2 * Copyright (c) 2009, 2010, 2011, 2012, 2013 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "classifier.h"
20 #include <netinet/in.h>
21 #include "byte-order.h"
22 #include "dynamic-string.h"
27 #include "ovs-thread.h"
31 VLOG_DEFINE_THIS_MODULE(classifier);
36 /* Ports trie depends on both ports sharing the same ovs_be32. */
37 #define TP_PORTS_OFS32 (offsetof(struct flow, tp_src) / 4)
38 BUILD_ASSERT_DECL(TP_PORTS_OFS32 == offsetof(struct flow, tp_dst) / 4);
40 /* Prefix trie for a 'field' */
42 const struct mf_field *field; /* Trie field, or NULL. */
43 struct trie_node *root; /* NULL if none. */
46 struct cls_subtable_entry {
47 struct cls_subtable *subtable;
49 unsigned int max_priority;
52 struct cls_subtable_cache {
53 struct cls_subtable_entry *subtables;
54 size_t alloc_size; /* Number of allocated elements. */
55 size_t size; /* One past last valid array element. */
59 CLS_MAX_INDICES = 3 /* Maximum number of lookup indices per subtable. */
62 struct cls_classifier {
63 int n_rules; /* Total number of rules. */
64 uint8_t n_flow_segments;
65 uint8_t flow_segments[CLS_MAX_INDICES]; /* Flow segment boundaries to use
66 * for staged lookup. */
67 struct hmap subtables; /* Contains "struct cls_subtable"s. */
68 struct cls_subtable_cache subtables_priority;
69 struct hmap partitions; /* Contains "struct cls_partition"s. */
70 struct cls_trie tries[CLS_MAX_TRIES]; /* Prefix tries. */
74 /* A set of rules that all have the same fields wildcarded. */
76 struct hmap_node hmap_node; /* Within struct cls_classifier 'subtables'
78 struct hmap rules; /* Contains "struct cls_rule"s. */
79 int n_rules; /* Number of rules, including duplicates. */
80 unsigned int max_priority; /* Max priority of any rule in the subtable. */
81 unsigned int max_count; /* Count of max_priority rules. */
82 tag_type tag; /* Tag generated from mask for partitioning. */
83 uint8_t n_indices; /* How many indices to use. */
84 uint8_t index_ofs[CLS_MAX_INDICES]; /* u32 flow segment boundaries. */
85 struct hindex indices[CLS_MAX_INDICES]; /* Staged lookup indices. */
86 unsigned int trie_plen[CLS_MAX_TRIES]; /* Trie prefix length in 'mask'. */
88 struct trie_node *ports_trie; /* NULL if none. */
89 struct minimask mask; /* Wildcards for fields. */
90 /* 'mask' must be the last field. */
93 /* Associates a metadata value (that is, a value of the OpenFlow 1.1+ metadata
94 * field) with tags for the "cls_subtable"s that contain rules that match that
96 struct cls_partition {
97 struct hmap_node hmap_node; /* In struct cls_classifier's 'partitions'
99 ovs_be64 metadata; /* metadata value for this partition. */
100 tag_type tags; /* OR of each flow's cls_subtable tag. */
101 struct tag_tracker tracker; /* Tracks the bits in 'tags'. */
104 /* Internal representation of a rule in a "struct cls_subtable". */
106 struct cls_rule *cls_rule;
107 struct hindex_node index_nodes[CLS_MAX_INDICES]; /* Within subtable's
109 struct hmap_node hmap_node; /* Within struct cls_subtable 'rules'. */
110 unsigned int priority; /* Larger numbers are higher priorities. */
111 struct cls_partition *partition;
112 struct list list; /* List of identical, lower-priority rules. */
113 struct miniflow flow; /* Matching rule. Mask is in the subtable. */
114 /* 'flow' must be the last field. */
117 static struct cls_match *
118 cls_match_alloc(struct cls_rule *rule)
120 int count = count_1bits(rule->match.flow.map);
122 struct cls_match *cls_match
123 = xmalloc(sizeof *cls_match - sizeof cls_match->flow.inline_values
124 + MINIFLOW_VALUES_SIZE(count));
126 cls_match->cls_rule = rule;
127 miniflow_clone_inline(&cls_match->flow, &rule->match.flow, count);
128 cls_match->priority = rule->priority;
129 rule->cls_match = cls_match;
134 static struct cls_subtable *find_subtable(const struct cls_classifier *,
135 const struct minimask *);
136 static struct cls_subtable *insert_subtable(struct cls_classifier *,
137 const struct minimask *);
139 static void destroy_subtable(struct cls_classifier *, struct cls_subtable *);
141 static void update_subtables_after_insertion(struct cls_classifier *,
142 struct cls_subtable *,
143 unsigned int new_priority);
144 static void update_subtables_after_removal(struct cls_classifier *,
145 struct cls_subtable *,
146 unsigned int del_priority);
148 static struct cls_match *find_match_wc(const struct cls_subtable *,
149 const struct flow *, struct trie_ctx *,
150 unsigned int n_tries,
151 struct flow_wildcards *);
152 static struct cls_match *find_equal(struct cls_subtable *,
153 const struct miniflow *, uint32_t hash);
154 static struct cls_match *insert_rule(struct cls_classifier *,
155 struct cls_subtable *, struct cls_rule *);
157 /* Iterates RULE over HEAD and all of the cls_rules on HEAD->list. */
158 #define FOR_EACH_RULE_IN_LIST(RULE, HEAD) \
159 for ((RULE) = (HEAD); (RULE) != NULL; (RULE) = next_rule_in_list(RULE))
160 #define FOR_EACH_RULE_IN_LIST_SAFE(RULE, NEXT, HEAD) \
161 for ((RULE) = (HEAD); \
162 (RULE) != NULL && ((NEXT) = next_rule_in_list(RULE), true); \
165 static struct cls_match *next_rule_in_list__(struct cls_match *);
166 static struct cls_match *next_rule_in_list(struct cls_match *);
168 static unsigned int minimask_get_prefix_len(const struct minimask *,
169 const struct mf_field *);
170 static void trie_init(struct cls_classifier *, int trie_idx,
171 const struct mf_field *);
172 static unsigned int trie_lookup(const struct cls_trie *, const struct flow *,
173 unsigned int *checkbits);
174 static unsigned int trie_lookup_value(const struct trie_node *,
175 const ovs_be32 value[],
176 unsigned int *checkbits);
177 static void trie_destroy(struct trie_node *);
178 static void trie_insert(struct cls_trie *, const struct cls_rule *, int mlen);
179 static void trie_insert_prefix(struct trie_node **, const ovs_be32 *prefix,
181 static void trie_remove(struct cls_trie *, const struct cls_rule *, int mlen);
182 static void trie_remove_prefix(struct trie_node **, const ovs_be32 *prefix,
184 static void mask_set_prefix_bits(struct flow_wildcards *, uint8_t be32ofs,
186 static bool mask_prefix_bits_set(const struct flow_wildcards *,
187 uint8_t be32ofs, unsigned int nbits);
190 cls_subtable_cache_init(struct cls_subtable_cache *array)
192 memset(array, 0, sizeof *array);
196 cls_subtable_cache_destroy(struct cls_subtable_cache *array)
198 free(array->subtables);
199 memset(array, 0, sizeof *array);
202 /* Array insertion. */
204 cls_subtable_cache_push_back(struct cls_subtable_cache *array,
205 struct cls_subtable_entry a)
207 if (array->size == array->alloc_size) {
208 array->subtables = x2nrealloc(array->subtables, &array->alloc_size,
212 array->subtables[array->size++] = a;
215 /* Move subtable entry at 'from' to 'to', shifting the elements in between
216 * (including the one at 'to') accordingly. */
218 cls_subtable_cache_move(struct cls_subtable_entry *to,
219 struct cls_subtable_entry *from)
222 struct cls_subtable_entry temp = *from;
225 /* Shift entries (from,to] backwards to make space at 'to'. */
226 memmove(from, from + 1, (to - from) * sizeof *to);
228 /* Shift entries [to,from) forward to make space at 'to'. */
229 memmove(to + 1, to, (from - to) * sizeof *to);
238 cls_subtable_cache_remove(struct cls_subtable_cache *array,
239 struct cls_subtable_entry *elem)
241 ssize_t size = (&array->subtables[array->size]
242 - (elem + 1)) * sizeof *elem;
244 memmove(elem, elem + 1, size);
249 #define CLS_SUBTABLE_CACHE_FOR_EACH(SUBTABLE, ITER, ARRAY) \
250 for (ITER = (ARRAY)->subtables; \
251 ITER < &(ARRAY)->subtables[(ARRAY)->size] \
252 && OVS_LIKELY(SUBTABLE = ITER->subtable); \
254 #define CLS_SUBTABLE_CACHE_FOR_EACH_CONTINUE(SUBTABLE, ITER, ARRAY) \
256 ITER < &(ARRAY)->subtables[(ARRAY)->size] \
257 && OVS_LIKELY(SUBTABLE = ITER->subtable); \
259 #define CLS_SUBTABLE_CACHE_FOR_EACH_REVERSE(SUBTABLE, ITER, ARRAY) \
260 for (ITER = &(ARRAY)->subtables[(ARRAY)->size]; \
261 ITER > (ARRAY)->subtables \
262 && OVS_LIKELY(SUBTABLE = (--ITER)->subtable);)
265 cls_subtable_cache_verify(struct cls_subtable_cache *array)
267 struct cls_subtable *table;
268 struct cls_subtable_entry *iter;
269 unsigned int priority = 0;
271 CLS_SUBTABLE_CACHE_FOR_EACH_REVERSE (table, iter, array) {
272 if (iter->max_priority != table->max_priority) {
273 VLOG_WARN("Subtable %p has mismatching priority in cache (%u != %u)",
274 table, iter->max_priority, table->max_priority);
276 if (iter->max_priority < priority) {
277 VLOG_WARN("Subtable cache is out of order (%u < %u)",
