2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014, 2015 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 #define CLASSIFIER_H 1
26 * A flow classifier holds any number of "rules", each of which specifies
27 * values to match for some fields or subfields and a priority. Each OpenFlow
28 * table is implemented as a flow classifier.
30 * The classifier has two primary design goals. The first is obvious: given a
31 * set of packet headers, as quickly as possible find the highest-priority rule
32 * that matches those headers. The following section describes the second
39 * A primary goal of the flow classifier is to produce, as a side effect of a
40 * packet lookup, a wildcard mask that indicates which bits of the packet
41 * headers were essential to the classification result. Ideally, a 1-bit in
42 * any position of this mask means that, if the corresponding bit in the packet
43 * header were flipped, then the classification result might change. A 0-bit
44 * means that changing the packet header bit would have no effect. Thus, the
45 * wildcarded bits are the ones that played no role in the classification
48 * Such a wildcard mask is useful with datapaths that support installing flows
49 * that wildcard fields or subfields. If an OpenFlow lookup for a TCP flow
50 * does not actually look at the TCP source or destination ports, for example,
51 * then the switch may install into the datapath a flow that wildcards the port
52 * numbers, which in turn allows the datapath to handle packets that arrive for
53 * other TCP source or destination ports without additional help from
54 * ovs-vswitchd. This is useful for the Open vSwitch software and,
55 * potentially, for ASIC-based switches as well.
57 * Some properties of the wildcard mask:
59 * - "False 1-bits" are acceptable, that is, setting a bit in the wildcard
60 * mask to 1 will never cause a packet to be forwarded the wrong way.
61 * As a corollary, a wildcard mask composed of all 1-bits will always
62 * yield correct (but often needlessly inefficient) behavior.
64 * - "False 0-bits" can cause problems, so they must be avoided. In the
65 * extreme case, a mask of all 0-bits is only correct if the classifier
66 * contains only a single flow that matches all packets.
68 * - 0-bits are desirable because they allow the datapath to act more
69 * autonomously, relying less on ovs-vswitchd to process flow setups,
70 * thereby improving performance.
72 * - We don't know a good way to generate wildcard masks with the maximum
73 * (correct) number of 0-bits. We use various approximations, described
76 * - Wildcard masks for lookups in a given classifier yield a
77 * non-overlapping set of rules. More specifically:
79 * Consider an classifier C1 filled with an arbitrary collection of rules
80 * and an empty classifier C2. Now take a set of packet headers H and
81 * look it up in C1, yielding a highest-priority matching rule R1 and
82 * wildcard mask M. Form a new classifier rule R2 out of packet headers
83 * H and mask M, and add R2 to C2 with a fixed priority. If one were to
84 * do this for every possible set of packet headers H, then this
85 * process would not attempt to add any overlapping rules to C2, that is,
86 * any packet lookup using the rules generated by this process matches at
87 * most one rule in C2.
89 * During the lookup process, the classifier starts out with a wildcard mask
90 * that is all 0-bits, that is, fully wildcarded. As lookup proceeds, each
91 * step tends to add constraints to the wildcard mask, that is, change
92 * wildcarded 0-bits into exact-match 1-bits. We call this "un-wildcarding".
93 * A lookup step that examines a particular field must un-wildcard that field.
94 * In general, un-wildcarding is necessary for correctness but undesirable for
98 * Basic Classifier Design
99 * =======================
101 * Suppose that all the rules in a classifier had the same form. For example,
102 * suppose that they all matched on the source and destination Ethernet address
103 * and wildcarded all the other fields. Then the obvious way to implement a
104 * classifier would be a hash table on the source and destination Ethernet
105 * addresses. If new classification rules came along with a different form,
106 * you could add a second hash table that hashed on the fields matched in those
107 * rules. With two hash tables, you look up a given flow in each hash table.
108 * If there are no matches, the classifier didn't contain a match; if you find
109 * a match in one of them, that's the result; if you find a match in both of
110 * them, then the result is the rule with the higher priority.
112 * This is how the classifier works. In a "struct classifier", each form of
113 * "struct cls_rule" present (based on its ->match.mask) goes into a separate
114 * "struct cls_subtable". A lookup does a hash lookup in every "struct
115 * cls_subtable" in the classifier and tracks the highest-priority match that
116 * it finds. The subtables are kept in a descending priority order according
117 * to the highest priority rule in each subtable, which allows lookup to skip
118 * over subtables that can't possibly have a higher-priority match than already
119 * found. Eliminating lookups through priority ordering aids both classifier
120 * primary design goals: skipping lookups saves time and avoids un-wildcarding
121 * fields that those lookups would have examined.
