X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=lib%2Fclassifier.c;h=67b2c280b7a3779f8c27fd86a42f85920c91466e;hb=ec988646afe6aee6a63d6894a3e9b50f715d5941;hp=33ade96050b6dfcb27c67a7352acc163e5925f61;hpb=476f36e83bc5b0ca7f11952a23bb0fef76d1cc4b;p=sliver-openvswitch.git diff --git a/lib/classifier.c b/lib/classifier.c index 33ade9605..67b2c280b 100644 --- a/lib/classifier.c +++ b/lib/classifier.c @@ -24,29 +24,96 @@ #include "hash.h" #include "odp-util.h" #include "ofp-util.h" -#include "packets.h" #include "ovs-thread.h" +#include "packets.h" +#include "vlog.h" + +VLOG_DEFINE_THIS_MODULE(classifier); -static struct cls_subtable *find_subtable(const struct classifier *, +struct trie_node; + +/* Prefix trie for a 'field' */ +struct cls_trie { + const struct mf_field *field; /* Trie field, or NULL. */ + struct trie_node *root; /* NULL if none. */ +}; + +struct cls_subtable_entry { + struct cls_subtable *subtable; + uint32_t *mask_values; + tag_type tag; + unsigned int max_priority; +}; + +struct cls_subtable_cache { + struct cls_subtable_entry *subtables; + size_t alloc_size; /* Number of allocated elements. */ + size_t size; /* One past last valid array element. */ +}; + +struct cls_classifier { + int n_rules; /* Total number of rules. */ + uint8_t n_flow_segments; + uint8_t flow_segments[CLS_MAX_INDICES]; /* Flow segment boundaries to use + * for staged lookup. */ + struct hmap subtables; /* Contains "struct cls_subtable"s. */ + struct cls_subtable_cache subtables_priority; + struct hmap partitions; /* Contains "struct cls_partition"s. */ + struct cls_trie tries[CLS_MAX_TRIES]; /* Prefix tries. */ + unsigned int n_tries; +}; + +/* A set of rules that all have the same fields wildcarded. */ +struct cls_subtable { + struct hmap_node hmap_node; /* Within struct cls_classifier 'subtables' + * hmap. */ + struct hmap rules; /* Contains "struct cls_rule"s. */ + struct minimask mask; /* Wildcards for fields. */ + int n_rules; /* Number of rules, including duplicates. */ + unsigned int max_priority; /* Max priority of any rule in the subtable. */ + unsigned int max_count; /* Count of max_priority rules. */ + tag_type tag; /* Tag generated from mask for partitioning. */ + uint8_t n_indices; /* How many indices to use. */ + uint8_t index_ofs[CLS_MAX_INDICES]; /* u32 flow segment boundaries. */ + struct hindex indices[CLS_MAX_INDICES]; /* Staged lookup indices. */ + unsigned int trie_plen[CLS_MAX_TRIES]; /* Trie prefix length in 'mask'. */ +}; + +/* Associates a metadata value (that is, a value of the OpenFlow 1.1+ metadata + * field) with tags for the "cls_subtable"s that contain rules that match that + * metadata value. */ +struct cls_partition { + struct hmap_node hmap_node; /* In struct cls_classifier's 'partitions' + * hmap. */ + ovs_be64 metadata; /* metadata value for this partition. */ + tag_type tags; /* OR of each flow's cls_subtable tag. */ + struct tag_tracker tracker; /* Tracks the bits in 'tags'. */ +}; + + + +struct trie_ctx; +static struct cls_subtable *find_subtable(const struct cls_classifier *, const struct minimask *); -static struct cls_subtable *insert_subtable(struct classifier *, +static struct cls_subtable *insert_subtable(struct cls_classifier *, const struct minimask *); -static void destroy_subtable(struct classifier *, struct cls_subtable *); +static void destroy_subtable(struct cls_classifier *, struct cls_subtable *); -static void update_subtables_after_insertion(struct classifier *, +static void update_subtables_after_insertion(struct cls_classifier *, struct cls_subtable *, unsigned int new_priority); -static void update_subtables_after_removal(struct classifier *, +static void update_subtables_after_removal(struct cls_classifier *, struct cls_subtable *, unsigned int del_priority); static struct cls_rule *find_match_wc(const struct cls_subtable *, - const struct flow *, + const struct flow *, struct trie_ctx *, + unsigned int n_tries, struct flow_wildcards *); static struct cls_rule *find_equal(struct cls_subtable *, const struct miniflow *, uint32_t hash); -static struct cls_rule *insert_rule(struct classifier *, +static struct cls_rule *insert_rule(struct cls_classifier *, struct cls_subtable *, struct cls_rule *); /* Iterates RULE over HEAD and all of the cls_rules on HEAD->list. */ @@ -59,6 +126,277 @@ static struct cls_rule *insert_rule(struct classifier *, static struct cls_rule *next_rule_in_list__(struct cls_rule *); static struct cls_rule *next_rule_in_list(struct cls_rule *); + +static unsigned int minimask_get_prefix_len(const struct minimask *, + const struct mf_field *); +static void trie_init(struct cls_classifier *, int trie_idx, + const struct mf_field *); +static unsigned int trie_lookup(const struct cls_trie *, const struct flow *, + unsigned int *checkbits); + +static void trie_destroy(struct trie_node *); +static void trie_insert(struct cls_trie *, const struct cls_rule *, int mlen); +static void trie_remove(struct cls_trie *, const struct cls_rule *, int mlen); +static void mask_set_prefix_bits(struct flow_wildcards *, uint8_t be32ofs, + unsigned int nbits); +static bool mask_prefix_bits_set(const struct flow_wildcards *, + uint8_t be32ofs, unsigned int nbits); + +static void +cls_subtable_cache_init(struct cls_subtable_cache *array) +{ + memset(array, 0, sizeof *array); +} + +static void +cls_subtable_cache_destroy(struct cls_subtable_cache *array) +{ + free(array->subtables); + memset(array, 0, sizeof *array); +} + +/* Array insertion. */ +static void +cls_subtable_cache_push_back(struct cls_subtable_cache *array, + struct cls_subtable_entry a) +{ + if (array->size == array->alloc_size) { + array->subtables = x2nrealloc(array->subtables, &array->alloc_size, + sizeof a); + } + + array->subtables[array->size++] = a; +} + +/* Only for rearranging entries in the same cache. */ +static inline void +cls_subtable_cache_splice(struct cls_subtable_entry *to, + struct cls_subtable_entry *start, + struct cls_subtable_entry *end) +{ + if (to > end) { + /* Same as splicing entries to (start) from [end, to). */ + struct cls_subtable_entry *temp = to; + to = start; start = end; end = temp; + } + if (to < start) { + while (start != end) { + struct cls_subtable_entry temp = *start; + + memmove(to + 1, to, (start - to) * sizeof *to); + *to = temp; + start++; + } + } /* Else nothing to be done. */ +} + +/* Array removal. */ +static inline void +cls_subtable_cache_remove(struct cls_subtable_cache *array, + struct cls_subtable_entry *elem) +{ + ssize_t size = (&array->subtables[array->size] + - (elem + 1)) * sizeof *elem; + if (size > 0) { + memmove(elem, elem + 1, size); + } + array->size--; +} + +#define CLS_SUBTABLE_CACHE_FOR_EACH(SUBTABLE, ITER, ARRAY) \ + for (ITER = (ARRAY)->subtables; \ + ITER < &(ARRAY)->subtables[(ARRAY)->size] \ + && OVS_LIKELY(SUBTABLE = ITER->subtable); \ + ++ITER) +#define CLS_SUBTABLE_CACHE_FOR_EACH_CONTINUE(SUBTABLE, ITER, ARRAY) \ + for (++ITER; \ + ITER < &(ARRAY)->subtables[(ARRAY)->size] \ + && OVS_LIKELY(SUBTABLE = ITER->subtable); \ + ++ITER) +#define CLS_SUBTABLE_CACHE_FOR_EACH_REVERSE(SUBTABLE, ITER, ARRAY) \ + for (ITER = &(ARRAY)->subtables[(ARRAY)->size]; \ + ITER > (ARRAY)->subtables \ + && OVS_LIKELY(SUBTABLE = (--ITER)->subtable);) + + +/* flow/miniflow/minimask/minimatch utilities. + * These are only used by the classifier, so place them here to allow + * for better optimization. */ + +static inline uint64_t +miniflow_get_map_in_range(const struct miniflow *miniflow, + uint8_t start, uint8_t end, unsigned int *offset) +{ + uint64_t map = miniflow->map; + *offset = 0; + + if (start > 0) { + uint64_t msk = (UINT64_C(1) << start) - 1; /* 'start' LSBs set */ + *offset = count_1bits(map & msk); + map &= ~msk; + } + if (end < FLOW_U32S) { + uint64_t msk = (UINT64_C(1) << end) - 1; /* 'end' LSBs set */ + map &= msk; + } + return map; +} + +/* Returns a hash value for the bits of 'flow' where there are 1-bits in + * 'mask', given 'basis'. + * + * The hash values returned by this function are the same as those returned by + * miniflow_hash_in_minimask(), only the form of the arguments differ. */ +static inline uint32_t +flow_hash_in_minimask(const struct flow *flow, const struct minimask *mask, + uint32_t basis) +{ + const uint32_t *flow_u32 = (const uint32_t *)flow; + const uint32_t *p = mask->masks.values; + uint32_t hash; + uint64_t map; + + hash = basis; + for (map = mask->masks.map; map; map = zero_rightmost_1bit(map)) { + hash = mhash_add(hash, flow_u32[raw_ctz(map)] & *p++); + } + + return mhash_finish(hash, (p - mask->masks.values) * 4); +} + +/* Returns a hash value for the bits of 'flow' where there are 1-bits in + * 'mask', given 'basis'. + * + * The hash values returned by this function are the same as those returned by + * flow_hash_in_minimask(), only the form of the arguments differ. */ +static inline uint32_t +miniflow_hash_in_minimask(const struct miniflow *flow, + const struct minimask *mask, uint32_t basis) +{ + const uint32_t *p = mask->masks.values; + uint32_t hash = basis; + uint32_t flow_u32; + + MINIFLOW_FOR_EACH_IN_MAP(flow_u32, flow, mask->masks.map) { + hash = mhash_add(hash, flow_u32 & *p++); + } + + return mhash_finish(hash, (p - mask->masks.values) * 4); +} + +/* Returns a hash value for the bits of range [start, end) in 'flow', + * where there are 1-bits in 'mask', given 'hash'. + * + * The hash values returned by this function are the same as those returned by + * minimatch_hash_range(), only the form of the arguments differ. */ +static inline uint32_t +flow_hash_in_minimask_range(const struct flow *flow, + const struct minimask *mask, + uint8_t start, uint8_t end, uint32_t *basis) +{ + const uint32_t *flow_u32 = (const uint32_t *)flow; + unsigned int offset; + uint64_t map = miniflow_get_map_in_range(&mask->masks, start, end, + &offset); + const uint32_t *p = mask->masks.values + offset; + uint32_t hash = *basis; + + for (; map; map = zero_rightmost_1bit(map)) { + hash = mhash_add(hash, flow_u32[raw_ctz(map)] & *p++); + } + + *basis = hash; /* Allow continuation from the unfinished value. */ + return mhash_finish(hash, (p - mask->masks.values) * 4); +} + +/* Fold minimask 'mask''s wildcard mask into 'wc's wildcard mask. */ +static inline void +flow_wildcards_fold_minimask(struct flow_wildcards *wc, + const struct minimask *mask) +{ + flow_union_with_miniflow(&wc->masks, &mask->masks); +} + +/* Fold minimask 'mask''s wildcard mask into 'wc's wildcard mask + * in range [start, end). */ +static inline void +flow_wildcards_fold_minimask_range(struct flow_wildcards *wc, + const struct minimask *mask, + uint8_t start, uint8_t end) +{ + uint32_t *dst_u32 = (uint32_t *)&wc->masks; + unsigned int offset; + uint64_t map = miniflow_get_map_in_range(&mask->masks, start, end, + &offset); + const uint32_t *p = mask->masks.values + offset; + + for (; map; map = zero_rightmost_1bit(map)) { + dst_u32[raw_ctz(map)] |= *p++; + } +} + +/* Returns a hash value for 'flow', given 'basis'. */ +static inline uint32_t +miniflow_hash(const struct miniflow *flow, uint32_t basis) +{ + const uint32_t *p = flow->values; + uint32_t hash = basis; + uint64_t hash_map = 0; + uint64_t map; + + for (map = flow->map; map; map = zero_rightmost_1bit(map)) { + if (*p) { + hash = mhash_add(hash, *p); + hash_map |= rightmost_1bit(map); + } + p++; + } + hash = mhash_add(hash, hash_map); + hash = mhash_add(hash, hash_map >> 32); + + return mhash_finish(hash, p - flow->values); +} + +/* Returns a hash value for 'mask', given 'basis'. */ +static inline uint32_t +minimask_hash(const struct minimask *mask, uint32_t basis) +{ + return miniflow_hash(&mask->masks, basis); +} + +/* Returns a hash value for 'match', given 'basis'. */ +static inline uint32_t +minimatch_hash(const struct minimatch *match, uint32_t basis) +{ + return miniflow_hash(&match->flow, minimask_hash(&match->mask, basis)); +} + +/* Returns a hash value for the bits of range [start, end) in 'minimatch', + * given 'basis'. + * + * The hash values returned by this function are the same as those returned by + * flow_hash_in_minimask_range(), only the form of the arguments differ. */ +static inline uint32_t +minimatch_hash_range(const struct minimatch *match, uint8_t start, uint8_t end, + uint32_t *basis) +{ + unsigned int offset; + const uint32_t *p, *q; + uint32_t hash = *basis; + int n, i; + + n = count_1bits(miniflow_get_map_in_range(&match->mask.masks, start, end, + &offset)); + q = match->mask.masks.values + offset; + p = match->flow.values + offset; + + for (i = 0; i < n; i++) { + hash = mhash_add(hash, p[i] & q[i]); + } + *basis = hash; /* Allow continuation from the unfinished value. */ + return mhash_finish(hash, (offset + n) * 4); +} + /* cls_rule. */ @@ -150,13 +488,18 @@ cls_rule_is_catchall(const struct cls_rule *rule) /* Initializes 