278 iter->max_priority, priority);
280 priority = iter->max_priority;
285 cls_subtable_cache_reset(struct cls_classifier *cls)
287 struct cls_subtable_cache old = cls->subtables_priority;
288 struct cls_subtable *subtable;
290 VLOG_WARN("Resetting subtable cache.");
292 cls_subtable_cache_verify(&cls->subtables_priority);
294 cls_subtable_cache_init(&cls->subtables_priority);
296 HMAP_FOR_EACH (subtable, hmap_node, &cls->subtables) {
297 struct cls_match *head;
298 struct cls_subtable_entry elem;
299 struct cls_subtable *table;
300 struct cls_subtable_entry *iter, *from = NULL;
301 unsigned int new_max = 0;
302 unsigned int max_count = 0;
305 /* Verify max_priority. */
306 HMAP_FOR_EACH (head, hmap_node, &subtable->rules) {
307 if (head->priority > new_max) {
308 new_max = head->priority;
310 } else if (head->priority == new_max) {
314 if (new_max != subtable->max_priority ||
315 max_count != subtable->max_count) {
316 VLOG_WARN("subtable %p (%u rules) has mismatching max_priority "
317 "(%u) or max_count (%u). Highest priority found was %u, "
319 subtable, subtable->n_rules, subtable->max_priority,
320 subtable->max_count, new_max, max_count);
321 subtable->max_priority = new_max;
322 subtable->max_count = max_count;
325 /* Locate the subtable from the old cache. */
327 CLS_SUBTABLE_CACHE_FOR_EACH (table, iter, &old) {
328 if (table == subtable) {
329 if (iter->max_priority != new_max) {
330 VLOG_WARN("Subtable %p has wrong max priority (%u != %u) "
332 subtable, iter->max_priority, new_max);
335 VLOG_WARN("Subtable %p duplicated in the old cache.",
342 VLOG_WARN("Subtable %p not found from the old cache.", subtable);
345 elem.subtable = subtable;
346 elem.tag = subtable->tag;
347 elem.max_priority = subtable->max_priority;
348 cls_subtable_cache_push_back(&cls->subtables_priority, elem);
350 /* Possibly move 'subtable' earlier in the priority array. If
351 * we break out of the loop, then the subtable (at 'from')
352 * should be moved to the position right after the current
353 * element. If the loop terminates normally, then 'iter' will
354 * be at the first array element and we'll move the subtable
355 * to the front of the array. */
356 CLS_SUBTABLE_CACHE_FOR_EACH_REVERSE (table, iter,
357 &cls->subtables_priority) {
358 if (table == subtable) {
359 from = iter; /* Locate the subtable as we go. */
360 } else if (table->max_priority >= new_max) {
361 ovs_assert(from != NULL);
362 iter++; /* After this. */
367 /* Move subtable at 'from' to 'iter'. */
368 cls_subtable_cache_move(iter, from);
371 /* Verify that the old and the new have the same size. */
372 if (old.size != cls->subtables_priority.size) {
373 VLOG_WARN("subtables cache sizes differ: old (%"PRIuSIZE
374 ") != new (%"PRIuSIZE").",
375 old.size, cls->subtables_priority.size);
378 cls_subtable_cache_destroy(&old);
380 cls_subtable_cache_verify(&cls->subtables_priority);
384 /* flow/miniflow/minimask/minimatch utilities.
385 * These are only used by the classifier, so place them here to allow
386 * for better optimization. */
388 static inline uint64_t
389 miniflow_get_map_in_range(const struct miniflow *miniflow,
390 uint8_t start, uint8_t end, unsigned int *offset)
392 uint64_t map = miniflow->map;
396 uint64_t msk = (UINT64_C(1) << start) - 1; /* 'start' LSBs set */
397 *offset = count_1bits(map & msk);
400 if (end < FLOW_U32S) {
401 uint64_t msk = (UINT64_C(1) << end) - 1; /* 'end' LSBs set */
407 /* Returns a hash value for the bits of 'flow' where there are 1-bits in
408 * 'mask', given 'basis'.
410 * The hash values returned by this function are the same as those returned by
411 * miniflow_hash_in_minimask(), only the form of the arguments differ. */
412 static inline uint32_t
413 flow_hash_in_minimask(const struct flow *flow, const struct minimask *mask,
416 const uint32_t *mask_values = miniflow_get_u32_values(&mask->masks);
417 const uint32_t *flow_u32 = (const uint32_t *)flow;
418 const uint32_t *p = mask_values;
423 for (map = mask->masks.map; map; map = zero_rightmost_1bit(map)) {
424 hash = mhash_add(hash, flow_u32[raw_ctz(map)] & *p++);
427 return mhash_finish(hash, (p - mask_values) * 4);
430 /* Returns a hash value for the bits of 'flow' where there are 1-bits in
431 * 'mask', given 'basis'.
433 * The hash values returned by this function are the same as those returned by
434 * flow_hash_in_minimask(), only the form of the arguments differ. */
435 static inline uint32_t
436 miniflow_hash_in_minimask(const struct miniflow *flow,
437 const struct minimask *mask, uint32_t basis)
439 const uint32_t *mask_values = miniflow_get_u32_values(&mask->masks);
440 const uint32_t *p = mask_values;
441 uint32_t hash = basis;
444 MINIFLOW_FOR_EACH_IN_MAP(flow_u32, flow, mask->masks.map) {
445 hash = mhash_add(hash, flow_u32 & *p++);
448 return mhash_finish(hash, (p - mask_values) * 4);
451 /* Returns a hash value for the bits of range [start, end) in 'flow',
452 * where there are 1-bits in 'mask', given 'hash'.
454 * The hash values returned by this function are the same as those returned by
455 * minimatch_hash_range(), only the form of the arguments differ. */
456 static inline uint32_t
457 flow_hash_in_minimask_range(const struct flow *flow,
458 const struct minimask *mask,
459 uint8_t start, uint8_t end, uint32_t *basis)
461 const uint32_t *mask_values = miniflow_get_u32_values(&mask->masks);
462 const uint32_t *flow_u32 = (const uint32_t *)flow;
464 uint64_t map = miniflow_get_map_in_range(&mask->masks, start, end,
466 const uint32_t *p = mask_values + offset;
467 uint32_t hash = *basis;
469 for (; map; map = zero_rightmost_1bit(map)) {
470 hash = mhash_add(hash, flow_u32[raw_ctz(map)] & *p++);
473 *basis = hash; /* Allow continuation from the unfinished value. */
474 return mhash_finish(hash, (p - mask_values) * 4);
477 /* Fold minimask 'mask''s wildcard mask into 'wc's wildcard mask. */
479 flow_wildcards_fold_minimask(struct flow_wildcards *wc,
480 const struct minimask *mask)
482 flow_union_with_miniflow(&wc->masks, &mask->masks);
485 /* Fold minimask 'mask''s wildcard mask into 'wc's wildcard mask
486 * in range [start, end). */
488 flow_wildcards_fold_minimask_range(struct flow_wildcards *wc,
489 const struct minimask *mask,
490 uint8_t start, uint8_t end)
492 uint32_t *dst_u32 = (uint32_t *)&wc->masks;
494 uint64_t map = miniflow_get_map_in_range(&mask->masks, start, end,
496 const uint32_t *p = miniflow_get_u32_values(&mask->masks) + offset;
498 for (; map; map = zero_rightmost_1bit(map)) {
499 dst_u32[raw_ctz(map)] |= *p++;
503 /* Returns a hash value for 'flow', given 'basis'. */
504 static inline uint32_t
505 miniflow_hash(const struct miniflow *flow, uint32_t basis)
507 const uint32_t *values = miniflow_get_u32_values(flow);
508 const uint32_t *p = values;
509 uint32_t hash = basis;
510 uint64_t hash_map = 0;
513 for (map = flow->map; map; map = zero_rightmost_1bit(map)) {
515 hash = mhash_add(hash, *p);
516 hash_map |= rightmost_1bit(map);
520 hash = mhash_add(hash, hash_map);
521 hash = mhash_add(hash, hash_map >> 32);
523 return mhash_finish(hash, p - values);
526 /* Returns a hash value for 'mask', given 'basis'. */
527 static inline uint32_t
528 minimask_hash(const struct minimask *mask, uint32_t basis)
530 return miniflow_hash(&mask->masks, basis);
533 /* Returns a hash value for 'match', given 'basis'. */
534 static inline uint32_t
535 minimatch_hash(const struct minimatch *match, uint32_t basis)
537 return miniflow_hash(&match->flow, minimask_hash(&match->mask, basis));
540 /* Returns a hash value for the bits of range [start, end) in 'minimatch',
543 * The hash values returned by this function are the same as those returned by
544 * flow_hash_in_minimask_range(), only the form of the arguments differ. */
545 static inline uint32_t
546 minimatch_hash_range(const struct minimatch *match, uint8_t start, uint8_t end,
550 const uint32_t *p, *q;
551 uint32_t hash = *basis;
554 n = count_1bits(miniflow_get_map_in_range(&match->mask.masks, start, end,
556 q = miniflow_get_u32_values(&match->mask.masks) + offset;
557 p = miniflow_get_u32_values(&match->flow) + offset;
559 for (i = 0; i < n; i++) {
560 hash = mhash_add(hash, p[i] & q[i]);
562 *basis = hash; /* Allow continuation from the unfinished value. */
563 return mhash_finish(hash, (offset + n) * 4);
569 /* Initializes 'rule' to match packets specified by 'match' at the given
570 * 'priority'. 'match' must satisfy the invariant described in the comment at
571 * the definition of struct match.
