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 #include "tag.h"
  25
  26 /* Classifier internal definitions, subject to change at any time. */
  27
  28 /* A set of rules that all have the same fields wildcarded. */
  29 struct cls_subtable {
  30     struct cmap_node cmap_node;    /* Within classifier's 'subtables_map'. */
  31
  32     /* These fields are only used by writers. */
  33     int max_priority;              /* Max priority of any rule in subtable. */
  34     unsigned int max_count;        /* Count of max_priority rules. */
  35
  36     /* Accessed by iterators. */
  37     struct rculist rules_list;              /* Unordered. */
  38
  39     /* Identical, but lower priority rules are not inserted to any of the
  40      * following data structures. */
  41
  42     /* These fields are accessed by readers who care about wildcarding. */
  43     const tag_type tag;       /* Tag generated from mask for partitioning. */
  44     const uint8_t n_indices;                   /* How many indices to use. */
  45     const struct flowmap index_maps[CLS_MAX_INDICES + 1]; /* Stage maps. */
  46     unsigned int trie_plen[CLS_MAX_TRIES];  /* Trie prefix length in 'mask'
  47                                              * (runtime configurable). */
  48     const int ports_mask_len;
  49     struct cmap indices[CLS_MAX_INDICES];   /* Staged lookup indices. */
  50     rcu_trie_ptr ports_trie;                /* NULL if none. */
  51
  52     /* These fields are accessed by all readers. */
  53     struct cmap rules;                      /* Contains 'cls_match'es. */
  54     const struct minimask mask;             /* Wildcards for fields. */
  55     /* 'mask' must be the last field. */
  56 };
  57
  58 /* Associates a metadata value (that is, a value of the OpenFlow 1.1+ metadata
  59  * field) with tags for the "cls_subtable"s that contain rules that match that
  60  * metadata value.  */
  61 struct cls_partition {
  62     struct cmap_node cmap_node; /* In struct classifier's 'partitions' map. */
  63     ovs_be64 metadata;          /* metadata value for this partition. */
  64     tag_type tags;              /* OR of each flow's cls_subtable tag. */
  65     struct tag_tracker tracker; /* Tracks the bits in 'tags'. */
  66 };
  67
  68 /* Internal representation of a rule in a "struct cls_subtable".
  69  *
  70  * The 'next' member is an element in a singly linked, null-terminated list.
  71  * This list links together identical "cls_match"es in order of decreasing
  72  * priority.  The classifier code maintains the invariant that at most one rule
  73  * of a given priority is visible for any given lookup version.
  74  */
  75 struct cls_match {
  76     /* Accessed by everybody. */
  77     OVSRCU_TYPE(struct cls_match *) next; /* Equal, lower-priority matches. */
  78     OVSRCU_TYPE(struct cls_conjunction_set *) conj_set;
  79
  80     /* Accessed only by writers. */
  81     struct cls_partition *partition;
  82
  83     /* Accessed by readers interested in wildcarding. */
  84     const int priority;         /* Larger numbers are higher priorities. */
  85     struct cmap_node index_nodes[CLS_MAX_INDICES]; /* Within subtable's
  86                                                     * 'indices'. */
  87     /* Accessed by all readers. */
  88     struct cmap_node cmap_node; /* Within struct cls_subtable 'rules'. */
  89
  90     /* Rule versioning.
  91      *
  92      * CLS_NOT_REMOVED_VERSION has a special meaning for 'remove_version',
  93      * meaningthat the rule has been added but not yet removed.
