1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Longest prefix match list implementation
4 *
5 * Copyright (c) 2016,2017 Daniel Mack
6 * Copyright (c) 2016 David Herrmann
7 */
8
9 #include <linux/bpf.h>
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>
15 #include <net/ipv6.h>
16 #include <uapi/linux/btf.h>
17 #include <linux/btf_ids.h>
18
19 /* Intermediate node */
20 #define LPM_TREE_NODE_FLAG_IM BIT(0)
21
22 struct lpm_trie_node;
23
24 struct lpm_trie_node {
25 struct rcu_head rcu;
26 struct lpm_trie_node __rcu *child[2];
27 u32 prefixlen;
28 u32 flags;
29 u8 data[];
30 };
31
32 struct lpm_trie {
33 struct bpf_map map;
34 struct lpm_trie_node __rcu *root;
35 size_t n_entries;
36 size_t max_prefixlen;
37 size_t data_size;
38 spinlock_t lock;
39 };
40
41 /* This trie implements a longest prefix match algorithm that can be used to
42 * match IP addresses to a stored set of ranges.
43 *
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.
46 *
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.
52 *
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.
58 *
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.
62 *
63 * +----------------+
64 * | (1) (R) |
65 * | 192.168.0.0/16 |
66 * | value: 1 |
67 * | [0] [1] |
68 * +----------------+
69 *
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
74 * child[0] of (1):
75 *
76 * +----------------+
77 * | (1) (R) |
78 * | 192.168.0.0/16 |
79 * | value: 1 |
80 * | [0] [1] |
81 * +----------------+
82 * |
83 * +----------------+
84 * | (2) |
85 * | 192.168.0.0/24 |
86 * | value: 2 |
87 * | [0] [1] |
88 * +----------------+
89 *
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,
92 * 192.168.128.0/24:
93 *
94 * +----------------+
95 * | (1) (R) |
96 * | 192.168.0.0/16 |
97 * | value: 1 |
98 * | [0] [1] |
99 * +----------------+
100 * | |
101 * +----------------+ +------------------+
102 * | (2) | | (3) |
103 * | 192.168.0.0/24 | | 192.168.128.0/24 |
104 * | value: 2 | | value: 3 |
105 * | [0] [1] | | [0] [1] |
106 * +----------------+ +------------------+
107 *
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:
116 *
117 * +----------------+
118 * | (1) (R) |
119 * | 192.168.0.0/16 |
120 * | value: 1 |
121 * | [0] [1] |
122 * +----------------+
123 * | |
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 * +----------------+ +------------------+
130 * | |
131 * +----------------+ +----------------+
132 * | (2) | | (5) |
133 * | 192.168.0.0/24 | | 192.168.1.0/24 |
134 * | value: 2 | | value: 5 |
135 * | [0] [1] | | [0] [1] |
136 * +----------------+ +----------------+
137 *
138 * 192.168.1.1/32 would be a child of (5) etc.
139 *
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.
142 *
143 * A fully populated trie would have a height of 32 nodes, as the trie was
144 * created with a prefix length of 32.
145 *
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
149 * returned.
150 */
151
extract_bit(const u8 * data,size_t index)152 static inline int extract_bit(const u8 *data, size_t index)
153 {
154 return !!(data[index / 8] & (1 << (7 - (index % 8))));
155 }
156
157 /**
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
162 *
163 * Determine the longest prefix of @node that matches the bits in @key.
164 */
165 static __always_inline
__longest_prefix_match(const struct lpm_trie * trie,const struct lpm_trie_node * node,const struct bpf_lpm_trie_key_u8 * key)166 size_t __longest_prefix_match(const struct lpm_trie *trie,
167 const struct lpm_trie_node *node,
168 const struct bpf_lpm_trie_key_u8 *key)
169 {
170 u32 limit = min(node->prefixlen, key->prefixlen);
171 u32 prefixlen = 0, i = 0;
172
173 BUILD_BUG_ON(offsetof(struct lpm_trie_node, data) % sizeof(u32));
174 BUILD_BUG_ON(offsetof(struct bpf_lpm_trie_key_u8, data) % sizeof(u32));
175
176 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && defined(CONFIG_64BIT)
177
178 /* data_size >= 16 has very small probability.
179 * We do not use a loop for optimal code generation.