573 * The caller must eventually destroy 'rule' with cls_rule_destroy().
575 * (OpenFlow uses priorities between 0 and UINT16_MAX, inclusive, but
576 * internally Open vSwitch supports a wider range.) */
578 cls_rule_init(struct cls_rule *rule,
579 const struct match *match, unsigned int priority)
581 minimatch_init(&rule->match, match);
582 rule->priority = priority;
583 rule->cls_match = NULL;
586 /* Same as cls_rule_init() for initialization from a "struct minimatch". */
588 cls_rule_init_from_minimatch(struct cls_rule *rule,
589 const struct minimatch *match,
590 unsigned int priority)
592 minimatch_clone(&rule->match, match);
593 rule->priority = priority;
594 rule->cls_match = NULL;
597 /* Initializes 'dst' as a copy of 'src'.
599 * The caller must eventually destroy 'dst' with cls_rule_destroy(). */
601 cls_rule_clone(struct cls_rule *dst, const struct cls_rule *src)
603 minimatch_clone(&dst->match, &src->match);
604 dst->priority = src->priority;
605 dst->cls_match = NULL;
608 /* Initializes 'dst' with the data in 'src', destroying 'src'.
610 * The caller must eventually destroy 'dst' with cls_rule_destroy(). */
612 cls_rule_move(struct cls_rule *dst, struct cls_rule *src)
614 minimatch_move(&dst->match, &src->match);
615 dst->priority = src->priority;
616 dst->cls_match = NULL;
619 /* Frees memory referenced by 'rule'. Doesn't free 'rule' itself (it's
620 * normally embedded into a larger structure).
622 * ('rule' must not currently be in a classifier.) */
624 cls_rule_destroy(struct cls_rule *rule)
626 ovs_assert(!rule->cls_match);
627 minimatch_destroy(&rule->match);
630 /* Returns true if 'a' and 'b' match the same packets at the same priority,
631 * false if they differ in some way. */
633 cls_rule_equal(const struct cls_rule *a, const struct cls_rule *b)
635 return a->priority == b->priority && minimatch_equal(&a->match, &b->match);
638 /* Returns a hash value for 'rule', folding in 'basis'. */
640 cls_rule_hash(const struct cls_rule *rule, uint32_t basis)
642 return minimatch_hash(&rule->match, hash_int(rule->priority, basis));
645 /* Appends a string describing 'rule' to 's'. */
647 cls_rule_format(const struct cls_rule *rule, struct ds *s)
649 minimatch_format(&rule->match, s, rule->priority);
652 /* Returns true if 'rule' matches every packet, false otherwise. */
654 cls_rule_is_catchall(const struct cls_rule *rule)
656 return minimask_is_catchall(&rule->match.mask);
659 /* Initializes 'cls' as a classifier that initially contains no classification
662 classifier_init(struct classifier *cls_, const uint8_t *flow_segments)
664 struct cls_classifier *cls = xmalloc(sizeof *cls);
666 fat_rwlock_init(&cls_->rwlock);
671 hmap_init(&cls->subtables);
672 cls_subtable_cache_init(&cls->subtables_priority);
673 hmap_init(&cls->partitions);
674 cls->n_flow_segments = 0;
676 while (cls->n_flow_segments < CLS_MAX_INDICES
677 && *flow_segments < FLOW_U32S) {
678 cls->flow_segments[cls->n_flow_segments++] = *flow_segments++;
684 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
685 * caller's responsibility. */
687 classifier_destroy(struct classifier *cls_)
690 struct cls_classifier *cls = cls_->cls;
691 struct cls_subtable *partition, *next_partition;
692 struct cls_subtable *subtable, *next_subtable;
695 fat_rwlock_destroy(&cls_->rwlock);
700 for (i = 0; i < cls->n_tries; i++) {
701 trie_destroy(cls->tries[i].root);
704 HMAP_FOR_EACH_SAFE (subtable, next_subtable, hmap_node,
706 destroy_subtable(cls, subtable);
708 hmap_destroy(&cls->subtables);
710 HMAP_FOR_EACH_SAFE (partition, next_partition, hmap_node,
712 hmap_remove(&cls->partitions, &partition->hmap_node);
715 hmap_destroy(&cls->partitions);
717 cls_subtable_cache_destroy(&cls->subtables_priority);
722 /* We use uint64_t as a set for the fields below. */
723 BUILD_ASSERT_DECL(MFF_N_IDS <= 64);
725 /* Set the fields for which prefix lookup should be performed. */
727 classifier_set_prefix_fields(struct classifier *cls_,
728 const enum mf_field_id *trie_fields,
729 unsigned int n_fields)
731 struct cls_classifier *cls = cls_->cls;
735 for (i = 0, trie = 0; i < n_fields && trie < CLS_MAX_TRIES; i++) {
736 const struct mf_field *field = mf_from_id(trie_fields[i]);
737 if (field->flow_be32ofs < 0 || field->n_bits % 32) {
738 /* Incompatible field. This is the only place where we
739 * enforce these requirements, but the rest of the trie code
740 * depends on the flow_be32ofs to be non-negative and the
741 * field length to be a multiple of 32 bits. */
745 if (fields & (UINT64_C(1) << trie_fields[i])) {
746 /* Duplicate field, there is no need to build more than
747 * one index for any one field. */
750 fields |= UINT64_C(1) << trie_fields[i];
752 if (trie >= cls->n_tries || field != cls->tries[trie].field) {
753 trie_init(cls, trie, field);
758 /* Destroy the rest. */
759 for (i = trie; i < cls->n_tries; i++) {
760 trie_init(cls, i, NULL);
766 trie_init(struct cls_classifier *cls, int trie_idx,
767 const struct mf_field *field)
769 struct cls_trie *trie = &cls->tries[trie_idx];
770 struct cls_subtable *subtable;
771 struct cls_subtable_entry *iter;
773 if (trie_idx < cls->n_tries) {
774 trie_destroy(trie->root);
779 /* Add existing rules to the trie. */
780 CLS_SUBTABLE_CACHE_FOR_EACH (subtable, iter, &cls->subtables_priority) {
783 plen = field ? minimask_get_prefix_len(&subtable->mask, field) : 0;
784 /* Initialize subtable's prefix length on this field. */
785 subtable->trie_plen[trie_idx] = plen;
788 struct cls_match *head;
790 HMAP_FOR_EACH (head, hmap_node, &subtable->rules) {
791 struct cls_match *match;
793 FOR_EACH_RULE_IN_LIST (match, head) {
794 trie_insert(trie, match->cls_rule, plen);
801 /* Returns true if 'cls' contains no classification rules, false otherwise. */
803 classifier_is_empty(const struct classifier *cls)
805 return cls->cls->n_rules == 0;
808 /* Returns the number of rules in 'cls'. */
810 classifier_count(const struct classifier *cls)
812 return cls->cls->n_rules;
816 hash_metadata(ovs_be64 metadata_)
818 uint64_t metadata = (OVS_FORCE uint64_t) metadata_;
819 return hash_uint64(metadata);
822 static struct cls_partition *
823 find_partition(const struct cls_classifier *cls, ovs_be64 metadata,
826 struct cls_partition *partition;
828 HMAP_FOR_EACH_IN_BUCKET (partition, hmap_node, hash, &cls->partitions) {
829 if (partition->metadata == metadata) {
837 static struct cls_partition *
838 create_partition(struct cls_classifier *cls, struct cls_subtable *subtable,
841 uint32_t hash = hash_metadata(metadata);
842 struct cls_partition *partition = find_partition(cls, metadata, hash);
844 partition = xmalloc(sizeof *partition);
845 partition->metadata = metadata;
847 tag_tracker_init(&partition->tracker);
848 hmap_insert(&cls->partitions, &partition->hmap_node, hash);
850 tag_tracker_add(&partition->tracker, &partition->tags, subtable->tag);
854 static inline ovs_be32 minimatch_get_ports(const struct minimatch *match)
856 /* Could optimize to use the same map if needed for fast path. */
857 return MINIFLOW_GET_BE32(&match->flow, tp_src)
858 & MINIFLOW_GET_BE32(&match->mask.masks, tp_src);
861 /* Inserts 'rule' into 'cls'. Until 'rule' is removed from 'cls', the caller
862 * must not modify or free it.
864 * If 'cls' already contains an identical rule (including wildcards, values of
865 * fixed fields, and priority), replaces the old rule by 'rule' and returns the
866 * rule that was replaced. The caller takes ownership of the returned rule and
867 * is thus responsible for destroying it with cls_rule_destroy(), freeing the
868 * memory block in which it resides, etc., as necessary.