  94      */
  95     const cls_version_t add_version;        /* Version rule was added in. */
  96     ATOMIC(cls_version_t) remove_version;   /* Version rule is removed in. */
  97
  98     const struct cls_rule *cls_rule;
  99     const struct miniflow flow; /* Matching rule. Mask is in the subtable. */
 100     /* 'flow' must be the last field. */
 101 };
 102
 103 /* Must be RCU postponed. */
 104 void cls_match_free_cb(struct cls_match *);
 105
 106 static inline void
 107 cls_match_set_remove_version(struct cls_match *rule, cls_version_t version)
 108 {
 109     atomic_store_relaxed(&rule->remove_version, version);
 110 }
 111
 112 static inline bool
 113 cls_match_visible_in_version(const struct cls_match *rule,
 114                              cls_version_t version)
 115 {
 116     cls_version_t remove_version;
 117
 118     /* C11 does not want to access an atomic via a const object pointer. */
 119     atomic_read_relaxed(&CONST_CAST(struct cls_match *, rule)->remove_version,
 120                         &remove_version);
 121
 122     return rule->add_version <= version && version < remove_version;
 123 }
 124
 125 static inline bool
 126 cls_match_is_eventually_invisible(const struct cls_match *rule)
 127 {
 128     cls_version_t remove_version;
 129
 130     /* C11 does not want to access an atomic via a const object pointer. */
 131     atomic_read_relaxed(&CONST_CAST(struct cls_match *, rule)->remove_version,
 132                         &remove_version);
 133
 134     return remove_version <= CLS_MAX_VERSION;
 135 }
 136
 137 \f
 138 /* cls_match 'next' */
 139
 140 static inline const struct cls_match *
 141 cls_match_next(const struct cls_match *rule)
 142 {
 143     return ovsrcu_get(struct cls_match *, &rule->next);
 144 }
 145
 146 static inline struct cls_match *
 147 cls_match_next_protected(const struct cls_match *rule)
 148 {
 149     return ovsrcu_get_protected(struct cls_match *, &rule->next);
 150 }
 151
 152 /* Puts 'rule' in the position between 'prev' and 'next'.  If 'prev' == NULL,
 153  * then the 'rule' is the new list head, and if 'next' == NULL, the rule is the
 154  * new list tail.
 155  * If there are any nodes between 'prev' and 'next', they are dropped from the
 156  * list. */
 157 static inline void
 158 cls_match_insert(struct cls_match *prev, struct cls_match *next,
 159                  struct cls_match *rule)
 160 {
 161     ovsrcu_set_hidden(&rule->next, next);
 162
 163     if (prev) {
 164         ovsrcu_set(&prev->next, rule);
 165     }
 166 }
 167
 168 /* Puts 'new_rule' in the position of 'old_rule', which is the next node after
 169  * 'prev'. If 'prev' == NULL, then the 'new_rule' is the new list head.
 170  *
 171  * The replaced cls_match still links to the later rules, and may still be
 172  * referenced by other threads until all other threads quiesce.  The replaced
 173  * rule may not be re-inserted, re-initialized, or deleted until after all
 174  * other threads have quiesced (use ovsrcu_postpone). */
 175 static inline void
 176 cls_match_replace(struct cls_match *prev,
 177                   struct cls_match *old_rule, struct cls_match *new_rule)
 178 {
 179     cls_match_insert(prev, cls_match_next_protected(old_rule), new_rule);
 180 }
 181
 182 /* Removes 'rule' following 'prev' from the list. If 'prev' is NULL, then the
 183  * 'rule' is a list head, and the caller is responsible for maintaining its
 184  * list head pointer (if any).
 185  *
 186  * Afterward, the removed rule is not linked to any more, but still links to
 187  * the following rules, and may still be referenced by other threads until all
 188  * other threads quiesce.  The removed rule may not be re-inserted,
 189  * re-initialized, or deleted until after all other threads have quiesced (use
 190  * ovsrcu_postpone).
