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>
18 /* Intermediate node */
19 #define LPM_TREE_NODE_FLAG_IM BIT(0)
23 struct lpm_trie_node
{
25 struct lpm_trie_node __rcu
*child
[2];
33 struct lpm_trie_node __rcu
*root
;
40 /* This trie implements a longest prefix match algorithm that can be used to
41 * match IP addresses to a stored set of ranges.
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.
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.
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.
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.
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
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,
100 * +----------------+ +------------------+
102 * | 192.168.0.0/24 | | 192.168.128.0/24 |
103 * | value: 2 | | value: 3 |
104 * | [0] [1] | | [0] [1] |
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
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:
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 * +----------------+ +------------------+
130 * +----------------+ +----------------+
132 * | 192.168.0.0/24 | | 192.168.1.0/24 |
133 * | value: 2 | | value: 5 |
134 * | [0] [1] | | [0] [1] |
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.
142 * A fully populated trie would have a height of 32 nodes, as the trie was
143 * created with a prefix length of 32.
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
151 static inline int extract_bit(const u8
*data
, size_t index
)
153 return !!(data
[index
/ 8] & (1 << (7 - (index
% 8))));
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
162 * Determine the longest prefix of @node that matches the bits in @key.
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
)
168 u32 limit
= min(node
->prefixlen
, key
->prefixlen
);
169 u32 prefixlen
= 0, i
= 0;
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
));
174 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && defined(CONFIG_64BIT)
176 /* data_size >= 16 has very small probability.
177 * We do not use a loop for optimal code generation.
179 if (trie
->data_size
>= 8) {
180 u64 diff
= be64_to_cpu(*(__be64
*)node
->data
^
181 *(__be64
*)key
->data
);
183 prefixlen
= 64 - fls64(diff
);
184 if (prefixlen
>= limit
)
192 while (trie
->data_size
>= i
+ 4) {
193 u32 diff
= be32_to_cpu(*(__be32
*)&node
->data
[i
] ^
194 *(__be32
*)&key
->data
[i
]);
196 prefixlen
+= 32 - fls(diff
);
197 if (prefixlen
>= limit
)
204 if (trie
->data_size
>= i
+ 2) {
205 u16 diff
= be16_to_cpu(*(__be16
*)&node
->data
[i
] ^
206 *(__be16
*)&key
->data
[i
]);
208 prefixlen
+= 16 - fls(diff
);
209 if (prefixlen
>= limit
)
216 if (trie
->data_size
>= i
+ 1) {
217 prefixlen
+= 8 - fls(node
->data
[i
] ^ key
->data
[i
]);
219 if (prefixlen
>= limit
)
226 /* Called from syscall or from eBPF program */
227 static void *trie_lookup_elem(struct bpf_map
*map
, void *_key
)
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
;
233 /* Start walking the trie from the root node ... */
235 for (node
= rcu_dereference(trie
->root
); node
;) {
236 unsigned int next_bit
;
239 /* Determine the longest prefix of @node that matches @key.
240 * If it's the maximum possible prefix for this trie, we have
241 * an exact match and can return it directly.
243 matchlen
= longest_prefix_match(trie
, node
, key
);
244 if (matchlen
== trie
->max_prefixlen
) {
249 /* If the number of bits that match is smaller than the prefix
250 * length of @node, bail out and return the node we have seen
251 * last in the traversal (ie, the parent).
253 if (matchlen
< node
->prefixlen
)
256 /* Consider this node as return candidate unless it is an
257 * artificially added intermediate one.