870 * Returns NULL if 'cls' does not contain a rule with an identical key, after
871 * inserting the new rule. In this case, no rules are displaced by the new
872 * rule, even rules that cannot have any effect because the new rule matches a
873 * superset of their flows and has higher priority. */
875 classifier_replace(struct classifier *cls_, struct cls_rule *rule)
877 struct cls_classifier *cls = cls_->cls;
878 struct cls_match *old_rule;
879 struct cls_subtable *subtable;
881 subtable = find_subtable(cls, &rule->match.mask);
883 subtable = insert_subtable(cls, &rule->match.mask);
886 old_rule = insert_rule(cls, subtable, rule);
890 rule->cls_match->partition = NULL;
891 if (minimask_get_metadata_mask(&rule->match.mask) == OVS_BE64_MAX) {
892 ovs_be64 metadata = miniflow_get_metadata(&rule->match.flow);
893 rule->cls_match->partition = create_partition(cls, subtable,
900 for (i = 0; i < cls->n_tries; i++) {
901 if (subtable->trie_plen[i]) {
902 trie_insert(&cls->tries[i], rule, subtable->trie_plen[i]);
907 if (subtable->ports_mask_len) {
908 /* We mask the value to be inserted to always have the wildcarded
909 * bits in known (zero) state, so we can include them in comparison
910 * and they will always match (== their original value does not
912 ovs_be32 masked_ports = minimatch_get_ports(&rule->match);
914 trie_insert_prefix(&subtable->ports_trie, &masked_ports,
915 subtable->ports_mask_len);
920 struct cls_rule *old_cls_rule = old_rule->cls_rule;
922 rule->cls_match->partition = old_rule->partition;
923 old_cls_rule->cls_match = NULL;
929 /* Inserts 'rule' into 'cls'. Until 'rule' is removed from 'cls', the caller
930 * must not modify or free it.
932 * 'cls' must not contain an identical rule (including wildcards, values of
933 * fixed fields, and priority). Use classifier_find_rule_exactly() to find
936 classifier_insert(struct classifier *cls, struct cls_rule *rule)
938 struct cls_rule *displaced_rule = classifier_replace(cls, rule);
939 ovs_assert(!displaced_rule);
942 /* Removes 'rule' from 'cls'. It is the caller's responsibility to destroy
943 * 'rule' with cls_rule_destroy(), freeing the memory block in which 'rule'
944 * resides, etc., as necessary. */
946 classifier_remove(struct classifier *cls_, struct cls_rule *rule)
948 struct cls_classifier *cls = cls_->cls;
949 struct cls_partition *partition;
950 struct cls_match *cls_match = rule->cls_match;
951 struct cls_match *head;
952 struct cls_subtable *subtable;
955 ovs_assert(cls_match);
957 subtable = find_subtable(cls, &rule->match.mask);
958 ovs_assert(subtable);
960 if (subtable->ports_mask_len) {
961 ovs_be32 masked_ports = minimatch_get_ports(&rule->match);
963 trie_remove_prefix(&subtable->ports_trie,
964 &masked_ports, subtable->ports_mask_len);
966 for (i = 0; i < cls->n_tries; i++) {
967 if (subtable->trie_plen[i]) {
968 trie_remove(&cls->tries[i], rule, subtable->trie_plen[i]);
972 /* Remove rule node from indices. */
973 for (i = 0; i < subtable->n_indices; i++) {
974 hindex_remove(&subtable->indices[i], &cls_match->index_nodes[i]);
977 head = find_equal(subtable, &rule->match.flow, cls_match->hmap_node.hash);
978 if (head != cls_match) {
979 list_remove(&cls_match->list);
980 } else if (list_is_empty(&cls_match->list)) {
981 hmap_remove(&subtable->rules, &cls_match->hmap_node);
983 struct cls_match *next = CONTAINER_OF(cls_match->list.next,
984 struct cls_match, list);
986 list_remove(&cls_match->list);
987 hmap_replace(&subtable->rules, &cls_match->hmap_node,
991 partition = cls_match->partition;
993 tag_tracker_subtract(&partition->tracker, &partition->tags,
995 if (!partition->tags) {
996 hmap_remove(&cls->partitions, &partition->hmap_node);
1001 if (--subtable->n_rules == 0) {
1002 destroy_subtable(cls, subtable);
1004 update_subtables_after_removal(cls, subtable, cls_match->priority);
1009 rule->cls_match = NULL;
1013 /* Prefix tree context. Valid when 'lookup_done' is true. Can skip all
1014 * subtables which have more than 'match_plen' bits in their corresponding
1015 * field at offset 'be32ofs'. If skipped, 'maskbits' prefix bits should be
1016 * unwildcarded to quarantee datapath flow matches only packets it should. */
1018 const struct cls_trie *trie;
1019 bool lookup_done; /* Status of the lookup. */
1020 uint8_t be32ofs; /* U32 offset of the field in question. */
1021 unsigned int match_plen; /* Longest prefix than could possibly match. */
1022 unsigned int maskbits; /* Prefix length needed to avoid false matches. */
1026 trie_ctx_init(struct trie_ctx *ctx, const struct cls_trie *trie)
1029 ctx->be32ofs = trie->field->flow_be32ofs;
1030 ctx->lookup_done = false;
1034 lookahead_subtable(const struct cls_subtable_entry *subtables)
1036 ovs_prefetch_range(subtables->subtable, sizeof *subtables->subtable);
1039 /* Finds and returns the highest-priority rule in 'cls' that matches 'flow'.
1040 * Returns a null pointer if no rules in 'cls' match 'flow'. If multiple rules
1041 * of equal priority match 'flow', returns one arbitrarily.
1043 * If a rule is found and 'wc' is non-null, bitwise-OR's 'wc' with the
1044 * set of bits that were significant in the lookup. At some point
1045 * earlier, 'wc' should have been initialized (e.g., by
1046 * flow_wildcards_init_catchall()). */
1048 classifier_lookup(const struct classifier *cls_, const struct flow *flow,
1049 struct flow_wildcards *wc)
1051 struct cls_classifier *cls = cls_->cls;
1052 const struct cls_partition *partition;
1054 struct cls_match *best;
1055 struct trie_ctx trie_ctx[CLS_MAX_TRIES];
1057 struct cls_subtable_entry *subtables = cls->subtables_priority.subtables;
1058 int n_subtables = cls->subtables_priority.size;
1059 int64_t best_priority = -1;
1061 /* Prefetch the subtables array. */
1062 ovs_prefetch_range(subtables, n_subtables * sizeof *subtables);
1064 /* Determine 'tags' such that, if 'subtable->tag' doesn't intersect them,
1065 * then 'flow' cannot possibly match in 'subtable':
1067 * - If flow->metadata maps to a given 'partition', then we can use
1068 * 'tags' for 'partition->tags'.
1070 * - If flow->metadata has no partition, then no rule in 'cls' has an
1071 * exact-match for flow->metadata. That means that we don't need to
1072 * search any subtable that includes flow->metadata in its mask.
1074 * In either case, we always need to search any cls_subtables that do not
1075 * include flow->metadata in its mask. One way to do that would be to
1076 * check the "cls_subtable"s explicitly for that, but that would require an
1077 * extra branch per subtable. Instead, we mark such a cls_subtable's
1078 * 'tags' as TAG_ALL and make sure that 'tags' is never empty. This means
1079 * that 'tags' always intersects such a cls_subtable's 'tags', so we don't
1080 * need a special case.
1082 partition = (hmap_is_empty(&cls->partitions)
1084 : find_partition(cls, flow->metadata,
1085 hash_metadata(flow->metadata)));
1086 tags = partition ? partition->tags : TAG_ARBITRARY;
1088 /* Initialize trie contexts for match_find_wc(). */
1089 for (i = 0; i < cls->n_tries; i++) {
1090 trie_ctx_init(&trie_ctx[i], &cls->tries[i]);
1093 /* Prefetch the first subtables. */
1094 if (n_subtables > 1) {
1095 lookahead_subtable(subtables);
1096 lookahead_subtable(subtables + 1);
1100 for (i = 0; OVS_LIKELY(i < n_subtables); i++) {
1101 struct cls_match *rule;
1103 if ((int64_t)subtables[i].max_priority <= best_priority) {
1104 /* Subtables are in descending priority order,
1105 * can not find anything better. */
1109 /* Prefetch a forthcoming subtable. */
1110 if (i + 2 < n_subtables) {
1111 lookahead_subtable(&subtables[i + 2]);
1114 if (!tag_intersects(tags, subtables[i].tag)) {
1118 rule = find_match_wc(subtables[i].subtable, flow, trie_ctx,
1120 if (rule && (int64_t)rule->priority > best_priority) {
1121 best_priority = (int64_t)rule->priority;
1126 return best ? best->cls_rule : NULL;
1129 /* Returns true if 'target' satisifies 'match', that is, if each bit for which
1130 * 'match' specifies a particular value has the correct value in 'target'.
1132 * 'flow' and 'mask' have the same mask! */
1134 miniflow_and_mask_matches_miniflow(const struct miniflow *flow,
1135 const struct minimask *mask,
1136 const struct miniflow *target)
1138 const uint32_t *flowp = miniflow_get_u32_values(flow);
1139 const uint32_t *maskp = miniflow_get_u32_values(&mask->masks);
1140 uint32_t target_u32;
1142 MINIFLOW_FOR_EACH_IN_MAP(target_u32, target, mask->masks.map) {
1143 if ((*flowp++ ^ target_u32) & *maskp++) {
1151 static inline struct cls_match *
1152 find_match_miniflow(const struct cls_subtable *subtable,
1153 const struct miniflow *flow,
1156 struct cls_match *rule;
1158 HMAP_FOR_EACH_WITH_HASH (rule, hmap_node, hash, &subtable->rules) {
1159 if (miniflow_and_mask_matches_miniflow(&rule->flow, &subtable->mask,
1168 /* Finds and returns the highest-priority rule in 'cls' that matches
1169 * 'miniflow'. Returns a null pointer if no rules in 'cls' match 'flow'.
1170 * If multiple rules of equal priority match 'flow', returns one arbitrarily.
1172 * This function is optimized for the userspace datapath, which only ever has
1173 * one priority value for it's flows!