123 * One detail: a classifier can contain multiple rules that are identical other
124 * than their priority. When this happens, only the highest priority rule out
125 * of a group of otherwise identical rules is stored directly in the "struct
126 * cls_subtable", with the other almost-identical rules chained off a linked
127 * list inside that highest-priority rule.
129 * The following sub-sections describe various optimizations over this simple
133 * Staged Lookup (Wildcard Optimization)
134 * -------------------------------------
136 * Subtable lookup is performed in ranges defined for struct flow, starting
137 * from metadata (registers, in_port, etc.), then L2 header, L3, and finally
138 * L4 ports. Whenever it is found that there are no matches in the current
139 * subtable, the rest of the subtable can be skipped.
141 * Staged lookup does not reduce lookup time, and it may increase it, because
142 * it changes a single hash table lookup into multiple hash table lookups.
143 * It reduces un-wildcarding significantly in important use cases.
146 * Prefix Tracking (Wildcard Optimization)
147 * ---------------------------------------
149 * Classifier uses prefix trees ("tries") for tracking the used
150 * address space, enabling skipping classifier tables containing
151 * longer masks than necessary for the given address. This reduces
152 * un-wildcarding for datapath flows in parts of the address space
153 * without host routes, but consulting extra data structures (the
154 * tries) may slightly increase lookup time.
156 * Trie lookup is interwoven with staged lookup, so that a trie is
157 * searched only when the configured trie field becomes relevant for
158 * the lookup. The trie lookup results are retained so that each trie
159 * is checked at most once for each classifier lookup.
161 * This implementation tracks the number of rules at each address
162 * prefix for the whole classifier. More aggressive table skipping
163 * would be possible by maintaining lists of tables that have prefixes
164 * at the lengths encountered on tree traversal, or by maintaining
165 * separate tries for subsets of rules separated by metadata fields.
167 * Prefix tracking is configured via OVSDB "Flow_Table" table,
168 * "fieldspec" column. "fieldspec" is a string map where a "prefix"
169 * key tells which fields should be used for prefix tracking. The
170 * value of the "prefix" key is a comma separated list of field names.
172 * There is a maximum number of fields that can be enabled for any one
173 * flow table. Currently this limit is 3.
176 * Partitioning (Lookup Time and Wildcard Optimization)
177 * ----------------------------------------------------
179 * Suppose that a given classifier is being used to handle multiple stages in a
180 * pipeline using "resubmit", with metadata (that is, the OpenFlow 1.1+ field
181 * named "metadata") distinguishing between the different stages. For example,
182 * metadata value 1 might identify ingress rules, metadata value 2 might
183 * identify ACLs, and metadata value 3 might identify egress rules. Such a
184 * classifier is essentially partitioned into multiple sub-classifiers on the
185 * basis of the metadata value.
187 * The classifier has a special optimization to speed up matching in this
190 * - Each cls_subtable that matches on metadata gets a tag derived from the
191 * subtable's mask, so that it is likely that each subtable has a unique
192 * tag. (Duplicate tags have a performance cost but do not affect
195 * - For each metadata value matched by any cls_rule, the classifier
196 * constructs a "struct cls_partition" indexed by the metadata value.
197 * The cls_partition has a 'tags' member whose value is the bitwise-OR of
198 * the tags of each cls_subtable that contains any rule that matches on
199 * the cls_partition's metadata value. In other words, struct
200 * cls_partition associates metadata values with subtables that need to
201 * be checked with flows with that specific metadata value.
203 * Thus, a flow lookup can start by looking up the partition associated with
204 * the flow's metadata, and then skip over any cls_subtable whose 'tag' does
205 * not intersect the partition's 'tags'. (The flow must also be looked up in
206 * any cls_subtable that doesn't match on metadata. We handle that by giving
207 * any such cls_subtable TAG_ALL as its 'tags' so that it matches any tag.)
209 * Partitioning saves lookup time by reducing the number of subtable lookups.
210 * Each eliminated subtable lookup also reduces the amount of un-wildcarding.
213 * Classifier Versioning
214 * =====================
216 * Classifier lookups are always done in a specific classifier version, where
217 * a version is defined to be a natural number.
219 * When a new rule is added to a classifier, it is set to become visible in a
220 * specific version. If the version number used at insert time is larger than
221 * any version number currently used in lookups, the new rule is said to be
222 * invisible to lookups. This means that lookups won't find the rule, but the
223 * rule is immediately available to classifier iterations.
225 * Similarly, a rule can be marked as to be deleted in a future version, or
226 * more precisely, to be visible upto a given version number. To delete a rule
227 * in a way to not remove the rule before all ongoing lookups are finished, the
228 * rule should be marked as "to be deleted" by setting the rule's visibility to
229 * the negation of the last version number in which it should be visible.
