lib/classifier-private.h

   1 /*
   2  * Copyright (c) 2014, 2015 Nicira, Inc.
   3  *
   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:
   7  *
   8  *     http://www.apache.org/licenses/LICENSE-2.0
   9  *
  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.
  15  */
  16
  17 #ifndef CLASSIFIER_PRIVATE_H
  18 #define CLASSIFIER_PRIVATE_H 1
  19
  20 #include "cmap.h"
  21 #include "flow.h"
  22 #include "hash.h"
  23 #include "rculist.h"
  24
  25 /* Classifier internal definitions, subject to change at any time. */
  26
  27 /* A set of rules that all have the same fields wildcarded. */
  28 struct cls_subtable {
  29     struct cmap_node cmap_node;    /* Within classifier's 'subtables_map'. */
  30
  31     /* These fields are only used by writers. */
  32     int max_priority;              /* Max priority of any rule in subtable. */
  33     unsigned int max_count;        /* Count of max_priority rules. */
  34
  35     /* Accessed by iterators. */
  36     struct rculist rules_list;              /* Unordered. */
  37
  38     /* Identical, but lower priority rules are not inserted to any of the
  39      * following data structures. */
  40
  41     /* These fields are accessed by readers who care about wildcarding. */
  42     const uint8_t n_indices;                   /* How many indices to use. */
  43     const struct flowmap index_maps[CLS_MAX_INDICES + 1]; /* Stage maps. */
  44     unsigned int trie_plen[CLS_MAX_TRIES];  /* Trie prefix length in 'mask'
  45                                              * (runtime configurable). */
  46     const int ports_mask_len;
  47     struct cmap indices[CLS_MAX_INDICES];   /* Staged lookup indices. */
  48     rcu_trie_ptr ports_trie;                /* NULL if none. */
  49
  50     /* These fields are accessed by all readers. */
  51     struct cmap rules;                      /* Contains 'cls_match'es. */
  52     const struct minimask mask;             /* Wildcards for fields. */
  53     /* 'mask' must be the last field. */
  54 };
  55
  56 /* Internal representation of a rule in a "struct cls_subtable".
  57  *
  58  * The 'next' member is an element in a singly linked, null-terminated list.
  59  * This list links together identical "cls_match"es in order of decreasing
  60  * priority.  The classifier code maintains the invariant that at most one rule
  61  * of a given priority is visible for any given lookup version.
  62  */
  63 struct cls_match {
  64     /* Accessed by everybody. */
  65     OVSRCU_TYPE(struct cls_match *) next; /* Equal, lower-priority matches. */
  66     OVSRCU_TYPE(struct cls_conjunction_set *) conj_set;
  67
  68     /* Accessed by readers interested in wildcarding. */
  69     const int priority;         /* Larger numbers are higher priorities. */
  70     struct cmap_node index_nodes[CLS_MAX_INDICES]; /* Within subtable's
  71                                                     * 'indices'. */
  72     /* Accessed by all readers. */
  73     struct cmap_node cmap_node; /* Within struct cls_subtable 'rules'. */
  74
  75     /* Rule versioning.
  76      *
  77      * CLS_NOT_REMOVED_VERSION has a special meaning for 'remove_version',
  78      * meaning that the rule has been added but not yet removed.