1175 struct cls_rule *classifier_lookup_miniflow_first(const struct classifier *cls_,
1176 const struct miniflow *flow)
1178 struct cls_classifier *cls = cls_->cls;
1179 struct cls_subtable *subtable;
1180 struct cls_subtable_entry *iter;
1182 CLS_SUBTABLE_CACHE_FOR_EACH (subtable, iter, &cls->subtables_priority) {
1183 struct cls_match *rule;
1185 rule = find_match_miniflow(subtable, flow,
1186 miniflow_hash_in_minimask(flow,
1190 return rule->cls_rule;
1197 /* Finds and returns a rule in 'cls' with exactly the same priority and
1198 * matching criteria as 'target'. Returns a null pointer if 'cls' doesn't
1199 * contain an exact match. */
1201 classifier_find_rule_exactly(const struct classifier *cls_,
1202 const struct cls_rule *target)
1204 struct cls_classifier *cls = cls_->cls;
1205 struct cls_match *head, *rule;
1206 struct cls_subtable *subtable;
1208 subtable = find_subtable(cls, &target->match.mask);
1213 /* Skip if there is no hope. */
1214 if (target->priority > subtable->max_priority) {
1218 head = find_equal(subtable, &target->match.flow,
1219 miniflow_hash_in_minimask(&target->match.flow,
1220 &target->match.mask, 0));
1221 FOR_EACH_RULE_IN_LIST (rule, head) {
1222 if (target->priority >= rule->priority) {
1223 return target->priority == rule->priority ? rule->cls_rule : NULL;
1229 /* Finds and returns a rule in 'cls' with priority 'priority' and exactly the
1230 * same matching criteria as 'target'. Returns a null pointer if 'cls' doesn't
1231 * contain an exact match. */
1233 classifier_find_match_exactly(const struct classifier *cls,
1234 const struct match *target,
1235 unsigned int priority)
1237 struct cls_rule *retval;
1240 cls_rule_init(&cr, target, priority);
1241 retval = classifier_find_rule_exactly(cls, &cr);
1242 cls_rule_destroy(&cr);
1247 /* Checks if 'target' would overlap any other rule in 'cls'. Two rules are
1248 * considered to overlap if both rules have the same priority and a packet
1249 * could match both. */
1251 classifier_rule_overlaps(const struct classifier *cls_,
1252 const struct cls_rule *target)
1254 struct cls_classifier *cls = cls_->cls;
1255 struct cls_subtable *subtable;
1256 struct cls_subtable_entry *iter;
1258 /* Iterate subtables in the descending max priority order. */
1259 CLS_SUBTABLE_CACHE_FOR_EACH (subtable, iter, &cls->subtables_priority) {
1260 uint32_t storage[FLOW_U32S];
1261 struct minimask mask;
1262 struct cls_match *head;
1264 if (target->priority > iter->max_priority) {
1265 break; /* Can skip this and the rest of the subtables. */
1268 minimask_combine(&mask, &target->match.mask, &subtable->mask, storage);
1269 HMAP_FOR_EACH (head, hmap_node, &subtable->rules) {
1270 struct cls_match *rule;
1272 FOR_EACH_RULE_IN_LIST (rule, head) {
1273 if (rule->priority < target->priority) {
1274 break; /* Rules in descending priority order. */
1276 if (rule->priority == target->priority
1277 && miniflow_equal_in_minimask(&target->match.flow,
1278 &rule->flow, &mask)) {
1288 /* Returns true if 'rule' exactly matches 'criteria' or if 'rule' is more
1289 * specific than 'criteria'. That is, 'rule' matches 'criteria' and this
1290 * function returns true if, for every field:
1292 * - 'criteria' and 'rule' specify the same (non-wildcarded) value for the
1295 * - 'criteria' wildcards the field,
1297 * Conversely, 'rule' does not match 'criteria' and this function returns false
1298 * if, for at least one field:
1300 * - 'criteria' and 'rule' specify different values for the field, or
1302 * - 'criteria' specifies a value for the field but 'rule' wildcards it.
1304 * Equivalently, the truth table for whether a field matches is:
1309 * r +---------+---------+
1310 * i wild | yes | yes |
1312 * e +---------+---------+
1313 * r exact | no |if values|
1315 * a +---------+---------+
1317 * This is the matching rule used by OpenFlow 1.0 non-strict OFPT_FLOW_MOD
1318 * commands and by OpenFlow 1.0 aggregate and flow stats.
1320 * Ignores rule->priority. */
1322 cls_rule_is_loose_match(const struct cls_rule *rule,
1323 const struct minimatch *criteria)
1325 return (!minimask_has_extra(&rule->match.mask, &criteria->mask)
1326 && miniflow_equal_in_minimask(&rule->match.flow, &criteria->flow,
1333 rule_matches(const struct cls_match *rule, const struct cls_rule *target)
1336 || miniflow_equal_in_minimask(&rule->flow,
1337 &target->match.flow,
1338 &target->match.mask));
1341 static struct cls_match *
1342 search_subtable(const struct cls_subtable *subtable,
1343 const struct cls_rule *target)
1345 if (!target || !minimask_has_extra(&subtable->mask, &target->match.mask)) {
1346 struct cls_match *rule;
1348 HMAP_FOR_EACH (rule, hmap_node, &subtable->rules) {
1349 if (rule_matches(rule, target)) {
1357 /* Initializes 'cursor' for iterating through rules in 'cls':
1359 * - If 'target' is null, the cursor will visit every rule in 'cls'.
1361 * - If 'target' is nonnull, the cursor will visit each 'rule' in 'cls'
1362 * such that cls_rule_is_loose_match(rule, target) returns true.
1364 * Ignores target->priority. */
1366 cls_cursor_init(struct cls_cursor *cursor, const struct classifier *cls,
1367 const struct cls_rule *target)
1369 cursor->cls = cls->cls;
1370 cursor->target = target && !cls_rule_is_catchall(target) ? target : NULL;
1373 /* Returns the first matching cls_rule in 'cursor''s iteration, or a null
1374 * pointer if there are no matches. */
1376 cls_cursor_first(struct cls_cursor *cursor)
1378 struct cls_subtable *subtable;
1380 HMAP_FOR_EACH (subtable, hmap_node, &cursor->cls->subtables) {
1381 struct cls_match *rule = search_subtable(subtable, cursor->target);
1383 cursor->subtable = subtable;
1384 return rule->cls_rule;
1391 /* Returns the next matching cls_rule in 'cursor''s iteration, or a null
1392 * pointer if there are no more matches. */
1394 cls_cursor_next(struct cls_cursor *cursor, const struct cls_rule *rule_)
1396 struct cls_match *rule = CONST_CAST(struct cls_match *, rule_->cls_match);
1397 const struct cls_subtable *subtable;
1398 struct cls_match *next;
1400 next = next_rule_in_list__(rule);
1401 if (next->priority < rule->priority) {
1402 return next->cls_rule;
1405 /* 'next' is the head of the list, that is, the rule that is included in
1406 * the subtable's hmap. (This is important when the classifier contains
1407 * rules that differ only in priority.) */
1409 HMAP_FOR_EACH_CONTINUE (rule, hmap_node, &cursor->subtable->rules) {
1410 if (rule_matches(rule, cursor->target)) {
1411 return rule->cls_rule;
1415 subtable = cursor->subtable;
1416 HMAP_FOR_EACH_CONTINUE (subtable, hmap_node, &cursor->cls->subtables) {
1417 rule = search_subtable(subtable, cursor->target);
1419 cursor->subtable = subtable;
1420 return rule->cls_rule;
1427 static struct cls_subtable *
1428 find_subtable(const struct cls_classifier *cls, const struct minimask *mask)
1430 struct cls_subtable *subtable;
1432 HMAP_FOR_EACH_IN_BUCKET (subtable, hmap_node, minimask_hash(mask, 0),
1434 if (minimask_equal(mask, &subtable->mask)) {
1441 static struct cls_subtable *
1442 insert_subtable(struct cls_classifier *cls, const struct minimask *mask)
1444 uint32_t hash = minimask_hash(mask, 0);
1445 struct cls_subtable *subtable;
1447 struct flow_wildcards old, new;
1449 struct cls_subtable_entry elem;
1450 int count = count_1bits(mask->masks.map);
1452 subtable = xzalloc(sizeof *subtable - sizeof mask->masks.inline_values
1453 + MINIFLOW_VALUES_SIZE(count));
1454 hmap_init(&subtable->rules);
1455 miniflow_clone_inline(&subtable->mask.masks, &mask->masks, count);
1457 /* Init indices for segmented lookup, if any. */
1458 flow_wildcards_init_catchall(&new);
1461 for (i = 0; i < cls->n_flow_segments; i++) {
1462 flow_wildcards_fold_minimask_range(&new, mask, prev,
1463 cls->flow_segments[i]);
1464 /* Add an index if it adds mask bits. */
1465 if (!flow_wildcards_equal(&new, &old)) {
1466 hindex_init(&subtable->indices[index]);
1467 subtable->index_ofs[index] = cls->flow_segments[i];
1471 prev = cls->flow_segments[i];
1473 /* Check if the rest of the subtable's mask adds any bits,
1474 * and remove the last index if it doesn't. */
1476 flow_wildcards_fold_minimask_range(&new, mask, prev, FLOW_U32S);
1477 if (flow_wildcards_equal(&new, &old)) {
1479 subtable->index_ofs[index] = 0;
1480 hindex_destroy(&subtable->indices[index]);
1483 subtable->n_indices = index;
1485 subtable->tag = (minimask_get_metadata_mask(mask) == OVS_BE64_MAX
1486 ? tag_create_deterministic(hash)
1489 for (i = 0; i < cls->n_tries; i++) {
1490 subtable->trie_plen[i] = minimask_get_prefix_len(mask,
1491 cls->tries[i].field);
1495 subtable->ports_trie = NULL;
1496 subtable->ports_mask_len
1497 = 32 - ctz32(ntohl(MINIFLOW_GET_BE32(&mask->masks, tp_src)));
1499 hmap_insert(&cls->subtables, &subtable->hmap_node, hash);
1500 elem.subtable = subtable;
1501 elem.tag = subtable->tag;
1502 elem.max_priority = subtable->max_priority;
1503 cls_subtable_cache_push_back(&cls->subtables_priority, elem);
1509 destroy_subtable(struct cls_classifier *cls, struct cls_subtable *subtable)
1512 struct cls_subtable *table = NULL;
1513 struct cls_subtable_entry *iter;
1515 CLS_SUBTABLE_CACHE_FOR_EACH (table, iter, &cls->subtables_priority) {
1516 if (table == subtable) {
1517 cls_subtable_cache_remove(&cls->subtables_priority, iter);
1522 trie_destroy(subtable->ports_trie);
1524 for (i = 0; i < subtable->n_indices; i++) {
1525 hindex_destroy(&subtable->indices[i]);
1527 minimask_destroy(&subtable->mask);
1528 hmap_remove(&cls->subtables, &subtable->hmap_node);
1529 hmap_destroy(&subtable->rules);
1533 /* This function performs the following updates for 'subtable' in 'cls'
1534 * following the addition of a new rule with priority 'new_priority' to
1537 * - Update 'subtable->max_priority' and 'subtable->max_count' if necessary.