 191  */
 192 static inline void
 193 cls_match_remove(struct cls_match *prev, struct cls_match *rule)
 194 {
 195     if (prev) {
 196         ovsrcu_set(&prev->next, cls_match_next_protected(rule));
 197     }
 198 }
 199
 200 #define CLS_MATCH_FOR_EACH(ITER, HEAD)                              \
 201     for ((ITER) = (HEAD); (ITER); (ITER) = cls_match_next(ITER))
 202
 203 #define CLS_MATCH_FOR_EACH_AFTER_HEAD(ITER, HEAD)   \
 204     CLS_MATCH_FOR_EACH(ITER, cls_match_next(HEAD))
 205
 206 /* Iterate cls_matches keeping the previous pointer for modifications. */
 207 #define FOR_EACH_RULE_IN_LIST_PROTECTED(ITER, PREV, HEAD)           \
 208     for ((PREV) = NULL, (ITER) = (HEAD);                            \
 209          (ITER);                                                    \
 210          (PREV) = (ITER), (ITER) = cls_match_next_protected(ITER))
 211
 212 \f
 213 /* A longest-prefix match tree. */
 214 struct trie_node {
 215     uint32_t prefix;           /* Prefix bits for this node, MSB first. */
 216     uint8_t  n_bits;           /* Never zero, except for the root node. */
 217     unsigned int n_rules;      /* Number of rules that have this prefix. */
 218     rcu_trie_ptr edges[2];     /* Both NULL if leaf. */
 219 };
 220
 221 /* Max bits per node.  Must fit in struct trie_node's 'prefix'.
 222  * Also tested with 16, 8, and 5 to stress the implementation. */
 223 #define TRIE_PREFIX_BITS 32
 224 \f
 225 /* flow/miniflow/minimask/minimatch utilities.
 226  * These are only used by the classifier, so place them here to allow
 227  * for better optimization. */
 228
 229 /* Returns a hash value for the bits of 'flow' where there are 1-bits in
 230  * 'mask', given 'basis'.
 231  *
 232  * The hash values returned by this function are the same as those returned by
 233  * miniflow_hash_in_minimask(), only the form of the arguments differ. */
 234 static inline uint32_t
 235 flow_hash_in_minimask(const struct flow *flow, const struct minimask *mask,
 236                       uint32_t basis)
 237 {
 238     const uint64_t *mask_values = miniflow_get_values(&mask->masks);
 239     const uint64_t *flow_u64 = (const uint64_t *)flow;
 240     const uint64_t *p = mask_values;
 241     uint32_t hash = basis;
 242     map_t map;
 243
 244     FLOWMAP_FOR_EACH_MAP (map, mask->masks.map) {
 245         size_t idx;
 246
 247         MAP_FOR_EACH_INDEX (idx, map) {
 248             hash = hash_add64(hash, flow_u64[idx] & *p++);
 249         }
 250         flow_u64 += MAP_T_BITS;
 251     }
 252
 253     return hash_finish(hash, (p - mask_values) * 8);
 254 }
 255
 256 /* Returns a hash value for the bits of 'flow' where there are 1-bits in
 257  * 'mask', given 'basis'.
 258  *
 259  * The hash values returned by this function are the same as those returned by
 260  * flow_hash_in_minimask(), only the form of the arguments differ. */
 261 static inline uint32_t
 262 miniflow_hash_in_minimask(const struct miniflow *flow,
 263                           const struct minimask *mask, uint32_t basis)
 264 {
 265     const uint64_t *mask_values = miniflow_get_values(&mask->masks);
 266     const uint64_t *p = mask_values;
 267     uint32_t hash = basis;
 268     uint64_t value;
 269
 270     MINIFLOW_FOR_EACH_IN_FLOWMAP(value, flow, mask->masks.map) {
 271         hash = hash_add64(hash, value & *p++);
 272     }
 273
 274     return hash_finish(hash, (p - mask_values) * 8);
 275 }
 276
 277 /* Returns a hash value for the values of 'flow', indicated by 'range', where
 278  * there are 1-bits in 'mask', given 'basis'.  'range' must be a continuous
 279  * subset of the bits in 'mask''s map, representing a continuous range of the
 280  * minimask's mask data.  '*offset' must be the number of 64-bit units of the
 281  * minimask's data to skip to get to the first unit covered by 'range'. On
 282  * return '*offset' is updated with the number of 64-bit units of the minimask
 283  * consumed.
 284  *
 285  * Typically this function is called for successive ranges of minimask's masks,
 286  * and the first invocation passes '*offset' as zero.