259 if (!(node
->flags
& LPM_TREE_NODE_FLAG_IM
))
262 /* If the node match is fully satisfied, let's see if we can
263 * become more specific. Determine the next bit in the key and
266 next_bit
= extract_bit(key
->data
, node
->prefixlen
);
267 node
= rcu_dereference(node
->child
[next_bit
]);
273 return found
->data
+ trie
->data_size
;
276 static struct lpm_trie_node
*lpm_trie_node_alloc(const struct lpm_trie
*trie
,
279 struct lpm_trie_node
*node
;
280 size_t size
= sizeof(struct lpm_trie_node
) + trie
->data_size
;
283 size
+= trie
->map
.value_size
;
285 node
= kmalloc_node(size
, GFP_ATOMIC
| __GFP_NOWARN
,
286 trie
->map
.numa_node
);
293 memcpy(node
->data
+ trie
->data_size
, value
,
294 trie
->map
.value_size
);
299 /* Called from syscall or from eBPF program */
300 static int trie_update_elem(struct bpf_map
*map
,
301 void *_key
, void *value
, u64 flags
)
303 struct lpm_trie
*trie
= container_of(map
, struct lpm_trie
, map
);
304 struct lpm_trie_node
*node
, *im_node
= NULL
, *new_node
= NULL
;
305 struct lpm_trie_node __rcu
**slot
;
306 struct bpf_lpm_trie_key
*key
= _key
;
307 unsigned long irq_flags
;
308 unsigned int next_bit
;
312 if (unlikely(flags
> BPF_EXIST
))
315 if (key
->prefixlen
> trie
->max_prefixlen
)
318 raw_spin_lock_irqsave(&trie
->lock
, irq_flags
);
320 /* Allocate and fill a new node */
322 if (trie
->n_entries
== trie
->map
.max_entries
) {
327 new_node
= lpm_trie_node_alloc(trie
, value
);
335 new_node
->prefixlen
= key
->prefixlen
;
336 RCU_INIT_POINTER(new_node
->child
[0], NULL
);
337 RCU_INIT_POINTER(new_node
->child
[1], NULL
);
338 memcpy(new_node
->data
, key
->data
, trie
->data_size
);
340 /* Now find a slot to attach the new node. To do that, walk the tree
341 * from the root and match as many bits as possible for each node until
342 * we either find an empty slot or a slot that needs to be replaced by
343 * an intermediate node.
347 while ((node
= rcu_dereference_protected(*slot
,
348 lockdep_is_held(&trie
->lock
)))) {
349 matchlen
= longest_prefix_match(trie
, node
, key
);
351 if (node
->prefixlen
!= matchlen
||
352 node
->prefixlen
== key
->prefixlen
||
353 node
->prefixlen
== trie
->max_prefixlen
)
356 next_bit
= extract_bit(key
->data
, node
->prefixlen
);
357 slot
= &node
->child
[next_bit
];
360 /* If the slot is empty (a free child pointer or an empty root),
361 * simply assign the @new_node to that slot and be done.
364 rcu_assign_pointer(*slot
, new_node
);
368 /* If the slot we picked already exists, replace it with @new_node
369 * which already has the correct data array set.
371 if (node
->prefixlen
== matchlen
) {
372 new_node
->child
[0] = node
->child
[0];
373 new_node
->child
[1] = node
->child
[1];
375 if (!(node
->flags
& LPM_TREE_NODE_FLAG_IM
))
378 rcu_assign_pointer(*slot
, new_node
);
379 kfree_rcu(node
, rcu
);
384 /* If the new node matches the prefix completely, it must be inserted
385 * as an ancestor. Simply insert it between @node and *@slot.
387 if (matchlen
== key
->prefixlen
) {
388 next_bit
= extract_bit(node
->data
, matchlen
);
389 rcu_assign_pointer(new_node
->child
[next_bit
], node
);
390 rcu_assign_pointer(*slot
, new_node
);
394 im_node
= lpm_trie_node_alloc(trie
, NULL
);
400 im_node
->prefixlen
= matchlen
;
401 im_node
->flags
|= LPM_TREE_NODE_FLAG_IM
;
402 memcpy(im_node
->data
, node
->data
, trie
->data_size
);
404 /* Now determine which child to install in which slot */
405 if (extract_bit(key
->data
, matchlen
)) {
406 rcu_assign_pointer(im_node
->child
[0], node
);
407 rcu_assign_pointer(im_node
->child
[1], new_node
);
409 rcu_assign_pointer(im_node
->child
[0], new_node
);
410 rcu_assign_pointer(im_node
->child
[1], node
);
413 /* Finally, assign the intermediate node to the determined spot */
414 rcu_assign_pointer(*slot
, im_node
);
425 raw_spin_unlock_irqrestore(&trie
->lock
, irq_flags
);
430 /* Called from syscall or from eBPF program */
431 static int trie_delete_elem(struct bpf_map
*map
, void *_key
)
433 struct lpm_trie
*trie
= container_of(map
, struct lpm_trie
, map
);
434 struct bpf_lpm_trie_key
*key
= _key
;
435 struct lpm_trie_node __rcu
**trim
, **trim2
;
436 struct lpm_trie_node
*node
, *parent
;
437 unsigned long irq_flags
;
438 unsigned int next_bit
;
442 if (key
->prefixlen
> trie
->max_prefixlen
)
445 raw_spin_lock_irqsave(&trie
->lock
, irq_flags
);
447 /* Walk the tree looking for an exact key/length match and keeping
448 * track of the path we traverse. We will need to know the node
449 * we wish to delete, and the slot that points to the node we want
450 * to delete. We may also need to know the nodes parent and the
451 * slot that contains it.
