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/err.h>
14 #include <linux/slab.h>
15 #include <linux/spinlock.h>
16 #include <linux/vmalloc.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 size_t prefixlen
= 0;
172 for (i
= 0; i
< trie
->data_size
; i
++) {
175 b
= 8 - fls(node
->data
[i
] ^ key
->data
[i
]);
178 if (prefixlen
>= node
->prefixlen
|| prefixlen
>= key
->prefixlen
)
179 return min(node
->prefixlen
, key
->prefixlen
);
188 /* Called from syscall or from eBPF program */
189 static void *trie_lookup_elem(struct bpf_map
*map
, void *_key
)
191 struct lpm_trie
*trie
= container_of(map
, struct lpm_trie
, map
);
192 struct lpm_trie_node
*node
, *found
= NULL
;
193 struct bpf_lpm_trie_key
*key
= _key
;
195 /* Start walking the trie from the root node ... */
197 for (node
= rcu_dereference(trie
->root
); node
;) {
198 unsigned int next_bit
;
201 /* Determine the longest prefix of @node that matches @key.
202 * If it's the maximum possible prefix for this trie, we have
203 * an exact match and can return it directly.
205 matchlen
= longest_prefix_match(trie
, node
, key
);
206 if (matchlen
== trie
->max_prefixlen
) {
211 /* If the number of bits that match is smaller than the prefix
212 * length of @node, bail out and return the node we have seen
213 * last in the traversal (ie, the parent).
215 if (matchlen
< node
->prefixlen
)
218 /* Consider this node as return candidate unless it is an
219 * artificially added intermediate one.
221 if (!(node
->flags
& LPM_TREE_NODE_FLAG_IM
))
224 /* If the node match is fully satisfied, let's see if we can
225 * become more specific. Determine the next bit in the key and
228 next_bit
= extract_bit(key
->data
, node
->prefixlen
);
229 node
= rcu_dereference(node
->child
[next_bit
]);
235 return found
->data
+ trie
->data_size
;
238 static struct lpm_trie_node
*lpm_trie_node_alloc(const struct lpm_trie
*trie
,
241 struct lpm_trie_node
*node
;
242 size_t size
= sizeof(struct lpm_trie_node
) + trie
->data_size
;
245 size
+= trie
->map
.value_size
;
247 node
= kmalloc_node(size
, GFP_ATOMIC
| __GFP_NOWARN
,
248 trie
->map
.numa_node
);
255 memcpy(node
->data
+ trie
->data_size
, value
,
256 trie
->map
.value_size
);
261 /* Called from syscall or from eBPF program */
262 static int trie_update_elem(struct bpf_map
*map
,
263 void *_key
, void *value
, u64 flags
)
265 struct lpm_trie
*trie
= container_of(map
, struct lpm_trie
, map
);
266 struct lpm_trie_node
*node
, *im_node
= NULL
, *new_node
= NULL
;
267 struct lpm_trie_node __rcu
**slot
;
268 struct bpf_lpm_trie_key
*key
= _key
;
269 unsigned long irq_flags
;
270 unsigned int next_bit
;
274 if (unlikely(flags
> BPF_EXIST
))
277 if (key
->prefixlen
> trie
->max_prefixlen
)
280 raw_spin_lock_irqsave(&trie
->lock
, irq_flags
);
282 /* Allocate and fill a new node */
284 if (trie
->n_entries
== trie
->map
.max_entries
) {
289 new_node
= lpm_trie_node_alloc(trie
, value
);
297 new_node
->prefixlen
= key
->prefixlen
;
298 RCU_INIT_POINTER(new_node
->child
[0], NULL
);
299 RCU_INIT_POINTER(new_node
->child
[1], NULL
);
300 memcpy(new_node
->data
, key
->data
, trie
->data_size
);
302 /* Now find a slot to attach the new node. To do that, walk the tree
303 * from the root and match as many bits as possible for each node until
304 * we either find an empty slot or a slot that needs to be replaced by
305 * an intermediate node.
309 while ((node
= rcu_dereference_protected(*slot
,
310 lockdep_is_held(&trie
->lock
)))) {
311 matchlen
= longest_prefix_match(trie
, node
, key
);
313 if (node
->prefixlen
!= matchlen
||
314 node
->prefixlen
== key
->prefixlen
||
315 node
->prefixlen
== trie
->max_prefixlen
)
318 next_bit
= extract_bit(key
->data
, node
->prefixlen
);
319 slot
= &node
->child
[next_bit
];
322 /* If the slot is empty (a free child pointer or an empty root),
323 * simply assign the @new_node to that slot and be done.