180 */
181 if (trie->data_size >= 8) {
182 u64 diff = be64_to_cpu(*(__be64 *)node->data ^
183 *(__be64 *)key->data);
184
185 prefixlen = 64 - fls64(diff);
186 if (prefixlen >= limit)
187 return limit;
188 if (diff)
189 return prefixlen;
190 i = 8;
191 }
192 #endif
193
194 while (trie->data_size >= i + 4) {
195 u32 diff = be32_to_cpu(*(__be32 *)&node->data[i] ^
196 *(__be32 *)&key->data[i]);
197
198 prefixlen += 32 - fls(diff);
199 if (prefixlen >= limit)
200 return limit;
201 if (diff)
202 return prefixlen;
203 i += 4;
204 }
205
206 if (trie->data_size >= i + 2) {
207 u16 diff = be16_to_cpu(*(__be16 *)&node->data[i] ^
208 *(__be16 *)&key->data[i]);
209
210 prefixlen += 16 - fls(diff);
211 if (prefixlen >= limit)
212 return limit;
213 if (diff)
214 return prefixlen;
215 i += 2;
216 }
217
218 if (trie->data_size >= i + 1) {
219 prefixlen += 8 - fls(node->data[i] ^ key->data[i]);
220
221 if (prefixlen >= limit)
222 return limit;
223 }
224
225 return prefixlen;
226 }
227
longest_prefix_match(const struct lpm_trie * trie,const struct lpm_trie_node * node,const struct bpf_lpm_trie_key_u8 * key)228 static size_t longest_prefix_match(const struct lpm_trie *trie,
229 const struct lpm_trie_node *node,
230 const struct bpf_lpm_trie_key_u8 *key)
231 {
232 return __longest_prefix_match(trie, node, key);
233 }
234
235 /* Called from syscall or from eBPF program */
trie_lookup_elem(struct bpf_map * map,void * _key)236 static void *trie_lookup_elem(struct bpf_map *map, void *_key)
237 {
238 struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
239 struct lpm_trie_node *node, *found = NULL;
240 struct bpf_lpm_trie_key_u8 *key = _key;
241
242 if (key->prefixlen > trie->max_prefixlen)
243 return NULL;
244
245 /* Start walking the trie from the root node ... */
246
247 for (node = rcu_dereference_check(trie->root, rcu_read_lock_bh_held());
248 node;) {
249 unsigned int next_bit;
250 size_t matchlen;
251
252 /* Determine the longest prefix of @node that matches @key.
253 * If it's the maximum possible prefix for this trie, we have
254 * an exact match and can return it directly.
255 */
256 matchlen = __longest_prefix_match(trie, node, key);
257 if (matchlen == trie->max_prefixlen) {
258 found = node;
259 break;
260 }
261
262 /* If the number of bits that match is smaller than the prefix
263 * length of @node, bail out and return the node we have seen
264 * last in the traversal (ie, the parent).
265 */
266 if (matchlen < node->prefixlen)
267 break;
268
269 /* Consider this node as return candidate unless it is an
270 * artificially added intermediate one.
271 */
272 if (!(node->flags & LPM_TREE_NODE_FLAG_IM))
273 found = node;
274
275 /* If the node match is fully satisfied, let's see if we can
276 * become more specific. Determine the next bit in the key and
277 * traverse down.