456 while ((node
= rcu_dereference_protected(
457 *trim
, lockdep_is_held(&trie
->lock
)))) {
458 matchlen
= longest_prefix_match(trie
, node
, key
);
460 if (node
->prefixlen
!= matchlen
||
461 node
->prefixlen
== key
->prefixlen
)
466 next_bit
= extract_bit(key
->data
, node
->prefixlen
);
467 trim
= &node
->child
[next_bit
];
470 if (!node
|| node
->prefixlen
!= key
->prefixlen
||
471 node
->prefixlen
!= matchlen
||
472 (node
->flags
& LPM_TREE_NODE_FLAG_IM
)) {
479 /* If the node we are removing has two children, simply mark it
480 * as intermediate and we are done.
482 if (rcu_access_pointer(node
->child
[0]) &&
483 rcu_access_pointer(node
->child
[1])) {
484 node
->flags
|= LPM_TREE_NODE_FLAG_IM
;
488 /* If the parent of the node we are about to delete is an intermediate
489 * node, and the deleted node doesn't have any children, we can delete
490 * the intermediate parent as well and promote its other child
491 * up the tree. Doing this maintains the invariant that all
492 * intermediate nodes have exactly 2 children and that there are no
493 * unnecessary intermediate nodes in the tree.
495 if (parent
&& (parent
->flags
& LPM_TREE_NODE_FLAG_IM
) &&
496 !node
->child
[0] && !node
->child
[1]) {
497 if (node
== rcu_access_pointer(parent
->child
[0]))
499 *trim2
, rcu_access_pointer(parent
->child
[1]));
502 *trim2
, rcu_access_pointer(parent
->child
[0]));
503 kfree_rcu(parent
, rcu
);
504 kfree_rcu(node
, rcu
);
508 /* The node we are removing has either zero or one child. If there
509 * is a child, move it into the removed node's slot then delete
510 * the node. Otherwise just clear the slot and delete the node.
513 rcu_assign_pointer(*trim
, rcu_access_pointer(node
->child
[0]));
514 else if (node
->child
[1])
515 rcu_assign_pointer(*trim
, rcu_access_pointer(node
->child
[1]));
517 RCU_INIT_POINTER(*trim
, NULL
);
518 kfree_rcu(node
, rcu
);
521 raw_spin_unlock_irqrestore(&trie
->lock
, irq_flags
);
526 #define LPM_DATA_SIZE_MAX 256
527 #define LPM_DATA_SIZE_MIN 1
529 #define LPM_VAL_SIZE_MAX (KMALLOC_MAX_SIZE - LPM_DATA_SIZE_MAX - \
530 sizeof(struct lpm_trie_node))
531 #define LPM_VAL_SIZE_MIN 1
533 #define LPM_KEY_SIZE(X) (sizeof(struct bpf_lpm_trie_key) + (X))
534 #define LPM_KEY_SIZE_MAX LPM_KEY_SIZE(LPM_DATA_SIZE_MAX)
535 #define LPM_KEY_SIZE_MIN LPM_KEY_SIZE(LPM_DATA_SIZE_MIN)
537 #define LPM_CREATE_FLAG_MASK (BPF_F_NO_PREALLOC | BPF_F_NUMA_NODE | \
540 static struct bpf_map
*trie_alloc(union bpf_attr
*attr
)
542 struct lpm_trie
*trie
;
543 u64 cost
= sizeof(*trie
), cost_per_node
;
546 if (!capable(CAP_SYS_ADMIN
))
547 return ERR_PTR(-EPERM
);
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
);
560 trie
= kzalloc(sizeof(*trie
), GFP_USER
| __GFP_NOWARN
);
562 return ERR_PTR(-ENOMEM
);
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;
570 cost_per_node
= sizeof(struct lpm_trie_node
) +
571 attr
->value_size
+ trie
->data_size
;
572 cost
+= (u64
) attr
->max_entries
* cost_per_node
;
574 ret
= bpf_map_charge_init(&trie
->map
.memory
, cost
);
578 raw_spin_lock_init(&trie
->lock
);
586 static void trie_free(struct bpf_map
*map
)
588 struct lpm_trie
*trie
= container_of(map
, struct lpm_trie
, map
);
589 struct lpm_trie_node __rcu
**slot
;
590 struct lpm_trie_node
*node
;
592 /* Wait for outstanding programs to complete
593 * update/lookup/delete/get_next_key and free the trie.