'cls' as a classifier that initially contains no classification * rules. */ void -classifier_init(struct classifier *cls, const uint8_t *flow_segments) +classifier_init(struct classifier *cls_, const uint8_t *flow_segments) { + struct cls_classifier *cls = xmalloc(sizeof *cls); + + fat_rwlock_init(&cls_->rwlock); + + cls_->cls = cls; + cls->n_rules = 0; hmap_init(&cls->subtables); - list_init(&cls->subtables_priority); + cls_subtable_cache_init(&cls->subtables_priority); hmap_init(&cls->partitions); - ovs_rwlock_init(&cls->rwlock); cls->n_flow_segments = 0; if (flow_segments) { while (cls->n_flow_segments < CLS_MAX_INDICES @@ -164,16 +507,28 @@ classifier_init(struct classifier *cls, const uint8_t *flow_segments) cls->flow_segments[cls->n_flow_segments++] = *flow_segments++; } } + cls->n_tries = 0; } /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the * caller's responsibility. */ void -classifier_destroy(struct classifier *cls) +classifier_destroy(struct classifier *cls_) { - if (cls) { + if (cls_) { + struct cls_classifier *cls = cls_->cls; struct cls_subtable *partition, *next_partition; struct cls_subtable *subtable, *next_subtable; + int i; + + fat_rwlock_destroy(&cls_->rwlock); + if (!cls) { + return; + } + + for (i = 0; i < cls->n_tries; i++) { + trie_destroy(cls->tries[i].root); + } HMAP_FOR_EACH_SAFE (subtable, next_subtable, hmap_node, &cls->subtables) { @@ -187,7 +542,88 @@ classifier_destroy(struct classifier *cls) free(partition); } hmap_destroy(&cls->partitions); - ovs_rwlock_destroy(&cls->rwlock); + + cls_subtable_cache_destroy(&cls->subtables_priority); + free(cls); + } +} + +/* We use uint64_t as a set for the fields below. */ +BUILD_ASSERT_DECL(MFF_N_IDS <= 64); + +/* Set the fields for which prefix lookup should be performed. */ +void +classifier_set_prefix_fields(struct classifier *cls_, + const enum mf_field_id *trie_fields, + unsigned int n_fields) +{ + struct cls_classifier *cls = cls_->cls; + uint64_t fields = 0; + int i, trie; + + for (i = 0, trie = 0; i < n_fields && trie < CLS_MAX_TRIES; i++) { + const struct mf_field *field = mf_from_id(trie_fields[i]); + if (field->flow_be32ofs < 0 || field->n_bits % 32) { + /* Incompatible field. This is the only place where we + * enforce these requirements, but the rest of the trie code + * depends on the flow_be32ofs to be non-negative and the + * field length to be a multiple of 32 bits. */ + continue; + } + + if (fields & (UINT64_C(1) << trie_fields[i])) { + /* Duplicate field, there is no need to build more than + * one index for any one field. */ + continue; + } + fields |= UINT64_C(1) << trie_fields[i]; + + if (trie >= cls->n_tries || field != cls->tries[trie].field) { + trie_init(cls, trie, field); + } + trie++; + } + + /* Destroy the rest. */ + for (i = trie; i < cls->n_tries; i++) { + trie_init(cls, i, NULL); + } + cls->n_tries = trie; +} + +static void +trie_init(struct cls_classifier *cls, int trie_idx, + const struct mf_field *field) +{ + struct cls_trie *trie = &cls->tries[trie_idx]; + struct cls_subtable *subtable; + struct cls_subtable_entry *iter; + + if (trie_idx < cls->n_tries) { + trie_destroy(trie->root); + } + trie->root = NULL; + trie->field = field; + + /* Add existing rules to the trie. */ + CLS_SUBTABLE_CACHE_FOR_EACH (subtable, iter, &cls->subtables_priority) { + unsigned int plen; + + plen = field ? minimask_get_prefix_len(&subtable->mask, field) : 0; + /* Initialize subtable's prefix length on this field. */ + subtable->trie_plen[trie_idx] = plen; + + if (plen) { + struct cls_rule *head; + + HMAP_FOR_EACH (head, hmap_node, &subtable->rules) { + struct cls_rule *rule; + + FOR_EACH_RULE_IN_LIST (rule, head) { + trie_insert(trie, rule, plen); + } + } + } } } @@ -195,25 +631,26 @@ classifier_destroy(struct classifier *cls) bool classifier_is_empty(const struct classifier *cls) { - return cls->n_rules == 0; + return cls->cls->n_rules == 0; } /* Returns the number of rules in 'cls'. */ int classifier_count(const struct classifier *cls) { - return cls->n_rules; + return cls->cls->n_rules; } static uint32_t hash_metadata(ovs_be64 metadata_) { uint64_t metadata = (OVS_FORCE uint64_t) metadata_; - return hash_2words(metadata, metadata >> 32); + return hash_uint64(metadata); } static struct cls_partition * -find_partition(const struct classifier *cls, ovs_be64 metadata, uint32_t hash) +find_partition(const struct cls_classifier *cls, ovs_be64 metadata, + uint32_t hash) { struct cls_partition *partition; @@ -227,7 +664,7 @@ find_partition(const struct classifier *cls, ovs_be64 metadata, uint32_t hash) } static struct cls_partition * -create_partition(struct classifier *cls, struct cls_subtable *subtable, +create_partition(struct cls_classifier *cls, struct cls_subtable *subtable, ovs_be64 metadata) { uint32_t hash = hash_metadata(metadata); @@ -257,8 +694,9 @@ create_partition(struct classifier *cls, struct cls_subtable *subtable, * rule, even rules that cannot have any effect because the new rule matches a * superset of their flows and has higher priority. */ struct cls_rule * -classifier_replace(struct classifier *cls, struct cls_rule *rule) +classifier_replace(struct classifier *cls_, struct cls_rule *rule) { + struct cls_classifier *cls = cls_->cls; struct cls_rule *old_rule; struct cls_subtable *subtable; @@ -269,6 +707,8 @@ classifier_replace(struct classifier *cls, struct cls_rule *rule) old_rule = insert_rule(cls, subtable, rule); if (!old_rule) { + int i; + if (minimask_get_metadata_mask(&rule->match.mask) == OVS_BE64_MAX) { ovs_be64 metadata = miniflow_get_metadata(&rule->match.flow); rule->partition = create_partition(cls, subtable, metadata); @@ -278,6 +718,12 @@ classifier_replace(struct classifier *cls, struct cls_rule *rule) subtable->n_rules++; cls->n_rules++; + + for (i = 0; i < cls->n_tries; i++) { + if (subtable->trie_plen[i]) { + trie_insert(&cls->tries[i], rule, subtable->trie_plen[i]); + } + } } else { rule->partition = old_rule->partition; } @@ -301,8 +747,9 @@ classifier_insert(struct classifier *cls, struct cls_rule *rule) * 'rule' with cls_rule_destroy(), freeing the memory block in which 'rule' * resides, etc., as necessary. */ void -classifier_remove(struct classifier *cls, struct cls_rule *rule) +classifier_remove(struct classifier *cls_, struct cls_rule *rule) { + struct cls_classifier *cls = cls_->cls; struct cls_partition *partition; struct cls_rule *head; struct cls_subtable *subtable; @@ -310,6 +757,12 @@ classifier_remove(struct classifier *cls, struct cls_rule *rule) subtable = find_subtable(cls, &rule->match.mask); + for (i = 0; i < cls->n_tries; i++) { + if (subtable->trie_plen[i]) { + trie_remove(&cls->tries[i], rule, subtable->trie_plen[i]); + } + } + /* Remove rule node from indices. */ for (i = 0; i < subtable->n_indices; i++) { hindex_remove(&subtable->indices[i], &rule->index_nodes[i]); @@ -343,9 +796,37 @@ classifier_remove(struct classifier *cls, struct cls_rule *rule) } else { update_subtables_after_removal(cls, subtable, rule->priority); } + cls->n_rules--; } +/* Prefix tree context. Valid when 'lookup_done' is true. Can skip all + * subtables which have more than 'match_plen' bits in their corresponding + * field at offset 'be32ofs'. If skipped, 'maskbits' prefix bits should be + * unwildcarded to quarantee datapath flow matches only packets it should. */ +struct trie_ctx { + const struct cls_trie *trie; + bool lookup_done; /* Status of the lookup. */ + uint8_t be32ofs; /* U32 offset of the field in question. */ + unsigned int match_plen; /* Longest prefix than could possibly match. */ + unsigned int maskbits; /* Prefix length needed to avoid false matches. */ +}; + +static void +trie_ctx_init(struct trie_ctx *ctx, const struct cls_trie *trie) +{ + ctx->trie = trie; + ctx->be32ofs = trie->field->flow_be32ofs; + ctx->lookup_done = false; +} + +static inline void +lookahead_subtable(const struct cls_subtable_entry *subtables) +{ + ovs_prefetch_range(subtables->subtable, sizeof *subtables->subtable); + ovs_prefetch_range(subtables->mask_values, 1); +} + /* Finds and returns the highest-priority rule in 'cls' that matches 'flow'. * Returns a null pointer if no rules in 'cls' match 'flow'. If multiple rules * of equal priority match 'flow', returns one arbitrarily. @@ -355,13 +836,21 @@ classifier_remove(struct classifier *cls, struct cls_rule *rule) * earlier, 'wc' should have been initialized (e.g., by * flow_wildcards_init_catchall()). */ struct cls_rule * -classifier_lookup(const struct classifier *cls, const struct flow *flow, +classifier_lookup(const struct classifier *cls_, const struct flow *flow, struct flow_wildcards *wc) { + struct cls_classifier *cls = cls_->cls; const struct cls_partition *partition; - struct cls_subtable *subtable; - struct cls_rule *best; tag_type tags; + struct cls_rule *best; + struct trie_ctx trie_ctx[CLS_MAX_TRIES]; + int i; + struct cls_subtable_entry *subtables = cls->subtables_priority.subtables; + int n_subtables = cls->subtables_priority.size; + int64_t best_priority = -1; + + /* Prefetch the subtables array. */ + ovs_prefetch_range(subtables, n_subtables * sizeof *subtables); /* Determine 'tags' such that, if 'subtable->tag' doesn't intersect them, * then 'flow' cannot possibly match in 'subtable': @@ -387,46 +876,119 @@ classifier_lookup(const struct classifier *cls, const struct flow *flow, hash_metadata(flow->metadata))); tags = partition ? partition->tags : TAG_ARBITRARY; + /* Initialize trie contexts for match_find_wc(). */ + for (i = 0; i < cls->n_tries; i++) { + trie_ctx_init(&trie_ctx[i], &cls->tries[i]); + } + + /* Prefetch the first subtables. */ + if (n_subtables > 1) { + lookahead_subtable(subtables); + lookahead_subtable(subtables + 1); + } + best = NULL; - LIST_FOR_EACH (subtable, list_node, &cls->subtables_priority) { + for (i = 0; OVS_LIKELY(i < n_subtables); i++) { struct cls_rule *rule; - if (!tag_intersects(tags, subtable->tag)) { + if ((int64_t)subtables[i].max_priority <= best_priority) { + /* Subtables are in descending priority order, + * can not find anything better. */ + break; + } + + /* Prefetch a forthcoming subtable. */ + if (i + 2 < n_subtables) { + lookahead_subtable(&subtables[i + 2]); + } + + if (!tag_intersects(tags, subtables[i].tag)) { continue; } - rule = find_match_wc(subtable, flow, wc); - if (rule) { + rule = find_match_wc(subtables[i].subtable, flow, trie_ctx, + cls->n_tries, wc); + if (rule && (int64_t)rule->priority > best_priority) { + best_priority = (int64_t)rule->priority; best = rule; - LIST_FOR_EACH_CONTINUE (subtable, list_node, - &cls->subtables_priority) { - if (subtable->max_priority <= best->priority) { - /* Subtables are in descending priority order, - * can not find anything better. */ - return best; - } - if (!tag_intersects(tags, subtable->tag)) { - continue; - } - - rule = find_match_wc(subtable, flow, wc); - if (rule && rule->priority > best->priority) { - best = rule; - } - } - break; } } + return best; } +/* Returns true if 'target' satisifies 'match', that is, if each bit for which + * 'match' specifies a particular value has the correct value in 'target'. */ +static bool +minimatch_matches_miniflow(const struct minimatch *match, + const struct miniflow *target) +{ + const uint32_t *flowp = (const uint32_t *)match->flow.values; + const uint32_t *maskp = (const uint32_t *)match->mask.masks.values; + uint32_t target_u32; + + MINIFLOW_FOR_EACH_IN_MAP(target_u32, target, match->mask.masks.map) { + if ((*flowp++ ^ target_u32) & *maskp++) { + return false; + } + } + + return true; +} + +static inline struct cls_rule * +find_match_miniflow(const struct cls_subtable *subtable, + const struct miniflow *flow, + uint32_t hash) +{ + struct cls_rule *rule; + + HMAP_FOR_EACH_WITH_HASH (rule, hmap_node, hash, &subtable->rules) { + if (minimatch_matches_miniflow(&rule->match, flow)) { + return rule; + } + } + + return NULL; +} + +/* Finds and returns the highest-priority rule in 'cls' that matches + * 'miniflow'. Returns a null pointer if no rules in 'cls' match 'flow'. + * If multiple rules of equal priority match 'flow', returns one arbitrarily. + * + * This function is optimized for the userspace datapath, which only ever has + * one priority value for it's flows! + */ +struct cls_rule *classifier_lookup_miniflow_first(const struct classifier *cls_, + const struct miniflow *flow) +{ + struct cls_classifier *cls = cls_->cls; + struct cls_subtable *subtable; + struct cls_subtable_entry *iter; + + CLS_SUBTABLE_CACHE_FOR_EACH (subtable, iter, &cls->subtables_priority) { + struct cls_rule *rule; + + rule = find_match_miniflow(subtable, flow, + miniflow_hash_in_minimask(flow, + &subtable->mask, + 0)); + if (rule) { + return rule; + } + } + + return NULL; +} + /* Finds and returns a rule in 'cls' with exactly the same priority and * matching criteria as 'target'. Returns a null pointer if 'cls' doesn't * contain an exact match. */ struct cls_rule * -classifier_find_rule_exactly(const struct classifier *cls, +classifier_find_rule_exactly(const struct classifier *cls_, const struct cls_rule *target) { + struct cls_classifier *cls = cls_->cls; struct cls_rule *head, *rule; struct cls_subtable *subtable; @@ -473,18 +1035,20 @@ classifier_find_match_exactly(const struct classifier *cls, * considered to overlap if both rules have the same priority and a packet * could match both. */ bool -classifier_rule_overlaps(const struct classifier *cls, +classifier_rule_overlaps(const struct classifier *cls_, const struct cls_rule *target) { + struct cls_classifier *cls = cls_->cls; struct cls_subtable *subtable; + struct cls_subtable_entry *iter; /* Iterate subtables in the descending max priority order. */ - LIST_FOR_EACH (subtable, list_node, &cls->subtables_priority) { + CLS_SUBTABLE_CACHE_FOR_EACH (subtable, iter, &cls->subtables_priority) { uint32_t storage[FLOW_U32S]; struct minimask mask; struct cls_rule *head; - if (target->priority > subtable->max_priority) { + if (target->priority > iter->max_priority) { break; /* Can skip this and the rest of the subtables. */ } @@ -589,7 +1153,7 @@ void cls_cursor_init(struct cls_cursor *cursor, const struct classifier *cls, const struct cls_rule *target) { - cursor->cls = cls; + cursor->cls = cls->cls; cursor->target = target && !cls_rule_is_catchall(target) ? target : NULL; } @@ -648,7 +1212,7 @@ cls_cursor_next(struct cls_cursor *cursor, const struct cls_rule *rule_) } static struct cls_subtable * -find_subtable(const struct classifier *cls, const struct minimask *mask) +find_subtable(const struct cls_classifier *cls, const struct minimask *mask) { struct cls_subtable *subtable; @@ -662,13 +1226,14 @@ find_subtable(const struct classifier *cls, const struct minimask *mask) } static struct cls_subtable * -insert_subtable(struct classifier *cls, const struct minimask *mask) +insert_subtable(struct cls_classifier *cls, const struct minimask *mask) { uint32_t hash = minimask_hash(mask, 0); struct cls_subtable *subtable; int i, index = 0; struct flow_wildcards old, new; uint8_t prev; + struct cls_subtable_entry elem; subtable = xzalloc(sizeof *subtable); hmap_init(&subtable->rules); @@ -702,19 +1267,38 @@ insert_subtable(struct classifier *cls, const struct minimask *mask) } subtable->n_indices = index; - hmap_insert(&cls->subtables, &subtable->hmap_node, hash); - list_push_back(&cls->subtables_priority, &subtable->list_node); subtable->tag = (minimask_get_metadata_mask(mask) == OVS_BE64_MAX ? tag_create_deterministic(hash) : TAG_ALL); + for (i = 0; i < cls->n_tries; i++) { + subtable->trie_plen[i] = minimask_get_prefix_len(mask, + cls->tries[i].field); + } + + hmap_insert(&cls->subtables, &subtable->hmap_node, hash); + elem.subtable = subtable; + elem.mask_values = subtable->mask.masks.values; + elem.tag = subtable->tag; + elem.max_priority = subtable->max_priority; + cls_subtable_cache_push_back(&cls->subtables_priority, elem); + return subtable; } static void -destroy_subtable(struct classifier *cls, struct cls_subtable *subtable) +destroy_subtable(struct cls_classifier *cls, struct cls_subtable *subtable) { int i; + struct cls_subtable *table = NULL; + struct cls_subtable_entry *iter; + + CLS_SUBTABLE_CACHE_FOR_EACH (table, iter, &cls->subtables_priority) { + if (table == subtable) { + cls_subtable_cache_remove(&cls->subtables_priority, iter); + break; + } + } for (i = 0; i < subtable->n_indices; i++) { hindex_destroy(&subtable->indices[i]); @@ -722,7 +1306,6 @@ destroy_subtable(struct classifier *cls, struct cls_subtable *subtable) minimask_destroy(&subtable->mask); hmap_remove(&cls->subtables, &subtable->hmap_node); hmap_destroy(&subtable->rules); - list_remove(&subtable->list_node); free(subtable); } @@ -737,35 +1320,38 @@ destroy_subtable(struct classifier *cls, struct cls_subtable *subtable) * This function should only be called after adding a new rule, not after * replacing a rule by an identical one or modifying a rule in-place. */ static void -update_subtables_after_insertion(struct classifier *cls, +update_subtables_after_insertion(struct cls_classifier *cls, struct cls_subtable *subtable, unsigned int new_priority) { if (new_priority == subtable->max_priority) { ++subtable->max_count; } else if (new_priority > subtable->max_priority) { - struct cls_subtable *iter; + struct cls_subtable *table; + struct cls_subtable_entry *iter, *subtable_iter = NULL; subtable->max_priority = new_priority; subtable->max_count = 1; /* Possibly move 'subtable' earlier in the priority list. If we break - * out of the loop, then 'subtable' should be moved just after that + * out of the loop, then 'subtable_iter' should be moved just before * 'iter'. If the loop terminates normally, then 'iter' will be the - * list head and we'll move subtable just after that (e.g. to the front - * of the list). */ - iter = subtable; - LIST_FOR_EACH_REVERSE_CONTINUE (iter, list_node, - &cls->subtables_priority) { - if (iter->max_priority >= subtable->max_priority) { + * first list element and we'll move subtable just before that + * (e.g. to the front of the list). */ + CLS_SUBTABLE_CACHE_FOR_EACH_REVERSE (table, iter, &cls->subtables_priority) { + if (table == subtable) { + subtable_iter = iter; /* Locate the subtable as we go. */ + iter->max_priority = new_priority; + } else if (table->max_priority >= new_priority) { + ovs_assert(subtable_iter != NULL); + iter++; break; } } - /* Move 'subtable' just after 'iter' (unless it's already there). */ - if (iter->list_node.next != &subtable->list_node) { - list_splice(iter->list_node.next, - &subtable->list_node, subtable->list_node.next); + /* Move 'subtable' just before 'iter' (unless it's already there). */ + if (iter != subtable_iter) { + cls_subtable_cache_splice(iter, subtable_iter, subtable_iter + 1); } } } @@ -781,14 +1367,14 @@ update_subtables_after_insertion(struct classifier *cls, * This function should only be called after removing a rule, not after * replacing a rule by an identical one or modifying a rule in-place. */ static void -update_subtables_after_removal(struct classifier *cls, +update_subtables_after_removal(struct cls_classifier *cls, struct cls_subtable *subtable, unsigned int del_priority) { - struct cls_subtable *iter; - if (del_priority == subtable->max_priority && --subtable->max_count == 0) { struct cls_rule *head; + struct cls_subtable *table; + struct cls_subtable_entry *iter, *subtable_iter = NULL; subtable->max_priority = 0; HMAP_FOR_EACH (head, hmap_node, &subtable->rules) { @@ -805,21 +1391,89 @@ update_subtables_after_removal(struct classifier *cls, * 'iter'. If the loop terminates normally, then 'iter' will be the * list head and we'll move subtable just before that (e.g. to the back * of the list). */ - iter = subtable; - LIST_FOR_EACH_CONTINUE (iter, list_node, &cls->subtables_priority) { - if (iter->max_priority <= subtable->max_priority) { + CLS_SUBTABLE_CACHE_FOR_EACH (table, iter, &cls->subtables_priority) { + if (table == subtable) { + subtable_iter = iter; /* Locate the subtable as we go. */ + iter->max_priority = subtable->max_priority; + } else if (table->max_priority <= subtable->max_priority) { + ovs_assert(subtable_iter != NULL); break; } } /* Move 'subtable' just before 'iter' (unless it's already there). */ - if (iter->list_node.prev != &subtable->list_node) { - list_splice(&iter->list_node, - &subtable->list_node, subtable->list_node.next); + if (iter != subtable_iter) { + cls_subtable_cache_splice(iter, subtable_iter, subtable_iter + 1); } } } +struct range { + uint8_t start; + uint8_t end; +}; + +/* Return 'true' if can skip rest of the subtable based on the prefix trie + * lookup results. */ +static inline bool +check_tries(struct trie_ctx trie_ctx[CLS_MAX_TRIES], unsigned int n_tries, + const unsigned int field_plen[CLS_MAX_TRIES], + const struct range ofs, const struct flow *flow, + struct flow_wildcards *wc) +{ + int j; + + /* Check if we could avoid fully unwildcarding the next level of + * fields using the prefix tries. The trie checks are done only as + * needed to avoid folding in additional bits to the wildcards mask. */ + for (j = 0; j < n_tries; j++) { + /* Is the trie field relevant for this subtable? */ + if (field_plen[j]) { + struct trie_ctx *ctx = &trie_ctx[j]; + uint8_t be32ofs = ctx->be32ofs; + + /* Is the trie field within the current range of fields? */ + if (be32ofs >= ofs.start && be32ofs < ofs.end) { + /* On-demand trie lookup. */ + if (!ctx->lookup_done) { + ctx->match_plen = trie_lookup(ctx->trie, flow, + &ctx->maskbits); + ctx->lookup_done = true; + } + /* Possible to skip the rest of the subtable if subtable's + * prefix on the field is longer than what is known to match + * based on the trie lookup. */ + if (field_plen[j] > ctx->match_plen) { + /* RFC: We want the trie lookup to never result in + * unwildcarding any bits that would not be unwildcarded + * otherwise. Since the trie is shared by the whole + * classifier, it is possible that the 'maskbits' contain + * bits that are irrelevant for the partition of the + * classifier relevant for the current flow. */ + + /* Can skip if the field is already unwildcarded. */ + if (mask_prefix_bits_set(wc, be32ofs, ctx->maskbits)) { + return true; + } + /* Check that the trie result will not unwildcard more bits + * than this stage will. */ + if (ctx->maskbits <= field_plen[j]) { + /* Unwildcard the bits and skip the rest. */ + mask_set_prefix_bits(wc, be32ofs, ctx->maskbits); + /* Note: Prerequisite already unwildcarded, as the only + * prerequisite of the supported trie lookup fields is + * the ethertype, which is currently always + * unwildcarded. + */ + return true; + } + } + } + } + } + return false; +} + static inline struct cls_rule * find_match(const struct cls_subtable *subtable, const struct flow *flow, uint32_t hash) @@ -837,32 +1491,37 @@ find_match(const struct cls_subtable *subtable, const struct flow *flow, static struct cls_rule * find_match_wc(const struct cls_subtable *subtable, const struct flow *flow, - struct flow_wildcards * wc) + struct trie_ctx trie_ctx[CLS_MAX_TRIES], unsigned int n_tries, + struct flow_wildcards *wc) { uint32_t basis = 0, hash; struct cls_rule *rule = NULL; - uint8_t prev_u32ofs = 0; int i; + struct range ofs; - if (!wc) { + if (OVS_UNLIKELY(!wc)) { return find_match(subtable, flow, flow_hash_in_minimask(flow, &subtable->mask, 0)); } + ofs.start = 0; /* Try to finish early by checking fields in segments. */ for (i = 0; i < subtable->n_indices; i++) { struct hindex_node *inode; + ofs.end = subtable->index_ofs[i]; - hash = flow_hash_in_minimask_range(flow, &subtable->mask, prev_u32ofs, - subtable->index_ofs[i], &basis); - prev_u32ofs = subtable->index_ofs[i]; + if (check_tries(trie_ctx, n_tries, subtable->trie_plen, ofs, flow, + wc)) { + goto range_out; + } + hash = flow_hash_in_minimask_range(flow, &subtable->mask, ofs.start, + ofs.end, &basis); + ofs.start = ofs.end; inode = hindex_node_with_hash(&subtable->indices[i], hash); if (!inode) { /* No match, can stop immediately, but must fold in the mask * covered so far. */ - flow_wildcards_fold_minimask_range(wc, &subtable->mask, 0, - prev_u32ofs); - return NULL; + goto range_out; } /* If we have narrowed down to a single rule already, check whether @@ -884,11 +1543,15 @@ find_match_wc(const struct cls_subtable *subtable, const struct flow *flow, } } } - + ofs.end = FLOW_U32S; + /* Trie check for the final range. */ + if (check_tries(trie_ctx, n_tries, subtable->trie_plen, ofs, flow, wc)) { + goto range_out; + } if (!rule) { /* Multiple potential matches exist, look for one. */ - hash = flow_hash_in_minimask_range(flow, &subtable->mask, prev_u32ofs, - FLOW_U32S, &basis); + hash = flow_hash_in_minimask_range(flow, &subtable->mask, ofs.start, + ofs.end, &basis); rule = find_match(subtable, flow, hash); } else { /* We already narrowed the matching candidates down to just 'rule', @@ -896,8 +1559,16 @@ find_match_wc(const struct cls_subtable *subtable, const struct flow *flow, rule = NULL; } out: + /* Must unwildcard all the fields, as they were looked at. */ flow_wildcards_fold_minimask(wc, &subtable->mask); return rule; + + range_out: + /* Must unwildcard the fields looked up so far, if any. */ + if (ofs.start) { + flow_wildcards_fold_minimask_range(wc, &subtable->mask, 0, ofs.start); + } + return NULL; } static struct cls_rule * @@ -915,23 +1586,23 @@ find_equal(struct cls_subtable *subtable, const struct miniflow *flow, } static struct cls_rule * -insert_rule(struct classifier *cls, struct cls_subtable *subtable, +insert_rule(struct cls_classifier *cls, struct cls_subtable *subtable, struct cls_rule *new) { struct cls_rule *head; struct cls_rule *old = NULL; int i; uint32_t basis = 0, hash; - uint8_t prev_u32ofs = 0; + uint8_t prev_be32ofs = 0; /* Add new node to segment indices. */ for (i = 0; i < subtable->n_indices; i++) { - hash = minimatch_hash_range(&new->match, prev_u32ofs, + hash = minimatch_hash_range(&new->match, prev_be32ofs, subtable->index_ofs[i], &basis); hindex_insert(&subtable->indices[i], &new->index_nodes[i], hash); - prev_u32ofs = subtable->index_ofs[i]; + prev_be32ofs = subtable->index_ofs[i]; } - hash = minimatch_hash_range(&new->match, prev_u32ofs, FLOW_U32S, &basis); + hash = minimatch_hash_range(&new->match, prev_be32ofs, FLOW_U32S, &basis); head = find_equal(subtable, &new->match.flow, hash); if (!head) { hmap_insert(&subtable->rules, &new->hmap_node, hash); @@ -992,3 +1663,393 @@ next_rule_in_list(struct cls_rule *rule) struct cls_rule *next = next_rule_in_list__(rule); return next->priority < rule->priority ? next : NULL; } + +/* A longest-prefix match tree. */ +struct trie_node { + uint32_t prefix; /* Prefix bits for this node, MSB first. */ + uint8_t nbits; /* Never zero, except for the root node. */ + unsigned int n_rules; /* Number of rules that have this prefix. */ + struct trie_node *edges[2]; /* Both NULL if leaf. */ +}; + +/* Max bits per node. Must fit in struct trie_node's 'prefix'. + * Also tested with 16, 8, and 5 to stress the implementation. */ +#define TRIE_PREFIX_BITS 32 + +/* Return at least 'plen' bits of the 'prefix', starting at bit offset 'ofs'. + * Prefixes are in the network byte order, and the offset 0 corresponds to + * the most significant bit of the first byte. The offset can be read as + * "how many bits to skip from the start of the prefix starting at 'pr'". */ +static uint32_t +raw_get_prefix(const ovs_be32 pr[], unsigned int ofs, unsigned int plen) +{ + uint32_t prefix; + + pr += ofs / 32; /* Where to start. */ + ofs %= 32; /* How many bits to skip at 'pr'. */ + + prefix = ntohl(*pr) << ofs; /* Get the first 32 - ofs bits. */ + if (plen > 32 - ofs) { /* Need more than we have already? */ + prefix |= ntohl(*++pr) >> (32 - ofs); + } + /* Return with possible unwanted bits at the end. */ + return prefix; +} + +/* Return min(TRIE_PREFIX_BITS, plen) bits of the 'prefix', starting at bit + * offset 'ofs'. Prefixes are in the network byte order, and the offset 0 + * corresponds to the most significant bit of the first byte. The offset can + * be read as "how many bits to skip from the start of the prefix starting at + * 'pr'". */ +static uint32_t +trie_get_prefix(const ovs_be32 pr[], unsigned int ofs, unsigned int plen) +{ + if (!plen) { + return 0; + } + if (plen > TRIE_PREFIX_BITS) { + plen = TRIE_PREFIX_BITS; /* Get at most TRIE_PREFIX_BITS. */ + } + /* Return with unwanted bits cleared. */ + return raw_get_prefix(pr, ofs, plen) & ~0u << (32 - plen); +} + +/* Return the number of equal bits in 'nbits' of 'prefix's MSBs and a 'value' + * starting at "MSB 0"-based offset 'ofs'. */ +static unsigned int +prefix_equal_bits(uint32_t prefix, unsigned int nbits, const ovs_be32 value[], + unsigned int ofs) +{ + uint64_t diff = prefix ^ raw_get_prefix(value, ofs, nbits); + /* Set the bit after the relevant bits to limit the result. */ + return raw_clz64(diff << 32 | UINT64_C(1) << (63 - nbits)); +} + +/* Return the number of equal bits in 'node' prefix and a 'prefix' of length + * 'plen', starting at "MSB 0"-based offset 'ofs'. */ +static unsigned int +trie_prefix_equal_bits(const struct trie_node *node, const ovs_be32 prefix[], + unsigned int ofs, unsigned int plen) +{ + return prefix_equal_bits(node->prefix, MIN(node->nbits, plen - ofs), + prefix, ofs); +} + +/* Return the bit at ("MSB 0"-based) offset 'ofs' as an int. 'ofs' can + * be greater than 31. */ +static unsigned int +be_get_bit_at(const ovs_be32 value[], unsigned int ofs) +{ + return (((const uint8_t *)value)[ofs / 8] >> (7 - ofs % 8)) & 1u; +} + +/* Return the bit at ("MSB 0"-based) offset 'ofs' as an int. 'ofs' must + * be between 0 and 31, inclusive. */ +static unsigned int +get_bit_at(const uint32_t prefix, unsigned int ofs) +{ + return (prefix >> (31 - ofs)) & 1u; +} + +/* Create new branch. */ +static struct trie_node * +trie_branch_create(const ovs_be32 *prefix, unsigned int ofs, unsigned int plen, + unsigned int n_rules) +{ + struct trie_node *node = xmalloc(sizeof *node); + + node->prefix = trie_get_prefix(prefix, ofs, plen); + + if (plen <= TRIE_PREFIX_BITS) { + node->nbits = plen; + node->edges[0] = NULL; + node->edges[1] = NULL; + node->n_rules = n_rules; + } else { /* Need intermediate nodes. */ + struct trie_node *subnode = trie_branch_create(prefix, + ofs + TRIE_PREFIX_BITS, + plen - TRIE_PREFIX_BITS, + n_rules); + int bit = get_bit_at(subnode->prefix, 0); + node->nbits = TRIE_PREFIX_BITS; + node->edges[bit] = subnode; + node->edges[!bit] = NULL; + node->n_rules = 0; + } + return node; +} + +static void +trie_node_destroy(struct trie_node *node) +{ + free(node); +} + +static void +trie_destroy(struct trie_node *node) +{ + if (node) { + trie_destroy(node->edges[0]); + trie_destroy(node->edges[1]); + free(node); + } +} + +static bool +trie_is_leaf(const struct trie_node *trie) +{ + return !trie->edges[0] && !trie->edges[1]; /* No children. */ +} + +static void +mask_set_prefix_bits(struct flow_wildcards *wc, uint8_t be32ofs, + unsigned int nbits) +{ + ovs_be32 *mask = &((ovs_be32 *)&wc->masks)[be32ofs]; + unsigned int i; + + for (i = 0; i < nbits / 32; i++) { + mask[i] = OVS_BE32_MAX; + } + if (nbits % 32) { + mask[i] |= htonl(~0u << (32 - nbits % 32)); + } +} + +static bool +mask_prefix_bits_set(const struct flow_wildcards *wc, uint8_t be32ofs, + unsigned int nbits) +{ + ovs_be32 *mask = &((ovs_be32 *)&wc->masks)[be32ofs]; + unsigned int i; + ovs_be32 zeroes = 0; + + for (i = 0; i < nbits / 32; i++) { + zeroes |= ~mask[i]; + } + if (nbits % 32) { + zeroes |= ~mask[i] & htonl(~0u << (32 - nbits % 32)); + } + + return !zeroes; /* All 'nbits' bits set. */ +} + +static struct trie_node ** +trie_next_edge(struct trie_node *node, const ovs_be32 value[], + unsigned int ofs) +{ + return node->edges + be_get_bit_at(value, ofs); +} + +static const struct trie_node * +trie_next_node(const struct trie_node *node, const ovs_be32 value[], + unsigned int ofs) +{ + return node->edges[be_get_bit_at(value, ofs)]; +} + +/* Return the prefix mask length necessary to find the longest-prefix match