1539 * - Update 'subtable''s position in 'cls->subtables_priority' if necessary.
1541 * This function should only be called after adding a new rule, not after
1542 * replacing a rule by an identical one or modifying a rule in-place. */
1544 update_subtables_after_insertion(struct cls_classifier *cls,
1545 struct cls_subtable *subtable,
1546 unsigned int new_priority)
1548 if (new_priority == subtable->max_priority) {
1549 ++subtable->max_count;
1550 } else if (new_priority > subtable->max_priority) {
1551 struct cls_subtable *table;
1552 struct cls_subtable_entry *iter, *from = NULL;
1554 subtable->max_priority = new_priority;
1555 subtable->max_count = 1;
1557 /* Possibly move 'subtable' earlier in the priority array. If
1558 * we break out of the loop, then the subtable (at 'from')
1559 * should be moved to the position right after the current
1560 * element. If the loop terminates normally, then 'iter' will
1561 * be at the first array element and we'll move the subtable
1562 * to the front of the array. */
1563 CLS_SUBTABLE_CACHE_FOR_EACH_REVERSE (table, iter,
1564 &cls->subtables_priority) {
1565 if (table == subtable) {
1566 from = iter; /* Locate the subtable as we go. */
1567 iter->max_priority = new_priority;
1568 } else if (table->max_priority >= new_priority) {
1570 /* Corrupted cache? */
1571 cls_subtable_cache_reset(cls);
1572 VLOG_ABORT("update_subtables_after_insertion(): Subtable priority list corrupted.");
1575 iter++; /* After this. */
1580 /* Move subtable at 'from' to 'iter'. */
1581 cls_subtable_cache_move(iter, from);
1585 /* This function performs the following updates for 'subtable' in 'cls'
1586 * following the deletion of a rule with priority 'del_priority' from
1589 * - Update 'subtable->max_priority' and 'subtable->max_count' if necessary.
1591 * - Update 'subtable''s position in 'cls->subtables_priority' if necessary.
1593 * This function should only be called after removing a rule, not after
1594 * replacing a rule by an identical one or modifying a rule in-place. */
1596 update_subtables_after_removal(struct cls_classifier *cls,
1597 struct cls_subtable *subtable,
1598 unsigned int del_priority)
1600 if (del_priority == subtable->max_priority && --subtable->max_count == 0) {
1601 struct cls_match *head;
1602 struct cls_subtable *table;
1603 struct cls_subtable_entry *iter, *from = NULL;
1605 subtable->max_priority = 0;
1606 HMAP_FOR_EACH (head, hmap_node, &subtable->rules) {
1607 if (head->priority > subtable->max_priority) {
1608 subtable->max_priority = head->priority;
1609 subtable->max_count = 1;
1610 } else if (head->priority == subtable->max_priority) {
1611 ++subtable->max_count;
1615 /* Possibly move 'subtable' later in the priority array.
1616 * After the loop the 'iter' will point right after the position
1617 * at which the subtable should be moved (either at a subtable
1618 * with an equal or lower priority, or just past the array),
1619 * so it is decremented once. */
1620 CLS_SUBTABLE_CACHE_FOR_EACH (table, iter, &cls->subtables_priority) {
1621 if (table == subtable) {
1622 from = iter; /* Locate the subtable as we go. */
1623 iter->max_priority = subtable->max_priority;
1624 } else if (table->max_priority <= subtable->max_priority) {
1626 /* Corrupted cache? */
1627 cls_subtable_cache_reset(cls);
1628 VLOG_ABORT("update_subtables_after_removal(): Subtable priority list corrupted.");
1634 /* Now at one past the destination. */
1637 /* Move subtable at 'from' to 'iter'. */
1638 cls_subtable_cache_move(iter, from);
1647 /* Return 'true' if can skip rest of the subtable based on the prefix trie
1648 * lookup results. */
1650 check_tries(struct trie_ctx trie_ctx[CLS_MAX_TRIES], unsigned int n_tries,
1651 const unsigned int field_plen[CLS_MAX_TRIES],
1652 const struct range ofs, const struct flow *flow,
1653 struct flow_wildcards *wc)
1657 /* Check if we could avoid fully unwildcarding the next level of
1658 * fields using the prefix tries. The trie checks are done only as
1659 * needed to avoid folding in additional bits to the wildcards mask. */
1660 for (j = 0; j < n_tries; j++) {
1661 /* Is the trie field relevant for this subtable? */
1662 if (field_plen[j]) {
1663 struct trie_ctx *ctx = &trie_ctx[j];
1664 uint8_t be32ofs = ctx->be32ofs;
1666 /* Is the trie field within the current range of fields? */
1667 if (be32ofs >= ofs.start && be32ofs < ofs.end) {
1668 /* On-demand trie lookup. */
1669 if (!ctx->lookup_done) {
1670 ctx->match_plen = trie_lookup(ctx->trie, flow,
1672 ctx->lookup_done = true;
1674 /* Possible to skip the rest of the subtable if subtable's
1675 * prefix on the field is longer than what is known to match
1676 * based on the trie lookup. */
1677 if (field_plen[j] > ctx->match_plen) {
1678 /* RFC: We want the trie lookup to never result in
1679 * unwildcarding any bits that would not be unwildcarded
1680 * otherwise. Since the trie is shared by the whole
1681 * classifier, it is possible that the 'maskbits' contain
1682 * bits that are irrelevant for the partition of the
1683 * classifier relevant for the current flow. */
1685 /* Can skip if the field is already unwildcarded. */
1686 if (mask_prefix_bits_set(wc, be32ofs, ctx->maskbits)) {
1689 /* Check that the trie result will not unwildcard more bits
1690 * than this stage will. */
1691 if (ctx->maskbits <= field_plen[j]) {
1692 /* Unwildcard the bits and skip the rest. */
1693 mask_set_prefix_bits(wc, be32ofs, ctx->maskbits);
1694 /* Note: Prerequisite already unwildcarded, as the only
1695 * prerequisite of the supported trie lookup fields is
1696 * the ethertype, which is currently always
1708 /* Returns true if 'target' satisifies 'flow'/'mask', that is, if each bit
1709 * for which 'flow', for which 'mask' has a bit set, specifies a particular
1710 * value has the correct value in 'target'.
1712 * This function is equivalent to miniflow_equal_flow_in_minimask(flow,
1713 * target, mask) but it is faster because of the invariant that
1714 * flow->map and mask->masks.map are the same. */
1716 miniflow_and_mask_matches_flow(const struct miniflow *flow,
1717 const struct minimask *mask,
1718 const struct flow *target)
1720 const uint32_t *flowp = miniflow_get_u32_values(flow);
1721 const uint32_t *maskp = miniflow_get_u32_values(&mask->masks);
1722 uint32_t target_u32;
1724 FLOW_FOR_EACH_IN_MAP(target_u32, target, mask->masks.map) {
1725 if ((*flowp++ ^ target_u32) & *maskp++) {
1733 static inline struct cls_match *
1734 find_match(const struct cls_subtable *subtable, const struct flow *flow,
1737 struct cls_match *rule;
1739 HMAP_FOR_EACH_WITH_HASH (rule, hmap_node, hash, &subtable->rules) {
1740 if (miniflow_and_mask_matches_flow(&rule->flow, &subtable->mask,
1749 static struct cls_match *
1750 find_match_wc(const struct cls_subtable *subtable, const struct flow *flow,
1751 struct trie_ctx trie_ctx[CLS_MAX_TRIES], unsigned int n_tries,
1752 struct flow_wildcards *wc)
1754 uint32_t basis = 0, hash;
1755 struct cls_match *rule = NULL;
1759 if (OVS_UNLIKELY(!wc)) {
1760 return find_match(subtable, flow,
1761 flow_hash_in_minimask(flow, &subtable->mask, 0));
1765 /* Try to finish early by checking fields in segments. */
1766 for (i = 0; i < subtable->n_indices; i++) {
1767 struct hindex_node *inode;
1768 ofs.end = subtable->index_ofs[i];
1770 if (check_tries(trie_ctx, n_tries, subtable->trie_plen, ofs, flow,
1774 hash = flow_hash_in_minimask_range(flow, &subtable->mask, ofs.start,
1776 ofs.start = ofs.end;
1777 inode = hindex_node_with_hash(&subtable->indices[i], hash);
1779 /* No match, can stop immediately, but must fold in the mask
1780 * covered so far. */
1784 /* If we have narrowed down to a single rule already, check whether
1785 * that rule matches. If it does match, then we're done. If it does
1786 * not match, then we know that we will never get a match, but we do
1787 * not yet know how many wildcards we need to fold into 'wc' so we
1788 * continue iterating through indices to find that out. (We won't
1789 * waste time calling miniflow_and_mask_matches_flow() again because
1790 * we've set 'rule' nonnull.)