  79      */
  80     const cls_version_t add_version;        /* Version rule was added in. */
  81     ATOMIC(cls_version_t) remove_version;   /* Version rule is removed in. */
  82
  83     const struct cls_rule *cls_rule;
  84     const struct miniflow flow; /* Matching rule. Mask is in the subtable. */
  85     /* 'flow' must be the last field. */
  86 };
  87
  88 /* Must be RCU postponed. */
  89 void cls_match_free_cb(struct cls_match *);
  90
  91 static inline void
  92 cls_match_set_remove_version(struct cls_match *rule, cls_version_t version)
  93 {
  94     atomic_store_relaxed(&rule->remove_version, version);
  95 }
  96
  97 static inline bool
  98 cls_match_visible_in_version(const struct cls_match *rule,
  99                              cls_version_t version)
 100 {
 101     cls_version_t remove_version;
 102
 103     /* C11 does not want to access an atomic via a const object pointer. */
 104     atomic_read_relaxed(&CONST_CAST(struct cls_match *, rule)->remove_version,
 105                         &remove_version);
 106
 107     return rule->add_version <= version && version < remove_version;
 108 }
 109
 110 static inline bool
 111 cls_match_is_eventually_invisible(const struct cls_match *rule)
 112 {
 113     cls_version_t remove_version;
 114
 115     /* C11 does not want to access an atomic via a const object pointer. */
 116     atomic_read_relaxed(&CONST_CAST(struct cls_match *, rule)->remove_version,
 117                         &remove_version);
 118
 119     return remove_version <= CLS_MAX_VERSION;
 120 }
 121
 122 \f
 123 /* cls_match 'next' */
 124
 125 static inline const struct cls_match *
 126 cls_match_next(const struct cls_match *rule)
 127 {
 128     return ovsrcu_get(struct cls_match *, &rule->next);
 129 }
 130
 131 static inline struct cls_match *
 132 cls_match_next_protected(const struct cls_match *rule)
 133 {
 134     return ovsrcu_get_protected(struct cls_match *, &rule->next);
 135 }
 136
 137 /* Puts 'rule' in the position between 'prev' and 'next'.  If 'prev' == NULL,
 138  * then the 'rule' is the new list head, and if 'next' == NULL, the rule is the
 139  * new list tail.
 140  * If there are any nodes between 'prev' and 'next', they are dropped from the
 141  * list. */
 142 static inline void
 143 cls_match_insert(struct cls_match *prev, struct cls_match *next,
 144                  struct cls_match *rule)
 145 {
 146     ovsrcu_set_hidden(&rule->next, next);
 147
 148     if (prev) {
 149         ovsrcu_set(&prev->next, rule);
 150     }
 151 }
 152
 153 /* Puts 'new_rule' in the position of 'old_rule', which is the next node after
 154  * 'prev'. If 'prev' == NULL, then the 'new_rule' is the new list head.
 155  *
 156  * The replaced cls_match still links to the later rules, and may still be
 157  * referenced by other threads until all other threads quiesce.  The replaced
 158  * rule may not be re-inserted, re-initialized, or deleted until after all
 159  * other threads have quiesced (use ovsrcu_postpone). */
 160 static inline void
 161 cls_match_replace(struct cls_match *prev,
 162                   struct cls_match *old_rule, struct cls_match *new_rule)
 163 {
 164     cls_match_insert(prev, cls_match_next_protected(old_rule), new_rule);
 165 }
 166
 167 /* Removes 'rule' following 'prev' from the list. If 'prev' is NULL, then the
 168  * 'rule' is a list head, and the caller is responsible for maintaining its
 169  * list head pointer (if any).
 170  *
 171  * Afterward, the removed rule is not linked to any more, but still links to
 172  * the following rules, and may still be referenced by other threads until all
 173  * other threads quiesce.  The removed rule may not be re-inserted,
 174  * re-initialized, or deleted until after all other threads have quiesced (use
 175  * ovsrcu_postpone).
 176  */
 177 static inline void
 178 cls_match_remove(struct cls_match *prev, struct cls_match *rule)
 179 {
 180     if (prev) {
 181         ovsrcu_set(&prev->next, cls_match_next_protected(rule));
 182     }
 183 }
 184
 185 #define CLS_MATCH_FOR_EACH(ITER, HEAD)                              \
 186     for ((ITER) = (HEAD); (ITER); (ITER) = cls_match_next(ITER))
 187
 188 #define CLS_MATCH_FOR_EACH_AFTER_HEAD(ITER, HEAD)   \
 189     CLS_MATCH_FOR_EACH(ITER, cls_match_next(HEAD))
 190
 191 /* Iterate cls_matches keeping the previous pointer for modifications. */
 192 #define FOR_EACH_RULE_IN_LIST_PROTECTED(ITER, PREV, HEAD)           \
 193     for ((PREV) = NULL, (ITER) = (HEAD);                            \
 194          (ITER);                                                    \
 195          (PREV) = (ITER), (ITER) = cls_match_next_protected(ITER))
 196
 197 \f
 198 /* A longest-prefix match tree. */
 199 struct trie_node {
 200     uint32_t prefix;           /* Prefix bits for this node, MSB first. */
 201     uint8_t  n_bits;           /* Never zero, except for the root node. */
 202     unsigned int n_rules;      /* Number of rules that have this prefix. */
 203     rcu_trie_ptr edges[2];     /* Both NULL if leaf. */
 204 };
 205
 206 /* Max bits per node.  Must fit in struct trie_node's 'prefix'.