for + * the '*value' in the prefix tree 'node'. + * '*checkbits' is set to the number of bits in the prefix mask necessary to + * determine a mismatch, in case there are longer prefixes in the tree below + * the one that matched. + */ +static unsigned int +trie_lookup_value(const struct trie_node *node, const ovs_be32 value[], + unsigned int *checkbits) +{ + unsigned int plen = 0, match_len = 0; + const struct trie_node *prev = NULL; + + for (; node; prev = node, node = trie_next_node(node, value, plen)) { + unsigned int eqbits; + /* Check if this edge can be followed. */ + eqbits = prefix_equal_bits(node->prefix, node->nbits, value, plen); + plen += eqbits; + if (eqbits < node->nbits) { /* Mismatch, nothing more to be found. */ + /* Bit at offset 'plen' differed. */ + *checkbits = plen + 1; /* Includes the first mismatching bit. */ + return match_len; + } + /* Full match, check if rules exist at this prefix length. */ + if (node->n_rules > 0) { + match_len = plen; + } + } + /* Dead end, exclude the other branch if it exists. */ + *checkbits = !prev || trie_is_leaf(prev) ? plen : plen + 1; + return match_len; +} + +static unsigned int +trie_lookup(const struct cls_trie *trie, const struct flow *flow, + unsigned int *checkbits) +{ + const struct mf_field *mf = trie->field; + + /* Check that current flow matches the prerequisites for the trie + * field. Some match fields are used for multiple purposes, so we + * must check that the trie is relevant for this flow. */ + if (mf_are_prereqs_ok(mf, flow)) { + return trie_lookup_value(trie->root, + &((ovs_be32 *)flow)[mf->flow_be32ofs], + checkbits); + } + *checkbits = 0; /* Value not used in this case. */ + return UINT_MAX; +} + +/* Returns the length of a prefix match mask for the field 'mf' in 'minimask'. + * Returns the u32 offset to the miniflow data in '*miniflow_index', if + * 'miniflow_index' is not NULL. */ +static unsigned int +minimask_get_prefix_len(const struct minimask *minimask, + const struct mf_field *mf) +{ + unsigned int nbits = 0, mask_tz = 0; /* Non-zero when end of mask seen. */ + uint8_t u32_ofs = mf->flow_be32ofs; + uint8_t u32_end = u32_ofs + mf->n_bytes / 4; + + for (; u32_ofs < u32_end; ++u32_ofs) { + uint32_t mask; + mask = ntohl((OVS_FORCE ovs_be32)minimask_get(minimask, u32_ofs)); + + /* Validate mask, count the mask length. */ + if (mask_tz) { + if (mask) { + return 0; /* No bits allowed after mask ended. */ + } + } else { + if (~mask & (~mask + 1)) { + return 0; /* Mask not contiguous. */ + } + mask_tz = ctz32(mask); + nbits += 32 - mask_tz; + } + } + + return nbits; +} + +/* + * This is called only when mask prefix is known to be CIDR and non-zero. + * Relies on the fact that the flow and mask have the same map, and since + * the mask is CIDR, the storage for the flow field exists even if it + * happened to be zeros. + */ +static const ovs_be32 * +minimatch_get_prefix(const struct minimatch *match, const struct mf_field *mf) +{ + return match->flow.values + + count_1bits(match->flow.map & ((UINT64_C(1) << mf->flow_be32ofs) - 1)); +} + +/* Insert rule in to the prefix tree. + * 'mlen' must be the (non-zero) CIDR prefix length of the 'trie->field' mask + * in 'rule'. */ +static void +trie_insert(struct cls_trie *trie, const struct cls_rule *rule, int mlen) +{ + const ovs_be32 *prefix = minimatch_get_prefix(&rule->match, trie->field); + struct trie_node *node; + struct trie_node **edge; + int ofs = 0; + + /* Walk the tree. */ + for (edge = &trie->root; + (node = *edge) != NULL; + edge = trie_next_edge(node, prefix, ofs)) { + unsigned int eqbits = trie_prefix_equal_bits(node, prefix, ofs, mlen); + ofs += eqbits; + if (eqbits < node->nbits) { + /* Mismatch, new node needs to be inserted above. */ + int old_branch = get_bit_at(node->prefix, eqbits); + + /* New parent node. */ + *edge = trie_branch_create(prefix, ofs - eqbits, eqbits, + ofs == mlen ? 1 : 0); + + /* Adjust old node for its new position in the tree. */ + node->prefix <<= eqbits; + node->nbits -= eqbits; + (*edge)->edges[old_branch] = node; + + /* Check if need a new branch for the new rule. */ + if (ofs < mlen) { + (*edge)->edges[!old_branch] + = trie_branch_create(prefix, ofs, mlen - ofs, 1); + } + return; + } + /* Full match so far. */ + + if (ofs == mlen) { + /* Full match at the current node, rule needs to be added here. */ + node->n_rules++; + return; + } + } + /* Must insert a new tree branch for the new rule. */ + *edge = trie_branch_create(prefix, ofs, mlen - ofs, 1); +} + +/* 'mlen' must be the (non-zero) CIDR prefix length of the 'trie->field' mask + * in 'rule'. */ +static void +trie_remove(struct cls_trie *trie, const struct cls_rule *rule, int mlen) +{ + const ovs_be32 *prefix = minimatch_get_prefix(&rule->match, trie->field); + struct trie_node *node; + struct trie_node **edges[sizeof(union mf_value) * 8]; + int depth = 0, ofs = 0; + + /* Walk the tree. */ + for (edges[depth] = &trie->root; + (node = *edges[depth]) != NULL; + edges[++depth] = trie_next_edge(node, prefix, ofs)) { + unsigned int eqbits = trie_prefix_equal_bits(node, prefix, ofs, mlen); + if (eqbits < node->nbits) { + /* Mismatch, nothing to be removed. This should never happen, as + * only rules in the classifier are ever removed. */ + break; /* Log a warning. */ + } + /* Full match so far. */ + ofs += eqbits; + + if (ofs == mlen) { + /* Full prefix match at the current node, remove rule here. */ + if (!node->n_rules) { + break; /* Log a warning. */ + } + node->n_rules--; + + /* Check if can prune the tree. */ + while (!node->n_rules && !(node->edges[0] && node->edges[1])) { + /* No rules and at most one child node, remove this node. */ + struct trie_node *next; + next = node->edges[0] ? node->edges[0] : node->edges[1]; + + if (next) { + if (node->nbits + next->nbits > TRIE_PREFIX_BITS) { + break; /* Cannot combine. */ + } + /* Combine node with next. */ + next->prefix = node->prefix | next->prefix >> node->nbits; + next->nbits += node->nbits; + } + trie_node_destroy(node); + /* Update the parent's edge. */ + *edges[depth] = next; + if (next || !depth) { + /* Branch not pruned or at root, nothing more to do. */ + break; + } + node = *edges[--depth]; + } + return; + } + } + /* Cannot go deeper. This should never happen, since only rules + * that actually exist in the classifier are ever removed. */ + VLOG_WARN("Trying to remove non-existing rule from a prefix trie."); +}