278 */
279 next_bit = extract_bit(key->data, node->prefixlen);
280 node = rcu_dereference_check(node->child[next_bit],
281 rcu_read_lock_bh_held());
282 }
283
284 if (!found)
285 return NULL;
286
287 return found->data + trie->data_size;
288 }
289
lpm_trie_node_alloc(const struct lpm_trie * trie,const void * value)290 static struct lpm_trie_node *lpm_trie_node_alloc(const struct lpm_trie *trie,
291 const void *value)
292 {
293 struct lpm_trie_node *node;
294 size_t size = sizeof(struct lpm_trie_node) + trie->data_size;
295
296 if (value)
297 size += trie->map.value_size;
298
299 node = bpf_map_kmalloc_node(&trie->map, size, GFP_NOWAIT | __GFP_NOWARN,
300 trie->map.numa_node);
301 if (!node)
302 return NULL;
303
304 node->flags = 0;
305
306 if (value)
307 memcpy(node->data + trie->data_size, value,
308 trie->map.value_size);
309
310 return node;
311 }
312
trie_check_add_elem(struct lpm_trie * trie,u64 flags)313 static int trie_check_add_elem(struct lpm_trie *trie, u64 flags)
314 {
315 if (flags == BPF_EXIST)
316 return -ENOENT;
317 if (trie->n_entries == trie->map.max_entries)
318 return -ENOSPC;
319 trie->n_entries++;
320 return 0;
321 }
322
323 /* Called from syscall or from eBPF program */
trie_update_elem(struct bpf_map * map,void * _key,void * value,u64 flags)324 static long trie_update_elem(struct bpf_map *map,
325 void *_key, void *value, u64 flags)
326 {
327 struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
328 struct lpm_trie_node *node, *im_node, *new_node = NULL;
329 struct lpm_trie_node *free_node = NULL;
330 struct lpm_trie_node __rcu **slot;
331 struct bpf_lpm_trie_key_u8 *key = _key;
332 unsigned long irq_flags;
333 unsigned int next_bit;
334 size_t matchlen = 0;
335 int ret = 0;
336
337 if (unlikely(flags > BPF_EXIST))
338 return -EINVAL;
339
340 if (key->prefixlen > trie->max_prefixlen)
341 return -EINVAL;
342
343 spin_lock_irqsave(&trie->lock, irq_flags);
344
345 /* Allocate and fill a new node */
346 new_node = lpm_trie_node_alloc(trie, value);
347 if (!new_node) {
348 ret = -ENOMEM;
349 goto out;
350 }
351
352 new_node->prefixlen = key->prefixlen;
353 RCU_INIT_POINTER(new_node->child[0], NULL);
354 RCU_INIT_POINTER(new_node->child[1], NULL);
355 memcpy(new_node->data, key->data, trie->data_size);
356
357 /* Now find a slot to attach the new node. To do that, walk the tree
358 * from the root and match as many bits as possible for each node until
359 * we either find an empty slot or a slot that needs to be replaced by
360 * an intermediate node.
361 */
362 slot = &trie->root;
363
364 while ((node = rcu_dereference_protected(*slot,
365 lockdep_is_held(&trie->lock)))) {
366 matchlen = longest_prefix_match(trie, node, key);
367
368 if (node->prefixlen != matchlen ||
369 node->prefixlen == key->prefixlen ||
370 node->prefixlen == trie->max_prefixlen)
371 break;
372
373 next_bit = extract_bit(key->data, node->prefixlen);
374 slot = &node->child[next_bit];
375 }
376
377 /* If the slot is empty (a free child pointer or an empty root),
378 * simply assign the @new_node to that slot and be done.
379 */
380 if (!node) {
381 ret = trie_check_add_elem(trie, flags);
382 if (ret)
383 goto out;
384
385 rcu_assign_pointer(*slot, new_node);
386 goto out;
387 }
388
389 /* If the slot we picked already exists, replace it with @new_node
390 * which already has the correct data array set.
391 */
392 if (node->prefixlen == matchlen) {
393 if (!(node->flags & LPM_TREE_NODE_FLAG_IM)) {
394 if (flags == BPF_NOEXIST) {
395 ret = -EEXIST;
396 goto out;
397 }
398 } else {
399 ret = trie_check_add_elem(trie, flags);
400 if (ret)
401 goto out;
402 }
403
404 new_node->child[0] = node->child[0];
405 new_node->child[1] = node->child[1];
406
407 rcu_assign_pointer(*slot, new_node);
408 free_node = node;
409
410 goto out;
411 }
412
413 ret = trie_check_add_elem(trie, flags);
414 if (ret)
415 goto out;
416
417 /* If the new node matches the prefix completely, it must be inserted
418 * as an ancestor. Simply insert it between @node and *@slot.