 207  * Also tested with 16, 8, and 5 to stress the implementation. */
 208 #define TRIE_PREFIX_BITS 32
 209 \f
 210 /* flow/miniflow/minimask/minimatch utilities.
 211  * These are only used by the classifier, so place them here to allow
 212  * for better optimization. */
 213
 214 /* Returns a hash value for the bits of 'flow' where there are 1-bits in
 215  * 'mask', given 'basis'.
 216  *
 217  * The hash values returned by this function are the same as those returned by
 218  * miniflow_hash_in_minimask(), only the form of the arguments differ. */
 219 static inline uint32_t
 220 flow_hash_in_minimask(const struct flow *flow, const struct minimask *mask,
 221                       uint32_t basis)
 222 {
 223     const uint64_t *mask_values = miniflow_get_values(&mask->masks);
 224     const uint64_t *flow_u64 = (const uint64_t *)flow;
 225     const uint64_t *p = mask_values;
 226     uint32_t hash = basis;
 227     map_t map;
 228
 229     FLOWMAP_FOR_EACH_MAP (map, mask->masks.map) {
 230         size_t idx;
 231
 232         MAP_FOR_EACH_INDEX (idx, map) {
 233             hash = hash_add64(hash, flow_u64[idx] & *p++);
 234         }
 235         flow_u64 += MAP_T_BITS;
 236     }
 237
 238     return hash_finish(hash, (p - mask_values) * 8);
 239 }
 240
 241 /* Returns a hash value for the bits of 'flow' where there are 1-bits in
 242  * 'mask', given 'basis'.
 243  *
 244  * The hash values returned by this function are the same as those returned by
 245  * flow_hash_in_minimask(), only the form of the arguments differ. */
 246 static inline uint32_t
 247 miniflow_hash_in_minimask(const struct miniflow *flow,
 248                           const struct minimask *mask, uint32_t basis)
 249 {
 250     const uint64_t *mask_values = miniflow_get_values(&mask->masks);
 251     const uint64_t *p = mask_values;
 252     uint32_t hash = basis;
 253     uint64_t value;
 254
 255     MINIFLOW_FOR_EACH_IN_FLOWMAP(value, flow, mask->masks.map) {
 256         hash = hash_add64(hash, value & *p++);
 257     }
 258
 259     return hash_finish(hash, (p - mask_values) * 8);
 260 }
 261
 262 /* Returns a hash value for the values of 'flow', indicated by 'range', where
 263  * there are 1-bits in 'mask', given 'basis'.  'range' must be a continuous
 264  * subset of the bits in 'mask''s map, representing a continuous range of the
 265  * minimask's mask data.  '*offset' must be the number of 64-bit units of the
 266  * minimask's data to skip to get to the first unit covered by 'range'. On
 267  * return '*offset' is updated with the number of 64-bit units of the minimask
 268  * consumed.
 269  *
 270  * Typically this function is called for successive ranges of minimask's masks,
 271  * and the first invocation passes '*offset' as zero.