230 * Then, when all the lookups use a later version number, the rule can be
231 * actually deleted from the classifier. A rule that is marked for deletion
232 * after a future version will not appear in iterations, although it will still
233 * be found by lookups using a lookup version number up to that future version
236 * Classifiers can hold duplicate rules (rules with the same match criteria and
237 * priority) when at most one of the duplicates with the same priority is
238 * visible in any given lookup version. The caller responsible for classifier
239 * modifications must maintain this invariant.
241 * The classifier supports versioning for two reasons:
243 * 1. Support for versioned modifications makes it possible to perform an
244 * arbitraty series of classifier changes as one atomic transaction,
245 * where intermediate versions of the classifier are not visible to any
246 * lookups. Also, when a rule is added for a future version, or marked
247 * for removal after the current version, such modifications can be
248 * reverted without any visible effects to any of the current lookups.
250 * 2. Performance: Adding (or deleting) a large set of rules can, in
251 * pathological cases, have a cost proportional to the number of rules
252 * already in the classifier. When multiple rules are being added (or
253 * deleted) in one go, though, this pathological case cost can be
254 * typically avoided, as long as it is OK for any new rules to be
255 * invisible until the batch change is complete.
257 * Note that the classifier_replace() function replaces a rule immediately, and
258 * is therefore not safe to use with versioning. It is still available for the
259 * users that do not use versioning.
262 * Deferred Publication
263 * ====================
265 * Removing large number of rules from classifier can be costly, as the
266 * supporting data structures are teared down, in many cases just to be
267 * re-instantiated right after. In the worst case, as when each rule has a
268 * different match pattern (mask), the maintenance of the match patterns can
269 * have cost O(N^2), where N is the number of different match patterns. To
270 * alleviate this, the classifier supports a "deferred mode", in which changes
271 * in internal data structures needed for future version lookups may not be
272 * fully computed yet. The computation is finalized when the deferred mode is
275 * This feature can be used with versioning such that all changes to future
276 * versions are made in the deferred mode. Then, right before making the new
277 * version visible to lookups, the deferred mode is turned off so that all the
278 * data structures are ready for lookups with the new version number.
280 * To use deferred publication, first call classifier_defer(). Then, modify
281 * the classifier via additions (classifier_insert() with a specific, future
282 * version number) and deletions (use cls_rule_make_removable_after_version()).
283 * Then call classifier_publish(), and after that, announce the new version
284 * number to be used in lookups.
290 * The classifier may safely be accessed by many reader threads concurrently
291 * and by a single writer, or by multiple writers when they guarantee mutually
292 * exlucive access to classifier modifications.
294 * Since the classifier rules are RCU protected, the rule destruction after
295 * removal from the classifier must be RCU postponed. Also, when versioning is
296 * used, the rule removal itself needs to be typically RCU postponed. In this
297 * case the rule destruction is doubly RCU postponed, i.e., the second
298 * ovsrcu_postpone() call to destruct the rule is called from the first RCU
299 * callback that removes the rule.
301 * Rules that have never been visible to lookups are an exeption to the above
302 * rule. Such rules can be removed immediately, but their destruction must
303 * still be RCU postponed, as the rule's visibility attribute may be examined
304 * parallel to the rule's removal. */
308 #include "meta-flow.h"
316 /* Classifier internal data structures. */
321 typedef OVSRCU_TYPE(struct trie_node *) rcu_trie_ptr;
323 /* Prefix trie for a 'field' */
325 const struct mf_field *field; /* Trie field, or NULL. */
326 rcu_trie_ptr root; /* NULL if none. */
330 CLS_MIN_VERSION = 1, /* Default version number to use. */
331 CLS_MAX_VERSION = LLONG_MAX, /* Last possible version number. */
332 CLS_MAX_INDICES = 3, /* Maximum number of lookup indices per subtable. */
333 CLS_MAX_TRIES = 3 /* Maximum number of prefix trees per classifier. */
336 /* A flow classifier. */
338 int n_rules; /* Total number of rules. */
339 uint8_t n_flow_segments;
340 uint8_t flow_segments[CLS_MAX_INDICES]; /* Flow segment boundaries to use
341 * for staged lookup. */
342 struct cmap subtables_map; /* Contains "struct cls_subtable"s. */