326 rcu_assign_pointer(*slot
, new_node
);
330 /* If the slot we picked already exists, replace it with @new_node
331 * which already has the correct data array set.
333 if (node
->prefixlen
== matchlen
) {
334 new_node
->child
[0] = node
->child
[0];
335 new_node
->child
[1] = node
->child
[1];
337 if (!(node
->flags
& LPM_TREE_NODE_FLAG_IM
))
340 rcu_assign_pointer(*slot
, new_node
);
341 kfree_rcu(node
, rcu
);
346 /* If the new node matches the prefix completely, it must be inserted
347 * as an ancestor. Simply insert it between @node and *@slot.
349 if (matchlen
== key
->prefixlen
) {
350 next_bit
= extract_bit(node
->data
, matchlen
);
351 rcu_assign_pointer(new_node
->child
[next_bit
], node
);
352 rcu_assign_pointer(*slot
, new_node
);
356 im_node
= lpm_trie_node_alloc(trie
, NULL
);
362 im_node
->prefixlen
= matchlen
;
363 im_node
->flags
|= LPM_TREE_NODE_FLAG_IM
;
364 memcpy(im_node
->data
, node
->data
, trie
->data_size
);
366 /* Now determine which child to install in which slot */
367 if (extract_bit(key
->data
, matchlen
)) {
368 rcu_assign_pointer(im_node
->child
[0], node
);
369 rcu_assign_pointer(im_node
->child
[1], new_node
);
371 rcu_assign_pointer(im_node
->child
[0], new_node
);
372 rcu_assign_pointer(im_node
->child
[1], node
);
375 /* Finally, assign the intermediate node to the determined spot */
376 rcu_assign_pointer(*slot
, im_node
);
387 raw_spin_unlock_irqrestore(&trie
->lock
, irq_flags
);
392 /* Called from syscall or from eBPF program */
393 static int trie_delete_elem(struct bpf_map
*map
, void *_key
)
395 struct lpm_trie
*trie
= container_of(map
, struct lpm_trie
, map
);
396 struct bpf_lpm_trie_key
*key
= _key
;
397 struct lpm_trie_node __rcu
**trim
, **trim2
;
398 struct lpm_trie_node
*node
, *parent
;
399 unsigned long irq_flags
;
400 unsigned int next_bit
;
404 if (key
->prefixlen
> trie
->max_prefixlen
)
407 raw_spin_lock_irqsave(&trie
->lock
, irq_flags
);
409 /* Walk the tree looking for an exact key/length match and keeping
410 * track of the path we traverse. We will need to know the node
411 * we wish to delete, and the slot that points to the node we want
412 * to delete. We may also need to know the nodes parent and the
413 * slot that contains it.
418 while ((node
= rcu_dereference_protected(
419 *trim
, lockdep_is_held(&trie
->lock
)))) {
420 matchlen
= longest_prefix_match(trie
, node
, key
);
422 if (node
->prefixlen
!= matchlen
||
423 node
->prefixlen
== key
->prefixlen
)
428 next_bit
= extract_bit(key
->data
, node
->prefixlen
);
429 trim
= &node
->child
[next_bit
];
432 if (!node
|| node
->prefixlen
!= key
->prefixlen
||
433 (node
->flags
& LPM_TREE_NODE_FLAG_IM
)) {
440 /* If the node we are removing has two children, simply mark it
441 * as intermediate and we are done.
443 if (rcu_access_pointer(node
->child
[0]) &&
444 rcu_access_pointer(node
->child
[1])) {
445 node
->flags
|= LPM_TREE_NODE_FLAG_IM
;
449 /* If the parent of the node we are about to delete is an intermediate
450 * node, and the deleted node doesn't have any children, we can delete
451 * the intermediate parent as well and promote its other child
452 * up the tree. Doing this maintains the invariant that all
453 * intermediate nodes have exactly 2 children and that there are no
454 * unnecessary intermediate nodes in the tree.
456 if (parent
&& (parent
->flags
& LPM_TREE_NODE_FLAG_IM
) &&
457 !node
->child
[0] && !node
->child
[1]) {
458 if (node
== rcu_access_pointer(parent
->child
[0]))
460 *trim2
, rcu_access_pointer(parent
->child
[1]));
463 *trim2
, rcu_access_pointer(parent
->child
[0]));
464 kfree_rcu(parent
, rcu
);
465 kfree_rcu(node
, rcu
);
469 /* The node we are removing has either zero or one child. If there
470 * is a child, move it into the removed node's slot then delete
471 * the node. Otherwise just clear the slot and delete the node.