1792 * This check shows a measurable benefit with non-trivial flow tables.
1794 * (Rare) hash collisions may cause us to miss the opportunity for this
1796 if (!inode->s && !rule) {
1797 ASSIGN_CONTAINER(rule, inode - i, index_nodes);
1798 if (miniflow_and_mask_matches_flow(&rule->flow, &subtable->mask,
1804 ofs.end = FLOW_U32S;
1805 /* Trie check for the final range. */
1806 if (check_tries(trie_ctx, n_tries, subtable->trie_plen, ofs, flow, wc)) {
1810 /* Multiple potential matches exist, look for one. */
1811 hash = flow_hash_in_minimask_range(flow, &subtable->mask, ofs.start,
1813 rule = find_match(subtable, flow, hash);
1815 /* We already narrowed the matching candidates down to just 'rule',
1816 * but it didn't match. */
1819 if (!rule && subtable->ports_mask_len) {
1820 /* Ports are always part of the final range, if any.
1821 * No match was found for the ports. Use the ports trie to figure out
1822 * which ports bits to unwildcard. */
1824 ovs_be32 value, mask;
1826 mask = MINIFLOW_GET_BE32(&subtable->mask.masks, tp_src);
1827 value = ((OVS_FORCE ovs_be32 *)flow)[TP_PORTS_OFS32] & mask;
1828 trie_lookup_value(subtable->ports_trie, &value, &mbits);
1830 ((OVS_FORCE ovs_be32 *)&wc->masks)[TP_PORTS_OFS32] |=
1831 mask & htonl(~0 << (32 - mbits));
1833 ofs.start = TP_PORTS_OFS32;
1837 /* Must unwildcard all the fields, as they were looked at. */
1838 flow_wildcards_fold_minimask(wc, &subtable->mask);
1842 /* Must unwildcard the fields looked up so far, if any. */
1844 flow_wildcards_fold_minimask_range(wc, &subtable->mask, 0, ofs.start);
1849 static struct cls_match *
1850 find_equal(struct cls_subtable *subtable, const struct miniflow *flow,
1853 struct cls_match *head;
1855 HMAP_FOR_EACH_WITH_HASH (head, hmap_node, hash, &subtable->rules) {
1856 if (miniflow_equal(&head->flow, flow)) {
1863 static struct cls_match *
1864 insert_rule(struct cls_classifier *cls, struct cls_subtable *subtable,
1865 struct cls_rule *new)
1867 struct cls_match *cls_match = cls_match_alloc(new);
1868 struct cls_match *head;
1869 struct cls_match *old = NULL;
1871 uint32_t basis = 0, hash;
1872 uint8_t prev_be32ofs = 0;
1874 /* Add new node to segment indices. */
1875 for (i = 0; i < subtable->n_indices; i++) {
1876 hash = minimatch_hash_range(&new->match, prev_be32ofs,
1877 subtable->index_ofs[i], &basis);
1878 hindex_insert(&subtable->indices[i], &cls_match->index_nodes[i], hash);
1879 prev_be32ofs = subtable->index_ofs[i];
1881 hash = minimatch_hash_range(&new->match, prev_be32ofs, FLOW_U32S, &basis);
1882 head = find_equal(subtable, &new->match.flow, hash);
1884 hmap_insert(&subtable->rules, &cls_match->hmap_node, hash);
1885 list_init(&cls_match->list);
1888 /* Scan the list for the insertion point that will keep the list in
1889 * order of decreasing priority. */
1890 struct cls_match *rule;
1892 cls_match->hmap_node.hash = hash; /* Otherwise done by hmap_insert. */
1894 FOR_EACH_RULE_IN_LIST (rule, head) {
1895 if (cls_match->priority >= rule->priority) {
1897 /* 'new' is the new highest-priority flow in the list. */
1898 hmap_replace(&subtable->rules,
1899 &rule->hmap_node, &cls_match->hmap_node);
1902 if (cls_match->priority == rule->priority) {
1903 list_replace(&cls_match->list, &rule->list);
1907 list_insert(&rule->list, &cls_match->list);
1913 /* Insert 'new' at the end of the list. */
1914 list_push_back(&head->list, &cls_match->list);
1919 update_subtables_after_insertion(cls, subtable, cls_match->priority);
1921 /* Remove old node from indices. */
1922 for (i = 0; i < subtable->n_indices; i++) {
1923 hindex_remove(&subtable->indices[i], &old->index_nodes[i]);
1929 static struct cls_match *
1930 next_rule_in_list__(struct cls_match *rule)
1932 struct cls_match *next = OBJECT_CONTAINING(rule->list.next, next, list);
1936 static struct cls_match *
1937 next_rule_in_list(struct cls_match *rule)
1939 struct cls_match *next = next_rule_in_list__(rule);
1940 return next->priority < rule->priority ? next : NULL;
1943 /* A longest-prefix match tree. */
1945 uint32_t prefix; /* Prefix bits for this node, MSB first. */
1946 uint8_t nbits; /* Never zero, except for the root node. */
1947 unsigned int n_rules; /* Number of rules that have this prefix. */
1948 struct trie_node *edges[2]; /* Both NULL if leaf. */
1951 /* Max bits per node. Must fit in struct trie_node's 'prefix'.
1952 * Also tested with 16, 8, and 5 to stress the implementation. */
1953 #define TRIE_PREFIX_BITS 32
1955 /* Return at least 'plen' bits of the 'prefix', starting at bit offset 'ofs'.
1956 * Prefixes are in the network byte order, and the offset 0 corresponds to
1957 * the most significant bit of the first byte. The offset can be read as
1958 * "how many bits to skip from the start of the prefix starting at 'pr'". */
1960 raw_get_prefix(const ovs_be32 pr[], unsigned int ofs, unsigned int plen)
1964 pr += ofs / 32; /* Where to start. */
1965 ofs %= 32; /* How many bits to skip at 'pr'. */
1967 prefix = ntohl(*pr) << ofs; /* Get the first 32 - ofs bits. */
1968 if (plen > 32 - ofs) { /* Need more than we have already? */
1969 prefix |= ntohl(*++pr) >> (32 - ofs);
1971 /* Return with possible unwanted bits at the end. */
1975 /* Return min(TRIE_PREFIX_BITS, plen) bits of the 'prefix', starting at bit
1976 * offset 'ofs'. Prefixes are in the network byte order, and the offset 0
1977 * corresponds to the most significant bit of the first byte. The offset can
1978 * be read as "how many bits to skip from the start of the prefix starting at
1981 trie_get_prefix(const ovs_be32 pr[], unsigned int ofs, unsigned int plen)
1986 if (plen > TRIE_PREFIX_BITS) {
1987 plen = TRIE_PREFIX_BITS; /* Get at most TRIE_PREFIX_BITS. */
1989 /* Return with unwanted bits cleared. */
1990 return raw_get_prefix(pr, ofs, plen) & ~0u << (32 - plen);
1993 /* Return the number of equal bits in 'nbits' of 'prefix's MSBs and a 'value'
1994 * starting at "MSB 0"-based offset 'ofs'. */
1996 prefix_equal_bits(uint32_t prefix, unsigned int nbits, const ovs_be32 value[],
1999 uint64_t diff = prefix ^ raw_get_prefix(value, ofs, nbits);
2000 /* Set the bit after the relevant bits to limit the result. */
2001 return raw_clz64(diff << 32 | UINT64_C(1) << (63 - nbits));
2004 /* Return the number of equal bits in 'node' prefix and a 'prefix' of length
2005 * 'plen', starting at "MSB 0"-based offset 'ofs'. */
2007 trie_prefix_equal_bits(const struct trie_node *node, const ovs_be32 prefix[],
2008 unsigned int ofs, unsigned int plen)
2010 return prefix_equal_bits(node->prefix, MIN(node->nbits, plen - ofs),
2014 /* Return the bit at ("MSB 0"-based) offset 'ofs' as an int. 'ofs' can
2015 * be greater than 31. */
2017 be_get_bit_at(const ovs_be32 value[], unsigned int ofs)
2019 return (((const uint8_t *)value)[ofs / 8] >> (7 - ofs % 8)) & 1u;
2022 /* Return the bit at ("MSB 0"-based) offset 'ofs' as an int. 'ofs' must
2023 * be between 0 and 31, inclusive. */
2025 get_bit_at(const uint32_t prefix, unsigned int ofs)
2027 return (prefix >> (31 - ofs)) & 1u;
2030 /* Create new branch. */
2031 static struct trie_node *
2032 trie_branch_create(const ovs_be32 *prefix, unsigned int ofs, unsigned int plen,
2033 unsigned int n_rules)
2035 struct trie_node *node = xmalloc(sizeof *node);
2037 node->prefix = trie_get_prefix(prefix, ofs, plen);
2039 if (plen <= TRIE_PREFIX_BITS) {
2041 node->edges[0] = NULL;
2042 node->edges[1] = NULL;
2043 node->n_rules = n_rules;
2044 } else { /* Need intermediate nodes. */
2045 struct trie_node *subnode = trie_branch_create(prefix,
2046 ofs + TRIE_PREFIX_BITS,
2047 plen - TRIE_PREFIX_BITS,
2049 int bit = get_bit_at(subnode->prefix, 0);
2050 node->nbits = TRIE_PREFIX_BITS;
2051 node->edges[bit] = subnode;
2052 node->edges[!bit] = NULL;
2059 trie_node_destroy(struct trie_node *node)
2065 trie_destroy(struct trie_node *node)
2068 trie_destroy(node->edges[0]);
2069 trie_destroy(node->edges[1]);
2075 trie_is_leaf(const struct trie_node *trie)
2077 return !trie->edges[0] && !trie->edges[1]; /* No children. */
2081 mask_set_prefix_bits(struct flow_wildcards *wc, uint8_t be32ofs,
2084 ovs_be32 *mask = &((ovs_be32 *)&wc->masks)[be32ofs];
2087 for (i = 0; i < nbits / 32; i++) {
2088 mask[i] = OVS_BE32_MAX;
2091 mask[i] |= htonl(~0u << (32 - nbits % 32));
2096 mask_prefix_bits_set(const struct flow_wildcards *wc, uint8_t be32ofs,
2099 ovs_be32 *mask = &((ovs_be32 *)&wc->masks)[be32ofs];
2101 ovs_be32 zeroes = 0;
2103 for (i = 0; i < nbits / 32; i++) {
2107 zeroes |= ~mask[i] & htonl(~0u << (32 - nbits % 32));
2110 return !zeroes; /* All 'nbits' bits set. */
2113 static struct trie_node **
2114 trie_next_edge(struct trie_node *node, const ovs_be32 value[],
2117 return node->edges + be_get_bit_at(value, ofs);
2120 static const struct trie_node *
2121 trie_next_node(const struct trie_node *node, const ovs_be32 value[],
2124 return node->edges[be_get_bit_at(value, ofs)];
2127 /* Return the prefix mask length necessary to find the longest-prefix match for
2128 * the '*value' in the prefix tree 'node'.