419 */
420 if (matchlen == key->prefixlen) {
421 next_bit = extract_bit(node->data, matchlen);
422 rcu_assign_pointer(new_node->child[next_bit], node);
423 rcu_assign_pointer(*slot, new_node);
424 goto out;
425 }
426
427 im_node = lpm_trie_node_alloc(trie, NULL);
428 if (!im_node) {
429 trie->n_entries--;
430 ret = -ENOMEM;
431 goto out;
432 }
433
434 im_node->prefixlen = matchlen;
435 im_node->flags |= LPM_TREE_NODE_FLAG_IM;
436 memcpy(im_node->data, node->data, trie->data_size);
437
438 /* Now determine which child to install in which slot */
439 if (extract_bit(key->data, matchlen)) {
440 rcu_assign_pointer(im_node->child[0], node);
441 rcu_assign_pointer(im_node->child[1], new_node);
442 } else {
443 rcu_assign_pointer(im_node->child[0], new_node);
444 rcu_assign_pointer(im_node->child[1], node);
445 }
446
447 /* Finally, assign the intermediate node to the determined slot */
448 rcu_assign_pointer(*slot, im_node);
449
450 out:
451 if (ret)
452 kfree(new_node);
453 spin_unlock_irqrestore(&trie->lock, irq_flags);
454 kfree_rcu(free_node, rcu);
455
456 return ret;
457 }
458
459 /* Called from syscall or from eBPF program */
trie_delete_elem(struct bpf_map * map,void * _key)460 static long trie_delete_elem(struct bpf_map *map, void *_key)
461 {
462 struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
463 struct lpm_trie_node *free_node = NULL, *free_parent = NULL;
464 struct bpf_lpm_trie_key_u8 *key = _key;
465 struct lpm_trie_node __rcu **trim, **trim2;
466 struct lpm_trie_node *node, *parent;
467 unsigned long irq_flags;
468 unsigned int next_bit;
469 size_t matchlen = 0;
470 int ret = 0;
471
472 if (key->prefixlen > trie->max_prefixlen)
473 return -EINVAL;
474
475 spin_lock_irqsave(&trie->lock, irq_flags);
476
477 /* Walk the tree looking for an exact key/length match and keeping
478 * track of the path we traverse. We will need to know the node
479 * we wish to delete, and the slot that points to the node we want
480 * to delete. We may also need to know the nodes parent and the
481 * slot that contains it.
482 */
483 trim = &trie->root;
484 trim2 = trim;
485 parent = NULL;
486 while ((node = rcu_dereference_protected(
487 *trim, lockdep_is_held(&trie->lock)))) {
488 matchlen = longest_prefix_match(trie, node, key);
489
490 if (node->prefixlen != matchlen ||
491 node->prefixlen == key->prefixlen)
492 break;
493
494 parent = node;
495 trim2 = trim;
496 next_bit = extract_bit(key->data, node->prefixlen);
497 trim = &node->child[next_bit];
498 }
499
500 if (!node || node->prefixlen != key->prefixlen ||
501 node->prefixlen != matchlen ||
502 (node->flags & LPM_TREE_NODE_FLAG_IM)) {
503 ret = -ENOENT;
504 goto out;
505 }
506
507 trie->n_entries--;
508
509 /* If the node we are removing has two children, simply mark it
510 * as intermediate and we are done.
511 */
512 if (rcu_access_pointer(node->child[0]) &&
513 rcu_access_pointer(node->child[1])) {
514 node->flags |= LPM_TREE_NODE_FLAG_IM;
515 goto out;
516 }
517
518 /* If the parent of the node we are about to delete is an intermediate
519 * node, and the deleted node doesn't have any children, we can delete
520 * the intermediate parent as well and promote its other child
521 * up the tree. Doing this maintains the invariant that all
522 * intermediate nodes have exactly 2 children and that there are no
523 * unnecessary intermediate nodes in the tree.
524 */
525 if (parent && (parent->flags & LPM_TREE_NODE_FLAG_IM) &&
526 !node->child[0] && !node->child[1]) {
527 if (node == rcu_access_pointer(parent->child[0]))
528 rcu_assign_pointer(
529 *trim2, rcu_access_pointer(parent->child[1]));
530 else
531 rcu_assign_pointer(
532 *trim2, rcu_access_pointer(parent->child[0]));
533 free_parent = parent;
534 free_node = node;
535 goto out;
536 }
537
538 /* The node we are removing has either zero or one child. If there
539 * is a child, move it into the removed node's slot then delete
540 * the node. Otherwise just clear the slot and delete the node.