 272  *
 273  * The hash values returned by this function are the same as those returned by
 274  * minimatch_hash_range(), only the form of the arguments differ. */
 275 static inline uint32_t
 276 flow_hash_in_minimask_range(const struct flow *flow,
 277                             const struct minimask *mask,
 278                             const struct flowmap range,
 279                             unsigned int *offset,
 280                             uint32_t *basis)
 281 {
 282     const uint64_t *mask_values = miniflow_get_values(&mask->masks);
 283     const uint64_t *flow_u64 = (const uint64_t *)flow;
 284     const uint64_t *p = mask_values + *offset;
 285     uint32_t hash = *basis;
 286     map_t map;
 287
 288     FLOWMAP_FOR_EACH_MAP (map, range) {
 289         size_t idx;
 290
 291         MAP_FOR_EACH_INDEX (idx, map) {
 292             hash = hash_add64(hash, flow_u64[idx] & *p++);
 293         }
 294         flow_u64 += MAP_T_BITS;
 295     }
 296
 297     *basis = hash; /* Allow continuation from the unfinished value. */
 298     *offset = p - mask_values;
 299     return hash_finish(hash, *offset * 8);
 300 }
 301
 302 /* Fold minimask 'mask''s wildcard mask into 'wc's wildcard mask. */
 303 static inline void
 304 flow_wildcards_fold_minimask(struct flow_wildcards *wc,
 305                              const struct minimask *mask)
 306 {
 307     flow_union_with_miniflow(&wc->masks, &mask->masks);
 308 }
 309
 310 /* Fold minimask 'mask''s wildcard mask into 'wc's wildcard mask for bits in
 311  * 'fmap'.  1-bits in 'fmap' are a subset of 1-bits in 'mask''s map. */
 312 static inline void
 313 flow_wildcards_fold_minimask_in_map(struct flow_wildcards *wc,
 314                                     const struct minimask *mask,
 315                                     const struct flowmap fmap)
 316 {
 317     flow_union_with_miniflow_subset(&wc->masks, &mask->masks, fmap);
 318 }
 319
 320 /* Returns a hash value for 'mask', given 'basis'. */
 321 static inline uint32_t
 322 minimask_hash(const struct minimask *mask, uint32_t basis)
 323 {
 324     const uint64_t *p = miniflow_get_values(&mask->masks);
 325     size_t n_values = miniflow_n_values(&mask->masks);
 326     uint32_t hash = basis;
 327
 328     for (size_t i = 0; i < n_values; i++) {
 329         hash = hash_add64(hash, *p++);
 330     }
 331
 332     map_t map;
 333     FLOWMAP_FOR_EACH_MAP (map, mask->masks.map) {
 334         hash = hash_add64(hash, map);
 335     }
 336
 337     return hash_finish(hash, n_values);
 338 }
 339
 340 /* Returns a hash value for the values of 'match->flow', indicated by 'range',
 341  * where there are 1-bits in 'match->mask', given 'basis'.  'range' must be a
 342  * continuous subset of the bits in the map of 'match', representing a
 343  * continuous range of the mask data of 'match'.  '*offset' must be the number
 344  * of 64-bit units of the match data to skip to get to the first unit covered
 345  * by 'range'.  On return '*offset' is updated with the number of 64-bit units
 346  * of the match consumed.
 347  *
 348  * Typically this function is called for successive ranges of minimask's masks,
 349  * and the first invocation passes '*offset' as zero.
 350  *
 351  * The hash values returned by this function are the same as those returned by
 352  * flow_hash_in_minimask_range(), only the form of the arguments differ. */
 353 static inline uint32_t
 354 minimatch_hash_range(const struct minimatch *match,
 355                      const struct flowmap range, unsigned int *offset,
 356                      uint32_t *basis)
 357 {
 358     const uint64_t *p = miniflow_get_values(match->flow) + *offset;
 359     const uint64_t *q = miniflow_get_values(&match->mask->masks) + *offset;
 360     unsigned int n = flowmap_n_1bits(range);
 361     uint32_t hash = *basis;
 362
 363     for (unsigned int i = 0; i < n; i++) {
 364         hash = hash_add64(hash, p[i] & q[i]);
 365     }
 366     *basis = hash; /* Allow continuation from the unfinished value. */
 367     *offset += n;
 368     return hash_finish(hash, *offset * 8);
 369 }
 370
 371 #endif