2129 * '*checkbits' is set to the number of bits in the prefix mask necessary to
2130 * determine a mismatch, in case there are longer prefixes in the tree below
2131 * the one that matched.
2134 trie_lookup_value(const struct trie_node *node, const ovs_be32 value[],
2135 unsigned int *checkbits)
2137 unsigned int ofs = 0, match_len = 0;
2138 const struct trie_node *prev = NULL;
2140 for (; node; prev = node, node = trie_next_node(node, value, ofs)) {
2141 unsigned int eqbits;
2142 /* Check if this edge can be followed. */
2143 eqbits = prefix_equal_bits(node->prefix, node->nbits, value, ofs);
2145 if (eqbits < node->nbits) { /* Mismatch, nothing more to be found. */
2146 /* Bit at offset 'ofs' differed. */
2147 *checkbits = ofs + 1; /* Includes the first mismatching bit. */
2150 /* Full match, check if rules exist at this prefix length. */
2151 if (node->n_rules > 0) {
2155 /* Dead end, exclude the other branch if it exists. */
2156 *checkbits = !prev || trie_is_leaf(prev) ? ofs : ofs + 1;
2161 trie_lookup(const struct cls_trie *trie, const struct flow *flow,
2162 unsigned int *checkbits)
2164 const struct mf_field *mf = trie->field;
2166 /* Check that current flow matches the prerequisites for the trie
2167 * field. Some match fields are used for multiple purposes, so we
2168 * must check that the trie is relevant for this flow. */
2169 if (mf_are_prereqs_ok(mf, flow)) {
2170 return trie_lookup_value(trie->root,
2171 &((ovs_be32 *)flow)[mf->flow_be32ofs],
2174 *checkbits = 0; /* Value not used in this case. */
2178 /* Returns the length of a prefix match mask for the field 'mf' in 'minimask'.
2179 * Returns the u32 offset to the miniflow data in '*miniflow_index', if
2180 * 'miniflow_index' is not NULL. */
2182 minimask_get_prefix_len(const struct minimask *minimask,
2183 const struct mf_field *mf)
2185 unsigned int nbits = 0, mask_tz = 0; /* Non-zero when end of mask seen. */
2186 uint8_t u32_ofs = mf->flow_be32ofs;
2187 uint8_t u32_end = u32_ofs + mf->n_bytes / 4;
2189 for (; u32_ofs < u32_end; ++u32_ofs) {
2191 mask = ntohl((OVS_FORCE ovs_be32)minimask_get(minimask, u32_ofs));
2193 /* Validate mask, count the mask length. */
2196 return 0; /* No bits allowed after mask ended. */
2199 if (~mask & (~mask + 1)) {
2200 return 0; /* Mask not contiguous. */
2202 mask_tz = ctz32(mask);
2203 nbits += 32 - mask_tz;
2211 * This is called only when mask prefix is known to be CIDR and non-zero.
2212 * Relies on the fact that the flow and mask have the same map, and since
2213 * the mask is CIDR, the storage for the flow field exists even if it
2214 * happened to be zeros.
2216 static const ovs_be32 *
2217 minimatch_get_prefix(const struct minimatch *match, const struct mf_field *mf)
2219 return miniflow_get_be32_values(&match->flow) +
2220 count_1bits(match->flow.map & ((UINT64_C(1) << mf->flow_be32ofs) - 1));
2223 /* Insert rule in to the prefix tree.
2224 * 'mlen' must be the (non-zero) CIDR prefix length of the 'trie->field' mask
2227 trie_insert(struct cls_trie *trie, const struct cls_rule *rule, int mlen)
2229 trie_insert_prefix(&trie->root,
2230 minimatch_get_prefix(&rule->match, trie->field), mlen);
2234 trie_insert_prefix(struct trie_node **edge, const ovs_be32 *prefix, int mlen)
2236 struct trie_node *node;
2239 /* Walk the tree. */
2240 for (; (node = *edge) != NULL;
2241 edge = trie_next_edge(node, prefix, ofs)) {
2242 unsigned int eqbits = trie_prefix_equal_bits(node, prefix, ofs, mlen);
2244 if (eqbits < node->nbits) {
2245 /* Mismatch, new node needs to be inserted above. */
2246 int old_branch = get_bit_at(node->prefix, eqbits);
2248 /* New parent node. */
2249 *edge = trie_branch_create(prefix, ofs - eqbits, eqbits,
2250 ofs == mlen ? 1 : 0);
2252 /* Adjust old node for its new position in the tree. */
2253 node->prefix <<= eqbits;
2254 node->nbits -= eqbits;
2255 (*edge)->edges[old_branch] = node;
2257 /* Check if need a new branch for the new rule. */
2259 (*edge)->edges[!old_branch]
2260 = trie_branch_create(prefix, ofs, mlen - ofs, 1);
2264 /* Full match so far. */
2267 /* Full match at the current node, rule needs to be added here. */
2272 /* Must insert a new tree branch for the new rule. */
2273 *edge = trie_branch_create(prefix, ofs, mlen - ofs, 1);
2276 /* 'mlen' must be the (non-zero) CIDR prefix length of the 'trie->field' mask
2279 trie_remove(struct cls_trie *trie, const struct cls_rule *rule, int mlen)
2281 trie_remove_prefix(&trie->root,
2282 minimatch_get_prefix(&rule->match, trie->field), mlen);
2285 /* 'mlen' must be the (non-zero) CIDR prefix length of the 'trie->field' mask
2288 trie_remove_prefix(struct trie_node **root, const ovs_be32 *prefix, int mlen)
2290 struct trie_node *node;
2291 struct trie_node **edges[sizeof(union mf_value) * 8];
2292 int depth = 0, ofs = 0;
2294 /* Walk the tree. */
2295 for (edges[0] = root;
2296 (node = *edges[depth]) != NULL;
2297 edges[++depth] = trie_next_edge(node, prefix, ofs)) {
2298 unsigned int eqbits = trie_prefix_equal_bits(node, prefix, ofs, mlen);
2300 if (eqbits < node->nbits) {
2301 /* Mismatch, nothing to be removed. This should never happen, as
2302 * only rules in the classifier are ever removed. */
2303 break; /* Log a warning. */
2305 /* Full match so far. */
2309 /* Full prefix match at the current node, remove rule here. */
2310 if (!node->n_rules) {
2311 break; /* Log a warning. */
2315 /* Check if can prune the tree. */
2316 while (!node->n_rules && !(node->edges[0] && node->edges[1])) {
2317 /* No rules and at most one child node, remove this node. */
2318 struct trie_node *next;
2319 next = node->edges[0] ? node->edges[0] : node->edges[1];
2322 if (node->nbits + next->nbits > TRIE_PREFIX_BITS) {
2323 break; /* Cannot combine. */
2325 /* Combine node with next. */
2326 next->prefix = node->prefix | next->prefix >> node->nbits;
2327 next->nbits += node->nbits;
2329 trie_node_destroy(node);
2330 /* Update the parent's edge. */
2331 *edges[depth] = next;
2332 if (next || !depth) {
2333 /* Branch not pruned or at root, nothing more to do. */
2336 node = *edges[--depth];
2341 /* Cannot go deeper. This should never happen, since only rules
2342 * that actually exist in the classifier are ever removed. */
2343 VLOG_WARN("Trying to remove non-existing rule from a prefix trie.");