 287  *
 288  * The hash values returned by this function are the same as those returned by
 289  * minimatch_hash_range(), only the form of the arguments differ. */
 290 static inline uint32_t
 291 flow_hash_in_minimask_range(const struct flow *flow,
 292                             const struct minimask *mask,
 293                             const struct flowmap range,
 294                             unsigned int *offset,
 295                             uint32_t *basis)
 296 {
 297     const uint64_t *mask_values = miniflow_get_values(&mask->masks);
 298     const uint64_t *flow_u64 = (const uint64_t *)flow;
 299     const uint64_t *p = mask_values + *offset;
 300     uint32_t hash = *basis;
 301     map_t map;
 302
 303     FLOWMAP_FOR_EACH_MAP (map, range) {
 304         size_t idx;
 305
 306         MAP_FOR_EACH_INDEX (idx, map) {
 307             hash = hash_add64(hash, flow_u64[idx] & *p++);
 308         }
 309         flow_u64 += MAP_T_BITS;
 310     }
 311
 312     *basis = hash; /* Allow continuation from the unfinished value. */
 313     *offset = p - mask_values;
 314     return hash_finish(hash, *offset * 8);
 315 }
 316
 317 /* Fold minimask 'mask''s wildcard mask into 'wc's wildcard mask. */
 318 static inline void
 319 flow_wildcards_fold_minimask(struct flow_wildcards *wc,
 320                              const struct minimask *mask)
 321 {
 322     flow_union_with_miniflow(&wc->masks, &mask->masks);
 323 }
 324
 325 /* Fold minimask 'mask''s wildcard mask into 'wc's wildcard mask for bits in
 326  * 'fmap'.  1-bits in 'fmap' are a subset of 1-bits in 'mask''s map. */
 327 static inline void
 328 flow_wildcards_fold_minimask_in_map(struct flow_wildcards *wc,
 329                                     const struct minimask *mask,
 330                                     const struct flowmap fmap)
 331 {
 332     flow_union_with_miniflow_subset(&wc->masks, &mask->masks, fmap);
 333 }
 334
 335 /* Returns a hash value for 'mask', given 'basis'. */
 336 static inline uint32_t
 337 minimask_hash(const struct minimask *mask, uint32_t basis)
 338 {
 339     const uint64_t *p = miniflow_get_values(&mask->masks);
 340     size_t n_values = miniflow_n_values(&mask->masks);
 341     uint32_t hash = basis;
 342
 343     for (size_t i = 0; i < n_values; i++) {
 344         hash = hash_add64(hash, *p++);
 345     }
 346
 347     map_t map;
 348     FLOWMAP_FOR_EACH_MAP (map, mask->masks.map) {
 349         hash = hash_add64(hash, map);
 350     }
 351
 352     return hash_finish(hash, n_values);
 353 }
 354
 355 /* Returns a hash value for the values of 'match->flow', indicated by 'range',
 356  * where there are 1-bits in 'match->mask', given 'basis'.  'range' must be a
 357  * continuous subset of the bits in the map of 'match', representing a
 358  * continuous range of the mask data of 'match'.  '*offset' must be the number
 359  * of 64-bit units of the match data to skip to get to the first unit covered
 360  * by 'range'.  On return '*offset' is updated with the number of 64-bit units
 361  * of the match consumed.
 362  *
 363  * Typically this function is called for successive ranges of minimask's masks,
 364  * and the first invocation passes '*offset' as zero.
 365  *
 366  * The hash values returned by this function are the same as those returned by
 367  * flow_hash_in_minimask_range(), only the form of the arguments differ. */
 368 static inline uint32_t
 369 minimatch_hash_range(const struct minimatch *match,
 370                      const struct flowmap range, unsigned int *offset,
 371                      uint32_t *basis)
 372 {
 373     const uint64_t *p = miniflow_get_values(match->flow) + *offset;
 374     const uint64_t *q = miniflow_get_values(&match->mask->masks) + *offset;
 375     unsigned int n = flowmap_n_1bits(range);
 376     uint32_t hash = *basis;
 377
 378     for (unsigned int i = 0; i < n; i++) {
 379         hash = hash_add64(hash, p[i] & q[i]);
 380     }
 381     *basis = hash; /* Allow continuation from the unfinished value. */
 382     *offset += n;
 383     return hash_finish(hash, *offset * 8);
 384 }
 385
 386 #endif