541 */
542 if (node->child[0])
543 rcu_assign_pointer(*trim, rcu_access_pointer(node->child[0]));
544 else if (node->child[1])
545 rcu_assign_pointer(*trim, rcu_access_pointer(node->child[1]));
546 else
547 RCU_INIT_POINTER(*trim, NULL);
548 free_node = node;
549
550 out:
551 spin_unlock_irqrestore(&trie->lock, irq_flags);
552 kfree_rcu(free_parent, rcu);
553 kfree_rcu(free_node, rcu);
554
555 return ret;
556 }
557
558 #define LPM_DATA_SIZE_MAX 256
559 #define LPM_DATA_SIZE_MIN 1
560
561 #define LPM_VAL_SIZE_MAX (KMALLOC_MAX_SIZE - LPM_DATA_SIZE_MAX - \
562 sizeof(struct lpm_trie_node))
563 #define LPM_VAL_SIZE_MIN 1
564
565 #define LPM_KEY_SIZE(X) (sizeof(struct bpf_lpm_trie_key_u8) + (X))
566 #define LPM_KEY_SIZE_MAX LPM_KEY_SIZE(LPM_DATA_SIZE_MAX)
567 #define LPM_KEY_SIZE_MIN LPM_KEY_SIZE(LPM_DATA_SIZE_MIN)
568
569 #define LPM_CREATE_FLAG_MASK (BPF_F_NO_PREALLOC | BPF_F_NUMA_NODE | \
570 BPF_F_ACCESS_MASK)
571
trie_alloc(union bpf_attr * attr)572 static struct bpf_map *trie_alloc(union bpf_attr *attr)
573 {
574 struct lpm_trie *trie;
575
576 /* check sanity of attributes */
577 if (attr->max_entries == 0 ||
578 !(attr->map_flags & BPF_F_NO_PREALLOC) ||
579 attr->map_flags & ~LPM_CREATE_FLAG_MASK ||
580 !bpf_map_flags_access_ok(attr->map_flags) ||
581 attr->key_size < LPM_KEY_SIZE_MIN ||
582 attr->key_size > LPM_KEY_SIZE_MAX ||
583 attr->value_size < LPM_VAL_SIZE_MIN ||
584 attr->value_size > LPM_VAL_SIZE_MAX)
585 return ERR_PTR(-EINVAL);
586
587 trie = bpf_map_area_alloc(sizeof(*trie), NUMA_NO_NODE);
588 if (!trie)
589 return ERR_PTR(-ENOMEM);
590
591 /* copy mandatory map attributes */
592 bpf_map_init_from_attr(&trie->map, attr);
593 trie->data_size = attr->key_size -
594 offsetof(struct bpf_lpm_trie_key_u8, data);
595 trie->max_prefixlen = trie->data_size * 8;
596
597 spin_lock_init(&trie->lock);
598
599 return &trie->map;
600 }
601
trie_free(struct bpf_map * map)602 static void trie_free(struct bpf_map *map)
603 {
604 struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
605 struct lpm_trie_node __rcu **slot;
606 struct lpm_trie_node *node;
607
608 /* Always start at the root and walk down to a node that has no
609 * children. Then free that node, nullify its reference in the parent
610 * and start over.
611 */
612
613 for (;;) {
614 slot = &trie->root;
615
616 for (;;) {
617 node = rcu_dereference_protected(*slot, 1);
618 if (!node)
619 goto out;
620
621 if (rcu_access_pointer(node->child[0])) {
622 slot = &node->child[0];
623 continue;
624 }
625
626 if (rcu_access_pointer(node->child[1])) {
627 slot = &node->child[1];
628 continue;
629 }
630
631 kfree(node);
632 RCU_INIT_POINTER(*slot, NULL);
633 break;
634 }
635 }
636
637 out:
638 bpf_map_area_free(trie);
639 }
640
trie_get_next_key(struct bpf_map * map,void * _key,void * _next_key)641 static int trie_get_next_key(struct bpf_map *map, void *_key, void *_next_key)
642 {
643 struct lpm_trie_node *node, *next_node = NULL, *parent, *search_root;
644 struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
645 struct bpf_lpm_trie_key_u8 *key = _key, *next_key = _next_key;
646 struct lpm_trie_node **node_stack = NULL;
647 int err = 0, stack_ptr = -1;
648 unsigned int next_bit;
649 size_t matchlen = 0;
650
651 /* The get_next_key follows postorder. For the 4 node example in
652 * the top of this file, the trie_get_next_key() returns the following
653 * one after another:
654 * 192.168.0.0/24
655 * 192.168.1.0/24
656 * 192.168.128.0/24
657 * 192.168.0.0/16
658 *
659 * The idea is to return more specific keys before less specific ones.