343 struct pvector subtables;
344 struct cmap partitions; /* Contains "struct cls_partition"s. */
345 struct cls_trie tries[CLS_MAX_TRIES]; /* Prefix tries. */
346 unsigned int n_tries;
347 bool publish; /* Make changes visible to lookups? */
350 struct cls_conjunction {
356 /* A rule to be inserted to the classifier. */
358 struct rculist node; /* In struct cls_subtable 'rules_list'. */
359 const int priority; /* Larger numbers are higher priorities. */
360 const long long version; /* Version in which the rule was added. */
361 struct cls_match *cls_match; /* NULL if not in a classifier. */
362 const struct minimatch match; /* Matching rule. */
365 void cls_rule_init(struct cls_rule *, const struct match *, int priority,
367 void cls_rule_init_from_minimatch(struct cls_rule *, const struct minimatch *,
368 int priority, long long version);
369 void cls_rule_clone(struct cls_rule *, const struct cls_rule *);
370 void cls_rule_move(struct cls_rule *dst, struct cls_rule *src);
371 void cls_rule_destroy(struct cls_rule *);
373 void cls_rule_set_conjunctions(struct cls_rule *,
374 const struct cls_conjunction *, size_t n);
376 bool cls_rule_equal(const struct cls_rule *, const struct cls_rule *);
377 uint32_t cls_rule_hash(const struct cls_rule *, uint32_t basis);
378 void cls_rule_format(const struct cls_rule *, struct ds *);
379 bool cls_rule_is_catchall(const struct cls_rule *);
380 bool cls_rule_is_loose_match(const struct cls_rule *rule,
381 const struct minimatch *criteria);
382 bool cls_rule_visible_in_version(const struct cls_rule *, long long version);
383 void cls_rule_make_invisible_in_version(const struct cls_rule *,
385 long long lookup_version);
386 void cls_rule_restore_visibility(const struct cls_rule *);
388 /* Constructor/destructor. Must run single-threaded. */
389 void classifier_init(struct classifier *, const uint8_t *flow_segments);
390 void classifier_destroy(struct classifier *);
392 /* Modifiers. Caller MUST exclude concurrent calls from other threads. */
393 bool classifier_set_prefix_fields(struct classifier *,
394 const enum mf_field_id *trie_fields,
395 unsigned int n_trie_fields);
396 void classifier_insert(struct classifier *, const struct cls_rule *,
397 const struct cls_conjunction *, size_t n_conjunctions);
398 const struct cls_rule *classifier_replace(struct classifier *,
399 const struct cls_rule *,
400 const struct cls_conjunction *,
401 size_t n_conjunctions);
402 const struct cls_rule *classifier_remove(struct classifier *,
403 const struct cls_rule *);
404 static inline void classifier_defer(struct classifier *);
405 static inline void classifier_publish(struct classifier *);
407 /* Lookups. These are RCU protected and may run concurrently with modifiers
409 const struct cls_rule *classifier_lookup(const struct classifier *,
410 long long version, struct flow *,
411 struct flow_wildcards *);
412 bool classifier_rule_overlaps(const struct classifier *,
413 const struct cls_rule *);
414 const struct cls_rule *classifier_find_rule_exactly(const struct classifier *,
415 const struct cls_rule *);
416 const struct cls_rule *classifier_find_match_exactly(const struct classifier *,
417 const struct match *,
420 bool classifier_is_empty(const struct classifier *);
421 int classifier_count(const struct classifier *);
425 * Iteration is lockless and RCU-protected. Concurrent threads may perform all
426 * kinds of concurrent modifications without ruining the iteration. Obviously,
427 * any modifications may or may not be visible to the concurrent iterator, but
428 * all the rules not deleted are visited by the iteration. The iterating
429 * thread may also modify the classifier rules itself.
431 * 'TARGET' iteration only iterates rules matching the 'TARGET' criteria.
432 * Rather than looping through all the rules and skipping ones that can't
433 * match, 'TARGET' iteration skips whole subtables, if the 'TARGET' happens to
434 * be more specific than the subtable. */
436 const struct classifier *cls;
437 const struct cls_subtable *subtable;
438 const struct cls_rule *target;
439 struct pvector_cursor subtables;
440 const struct cls_rule *rule;
443 struct cls_cursor cls_cursor_start(const struct classifier *cls,
444 const struct cls_rule *target);
445 void cls_cursor_advance(struct cls_cursor *);
447 #define CLS_FOR_EACH(RULE, MEMBER, CLS) \
448 CLS_FOR_EACH_TARGET(RULE, MEMBER, CLS, NULL)
449 #define CLS_FOR_EACH_TARGET(RULE, MEMBER, CLS, TARGET) \
450 for (struct cls_cursor cursor__ = cls_cursor_start(CLS, TARGET); \
452 ? (INIT_CONTAINER(RULE, cursor__.rule, MEMBER), \
453 cls_cursor_advance(&cursor__), \
463 classifier_defer(struct classifier *cls)
465 cls->publish = false;
469 classifier_publish(struct classifier *cls)
472 pvector_publish(&cls->subtables);
474 #endif /* classifier.h */