474 rcu_assign_pointer(*trim
, rcu_access_pointer(node
->child
[0]));
475 else if (node
->child
[1])
476 rcu_assign_pointer(*trim
, rcu_access_pointer(node
->child
[1]));
478 RCU_INIT_POINTER(*trim
, NULL
);
479 kfree_rcu(node
, rcu
);
482 raw_spin_unlock_irqrestore(&trie
->lock
, irq_flags
);
487 #define LPM_DATA_SIZE_MAX 256
488 #define LPM_DATA_SIZE_MIN 1
490 #define LPM_VAL_SIZE_MAX (KMALLOC_MAX_SIZE - LPM_DATA_SIZE_MAX - \
491 sizeof(struct lpm_trie_node))
492 #define LPM_VAL_SIZE_MIN 1
494 #define LPM_KEY_SIZE(X) (sizeof(struct bpf_lpm_trie_key) + (X))
495 #define LPM_KEY_SIZE_MAX LPM_KEY_SIZE(LPM_DATA_SIZE_MAX)
496 #define LPM_KEY_SIZE_MIN LPM_KEY_SIZE(LPM_DATA_SIZE_MIN)
498 #define LPM_CREATE_FLAG_MASK (BPF_F_NO_PREALLOC | BPF_F_NUMA_NODE | \
499 BPF_F_RDONLY | BPF_F_WRONLY)
501 static struct bpf_map
*trie_alloc(union bpf_attr
*attr
)
503 struct lpm_trie
*trie
;
504 u64 cost
= sizeof(*trie
), cost_per_node
;
507 if (!capable(CAP_SYS_ADMIN
))
508 return ERR_PTR(-EPERM
);
510 /* check sanity of attributes */
511 if (attr
->max_entries
== 0 ||
512 !(attr
->map_flags
& BPF_F_NO_PREALLOC
) ||
513 attr
->map_flags
& ~LPM_CREATE_FLAG_MASK
||
514 attr
->key_size
< LPM_KEY_SIZE_MIN
||
515 attr
->key_size
> LPM_KEY_SIZE_MAX
||
516 attr
->value_size
< LPM_VAL_SIZE_MIN
||
517 attr
->value_size
> LPM_VAL_SIZE_MAX
)
518 return ERR_PTR(-EINVAL
);
520 trie
= kzalloc(sizeof(*trie
), GFP_USER
| __GFP_NOWARN
);
522 return ERR_PTR(-ENOMEM
);
524 /* copy mandatory map attributes */
525 trie
->map
.map_type
= attr
->map_type
;
526 trie
->map
.key_size
= attr
->key_size
;
527 trie
->map
.value_size
= attr
->value_size
;
528 trie
->map
.max_entries
= attr
->max_entries
;
529 trie
->map
.map_flags
= attr
->map_flags
;
530 trie
->map
.numa_node
= bpf_map_attr_numa_node(attr
);
531 trie
->data_size
= attr
->key_size
-
532 offsetof(struct bpf_lpm_trie_key
, data
);
533 trie
->max_prefixlen
= trie
->data_size
* 8;
535 cost_per_node
= sizeof(struct lpm_trie_node
) +
536 attr
->value_size
+ trie
->data_size
;
537 cost
+= (u64
) attr
->max_entries
* cost_per_node
;
538 if (cost
>= U32_MAX
- PAGE_SIZE
) {
543 trie
->map
.pages
= round_up(cost
, PAGE_SIZE
) >> PAGE_SHIFT
;
545 ret
= bpf_map_precharge_memlock(trie
->map
.pages
);
549 raw_spin_lock_init(&trie
->lock
);
557 static void trie_free(struct bpf_map
*map
)
559 struct lpm_trie
*trie
= container_of(map
, struct lpm_trie
, map
);
560 struct lpm_trie_node __rcu
**slot
;
561 struct lpm_trie_node
*node
;
563 raw_spin_lock(&trie
->lock
);
565 /* Always start at the root and walk down to a node that has no
566 * children. Then free that node, nullify its reference in the parent
574 node
= rcu_dereference_protected(*slot
,
575 lockdep_is_held(&trie
->lock
));
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
);
596 raw_spin_unlock(&trie
->lock
);
599 static int trie_get_next_key(struct bpf_map
*map
, void *key
, void *next_key
)
604 const struct bpf_map_ops trie_map_ops
= {
605 .map_alloc
= trie_alloc
,
606 .map_free
= trie_free
,
607 .map_get_next_key
= trie_get_next_key
,
608 .map_lookup_elem
= trie_lookup_elem
,
609 .map_update_elem
= trie_update_elem
,
610 .map_delete_elem
= trie_delete_elem
,