_NODE_FLAG_IM;  in trie_delete_elem()
490 	/* If the parent of the node we are about to delete is an intermediate  in trie_delete_elem()
491 	 * node, and the deleted node doesn't have any children, we can delete  in trie_delete_elem()
492 	 * the intermediate parent as well and promote its other child  in trie_delete_elem()
497 	if (parent && (parent->flags & LPM_TREE_NODE_FLAG_IM) &&  in trie_delete_elem()
498 	    !node->child[0] && !node->child[1]) {  in trie_delete_elem()
499 		if (node == rcu_access_pointer(parent->child[0]))  in trie_delete_elem()
501 				*trim2, rcu_access_pointer(parent->child[1]));  in trie_delete_elem()
504 				*trim2, rcu_access_pointer(parent->child[0]));  in trie_delete_elem()
506 		kfree_rcu(node, rcu);  in trie_delete_elem()
510 	/* The node we are removing has either zero or one child. If there  in trie_delete_elem()
511 	 * is a child, move it into the removed node's slot then delete  in trie_delete_elem()
512 	 * the node.  Otherwise just clear the slot and delete the node.  in trie_delete_elem()
514 	if (node->child[0])  in trie_delete_elem()
515 		rcu_assign_pointer(*trim, rcu_access_pointer(node->child[0]));  in trie_delete_elem()
516 	else if (node->child[1])  in trie_delete_elem()
517 		rcu_assign_pointer(*trim, rcu_access_pointer(node->child[1]));  in trie_delete_elem()
520 	kfree_rcu(node, rcu);  in trie_delete_elem()
523 	spin_unlock_irqrestore(&trie->lock, irq_flags);  in trie_delete_elem()
531 #define LPM_VAL_SIZE_MAX	(KMALLOC_MAX_SIZE - LPM_DATA_SIZE_MAX - \
547 		return ERR_PTR(-EPERM);  in trie_alloc()
550 	if (attr->max_entries == 0 ||  in trie_alloc()
551 	    !(attr->map_flags & BPF_F_NO_PREALLOC) ||  in trie_alloc()
552 	    attr->map_flags & ~LPM_CREATE_FLAG_MASK ||  in trie_alloc()
553 	    !bpf_map_flags_access_ok(attr->map_flags) ||  in trie_alloc()
554 	    attr->key_size < LPM_KEY_SIZE_MIN ||  in trie_alloc()
555 	    attr->key_size > LPM_KEY_SIZE_MAX ||  in trie_alloc()
556 	    attr->value_size < LPM_VAL_SIZE_MIN ||  in trie_alloc()
557 	    attr->value_size > LPM_VAL_SIZE_MAX)  in trie_alloc()
558 		return ERR_PTR(-EINVAL);  in trie_alloc()
562 		return ERR_PTR(-ENOMEM);  in trie_alloc()
565 	bpf_map_init_from_attr(&trie->map, attr);  in trie_alloc()
566 	trie->data_size = attr->key_size -  in trie_alloc()
568 	trie->max_prefixlen = trie->data_size * 8;  in trie_alloc()
570 	spin_lock_init(&trie->lock);  in trie_alloc()
572 	return &trie->map;  in trie_alloc()
579 	struct lpm_trie_node *node;  in trie_free()  local
581 	/* Always start at the root and walk down to a node that has no  in trie_free()
582 	 * children. Then free that node, nullify its reference in the parent  in trie_free()
587 		slot = &trie->root;  in trie_free()
590 			node = rcu_dereference_protected(*slot, 1);  in trie_free()
591 			if (!node)  in trie_free()
594 			if (rcu_access_pointer(node->child[0])) {  in trie_free()
595 				slot = &node->child[0];  in trie_free()
599 			if (rcu_access_pointer(node->child[1])) {  in trie_free()
600 				slot = &node->child[1];  in trie_free()
604 			kfree(node);  in trie_free()
616 	struct lpm_trie_node *node, *next_node = NULL, *parent, *search_root;  in trie_get_next_key()  local
620 	int err = 0, stack_ptr = -1;  in trie_get_next_key()
624 	/* The get_next_key follows postorder. For the 4 node example in  in trie_get_next_key()
636 	search_root = rcu_dereference(trie->root);  in trie_get_next_key()
638 		return -ENOENT;  in trie_get_next_key()
640 	/* For invalid key, find the leftmost node in the trie */  in trie_get_next_key()
641 	if (!key || key->prefixlen > trie->max_prefixlen)  in trie_get_next_key()
644 	node_stack = kmalloc_array(trie->max_prefixlen,  in trie_get_next_key()
648 		return -ENOMEM;  in trie_get_next_key()
650 	/* Try to find the exact node for the given key */  in trie_get_next_key()
651 	for (node = search_root; node;) {  in trie_get_next_key()
652 		node_stack[++stack_ptr] = node;  in trie_get_next_key()
653 		matchlen = longest_prefix_match(trie, node, key);  in trie_get_next_key()
654 		if (node->prefixlen != matchlen ||  in trie_get_next_key()
655 		    node->prefixlen == key->prefixlen)  in trie_get_next_key()
658 		next_bit = extract_bit(key->data, node->prefixlen);  in trie_get_next_key()
659 		node = rcu_dereference(node->child[next_bit]);  in trie_get_next_key()
661 	if (!node || node->prefixlen != key->prefixlen ||  in trie_get_next_key()
662 	    (node->flags & LPM_TREE_NODE_FLAG_IM))  in trie_get_next_key()
665 	/* The node with the exactly-matching key has been found,  in trie_get_next_key()
666 	 * find the first node in postorder after the matched node.  in trie_get_next_key()
668 	node = node_stack[stack_ptr];  in trie_get_next_key()
670 		parent = node_stack[stack_ptr - 1];  in trie_get_next_key()
671 		if (rcu_dereference(parent->child[0]) == node) {  in trie_get_next_key()
672 			search_root = rcu_dereference(parent->child[1]);  in trie_get_next_key()
676 		if (!(parent->flags & LPM_TREE_NODE_FLAG_IM)) {  in trie_get_next_key()
681 		node = parent;  in trie_get_next_key()
682 		stack_ptr--;  in trie_get_next_key()
686 	err = -ENOENT;  in trie_get_next_key()
690 	/* Find the leftmost non-intermediate node, all intermediate nodes  in trie_get_next_key()
693 	for (node = search_root; node;) {  in trie_get_next_key()
694 		if (node->flags & LPM_TREE_NODE_FLAG_IM) {  in trie_get_next_key()
695 			node = rcu_dereference(node->child[0]);  in trie_get_next_key()
697 			next_node = node;  in trie_get_next_key()
698 			node = rcu_dereference(node->child[0]);  in trie_get_next_key()
699 			if (!node)  in trie_get_next_key()
700 				node = rcu_dereference(next_node->child[1]);  in trie_get_next_key()
704 	next_key->prefixlen = next_node->prefixlen;  in trie_get_next_key()
706 	       next_node->data, trie->data_size);  in trie_get_next_key()
718 	return BTF_INFO_KIND(key_type->info) != BTF_KIND_STRUCT ?  in trie_check_btf()
719 	       -EINVAL : 0;  in trie_check_btf()