660 */
661
662 /* Empty trie */
663 search_root = rcu_dereference(trie->root);
664 if (!search_root)
665 return -ENOENT;
666
667 /* For invalid key, find the leftmost node in the trie */
668 if (!key || key->prefixlen > trie->max_prefixlen)
669 goto find_leftmost;
670
671 node_stack = kmalloc_array(trie->max_prefixlen + 1,
672 sizeof(struct lpm_trie_node *),
673 GFP_ATOMIC | __GFP_NOWARN);
674 if (!node_stack)
675 return -ENOMEM;
676
677 /* Try to find the exact node for the given key */
678 for (node = search_root; node;) {
679 node_stack[++stack_ptr] = node;
680 matchlen = longest_prefix_match(trie, node, key);
681 if (node->prefixlen != matchlen ||
682 node->prefixlen == key->prefixlen)
683 break;
684
685 next_bit = extract_bit(key->data, node->prefixlen);
686 node = rcu_dereference(node->child[next_bit]);
687 }
688 if (!node || node->prefixlen != matchlen ||
689 (node->flags & LPM_TREE_NODE_FLAG_IM))
690 goto find_leftmost;
691
692 /* The node with the exactly-matching key has been found,
693 * find the first node in postorder after the matched node.
694 */
695 node = node_stack[stack_ptr];
696 while (stack_ptr > 0) {
697 parent = node_stack[stack_ptr - 1];
698 if (rcu_dereference(parent->child[0]) == node) {
699 search_root = rcu_dereference(parent->child[1]);
700 if (search_root)
701 goto find_leftmost;
702 }
703 if (!(parent->flags & LPM_TREE_NODE_FLAG_IM)) {
704 next_node = parent;
705 goto do_copy;
706 }
707
708 node = parent;
709 stack_ptr--;
710 }
711
712 /* did not find anything */
713 err = -ENOENT;
714 goto free_stack;
715
716 find_leftmost:
717 /* Find the leftmost non-intermediate node, all intermediate nodes
718 * have exact two children, so this function will never return NULL.
719 */
720 for (node = search_root; node;) {
721 if (node->flags & LPM_TREE_NODE_FLAG_IM) {
722 node = rcu_dereference(node->child[0]);
723 } else {
724 next_node = node;
725 node = rcu_dereference(node->child[0]);
726 if (!node)
727 node = rcu_dereference(next_node->child[1]);
728 }
729 }
730 do_copy:
731 next_key->prefixlen = next_node->prefixlen;
732 memcpy((void *)next_key + offsetof(struct bpf_lpm_trie_key_u8, data),
733 next_node->data, trie->data_size);
734 free_stack:
735 kfree(node_stack);
736 return err;
737 }
738
trie_check_btf(const struct bpf_map * map,const struct btf * btf,const struct btf_type * key_type,const struct btf_type * value_type)739 static int trie_check_btf(const struct bpf_map *map,
740 const struct btf *btf,
741 const struct btf_type *key_type,
742 const struct btf_type *value_type)
743 {
744 /* Keys must have struct bpf_lpm_trie_key_u8 embedded. */
745 return BTF_INFO_KIND(key_type->info) != BTF_KIND_STRUCT ?
746 -EINVAL : 0;
747 }
748
trie_mem_usage(const struct bpf_map * map)749 static u64 trie_mem_usage(const struct bpf_map *map)
750 {
751 struct lpm_trie *trie = container_of(map, struct lpm_trie, map);
752 u64 elem_size;
753
754 elem_size = sizeof(struct lpm_trie_node) + trie->data_size +
755 trie->map.value_size;
756 return elem_size * READ_ONCE(trie->n_entries);
757 }
758
759 BTF_ID_LIST_SINGLE(trie_map_btf_ids, struct, lpm_trie)
760 const struct bpf_map_ops trie_map_ops = {
761 .map_meta_equal = bpf_map_meta_equal,
762 .map_alloc = trie_alloc,
763 .map_free = trie_free,
764 .map_get_next_key = trie_get_next_key,
765 .map_lookup_elem = trie_lookup_elem,
766 .map_update_elem = trie_update_elem,
767 .map_delete_elem = trie_delete_elem,
768 .map_lookup_batch = generic_map_lookup_batch,
769 .map_update_batch = generic_map_update_batch,
770 .map_delete_batch = generic_map_delete_batch,
771 .map_check_btf = trie_check_btf,
772 .map_mem_usage = trie_mem_usage,
773 .map_btf_id = &trie_map_btf_ids[0],
774 };
775