597 /* Always start at the root and walk down to a node that has no
598 * children. Then free that node, nullify its reference in the parent
606 node
= rcu_dereference_protected(*slot
, 1);
610 if (rcu_access_pointer(node
->child
[0])) {
611 slot
= &node
->child
[0];
615 if (rcu_access_pointer(node
->child
[1])) {
616 slot
= &node
->child
[1];
621 RCU_INIT_POINTER(*slot
, NULL
);
630 static int trie_get_next_key(struct bpf_map
*map
, void *_key
, void *_next_key
)
632 struct lpm_trie_node
*node
, *next_node
= NULL
, *parent
, *search_root
;
633 struct lpm_trie
*trie
= container_of(map
, struct lpm_trie
, map
);
634 struct bpf_lpm_trie_key
*key
= _key
, *next_key
= _next_key
;
635 struct lpm_trie_node
**node_stack
= NULL
;
636 int err
= 0, stack_ptr
= -1;
637 unsigned int next_bit
;
640 /* The get_next_key follows postorder. For the 4 node example in
641 * the top of this file, the trie_get_next_key() returns the following
648 * The idea is to return more specific keys before less specific ones.
652 search_root
= rcu_dereference(trie
->root
);
656 /* For invalid key, find the leftmost node in the trie */
657 if (!key
|| key
->prefixlen
> trie
->max_prefixlen
)
660 node_stack
= kmalloc_array(trie
->max_prefixlen
,
661 sizeof(struct lpm_trie_node
*),
662 GFP_ATOMIC
| __GFP_NOWARN
);
666 /* Try to find the exact node for the given key */
667 for (node
= search_root
; node
;) {
668 node_stack
[++stack_ptr
] = node
;
669 matchlen
= longest_prefix_match(trie
, node
, key
);
670 if (node
->prefixlen
!= matchlen
||
671 node
->prefixlen
== key
->prefixlen
)
674 next_bit
= extract_bit(key
->data
, node
->prefixlen
);
675 node
= rcu_dereference(node
->child
[next_bit
]);
677 if (!node
|| node
->prefixlen
!= key
->prefixlen
||
678 (node
->flags
& LPM_TREE_NODE_FLAG_IM
))
681 /* The node with the exactly-matching key has been found,
682 * find the first node in postorder after the matched node.
684 node
= node_stack
[stack_ptr
];
685 while (stack_ptr
> 0) {
686 parent
= node_stack
[stack_ptr
- 1];
687 if (rcu_dereference(parent
->child
[0]) == node
) {
688 search_root
= rcu_dereference(parent
->child
[1]);
692 if (!(parent
->flags
& LPM_TREE_NODE_FLAG_IM
)) {
701 /* did not find anything */
706 /* Find the leftmost non-intermediate node, all intermediate nodes
707 * have exact two children, so this function will never return NULL.
709 for (node
= search_root
; node
;) {
710 if (node
->flags
& LPM_TREE_NODE_FLAG_IM
) {
711 node
= rcu_dereference(node
->child
[0]);
714 node
= rcu_dereference(node
->child
[0]);
716 node
= rcu_dereference(next_node
->child
[1]);
720 next_key
->prefixlen
= next_node
->prefixlen
;
721 memcpy((void *)next_key
+ offsetof(struct bpf_lpm_trie_key
, data
),
722 next_node
->data
, trie
->data_size
);
728 static int trie_check_btf(const struct bpf_map
*map
,
729 const struct btf
*btf
,
730 const struct btf_type
*key_type
,
731 const struct btf_type
*value_type
)
733 /* Keys must have struct bpf_lpm_trie_key embedded. */
734 return BTF_INFO_KIND(key_type
->info
) != BTF_KIND_STRUCT
?
738 const struct bpf_map_ops trie_map_ops
= {
739 .map_alloc
= trie_alloc
,
740 .map_free
= trie_free
,
741 .map_get_next_key
= trie_get_next_key
,
742 .map_lookup_elem
= trie_lookup_elem
,
743 .map_update_elem
= trie_update_elem
,
744 .map_delete_elem
= trie_delete_elem
,
745 .map_check_btf
= trie_check_btf
,